You cannot select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
FreeRTOS-Kernel/portable/ARMv8M/non_secure/port.c

2140 lines
93 KiB
C

/*
* FreeRTOS Kernel <DEVELOPMENT BRANCH>
* Copyright (C) 2021 Amazon.com, Inc. or its affiliates. All Rights Reserved.
*
* SPDX-License-Identifier: MIT
*
* Permission is hereby granted, free of charge, to any person obtaining a copy of
* this software and associated documentation files (the "Software"), to deal in
* the Software without restriction, including without limitation the rights to
* use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
* the Software, and to permit persons to whom the Software is furnished to do so,
* subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
* FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
* COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
* IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
* CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
*
* https://www.FreeRTOS.org
* https://github.com/FreeRTOS
*
*/
/* Defining MPU_WRAPPERS_INCLUDED_FROM_API_FILE prevents task.h from redefining
* all the API functions to use the MPU wrappers. That should only be done when
* task.h is included from an application file. */
#define MPU_WRAPPERS_INCLUDED_FROM_API_FILE
/* Scheduler includes. */
#include "FreeRTOS.h"
#include "task.h"
/* MPU wrappers includes. */
#include "mpu_wrappers.h"
/* Portasm includes. */
#include "portasm.h"
#if ( configENABLE_TRUSTZONE == 1 )
/* Secure components includes. */
#include "secure_context.h"
#include "secure_init.h"
#endif /* configENABLE_TRUSTZONE */
#undef MPU_WRAPPERS_INCLUDED_FROM_API_FILE
/**
* The FreeRTOS Cortex M33 port can be configured to run on the Secure Side only
* i.e. the processor boots as secure and never jumps to the non-secure side.
* The Trust Zone support in the port must be disabled in order to run FreeRTOS
* on the secure side. The following are the valid configuration seetings:
*
* 1. Run FreeRTOS on the Secure Side:
* configRUN_FREERTOS_SECURE_ONLY = 1 and configENABLE_TRUSTZONE = 0
*
* 2. Run FreeRTOS on the Non-Secure Side with Secure Side function call support:
* configRUN_FREERTOS_SECURE_ONLY = 0 and configENABLE_TRUSTZONE = 1
*
* 3. Run FreeRTOS on the Non-Secure Side only i.e. no Secure Side function call support:
* configRUN_FREERTOS_SECURE_ONLY = 0 and configENABLE_TRUSTZONE = 0
*/
#if ( ( configRUN_FREERTOS_SECURE_ONLY == 1 ) && ( configENABLE_TRUSTZONE == 1 ) )
#error TrustZone needs to be disabled in order to run FreeRTOS on the Secure Side.
#endif
/**
* Cortex-M23 does not have non-secure PSPLIM. We should use PSPLIM on Cortex-M23
* only when FreeRTOS runs on secure side.
*/
#if ( ( portHAS_ARMV8M_MAIN_EXTENSION == 0 ) && ( configRUN_FREERTOS_SECURE_ONLY == 0 ) )
#define portUSE_PSPLIM_REGISTER 0
#else
#define portUSE_PSPLIM_REGISTER 1
#endif
/*-----------------------------------------------------------*/
/**
* @brief Constants required to manipulate the NVIC.
*/
#define portNVIC_SYSTICK_CTRL_REG ( *( ( volatile uint32_t * ) 0xe000e010 ) )
#define portNVIC_SYSTICK_LOAD_REG ( *( ( volatile uint32_t * ) 0xe000e014 ) )
#define portNVIC_SYSTICK_CURRENT_VALUE_REG ( *( ( volatile uint32_t * ) 0xe000e018 ) )
#define portNVIC_SHPR3_REG ( *( ( volatile uint32_t * ) 0xe000ed20 ) )
#define portNVIC_SYSTICK_ENABLE_BIT ( 1UL << 0UL )
#define portNVIC_SYSTICK_INT_BIT ( 1UL << 1UL )
Tickless idle fixes/improvement (#59) * Fix tickless idle when stopping systick on zero... ...and don't stop SysTick at all in the eAbortSleep case. Prior to this commit, if vPortSuppressTicksAndSleep() happens to stop the SysTick on zero, then after tickless idle ends, xTickCount advances one full tick more than the time that actually elapsed as measured by the SysTick. See "bug 1" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - Idle task calls vPortSuppressTicksAndSleep(xExpectedIdleTime = 2). [Doesn't have to be "2" -- could be any number.] - vPortSuppressTicksAndSleep() stops SysTick, and the current-count register happens to stop on zero. - SysTick ISR executes, setting xPendedTicks = 1 - vPortSuppressTicksAndSleep() masks interrupts and calls eTaskConfirmSleepModeStatus() which confirms the sleep operation. *** - vPortSuppressTicksAndSleep() configures SysTick for 1 full tick (xExpectedIdleTime - 1) plus the current-count register (which is 0) - One tick period elapses in sleep. - SysTick wakes CPU, ISR executes and increments xPendedTicks to 2. - vPortSuppressTicksAndSleep() calls vTaskStepTick(1), then returns. - Idle task resumes scheduler, which increments xTickCount twice (for xPendedTicks = 2) In the end, two ticks elapsed as measured by SysTick, but the code increments xTickCount three times. The root cause is that the code assumes the SysTick current-count register always contains the number of SysTick counts remaining in the current tick period. However, when the current-count register is zero, there are ulTimerCountsForOneTick counts remaining, not zero. This error is not the kind of time slippage normally associated with tickless idle. *** Note that a recent commit https://github.com/FreeRTOS/FreeRTOS-Kernel/commit/e1b98f0 results in eAbortSleep in this case, due to xPendedTicks != 0. That commit does mostly resolve this bug without specifically mentioning it, and without this commit. But that resolution allows the code in port.c not to directly address the special case of stopping SysTick on zero in any code or comments. That commit also generates additional instances of eAbortSleep, and a second purpose of this commit is to optimize how vPortSuppressTicksAndSleep() behaves for eAbortSleep, as noted below. This commit also includes an optimization to avoid stopping the SysTick when eTaskConfirmSleepModeStatus() returns eAbortSleep. This optimization belongs with this fix because the method of handling the SysTick being stopped on zero changes with this optimization. * Fix imminent tick rescheduled after tickless idle Prior to this commit, if something other than systick wakes the CPU from tickless idle, vPortSuppressTicksAndSleep() might cause xTickCount to increment once too many times. See "bug 2" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - CPU is sleeping in vPortSuppressTicksAndSleep() - Something other than the SysTick wakes the CPU. - vPortSuppressTicksAndSleep() calculates the number of SysTick counts until the next tick. The bug occurs only if this number is small. - vPortSuppressTicksAndSleep() puts this small number into the SysTick reload register, and starts SysTick. - vPortSuppressTicksAndSleep() calls vTaskStepTick() - While vTaskStepTick() executes, the SysTick expires. The ISR pends because interrupts are masked, and SysTick starts a 2nd period still based on the small number of counts in its reload register. This 2nd period is undesirable and is likely to cause the error noted below. - vPortSuppressTicksAndSleep() puts the normal tick duration into the SysTick's reload register. - vPortSuppressTicksAndSleep() unmasks interrupts before the SysTick starts a new period based on the new value in the reload register. [This is a race condition that can go either way, but for the bug to occur, the race must play out this way.] - The pending SysTick ISR executes and increments xPendedTicks. - The SysTick expires again, finishing the second very small period, and starts a new period this time based on the full tick duration. - The SysTick ISR increments xPendedTicks (or xTickCount) even though only a tiny fraction of a tick period has elapsed since the previous tick. The bug occurs when *two* consecutive small periods of the SysTick are both counted as ticks. The root cause is a race caused by the small SysTick period. If vPortSuppressTicksAndSleep() unmasks interrupts *after* the small period expires but *before* the SysTick starts a period based on the full tick period, then two small periods are counted as ticks when only one should be counted. The end result is xTickCount advancing nearly one full tick more than time actually elapsed as measured by the SysTick. This is not the kind of time slippage normally associated with tickless idle. After this commit the code starts the SysTick and then immediately modifies the reload register to ensure the very short cycle (if any) is conducted only once. This strategy requires special consideration for the build option that configures SysTick to use a divided clock. To avoid waiting around for the SysTick to load value from the reload register, the new code temporarily configures the SysTick to use the undivided clock. The resulting timing error is typical for tickless idle. The error (commonly known as drift or slippage in kernel time) caused by this strategy is equivalent to one or two counts in ulStoppedTimerCompensation. This commit also updates comments and #define symbols related to the SysTick clock option. The SysTick can optionally be clocked by a divided version of the CPU clock (commonly divide-by-8). The new code in this commit adjusts these comments and symbols to make them clearer and more useful in configurations that use the divided clock. The fix made in this commit requires the use of these symbols, as noted in the code comments. * Fix tickless idle with alternate systick clocking Prior to this commit, in configurations using the alternate SysTick clocking, vPortSuppressTicksAndSleep() might cause xTickCount to jump ahead as much as the entire expected idle time or fall behind as much as one full tick compared to time as measured by the SysTick. SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. The SysTick has a configuration option to be clocked by an alternate clock besides the core clock. This alternate clock is MCU dependent. Scenarios Fixed --------------- The new code in this commit handles the following scenarios that were not handled correctly prior to this commit. 1. Before the sleep, vPortSuppressTicksAndSleep() stops the SysTick on zero, long after SysTick reached zero. Prior to this commit, this scenario caused xTickCount to jump ahead one full tick for the same reason documented here: https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/0c7b04bd3a745c52151abebc882eed3f811c4c81 2. After the sleep, vPortSuppressTicksAndSleep() stops the SysTick before it loads the counter from the reload register. Prior to this commit, this scenario caused xTickCount to jump ahead by the entire expected idle time (xExpectedIdleTime) because the current-count register is zero before it loads from the reload register. 3. Prior to return, vPortSuppressTicksAndSleep() attempts to start a short SysTick period when the current SysTick clock cycle has a lot of time remaining. Prior to this commit, this scenario could cause xTickCount to fall behind by as much as nearly one full tick because the short SysTick cycle never started. Note that #3 is partially fixed by https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/967acc9b200d3d4beeb289d9da9e88798074b431 even though that commit addresses a different issue. So this commit completes the partial fix. * Improve comments and name of preprocessor symbol Add a note in the code comments that SysTick requests an interrupt when decrementing from 1 to 0, so that's why stopping SysTick on zero is a special case. Readers might unknowingly assume that SysTick requests an interrupt when wrapping from 0 back to the load-register value. Reconsider new "_SETTING" suffix since "_CONFIG" suffix seems more descriptive. The code relies on *both* of these preprocessor symbols: portNVIC_SYSTICK_CLK_BIT portNVIC_SYSTICK_CLK_BIT_CONFIG **new** A meaningful suffix is really helpful to distinguish the two symbols. * Revert introduction of 2nd name for NVIC register When I added portNVIC_ICSR_REG I didn't realize there was already a portNVIC_INT_CTRL_REG, which identifies the same register. Not good to have both. Note that portNVIC_INT_CTRL_REG is defined in portmacro.h and is already used in this file (port.c). * Replicate to other Cortex M ports Also set a new fiddle factor based on tests with a CM4F. I used gcc, optimizing at -O1. Users can fine-tune as needed. Also add configSYSTICK_CLOCK_HZ to the CM0 ports to be just like the other Cortex M ports. This change allowed uniformity in the default tickless implementations across all Cortex M ports. And CM0 is likely to benefit from configSYSTICK_CLOCK_HZ, especially considering new CM0 devices with very fast CPU clock speeds. * Revert changes to IAR-CM0-portmacro.h portNVIC_INT_CTRL_REG was already defined in port.c. No need to define it in portmacro.h. * Handle edge cases with slow SysTick clock Co-authored-by: Cobus van Eeden <35851496+cobusve@users.noreply.github.com> Co-authored-by: abhidixi11 <44424462+abhidixi11@users.noreply.github.com> Co-authored-by: Joseph Julicher <jjulicher@mac.com> Co-authored-by: alfred gedeon <28123637+alfred2g@users.noreply.github.com>
2 years ago
#define portNVIC_SYSTICK_CLK_BIT ( 1UL << 2UL )
#define portNVIC_SYSTICK_COUNT_FLAG_BIT ( 1UL << 16UL )
Tickless idle fixes/improvement (#59) * Fix tickless idle when stopping systick on zero... ...and don't stop SysTick at all in the eAbortSleep case. Prior to this commit, if vPortSuppressTicksAndSleep() happens to stop the SysTick on zero, then after tickless idle ends, xTickCount advances one full tick more than the time that actually elapsed as measured by the SysTick. See "bug 1" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - Idle task calls vPortSuppressTicksAndSleep(xExpectedIdleTime = 2). [Doesn't have to be "2" -- could be any number.] - vPortSuppressTicksAndSleep() stops SysTick, and the current-count register happens to stop on zero. - SysTick ISR executes, setting xPendedTicks = 1 - vPortSuppressTicksAndSleep() masks interrupts and calls eTaskConfirmSleepModeStatus() which confirms the sleep operation. *** - vPortSuppressTicksAndSleep() configures SysTick for 1 full tick (xExpectedIdleTime - 1) plus the current-count register (which is 0) - One tick period elapses in sleep. - SysTick wakes CPU, ISR executes and increments xPendedTicks to 2. - vPortSuppressTicksAndSleep() calls vTaskStepTick(1), then returns. - Idle task resumes scheduler, which increments xTickCount twice (for xPendedTicks = 2) In the end, two ticks elapsed as measured by SysTick, but the code increments xTickCount three times. The root cause is that the code assumes the SysTick current-count register always contains the number of SysTick counts remaining in the current tick period. However, when the current-count register is zero, there are ulTimerCountsForOneTick counts remaining, not zero. This error is not the kind of time slippage normally associated with tickless idle. *** Note that a recent commit https://github.com/FreeRTOS/FreeRTOS-Kernel/commit/e1b98f0 results in eAbortSleep in this case, due to xPendedTicks != 0. That commit does mostly resolve this bug without specifically mentioning it, and without this commit. But that resolution allows the code in port.c not to directly address the special case of stopping SysTick on zero in any code or comments. That commit also generates additional instances of eAbortSleep, and a second purpose of this commit is to optimize how vPortSuppressTicksAndSleep() behaves for eAbortSleep, as noted below. This commit also includes an optimization to avoid stopping the SysTick when eTaskConfirmSleepModeStatus() returns eAbortSleep. This optimization belongs with this fix because the method of handling the SysTick being stopped on zero changes with this optimization. * Fix imminent tick rescheduled after tickless idle Prior to this commit, if something other than systick wakes the CPU from tickless idle, vPortSuppressTicksAndSleep() might cause xTickCount to increment once too many times. See "bug 2" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - CPU is sleeping in vPortSuppressTicksAndSleep() - Something other than the SysTick wakes the CPU. - vPortSuppressTicksAndSleep() calculates the number of SysTick counts until the next tick. The bug occurs only if this number is small. - vPortSuppressTicksAndSleep() puts this small number into the SysTick reload register, and starts SysTick. - vPortSuppressTicksAndSleep() calls vTaskStepTick() - While vTaskStepTick() executes, the SysTick expires. The ISR pends because interrupts are masked, and SysTick starts a 2nd period still based on the small number of counts in its reload register. This 2nd period is undesirable and is likely to cause the error noted below. - vPortSuppressTicksAndSleep() puts the normal tick duration into the SysTick's reload register. - vPortSuppressTicksAndSleep() unmasks interrupts before the SysTick starts a new period based on the new value in the reload register. [This is a race condition that can go either way, but for the bug to occur, the race must play out this way.] - The pending SysTick ISR executes and increments xPendedTicks. - The SysTick expires again, finishing the second very small period, and starts a new period this time based on the full tick duration. - The SysTick ISR increments xPendedTicks (or xTickCount) even though only a tiny fraction of a tick period has elapsed since the previous tick. The bug occurs when *two* consecutive small periods of the SysTick are both counted as ticks. The root cause is a race caused by the small SysTick period. If vPortSuppressTicksAndSleep() unmasks interrupts *after* the small period expires but *before* the SysTick starts a period based on the full tick period, then two small periods are counted as ticks when only one should be counted. The end result is xTickCount advancing nearly one full tick more than time actually elapsed as measured by the SysTick. This is not the kind of time slippage normally associated with tickless idle. After this commit the code starts the SysTick and then immediately modifies the reload register to ensure the very short cycle (if any) is conducted only once. This strategy requires special consideration for the build option that configures SysTick to use a divided clock. To avoid waiting around for the SysTick to load value from the reload register, the new code temporarily configures the SysTick to use the undivided clock. The resulting timing error is typical for tickless idle. The error (commonly known as drift or slippage in kernel time) caused by this strategy is equivalent to one or two counts in ulStoppedTimerCompensation. This commit also updates comments and #define symbols related to the SysTick clock option. The SysTick can optionally be clocked by a divided version of the CPU clock (commonly divide-by-8). The new code in this commit adjusts these comments and symbols to make them clearer and more useful in configurations that use the divided clock. The fix made in this commit requires the use of these symbols, as noted in the code comments. * Fix tickless idle with alternate systick clocking Prior to this commit, in configurations using the alternate SysTick clocking, vPortSuppressTicksAndSleep() might cause xTickCount to jump ahead as much as the entire expected idle time or fall behind as much as one full tick compared to time as measured by the SysTick. SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. The SysTick has a configuration option to be clocked by an alternate clock besides the core clock. This alternate clock is MCU dependent. Scenarios Fixed --------------- The new code in this commit handles the following scenarios that were not handled correctly prior to this commit. 1. Before the sleep, vPortSuppressTicksAndSleep() stops the SysTick on zero, long after SysTick reached zero. Prior to this commit, this scenario caused xTickCount to jump ahead one full tick for the same reason documented here: https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/0c7b04bd3a745c52151abebc882eed3f811c4c81 2. After the sleep, vPortSuppressTicksAndSleep() stops the SysTick before it loads the counter from the reload register. Prior to this commit, this scenario caused xTickCount to jump ahead by the entire expected idle time (xExpectedIdleTime) because the current-count register is zero before it loads from the reload register. 3. Prior to return, vPortSuppressTicksAndSleep() attempts to start a short SysTick period when the current SysTick clock cycle has a lot of time remaining. Prior to this commit, this scenario could cause xTickCount to fall behind by as much as nearly one full tick because the short SysTick cycle never started. Note that #3 is partially fixed by https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/967acc9b200d3d4beeb289d9da9e88798074b431 even though that commit addresses a different issue. So this commit completes the partial fix. * Improve comments and name of preprocessor symbol Add a note in the code comments that SysTick requests an interrupt when decrementing from 1 to 0, so that's why stopping SysTick on zero is a special case. Readers might unknowingly assume that SysTick requests an interrupt when wrapping from 0 back to the load-register value. Reconsider new "_SETTING" suffix since "_CONFIG" suffix seems more descriptive. The code relies on *both* of these preprocessor symbols: portNVIC_SYSTICK_CLK_BIT portNVIC_SYSTICK_CLK_BIT_CONFIG **new** A meaningful suffix is really helpful to distinguish the two symbols. * Revert introduction of 2nd name for NVIC register When I added portNVIC_ICSR_REG I didn't realize there was already a portNVIC_INT_CTRL_REG, which identifies the same register. Not good to have both. Note that portNVIC_INT_CTRL_REG is defined in portmacro.h and is already used in this file (port.c). * Replicate to other Cortex M ports Also set a new fiddle factor based on tests with a CM4F. I used gcc, optimizing at -O1. Users can fine-tune as needed. Also add configSYSTICK_CLOCK_HZ to the CM0 ports to be just like the other Cortex M ports. This change allowed uniformity in the default tickless implementations across all Cortex M ports. And CM0 is likely to benefit from configSYSTICK_CLOCK_HZ, especially considering new CM0 devices with very fast CPU clock speeds. * Revert changes to IAR-CM0-portmacro.h portNVIC_INT_CTRL_REG was already defined in port.c. No need to define it in portmacro.h. * Handle edge cases with slow SysTick clock Co-authored-by: Cobus van Eeden <35851496+cobusve@users.noreply.github.com> Co-authored-by: abhidixi11 <44424462+abhidixi11@users.noreply.github.com> Co-authored-by: Joseph Julicher <jjulicher@mac.com> Co-authored-by: alfred gedeon <28123637+alfred2g@users.noreply.github.com>
2 years ago
#define portNVIC_PEND_SYSTICK_CLEAR_BIT ( 1UL << 25UL )
#define portNVIC_PEND_SYSTICK_SET_BIT ( 1UL << 26UL )
#define portMIN_INTERRUPT_PRIORITY ( 255UL )
#define portNVIC_PENDSV_PRI ( portMIN_INTERRUPT_PRIORITY << 16UL )
#define portNVIC_SYSTICK_PRI ( portMIN_INTERRUPT_PRIORITY << 24UL )
/*-----------------------------------------------------------*/
/**
* @brief Constants required to manipulate the SCB.
*/
#define portSCB_SYS_HANDLER_CTRL_STATE_REG ( *( volatile uint32_t * ) 0xe000ed24 )
#define portSCB_MEM_FAULT_ENABLE_BIT ( 1UL << 16UL )
/*-----------------------------------------------------------*/
Armv8-M (except Cortex-M23) interrupt priority checking (#673) * Armv8-M: Formatting changes Signed-off-by: Devaraj Ranganna <devaraj.ranganna@arm.com> * Armv8-M: Add support for interrupt priority check FreeRTOS provides `FromISR` system calls which can be called directly from interrupt service routines. It is crucial that the priority of these ISRs is set to same or lower value (numerically higher) than that of `configMAX_SYSCALL_INTERRUPT_PRIORITY`. For more information refer to https://www.FreeRTOS.org/RTOS-Cortex-M3-M4.html. Add a check to trigger an assert when an ISR with priority higher (numerically lower) than `configMAX_SYSCALL_INTERRUPT_PRIORITY` calls `FromISR` system calls if `configASSERT` macro is defined. In addition, add a config option `configQEMU_DISABLE_INTERRUPT_PRIO_BITS_CHECK` to disable interrupt priority check while running on QEMU. Based on the discussion https://gitlab.com/qemu-project/qemu/-/issues/1122, The interrupt priority bits in QEMU do not match the real hardware. Therefore the assert that checks the number of implemented bits and __NVIC_PRIO_BITS will always fail. The config option `configQEMU_DISABLE_INTERRUPT_PRIO_BITS_CHECK` should be defined in the `FreeRTOSConfig.h` for QEMU targets. Signed-off-by: Devaraj Ranganna <devaraj.ranganna@arm.com> * Use SHPR2 for calculating interrupt priority bits This removes the dependency on the secure software to mark the interrupt as non-secure. Signed-off-by: Gaurav Aggarwal <aggarg@amazon.com> --------- Signed-off-by: Devaraj Ranganna <devaraj.ranganna@arm.com> Signed-off-by: Gaurav Aggarwal <aggarg@amazon.com> Co-authored-by: Gaurav Aggarwal <aggarg@amazon.com> Co-authored-by: Gaurav-Aggarwal-AWS <33462878+aggarg@users.noreply.github.com>
2 years ago
/**
* @brief Constants required to check the validity of an interrupt priority.
*/
#define portNVIC_SHPR2_REG ( *( ( volatile uint32_t * ) 0xE000ED1C ) )
#define portFIRST_USER_INTERRUPT_NUMBER ( 16 )
#define portNVIC_IP_REGISTERS_OFFSET_16 ( 0xE000E3F0 )
#define portAIRCR_REG ( *( ( volatile uint32_t * ) 0xE000ED0C ) )
#define portTOP_BIT_OF_BYTE ( ( uint8_t ) 0x80 )
#define portMAX_PRIGROUP_BITS ( ( uint8_t ) 7 )
#define portPRIORITY_GROUP_MASK ( 0x07UL << 8UL )
#define portPRIGROUP_SHIFT ( 8UL )
/*-----------------------------------------------------------*/
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
/**
* @brief Constants used during system call enter and exit.
*/
#define portPSR_STACK_PADDING_MASK ( 1UL << 9UL )
#define portEXC_RETURN_STACK_FRAME_TYPE_MASK ( 1UL << 4UL )
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
/*-----------------------------------------------------------*/
/**
* @brief Constants required to manipulate the FPU.
*/
#define portCPACR ( ( volatile uint32_t * ) 0xe000ed88 ) /* Coprocessor Access Control Register. */
#define portCPACR_CP10_VALUE ( 3UL )
#define portCPACR_CP11_VALUE portCPACR_CP10_VALUE
#define portCPACR_CP10_POS ( 20UL )
#define portCPACR_CP11_POS ( 22UL )
#define portFPCCR ( ( volatile uint32_t * ) 0xe000ef34 ) /* Floating Point Context Control Register. */
#define portFPCCR_ASPEN_POS ( 31UL )
#define portFPCCR_ASPEN_MASK ( 1UL << portFPCCR_ASPEN_POS )
#define portFPCCR_LSPEN_POS ( 30UL )
#define portFPCCR_LSPEN_MASK ( 1UL << portFPCCR_LSPEN_POS )
/*-----------------------------------------------------------*/
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
/**
* @brief Offsets in the stack to the parameters when inside the SVC handler.
*/
#define portOFFSET_TO_LR ( 5 )
#define portOFFSET_TO_PC ( 6 )
#define portOFFSET_TO_PSR ( 7 )
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
/*-----------------------------------------------------------*/
/**
* @brief Constants required to manipulate the MPU.
*/
#define portMPU_TYPE_REG ( *( ( volatile uint32_t * ) 0xe000ed90 ) )
#define portMPU_CTRL_REG ( *( ( volatile uint32_t * ) 0xe000ed94 ) )
#define portMPU_RNR_REG ( *( ( volatile uint32_t * ) 0xe000ed98 ) )
#define portMPU_RBAR_REG ( *( ( volatile uint32_t * ) 0xe000ed9c ) )
#define portMPU_RLAR_REG ( *( ( volatile uint32_t * ) 0xe000eda0 ) )
#define portMPU_RBAR_A1_REG ( *( ( volatile uint32_t * ) 0xe000eda4 ) )
#define portMPU_RLAR_A1_REG ( *( ( volatile uint32_t * ) 0xe000eda8 ) )
#define portMPU_RBAR_A2_REG ( *( ( volatile uint32_t * ) 0xe000edac ) )
#define portMPU_RLAR_A2_REG ( *( ( volatile uint32_t * ) 0xe000edb0 ) )
#define portMPU_RBAR_A3_REG ( *( ( volatile uint32_t * ) 0xe000edb4 ) )
#define portMPU_RLAR_A3_REG ( *( ( volatile uint32_t * ) 0xe000edb8 ) )
#define portMPU_MAIR0_REG ( *( ( volatile uint32_t * ) 0xe000edc0 ) )
#define portMPU_MAIR1_REG ( *( ( volatile uint32_t * ) 0xe000edc4 ) )
#define portMPU_RBAR_ADDRESS_MASK ( 0xffffffe0 ) /* Must be 32-byte aligned. */
#define portMPU_RLAR_ADDRESS_MASK ( 0xffffffe0 ) /* Must be 32-byte aligned. */
#define portMPU_RBAR_ACCESS_PERMISSIONS_MASK ( 3UL << 1UL )
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
#define portMPU_MAIR_ATTR0_POS ( 0UL )
#define portMPU_MAIR_ATTR0_MASK ( 0x000000ff )
#define portMPU_MAIR_ATTR1_POS ( 8UL )
#define portMPU_MAIR_ATTR1_MASK ( 0x0000ff00 )
#define portMPU_MAIR_ATTR2_POS ( 16UL )
#define portMPU_MAIR_ATTR2_MASK ( 0x00ff0000 )
#define portMPU_MAIR_ATTR3_POS ( 24UL )
#define portMPU_MAIR_ATTR3_MASK ( 0xff000000 )
#define portMPU_MAIR_ATTR4_POS ( 0UL )
#define portMPU_MAIR_ATTR4_MASK ( 0x000000ff )
#define portMPU_MAIR_ATTR5_POS ( 8UL )
#define portMPU_MAIR_ATTR5_MASK ( 0x0000ff00 )
#define portMPU_MAIR_ATTR6_POS ( 16UL )
#define portMPU_MAIR_ATTR6_MASK ( 0x00ff0000 )
#define portMPU_MAIR_ATTR7_POS ( 24UL )
#define portMPU_MAIR_ATTR7_MASK ( 0xff000000 )
#define portMPU_RLAR_ATTR_INDEX0 ( 0UL << 1UL )
#define portMPU_RLAR_ATTR_INDEX1 ( 1UL << 1UL )
#define portMPU_RLAR_ATTR_INDEX2 ( 2UL << 1UL )
#define portMPU_RLAR_ATTR_INDEX3 ( 3UL << 1UL )
#define portMPU_RLAR_ATTR_INDEX4 ( 4UL << 1UL )
#define portMPU_RLAR_ATTR_INDEX5 ( 5UL << 1UL )
#define portMPU_RLAR_ATTR_INDEX6 ( 6UL << 1UL )
#define portMPU_RLAR_ATTR_INDEX7 ( 7UL << 1UL )
#define portMPU_RLAR_REGION_ENABLE ( 1UL )
/* Enable privileged access to unmapped region. */
#define portMPU_PRIV_BACKGROUND_ENABLE_BIT ( 1UL << 2UL )
/* Enable MPU. */
#define portMPU_ENABLE_BIT ( 1UL << 0UL )
/* Expected value of the portMPU_TYPE register. */
#define portEXPECTED_MPU_TYPE_VALUE ( configTOTAL_MPU_REGIONS << 8UL )
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
/* Extract first address of the MPU region as encoded in the
* RBAR (Region Base Address Register) value. */
#define portEXTRACT_FIRST_ADDRESS_FROM_RBAR( rbar ) \
( ( rbar ) & portMPU_RBAR_ADDRESS_MASK )
/* Extract last address of the MPU region as encoded in the
* RLAR (Region Limit Address Register) value. */
#define portEXTRACT_LAST_ADDRESS_FROM_RLAR( rlar ) \
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
( ( ( rlar ) & portMPU_RLAR_ADDRESS_MASK ) | ~portMPU_RLAR_ADDRESS_MASK )
/* Does addr lies within [start, end] address range? */
#define portIS_ADDRESS_WITHIN_RANGE( addr, start, end ) \
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
( ( ( addr ) >= ( start ) ) && ( ( addr ) <= ( end ) ) )
/* Is the access request satisfied by the available permissions? */
#define portIS_AUTHORIZED( accessRequest, permissions ) \
( ( ( permissions ) & ( accessRequest ) ) == accessRequest )
/* Max value that fits in a uint32_t type. */
#define portUINT32_MAX ( ~( ( uint32_t ) 0 ) )
/* Check if adding a and b will result in overflow. */
#define portADD_UINT32_WILL_OVERFLOW( a, b ) ( ( a ) > ( portUINT32_MAX - ( b ) ) )
/*-----------------------------------------------------------*/
/**
* @brief The maximum 24-bit number.
*
* It is needed because the systick is a 24-bit counter.
*/
#define portMAX_24_BIT_NUMBER ( 0xffffffUL )
/**
* @brief A fiddle factor to estimate the number of SysTick counts that would
* have occurred while the SysTick counter is stopped during tickless idle
* calculations.
*/
Tickless idle fixes/improvement (#59) * Fix tickless idle when stopping systick on zero... ...and don't stop SysTick at all in the eAbortSleep case. Prior to this commit, if vPortSuppressTicksAndSleep() happens to stop the SysTick on zero, then after tickless idle ends, xTickCount advances one full tick more than the time that actually elapsed as measured by the SysTick. See "bug 1" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - Idle task calls vPortSuppressTicksAndSleep(xExpectedIdleTime = 2). [Doesn't have to be "2" -- could be any number.] - vPortSuppressTicksAndSleep() stops SysTick, and the current-count register happens to stop on zero. - SysTick ISR executes, setting xPendedTicks = 1 - vPortSuppressTicksAndSleep() masks interrupts and calls eTaskConfirmSleepModeStatus() which confirms the sleep operation. *** - vPortSuppressTicksAndSleep() configures SysTick for 1 full tick (xExpectedIdleTime - 1) plus the current-count register (which is 0) - One tick period elapses in sleep. - SysTick wakes CPU, ISR executes and increments xPendedTicks to 2. - vPortSuppressTicksAndSleep() calls vTaskStepTick(1), then returns. - Idle task resumes scheduler, which increments xTickCount twice (for xPendedTicks = 2) In the end, two ticks elapsed as measured by SysTick, but the code increments xTickCount three times. The root cause is that the code assumes the SysTick current-count register always contains the number of SysTick counts remaining in the current tick period. However, when the current-count register is zero, there are ulTimerCountsForOneTick counts remaining, not zero. This error is not the kind of time slippage normally associated with tickless idle. *** Note that a recent commit https://github.com/FreeRTOS/FreeRTOS-Kernel/commit/e1b98f0 results in eAbortSleep in this case, due to xPendedTicks != 0. That commit does mostly resolve this bug without specifically mentioning it, and without this commit. But that resolution allows the code in port.c not to directly address the special case of stopping SysTick on zero in any code or comments. That commit also generates additional instances of eAbortSleep, and a second purpose of this commit is to optimize how vPortSuppressTicksAndSleep() behaves for eAbortSleep, as noted below. This commit also includes an optimization to avoid stopping the SysTick when eTaskConfirmSleepModeStatus() returns eAbortSleep. This optimization belongs with this fix because the method of handling the SysTick being stopped on zero changes with this optimization. * Fix imminent tick rescheduled after tickless idle Prior to this commit, if something other than systick wakes the CPU from tickless idle, vPortSuppressTicksAndSleep() might cause xTickCount to increment once too many times. See "bug 2" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - CPU is sleeping in vPortSuppressTicksAndSleep() - Something other than the SysTick wakes the CPU. - vPortSuppressTicksAndSleep() calculates the number of SysTick counts until the next tick. The bug occurs only if this number is small. - vPortSuppressTicksAndSleep() puts this small number into the SysTick reload register, and starts SysTick. - vPortSuppressTicksAndSleep() calls vTaskStepTick() - While vTaskStepTick() executes, the SysTick expires. The ISR pends because interrupts are masked, and SysTick starts a 2nd period still based on the small number of counts in its reload register. This 2nd period is undesirable and is likely to cause the error noted below. - vPortSuppressTicksAndSleep() puts the normal tick duration into the SysTick's reload register. - vPortSuppressTicksAndSleep() unmasks interrupts before the SysTick starts a new period based on the new value in the reload register. [This is a race condition that can go either way, but for the bug to occur, the race must play out this way.] - The pending SysTick ISR executes and increments xPendedTicks. - The SysTick expires again, finishing the second very small period, and starts a new period this time based on the full tick duration. - The SysTick ISR increments xPendedTicks (or xTickCount) even though only a tiny fraction of a tick period has elapsed since the previous tick. The bug occurs when *two* consecutive small periods of the SysTick are both counted as ticks. The root cause is a race caused by the small SysTick period. If vPortSuppressTicksAndSleep() unmasks interrupts *after* the small period expires but *before* the SysTick starts a period based on the full tick period, then two small periods are counted as ticks when only one should be counted. The end result is xTickCount advancing nearly one full tick more than time actually elapsed as measured by the SysTick. This is not the kind of time slippage normally associated with tickless idle. After this commit the code starts the SysTick and then immediately modifies the reload register to ensure the very short cycle (if any) is conducted only once. This strategy requires special consideration for the build option that configures SysTick to use a divided clock. To avoid waiting around for the SysTick to load value from the reload register, the new code temporarily configures the SysTick to use the undivided clock. The resulting timing error is typical for tickless idle. The error (commonly known as drift or slippage in kernel time) caused by this strategy is equivalent to one or two counts in ulStoppedTimerCompensation. This commit also updates comments and #define symbols related to the SysTick clock option. The SysTick can optionally be clocked by a divided version of the CPU clock (commonly divide-by-8). The new code in this commit adjusts these comments and symbols to make them clearer and more useful in configurations that use the divided clock. The fix made in this commit requires the use of these symbols, as noted in the code comments. * Fix tickless idle with alternate systick clocking Prior to this commit, in configurations using the alternate SysTick clocking, vPortSuppressTicksAndSleep() might cause xTickCount to jump ahead as much as the entire expected idle time or fall behind as much as one full tick compared to time as measured by the SysTick. SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. The SysTick has a configuration option to be clocked by an alternate clock besides the core clock. This alternate clock is MCU dependent. Scenarios Fixed --------------- The new code in this commit handles the following scenarios that were not handled correctly prior to this commit. 1. Before the sleep, vPortSuppressTicksAndSleep() stops the SysTick on zero, long after SysTick reached zero. Prior to this commit, this scenario caused xTickCount to jump ahead one full tick for the same reason documented here: https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/0c7b04bd3a745c52151abebc882eed3f811c4c81 2. After the sleep, vPortSuppressTicksAndSleep() stops the SysTick before it loads the counter from the reload register. Prior to this commit, this scenario caused xTickCount to jump ahead by the entire expected idle time (xExpectedIdleTime) because the current-count register is zero before it loads from the reload register. 3. Prior to return, vPortSuppressTicksAndSleep() attempts to start a short SysTick period when the current SysTick clock cycle has a lot of time remaining. Prior to this commit, this scenario could cause xTickCount to fall behind by as much as nearly one full tick because the short SysTick cycle never started. Note that #3 is partially fixed by https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/967acc9b200d3d4beeb289d9da9e88798074b431 even though that commit addresses a different issue. So this commit completes the partial fix. * Improve comments and name of preprocessor symbol Add a note in the code comments that SysTick requests an interrupt when decrementing from 1 to 0, so that's why stopping SysTick on zero is a special case. Readers might unknowingly assume that SysTick requests an interrupt when wrapping from 0 back to the load-register value. Reconsider new "_SETTING" suffix since "_CONFIG" suffix seems more descriptive. The code relies on *both* of these preprocessor symbols: portNVIC_SYSTICK_CLK_BIT portNVIC_SYSTICK_CLK_BIT_CONFIG **new** A meaningful suffix is really helpful to distinguish the two symbols. * Revert introduction of 2nd name for NVIC register When I added portNVIC_ICSR_REG I didn't realize there was already a portNVIC_INT_CTRL_REG, which identifies the same register. Not good to have both. Note that portNVIC_INT_CTRL_REG is defined in portmacro.h and is already used in this file (port.c). * Replicate to other Cortex M ports Also set a new fiddle factor based on tests with a CM4F. I used gcc, optimizing at -O1. Users can fine-tune as needed. Also add configSYSTICK_CLOCK_HZ to the CM0 ports to be just like the other Cortex M ports. This change allowed uniformity in the default tickless implementations across all Cortex M ports. And CM0 is likely to benefit from configSYSTICK_CLOCK_HZ, especially considering new CM0 devices with very fast CPU clock speeds. * Revert changes to IAR-CM0-portmacro.h portNVIC_INT_CTRL_REG was already defined in port.c. No need to define it in portmacro.h. * Handle edge cases with slow SysTick clock Co-authored-by: Cobus van Eeden <35851496+cobusve@users.noreply.github.com> Co-authored-by: abhidixi11 <44424462+abhidixi11@users.noreply.github.com> Co-authored-by: Joseph Julicher <jjulicher@mac.com> Co-authored-by: alfred gedeon <28123637+alfred2g@users.noreply.github.com>
2 years ago
#define portMISSED_COUNTS_FACTOR ( 94UL )
/*-----------------------------------------------------------*/
/**
* @brief Constants required to set up the initial stack.
*/
#define portINITIAL_XPSR ( 0x01000000 )
#if ( configRUN_FREERTOS_SECURE_ONLY == 1 )
/**
* @brief Initial EXC_RETURN value.
*
* FF FF FF FD
* 1111 1111 1111 1111 1111 1111 1111 1101
*
* Bit[6] - 1 --> The exception was taken from the Secure state.
* Bit[5] - 1 --> Do not skip stacking of additional state context.
* Bit[4] - 1 --> The PE did not allocate space on the stack for FP context.
* Bit[3] - 1 --> Return to the Thread mode.
* Bit[2] - 1 --> Restore registers from the process stack.
* Bit[1] - 0 --> Reserved, 0.
* Bit[0] - 1 --> The exception was taken to the Secure state.
*/
#define portINITIAL_EXC_RETURN ( 0xfffffffd )
#else
/**
* @brief Initial EXC_RETURN value.
*
* FF FF FF BC
* 1111 1111 1111 1111 1111 1111 1011 1100
*
* Bit[6] - 0 --> The exception was taken from the Non-Secure state.
* Bit[5] - 1 --> Do not skip stacking of additional state context.
* Bit[4] - 1 --> The PE did not allocate space on the stack for FP context.
* Bit[3] - 1 --> Return to the Thread mode.
* Bit[2] - 1 --> Restore registers from the process stack.
* Bit[1] - 0 --> Reserved, 0.
* Bit[0] - 0 --> The exception was taken to the Non-Secure state.
*/
#define portINITIAL_EXC_RETURN ( 0xffffffbc )
#endif /* configRUN_FREERTOS_SECURE_ONLY */
/**
* @brief CONTROL register privileged bit mask.
*
* Bit[0] in CONTROL register tells the privilege:
* Bit[0] = 0 ==> The task is privileged.
* Bit[0] = 1 ==> The task is not privileged.
*/
#define portCONTROL_PRIVILEGED_MASK ( 1UL << 0UL )
/**
* @brief Initial CONTROL register values.
*/
#define portINITIAL_CONTROL_UNPRIVILEGED ( 0x3 )
#define portINITIAL_CONTROL_PRIVILEGED ( 0x2 )
Tickless idle fixes/improvement (#59) * Fix tickless idle when stopping systick on zero... ...and don't stop SysTick at all in the eAbortSleep case. Prior to this commit, if vPortSuppressTicksAndSleep() happens to stop the SysTick on zero, then after tickless idle ends, xTickCount advances one full tick more than the time that actually elapsed as measured by the SysTick. See "bug 1" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - Idle task calls vPortSuppressTicksAndSleep(xExpectedIdleTime = 2). [Doesn't have to be "2" -- could be any number.] - vPortSuppressTicksAndSleep() stops SysTick, and the current-count register happens to stop on zero. - SysTick ISR executes, setting xPendedTicks = 1 - vPortSuppressTicksAndSleep() masks interrupts and calls eTaskConfirmSleepModeStatus() which confirms the sleep operation. *** - vPortSuppressTicksAndSleep() configures SysTick for 1 full tick (xExpectedIdleTime - 1) plus the current-count register (which is 0) - One tick period elapses in sleep. - SysTick wakes CPU, ISR executes and increments xPendedTicks to 2. - vPortSuppressTicksAndSleep() calls vTaskStepTick(1), then returns. - Idle task resumes scheduler, which increments xTickCount twice (for xPendedTicks = 2) In the end, two ticks elapsed as measured by SysTick, but the code increments xTickCount three times. The root cause is that the code assumes the SysTick current-count register always contains the number of SysTick counts remaining in the current tick period. However, when the current-count register is zero, there are ulTimerCountsForOneTick counts remaining, not zero. This error is not the kind of time slippage normally associated with tickless idle. *** Note that a recent commit https://github.com/FreeRTOS/FreeRTOS-Kernel/commit/e1b98f0 results in eAbortSleep in this case, due to xPendedTicks != 0. That commit does mostly resolve this bug without specifically mentioning it, and without this commit. But that resolution allows the code in port.c not to directly address the special case of stopping SysTick on zero in any code or comments. That commit also generates additional instances of eAbortSleep, and a second purpose of this commit is to optimize how vPortSuppressTicksAndSleep() behaves for eAbortSleep, as noted below. This commit also includes an optimization to avoid stopping the SysTick when eTaskConfirmSleepModeStatus() returns eAbortSleep. This optimization belongs with this fix because the method of handling the SysTick being stopped on zero changes with this optimization. * Fix imminent tick rescheduled after tickless idle Prior to this commit, if something other than systick wakes the CPU from tickless idle, vPortSuppressTicksAndSleep() might cause xTickCount to increment once too many times. See "bug 2" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - CPU is sleeping in vPortSuppressTicksAndSleep() - Something other than the SysTick wakes the CPU. - vPortSuppressTicksAndSleep() calculates the number of SysTick counts until the next tick. The bug occurs only if this number is small. - vPortSuppressTicksAndSleep() puts this small number into the SysTick reload register, and starts SysTick. - vPortSuppressTicksAndSleep() calls vTaskStepTick() - While vTaskStepTick() executes, the SysTick expires. The ISR pends because interrupts are masked, and SysTick starts a 2nd period still based on the small number of counts in its reload register. This 2nd period is undesirable and is likely to cause the error noted below. - vPortSuppressTicksAndSleep() puts the normal tick duration into the SysTick's reload register. - vPortSuppressTicksAndSleep() unmasks interrupts before the SysTick starts a new period based on the new value in the reload register. [This is a race condition that can go either way, but for the bug to occur, the race must play out this way.] - The pending SysTick ISR executes and increments xPendedTicks. - The SysTick expires again, finishing the second very small period, and starts a new period this time based on the full tick duration. - The SysTick ISR increments xPendedTicks (or xTickCount) even though only a tiny fraction of a tick period has elapsed since the previous tick. The bug occurs when *two* consecutive small periods of the SysTick are both counted as ticks. The root cause is a race caused by the small SysTick period. If vPortSuppressTicksAndSleep() unmasks interrupts *after* the small period expires but *before* the SysTick starts a period based on the full tick period, then two small periods are counted as ticks when only one should be counted. The end result is xTickCount advancing nearly one full tick more than time actually elapsed as measured by the SysTick. This is not the kind of time slippage normally associated with tickless idle. After this commit the code starts the SysTick and then immediately modifies the reload register to ensure the very short cycle (if any) is conducted only once. This strategy requires special consideration for the build option that configures SysTick to use a divided clock. To avoid waiting around for the SysTick to load value from the reload register, the new code temporarily configures the SysTick to use the undivided clock. The resulting timing error is typical for tickless idle. The error (commonly known as drift or slippage in kernel time) caused by this strategy is equivalent to one or two counts in ulStoppedTimerCompensation. This commit also updates comments and #define symbols related to the SysTick clock option. The SysTick can optionally be clocked by a divided version of the CPU clock (commonly divide-by-8). The new code in this commit adjusts these comments and symbols to make them clearer and more useful in configurations that use the divided clock. The fix made in this commit requires the use of these symbols, as noted in the code comments. * Fix tickless idle with alternate systick clocking Prior to this commit, in configurations using the alternate SysTick clocking, vPortSuppressTicksAndSleep() might cause xTickCount to jump ahead as much as the entire expected idle time or fall behind as much as one full tick compared to time as measured by the SysTick. SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. The SysTick has a configuration option to be clocked by an alternate clock besides the core clock. This alternate clock is MCU dependent. Scenarios Fixed --------------- The new code in this commit handles the following scenarios that were not handled correctly prior to this commit. 1. Before the sleep, vPortSuppressTicksAndSleep() stops the SysTick on zero, long after SysTick reached zero. Prior to this commit, this scenario caused xTickCount to jump ahead one full tick for the same reason documented here: https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/0c7b04bd3a745c52151abebc882eed3f811c4c81 2. After the sleep, vPortSuppressTicksAndSleep() stops the SysTick before it loads the counter from the reload register. Prior to this commit, this scenario caused xTickCount to jump ahead by the entire expected idle time (xExpectedIdleTime) because the current-count register is zero before it loads from the reload register. 3. Prior to return, vPortSuppressTicksAndSleep() attempts to start a short SysTick period when the current SysTick clock cycle has a lot of time remaining. Prior to this commit, this scenario could cause xTickCount to fall behind by as much as nearly one full tick because the short SysTick cycle never started. Note that #3 is partially fixed by https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/967acc9b200d3d4beeb289d9da9e88798074b431 even though that commit addresses a different issue. So this commit completes the partial fix. * Improve comments and name of preprocessor symbol Add a note in the code comments that SysTick requests an interrupt when decrementing from 1 to 0, so that's why stopping SysTick on zero is a special case. Readers might unknowingly assume that SysTick requests an interrupt when wrapping from 0 back to the load-register value. Reconsider new "_SETTING" suffix since "_CONFIG" suffix seems more descriptive. The code relies on *both* of these preprocessor symbols: portNVIC_SYSTICK_CLK_BIT portNVIC_SYSTICK_CLK_BIT_CONFIG **new** A meaningful suffix is really helpful to distinguish the two symbols. * Revert introduction of 2nd name for NVIC register When I added portNVIC_ICSR_REG I didn't realize there was already a portNVIC_INT_CTRL_REG, which identifies the same register. Not good to have both. Note that portNVIC_INT_CTRL_REG is defined in portmacro.h and is already used in this file (port.c). * Replicate to other Cortex M ports Also set a new fiddle factor based on tests with a CM4F. I used gcc, optimizing at -O1. Users can fine-tune as needed. Also add configSYSTICK_CLOCK_HZ to the CM0 ports to be just like the other Cortex M ports. This change allowed uniformity in the default tickless implementations across all Cortex M ports. And CM0 is likely to benefit from configSYSTICK_CLOCK_HZ, especially considering new CM0 devices with very fast CPU clock speeds. * Revert changes to IAR-CM0-portmacro.h portNVIC_INT_CTRL_REG was already defined in port.c. No need to define it in portmacro.h. * Handle edge cases with slow SysTick clock Co-authored-by: Cobus van Eeden <35851496+cobusve@users.noreply.github.com> Co-authored-by: abhidixi11 <44424462+abhidixi11@users.noreply.github.com> Co-authored-by: Joseph Julicher <jjulicher@mac.com> Co-authored-by: alfred gedeon <28123637+alfred2g@users.noreply.github.com>
2 years ago
/**
* @brief Let the user override the default SysTick clock rate. If defined by the
* user, this symbol must equal the SysTick clock rate when the CLK bit is 0 in the
* configuration register.
*/
#ifndef configSYSTICK_CLOCK_HZ
#define configSYSTICK_CLOCK_HZ ( configCPU_CLOCK_HZ )
/* Ensure the SysTick is clocked at the same frequency as the core. */
#define portNVIC_SYSTICK_CLK_BIT_CONFIG ( portNVIC_SYSTICK_CLK_BIT )
#else
/* Select the option to clock SysTick not at the same frequency as the core. */
#define portNVIC_SYSTICK_CLK_BIT_CONFIG ( 0 )
#endif
/**
* @brief Let the user override the pre-loading of the initial LR with the
* address of prvTaskExitError() in case it messes up unwinding of the stack
* in the debugger.
*/
#ifdef configTASK_RETURN_ADDRESS
#define portTASK_RETURN_ADDRESS configTASK_RETURN_ADDRESS
#else
#define portTASK_RETURN_ADDRESS prvTaskExitError
#endif
/**
* @brief If portPRELOAD_REGISTERS then registers will be given an initial value
* when a task is created. This helps in debugging at the cost of code size.
*/
#define portPRELOAD_REGISTERS 1
/**
* @brief A task is created without a secure context, and must call
* portALLOCATE_SECURE_CONTEXT() to give itself a secure context before it makes
* any secure calls.
*/
#define portNO_SECURE_CONTEXT 0
/*-----------------------------------------------------------*/
/**
* @brief Used to catch tasks that attempt to return from their implementing
* function.
*/
static void prvTaskExitError( void );
#if ( configENABLE_MPU == 1 )
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
/**
* @brief Extract MPU region's access permissions from the Region Base Address
* Register (RBAR) value.
*
* @param ulRBARValue RBAR value for the MPU region.
*
* @return uint32_t Access permissions.
*/
static uint32_t prvGetRegionAccessPermissions( uint32_t ulRBARValue ) PRIVILEGED_FUNCTION;
#endif /* configENABLE_MPU */
#if ( configENABLE_MPU == 1 )
/**
* @brief Setup the Memory Protection Unit (MPU).
*/
static void prvSetupMPU( void ) PRIVILEGED_FUNCTION;
#endif /* configENABLE_MPU */
#if ( configENABLE_FPU == 1 )
/**
* @brief Setup the Floating Point Unit (FPU).
*/
static void prvSetupFPU( void ) PRIVILEGED_FUNCTION;
#endif /* configENABLE_FPU */
/**
* @brief Setup the timer to generate the tick interrupts.
*
* The implementation in this file is weak to allow application writers to
* change the timer used to generate the tick interrupt.
*/
void vPortSetupTimerInterrupt( void ) PRIVILEGED_FUNCTION;
/**
* @brief Checks whether the current execution context is interrupt.
*
* @return pdTRUE if the current execution context is interrupt, pdFALSE
* otherwise.
*/
BaseType_t xPortIsInsideInterrupt( void );
/**
* @brief Yield the processor.
*/
void vPortYield( void ) PRIVILEGED_FUNCTION;
/**
* @brief Enter critical section.
*/
void vPortEnterCritical( void ) PRIVILEGED_FUNCTION;
/**
* @brief Exit from critical section.
*/
void vPortExitCritical( void ) PRIVILEGED_FUNCTION;
/**
* @brief SysTick handler.
*/
void SysTick_Handler( void ) PRIVILEGED_FUNCTION;
/**
* @brief C part of SVC handler.
*/
portDONT_DISCARD void vPortSVCHandler_C( uint32_t * pulCallerStackAddress ) PRIVILEGED_FUNCTION;
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
#if ( ( configENABLE_MPU == 1 ) && ( configUSE_MPU_WRAPPERS_V1 == 0 ) )
/**
* @brief Sets up the system call stack so that upon returning from
* SVC, the system call stack is used.
*
* It is used for the system calls with up to 4 parameters.
*
* @param pulTaskStack The current SP when the SVC was raised.
* @param ulLR The value of Link Register (EXC_RETURN) in the SVC handler.
*/
void vSystemCallEnter( uint32_t * pulTaskStack,
uint32_t ulLR ) PRIVILEGED_FUNCTION;
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
#endif /* ( configENABLE_MPU == 1 ) && ( configUSE_MPU_WRAPPERS_V1 == 0 ) */
#if ( ( configENABLE_MPU == 1 ) && ( configUSE_MPU_WRAPPERS_V1 == 0 ) )
/**
* @brief Sets up the system call stack so that upon returning from
* SVC, the system call stack is used.
*
* It is used for the system calls with 5 parameters.
*
* @param pulTaskStack The current SP when the SVC was raised.
* @param ulLR The value of Link Register (EXC_RETURN) in the SVC handler.
*/
void vSystemCallEnter_1( uint32_t * pulTaskStack,
uint32_t ulLR ) PRIVILEGED_FUNCTION;
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
#endif /* ( configENABLE_MPU == 1 ) && ( configUSE_MPU_WRAPPERS_V1 == 0 ) */
#if ( ( configENABLE_MPU == 1 ) && ( configUSE_MPU_WRAPPERS_V1 == 0 ) )
/**
* @brief Sets up the task stack so that upon returning from
* SVC, the task stack is used again.
*
* @param pulSystemCallStack The current SP when the SVC was raised.
* @param ulLR The value of Link Register (EXC_RETURN) in the SVC handler.
*/
void vSystemCallExit( uint32_t * pulSystemCallStack,
uint32_t ulLR ) PRIVILEGED_FUNCTION;
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
#endif /* ( configENABLE_MPU == 1 ) && ( configUSE_MPU_WRAPPERS_V1 == 0 ) */
#if ( configENABLE_MPU == 1 )
/**
* @brief Checks whether or not the calling task is privileged.
*
* @return pdTRUE if the calling task is privileged, pdFALSE otherwise.
*/
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
BaseType_t xPortIsTaskPrivileged( void ) PRIVILEGED_FUNCTION;
#endif /* configENABLE_MPU == 1 */
/*-----------------------------------------------------------*/
/**
* @brief Each task maintains its own interrupt status in the critical nesting
* variable.
*/
PRIVILEGED_DATA static volatile uint32_t ulCriticalNesting = 0xaaaaaaaaUL;
#if ( configENABLE_TRUSTZONE == 1 )
/**
* @brief Saved as part of the task context to indicate which context the
* task is using on the secure side.
*/
PRIVILEGED_DATA portDONT_DISCARD volatile SecureContextHandle_t xSecureContext = portNO_SECURE_CONTEXT;
#endif /* configENABLE_TRUSTZONE */
Armv8-M (except Cortex-M23) interrupt priority checking (#673) * Armv8-M: Formatting changes Signed-off-by: Devaraj Ranganna <devaraj.ranganna@arm.com> * Armv8-M: Add support for interrupt priority check FreeRTOS provides `FromISR` system calls which can be called directly from interrupt service routines. It is crucial that the priority of these ISRs is set to same or lower value (numerically higher) than that of `configMAX_SYSCALL_INTERRUPT_PRIORITY`. For more information refer to https://www.FreeRTOS.org/RTOS-Cortex-M3-M4.html. Add a check to trigger an assert when an ISR with priority higher (numerically lower) than `configMAX_SYSCALL_INTERRUPT_PRIORITY` calls `FromISR` system calls if `configASSERT` macro is defined. In addition, add a config option `configQEMU_DISABLE_INTERRUPT_PRIO_BITS_CHECK` to disable interrupt priority check while running on QEMU. Based on the discussion https://gitlab.com/qemu-project/qemu/-/issues/1122, The interrupt priority bits in QEMU do not match the real hardware. Therefore the assert that checks the number of implemented bits and __NVIC_PRIO_BITS will always fail. The config option `configQEMU_DISABLE_INTERRUPT_PRIO_BITS_CHECK` should be defined in the `FreeRTOSConfig.h` for QEMU targets. Signed-off-by: Devaraj Ranganna <devaraj.ranganna@arm.com> * Use SHPR2 for calculating interrupt priority bits This removes the dependency on the secure software to mark the interrupt as non-secure. Signed-off-by: Gaurav Aggarwal <aggarg@amazon.com> --------- Signed-off-by: Devaraj Ranganna <devaraj.ranganna@arm.com> Signed-off-by: Gaurav Aggarwal <aggarg@amazon.com> Co-authored-by: Gaurav Aggarwal <aggarg@amazon.com> Co-authored-by: Gaurav-Aggarwal-AWS <33462878+aggarg@users.noreply.github.com>
2 years ago
/**
* @brief Used by the portASSERT_IF_INTERRUPT_PRIORITY_INVALID() macro to ensure
* FreeRTOS API functions are not called from interrupts that have been assigned
* a priority above configMAX_SYSCALL_INTERRUPT_PRIORITY.
*/
#if ( ( configASSERT_DEFINED == 1 ) && ( portHAS_ARMV8M_MAIN_EXTENSION == 1 ) )
Armv8-M (except Cortex-M23) interrupt priority checking (#673) * Armv8-M: Formatting changes Signed-off-by: Devaraj Ranganna <devaraj.ranganna@arm.com> * Armv8-M: Add support for interrupt priority check FreeRTOS provides `FromISR` system calls which can be called directly from interrupt service routines. It is crucial that the priority of these ISRs is set to same or lower value (numerically higher) than that of `configMAX_SYSCALL_INTERRUPT_PRIORITY`. For more information refer to https://www.FreeRTOS.org/RTOS-Cortex-M3-M4.html. Add a check to trigger an assert when an ISR with priority higher (numerically lower) than `configMAX_SYSCALL_INTERRUPT_PRIORITY` calls `FromISR` system calls if `configASSERT` macro is defined. In addition, add a config option `configQEMU_DISABLE_INTERRUPT_PRIO_BITS_CHECK` to disable interrupt priority check while running on QEMU. Based on the discussion https://gitlab.com/qemu-project/qemu/-/issues/1122, The interrupt priority bits in QEMU do not match the real hardware. Therefore the assert that checks the number of implemented bits and __NVIC_PRIO_BITS will always fail. The config option `configQEMU_DISABLE_INTERRUPT_PRIO_BITS_CHECK` should be defined in the `FreeRTOSConfig.h` for QEMU targets. Signed-off-by: Devaraj Ranganna <devaraj.ranganna@arm.com> * Use SHPR2 for calculating interrupt priority bits This removes the dependency on the secure software to mark the interrupt as non-secure. Signed-off-by: Gaurav Aggarwal <aggarg@amazon.com> --------- Signed-off-by: Devaraj Ranganna <devaraj.ranganna@arm.com> Signed-off-by: Gaurav Aggarwal <aggarg@amazon.com> Co-authored-by: Gaurav Aggarwal <aggarg@amazon.com> Co-authored-by: Gaurav-Aggarwal-AWS <33462878+aggarg@users.noreply.github.com>
2 years ago
static uint8_t ucMaxSysCallPriority = 0;
static uint32_t ulMaxPRIGROUPValue = 0;
static const volatile uint8_t * const pcInterruptPriorityRegisters = ( const volatile uint8_t * ) portNVIC_IP_REGISTERS_OFFSET_16;
#endif /* #if ( ( configASSERT_DEFINED == 1 ) && ( portHAS_ARMV8M_MAIN_EXTENSION == 1 ) ) */
Armv8-M (except Cortex-M23) interrupt priority checking (#673) * Armv8-M: Formatting changes Signed-off-by: Devaraj Ranganna <devaraj.ranganna@arm.com> * Armv8-M: Add support for interrupt priority check FreeRTOS provides `FromISR` system calls which can be called directly from interrupt service routines. It is crucial that the priority of these ISRs is set to same or lower value (numerically higher) than that of `configMAX_SYSCALL_INTERRUPT_PRIORITY`. For more information refer to https://www.FreeRTOS.org/RTOS-Cortex-M3-M4.html. Add a check to trigger an assert when an ISR with priority higher (numerically lower) than `configMAX_SYSCALL_INTERRUPT_PRIORITY` calls `FromISR` system calls if `configASSERT` macro is defined. In addition, add a config option `configQEMU_DISABLE_INTERRUPT_PRIO_BITS_CHECK` to disable interrupt priority check while running on QEMU. Based on the discussion https://gitlab.com/qemu-project/qemu/-/issues/1122, The interrupt priority bits in QEMU do not match the real hardware. Therefore the assert that checks the number of implemented bits and __NVIC_PRIO_BITS will always fail. The config option `configQEMU_DISABLE_INTERRUPT_PRIO_BITS_CHECK` should be defined in the `FreeRTOSConfig.h` for QEMU targets. Signed-off-by: Devaraj Ranganna <devaraj.ranganna@arm.com> * Use SHPR2 for calculating interrupt priority bits This removes the dependency on the secure software to mark the interrupt as non-secure. Signed-off-by: Gaurav Aggarwal <aggarg@amazon.com> --------- Signed-off-by: Devaraj Ranganna <devaraj.ranganna@arm.com> Signed-off-by: Gaurav Aggarwal <aggarg@amazon.com> Co-authored-by: Gaurav Aggarwal <aggarg@amazon.com> Co-authored-by: Gaurav-Aggarwal-AWS <33462878+aggarg@users.noreply.github.com>
2 years ago
#if ( configUSE_TICKLESS_IDLE == 1 )
/**
* @brief The number of SysTick increments that make up one tick period.
*/
PRIVILEGED_DATA static uint32_t ulTimerCountsForOneTick = 0;
/**
* @brief The maximum number of tick periods that can be suppressed is
* limited by the 24 bit resolution of the SysTick timer.
*/
PRIVILEGED_DATA static uint32_t xMaximumPossibleSuppressedTicks = 0;
/**
* @brief Compensate for the CPU cycles that pass while the SysTick is
* stopped (low power functionality only).
*/
PRIVILEGED_DATA static uint32_t ulStoppedTimerCompensation = 0;
#endif /* configUSE_TICKLESS_IDLE */
/*-----------------------------------------------------------*/
#if ( configUSE_TICKLESS_IDLE == 1 )
__attribute__( ( weak ) ) void vPortSuppressTicksAndSleep( TickType_t xExpectedIdleTime )
{
Tickless idle fixes/improvement (#59) * Fix tickless idle when stopping systick on zero... ...and don't stop SysTick at all in the eAbortSleep case. Prior to this commit, if vPortSuppressTicksAndSleep() happens to stop the SysTick on zero, then after tickless idle ends, xTickCount advances one full tick more than the time that actually elapsed as measured by the SysTick. See "bug 1" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - Idle task calls vPortSuppressTicksAndSleep(xExpectedIdleTime = 2). [Doesn't have to be "2" -- could be any number.] - vPortSuppressTicksAndSleep() stops SysTick, and the current-count register happens to stop on zero. - SysTick ISR executes, setting xPendedTicks = 1 - vPortSuppressTicksAndSleep() masks interrupts and calls eTaskConfirmSleepModeStatus() which confirms the sleep operation. *** - vPortSuppressTicksAndSleep() configures SysTick for 1 full tick (xExpectedIdleTime - 1) plus the current-count register (which is 0) - One tick period elapses in sleep. - SysTick wakes CPU, ISR executes and increments xPendedTicks to 2. - vPortSuppressTicksAndSleep() calls vTaskStepTick(1), then returns. - Idle task resumes scheduler, which increments xTickCount twice (for xPendedTicks = 2) In the end, two ticks elapsed as measured by SysTick, but the code increments xTickCount three times. The root cause is that the code assumes the SysTick current-count register always contains the number of SysTick counts remaining in the current tick period. However, when the current-count register is zero, there are ulTimerCountsForOneTick counts remaining, not zero. This error is not the kind of time slippage normally associated with tickless idle. *** Note that a recent commit https://github.com/FreeRTOS/FreeRTOS-Kernel/commit/e1b98f0 results in eAbortSleep in this case, due to xPendedTicks != 0. That commit does mostly resolve this bug without specifically mentioning it, and without this commit. But that resolution allows the code in port.c not to directly address the special case of stopping SysTick on zero in any code or comments. That commit also generates additional instances of eAbortSleep, and a second purpose of this commit is to optimize how vPortSuppressTicksAndSleep() behaves for eAbortSleep, as noted below. This commit also includes an optimization to avoid stopping the SysTick when eTaskConfirmSleepModeStatus() returns eAbortSleep. This optimization belongs with this fix because the method of handling the SysTick being stopped on zero changes with this optimization. * Fix imminent tick rescheduled after tickless idle Prior to this commit, if something other than systick wakes the CPU from tickless idle, vPortSuppressTicksAndSleep() might cause xTickCount to increment once too many times. See "bug 2" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - CPU is sleeping in vPortSuppressTicksAndSleep() - Something other than the SysTick wakes the CPU. - vPortSuppressTicksAndSleep() calculates the number of SysTick counts until the next tick. The bug occurs only if this number is small. - vPortSuppressTicksAndSleep() puts this small number into the SysTick reload register, and starts SysTick. - vPortSuppressTicksAndSleep() calls vTaskStepTick() - While vTaskStepTick() executes, the SysTick expires. The ISR pends because interrupts are masked, and SysTick starts a 2nd period still based on the small number of counts in its reload register. This 2nd period is undesirable and is likely to cause the error noted below. - vPortSuppressTicksAndSleep() puts the normal tick duration into the SysTick's reload register. - vPortSuppressTicksAndSleep() unmasks interrupts before the SysTick starts a new period based on the new value in the reload register. [This is a race condition that can go either way, but for the bug to occur, the race must play out this way.] - The pending SysTick ISR executes and increments xPendedTicks. - The SysTick expires again, finishing the second very small period, and starts a new period this time based on the full tick duration. - The SysTick ISR increments xPendedTicks (or xTickCount) even though only a tiny fraction of a tick period has elapsed since the previous tick. The bug occurs when *two* consecutive small periods of the SysTick are both counted as ticks. The root cause is a race caused by the small SysTick period. If vPortSuppressTicksAndSleep() unmasks interrupts *after* the small period expires but *before* the SysTick starts a period based on the full tick period, then two small periods are counted as ticks when only one should be counted. The end result is xTickCount advancing nearly one full tick more than time actually elapsed as measured by the SysTick. This is not the kind of time slippage normally associated with tickless idle. After this commit the code starts the SysTick and then immediately modifies the reload register to ensure the very short cycle (if any) is conducted only once. This strategy requires special consideration for the build option that configures SysTick to use a divided clock. To avoid waiting around for the SysTick to load value from the reload register, the new code temporarily configures the SysTick to use the undivided clock. The resulting timing error is typical for tickless idle. The error (commonly known as drift or slippage in kernel time) caused by this strategy is equivalent to one or two counts in ulStoppedTimerCompensation. This commit also updates comments and #define symbols related to the SysTick clock option. The SysTick can optionally be clocked by a divided version of the CPU clock (commonly divide-by-8). The new code in this commit adjusts these comments and symbols to make them clearer and more useful in configurations that use the divided clock. The fix made in this commit requires the use of these symbols, as noted in the code comments. * Fix tickless idle with alternate systick clocking Prior to this commit, in configurations using the alternate SysTick clocking, vPortSuppressTicksAndSleep() might cause xTickCount to jump ahead as much as the entire expected idle time or fall behind as much as one full tick compared to time as measured by the SysTick. SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. The SysTick has a configuration option to be clocked by an alternate clock besides the core clock. This alternate clock is MCU dependent. Scenarios Fixed --------------- The new code in this commit handles the following scenarios that were not handled correctly prior to this commit. 1. Before the sleep, vPortSuppressTicksAndSleep() stops the SysTick on zero, long after SysTick reached zero. Prior to this commit, this scenario caused xTickCount to jump ahead one full tick for the same reason documented here: https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/0c7b04bd3a745c52151abebc882eed3f811c4c81 2. After the sleep, vPortSuppressTicksAndSleep() stops the SysTick before it loads the counter from the reload register. Prior to this commit, this scenario caused xTickCount to jump ahead by the entire expected idle time (xExpectedIdleTime) because the current-count register is zero before it loads from the reload register. 3. Prior to return, vPortSuppressTicksAndSleep() attempts to start a short SysTick period when the current SysTick clock cycle has a lot of time remaining. Prior to this commit, this scenario could cause xTickCount to fall behind by as much as nearly one full tick because the short SysTick cycle never started. Note that #3 is partially fixed by https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/967acc9b200d3d4beeb289d9da9e88798074b431 even though that commit addresses a different issue. So this commit completes the partial fix. * Improve comments and name of preprocessor symbol Add a note in the code comments that SysTick requests an interrupt when decrementing from 1 to 0, so that's why stopping SysTick on zero is a special case. Readers might unknowingly assume that SysTick requests an interrupt when wrapping from 0 back to the load-register value. Reconsider new "_SETTING" suffix since "_CONFIG" suffix seems more descriptive. The code relies on *both* of these preprocessor symbols: portNVIC_SYSTICK_CLK_BIT portNVIC_SYSTICK_CLK_BIT_CONFIG **new** A meaningful suffix is really helpful to distinguish the two symbols. * Revert introduction of 2nd name for NVIC register When I added portNVIC_ICSR_REG I didn't realize there was already a portNVIC_INT_CTRL_REG, which identifies the same register. Not good to have both. Note that portNVIC_INT_CTRL_REG is defined in portmacro.h and is already used in this file (port.c). * Replicate to other Cortex M ports Also set a new fiddle factor based on tests with a CM4F. I used gcc, optimizing at -O1. Users can fine-tune as needed. Also add configSYSTICK_CLOCK_HZ to the CM0 ports to be just like the other Cortex M ports. This change allowed uniformity in the default tickless implementations across all Cortex M ports. And CM0 is likely to benefit from configSYSTICK_CLOCK_HZ, especially considering new CM0 devices with very fast CPU clock speeds. * Revert changes to IAR-CM0-portmacro.h portNVIC_INT_CTRL_REG was already defined in port.c. No need to define it in portmacro.h. * Handle edge cases with slow SysTick clock Co-authored-by: Cobus van Eeden <35851496+cobusve@users.noreply.github.com> Co-authored-by: abhidixi11 <44424462+abhidixi11@users.noreply.github.com> Co-authored-by: Joseph Julicher <jjulicher@mac.com> Co-authored-by: alfred gedeon <28123637+alfred2g@users.noreply.github.com>
2 years ago
uint32_t ulReloadValue, ulCompleteTickPeriods, ulCompletedSysTickDecrements, ulSysTickDecrementsLeft;
TickType_t xModifiableIdleTime;
/* Make sure the SysTick reload value does not overflow the counter. */
if( xExpectedIdleTime > xMaximumPossibleSuppressedTicks )
{
xExpectedIdleTime = xMaximumPossibleSuppressedTicks;
}
/* Enter a critical section but don't use the taskENTER_CRITICAL()
* method as that will mask interrupts that should exit sleep mode. */
__asm volatile ( "cpsid i" ::: "memory" );
__asm volatile ( "dsb" );
__asm volatile ( "isb" );
/* If a context switch is pending or a task is waiting for the scheduler
Tickless idle fixes/improvement (#59) * Fix tickless idle when stopping systick on zero... ...and don't stop SysTick at all in the eAbortSleep case. Prior to this commit, if vPortSuppressTicksAndSleep() happens to stop the SysTick on zero, then after tickless idle ends, xTickCount advances one full tick more than the time that actually elapsed as measured by the SysTick. See "bug 1" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - Idle task calls vPortSuppressTicksAndSleep(xExpectedIdleTime = 2). [Doesn't have to be "2" -- could be any number.] - vPortSuppressTicksAndSleep() stops SysTick, and the current-count register happens to stop on zero. - SysTick ISR executes, setting xPendedTicks = 1 - vPortSuppressTicksAndSleep() masks interrupts and calls eTaskConfirmSleepModeStatus() which confirms the sleep operation. *** - vPortSuppressTicksAndSleep() configures SysTick for 1 full tick (xExpectedIdleTime - 1) plus the current-count register (which is 0) - One tick period elapses in sleep. - SysTick wakes CPU, ISR executes and increments xPendedTicks to 2. - vPortSuppressTicksAndSleep() calls vTaskStepTick(1), then returns. - Idle task resumes scheduler, which increments xTickCount twice (for xPendedTicks = 2) In the end, two ticks elapsed as measured by SysTick, but the code increments xTickCount three times. The root cause is that the code assumes the SysTick current-count register always contains the number of SysTick counts remaining in the current tick period. However, when the current-count register is zero, there are ulTimerCountsForOneTick counts remaining, not zero. This error is not the kind of time slippage normally associated with tickless idle. *** Note that a recent commit https://github.com/FreeRTOS/FreeRTOS-Kernel/commit/e1b98f0 results in eAbortSleep in this case, due to xPendedTicks != 0. That commit does mostly resolve this bug without specifically mentioning it, and without this commit. But that resolution allows the code in port.c not to directly address the special case of stopping SysTick on zero in any code or comments. That commit also generates additional instances of eAbortSleep, and a second purpose of this commit is to optimize how vPortSuppressTicksAndSleep() behaves for eAbortSleep, as noted below. This commit also includes an optimization to avoid stopping the SysTick when eTaskConfirmSleepModeStatus() returns eAbortSleep. This optimization belongs with this fix because the method of handling the SysTick being stopped on zero changes with this optimization. * Fix imminent tick rescheduled after tickless idle Prior to this commit, if something other than systick wakes the CPU from tickless idle, vPortSuppressTicksAndSleep() might cause xTickCount to increment once too many times. See "bug 2" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - CPU is sleeping in vPortSuppressTicksAndSleep() - Something other than the SysTick wakes the CPU. - vPortSuppressTicksAndSleep() calculates the number of SysTick counts until the next tick. The bug occurs only if this number is small. - vPortSuppressTicksAndSleep() puts this small number into the SysTick reload register, and starts SysTick. - vPortSuppressTicksAndSleep() calls vTaskStepTick() - While vTaskStepTick() executes, the SysTick expires. The ISR pends because interrupts are masked, and SysTick starts a 2nd period still based on the small number of counts in its reload register. This 2nd period is undesirable and is likely to cause the error noted below. - vPortSuppressTicksAndSleep() puts the normal tick duration into the SysTick's reload register. - vPortSuppressTicksAndSleep() unmasks interrupts before the SysTick starts a new period based on the new value in the reload register. [This is a race condition that can go either way, but for the bug to occur, the race must play out this way.] - The pending SysTick ISR executes and increments xPendedTicks. - The SysTick expires again, finishing the second very small period, and starts a new period this time based on the full tick duration. - The SysTick ISR increments xPendedTicks (or xTickCount) even though only a tiny fraction of a tick period has elapsed since the previous tick. The bug occurs when *two* consecutive small periods of the SysTick are both counted as ticks. The root cause is a race caused by the small SysTick period. If vPortSuppressTicksAndSleep() unmasks interrupts *after* the small period expires but *before* the SysTick starts a period based on the full tick period, then two small periods are counted as ticks when only one should be counted. The end result is xTickCount advancing nearly one full tick more than time actually elapsed as measured by the SysTick. This is not the kind of time slippage normally associated with tickless idle. After this commit the code starts the SysTick and then immediately modifies the reload register to ensure the very short cycle (if any) is conducted only once. This strategy requires special consideration for the build option that configures SysTick to use a divided clock. To avoid waiting around for the SysTick to load value from the reload register, the new code temporarily configures the SysTick to use the undivided clock. The resulting timing error is typical for tickless idle. The error (commonly known as drift or slippage in kernel time) caused by this strategy is equivalent to one or two counts in ulStoppedTimerCompensation. This commit also updates comments and #define symbols related to the SysTick clock option. The SysTick can optionally be clocked by a divided version of the CPU clock (commonly divide-by-8). The new code in this commit adjusts these comments and symbols to make them clearer and more useful in configurations that use the divided clock. The fix made in this commit requires the use of these symbols, as noted in the code comments. * Fix tickless idle with alternate systick clocking Prior to this commit, in configurations using the alternate SysTick clocking, vPortSuppressTicksAndSleep() might cause xTickCount to jump ahead as much as the entire expected idle time or fall behind as much as one full tick compared to time as measured by the SysTick. SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. The SysTick has a configuration option to be clocked by an alternate clock besides the core clock. This alternate clock is MCU dependent. Scenarios Fixed --------------- The new code in this commit handles the following scenarios that were not handled correctly prior to this commit. 1. Before the sleep, vPortSuppressTicksAndSleep() stops the SysTick on zero, long after SysTick reached zero. Prior to this commit, this scenario caused xTickCount to jump ahead one full tick for the same reason documented here: https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/0c7b04bd3a745c52151abebc882eed3f811c4c81 2. After the sleep, vPortSuppressTicksAndSleep() stops the SysTick before it loads the counter from the reload register. Prior to this commit, this scenario caused xTickCount to jump ahead by the entire expected idle time (xExpectedIdleTime) because the current-count register is zero before it loads from the reload register. 3. Prior to return, vPortSuppressTicksAndSleep() attempts to start a short SysTick period when the current SysTick clock cycle has a lot of time remaining. Prior to this commit, this scenario could cause xTickCount to fall behind by as much as nearly one full tick because the short SysTick cycle never started. Note that #3 is partially fixed by https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/967acc9b200d3d4beeb289d9da9e88798074b431 even though that commit addresses a different issue. So this commit completes the partial fix. * Improve comments and name of preprocessor symbol Add a note in the code comments that SysTick requests an interrupt when decrementing from 1 to 0, so that's why stopping SysTick on zero is a special case. Readers might unknowingly assume that SysTick requests an interrupt when wrapping from 0 back to the load-register value. Reconsider new "_SETTING" suffix since "_CONFIG" suffix seems more descriptive. The code relies on *both* of these preprocessor symbols: portNVIC_SYSTICK_CLK_BIT portNVIC_SYSTICK_CLK_BIT_CONFIG **new** A meaningful suffix is really helpful to distinguish the two symbols. * Revert introduction of 2nd name for NVIC register When I added portNVIC_ICSR_REG I didn't realize there was already a portNVIC_INT_CTRL_REG, which identifies the same register. Not good to have both. Note that portNVIC_INT_CTRL_REG is defined in portmacro.h and is already used in this file (port.c). * Replicate to other Cortex M ports Also set a new fiddle factor based on tests with a CM4F. I used gcc, optimizing at -O1. Users can fine-tune as needed. Also add configSYSTICK_CLOCK_HZ to the CM0 ports to be just like the other Cortex M ports. This change allowed uniformity in the default tickless implementations across all Cortex M ports. And CM0 is likely to benefit from configSYSTICK_CLOCK_HZ, especially considering new CM0 devices with very fast CPU clock speeds. * Revert changes to IAR-CM0-portmacro.h portNVIC_INT_CTRL_REG was already defined in port.c. No need to define it in portmacro.h. * Handle edge cases with slow SysTick clock Co-authored-by: Cobus van Eeden <35851496+cobusve@users.noreply.github.com> Co-authored-by: abhidixi11 <44424462+abhidixi11@users.noreply.github.com> Co-authored-by: Joseph Julicher <jjulicher@mac.com> Co-authored-by: alfred gedeon <28123637+alfred2g@users.noreply.github.com>
2 years ago
* to be unsuspended then abandon the low power entry. */
if( eTaskConfirmSleepModeStatus() == eAbortSleep )
{
Tickless idle fixes/improvement (#59) * Fix tickless idle when stopping systick on zero... ...and don't stop SysTick at all in the eAbortSleep case. Prior to this commit, if vPortSuppressTicksAndSleep() happens to stop the SysTick on zero, then after tickless idle ends, xTickCount advances one full tick more than the time that actually elapsed as measured by the SysTick. See "bug 1" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - Idle task calls vPortSuppressTicksAndSleep(xExpectedIdleTime = 2). [Doesn't have to be "2" -- could be any number.] - vPortSuppressTicksAndSleep() stops SysTick, and the current-count register happens to stop on zero. - SysTick ISR executes, setting xPendedTicks = 1 - vPortSuppressTicksAndSleep() masks interrupts and calls eTaskConfirmSleepModeStatus() which confirms the sleep operation. *** - vPortSuppressTicksAndSleep() configures SysTick for 1 full tick (xExpectedIdleTime - 1) plus the current-count register (which is 0) - One tick period elapses in sleep. - SysTick wakes CPU, ISR executes and increments xPendedTicks to 2. - vPortSuppressTicksAndSleep() calls vTaskStepTick(1), then returns. - Idle task resumes scheduler, which increments xTickCount twice (for xPendedTicks = 2) In the end, two ticks elapsed as measured by SysTick, but the code increments xTickCount three times. The root cause is that the code assumes the SysTick current-count register always contains the number of SysTick counts remaining in the current tick period. However, when the current-count register is zero, there are ulTimerCountsForOneTick counts remaining, not zero. This error is not the kind of time slippage normally associated with tickless idle. *** Note that a recent commit https://github.com/FreeRTOS/FreeRTOS-Kernel/commit/e1b98f0 results in eAbortSleep in this case, due to xPendedTicks != 0. That commit does mostly resolve this bug without specifically mentioning it, and without this commit. But that resolution allows the code in port.c not to directly address the special case of stopping SysTick on zero in any code or comments. That commit also generates additional instances of eAbortSleep, and a second purpose of this commit is to optimize how vPortSuppressTicksAndSleep() behaves for eAbortSleep, as noted below. This commit also includes an optimization to avoid stopping the SysTick when eTaskConfirmSleepModeStatus() returns eAbortSleep. This optimization belongs with this fix because the method of handling the SysTick being stopped on zero changes with this optimization. * Fix imminent tick rescheduled after tickless idle Prior to this commit, if something other than systick wakes the CPU from tickless idle, vPortSuppressTicksAndSleep() might cause xTickCount to increment once too many times. See "bug 2" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - CPU is sleeping in vPortSuppressTicksAndSleep() - Something other than the SysTick wakes the CPU. - vPortSuppressTicksAndSleep() calculates the number of SysTick counts until the next tick. The bug occurs only if this number is small. - vPortSuppressTicksAndSleep() puts this small number into the SysTick reload register, and starts SysTick. - vPortSuppressTicksAndSleep() calls vTaskStepTick() - While vTaskStepTick() executes, the SysTick expires. The ISR pends because interrupts are masked, and SysTick starts a 2nd period still based on the small number of counts in its reload register. This 2nd period is undesirable and is likely to cause the error noted below. - vPortSuppressTicksAndSleep() puts the normal tick duration into the SysTick's reload register. - vPortSuppressTicksAndSleep() unmasks interrupts before the SysTick starts a new period based on the new value in the reload register. [This is a race condition that can go either way, but for the bug to occur, the race must play out this way.] - The pending SysTick ISR executes and increments xPendedTicks. - The SysTick expires again, finishing the second very small period, and starts a new period this time based on the full tick duration. - The SysTick ISR increments xPendedTicks (or xTickCount) even though only a tiny fraction of a tick period has elapsed since the previous tick. The bug occurs when *two* consecutive small periods of the SysTick are both counted as ticks. The root cause is a race caused by the small SysTick period. If vPortSuppressTicksAndSleep() unmasks interrupts *after* the small period expires but *before* the SysTick starts a period based on the full tick period, then two small periods are counted as ticks when only one should be counted. The end result is xTickCount advancing nearly one full tick more than time actually elapsed as measured by the SysTick. This is not the kind of time slippage normally associated with tickless idle. After this commit the code starts the SysTick and then immediately modifies the reload register to ensure the very short cycle (if any) is conducted only once. This strategy requires special consideration for the build option that configures SysTick to use a divided clock. To avoid waiting around for the SysTick to load value from the reload register, the new code temporarily configures the SysTick to use the undivided clock. The resulting timing error is typical for tickless idle. The error (commonly known as drift or slippage in kernel time) caused by this strategy is equivalent to one or two counts in ulStoppedTimerCompensation. This commit also updates comments and #define symbols related to the SysTick clock option. The SysTick can optionally be clocked by a divided version of the CPU clock (commonly divide-by-8). The new code in this commit adjusts these comments and symbols to make them clearer and more useful in configurations that use the divided clock. The fix made in this commit requires the use of these symbols, as noted in the code comments. * Fix tickless idle with alternate systick clocking Prior to this commit, in configurations using the alternate SysTick clocking, vPortSuppressTicksAndSleep() might cause xTickCount to jump ahead as much as the entire expected idle time or fall behind as much as one full tick compared to time as measured by the SysTick. SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. The SysTick has a configuration option to be clocked by an alternate clock besides the core clock. This alternate clock is MCU dependent. Scenarios Fixed --------------- The new code in this commit handles the following scenarios that were not handled correctly prior to this commit. 1. Before the sleep, vPortSuppressTicksAndSleep() stops the SysTick on zero, long after SysTick reached zero. Prior to this commit, this scenario caused xTickCount to jump ahead one full tick for the same reason documented here: https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/0c7b04bd3a745c52151abebc882eed3f811c4c81 2. After the sleep, vPortSuppressTicksAndSleep() stops the SysTick before it loads the counter from the reload register. Prior to this commit, this scenario caused xTickCount to jump ahead by the entire expected idle time (xExpectedIdleTime) because the current-count register is zero before it loads from the reload register. 3. Prior to return, vPortSuppressTicksAndSleep() attempts to start a short SysTick period when the current SysTick clock cycle has a lot of time remaining. Prior to this commit, this scenario could cause xTickCount to fall behind by as much as nearly one full tick because the short SysTick cycle never started. Note that #3 is partially fixed by https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/967acc9b200d3d4beeb289d9da9e88798074b431 even though that commit addresses a different issue. So this commit completes the partial fix. * Improve comments and name of preprocessor symbol Add a note in the code comments that SysTick requests an interrupt when decrementing from 1 to 0, so that's why stopping SysTick on zero is a special case. Readers might unknowingly assume that SysTick requests an interrupt when wrapping from 0 back to the load-register value. Reconsider new "_SETTING" suffix since "_CONFIG" suffix seems more descriptive. The code relies on *both* of these preprocessor symbols: portNVIC_SYSTICK_CLK_BIT portNVIC_SYSTICK_CLK_BIT_CONFIG **new** A meaningful suffix is really helpful to distinguish the two symbols. * Revert introduction of 2nd name for NVIC register When I added portNVIC_ICSR_REG I didn't realize there was already a portNVIC_INT_CTRL_REG, which identifies the same register. Not good to have both. Note that portNVIC_INT_CTRL_REG is defined in portmacro.h and is already used in this file (port.c). * Replicate to other Cortex M ports Also set a new fiddle factor based on tests with a CM4F. I used gcc, optimizing at -O1. Users can fine-tune as needed. Also add configSYSTICK_CLOCK_HZ to the CM0 ports to be just like the other Cortex M ports. This change allowed uniformity in the default tickless implementations across all Cortex M ports. And CM0 is likely to benefit from configSYSTICK_CLOCK_HZ, especially considering new CM0 devices with very fast CPU clock speeds. * Revert changes to IAR-CM0-portmacro.h portNVIC_INT_CTRL_REG was already defined in port.c. No need to define it in portmacro.h. * Handle edge cases with slow SysTick clock Co-authored-by: Cobus van Eeden <35851496+cobusve@users.noreply.github.com> Co-authored-by: abhidixi11 <44424462+abhidixi11@users.noreply.github.com> Co-authored-by: Joseph Julicher <jjulicher@mac.com> Co-authored-by: alfred gedeon <28123637+alfred2g@users.noreply.github.com>
2 years ago
/* Re-enable interrupts - see comments above the cpsid instruction
* above. */
__asm volatile ( "cpsie i" ::: "memory" );
}
else
{
Tickless idle fixes/improvement (#59) * Fix tickless idle when stopping systick on zero... ...and don't stop SysTick at all in the eAbortSleep case. Prior to this commit, if vPortSuppressTicksAndSleep() happens to stop the SysTick on zero, then after tickless idle ends, xTickCount advances one full tick more than the time that actually elapsed as measured by the SysTick. See "bug 1" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - Idle task calls vPortSuppressTicksAndSleep(xExpectedIdleTime = 2). [Doesn't have to be "2" -- could be any number.] - vPortSuppressTicksAndSleep() stops SysTick, and the current-count register happens to stop on zero. - SysTick ISR executes, setting xPendedTicks = 1 - vPortSuppressTicksAndSleep() masks interrupts and calls eTaskConfirmSleepModeStatus() which confirms the sleep operation. *** - vPortSuppressTicksAndSleep() configures SysTick for 1 full tick (xExpectedIdleTime - 1) plus the current-count register (which is 0) - One tick period elapses in sleep. - SysTick wakes CPU, ISR executes and increments xPendedTicks to 2. - vPortSuppressTicksAndSleep() calls vTaskStepTick(1), then returns. - Idle task resumes scheduler, which increments xTickCount twice (for xPendedTicks = 2) In the end, two ticks elapsed as measured by SysTick, but the code increments xTickCount three times. The root cause is that the code assumes the SysTick current-count register always contains the number of SysTick counts remaining in the current tick period. However, when the current-count register is zero, there are ulTimerCountsForOneTick counts remaining, not zero. This error is not the kind of time slippage normally associated with tickless idle. *** Note that a recent commit https://github.com/FreeRTOS/FreeRTOS-Kernel/commit/e1b98f0 results in eAbortSleep in this case, due to xPendedTicks != 0. That commit does mostly resolve this bug without specifically mentioning it, and without this commit. But that resolution allows the code in port.c not to directly address the special case of stopping SysTick on zero in any code or comments. That commit also generates additional instances of eAbortSleep, and a second purpose of this commit is to optimize how vPortSuppressTicksAndSleep() behaves for eAbortSleep, as noted below. This commit also includes an optimization to avoid stopping the SysTick when eTaskConfirmSleepModeStatus() returns eAbortSleep. This optimization belongs with this fix because the method of handling the SysTick being stopped on zero changes with this optimization. * Fix imminent tick rescheduled after tickless idle Prior to this commit, if something other than systick wakes the CPU from tickless idle, vPortSuppressTicksAndSleep() might cause xTickCount to increment once too many times. See "bug 2" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - CPU is sleeping in vPortSuppressTicksAndSleep() - Something other than the SysTick wakes the CPU. - vPortSuppressTicksAndSleep() calculates the number of SysTick counts until the next tick. The bug occurs only if this number is small. - vPortSuppressTicksAndSleep() puts this small number into the SysTick reload register, and starts SysTick. - vPortSuppressTicksAndSleep() calls vTaskStepTick() - While vTaskStepTick() executes, the SysTick expires. The ISR pends because interrupts are masked, and SysTick starts a 2nd period still based on the small number of counts in its reload register. This 2nd period is undesirable and is likely to cause the error noted below. - vPortSuppressTicksAndSleep() puts the normal tick duration into the SysTick's reload register. - vPortSuppressTicksAndSleep() unmasks interrupts before the SysTick starts a new period based on the new value in the reload register. [This is a race condition that can go either way, but for the bug to occur, the race must play out this way.] - The pending SysTick ISR executes and increments xPendedTicks. - The SysTick expires again, finishing the second very small period, and starts a new period this time based on the full tick duration. - The SysTick ISR increments xPendedTicks (or xTickCount) even though only a tiny fraction of a tick period has elapsed since the previous tick. The bug occurs when *two* consecutive small periods of the SysTick are both counted as ticks. The root cause is a race caused by the small SysTick period. If vPortSuppressTicksAndSleep() unmasks interrupts *after* the small period expires but *before* the SysTick starts a period based on the full tick period, then two small periods are counted as ticks when only one should be counted. The end result is xTickCount advancing nearly one full tick more than time actually elapsed as measured by the SysTick. This is not the kind of time slippage normally associated with tickless idle. After this commit the code starts the SysTick and then immediately modifies the reload register to ensure the very short cycle (if any) is conducted only once. This strategy requires special consideration for the build option that configures SysTick to use a divided clock. To avoid waiting around for the SysTick to load value from the reload register, the new code temporarily configures the SysTick to use the undivided clock. The resulting timing error is typical for tickless idle. The error (commonly known as drift or slippage in kernel time) caused by this strategy is equivalent to one or two counts in ulStoppedTimerCompensation. This commit also updates comments and #define symbols related to the SysTick clock option. The SysTick can optionally be clocked by a divided version of the CPU clock (commonly divide-by-8). The new code in this commit adjusts these comments and symbols to make them clearer and more useful in configurations that use the divided clock. The fix made in this commit requires the use of these symbols, as noted in the code comments. * Fix tickless idle with alternate systick clocking Prior to this commit, in configurations using the alternate SysTick clocking, vPortSuppressTicksAndSleep() might cause xTickCount to jump ahead as much as the entire expected idle time or fall behind as much as one full tick compared to time as measured by the SysTick. SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. The SysTick has a configuration option to be clocked by an alternate clock besides the core clock. This alternate clock is MCU dependent. Scenarios Fixed --------------- The new code in this commit handles the following scenarios that were not handled correctly prior to this commit. 1. Before the sleep, vPortSuppressTicksAndSleep() stops the SysTick on zero, long after SysTick reached zero. Prior to this commit, this scenario caused xTickCount to jump ahead one full tick for the same reason documented here: https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/0c7b04bd3a745c52151abebc882eed3f811c4c81 2. After the sleep, vPortSuppressTicksAndSleep() stops the SysTick before it loads the counter from the reload register. Prior to this commit, this scenario caused xTickCount to jump ahead by the entire expected idle time (xExpectedIdleTime) because the current-count register is zero before it loads from the reload register. 3. Prior to return, vPortSuppressTicksAndSleep() attempts to start a short SysTick period when the current SysTick clock cycle has a lot of time remaining. Prior to this commit, this scenario could cause xTickCount to fall behind by as much as nearly one full tick because the short SysTick cycle never started. Note that #3 is partially fixed by https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/967acc9b200d3d4beeb289d9da9e88798074b431 even though that commit addresses a different issue. So this commit completes the partial fix. * Improve comments and name of preprocessor symbol Add a note in the code comments that SysTick requests an interrupt when decrementing from 1 to 0, so that's why stopping SysTick on zero is a special case. Readers might unknowingly assume that SysTick requests an interrupt when wrapping from 0 back to the load-register value. Reconsider new "_SETTING" suffix since "_CONFIG" suffix seems more descriptive. The code relies on *both* of these preprocessor symbols: portNVIC_SYSTICK_CLK_BIT portNVIC_SYSTICK_CLK_BIT_CONFIG **new** A meaningful suffix is really helpful to distinguish the two symbols. * Revert introduction of 2nd name for NVIC register When I added portNVIC_ICSR_REG I didn't realize there was already a portNVIC_INT_CTRL_REG, which identifies the same register. Not good to have both. Note that portNVIC_INT_CTRL_REG is defined in portmacro.h and is already used in this file (port.c). * Replicate to other Cortex M ports Also set a new fiddle factor based on tests with a CM4F. I used gcc, optimizing at -O1. Users can fine-tune as needed. Also add configSYSTICK_CLOCK_HZ to the CM0 ports to be just like the other Cortex M ports. This change allowed uniformity in the default tickless implementations across all Cortex M ports. And CM0 is likely to benefit from configSYSTICK_CLOCK_HZ, especially considering new CM0 devices with very fast CPU clock speeds. * Revert changes to IAR-CM0-portmacro.h portNVIC_INT_CTRL_REG was already defined in port.c. No need to define it in portmacro.h. * Handle edge cases with slow SysTick clock Co-authored-by: Cobus van Eeden <35851496+cobusve@users.noreply.github.com> Co-authored-by: abhidixi11 <44424462+abhidixi11@users.noreply.github.com> Co-authored-by: Joseph Julicher <jjulicher@mac.com> Co-authored-by: alfred gedeon <28123637+alfred2g@users.noreply.github.com>
2 years ago
/* Stop the SysTick momentarily. The time the SysTick is stopped for
* is accounted for as best it can be, but using the tickless mode will
* inevitably result in some tiny drift of the time maintained by the
* kernel with respect to calendar time. */
portNVIC_SYSTICK_CTRL_REG = ( portNVIC_SYSTICK_CLK_BIT_CONFIG | portNVIC_SYSTICK_INT_BIT );
/* Use the SysTick current-value register to determine the number of
* SysTick decrements remaining until the next tick interrupt. If the
* current-value register is zero, then there are actually
* ulTimerCountsForOneTick decrements remaining, not zero, because the
* SysTick requests the interrupt when decrementing from 1 to 0. */
ulSysTickDecrementsLeft = portNVIC_SYSTICK_CURRENT_VALUE_REG;
if( ulSysTickDecrementsLeft == 0 )
{
ulSysTickDecrementsLeft = ulTimerCountsForOneTick;
}
/* Calculate the reload value required to wait xExpectedIdleTime
* tick periods. -1 is used because this code normally executes part
* way through the first tick period. But if the SysTick IRQ is now
* pending, then clear the IRQ, suppressing the first tick, and correct
* the reload value to reflect that the second tick period is already
* underway. The expected idle time is always at least two ticks. */
ulReloadValue = ulSysTickDecrementsLeft + ( ulTimerCountsForOneTick * ( xExpectedIdleTime - 1UL ) );
if( ( portNVIC_INT_CTRL_REG & portNVIC_PEND_SYSTICK_SET_BIT ) != 0 )
{
portNVIC_INT_CTRL_REG = portNVIC_PEND_SYSTICK_CLEAR_BIT;
ulReloadValue -= ulTimerCountsForOneTick;
}
if( ulReloadValue > ulStoppedTimerCompensation )
{
ulReloadValue -= ulStoppedTimerCompensation;
}
/* Set the new reload value. */
portNVIC_SYSTICK_LOAD_REG = ulReloadValue;
/* Clear the SysTick count flag and set the count value back to
* zero. */
portNVIC_SYSTICK_CURRENT_VALUE_REG = 0UL;
/* Restart SysTick. */
portNVIC_SYSTICK_CTRL_REG |= portNVIC_SYSTICK_ENABLE_BIT;
Tickless idle fixes/improvement (#59) * Fix tickless idle when stopping systick on zero... ...and don't stop SysTick at all in the eAbortSleep case. Prior to this commit, if vPortSuppressTicksAndSleep() happens to stop the SysTick on zero, then after tickless idle ends, xTickCount advances one full tick more than the time that actually elapsed as measured by the SysTick. See "bug 1" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - Idle task calls vPortSuppressTicksAndSleep(xExpectedIdleTime = 2). [Doesn't have to be "2" -- could be any number.] - vPortSuppressTicksAndSleep() stops SysTick, and the current-count register happens to stop on zero. - SysTick ISR executes, setting xPendedTicks = 1 - vPortSuppressTicksAndSleep() masks interrupts and calls eTaskConfirmSleepModeStatus() which confirms the sleep operation. *** - vPortSuppressTicksAndSleep() configures SysTick for 1 full tick (xExpectedIdleTime - 1) plus the current-count register (which is 0) - One tick period elapses in sleep. - SysTick wakes CPU, ISR executes and increments xPendedTicks to 2. - vPortSuppressTicksAndSleep() calls vTaskStepTick(1), then returns. - Idle task resumes scheduler, which increments xTickCount twice (for xPendedTicks = 2) In the end, two ticks elapsed as measured by SysTick, but the code increments xTickCount three times. The root cause is that the code assumes the SysTick current-count register always contains the number of SysTick counts remaining in the current tick period. However, when the current-count register is zero, there are ulTimerCountsForOneTick counts remaining, not zero. This error is not the kind of time slippage normally associated with tickless idle. *** Note that a recent commit https://github.com/FreeRTOS/FreeRTOS-Kernel/commit/e1b98f0 results in eAbortSleep in this case, due to xPendedTicks != 0. That commit does mostly resolve this bug without specifically mentioning it, and without this commit. But that resolution allows the code in port.c not to directly address the special case of stopping SysTick on zero in any code or comments. That commit also generates additional instances of eAbortSleep, and a second purpose of this commit is to optimize how vPortSuppressTicksAndSleep() behaves for eAbortSleep, as noted below. This commit also includes an optimization to avoid stopping the SysTick when eTaskConfirmSleepModeStatus() returns eAbortSleep. This optimization belongs with this fix because the method of handling the SysTick being stopped on zero changes with this optimization. * Fix imminent tick rescheduled after tickless idle Prior to this commit, if something other than systick wakes the CPU from tickless idle, vPortSuppressTicksAndSleep() might cause xTickCount to increment once too many times. See "bug 2" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - CPU is sleeping in vPortSuppressTicksAndSleep() - Something other than the SysTick wakes the CPU. - vPortSuppressTicksAndSleep() calculates the number of SysTick counts until the next tick. The bug occurs only if this number is small. - vPortSuppressTicksAndSleep() puts this small number into the SysTick reload register, and starts SysTick. - vPortSuppressTicksAndSleep() calls vTaskStepTick() - While vTaskStepTick() executes, the SysTick expires. The ISR pends because interrupts are masked, and SysTick starts a 2nd period still based on the small number of counts in its reload register. This 2nd period is undesirable and is likely to cause the error noted below. - vPortSuppressTicksAndSleep() puts the normal tick duration into the SysTick's reload register. - vPortSuppressTicksAndSleep() unmasks interrupts before the SysTick starts a new period based on the new value in the reload register. [This is a race condition that can go either way, but for the bug to occur, the race must play out this way.] - The pending SysTick ISR executes and increments xPendedTicks. - The SysTick expires again, finishing the second very small period, and starts a new period this time based on the full tick duration. - The SysTick ISR increments xPendedTicks (or xTickCount) even though only a tiny fraction of a tick period has elapsed since the previous tick. The bug occurs when *two* consecutive small periods of the SysTick are both counted as ticks. The root cause is a race caused by the small SysTick period. If vPortSuppressTicksAndSleep() unmasks interrupts *after* the small period expires but *before* the SysTick starts a period based on the full tick period, then two small periods are counted as ticks when only one should be counted. The end result is xTickCount advancing nearly one full tick more than time actually elapsed as measured by the SysTick. This is not the kind of time slippage normally associated with tickless idle. After this commit the code starts the SysTick and then immediately modifies the reload register to ensure the very short cycle (if any) is conducted only once. This strategy requires special consideration for the build option that configures SysTick to use a divided clock. To avoid waiting around for the SysTick to load value from the reload register, the new code temporarily configures the SysTick to use the undivided clock. The resulting timing error is typical for tickless idle. The error (commonly known as drift or slippage in kernel time) caused by this strategy is equivalent to one or two counts in ulStoppedTimerCompensation. This commit also updates comments and #define symbols related to the SysTick clock option. The SysTick can optionally be clocked by a divided version of the CPU clock (commonly divide-by-8). The new code in this commit adjusts these comments and symbols to make them clearer and more useful in configurations that use the divided clock. The fix made in this commit requires the use of these symbols, as noted in the code comments. * Fix tickless idle with alternate systick clocking Prior to this commit, in configurations using the alternate SysTick clocking, vPortSuppressTicksAndSleep() might cause xTickCount to jump ahead as much as the entire expected idle time or fall behind as much as one full tick compared to time as measured by the SysTick. SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. The SysTick has a configuration option to be clocked by an alternate clock besides the core clock. This alternate clock is MCU dependent. Scenarios Fixed --------------- The new code in this commit handles the following scenarios that were not handled correctly prior to this commit. 1. Before the sleep, vPortSuppressTicksAndSleep() stops the SysTick on zero, long after SysTick reached zero. Prior to this commit, this scenario caused xTickCount to jump ahead one full tick for the same reason documented here: https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/0c7b04bd3a745c52151abebc882eed3f811c4c81 2. After the sleep, vPortSuppressTicksAndSleep() stops the SysTick before it loads the counter from the reload register. Prior to this commit, this scenario caused xTickCount to jump ahead by the entire expected idle time (xExpectedIdleTime) because the current-count register is zero before it loads from the reload register. 3. Prior to return, vPortSuppressTicksAndSleep() attempts to start a short SysTick period when the current SysTick clock cycle has a lot of time remaining. Prior to this commit, this scenario could cause xTickCount to fall behind by as much as nearly one full tick because the short SysTick cycle never started. Note that #3 is partially fixed by https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/967acc9b200d3d4beeb289d9da9e88798074b431 even though that commit addresses a different issue. So this commit completes the partial fix. * Improve comments and name of preprocessor symbol Add a note in the code comments that SysTick requests an interrupt when decrementing from 1 to 0, so that's why stopping SysTick on zero is a special case. Readers might unknowingly assume that SysTick requests an interrupt when wrapping from 0 back to the load-register value. Reconsider new "_SETTING" suffix since "_CONFIG" suffix seems more descriptive. The code relies on *both* of these preprocessor symbols: portNVIC_SYSTICK_CLK_BIT portNVIC_SYSTICK_CLK_BIT_CONFIG **new** A meaningful suffix is really helpful to distinguish the two symbols. * Revert introduction of 2nd name for NVIC register When I added portNVIC_ICSR_REG I didn't realize there was already a portNVIC_INT_CTRL_REG, which identifies the same register. Not good to have both. Note that portNVIC_INT_CTRL_REG is defined in portmacro.h and is already used in this file (port.c). * Replicate to other Cortex M ports Also set a new fiddle factor based on tests with a CM4F. I used gcc, optimizing at -O1. Users can fine-tune as needed. Also add configSYSTICK_CLOCK_HZ to the CM0 ports to be just like the other Cortex M ports. This change allowed uniformity in the default tickless implementations across all Cortex M ports. And CM0 is likely to benefit from configSYSTICK_CLOCK_HZ, especially considering new CM0 devices with very fast CPU clock speeds. * Revert changes to IAR-CM0-portmacro.h portNVIC_INT_CTRL_REG was already defined in port.c. No need to define it in portmacro.h. * Handle edge cases with slow SysTick clock Co-authored-by: Cobus van Eeden <35851496+cobusve@users.noreply.github.com> Co-authored-by: abhidixi11 <44424462+abhidixi11@users.noreply.github.com> Co-authored-by: Joseph Julicher <jjulicher@mac.com> Co-authored-by: alfred gedeon <28123637+alfred2g@users.noreply.github.com>
2 years ago
/* Sleep until something happens. configPRE_SLEEP_PROCESSING() can
* set its parameter to 0 to indicate that its implementation contains
* its own wait for interrupt or wait for event instruction, and so wfi
* should not be executed again. However, the original expected idle
* time variable must remain unmodified, so a copy is taken. */
xModifiableIdleTime = xExpectedIdleTime;
configPRE_SLEEP_PROCESSING( xModifiableIdleTime );
if( xModifiableIdleTime > 0 )
{
__asm volatile ( "dsb" ::: "memory" );
__asm volatile ( "wfi" );
__asm volatile ( "isb" );
}
configPOST_SLEEP_PROCESSING( xExpectedIdleTime );
/* Re-enable interrupts to allow the interrupt that brought the MCU
Tickless idle fixes/improvement (#59) * Fix tickless idle when stopping systick on zero... ...and don't stop SysTick at all in the eAbortSleep case. Prior to this commit, if vPortSuppressTicksAndSleep() happens to stop the SysTick on zero, then after tickless idle ends, xTickCount advances one full tick more than the time that actually elapsed as measured by the SysTick. See "bug 1" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - Idle task calls vPortSuppressTicksAndSleep(xExpectedIdleTime = 2). [Doesn't have to be "2" -- could be any number.] - vPortSuppressTicksAndSleep() stops SysTick, and the current-count register happens to stop on zero. - SysTick ISR executes, setting xPendedTicks = 1 - vPortSuppressTicksAndSleep() masks interrupts and calls eTaskConfirmSleepModeStatus() which confirms the sleep operation. *** - vPortSuppressTicksAndSleep() configures SysTick for 1 full tick (xExpectedIdleTime - 1) plus the current-count register (which is 0) - One tick period elapses in sleep. - SysTick wakes CPU, ISR executes and increments xPendedTicks to 2. - vPortSuppressTicksAndSleep() calls vTaskStepTick(1), then returns. - Idle task resumes scheduler, which increments xTickCount twice (for xPendedTicks = 2) In the end, two ticks elapsed as measured by SysTick, but the code increments xTickCount three times. The root cause is that the code assumes the SysTick current-count register always contains the number of SysTick counts remaining in the current tick period. However, when the current-count register is zero, there are ulTimerCountsForOneTick counts remaining, not zero. This error is not the kind of time slippage normally associated with tickless idle. *** Note that a recent commit https://github.com/FreeRTOS/FreeRTOS-Kernel/commit/e1b98f0 results in eAbortSleep in this case, due to xPendedTicks != 0. That commit does mostly resolve this bug without specifically mentioning it, and without this commit. But that resolution allows the code in port.c not to directly address the special case of stopping SysTick on zero in any code or comments. That commit also generates additional instances of eAbortSleep, and a second purpose of this commit is to optimize how vPortSuppressTicksAndSleep() behaves for eAbortSleep, as noted below. This commit also includes an optimization to avoid stopping the SysTick when eTaskConfirmSleepModeStatus() returns eAbortSleep. This optimization belongs with this fix because the method of handling the SysTick being stopped on zero changes with this optimization. * Fix imminent tick rescheduled after tickless idle Prior to this commit, if something other than systick wakes the CPU from tickless idle, vPortSuppressTicksAndSleep() might cause xTickCount to increment once too many times. See "bug 2" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - CPU is sleeping in vPortSuppressTicksAndSleep() - Something other than the SysTick wakes the CPU. - vPortSuppressTicksAndSleep() calculates the number of SysTick counts until the next tick. The bug occurs only if this number is small. - vPortSuppressTicksAndSleep() puts this small number into the SysTick reload register, and starts SysTick. - vPortSuppressTicksAndSleep() calls vTaskStepTick() - While vTaskStepTick() executes, the SysTick expires. The ISR pends because interrupts are masked, and SysTick starts a 2nd period still based on the small number of counts in its reload register. This 2nd period is undesirable and is likely to cause the error noted below. - vPortSuppressTicksAndSleep() puts the normal tick duration into the SysTick's reload register. - vPortSuppressTicksAndSleep() unmasks interrupts before the SysTick starts a new period based on the new value in the reload register. [This is a race condition that can go either way, but for the bug to occur, the race must play out this way.] - The pending SysTick ISR executes and increments xPendedTicks. - The SysTick expires again, finishing the second very small period, and starts a new period this time based on the full tick duration. - The SysTick ISR increments xPendedTicks (or xTickCount) even though only a tiny fraction of a tick period has elapsed since the previous tick. The bug occurs when *two* consecutive small periods of the SysTick are both counted as ticks. The root cause is a race caused by the small SysTick period. If vPortSuppressTicksAndSleep() unmasks interrupts *after* the small period expires but *before* the SysTick starts a period based on the full tick period, then two small periods are counted as ticks when only one should be counted. The end result is xTickCount advancing nearly one full tick more than time actually elapsed as measured by the SysTick. This is not the kind of time slippage normally associated with tickless idle. After this commit the code starts the SysTick and then immediately modifies the reload register to ensure the very short cycle (if any) is conducted only once. This strategy requires special consideration for the build option that configures SysTick to use a divided clock. To avoid waiting around for the SysTick to load value from the reload register, the new code temporarily configures the SysTick to use the undivided clock. The resulting timing error is typical for tickless idle. The error (commonly known as drift or slippage in kernel time) caused by this strategy is equivalent to one or two counts in ulStoppedTimerCompensation. This commit also updates comments and #define symbols related to the SysTick clock option. The SysTick can optionally be clocked by a divided version of the CPU clock (commonly divide-by-8). The new code in this commit adjusts these comments and symbols to make them clearer and more useful in configurations that use the divided clock. The fix made in this commit requires the use of these symbols, as noted in the code comments. * Fix tickless idle with alternate systick clocking Prior to this commit, in configurations using the alternate SysTick clocking, vPortSuppressTicksAndSleep() might cause xTickCount to jump ahead as much as the entire expected idle time or fall behind as much as one full tick compared to time as measured by the SysTick. SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. The SysTick has a configuration option to be clocked by an alternate clock besides the core clock. This alternate clock is MCU dependent. Scenarios Fixed --------------- The new code in this commit handles the following scenarios that were not handled correctly prior to this commit. 1. Before the sleep, vPortSuppressTicksAndSleep() stops the SysTick on zero, long after SysTick reached zero. Prior to this commit, this scenario caused xTickCount to jump ahead one full tick for the same reason documented here: https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/0c7b04bd3a745c52151abebc882eed3f811c4c81 2. After the sleep, vPortSuppressTicksAndSleep() stops the SysTick before it loads the counter from the reload register. Prior to this commit, this scenario caused xTickCount to jump ahead by the entire expected idle time (xExpectedIdleTime) because the current-count register is zero before it loads from the reload register. 3. Prior to return, vPortSuppressTicksAndSleep() attempts to start a short SysTick period when the current SysTick clock cycle has a lot of time remaining. Prior to this commit, this scenario could cause xTickCount to fall behind by as much as nearly one full tick because the short SysTick cycle never started. Note that #3 is partially fixed by https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/967acc9b200d3d4beeb289d9da9e88798074b431 even though that commit addresses a different issue. So this commit completes the partial fix. * Improve comments and name of preprocessor symbol Add a note in the code comments that SysTick requests an interrupt when decrementing from 1 to 0, so that's why stopping SysTick on zero is a special case. Readers might unknowingly assume that SysTick requests an interrupt when wrapping from 0 back to the load-register value. Reconsider new "_SETTING" suffix since "_CONFIG" suffix seems more descriptive. The code relies on *both* of these preprocessor symbols: portNVIC_SYSTICK_CLK_BIT portNVIC_SYSTICK_CLK_BIT_CONFIG **new** A meaningful suffix is really helpful to distinguish the two symbols. * Revert introduction of 2nd name for NVIC register When I added portNVIC_ICSR_REG I didn't realize there was already a portNVIC_INT_CTRL_REG, which identifies the same register. Not good to have both. Note that portNVIC_INT_CTRL_REG is defined in portmacro.h and is already used in this file (port.c). * Replicate to other Cortex M ports Also set a new fiddle factor based on tests with a CM4F. I used gcc, optimizing at -O1. Users can fine-tune as needed. Also add configSYSTICK_CLOCK_HZ to the CM0 ports to be just like the other Cortex M ports. This change allowed uniformity in the default tickless implementations across all Cortex M ports. And CM0 is likely to benefit from configSYSTICK_CLOCK_HZ, especially considering new CM0 devices with very fast CPU clock speeds. * Revert changes to IAR-CM0-portmacro.h portNVIC_INT_CTRL_REG was already defined in port.c. No need to define it in portmacro.h. * Handle edge cases with slow SysTick clock Co-authored-by: Cobus van Eeden <35851496+cobusve@users.noreply.github.com> Co-authored-by: abhidixi11 <44424462+abhidixi11@users.noreply.github.com> Co-authored-by: Joseph Julicher <jjulicher@mac.com> Co-authored-by: alfred gedeon <28123637+alfred2g@users.noreply.github.com>
2 years ago
* out of sleep mode to execute immediately. See comments above
* the cpsid instruction above. */
__asm volatile ( "cpsie i" ::: "memory" );
__asm volatile ( "dsb" );
__asm volatile ( "isb" );
/* Disable interrupts again because the clock is about to be stopped
Tickless idle fixes/improvement (#59) * Fix tickless idle when stopping systick on zero... ...and don't stop SysTick at all in the eAbortSleep case. Prior to this commit, if vPortSuppressTicksAndSleep() happens to stop the SysTick on zero, then after tickless idle ends, xTickCount advances one full tick more than the time that actually elapsed as measured by the SysTick. See "bug 1" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - Idle task calls vPortSuppressTicksAndSleep(xExpectedIdleTime = 2). [Doesn't have to be "2" -- could be any number.] - vPortSuppressTicksAndSleep() stops SysTick, and the current-count register happens to stop on zero. - SysTick ISR executes, setting xPendedTicks = 1 - vPortSuppressTicksAndSleep() masks interrupts and calls eTaskConfirmSleepModeStatus() which confirms the sleep operation. *** - vPortSuppressTicksAndSleep() configures SysTick for 1 full tick (xExpectedIdleTime - 1) plus the current-count register (which is 0) - One tick period elapses in sleep. - SysTick wakes CPU, ISR executes and increments xPendedTicks to 2. - vPortSuppressTicksAndSleep() calls vTaskStepTick(1), then returns. - Idle task resumes scheduler, which increments xTickCount twice (for xPendedTicks = 2) In the end, two ticks elapsed as measured by SysTick, but the code increments xTickCount three times. The root cause is that the code assumes the SysTick current-count register always contains the number of SysTick counts remaining in the current tick period. However, when the current-count register is zero, there are ulTimerCountsForOneTick counts remaining, not zero. This error is not the kind of time slippage normally associated with tickless idle. *** Note that a recent commit https://github.com/FreeRTOS/FreeRTOS-Kernel/commit/e1b98f0 results in eAbortSleep in this case, due to xPendedTicks != 0. That commit does mostly resolve this bug without specifically mentioning it, and without this commit. But that resolution allows the code in port.c not to directly address the special case of stopping SysTick on zero in any code or comments. That commit also generates additional instances of eAbortSleep, and a second purpose of this commit is to optimize how vPortSuppressTicksAndSleep() behaves for eAbortSleep, as noted below. This commit also includes an optimization to avoid stopping the SysTick when eTaskConfirmSleepModeStatus() returns eAbortSleep. This optimization belongs with this fix because the method of handling the SysTick being stopped on zero changes with this optimization. * Fix imminent tick rescheduled after tickless idle Prior to this commit, if something other than systick wakes the CPU from tickless idle, vPortSuppressTicksAndSleep() might cause xTickCount to increment once too many times. See "bug 2" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - CPU is sleeping in vPortSuppressTicksAndSleep() - Something other than the SysTick wakes the CPU. - vPortSuppressTicksAndSleep() calculates the number of SysTick counts until the next tick. The bug occurs only if this number is small. - vPortSuppressTicksAndSleep() puts this small number into the SysTick reload register, and starts SysTick. - vPortSuppressTicksAndSleep() calls vTaskStepTick() - While vTaskStepTick() executes, the SysTick expires. The ISR pends because interrupts are masked, and SysTick starts a 2nd period still based on the small number of counts in its reload register. This 2nd period is undesirable and is likely to cause the error noted below. - vPortSuppressTicksAndSleep() puts the normal tick duration into the SysTick's reload register. - vPortSuppressTicksAndSleep() unmasks interrupts before the SysTick starts a new period based on the new value in the reload register. [This is a race condition that can go either way, but for the bug to occur, the race must play out this way.] - The pending SysTick ISR executes and increments xPendedTicks. - The SysTick expires again, finishing the second very small period, and starts a new period this time based on the full tick duration. - The SysTick ISR increments xPendedTicks (or xTickCount) even though only a tiny fraction of a tick period has elapsed since the previous tick. The bug occurs when *two* consecutive small periods of the SysTick are both counted as ticks. The root cause is a race caused by the small SysTick period. If vPortSuppressTicksAndSleep() unmasks interrupts *after* the small period expires but *before* the SysTick starts a period based on the full tick period, then two small periods are counted as ticks when only one should be counted. The end result is xTickCount advancing nearly one full tick more than time actually elapsed as measured by the SysTick. This is not the kind of time slippage normally associated with tickless idle. After this commit the code starts the SysTick and then immediately modifies the reload register to ensure the very short cycle (if any) is conducted only once. This strategy requires special consideration for the build option that configures SysTick to use a divided clock. To avoid waiting around for the SysTick to load value from the reload register, the new code temporarily configures the SysTick to use the undivided clock. The resulting timing error is typical for tickless idle. The error (commonly known as drift or slippage in kernel time) caused by this strategy is equivalent to one or two counts in ulStoppedTimerCompensation. This commit also updates comments and #define symbols related to the SysTick clock option. The SysTick can optionally be clocked by a divided version of the CPU clock (commonly divide-by-8). The new code in this commit adjusts these comments and symbols to make them clearer and more useful in configurations that use the divided clock. The fix made in this commit requires the use of these symbols, as noted in the code comments. * Fix tickless idle with alternate systick clocking Prior to this commit, in configurations using the alternate SysTick clocking, vPortSuppressTicksAndSleep() might cause xTickCount to jump ahead as much as the entire expected idle time or fall behind as much as one full tick compared to time as measured by the SysTick. SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. The SysTick has a configuration option to be clocked by an alternate clock besides the core clock. This alternate clock is MCU dependent. Scenarios Fixed --------------- The new code in this commit handles the following scenarios that were not handled correctly prior to this commit. 1. Before the sleep, vPortSuppressTicksAndSleep() stops the SysTick on zero, long after SysTick reached zero. Prior to this commit, this scenario caused xTickCount to jump ahead one full tick for the same reason documented here: https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/0c7b04bd3a745c52151abebc882eed3f811c4c81 2. After the sleep, vPortSuppressTicksAndSleep() stops the SysTick before it loads the counter from the reload register. Prior to this commit, this scenario caused xTickCount to jump ahead by the entire expected idle time (xExpectedIdleTime) because the current-count register is zero before it loads from the reload register. 3. Prior to return, vPortSuppressTicksAndSleep() attempts to start a short SysTick period when the current SysTick clock cycle has a lot of time remaining. Prior to this commit, this scenario could cause xTickCount to fall behind by as much as nearly one full tick because the short SysTick cycle never started. Note that #3 is partially fixed by https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/967acc9b200d3d4beeb289d9da9e88798074b431 even though that commit addresses a different issue. So this commit completes the partial fix. * Improve comments and name of preprocessor symbol Add a note in the code comments that SysTick requests an interrupt when decrementing from 1 to 0, so that's why stopping SysTick on zero is a special case. Readers might unknowingly assume that SysTick requests an interrupt when wrapping from 0 back to the load-register value. Reconsider new "_SETTING" suffix since "_CONFIG" suffix seems more descriptive. The code relies on *both* of these preprocessor symbols: portNVIC_SYSTICK_CLK_BIT portNVIC_SYSTICK_CLK_BIT_CONFIG **new** A meaningful suffix is really helpful to distinguish the two symbols. * Revert introduction of 2nd name for NVIC register When I added portNVIC_ICSR_REG I didn't realize there was already a portNVIC_INT_CTRL_REG, which identifies the same register. Not good to have both. Note that portNVIC_INT_CTRL_REG is defined in portmacro.h and is already used in this file (port.c). * Replicate to other Cortex M ports Also set a new fiddle factor based on tests with a CM4F. I used gcc, optimizing at -O1. Users can fine-tune as needed. Also add configSYSTICK_CLOCK_HZ to the CM0 ports to be just like the other Cortex M ports. This change allowed uniformity in the default tickless implementations across all Cortex M ports. And CM0 is likely to benefit from configSYSTICK_CLOCK_HZ, especially considering new CM0 devices with very fast CPU clock speeds. * Revert changes to IAR-CM0-portmacro.h portNVIC_INT_CTRL_REG was already defined in port.c. No need to define it in portmacro.h. * Handle edge cases with slow SysTick clock Co-authored-by: Cobus van Eeden <35851496+cobusve@users.noreply.github.com> Co-authored-by: abhidixi11 <44424462+abhidixi11@users.noreply.github.com> Co-authored-by: Joseph Julicher <jjulicher@mac.com> Co-authored-by: alfred gedeon <28123637+alfred2g@users.noreply.github.com>
2 years ago
* and interrupts that execute while the clock is stopped will increase
* any slippage between the time maintained by the RTOS and calendar
* time. */
__asm volatile ( "cpsid i" ::: "memory" );
__asm volatile ( "dsb" );
__asm volatile ( "isb" );
/* Disable the SysTick clock without reading the
* portNVIC_SYSTICK_CTRL_REG register to ensure the
Tickless idle fixes/improvement (#59) * Fix tickless idle when stopping systick on zero... ...and don't stop SysTick at all in the eAbortSleep case. Prior to this commit, if vPortSuppressTicksAndSleep() happens to stop the SysTick on zero, then after tickless idle ends, xTickCount advances one full tick more than the time that actually elapsed as measured by the SysTick. See "bug 1" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - Idle task calls vPortSuppressTicksAndSleep(xExpectedIdleTime = 2). [Doesn't have to be "2" -- could be any number.] - vPortSuppressTicksAndSleep() stops SysTick, and the current-count register happens to stop on zero. - SysTick ISR executes, setting xPendedTicks = 1 - vPortSuppressTicksAndSleep() masks interrupts and calls eTaskConfirmSleepModeStatus() which confirms the sleep operation. *** - vPortSuppressTicksAndSleep() configures SysTick for 1 full tick (xExpectedIdleTime - 1) plus the current-count register (which is 0) - One tick period elapses in sleep. - SysTick wakes CPU, ISR executes and increments xPendedTicks to 2. - vPortSuppressTicksAndSleep() calls vTaskStepTick(1), then returns. - Idle task resumes scheduler, which increments xTickCount twice (for xPendedTicks = 2) In the end, two ticks elapsed as measured by SysTick, but the code increments xTickCount three times. The root cause is that the code assumes the SysTick current-count register always contains the number of SysTick counts remaining in the current tick period. However, when the current-count register is zero, there are ulTimerCountsForOneTick counts remaining, not zero. This error is not the kind of time slippage normally associated with tickless idle. *** Note that a recent commit https://github.com/FreeRTOS/FreeRTOS-Kernel/commit/e1b98f0 results in eAbortSleep in this case, due to xPendedTicks != 0. That commit does mostly resolve this bug without specifically mentioning it, and without this commit. But that resolution allows the code in port.c not to directly address the special case of stopping SysTick on zero in any code or comments. That commit also generates additional instances of eAbortSleep, and a second purpose of this commit is to optimize how vPortSuppressTicksAndSleep() behaves for eAbortSleep, as noted below. This commit also includes an optimization to avoid stopping the SysTick when eTaskConfirmSleepModeStatus() returns eAbortSleep. This optimization belongs with this fix because the method of handling the SysTick being stopped on zero changes with this optimization. * Fix imminent tick rescheduled after tickless idle Prior to this commit, if something other than systick wakes the CPU from tickless idle, vPortSuppressTicksAndSleep() might cause xTickCount to increment once too many times. See "bug 2" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - CPU is sleeping in vPortSuppressTicksAndSleep() - Something other than the SysTick wakes the CPU. - vPortSuppressTicksAndSleep() calculates the number of SysTick counts until the next tick. The bug occurs only if this number is small. - vPortSuppressTicksAndSleep() puts this small number into the SysTick reload register, and starts SysTick. - vPortSuppressTicksAndSleep() calls vTaskStepTick() - While vTaskStepTick() executes, the SysTick expires. The ISR pends because interrupts are masked, and SysTick starts a 2nd period still based on the small number of counts in its reload register. This 2nd period is undesirable and is likely to cause the error noted below. - vPortSuppressTicksAndSleep() puts the normal tick duration into the SysTick's reload register. - vPortSuppressTicksAndSleep() unmasks interrupts before the SysTick starts a new period based on the new value in the reload register. [This is a race condition that can go either way, but for the bug to occur, the race must play out this way.] - The pending SysTick ISR executes and increments xPendedTicks. - The SysTick expires again, finishing the second very small period, and starts a new period this time based on the full tick duration. - The SysTick ISR increments xPendedTicks (or xTickCount) even though only a tiny fraction of a tick period has elapsed since the previous tick. The bug occurs when *two* consecutive small periods of the SysTick are both counted as ticks. The root cause is a race caused by the small SysTick period. If vPortSuppressTicksAndSleep() unmasks interrupts *after* the small period expires but *before* the SysTick starts a period based on the full tick period, then two small periods are counted as ticks when only one should be counted. The end result is xTickCount advancing nearly one full tick more than time actually elapsed as measured by the SysTick. This is not the kind of time slippage normally associated with tickless idle. After this commit the code starts the SysTick and then immediately modifies the reload register to ensure the very short cycle (if any) is conducted only once. This strategy requires special consideration for the build option that configures SysTick to use a divided clock. To avoid waiting around for the SysTick to load value from the reload register, the new code temporarily configures the SysTick to use the undivided clock. The resulting timing error is typical for tickless idle. The error (commonly known as drift or slippage in kernel time) caused by this strategy is equivalent to one or two counts in ulStoppedTimerCompensation. This commit also updates comments and #define symbols related to the SysTick clock option. The SysTick can optionally be clocked by a divided version of the CPU clock (commonly divide-by-8). The new code in this commit adjusts these comments and symbols to make them clearer and more useful in configurations that use the divided clock. The fix made in this commit requires the use of these symbols, as noted in the code comments. * Fix tickless idle with alternate systick clocking Prior to this commit, in configurations using the alternate SysTick clocking, vPortSuppressTicksAndSleep() might cause xTickCount to jump ahead as much as the entire expected idle time or fall behind as much as one full tick compared to time as measured by the SysTick. SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. The SysTick has a configuration option to be clocked by an alternate clock besides the core clock. This alternate clock is MCU dependent. Scenarios Fixed --------------- The new code in this commit handles the following scenarios that were not handled correctly prior to this commit. 1. Before the sleep, vPortSuppressTicksAndSleep() stops the SysTick on zero, long after SysTick reached zero. Prior to this commit, this scenario caused xTickCount to jump ahead one full tick for the same reason documented here: https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/0c7b04bd3a745c52151abebc882eed3f811c4c81 2. After the sleep, vPortSuppressTicksAndSleep() stops the SysTick before it loads the counter from the reload register. Prior to this commit, this scenario caused xTickCount to jump ahead by the entire expected idle time (xExpectedIdleTime) because the current-count register is zero before it loads from the reload register. 3. Prior to return, vPortSuppressTicksAndSleep() attempts to start a short SysTick period when the current SysTick clock cycle has a lot of time remaining. Prior to this commit, this scenario could cause xTickCount to fall behind by as much as nearly one full tick because the short SysTick cycle never started. Note that #3 is partially fixed by https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/967acc9b200d3d4beeb289d9da9e88798074b431 even though that commit addresses a different issue. So this commit completes the partial fix. * Improve comments and name of preprocessor symbol Add a note in the code comments that SysTick requests an interrupt when decrementing from 1 to 0, so that's why stopping SysTick on zero is a special case. Readers might unknowingly assume that SysTick requests an interrupt when wrapping from 0 back to the load-register value. Reconsider new "_SETTING" suffix since "_CONFIG" suffix seems more descriptive. The code relies on *both* of these preprocessor symbols: portNVIC_SYSTICK_CLK_BIT portNVIC_SYSTICK_CLK_BIT_CONFIG **new** A meaningful suffix is really helpful to distinguish the two symbols. * Revert introduction of 2nd name for NVIC register When I added portNVIC_ICSR_REG I didn't realize there was already a portNVIC_INT_CTRL_REG, which identifies the same register. Not good to have both. Note that portNVIC_INT_CTRL_REG is defined in portmacro.h and is already used in this file (port.c). * Replicate to other Cortex M ports Also set a new fiddle factor based on tests with a CM4F. I used gcc, optimizing at -O1. Users can fine-tune as needed. Also add configSYSTICK_CLOCK_HZ to the CM0 ports to be just like the other Cortex M ports. This change allowed uniformity in the default tickless implementations across all Cortex M ports. And CM0 is likely to benefit from configSYSTICK_CLOCK_HZ, especially considering new CM0 devices with very fast CPU clock speeds. * Revert changes to IAR-CM0-portmacro.h portNVIC_INT_CTRL_REG was already defined in port.c. No need to define it in portmacro.h. * Handle edge cases with slow SysTick clock Co-authored-by: Cobus van Eeden <35851496+cobusve@users.noreply.github.com> Co-authored-by: abhidixi11 <44424462+abhidixi11@users.noreply.github.com> Co-authored-by: Joseph Julicher <jjulicher@mac.com> Co-authored-by: alfred gedeon <28123637+alfred2g@users.noreply.github.com>
2 years ago
* portNVIC_SYSTICK_COUNT_FLAG_BIT is not cleared if it is set. Again,
* the time the SysTick is stopped for is accounted for as best it can
* be, but using the tickless mode will inevitably result in some tiny
* drift of the time maintained by the kernel with respect to calendar
* time*/
portNVIC_SYSTICK_CTRL_REG = ( portNVIC_SYSTICK_CLK_BIT_CONFIG | portNVIC_SYSTICK_INT_BIT );
/* Determine whether the SysTick has already counted to zero. */
if( ( portNVIC_SYSTICK_CTRL_REG & portNVIC_SYSTICK_COUNT_FLAG_BIT ) != 0 )
{
uint32_t ulCalculatedLoadValue;
Tickless idle fixes/improvement (#59) * Fix tickless idle when stopping systick on zero... ...and don't stop SysTick at all in the eAbortSleep case. Prior to this commit, if vPortSuppressTicksAndSleep() happens to stop the SysTick on zero, then after tickless idle ends, xTickCount advances one full tick more than the time that actually elapsed as measured by the SysTick. See "bug 1" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - Idle task calls vPortSuppressTicksAndSleep(xExpectedIdleTime = 2). [Doesn't have to be "2" -- could be any number.] - vPortSuppressTicksAndSleep() stops SysTick, and the current-count register happens to stop on zero. - SysTick ISR executes, setting xPendedTicks = 1 - vPortSuppressTicksAndSleep() masks interrupts and calls eTaskConfirmSleepModeStatus() which confirms the sleep operation. *** - vPortSuppressTicksAndSleep() configures SysTick for 1 full tick (xExpectedIdleTime - 1) plus the current-count register (which is 0) - One tick period elapses in sleep. - SysTick wakes CPU, ISR executes and increments xPendedTicks to 2. - vPortSuppressTicksAndSleep() calls vTaskStepTick(1), then returns. - Idle task resumes scheduler, which increments xTickCount twice (for xPendedTicks = 2) In the end, two ticks elapsed as measured by SysTick, but the code increments xTickCount three times. The root cause is that the code assumes the SysTick current-count register always contains the number of SysTick counts remaining in the current tick period. However, when the current-count register is zero, there are ulTimerCountsForOneTick counts remaining, not zero. This error is not the kind of time slippage normally associated with tickless idle. *** Note that a recent commit https://github.com/FreeRTOS/FreeRTOS-Kernel/commit/e1b98f0 results in eAbortSleep in this case, due to xPendedTicks != 0. That commit does mostly resolve this bug without specifically mentioning it, and without this commit. But that resolution allows the code in port.c not to directly address the special case of stopping SysTick on zero in any code or comments. That commit also generates additional instances of eAbortSleep, and a second purpose of this commit is to optimize how vPortSuppressTicksAndSleep() behaves for eAbortSleep, as noted below. This commit also includes an optimization to avoid stopping the SysTick when eTaskConfirmSleepModeStatus() returns eAbortSleep. This optimization belongs with this fix because the method of handling the SysTick being stopped on zero changes with this optimization. * Fix imminent tick rescheduled after tickless idle Prior to this commit, if something other than systick wakes the CPU from tickless idle, vPortSuppressTicksAndSleep() might cause xTickCount to increment once too many times. See "bug 2" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - CPU is sleeping in vPortSuppressTicksAndSleep() - Something other than the SysTick wakes the CPU. - vPortSuppressTicksAndSleep() calculates the number of SysTick counts until the next tick. The bug occurs only if this number is small. - vPortSuppressTicksAndSleep() puts this small number into the SysTick reload register, and starts SysTick. - vPortSuppressTicksAndSleep() calls vTaskStepTick() - While vTaskStepTick() executes, the SysTick expires. The ISR pends because interrupts are masked, and SysTick starts a 2nd period still based on the small number of counts in its reload register. This 2nd period is undesirable and is likely to cause the error noted below. - vPortSuppressTicksAndSleep() puts the normal tick duration into the SysTick's reload register. - vPortSuppressTicksAndSleep() unmasks interrupts before the SysTick starts a new period based on the new value in the reload register. [This is a race condition that can go either way, but for the bug to occur, the race must play out this way.] - The pending SysTick ISR executes and increments xPendedTicks. - The SysTick expires again, finishing the second very small period, and starts a new period this time based on the full tick duration. - The SysTick ISR increments xPendedTicks (or xTickCount) even though only a tiny fraction of a tick period has elapsed since the previous tick. The bug occurs when *two* consecutive small periods of the SysTick are both counted as ticks. The root cause is a race caused by the small SysTick period. If vPortSuppressTicksAndSleep() unmasks interrupts *after* the small period expires but *before* the SysTick starts a period based on the full tick period, then two small periods are counted as ticks when only one should be counted. The end result is xTickCount advancing nearly one full tick more than time actually elapsed as measured by the SysTick. This is not the kind of time slippage normally associated with tickless idle. After this commit the code starts the SysTick and then immediately modifies the reload register to ensure the very short cycle (if any) is conducted only once. This strategy requires special consideration for the build option that configures SysTick to use a divided clock. To avoid waiting around for the SysTick to load value from the reload register, the new code temporarily configures the SysTick to use the undivided clock. The resulting timing error is typical for tickless idle. The error (commonly known as drift or slippage in kernel time) caused by this strategy is equivalent to one or two counts in ulStoppedTimerCompensation. This commit also updates comments and #define symbols related to the SysTick clock option. The SysTick can optionally be clocked by a divided version of the CPU clock (commonly divide-by-8). The new code in this commit adjusts these comments and symbols to make them clearer and more useful in configurations that use the divided clock. The fix made in this commit requires the use of these symbols, as noted in the code comments. * Fix tickless idle with alternate systick clocking Prior to this commit, in configurations using the alternate SysTick clocking, vPortSuppressTicksAndSleep() might cause xTickCount to jump ahead as much as the entire expected idle time or fall behind as much as one full tick compared to time as measured by the SysTick. SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. The SysTick has a configuration option to be clocked by an alternate clock besides the core clock. This alternate clock is MCU dependent. Scenarios Fixed --------------- The new code in this commit handles the following scenarios that were not handled correctly prior to this commit. 1. Before the sleep, vPortSuppressTicksAndSleep() stops the SysTick on zero, long after SysTick reached zero. Prior to this commit, this scenario caused xTickCount to jump ahead one full tick for the same reason documented here: https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/0c7b04bd3a745c52151abebc882eed3f811c4c81 2. After the sleep, vPortSuppressTicksAndSleep() stops the SysTick before it loads the counter from the reload register. Prior to this commit, this scenario caused xTickCount to jump ahead by the entire expected idle time (xExpectedIdleTime) because the current-count register is zero before it loads from the reload register. 3. Prior to return, vPortSuppressTicksAndSleep() attempts to start a short SysTick period when the current SysTick clock cycle has a lot of time remaining. Prior to this commit, this scenario could cause xTickCount to fall behind by as much as nearly one full tick because the short SysTick cycle never started. Note that #3 is partially fixed by https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/967acc9b200d3d4beeb289d9da9e88798074b431 even though that commit addresses a different issue. So this commit completes the partial fix. * Improve comments and name of preprocessor symbol Add a note in the code comments that SysTick requests an interrupt when decrementing from 1 to 0, so that's why stopping SysTick on zero is a special case. Readers might unknowingly assume that SysTick requests an interrupt when wrapping from 0 back to the load-register value. Reconsider new "_SETTING" suffix since "_CONFIG" suffix seems more descriptive. The code relies on *both* of these preprocessor symbols: portNVIC_SYSTICK_CLK_BIT portNVIC_SYSTICK_CLK_BIT_CONFIG **new** A meaningful suffix is really helpful to distinguish the two symbols. * Revert introduction of 2nd name for NVIC register When I added portNVIC_ICSR_REG I didn't realize there was already a portNVIC_INT_CTRL_REG, which identifies the same register. Not good to have both. Note that portNVIC_INT_CTRL_REG is defined in portmacro.h and is already used in this file (port.c). * Replicate to other Cortex M ports Also set a new fiddle factor based on tests with a CM4F. I used gcc, optimizing at -O1. Users can fine-tune as needed. Also add configSYSTICK_CLOCK_HZ to the CM0 ports to be just like the other Cortex M ports. This change allowed uniformity in the default tickless implementations across all Cortex M ports. And CM0 is likely to benefit from configSYSTICK_CLOCK_HZ, especially considering new CM0 devices with very fast CPU clock speeds. * Revert changes to IAR-CM0-portmacro.h portNVIC_INT_CTRL_REG was already defined in port.c. No need to define it in portmacro.h. * Handle edge cases with slow SysTick clock Co-authored-by: Cobus van Eeden <35851496+cobusve@users.noreply.github.com> Co-authored-by: abhidixi11 <44424462+abhidixi11@users.noreply.github.com> Co-authored-by: Joseph Julicher <jjulicher@mac.com> Co-authored-by: alfred gedeon <28123637+alfred2g@users.noreply.github.com>
2 years ago
/* The tick interrupt ended the sleep (or is now pending), and
* a new tick period has started. Reset portNVIC_SYSTICK_LOAD_REG
* with whatever remains of the new tick period. */
ulCalculatedLoadValue = ( ulTimerCountsForOneTick - 1UL ) - ( ulReloadValue - portNVIC_SYSTICK_CURRENT_VALUE_REG );
/* Don't allow a tiny value, or values that have somehow
* underflowed because the post sleep hook did something
Tickless idle fixes/improvement (#59) * Fix tickless idle when stopping systick on zero... ...and don't stop SysTick at all in the eAbortSleep case. Prior to this commit, if vPortSuppressTicksAndSleep() happens to stop the SysTick on zero, then after tickless idle ends, xTickCount advances one full tick more than the time that actually elapsed as measured by the SysTick. See "bug 1" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - Idle task calls vPortSuppressTicksAndSleep(xExpectedIdleTime = 2). [Doesn't have to be "2" -- could be any number.] - vPortSuppressTicksAndSleep() stops SysTick, and the current-count register happens to stop on zero. - SysTick ISR executes, setting xPendedTicks = 1 - vPortSuppressTicksAndSleep() masks interrupts and calls eTaskConfirmSleepModeStatus() which confirms the sleep operation. *** - vPortSuppressTicksAndSleep() configures SysTick for 1 full tick (xExpectedIdleTime - 1) plus the current-count register (which is 0) - One tick period elapses in sleep. - SysTick wakes CPU, ISR executes and increments xPendedTicks to 2. - vPortSuppressTicksAndSleep() calls vTaskStepTick(1), then returns. - Idle task resumes scheduler, which increments xTickCount twice (for xPendedTicks = 2) In the end, two ticks elapsed as measured by SysTick, but the code increments xTickCount three times. The root cause is that the code assumes the SysTick current-count register always contains the number of SysTick counts remaining in the current tick period. However, when the current-count register is zero, there are ulTimerCountsForOneTick counts remaining, not zero. This error is not the kind of time slippage normally associated with tickless idle. *** Note that a recent commit https://github.com/FreeRTOS/FreeRTOS-Kernel/commit/e1b98f0 results in eAbortSleep in this case, due to xPendedTicks != 0. That commit does mostly resolve this bug without specifically mentioning it, and without this commit. But that resolution allows the code in port.c not to directly address the special case of stopping SysTick on zero in any code or comments. That commit also generates additional instances of eAbortSleep, and a second purpose of this commit is to optimize how vPortSuppressTicksAndSleep() behaves for eAbortSleep, as noted below. This commit also includes an optimization to avoid stopping the SysTick when eTaskConfirmSleepModeStatus() returns eAbortSleep. This optimization belongs with this fix because the method of handling the SysTick being stopped on zero changes with this optimization. * Fix imminent tick rescheduled after tickless idle Prior to this commit, if something other than systick wakes the CPU from tickless idle, vPortSuppressTicksAndSleep() might cause xTickCount to increment once too many times. See "bug 2" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - CPU is sleeping in vPortSuppressTicksAndSleep() - Something other than the SysTick wakes the CPU. - vPortSuppressTicksAndSleep() calculates the number of SysTick counts until the next tick. The bug occurs only if this number is small. - vPortSuppressTicksAndSleep() puts this small number into the SysTick reload register, and starts SysTick. - vPortSuppressTicksAndSleep() calls vTaskStepTick() - While vTaskStepTick() executes, the SysTick expires. The ISR pends because interrupts are masked, and SysTick starts a 2nd period still based on the small number of counts in its reload register. This 2nd period is undesirable and is likely to cause the error noted below. - vPortSuppressTicksAndSleep() puts the normal tick duration into the SysTick's reload register. - vPortSuppressTicksAndSleep() unmasks interrupts before the SysTick starts a new period based on the new value in the reload register. [This is a race condition that can go either way, but for the bug to occur, the race must play out this way.] - The pending SysTick ISR executes and increments xPendedTicks. - The SysTick expires again, finishing the second very small period, and starts a new period this time based on the full tick duration. - The SysTick ISR increments xPendedTicks (or xTickCount) even though only a tiny fraction of a tick period has elapsed since the previous tick. The bug occurs when *two* consecutive small periods of the SysTick are both counted as ticks. The root cause is a race caused by the small SysTick period. If vPortSuppressTicksAndSleep() unmasks interrupts *after* the small period expires but *before* the SysTick starts a period based on the full tick period, then two small periods are counted as ticks when only one should be counted. The end result is xTickCount advancing nearly one full tick more than time actually elapsed as measured by the SysTick. This is not the kind of time slippage normally associated with tickless idle. After this commit the code starts the SysTick and then immediately modifies the reload register to ensure the very short cycle (if any) is conducted only once. This strategy requires special consideration for the build option that configures SysTick to use a divided clock. To avoid waiting around for the SysTick to load value from the reload register, the new code temporarily configures the SysTick to use the undivided clock. The resulting timing error is typical for tickless idle. The error (commonly known as drift or slippage in kernel time) caused by this strategy is equivalent to one or two counts in ulStoppedTimerCompensation. This commit also updates comments and #define symbols related to the SysTick clock option. The SysTick can optionally be clocked by a divided version of the CPU clock (commonly divide-by-8). The new code in this commit adjusts these comments and symbols to make them clearer and more useful in configurations that use the divided clock. The fix made in this commit requires the use of these symbols, as noted in the code comments. * Fix tickless idle with alternate systick clocking Prior to this commit, in configurations using the alternate SysTick clocking, vPortSuppressTicksAndSleep() might cause xTickCount to jump ahead as much as the entire expected idle time or fall behind as much as one full tick compared to time as measured by the SysTick. SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. The SysTick has a configuration option to be clocked by an alternate clock besides the core clock. This alternate clock is MCU dependent. Scenarios Fixed --------------- The new code in this commit handles the following scenarios that were not handled correctly prior to this commit. 1. Before the sleep, vPortSuppressTicksAndSleep() stops the SysTick on zero, long after SysTick reached zero. Prior to this commit, this scenario caused xTickCount to jump ahead one full tick for the same reason documented here: https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/0c7b04bd3a745c52151abebc882eed3f811c4c81 2. After the sleep, vPortSuppressTicksAndSleep() stops the SysTick before it loads the counter from the reload register. Prior to this commit, this scenario caused xTickCount to jump ahead by the entire expected idle time (xExpectedIdleTime) because the current-count register is zero before it loads from the reload register. 3. Prior to return, vPortSuppressTicksAndSleep() attempts to start a short SysTick period when the current SysTick clock cycle has a lot of time remaining. Prior to this commit, this scenario could cause xTickCount to fall behind by as much as nearly one full tick because the short SysTick cycle never started. Note that #3 is partially fixed by https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/967acc9b200d3d4beeb289d9da9e88798074b431 even though that commit addresses a different issue. So this commit completes the partial fix. * Improve comments and name of preprocessor symbol Add a note in the code comments that SysTick requests an interrupt when decrementing from 1 to 0, so that's why stopping SysTick on zero is a special case. Readers might unknowingly assume that SysTick requests an interrupt when wrapping from 0 back to the load-register value. Reconsider new "_SETTING" suffix since "_CONFIG" suffix seems more descriptive. The code relies on *both* of these preprocessor symbols: portNVIC_SYSTICK_CLK_BIT portNVIC_SYSTICK_CLK_BIT_CONFIG **new** A meaningful suffix is really helpful to distinguish the two symbols. * Revert introduction of 2nd name for NVIC register When I added portNVIC_ICSR_REG I didn't realize there was already a portNVIC_INT_CTRL_REG, which identifies the same register. Not good to have both. Note that portNVIC_INT_CTRL_REG is defined in portmacro.h and is already used in this file (port.c). * Replicate to other Cortex M ports Also set a new fiddle factor based on tests with a CM4F. I used gcc, optimizing at -O1. Users can fine-tune as needed. Also add configSYSTICK_CLOCK_HZ to the CM0 ports to be just like the other Cortex M ports. This change allowed uniformity in the default tickless implementations across all Cortex M ports. And CM0 is likely to benefit from configSYSTICK_CLOCK_HZ, especially considering new CM0 devices with very fast CPU clock speeds. * Revert changes to IAR-CM0-portmacro.h portNVIC_INT_CTRL_REG was already defined in port.c. No need to define it in portmacro.h. * Handle edge cases with slow SysTick clock Co-authored-by: Cobus van Eeden <35851496+cobusve@users.noreply.github.com> Co-authored-by: abhidixi11 <44424462+abhidixi11@users.noreply.github.com> Co-authored-by: Joseph Julicher <jjulicher@mac.com> Co-authored-by: alfred gedeon <28123637+alfred2g@users.noreply.github.com>
2 years ago
* that took too long or because the SysTick current-value register
* is zero. */
if( ( ulCalculatedLoadValue <= ulStoppedTimerCompensation ) || ( ulCalculatedLoadValue > ulTimerCountsForOneTick ) )
{
ulCalculatedLoadValue = ( ulTimerCountsForOneTick - 1UL );
}
portNVIC_SYSTICK_LOAD_REG = ulCalculatedLoadValue;
/* As the pending tick will be processed as soon as this
Tickless idle fixes/improvement (#59) * Fix tickless idle when stopping systick on zero... ...and don't stop SysTick at all in the eAbortSleep case. Prior to this commit, if vPortSuppressTicksAndSleep() happens to stop the SysTick on zero, then after tickless idle ends, xTickCount advances one full tick more than the time that actually elapsed as measured by the SysTick. See "bug 1" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - Idle task calls vPortSuppressTicksAndSleep(xExpectedIdleTime = 2). [Doesn't have to be "2" -- could be any number.] - vPortSuppressTicksAndSleep() stops SysTick, and the current-count register happens to stop on zero. - SysTick ISR executes, setting xPendedTicks = 1 - vPortSuppressTicksAndSleep() masks interrupts and calls eTaskConfirmSleepModeStatus() which confirms the sleep operation. *** - vPortSuppressTicksAndSleep() configures SysTick for 1 full tick (xExpectedIdleTime - 1) plus the current-count register (which is 0) - One tick period elapses in sleep. - SysTick wakes CPU, ISR executes and increments xPendedTicks to 2. - vPortSuppressTicksAndSleep() calls vTaskStepTick(1), then returns. - Idle task resumes scheduler, which increments xTickCount twice (for xPendedTicks = 2) In the end, two ticks elapsed as measured by SysTick, but the code increments xTickCount three times. The root cause is that the code assumes the SysTick current-count register always contains the number of SysTick counts remaining in the current tick period. However, when the current-count register is zero, there are ulTimerCountsForOneTick counts remaining, not zero. This error is not the kind of time slippage normally associated with tickless idle. *** Note that a recent commit https://github.com/FreeRTOS/FreeRTOS-Kernel/commit/e1b98f0 results in eAbortSleep in this case, due to xPendedTicks != 0. That commit does mostly resolve this bug without specifically mentioning it, and without this commit. But that resolution allows the code in port.c not to directly address the special case of stopping SysTick on zero in any code or comments. That commit also generates additional instances of eAbortSleep, and a second purpose of this commit is to optimize how vPortSuppressTicksAndSleep() behaves for eAbortSleep, as noted below. This commit also includes an optimization to avoid stopping the SysTick when eTaskConfirmSleepModeStatus() returns eAbortSleep. This optimization belongs with this fix because the method of handling the SysTick being stopped on zero changes with this optimization. * Fix imminent tick rescheduled after tickless idle Prior to this commit, if something other than systick wakes the CPU from tickless idle, vPortSuppressTicksAndSleep() might cause xTickCount to increment once too many times. See "bug 2" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - CPU is sleeping in vPortSuppressTicksAndSleep() - Something other than the SysTick wakes the CPU. - vPortSuppressTicksAndSleep() calculates the number of SysTick counts until the next tick. The bug occurs only if this number is small. - vPortSuppressTicksAndSleep() puts this small number into the SysTick reload register, and starts SysTick. - vPortSuppressTicksAndSleep() calls vTaskStepTick() - While vTaskStepTick() executes, the SysTick expires. The ISR pends because interrupts are masked, and SysTick starts a 2nd period still based on the small number of counts in its reload register. This 2nd period is undesirable and is likely to cause the error noted below. - vPortSuppressTicksAndSleep() puts the normal tick duration into the SysTick's reload register. - vPortSuppressTicksAndSleep() unmasks interrupts before the SysTick starts a new period based on the new value in the reload register. [This is a race condition that can go either way, but for the bug to occur, the race must play out this way.] - The pending SysTick ISR executes and increments xPendedTicks. - The SysTick expires again, finishing the second very small period, and starts a new period this time based on the full tick duration. - The SysTick ISR increments xPendedTicks (or xTickCount) even though only a tiny fraction of a tick period has elapsed since the previous tick. The bug occurs when *two* consecutive small periods of the SysTick are both counted as ticks. The root cause is a race caused by the small SysTick period. If vPortSuppressTicksAndSleep() unmasks interrupts *after* the small period expires but *before* the SysTick starts a period based on the full tick period, then two small periods are counted as ticks when only one should be counted. The end result is xTickCount advancing nearly one full tick more than time actually elapsed as measured by the SysTick. This is not the kind of time slippage normally associated with tickless idle. After this commit the code starts the SysTick and then immediately modifies the reload register to ensure the very short cycle (if any) is conducted only once. This strategy requires special consideration for the build option that configures SysTick to use a divided clock. To avoid waiting around for the SysTick to load value from the reload register, the new code temporarily configures the SysTick to use the undivided clock. The resulting timing error is typical for tickless idle. The error (commonly known as drift or slippage in kernel time) caused by this strategy is equivalent to one or two counts in ulStoppedTimerCompensation. This commit also updates comments and #define symbols related to the SysTick clock option. The SysTick can optionally be clocked by a divided version of the CPU clock (commonly divide-by-8). The new code in this commit adjusts these comments and symbols to make them clearer and more useful in configurations that use the divided clock. The fix made in this commit requires the use of these symbols, as noted in the code comments. * Fix tickless idle with alternate systick clocking Prior to this commit, in configurations using the alternate SysTick clocking, vPortSuppressTicksAndSleep() might cause xTickCount to jump ahead as much as the entire expected idle time or fall behind as much as one full tick compared to time as measured by the SysTick. SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. The SysTick has a configuration option to be clocked by an alternate clock besides the core clock. This alternate clock is MCU dependent. Scenarios Fixed --------------- The new code in this commit handles the following scenarios that were not handled correctly prior to this commit. 1. Before the sleep, vPortSuppressTicksAndSleep() stops the SysTick on zero, long after SysTick reached zero. Prior to this commit, this scenario caused xTickCount to jump ahead one full tick for the same reason documented here: https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/0c7b04bd3a745c52151abebc882eed3f811c4c81 2. After the sleep, vPortSuppressTicksAndSleep() stops the SysTick before it loads the counter from the reload register. Prior to this commit, this scenario caused xTickCount to jump ahead by the entire expected idle time (xExpectedIdleTime) because the current-count register is zero before it loads from the reload register. 3. Prior to return, vPortSuppressTicksAndSleep() attempts to start a short SysTick period when the current SysTick clock cycle has a lot of time remaining. Prior to this commit, this scenario could cause xTickCount to fall behind by as much as nearly one full tick because the short SysTick cycle never started. Note that #3 is partially fixed by https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/967acc9b200d3d4beeb289d9da9e88798074b431 even though that commit addresses a different issue. So this commit completes the partial fix. * Improve comments and name of preprocessor symbol Add a note in the code comments that SysTick requests an interrupt when decrementing from 1 to 0, so that's why stopping SysTick on zero is a special case. Readers might unknowingly assume that SysTick requests an interrupt when wrapping from 0 back to the load-register value. Reconsider new "_SETTING" suffix since "_CONFIG" suffix seems more descriptive. The code relies on *both* of these preprocessor symbols: portNVIC_SYSTICK_CLK_BIT portNVIC_SYSTICK_CLK_BIT_CONFIG **new** A meaningful suffix is really helpful to distinguish the two symbols. * Revert introduction of 2nd name for NVIC register When I added portNVIC_ICSR_REG I didn't realize there was already a portNVIC_INT_CTRL_REG, which identifies the same register. Not good to have both. Note that portNVIC_INT_CTRL_REG is defined in portmacro.h and is already used in this file (port.c). * Replicate to other Cortex M ports Also set a new fiddle factor based on tests with a CM4F. I used gcc, optimizing at -O1. Users can fine-tune as needed. Also add configSYSTICK_CLOCK_HZ to the CM0 ports to be just like the other Cortex M ports. This change allowed uniformity in the default tickless implementations across all Cortex M ports. And CM0 is likely to benefit from configSYSTICK_CLOCK_HZ, especially considering new CM0 devices with very fast CPU clock speeds. * Revert changes to IAR-CM0-portmacro.h portNVIC_INT_CTRL_REG was already defined in port.c. No need to define it in portmacro.h. * Handle edge cases with slow SysTick clock Co-authored-by: Cobus van Eeden <35851496+cobusve@users.noreply.github.com> Co-authored-by: abhidixi11 <44424462+abhidixi11@users.noreply.github.com> Co-authored-by: Joseph Julicher <jjulicher@mac.com> Co-authored-by: alfred gedeon <28123637+alfred2g@users.noreply.github.com>
2 years ago
* function exits, the tick value maintained by the tick is stepped
* forward by one less than the time spent waiting. */
ulCompleteTickPeriods = xExpectedIdleTime - 1UL;
}
else
{
Tickless idle fixes/improvement (#59) * Fix tickless idle when stopping systick on zero... ...and don't stop SysTick at all in the eAbortSleep case. Prior to this commit, if vPortSuppressTicksAndSleep() happens to stop the SysTick on zero, then after tickless idle ends, xTickCount advances one full tick more than the time that actually elapsed as measured by the SysTick. See "bug 1" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - Idle task calls vPortSuppressTicksAndSleep(xExpectedIdleTime = 2). [Doesn't have to be "2" -- could be any number.] - vPortSuppressTicksAndSleep() stops SysTick, and the current-count register happens to stop on zero. - SysTick ISR executes, setting xPendedTicks = 1 - vPortSuppressTicksAndSleep() masks interrupts and calls eTaskConfirmSleepModeStatus() which confirms the sleep operation. *** - vPortSuppressTicksAndSleep() configures SysTick for 1 full tick (xExpectedIdleTime - 1) plus the current-count register (which is 0) - One tick period elapses in sleep. - SysTick wakes CPU, ISR executes and increments xPendedTicks to 2. - vPortSuppressTicksAndSleep() calls vTaskStepTick(1), then returns. - Idle task resumes scheduler, which increments xTickCount twice (for xPendedTicks = 2) In the end, two ticks elapsed as measured by SysTick, but the code increments xTickCount three times. The root cause is that the code assumes the SysTick current-count register always contains the number of SysTick counts remaining in the current tick period. However, when the current-count register is zero, there are ulTimerCountsForOneTick counts remaining, not zero. This error is not the kind of time slippage normally associated with tickless idle. *** Note that a recent commit https://github.com/FreeRTOS/FreeRTOS-Kernel/commit/e1b98f0 results in eAbortSleep in this case, due to xPendedTicks != 0. That commit does mostly resolve this bug without specifically mentioning it, and without this commit. But that resolution allows the code in port.c not to directly address the special case of stopping SysTick on zero in any code or comments. That commit also generates additional instances of eAbortSleep, and a second purpose of this commit is to optimize how vPortSuppressTicksAndSleep() behaves for eAbortSleep, as noted below. This commit also includes an optimization to avoid stopping the SysTick when eTaskConfirmSleepModeStatus() returns eAbortSleep. This optimization belongs with this fix because the method of handling the SysTick being stopped on zero changes with this optimization. * Fix imminent tick rescheduled after tickless idle Prior to this commit, if something other than systick wakes the CPU from tickless idle, vPortSuppressTicksAndSleep() might cause xTickCount to increment once too many times. See "bug 2" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - CPU is sleeping in vPortSuppressTicksAndSleep() - Something other than the SysTick wakes the CPU. - vPortSuppressTicksAndSleep() calculates the number of SysTick counts until the next tick. The bug occurs only if this number is small. - vPortSuppressTicksAndSleep() puts this small number into the SysTick reload register, and starts SysTick. - vPortSuppressTicksAndSleep() calls vTaskStepTick() - While vTaskStepTick() executes, the SysTick expires. The ISR pends because interrupts are masked, and SysTick starts a 2nd period still based on the small number of counts in its reload register. This 2nd period is undesirable and is likely to cause the error noted below. - vPortSuppressTicksAndSleep() puts the normal tick duration into the SysTick's reload register. - vPortSuppressTicksAndSleep() unmasks interrupts before the SysTick starts a new period based on the new value in the reload register. [This is a race condition that can go either way, but for the bug to occur, the race must play out this way.] - The pending SysTick ISR executes and increments xPendedTicks. - The SysTick expires again, finishing the second very small period, and starts a new period this time based on the full tick duration. - The SysTick ISR increments xPendedTicks (or xTickCount) even though only a tiny fraction of a tick period has elapsed since the previous tick. The bug occurs when *two* consecutive small periods of the SysTick are both counted as ticks. The root cause is a race caused by the small SysTick period. If vPortSuppressTicksAndSleep() unmasks interrupts *after* the small period expires but *before* the SysTick starts a period based on the full tick period, then two small periods are counted as ticks when only one should be counted. The end result is xTickCount advancing nearly one full tick more than time actually elapsed as measured by the SysTick. This is not the kind of time slippage normally associated with tickless idle. After this commit the code starts the SysTick and then immediately modifies the reload register to ensure the very short cycle (if any) is conducted only once. This strategy requires special consideration for the build option that configures SysTick to use a divided clock. To avoid waiting around for the SysTick to load value from the reload register, the new code temporarily configures the SysTick to use the undivided clock. The resulting timing error is typical for tickless idle. The error (commonly known as drift or slippage in kernel time) caused by this strategy is equivalent to one or two counts in ulStoppedTimerCompensation. This commit also updates comments and #define symbols related to the SysTick clock option. The SysTick can optionally be clocked by a divided version of the CPU clock (commonly divide-by-8). The new code in this commit adjusts these comments and symbols to make them clearer and more useful in configurations that use the divided clock. The fix made in this commit requires the use of these symbols, as noted in the code comments. * Fix tickless idle with alternate systick clocking Prior to this commit, in configurations using the alternate SysTick clocking, vPortSuppressTicksAndSleep() might cause xTickCount to jump ahead as much as the entire expected idle time or fall behind as much as one full tick compared to time as measured by the SysTick. SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. The SysTick has a configuration option to be clocked by an alternate clock besides the core clock. This alternate clock is MCU dependent. Scenarios Fixed --------------- The new code in this commit handles the following scenarios that were not handled correctly prior to this commit. 1. Before the sleep, vPortSuppressTicksAndSleep() stops the SysTick on zero, long after SysTick reached zero. Prior to this commit, this scenario caused xTickCount to jump ahead one full tick for the same reason documented here: https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/0c7b04bd3a745c52151abebc882eed3f811c4c81 2. After the sleep, vPortSuppressTicksAndSleep() stops the SysTick before it loads the counter from the reload register. Prior to this commit, this scenario caused xTickCount to jump ahead by the entire expected idle time (xExpectedIdleTime) because the current-count register is zero before it loads from the reload register. 3. Prior to return, vPortSuppressTicksAndSleep() attempts to start a short SysTick period when the current SysTick clock cycle has a lot of time remaining. Prior to this commit, this scenario could cause xTickCount to fall behind by as much as nearly one full tick because the short SysTick cycle never started. Note that #3 is partially fixed by https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/967acc9b200d3d4beeb289d9da9e88798074b431 even though that commit addresses a different issue. So this commit completes the partial fix. * Improve comments and name of preprocessor symbol Add a note in the code comments that SysTick requests an interrupt when decrementing from 1 to 0, so that's why stopping SysTick on zero is a special case. Readers might unknowingly assume that SysTick requests an interrupt when wrapping from 0 back to the load-register value. Reconsider new "_SETTING" suffix since "_CONFIG" suffix seems more descriptive. The code relies on *both* of these preprocessor symbols: portNVIC_SYSTICK_CLK_BIT portNVIC_SYSTICK_CLK_BIT_CONFIG **new** A meaningful suffix is really helpful to distinguish the two symbols. * Revert introduction of 2nd name for NVIC register When I added portNVIC_ICSR_REG I didn't realize there was already a portNVIC_INT_CTRL_REG, which identifies the same register. Not good to have both. Note that portNVIC_INT_CTRL_REG is defined in portmacro.h and is already used in this file (port.c). * Replicate to other Cortex M ports Also set a new fiddle factor based on tests with a CM4F. I used gcc, optimizing at -O1. Users can fine-tune as needed. Also add configSYSTICK_CLOCK_HZ to the CM0 ports to be just like the other Cortex M ports. This change allowed uniformity in the default tickless implementations across all Cortex M ports. And CM0 is likely to benefit from configSYSTICK_CLOCK_HZ, especially considering new CM0 devices with very fast CPU clock speeds. * Revert changes to IAR-CM0-portmacro.h portNVIC_INT_CTRL_REG was already defined in port.c. No need to define it in portmacro.h. * Handle edge cases with slow SysTick clock Co-authored-by: Cobus van Eeden <35851496+cobusve@users.noreply.github.com> Co-authored-by: abhidixi11 <44424462+abhidixi11@users.noreply.github.com> Co-authored-by: Joseph Julicher <jjulicher@mac.com> Co-authored-by: alfred gedeon <28123637+alfred2g@users.noreply.github.com>
2 years ago
/* Something other than the tick interrupt ended the sleep. */
/* Use the SysTick current-value register to determine the
* number of SysTick decrements remaining until the expected idle
* time would have ended. */
ulSysTickDecrementsLeft = portNVIC_SYSTICK_CURRENT_VALUE_REG;
#if ( portNVIC_SYSTICK_CLK_BIT_CONFIG != portNVIC_SYSTICK_CLK_BIT )
{
/* If the SysTick is not using the core clock, the current-
* value register might still be zero here. In that case, the
* SysTick didn't load from the reload register, and there are
* ulReloadValue decrements remaining in the expected idle
* time, not zero. */
if( ulSysTickDecrementsLeft == 0 )
{
ulSysTickDecrementsLeft = ulReloadValue;
}
}
#endif /* portNVIC_SYSTICK_CLK_BIT_CONFIG */
/* Work out how long the sleep lasted rounded to complete tick
* periods (not the ulReload value which accounted for part
* ticks). */
Tickless idle fixes/improvement (#59) * Fix tickless idle when stopping systick on zero... ...and don't stop SysTick at all in the eAbortSleep case. Prior to this commit, if vPortSuppressTicksAndSleep() happens to stop the SysTick on zero, then after tickless idle ends, xTickCount advances one full tick more than the time that actually elapsed as measured by the SysTick. See "bug 1" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - Idle task calls vPortSuppressTicksAndSleep(xExpectedIdleTime = 2). [Doesn't have to be "2" -- could be any number.] - vPortSuppressTicksAndSleep() stops SysTick, and the current-count register happens to stop on zero. - SysTick ISR executes, setting xPendedTicks = 1 - vPortSuppressTicksAndSleep() masks interrupts and calls eTaskConfirmSleepModeStatus() which confirms the sleep operation. *** - vPortSuppressTicksAndSleep() configures SysTick for 1 full tick (xExpectedIdleTime - 1) plus the current-count register (which is 0) - One tick period elapses in sleep. - SysTick wakes CPU, ISR executes and increments xPendedTicks to 2. - vPortSuppressTicksAndSleep() calls vTaskStepTick(1), then returns. - Idle task resumes scheduler, which increments xTickCount twice (for xPendedTicks = 2) In the end, two ticks elapsed as measured by SysTick, but the code increments xTickCount three times. The root cause is that the code assumes the SysTick current-count register always contains the number of SysTick counts remaining in the current tick period. However, when the current-count register is zero, there are ulTimerCountsForOneTick counts remaining, not zero. This error is not the kind of time slippage normally associated with tickless idle. *** Note that a recent commit https://github.com/FreeRTOS/FreeRTOS-Kernel/commit/e1b98f0 results in eAbortSleep in this case, due to xPendedTicks != 0. That commit does mostly resolve this bug without specifically mentioning it, and without this commit. But that resolution allows the code in port.c not to directly address the special case of stopping SysTick on zero in any code or comments. That commit also generates additional instances of eAbortSleep, and a second purpose of this commit is to optimize how vPortSuppressTicksAndSleep() behaves for eAbortSleep, as noted below. This commit also includes an optimization to avoid stopping the SysTick when eTaskConfirmSleepModeStatus() returns eAbortSleep. This optimization belongs with this fix because the method of handling the SysTick being stopped on zero changes with this optimization. * Fix imminent tick rescheduled after tickless idle Prior to this commit, if something other than systick wakes the CPU from tickless idle, vPortSuppressTicksAndSleep() might cause xTickCount to increment once too many times. See "bug 2" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - CPU is sleeping in vPortSuppressTicksAndSleep() - Something other than the SysTick wakes the CPU. - vPortSuppressTicksAndSleep() calculates the number of SysTick counts until the next tick. The bug occurs only if this number is small. - vPortSuppressTicksAndSleep() puts this small number into the SysTick reload register, and starts SysTick. - vPortSuppressTicksAndSleep() calls vTaskStepTick() - While vTaskStepTick() executes, the SysTick expires. The ISR pends because interrupts are masked, and SysTick starts a 2nd period still based on the small number of counts in its reload register. This 2nd period is undesirable and is likely to cause the error noted below. - vPortSuppressTicksAndSleep() puts the normal tick duration into the SysTick's reload register. - vPortSuppressTicksAndSleep() unmasks interrupts before the SysTick starts a new period based on the new value in the reload register. [This is a race condition that can go either way, but for the bug to occur, the race must play out this way.] - The pending SysTick ISR executes and increments xPendedTicks. - The SysTick expires again, finishing the second very small period, and starts a new period this time based on the full tick duration. - The SysTick ISR increments xPendedTicks (or xTickCount) even though only a tiny fraction of a tick period has elapsed since the previous tick. The bug occurs when *two* consecutive small periods of the SysTick are both counted as ticks. The root cause is a race caused by the small SysTick period. If vPortSuppressTicksAndSleep() unmasks interrupts *after* the small period expires but *before* the SysTick starts a period based on the full tick period, then two small periods are counted as ticks when only one should be counted. The end result is xTickCount advancing nearly one full tick more than time actually elapsed as measured by the SysTick. This is not the kind of time slippage normally associated with tickless idle. After this commit the code starts the SysTick and then immediately modifies the reload register to ensure the very short cycle (if any) is conducted only once. This strategy requires special consideration for the build option that configures SysTick to use a divided clock. To avoid waiting around for the SysTick to load value from the reload register, the new code temporarily configures the SysTick to use the undivided clock. The resulting timing error is typical for tickless idle. The error (commonly known as drift or slippage in kernel time) caused by this strategy is equivalent to one or two counts in ulStoppedTimerCompensation. This commit also updates comments and #define symbols related to the SysTick clock option. The SysTick can optionally be clocked by a divided version of the CPU clock (commonly divide-by-8). The new code in this commit adjusts these comments and symbols to make them clearer and more useful in configurations that use the divided clock. The fix made in this commit requires the use of these symbols, as noted in the code comments. * Fix tickless idle with alternate systick clocking Prior to this commit, in configurations using the alternate SysTick clocking, vPortSuppressTicksAndSleep() might cause xTickCount to jump ahead as much as the entire expected idle time or fall behind as much as one full tick compared to time as measured by the SysTick. SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. The SysTick has a configuration option to be clocked by an alternate clock besides the core clock. This alternate clock is MCU dependent. Scenarios Fixed --------------- The new code in this commit handles the following scenarios that were not handled correctly prior to this commit. 1. Before the sleep, vPortSuppressTicksAndSleep() stops the SysTick on zero, long after SysTick reached zero. Prior to this commit, this scenario caused xTickCount to jump ahead one full tick for the same reason documented here: https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/0c7b04bd3a745c52151abebc882eed3f811c4c81 2. After the sleep, vPortSuppressTicksAndSleep() stops the SysTick before it loads the counter from the reload register. Prior to this commit, this scenario caused xTickCount to jump ahead by the entire expected idle time (xExpectedIdleTime) because the current-count register is zero before it loads from the reload register. 3. Prior to return, vPortSuppressTicksAndSleep() attempts to start a short SysTick period when the current SysTick clock cycle has a lot of time remaining. Prior to this commit, this scenario could cause xTickCount to fall behind by as much as nearly one full tick because the short SysTick cycle never started. Note that #3 is partially fixed by https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/967acc9b200d3d4beeb289d9da9e88798074b431 even though that commit addresses a different issue. So this commit completes the partial fix. * Improve comments and name of preprocessor symbol Add a note in the code comments that SysTick requests an interrupt when decrementing from 1 to 0, so that's why stopping SysTick on zero is a special case. Readers might unknowingly assume that SysTick requests an interrupt when wrapping from 0 back to the load-register value. Reconsider new "_SETTING" suffix since "_CONFIG" suffix seems more descriptive. The code relies on *both* of these preprocessor symbols: portNVIC_SYSTICK_CLK_BIT portNVIC_SYSTICK_CLK_BIT_CONFIG **new** A meaningful suffix is really helpful to distinguish the two symbols. * Revert introduction of 2nd name for NVIC register When I added portNVIC_ICSR_REG I didn't realize there was already a portNVIC_INT_CTRL_REG, which identifies the same register. Not good to have both. Note that portNVIC_INT_CTRL_REG is defined in portmacro.h and is already used in this file (port.c). * Replicate to other Cortex M ports Also set a new fiddle factor based on tests with a CM4F. I used gcc, optimizing at -O1. Users can fine-tune as needed. Also add configSYSTICK_CLOCK_HZ to the CM0 ports to be just like the other Cortex M ports. This change allowed uniformity in the default tickless implementations across all Cortex M ports. And CM0 is likely to benefit from configSYSTICK_CLOCK_HZ, especially considering new CM0 devices with very fast CPU clock speeds. * Revert changes to IAR-CM0-portmacro.h portNVIC_INT_CTRL_REG was already defined in port.c. No need to define it in portmacro.h. * Handle edge cases with slow SysTick clock Co-authored-by: Cobus van Eeden <35851496+cobusve@users.noreply.github.com> Co-authored-by: abhidixi11 <44424462+abhidixi11@users.noreply.github.com> Co-authored-by: Joseph Julicher <jjulicher@mac.com> Co-authored-by: alfred gedeon <28123637+alfred2g@users.noreply.github.com>
2 years ago
ulCompletedSysTickDecrements = ( xExpectedIdleTime * ulTimerCountsForOneTick ) - ulSysTickDecrementsLeft;
/* How many complete tick periods passed while the processor
* was waiting? */
ulCompleteTickPeriods = ulCompletedSysTickDecrements / ulTimerCountsForOneTick;
/* The reload value is set to whatever fraction of a single tick
* period remains. */
portNVIC_SYSTICK_LOAD_REG = ( ( ulCompleteTickPeriods + 1UL ) * ulTimerCountsForOneTick ) - ulCompletedSysTickDecrements;
}
Tickless idle fixes/improvement (#59) * Fix tickless idle when stopping systick on zero... ...and don't stop SysTick at all in the eAbortSleep case. Prior to this commit, if vPortSuppressTicksAndSleep() happens to stop the SysTick on zero, then after tickless idle ends, xTickCount advances one full tick more than the time that actually elapsed as measured by the SysTick. See "bug 1" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - Idle task calls vPortSuppressTicksAndSleep(xExpectedIdleTime = 2). [Doesn't have to be "2" -- could be any number.] - vPortSuppressTicksAndSleep() stops SysTick, and the current-count register happens to stop on zero. - SysTick ISR executes, setting xPendedTicks = 1 - vPortSuppressTicksAndSleep() masks interrupts and calls eTaskConfirmSleepModeStatus() which confirms the sleep operation. *** - vPortSuppressTicksAndSleep() configures SysTick for 1 full tick (xExpectedIdleTime - 1) plus the current-count register (which is 0) - One tick period elapses in sleep. - SysTick wakes CPU, ISR executes and increments xPendedTicks to 2. - vPortSuppressTicksAndSleep() calls vTaskStepTick(1), then returns. - Idle task resumes scheduler, which increments xTickCount twice (for xPendedTicks = 2) In the end, two ticks elapsed as measured by SysTick, but the code increments xTickCount three times. The root cause is that the code assumes the SysTick current-count register always contains the number of SysTick counts remaining in the current tick period. However, when the current-count register is zero, there are ulTimerCountsForOneTick counts remaining, not zero. This error is not the kind of time slippage normally associated with tickless idle. *** Note that a recent commit https://github.com/FreeRTOS/FreeRTOS-Kernel/commit/e1b98f0 results in eAbortSleep in this case, due to xPendedTicks != 0. That commit does mostly resolve this bug without specifically mentioning it, and without this commit. But that resolution allows the code in port.c not to directly address the special case of stopping SysTick on zero in any code or comments. That commit also generates additional instances of eAbortSleep, and a second purpose of this commit is to optimize how vPortSuppressTicksAndSleep() behaves for eAbortSleep, as noted below. This commit also includes an optimization to avoid stopping the SysTick when eTaskConfirmSleepModeStatus() returns eAbortSleep. This optimization belongs with this fix because the method of handling the SysTick being stopped on zero changes with this optimization. * Fix imminent tick rescheduled after tickless idle Prior to this commit, if something other than systick wakes the CPU from tickless idle, vPortSuppressTicksAndSleep() might cause xTickCount to increment once too many times. See "bug 2" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - CPU is sleeping in vPortSuppressTicksAndSleep() - Something other than the SysTick wakes the CPU. - vPortSuppressTicksAndSleep() calculates the number of SysTick counts until the next tick. The bug occurs only if this number is small. - vPortSuppressTicksAndSleep() puts this small number into the SysTick reload register, and starts SysTick. - vPortSuppressTicksAndSleep() calls vTaskStepTick() - While vTaskStepTick() executes, the SysTick expires. The ISR pends because interrupts are masked, and SysTick starts a 2nd period still based on the small number of counts in its reload register. This 2nd period is undesirable and is likely to cause the error noted below. - vPortSuppressTicksAndSleep() puts the normal tick duration into the SysTick's reload register. - vPortSuppressTicksAndSleep() unmasks interrupts before the SysTick starts a new period based on the new value in the reload register. [This is a race condition that can go either way, but for the bug to occur, the race must play out this way.] - The pending SysTick ISR executes and increments xPendedTicks. - The SysTick expires again, finishing the second very small period, and starts a new period this time based on the full tick duration. - The SysTick ISR increments xPendedTicks (or xTickCount) even though only a tiny fraction of a tick period has elapsed since the previous tick. The bug occurs when *two* consecutive small periods of the SysTick are both counted as ticks. The root cause is a race caused by the small SysTick period. If vPortSuppressTicksAndSleep() unmasks interrupts *after* the small period expires but *before* the SysTick starts a period based on the full tick period, then two small periods are counted as ticks when only one should be counted. The end result is xTickCount advancing nearly one full tick more than time actually elapsed as measured by the SysTick. This is not the kind of time slippage normally associated with tickless idle. After this commit the code starts the SysTick and then immediately modifies the reload register to ensure the very short cycle (if any) is conducted only once. This strategy requires special consideration for the build option that configures SysTick to use a divided clock. To avoid waiting around for the SysTick to load value from the reload register, the new code temporarily configures the SysTick to use the undivided clock. The resulting timing error is typical for tickless idle. The error (commonly known as drift or slippage in kernel time) caused by this strategy is equivalent to one or two counts in ulStoppedTimerCompensation. This commit also updates comments and #define symbols related to the SysTick clock option. The SysTick can optionally be clocked by a divided version of the CPU clock (commonly divide-by-8). The new code in this commit adjusts these comments and symbols to make them clearer and more useful in configurations that use the divided clock. The fix made in this commit requires the use of these symbols, as noted in the code comments. * Fix tickless idle with alternate systick clocking Prior to this commit, in configurations using the alternate SysTick clocking, vPortSuppressTicksAndSleep() might cause xTickCount to jump ahead as much as the entire expected idle time or fall behind as much as one full tick compared to time as measured by the SysTick. SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. The SysTick has a configuration option to be clocked by an alternate clock besides the core clock. This alternate clock is MCU dependent. Scenarios Fixed --------------- The new code in this commit handles the following scenarios that were not handled correctly prior to this commit. 1. Before the sleep, vPortSuppressTicksAndSleep() stops the SysTick on zero, long after SysTick reached zero. Prior to this commit, this scenario caused xTickCount to jump ahead one full tick for the same reason documented here: https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/0c7b04bd3a745c52151abebc882eed3f811c4c81 2. After the sleep, vPortSuppressTicksAndSleep() stops the SysTick before it loads the counter from the reload register. Prior to this commit, this scenario caused xTickCount to jump ahead by the entire expected idle time (xExpectedIdleTime) because the current-count register is zero before it loads from the reload register. 3. Prior to return, vPortSuppressTicksAndSleep() attempts to start a short SysTick period when the current SysTick clock cycle has a lot of time remaining. Prior to this commit, this scenario could cause xTickCount to fall behind by as much as nearly one full tick because the short SysTick cycle never started. Note that #3 is partially fixed by https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/967acc9b200d3d4beeb289d9da9e88798074b431 even though that commit addresses a different issue. So this commit completes the partial fix. * Improve comments and name of preprocessor symbol Add a note in the code comments that SysTick requests an interrupt when decrementing from 1 to 0, so that's why stopping SysTick on zero is a special case. Readers might unknowingly assume that SysTick requests an interrupt when wrapping from 0 back to the load-register value. Reconsider new "_SETTING" suffix since "_CONFIG" suffix seems more descriptive. The code relies on *both* of these preprocessor symbols: portNVIC_SYSTICK_CLK_BIT portNVIC_SYSTICK_CLK_BIT_CONFIG **new** A meaningful suffix is really helpful to distinguish the two symbols. * Revert introduction of 2nd name for NVIC register When I added portNVIC_ICSR_REG I didn't realize there was already a portNVIC_INT_CTRL_REG, which identifies the same register. Not good to have both. Note that portNVIC_INT_CTRL_REG is defined in portmacro.h and is already used in this file (port.c). * Replicate to other Cortex M ports Also set a new fiddle factor based on tests with a CM4F. I used gcc, optimizing at -O1. Users can fine-tune as needed. Also add configSYSTICK_CLOCK_HZ to the CM0 ports to be just like the other Cortex M ports. This change allowed uniformity in the default tickless implementations across all Cortex M ports. And CM0 is likely to benefit from configSYSTICK_CLOCK_HZ, especially considering new CM0 devices with very fast CPU clock speeds. * Revert changes to IAR-CM0-portmacro.h portNVIC_INT_CTRL_REG was already defined in port.c. No need to define it in portmacro.h. * Handle edge cases with slow SysTick clock Co-authored-by: Cobus van Eeden <35851496+cobusve@users.noreply.github.com> Co-authored-by: abhidixi11 <44424462+abhidixi11@users.noreply.github.com> Co-authored-by: Joseph Julicher <jjulicher@mac.com> Co-authored-by: alfred gedeon <28123637+alfred2g@users.noreply.github.com>
2 years ago
/* Restart SysTick so it runs from portNVIC_SYSTICK_LOAD_REG again,
* then set portNVIC_SYSTICK_LOAD_REG back to its standard value. If
* the SysTick is not using the core clock, temporarily configure it to
* use the core clock. This configuration forces the SysTick to load
* from portNVIC_SYSTICK_LOAD_REG immediately instead of at the next
* cycle of the other clock. Then portNVIC_SYSTICK_LOAD_REG is ready
* to receive the standard value immediately. */
portNVIC_SYSTICK_CURRENT_VALUE_REG = 0UL;
Tickless idle fixes/improvement (#59) * Fix tickless idle when stopping systick on zero... ...and don't stop SysTick at all in the eAbortSleep case. Prior to this commit, if vPortSuppressTicksAndSleep() happens to stop the SysTick on zero, then after tickless idle ends, xTickCount advances one full tick more than the time that actually elapsed as measured by the SysTick. See "bug 1" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - Idle task calls vPortSuppressTicksAndSleep(xExpectedIdleTime = 2). [Doesn't have to be "2" -- could be any number.] - vPortSuppressTicksAndSleep() stops SysTick, and the current-count register happens to stop on zero. - SysTick ISR executes, setting xPendedTicks = 1 - vPortSuppressTicksAndSleep() masks interrupts and calls eTaskConfirmSleepModeStatus() which confirms the sleep operation. *** - vPortSuppressTicksAndSleep() configures SysTick for 1 full tick (xExpectedIdleTime - 1) plus the current-count register (which is 0) - One tick period elapses in sleep. - SysTick wakes CPU, ISR executes and increments xPendedTicks to 2. - vPortSuppressTicksAndSleep() calls vTaskStepTick(1), then returns. - Idle task resumes scheduler, which increments xTickCount twice (for xPendedTicks = 2) In the end, two ticks elapsed as measured by SysTick, but the code increments xTickCount three times. The root cause is that the code assumes the SysTick current-count register always contains the number of SysTick counts remaining in the current tick period. However, when the current-count register is zero, there are ulTimerCountsForOneTick counts remaining, not zero. This error is not the kind of time slippage normally associated with tickless idle. *** Note that a recent commit https://github.com/FreeRTOS/FreeRTOS-Kernel/commit/e1b98f0 results in eAbortSleep in this case, due to xPendedTicks != 0. That commit does mostly resolve this bug without specifically mentioning it, and without this commit. But that resolution allows the code in port.c not to directly address the special case of stopping SysTick on zero in any code or comments. That commit also generates additional instances of eAbortSleep, and a second purpose of this commit is to optimize how vPortSuppressTicksAndSleep() behaves for eAbortSleep, as noted below. This commit also includes an optimization to avoid stopping the SysTick when eTaskConfirmSleepModeStatus() returns eAbortSleep. This optimization belongs with this fix because the method of handling the SysTick being stopped on zero changes with this optimization. * Fix imminent tick rescheduled after tickless idle Prior to this commit, if something other than systick wakes the CPU from tickless idle, vPortSuppressTicksAndSleep() might cause xTickCount to increment once too many times. See "bug 2" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - CPU is sleeping in vPortSuppressTicksAndSleep() - Something other than the SysTick wakes the CPU. - vPortSuppressTicksAndSleep() calculates the number of SysTick counts until the next tick. The bug occurs only if this number is small. - vPortSuppressTicksAndSleep() puts this small number into the SysTick reload register, and starts SysTick. - vPortSuppressTicksAndSleep() calls vTaskStepTick() - While vTaskStepTick() executes, the SysTick expires. The ISR pends because interrupts are masked, and SysTick starts a 2nd period still based on the small number of counts in its reload register. This 2nd period is undesirable and is likely to cause the error noted below. - vPortSuppressTicksAndSleep() puts the normal tick duration into the SysTick's reload register. - vPortSuppressTicksAndSleep() unmasks interrupts before the SysTick starts a new period based on the new value in the reload register. [This is a race condition that can go either way, but for the bug to occur, the race must play out this way.] - The pending SysTick ISR executes and increments xPendedTicks. - The SysTick expires again, finishing the second very small period, and starts a new period this time based on the full tick duration. - The SysTick ISR increments xPendedTicks (or xTickCount) even though only a tiny fraction of a tick period has elapsed since the previous tick. The bug occurs when *two* consecutive small periods of the SysTick are both counted as ticks. The root cause is a race caused by the small SysTick period. If vPortSuppressTicksAndSleep() unmasks interrupts *after* the small period expires but *before* the SysTick starts a period based on the full tick period, then two small periods are counted as ticks when only one should be counted. The end result is xTickCount advancing nearly one full tick more than time actually elapsed as measured by the SysTick. This is not the kind of time slippage normally associated with tickless idle. After this commit the code starts the SysTick and then immediately modifies the reload register to ensure the very short cycle (if any) is conducted only once. This strategy requires special consideration for the build option that configures SysTick to use a divided clock. To avoid waiting around for the SysTick to load value from the reload register, the new code temporarily configures the SysTick to use the undivided clock. The resulting timing error is typical for tickless idle. The error (commonly known as drift or slippage in kernel time) caused by this strategy is equivalent to one or two counts in ulStoppedTimerCompensation. This commit also updates comments and #define symbols related to the SysTick clock option. The SysTick can optionally be clocked by a divided version of the CPU clock (commonly divide-by-8). The new code in this commit adjusts these comments and symbols to make them clearer and more useful in configurations that use the divided clock. The fix made in this commit requires the use of these symbols, as noted in the code comments. * Fix tickless idle with alternate systick clocking Prior to this commit, in configurations using the alternate SysTick clocking, vPortSuppressTicksAndSleep() might cause xTickCount to jump ahead as much as the entire expected idle time or fall behind as much as one full tick compared to time as measured by the SysTick. SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. The SysTick has a configuration option to be clocked by an alternate clock besides the core clock. This alternate clock is MCU dependent. Scenarios Fixed --------------- The new code in this commit handles the following scenarios that were not handled correctly prior to this commit. 1. Before the sleep, vPortSuppressTicksAndSleep() stops the SysTick on zero, long after SysTick reached zero. Prior to this commit, this scenario caused xTickCount to jump ahead one full tick for the same reason documented here: https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/0c7b04bd3a745c52151abebc882eed3f811c4c81 2. After the sleep, vPortSuppressTicksAndSleep() stops the SysTick before it loads the counter from the reload register. Prior to this commit, this scenario caused xTickCount to jump ahead by the entire expected idle time (xExpectedIdleTime) because the current-count register is zero before it loads from the reload register. 3. Prior to return, vPortSuppressTicksAndSleep() attempts to start a short SysTick period when the current SysTick clock cycle has a lot of time remaining. Prior to this commit, this scenario could cause xTickCount to fall behind by as much as nearly one full tick because the short SysTick cycle never started. Note that #3 is partially fixed by https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/967acc9b200d3d4beeb289d9da9e88798074b431 even though that commit addresses a different issue. So this commit completes the partial fix. * Improve comments and name of preprocessor symbol Add a note in the code comments that SysTick requests an interrupt when decrementing from 1 to 0, so that's why stopping SysTick on zero is a special case. Readers might unknowingly assume that SysTick requests an interrupt when wrapping from 0 back to the load-register value. Reconsider new "_SETTING" suffix since "_CONFIG" suffix seems more descriptive. The code relies on *both* of these preprocessor symbols: portNVIC_SYSTICK_CLK_BIT portNVIC_SYSTICK_CLK_BIT_CONFIG **new** A meaningful suffix is really helpful to distinguish the two symbols. * Revert introduction of 2nd name for NVIC register When I added portNVIC_ICSR_REG I didn't realize there was already a portNVIC_INT_CTRL_REG, which identifies the same register. Not good to have both. Note that portNVIC_INT_CTRL_REG is defined in portmacro.h and is already used in this file (port.c). * Replicate to other Cortex M ports Also set a new fiddle factor based on tests with a CM4F. I used gcc, optimizing at -O1. Users can fine-tune as needed. Also add configSYSTICK_CLOCK_HZ to the CM0 ports to be just like the other Cortex M ports. This change allowed uniformity in the default tickless implementations across all Cortex M ports. And CM0 is likely to benefit from configSYSTICK_CLOCK_HZ, especially considering new CM0 devices with very fast CPU clock speeds. * Revert changes to IAR-CM0-portmacro.h portNVIC_INT_CTRL_REG was already defined in port.c. No need to define it in portmacro.h. * Handle edge cases with slow SysTick clock Co-authored-by: Cobus van Eeden <35851496+cobusve@users.noreply.github.com> Co-authored-by: abhidixi11 <44424462+abhidixi11@users.noreply.github.com> Co-authored-by: Joseph Julicher <jjulicher@mac.com> Co-authored-by: alfred gedeon <28123637+alfred2g@users.noreply.github.com>
2 years ago
portNVIC_SYSTICK_CTRL_REG = portNVIC_SYSTICK_CLK_BIT | portNVIC_SYSTICK_INT_BIT | portNVIC_SYSTICK_ENABLE_BIT;
#if ( portNVIC_SYSTICK_CLK_BIT_CONFIG == portNVIC_SYSTICK_CLK_BIT )
{
portNVIC_SYSTICK_LOAD_REG = ulTimerCountsForOneTick - 1UL;
}
#else
{
/* The temporary usage of the core clock has served its purpose,
* as described above. Resume usage of the other clock. */
portNVIC_SYSTICK_CTRL_REG = portNVIC_SYSTICK_CLK_BIT | portNVIC_SYSTICK_INT_BIT;
if( ( portNVIC_SYSTICK_CTRL_REG & portNVIC_SYSTICK_COUNT_FLAG_BIT ) != 0 )
{
/* The partial tick period already ended. Be sure the SysTick
* counts it only once. */
portNVIC_SYSTICK_CURRENT_VALUE_REG = 0;
}
portNVIC_SYSTICK_LOAD_REG = ulTimerCountsForOneTick - 1UL;
portNVIC_SYSTICK_CTRL_REG = portNVIC_SYSTICK_CLK_BIT_CONFIG | portNVIC_SYSTICK_INT_BIT | portNVIC_SYSTICK_ENABLE_BIT;
}
#endif /* portNVIC_SYSTICK_CLK_BIT_CONFIG */
/* Step the tick to account for any tick periods that elapsed. */
vTaskStepTick( ulCompleteTickPeriods );
/* Exit with interrupts enabled. */
__asm volatile ( "cpsie i" ::: "memory" );
}
}
#endif /* configUSE_TICKLESS_IDLE */
/*-----------------------------------------------------------*/
__attribute__( ( weak ) ) void vPortSetupTimerInterrupt( void ) /* PRIVILEGED_FUNCTION */
{
/* Calculate the constants required to configure the tick interrupt. */
#if ( configUSE_TICKLESS_IDLE == 1 )
Tickless idle fixes/improvement (#59) * Fix tickless idle when stopping systick on zero... ...and don't stop SysTick at all in the eAbortSleep case. Prior to this commit, if vPortSuppressTicksAndSleep() happens to stop the SysTick on zero, then after tickless idle ends, xTickCount advances one full tick more than the time that actually elapsed as measured by the SysTick. See "bug 1" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - Idle task calls vPortSuppressTicksAndSleep(xExpectedIdleTime = 2). [Doesn't have to be "2" -- could be any number.] - vPortSuppressTicksAndSleep() stops SysTick, and the current-count register happens to stop on zero. - SysTick ISR executes, setting xPendedTicks = 1 - vPortSuppressTicksAndSleep() masks interrupts and calls eTaskConfirmSleepModeStatus() which confirms the sleep operation. *** - vPortSuppressTicksAndSleep() configures SysTick for 1 full tick (xExpectedIdleTime - 1) plus the current-count register (which is 0) - One tick period elapses in sleep. - SysTick wakes CPU, ISR executes and increments xPendedTicks to 2. - vPortSuppressTicksAndSleep() calls vTaskStepTick(1), then returns. - Idle task resumes scheduler, which increments xTickCount twice (for xPendedTicks = 2) In the end, two ticks elapsed as measured by SysTick, but the code increments xTickCount three times. The root cause is that the code assumes the SysTick current-count register always contains the number of SysTick counts remaining in the current tick period. However, when the current-count register is zero, there are ulTimerCountsForOneTick counts remaining, not zero. This error is not the kind of time slippage normally associated with tickless idle. *** Note that a recent commit https://github.com/FreeRTOS/FreeRTOS-Kernel/commit/e1b98f0 results in eAbortSleep in this case, due to xPendedTicks != 0. That commit does mostly resolve this bug without specifically mentioning it, and without this commit. But that resolution allows the code in port.c not to directly address the special case of stopping SysTick on zero in any code or comments. That commit also generates additional instances of eAbortSleep, and a second purpose of this commit is to optimize how vPortSuppressTicksAndSleep() behaves for eAbortSleep, as noted below. This commit also includes an optimization to avoid stopping the SysTick when eTaskConfirmSleepModeStatus() returns eAbortSleep. This optimization belongs with this fix because the method of handling the SysTick being stopped on zero changes with this optimization. * Fix imminent tick rescheduled after tickless idle Prior to this commit, if something other than systick wakes the CPU from tickless idle, vPortSuppressTicksAndSleep() might cause xTickCount to increment once too many times. See "bug 2" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - CPU is sleeping in vPortSuppressTicksAndSleep() - Something other than the SysTick wakes the CPU. - vPortSuppressTicksAndSleep() calculates the number of SysTick counts until the next tick. The bug occurs only if this number is small. - vPortSuppressTicksAndSleep() puts this small number into the SysTick reload register, and starts SysTick. - vPortSuppressTicksAndSleep() calls vTaskStepTick() - While vTaskStepTick() executes, the SysTick expires. The ISR pends because interrupts are masked, and SysTick starts a 2nd period still based on the small number of counts in its reload register. This 2nd period is undesirable and is likely to cause the error noted below. - vPortSuppressTicksAndSleep() puts the normal tick duration into the SysTick's reload register. - vPortSuppressTicksAndSleep() unmasks interrupts before the SysTick starts a new period based on the new value in the reload register. [This is a race condition that can go either way, but for the bug to occur, the race must play out this way.] - The pending SysTick ISR executes and increments xPendedTicks. - The SysTick expires again, finishing the second very small period, and starts a new period this time based on the full tick duration. - The SysTick ISR increments xPendedTicks (or xTickCount) even though only a tiny fraction of a tick period has elapsed since the previous tick. The bug occurs when *two* consecutive small periods of the SysTick are both counted as ticks. The root cause is a race caused by the small SysTick period. If vPortSuppressTicksAndSleep() unmasks interrupts *after* the small period expires but *before* the SysTick starts a period based on the full tick period, then two small periods are counted as ticks when only one should be counted. The end result is xTickCount advancing nearly one full tick more than time actually elapsed as measured by the SysTick. This is not the kind of time slippage normally associated with tickless idle. After this commit the code starts the SysTick and then immediately modifies the reload register to ensure the very short cycle (if any) is conducted only once. This strategy requires special consideration for the build option that configures SysTick to use a divided clock. To avoid waiting around for the SysTick to load value from the reload register, the new code temporarily configures the SysTick to use the undivided clock. The resulting timing error is typical for tickless idle. The error (commonly known as drift or slippage in kernel time) caused by this strategy is equivalent to one or two counts in ulStoppedTimerCompensation. This commit also updates comments and #define symbols related to the SysTick clock option. The SysTick can optionally be clocked by a divided version of the CPU clock (commonly divide-by-8). The new code in this commit adjusts these comments and symbols to make them clearer and more useful in configurations that use the divided clock. The fix made in this commit requires the use of these symbols, as noted in the code comments. * Fix tickless idle with alternate systick clocking Prior to this commit, in configurations using the alternate SysTick clocking, vPortSuppressTicksAndSleep() might cause xTickCount to jump ahead as much as the entire expected idle time or fall behind as much as one full tick compared to time as measured by the SysTick. SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. The SysTick has a configuration option to be clocked by an alternate clock besides the core clock. This alternate clock is MCU dependent. Scenarios Fixed --------------- The new code in this commit handles the following scenarios that were not handled correctly prior to this commit. 1. Before the sleep, vPortSuppressTicksAndSleep() stops the SysTick on zero, long after SysTick reached zero. Prior to this commit, this scenario caused xTickCount to jump ahead one full tick for the same reason documented here: https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/0c7b04bd3a745c52151abebc882eed3f811c4c81 2. After the sleep, vPortSuppressTicksAndSleep() stops the SysTick before it loads the counter from the reload register. Prior to this commit, this scenario caused xTickCount to jump ahead by the entire expected idle time (xExpectedIdleTime) because the current-count register is zero before it loads from the reload register. 3. Prior to return, vPortSuppressTicksAndSleep() attempts to start a short SysTick period when the current SysTick clock cycle has a lot of time remaining. Prior to this commit, this scenario could cause xTickCount to fall behind by as much as nearly one full tick because the short SysTick cycle never started. Note that #3 is partially fixed by https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/967acc9b200d3d4beeb289d9da9e88798074b431 even though that commit addresses a different issue. So this commit completes the partial fix. * Improve comments and name of preprocessor symbol Add a note in the code comments that SysTick requests an interrupt when decrementing from 1 to 0, so that's why stopping SysTick on zero is a special case. Readers might unknowingly assume that SysTick requests an interrupt when wrapping from 0 back to the load-register value. Reconsider new "_SETTING" suffix since "_CONFIG" suffix seems more descriptive. The code relies on *both* of these preprocessor symbols: portNVIC_SYSTICK_CLK_BIT portNVIC_SYSTICK_CLK_BIT_CONFIG **new** A meaningful suffix is really helpful to distinguish the two symbols. * Revert introduction of 2nd name for NVIC register When I added portNVIC_ICSR_REG I didn't realize there was already a portNVIC_INT_CTRL_REG, which identifies the same register. Not good to have both. Note that portNVIC_INT_CTRL_REG is defined in portmacro.h and is already used in this file (port.c). * Replicate to other Cortex M ports Also set a new fiddle factor based on tests with a CM4F. I used gcc, optimizing at -O1. Users can fine-tune as needed. Also add configSYSTICK_CLOCK_HZ to the CM0 ports to be just like the other Cortex M ports. This change allowed uniformity in the default tickless implementations across all Cortex M ports. And CM0 is likely to benefit from configSYSTICK_CLOCK_HZ, especially considering new CM0 devices with very fast CPU clock speeds. * Revert changes to IAR-CM0-portmacro.h portNVIC_INT_CTRL_REG was already defined in port.c. No need to define it in portmacro.h. * Handle edge cases with slow SysTick clock Co-authored-by: Cobus van Eeden <35851496+cobusve@users.noreply.github.com> Co-authored-by: abhidixi11 <44424462+abhidixi11@users.noreply.github.com> Co-authored-by: Joseph Julicher <jjulicher@mac.com> Co-authored-by: alfred gedeon <28123637+alfred2g@users.noreply.github.com>
2 years ago
{
ulTimerCountsForOneTick = ( configSYSTICK_CLOCK_HZ / configTICK_RATE_HZ );
xMaximumPossibleSuppressedTicks = portMAX_24_BIT_NUMBER / ulTimerCountsForOneTick;
ulStoppedTimerCompensation = portMISSED_COUNTS_FACTOR / ( configCPU_CLOCK_HZ / configSYSTICK_CLOCK_HZ );
}
#endif /* configUSE_TICKLESS_IDLE */
/* Stop and reset SysTick.
*
* QEMU versions older than 7.0.0 contain a bug which causes an error if we
* enable SysTick without first selecting a valid clock source. We trigger
* the bug if we change clock sources from a clock with a zero clock period
* to one with a nonzero clock period and enable Systick at the same time.
* So we configure the CLKSOURCE bit here, prior to setting the ENABLE bit.
* This workaround avoids the bug in QEMU versions older than 7.0.0. */
portNVIC_SYSTICK_CTRL_REG = portNVIC_SYSTICK_CLK_BIT_CONFIG;
portNVIC_SYSTICK_CURRENT_VALUE_REG = 0UL;
/* Configure SysTick to interrupt at the requested rate. */
portNVIC_SYSTICK_LOAD_REG = ( configSYSTICK_CLOCK_HZ / configTICK_RATE_HZ ) - 1UL;
Tickless idle fixes/improvement (#59) * Fix tickless idle when stopping systick on zero... ...and don't stop SysTick at all in the eAbortSleep case. Prior to this commit, if vPortSuppressTicksAndSleep() happens to stop the SysTick on zero, then after tickless idle ends, xTickCount advances one full tick more than the time that actually elapsed as measured by the SysTick. See "bug 1" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - Idle task calls vPortSuppressTicksAndSleep(xExpectedIdleTime = 2). [Doesn't have to be "2" -- could be any number.] - vPortSuppressTicksAndSleep() stops SysTick, and the current-count register happens to stop on zero. - SysTick ISR executes, setting xPendedTicks = 1 - vPortSuppressTicksAndSleep() masks interrupts and calls eTaskConfirmSleepModeStatus() which confirms the sleep operation. *** - vPortSuppressTicksAndSleep() configures SysTick for 1 full tick (xExpectedIdleTime - 1) plus the current-count register (which is 0) - One tick period elapses in sleep. - SysTick wakes CPU, ISR executes and increments xPendedTicks to 2. - vPortSuppressTicksAndSleep() calls vTaskStepTick(1), then returns. - Idle task resumes scheduler, which increments xTickCount twice (for xPendedTicks = 2) In the end, two ticks elapsed as measured by SysTick, but the code increments xTickCount three times. The root cause is that the code assumes the SysTick current-count register always contains the number of SysTick counts remaining in the current tick period. However, when the current-count register is zero, there are ulTimerCountsForOneTick counts remaining, not zero. This error is not the kind of time slippage normally associated with tickless idle. *** Note that a recent commit https://github.com/FreeRTOS/FreeRTOS-Kernel/commit/e1b98f0 results in eAbortSleep in this case, due to xPendedTicks != 0. That commit does mostly resolve this bug without specifically mentioning it, and without this commit. But that resolution allows the code in port.c not to directly address the special case of stopping SysTick on zero in any code or comments. That commit also generates additional instances of eAbortSleep, and a second purpose of this commit is to optimize how vPortSuppressTicksAndSleep() behaves for eAbortSleep, as noted below. This commit also includes an optimization to avoid stopping the SysTick when eTaskConfirmSleepModeStatus() returns eAbortSleep. This optimization belongs with this fix because the method of handling the SysTick being stopped on zero changes with this optimization. * Fix imminent tick rescheduled after tickless idle Prior to this commit, if something other than systick wakes the CPU from tickless idle, vPortSuppressTicksAndSleep() might cause xTickCount to increment once too many times. See "bug 2" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - CPU is sleeping in vPortSuppressTicksAndSleep() - Something other than the SysTick wakes the CPU. - vPortSuppressTicksAndSleep() calculates the number of SysTick counts until the next tick. The bug occurs only if this number is small. - vPortSuppressTicksAndSleep() puts this small number into the SysTick reload register, and starts SysTick. - vPortSuppressTicksAndSleep() calls vTaskStepTick() - While vTaskStepTick() executes, the SysTick expires. The ISR pends because interrupts are masked, and SysTick starts a 2nd period still based on the small number of counts in its reload register. This 2nd period is undesirable and is likely to cause the error noted below. - vPortSuppressTicksAndSleep() puts the normal tick duration into the SysTick's reload register. - vPortSuppressTicksAndSleep() unmasks interrupts before the SysTick starts a new period based on the new value in the reload register. [This is a race condition that can go either way, but for the bug to occur, the race must play out this way.] - The pending SysTick ISR executes and increments xPendedTicks. - The SysTick expires again, finishing the second very small period, and starts a new period this time based on the full tick duration. - The SysTick ISR increments xPendedTicks (or xTickCount) even though only a tiny fraction of a tick period has elapsed since the previous tick. The bug occurs when *two* consecutive small periods of the SysTick are both counted as ticks. The root cause is a race caused by the small SysTick period. If vPortSuppressTicksAndSleep() unmasks interrupts *after* the small period expires but *before* the SysTick starts a period based on the full tick period, then two small periods are counted as ticks when only one should be counted. The end result is xTickCount advancing nearly one full tick more than time actually elapsed as measured by the SysTick. This is not the kind of time slippage normally associated with tickless idle. After this commit the code starts the SysTick and then immediately modifies the reload register to ensure the very short cycle (if any) is conducted only once. This strategy requires special consideration for the build option that configures SysTick to use a divided clock. To avoid waiting around for the SysTick to load value from the reload register, the new code temporarily configures the SysTick to use the undivided clock. The resulting timing error is typical for tickless idle. The error (commonly known as drift or slippage in kernel time) caused by this strategy is equivalent to one or two counts in ulStoppedTimerCompensation. This commit also updates comments and #define symbols related to the SysTick clock option. The SysTick can optionally be clocked by a divided version of the CPU clock (commonly divide-by-8). The new code in this commit adjusts these comments and symbols to make them clearer and more useful in configurations that use the divided clock. The fix made in this commit requires the use of these symbols, as noted in the code comments. * Fix tickless idle with alternate systick clocking Prior to this commit, in configurations using the alternate SysTick clocking, vPortSuppressTicksAndSleep() might cause xTickCount to jump ahead as much as the entire expected idle time or fall behind as much as one full tick compared to time as measured by the SysTick. SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. The SysTick has a configuration option to be clocked by an alternate clock besides the core clock. This alternate clock is MCU dependent. Scenarios Fixed --------------- The new code in this commit handles the following scenarios that were not handled correctly prior to this commit. 1. Before the sleep, vPortSuppressTicksAndSleep() stops the SysTick on zero, long after SysTick reached zero. Prior to this commit, this scenario caused xTickCount to jump ahead one full tick for the same reason documented here: https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/0c7b04bd3a745c52151abebc882eed3f811c4c81 2. After the sleep, vPortSuppressTicksAndSleep() stops the SysTick before it loads the counter from the reload register. Prior to this commit, this scenario caused xTickCount to jump ahead by the entire expected idle time (xExpectedIdleTime) because the current-count register is zero before it loads from the reload register. 3. Prior to return, vPortSuppressTicksAndSleep() attempts to start a short SysTick period when the current SysTick clock cycle has a lot of time remaining. Prior to this commit, this scenario could cause xTickCount to fall behind by as much as nearly one full tick because the short SysTick cycle never started. Note that #3 is partially fixed by https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/967acc9b200d3d4beeb289d9da9e88798074b431 even though that commit addresses a different issue. So this commit completes the partial fix. * Improve comments and name of preprocessor symbol Add a note in the code comments that SysTick requests an interrupt when decrementing from 1 to 0, so that's why stopping SysTick on zero is a special case. Readers might unknowingly assume that SysTick requests an interrupt when wrapping from 0 back to the load-register value. Reconsider new "_SETTING" suffix since "_CONFIG" suffix seems more descriptive. The code relies on *both* of these preprocessor symbols: portNVIC_SYSTICK_CLK_BIT portNVIC_SYSTICK_CLK_BIT_CONFIG **new** A meaningful suffix is really helpful to distinguish the two symbols. * Revert introduction of 2nd name for NVIC register When I added portNVIC_ICSR_REG I didn't realize there was already a portNVIC_INT_CTRL_REG, which identifies the same register. Not good to have both. Note that portNVIC_INT_CTRL_REG is defined in portmacro.h and is already used in this file (port.c). * Replicate to other Cortex M ports Also set a new fiddle factor based on tests with a CM4F. I used gcc, optimizing at -O1. Users can fine-tune as needed. Also add configSYSTICK_CLOCK_HZ to the CM0 ports to be just like the other Cortex M ports. This change allowed uniformity in the default tickless implementations across all Cortex M ports. And CM0 is likely to benefit from configSYSTICK_CLOCK_HZ, especially considering new CM0 devices with very fast CPU clock speeds. * Revert changes to IAR-CM0-portmacro.h portNVIC_INT_CTRL_REG was already defined in port.c. No need to define it in portmacro.h. * Handle edge cases with slow SysTick clock Co-authored-by: Cobus van Eeden <35851496+cobusve@users.noreply.github.com> Co-authored-by: abhidixi11 <44424462+abhidixi11@users.noreply.github.com> Co-authored-by: Joseph Julicher <jjulicher@mac.com> Co-authored-by: alfred gedeon <28123637+alfred2g@users.noreply.github.com>
2 years ago
portNVIC_SYSTICK_CTRL_REG = portNVIC_SYSTICK_CLK_BIT_CONFIG | portNVIC_SYSTICK_INT_BIT | portNVIC_SYSTICK_ENABLE_BIT;
}
/*-----------------------------------------------------------*/
static void prvTaskExitError( void )
{
volatile uint32_t ulDummy = 0UL;
/* A function that implements a task must not exit or attempt to return to
* its caller as there is nothing to return to. If a task wants to exit it
* should instead call vTaskDelete( NULL ). Artificially force an assert()
* to be triggered if configASSERT() is defined, then stop here so
* application writers can catch the error. */
configASSERT( ulCriticalNesting == ~0UL );
portDISABLE_INTERRUPTS();
while( ulDummy == 0 )
{
/* This file calls prvTaskExitError() after the scheduler has been
* started to remove a compiler warning about the function being
* defined but never called. ulDummy is used purely to quieten other
* warnings about code appearing after this function is called - making
* ulDummy volatile makes the compiler think the function could return
* and therefore not output an 'unreachable code' warning for code that
* appears after it. */
}
}
/*-----------------------------------------------------------*/
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
#if ( configENABLE_MPU == 1 )
static uint32_t prvGetRegionAccessPermissions( uint32_t ulRBARValue ) /* PRIVILEGED_FUNCTION */
{
uint32_t ulAccessPermissions = 0;
if( ( ulRBARValue & portMPU_RBAR_ACCESS_PERMISSIONS_MASK ) == portMPU_REGION_READ_ONLY )
{
ulAccessPermissions = tskMPU_READ_PERMISSION;
}
if( ( ulRBARValue & portMPU_RBAR_ACCESS_PERMISSIONS_MASK ) == portMPU_REGION_READ_WRITE )
{
ulAccessPermissions = ( tskMPU_READ_PERMISSION | tskMPU_WRITE_PERMISSION );
}
return ulAccessPermissions;
}
#endif /* configENABLE_MPU */
/*-----------------------------------------------------------*/
#if ( configENABLE_MPU == 1 )
static void prvSetupMPU( void ) /* PRIVILEGED_FUNCTION */
{
#if defined( __ARMCC_VERSION )
/* Declaration when these variable are defined in code instead of being
* exported from linker scripts. */
extern uint32_t * __privileged_functions_start__;
extern uint32_t * __privileged_functions_end__;
extern uint32_t * __syscalls_flash_start__;
extern uint32_t * __syscalls_flash_end__;
extern uint32_t * __unprivileged_flash_start__;
extern uint32_t * __unprivileged_flash_end__;
extern uint32_t * __privileged_sram_start__;
extern uint32_t * __privileged_sram_end__;
#else /* if defined( __ARMCC_VERSION ) */
/* Declaration when these variable are exported from linker scripts. */
extern uint32_t __privileged_functions_start__[];
extern uint32_t __privileged_functions_end__[];
extern uint32_t __syscalls_flash_start__[];
extern uint32_t __syscalls_flash_end__[];
extern uint32_t __unprivileged_flash_start__[];
extern uint32_t __unprivileged_flash_end__[];
extern uint32_t __privileged_sram_start__[];
extern uint32_t __privileged_sram_end__[];
#endif /* defined( __ARMCC_VERSION ) */
/* The only permitted number of regions are 8 or 16. */
configASSERT( ( configTOTAL_MPU_REGIONS == 8 ) || ( configTOTAL_MPU_REGIONS == 16 ) );
/* Ensure that the configTOTAL_MPU_REGIONS is configured correctly. */
configASSERT( portMPU_TYPE_REG == portEXPECTED_MPU_TYPE_VALUE );
/* Check that the MPU is present. */
if( portMPU_TYPE_REG == portEXPECTED_MPU_TYPE_VALUE )
{
/* MAIR0 - Index 0. */
portMPU_MAIR0_REG |= ( ( portMPU_NORMAL_MEMORY_BUFFERABLE_CACHEABLE << portMPU_MAIR_ATTR0_POS ) & portMPU_MAIR_ATTR0_MASK );
/* MAIR0 - Index 1. */
portMPU_MAIR0_REG |= ( ( portMPU_DEVICE_MEMORY_nGnRE << portMPU_MAIR_ATTR1_POS ) & portMPU_MAIR_ATTR1_MASK );
/* Setup privileged flash as Read Only so that privileged tasks can
* read it but not modify. */
portMPU_RNR_REG = portPRIVILEGED_FLASH_REGION;
portMPU_RBAR_REG = ( ( ( uint32_t ) __privileged_functions_start__ ) & portMPU_RBAR_ADDRESS_MASK ) |
( portMPU_REGION_NON_SHAREABLE ) |
( portMPU_REGION_PRIVILEGED_READ_ONLY );
portMPU_RLAR_REG = ( ( ( uint32_t ) __privileged_functions_end__ ) & portMPU_RLAR_ADDRESS_MASK ) |
( portMPU_RLAR_ATTR_INDEX0 ) |
( portMPU_RLAR_REGION_ENABLE );
/* Setup unprivileged flash as Read Only by both privileged and
* unprivileged tasks. All tasks can read it but no-one can modify. */
portMPU_RNR_REG = portUNPRIVILEGED_FLASH_REGION;
portMPU_RBAR_REG = ( ( ( uint32_t ) __unprivileged_flash_start__ ) & portMPU_RBAR_ADDRESS_MASK ) |
( portMPU_REGION_NON_SHAREABLE ) |
( portMPU_REGION_READ_ONLY );
portMPU_RLAR_REG = ( ( ( uint32_t ) __unprivileged_flash_end__ ) & portMPU_RLAR_ADDRESS_MASK ) |
( portMPU_RLAR_ATTR_INDEX0 ) |
( portMPU_RLAR_REGION_ENABLE );
/* Setup unprivileged syscalls flash as Read Only by both privileged
* and unprivileged tasks. All tasks can read it but no-one can modify. */
portMPU_RNR_REG = portUNPRIVILEGED_SYSCALLS_REGION;
portMPU_RBAR_REG = ( ( ( uint32_t ) __syscalls_flash_start__ ) & portMPU_RBAR_ADDRESS_MASK ) |
( portMPU_REGION_NON_SHAREABLE ) |
( portMPU_REGION_READ_ONLY );
portMPU_RLAR_REG = ( ( ( uint32_t ) __syscalls_flash_end__ ) & portMPU_RLAR_ADDRESS_MASK ) |
( portMPU_RLAR_ATTR_INDEX0 ) |
( portMPU_RLAR_REGION_ENABLE );
/* Setup RAM containing kernel data for privileged access only. */
portMPU_RNR_REG = portPRIVILEGED_RAM_REGION;
portMPU_RBAR_REG = ( ( ( uint32_t ) __privileged_sram_start__ ) & portMPU_RBAR_ADDRESS_MASK ) |
( portMPU_REGION_NON_SHAREABLE ) |
( portMPU_REGION_PRIVILEGED_READ_WRITE ) |
( portMPU_REGION_EXECUTE_NEVER );
portMPU_RLAR_REG = ( ( ( uint32_t ) __privileged_sram_end__ ) & portMPU_RLAR_ADDRESS_MASK ) |
( portMPU_RLAR_ATTR_INDEX0 ) |
( portMPU_RLAR_REGION_ENABLE );
/* Enable mem fault. */
portSCB_SYS_HANDLER_CTRL_STATE_REG |= portSCB_MEM_FAULT_ENABLE_BIT;
/* Enable MPU with privileged background access i.e. unmapped
* regions have privileged access. */
portMPU_CTRL_REG |= ( portMPU_PRIV_BACKGROUND_ENABLE_BIT | portMPU_ENABLE_BIT );
}
}
#endif /* configENABLE_MPU */
/*-----------------------------------------------------------*/
#if ( configENABLE_FPU == 1 )
static void prvSetupFPU( void ) /* PRIVILEGED_FUNCTION */
{
#if ( configENABLE_TRUSTZONE == 1 )
Tickless idle fixes/improvement (#59) * Fix tickless idle when stopping systick on zero... ...and don't stop SysTick at all in the eAbortSleep case. Prior to this commit, if vPortSuppressTicksAndSleep() happens to stop the SysTick on zero, then after tickless idle ends, xTickCount advances one full tick more than the time that actually elapsed as measured by the SysTick. See "bug 1" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - Idle task calls vPortSuppressTicksAndSleep(xExpectedIdleTime = 2). [Doesn't have to be "2" -- could be any number.] - vPortSuppressTicksAndSleep() stops SysTick, and the current-count register happens to stop on zero. - SysTick ISR executes, setting xPendedTicks = 1 - vPortSuppressTicksAndSleep() masks interrupts and calls eTaskConfirmSleepModeStatus() which confirms the sleep operation. *** - vPortSuppressTicksAndSleep() configures SysTick for 1 full tick (xExpectedIdleTime - 1) plus the current-count register (which is 0) - One tick period elapses in sleep. - SysTick wakes CPU, ISR executes and increments xPendedTicks to 2. - vPortSuppressTicksAndSleep() calls vTaskStepTick(1), then returns. - Idle task resumes scheduler, which increments xTickCount twice (for xPendedTicks = 2) In the end, two ticks elapsed as measured by SysTick, but the code increments xTickCount three times. The root cause is that the code assumes the SysTick current-count register always contains the number of SysTick counts remaining in the current tick period. However, when the current-count register is zero, there are ulTimerCountsForOneTick counts remaining, not zero. This error is not the kind of time slippage normally associated with tickless idle. *** Note that a recent commit https://github.com/FreeRTOS/FreeRTOS-Kernel/commit/e1b98f0 results in eAbortSleep in this case, due to xPendedTicks != 0. That commit does mostly resolve this bug without specifically mentioning it, and without this commit. But that resolution allows the code in port.c not to directly address the special case of stopping SysTick on zero in any code or comments. That commit also generates additional instances of eAbortSleep, and a second purpose of this commit is to optimize how vPortSuppressTicksAndSleep() behaves for eAbortSleep, as noted below. This commit also includes an optimization to avoid stopping the SysTick when eTaskConfirmSleepModeStatus() returns eAbortSleep. This optimization belongs with this fix because the method of handling the SysTick being stopped on zero changes with this optimization. * Fix imminent tick rescheduled after tickless idle Prior to this commit, if something other than systick wakes the CPU from tickless idle, vPortSuppressTicksAndSleep() might cause xTickCount to increment once too many times. See "bug 2" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - CPU is sleeping in vPortSuppressTicksAndSleep() - Something other than the SysTick wakes the CPU. - vPortSuppressTicksAndSleep() calculates the number of SysTick counts until the next tick. The bug occurs only if this number is small. - vPortSuppressTicksAndSleep() puts this small number into the SysTick reload register, and starts SysTick. - vPortSuppressTicksAndSleep() calls vTaskStepTick() - While vTaskStepTick() executes, the SysTick expires. The ISR pends because interrupts are masked, and SysTick starts a 2nd period still based on the small number of counts in its reload register. This 2nd period is undesirable and is likely to cause the error noted below. - vPortSuppressTicksAndSleep() puts the normal tick duration into the SysTick's reload register. - vPortSuppressTicksAndSleep() unmasks interrupts before the SysTick starts a new period based on the new value in the reload register. [This is a race condition that can go either way, but for the bug to occur, the race must play out this way.] - The pending SysTick ISR executes and increments xPendedTicks. - The SysTick expires again, finishing the second very small period, and starts a new period this time based on the full tick duration. - The SysTick ISR increments xPendedTicks (or xTickCount) even though only a tiny fraction of a tick period has elapsed since the previous tick. The bug occurs when *two* consecutive small periods of the SysTick are both counted as ticks. The root cause is a race caused by the small SysTick period. If vPortSuppressTicksAndSleep() unmasks interrupts *after* the small period expires but *before* the SysTick starts a period based on the full tick period, then two small periods are counted as ticks when only one should be counted. The end result is xTickCount advancing nearly one full tick more than time actually elapsed as measured by the SysTick. This is not the kind of time slippage normally associated with tickless idle. After this commit the code starts the SysTick and then immediately modifies the reload register to ensure the very short cycle (if any) is conducted only once. This strategy requires special consideration for the build option that configures SysTick to use a divided clock. To avoid waiting around for the SysTick to load value from the reload register, the new code temporarily configures the SysTick to use the undivided clock. The resulting timing error is typical for tickless idle. The error (commonly known as drift or slippage in kernel time) caused by this strategy is equivalent to one or two counts in ulStoppedTimerCompensation. This commit also updates comments and #define symbols related to the SysTick clock option. The SysTick can optionally be clocked by a divided version of the CPU clock (commonly divide-by-8). The new code in this commit adjusts these comments and symbols to make them clearer and more useful in configurations that use the divided clock. The fix made in this commit requires the use of these symbols, as noted in the code comments. * Fix tickless idle with alternate systick clocking Prior to this commit, in configurations using the alternate SysTick clocking, vPortSuppressTicksAndSleep() might cause xTickCount to jump ahead as much as the entire expected idle time or fall behind as much as one full tick compared to time as measured by the SysTick. SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. The SysTick has a configuration option to be clocked by an alternate clock besides the core clock. This alternate clock is MCU dependent. Scenarios Fixed --------------- The new code in this commit handles the following scenarios that were not handled correctly prior to this commit. 1. Before the sleep, vPortSuppressTicksAndSleep() stops the SysTick on zero, long after SysTick reached zero. Prior to this commit, this scenario caused xTickCount to jump ahead one full tick for the same reason documented here: https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/0c7b04bd3a745c52151abebc882eed3f811c4c81 2. After the sleep, vPortSuppressTicksAndSleep() stops the SysTick before it loads the counter from the reload register. Prior to this commit, this scenario caused xTickCount to jump ahead by the entire expected idle time (xExpectedIdleTime) because the current-count register is zero before it loads from the reload register. 3. Prior to return, vPortSuppressTicksAndSleep() attempts to start a short SysTick period when the current SysTick clock cycle has a lot of time remaining. Prior to this commit, this scenario could cause xTickCount to fall behind by as much as nearly one full tick because the short SysTick cycle never started. Note that #3 is partially fixed by https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/967acc9b200d3d4beeb289d9da9e88798074b431 even though that commit addresses a different issue. So this commit completes the partial fix. * Improve comments and name of preprocessor symbol Add a note in the code comments that SysTick requests an interrupt when decrementing from 1 to 0, so that's why stopping SysTick on zero is a special case. Readers might unknowingly assume that SysTick requests an interrupt when wrapping from 0 back to the load-register value. Reconsider new "_SETTING" suffix since "_CONFIG" suffix seems more descriptive. The code relies on *both* of these preprocessor symbols: portNVIC_SYSTICK_CLK_BIT portNVIC_SYSTICK_CLK_BIT_CONFIG **new** A meaningful suffix is really helpful to distinguish the two symbols. * Revert introduction of 2nd name for NVIC register When I added portNVIC_ICSR_REG I didn't realize there was already a portNVIC_INT_CTRL_REG, which identifies the same register. Not good to have both. Note that portNVIC_INT_CTRL_REG is defined in portmacro.h and is already used in this file (port.c). * Replicate to other Cortex M ports Also set a new fiddle factor based on tests with a CM4F. I used gcc, optimizing at -O1. Users can fine-tune as needed. Also add configSYSTICK_CLOCK_HZ to the CM0 ports to be just like the other Cortex M ports. This change allowed uniformity in the default tickless implementations across all Cortex M ports. And CM0 is likely to benefit from configSYSTICK_CLOCK_HZ, especially considering new CM0 devices with very fast CPU clock speeds. * Revert changes to IAR-CM0-portmacro.h portNVIC_INT_CTRL_REG was already defined in port.c. No need to define it in portmacro.h. * Handle edge cases with slow SysTick clock Co-authored-by: Cobus van Eeden <35851496+cobusve@users.noreply.github.com> Co-authored-by: abhidixi11 <44424462+abhidixi11@users.noreply.github.com> Co-authored-by: Joseph Julicher <jjulicher@mac.com> Co-authored-by: alfred gedeon <28123637+alfred2g@users.noreply.github.com>
2 years ago
{
/* Enable non-secure access to the FPU. */
SecureInit_EnableNSFPUAccess();
}
#endif /* configENABLE_TRUSTZONE */
/* CP10 = 11 ==> Full access to FPU i.e. both privileged and
* unprivileged code should be able to access FPU. CP11 should be
* programmed to the same value as CP10. */
*( portCPACR ) |= ( ( portCPACR_CP10_VALUE << portCPACR_CP10_POS ) |
( portCPACR_CP11_VALUE << portCPACR_CP11_POS )
);
/* ASPEN = 1 ==> Hardware should automatically preserve floating point
* context on exception entry and restore on exception return.
* LSPEN = 1 ==> Enable lazy context save of FP state. */
*( portFPCCR ) |= ( portFPCCR_ASPEN_MASK | portFPCCR_LSPEN_MASK );
}
#endif /* configENABLE_FPU */
/*-----------------------------------------------------------*/
void vPortYield( void ) /* PRIVILEGED_FUNCTION */
{
/* Set a PendSV to request a context switch. */
portNVIC_INT_CTRL_REG = portNVIC_PENDSVSET_BIT;
/* Barriers are normally not required but do ensure the code is
* completely within the specified behaviour for the architecture. */
__asm volatile ( "dsb" ::: "memory" );
__asm volatile ( "isb" );
}
/*-----------------------------------------------------------*/
void vPortEnterCritical( void ) /* PRIVILEGED_FUNCTION */
{
portDISABLE_INTERRUPTS();
ulCriticalNesting++;
/* Barriers are normally not required but do ensure the code is
* completely within the specified behaviour for the architecture. */
__asm volatile ( "dsb" ::: "memory" );
__asm volatile ( "isb" );
}
/*-----------------------------------------------------------*/
void vPortExitCritical( void ) /* PRIVILEGED_FUNCTION */
{
configASSERT( ulCriticalNesting );
ulCriticalNesting--;
if( ulCriticalNesting == 0 )
{
portENABLE_INTERRUPTS();
}
}
/*-----------------------------------------------------------*/
void SysTick_Handler( void ) /* PRIVILEGED_FUNCTION */
{
uint32_t ulPreviousMask;
ulPreviousMask = portSET_INTERRUPT_MASK_FROM_ISR();
traceISR_ENTER();
{
/* Increment the RTOS tick. */
if( xTaskIncrementTick() != pdFALSE )
{
traceISR_EXIT_TO_SCHEDULER();
/* Pend a context switch. */
portNVIC_INT_CTRL_REG = portNVIC_PENDSVSET_BIT;
}
else
{
traceISR_EXIT();
}
}
portCLEAR_INTERRUPT_MASK_FROM_ISR( ulPreviousMask );
}
/*-----------------------------------------------------------*/
void vPortSVCHandler_C( uint32_t * pulCallerStackAddress ) /* PRIVILEGED_FUNCTION portDONT_DISCARD */
{
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
#if ( ( configENABLE_MPU == 1 ) && ( configUSE_MPU_WRAPPERS_V1 == 1 ) )
#if defined( __ARMCC_VERSION )
/* Declaration when these variable are defined in code instead of being
* exported from linker scripts. */
extern uint32_t * __syscalls_flash_start__;
extern uint32_t * __syscalls_flash_end__;
#else
/* Declaration when these variable are exported from linker scripts. */
extern uint32_t __syscalls_flash_start__[];
extern uint32_t __syscalls_flash_end__[];
#endif /* defined( __ARMCC_VERSION ) */
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
#endif /* ( configENABLE_MPU == 1 ) && ( configUSE_MPU_WRAPPERS_V1 == 1 ) */
uint32_t ulPC;
#if ( configENABLE_TRUSTZONE == 1 )
uint32_t ulR0, ulR1;
extern TaskHandle_t pxCurrentTCB;
#if ( configENABLE_MPU == 1 )
uint32_t ulControl, ulIsTaskPrivileged;
#endif /* configENABLE_MPU */
#endif /* configENABLE_TRUSTZONE */
uint8_t ucSVCNumber;
/* Register are stored on the stack in the following order - R0, R1, R2, R3,
* R12, LR, PC, xPSR. */
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
ulPC = pulCallerStackAddress[ portOFFSET_TO_PC ];
ucSVCNumber = ( ( uint8_t * ) ulPC )[ -2 ];
switch( ucSVCNumber )
{
#if ( configENABLE_TRUSTZONE == 1 )
case portSVC_ALLOCATE_SECURE_CONTEXT:
/* R0 contains the stack size passed as parameter to the
* vPortAllocateSecureContext function. */
ulR0 = pulCallerStackAddress[ 0 ];
#if ( configENABLE_MPU == 1 )
Tickless idle fixes/improvement (#59) * Fix tickless idle when stopping systick on zero... ...and don't stop SysTick at all in the eAbortSleep case. Prior to this commit, if vPortSuppressTicksAndSleep() happens to stop the SysTick on zero, then after tickless idle ends, xTickCount advances one full tick more than the time that actually elapsed as measured by the SysTick. See "bug 1" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - Idle task calls vPortSuppressTicksAndSleep(xExpectedIdleTime = 2). [Doesn't have to be "2" -- could be any number.] - vPortSuppressTicksAndSleep() stops SysTick, and the current-count register happens to stop on zero. - SysTick ISR executes, setting xPendedTicks = 1 - vPortSuppressTicksAndSleep() masks interrupts and calls eTaskConfirmSleepModeStatus() which confirms the sleep operation. *** - vPortSuppressTicksAndSleep() configures SysTick for 1 full tick (xExpectedIdleTime - 1) plus the current-count register (which is 0) - One tick period elapses in sleep. - SysTick wakes CPU, ISR executes and increments xPendedTicks to 2. - vPortSuppressTicksAndSleep() calls vTaskStepTick(1), then returns. - Idle task resumes scheduler, which increments xTickCount twice (for xPendedTicks = 2) In the end, two ticks elapsed as measured by SysTick, but the code increments xTickCount three times. The root cause is that the code assumes the SysTick current-count register always contains the number of SysTick counts remaining in the current tick period. However, when the current-count register is zero, there are ulTimerCountsForOneTick counts remaining, not zero. This error is not the kind of time slippage normally associated with tickless idle. *** Note that a recent commit https://github.com/FreeRTOS/FreeRTOS-Kernel/commit/e1b98f0 results in eAbortSleep in this case, due to xPendedTicks != 0. That commit does mostly resolve this bug without specifically mentioning it, and without this commit. But that resolution allows the code in port.c not to directly address the special case of stopping SysTick on zero in any code or comments. That commit also generates additional instances of eAbortSleep, and a second purpose of this commit is to optimize how vPortSuppressTicksAndSleep() behaves for eAbortSleep, as noted below. This commit also includes an optimization to avoid stopping the SysTick when eTaskConfirmSleepModeStatus() returns eAbortSleep. This optimization belongs with this fix because the method of handling the SysTick being stopped on zero changes with this optimization. * Fix imminent tick rescheduled after tickless idle Prior to this commit, if something other than systick wakes the CPU from tickless idle, vPortSuppressTicksAndSleep() might cause xTickCount to increment once too many times. See "bug 2" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - CPU is sleeping in vPortSuppressTicksAndSleep() - Something other than the SysTick wakes the CPU. - vPortSuppressTicksAndSleep() calculates the number of SysTick counts until the next tick. The bug occurs only if this number is small. - vPortSuppressTicksAndSleep() puts this small number into the SysTick reload register, and starts SysTick. - vPortSuppressTicksAndSleep() calls vTaskStepTick() - While vTaskStepTick() executes, the SysTick expires. The ISR pends because interrupts are masked, and SysTick starts a 2nd period still based on the small number of counts in its reload register. This 2nd period is undesirable and is likely to cause the error noted below. - vPortSuppressTicksAndSleep() puts the normal tick duration into the SysTick's reload register. - vPortSuppressTicksAndSleep() unmasks interrupts before the SysTick starts a new period based on the new value in the reload register. [This is a race condition that can go either way, but for the bug to occur, the race must play out this way.] - The pending SysTick ISR executes and increments xPendedTicks. - The SysTick expires again, finishing the second very small period, and starts a new period this time based on the full tick duration. - The SysTick ISR increments xPendedTicks (or xTickCount) even though only a tiny fraction of a tick period has elapsed since the previous tick. The bug occurs when *two* consecutive small periods of the SysTick are both counted as ticks. The root cause is a race caused by the small SysTick period. If vPortSuppressTicksAndSleep() unmasks interrupts *after* the small period expires but *before* the SysTick starts a period based on the full tick period, then two small periods are counted as ticks when only one should be counted. The end result is xTickCount advancing nearly one full tick more than time actually elapsed as measured by the SysTick. This is not the kind of time slippage normally associated with tickless idle. After this commit the code starts the SysTick and then immediately modifies the reload register to ensure the very short cycle (if any) is conducted only once. This strategy requires special consideration for the build option that configures SysTick to use a divided clock. To avoid waiting around for the SysTick to load value from the reload register, the new code temporarily configures the SysTick to use the undivided clock. The resulting timing error is typical for tickless idle. The error (commonly known as drift or slippage in kernel time) caused by this strategy is equivalent to one or two counts in ulStoppedTimerCompensation. This commit also updates comments and #define symbols related to the SysTick clock option. The SysTick can optionally be clocked by a divided version of the CPU clock (commonly divide-by-8). The new code in this commit adjusts these comments and symbols to make them clearer and more useful in configurations that use the divided clock. The fix made in this commit requires the use of these symbols, as noted in the code comments. * Fix tickless idle with alternate systick clocking Prior to this commit, in configurations using the alternate SysTick clocking, vPortSuppressTicksAndSleep() might cause xTickCount to jump ahead as much as the entire expected idle time or fall behind as much as one full tick compared to time as measured by the SysTick. SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. The SysTick has a configuration option to be clocked by an alternate clock besides the core clock. This alternate clock is MCU dependent. Scenarios Fixed --------------- The new code in this commit handles the following scenarios that were not handled correctly prior to this commit. 1. Before the sleep, vPortSuppressTicksAndSleep() stops the SysTick on zero, long after SysTick reached zero. Prior to this commit, this scenario caused xTickCount to jump ahead one full tick for the same reason documented here: https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/0c7b04bd3a745c52151abebc882eed3f811c4c81 2. After the sleep, vPortSuppressTicksAndSleep() stops the SysTick before it loads the counter from the reload register. Prior to this commit, this scenario caused xTickCount to jump ahead by the entire expected idle time (xExpectedIdleTime) because the current-count register is zero before it loads from the reload register. 3. Prior to return, vPortSuppressTicksAndSleep() attempts to start a short SysTick period when the current SysTick clock cycle has a lot of time remaining. Prior to this commit, this scenario could cause xTickCount to fall behind by as much as nearly one full tick because the short SysTick cycle never started. Note that #3 is partially fixed by https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/967acc9b200d3d4beeb289d9da9e88798074b431 even though that commit addresses a different issue. So this commit completes the partial fix. * Improve comments and name of preprocessor symbol Add a note in the code comments that SysTick requests an interrupt when decrementing from 1 to 0, so that's why stopping SysTick on zero is a special case. Readers might unknowingly assume that SysTick requests an interrupt when wrapping from 0 back to the load-register value. Reconsider new "_SETTING" suffix since "_CONFIG" suffix seems more descriptive. The code relies on *both* of these preprocessor symbols: portNVIC_SYSTICK_CLK_BIT portNVIC_SYSTICK_CLK_BIT_CONFIG **new** A meaningful suffix is really helpful to distinguish the two symbols. * Revert introduction of 2nd name for NVIC register When I added portNVIC_ICSR_REG I didn't realize there was already a portNVIC_INT_CTRL_REG, which identifies the same register. Not good to have both. Note that portNVIC_INT_CTRL_REG is defined in portmacro.h and is already used in this file (port.c). * Replicate to other Cortex M ports Also set a new fiddle factor based on tests with a CM4F. I used gcc, optimizing at -O1. Users can fine-tune as needed. Also add configSYSTICK_CLOCK_HZ to the CM0 ports to be just like the other Cortex M ports. This change allowed uniformity in the default tickless implementations across all Cortex M ports. And CM0 is likely to benefit from configSYSTICK_CLOCK_HZ, especially considering new CM0 devices with very fast CPU clock speeds. * Revert changes to IAR-CM0-portmacro.h portNVIC_INT_CTRL_REG was already defined in port.c. No need to define it in portmacro.h. * Handle edge cases with slow SysTick clock Co-authored-by: Cobus van Eeden <35851496+cobusve@users.noreply.github.com> Co-authored-by: abhidixi11 <44424462+abhidixi11@users.noreply.github.com> Co-authored-by: Joseph Julicher <jjulicher@mac.com> Co-authored-by: alfred gedeon <28123637+alfred2g@users.noreply.github.com>
2 years ago
{
/* Read the CONTROL register value. */
__asm volatile ( "mrs %0, control" : "=r" ( ulControl ) );
Tickless idle fixes/improvement (#59) * Fix tickless idle when stopping systick on zero... ...and don't stop SysTick at all in the eAbortSleep case. Prior to this commit, if vPortSuppressTicksAndSleep() happens to stop the SysTick on zero, then after tickless idle ends, xTickCount advances one full tick more than the time that actually elapsed as measured by the SysTick. See "bug 1" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - Idle task calls vPortSuppressTicksAndSleep(xExpectedIdleTime = 2). [Doesn't have to be "2" -- could be any number.] - vPortSuppressTicksAndSleep() stops SysTick, and the current-count register happens to stop on zero. - SysTick ISR executes, setting xPendedTicks = 1 - vPortSuppressTicksAndSleep() masks interrupts and calls eTaskConfirmSleepModeStatus() which confirms the sleep operation. *** - vPortSuppressTicksAndSleep() configures SysTick for 1 full tick (xExpectedIdleTime - 1) plus the current-count register (which is 0) - One tick period elapses in sleep. - SysTick wakes CPU, ISR executes and increments xPendedTicks to 2. - vPortSuppressTicksAndSleep() calls vTaskStepTick(1), then returns. - Idle task resumes scheduler, which increments xTickCount twice (for xPendedTicks = 2) In the end, two ticks elapsed as measured by SysTick, but the code increments xTickCount three times. The root cause is that the code assumes the SysTick current-count register always contains the number of SysTick counts remaining in the current tick period. However, when the current-count register is zero, there are ulTimerCountsForOneTick counts remaining, not zero. This error is not the kind of time slippage normally associated with tickless idle. *** Note that a recent commit https://github.com/FreeRTOS/FreeRTOS-Kernel/commit/e1b98f0 results in eAbortSleep in this case, due to xPendedTicks != 0. That commit does mostly resolve this bug without specifically mentioning it, and without this commit. But that resolution allows the code in port.c not to directly address the special case of stopping SysTick on zero in any code or comments. That commit also generates additional instances of eAbortSleep, and a second purpose of this commit is to optimize how vPortSuppressTicksAndSleep() behaves for eAbortSleep, as noted below. This commit also includes an optimization to avoid stopping the SysTick when eTaskConfirmSleepModeStatus() returns eAbortSleep. This optimization belongs with this fix because the method of handling the SysTick being stopped on zero changes with this optimization. * Fix imminent tick rescheduled after tickless idle Prior to this commit, if something other than systick wakes the CPU from tickless idle, vPortSuppressTicksAndSleep() might cause xTickCount to increment once too many times. See "bug 2" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - CPU is sleeping in vPortSuppressTicksAndSleep() - Something other than the SysTick wakes the CPU. - vPortSuppressTicksAndSleep() calculates the number of SysTick counts until the next tick. The bug occurs only if this number is small. - vPortSuppressTicksAndSleep() puts this small number into the SysTick reload register, and starts SysTick. - vPortSuppressTicksAndSleep() calls vTaskStepTick() - While vTaskStepTick() executes, the SysTick expires. The ISR pends because interrupts are masked, and SysTick starts a 2nd period still based on the small number of counts in its reload register. This 2nd period is undesirable and is likely to cause the error noted below. - vPortSuppressTicksAndSleep() puts the normal tick duration into the SysTick's reload register. - vPortSuppressTicksAndSleep() unmasks interrupts before the SysTick starts a new period based on the new value in the reload register. [This is a race condition that can go either way, but for the bug to occur, the race must play out this way.] - The pending SysTick ISR executes and increments xPendedTicks. - The SysTick expires again, finishing the second very small period, and starts a new period this time based on the full tick duration. - The SysTick ISR increments xPendedTicks (or xTickCount) even though only a tiny fraction of a tick period has elapsed since the previous tick. The bug occurs when *two* consecutive small periods of the SysTick are both counted as ticks. The root cause is a race caused by the small SysTick period. If vPortSuppressTicksAndSleep() unmasks interrupts *after* the small period expires but *before* the SysTick starts a period based on the full tick period, then two small periods are counted as ticks when only one should be counted. The end result is xTickCount advancing nearly one full tick more than time actually elapsed as measured by the SysTick. This is not the kind of time slippage normally associated with tickless idle. After this commit the code starts the SysTick and then immediately modifies the reload register to ensure the very short cycle (if any) is conducted only once. This strategy requires special consideration for the build option that configures SysTick to use a divided clock. To avoid waiting around for the SysTick to load value from the reload register, the new code temporarily configures the SysTick to use the undivided clock. The resulting timing error is typical for tickless idle. The error (commonly known as drift or slippage in kernel time) caused by this strategy is equivalent to one or two counts in ulStoppedTimerCompensation. This commit also updates comments and #define symbols related to the SysTick clock option. The SysTick can optionally be clocked by a divided version of the CPU clock (commonly divide-by-8). The new code in this commit adjusts these comments and symbols to make them clearer and more useful in configurations that use the divided clock. The fix made in this commit requires the use of these symbols, as noted in the code comments. * Fix tickless idle with alternate systick clocking Prior to this commit, in configurations using the alternate SysTick clocking, vPortSuppressTicksAndSleep() might cause xTickCount to jump ahead as much as the entire expected idle time or fall behind as much as one full tick compared to time as measured by the SysTick. SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. The SysTick has a configuration option to be clocked by an alternate clock besides the core clock. This alternate clock is MCU dependent. Scenarios Fixed --------------- The new code in this commit handles the following scenarios that were not handled correctly prior to this commit. 1. Before the sleep, vPortSuppressTicksAndSleep() stops the SysTick on zero, long after SysTick reached zero. Prior to this commit, this scenario caused xTickCount to jump ahead one full tick for the same reason documented here: https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/0c7b04bd3a745c52151abebc882eed3f811c4c81 2. After the sleep, vPortSuppressTicksAndSleep() stops the SysTick before it loads the counter from the reload register. Prior to this commit, this scenario caused xTickCount to jump ahead by the entire expected idle time (xExpectedIdleTime) because the current-count register is zero before it loads from the reload register. 3. Prior to return, vPortSuppressTicksAndSleep() attempts to start a short SysTick period when the current SysTick clock cycle has a lot of time remaining. Prior to this commit, this scenario could cause xTickCount to fall behind by as much as nearly one full tick because the short SysTick cycle never started. Note that #3 is partially fixed by https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/967acc9b200d3d4beeb289d9da9e88798074b431 even though that commit addresses a different issue. So this commit completes the partial fix. * Improve comments and name of preprocessor symbol Add a note in the code comments that SysTick requests an interrupt when decrementing from 1 to 0, so that's why stopping SysTick on zero is a special case. Readers might unknowingly assume that SysTick requests an interrupt when wrapping from 0 back to the load-register value. Reconsider new "_SETTING" suffix since "_CONFIG" suffix seems more descriptive. The code relies on *both* of these preprocessor symbols: portNVIC_SYSTICK_CLK_BIT portNVIC_SYSTICK_CLK_BIT_CONFIG **new** A meaningful suffix is really helpful to distinguish the two symbols. * Revert introduction of 2nd name for NVIC register When I added portNVIC_ICSR_REG I didn't realize there was already a portNVIC_INT_CTRL_REG, which identifies the same register. Not good to have both. Note that portNVIC_INT_CTRL_REG is defined in portmacro.h and is already used in this file (port.c). * Replicate to other Cortex M ports Also set a new fiddle factor based on tests with a CM4F. I used gcc, optimizing at -O1. Users can fine-tune as needed. Also add configSYSTICK_CLOCK_HZ to the CM0 ports to be just like the other Cortex M ports. This change allowed uniformity in the default tickless implementations across all Cortex M ports. And CM0 is likely to benefit from configSYSTICK_CLOCK_HZ, especially considering new CM0 devices with very fast CPU clock speeds. * Revert changes to IAR-CM0-portmacro.h portNVIC_INT_CTRL_REG was already defined in port.c. No need to define it in portmacro.h. * Handle edge cases with slow SysTick clock Co-authored-by: Cobus van Eeden <35851496+cobusve@users.noreply.github.com> Co-authored-by: abhidixi11 <44424462+abhidixi11@users.noreply.github.com> Co-authored-by: Joseph Julicher <jjulicher@mac.com> Co-authored-by: alfred gedeon <28123637+alfred2g@users.noreply.github.com>
2 years ago
/* The task that raised the SVC is privileged if Bit[0]
* in the CONTROL register is 0. */
ulIsTaskPrivileged = ( ( ulControl & portCONTROL_PRIVILEGED_MASK ) == 0 );
Tickless idle fixes/improvement (#59) * Fix tickless idle when stopping systick on zero... ...and don't stop SysTick at all in the eAbortSleep case. Prior to this commit, if vPortSuppressTicksAndSleep() happens to stop the SysTick on zero, then after tickless idle ends, xTickCount advances one full tick more than the time that actually elapsed as measured by the SysTick. See "bug 1" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - Idle task calls vPortSuppressTicksAndSleep(xExpectedIdleTime = 2). [Doesn't have to be "2" -- could be any number.] - vPortSuppressTicksAndSleep() stops SysTick, and the current-count register happens to stop on zero. - SysTick ISR executes, setting xPendedTicks = 1 - vPortSuppressTicksAndSleep() masks interrupts and calls eTaskConfirmSleepModeStatus() which confirms the sleep operation. *** - vPortSuppressTicksAndSleep() configures SysTick for 1 full tick (xExpectedIdleTime - 1) plus the current-count register (which is 0) - One tick period elapses in sleep. - SysTick wakes CPU, ISR executes and increments xPendedTicks to 2. - vPortSuppressTicksAndSleep() calls vTaskStepTick(1), then returns. - Idle task resumes scheduler, which increments xTickCount twice (for xPendedTicks = 2) In the end, two ticks elapsed as measured by SysTick, but the code increments xTickCount three times. The root cause is that the code assumes the SysTick current-count register always contains the number of SysTick counts remaining in the current tick period. However, when the current-count register is zero, there are ulTimerCountsForOneTick counts remaining, not zero. This error is not the kind of time slippage normally associated with tickless idle. *** Note that a recent commit https://github.com/FreeRTOS/FreeRTOS-Kernel/commit/e1b98f0 results in eAbortSleep in this case, due to xPendedTicks != 0. That commit does mostly resolve this bug without specifically mentioning it, and without this commit. But that resolution allows the code in port.c not to directly address the special case of stopping SysTick on zero in any code or comments. That commit also generates additional instances of eAbortSleep, and a second purpose of this commit is to optimize how vPortSuppressTicksAndSleep() behaves for eAbortSleep, as noted below. This commit also includes an optimization to avoid stopping the SysTick when eTaskConfirmSleepModeStatus() returns eAbortSleep. This optimization belongs with this fix because the method of handling the SysTick being stopped on zero changes with this optimization. * Fix imminent tick rescheduled after tickless idle Prior to this commit, if something other than systick wakes the CPU from tickless idle, vPortSuppressTicksAndSleep() might cause xTickCount to increment once too many times. See "bug 2" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - CPU is sleeping in vPortSuppressTicksAndSleep() - Something other than the SysTick wakes the CPU. - vPortSuppressTicksAndSleep() calculates the number of SysTick counts until the next tick. The bug occurs only if this number is small. - vPortSuppressTicksAndSleep() puts this small number into the SysTick reload register, and starts SysTick. - vPortSuppressTicksAndSleep() calls vTaskStepTick() - While vTaskStepTick() executes, the SysTick expires. The ISR pends because interrupts are masked, and SysTick starts a 2nd period still based on the small number of counts in its reload register. This 2nd period is undesirable and is likely to cause the error noted below. - vPortSuppressTicksAndSleep() puts the normal tick duration into the SysTick's reload register. - vPortSuppressTicksAndSleep() unmasks interrupts before the SysTick starts a new period based on the new value in the reload register. [This is a race condition that can go either way, but for the bug to occur, the race must play out this way.] - The pending SysTick ISR executes and increments xPendedTicks. - The SysTick expires again, finishing the second very small period, and starts a new period this time based on the full tick duration. - The SysTick ISR increments xPendedTicks (or xTickCount) even though only a tiny fraction of a tick period has elapsed since the previous tick. The bug occurs when *two* consecutive small periods of the SysTick are both counted as ticks. The root cause is a race caused by the small SysTick period. If vPortSuppressTicksAndSleep() unmasks interrupts *after* the small period expires but *before* the SysTick starts a period based on the full tick period, then two small periods are counted as ticks when only one should be counted. The end result is xTickCount advancing nearly one full tick more than time actually elapsed as measured by the SysTick. This is not the kind of time slippage normally associated with tickless idle. After this commit the code starts the SysTick and then immediately modifies the reload register to ensure the very short cycle (if any) is conducted only once. This strategy requires special consideration for the build option that configures SysTick to use a divided clock. To avoid waiting around for the SysTick to load value from the reload register, the new code temporarily configures the SysTick to use the undivided clock. The resulting timing error is typical for tickless idle. The error (commonly known as drift or slippage in kernel time) caused by this strategy is equivalent to one or two counts in ulStoppedTimerCompensation. This commit also updates comments and #define symbols related to the SysTick clock option. The SysTick can optionally be clocked by a divided version of the CPU clock (commonly divide-by-8). The new code in this commit adjusts these comments and symbols to make them clearer and more useful in configurations that use the divided clock. The fix made in this commit requires the use of these symbols, as noted in the code comments. * Fix tickless idle with alternate systick clocking Prior to this commit, in configurations using the alternate SysTick clocking, vPortSuppressTicksAndSleep() might cause xTickCount to jump ahead as much as the entire expected idle time or fall behind as much as one full tick compared to time as measured by the SysTick. SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. The SysTick has a configuration option to be clocked by an alternate clock besides the core clock. This alternate clock is MCU dependent. Scenarios Fixed --------------- The new code in this commit handles the following scenarios that were not handled correctly prior to this commit. 1. Before the sleep, vPortSuppressTicksAndSleep() stops the SysTick on zero, long after SysTick reached zero. Prior to this commit, this scenario caused xTickCount to jump ahead one full tick for the same reason documented here: https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/0c7b04bd3a745c52151abebc882eed3f811c4c81 2. After the sleep, vPortSuppressTicksAndSleep() stops the SysTick before it loads the counter from the reload register. Prior to this commit, this scenario caused xTickCount to jump ahead by the entire expected idle time (xExpectedIdleTime) because the current-count register is zero before it loads from the reload register. 3. Prior to return, vPortSuppressTicksAndSleep() attempts to start a short SysTick period when the current SysTick clock cycle has a lot of time remaining. Prior to this commit, this scenario could cause xTickCount to fall behind by as much as nearly one full tick because the short SysTick cycle never started. Note that #3 is partially fixed by https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/967acc9b200d3d4beeb289d9da9e88798074b431 even though that commit addresses a different issue. So this commit completes the partial fix. * Improve comments and name of preprocessor symbol Add a note in the code comments that SysTick requests an interrupt when decrementing from 1 to 0, so that's why stopping SysTick on zero is a special case. Readers might unknowingly assume that SysTick requests an interrupt when wrapping from 0 back to the load-register value. Reconsider new "_SETTING" suffix since "_CONFIG" suffix seems more descriptive. The code relies on *both* of these preprocessor symbols: portNVIC_SYSTICK_CLK_BIT portNVIC_SYSTICK_CLK_BIT_CONFIG **new** A meaningful suffix is really helpful to distinguish the two symbols. * Revert introduction of 2nd name for NVIC register When I added portNVIC_ICSR_REG I didn't realize there was already a portNVIC_INT_CTRL_REG, which identifies the same register. Not good to have both. Note that portNVIC_INT_CTRL_REG is defined in portmacro.h and is already used in this file (port.c). * Replicate to other Cortex M ports Also set a new fiddle factor based on tests with a CM4F. I used gcc, optimizing at -O1. Users can fine-tune as needed. Also add configSYSTICK_CLOCK_HZ to the CM0 ports to be just like the other Cortex M ports. This change allowed uniformity in the default tickless implementations across all Cortex M ports. And CM0 is likely to benefit from configSYSTICK_CLOCK_HZ, especially considering new CM0 devices with very fast CPU clock speeds. * Revert changes to IAR-CM0-portmacro.h portNVIC_INT_CTRL_REG was already defined in port.c. No need to define it in portmacro.h. * Handle edge cases with slow SysTick clock Co-authored-by: Cobus van Eeden <35851496+cobusve@users.noreply.github.com> Co-authored-by: abhidixi11 <44424462+abhidixi11@users.noreply.github.com> Co-authored-by: Joseph Julicher <jjulicher@mac.com> Co-authored-by: alfred gedeon <28123637+alfred2g@users.noreply.github.com>
2 years ago
/* Allocate and load a context for the secure task. */
xSecureContext = SecureContext_AllocateContext( ulR0, ulIsTaskPrivileged, pxCurrentTCB );
}
#else /* if ( configENABLE_MPU == 1 ) */
Tickless idle fixes/improvement (#59) * Fix tickless idle when stopping systick on zero... ...and don't stop SysTick at all in the eAbortSleep case. Prior to this commit, if vPortSuppressTicksAndSleep() happens to stop the SysTick on zero, then after tickless idle ends, xTickCount advances one full tick more than the time that actually elapsed as measured by the SysTick. See "bug 1" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - Idle task calls vPortSuppressTicksAndSleep(xExpectedIdleTime = 2). [Doesn't have to be "2" -- could be any number.] - vPortSuppressTicksAndSleep() stops SysTick, and the current-count register happens to stop on zero. - SysTick ISR executes, setting xPendedTicks = 1 - vPortSuppressTicksAndSleep() masks interrupts and calls eTaskConfirmSleepModeStatus() which confirms the sleep operation. *** - vPortSuppressTicksAndSleep() configures SysTick for 1 full tick (xExpectedIdleTime - 1) plus the current-count register (which is 0) - One tick period elapses in sleep. - SysTick wakes CPU, ISR executes and increments xPendedTicks to 2. - vPortSuppressTicksAndSleep() calls vTaskStepTick(1), then returns. - Idle task resumes scheduler, which increments xTickCount twice (for xPendedTicks = 2) In the end, two ticks elapsed as measured by SysTick, but the code increments xTickCount three times. The root cause is that the code assumes the SysTick current-count register always contains the number of SysTick counts remaining in the current tick period. However, when the current-count register is zero, there are ulTimerCountsForOneTick counts remaining, not zero. This error is not the kind of time slippage normally associated with tickless idle. *** Note that a recent commit https://github.com/FreeRTOS/FreeRTOS-Kernel/commit/e1b98f0 results in eAbortSleep in this case, due to xPendedTicks != 0. That commit does mostly resolve this bug without specifically mentioning it, and without this commit. But that resolution allows the code in port.c not to directly address the special case of stopping SysTick on zero in any code or comments. That commit also generates additional instances of eAbortSleep, and a second purpose of this commit is to optimize how vPortSuppressTicksAndSleep() behaves for eAbortSleep, as noted below. This commit also includes an optimization to avoid stopping the SysTick when eTaskConfirmSleepModeStatus() returns eAbortSleep. This optimization belongs with this fix because the method of handling the SysTick being stopped on zero changes with this optimization. * Fix imminent tick rescheduled after tickless idle Prior to this commit, if something other than systick wakes the CPU from tickless idle, vPortSuppressTicksAndSleep() might cause xTickCount to increment once too many times. See "bug 2" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - CPU is sleeping in vPortSuppressTicksAndSleep() - Something other than the SysTick wakes the CPU. - vPortSuppressTicksAndSleep() calculates the number of SysTick counts until the next tick. The bug occurs only if this number is small. - vPortSuppressTicksAndSleep() puts this small number into the SysTick reload register, and starts SysTick. - vPortSuppressTicksAndSleep() calls vTaskStepTick() - While vTaskStepTick() executes, the SysTick expires. The ISR pends because interrupts are masked, and SysTick starts a 2nd period still based on the small number of counts in its reload register. This 2nd period is undesirable and is likely to cause the error noted below. - vPortSuppressTicksAndSleep() puts the normal tick duration into the SysTick's reload register. - vPortSuppressTicksAndSleep() unmasks interrupts before the SysTick starts a new period based on the new value in the reload register. [This is a race condition that can go either way, but for the bug to occur, the race must play out this way.] - The pending SysTick ISR executes and increments xPendedTicks. - The SysTick expires again, finishing the second very small period, and starts a new period this time based on the full tick duration. - The SysTick ISR increments xPendedTicks (or xTickCount) even though only a tiny fraction of a tick period has elapsed since the previous tick. The bug occurs when *two* consecutive small periods of the SysTick are both counted as ticks. The root cause is a race caused by the small SysTick period. If vPortSuppressTicksAndSleep() unmasks interrupts *after* the small period expires but *before* the SysTick starts a period based on the full tick period, then two small periods are counted as ticks when only one should be counted. The end result is xTickCount advancing nearly one full tick more than time actually elapsed as measured by the SysTick. This is not the kind of time slippage normally associated with tickless idle. After this commit the code starts the SysTick and then immediately modifies the reload register to ensure the very short cycle (if any) is conducted only once. This strategy requires special consideration for the build option that configures SysTick to use a divided clock. To avoid waiting around for the SysTick to load value from the reload register, the new code temporarily configures the SysTick to use the undivided clock. The resulting timing error is typical for tickless idle. The error (commonly known as drift or slippage in kernel time) caused by this strategy is equivalent to one or two counts in ulStoppedTimerCompensation. This commit also updates comments and #define symbols related to the SysTick clock option. The SysTick can optionally be clocked by a divided version of the CPU clock (commonly divide-by-8). The new code in this commit adjusts these comments and symbols to make them clearer and more useful in configurations that use the divided clock. The fix made in this commit requires the use of these symbols, as noted in the code comments. * Fix tickless idle with alternate systick clocking Prior to this commit, in configurations using the alternate SysTick clocking, vPortSuppressTicksAndSleep() might cause xTickCount to jump ahead as much as the entire expected idle time or fall behind as much as one full tick compared to time as measured by the SysTick. SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. The SysTick has a configuration option to be clocked by an alternate clock besides the core clock. This alternate clock is MCU dependent. Scenarios Fixed --------------- The new code in this commit handles the following scenarios that were not handled correctly prior to this commit. 1. Before the sleep, vPortSuppressTicksAndSleep() stops the SysTick on zero, long after SysTick reached zero. Prior to this commit, this scenario caused xTickCount to jump ahead one full tick for the same reason documented here: https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/0c7b04bd3a745c52151abebc882eed3f811c4c81 2. After the sleep, vPortSuppressTicksAndSleep() stops the SysTick before it loads the counter from the reload register. Prior to this commit, this scenario caused xTickCount to jump ahead by the entire expected idle time (xExpectedIdleTime) because the current-count register is zero before it loads from the reload register. 3. Prior to return, vPortSuppressTicksAndSleep() attempts to start a short SysTick period when the current SysTick clock cycle has a lot of time remaining. Prior to this commit, this scenario could cause xTickCount to fall behind by as much as nearly one full tick because the short SysTick cycle never started. Note that #3 is partially fixed by https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/967acc9b200d3d4beeb289d9da9e88798074b431 even though that commit addresses a different issue. So this commit completes the partial fix. * Improve comments and name of preprocessor symbol Add a note in the code comments that SysTick requests an interrupt when decrementing from 1 to 0, so that's why stopping SysTick on zero is a special case. Readers might unknowingly assume that SysTick requests an interrupt when wrapping from 0 back to the load-register value. Reconsider new "_SETTING" suffix since "_CONFIG" suffix seems more descriptive. The code relies on *both* of these preprocessor symbols: portNVIC_SYSTICK_CLK_BIT portNVIC_SYSTICK_CLK_BIT_CONFIG **new** A meaningful suffix is really helpful to distinguish the two symbols. * Revert introduction of 2nd name for NVIC register When I added portNVIC_ICSR_REG I didn't realize there was already a portNVIC_INT_CTRL_REG, which identifies the same register. Not good to have both. Note that portNVIC_INT_CTRL_REG is defined in portmacro.h and is already used in this file (port.c). * Replicate to other Cortex M ports Also set a new fiddle factor based on tests with a CM4F. I used gcc, optimizing at -O1. Users can fine-tune as needed. Also add configSYSTICK_CLOCK_HZ to the CM0 ports to be just like the other Cortex M ports. This change allowed uniformity in the default tickless implementations across all Cortex M ports. And CM0 is likely to benefit from configSYSTICK_CLOCK_HZ, especially considering new CM0 devices with very fast CPU clock speeds. * Revert changes to IAR-CM0-portmacro.h portNVIC_INT_CTRL_REG was already defined in port.c. No need to define it in portmacro.h. * Handle edge cases with slow SysTick clock Co-authored-by: Cobus van Eeden <35851496+cobusve@users.noreply.github.com> Co-authored-by: abhidixi11 <44424462+abhidixi11@users.noreply.github.com> Co-authored-by: Joseph Julicher <jjulicher@mac.com> Co-authored-by: alfred gedeon <28123637+alfred2g@users.noreply.github.com>
2 years ago
{
/* Allocate and load a context for the secure task. */
xSecureContext = SecureContext_AllocateContext( ulR0, pxCurrentTCB );
}
#endif /* configENABLE_MPU */
configASSERT( xSecureContext != securecontextINVALID_CONTEXT_ID );
SecureContext_LoadContext( xSecureContext, pxCurrentTCB );
break;
case portSVC_FREE_SECURE_CONTEXT:
Tickless idle fixes/improvement (#59) * Fix tickless idle when stopping systick on zero... ...and don't stop SysTick at all in the eAbortSleep case. Prior to this commit, if vPortSuppressTicksAndSleep() happens to stop the SysTick on zero, then after tickless idle ends, xTickCount advances one full tick more than the time that actually elapsed as measured by the SysTick. See "bug 1" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - Idle task calls vPortSuppressTicksAndSleep(xExpectedIdleTime = 2). [Doesn't have to be "2" -- could be any number.] - vPortSuppressTicksAndSleep() stops SysTick, and the current-count register happens to stop on zero. - SysTick ISR executes, setting xPendedTicks = 1 - vPortSuppressTicksAndSleep() masks interrupts and calls eTaskConfirmSleepModeStatus() which confirms the sleep operation. *** - vPortSuppressTicksAndSleep() configures SysTick for 1 full tick (xExpectedIdleTime - 1) plus the current-count register (which is 0) - One tick period elapses in sleep. - SysTick wakes CPU, ISR executes and increments xPendedTicks to 2. - vPortSuppressTicksAndSleep() calls vTaskStepTick(1), then returns. - Idle task resumes scheduler, which increments xTickCount twice (for xPendedTicks = 2) In the end, two ticks elapsed as measured by SysTick, but the code increments xTickCount three times. The root cause is that the code assumes the SysTick current-count register always contains the number of SysTick counts remaining in the current tick period. However, when the current-count register is zero, there are ulTimerCountsForOneTick counts remaining, not zero. This error is not the kind of time slippage normally associated with tickless idle. *** Note that a recent commit https://github.com/FreeRTOS/FreeRTOS-Kernel/commit/e1b98f0 results in eAbortSleep in this case, due to xPendedTicks != 0. That commit does mostly resolve this bug without specifically mentioning it, and without this commit. But that resolution allows the code in port.c not to directly address the special case of stopping SysTick on zero in any code or comments. That commit also generates additional instances of eAbortSleep, and a second purpose of this commit is to optimize how vPortSuppressTicksAndSleep() behaves for eAbortSleep, as noted below. This commit also includes an optimization to avoid stopping the SysTick when eTaskConfirmSleepModeStatus() returns eAbortSleep. This optimization belongs with this fix because the method of handling the SysTick being stopped on zero changes with this optimization. * Fix imminent tick rescheduled after tickless idle Prior to this commit, if something other than systick wakes the CPU from tickless idle, vPortSuppressTicksAndSleep() might cause xTickCount to increment once too many times. See "bug 2" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - CPU is sleeping in vPortSuppressTicksAndSleep() - Something other than the SysTick wakes the CPU. - vPortSuppressTicksAndSleep() calculates the number of SysTick counts until the next tick. The bug occurs only if this number is small. - vPortSuppressTicksAndSleep() puts this small number into the SysTick reload register, and starts SysTick. - vPortSuppressTicksAndSleep() calls vTaskStepTick() - While vTaskStepTick() executes, the SysTick expires. The ISR pends because interrupts are masked, and SysTick starts a 2nd period still based on the small number of counts in its reload register. This 2nd period is undesirable and is likely to cause the error noted below. - vPortSuppressTicksAndSleep() puts the normal tick duration into the SysTick's reload register. - vPortSuppressTicksAndSleep() unmasks interrupts before the SysTick starts a new period based on the new value in the reload register. [This is a race condition that can go either way, but for the bug to occur, the race must play out this way.] - The pending SysTick ISR executes and increments xPendedTicks. - The SysTick expires again, finishing the second very small period, and starts a new period this time based on the full tick duration. - The SysTick ISR increments xPendedTicks (or xTickCount) even though only a tiny fraction of a tick period has elapsed since the previous tick. The bug occurs when *two* consecutive small periods of the SysTick are both counted as ticks. The root cause is a race caused by the small SysTick period. If vPortSuppressTicksAndSleep() unmasks interrupts *after* the small period expires but *before* the SysTick starts a period based on the full tick period, then two small periods are counted as ticks when only one should be counted. The end result is xTickCount advancing nearly one full tick more than time actually elapsed as measured by the SysTick. This is not the kind of time slippage normally associated with tickless idle. After this commit the code starts the SysTick and then immediately modifies the reload register to ensure the very short cycle (if any) is conducted only once. This strategy requires special consideration for the build option that configures SysTick to use a divided clock. To avoid waiting around for the SysTick to load value from the reload register, the new code temporarily configures the SysTick to use the undivided clock. The resulting timing error is typical for tickless idle. The error (commonly known as drift or slippage in kernel time) caused by this strategy is equivalent to one or two counts in ulStoppedTimerCompensation. This commit also updates comments and #define symbols related to the SysTick clock option. The SysTick can optionally be clocked by a divided version of the CPU clock (commonly divide-by-8). The new code in this commit adjusts these comments and symbols to make them clearer and more useful in configurations that use the divided clock. The fix made in this commit requires the use of these symbols, as noted in the code comments. * Fix tickless idle with alternate systick clocking Prior to this commit, in configurations using the alternate SysTick clocking, vPortSuppressTicksAndSleep() might cause xTickCount to jump ahead as much as the entire expected idle time or fall behind as much as one full tick compared to time as measured by the SysTick. SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. The SysTick has a configuration option to be clocked by an alternate clock besides the core clock. This alternate clock is MCU dependent. Scenarios Fixed --------------- The new code in this commit handles the following scenarios that were not handled correctly prior to this commit. 1. Before the sleep, vPortSuppressTicksAndSleep() stops the SysTick on zero, long after SysTick reached zero. Prior to this commit, this scenario caused xTickCount to jump ahead one full tick for the same reason documented here: https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/0c7b04bd3a745c52151abebc882eed3f811c4c81 2. After the sleep, vPortSuppressTicksAndSleep() stops the SysTick before it loads the counter from the reload register. Prior to this commit, this scenario caused xTickCount to jump ahead by the entire expected idle time (xExpectedIdleTime) because the current-count register is zero before it loads from the reload register. 3. Prior to return, vPortSuppressTicksAndSleep() attempts to start a short SysTick period when the current SysTick clock cycle has a lot of time remaining. Prior to this commit, this scenario could cause xTickCount to fall behind by as much as nearly one full tick because the short SysTick cycle never started. Note that #3 is partially fixed by https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/967acc9b200d3d4beeb289d9da9e88798074b431 even though that commit addresses a different issue. So this commit completes the partial fix. * Improve comments and name of preprocessor symbol Add a note in the code comments that SysTick requests an interrupt when decrementing from 1 to 0, so that's why stopping SysTick on zero is a special case. Readers might unknowingly assume that SysTick requests an interrupt when wrapping from 0 back to the load-register value. Reconsider new "_SETTING" suffix since "_CONFIG" suffix seems more descriptive. The code relies on *both* of these preprocessor symbols: portNVIC_SYSTICK_CLK_BIT portNVIC_SYSTICK_CLK_BIT_CONFIG **new** A meaningful suffix is really helpful to distinguish the two symbols. * Revert introduction of 2nd name for NVIC register When I added portNVIC_ICSR_REG I didn't realize there was already a portNVIC_INT_CTRL_REG, which identifies the same register. Not good to have both. Note that portNVIC_INT_CTRL_REG is defined in portmacro.h and is already used in this file (port.c). * Replicate to other Cortex M ports Also set a new fiddle factor based on tests with a CM4F. I used gcc, optimizing at -O1. Users can fine-tune as needed. Also add configSYSTICK_CLOCK_HZ to the CM0 ports to be just like the other Cortex M ports. This change allowed uniformity in the default tickless implementations across all Cortex M ports. And CM0 is likely to benefit from configSYSTICK_CLOCK_HZ, especially considering new CM0 devices with very fast CPU clock speeds. * Revert changes to IAR-CM0-portmacro.h portNVIC_INT_CTRL_REG was already defined in port.c. No need to define it in portmacro.h. * Handle edge cases with slow SysTick clock Co-authored-by: Cobus van Eeden <35851496+cobusve@users.noreply.github.com> Co-authored-by: abhidixi11 <44424462+abhidixi11@users.noreply.github.com> Co-authored-by: Joseph Julicher <jjulicher@mac.com> Co-authored-by: alfred gedeon <28123637+alfred2g@users.noreply.github.com>
2 years ago
/* R0 contains TCB being freed and R1 contains the secure
* context handle to be freed. */
ulR0 = pulCallerStackAddress[ 0 ];
ulR1 = pulCallerStackAddress[ 1 ];
/* Free the secure context. */
SecureContext_FreeContext( ( SecureContextHandle_t ) ulR1, ( void * ) ulR0 );
break;
#endif /* configENABLE_TRUSTZONE */
case portSVC_START_SCHEDULER:
#if ( configENABLE_TRUSTZONE == 1 )
Tickless idle fixes/improvement (#59) * Fix tickless idle when stopping systick on zero... ...and don't stop SysTick at all in the eAbortSleep case. Prior to this commit, if vPortSuppressTicksAndSleep() happens to stop the SysTick on zero, then after tickless idle ends, xTickCount advances one full tick more than the time that actually elapsed as measured by the SysTick. See "bug 1" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - Idle task calls vPortSuppressTicksAndSleep(xExpectedIdleTime = 2). [Doesn't have to be "2" -- could be any number.] - vPortSuppressTicksAndSleep() stops SysTick, and the current-count register happens to stop on zero. - SysTick ISR executes, setting xPendedTicks = 1 - vPortSuppressTicksAndSleep() masks interrupts and calls eTaskConfirmSleepModeStatus() which confirms the sleep operation. *** - vPortSuppressTicksAndSleep() configures SysTick for 1 full tick (xExpectedIdleTime - 1) plus the current-count register (which is 0) - One tick period elapses in sleep. - SysTick wakes CPU, ISR executes and increments xPendedTicks to 2. - vPortSuppressTicksAndSleep() calls vTaskStepTick(1), then returns. - Idle task resumes scheduler, which increments xTickCount twice (for xPendedTicks = 2) In the end, two ticks elapsed as measured by SysTick, but the code increments xTickCount three times. The root cause is that the code assumes the SysTick current-count register always contains the number of SysTick counts remaining in the current tick period. However, when the current-count register is zero, there are ulTimerCountsForOneTick counts remaining, not zero. This error is not the kind of time slippage normally associated with tickless idle. *** Note that a recent commit https://github.com/FreeRTOS/FreeRTOS-Kernel/commit/e1b98f0 results in eAbortSleep in this case, due to xPendedTicks != 0. That commit does mostly resolve this bug without specifically mentioning it, and without this commit. But that resolution allows the code in port.c not to directly address the special case of stopping SysTick on zero in any code or comments. That commit also generates additional instances of eAbortSleep, and a second purpose of this commit is to optimize how vPortSuppressTicksAndSleep() behaves for eAbortSleep, as noted below. This commit also includes an optimization to avoid stopping the SysTick when eTaskConfirmSleepModeStatus() returns eAbortSleep. This optimization belongs with this fix because the method of handling the SysTick being stopped on zero changes with this optimization. * Fix imminent tick rescheduled after tickless idle Prior to this commit, if something other than systick wakes the CPU from tickless idle, vPortSuppressTicksAndSleep() might cause xTickCount to increment once too many times. See "bug 2" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - CPU is sleeping in vPortSuppressTicksAndSleep() - Something other than the SysTick wakes the CPU. - vPortSuppressTicksAndSleep() calculates the number of SysTick counts until the next tick. The bug occurs only if this number is small. - vPortSuppressTicksAndSleep() puts this small number into the SysTick reload register, and starts SysTick. - vPortSuppressTicksAndSleep() calls vTaskStepTick() - While vTaskStepTick() executes, the SysTick expires. The ISR pends because interrupts are masked, and SysTick starts a 2nd period still based on the small number of counts in its reload register. This 2nd period is undesirable and is likely to cause the error noted below. - vPortSuppressTicksAndSleep() puts the normal tick duration into the SysTick's reload register. - vPortSuppressTicksAndSleep() unmasks interrupts before the SysTick starts a new period based on the new value in the reload register. [This is a race condition that can go either way, but for the bug to occur, the race must play out this way.] - The pending SysTick ISR executes and increments xPendedTicks. - The SysTick expires again, finishing the second very small period, and starts a new period this time based on the full tick duration. - The SysTick ISR increments xPendedTicks (or xTickCount) even though only a tiny fraction of a tick period has elapsed since the previous tick. The bug occurs when *two* consecutive small periods of the SysTick are both counted as ticks. The root cause is a race caused by the small SysTick period. If vPortSuppressTicksAndSleep() unmasks interrupts *after* the small period expires but *before* the SysTick starts a period based on the full tick period, then two small periods are counted as ticks when only one should be counted. The end result is xTickCount advancing nearly one full tick more than time actually elapsed as measured by the SysTick. This is not the kind of time slippage normally associated with tickless idle. After this commit the code starts the SysTick and then immediately modifies the reload register to ensure the very short cycle (if any) is conducted only once. This strategy requires special consideration for the build option that configures SysTick to use a divided clock. To avoid waiting around for the SysTick to load value from the reload register, the new code temporarily configures the SysTick to use the undivided clock. The resulting timing error is typical for tickless idle. The error (commonly known as drift or slippage in kernel time) caused by this strategy is equivalent to one or two counts in ulStoppedTimerCompensation. This commit also updates comments and #define symbols related to the SysTick clock option. The SysTick can optionally be clocked by a divided version of the CPU clock (commonly divide-by-8). The new code in this commit adjusts these comments and symbols to make them clearer and more useful in configurations that use the divided clock. The fix made in this commit requires the use of these symbols, as noted in the code comments. * Fix tickless idle with alternate systick clocking Prior to this commit, in configurations using the alternate SysTick clocking, vPortSuppressTicksAndSleep() might cause xTickCount to jump ahead as much as the entire expected idle time or fall behind as much as one full tick compared to time as measured by the SysTick. SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. The SysTick has a configuration option to be clocked by an alternate clock besides the core clock. This alternate clock is MCU dependent. Scenarios Fixed --------------- The new code in this commit handles the following scenarios that were not handled correctly prior to this commit. 1. Before the sleep, vPortSuppressTicksAndSleep() stops the SysTick on zero, long after SysTick reached zero. Prior to this commit, this scenario caused xTickCount to jump ahead one full tick for the same reason documented here: https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/0c7b04bd3a745c52151abebc882eed3f811c4c81 2. After the sleep, vPortSuppressTicksAndSleep() stops the SysTick before it loads the counter from the reload register. Prior to this commit, this scenario caused xTickCount to jump ahead by the entire expected idle time (xExpectedIdleTime) because the current-count register is zero before it loads from the reload register. 3. Prior to return, vPortSuppressTicksAndSleep() attempts to start a short SysTick period when the current SysTick clock cycle has a lot of time remaining. Prior to this commit, this scenario could cause xTickCount to fall behind by as much as nearly one full tick because the short SysTick cycle never started. Note that #3 is partially fixed by https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/967acc9b200d3d4beeb289d9da9e88798074b431 even though that commit addresses a different issue. So this commit completes the partial fix. * Improve comments and name of preprocessor symbol Add a note in the code comments that SysTick requests an interrupt when decrementing from 1 to 0, so that's why stopping SysTick on zero is a special case. Readers might unknowingly assume that SysTick requests an interrupt when wrapping from 0 back to the load-register value. Reconsider new "_SETTING" suffix since "_CONFIG" suffix seems more descriptive. The code relies on *both* of these preprocessor symbols: portNVIC_SYSTICK_CLK_BIT portNVIC_SYSTICK_CLK_BIT_CONFIG **new** A meaningful suffix is really helpful to distinguish the two symbols. * Revert introduction of 2nd name for NVIC register When I added portNVIC_ICSR_REG I didn't realize there was already a portNVIC_INT_CTRL_REG, which identifies the same register. Not good to have both. Note that portNVIC_INT_CTRL_REG is defined in portmacro.h and is already used in this file (port.c). * Replicate to other Cortex M ports Also set a new fiddle factor based on tests with a CM4F. I used gcc, optimizing at -O1. Users can fine-tune as needed. Also add configSYSTICK_CLOCK_HZ to the CM0 ports to be just like the other Cortex M ports. This change allowed uniformity in the default tickless implementations across all Cortex M ports. And CM0 is likely to benefit from configSYSTICK_CLOCK_HZ, especially considering new CM0 devices with very fast CPU clock speeds. * Revert changes to IAR-CM0-portmacro.h portNVIC_INT_CTRL_REG was already defined in port.c. No need to define it in portmacro.h. * Handle edge cases with slow SysTick clock Co-authored-by: Cobus van Eeden <35851496+cobusve@users.noreply.github.com> Co-authored-by: abhidixi11 <44424462+abhidixi11@users.noreply.github.com> Co-authored-by: Joseph Julicher <jjulicher@mac.com> Co-authored-by: alfred gedeon <28123637+alfred2g@users.noreply.github.com>
2 years ago
{
/* De-prioritize the non-secure exceptions so that the
* non-secure pendSV runs at the lowest priority. */
SecureInit_DePrioritizeNSExceptions();
Tickless idle fixes/improvement (#59) * Fix tickless idle when stopping systick on zero... ...and don't stop SysTick at all in the eAbortSleep case. Prior to this commit, if vPortSuppressTicksAndSleep() happens to stop the SysTick on zero, then after tickless idle ends, xTickCount advances one full tick more than the time that actually elapsed as measured by the SysTick. See "bug 1" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - Idle task calls vPortSuppressTicksAndSleep(xExpectedIdleTime = 2). [Doesn't have to be "2" -- could be any number.] - vPortSuppressTicksAndSleep() stops SysTick, and the current-count register happens to stop on zero. - SysTick ISR executes, setting xPendedTicks = 1 - vPortSuppressTicksAndSleep() masks interrupts and calls eTaskConfirmSleepModeStatus() which confirms the sleep operation. *** - vPortSuppressTicksAndSleep() configures SysTick for 1 full tick (xExpectedIdleTime - 1) plus the current-count register (which is 0) - One tick period elapses in sleep. - SysTick wakes CPU, ISR executes and increments xPendedTicks to 2. - vPortSuppressTicksAndSleep() calls vTaskStepTick(1), then returns. - Idle task resumes scheduler, which increments xTickCount twice (for xPendedTicks = 2) In the end, two ticks elapsed as measured by SysTick, but the code increments xTickCount three times. The root cause is that the code assumes the SysTick current-count register always contains the number of SysTick counts remaining in the current tick period. However, when the current-count register is zero, there are ulTimerCountsForOneTick counts remaining, not zero. This error is not the kind of time slippage normally associated with tickless idle. *** Note that a recent commit https://github.com/FreeRTOS/FreeRTOS-Kernel/commit/e1b98f0 results in eAbortSleep in this case, due to xPendedTicks != 0. That commit does mostly resolve this bug without specifically mentioning it, and without this commit. But that resolution allows the code in port.c not to directly address the special case of stopping SysTick on zero in any code or comments. That commit also generates additional instances of eAbortSleep, and a second purpose of this commit is to optimize how vPortSuppressTicksAndSleep() behaves for eAbortSleep, as noted below. This commit also includes an optimization to avoid stopping the SysTick when eTaskConfirmSleepModeStatus() returns eAbortSleep. This optimization belongs with this fix because the method of handling the SysTick being stopped on zero changes with this optimization. * Fix imminent tick rescheduled after tickless idle Prior to this commit, if something other than systick wakes the CPU from tickless idle, vPortSuppressTicksAndSleep() might cause xTickCount to increment once too many times. See "bug 2" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - CPU is sleeping in vPortSuppressTicksAndSleep() - Something other than the SysTick wakes the CPU. - vPortSuppressTicksAndSleep() calculates the number of SysTick counts until the next tick. The bug occurs only if this number is small. - vPortSuppressTicksAndSleep() puts this small number into the SysTick reload register, and starts SysTick. - vPortSuppressTicksAndSleep() calls vTaskStepTick() - While vTaskStepTick() executes, the SysTick expires. The ISR pends because interrupts are masked, and SysTick starts a 2nd period still based on the small number of counts in its reload register. This 2nd period is undesirable and is likely to cause the error noted below. - vPortSuppressTicksAndSleep() puts the normal tick duration into the SysTick's reload register. - vPortSuppressTicksAndSleep() unmasks interrupts before the SysTick starts a new period based on the new value in the reload register. [This is a race condition that can go either way, but for the bug to occur, the race must play out this way.] - The pending SysTick ISR executes and increments xPendedTicks. - The SysTick expires again, finishing the second very small period, and starts a new period this time based on the full tick duration. - The SysTick ISR increments xPendedTicks (or xTickCount) even though only a tiny fraction of a tick period has elapsed since the previous tick. The bug occurs when *two* consecutive small periods of the SysTick are both counted as ticks. The root cause is a race caused by the small SysTick period. If vPortSuppressTicksAndSleep() unmasks interrupts *after* the small period expires but *before* the SysTick starts a period based on the full tick period, then two small periods are counted as ticks when only one should be counted. The end result is xTickCount advancing nearly one full tick more than time actually elapsed as measured by the SysTick. This is not the kind of time slippage normally associated with tickless idle. After this commit the code starts the SysTick and then immediately modifies the reload register to ensure the very short cycle (if any) is conducted only once. This strategy requires special consideration for the build option that configures SysTick to use a divided clock. To avoid waiting around for the SysTick to load value from the reload register, the new code temporarily configures the SysTick to use the undivided clock. The resulting timing error is typical for tickless idle. The error (commonly known as drift or slippage in kernel time) caused by this strategy is equivalent to one or two counts in ulStoppedTimerCompensation. This commit also updates comments and #define symbols related to the SysTick clock option. The SysTick can optionally be clocked by a divided version of the CPU clock (commonly divide-by-8). The new code in this commit adjusts these comments and symbols to make them clearer and more useful in configurations that use the divided clock. The fix made in this commit requires the use of these symbols, as noted in the code comments. * Fix tickless idle with alternate systick clocking Prior to this commit, in configurations using the alternate SysTick clocking, vPortSuppressTicksAndSleep() might cause xTickCount to jump ahead as much as the entire expected idle time or fall behind as much as one full tick compared to time as measured by the SysTick. SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. The SysTick has a configuration option to be clocked by an alternate clock besides the core clock. This alternate clock is MCU dependent. Scenarios Fixed --------------- The new code in this commit handles the following scenarios that were not handled correctly prior to this commit. 1. Before the sleep, vPortSuppressTicksAndSleep() stops the SysTick on zero, long after SysTick reached zero. Prior to this commit, this scenario caused xTickCount to jump ahead one full tick for the same reason documented here: https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/0c7b04bd3a745c52151abebc882eed3f811c4c81 2. After the sleep, vPortSuppressTicksAndSleep() stops the SysTick before it loads the counter from the reload register. Prior to this commit, this scenario caused xTickCount to jump ahead by the entire expected idle time (xExpectedIdleTime) because the current-count register is zero before it loads from the reload register. 3. Prior to return, vPortSuppressTicksAndSleep() attempts to start a short SysTick period when the current SysTick clock cycle has a lot of time remaining. Prior to this commit, this scenario could cause xTickCount to fall behind by as much as nearly one full tick because the short SysTick cycle never started. Note that #3 is partially fixed by https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/967acc9b200d3d4beeb289d9da9e88798074b431 even though that commit addresses a different issue. So this commit completes the partial fix. * Improve comments and name of preprocessor symbol Add a note in the code comments that SysTick requests an interrupt when decrementing from 1 to 0, so that's why stopping SysTick on zero is a special case. Readers might unknowingly assume that SysTick requests an interrupt when wrapping from 0 back to the load-register value. Reconsider new "_SETTING" suffix since "_CONFIG" suffix seems more descriptive. The code relies on *both* of these preprocessor symbols: portNVIC_SYSTICK_CLK_BIT portNVIC_SYSTICK_CLK_BIT_CONFIG **new** A meaningful suffix is really helpful to distinguish the two symbols. * Revert introduction of 2nd name for NVIC register When I added portNVIC_ICSR_REG I didn't realize there was already a portNVIC_INT_CTRL_REG, which identifies the same register. Not good to have both. Note that portNVIC_INT_CTRL_REG is defined in portmacro.h and is already used in this file (port.c). * Replicate to other Cortex M ports Also set a new fiddle factor based on tests with a CM4F. I used gcc, optimizing at -O1. Users can fine-tune as needed. Also add configSYSTICK_CLOCK_HZ to the CM0 ports to be just like the other Cortex M ports. This change allowed uniformity in the default tickless implementations across all Cortex M ports. And CM0 is likely to benefit from configSYSTICK_CLOCK_HZ, especially considering new CM0 devices with very fast CPU clock speeds. * Revert changes to IAR-CM0-portmacro.h portNVIC_INT_CTRL_REG was already defined in port.c. No need to define it in portmacro.h. * Handle edge cases with slow SysTick clock Co-authored-by: Cobus van Eeden <35851496+cobusve@users.noreply.github.com> Co-authored-by: abhidixi11 <44424462+abhidixi11@users.noreply.github.com> Co-authored-by: Joseph Julicher <jjulicher@mac.com> Co-authored-by: alfred gedeon <28123637+alfred2g@users.noreply.github.com>
2 years ago
/* Initialize the secure context management system. */
SecureContext_Init();
}
#endif /* configENABLE_TRUSTZONE */
#if ( configENABLE_FPU == 1 )
Tickless idle fixes/improvement (#59) * Fix tickless idle when stopping systick on zero... ...and don't stop SysTick at all in the eAbortSleep case. Prior to this commit, if vPortSuppressTicksAndSleep() happens to stop the SysTick on zero, then after tickless idle ends, xTickCount advances one full tick more than the time that actually elapsed as measured by the SysTick. See "bug 1" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - Idle task calls vPortSuppressTicksAndSleep(xExpectedIdleTime = 2). [Doesn't have to be "2" -- could be any number.] - vPortSuppressTicksAndSleep() stops SysTick, and the current-count register happens to stop on zero. - SysTick ISR executes, setting xPendedTicks = 1 - vPortSuppressTicksAndSleep() masks interrupts and calls eTaskConfirmSleepModeStatus() which confirms the sleep operation. *** - vPortSuppressTicksAndSleep() configures SysTick for 1 full tick (xExpectedIdleTime - 1) plus the current-count register (which is 0) - One tick period elapses in sleep. - SysTick wakes CPU, ISR executes and increments xPendedTicks to 2. - vPortSuppressTicksAndSleep() calls vTaskStepTick(1), then returns. - Idle task resumes scheduler, which increments xTickCount twice (for xPendedTicks = 2) In the end, two ticks elapsed as measured by SysTick, but the code increments xTickCount three times. The root cause is that the code assumes the SysTick current-count register always contains the number of SysTick counts remaining in the current tick period. However, when the current-count register is zero, there are ulTimerCountsForOneTick counts remaining, not zero. This error is not the kind of time slippage normally associated with tickless idle. *** Note that a recent commit https://github.com/FreeRTOS/FreeRTOS-Kernel/commit/e1b98f0 results in eAbortSleep in this case, due to xPendedTicks != 0. That commit does mostly resolve this bug without specifically mentioning it, and without this commit. But that resolution allows the code in port.c not to directly address the special case of stopping SysTick on zero in any code or comments. That commit also generates additional instances of eAbortSleep, and a second purpose of this commit is to optimize how vPortSuppressTicksAndSleep() behaves for eAbortSleep, as noted below. This commit also includes an optimization to avoid stopping the SysTick when eTaskConfirmSleepModeStatus() returns eAbortSleep. This optimization belongs with this fix because the method of handling the SysTick being stopped on zero changes with this optimization. * Fix imminent tick rescheduled after tickless idle Prior to this commit, if something other than systick wakes the CPU from tickless idle, vPortSuppressTicksAndSleep() might cause xTickCount to increment once too many times. See "bug 2" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - CPU is sleeping in vPortSuppressTicksAndSleep() - Something other than the SysTick wakes the CPU. - vPortSuppressTicksAndSleep() calculates the number of SysTick counts until the next tick. The bug occurs only if this number is small. - vPortSuppressTicksAndSleep() puts this small number into the SysTick reload register, and starts SysTick. - vPortSuppressTicksAndSleep() calls vTaskStepTick() - While vTaskStepTick() executes, the SysTick expires. The ISR pends because interrupts are masked, and SysTick starts a 2nd period still based on the small number of counts in its reload register. This 2nd period is undesirable and is likely to cause the error noted below. - vPortSuppressTicksAndSleep() puts the normal tick duration into the SysTick's reload register. - vPortSuppressTicksAndSleep() unmasks interrupts before the SysTick starts a new period based on the new value in the reload register. [This is a race condition that can go either way, but for the bug to occur, the race must play out this way.] - The pending SysTick ISR executes and increments xPendedTicks. - The SysTick expires again, finishing the second very small period, and starts a new period this time based on the full tick duration. - The SysTick ISR increments xPendedTicks (or xTickCount) even though only a tiny fraction of a tick period has elapsed since the previous tick. The bug occurs when *two* consecutive small periods of the SysTick are both counted as ticks. The root cause is a race caused by the small SysTick period. If vPortSuppressTicksAndSleep() unmasks interrupts *after* the small period expires but *before* the SysTick starts a period based on the full tick period, then two small periods are counted as ticks when only one should be counted. The end result is xTickCount advancing nearly one full tick more than time actually elapsed as measured by the SysTick. This is not the kind of time slippage normally associated with tickless idle. After this commit the code starts the SysTick and then immediately modifies the reload register to ensure the very short cycle (if any) is conducted only once. This strategy requires special consideration for the build option that configures SysTick to use a divided clock. To avoid waiting around for the SysTick to load value from the reload register, the new code temporarily configures the SysTick to use the undivided clock. The resulting timing error is typical for tickless idle. The error (commonly known as drift or slippage in kernel time) caused by this strategy is equivalent to one or two counts in ulStoppedTimerCompensation. This commit also updates comments and #define symbols related to the SysTick clock option. The SysTick can optionally be clocked by a divided version of the CPU clock (commonly divide-by-8). The new code in this commit adjusts these comments and symbols to make them clearer and more useful in configurations that use the divided clock. The fix made in this commit requires the use of these symbols, as noted in the code comments. * Fix tickless idle with alternate systick clocking Prior to this commit, in configurations using the alternate SysTick clocking, vPortSuppressTicksAndSleep() might cause xTickCount to jump ahead as much as the entire expected idle time or fall behind as much as one full tick compared to time as measured by the SysTick. SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. The SysTick has a configuration option to be clocked by an alternate clock besides the core clock. This alternate clock is MCU dependent. Scenarios Fixed --------------- The new code in this commit handles the following scenarios that were not handled correctly prior to this commit. 1. Before the sleep, vPortSuppressTicksAndSleep() stops the SysTick on zero, long after SysTick reached zero. Prior to this commit, this scenario caused xTickCount to jump ahead one full tick for the same reason documented here: https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/0c7b04bd3a745c52151abebc882eed3f811c4c81 2. After the sleep, vPortSuppressTicksAndSleep() stops the SysTick before it loads the counter from the reload register. Prior to this commit, this scenario caused xTickCount to jump ahead by the entire expected idle time (xExpectedIdleTime) because the current-count register is zero before it loads from the reload register. 3. Prior to return, vPortSuppressTicksAndSleep() attempts to start a short SysTick period when the current SysTick clock cycle has a lot of time remaining. Prior to this commit, this scenario could cause xTickCount to fall behind by as much as nearly one full tick because the short SysTick cycle never started. Note that #3 is partially fixed by https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/967acc9b200d3d4beeb289d9da9e88798074b431 even though that commit addresses a different issue. So this commit completes the partial fix. * Improve comments and name of preprocessor symbol Add a note in the code comments that SysTick requests an interrupt when decrementing from 1 to 0, so that's why stopping SysTick on zero is a special case. Readers might unknowingly assume that SysTick requests an interrupt when wrapping from 0 back to the load-register value. Reconsider new "_SETTING" suffix since "_CONFIG" suffix seems more descriptive. The code relies on *both* of these preprocessor symbols: portNVIC_SYSTICK_CLK_BIT portNVIC_SYSTICK_CLK_BIT_CONFIG **new** A meaningful suffix is really helpful to distinguish the two symbols. * Revert introduction of 2nd name for NVIC register When I added portNVIC_ICSR_REG I didn't realize there was already a portNVIC_INT_CTRL_REG, which identifies the same register. Not good to have both. Note that portNVIC_INT_CTRL_REG is defined in portmacro.h and is already used in this file (port.c). * Replicate to other Cortex M ports Also set a new fiddle factor based on tests with a CM4F. I used gcc, optimizing at -O1. Users can fine-tune as needed. Also add configSYSTICK_CLOCK_HZ to the CM0 ports to be just like the other Cortex M ports. This change allowed uniformity in the default tickless implementations across all Cortex M ports. And CM0 is likely to benefit from configSYSTICK_CLOCK_HZ, especially considering new CM0 devices with very fast CPU clock speeds. * Revert changes to IAR-CM0-portmacro.h portNVIC_INT_CTRL_REG was already defined in port.c. No need to define it in portmacro.h. * Handle edge cases with slow SysTick clock Co-authored-by: Cobus van Eeden <35851496+cobusve@users.noreply.github.com> Co-authored-by: abhidixi11 <44424462+abhidixi11@users.noreply.github.com> Co-authored-by: Joseph Julicher <jjulicher@mac.com> Co-authored-by: alfred gedeon <28123637+alfred2g@users.noreply.github.com>
2 years ago
{
/* Setup the Floating Point Unit (FPU). */
prvSetupFPU();
}
#endif /* configENABLE_FPU */
/* Setup the context of the first task so that the first task starts
* executing. */
vRestoreContextOfFirstTask();
break;
#if ( ( configENABLE_MPU == 1 ) && ( configUSE_MPU_WRAPPERS_V1 == 1 ) )
case portSVC_RAISE_PRIVILEGE:
/* Only raise the privilege, if the svc was raised from any of
* the system calls. */
if( ( ulPC >= ( uint32_t ) __syscalls_flash_start__ ) &&
( ulPC <= ( uint32_t ) __syscalls_flash_end__ ) )
{
vRaisePrivilege();
}
break;
#endif /* ( configENABLE_MPU == 1 ) && ( configUSE_MPU_WRAPPERS_V1 == 1 ) */
default:
/* Incorrect SVC call. */
configASSERT( pdFALSE );
}
}
/*-----------------------------------------------------------*/
Armv8-M (except Cortex-M23) interrupt priority checking (#673) * Armv8-M: Formatting changes Signed-off-by: Devaraj Ranganna <devaraj.ranganna@arm.com> * Armv8-M: Add support for interrupt priority check FreeRTOS provides `FromISR` system calls which can be called directly from interrupt service routines. It is crucial that the priority of these ISRs is set to same or lower value (numerically higher) than that of `configMAX_SYSCALL_INTERRUPT_PRIORITY`. For more information refer to https://www.FreeRTOS.org/RTOS-Cortex-M3-M4.html. Add a check to trigger an assert when an ISR with priority higher (numerically lower) than `configMAX_SYSCALL_INTERRUPT_PRIORITY` calls `FromISR` system calls if `configASSERT` macro is defined. In addition, add a config option `configQEMU_DISABLE_INTERRUPT_PRIO_BITS_CHECK` to disable interrupt priority check while running on QEMU. Based on the discussion https://gitlab.com/qemu-project/qemu/-/issues/1122, The interrupt priority bits in QEMU do not match the real hardware. Therefore the assert that checks the number of implemented bits and __NVIC_PRIO_BITS will always fail. The config option `configQEMU_DISABLE_INTERRUPT_PRIO_BITS_CHECK` should be defined in the `FreeRTOSConfig.h` for QEMU targets. Signed-off-by: Devaraj Ranganna <devaraj.ranganna@arm.com> * Use SHPR2 for calculating interrupt priority bits This removes the dependency on the secure software to mark the interrupt as non-secure. Signed-off-by: Gaurav Aggarwal <aggarg@amazon.com> --------- Signed-off-by: Devaraj Ranganna <devaraj.ranganna@arm.com> Signed-off-by: Gaurav Aggarwal <aggarg@amazon.com> Co-authored-by: Gaurav Aggarwal <aggarg@amazon.com> Co-authored-by: Gaurav-Aggarwal-AWS <33462878+aggarg@users.noreply.github.com>
2 years ago
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
#if ( ( configENABLE_MPU == 1 ) && ( configUSE_MPU_WRAPPERS_V1 == 0 ) )
void vSystemCallEnter( uint32_t * pulTaskStack,
uint32_t ulLR ) /* PRIVILEGED_FUNCTION */
Tickless idle fixes/improvement (#59) * Fix tickless idle when stopping systick on zero... ...and don't stop SysTick at all in the eAbortSleep case. Prior to this commit, if vPortSuppressTicksAndSleep() happens to stop the SysTick on zero, then after tickless idle ends, xTickCount advances one full tick more than the time that actually elapsed as measured by the SysTick. See "bug 1" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - Idle task calls vPortSuppressTicksAndSleep(xExpectedIdleTime = 2). [Doesn't have to be "2" -- could be any number.] - vPortSuppressTicksAndSleep() stops SysTick, and the current-count register happens to stop on zero. - SysTick ISR executes, setting xPendedTicks = 1 - vPortSuppressTicksAndSleep() masks interrupts and calls eTaskConfirmSleepModeStatus() which confirms the sleep operation. *** - vPortSuppressTicksAndSleep() configures SysTick for 1 full tick (xExpectedIdleTime - 1) plus the current-count register (which is 0) - One tick period elapses in sleep. - SysTick wakes CPU, ISR executes and increments xPendedTicks to 2. - vPortSuppressTicksAndSleep() calls vTaskStepTick(1), then returns. - Idle task resumes scheduler, which increments xTickCount twice (for xPendedTicks = 2) In the end, two ticks elapsed as measured by SysTick, but the code increments xTickCount three times. The root cause is that the code assumes the SysTick current-count register always contains the number of SysTick counts remaining in the current tick period. However, when the current-count register is zero, there are ulTimerCountsForOneTick counts remaining, not zero. This error is not the kind of time slippage normally associated with tickless idle. *** Note that a recent commit https://github.com/FreeRTOS/FreeRTOS-Kernel/commit/e1b98f0 results in eAbortSleep in this case, due to xPendedTicks != 0. That commit does mostly resolve this bug without specifically mentioning it, and without this commit. But that resolution allows the code in port.c not to directly address the special case of stopping SysTick on zero in any code or comments. That commit also generates additional instances of eAbortSleep, and a second purpose of this commit is to optimize how vPortSuppressTicksAndSleep() behaves for eAbortSleep, as noted below. This commit also includes an optimization to avoid stopping the SysTick when eTaskConfirmSleepModeStatus() returns eAbortSleep. This optimization belongs with this fix because the method of handling the SysTick being stopped on zero changes with this optimization. * Fix imminent tick rescheduled after tickless idle Prior to this commit, if something other than systick wakes the CPU from tickless idle, vPortSuppressTicksAndSleep() might cause xTickCount to increment once too many times. See "bug 2" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - CPU is sleeping in vPortSuppressTicksAndSleep() - Something other than the SysTick wakes the CPU. - vPortSuppressTicksAndSleep() calculates the number of SysTick counts until the next tick. The bug occurs only if this number is small. - vPortSuppressTicksAndSleep() puts this small number into the SysTick reload register, and starts SysTick. - vPortSuppressTicksAndSleep() calls vTaskStepTick() - While vTaskStepTick() executes, the SysTick expires. The ISR pends because interrupts are masked, and SysTick starts a 2nd period still based on the small number of counts in its reload register. This 2nd period is undesirable and is likely to cause the error noted below. - vPortSuppressTicksAndSleep() puts the normal tick duration into the SysTick's reload register. - vPortSuppressTicksAndSleep() unmasks interrupts before the SysTick starts a new period based on the new value in the reload register. [This is a race condition that can go either way, but for the bug to occur, the race must play out this way.] - The pending SysTick ISR executes and increments xPendedTicks. - The SysTick expires again, finishing the second very small period, and starts a new period this time based on the full tick duration. - The SysTick ISR increments xPendedTicks (or xTickCount) even though only a tiny fraction of a tick period has elapsed since the previous tick. The bug occurs when *two* consecutive small periods of the SysTick are both counted as ticks. The root cause is a race caused by the small SysTick period. If vPortSuppressTicksAndSleep() unmasks interrupts *after* the small period expires but *before* the SysTick starts a period based on the full tick period, then two small periods are counted as ticks when only one should be counted. The end result is xTickCount advancing nearly one full tick more than time actually elapsed as measured by the SysTick. This is not the kind of time slippage normally associated with tickless idle. After this commit the code starts the SysTick and then immediately modifies the reload register to ensure the very short cycle (if any) is conducted only once. This strategy requires special consideration for the build option that configures SysTick to use a divided clock. To avoid waiting around for the SysTick to load value from the reload register, the new code temporarily configures the SysTick to use the undivided clock. The resulting timing error is typical for tickless idle. The error (commonly known as drift or slippage in kernel time) caused by this strategy is equivalent to one or two counts in ulStoppedTimerCompensation. This commit also updates comments and #define symbols related to the SysTick clock option. The SysTick can optionally be clocked by a divided version of the CPU clock (commonly divide-by-8). The new code in this commit adjusts these comments and symbols to make them clearer and more useful in configurations that use the divided clock. The fix made in this commit requires the use of these symbols, as noted in the code comments. * Fix tickless idle with alternate systick clocking Prior to this commit, in configurations using the alternate SysTick clocking, vPortSuppressTicksAndSleep() might cause xTickCount to jump ahead as much as the entire expected idle time or fall behind as much as one full tick compared to time as measured by the SysTick. SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. The SysTick has a configuration option to be clocked by an alternate clock besides the core clock. This alternate clock is MCU dependent. Scenarios Fixed --------------- The new code in this commit handles the following scenarios that were not handled correctly prior to this commit. 1. Before the sleep, vPortSuppressTicksAndSleep() stops the SysTick on zero, long after SysTick reached zero. Prior to this commit, this scenario caused xTickCount to jump ahead one full tick for the same reason documented here: https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/0c7b04bd3a745c52151abebc882eed3f811c4c81 2. After the sleep, vPortSuppressTicksAndSleep() stops the SysTick before it loads the counter from the reload register. Prior to this commit, this scenario caused xTickCount to jump ahead by the entire expected idle time (xExpectedIdleTime) because the current-count register is zero before it loads from the reload register. 3. Prior to return, vPortSuppressTicksAndSleep() attempts to start a short SysTick period when the current SysTick clock cycle has a lot of time remaining. Prior to this commit, this scenario could cause xTickCount to fall behind by as much as nearly one full tick because the short SysTick cycle never started. Note that #3 is partially fixed by https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/967acc9b200d3d4beeb289d9da9e88798074b431 even though that commit addresses a different issue. So this commit completes the partial fix. * Improve comments and name of preprocessor symbol Add a note in the code comments that SysTick requests an interrupt when decrementing from 1 to 0, so that's why stopping SysTick on zero is a special case. Readers might unknowingly assume that SysTick requests an interrupt when wrapping from 0 back to the load-register value. Reconsider new "_SETTING" suffix since "_CONFIG" suffix seems more descriptive. The code relies on *both* of these preprocessor symbols: portNVIC_SYSTICK_CLK_BIT portNVIC_SYSTICK_CLK_BIT_CONFIG **new** A meaningful suffix is really helpful to distinguish the two symbols. * Revert introduction of 2nd name for NVIC register When I added portNVIC_ICSR_REG I didn't realize there was already a portNVIC_INT_CTRL_REG, which identifies the same register. Not good to have both. Note that portNVIC_INT_CTRL_REG is defined in portmacro.h and is already used in this file (port.c). * Replicate to other Cortex M ports Also set a new fiddle factor based on tests with a CM4F. I used gcc, optimizing at -O1. Users can fine-tune as needed. Also add configSYSTICK_CLOCK_HZ to the CM0 ports to be just like the other Cortex M ports. This change allowed uniformity in the default tickless implementations across all Cortex M ports. And CM0 is likely to benefit from configSYSTICK_CLOCK_HZ, especially considering new CM0 devices with very fast CPU clock speeds. * Revert changes to IAR-CM0-portmacro.h portNVIC_INT_CTRL_REG was already defined in port.c. No need to define it in portmacro.h. * Handle edge cases with slow SysTick clock Co-authored-by: Cobus van Eeden <35851496+cobusve@users.noreply.github.com> Co-authored-by: abhidixi11 <44424462+abhidixi11@users.noreply.github.com> Co-authored-by: Joseph Julicher <jjulicher@mac.com> Co-authored-by: alfred gedeon <28123637+alfred2g@users.noreply.github.com>
2 years ago
{
extern TaskHandle_t pxCurrentTCB;
xMPU_SETTINGS * pxMpuSettings;
uint32_t * pulSystemCallStack;
uint32_t ulStackFrameSize, ulSystemCallLocation, i;
#if defined( __ARMCC_VERSION )
/* Declaration when these variable are defined in code instead of being
* exported from linker scripts. */
extern uint32_t * __syscalls_flash_start__;
extern uint32_t * __syscalls_flash_end__;
#else
/* Declaration when these variable are exported from linker scripts. */
extern uint32_t __syscalls_flash_start__[];
extern uint32_t __syscalls_flash_end__[];
#endif /* #if defined( __ARMCC_VERSION ) */
Tickless idle fixes/improvement (#59) * Fix tickless idle when stopping systick on zero... ...and don't stop SysTick at all in the eAbortSleep case. Prior to this commit, if vPortSuppressTicksAndSleep() happens to stop the SysTick on zero, then after tickless idle ends, xTickCount advances one full tick more than the time that actually elapsed as measured by the SysTick. See "bug 1" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - Idle task calls vPortSuppressTicksAndSleep(xExpectedIdleTime = 2). [Doesn't have to be "2" -- could be any number.] - vPortSuppressTicksAndSleep() stops SysTick, and the current-count register happens to stop on zero. - SysTick ISR executes, setting xPendedTicks = 1 - vPortSuppressTicksAndSleep() masks interrupts and calls eTaskConfirmSleepModeStatus() which confirms the sleep operation. *** - vPortSuppressTicksAndSleep() configures SysTick for 1 full tick (xExpectedIdleTime - 1) plus the current-count register (which is 0) - One tick period elapses in sleep. - SysTick wakes CPU, ISR executes and increments xPendedTicks to 2. - vPortSuppressTicksAndSleep() calls vTaskStepTick(1), then returns. - Idle task resumes scheduler, which increments xTickCount twice (for xPendedTicks = 2) In the end, two ticks elapsed as measured by SysTick, but the code increments xTickCount three times. The root cause is that the code assumes the SysTick current-count register always contains the number of SysTick counts remaining in the current tick period. However, when the current-count register is zero, there are ulTimerCountsForOneTick counts remaining, not zero. This error is not the kind of time slippage normally associated with tickless idle. *** Note that a recent commit https://github.com/FreeRTOS/FreeRTOS-Kernel/commit/e1b98f0 results in eAbortSleep in this case, due to xPendedTicks != 0. That commit does mostly resolve this bug without specifically mentioning it, and without this commit. But that resolution allows the code in port.c not to directly address the special case of stopping SysTick on zero in any code or comments. That commit also generates additional instances of eAbortSleep, and a second purpose of this commit is to optimize how vPortSuppressTicksAndSleep() behaves for eAbortSleep, as noted below. This commit also includes an optimization to avoid stopping the SysTick when eTaskConfirmSleepModeStatus() returns eAbortSleep. This optimization belongs with this fix because the method of handling the SysTick being stopped on zero changes with this optimization. * Fix imminent tick rescheduled after tickless idle Prior to this commit, if something other than systick wakes the CPU from tickless idle, vPortSuppressTicksAndSleep() might cause xTickCount to increment once too many times. See "bug 2" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - CPU is sleeping in vPortSuppressTicksAndSleep() - Something other than the SysTick wakes the CPU. - vPortSuppressTicksAndSleep() calculates the number of SysTick counts until the next tick. The bug occurs only if this number is small. - vPortSuppressTicksAndSleep() puts this small number into the SysTick reload register, and starts SysTick. - vPortSuppressTicksAndSleep() calls vTaskStepTick() - While vTaskStepTick() executes, the SysTick expires. The ISR pends because interrupts are masked, and SysTick starts a 2nd period still based on the small number of counts in its reload register. This 2nd period is undesirable and is likely to cause the error noted below. - vPortSuppressTicksAndSleep() puts the normal tick duration into the SysTick's reload register. - vPortSuppressTicksAndSleep() unmasks interrupts before the SysTick starts a new period based on the new value in the reload register. [This is a race condition that can go either way, but for the bug to occur, the race must play out this way.] - The pending SysTick ISR executes and increments xPendedTicks. - The SysTick expires again, finishing the second very small period, and starts a new period this time based on the full tick duration. - The SysTick ISR increments xPendedTicks (or xTickCount) even though only a tiny fraction of a tick period has elapsed since the previous tick. The bug occurs when *two* consecutive small periods of the SysTick are both counted as ticks. The root cause is a race caused by the small SysTick period. If vPortSuppressTicksAndSleep() unmasks interrupts *after* the small period expires but *before* the SysTick starts a period based on the full tick period, then two small periods are counted as ticks when only one should be counted. The end result is xTickCount advancing nearly one full tick more than time actually elapsed as measured by the SysTick. This is not the kind of time slippage normally associated with tickless idle. After this commit the code starts the SysTick and then immediately modifies the reload register to ensure the very short cycle (if any) is conducted only once. This strategy requires special consideration for the build option that configures SysTick to use a divided clock. To avoid waiting around for the SysTick to load value from the reload register, the new code temporarily configures the SysTick to use the undivided clock. The resulting timing error is typical for tickless idle. The error (commonly known as drift or slippage in kernel time) caused by this strategy is equivalent to one or two counts in ulStoppedTimerCompensation. This commit also updates comments and #define symbols related to the SysTick clock option. The SysTick can optionally be clocked by a divided version of the CPU clock (commonly divide-by-8). The new code in this commit adjusts these comments and symbols to make them clearer and more useful in configurations that use the divided clock. The fix made in this commit requires the use of these symbols, as noted in the code comments. * Fix tickless idle with alternate systick clocking Prior to this commit, in configurations using the alternate SysTick clocking, vPortSuppressTicksAndSleep() might cause xTickCount to jump ahead as much as the entire expected idle time or fall behind as much as one full tick compared to time as measured by the SysTick. SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. The SysTick has a configuration option to be clocked by an alternate clock besides the core clock. This alternate clock is MCU dependent. Scenarios Fixed --------------- The new code in this commit handles the following scenarios that were not handled correctly prior to this commit. 1. Before the sleep, vPortSuppressTicksAndSleep() stops the SysTick on zero, long after SysTick reached zero. Prior to this commit, this scenario caused xTickCount to jump ahead one full tick for the same reason documented here: https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/0c7b04bd3a745c52151abebc882eed3f811c4c81 2. After the sleep, vPortSuppressTicksAndSleep() stops the SysTick before it loads the counter from the reload register. Prior to this commit, this scenario caused xTickCount to jump ahead by the entire expected idle time (xExpectedIdleTime) because the current-count register is zero before it loads from the reload register. 3. Prior to return, vPortSuppressTicksAndSleep() attempts to start a short SysTick period when the current SysTick clock cycle has a lot of time remaining. Prior to this commit, this scenario could cause xTickCount to fall behind by as much as nearly one full tick because the short SysTick cycle never started. Note that #3 is partially fixed by https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/967acc9b200d3d4beeb289d9da9e88798074b431 even though that commit addresses a different issue. So this commit completes the partial fix. * Improve comments and name of preprocessor symbol Add a note in the code comments that SysTick requests an interrupt when decrementing from 1 to 0, so that's why stopping SysTick on zero is a special case. Readers might unknowingly assume that SysTick requests an interrupt when wrapping from 0 back to the load-register value. Reconsider new "_SETTING" suffix since "_CONFIG" suffix seems more descriptive. The code relies on *both* of these preprocessor symbols: portNVIC_SYSTICK_CLK_BIT portNVIC_SYSTICK_CLK_BIT_CONFIG **new** A meaningful suffix is really helpful to distinguish the two symbols. * Revert introduction of 2nd name for NVIC register When I added portNVIC_ICSR_REG I didn't realize there was already a portNVIC_INT_CTRL_REG, which identifies the same register. Not good to have both. Note that portNVIC_INT_CTRL_REG is defined in portmacro.h and is already used in this file (port.c). * Replicate to other Cortex M ports Also set a new fiddle factor based on tests with a CM4F. I used gcc, optimizing at -O1. Users can fine-tune as needed. Also add configSYSTICK_CLOCK_HZ to the CM0 ports to be just like the other Cortex M ports. This change allowed uniformity in the default tickless implementations across all Cortex M ports. And CM0 is likely to benefit from configSYSTICK_CLOCK_HZ, especially considering new CM0 devices with very fast CPU clock speeds. * Revert changes to IAR-CM0-portmacro.h portNVIC_INT_CTRL_REG was already defined in port.c. No need to define it in portmacro.h. * Handle edge cases with slow SysTick clock Co-authored-by: Cobus van Eeden <35851496+cobusve@users.noreply.github.com> Co-authored-by: abhidixi11 <44424462+abhidixi11@users.noreply.github.com> Co-authored-by: Joseph Julicher <jjulicher@mac.com> Co-authored-by: alfred gedeon <28123637+alfred2g@users.noreply.github.com>
2 years ago
ulSystemCallLocation = pulTaskStack[ portOFFSET_TO_PC ];
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
/* If the request did not come from the system call section, do nothing. */
if( ( ulSystemCallLocation >= ( uint32_t ) __syscalls_flash_start__ ) &&
( ulSystemCallLocation <= ( uint32_t ) __syscalls_flash_end__ ) )
{
pxMpuSettings = xTaskGetMPUSettings( pxCurrentTCB );
pulSystemCallStack = pxMpuSettings->xSystemCallStackInfo.pulSystemCallStack;
/* This is not NULL only for the duration of the system call. */
configASSERT( pxMpuSettings->xSystemCallStackInfo.pulTaskStack == NULL );
#if ( ( configENABLE_FPU == 1 ) || ( configENABLE_MVE == 1 ) )
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
{
if( ( ulLR & portEXC_RETURN_STACK_FRAME_TYPE_MASK ) == 0UL )
{
/* Extended frame i.e. FPU in use. */
ulStackFrameSize = 26;
__asm volatile (
" vpush {s0} \n" /* Trigger lazy stacking. */
" vpop {s0} \n" /* Nullify the affect of the above instruction. */
::: "memory"
);
}
else
{
/* Standard frame i.e. FPU not in use. */
ulStackFrameSize = 8;
}
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
}
#else /* if ( ( configENABLE_FPU == 1 ) || ( configENABLE_MVE == 1 ) ) */
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
{
ulStackFrameSize = 8;
}
#endif /* configENABLE_FPU || configENABLE_MVE */
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
/* Make space on the system call stack for the stack frame. */
pulSystemCallStack = pulSystemCallStack - ulStackFrameSize;
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
/* Copy the stack frame. */
for( i = 0; i < ulStackFrameSize; i++ )
{
pulSystemCallStack[ i ] = pulTaskStack[ i ];
}
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
/* Store the value of the LR and PSPLIM registers before the SVC was raised. We need to
* restore it when we exit from the system call. */
pxMpuSettings->xSystemCallStackInfo.ulLinkRegisterAtSystemCallEntry = pulTaskStack[ portOFFSET_TO_LR ];
#if ( portUSE_PSPLIM_REGISTER == 1 )
{
__asm volatile ( "mrs %0, psplim" : "=r" ( pxMpuSettings->xSystemCallStackInfo.ulStackLimitRegisterAtSystemCallEntry ) );
}
#endif
/* Use the pulSystemCallStack in thread mode. */
__asm volatile ( "msr psp, %0" : : "r" ( pulSystemCallStack ) );
#if ( portUSE_PSPLIM_REGISTER == 1 )
{
__asm volatile ( "msr psplim, %0" : : "r" ( pxMpuSettings->xSystemCallStackInfo.pulSystemCallStackLimit ) );
}
#endif
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
/* Remember the location where we should copy the stack frame when we exit from
* the system call. */
pxMpuSettings->xSystemCallStackInfo.pulTaskStack = pulTaskStack + ulStackFrameSize;
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
/* Record if the hardware used padding to force the stack pointer
* to be double word aligned. */
if( ( pulTaskStack[ portOFFSET_TO_PSR ] & portPSR_STACK_PADDING_MASK ) == portPSR_STACK_PADDING_MASK )
{
pxMpuSettings->ulTaskFlags |= portSTACK_FRAME_HAS_PADDING_FLAG;
}
else
{
pxMpuSettings->ulTaskFlags &= ( ~portSTACK_FRAME_HAS_PADDING_FLAG );
}
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
/* We ensure in pxPortInitialiseStack that the system call stack is
* double word aligned and therefore, there is no need of padding.
* Clear the bit[9] of stacked xPSR. */
pulSystemCallStack[ portOFFSET_TO_PSR ] &= ( ~portPSR_STACK_PADDING_MASK );
/* Raise the privilege for the duration of the system call. */
__asm volatile (
" mrs r0, control \n" /* Obtain current control value. */
" movs r1, #1 \n" /* r1 = 1. */
" bics r0, r1 \n" /* Clear nPRIV bit. */
" msr control, r0 \n" /* Write back new control value. */
::: "r0", "r1", "memory"
);
}
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
}
#endif /* ( configENABLE_MPU == 1 ) && ( configUSE_MPU_WRAPPERS_V1 == 0 ) */
/*-----------------------------------------------------------*/
#if ( ( configENABLE_MPU == 1 ) && ( configUSE_MPU_WRAPPERS_V1 == 0 ) )
void vSystemCallEnter_1( uint32_t * pulTaskStack,
uint32_t ulLR ) /* PRIVILEGED_FUNCTION */
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
{
extern TaskHandle_t pxCurrentTCB;
xMPU_SETTINGS * pxMpuSettings;
uint32_t * pulSystemCallStack;
uint32_t ulStackFrameSize, ulSystemCallLocation, i;
#if defined( __ARMCC_VERSION )
/* Declaration when these variable are defined in code instead of being
* exported from linker scripts. */
extern uint32_t * __syscalls_flash_start__;
extern uint32_t * __syscalls_flash_end__;
#else
/* Declaration when these variable are exported from linker scripts. */
extern uint32_t __syscalls_flash_start__[];
extern uint32_t __syscalls_flash_end__[];
#endif /* #if defined( __ARMCC_VERSION ) */
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
ulSystemCallLocation = pulTaskStack[ portOFFSET_TO_PC ];
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
/* If the request did not come from the system call section, do nothing. */
if( ( ulSystemCallLocation >= ( uint32_t ) __syscalls_flash_start__ ) &&
( ulSystemCallLocation <= ( uint32_t ) __syscalls_flash_end__ ) )
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
{
pxMpuSettings = xTaskGetMPUSettings( pxCurrentTCB );
pulSystemCallStack = pxMpuSettings->xSystemCallStackInfo.pulSystemCallStack;
/* This is not NULL only for the duration of the system call. */
configASSERT( pxMpuSettings->xSystemCallStackInfo.pulTaskStack == NULL );
#if ( ( configENABLE_FPU == 1 ) || ( configENABLE_MVE == 1 ) )
Tickless idle fixes/improvement (#59) * Fix tickless idle when stopping systick on zero... ...and don't stop SysTick at all in the eAbortSleep case. Prior to this commit, if vPortSuppressTicksAndSleep() happens to stop the SysTick on zero, then after tickless idle ends, xTickCount advances one full tick more than the time that actually elapsed as measured by the SysTick. See "bug 1" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - Idle task calls vPortSuppressTicksAndSleep(xExpectedIdleTime = 2). [Doesn't have to be "2" -- could be any number.] - vPortSuppressTicksAndSleep() stops SysTick, and the current-count register happens to stop on zero. - SysTick ISR executes, setting xPendedTicks = 1 - vPortSuppressTicksAndSleep() masks interrupts and calls eTaskConfirmSleepModeStatus() which confirms the sleep operation. *** - vPortSuppressTicksAndSleep() configures SysTick for 1 full tick (xExpectedIdleTime - 1) plus the current-count register (which is 0) - One tick period elapses in sleep. - SysTick wakes CPU, ISR executes and increments xPendedTicks to 2. - vPortSuppressTicksAndSleep() calls vTaskStepTick(1), then returns. - Idle task resumes scheduler, which increments xTickCount twice (for xPendedTicks = 2) In the end, two ticks elapsed as measured by SysTick, but the code increments xTickCount three times. The root cause is that the code assumes the SysTick current-count register always contains the number of SysTick counts remaining in the current tick period. However, when the current-count register is zero, there are ulTimerCountsForOneTick counts remaining, not zero. This error is not the kind of time slippage normally associated with tickless idle. *** Note that a recent commit https://github.com/FreeRTOS/FreeRTOS-Kernel/commit/e1b98f0 results in eAbortSleep in this case, due to xPendedTicks != 0. That commit does mostly resolve this bug without specifically mentioning it, and without this commit. But that resolution allows the code in port.c not to directly address the special case of stopping SysTick on zero in any code or comments. That commit also generates additional instances of eAbortSleep, and a second purpose of this commit is to optimize how vPortSuppressTicksAndSleep() behaves for eAbortSleep, as noted below. This commit also includes an optimization to avoid stopping the SysTick when eTaskConfirmSleepModeStatus() returns eAbortSleep. This optimization belongs with this fix because the method of handling the SysTick being stopped on zero changes with this optimization. * Fix imminent tick rescheduled after tickless idle Prior to this commit, if something other than systick wakes the CPU from tickless idle, vPortSuppressTicksAndSleep() might cause xTickCount to increment once too many times. See "bug 2" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - CPU is sleeping in vPortSuppressTicksAndSleep() - Something other than the SysTick wakes the CPU. - vPortSuppressTicksAndSleep() calculates the number of SysTick counts until the next tick. The bug occurs only if this number is small. - vPortSuppressTicksAndSleep() puts this small number into the SysTick reload register, and starts SysTick. - vPortSuppressTicksAndSleep() calls vTaskStepTick() - While vTaskStepTick() executes, the SysTick expires. The ISR pends because interrupts are masked, and SysTick starts a 2nd period still based on the small number of counts in its reload register. This 2nd period is undesirable and is likely to cause the error noted below. - vPortSuppressTicksAndSleep() puts the normal tick duration into the SysTick's reload register. - vPortSuppressTicksAndSleep() unmasks interrupts before the SysTick starts a new period based on the new value in the reload register. [This is a race condition that can go either way, but for the bug to occur, the race must play out this way.] - The pending SysTick ISR executes and increments xPendedTicks. - The SysTick expires again, finishing the second very small period, and starts a new period this time based on the full tick duration. - The SysTick ISR increments xPendedTicks (or xTickCount) even though only a tiny fraction of a tick period has elapsed since the previous tick. The bug occurs when *two* consecutive small periods of the SysTick are both counted as ticks. The root cause is a race caused by the small SysTick period. If vPortSuppressTicksAndSleep() unmasks interrupts *after* the small period expires but *before* the SysTick starts a period based on the full tick period, then two small periods are counted as ticks when only one should be counted. The end result is xTickCount advancing nearly one full tick more than time actually elapsed as measured by the SysTick. This is not the kind of time slippage normally associated with tickless idle. After this commit the code starts the SysTick and then immediately modifies the reload register to ensure the very short cycle (if any) is conducted only once. This strategy requires special consideration for the build option that configures SysTick to use a divided clock. To avoid waiting around for the SysTick to load value from the reload register, the new code temporarily configures the SysTick to use the undivided clock. The resulting timing error is typical for tickless idle. The error (commonly known as drift or slippage in kernel time) caused by this strategy is equivalent to one or two counts in ulStoppedTimerCompensation. This commit also updates comments and #define symbols related to the SysTick clock option. The SysTick can optionally be clocked by a divided version of the CPU clock (commonly divide-by-8). The new code in this commit adjusts these comments and symbols to make them clearer and more useful in configurations that use the divided clock. The fix made in this commit requires the use of these symbols, as noted in the code comments. * Fix tickless idle with alternate systick clocking Prior to this commit, in configurations using the alternate SysTick clocking, vPortSuppressTicksAndSleep() might cause xTickCount to jump ahead as much as the entire expected idle time or fall behind as much as one full tick compared to time as measured by the SysTick. SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. The SysTick has a configuration option to be clocked by an alternate clock besides the core clock. This alternate clock is MCU dependent. Scenarios Fixed --------------- The new code in this commit handles the following scenarios that were not handled correctly prior to this commit. 1. Before the sleep, vPortSuppressTicksAndSleep() stops the SysTick on zero, long after SysTick reached zero. Prior to this commit, this scenario caused xTickCount to jump ahead one full tick for the same reason documented here: https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/0c7b04bd3a745c52151abebc882eed3f811c4c81 2. After the sleep, vPortSuppressTicksAndSleep() stops the SysTick before it loads the counter from the reload register. Prior to this commit, this scenario caused xTickCount to jump ahead by the entire expected idle time (xExpectedIdleTime) because the current-count register is zero before it loads from the reload register. 3. Prior to return, vPortSuppressTicksAndSleep() attempts to start a short SysTick period when the current SysTick clock cycle has a lot of time remaining. Prior to this commit, this scenario could cause xTickCount to fall behind by as much as nearly one full tick because the short SysTick cycle never started. Note that #3 is partially fixed by https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/967acc9b200d3d4beeb289d9da9e88798074b431 even though that commit addresses a different issue. So this commit completes the partial fix. * Improve comments and name of preprocessor symbol Add a note in the code comments that SysTick requests an interrupt when decrementing from 1 to 0, so that's why stopping SysTick on zero is a special case. Readers might unknowingly assume that SysTick requests an interrupt when wrapping from 0 back to the load-register value. Reconsider new "_SETTING" suffix since "_CONFIG" suffix seems more descriptive. The code relies on *both* of these preprocessor symbols: portNVIC_SYSTICK_CLK_BIT portNVIC_SYSTICK_CLK_BIT_CONFIG **new** A meaningful suffix is really helpful to distinguish the two symbols. * Revert introduction of 2nd name for NVIC register When I added portNVIC_ICSR_REG I didn't realize there was already a portNVIC_INT_CTRL_REG, which identifies the same register. Not good to have both. Note that portNVIC_INT_CTRL_REG is defined in portmacro.h and is already used in this file (port.c). * Replicate to other Cortex M ports Also set a new fiddle factor based on tests with a CM4F. I used gcc, optimizing at -O1. Users can fine-tune as needed. Also add configSYSTICK_CLOCK_HZ to the CM0 ports to be just like the other Cortex M ports. This change allowed uniformity in the default tickless implementations across all Cortex M ports. And CM0 is likely to benefit from configSYSTICK_CLOCK_HZ, especially considering new CM0 devices with very fast CPU clock speeds. * Revert changes to IAR-CM0-portmacro.h portNVIC_INT_CTRL_REG was already defined in port.c. No need to define it in portmacro.h. * Handle edge cases with slow SysTick clock Co-authored-by: Cobus van Eeden <35851496+cobusve@users.noreply.github.com> Co-authored-by: abhidixi11 <44424462+abhidixi11@users.noreply.github.com> Co-authored-by: Joseph Julicher <jjulicher@mac.com> Co-authored-by: alfred gedeon <28123637+alfred2g@users.noreply.github.com>
2 years ago
{
if( ( ulLR & portEXC_RETURN_STACK_FRAME_TYPE_MASK ) == 0UL )
{
/* Extended frame i.e. FPU in use. */
ulStackFrameSize = 26;
__asm volatile (
" vpush {s0} \n" /* Trigger lazy stacking. */
" vpop {s0} \n" /* Nullify the affect of the above instruction. */
::: "memory"
);
}
else
{
/* Standard frame i.e. FPU not in use. */
ulStackFrameSize = 8;
}
Tickless idle fixes/improvement (#59) * Fix tickless idle when stopping systick on zero... ...and don't stop SysTick at all in the eAbortSleep case. Prior to this commit, if vPortSuppressTicksAndSleep() happens to stop the SysTick on zero, then after tickless idle ends, xTickCount advances one full tick more than the time that actually elapsed as measured by the SysTick. See "bug 1" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - Idle task calls vPortSuppressTicksAndSleep(xExpectedIdleTime = 2). [Doesn't have to be "2" -- could be any number.] - vPortSuppressTicksAndSleep() stops SysTick, and the current-count register happens to stop on zero. - SysTick ISR executes, setting xPendedTicks = 1 - vPortSuppressTicksAndSleep() masks interrupts and calls eTaskConfirmSleepModeStatus() which confirms the sleep operation. *** - vPortSuppressTicksAndSleep() configures SysTick for 1 full tick (xExpectedIdleTime - 1) plus the current-count register (which is 0) - One tick period elapses in sleep. - SysTick wakes CPU, ISR executes and increments xPendedTicks to 2. - vPortSuppressTicksAndSleep() calls vTaskStepTick(1), then returns. - Idle task resumes scheduler, which increments xTickCount twice (for xPendedTicks = 2) In the end, two ticks elapsed as measured by SysTick, but the code increments xTickCount three times. The root cause is that the code assumes the SysTick current-count register always contains the number of SysTick counts remaining in the current tick period. However, when the current-count register is zero, there are ulTimerCountsForOneTick counts remaining, not zero. This error is not the kind of time slippage normally associated with tickless idle. *** Note that a recent commit https://github.com/FreeRTOS/FreeRTOS-Kernel/commit/e1b98f0 results in eAbortSleep in this case, due to xPendedTicks != 0. That commit does mostly resolve this bug without specifically mentioning it, and without this commit. But that resolution allows the code in port.c not to directly address the special case of stopping SysTick on zero in any code or comments. That commit also generates additional instances of eAbortSleep, and a second purpose of this commit is to optimize how vPortSuppressTicksAndSleep() behaves for eAbortSleep, as noted below. This commit also includes an optimization to avoid stopping the SysTick when eTaskConfirmSleepModeStatus() returns eAbortSleep. This optimization belongs with this fix because the method of handling the SysTick being stopped on zero changes with this optimization. * Fix imminent tick rescheduled after tickless idle Prior to this commit, if something other than systick wakes the CPU from tickless idle, vPortSuppressTicksAndSleep() might cause xTickCount to increment once too many times. See "bug 2" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - CPU is sleeping in vPortSuppressTicksAndSleep() - Something other than the SysTick wakes the CPU. - vPortSuppressTicksAndSleep() calculates the number of SysTick counts until the next tick. The bug occurs only if this number is small. - vPortSuppressTicksAndSleep() puts this small number into the SysTick reload register, and starts SysTick. - vPortSuppressTicksAndSleep() calls vTaskStepTick() - While vTaskStepTick() executes, the SysTick expires. The ISR pends because interrupts are masked, and SysTick starts a 2nd period still based on the small number of counts in its reload register. This 2nd period is undesirable and is likely to cause the error noted below. - vPortSuppressTicksAndSleep() puts the normal tick duration into the SysTick's reload register. - vPortSuppressTicksAndSleep() unmasks interrupts before the SysTick starts a new period based on the new value in the reload register. [This is a race condition that can go either way, but for the bug to occur, the race must play out this way.] - The pending SysTick ISR executes and increments xPendedTicks. - The SysTick expires again, finishing the second very small period, and starts a new period this time based on the full tick duration. - The SysTick ISR increments xPendedTicks (or xTickCount) even though only a tiny fraction of a tick period has elapsed since the previous tick. The bug occurs when *two* consecutive small periods of the SysTick are both counted as ticks. The root cause is a race caused by the small SysTick period. If vPortSuppressTicksAndSleep() unmasks interrupts *after* the small period expires but *before* the SysTick starts a period based on the full tick period, then two small periods are counted as ticks when only one should be counted. The end result is xTickCount advancing nearly one full tick more than time actually elapsed as measured by the SysTick. This is not the kind of time slippage normally associated with tickless idle. After this commit the code starts the SysTick and then immediately modifies the reload register to ensure the very short cycle (if any) is conducted only once. This strategy requires special consideration for the build option that configures SysTick to use a divided clock. To avoid waiting around for the SysTick to load value from the reload register, the new code temporarily configures the SysTick to use the undivided clock. The resulting timing error is typical for tickless idle. The error (commonly known as drift or slippage in kernel time) caused by this strategy is equivalent to one or two counts in ulStoppedTimerCompensation. This commit also updates comments and #define symbols related to the SysTick clock option. The SysTick can optionally be clocked by a divided version of the CPU clock (commonly divide-by-8). The new code in this commit adjusts these comments and symbols to make them clearer and more useful in configurations that use the divided clock. The fix made in this commit requires the use of these symbols, as noted in the code comments. * Fix tickless idle with alternate systick clocking Prior to this commit, in configurations using the alternate SysTick clocking, vPortSuppressTicksAndSleep() might cause xTickCount to jump ahead as much as the entire expected idle time or fall behind as much as one full tick compared to time as measured by the SysTick. SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. The SysTick has a configuration option to be clocked by an alternate clock besides the core clock. This alternate clock is MCU dependent. Scenarios Fixed --------------- The new code in this commit handles the following scenarios that were not handled correctly prior to this commit. 1. Before the sleep, vPortSuppressTicksAndSleep() stops the SysTick on zero, long after SysTick reached zero. Prior to this commit, this scenario caused xTickCount to jump ahead one full tick for the same reason documented here: https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/0c7b04bd3a745c52151abebc882eed3f811c4c81 2. After the sleep, vPortSuppressTicksAndSleep() stops the SysTick before it loads the counter from the reload register. Prior to this commit, this scenario caused xTickCount to jump ahead by the entire expected idle time (xExpectedIdleTime) because the current-count register is zero before it loads from the reload register. 3. Prior to return, vPortSuppressTicksAndSleep() attempts to start a short SysTick period when the current SysTick clock cycle has a lot of time remaining. Prior to this commit, this scenario could cause xTickCount to fall behind by as much as nearly one full tick because the short SysTick cycle never started. Note that #3 is partially fixed by https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/967acc9b200d3d4beeb289d9da9e88798074b431 even though that commit addresses a different issue. So this commit completes the partial fix. * Improve comments and name of preprocessor symbol Add a note in the code comments that SysTick requests an interrupt when decrementing from 1 to 0, so that's why stopping SysTick on zero is a special case. Readers might unknowingly assume that SysTick requests an interrupt when wrapping from 0 back to the load-register value. Reconsider new "_SETTING" suffix since "_CONFIG" suffix seems more descriptive. The code relies on *both* of these preprocessor symbols: portNVIC_SYSTICK_CLK_BIT portNVIC_SYSTICK_CLK_BIT_CONFIG **new** A meaningful suffix is really helpful to distinguish the two symbols. * Revert introduction of 2nd name for NVIC register When I added portNVIC_ICSR_REG I didn't realize there was already a portNVIC_INT_CTRL_REG, which identifies the same register. Not good to have both. Note that portNVIC_INT_CTRL_REG is defined in portmacro.h and is already used in this file (port.c). * Replicate to other Cortex M ports Also set a new fiddle factor based on tests with a CM4F. I used gcc, optimizing at -O1. Users can fine-tune as needed. Also add configSYSTICK_CLOCK_HZ to the CM0 ports to be just like the other Cortex M ports. This change allowed uniformity in the default tickless implementations across all Cortex M ports. And CM0 is likely to benefit from configSYSTICK_CLOCK_HZ, especially considering new CM0 devices with very fast CPU clock speeds. * Revert changes to IAR-CM0-portmacro.h portNVIC_INT_CTRL_REG was already defined in port.c. No need to define it in portmacro.h. * Handle edge cases with slow SysTick clock Co-authored-by: Cobus van Eeden <35851496+cobusve@users.noreply.github.com> Co-authored-by: abhidixi11 <44424462+abhidixi11@users.noreply.github.com> Co-authored-by: Joseph Julicher <jjulicher@mac.com> Co-authored-by: alfred gedeon <28123637+alfred2g@users.noreply.github.com>
2 years ago
}
#else /* if ( ( configENABLE_FPU == 1 ) || ( configENABLE_MVE == 1 ) ) */
Tickless idle fixes/improvement (#59) * Fix tickless idle when stopping systick on zero... ...and don't stop SysTick at all in the eAbortSleep case. Prior to this commit, if vPortSuppressTicksAndSleep() happens to stop the SysTick on zero, then after tickless idle ends, xTickCount advances one full tick more than the time that actually elapsed as measured by the SysTick. See "bug 1" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - Idle task calls vPortSuppressTicksAndSleep(xExpectedIdleTime = 2). [Doesn't have to be "2" -- could be any number.] - vPortSuppressTicksAndSleep() stops SysTick, and the current-count register happens to stop on zero. - SysTick ISR executes, setting xPendedTicks = 1 - vPortSuppressTicksAndSleep() masks interrupts and calls eTaskConfirmSleepModeStatus() which confirms the sleep operation. *** - vPortSuppressTicksAndSleep() configures SysTick for 1 full tick (xExpectedIdleTime - 1) plus the current-count register (which is 0) - One tick period elapses in sleep. - SysTick wakes CPU, ISR executes and increments xPendedTicks to 2. - vPortSuppressTicksAndSleep() calls vTaskStepTick(1), then returns. - Idle task resumes scheduler, which increments xTickCount twice (for xPendedTicks = 2) In the end, two ticks elapsed as measured by SysTick, but the code increments xTickCount three times. The root cause is that the code assumes the SysTick current-count register always contains the number of SysTick counts remaining in the current tick period. However, when the current-count register is zero, there are ulTimerCountsForOneTick counts remaining, not zero. This error is not the kind of time slippage normally associated with tickless idle. *** Note that a recent commit https://github.com/FreeRTOS/FreeRTOS-Kernel/commit/e1b98f0 results in eAbortSleep in this case, due to xPendedTicks != 0. That commit does mostly resolve this bug without specifically mentioning it, and without this commit. But that resolution allows the code in port.c not to directly address the special case of stopping SysTick on zero in any code or comments. That commit also generates additional instances of eAbortSleep, and a second purpose of this commit is to optimize how vPortSuppressTicksAndSleep() behaves for eAbortSleep, as noted below. This commit also includes an optimization to avoid stopping the SysTick when eTaskConfirmSleepModeStatus() returns eAbortSleep. This optimization belongs with this fix because the method of handling the SysTick being stopped on zero changes with this optimization. * Fix imminent tick rescheduled after tickless idle Prior to this commit, if something other than systick wakes the CPU from tickless idle, vPortSuppressTicksAndSleep() might cause xTickCount to increment once too many times. See "bug 2" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - CPU is sleeping in vPortSuppressTicksAndSleep() - Something other than the SysTick wakes the CPU. - vPortSuppressTicksAndSleep() calculates the number of SysTick counts until the next tick. The bug occurs only if this number is small. - vPortSuppressTicksAndSleep() puts this small number into the SysTick reload register, and starts SysTick. - vPortSuppressTicksAndSleep() calls vTaskStepTick() - While vTaskStepTick() executes, the SysTick expires. The ISR pends because interrupts are masked, and SysTick starts a 2nd period still based on the small number of counts in its reload register. This 2nd period is undesirable and is likely to cause the error noted below. - vPortSuppressTicksAndSleep() puts the normal tick duration into the SysTick's reload register. - vPortSuppressTicksAndSleep() unmasks interrupts before the SysTick starts a new period based on the new value in the reload register. [This is a race condition that can go either way, but for the bug to occur, the race must play out this way.] - The pending SysTick ISR executes and increments xPendedTicks. - The SysTick expires again, finishing the second very small period, and starts a new period this time based on the full tick duration. - The SysTick ISR increments xPendedTicks (or xTickCount) even though only a tiny fraction of a tick period has elapsed since the previous tick. The bug occurs when *two* consecutive small periods of the SysTick are both counted as ticks. The root cause is a race caused by the small SysTick period. If vPortSuppressTicksAndSleep() unmasks interrupts *after* the small period expires but *before* the SysTick starts a period based on the full tick period, then two small periods are counted as ticks when only one should be counted. The end result is xTickCount advancing nearly one full tick more than time actually elapsed as measured by the SysTick. This is not the kind of time slippage normally associated with tickless idle. After this commit the code starts the SysTick and then immediately modifies the reload register to ensure the very short cycle (if any) is conducted only once. This strategy requires special consideration for the build option that configures SysTick to use a divided clock. To avoid waiting around for the SysTick to load value from the reload register, the new code temporarily configures the SysTick to use the undivided clock. The resulting timing error is typical for tickless idle. The error (commonly known as drift or slippage in kernel time) caused by this strategy is equivalent to one or two counts in ulStoppedTimerCompensation. This commit also updates comments and #define symbols related to the SysTick clock option. The SysTick can optionally be clocked by a divided version of the CPU clock (commonly divide-by-8). The new code in this commit adjusts these comments and symbols to make them clearer and more useful in configurations that use the divided clock. The fix made in this commit requires the use of these symbols, as noted in the code comments. * Fix tickless idle with alternate systick clocking Prior to this commit, in configurations using the alternate SysTick clocking, vPortSuppressTicksAndSleep() might cause xTickCount to jump ahead as much as the entire expected idle time or fall behind as much as one full tick compared to time as measured by the SysTick. SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. The SysTick has a configuration option to be clocked by an alternate clock besides the core clock. This alternate clock is MCU dependent. Scenarios Fixed --------------- The new code in this commit handles the following scenarios that were not handled correctly prior to this commit. 1. Before the sleep, vPortSuppressTicksAndSleep() stops the SysTick on zero, long after SysTick reached zero. Prior to this commit, this scenario caused xTickCount to jump ahead one full tick for the same reason documented here: https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/0c7b04bd3a745c52151abebc882eed3f811c4c81 2. After the sleep, vPortSuppressTicksAndSleep() stops the SysTick before it loads the counter from the reload register. Prior to this commit, this scenario caused xTickCount to jump ahead by the entire expected idle time (xExpectedIdleTime) because the current-count register is zero before it loads from the reload register. 3. Prior to return, vPortSuppressTicksAndSleep() attempts to start a short SysTick period when the current SysTick clock cycle has a lot of time remaining. Prior to this commit, this scenario could cause xTickCount to fall behind by as much as nearly one full tick because the short SysTick cycle never started. Note that #3 is partially fixed by https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/967acc9b200d3d4beeb289d9da9e88798074b431 even though that commit addresses a different issue. So this commit completes the partial fix. * Improve comments and name of preprocessor symbol Add a note in the code comments that SysTick requests an interrupt when decrementing from 1 to 0, so that's why stopping SysTick on zero is a special case. Readers might unknowingly assume that SysTick requests an interrupt when wrapping from 0 back to the load-register value. Reconsider new "_SETTING" suffix since "_CONFIG" suffix seems more descriptive. The code relies on *both* of these preprocessor symbols: portNVIC_SYSTICK_CLK_BIT portNVIC_SYSTICK_CLK_BIT_CONFIG **new** A meaningful suffix is really helpful to distinguish the two symbols. * Revert introduction of 2nd name for NVIC register When I added portNVIC_ICSR_REG I didn't realize there was already a portNVIC_INT_CTRL_REG, which identifies the same register. Not good to have both. Note that portNVIC_INT_CTRL_REG is defined in portmacro.h and is already used in this file (port.c). * Replicate to other Cortex M ports Also set a new fiddle factor based on tests with a CM4F. I used gcc, optimizing at -O1. Users can fine-tune as needed. Also add configSYSTICK_CLOCK_HZ to the CM0 ports to be just like the other Cortex M ports. This change allowed uniformity in the default tickless implementations across all Cortex M ports. And CM0 is likely to benefit from configSYSTICK_CLOCK_HZ, especially considering new CM0 devices with very fast CPU clock speeds. * Revert changes to IAR-CM0-portmacro.h portNVIC_INT_CTRL_REG was already defined in port.c. No need to define it in portmacro.h. * Handle edge cases with slow SysTick clock Co-authored-by: Cobus van Eeden <35851496+cobusve@users.noreply.github.com> Co-authored-by: abhidixi11 <44424462+abhidixi11@users.noreply.github.com> Co-authored-by: Joseph Julicher <jjulicher@mac.com> Co-authored-by: alfred gedeon <28123637+alfred2g@users.noreply.github.com>
2 years ago
{
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
ulStackFrameSize = 8;
Tickless idle fixes/improvement (#59) * Fix tickless idle when stopping systick on zero... ...and don't stop SysTick at all in the eAbortSleep case. Prior to this commit, if vPortSuppressTicksAndSleep() happens to stop the SysTick on zero, then after tickless idle ends, xTickCount advances one full tick more than the time that actually elapsed as measured by the SysTick. See "bug 1" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - Idle task calls vPortSuppressTicksAndSleep(xExpectedIdleTime = 2). [Doesn't have to be "2" -- could be any number.] - vPortSuppressTicksAndSleep() stops SysTick, and the current-count register happens to stop on zero. - SysTick ISR executes, setting xPendedTicks = 1 - vPortSuppressTicksAndSleep() masks interrupts and calls eTaskConfirmSleepModeStatus() which confirms the sleep operation. *** - vPortSuppressTicksAndSleep() configures SysTick for 1 full tick (xExpectedIdleTime - 1) plus the current-count register (which is 0) - One tick period elapses in sleep. - SysTick wakes CPU, ISR executes and increments xPendedTicks to 2. - vPortSuppressTicksAndSleep() calls vTaskStepTick(1), then returns. - Idle task resumes scheduler, which increments xTickCount twice (for xPendedTicks = 2) In the end, two ticks elapsed as measured by SysTick, but the code increments xTickCount three times. The root cause is that the code assumes the SysTick current-count register always contains the number of SysTick counts remaining in the current tick period. However, when the current-count register is zero, there are ulTimerCountsForOneTick counts remaining, not zero. This error is not the kind of time slippage normally associated with tickless idle. *** Note that a recent commit https://github.com/FreeRTOS/FreeRTOS-Kernel/commit/e1b98f0 results in eAbortSleep in this case, due to xPendedTicks != 0. That commit does mostly resolve this bug without specifically mentioning it, and without this commit. But that resolution allows the code in port.c not to directly address the special case of stopping SysTick on zero in any code or comments. That commit also generates additional instances of eAbortSleep, and a second purpose of this commit is to optimize how vPortSuppressTicksAndSleep() behaves for eAbortSleep, as noted below. This commit also includes an optimization to avoid stopping the SysTick when eTaskConfirmSleepModeStatus() returns eAbortSleep. This optimization belongs with this fix because the method of handling the SysTick being stopped on zero changes with this optimization. * Fix imminent tick rescheduled after tickless idle Prior to this commit, if something other than systick wakes the CPU from tickless idle, vPortSuppressTicksAndSleep() might cause xTickCount to increment once too many times. See "bug 2" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - CPU is sleeping in vPortSuppressTicksAndSleep() - Something other than the SysTick wakes the CPU. - vPortSuppressTicksAndSleep() calculates the number of SysTick counts until the next tick. The bug occurs only if this number is small. - vPortSuppressTicksAndSleep() puts this small number into the SysTick reload register, and starts SysTick. - vPortSuppressTicksAndSleep() calls vTaskStepTick() - While vTaskStepTick() executes, the SysTick expires. The ISR pends because interrupts are masked, and SysTick starts a 2nd period still based on the small number of counts in its reload register. This 2nd period is undesirable and is likely to cause the error noted below. - vPortSuppressTicksAndSleep() puts the normal tick duration into the SysTick's reload register. - vPortSuppressTicksAndSleep() unmasks interrupts before the SysTick starts a new period based on the new value in the reload register. [This is a race condition that can go either way, but for the bug to occur, the race must play out this way.] - The pending SysTick ISR executes and increments xPendedTicks. - The SysTick expires again, finishing the second very small period, and starts a new period this time based on the full tick duration. - The SysTick ISR increments xPendedTicks (or xTickCount) even though only a tiny fraction of a tick period has elapsed since the previous tick. The bug occurs when *two* consecutive small periods of the SysTick are both counted as ticks. The root cause is a race caused by the small SysTick period. If vPortSuppressTicksAndSleep() unmasks interrupts *after* the small period expires but *before* the SysTick starts a period based on the full tick period, then two small periods are counted as ticks when only one should be counted. The end result is xTickCount advancing nearly one full tick more than time actually elapsed as measured by the SysTick. This is not the kind of time slippage normally associated with tickless idle. After this commit the code starts the SysTick and then immediately modifies the reload register to ensure the very short cycle (if any) is conducted only once. This strategy requires special consideration for the build option that configures SysTick to use a divided clock. To avoid waiting around for the SysTick to load value from the reload register, the new code temporarily configures the SysTick to use the undivided clock. The resulting timing error is typical for tickless idle. The error (commonly known as drift or slippage in kernel time) caused by this strategy is equivalent to one or two counts in ulStoppedTimerCompensation. This commit also updates comments and #define symbols related to the SysTick clock option. The SysTick can optionally be clocked by a divided version of the CPU clock (commonly divide-by-8). The new code in this commit adjusts these comments and symbols to make them clearer and more useful in configurations that use the divided clock. The fix made in this commit requires the use of these symbols, as noted in the code comments. * Fix tickless idle with alternate systick clocking Prior to this commit, in configurations using the alternate SysTick clocking, vPortSuppressTicksAndSleep() might cause xTickCount to jump ahead as much as the entire expected idle time or fall behind as much as one full tick compared to time as measured by the SysTick. SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. The SysTick has a configuration option to be clocked by an alternate clock besides the core clock. This alternate clock is MCU dependent. Scenarios Fixed --------------- The new code in this commit handles the following scenarios that were not handled correctly prior to this commit. 1. Before the sleep, vPortSuppressTicksAndSleep() stops the SysTick on zero, long after SysTick reached zero. Prior to this commit, this scenario caused xTickCount to jump ahead one full tick for the same reason documented here: https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/0c7b04bd3a745c52151abebc882eed3f811c4c81 2. After the sleep, vPortSuppressTicksAndSleep() stops the SysTick before it loads the counter from the reload register. Prior to this commit, this scenario caused xTickCount to jump ahead by the entire expected idle time (xExpectedIdleTime) because the current-count register is zero before it loads from the reload register. 3. Prior to return, vPortSuppressTicksAndSleep() attempts to start a short SysTick period when the current SysTick clock cycle has a lot of time remaining. Prior to this commit, this scenario could cause xTickCount to fall behind by as much as nearly one full tick because the short SysTick cycle never started. Note that #3 is partially fixed by https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/967acc9b200d3d4beeb289d9da9e88798074b431 even though that commit addresses a different issue. So this commit completes the partial fix. * Improve comments and name of preprocessor symbol Add a note in the code comments that SysTick requests an interrupt when decrementing from 1 to 0, so that's why stopping SysTick on zero is a special case. Readers might unknowingly assume that SysTick requests an interrupt when wrapping from 0 back to the load-register value. Reconsider new "_SETTING" suffix since "_CONFIG" suffix seems more descriptive. The code relies on *both* of these preprocessor symbols: portNVIC_SYSTICK_CLK_BIT portNVIC_SYSTICK_CLK_BIT_CONFIG **new** A meaningful suffix is really helpful to distinguish the two symbols. * Revert introduction of 2nd name for NVIC register When I added portNVIC_ICSR_REG I didn't realize there was already a portNVIC_INT_CTRL_REG, which identifies the same register. Not good to have both. Note that portNVIC_INT_CTRL_REG is defined in portmacro.h and is already used in this file (port.c). * Replicate to other Cortex M ports Also set a new fiddle factor based on tests with a CM4F. I used gcc, optimizing at -O1. Users can fine-tune as needed. Also add configSYSTICK_CLOCK_HZ to the CM0 ports to be just like the other Cortex M ports. This change allowed uniformity in the default tickless implementations across all Cortex M ports. And CM0 is likely to benefit from configSYSTICK_CLOCK_HZ, especially considering new CM0 devices with very fast CPU clock speeds. * Revert changes to IAR-CM0-portmacro.h portNVIC_INT_CTRL_REG was already defined in port.c. No need to define it in portmacro.h. * Handle edge cases with slow SysTick clock Co-authored-by: Cobus van Eeden <35851496+cobusve@users.noreply.github.com> Co-authored-by: abhidixi11 <44424462+abhidixi11@users.noreply.github.com> Co-authored-by: Joseph Julicher <jjulicher@mac.com> Co-authored-by: alfred gedeon <28123637+alfred2g@users.noreply.github.com>
2 years ago
}
#endif /* configENABLE_FPU || configENABLE_MVE */
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
/* Make space on the system call stack for the stack frame and
* the parameter passed on the stack. We only need to copy one
* parameter but we still reserve 2 spaces to keep the stack
* double word aligned. */
pulSystemCallStack = pulSystemCallStack - ulStackFrameSize - 2UL;
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
/* Copy the stack frame. */
for( i = 0; i < ulStackFrameSize; i++ )
{
pulSystemCallStack[ i ] = pulTaskStack[ i ];
}
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
/* Copy the parameter which is passed the stack. */
if( ( pulTaskStack[ portOFFSET_TO_PSR ] & portPSR_STACK_PADDING_MASK ) == portPSR_STACK_PADDING_MASK )
{
pulSystemCallStack[ ulStackFrameSize ] = pulTaskStack[ ulStackFrameSize + 1 ];
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
/* Record if the hardware used padding to force the stack pointer
* to be double word aligned. */
pxMpuSettings->ulTaskFlags |= portSTACK_FRAME_HAS_PADDING_FLAG;
}
else
{
pulSystemCallStack[ ulStackFrameSize ] = pulTaskStack[ ulStackFrameSize ];
/* Record if the hardware used padding to force the stack pointer
* to be double word aligned. */
pxMpuSettings->ulTaskFlags &= ( ~portSTACK_FRAME_HAS_PADDING_FLAG );
}
/* Store the value of the LR and PSPLIM registers before the SVC was raised.
* We need to restore it when we exit from the system call. */
pxMpuSettings->xSystemCallStackInfo.ulLinkRegisterAtSystemCallEntry = pulTaskStack[ portOFFSET_TO_LR ];
#if ( portUSE_PSPLIM_REGISTER == 1 )
{
__asm volatile ( "mrs %0, psplim" : "=r" ( pxMpuSettings->xSystemCallStackInfo.ulStackLimitRegisterAtSystemCallEntry ) );
}
#endif
/* Use the pulSystemCallStack in thread mode. */
__asm volatile ( "msr psp, %0" : : "r" ( pulSystemCallStack ) );
#if ( portUSE_PSPLIM_REGISTER == 1 )
{
__asm volatile ( "msr psplim, %0" : : "r" ( pxMpuSettings->xSystemCallStackInfo.pulSystemCallStackLimit ) );
}
#endif
/* Remember the location where we should copy the stack frame when we exit from
* the system call. */
pxMpuSettings->xSystemCallStackInfo.pulTaskStack = pulTaskStack + ulStackFrameSize;
/* We ensure in pxPortInitialiseStack that the system call stack is
* double word aligned and therefore, there is no need of padding.
* Clear the bit[9] of stacked xPSR. */
pulSystemCallStack[ portOFFSET_TO_PSR ] &= ( ~portPSR_STACK_PADDING_MASK );
/* Raise the privilege for the duration of the system call. */
__asm volatile (
" mrs r0, control \n" /* Obtain current control value. */
" movs r1, #1 \n" /* r1 = 1. */
" bics r0, r1 \n" /* Clear nPRIV bit. */
" msr control, r0 \n" /* Write back new control value. */
::: "r0", "r1", "memory"
);
}
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
}
#endif /* ( configENABLE_MPU == 1 ) && ( configUSE_MPU_WRAPPERS_V1 == 0 ) */
/*-----------------------------------------------------------*/
#if ( ( configENABLE_MPU == 1 ) && ( configUSE_MPU_WRAPPERS_V1 == 0 ) )
void vSystemCallExit( uint32_t * pulSystemCallStack,
uint32_t ulLR ) /* PRIVILEGED_FUNCTION */
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
{
extern TaskHandle_t pxCurrentTCB;
xMPU_SETTINGS * pxMpuSettings;
uint32_t * pulTaskStack;
uint32_t ulStackFrameSize, ulSystemCallLocation, i;
#if defined( __ARMCC_VERSION )
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
/* Declaration when these variable are defined in code instead of being
* exported from linker scripts. */
extern uint32_t * __syscalls_flash_start__;
extern uint32_t * __syscalls_flash_end__;
#else
/* Declaration when these variable are exported from linker scripts. */
extern uint32_t __syscalls_flash_start__[];
extern uint32_t __syscalls_flash_end__[];
#endif /* #if defined( __ARMCC_VERSION ) */
ulSystemCallLocation = pulSystemCallStack[ portOFFSET_TO_PC ];
/* If the request did not come from the system call section, do nothing. */
if( ( ulSystemCallLocation >= ( uint32_t ) __syscalls_flash_start__ ) &&
( ulSystemCallLocation <= ( uint32_t ) __syscalls_flash_end__ ) )
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
{
pxMpuSettings = xTaskGetMPUSettings( pxCurrentTCB );
pulTaskStack = pxMpuSettings->xSystemCallStackInfo.pulTaskStack;
#if ( ( configENABLE_FPU == 1 ) || ( configENABLE_MVE == 1 ) )
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
{
if( ( ulLR & portEXC_RETURN_STACK_FRAME_TYPE_MASK ) == 0UL )
{
/* Extended frame i.e. FPU in use. */
ulStackFrameSize = 26;
__asm volatile (
" vpush {s0} \n" /* Trigger lazy stacking. */
" vpop {s0} \n" /* Nullify the affect of the above instruction. */
::: "memory"
);
}
else
{
/* Standard frame i.e. FPU not in use. */
ulStackFrameSize = 8;
}
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
}
#else /* if ( ( configENABLE_FPU == 1 ) || ( configENABLE_MVE == 1 ) ) */
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
{
ulStackFrameSize = 8;
}
#endif /* configENABLE_FPU || configENABLE_MVE */
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
/* Make space on the task stack for the stack frame. */
pulTaskStack = pulTaskStack - ulStackFrameSize;
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
/* Copy the stack frame. */
for( i = 0; i < ulStackFrameSize; i++ )
{
pulTaskStack[ i ] = pulSystemCallStack[ i ];
}
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
/* Use the pulTaskStack in thread mode. */
__asm volatile ( "msr psp, %0" : : "r" ( pulTaskStack ) );
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
/* Restore the LR and PSPLIM to what they were at the time of
* system call entry. */
pulTaskStack[ portOFFSET_TO_LR ] = pxMpuSettings->xSystemCallStackInfo.ulLinkRegisterAtSystemCallEntry;
#if ( portUSE_PSPLIM_REGISTER == 1 )
{
__asm volatile ( "msr psplim, %0" : : "r" ( pxMpuSettings->xSystemCallStackInfo.ulStackLimitRegisterAtSystemCallEntry ) );
}
#endif
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
/* If the hardware used padding to force the stack pointer
* to be double word aligned, set the stacked xPSR bit[9],
* otherwise clear it. */
if( ( pxMpuSettings->ulTaskFlags & portSTACK_FRAME_HAS_PADDING_FLAG ) == portSTACK_FRAME_HAS_PADDING_FLAG )
{
pulTaskStack[ portOFFSET_TO_PSR ] |= portPSR_STACK_PADDING_MASK;
}
else
{
pulTaskStack[ portOFFSET_TO_PSR ] &= ( ~portPSR_STACK_PADDING_MASK );
}
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
/* This is not NULL only for the duration of the system call. */
pxMpuSettings->xSystemCallStackInfo.pulTaskStack = NULL;
/* Drop the privilege before returning to the thread mode. */
__asm volatile (
" mrs r0, control \n" /* Obtain current control value. */
" movs r1, #1 \n" /* r1 = 1. */
" orrs r0, r1 \n" /* Set nPRIV bit. */
" msr control, r0 \n" /* Write back new control value. */
::: "r0", "r1", "memory"
);
}
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
}
#endif /* ( configENABLE_MPU == 1 ) && ( configUSE_MPU_WRAPPERS_V1 == 0 ) */
/*-----------------------------------------------------------*/
#if ( configENABLE_MPU == 1 )
BaseType_t xPortIsTaskPrivileged( void ) /* PRIVILEGED_FUNCTION */
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
{
BaseType_t xTaskIsPrivileged = pdFALSE;
const xMPU_SETTINGS * xTaskMpuSettings = xTaskGetMPUSettings( NULL ); /* Calling task's MPU settings. */
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
if( ( xTaskMpuSettings->ulTaskFlags & portTASK_IS_PRIVILEGED_FLAG ) == portTASK_IS_PRIVILEGED_FLAG )
{
xTaskIsPrivileged = pdTRUE;
}
return xTaskIsPrivileged;
}
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
#endif /* configENABLE_MPU == 1 */
/*-----------------------------------------------------------*/
#if ( configENABLE_MPU == 1 )
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
StackType_t * pxPortInitialiseStack( StackType_t * pxTopOfStack,
StackType_t * pxEndOfStack,
TaskFunction_t pxCode,
void * pvParameters,
BaseType_t xRunPrivileged,
xMPU_SETTINGS * xMPUSettings ) /* PRIVILEGED_FUNCTION */
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
{
uint32_t ulIndex = 0;
xMPUSettings->ulContext[ ulIndex ] = 0x04040404; /* r4. */
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
ulIndex++;
xMPUSettings->ulContext[ ulIndex ] = 0x05050505; /* r5. */
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
ulIndex++;
xMPUSettings->ulContext[ ulIndex ] = 0x06060606; /* r6. */
ulIndex++;
xMPUSettings->ulContext[ ulIndex ] = 0x07070707; /* r7. */
ulIndex++;
xMPUSettings->ulContext[ ulIndex ] = 0x08080808; /* r8. */
ulIndex++;
xMPUSettings->ulContext[ ulIndex ] = 0x09090909; /* r9. */
ulIndex++;
xMPUSettings->ulContext[ ulIndex ] = 0x10101010; /* r10. */
ulIndex++;
xMPUSettings->ulContext[ ulIndex ] = 0x11111111; /* r11. */
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
ulIndex++;
xMPUSettings->ulContext[ ulIndex ] = ( uint32_t ) pvParameters; /* r0. */
ulIndex++;
xMPUSettings->ulContext[ ulIndex ] = 0x01010101; /* r1. */
ulIndex++;
xMPUSettings->ulContext[ ulIndex ] = 0x02020202; /* r2. */
ulIndex++;
xMPUSettings->ulContext[ ulIndex ] = 0x03030303; /* r3. */
ulIndex++;
xMPUSettings->ulContext[ ulIndex ] = 0x12121212; /* r12. */
ulIndex++;
xMPUSettings->ulContext[ ulIndex ] = ( uint32_t ) portTASK_RETURN_ADDRESS; /* LR. */
ulIndex++;
xMPUSettings->ulContext[ ulIndex ] = ( uint32_t ) pxCode; /* PC. */
ulIndex++;
xMPUSettings->ulContext[ ulIndex ] = portINITIAL_XPSR; /* xPSR. */
ulIndex++;
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
#if ( configENABLE_TRUSTZONE == 1 )
{
xMPUSettings->ulContext[ ulIndex ] = portNO_SECURE_CONTEXT; /* xSecureContext. */
ulIndex++;
}
#endif /* configENABLE_TRUSTZONE */
xMPUSettings->ulContext[ ulIndex ] = ( uint32_t ) ( pxTopOfStack - 8 ); /* PSP with the hardware saved stack. */
ulIndex++;
xMPUSettings->ulContext[ ulIndex ] = ( uint32_t ) pxEndOfStack; /* PSPLIM. */
ulIndex++;
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
if( xRunPrivileged == pdTRUE )
{
xMPUSettings->ulTaskFlags |= portTASK_IS_PRIVILEGED_FLAG;
xMPUSettings->ulContext[ ulIndex ] = ( uint32_t ) portINITIAL_CONTROL_PRIVILEGED; /* CONTROL. */
ulIndex++;
}
else
{
xMPUSettings->ulTaskFlags &= ( ~portTASK_IS_PRIVILEGED_FLAG );
xMPUSettings->ulContext[ ulIndex ] = ( uint32_t ) portINITIAL_CONTROL_UNPRIVILEGED; /* CONTROL. */
ulIndex++;
}
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
xMPUSettings->ulContext[ ulIndex ] = portINITIAL_EXC_RETURN; /* LR (EXC_RETURN). */
ulIndex++;
#if ( configUSE_MPU_WRAPPERS_V1 == 0 )
Tickless idle fixes/improvement (#59) * Fix tickless idle when stopping systick on zero... ...and don't stop SysTick at all in the eAbortSleep case. Prior to this commit, if vPortSuppressTicksAndSleep() happens to stop the SysTick on zero, then after tickless idle ends, xTickCount advances one full tick more than the time that actually elapsed as measured by the SysTick. See "bug 1" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - Idle task calls vPortSuppressTicksAndSleep(xExpectedIdleTime = 2). [Doesn't have to be "2" -- could be any number.] - vPortSuppressTicksAndSleep() stops SysTick, and the current-count register happens to stop on zero. - SysTick ISR executes, setting xPendedTicks = 1 - vPortSuppressTicksAndSleep() masks interrupts and calls eTaskConfirmSleepModeStatus() which confirms the sleep operation. *** - vPortSuppressTicksAndSleep() configures SysTick for 1 full tick (xExpectedIdleTime - 1) plus the current-count register (which is 0) - One tick period elapses in sleep. - SysTick wakes CPU, ISR executes and increments xPendedTicks to 2. - vPortSuppressTicksAndSleep() calls vTaskStepTick(1), then returns. - Idle task resumes scheduler, which increments xTickCount twice (for xPendedTicks = 2) In the end, two ticks elapsed as measured by SysTick, but the code increments xTickCount three times. The root cause is that the code assumes the SysTick current-count register always contains the number of SysTick counts remaining in the current tick period. However, when the current-count register is zero, there are ulTimerCountsForOneTick counts remaining, not zero. This error is not the kind of time slippage normally associated with tickless idle. *** Note that a recent commit https://github.com/FreeRTOS/FreeRTOS-Kernel/commit/e1b98f0 results in eAbortSleep in this case, due to xPendedTicks != 0. That commit does mostly resolve this bug without specifically mentioning it, and without this commit. But that resolution allows the code in port.c not to directly address the special case of stopping SysTick on zero in any code or comments. That commit also generates additional instances of eAbortSleep, and a second purpose of this commit is to optimize how vPortSuppressTicksAndSleep() behaves for eAbortSleep, as noted below. This commit also includes an optimization to avoid stopping the SysTick when eTaskConfirmSleepModeStatus() returns eAbortSleep. This optimization belongs with this fix because the method of handling the SysTick being stopped on zero changes with this optimization. * Fix imminent tick rescheduled after tickless idle Prior to this commit, if something other than systick wakes the CPU from tickless idle, vPortSuppressTicksAndSleep() might cause xTickCount to increment once too many times. See "bug 2" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - CPU is sleeping in vPortSuppressTicksAndSleep() - Something other than the SysTick wakes the CPU. - vPortSuppressTicksAndSleep() calculates the number of SysTick counts until the next tick. The bug occurs only if this number is small. - vPortSuppressTicksAndSleep() puts this small number into the SysTick reload register, and starts SysTick. - vPortSuppressTicksAndSleep() calls vTaskStepTick() - While vTaskStepTick() executes, the SysTick expires. The ISR pends because interrupts are masked, and SysTick starts a 2nd period still based on the small number of counts in its reload register. This 2nd period is undesirable and is likely to cause the error noted below. - vPortSuppressTicksAndSleep() puts the normal tick duration into the SysTick's reload register. - vPortSuppressTicksAndSleep() unmasks interrupts before the SysTick starts a new period based on the new value in the reload register. [This is a race condition that can go either way, but for the bug to occur, the race must play out this way.] - The pending SysTick ISR executes and increments xPendedTicks. - The SysTick expires again, finishing the second very small period, and starts a new period this time based on the full tick duration. - The SysTick ISR increments xPendedTicks (or xTickCount) even though only a tiny fraction of a tick period has elapsed since the previous tick. The bug occurs when *two* consecutive small periods of the SysTick are both counted as ticks. The root cause is a race caused by the small SysTick period. If vPortSuppressTicksAndSleep() unmasks interrupts *after* the small period expires but *before* the SysTick starts a period based on the full tick period, then two small periods are counted as ticks when only one should be counted. The end result is xTickCount advancing nearly one full tick more than time actually elapsed as measured by the SysTick. This is not the kind of time slippage normally associated with tickless idle. After this commit the code starts the SysTick and then immediately modifies the reload register to ensure the very short cycle (if any) is conducted only once. This strategy requires special consideration for the build option that configures SysTick to use a divided clock. To avoid waiting around for the SysTick to load value from the reload register, the new code temporarily configures the SysTick to use the undivided clock. The resulting timing error is typical for tickless idle. The error (commonly known as drift or slippage in kernel time) caused by this strategy is equivalent to one or two counts in ulStoppedTimerCompensation. This commit also updates comments and #define symbols related to the SysTick clock option. The SysTick can optionally be clocked by a divided version of the CPU clock (commonly divide-by-8). The new code in this commit adjusts these comments and symbols to make them clearer and more useful in configurations that use the divided clock. The fix made in this commit requires the use of these symbols, as noted in the code comments. * Fix tickless idle with alternate systick clocking Prior to this commit, in configurations using the alternate SysTick clocking, vPortSuppressTicksAndSleep() might cause xTickCount to jump ahead as much as the entire expected idle time or fall behind as much as one full tick compared to time as measured by the SysTick. SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. The SysTick has a configuration option to be clocked by an alternate clock besides the core clock. This alternate clock is MCU dependent. Scenarios Fixed --------------- The new code in this commit handles the following scenarios that were not handled correctly prior to this commit. 1. Before the sleep, vPortSuppressTicksAndSleep() stops the SysTick on zero, long after SysTick reached zero. Prior to this commit, this scenario caused xTickCount to jump ahead one full tick for the same reason documented here: https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/0c7b04bd3a745c52151abebc882eed3f811c4c81 2. After the sleep, vPortSuppressTicksAndSleep() stops the SysTick before it loads the counter from the reload register. Prior to this commit, this scenario caused xTickCount to jump ahead by the entire expected idle time (xExpectedIdleTime) because the current-count register is zero before it loads from the reload register. 3. Prior to return, vPortSuppressTicksAndSleep() attempts to start a short SysTick period when the current SysTick clock cycle has a lot of time remaining. Prior to this commit, this scenario could cause xTickCount to fall behind by as much as nearly one full tick because the short SysTick cycle never started. Note that #3 is partially fixed by https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/967acc9b200d3d4beeb289d9da9e88798074b431 even though that commit addresses a different issue. So this commit completes the partial fix. * Improve comments and name of preprocessor symbol Add a note in the code comments that SysTick requests an interrupt when decrementing from 1 to 0, so that's why stopping SysTick on zero is a special case. Readers might unknowingly assume that SysTick requests an interrupt when wrapping from 0 back to the load-register value. Reconsider new "_SETTING" suffix since "_CONFIG" suffix seems more descriptive. The code relies on *both* of these preprocessor symbols: portNVIC_SYSTICK_CLK_BIT portNVIC_SYSTICK_CLK_BIT_CONFIG **new** A meaningful suffix is really helpful to distinguish the two symbols. * Revert introduction of 2nd name for NVIC register When I added portNVIC_ICSR_REG I didn't realize there was already a portNVIC_INT_CTRL_REG, which identifies the same register. Not good to have both. Note that portNVIC_INT_CTRL_REG is defined in portmacro.h and is already used in this file (port.c). * Replicate to other Cortex M ports Also set a new fiddle factor based on tests with a CM4F. I used gcc, optimizing at -O1. Users can fine-tune as needed. Also add configSYSTICK_CLOCK_HZ to the CM0 ports to be just like the other Cortex M ports. This change allowed uniformity in the default tickless implementations across all Cortex M ports. And CM0 is likely to benefit from configSYSTICK_CLOCK_HZ, especially considering new CM0 devices with very fast CPU clock speeds. * Revert changes to IAR-CM0-portmacro.h portNVIC_INT_CTRL_REG was already defined in port.c. No need to define it in portmacro.h. * Handle edge cases with slow SysTick clock Co-authored-by: Cobus van Eeden <35851496+cobusve@users.noreply.github.com> Co-authored-by: abhidixi11 <44424462+abhidixi11@users.noreply.github.com> Co-authored-by: Joseph Julicher <jjulicher@mac.com> Co-authored-by: alfred gedeon <28123637+alfred2g@users.noreply.github.com>
2 years ago
{
/* Ensure that the system call stack is double word aligned. */
xMPUSettings->xSystemCallStackInfo.pulSystemCallStack = &( xMPUSettings->xSystemCallStackInfo.ulSystemCallStackBuffer[ configSYSTEM_CALL_STACK_SIZE - 1 ] );
xMPUSettings->xSystemCallStackInfo.pulSystemCallStack = ( uint32_t * ) ( ( uint32_t ) ( xMPUSettings->xSystemCallStackInfo.pulSystemCallStack ) &
( uint32_t ) ( ~( portBYTE_ALIGNMENT_MASK ) ) );
xMPUSettings->xSystemCallStackInfo.pulSystemCallStackLimit = &( xMPUSettings->xSystemCallStackInfo.ulSystemCallStackBuffer[ 0 ] );
xMPUSettings->xSystemCallStackInfo.pulSystemCallStackLimit = ( uint32_t * ) ( ( ( uint32_t ) ( xMPUSettings->xSystemCallStackInfo.pulSystemCallStackLimit ) +
( uint32_t ) ( portBYTE_ALIGNMENT - 1 ) ) &
( uint32_t ) ( ~( portBYTE_ALIGNMENT_MASK ) ) );
/* This is not NULL only for the duration of a system call. */
xMPUSettings->xSystemCallStackInfo.pulTaskStack = NULL;
}
#endif /* configUSE_MPU_WRAPPERS_V1 == 0 */
return &( xMPUSettings->ulContext[ ulIndex ] );
Tickless idle fixes/improvement (#59) * Fix tickless idle when stopping systick on zero... ...and don't stop SysTick at all in the eAbortSleep case. Prior to this commit, if vPortSuppressTicksAndSleep() happens to stop the SysTick on zero, then after tickless idle ends, xTickCount advances one full tick more than the time that actually elapsed as measured by the SysTick. See "bug 1" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - Idle task calls vPortSuppressTicksAndSleep(xExpectedIdleTime = 2). [Doesn't have to be "2" -- could be any number.] - vPortSuppressTicksAndSleep() stops SysTick, and the current-count register happens to stop on zero. - SysTick ISR executes, setting xPendedTicks = 1 - vPortSuppressTicksAndSleep() masks interrupts and calls eTaskConfirmSleepModeStatus() which confirms the sleep operation. *** - vPortSuppressTicksAndSleep() configures SysTick for 1 full tick (xExpectedIdleTime - 1) plus the current-count register (which is 0) - One tick period elapses in sleep. - SysTick wakes CPU, ISR executes and increments xPendedTicks to 2. - vPortSuppressTicksAndSleep() calls vTaskStepTick(1), then returns. - Idle task resumes scheduler, which increments xTickCount twice (for xPendedTicks = 2) In the end, two ticks elapsed as measured by SysTick, but the code increments xTickCount three times. The root cause is that the code assumes the SysTick current-count register always contains the number of SysTick counts remaining in the current tick period. However, when the current-count register is zero, there are ulTimerCountsForOneTick counts remaining, not zero. This error is not the kind of time slippage normally associated with tickless idle. *** Note that a recent commit https://github.com/FreeRTOS/FreeRTOS-Kernel/commit/e1b98f0 results in eAbortSleep in this case, due to xPendedTicks != 0. That commit does mostly resolve this bug without specifically mentioning it, and without this commit. But that resolution allows the code in port.c not to directly address the special case of stopping SysTick on zero in any code or comments. That commit also generates additional instances of eAbortSleep, and a second purpose of this commit is to optimize how vPortSuppressTicksAndSleep() behaves for eAbortSleep, as noted below. This commit also includes an optimization to avoid stopping the SysTick when eTaskConfirmSleepModeStatus() returns eAbortSleep. This optimization belongs with this fix because the method of handling the SysTick being stopped on zero changes with this optimization. * Fix imminent tick rescheduled after tickless idle Prior to this commit, if something other than systick wakes the CPU from tickless idle, vPortSuppressTicksAndSleep() might cause xTickCount to increment once too many times. See "bug 2" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - CPU is sleeping in vPortSuppressTicksAndSleep() - Something other than the SysTick wakes the CPU. - vPortSuppressTicksAndSleep() calculates the number of SysTick counts until the next tick. The bug occurs only if this number is small. - vPortSuppressTicksAndSleep() puts this small number into the SysTick reload register, and starts SysTick. - vPortSuppressTicksAndSleep() calls vTaskStepTick() - While vTaskStepTick() executes, the SysTick expires. The ISR pends because interrupts are masked, and SysTick starts a 2nd period still based on the small number of counts in its reload register. This 2nd period is undesirable and is likely to cause the error noted below. - vPortSuppressTicksAndSleep() puts the normal tick duration into the SysTick's reload register. - vPortSuppressTicksAndSleep() unmasks interrupts before the SysTick starts a new period based on the new value in the reload register. [This is a race condition that can go either way, but for the bug to occur, the race must play out this way.] - The pending SysTick ISR executes and increments xPendedTicks. - The SysTick expires again, finishing the second very small period, and starts a new period this time based on the full tick duration. - The SysTick ISR increments xPendedTicks (or xTickCount) even though only a tiny fraction of a tick period has elapsed since the previous tick. The bug occurs when *two* consecutive small periods of the SysTick are both counted as ticks. The root cause is a race caused by the small SysTick period. If vPortSuppressTicksAndSleep() unmasks interrupts *after* the small period expires but *before* the SysTick starts a period based on the full tick period, then two small periods are counted as ticks when only one should be counted. The end result is xTickCount advancing nearly one full tick more than time actually elapsed as measured by the SysTick. This is not the kind of time slippage normally associated with tickless idle. After this commit the code starts the SysTick and then immediately modifies the reload register to ensure the very short cycle (if any) is conducted only once. This strategy requires special consideration for the build option that configures SysTick to use a divided clock. To avoid waiting around for the SysTick to load value from the reload register, the new code temporarily configures the SysTick to use the undivided clock. The resulting timing error is typical for tickless idle. The error (commonly known as drift or slippage in kernel time) caused by this strategy is equivalent to one or two counts in ulStoppedTimerCompensation. This commit also updates comments and #define symbols related to the SysTick clock option. The SysTick can optionally be clocked by a divided version of the CPU clock (commonly divide-by-8). The new code in this commit adjusts these comments and symbols to make them clearer and more useful in configurations that use the divided clock. The fix made in this commit requires the use of these symbols, as noted in the code comments. * Fix tickless idle with alternate systick clocking Prior to this commit, in configurations using the alternate SysTick clocking, vPortSuppressTicksAndSleep() might cause xTickCount to jump ahead as much as the entire expected idle time or fall behind as much as one full tick compared to time as measured by the SysTick. SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. The SysTick has a configuration option to be clocked by an alternate clock besides the core clock. This alternate clock is MCU dependent. Scenarios Fixed --------------- The new code in this commit handles the following scenarios that were not handled correctly prior to this commit. 1. Before the sleep, vPortSuppressTicksAndSleep() stops the SysTick on zero, long after SysTick reached zero. Prior to this commit, this scenario caused xTickCount to jump ahead one full tick for the same reason documented here: https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/0c7b04bd3a745c52151abebc882eed3f811c4c81 2. After the sleep, vPortSuppressTicksAndSleep() stops the SysTick before it loads the counter from the reload register. Prior to this commit, this scenario caused xTickCount to jump ahead by the entire expected idle time (xExpectedIdleTime) because the current-count register is zero before it loads from the reload register. 3. Prior to return, vPortSuppressTicksAndSleep() attempts to start a short SysTick period when the current SysTick clock cycle has a lot of time remaining. Prior to this commit, this scenario could cause xTickCount to fall behind by as much as nearly one full tick because the short SysTick cycle never started. Note that #3 is partially fixed by https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/967acc9b200d3d4beeb289d9da9e88798074b431 even though that commit addresses a different issue. So this commit completes the partial fix. * Improve comments and name of preprocessor symbol Add a note in the code comments that SysTick requests an interrupt when decrementing from 1 to 0, so that's why stopping SysTick on zero is a special case. Readers might unknowingly assume that SysTick requests an interrupt when wrapping from 0 back to the load-register value. Reconsider new "_SETTING" suffix since "_CONFIG" suffix seems more descriptive. The code relies on *both* of these preprocessor symbols: portNVIC_SYSTICK_CLK_BIT portNVIC_SYSTICK_CLK_BIT_CONFIG **new** A meaningful suffix is really helpful to distinguish the two symbols. * Revert introduction of 2nd name for NVIC register When I added portNVIC_ICSR_REG I didn't realize there was already a portNVIC_INT_CTRL_REG, which identifies the same register. Not good to have both. Note that portNVIC_INT_CTRL_REG is defined in portmacro.h and is already used in this file (port.c). * Replicate to other Cortex M ports Also set a new fiddle factor based on tests with a CM4F. I used gcc, optimizing at -O1. Users can fine-tune as needed. Also add configSYSTICK_CLOCK_HZ to the CM0 ports to be just like the other Cortex M ports. This change allowed uniformity in the default tickless implementations across all Cortex M ports. And CM0 is likely to benefit from configSYSTICK_CLOCK_HZ, especially considering new CM0 devices with very fast CPU clock speeds. * Revert changes to IAR-CM0-portmacro.h portNVIC_INT_CTRL_REG was already defined in port.c. No need to define it in portmacro.h. * Handle edge cases with slow SysTick clock Co-authored-by: Cobus van Eeden <35851496+cobusve@users.noreply.github.com> Co-authored-by: abhidixi11 <44424462+abhidixi11@users.noreply.github.com> Co-authored-by: Joseph Julicher <jjulicher@mac.com> Co-authored-by: alfred gedeon <28123637+alfred2g@users.noreply.github.com>
2 years ago
}
#else /* configENABLE_MPU */
StackType_t * pxPortInitialiseStack( StackType_t * pxTopOfStack,
StackType_t * pxEndOfStack,
TaskFunction_t pxCode,
void * pvParameters ) /* PRIVILEGED_FUNCTION */
Tickless idle fixes/improvement (#59) * Fix tickless idle when stopping systick on zero... ...and don't stop SysTick at all in the eAbortSleep case. Prior to this commit, if vPortSuppressTicksAndSleep() happens to stop the SysTick on zero, then after tickless idle ends, xTickCount advances one full tick more than the time that actually elapsed as measured by the SysTick. See "bug 1" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - Idle task calls vPortSuppressTicksAndSleep(xExpectedIdleTime = 2). [Doesn't have to be "2" -- could be any number.] - vPortSuppressTicksAndSleep() stops SysTick, and the current-count register happens to stop on zero. - SysTick ISR executes, setting xPendedTicks = 1 - vPortSuppressTicksAndSleep() masks interrupts and calls eTaskConfirmSleepModeStatus() which confirms the sleep operation. *** - vPortSuppressTicksAndSleep() configures SysTick for 1 full tick (xExpectedIdleTime - 1) plus the current-count register (which is 0) - One tick period elapses in sleep. - SysTick wakes CPU, ISR executes and increments xPendedTicks to 2. - vPortSuppressTicksAndSleep() calls vTaskStepTick(1), then returns. - Idle task resumes scheduler, which increments xTickCount twice (for xPendedTicks = 2) In the end, two ticks elapsed as measured by SysTick, but the code increments xTickCount three times. The root cause is that the code assumes the SysTick current-count register always contains the number of SysTick counts remaining in the current tick period. However, when the current-count register is zero, there are ulTimerCountsForOneTick counts remaining, not zero. This error is not the kind of time slippage normally associated with tickless idle. *** Note that a recent commit https://github.com/FreeRTOS/FreeRTOS-Kernel/commit/e1b98f0 results in eAbortSleep in this case, due to xPendedTicks != 0. That commit does mostly resolve this bug without specifically mentioning it, and without this commit. But that resolution allows the code in port.c not to directly address the special case of stopping SysTick on zero in any code or comments. That commit also generates additional instances of eAbortSleep, and a second purpose of this commit is to optimize how vPortSuppressTicksAndSleep() behaves for eAbortSleep, as noted below. This commit also includes an optimization to avoid stopping the SysTick when eTaskConfirmSleepModeStatus() returns eAbortSleep. This optimization belongs with this fix because the method of handling the SysTick being stopped on zero changes with this optimization. * Fix imminent tick rescheduled after tickless idle Prior to this commit, if something other than systick wakes the CPU from tickless idle, vPortSuppressTicksAndSleep() might cause xTickCount to increment once too many times. See "bug 2" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - CPU is sleeping in vPortSuppressTicksAndSleep() - Something other than the SysTick wakes the CPU. - vPortSuppressTicksAndSleep() calculates the number of SysTick counts until the next tick. The bug occurs only if this number is small. - vPortSuppressTicksAndSleep() puts this small number into the SysTick reload register, and starts SysTick. - vPortSuppressTicksAndSleep() calls vTaskStepTick() - While vTaskStepTick() executes, the SysTick expires. The ISR pends because interrupts are masked, and SysTick starts a 2nd period still based on the small number of counts in its reload register. This 2nd period is undesirable and is likely to cause the error noted below. - vPortSuppressTicksAndSleep() puts the normal tick duration into the SysTick's reload register. - vPortSuppressTicksAndSleep() unmasks interrupts before the SysTick starts a new period based on the new value in the reload register. [This is a race condition that can go either way, but for the bug to occur, the race must play out this way.] - The pending SysTick ISR executes and increments xPendedTicks. - The SysTick expires again, finishing the second very small period, and starts a new period this time based on the full tick duration. - The SysTick ISR increments xPendedTicks (or xTickCount) even though only a tiny fraction of a tick period has elapsed since the previous tick. The bug occurs when *two* consecutive small periods of the SysTick are both counted as ticks. The root cause is a race caused by the small SysTick period. If vPortSuppressTicksAndSleep() unmasks interrupts *after* the small period expires but *before* the SysTick starts a period based on the full tick period, then two small periods are counted as ticks when only one should be counted. The end result is xTickCount advancing nearly one full tick more than time actually elapsed as measured by the SysTick. This is not the kind of time slippage normally associated with tickless idle. After this commit the code starts the SysTick and then immediately modifies the reload register to ensure the very short cycle (if any) is conducted only once. This strategy requires special consideration for the build option that configures SysTick to use a divided clock. To avoid waiting around for the SysTick to load value from the reload register, the new code temporarily configures the SysTick to use the undivided clock. The resulting timing error is typical for tickless idle. The error (commonly known as drift or slippage in kernel time) caused by this strategy is equivalent to one or two counts in ulStoppedTimerCompensation. This commit also updates comments and #define symbols related to the SysTick clock option. The SysTick can optionally be clocked by a divided version of the CPU clock (commonly divide-by-8). The new code in this commit adjusts these comments and symbols to make them clearer and more useful in configurations that use the divided clock. The fix made in this commit requires the use of these symbols, as noted in the code comments. * Fix tickless idle with alternate systick clocking Prior to this commit, in configurations using the alternate SysTick clocking, vPortSuppressTicksAndSleep() might cause xTickCount to jump ahead as much as the entire expected idle time or fall behind as much as one full tick compared to time as measured by the SysTick. SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. The SysTick has a configuration option to be clocked by an alternate clock besides the core clock. This alternate clock is MCU dependent. Scenarios Fixed --------------- The new code in this commit handles the following scenarios that were not handled correctly prior to this commit. 1. Before the sleep, vPortSuppressTicksAndSleep() stops the SysTick on zero, long after SysTick reached zero. Prior to this commit, this scenario caused xTickCount to jump ahead one full tick for the same reason documented here: https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/0c7b04bd3a745c52151abebc882eed3f811c4c81 2. After the sleep, vPortSuppressTicksAndSleep() stops the SysTick before it loads the counter from the reload register. Prior to this commit, this scenario caused xTickCount to jump ahead by the entire expected idle time (xExpectedIdleTime) because the current-count register is zero before it loads from the reload register. 3. Prior to return, vPortSuppressTicksAndSleep() attempts to start a short SysTick period when the current SysTick clock cycle has a lot of time remaining. Prior to this commit, this scenario could cause xTickCount to fall behind by as much as nearly one full tick because the short SysTick cycle never started. Note that #3 is partially fixed by https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/967acc9b200d3d4beeb289d9da9e88798074b431 even though that commit addresses a different issue. So this commit completes the partial fix. * Improve comments and name of preprocessor symbol Add a note in the code comments that SysTick requests an interrupt when decrementing from 1 to 0, so that's why stopping SysTick on zero is a special case. Readers might unknowingly assume that SysTick requests an interrupt when wrapping from 0 back to the load-register value. Reconsider new "_SETTING" suffix since "_CONFIG" suffix seems more descriptive. The code relies on *both* of these preprocessor symbols: portNVIC_SYSTICK_CLK_BIT portNVIC_SYSTICK_CLK_BIT_CONFIG **new** A meaningful suffix is really helpful to distinguish the two symbols. * Revert introduction of 2nd name for NVIC register When I added portNVIC_ICSR_REG I didn't realize there was already a portNVIC_INT_CTRL_REG, which identifies the same register. Not good to have both. Note that portNVIC_INT_CTRL_REG is defined in portmacro.h and is already used in this file (port.c). * Replicate to other Cortex M ports Also set a new fiddle factor based on tests with a CM4F. I used gcc, optimizing at -O1. Users can fine-tune as needed. Also add configSYSTICK_CLOCK_HZ to the CM0 ports to be just like the other Cortex M ports. This change allowed uniformity in the default tickless implementations across all Cortex M ports. And CM0 is likely to benefit from configSYSTICK_CLOCK_HZ, especially considering new CM0 devices with very fast CPU clock speeds. * Revert changes to IAR-CM0-portmacro.h portNVIC_INT_CTRL_REG was already defined in port.c. No need to define it in portmacro.h. * Handle edge cases with slow SysTick clock Co-authored-by: Cobus van Eeden <35851496+cobusve@users.noreply.github.com> Co-authored-by: abhidixi11 <44424462+abhidixi11@users.noreply.github.com> Co-authored-by: Joseph Julicher <jjulicher@mac.com> Co-authored-by: alfred gedeon <28123637+alfred2g@users.noreply.github.com>
2 years ago
{
/* Simulate the stack frame as it would be created by a context switch
* interrupt. */
#if ( portPRELOAD_REGISTERS == 0 )
{
pxTopOfStack--; /* Offset added to account for the way the MCU uses the stack on entry/exit of interrupts. */
*pxTopOfStack = portINITIAL_XPSR; /* xPSR. */
pxTopOfStack--;
*pxTopOfStack = ( StackType_t ) pxCode; /* PC. */
pxTopOfStack--;
*pxTopOfStack = ( StackType_t ) portTASK_RETURN_ADDRESS; /* LR. */
pxTopOfStack -= 5; /* R12, R3, R2 and R1. */
*pxTopOfStack = ( StackType_t ) pvParameters; /* R0. */
pxTopOfStack -= 9; /* R11..R4, EXC_RETURN. */
*pxTopOfStack = portINITIAL_EXC_RETURN;
pxTopOfStack--;
*pxTopOfStack = ( StackType_t ) pxEndOfStack; /* Slot used to hold this task's PSPLIM value. */
#if ( configENABLE_TRUSTZONE == 1 )
{
pxTopOfStack--;
*pxTopOfStack = portNO_SECURE_CONTEXT; /* Slot used to hold this task's xSecureContext value. */
}
#endif /* configENABLE_TRUSTZONE */
}
#else /* portPRELOAD_REGISTERS */
Tickless idle fixes/improvement (#59) * Fix tickless idle when stopping systick on zero... ...and don't stop SysTick at all in the eAbortSleep case. Prior to this commit, if vPortSuppressTicksAndSleep() happens to stop the SysTick on zero, then after tickless idle ends, xTickCount advances one full tick more than the time that actually elapsed as measured by the SysTick. See "bug 1" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - Idle task calls vPortSuppressTicksAndSleep(xExpectedIdleTime = 2). [Doesn't have to be "2" -- could be any number.] - vPortSuppressTicksAndSleep() stops SysTick, and the current-count register happens to stop on zero. - SysTick ISR executes, setting xPendedTicks = 1 - vPortSuppressTicksAndSleep() masks interrupts and calls eTaskConfirmSleepModeStatus() which confirms the sleep operation. *** - vPortSuppressTicksAndSleep() configures SysTick for 1 full tick (xExpectedIdleTime - 1) plus the current-count register (which is 0) - One tick period elapses in sleep. - SysTick wakes CPU, ISR executes and increments xPendedTicks to 2. - vPortSuppressTicksAndSleep() calls vTaskStepTick(1), then returns. - Idle task resumes scheduler, which increments xTickCount twice (for xPendedTicks = 2) In the end, two ticks elapsed as measured by SysTick, but the code increments xTickCount three times. The root cause is that the code assumes the SysTick current-count register always contains the number of SysTick counts remaining in the current tick period. However, when the current-count register is zero, there are ulTimerCountsForOneTick counts remaining, not zero. This error is not the kind of time slippage normally associated with tickless idle. *** Note that a recent commit https://github.com/FreeRTOS/FreeRTOS-Kernel/commit/e1b98f0 results in eAbortSleep in this case, due to xPendedTicks != 0. That commit does mostly resolve this bug without specifically mentioning it, and without this commit. But that resolution allows the code in port.c not to directly address the special case of stopping SysTick on zero in any code or comments. That commit also generates additional instances of eAbortSleep, and a second purpose of this commit is to optimize how vPortSuppressTicksAndSleep() behaves for eAbortSleep, as noted below. This commit also includes an optimization to avoid stopping the SysTick when eTaskConfirmSleepModeStatus() returns eAbortSleep. This optimization belongs with this fix because the method of handling the SysTick being stopped on zero changes with this optimization. * Fix imminent tick rescheduled after tickless idle Prior to this commit, if something other than systick wakes the CPU from tickless idle, vPortSuppressTicksAndSleep() might cause xTickCount to increment once too many times. See "bug 2" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - CPU is sleeping in vPortSuppressTicksAndSleep() - Something other than the SysTick wakes the CPU. - vPortSuppressTicksAndSleep() calculates the number of SysTick counts until the next tick. The bug occurs only if this number is small. - vPortSuppressTicksAndSleep() puts this small number into the SysTick reload register, and starts SysTick. - vPortSuppressTicksAndSleep() calls vTaskStepTick() - While vTaskStepTick() executes, the SysTick expires. The ISR pends because interrupts are masked, and SysTick starts a 2nd period still based on the small number of counts in its reload register. This 2nd period is undesirable and is likely to cause the error noted below. - vPortSuppressTicksAndSleep() puts the normal tick duration into the SysTick's reload register. - vPortSuppressTicksAndSleep() unmasks interrupts before the SysTick starts a new period based on the new value in the reload register. [This is a race condition that can go either way, but for the bug to occur, the race must play out this way.] - The pending SysTick ISR executes and increments xPendedTicks. - The SysTick expires again, finishing the second very small period, and starts a new period this time based on the full tick duration. - The SysTick ISR increments xPendedTicks (or xTickCount) even though only a tiny fraction of a tick period has elapsed since the previous tick. The bug occurs when *two* consecutive small periods of the SysTick are both counted as ticks. The root cause is a race caused by the small SysTick period. If vPortSuppressTicksAndSleep() unmasks interrupts *after* the small period expires but *before* the SysTick starts a period based on the full tick period, then two small periods are counted as ticks when only one should be counted. The end result is xTickCount advancing nearly one full tick more than time actually elapsed as measured by the SysTick. This is not the kind of time slippage normally associated with tickless idle. After this commit the code starts the SysTick and then immediately modifies the reload register to ensure the very short cycle (if any) is conducted only once. This strategy requires special consideration for the build option that configures SysTick to use a divided clock. To avoid waiting around for the SysTick to load value from the reload register, the new code temporarily configures the SysTick to use the undivided clock. The resulting timing error is typical for tickless idle. The error (commonly known as drift or slippage in kernel time) caused by this strategy is equivalent to one or two counts in ulStoppedTimerCompensation. This commit also updates comments and #define symbols related to the SysTick clock option. The SysTick can optionally be clocked by a divided version of the CPU clock (commonly divide-by-8). The new code in this commit adjusts these comments and symbols to make them clearer and more useful in configurations that use the divided clock. The fix made in this commit requires the use of these symbols, as noted in the code comments. * Fix tickless idle with alternate systick clocking Prior to this commit, in configurations using the alternate SysTick clocking, vPortSuppressTicksAndSleep() might cause xTickCount to jump ahead as much as the entire expected idle time or fall behind as much as one full tick compared to time as measured by the SysTick. SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. The SysTick has a configuration option to be clocked by an alternate clock besides the core clock. This alternate clock is MCU dependent. Scenarios Fixed --------------- The new code in this commit handles the following scenarios that were not handled correctly prior to this commit. 1. Before the sleep, vPortSuppressTicksAndSleep() stops the SysTick on zero, long after SysTick reached zero. Prior to this commit, this scenario caused xTickCount to jump ahead one full tick for the same reason documented here: https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/0c7b04bd3a745c52151abebc882eed3f811c4c81 2. After the sleep, vPortSuppressTicksAndSleep() stops the SysTick before it loads the counter from the reload register. Prior to this commit, this scenario caused xTickCount to jump ahead by the entire expected idle time (xExpectedIdleTime) because the current-count register is zero before it loads from the reload register. 3. Prior to return, vPortSuppressTicksAndSleep() attempts to start a short SysTick period when the current SysTick clock cycle has a lot of time remaining. Prior to this commit, this scenario could cause xTickCount to fall behind by as much as nearly one full tick because the short SysTick cycle never started. Note that #3 is partially fixed by https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/967acc9b200d3d4beeb289d9da9e88798074b431 even though that commit addresses a different issue. So this commit completes the partial fix. * Improve comments and name of preprocessor symbol Add a note in the code comments that SysTick requests an interrupt when decrementing from 1 to 0, so that's why stopping SysTick on zero is a special case. Readers might unknowingly assume that SysTick requests an interrupt when wrapping from 0 back to the load-register value. Reconsider new "_SETTING" suffix since "_CONFIG" suffix seems more descriptive. The code relies on *both* of these preprocessor symbols: portNVIC_SYSTICK_CLK_BIT portNVIC_SYSTICK_CLK_BIT_CONFIG **new** A meaningful suffix is really helpful to distinguish the two symbols. * Revert introduction of 2nd name for NVIC register When I added portNVIC_ICSR_REG I didn't realize there was already a portNVIC_INT_CTRL_REG, which identifies the same register. Not good to have both. Note that portNVIC_INT_CTRL_REG is defined in portmacro.h and is already used in this file (port.c). * Replicate to other Cortex M ports Also set a new fiddle factor based on tests with a CM4F. I used gcc, optimizing at -O1. Users can fine-tune as needed. Also add configSYSTICK_CLOCK_HZ to the CM0 ports to be just like the other Cortex M ports. This change allowed uniformity in the default tickless implementations across all Cortex M ports. And CM0 is likely to benefit from configSYSTICK_CLOCK_HZ, especially considering new CM0 devices with very fast CPU clock speeds. * Revert changes to IAR-CM0-portmacro.h portNVIC_INT_CTRL_REG was already defined in port.c. No need to define it in portmacro.h. * Handle edge cases with slow SysTick clock Co-authored-by: Cobus van Eeden <35851496+cobusve@users.noreply.github.com> Co-authored-by: abhidixi11 <44424462+abhidixi11@users.noreply.github.com> Co-authored-by: Joseph Julicher <jjulicher@mac.com> Co-authored-by: alfred gedeon <28123637+alfred2g@users.noreply.github.com>
2 years ago
{
pxTopOfStack--; /* Offset added to account for the way the MCU uses the stack on entry/exit of interrupts. */
*pxTopOfStack = portINITIAL_XPSR; /* xPSR. */
pxTopOfStack--;
*pxTopOfStack = ( StackType_t ) pxCode; /* PC. */
pxTopOfStack--;
*pxTopOfStack = ( StackType_t ) portTASK_RETURN_ADDRESS; /* LR. */
pxTopOfStack--;
*pxTopOfStack = ( StackType_t ) 0x12121212UL; /* R12. */
pxTopOfStack--;
*pxTopOfStack = ( StackType_t ) 0x03030303UL; /* R3. */
pxTopOfStack--;
*pxTopOfStack = ( StackType_t ) 0x02020202UL; /* R2. */
pxTopOfStack--;
*pxTopOfStack = ( StackType_t ) 0x01010101UL; /* R1. */
pxTopOfStack--;
*pxTopOfStack = ( StackType_t ) pvParameters; /* R0. */
pxTopOfStack--;
*pxTopOfStack = ( StackType_t ) 0x11111111UL; /* R11. */
pxTopOfStack--;
*pxTopOfStack = ( StackType_t ) 0x10101010UL; /* R10. */
pxTopOfStack--;
*pxTopOfStack = ( StackType_t ) 0x09090909UL; /* R09. */
pxTopOfStack--;
*pxTopOfStack = ( StackType_t ) 0x08080808UL; /* R08. */
pxTopOfStack--;
*pxTopOfStack = ( StackType_t ) 0x07070707UL; /* R07. */
pxTopOfStack--;
*pxTopOfStack = ( StackType_t ) 0x06060606UL; /* R06. */
pxTopOfStack--;
*pxTopOfStack = ( StackType_t ) 0x05050505UL; /* R05. */
pxTopOfStack--;
*pxTopOfStack = ( StackType_t ) 0x04040404UL; /* R04. */
pxTopOfStack--;
*pxTopOfStack = portINITIAL_EXC_RETURN; /* EXC_RETURN. */
pxTopOfStack--;
*pxTopOfStack = ( StackType_t ) pxEndOfStack; /* Slot used to hold this task's PSPLIM value. */
#if ( configENABLE_TRUSTZONE == 1 )
{
pxTopOfStack--;
*pxTopOfStack = portNO_SECURE_CONTEXT; /* Slot used to hold this task's xSecureContext value. */
}
#endif /* configENABLE_TRUSTZONE */
}
#endif /* portPRELOAD_REGISTERS */
return pxTopOfStack;
}
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
#endif /* configENABLE_MPU */
/*-----------------------------------------------------------*/
BaseType_t xPortStartScheduler( void ) /* PRIVILEGED_FUNCTION */
{
#if ( ( configASSERT_DEFINED == 1 ) && ( portHAS_ARMV8M_MAIN_EXTENSION == 1 ) )
Armv8-M (except Cortex-M23) interrupt priority checking (#673) * Armv8-M: Formatting changes Signed-off-by: Devaraj Ranganna <devaraj.ranganna@arm.com> * Armv8-M: Add support for interrupt priority check FreeRTOS provides `FromISR` system calls which can be called directly from interrupt service routines. It is crucial that the priority of these ISRs is set to same or lower value (numerically higher) than that of `configMAX_SYSCALL_INTERRUPT_PRIORITY`. For more information refer to https://www.FreeRTOS.org/RTOS-Cortex-M3-M4.html. Add a check to trigger an assert when an ISR with priority higher (numerically lower) than `configMAX_SYSCALL_INTERRUPT_PRIORITY` calls `FromISR` system calls if `configASSERT` macro is defined. In addition, add a config option `configQEMU_DISABLE_INTERRUPT_PRIO_BITS_CHECK` to disable interrupt priority check while running on QEMU. Based on the discussion https://gitlab.com/qemu-project/qemu/-/issues/1122, The interrupt priority bits in QEMU do not match the real hardware. Therefore the assert that checks the number of implemented bits and __NVIC_PRIO_BITS will always fail. The config option `configQEMU_DISABLE_INTERRUPT_PRIO_BITS_CHECK` should be defined in the `FreeRTOSConfig.h` for QEMU targets. Signed-off-by: Devaraj Ranganna <devaraj.ranganna@arm.com> * Use SHPR2 for calculating interrupt priority bits This removes the dependency on the secure software to mark the interrupt as non-secure. Signed-off-by: Gaurav Aggarwal <aggarg@amazon.com> --------- Signed-off-by: Devaraj Ranganna <devaraj.ranganna@arm.com> Signed-off-by: Gaurav Aggarwal <aggarg@amazon.com> Co-authored-by: Gaurav Aggarwal <aggarg@amazon.com> Co-authored-by: Gaurav-Aggarwal-AWS <33462878+aggarg@users.noreply.github.com>
2 years ago
{
volatile uint32_t ulOriginalPriority;
volatile uint32_t ulImplementedPrioBits = 0;
volatile uint8_t ucMaxPriorityValue;
/* Determine the maximum priority from which ISR safe FreeRTOS API
* functions can be called. ISR safe functions are those that end in
* "FromISR". FreeRTOS maintains separate thread and ISR API functions to
* ensure interrupt entry is as fast and simple as possible.
*
* Save the interrupt priority value that is about to be clobbered. */
ulOriginalPriority = portNVIC_SHPR2_REG;
/* Determine the number of priority bits available. First write to all
* possible bits. */
portNVIC_SHPR2_REG = 0xFF000000;
/* Read the value back to see how many bits stuck. */
ucMaxPriorityValue = ( uint8_t ) ( ( portNVIC_SHPR2_REG & 0xFF000000 ) >> 24 );
/* Use the same mask on the maximum system call priority. */
ucMaxSysCallPriority = configMAX_SYSCALL_INTERRUPT_PRIORITY & ucMaxPriorityValue;
/* Check that the maximum system call priority is nonzero after
* accounting for the number of priority bits supported by the
* hardware. A priority of 0 is invalid because setting the BASEPRI
* register to 0 unmasks all interrupts, and interrupts with priority 0
* cannot be masked using BASEPRI.
* See https://www.FreeRTOS.org/RTOS-Cortex-M3-M4.html */
configASSERT( ucMaxSysCallPriority );
/* Check that the bits not implemented in hardware are zero in
* configMAX_SYSCALL_INTERRUPT_PRIORITY. */
configASSERT( ( configMAX_SYSCALL_INTERRUPT_PRIORITY & ( uint8_t ) ( ~( uint32_t ) ucMaxPriorityValue ) ) == 0U );
Armv8-M (except Cortex-M23) interrupt priority checking (#673) * Armv8-M: Formatting changes Signed-off-by: Devaraj Ranganna <devaraj.ranganna@arm.com> * Armv8-M: Add support for interrupt priority check FreeRTOS provides `FromISR` system calls which can be called directly from interrupt service routines. It is crucial that the priority of these ISRs is set to same or lower value (numerically higher) than that of `configMAX_SYSCALL_INTERRUPT_PRIORITY`. For more information refer to https://www.FreeRTOS.org/RTOS-Cortex-M3-M4.html. Add a check to trigger an assert when an ISR with priority higher (numerically lower) than `configMAX_SYSCALL_INTERRUPT_PRIORITY` calls `FromISR` system calls if `configASSERT` macro is defined. In addition, add a config option `configQEMU_DISABLE_INTERRUPT_PRIO_BITS_CHECK` to disable interrupt priority check while running on QEMU. Based on the discussion https://gitlab.com/qemu-project/qemu/-/issues/1122, The interrupt priority bits in QEMU do not match the real hardware. Therefore the assert that checks the number of implemented bits and __NVIC_PRIO_BITS will always fail. The config option `configQEMU_DISABLE_INTERRUPT_PRIO_BITS_CHECK` should be defined in the `FreeRTOSConfig.h` for QEMU targets. Signed-off-by: Devaraj Ranganna <devaraj.ranganna@arm.com> * Use SHPR2 for calculating interrupt priority bits This removes the dependency on the secure software to mark the interrupt as non-secure. Signed-off-by: Gaurav Aggarwal <aggarg@amazon.com> --------- Signed-off-by: Devaraj Ranganna <devaraj.ranganna@arm.com> Signed-off-by: Gaurav Aggarwal <aggarg@amazon.com> Co-authored-by: Gaurav Aggarwal <aggarg@amazon.com> Co-authored-by: Gaurav-Aggarwal-AWS <33462878+aggarg@users.noreply.github.com>
2 years ago
/* Calculate the maximum acceptable priority group value for the number
* of bits read back. */
while( ( ucMaxPriorityValue & portTOP_BIT_OF_BYTE ) == portTOP_BIT_OF_BYTE )
{
ulImplementedPrioBits++;
ucMaxPriorityValue <<= ( uint8_t ) 0x01;
}
if( ulImplementedPrioBits == 8 )
{
/* When the hardware implements 8 priority bits, there is no way for
* the software to configure PRIGROUP to not have sub-priorities. As
* a result, the least significant bit is always used for sub-priority
* and there are 128 preemption priorities and 2 sub-priorities.
*
* This may cause some confusion in some cases - for example, if
* configMAX_SYSCALL_INTERRUPT_PRIORITY is set to 5, both 5 and 4
* priority interrupts will be masked in Critical Sections as those
* are at the same preemption priority. This may appear confusing as
* 4 is higher (numerically lower) priority than
* configMAX_SYSCALL_INTERRUPT_PRIORITY and therefore, should not
* have been masked. Instead, if we set configMAX_SYSCALL_INTERRUPT_PRIORITY
* to 4, this confusion does not happen and the behaviour remains the same.
*
* The following assert ensures that the sub-priority bit in the
* configMAX_SYSCALL_INTERRUPT_PRIORITY is clear to avoid the above mentioned
* confusion. */
configASSERT( ( configMAX_SYSCALL_INTERRUPT_PRIORITY & 0x1U ) == 0U );
ulMaxPRIGROUPValue = 0;
}
else
{
ulMaxPRIGROUPValue = portMAX_PRIGROUP_BITS - ulImplementedPrioBits;
}
/* Shift the priority group value back to its position within the AIRCR
* register. */
ulMaxPRIGROUPValue <<= portPRIGROUP_SHIFT;
ulMaxPRIGROUPValue &= portPRIORITY_GROUP_MASK;
/* Restore the clobbered interrupt priority register to its original
* value. */
portNVIC_SHPR2_REG = ulOriginalPriority;
}
#endif /* #if ( ( configASSERT_DEFINED == 1 ) && ( portHAS_ARMV8M_MAIN_EXTENSION == 1 ) ) */
Armv8-M (except Cortex-M23) interrupt priority checking (#673) * Armv8-M: Formatting changes Signed-off-by: Devaraj Ranganna <devaraj.ranganna@arm.com> * Armv8-M: Add support for interrupt priority check FreeRTOS provides `FromISR` system calls which can be called directly from interrupt service routines. It is crucial that the priority of these ISRs is set to same or lower value (numerically higher) than that of `configMAX_SYSCALL_INTERRUPT_PRIORITY`. For more information refer to https://www.FreeRTOS.org/RTOS-Cortex-M3-M4.html. Add a check to trigger an assert when an ISR with priority higher (numerically lower) than `configMAX_SYSCALL_INTERRUPT_PRIORITY` calls `FromISR` system calls if `configASSERT` macro is defined. In addition, add a config option `configQEMU_DISABLE_INTERRUPT_PRIO_BITS_CHECK` to disable interrupt priority check while running on QEMU. Based on the discussion https://gitlab.com/qemu-project/qemu/-/issues/1122, The interrupt priority bits in QEMU do not match the real hardware. Therefore the assert that checks the number of implemented bits and __NVIC_PRIO_BITS will always fail. The config option `configQEMU_DISABLE_INTERRUPT_PRIO_BITS_CHECK` should be defined in the `FreeRTOSConfig.h` for QEMU targets. Signed-off-by: Devaraj Ranganna <devaraj.ranganna@arm.com> * Use SHPR2 for calculating interrupt priority bits This removes the dependency on the secure software to mark the interrupt as non-secure. Signed-off-by: Gaurav Aggarwal <aggarg@amazon.com> --------- Signed-off-by: Devaraj Ranganna <devaraj.ranganna@arm.com> Signed-off-by: Gaurav Aggarwal <aggarg@amazon.com> Co-authored-by: Gaurav Aggarwal <aggarg@amazon.com> Co-authored-by: Gaurav-Aggarwal-AWS <33462878+aggarg@users.noreply.github.com>
2 years ago
/* Make PendSV, CallSV and SysTick the same priority as the kernel. */
portNVIC_SHPR3_REG |= portNVIC_PENDSV_PRI;
portNVIC_SHPR3_REG |= portNVIC_SYSTICK_PRI;
#if ( configENABLE_MPU == 1 )
Tickless idle fixes/improvement (#59) * Fix tickless idle when stopping systick on zero... ...and don't stop SysTick at all in the eAbortSleep case. Prior to this commit, if vPortSuppressTicksAndSleep() happens to stop the SysTick on zero, then after tickless idle ends, xTickCount advances one full tick more than the time that actually elapsed as measured by the SysTick. See "bug 1" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - Idle task calls vPortSuppressTicksAndSleep(xExpectedIdleTime = 2). [Doesn't have to be "2" -- could be any number.] - vPortSuppressTicksAndSleep() stops SysTick, and the current-count register happens to stop on zero. - SysTick ISR executes, setting xPendedTicks = 1 - vPortSuppressTicksAndSleep() masks interrupts and calls eTaskConfirmSleepModeStatus() which confirms the sleep operation. *** - vPortSuppressTicksAndSleep() configures SysTick for 1 full tick (xExpectedIdleTime - 1) plus the current-count register (which is 0) - One tick period elapses in sleep. - SysTick wakes CPU, ISR executes and increments xPendedTicks to 2. - vPortSuppressTicksAndSleep() calls vTaskStepTick(1), then returns. - Idle task resumes scheduler, which increments xTickCount twice (for xPendedTicks = 2) In the end, two ticks elapsed as measured by SysTick, but the code increments xTickCount three times. The root cause is that the code assumes the SysTick current-count register always contains the number of SysTick counts remaining in the current tick period. However, when the current-count register is zero, there are ulTimerCountsForOneTick counts remaining, not zero. This error is not the kind of time slippage normally associated with tickless idle. *** Note that a recent commit https://github.com/FreeRTOS/FreeRTOS-Kernel/commit/e1b98f0 results in eAbortSleep in this case, due to xPendedTicks != 0. That commit does mostly resolve this bug without specifically mentioning it, and without this commit. But that resolution allows the code in port.c not to directly address the special case of stopping SysTick on zero in any code or comments. That commit also generates additional instances of eAbortSleep, and a second purpose of this commit is to optimize how vPortSuppressTicksAndSleep() behaves for eAbortSleep, as noted below. This commit also includes an optimization to avoid stopping the SysTick when eTaskConfirmSleepModeStatus() returns eAbortSleep. This optimization belongs with this fix because the method of handling the SysTick being stopped on zero changes with this optimization. * Fix imminent tick rescheduled after tickless idle Prior to this commit, if something other than systick wakes the CPU from tickless idle, vPortSuppressTicksAndSleep() might cause xTickCount to increment once too many times. See "bug 2" in this forum post: https://forums.freertos.org/t/ultasknotifytake-timeout-accuracy/9629/40 SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. To get periodic interrupts every N SysTick clock cycles, the reload register must be N - 1. Bug Example ----------- - CPU is sleeping in vPortSuppressTicksAndSleep() - Something other than the SysTick wakes the CPU. - vPortSuppressTicksAndSleep() calculates the number of SysTick counts until the next tick. The bug occurs only if this number is small. - vPortSuppressTicksAndSleep() puts this small number into the SysTick reload register, and starts SysTick. - vPortSuppressTicksAndSleep() calls vTaskStepTick() - While vTaskStepTick() executes, the SysTick expires. The ISR pends because interrupts are masked, and SysTick starts a 2nd period still based on the small number of counts in its reload register. This 2nd period is undesirable and is likely to cause the error noted below. - vPortSuppressTicksAndSleep() puts the normal tick duration into the SysTick's reload register. - vPortSuppressTicksAndSleep() unmasks interrupts before the SysTick starts a new period based on the new value in the reload register. [This is a race condition that can go either way, but for the bug to occur, the race must play out this way.] - The pending SysTick ISR executes and increments xPendedTicks. - The SysTick expires again, finishing the second very small period, and starts a new period this time based on the full tick duration. - The SysTick ISR increments xPendedTicks (or xTickCount) even though only a tiny fraction of a tick period has elapsed since the previous tick. The bug occurs when *two* consecutive small periods of the SysTick are both counted as ticks. The root cause is a race caused by the small SysTick period. If vPortSuppressTicksAndSleep() unmasks interrupts *after* the small period expires but *before* the SysTick starts a period based on the full tick period, then two small periods are counted as ticks when only one should be counted. The end result is xTickCount advancing nearly one full tick more than time actually elapsed as measured by the SysTick. This is not the kind of time slippage normally associated with tickless idle. After this commit the code starts the SysTick and then immediately modifies the reload register to ensure the very short cycle (if any) is conducted only once. This strategy requires special consideration for the build option that configures SysTick to use a divided clock. To avoid waiting around for the SysTick to load value from the reload register, the new code temporarily configures the SysTick to use the undivided clock. The resulting timing error is typical for tickless idle. The error (commonly known as drift or slippage in kernel time) caused by this strategy is equivalent to one or two counts in ulStoppedTimerCompensation. This commit also updates comments and #define symbols related to the SysTick clock option. The SysTick can optionally be clocked by a divided version of the CPU clock (commonly divide-by-8). The new code in this commit adjusts these comments and symbols to make them clearer and more useful in configurations that use the divided clock. The fix made in this commit requires the use of these symbols, as noted in the code comments. * Fix tickless idle with alternate systick clocking Prior to this commit, in configurations using the alternate SysTick clocking, vPortSuppressTicksAndSleep() might cause xTickCount to jump ahead as much as the entire expected idle time or fall behind as much as one full tick compared to time as measured by the SysTick. SysTick ------- The SysTick is the hardware timer that provides the OS tick interrupt in the official ports for Cortex M. SysTick starts counting down from the value stored in its reload register. When SysTick reaches zero, it requests an interrupt. On the next SysTick clock cycle, it loads the counter again from the reload register. The SysTick has a configuration option to be clocked by an alternate clock besides the core clock. This alternate clock is MCU dependent. Scenarios Fixed --------------- The new code in this commit handles the following scenarios that were not handled correctly prior to this commit. 1. Before the sleep, vPortSuppressTicksAndSleep() stops the SysTick on zero, long after SysTick reached zero. Prior to this commit, this scenario caused xTickCount to jump ahead one full tick for the same reason documented here: https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/0c7b04bd3a745c52151abebc882eed3f811c4c81 2. After the sleep, vPortSuppressTicksAndSleep() stops the SysTick before it loads the counter from the reload register. Prior to this commit, this scenario caused xTickCount to jump ahead by the entire expected idle time (xExpectedIdleTime) because the current-count register is zero before it loads from the reload register. 3. Prior to return, vPortSuppressTicksAndSleep() attempts to start a short SysTick period when the current SysTick clock cycle has a lot of time remaining. Prior to this commit, this scenario could cause xTickCount to fall behind by as much as nearly one full tick because the short SysTick cycle never started. Note that #3 is partially fixed by https://github.com/FreeRTOS/FreeRTOS-Kernel/pull/59/commits/967acc9b200d3d4beeb289d9da9e88798074b431 even though that commit addresses a different issue. So this commit completes the partial fix. * Improve comments and name of preprocessor symbol Add a note in the code comments that SysTick requests an interrupt when decrementing from 1 to 0, so that's why stopping SysTick on zero is a special case. Readers might unknowingly assume that SysTick requests an interrupt when wrapping from 0 back to the load-register value. Reconsider new "_SETTING" suffix since "_CONFIG" suffix seems more descriptive. The code relies on *both* of these preprocessor symbols: portNVIC_SYSTICK_CLK_BIT portNVIC_SYSTICK_CLK_BIT_CONFIG **new** A meaningful suffix is really helpful to distinguish the two symbols. * Revert introduction of 2nd name for NVIC register When I added portNVIC_ICSR_REG I didn't realize there was already a portNVIC_INT_CTRL_REG, which identifies the same register. Not good to have both. Note that portNVIC_INT_CTRL_REG is defined in portmacro.h and is already used in this file (port.c). * Replicate to other Cortex M ports Also set a new fiddle factor based on tests with a CM4F. I used gcc, optimizing at -O1. Users can fine-tune as needed. Also add configSYSTICK_CLOCK_HZ to the CM0 ports to be just like the other Cortex M ports. This change allowed uniformity in the default tickless implementations across all Cortex M ports. And CM0 is likely to benefit from configSYSTICK_CLOCK_HZ, especially considering new CM0 devices with very fast CPU clock speeds. * Revert changes to IAR-CM0-portmacro.h portNVIC_INT_CTRL_REG was already defined in port.c. No need to define it in portmacro.h. * Handle edge cases with slow SysTick clock Co-authored-by: Cobus van Eeden <35851496+cobusve@users.noreply.github.com> Co-authored-by: abhidixi11 <44424462+abhidixi11@users.noreply.github.com> Co-authored-by: Joseph Julicher <jjulicher@mac.com> Co-authored-by: alfred gedeon <28123637+alfred2g@users.noreply.github.com>
2 years ago
{
/* Setup the Memory Protection Unit (MPU). */
prvSetupMPU();
}
#endif /* configENABLE_MPU */
/* Start the timer that generates the tick ISR. Interrupts are disabled
* here already. */
vPortSetupTimerInterrupt();
/* Initialize the critical nesting count ready for the first task. */
ulCriticalNesting = 0;
/* Start the first task. */
vStartFirstTask();
/* Should never get here as the tasks will now be executing. Call the task
* exit error function to prevent compiler warnings about a static function
* not being called in the case that the application writer overrides this
* functionality by defining configTASK_RETURN_ADDRESS. Call
* vTaskSwitchContext() so link time optimization does not remove the
* symbol. */
vTaskSwitchContext();
prvTaskExitError();
/* Should not get here. */
return 0;
}
/*-----------------------------------------------------------*/
void vPortEndScheduler( void ) /* PRIVILEGED_FUNCTION */
{
/* Not implemented in ports where there is nothing to return to.
* Artificially force an assert. */
configASSERT( ulCriticalNesting == 1000UL );
}
/*-----------------------------------------------------------*/
#if ( configENABLE_MPU == 1 )
void vPortStoreTaskMPUSettings( xMPU_SETTINGS * xMPUSettings,
const struct xMEMORY_REGION * const xRegions,
StackType_t * pxBottomOfStack,
uint32_t ulStackDepth )
{
uint32_t ulRegionStartAddress, ulRegionEndAddress, ulRegionNumber;
int32_t lIndex = 0;
#if defined( __ARMCC_VERSION )
/* Declaration when these variable are defined in code instead of being
* exported from linker scripts. */
extern uint32_t * __privileged_sram_start__;
extern uint32_t * __privileged_sram_end__;
#else
/* Declaration when these variable are exported from linker scripts. */
extern uint32_t __privileged_sram_start__[];
extern uint32_t __privileged_sram_end__[];
#endif /* defined( __ARMCC_VERSION ) */
/* Setup MAIR0. */
xMPUSettings->ulMAIR0 = ( ( portMPU_NORMAL_MEMORY_BUFFERABLE_CACHEABLE << portMPU_MAIR_ATTR0_POS ) & portMPU_MAIR_ATTR0_MASK );
xMPUSettings->ulMAIR0 |= ( ( portMPU_DEVICE_MEMORY_nGnRE << portMPU_MAIR_ATTR1_POS ) & portMPU_MAIR_ATTR1_MASK );
/* This function is called automatically when the task is created - in
* which case the stack region parameters will be valid. At all other
* times the stack parameters will not be valid and it is assumed that
* the stack region has already been configured. */
if( ulStackDepth > 0 )
{
ulRegionStartAddress = ( uint32_t ) pxBottomOfStack;
ulRegionEndAddress = ( uint32_t ) pxBottomOfStack + ( ulStackDepth * ( uint32_t ) sizeof( StackType_t ) ) - 1;
/* If the stack is within the privileged SRAM, do not protect it
* using a separate MPU region. This is needed because privileged
* SRAM is already protected using an MPU region and ARMv8-M does
* not allow overlapping MPU regions. */
if( ( ulRegionStartAddress >= ( uint32_t ) __privileged_sram_start__ ) &&
( ulRegionEndAddress <= ( uint32_t ) __privileged_sram_end__ ) )
{
xMPUSettings->xRegionsSettings[ 0 ].ulRBAR = 0;
xMPUSettings->xRegionsSettings[ 0 ].ulRLAR = 0;
}
else
{
/* Define the region that allows access to the stack. */
ulRegionStartAddress &= portMPU_RBAR_ADDRESS_MASK;
ulRegionEndAddress &= portMPU_RLAR_ADDRESS_MASK;
xMPUSettings->xRegionsSettings[ 0 ].ulRBAR = ( ulRegionStartAddress ) |
( portMPU_REGION_NON_SHAREABLE ) |
( portMPU_REGION_READ_WRITE ) |
( portMPU_REGION_EXECUTE_NEVER );
xMPUSettings->xRegionsSettings[ 0 ].ulRLAR = ( ulRegionEndAddress ) |
( portMPU_RLAR_ATTR_INDEX0 ) |
( portMPU_RLAR_REGION_ENABLE );
}
}
/* User supplied configurable regions. */
for( ulRegionNumber = 1; ulRegionNumber <= portNUM_CONFIGURABLE_REGIONS; ulRegionNumber++ )
{
/* If xRegions is NULL i.e. the task has not specified any MPU
* region, the else part ensures that all the configurable MPU
* regions are invalidated. */
if( ( xRegions != NULL ) && ( xRegions[ lIndex ].ulLengthInBytes > 0UL ) )
{
/* Translate the generic region definition contained in xRegions
* into the ARMv8 specific MPU settings that are then stored in
* xMPUSettings. */
ulRegionStartAddress = ( ( uint32_t ) xRegions[ lIndex ].pvBaseAddress ) & portMPU_RBAR_ADDRESS_MASK;
ulRegionEndAddress = ( uint32_t ) xRegions[ lIndex ].pvBaseAddress + xRegions[ lIndex ].ulLengthInBytes - 1;
ulRegionEndAddress &= portMPU_RLAR_ADDRESS_MASK;
/* Start address. */
xMPUSettings->xRegionsSettings[ ulRegionNumber ].ulRBAR = ( ulRegionStartAddress ) |
( portMPU_REGION_NON_SHAREABLE );
/* RO/RW. */
if( ( xRegions[ lIndex ].ulParameters & tskMPU_REGION_READ_ONLY ) != 0 )
{
xMPUSettings->xRegionsSettings[ ulRegionNumber ].ulRBAR |= ( portMPU_REGION_READ_ONLY );
}
else
{
xMPUSettings->xRegionsSettings[ ulRegionNumber ].ulRBAR |= ( portMPU_REGION_READ_WRITE );
}
/* XN. */
if( ( xRegions[ lIndex ].ulParameters & tskMPU_REGION_EXECUTE_NEVER ) != 0 )
{
xMPUSettings->xRegionsSettings[ ulRegionNumber ].ulRBAR |= ( portMPU_REGION_EXECUTE_NEVER );
}
/* End Address. */
xMPUSettings->xRegionsSettings[ ulRegionNumber ].ulRLAR = ( ulRegionEndAddress ) |
( portMPU_RLAR_REGION_ENABLE );
/* Normal memory/ Device memory. */
if( ( xRegions[ lIndex ].ulParameters & tskMPU_REGION_DEVICE_MEMORY ) != 0 )
{
/* Attr1 in MAIR0 is configured as device memory. */
xMPUSettings->xRegionsSettings[ ulRegionNumber ].ulRLAR |= portMPU_RLAR_ATTR_INDEX1;
}
else
{
/* Attr0 in MAIR0 is configured as normal memory. */
xMPUSettings->xRegionsSettings[ ulRegionNumber ].ulRLAR |= portMPU_RLAR_ATTR_INDEX0;
}
}
else
{
/* Invalidate the region. */
xMPUSettings->xRegionsSettings[ ulRegionNumber ].ulRBAR = 0UL;
xMPUSettings->xRegionsSettings[ ulRegionNumber ].ulRLAR = 0UL;
}
lIndex++;
}
}
#endif /* configENABLE_MPU */
/*-----------------------------------------------------------*/
Memory Protection Unit (MPU) Enhancements (#705) Memory Protection Unit (MPU) Enhancements This commit introduces a new MPU wrapper that places additional restrictions on unprivileged tasks. The following is the list of changes introduced with the new MPU wrapper: 1. Opaque and indirectly verifiable integers for kernel object handles: All the kernel object handles (for example, queue handles) are now opaque integers. Previously object handles were raw pointers. 2. Saving the task context in Task Control Block (TCB): When a task is swapped out by the scheduler, the task's context is now saved in its TCB. Previously the task's context was saved on its stack. 3. Execute system calls on a separate privileged only stack: FreeRTOS system calls, which execute with elevated privilege, now use a separate privileged only stack. Previously system calls used the calling task's stack. The application writer can control the size of the system call stack using new configSYSTEM_CALL_STACK_SIZE config macro. 4. Memory bounds checks: FreeRTOS system calls which accept a pointer and de-reference it, now verify that the calling task has required permissions to access the memory location referenced by the pointer. 5. System call restrictions: The following system calls are no longer available to unprivileged tasks: - vQueueDelete - xQueueCreateMutex - xQueueCreateMutexStatic - xQueueCreateCountingSemaphore - xQueueCreateCountingSemaphoreStatic - xQueueGenericCreate - xQueueGenericCreateStatic - xQueueCreateSet - xQueueRemoveFromSet - xQueueGenericReset - xTaskCreate - xTaskCreateStatic - vTaskDelete - vTaskPrioritySet - vTaskSuspendAll - xTaskResumeAll - xTaskGetHandle - xTaskCallApplicationTaskHook - vTaskList - vTaskGetRunTimeStats - xTaskCatchUpTicks - xEventGroupCreate - xEventGroupCreateStatic - vEventGroupDelete - xStreamBufferGenericCreate - xStreamBufferGenericCreateStatic - vStreamBufferDelete - xStreamBufferReset Also, an unprivileged task can no longer use vTaskSuspend to suspend any task other than itself. We thank the following people for their inputs in these enhancements: - David Reiss of Meta Platforms, Inc. - Lan Luo, Xinhui Shao, Yumeng Wei, Zixia Liu, Huaiyu Yan and Zhen Ling of School of Computer Science and Engineering, Southeast University, China. - Xinwen Fu of Department of Computer Science, University of Massachusetts Lowell, USA. - Yuequi Chen, Zicheng Wang, Minghao Lin of University of Colorado Boulder, USA.
2 years ago
#if ( configENABLE_MPU == 1 )
BaseType_t xPortIsAuthorizedToAccessBuffer( const void * pvBuffer,
uint32_t ulBufferLength,
uint32_t ulAccessRequested ) /* PRIVILEGED_FUNCTION */
{
uint32_t i, ulBufferStartAddress, ulBufferEndAddress;
BaseType_t xAccessGranted = pdFALSE;
const xMPU_SETTINGS * xTaskMpuSettings = xTaskGetMPUSettings( NULL ); /* Calling task's MPU settings. */
if( ( xTaskMpuSettings->ulTaskFlags & portTASK_IS_PRIVILEGED_FLAG ) == portTASK_IS_PRIVILEGED_FLAG )
{
xAccessGranted = pdTRUE;
}
else
{
if( portADD_UINT32_WILL_OVERFLOW( ( ( uint32_t ) pvBuffer ), ( ulBufferLength - 1UL ) ) == pdFALSE )
{
ulBufferStartAddress = ( uint32_t ) pvBuffer;
ulBufferEndAddress = ( ( ( uint32_t ) pvBuffer ) + ulBufferLength - 1UL );
for( i = 0; i < portTOTAL_NUM_REGIONS; i++ )
{
/* Is the MPU region enabled? */
if( ( xTaskMpuSettings->xRegionsSettings[ i ].ulRLAR & portMPU_RLAR_REGION_ENABLE ) == portMPU_RLAR_REGION_ENABLE )
{
if( portIS_ADDRESS_WITHIN_RANGE( ulBufferStartAddress,
portEXTRACT_FIRST_ADDRESS_FROM_RBAR( xTaskMpuSettings->xRegionsSettings[ i ].ulRBAR ),
portEXTRACT_LAST_ADDRESS_FROM_RLAR( xTaskMpuSettings->xRegionsSettings[ i ].ulRLAR ) ) &&
portIS_ADDRESS_WITHIN_RANGE( ulBufferEndAddress,
portEXTRACT_FIRST_ADDRESS_FROM_RBAR( xTaskMpuSettings->xRegionsSettings[ i ].ulRBAR ),
portEXTRACT_LAST_ADDRESS_FROM_RLAR( xTaskMpuSettings->xRegionsSettings[ i ].ulRLAR ) ) &&
portIS_AUTHORIZED( ulAccessRequested,
prvGetRegionAccessPermissions( xTaskMpuSettings->xRegionsSettings[ i ].ulRBAR ) ) )
{
xAccessGranted = pdTRUE;
break;
}
}
}
}
}
return xAccessGranted;
}
#endif /* configENABLE_MPU */
/*-----------------------------------------------------------*/
BaseType_t xPortIsInsideInterrupt( void )
{
uint32_t ulCurrentInterrupt;
BaseType_t xReturn;
/* Obtain the number of the currently executing interrupt. Interrupt Program
* Status Register (IPSR) holds the exception number of the currently-executing
* exception or zero for Thread mode.*/
__asm volatile ( "mrs %0, ipsr" : "=r" ( ulCurrentInterrupt )::"memory" );
if( ulCurrentInterrupt == 0 )
{
xReturn = pdFALSE;
}
else
{
xReturn = pdTRUE;
}
return xReturn;
}
/*-----------------------------------------------------------*/
Armv8-M (except Cortex-M23) interrupt priority checking (#673) * Armv8-M: Formatting changes Signed-off-by: Devaraj Ranganna <devaraj.ranganna@arm.com> * Armv8-M: Add support for interrupt priority check FreeRTOS provides `FromISR` system calls which can be called directly from interrupt service routines. It is crucial that the priority of these ISRs is set to same or lower value (numerically higher) than that of `configMAX_SYSCALL_INTERRUPT_PRIORITY`. For more information refer to https://www.FreeRTOS.org/RTOS-Cortex-M3-M4.html. Add a check to trigger an assert when an ISR with priority higher (numerically lower) than `configMAX_SYSCALL_INTERRUPT_PRIORITY` calls `FromISR` system calls if `configASSERT` macro is defined. In addition, add a config option `configQEMU_DISABLE_INTERRUPT_PRIO_BITS_CHECK` to disable interrupt priority check while running on QEMU. Based on the discussion https://gitlab.com/qemu-project/qemu/-/issues/1122, The interrupt priority bits in QEMU do not match the real hardware. Therefore the assert that checks the number of implemented bits and __NVIC_PRIO_BITS will always fail. The config option `configQEMU_DISABLE_INTERRUPT_PRIO_BITS_CHECK` should be defined in the `FreeRTOSConfig.h` for QEMU targets. Signed-off-by: Devaraj Ranganna <devaraj.ranganna@arm.com> * Use SHPR2 for calculating interrupt priority bits This removes the dependency on the secure software to mark the interrupt as non-secure. Signed-off-by: Gaurav Aggarwal <aggarg@amazon.com> --------- Signed-off-by: Devaraj Ranganna <devaraj.ranganna@arm.com> Signed-off-by: Gaurav Aggarwal <aggarg@amazon.com> Co-authored-by: Gaurav Aggarwal <aggarg@amazon.com> Co-authored-by: Gaurav-Aggarwal-AWS <33462878+aggarg@users.noreply.github.com>
2 years ago
#if ( ( configASSERT_DEFINED == 1 ) && ( portHAS_ARMV8M_MAIN_EXTENSION == 1 ) )
Armv8-M (except Cortex-M23) interrupt priority checking (#673) * Armv8-M: Formatting changes Signed-off-by: Devaraj Ranganna <devaraj.ranganna@arm.com> * Armv8-M: Add support for interrupt priority check FreeRTOS provides `FromISR` system calls which can be called directly from interrupt service routines. It is crucial that the priority of these ISRs is set to same or lower value (numerically higher) than that of `configMAX_SYSCALL_INTERRUPT_PRIORITY`. For more information refer to https://www.FreeRTOS.org/RTOS-Cortex-M3-M4.html. Add a check to trigger an assert when an ISR with priority higher (numerically lower) than `configMAX_SYSCALL_INTERRUPT_PRIORITY` calls `FromISR` system calls if `configASSERT` macro is defined. In addition, add a config option `configQEMU_DISABLE_INTERRUPT_PRIO_BITS_CHECK` to disable interrupt priority check while running on QEMU. Based on the discussion https://gitlab.com/qemu-project/qemu/-/issues/1122, The interrupt priority bits in QEMU do not match the real hardware. Therefore the assert that checks the number of implemented bits and __NVIC_PRIO_BITS will always fail. The config option `configQEMU_DISABLE_INTERRUPT_PRIO_BITS_CHECK` should be defined in the `FreeRTOSConfig.h` for QEMU targets. Signed-off-by: Devaraj Ranganna <devaraj.ranganna@arm.com> * Use SHPR2 for calculating interrupt priority bits This removes the dependency on the secure software to mark the interrupt as non-secure. Signed-off-by: Gaurav Aggarwal <aggarg@amazon.com> --------- Signed-off-by: Devaraj Ranganna <devaraj.ranganna@arm.com> Signed-off-by: Gaurav Aggarwal <aggarg@amazon.com> Co-authored-by: Gaurav Aggarwal <aggarg@amazon.com> Co-authored-by: Gaurav-Aggarwal-AWS <33462878+aggarg@users.noreply.github.com>
2 years ago
void vPortValidateInterruptPriority( void )
{
uint32_t ulCurrentInterrupt;
uint8_t ucCurrentPriority;
/* Obtain the number of the currently executing interrupt. */
__asm volatile ( "mrs %0, ipsr" : "=r" ( ulCurrentInterrupt )::"memory" );
/* Is the interrupt number a user defined interrupt? */
if( ulCurrentInterrupt >= portFIRST_USER_INTERRUPT_NUMBER )
{
/* Look up the interrupt's priority. */
ucCurrentPriority = pcInterruptPriorityRegisters[ ulCurrentInterrupt ];
/* The following assertion will fail if a service routine (ISR) for
* an interrupt that has been assigned a priority above
* configMAX_SYSCALL_INTERRUPT_PRIORITY calls an ISR safe FreeRTOS API
* function. ISR safe FreeRTOS API functions must *only* be called
* from interrupts that have been assigned a priority at or below
* configMAX_SYSCALL_INTERRUPT_PRIORITY.
*
* Numerically low interrupt priority numbers represent logically high
* interrupt priorities, therefore the priority of the interrupt must
* be set to a value equal to or numerically *higher* than
* configMAX_SYSCALL_INTERRUPT_PRIORITY.
*
* Interrupts that use the FreeRTOS API must not be left at their
* default priority of zero as that is the highest possible priority,
* which is guaranteed to be above configMAX_SYSCALL_INTERRUPT_PRIORITY,
* and therefore also guaranteed to be invalid.
*
* FreeRTOS maintains separate thread and ISR API functions to ensure
* interrupt entry is as fast and simple as possible.
*
* The following links provide detailed information:
* https://www.FreeRTOS.org/RTOS-Cortex-M3-M4.html
* https://www.FreeRTOS.org/FAQHelp.html */
configASSERT( ucCurrentPriority >= ucMaxSysCallPriority );
}
/* Priority grouping: The interrupt controller (NVIC) allows the bits
* that define each interrupt's priority to be split between bits that
* define the interrupt's pre-emption priority bits and bits that define
* the interrupt's sub-priority. For simplicity all bits must be defined
* to be pre-emption priority bits. The following assertion will fail if
* this is not the case (if some bits represent a sub-priority).
*
* If the application only uses CMSIS libraries for interrupt
* configuration then the correct setting can be achieved on all Cortex-M
* devices by calling NVIC_SetPriorityGrouping( 0 ); before starting the
* scheduler. Note however that some vendor specific peripheral libraries
* assume a non-zero priority group setting, in which cases using a value
* of zero will result in unpredictable behaviour. */
configASSERT( ( portAIRCR_REG & portPRIORITY_GROUP_MASK ) <= ulMaxPRIGROUPValue );
}
#endif /* #if ( ( configASSERT_DEFINED == 1 ) && ( portHAS_ARMV8M_MAIN_EXTENSION == 1 ) ) */
Armv8-M (except Cortex-M23) interrupt priority checking (#673) * Armv8-M: Formatting changes Signed-off-by: Devaraj Ranganna <devaraj.ranganna@arm.com> * Armv8-M: Add support for interrupt priority check FreeRTOS provides `FromISR` system calls which can be called directly from interrupt service routines. It is crucial that the priority of these ISRs is set to same or lower value (numerically higher) than that of `configMAX_SYSCALL_INTERRUPT_PRIORITY`. For more information refer to https://www.FreeRTOS.org/RTOS-Cortex-M3-M4.html. Add a check to trigger an assert when an ISR with priority higher (numerically lower) than `configMAX_SYSCALL_INTERRUPT_PRIORITY` calls `FromISR` system calls if `configASSERT` macro is defined. In addition, add a config option `configQEMU_DISABLE_INTERRUPT_PRIO_BITS_CHECK` to disable interrupt priority check while running on QEMU. Based on the discussion https://gitlab.com/qemu-project/qemu/-/issues/1122, The interrupt priority bits in QEMU do not match the real hardware. Therefore the assert that checks the number of implemented bits and __NVIC_PRIO_BITS will always fail. The config option `configQEMU_DISABLE_INTERRUPT_PRIO_BITS_CHECK` should be defined in the `FreeRTOSConfig.h` for QEMU targets. Signed-off-by: Devaraj Ranganna <devaraj.ranganna@arm.com> * Use SHPR2 for calculating interrupt priority bits This removes the dependency on the secure software to mark the interrupt as non-secure. Signed-off-by: Gaurav Aggarwal <aggarg@amazon.com> --------- Signed-off-by: Devaraj Ranganna <devaraj.ranganna@arm.com> Signed-off-by: Gaurav Aggarwal <aggarg@amazon.com> Co-authored-by: Gaurav Aggarwal <aggarg@amazon.com> Co-authored-by: Gaurav-Aggarwal-AWS <33462878+aggarg@users.noreply.github.com>
2 years ago
/*-----------------------------------------------------------*/
Add Access Control List to MPU ports (#765) This PR adds Access Control to kernel objects on a per task basis to MPU ports. The following needs to be defined in the `FreeRTOSConfig.h` to enable this feature: ```c #define configUSE_MPU_WRAPPERS_V1 0 #define configENABLE_ACCESS_CONTROL_LIST 1 ``` This PR adds the following new APIs: ```c void vGrantAccessToTask( TaskHandle_t xTask, TaskHandle_t xTaskToGrantAccess ); void vRevokeAccessToTask( TaskHandle_t xTask, TaskHandle_t xTaskToRevokeAccess ); void vGrantAccessToSemaphore( TaskHandle_t xTask, SemaphoreHandle_t xSemaphoreToGrantAccess ); void vRevokeAccessToSemaphore( TaskHandle_t xTask, SemaphoreHandle_t xSemaphoreToRevokeAccess ); void vGrantAccessToQueue( TaskHandle_t xTask, QueueHandle_t xQueueToGrantAccess ); void vRevokeAccessToQueue( TaskHandle_t xTask, QueueHandle_t xQueueToRevokeAccess ); void vGrantAccessToQueueSet( TaskHandle_t xTask, QueueSetHandle_t xQueueSetToGrantAccess ); void vRevokeAccessToQueueSet( TaskHandle_t xTask, QueueSetHandle_t xQueueSetToRevokeAccess ); void vGrantAccessToEventGroup( TaskHandle_t xTask, EventGroupHandle_t xEventGroupToGrantAccess ); void vRevokeAccessToEventGroup( TaskHandle_t xTask, EventGroupHandle_t xEventGroupToRevokeAccess ); void vGrantAccessToStreamBuffer( TaskHandle_t xTask, StreamBufferHandle_t xStreamBufferToGrantAccess ); void vRevokeAccessToStreamBuffer( TaskHandle_t xTask, StreamBufferHandle_t xStreamBufferToRevokeAccess ); void vGrantAccessToMessageBuffer( TaskHandle_t xTask, MessageBufferHandle_t xMessageBufferToGrantAccess ); void vRevokeAccessToMessageBuffer( TaskHandle_t xTask, MessageBufferHandle_t xMessageBufferToRevokeAccess ); void vGrantAccessToTimer( TaskHandle_t xTask, TimerHandle_t xTimerToGrantAccess ); void vRevokeAccessToTimer( TaskHandle_t xTask, TimerHandle_t xTimerToRevokeAccess ); ``` An unprivileged task by default has access to itself only and no other kernel object. The application writer needs to explicitly grant an unprivileged task access to all the kernel objects it needs. The best place to do that is before starting the scheduler when all the kernel objects are created. For example, let's say an unprivileged tasks needs access to a queue and an event group, the application writer needs to do the following: ```c vGrantAccessToQueue( xUnprivilegedTaskHandle, xQueue ); vGrantAccessToEventGroup( xUnprivilegedTaskHandle, xEventGroup ); ``` The application writer MUST revoke all the accesses before deleting a task. Failing to do so will result in undefined behavior. In the above example, the application writer needs to make the following 2 calls before deleting the task: ```c vRevokeAccessToQueue( xUnprivilegedTaskHandle, xQueue ); vRevokeAccessToEventGroup( xUnprivilegedTaskHandle, xEventGroup ); ```
1 year ago
#if ( ( configUSE_MPU_WRAPPERS_V1 == 0 ) && ( configENABLE_ACCESS_CONTROL_LIST == 1 ) )
void vPortGrantAccessToKernelObject( TaskHandle_t xInternalTaskHandle,
int32_t lInternalIndexOfKernelObject ) /* PRIVILEGED_FUNCTION */
{
uint32_t ulAccessControlListEntryIndex, ulAccessControlListEntryBit;
xMPU_SETTINGS * xTaskMpuSettings;
ulAccessControlListEntryIndex = ( ( uint32_t ) lInternalIndexOfKernelObject / portACL_ENTRY_SIZE_BITS );
ulAccessControlListEntryBit = ( ( uint32_t ) lInternalIndexOfKernelObject % portACL_ENTRY_SIZE_BITS );
xTaskMpuSettings = xTaskGetMPUSettings( xInternalTaskHandle );
xTaskMpuSettings->ulAccessControlList[ ulAccessControlListEntryIndex ] |= ( 1U << ulAccessControlListEntryBit );
}
#endif /* #if ( ( configUSE_MPU_WRAPPERS_V1 == 0 ) && ( configENABLE_ACCESS_CONTROL_LIST == 1 ) ) */
/*-----------------------------------------------------------*/
#if ( ( configUSE_MPU_WRAPPERS_V1 == 0 ) && ( configENABLE_ACCESS_CONTROL_LIST == 1 ) )
void vPortRevokeAccessToKernelObject( TaskHandle_t xInternalTaskHandle,
int32_t lInternalIndexOfKernelObject ) /* PRIVILEGED_FUNCTION */
{
uint32_t ulAccessControlListEntryIndex, ulAccessControlListEntryBit;
xMPU_SETTINGS * xTaskMpuSettings;
ulAccessControlListEntryIndex = ( ( uint32_t ) lInternalIndexOfKernelObject / portACL_ENTRY_SIZE_BITS );
ulAccessControlListEntryBit = ( ( uint32_t ) lInternalIndexOfKernelObject % portACL_ENTRY_SIZE_BITS );
xTaskMpuSettings = xTaskGetMPUSettings( xInternalTaskHandle );
xTaskMpuSettings->ulAccessControlList[ ulAccessControlListEntryIndex ] &= ~( 1U << ulAccessControlListEntryBit );
}
#endif /* #if ( ( configUSE_MPU_WRAPPERS_V1 == 0 ) && ( configENABLE_ACCESS_CONTROL_LIST == 1 ) ) */
/*-----------------------------------------------------------*/
#if ( configUSE_MPU_WRAPPERS_V1 == 0 )
#if ( configENABLE_ACCESS_CONTROL_LIST == 1 )
BaseType_t xPortIsAuthorizedToAccessKernelObject( int32_t lInternalIndexOfKernelObject ) /* PRIVILEGED_FUNCTION */
{
uint32_t ulAccessControlListEntryIndex, ulAccessControlListEntryBit;
BaseType_t xAccessGranted = pdFALSE;
const xMPU_SETTINGS * xTaskMpuSettings = xTaskGetMPUSettings( NULL ); /* Calling task's MPU settings. */
ulAccessControlListEntryIndex = ( ( uint32_t ) lInternalIndexOfKernelObject / portACL_ENTRY_SIZE_BITS );
ulAccessControlListEntryBit = ( ( uint32_t ) lInternalIndexOfKernelObject % portACL_ENTRY_SIZE_BITS );
if( ( xTaskMpuSettings->ulTaskFlags & portTASK_IS_PRIVILEGED_FLAG ) == portTASK_IS_PRIVILEGED_FLAG )
{
xAccessGranted = pdTRUE;
}
else
{
if( ( xTaskMpuSettings->ulAccessControlList[ ulAccessControlListEntryIndex ] & ( 1U << ulAccessControlListEntryBit ) ) != 0 )
{
xAccessGranted = pdTRUE;
}
}
return xAccessGranted;
}
#else /* #if ( configENABLE_ACCESS_CONTROL_LIST == 1 ) */
BaseType_t xPortIsAuthorizedToAccessKernelObject( int32_t lInternalIndexOfKernelObject ) /* PRIVILEGED_FUNCTION */
{
( void ) lInternalIndexOfKernelObject;
/* If Access Control List feature is not used, all the tasks have
* access to all the kernel objects. */
return pdTRUE;
}
#endif /* #if ( configENABLE_ACCESS_CONTROL_LIST == 1 ) */
#endif /* #if ( configUSE_MPU_WRAPPERS_V1 == 0 ) */
/*-----------------------------------------------------------*/