ESP-IDF FreeRTOS SMP Changes

Overview

The vanilla FreeRTOS is designed to run on a single core. However the ESP32 is dual core containing a Protocol CPU (known as CPU 0 or PRO_CPU) and an Application CPU (known as CPU 1 or APP_CPU). The two cores are identical in practice and share the same memory. This allows the two cores to run tasks interchangeably between them.

The ESP-IDF FreeRTOS is a modified version of vanilla FreeRTOS which supports symmetric multiprocessing (SMP). ESP-IDF FreeRTOS is based on the Xtensa port of FreeRTOS v10.2.0. This guide outlines the major differences between vanilla FreeRTOS and ESP-IDF FreeRTOS. The API reference for vanilla FreeRTOS can be found via https://www.freertos.org/a00106.html

For information regarding features that are exclusive to ESP-IDF FreeRTOS, see ESP-IDF FreeRTOS Additions.

Tasks and Task Creation: Use xTaskCreatePinnedToCore() or xTaskCreateStaticPinnedToCore() to create tasks in ESP-IDF FreeRTOS. The last parameter of the two functions is xCoreID. This parameter specifies which core the task is pinned to. Acceptable values are 0 for PRO_CPU, 1 for APP_CPU, or tskNO_AFFINITY which allows the task to run on both.

Round Robin Scheduling: The ESP-IDF FreeRTOS scheduler will skip tasks when implementing Round-Robin scheduling between multiple tasks in the Ready state that are of the same priority. To avoid this behavior, ensure that those tasks either enter a blocked state, or are distributed across a wider range of priorities.

Scheduler Suspension: Suspending the scheduler in ESP-IDF FreeRTOS will only affect the scheduler on the the calling core. In other words, calling vTaskSuspendAll() on PRO_CPU will not prevent APP_CPU from scheduling, and vice versa. Use critical sections or semaphores instead for simultaneous access protection.

Tick Interrupt Synchronicity: Tick interrupts of PRO_CPU and APP_CPU are not synchronized. Do not expect to use vTaskDelay() or vTaskDelayUntil() as an accurate method of synchronizing task execution between the two cores. Use a counting semaphore instead as their context switches are not tied to tick interrupts due to preemption.

Critical Sections & Disabling Interrupts: In ESP-IDF FreeRTOS, critical sections are implemented using mutexes. Entering critical sections involve taking a mutex, then disabling the scheduler and interrupts of the calling core. However the other core is left unaffected. If the other core attemps to take same mutex, it will spin until the calling core has released the mutex by exiting the critical section.

Floating Point Arithmetic: The ESP32 supports hardware acceleration of single precision floating point arithmetic (float). However the use of hardware acceleration leads to some behavioral restrictions in ESP-IDF FreeRTOS. Therefore, tasks that utilize float will automatically be pinned to a core if not done so already. Furthermore, float cannot be used in interrupt service routines.

Thread Local Storage Pointers & Deletion Callbacks: Deletion callbacks are called automatically during task deletion and are used to free memory pointed to by TLSP. Call vTaskSetThreadLocalStoragePointerAndDelCallback() to set TLSP and Deletion Callbacks.

Configuring ESP-IDF FreeRTOS: Several aspects of ESP-IDF FreeRTOS can be set in the project configuration (idf.py menuconfig) such as running ESP-IDF in Unicore (single core) Mode, or configuring the number of Thread Local Storage Pointers each task will have.

It is not necessary to manually start the FreeRTOS scheduler by calling vTaskStartScheduler(). In ESP-IDF the scheduler is started by the 应用程序的启动流程 and is already running when the app_main function is called (see Running the main task for details).

Tasks and Task Creation

Tasks in ESP-IDF FreeRTOS are designed to run on a particular core, therefore two new task creation functions have been added to ESP-IDF FreeRTOS by appending PinnedToCore to the names of the task creation functions in vanilla FreeRTOS. The vanilla FreeRTOS functions of xTaskCreate() and xTaskCreateStatic() have led to the addition of xTaskCreatePinnedToCore() and xTaskCreateStaticPinnedToCore() in ESP-IDF FreeRTOS

For more details see freertos/tasks.c

The ESP-IDF FreeRTOS task creation functions are nearly identical to their vanilla counterparts with the exception of the extra parameter known as xCoreID. This parameter specifies the core on which the task should run on and can be one of the following values.

  • 0 pins the task to PRO_CPU

  • 1 pins the task to APP_CPU

  • tskNO_AFFINITY allows the task to be run on both CPUs

For example xTaskCreatePinnedToCore(tsk_callback, “APP_CPU Task”, 1000, NULL, 10, NULL, 1) creates a task of priority 10 that is pinned to APP_CPU with a stack size of 1000 bytes. It should be noted that the uxStackDepth parameter in vanilla FreeRTOS specifies a task’s stack depth in terms of the number of words, whereas ESP-IDF FreeRTOS specifies the stack depth in terms of bytes.

Note that the vanilla FreeRTOS functions xTaskCreate() and xTaskCreateStatic() have been defined in ESP-IDF FreeRTOS as inline functions which call xTaskCreatePinnedToCore() and xTaskCreateStaticPinnedToCore() respectively with tskNO_AFFINITY as the xCoreID value.

Each Task Control Block (TCB) in ESP-IDF stores the xCoreID as a member. Hence when each core calls the scheduler to select a task to run, the xCoreID member will allow the scheduler to determine if a given task is permitted to run on the core that called it.

Scheduling

The vanilla FreeRTOS implements scheduling in the vTaskSwitchContext() function. This function is responsible for selecting the highest priority task to run from a list of tasks in the Ready state known as the Ready Tasks List (described in the next section). In ESP-IDF FreeRTOS, each core will call vTaskSwitchContext() independently to select a task to run from the Ready Tasks List which is shared between both cores. There are several differences in scheduling behavior between vanilla and ESP-IDF FreeRTOS such as differences in Round Robin scheduling, scheduler suspension, and tick interrupt synchronicity.

Round Robin Scheduling

Given multiple tasks in the Ready state and of the same priority, vanilla FreeRTOS implements Round Robin scheduling between each task. This will result in running those tasks in turn each time the scheduler is called (e.g. every tick interrupt). On the other hand, the ESP-IDF FreeRTOS scheduler may skip tasks when Round Robin scheduling multiple Ready state tasks of the same priority.

The issue of skipping tasks during Round Robin scheduling arises from the way the Ready Tasks List is implemented in FreeRTOS. In vanilla FreeRTOS, pxReadyTasksList is used to store a list of tasks that are in the Ready state. The list is implemented as an array of length configMAX_PRIORITIES where each element of the array is a linked list. Each linked list is of type List_t and contains TCBs of tasks of the same priority that are in the Ready state. The following diagram illustrates the pxReadyTasksList structure.

Vanilla FreeRTOS Ready Task List Structure

Illustration of FreeRTOS Ready Task List Data Structure

Each linked list also contains a pxIndex which points to the last TCB returned when the list was queried. This index allows the vTaskSwitchContext() to start traversing the list at the TCB immediately after pxIndex hence implementing Round Robin Scheduling between tasks of the same priority.

In ESP-IDF FreeRTOS, the Ready Tasks List is shared between cores hence pxReadyTasksList will contain tasks pinned to different cores. When a core calls the scheduler, it is able to look at the xCoreID member of each TCB in the list to determine if a task is allowed to run on calling the core. The ESP-IDF FreeRTOS pxReadyTasksList is illustrated below.

ESP-IDF FreeRTOS Ready Task List Structure

Illustration of FreeRTOS Ready Task List Data Structure in ESP-IDF

Therefore when PRO_CPU calls the scheduler, it will only consider the tasks in blue or purple. Whereas when APP_CPU calls the scheduler, it will only consider the tasks in orange or purple.

Although each TCB has an xCoreID in ESP-IDF FreeRTOS, the linked list of each priority only has a single pxIndex. Therefore when the scheduler is called from a particular core and traverses the linked list, it will skip all TCBs pinned to the other core and point the pxIndex at the selected task. If the other core then calls the scheduler, it will traverse the linked list starting at the TCB immediately after pxIndex. Therefore, TCBs skipped on the previous scheduler call from the other core would not be considered on the current scheduler call. This issue is demonstrated in the following illustration.

ESP-IDF pxIndex Behavior

Illustration of pxIndex behavior in ESP-IDF FreeRTOS

Referring to the illustration above, assume that priority 9 is the highest priority, and none of the tasks in priority 9 will block hence will always be either in the running or Ready state.

1) PRO_CPU calls the scheduler and selects Task A to run, hence moves pxIndex to point to Task A

2) APP_CPU calls the scheduler and starts traversing from the task after pxIndex which is Task B. However Task B is not selected to run as it is not pinned to APP_CPU hence it is skipped and Task C is selected instead. pxIndex now points to Task C

3) PRO_CPU calls the scheduler and starts traversing from Task D. It skips Task D and selects Task E to run and points pxIndex to Task E. Notice that Task B isn’t traversed because it was skipped the last time APP_CPU called the scheduler to traverse the list.

4) The same situation with Task D will occur if APP_CPU calls the scheduler again as pxIndex now points to Task E

One solution to the issue of task skipping is to ensure that every task will enter a blocked state so that they are removed from the Ready Task List. Another solution is to distribute tasks across multiple priorities such that a given priority will not be assigned multiple tasks that are pinned to different cores.

Scheduler Suspension

In vanilla FreeRTOS, suspending the scheduler via vTaskSuspendAll() will prevent calls of vTaskSwitchContext from context switching until the scheduler has been resumed with xTaskResumeAll(). However servicing ISRs are still permitted. Therefore any changes in task states as a result from the current running task or ISRs will not be executed until the scheduler is resumed. Scheduler suspension in vanilla FreeRTOS is a common protection method against simultaneous access of data shared between tasks, whilst still allowing ISRs to be serviced.

In ESP-IDF FreeRTOS, xTaskSuspendAll() will only prevent calls of vTaskSwitchContext() from switching contexts on the core that called for the suspension. Hence if PRO_CPU calls vTaskSuspendAll(), APP_CPU will still be able to switch contexts. If data is shared between tasks that are pinned to different cores, scheduler suspension is NOT a valid method of protection against simultaneous access. Consider using critical sections (disables interrupts) or semaphores (does not disable interrupts) instead when protecting shared resources in ESP-IDF FreeRTOS.

In general, it’s better to use other RTOS primitives like mutex semaphores to protect against data shared between tasks, rather than vTaskSuspendAll().

Tick Interrupt Synchronicity

In ESP-IDF FreeRTOS, tasks on different cores that unblock on the same tick count might not run at exactly the same time due to the scheduler calls from each core being independent, and the tick interrupts to each core being unsynchronized.

In vanilla FreeRTOS the tick interrupt triggers a call to xTaskIncrementTick() which is responsible for incrementing the tick counter, checking if tasks which have called vTaskDelay() have fulfilled their delay period, and moving those tasks from the Delayed Task List to the Ready Task List. The tick interrupt will then call the scheduler if a context switch is necessary.

In ESP-IDF FreeRTOS, delayed tasks are unblocked with reference to the tick interrupt on PRO_CPU as PRO_CPU is responsible for incrementing the shared tick count. However tick interrupts to each core might not be synchronized (same frequency but out of phase) hence when PRO_CPU receives a tick interrupt, APP_CPU might not have received it yet. Therefore if multiple tasks of the same priority are unblocked on the same tick count, the task pinned to PRO_CPU will run immediately whereas the task pinned to APP_CPU must wait until APP_CPU receives its out of sync tick interrupt. Upon receiving the tick interrupt, APP_CPU will then call for a context switch and finally switches contexts to the newly unblocked task.

Therefore, task delays should NOT be used as a method of synchronization between tasks in ESP-IDF FreeRTOS. Instead, consider using a counting semaphore to unblock multiple tasks at the same time.

Critical Sections & Disabling Interrupts

Vanilla FreeRTOS implements critical sections with taskENTER_CRITICAL() which calls portDISABLE_INTERRUPTS(). This prevents preemptive context switches and servicing of ISRs during a critical section. Therefore, critical sections are used as a valid protection method against simultaneous access in vanilla FreeRTOS.

On the other hand, ESP32 has no hardware method for cores to disable each other’s interrupts. Calling portDISABLE_INTERRUPTS() will have no effect on the interrupts of the other core. Therefore, disabling interrupts is NOT a valid protection method against simultaneous access to shared data as it leaves the other core free to access the data even if the current core has disabled its own interrupts.

For this reason, ESP-IDF FreeRTOS implements critical sections using special mutexes, referred by portMUX_Type objects. These are implemented on top of a specific spinlock component. Calls to taskENTER_CRITICAL or taskEXIT_CRITICAL each provide a spinlock object as an argument. The spinlock is associated with a shared resource requiring access protection. When entering a critical section in ESP-IDF FreeRTOS, the calling core will disable interrupts similar to the vanilla FreeRTOS implementation, and will then take the spinlock and enter the critical section. The other core is unaffected at this point, unless it enters its own critical section and attempts to take the same spinlock. In that case it will spin until the lock is released. Therefore, the ESP-IDF FreeRTOS implementation of critical sections allows a core to have protected access to a shared resource without disabling the other core. The other core will only be affected if it tries to concurrently access the same resource.

The ESP-IDF FreeRTOS critical section functions have been modified as follows…

  • taskENTER_CRITICAL(mux), taskENTER_CRITICAL_ISR(mux), portENTER_CRITICAL(mux), portENTER_CRITICAL_ISR(mux) are all macro defined to call internal function vPortEnterCritical()

  • taskEXIT_CRITICAL(mux), taskEXIT_CRITICAL_ISR(mux), portEXIT_CRITICAL(mux), portEXIT_CRITICAL_ISR(mux) are all macro defined to call internal function vPortExitCritical()

  • portENTER_CRITICAL_SAFE(mux), portEXIT_CRITICAL_SAFE(mux) macro identifies the context of execution, i.e ISR or Non-ISR, and calls appropriate critical section functions (port*_CRITICAL in Non-ISR and port*_CRITICAL_ISR in ISR) in order to be in compliance with Vanilla FreeRTOS.

For more details see esp_hw_support/include/soc/spinlock.h, freertos/include/freertos/task.h, and freertos/tasks.c

It should be noted that when modifying vanilla FreeRTOS code to be ESP-IDF FreeRTOS compatible, it is trivial to modify the type of critical section called as they are all defined to call the same function. As long as the same spinlock is provided upon entering and exiting, the exact macro or function used for the call should not matter.

Floating Point Arithmetic

ESP-IDF FreeRTOS implements Lazy Context Switching for FPUs. In other words, the state of a core’s FPU registers are not immediately saved when a context switch occurs. Therefore, tasks that utilize float must be pinned to a particular core upon creation. If not, ESP-IDF FreeRTOS will automatically pin the task in question to whichever core the task was running on upon the task’s first use of float. Likewise due to Lazy Context Switching, only interrupt service routines of lowest priority (that is it the Level 1) can use float, higher priority interrupts do not support FPU usage.

ESP32 does not support hardware acceleration for double precision floating point arithmetic (double). Instead double is implemented via software hence the behavioral restrictions with regards to float do not apply to double. Note that due to the lack of hardware acceleration, double operations may consume significantly larger amount of CPU time in comparison to float.

Task Deletion

In FreeRTOS task deletion the freeing of task memory will occur immediately (within vTaskDelete()) if the task being deleted is not currently running or is not pinned to the other core (with respect to the core vTaskDelete() is called on). TLSP deletion callbacks will also run immediately if the same conditions are met.

However, calling vTaskDelete() to delete a task that is either currently running or pinned to the other core will still result in the freeing of memory being delegated to the Idle Task.

Thread Local Storage Pointers & Deletion Callbacks

Thread Local Storage Pointers (TLSP) are pointers stored directly in the TCB. TLSP allow each task to have its own unique set of pointers to data structures. However task deletion behavior in vanilla FreeRTOS does not automatically free the memory pointed to by TLSP. Therefore if the memory pointed to by TLSP is not explicitly freed by the user before task deletion, memory leak will occur.

ESP-IDF FreeRTOS provides the added feature of Deletion Callbacks. Deletion Callbacks are called automatically during task deletion to free memory pointed to by TLSP. Each TLSP can have its own Deletion Callback. Note that due to the to Task Deletion behavior, there can be instances where Deletion Callbacks are called in the context of the Idle Tasks. Therefore Deletion Callbacks should never attempt to block and critical sections should be kept as short as possible to minimize priority inversion.

Deletion callbacks are of type void (*TlsDeleteCallbackFunction_t)( int, void * ) where the first parameter is the index number of the associated TLSP, and the second parameter is the TLSP itself.

Deletion callbacks are set alongside TLSP by calling vTaskSetThreadLocalStoragePointerAndDelCallback(). Calling the vanilla FreeRTOS function vTaskSetThreadLocalStoragePointer() will simply set the TLSP’s associated Deletion Callback to NULL meaning that no callback will be called for that TLSP during task deletion. If a deletion callback is NULL, users should manually free the memory pointed to by the associated TLSP before task deletion in order to avoid memory leak.

For more details see FreeRTOS API reference.

Configuring ESP-IDF FreeRTOS

The ESP-IDF FreeRTOS can be configured in the project configuration menu (idf.py menuconfig) under Component Config/FreeRTOS. The following section highlights some of the ESP-IDF FreeRTOS configuration options. For a full list of ESP-IDF FreeRTOS configurations, see FreeRTOS

CONFIG_FREERTOS_UNICORE will run ESP-IDF FreeRTOS only on PRO_CPU. Note that this is not equivalent to running vanilla FreeRTOS. Note that this option may affect behavior of components other than freertos. For more details regarding the effects of running ESP-IDF FreeRTOS on a single core, search for occurences of CONFIG_FREERTOS_UNICORE in the ESP-IDF components.

CONFIG_FREERTOS_ASSERT_ON_UNTESTED_FUNCTION will trigger a halt in particular functions in ESP-IDF FreeRTOS which have not been fully tested in an SMP context.

CONFIG_FREERTOS_TASK_FUNCTION_WRAPPER will enclose all task functions within a wrapper function. In the case that a task function mistakenly returns (i.e. does not call vTaskDelete()), the call flow will return to the wrapper function. The wrapper function will then log an error and abort the application, as illustrated below:

E (25) FreeRTOS: FreeRTOS task should not return. Aborting now!
abort() was called at PC 0x40085c53 on core 0