The kernel’s priority-based scheduler allows an application’s threads to share the CPU.
The scheduler determines which thread is allowed to execute at any point in time; this thread is known as the current thread.
There are various points in time when the scheduler is given an opportunity to change the identity of the current thread. These points are called reschedule points. Some potential reschedule points are:
return to thread context after processing an interrupt
when a running thread invokes
Whenever the scheduler changes the identity of the current thread, or when execution of the current thread is replaced by an ISR, the kernel first saves the current thread’s CPU register values. These register values get restored when the thread later resumes execution.
The kernel’s scheduler selects the highest priority ready thread to be the current thread. When multiple ready threads of the same priority exist, the scheduler chooses the one that has been waiting longest.
Execution of ISRs takes precedence over thread execution, so the execution of the current thread may be replaced by an ISR at any time unless interrupts have been masked. This applies to both cooperative threads and preemptive threads.
The kernel can be built with one of several choices for the ready queue implementation, offering different choices between code size, constant factor runtime overhead and performance scaling when many threads are added.
Simple linked-list ready queue (
The scheduler ready queue will be implemented as a simple unordered list, with very fast constant time performance for single threads and very low code size. This implementation should be selected on systems with constrained code size that will never see more than a small number (3, maybe) of runnable threads in the queue at any given time. On most platforms (that are not otherwise using the red/black tree) this results in a savings of ~2k of code size.
Red/black tree ready queue (
The scheduler ready queue will be implemented as a red/black tree. This has rather slower constant-time insertion and removal overhead, and on most platforms (that are not otherwise using the red/black tree somewhere) requires an extra ~2kb of code. The resulting behavior will scale cleanly and quickly into the many thousands of threads.
Use this for applications needing many concurrent runnable threads (> 20 or so). Most applications won’t need this ready queue implementation.
Traditional multi-queue ready queue (
When selected, the scheduler ready queue will be implemented as the classic/textbook array of lists, one per priority (max 32 priorities).
This corresponds to the scheduler algorithm used in Zephyr versions prior to 1.12.
It incurs only a tiny code size overhead vs. the “dumb” scheduler and runs in O(1) time in almost all circumstances with very low constant factor. But it requires a fairly large RAM budget to store those list heads, and the limited features make it incompatible with features like deadline scheduling that need to sort threads more finely, and SMP affinity which need to traverse the list of threads.
Typical applications with small numbers of runnable threads probably want the DUMB scheduler.
The wait_q abstraction used in IPC primitives to pend threads for later wakeup shares the same backend data structure choices as the scheduler, and can use the same options.
Scalable wait_q implementation (
When selected, the wait_q will be implemented with a balanced tree. Choose this if you expect to have many threads waiting on individual primitives. There is a ~2kb code size increase over
CONFIG_WAITQ_DUMB(which may be shared with
CONFIG_SCHED_SCALABLE) if the red/black tree is not used elsewhere in the application, and pend/unpend operations on “small” queues will be somewhat slower (though this is not generally a performance path).
Simple linked-list wait_q (
When selected, the wait_q will be implemented with a doubly-linked list. Choose this if you expect to have only a few threads blocked on any single IPC primitive.
Cooperative Time Slicing¶
Once a cooperative thread becomes the current thread, it remains the current thread until it performs an action that makes it unready. Consequently, if a cooperative thread performs lengthy computations, it may cause an unacceptable delay in the scheduling of other threads, including those of higher priority and equal priority.
To overcome such problems, a cooperative thread can voluntarily relinquish the CPU from time to time to permit other threads to execute. A thread can relinquish the CPU in two ways:
k_yield()puts the thread at the back of the scheduler’s prioritized list of ready threads, and then invokes the scheduler. All ready threads whose priority is higher or equal to that of the yielding thread are then allowed to execute before the yielding thread is rescheduled. If no such ready threads exist, the scheduler immediately reschedules the yielding thread without context switching.
k_sleep()makes the thread unready for a specified time period. Ready threads of all priorities are then allowed to execute; however, there is no guarantee that threads whose priority is lower than that of the sleeping thread will actually be scheduled before the sleeping thread becomes ready once again.
Preemptive Time Slicing¶
Once a preemptive thread becomes the current thread, it remains the current thread until a higher priority thread becomes ready, or until the thread performs an action that makes it unready. Consequently, if a preemptive thread performs lengthy computations, it may cause an unacceptable delay in the scheduling of other threads, including those of equal priority.
To overcome such problems, a preemptive thread can perform cooperative time slicing (as described above), or the scheduler’s time slicing capability can be used to allow other threads of the same priority to execute.
The scheduler divides time into a series of time slices, where slices are measured in system clock ticks. The time slice size is configurable, but this size can be changed while the application is running.
At the end of every time slice, the scheduler checks to see if the current
thread is preemptible and, if so, implicitly invokes
on behalf of the thread. This gives other ready threads of the same priority
the opportunity to execute before the current thread is scheduled again.
If no threads of equal priority are ready, the current thread remains
the current thread.
Threads with a priority higher than specified limit are exempt from preemptive time slicing, and are never preempted by a thread of equal priority. This allows an application to use preemptive time slicing only when dealing with lower priority threads that are less time-sensitive.
The kernel’s time slicing algorithm does not ensure that a set of equal-priority threads receive an equitable amount of CPU time, since it does not measure the amount of time a thread actually gets to execute. For example, a thread may become the current thread just before the end of a time slice and then immediately have to yield the CPU. However, the algorithm does ensure that a thread never executes for longer than a single time slice without being required to yield.
A preemptible thread that does not wish to be preempted while performing
a critical operation can instruct the scheduler to temporarily treat it
as a cooperative thread by calling
k_sched_lock(). This prevents
other threads from interfering while the critical operation is being performed.
Once the critical operation is complete the preemptible thread must call
k_sched_unlock() to restore its normal, preemptible status.
If a thread calls
k_sched_lock() and subsequently performs an
action that makes it unready, the scheduler will switch the locking thread out
and allow other threads to execute. When the locking thread again
becomes the current thread, its non-preemptible status is maintained.
Locking out the scheduler is a more efficient way for a preemptible thread to prevent preemption than changing its priority level to a negative value.
When enabled (see
CONFIG_NUM_METAIRQ_PRIORITIES), there is a special
subclass of cooperative priorities at the highest (numerically lowest)
end of the priority space: meta-IRQ threads. These are scheduled
according to their normal priority, but also have the special ability
to preempt all other threads (and other meta-irq threads) at lower
priorities, even if those threads are cooperative and/or have taken a
This behavior makes the act of unblocking a meta-IRQ thread (by any means, e.g. creating it, calling k_sem_give(), etc.) into the equivalent of a synchronous system call when done by a lower priority thread, or an ARM-like “pended IRQ” when done from true interrupt context. The intent is that this feature will be used to implement interrupt “bottom half” processing and/or “tasklet” features in driver subsystems. The thread, once woken, will be guaranteed to run before the current CPU returns into application code.
Unlike similar features in other OSes, meta-IRQ threads are true threads and run on their own stack (which must be allocated normally), not the per-CPU interrupt stack. Design work to enable the use of the IRQ stack on supported architectures is pending.
Note that because this breaks the promise made to cooperative threads by the Zephyr API (namely that the OS won’t schedule other thread until the current thread deliberately blocks), it should be used only with great care from application code. These are not simply very high priority threads and should not be used as such.
A thread can call
k_sleep() to delay its processing
for a specified time period. During the time the thread is sleeping
the CPU is relinquished to allow other ready threads to execute.
Once the specified delay has elapsed the thread becomes ready
and is eligible to be scheduled once again.
A sleeping thread can be woken up prematurely by another thread using
k_wakeup(). This technique can sometimes be used
to permit the secondary thread to signal the sleeping thread
that something has occurred without requiring the threads
to define a kernel synchronization object, such as a semaphore.
Waking up a thread that is not sleeping is allowed, but has no effect.
A thread can call
k_busy_wait() to perform a
that delays its processing for a specified time period
without relinquishing the CPU to another ready thread.
A busy wait is typically used instead of thread sleeping when the required delay is too short to warrant having the scheduler context switch from the current thread to another thread and then back again.
Use cooperative threads for device drivers and other performance-critical work.
Use cooperative threads to implement mutually exclusion without the need for a kernel object, such as a mutex.
Use preemptive threads to give priority to time-sensitive processing over less time-sensitive processing.