path: root/c-user/scheduling_concepts.rst
diff options
Diffstat (limited to 'c-user/scheduling_concepts.rst')
1 files changed, 437 insertions, 0 deletions
diff --git a/c-user/scheduling_concepts.rst b/c-user/scheduling_concepts.rst
new file mode 100644
index 0000000..c7ccb2d
--- /dev/null
+++ b/c-user/scheduling_concepts.rst
@@ -0,0 +1,437 @@
+.. comment SPDX-License-Identifier: CC-BY-SA-4.0
+.. COMMENT: COPYRIGHT (c) 1988-2008.
+.. COMMENT: On-Line Applications Research Corporation (OAR).
+.. COMMENT: All rights reserved.
+Scheduling Concepts
+.. index:: scheduling
+.. index:: task scheduling
+The concept of scheduling in real-time systems dictates the ability to provide
+immediate response to specific external events, particularly the necessity of
+scheduling tasks to run within a specified time limit after the occurrence of
+an event. For example, software embedded in life-support systems used to
+monitor hospital patients must take instant action if a change in the patient's
+status is detected.
+The component of RTEMS responsible for providing this capability is
+appropriately called the scheduler. The scheduler's sole purpose is to
+allocate the all important resource of processor time to the various tasks
+competing for attention.
+Scheduling Algorithms
+.. index:: scheduling algorithms
+RTEMS provides a plugin framework which allows it to support multiple
+scheduling algorithms. RTEMS now includes multiple scheduling algorithms in the
+SuperCore and the user can select which of these they wish to use in their
+application. In addition, the user can implement their own scheduling
+algorithm and configure RTEMS to use it.
+Supporting multiple scheduling algorithms gives the end user the option to
+select the algorithm which is most appropriate to their use case. Most
+real-time operating systems schedule tasks using a priority based algorithm,
+possibly with preemption control. The classic RTEMS scheduling algorithm which
+was the only algorithm available in RTEMS 4.10 and earlier, is a priority based
+scheduling algorithm. This scheduling algoritm is suitable for single core
+(e.g. non-SMP) systems and is now known as the *Deterministic Priority
+Scheduler*. Unless the user configures another scheduling algorithm, RTEMS
+will use this on single core systems.
+Priority Scheduling
+.. index:: priority scheduling
+When using priority based scheduling, RTEMS allocates the processor using a
+priority-based, preemptive algorithm augmented to provide round-robin
+characteristics within individual priority groups. The goal of this algorithm
+is to guarantee that the task which is executing on the processor at any point
+in time is the one with the highest priority among all tasks in the ready
+When a task is added to the ready chain, it is placed behind all other tasks of
+the same priority. This rule provides a round-robin within priority group
+scheduling characteristic. This means that in a group of equal priority tasks,
+tasks will execute in the order they become ready or FIFO order. Even though
+there are ways to manipulate and adjust task priorities, the most important
+rule to remember is:
+.. note::
+ Priority based scheduling algorithms will always select the highest priority
+ task that is ready to run when allocating the processor to a task.
+Priority scheduling is the most commonly used scheduling algorithm. It should
+be used by applications in which multiple tasks contend for CPU time or other
+resources and there is a need to ensure certain tasks are given priority over
+other tasks.
+There are a few common methods of accomplishing the mechanics of this
+algorithm. These ways involve a list or chain of tasks in the ready state.
+- The least efficient method is to randomly place tasks in the ready chain
+ forcing the scheduler to scan the entire chain to determine which task
+ receives the processor.
+- A more efficient method is to schedule the task by placing it in the proper
+ place on the ready chain based on the designated scheduling criteria at the
+ time it enters the ready state. Thus, when the processor is free, the first
+ task on the ready chain is allocated the processor.
+- Another mechanism is to maintain a list of FIFOs per priority. When a task
+ is readied, it is placed on the rear of the FIFO for its priority. This
+ method is often used with a bitmap to assist in locating which FIFOs have
+ ready tasks on them.
+RTEMS currently includes multiple priority based scheduling algorithms as well
+as other algorithms which incorporate deadline. Each algorithm is discussed in
+the following sections.
+Deterministic Priority Scheduler
+This is the scheduler implementation which has always been in RTEMS. After the
+4.10 release series, it was factored into pluggable scheduler selection. It
+schedules tasks using a priority based algorithm which takes into account
+preemption. It is implemented using an array of FIFOs with a FIFO per
+priority. It maintains a bitmap which is used to track which priorities have
+ready tasks.
+This algorithm is deterministic (e.g. predictable and fixed) in execution time.
+This comes at the cost of using slightly over three (3) kilobytes of RAM on a
+system configured to support 256 priority levels.
+This scheduler is only aware of a single core.
+Simple Priority Scheduler
+This scheduler implementation has the same behaviour as the Deterministic
+Priority Scheduler but uses only one linked list to manage all ready tasks.
+When a task is readied, a linear search of that linked list is performed to
+determine where to insert the newly readied task.
+This algorithm uses much less RAM than the Deterministic Priority Scheduler but
+is *O(n)* where *n* is the number of ready tasks. In a small system with a
+small number of tasks, this will not be a performance issue. Reducing RAM
+consumption is often critical in small systems which are incapable of
+supporting a large number of tasks.
+This scheduler is only aware of a single core.
+Simple SMP Priority Scheduler
+This scheduler is based upon the Simple Priority Scheduler and is designed to
+have the same behaviour on a single core system. But this scheduler is capable
+of scheduling threads across multiple cores in an SMP system. When given a
+choice of replacing one of two threads at equal priority on different cores,
+this algorithm favors replacing threads which are preemptible and have executed
+the longest.
+This algorithm is non-deterministic. When scheduling, it must consider which
+tasks are to be executed on each core while avoiding superfluous task
+Earliest Deadline First Scheduler
+.. index:: earliest deadline first scheduling
+This is an alternative scheduler in RTEMS for single core applications. The
+primary EDF advantage is high total CPU utilization (theoretically up to
+100%). It assumes that tasks have priorities equal to deadlines.
+This EDF is initially preemptive, however, individual tasks may be declared
+not-preemptive. Deadlines are declared using only Rate Monotonic manager which
+goal is to handle periodic behavior. Period is always equal to deadline. All
+ready tasks reside in a single ready queue implemented using a red-black tree.
+This implementation of EDF schedules two different types of task priority types
+while each task may switch between the two types within its execution. If a
+task does have a deadline declared using the Rate Monotonic manager, the task
+is deadline-driven and its priority is equal to deadline. On the contrary if a
+task does not have any deadline or the deadline is cancelled using the Rate
+Monotonic manager, the task is considered a background task with priority equal
+to that assigned upon initialization in the same manner as for priority
+scheduler. Each background task is of a lower importance than each
+deadline-driven one and is scheduled when no deadline-driven task and no higher
+priority background task is ready to run.
+Every deadline-driven scheduling algorithm requires means for tasks to claim a
+deadline. The Rate Monotonic Manager is responsible for handling periodic
+execution. In RTEMS periods are equal to deadlines, thus if a task announces a
+period, it has to be finished until the end of this period. The call of
+``rtems_rate_monotonic_period`` passes the scheduler the length of oncoming
+deadline. Moreover, the ``rtems_rate_monotonic_cancel`` and
+``rtems_rate_monotonic_delete`` calls clear the deadlines assigned to the task.
+Constant Bandwidth Server Scheduling (CBS)
+.. index:: constant bandwidth server scheduling
+This is an alternative scheduler in RTEMS for single core applications. The
+CBS is a budget aware extension of EDF scheduler. The main goal of this
+scheduler is to ensure temporal isolation of tasks meaning that a task's
+execution in terms of meeting deadlines must not be influenced by other tasks
+as if they were run on multiple independent processors.
+Each task can be assigned a server (current implementation supports only one
+task per server). The server is characterized by period (deadline) and
+computation time (budget). The ratio budget/period yields bandwidth, which is
+the fraction of CPU to be reserved by the scheduler for each subsequent period.
+The CBS is equipped with a set of rules applied to tasks attached to servers
+ensuring that deadline miss because of another task cannot occur. In case a
+task breaks one of the rules, its priority is pulled to background until the
+end of its period and then restored again. The rules are:
+- Task cannot exceed its registered budget,
+- Task cannot be unblocked when a ratio between remaining budget and remaining
+ deadline is higher than declared bandwidth.
+The CBS provides an extensive API. Unlike EDF, the
+``rtems_rate_monotonic_period`` does not declare a deadline because it is
+carried out using CBS API. This call only announces next period.
+Scheduling Modification Mechanisms
+.. index:: scheduling mechanisms
+RTEMS provides four mechanisms which allow the user to alter the task
+scheduling decisions:
+- user-selectable task priority level
+- task preemption control
+- task timeslicing control
+- manual round-robin selection
+Each of these methods provides a powerful capability to customize sets of tasks
+to satisfy the unique and particular requirements encountered in custom
+real-time applications. Although each mechanism operates independently, there
+is a precedence relationship which governs the effects of scheduling
+modifications. The evaluation order for scheduling characteristics is always
+priority, preemption mode, and timeslicing. When reading the descriptions of
+timeslicing and manual round-robin it is important to keep in mind that
+preemption (if enabled) of a task by higher priority tasks will occur as
+required, overriding the other factors presented in the description.
+Task Priority and Scheduling
+.. index:: task priority
+The most significant task scheduling modification mechanism is the ability for
+the user to assign a priority level to each individual task when it is created
+and to alter a task's priority at run-time. RTEMS supports up to 255 priority
+levels. Level 255 is the lowest priority and level 1 is the highest.
+.. index:: preemption
+Another way the user can alter the basic scheduling algorithm is by
+manipulating the preemption mode flag (``RTEMS_PREEMPT_MASK``) of individual
+tasks. If preemption is disabled for a task (``RTEMS_NO_PREEMPT``), then the
+task will not relinquish control of the processor until it terminates, blocks,
+or re-enables preemption. Even tasks which become ready to run and possess
+higher priority levels will not be allowed to execute. Note that the
+preemption setting has no effect on the manner in which a task is scheduled.
+It only applies once a task has control of the processor.
+.. index:: timeslicing
+.. index:: round robin scheduling
+Timeslicing or round-robin scheduling is an additional method which can be used
+to alter the basic scheduling algorithm. Like preemption, timeslicing is
+specified on a task by task basis using the timeslicing mode flag
+(``RTEMS_TIMESLICE_MASK``). If timeslicing is enabled for a task
+(``RTEMS_TIMESLICE``), then RTEMS will limit the amount of time the task can
+execute before the processor is allocated to another task. Each tick of the
+real-time clock reduces the currently running task's timeslice. When the
+execution time equals the timeslice, RTEMS will dispatch another task of the
+same priority to execute. If there are no other tasks of the same priority
+ready to execute, then the current task is allocated an additional timeslice
+and continues to run. Remember that a higher priority task will preempt the
+task (unless preemption is disabled) as soon as it is ready to run, even if the
+task has not used up its entire timeslice.
+Manual Round-Robin
+.. index:: manual round robin
+The final mechanism for altering the RTEMS scheduling algorithm is called
+manual round-robin. Manual round-robin is invoked by using
+the ``rtems_task_wake_after`` directive with a time interval of
+``RTEMS_YIELD_PROCESSOR``. This allows a task to give up the processor and be
+immediately returned to the ready chain at the end of its priority group. If
+no other tasks of the same priority are ready to run, then the task does not
+lose control of the processor.
+Dispatching Tasks
+.. index:: dispatching
+The dispatcher is the RTEMS component responsible for allocating the processor
+to a ready task. In order to allocate the processor to one task, it must be
+deallocated or retrieved from the task currently using it. This involves a
+concept called a context switch. To perform a context switch, the dispatcher
+saves the context of the current task and restores the context of the task
+which has been allocated to the processor. Saving and restoring a task's
+context is the storing/loading of all the essential information about a task to
+enable it to continue execution without any effects of the interruption. For
+example, the contents of a task's register set must be the same when it is
+given the processor as they were when it was taken away. All of the
+information that must be saved or restored for a context switch is located
+either in the TCB or on the task's stacks.
+Tasks that utilize a numeric coprocessor and are created with the
+``RTEMS_FLOATING_POINT`` attribute require additional operations during a
+context switch. These additional operations are necessary to save and restore
+the floating point context of ``RTEMS_FLOATING_POINT`` tasks. To avoid
+unnecessary save and restore operations, the state of the numeric coprocessor
+is only saved when a ``RTEMS_FLOATING_POINT`` task is dispatched and that task
+was not the last task to utilize the coprocessor.
+Task State Transitions
+.. index:: task state transitions
+Tasks in an RTEMS system must always be in one of the five allowable task
+states. These states are: executing, ready, blocked, dormant, and
+A task occupies the non-existent state before a ``rtems_task_create`` has been
+issued on its behalf. A task enters the non-existent state from any other
+state in the system when it is deleted with the ``rtems_task_delete``
+directive. While a task occupies this state it does not have a TCB or a task
+ID assigned to it; therefore, no other tasks in the system may reference this
+When a task is created via the ``rtems_task_create`` directive it enters the
+dormant state. This state is not entered through any other means. Although
+the task exists in the system, it cannot actively compete for system resources.
+It will remain in the dormant state until it is started via the
+``rtems_task_start`` directive, at which time it enters the ready state. The
+task is now permitted to be scheduled for the processor and to compete for
+other system resources.
+.. figure:: ../images/c_user/states.png
+ :width: 70%
+ :align: center
+ :alt: Task State Transitions
+A task occupies the blocked state whenever it is unable to be scheduled to run.
+A running task may block itself or be blocked by other tasks in the system.
+The running task blocks itself through voluntary operations that cause the task
+to wait. The only way a task can block a task other than itself is with the
+``rtems_task_suspend`` directive. A task enters the blocked state due to any
+of the following conditions:
+- A task issues a ``rtems_task_suspend`` directive which blocks either itself
+ or another task in the system.
+- The running task issues a ``rtems_barrier_wait`` directive.
+- The running task issues a ``rtems_message_queue_receive`` directive with the
+ wait option and the message queue is empty.
+- The running task issues an ``rtems_event_receive`` directive with the wait
+ option and the currently pending events do not satisfy the request.
+- The running task issues a ``rtems_semaphore_obtain`` directive with the wait
+ option and the requested semaphore is unavailable.
+- The running task issues a ``rtems_task_wake_after`` directive which blocks
+ the task for the given time interval. If the time interval specified is
+ zero, the task yields the processor and remains in the ready state.
+- The running task issues a ``rtems_task_wake_when`` directive which blocks the
+ task until the requested date and time arrives.
+- The running task issues a ``rtems_rate_monotonic_period`` directive and must
+ wait for the specified rate monotonic period to conclude.
+- The running task issues a ``rtems_region_get_segment`` directive with the
+ wait option and there is not an available segment large enough to satisfy the
+ task's request.
+A blocked task may also be suspended. Therefore, both the suspension and the
+blocking condition must be removed before the task becomes ready to run again.
+A task occupies the ready state when it is able to be scheduled to run, but
+currently does not have control of the processor. Tasks of the same or higher
+priority will yield the processor by either becoming blocked, completing their
+timeslice, or being deleted. All tasks with the same priority will execute in
+FIFO order. A task enters the ready state due to any of the following
+- A running task issues a ``rtems_task_resume`` directive for a task that is
+ suspended and the task is not blocked waiting on any resource.
+- A running task issues a ``rtems_message_queue_send``,
+ ``rtems_message_queue_broadcast``, or a ``rtems_message_queue_urgent``
+ directive which posts a message to the queue on which the blocked task is
+ waiting.
+- A running task issues an ``rtems_event_send`` directive which sends an event
+ condition to a task which is blocked waiting on that event condition.
+- A running task issues a ``rtems_semaphore_release`` directive which releases
+ the semaphore on which the blocked task is waiting.
+- A timeout interval expires for a task which was blocked by a call to the
+ ``rtems_task_wake_after`` directive.
+- A timeout period expires for a task which blocked by a call to the
+ ``rtems_task_wake_when`` directive.
+- A running task issues a ``rtems_region_return_segment`` directive which
+ releases a segment to the region on which the blocked task is waiting and a
+ resulting segment is large enough to satisfy the task's request.
+- A rate monotonic period expires for a task which blocked by a call to the
+ ``rtems_rate_monotonic_period`` directive.
+- A timeout interval expires for a task which was blocked waiting on a message,
+ event, semaphore, or segment with a timeout specified.
+- A running task issues a directive which deletes a message queue, a semaphore,
+ or a region on which the blocked task is waiting.
+- A running task issues a ``rtems_task_restart`` directive for the blocked
+ task.
+- The running task, with its preemption mode enabled, may be made ready by
+ issuing any of the directives that may unblock a task with a higher priority.
+ This directive may be issued from the running task itself or from an ISR. A
+ ready task occupies the executing state when it has control of the CPU. A
+ task enters the executing state due to any of the following conditions:
+- The task is the highest priority ready task in the system.
+- The running task blocks and the task is next in the scheduling queue. The
+ task may be of equal priority as in round-robin scheduling or the task may
+ possess the highest priority of the remaining ready tasks.
+- The running task may reenable its preemption mode and a task exists in the
+ ready queue that has a higher priority than the running task.
+- The running task lowers its own priority and another task is of higher
+ priority as a result.
+- The running task raises the priority of a task above its own and the running
+ task is in preemption mode.