.. COMMENT: COPYRIGHT (c) 1988-2008. .. COMMENT: On-Line Applications Research Corporation (OAR). .. COMMENT: All rights reserved. Scheduling Concepts ################### .. index:: scheduling .. index:: task scheduling Introduction ============ 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 state. 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 migrations. 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. Preemption ---------- .. 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. Timeslicing ----------- .. 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 non-existent. 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 task. 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 conditions: - 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.