Real-time operating system

A real-time operating system (RTOS) is a class of operating system intended for real-time applications. Examples include embedded applications (programmable thermostats, household appliance controllers, mobile telephones), industrial robots, industrial control (see SCADA), and scientific research equipment.

A RTOS facilitates the creation of a real-time system, but does not guarantee the finished product will be real-time; this requires correct development of the software. A RTOS does not necessarily have high throughput; rather, a RTOS provides facilities which, if used properly, guarantee that system deadlines can be met generally ("soft real-time") or deterministically ("hard real-time"). A RTOS will typically use specialized scheduling algorithms in order to provide the real-time developer with the tools necessary to produce deterministic behavior in the final system. A RTOS is valued more for how quickly and/or predictably it can respond to a particular event than for the given amount of work it can perform over time. Key factors in an RTOS are therefore minimal interrupt and thread switching latency.

An early example of a large-scale real-time operating system was the so-called "control program" developed by American Airlines and IBM for the Sabre Airline Reservations System.

Debate exists about what actually constitutes real-time computing.

Design philosophies

There are two basic designs:

An event-driven operating system only changes tasks when an event requires service.
A time-sharing design switches tasks on a clock interrupt, as well as on events.

The time-sharing design switches tasks more often than is strictly necessary, but gives "smoother", more deterministic multitasking, the illusion that a user has sole use of a machine.

Early CPU designs required many cycles to switch tasks, during which the CPU could do nothing useful. So early OSes tried to minimize wasting CPU time by avoiding unnecessary task-switches as much as possible.

More recent CPUs take far less time to switch from one task to another (the extreme case is barrel processors that switch from one task to the next in zero cycles). More recent RTOSes almost invariably implement time-sharing scheduling in addition to priority driven pre-emptive scheduling.

Scheduling

In typical designs, a task has three states: running, ready and blocked. Most tasks are blocked, most of the time. Only one task per CPU is running. In simpler systems, the ready list is usually short, two or three tasks at most.

The real trick is designing the scheduler. Usually the data structure of the ready list in the scheduler is designed so as to minimize the worst-case length of time spent in the scheduler's critical section (during which preemption is inhibited, and, in some cases, all interrupts are disabled). But the choice of data structure depends also on the maximum number of tasks that can be on the ready list.

If there are never more than a few tasks on the ready list, then a simple unsorted bidirectional linked list of ready tasks is likely optimal. If the ready list usually contains only a few tasks but occasionally contains more, then the list should be sorted by priority, so that finding the highest priority task to run does not require iterating through the entire list. Inserting a task then requires walking the ready list until reaching either the end of the list, or a task of lower priority than that of the task being inserted. Care must be taken not to inhibit preemption during this entire search; the otherwise-long critical section should probably be divided into small pieces, so that if, during the insertion of a low priority task, an interrupt occurs that makes a high priority task ready, that high priority task can be inserted and run immediately (before the low priority task is inserted).

The critical response time, sometimes called the flyback time, is the time it takes to queue a new ready task and restore the state of the highest priority task. In a well-designed RTOS, readying a new task will take 3-20 instructions per ready queue entry, and restoration of the highest-priority ready task will take 5-30 instructions. On a 20MHz 68000 processor, task switch times run about 20 microseconds with two tasks ready. 100 MHz ARM CPUs switch in a few microseconds.

In more advanced real-time systems, real-time tasks share computing resources with many non-real-time tasks, and the ready list can be arbitrarily long. In such systems, a scheduler ready list implemented as a linked list would be inadequate.

Intertask communication and resource sharing

A significant problem that multitasking systems must address is sharing data and hardware resources among multiple tasks. It is usually "unsafe" for two tasks to access the same specific data or hardware resource simultaneously. ("Unsafe" means the results are inconsistent or unpredictable, particularly when one task is in the midst of changing a data collection. The view by another task is best done either before any change begins, or after changes are completely finished.) There are three common approaches to resolve this problem:

General-purpose operating systems usually do not allow user programs to mask (disable) interrupts, because the user program could control the CPU for as long as it wished. Modern CPUs make the interrupt disable control bit (or instruction) inaccessible in user mode to allow operating systems to prevent user tasks from doing this. Many embedded systems and RTOSs, however, allow the application itself to run in kernel mode for greater system call efficiency and also to permit the application to have greater control of the operating environment without requiring OS intervention.

On single-processor systems, if the application runs in kernel mode and can mask interrupts, often that is the best (lowest overhead) solution to preventing simultaneous access to a shared resource. While interrupts are masked, the current task has exclusive use of the CPU; no other task or interrupt can take control, so the critical section is effectively protected. When the task exits its critical section, it must unmask interrupts; pending interrupts, if any, will then execute. Temporarily masking interrupts should only be done when the longest path through the critical section is shorter than the desired maximum interrupt latency, or else this method will increase the system's maximum interrupt latency. Typically this method of protection is used only when the critical section is just a few source code lines long and contains no loops. This method is ideal for protecting hardware bitmapped registers when the bits are controlled by different tasks.

When the critical section is longer than a few source code lines or involves lengthy looping, an embedded/real-time programmer must resort to using mechanisms identical or similar to those available on general-purpose operating systems, such as semaphores and OS-supervised interprocess messaging. Such mechanisms involve system calls, and usually invoke the OS's dispatcher code on exit, so they can take many hundreds of CPU instructions to execute, while masking interrupts may take as few as three instructions on some processors. But for longer critical sections, there may be no choice; interrupts cannot be masked for long periods without increasing the system's interrupt latency.

A binary semaphore is either locked or unlocked. When it is locked, a queue of tasks can wait for the semaphore. Typically a task can set a timeout on its wait for a semaphore. Problems with semaphore designs are well known: priority inversion and deadlocks.

In priority inversion, a high priority task waits because a low priority task has a semaphore. A typical solution is to have the task that has a semaphore run at (inherit) the priority of the highest waiting task. But this simplistic approach fails when there are multiple levels of waiting (A waits for a binary semaphore locked by B, which waits for a binary semaphore locked by C). Handling multiple levels of inheritance without introducing instability in cycles is not straightforward.

In a deadlock, two or more tasks lock the same number of binary semaphores and then wait forever (no timeout) for other binary semaphores, creating a cyclic dependency graph. The simplest deadlock scenario occurs when two tasks lock two semaphores in lockstep, but in the opposite order. Deadlock is usually prevented by careful design, or by having floored semaphores (which pass control of a semaphore to the higher priority task on defined conditions).

The other approach to resource sharing is for tasks to send messages. In this paradigm, the resource is managed directly by only one task; when another task wants to interrogate or manipulate the resource, it sends a message to the managing task. This paradigm suffers from similar problems as binary semaphores: Priority inversion occurs when a task is working on a low-priority message, and ignores a higher-priority message (or a message originating indirectly from a high priority task) in its in-box. Protocol deadlocks occur when two or more tasks wait for each other to send response messages.

Although their real-time behavior is less crisp than semaphore systems, simple message-based systems usually do not have protocol deadlock hazards, and are generally better-behaved than semaphore systems.

Interrupt handlers and the scheduler

Since an interrupt handler blocks the highest priority task from running, and since real time operating systems are designed to keep thread latency to a minimum, interrupt handlers are typically kept as short as possible. The interrupt handler defers all interaction with the hardware as long as possible; typically all that is necessary is to acknowledge or disable the interrupt (so that it won't occur again when the interrupt handler returns). The interrupt handler then queues work to be done at a lower priority level, often by unblocking a driver task (through releasing a semaphore or sending a message). The scheduler often provides the ability to unblock a task from interrupt handler context.

Memory allocation

Memory allocation is even more critical in a RTOS than in other operating systems.

Firstly, speed of allocation is important. A standard memory allocation scheme scans a linked list of indeterminate length to find a suitable free memory block; however, this is unacceptable as memory allocation has to occur in a fixed time in a RTOS.

Secondly, memory can become fragmented as free regions become separated by regions that are in use. This can cause a program to stall, unable to get memory, even though there is theoretically enough available. Memory allocation algorithms that slowly accumulate fragmentation may work fine for desktop machines—when rebooted every month or so—but are unacceptable for embedded systems that often run for years without rebooting.

The simple fixed-size-blocks algorithm works astonishingly well for simple embedded systems.

See memory allocation for more details.

Use of memory in deeply embedded systems

Most embedded operating systems store an entire binary code image in ROM, from which it also executes. That is, unlike PC-style operating systems, code is never stored in RAM. The reason for this is that the code is never expected to change, as embedded devices usually are dedicated to a specific purpose. The behaviour of the system is completely specified at design-time.

Example RTOSes

Open source:

Proprietary:

Various companies also sell customized versions of Linux with added real-time functionality. On October 8, 2004, Montavista submitted patches to the Linux kernel mailing list to enhance Linux as a real time operating system.

Ingo Molnar is maintaining a realtime preemption patch that further increases Linux's realtime capabilities.

See also: List of operating systems

External links

Embedded systems | Operating system technology | Real-time operating systems

Gourt Home::Gourt :: Real-time operating system