7.3 Priority Inheritance Pattern
The Priority Inheritance Pattern reduces priority inversion by manipulating the executing priorities of tasks that lock resources. While not an ideal solution, it significantly reduces priority inversion at a relatively low run-time overhead cost.
The problem of unbounded priority inversion is a very real one and has accounted for many difficult-to-identify system failures. In systems running many tasks, such problems may not be at all obvious, and typically the only symptom is that occasionally the system fails to meet one or more deadlines. The Priority Inheritance Pattern is a simple, low-overhead solution for limiting the priority inversion to at most a single levelthat is, at most, a task will only be blocked by a single, lower-priority task owning a needed resource.
The unbounded priority inversion problem is discussed in the chapter introduction in some detail. The problem addressed by this pattern is to bound the maximum amount of priority inversion.
7.3.3 Pattern Structure
Figure 7-6 shows the structure of the pattern. The basic elements of this pattern are familiar: Scheduler, Abstract Task, Task Control Block, and so on. This can be thought of as an elaborated subset of the Static Priority Pattern, presented in Chapter 5. Note the use of the «frozen» constraint applied to the Task Control Block's nominalPriority attribute. This means the attribute is unchangeable once the object is created.
Figure 7-6: Priority Inheritance Pattern
7.3.4 Collaboration Roles
The Abstract Thread class is an abstract (noninstantiable) superclass for Concrete Thread. Abstract Thread associates with the Scheduler. Since Concrete Thread is a subclass, it has the same interface to the Scheduler as the Abstract Thread. This enforces interface compliance. The Abstract Thread is an «active» object, meaning that when it is created, it creates an OS thread in which to run. It contains (that is, it has composition relations with) more primitive application objects that execute in the thread of the composite «active» object.
The Concrete Thread is an «active» object most typically constructed to contain passive "semantic" objects (via the composition relation) that do the real work of the system. The Concrete Thread object provides a straightforward means of attaching these semantic objects into the concurrency architecture. Concrete Thread is an instantiable subclass of Abstract Thread.
The Mutex is a mutual exclusion semaphore object that permits only a single caller through at a time. The operations of the Shared Resource invoke it whenever a relevant service is called, locking it prior to starting the service and unlocking it once the service is complete. Threads that attempt to invoke a service when the services are already locked become blocked until the Mutex is in its unlocked state. This is done by the Mutex semaphore signaling the Scheduler that a call attempt was made by the currently active thread, the Mutex ID (necessary to unlock it later when the Mutex is released), and the entry pointthe place at which to continue execution of the Thread.
This object orchestrates the execution of multiple threads based on their priority according to a simple rule: Always run the ready thread with the highest priority. When the «active» Thread object is created, it (or its creator) calls the createThread operation to create a thread for the «active» object. Whenever this thread is executed by the Scheduler, it calls the StartAddr address (except when the thread has been blocked or preempted, in which case it calls the EntryPoint address).
In this pattern, the Scheduler has some special duties when the Mutex signals an attempt to access a locked resource: Specifically, it must block the requesting task (done by stopping that task and placing a reference to it in the Blocked Queue (not shownfor details of the Blocked Queue, see Static Priority Pattern in Chapter 5), and it must elevate the priority of the task owning the resource to that of the highest priority Thread being blocked. This is easy to determine since the Blocked Queue is a priority FIFOthe highest-priority blocked task is the first one in that queue. Similarly, when the Thread releases the resource, the Scheduler must lower its priority back to its nominal priority.
A Shared Resource is an object shared by one or more Threads. For the system to operate properly in all cases, all shared resources must either be reentrant (meaning that corruption from simultaneous access cannot occur) or they must be protected. In the case of a protected resource, when a Thread attempts to use the resource, the associated Mutex semaphore is checked, and if locked, the calling task is placed into the Blocked Queue. The task is terminated with its reentry point noted in the TCB.
Task Control Block
The TCB contains the scheduling information for its corresponding Thread object. This includes the priority of the thread, the default start address and the current entry address, if it was preempted or blocked prior to completion. The Scheduler maintains a TCB object for each existing Thread. Note that TCB typically also has a reference off to a call and parameter stack for its Thread, but that level of detail is not shown in Figure 7-6. The TCB tracks both the current priority of the thread (which may have been elevated due to resource access and blocking) and its nominal priority.
The Priority Inheritance Pattern handles well the problem of priority inversion when at most a single resource is locked at any given time and prevents unbounded priority inversion in this case. This is illustrated in Figure 7-7. With naïve priority management, Task 1, the highest-priority task in the system, is delayed from execution until Task 2 has completed. Using the Priority Inheritance Pattern, Task 1 completes as early as possible.
When there are multiple resources that may be locked at any time, this pattern exhibits behavior called chain blocking. That is, one task may block another, which blocks another, and so on. This is illustrated in the only slightly more complex example in Figure 7-8. The timing diagram in Figure 7-8b shows that Task 1 is blocked by Task 2 and Task 3 at Point G.
Figure 7-7: Priority Inheritance Pattern
Figure 7-8: Priority Inheritance Pattern
In general, the Priority Inheritance Pattern greatly reduces unbounded blocking. In fact, though, the number of blocked tasks at any given time is bounded only by the lesser of the number of tasks and the number of currently locked resources. There is a small amount of overhead to pay when tasks are blocked or unblocked to manage the elevation or depression of the priority of the tasks involved. Computation of a single task's worst-case blocking time involves computation of the worst-case chain blocking of all tasks of lesser priority.
This pattern does not address deadlock issues at all, so it is still possible to construct task models using this pattern that have deadlock.
Another consequence of the use of the priority inheritance patterns (Priority Inheritance Pattern, Highest Locker Pattern, and Priority Ceiling Pattern) is the overhead. The use of semaphores and blocking involves task switching whenever a locked mutex is requested and another task switch whenever a waited-for mutex is released. In addition, the acts of blocking and unblocking tasks during those task context switches involves the manipulation of priority queues. Further, the use of priority inheritance means that there is some overhead in the escalation and deescalation of priorities. If blocking occurs infrequently, then this overhead will be slight, but if there is a great deal of contention for resources, then the overhead can be severe.
7.3.6 Implementation Strategies
Some RTOS directly support the notion of priority inheritance, and so it is very little work to use this pattern with such an RTOS. If you are using an RTOS that does not support it, or if you are writing your own RTOS, then you must extend the RTOS (many RTOSs have API for just this purpose) to call your own function when the mutex blocks a task on a resource. The Scheduler must be able to identify the priority of the thread being blocked (a simple matter because it is in the Task Control Block for the task) in order to elevate the priority of the task currently owning the resource.
It is possible to build in the nominal priority as a constant attribute of the Concrete Thread. When the Concrete Thread always runs at a given priority, then the constructor of the «active» object should do exactly that. Otherwise, the creator of that active object should specify the priority at which that task should run.
In virtually all other ways, the implementation is very similar to the implementation of standard concurrency patterns, such as the Static Priority Pattern presented in Chapter 5.
7.3.7 Related Patterns
The Priority Inheritance Pattern exists to help solve a particular problem peculiar to priority-based preemption multitasking, so all of the concurrency patterns having to do with that style of multitasking can be mixed with this pattern.
While this pattern is lightweight, it greatly reduces priority inversion in multitasking systems. However, there are other approaches that can reduce it further, such as Priority Ceiling Pattern and Highest Locker Pattern. In addition, Priority Ceiling Pattern also removes the possibility of deadlock.
7.3.8 Sample Model
Figure 7-9 provides an example to illustrate how the Priority Inheritance Pattern works. States of the objects are shown using standard UMLthat is, as state marks on the instance lifelines. Some of the returns are shown, again using standard UML dashed lines. Showing that a call cannot complete is indicated with a large X on the callnot standard UML, but clear as to its interpretation.
Figure 7-9: Priority Inheritance Pattern
The flow of the scenario in Figure 7-9b is straightforward. All tasks begin the scenario in the Idle state. Then, at point A, the FilteringThread task becomes ready to run. It runs at its nominal priority, which is LOW (the priority of the thread is shown inside square brackets in the Running state markagain, not quite standard UML, but parsimonious). It then calls the resource SensorData that then enters the Locked state.
At point B, the ValveMonitor task becomes ready to run. It preempts the FilteringThread because the former is of higher priority. The ValveMonitor task runs for a while, but at point C, task DataAcqThread becomes ready to run. Since it is the highest priority, it preempts the ValveMonitor thread. DataAcqThread object then tries to access the SensorData object and finds that it cannot because the latter is locked with a Mutex semaphore (not shown in the scenario). The Scheduler then blocks the DataAcqThread thread and runs the FilteringThread at the same priority as DataAcqThread because the FilteringThread inherits the priority from the highest blocking taskin this case the DataAcqThread task. Note at this point, the medium-priority task, ValveMonitor, is in the state Waiting. Without priority inheritance, if DataAcqThread is blocked, the ValveMonitor would run because it has the next highest priority.
At point D, FilteringThread's use of the resource is complete, and it releases the resource (done at the end of the SensorData.gimme operation). As it returns, the Mutex signals the Scheduler that it is now available, so the Scheduler deescalates FilteringThread's priority to its nominal value (LOW) and unblocks the highest-priority task, DataAcqThread. This task now runs to completion and returns. The Scheduler then runs the next highest-priority waiting task, ValveMonitor, which runs until it is done and returns. Finally, the lowest-priority task, FilteringThread, gets to complete.
The worst-case blocking time for the DataAcqThread task is then the amount of time that FilteringThread locks the SensorData resource. Without the Priority Inheritance Pattern, the worst-case blocking for DataAcqThread task would be the amount of time FilteringThread locks the SensorData resource plus the amount of time that ValveMonitor executes.