9.5.3 Entry Calls

Reason: An exception during barrier evaluation is considered essentially a fatal error. All current entry callers are notified with a Program_Error. In a fault-tolerant system, a protected object might provide a Reset protected procedure, or equivalent, to support attempts to restore such a "broken" protected object to a reasonable state. 

Discussion: Note that the definition of when a task entry is open is based on the state of the (accepting) task, whereas the "openness" of a protected entry is defined only when it is explicitly checked, since the barrier expression needs to be evaluated. Implementation permissions are given (below) to allow implementations to evaluate the barrier expression more or less often than it is checked, but the basic semantic model presumes it is evaluated at the times when it is checked. 

Ramification: The return to the caller will generally not occur until the protected action completes, unless some other thread of control is given the job of completing the protected action and releasing the associated execution resource. 

Discussion: While servicing the entry queues of a protected object, no new calls can be added to any entry queue of the object, except due to an internal requeue (see 9.5.4). This is because the first step of a call on a protected entry is to start a new protected action, which implies acquiring (for exclusive read-write access) the execution resource associated with the protected object, which cannot be done while another protected action is already in progress.

Ramification: For a call on a protected entry, the caller is not blocked if the call is selected immediately, unless a requeue causes the call to be queued.

Implementation Note: {AI95-00114-01} In the case of an attempted cancellation due to abort, this removal might have to be performed by the calling task itself if the ceiling priority of the protected object is lower than the priority of the task initiating the abort. 

Reason: These permissions are intended to allow flexibility for implementations on multiprocessors. On a monoprocessor, which thread of control executes the protected action is essentially invisible, since the thread is not abortable in any case, and the "current_task" function is not guaranteed to work during a protected action (see C.7.1 C.7).

Ramification: Changes to variables referenced by an entry barrier that result from actions outside of a protected procedure or entry call on the protected object need not be "noticed." For example, if a global variable is referenced by an entry barrier, it should not be altered (except as part of a protected action on the object) any time after the barrier is first evaluated. In other words, globals can be used to "parameterize" a protected object, but they cannot reliably be used to control it after the first use of the protected object. 

Implementation Note: Note that even if a global variable is volatile, the implementation need only reevaluate a barrier if the global is updated during a protected action on the protected object. This ensures that an entry-open bit-vector implementation approach is possible, where the bit-vector is computed at the end of a protected action, rather than upon each entry call. 

Ramification: In other words, any side-effects of evaluating an entry barrier should be innocuous, since an entry barrier might be evaluated more or less often than is implied by the "official" dynamic semantics. 

Implementation Note: It is anticipated that when the number of entries is known to be small, all barriers will be evaluated any time one of them needs to be, to produce an "entry-open bit-vector." The appropriate bit will be tested when the entry is called, and only if the bit is false will a check be made to see whether the bit-vector might need to be recomputed. This should allow an implementation to maximize the performance of a call on an open entry, which seems like the most important case.

In addition to the entry-open bit-vector, an "is-valid" bit is needed per object, which indicates whether the current bit-vector setting is valid. A "depends-on-Count-attribute" bit is needed per type. The "is-valid" bit is set to false (as are all the bits of the bit-vector) when the protected object is first created, as well as any time an exception is propagated from computing the bit-vector. Is-valid would also be set false any time the Count is changed and "depends-on-Count-attribute" is true for the type, or a protected procedure or entry returns indicating it might have updated a variable referenced in some barrier.

A single procedure can be compiled to evaluate all of the barriers, set the entry-open bit-vector accordingly, and set the is-valid bit to true. It could have a "when others" handler to set them all false, and call a routine to propagate Program_Error to all queued callers.

For protected types where the number of entries is not known to be small, it makes more sense to evaluate a barrier only when the corresponding entry is checked to see if it is open. It isn't worth saving the state of the entry between checks, because of the space that would be required. Furthermore, the entry queues probably want to take up space only when there is actually a caller on them, so rather than an array of all entry queues, a linked list of nonempty entry queues make the most sense in this case, with the first caller on each entry queue acting as the queue header. 

Reason: Because cancellation of a protected entry call is a protected action (which helps make the Count attribute of a protected entry meaningful), it might not be practical to attempt the cancellation from the thread of control that initiated the cancellation. For example, if the cancellation is due to the expiration of a delay, it is unlikely that the handler of the timer interrupt could perform the necessary protected action itself (due to being on the interrupt level). Similarly, if the cancellation is due to an abort, it is possible that the task initiating the abort has a priority higher than the ceiling priority of the protected object (for implementations that support ceiling priorities). Similar considerations could apply in a multiprocessor situation.

Ramification: Given this, extra care is required if a reference to the Count attribute of an entry appears in the entry's own barrier. 

Reason: An entry is checked to see if it is open prior to queuing to maximize the performance of a call on an open entry. 

Ramification: A task can call its own entries, but the task will deadlock if the call is a simple entry call.