Contents   Index   Search   Previous   Next

D.4 Entry Queuing Policies

     {8652/0074} [{queuing policy} This clause specifies a mechanism for a user to choose an entry queuing policy. It also defines two one such policies y. Other policies are implementation defined.]
Implementation defined: Implementation-defined queuing policies.


The form of a pragma Queuing_Policy is as follows:
  pragma Queuing_Policy(policy_identifier);

Legality Rules

   The policy_identifier shall be either FIFO_Queuing, Priority_Queuing or an implementation-defined identifier.

Post-Compilation Rules

   {configuration pragma (Queuing_Policy) [partial]} {pragma, configuration (Queuing_Policy) [partial]} A Queuing_Policy pragma is a configuration pragma.

Dynamic Semantics

   {queuing policy} [A queuing policy governs the order in which tasks are queued for entry service, and the order in which different entry queues are considered for service.] The queuing policy is specified by a Queuing_Policy pragma.
Ramification: The queuing policy includes entry queuing order, the choice among open alternatives of a selective_accept, and the choice among queued entry calls of a protected object when more than one entry_barrier condition is True.
   Two queuing policies, FIFO_Queuing and Priority_Queuing, are language defined. If no Queuing_Policy pragma appears in any of the program units comprising the partition, the queuing policy for that partition is FIFO_Queuing. The rules for this policy are specified in 9.5.3 and 9.7.1.
   The Priority_Queuing policy is defined as follows:
Reason: A task is blocked on an entry call if the entry call is simple, conditional, or timed. If the call came from the triggering_statement of an asynchronous_select, or a requeue thereof, then the task is not blocked on that call; such calls do not have their priority updated. Thus, there can exist many queued calls from a given task (caused by many nested ATC's), but a task can be blocked on only one call at a time.
A previous version of Ada 9X required queue reordering in the asynchronous_select case as well. If the call corresponds to a ``synchronous'' entry call, then the task is blocked while queued, and it makes good sense to move it up in the queue if its priority is raised.
However, if the entry call is ``asynchronous,'' that is, it is due to an asynchronous_select whose triggering_statement is an entry call, then the task is not waiting for this entry call, so the placement of the entry call on the queue is irrelevant to the rate at which the task proceeds.
Furthermore, when an entry is used for asynchronous_selects, it is almost certain to be a ``broadcast'' entry or have only one caller at a time. For example, if the entry is used to notify tasks of a mode switch, then all tasks on the entry queue would be signaled when the mode changes. Similarly, if it is indicating some interrupting event such as a control-C, all tasks sensitive to the interrupt will want to be informed that the event occurred. Hence, the order on such a queue is essentially irrelevant.
Given the above, it seems an unnecessary semantic and implementation complexity to specify that asynchronous queued calls are moved in response to dynamic priority changes. Furthermore, it is somewhat inconsistent, since the call was originally queued based on the active priority of the task, but dynamic priority changes are changing the base priority of the task, and only indirectly the active priority. We say explicitly that asynchronous queued calls are not affected by normal changes in active priority during the execution of an abortable_part. Saying that, if a change in the base priority affects the active priority, then we do want the calls reordered, would be inconsistent. It would also require the implementation to maintain a readily accessible list of all queued calls which would not otherwise be necessary.
Several rules were removed or simplified when we changed the rules so that calls due to asynchronous_selects are never moved due to intervening changes in active priority, be they due to protected actions, some other priority inheritance, or changes in the base priority.

Implementation Permissions

    Implementations are allowed to define other queuing policies, but need not support more than one such policy per partition.
Discussion: {8652/0116} This rule is really redundant, as 10.1.5 allows an implementation to limit the use of configuration pragmas to an empty environment. In that case, there would be no way to have multiple policies in a partition. In any case, the wording here really ought to be "...more than one queuing policy per partition.", since this part of the rule applies to all queuing policies, not just implementation-defined ones.

Implementation Advice

    The implementation should use names that end with ``_Queuing'' for implementation-defined queuing policies.

Contents   Index   Search   Previous   Next   Legal