MFELDMAN@GWUVM.BITNET (Mike Feldman) (01/17/88)
First let me express my joy and satisfaction at the high technical plane of the list these days. Now to the issue: Flynn asserts that the Ada standard requires a pre-emptive scheduler, implying that a running task must be interrupted by a ready-to-run task of higher priority. I have always been under the impression that a "run-until-blocked" model was legal, that is, that a _running_ task was not required to be interrupted by another (even of higher priority) that merely became _ready_. On a uniprocessor, for example, is the scheduler required to _interrupt_ all running tasks and reschedule just because one task's delay has expired? Indeed, it seems possible that we won't even know the delay has expired until the currently running task "goes to sleep" for some reason. The ALRM (p. 9-16) is a bit vague here: If two tasks with different priorities are both _eligible for execution_ (emphasis mine) could be sensibly executed using the same physical proc- essors and the same other processing resources, then it cannot be the case that the task with the lower priority is executing while the task with the higher priority is not. The question: what does the ALRM mean by "eligible for execution"? At what point do we determine that the higher-priority task is eligible? Suppose a lower-priority task has had control of the CPU and no timeslicing is in effect; suppose that task doesn't do anything to block itself; suppose the higher priority task was waiting on a delay which expires. Now - does the paragraph above require that we immediately discover the "eligibility to run"? I have always understood the answer was "no" - that the decision is allowed to be deferred until the next "synchronization point". At _that_ point surely the higher-priority task will be scheduled. BTW - I have never seen a clearly-stated definition of the points at which - at a minimum - rescheduling _must_ be done. Clearly there are many possibilities, like expiration of a time slice, that are definitely _not_ required (timeslicing isn't required at all). But where are the points at which rescheduling _is_ required? The closest I have come to it in the ALRM is p. 9-18, in the section on ABORT, of all places (sect 9.10, paragraph 6, too long to copy here). But this seems like an obscure place to bury such an important point. Robert Firth (or whoever) - can you comment for us? Susan, how about you?