RE: [patch] aio: add per task aio wait event condition

["Chen, Kenneth W" <kenneth.w.chen@intel.com>]

Zach Brown wrote on Tuesday, January 02, 2007 6:06 PM
> On Jan 2, 2007, at 5:50 PM, Chen, Kenneth W wrote:
> > Zach Brown wrote on Tuesday, January 02, 2007 5:24 PM
> >>> That is not possible because when multiple tasks waiting for
> >>> events, they
> >>> enter the wait queue in FIFO order, prepare_to_wait_exclusive() does
> >>> __add_wait_queue_tail().  So first io_getevents() with min_nr of 2
> >>> will be woken up when 2 ops completes.
> >>
> >> So switch the order of the two sleepers in the example?
> >
> > Not sure why that would be a problem though:  whoever sleep first will
> > be woken up first.
> 
> Why would the min_nr = 3 sleeper be woken up in that case?  Only 2  
> ios were issued.
> 
> Maybe the app was relying on the min_nr = 2 completion to issue 3  
> more ios for the min_nr = 3 sleeper, who knows.
> 
> Does that clear up the confusion?


Not really. I don't think I understand your concern. You gave an example:

issue 2 ops
first io_getevents sleeps with a min_nr of 2
second io_getevents sleeps with min_nr of 3
2 ops complete but only test the second sleeper's min_nr of 3
first sleeper twiddles thumbs

Or:

issue 2 ops
first io_getevents sleeps with a min_nr of 3
second io_getevents sleeps with min_nr of 2
2 ops complete but only test the second sleeper's min_nr of 2
first sleeper twiddles thumbs


First scenario doesn't exist because in the new scheme, we test first
sleeper (as in head of the queue) when 2 ops complete. It wakes up first.

2nd scenario is OK to me because first sleeper waiting for 3 events,
and there are only 2 ops completed, so it waits.

The one scenario that I can think of that breaks down is that one task
sleeps with min_nr of 100.  Then 50 ops completed.  Comes along 2nd
thread does a io_getevents and it will take all 50 events in the 2nd
thread.  Is that what you are talking about?  It doesn't involve two
sleepers.  That I can fix by testing whether wait queue is active or
not at the beginning of fast path in read_events().

The bigger question is: what is the semantics on event reap order for
thread? Random, FIFO or round robin?  It is not specified anywhere.
What would be the most optimal policy?