Messages in this thread | | | Date | Thu, 2 May 2013 14:54:26 +0200 | From | Jens Axboe <> | Subject | Re: [PATCH] wait: fix false timeouts when using wait_event_timeout() |
| |
On Thu, May 02 2013, Imre Deak wrote: > On Thu, 2013-05-02 at 14:23 +0200, Jens Axboe wrote: > > On Thu, May 02 2013, Daniel Vetter wrote: > > > On Thu, May 2, 2013 at 12:29 PM, David Howells <dhowells@redhat.com> wrote: > > > >> Fix this by returning at least 1 if the condition becomes true. This > > > >> semantic is in line with what wait_for_condition_timeout() does; see > > > >> commit bb10ed09 - "sched: fix wait_for_completion_timeout() spurious > > > >> failure under heavy load". > > > > > > > > But now you can't distinguish the timer expiring first, if the thread doing > > > > the waiting gets delayed sufficiently long for the event to happen. > > > > > > That can already happen, e.g. > > > > > > 1. wakeup happens and condition is true. > > > 2. we compute remaining jiffies > 0 > > > -> preempt > > > 3. now wait_for_event_timeout returns. > > > > > > Only difference is that the delay/preempt happens in between 1. and > > > 2., and then suddenly the wake up didn't happen in time (with the > > > current return code semantics). > > > > > > So imo the current behaviour is simply a bug and will miss timely > > > wakeups in some cases. > > > > > > The other way round, namely wait_for_event_timeout taking longer than > > > the timeout is expected (and part of the interface for every timeout > > > function). So all current callers already need to be able to cope with > > > random preemption/delays pushing the total time before the call to > > > wait_for_event and checking the return value over the timeout, even > > > when condition was signalled in time. > > > > > > If there's any case which relies on accurate timeout detection that > > > simply won't work with wait_for_event (they need an nmi or a hw > > > timestamp counter or something similar). > > > > I seriously doubt that anyone is depending on any sort of accuracy on > > the return. 1 jiffy is not going to make or break anything - in fact, > > jiffies could be incremented nsecs after the initial call. So a > > granularity of at least 1 is going to be expected in any case. > > > > The important bit here is that the API should behave as expected. And > > the most logical way to code that is to check the return value. I can > > easily see people forgetting to re-check the condition, hence you get a > > bug. The fact that you and the original reporter already had accidents > > with this is a clear sign that the logical way to use the API is not the > > correct one. > > > > IMHO, the change definitely makes sense. > > Ok, so taking courage of this answer ;P How about also the following? > > diff --git a/kernel/timer.c b/kernel/timer.c > index dbf7a78..5a62456 100644 > --- a/kernel/timer.c > +++ b/kernel/timer.c > @@ -1515,7 +1515,11 @@ signed long __sched schedule_timeout(signed long > timeout) > } > } > > - expire = timeout + jiffies; > + /* > + * We can't be sure how close we are to the next tick, so +1 to > + * guarantee that we wait at least timeout amount. > + */ > + expire = timeout + jiffies + 1; > > setup_timer_on_stack(&timer, process_timeout, (unsigned long)current); > __mod_timer(&timer, expire, false, TIMER_NOT_PINNED); > > > It'd plug a similar hole for wait_event_timeout() and similar users, who > don't compensate for the above..
Any jiffy based API is going to have this issue. I think it's different from the original patch, which just makes the API potentially return something that is confusing.
So not sure on the above, sorry.
-- Jens Axboe
| |