Gilles Chanteperdrix wrote:
> Gilles Chanteperdrix wrote:
>> Jan Kiszka wrote:
>>> Gilles Chanteperdrix wrote:
>>>> Jan Kiszka wrote:
>>>>> Gilles Chanteperdrix wrote:
>>>>>> Jan Kiszka wrote:
>>>>>>> Hi,
>>>>>>> I'm hitting that bug check in __xnpod_schedule after
>>>>>>> xnintr_clock_handler issued a xnpod_schedule like this:
>>>>>>>         if (--sched->inesting == 0) {
>>>>>>>                 __clrbits(sched->status, XNINIRQ);
>>>>>>>                 xnpod_schedule();
>>>>>>>         }
>>>>>>> Either the assumption behind the bug check is no longer correct (no call
>>>>>>> to xnpod_schedule() without a real need), or we should check for
>>>>>>> __xnpod_test_resched(sched) in xnintr_clock_handler (but under nklock 
>>>>>>> then).
>>>>>>> Comments?
>>>>>> You probably have a real bug. This BUG_ON means that the scheduler is
>>>>>> about to switch context for real, whereas the resched bit is not set,
>>>>>> which is wrong.
>>>>> This happened over my 2.6.35 port - maybe some spurious IRQ enabling.
>>>>> Debugging further...
>>>> You should look for something which changes the scheduler state without
>>>> setting the resched bit, or for something which clears the bit without
>>>> taking the scheduler changes into account.
>>> It looks like a generic Xenomai issue on SMP boxes, though a mostly
>>> harmless one:
>>> The task that was scheduled in without XNRESCHED set locally has been
>>> woken up by a remote CPU. The waker requeued the task and set the
>>> resched condition for itself and in the resched proxy mask for the
>>> remote CPU. But there is at least one place in the Xenomai code where we
>>> drop the nklock between xnsched_set_resched and xnpod_schedule:
>>> do_taskexit_event (I bet there are even more). Now the resched target
>>> CPU runs into a timer handler, issues xnpod_schedule unconditionally,
>>> and happens to find the woken-up task before it is actually informed via
>>> an IPI.
>>> I think this is a harmless race, but it ruins the debug assertion
>>> "need_resched != 0".
>> Not that harmless, since without the debugging code, we would miss the
>> reschedule too...
> Ok. But we would finally reschedule when handling the IPI. So, the
> effect we see is a useless delay in the rescheduling.

Depends on the POV: The interrupt or context switch between set_resched
and xnpod_reschedule that may defer rescheduling may also hit us before
we were able to wake up the thread at all. The worst case should not
differ significantly.


Siemens AG, Corporate Technology, CT T DE IT 1
Corporate Competence Center Embedded Linux

Xenomai-core mailing list

Reply via email to