Re: [Xenomai-core] Houston, we have a circular problem

[Philippe Gerum <rpm@xenomai.org>]

Jan Kiszka wrote:
> Philippe Gerum wrote:
>> Jan Kiszka wrote:
>>> Gilles Chanteperdrix wrote:
>>>> On Mon, May 5, 2008 at 6:08 PM, Philippe Gerum <rpm@xenomai.org> wrote:
>>>>>  do_schedule_event() is the culprit when it reads the pending signals on the
>>>>>  shared queue (XNDEBUG check for rearming the timers),
>>>> A stupid suggestion: if we know that the spinlock is always locked
>>>> when calling do_schedule_event, maybe we can simply avoid the lock
>>>> there ?
>>> Would be the best solution - but I don't think so. After reading a bit
>>> more into the lockdep output, I think the issue is that _some_other_
>>> task my hold the siglock and then acquire our rq_lock, but not
>>> necessarily along a similar code path we took to acquire the siglock now.
>>>
>> Actually, this locking around the sigmask retrieval looks overkill, since we
>> only address ptracing signals here, and those should go through the shared
>> pending set, not through the task's private one. I.e. There should be no way to
>> get fooled by any asynchronous update of that mask.
> 
> This is a debug helper anyway, so we risk (if I got this right) at worst
> a spurious unfreeze of the Xenomai timers. Does not really compare to
> the current deadlock...
>

As a matter of fact, we don't test any condition under the protection of this
lock, so aside of the memory barrier induced on SMP, this lock does not buy us
anything. Except a deadlock, that is...

> I will let my colleagues run the hunk below tomorrow (which works for me) -
> let's see if they manage to crash it again :P (they are experts in this!).
> 
> Jan
> 
> Index: xenomai-2.4.x/ksrc/nucleus/shadow.c
> ===================================================================
> --- xenomai-2.4.x/ksrc/nucleus/shadow.c	(Revision 3734)
> +++ xenomai-2.4.x/ksrc/nucleus/shadow.c	(Arbeitskopie)
> @@ -2194,9 +2194,7 @@ static inline void do_schedule_event(str
>  			if (signal_pending(next)) {
>  				sigset_t pending;
>  
> -				spin_lock(&wrap_sighand_lock(next));	/* Already interrupt-safe. */
>  				wrap_get_sigpending(&pending, next);
> -				spin_unlock(&wrap_sighand_lock(next));
>  
>  				if (sigismember(&pending, SIGSTOP) ||
>  				    sigismember(&pending, SIGINT))
> 


-- 
Philippe.