On 07/17/2015 02:52 PM, git repository hosting wrote:
>
> diff --git a/kernel/cobalt/synch.c b/kernel/cobalt/synch.c
> index 7142fd2..ae4ceef 100644
> --- a/kernel/cobalt/synch.c
> +++ b/kernel/cobalt/synch.c
> @@ -443,7 +443,9 @@ redo:
> if (likely(h == XN_NO_HANDLE)) {
> xnsynch_set_owner(synch, curr);
> xnthread_get_resource(curr);
> + xnlock_get_irqsave(&nklock, s);
> xnthread_clear_info(curr, XNRMID | XNTIMEO | XNBREAK);
> + xnlock_put_irqrestore(&nklock, s);
> return 0;
> }
>
Instead of serializing in the fast path only for clearing those bits, we
could convert the remaining call sites to use xnsynch_acquire's return
value instead. At any rate, those info bits are meant to be tested
immediately upon return from the blocking call. I did most of the
conversions already, except __cobalt_mutex_acquire_unchecked(). Fixing
this one should allow us to drop the requirement for clearing the bits
in the info word, and the need for serializing in the same move.
> diff --git a/kernel/cobalt/thread.c b/kernel/cobalt/thread.c
> index 4cec6e7..7c8cb3a 100644
> --- a/kernel/cobalt/thread.c
> +++ b/kernel/cobalt/thread.c
> @@ -1989,6 +1989,7 @@ void xnthread_relax(int notify, int reason)
> struct task_struct *p = current;
> int cpu __maybe_unused;
> siginfo_t si;
> + spl_t s;
>
> primary_mode_only();
>
> @@ -2045,7 +2046,9 @@ void xnthread_relax(int notify, int reason)
> si.si_int = reason | sigdebug_marker;
> send_sig_info(SIGDEBUG, &si, p);
> }
> + xnlock_get_irqsave(&nklock, s);
> xnsynch_detect_claimed_relax(thread);
> + xnlock_put_irqrestore(&nklock, s);
> }
Sigh, this was a bad one, scanning two queues with preemption enabled:
scary. I would suggest to move the serialization to
xnsynch_detect_claimed_relax() directly, since that code can be
optimized out when !XENO_DEBUG(MUTEX_RELAXED).
--
Philippe.
_______________________________________________
Xenomai mailing list
[email protected]
http://xenomai.org/mailman/listinfo/xenomai
