Re: NPTL mutex and the scheduling priority

[Jakub Jelinek <jakub@redhat.com>]

On Tue, Jun 13, 2006 at 02:04:32PM +0200, S?bastien Dugu? wrote:
> > Now not sure what do you mean by "use PI futexes all along in glibc",
> > certainly you don't mean using them for normal mutexes, right? 
> > FUTEX_LOCK_PI has effects the normal futexes shouldn't have.
> > The condvars can be also used with PP mutexes and using PI for the cv
> > internal lock unconditionally wouldn't be the right thing either.
> 
>   I effectively meant using a PI futex for the cv __data.__futex but now
> I realize it's a Really Bad Idea.

For __data.__futex?  That is not just a really bad idea, that's not possible
at all.  PI futex has a hardcoded format (owner tid plus 2 control bits in
the MSB), while cv in __data.__futex needs to have an application controlled
format (a counter).

>   To summarize (correct me if I'm wrong), we need a way in the broadcast
> case to promote the cv __data.__futex type to the type of the external
> mutex (PI, PP, normal) in the requeue path. Therefore we need the
> ability to requeue waiters on a regular futex onto a PI futex.

We need more things:
Have FUTEX_WAKE/FUTEX_REQUEUE/FUTEX_WAKE_OP honor the scheduling policy/priorities
rather than use FIFO (this is needed not just for proper behavior
of pthread_mutex_unlock, but also for __data.__futex).

FUTEX_REQUEUE is used by pthread_cond_signal to requeue the __data.__futex
onto __data.__lock.  So it all depends on what futex type is used by the
__data.__lock.  A CV doesn't have a mutex associated to it at all times,
when there is no contention on it, it really doesn't matter.  Maybe
it would be possible to use PI futex always (well, if FUTEX_LOCK_PI is
supported, otherwise of course only use normal lock) for the internal lock
though.  All we want to ensure is that for the short time it is held
(no blocking operation should be done while __data.__lock is held
unless interrupted by signal) if the owning thread of __data.__lock is
scheduled away (or interrupted by signal) it doesn't cause priority
inversion.  That even includes e.g. two threads with different priorities
calling pthread_signal or pthread_broadcast concurrently (and in that
case, we might not have an associated mutex at all).
But, for PI __data.__lock, we need:
1) FUTEX_REQUEUE being able to requeue non-PI futex to PI futex
2) FUTEX_WAKE_OP alternative that allows the target futex be a PI futex:
   ATM NPTL uses FUTEX_OP_CLEAR_WAKE_IF_GT_ONE   ((4 << 24) | 1)
   FUTEX_WAKE_OP operation, i.e. atomically
   FUTEX_WAKE (futex); int old = *lock; *lock = 0; if (old > 1) FUTEX_WAKE (lock);
   but with PI __data.__lock, we want instead atomically:
   FUTEX_WAKE (futex); /* This one is normal, non-PI */ FUTEX_UNLOCK_PI (lock);
   (or perhaps:
   FUTEX_WAKE (futex); int old = *lock;
   if (old & FUTEX_WAITERS)
     FUTEX_UNLOCK_PI (lock);
   else if (old == gettid ())
     *lock = 0;
   else
     FUTEX_UNLOCK_PI (lock);
   ).

	Jakub
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/