Re: [RFC][PATCH RT 0/2] futex priority based wakeup

[Sébastien Dugué]

On Wed, 2006-05-10 at 12:08 +0200, Ingo Molnar wrote:
> * Sébastien Dugué <sebastien.dugue@xxxxxxxx> wrote:
> 
> >   in the current futex implementation, tasks are woken up in FIFO 
> > order, (i.e. in the order they were put to sleep). For realtime 
> > systems needing system wide strict realtime priority scheduling, tasks 
> > should be woken up in priority order.
> > 
> >   This patchset achieves this by changing the futex hash bucket list 
> > into a plist. Tasks waiting on a futex are enqueued in this plist 
> > based on their priority so that they can be woken up in priority 
> > order.
> 
> hm, i dont think this is enough. Basically, waking up in priority order 
> is just the (easier) half of the story - what you want is to also 
> propagate priorities when you block. We provided a complete solution via 
> the PI-futex patchset (currently included in -mm).
> 
> In other words: as long as locking primitives go, i dont think real-time 
> applications should use wakeup-priority-ordered futexes, they should use 
> the real thing, PI futexes.

  Yeah, that's right as long as userland pthread_mutexes are used, but
currently, pthread_condvars do not use PI-futexes. Therefore when we
have some threads sleeping on a condvar and the condvar is broadcasted,
those blocked threads are woken up through futex_requeue() in the order
they were put to sleep and not in their priority order.

  Maybe the pthread_cond_*() function should be made to use the
PI-futexes support in glibc.

> 
> There is one exception: when a normal futex is used as a waitqueue 
> without any contention properties. (for example a waitqueue for worker 
> threads) But those are both rare, and typically dont muster tasks with 
> different priorities - i.e. FIFO is good enough.
> 
> Also, there's a performance cost to this. Could you try to measure the 
> impact to SCHED_OTHER tasks via some pthread locking benchmark?

  Right, I will try to quantify the impact for SCHED_OTHER tasks. Any
pointers to such a benchmark?

  Thanks,

  Sébastien.


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