Re: robust futex heap support patch

["Randy.Dunlap" <rdunlap@xenotime.net>]

On Tue, 6 Dec 2005 15:06:12 +0100 (MET) Esben Nielsen wrote:

> On Mon, 5 Dec 2005, david singleton wrote:
> 
> > 
> > On Dec 5, 2005, at 2:30 PM, Esben Nielsen wrote:
> > 
> > I'm currently trying to close a race condition between futex_wait_robust
> > and futex_wake_robust that Dave Carlson is seeing on his SMP system.
> > 
> > The scenario is as follows:
> > 
> > Thread A locks an pthread_mutex via the fast path and does not enter 
> > the kernel.
> > 
> > Thread B tries to lock the lock and sees it is already locked.  Thread 
> > B sets the
> > waiters flag in the lock and enters the kernel to lock the lock on 
> > behalf of
> > thread A and then block on the mutex waiting for it's release.
> > 
> > Thread A unlocks the lock and sees the waiters flag set.  Thread A gets
> > to the futex_wake_robust before Thread B can get to futex_wait_robust.
> > 
> > Thread A sees that it does not own the lock in the kernel and was 
> > returning EINVAL.
> > 
> > patch-2.6.14-rt21-rf8 was a preliminary patch for Dave Carlson to try 
> > and get more information about
> > the race condition.   ( rf8 and rf9 are still returning EAGAIN from 
> > thread B trying to
> > do the futex_wait_robust and the library should be retrying with 
> > EAGAIN, but it currently isn't).
> > 
> 
> *nod*
> I just pointed out that you can't make thread A loop the way you do.
> 
> What I would do was to do the user space flag checks while having the
> raw spinlock of the rt_mutex. That way you are sure that stuff in the
> kernel is serialized. But I don't know what to do exactly to do in
> there....
> 
> 
> > When I get the race condition closed I'll post the patch and notify 
> > everyone on lkml
> > and the robustmutexes mailing lists.
> 
> Where can I sign up to that mailing list? I would like to follow the
> development, although I don't have much time to contibute.

http://lists.osdl.org/mailman/listinfo/robustmutexes/


---
~Randy