Re: [BUG] futex_unlock_pi() hurts my brain and may cause application deadlock

[john stultz <johnstul@us.ibm.com>]

On Thu, 2007-05-31 at 16:55 +0200, Ingo Molnar wrote:
> * john stultz <johnstul@us.ibm.com> wrote:
> 
> > > > --- a/kernel/futex.c
> > > > +++ b/kernel/futex.c
> > > > @@ -2011,6 +2011,7 @@ pi_faulted:
> > > >  					 attempt);
> > > >  		if (ret)
> > > >  			goto out_unlock;
> > > > +		uval = 0;
> > > >  		goto retry_locked;
> > > 
> > > this looks good to me. Oleg has posted a few more fixes as well - 
> > > have you tried those too?
> > 
> > Do you have a URL or subject line for the discussion? I'll take a look 
> > and give them a whirl.
> 
> i meant to say Alexey, not Oleg :-) Here's the thread:
> 
>    http://lkml.org/lkml/2007/5/7/129
> 
> i've Cc:-ed Alexey too.

I haven't had a chance to run it, but reviewing at Alexey's changes,
they do not look to be sufficient to the issue we were hitting. We still
need to clear uval to avoid  -EFAULT &'ing w/ FUTEX_OWNER_DIED.


Here's the patch against -rt9 (which also contains Alexey's fix)


diff -rup 2.6-rt/kernel/futex.c 2.6-rtnew/kernel/futex.c
--- 2.6-rt/kernel/futex.c	2007-05-31 06:38:00.000000000 -0500
+++ 2.6-rtnew/kernel/futex.c	2007-05-31 06:43:08.000000000 -0500
@@ -1954,6 +1954,7 @@ pi_faulted:
 			ret = -EFAULT;
 			goto out_unlock;
 		}
+		uval = 0;
 		goto retry_locked;
 	}
 
@@ -3503,6 +3504,7 @@ pi_faulted:
 			ret = -EFAULT;
 			goto out_unlock;
 		}
+		uval = 0;
 		goto retry_locked;
 	}