Re: 3.5-rc6 futex_wait_requeue_pi oops.

[Darren Hart <dvhart@linux.intel.com>]

On 07/19/2012 04:22 PM, Darren Hart wrote:
> 
> 
> On 07/13/2012 11:54 AM, Dave Jones wrote:
>> On Fri, Jul 13, 2012 at 08:47:38PM +0200, Thomas Gleixner wrote:
>>  > On Fri, 13 Jul 2012, Dave Jones wrote:
>>  > 
>>  > > Looks like calling futex() with garbage makes things unhappy.
>>  > 
>>  >                 WARN_ON(!&q.pi_state);
>>  >                 pi_mutex = &q.pi_state->pi_mutex;
>>  >                 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter, 1);
>>  >                 debug_rt_mutex_free_waiter(&rt_waiter);
>>  > 
>>  > So there is some weird way which causes q.pi_state = NULL. Dave, did
>>  > you see the warning before the oops happened ?
>>
>> No, that didn't seem to trigger.
> 
> Well I don't have a fix yet, but I can explain this not triggering.
> 
> q is on the stack, so the ADDRESS for q.pi_state is never going to be
> NULL. However, properly instrumented, we do see this:
> 
> [   23.621501] ---[ end trace 20bdfb44db182a17 ]---
> [   23.622425] q.pi_state @           (null)
> [   23.623272] &q.pi_state @ ffff880185e2dca8
> [   23.624119] ------------[ cut here ]------------
> 
> Duh.
> 
> I'll add a fix to that WARN_ON in my futex-fixes branch along with the
> fix for the bug Dan found.
> 

I think I have root cause. futex_wait_requeue_pi() doesn't like having
uaddr == uaddr2. The handle_early_wakeup() doesn't detect a problem
because key2 IS the same as key1, I think. I've just discovered this and
quickly hacked in a "if (uaddr==uaddr2) return -EINVAL" fix and the test
continues to run (with just ops 0, 11, 12) for several minutes now
(typically fails in a few seconds). I'll let it run for a few hours and
contemplate the proper fix.

-- 
Darren Hart
Intel Open Source Technology Center
Yocto Project - Linux Kernel