linux-kernel - Re: [PATCH 1/4] locking/ww_mutex: Fix a deadlock affecting ww

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Date:   Wed, 23 Nov 2016 15:25:25 +0100
From:   Daniel Vetter <daniel@...ll.ch>
To:     Peter Zijlstra <peterz@...radead.org>
Cc:     Nicolai Hähnle <nhaehnle@...il.com>,
        Nicolai Hähnle <Nicolai.Haehnle@....com>,
        linux-kernel@...r.kernel.org, dri-devel@...ts.freedesktop.org,
        Ingo Molnar <mingo@...hat.com>, stable@...r.kernel.org,
        Maarten Lankhorst <maarten.lankhorst@...onical.com>
Subject: Re: [PATCH 1/4] locking/ww_mutex: Fix a deadlock affecting ww_mutexes

On Wed, Nov 23, 2016 at 03:03:36PM +0100, Peter Zijlstra wrote:
> On Wed, Nov 23, 2016 at 12:25:22PM +0100, Nicolai Hähnle wrote:
> > @@ -473,7 +476,14 @@ void __sched ww_mutex_unlock(struct ww_mutex *lock)
> >  	 */
> >  	mutex_clear_owner(&lock->base);
> >  #endif
> > -	__mutex_fastpath_unlock(&lock->base.count, __mutex_unlock_slowpath);
> > +	/*
> > +	 * A previously _not_ waiting task may acquire the lock via the fast
> > +	 * path during our unlock. In that case, already waiting tasks may have
> > +	 * to back off to avoid a deadlock. Wake up all waiters so that they
> > +	 * can check their acquire context stamp against the new owner.
> > +	 */
> > +	__mutex_fastpath_unlock(&lock->base.count,
> > +				__mutex_unlock_slowpath_wakeall);
> >  }
> 
> So doing a wake-all has obvious issues with thundering herd etc.. Also,
> with the new mutex, you'd not be able to do hand-off, which would
> introduce starvation cases.
> 
> Ideally we'd iterate the blocked list and pick the waiter with the
> earliest stamp, or we'd maintain the list in stamp order instead of
> FIFO, for ww_mutex.

Not sure we'll win that much, at least I think we still need to wake up
everyone with earlier stamp than the one of the task that just released
the lock. Otherwise there's deadlocks. So just cuts the wakeups in half,
on average.

What we could do is do a handoff-hint with the timestamp of earliest task
we believe should get the lock. Everyone with a later timestamp that gets
woken then knows that they definitely have a deadlock situation and need
to back off (thread 2 in the example).

thread 1 would get woken, and would be able to take the lock, except when
thread 0 successfully raced it and stole the lock. And anyone else racing
in with later timestamps would also immediately back off, ensuring
fairness.

Without thinking it through in detail this is a PI issue, except that we
replace boosting with wakeup&back-off. Could we perhaps steal something
from rt mutexes to make it fair&efficient?
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch