Re: [RFC 3/3] mutex: When there is no owner, stop spinning after too many tries

[Peter Zijlstra <peterz@infradead.org>]

On Wed, Jan 15, 2014 at 10:46:17PM -0800, Jason Low wrote:
> On Thu, 2014-01-16 at 10:14 +0700, Linus Torvalds wrote:
> > On Thu, Jan 16, 2014 at 9:45 AM, Jason Low <jason.low2@hp.com> wrote:
> > >
> > > Any comments on the below change which unlocks the mutex before taking
> > > the lock->wait_lock to wake up a waiter? Thanks.
> > 
> > Hmm. Doesn't that mean that a new lock owner can come in *before*
> > you've called debug_mutex_unlock and the lockdep stuff, and get the
> > lock? And then debug_mutex_lock() will be called *before* the unlocker
> > called debug_mutex_unlock(), which I'm sure confuses things.
> 
> If obtaining the wait_lock for debug_mutex_unlock is the issue, then
> perhaps we can address that by taking care of
> #ifdef CONFIG_DEBUG_MUTEXES. In the CONFIG_DEBUG_MUTEXES case, we can
> take the wait_lock first, and in the regular case, take the wait_lock
> after releasing the mutex.

I think we're already good for DEBUG_MUTEXES, because DEBUG_MUTEXES has
to work for archs that have !__mutex_slowpath_needs_to_unlock() and also
the DEBUG_MUTEXES code is entirely serialized on ->wait_lock.

Note that we cannot do the optimistic spinning for DEBUG_MUTEXES exactly
because of this reason.