Re: [PATCH 4/4] locking: Introduce smp_cond_acquire()

[Paul E. McKenney]

On Mon, Nov 16, 2015 at 04:46:36PM +0000, Will Deacon wrote:
> On Mon, Nov 16, 2015 at 08:44:43AM -0800, Paul E. McKenney wrote:
> > On Mon, Nov 16, 2015 at 04:24:53PM +0000, Will Deacon wrote:
> > > On Mon, Nov 16, 2015 at 05:04:45PM +0100, Peter Zijlstra wrote:
> > > > On Mon, Nov 16, 2015 at 04:56:58PM +0100, Peter Zijlstra wrote:
> > > > > On Thu, Nov 12, 2015 at 10:21:39AM -0800, Linus Torvalds wrote:
> > > > > > Now, the point of spin_unlock_wait() (and "spin_is_locked()") should
> > > > > > generally be that you have some external ordering guarantee that
> > > > > > guarantees that the lock has been taken. For example, for the IPC
> > > > > > semaphores, we do either one of:
> > > > > > 
> > > > > >  (a) get large lock, then - once you hold that lock - wait for each small lock
> > > > > > 
> > > > > > or
> > > > > > 
> > > > > >  (b) get small lock, then - once you hold that lock - check that the
> > > > > > largo lock is unlocked
> > > > > > 
> > > > > > and that's the case we should really worry about.  The other uses of
> > > > > > spin_unlock_wait() should have similar "I have other reasons to know
> > > > > > I've seen that the lock was taken, or will never be taken after this
> > > > > > because XYZ".
> > > > > 
> > > > > I don't think this is true for the usage in do_exit(), we have no
> > > > > knowledge on if pi_lock is taken or not. We just want to make sure that
> > > > > _if_ it were taken, we wait until it is released.
> > > > 
> > > > And unless PPC would move to using RCsc locks with a SYNC in
> > > > spin_lock(), I don't think it makes sense to add
> > > > smp_mb__after_unlock_lock() to all tsk->pi_lock instances to fix this.
> > > > As that is far more expensive than flipping the exit path to do
> > > > spin_lock()+spin_unlock().
> > > 
> > > ... or we upgrade spin_unlock_wait to a LOCK operation, which might be
> > > slightly cheaper than spin_lock()+spin_unlock().
> > 
> > Or we supply a heavyweight version of spin_unlock_wait() that forces
> > the cache miss.  But I bet that the difference in overhead between
> > spin_lock()+spin_unlock() and the heavyweight version would be down in
> > the noise.
> 
> I'm not so sure. If the lock is ticket-based, then spin_lock() has to
> queue for its turn, whereas spin_unlock_wait could just wait for the
> next unlock.

Fair point, and it actually applies to high-contention spinlocks as well,
just a bit less deterministically.  OK, given that I believe that we do
see high contention on the lock in question, I withdraw any objections
to a heavy-weight form of spin_unlock_wait().

							Thanx, Paul

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/