Re: [locks] 6d390e4b5d: will-it-scale.per_process_ops -96.6% regression

[Jeff Layton <jlayton@kernel.org>]

On Mon, 2020-03-09 at 15:09 -0400, Jeff Layton wrote:
> On Mon, 2020-03-09 at 13:22 -0400, Jeff Layton wrote:
> > On Mon, 2020-03-09 at 08:52 -0700, Linus Torvalds wrote:
> > > On Mon, Mar 9, 2020 at 7:36 AM Jeff Layton <jlayton@kernel.org> wrote:
> > > > On Sun, 2020-03-08 at 22:03 +0800, kernel test robot wrote:
> > > > > FYI, we noticed a -96.6% regression of will-it-scale.per_process_ops due to commit:
> > > > 
> > > > This is not completely unexpected as we're banging on the global
> > > > blocked_lock_lock now for every unlock. This test just thrashes file
> > > > locks and unlocks without doing anything in between, so the workload
> > > > looks pretty artificial [1].
> > > > 
> > > > It would be nice to avoid the global lock in this codepath, but it
> > > > doesn't look simple to do. I'll keep thinking about it, but for now I'm
> > > > inclined to ignore this result unless we see a problem in more realistic
> > > > workloads.
> > > 
> > > That is a _huge_ regression, though.
> > > 
> > > What about something like the attached? Wouldn't that work? And make
> > > the code actually match the old comment about wow "fl_blocker" being
> > > NULL being special.
> > > 
> > > The old code seemed to not know about things like memory ordering either.
> > > 
> > > Patch is entirely untested, but aims to have that "smp_store_release()
> > > means I'm done and not going to touch it any more", making that
> > > smp_load_acquire() test hopefully be valid as per the comment..
> > 
> > Yeah, something along those lines maybe. I don't think we can use
> > fl_blocker that way though, as the wait_event_interruptible is waiting
> > on it to go to NULL, and the wake_up happens before fl_blocker is
> > cleared.
> > 
> > Maybe we need to mix in some sort of FL_BLOCK_ACTIVE flag and use that
> > instead of testing for !fl_blocker to see whether we can avoid the
> > blocked_lock_lock?
> >   
> 
> How about something like this instead? (untested other than for
> compilation)
> 
> Basically, this just switches the waiters over to wait for
> fl_blocked_member to go empty. That still happens before the wakeup, so
> it should be ok to wait on that.
> 
> I think we can also eliminate the lockless list_empty check in
> locks_delete_block, as the fl_blocker check should be sufficient now.

Actually, no -- we need to keep that check in. The rest should work
though. I'll do some testing with it and see if the perf issue goes
away.

Thanks,
-- 
Jeff Layton <jlayton@kernel.org>