Re: [lkp-robot] [fs/locks] 9d21d181d0: will-it-scale.per_process_ops -14.1% regression

[Jeff Layton <jlayton@redhat.com>]

On Thu, 2017-06-01 at 11:14 -0400, J. Bruce Fields wrote:
> On Thu, Jun 01, 2017 at 08:59:21AM -0400, Jeff Layton wrote:
> > On Thu, 2017-06-01 at 07:49 -0400, Benjamin Coddington wrote:
> > > On 1 Jun 2017, at 7:41, Jeff Layton wrote:
> > > 
> > > > On Thu, 2017-06-01 at 10:05 +0800, kernel test robot wrote:
> > > > > Greeting,
> > > > > 
> > > > > FYI, we noticed a -14.1% regression of will-it-scale.per_process_ops 
> > > > > due to commit:
> > > > > 
> > > > > 
> > > > > commit: 9d21d181d06acab9a8e80eac2ec4eed77b656793 ("fs/locks: Set 
> > > > > fl_nspid at file_lock allocation")
> > > > > url: 
> > > > > https://github.com/0day-ci/linux/commits/Benjamin-Coddington/fs-locks-Alloc-file_lock-where-practical/20170527-050700
> > > > > 
> > > > > 
> > > > 
> > > > Ouch, that's a rather nasty performance hit. In hindsight, maybe we
> > > > shouldn't move those off the stack after all? Heck, if it's that
> > > > significant, maybe we should move the F_SETLK callers to allocate 
> > > > these
> > > > on the stack as well?
> > > 
> > > We can do that.  But, I think this is picking up the 
> > > locks_mandatory_area()
> > > allocation which is now removed.  The attached config has
> > > CONFIG_MANDATORY_FILE_LOCKING=y,  so there's allocation on every 
> > > read/write.
> > > 
> > 
> > I'm not so sure. That would only be the case if the thing were marked
> > for manadatory locking (a really rare thing).
> > 
> > The test is really simple and I don't think any read/write activity is
> > involved:
> > 
> >     https://github.com/antonblanchard/will-it-scale/blob/master/tests/lock1.c
> 
> So it's just F_WRLCK/F_UNLCK in a loop spread across multiple cores?
> I'd think real workloads do some work while holding the lock, and a 15%
> regression on just the pure lock/unlock loop might not matter?  But best
> to be careful, I guess.
> 
> --b.
> 

Yeah, that's my take.

I was assuming that getting a pid reference would be essentially free,
but it doesn't seem to be.

So, I think we probably want to avoid taking it for a file_lock that we
use to request a lock, but do take it for a file_lock that is used to
record a lock. How best to code that up, I'm not quite sure...

> > 
> > ...and the 0 day bisected it down to this patch, IIUC:
> > 
> >     https://github.com/0day-ci/linux/commit/9d21d181d06acab9a8e80eac2ec4eed77b656793
> > 
> > It seems likely that it's the extra get_pid/put_pid in the allocation
> > and free codepath. I expected those to be pretty cheap, but maybe
> > they're not?
> 
> 

-- 
Jeff Layton <jlayton@redhat.com>