Re: Another Performance Regression in write() syscall

[Dave Hansen <dave@linux.vnet.ibm.com>]

On Tue, 2009-02-24 at 19:47 +1100, Nick Piggin wrote:
> On Tuesday 24 February 2009 17:25:45 Dave Hansen wrote:
> > On Mon, 2009-02-23 at 22:05 -0800, Salman Qazi wrote:
> > > Analysis of profile data has led us to believe that the commit
> > > 3d733633a633065729c9e4e254b2e5442c00ef7e has caused a performance
> > > regression.  This commit provides for tracking of writers so that read
> > > only bind mounts function correctly.
> > >
> > > We can verify this regression by applying the following patch to
> > > partially disable the above-mentioned commit and then running the fstime
> > > component of Unixbench.  The settings used were 256 byte writes with
> > > MAX_BLOCK of 2000.
> >
> > I'm a bit surprised that write() is what is regressing.  Unless I
> > screwed up, we do all the expensive accounting at open()/close() time.
> > Is this a test that gets run in parallel on multiple cpus?
> 
> Don't forget touch_atime...

Yeah, that's a good point.  Are we sure that's what is happening here,
though?  That's one thing a profile would hopefully help with.

> Still, open/close isn't unimportant either.

Yeah, that's true.  But, what I noticed was that all of the other
open/close activity masked out any overhead from mnt_want/drop_write()
since a big chunk of the overhead was just going and bringing the
vfsmount pieces into the cache.

> > Could you take a look at Nick's patches to speed this stuff up?
> >
> > 	http://thread.gmane.org/gmane.linux.file-systems/28186
> >
> > We may need to dust those off, although I'm still a bit worried about
> > the complexities of open-coding all the barriers.
> 
> I really need to do something about trying to push them upstream again
> actually because we've got them in SLES11 tree.

Were the patches that you integrated any different from the ones you
posted a few months ago?

-- Dave