Re: [PATCH] fs-writeback: drop wb->list_lock during blk_finish_plug()

From: Chris Mason <clm@fb.com>
To: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Josef Bacik <jbacik@fb.com>, LKML <linux-kernel@vger.kernel.org>,
	linux-fsdevel <linux-fsdevel@vger.kernel.org>,
	Dave Chinner <david@fromorbit.com>, Neil Brown <neilb@suse.de>,
	Jan Kara <jack@suse.cz>, Christoph Hellwig <hch@lst.de>
Subject: Re: [PATCH] fs-writeback: drop wb->list_lock during blk_finish_plug()
Date: Sat, 12 Sep 2015 19:46:32 -0400	[thread overview]
Message-ID: <20150912234632.GF4150@ret.masoncoding.com> (raw)
In-Reply-To: <CA+55aFzbniZkd4EegyKqNmyXWQF6dPypvdovP4K7RdeBrdZDVg@mail.gmail.com>

On Sat, Sep 12, 2015 at 04:29:06PM -0700, Linus Torvalds wrote:
> On Sat, Sep 12, 2015 at 4:00 PM, Chris Mason <clm@fb.com> wrote:
> >
> > I did the plain revert as well, just to have a baseline.
> 
> Ahh, I ended up not expecting you to get this done until after rc1 was
> out, so I in the meantime just merged my fix instead rather than leave
> the expected scheduling-while-atomic problem.

Yeah, I wasn't sure I'd be able to do the runs, but it was a rainy
afternoon and this was more fun than cleaning.  Really glad something
got in for rc1 either way.

> 
> And just as well that you did a baseline, since apparently the numbers
> are all over the map. I don't see how your hack and dave's original
> can _possibly_ differ that much, but they clearly did on your xfs
> test. So there's probably huge variance that depends on random
> details.

I don't think the XFS numbers can be trusted too much since it was
basically bottlenecked behind that single pegged CPU.  It was bouncing
around and I couldn't quite track it down to a process name (or perf
profile).

The btrfs numbers were much more consistent, but your patch is still a
win over plain 4.2.

> 
> I'll leave things as they are until we have something that looks a bit
> more believable ;)

We can build from here, thanks Linus.

-chris