Re: dcache shrink list corruption?

From: Al Viro <viro@ZenIV.linux.org.uk>
To: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Miklos Szeredi <miklos@szeredi.hu>,
	Dave Chinner <david@fromorbit.com>,
	Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
	linux-fsdevel <linux-fsdevel@vger.kernel.org>
Subject: Re: dcache shrink list corruption?
Date: Thu, 1 May 2014 04:12:51 +0100	[thread overview]
Message-ID: <20140501031251.GZ18016@ZenIV.linux.org.uk> (raw)
In-Reply-To: <CA+55aFzN=ERrtWUwAQv0ry-ZzX2k7dz3JZQT6q5zjp_ABwMg6Q@mail.gmail.com>

On Wed, Apr 30, 2014 at 07:59:43PM -0700, Linus Torvalds wrote:
> On Wed, Apr 30, 2014 at 7:51 PM, Al Viro <viro@zeniv.linux.org.uk> wrote:
> >
> >       Help with profiling is needed; the loads to watch are
> > the ones where dentry_kill() + dentry_free() are sufficiently high in profiles
> > for the differences to matter.  Any takers?
> 
> I really hope there are no such loads, my "lock/unlock pairing"
> suggestion was mostly so that the pairing is clearer, not necessarily
> for performance.
> 
> That said, I'd assume that it migth be worth testing at least the
> "lots of concurrent lookups of 'simple_dentry_operations' dentries".
> So most of /proc, most uses of simple_lookup(). That at least triggers
> the dentry_kill() path in dput(), so it should be fairly easy to get
> profiles.
> 
> But real loads with real filesystems? That sounds harder.

Well, the simplest way to do that is a bunch of open/unlink/close.  Another
one is cp -rl / rm -rf of a large tree (rmdir(2) does shrink_dcache_parent()). 
For that matter, unmapping an anon shared mapping will trigger the same path.