Re: dcache shrink list corruption?

[Al Viro <viro@ZenIV.linux.org.uk>]

On Wed, Apr 30, 2014 at 07:33:38PM +0200, Miklos Szeredi wrote:
> On Wed, Apr 30, 2014 at 6:03 PM, Al Viro <viro@zeniv.linux.org.uk> wrote:
> > On Wed, Apr 30, 2014 at 05:49:58PM +0200, Miklos Szeredi wrote:
> >> > FWIW, the first two are really straightforward expanding the function
> >> > into its only callsite.  The last needs more splitup.  Not sure if the
> >> > following is good enough, but it ought to be at least somewhat cleaner.
> >> > Combined change is identical to the original, so it doesn't invalidate
> >> > the testing so far...
> >>
> >> Hmm, patches look okay, but I'm wondering if we really need the morgue list and
> >> the waiting.  Why not just skip dentries that are presently being handled by
> >> dentry_kill()?
> >
> > Who will be freeing them?  If we do that from dentry_kill(), we are back to
> > needing them removed from shrink list by something called by dput().  And
> > if we do that from shrink_dentry_list(), we'd damn better wait for
> > dentry_iput() et.al. to finish.
> 
> We can do it from dput if the shrinker gets there first and from the
> shrinker if dput manages to finish before.  See the updated patch in
> the previous mail.

Er?  The only patch I see is removal of RCU from shrink_dentry_list(), which
is fine, but doesn't do anything of that sort.  What was the Message-ID?

Let me see if I understand what you are proposing:
	dentry_kill(dentry, 0) seeing DCACHE_DENTRY_KILLED
		check DCACHE_MAY_FREE, free it's been set
	dentry_kill(dentry, 1) seeing that we are on shrinker list
		leave on the list, do the work on killing, retake ->d_lock,
		if we are still on shrinker list
			set DCACHE_MAY_FREE,
		else
			free it

That would probably work...