Re: dcache shrink list corruption?

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

On Wed, Apr 30, 2014 at 01:57:05PM -0700, Linus Torvalds wrote:
> On Wed, Apr 30, 2014 at 1:38 PM, Al Viro <viro@zeniv.linux.org.uk> wrote:
> >
> > We do not (and cannot) call dentry_kill() with rcu_read_lock held - it can
> > trigger any amount of IO, for one thing.  We can take it around the
> > couple of places where do that spin_unlock(&dentry->d_lock) (along with
> > setting DCACHE_RCUACCESS) - that's what I'd been refering to.
> 
> Just the last spin_unlock() would be the case that matters, if the
> spin_unlock() is done on something that could be freed immediately and
> the lock protects and is inside the entity that gets freed.

*nod*

There are two such spin_unlock (handover from shrink_dentry_list() to
dput() and the opposite one), but they are all that needs protection -
->d_flags update is outside the rcu-critical area.  I really wonder
if we *can* get there without DCACHE_RCUACCESS having been set, though;
dentry would have to be
	* picked into shrink list (i.e. have had zero refcount at some point)
	* never had been through __d_rehash()
shrink_dentry_list() definitely counts on that being impossible, and it
probably is, but I'm feeling seriously paranoid about the whole area.
I'll finish grepping through the tree and probably drop setting
DCACHE_RCUACCESS from the patch - either that, or set it in d_shrink_add()
it it turns out that it is possible and shrink_dentry_list() is fucked...

Tests seem to be running fine so far...

> >   BTW, is there any convenient
> > way to tell git commit --amend to update the commit date?  Something
> > like --date=now would be nice, but it isn't accepted...
> 
> --date="$(date)" works.

Thanks...