On Fri, Nov 29, 2013 at 04:06:59AM +0000, Al Viro wrote:
> On Fri, Nov 29, 2013 at 03:59:39AM +0000, Al Viro wrote:
> > On Fri, Nov 29, 2013 at 02:41:21AM +0000, Al Viro wrote:
> > > On Thu, Nov 28, 2013 at 06:07:27PM -0800, Linus Torvalds wrote:
> > >
> > > > HOWEVER. It's certainly *not* valid if "current->fs->root/pwd" points
> > > > to it. So yeah, there must have been an extra dput() somewhere. Or,
> > > > more likely, I think, we don't get the refcount to some dentry
> > > > properly any more.
> > > >
> > > > I don't see where, though. You did change where "LOOKUP_RCU" is
> > > > cleared in unlazy_walk() but you did add that
> > > >
> > > > nd->path.dentry = NULL;
> > > >
> > > > and that looks like it should be ok. And I don't see what else would care.
> > >
> > > *nod*
> > >
> > > BTW, vfsmount refcount is 12, so we *definitely* nowhere near the
> > > final mntput(), etc. and mnt->mnt_root itself should also have
> > > contributed.
> > >
> > > I'm going to try to find out _which_ test buggers the refcount - at
> > > least that way I'll have something resembling a usable reproducer...
> >
> > OK, we have a winner. generic/234 drops refcount of root dentry by about
> > 20 (and yes, I should've started with that one, what with Ted's report).
> > Run it several times (4 should suffice nicely) and the damn thing triggers
> > right there. Uff... At least that takes under a minute instead of a couple
> > of hours, which makes debugging that shite much more tolerable...
>
> I think I see what's going on; it *is* unlazy_walk(), but not nd->path.
> It's nd->root. IOW, the relevant fix to fs/namei.c is
>
> @@ -513,8 +513,7 @@ static int unlazy_walk(struct nameidata *nd, struct dentry *dentry)
>
> if (!lockref_get_not_dead(&parent->d_lockref)) {
> nd->path.dentry = NULL;
> - rcu_read_unlock();
> - return -ECHILD;
> + goto out;
> }
>
> /*
>
> (in addition to other pieces of fun found in process). I'll test and post
> results in a few...
And yes, it has fixed the problem with generic/234. I'll do full xfstests
run to see if there's anything else, but this one is obviously needed.
I'll send it with sane commit message (along with follow_dotdot_rcu()
fix) later tonight. path_init() race is a separate story - that one should
probably go separately, since we'll want it in all branches starting with
early 2011 or so.
_______________________________________________
xfs mailing list
xfs@xxxxxxxxxxx
http://oss.sgi.com/mailman/listinfo/xfs
