Re: 3.5-rc3: BUG: Dentry still in use (1) [unmount of cgroup cgroup]

[Al Viro <viro-3bDd1+5oDREiFSDQTTA3OLVCufUGDwFn@public.gmane.org>]

On Sat, Jun 30, 2012 at 12:04:36AM -0700, Tejun Heo wrote:
> Hello, Al.
> 
> On Fri, Jun 29, 2012 at 11:47 PM, Al Viro <viro-RmSDqhL/yNMiFSDQTTA3OLVCufUGDwFn@public.gmane.org> wrote:
> > On Sat, Jun 30, 2012 at 02:13:02PM +0800, Li Zefan wrote:
> >> So it's bad to have dentry refcnts dangling after umount.
> >
> > No shit. ?Yes, it is bad. ?What on the Earth is cgroup code doing with
> > those? ?And what could it possibly want to do with dentry reference
> > after the filesystem has been shut down, assuming it could hold one
> > in the first place?
> 
> cgroup interface code was copied from sysfs back when it was
> piggybacking internal data structures to dentries, so, unfortunately,
> sysfs is still using dentries to manage internal data structures and
> propagates internal refs to dentry refs. There seem to be several
> places where dentry ref is held w/o active super ref triggering BUG on
> umount. Longer term, it should be updated to share sysfs code, I
> guess.

Now that I've looked at that code again...  What's the story with
                simple_unlink(d->d_inode, d);
in cgroup_rm_file()?  Wrong parent inode, at the very least...

While we are at it, what the hell is going on in
static void cgroup_clear_directory(struct dentry *dir)
{
        struct cgroup *cgrp = __d_cgrp(dir);

        while (!list_empty(&cgrp->files))
                cgroup_rm_file(cgrp, NULL);
}
Are you fighting some kind of race against somebody adding stuff
there?  Unless I'm seriously misreading cgroup_rm_file(), it'll
do all the work on the first call...