Re: [patch 3/6] vfs: mountinfo stable peer group id

[Ram Pai <linuxram@us.ibm.com>]

On Sat, 2008-03-22 at 04:11 +0000, Al Viro wrote:
> On Sat, Mar 22, 2008 at 03:49:50AM +0000, Al Viro wrote:
> 
> >       Shifting increment from mnt_set_mountpoint() and commit_tree()
> > to theirs callers and collapsing where possible, we get the
> following:
> >       * decrement in release_mounts() when resetting ->mnt_parent
> >       * increment in propagate_mnt() after call of
> mnt_set_mountpoint()
> >       * decrement in attach_recursive_mnt() in the loop calling
> > commit_tree() for clones (on mountpoint of each clone).
> >       * increment in umount_tree() at the point where we update
> d_mounted.
> 
> ... except that it'd give a leak in case of mount to shared mountpoint
> failing halfway through - we'll get double increments since
> umount_tree()
> would hit the mountpoints of cloned trees with extra increment, even
> though 
> reference from root of cloned to its mountpoint is _already_ a ghost.

> OTOH, we probably don't want to bother with counting those anyway -
> i.e.
> it's simply a bad definition and the right one would be along the
> lines of
> "number of vfsmounts that are doomed to be eaten by release_mounts()
> and
> that have ->mnt_parent pointing to us".  IOW, dropping the 2nd and 3rd
> in the above would do the right thing - anything chewed by
> umount_tree()
> *will* go to release_mounts() and ones in flight are what we are
> interested
> in... 

By not accounting for the ghost reference created in propagate_mnt(),
i.e case 2 and 3; the race is still on with shrink_mounts. But I think,
you are right. We don't want the shrink_mounts and friends to think that
the mounts are available to be purged, by accounting them into
mnt_ghosts.

RP