Re: processes hung after sys_renameat, and 'missing' processes

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

On Wed, Jun 06, 2012 at 06:45:35PM -0700, Linus Torvalds wrote:
> On Wed, Jun 6, 2012 at 6:31 PM, Al Viro <viro@zeniv.linux.org.uk> wrote:
> >
> > And you want to check i_mutex on old parent, not the file being moved
> > itself. ?IOW, the second one should be
> >
> > if (dentry->d_parent != dentry)
> > ? ? ? ?WARN_ON_ONCE(!mutex_is_locked(&dentry->d_parent->d_inode->i_mutex));
> 
> Do you even need the "dentry->d_parent != dentry" test? When do we
> even rename root dentries?

Root of the filesystem - never; root of disconnected subtree - sure,
that's how they become connected to the tree.  See d_materialize_unique(),
for example...

BTW, I really need more coffee - those checks belong in __d_move(),
not d_move(); aforementioned d_materialize_unique() doesn't use d_move(),
it calls __d_move() directly.  Sorry.