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

["J. Bruce Fields" <bfields@fieldses.org>]

On Thu, Jun 07, 2012 at 08:36:07PM +0100, Al Viro wrote:
> 	* d_splice_alias() - broken.  Called without any locking on the
> old directory, nevermind ->s_vfs_rename_mutex.

I was assuming that the callers were in lookup, held i_mutex on a
parent, and that in the case of a directory, existence of an alias with
a different parent could only result from a filesystem bug.

> I really believe that this pair of commits needs to be reverted.  The
> earlier code used to guarantee that alias would be detached.

In the case that prompted that first commit, the directory in question
had an alias that was detached (which I'm taking to mean IS_ROOT(dentry)
was true?), but not flagged DISCONNECTED.  The particular case was only
reproduceable on an older kernel, and I couldn't find a similar
reproducer on recent upstream, but I also couldn't convince myself it
was impossible.

So, maybe the correct thing is to revert that change.  Or maybe it
should be picking an IS_ROOT dentry instead of a DISCONNECTED one?

There's some previous discussion at

	http://marc.info/?i=<20110310105821.GE22723@ZenIV.linux.org.uk>

(in particular a long post from Neil:

	http://marc.info/?i=<20110311150749.2fa2be66@notabene.brown>

that I should review.)

--b.