Re: [git pull] vfs and fs fixes

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

On Tue, Apr 17, 2012 at 08:01:48AM -0700, Linus Torvalds wrote:
> On Mon, Apr 16, 2012 at 10:25 PM, Al Viro <viro@zeniv.linux.org.uk> wrote:
> > ? ? ? ?A bunch of endianness fixes plus a patch from bfields untangling
> > dependencies between vfs and nfsd trees; in principle, we could keep it
> > in nfsd tree (along with a bunch of followups that definitely belong there),
> > but Miklos' stuff in fs/namei.c steps fairly close to it and overlayfs
> > and unionfs series - even closer, so that would create serious PITA for
> > both, whichever tree it would sit in.
> 
> Why is that double mutex taking in vfs_rename_other() safe from ABBA?
> 
> We aren't guaranteed to hold the s_vfs_rename_mutex, since the parent
> directories may be the same.
> 
> And yes, we hold the i_mutex on that shared parent, but the inodes may
> exist (hardlinked) in another directory, so another rename could be
> doing the i_mutex in the reverse order.
> 
> Maybe there is some reason why that double lock is safe, but I don't
> see it, and I want it clearly documented. So I'm not pulling this.

It isn't.  Hell knows - I wonder if taking s_vfs_rename_mutex in all cases
in lock_rename() would be the right thing to do; it would remove the
problem, but might cost us too much contention...