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

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

On Mon, Jun 04, 2012 at 12:40:42AM +0100, Al Viro wrote:
> On Mon, Jun 04, 2012 at 12:28:20AM +0100, Al Viro wrote:
> 
> > Everything in lock_rename() appears to be at lock_rename+0x3e.  Unless
> > there's a really huge amount of filesystems on that box, this has to
> > be
> >                 mutex_lock_nested(&p1->d_inode->i_mutex, I_MUTEX_PARENT);
> > and everything on that sucker is not holding any locks yet.  IOW, that's
> > the tail hanging off whatever deadlock is there.
> 
> Er...  After another look, probably not - it's ->s_vfs_rename_mutex,
> so we are seeing one cross-directory rename stuck on something with
> all subsequent ones blocked on attempt to grab said mutex.
> 
> The interesting one is the guy stuck at lock_rename+0xc9/0xf0, everything
> else in lock_rename() is the consequence.

BTW, another suspicious patch is d_splice_alias() one; note that if we
_ever_ pick a dentry that isn't disconnected, we are deeply fucked.
d_move() without the old parent locked is a Bad Thing(tm).  I don't see
how that could've triggered without another bug somewhere, but what's
happening in d_splice_alias() right now is wrong.