Re: [PATCH v7 1/4] spinlock: A new lockref structure for lockless update of refcount

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

On Fri, Aug 30, 2013 at 05:10:45PM -0400, Waiman Long wrote:
> On 08/30/2013 04:54 PM, Al Viro wrote:
> >On Fri, Aug 30, 2013 at 01:43:11PM -0700, Linus Torvalds wrote:
> >>On Fri, Aug 30, 2013 at 1:15 PM, Waiman Long<waiman.long@hp.com>  wrote:
> >>>The prepend_path() isn't all due to getcwd. The correct profile should be
> >>Ugh. I really think that prepend_path() should just be rewritten to
> >>run entirely under RCU.
> >>
> >>Then we can remove *all* the stupid locking, and replace it with doing
                                                     ^^^^^^^^^^^^^^^^^^^^^
> >>a read-lock on the rename sequence count, and repeating if requited.
    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
> >>
> >>That shouldn't even be hard to do, it just requires mindless massaging
> >>and being careful.
> >Not really.  Sure, you'll retry it if you race with d_move(); that's not
> >the real problem - access past the end of the object containing ->d_name.name
> >would screw you and that's what ->d_lock is preventing there.  Delayed freeing
> >of what ->d_name is pointing into is fine, but it's not the only way to get
> >hurt there...
> 
> Actually, prepend_path() was called with rename_lock taken. So
> d_move() couldn't be run at the same time. Am I right?

See above.  You are right, but if Linus wants to turn that sucker into
reader (which is possible - see e.g. cifs build_path_from_dentry() and its
ilk), d_move() races will start to play.