Re: fs/namei.c: Misuse of sequence counts?

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

On Sat, Oct 11, 2014 at 11:01:42PM -0500, Eric Biggers wrote:
> On Sun, Oct 12, 2014 at 01:12:59AM +0100, Al Viro wrote:
> > 
> > Huh?  What's to guarantee that dentry hasn't become negative since the
> > moment we'd fetched the seqcount?  _That_ is the problem we are dealing
> > with here - link_path_walk() relies on nd->inode being non-NULL.
> 
> Hmm, I guess that makes sense.  So the code is actually verifying that the inode
> is still the inode that was referenced from the current or root directory when
> nd->path was set.  But couldn't the problem also be solved by setting nd->inode
> directly in the fs->seq retry loops?

Gets clumsy in set_root_rcu() - you do *not* want it to bugger nd->inode
when done by follow_dotdot_rcu(), so we'd need either some indication which
caller it is, or something like struct inode **inode in argument list,
with NULL passed from follow_dotdot_rcu(), while path_init() would give
it &nd->inode...

Doable, but unpleasant.  And the price of that check is trivial - after all,
in case we *don't* bugger off immediately, we have that ->d_seq in cache -
we'd fetched it just before.