Re: [announce] vfs-scale git tree update

[Ian Kent <raven@themaw.net>]

On Wed, 2011-01-12 at 12:41 +0800, Ian Kent wrote:
> On Tue, 2011-01-11 at 20:06 -0800, Linus Torvalds wrote:
> > On Tue, Jan 11, 2011 at 7:59 PM, Ian Kent <raven@themaw.net> wrote:
> > >
> > > Isn't the parent i_mutex held during mkdir()?
> > 
> > Yes, but a lookup that hits in the dentry cache won't actually take
> > the parent mutex.
> > 
> > So as far as I can tell, doing the "d_add()" before setting d_op can
> > result in another CPU coming in and seeing the newly added dentry
> > before d_op has actually been initialized. Exactly because it will do
> > the dentry lookup without holding any mutex.
> > 
> > Of course, it's a very small window, so it probably doesn't matter in practice.
> > 
> > >> Looking at it quickly, I don't think that would matter for
> > >> the case at hand.  I.e., that might be safer but it doesn't
> > >> address the fact that these fields are getting initialized
> > >> multiple times.
> > >
> > > Yeah, a hangover from changes done over time.
> > > Not setting the dentry op in ->lookup() should fix this.
> > 
> > Alex, care to test just removing the d_set_d_op() call from autofs4_lookup()?
> > 
> > (That code is a bit scary, though - it explicitly makes it a negative
> > dentry with a d_instantiate(dentry, NULL), and then hides the inode
> > information away separately. Scary scary)
> 
> Yeah, but the expire to mount races with autofs are difficult to handle
> and this approach has worked well under heavy stress testing. It's true
> that this would almost certainly be bad for a file system that supported
> a full range of functionality but that's not so for autofs.

I think I have to partly take this back.
With Nick's recent vfs-scale patches this may not be OK any more since
the dcache_lock has gone away and, at first glance, it looks like the
added autofs4_lock spin lock doesn't provide the needed protection.

Ian