Re: [git pull] vfs and fs fixes

["J. Bruce Fields" <bfields@fieldses.org>]

On Tue, Apr 17, 2012 at 03:08:26PM -0700, Linus Torvalds wrote:
> On Tue, Apr 17, 2012 at 2:14 PM, J. Bruce Fields <bfields@fieldses.org> wrote:
> > On Tue, Apr 17, 2012 at 07:28:26PM +0100, Al Viro wrote:>
> > Maybe instead I could continue using the i_mutex but handle rename some
> > other way; e.g. in delegation code:
> >
> >        if (!mutex_trylock(inode->i_mutex))
> >                return -EAGAIN;
> >        if (atomic_read(inode->i_renames_in_progress))
> >                return -EAGAIN;
> >
> > and add an
> >
> >        atomic_inc(inode->i_renames_in_progress);
> >        atomic_dec(inode->i_renames_in_progress);
> >
> > pair around rename.
> 
> Please don't make up your own locking. Plus it's broken anyway, since
> a rename could come in directly after your atomic_read (and this is
> *why* people shouldn't make up their own locks - they are invariably
> broken).

Doh, yes, sounds like a good rule.  (I was misremembering some previous
attempt at this--which admittedly may just have failed in some more
complicated way.)

--b.

> > Or I could increment that counter for all the conflicting operations and
> > rely on it instead of the i_mutex.  I was trying to avoid adding
> > something like that (an inc, a dec, another error path) to every
> > operation.  And hoping to avoid adding another field to struct inode.
> > Oh well.
> 
> We could just say that we can do a double inode lock, but then
> standardize on the order. And the only sane order is comparing inode
> pointers, not inode numbers like ext4 apparently does.
> 
> With a standard order, I don't think it would be at all wrong to just
> take the inode lock on rename.
> 
>                         Linus