Re: fs_stack/eCryptfs: remove 3rd arg of copy_attr_all, add locking to copy_inode_size

[Hugh Dickins <hugh@veritas.com>]

On Thu, 3 Apr 2008, Josef 'Jeff' Sipek wrote:
> On Wed, Apr 02, 2008 at 09:49:11PM -0400, Erez Zadok wrote:
> ...
> > +#if BITS_PER_LONG == 32 && defined(CONFIG_SMP)
> > +	spin_lock(&dst->i_lock);
> > +#endif
> 
> I think you need to check CONFIG_PREEMPT as well.

Don't i_size_read() and i_size_write() handle that case adequately
themselves with preempt_dis/enable?

If you mean it'd be better to have the preempt_dis/enable bracketing
around the i_size_read() along with the i_size_write(): well, could
do, but if lower i_size is changing underneath us in that way,
it can just as well change again a moment later, so no point.

The bit that worries me, that prevents me from saying Acked-by,
is that I'm unsure of the extent to which lower filesystems are
relying on i_mutex or something else for their i_size serialization.
Using the i_lock here plays well with unionfs internally, and fixes
the hangs I used to see; but relying on one locking scheme here and
another below is uncomfortable for me.  But it is an improvement.

Hugh