Re: [PATCH 2/7] fs, xfs: convert xfs_buf.b_hold and xfs_buf.b_lru_ref from atomic_t to refcount_t

[Dave Chinner]

On Tue, Feb 21, 2017 at 05:04:08PM +0100, Peter Zijlstra wrote:
> On Tue, Feb 21, 2017 at 05:49:02PM +0200, Elena Reshetova wrote:
> > @@ -1684,10 +1684,11 @@ xfs_buftarg_isolate(
> >  	 * zero. If the value is already zero, we need to reclaim the
> >  	 * buffer, otherwise it gets another trip through the LRU.
> >  	 */
> > -	if (!atomic_add_unless(&bp->b_lru_ref, -1, 0)) {
> > +	if (!refcount_read(&bp->b_lru_ref)) {
> >  		spin_unlock(&bp->b_lock);
> >  		return LRU_ROTATE;
> >  	}
> > +	refcount_dec_and_test(&bp->b_lru_ref);
> >  
> >  	bp->b_state |= XFS_BSTATE_DISPOSE;
> >  	list_lru_isolate_move(lru, item, dispose);
> 
> This should never have passed testing.. refcount_dec_and_test() has a
> __must_check.
> 
> Furthermore the above seems to suggest thingies can live with a 0
> refcount, so a straight conversion cannot work.

Yes, 0 is a valid value - the buffer lru reference is *not an object
lifecycle reference count*. A value of zero means reclaim needs to
take action if it sees that value - it does not mean that the object
is not referenced by anyone (that's b_hold). i.e.  b_lru_ref is an
"active reference weighting" used to provide a heirarchical reclaim
bias toward less important metadata objects, and has no bearing on
the actual active users of the object.

Cheers,

Dave.
-- 
Dave Chinner
david@xxxxxxxxxxxxx