Re: [PATCH 1/3] xfs: fix the AG loop in xfs_alloc_vextent for busy extents

[Brian Foster <bfoster@redhat.com>]

On Fri, Oct 21, 2016 at 02:48:24PM +0200, Christoph Hellwig wrote:
> On Wed, Oct 19, 2016 at 09:48:08AM -0400, Brian Foster wrote:
> > >  			if (args->agbp) {
> > > -				if ((error = xfs_alloc_ag_vextent(args)))
> > > +				error = xfs_alloc_ag_vextent(args);
> > > +				if (error)
> > >  					goto error0;
> > > -				break;
> > > +				if (args->agbno != NULLAGBLOCK)
> > > +					break;
> > > +				xfs_trans_brelse(args->tp, args->agbp);
> > 
> > How is this safe with respect to xfs_alloc_fix_freelist() potentially
> > dirtying the agf? Haven't we had deadlock and/or other problems due to
> > xfs_alloc_fix_freelist() succeeding when an allocation ultimately fails,
> > and subsequently rotating to a potentially lower agno?
> 
> We've had the case where the allocation would fail despite
> xfs_alloc_fix_freelist succeeding forever, it's just that with
> discard in general and async discard in particular we can hit it
> much more easily.
> 

Perhaps, but I think that is beside the point. IME, any time that has
resulted in such a deadlock, we attribute it to the fact that
xfs_alloc_fix_freelist() succeeded in a case where it shouldn't have. I
don't think we should introduce more of such cases if we can help it.

> The only way to make the allocation no fail if xfs_alloc_fix_freelist
> succeeded would be to force out any busy extents at the low level
> if we are tigh on space, I'll have to see how doable that would be.
> 
> The other option would be to roll around the transaction when switching
> to a different AG so that we can release the locks.  That sounds a lot
> easier, and also less fragile in the long run.

That sounds reasonable provided we don't have any partial modifications
or whatnot in the transaction. Another angle might be to find a way to
take the busy extents into consideration before xfs_alloc_fix_freelist()
actually makes any changes, but I also don't know how straightforward
that might be either.

Brian

> --
> To unsubscribe from this list: send the line "unsubscribe linux-xfs" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html