[PATCH 1/4] xfs: add online discard support

[Christoph Hellwig]

On Thu, May 19, 2011 at 04:53:44PM -0500, Alex Elder wrote:
> The first is, why not support it for non-delaylog?

Because:

 a) performance is going to suck even more horribly with the
    amount of trim commands needed, with no chance of actually
    fixing it
 b) the async discard code in patch 3 not easily applyable to
    the non-delaylog case, we'd need to keep two parallel codebases,
    one of them guaranteed to be untested.

> Second, why is it a two phase operation (marking an
> extent for discard, then doing all the discards at
> once)?  Is it just so you can do the discards without
> holding the perag lock?

Because we must prevent the allocation code from reusing an extent
that is undergoing a discard right now to prevent corruption, thus
we need to mark it as do not touch first. 

> >  	xfs_trans_committed_bulk(ctx->cil->xc_log->l_ailp, ctx->lv_chain,
> >  					ctx->start_lsn, abort);
> >  
> >  	xfs_alloc_busy_sort(&ctx->busy_extents);
> 
> I still think sorting the list belongs inside xfs_alloc_busy_clear().
> I see that list_sort() is not necessarily trivial for an already
> sorted list though...

It's a bad idea to do the sort twice for no good reason, and adding
another parameter to further overload xfs_alloc_busy_clear behaviour
doesn't seem smart either.

> 		if (error == EOPNOTSUPP) {
> 			/*
> 			 * Report this once per mount point somehow?
> 			 * If so, turn off the mount option?
> 			 */
> 			break;

We've been through this discussion again lately with dm and ext4
folks, and the conclusion is that EOPNOTSUPP is perfectly fine to happen
here.

> > +	 * performing the discard a chance to mark the extent unbusy
> > +	 * and retry.
> > +	 */
> > +	if (busyp->flags & XFS_ALLOC_BUSY_DISCARDED) {
> > +		spin_unlock(&pag->pagb_lock);
> > +		delay(1);
> 
> I hate seeing calls to delay() although sometimes
> it's the right thing to do...  I don't have a feel
> for how long a discard is likely to take so I don't
> know whether waiting here instead would be worth
> the effort.

It's not nice, but if the block layer gets fixed and we do asynchronous
discards it simply goes away.

> If this option is to only be available for delaylog, it should
> say so here (and maybe report that it's being ignored if it's
> supplied with nodelaylog at mount time).

ok.