Re: [PATCH 15/25] mm, compaction: Finish pageblock scanning on contention

[Mel Gorman <mgorman@techsingularity.net>]

On Thu, Jan 17, 2019 at 05:38:36PM +0100, Vlastimil Babka wrote:
> > rate but also by the fact that the scanners do not meet for longer when
> > pageblocks are actually used. Overall this is justified and completing
> > a pageblock scan is very important for later patches.
> > 
> > Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
> 
> Acked-by: Vlastimil Babka <vbabka@suse.cz>
> 
> Some comments below.
> 

Thanks

> > @@ -538,18 +535,8 @@ static unsigned long isolate_freepages_block(struct compact_control *cc,
> >  		 * recheck as well.
> >  		 */
> >  		if (!locked) {
> > -			/*
> > -			 * The zone lock must be held to isolate freepages.
> > -			 * Unfortunately this is a very coarse lock and can be
> > -			 * heavily contended if there are parallel allocations
> > -			 * or parallel compactions. For async compaction do not
> > -			 * spin on the lock and we acquire the lock as late as
> > -			 * possible.
> > -			 */
> > -			locked = compact_trylock_irqsave(&cc->zone->lock,
> > +			locked = compact_lock_irqsave(&cc->zone->lock,
> >  								&flags, cc);
> > -			if (!locked)
> > -				break;
> 
> Seems a bit dangerous to continue compact_lock_irqsave() to return bool that
> however now always returns true, and remove the safety checks that test the
> result. Easy for somebody in the future to reintroduce some 'return false'
> condition (even though the name now says lock and not trylock) and start
> crashing. I would either change it to return void, or leave the checks in place.
> 

I considered changing it from bool at the same time as "Rework
compact_should_abort as compact_check_resched". It turned out to be a
bit clumsy because the locked state must be explicitly updated in the
caller then. e.g.

locked = compact_lock_irqsave(...)

becomes

compact_lock_irqsave(...)
locked = true

I didn't think the result looked that great to be honest but maybe it's
worth revisiting as a cleanup patch like "Rework compact_should_abort as
compact_check_resched" on top.

> > 
> > @@ -1411,12 +1395,8 @@ static void isolate_freepages(struct compact_control *cc)
> >  		isolate_freepages_block(cc, &isolate_start_pfn, block_end_pfn,
> >  					freelist, false);
> >  
> > -		/*
> > -		 * If we isolated enough freepages, or aborted due to lock
> > -		 * contention, terminate.
> > -		 */
> > -		if ((cc->nr_freepages >= cc->nr_migratepages)
> > -							|| cc->contended) {
> 
> Does it really make sense to continue in the case of free scanner, when we know
> we will just return back the extra pages in the end? release_freepages() will
> update the cached pfns, but the pageblock skip bit will stay, so we just leave
> those pages behind. Unless finishing the block is important for the later
> patches (as changelog mentions) even in the case of free scanner, but then we
> can just skip the rest of it, as truly scanning it can't really help anything?
> 

Finishing is important for later patches is one factor but not the only
factor. While we eventually return all pages, we do not know at this
point in time how many free pages are needed. Remember the migration
source isolates COMPACT_CLUSTER_MAX pages and then looks for migration
targets.  If the source isolates 32 pages, free might isolate more from
one pageblock but that's ok as the migration source may need more free
pages in the immediate future. It's less wasteful than it looks at first
glance (or second or even third glance).

However, if we isolated exactly enough targets, and the pageblock gets
marked skipped, then each COMPACT_CLUSTER_MAX isolation from the target
could potentially marge one new pageblock unnecessarily and increase
scanning+resets overall. That would be bad.

There still can be waste because we do not know in advance exactly how
many migration sources there will be -- sure, we could calculate it but
that involves scanning the source pageblock twice which is wasteful.
I did try estimating it based on the remaining number of pages in the
pageblock but the additional complexity did not appear to help.

Does that make sense?

-- 
Mel Gorman
SUSE Labs