Re: [PATCH -mm] mm: have order>0 compaction start off where it left

[Rik van Riel]

On 06/28/2012 06:29 AM, Mel Gorman wrote:

> Lets say there are two parallel compactions running. Process A meets
> the migration PFN and moves to the end of the zone to restart. Process B
> finishes scanning mid-way through the zone and updates last_free_pfn. This
> will cause Process A to "jump" to where Process B left off which is not
> necessarily desirable.
>
> Another side effect is that a workload that allocations/frees
> aggressively will not compact as well as the "free" scanner is not
> scanning the end of the zone each time. It would be better if
> last_free_pfn was updated when a full pageblock was encountered
>
> So;
>
> 1. Initialise last_free_pfn to the end of the zone
> 2. On compaction, scan from last_free_pfn and record where it started
> 3. If a pageblock is full, update last_free_pfn
> 4. If the migration and free scanner meet, reset last_free_pfn and
>     the free scanner. Abort if the free scanner wraps to where it started
>
> Does that make sense?

Yes, that makes sense.  We still have to keep track
of whether we have wrapped around, but I guess that
allows for a better name for the bool :)

Maybe cc->wrapped?

Does anyone have a better name?

As for point (4), should we abort when we wrap
around to where we started, or should we abort
when free_pfn and migrate_pfn meet after we
wrapped around?

> > diff --git a/mm/internal.h b/mm/internal.h
> > index 2ba87fb..b041874 100644
> > --- a/mm/internal.h
> > +++ b/mm/internal.h
> > @@ -120,6 +120,7 @@ struct compact_control {
> >   	unsigned long free_pfn;		/* isolate_freepages search base */
> >   	unsigned long migrate_pfn;	/* isolate_migratepages search base */
> >   	bool sync;			/* Synchronous migration */
> > +	bool last_round;		/* Last round for order>0 compaction */
>
> I don't get what you mean by last_round. Did you mean "wrapped". When
> false, it means the free scanner started from last_pfn and when true it
> means it started from last_pfn, met the migrate scanner and wrapped
> around to the end of the zone?

Yes, I do mean "wrapped" :)

--
All rights reversed
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/