lkml.org 
[lkml]   [2010]   [Mar]   [25]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 07/11] Memory compaction core
On Wed, Mar 24, 2010 at 01:33:47PM -0700, Andrew Morton wrote:
> On Tue, 23 Mar 2010 12:25:42 +0000
> Mel Gorman <mel@csn.ul.ie> wrote:
>
> > This patch is the core of a mechanism which compacts memory in a zone by
> > relocating movable pages towards the end of the zone.
> >
> > A single compaction run involves a migration scanner and a free scanner.
> > Both scanners operate on pageblock-sized areas in the zone. The migration
> > scanner starts at the bottom of the zone and searches for all movable pages
> > within each area, isolating them onto a private list called migratelist.
> > The free scanner starts at the top of the zone and searches for suitable
> > areas and consumes the free pages within making them available for the
> > migration scanner. The pages isolated for migration are then migrated to
> > the newly isolated free pages.
>
> General comment: it looks like there are some codepaths which could
> hold zone->lock for a long time. It's unclear that they're all
> constrained by COMPACT_CLUSTER_MAX. Is there a a latency issue here?
>

I don't think so. There are two points where zone-related locks are
held.

zone->lock is held in isolate_freepages() while it gets the free pages
necessary for migration to complete. The size of the list of pages
being migrated is constrained by COMPACT_CLUSTER_MAX so it is bounded
by that. Worst case scenario is the zone is almost fully
scanned.

zone->lru_lock is held in isolate_migratepages) while it gets pages for
migration. It's released if COMPACT_CLUSTER_MAX pages are
isolated. Again, worst case scenario is that the zone is
almost fully scanned.

The worst-case scenario in both cases is the lock is held while the zone
is scanned. The concern would be if we managed to scan almost a full
zone and that zone is very large. I could add an additional check to
release the lock when a large number of pages has been scanned but I
don't think it's necessary. I find it very unlikely that a large zone
would not have COMPACT_CLUSTER_MAX pages found quickly for isolation.

> >
> > ...
> >
> > +static struct page *compaction_alloc(struct page *migratepage,
> > + unsigned long data,
> > + int **result)
> > +{
> > + struct compact_control *cc = (struct compact_control *)data;
> > + struct page *freepage;
> > +
> > + VM_BUG_ON(cc == NULL);
>
> It's a bit strange to test this when we're about to oops anyway. The
> oops will tell us the same thing.
>

It was paranoia after the bugs related to NULL-offsets but unnecessary
paranoia in this case. It would require migration to be very broken for it to
trigger. Even if it was, I cannot imagine a case where it would be exploited
because it's a small structure and not offset by any userspace-supplied
piece of data. I will drop the check.

> > + /* Isolate free pages if necessary */
> > + if (list_empty(&cc->freepages)) {
> > + isolate_freepages(cc->zone, cc);
> > +
> > + if (list_empty(&cc->freepages))
> > + return NULL;
> > + }
> > +
> > + freepage = list_entry(cc->freepages.next, struct page, lru);
> > + list_del(&freepage->lru);
> > + cc->nr_freepages--;
> > +
> > + return freepage;
> > +}
>

--
Mel Gorman
Part-time Phd Student Linux Technology Center
University of Limerick IBM Dublin Software Lab


\
 
 \ /
  Last update: 2010-03-25 10:17    [from the cache]
©2003-2014 Jasper Spaans. Advertise on this site