On 01/20/2015 04:18 PM, Zhang Yanfei wrote:
Hello Vlastimil
在 2015/1/19 18:05, Vlastimil Babka 写道:
Compaction employs two page scanners - migration scanner isolates pages to be
the source of migration, free page scanner isolates pages to be the target of
migration. Currently, migration scanner starts at the zone's first pageblock
and progresses towards the last one. Free scanner starts at the last pageblock
and progresses towards the first one. Within a pageblock, each scanner scans
pages from the first to the last one. When the scanners meet within the same
pageblock, compaction terminates.
One consequence of the current scheme, that turns out to be unfortunate, is
that the migration scanner does not encounter the pageblocks which were
scanned by the free scanner. In a test with stress-highalloc from mmtests,
the scanners were observed to meet around the middle of the zone in first two
phases (with background memory pressure) of the test when executed after fresh
reboot. On further executions without reboot, the meeting point shifts to
roughly third of the zone, and compaction activity as well as allocation
success rates deteriorates compared to the run after fresh reboot.
It turns out that the deterioration is indeed due to the migration scanner
processing only a small part of the zone. Compaction also keeps making this
bias worse by its activity - by moving all migratable pages towards end of the
zone, the free scanner has to scan a lot of full pageblocks to find more free
pages. The beginning of the zone contains pageblocks that have been compacted
as much as possible, but the free pages there cannot be further merged into
larger orders due to unmovable pages. The rest of the zone might contain more
suitable pageblocks, but the migration scanner will not reach them. It also
isn't be able to move movable pages out of unmovable pageblocks there, which
affects fragmentation.
This patch is the first step to remove this bias. It allows the compaction
scanners to start at arbitrary pfn (aligned to pageblock for practical
purposes), called pivot, within the zone. The migration scanner starts at the
exact pfn, the free scanner starts at the pageblock preceding the pivot. The
direction of scanning is unaffected, but when the migration scanner reaches
the last pageblock of the zone, or the free scanner reaches the first
pageblock, they wrap and continue with the first or last pageblock,
respectively. Compaction terminates when any of the scanners wrap and both
meet within the same pageblock.
For easier bisection of potential regressions, this patch always uses the
first zone's pfn as the pivot. That means the free scanner immediately wraps
to the last pageblock and the operation of scanners is thus unchanged. The
actual pivot changing is done by the next patch.
Signed-off-by: Vlastimil Babka <vbabka@xxxxxxx>
I read through the whole patch, and you can feel free to add:
Acked-by: Zhang Yanfei <zhangyanfei@xxxxxxxxxxxxxx>
Thanks.
I agree with you and the approach to improve the current scheme. One thing
I think should be carefully treated is how to avoid migrating back and forth
since the pivot pfn can be changed. I see patch 5 has introduced a policy to
change the pivot so we can have a careful observation on it.
(The changes in the patch make the code more difficult to understand now...
and I just find a tiny mistake, please see below)
Cc: Minchan Kim <minchan@xxxxxxxxxx>
Cc: Mel Gorman <mgorman@xxxxxxx>
Cc: Joonsoo Kim <iamjoonsoo.kim@xxxxxxx>
Cc: Michal Nazarewicz <mina86@xxxxxxxxxx>
Cc: Naoya Horiguchi <n-horiguchi@xxxxxxxxxxxxx>
Cc: Christoph Lameter <cl@xxxxxxxxx>
Cc: Rik van Riel <riel@xxxxxxxxxx>
Cc: David Rientjes <rientjes@xxxxxxxxxx>
---
include/linux/mmzone.h | 2 +
mm/compaction.c | 204 +++++++++++++++++++++++++++++++++++++++++++------
mm/internal.h | 1 +
3 files changed, 182 insertions(+), 25 deletions(-)
diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
index 2f0856d..47aa181 100644
--- a/include/linux/mmzone.h
+++ b/include/linux/mmzone.h
@@ -503,6 +503,8 @@ struct zone {
unsigned long percpu_drift_mark;
#if defined CONFIG_COMPACTION || defined CONFIG_CMA
+ /* pfn where compaction scanners have initially started last time */
+ unsigned long compact_cached_pivot_pfn;
/* pfn where compaction free scanner should start */
unsigned long compact_cached_free_pfn;
/* pfn where async and sync compaction migration scanner should start */
diff --git a/mm/compaction.c b/mm/compaction.c
index 5626220..abae89a 100644
--- a/mm/compaction.c
+++ b/mm/compaction.c
@@ -123,11 +123,16 @@ static inline bool isolation_suitable(struct compact_control *cc,
return !get_pageblock_skip(page);
}
+/*
+ * Invalidate cached compaction scanner positions, so that compact_zone()
+ * will reinitialize them on the next compaction.
+ */
static void reset_cached_positions(struct zone *zone)
{
- zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn;
- zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn;
- zone->compact_cached_free_pfn = zone_end_pfn(zone);
+ /* Invalid values are re-initialized in compact_zone */
+ zone->compact_cached_migrate_pfn[0] = 0;
+ zone->compact_cached_migrate_pfn[1] = 0;
+ zone->compact_cached_free_pfn = 0;
}
/*
@@ -172,11 +177,35 @@ void reset_isolation_suitable(pg_data_t *pgdat)
/* Only flush if a full compaction finished recently */
if (zone->compact_blockskip_flush) {
__reset_isolation_suitable(zone);
- reset_cached_positions(zone);
+ reset_cached_positions(zone, false);
The second argument should be in patch 5 instead of here, right? ^-^
Yes, a mistake with some last-minute rebasing :)
Thanks.
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>