On 1/25/23 14:44, Mel Gorman wrote:
> Commit 7efc3b726103 ("mm/compaction: fix set skip in
> fast_find_migrateblock") address an issue where a pageblock selected
> by fast_find_migrateblock() was ignored. Unfortunately, the same fix
> resulted in numerous reports of khugepaged or kcompactd stalling for
> long periods of time or consuming 100% of CPU.
>
> Tracing showed that there was a lot of rescanning between a small subset
> of pageblocks because the conditions for marking the block skip are not
> met. The scan is not reaching the end of the pageblock because enough
> pages were isolated but none were migrated successfully. Eventually it
> circles back to the same block.
>
> Pageblock skip tracking tries to minimise both latency and excessive
> scanning but tracking exactly when a block is fully scanned requires an
> excessive amount of state. This patch forcibly rescans a pageblock when
> all isolated pages fail to migrate even though it could be for transient
> reasons such as page writeback or page dirty. This will sometimes migrate
> too many pages but pageblocks will be marked skip and forward progress
> will be made.
>
> "Usemen" from the mmtests configuration
> workload-usemem-stress-numa-compact was used to stress compaction.
> The compaction trace events were recorded using a 6.2-rc5 kernel that
> includes commit 7efc3b726103 and count of unique ranges were measured. The
> top 5 ranges were
>
> 3076 range=(0x10ca00-0x10cc00)
> 3076 range=(0x110a00-0x110c00)
> 3098 range=(0x13b600-0x13b800)
> 3104 range=(0x141c00-0x141e00)
> 11424 range=(0x11b600-0x11b800)
>
> While this workload is very different than what the bugs reported,
> the pattern of the same subset of blocks being repeatedly scanned is
> observed. At one point, *only* the range range=(0x11b600 ~ 0x11b800)
> was scanned for 2 seconds. 14 seconds passed between the first
> migration-related event and the last.
>
> With the series applied including this patch, the top 5 ranges were
>
> 1 range=(0x11607e-0x116200)
> 1 range=(0x116200-0x116278)
> 1 range=(0x116278-0x116400)
> 1 range=(0x116400-0x116424)
> 1 range=(0x116424-0x116600)
>
> Only unique ranges were scanned and the time between the first
> migration-related event was 0.11 milliseconds.
>
> Fixes: 7efc3b726103 ("mm/compaction: fix set skip in fast_find_migrateblock")
> Signed-off-by: Mel Gorman <mgorman@xxxxxxxxxxxxxxxxxxx>
While this seems like it will mostly prevent the issue at hand (I think
kcompactd is still a hazard, see below), I'm not super happy about some of
the implementation details.
1. it modifies code that was meant to quickly skip an order-aligned block
where a page migration failed during MIGRATE_ASYNC direct compaction, as
it's very unlikely to sucessfully form a free page of that order in that
block. Now instead it will finish the whole pageblock in that case, which
could be lots of useless work and thus exactly the opposite of what we
wanted for MIGRATE_ASYNC direct compaction.
2. The conditions "cc->direct_compaction" and "(cc->mode < MIGRATE_SYNC)"
seem a bit hazardous. I think we can have a compaction where these are not
true, and yet it uses fast_find_migrateblock() and thus can exhibit the bug
but won't be forced to rescan?
AFAICS kcompactd_do_work()
- is MIGRATE_SYNC_LIGHT
- has ignore_skip_hint = false
- has direct_compaction = false
so AFAICS it will use fast_find_migrateblock() and not bail out in one of
the preconditions. But the cc->direct_compaction condition here won't apply.
So it might be better to leave the current "skip the rest of block" check
alone, and add a separate check for the finish_pageblock rescan that will
not miss some cases where it should apply - maybe it could check for a
complete migration failure specifically as well?
> ---
> mm/compaction.c | 30 ++++++++++++++++++++++--------
> 1 file changed, 22 insertions(+), 8 deletions(-)
>
> diff --git a/mm/compaction.c b/mm/compaction.c
> index 4b3a0238879c..937ec2f05f2c 100644
> --- a/mm/compaction.c
> +++ b/mm/compaction.c
> @@ -2394,6 +2394,7 @@ compact_zone(struct compact_control *cc, struct capture_control *capc)
> cc->finish_pageblock = true;
> }
>
> +rescan:
> switch (isolate_migratepages(cc)) {
> case ISOLATE_ABORT:
> ret = COMPACT_CONTENDED;
> @@ -2436,15 +2437,28 @@ compact_zone(struct compact_control *cc, struct capture_control *capc)
> goto out;
> }
> /*
> - * We failed to migrate at least one page in the current
> - * order-aligned block, so skip the rest of it.
> + * If an ASYNC or SYNC_LIGHT fails to migrate a page
> + * within the current order-aligned block, scan the
> + * remainder of the pageblock. This will mark the
> + * pageblock "skip" to avoid rescanning in the near
> + * future. This will isolate more pages than necessary
> + * for the request but avoid loops due to
> + * fast_find_migrateblock revisiting blocks that were
> + * recently partially scanned.
> */
> - if (cc->direct_compaction &&
> - (cc->mode == MIGRATE_ASYNC)) {
> - cc->migrate_pfn = block_end_pfn(
> - cc->migrate_pfn - 1, cc->order);
> - /* Draining pcplists is useless in this case */
> - last_migrated_pfn = 0;
> + if (cc->direct_compaction && !cc->finish_pageblock &&
> + (cc->mode < MIGRATE_SYNC)) {
> + cc->finish_pageblock = true;
> +
> + /*
> + * Draining pcplists does not help THP if
> + * any page failed to migrate. Even after
> + * drain, the pageblock will not be free.
> + */
> + if (cc->order == COMPACTION_HPAGE_ORDER)
> + last_migrated_pfn = 0;
> +
> + goto rescan;
> }
> }
>
