Skip to contentSkip to conversation

Follow-up "Fix excessive CPU usage during compaction"

PATCH In review Mel Gorman 4 participants 20 messages May 15, 2023 – Jun 7, 2023

Subsystems (2)
lore
Share & tools
lore

Mel Gorman

author
#0 lore raw

cover Follow-up "Fix excessive CPU usage during compaction"

The series "Fix excessive CPU usage during compaction" [1] attempted to fix a bug [2] but Vlastimil noted that the fix was incomplete [3]. While the series was merged, fast_find_migrateblock was still disabled. This series should fix the corner cases and allow 95e7a450b819 ("Revert "mm/compaction: fix set skip in fast_find_migrateblock"") to be safely reverted. Details on how many pageblocks are rescanned are in the changelog of the last patch.

[1] https://lore.kernel.org/r/[email protected]
[2] https://bugzilla.suse.com/show_bug.cgi?id=1206848
[3] https://lore.kernel.org/r/[email protected]

 mm/compaction.c | 24 ++++++++++--------------
 1 file changed, 10 insertions(+), 14 deletions(-)
Signature · 1 line
2.35.3
author
#1 lore raw

PATCH 1/4 mm: compaction: Ensure rescanning only happens on partially scanned pageblocks

compact_zone() intends to rescan pageblocks if there is a failure to migrate "within the current order-aligned block". However, the pageblock scan may already be complete and moved to the next block causing the next pageblock to be "rescanned". Ensure only the most recent pageblock is rescanned.

---
 mm/compaction.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)
mm/compaction.c +3 -2
index c8bcdea15f5f..81791c124bb8 100644
@@ -2464,8 +2464,9 @@ compact_zone(struct compact_control *cc, struct capture_control *capc)
 			 * fast_find_migrateblock revisiting blocks that were
 			 * recently partially scanned.
 			 */
-			if (cc->direct_compaction && !cc->finish_pageblock &&
-						(cc->mode < MIGRATE_SYNC)) {
+			if (!pageblock_aligned(cc->migrate_pfn) &&
+			    cc->direct_compaction && !cc->finish_pageblock &&
+			    (cc->mode < MIGRATE_SYNC)) {
 				cc->finish_pageblock = true;
 
 				/*
Signature · 1 line
2.35.3
Acked-by on v1 1/4
#2 lore raw

Re: [PATCH 1/4] mm: compaction: Ensure rescanning only happens on partially scanned pageblocks

Quoted 8 lines · Mel Gorman

compact_zone() intends to rescan pageblocks if there is a failure to migrate "within the current order-aligned block". However, the pageblock scan may already be complete and moved to the next block causing the next pageblock to be "rescanned". Ensure only the most recent pageblock is rescanned.

Quoted 20 lines
---
 mm/compaction.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)
mm/compaction.c +3 -2
index c8bcdea15f5f..81791c124bb8 100644
@@ -2464,8 +2464,9 @@ compact_zone(struct compact_control *cc, struct capture_control *capc)
 			 * fast_find_migrateblock revisiting blocks that were
 			 * recently partially scanned.
 			 */
-			if (cc->direct_compaction && !cc->finish_pageblock &&
-						(cc->mode < MIGRATE_SYNC)) {
+			if (!pageblock_aligned(cc->migrate_pfn) &&
+			    cc->direct_compaction && !cc->finish_pageblock &&
+			    (cc->mode < MIGRATE_SYNC)) {
 				cc->finish_pageblock = true;
 
 				/*
author
#3 lore raw

PATCH 2/4 mm: compaction: Only force pageblock scan completion when skip hints are obeyed

In reply to Mel Gorman

fast_find_migrateblock relies on skip hints to avoid rescanning a recently selected pageblock but compact_zone() only forces the pageblock scan completion to set the skip hint if in direct compaction. While this prevents direct compaction repeatedly scanning a subset of blocks due to fast_find_migrateblock(), it does not prevent proactive compaction, node compaction and kcompactd encountering the same problem described in commit cfccd2e63e7e ("mm, compaction: finish pageblocks on complete migration failure").

Force the scan completion of a pageblock to set the skip hint if skip hints are obeyed to prevent fast_find_migrateblock() repeatedly selecting a subset of pageblocks.

---
 mm/compaction.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)
mm/compaction.c +3 -2
index 81791c124bb8..accc6568091a 100644
@@ -2456,7 +2456,8 @@ compact_zone(struct compact_control *cc, struct capture_control *capc)
 			}
 			/*
 			 * If an ASYNC or SYNC_LIGHT fails to migrate a page
-			 * within the current order-aligned block, scan the
+			 * within the current order-aligned block and
+			 * fast_find_migrateblock may be used then 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
@@ -2465,7 +2466,7 @@ compact_zone(struct compact_control *cc, struct capture_control *capc)
 			 * recently partially scanned.
 			 */
 			if (!pageblock_aligned(cc->migrate_pfn) &&
-			    cc->direct_compaction && !cc->finish_pageblock &&
+			    !cc->ignore_skip_hint && !cc->finish_pageblock &&
 			    (cc->mode < MIGRATE_SYNC)) {
 				cc->finish_pageblock = true;
 
Signature · 1 line
2.35.3
Acked-by on v1 2/4
#4 lore raw

Re: [PATCH 2/4] mm: compaction: Only force pageblock scan completion when skip hints are obeyed

Quoted 9 lines · Mel Gorman

fast_find_migrateblock relies on skip hints to avoid rescanning a recently selected pageblock but compact_zone() only forces the pageblock scan completion to set the skip hint if in direct compaction. While this prevents direct compaction repeatedly scanning a subset of blocks due to fast_find_migrateblock(), it does not prevent proactive compaction, node compaction and kcompactd encountering the same problem described in commit cfccd2e63e7e ("mm, compaction: finish pageblocks on complete migration failure").

Quoted 6 lines

Force the scan completion of a pageblock to set the skip hint if skip hints are obeyed to prevent fast_find_migrateblock() repeatedly selecting a subset of pageblocks.

Quoted 27 lines
---
 mm/compaction.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)
mm/compaction.c +3 -2
index 81791c124bb8..accc6568091a 100644
@@ -2456,7 +2456,8 @@ compact_zone(struct compact_control *cc, struct capture_control *capc)
 			}
 			/*
 			 * If an ASYNC or SYNC_LIGHT fails to migrate a page
-			 * within the current order-aligned block, scan the
+			 * within the current order-aligned block and
+			 * fast_find_migrateblock may be used then 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
@@ -2465,7 +2466,7 @@ compact_zone(struct compact_control *cc, struct capture_control *capc)
 			 * recently partially scanned.
 			 */
 			if (!pageblock_aligned(cc->migrate_pfn) &&
-			    cc->direct_compaction && !cc->finish_pageblock &&
+			    !cc->ignore_skip_hint && !cc->finish_pageblock &&
 			    (cc->mode < MIGRATE_SYNC)) {
 				cc->finish_pageblock = true;
 
author
#5 lore raw

PATCH 3/4 mm: compaction: Update pageblock skip when first migration candidate is not at the start

In reply to Mel Gorman

isolate_migratepages_block should mark a pageblock as skip if scanning started on an aligned pageblock boundary but it only updates the skip flag if the first migration candidate is also aligned. Tracing during a compaction stress load (mmtests: workload-usemem-stress-numa-compact) that many pageblocks are not marked skip causing excessive scanning of blocks that had been recently checked. Update pageblock skip based on "valid_page" which is set if scanning started on a pageblock boundary.

---
 mm/compaction.c | 15 +++++----------
 1 file changed, 5 insertions(+), 10 deletions(-)
mm/compaction.c +5 -10
index accc6568091a..d7be990b1d60 100644
@@ -392,18 +392,14 @@ void reset_isolation_suitable(pg_data_t *pgdat)
  * Sets the pageblock skip bit if it was clear. Note that this is a hint as
  * locks are not required for read/writers. Returns true if it was already set.
  */
-static bool test_and_set_skip(struct compact_control *cc, struct page *page,
-							unsigned long pfn)
+static bool test_and_set_skip(struct compact_control *cc, struct page *page)
 {
 	bool skip;
 
-	/* Do no update if skip hint is being ignored */
+	/* Do not update if skip hint is being ignored */
 	if (cc->ignore_skip_hint)
 		return false;
 
-	if (!pageblock_aligned(pfn))
-		return false;
-
 	skip = get_pageblock_skip(page);
 	if (!skip && !cc->no_set_skip_hint)
 		set_pageblock_skip(page);
@@ -470,8 +466,7 @@ static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
 {
 }
 
-static bool test_and_set_skip(struct compact_control *cc, struct page *page,
-							unsigned long pfn)
+static bool test_and_set_skip(struct compact_control *cc, struct page *page)
 {
 	return false;
 }
@@ -1075,9 +1070,9 @@ isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
 			lruvec_memcg_debug(lruvec, page_folio(page));
 
 			/* Try get exclusive access under lock */
-			if (!skip_updated) {
+			if (!skip_updated && valid_page) {
 				skip_updated = true;
-				if (test_and_set_skip(cc, page, low_pfn))
+				if (test_and_set_skip(cc, valid_page))
 					goto isolate_abort;
 			}
 
Signature · 1 line
2.35.3
#6 lore raw

Re: [PATCH 3/4] mm: compaction: Update pageblock skip when first migration candidate is not at the start

Quoted 9 lines · Mel Gorman

isolate_migratepages_block should mark a pageblock as skip if scanning started on an aligned pageblock boundary but it only updates the skip flag if the first migration candidate is also aligned. Tracing during a compaction stress load (mmtests: workload-usemem-stress-numa-compact) that many pageblocks are not marked skip causing excessive scanning of blocks that had been recently checked. Update pageblock skip based on "valid_page" which is set if scanning started on a pageblock boundary.

I wonder if this has an unintended side-effect that if we resume isolate_migratepages_block() of a partially compacted pageblock to finish it, test_and_set_skip() will now tell us to abort, because we already set the skip bit in the previous call. This would include the cc->finish_pageblock rescan cases.

So unless I miss something that already prevents that, I agree we should not tie setting the skip bit to pageblock_aligned(pfn), but maybe if we are not pageblock aligned, we should ignore the already-set skip bit, as it was most likely being set by us in the previous iteration and should not prevent us from finishing the pageblock?

Quoted 51 lines
---
 mm/compaction.c | 15 +++++----------
 1 file changed, 5 insertions(+), 10 deletions(-)
mm/compaction.c +5 -10
index accc6568091a..d7be990b1d60 100644
@@ -392,18 +392,14 @@ void reset_isolation_suitable(pg_data_t *pgdat)
  * Sets the pageblock skip bit if it was clear. Note that this is a hint as
  * locks are not required for read/writers. Returns true if it was already set.
  */
-static bool test_and_set_skip(struct compact_control *cc, struct page *page,
-							unsigned long pfn)
+static bool test_and_set_skip(struct compact_control *cc, struct page *page)
 {
 	bool skip;
 
-	/* Do no update if skip hint is being ignored */
+	/* Do not update if skip hint is being ignored */
 	if (cc->ignore_skip_hint)
 		return false;
 
-	if (!pageblock_aligned(pfn))
-		return false;
-
 	skip = get_pageblock_skip(page);
 	if (!skip && !cc->no_set_skip_hint)
 		set_pageblock_skip(page);
@@ -470,8 +466,7 @@ static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
 {
 }
 
-static bool test_and_set_skip(struct compact_control *cc, struct page *page,
-							unsigned long pfn)
+static bool test_and_set_skip(struct compact_control *cc, struct page *page)
 {
 	return false;
 }
@@ -1075,9 +1070,9 @@ isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
 			lruvec_memcg_debug(lruvec, page_folio(page));
 
 			/* Try get exclusive access under lock */
-			if (!skip_updated) {
+			if (!skip_updated && valid_page) {
 				skip_updated = true;
-				if (test_and_set_skip(cc, page, low_pfn))
+				if (test_and_set_skip(cc, valid_page))
 					goto isolate_abort;
 			}
 
author
#7 lore raw

Re: [PATCH 3/4] mm: compaction: Update pageblock skip when first migration candidate is not at the start

Quoted 17 lines · Vlastimil Babka
Quoted 9 lines · Mel Gorman

isolate_migratepages_block should mark a pageblock as skip if scanning started on an aligned pageblock boundary but it only updates the skip flag if the first migration candidate is also aligned. Tracing during a compaction stress load (mmtests: workload-usemem-stress-numa-compact) that many pageblocks are not marked skip causing excessive scanning of blocks that had been recently checked. Update pageblock skip based on "valid_page" which is set if scanning started on a pageblock boundary.

I wonder if this has an unintended side-effect that if we resume isolate_migratepages_block() of a partially compacted pageblock to finish it, test_and_set_skip() will now tell us to abort, because we already set the skip bit in the previous call. This would include the cc->finish_pageblock rescan cases.

Quoted 6 lines

So unless I miss something that already prevents that, I agree we should not tie setting the skip bit to pageblock_aligned(pfn), but maybe if we are not pageblock aligned, we should ignore the already-set skip bit, as it was most likely being set by us in the previous iteration and should not prevent us from finishing the pageblock?

Hmm, I think you're right. While it should not hit the original bug, migration candidates are missed until the next compaction scan which could be tricky to detect. Something like this as a separate patch? Build tested only but the intent is for an unaligned start to set the skip bet if already unset but otherwise complete the scan. Like earlier fixes, this might overscan some pageblocks in a given context but we are probably hitting the limits on how compaction can run efficiently in the current scheme without causing other side-effects :(

mm/compaction.c +9 -2
index 91af6a8b7a98..761a2dd7d78a 100644
@@ -792,6 +792,7 @@ isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
 	bool skip_on_failure = false;
 	unsigned long next_skip_pfn = 0;
 	bool skip_updated = false;
+	bool start_aligned;
 	int ret = 0;
 
 	cc->migrate_pfn = low_pfn;
@@ -824,6 +825,7 @@ isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
 	}
 
 	/* Time to isolate some pages for migration */
+	start_aligned = pageblock_aligned(start_pfn);
 	for (; low_pfn < end_pfn; low_pfn++) {
 
 		if (skip_on_failure && low_pfn >= next_skip_pfn) {
@@ -1069,10 +1071,15 @@ isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
 
 			lruvec_memcg_debug(lruvec, page_folio(page));
 
-			/* Try get exclusive access under lock */
+			/* Try get exclusive access under lock. Isolation is
+			 * only aborted if the start was pageblock aligned
+			 * as this may be a partial resumed scan that set
+			 * the bit on a recent scan but the scan must reach
+			 * the end of the pageblock.
+			 */
 			if (!skip_updated && valid_page) {
 				skip_updated = true;
-				if (test_and_set_skip(cc, valid_page))
+				if (test_and_set_skip(cc, valid_page) && start_aligned)
 					goto isolate_abort;
 			}
 
#8 lore raw

Re: [PATCH 3/4] mm: compaction: Update pageblock skip when first migration candidate is not at the start

Quoted 25 lines · Mel Gorman
Quoted 17 lines · Vlastimil Babka
Quoted 9 lines · Mel Gorman

isolate_migratepages_block should mark a pageblock as skip if scanning started on an aligned pageblock boundary but it only updates the skip flag if the first migration candidate is also aligned. Tracing during a compaction stress load (mmtests: workload-usemem-stress-numa-compact) that many pageblocks are not marked skip causing excessive scanning of blocks that had been recently checked. Update pageblock skip based on "valid_page" which is set if scanning started on a pageblock boundary.

I wonder if this has an unintended side-effect that if we resume isolate_migratepages_block() of a partially compacted pageblock to finish it, test_and_set_skip() will now tell us to abort, because we already set the skip bit in the previous call. This would include the cc->finish_pageblock rescan cases.

Quoted 6 lines

So unless I miss something that already prevents that, I agree we should not tie setting the skip bit to pageblock_aligned(pfn), but maybe if we are not pageblock aligned, we should ignore the already-set skip bit, as it was most likely being set by us in the previous iteration and should not prevent us from finishing the pageblock?

Quoted 8 lines

Hmm, I think you're right. While it should not hit the original bug, migration candidates are missed until the next compaction scan which could be tricky to detect. Something like this as a separate patch? Build tested only but the intent is for an unaligned start to set the skip bet if already unset but otherwise complete the scan. Like earlier fixes, this might overscan some pageblocks in a given context but we are probably hitting the limits on how compaction can run efficiently in the current scheme without causing other side-effects :(

Yeah that should work! I think it should be even folded to 3/4 but if you want separate, fine too.

Quoted 38 lines
mm/compaction.c +9 -2
index 91af6a8b7a98..761a2dd7d78a 100644
@@ -792,6 +792,7 @@ isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
 	bool skip_on_failure = false;
 	unsigned long next_skip_pfn = 0;
 	bool skip_updated = false;
+	bool start_aligned;
 	int ret = 0;
 
 	cc->migrate_pfn = low_pfn;
@@ -824,6 +825,7 @@ isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
 	}
 
 	/* Time to isolate some pages for migration */
+	start_aligned = pageblock_aligned(start_pfn);
 	for (; low_pfn < end_pfn; low_pfn++) {
 
 		if (skip_on_failure && low_pfn >= next_skip_pfn) {
@@ -1069,10 +1071,15 @@ isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
 
 			lruvec_memcg_debug(lruvec, page_folio(page));
 
-			/* Try get exclusive access under lock */
+			/* Try get exclusive access under lock. Isolation is
+			 * only aborted if the start was pageblock aligned
+			 * as this may be a partial resumed scan that set
+			 * the bit on a recent scan but the scan must reach
+			 * the end of the pageblock.
+			 */
 			if (!skip_updated && valid_page) {
 				skip_updated = true;
-				if (test_and_set_skip(cc, valid_page))
+				if (test_and_set_skip(cc, valid_page) && start_aligned)
 					goto isolate_abort;
 			}
 
author
#9 lore raw

Re: [PATCH 3/4] mm: compaction: Update pageblock skip when first migration candidate is not at the start

Quoted 35 lines · Vlastimil Babka
Quoted 25 lines · Mel Gorman
Quoted 17 lines · Vlastimil Babka
Quoted 9 lines · Mel Gorman

isolate_migratepages_block should mark a pageblock as skip if scanning started on an aligned pageblock boundary but it only updates the skip flag if the first migration candidate is also aligned. Tracing during a compaction stress load (mmtests: workload-usemem-stress-numa-compact) that many pageblocks are not marked skip causing excessive scanning of blocks that had been recently checked. Update pageblock skip based on "valid_page" which is set if scanning started on a pageblock boundary.

I wonder if this has an unintended side-effect that if we resume isolate_migratepages_block() of a partially compacted pageblock to finish it, test_and_set_skip() will now tell us to abort, because we already set the skip bit in the previous call. This would include the cc->finish_pageblock rescan cases.

Quoted 6 lines

So unless I miss something that already prevents that, I agree we should not tie setting the skip bit to pageblock_aligned(pfn), but maybe if we are not pageblock aligned, we should ignore the already-set skip bit, as it was most likely being set by us in the previous iteration and should not prevent us from finishing the pageblock?

Quoted 8 lines

Hmm, I think you're right. While it should not hit the original bug, migration candidates are missed until the next compaction scan which could be tricky to detect. Something like this as a separate patch? Build tested only but the intent is for an unaligned start to set the skip bet if already unset but otherwise complete the scan. Like earlier fixes, this might overscan some pageblocks in a given context but we are probably hitting the limits on how compaction can run efficiently in the current scheme without causing other side-effects :(

Quoted 3 lines

Yeah that should work! I think it should be even folded to 3/4 but if you want separate, fine too.

I was not happy with the test results so limited the scope of the patch which performed much better both in terms of absolute performance and compaction activity. Are you still ok with this version?

Thanks

--8<--
mm: compaction: Update pageblock skip when first migration candidate is not at the start -fix

Vlastimil Babka pointed out the following problem with mm-compaction-update-pageblock-skip-when-first-migration-candidate-is-not-at-the-start.patch

I wonder if this has an unintended side-effect that if we resume isolate_migratepages_block() of a partially compacted pageblock to finish it, test_and_set_skip() will now tell us to abort, because we already set the skip bit in the previous call. This would include the cc->finish_pageblock rescan cases.

He is correct and a partial rescan as implemented in "mm, compaction:
finish pageblocks on complete migration failure" would abort prematurely.

Test and set the skip bit when acquiring "exclusive access" to a pageblock for migration but only abort if the calling context is rescanning to finish a pageblock.

This is a fix for the mmotm patch mm-compaction-update-pageblock-skip-when-first-migration-candidate-is-not-at-the-start.patch

---
 mm/compaction.c | 10 ++++++++--
 1 file changed, 8 insertions(+), 2 deletions(-)
mm/compaction.c +8 -2
index 91af6a8b7a98..300aa968a0cf 100644
@@ -1069,11 +1069,17 @@ isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
 
 			lruvec_memcg_debug(lruvec, page_folio(page));
 
-			/* Try get exclusive access under lock */
+			/*
+			 * Try get exclusive access under lock. If marked for
+			 * skip, the scan is aborted unless the current context
+			 * is a rescan to reach the end of the pageblock.
+			 */
 			if (!skip_updated && valid_page) {
 				skip_updated = true;
-				if (test_and_set_skip(cc, valid_page))
+				if (test_and_set_skip(cc, valid_page) &&
+				    !cc->finish_pageblock) {
 					goto isolate_abort;
+				}
 			}
 
 			/*
#10 lore raw

Re: [PATCH 3/4] mm: compaction: Update pageblock skip when first migration candidate is not at the start

Quoted 36 lines · Mel Gorman
Quoted 35 lines · Vlastimil Babka
Quoted 33 lines · Mel Gorman
Quoted 17 lines · Vlastimil Babka
Quoted 9 lines · Mel Gorman

isolate_migratepages_block should mark a pageblock as skip if scanning started on an aligned pageblock boundary but it only updates the skip flag if the first migration candidate is also aligned. Tracing during a compaction stress load (mmtests: workload-usemem-stress-numa-compact) that many pageblocks are not marked skip causing excessive scanning of blocks that had been recently checked. Update pageblock skip based on "valid_page" which is set if scanning started on a pageblock boundary.

I wonder if this has an unintended side-effect that if we resume isolate_migratepages_block() of a partially compacted pageblock to finish it, test_and_set_skip() will now tell us to abort, because we already set the skip bit in the previous call. This would include the cc->finish_pageblock rescan cases.

Quoted 6 lines

So unless I miss something that already prevents that, I agree we should not tie setting the skip bit to pageblock_aligned(pfn), but maybe if we are not pageblock aligned, we should ignore the already-set skip bit, as it was most likely being set by us in the previous iteration and should not prevent us from finishing the pageblock?

Hmm, I think you're right. While it should not hit the original bug, migration candidates are missed until the next compaction scan which could be tricky to detect. Something like this as a separate patch? Build tested only but the intent is for an unaligned start to set the skip bet if already unset but otherwise complete the scan. Like earlier fixes, this might overscan some pageblocks in a given context but we are probably hitting the limits on how compaction can run efficiently in the current scheme without causing other side-effects :(

Quoted 7 lines
Quoted 3 lines

Yeah that should work! I think it should be even folded to 3/4 but if you want separate, fine too.

I was not happy with the test results so limited the scope of the patch which performed much better both in terms of absolute performance and compaction activity.

That's surprising. Does that mean that if we isolate COMPACT_CLUSTER_MAX pages, migrate them without failing, but it's not enough to succeed (i.e. there are more pages we need to migrate to free up a whole pageblock), it's better to give up on the rest of the pageblock rather than continue? As that's AFAIU the scenario where cc->finish_pageblock is false when we revisit an unfinished pageblock.

Are you still ok with this version?

It's better than previously in that cc->finish_pageblock == true case is not sabotaged anymore. But the result as described above seems to be a weird non-intuitive and non-obvious heuristic. How did the test differences look like?

Quoted 14 lines

Thanks

--8<--
mm: compaction: Update pageblock skip when first migration candidate is not at the start -fix

Vlastimil Babka pointed out the following problem with mm-compaction-update-pageblock-skip-when-first-migration-candidate-is-not-at-the-start.patch

I wonder if this has an unintended side-effect that if we resume isolate_migratepages_block() of a partially compacted pageblock to finish it, test_and_set_skip() will now tell us to abort, because we already set the skip bit in the previous call. This would include the cc->finish_pageblock rescan cases.

Quoted 7 lines

He is correct and a partial rescan as implemented in "mm, compaction:
finish pageblocks on complete migration failure" would abort prematurely.

Test and set the skip bit when acquiring "exclusive access" to a pageblock for migration but only abort if the calling context is rescanning to finish a pageblock.

Should it say NOT rescanning to finish a pageblock?

Quoted 32 lines

This is a fix for the mmotm patch mm-compaction-update-pageblock-skip-when-first-migration-candidate-is-not-at-the-start.patch

---
 mm/compaction.c | 10 ++++++++--
 1 file changed, 8 insertions(+), 2 deletions(-)
mm/compaction.c +8 -2
index 91af6a8b7a98..300aa968a0cf 100644
@@ -1069,11 +1069,17 @@ isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
 
 			lruvec_memcg_debug(lruvec, page_folio(page));
 
-			/* Try get exclusive access under lock */
+			/*
+			 * Try get exclusive access under lock. If marked for
+			 * skip, the scan is aborted unless the current context
+			 * is a rescan to reach the end of the pageblock.
+			 */
 			if (!skip_updated && valid_page) {
 				skip_updated = true;
-				if (test_and_set_skip(cc, valid_page))
+				if (test_and_set_skip(cc, valid_page) &&
+				    !cc->finish_pageblock) {
 					goto isolate_abort;
+				}
 			}
 
 			/*
author
#11 lore raw

Re: [PATCH 3/4] mm: compaction: Update pageblock skip when first migration candidate is not at the start

Quoted 41 lines · Vlastimil Babka
Quoted 40 lines · Mel Gorman
Quoted 38 lines · Vlastimil Babka
Quoted 25 lines · Mel Gorman
Quoted 17 lines · Vlastimil Babka
Quoted 9 lines · Mel Gorman

isolate_migratepages_block should mark a pageblock as skip if scanning started on an aligned pageblock boundary but it only updates the skip flag if the first migration candidate is also aligned. Tracing during a compaction stress load (mmtests: workload-usemem-stress-numa-compact) that many pageblocks are not marked skip causing excessive scanning of blocks that had been recently checked. Update pageblock skip based on "valid_page" which is set if scanning started on a pageblock boundary.

I wonder if this has an unintended side-effect that if we resume isolate_migratepages_block() of a partially compacted pageblock to finish it, test_and_set_skip() will now tell us to abort, because we already set the skip bit in the previous call. This would include the cc->finish_pageblock rescan cases.

Quoted 6 lines

So unless I miss something that already prevents that, I agree we should not tie setting the skip bit to pageblock_aligned(pfn), but maybe if we are not pageblock aligned, we should ignore the already-set skip bit, as it was most likely being set by us in the previous iteration and should not prevent us from finishing the pageblock?

Quoted 8 lines

Hmm, I think you're right. While it should not hit the original bug, migration candidates are missed until the next compaction scan which could be tricky to detect. Something like this as a separate patch? Build tested only but the intent is for an unaligned start to set the skip bet if already unset but otherwise complete the scan. Like earlier fixes, this might overscan some pageblocks in a given context but we are probably hitting the limits on how compaction can run efficiently in the current scheme without causing other side-effects :(

Yeah that should work! I think it should be even folded to 3/4 but if you want separate, fine too.

Quoted 8 lines
Quoted 3 lines

I was not happy with the test results so limited the scope of the patch which performed much better both in terms of absolute performance and compaction activity.

That's surprising. Does that mean that if we isolate COMPACT_CLUSTER_MAX pages, migrate them without failing, but it's not enough to succeed (i.e. there are more pages we need to migrate to free up a whole pageblock), it's better to give up on the rest of the pageblock rather than continue?

I don't have precise enough data to answer that with certainty but probably yes, at least in terms of scan activity. The first version had spikes of pages scanned for migration that are not always reproducible and not on all machines.

Quoted 10 lines

As
that's AFAIU the scenario where cc->finish_pageblock is false when we revisit an unfinished pageblock.

Quoted 1 line

Are you still ok with this version?

It's better than previously in that cc->finish_pageblock == true case is not sabotaged anymore. But the result as described above seems to be a weird non-intuitive and non-obvious heuristic. How did the test differences look like?

5 machines were tested in all (different Intel and AMD generations), 2 showed unexpected spikes in scan activity. 1 showed high migration scan counts for workloads workload-thpchallenge-kernel-build-xfs and workload-usemem-stress-numa-compact. They are both basically compaction stressors with the first one using a kernel build
in the background to generate noise. For the second machine, only workload-usemem-stress-numa-compact was affected. In all other test cases and machines, the patches were equivalent in terms of behaviour but the extreme counter-examples led be to conclude the fix should be as constrained as possible unless there is a good reason to do otherwise.

Quoted 13 lines
Quoted 13 lines

Vlastimil Babka pointed out the following problem with mm-compaction-update-pageblock-skip-when-first-migration-candidate-is-not-at-the-start.patch

I wonder if this has an unintended side-effect that if we resume isolate_migratepages_block() of a partially compacted pageblock to finish it, test_and_set_skip() will now tell us to abort, because we already set the skip bit in the previous call. This would include the cc->finish_pageblock rescan cases.

He is correct and a partial rescan as implemented in "mm, compaction:
finish pageblocks on complete migration failure" would abort prematurely.

Quoted 6 lines
Quoted 3 lines

Test and set the skip bit when acquiring "exclusive access" to a pageblock for migration but only abort if the calling context is rescanning to finish a pageblock.

Should it say NOT rescanning to finish a pageblock?

Yep, it should. The sentence was a last minute update before pushing send :(

Signature · 2 lines
Mel Gorman
SUSE Labs
#12 lore raw

Re: [PATCH 3/4] mm: compaction: Update pageblock skip when first migration candidate is not at the start

Quoted 51 lines · Mel Gorman
Quoted 41 lines · Vlastimil Babka
Quoted 40 lines · Mel Gorman
Quoted 38 lines · Vlastimil Babka
Quoted 25 lines · Mel Gorman
Quoted 17 lines · Vlastimil Babka
Quoted 9 lines · Mel Gorman

isolate_migratepages_block should mark a pageblock as skip if scanning started on an aligned pageblock boundary but it only updates the skip flag if the first migration candidate is also aligned. Tracing during a compaction stress load (mmtests: workload-usemem-stress-numa-compact) that many pageblocks are not marked skip causing excessive scanning of blocks that had been recently checked. Update pageblock skip based on "valid_page" which is set if scanning started on a pageblock boundary.

I wonder if this has an unintended side-effect that if we resume isolate_migratepages_block() of a partially compacted pageblock to finish it, test_and_set_skip() will now tell us to abort, because we already set the skip bit in the previous call. This would include the cc->finish_pageblock rescan cases.

Quoted 6 lines

So unless I miss something that already prevents that, I agree we should not tie setting the skip bit to pageblock_aligned(pfn), but maybe if we are not pageblock aligned, we should ignore the already-set skip bit, as it was most likely being set by us in the previous iteration and should not prevent us from finishing the pageblock?

Quoted 8 lines

Hmm, I think you're right. While it should not hit the original bug, migration candidates are missed until the next compaction scan which could be tricky to detect. Something like this as a separate patch? Build tested only but the intent is for an unaligned start to set the skip bet if already unset but otherwise complete the scan. Like earlier fixes, this might overscan some pageblocks in a given context but we are probably hitting the limits on how compaction can run efficiently in the current scheme without causing other side-effects :(

Yeah that should work! I think it should be even folded to 3/4 but if you want separate, fine too.

Quoted 8 lines
Quoted 3 lines

I was not happy with the test results so limited the scope of the patch which performed much better both in terms of absolute performance and compaction activity.

That's surprising. Does that mean that if we isolate COMPACT_CLUSTER_MAX pages, migrate them without failing, but it's not enough to succeed (i.e. there are more pages we need to migrate to free up a whole pageblock), it's better to give up on the rest of the pageblock rather than continue?

Quoted 4 lines

I don't have precise enough data to answer that with certainty but probably yes, at least in terms of scan activity. The first version had spikes of pages scanned for migration that are not always reproducible and not on all machines.

Well, that kinda sucks. So for the patch (with adding the missing NOT below).

But in raises a question whether we should terminate compaction under the right conditions after a successful migration immediately, rather than invoke another iteration of isolate_migratepages_block() where we could skip over some pages uselessly only to abort at first valid page due to the skip bit. It would save some cycles and be much more obvious than now, where anyone trying to understand how it works in detail might conclude it's an oversight?

Quoted 45 lines
Quoted 10 lines

As
that's AFAIU the scenario where cc->finish_pageblock is false when we revisit an unfinished pageblock.

Quoted 1 line

Are you still ok with this version?

It's better than previously in that cc->finish_pageblock == true case is not sabotaged anymore. But the result as described above seems to be a weird non-intuitive and non-obvious heuristic. How did the test differences look like?

5 machines were tested in all (different Intel and AMD generations), 2 showed unexpected spikes in scan activity. 1 showed high migration scan counts for workloads workload-thpchallenge-kernel-build-xfs and workload-usemem-stress-numa-compact. They are both basically compaction stressors with the first one using a kernel build
in the background to generate noise. For the second machine, only workload-usemem-stress-numa-compact was affected. In all other test cases and machines, the patches were equivalent in terms of behaviour but the extreme counter-examples led be to conclude the fix should be as constrained as possible unless there is a good reason to do otherwise.

Quoted 13 lines
Quoted 13 lines

Vlastimil Babka pointed out the following problem with mm-compaction-update-pageblock-skip-when-first-migration-candidate-is-not-at-the-start.patch

I wonder if this has an unintended side-effect that if we resume isolate_migratepages_block() of a partially compacted pageblock to finish it, test_and_set_skip() will now tell us to abort, because we already set the skip bit in the previous call. This would include the cc->finish_pageblock rescan cases.

He is correct and a partial rescan as implemented in "mm, compaction:
finish pageblocks on complete migration failure" would abort prematurely.

Quoted 6 lines
Quoted 3 lines

Test and set the skip bit when acquiring "exclusive access" to a pageblock for migration but only abort if the calling context is rescanning to finish a pageblock.

Should it say NOT rescanning to finish a pageblock?

Yep, it should. The sentence was a last minute update before pushing send :(

#13 lore raw

Re: [PATCH 3/4] mm: compaction: Update pageblock skip when first migration candidate is not at the start

Quoted 59 lines · Vlastimil Babka
Quoted 51 lines · Mel Gorman
Quoted 41 lines · Vlastimil Babka
Quoted 40 lines · Mel Gorman
Quoted 38 lines · Vlastimil Babka
Quoted 25 lines · Mel Gorman
Quoted 17 lines · Vlastimil Babka
Quoted 9 lines · Mel Gorman

isolate_migratepages_block should mark a pageblock as skip if scanning started on an aligned pageblock boundary but it only updates the skip flag if the first migration candidate is also aligned. Tracing during a compaction stress load (mmtests: workload-usemem-stress-numa-compact) that many pageblocks are not marked skip causing excessive scanning of blocks that had been recently checked. Update pageblock skip based on "valid_page" which is set if scanning started on a pageblock boundary.

I wonder if this has an unintended side-effect that if we resume isolate_migratepages_block() of a partially compacted pageblock to finish it, test_and_set_skip() will now tell us to abort, because we already set the skip bit in the previous call. This would include the cc->finish_pageblock rescan cases.

Quoted 6 lines

So unless I miss something that already prevents that, I agree we should not tie setting the skip bit to pageblock_aligned(pfn), but maybe if we are not pageblock aligned, we should ignore the already-set skip bit, as it was most likely being set by us in the previous iteration and should not prevent us from finishing the pageblock?

Quoted 8 lines

Hmm, I think you're right. While it should not hit the original bug, migration candidates are missed until the next compaction scan which could be tricky to detect. Something like this as a separate patch? Build tested only but the intent is for an unaligned start to set the skip bet if already unset but otherwise complete the scan. Like earlier fixes, this might overscan some pageblocks in a given context but we are probably hitting the limits on how compaction can run efficiently in the current scheme without causing other side-effects :(

Yeah that should work! I think it should be even folded to 3/4 but if you want separate, fine too.

Quoted 8 lines
Quoted 3 lines

I was not happy with the test results so limited the scope of the patch which performed much better both in terms of absolute performance and compaction activity.

That's surprising. Does that mean that if we isolate COMPACT_CLUSTER_MAX pages, migrate them without failing, but it's not enough to succeed (i.e. there are more pages we need to migrate to free up a whole pageblock), it's better to give up on the rest of the pageblock rather than continue?

Quoted 4 lines

I don't have precise enough data to answer that with certainty but probably yes, at least in terms of scan activity. The first version had spikes of pages scanned for migration that are not always reproducible and not on all machines.

Well, that kinda sucks. So for the patch (with adding the missing NOT below).

Quoted 8 lines

But in raises a question whether we should terminate compaction under the right conditions after a successful migration immediately, rather than invoke another iteration of isolate_migratepages_block() where we could skip over some pages uselessly only to abort at first valid page due to the skip bit. It would save some cycles and be much more obvious than now, where anyone trying to understand how it works in detail might conclude it's an oversight?

IIUC, for direct compactor, we can terminate compaction if there is a suitable free page, that means other pages scanning in this pageblock are uselessly like you said, and we can skip them to save cpu cycles. So we can remove below COMPACT_CONTINUE validation in __compact_finished():

	/*
	 * Always finish scanning a pageblock to reduce the possibility of
	 * fallbacks in the future. This is particularly important when
	 * migration source is unmovable/reclaimable but it's not worth
	 * special casing.
	 */
	if (!pageblock_aligned(cc->migrate_pfn))
		return COMPACT_CONTINUE;

And in this case, we also should call update_cached_migrate() to record the partial cc->migrate_pfn, which is the start position for next scanning. Please correct me if I missed something. Thanks.

Quoted 46 lines
Quoted 45 lines
Quoted 10 lines

As
that's AFAIU the scenario where cc->finish_pageblock is false when we revisit an unfinished pageblock.

Quoted 1 line

Are you still ok with this version?

It's better than previously in that cc->finish_pageblock == true case is not sabotaged anymore. But the result as described above seems to be a weird non-intuitive and non-obvious heuristic. How did the test differences look like?

5 machines were tested in all (different Intel and AMD generations), 2 showed unexpected spikes in scan activity. 1 showed high migration scan counts for workloads workload-thpchallenge-kernel-build-xfs and workload-usemem-stress-numa-compact. They are both basically compaction stressors with the first one using a kernel build
in the background to generate noise. For the second machine, only workload-usemem-stress-numa-compact was affected. In all other test cases and machines, the patches were equivalent in terms of behaviour but the extreme counter-examples led be to conclude the fix should be as constrained as possible unless there is a good reason to do otherwise.

Quoted 13 lines
Quoted 13 lines

Vlastimil Babka pointed out the following problem with mm-compaction-update-pageblock-skip-when-first-migration-candidate-is-not-at-the-start.patch

I wonder if this has an unintended side-effect that if we resume isolate_migratepages_block() of a partially compacted pageblock to finish it, test_and_set_skip() will now tell us to abort, because we already set the skip bit in the previous call. This would include the cc->finish_pageblock rescan cases.

He is correct and a partial rescan as implemented in "mm, compaction:
finish pageblocks on complete migration failure" would abort prematurely.

Quoted 6 lines
Quoted 3 lines

Test and set the skip bit when acquiring "exclusive access" to a pageblock for migration but only abort if the calling context is rescanning to finish a pageblock.

Should it say NOT rescanning to finish a pageblock?

Yep, it should. The sentence was a last minute update before pushing send :(

author
#14 lore raw

Re: [PATCH 3/4] mm: compaction: Update pageblock skip when first migration candidate is not at the start

In reply to Vlastimil Babka
Quoted 59 lines · Vlastimil Babka
Quoted 51 lines · Mel Gorman
Quoted 41 lines · Vlastimil Babka
Quoted 40 lines · Mel Gorman
Quoted 38 lines · Vlastimil Babka
Quoted 25 lines · Mel Gorman
Quoted 17 lines · Vlastimil Babka
Quoted 9 lines · Mel Gorman

isolate_migratepages_block should mark a pageblock as skip if scanning started on an aligned pageblock boundary but it only updates the skip flag if the first migration candidate is also aligned. Tracing during a compaction stress load (mmtests: workload-usemem-stress-numa-compact) that many pageblocks are not marked skip causing excessive scanning of blocks that had been recently checked. Update pageblock skip based on "valid_page" which is set if scanning started on a pageblock boundary.

I wonder if this has an unintended side-effect that if we resume isolate_migratepages_block() of a partially compacted pageblock to finish it, test_and_set_skip() will now tell us to abort, because we already set the skip bit in the previous call. This would include the cc->finish_pageblock rescan cases.

Quoted 6 lines

So unless I miss something that already prevents that, I agree we should not tie setting the skip bit to pageblock_aligned(pfn), but maybe if we are not pageblock aligned, we should ignore the already-set skip bit, as it was most likely being set by us in the previous iteration and should not prevent us from finishing the pageblock?

Quoted 8 lines

Hmm, I think you're right. While it should not hit the original bug, migration candidates are missed until the next compaction scan which could be tricky to detect. Something like this as a separate patch? Build tested only but the intent is for an unaligned start to set the skip bet if already unset but otherwise complete the scan. Like earlier fixes, this might overscan some pageblocks in a given context but we are probably hitting the limits on how compaction can run efficiently in the current scheme without causing other side-effects :(

Yeah that should work! I think it should be even folded to 3/4 but if you want separate, fine too.

Quoted 8 lines
Quoted 3 lines

I was not happy with the test results so limited the scope of the patch which performed much better both in terms of absolute performance and compaction activity.

That's surprising. Does that mean that if we isolate COMPACT_CLUSTER_MAX pages, migrate them without failing, but it's not enough to succeed (i.e. there are more pages we need to migrate to free up a whole pageblock), it's better to give up on the rest of the pageblock rather than continue?

Quoted 4 lines

I don't have precise enough data to answer that with certainty but probably yes, at least in terms of scan activity. The first version had spikes of pages scanned for migration that are not always reproducible and not on all machines.

Well, that kinda sucks. So for the patch (with adding the missing NOT below).

Quoted 9 lines

But in raises a question whether we should terminate compaction under the right conditions after a successful migration immediately, rather than invoke another iteration of isolate_migratepages_block() where we could skip over some pages uselessly only to abort at first valid page due to the skip bit. It would save some cycles and be much more obvious than now, where anyone trying to understand how it works in detail might conclude it's an oversight?

It sounds like a solid idea and would be a good standalone patch with the usual supporting data. At a quick glance, the check for a page stored on compact_control happens surprisingly late which makes me think we probably over-compact in direct compaction in particular. It would need supporting data because it probably means that compaction cost gets spread over multiple tasks requiring high-order pages instead of one unlucky task doing compaction works that unrelated tasks indirectly benefit from. It's probably more sensible behaviour that tasks requiring high-order pages pay the cost if kcompactd cannot keep up but supporting data would tell us one way or the other.

Signature · 2 lines
Mel Gorman
SUSE Labs
author
#15 lore raw

PATCH 4/4 Revert "Revert "mm/compaction: fix set skip in fast_find_migrateblock""

In reply to Mel Gorman

This reverts commit 95e7a450b819 ("Revert "mm/compaction: fix set skip in fast_find_migrateblock"").

Commit 7efc3b726103 ("mm/compaction: fix set skip in fast_find_migrateblock") was reverted due to bug reports about khugepaged consuming large amounts of CPU without making progress. The underlying bug was partially fixed by commit cfccd2e63e7e ("mm, compaction: finish pageblocks on complete migration failure") but it only mitigated the problem and Vlastimil Babka pointing out the same issue could theoretically happen to kcompactd.

As pageblocks containing pages that fail to migrate should now be forcibly rescanned to set the skip hint if skip hints are used, fast_find_migrateblock() should no longer loop on a small subset of pageblocks for prolonged periods of time. Revert the revert so fast_find_migrateblock() is effective again.

Using the mmtests config workload-usemem-stress-numa-compact, the number of unique ranges scanned was analysed for both kcompactd and !kcompactd activity.

6.4.0-rc1-vanilla
kcompactd

      7 range=(0x10d600~0x10d800)
      7 range=(0x110c00~0x110e00)
      7 range=(0x110e00~0x111000)
      7 range=(0x111800~0x111a00)
      7 range=(0x111a00~0x111c00)
!kcompactd
      1 range=(0x113e00~0x114000)
      1 range=(0x114000~0x114020)
      1 range=(0x114400~0x114489)
      1 range=(0x114489~0x1144aa)
      1 range=(0x1144aa~0x114600)

6.4.0-rc1-mm-revertfastmigrate
kcompactd

     17 range=(0x104200~0x104400)
     17 range=(0x104400~0x104600)
     17 range=(0x104600~0x104800)
     17 range=(0x104800~0x104a00)
     17 range=(0x104a00~0x104c00)

!kcompactd
1793 range=(0x15c200~0x15c400)
5436 range=(0x105800~0x105a00)
19826 range=(0x150a00~0x150c00)
19833 range=(0x150800~0x150a00)
19834 range=(0x11ce00~0x11d000)

6.4.0-rc1-mm-follupfastfind
kcompactd

     22 range=(0x107200~0x107400)
     23 range=(0x107400~0x107600)
     23 range=(0x107600~0x107800)
     23 range=(0x107c00~0x107e00)
     23 range=(0x107e00~0x108000)
!kcompactd
      3 range=(0x890240~0x890400)
      5 range=(0x886e00~0x887000)
      5 range=(0x88a400~0x88a600)
      6 range=(0x88f800~0x88fa00)
      9 range=(0x88a400~0x88a420)

Note that the vanilla kernel and the full series had some duplication of ranges scanned but it was not severe and would be in line with compaction resets when the skip hints are cleared. Just a revert of commit 7efc3b726103 ("mm/compaction: fix set skip in fast_find_migrateblock") showed excessive rescans of the same ranges so the series should not reintroduce bug 1206848.

---
 mm/compaction.c | 1 -
 1 file changed, 1 deletion(-)
mm/compaction.c +0 -1
index d7be990b1d60..91af6a8b7a98 100644
@@ -1859,7 +1859,6 @@ static unsigned long fast_find_migrateblock(struct compact_control *cc)
 					pfn = cc->zone->zone_start_pfn;
 				cc->fast_search_fail = 0;
 				found_block = true;
-				set_pageblock_skip(freepage);
 				break;
 			}
 		}
Signature · 1 line
2.35.3
Acked-by on v1 4/4
#16 lore raw

Re: [PATCH 4/4] Revert "Revert "mm/compaction: fix set skip in fast_find_migrateblock""

Quoted 62 lines · Mel Gorman

This reverts commit 95e7a450b819 ("Revert "mm/compaction: fix set skip in fast_find_migrateblock"").

Commit 7efc3b726103 ("mm/compaction: fix set skip in fast_find_migrateblock") was reverted due to bug reports about khugepaged consuming large amounts of CPU without making progress. The underlying bug was partially fixed by commit cfccd2e63e7e ("mm, compaction: finish pageblocks on complete migration failure") but it only mitigated the problem and Vlastimil Babka pointing out the same issue could theoretically happen to kcompactd.

As pageblocks containing pages that fail to migrate should now be forcibly rescanned to set the skip hint if skip hints are used, fast_find_migrateblock() should no longer loop on a small subset of pageblocks for prolonged periods of time. Revert the revert so fast_find_migrateblock() is effective again.

Using the mmtests config workload-usemem-stress-numa-compact, the number of unique ranges scanned was analysed for both kcompactd and !kcompactd activity.

6.4.0-rc1-vanilla
kcompactd

      7 range=(0x10d600~0x10d800)
      7 range=(0x110c00~0x110e00)
      7 range=(0x110e00~0x111000)
      7 range=(0x111800~0x111a00)
      7 range=(0x111a00~0x111c00)
!kcompactd
      1 range=(0x113e00~0x114000)
      1 range=(0x114000~0x114020)
      1 range=(0x114400~0x114489)
      1 range=(0x114489~0x1144aa)
      1 range=(0x1144aa~0x114600)

6.4.0-rc1-mm-revertfastmigrate
kcompactd

     17 range=(0x104200~0x104400)
     17 range=(0x104400~0x104600)
     17 range=(0x104600~0x104800)
     17 range=(0x104800~0x104a00)
     17 range=(0x104a00~0x104c00)

!kcompactd
1793 range=(0x15c200~0x15c400)
5436 range=(0x105800~0x105a00)
19826 range=(0x150a00~0x150c00)
19833 range=(0x150800~0x150a00)
19834 range=(0x11ce00~0x11d000)

6.4.0-rc1-mm-follupfastfind
kcompactd

     22 range=(0x107200~0x107400)
     23 range=(0x107400~0x107600)
     23 range=(0x107600~0x107800)
     23 range=(0x107c00~0x107e00)
     23 range=(0x107e00~0x108000)
!kcompactd
      3 range=(0x890240~0x890400)
      5 range=(0x886e00~0x887000)
      5 range=(0x88a400~0x88a600)
      6 range=(0x88f800~0x88fa00)
      9 range=(0x88a400~0x88a420)
Quoted 8 lines

Note that the vanilla kernel and the full series had some duplication of ranges scanned but it was not severe and would be in line with compaction resets when the skip hints are cleared. Just a revert of commit 7efc3b726103 ("mm/compaction: fix set skip in fast_find_migrateblock") showed excessive rescans of the same ranges so the series should not reintroduce bug 1206848.

My concerns about patch 3/4 don't affect this part so

Quoted 16 lines
---
 mm/compaction.c | 1 -
 1 file changed, 1 deletion(-)
mm/compaction.c +0 -1
index d7be990b1d60..91af6a8b7a98 100644
@@ -1859,7 +1859,6 @@ static unsigned long fast_find_migrateblock(struct compact_control *cc)
 					pfn = cc->zone->zone_start_pfn;
 				cc->fast_search_fail = 0;
 				found_block = true;
-				set_pageblock_skip(freepage);
 				break;
 			}
 		}
#17 lore raw
In reply to Mel Gorman
Quoted 8 lines · Mel Gorman

The series "Fix excessive CPU usage during compaction" [1] attempted to fix a bug [2] but Vlastimil noted that the fix was incomplete [3]. While the series was merged, fast_find_migrateblock was still disabled. This series should fix the corner cases and allow 95e7a450b819 ("Revert "mm/compaction: fix set skip in fast_find_migrateblock"") to be safely reverted. Details on how many pageblocks are rescanned are in the changelog of the last patch.

Quoted 7 lines

[1] https://lore.kernel.org/r/[email protected]
[2] https://bugzilla.suse.com/show_bug.cgi?id=1206848
[3] https://lore.kernel.org/r/[email protected]

  mm/compaction.c | 24 ++++++++++--------------
  1 file changed, 10 insertions(+), 14 deletions(-)

Hello Mel,

Not sure how much this info would help, (also I saw it is in Andrew's tree already) But was curious to evaluate the patchset from perf perspective, and I have run mmtest usemem so here is the result (Only for compact cases).

kernel configuration:
1) base 6.4.0-rc2
2) patchseries
3) base - revert of 7efc3b726103

Summary,
The result shows decent improvement from perf perspective as well as compaction related data.

$ cat kcompact_fix_result_thp_madv

SUT : 256 cpu two node milan

usemem

                      6.4.0-rc2              6.4.0-rc2-compactfix 
6.4.0-rc2-revert
Amean     syst-1      102.18 (   0.00%)       98.98 *   3.13%* 96.96 *   5.10%*
Amean     elsp-1      220.57 (   0.00%)      215.96 *   2.09%* 212.49 *   3.67%*
Amean     syst-3      126.34 (   0.00%)      122.39 *   3.13%* 124.09 *   1.78%*
Amean     elsp-3       84.72 (   0.00%)       81.91 *   3.31%* 82.46 *   2.66%*
Amean     syst-4      142.64 (   0.00%)      130.69 *   8.38%* 131.19 *   8.03%*
Amean     elsp-4       79.48 (   0.00%)       66.60 *  16.21%* 66.37 *  16.49%*
                     6.4.0-rc2  6.4.0-rc2-compactfix    6.4.0-rc2-revert
Duration User        2106.24             2065.61        2047.87
Duration System      2598.68             2464.98        2466.33
Duration Elapsed     2693.98             2551.89        2529.83
                                6.4.0-rc2 
6.4.0-rc2-compactfix            6.4.0-rc2-revert
Ops Minor Faults                  2440289187.00          2356626072.00 
        2354663705.00
Ops Sector Reads                         556.00                1028.00 
              2400.00
Ops Sector Writes                      20388.00               15556.00 
             16744.00
Ops Page migrate success            94642602.00            40234601.00 
          38486059.00
Ops Page migrate failure                2161.00                 467.00 
               725.00
Ops Compaction pages isolated      116977416.00            80467188.00 
          76966075.00
Ops Compaction migrate scanned     136909038.00            60060978.00 
          60479389.00
Ops Compaction free scanned        165907615.00            95661742.00 
          89794809.00
Ops Compact scan efficiency               82.52                  62.78 
                67.35
Ops Compaction cost                   101419.99               43712.83 
             41834.25
Ops Kcompactd wake                         0.00                   0.00 
                 0.00
Ops Kcompactd migrate scanned      136909038.00            60060978.00 
          60479389.00
Ops Kcompactd free scanned         165907615.00            95661742.00 
          89794809.00
Ops NUMA alloc hit                1834950351.00          1853285684.00 
        1867258209.00
Ops NUMA alloc miss                413982625.00           359494012.00 
         345537623.00
Ops NUMA alloc local              1834950349.00          1853285649.00 
        1867258632.00
Ops NUMA base-page range updates   202326681.00           149242271.00 
         147460952.00
Ops NUMA PTE updates               202326681.00           149242271.00 
         147460952.00
Ops NUMA hint faults               195595665.00           148152760.00 
         146173618.00
Ops NUMA hint local faults %       159439097.00           148150869.00 
         146169658.00
Ops NUMA hint local percent               81.51                 100.00 
               100.00
Ops NUMA pages migrated             36155998.00                1454.00 
              3596.00
Ops AutoNUMA cost                     980081.58              741808.52 
            731900.38

Thanks

author
#18 lore raw
Quoted 17 lines · Raghavendra K T
Quoted 8 lines · Mel Gorman

The series "Fix excessive CPU usage during compaction" [1] attempted to fix a bug [2] but Vlastimil noted that the fix was incomplete [3]. While the series was merged, fast_find_migrateblock was still disabled. This series should fix the corner cases and allow 95e7a450b819 ("Revert "mm/compaction: fix set skip in fast_find_migrateblock"") to be safely reverted. Details on how many pageblocks are rescanned are in the changelog of the last patch.

Quoted 7 lines

[1] https://lore.kernel.org/r/[email protected]
[2] https://bugzilla.suse.com/show_bug.cgi?id=1206848
[3] https://lore.kernel.org/r/[email protected]

  mm/compaction.c | 24 ++++++++++--------------
  1 file changed, 10 insertions(+), 14 deletions(-)
Quoted 9 lines

Hello Mel,

Not sure how much this info would help, (also I saw it is in Andrew's tree already) But was curious to evaluate the patchset from perf perspective, and I have run mmtest usemem so here is the result (Only for compact cases).

<SNIP>

Thanks Raghavendra!

Signature · 2 lines
Mel Gorman
SUSE Labs
Tested-by on the v1 series
#19 lore raw
In reply to Mel Gorman

Hi Mel,

Quoted 11 lines · Mel Gorman

The series "Fix excessive CPU usage during compaction" [1] attempted to fix a bug [2] but Vlastimil noted that the fix was incomplete [3]. While the series was merged, fast_find_migrateblock was still disabled. This series should fix the corner cases and allow 95e7a450b819 ("Revert "mm/compaction: fix set skip in fast_find_migrateblock"") to be safely reverted. Details on how many pageblocks are rescanned are in the changelog of the last patch.

[1] https://lore.kernel.org/r/[email protected]
[2] https://bugzilla.suse.com/show_bug.cgi?id=1206848
[3] https://lore.kernel.org/r/[email protected]

I've run this series on my machine, which can solve the regression I met before [1], and no other issues found. So please feel free to add:

[1] https://lore.kernel.org/all/3576e3520c044beb2a81860aecb2d4f597089300.1682521303.git.baolin.wang@linux.alibaba.com/

Quoted 4 lines
  mm/compaction.c | 24 ++++++++++--------------
  1 file changed, 10 insertions(+), 14 deletions(-)