* [PATCH 1/4] mm: compaction: Ensure rescanning only happens on partially scanned pageblocks
2023-05-15 11:33 [PATCH 0/4] Follow-up "Fix excessive CPU usage during compaction" Mel Gorman
@ 2023-05-15 11:33 ` Mel Gorman
2023-05-25 9:57 ` Vlastimil Babka
2023-05-15 11:33 ` [PATCH 2/4] mm: compaction: Only force pageblock scan completion when skip hints are obeyed Mel Gorman
` (4 subsequent siblings)
5 siblings, 1 reply; 16+ messages in thread
From: Mel Gorman @ 2023-05-15 11:33 UTC (permalink / raw)
To: Vlastimil Babka
Cc: Andrew Morton, Jiri Slaby, Maxim Levitsky, Michal Hocko,
Pedro Falcato, Paolo Bonzini, Chuyi Zhou, Linux-MM, LKML,
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.
Reported-by: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
---
mm/compaction.c | 5 +++--
1 file changed, 3 insertions(+), 2 deletions(-)
diff --git a/mm/compaction.c b/mm/compaction.c
index c8bcdea15f5f..81791c124bb8 100644
--- a/mm/compaction.c
+++ b/mm/compaction.c
@@ -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;
/*
--
2.35.3
^ permalink raw reply [flat|nested] 16+ messages in thread* Re: [PATCH 1/4] mm: compaction: Ensure rescanning only happens on partially scanned pageblocks
2023-05-15 11:33 ` [PATCH 1/4] mm: compaction: Ensure rescanning only happens on partially scanned pageblocks Mel Gorman
@ 2023-05-25 9:57 ` Vlastimil Babka
0 siblings, 0 replies; 16+ messages in thread
From: Vlastimil Babka @ 2023-05-25 9:57 UTC (permalink / raw)
To: Mel Gorman
Cc: Andrew Morton, Jiri Slaby, Maxim Levitsky, Michal Hocko,
Pedro Falcato, Paolo Bonzini, Chuyi Zhou, Linux-MM, LKML
On 5/15/23 13:33, Mel Gorman wrote:
> 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.
>
> Reported-by: Vlastimil Babka <vbabka@suse.cz>
> Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
> ---
> mm/compaction.c | 5 +++--
> 1 file changed, 3 insertions(+), 2 deletions(-)
>
> diff --git a/mm/compaction.c b/mm/compaction.c
> index c8bcdea15f5f..81791c124bb8 100644
> --- a/mm/compaction.c
> +++ b/mm/compaction.c
> @@ -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;
>
> /*
^ permalink raw reply [flat|nested] 16+ messages in thread
* [PATCH 2/4] mm: compaction: Only force pageblock scan completion when skip hints are obeyed
2023-05-15 11:33 [PATCH 0/4] Follow-up "Fix excessive CPU usage during compaction" Mel Gorman
2023-05-15 11:33 ` [PATCH 1/4] mm: compaction: Ensure rescanning only happens on partially scanned pageblocks Mel Gorman
@ 2023-05-15 11:33 ` Mel Gorman
2023-05-25 10:01 ` Vlastimil Babka
2023-05-15 11:33 ` [PATCH 3/4] mm: compaction: Update pageblock skip when first migration candidate is not at the start Mel Gorman
` (3 subsequent siblings)
5 siblings, 1 reply; 16+ messages in thread
From: Mel Gorman @ 2023-05-15 11:33 UTC (permalink / raw)
To: Vlastimil Babka
Cc: Andrew Morton, Jiri Slaby, Maxim Levitsky, Michal Hocko,
Pedro Falcato, Paolo Bonzini, Chuyi Zhou, Linux-MM, LKML,
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.
Suggested-by: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
---
mm/compaction.c | 5 +++--
1 file changed, 3 insertions(+), 2 deletions(-)
diff --git a/mm/compaction.c b/mm/compaction.c
index 81791c124bb8..accc6568091a 100644
--- a/mm/compaction.c
+++ b/mm/compaction.c
@@ -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;
--
2.35.3
^ permalink raw reply [flat|nested] 16+ messages in thread* Re: [PATCH 2/4] mm: compaction: Only force pageblock scan completion when skip hints are obeyed
2023-05-15 11:33 ` [PATCH 2/4] mm: compaction: Only force pageblock scan completion when skip hints are obeyed Mel Gorman
@ 2023-05-25 10:01 ` Vlastimil Babka
0 siblings, 0 replies; 16+ messages in thread
From: Vlastimil Babka @ 2023-05-25 10:01 UTC (permalink / raw)
To: Mel Gorman
Cc: Andrew Morton, Jiri Slaby, Maxim Levitsky, Michal Hocko,
Pedro Falcato, Paolo Bonzini, Chuyi Zhou, Linux-MM, LKML
On 5/15/23 13:33, Mel Gorman wrote:
> 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.
>
> Suggested-by: Vlastimil Babka <vbabka@suse.cz>
> Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
> ---
> mm/compaction.c | 5 +++--
> 1 file changed, 3 insertions(+), 2 deletions(-)
>
> diff --git a/mm/compaction.c b/mm/compaction.c
> index 81791c124bb8..accc6568091a 100644
> --- a/mm/compaction.c
> +++ b/mm/compaction.c
> @@ -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;
>
^ permalink raw reply [flat|nested] 16+ messages in thread
* [PATCH 3/4] mm: compaction: Update pageblock skip when first migration candidate is not at the start
2023-05-15 11:33 [PATCH 0/4] Follow-up "Fix excessive CPU usage during compaction" Mel Gorman
2023-05-15 11:33 ` [PATCH 1/4] mm: compaction: Ensure rescanning only happens on partially scanned pageblocks Mel Gorman
2023-05-15 11:33 ` [PATCH 2/4] mm: compaction: Only force pageblock scan completion when skip hints are obeyed Mel Gorman
@ 2023-05-15 11:33 ` Mel Gorman
2023-05-25 13:37 ` Vlastimil Babka
2023-05-15 11:33 ` [PATCH 4/4] Revert "Revert "mm/compaction: fix set skip in fast_find_migrateblock"" Mel Gorman
` (2 subsequent siblings)
5 siblings, 1 reply; 16+ messages in thread
From: Mel Gorman @ 2023-05-15 11:33 UTC (permalink / raw)
To: Vlastimil Babka
Cc: Andrew Morton, Jiri Slaby, Maxim Levitsky, Michal Hocko,
Pedro Falcato, Paolo Bonzini, Chuyi Zhou, Linux-MM, LKML,
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.
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
---
mm/compaction.c | 15 +++++----------
1 file changed, 5 insertions(+), 10 deletions(-)
diff --git a/mm/compaction.c b/mm/compaction.c
index accc6568091a..d7be990b1d60 100644
--- a/mm/compaction.c
+++ b/mm/compaction.c
@@ -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;
}
--
2.35.3
^ permalink raw reply [flat|nested] 16+ messages in thread* Re: [PATCH 3/4] mm: compaction: Update pageblock skip when first migration candidate is not at the start
2023-05-15 11:33 ` [PATCH 3/4] mm: compaction: Update pageblock skip when first migration candidate is not at the start Mel Gorman
@ 2023-05-25 13:37 ` Vlastimil Babka
2023-05-29 10:33 ` Mel Gorman
0 siblings, 1 reply; 16+ messages in thread
From: Vlastimil Babka @ 2023-05-25 13:37 UTC (permalink / raw)
To: Mel Gorman
Cc: Andrew Morton, Jiri Slaby, Maxim Levitsky, Michal Hocko,
Pedro Falcato, Paolo Bonzini, Chuyi Zhou, Linux-MM, LKML
On 5/15/23 13:33, Mel Gorman wrote:
> 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.
>
> Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
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?
> ---
> mm/compaction.c | 15 +++++----------
> 1 file changed, 5 insertions(+), 10 deletions(-)
>
> diff --git a/mm/compaction.c b/mm/compaction.c
> index accc6568091a..d7be990b1d60 100644
> --- a/mm/compaction.c
> +++ b/mm/compaction.c
> @@ -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;
> }
>
^ permalink raw reply [flat|nested] 16+ messages in thread* Re: [PATCH 3/4] mm: compaction: Update pageblock skip when first migration candidate is not at the start
2023-05-25 13:37 ` Vlastimil Babka
@ 2023-05-29 10:33 ` Mel Gorman
2023-05-29 12:43 ` Vlastimil Babka
0 siblings, 1 reply; 16+ messages in thread
From: Mel Gorman @ 2023-05-29 10:33 UTC (permalink / raw)
To: Vlastimil Babka
Cc: Andrew Morton, Jiri Slaby, Maxim Levitsky, Michal Hocko,
Pedro Falcato, Paolo Bonzini, Chuyi Zhou, Linux-MM, LKML
On Thu, May 25, 2023 at 03:37:43PM +0200, Vlastimil Babka wrote:
> On 5/15/23 13:33, Mel Gorman wrote:
> > 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.
> >
> > Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
>
> 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?
>
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 :(
diff --git a/mm/compaction.c b/mm/compaction.c
index 91af6a8b7a98..761a2dd7d78a 100644
--- a/mm/compaction.c
+++ b/mm/compaction.c
@@ -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;
}
^ permalink raw reply [flat|nested] 16+ messages in thread* Re: [PATCH 3/4] mm: compaction: Update pageblock skip when first migration candidate is not at the start
2023-05-29 10:33 ` Mel Gorman
@ 2023-05-29 12:43 ` Vlastimil Babka
[not found] ` <20230602111622.swtxhn6lu2qwgrwq@techsingularity.net>
0 siblings, 1 reply; 16+ messages in thread
From: Vlastimil Babka @ 2023-05-29 12:43 UTC (permalink / raw)
To: Mel Gorman
Cc: Andrew Morton, Jiri Slaby, Maxim Levitsky, Michal Hocko,
Pedro Falcato, Paolo Bonzini, Chuyi Zhou, Linux-MM, LKML
On 5/29/23 12:33, Mel Gorman wrote:
> On Thu, May 25, 2023 at 03:37:43PM +0200, Vlastimil Babka wrote:
>> On 5/15/23 13:33, Mel Gorman wrote:
>> > 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.
>> >
>> > Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
>>
>> 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?
>>
>
> 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.
> diff --git a/mm/compaction.c b/mm/compaction.c
> index 91af6a8b7a98..761a2dd7d78a 100644
> --- a/mm/compaction.c
> +++ b/mm/compaction.c
> @@ -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;
> }
>
^ permalink raw reply [flat|nested] 16+ messages in thread
* [PATCH 4/4] Revert "Revert "mm/compaction: fix set skip in fast_find_migrateblock""
2023-05-15 11:33 [PATCH 0/4] Follow-up "Fix excessive CPU usage during compaction" Mel Gorman
` (2 preceding siblings ...)
2023-05-15 11:33 ` [PATCH 3/4] mm: compaction: Update pageblock skip when first migration candidate is not at the start Mel Gorman
@ 2023-05-15 11:33 ` Mel Gorman
2023-05-25 13:42 ` Vlastimil Babka
2023-05-19 6:43 ` [PATCH 0/4] Follow-up "Fix excessive CPU usage during compaction" Raghavendra K T
2023-05-23 13:47 ` Baolin Wang
5 siblings, 1 reply; 16+ messages in thread
From: Mel Gorman @ 2023-05-15 11:33 UTC (permalink / raw)
To: Vlastimil Babka
Cc: Andrew Morton, Jiri Slaby, Maxim Levitsky, Michal Hocko,
Pedro Falcato, Paolo Bonzini, Chuyi Zhou, Linux-MM, LKML,
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.
Link: https://bugzilla.suse.com/show_bug.cgi?id=1206848
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
---
mm/compaction.c | 1 -
1 file changed, 1 deletion(-)
diff --git a/mm/compaction.c b/mm/compaction.c
index d7be990b1d60..91af6a8b7a98 100644
--- a/mm/compaction.c
+++ b/mm/compaction.c
@@ -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;
}
}
--
2.35.3
^ permalink raw reply [flat|nested] 16+ messages in thread* Re: [PATCH 4/4] Revert "Revert "mm/compaction: fix set skip in fast_find_migrateblock""
2023-05-15 11:33 ` [PATCH 4/4] Revert "Revert "mm/compaction: fix set skip in fast_find_migrateblock"" Mel Gorman
@ 2023-05-25 13:42 ` Vlastimil Babka
0 siblings, 0 replies; 16+ messages in thread
From: Vlastimil Babka @ 2023-05-25 13:42 UTC (permalink / raw)
To: Mel Gorman
Cc: Andrew Morton, Jiri Slaby, Maxim Levitsky, Michal Hocko,
Pedro Falcato, Paolo Bonzini, Chuyi Zhou, Linux-MM, LKML
On 5/15/23 13:33, Mel Gorman wrote:
> 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.
>
> Link: https://bugzilla.suse.com/show_bug.cgi?id=1206848
> Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
My concerns about patch 3/4 don't affect this part so
Acked-by: Vlastimil Babka <vbabka@suse.cz>
> ---
> mm/compaction.c | 1 -
> 1 file changed, 1 deletion(-)
>
> diff --git a/mm/compaction.c b/mm/compaction.c
> index d7be990b1d60..91af6a8b7a98 100644
> --- a/mm/compaction.c
> +++ b/mm/compaction.c
> @@ -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;
> }
> }
^ permalink raw reply [flat|nested] 16+ messages in thread
* Re: [PATCH 0/4] Follow-up "Fix excessive CPU usage during compaction"
2023-05-15 11:33 [PATCH 0/4] Follow-up "Fix excessive CPU usage during compaction" Mel Gorman
` (3 preceding siblings ...)
2023-05-15 11:33 ` [PATCH 4/4] Revert "Revert "mm/compaction: fix set skip in fast_find_migrateblock"" Mel Gorman
@ 2023-05-19 6:43 ` Raghavendra K T
2023-05-21 19:20 ` Mel Gorman
2023-05-23 13:47 ` Baolin Wang
5 siblings, 1 reply; 16+ messages in thread
From: Raghavendra K T @ 2023-05-19 6:43 UTC (permalink / raw)
To: Mel Gorman, Vlastimil Babka
Cc: Andrew Morton, Jiri Slaby, Maxim Levitsky, Michal Hocko,
Pedro Falcato, Paolo Bonzini, Chuyi Zhou, Linux-MM, LKML
On 5/15/2023 5:03 PM, Mel Gorman wrote:
> 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/20230125134434.18017-1-mgorman@techsingularity.net
> [2] https://bugzilla.suse.com/show_bug.cgi?id=1206848
> [3] https://lore.kernel.org/r/a55cf026-a2f9-ef01-9a4c-398693e048ea@suse.cz
>
> 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
^ permalink raw reply [flat|nested] 16+ messages in thread* Re: [PATCH 0/4] Follow-up "Fix excessive CPU usage during compaction"
2023-05-19 6:43 ` [PATCH 0/4] Follow-up "Fix excessive CPU usage during compaction" Raghavendra K T
@ 2023-05-21 19:20 ` Mel Gorman
0 siblings, 0 replies; 16+ messages in thread
From: Mel Gorman @ 2023-05-21 19:20 UTC (permalink / raw)
To: Raghavendra K T
Cc: Vlastimil Babka, Andrew Morton, Jiri Slaby, Maxim Levitsky,
Michal Hocko, Pedro Falcato, Paolo Bonzini, Chuyi Zhou, Linux-MM,
LKML
On Fri, May 19, 2023 at 12:13:11PM +0530, Raghavendra K T wrote:
> On 5/15/2023 5:03 PM, Mel Gorman wrote:
> > 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/20230125134434.18017-1-mgorman@techsingularity.net
> > [2] https://bugzilla.suse.com/show_bug.cgi?id=1206848
> > [3] https://lore.kernel.org/r/a55cf026-a2f9-ef01-9a4c-398693e048ea@suse.cz
> >
> > 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).
>
> <SNIP>
>
Thanks Raghavendra!
--
Mel Gorman
SUSE Labs
^ permalink raw reply [flat|nested] 16+ messages in thread
* Re: [PATCH 0/4] Follow-up "Fix excessive CPU usage during compaction"
2023-05-15 11:33 [PATCH 0/4] Follow-up "Fix excessive CPU usage during compaction" Mel Gorman
` (4 preceding siblings ...)
2023-05-19 6:43 ` [PATCH 0/4] Follow-up "Fix excessive CPU usage during compaction" Raghavendra K T
@ 2023-05-23 13:47 ` Baolin Wang
5 siblings, 0 replies; 16+ messages in thread
From: Baolin Wang @ 2023-05-23 13:47 UTC (permalink / raw)
To: Mel Gorman, Vlastimil Babka
Cc: Andrew Morton, Jiri Slaby, Maxim Levitsky, Michal Hocko,
Pedro Falcato, Paolo Bonzini, Chuyi Zhou, Linux-MM, LKML
Hi Mel,
On 5/15/2023 7:33 PM, Mel Gorman wrote:
> 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/20230125134434.18017-1-mgorman@techsingularity.net
> [2] https://bugzilla.suse.com/show_bug.cgi?id=1206848
> [3] https://lore.kernel.org/r/a55cf026-a2f9-ef01-9a4c-398693e048ea@suse.cz
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:
Tested-by: Baolin Wang <baolin.wang@linux.alibaba.com>
[1]
https://lore.kernel.org/all/3576e3520c044beb2a81860aecb2d4f597089300.1682521303.git.baolin.wang@linux.alibaba.com/
>
> mm/compaction.c | 24 ++++++++++--------------
> 1 file changed, 10 insertions(+), 14 deletions(-)
>
^ permalink raw reply [flat|nested] 16+ messages in thread