Skip to contentSkip to conversation

mm: Fix for negative counter: nr_file_hugepages

PATCH Superseded Stefan Roesch 4 participants 5 messages Nov 6, 2023

Subsystems (2)
Memory management THP

v1 of 3 Superseded Latest is v3 →

lore
Share & tools
lore

Stefan Roesch

author
#0 lore raw

PATCH 1/1 mm: Fix for negative counter: nr_file_hugepages

While qualifiying the 6.4 release, the following warning was detected in messages:

vmstat_refresh: nr_file_hugepages -15664

The warning is caused by the incorrect updating of the NR_FILE_THPS counter in the function split_huge_page_to_list. The if case is checking for folio_test_swapbacked, but the else case is missing the check for folio_test_pmd_mappable. The other functions that manipulate the counter like __filemap_add_folio and filemap_unaccount_folio have the corresponding check.

I have a test case, which reproduces the problem. It can be found here:
https://github.com/sroeschus/testcase/blob/main/vmstat_refresh/madv.c

The test case reproduces on an XFS filesystem. Running the same test case on a BTRFS filesystem does not reproduce the problem.

AFAIK version 6.1 until 6.6 are affected by this problem.

Signed-off-by Stefan Roesch <[email protected]>
Co-debugged-by Johannes Weiner <[email protected]>
---
 mm/huge_memory.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
mm/huge_memory.c +2 -1
index 064fbd90822b4..ea6bee675c4d3 100644
@@ -2740,7 +2740,8 @@ int split_huge_page_to_list(struct page *page, struct list_head *list)
 			if (folio_test_swapbacked(folio)) {
 				__lruvec_stat_mod_folio(folio, NR_SHMEM_THPS,
 							-nr);
-			} else {
+			} else if (folio_test_pmd_mappable(folio)) {
+
 				__lruvec_stat_mod_folio(folio, NR_FILE_THPS,
 							-nr);
 				filemap_nr_thps_dec(mapping);

base-commit: ffc253263a1375a65fa6c9f62a893e9767fbebfa

Signature · 1 line
2.39.3
maintainer
#1 lore raw
Quoted 12 lines · On Mon, 6 Nov 2023 10:19:18 -0800 Stefan Roesch <[email protected]> wrote:

While qualifiying the 6.4 release, the following warning was detected in messages:

vmstat_refresh: nr_file_hugepages -15664

The warning is caused by the incorrect updating of the NR_FILE_THPS counter in the function split_huge_page_to_list. The if case is checking for folio_test_swapbacked, but the else case is missing the check for folio_test_pmd_mappable. The other functions that manipulate the counter like __filemap_add_folio and filemap_unaccount_folio have the corresponding check.

Quoted 7 lines

I have a test case, which reproduces the problem. It can be found here:
https://github.com/sroeschus/testcase/blob/main/vmstat_refresh/madv.c

The test case reproduces on an XFS filesystem. Running the same test case on a BTRFS filesystem does not reproduce the problem.

AFAIK version 6.1 until 6.6 are affected by this problem.

I'm thinking a cc:stable is justified.

mm/huge_memory.c int split_huge_page_to_list(struct page *page, struct list_head *list) · line 2747

Quoted 13 lines
mm/huge_memory.c +2 -1
@@ -2740,7 +2740,8 @@ int split_huge_page_to_list(struct page *page, struct list_head *list)
 			if (folio_test_swapbacked(folio)) {
 				__lruvec_stat_mod_folio(folio, NR_SHMEM_THPS,
 							-nr);
-			} else {
+			} else if (folio_test_pmd_mappable(folio)) {
+
 				__lruvec_stat_mod_folio(folio, NR_FILE_THPS,
 							-nr);
 				filemap_nr_thps_dec(mapping);

I expect this will backport OK until it hits 3e9a13daa ("huge_memory:
convert split_huge_page_to_list() to use a folio") at which point the -stable maintainers might request a reworked version.

#2 lore raw
In reply to Stefan Roesch

mm/huge_memory.c int split_huge_page_to_list(struct page *page, struct list_head *list) · line 2747

Quoted 11 lines · Stefan Roesch

+++ b/mm/huge_memory.c

mm/huge_memory.c +2 -1
@@ -2740,7 +2740,8 @@ int split_huge_page_to_list(struct page *page, struct list_head *list)
 			if (folio_test_swapbacked(folio)) {
 				__lruvec_stat_mod_folio(folio, NR_SHMEM_THPS,
 							-nr);
-			} else {
+			} else if (folio_test_pmd_mappable(folio)) {
+
 				__lruvec_stat_mod_folio(folio, NR_FILE_THPS,
 							-nr);
 				filemap_nr_thps_dec(mapping);

Good catch. Two things:

1. No blank line after the 'else if'.

2. We're leaving a bit of a landmine for shmem when it gets support for arbitrary folio sizes. Really all of this should be under a test_pmd_mappable.

#3 lore raw
Quoted 15 lines · Matthew Wilcox

mm/huge_memory.c int split_huge_page_to_list(struct page *page, struct list_head *list) · line 2747

Quoted 11 lines · Stefan Roesch

+++ b/mm/huge_memory.c

mm/huge_memory.c +2 -1
@@ -2740,7 +2740,8 @@ int split_huge_page_to_list(struct page *page, struct list_head *list)
 			if (folio_test_swapbacked(folio)) {
 				__lruvec_stat_mod_folio(folio, NR_SHMEM_THPS,
 							-nr);
-			} else {
+			} else if (folio_test_pmd_mappable(folio)) {
+
 				__lruvec_stat_mod_folio(folio, NR_FILE_THPS,
 							-nr);
 				filemap_nr_thps_dec(mapping);

Good catch. Two things:

Quoted 5 lines

1. No blank line after the 'else if'.

2. We're leaving a bit of a landmine for shmem when it gets support for arbitrary folio sizes. Really all of this should be under a test_pmd_mappable.

I was wondering if we want to keep NR_FILE_THPS permanently for original flavor 512 basepage THPs, or whether they should account large folios as well? Same for NR_ANON_THPS and NR_SHMEM_THPS.

If so, then I agree this should all be conditional on pmdmapped. I suppose the same in filemap_unaccount_folio().

Acked-by on v1 1/1
#4 lore raw
In reply to Stefan Roesch
Quoted 15 lines · Stefan Roesch

While qualifiying the 6.4 release, the following warning was detected in messages:

vmstat_refresh: nr_file_hugepages -15664

The warning is caused by the incorrect updating of the NR_FILE_THPS counter in the function split_huge_page_to_list. The if case is checking for folio_test_swapbacked, but the else case is missing the check for folio_test_pmd_mappable. The other functions that manipulate the counter like __filemap_add_folio and filemap_unaccount_folio have the corresponding check.

I have a test case, which reproduces the problem. It can be found here:
https://github.com/sroeschus/testcase/blob/main/vmstat_refresh/madv.c

Quoted 7 lines

The test case reproduces on an XFS filesystem. Running the same test case on a BTRFS filesystem does not reproduce the problem.

AFAIK version 6.1 until 6.6 are affected by this problem.

Signed-off-by Stefan Roesch <[email protected]>
Co-debugged-by Johannes Weiner <[email protected]>

With the newline fix Willy pointed out, and CC: stable:

Acked-by Johannes Weiner <[email protected]>