Skip to contentSkip to conversation

mm/vmscan: restore allowed mask in alloc_demote_folio()

PATCH In review Bing Jiao 3 participants 6 messages Mar 2 – Mar 3

Subsystems (3)
lore
Share & tools
lore

Bing Jiao

author
#0 lore raw

PATCH 1/1 mm/vmscan: restore allowed mask in alloc_demote_folio()

In alloc_demote_folio(), mtc->nmask is set to NULL for the first allocation. If that succeeds, it returns without restoring mtc->nmask to allowed_mask. For subsequent allocations from the migrate_pages() batch, mtc->nmask will be NULL. If the target node then becomes full, the fallback allocation will use nmask = NULL, allocating from any node allowed by the task cpuset, which for kswapd is all nodes.

To address this issue, restore the mtc->nmask to its original allowed nodemask after the first allocation.

---
 mm/vmscan.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
mm/vmscan.c +1 -1
index cbffc0a27824..b42abd17aee7 100644
@@ -985,11 +985,11 @@ static struct folio *alloc_demote_folio(struct folio *src,
 	mtc->nmask = NULL;
 	mtc->gfp_mask |= __GFP_THISNODE;
 	dst = alloc_migration_target(src, (unsigned long)mtc);
+	mtc->nmask = allowed_mask;
 	if (dst)
 		return dst;
 
 	mtc->gfp_mask &= ~__GFP_THISNODE;
-	mtc->nmask = allowed_mask;
 
 	return alloc_migration_target(src, (unsigned long)mtc);
 }

--
2.53.0.473.g4a7958ca14-goog

#1 lore raw
Quoted 10 lines · Bing Jiao

In alloc_demote_folio(), mtc->nmask is set to NULL for the first allocation. If that succeeds, it returns without restoring mtc->nmask to allowed_mask. For subsequent allocations from the migrate_pages() batch, mtc->nmask will be NULL. If the target node then becomes full, the fallback allocation will use nmask = NULL, allocating from any node allowed by the task cpuset, which for kswapd is all nodes.

To address this issue, restore the mtc->nmask to its original allowed nodemask after the first allocation.

That would be

Fixes 320080272892 ("mm/demotion: demote pages according to allocation fallback order")

?

Quoted 23 lines
---
 mm/vmscan.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
mm/vmscan.c +1 -2
index cbffc0a27824..b42abd17aee7 100644
@@ -985,11 +985,11 @@ static struct folio *alloc_demote_folio(struct folio *src,
 	mtc->nmask = NULL;
 	mtc->gfp_mask |= __GFP_THISNODE;
 	dst = alloc_migration_target(src, (unsigned long)mtc);
+	mtc->nmask = allowed_mask;
 	if (dst)
 		return dst;
 
 	mtc->gfp_mask &= ~__GFP_THISNODE;
-	mtc->nmask = allowed_mask;
 
 	return alloc_migration_target(src, (unsigned long)mtc);
 }
--
Quoted 2 lines

2.53.0.473.g4a7958ca14-goog

Maybe we should just not touch the original mtc?

mm/vmscan.c +5 -8
index de62225b381a..f07716e5389e 100644
@@ -985,9 +985,9 @@ static void folio_check_dirty_writeback(struct folio *folio,
 static struct folio *alloc_demote_folio(struct folio *src,
                unsigned long private)
 {
+       struct migration_target_control *mtc, target_nid_mtc;
        struct folio *dst;
        nodemask_t *allowed_mask;
-       struct migration_target_control *mtc;
 
        mtc = (struct migration_target_control *)private;
 
@@ -1001,15 +1001,12 @@ static struct folio *alloc_demote_folio(struct folio *src,
         * a demotion of cold pages from the target memtier. This can result
         * in the kernel placing hot pages in slower(lower) memory tiers.
         */
-       mtc->nmask = NULL;
-       mtc->gfp_mask |= __GFP_THISNODE;
-       dst = alloc_migration_target(src, (unsigned long)mtc);
+       target_nid_mtc = *mtc;
+       target_nid_mtc.nmask = NULL;
+       target_nid_mtc.gfp_mask |= __GFP_THISNODE;
+       dst = alloc_migration_target(src, (unsigned long)&target_nid_mtc);
        if (dst)
                return dst;
-
-       mtc->gfp_mask &= ~__GFP_THISNODE;
-       mtc->nmask = allowed_mask;
-
        return alloc_migration_target(src, (unsigned long)mtc);
 }
 
Signature · 3 lines
Cheers,

David
author
#2 lore raw
Quoted 12 lines · David Hildenbrand (Arm)
Quoted 10 lines · Bing Jiao

In alloc_demote_folio(), mtc->nmask is set to NULL for the first allocation. If that succeeds, it returns without restoring mtc->nmask to allowed_mask. For subsequent allocations from the migrate_pages() batch, mtc->nmask will be NULL. If the target node then becomes full, the fallback allocation will use nmask = NULL, allocating from any node allowed by the task cpuset, which for kswapd is all nodes.

To address this issue, restore the mtc->nmask to its original allowed nodemask after the first allocation.

Quoted 5 lines

That would be

Fixes 320080272892 ("mm/demotion: demote pages according to allocation fallback order")

?

Thanks for pointing it out. Will add it in the new patch.

Quoted 68 lines
Quoted 23 lines
---
 mm/vmscan.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
mm/vmscan.c +1 -2
index cbffc0a27824..b42abd17aee7 100644
@@ -985,11 +985,11 @@ static struct folio *alloc_demote_folio(struct folio *src,
 	mtc->nmask = NULL;
 	mtc->gfp_mask |= __GFP_THISNODE;
 	dst = alloc_migration_target(src, (unsigned long)mtc);
+	mtc->nmask = allowed_mask;
 	if (dst)
 		return dst;
 
 	mtc->gfp_mask &= ~__GFP_THISNODE;
-	mtc->nmask = allowed_mask;
 
 	return alloc_migration_target(src, (unsigned long)mtc);
 }
--
Quoted 2 lines

2.53.0.473.g4a7958ca14-goog

Maybe we should just not touch the original mtc?

mm/vmscan.c +5 -9
index de62225b381a..f07716e5389e 100644
@@ -985,9 +985,9 @@ static void folio_check_dirty_writeback(struct folio *folio,
 static struct folio *alloc_demote_folio(struct folio *src,
                unsigned long private)
 {
+       struct migration_target_control *mtc, target_nid_mtc;
        struct folio *dst;
        nodemask_t *allowed_mask;
-       struct migration_target_control *mtc;
 
        mtc = (struct migration_target_control *)private;
 
@@ -1001,15 +1001,12 @@ static struct folio *alloc_demote_folio(struct folio *src,
         * a demotion of cold pages from the target memtier. This can result
         * in the kernel placing hot pages in slower(lower) memory tiers.
         */
-       mtc->nmask = NULL;
-       mtc->gfp_mask |= __GFP_THISNODE;
-       dst = alloc_migration_target(src, (unsigned long)mtc);
+       target_nid_mtc = *mtc;
+       target_nid_mtc.nmask = NULL;
+       target_nid_mtc.gfp_mask |= __GFP_THISNODE;
+       dst = alloc_migration_target(src, (unsigned long)&target_nid_mtc);
        if (dst)
                return dst;
-
-       mtc->gfp_mask &= ~__GFP_THISNODE;
-       mtc->nmask = allowed_mask;
-
        return alloc_migration_target(src, (unsigned long)mtc);
 }
 
 
 
--

Cheers,

Quoted 1 line

David

Thank you for the suggestion, David.

I agree that not touching the original mtc is a better. It makes the distinction between the two allocation attempts much clearer and avoids the side-effect bug. Will update it then.

Best,
Bing

author
#3 lore raw

PATCH v2 1/1 mm/vmscan: fix unintended mtc->nmask mutation in alloc_demote_folio()

In reply to Bing Jiao

In alloc_demote_folio(), mtc->nmask is set to NULL for the first allocation. If that succeeds, it returns without restoring mtc->nmask to allowed_mask. For subsequent allocations from the migrate_pages() batch, mtc->nmask will be NULL. If the target node then becomes full, the fallback allocation will use nmask = NULL, allocating from any node allowed by the task cpuset, which for kswapd is all nodes.

To address this issue, use a local copy of the mtc structure with nmask = NULL for the first allocation attempt specifically, ensuring the original mtc remains unmodified.

Fixes 320080272892 ("mm/demotion: demote pages according to allocation fallback order")
---
 mm/vmscan.c | 14 +++++---------
 1 file changed, 5 insertions(+), 9 deletions(-)
mm/vmscan.c +5 -9
index cbffc0a27824..c4e0ce737e03 100644
@@ -966,13 +966,11 @@ static void folio_check_dirty_writeback(struct folio *folio,
 static struct folio *alloc_demote_folio(struct folio *src,
 		unsigned long private)
 {
+	struct migration_target_control *mtc, target_nid_mtc;
 	struct folio *dst;
-	nodemask_t *allowed_mask;
-	struct migration_target_control *mtc;
 
 	mtc = (struct migration_target_control *)private;
 
-	allowed_mask = mtc->nmask;
 	/*
 	 * make sure we allocate from the target node first also trying to
 	 * demote or reclaim pages from the target node via kswapd if we are
@@ -982,15 +980,13 @@ static struct folio *alloc_demote_folio(struct folio *src,
 	 * a demotion of cold pages from the target memtier. This can result
 	 * in the kernel placing hot pages in slower(lower) memory tiers.
 	 */
-	mtc->nmask = NULL;
-	mtc->gfp_mask |= __GFP_THISNODE;
-	dst = alloc_migration_target(src, (unsigned long)mtc);
+	target_nid_mtc = *mtc;
+	target_nid_mtc.nmask = NULL;
+	target_nid_mtc.gfp_mask |= __GFP_THISNODE;
+	dst = alloc_migration_target(src, (unsigned long)&target_nid_mtc);
 	if (dst)
 		return dst;
 
-	mtc->gfp_mask &= ~__GFP_THISNODE;
-	mtc->nmask = allowed_mask;
-
 	return alloc_migration_target(src, (unsigned long)mtc);
 }
 

--
2.53.0.473.g4a7958ca14-goog

reviewer Acked-by on v2 1/1
#4 lore raw

Re: [PATCH v2] mm/vmscan: fix unintended mtc->nmask mutation in alloc_demote_folio()

Quoted 11 lines · Bing Jiao

In alloc_demote_folio(), mtc->nmask is set to NULL for the first allocation. If that succeeds, it returns without restoring mtc->nmask to allowed_mask. For subsequent allocations from the migrate_pages() batch, mtc->nmask will be NULL. If the target node then becomes full, the fallback allocation will use nmask = NULL, allocating from any node allowed by the task cpuset, which for kswapd is all nodes.

To address this issue, use a local copy of the mtc structure with nmask = NULL for the first allocation attempt specifically, ensuring the original mtc remains unmodified.

Quoted 5 lines
Fixes 320080272892 ("mm/demotion: demote pages according to allocation fallback order")
---
 mm/vmscan.c | 14 +++++---------
 1 file changed, 5 insertions(+), 9 deletions(-)
Signature · 3 lines
Cheers,

David
Reviewed-by on v2 1/1
#5 lore raw

Re: [PATCH v2] mm/vmscan: fix unintended mtc->nmask mutation in alloc_demote_folio()

In reply to Bing Jiao
Quoted 7 lines · Bing Jiao

In alloc_demote_folio(), mtc->nmask is set to NULL for the first allocation. If that succeeds, it returns without restoring mtc->nmask to allowed_mask. For subsequent allocations from the migrate_pages() batch, mtc->nmask will be NULL. If the target node then becomes full, the fallback allocation will use nmask = NULL, allocating from any node allowed by the task cpuset, which for kswapd is all nodes.

Quoted 6 lines

To address this issue, use a local copy of the mtc structure with nmask = NULL for the first allocation attempt specifically, ensuring the original mtc remains unmodified.

Fixes 320080272892 ("mm/demotion: demote pages according to allocation fallback order")

LGTM, so:

Quoted 46 lines
---
 mm/vmscan.c | 14 +++++---------
 1 file changed, 5 insertions(+), 9 deletions(-)
mm/vmscan.c +5 -10
index cbffc0a27824..c4e0ce737e03 100644
@@ -966,13 +966,11 @@ static void folio_check_dirty_writeback(struct folio *folio,
 static struct folio *alloc_demote_folio(struct folio *src,
 		unsigned long private)
 {
+	struct migration_target_control *mtc, target_nid_mtc;
 	struct folio *dst;
-	nodemask_t *allowed_mask;
-	struct migration_target_control *mtc;
 
 	mtc = (struct migration_target_control *)private;
 
-	allowed_mask = mtc->nmask;
 	/*
 	 * make sure we allocate from the target node first also trying to
 	 * demote or reclaim pages from the target node via kswapd if we are
@@ -982,15 +980,13 @@ static struct folio *alloc_demote_folio(struct folio *src,
 	 * a demotion of cold pages from the target memtier. This can result
 	 * in the kernel placing hot pages in slower(lower) memory tiers.
 	 */
-	mtc->nmask = NULL;
-	mtc->gfp_mask |= __GFP_THISNODE;
-	dst = alloc_migration_target(src, (unsigned long)mtc);
+	target_nid_mtc = *mtc;
+	target_nid_mtc.nmask = NULL;
+	target_nid_mtc.gfp_mask |= __GFP_THISNODE;
+	dst = alloc_migration_target(src, (unsigned long)&target_nid_mtc);
 	if (dst)
 		return dst;
 
-	mtc->gfp_mask &= ~__GFP_THISNODE;
-	mtc->nmask = allowed_mask;
-
 	return alloc_migration_target(src, (unsigned long)mtc);
 }
 
--

2.53.0.473.g4a7958ca14-goog