Skip to contentSkip to conversation

mm: avoid KCSAN false positive in page_to_nid()

PATCH Superseded Hui Zhu 3 participants 4 messages Jun 23 – Jun 24

Subsystems (2)

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore

Hui Zhu

author
#0 lore raw

PATCH 1/1 mm: avoid KCSAN false positive in page_to_nid()

From: Hui Zhu <[email protected]>

KCSAN reports a data race between page_to_nid() reading page->flags and folio_trylock()/folio_lock() doing test_and_set_bit_lock(PG_locked, ...) on the same word from another CPU, e.g.:

  BUG: KCSAN: data-race in __lruvec_stat_mod_folio / shmem_get_folio_gfp

The node id occupies a fixed, high bit-range of page->flags that is set once when the page is initialized and never modified afterwards, so it can never overlap with the low PG_locked/PG_waiters bits touched by the folio lock path. The race is therefore harmless: page_to_nid() always returns a consistent value regardless of how the read interleaves with the lock bit ops.

Wrap the flags read with data_race() to tell KCSAN this race is intentional and benign, consistent with how page->page_type is already annotated for similar packed-field accesses.

---
 include/linux/mm.h | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)
include/linux/mm.h +8 -1
index 485df9c2dbdd..122d3b39369f 100644
@@ -2296,7 +2296,14 @@ static inline int memdesc_nid(memdesc_flags_t mdf)
 
 static inline int page_to_nid(const struct page *page)
 {
-	return memdesc_nid(PF_POISONED_CHECK(page)->flags);
+	/*
+	 * The node id occupies a fixed high bit-range of page->flags
+	 * that is set once at page init and never changed afterwards.
+	 * It cannot overlap with the low PG_locked/PG_waiters bits
+	 * that folio_lock()/folio_unlock() concurrently update, so
+	 * this data race is benign.
+	 */
+	return memdesc_nid(data_race(PF_POISONED_CHECK(page)->flags));
 }
 
 static inline int folio_nid(const struct folio *folio)
Signature · 1 line
2.43.0
#1 lore raw
Quoted 30 lines · Hui Zhu

From: Hui Zhu <[email protected]>

KCSAN reports a data race between page_to_nid() reading page->flags and folio_trylock()/folio_lock() doing test_and_set_bit_lock(PG_locked, ...) on the same word from another CPU, e.g.:

  BUG: KCSAN: data-race in __lruvec_stat_mod_folio / shmem_get_folio_gfp

The node id occupies a fixed, high bit-range of page->flags that is set once when the page is initialized and never modified afterwards, so it can never overlap with the low PG_locked/PG_waiters bits touched by the folio lock path. The race is therefore harmless: page_to_nid() always returns a consistent value regardless of how the read interleaves with the lock bit ops.

Wrap the flags read with data_race() to tell KCSAN this race is intentional and benign, consistent with how page->page_type is already annotated for similar packed-field accesses.

---
 include/linux/mm.h | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)
include/linux/mm.h +0 -0
index 485df9c2dbdd..122d3b39369f 100644
@@ -2296,7 +2296,14 @@ static inline int memdesc_nid(memdesc_flags_t mdf)
 

include/linux/mm.h static inline int memdesc_nid(memdesc_flags_t mdf) · line 2305

Quoted 10 lines
include/linux/mm.h +7 -1
 static inline int page_to_nid(const struct page *page)
 {
-	return memdesc_nid(PF_POISONED_CHECK(page)->flags);
+	/*
+	 * The node id occupies a fixed high bit-range of page->flags
+	 * that is set once at page init and never changed afterwards.
+	 * It cannot overlap with the low PG_locked/PG_waiters bits
+	 * that folio_lock()/folio_unlock() concurrently update, so
+	 * this data race is benign.
+	 */

Do we really need this excessive comment?

include/linux/mm.h static inline int memdesc_nid(memdesc_flags_t mdf) · line 2306

Quoted 1 line
include/linux/mm.h +1 -0
+	return memdesc_nid(data_race(PF_POISONED_CHECK(page)->flags));

In memdesc_zonenum() we use ASSERT_EXCLUSIVE_BITS.

Can we do the same here inside memdesc_nid?

include/linux/mm.h +1 -0
index 69daeeab7fe8f..76d3bb54be844 100644
@@ -2290,6 +2290,7 @@ int memdesc_nid(memdesc_flags_t mdf);
 #else
 static inline int memdesc_nid(memdesc_flags_t mdf)
 {
+       ASSERT_EXCLUSIVE_BITS(mdf.f, NODES_MASK << NODES_PGSHIFT);
        return (mdf.f >> NODES_PGSHIFT) & NODES_MASK;
 }
 #endif
Signature · 3 lines
Cheers,

David
#2 lore raw
Quoted 16 lines · David Hildenbrand (Arm)
Quoted 15 lines · Hui Zhu

From: Hui Zhu <[email protected]>

KCSAN reports a data race between page_to_nid() reading page->flags and folio_trylock()/folio_lock() doing test_and_set_bit_lock(PG_locked, ...) on the same word from another CPU, e.g.:

  BUG: KCSAN: data-race in __lruvec_stat_mod_folio / shmem_get_folio_gfp

The node id occupies a fixed, high bit-range of page->flags that is set once when the page is initialized and never modified afterwards, so it can never overlap with the low PG_locked/PG_waiters bits touched by the folio lock path. The race is therefore harmless: page_to_nid() always returns a consistent value regardless of how the read interleaves with the lock bit ops.

Quoted 5 lines
Quoted 5 lines

Wrap the flags read with data_race() to tell KCSAN this race is intentional and benign, consistent with how page->page_type is already annotated for similar packed-field accesses.

How did you discover this?

A syzbot report? If so please include Reported-by, Closes tags.

Quoted 10 lines
Quoted 10 lines
---
 include/linux/mm.h | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)
include/linux/mm.h +0 -0
index 485df9c2dbdd..122d3b39369f 100644
@@ -2296,7 +2296,14 @@ static inline int memdesc_nid(memdesc_flags_t mdf)
 
Quoted 12 lines

include/linux/mm.h static inline int memdesc_nid(memdesc_flags_t mdf) · line 2305

Quoted 10 lines
include/linux/mm.h +7 -1
 static inline int page_to_nid(const struct page *page)
 {
-	return memdesc_nid(PF_POISONED_CHECK(page)->flags);
+	/*
+	 * The node id occupies a fixed high bit-range of page->flags
+	 * that is set once at page init and never changed afterwards.
+	 * It cannot overlap with the low PG_locked/PG_waiters bits
+	 * that folio_lock()/folio_unlock() concurrently update, so
+	 * this data race is benign.
+	 */

Do we really need this excessive comment?

Agreed, just delete it. For a trivial benign data race it's a bit much, and the commit message can cover it off for those who are curious.

Quoted 6 lines

include/linux/mm.h static inline int memdesc_nid(memdesc_flags_t mdf) · line 2306

Quoted 2 lines
include/linux/mm.h +1 -0
+	return memdesc_nid(data_race(PF_POISONED_CHECK(page)->flags));

In memdesc_zonenum() we use ASSERT_EXCLUSIVE_BITS.

Can we do the same here inside memdesc_nid?

Also agreed

Quoted 18 lines
include/linux/mm.h +1 -1
index 69daeeab7fe8f..76d3bb54be844 100644
@@ -2290,6 +2290,7 @@ int memdesc_nid(memdesc_flags_t mdf);
 #else
 static inline int memdesc_nid(memdesc_flags_t mdf)
 {
+       ASSERT_EXCLUSIVE_BITS(mdf.f, NODES_MASK << NODES_PGSHIFT);
        return (mdf.f >> NODES_PGSHIFT) & NODES_MASK;
 }
 #endif
 
 
--

Cheers,

Quoted 1 line

David

Thanks, Lorenzo

author
#3 lore raw
Quoted 24 lines · David Hildenbrand (Arm)
Quoted 22 lines · Hui Zhu

From: Hui Zhu <[email protected]>

KCSAN reports a data race between page_to_nid() reading page->flags and folio_trylock()/folio_lock() doing test_and_set_bit_lock(PG_locked, ...) on the same word from another CPU, e.g.:

 BUG: KCSAN: data-race in __lruvec_stat_mod_folio / shmem_get_folio_gfp

The node id occupies a fixed, high bit-range of page->flags that is set once when the page is initialized and never modified afterwards, so it can never overlap with the low PG_locked/PG_waiters bits touched by the folio lock path. The race is therefore harmless: page_to_nid() always returns a consistent value regardless of how the read interleaves with the lock bit ops.

Wrap the flags read with data_race() to tell KCSAN this race is intentional and benign, consistent with how page->page_type is already annotated for similar packed-field accesses.

Quoted 5 lines
Quoted 2 lines

How did you discover this?

A syzbot report? If so please include Reported-by, Closes tags.

This wasn't reported by syzbot - it was found by our internal fuzzing CI on a KCSAN-enabled build of our downstream tree (based on v6.6.121), triggered by a multi-threaded mlockall() fuzz testcase.
There's no public bug id for it, so Reported-by/Closes don't apply here.

Best,
Hui

Quoted 58 lines
Quoted 23 lines

---
 include/linux/mm.h | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)

diff --git a/include/linux/mm.h b/include/linux/mm.h
index 485df9c2dbdd..122d3b39369f 100644
--- a/include/linux/mm.h
+++ b/include/linux/mm.h
@@ -2296,7 +2296,14 @@ static inline int memdesc_nid(memdesc_flags_t mdf)

static inline int page_to_nid(const struct page *page) {
- return memdesc_nid(PF_POISONED_CHECK(page)->flags);
+ /*
+ * The node id occupies a fixed high bit-range of page->flags
+ * that is set once at page init and never changed afterwards.
+ * It cannot overlap with the low PG_locked/PG_waiters bits
+ * that folio_lock()/folio_unlock() concurrently update, so
+ * this data race is benign.
+ */

Quoted 2 lines

Do we really need this excessive comment?

Agreed, just delete it. For a trivial benign data race it's a bit much, and the commit message can cover it off for those who are curious.

Quoted 7 lines

+ return memdesc_nid(data_race(PF_POISONED_CHECK(page)->flags));

In memdesc_zonenum() we use ASSERT_EXCLUSIVE_BITS.

Can we do the same here inside memdesc_nid?

Also agreed

Quoted 9 lines
include/linux/mm.h +0 -0
 index 69daeeab7fe8f..76d3bb54be844 100644
 --- a/include/linux/mm.h
 +++ b/include/linux/mm.h
 @@ -2290,6 +2290,7 @@ int memdesc_nid(memdesc_flags_t mdf);
 #else
 static inline int memdesc_nid(memdesc_flags_t mdf)

include/linux/mm.h

Quoted 10 lines
include/linux/mm.h +0 -0
 {
 + ASSERT_EXCLUSIVE_BITS(mdf.f, NODES_MASK << NODES_PGSHIFT);
 return (mdf.f >> NODES_PGSHIFT) & NODES_MASK;
 }
 #endif
 
 --
 Cheers,
 
 David

Thanks, Lorenzo