Skip to contentSkip to conversation

mm: madvise: use per_vma lock for MADV_FREE

PATCH Applied Barry Song 2 participants 2 messages Jun 11, 2025

Subsystems (3)
lore
Share & tools
lore

Barry Song

author
#0 lore raw

PATCH v2 1/1 mm: madvise: use per_vma lock for MADV_FREE

From: Barry Song <[email protected]>

MADV_FREE is another option, besides MADV_DONTNEED, for dynamic memory freeing in user-space native or Java heap memory management. For example, jemalloc can be configured to use MADV_FREE, and recent versions of the Android Java heap have also increasingly adopted MADV_FREE. Supporting per-VMA locking for MADV_FREE thus appears increasingly necessary.

We have replaced walk_page_range() with walk_page_range_vma(). Along with the proposed madvise_lock_mode by Lorenzo, the necessary infrastructure is now in place to begin exploring per-VMA locking support for MADV_FREE and potentially other madvise using walk_page_range_vma().

This patch adds support for the PGWALK_VMA_RDLOCK walk_lock mode in walk_page_range_vma(), and leverages madvise_lock_mode from madv_behavior to select the appropriate walk_lock—either mmap_lock or per-VMA lock—based on the context.

Because we now dynamically update the walk_ops->walk_lock field, we must ensure this is thread-safe. The madvise_free_walk_ops is now defined as a stack variable instead of a global constant.

Cc "Liam R. Howlett" <[email protected]>
Cc Vlastimil Babka <[email protected]>
Cc Jann Horn <[email protected]>
Cc Suren Baghdasaryan <[email protected]>
Cc Lokesh Gidra <[email protected]>
Cc Mike Rapoport <[email protected]>
Cc Michal Hocko <[email protected]>
Cc Tangquan Zheng <[email protected]>
Cc Qi Zheng <[email protected]>

---
-v2:
* Collect David's acked-by and Lorenzo's reviewed-by;
* refine changelog and code cleanup according to David and Lorenzo

 include/linux/pagewalk.h |  2 ++
 mm/madvise.c             | 25 +++++++++++++++++++------
 mm/pagewalk.c            |  5 ++++-
 3 files changed, 25 insertions(+), 7 deletions(-)
include/linux/pagewalk.h +2 -0
index 9700a29f8afb..a4afa64ef0ab 100644
@@ -14,6 +14,8 @@ enum page_walk_lock {
 	PGWALK_WRLOCK = 1,
 	/* vma is expected to be already write-locked during the walk */
 	PGWALK_WRLOCK_VERIFY = 2,
+	/* vma is expected to be already read-locked during the walk */
+	PGWALK_VMA_RDLOCK_VERIFY = 3,
 };
 
 /**
mm/madvise.c +19 -6
index 790c238d04d4..267d8e4adf31 100644
@@ -777,10 +777,19 @@ static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
 	return 0;
 }
 
-static const struct mm_walk_ops madvise_free_walk_ops = {
-	.pmd_entry		= madvise_free_pte_range,
-	.walk_lock		= PGWALK_RDLOCK,
-};
+static inline enum page_walk_lock get_walk_lock(enum madvise_lock_mode mode)
+{
+	switch (mode) {
+	case MADVISE_VMA_READ_LOCK:
+		return PGWALK_VMA_RDLOCK_VERIFY;
+	case MADVISE_MMAP_READ_LOCK:
+		return PGWALK_RDLOCK;
+	default:
+		/* Other modes don't require fixing up the walk_lock */
+		WARN_ON_ONCE(1);
+		return PGWALK_RDLOCK;
+	}
+}
 
 static int madvise_free_single_vma(struct madvise_behavior *madv_behavior,
 			struct vm_area_struct *vma,
@@ -789,6 +798,9 @@ static int madvise_free_single_vma(struct madvise_behavior *madv_behavior,
 	struct mm_struct *mm = vma->vm_mm;
 	struct mmu_notifier_range range;
 	struct mmu_gather *tlb = madv_behavior->tlb;
+	struct mm_walk_ops walk_ops = {
+		.pmd_entry		= madvise_free_pte_range,
+	};
 
 	/* MADV_FREE works for only anon vma at the moment */
 	if (!vma_is_anonymous(vma))
@@ -808,8 +820,9 @@ static int madvise_free_single_vma(struct madvise_behavior *madv_behavior,
 
 	mmu_notifier_invalidate_range_start(&range);
 	tlb_start_vma(tlb, vma);
+	walk_ops.walk_lock = get_walk_lock(madv_behavior->lock_mode);
 	walk_page_range_vma(vma, range.start, range.end,
-			&madvise_free_walk_ops, tlb);
+			&walk_ops, tlb);
 	tlb_end_vma(tlb, vma);
 	mmu_notifier_invalidate_range_end(&range);
 	return 0;
@@ -1655,7 +1668,6 @@ static enum madvise_lock_mode get_lock_mode(struct madvise_behavior *madv_behavi
 	case MADV_WILLNEED:
 	case MADV_COLD:
 	case MADV_PAGEOUT:
-	case MADV_FREE:
 	case MADV_POPULATE_READ:
 	case MADV_POPULATE_WRITE:
 	case MADV_COLLAPSE:
@@ -1664,6 +1676,7 @@ static enum madvise_lock_mode get_lock_mode(struct madvise_behavior *madv_behavi
 		return MADVISE_MMAP_READ_LOCK;
 	case MADV_DONTNEED:
 	case MADV_DONTNEED_LOCKED:
+	case MADV_FREE:
 		return MADVISE_VMA_READ_LOCK;
 	default:
 		return MADVISE_MMAP_WRITE_LOCK;
mm/pagewalk.c +4 -1
index e478777c86e1..74f623159f7b 100644
@@ -422,7 +422,7 @@ static inline void process_mm_walk_lock(struct mm_struct *mm,
 {
 	if (walk_lock == PGWALK_RDLOCK)
 		mmap_assert_locked(mm);
-	else
+	else if (walk_lock != PGWALK_VMA_RDLOCK_VERIFY)
 		mmap_assert_write_locked(mm);
 }
 
@@ -437,6 +437,9 @@ static inline void process_vma_walk_lock(struct vm_area_struct *vma,
 	case PGWALK_WRLOCK_VERIFY:
 		vma_assert_write_locked(vma);
 		break;
+	case PGWALK_VMA_RDLOCK_VERIFY:
+		vma_assert_locked(vma);
+		break;
 	case PGWALK_RDLOCK:
 		/* PGWALK_RDLOCK is handled by process_mm_walk_lock */
 		break;
Signature · 1 line
2.39.3 (Apple Git-146)
Acked-by on v2 1/1
#1 lore raw
Quoted 22 lines · On Wed, 11 Jun 2025 22:47:45 +1200 Barry Song <[email protected]> wrote:

From: Barry Song <[email protected]>

MADV_FREE is another option, besides MADV_DONTNEED, for dynamic memory freeing in user-space native or Java heap memory management. For example, jemalloc can be configured to use MADV_FREE, and recent versions of the Android Java heap have also increasingly adopted MADV_FREE. Supporting per-VMA locking for MADV_FREE thus appears increasingly necessary.

We have replaced walk_page_range() with walk_page_range_vma(). Along with the proposed madvise_lock_mode by Lorenzo, the necessary infrastructure is now in place to begin exploring per-VMA locking support for MADV_FREE and potentially other madvise using walk_page_range_vma().

This patch adds support for the PGWALK_VMA_RDLOCK walk_lock mode in walk_page_range_vma(), and leverages madvise_lock_mode from madv_behavior to select the appropriate walk_lock—either mmap_lock or per-VMA lock—based on the context.

Because we now dynamically update the walk_ops->walk_lock field, we must ensure this is thread-safe. The madvise_free_walk_ops is now defined as a stack variable instead of a global constant.

Quoted 12 lines
Cc "Liam R. Howlett" <[email protected]>
Cc Vlastimil Babka <[email protected]>
Cc Jann Horn <[email protected]>
Cc Suren Baghdasaryan <[email protected]>
Cc Lokesh Gidra <[email protected]>
Cc Mike Rapoport <[email protected]>
Cc Michal Hocko <[email protected]>
Cc Tangquan Zheng <[email protected]>
Cc Qi Zheng <[email protected]>

Thanks,
SJ

[...]