[PATCH 0/2] mm/mremap: fix two issues with MREMAP_DONTUNMAP

[PATCH 0/2] mm/mremap: fix two issues with MREMAP_DONTUNMAP

[Lorenzo Stoakes (ARM)]

The MREMAP_DONTUNMAP feature is highly unusual in that it permits mremap()
operations that keep the original VMA in place.

Historically this has led to a lot of bugs where non-obvious interactions
occur between existing mremap() operations and the original VMA.

Commit 397432cab17b ("mm/mremap: account mm->locked_vm correctly for
MREMAP_DONTUNMAP") fixed an accidentally introduced bug around
mm->locked_vm accounting, but this wasn't the only issue.

And thus history repeats itself, as it turns out that mm->locked_vm
accounting is broken by MREMAP_DONTUNMAP yet again by two further cases,
and has been broken ever since the feature was introduced.

Both relate to the fact that VMA_LOCKED_BIT is cleared on the source
VMA (it has to be as all page tables are moved):

1. If an unfaulted VMA_LOCKONFAULT_BIT anonymous VMA self-merges it
   clears the VMA_LOCKED_BIT flag and permanently leaks mm->locked_vm
   pages.

2. If a partial mremap() is performed on a locked VMA there is a leak equal
   to the number of pages not copied.

(Both for MREMAP_DONTUNMAP operations only)

Both issues can be fixed by treating the source range as distinct from the
destination range, which is the definition of what MREMAP_DONTUNMAP does so
is appropriate.

In case 1, simply disallow the self-merge, keeping adjacent source and
destination VMAs distinct.

In case 2, split the source range ahead of time, so accounting is always
correct.

Both changes were tested locally and confirmed to fix the issues.

For the purposes of a backport, the fixes are kept distinct, a follow-up
series can add self-tests.

Signed-off-by: Lorenzo Stoakes (ARM) <ljs@kernel.org>
---
Lorenzo Stoakes (ARM) (2):
      mm/mremap: fix locked_vm leak from MREMAP_DONTUNMAP self-merge
      mm/mremap: fix locked_vm leak by splitting VMA for MREMAP_DONTUNMAP

 mm/mremap.c                   | 87 +++++++++++++++++++++++++++----------------
 mm/vma.c                      | 19 +++++++++-
 mm/vma.h                      |  7 +++-
 tools/testing/vma/tests/vma.c | 10 ++---
 4 files changed, 83 insertions(+), 40 deletions(-)
---
base-commit: a3d0117e02bfa1c3bb6c50e7afe42bbecdbb3459
change-id: 20260920-fix-dontunmap-partial-self-merge-74a98b1d5a65

Best regards,
-- 
Lorenzo Stoakes (ARM) <ljs@kernel.org>

[PATCH 1/2] mm/mremap: fix locked_vm leak from MREMAP_DONTUNMAP self-merge

[Lorenzo Stoakes (ARM)]

The MREMAP_DONTUNMAP feature is highly unusual in that it permits mremap()
operations that keep the original VMA in place.

Historically this has led to a lot of bugs where non-obvious interactions
occur between existing mremap() operations and the original VMA.

Fix another of these - self-merge.

Self-merge occurs when a VMA is moved in front of or behind itself and the
attributes of the VMA permit such a merge.

Practically this can only happen for unfaulted anonymous VMAs due to the
page offset equality requirement for merge:

                           |------------|
                           |            |
                           |            v
        |...........||-----------||...........|
        |           || unfaulted ||           |
        |...........||-----------||...........|
              ^            |
              |            |
              |------------|

This becomes problematic if the VMA is configured by the user to
mlock-on-fault, i.e. the VMA_LOCKED_BIT, VMA_LOCKONFAULT_BIT VMA flags are
set.

MREMAP_DONTUNMAP clears mlock flags for the source VMA and maintains them
for the destination VMA.

Self-merge makes this impossible (there is only one VMA) and incorrectly
clears the destination VMA's mlock flags.

This causes a leak in mm->locked_vm as clearing this flag does not
decrement the counter and the VMA no longer has VMA_LOCKED_BIT set so it
is not decremented on unmap.

Resolve this by simply disallowing a self-merge in this case - the source
and destination VMAs are kept distinct and then are able to have distinct
mlock() flags.

Update dontunmap_complete() to make the now-redundant self-merge check a
VM_WARN_ON_ONCE() instead to guard against future regressions.

Also update the VMA userland tests to reflect the change.

Fixes: e346b3813067 ("mm/mremap: add MREMAP_DONTUNMAP to mremap()")
Cc: <stable@vger.kernel.org>
Signed-off-by: Lorenzo Stoakes (ARM) <ljs@kernel.org>
---
 mm/mremap.c                   |  8 ++++++--
 mm/vma.c                      | 17 ++++++++++++++++-
 mm/vma.h                      |  2 +-
 tools/testing/vma/tests/vma.c | 10 +++++-----
 4 files changed, 28 insertions(+), 9 deletions(-)

diff --git a/mm/mremap.c b/mm/mremap.c
index 1122282a1d6a..73f52c45705c 100644
--- a/mm/mremap.c
+++ b/mm/mremap.c
@@ -1275,7 +1275,8 @@ static int copy_vma_and_data(struct vma_remap_struct *vrm,
 	PAGETABLE_MOVE(pmc, NULL, NULL, vrm->addr, vrm->new_addr, vrm->old_len);
 
 	new_vma = copy_vma(&vma, vrm->new_addr, vrm->new_len, new_pgoff,
-			   new_anon_pgoff, &pmc.need_rmap_locks);
+			   new_anon_pgoff, &pmc.need_rmap_locks,
+			   vrm->flags & MREMAP_DONTUNMAP);
 	if (!new_vma) {
 		vrm_uncharge(vrm);
 		*new_vma_ptr = NULL;
@@ -1335,6 +1336,9 @@ static void dontunmap_complete(struct vma_remap_struct *vrm,
 	unsigned long old_start = vma->vm_start;
 	unsigned long old_end = vma->vm_end;
 
+	/* Self-merge is disallowed. */
+	VM_WARN_ON_ONCE(new_vma == vma);
+
 	/* We always clear VMA_LOCKED[ONFAULT]_BIT on the old VMA. */
 	vma_clear_flags_mask(vma, VMA_LOCKED_MASK);
 
@@ -1342,7 +1346,7 @@ static void dontunmap_complete(struct vma_remap_struct *vrm,
 	 * anon_vma links of the old vma is no longer needed after its page
 	 * table has been moved.
 	 */
-	if (new_vma != vma && start == old_start && end == old_end) {
+	if (start == old_start && end == old_end) {
 		const pgoff_t pgoff_unfaulted = vma->vm_start >> PAGE_SHIFT;
 
 		unlink_anon_vmas(vma);
diff --git a/mm/vma.c b/mm/vma.c
index 0db3f1222cad..55d4d0939129 100644
--- a/mm/vma.c
+++ b/mm/vma.c
@@ -1946,7 +1946,7 @@ static int vma_link(struct mm_struct *mm, struct vm_area_struct *vma)
  */
 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
 	unsigned long addr, unsigned long len, pgoff_t pgoff,
-	pgoff_t anon_pgoff, bool *need_rmap_locks)
+	pgoff_t anon_pgoff, bool *need_rmap_locks, bool keep_source)
 {
 	struct vm_area_struct *vma = *vmap;
 	unsigned long old_vma_start = vma->vm_start;
@@ -1984,6 +1984,21 @@ struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
 	vmg.pgoff = pgoff;
 	vmg.anon_pgoff = anon_pgoff;
 	vmg.next = vma_iter_next_rewind(&vmi, NULL);
+
+	/*
+	 * If the original VMA is kept (MREMAP_DONTUNMAP), the source and
+	 * destination VMA must be treated distinctly.
+	 *
+	 * A merge violates this, so in this case disallow a self-merge.
+	 */
+	if (can_self_merge && keep_source) {
+		if (vmg.prev == vma)
+			vmg.prev = NULL;
+		if (vmg.next == vma)
+			vmg.next = NULL;
+		can_self_merge = false;
+	}
+
 	new_vma = vma_merge_copied_range(&vmg);
 
 	if (new_vma) {
diff --git a/mm/vma.h b/mm/vma.h
index b2c3bc832a48..03ed8afd0c1f 100644
--- a/mm/vma.h
+++ b/mm/vma.h
@@ -534,7 +534,7 @@ void unlink_file_vma_batch_add(struct unlink_vma_file_batch *vb,
 
 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
 	unsigned long addr, unsigned long len, pgoff_t pgoff,
-	pgoff_t anon_pgoff, bool *need_rmap_locks);
+	pgoff_t anon_pgoff, bool *need_rmap_locks, bool keep_source);
 
 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma);
 
diff --git a/tools/testing/vma/tests/vma.c b/tools/testing/vma/tests/vma.c
index c8ef7b8cd46b..e973e0a6d1a8 100644
--- a/tools/testing/vma/tests/vma.c
+++ b/tools/testing/vma/tests/vma.c
@@ -40,7 +40,7 @@ static bool test_copy_vma(void)
 	vma = alloc_and_link_vma(&mm, 0x1000, 0x2000, 1, vma_flags);
 	vma_set_anonymous(vma);
 	vma_orig = vma;
-	vma_new = copy_vma(&vma, 0x2000, 0x1000, 1, 1, &need_locks);
+	vma_new = copy_vma(&vma, 0x2000, 0x1000, 1, 1, &need_locks, false);
 	ASSERT_EQ(vma_new, vma_orig);
 	ASSERT_EQ(vma, vma_orig);
 	ASSERT_EQ(vma_new->vm_start, 0x1000);
@@ -53,7 +53,7 @@ static bool test_copy_vma(void)
 	vma = alloc_and_link_vma(&mm, 0x2000, 0x3000, 2, vma_flags);
 	vma_set_anonymous(vma);
 	vma_orig = vma;
-	vma_new = copy_vma(&vma, 0x1000, 0x1000, 2, 2, &need_locks);
+	vma_new = copy_vma(&vma, 0x1000, 0x1000, 2, 2, &need_locks, false);
 	ASSERT_EQ(vma_new, vma_orig);
 	ASSERT_EQ(vma, vma_orig);
 	ASSERT_EQ(vma_new->vm_start, 0x1000);
@@ -71,7 +71,7 @@ static bool test_copy_vma(void)
 	vma = alloc_and_link_vma(&mm, 0x3000, 0x4000, 3, vma_flags);
 	vma_set_anonymous(vma);
 	vma_orig = vma;
-	vma_new = copy_vma(&vma, 0x2000, 0x1000, 3, 3, &need_locks);
+	vma_new = copy_vma(&vma, 0x2000, 0x1000, 3, 3, &need_locks, false);
 	ASSERT_NE(vma_new, vma_orig);
 	ASSERT_EQ(vma_new, vma);
 	ASSERT_EQ(vma_new->vm_start, 0x1000);
@@ -82,7 +82,7 @@ static bool test_copy_vma(void)
 	/* Move backwards and do not merge. */
 
 	vma = alloc_and_link_vma(&mm, 0x3000, 0x5000, 3, vma_flags);
-	vma_new = copy_vma(&vma, 0, 0x2000, 0, 3, &need_locks);
+	vma_new = copy_vma(&vma, 0, 0x2000, 0, 3, &need_locks, false);
 	ASSERT_NE(vma_new, vma);
 	ASSERT_EQ(vma_new->vm_start, 0);
 	ASSERT_EQ(vma_new->vm_end, 0x2000);
@@ -95,7 +95,7 @@ static bool test_copy_vma(void)
 
 	vma = alloc_and_link_vma(&mm, 0, 0x2000, 0, vma_flags);
 	vma_next = alloc_and_link_vma(&mm, 0x6000, 0x8000, 6, vma_flags);
-	vma_new = copy_vma(&vma, 0x4000, 0x2000, 4, 4, &need_locks);
+	vma_new = copy_vma(&vma, 0x4000, 0x2000, 4, 4, &need_locks, false);
 	vma_assert_attached(vma_new);
 
 	ASSERT_EQ(vma_new, vma_next);

-- 
2.55.0

[PATCH 2/2] mm/mremap: fix locked_vm leak by splitting VMA for MREMAP_DONTUNMAP

[Lorenzo Stoakes (ARM)]

The MREMAP_DONTUNMAP feature is highly unusual in that it permits mremap()
operations that keep the original VMA in place.

Historically this has led to a lot of bugs where non-obvious interactions
occur between existing mremap() operations and the original VMA.

Fix another of these - partial copies.

The long-standing mremap() partial VMA logic has the baked-in assumption
that the originating VMA is unmapped and thus moved.

However MREMAP_DONTUNMAP defeats this by performing a partial copy
instead since it keeps the source VMA around.

An mremap(..., MREMAP_DONTUNMAP) operation disallows resizing of the VMA,
but the operation can be performed partially:

                           |-----------------|
                           |                 |
                           |                 v
                        <------>          <------>
                        .new_sz.           new_sz
                     |--.------.--|       |------|
                     |  .source.  |       | dest |
                     |--.------.--|       |------|
                     <------------>
                         old_sz

The page tables in the specified range are moved, but the original VMA is
kept intact.

This interacts poorly with mlock()'d VMAs, as the VMA_LOCKED_BIT flag is
cleared for the entire source VMA and set for the entire destination VMA.

This results in an mm->locked_vm leak as the change is therefore not
accounted correctly.

The clear solution here is to make the portion of the source VMA which is
mremap()'d distinct from the rest of it, a.k.a. split it.

Therefore resolve this issue by splitting it ahead of the rest of the
mremap() operation.

In order to make this change re-expose split_vma() in vma.h for
CONFIG_MMU (nommu doesn't compile mremap.c and uses a static helper
instead).

A quick search of how MREMAP_DONTUNMAP is used in the wild suggests that
the partial case is either unused or rarely used, so this should not result
in unreasonable VMA proliferation.

Since this makes every mremap() MREMAP_DONTUNMAP operation operate across
an entire, distinct, VMA, also eliminate now-redundant code checking for
this in dontunmap_complete().

Finally, update the sys_map_count check to account for this case.

Fixes: e346b3813067 ("mm/mremap: add MREMAP_DONTUNMAP to mremap()")
Cc: <stable@vger.kernel.org>
Signed-off-by: Lorenzo Stoakes (ARM) <ljs@kernel.org>
---
 mm/mremap.c | 81 ++++++++++++++++++++++++++++++++++++++-----------------------
 mm/vma.c    |  2 +-
 mm/vma.h    |  5 ++++
 3 files changed, 56 insertions(+), 32 deletions(-)

diff --git a/mm/mremap.c b/mm/mremap.c
index 73f52c45705c..49dc25d8a34d 100644
--- a/mm/mremap.c
+++ b/mm/mremap.c
@@ -1037,6 +1037,7 @@ static void vrm_stat_account(struct vma_remap_struct *vrm,
 }
 
 static bool __check_map_count_against_split(struct mm_struct *mm,
+					    bool is_dontunmap,
 					    bool before_unmaps)
 {
 	const int sys_map_count = get_sysctl_max_map_count();
@@ -1088,26 +1089,38 @@ static bool __check_map_count_against_split(struct mm_struct *mm,
 	 * Therefore we must check to ensure we have headroom of 2 additional
 	 * VMAs.
 	 */
-	return map_count + 2 <= sys_map_count;
+	map_count += 2;
+
+	/*
+	 * If MREMAP_DONTUNMAP is set and a partial operation is performed,
+	 * the VMA is split ahead of time and the -1 observed above doesn't
+	 * apply.
+	 */
+	if (is_dontunmap)
+		map_count++;
+
+	return map_count <= sys_map_count;
 }
 
 /* Do we violate the map count limit if we split VMAs when moving the VMA? */
-static bool check_map_count_against_split(void)
+static bool check_map_count_against_split(struct vma_remap_struct *vrm)
 {
 	return __check_map_count_against_split(current->mm,
+					       vrm->flags & MREMAP_DONTUNMAP,
 					       /*before_unmaps=*/false);
 }
 
 /* Do we violate the map count limit if we split VMAs prior to early unmaps? */
-static bool check_map_count_against_split_early(void)
+static bool check_map_count_against_split_early(struct vma_remap_struct *vrm)
 {
 	return __check_map_count_against_split(current->mm,
+					       vrm->flags & MREMAP_DONTUNMAP,
 					       /*before_unmaps=*/true);
 }
 
 /*
- * Perform checks before attempting to write a VMA prior to it being
- * moved.
+ * Perform checks and preparation before attempting to write a VMA prior to it
+ * being moved.
  */
 static unsigned long prep_move_vma(struct vma_remap_struct *vrm)
 {
@@ -1116,19 +1129,17 @@ static unsigned long prep_move_vma(struct vma_remap_struct *vrm)
 	unsigned long old_addr = vrm->addr;
 	unsigned long old_len = vrm->old_len;
 	vm_flags_t dummy = vma->vm_flags;
+	const bool split_before = vma->vm_start != old_addr;
+	const bool split_after = vma->vm_end != old_addr + old_len;
 
-	/*
-	 * We'd prefer to avoid failure later on in do_munmap: we copy a VMA,
-	 * which may not merge, then (if MREMAP_DONTUNMAP is not set) unmap the
-	 * source, which may split, causing a net increase of 2 mappings.
-	 */
-	if (!check_map_count_against_split())
+	/* Avoid failure later on. */
+	if (!check_map_count_against_split(vrm))
 		return -ENOMEM;
 
 	if (vma->vm_ops && vma->vm_ops->may_split) {
-		if (vma->vm_start != old_addr)
+		if (split_before)
 			err = vma->vm_ops->may_split(vma, old_addr);
-		if (!err && vma->vm_end != old_addr + old_len)
+		if (!err && split_after)
 			err = vma->vm_ops->may_split(vma, old_addr + old_len);
 		if (err)
 			return err;
@@ -1146,6 +1157,21 @@ static unsigned long prep_move_vma(struct vma_remap_struct *vrm)
 	if (err)
 		return err;
 
+	/*
+	 * To account mlock()'d pages correctly in the MREMAP_DONTUNMAP
+	 * case perform any split ahead of time.
+	 */
+	if (vrm->flags & MREMAP_DONTUNMAP) {
+		VMA_ITERATOR(vmi, vma->vm_mm, old_addr);
+
+		if (split_before)
+			err = split_vma(&vmi, vma, old_addr, 1);
+		if (!err && split_after)
+			err = split_vma(&vmi, vma, old_addr + old_len, 0);
+		vrm->vmi_needs_invalidate = true;
+		return err;
+	}
+
 	return 0;
 }
 
@@ -1330,11 +1356,8 @@ static int copy_vma_and_data(struct vma_remap_struct *vrm,
 static void dontunmap_complete(struct vma_remap_struct *vrm,
 			       struct vm_area_struct *new_vma)
 {
-	unsigned long start = vrm->addr;
-	unsigned long end = vrm->addr + vrm->old_len;
 	struct vm_area_struct *vma = vrm->vma;
-	unsigned long old_start = vma->vm_start;
-	unsigned long old_end = vma->vm_end;
+	const pgoff_t pgoff_unfaulted = vma->vm_start >> PAGE_SHIFT;
 
 	/* Self-merge is disallowed. */
 	VM_WARN_ON_ONCE(new_vma == vma);
@@ -1346,19 +1369,15 @@ static void dontunmap_complete(struct vma_remap_struct *vrm,
 	 * anon_vma links of the old vma is no longer needed after its page
 	 * table has been moved.
 	 */
-	if (start == old_start && end == old_end) {
-		const pgoff_t pgoff_unfaulted = vma->vm_start >> PAGE_SHIFT;
-
-		unlink_anon_vmas(vma);
-		/*
-		 * The VMA is now unfaulted and it is an invariant that
-		 * unfaulted anonymous VMAs have page offset equal to
-		 * vma->vm_start >> PAGE_SHIFT.
-		 */
-		vma_set_anon_pgoff(vma, pgoff_unfaulted);
-		if (vma_is_anonymous(vma) && !vma->vm_file)
-			vma_set_pgoff(vma, pgoff_unfaulted);
-	}
+	unlink_anon_vmas(vma);
+	/*
+	 * The VMA is now unfaulted and it is an invariant that
+	 * unfaulted anonymous VMAs have page offset equal to
+	 * vma->vm_start >> PAGE_SHIFT.
+	 */
+	vma_set_anon_pgoff(vma, pgoff_unfaulted);
+	if (vma_is_anonymous(vma) && !vma->vm_file)
+		vma_set_pgoff(vma, pgoff_unfaulted);
 }
 
 static unsigned long move_vma(struct vma_remap_struct *vrm)
@@ -2005,7 +2024,7 @@ static unsigned long do_mremap(struct vma_remap_struct *vrm)
 		return -EINTR;
 	vrm->mmap_locked = true;
 
-	if (!check_map_count_against_split_early()) {
+	if (!check_map_count_against_split_early(vrm)) {
 		mmap_write_unlock(mm);
 		return -ENOMEM;
 	}
diff --git a/mm/vma.c b/mm/vma.c
index 55d4d0939129..92a2fafd66ae 100644
--- a/mm/vma.c
+++ b/mm/vma.c
@@ -634,7 +634,7 @@ __split_vma(struct vma_iterator *vmi, struct vm_area_struct *vma,
  * Split a vma into two pieces at address 'addr', a new vma is allocated
  * either for the first part or the tail.
  */
-static int split_vma(struct vma_iterator *vmi, struct vm_area_struct *vma,
+int split_vma(struct vma_iterator *vmi, struct vm_area_struct *vma,
 		     unsigned long addr, int new_below)
 {
 	if (vma->vm_mm->map_count >= get_sysctl_max_map_count())
diff --git a/mm/vma.h b/mm/vma.h
index 03ed8afd0c1f..b9b99fa02a86 100644
--- a/mm/vma.h
+++ b/mm/vma.h
@@ -555,6 +555,11 @@ int do_brk_flags(struct vma_iterator *vmi, struct vm_area_struct *brkvma,
 unsigned long unmapped_area(struct vm_unmapped_area_info *info);
 unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info);
 
+#ifdef CONFIG_MMU
+int split_vma(struct vma_iterator *vmi, struct vm_area_struct *vma,
+	      unsigned long addr, int new_below);
+#endif
+
 static inline bool vma_wants_manual_pte_write_upgrade(struct vm_area_struct *vma)
 {
 	/*

-- 
2.55.0

Re: [PATCH 0/2] mm/mremap: fix two issues with MREMAP_DONTUNMAP

[Andrew Morton]

On Sun, 20 Sep 2026 15:13:09 +0100 "Lorenzo Stoakes (ARM)" <ljs@kernel.org> wrote:

> The MREMAP_DONTUNMAP feature is highly unusual in that it permits mremap()
> operations that keep the original VMA in place.
> 
> Historically this has led to a lot of bugs where non-obvious interactions
> occur between existing mremap() operations and the original VMA.
> 
> ...
> 
> Both issues can be fixed by treating the source range as distinct from the
> destination range, which is the definition of what MREMAP_DONTUNMAP does so
> is appropriate.

Can we please have some review of these two cc:stable hotfixes?

Thanks.