[PATCH v4 07/20] mm/vma: fix self-merge check in copy_vma()

"Lorenzo Stoakes (ARM)" <[email protected]> Thu, 06 Aug 2026 21:21:32 +0100
Newsgroups org.kernel.vger.linux-trace-kernel,org.freedesktop.lists.amd-gfx,org.freedesktop.lists.dri-devel,org.freedesktop.lists.intel-xe,org.kernel.vger.kvm,org.kernel.vger.linux-fsdevel,org.kernel.vger.linux-kernel,org.kernel.vger.linux-kselftest,org.kernel.vger.linux-perf-users,org.kernel.vger.linux-s390,org.kvack.linux-mm
Message-ID <[email protected]>
The existing logic is very confusing so improve things. Firstly rename the
confusing faulted_in_anon_vma variable to can_self_merge and update this
when the page offset is updated.

What is being checked for is a 'self-merge' - that is between the VMA being
remapped and its prior VMA (remember that this is copy_vma() - if a
non-MREMAP_DONTUNMAP remap the original VMA is only removed afterwards).

This can happen if the VMA is moved immediately adjacent to
itself, either before or after it:

		|----------------|----------------|
		|		 |                |
		v		 |		  v
	|...............||---------------||...............|
	|      new      ||      old      ||      new      |
	|...............||---------------||---------------|

In these cases the old VMA is simply expanded to cover the new range.

It is also possible for the move to both self-merge and merge with a prior
VMA if it is placed between a preceding VMA and its old self:

				|---------------|
				|		|
				v		|
	|---------------||...............||---------------|
	|      prev     ||     new       ||     old       |
	|---------------||...............||---------------|

In this case, the old VMA is removed and 'prev' is expanded and replaces
it.

Since copy_vma_and_data() which calls copy_vma() intends to reference the
old VMA after the merge, it must have this pointer updated.

This kind of self-merge is not possible with a succeeding merge, as the
merge always prefers to expand the preceding VMA if possible.

copy_vma() accounts for this by explicitly checking to see if a self-merge
occurred and updating the vmap pointer if so. However it incorrect did so
even for a subsequent merge (this is simply a noop so it had no impact).

So change this to only check for the case which matters - a backwards
merge - and rearrange the parameters to make it clearer we're doing that -
i.e. check new_vma->vm_start < old_vma_start (having already renamed
vma_start to old_vma_start to make it clear this is the previous VMA).

Also update the existing wall-of-text comment to be a lot clearer.

While we're here, replace the VM_BUG_ON_VMA() with a VM_WARN_ON_ONCE_VMA()
and update the VMA userland tests accordingly.

No functional change intended.

Signed-off-by: Lorenzo Stoakes (ARM) <[email protected]>
---
 mm/vma.c                         | 35 ++++++++++++++++-------------------
 tools/testing/vma/vma_internal.h |  1 +
 2 files changed, 17 insertions(+), 19 deletions(-)

diff --git a/mm/vma.c b/mm/vma.c
index b5bc3eec961c..18c8f2546765 100644
--- a/mm/vma.c
+++ b/mm/vma.c
@@ -1911,10 +1911,10 @@ struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
 	bool *need_rmap_locks)
 {
 	struct vm_area_struct *vma = *vmap;
-	unsigned long vma_start = vma->vm_start;
+	unsigned long old_vma_start = vma->vm_start;
 	struct mm_struct *mm = vma->vm_mm;
 	struct vm_area_struct *new_vma;
-	bool faulted_in_anon_vma = true;
+	bool can_self_merge = false;
 	VMA_ITERATOR(vmi, mm, addr);
 	VMG_VMA_STATE(vmg, &vmi, NULL, vma, addr, addr + len);
 
@@ -1924,7 +1924,7 @@ struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
 	 */
 	if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
 		pgoff = addr >> PAGE_SHIFT;
-		faulted_in_anon_vma = false;
+		can_self_merge = true;
 	}
 
 	/*
@@ -1944,24 +1944,21 @@ struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
 	new_vma = vma_merge_copied_range(&vmg);
 
 	if (new_vma) {
-		/*
-		 * Source vma may have been merged into new_vma
-		 */
-		if (unlikely(vma_start >= new_vma->vm_start &&
-			     vma_start < new_vma->vm_end)) {
+		/* Self-merged and VMA replaced. */
+		if (unlikely(new_vma->vm_start < old_vma_start &&
+			     new_vma->vm_end > old_vma_start)) {
 			/*
-			 * The only way we can get a vma_merge with
-			 * self during an mremap is if the vma hasn't
-			 * been faulted in yet and we were allowed to
-			 * reset the dst vma->vm_pgoff to the
-			 * destination address of the mremap to allow
-			 * the merge to happen. mremap must change the
-			 * vm_pgoff linearity between src and dst vmas
-			 * (in turn preventing a vma_merge) to be
-			 * safe. It is only safe to keep the vm_pgoff
-			 * linear if there are no pages mapped yet.
+			 * The only way a VMA can both self-merge and be
+			 * replaced is if the remap places the new VMA
+			 * immediately prior to its old self ('next') and
+			 * immediately after another VMA ('prev') causing the
+			 * next to be removed and prev to be expanded to cover
+			 * the entire range.
+			 *
+			 * This should only be possible if the page offset was
+			 * updated, i.e. the VMA is unfaulted.
 			 */
-			VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
+			VM_WARN_ON_ONCE_VMA(!can_self_merge, new_vma);
 			*vmap = vma = new_vma;
 		}
 		*need_rmap_locks =
diff --git a/tools/testing/vma/vma_internal.h b/tools/testing/vma/vma_internal.h
index 4f6c5666ac07..8a48b231aa7a 100644
--- a/tools/testing/vma/vma_internal.h
+++ b/tools/testing/vma/vma_internal.h
@@ -53,6 +53,7 @@ typedef __bitwise unsigned int vm_fault_t;
 
 #define VM_WARN_ON(_expr) (WARN_ON(_expr))
 #define VM_WARN_ON_ONCE(_expr) (WARN_ON_ONCE(_expr))
+#define VM_WARN_ON_ONCE_VMA(_expr, _vma) (WARN_ON_ONCE(_expr))
 #define VM_WARN_ON_VMG(_expr, _vmg) (WARN_ON(_expr))
 #define VM_BUG_ON(_expr) (BUG_ON(_expr))
 #define VM_BUG_ON_VMA(_expr, _vma) (BUG_ON(_expr))

-- 
2.55.0