Skip to contentSkip to conversation

mm/vmalloc: fix vmalloc_dump_obj VA lookup

PATCH Superseded Ye Liu 2 participants 4 messages Sep 24 – Sep 25

Subsystems (3)
Memory management RISC-V vmalloc

v3 of 4 Superseded Latest is v4 →

lore
Share & tools
lore

Ye Liu

author
#0 lore raw

v3 cover mm/vmalloc: fix vmalloc_dump_obj VA lookup

vmalloc_dump_obj() has two bugs that cause it to miss vmalloc allocations.

1. PAGE_ALIGN() rounds up, pushing last-page addresses to va_end and outside the VA lookup range.

2. The function searches only one vmap node, but allocations larger than 64 KiB may span multiple vmap zones whose VA is stored in a different node's rb-tree.

Patch 1 fixes the alignment, patch 2 fixes the cross-zone search.

Signed-off-by Ye Liu <[email protected]>

---
Changes in v3:
- Patch3 drop.
- Patch2 Extract find_vmap_area_lock as a helper to simplify the code.
- Link to v2: https://lore.kernel.org/r/[email protected]

Changes in v2:
- Use is_vmalloc_or_module_addr() instead of is_vmalloc_addr() to
avoid filtering module/BPF/execmem addresses. (sashiko-bot)
- Link to v1: https://lore.kernel.org/r/[email protected]

---
Ye Liu (2):
mm/vmalloc: fix vmalloc_dump_obj address alignment for last-page lookups mm/vmalloc: fix vmalloc_dump_obj cross-zone VA lookup

 mm/vmalloc.c | 113 ++++++++++++++++++++++++++++++++++++-----------------------
 1 file changed, 70 insertions(+), 43 deletions(-)

---
base-commit: e6e35979777d646fe3c7c94dca7dd32fb25d45f4
change-id: 20260916-vmalloc_dump_obj-d80908447369

Best regards,

Signature · 1 line
Ye Liu <[email protected]>
author
#1 lore raw

PATCH v3 1/2 mm/vmalloc: fix vmalloc_dump_obj address alignment for last-page lookups

From: Ye Liu <[email protected]>

vmalloc_dump_obj() uses PAGE_ALIGN() to normalize the input address before looking it up in the per-node busy tree. PAGE_ALIGN() rounds up, which can push an address in the last page of a vmalloc allocation to va_end -- outside the [va_start, va_end) range that __find_vmap_area() searches. This causes the lookup to miss the VA and return false, degrading diagnostic output in OOM dumps and KASAN reports to the less informative "vmalloc memory" fallback.

The upward alignment can also change the addr_to_node() mapping when the page boundary crosses a vmap zone boundary, causing the search to hit the wrong node entirely.

Use PAGE_ALIGN_DOWN() instead, which rounds down to the page containing the address. This keeps the address within the VA range and preserves the correct node mapping.

Signed-off-by Ye Liu <[email protected]>
Reviewed-by Uladzislau Rezki (Sony) <[email protected]>
---
 mm/vmalloc.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
mm/vmalloc.c +1 -1
index 859e6d2d57a3..df42d8a6f058 100644
@@ -5277,7 +5277,7 @@ bool vmalloc_dump_obj(void *object)
 	unsigned long addr;
 	unsigned long nr_pages;
 
-	addr = PAGE_ALIGN((unsigned long) object);
+	addr = PAGE_ALIGN_DOWN((unsigned long) object);
 	vn = addr_to_node(addr);
 
 	if (!spin_trylock(&vn->busy.lock))
Signature · 1 line
2.25.1
author
#2 lore raw

PATCH v3 2/2 mm/vmalloc: fix vmalloc_dump_obj cross-zone VA lookup

In reply to Ye Liu

From: Ye Liu <[email protected]>

vmalloc_dump_obj() searches only one vmap node (addr_to_node(addr)), but a vmalloc allocation may span multiple vmap zones. The VA is stored in only one node's rb-tree (addr_to_node(va_start)), so an object pointer in a different zone than va_start maps to a different node and the search misses. This affects any allocation larger than vmap_zone_size (64 KiB) on multi-CPU systems.

Extract find_vmap_area_lock() from find_vmap_area() to share the cross-node iteration logic. The helper supports both spin_lock and spin_trylock, the latter for atomic dump contexts (OOM, KASAN, RCU).

Signed-off-by Ye Liu <[email protected]>
---
 mm/vmalloc.c | 111 +++++++++++++++++++++++++++++++++++++----------------------
 1 file changed, 69 insertions(+), 42 deletions(-)
mm/vmalloc.c +69 -42
index df42d8a6f058..e5b465de1559 100644
@@ -2517,39 +2517,81 @@ static void free_unmap_vmap_area(struct vmap_area *va)
 	free_vmap_area_noflush(va);
 }
 
-struct vmap_area *find_vmap_area(unsigned long addr)
+static inline int next_vmap_node_id(int i)
+{
+	return (i + nr_vmap_nodes - 1) % nr_vmap_nodes;
+}
+
+enum vmap_lock_mode {
+	VMAP_LOCK,
+	VMAP_TRYLOCK,
+};
+
+/*
+ * Search for a vmap_area at @addr across all vmap nodes.  An
+ * addr_to_node_id(addr) converts an address to a node index where
+ * a VA is located. If VA spans several zones and passed addr is not
+ * the same as va->va_start, what is not common, we may need to scan
+ * extra nodes. See an example:
+ *
+ *      <----va---->
+ * -|-----|-----|-----|-----|-
+ *     1     2     0     1
+ *
+ * VA resides in node 1 whereas it spans 1, 2 an 0. If passed addr
+ * is within 2 or 0 nodes we should do extra work.
+ *
+ * Returns the VA with @locked_vn->busy.lock held; the caller must
+ * release it. If @mode is VMAP_TRYLOCK, nodes that cannot be locked
+ * are skipped.
+ */
+static struct vmap_area *
+find_vmap_area_lock(unsigned long addr, struct vmap_node **locked_vn,
+		enum vmap_lock_mode mode)
 {
 	struct vmap_node *vn;
 	struct vmap_area *va;
 	int i, j;
 
-	if (unlikely(!vmap_initialized))
+	if (unlikely(!vmap_initialized)) {
+		*locked_vn = NULL;
 		return NULL;
+	}
 
-	/*
-	 * An addr_to_node_id(addr) converts an address to a node index
-	 * where a VA is located. If VA spans several zones and passed
-	 * addr is not the same as va->va_start, what is not common, we
-	 * may need to scan extra nodes. See an example:
-	 *
-	 *      <----va---->
-	 * -|-----|-----|-----|-----|-
-	 *     1     2     0     1
-	 *
-	 * VA resides in node 1 whereas it spans 1, 2 an 0. If passed
-	 * addr is within 2 or 0 nodes we should do extra work.
-	 */
 	i = j = addr_to_node_id(addr);
 	do {
 		vn = &vmap_nodes[i];
 
-		spin_lock(&vn->busy.lock);
+		if (mode == VMAP_LOCK) {
+			spin_lock(&vn->busy.lock);
+		} else {
+			if (!spin_trylock(&vn->busy.lock))
+				continue;
+		}
+
 		va = __find_vmap_area(addr, &vn->busy.root);
+		if (va) {
+			*locked_vn = vn;
+			return va;
+		}
+
 		spin_unlock(&vn->busy.lock);
+	} while ((i = next_vmap_node_id(i)) != j);
 
-		if (va)
-			return va;
-	} while ((i = (i + nr_vmap_nodes - 1) % nr_vmap_nodes) != j);
+	*locked_vn = NULL;
+	return NULL;
+}
+
+struct vmap_area *find_vmap_area(unsigned long addr)
+{
+	struct vmap_node *vn;
+	struct vmap_area *va;
+
+	va = find_vmap_area_lock(addr, &vn, VMAP_LOCK);
+	if (va) {
+		spin_unlock(&vn->busy.lock);
+		return va;
+	}
 
 	return NULL;
 }

@@ -2558,26 +2600,14 @@ static struct vmap_area *find_unlink_vmap_area(unsigned long addr)
 {
 	struct vmap_node *vn;
 	struct vmap_area *va;
-	int i, j;
-
-	/*
-	 * Check the comment in the find_vmap_area() about the loop.
-	 */
-	i = j = addr_to_node_id(addr);
-	do {
-		vn = &vmap_nodes[i];
 
-		spin_lock(&vn->busy.lock);
-		va = __find_vmap_area(addr, &vn->busy.root);
-		if (va)
-			unlink_va(va, &vn->busy.root);
+	va = find_vmap_area_lock(addr, &vn, VMAP_LOCK);
+	if (va) {
+		unlink_va(va, &vn->busy.root);
 		spin_unlock(&vn->busy.lock);
+	}
 
-		if (va)
-			return va;
-	} while ((i = (i + nr_vmap_nodes - 1) % nr_vmap_nodes) != j);
-
-	return NULL;
+	return va;
 }
 
 /*** Per cpu kva allocator ***/

@@ -5278,14 +5308,11 @@ bool vmalloc_dump_obj(void *object)
 	unsigned long nr_pages;
 
 	addr = PAGE_ALIGN_DOWN((unsigned long) object);
-	vn = addr_to_node(addr);
-
-	if (!spin_trylock(&vn->busy.lock))
-		return false;
 
-	va = __find_vmap_area(addr, &vn->busy.root);
+	va = find_vmap_area_lock(addr, &vn, VMAP_TRYLOCK);
 	if (!va || !va->vm) {
-		spin_unlock(&vn->busy.lock);
+		if (va)
+			spin_unlock(&vn->busy.lock);
 		return false;
 	}
Signature · 1 line
2.25.1
maintainer
#3 lore raw

Re: [PATCH v3 2/2] mm/vmalloc: fix vmalloc_dump_obj cross-zone VA lookup

Quoted 57 lines · Ye Liu

From: Ye Liu <[email protected]>

vmalloc_dump_obj() searches only one vmap node (addr_to_node(addr)), but a vmalloc allocation may span multiple vmap zones. The VA is stored in only one node's rb-tree (addr_to_node(va_start)), so an object pointer in a different zone than va_start maps to a different node and the search misses. This affects any allocation larger than vmap_zone_size (64 KiB) on multi-CPU systems.

Extract find_vmap_area_lock() from find_vmap_area() to share the cross-node iteration logic. The helper supports both spin_lock and spin_trylock, the latter for atomic dump contexts (OOM, KASAN, RCU).

Signed-off-by Ye Liu <[email protected]>
---
 mm/vmalloc.c | 111 +++++++++++++++++++++++++++++++++++++----------------------
 1 file changed, 69 insertions(+), 42 deletions(-)
mm/vmalloc.c +30 -1
index df42d8a6f058..e5b465de1559 100644
@@ -2517,39 +2517,81 @@ static void free_unmap_vmap_area(struct vmap_area *va)
 	free_vmap_area_noflush(va);
 }
 
-struct vmap_area *find_vmap_area(unsigned long addr)
+static inline int next_vmap_node_id(int i)
+{
+	return (i + nr_vmap_nodes - 1) % nr_vmap_nodes;
+}
+
+enum vmap_lock_mode {
+	VMAP_LOCK,
+	VMAP_TRYLOCK,
+};
+
+/*
+ * Search for a vmap_area at @addr across all vmap nodes.  An
+ * addr_to_node_id(addr) converts an address to a node index where
+ * a VA is located. If VA spans several zones and passed addr is not
+ * the same as va->va_start, what is not common, we may need to scan
+ * extra nodes. See an example:
+ *
+ *      <----va---->
+ * -|-----|-----|-----|-----|-
+ *     1     2     0     1
+ *
+ * VA resides in node 1 whereas it spans 1, 2 an 0. If passed addr
+ * is within 2 or 0 nodes we should do extra work.
+ *
+ * Returns the VA with @locked_vn->busy.lock held; the caller must
+ * release it. If @mode is VMAP_TRYLOCK, nodes that cannot be locked
+ * are skipped.
+ */
+static struct vmap_area *
+find_vmap_area_lock(unsigned long addr, struct vmap_node **locked_vn,

mm/vmalloc.c static void free_unmap_vmap_area(struct vmap_area *va) · line 2558

Quoted 10 lines
mm/vmalloc.c +3 -1
+		enum vmap_lock_mode mode)
 {
 	struct vmap_node *vn;
 	struct vmap_area *va;
 	int i, j;
 
-	if (unlikely(!vmap_initialized))
+	if (unlikely(!vmap_initialized)) {
+		*locked_vn = NULL;

Just set it to NULL once on entry?

Quoted 45 lines
mm/vmalloc.c +19 -17
 		return NULL;
+	}
 
-	/*
-	 * An addr_to_node_id(addr) converts an address to a node index
-	 * where a VA is located. If VA spans several zones and passed
-	 * addr is not the same as va->va_start, what is not common, we
-	 * may need to scan extra nodes. See an example:
-	 *
-	 *      <----va---->
-	 * -|-----|-----|-----|-----|-
-	 *     1     2     0     1
-	 *
-	 * VA resides in node 1 whereas it spans 1, 2 an 0. If passed
-	 * addr is within 2 or 0 nodes we should do extra work.
-	 */
 	i = j = addr_to_node_id(addr);
 	do {
 		vn = &vmap_nodes[i];
 
-		spin_lock(&vn->busy.lock);
+		if (mode == VMAP_LOCK) {
+			spin_lock(&vn->busy.lock);
+		} else {
+			if (!spin_trylock(&vn->busy.lock))
+				continue;
+		}
+
 		va = __find_vmap_area(addr, &vn->busy.root);
+		if (va) {
+			*locked_vn = vn;
+			return va;
+		}
+
 		spin_unlock(&vn->busy.lock);
+	} while ((i = next_vmap_node_id(i)) != j);
 
-		if (va)
-			return va;
-	} while ((i = (i + nr_vmap_nodes - 1) % nr_vmap_nodes) != j);
+	*locked_vn = NULL;
+	return NULL;
+}
+
+struct vmap_area *find_vmap_area(unsigned long addr)

mm/vmalloc.c static void free_unmap_vmap_area(struct vmap_area *va) · line 2595

Quoted 10 lines
mm/vmalloc.c +9 -0
+{
+	struct vmap_node *vn;
+	struct vmap_area *va;
+
+	va = find_vmap_area_lock(addr, &vn, VMAP_LOCK);
+	if (va) {
+		spin_unlock(&vn->busy.lock);
+		return va;
+	}

Can we simplify like?
...

	va = find_vmap_area_lock(addr, &vn, VMAP_LOCK);
	if (va)
		spin_unlock(&vn->busy.lock);

	return va;

...

Thanks!

--
Uladzislau Rezki