Skip to contentSkip to conversation

btrfs: reject names longer than BTRFS_NAME_LEN in btrfs_get_name()

PATCH In review Weiming Shi 3 participants 4 messages Jun 7 – Jun 8 Btrfs file system

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 btrfs: reject names longer than BTRFS_NAME_LEN in btrfs_get_name()

btrfs_get_name() reads the name length straight from the on-disk inode_ref (or root_ref) and copies that many bytes into the caller's buffer with no upper bound. The caller (exportfs_get_name()) supplies a fixed NAME_MAX + 1 byte stack buffer, but name_len is a __le16 read from the leaf and the tree-checker only bounds it to the item size, not to BTRFS_NAME_LEN. A crafted leaf with name_len = 4096 therefore overflows the 256-byte buffer with attacker-controlled bytes. It is reachable from a mounted untrusted btrfs image via open_by_handle_at(), and on btrfs filesystems exported over NFS.

 BUG: KASAN: stack-out-of-bounds in read_extent_buffer (fs/btrfs/extent_io.c:3742)
 Write of size 633 at addr ffffc90006c9fc40 by task exploit/5192
  read_extent_buffer (fs/btrfs/extent_io.c:3742)
  btrfs_get_name (fs/btrfs/export.c:289)
  reconnect_path (fs/exportfs/expfs.c:222)
  exportfs_decode_fh_raw (fs/exportfs/expfs.c:473)
  do_handle_open (fs/fhandle.c:230)
  do_syscall_64 (arch/x86/entry/syscall_64.c:94)
 Kernel panic - not syncing: stack-protector: Kernel stack is corrupted

Reject any name_len greater than BTRFS_NAME_LEN before the copy. Such a name is never valid on disk, so this only rejects corrupted leaves and leaves valid names unchanged.

Fixes 2ede0daf0154 ("Btrfs: handle NFS lookups properly")
Assisted-by Claude:claude-opus-4-8
---
 fs/btrfs/export.c | 10 ++++++++++
 1 file changed, 10 insertions(+)
fs/btrfs/export.c +10 -0
index c403117ac..a54c6e8b3 100644
@@ -285,6 +285,16 @@ static int btrfs_get_name(struct dentry *parent, char *name,
 		name_len = btrfs_inode_ref_name_len(leaf, iref);
 	}
 
+	/* The caller's buffer is only NAME_MAX + 1 bytes. */
+	if (name_len > BTRFS_NAME_LEN) {
+		btrfs_err(fs_info,
+			  "corrupt name length %d for inode %llu in root %llu, max %d",
+			  name_len, btrfs_ino(BTRFS_I(inode)),
+			  btrfs_root_id(BTRFS_I(inode)->root), BTRFS_NAME_LEN);
+		btrfs_free_path(path);
+		return -EUCLEAN;
+	}
+
 	read_extent_buffer(leaf, name, name_ptr, name_len);
 	btrfs_free_path(path);
 
Signature · 1 line
2.43.0
#1 lore raw

Hi,

Weiming Shi @ 2026-06-07 05:14 -07:

Quoted 37 lines

btrfs_get_name() reads the name length straight from the on-disk inode_ref (or root_ref) and copies that many bytes into the caller's buffer with no upper bound. The caller (exportfs_get_name()) supplies a fixed NAME_MAX + 1 byte stack buffer, but name_len is a __le16 read from the leaf and the tree-checker only bounds it to the item size, not to BTRFS_NAME_LEN. A crafted leaf with name_len = 4096 therefore overflows the 256-byte buffer with attacker-controlled bytes. It is reachable from a mounted untrusted btrfs image via open_by_handle_at(), and on btrfs filesystems exported over NFS.

 BUG: KASAN: stack-out-of-bounds in read_extent_buffer (fs/btrfs/extent_io.c:3742)
 Write of size 633 at addr ffffc90006c9fc40 by task exploit/5192
  read_extent_buffer (fs/btrfs/extent_io.c:3742)
  btrfs_get_name (fs/btrfs/export.c:289)
  reconnect_path (fs/exportfs/expfs.c:222)
  exportfs_decode_fh_raw (fs/exportfs/expfs.c:473)
  do_handle_open (fs/fhandle.c:230)
  do_syscall_64 (arch/x86/entry/syscall_64.c:94)
 Kernel panic - not syncing: stack-protector: Kernel stack is corrupted

Reject any name_len greater than BTRFS_NAME_LEN before the copy. Such a name is never valid on disk, so this only rejects corrupted leaves and leaves valid names unchanged.

Fixes 2ede0daf0154 ("Btrfs: handle NFS lookups properly")
Assisted-by Claude:claude-opus-4-8
---
 fs/btrfs/export.c | 10 ++++++++++
 1 file changed, 10 insertions(+)
fs/btrfs/export.c +0 -0
index c403117ac..a54c6e8b3 100644

fs/btrfs/export.c static int btrfs_get_name(struct dentry *parent, char *name, · line 294

Quoted 10 lines
fs/btrfs/export.c +7 -0
@@ -285,6 +285,16 @@ static int btrfs_get_name(struct dentry *parent, char *name,
 		name_len = btrfs_inode_ref_name_len(leaf, iref);
 	}
 
+	/* The caller's buffer is only NAME_MAX + 1 bytes. */
+	if (name_len > BTRFS_NAME_LEN) {
+		btrfs_err(fs_info,
+			  "corrupt name length %d for inode %llu in root %llu, max %d",
+			  name_len, btrfs_ino(BTRFS_I(inode)),
+			  btrfs_root_id(BTRFS_I(inode)->root), BTRFS_NAME_LEN);
+		btrfs_free_path(path);

Nitpicking: this is not necessary as 'path' is declared with a cleanup attribute.

Quoted 5 lines
fs/btrfs/export.c +3 -0
+		return -EUCLEAN;
+	}
+
 	read_extent_buffer(leaf, name, name_ptr, name_len);
 	btrfs_free_path(path);
#2 lore raw
In reply to Weiming Shi
Quoted 37 lines · Weiming Shi

btrfs_get_name() reads the name length straight from the on-disk inode_ref (or root_ref) and copies that many bytes into the caller's buffer with no upper bound. The caller (exportfs_get_name()) supplies a fixed NAME_MAX + 1 byte stack buffer, but name_len is a __le16 read from the leaf and the tree-checker only bounds it to the item size, not to BTRFS_NAME_LEN. A crafted leaf with name_len = 4096 therefore overflows the 256-byte buffer with attacker-controlled bytes. It is reachable from a mounted untrusted btrfs image via open_by_handle_at(), and on btrfs filesystems exported over NFS.

  BUG: KASAN: stack-out-of-bounds in read_extent_buffer (fs/btrfs/extent_io.c:3742)
  Write of size 633 at addr ffffc90006c9fc40 by task exploit/5192
   read_extent_buffer (fs/btrfs/extent_io.c:3742)
   btrfs_get_name (fs/btrfs/export.c:289)
   reconnect_path (fs/exportfs/expfs.c:222)
   exportfs_decode_fh_raw (fs/exportfs/expfs.c:473)
   do_handle_open (fs/fhandle.c:230)
   do_syscall_64 (arch/x86/entry/syscall_64.c:94)
  Kernel panic - not syncing: stack-protector: Kernel stack is corrupted

Reject any name_len greater than BTRFS_NAME_LEN before the copy. Such a name is never valid on disk, so this only rejects corrupted leaves and leaves valid names unchanged.

Fixes 2ede0daf0154 ("Btrfs: handle NFS lookups properly")
Assisted-by Claude:claude-opus-4-8
---
  fs/btrfs/export.c | 10 ++++++++++
  1 file changed, 10 insertions(+)
fs/btrfs/export.c +0 -0
index c403117ac..a54c6e8b3 100644

fs/btrfs/export.c static int btrfs_get_name(struct dentry *parent, char *name, · line 289

Quoted 5 lines
fs/btrfs/export.c +2 -0
@@ -285,6 +285,16 @@ static int btrfs_get_name(struct dentry *parent, char *name,
  		name_len = btrfs_inode_ref_name_len(leaf, iref);
  	}
 
+	/* The caller's buffer is only NAME_MAX + 1 bytes. */
+	if (name_len > BTRFS_NAME_LEN) {

We do not play the whac-a-mole game by add random checks ad-hoc.

Do the proper ROOT_REF_KEY and INODE_REF_KEY verification inside tree-checker.

Quoted 11 lines
fs/btrfs/export.c +8 -0
+		btrfs_err(fs_info,
+			  "corrupt name length %d for inode %llu in root %llu, max %d",
+			  name_len, btrfs_ino(BTRFS_I(inode)),
+			  btrfs_root_id(BTRFS_I(inode)->root), BTRFS_NAME_LEN);
+		btrfs_free_path(path);
+		return -EUCLEAN;
+	}
+
  	read_extent_buffer(leaf, name, name_ptr, name_len);
  	btrfs_free_path(path);
 
#3 lore raw
Quoted 50 lines · Qu Wenruo
Quoted 37 lines · Weiming Shi

btrfs_get_name() reads the name length straight from the on-disk inode_ref (or root_ref) and copies that many bytes into the caller's buffer with no upper bound. The caller (exportfs_get_name()) supplies a fixed NAME_MAX + 1 byte stack buffer, but name_len is a __le16 read from the leaf and the tree-checker only bounds it to the item size, not to BTRFS_NAME_LEN. A crafted leaf with name_len = 4096 therefore overflows the 256-byte buffer with attacker-controlled bytes. It is reachable from a mounted untrusted btrfs image via open_by_handle_at(), and on btrfs filesystems exported over NFS.

  BUG: KASAN: stack-out-of-bounds in read_extent_buffer (fs/btrfs/extent_io.c:3742)
  Write of size 633 at addr ffffc90006c9fc40 by task exploit/5192
   read_extent_buffer (fs/btrfs/extent_io.c:3742)
   btrfs_get_name (fs/btrfs/export.c:289)
   reconnect_path (fs/exportfs/expfs.c:222)
   exportfs_decode_fh_raw (fs/exportfs/expfs.c:473)
   do_handle_open (fs/fhandle.c:230)
   do_syscall_64 (arch/x86/entry/syscall_64.c:94)
  Kernel panic - not syncing: stack-protector: Kernel stack is corrupted

Reject any name_len greater than BTRFS_NAME_LEN before the copy. Such a name is never valid on disk, so this only rejects corrupted leaves and leaves valid names unchanged.

Fixes 2ede0daf0154 ("Btrfs: handle NFS lookups properly")
Assisted-by Claude:claude-opus-4-8
---
  fs/btrfs/export.c | 10 ++++++++++
  1 file changed, 10 insertions(+)
fs/btrfs/export.c +0 -0
index c403117ac..a54c6e8b3 100644

fs/btrfs/export.c static int btrfs_get_name(struct dentry *parent, char *name, · line 288

Quoted 4 lines
fs/btrfs/export.c +2 -0
@@ -285,6 +285,16 @@ static int btrfs_get_name(struct dentry *parent, char *name,
  		name_len = btrfs_inode_ref_name_len(leaf, iref);
  	}
+	/* The caller's buffer is only NAME_MAX + 1 bytes. */
+	if (name_len > BTRFS_NAME_LEN) {

We do not play the whac-a-mole game by add random checks ad-hoc.

Do the proper ROOT_REF_KEY and INODE_REF_KEY verification inside tree-checker.

Quoted 11 lines

fs/btrfs/export.c static int btrfs_get_name(struct dentry *parent, char *name, · line 299

Quoted 10 lines
fs/btrfs/export.c +8 -0
+		btrfs_err(fs_info,
+			  "corrupt name length %d for inode %llu in root %llu, max %d",
+			  name_len, btrfs_ino(BTRFS_I(inode)),
+			  btrfs_root_id(BTRFS_I(inode)->root), BTRFS_NAME_LEN);
+		btrfs_free_path(path);
+		return -EUCLEAN;
+	}
+
  	read_extent_buffer(leaf, name, name_ptr, name_len);
  	btrfs_free_path(path);

Thanks for the review. v2 sent.