[PATCH] ubifs: fix integer overflow in pad node length validation

[PATCH] ubifs: fix integer overflow in pad node length validation

[Mark Payne]

ubifs_scan_a_node() validates the padding length of a UBIFS_PAD_NODE
with:

   if (pad_len < 0 || offs + node_len + pad_len > c->leb_size)

offs, node_len and pad_len are all signed 32-bit ints, so the sum can
overflow. ubifs_check_node() has already bounded node_len (28 bytes for
a pad node) and guarantees offs + node_len <= c->leb_size, but pad_len
is taken straight from the image. With offs = 0x20 and
pad_len = 0x7fffffc4:

   offs + node_len + pad_len = 0x80000000 (INT_MIN) -> passes the check
   node_len + pad_len        = 0x7fffffe0 (positive, 8-byte aligned)

so the following alignment check also passes and ubifs_scan_a_node()
returns 0x7fffffe0.

ubifs_scan() then does offs += ret, buf += ret and len -= ret, leaving
buf 2 GiB past the scan buffer and len negative. The trailing
empty-space loop (for (; len; offs++, buf++, len--)) dereferences buf
and faults on unmapped memory:

  BUG: unable to handle page fault for address: ff799130402f9000
  #PF: supervisor read access in kernel mode
  RIP: 0010:ubifs_scan+0x24f/0x2a0
  Call Trace:
   ubifs_replay_journal+0x13e/0x18a0
   ubifs_get_tree+0x5f7/0x2410
   vfs_get_tree+0x2a/0x100
   path_mount+0x36d/0xcf0
   __x64_sys_mount+0x12b/0x160

On an unclean image the same wrapped length reaches ubifs_recover_leb(),
where clean_buf() computes c->leb_size - empty_offs with empty_offs ==
INT_MIN; this sign-extends to a huge size_t and the following
memset(*buf, 0xff, ...) writes out of bounds. In practice the read
fault above is hit first.

Bound pad_len by the remaining space instead:

   pad_len > c->leb_size - offs - node_len

offs + node_len <= c->leb_size already holds after ubifs_check_node(),
so the right-hand side cannot underflow, and the comparison no longer
depends on a sum that can overflow.

Built and boot-tested on Linux 7.4.0-rc1 under QEMU with nandsim
(128 MiB Samsung, 128 KiB PEB, 2 KiB page). On the unpatched kernel,
mounting the crafted image oopses in ubifs_scan(); with the fix the pad
node is rejected ("bad pad node at LEB 3:32") and mount fails cleanly
with -EUCLEAN, while a normally-built UBIFS image still mounts and scans
without error.

Fixes: 1e51764a3c2a ("UBIFS: add new flash file system")
Cc: stable@vger.kernel.org
Signed-off-by: Mark Payne <mnp@pobox.com>
---
 fs/ubifs/scan.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/fs/ubifs/scan.c b/fs/ubifs/scan.c
index cbff2d800e27..441b2f04a0ae 100644
--- a/fs/ubifs/scan.c
+++ b/fs/ubifs/scan.c
@@ -86,7 +86,7 @@ int ubifs_scan_a_node(const struct ubifs_info *c, void *buf, int len, int lnum,
 
 		/* Validate the padding node */
 		if (pad_len < 0 ||
-		    offs + node_len + pad_len > c->leb_size) {
+		    pad_len > c->leb_size - offs - node_len) {
 			if (!quiet) {
 				ubifs_err(c, "bad pad node at LEB %d:%d",
 					  lnum, offs);
-- 
2.43.0