* [PATCH] qnx6: validate di_filelevels against the block size
[not found] <c4-qnx6-filelevels-ubsan-20261031@qq.com>
@ 2026-11-08 2:01 ` Guikun Yang
2026-11-17 13:49 ` Christian Brauner
0 siblings, 1 reply; 2+ messages in thread
From: Guikun Yang @ 2026-11-08 2:01 UTC (permalink / raw)
To: Christian Brauner; +Cc: linux-fsdevel, linux-kernel
qnx6_map_block() computes the direct-pointer slot as
bitdelta = ptrbits * depth;
levelptr = no >> bitdelta;
with depth coming from the on-disk di_filelevels byte. Nothing bounds
depth against the shift width: with ptrbits = 7 (512-byte blocks) a
di_filelevels of 5 already shifts a 32-bit block number by 35, which
is undefined, and the resulting levelptr indexes di_block_ptr[] out of
range. The superblock check limits the inode and longfilename tree
levels to the constant QNX6_PTR_MAX_LEVELS (5), which is not
sufficient for small block sizes either, and the per-inode
di_filelevels byte was never checked at all.
The usable maximum is 31 / ptrbits. Enforce it both where the
superblock levels are validated and in qnx6_iget().
Reported-by: Guikun Yang <yangguikun19@qq.com> (bug report of 2026-10-31, <c4-qnx6-filelevels-ubsan-20261031@qq.com>)
Signed-off-by: Guikun Yang <yangguikun19@qq.com>
---
fs/qnx6/inode.c | 22 +++++++++++++++++++++-
1 file changed, 21 insertions(+), 1 deletion(-)
diff --git a/fs/qnx6/inode.c b/fs/qnx6/inode.c
index 1111111..2222222 100644
--- a/fs/qnx6/inode.c
+++ b/fs/qnx6/inode.c
@@ -406,6 +406,7 @@ static struct buffer_head *qnx6_check_first_superblock(struct super_block *s,
struct qnx6_super_block *sb1)
{
...
+ int maxlevels;
/* sanity check - limit maximum indirect pointer levels */
- if (sb1->Inode.levels > QNX6_PTR_MAX_LEVELS) {
+ sbi = QNX6_SB(s);
+ sbi->s_ptrbits = ilog2(s->s_blocksize / 4);
+ maxlevels = min_t(int, QNX6_PTR_MAX_LEVELS, 31 / sbi->s_ptrbits);
+ if (sb1->Inode.levels > maxlevels) {
pr_err("too many inode levels (max %i, sb %i)\n",
- QNX6_PTR_MAX_LEVELS, sb1->Inode.levels);
+ maxlevels, sb1->Inode.levels);
goto out;
}
- if (sb1->Longfile.levels > QNX6_PTR_MAX_LEVELS) {
+ if (sb1->Longfile.levels > maxlevels) {
pr_err("too many longfilename levels (max %i, sb %i)\n",
- QNX6_PTR_MAX_LEVELS, sb1->Longfile.levels);
+ maxlevels, sb1->Longfile.levels);
goto out;
}
...
- sbi = QNX6_SB(s);
- sbi->s_ptrbits = ilog2(s->s_blocksize / 4);
/* (moved above the level checks) */
@@ -561,6 +562,14 @@ struct inode *qnx6_iget(struct super_block *sb, unsigned ino)
memcpy(&ei->di_block_ptr, &raw_inode->di_block_ptr,
sizeof(raw_inode->di_block_ptr));
ei->di_filelevels = raw_inode->di_filelevels;
+ if (ei->di_filelevels > 31 / sbi->s_ptrbits) {
+ pr_err("inode %u on dev %s: too many file levels (%u > %lu)\n",
+ ino, sb->s_id, ei->di_filelevels, 31 / sbi->s_ptrbits);
+ folio_release_kmap(folio, raw_inode);
+ iget_failed(inode);
+ return ERR_PTR(-EIO);
+ }
Tested on the image from the bug report: the crafted di_filelevels is
now rejected at iget with the message above instead of shifting out of
bounds in qnx6_map_block(); the unmodified image still mounts.
Fixes: 3735188fae42 ("qnx6: add support")
--
2.43.0
^ permalink raw reply [flat|nested] 2+ messages in thread