xfs: reject log records with v2 size but v1 header version to avoid OOB

[PATCH] xfs: reject log records with v2 size but v1 header version to avoid OOB

Posted by Raphael Pinsonneault-Thibeault 2 months, 3 weeks ago

In xlog_do_recovery_pass(),
commit 45cf976008dd ("xfs: fix log recovery buffer allocation for the
legacy h_size fixup")
added a fix to take the corrected h_size (from the xfsprogs bug
workaround) into consideration for the log recovery buffer calculation.
Without it, we would still allocate the buffer based on the incorrect
on-disk size.

However, in a scenario similar to 45cf976008dd, syzbot creates a fuzzed
record where xfs_has_logv2() but the xlog_rec_header h_version !=
XLOG_VERSION_2. Meaning, we skip the log recover buffer calculation
fix and allocate the buffer based on the incorrect on-disk size. Hence,
a KASAN: slab-out-of-bounds read in xlog_do_recovery_pass() ->
xlog_recover_process() -> xlog_cksum().

Fix by rejecting the record header for
h_size > XLOG_HEADER_CYCLE_SIZE && !XLOG_VERSION_2
since the larger h_size cannot work for v1 logs, and the log stripe unit
adjustment is only a v2 feature.

Reported-by: syzbot+9f6d080dece587cfdd4c@syzkaller.appspotmail.com
Closes: https://syzkaller.appspot.com/bug?extid=9f6d080dece587cfdd4c
Tested-by: syzbot+9f6d080dece587cfdd4c@syzkaller.appspotmail.com
Fixes: 45cf976008dd ("xfs: fix log recovery buffer allocation for the legacy h_size fixup")
Signed-off-by: Raphael Pinsonneault-Thibeault <rpthibeault@gmail.com>
---
changelog
v1 -> v2: 
- reject the mount for h_size > XLOG_HEADER_CYCLE_SIZE && !XLOG_VERSION_2
- update commit subject and message

 fs/xfs/xfs_log_recover.c | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)

diff --git a/fs/xfs/xfs_log_recover.c b/fs/xfs/xfs_log_recover.c
index e6ed9e09c027..99a903e01869 100644
--- a/fs/xfs/xfs_log_recover.c
+++ b/fs/xfs/xfs_log_recover.c
@@ -3064,8 +3064,12 @@ xlog_do_recovery_pass(
 		 * still allocate the buffer based on the incorrect on-disk
 		 * size.
 		 */
-		if (h_size > XLOG_HEADER_CYCLE_SIZE &&
-		    (rhead->h_version & cpu_to_be32(XLOG_VERSION_2))) {
+		if (h_size > XLOG_HEADER_CYCLE_SIZE) {
+			if (!(rhead->h_version & cpu_to_be32(XLOG_VERSION_2))) {
+				error = -EFSCORRUPTED;
+				goto bread_err1;
+			}
+
 			hblks = DIV_ROUND_UP(h_size, XLOG_HEADER_CYCLE_SIZE);
 			if (hblks > 1) {
 				kvfree(hbp);
-- 
2.43.0

Re: [PATCH] xfs: reject log records with v2 size but v1 header version to avoid OOB

Posted by Darrick J. Wong 2 months, 3 weeks ago

On Wed, Nov 12, 2025 at 01:18:18PM -0500, Raphael Pinsonneault-Thibeault wrote:
> In xlog_do_recovery_pass(),
> commit 45cf976008dd ("xfs: fix log recovery buffer allocation for the
> legacy h_size fixup")
> added a fix to take the corrected h_size (from the xfsprogs bug
> workaround) into consideration for the log recovery buffer calculation.
> Without it, we would still allocate the buffer based on the incorrect
> on-disk size.
> 
> However, in a scenario similar to 45cf976008dd, syzbot creates a fuzzed
> record where xfs_has_logv2() but the xlog_rec_header h_version !=
> XLOG_VERSION_2. Meaning, we skip the log recover buffer calculation
> fix and allocate the buffer based on the incorrect on-disk size. Hence,
> a KASAN: slab-out-of-bounds read in xlog_do_recovery_pass() ->
> xlog_recover_process() -> xlog_cksum().
> 
> Fix by rejecting the record header for
> h_size > XLOG_HEADER_CYCLE_SIZE && !XLOG_VERSION_2
> since the larger h_size cannot work for v1 logs, and the log stripe unit
> adjustment is only a v2 feature.
> 
> Reported-by: syzbot+9f6d080dece587cfdd4c@syzkaller.appspotmail.com
> Closes: https://syzkaller.appspot.com/bug?extid=9f6d080dece587cfdd4c
> Tested-by: syzbot+9f6d080dece587cfdd4c@syzkaller.appspotmail.com
> Fixes: 45cf976008dd ("xfs: fix log recovery buffer allocation for the legacy h_size fixup")
> Signed-off-by: Raphael Pinsonneault-Thibeault <rpthibeault@gmail.com>
> ---
> changelog
> v1 -> v2: 
> - reject the mount for h_size > XLOG_HEADER_CYCLE_SIZE && !XLOG_VERSION_2
> - update commit subject and message
> 
>  fs/xfs/xfs_log_recover.c | 8 ++++++--
>  1 file changed, 6 insertions(+), 2 deletions(-)
> 
> diff --git a/fs/xfs/xfs_log_recover.c b/fs/xfs/xfs_log_recover.c
> index e6ed9e09c027..99a903e01869 100644
> --- a/fs/xfs/xfs_log_recover.c
> +++ b/fs/xfs/xfs_log_recover.c
> @@ -3064,8 +3064,12 @@ xlog_do_recovery_pass(
>  		 * still allocate the buffer based on the incorrect on-disk
>  		 * size.
>  		 */
> -		if (h_size > XLOG_HEADER_CYCLE_SIZE &&
> -		    (rhead->h_version & cpu_to_be32(XLOG_VERSION_2))) {

Just out of curiosity, why is this a bit flag test?  Did XFS ever emit a
log record with both XLOG_VERSION_2 *and* XLOG_VERSION_1 set?  The code
that writes new log records only sets h_version to 1 or 2, not 3.

(I can't tell if this is a hysterical raisins compatibility thing, or
just bugs)

--D

> +		if (h_size > XLOG_HEADER_CYCLE_SIZE) {
> +			if (!(rhead->h_version & cpu_to_be32(XLOG_VERSION_2))) {
> +				error = -EFSCORRUPTED;
> +				goto bread_err1;
> +			}
> +
>  			hblks = DIV_ROUND_UP(h_size, XLOG_HEADER_CYCLE_SIZE);
>  			if (hblks > 1) {
>  				kvfree(hbp);
> -- 
> 2.43.0
> 
>

Re: [PATCH] xfs: reject log records with v2 size but v1 header version to avoid OOB

Posted by Christoph Hellwig 2 months, 3 weeks ago

On Wed, Nov 12, 2025 at 10:45:04AM -0800, Darrick J. Wong wrote:
> > @@ -3064,8 +3064,12 @@ xlog_do_recovery_pass(
> >  		 * still allocate the buffer based on the incorrect on-disk
> >  		 * size.
> >  		 */
> > -		if (h_size > XLOG_HEADER_CYCLE_SIZE &&
> > -		    (rhead->h_version & cpu_to_be32(XLOG_VERSION_2))) {
> 
> Just out of curiosity, why is this a bit flag test?  Did XFS ever emit a
> log record with both XLOG_VERSION_2 *and* XLOG_VERSION_1 set?  The code
> that writes new log records only sets h_version to 1 or 2, not 3.

Yeah.  This particular instance got added by me, but it is a copy and
paste from xlog_logrec_hblks, which again consolidate multiple chunks
of this style of code, which were moved around a few times.

I think originally this came from Nathan fixing this:

-               if ((h_version && XLOG_VERSION_2) &&
+
+               if ((h_version & XLOG_VERSION_2) &&

or in other words, this was a mess all the way back.