Skip to contentSkip to conversation

xfs: fix NULL deref in log recovery reorder

PATCH Superseded Weiming Shi 3 participants 12 messages Jul 1 – Jul 2 XFS filesystem

v2 of 3 Superseded Latest is v3 →

lore
Share & tools
lore
#0 lore raw

v2 cover xfs: fix NULL deref in log recovery reorder

A crafted XFS image can commit a transaction whose only item is a bare transaction header (ri_cnt == 0, ri_buf == NULL). xlog_recover_reorder_trans() then runs ITEM_TYPE() on it and dereferences the NULL ri_buf, faulting the kernel at mount time.

v1 was a single patch that added the check but kept an ASSERT(0) on the path and duplicated the sort_list splice. Per Christoph's review, v2 splits it:

1: drop the existing ASSERT(0) on an unrecognized item type 2: move the sort_list splice out of the loop (no functional change) 3: add the check that rejects a committed item with no regions

Patch 3 also notes that, since the log is CRC-checked, this is a crafted image rather than media corruption.

Tested on xfs-7.2-fixes with KASAN: the unpatched kernel oopses in xlog_recover_reorder_trans; the patched kernel fails the mount with -EFSCORRUPTED and no splat.

Weiming Shi (3):
xfs: drop ASSERT(0) on unrecognized log item type
xfs: splice unsorted log items back to the transaction after the loop
xfs: fail recovery on a committed log item with no regions

 fs/xfs/xfs_log_recover.c | 26 ++++++++++++++++++--------
 1 file changed, 18 insertions(+), 8 deletions(-)
Signature · 1 line
2.43.0
#1 lore raw

PATCH v2 1/3 xfs: drop ASSERT(0) on unrecognized log item type

The item type passed to ITEM_TYPE() comes from the on-disk log, so a fuzzed or crafted image can reach the "unrecognized type" path in xlog_recover_reorder_trans() and trip its ASSERT(0) on a CONFIG_XFS_DEBUG kernel. The -EFSCORRUPTED return handles it fine; drop the assert.

---
 fs/xfs/xfs_log_recover.c | 1 -
 1 file changed, 1 deletion(-)
fs/xfs/xfs_log_recover.c +0 -1
index 5f984bf5698a..a1b373c68f0e 100644
@@ -1912,7 +1912,6 @@ xlog_recover_reorder_trans(
 			xfs_warn(log->l_mp,
 				"%s: unrecognized type of log operation (%d)",
 				__func__, ITEM_TYPE(item));
-			ASSERT(0);
 			/*
 			 * return the remaining items back to the transaction
 			 * item list so they can be freed in caller.
Signature · 1 line
2.43.0
Reviewed-by on v2 1/3
#2 lore raw

Re: [PATCH v2 1/3] xfs: drop ASSERT(0) on unrecognized log item type

Quoted 7 lines · Weiming Shi

The item type passed to ITEM_TYPE() comes from the on-disk log, so a fuzzed or crafted image can reach the "unrecognized type" path in xlog_recover_reorder_trans() and trip its ASSERT(0) on a CONFIG_XFS_DEBUG kernel. The -EFSCORRUPTED return handles it fine; drop the assert.

<nod> this is a validation of ondisk artifacts, so no assertions

--D

Quoted 20 lines
---
 fs/xfs/xfs_log_recover.c | 1 -
 1 file changed, 1 deletion(-)
fs/xfs/xfs_log_recover.c +0 -1
index 5f984bf5698a..a1b373c68f0e 100644
@@ -1912,7 +1912,6 @@ xlog_recover_reorder_trans(
 			xfs_warn(log->l_mp,
 				"%s: unrecognized type of log operation (%d)",
 				__func__, ITEM_TYPE(item));
-			ASSERT(0);
 			/*
 			 * return the remaining items back to the transaction
 			 * item list so they can be freed in caller.
Signature · 1 line
2.43.0
#4 lore raw

PATCH v2 2/3 xfs: splice unsorted log items back to the transaction after the loop

In reply to Weiming Shi

On error, xlog_recover_reorder_trans() splices the leftover sort_list items back to trans->r_itemq inside the loop before breaking out. The loop tail already splices the per-fate lists back, so do sort_list there too, guarded by the assert that used to sit after the loop.

No functional change. It drops the duplicated splice so the next patch can add another error case without repeating it.

---
 fs/xfs/xfs_log_recover.c | 16 +++++++++-------
 1 file changed, 9 insertions(+), 7 deletions(-)
fs/xfs/xfs_log_recover.c +9 -7
index a1b373c68f0e..5347f6a5ec42 100644
@@ -1912,12 +1912,6 @@ xlog_recover_reorder_trans(
 			xfs_warn(log->l_mp,
 				"%s: unrecognized type of log operation (%d)",
 				__func__, ITEM_TYPE(item));
-			/*
-			 * return the remaining items back to the transaction
-			 * item list so they can be freed in caller.
-			 */
-			if (!list_empty(&sort_list))
-				list_splice_init(&sort_list, &trans->r_itemq);
 			error = -EFSCORRUPTED;
 			break;
 		}
@@ -1945,7 +1939,15 @@ xlog_recover_reorder_trans(
 		}
 	}
 
-	ASSERT(list_empty(&sort_list));
+	/*
+	 * Return the remaining items back to the transaction item list so they
+	 * can be freed in caller.  This should only happen when we encountered
+	 * an error.
+	 */
+	if (!list_empty(&sort_list)) {
+		ASSERT(error);
+		list_splice_init(&sort_list, &trans->r_itemq);
+	}
 	if (!list_empty(&buffer_list))
 		list_splice(&buffer_list, &trans->r_itemq);
 	if (!list_empty(&item_list))
Signature · 1 line
2.43.0
Reviewed-by on v2 2/3
#5 lore raw

Re: [PATCH v2 2/3] xfs: splice unsorted log items back to the transaction after the loop

Quoted 31 lines · Weiming Shi

On error, xlog_recover_reorder_trans() splices the leftover sort_list items back to trans->r_itemq inside the loop before breaking out. The loop tail already splices the per-fate lists back, so do sort_list there too, guarded by the assert that used to sit after the loop.

No functional change. It drops the duplicated splice so the next patch can add another error case without repeating it.

---
 fs/xfs/xfs_log_recover.c | 16 +++++++++-------
 1 file changed, 9 insertions(+), 7 deletions(-)
fs/xfs/xfs_log_recover.c +0 -6
index a1b373c68f0e..5347f6a5ec42 100644
@@ -1912,12 +1912,6 @@ xlog_recover_reorder_trans(
 			xfs_warn(log->l_mp,
 				"%s: unrecognized type of log operation (%d)",
 				__func__, ITEM_TYPE(item));
-			/*
-			 * return the remaining items back to the transaction
-			 * item list so they can be freed in caller.
-			 */
-			if (!list_empty(&sort_list))
-				list_splice_init(&sort_list, &trans->r_itemq);
 			error = -EFSCORRUPTED;
 			break;
 		}

fs/xfs/xfs_log_recover.c xlog_recover_reorder_trans( · line 1944

Quoted 7 lines
fs/xfs/xfs_log_recover.c +3 -1
@@ -1945,7 +1939,15 @@ xlog_recover_reorder_trans(
 		}
 	}
 
-	ASSERT(list_empty(&sort_list));
+	/*
+	 * Return the remaining items back to the transaction item list so they
+	 * can be freed in caller.  This should only happen when we encountered

s/encountered/encounter/

(same verb tense across a sentence)

fs/xfs/xfs_log_recover.c xlog_recover_reorder_trans( · line 1948

Quoted 4 lines
fs/xfs/xfs_log_recover.c +4 -0
+	 * an error.
+	 */
+	if (!list_empty(&sort_list)) {
+		ASSERT(error);

This otherwise looks fine to me.

--D

Quoted 9 lines
fs/xfs/xfs_log_recover.c +2 -0
+		list_splice_init(&sort_list, &trans->r_itemq);
+	}
 	if (!list_empty(&buffer_list))
 		list_splice(&buffer_list, &trans->r_itemq);
 	if (!list_empty(&item_list))
Signature · 1 line
2.43.0
Reviewed-by
#6 lore raw

Re: [PATCH v2 2/3] xfs: splice unsorted log items back to the transaction after the loop

Quoted 7 lines · Darrick J. Wong

fs/xfs/xfs_log_recover.c xlog_recover_reorder_trans( · line 1944

Quoted 3 lines
fs/xfs/xfs_log_recover.c +3 -0
+	/*
+	 * Return the remaining items back to the transaction item list so they
+	 * can be freed in caller.  This should only happen when we encountered

s/encountered/encounter/

(same verb tense across a sentence)

That might have been my fault :)

Otherwise looks good:

#7 lore raw

PATCH v2 3/3 xfs: fail recovery on a committed log item with no regions

In reply to Weiming Shi

If the first op of a transaction is a bare transaction header (len == sizeof(struct xfs_trans_header)), xlog_recover_add_to_trans() adds a recovery item but no region, leaving it on r_itemq with ri_cnt == 0 and ri_buf == NULL.

If a commit op follows, xlog_recover_reorder_trans() calls ITEM_TYPE() on the item, which reads *(unsigned short *)item->ri_buf[0].iov_base and faults on the NULL ri_buf.

The log is CRC-checked, so this op sequence comes from a crafted image, not media corruption. It faults the kernel when such an image is mounted:

 KASAN: null-ptr-deref in range [0x0000000000000000-0x0000000000000007]
 RIP: 0010:xlog_recover_reorder_trans (fs/xfs/xfs_log_recover.c:1836)
  xlog_recover_commit_trans (fs/xfs/xfs_log_recover.c:2043)
  xlog_recover_process_data (fs/xfs/xfs_log_recover.c:2501)
  xlog_do_recovery_pass (fs/xfs/xfs_log_recover.c:3244)
  xlog_recover (fs/xfs/xfs_log_recover.c:3493)
  xfs_log_mount (fs/xfs/xfs_log.c:618)
  xfs_mountfs (fs/xfs/xfs_mount.c:1034)
  xfs_fs_fill_super (fs/xfs/xfs_super.c:1938)
  vfs_get_tree (fs/super.c:1695)
  path_mount (fs/namespace.c:4161)
  __x64_sys_mount (fs/namespace.c:4367)

A committed item always carries its format descriptor in ri_buf[0], so one with no regions is invalid. Reject it with -EFSCORRUPTED, like the unrecognized item type below.

Fixes 89cebc847729 ("xfs: validate transaction header length on log recovery")
Assisted-by Claude:claude-opus-4-8
---
 fs/xfs/xfs_log_recover.c | 9 +++++++++
 1 file changed, 9 insertions(+)
fs/xfs/xfs_log_recover.c +9 -0
index 5347f6a5ec42..461e847c32a2 100644
@@ -1907,6 +1907,15 @@ xlog_recover_reorder_trans(
 	list_for_each_entry_safe(item, n, &sort_list, ri_list) {
 		enum xlog_recover_reorder	fate = XLOG_REORDER_ITEM_LIST;
 
+		/* a committed item with no regions has a NULL ri_buf[0] */
+		if (!item->ri_cnt || !item->ri_buf) {
+			xfs_warn(log->l_mp,
+				"%s: committed log item has no regions",
+				__func__);
+			error = -EFSCORRUPTED;
+			break;
+		}
+
 		item->ri_ops = xlog_find_item_ops(item);
 		if (!item->ri_ops) {
 			xfs_warn(log->l_mp,
Signature · 1 line
2.43.0
#8 lore raw

Re: [PATCH v2 3/3] xfs: fail recovery on a committed log item with no regions

Quoted 8 lines · Weiming Shi

If the first op of a transaction is a bare transaction header (len == sizeof(struct xfs_trans_header)), xlog_recover_add_to_trans() adds a recovery item but no region, leaving it on r_itemq with ri_cnt == 0 and ri_buf == NULL.

If a commit op follows, xlog_recover_reorder_trans() calls ITEM_TYPE() on the item, which reads *(unsigned short *)item->ri_buf[0].iov_base and faults on the NULL ri_buf.

How do we escape xlog_recover_add_to_trans with item->ri_cnt==0 when there's a bare transaction header? Is this the "!len" case at the top of xlog_recover_add_to_trans?

Why don't we abort log recovery right then and there? Are we hoping that a subsequent "continue" section will supply some items, and that's why we can only check ri_cnt/ri_buf if the ondisk transaction is committed?

Quoted 17 lines

The log is CRC-checked, so this op sequence comes from a crafted image, not media corruption. It faults the kernel when such an image is mounted:

 KASAN: null-ptr-deref in range [0x0000000000000000-0x0000000000000007]
 RIP: 0010:xlog_recover_reorder_trans (fs/xfs/xfs_log_recover.c:1836)
  xlog_recover_commit_trans (fs/xfs/xfs_log_recover.c:2043)
  xlog_recover_process_data (fs/xfs/xfs_log_recover.c:2501)
  xlog_do_recovery_pass (fs/xfs/xfs_log_recover.c:3244)
  xlog_recover (fs/xfs/xfs_log_recover.c:3493)
  xfs_log_mount (fs/xfs/xfs_log.c:618)
  xfs_mountfs (fs/xfs/xfs_mount.c:1034)
  xfs_fs_fill_super (fs/xfs/xfs_super.c:1938)
  vfs_get_tree (fs/super.c:1695)
  path_mount (fs/namespace.c:4161)
  __x64_sys_mount (fs/namespace.c:4367)
Quoted 8 lines

A committed item always carries its format descriptor in ri_buf[0], so one with no regions is invalid. Reject it with -EFSCORRUPTED, like the unrecognized item type below.

Fixes 89cebc847729 ("xfs: validate transaction header length on log recovery")
Assisted-by Claude:claude-opus-4-8
Cc <[email protected]> # v4.3

I agree that a totally empty log item is a sign of corruption, or at least something going seriously wrong. Can the runtime log code ever be tricked into emitting a bare transaction header?

--D

Quoted 28 lines
---
 fs/xfs/xfs_log_recover.c | 9 +++++++++
 1 file changed, 9 insertions(+)
fs/xfs/xfs_log_recover.c +9 -0
index 5347f6a5ec42..461e847c32a2 100644
@@ -1907,6 +1907,15 @@ xlog_recover_reorder_trans(
 	list_for_each_entry_safe(item, n, &sort_list, ri_list) {
 		enum xlog_recover_reorder	fate = XLOG_REORDER_ITEM_LIST;
 
+		/* a committed item with no regions has a NULL ri_buf[0] */
+		if (!item->ri_cnt || !item->ri_buf) {
+			xfs_warn(log->l_mp,
+				"%s: committed log item has no regions",
+				__func__);
+			error = -EFSCORRUPTED;
+			break;
+		}
+
 		item->ri_ops = xlog_find_item_ops(item);
 		if (!item->ri_ops) {
 			xfs_warn(log->l_mp,
Signature · 1 line
2.43.0
#9 lore raw

Re: [PATCH v2 3/3] xfs: fail recovery on a committed log item with no regions

Quoted 3 lines · Darrick J. Wong

I agree that a totally empty log item is a sign of corruption, or at least something going seriously wrong. Can the runtime log code ever be tricked into emitting a bare transaction header?

I don't think so. This looks like an automated search for bugs in the parser. And while I'm happy about tightening up all parsing, I still think this should be stated in the commit log much more clearly.

#10 lore raw

Re: [PATCH v2 3/3] xfs: fail recovery on a committed log item with no regions

Quoted 8 lines · Christoph Hellwig
Quoted 3 lines · Darrick J. Wong

I agree that a totally empty log item is a sign of corruption, or at least something going seriously wrong. Can the runtime log code ever be tricked into emitting a bare transaction header?

I don't think so. This looks like an automated search for bugs in the parser. And while I'm happy about tightening up all parsing, I still think this should be stated in the commit log much more clearly.

Yes, it could at least state whether this is a code change because the author hit a crash in practice vs. running some tool that pointed out a gap and emitted a reproducer.

--D

#11 lore raw

Re: [PATCH v2 3/3] xfs: fail recovery on a committed log item with no regions

Quoted 11 lines · Darrick J. Wong <[email protected]> 于2026年7月2日周四 23:32写道:
Quoted 8 lines · Christoph Hellwig
Quoted 3 lines · Darrick J. Wong

I agree that a totally empty log item is a sign of corruption, or at least something going seriously wrong. Can the runtime log code ever be tricked into emitting a bare transaction header?

I don't think so. This looks like an automated search for bugs in the parser. And while I'm happy about tightening up all parsing, I still think this should be stated in the commit log much more clearly.

Quoted 5 lines

Yes, it could at least state whether this is a code change because the author hit a crash in practice vs. running some tool that pointed out a gap and emitted a reproducer.

--D

Hi,

It came from an AI-assisted code audit of the log recovery parser. I'll state that in the v3 changelog.

Best,
Weiming Shi