Skip to contentSkip to conversation

mdrestore: fix restore_v2() superblock length check

PATCH Superseded Pavel Reichl 4 participants 7 messages Dec 9, 2025 – Dec 10, 2025

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

cover mdrestore: fix restore_v2() superblock length check

Hi all,

On s390x (big-endian), running xfstests -g metadump currently fails 6 out of 9 tests. The failure is triggered by the superblock extent-length validation in restore_v2(). The code rejects xme_len == 1, but a length of 1 is the correct and expected value, since the superblock fits within a single 512-byte sector.

On big-endian systems, this length decodes to 1 and the check aborts the restore. On little-endian systems, the same on-disk bytes are interpreted as 16777216, so the faulty logic never triggers there.

The patch removes the incorrect rejection of a valid extent length of 1 and applies proper logic so that the superblock extent length is validated consistently across all architectures.

The outline of the fix was discussed with Chandan (thanks).

Pavel Reichl (1):
mdrestore: fix restore_v2() superblock length check

 mdrestore/xfs_mdrestore.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
Signature · 1 line
2.52.0
#1 lore raw

PATCH 1/1 mdrestore: fix restore_v2() superblock length check

On big-endian architectures (e.g. s390x), restoring a filesystem from a v2 metadump fails with "Invalid superblock disk address/length". This is caused by restore_v2() treating a superblock extent length of 1 as an error, even though a length of 1 is expected because the superblock fits within a 512-byte sector.

On little-endian systems, the same raw extent length bytes that represent a value of 1 on big-endian are misinterpreted as 16777216 due to byte ordering, so the faulty check never triggers there and the bug is hidden.

Fix the issue by using an endian-correct comparison of xme_len so that the superblock extent length is validated properly and consistently on all architectures.

---
 mdrestore/xfs_mdrestore.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
mdrestore/xfs_mdrestore.c +1 -1
index f10c4bef..71c2bb9a 100644
@@ -437,7 +437,7 @@ restore_v2(
 	if (fread(&xme, sizeof(xme), 1, md_fp) != 1)
 		fatal("error reading from metadump file\n");
 
-	if (xme.xme_addr != 0 || xme.xme_len == 1 ||
+	if (xme.xme_addr != 0 || cpu_to_be32(xme.xme_len) != 1 ||
 	    (be64_to_cpu(xme.xme_addr) & XME_ADDR_DEVICE_MASK) !=
 			XME_ADDR_DATA_DEVICE)
 		fatal("Invalid superblock disk address/length\n");
Signature · 1 line
2.52.0
#2 lore raw
Quoted 24 lines · Pavel Reichl

On big-endian architectures (e.g. s390x), restoring a filesystem from a v2 metadump fails with "Invalid superblock disk address/length". This is caused by restore_v2() treating a superblock extent length of 1 as an error, even though a length of 1 is expected because the superblock fits within a 512-byte sector.

On little-endian systems, the same raw extent length bytes that represent a value of 1 on big-endian are misinterpreted as 16777216 due to byte ordering, so the faulty check never triggers there and the bug is hidden.

Fix the issue by using an endian-correct comparison of xme_len so that the superblock extent length is validated properly and consistently on all architectures.

---
 mdrestore/xfs_mdrestore.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
mdrestore/xfs_mdrestore.c +0 -0
index f10c4bef..71c2bb9a 100644

mdrestore/xfs_mdrestore.c restore_v2( · line 440

Quoted 5 lines
mdrestore/xfs_mdrestore.c +1 -1
@@ -437,7 +437,7 @@ restore_v2(
 	if (fread(&xme, sizeof(xme), 1, md_fp) != 1)
 		fatal("error reading from metadump file\n");
 
-	if (xme.xme_addr != 0 || xme.xme_len == 1 ||
+	if (xme.xme_addr != 0 || cpu_to_be32(xme.xme_len) != 1 ||

xme.xme_len is the ondisk value, so that should be be32_to_cpu().

Otherwise the patch looks ok.

--D

Quoted 7 lines
mdrestore/xfs_mdrestore.c +0 -0
 	    (be64_to_cpu(xme.xme_addr) & XME_ADDR_DEVICE_MASK) !=
 			XME_ADDR_DATA_DEVICE)
 		fatal("Invalid superblock disk address/length\n");
Signature · 1 line
2.52.0
#3 lore raw
Quoted 6 lines · Darrick J. Wong

mdrestore/xfs_mdrestore.c restore_v2( · line 440

Quoted 2 lines
mdrestore/xfs_mdrestore.c +1 -1
-	if (xme.xme_addr != 0 || xme.xme_len == 1 ||
+	if (xme.xme_addr != 0 || cpu_to_be32(xme.xme_len) != 1 ||

xme.xme_len is the ondisk value, so that should be be32_to_cpu().

Otherwise the patch looks ok.

We really need to bring back regular sparse runs on the userspace code. Let's see if I can get it back working..

#4 lore raw
Quoted 10 lines · Christoph Hellwig
Quoted 6 lines · Darrick J. Wong

mdrestore/xfs_mdrestore.c restore_v2( · line 440

Quoted 2 lines
mdrestore/xfs_mdrestore.c +1 -1
-	if (xme.xme_addr != 0 || xme.xme_len == 1 ||
+	if (xme.xme_addr != 0 || cpu_to_be32(xme.xme_len) != 1 ||

xme.xme_len is the ondisk value, so that should be be32_to_cpu().

Otherwise the patch looks ok.

We really need to bring back regular sparse runs on the userspace code. Let's see if I can get it back working..

I just gave it a try, and make CC=cgcc still works in theory. But between the urcu headers making it throw up, issues in the Linux UAPI headers and our own redefinition of the __be32/__be16 types it generates so much noise that it stops reporting before any real issues including this one. Sigh. I'll see if there is a way to clean some of this up and get useful results.

#5 lore raw
Quoted 12 lines · Christoph Hellwig
Quoted 10 lines · Christoph Hellwig
Quoted 6 lines · Darrick J. Wong

mdrestore/xfs_mdrestore.c restore_v2( · line 440

Quoted 2 lines
mdrestore/xfs_mdrestore.c +1 -1
-	if (xme.xme_addr != 0 || xme.xme_len == 1 ||
+	if (xme.xme_addr != 0 || cpu_to_be32(xme.xme_len) != 1 ||

xme.xme_len is the ondisk value, so that should be be32_to_cpu().

Otherwise the patch looks ok.

We really need to bring back regular sparse runs on the userspace code. Let's see if I can get it back working..

Quoted 6 lines

I just gave it a try, and make CC=cgcc still works in theory. But between the urcu headers making it throw up, issues in the Linux UAPI headers and our own redefinition of the __be32/__be16 types it generates so much noise that it stops reporting before any real issues including this one. Sigh. I'll see if there is a way to clean some of this up and get useful results.

I usually just grep -v out the macro crap and squint real hard to find the real issues. Oftentimes Dan Carpenter would help out, but I hear that his work on that is coming to an end.

--D

#6 lore raw
In reply to Christoph Hellwig
Quoted 12 lines · Christoph Hellwig
Quoted 10 lines · Christoph Hellwig
Quoted 6 lines · Darrick J. Wong

mdrestore/xfs_mdrestore.c restore_v2( · line 440

Quoted 2 lines
mdrestore/xfs_mdrestore.c +1 -1
-	if (xme.xme_addr != 0 || xme.xme_len == 1 ||
+	if (xme.xme_addr != 0 || cpu_to_be32(xme.xme_len) != 1 ||

xme.xme_len is the ondisk value, so that should be be32_to_cpu().

Otherwise the patch looks ok.

We really need to bring back regular sparse runs on the userspace code. Let's see if I can get it back working..

Quoted 7 lines

I just gave it a try, and make CC=cgcc still works in theory. But between the urcu headers making it throw up, issues in the Linux UAPI headers and our own redefinition of the __be32/__be16 types it generates so much noise that it stops reporting before any real issues including this one. Sigh. I'll see if there is a way to clean some of this up and get useful results.

"make C=1 / C=2" worked once but when I ran it after seeing this patch, it didn't seem to catch any errors. It spewed a lot of other things though, as you mention (urcu, ugh).

I didn't realize that those results could swamp out other reports. :(

-Eric