Skip to contentSkip to conversation

libceph: fix potential out-of-bounds read in decode_new_up_state_weight()

PATCH In review Zhenhao Wan 3 participants 3 messages Jun 6 – Jul 2

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 libceph: fix potential out-of-bounds read in decode_new_up_state_weight()

The new_state section of an incremental OSD map is validated and skipped using a byte count computed as

	len *= sizeof(u32) + (struct_v >= 5 ? sizeof(u32) : sizeof(u8));

The multiplication is evaluated in size_t, but the result is stored back into the u32 "len", truncating it. A malicious or corrupted incremental map can supply a new_state element count >= 0x20000000 (struct_v >= 5) so that len * 8 wraps modulo 2^32 to a small value. The following ceph_decode_need() then validates far fewer bytes than the section actually occupies.

new_state is then reprocessed with the unchecked ceph_decode_32() and ceph_decode_8() helpers, which have no per-iteration bounds check and rely entirely on that truncated up-front validation. This can lead to a kernel out-of-bounds read past "end".

Compute the byte count in u64 and bounds-check it against the remaining buffer before skipping, mirroring the size_t-typed length checks used elsewhere in this file (e.g. decode_crush_names(), decode_pg_mapping()). The osd index used for the osd_state[] write is already bounds-checked against map->max_osd, so this is an out-of-bounds read, not a write.

Fixes 930c53286977 ("libceph: apply new_state before new_up_client on incrementals")
---
 net/ceph/osdmap.c | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)
net/ceph/osdmap.c +5 -3
index 8b5b0587a0cf..dd3023fe821e 100644
@@ -1842,6 +1842,7 @@ static int decode_new_up_state_weight(void **p, void *end, u8 struct_v,
 	void *new_up_client;
 	void *new_state;
 	void *new_weight_end;
+	u64 skip_len;
 	u32 len;
 	int ret;
 	int i;
@@ -1862,9 +1863,10 @@ static int decode_new_up_state_weight(void **p, void *end, u8 struct_v,
 
 	new_state = *p;
 	ceph_decode_32_safe(p, end, len, e_inval);
-	len *= sizeof(u32) + (struct_v >= 5 ? sizeof(u32) : sizeof(u8));
-	ceph_decode_need(p, end, len, e_inval);
-	*p += len;
+	skip_len = (u64)len * (sizeof(u32) + (struct_v >= 5 ? sizeof(u32) : sizeof(u8)));
+	if (skip_len > end - *p)
+		goto e_inval;
+	*p += skip_len;
 
 	/* new_weight */
 	ceph_decode_32_safe(p, end, len, e_inval);

---
base-commit: dbe8d05c9750b107b10c15361aad40fbb350bedb
change-id: 20260606-ceph-fix-final-16f4e1df5a5e

Best regards,

Signature · 1 line
Zhenhao Wan <[email protected]>
#1 lore raw
Quoted 62 lines · Zhenhao Wan

The new_state section of an incremental OSD map is validated and skipped using a byte count computed as

	len *= sizeof(u32) + (struct_v >= 5 ? sizeof(u32) : sizeof(u8));

The multiplication is evaluated in size_t, but the result is stored back into the u32 "len", truncating it. A malicious or corrupted incremental map can supply a new_state element count >= 0x20000000 (struct_v >= 5) so that len * 8 wraps modulo 2^32 to a small value. The following ceph_decode_need() then validates far fewer bytes than the section actually occupies.

new_state is then reprocessed with the unchecked ceph_decode_32() and ceph_decode_8() helpers, which have no per-iteration bounds check and rely entirely on that truncated up-front validation. This can lead to a kernel out-of-bounds read past "end".

Compute the byte count in u64 and bounds-check it against the remaining buffer before skipping, mirroring the size_t-typed length checks used elsewhere in this file (e.g. decode_crush_names(), decode_pg_mapping()). The osd index used for the osd_state[] write is already bounds-checked against map->max_osd, so this is an out-of-bounds read, not a write.

Fixes 930c53286977 ("libceph: apply new_state before new_up_client on incrementals")
---
 net/ceph/osdmap.c | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)
net/ceph/osdmap.c +5 -4
index 8b5b0587a0cf..dd3023fe821e 100644
@@ -1842,6 +1842,7 @@ static int decode_new_up_state_weight(void **p, void *end, u8 struct_v,
 	void *new_up_client;
 	void *new_state;
 	void *new_weight_end;
+	u64 skip_len;
 	u32 len;
 	int ret;
 	int i;
@@ -1862,9 +1863,10 @@ static int decode_new_up_state_weight(void **p, void *end, u8 struct_v,
 
 	new_state = *p;
 	ceph_decode_32_safe(p, end, len, e_inval);
-	len *= sizeof(u32) + (struct_v >= 5 ? sizeof(u32) : sizeof(u8));
-	ceph_decode_need(p, end, len, e_inval);
-	*p += len;
+	skip_len = (u64)len * (sizeof(u32) + (struct_v >= 5 ? sizeof(u32) : sizeof(u8)));
+	if (skip_len > end - *p)
+		goto e_inval;
+	*p += skip_len;
 
 	/* new_weight */
 	ceph_decode_32_safe(p, end, len, e_inval);
 
---

base-commit: dbe8d05c9750b107b10c15361aad40fbb350bedb
change-id: 20260606-ceph-fix-final-16f4e1df5a5e

Quoted 3 lines

Best regards,
--
Zhenhao Wan <[email protected]>

Hi,
a patch for this issue has already been discussed on the ceph-devel mailing list here:
https://lore.kernel.org/ceph-devel/[email protected]/T/#t

Best regards,
Raphael

maintainer
#2 lore raw
In reply to Zhenhao Wan

Hi Zhenhao,

NACK for now. Thanks for the patch, but I have some concerns that need a v2:

1. [major] (process): Competing patch unresolved on-list
2. [nit] net/ceph/osdmap.c:1866: Open-coded bounds check deserves a comment

Signature · 1 line
Alex Markuze