Skip to contentSkip to conversation

NFSv4.1/pnfs: bound GETDEVICEINFO notification bitmap length

PATCH New Michael Bommarito 1 participant 3 messages Jun 14

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

cover NFSv4.1/pnfs: bound GETDEVICEINFO notification bitmap length

decode_getdeviceinfo() scales a server-supplied notification bitmap length by four for xdr_inline_decode(). The 32-bit multiply wraps to 0 for len >= 0x40000000, so the bounds check passes on an empty request and the verify loop reads up to ~2^30 words past the inline XDR buffer. A malicious or compromised pNFS server, or a man in the middle on an unprotected mount, can drive this out-of-bounds read while the client decodes a GETDEVICEINFO reply. It is a read, not a leak: the loop only checks each word for nonzero and returns -EIO on the first, so over-read contents are never returned.

A conformant reply uses a single notification word. Patch 1 bounds the length before the scale so it cannot wrap. Patch 2 adds KUnit coverage and is offered separately so it can be taken or dropped on its own.

Tested on QEMU x86_64 with KASAN: a Level-2 KUnit case drives the real decode_getdeviceinfo() and reports a slab-out-of-bounds read on stock, returning -EIO after patch 1; two benign controls drive the same decoder in bounds and pass on both stock and fixed trees.

Michael Bommarito (2):
NFSv4.1/pnfs: bound notification bitmap length in decode_getdeviceinfo NFSv4.1/pnfs: add KUnit coverage for GETDEVICEINFO notification decode

 fs/nfs/Kconfig                      |  14 ++++
 fs/nfs/getdeviceinfo_notify_kunit.c | 110 ++++++++++++++++++++++++++++
 fs/nfs/nfs4xdr.c                    |  10 ++-
 include/linux/nfs4.h                |   3 +
 4 files changed, 136 insertions(+), 1 deletion(-)
 create mode 100644 fs/nfs/getdeviceinfo_notify_kunit.c

base-commit: 5200f5f493f79f14bbdc349e402a40dfb32f23c8

Signature · 1 line
2.53.0
#1 lore raw

PATCH 1/2 NFSv4.1/pnfs: bound notification bitmap length in decode_getdeviceinfo

decode_getdeviceinfo() reads a server-supplied notification bitmap length into a u32 and scales it by four for xdr_inline_decode(). The multiply is 32-bit and wraps to 0 for len >= 0x40000000, so the !p bounds check passes on an empty request and the verify loop reads up to ~2^30 words past the inline XDR buffer. A malicious or compromised pNFS server, or a man in the middle on an unprotected mount, can drive this out-of-bounds read while the client decodes a GETDEVICEINFO reply.

RFC 8881 defines only the CHANGE and DELETE deviceid notification bits, both in word 0, so a conformant reply uses a single word. Bound len to NFS4_GETDEVICEINFO_NOTIFY_MAXWORDS before scaling, which cannot wrap and still tolerates trailing zero words the decoder keeps verifying.

Reproduced under KASAN on QEMU via a KUnit case driving the real decode_getdeviceinfo(); the out-of-bounds read is gone after this change.

Fixes b1f69b754ee3 ("NFSv4.1: pnfs: add LAYOUTGET and GETDEVICEINFO infrastructure")
Assisted-by Claude:claude-opus-4-8

---

Reproduction (v7.1-rc4, x86_64 QEMU/KVM, KASAN):

  BUG: KASAN: slab-out-of-bounds in decode_getdeviceinfo

A KUnit case (patch 2) drives the real decode_getdeviceinfo() with a 0x40000000 notification length placed at a page edge; the wrapped 4*len multiply over-reads the receive buffer on stock and faults. Patched, the length is rejected with -EIO. Two benign in-bounds controls pass on both stock and patched. fs/nfs has no existing kselftest for this decoder.

 fs/nfs/nfs4xdr.c     | 6 +++++-
 include/linux/nfs4.h | 3 +++
 2 files changed, 8 insertions(+), 1 deletion(-)
fs/nfs/nfs4xdr.c +5 -1
index c23c2eee1b5c4..ca84d0c872a6c 100644
@@ -6089,7 +6089,11 @@ static int decode_getdeviceinfo(struct xdr_stream *xdr,
 	if (len) {
 		uint32_t i;
 
-		p = xdr_inline_decode(xdr, 4 * len);
+		/* Bound len so len << 2 cannot wrap and defeat the check below. */
+		if (len > NFS4_GETDEVICEINFO_NOTIFY_MAXWORDS)
+			return -EIO;
+
+		p = xdr_inline_decode(xdr, len << 2);
 		if (unlikely(!p))
 			return -EIO;
 
include/linux/nfs4.h +3 -0
index d87be1f25273a..d6a84c4545864 100644
@@ -752,6 +752,9 @@ enum pnfs_notify_deviceid_type4 {
 	NOTIFY_DEVICEID4_DELETE = 1 << 2,
 };
 
+/* All defined deviceid notification bits live in word 0; bound the decoder. */
+#define NFS4_GETDEVICEINFO_NOTIFY_MAXWORDS	4
+
 enum pnfs_block_volume_type {
 	PNFS_BLOCK_VOLUME_SIMPLE	= 0,
 	PNFS_BLOCK_VOLUME_SLICE		= 1,
Signature · 1 line
2.53.0
#2 lore raw

PATCH 2/2 NFSv4.1/pnfs: add KUnit coverage for GETDEVICEINFO notification decode

In reply to Michael Bommarito

Add a KUnit suite driving the real file-local decode_getdeviceinfo() over a crafted GETDEVICEINFO reply to cover the notification-bitmap length pass. It is included into nfs4xdr.c (gated by CONFIG_NFS_GETDEVICEINFO_KUNIT_TEST) to reach the static decoder without exporting it.

A trigger supplies the wrapping length and two benign controls drive the same decoder in bounds. Integer overflow has no sanitizer, so the oracle is the downstream KASAN slab-out-of-bounds read: on QEMU x86_64 with KASAN the trigger faults on stock and passes after patch 1, while both controls pass on stock and fixed trees.

Assisted-by Claude:claude-opus-4-8

---

On QEMU x86_64 with KASAN: the trigger faults on stock (slab-out-of- bounds READ) and passes after patch 1; both benign controls pass on stock and patched. The suite builds the reply in a kmalloc(PAGE_SIZE) buffer with the notification fields in the XDR tail.

 fs/nfs/Kconfig                      |  14 ++++
 fs/nfs/getdeviceinfo_notify_kunit.c | 110 ++++++++++++++++++++++++++++
 fs/nfs/nfs4xdr.c                    |   4 +
 3 files changed, 128 insertions(+)
 create mode 100644 fs/nfs/getdeviceinfo_notify_kunit.c
fs/nfs/Kconfig +14 -0
index 6bb30543eff00..73cdf201ebb23 100644
@@ -215,3 +215,17 @@ config NFS_V4_2_READ_PLUS
 	default y
 	help
 	 Choose Y here to enable use of the NFS v4.2 READ_PLUS operation.
+
+config NFS_GETDEVICEINFO_KUNIT_TEST
+	tristate "KUnit test for pNFS GETDEVICEINFO notification decode" if !KUNIT_ALL_TESTS
+	depends on NFS_V4 && KUNIT
+	default KUNIT_ALL_TESTS
+	help
+	  Builds KUnit coverage for the notification-bitmap length pass in
+	  the NFS client pNFS GETDEVICEINFO reply decoder. The test drives
+	  the real decode_getdeviceinfo() over a crafted reply and, on an
+	  unfixed kernel built with CONFIG_KASAN, reports the slab-out-of-
+	  bounds read caused by a 32-bit overflow in the notification-bitmap
+	  length handling.
+
+	  If unsure, say N.
fs/nfs/getdeviceinfo_notify_kunit.c new file +100 -0

100 lines not on this page · raw message

fs/nfs/nfs4xdr.c +4 -0
index ca84d0c872a6c..42eb82ab0346f 100644
@@ -7788,3 +7788,7 @@ const struct rpc_version nfs_version4 = {
 	.procs			= nfs4_procedures,
 	.counts			= nfs_version4_counts,
 };
+
+#if IS_ENABLED(CONFIG_NFS_GETDEVICEINFO_KUNIT_TEST)
+#include "getdeviceinfo_notify_kunit.c"
+#endif
Signature · 1 line
2.53.0