mptcp: fix request migration ownership

From: Ruide Cao <[email protected]>

Hi,

TCP request migration clones pending requests with inet_reqsk_clone(). Some MPTCP request fields carry ownership which cannot be transferred by a plain byte copy.

For MP_JOIN requests, subflow_req->msk holds a socket reference. The clone inherits the pointer without acquiring its own reference, so the original and cloned requests can drop the same reference. Patch 1 lets the cloned request take an additional reference.

For MP_CAPABLE requests, token_node belongs to the original request and is hashed in the token table. Copying the node gives the clone invalid hash state, while concurrent request migration and RX processing can observe different owners in the ehash and token tables. Patch 2 moves token-table ownership from the original request to the clone under the bucket lock, and makes token acceptance and destruction tolerate a token that has already been removed.

--------------------
Changes in v3:

- Split MP_JOIN and MP_CAPABLE into two patches.
- Reworked MP_CAPABLE token migration to happen during MPTCP request cloning.
- Made MP_CAPABLE accept/destroy paths tolerant of already moved or removed request tokens.
- Added a packetdrill MP_CAPABLE reproducer and decoded warning.
- Dropped the redundant IPPROTO_TCP guard around the direct MPTCP clone call, as request migration is TCP-only.
- Based v3 on the latest net tree rather than the current mptcp_net-next export, which has not yet merged commit a0ab2ba83e35 ("tcp: fix TFO max_qlen accounting across reuseport migration") and therefore lacks the overlapping inet_reqsk_clone() changes.
- v2 link:
https://lore.kernel.org/all/40fd38e7a368e5b7bc9bc83364a32241f977d53f.1778404619.git.caoruide123@gmail.com/

Changes in v2:

- drop the generic request_sock clone callback
- call MPTCP directly from inet_reqsk_clone() under the TCP protocol check
- keep cloned MP_JOIN requests holding an msk reference
- clear raw-copied MP_CAPABLE token hash state in the clone
- move MP_CAPABLE token ownership only after successful req migration
- avoid exposing token internals to inet_connection_sock.c
- update the commit message accordingly
- v1 link:
https://lore.kernel.org/all/86e2514b533bf4d55d4aa2fdbf1404022e8c9430.1776149210.git.caoruide123@gmail.com/

------------------------------

// poc for MP_JOIN:

// Minimal reproducer for a stale subflow_req->msk after reqsk migration. --tolerance_usecs=200000
--non_fatal=packet

`sysctl -q net.mptcp.enabled=1
sysctl -q net.ipv4.tcp_migrate_req=1
sysctl -q net.ipv4.tcp_synack_retries=1`

// Listener A and the owning MPTCP connection.

+0     socket(..., SOCK_STREAM, IPPROTO_MPTCP) = 3
+0     setsockopt(3, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0
+0     setsockopt(3, SOL_SOCKET, SO_REUSEPORT, [1], 4) = 0
+0     bind(3, ..., ...) = 0
+0     listen(3, 8) = 0

+0.0   <  addr[caddr0] > addr[saddr0]  S   0:0(0)         win 65535  <mss 1460, sackOK, TS val 1000 ecr 0,    nop, wscale 8, mpcapable v1 flags[flag_h] nokey>
+0.0   >                               S.  0:0(0)  ack 1             <mss 1460, sackOK, TS val 1000 ecr 1000, nop, wscale 8, mpcapable v1 flags[flag_h] key[skey]>
+0.1   <                                .  1:1(0)  ack 1  win 256    <nop, nop, TS val 1000 ecr 1000, mpcapable v1 flags[flag_h] key[ckey=2, skey]>
+0     accept(3, ..., ...) = 4

// Make the MPTCP socket fully established so it accepts MP_JOIN.

+0.1   <                               P.  1:3(2)  ack 1  win 256    <nop, nop, TS val 1001 ecr 1000, mpcapable v1 flags[flag_h] key[skey, ckey] mpcdatalen 2, nop, nop>
+0.0   >                                .  1:1(0)  ack 3             <nop, nop, TS val 1001 ecr 1001, dss dack8=3 dll=0 nocs>

// Leave exactly one MP_JOIN request half-open.

+0.1   <  addr[caddr1] > addr[saddr0]  S   0:0(0)         win 65535  <mss 1460, sackOK, TS val 2000 ecr 0,    nop, wscale 8, mp_join_syn address_id=1 token=sha256_32(skey)>
+0.0   >                               S.  0:0(0)  ack 1             <mss 1460, sackOK, TS val 2000 ecr 2000, nop, wscale 8, mp_join_syn_ack address_id=0 sender_hmac=auto>

// Listener B joins the reuseport group, then A is closed. The next request // timer clones and migrates the half-open MP_JOIN request to B.

+0.1   socket(..., SOCK_STREAM, IPPROTO_MPTCP) = 5
+0     setsockopt(5, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0
+0     setsockopt(5, SOL_SOCKET, SO_REUSEPORT, [1], 4) = 0
+0     bind(5, ..., ...) = 0
+0     listen(5, 8) = 0
+0     close(3) = 0

// Wait past the first SYN+ACK RTO, then release the owning MPTCP socket.
+1.5 setsockopt(4, SOL_SOCKET, SO_LINGER, {onoff=1, linger=0}, 8) = 0
+0 close(4) = 0

// The migrated request expires on its next timer and its destructor uses msk.
+4.0 true

------------------------------
crash log of MP_JOIN

[  280.449259] [      C0] BUG: KASAN: slab-use-after-free in subflow_req_destructor (net/mptcp/subflow.c:45)
[  280.449417] [      C0] Write of size 4 at addr ff1100010e008d40 by task swapper/0/0
[  280.449525] [      C0] CPU: 0 UID: 0 PID: 0 Comm: swapper/0 Not tainted 7.2.0-rc5-00353-gc27e36054537 #10 PREEMPT(full)

[  280.449637] [      C0] Call Trace:
[  280.450422] [      C0]  subflow_req_destructor (net/mptcp/subflow.c:45)
[  280.450504] [      C0]  subflow_v4_req_destructor (net/mptcp/subflow.c:694)
[  280.450581] [      C0]  __reqsk_free (net/ipv4/inet_connection_sock.c:906)
[  280.450681] [      C0]  reqsk_timer_handler (include/net/request_sock.h:137 net/ipv4/inet_connection_sock.c:1147)

[  280.454937] [      C0] Allocated by task 10014:
[  280.455381] [      C0]  sk_prot_alloc (net/core/sock.c:2246)
[  280.455516] [      C0]  sk_clone (net/core/sock.c:2488)
[  280.455611] [      C0]  mptcp_sk_clone_init (include/net/sock.h:1848 net/mptcp/protocol.c:3564)
[  280.455683] [      C0]  subflow_syn_recv_sock (net/mptcp/subflow.c:883)
[  280.455772] [      C0]  tcp_check_req (net/ipv4/tcp_minisocks.c:934)

[  280.457177] [      C0] Freed by task 0:
[  280.457603] [      C0]  slab_free_after_rcu_debug (include/linux/kasan.h:235 mm/slub.c:2677 mm/slub.c:6439)
[  280.457688] [      C0]  rcu_core (kernel/rcu/tree.c:2645 kernel/rcu/tree.c:2897)

[  280.458199] [      C0] Last potentially related work creation:
[  280.458426] [      C0]  kmem_cache_free (mm/slub.c:2638 mm/slub.c:6377 mm/slub.c:6504)
[  280.458518] [      C0]  __sk_destruct (net/core/sock.c:2289 net/core/sock.c:2391)
[  280.458607] [      C0]  sk_destruct (net/core/sock.c:2419)
[  280.458700] [      C0]  __sk_free (net/core/sock.c:2430)
[  280.458793] [      C0]  sk_free (net/core/sock.c:2441)
[  280.458885] [      C0]  mptcp_close (include/net/sock.h:2020 net/mptcp/protocol.c:3399)
[  280.458969] [      C0]  inet_release (net/ipv4/af_inet.c:442)

[  280.459536] [      C0] The buggy address belongs to the cache MPTCP of size 2968
[  280.459593] [      C0] The buggy address is 128 bytes inside a freed 2968-byte region

------------------------------

MP_CAPABLE packetdrill reproducer:

// Reproducer for MP_CAPABLE request token ownership during TCP req migration. //
// The first listener owns the request created by the MP_CAPABLE SYN. A second // SO_REUSEPORT listener is added only after that SYN, then the first listener is // closed. The SYN+ACK retransmission timer migrates the request to the second // listener, and a later request timer destroys the migrated request. //
// On a vulnerable kernel, inet_reqsk_clone() raw-copies token_node. The clone // is not the token table owner, so destroying the migrated request triggers the // MPTCP token ownership bug.
--tolerance_usecs=250000

+0     `sysctl -q net.mptcp.enabled=1`
+0     `sysctl -q net.ipv4.tcp_migrate_req=1`
+0     `sysctl -q net.ipv4.tcp_synack_retries=2`
+0     `sysctl -q net.ipv4.tcp_timestamps=1`
+0     `sysctl -q kernel.panic_on_warn=0`
+0     `sysctl -q kernel.panic_on_oops=0`
+0     `ip tcp_metrics flush all >/dev/null 2>&1 || true`
+0     `tc qdisc replace dev tun0 root pfifo >/dev/null 2>&1 || true`

+0     socket(..., SOCK_STREAM, IPPROTO_MPTCP) = 3
+0     setsockopt(3, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0
+0     setsockopt(3, SOL_SOCKET, SO_REUSEPORT, [1], 4) = 0
+0     getsockopt(3, SOL_TCP, TCP_IS_MPTCP, [1], [4]) = 0
+0     bind(3, ..., ...) = 0
+0     listen(3, 1) = 0

+0       <  S   0:0(0)         win 32792  <mss 1000, sackOK, nop, nop, nop, wscale 7, mpcapable v1 flags[flag_h] nokey>
+0       >  S.  0:0(0)  ack 1             <mss 1460, nop, nop, sackOK, nop, wscale 9, mpcapable v1 flags[flag_h] key[skey]>

+0     socket(..., SOCK_STREAM, IPPROTO_MPTCP) = 4
+0     setsockopt(4, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0
+0     setsockopt(4, SOL_SOCKET, SO_REUSEPORT, [1], 4) = 0
+0     getsockopt(4, SOL_TCP, TCP_IS_MPTCP, [1], [4]) = 0
+0     bind(4, ..., ...) = 0
+0     listen(4, 1) = 0

+0 close(3) = 0

// Let the retransmission timer migrate the request to fd 4 and let the migrated // request expire. On vulnerable kernels its raw-copied token_node is not the // token-table owner, so the request timer trips the token owner assertion.
+8.0 true

+0 close(4) = 0

------------------------------
decoded warning from MP_CAPABLE reproducer:

[  314.661418] [      C2] ------------[ cut here ]------------
[  314.661636] [      C2] WARNING: net/mptcp/token.c:364 at mptcp_token_destroy_request+0x2b0/0x330, CPU#2: swapper/2/0
[  314.661931] [      C2] CPU: 2 UID: 0 PID: 0 Comm: swapper/2 Not tainted 7.2.0-rc5-00353-gc27e36054537 #10 PREEMPT(full)
[  314.662092] [      C2] RIP: 0010:mptcp_token_destroy_request (build/../net/mptcp/token.c:364 (discriminator 1))
[  314.663001] [      C2] Call Trace:
[  314.663043] [      C2]  <IRQ>
[  314.663107] [      C2]  subflow_v4_req_destructor (build/../net/mptcp/subflow.c:694)
[  314.663213] [      C2]  __reqsk_free (build/../net/ipv4/inet_connection_sock.c:906)
[  314.663325] [      C2]  reqsk_timer_handler (build/../include/net/request_sock.h:137 build/../net/ipv4/inet_connection_sock.c:1147)
[  314.663762] [      C2]  call_timer_fn (build/../kernel/time/timer.c:1748)
[  314.668915] [      C2] ---[ end trace 0000000000000000 ]---

Ruide Cao (2):
mptcp: hold MP_JOIN msk ref when cloning reqsk
mptcp: fix MP_CAPABLE token migration when cloning reqsk

 include/net/mptcp.h             |  7 ++++
 net/ipv4/inet_connection_sock.c |  4 +++
 net/mptcp/protocol.c            | 12 ++++---
 net/mptcp/protocol.h            |  4 ++-
 net/mptcp/subflow.c             | 13 +++++++
 net/mptcp/token.c               | 61 +++++++++++++++++++++++++++++----
 net/mptcp/token_test.c          |  4 +--
 7 files changed, 90 insertions(+), 15 deletions(-)

base-commit: 44871eadd07a7f004aa00cb87399461eea08c630

2.43.0

From: Ruide Cao <[email protected]>

TCP request migration clones pending request sockets with inet_reqsk_clone(). For MPTCP MP_JOIN requests this byte-copies subflow_req->msk, but the clone does not take its own reference.

The original and cloned request can consequently drop the same msk reference from subflow_req_destructor(), leaving one request with a dangling pointer after the other is released.

Add an MPTCP clone helper on the TCP request migration path and let cloned subflow requests grab their own msk reference when one is present. The clone's normal destructor balances this reference on both successful migration and failed clone/insert paths.

---
 include/net/mptcp.h             |  7 +++++++
 net/ipv4/inet_connection_sock.c |  4 ++++
 net/mptcp/subflow.c             | 11 +++++++++++
 3 files changed, 22 insertions(+)

index 71b9fc5a5796..0a02ac1ed22d 100644
@@ -223,6 +223,8 @@ int mptcp_subflow_init_cookie_req(struct request_sock *req,
 struct request_sock *mptcp_subflow_reqsk_alloc(const struct request_sock_ops *ops,
 					       struct sock *sk_listener,
 					       bool attach_listener);
+void mptcp_subflow_reqsk_clone(struct request_sock *req,
+			       struct request_sock *new_req);
 
 __be32 mptcp_get_reset_option(const struct sk_buff *skb);
 

@@ -309,6 +311,11 @@ static inline struct request_sock *mptcp_subflow_reqsk_alloc(const struct reques
 	return NULL;
 }
 
+static inline void mptcp_subflow_reqsk_clone(struct request_sock *req,
+					     struct request_sock *new_req)
+{
+}
+
 static inline __be32 mptcp_reset_option(const struct sk_buff *skb)  { return htonl(0u); }
 
 static inline void mptcp_active_detect_blackhole(struct sock *sk, bool expired) { }

index 6257459bcee2..896f472dcba2 100644
@@ -21,6 +21,7 @@
 #include <net/xfrm.h>
 #include <net/tcp.h>
 #include <net/tcp_ecn.h>
+#include <net/mptcp.h>
 #include <net/sock_reuseport.h>
 #include <net/addrconf.h>
 

@@ -961,6 +962,9 @@ static struct request_sock *inet_reqsk_clone(struct request_sock *req,
 		rcu_assign_pointer(tcp_sk(nreq->sk)->fastopen_rsk, nreq);
 	}
 
+	if (rsk_is_mptcp(req))
+		mptcp_subflow_reqsk_clone(req, nreq);
+
 	return nreq;
 }
 

index 8e386899ceb9..be7260821566 100644
@@ -47,6 +47,17 @@ static void subflow_req_destructor(struct request_sock *req)
 	mptcp_token_destroy_request(req);
 }
 
+void mptcp_subflow_reqsk_clone(struct request_sock *req,
+			       struct request_sock *new_req)
+{
+	struct mptcp_subflow_request_sock *subflow_req;
+
+	subflow_req = mptcp_subflow_rsk(new_req);
+
+	if (subflow_req->msk)
+		sock_hold((struct sock *)subflow_req->msk);
+}
+
 static void subflow_generate_hmac(u64 key1, u64 key2, u32 nonce1, u32 nonce2,
 				  void *hmac)
 {

2.43.0

From: Ruide Cao <[email protected]>

TCP request migration clones pending request sockets with inet_reqsk_clone(). For MPTCP MP_CAPABLE requests this byte-copies the token_node hlist state into the clone even though the token table still names the original request.

Moving the token only after inet_ehash_insert() succeeds leaves a window where the cloned request is already globally visible from the ehash but the token table still points at the original request. Reordering the move after clone but before ehash exposure closes that window, but concurrent RX can still race on the old request and observe that the token was already moved.

Move MP_CAPABLE token request ownership during MPTCP request cloning under the token bucket lock. Make mptcp_token_accept() and mptcp_token_destroy_request() re-check token_node under the same lock and treat an already moved or removed request as a normal race instead of warning.

If the passive MP_CAPABLE socket cannot claim the token, fail mptcp_sk_clone_init() and let the subflow fall back instead of installing a socket with mismatched token ownership.

---
 net/mptcp/protocol.c   | 12 +++++----
 net/mptcp/protocol.h   |  4 ++-
 net/mptcp/subflow.c    |  2 ++
 net/mptcp/token.c      | 61 +++++++++++++++++++++++++++++++++++++-----
 net/mptcp/token_test.c |  4 +--
 5 files changed, 68 insertions(+), 15 deletions(-)

index ca644ec53eed..07eaa6858d05 100644
@@ -3608,17 +3608,19 @@ struct sock *mptcp_sk_clone_init(const struct sock *sk,
 	 */
 	mptcp_set_state(nsk, TCP_ESTABLISHED);
 
+	if (!mptcp_token_accept(subflow_req, msk)) {
+		mptcp_release_sched(msk);
+		inet_csk_prepare_forced_close(nsk);
+		tcp_done(nsk);
+		return NULL;
+	}
+
 	/* The msk maintain a ref to each subflow in the connections list */
 	WRITE_ONCE(msk->first, ssk);
 	subflow = mptcp_subflow_ctx(ssk);
 	list_add(&subflow->node, &msk->conn_list);
 	sock_hold(ssk);
 
-	/* new mpc subflow takes ownership of the newly
-	 * created mptcp socket
-	 */
-	mptcp_token_accept(subflow_req, msk);
-
 	/* set msk addresses early to ensure mptcp_pm_get_local_id()
 	 * uses the correct data
 	 */

index 4a2d40cd7b13..c58cfb1326b7 100644
@@ -1074,9 +1074,11 @@ static inline void mptcp_token_init_request(struct request_sock *req)
 }
 
 int mptcp_token_new_request(struct request_sock *req);
+void mptcp_token_move_request(struct request_sock *req,
+			      struct request_sock *new_req);
 void mptcp_token_destroy_request(struct request_sock *req);
 int mptcp_token_new_connect(struct sock *ssk);
-void mptcp_token_accept(struct mptcp_subflow_request_sock *r,
+bool mptcp_token_accept(struct mptcp_subflow_request_sock *r,
 			struct mptcp_sock *msk);
 bool mptcp_token_exists(u32 token);
 struct mptcp_sock *mptcp_token_get_sock(struct net *net, u32 token);

index be7260821566..65d876a08233 100644
@@ -56,6 +56,8 @@ void mptcp_subflow_reqsk_clone(struct request_sock *req,
 
 	if (subflow_req->msk)
 		sock_hold((struct sock *)subflow_req->msk);
+
+	mptcp_token_move_request(req, new_req);
 }
 
 static void subflow_generate_hmac(u64 key1, u64 key2, u32 nonce1, u32 nonce2,

index 4fc39fa2e262..be9acce8a567 100644
@@ -99,7 +99,7 @@ static void mptcp_token_test_accept(struct kunit *test)
 	KUNIT_ASSERT_EQ(test, 0,
 			mptcp_token_new_request((struct request_sock *)req));
 	msk->token = req->token;
-	mptcp_token_accept(req, msk);
+	KUNIT_EXPECT_TRUE(test, mptcp_token_accept(req, msk));
 	KUNIT_EXPECT_PTR_EQ(test, msk, mptcp_token_get_sock(&init_net, msk->token));
 
 	/* this is now a no-op */

@@ -122,7 +122,7 @@ static void mptcp_token_test_destroyed(struct kunit *test)
 	KUNIT_ASSERT_EQ(test, 0,
 			mptcp_token_new_request((struct request_sock *)req));
 	msk->token = req->token;
-	mptcp_token_accept(req, msk);
+	KUNIT_EXPECT_TRUE(test, mptcp_token_accept(req, msk));
 
 	/* simulate race on removal */
 	refcount_set(&sk->sk_refcnt, 0);

2.43.0

Hi Ren, Ruide,

Thank you for the new version!

I have one major comment, please see below:

Sponsored by the NGI0 Core fund.
pw-bot: cr

Hi Matt,

Thank you for your comment. We are adding a
mptcp_sk_clone_destroy(nsk) in our v4.

Regarding the comments from Clashiko, my reply is as follows:

Comment 1:

I agree. Patch 1 only fixes the MP_JOIN msk reference ownership problem, but the same request clone helper also needs to handle MP_CAPABLE token_node ownership.

Comment 2:

I agree that passing nsk to tcp_done() was wrong. nsk is an MPTCP master socket, not a TCP socket. tcp_done() and inet_csk_prepare_forced_close() are TCP-only here because they use tcp_sk(), so they can interpret the private MPTCP socket layout as struct tcp_sock after the shared inet_connection_sock header.

We are reviewing v4, and will send it out soon.

Best,
Ruide

Hi Ruide,

Thank you for your modifications, that's great!

Our CI did some validations and here is its report:

- KVM Validation: normal (except selftest_mptcp_join): Success! ✅
- KVM Validation: normal (only selftest_mptcp_join): Success! ✅
- KVM Validation: debug (except selftest_mptcp_join): Success! ✅
- KVM Validation: debug (only selftest_mptcp_join): Success! ✅
- KVM Validation: btf-normal (only bpftest_all): Success! ✅
- KVM Validation: btf-debug (only bpftest_all): Success! ✅
- Task: https://github.com/multipath-tcp/mptcp_net-next/actions/runs/31098302528

Initiator: Patchew Applier
Commits: https://github.com/multipath-tcp/mptcp_net-next/commits/f809c35b85ca
Patchwork: https://patchwork.kernel.org/project/mptcp/list/?series=1141434

If there are some issues, you can reproduce them using the same environment as the one used by the CI thanks to a docker image, e.g.:

    $ cd [kernel source code]
    $ docker run -v "${PWD}:${PWD}:rw" -w "${PWD}" --privileged --rm -it \
        --pull always mptcp/mptcp-upstream-virtme-docker:latest \
        auto-normal

For more details:

https://github.com/multipath-tcp/mptcp-upstream-virtme-docker

Please note that despite all the efforts that have been already done to have a stable tests suite when executed on a public CI like here, it is possible some reported issues are not due to your modifications. Still, do not hesitate to help us improve that ;-)

Cheers,
MPTCP GH Action bot
Bot operated by Matthieu Baerts (NGI0 Core)

Hi Ren, Ruide,

Thank you for the v3! Here is a quick reaction, only about the reproducers, not about the kernel code:

On 06/08/2026 13:14, Ren Wei wrote:

(...)

(...)

Nice, very appreciated, thank you!

I confirm they work well!

Small details to help you next time to look at packetdrill reproducers for MPTCP:

- You can add new tests in gtests/net/mptcp/<dir>/<file>.pkt from the MPTCP packetdrill repository.
- Then here above, you can add: ../common/defaults.sh.
- And execute your tests with run_all.py from the gtests/net dir, e.g.:
./packetdrill/run_all.py -lv4 mptcp/<dir>/<file>.pkt

These tests will then be executed from a dedicated netns, using many sysctl settings: this helps to reproduce issues in a more controlled environment.

(detail) Here, tcp_synack_retries can be set to 0, then the sleep at the end can be reduced.

Also for next time, when validating the listener side, feel free to drop the TCP Timestamps option if it is not required (like what you did with the other test): that makes the .pkt file more readable.

(...)

Same here for ../common/defaults.sh. Then I guess you only need tcp_migrate_req=1 and tcp_synack_retries=2 (set to 0 to reduce the last wait), and "wscale" will always be 8 in the S(.) instead of depending on the environment.

Cheers,
Matt

Sponsored by the NGI0 Core fund.

Thanks so much for the detailed suggestions! We learned a lot from them, and we’ll definitely take them into account for our future patches.

We’ve also tried using packetdrill as the PoC for patches in other parts of the networking stack, but some cases don’t seem to be fully expressible with packetdrill. Anyway, we’ll try our best to use it whenever possible.