[PATCH net v3 0/2] tcp: fix listener wakeup after reuseport migration

[PATCH net v3 0/2] tcp: fix listener wakeup after reuseport migration

[Zhenzhong Wu]

This series fixes a missing wakeup when inet_csk_listen_stop() migrates
an established child socket from a closing listener to another socket
in the same SO_REUSEPORT group after the child has already been queued
for accept.

The target listener receives the migrated accept-queue entry via
inet_csk_reqsk_queue_add(), but its waiters are not notified.
Nonblocking accept() still succeeds because it checks the accept queue
directly, but readiness-based waiters can remain asleep until another
connection generates a wakeup.

Patch 1 notifies the target listener after a successful migration in
inet_csk_listen_stop() and protects the post-queue_add() nsk accesses
with rcu_read_lock()/rcu_read_unlock().

Patch 2 extends the existing migrate_reuseport BPF selftest with epoll
readiness checks inside migrate_dance(), around shutdown() where the
migration happens. The test now verifies that the target listener is
not ready before migration and becomes ready immediately after it, for
both TCP_ESTABLISHED and TCP_SYN_RECV. TCP_NEW_SYN_RECV remains
excluded because it still depends on later handshake completion.

Testing:
- On a local unpatched kernel, the focused migrate_reuseport test
  fails for the listener-migration cases and passes for the
  TCP_NEW_SYN_RECV cases:
    not ok 1 IPv4 TCP_ESTABLISHED  inet_csk_listen_stop
    not ok 2 IPv4 TCP_SYN_RECV     inet_csk_listen_stop
    ok 3 IPv4 TCP_NEW_SYN_RECV reqsk_timer_handler
    ok 4 IPv4 TCP_NEW_SYN_RECV inet_csk_complete_hashdance
    not ok 5 IPv6 TCP_ESTABLISHED  inet_csk_listen_stop
    not ok 6 IPv6 TCP_SYN_RECV     inet_csk_listen_stop
    ok 7 IPv6 TCP_NEW_SYN_RECV reqsk_timer_handler
    ok 8 IPv6 TCP_NEW_SYN_RECV inet_csk_complete_hashdance
- On a patched kernel booted under QEMU, the full migrate_reuseport
  selftest passes:
    ok 1 IPv4 TCP_ESTABLISHED  inet_csk_listen_stop
    ok 2 IPv4 TCP_SYN_RECV     inet_csk_listen_stop
    ok 3 IPv4 TCP_NEW_SYN_RECV reqsk_timer_handler
    ok 4 IPv4 TCP_NEW_SYN_RECV inet_csk_complete_hashdance
    ok 5 IPv6 TCP_ESTABLISHED  inet_csk_listen_stop
    ok 6 IPv6 TCP_SYN_RECV     inet_csk_listen_stop
    ok 7 IPv6 TCP_NEW_SYN_RECV reqsk_timer_handler
    ok 8 IPv6 TCP_NEW_SYN_RECV inet_csk_complete_hashdance
    SELFTEST_RC=0

---

v3:
- update the selftest patch as suggested by Kuniyuki Iwashima
- update the test flow comment to match the new epoll checks
- initialize epoll to -1 to avoid a compiler warning in the selftest

v2:
  https://lore.kernel.org/netdev/20260418181333.1713389-1-jt26wzz@gmail.com/

v1:
  https://lore.kernel.org/netdev/20260418041633.691435-1-jt26wzz@gmail.com/

Zhenzhong Wu (2):
  tcp: call sk_data_ready() after listener migration
  selftests/bpf: check epoll readiness during reuseport migration

 net/ipv4/inet_connection_sock.c               |  3 ++
 .../bpf/prog_tests/migrate_reuseport.c        | 46 ++++++++++++++++---
 2 files changed, 43 insertions(+), 6 deletions(-)


base-commit: 52bcb57a4e8a0865a76c587c2451906342ae1b2d
-- 
2.43.0

[PATCH net v3 1/2] tcp: call sk_data_ready() after listener migration

[Zhenzhong Wu]

When inet_csk_listen_stop() migrates an established child socket from
a closing listener to another socket in the same SO_REUSEPORT group,
the target listener gets a new accept-queue entry via
inet_csk_reqsk_queue_add(), but that path never notifies the target
listener's waiters. A nonblocking accept() still works because it
checks the queue directly, but poll()/epoll_wait() waiters and
blocking accept() callers can also remain asleep indefinitely.

Call READ_ONCE(nsk->sk_data_ready)(nsk) after a successful migration
in inet_csk_listen_stop().

However, after inet_csk_reqsk_queue_add() succeeds, the ref acquired
in reuseport_migrate_sock() is effectively transferred to
nreq->rsk_listener. Another CPU can then dequeue nreq via accept()
or listener shutdown, hit reqsk_put(), and drop that listener ref.
Since listeners are SOCK_RCU_FREE, wrap the post-queue_add()
dereferences of nsk in rcu_read_lock()/rcu_read_unlock(), which also
covers the existing sock_net(nsk) access in that path.

The reqsk_timer_handler() path does not need the same changes for two
reasons: half-open requests become readable only after the final ACK,
where tcp_child_process() already wakes the listener; and once nreq is
visible via inet_ehash_insert(), the success path no longer touches
nsk directly.

Fixes: 54b92e841937 ("tcp: Migrate TCP_ESTABLISHED/TCP_SYN_RECV sockets in accept queues.")
Cc: stable@vger.kernel.org
Suggested-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Zhenzhong Wu <jt26wzz@gmail.com>
---
 net/ipv4/inet_connection_sock.c | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/net/ipv4/inet_connection_sock.c b/net/ipv4/inet_connection_sock.c
index 4ac3ae1bc..928654c34 100644
--- a/net/ipv4/inet_connection_sock.c
+++ b/net/ipv4/inet_connection_sock.c
@@ -1479,16 +1479,19 @@ void inet_csk_listen_stop(struct sock *sk)
 			if (nreq) {
 				refcount_set(&nreq->rsk_refcnt, 1);
 
+				rcu_read_lock();
 				if (inet_csk_reqsk_queue_add(nsk, nreq, child)) {
 					__NET_INC_STATS(sock_net(nsk),
 							LINUX_MIB_TCPMIGRATEREQSUCCESS);
 					reqsk_migrate_reset(req);
+					READ_ONCE(nsk->sk_data_ready)(nsk);
 				} else {
 					__NET_INC_STATS(sock_net(nsk),
 							LINUX_MIB_TCPMIGRATEREQFAILURE);
 					reqsk_migrate_reset(nreq);
 					__reqsk_free(nreq);
 				}
+				rcu_read_unlock();
 
 				/* inet_csk_reqsk_queue_add() has already
 				 * called inet_child_forget() on failure case.
-- 
2.43.0

[PATCH net v3 2/2] selftests/bpf: check epoll readiness during reuseport migration

[Zhenzhong Wu]

Inside migrate_dance(), add epoll checks around shutdown() to
verify that the target listener is not ready before shutdown()
and becomes ready immediately after shutdown() triggers migration.

Cover TCP_ESTABLISHED and TCP_SYN_RECV. Exclude TCP_NEW_SYN_RECV
as it depends on later handshake completion.

Suggested-by: Kuniyuki Iwashima <kuniyu@google.com>
Signed-off-by: Zhenzhong Wu <jt26wzz@gmail.com>
---
 .../bpf/prog_tests/migrate_reuseport.c        | 46 ++++++++++++++++---
 1 file changed, 40 insertions(+), 6 deletions(-)

diff --git a/tools/testing/selftests/bpf/prog_tests/migrate_reuseport.c b/tools/testing/selftests/bpf/prog_tests/migrate_reuseport.c
index 653b0a20f..6180a79a7 100644
--- a/tools/testing/selftests/bpf/prog_tests/migrate_reuseport.c
+++ b/tools/testing/selftests/bpf/prog_tests/migrate_reuseport.c
@@ -7,24 +7,29 @@
  *   3. call listen() for 1 server socket. (migration target)
  *   4. update a map to migrate all child sockets
  *        to the last server socket (migrate_map[cookie] = 4)
- *   5. call shutdown() for first 4 server sockets
+ *   5. for TCP_ESTABLISHED and TCP_SYN_RECV cases, verify via epoll
+ *        that the last server socket is not ready before migration.
+ *   6. call shutdown() for first 4 server sockets
  *        and migrate the requests in the accept queue
  *        to the last server socket.
- *   6. call listen() for the second server socket.
- *   7. call shutdown() for the last server
+ *   7. for TCP_ESTABLISHED and TCP_SYN_RECV cases, verify via epoll
+ *        that the last server socket is ready after migration.
+ *   8. call listen() for the second server socket.
+ *   9. call shutdown() for the last server
  *        and migrate the requests in the accept queue
  *        to the second server socket.
- *   8. call listen() for the last server.
- *   9. call shutdown() for the second server
+ *  10. call listen() for the last server.
+ *  11. call shutdown() for the second server
  *        and migrate the requests in the accept queue
  *        to the last server socket.
- *  10. call accept() for the last server socket.
+ *  12. call accept() for the last server socket.
  *
  * Author: Kuniyuki Iwashima <kuniyu@amazon.co.jp>
  */
 
 #include <bpf/bpf.h>
 #include <bpf/libbpf.h>
+#include <sys/epoll.h>
 
 #include "test_progs.h"
 #include "test_migrate_reuseport.skel.h"
@@ -350,8 +355,28 @@ static int update_maps(struct migrate_reuseport_test_case *test_case,
 
 static int migrate_dance(struct migrate_reuseport_test_case *test_case)
 {
+	struct epoll_event ev = {
+		.events = EPOLLIN,
+	};
+	int epoll = -1, nfds;
 	int i, err;
 
+	if (test_case->state != BPF_TCP_NEW_SYN_RECV) {
+		epoll = epoll_create1(0);
+		if (!ASSERT_NEQ(epoll, -1, "epoll_create1"))
+			return -1;
+
+		ev.data.fd = test_case->servers[MIGRATED_TO];
+		if (!ASSERT_OK(epoll_ctl(epoll, EPOLL_CTL_ADD,
+					 test_case->servers[MIGRATED_TO], &ev),
+			       "epoll_ctl"))
+			goto close_epoll;
+
+		nfds = epoll_wait(epoll, &ev, 1, 0);
+		if (!ASSERT_EQ(nfds, 0, "epoll_wait 1"))
+			goto close_epoll;
+	}
+
 	/* Migrate TCP_ESTABLISHED and TCP_SYN_RECV requests
 	 * to the last listener based on eBPF.
 	 */
@@ -365,6 +390,15 @@ static int migrate_dance(struct migrate_reuseport_test_case *test_case)
 	if (test_case->state == BPF_TCP_NEW_SYN_RECV)
 		return 0;
 
+	nfds = epoll_wait(epoll, &ev, 1, 0);
+	if (!ASSERT_EQ(nfds, 1, "epoll_wait 2")) {
+close_epoll:
+		close(epoll);
+		return -1;
+	}
+
+	close(epoll);
+
 	/* Note that we use the second listener instead of the
 	 * first one here.
 	 *
-- 
2.43.0