From: Menglong Dong <imagedong@tencent.com>
The mptcp socket and its subflow sockets in accept queue can't be
released after the process exit.
While the release of a mptcp socket in listening state, the
corresponding tcp socket will be released too. Meanwhile, the tcp
socket in the unaccept queue will be released too. However, only init
subflow is in the unaccept queue, and the joined subflow is not in the
unaccept queue, which makes the joined subflow won't be released, and
therefore the corresponding unaccepted mptcp socket will not be released
to.
This can be reproduced easily with following steps:
1. create 2 namespace and veth:
$ ip netns add mptcp-client
$ ip netns add mptcp-server
$ sysctl -w net.ipv4.conf.all.rp_filter=0
$ ip netns exec mptcp-client sysctl -w net.mptcp.enabled=1
$ ip netns exec mptcp-server sysctl -w net.mptcp.enabled=1
$ ip link add red-client netns mptcp-client type veth peer red-server \
netns mptcp-server
$ ip -n mptcp-server address add 10.0.0.1/24 dev red-server
$ ip -n mptcp-server address add 192.168.0.1/24 dev red-server
$ ip -n mptcp-client address add 10.0.0.2/24 dev red-client
$ ip -n mptcp-client address add 192.168.0.2/24 dev red-client
$ ip -n mptcp-server link set red-server up
$ ip -n mptcp-client link set red-client up
2. configure the endpoint and limit for client and server:
$ ip -n mptcp-server mptcp endpoint flush
$ ip -n mptcp-server mptcp limits set subflow 2 add_addr_accepted 2
$ ip -n mptcp-client mptcp endpoint flush
$ ip -n mptcp-client mptcp limits set subflow 2 add_addr_accepted 2
$ ip -n mptcp-client mptcp endpoint add 192.168.0.2 dev red-client id \
1 subflow
3. listen and accept on a port, such as 9999. The nc command we used
here is modified, which makes it use mptcp protocol by default.
$ ip netns exec mptcp-server nc -l -k -p 9999
4. open another *two* terminal and use each of them to connect to the
server with the following command:
$ ip netns exec mptcp-client nc 10.0.0.1 9999
Input something after connect to triger the connection of the second
subflow. So that there are two established mptcp connections, with the
second one still unaccepted.
5. exit all the nc command, and check the tcp socket in server namespace.
And you will find that there is one tcp socket in CLOSE_WAIT state
and can't release forever.
Fix this by closing all of the unaccepted mptcp socket in
mptcp_subflow_queue_clean() with __mptcp_close().
Now, we can ensure that all unaccepted mptcp sockets will be cleaned by
__mptcp_close() before they are released, so mptcp_sock_destruct(), which
is used to clean the unaccepted mptcp socket, is not needed anymore.
The selftests for mptcp is ran for this commit, and no new failures.
Fixes: f296234c98a8 ("mptcp: Add handling of incoming MP_JOIN requests")
Reviewed-by: Jiang Biao <benbjiang@tencent.com>
Reviewed-by: Mengen Sun <mengensun@tencent.com>
Signed-off-by: Menglong Dong <imagedong@tencent.com>
---
v4:
- factor-out __mptcp_close() and use it in this commit.
v3:
- remove mptcp_close_nolock() and call mptcp_close() directly in
mptcp_subflow_queue_clean(), as mptcp_close_nolock() will cause
dead lock.
v2:
- remove mptcp_sock_destruct()
- introduce mptcp_close_nolock() and replace mptcp_close() with it in
mptcp_subflow_queue_clean()
---
net/mptcp/subflow.c | 37 +++++++++++--------------------------
1 file changed, 11 insertions(+), 26 deletions(-)
diff --git a/net/mptcp/subflow.c b/net/mptcp/subflow.c
index c7d49fb6e7bd..7c437debc368 100644
--- a/net/mptcp/subflow.c
+++ b/net/mptcp/subflow.c
@@ -602,30 +602,6 @@ static bool subflow_hmac_valid(const struct request_sock *req,
return !crypto_memneq(hmac, mp_opt->hmac, MPTCPOPT_HMAC_LEN);
}
-static void mptcp_sock_destruct(struct sock *sk)
-{
- /* if new mptcp socket isn't accepted, it is free'd
- * from the tcp listener sockets request queue, linked
- * from req->sk. The tcp socket is released.
- * This calls the ULP release function which will
- * also remove the mptcp socket, via
- * sock_put(ctx->conn).
- *
- * Problem is that the mptcp socket will be in
- * ESTABLISHED state and will not have the SOCK_DEAD flag.
- * Both result in warnings from inet_sock_destruct.
- */
- if ((1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
- sk->sk_state = TCP_CLOSE;
- WARN_ON_ONCE(sk->sk_socket);
- sock_orphan(sk);
- }
-
- /* We don't need to clear msk->subflow, as it's still NULL at this point */
- mptcp_destroy_common(mptcp_sk(sk), 0);
- inet_sock_destruct(sk);
-}
-
static void mptcp_force_close(struct sock *sk)
{
/* the msk is not yet exposed to user-space */
@@ -768,7 +744,6 @@ static struct sock *subflow_syn_recv_sock(const struct sock *sk,
/* new mpc subflow takes ownership of the newly
* created mptcp socket
*/
- new_msk->sk_destruct = mptcp_sock_destruct;
mptcp_sk(new_msk)->setsockopt_seq = ctx->setsockopt_seq;
mptcp_pm_new_connection(mptcp_sk(new_msk), child, 1);
mptcp_token_accept(subflow_req, mptcp_sk(new_msk));
@@ -1763,13 +1738,23 @@ void mptcp_subflow_queue_clean(struct sock *listener_ssk)
for (msk = head; msk; msk = next) {
struct sock *sk = (struct sock *)msk;
- bool slow;
+ bool slow, cancel_work;
+ sock_hold(sk);
slow = lock_sock_fast_nested(sk);
next = msk->dl_next;
msk->first = NULL;
msk->dl_next = NULL;
+
+ /* __mptcp_close() will put a extra reference on sk,
+ * so we hold one here.
+ */
+ sock_hold(sk);
+ cancel_work = __mptcp_close(sk, 0);
unlock_sock_fast(sk, slow);
+ if (cancel_work)
+ mptcp_cancel_work(sk);
+ sock_put(sk);
}
/* we are still under the listener msk socket lock */
--
2.37.2On Fri, 2022-09-09 at 13:49 +0800, menglong8.dong@gmail.com wrote:
> From: Menglong Dong <imagedong@tencent.com>
>
> The mptcp socket and its subflow sockets in accept queue can't be
> released after the process exit.
>
> While the release of a mptcp socket in listening state, the
> corresponding tcp socket will be released too. Meanwhile, the tcp
> socket in the unaccept queue will be released too. However, only init
> subflow is in the unaccept queue, and the joined subflow is not in the
> unaccept queue, which makes the joined subflow won't be released, and
> therefore the corresponding unaccepted mptcp socket will not be released
> to.
>
> This can be reproduced easily with following steps:
>
> 1. create 2 namespace and veth:
> $ ip netns add mptcp-client
> $ ip netns add mptcp-server
> $ sysctl -w net.ipv4.conf.all.rp_filter=0
> $ ip netns exec mptcp-client sysctl -w net.mptcp.enabled=1
> $ ip netns exec mptcp-server sysctl -w net.mptcp.enabled=1
> $ ip link add red-client netns mptcp-client type veth peer red-server \
> netns mptcp-server
> $ ip -n mptcp-server address add 10.0.0.1/24 dev red-server
> $ ip -n mptcp-server address add 192.168.0.1/24 dev red-server
> $ ip -n mptcp-client address add 10.0.0.2/24 dev red-client
> $ ip -n mptcp-client address add 192.168.0.2/24 dev red-client
> $ ip -n mptcp-server link set red-server up
> $ ip -n mptcp-client link set red-client up
>
> 2. configure the endpoint and limit for client and server:
> $ ip -n mptcp-server mptcp endpoint flush
> $ ip -n mptcp-server mptcp limits set subflow 2 add_addr_accepted 2
> $ ip -n mptcp-client mptcp endpoint flush
> $ ip -n mptcp-client mptcp limits set subflow 2 add_addr_accepted 2
> $ ip -n mptcp-client mptcp endpoint add 192.168.0.2 dev red-client id \
> 1 subflow
>
> 3. listen and accept on a port, such as 9999. The nc command we used
> here is modified, which makes it use mptcp protocol by default.
> $ ip netns exec mptcp-server nc -l -k -p 9999
>
> 4. open another *two* terminal and use each of them to connect to the
> server with the following command:
> $ ip netns exec mptcp-client nc 10.0.0.1 9999
> Input something after connect to triger the connection of the second
> subflow. So that there are two established mptcp connections, with the
> second one still unaccepted.
>
> 5. exit all the nc command, and check the tcp socket in server namespace.
> And you will find that there is one tcp socket in CLOSE_WAIT state
> and can't release forever.
>
> Fix this by closing all of the unaccepted mptcp socket in
> mptcp_subflow_queue_clean() with __mptcp_close().
>
> Now, we can ensure that all unaccepted mptcp sockets will be cleaned by
> __mptcp_close() before they are released, so mptcp_sock_destruct(), which
> is used to clean the unaccepted mptcp socket, is not needed anymore.
>
> The selftests for mptcp is ran for this commit, and no new failures.
>
> Fixes: f296234c98a8 ("mptcp: Add handling of incoming MP_JOIN requests")
> Reviewed-by: Jiang Biao <benbjiang@tencent.com>
> Reviewed-by: Mengen Sun <mengensun@tencent.com>
> Signed-off-by: Menglong Dong <imagedong@tencent.com>
> ---
> v4:
> - factor-out __mptcp_close() and use it in this commit.
>
> v3:
> - remove mptcp_close_nolock() and call mptcp_close() directly in
> mptcp_subflow_queue_clean(), as mptcp_close_nolock() will cause
> dead lock.
>
> v2:
> - remove mptcp_sock_destruct()
> - introduce mptcp_close_nolock() and replace mptcp_close() with it in
> mptcp_subflow_queue_clean()
> ---
> net/mptcp/subflow.c | 37 +++++++++++--------------------------
> 1 file changed, 11 insertions(+), 26 deletions(-)
>
> diff --git a/net/mptcp/subflow.c b/net/mptcp/subflow.c
> index c7d49fb6e7bd..7c437debc368 100644
> --- a/net/mptcp/subflow.c
> +++ b/net/mptcp/subflow.c
> @@ -602,30 +602,6 @@ static bool subflow_hmac_valid(const struct request_sock *req,
> return !crypto_memneq(hmac, mp_opt->hmac, MPTCPOPT_HMAC_LEN);
> }
>
> -static void mptcp_sock_destruct(struct sock *sk)
> -{
> - /* if new mptcp socket isn't accepted, it is free'd
> - * from the tcp listener sockets request queue, linked
> - * from req->sk. The tcp socket is released.
> - * This calls the ULP release function which will
> - * also remove the mptcp socket, via
> - * sock_put(ctx->conn).
> - *
> - * Problem is that the mptcp socket will be in
> - * ESTABLISHED state and will not have the SOCK_DEAD flag.
> - * Both result in warnings from inet_sock_destruct.
> - */
> - if ((1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
> - sk->sk_state = TCP_CLOSE;
> - WARN_ON_ONCE(sk->sk_socket);
> - sock_orphan(sk);
> - }
> -
> - /* We don't need to clear msk->subflow, as it's still NULL at this point */
> - mptcp_destroy_common(mptcp_sk(sk), 0);
> - inet_sock_destruct(sk);
> -}
> -
> static void mptcp_force_close(struct sock *sk)
> {
> /* the msk is not yet exposed to user-space */
> @@ -768,7 +744,6 @@ static struct sock *subflow_syn_recv_sock(const struct sock *sk,
> /* new mpc subflow takes ownership of the newly
> * created mptcp socket
> */
> - new_msk->sk_destruct = mptcp_sock_destruct;
> mptcp_sk(new_msk)->setsockopt_seq = ctx->setsockopt_seq;
> mptcp_pm_new_connection(mptcp_sk(new_msk), child, 1);
> mptcp_token_accept(subflow_req, mptcp_sk(new_msk));
> @@ -1763,13 +1738,23 @@ void mptcp_subflow_queue_clean(struct sock *listener_ssk)
>
> for (msk = head; msk; msk = next) {
> struct sock *sk = (struct sock *)msk;
> - bool slow;
> + bool slow, cancel_work;
I think is better to rename 'cancel_work' to 'do_cancel_work'
>
> + sock_hold(sk);
> slow = lock_sock_fast_nested(sk);
> next = msk->dl_next;
> msk->first = NULL;
> msk->dl_next = NULL;
> +
> + /* __mptcp_close() will put a extra reference on sk,
> + * so we hold one here.
> + */
> + sock_hold(sk);
With the suggested changes to patch 1 one of the above 2 sock_hold()
will not be needed anymore.
Thanks,
Paolo
© 2016 - 2026 Red Hat, Inc.