[RFC PATCH 0/3] mptcp: improve accept() and disconnect()

[RFC PATCH 0/3] mptcp: improve accept() and disconnect()

[Paolo Abeni]

As outlined in the public mtg, mptcp_accept() is currently quite
suboptimal, both from performance and code complexity

This series tries to clean it up, enforcing a wider lifetime for
the initial subflow, so that we don't need to acquire additional
references there.

To reach such goal we need to properly define the disconnect()
behavior, which is currently quite incomplete.

Some additional self-tests is likely required, 

Paolo Abeni (3):
  mptcp: never allow the PM to close a listener subflow
  mptcp: full disconnect implementation
  mptcp: cleanup accept and poll

 net/mptcp/pm.c         |  10 ++--
 net/mptcp/pm_netlink.c |   3 ++
 net/mptcp/protocol.c   | 120 ++++++++++++++++++++++++++---------------
 net/mptcp/protocol.h   |  15 +++++-
 net/mptcp/subflow.c    |   1 -
 net/mptcp/token.c      |   1 +
 6 files changed, 102 insertions(+), 48 deletions(-)

-- 
2.26.3

[RFC PATCH 1/3] mptcp: never allow the PM to close a listener subflow

[Paolo Abeni]

Currently, when deleting an endpoint the netlink PM treverses
all the local MPTCP sockets, regardless of their status.

If an MPTCP listener socket is bound to the IP matching the
delete endpoint, the listener TCP socket will be closed.
That is unexpected, the PM should only affect data subflows.

Fix the issue explicitly skipping MPTCP socket in TCP_LISTEN
status.

Signed-off-by: Paolo Abeni <pabeni@redhat.com>
---
 net/mptcp/pm_netlink.c | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/net/mptcp/pm_netlink.c b/net/mptcp/pm_netlink.c
index 7b96be1e9f14..f523051f5aef 100644
--- a/net/mptcp/pm_netlink.c
+++ b/net/mptcp/pm_netlink.c
@@ -700,6 +700,9 @@ static void mptcp_pm_nl_rm_addr_or_subflow(struct mptcp_sock *msk,
 
 	msk_owned_by_me(msk);
 
+	if (sk->sk_state == TCP_LISTEN)
+		return;
+
 	if (!rm_list->nr)
 		return;
 
-- 
2.26.3

[RFC PATCH 2/3] mptcp: full disconnect implementation

[Paolo Abeni]

The current mptcp_disconnect() implementation lacks several
steps, we additionally need to reset the msk socket state
and flush the subflow list.

Factor out the needed helper to avoid code duplication.

Additionally ensure that the initial subflow is disposed
only after mptcp_close(), just reset it at disconnect time.

Signed-off-by: Paolo Abeni <pabeni@redhat.com>
---
 net/mptcp/pm.c       | 10 +++--
 net/mptcp/protocol.c | 95 ++++++++++++++++++++++++++++++++------------
 net/mptcp/protocol.h | 14 +++++++
 net/mptcp/token.c    |  1 +
 4 files changed, 92 insertions(+), 28 deletions(-)

diff --git a/net/mptcp/pm.c b/net/mptcp/pm.c
index 6ab386ff3294..6b2bfe89d445 100644
--- a/net/mptcp/pm.c
+++ b/net/mptcp/pm.c
@@ -356,7 +356,7 @@ void mptcp_pm_subflow_chk_stale(const struct mptcp_sock *msk, struct sock *ssk)
 	}
 }
 
-void mptcp_pm_data_init(struct mptcp_sock *msk)
+void mptcp_pm_data_reset(struct mptcp_sock *msk)
 {
 	msk->pm.add_addr_signaled = 0;
 	msk->pm.add_addr_accepted = 0;
@@ -371,10 +371,14 @@ void mptcp_pm_data_init(struct mptcp_sock *msk)
 	WRITE_ONCE(msk->pm.remote_deny_join_id0, false);
 	msk->pm.status = 0;
 
+	mptcp_pm_nl_data_init(msk);
+}
+
+void mptcp_pm_data_init(struct mptcp_sock *msk)
+{
 	spin_lock_init(&msk->pm.lock);
 	INIT_LIST_HEAD(&msk->pm.anno_list);
-
-	mptcp_pm_nl_data_init(msk);
+	mptcp_pm_data_reset(msk);
 }
 
 void __init mptcp_pm_init(void)
diff --git a/net/mptcp/protocol.c b/net/mptcp/protocol.c
index 1ee3530ac2a3..5fe19adb88a5 100644
--- a/net/mptcp/protocol.c
+++ b/net/mptcp/protocol.c
@@ -2226,22 +2226,33 @@ bool __mptcp_retransmit_pending_data(struct sock *sk)
  * parent socket.
  */
 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
-			      struct mptcp_subflow_context *subflow)
+			      struct mptcp_subflow_context *subflow,
+			      bool check_pending)
 {
 	struct mptcp_sock *msk = mptcp_sk(sk);
-	bool need_push;
+	bool need_push, dispose_it;
 
-	list_del(&subflow->node);
+	dispose_it = !msk->subflow || ssk != msk->subflow->sk;
+	if (dispose_it)
+		list_del(&subflow->node);
 
 	lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
 
+	need_push = check_pending && __mptcp_retransmit_pending_data(sk);
+	if (!dispose_it) {
+		tcp_disconnect(ssk, 0);
+		mptcp_subflow_ctx_reset(subflow);
+		release_sock(ssk);
+
+		goto out;
+	}
+
 	/* if we are invoked by the msk cleanup code, the subflow is
 	 * already orphaned
 	 */
 	if (ssk->sk_socket)
 		sock_orphan(ssk);
 
-	need_push = __mptcp_retransmit_pending_data(sk);
 	subflow->disposable = 1;
 
 	/* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
@@ -2261,14 +2272,12 @@ static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
 
 	sock_put(ssk);
 
-	if (ssk == msk->last_snd)
-		msk->last_snd = NULL;
-
 	if (ssk == msk->first)
 		msk->first = NULL;
 
-	if (msk->subflow && ssk == msk->subflow->sk)
-		mptcp_dispose_initial_subflow(msk);
+out:
+	if (ssk == msk->last_snd)
+		msk->last_snd = NULL;
 
 	if (need_push)
 		__mptcp_push_pending(sk, 0);
@@ -2279,7 +2288,7 @@ void mptcp_close_ssk(struct sock *sk, struct sock *ssk,
 {
 	if (sk->sk_state == TCP_ESTABLISHED)
 		mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL);
-	__mptcp_close_ssk(sk, ssk, subflow);
+	__mptcp_close_ssk(sk, ssk, subflow, true);
 }
 
 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
@@ -2501,9 +2510,20 @@ static int __mptcp_init_sock(struct sock *sk)
 	return 0;
 }
 
-static int mptcp_init_sock(struct sock *sk)
+static void mptcp_ca_reset(struct sock *sk)
 {
 	struct inet_connection_sock *icsk = inet_csk(sk);
+
+	tcp_assign_congestion_control(sk);
+	strcpy(mptcp_sk(sk)->ca_name, icsk->icsk_ca_ops->name);
+
+	/* no need to keep a reference to the ops, the name will suffice */
+	tcp_cleanup_congestion_control(sk);
+	icsk->icsk_ca_ops = NULL;
+}
+
+static int mptcp_init_sock(struct sock *sk)
+{
 	struct net *net = sock_net(sk);
 	int ret;
 
@@ -2524,12 +2544,7 @@ static int mptcp_init_sock(struct sock *sk)
 	/* fetch the ca name; do it outside __mptcp_init_sock(), so that clone will
 	 * propagate the correct value
 	 */
-	tcp_assign_congestion_control(sk);
-	strcpy(mptcp_sk(sk)->ca_name, icsk->icsk_ca_ops->name);
-
-	/* no need to keep a reference to the ops, the name will suffice */
-	tcp_cleanup_congestion_control(sk);
-	icsk->icsk_ca_ops = NULL;
+	mptcp_ca_reset(sk);
 
 	sk_sockets_allocated_inc(sk);
 	sk->sk_rcvbuf = sock_net(sk)->ipv4.sysctl_tcp_rmem[1];
@@ -2687,9 +2702,13 @@ static void __mptcp_destroy_sock(struct sock *sk)
 	sk_stop_timer(sk, &sk->sk_timer);
 	msk->pm.status = 0;
 
+	/* clears msk->subflow, allowing the following loop to close
+	 * even the initial subflow
+	 */
+	mptcp_dispose_initial_subflow(msk);
 	list_for_each_entry_safe(subflow, tmp, &conn_list, node) {
 		struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
-		__mptcp_close_ssk(sk, ssk, subflow);
+		__mptcp_close_ssk(sk, ssk, subflow, false);
 	}
 
 	sk->sk_prot->destroy(sk);
@@ -2700,7 +2719,6 @@ static void __mptcp_destroy_sock(struct sock *sk)
 	xfrm_sk_free_policy(sk);
 
 	sk_refcnt_debug_release(sk);
-	mptcp_dispose_initial_subflow(msk);
 	sock_put(sk);
 }
 
@@ -2736,6 +2754,9 @@ static void mptcp_close(struct sock *sk, long timeout)
 
 	sock_hold(sk);
 	pr_debug("msk=%p state=%d", sk, sk->sk_state);
+	if (mptcp_sk(sk)->token)
+		mptcp_event(MPTCP_EVENT_CLOSED, mptcp_sk(sk), NULL, GFP_KERNEL);
+
 	if (sk->sk_state == TCP_CLOSE) {
 		__mptcp_destroy_sock(sk);
 		do_cancel_work = true;
@@ -2746,9 +2767,6 @@ static void mptcp_close(struct sock *sk, long timeout)
 	if (do_cancel_work)
 		mptcp_cancel_work(sk);
 
-	if (mptcp_sk(sk)->token)
-		mptcp_event(MPTCP_EVENT_CLOSED, mptcp_sk(sk), NULL, GFP_KERNEL);
-
 	sock_put(sk);
 }
 
@@ -2782,13 +2800,40 @@ static int mptcp_disconnect(struct sock *sk, int flags)
 
 	mptcp_do_flush_join_list(msk);
 
+	inet_sk_state_store(sk, TCP_CLOSE);
+
 	mptcp_for_each_subflow(msk, subflow) {
 		struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
 
-		lock_sock(ssk);
-		tcp_disconnect(ssk, flags);
-		release_sock(ssk);
+		__mptcp_close_ssk(sk, ssk, subflow, false);
 	}
+
+	sk_stop_timer(sk, &msk->sk.icsk_retransmit_timer);
+	sk_stop_timer(sk, &sk->sk_timer);
+
+	if (mptcp_sk(sk)->token)
+		mptcp_event(MPTCP_EVENT_CLOSED, mptcp_sk(sk), NULL, GFP_KERNEL);
+
+	mptcp_destroy_common(msk);
+	msk->last_snd = 0;
+	msk->flags = 0;
+	msk->recovery = false;
+	msk->can_ack = false;
+	msk->fully_established = false;
+	msk->rcv_data_fin = false;
+	msk->snd_data_fin_enable = false;
+	msk->rcv_fastclose = false;
+	msk->use_64bit_ack = false;
+	WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
+	mptcp_pm_data_reset(msk);
+	mptcp_ca_reset(sk);
+
+#if IS_ENABLED(CONFIG_KASAN)
+	sock_set_flag(sk, SOCK_RCU_FREE);
+#else
+	sock_reset_flag(sk, SOCK_RCU_FREE);
+#endif
+
 	return 0;
 }
 
diff --git a/net/mptcp/protocol.h b/net/mptcp/protocol.h
index f6f4054e19ee..188b9e044672 100644
--- a/net/mptcp/protocol.h
+++ b/net/mptcp/protocol.h
@@ -392,6 +392,9 @@ DECLARE_PER_CPU(struct mptcp_delegated_action, mptcp_delegated_actions);
 /* MPTCP subflow context */
 struct mptcp_subflow_context {
 	struct	list_head node;/* conn_list of subflows */
+
+	char	reset_start[0];
+
 	unsigned long avg_pacing_rate; /* protected by msk socket lock */
 	u64	local_key;
 	u64	remote_key;
@@ -438,6 +441,9 @@ struct mptcp_subflow_context {
 	u8	stale_count;
 
 	long	delegated_status;
+
+	char	reset_end[0];
+
 	struct	list_head delegated_node;   /* link into delegated_action, protected by local BH */
 
 	u32	setsockopt_seq;
@@ -469,6 +475,13 @@ mptcp_subflow_tcp_sock(const struct mptcp_subflow_context *subflow)
 	return subflow->tcp_sock;
 }
 
+static inline void
+mptcp_subflow_ctx_reset(struct mptcp_subflow_context *subflow)
+{
+	memset(subflow->reset_start, 0, subflow->reset_end - subflow->reset_start);
+	subflow->request_mptcp = 1;
+}
+
 static inline u64
 mptcp_subflow_get_map_offset(const struct mptcp_subflow_context *subflow)
 {
@@ -708,6 +721,7 @@ void mptcp_crypto_hmac_sha(u64 key1, u64 key2, u8 *msg, int len, void *hmac);
 
 void __init mptcp_pm_init(void);
 void mptcp_pm_data_init(struct mptcp_sock *msk);
+void mptcp_pm_data_reset(struct mptcp_sock *msk);
 void mptcp_pm_subflow_chk_stale(const struct mptcp_sock *msk, struct sock *ssk);
 void mptcp_pm_nl_subflow_chk_stale(const struct mptcp_sock *msk, struct sock *ssk);
 void mptcp_pm_new_connection(struct mptcp_sock *msk, const struct sock *ssk, int server_side);
diff --git a/net/mptcp/token.c b/net/mptcp/token.c
index e581b341c5be..f52ee7b26aed 100644
--- a/net/mptcp/token.c
+++ b/net/mptcp/token.c
@@ -384,6 +384,7 @@ void mptcp_token_destroy(struct mptcp_sock *msk)
 		bucket->chain_len--;
 	}
 	spin_unlock_bh(&bucket->lock);
+	WRITE_ONCE(msk->token, 0);
 }
 
 void __init mptcp_token_init(void)
-- 
2.26.3

[RFC PATCH 3/3] mptcp: cleanup accept and poll

[Paolo Abeni]

After the previous patch,  msk->subflow will never be deleted during
the whole msk lifetime. We don't need anymore to acquire references to
it in mptcp_stream_accept() and we can use the listener subflow accept
queue to simplify mptcp_poll() for listener socket.

Overall this removes a lock pair and 4 more atomic operations per accept().

Signed-off-by: Paolo Abeni <pabeni@redhat.com>
---
 net/mptcp/protocol.c | 25 +++++++------------------
 net/mptcp/protocol.h |  1 -
 net/mptcp/subflow.c  |  1 -
 3 files changed, 7 insertions(+), 20 deletions(-)

diff --git a/net/mptcp/protocol.c b/net/mptcp/protocol.c
index 5fe19adb88a5..38f5b4bcba69 100644
--- a/net/mptcp/protocol.c
+++ b/net/mptcp/protocol.c
@@ -3402,17 +3402,9 @@ static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
 
 	pr_debug("msk=%p", msk);
 
-	lock_sock(sock->sk);
-	if (sock->sk->sk_state != TCP_LISTEN)
-		goto unlock_fail;
-
 	ssock = __mptcp_nmpc_socket(msk);
 	if (!ssock)
-		goto unlock_fail;
-
-	clear_bit(MPTCP_DATA_READY, &msk->flags);
-	sock_hold(ssock->sk);
-	release_sock(sock->sk);
+		return -EINVAL;
 
 	err = ssock->ops->accept(sock, newsock, flags, kern);
 	if (err == 0 && !mptcp_is_tcpsk(newsock->sk)) {
@@ -3452,14 +3444,7 @@ static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
 		release_sock(newsk);
 	}
 
-	if (inet_csk_listen_poll(ssock->sk))
-		set_bit(MPTCP_DATA_READY, &msk->flags);
-	sock_put(ssock->sk);
 	return err;
-
-unlock_fail:
-	release_sock(sock->sk);
-	return -EINVAL;
 }
 
 static __poll_t mptcp_check_readable(struct mptcp_sock *msk)
@@ -3505,8 +3490,12 @@ static __poll_t mptcp_poll(struct file *file, struct socket *sock,
 
 	state = inet_sk_state_load(sk);
 	pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags);
-	if (state == TCP_LISTEN)
-		return test_bit(MPTCP_DATA_READY, &msk->flags) ? EPOLLIN | EPOLLRDNORM : 0;
+	if (state == TCP_LISTEN) {
+		if (WARN_ON_ONCE(!msk->subflow || !msk->subflow->sk))
+			return 0;
+
+		return inet_csk_listen_poll(msk->subflow->sk);
+	}
 
 	if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
 		mask |= mptcp_check_readable(msk);
diff --git a/net/mptcp/protocol.h b/net/mptcp/protocol.h
index 188b9e044672..b2d016f8803a 100644
--- a/net/mptcp/protocol.h
+++ b/net/mptcp/protocol.h
@@ -111,7 +111,6 @@
 #define MPTCP_RST_TRANSIENT	BIT(0)
 
 /* MPTCP socket flags */
-#define MPTCP_DATA_READY	0
 #define MPTCP_NOSPACE		1
 #define MPTCP_WORK_RTX		2
 #define MPTCP_WORK_EOF		3
diff --git a/net/mptcp/subflow.c b/net/mptcp/subflow.c
index 72ccbd85941b..4e7395cfab9b 100644
--- a/net/mptcp/subflow.c
+++ b/net/mptcp/subflow.c
@@ -1293,7 +1293,6 @@ static void subflow_data_ready(struct sock *sk)
 		if (reqsk_queue_empty(&inet_csk(sk)->icsk_accept_queue))
 			return;
 
-		set_bit(MPTCP_DATA_READY, &msk->flags);
 		parent->sk_data_ready(parent);
 		return;
 	}
-- 
2.26.3

Re: mptcp: cleanup accept and poll: Build Failure

[MPTCP CI]

Hi Paolo,

Thank you for your modifications, that's great!

But sadly, our CI spotted some issues with it when trying to build it.

You can find more details there:

  https://patchwork.kernel.org/project/mptcp/patch/bcf5c3220ac27570e942835d97b4b10b8e448eb8.1636107361.git.pabeni@redhat.com/
  https://github.com/multipath-tcp/mptcp_net-next/actions/runs/1425370804

Status: failure
Initiator: MPTCPimporter
Commits: https://github.com/multipath-tcp/mptcp_net-next/commits/f9fa79e12d45

Feel free to reply to this email if you cannot access logs, if you need
some support to fix the error, if this doesn't seem to be caused by your
modifications or if the error is a false positive one.

Cheers,
MPTCP GH Action bot

Re: [RFC PATCH 0/3] mptcp: improve accept() and disconnect()

[Mat Martineau]

On Fri, 5 Nov 2021, Paolo Abeni wrote:

> As outlined in the public mtg, mptcp_accept() is currently quite
> suboptimal, both from performance and code complexity
>
> This series tries to clean it up, enforcing a wider lifetime for
> the initial subflow, so that we don't need to acquire additional
> references there.
>
> To reach such goal we need to properly define the disconnect()
> behavior, which is currently quite incomplete.
>
> Some additional self-tests is likely required,
>

Hi Paolo -

I don't have any changes to suggest right now except for the 
"msk->last_snd = 0" (vs NULL) that the CI found.

And this finally expunges MPTCP_DATA_READY too!

I agree that some kind of test that tries to close/destroy the msk while a 
listen() (or other calls) are pending would be very useful.

--
Mat Martineau
Intel

Re: [RFC PATCH 1/3] mptcp: never allow the PM to close a listener subflow

[Matthieu Baerts]

Hi Paolo,

On 05/11/2021 11:49, Paolo Abeni wrote:
> Currently, when deleting an endpoint the netlink PM treverses
> all the local MPTCP sockets, regardless of their status.
> 
> If an MPTCP listener socket is bound to the IP matching the
> delete endpoint, the listener TCP socket will be closed.
> That is unexpected, the PM should only affect data subflows.
> 
> Fix the issue explicitly skipping MPTCP socket in TCP_LISTEN
> status.

Good catch!

Should we add a "Fixes" tag and have this backported in stable versions?

Cheers,
Matt
-- 
Tessares | Belgium | Hybrid Access Solutions
www.tessares.net

Re: [RFC PATCH 1/3] mptcp: never allow the PM to close a listener subflow

[Paolo Abeni]

On Mon, 2021-11-08 at 17:56 +0100, Matthieu Baerts wrote:
> Hi Paolo,
> 
> On 05/11/2021 11:49, Paolo Abeni wrote:
> > Currently, when deleting an endpoint the netlink PM treverses
> > all the local MPTCP sockets, regardless of their status.
> > 
> > If an MPTCP listener socket is bound to the IP matching the
> > delete endpoint, the listener TCP socket will be closed.
> > That is unexpected, the PM should only affect data subflows.
> > 
> > Fix the issue explicitly skipping MPTCP socket in TCP_LISTEN
> > status.
> 
> Good catch!
> 
> Should we add a "Fixes" tag and have this backported in stable versions?

We need this patch to be in the same series of the following two - or
at least to enter the tree before the others. I did not had the Fixes
tag to allow this patch to target the -net-next tree as the following
ones.

BTW I'm working to add self-tests here (mptcp_connect based [!!!]). It
looks like tha a good disconnect() implementation should use FASTCLOSE
to shutdown the MPTCP-level connection, so I'm going to add the tx part
of it, too.

Cheers,

Paolo

Re: [RFC PATCH 1/3] mptcp: never allow the PM to close a listener subflow

[Matthieu Baerts]

On 09/11/2021 15:38, Paolo Abeni wrote:
> On Mon, 2021-11-08 at 17:56 +0100, Matthieu Baerts wrote:
>> Hi Paolo,
>>
>> On 05/11/2021 11:49, Paolo Abeni wrote:
>>> Currently, when deleting an endpoint the netlink PM treverses
>>> all the local MPTCP sockets, regardless of their status.
>>>
>>> If an MPTCP listener socket is bound to the IP matching the
>>> delete endpoint, the listener TCP socket will be closed.
>>> That is unexpected, the PM should only affect data subflows.
>>>
>>> Fix the issue explicitly skipping MPTCP socket in TCP_LISTEN
>>> status.
>>
>> Good catch!
>>
>> Should we add a "Fixes" tag and have this backported in stable versions?
> 
> We need this patch to be in the same series of the following two - or
> at least to enter the tree before the others. I did not had the Fixes
> tag to allow this patch to target the -net-next tree as the following
> ones.

Can we not add the Fixes tag and send it to net-next?
Or we already send this one to -net while net-next is closed?

> BTW I'm working to add self-tests here (mptcp_connect based [!!!]). It

I thought only Florian was allowed to modify mptcp_connect! :-O

> looks like tha a good disconnect() implementation should use FASTCLOSE
> to shutdown the MPTCP-level connection, so I'm going to add the tx part
> of it, too.

Good idea, thank you!

Cheers,
Matt
-- 
Tessares | Belgium | Hybrid Access Solutions
www.tessares.net