[PATCH 6.6 28/49] mptcp: refactor sndbuf auto-tuning

[Greg Kroah-Hartman <gregkh@linuxfoundation.org>]

6.6-stable review patch.  If anyone has any objections, please let me know.

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

From: Paolo Abeni <pabeni@redhat.com>

[ Upstream commit 8005184fd1ca6aeb3fea36f4eb9463fc1b90c114 ]

The MPTCP protocol account for the data enqueued on all the subflows
to the main socket send buffer, while the send buffer auto-tuning
algorithm set the main socket send buffer size as the max size among
the subflows.

That causes bad performances when at least one subflow is sndbuf
limited, e.g. due to very high latency, as the MPTCP scheduler can't
even fill such buffer.

Change the send-buffer auto-tuning algorithm to compute the main socket
send buffer size as the sum of all the subflows buffer size.

Reviewed-by: Mat Martineau <martineau@kernel.org>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: Mat Martineau <martineau@kernel.org>
Link: https://lore.kernel.org/r/20231023-send-net-next-20231023-2-v1-9-9dc60939d371@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Stable-dep-of: 4fd19a307016 ("mptcp: fix inconsistent state on fastopen race")
Signed-off-by: Sasha Levin <sashal@kernel.org>
---
 net/mptcp/protocol.c | 18 +++++++++++++--
 net/mptcp/protocol.h | 54 ++++++++++++++++++++++++++++++++++++++++----
 net/mptcp/sockopt.c  |  5 +++-
 net/mptcp/subflow.c  |  3 +--
 4 files changed, 70 insertions(+), 10 deletions(-)

diff --git a/net/mptcp/protocol.c b/net/mptcp/protocol.c
index c1527f520dce3..44499e49d76e6 100644
--- a/net/mptcp/protocol.c
+++ b/net/mptcp/protocol.c
@@ -893,6 +893,7 @@ static bool __mptcp_finish_join(struct mptcp_sock *msk, struct sock *ssk)
 	mptcp_sockopt_sync_locked(msk, ssk);
 	mptcp_subflow_joined(msk, ssk);
 	mptcp_stop_tout_timer(sk);
+	__mptcp_propagate_sndbuf(sk, ssk);
 	return true;
 }
 
@@ -1079,15 +1080,16 @@ static void mptcp_enter_memory_pressure(struct sock *sk)
 	struct mptcp_sock *msk = mptcp_sk(sk);
 	bool first = true;
 
-	sk_stream_moderate_sndbuf(sk);
 	mptcp_for_each_subflow(msk, subflow) {
 		struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
 
 		if (first)
 			tcp_enter_memory_pressure(ssk);
 		sk_stream_moderate_sndbuf(ssk);
+
 		first = false;
 	}
+	__mptcp_sync_sndbuf(sk);
 }
 
 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of
@@ -2452,6 +2454,7 @@ static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
 		WRITE_ONCE(msk->first, NULL);
 
 out:
+	__mptcp_sync_sndbuf(sk);
 	if (need_push)
 		__mptcp_push_pending(sk, 0);
 
@@ -3223,7 +3226,7 @@ struct sock *mptcp_sk_clone_init(const struct sock *sk,
 	 * uses the correct data
 	 */
 	mptcp_copy_inaddrs(nsk, ssk);
-	mptcp_propagate_sndbuf(nsk, ssk);
+	__mptcp_propagate_sndbuf(nsk, ssk);
 
 	mptcp_rcv_space_init(msk, ssk);
 	bh_unlock_sock(nsk);
@@ -3401,6 +3404,8 @@ static void mptcp_release_cb(struct sock *sk)
 			__mptcp_set_connected(sk);
 		if (__test_and_clear_bit(MPTCP_ERROR_REPORT, &msk->cb_flags))
 			__mptcp_error_report(sk);
+		if (__test_and_clear_bit(MPTCP_SYNC_SNDBUF, &msk->cb_flags))
+			__mptcp_sync_sndbuf(sk);
 	}
 
 	__mptcp_update_rmem(sk);
@@ -3445,6 +3450,14 @@ void mptcp_subflow_process_delegated(struct sock *ssk, long status)
 			__set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
 		mptcp_data_unlock(sk);
 	}
+	if (status & BIT(MPTCP_DELEGATE_SNDBUF)) {
+		mptcp_data_lock(sk);
+		if (!sock_owned_by_user(sk))
+			__mptcp_sync_sndbuf(sk);
+		else
+			__set_bit(MPTCP_SYNC_SNDBUF, &mptcp_sk(sk)->cb_flags);
+		mptcp_data_unlock(sk);
+	}
 	if (status & BIT(MPTCP_DELEGATE_ACK))
 		schedule_3rdack_retransmission(ssk);
 }
@@ -3529,6 +3542,7 @@ bool mptcp_finish_join(struct sock *ssk)
 	/* active subflow, already present inside the conn_list */
 	if (!list_empty(&subflow->node)) {
 		mptcp_subflow_joined(msk, ssk);
+		mptcp_propagate_sndbuf(parent, ssk);
 		return true;
 	}
 
diff --git a/net/mptcp/protocol.h b/net/mptcp/protocol.h
index 3612545fa62e0..40866acd91ad5 100644
--- a/net/mptcp/protocol.h
+++ b/net/mptcp/protocol.h
@@ -123,6 +123,7 @@
 #define MPTCP_RETRANSMIT	4
 #define MPTCP_FLUSH_JOIN_LIST	5
 #define MPTCP_CONNECTED		6
+#define MPTCP_SYNC_SNDBUF	7
 
 struct mptcp_skb_cb {
 	u64 map_seq;
@@ -447,6 +448,7 @@ DECLARE_PER_CPU(struct mptcp_delegated_action, mptcp_delegated_actions);
 #define MPTCP_DELEGATE_SCHEDULED	0
 #define MPTCP_DELEGATE_SEND		1
 #define MPTCP_DELEGATE_ACK		2
+#define MPTCP_DELEGATE_SNDBUF		3
 
 #define MPTCP_DELEGATE_ACTIONS_MASK	(~BIT(MPTCP_DELEGATE_SCHEDULED))
 /* MPTCP subflow context */
@@ -520,6 +522,9 @@ struct mptcp_subflow_context {
 
 	u32	setsockopt_seq;
 	u32	stale_rcv_tstamp;
+	int     cached_sndbuf;	    /* sndbuf size when last synced with the msk sndbuf,
+				     * protected by the msk socket lock
+				     */
 
 	struct	sock *tcp_sock;	    /* tcp sk backpointer */
 	struct	sock *conn;	    /* parent mptcp_sock */
@@ -762,13 +767,52 @@ static inline bool mptcp_data_fin_enabled(const struct mptcp_sock *msk)
 	       READ_ONCE(msk->write_seq) == READ_ONCE(msk->snd_nxt);
 }
 
-static inline bool mptcp_propagate_sndbuf(struct sock *sk, struct sock *ssk)
+static inline void __mptcp_sync_sndbuf(struct sock *sk)
 {
-	if ((sk->sk_userlocks & SOCK_SNDBUF_LOCK) || ssk->sk_sndbuf <= READ_ONCE(sk->sk_sndbuf))
-		return false;
+	struct mptcp_subflow_context *subflow;
+	int ssk_sndbuf, new_sndbuf;
+
+	if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
+		return;
+
+	new_sndbuf = sock_net(sk)->ipv4.sysctl_tcp_wmem[0];
+	mptcp_for_each_subflow(mptcp_sk(sk), subflow) {
+		ssk_sndbuf =  READ_ONCE(mptcp_subflow_tcp_sock(subflow)->sk_sndbuf);
+
+		subflow->cached_sndbuf = ssk_sndbuf;
+		new_sndbuf += ssk_sndbuf;
+	}
+
+	/* the msk max wmem limit is <nr_subflows> * tcp wmem[2] */
+	WRITE_ONCE(sk->sk_sndbuf, new_sndbuf);
+}
+
+/* The called held both the msk socket and the subflow socket locks,
+ * possibly under BH
+ */
+static inline void __mptcp_propagate_sndbuf(struct sock *sk, struct sock *ssk)
+{
+	struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
+
+	if (READ_ONCE(ssk->sk_sndbuf) != subflow->cached_sndbuf)
+		__mptcp_sync_sndbuf(sk);
+}
+
+/* the caller held only the subflow socket lock, either in process or
+ * BH context. Additionally this can be called under the msk data lock,
+ * so we can't acquire such lock here: let the delegate action acquires
+ * the needed locks in suitable order.
+ */
+static inline void mptcp_propagate_sndbuf(struct sock *sk, struct sock *ssk)
+{
+	struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
+
+	if (likely(READ_ONCE(ssk->sk_sndbuf) == subflow->cached_sndbuf))
+		return;
 
-	WRITE_ONCE(sk->sk_sndbuf, ssk->sk_sndbuf);
-	return true;
+	local_bh_disable();
+	mptcp_subflow_delegate(subflow, MPTCP_DELEGATE_SNDBUF);
+	local_bh_enable();
 }
 
 static inline void mptcp_write_space(struct sock *sk)
diff --git a/net/mptcp/sockopt.c b/net/mptcp/sockopt.c
index 7539b9c8c2fb4..116e3008231bd 100644
--- a/net/mptcp/sockopt.c
+++ b/net/mptcp/sockopt.c
@@ -95,6 +95,7 @@ static void mptcp_sol_socket_sync_intval(struct mptcp_sock *msk, int optname, in
 		case SO_SNDBUFFORCE:
 			ssk->sk_userlocks |= SOCK_SNDBUF_LOCK;
 			WRITE_ONCE(ssk->sk_sndbuf, sk->sk_sndbuf);
+			mptcp_subflow_ctx(ssk)->cached_sndbuf = sk->sk_sndbuf;
 			break;
 		case SO_RCVBUF:
 		case SO_RCVBUFFORCE:
@@ -1418,8 +1419,10 @@ static void sync_socket_options(struct mptcp_sock *msk, struct sock *ssk)
 
 	if (sk->sk_userlocks & tx_rx_locks) {
 		ssk->sk_userlocks |= sk->sk_userlocks & tx_rx_locks;
-		if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
+		if (sk->sk_userlocks & SOCK_SNDBUF_LOCK) {
 			WRITE_ONCE(ssk->sk_sndbuf, sk->sk_sndbuf);
+			mptcp_subflow_ctx(ssk)->cached_sndbuf = sk->sk_sndbuf;
+		}
 		if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
 			WRITE_ONCE(ssk->sk_rcvbuf, sk->sk_rcvbuf);
 	}
diff --git a/net/mptcp/subflow.c b/net/mptcp/subflow.c
index 9c1f8d1d63d24..d8827427ffc84 100644
--- a/net/mptcp/subflow.c
+++ b/net/mptcp/subflow.c
@@ -421,6 +421,7 @@ static bool subflow_use_different_dport(struct mptcp_sock *msk, const struct soc
 
 void __mptcp_set_connected(struct sock *sk)
 {
+	__mptcp_propagate_sndbuf(sk, mptcp_sk(sk)->first);
 	if (sk->sk_state == TCP_SYN_SENT) {
 		inet_sk_state_store(sk, TCP_ESTABLISHED);
 		sk->sk_state_change(sk);
@@ -472,7 +473,6 @@ static void subflow_finish_connect(struct sock *sk, const struct sk_buff *skb)
 		return;
 
 	msk = mptcp_sk(parent);
-	mptcp_propagate_sndbuf(parent, sk);
 	subflow->rel_write_seq = 1;
 	subflow->conn_finished = 1;
 	subflow->ssn_offset = TCP_SKB_CB(skb)->seq;
@@ -1728,7 +1728,6 @@ static void subflow_state_change(struct sock *sk)
 
 	msk = mptcp_sk(parent);
 	if (subflow_simultaneous_connect(sk)) {
-		mptcp_propagate_sndbuf(parent, sk);
 		mptcp_do_fallback(sk);
 		mptcp_rcv_space_init(msk, sk);
 		pr_fallback(msk);
-- 
2.43.0