[PATCH v6 bpf-next 06/11] tcp: Migrate TCP_NEW_SYN_RECV requests at retransmitting SYN+ACKs.

From: Kuniyuki Iwashima <kuniyu@amazon.co.jp>
To: "David S . Miller" <davem@davemloft.net>,
	Jakub Kicinski <kuba@kernel.org>,
	Eric Dumazet <edumazet@google.com>,
	Alexei Starovoitov <ast@kernel.org>,
	Daniel Borkmann <daniel@iogearbox.net>,
	Andrii Nakryiko <andrii@kernel.org>,
	Martin KaFai Lau <kafai@fb.com>
Cc: Benjamin Herrenschmidt <benh@amazon.com>,
	Kuniyuki Iwashima <kuniyu@amazon.co.jp>,
	Kuniyuki Iwashima <kuni1840@gmail.com>, <bpf@vger.kernel.org>,
	<netdev@vger.kernel.org>, <linux-kernel@vger.kernel.org>
Subject: [PATCH v6 bpf-next 06/11] tcp: Migrate TCP_NEW_SYN_RECV requests at retransmitting SYN+ACKs.
Date: Mon, 17 May 2021 09:22:53 +0900	[thread overview]
Message-ID: <20210517002258.75019-7-kuniyu@amazon.co.jp> (raw)
In-Reply-To: <20210517002258.75019-1-kuniyu@amazon.co.jp>

As with the preceding patch, this patch changes reqsk_timer_handler() to
call reuseport_migrate_sock() and reqsk_clone() to migrate in-flight
requests at retransmitting SYN+ACKs. If we can select a new listener and
clone the request, we resume setting the SYN+ACK timer for the new req. If
we can set the timer, we call inet_ehash_insert() to unhash the old req and
put the new req into ehash.

The noteworthy point here is that by unhashing the old req, another CPU
processing it may lose the "own_req" race in tcp_v[46]_syn_recv_sock() and
drop the final ACK packet. However, the new timer will recover this
situation.

Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.co.jp>
---
 net/ipv4/inet_connection_sock.c | 75 ++++++++++++++++++++++++++++++---
 1 file changed, 68 insertions(+), 7 deletions(-)

diff --git a/net/ipv4/inet_connection_sock.c b/net/ipv4/inet_connection_sock.c
index 07e97b2f3635..c1f068464363 100644
--- a/net/ipv4/inet_connection_sock.c
+++ b/net/ipv4/inet_connection_sock.c
@@ -735,10 +735,20 @@ static struct request_sock *inet_reqsk_clone(struct request_sock *req,
 	return nreq;
 }
 
+static void reqsk_queue_migrated(struct request_sock_queue *queue,
+				 const struct request_sock *req)
+{
+	if (req->num_timeout == 0)
+		atomic_inc(&queue->young);
+	atomic_inc(&queue->qlen);
+}
+
 static void reqsk_migrate_reset(struct request_sock *req)
 {
+	req->saved_syn = NULL;
+	inet_rsk(req)->ireq_opt = NULL;
 #if IS_ENABLED(CONFIG_IPV6)
-	inet_rsk(req)->ipv6_opt = NULL;
+	inet_rsk(req)->pktopts = NULL;
 #endif
 }
 
@@ -782,15 +792,39 @@ EXPORT_SYMBOL(inet_csk_reqsk_queue_drop_and_put);
 static void reqsk_timer_handler(struct timer_list *t)
 {
 	struct request_sock *req = from_timer(req, t, rsk_timer);
+	struct request_sock *nreq = NULL, *oreq = req;
 	struct sock *sk_listener = req->rsk_listener;
-	struct net *net = sock_net(sk_listener);
-	struct inet_connection_sock *icsk = inet_csk(sk_listener);
-	struct request_sock_queue *queue = &icsk->icsk_accept_queue;
+	struct inet_connection_sock *icsk;
+	struct request_sock_queue *queue;
+	struct net *net;
 	int max_syn_ack_retries, qlen, expire = 0, resend = 0;
 
-	if (inet_sk_state_load(sk_listener) != TCP_LISTEN)
-		goto drop;
+	if (inet_sk_state_load(sk_listener) != TCP_LISTEN) {
+		struct sock *nsk;
+
+		nsk = reuseport_migrate_sock(sk_listener, req_to_sk(req), NULL);
+		if (!nsk)
+			goto drop;
+
+		nreq = inet_reqsk_clone(req, nsk);
+		if (!nreq)
+			goto drop;
+
+		/* The new timer for the cloned req can decrease the 2
+		 * by calling inet_csk_reqsk_queue_drop_and_put(), so
+		 * hold another count to prevent use-after-free and
+		 * call reqsk_put() just before return.
+		 */
+		refcount_set(&nreq->rsk_refcnt, 2 + 1);
+		timer_setup(&nreq->rsk_timer, reqsk_timer_handler, TIMER_PINNED);
+		reqsk_queue_migrated(&inet_csk(nsk)->icsk_accept_queue, req);
+
+		req = nreq;
+		sk_listener = nsk;
+	}
 
+	icsk = inet_csk(sk_listener);
+	net = sock_net(sk_listener);
 	max_syn_ack_retries = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_synack_retries;
 	/* Normally all the openreqs are young and become mature
 	 * (i.e. converted to established socket) for first timeout.
@@ -809,6 +843,7 @@ static void reqsk_timer_handler(struct timer_list *t)
 	 * embrions; and abort old ones without pity, if old
 	 * ones are about to clog our table.
 	 */
+	queue = &icsk->icsk_accept_queue;
 	qlen = reqsk_queue_len(queue);
 	if ((qlen << 1) > max(8U, READ_ONCE(sk_listener->sk_max_ack_backlog))) {
 		int young = reqsk_queue_len_young(queue) << 1;
@@ -833,10 +868,36 @@ static void reqsk_timer_handler(struct timer_list *t)
 			atomic_dec(&queue->young);
 		timeo = min(TCP_TIMEOUT_INIT << req->num_timeout, TCP_RTO_MAX);
 		mod_timer(&req->rsk_timer, jiffies + timeo);
+
+		if (!nreq)
+			return;
+
+		if (!inet_ehash_insert(req_to_sk(nreq), req_to_sk(oreq), NULL)) {
+			/* delete timer */
+			inet_csk_reqsk_queue_drop(sk_listener, nreq);
+			goto drop;
+		}
+
+		reqsk_migrate_reset(oreq);
+		reqsk_queue_removed(&inet_csk(oreq->rsk_listener)->icsk_accept_queue, oreq);
+		reqsk_put(oreq);
+
+		reqsk_put(nreq);
 		return;
 	}
+
 drop:
-	inet_csk_reqsk_queue_drop_and_put(sk_listener, req);
+	/* Even if we can clone the req, we may need not retransmit any more
+	 * SYN+ACKs (nreq->num_timeout > max_syn_ack_retries, etc), or another
+	 * CPU may win the "own_req" race so that inet_ehash_insert() fails.
+	 */
+	if (nreq) {
+		reqsk_migrate_reset(nreq);
+		reqsk_queue_removed(queue, nreq);
+		__reqsk_free(nreq);
+	}
+
+	inet_csk_reqsk_queue_drop_and_put(oreq->rsk_listener, oreq);
 }
 
 static void reqsk_queue_hash_req(struct request_sock *req,
-- 
2.30.2

