tcp/dccp: install syn_recv requests into ehash table

In this patch, we insert request sockets into TCP/DCCP
regular ehash table (where ESTABLISHED and TIMEWAIT sockets
are) instead of using the per listener hash table.

ACK packets find SYN_RECV pseudo sockets without having
to find and lock the listener.

In nominal conditions, this halves pressure on listener lock.

Note that this will allow for SO_REUSEPORT refinements,
so that we can select a listener using cpu/numa affinities instead
of the prior 'consistent hash', since only SYN packets will
apply this selection logic.

We will shrink listen_sock in the following patch to ease
code review.

Signed-off-by: Eric Dumazet <[email protected]>
Cc: Ying Cai <[email protected]>
Cc: Willem de Bruijn <[email protected]>
Signed-off-by: David S. Miller <[email protected]>

Author: Eric Dumazet

include/net/inet_connection_sock.h

@@ -258,10 +258,6 @@ inet_csk_rto_backoff(const struct inet_connection_sock *icsk,
 
 struct sock *inet_csk_accept(struct sock *sk, int flags, int *err);
 
-struct request_sock *inet_csk_search_req(struct sock *sk,
-					 const __be16 rport,
-					 const __be32 raddr,
-					 const __be32 laddr);
 int inet_csk_bind_conflict(const struct sock *sk,
 			   const struct inet_bind_bucket *tb, bool relax);
 int inet_csk_get_port(struct sock *sk, unsigned short snum);

include/net/inet_hashtables.h

@@ -205,6 +205,7 @@ void inet_put_port(struct sock *sk);
 
 void inet_hashinfo_init(struct inet_hashinfo *h);
 
+int inet_ehash_insert(struct sock *sk, struct sock *osk);
 void __inet_hash_nolisten(struct sock *sk, struct sock *osk);
 void __inet_hash(struct sock *sk, struct sock *osk);
 void inet_hash(struct sock *sk);

include/net/request_sock.h

@@ -266,8 +266,4 @@ static inline int reqsk_queue_is_full(const struct request_sock_queue *queue)
 	return reqsk_queue_len(queue) >> queue->listen_opt->max_qlen_log;
 }
 
-void reqsk_queue_hash_req(struct request_sock_queue *queue,
-			  u32 hash, struct request_sock *req,
-			  unsigned long timeout);
-
 #endif /* _REQUEST_SOCK_H */

include/net/tcp.h

@@ -1618,7 +1618,6 @@ static inline bool tcp_stream_is_thin(struct tcp_sock *tp)
 /* /proc */
 enum tcp_seq_states {
 	TCP_SEQ_STATE_LISTENING,
-	TCP_SEQ_STATE_OPENREQ,
 	TCP_SEQ_STATE_ESTABLISHED,
 };
 
@@ -1717,8 +1716,6 @@ struct tcp_request_sock_ops {
 	int (*send_synack)(const struct sock *sk, struct dst_entry *dst,
 			   struct flowi *fl, struct request_sock *req,
 			   u16 queue_mapping, struct tcp_fastopen_cookie *foc);
-	void (*queue_hash_add)(struct sock *sk, struct request_sock *req,
-			       const unsigned long timeout);
 };
 
 #ifdef CONFIG_SYN_COOKIES

net/core/request_sock.c

@@ -99,35 +99,9 @@ static inline struct listen_sock *reqsk_queue_yank_listen_sk(
 
 void reqsk_queue_destroy(struct request_sock_queue *queue)
 {
-	/* make all the listen_opt local to us */
 	struct listen_sock *lopt = reqsk_queue_yank_listen_sk(queue);
 
-	if (reqsk_queue_len(queue) != 0) {
-		unsigned int i;
-
-		for (i = 0; i < lopt->nr_table_entries; i++) {
-			struct request_sock *req;
-
-			spin_lock_bh(&queue->syn_wait_lock);
-			while ((req = lopt->syn_table[i]) != NULL) {
-				lopt->syn_table[i] = req->dl_next;
-				/* Because of following del_timer_sync(),
-				 * we must release the spinlock here
-				 * or risk a dead lock.
-				 */
-				spin_unlock_bh(&queue->syn_wait_lock);
-				atomic_dec(&queue->qlen);
-				if (del_timer_sync(&req->rsk_timer))
-					reqsk_put(req);
-				reqsk_put(req);
-				spin_lock_bh(&queue->syn_wait_lock);
-			}
-			spin_unlock_bh(&queue->syn_wait_lock);
-		}
-	}
-
-	if (WARN_ON(reqsk_queue_len(queue) != 0))
-		pr_err("qlen %u\n", reqsk_queue_len(queue));
+	/* cleaning is done by req timers */
 	kvfree(lopt);
 }
 

net/dccp/ipv4.c

@@ -444,36 +444,6 @@ struct sock *dccp_v4_request_recv_sock(const struct sock *sk,
 }
 EXPORT_SYMBOL_GPL(dccp_v4_request_recv_sock);
 
-static struct sock *dccp_v4_hnd_req(struct sock *sk, struct sk_buff *skb)
-{
-	const struct dccp_hdr *dh = dccp_hdr(skb);
-	const struct iphdr *iph = ip_hdr(skb);
-	struct sock *nsk;
-	/* Find possible connection requests. */
-	struct request_sock *req = inet_csk_search_req(sk, dh->dccph_sport,
-						       iph->saddr, iph->daddr);
-	if (req) {
-		nsk = dccp_check_req(sk, skb, req);
-		if (!nsk)
-			reqsk_put(req);
-		return nsk;
-	}
-	nsk = inet_lookup_established(sock_net(sk), &dccp_hashinfo,
-				      iph->saddr, dh->dccph_sport,
-				      iph->daddr, dh->dccph_dport,
-				      inet_iif(skb));
-	if (nsk != NULL) {
-		if (nsk->sk_state != DCCP_TIME_WAIT) {
-			bh_lock_sock(nsk);
-			return nsk;
-		}
-		inet_twsk_put(inet_twsk(nsk));
-		return NULL;
-	}
-
-	return sk;
-}
-
 static struct dst_entry* dccp_v4_route_skb(struct net *net, struct sock *sk,
 					   struct sk_buff *skb)
 {
@@ -705,26 +675,13 @@ int dccp_v4_do_rcv(struct sock *sk, struct sk_buff *skb)
 	 * NOTE: the check for the packet types is done in
 	 *	 dccp_rcv_state_process
 	 */
-	if (sk->sk_state == DCCP_LISTEN) {
-		struct sock *nsk = dccp_v4_hnd_req(sk, skb);
-
-		if (nsk == NULL)
-			goto discard;
-
-		if (nsk != sk) {
-			if (dccp_child_process(sk, nsk, skb))
-				goto reset;
-			return 0;
-		}
-	}
 
 	if (dccp_rcv_state_process(sk, skb, dh, skb->len))
 		goto reset;
 	return 0;
 
 reset:
 	dccp_v4_ctl_send_reset(sk, skb);
-discard:
 	kfree_skb(skb);
 	return 0;
 }
@@ -868,6 +825,27 @@ static int dccp_v4_rcv(struct sk_buff *skb)
 		goto no_dccp_socket;
 	}
 
+	if (sk->sk_state == DCCP_NEW_SYN_RECV) {
+		struct request_sock *req = inet_reqsk(sk);
+		struct sock *nsk = NULL;
+
+		sk = req->rsk_listener;
+		if (sk->sk_state == DCCP_LISTEN)
+			nsk = dccp_check_req(sk, skb, req);
+		if (!nsk) {
+			reqsk_put(req);
+			goto discard_it;
+		}
+		if (nsk == sk) {
+			sock_hold(sk);
+			reqsk_put(req);
+		} else if (dccp_child_process(sk, nsk, skb)) {
+			dccp_v4_ctl_send_reset(sk, skb);
+			goto discard_it;
+		} else {
+			return 0;
+		}
+	}
 	/*
 	 * RFC 4340, sec. 9.2.1: Minimum Checksum Coverage
 	 *	o if MinCsCov = 0, only packets with CsCov = 0 are accepted

net/dccp/ipv6.c

@@ -290,37 +290,6 @@ static struct request_sock_ops dccp6_request_sock_ops = {
 	.syn_ack_timeout = dccp_syn_ack_timeout,
 };
 
-static struct sock *dccp_v6_hnd_req(struct sock *sk,struct sk_buff *skb)
-{
-	const struct dccp_hdr *dh = dccp_hdr(skb);
-	const struct ipv6hdr *iph = ipv6_hdr(skb);
-	struct request_sock *req;
-	struct sock *nsk;
-
-	req = inet6_csk_search_req(sk, dh->dccph_sport, &iph->saddr,
-				   &iph->daddr, inet6_iif(skb));
-	if (req) {
-		nsk = dccp_check_req(sk, skb, req);
-		if (!nsk)
-			reqsk_put(req);
-		return nsk;
-	}
-	nsk = __inet6_lookup_established(sock_net(sk), &dccp_hashinfo,
-					 &iph->saddr, dh->dccph_sport,
-					 &iph->daddr, ntohs(dh->dccph_dport),
-					 inet6_iif(skb));
-	if (nsk != NULL) {
-		if (nsk->sk_state != DCCP_TIME_WAIT) {
-			bh_lock_sock(nsk);
-			return nsk;
-		}
-		inet_twsk_put(inet_twsk(nsk));
-		return NULL;
-	}
-
-	return sk;
-}
-
 static int dccp_v6_conn_request(struct sock *sk, struct sk_buff *skb)
 {
 	struct request_sock *req;
@@ -398,7 +367,7 @@ static int dccp_v6_conn_request(struct sock *sk, struct sk_buff *skb)
 	if (dccp_v6_send_response(sk, req))
 		goto drop_and_free;
 
-	inet6_csk_reqsk_queue_hash_add(sk, req, DCCP_TIMEOUT_INIT);
+	inet_csk_reqsk_queue_hash_add(sk, req, DCCP_TIMEOUT_INIT);
 	return 0;
 
 drop_and_free:
@@ -641,24 +610,6 @@ static int dccp_v6_do_rcv(struct sock *sk, struct sk_buff *skb)
 	 * NOTE: the check for the packet types is done in
 	 *	 dccp_rcv_state_process
 	 */
-	if (sk->sk_state == DCCP_LISTEN) {
-		struct sock *nsk = dccp_v6_hnd_req(sk, skb);
-
-		if (nsk == NULL)
-			goto discard;
-		/*
-		 * Queue it on the new socket if the new socket is active,
-		 * otherwise we just shortcircuit this and continue with
-		 * the new socket..
-		 */
-		if (nsk != sk) {
-			if (dccp_child_process(sk, nsk, skb))
-				goto reset;
-			if (opt_skb != NULL)
-				__kfree_skb(opt_skb);
-			return 0;
-		}
-	}
 
 	if (dccp_rcv_state_process(sk, skb, dccp_hdr(skb), skb->len))
 		goto reset;
@@ -732,6 +683,27 @@ static int dccp_v6_rcv(struct sk_buff *skb)
 		goto no_dccp_socket;
 	}
 
+	if (sk->sk_state == DCCP_NEW_SYN_RECV) {
+		struct request_sock *req = inet_reqsk(sk);
+		struct sock *nsk = NULL;
+
+		sk = req->rsk_listener;
+		if (sk->sk_state == DCCP_LISTEN)
+			nsk = dccp_check_req(sk, skb, req);
+		if (!nsk) {
+			reqsk_put(req);
+			goto discard_it;
+		}
+		if (nsk == sk) {
+			sock_hold(sk);
+			reqsk_put(req);
+		} else if (dccp_child_process(sk, nsk, skb)) {
+			dccp_v6_ctl_send_reset(sk, skb);
+			goto discard_it;
+		} else {
+			return 0;
+		}
+	}
 	/*
 	 * RFC 4340, sec. 9.2.1: Minimum Checksum Coverage
 	 *	o if MinCsCov = 0, only packets with CsCov = 0 are accepted