Merge branch 'tcp-preempt'

Eric Dumazet says:

====================
net: make TCP preemptible

Most of TCP stack assumed it was running from BH handler.

This is great for most things, as TCP behavior is very sensitive
to scheduling artifacts.

However, the prequeue and backlog processing are problematic,
as they need to be flushed with BH being blocked.

To cope with modern needs, TCP sockets have big sk_rcvbuf values,
in the order of 16 MB, and soon 32 MB.
This means that backlog can hold thousands of packets, and things
like TCP coalescing or collapsing on this amount of packets can
lead to insane latency spikes, since BH are blocked for too long.

It is time to make UDP/TCP stacks preemptible.

Note that fast path still runs from BH handler.

v2: Added "tcp: make tcp_sendmsg() aware of socket backlog"
    to reduce latency problems of large sends.

v3: Fixed a typo in tcp_cdg.c
====================

Signed-off-by: David S. Miller <davem@davemloft.net>

include/net/sock.h

@@ -926,6 +926,17 @@ void sk_stream_kill_queues(struct sock *sk);
 void sk_set_memalloc(struct sock *sk);
 void sk_clear_memalloc(struct sock *sk);
 
+void __sk_flush_backlog(struct sock *sk);
+
+static inline bool sk_flush_backlog(struct sock *sk)
+{
+	if (unlikely(READ_ONCE(sk->sk_backlog.tail))) {
+		__sk_flush_backlog(sk);
+		return true;
+	}
+	return false;
+}
+
 int sk_wait_data(struct sock *sk, long *timeo, const struct sk_buff *skb);
 
 struct request_sock_ops;

net/core/sock.c

@@ -2019,33 +2019,27 @@ static void __release_sock(struct sock *sk)
 	__releases(&sk->sk_lock.slock)
 	__acquires(&sk->sk_lock.slock)
 {
-	struct sk_buff *skb = sk->sk_backlog.head;
+	struct sk_buff *skb, *next;
 
-	do {
+	while ((skb = sk->sk_backlog.head) != NULL) {
 		sk->sk_backlog.head = sk->sk_backlog.tail = NULL;
-		bh_unlock_sock(sk);
 
-		do {
-			struct sk_buff *next = skb->next;
+		spin_unlock_bh(&sk->sk_lock.slock);
 
+		do {
+			next = skb->next;
 			prefetch(next);
 			WARN_ON_ONCE(skb_dst_is_noref(skb));
 			skb->next = NULL;
 			sk_backlog_rcv(sk, skb);
 
-			/*
-			 * We are in process context here with softirqs
-			 * disabled, use cond_resched_softirq() to preempt.
-			 * This is safe to do because we've taken the backlog
-			 * queue private:
-			 */
-			cond_resched_softirq();
+			cond_resched();
 
 			skb = next;
 		} while (skb != NULL);
 
-		bh_lock_sock(sk);
-	} while ((skb = sk->sk_backlog.head) != NULL);
+		spin_lock_bh(&sk->sk_lock.slock);
+	}
 
 	/*
 	 * Doing the zeroing here guarantee we can not loop forever
@@ -2054,6 +2048,13 @@ static void __release_sock(struct sock *sk)
 	sk->sk_backlog.len = 0;
 }
 
+void __sk_flush_backlog(struct sock *sk)
+{
+	spin_lock_bh(&sk->sk_lock.slock);
+	__release_sock(sk);
+	spin_unlock_bh(&sk->sk_lock.slock);
+}
+
 /**
  * sk_wait_data - wait for data to arrive at sk_receive_queue
  * @sk:    sock to wait on

net/dccp/input.c

@@ -359,7 +359,7 @@ static int __dccp_rcv_established(struct sock *sk, struct sk_buff *skb,
 		goto discard;
 	}
 
-	__DCCP_INC_STATS(DCCP_MIB_INERRS);
+	DCCP_INC_STATS(DCCP_MIB_INERRS);
 discard:
 	__kfree_skb(skb);
 	return 0;

net/dccp/ipv4.c

@@ -533,8 +533,8 @@ static void dccp_v4_ctl_send_reset(const struct sock *sk, struct sk_buff *rxskb)
 	bh_unlock_sock(ctl_sk);
 
 	if (net_xmit_eval(err) == 0) {
-		__DCCP_INC_STATS(DCCP_MIB_OUTSEGS);
-		__DCCP_INC_STATS(DCCP_MIB_OUTRSTS);
+		DCCP_INC_STATS(DCCP_MIB_OUTSEGS);
+		DCCP_INC_STATS(DCCP_MIB_OUTRSTS);
 	}
 out:
 	 dst_release(dst);

net/dccp/ipv6.c

@@ -277,8 +277,8 @@ static void dccp_v6_ctl_send_reset(const struct sock *sk, struct sk_buff *rxskb)
 	if (!IS_ERR(dst)) {
 		skb_dst_set(skb, dst);
 		ip6_xmit(ctl_sk, skb, &fl6, NULL, 0);
-		__DCCP_INC_STATS(DCCP_MIB_OUTSEGS);
-		__DCCP_INC_STATS(DCCP_MIB_OUTRSTS);
+		DCCP_INC_STATS(DCCP_MIB_OUTSEGS);
+		DCCP_INC_STATS(DCCP_MIB_OUTRSTS);
 		return;
 	}
 

net/dccp/options.c

@@ -253,7 +253,7 @@ int dccp_parse_options(struct sock *sk, struct dccp_request_sock *dreq,
 	return 0;
 
 out_invalid_option:
-	__DCCP_INC_STATS(DCCP_MIB_INVALIDOPT);
+	DCCP_INC_STATS(DCCP_MIB_INVALIDOPT);
 	rc = DCCP_RESET_CODE_OPTION_ERROR;
 out_featneg_failed:
 	DCCP_WARN("DCCP(%p): Option %d (len=%d) error=%u\n", sk, opt, len, rc);

net/ipv4/tcp.c

@@ -1136,11 +1136,12 @@ int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
 	/* This should be in poll */
 	sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
 
-	mss_now = tcp_send_mss(sk, &size_goal, flags);
-
 	/* Ok commence sending. */
 	copied = 0;
 
+restart:
+	mss_now = tcp_send_mss(sk, &size_goal, flags);
+
 	err = -EPIPE;
 	if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
 		goto out_err;
@@ -1166,6 +1167,9 @@ int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
 			if (!sk_stream_memory_free(sk))
 				goto wait_for_sndbuf;
 
+			if (sk_flush_backlog(sk))
+				goto restart;
+
 			skb = sk_stream_alloc_skb(sk,
 						  select_size(sk, sg),
 						  sk->sk_allocation,
@@ -1449,12 +1453,8 @@ static void tcp_prequeue_process(struct sock *sk)
 
 	NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
 
-	/* RX process wants to run with disabled BHs, though it is not
-	 * necessary */
-	local_bh_disable();
 	while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
 		sk_backlog_rcv(sk, skb);
-	local_bh_enable();
 
 	/* Clear memory counter. */
 	tp->ucopy.memory = 0;
@@ -3095,7 +3095,7 @@ void tcp_done(struct sock *sk)
 	struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
 
 	if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
-		__TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
+		TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
 
 	tcp_set_state(sk, TCP_CLOSE);
 	tcp_clear_xmit_timers(sk);

net/ipv4/tcp_cdg.c

@@ -155,11 +155,11 @@ static void tcp_cdg_hystart_update(struct sock *sk)
 
 			ca->last_ack = now_us;
 			if (after(now_us, ca->round_start + base_owd)) {
-				__NET_INC_STATS(sock_net(sk),
-						LINUX_MIB_TCPHYSTARTTRAINDETECT);
-				__NET_ADD_STATS(sock_net(sk),
-						LINUX_MIB_TCPHYSTARTTRAINCWND,
-						tp->snd_cwnd);
+				NET_INC_STATS(sock_net(sk),
+					      LINUX_MIB_TCPHYSTARTTRAINDETECT);
+				NET_ADD_STATS(sock_net(sk),
+					      LINUX_MIB_TCPHYSTARTTRAINCWND,
+					      tp->snd_cwnd);
 				tp->snd_ssthresh = tp->snd_cwnd;
 				return;
 			}
@@ -174,11 +174,11 @@ static void tcp_cdg_hystart_update(struct sock *sk)
 					 125U);
 
 			if (ca->rtt.min > thresh) {
-				__NET_INC_STATS(sock_net(sk),
-						LINUX_MIB_TCPHYSTARTDELAYDETECT);
-				__NET_ADD_STATS(sock_net(sk),
-						LINUX_MIB_TCPHYSTARTDELAYCWND,
-						tp->snd_cwnd);
+				NET_INC_STATS(sock_net(sk),
+					      LINUX_MIB_TCPHYSTARTDELAYDETECT);
+				NET_ADD_STATS(sock_net(sk),
+					      LINUX_MIB_TCPHYSTARTDELAYCWND,
+					      tp->snd_cwnd);
 				tp->snd_ssthresh = tp->snd_cwnd;
 			}
 		}

net/ipv4/tcp_cubic.c

@@ -402,11 +402,11 @@ static void hystart_update(struct sock *sk, u32 delay)
 			ca->last_ack = now;
 			if ((s32)(now - ca->round_start) > ca->delay_min >> 4) {
 				ca->found |= HYSTART_ACK_TRAIN;
-				__NET_INC_STATS(sock_net(sk),
-						LINUX_MIB_TCPHYSTARTTRAINDETECT);
-				__NET_ADD_STATS(sock_net(sk),
-						LINUX_MIB_TCPHYSTARTTRAINCWND,
-						tp->snd_cwnd);
+				NET_INC_STATS(sock_net(sk),
+					      LINUX_MIB_TCPHYSTARTTRAINDETECT);
+				NET_ADD_STATS(sock_net(sk),
+					      LINUX_MIB_TCPHYSTARTTRAINCWND,
+					      tp->snd_cwnd);
 				tp->snd_ssthresh = tp->snd_cwnd;
 			}
 		}
@@ -423,11 +423,11 @@ static void hystart_update(struct sock *sk, u32 delay)
 			if (ca->curr_rtt > ca->delay_min +
 			    HYSTART_DELAY_THRESH(ca->delay_min >> 3)) {
 				ca->found |= HYSTART_DELAY;
-				__NET_INC_STATS(sock_net(sk),
-						LINUX_MIB_TCPHYSTARTDELAYDETECT);
-				__NET_ADD_STATS(sock_net(sk),
-						LINUX_MIB_TCPHYSTARTDELAYCWND,
-						tp->snd_cwnd);
+				NET_INC_STATS(sock_net(sk),
+					      LINUX_MIB_TCPHYSTARTDELAYDETECT);
+				NET_ADD_STATS(sock_net(sk),
+					      LINUX_MIB_TCPHYSTARTDELAYCWND,
+					      tp->snd_cwnd);
 				tp->snd_ssthresh = tp->snd_cwnd;
 			}
 		}

net/ipv4/tcp_fastopen.c

@@ -255,9 +255,9 @@ static bool tcp_fastopen_queue_check(struct sock *sk)
 		spin_lock(&fastopenq->lock);
 		req1 = fastopenq->rskq_rst_head;
 		if (!req1 || time_after(req1->rsk_timer.expires, jiffies)) {
-			spin_unlock(&fastopenq->lock);
 			__NET_INC_STATS(sock_net(sk),
 					LINUX_MIB_TCPFASTOPENLISTENOVERFLOW);
+			spin_unlock(&fastopenq->lock);
 			return false;
 		}
 		fastopenq->rskq_rst_head = req1->dl_next;
@@ -282,7 +282,7 @@ struct sock *tcp_try_fastopen(struct sock *sk, struct sk_buff *skb,
 	struct sock *child;
 
 	if (foc->len == 0) /* Client requests a cookie */
-		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENCOOKIEREQD);
+		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENCOOKIEREQD);
 
 	if (!((sysctl_tcp_fastopen & TFO_SERVER_ENABLE) &&
 	      (syn_data || foc->len >= 0) &&
@@ -311,13 +311,13 @@ struct sock *tcp_try_fastopen(struct sock *sk, struct sk_buff *skb,
 		child = tcp_fastopen_create_child(sk, skb, dst, req);
 		if (child) {
 			foc->len = -1;
-			__NET_INC_STATS(sock_net(sk),
-					LINUX_MIB_TCPFASTOPENPASSIVE);
+			NET_INC_STATS(sock_net(sk),
+				      LINUX_MIB_TCPFASTOPENPASSIVE);
 			return child;
 		}
-		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENPASSIVEFAIL);
+		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENPASSIVEFAIL);
 	} else if (foc->len > 0) /* Client presents an invalid cookie */
-		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENPASSIVEFAIL);
+		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENPASSIVEFAIL);
 
 	valid_foc.exp = foc->exp;
 	*foc = valid_foc;