Merge branch 'tcp-tsq-perf'

Eric Dumazet says:

====================
tcp: tsq: performance series

Under very high TX stress, CPU handling NIC TX completions can spend
considerable amount of cycles handling TSQ (TCP Small Queues) logic.

This patch series avoids some atomic operations, but most notable
patch is the 3rd one, allowing other cpus processing ACK packets and
calling tcp_write_xmit() to grab TCP_TSQ_DEFERRED so that
tcp_tasklet_func() can skip already processed sockets.

This avoid lots of lock acquisitions and cache lines accesses,
particularly under load.

In v2, I added :

- tcp_small_queue_check() change to allow 1st and 2nd packets
  in write queue to be sent, even in the case TX completion of
    already acknowledged packets did not happen yet.
      This helps when TX completion coalescing parameters are set
        even to insane values, and/or busy polling is used.

- A reorganization of struct sock fields to
  lower false sharing and increase data locality.

- Then I moved tsq_flags from tcp_sock to struct sock also
  to reduce cache line misses during TX completions.

I measured an overall throughput gain of 22 % for heavy TCP use
over a single TX queue.
====================

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

7 files changed, 98 insertions, 68 deletions

diff --git a/include/linux/tcp.h b/include/linux/tcp.h
index 734bab4c3bef..fc5848dad7a4 100644
--- a/include/linux/tcp.h
+++ b/include/linux/tcp.h
@@ -186,7 +186,6 @@ struct tcp_sock {
 	u32	tsoffset;	/* timestamp offset */
 
 	struct list_head tsq_node; /* anchor in tsq_tasklet.head list */
-	unsigned long	tsq_flags;
 
 	/* Data for direct copy to user */
 	struct {
@@ -364,7 +363,7 @@ struct tcp_sock {
 	u32	*saved_syn;
 };
 
-enum tsq_flags {
+enum tsq_enum {
 	TSQ_THROTTLED,
 	TSQ_QUEUED,
 	TCP_TSQ_DEFERRED,	   /* tcp_tasklet_func() found socket was owned */
@@ -375,6 +374,15 @@ enum tsq_flags {
 				    */
 };
 
+enum tsq_flags {
+	TSQF_THROTTLED			= (1UL << TSQ_THROTTLED),
+	TSQF_QUEUED			= (1UL << TSQ_QUEUED),
+	TCPF_TSQ_DEFERRED		= (1UL << TCP_TSQ_DEFERRED),
+	TCPF_WRITE_TIMER_DEFERRED	= (1UL << TCP_WRITE_TIMER_DEFERRED),
+	TCPF_DELACK_TIMER_DEFERRED	= (1UL << TCP_DELACK_TIMER_DEFERRED),
+	TCPF_MTU_REDUCED_DEFERRED	= (1UL << TCP_MTU_REDUCED_DEFERRED),
+};
+
 static inline struct tcp_sock *tcp_sk(const struct sock *sk)
 {
 	return (struct tcp_sock *)sk;
diff --git a/include/net/sock.h b/include/net/sock.h
index 69afda6bea15..6dfe3aa22b97 100644
--- a/include/net/sock.h
+++ b/include/net/sock.h
@@ -343,6 +343,9 @@ struct sock {
 #define sk_rxhash		__sk_common.skc_rxhash
 
 	socket_lock_t		sk_lock;
+	atomic_t		sk_drops;
+	int			sk_rcvlowat;
+	struct sk_buff_head	sk_error_queue;
 	struct sk_buff_head	sk_receive_queue;
 	/*
 	 * The backlog queue is special, it is always used with
@@ -359,14 +362,13 @@ struct sock {
 		struct sk_buff	*tail;
 	} sk_backlog;
 #define sk_rmem_alloc sk_backlog.rmem_alloc
-	int			sk_forward_alloc;
 
-	__u32			sk_txhash;
+	int			sk_forward_alloc;
 #ifdef CONFIG_NET_RX_BUSY_POLL
-	unsigned int		sk_napi_id;
 	unsigned int		sk_ll_usec;
+	/* ===== mostly read cache line ===== */
+	unsigned int		sk_napi_id;
 #endif
-	atomic_t		sk_drops;
 	int			sk_rcvbuf;
 
 	struct sk_filter __rcu	*sk_filter;
@@ -379,11 +381,30 @@ struct sock {
 #endif
 	struct dst_entry	*sk_rx_dst;
 	struct dst_entry __rcu	*sk_dst_cache;
-	/* Note: 32bit hole on 64bit arches */
-	atomic_t		sk_wmem_alloc;
 	atomic_t		sk_omem_alloc;
 	int			sk_sndbuf;
+
+	/* ===== cache line for TX ===== */
+	int			sk_wmem_queued;
+	atomic_t		sk_wmem_alloc;
+	unsigned long		sk_tsq_flags;
+	struct sk_buff		*sk_send_head;
 	struct sk_buff_head	sk_write_queue;
+	__s32			sk_peek_off;
+	int			sk_write_pending;
+	long			sk_sndtimeo;
+	struct timer_list	sk_timer;
+	__u32			sk_priority;
+	__u32			sk_mark;
+	u32			sk_pacing_rate; /* bytes per second */
+	u32			sk_max_pacing_rate;
+	struct page_frag	sk_frag;
+	netdev_features_t	sk_route_caps;
+	netdev_features_t	sk_route_nocaps;
+	int			sk_gso_type;
+	unsigned int		sk_gso_max_size;
+	gfp_t			sk_allocation;
+	__u32			sk_txhash;
 
 	/*
 	 * Because of non atomicity rules, all
@@ -414,42 +435,24 @@ struct sock {
 #define SK_PROTOCOL_MAX U8_MAX
 	kmemcheck_bitfield_end(flags);
 
-	int			sk_wmem_queued;
-	gfp_t			sk_allocation;
-	u32			sk_pacing_rate; /* bytes per second */
-	u32			sk_max_pacing_rate;
-	netdev_features_t	sk_route_caps;
-	netdev_features_t	sk_route_nocaps;
-	int			sk_gso_type;
-	unsigned int		sk_gso_max_size;
 	u16			sk_gso_max_segs;
-	int			sk_rcvlowat;
 	unsigned long	        sk_lingertime;
-	struct sk_buff_head	sk_error_queue;
 	struct proto		*sk_prot_creator;
 	rwlock_t		sk_callback_lock;
 	int			sk_err,
 				sk_err_soft;
 	u32			sk_ack_backlog;
 	u32			sk_max_ack_backlog;
-	__u32			sk_priority;
-	__u32			sk_mark;
 	kuid_t			sk_uid;
 	struct pid		*sk_peer_pid;
 	const struct cred	*sk_peer_cred;
 	long			sk_rcvtimeo;
-	long			sk_sndtimeo;
-	struct timer_list	sk_timer;
 	ktime_t			sk_stamp;
 	u16			sk_tsflags;
 	u8			sk_shutdown;
 	u32			sk_tskey;
 	struct socket		*sk_socket;
 	void			*sk_user_data;
-	struct page_frag	sk_frag;
-	struct sk_buff		*sk_send_head;
-	__s32			sk_peek_off;
-	int			sk_write_pending;
 #ifdef CONFIG_SECURITY
 	void			*sk_security;
 #endif
diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
index 1149b48700a1..1ef3165114ba 100644
--- a/net/ipv4/tcp.c
+++ b/net/ipv4/tcp.c
@@ -663,9 +663,9 @@ static void tcp_push(struct sock *sk, int flags, int mss_now,
 	if (tcp_should_autocork(sk, skb, size_goal)) {
 
 		/* avoid atomic op if TSQ_THROTTLED bit is already set */
-		if (!test_bit(TSQ_THROTTLED, &tp->tsq_flags)) {
+		if (!test_bit(TSQ_THROTTLED, &sk->sk_tsq_flags)) {
 			NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
-			set_bit(TSQ_THROTTLED, &tp->tsq_flags);
+			set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
 		}
 		/* It is possible TX completion already happened
 		 * before we set TSQ_THROTTLED.
diff --git a/net/ipv4/tcp_ipv4.c b/net/ipv4/tcp_ipv4.c
index b50f05905ced..30d81f533ada 100644
--- a/net/ipv4/tcp_ipv4.c
+++ b/net/ipv4/tcp_ipv4.c
@@ -443,7 +443,7 @@ void tcp_v4_err(struct sk_buff *icmp_skb, u32 info)
 			if (!sock_owned_by_user(sk)) {
 				tcp_v4_mtu_reduced(sk);
 			} else {
-				if (!test_and_set_bit(TCP_MTU_REDUCED_DEFERRED, &tp->tsq_flags))
+				if (!test_and_set_bit(TCP_MTU_REDUCED_DEFERRED, &sk->sk_tsq_flags))
 					sock_hold(sk);
 			}
 			goto out;
diff --git a/net/ipv4/tcp_output.c b/net/ipv4/tcp_output.c
index c7adcb57654e..b45101f3d2bd 100644
--- a/net/ipv4/tcp_output.c
+++ b/net/ipv4/tcp_output.c
@@ -769,25 +769,26 @@ static void tcp_tasklet_func(unsigned long data)
 		list_del(&tp->tsq_node);
 
 		sk = (struct sock *)tp;
-		bh_lock_sock(sk);
-
-		if (!sock_owned_by_user(sk)) {
-			tcp_tsq_handler(sk);
-		} else {
-			/* defer the work to tcp_release_cb() */
-			set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags);
+		clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags);
+
+		if (!sk->sk_lock.owned &&
+		    test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) {
+			bh_lock_sock(sk);
+			if (!sock_owned_by_user(sk)) {
+				clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags);
+				tcp_tsq_handler(sk);
+			}
+			bh_unlock_sock(sk);
 		}
-		bh_unlock_sock(sk);
 
-		clear_bit(TSQ_QUEUED, &tp->tsq_flags);
 		sk_free(sk);
 	}
 }
 
-#define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) |		\
-			  (1UL << TCP_WRITE_TIMER_DEFERRED) |	\
-			  (1UL << TCP_DELACK_TIMER_DEFERRED) |	\
-			  (1UL << TCP_MTU_REDUCED_DEFERRED))
+#define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED |		\
+			  TCPF_WRITE_TIMER_DEFERRED |	\
+			  TCPF_DELACK_TIMER_DEFERRED |	\
+			  TCPF_MTU_REDUCED_DEFERRED)
 /**
  * tcp_release_cb - tcp release_sock() callback
  * @sk: socket
@@ -797,18 +798,17 @@ static void tcp_tasklet_func(unsigned long data)
  */
 void tcp_release_cb(struct sock *sk)
 {
-	struct tcp_sock *tp = tcp_sk(sk);
 	unsigned long flags, nflags;
 
 	/* perform an atomic operation only if at least one flag is set */
 	do {
-		flags = tp->tsq_flags;
+		flags = sk->sk_tsq_flags;
 		if (!(flags & TCP_DEFERRED_ALL))
 			return;
 		nflags = flags & ~TCP_DEFERRED_ALL;
-	} while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags);
+	} while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
 
-	if (flags & (1UL << TCP_TSQ_DEFERRED))
+	if (flags & TCPF_TSQ_DEFERRED)
 		tcp_tsq_handler(sk);
 
 	/* Here begins the tricky part :
@@ -822,15 +822,15 @@ void tcp_release_cb(struct sock *sk)
 	 */
 	sock_release_ownership(sk);
 
-	if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) {
+	if (flags & TCPF_WRITE_TIMER_DEFERRED) {
 		tcp_write_timer_handler(sk);
 		__sock_put(sk);
 	}
-	if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) {
+	if (flags & TCPF_DELACK_TIMER_DEFERRED) {
 		tcp_delack_timer_handler(sk);
 		__sock_put(sk);
 	}
-	if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) {
+	if (flags & TCPF_MTU_REDUCED_DEFERRED) {
 		inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
 		__sock_put(sk);
 	}
@@ -860,6 +860,7 @@ void tcp_wfree(struct sk_buff *skb)
 {
 	struct sock *sk = skb->sk;
 	struct tcp_sock *tp = tcp_sk(sk);
+	unsigned long flags, nval, oval;
 	int wmem;
 
 	/* Keep one reference on sk_wmem_alloc.
@@ -877,16 +878,25 @@ void tcp_wfree(struct sk_buff *skb)
 	if (wmem >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
 		goto out;
 
-	if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) &&
-	    !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) {
-		unsigned long flags;
+	for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
 		struct tsq_tasklet *tsq;
+		bool empty;
+
+		if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
+			goto out;
+
+		nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED | TCPF_TSQ_DEFERRED;
+		nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
+		if (nval != oval)
+			continue;
 
 		/* queue this socket to tasklet queue */
 		local_irq_save(flags);
 		tsq = this_cpu_ptr(&tsq_tasklet);
+		empty = list_empty(&tsq->head);
 		list_add(&tp->tsq_node, &tsq->head);
-		tasklet_schedule(&tsq->tasklet);
+		if (empty)
+			tasklet_schedule(&tsq->tasklet);
 		local_irq_restore(flags);
 		return;
 	}
@@ -1922,26 +1932,26 @@ static inline void tcp_mtu_check_reprobe(struct sock *sk)
  */
 static int tcp_mtu_probe(struct sock *sk)
 {
-	struct tcp_sock *tp = tcp_sk(sk);
 	struct inet_connection_sock *icsk = inet_csk(sk);
+	struct tcp_sock *tp = tcp_sk(sk);
 	struct sk_buff *skb, *nskb, *next;
 	struct net *net = sock_net(sk);
-	int len;
 	int probe_size;
 	int size_needed;
-	int copy;
+	int copy, len;
 	int mss_now;
 	int interval;
 
 	/* Not currently probing/verifying,
 	 * not in recovery,
 	 * have enough cwnd, and
-	 * not SACKing (the variable headers throw things off) */
-	if (!icsk->icsk_mtup.enabled ||
-	    icsk->icsk_mtup.probe_size ||
-	    inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
-	    tp->snd_cwnd < 11 ||
-	    tp->rx_opt.num_sacks || tp->rx_opt.dsack)
+	 * not SACKing (the variable headers throw things off)
+	 */
+	if (likely(!icsk->icsk_mtup.enabled ||
+		   icsk->icsk_mtup.probe_size ||
+		   inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
+		   tp->snd_cwnd < 11 ||
+		   tp->rx_opt.num_sacks || tp->rx_opt.dsack))
 		return -1;
 
 	/* Use binary search for probe_size between tcp_mss_base,
@@ -2081,7 +2091,16 @@ static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
 	limit <<= factor;
 
 	if (atomic_read(&sk->sk_wmem_alloc) > limit) {
-		set_bit(TSQ_THROTTLED, &tcp_sk(sk)->tsq_flags);
+		/* Always send the 1st or 2nd skb in write queue.
+		 * No need to wait for TX completion to call us back,
+		 * after softirq/tasklet schedule.
+		 * This helps when TX completions are delayed too much.
+		 */
+		if (skb == sk->sk_write_queue.next ||
+		    skb->prev == sk->sk_write_queue.next)
+			return false;
+
+		set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
 		/* It is possible TX completion already happened
 		 * before we set TSQ_THROTTLED, so we must
 		 * test again the condition.
@@ -2222,6 +2241,8 @@ static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
 		    unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
 			break;
 
+		if (test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags))
+			clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags);
 		if (tcp_small_queue_check(sk, skb, 0))
 			break;
 
@@ -3524,8 +3545,6 @@ void tcp_send_ack(struct sock *sk)
 	/* We do not want pure acks influencing TCP Small Queues or fq/pacing
 	 * too much.
 	 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
-	 * We also avoid tcp_wfree() overhead (cache line miss accessing
-	 * tp->tsq_flags) by using regular sock_wfree()
 	 */
 	skb_set_tcp_pure_ack(buff);
 
diff --git a/net/ipv4/tcp_timer.c b/net/ipv4/tcp_timer.c
index 3ea1cf804748..3705075f42c3 100644
--- a/net/ipv4/tcp_timer.c
+++ b/net/ipv4/tcp_timer.c
@@ -310,7 +310,7 @@ static void tcp_delack_timer(unsigned long data)
 		inet_csk(sk)->icsk_ack.blocked = 1;
 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED);
 		/* deleguate our work to tcp_release_cb() */
-		if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED, &tcp_sk(sk)->tsq_flags))
+		if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED, &sk->sk_tsq_flags))
 			sock_hold(sk);
 	}
 	bh_unlock_sock(sk);
@@ -592,7 +592,7 @@ static void tcp_write_timer(unsigned long data)
 		tcp_write_timer_handler(sk);
 	} else {
 		/* delegate our work to tcp_release_cb() */
-		if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED, &tcp_sk(sk)->tsq_flags))
+		if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED, &sk->sk_tsq_flags))
 			sock_hold(sk);
 	}
 	bh_unlock_sock(sk);
diff --git a/net/ipv6/tcp_ipv6.c b/net/ipv6/tcp_ipv6.c
index a2185a214abc..73bc8fc68acd 100644
--- a/net/ipv6/tcp_ipv6.c
+++ b/net/ipv6/tcp_ipv6.c
@@ -399,7 +399,7 @@ static void tcp_v6_err(struct sk_buff *skb, struct inet6_skb_parm *opt,
 		if (!sock_owned_by_user(sk))
 			tcp_v6_mtu_reduced(sk);
 		else if (!test_and_set_bit(TCP_MTU_REDUCED_DEFERRED,
-					   &tp->tsq_flags))
+					   &sk->sk_tsq_flags))
 			sock_hold(sk);
 		goto out;
 	}