tcp: Revert "tcp: remove header prediction"

This reverts commit 45f119bf93.

Eric Dumazet says:
  We found at Google a significant regression caused by
  45f119bf93 tcp: remove header prediction

  In typical RPC  (TCP_RR), when a TCP socket receives data, we now call
  tcp_ack() while we used to not call it.

  This touches enough cache lines to cause a slowdown.

so problem does not seem to be HP removal itself but the tcp_ack()
call.  Therefore, it might be possible to remove HP after all, provided
one finds a way to elide tcp_ack for most cases.

Reported-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: David S. Miller <davem@davemloft.net>

include/linux/tcp.h

@@ -147,6 +147,12 @@ struct tcp_sock {
 	u16	tcp_header_len;	/* Bytes of tcp header to send		*/
 	u16	gso_segs;	/* Max number of segs per GSO packet	*/
 
+/*
+ *	Header prediction flags
+ *	0x5?10 << 16 + snd_wnd in net byte order
+ */
+	__be32	pred_flags;
+
 /*
  *	RFC793 variables by their proper names. This means you can
  *	read the code and the spec side by side (and laugh ...)

include/net/tcp.h

@@ -634,6 +634,29 @@ static inline u32 __tcp_set_rto(const struct tcp_sock *tp)
 	return usecs_to_jiffies((tp->srtt_us >> 3) + tp->rttvar_us);
 }
 
+static inline void __tcp_fast_path_on(struct tcp_sock *tp, u32 snd_wnd)
+{
+	tp->pred_flags = htonl((tp->tcp_header_len << 26) |
+			       ntohl(TCP_FLAG_ACK) |
+			       snd_wnd);
+}
+
+static inline void tcp_fast_path_on(struct tcp_sock *tp)
+{
+	__tcp_fast_path_on(tp, tp->snd_wnd >> tp->rx_opt.snd_wscale);
+}
+
+static inline void tcp_fast_path_check(struct sock *sk)
+{
+	struct tcp_sock *tp = tcp_sk(sk);
+
+	if (RB_EMPTY_ROOT(&tp->out_of_order_queue) &&
+	    tp->rcv_wnd &&
+	    atomic_read(&sk->sk_rmem_alloc) < sk->sk_rcvbuf &&
+	    !tp->urg_data)
+		tcp_fast_path_on(tp);
+}
+
 /* Compute the actual rto_min value */
 static inline u32 tcp_rto_min(struct sock *sk)
 {

include/uapi/linux/snmp.h

@@ -184,7 +184,9 @@ enum
 	LINUX_MIB_DELAYEDACKLOST,		/* DelayedACKLost */
 	LINUX_MIB_LISTENOVERFLOWS,		/* ListenOverflows */
 	LINUX_MIB_LISTENDROPS,			/* ListenDrops */
+	LINUX_MIB_TCPHPHITS,			/* TCPHPHits */
 	LINUX_MIB_TCPPUREACKS,			/* TCPPureAcks */
+	LINUX_MIB_TCPHPACKS,			/* TCPHPAcks */
 	LINUX_MIB_TCPRENORECOVERY,		/* TCPRenoRecovery */
 	LINUX_MIB_TCPSACKRECOVERY,		/* TCPSackRecovery */
 	LINUX_MIB_TCPSACKRENEGING,		/* TCPSACKReneging */

net/ipv4/proc.c

@@ -206,7 +206,9 @@ static const struct snmp_mib snmp4_net_list[] = {
 	SNMP_MIB_ITEM("DelayedACKLost", LINUX_MIB_DELAYEDACKLOST),
 	SNMP_MIB_ITEM("ListenOverflows", LINUX_MIB_LISTENOVERFLOWS),
 	SNMP_MIB_ITEM("ListenDrops", LINUX_MIB_LISTENDROPS),
+	SNMP_MIB_ITEM("TCPHPHits", LINUX_MIB_TCPHPHITS),
 	SNMP_MIB_ITEM("TCPPureAcks", LINUX_MIB_TCPPUREACKS),
+	SNMP_MIB_ITEM("TCPHPAcks", LINUX_MIB_TCPHPACKS),
 	SNMP_MIB_ITEM("TCPRenoRecovery", LINUX_MIB_TCPRENORECOVERY),
 	SNMP_MIB_ITEM("TCPSackRecovery", LINUX_MIB_TCPSACKRECOVERY),
 	SNMP_MIB_ITEM("TCPSACKReneging", LINUX_MIB_TCPSACKRENEGING),

net/ipv4/tcp.c

@@ -1963,8 +1963,10 @@ int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock,
 		tcp_rcv_space_adjust(sk);
 
 skip_copy:
-		if (tp->urg_data && after(tp->copied_seq, tp->urg_seq))
+		if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
 			tp->urg_data = 0;
+			tcp_fast_path_check(sk);
+		}
 		if (used + offset < skb->len)
 			continue;
 

net/ipv4/tcp_input.c

@@ -103,6 +103,7 @@ int sysctl_tcp_invalid_ratelimit __read_mostly = HZ/2;
 #define FLAG_DATA_SACKED	0x20 /* New SACK.				*/
 #define FLAG_ECE		0x40 /* ECE in this ACK				*/
 #define FLAG_LOST_RETRANS	0x80 /* This ACK marks some retransmission lost */
+#define FLAG_SLOWPATH		0x100 /* Do not skip RFC checks for window update.*/
 #define FLAG_ORIG_SACK_ACKED	0x200 /* Never retransmitted data are (s)acked	*/
 #define FLAG_SND_UNA_ADVANCED	0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
 #define FLAG_DSACKING_ACK	0x800 /* SACK blocks contained D-SACK info */
@@ -3371,6 +3372,12 @@ static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32
 		if (tp->snd_wnd != nwin) {
 			tp->snd_wnd = nwin;
 
+			/* Note, it is the only place, where
+			 * fast path is recovered for sending TCP.
+			 */
+			tp->pred_flags = 0;
+			tcp_fast_path_check(sk);
+
 			if (tcp_send_head(sk))
 				tcp_slow_start_after_idle_check(sk);
 
@@ -3592,7 +3599,19 @@ static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
 	if (flag & FLAG_UPDATE_TS_RECENT)
 		tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
 
-	{
+	if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
+		/* Window is constant, pure forward advance.
+		 * No more checks are required.
+		 * Note, we use the fact that SND.UNA>=SND.WL2.
+		 */
+		tcp_update_wl(tp, ack_seq);
+		tcp_snd_una_update(tp, ack);
+		flag |= FLAG_WIN_UPDATE;
+
+		tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
+
+		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
+	} else {
 		u32 ack_ev_flags = CA_ACK_SLOWPATH;
 
 		if (ack_seq != TCP_SKB_CB(skb)->end_seq)
@@ -4407,6 +4426,8 @@ static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
 	if (TCP_SKB_CB(skb)->has_rxtstamp)
 		TCP_SKB_CB(skb)->swtstamp = skb->tstamp;
 
+	/* Disable header prediction. */
+	tp->pred_flags = 0;
 	inet_csk_schedule_ack(sk);
 
 	NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
@@ -4647,6 +4668,8 @@ queue_and_out:
 		if (tp->rx_opt.num_sacks)
 			tcp_sack_remove(tp);
 
+		tcp_fast_path_check(sk);
+
 		if (eaten > 0)
 			kfree_skb_partial(skb, fragstolen);
 		if (!sock_flag(sk, SOCK_DEAD))
@@ -4972,6 +4995,7 @@ static int tcp_prune_queue(struct sock *sk)
 	NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
 
 	/* Massive buffer overcommit. */
+	tp->pred_flags = 0;
 	return -1;
 }
 
@@ -5143,6 +5167,9 @@ static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
 
 	tp->urg_data = TCP_URG_NOTYET;
 	tp->urg_seq = ptr;
+
+	/* Disable header prediction. */
+	tp->pred_flags = 0;
 }
 
 /* This is the 'fast' part of urgent handling. */
@@ -5301,6 +5328,26 @@ discard:
 
 /*
  *	TCP receive function for the ESTABLISHED state.
+ *
+ *	It is split into a fast path and a slow path. The fast path is
+ * 	disabled when:
+ *	- A zero window was announced from us - zero window probing
+ *        is only handled properly in the slow path.
+ *	- Out of order segments arrived.
+ *	- Urgent data is expected.
+ *	- There is no buffer space left
+ *	- Unexpected TCP flags/window values/header lengths are received
+ *	  (detected by checking the TCP header against pred_flags)
+ *	- Data is sent in both directions. Fast path only supports pure senders
+ *	  or pure receivers (this means either the sequence number or the ack
+ *	  value must stay constant)
+ *	- Unexpected TCP option.
+ *
+ *	When these conditions are not satisfied it drops into a standard
+ *	receive procedure patterned after RFC793 to handle all cases.
+ *	The first three cases are guaranteed by proper pred_flags setting,
+ *	the rest is checked inline. Fast processing is turned on in
+ *	tcp_data_queue when everything is OK.
  */
 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
 			 const struct tcphdr *th)
@@ -5311,19 +5358,144 @@ void tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
 	tcp_mstamp_refresh(tp);
 	if (unlikely(!sk->sk_rx_dst))
 		inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
+	/*
+	 *	Header prediction.
+	 *	The code loosely follows the one in the famous
+	 *	"30 instruction TCP receive" Van Jacobson mail.
+	 *
+	 *	Van's trick is to deposit buffers into socket queue
+	 *	on a device interrupt, to call tcp_recv function
+	 *	on the receive process context and checksum and copy
+	 *	the buffer to user space. smart...
+	 *
+	 *	Our current scheme is not silly either but we take the
+	 *	extra cost of the net_bh soft interrupt processing...
+	 *	We do checksum and copy also but from device to kernel.
+	 */
 
 	tp->rx_opt.saw_tstamp = 0;
 
+	/*	pred_flags is 0xS?10 << 16 + snd_wnd
+	 *	if header_prediction is to be made
+	 *	'S' will always be tp->tcp_header_len >> 2
+	 *	'?' will be 0 for the fast path, otherwise pred_flags is 0 to
+	 *  turn it off	(when there are holes in the receive
+	 *	 space for instance)
+	 *	PSH flag is ignored.
+	 */
+
+	if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
+	    TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
+	    !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
+		int tcp_header_len = tp->tcp_header_len;
+
+		/* Timestamp header prediction: tcp_header_len
+		 * is automatically equal to th->doff*4 due to pred_flags
+		 * match.
+		 */
+
+		/* Check timestamp */
+		if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
+			/* No? Slow path! */
+			if (!tcp_parse_aligned_timestamp(tp, th))
+				goto slow_path;
+
+			/* If PAWS failed, check it more carefully in slow path */
+			if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
+				goto slow_path;
+
+			/* DO NOT update ts_recent here, if checksum fails
+			 * and timestamp was corrupted part, it will result
+			 * in a hung connection since we will drop all
+			 * future packets due to the PAWS test.
+			 */
+		}
+
+		if (len <= tcp_header_len) {
+			/* Bulk data transfer: sender */
+			if (len == tcp_header_len) {
+				/* Predicted packet is in window by definition.
+				 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
+				 * Hence, check seq<=rcv_wup reduces to:
+				 */
+				if (tcp_header_len ==
+				    (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
+				    tp->rcv_nxt == tp->rcv_wup)
+					tcp_store_ts_recent(tp);
+
+				/* We know that such packets are checksummed
+				 * on entry.
+				 */
+				tcp_ack(sk, skb, 0);
+				__kfree_skb(skb);
+				tcp_data_snd_check(sk);
+				return;
+			} else { /* Header too small */
+				TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
+				goto discard;
+			}
+		} else {
+			int eaten = 0;
+			bool fragstolen = false;
+
+			if (tcp_checksum_complete(skb))
+				goto csum_error;
+
+			if ((int)skb->truesize > sk->sk_forward_alloc)
+				goto step5;
+
+			/* Predicted packet is in window by definition.
+			 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
+			 * Hence, check seq<=rcv_wup reduces to:
+			 */
+			if (tcp_header_len ==
+			    (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
+			    tp->rcv_nxt == tp->rcv_wup)
+				tcp_store_ts_recent(tp);
+
+			tcp_rcv_rtt_measure_ts(sk, skb);
+
+			NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
+
+			/* Bulk data transfer: receiver */
+			eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
+					      &fragstolen);
+
+			tcp_event_data_recv(sk, skb);
+
+			if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
+				/* Well, only one small jumplet in fast path... */
+				tcp_ack(sk, skb, FLAG_DATA);
+				tcp_data_snd_check(sk);
+				if (!inet_csk_ack_scheduled(sk))
+					goto no_ack;
+			}
+
+			__tcp_ack_snd_check(sk, 0);
+no_ack:
+			if (eaten)
+				kfree_skb_partial(skb, fragstolen);
+			sk->sk_data_ready(sk);
+			return;
+		}
+	}
+
+slow_path:
 	if (len < (th->doff << 2) || tcp_checksum_complete(skb))
 		goto csum_error;
 
 	if (!th->ack && !th->rst && !th->syn)
 		goto discard;
 
+	/*
+	 *	Standard slow path.
+	 */
+
 	if (!tcp_validate_incoming(sk, skb, th, 1))
 		return;
 
-	if (tcp_ack(sk, skb, FLAG_UPDATE_TS_RECENT) < 0)
+step5:
+	if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
 		goto discard;
 
 	tcp_rcv_rtt_measure_ts(sk, skb);
@@ -5376,6 +5548,11 @@ void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
 
 	if (sock_flag(sk, SOCK_KEEPOPEN))
 		inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
+
+	if (!tp->rx_opt.snd_wscale)
+		__tcp_fast_path_on(tp, tp->snd_wnd);
+	else
+		tp->pred_flags = 0;
 }
 
 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
@@ -5504,7 +5681,7 @@ static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
 		tcp_ecn_rcv_synack(tp, th);
 
 		tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
-		tcp_ack(sk, skb, 0);
+		tcp_ack(sk, skb, FLAG_SLOWPATH);
 
 		/* Ok.. it's good. Set up sequence numbers and
 		 * move to established.
@@ -5740,8 +5917,8 @@ int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
 		return 0;
 
 	/* step 5: check the ACK field */
-
+	acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
-	acceptable = tcp_ack(sk, skb, FLAG_UPDATE_TS_RECENT |
+				      FLAG_UPDATE_TS_RECENT |
 				      FLAG_NO_CHALLENGE_ACK) > 0;
 
 	if (!acceptable) {
@@ -5809,6 +5986,7 @@ int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
 		tp->lsndtime = tcp_jiffies32;
 
 		tcp_initialize_rcv_mss(sk);
+		tcp_fast_path_on(tp);
 		break;
 
 	case TCP_FIN_WAIT1: {

net/ipv4/tcp_minisocks.c

@@ -436,6 +436,8 @@ struct sock *tcp_create_openreq_child(const struct sock *sk,
 		struct tcp_sock *newtp = tcp_sk(newsk);
 
 		/* Now setup tcp_sock */
+		newtp->pred_flags = 0;
+
 		newtp->rcv_wup = newtp->copied_seq =
 		newtp->rcv_nxt = treq->rcv_isn + 1;
 		newtp->segs_in = 1;

net/ipv4/tcp_output.c

@@ -295,7 +295,9 @@ static u16 tcp_select_window(struct sock *sk)
 	/* RFC1323 scaling applied */
 	new_win >>= tp->rx_opt.rcv_wscale;
 
+	/* If we advertise zero window, disable fast path. */
 	if (new_win == 0) {
+		tp->pred_flags = 0;
 		if (old_win)
 			NET_INC_STATS(sock_net(sk),
 				      LINUX_MIB_TCPTOZEROWINDOWADV);