tcp: implement coalescing on backlog queue

In case GRO is not as efficient as it should be or disabled,
we might have a user thread trapped in __release_sock() while
softirq handler flood packets up to the point we have to drop.

This patch balances work done from user thread and softirq,
to give more chances to __release_sock() to complete its work
before new packets are added the the backlog.

This also helps if we receive many ACK packets, since GRO
does not aggregate them.

This patch brings ~60% throughput increase on a receiver
without GRO, but the spectacular gain is really on
1000x release_sock() latency reduction I have measured.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

3 files changed, 88 insertions, 6 deletions

diff --git a/include/uapi/linux/snmp.h b/include/uapi/linux/snmp.h
index f80135e5feaa..86dc24a96c90 100644
--- a/include/uapi/linux/snmp.h
+++ b/include/uapi/linux/snmp.h
@@ -243,6 +243,7 @@ enum
 	LINUX_MIB_TCPREQQFULLDROP,		/* TCPReqQFullDrop */
 	LINUX_MIB_TCPRETRANSFAIL,		/* TCPRetransFail */
 	LINUX_MIB_TCPRCVCOALESCE,		/* TCPRcvCoalesce */
+	LINUX_MIB_TCPBACKLOGCOALESCE,		/* TCPBacklogCoalesce */
 	LINUX_MIB_TCPOFOQUEUE,			/* TCPOFOQueue */
 	LINUX_MIB_TCPOFODROP,			/* TCPOFODrop */
 	LINUX_MIB_TCPOFOMERGE,			/* TCPOFOMerge */
diff --git a/net/ipv4/proc.c b/net/ipv4/proc.c
index 70289682a670..c3610b37bb4c 100644
--- a/net/ipv4/proc.c
+++ b/net/ipv4/proc.c
@@ -219,6 +219,7 @@ static const struct snmp_mib snmp4_net_list[] = {
 	SNMP_MIB_ITEM("TCPRenoRecoveryFail", LINUX_MIB_TCPRENORECOVERYFAIL),
 	SNMP_MIB_ITEM("TCPSackRecoveryFail", LINUX_MIB_TCPSACKRECOVERYFAIL),
 	SNMP_MIB_ITEM("TCPRcvCollapsed", LINUX_MIB_TCPRCVCOLLAPSED),
+	SNMP_MIB_ITEM("TCPBacklogCoalesce", LINUX_MIB_TCPBACKLOGCOALESCE),
 	SNMP_MIB_ITEM("TCPDSACKOldSent", LINUX_MIB_TCPDSACKOLDSENT),
 	SNMP_MIB_ITEM("TCPDSACKOfoSent", LINUX_MIB_TCPDSACKOFOSENT),
 	SNMP_MIB_ITEM("TCPDSACKRecv", LINUX_MIB_TCPDSACKRECV),
diff --git a/net/ipv4/tcp_ipv4.c b/net/ipv4/tcp_ipv4.c
index 795605a23275..4904250a9aac 100644
--- a/net/ipv4/tcp_ipv4.c
+++ b/net/ipv4/tcp_ipv4.c
@@ -1619,12 +1619,14 @@ int tcp_v4_early_demux(struct sk_buff *skb)
 bool tcp_add_backlog(struct sock *sk, struct sk_buff *skb)
 {
 	u32 limit = sk->sk_rcvbuf + sk->sk_sndbuf;
-
-	/* Only socket owner can try to collapse/prune rx queues
-	 * to reduce memory overhead, so add a little headroom here.
-	 * Few sockets backlog are possibly concurrently non empty.
-	 */
-	limit += 64*1024;
+	struct skb_shared_info *shinfo;
+	const struct tcphdr *th;
+	struct tcphdr *thtail;
+	struct sk_buff *tail;
+	unsigned int hdrlen;
+	bool fragstolen;
+	u32 gso_segs;
+	int delta;
 
 	/* In case all data was pulled from skb frags (in __pskb_pull_tail()),
 	 * we can fix skb->truesize to its real value to avoid future drops.
@@ -1636,6 +1638,84 @@ bool tcp_add_backlog(struct sock *sk, struct sk_buff *skb)
 
 	skb_dst_drop(skb);
 
+	if (unlikely(tcp_checksum_complete(skb))) {
+		bh_unlock_sock(sk);
+		__TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
+		__TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
+		return true;
+	}
+
+	/* Attempt coalescing to last skb in backlog, even if we are
+	 * above the limits.
+	 * This is okay because skb capacity is limited to MAX_SKB_FRAGS.
+	 */
+	th = (const struct tcphdr *)skb->data;
+	hdrlen = th->doff * 4;
+	shinfo = skb_shinfo(skb);
+
+	if (!shinfo->gso_size)
+		shinfo->gso_size = skb->len - hdrlen;
+
+	if (!shinfo->gso_segs)
+		shinfo->gso_segs = 1;
+
+	tail = sk->sk_backlog.tail;
+	if (!tail)
+		goto no_coalesce;
+	thtail = (struct tcphdr *)tail->data;
+
+	if (TCP_SKB_CB(tail)->end_seq != TCP_SKB_CB(skb)->seq ||
+	    TCP_SKB_CB(tail)->ip_dsfield != TCP_SKB_CB(skb)->ip_dsfield ||
+	    ((TCP_SKB_CB(tail)->tcp_flags |
+	      TCP_SKB_CB(skb)->tcp_flags) & TCPHDR_URG) ||
+	    ((TCP_SKB_CB(tail)->tcp_flags ^
+	      TCP_SKB_CB(skb)->tcp_flags) & (TCPHDR_ECE | TCPHDR_CWR)) ||
+#ifdef CONFIG_TLS_DEVICE
+	    tail->decrypted != skb->decrypted ||
+#endif
+	    thtail->doff != th->doff ||
+	    memcmp(thtail + 1, th + 1, hdrlen - sizeof(*th)))
+		goto no_coalesce;
+
+	__skb_pull(skb, hdrlen);
+	if (skb_try_coalesce(tail, skb, &fragstolen, &delta)) {
+		thtail->window = th->window;
+
+		TCP_SKB_CB(tail)->end_seq = TCP_SKB_CB(skb)->end_seq;
+
+		if (after(TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(tail)->ack_seq))
+			TCP_SKB_CB(tail)->ack_seq = TCP_SKB_CB(skb)->ack_seq;
+
+		TCP_SKB_CB(tail)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
+
+		if (TCP_SKB_CB(skb)->has_rxtstamp) {
+			TCP_SKB_CB(tail)->has_rxtstamp = true;
+			tail->tstamp = skb->tstamp;
+			skb_hwtstamps(tail)->hwtstamp = skb_hwtstamps(skb)->hwtstamp;
+		}
+
+		/* Not as strict as GRO. We only need to carry mss max value */
+		skb_shinfo(tail)->gso_size = max(shinfo->gso_size,
+						 skb_shinfo(tail)->gso_size);
+
+		gso_segs = skb_shinfo(tail)->gso_segs + shinfo->gso_segs;
+		skb_shinfo(tail)->gso_segs = min_t(u32, gso_segs, 0xFFFF);
+
+		sk->sk_backlog.len += delta;
+		__NET_INC_STATS(sock_net(sk),
+				LINUX_MIB_TCPBACKLOGCOALESCE);
+		kfree_skb_partial(skb, fragstolen);
+		return false;
+	}
+	__skb_push(skb, hdrlen);
+
+no_coalesce:
+	/* Only socket owner can try to collapse/prune rx queues
+	 * to reduce memory overhead, so add a little headroom here.
+	 * Few sockets backlog are possibly concurrently non empty.
+	 */
+	limit += 64*1024;
+
 	if (unlikely(sk_add_backlog(sk, skb, limit))) {
 		bh_unlock_sock(sk);
 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPBACKLOGDROP);