tcp: implement coalescing on backlog queue
authorEric Dumazet <edumazet@google.com>
Tue, 27 Nov 2018 22:42:03 +0000 (14:42 -0800)
committerDavid S. Miller <davem@davemloft.net>
Fri, 30 Nov 2018 21:26:54 +0000 (13:26 -0800)
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>
include/uapi/linux/snmp.h
net/ipv4/proc.c
net/ipv4/tcp_ipv4.c

index f80135e5feaa886000009db6dff75b2bc2d637b2..86dc24a96c90ab047d5173d625450facd6c6dd79 100644 (file)
@@ -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 */
index 70289682a6701438aed99a00a9705c39fa4394d3..c3610b37bb4ce665b1976d8cc907b6dd0de42ab9 100644 (file)
@@ -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),
index 795605a2327504b8a025405826e7e0ca8dc8501d..4904250a9aac5001410f9454258cbb8978bb8202 100644 (file)
@@ -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);