dctcp: more accurate tracking of packets delivery
authorEric Dumazet <edumazet@google.com>
Thu, 11 Apr 2019 12:55:23 +0000 (05:55 -0700)
committerDavid S. Miller <davem@davemloft.net>
Fri, 12 Apr 2019 04:31:03 +0000 (21:31 -0700)
After commit e21db6f69a95 ("tcp: track total bytes delivered with ECN CE marks")
core TCP stack does a very good job tracking ECN signals.

The "sender's best estimate of CE information" Yuchung mentioned in his
patch is indeed the best we can do.

DCTCP can use tp->delivered_ce and tp->delivered to not duplicate the logic,
and use the existing best estimate.

This solves some problems, since current DCTCP logic does not deal with losses
and/or GRO or ack aggregation very well.

This also removes a dubious use of inet_csk(sk)->icsk_ack.rcv_mss
(this should have been tp->mss_cache), and a 64 bit divide.

Finally, we can see that the DCTCP logic, calling dctcp_update_alpha() for
every ACK could be done differently, calling it only once per RTT.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Soheil Hassas Yeganeh <soheil@google.com>
Cc: Florian Westphal <fw@strlen.de>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Lawrence Brakmo <brakmo@fb.com>
Cc: Abdul Kabbani <akabbani@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/ipv4/tcp_dctcp.c

index 359da68d7c0628360d5f1b727c878f678e28d39a..477cb4aa456c11c70185a982cbadafba857d3619 100644 (file)
@@ -49,9 +49,8 @@
 #define DCTCP_MAX_ALPHA        1024U
 
 struct dctcp {
-       u32 acked_bytes_ecn;
-       u32 acked_bytes_total;
-       u32 prior_snd_una;
+       u32 old_delivered;
+       u32 old_delivered_ce;
        u32 prior_rcv_nxt;
        u32 dctcp_alpha;
        u32 next_seq;
@@ -73,8 +72,8 @@ static void dctcp_reset(const struct tcp_sock *tp, struct dctcp *ca)
 {
        ca->next_seq = tp->snd_nxt;
 
-       ca->acked_bytes_ecn = 0;
-       ca->acked_bytes_total = 0;
+       ca->old_delivered = tp->delivered;
+       ca->old_delivered_ce = tp->delivered_ce;
 }
 
 static void dctcp_init(struct sock *sk)
@@ -86,7 +85,6 @@ static void dctcp_init(struct sock *sk)
             sk->sk_state == TCP_CLOSE)) {
                struct dctcp *ca = inet_csk_ca(sk);
 
-               ca->prior_snd_una = tp->snd_una;
                ca->prior_rcv_nxt = tp->rcv_nxt;
 
                ca->dctcp_alpha = min(dctcp_alpha_on_init, DCTCP_MAX_ALPHA);
@@ -118,37 +116,25 @@ static void dctcp_update_alpha(struct sock *sk, u32 flags)
 {
        const struct tcp_sock *tp = tcp_sk(sk);
        struct dctcp *ca = inet_csk_ca(sk);
-       u32 acked_bytes = tp->snd_una - ca->prior_snd_una;
-
-       /* If ack did not advance snd_una, count dupack as MSS size.
-        * If ack did update window, do not count it at all.
-        */
-       if (acked_bytes == 0 && !(flags & CA_ACK_WIN_UPDATE))
-               acked_bytes = inet_csk(sk)->icsk_ack.rcv_mss;
-       if (acked_bytes) {
-               ca->acked_bytes_total += acked_bytes;
-               ca->prior_snd_una = tp->snd_una;
-
-               if (flags & CA_ACK_ECE)
-                       ca->acked_bytes_ecn += acked_bytes;
-       }
 
        /* Expired RTT */
        if (!before(tp->snd_una, ca->next_seq)) {
-               u64 bytes_ecn = ca->acked_bytes_ecn;
+               u32 delivered_ce = tp->delivered_ce - ca->old_delivered_ce;
                u32 alpha = ca->dctcp_alpha;
 
                /* alpha = (1 - g) * alpha + g * F */
 
                alpha -= min_not_zero(alpha, alpha >> dctcp_shift_g);
-               if (bytes_ecn) {
+               if (delivered_ce) {
+                       u32 delivered = tp->delivered - ca->old_delivered;
+
                        /* If dctcp_shift_g == 1, a 32bit value would overflow
-                        * after 8 Mbytes.
+                        * after 8 M packets.
                         */
-                       bytes_ecn <<= (10 - dctcp_shift_g);
-                       do_div(bytes_ecn, max(1U, ca->acked_bytes_total));
+                       delivered_ce <<= (10 - dctcp_shift_g);
+                       delivered_ce /= max(1U, delivered);
 
-                       alpha = min(alpha + (u32)bytes_ecn, DCTCP_MAX_ALPHA);
+                       alpha = min(alpha + delivered_ce, DCTCP_MAX_ALPHA);
                }
                /* dctcp_alpha can be read from dctcp_get_info() without
                 * synchro, so we ask compiler to not use dctcp_alpha
@@ -200,6 +186,7 @@ static size_t dctcp_get_info(struct sock *sk, u32 ext, int *attr,
                             union tcp_cc_info *info)
 {
        const struct dctcp *ca = inet_csk_ca(sk);
+       const struct tcp_sock *tp = tcp_sk(sk);
 
        /* Fill it also in case of VEGASINFO due to req struct limits.
         * We can still correctly retrieve it later.
@@ -211,8 +198,10 @@ static size_t dctcp_get_info(struct sock *sk, u32 ext, int *attr,
                        info->dctcp.dctcp_enabled = 1;
                        info->dctcp.dctcp_ce_state = (u16) ca->ce_state;
                        info->dctcp.dctcp_alpha = ca->dctcp_alpha;
-                       info->dctcp.dctcp_ab_ecn = ca->acked_bytes_ecn;
-                       info->dctcp.dctcp_ab_tot = ca->acked_bytes_total;
+                       info->dctcp.dctcp_ab_ecn = tp->mss_cache *
+                                                  (tp->delivered_ce - ca->old_delivered_ce);
+                       info->dctcp.dctcp_ab_tot = tp->mss_cache *
+                                                  (tp->delivered - ca->old_delivered);
                }
 
                *attr = INET_DIAG_DCTCPINFO;