tcp: add net.ipv4.tcp_rcvbuf_low_rtt

author Eric Dumazet <edumazet@google.com>

Wed, 19 Nov 2025 08:48:13 +0000 (08:48 +0000)

committer Jakub Kicinski <kuba@kernel.org>

Fri, 21 Nov 2025 01:44:23 +0000 (17:44 -0800)
author Eric Dumazet <edumazet@google.com>
Wed, 19 Nov 2025 08:48:13 +0000 (08:48 +0000)
committer Jakub Kicinski <kuba@kernel.org>
Fri, 21 Nov 2025 01:44:23 +0000 (17:44 -0800)
diff --git a/Documentation/networking/ip-sysctl.rst b/Documentation/networking/ip-sysctl.rst

index f4ad739a6b532914e4091c425828b329ee342bc6..bc9a01606daf5a3c86477d74fdc8565eb44fd508 100644 (file)
--- a/Documentation/networking/ip-sysctl.rst
+++ b/Documentation/networking/ip-sysctl.rst
@@ -673,6 +673,16 @@ tcp_moderate_rcvbuf - BOOLEAN
  
         Default: 1 (enabled)
  
+tcp_rcvbuf_low_rtt - INTEGER
+       rcvbuf autotuning can over estimate final socket rcvbuf, which
+       can lead to cache trashing for high throughput flows.
+
+       For small RTT flows (below tcp_rcvbuf_low_rtt usecs), we can relax
+       rcvbuf growth: Few additional ms to reach the final (and smaller)
+       rcvbuf is a good tradeoff.
+
+       Default : 1000 (1 ms)
+
  tcp_mtu_probing - INTEGER
         Controls TCP Packetization-Layer Path MTU Discovery.  Takes three
         values:
diff --git a/Documentation/networking/net_cachelines/netns_ipv4_sysctl.rst b/Documentation/networking/net_cachelines/netns_ipv4_sysctl.rst

index 5d5d54fb6ab1b2697d06e0b0ba8c0a91b5dbd438..beaf1880a19bf4cfa578e162c571e09f7a9dffbe 100644 (file)
--- a/Documentation/networking/net_cachelines/netns_ipv4_sysctl.rst
+++ b/Documentation/networking/net_cachelines/netns_ipv4_sysctl.rst
@@ -103,6 +103,7 @@ u8                              sysctl_tcp_frto
  u8                              sysctl_tcp_nometrics_save                                                            TCP_LAST_ACK/tcp_update_metrics
  u8                              sysctl_tcp_no_ssthresh_metrics_save                                                  TCP_LAST_ACK/tcp_(update/init)_metrics
  u8                              sysctl_tcp_moderate_rcvbuf                                       read_mostly         tcp_rcvbuf_grow()
+u32                             sysctl_tcp_rcvbuf_low_rtt                                        read_mostly         tcp_rcvbuf_grow()
  u8                              sysctl_tcp_tso_win_divisor                   read_mostly                             tcp_tso_should_defer(tcp_write_xmit)
  u8                              sysctl_tcp_workaround_signed_windows                                                 tcp_select_window
  int                             sysctl_tcp_limit_output_bytes                read_mostly                             tcp_small_queue_check(tcp_write_xmit)
diff --git a/include/net/netns/ipv4.h b/include/net/netns/ipv4.h

index 11837d3ccc0ab6dbd6eaacc32536c912b3752202..2dbd46fc4734b78201818ecf6065237d475101ce 100644 (file)
--- a/include/net/netns/ipv4.h
+++ b/include/net/netns/ipv4.h
@@ -85,6 +85,7 @@ struct netns_ipv4 {
         /* 3 bytes hole, try to pack */
         int sysctl_tcp_reordering;
         int sysctl_tcp_rmem[3];
+       int sysctl_tcp_rcvbuf_low_rtt;
         __cacheline_group_end(netns_ipv4_read_rx);
  
         struct inet_timewait_death_row tcp_death_row;
diff --git a/net/core/net_namespace.c b/net/core/net_namespace.c

index c8adbbe014518602857b5f36b90da64333fbeafd..dfad7c03b80945d8be0444caa202f127ab3854e3 100644 (file)
--- a/net/core/net_namespace.c
+++ b/net/core/net_namespace.c
@@ -1227,6 +1227,8 @@ static void __init netns_ipv4_struct_check(void)
         /* RX readonly hotpath cache line */
         CACHELINE_ASSERT_GROUP_MEMBER(struct netns_ipv4, netns_ipv4_read_rx,
                                       sysctl_tcp_moderate_rcvbuf);
+       CACHELINE_ASSERT_GROUP_MEMBER(struct netns_ipv4, netns_ipv4_read_rx,
+                                     sysctl_tcp_rcvbuf_low_rtt);
         CACHELINE_ASSERT_GROUP_MEMBER(struct netns_ipv4, netns_ipv4_read_rx,
                                       sysctl_ip_early_demux);
         CACHELINE_ASSERT_GROUP_MEMBER(struct netns_ipv4, netns_ipv4_read_rx,
diff --git a/net/ipv4/sysctl_net_ipv4.c b/net/ipv4/sysctl_net_ipv4.c

index 35367f8e2da32f2c7de5a06164f5e47c8929c8f1..a1a50a5c80dc11eac19aa5c4d2ef0d48ab12a071 100644 (file)
--- a/net/ipv4/sysctl_net_ipv4.c
+++ b/net/ipv4/sysctl_net_ipv4.c
@@ -1342,6 +1342,15 @@ static struct ctl_table ipv4_net_table[] = {
                 .mode           = 0644,
                 .proc_handler   = proc_dou8vec_minmax,
         },
+       {
+               .procname       = "tcp_rcvbuf_low_rtt",
+               .data           = &init_net.ipv4.sysctl_tcp_rcvbuf_low_rtt,
+               .maxlen         = sizeof(int),
+               .mode           = 0644,
+               .proc_handler   = proc_dointvec_minmax,
+               .extra1         = SYSCTL_ZERO,
+               .extra2         = SYSCTL_INT_MAX,
+       },
         {
                 .procname       = "tcp_tso_win_divisor",
                 .data           = &init_net.ipv4.sysctl_tcp_tso_win_divisor,
diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c

index 9df5d75156057e6ba6b64ff7a0517809e8d1d49a..198f8a0d37be04f78da9268a230c9494b50b672a 100644 (file)
--- a/net/ipv4/tcp_input.c
+++ b/net/ipv4/tcp_input.c
@@ -896,6 +896,7 @@ void tcp_rcvbuf_grow(struct sock *sk, u32 newval)
         const struct net *net = sock_net(sk);
         struct tcp_sock *tp = tcp_sk(sk);
         u32 rcvwin, rcvbuf, cap, oldval;
+       u32 rtt_threshold, rtt_us;
         u64 grow;
  
         oldval = tp->rcvq_space.space;
@@ -908,10 +909,19 @@ void tcp_rcvbuf_grow(struct sock *sk, u32 newval)
         /* DRS is always one RTT late. */
         rcvwin = newval << 1;
  
-       /* slow start: allow the sender to double its rate. */
-       grow = (u64)rcvwin * (newval - oldval);
-       do_div(grow, oldval);
-       rcvwin += grow << 1;
+       rtt_us = tp->rcv_rtt_est.rtt_us >> 3;
+       rtt_threshold = READ_ONCE(net->ipv4.sysctl_tcp_rcvbuf_low_rtt);
+       if (rtt_us < rtt_threshold) {
+               /* For small RTT, we set @grow to rcvwin * rtt_us/rtt_threshold.
+                * It might take few additional ms to reach 'line rate',
+                * but will avoid sk_rcvbuf inflation and poor cache use.
+                */
+               grow = div_u64((u64)rcvwin * rtt_us, rtt_threshold);
+       } else {
+               /* slow start: allow the sender to double its rate. */
+               grow = div_u64(((u64)rcvwin << 1) * (newval - oldval), oldval);
+       }
+       rcvwin += grow;
  
         if (!RB_EMPTY_ROOT(&tp->out_of_order_queue))
                 rcvwin += TCP_SKB_CB(tp->ooo_last_skb)->end_seq - tp->rcv_nxt;
diff --git a/net/ipv4/tcp_ipv4.c b/net/ipv4/tcp_ipv4.c

index 6fcaecb67284ecade97b623d955dbbe2cd02a831..e0bb8d9e2d9c8c4a49519655d627e6e4d1b1cbac 100644 (file)
--- a/net/ipv4/tcp_ipv4.c
+++ b/net/ipv4/tcp_ipv4.c
@@ -3566,6 +3566,7 @@ static int __net_init tcp_sk_init(struct net *net)
         net->ipv4.sysctl_tcp_adv_win_scale = 1;
         net->ipv4.sysctl_tcp_frto = 2;
         net->ipv4.sysctl_tcp_moderate_rcvbuf = 1;
+       net->ipv4.sysctl_tcp_rcvbuf_low_rtt = USEC_PER_MSEC;
         /* This limits the percentage of the congestion window which we
          * will allow a single TSO frame to consume.  Building TSO frames
          * which are too large can cause TCP streams to be bursty.
author	Eric Dumazet <edumazet@google.com>
	Wed, 19 Nov 2025 08:48:13 +0000 (08:48 +0000)
committer	Jakub Kicinski <kuba@kernel.org>
	Fri, 21 Nov 2025 01:44:23 +0000 (17:44 -0800)
Documentation/networking/ip-sysctl.rst		patch \| blob \| blame \| history
Documentation/networking/net_cachelines/netns_ipv4_sysctl.rst		patch \| blob \| blame \| history
include/net/netns/ipv4.h		patch \| blob \| blame \| history
net/core/net_namespace.c		patch \| blob \| blame \| history
net/ipv4/sysctl_net_ipv4.c		patch \| blob \| blame \| history
net/ipv4/tcp_input.c		patch \| blob \| blame \| history
net/ipv4/tcp_ipv4.c		patch \| blob \| blame \| history