tcp: adjust TSO packet sizes based on min_rtt

Back when tcp_tso_autosize() and TCP pacing were introduced,
our focus was really to reduce burst sizes for long distance
flows.

The simple heuristic of using sk_pacing_rate/1024 has worked
well, but can lead to too small packets for hosts in the same
rack/cluster, when thousands of flows compete for the bottleneck.

Neal Cardwell had the idea of making the TSO burst size
a function of both sk_pacing_rate and tcp_min_rtt()

Indeed, for local flows, sending bigger bursts is better
to reduce cpu costs, as occasional losses can be repaired
quite fast.

This patch is based on Neal Cardwell implementation
done more than two years ago.
bbr is adjusting max_pacing_rate based on measured bandwidth,
while cubic would over estimate max_pacing_rate.

/proc/sys/net/ipv4/tcp_tso_rtt_log can be used to tune or disable
this new feature, in logarithmic steps.

Tested:

100Gbit NIC, two hosts in the same rack, 4K MTU.
600 flows rate-limited to 20000000 bytes per second.

Before patch: (TSO sizes would be limited to 20000000/1024/4096 -> 4 segments per TSO)

~# echo 0 >/proc/sys/net/ipv4/tcp_tso_rtt_log
~# nstat -n;perf stat ./super_netperf 600 -H otrv6 -l 20 -- -K dctcp -q 20000000;nstat|egrep "TcpInSegs|TcpOutSegs|TcpRetransSegs|Delivered"
  96005

 Performance counter stats for './super_netperf 600 -H otrv6 -l 20 -- -K dctcp -q 20000000':

         65,945.29 msec task-clock                #    2.845 CPUs utilized
         1,314,632      context-switches          # 19935.279 M/sec
             5,292      cpu-migrations            #   80.249 M/sec
           940,641      page-faults               # 14264.023 M/sec
   201,117,030,926      cycles                    # 3049769.216 GHz                   (83.45%)
    17,699,435,405      stalled-cycles-frontend   #    8.80% frontend cycles idle     (83.48%)
   136,584,015,071      stalled-cycles-backend    #   67.91% backend cycles idle      (83.44%)
    53,809,530,436      instructions              #    0.27  insn per cycle
                                                  #    2.54  stalled cycles per insn  (83.36%)
     9,062,315,523      branches                  # 137422329.563 M/sec               (83.22%)
       153,008,621      branch-misses             #    1.69% of all branches          (83.32%)

      23.182970846 seconds time elapsed

TcpInSegs                       15648792           0.0
TcpOutSegs                      58659110           0.0  # Average of 3.7 4K segments per TSO packet
TcpExtTCPDelivered              58654791           0.0
TcpExtTCPDeliveredCE            19                 0.0

After patch:

~# echo 9 >/proc/sys/net/ipv4/tcp_tso_rtt_log
~# nstat -n;perf stat ./super_netperf 600 -H otrv6 -l 20 -- -K dctcp -q 20000000;nstat|egrep "TcpInSegs|TcpOutSegs|TcpRetransSegs|Delivered"
  96046

 Performance counter stats for './super_netperf 600 -H otrv6 -l 20 -- -K dctcp -q 20000000':

         48,982.58 msec task-clock                #    2.104 CPUs utilized
           186,014      context-switches          # 3797.599 M/sec
             3,109      cpu-migrations            #   63.472 M/sec
           941,180      page-faults               # 19214.814 M/sec
   153,459,763,868      cycles                    # 3132982.807 GHz                   (83.56%)
    12,069,861,356      stalled-cycles-frontend   #    7.87% frontend cycles idle     (83.32%)
   120,485,917,953      stalled-cycles-backend    #   78.51% backend cycles idle      (83.24%)
    36,803,672,106      instructions              #    0.24  insn per cycle
                                                  #    3.27  stalled cycles per insn  (83.18%)
     5,947,266,275      branches                  # 121417383.427 M/sec               (83.64%)
        87,984,616      branch-misses             #    1.48% of all branches          (83.43%)

      23.281200256 seconds time elapsed

TcpInSegs                       1434706            0.0
TcpOutSegs                      58883378           0.0  # Average of 41 4K segments per TSO packet
TcpExtTCPDelivered              58878971           0.0
TcpExtTCPDeliveredCE            9664               0.0
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Neal Cardwell <ncardwell@google.com>
Link: https://lore.kernel.org/r/20220309015757.2532973-1-eric.dumazet@gmail.comSigned-off-by: Jakub Kicinski <kuba@kernel.org>

Documentation/networking/ip-sysctl.rst

@@ -878,6 +878,29 @@ tcp_min_tso_segs - INTEGER
 
 	Default: 2
 
+tcp_tso_rtt_log - INTEGER
+	Adjustment of TSO packet sizes based on min_rtt
+
+	Starting from linux-5.18, TCP autosizing can be tweaked
+	for flows having small RTT.
+
+	Old autosizing was splitting the pacing budget to send 1024 TSO
+	per second.
+
+	tso_packet_size = sk->sk_pacing_rate / 1024;
+
+	With the new mechanism, we increase this TSO sizing using:
+
+	distance = min_rtt_usec / (2^tcp_tso_rtt_log)
+	tso_packet_size += gso_max_size >> distance;
+
+	This means that flows between very close hosts can use bigger
+	TSO packets, reducing their cpu costs.
+
+	If you want to use the old autosizing, set this sysctl to 0.
+
+	Default: 9  (2^9 = 512 usec)
+
 tcp_pacing_ss_ratio - INTEGER
 	sk->sk_pacing_rate is set by TCP stack using a ratio applied
 	to current rate. (current_rate = cwnd * mss / srtt)

include/net/netns/ipv4.h

@@ -127,6 +127,7 @@ struct netns_ipv4 {
 	u8 sysctl_tcp_synack_retries;
 	u8 sysctl_tcp_syncookies;
 	u8 sysctl_tcp_migrate_req;
+	u8 sysctl_tcp_comp_sack_nr;
 	int sysctl_tcp_reordering;
 	u8 sysctl_tcp_retries1;
 	u8 sysctl_tcp_retries2;
@@ -160,9 +161,9 @@ struct netns_ipv4 {
 	int sysctl_tcp_challenge_ack_limit;
 	int sysctl_tcp_min_rtt_wlen;
 	u8 sysctl_tcp_min_tso_segs;
+	u8 sysctl_tcp_tso_rtt_log;
 	u8 sysctl_tcp_autocorking;
 	u8 sysctl_tcp_reflect_tos;
-	u8 sysctl_tcp_comp_sack_nr;
 	int sysctl_tcp_invalid_ratelimit;
 	int sysctl_tcp_pacing_ss_ratio;
 	int sysctl_tcp_pacing_ca_ratio;

net/ipv4/sysctl_net_ipv4.c

@@ -1271,6 +1271,13 @@ static struct ctl_table ipv4_net_table[] = {
 		.proc_handler	= proc_dou8vec_minmax,
 		.extra1		= SYSCTL_ONE,
 	},
+	{
+		.procname	= "tcp_tso_rtt_log",
+		.data		= &init_net.ipv4.sysctl_tcp_tso_rtt_log,
+		.maxlen		= sizeof(u8),
+		.mode		= 0644,
+		.proc_handler	= proc_dou8vec_minmax,
+	},
 	{
 		.procname	= "tcp_min_rtt_wlen",
 		.data		= &init_net.ipv4.sysctl_tcp_min_rtt_wlen,

net/ipv4/tcp_ipv4.c

@@ -3137,6 +3137,7 @@ static int __net_init tcp_sk_init(struct net *net)
 	/* rfc5961 challenge ack rate limiting */
 	net->ipv4.sysctl_tcp_challenge_ack_limit = 1000;
 	net->ipv4.sysctl_tcp_min_tso_segs = 2;
+	net->ipv4.sysctl_tcp_tso_rtt_log = 9;  /* 2^9 = 512 usec */
 	net->ipv4.sysctl_tcp_min_rtt_wlen = 300;
 	net->ipv4.sysctl_tcp_autocorking = 1;
 	net->ipv4.sysctl_tcp_invalid_ratelimit = HZ/2;

net/ipv4/tcp_output.c

@@ -1951,25 +1951,34 @@ static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
 }
 
 /* Return how many segs we'd like on a TSO packet,
- * to send one TSO packet per ms
+ * depending on current pacing rate, and how close the peer is.
+ *
+ * Rationale is:
+ * - For close peers, we rather send bigger packets to reduce
+ *   cpu costs, because occasional losses will be repaired fast.
+ * - For long distance/rtt flows, we would like to get ACK clocking
+ *   with 1 ACK per ms.
+ *
+ * Use min_rtt to help adapt TSO burst size, with smaller min_rtt resulting
+ * in bigger TSO bursts. We we cut the RTT-based allowance in half
+ * for every 2^9 usec (aka 512 us) of RTT, so that the RTT-based allowance
+ * is below 1500 bytes after 6 * ~500 usec = 3ms.
  */
 static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
 			    int min_tso_segs)
 {
-	u32 bytes, segs;
+	unsigned long bytes;
+	u32 r;
 
-	bytes = min_t(unsigned long,
-		      sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift),
-		      sk->sk_gso_max_size);
+	bytes = sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift);
 
-	/* Goal is to send at least one packet per ms,
-	 * not one big TSO packet every 100 ms.
-	 * This preserves ACK clocking and is consistent
-	 * with tcp_tso_should_defer() heuristic.
-	 */
-	segs = max_t(u32, bytes / mss_now, min_tso_segs);
+	r = tcp_min_rtt(tcp_sk(sk)) >> sock_net(sk)->ipv4.sysctl_tcp_tso_rtt_log;
+	if (r < BITS_PER_TYPE(sk->sk_gso_max_size))
+		bytes += sk->sk_gso_max_size >> r;
+
+	bytes = min_t(unsigned long, bytes, sk->sk_gso_max_size);
 
-	return segs;
+	return max_t(u32, bytes / mss_now, min_tso_segs);
 }
 
 /* Return the number of segments we want in the skb we are transmitting.