tcp: try to send bigger TSO packets

While investigating TCP performance, I found that TCP would
sometimes send big skbs followed by a single MSS skb,
in a 'locked' pattern.

For instance, BIG TCP is enabled, MSS is set to have 4096 bytes
of payload per segment. gso_max_size is set to 181000.

This means that an optimal TCP packet size should contain
44 * 4096 = 180224 bytes of payload,

However, I was seeing packets sizes interleaved in this pattern:

172032, 8192, 172032, 8192, 172032, 8192, <repeat>

tcp_tso_should_defer() heuristic is defeated, because after a split of
a packet in write queue for whatever reason (this might be a too small
CWND or a small enough pacing_rate),
the leftover packet in the queue is smaller than the optimal size.

It is time to try to make 'leftover packets' bigger so that
tcp_tso_should_defer() can give its full potential.

After this patch, we can see the following output:

14:13:34.009273 IP6 sender > receiver: Flags [P.], seq 4048380:4098360, ack 1, win 256, options [nop,nop,TS val 3425678144 ecr 1561784500], length 49980
14:13:34.010272 IP6 sender > receiver: Flags [P.], seq 4098360:4148340, ack 1, win 256, options [nop,nop,TS val 3425678145 ecr 1561784501], length 49980
14:13:34.011271 IP6 sender > receiver: Flags [P.], seq 4148340:4198320, ack 1, win 256, options [nop,nop,TS val 3425678146 ecr 1561784502], length 49980
14:13:34.012271 IP6 sender > receiver: Flags [P.], seq 4198320:4248300, ack 1, win 256, options [nop,nop,TS val 3425678147 ecr 1561784503], length 49980
14:13:34.013272 IP6 sender > receiver: Flags [P.], seq 4248300:4298280, ack 1, win 256, options [nop,nop,TS val 3425678148 ecr 1561784504], length 49980
14:13:34.014271 IP6 sender > receiver: Flags [P.], seq 4298280:4348260, ack 1, win 256, options [nop,nop,TS val 3425678149 ecr 1561784505], length 49980
14:13:34.015272 IP6 sender > receiver: Flags [P.], seq 4348260:4398240, ack 1, win 256, options [nop,nop,TS val 3425678150 ecr 1561784506], length 49980
14:13:34.016270 IP6 sender > receiver: Flags [P.], seq 4398240:4448220, ack 1, win 256, options [nop,nop,TS val 3425678151 ecr 1561784507], length 49980
14:13:34.017269 IP6 sender > receiver: Flags [P.], seq 4448220:4498200, ack 1, win 256, options [nop,nop,TS val 3425678152 ecr 1561784508], length 49980
14:13:34.018276 IP6 sender > receiver: Flags [P.], seq 4498200:4548180, ack 1, win 256, options [nop,nop,TS val 3425678153 ecr 1561784509], length 49980
14:13:34.019259 IP6 sender > receiver: Flags [P.], seq 4548180:4598160, ack 1, win 256, options [nop,nop,TS val 3425678154 ecr 1561784510], length 49980

With 200 concurrent flows on a 100Gbit NIC, we can see a reduction
of TSO packets (and ACK packets) of about 30 %.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://lore.kernel.org/r/20240418214600.1291486-4-edumazet@google.comSigned-off-by: Jakub Kicinski <kuba@kernel.org>

net/ipv4/tcp_output.c

@@ -2683,6 +2683,36 @@ void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
 		tcp_chrono_set(tp, TCP_CHRONO_BUSY);
 }
 
+/* First skb in the write queue is smaller than ideal packet size.
+ * Check if we can move payload from the second skb in the queue.
+ */
+static void tcp_grow_skb(struct sock *sk, struct sk_buff *skb, int amount)
+{
+	struct sk_buff *next_skb = skb->next;
+	unsigned int nlen;
+
+	if (tcp_skb_is_last(sk, skb))
+		return;
+
+	if (!tcp_skb_can_collapse(skb, next_skb))
+		return;
+
+	nlen = min_t(u32, amount, next_skb->len);
+	if (!nlen || !skb_shift(skb, next_skb, nlen))
+		return;
+
+	TCP_SKB_CB(skb)->end_seq += nlen;
+	TCP_SKB_CB(next_skb)->seq += nlen;
+
+	if (!next_skb->len) {
+		TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
+		TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
+		TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
+		tcp_unlink_write_queue(next_skb, sk);
+		tcp_wmem_free_skb(sk, next_skb);
+	}
+}
+
 /* This routine writes packets to the network.  It advances the
  * send_head.  This happens as incoming acks open up the remote
  * window for us.
@@ -2723,6 +2753,7 @@ static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
 	max_segs = tcp_tso_segs(sk, mss_now);
 	while ((skb = tcp_send_head(sk))) {
 		unsigned int limit;
+		int missing_bytes;
 
 		if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
 			/* "skb_mstamp_ns" is used as a start point for the retransmit timer */
@@ -2744,6 +2775,10 @@ static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
 			else
 				break;
 		}
+		cwnd_quota = min(cwnd_quota, max_segs);
+		missing_bytes = cwnd_quota * mss_now - skb->len;
+		if (missing_bytes > 0)
+			tcp_grow_skb(sk, skb, missing_bytes);
 
 		tso_segs = tcp_set_skb_tso_segs(skb, mss_now);
 
@@ -2767,8 +2802,7 @@ static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
 		limit = mss_now;
 		if (tso_segs > 1 && !tcp_urg_mode(tp))
 			limit = tcp_mss_split_point(sk, skb, mss_now,
-						    min(cwnd_quota,
-							max_segs),
+						    cwnd_quota,
 						    nonagle);
 
 		if (skb->len > limit &&