rxrpc: Move the cwnd degradation after transmitting packets

When we've gone for >1RTT without transmitting a packet, we should reduce
the ssthresh and cut the cwnd by half (as suggested in RFC2861 sec 3.1).

However, we may receive ACK packets in a batch and the first of these may
cut the cwnd, preventing further transmission, and each subsequent one cuts
the cwnd yet further, reducing it to the floor and killing performance.

Fix this by moving the cwnd reset to after doing the transmission and
resetting the base time such that we don't cut the cwnd by half again for
at least another RTT.

Signed-off-by: David Howells <dhowells@redhat.com>
cc: Marc Dionne <marc.dionne@auristor.com>
cc: linux-afs@lists.infradead.org

4 files changed, 39 insertions, 24 deletions

diff --git a/net/rxrpc/ar-internal.h b/net/rxrpc/ar-internal.h
index 6cfe366ee2244..785cd0dd1eea1 100644
--- a/net/rxrpc/ar-internal.h
+++ b/net/rxrpc/ar-internal.h
@@ -666,6 +666,7 @@ struct rxrpc_call {
 	 * packets) rather than bytes.
 	 */
 #define RXRPC_TX_SMSS		RXRPC_JUMBO_DATALEN
+#define RXRPC_MIN_CWND		(RXRPC_TX_SMSS > 2190 ? 2 : RXRPC_TX_SMSS > 1095 ? 3 : 4)
 	u8			cong_cwnd;	/* Congestion window size */
 	u8			cong_extra;	/* Extra to send for congestion management */
 	u8			cong_ssthresh;	/* Slow-start threshold */
@@ -953,6 +954,7 @@ void rxrpc_unpublish_service_conn(struct rxrpc_connection *);
 /*
  * input.c
  */
+void rxrpc_congestion_degrade(struct rxrpc_call *);
 void rxrpc_input_call_packet(struct rxrpc_call *, struct sk_buff *);
 void rxrpc_implicit_end_call(struct rxrpc_call *, struct sk_buff *);
 
diff --git a/net/rxrpc/call_event.c b/net/rxrpc/call_event.c
index 9f1e490ab9761..fd122e3726bd3 100644
--- a/net/rxrpc/call_event.c
+++ b/net/rxrpc/call_event.c
@@ -427,6 +427,13 @@ void rxrpc_input_call_event(struct rxrpc_call *call, struct sk_buff *skb)
 
 	rxrpc_transmit_some_data(call);
 
+	if (skb) {
+		struct rxrpc_skb_priv *sp = rxrpc_skb(skb);
+
+		if (sp->hdr.type == RXRPC_PACKET_TYPE_ACK)
+			rxrpc_congestion_degrade(call);
+	}
+
 	if (test_and_clear_bit(RXRPC_CALL_EV_INITIAL_PING, &call->events))
 		rxrpc_send_initial_ping(call);
 
diff --git a/net/rxrpc/input.c b/net/rxrpc/input.c
index 2988e3d0c1f6f..d0e20e946e48d 100644
--- a/net/rxrpc/input.c
+++ b/net/rxrpc/input.c
@@ -27,7 +27,6 @@ static void rxrpc_congestion_management(struct rxrpc_call *call,
 	enum rxrpc_congest_change change = rxrpc_cong_no_change;
 	unsigned int cumulative_acks = call->cong_cumul_acks;
 	unsigned int cwnd = call->cong_cwnd;
-	ktime_t now;
 	bool resend = false;
 
 	summary->flight_size =
@@ -57,27 +56,6 @@ static void rxrpc_congestion_management(struct rxrpc_call *call,
 	summary->cumulative_acks = cumulative_acks;
 	summary->dup_acks = call->cong_dup_acks;
 
-	/* If we haven't transmitted anything for >1RTT, we should reset the
-	 * congestion management state.
-	 */
-	now = ktime_get_real();
-	if ((call->cong_mode == RXRPC_CALL_SLOW_START ||
-	     call->cong_mode == RXRPC_CALL_CONGEST_AVOIDANCE) &&
-	    ktime_before(ktime_add_us(call->tx_last_sent,
-				      call->peer->srtt_us >> 3), now)
-	    ) {
-		trace_rxrpc_reset_cwnd(call, now);
-		change = rxrpc_cong_idle_reset;
-		rxrpc_inc_stat(call->rxnet, stat_tx_data_cwnd_reset);
-		summary->mode = RXRPC_CALL_SLOW_START;
-		if (RXRPC_TX_SMSS > 2190)
-			summary->cwnd = 2;
-		else if (RXRPC_TX_SMSS > 1095)
-			summary->cwnd = 3;
-		else
-			summary->cwnd = 4;
-	}
-
 	switch (call->cong_mode) {
 	case RXRPC_CALL_SLOW_START:
 		if (summary->saw_nacks)
@@ -198,6 +176,33 @@ send_extra_data:
 }
 
 /*
+ * Degrade the congestion window if we haven't transmitted a packet for >1RTT.
+ */
+void rxrpc_congestion_degrade(struct rxrpc_call *call)
+{
+	ktime_t rtt, now;
+
+	if (call->cong_mode != RXRPC_CALL_SLOW_START &&
+	    call->cong_mode != RXRPC_CALL_CONGEST_AVOIDANCE)
+		return;
+	if (call->state == RXRPC_CALL_CLIENT_AWAIT_REPLY)
+		return;
+
+	rtt = ns_to_ktime(call->peer->srtt_us * (1000 / 8));
+	now = ktime_get_real();
+	if (!ktime_before(ktime_add(call->tx_last_sent, rtt), now))
+		return;
+
+	trace_rxrpc_reset_cwnd(call, now);
+	rxrpc_inc_stat(call->rxnet, stat_tx_data_cwnd_reset);
+	call->tx_last_sent = now;
+	call->cong_mode = RXRPC_CALL_SLOW_START;
+	call->cong_ssthresh = max_t(unsigned int, call->cong_ssthresh,
+				    call->cong_cwnd * 3 / 4);
+	call->cong_cwnd = max_t(unsigned int, call->cong_cwnd / 2, RXRPC_MIN_CWND);
+}
+
+/*
  * Apply a hard ACK by advancing the Tx window.
  */
 static bool rxrpc_rotate_tx_window(struct rxrpc_call *call, rxrpc_seq_t to,
diff --git a/net/rxrpc/proc.c b/net/rxrpc/proc.c
index 6816934cb4cf1..3a59591ec0615 100644
--- a/net/rxrpc/proc.c
+++ b/net/rxrpc/proc.c
@@ -61,7 +61,7 @@ static int rxrpc_call_seq_show(struct seq_file *seq, void *v)
 			 "Proto Local                                          "
 			 " Remote                                         "
 			 " SvID ConnID   CallID   End Use State    Abort   "
-			 " DebugId  TxSeq    TW RxSeq    RW RxSerial RxTimo\n");
+			 " DebugId  TxSeq    TW RxSeq    RW RxSerial CW RxTimo\n");
 		return 0;
 	}
 
@@ -84,7 +84,7 @@ static int rxrpc_call_seq_show(struct seq_file *seq, void *v)
 	wtmp   = atomic64_read_acquire(&call->ackr_window);
 	seq_printf(seq,
 		   "UDP   %-47.47s %-47.47s %4x %08x %08x %s %3u"
-		   " %-8.8s %08x %08x %08x %02x %08x %02x %08x %06lx\n",
+		   " %-8.8s %08x %08x %08x %02x %08x %02x %08x %02x %06lx\n",
 		   lbuff,
 		   rbuff,
 		   call->dest_srx.srx_service,
@@ -98,6 +98,7 @@ static int rxrpc_call_seq_show(struct seq_file *seq, void *v)
 		   acks_hard_ack, READ_ONCE(call->tx_top) - acks_hard_ack,
 		   lower_32_bits(wtmp), upper_32_bits(wtmp) - lower_32_bits(wtmp),
 		   call->rx_serial,
+		   call->cong_cwnd,
 		   timeout);
 
 	return 0;