rxrpc: Fix RTT gathering

Fix RTT information gathering in AF_RXRPC by the following means:

 (1) Enable Rx timestamping on the transport socket with SO_TIMESTAMPNS.

 (2) If the sk_buff doesn't have a timestamp set when rxrpc_data_ready()
     collects it, set it at that point.

 (3) Allow ACKs to be requested on the last packet of a client call, but
     not a service call.  We need to be careful lest we undo:

	bf7d620abf22c321208a4da4f435e7af52551a21
	Author: David Howells <dhowells@redhat.com>
	Date:   Thu Oct 6 08:11:51 2016 +0100
	rxrpc: Don't request an ACK on the last DATA packet of a call's Tx phase

     but that only really applies to service calls that we're handling,
     since the client side gets to send the final ACK (or not).

 (4) When about to transmit an ACK or DATA packet, record the Tx timestamp
     before only; don't update the timestamp afterwards.

 (5) Switch the ordering between recording the serial and recording the
     timestamp to always set the serial number first.  The serial number
     shouldn't be seen referenced by an ACK packet until we've transmitted
     the packet bearing it - so in the Rx path, we don't need the timestamp
     until we've checked the serial number.

Fixes: cf1a6474f807 ("rxrpc: Add per-peer RTT tracker")
Signed-off-by: David Howells <dhowells@redhat.com>

net/rxrpc/input.c

@@ -622,13 +622,14 @@ static void rxrpc_input_requested_ack(struct rxrpc_call *call,
 		if (!skb)
 			continue;
 
+		sent_at = skb->tstamp;
+		smp_rmb(); /* Read timestamp before serial. */
 		sp = rxrpc_skb(skb);
 		if (sp->hdr.serial != orig_serial)
 			continue;
-		smp_rmb();
-		sent_at = skb->tstamp;
 		goto found;
 	}
+
 	return;
 
 found:
@@ -1143,6 +1144,9 @@ void rxrpc_data_ready(struct sock *udp_sk)
 		return;
 	}
 
+	if (skb->tstamp == 0)
+		skb->tstamp = ktime_get_real();
+
 	rxrpc_new_skb(skb, rxrpc_skb_rx_received);
 
 	_net("recv skb %p", skb);

net/rxrpc/local_object.c

@@ -173,6 +173,15 @@ static int rxrpc_open_socket(struct rxrpc_local *local, struct net *net)
 			_debug("setsockopt failed");
 			goto error;
 		}
+
+		/* We want receive timestamps. */
+		opt = 1;
+		ret = kernel_setsockopt(local->socket, SOL_SOCKET, SO_TIMESTAMPNS,
+					(char *)&opt, sizeof(opt));
+		if (ret < 0) {
+			_debug("setsockopt failed");
+			goto error;
+		}
 		break;
 
 	default:

net/rxrpc/output.c

@@ -124,7 +124,6 @@ int rxrpc_send_ack_packet(struct rxrpc_call *call, bool ping,
 	struct kvec iov[2];
 	rxrpc_serial_t serial;
 	rxrpc_seq_t hard_ack, top;
-	ktime_t now;
 	size_t len, n;
 	int ret;
 	u8 reason;
@@ -196,9 +195,7 @@ int rxrpc_send_ack_packet(struct rxrpc_call *call, bool ping,
 		/* We need to stick a time in before we send the packet in case
 		 * the reply gets back before kernel_sendmsg() completes - but
 		 * asking UDP to send the packet can take a relatively long
-		 * time, so we update the time after, on the assumption that
-		 * the packet transmission is more likely to happen towards the
-		 * end of the kernel_sendmsg() call.
+		 * time.
 		 */
 		call->ping_time = ktime_get_real();
 		set_bit(RXRPC_CALL_PINGING, &call->flags);
@@ -206,9 +203,6 @@ int rxrpc_send_ack_packet(struct rxrpc_call *call, bool ping,
 	}
 
 	ret = kernel_sendmsg(conn->params.local->socket, &msg, iov, 2, len);
-	now = ktime_get_real();
-	if (ping)
-		call->ping_time = now;
 	conn->params.peer->last_tx_at = ktime_get_seconds();
 	if (ret < 0)
 		trace_rxrpc_tx_fail(call->debug_id, serial, ret,
@@ -363,8 +357,14 @@ int rxrpc_send_data_packet(struct rxrpc_call *call, struct sk_buff *skb,
 
 	/* If our RTT cache needs working on, request an ACK.  Also request
 	 * ACKs if a DATA packet appears to have been lost.
+	 *
+	 * However, we mustn't request an ACK on the last reply packet of a
+	 * service call, lest OpenAFS incorrectly send us an ACK with some
+	 * soft-ACKs in it and then never follow up with a proper hard ACK.
 	 */
-	if (!(sp->hdr.flags & RXRPC_LAST_PACKET) &&
+	if ((!(sp->hdr.flags & RXRPC_LAST_PACKET) ||
+	     rxrpc_to_server(sp)
+	     ) &&
 	    (test_and_clear_bit(RXRPC_CALL_EV_ACK_LOST, &call->events) ||
 	     retrans ||
 	     call->cong_mode == RXRPC_CALL_SLOW_START ||
@@ -390,6 +390,11 @@ int rxrpc_send_data_packet(struct rxrpc_call *call, struct sk_buff *skb,
 		goto send_fragmentable;
 
 	down_read(&conn->params.local->defrag_sem);
+
+	sp->hdr.serial = serial;
+	smp_wmb(); /* Set serial before timestamp */
+	skb->tstamp = ktime_get_real();
+
 	/* send the packet by UDP
 	 * - returns -EMSGSIZE if UDP would have to fragment the packet
 	 *   to go out of the interface
@@ -413,12 +418,8 @@ done:
 	trace_rxrpc_tx_data(call, sp->hdr.seq, serial, whdr.flags,
 			    retrans, lost);
 	if (ret >= 0) {
-		ktime_t now = ktime_get_real();
-		skb->tstamp = now;
-		smp_wmb();
-		sp->hdr.serial = serial;
 		if (whdr.flags & RXRPC_REQUEST_ACK) {
-			call->peer->rtt_last_req = now;
+			call->peer->rtt_last_req = skb->tstamp;
 			trace_rxrpc_rtt_tx(call, rxrpc_rtt_tx_data, serial);
 			if (call->peer->rtt_usage > 1) {
 				unsigned long nowj = jiffies, ack_lost_at;
@@ -457,6 +458,10 @@ send_fragmentable:
 
 	down_write(&conn->params.local->defrag_sem);
 
+	sp->hdr.serial = serial;
+	smp_wmb(); /* Set serial before timestamp */
+	skb->tstamp = ktime_get_real();
+
 	switch (conn->params.local->srx.transport.family) {
 	case AF_INET:
 		opt = IP_PMTUDISC_DONT;