Slight change to the semantics of the serial number field in the

ack packet, which results in more reliable RTT computation.

(cherry picked from commit 70872e9d3cba906a28fa4cf78a080819c822f5f6)

diff --git a/src/rx/rx.c b/src/rx/rx.c
index d6b182e8fe32b4d881f7257721112724f8822cec..dba5ddd7cfcb4b3699ce6dd47861e84228f9c979 100644 (file)
--- a/src/rx/rx.c
+++ b/src/rx/rx.c
@@ -2782,8 +2782,11 @@ struct rx_packet *rxi_ReceivePacket(np, socket, host, port, tnop, newcallp)
            /* Respond immediately to ack packets requesting acknowledgement
              * (ping packets) */
            if (np->header.flags & RX_REQUEST_ACK) {
-               if (call->error) (void) rxi_SendCallAbort(call, 0, 1, 0);
-               else (void) rxi_SendAck(call, 0, 0, 0, 0, RX_ACK_PING_RESPONSE, 1);
+               if (call->error)
+                   (void) rxi_SendCallAbort(call, 0, 1, 0);
+               else
+                   (void) rxi_SendAck(call, 0, 0, np->header.serial, 0,
+                                      RX_ACK_PING_RESPONSE, 1);
            }
            np = rxi_ReceiveAckPacket(call, np, 1);
            break;
@@ -3382,6 +3385,27 @@ static void rxi_UpdatePeerReach(conn, acall)
        MUTEX_EXIT(&conn->conn_data_lock);
 }
 
+/* rxi_ComputePeerNetStats
+ *
+ * Called exclusively by rxi_ReceiveAckPacket to compute network link
+ * estimates (like RTT and throughput) based on ack packets.  Caller
+ * must ensure that the packet in question is the right one (i.e.
+ * serial number matches).
+ */
+static void
+rxi_ComputePeerNetStats(struct rx_call *call, struct rx_packet *p,
+       struct rx_ackPacket *ap, struct rx_packet *np)
+{
+    struct rx_peer *peer = call->conn->peer;
+
+    /* Use RTT if not delayed by client. */
+    if (ap->reason != RX_ACK_DELAY)
+       rxi_ComputeRoundTripTime(p, &p->timeSent, peer);
+#ifdef ADAPT_WINDOW
+    rxi_ComputeRate(peer, call, p, np, ap->reason);
+#endif
+}
+
 /* The real smarts of the whole thing.  */
 struct rx_packet *rxi_ReceiveAckPacket(call, np, istack)
     register struct rx_call *call;
@@ -3449,15 +3473,6 @@ struct rx_packet *rxi_ReceiveAckPacket(call, np, istack)
     }
 #endif
 
-    /* if a server connection has been re-created, it doesn't remember what
-       serial # it was up to.  An ack will tell us, since the serial field
-       contains the largest serial received by the other side */
-    MUTEX_ENTER(&conn->conn_data_lock);
-    if ((conn->type == RX_SERVER_CONNECTION) && (conn->serial < serial)) {
-       conn->serial = serial+1;
-    }
-    MUTEX_EXIT(&conn->conn_data_lock);
-
     /* Update the outgoing packet skew value to the latest value of
      * the peer's incoming packet skew value.  The ack packet, of
      * course, could arrive out of order, but that won't affect things
@@ -3473,22 +3488,9 @@ struct rx_packet *rxi_ReceiveAckPacket(call, np, istack)
     for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
        if (tp->header.seq >= first) break;
        call->tfirst = tp->header.seq + 1;
-       if (tp->header.serial == serial) {
-         /* Use RTT if not delayed by client. */
-         if (ap->reason != RX_ACK_DELAY)
-             rxi_ComputeRoundTripTime(tp, &tp->timeSent, peer);
-#ifdef ADAPT_WINDOW
-         rxi_ComputeRate(peer, call, tp, np, ap->reason);
-#endif
-       }
-       else if (tp->firstSerial == serial) {
-           /* Use RTT if not delayed by client. */
-           if (ap->reason != RX_ACK_DELAY)
-               rxi_ComputeRoundTripTime(tp, &tp->firstSent, peer);
-#ifdef ADAPT_WINDOW
-         rxi_ComputeRate(peer, call, tp, np, ap->reason);
-#endif
-       }
+       if (serial && (tp->header.serial == serial ||
+                      tp->firstSerial == serial))
+           rxi_ComputePeerNetStats(call, tp, ap, np);
 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
     /* XXX Hack. Because we have to release the global rx lock when sending
      * packets (osi_NetSend) we drop all acks while we're traversing the tq
@@ -3541,30 +3543,12 @@ struct rx_packet *rxi_ReceiveAckPacket(call, np, istack)
          * of this packet */
 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
 #ifdef RX_ENABLE_LOCKS
-       if (tp->header.seq >= first) {
-#endif /* RX_ENABLE_LOCKS */
-#endif /* AFS_GLOBAL_RXLOCK_KERNEL */
-       if (tp->header.serial == serial) {
-           /* Use RTT if not delayed by client. */
-           if (ap->reason != RX_ACK_DELAY)
-               rxi_ComputeRoundTripTime(tp, &tp->timeSent, peer);
-#ifdef ADAPT_WINDOW
-         rxi_ComputeRate(peer, call, tp, np, ap->reason);
-#endif
-       }
-       else if ((tp->firstSerial == serial)) {
-           /* Use RTT if not delayed by client. */
-           if (ap->reason != RX_ACK_DELAY)
-               rxi_ComputeRoundTripTime(tp, &tp->firstSent, peer);
-#ifdef ADAPT_WINDOW
-         rxi_ComputeRate(peer, call, tp, np, ap->reason);
-#endif
-       }
-#ifdef AFS_GLOBAL_RXLOCK_KERNEL
-#ifdef RX_ENABLE_LOCKS
-       }
+       if (tp->header.seq >= first)
 #endif /* RX_ENABLE_LOCKS */
 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
+           if (serial && (tp->header.serial == serial ||
+                          tp->firstSerial == serial))
+               rxi_ComputePeerNetStats(call, tp, ap, np);
 
        /* Set the acknowledge flag per packet based on the
          * information in the ack packet. An acknowlegded packet can
@@ -4573,7 +4557,7 @@ struct rx_packet *rxi_SendAck(call, optionalPacket, seq, serial, pflags, reason,
 
     /* The skew computation used to be bogus, I think it's better now. */
     /* We should start paying attention to skew.    XXX  */
-    ap->serial = htonl(call->conn->maxSerial);
+    ap->serial = htonl(serial);
     ap->maxSkew        = 0;    /* used to be peer->inPacketSkew */
 
     ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */