Problem:

When a FreeBSD host is sending transmitting at a high packet rate to
another host, it's packets may be dropped at the sender when the ARP
cache entry times out.  When the ARP entry has timed out, and ARP
request is sent for the receiving machine on the local network.  In
the period between the request and the response it elicits, outbound
packets are pushed into a buffer that provides storage for just one
packet.  Therefore, if N packets are pushed down for transmission
between the ARP request and it's response, N-1 of them are dropped at
the sender.

Proposed solution:

Outbound packets pass through arpresolve (netinet/if_ether.c).  This
function makes the decision about whether a packet should be forwarded
or buffered awaiting an arp response.  It does this on the basis on
whether there is a valid ARP cache entry for the receiving machine on
the local network.  Since it fetches the ARP cache entry and the ARP
cache entry has field stating when it will expire, we can trigger an
ARP request packet if we are within a finite interval of the expiry
time.  The response for this pre-expiry ARP request will update the
ARP cache entry and push back the expiration time.

The additional ARP requests are data-driven, so pre-expiration
requests are only sent for hosts that we are sending data to.  Entries
in the ARP cache that the host is not currently sending to are timed
out in the usual manner.

The patch referenced below uses a counter to send ARP requests
triggered by data packets.  The counter ensures that we send at most 1
ARP packet per arpt_down (default = 20) seconds.  Thus, if an ARP
response is lost or corrupted we have several well spaced attempts at
updating the ARP cache entry before it expires.  The number of
pre-expiry attempts is set to the number of regular ARP request
attempts (arp_maxtries, default = 5).

The counter variable used is the same counter ARP uses for sending
initial ARP requests.  It's state is already reset when ARP responses
are received so there is no problem in re-using the variable for
entries that have not yet expired.

A solution very similar to this is discussed in the last three
paragraphs of the original ARP RFC (RFC826) and by Casner in
kern/25517.

Index: if_ether.c
===================================================================
RCS file: /home/ncvs/src/sys/netinet/if_ether.c,v
retrieving revision 1.64.2.17
diff -u -r1.64.2.17 if_ether.c
--- if_ether.c	20 Feb 2002 23:34:09 -0000	1.64.2.17
+++ if_ether.c	9 Mar 2002 21:40:02 -0000
@@ -432,6 +431,21 @@
 	 */
 	if ((rt->rt_expire == 0 || rt->rt_expire > time_second) &&
 	    sdl->sdl_family == AF_LINK && sdl->sdl_alen != 0) {
+		/*
+		 * If entry has an expiry time and it is approaching,
+		 * see if we need to send an ARP request within this
+		 * arpt_down interval.
+		 */
+		if ((rt->rt_expire != 0) &&
+		    (time_second + (arp_maxtries - la->la_asked) * arpt_down >
+		     rt->rt_expire)) {
+			arprequest(ifp,
+				   &SIN(rt->rt_ifa->ifa_addr)->sin_addr,
+				   &SIN(dst)->sin_addr,
+				   IF_LLADDR(ifp));
+			la->la_asked++;
+		} 
+
 		bcopy(LLADDR(sdl), desten, sdl->sdl_alen);
 		return 1;
 	}