From: Dennis Ferguson <dennis@gw.ccie.utoronto.ca>
To: ntp@TRANTOR.UMD.EDU
Subject: Bug fix for ntpd's clock selection code
Date: 	Sat, 25 Feb 89 00:19:36 EST

There is a misplaced initialization in ntpd's clock selection
procedure which causes the worst of the five-or-fewer candidates
which make it to the final stage of the clock selection procedure
("falseticker detection") to always be selected.  Worst, in this
case, means the highest stratum candidate with the longest delay.

I think this might repair Doug Kingston's problems.  This is almost
certainly why his clients tended to form into long chains of peers
of increasing strata rather than peering with the best stratum
server.  Also, since the faulty code wasn't really doing falseticker
detection and was always selecting the clock most likely to be out
to lunch, I suspect a fair amount of peer jumping might also be
caused by this in some situations.

A patch to ntp_proto.c follows.  In case it's not clear the variable
of concern is weight, in select_clock().

Dennis

*** /tmp/,RCSt1007392	Sat Feb 25 00:04:59 1989
--- ntp_proto.c	Fri Feb 24 23:25:17 1989
***************
*** 741,747 ****
  		}
  	}
  	while (candidates > 1) {
! 		double maxdispersion = 0.0, dispersion, weight = 1.0;
  		short worst;
  #ifdef	DEBUG
  		if (debug > 3)
--- 746,752 ----
  		}
  	}
  	while (candidates > 1) {
! 		double maxdispersion = 0.0, dispersion, weight;
  		short worst;
  #ifdef	DEBUG
  		if (debug > 3)
***************
*** 751,756 ****
--- 756,762 ----
  			/* compute dispersion of candidate `i' relative to the
  			   rest of the candidates */
  			dispersion = 0.0;
+ 			weight = 1.0;
  			for (j = 0; j < candidates; j++) {
  				dtmp = sel_lst[j].peer->estoffset -
  					sel_lst[i].peer->estoffset;