when those timers are actually running. due to the way ntpd's logic works
this does not really make a difference, but correctness is good.
spotted by me, joerg agrees
do so when
-we tried to send at least one query (that is the change)
-we could not send ou a single one without failure (this was already in
place but catched too much)
problem independently noticed by nick and danh, ok mickey danh, testing by
many
a reply to that query.
if we get errors for all queries and the initial settime() is still due
and thus the parent process still waits (not yet daemonized!), send an
IMSG_SETTIME with offset 0.
shortens the delay dramatically when you boot without network
idea from a discussion with theo
inherit rootdelay from the delay from the last client update from the peer
that we picked last time to adjust the local clock.
in some cases we use the average offset between two peers' client updates,
then use the average delay between the two as well.
code path for NTP4 remains unchanged, we already set refid correctly there.
NTP3 and older uses an IPv4 address as refid.
use the IP of the server we last synced to if it was a IPv4 one.
sometimes we use the average offset between two, in that case just pick
one for the IP.
this scheme naturally fails when we query IPv6 servers and have to reply
to IPv4 NTP3 (or even older NTP versions) clients - refid stays at 0 then.
this is a protocol limitation, nothing we can do about it.
might have become invalid (because the peer showed up, dns request sent to
parent, peer vanishes, and then the reply comes back), so do not fatal() in
that case but just log_warnx(). provoked by brad
offset to correct the local clock, but use the median.
given a reasonable sized set of servers this makes us nearly immune against
outliers or flasetickers, without the need for a horribly complicated outliers
detection which does not yield to better results anyway.
test & ok otto
once at startup. ntpd delays daemonizing until it has done the intial
time setting (or ran into the timeout) in this mode to make sure stuff started
later in rc is not subject to time jumps.
this eleminates the need to run rdate -n beforehands.
with some input from & ok ryan and bob, march music from mickey
-kill the _pid flavors of imsg_create and imsg_compose, and just add pid as
argument to those
-use imsg_create in imsg_compose instead of duplicating code
-check for datalen overflow
asking the privileged one to do it. sends back an imsg with the
resulting addresses in a bunch of struct sockaddr_storage in the data
part.
this should fix all remaining issues with dns (non-)availability at
ntpd startup, be it due to named on localhost or something else.
tested by marco@ and Chris Paul <chris.paul@sentinare.com>
1) base the interval calculation on the offset from the last reply, not
from the last peer update.
Allows us to send more queries again faster when the local clock
diverges too much
2) every time we form a peer update (for which we need 8 replies)
check wether we have a ready peer update for all peers that are
currently trusted, and if so, calculate the total offset and call
adjtime().
that means that adjtime is no longer called in fixed intervals
but whenever we have enough data to reliably calculate the local
clock offset.
In practice, that means we call adjtime() less often, but with
probably better data.
3) invalidate peer updates after beeing used. no point in re-using them
- this resulted in calling adjtime() multiple times with the same
offset, which doesn't make sense
tested by many
a close-to-reality stratum, a real reference time, and a leap indicator
that will indicate if the local clock isn't synchronized.
This also means that until the server feels it's synchronized, it will
tell the clients it isn't. This is normal, and correct.
ok henning@
we know have both a "server" and "servers" keyword. they differ when the
hostname resolves to more than one IP, server picks one and servers expands
to all.
that means no longer stuffing a sockaddr_storage into ntp_peer but a pointer
to a linked list of ntp_addr structs.
in the "servers" case the list of n addresses returned by host() is expanded
into n ntp_peer structs and thus n individual peers.
in the "server" case the whole list is attached to ntp_peer, and whenever we
do not receive a reply in time we traverse the list one further, so that
hosts with both AAAA and A records are first tried with the AAAA one but
we gracefully fall back to the A one.
semantics with theo; hacked up on the Montreal->Frankfurt flight.
again Air Canada surprised me, that older 767 hat pretty decent seats.
* Respond to the query with a reasonable received time (which
will help clients get better accuracy).
* Consolidate the server response code in preparation for a
completely 'proper' response to the client.
tips and ok from henning@
a lot of credit for not having supplied us with enough data within an
adjtime run interval, and get a little credit each time we get a good
reply packet. if a peer is below 20%, only send a packet occasionally to
see wether it is back. send out queries much more often between 20 and 80%
to (re-)sync quickly, and above 80% usethe regular interval.
do not use peers < 60% for calculating teh local clock offset.
designed with theo at the pho, alexander ok
form the last 8 replied received from a peer, find the one with the lowest
delay. Use that as the peer's update taken into account for calculating
the local clock's offset.
Invalidate that reply and all ones received earlier than it so that they do
not get used again.
dtucker