instead of kind-of manual copyin/out. increases accuracy in server mode.
collecting dust in my tree for some time, result of a conversation with
somebody i really want to give credit to, but I can't find the mails now :(
okey dokey sez theo
do not have more than one dns request outstanding per peer. resolves
slow recovery when resolving fails initially, without clogging the
pipe with lots of dns requests; tested by Jason George; ok deraadt@
improves sensor timekeeping significantly.
Before this patch my test system's frequency adjustment would range
between -350 and +250, with timedelta rarely getting close to 0. After,
frequency adjustmens is on the order of +/- 0.05ppm, with time +/- a
few microseconds away from 0
ok henning, mbalmer, otto
With this patch, we give up without bothering sysctl
kern_sensors.c::sensor_find() unless we know for sure that timedelta
sensor is present.
From: "Constantine A. Murenin" <cnst+openbsd@bugmail.mojo.ru>
symbol, follow the guidelines from K&R: only one definition of a
global symbol (and possibly more declarations). Rename some vars
here and there to avoid shadowing. ok henning@
when we adjust time. This prevents ntpd from going wild when using
sensor time sources; ok henning@ (on an earlier version) and a LOT
of testing by naddy@
of the last sensor update time got broken, doesn't show up with gps since
it updates often (more often than we read), but naddy ran into it with dcf.
record time of last sensor datum seperately. ok naddy balmer
time, and make ntpd use that to send the next uery to an ntp peer and the
like. this has the advantage that changes to the clock do not interfere
with the intervals. for example, when we start on machines without an
RTC and the initial settime (-s) kicks in, intervals were strange.
idea from amandal@entrisphere.com, this implementation by me
tested ckuethe, phessler, mbalmer, ok mbalmer
them only every 30 seconds. now query them every 5,and take the median
value from 7 queries as sensor value. this takes outliers out of the
equation and makes the overall result much better, especially for
sensors with heavy jitter (like nmea for now)
-1 because of error, it may still participate in poll()
causing poll() to repeatedly wake up on error fd.
so make sure w edon't add -1 fds to pollevents to avoid unnecessary wakeups
From: amandal@entrisphere.com
takes the offset it adjtime() is already correcting for into account
when taking the offset from a sensor, we need to correct it by the
offset between system time and ntpd view.
only do frequency compensation if the clock is synced, and a slightly
diffent way of computing the linear regression.
You'll need a recent kernel and libc to use this.
Testing by naddy@ and ckuethe@ and others, thanks!
ok henning@
jmc