As of today, NTPsec has a capability you might have thought would be there all along in the NTP lineage but never was before - the ability to run fully autonomously, fed only by multiple local reference clocks for robustness, without requiring network check peers.
It’s been general folklore among NTP operators for a long time that you can’t run without remote check peers, but there was never any specificity about why. Yesterday I found out why.
It started out innocently enough. I was writing some documentation on what happens, and how to recover, in the case of device rollovers - like when your GPS’s time counter does its 1024-week thing and rolls over to zero, or you have a device that reports only 2-digit years and you hit a century boundary.
While reading code to research this, I noticed a very curious thing. When they ship timestamps to the sample-processing code, some of our reference-clock drivers pass full 4-digit years. Others pass only two digits, range of 0 to 99.
* Compute the timecode timestamp from the days, hours, minutes, * seconds and milliseconds/microseconds of the timecode. Use * clocktime() for the aggregate seconds and the msec/usec for * the fraction, when present. Note that this code relies on the * filesystem time for the years and does not use the years of * the timecode.
Say what? Is ntpd ignoring the year part of clock samples entirely?
Yes, as it turns out. What it was doing was a complicated little dance that in effect deduced the year of the sample from its arrival time stamp - that is, the value of the system clock when it came in.
But…but…you can’t use a reference clock to set the time of the system clock if processing the reference clock sample requires the system clock to be synced already! That’s circular!
In particular, consider initial sync at boot time if the battery backup on your clock has failed. The clock is zero; your machine thinks it’s 1970. A clock sample comes in. For good reasons, there’s clipping logic that throws out a sample if it’s more than 4 hours divergent from system time. Either the sample is rejected (most likely) or your clock is set to some time within 4 hours of the Unix epoch. Hilarity ensues.
This, in particular, seems to be the reason you’re always told you need remote check peers even if you have good local clocks. It’s so the sample processing eventually gets the message that it shouldn’t be tossing out your local refclock reports just because the system clock has gotten time-warped back to the days of love-beads and bell-bottoms. (Er, maybe it was the brown acid?)
Some days, you find a bug or design error and go "What? How did this persist so long?" This was like that, because the fix is obvious: if the reference clock hands you a 4-digit year, ignore the system clock and use it. It’s two lines of code. The unit test for the fix was bigger than the fix.
Small fix, big implications. Let’s say you run an oil refinery or a a military base. You need accurate time, but you want to open as few ports to the net as possible, and that definitely excludes NTP ports that have a long history as a DoS hazard.
What you really want to do is set up NTP with multiple redundant local clocks for redundancy - a couple of GPSes with different vendors and different firmware rev dates, a time radio, maybe even your own cesium clock. And no network peers at all.
You couldn’t do that reliably with legacy NTP, even if all your clock sources returned 4-digit years. Now, with NTPsec, you can do that. Our driver manuals now warn you which clock sources are unsafe. With any of the rest, go to it; autonomy is yours.