[gpsd-dev] GPSD's assumptions about time

[Eric S. Raymond]

The following is the header comment of timebase.c.  Please criticize,
amplify and correct.  In particular, I need to know (and document) how
often the leap second is actually transmitted.

All of gpsd's assumptions about time and GPS time reporting live in this file.

This is a work in progress.  Currently GPSD requires that the host system
clock be accurate to within one second.  It would be nice to relax this
to "accurate within one GPS rollover period" for receivers reporting
GPS week+TOW, but isn't possible in general.

Date and time in GPS is represented as number of weeks from the start
of zero second of 6 January 1980, plus number of seconds into the
week.  GPS time is not leap-second corrected, though satellites also
broadcast a current leap-second correction which is updated on
six-month boundaries according to rotational bulletins issued by the
International Earth Rotation and Reference Systems Service (IERS).

The leap-second correction is only included in the satellite subframe
broadcast, roughly once ever 20 minutes.  While the satellites do
notify GPSes of upcoming leap-seconds, this notification is not
necessarily processed correctly on consumer-grade devices, and will
not be available at all when a GPS receiver has just
cold-booted. Thus, UTC time reported from NMEA devices may be slightly
inaccurate between a cold boot or leap second and the following
subframe broadcast. 

It might be best not to trust time for 20 minutes after GPSD startup
if it is more than 500ms from current system time (that is long enough
for an ephemeris to load) but this isn't actually implemented as the
divergence will normally be only one second or less.

GPS date and time are subject to a rollover problem in the 10-bit week
number counter, which will re-zero every 1024 weeks (roughly every 20
years). The last rollover (and the first since GPS went live in 1980)
was 0000 22 August 1999; the next would fall in 2019, but plans are
afoot to upgrade the satellite counters to 13 bits; this will delay
the next rollover until 2173.

For accurate time reporting, therefore, a GPS requires a supplemental
time references sufficient to identify the current rollover period,
e.g. accurate to within 512 weeks.  Many NMEA GPSes have a wired-in
assumption about the UTC time of the last rollover and will thus report
incorrect times outside the rollover period they were designed in.

These conditions leave gpsd in a serious hole.  Actually there are several
interrelated problems:

1) Every NMEA device has some assumption about base epoch (date of
last rollover) that we don't have access to.  Thus, there's no way to
check whether a rollover the device wasn't prepared for has occurred
before gpsd startup time (making the reported UTC date invalid)
without some other time source.  (Some NMEA devices may keep a
rollover count in RAM and avoid the problem; we can't tell when that's
happening, either.)

2) Many NMEA devices - in fact, all that don't report ZDA - never tell
us what century they think it is. Those that do report century are
still subject to rollover problems. We need an external time reference
for this, too.

3) Supposing we're looking at a binary protocol that returns week/tow,
we can't know which rollover period we're in without an external time
source.

4) Only one external time source, the host system clock, is reliably
available.

5) Another source *may* be available - the GPS leap second count, if we can
get the device to report it. The latter is not a given; SiRFs before
firmware rev 2.3.2 don't report it unless special subframe data reporting
is enabled, which requires 38400bps. Evermore GPSes can't be made to
report it at all. Furthermore, before the almanac load the GPS may report
a fixed (and possibly out of date) offset.

Conclusion: if the system clock isn't accurate enough that we can deduce
what rollover period we're in, we're utterly hosed. Furthermore, if it's
not accurate to within a second and only NMEA devices are reporting,
we don't even know what century it is!

Therefore, we must assume the system clock is reliable to within a second.

However, none of these caveats affect the usefulness of PPS, which
tells us top of second to theoretical 50ns accuracy (actually about 1
microsecond over RS232 and roughly one poll interval over USB) and can
be made to condition a local NTP instance that does *not* rely on the
system clock. The combination of PPS with NTP time should be reliable
regardless of what the local system clock gets up to. That is, unless
NTP clock skew goes over 1 second, but this is unlikely to ever happen
- and if it does the reasons will have nothing to do with GPS
idiosyncracies.


-- 
                <a href="http://www.catb.org/~esr/";>Eric S. Raymond</a>

A ``decay in the social contract'' is detectable; there is a growing
feeling, particularly among middle-income taxpayers, that they are not
getting back, from society and government, their money's worth for
taxes paid. The tendency is for taxpayers to try to take more control
of their finances...    -- IRS Strategic Plan, (May 1984)