GPSD JSON client driver

This driver is a client driver to the GPSD daemon, which over time has become increasingly popular for UN*Xish platforms. GPSD can manage several devices in parallel, aggregate information, and acts as a data hub for client applications. GPSD can also auto-detect and handle PPS hardware signals on serial ports. Have a look at the GPSD project page.

It is important to understand that this driver works best using a GPS device with PPS support.

The GPSD protocol is text based, using JSON notation to transfer records in the form of JSON objects. The driver uses a TCP/IP connection to localhost:gpsd to connect to the daemon and then requests the GPS device /dev/gpsu to be watched. (Different clock units use different devices, and GPSD is able to give only the relevant information to a clock instance.)

This driver does not expect GPSD to be running or the clock device to be present a priori; it will try to re-establish a lost or hitherto unsuccessful connection and will wait for the device to come up in GPSD. There is an initial 10 seconds delay between a connection loss or failed attempt and the next reconnect attempt; this ensures that there is no thrashing on the network layer. If the connection fails again, an exponential back off is used with an upper limit of approximately 10 minutes.

The overall accuracy depends on the receiver used. The driver uses the error estimations (95% probability limits) provided by GPSD to set the clock precision dynamically according to these readings.

The driver needs the VERSION, TPV, PPS, WATCH and TOFF objects of the GPSD protocol. (Others are quietly ignored.) The driver can operate without the TOFF objects, which are available with the protocol version 3.10 and above. (Not to be confused with the release version of GPSD!) Running without TOFF objects has a negative impact on the jitter and offset of the serial timing information; if possible, a version of GPSD with support for TOFF objects should be used.

The acronym IBT is used here as a synonym for "in-band time" from the data channel of the receiver, no matter what objects were used to obtain it.

The GPSD driver uses the same device name as the NMEA driver, namely /dev/gpsu. There is a simple reason for that: While the NMEA driver and the GPSD driver can be active at the same time for different devices, they cannot access the same device at the same time. Having the same name helps prevent that collision. It also eases migration from using NMEA directly to using GPSD, as no new links etc. need to be created.

GPSD is normally started with the device name to access; it can also be instructed by hot-plug scripts to add or remove devices from its device pool. Luckily, the symlinks used by the NMEA driver are happily accepted and used by GPSD; this makes it possible to use the symlink names as device identification. This makes the migration from the built-in NMEA driver a bit easier.

Note: GPSD (as of version 3.10) cannot use kernel mode PPS on devices that are hot-plugged. This would require to attach the PPS line discipline to the character special file, which is not possible when running with root privileges already dropped. This is not likely to change in the future.

A few operational modes can be selected with the mode word.

This driver can create a lot of syslog messages when things go wrong, and cluttering the log files is frowned upon. So we attempt to log persistent or recurring errors only once per hour. On the other hand, when tracking a problem, the syslog flood throttle can get in the way.

Therefore, option flag3 can be used to disable the flood throttle at any time; the throttle is engaged by default. Running with the syslog flood throttle disabled for extended periods is not recommended unless the log files are closely monitored.

Units with numbers ≥128 act as secondary clock unit for the primary clock unit (u mod 128). A secondary unit processes only the PPS data from GPSD and needs the corresponding master unit to work¹. Use the noselect keyword on the primary unit if you are not interested in its data.

The way to use a secondary clock unit is to pair it with a primary in "tandem mode": see the configuration examples below.

The secondary unit employs the usual precautions before feeding clock samples:

If option flag1 is set for the secondary unit, the unit asserts the PPS flag on the clock as long as PPS data is available. This makes the unit eligible as PPS peer and should only be used if the GPS receiver can be trusted for the quality of its PPS signal². The PPS flag gets cleared if no PPS records can be acquired for some time. The unit also flushes the sample buffer at this point to avoid the use of stale PPS data.

Attention: This unit uses its own PPS fudge value, which must be set as fudge time1. Only the fudge values time1 and the flag1 option have an impact on secondary units.

If flag4 is set when the driver is polled, a clockstats record is written for the primary clock unit. (The secondary PPS unit does not provide clock stats on its own.) The first 3 fields are the normal date, time, and IP address common to all clockstats records.

¹)Data transmission and decoding is done only once by the primary unit. The decoded data is then processed independently in both clock units. This avoids double transmission over two sockets and decoding the same data twice, but the primary unit is always needed as a downside of this approach.

²)The clock driver suppresses the processing of PPS records when the TPV/TIME data indicates the receiver has no fix. It can also deal with situations where the PPS signal is not delivered to GPSD. But once it is available, it is also processed and used to create samples. If a receiver cannot be trusted for the precision of its PPS signal, it should not be used to create a possible PPS peer: These get extra clout and can effectively become the sole source of input for the control loop. You do not want to use sloppy data for that.

If your GPS has firmware made more than 1024 weeks (19 years and 36 weeks) in the past, its internal date counter may well wrap around and generate spurious timestamps. Some newer GPSes may have a longer wraparound (8192 weeks, or 157 years and 28 weeks) but this is almost never documented and not safe to bet on. GPSD cannot tell what the device’s wraparound time is, nor what any base-date assumption in its firmware might be.

This problem is fundamental and cannot be compensated for in code without relying on the accuracy of the local system clock, which is what GPSD (primarily designed for geospatial location) does. This choice creates a risk of perverse GIGO failure modes (especially at startup time). The only remedy is (a) not to use ancient GPS hardware, and (b) not to rely on GPSD to manage a time source if there is any serious risk that your NTP host might boot with a system clock time that is wildly off.

Even with these constraints there is some risk of flaky failures in a time window around GPS defined by your host system’s worst-case clock skew from UTC at boot time.