|
|
You need a little magic.
Last update:
NTP reference clock support maintains the fiction that the clock is actually an ordinary peer in the NTP tradition, but operating at a synthetic stratum of zero. The entire suite of algorithms used to filter the received data, select the best clocks or peers and combine them to produce a system clock correction operate just like ordinary NTP peers. In this way, defective clocks can be detected and removed from the peer population. As no packets are exchanged with a reference clock; however, the transmit, receive and packet procedures are replaced with separate code to simulate them.
It is important to understand how the NTP clock driver interface works. The driver assumes three timescales: standard time maintained by a distant laboratory such as USNO or NIST, reference time maintained by the external radio and the system time maintained by NTP. The radio synchronizes reference time and frequency to standard time via radio, satellite or modem. As the transmission means may not always be reliable, most radios continue to provide clock updates for some time after signal loss using an internal reference oscillator. In such cases the radio may or may not reveal the time since last synchronized and/or the estimated time error.
All three timescales run only in Coordinated Universal Time (UTC), 24-hour format, and are not adjusted for local timezone or standard/daylight time. The local timezone, standard/daylight indicator and year, if provided, are ignored. However, it is important to determine whether a leap second is to be inserted in the UTC timescale in the near future so NTP can insert it in the system timescale at the appropriate epoch.
The NTP clock driver synchronizes the system time and frequency to the radio via serial or parallel port, PPS signal or other means. The driver routinely checks the radio timecode string or status indicators to determine whether it is operating correctly or not. If it is, the driver decodes the radio timecode in days, hours, minutes, seconds and nanoseconds and provides these data with the NTP receive timestamp corresponding to the on-time epoch of the timecode. The driver interface computes the difference between the timecode time and NTP timestamp and saves the difference in a circular buffer for later processing. Once each poll interval, usually 64 s, the driver provides ancillary data including leap bits and last reference time to the interface. The interface processes the circular buffer using a median/trimmed mean algorithm to extract the best estimate and provides this and the ancillary data to the clock filter as with ordinary NTP peers.
The audio drivers are designed to look like a typical external radio in that the reference oscillator is derived from the audio codec oscillator and separate from the system clock oscillator. In the WWV and IRIG drivers, the codec oscillator is disciplined in frequency to the standard timescale via radio or local sources and can be assumed to have the same reliability and accuracy as an external radio. In these cases the driver continues to provide updates to the clock filter even if the WWV or IRIG signals are lost. However, the interface is provided the last reference time when the signals were received and increases the dispersion as expected with an ordinary peer.
The best way to understand how the clock drivers work is to study the ntp_refclock.c module and one of the drivers already implemented, such as refclock_wwvb.c. Routines refclock_transmit() and refclock_receive() maintain the peer variables in a state analogous to a network peer and pass received data on through the clock filters. Routines refclock_peer() and refclock_unpeer() initialize and terminate reference clock associations, should this ever be necessary. A set of utility routines is included to open serial devices, process sample data, edit input lines to extract embedded timestamps and to perform various debugging functions.
The main interface used by these routines is the refclockproc structure, which contains for most drivers the decimal equivalents of the year, day, month, hour, second and nanosecond decoded from the radio timecode. Additional information includes the receive timestamp, reference timestamp, exception reports, statistics tallies, etc. The support routines are passed a pointer to the peer structure, which is used for all peer-specific processing and contains a pointer to the refclockproc structure, which in turn contains a pointer to the unit structure, if used. For legacy purposes, a table typeunit[type][unit] contains the peer structure pointer for each configured clock type and unit. This structure should not be used for new implementations.
The reference clock interface supports auxiliary functions to support in-stream timestamping, pulse-per-second (PPS) interfacing and precision time kernel support. In most cases the drivers do not need to be aware of them, since they are detected at autoconfigure time and loaded automatically when the device is opened. These include the tty_clk STREAMS module and ppsapi PPS interface described in the Line Disciplines and Streams Modules page. The tty_clk module reduces latency errors due to the operating system and serial port code in slower systems. The ppsapi PPS interface replaces the ppsclock STREAMS module and is expected to become the IETF standard cross-platform interface for PPS signals. In either case, the PPS signal can be connected via a level converter/pulse generator described in the Pulse-per-second (PPS) Signal Interfacing page.
Radio and modem reference clocks by convention have addresses in the form 127.127.t.u, where t is the clock type and u in the range 0-3 is used to distinguish multiple instances of clocks of the same type. Most clocks require a serial or parallel port or special bus peripheral. The particular device is normally specified by adding a soft link /dev/devicedd to the particular hardware device involved, where d corresponds to the unit number.
By convention, reference clock drivers are named in the form refclock_xxxx.c, where xxxx is a unique string. Each driver is assigned a unique type number, long-form driver name, short-form driver name and device name. The existing assignments are in the Reference Clock Drivers page and its dependencies. All drivers supported by the particular hardware and operating system are automatically detected in the autoconfigure phase and conditionally compiled. They are configured when the daemon is started according to the configuration file, as described in the Configuration Options page.
The standard clock driver interface includes a set of common support routines some of which do such things as start and stop the device, open the serial port, and establish special functions such as PPS signal support. Other routines read and write data to the device and process time values. Most drivers need only a little customizing code to, for instance, transform idiosyncratic timecode formats to standard form, poll the device as necessary, and handle exception conditions. A standard interface is available for remote debugging and monitoring programs, such as ntpq and ntpdc, as well as the filegen facility, which can be used to record device status on a continuous basis.
The general organization of a typical clock driver includes a receive-interrupt routine to read a timecode from the I/O buffer and convert to internal format, generally in days, hours, minutes, seconds and fraction. Some timecode formats include provisions for leap-second warning and determine the clock hardware and software health. The interrupt routine then calls refclock_process() with these data and the timestamp captured at the on-time character of the timecode. This routine saves each sample as received in a circular buffer, which can store from a few up to 60 samples, in cases where the timecodes arrive one per second.
The refclock_transmit() routine in the interface is called by the system at intervals defined by the poll interval in the peer structure, generally 64 s. This routine in turn calls the transmit poll routine in the driver. In the intended design, the driver calls the refclock_receive() to process the offset samples that have accumulated since the last poll and produce the final offset and variance. The samples are processed by recursively discarding median outlyers until about 60 percent of samples remain, then averaging the surviving samples. When a reference clock must be explicitly polled to produce a timecode, the driver can reset the poll interval so that the poll routine is called a specified number of times at 1-s intervals.
The interface code and this documentation have been developed over some time and required not a little hard work converting old drivers, etc. Should you find success writing a driver for a new radio or modem service, please consider contributing it to the common good. Send the driver file itself and patches for the other files to Dave Mills (mills@udel.edu).
Most drivers support manual or automatic calibration for systematic offset bias using values encoded in the fudge configuration command. By convention, the time1 value defines the calibration offset in seconds. For those drivers that support statistics collection using the filegen utility and the clockstats file, the flag4 switch enables the utility. When a PPS signal is available, a special automatic calibration facility is provided. If the flag1 switch is set and the PPS signal is actively disciplining the system time, the calibration value is automatically adjusted to maintain a residual offset of zero. Should the PPS signal or the prefer peer fail, the adjustment is frozen and the remaining drivers continue to discipline the system clock with a minimum of residual error.
A new reference clock implementation needs to supply, in addition to the driver itself, several changes to existing files.
AC_MSG_CHECKING(FOO clock_description) AC_ARG_ENABLE(FOO, AC_HELP_STRING([--enable-FOO], [x clock_description]), [ntp_ok=$enableval], [ntp_ok=$ntp_eac]) if test "$ntp_ok" = "yes"; then ntp_refclock=yes AC_DEFINE(CLOCK_FOO, 1, [Foo clock?]) fi AC_MSG_RESULT($ntp_ok)
autoreconf configure