This program gives you raw access to the kernel time variables. Anyone may print out the time variables, but only the superuser may change them.
Your computer has two clocks - the "hardware clock" that runs all the time, and the system clock that runs only while the computer is on. Normally, "hwclock --hctosys" should be run at startup to initialize the system clock. The system clock has much better precision (approximately 1 usec), but the hardware clock probably has better long-term stability. There are three basic strategies for managing these clocks.
For a machine connected to the Internet, or equipped with a precision oscillator or radio clock, the best way is to regulate the system clock with ntpd(8). The kernel will automatically update the hardware clock every eleven minutes.
In addition, hwclock(8) can be used to approximately correct for a constant drift in the hardware clock. In this case, "hwclock --adjust" is run occasionally. hwclock notes how long it has been since the last adjustment, and nudges the hardware clock forward or back by the appropriate amount. The user needs to set the time with "hwclock --set" several times over the course of a few days so hwclock can estimate the drift rate. During that time, ntpd should not be running, or else hwclock will conclude the hardware clock does not drift at all. After you have run "hwclock --set" for the last time, it's okay to start ntpd. Then, "hwclock --systohc" should be run when the machine is shut down. (To see why, suppose the machine runs for a week with ntpd, is shut down for a day, is restarted, and "hwclock --adjust" is run by a startup script. It should only correct for one day's worth of drift. However, it has no way of knowing that ntpd has been adjusting the hardware clock, so it bases its adjustment on the last time hwclock was run.)
For a standalone or intermittently connected machine, where it's not possible to run ntpd, you may use adjtimex instead to correct the system clock for systematic drift.
There are several ways to estimate the drift rate. If your computer can be connected to the net, you might run ntpd for at least several hours and run "adjtimex --print" to learn what values of tick and freq it settled on. Alternately, you could estimate values using as a reference the CMOS clock (see the --compare and --adjust switches), another host (see --host and --review), or some other source of time (see --watch and --review). You could then add a line to rc.local invoking adjtimex, or configure /etc/init.d/adjtimex or /etc/default/adjtimex, to set those parameters each time you reboot.
Options may be introduced by either - or --, and unique abbreviations may be used.
Here is a summary of the options, grouped by type. Explanations follow.
- Get/Set Kernel Time Parameters
-p --print -t --tick val -f newfreq --frequency newfreq -o val --offset val -s adjustment --singleshot adjustment -S status --status status -m val -R --reset --maxerror val -e val --esterror val -T val --timeconstant val -a[count] --adjust[=count]
- Estimate Systematic Drifts
-c[count] --compare[=count] -i tim --interval tim -l file --log file -h timeserver --host timeserver -w --watch -r[file] --review[=file] -u --utc -d --directisa -n --nointerrupt
- Informative Output
--help -v --version -V --verbose
- -p, --print
Print the current values of the kernel time variables. NOTE: The time is "raw", and may be off by up to one timer tick (10 msec). "status" gives the value of the time_status variable in the kernel. For Linux 1.0 and 1.2 kernels, the value is as follows:
0 clock is synchronized (so the kernel should periodically set the CMOS clock to match the system clock) 1 inserting a leap second at midnight 2 deleting a leap second at midnight 3 leap second in progress 4 leap second has occurred 5 clock not externally synchronized (so the kernel should leave the CMOS clock alone)
For Linux kernels 2.0 through 2.6, the value is a sum of these:
1 PLL updates enabled 2 PPS freq discipline enabled 4 PPS time discipline enabled 8 frequency-lock mode enabled 16 inserting leap second 32 deleting leap second 64 clock unsynchronized 128 holding frequency 256 PPS signal present 512 PPS signal jitter exceeded 1024 PPS signal wander exceeded 2048 PPS signal calibration error 4096 clock hardware fault
- -t val, --tick val
Set the number of microseconds that should be added to the system time for each kernel tick interrupt. For a kernel with USER_HZ=100, there are supposed to be 100 ticks per second, so val should be close to 10000. Increasing val by 1 speeds up the system clock by about 100 ppm, or 8.64 sec/day. tick must be in the range 900000/USER_HZ...1100000/USER_HZ. If val is rejected by the kernel, adjtimex will determine the acceptable range through trial and error and print it. (After completing the search, it will restore the original value.)
- -f newfreq, --frequency newfreq
Set the system clock frequency offset to newfreq. newfreq can be negative or positive, and gives a much finer adjustment than the --tick switch. When USER_HZ=100, the value is scaled such that newfreq = 65536 speeds up the system clock by about 1 ppm, or .0864 sec/day. Thus, all of these are about the same:
--tick 9995 --frequency 32768000 --tick 10000 --frequency 6553600 --tick 10001 --frequency 0 --tick 10002 --frequency -6553600 --tick 10005 --frequency -32768000
To see the acceptable range for newfreq, use --print and look at "tolerance", or try an illegal value (e.g. --tick 0).
- -s adj, --singleshot adj
Slew the system clock by adj usec. (Its rate is changed temporarily by about 1 part in 2000.)
- -o adj, --offset adj
Add a time offset of adj usec. The kernel code adjusts the time gradually by adj, notes how long it has been since the last time offset, and then adjusts the frequency offset to correct for the apparent drift. adj must be in the range -512000...512000.
- -S status, --status status
Set kernel system clock status register to value status. Look here above at the --print switch section for the meaning of status, depending on your kernel.
- -R, --reset
Reset clock status after setting a clock parameter. For early Linux kernels, using the adjtimex(2) system call to set any time parameter the kernel think the clock is synchronized with an external time source, so it sets the kernel variable time_status to TIME_OK. Thereafter, at 11 minute intervals, it will adjust the CMOS clock to match. We prevent this "eleven minute mode" by setting the clock, because that has the side effect of resetting time_status to TIME_BAD. We try not to actually change the clock setting. Kernel versions 2.0.40 and later apparently don't need this. If your kernel does require it, use this option with: -t -T -t -e -m -f -s -o -c -r.
- -m val, --maxerror val
Set maximum error (usec).
- -e val, --esterror val
Set estimated error (usec). The maximum and estimated error are not used by the kernel. They are merely made available to user processes via the adjtimex(2) system call.
- -T val, --timeconstant val
Set phase locked loop (PLL) time constant. val determines the bandwidth or "stiffness" of the PLL. The effective PLL time constant will be a multiple of (2^val). For room-temperature quartz oscillators, David Mills recommends the value 2, which corresponds to a PLL time constant of about 900 sec and a maximum update interval of about 64 sec. The maximum update interval scales directly with the time constant, so that at the maximum time constant of 6, the update interval can be as large as 1024 sec.
Values of val between zero and 2 give quick convergence; values between 2 and 6 can be used to reduce network load, but at a modest cost in accuracy.
- -c[count], --compare[=count]
Periodically compare the system clock with the CMOS clock. After the first two calls, print values for tick and frequency offset that would bring the system clock into approximate agreement with the CMOS clock. CMOS clock readings are adjusted for systematic drift using using the correction in /etc/adjtime — see hwclock(8). The interval between comparisons is 10 seconds, unless changed by the --interval switch. The optional argument is the number of comparisons. (If the argument is supplied, the "=" is required.) If the CMOS clock and the system clock differ by more than six minutes, adjtimex will try shifting the time from the CMOS clock by some multiple of one hour, up to plus or minus 13 hours in all. This should allow correct operation, including logging, if the --utc switch was used when the CMOS clock is set to local time (or vice-versa), or if summer time has started or stopped since the CMOS clock was last set.
- -a[count], --adjust[=count]
By itself, same as --compare, except the recommended values are actually installed after every third comparison. With --review, the tick and frequency are set to the least-squares estimates. (In the latter case, any count value is ignored.)
Override the sanity check that prevents changing the clock rate by more than 500 ppm.
- -i tim, --interval tim
Set the interval in seconds between clock comparisons for the --compare and --adjust options.
- -u, --utc
The CMOS clock is set to UTC (universal time) rather than local time.
- -d, --directisa
To read the CMOS clock accurately, adjtimex usually accesses the clock via the /dev/rtc device driver of the kernel, and makes use of its CMOS update-ended interrupt to detect the beginning of seconds. It will also try /dev/rtc0 (for udev), /dev/misc/rtc (for the obsolete devfs) and possibly others. When the /dev/rtc driver is absent, or when the interrupt is not available, adjtimex can sometimes automatically fallback to a direct access method. This method detects the start of seconds by polling the update-in-progress (UIP) flag of the CMOS clock. You can force this direct access to the CMOS chip with the --directisa switch.
Note that the /dev/rtc interrupt method is more accurate, less sensible to perturbations due to system load, cleaner, cheaper, and is generally better than the direct access method. It is advisable to not use the --directisa switch, unless the CMOS chip or the motherboard don't properly provide the necessary interrupt.
- -n, --nointerrupt
Force immediate use of busywait access method, without first waiting for the interrupt timeout.
- -l[file], --log[=file]
Save the current values of the system and CMOS clocks, and optionally a reference time, to file (default /var/log/clocks.log). The reference time is taken from a network timeserver (see the --host switch) or supplied by the user (see the --watch switch).
- -h timeserver, --host timeserver
Use ntpdate to query the given timeserver for the current time. This will fail if timeserver is not running a Network Time Protocol (NTP) server, or if that server is not synchronized. Implies --log.
- -w, --watch
Ask for a keypress when the user knows the time, then ask what that time was, and its approximate accuracy. Implies --log.
- -r[file], --review[=file]
Review the clock log file (default /var/log/clocks.log) and estimate, if possible, the rates of the CMOS and system clocks. Calculate least-squares rates using all suitable log entries. Suggest corrections to adjust for systematic drift. With --adjust, the frequency and tick are set to the suggested values. (The CMOS clock correction is not changed.)
- -V, --verbose
Print the program options.
- -v, --version
Print the program version.
If your system clock gained 8 seconds in 24 hours, you could set the tick to 9999, and then it would lose 0.64 seconds a day (that is, 1 tick unit = 8.64 seconds per day). To correct the rest of the error, you could set the frequency offset to (2^16)*0.64/.0864 = 485452. Thus, putting the following in rc.local would approximately correct the system clock:
adjtimex --tick 9999 --frequency 485452
adjtimex adjusts only the system clock — the one that runs while the computer is powered up. To set or regulate the CMOS clock, see hwclock(8).
Steven S. Dick <ssd at nevets.oau.org>, Jim Van Zandt <jrv at comcast.net>.
date(1L), gettimeofday(2), settimeofday(2), hwclock(8), ntpdate(8), ntpd(8), /usr/src/linux/include/linux/timex.h, /usr/src/linux/include/linux/sched.h, /usr/src/linux/kernel/time.c, /usr/src/linux/kernel/sched.c