Package: adjtimex Version: 1.29-2.1 There are two (minor) problems that cause reduced precision when calculating the frequency error of the system clock by comparing the system clock with an external reference clock (NTP) and the CMOS clock (the first problem is more important):
Problem 1) Integer arithmetics truncate the value of the frequency kernel variable to full PPMs. Source code fragments from adjtimex_1.29-2.1, file adjtimex.c: 92 int freq; /* "freq" system parameter */ ... 1249 #define SHIFT (1<<16) ... 1980 hacks[i]->relative_rate = 1981 diff_ppm = 1.e6*(sys_time - cmos_time)/sys_time 1982 - 100*(hacks[i]->tick - tick_mid) - hacks[i]->freq/SHIFT; ... 2047 hacks[i]->sys_rate = 2048 diff_ppm = 1.e6*(sys_time - ref_time)/sys_time 2049 - 100*(hacks[i]->tick - tick_mid) - hacks[i]->freq/SHIFT; SHIFT (=65536) and hacks[i]->freq are integer values; therefore, hacks[i]->freq/SHIFT is computed using integer arithmetics, causing the current frequency correction in the kernel to be truncated to full PPMs when adding it to the frequency error calculated from clock comparisons. Required correction: Convert SHIFT or hacks[i]->freq to double by casting, e.g. replace hacks[i]->freq/SHIFT with ((double)hacks[i]->freq)/SHIFT. Problem 2) In the code that compares the system clock with a timeserver, the system clock is only used with a precision of seconds instead of (readily available) microseconds. Source code fragment from adjtimex_1.29-2.1, file adjtimex.c: 1459 ftime_sys = tv_sys.tv_sec; However, in the branch with label /* no absolute time reference */ the full precision is used: 1506 ftime_sys = tv_sys.tv_sec + tv_sys.tv_usec*.000001; When comparing the NTP reference clock with the system clock in intervals of 15 minutes, the error might be as large as 1/900, or just below 0.1%, so this effect is rather minor. As fetching the time offset from NTP by calling ntpdate (or even polling the RTC!) also takes some time (which, if I am not mistaken, is not taken into account currently), there are already other sources that degrade precision. However, the fix is really trivial (copy line 1506 to line 1459), so I recommend to fix this, too. -- To UNSUBSCRIBE, email to debian-bugs-dist-requ...@lists.debian.org with a subject of "unsubscribe". Trouble? Contact listmas...@lists.debian.org
