Re: settimeofday function taking 24 - 30 minutes to complete

On 12/20/06, M. Warner Losh <imp@xxxxxxxxxx> wrote:
It does run almost instantly if you use a time from 2000. The date
command also causes the lockup if you say 'date 1970010100140' as
well. Looks like there's a loop in the kernel to do division, but I
can't quite find where it is.

If you have a good test setup, maybe you can do a binary search in the
settimeofday code to find why a large leap backwards causes problems.

Found the source of the problem by following the code in
kern_settimeofday (sys/kern/kern_time.c), settime
(sys/kern/kern_time.c), and finally in resettodr
(sys/amd64/isa/clock.c).

/*
* Write system time back to RTC
*/
void
resettodr()
{
:
s = splclock();
tm = time_second;
splx(s);
:
/* Calculate local time to put in RTC */
tm -= utc_offset();

:
/* We have now the days since 01-01-1970 in tm */
writertc(RTC_WDAY, (tm + 4) % 7 + 1); /* Write back Weekday */
for (y = 1970, m = DAYSPERYEAR + LEAPYEAR(y);
tm >= m;
y++, m = DAYSPERYEAR + LEAPYEAR(y))
tm -= m;
:
}

After adding some printf's to the resettodor, I was able to get some
interesting results:

hp010# ./t1
INFO: test running as root
INFO: Saved current time
settime: delta.tv_sec = - 1166842289
resettodr: tm = 100
resettodr: tm = -21500, utc_offset = 21600 <- value after subtracting
utc_offset()
resettodr: tm= 229, y = 426495804 <- the hang occurs in the
calculation for year.
INFO: settimeofday completed successfully
settime: delta.tv_sec = 1166840819
resettodr: tm = 1166842388
resettodr: tm = 1166820788
resettodr: tm = 355, y = 2006
INFO: Reset time to original value


By making the following change, the test program no longer hangs the system:

/*
* Calculate local time to put in RTC
* Ignore UTC offset, if it would cause the time < Jan 1, 1970 00:00.
*/
if (tm >= utc_offset())
tm -= utc_offset();
else
printf("resettodr: utc_offset > tm");

We probably need to change all the locations that use utc_offset(), to
have a similar "if (x >= utc_offset()) x -= utc_offset();".

Scot
--
DISCLAIMER:
No electrons were mamed while sending this message. Only slightly bruised.
_______________________________________________
freebsd-current@xxxxxxxxxxx mailing list
http://lists.freebsd.org/mailman/listinfo/freebsd-current
To unsubscribe, send any mail to "freebsd-current-unsubscribe@xxxxxxxxxxx"