mktime + UTC problem

[Daniel Drake <ddrake at brontes3d dot com>]

Hi,

I'm trying to diagnose a problem where hwclock is restoring an incorrect
time from my hardware clock. I have traced this down to mktime() not
behaving as described.

In this context, hwclock is dealing with a UTC time value in a tm
structure. It then tries to convert this to UTC seconds from the epoch
using setenv("TZ", "", 1); tzset(); then mktime(). However, the
seconds-from-epoch value being returned here is wrong.

I have reproduced this in a small stand-alone test program:

#include <time.h>
#include <stdlib.h>
#include <string.h>
#include <stdio.h>

int main(void)
{
	struct tm tm;

	setenv("TZ", "", 1);
	tzset();

	memset(&tm, 0, sizeof(tm));
	tm.tm_sec = 0;
	tm.tm_min = 0;
	tm.tm_hour = 18;
	tm.tm_mday = 3;
	tm.tm_mon = 10;
	tm.tm_year = 106;
	tm.tm_yday = -1;
	tm.tm_isdst = 0;

	time_t ret = mktime(&tm);
	printf("ret %ld\n", ret);
	ret = timegm(&tm);
	printf("ret %ld\n", ret);
}

The above program attempts to interpret the tm time as UTC using
mktime(), then does the same with timegm().

On a normal system, this should output:

1162576800
1162576800

On this one system, I get:

1162594800
1162576800

timegm() is doing as it should but mktime() is not. I note that the
timegm man page has a code snippet for mktime very similar to what I am
using above.

Interestingly enough if I change:

	setenv("TZ", "", 1);

to:

	setenv("TZ", "GMT0", 1);

then the test program produces the expected results.

Am I missing something, or does this look like a bug?

Thanks!
Daniel