Re: [PATCH 2/3] Allow Git::get_tz_offset to properly handle DST boundary times

[Junio C Hamano <gitster@pobox.com>]

Ben Walton <bdwalton@gmail.com> writes:

> The Git::get_tz_offset is meant to provide a workalike replacement for
> the GNU strftime %z format specifier.  The algorithm used failed to
> properly handle DST boundary cases.
>
> For example, the unix time 1162105199 in CST6CDT saw set_tz_offset
> improperly return -0600 instead of -0500.
>
> TZ=CST6CDT date -d @1162105199 +"%c %z"
> Sun 29 Oct 2006 01:59:59 AM CDT -0500
>
> $ zdump -v /usr/share/zoneinfo/CST6CDT | grep 2006
> /usr/share/zoneinfo/CST6CDT  Sun Apr  2 07:59:59 2006 UTC = Sun Apr  2
> 01:59:59 2006 CST isdst=0 gmtoff=-21600
> /usr/share/zoneinfo/CST6CDT  Sun Apr  2 08:00:00 2006 UTC = Sun Apr  2
> 03:00:00 2006 CDT isdst=1 gmtoff=-18000
> /usr/share/zoneinfo/CST6CDT  Sun Oct 29 06:59:59 2006 UTC = Sun Oct 29
> 01:59:59 2006 CDT isdst=1 gmtoff=-18000
> /usr/share/zoneinfo/CST6CDT  Sun Oct 29 07:00:00 2006 UTC = Sun Oct 29
> 01:00:00 2006 CST isdst=0 gmtoff=-21600
>
> To determine how many hours/minutes away from GMT a particular time
> was, we calculated the gmtime() of the requested time value and then
> used Time::Local's timelocal() function to turn the GMT-based time
> back into a scalar value representing seconds from the epoch.  Because
> GMT has no daylight savings time, timelocal() cannot handle the
> ambiguous times that occur at DST boundaries since there are two
> possible correct results.
>
> To work around the ambiguity at these boundaries, we must take into
> account the pre and post conversion values for is_dst as provided by
> both the original time value and the value that has been run through
> timelocal().  If the is_dst field of the two times disagree then we
> must modify the value returned from timelocal() by an hour in the
> correct direction.

It seems to me that it is a very roundabout way.  It may be correct,
but it is unclear why the magic +/-3600 shift is the right solution
and I suspect even you wouldn't notice if I sent you back your patch
with a slight change to swap $gm += 3600 and $gm -= 3600 lines ;-).

For that timestamp in question, the human-readable representation of
gmtime($t) and localtime($t) look like these two strings:

	my $t = 1162105199;
	print gmtime($t), "\n";    # Sun Oct 29 06:59:59 2006
        print localtime($t), "\n"; # Sun Oct 29 01:59:59 2006

As a human, you can easily see that these two stringified timestamps
look 5 hours apart.  Think how you managed to do so.

If we convert these back to the seconds-since-epoch, as if these
broken-down times were both in a single timezone that does not have
any DST issues, you can get the offset (in seconds) by subtraction,
and that is essentially the same as the way in which your eyes saw
they are 5 hours apart, no?  In other words, why do you need to run
timelocal() at all?

	my $t = 1162105199;
        my $lct = timegm(localtime($t)); 
        # of course, timegm(gmtime($t)) == $t

	my $minutes = int(($lct - $t)/60);
        my $sign "+";
        if ($minutes < 0) {
		$sign = "-";
                $minutes = -$minutes;
	}
        my $hours = int($minutes/60);
        $minutes -= $hours * 60;
        sprintf("%s%02d%02d", $sign, $hours, $minutes);

Confused...

>
> Signed-off-by: Ben Walton <bdwalton@gmail.com>
> ---
>  perl/Git.pm |   20 ++++++++++++++++++++
>  1 file changed, 20 insertions(+)
>
> diff --git a/perl/Git.pm b/perl/Git.pm
> index 5649bcc..788b9b4 100644
> --- a/perl/Git.pm
> +++ b/perl/Git.pm
> @@ -528,7 +528,27 @@ If TIME is not supplied, the current local time is used.
>  sub get_tz_offset {
>  	# some systmes don't handle or mishandle %z, so be creative.
>  	my $t = shift || time;
> +	# timelocal() has a problem when it comes to DST ambiguity so
> +	# times that are on a DST boundary cannot be properly converted
> +	# using it.  we will possibly adjust its result depending on whehter
> +	# pre and post conversions agree on DST
>  	my $gm = timelocal(gmtime($t));
> +
> +	# we need to know whether we were originally in DST or not
> +	my $orig_dst = (localtime($t))[8];
> +	# and also whether timelocal thinks we're in DST
> +	my $conv_dst = (localtime($gm))[8];
> +
> +	# re-adjust $gm based on the DST value for the two times we're
> +	# handling.
> +	if ($orig_dst != $conv_dst) {
> +		if ($orig_dst == 1) {
> +			$gm -= 3600;
> +		} else {
> +			$gm += 3600;
> +		}
> +	}
> +
>  	my $sign = qw( + + - )[ $t <=> $gm ];
>  	return sprintf("%s%02d%02d", $sign, (gmtime(abs($t - $gm)))[2,1]);
>  }