As I have written in my previous post there was a timezone related problem triggered by one of the unit test programs of syslog-ng. Apart from a minor issue in the testprogram itself, it turned out that there's a timezone conversion problem at the reception of messages. Syslog-ng 1.9.x has support for messages that use the ISO 6501 timestamp. As an example the current local time here right now in ISO 6501 is: 2006-04-09T22:43:24+02:00. The important part is that it includes an explicit timezone offset. This offset is processed by syslog-ng and it can convert timezones when necessary.
I spent about half a day to fix timezone conversion, I even used a pen and a sheet of paper to do some calculations. All I can say that there's an important building block missing from the POSIX time handling functions, which would have made my job as an application developer way easier: one that converts a broken down time representation to a UNIX time_t value, where the time to be converted is NOT in the local timezone, but in GMT. The other side of this conversion exists: localtime() converts a UNIX timestamp to the local timezone, and gmtime() does the same but instead of using the local timezone and daylight saving settings, it uses GMT as timezone.
The only portable function to convert a human readable timestamp to UNIX timestamp is mktime(3), which assumes that the converted timestamp is in local time. At first blick this can be easily used in place of our imaginery mktimegm() function: mktime() returns a value offseted by the local timezone, but we also know the local timezone offset, so we substract this from the return value of mktime() and we have a stamp in GMT, right? No, not right.
There are cases when mktime() changes its incoming broken down time representation when Daylight Saving kicks in: the value of "2006-03-26 02:00:00 CET" does not exist, it is equal to "2006-03-26 03:00:00 CEST" (CET is +01:00, CEST the daylight saving time is +02:00), and this happens to every value in this time interval, e.g. 2:33 CET becomes 3:33 CEST.
Remember, I have a timestamp with an explicitly specified timezone offset where the daylight saving settings of the syslog-ng process should not count, e.g. the sender sends something like 2006-03-26 02:00:00 +02:00, which is converted to 2006-03-26 03:00:00 +02:00 by the mktime() function, e.g. it is off one hour. And all this happens only in the transition hour. Good, heh?
The solution was to check this change in the time by mktime() and adjust the returned value, this seems to work reasonably well for the transition hour.
While writing this post I have found that there is a GNU extension defined, a function named timegm(3), which seems to do exactly what I have wanted. The problem that this function does not seem to be too portable. The notes in the manpage say that for achieve timegm() functionality, the application should change its own environment, set the TZ environment variable, call mktime(), and reset the environment variable. This does not look too clean I would even call that ugly. IIRC setenv() allocates memory, I would need to call this kludge for each and every incoming message.
I think this important hole in the API should be plugged, there are a lot of applications that need to work with various timezones and I have a bet that a lot of those work incorrectly in daylight saving transition hours.
I already have one example: GNU date program also allows specifying an explicit timezone offset:
bazsi@bzorp:~$ date -d "2006-03-26 01:59:59 +0100"
Sun Mar 26 01:59:59 CET 2006
bazsi@bzorp:~$ date -d "2006-03-26 02:00:00 +0100"
Sun Mar 26 04:00:00 CEST 2006
The second one should only be one second later than the first, e.g. it should be 03:00:00 CEST, and not 04:00:00 CEST. Try it with your favourite application :)