There is a bug in datetime.fromtimestamp(), whereby if it is called with -1, it fails with "mktime argument out of range" when it should not (see attached test program to reproduce the problem).

The bug is that the way that mktime() signals an error code is subtle and error-prone: you need to set a sentinel in the tm's wday or yday and not only check the return value of mktime, but also check if those values have been modified; it figures: -1 is a valid value in the return domain of mktime() and is not a sufficient condition for signaling an error.

Here is the relevant excerpt from the Linux man page:


       The mktime() function converts a broken-down time structure,  expressed
       as  local  time, to calendar time representation.  The function ignores
       the specified contents of the structure members tm_wday and tm_yday and
       recomputes  them  from  the  other  information in the broken-down time
       structure.  If structure members are outside their legal interval, they
       will  be normalized (so that, e.g., 40 October is changed into 9 Novem-
       ber).  Calling mktime() also sets the  external  variable  tzname  with
       information  about the current time zone.  If the specified broken-down
       time cannot be represented as calendar time (seconds since the  epoch),
       mktime() returns a value of (time_t)(-1) and does not alter the tm_wday
       and tm_yday members of the broken-down time structure.


This was found under Linux, I do not know if this bug also occurs on Windows or the Mac.

I attached a couple of files:

- timebug.py: reproduce the bug
- timebug.c: tests that mktime()'s behaviour is as wicked as expected
- timebug.patch: the fix to the datetime module.



P.S. I hit this bug in a graphics application while zooming in/out of a viewer rendering time-based data. Sheer luck.

File Added: timebug.py

File Added: timebug.c

File Added: timebug.patch

(Additional note: this bug was found and fixed with Peter Wang from Enthought.)

Reclassifying as a patch.

The patch is correct. I tried to use errno, but errno is unchanged on 
error. Here is a new patch with regression tests.

Can anyone review the last patch?

on Windows (with Visual Studio), mktime() also sets tm_wday only if 
successful.

But negative time_t are still not allowed by the Microsoft CRT, the 
tests fail.
There are workaround to this - for example python could use techniques 
similar to http://robertinventor.com/software/t64/
 
OTOH, the docs of the time module explicitly says that dates before the 
Epoch are not handled. Do you want to change this? in other words: is 
this a bug or a feature request?
http://docs.python.org/library/time.html

> But negative time_t are still not allowed by the Microsoft CRT, 
> the tests fail.
> (...)
> is this a bug or a feature request?

Linux mktime() supports any timestamp from 1901..2038. Should we limit 
the timestamp to 1970 just because of Microsoft? Test tm_wday fixes a 
bug on Linux and doesn't change the behaviour on Windows. So the 
problem is just the unit test: the test should be different on Windows 
(make sure that -1 raises an error).

My test included in mktime_fix_and_tests.patch has a problem: the 
timezone is constant and it's mine (GMT+1). I don't know how to write 
a generic test working on any time zone. I can't use 
datetime.fromtimestamp() because datetime.fromtimestamp() uses 
time.mktime() :-)

New version of my fix:
 - the test doesn't depend on _my_ local anymore: it uses localtime() 
to get the time tuple in the host local
 - ignore the test if mktime(-2) raise an OverflowError: avoid the 
test on Windows

Is it now ok for everyone?

Is the "break" intended in the test function? it seems that this will
skip the whole test. Isn't "continue" better?

@Amaury: You wrote:

 << But negative time_t are still not allowed by the Microsoft CRT, 
the tests fail.>>

So I choosed to skip mktime(-1) test if mktime(-2) fails.

I don't have Windows to test my patch nor current behaviour.