Re: [PATCH 1/2] t/test-lib.sh: Let test_must_fail fail on signals only

From: Jeff King <peff@peff.net>
Date: Jul 17, 2008, 05:18 UTC
Message-ID: <20080717051833.GA3100@sigio.intra.peff.net>
In-Reply-To: <7v4p6qwezy.fsf@gitster.siamese.dyndns.org>

On Tue, Jul 15, 2008 at 10:54:25PM -0700, Junio C Hamano wrote:

> Anything that returns error() from its cmd_xxx() routine, for example,
> would end up exiting with (-1).  Is it "such bogus" error codes, though?

I think it is bogus, because it is being implicitly truncated to an unsigned 8-bit value (at least on Linux -- I have no idea what other platforms do). So your -1 is actually 255. Portably speaking, C defines only the macros EXIT_SUCCESS and EXIT_FAILURE; in practice, I don't know what is most common.

Bad side effects of not treating your exit codes as unsigned 8-bit integers:

  - the exit values are easily confused with other things, like signal
    death. As in this case. We have modified our checking code in the
    test scripts, but there may be other, less robust code out there.

  - other exit values can be mistaken as success. Obviously 256, -256,
    512, -512, etc all produce an erroneous "success". Now we aren't
    doing this, as we are using "-1", but it just seems a bit cleaner to
    be up front about what is happening (and the 255 we end up with is
    unnecessarily confusing; some documents, like this one:

      http://tldp.org/LDP/abs/html/exitcodes.html

    claim 255 as "out of range").

So what we are doing now isn't terrible, but since it was noted (and did in fact cause a problem!), I just expected a "let's stop doing that" patch in the original series.

-Peff