Re: git segfaults on older Solaris releases

From: Jeff King <peff@peff.net>
Date: Apr 9, 2016, 17:39 UTC
Message-ID: <20160409173904.GA5127@sigill.intra.peff.net>
In-Reply-To: <5708A90E.1050705@jupiterrise.com>

On Sat, Apr 09, 2016 at 09:02:38AM +0200, Tom G. Christensen wrote:

> I've finished testing 2.8.1 and I found one more case where a null is being
> printed and causing a segfault. This happens even on Solaris 8 and 9.
> The failling test is t3200.63.

Oh good, this one wasn't me. :)

It's just a normal "oops, we feed NULL and nobody on glibc noticed because it silently replaced it with "(null)" case. I did find a few other oddities while fixing it, though. +cc Patrick, who worked in this area most recently.

  [1/3]: config: lower-case first word of error strings
  [2/3]: git_config_set_multivar_in_file: all non-zero returns are errors
  [3/3]: git_config_set_multivar_in_file: handle "unset" errors

I think we may want some additional improvements. While doing 1/3, I noticed that many of these error messages could stand to be marked for translation. As other people are already looking at mass-conversion, I stopped short of doing it here (and merely contented myself with throwing a conflict into their patches ;) ).

The other thing is that 2/3 notices the error return from the config-setting functions is weird. It's sometimes negative and sometimes positive. I fixed this caller, but I think it's possible for the negative values to leak into our exit codes:

  $ touch .git/config
  $ git config foo.bar baz
  error: could not lock config file .git/config: File exists
  $ echo $?
  255

I seem to recall some systems having trouble with negative error codes, so we may want to make that more consistent.

-Peff