Re: bug & patch: exit codes from internal commands are handled incorrectly

On Dec 18, 2014, at 2:18 PM, Junio C Hamano [off-list ref] wrote:

Kenneth Lorber [off-list ref] writes:

quoted

quoted

Bug: exit codes from (at least) internal commands are handled incorrectly.
E.g. git-merge-file, docs say:
      The exit value of this program is negative on error, and the number of
      conflicts otherwise. If the merge was clean, the exit value is 0.
But only 8 bits get carried through exit, so 256 conflicts gives
exit(0), which means the merge was clean.

Wouldn't any cmd_foo() that returns negative to main() be buggy?

Yes.

Your change sweeps such problems under the rug, which is not a
healthy thing to do.

Agreed.  (See below.)

Expecting that the exit code can signal small positive integers and
other generic kinds of failures is a losing proposition.  I think it
is a better fix to update cmd_merge_file() to return 1 (when ret is
positive), 0 (when ret is zero) or 128 (when ret is negative), or
something simple like that, and update the documentation to match
that, without touching git.c::main().

The new patch uses 0, 1, and 2 to match diff(1).

Among the in-tree users, I notice git-cvsserver.perl is already
using the command incorrectly.  It does this:

           my $return = system("git", "merge-file", $file_local, $file_old, $file_new);
           $return >>= 8;

           cleanupTmpDir();

           if ( $return == 0 )
           {
               $log->info("Merged successfully");
               ...
           }
           elsif ( $return == 1 )
           {
               $log->info("Merged with conflicts");
               ...
           }
           else
           {
               $log->warn("Merge failed");
               next;
           }

which assumes $return == 1 is special "half-good", which is not
consistent with the documented interface.  It will incorrectly
say "Merge failed" when there are two or more conflicts.

And with such a "1 or 0 or -1" change, you will fix that caller as
well ;-)

I'll call that a win.

The attached patch covers the following:
- fixes all places where "make test" attempts to call exit() with a value <0 or >255
- changes git-merge-file.txt to reflect the change in exit codes
- fixes the exit codes for merge-file
- adds a warning (which should probably be changed to something clearer) if git.c:run_builtin() is going to cause an out-of-bounds exit
- fixes a typo in t7007-show.sh
- replaces return(0) with exit(0) in test-parse-options.c

The diff is cut against 15598cf41beed0d86cd2ac443e0f69c5a3b40321 (2.3.0-rc2)

Thanks,
Keni