On Tue, Nov 22, 2016 at 11:10:25AM -0800, Junio C Hamano wrote:
> Archive & Upload-archive:
>
> "cd Documentation && git archive --remote=origin" immediately hits
> "BUG: setup_git_env called without repository" if your Git is built
> with b1ef400eec ("setup_git_env: avoid blind fall-back to ".git"",
> 2016-10-20), which will not be part of the upcoming release. And
> 'origin' will probably not be understood from the local config.
Yeah, I think "cd Documentation && git archive --remote" has always been
broken. It doesn't barf until b1ef400eec, but yes, v2.11-rc2 breaks
finding "origin" in the local config.
I agree that doing the "gently" setup will fix the config problem, but
there's more to do to make it work with b1ef400eec outside of a repo.
It looks like read_remotes_file() and read_branches_file() in remote.c
need to learn to silently return when we are not in a git repository (or
alternatively, remote_get_1() should learn not to call them in such a
case).
I doubt this is a critical case (it only kicks in if you are outside a
repo _and_ your name looks like a remote name and not a URL, _and_ it is
not defined by your config, and the result is that we die("BUG") rather
than reporting "you don't have such a remote"). But any time it is
possible to hit a die("BUG"), we should be fixing it.
> I think we can do the "gently" thing there, as we may be retrieving
> a remote archive outside a local repository. We'd need to tweak
> "output" with prefix to compensate for the case in which the command
> is run from a subdirectory, and probably we need to futz with the
> setup_prefix parameter to write_archive(), as a local caller now
> will know if we are in nongit situation.
Makes sense. I think OPT_FILENAME() would handle "output", but yeah, it
looks like write_archive() already tries to be clever about the setup.
> On the upload-archive side that serves "--remote" request, there is
> enter_repo() so we should be covered OK.
Right, that has to run in a repo, and enter_repo() should handle it.
> Mailinfo:
>
> We may want a "gently" thing there to pick up local configuration.
> i18n.commitencoding and mailinfo.scissors in local repository would
> be ignored otherwise.
Yeah, agreed.
> Verify-pack:
>
> This calls git_config() but these days farms out its operation to
> "index-pack", so we should be OK. We may even want to lose the call
> to git_config() which does not affect anything.
I'd be slightly worried that there is a core.* config option that
affects us in a subtle way (e.g., some portability flag for how we do
run_command()). I don't think there is one now, but it seems like a
potential maintenance pitfall. IOW, I think I'd generally prefer to err
on the side of having everything do a gentle setup and load at least the
default config, to avoid surprises.
I don't think it's a high priority, though.
So what do you want to do for v2.11? I think there are minor regressions
in stripspace, archive, and mailinfo with respect to reading
.git/config. The right solution in each case is to do a gentle repo
setup. We have patches for stripspace already, but not yet for the other
two. They _should_ be fairly trivial, but we are getting awfully close
to the release. Do you want to do another round of -rc3? Ship with the
minor regressions and fix them up in v2.11.1? Last-minute revert the
original config topic that introduced the regressions (I'm biased
against that as the author, but I also think it's dangerous; it's a big
enough topic that it might introduce new problems)?
-Peff