On Wed, Mar 24, 2021 at 02:39:51PM -0400, Jeff King wrote:
> On Tue, Mar 23, 2021 at 03:35:07PM -0700, Junio C Hamano wrote:
> 
> > > So, I think this direction is good. I imagine a full patch would
> > > include also dropping the specialized helper function that is no
> > > longer needed, and possibly adding new tests for the behavior of
> > > GIT_CONFIG_COUNT?
> > 
> > Yeah, coding that is the easiest part.  Thinking through
> > ramifications of making (or not making) such a change is much
> > harder.
> > 
> > I said "assuming" number of times, because I am not so sure if the
> > subprocesses spawned from other codepaths do or do not want to see
> > the one-shot custom configuration settings.  If that assumption
> > turns out to be wrong and the processes spawned using the helper in
> > various helper functions in submodule.c are the oddball cases that
> > want to see the custom configuration, then such a change would break
> > existing users.
> 
> I think it really depends on the command being spawned. But keep in mind
> that the local_repo_env list is not limited just to callers inside of
> Git. We expose it to the user via rev-parse, so scripts can do:
> 
>   unset $(git rev-parse --local-env-vars)
>   cd /some/other/repo
> 
> I'm hesitant to change the output there, since we don't know exactly how
> it's used in the wild[1]. Changing what our internal callers do is less
> risky, though I'd generally avoid doing so unless there is a known
> benefit. And I'm sure what the benefit is; I think this came up mostly
> because you were looking at harmonizing the behavior of the two config
> systems (and I think that _is_ worth doing, but I'd probably choose the
> historical behavior for the new system).

Agreed. I wasn't aware of this helper function at all, and aligning both
config systems so they have the same behaviour there seems like the
right thing to do to me.

> I also think it really depends on the specific config the user is
> expecting to get passed. Remember we used to have a whitelist for "this
> config is OK to pass to submodules", but it was such a mess that we did
> away with it in 89044baa8b (submodule: stop sanitizing config options,
> 2016-05-04).

That also came to my mind while this thread. I can see why it would be
useful if e.g. `gc.auto=0` gets passed down to all subcommands spawned
by git. But if the user for example injects remote configuration via
config envvars, then it'd certainly be unexpected if submodules would
try to fetch from the same in-memory remote as the parent on a recursive
fetch.

> > I _think_ the one in connect.c, which runs either the ssh transport
> > (for which the processes that run on the other side in the other
> > repository won't be affected by our environment anyway) or the file
> > transport that runs another process and talks with it over a pipe is
> > probably OK if the configuration on the "client" side leaks through
> > to the "server" side, e.g.
> > 
> >     $ git -c advice.ignoredHook=false clone file:///the/repo.git/ here
> > 
> > would probably want the other end (i.e. the one that runs upload-pack
> > in /the/repo.git/ directory) to see the one-shot configuration, too.
> 
> That example is one of the reasons I prefer _not_ to pass config here.
> It only works over local-process invocations! Not over ssh://, nor
> git://, nor https://. Even though it will do what you want in this case,
> the overall behavior is more confusing.
> 
> The more-consistent (or less inconsistent, perhaps) way is:
> 
>   git clone -u 'git -c advice.ignoredHook=false upload-pack' \
>     file:///the/repo.git
> 
> which also works with ssh. It of course _doesn't_ work with other
> protocols, but I think the technique at least makes it more clear why
> that is the case (you do not get to specify arbitrary shell commands to
> https servers).
> 
> > I do not think it makes much difference to the use of local_repo_env
> > in object-file.c::for_each_alternate_ref() either way; it could be
> > used (via core.alternateRefsCommand) an arbitrary command in each
> > alternate repository, but by default it runs for-each-ref in them,
> > and I do not think of any configuration variables that would be
> > useful on "the other side".
> > 
> > And I suspect that trailers.c::apply_command() excludes these
> > environment variables just out of habit without much deep thinking.
> > It is not going in a different repository to run the command, and
> > santitizing the environment that pertains to this repository should
> > not have any meaningful effect [*].
> > 
> > So, I would not be surprised if it were a totally safe change, but I
> > am not yet sure.
> 
> My suspicion is that for most cases, nobody cares that much either way
> (which is why we have not seen people ask "hey, why is my config not
> passed down" in any context _except_ submodules).

Probably not, but it may be a good idea to document config boundaries
such that nobody is caught by surprise there.

Patrick