Re: [PATCH v10 0/9] submodule inline diff format

From: Jacob Keller <jacob.keller@gmail.com>
Date: Aug 25, 2016, 22:46 UTC
Message-ID: <CA+P7+xqDRD4akNShqxs4D2dDUBK0E5Eyp2Y3-8xK6wEb5gCU6w@mail.gmail.com>
In-Reply-To: <xmqqr39cjwxx.fsf@gitster.mtv.corp.google.com>

On Thu, Aug 25, 2016 at 3:38 PM, Junio C Hamano <gitster@pobox.com> wrote:

Show 19 quoted lines

> Jacob Keller <jacob.keller@gmail.com> writes:
>
>> So we should support the gitlink to a repository stored at <path>
>> without stuff inside the .git/modules, and we should support submodule
>> gitlinks with a proper .gitmodules setup. I don't think we should
>> die() but we should error properly so I will introduce a _gently()
>> variant of these functions, and die properly in the regular flow.
>
> Because "git diff [--cached] [<tree-ish>]" in the top-level is
> driven by a gitlink in the index, immediately after adding a new
> submodule to the index but before describing it in .gitmodules you
> might not have a name (and you know in that case the path will
> become the name when adding it to .gitmodules).  Also a gitlink in
> the index may correspond to a submodule the user of the top-level is
> not interested in, so there may not be anything in .git/modules/
> that corresponds to it.  In these cases, I suspect that you do not
> want to die, but you can just tell the user "I do not have enough
> information to tell you a useful story yet".
>

Right. submodule_from_path() fails to find a config. I don't think die() is right here, because there is no easy way to make this into a gently() variant.... I can still do it if we think a die() is worthwhile otherwise for the other callers of do_submodule_path...

However, I think the safest thing is to just:

a) read_gitfile on <path>/.git b) if read_gitfile succeeds, use it's contents, otherwise use <path>/.git for next steps c) check if the resulting file is a git directory, we're fine.. we found a gitdir, so stop. d) otherwise, empty the buffer, then lookup submodules e) when submodules lookup succeeds.. see if we found a name. If so, use that. f) if we didn't just exit with an empty buffer.

That empty buffer *should* trigger revision error codes since it won't point to any valid path and it also triggers the regular error code in add_submodule_odb so it handles that with showing not initizlied.

This method is less work then re-implementing a _gently() variant for all of these functions.

Stefan, does this make sense and seem reasonable?

If we just die when submodule_from_path fails I think it's bad for the submodule=* formats beside short. I don't know if it causes problems for revision code or not... I think falling back to resetting the buffer as our way of indicating error is reasonable enough...

Thanks, Jake