Re: [PATCH 1/2] config: allow config_with_options() to handle any repo

From: Duy Nguyen <pclouds@gmail.com>
To: Matheus Tavares Bernardino <matheus.bernardino@usp.br>
Cc: Git Mailing List <git@vger.kernel.org>,
	Brandon Williams <bwilliams.eng@gmail.com>,
	Denton Liu <liu.denton@gmail.com>, Jeff King <peff@peff.net>,
	Junio C Hamano <gitster@pobox.com>,
	"brian m. carlson" <sandals@crustytoothpaste.net>
Subject: Re: [PATCH 1/2] config: allow config_with_options() to handle any repo
Date: Thu, 29 Aug 2019 16:31:34 +0700	[thread overview]
Message-ID: <CACsJy8DpTxpejkOHCYPnt3saC-h-3Ez0TthAPnPvHHThaG64bQ@mail.gmail.com> (raw)
In-Reply-To: <CAHd-oW6doh=06nxUkMLWZOwNMyUOogLRbLstD0bJVQxLauR_Aw@mail.gmail.com>

On Thu, Aug 29, 2019 at 11:24 AM Matheus Tavares Bernardino
<matheus.bernardino@usp.br> wrote:
>
> On Tue, Aug 27, 2019 at 8:46 PM Matheus Tavares Bernardino
> <matheus.bernardino@usp.br> wrote:
> >
> > Hi, Duy
> >
> > On Tue, Aug 27, 2019 at 6:26 AM Duy Nguyen <pclouds@gmail.com> wrote:
> > >
> > > On Tue, Aug 27, 2019 at 6:57 AM Matheus Tavares
> > > <matheus.bernardino@usp.br> wrote:
> > > >
> > > > Currently, config_with_options() relies on the global the_repository
> > > > when it has to configure from a blob.
> > >
> > > Not really reading the patch, but my last experience with moving
> > > config.c away from the_repo [1] shows that there are more hidden
> > > dependencies, in git_path() and particularly the git_config_clear()
> > > call in git_config_set_multivar_... Not really sure if those deps
> > > really affect your goals or not. Have a look at that branch, filtering
> > > on config.c for more info (and if you want to pick up some patches
> > > from that, you have my sign-off).
> >
> > Thanks for the advice. Indeed, I see now that do_git_config_sequence()
> > may call git_pathdup(), which relies on the_repo. For my use in patch
> > 2/2, repo_config_with_options() won't ever get to call
> > do_git_config_sequence(), so that's fine. But in other use cases it
> > may have to, so I'll need to check that.
>
> While working on this, I think I may have found a bug: The
> repo_read_config() function takes a repository R as parameter and
> calls this chain of functions:
>
> repo_read_config(struct repository *R) > config_with_options() >
> do_git_config_sequence() > git_pathdup("config.worktree")
>
> Shouldn't, however, the last call consider R instead of using
> the_repository? i.e., use repo_git_path(R, "config.worktree"),
> instead?

Yes. You just found one of the plenty traps because the_repository is
still hidden in many core functions.

> If so, how could we get R there? I mean, we could pass it through this
> chain, but the chain already passes a "struct config_options", which
> carries the "commondir" and "git_dir" fields. So it would probably be
> confusing to have them and an extra repository parameter (which also
> has "commondir" and "git_dir"), right? Any ideas on how to better
> approach this?

I would change 'struct config_options' to carry 'struct repository'
which also contains git_dir and other info inside. Though I have no
idea how big that change would be (didn't check the code). Config code
relies on plenty callbacks without "void *cb_data" so relying on
global state is the only way in some cases.
-- 
Duy