Re: [GSoC Proposal] Refactoring in order to reduce Git's global state

[Christian Couder <christian.couder@gmail.com>]

Hi,

On Tue, Mar 17, 2026 at 6:54 PM Francesco Paparatto
<francescopaparatto@gmail.com> wrote:

[...]

> So far, I have made the following contributions:
>
> * [GSoC PATCH v2] t3310: replace test -f/-d with
>   test_path_is_file/test_path_is_dir
>   Link: https://lore.kernel.org/git/20260228005939.9012-1-francescopaparatto@gmail.com/
>   Status: Graduated to 'master'.

For commits which graduated to master, please give the commit ID of
either the commits you authored or the merge commit that merged your
commit(s) into master.

> * [PATCH v4] t3310: avoid hiding failures from rev-parse in
>   command substitutions
>   Link: https://lore.kernel.org/git/20260307103631.89829-1-francescopaparatto@gmail.com/
>   Status: Will merge to 'master'.

[...]

> Context and Prior Work
> -----------------------
> Not all config variables can be treated in the same way. There is
> a fundamental distinction between eagerly and lazily parsed
> variables, and conflating the two causes regressions.
>
> Variables set in git_default_core_config() are eagerly parsed. They
> are read at startup, and if a value is invalid, Git calls die()
> immediately with a clear error before doing any real work. The user
> gets early feedback and can fix their config.
>
> Variables in struct repo_settings are lazily parsed. They are
> populated on first access via prepare_repo_settings(). If an eagerly
> parsed variable is naively moved into this struct, invalid config
> that used to crash at startup now crashes mid-operation — the user
> may have already started work that is now lost.
>
> During GSoC 2025, Ayush Chandekar moved several global configuration
> variables into repository-scoped structures [3]. Through this work
> and subsequent review discussions, the eager/lazy problem became
> visible [4].
>
> Ayush's work also surfaced the getter/setter debate. When he
> introduced getter and setter functions for repo_settings fields,
> reviewers pointed out they added no value without calling
> prepare_repo_settings() internally. From this discussion, Junio
> suggested two approaches for repo_settings variables that must
> not be mixed [5]:
>
> - Common variables: populated in prepare_repo_settings(), accessed
>   directly via repo->settings.foo. No getter, no setter.
> - Rare variables: prepare_repo_settings() does not touch the field.
>   A lazy getter checks a sentinel value (e.g. -1), reads from
>   config on first access, and caches the result.
>
> The appropriate pattern for each variable will require reasoning
> and discussion on the mailing list.
>
> Phillip Wood suggested a third approach: passing a
> repository pointer through git_default_config() via the void *cb
> callback data parameter, so handlers can populate per-repo structs
> without touching globals [6].
>
> Building on these lessons, Olamide Bello during the Outreachy
> program introduced struct repo_config_values [2], a structure
> linked to struct repository that stores eagerly parsed configuration
> values while preserving their startup-time error detection. An
> accessor function repo_config_values() enforces safety by preventing
> access from uninitialized repositories and guarding against access
> from secondary repository instances that do not yet have their
> config populated.
>
> So we now have two structs living inside struct repository:
> repo_settings for lazily parsed variables, and repo_config_values
> for eagerly parsed variables.
>
> Approach
> --------
> I will follow the pattern established in Olamide Bello's approved
> patch series [2], which provides the concrete workflow for each
> variable:
>
> 1. Add a new field to struct repo_config_values in environment.h.
> 2. Initialize the field in repo_config_values_init().
> 3. Update the config callback: get cfg via
>    repo_config_values(the_repository), write to cfg->field instead
>    of the global.
> 4. Update all call sites: replace the global with cfg->field.
> 5. Remove the global from environment.c and the extern from
>    environment.h.
> 6. Run tests and check fuzz targets.

By the way there is also this series from Olamide Bello:

https://lore.kernel.org/git/cover.1773127785.git.belkid98@gmail.com/

> Timeline
> --------
> Project size: 175 hours.
>
> Community Bonding (May 1 - May 25):
> - Discuss project direction and design approaches with mentors.
> - Study Bello Caleb's and Ayush Chandekar's patches in depth.
>   Review remaining repo_config_values work and identify
>   unfinished tasks.

It would be nice if your proposal started to look at the remaining
repo_config_values work already.

Thanks for your interest in Git and this project.

Best,
Christian.