Re: [PATCH v2 3/3] repository: move 'repository_format_worktree_config' to repo scope

From: Glen Choo <chooglen@google.com>
Date: Jun 12, 2023, 18:10 UTC
Message-ID: <kl6lttvcft59.fsf@chooglen-macbookpro.roam.corp.google.com>
In-Reply-To: <49509708-c0a1-2439-a551-cab05d944b66@github.com>

Victoria Dye <vdye@github.com> writes:

Show 6 quoted lines

> For example, in addition to what you mentioned here w.r.t. '.hash_algo',
> there are also differences in how 'repository_format_partial_clone' is
> assigned: it's deep-copied in 'check_repository_format', but shallow-copied
> (then subsequently NULL'd in the 'struct repository_format' to avoid freeing
> the pointer when the struct is disposed of) in 'discover_git_directory()' &
> 'setup_git_directory_gently()'.

Thanks for the analysis and explanation. It's quite a pain that the various sites are similar but subtly different.

Show 18 quoted lines

> If we were to settle on a single "copy repository format settings" function,
> it's not obvious what the "right" approach is. We could change
> 'check_repository_format()' to the shallow-copy-then-null like the others:
> its two callers (in 'init-db.c' and 'path.c') don't use the value of
> 'repository_format_partial_clone' in 'struct repository_format' after
> calling 'check_repository_format()'. But, if we did that, it'd introduce a
> side effect to the input 'struct repository_format', which IMO would be
> surprising behavior for a function called 'check_<something>()'. Conversely,
> unifying on a deep copy or adding a flag to toggle deep vs. shallow copy
> feels like unnecessary complexity if we don't actually need a deep copy.
>
> Beyond the smaller subtleties, there's the larger question (that you sort of
> get at with the questions around 'discover_git_directory()') as to whether
> we should more heavily refactor or consolidate these setup functions. The
> similar code implies "yes", but such a refactor feels firmly out-of-scope
> for this series. A smaller change (e.g. just moving the assignments into
> their own function) could be less of a diversion, but any benefit seems like
> it'd be outweighed by the added churn/complexity of a new function.

I don't agree that this refactor is out of scope. I think we agree that the refactor is desirable, but if we apply the same heuristics in the future, the next author to copy a member from 'repository_format' to 'repository' could do the same and we'd never end up with the refactor we wanted. I strongly feel that if we don't put in a concerted effort into such refactors along the way, we end up creating more of the churn that made our lives harder in the first place.

I sympathize with the 'out-of-scope' sentiment, though, and I find it frustrating when a simple change starts growing in scope because a reviewer suggests fixing oddities in the codebase that I didn't think were in scope. In that vein, I think the helper function can simplify the in-scope things even if we punt on the difficult-to-reason-about parts.

E.g. we could support both deep and shallow copying, like:

  /*
   * Copy members from a repository_format to repository.
   *
   * If 'src' will no longer be read after copying (e.g. it will be
   * cleared soon), pass a nonzero value so that pointer members will be
   * moved to 'dest' (NULL-ed and shallow copied) instead of being deep
   * copied.
   */
  void copy_repository_format(struct repository *dest,
                              struct repository_format *src,
                              int take_ownership);

And in discover_git_directory(), where we don't copy .hash_algo, we could leave the code as-is and put a FIXME to figure out if we should use the helper function or drop the copying entirely.

(I'm somewhat convinced that we can just do shallow copying, though. Inspecting check_repository_format() shows that it calls clear_repository_format() right afterwards, so we really don't need the deep copy there. Using shallow copying seems to work just fine here [1]. I'll ping Jonathan Tan to see if there was a good reason to deep copy.)

[1] https://github.com/chooglen/git/actions/runs/5246795137/jobs/9476098535

Show 5 quoted lines

> In any case, sorry for the long-winded response. I'd initially tried to
> implement your feedback, but every time I did I'd get stopped up on the
> things I mentioned above. So, rather than continue to put off responding to
> this thread, I tried to capture what kept stopping me from moving forward -
> hopefully it makes (at least a little bit of) sense!

Thanks for being receptive to the feedback in the first round. I really appreciate the response, and I agree that discussing this was a better way forward than being stuck.