Re: [RFC PATCH] repo-settings: set defaults even when not in a repo

> On Wed, Mar 23, 2022 at 03:22:13PM -0400, Derrick Stolee wrote:
> > On 3/23/2022 2:03 PM, Josh Steadmon wrote:
> > > prepare_repo_settings() initializes a `struct repository` with various
> > > default config options and settings read from a repository-local config
> > > file. In 44c7e62 (2021-12-06, repo-settings:prepare_repo_settings only
> > > in git repos), prepare_repo_settings was changed to issue a BUG() if it
> > > is called by a process whose CWD is not a Git repository. This approach
> > > was suggested in [1].
> > >
> > > This breaks fuzz-commit-graph, which attempts to parse arbitrary
> > > fuzzing-engine-provided bytes as a commit graph file.
> > > commit-graph.c:parse_commit_graph() calls prepare_repo_settings(), but
> > > since we run the fuzz tests without a valid repository, we are hitting
> > > the BUG() from 44c7e62 for every test case.
> > >
> > > Fix this by refactoring prepare_repo_settings() such that it sets
> > > default options unconditionally; if its process is in a Git repository,
> > > it will also load settings from the local config. This eliminates the
> > > need for a BUG() when not in a repository.
> >
> > I think you have the right idea and this can work.
> 
> Hmmm. To me this feels like bending over backwards in
> `prepare_repo_settings()` to accommodate one particular caller. I'm not
> necessarily opposed to it, but it does feel strange to make
> `prepare_repo_settings()` a noop here, since I would expect that any
> callers who do want to call `prepare_repo_settings()` are likely
> convinced that they are inside of a repository, and it probably should
> be a BUG() if they aren't.
> 
> I was initially thinking that Josh's alternative of introducing and
> calling a lower-level version of `prepare_commit_graph()` that doesn't
> require the use of any repository settings would make sense.
> 
> But when I started looking at implementing it, I became confused at how
> this is supposed to work at all without using a repository. We depend on
> the values parsed from:
> 
>   - commitGraph.generationVersion
>   - commitGraph.readChangedPaths
> 
> to determine which part(s) of the commit graph we do and don't read.
> 
> Looking around, I think I probably inadvertently broke this in
> ab14d0676c (commit-graph: pass a 'struct repository *' in more places,
> 2020-09-09). But prior to ab14d0676c, neither of those settings existed,
> so parsing the commit graph was a pure function of the commit graph's
> contents alone, and didn't rely on the existence of a repository.

> I assume that libFuzzer doesn't give us a convenient way to stub out a
> repository. Maybe we should have a variant of `parse_commit_graph` that
> instead takes a `struct repo_settings` filled out with the defaults?
> 
> We could use that to teach libFuzzer about additional states that the
> commit graph parser can be in, too (though probably outside the scope of
> this patch).
> 
> I tried to sketch all of this out, which seems to work. Let me know what
> you think: