Re: [PATCH v3 4/9] tree: handle submodule case for read_tree_at properly

From: Glen Choo <chooglen@google.com>
Date: Oct 26, 2022, 22:48 UTC
Message-ID: <kl6lk04mqk0j.fsf@chooglen-macbookpro.roam.corp.google.com>
In-Reply-To: <d3d1738e670d5dbf1378fc5c3209b2e98234a771.1665973401.git.gitgitgadget@gmail.com>

"Heather Lapointe via GitGitGadget" <gitgitgadget@gmail.com> writes:

Show 98 quoted lines

> @@ -47,40 +48,73 @@ int read_tree_at(struct repository *r,
>  			return -1;
>  		}
>  
> -		if (S_ISDIR(entry.mode))
> +		if (S_ISDIR(entry.mode)) {
>  			oidcpy(&oid, &entry.oid);
> -		else if (S_ISGITLINK(entry.mode)) {
> -			struct commit *commit;
>  
> -			commit = lookup_commit(r, &entry.oid);
> +			len = tree_entry_len(&entry);
> +			strbuf_add(base, entry.path, len);
> +			strbuf_addch(base, '/');
> +			retval = read_tree_at(r, lookup_tree(r, &oid),
> +						base, pathspec,
> +						fn, context);
> +			strbuf_setlen(base, oldlen);
> +			if (retval)
> +				return -1;
> +		} else if (pathspec->recurse_submodules && S_ISGITLINK(entry.mode)) {
> +			struct commit *commit;
> +			struct repository subrepo;
> +			struct repository* subrepo_p = &subrepo;
> +			struct tree* submodule_tree;
> +			char *submodule_rel_path;
> +			int name_base_len = 0;
> +
> +			len = tree_entry_len(&entry);
> +			strbuf_add(base, entry.path, len);
> +			submodule_rel_path = base->buf;
> +			// repo_submodule_init expects a path relative to submodule_prefix
> +			if (r->submodule_prefix) {
> +				name_base_len = strlen(r->submodule_prefix);
> +				// we should always expect to start with submodule_prefix
> +				assert(!strncmp(submodule_rel_path, r->submodule_prefix, name_base_len));
> +				// strip the prefix
> +				submodule_rel_path += name_base_len;
> +				// if submodule_prefix doesn't end with a /, we want to get rid of that too
> +				if (is_dir_sep(submodule_rel_path[0])) {
> +					submodule_rel_path++;
> +				}
> +			}
> +
> +			if (repo_submodule_init(subrepo_p, r, submodule_rel_path, null_oid()))
> +				die("couldn't init submodule %s", base->buf);
> +
> +			if (repo_read_index(subrepo_p) < 0)
> +				die("index file corrupt");
> +
> +			commit = lookup_commit(subrepo_p, &entry.oid);
>  			if (!commit)
> -				die("Commit %s in submodule path %s%s not found",
> +				die("Commit %s in submodule path %s not found",
>  				    oid_to_hex(&entry.oid),
> -				    base->buf, entry.path);
> -
> -			// FIXME: This is the wrong repo instance (it refers to the superproject)
> -			// it will always fail as is (will fix in later patch)
> -			// This current codepath isn't executed by any existing callbacks
> -			// so it wouldn't show up as an issue at this time.
> -			if (repo_parse_commit(r, commit))
> -				die("Invalid commit %s in submodule path %s%s",
> +				    base->buf);
> +
> +			if (repo_parse_commit(subrepo_p, commit))
> +				die("Invalid commit %s in submodule path %s",
>  				    oid_to_hex(&entry.oid),
> -				    base->buf, entry.path);
> +				    base->buf);
>  
> -			oidcpy(&oid, get_commit_tree_oid(commit));
> -		}
> -		else
> -			continue;
> +			submodule_tree = repo_get_commit_tree(subrepo_p, commit);
> +			oidcpy(&oid, submodule_tree ? &submodule_tree->object.oid : NULL);
>  
> -		len = tree_entry_len(&entry);
> -		strbuf_add(base, entry.path, len);
> -		strbuf_addch(base, '/');
> -		retval = read_tree_at(r, lookup_tree(r, &oid),
> -				      base, pathspec,
> -				      fn, context);
> -		strbuf_setlen(base, oldlen);
> -		if (retval)
> -			return -1;
> +			strbuf_addch(base, '/');
> +
> +			retval = read_tree_at(subrepo_p, lookup_tree(subrepo_p, &oid),
> +						base, pathspec,
> +						fn, context);
> +			if (retval)
> +			    die("failed to read tree for %s", base->buf);
> +			strbuf_setlen(base, oldlen);
> +			repo_clear(subrepo_p);
> +		}
> +		// else, this is a file (or a submodule, but no pathspec->recurse_submodules)

In this patch, we say that we can ignore a submodule when pathspec->recurse_submodules is 0, but unless I'm missing something, I don't think that's the case. The preimage is:

		else if (S_ISGITLINK(entry.mode)) {
			struct commit *commit;

			commit = lookup_commit(r, &entry.oid);
			if (!commit)
				die("Commit %s in submodule path %s%s not found",
				    oid_to_hex(&entry.oid),
				    base->buf, entry.path);

      /* ... */
			if (repo_parse_commit(r, commit))
				die("Invalid commit %s in submodule path %s%s",
				    oid_to_hex(&entry.oid),
				    base->buf, entry.path);

			oidcpy(&oid, get_commit_tree_oid(commit));
		}
		else
			continue;

		len = tree_entry_len(&entry);
		strbuf_add(base, entry.path, len);
		strbuf_addch(base, '/');
		retval = read_tree_at(r, lookup_tree(r, &oid),
				      base, pathspec,
				      fn, context);

which isn't a no-op since we actually do recurse into the gitlink. I don't know whether the subsequent call actually succeeds though (e.g. maybe it always failed and it was just a de facto no-op?), but that's much harder to prove. Since this function has callers outside of "git archive", it would be better to be conservative and keep the original behavior in the S_ISGITLINK(entry.mode) && !pathspec->recurse_submodules case.

Show 5 quoted lines

>  	}
>  	return 0;
>  }
> -- 
> gitgitgadget