Re: [PATCH v5 1/5] hiderefs: add hide-refs hook to hide refs dynamically

From: Junio C Hamano <gitster@pobox.com>
Date: Sep 13, 2022, 17:01 UTC
Message-ID: <xmqqedwf9pjr.fsf@gitster.g>
In-Reply-To: <278bd185aec26285f8c00aca838f89e5f3877748.1662735985.git.gitgitgadget@gmail.com>

"Sun Chao via GitGitGadget" <gitgitgadget@gmail.com> writes:

Show 7 quoted lines

> From: Sun Chao <sunchao9@huawei.com>
>
> Gerrit is implemented by JGit and is known as a centralized workflow system
> which supports reference-level access control for repository. If we choose
> to work in centralized workflow like what Gerrit provided, reference-level
> access control is needed and we might add a reference filter hook
> `hide-refs` to hide the private data.

Please rewrite the above so that it does not sound like "Gerrit supports it, there are tons of users of Gerrit, we must support it, too". If this feature is meaningful for us, even if Gerrit folks were deprecating and planning to remove the support of it, we would add it. If it is not, even if Gerrit folks support it, we wouldn't.

Show 6 quoted lines

> +
> +		/*
> +		 * the prefix 'hook:' means that the matched refs will be
> +		 * checked by the hide-refs hook dynamically, we need to put
> +		 * the 'ref' string to the hook_hide_refs list
> +		 */

I am not sure if this deserves a five-line comment. We didn't need to have a comment that says "value without hook: means the matched refs will be hidden and we need to remember them in the hide_refs string_list" for over 10 years after all.

> +		if (skip_prefix(value, "hook:", &value)) {
> +			if (!strlen(value))
> +				return error(_("missing value for '%s' after hook option"), var);

I am not sure it is a good idea to special case an empty string, especially here at this point in the code flow. There would be strings that cannot be a refname prefix (e.g. "foo..bar") and such a check is better done at the place where the accumuldated list of ref patterns are actually used. If you are using prefix match, a value of an empty string here would be a very natural way to say "we pass all the refs through our hook".

By the way, how does the negated entry work with this new one? For static ones,

	[transfer] hiderefs = !refs/heads/

would hide everything other than refs/heads/ hierarchy, I suppose. Would we spell

	[transfer] hiderefs = hook:!refs/heads/

or

	[transfer] hiderefs = !hook:refs/heads/

to say "send everything outside the branches to hook"? If the former, you'd also need to special case "!" the same way as you special case an empty string (in short, I am saying that the special case only for an empty string does not make much sense).

How does this mechanism work with gitnamespaces (see "git config --help" and read on transfer.hideRerfs)?

Show 26 quoted lines

> +			hook = 1;
> +		}
> +
>  		ref = xstrdup(value);
>  		len = strlen(ref);
>  		while (len && ref[len - 1] == '/')
>  			ref[--len] = '\0';
> -		if (!hide_refs) {
> -			CALLOC_ARRAY(hide_refs, 1);
> -			hide_refs->strdup_strings = 1;
> +
> +		if (hook) {
> +			if (!hook_hide_refs) {
> +				CALLOC_ARRAY(hook_hide_refs, 1);
> +				hook_hide_refs->strdup_strings = 1;
> +			}
> +			string_list_append(hook_hide_refs, ref);
> +		} else {
> +			if (!hide_refs) {
> +				CALLOC_ARRAY(hide_refs, 1);
> +				hide_refs->strdup_strings = 1;
> +			}
> +			string_list_append(hide_refs, ref);
>  		}
> -		string_list_append(hide_refs, ref);
>  	}

That's a somewhat duplicated code.  I wonder

	/* no need for "hook" variable anymore */
	struct string_list **refs_list= &hide_refs;

	if (strip "hook:" prefix from value)
		refs_list = &hook_hide_refs;
		...
	if (!*refs_list) {
        	*refs_list = xcalloc(1, sizeof(*refs_list));
		(*refs_list)->strdup_strings = 1;
	}
	string_list_append(*refs_list, ref);
		
would be a better organization.  I dunno.

Show 7 quoted lines

> +
> +	/*
> +	 * Once hide-refs hook is invoked, Git need to do version negotiation,
> +	 * with it, version number and process name ('uploadpack' or 'receive')
> +	 * will send to it in pkt-line format, the proccess name is recorded
> +	 * by hide_refs_section
> +	 */

Grammar.

> +	if (hook && hide_refs_section.len == 0)
> +		strbuf_addstr(&hide_refs_section, section);
> +

I am not sure if this is correct at all, but because the 1/N patch has only code without documentation I cannot guess the intention.

The first conditional to parse the configuration variable name var tries to handle both generic transfer.hideRefs and direction specific {receive,uploadpack}.hideRefs and "section" at this point has "transfer", "receive" or "uploadpack", doesn't it?

As this is a git_config() callback, when we have

	[receive] hiderefs = hook:refs/foo
	[uploadpack] hiderefs = hook:refs/bar
	[transfer] hiderefs = hook:refs/baz

we would want to send refs/bar and refs/baz to the hook if we are a "uploadpack" process. But because the above code records the first section we happen to see (which is "receive"), hide_refs_section has that value. I am not sure how a code that later user that piece of information can behave any sensibly. Does it say "We are a 'uploadpack', but hide_refs_section says 'receive', so we should ignore what is in hook_hide_refs string list"?

I'll stop reading at this point for now, as it is not a good use of our time to review the implementation until we know the basic design is sound, which I do not quite see from what we saw up to this point. It might have made sense if each string list element had the ref pattern to match as its value and stored extra info, like "is this negated?", "is this hook pattern or static?", "is this transfer, receive, or uploadpack?" in its .util member, for example.

Thanks.