Re: [PATCH 1/2] upload-pack: send shallow-info before wanted-refs in protocol v2
On Wed, Dec 24, 2025 at 01:35:03AM +0100, Matthew Dodd wrote:
Show 21 quoted lines
> From: Mats-Dodd <mats.dodd12@gmail.com>
>
> The protocol v2 specification (Documentation/gitprotocol-v2.adoc) defines
> the ordering of optional sections in the fetch response as:
>
> [acknowledgments delim-pkt] [shallow-info delim-pkt]
> [wanted-refs delim-pkt] [packfile-uris delim-pkt]
> packfile flush-pkt
>
> However, since the ref-in-want feature was introduced in 516e2b76bdc
> (upload-pack: implement ref-in-want, 2018-06-27), the server sends
> wanted-refs before shallow-info. This violates the specification and
> breaks the client (fetch-pack.c), which expects shallow-info first.
>
> When a client performs a shallow clone/fetch against a server with
> uploadpack.allowRefInWant=true, the client receives sections in the
> wrong order and fails with:
>
> fatal: expected 'packfile', received 'shallow-info'
>
> Fix by swapping the order of send_shallow_info() and
Nit: is there a word missing here? E.g. "Fix this by..."
Show 11 quoted lines
> diff --git a/upload-pack.c b/upload-pack.c
> index 1e87ae9559..029ca93e69 100644
> --- a/upload-pack.c
> +++ b/upload-pack.c
> @@ -1830,8 +1830,8 @@ int upload_pack_v2(struct repository *r, struct packet_reader *request)
> state = UPLOAD_DONE;
> break;
> case UPLOAD_SEND_PACK:
> - send_wanted_ref_info(&data);
> send_shallow_info(&data);
> + send_wanted_ref_info(&data);
Indeed. The accompanying code in "fetch-pack.c" expects information the other way round:
if (process_section_header(&reader, "shallow-info", 1))
receive_shallow_info(args, &reader, shallows, si);
if (process_section_header(&reader, "wanted-refs", 1))
receive_wanted_refs(&reader, sought, nr_sought);
The bug seems to exist since the inception of this feature. 516e2b76bd (upload-pack: implement ref-in-want, 2018-06-27) implements the server side in the current-broken way, and 733020517a (fetch-pack: implement ref-in-want, 2018-06-27) implements the client side in the correct way. So this combination has always been broken, and the fix looks obviously correct to me indeed.
One nit though: I don't really think it's necessary to split up this series into two patches. The new test can simply be added to this commit here.
Thanks!
Patrick
