Re: [PATCH 3/3] archive: allow archive over HTTP(S) with proto v2

[Jonathan Nieder <jrnieder@gmail.com>]

Hi,

Josh Steadmon wrote:

> Subject: archive: allow archive over HTTP(S) with proto v2

It's interesting how little this has to touch the client.

> Signed-off-by: Josh Steadmon <steadmon@google.com>
> ---
>  builtin/archive.c  |  8 +++++++-
>  http-backend.c     | 10 +++++++++-
>  transport-helper.c |  5 +++--
>  3 files changed, 19 insertions(+), 4 deletions(-)
[....]
> --- a/builtin/archive.c
> +++ b/builtin/archive.c
> @@ -87,7 +87,13 @@ static int run_remote_archiver(int argc, const char **argv,
>  		status = packet_reader_read(&reader);
>  		if (status != PACKET_READ_FLUSH)
>  			die(_("git archive: expected a flush"));
> -	}
> +	} else if (version == protocol_v2 &&
> +		   starts_with(transport->url, "http"))

As Stefan noticed, this starts_with test seems a bit too loose.  For
example, what happens if I try an scp-style SSH URL like
http.example.com:path/to/repo, a local path like http/foo/bar, or a
custom protocol like httplikebutbetter://path/to/repo (honest
question: I haven't tried)?

> +		/*
> +		 * Commands over HTTP require two requests, so there's an
> +		 * additional server response to parse.
> +		 */
> +		discover_version(&reader);

Can this be made consistent with the non-http case?  The original
capabilities (/info/refs) response told us what protocol version the
server wants to use, which means that a hypothetical protocol v3 could
use a completely different request format for the followup commands:
so could the server omit the protocol version in the v2
/git-upload-archive response?  Alternatively, if we want to include
the protocol version again, could we do that in stateful protocols as
well?

Related question: what should happen if the two responses declare
different protocol versions?  Should we diagnose that as a protocol
error?

[...]
> --- a/http-backend.c
> +++ b/http-backend.c
> @@ -32,6 +32,7 @@ struct rpc_service {
>  static struct rpc_service rpc_service[] = {
>  	{ "upload-pack", "uploadpack", 1, 1 },
>  	{ "receive-pack", "receivepack", 0, -1 },
> +	{ "upload-archive", "uploadarchive", 1, 1 },

shell.c orders these in almost-alphabetical order (receive-pack,
upload-pack, upload-archive).  I guess they should both use actual
alphabetical order?  (If you agree, then please feel free to do that
in a separate patch.)

[...]
> @@ -637,6 +638,12 @@ static void service_rpc(struct strbuf *hdr, char *service_name)
>  	struct rpc_service *svc = select_service(hdr, service_name);
>  	struct strbuf buf = STRBUF_INIT;
>  
> +	if (!strcmp(service_name, "git-upload-archive")) {
> +		/* git-upload-archive doesn't need --stateless-rpc */

This comment doesn't seem actionable.  Can it say why?  E.g. "[...]
because an upload-archive command always involves a single
round-trip".  Or alternatively, I think it's fine to omit the comment.

> +		argv[1] = ".";
> +		argv[2] = NULL;
> +	}
[...]
> @@ -713,7 +720,8 @@ static struct service_cmd {
>  	{"GET", "/objects/pack/pack-[0-9a-f]{40}\\.idx$", get_idx_file},
>  
>  	{"POST", "/git-upload-pack$", service_rpc},
> -	{"POST", "/git-receive-pack$", service_rpc}
> +	{"POST", "/git-receive-pack$", service_rpc},
> +	{"POST", "/git-upload-archive$", service_rpc},

Same comment about services seeming to be in a randomish order.

[...]
> --- a/transport-helper.c
> +++ b/transport-helper.c
> @@ -605,7 +605,8 @@ static int process_connect_service(struct transport *transport,
>  		ret = run_connect(transport, &cmdbuf);
>  	} else if (data->stateless_connect &&
>  		   (get_protocol_version_config() == protocol_v2) &&

(not about this patch) These parens don't help --- they make it harder
for me to read, especially with the new parens to try to match them up
with.

> -		   !strcmp("git-upload-pack", name)) {
> +		   (!strcmp("git-upload-pack", name) ||
> +		    !strcmp("git-upload-archive", name))) {

A part of me wonders about the wasted cycles comparing to
"git-upload-" twice, but (1) it is tiny relative to actually serving
the request and (2) if we're lucky, the compiler (or a compiler of the
future) inlines the strcmp call and could optimize it out.

[...]
> @@ -639,7 +640,7 @@ static int connect_helper(struct transport *transport, const char *name,
>  
>  	/* Get_helper so connect is inited. */
>  	get_helper(transport);
> -	if (!data->connect)
> +	if (!data->connect && !data->stateless_connect)
>  		die(_("operation not supported by protocol"));

I don't understand this part.  Can you explain it further (possibly by
putting it in its own patch)?

Thanks for a pleasant read,
Jonathan