Issues with git clone over HTTP/2 and closed connections

Hi,

I just tried to clone a repo from a server over HTTPS, which failed with a message like this:

  error:  (curl_result = 55, http_code = 0, sha1 = <XYZ>
  error: Unable to find <XYZ> under https://example.com/myrepo.git
  Fetching objects: 20790, done.
  Cannot obtain needed tree <XYZ>
  while processing commit <ABC>
  error: fetch failed.

Every time I retried cloning, <XYZ> and <ABC> changed, but the error message was the same.

By running "GIT_CURL_VERBOSE=1 git clone https://example.com/myrepo.git", I noticed that:

  a) HTTP/2 was being used; and
  b) just before the error the server returned a GOAWAY [1]:
     "== Info: received GOAWAY, error=0, last_stream=1999"

On the client side I'm using Debian Unstable (libcurl 8.3.0, git 2.40.1), and the server is running Debian Stable (nginx 1.22.1-9).

nginx will, by default, close HTTP/2 connections after "http2_max_requests", (default: 1000, i.e. 1999 streams, note that the error message above says last_stream=1999) and it seems that it is using GOAWAY to do so, which seems to confuse git/libcurl.

And sure enough, after running "git config --global http.version HTTP/1.1" on the client and trying again, the "git clone" was successful (I'm guessing I could/should also bump http2_max_requests on the server).

From what I understand, git should close the connection, try to open a new one and resume the clone operation before erroring out (because the GOAWAY message could mean anything).

Is this a known bug and is it something that would need to be fixed in libcurl or in git?

Cheers, David

PS. Not subscribed, please CC: me on any replies.

[1] https://www.rfc-editor.org/rfc/rfc7540#section-6.8 [2] http://nginx.org/en/docs/http/ngx_http_v2_module.html#http2_max_requests

On Sat, Sep 23, 2023 at 12:58:09PM +0000, David Härdeman wrote:

Thanks for a detailed report. Your analysis all makes sense to me.

I don't think we've heard of such a problem before with Git. I don't know enough about GOAWAY to comment on the correct behavior, but this is almost certainly a curl issue, not a Git one. All of the connection handling, reuse, etc, is happening invisibly at the curl layer.

It's probably worth poking around libcurl's issue tracker. This seems like it might be related:

  https://github.com/curl/curl/issues/11859

And one final comment: 2000 is a lot of requests for one clone. That plus the error you are seeing from Git makes me think you're using the "dumb" http protocol (i.e., your webserver is not set up to run the server side of Git's smart protocol, so it is just serving files blindly).

I don't know if using it is intentional or not. But the smart protocol is much more efficient, and in general I would expect it to have fewer corner cases (none of the major forges allow dumb-http at all).

You can find more details on setting it up in "git help http-backend".

If you do want to keep using the dumb protocol, consider running "git gc" on the server side repository. 2000 requests implies you have many loose objects, which could be served much more efficiently as a single pack.

-Peff

September 24, 2023 at 5:50 AM, "Jeff King" <peff@peff.net> wrote:

Yeah, looks very relevant. I'll keep an eye on that issue instead.

Thanks for the prompt feedback.

Yeah, thanks for the advice...I've already sorted this out so that I'm not affected, but I wanted to make sure that I posted the bug report before I forgot all the details.

Cheers, David