git urls are incorrectly transformed from relative to absolute (unable to clone git repo from server when the repo is relative to the git user's home dir)

[alexcb]

#4142 introduced a regression, which makes it impossible to access a git repo from a ssh server that is configured to store the hosted git repos out of a user's home directory.

Here's a reproduction case:

First setup a git repo on a server:

$ ssh alex@192.168.0.2
Welcome to Ubuntu 21.04 (GNU/Linux 5.11.0-1027-raspi aarch64)
...
alex@coho:~$ mkdir repo.git
alex@coho:~$ cd repo.git
alex@coho:~/repo.git$ git init --bare .
...
Initialized empty Git repository in /home/alex/repo.git/
alex@coho:~/repo.git$ exit
logout
Connection to 192.168.0.2 closed.

Then add something to it:

alex@surfperch:/tmp$ git clone alex@192.168.0.2:repo.git
Cloning into 'repo'...
warning: You appear to have cloned an empty repository.
alex@surfperch:/tmp$ cd repo/
alex@surfperch:/tmp/repo$ echo -e "FROM alpine\nRUN echo hello"
FROM alpine
RUN echo hello
alex@surfperch:/tmp/repo$ echo -e "FROM alpine\nRUN echo hello" > Dockerfile
alex@surfperch:/tmp/repo$ git add Dockerfile 
alex@surfperch:/tmp/repo$ git commit -m 'my initial commit'
[master (root-commit) b696fd5] my initial commit
 Author: Alex Couture-Beil <alex@earthly.dev>
 1 file changed, 2 insertions(+)
 create mode 100644 Dockerfile
alex@surfperch:/tmp/repo$ git push
Enumerating objects: 3, done.
Counting objects: 100% (3/3), done.
Writing objects: 100% (3/3), 773 bytes | 773.00 KiB/s, done.
Total 3 (delta 0), reused 0 (delta 0), pack-reused 0
To 192.168.0.2:repo.git
 * [new branch]      master -> master

This works under docker (which might be an older copy without #4142)

alex@surfperch:~/gh/earthly/earthly$ DOCKER_BUILDKIT=1 docker build alex@192.168.0.2:repo.git
[+] Building 13.0s (5/5) FINISHED                                                                                                                   docker:default
 => [internal] load git source alex@192.168.0.2:repo.git                                                                                                     12.7s
 => [internal] load metadata for docker.io/library/alpine:latest                                                                                              0.0s
 => CACHED [1/2] FROM docker.io/library/alpine                                                                                                                0.0s
 => [2/2] RUN echo hello                                                                                                                                      0.3s
 => exporting to image                                                                                                                                        0.0s 
 => => exporting layers                                                                                                                                       0.0s 
 => => writing image sha256:b73f8eb80a7ed8374984c53627794db5dba9716a9cb0dae66e282cdc004185fc                                                                  0.0s 
alex@surfperch:~/gh/earthly/earthly$ docker --version
Docker version 24.0.5, build ced0996

However when I use the latest version of buildkit (which includes #4142), it fails:

$ sudo -E ./buildctl build --ssh=default --frontend=dockerfile.v0 --opt source=docker/dockerfile --opt context=alex@192.168.0.2:repo.git
[+] Building 3.1s (1/1) FINISHED                                                                                                                                                                                                                     
 => ERROR [internal] load git source alex@192.168.0.2:repo.git                                                                                                                                                                                  3.0s
------                                                                                                                                                                                                                                               
 > [internal] load git source alex@192.168.0.2:repo.git:
2.991 fatal: '/repo.git' does not appear to be a git repository
2.993 fatal: Could not read from remote repository.
2.993 
2.993 Please make sure you have the correct access rights
2.993 and the repository exists.
------
error: failed to solve: failed to read dockerfile: failed to load cache key: error fetching default branch for repository ssh://xxxxx@192.168.0.2/repo.git: git error: exit status 128
stderr:
fatal: '/repo.git' does not appear to be a git repository
fatal: Could not read from remote repository.

Please make sure you have the correct access rights
and the repository exists.

This is because buildkit has incorrectly transformed alex@192.168.0.2:repo.git into ssh://alex@192.168.0.2/repo.git.

These two git URLs are not the same, the former is relative to my home director, and the later is absolute to the / filesystem.

[alexcb]

cc @jedevc

[jedevc]

So does that imply you cannot construct a URL that has a home directory set up with git, as in this issue, that has a custom port number as in #4069?

[alexcb]

you can embed a ~ in the ssh:// form, which would then allow you to use a custom port:

e.g. this works on my server (Ubuntu with OpenSSH)

git clone ssh://alex@192.168.0.2:22/~/repo.git

unfortunately you can't assume all git hosts (e.g. github), support repos with a path like that.

e.g. this does not work:

$ git clone ssh://git@github.com:22/~/moby/buildkit.git
Cloning into 'buildkit'...
fatal: remote error: 
  is not a valid repository name
Visit https://support.github.com/ for help

Ultimately if a user gives you an implicit git url, I think it needs to kept in it's implicit form.