gitssh-demo #

Absolutely the simplest SSH server that implements basics of Git SSH protocol to allow full git clone to work.

This server requires repository to be packed into a single packfile, however it doesn't just stream that file to the client, but reads each object header and contents, and streams each object to the client.

It does not read full objects into memory, so this should well with pretty large files too. Max supported packfile size is 2Gi.

It should be relatively easy to extend it support multiple packfiles, but special care must be taken about duplicate objects. Here OFS deltas add a bit more complexity, because skipping or adding objects requires OFS deltas to be rewritten accordingly.

Additional logic can be added to filter and only send the requested objects, this requires inspecting each object to ensure their dependencies are sent too. For example, commits refer to trees and blobs. Also, some objects are stored as deltas (i.e. diffs based on other objects), which also creates a dependency link.

Git docs on file formats are here.

gc25{XVqFlsmlZZnL}

Limitations #

• Only one *.{idx,pack} file is supported. Run git repack -ad to repack it
• It does not support loose objects
• Only one branch named main is supported
• Only full clones, partial git fetch would send the whole repository
• Only supports packfiles less than 2Gi
• No shallow clones etc
• No v2 protocol support
• No SSH auth. Do not expose this to the internet :)

Usage #

Clone this repository and run git repack -ad, then:

go run main.go

Then:

git clone ssh://git@localhost:2222/demo

should clone the source code of this repository.

Using default port #

Add this to your ~/.ssh/config:

Host localhost
  Port 2222

Now you should be able to clone it using shorter address:

git clone git@localhost:demo

FAQ #

Q: Can we stream the entire *.pack file to the client?

A: Yes, this should work if we only have one packfile, but would not help if we needed to extend this demo to support partial clones or multiple packfiles.

Q: Why does it need to decompress every object with zlib and then discards the result?

A: Git packfiles do not store information about the compressed size of the object in *.pack files. So decompression is necessary to find out how many bytes to send. *.rev reverse indexes indirectly contain this information and can be used as an alternative.

Q: Why are we reading the entire list of objects from the *.idx file?

A: This is done to make this demo simpler, and since we're sending all objects anyway. Real implementation would use binary search over the index table to select only the necessary objects.

License #

MIT