ch-run: run oci bundles

[kchilleri]

If option --oci or --o is provided/enabled and the image path is the path to an OCI bundle containing a directory named rootfs and a spec file config.json then parse the JSON file and distribute the arguments to the corresponding Charliecloud container arguments.

Example:

kchilleri@dev:/usr/local/src/charliecloud$ ls ~/examples/redis
config.json  rootfs
kchilleri@dev:/usr/local/src/charliecloud$ ls ~/examples/redis/rootfs/
bin  boot  data  dev  etc  home  lib  media  mnt  opt  proc  root  run  sbin  srv  sys  tmp  usr  var
kchilleri@dev:/usr/local/src/charliecloud$ cat ~/examples/redis/config.json 
{
        "ociVersion": "1.0.2-dev",
        "process": {
                "terminal": true,
                "user": {}, 

                "args": [
                        "sh", "redis-server", "--bind", "0.0.0.0"
                ],
                "env": [
                        "PATH=/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin",
                        "TERM=xterm"
                ],
                "cwd": "/",
                "capabilities": {
                        "bounding": [
                                "CAP_AUDIT_WRITE",
                                "CAP_KILL",
                                "CAP_NET_BIND_SERVICE"
                        ],
...
kchilleri@dev:/usr/local/src/charliecloud$ ch-run --oci ~/examples/redis -- true

closes #1754

[reidpr]

Nice work.

I’d like to also see documentation for this. In particular, as written, the PR ignores config.json. We’ll need to justify this choice. Looking briefly over the config.json spec, I do see a lot of things that correspond to ch-run arguments, as well as a lot of stuff that we can’t support, e.g. complex UID/GID mappings. We’ll need to decide what to do with all that.

There are quite a few JSON libraries for C, so we could consider those. A Python wrapper to process config.json would also be a reasonable choice.

[reidpr]

e.g. this one is BSD licensed and recommends “just drop json.c and json.h into your project”.

[reidpr]

this one similarly

[j-ogas]

FWIW, we should embrace JSON parsing sooner, rather than later. I suspect it would be very useful for things like oci-hooks for GPU injection.

[reidpr]

@j-ogas do you have an opinion on whether it lives in ch-run or a Python wrapper?

[j-ogas]

@j-ogas do you have an opinion on whether it lives in ch-run or a Python wrapper?

Python may offer the path of the least resistance. Aside from that, not really.

[reidpr]

Krishna, great work here! I think we might want to alter the approach a bit, and some suggestions in-line on that, but this is a solid start that avoids some common pitfalls.

[reidpr]

Nice progress! Discussed offline.

[kchilleri]

TO DO:

1. Add documentation
2. Perhaps more error checking

[reidpr]

Solid! Looks like the actual logic isn’t too complicated.

The major missing feature might be priority (defaults, environment variables, cmd line, now config.json). One way to approach that might be to create an arguments struct for each of those, then merge.

I think the path forward on this might be to table it for a little while until my CDI stuff is merged — there’s some duplicate code.

[reidpr]

Makefiles are generated, so shouldn’t include in commit.

[reidpr]

We’ll want to auto-detect that it’s a bundle rather than adding an option. In image_type() I think it would be a straightforward addition to the directory case.