Add go bindings or go implementation about libnvidia-container HOT 6 OPEN

The API is fairly simple and should already be Cgo friendly: https://github.com/NVIDIA/libnvidia-container/blob/master/src/nvc.h

I would prefer not to provide official bindings unless we feel that they are really required, as there are potentially many drawbacks (opinionated implementation, additional maintenance, ABI stability...)

from libnvidia-container.

How would those drawbacks be fixed with unofficial bindings or be any different from using this as a C lib?

from libnvidia-container.

Well for example, most Go runtimes already have a dedicated process in place to work around the multithreading-setns issue, so they could just reuse it instead having us hack it in the bindings.

And as I said, maintaining bindings with yet another ABI is a lot of work when (arguably) only one runtime would use it.

from libnvidia-container.

@3XX0 I can see go bindings being useful for k8s device plugins too where as @crosbymichael mentioned we can perform all static checks (potentially) at the device plugin level.
cc @jiayingz

from libnvidia-container.

Like you said the Go threading issues are a problem and I don't think we need to support that.

Some of the functionality with this lib is determining the lib paths, versions, device paths and what devices that need to go into a container. This stuff should be split out so that we can depend on it and make decisions before invoking the runtime. Smaller surface area could help more.

from libnvidia-container.

What do you mean by checks and validation exactly, are you referring to the environment variables NVIDIA_REQUIRE_*?

As for the API, the calls for GPU information and injection are already split: nvc_*_info vs nvc_*_mount.

from libnvidia-container.