program: retrieve xlated Instructions from Program

[ti-mo]

Allow introspecting instructions after they have been translated by and loaded into the kernel.

Just putting this up here for discussion, I remember there being some prior discussion around this a while back.

---

There's also the following cosmetic bug:

tail_main:
	0: LoadMapPtr dst: r2 fd: 226

226 is the map's BPF ID, not its fd. Not sure if this needs to be taken into account.

[klueska]

I see the comment that the instructions read out cannot be written back in. My original use case from #444 was to do exactly this (i.e. read out, prepend to, write back).

Is there someway to retrieve the original stream of instructions to accomplish what I need from what’s been added here?

[ti-mo]

@klueska I think this implementation does what you asked for, at least it provides the API for you to build your use case on top of.

xlated instructions have BPF map IDs where map FDs used to be during load time. Walking the insns, calling NewMapFromID on all map IDs you encounter (maybe taking care you don't open the same map ID more than once) and rewriting those map IDs with the newly-opened FDs should work. There might be some other differences that I'm not aware of, but this should you most of the way there. 🙂

At least now, you can also diff your ProgramSpec insns with the loaded Program insns and see what changed. Good luck!

[klueska]

I have verified that moving from my custom code to this implementation works identically.

The diff is fairly small: https://gitlab.com/nvidia/container-toolkit/libnvidia-container/-/merge_requests/127/diffs?commit_id=d21ac33cb0459e9277f61622ef99fc8d8fa91ca5