So I want to be clear. This work is very much in the "fantasy" stage. I haven't ran this at scale, and there is a lot I would do before I blindly rolled something like this out. As far performance impact, it obviously would depend on how it's implemented however I think a reasonable amount of "tracer packets" being sent out every 15 seconds or so to each hop shouldn't be too disruptive to the network, or to the host machine issuing the request. In other words if you take the `ptrace(2)` concern out of the equation (this is the big one in my mind) its going to be negligible.

> And, the rootkit question, how would eBPF notice you doing this?

I suppose it would depend on what your strategy with eBPF is and where you were looking. Reminding yourself that the pidfd_getfd and pidfd_open functions are system calls, I think any modern Linux auditing system would see something like this "a mile away".

In other words, I don't see somebody using this tactic to steal FDs in production without quickly alerting most security systems. I suspect there would be ways of hiding this however... https://github.com/krisnova/boopkit...