opennetworkinglab / int-host-reporter Goto Github PK

The current implementation may be buggy and may lead to reporting false-positives. We should implement more reliable mechanism.

We currently use dataInterface to define fabric network interface on the server. This solution requires the same interface name on each of the servers and that is not always the case (we should not introduce such a requirement).

The current idea to solve this is to provide “K8s cluster subnet” in the CIDR format and use it to determine data interface on each server.

TODO:

• add prefix to log messages from INT programs to help filtering logs
• log messages lacks new line

Calico uses the following name format for containers' interfaces: caliXXXX. This is different from the naming convention used by Cilium (lxcXXXX) and possibly other CNIs.

We should come up with a way to configure naming scheme of CNI, so that INT Host Reporter will be able to load BPF programs to containers' interfaces.

We currently compile INT-aware BPF program while the Docker image is built. However, the are some per-node metadata required to be a part of BPF program (e.g. Host IP, interface name, etc.). Therefore, we should compile BPF program at startup. This is also the approach that is taken by Cilium or Calico.

Currently, log level is always set as "Debug". We should have option to change that at startup time.

We currently have fixed configuration to listen only to the events on 2 CPUs, which may cause loosing some data plane reports. Once we'll have BPF programs compiled at load time, we should pass number of available CPUs as compiler flag and configure BPF_PERF_EVENT_ARRAY to listen on all CPUs.

This can reduce the eBPF program's size (the number of eBPF instructions). See: https://pchaigno.github.io/bpf/2021/10/20/ebpf-instruction-sets.html.

TODO:

• cancel all goroutines
• remove BPF maps
• remove eBPF programs attached to TC filters