netkit, bpf: Add bpf programmable net device

This work adds a new, minimal BPF-programmable device called "netkit" (former PoC code-name "meta") we recently presented at LSF/MM/BPF. The core idea is that BPF programs are executed within the drivers xmit routine and therefore e.g. in case of containers/Pods moving BPF processing closer to the source. One of the goals was that in case of Pod egress traffic, this allows to move BPF programs from hostns tcx ingress into the device itself, providing earlier drop or forward mechanisms, for example, if the BPF program determines that the skb must be sent out of the node, then a redirect to the physical device can take place directly without going through per-CPU backlog queue. This helps to shift processing for such traffic from softirq to process context, leading to better scheduling decisions/performance (see measurements in the slides). In this initial version, the netkit device ships as a pair, but we plan to extend this further so it can also operate in single device mode. The pair comes with a primary and a peer device. Only the primary device, typically residing in hostns, can manage BPF programs for itself and its peer. The peer device is designated for containers/Pods and cannot attach/detach BPF programs. Upon the device creation, the user can set the default policy to 'pass' or 'drop' for the case when no BPF program is attached. Additionally, the device can be operated in L3 (default) or L2 mode. The management of BPF programs is done via bpf_mprog, so that multi-attach is supported right from the beginning with similar API and dependency controls as tcx. For details on the latter see commit 053c8e1f235d ("bpf: Add generic attach/detach/query API for multi-progs"). tc BPF compatibility is provided, so that existing programs can be easily migrated. Going forward, we plan to use netkit devices in Cilium as the main device type for connecting Pods. They will be operated in L3 mode in order to simplify a Pod's neighbor management and the peer will operate in default drop mode, so that no traffic is leaving between the time when a Pod is brought up by the CNI plugin and programs attached by the agent. Additionally, the programs we attach via tcx on the physical devices are using bpf_redirect_peer() for inbound traffic into netkit device, hence the latter is also supporting the ndo_get_peer_dev callback. Similarly, we use bpf_redirect_neigh() for the way out, pushing from netkit peer to phys device directly. Also, BIG TCP is supported on netkit device. For the follow-up work in single device mode, we plan to convert Cilium's cilium_host/_net devices into a single one. An extensive test suite for checking device operations and the BPF program and link management API comes as BPF selftests in this series.

DBDaniel Borkmann committed 3 years ago