wenxin-wang/flowlabel-switching

flowlabel-base label switching for IPv6

Edge

1. labeling: lwt xmit
2. switching: like backbone
3. unlabeling: lwt xmit

Labeling

lwt xmit bpf cannot have arguments

1. lpm match for a label value
2. if not match, unset flowlabel
3. if match, change flowlabel
4. continue routing

Unlabeling

Just set flowlabel to 0

Backbone

xdp bpf

Fallback

unlabel + routing: switching fallback to this when seeing 0-label or unknown label (sets label to 0 before routing)

Build

Prerequisites

1. linux-kernel >= 4.18 (for fib_lookup): See BPF Features by Linux Kernel Version
2. bcc: for libbpf

Build & Install

cd flowlabel-switching/
make

Examples

See test/.

About ip netns exec

The default bpf filesystem, used for pinning bpf programs and maps, is mounted on /sys/fs/bpf/ by iproute2 programs. However, ip netns exec happens to unmount /sys before exec, so the called program never sees that bpf filesystem.

And it seems that bpf filesystem cannot be mounted by any command invoked with ip netns exec. I don't know the reason behind.

So before calling ip netns exec, mount a bpf filesystem somewhere other than /sys, e.g. /run/flsw/bpffs/, and ip netns exec will see that filesystem.

And by setting TC_BPF_MNT environment variable to subdirectories of the previously mounted bpf filesystem. (iproute2 tries to mount bpf filesystem whenever TC_BPF_MNT is not a bpf mount point, whether it's a subdirectory of a bpf fs or not, probably with good reason) different mount points of bpf filesystems, we can have different bpf objects with the same name for different netns.