Offloading eBPF programs attached to TC (traffic control interface) or XDP (at the driver level) is possible since kernel 4.9, and keeps getting better (more complete, and easier). [Disclaimer: I work for the company working on this.]
Tools in iproute2 package are being updated too, so typically you would attach and offload programs to hardware with `tc`- or `ip`-based command lines.
You work at Netronome? I’m curious if we will ever see your chips in traditional data planes like switches.
That's stated in the discussion as one of the major points of bpfilter against netfilter's nftables.
Re:BPF comes to firewalls
by 0xdeadbeefbabe on Tuesday February 20, 2018 @11:32AM
(#16421549)
From what I can tell snabb might be able to do this, but it bypasses the kernel: https://github.com/snabbco/snabb
Re:BPF comes to firewalls
by chrissnell on Tuesday February 20, 2018 @09:05AM
(#16420351)
In this age where major security flaws are being discovered in hardware, do we really want to offload network filtering to what might be black box hardware with limited/no auditability?
We already have plenty of insecure non-auditable black boxes on our networks [1,2] and in our CPUs [3]. A dodgy offload chip seems like a minor concern until those are solved
[1] https://www.networkworld.com/article/3016992/security/junipe...
[2] https://www.theregister.co.uk/2015/09/15/compromised_cisco_r...
[3] https://en.wikipedia.org/wiki/Intel_Management_Engine#Securi...
There are various levels of offloading. There are basic operations like checksum offloading that existed for a long time and it's optional. Not sure anyone would complain about the auditability of that one. (you would see errors on the next hop if it didn't work correctly)
I interpreted OP's point as moving that kind of complexity to a black box would open it up to being compromised and then leveraged as a backdoor into your system, or as a botnet, etc.
Without hacking up everything, like it is done in consumer routers these days.
Will BPF handle that better than nftables?