Re: [RFC 2/2] net: bridge: add a software fast-path implementation

[Nikolay Aleksandrov <razor@blackwall.org>]

On 28/03/2022 18:15, Felix Fietkau wrote:
> 
> Hi Nik,
> 
> I'd like to follow up on our discussion regarding bridge offloading.
> I managed to come up with a user space + eBPF implementation that replaces this code and shows mostly the same performance gain as my previous kernel space implementation.
> 
> At first I tried to use generic XDP, but after getting it working, performance was pretty bad (due to headroom issues) and I was told that this is by design and nobody should use it in production.
> 
> Then I reworked the code to use tc classifier instead and it worked much better.
> 
> It's not fully ready yet, I need to add some more tests for incompatible features, but I'm getting there...
> The code is here: https://github.com/nbd168/bridger
> 
> There's one thing I haven't been able to figure out yet: What's the proper way to keep bridge fdb entries alive from user space without modifying them in any other way?
> 
> - Felix

Hi Felix,
That's very nice! Interesting work. One way it's usually done is through periodic NTF_USE (refresh),
another would be to mark them externally learned and delete them yourself (user-space aging).
It really depends on the exact semantics you'd like.

Cheers,
 Nik