Re: [PATCH RFC net-next] openvswitch: Queue upcalls to userspace in per-port round-robin order

[Stefano Brivio <sbrivio@redhat.com>]

Hi William,

On Fri, 10 Aug 2018 07:11:01 -0700
William Tu <u9012063@gmail.com> wrote:

> > int rr_select_srcport(struct dp_upcall_info *upcall)
> > {
> >         /* look up source port from upcall->skb... */
> > }
> >
> > And we could then easily extend this to use BPF with maps one day.
> >
> >  
> Hi Stefano,
> 
> If you want to experiment with BPF, Joe and I have some prototype.
> We implemented the upcall mechanism using BPF perf event helper function
> https://github.com/williamtu/ovs-ebpf/blob/master/bpf/datapath.c#L62
> 
> And there are threads polling the perf ring buffer to receive packets from
> BPF.
> https://github.com/williamtu/ovs-ebpf/blob/master/lib/perf-event.c#L232

Interesting, thanks for the pointers!

> If I follow the discussion correctly, before upcall, you need to queue
> packets based on different configurations (vport/hash/vni/5-tuple/...)
> and queue to different buckets when congestion happens.

Yes, correct.

> In this case, you
> probably needs a BPF map to enqueue/dequeue the packet.BPF queue map is
> not supported yet, but there is patch available:
> [iovisor-dev] [RFC PATCH 1/3] bpf: add bpf queue map
> 
> So how to enqueue and dequeue packets depends on user's BPF implementation.
> This allows fairness scheme to be extensible.

For the moment being we'll try to ensure that BPF can be plugged there
rather easily. I see the advantage, but I'd rather do this as a second
step.

-- 
Stefano