Re: [RFC Patch net-next] net_sched: introduce eBPF based Qdisc

[Cong Wang <xiyou.wangcong@gmail.com>]

On Wed, Sep 1, 2021 at 3:42 AM Toke Høiland-Jørgensen <toke@redhat.com> wrote:
>
> John Fastabend <john.fastabend@gmail.com> writes:
>
> > Cong Wang wrote:
> >> On Tue, Aug 24, 2021 at 4:47 PM Martin KaFai Lau <kafai@fb.com> wrote:
> >> > Please explain more on this.  What is currently missing
> >> > to make qdisc in struct_ops possible?
> >>
> >> I think you misunderstand this point. The reason why I avoid it is
> >> _not_ anything is missing, quite oppositely, it is because it requires
> >> a lot of work to implement a Qdisc with struct_ops approach, literally
> >> all those struct Qdisc_ops (not to mention struct Qdisc_class_ops).
> >> WIth current approach, programmers only need to implement two
> >> eBPF programs (enqueue and dequeue).
> >>
> >> Thanks.
> >
> > Another idea. Rather than work with qdisc objects which creates all
> > these issues with how to work with existing interfaces, filters, etc.
> > Why not create an sk_buff map? Then this can be used from the existing
> > egress/ingress hooks independent of the actual qdisc being used.
>
> I agree. In fact, I'm working on doing just this for XDP, and I see no
> reason why the map type couldn't be reused for skbs as well. Doing it
> this way has a couple of benefits:

I do see a lot of reasons, for starters, struct skb_buff is very different
from struct xdp_buff, any specialized map can not be reused. I guess you
are using a generic one, how do you handle the refcnt at least for skb?

>
> - It leaves more flexibility to BPF: want a simple FIFO queue? just
>   implement that with a single queue map. Or do you want to build a full
>   hierarchical queueing structure? Just instantiate as many queue maps
>   as you need to achieve this. Etc.

Please give an example without a queue. ;) Queue is too simple, show us
something more useful please. How do you plan to re-implement EDT with
just queues?

>
> - The behaviour is defined entirely by BPF program behaviour, and does
>   not require setting up a qdisc hierarchy in addition to writing BPF
>   code.

I have no idea why you call this a benefit, because my goal is to replace
Qdisc's, not to replace any other things. You know there are plenty of Qdisc's
which are not implemented in Linux kernel.

>
> - It should be possible to structure the hooks in a way that allows
>   reusing queueing algorithm implementations between the qdisc and XDP
>   layers.

XDP has no skb but xdp_buff, no? And again, why only queues?

Thanks.