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

[Jamal Hadi Salim <jhs@mojatatu.com>]

On 2021-09-03 10:44 a.m., Toke Høiland-Jørgensen wrote:
> Martin KaFai Lau <kafai@fb.com> writes:
> 
>> On Fri, Sep 03, 2021 at 12:27:52AM +0200, Toke Høiland-Jørgensen wrote:
>>>>> The question is if it's useful to provide the full struct_ops for
>>>>> qdiscs? Having it would allow a BPF program to implement that interface
>>>>> towards userspace (things like statistics, classes etc), but the
>>>>> question is if anyone is going to bother with that given the wealth of
>>>>> BPF-specific introspection tools already available?
>> Instead of bpftool can only introspect bpf qdisc and the existing tc
>> can only introspect kernel qdisc,  it will be nice to have bpf
>> qdisc work as other qdisc and showing details together with others
>> in tc.  e.g. a bpf qdisc export its data/stats with its btf-id
>> to tc and have tc print it out in a generic way?
> 
> I'm not opposed to the idea, certainly. I just wonder if people who go
> to the trouble of writing a custom qdisc in BPF will feel it's worth it
> to do the extra work to make this available via a second API. We could
> certainly encourage it, and some things are easy (drop and pkt counters,
> etc), but other things (like class stats) will depend on the semantics
> of the qdisc being implemented, so will require extra work from the BPF
> qdisc developer...

The idea of using btf to overcome the domain difference is _very_
appealing but sounds like a lot of work? Havent delved enough
into btf - but wondering if the same could be stated for filters
and actions...Note:
Aside from current existing tooling being well understood,
challenges  you will be faced with is reinventing all the
infrastructure that tc qdiscs have taken care of over the years,
example:
the proper integrations with softirqs and multiprocessor protections,
irqs, timers etc which take care of smooth triggering of
enqueue/dequeue, taking care of defering things when the target
device/hw is busy, hierarchies, etc, etc;
not saying it is the most perfect or performant but it is one of
those 'day 3' deployments i.e a lot of corner cases taken care of.
I noticed you mentioned some of those things in one of your emails.
For this reason - Cong's approach looks appealing because it
reuses said infra. Main thing that needs to have extensibility is
the de/enqueue ops as ebpf progs. Allowing enq/deq to be ebpf specific
sounds like will allow one scheme that works for both tc and XDP
(with enq/deq taking care of the buffer contextual differences).
I admit XDP is a little harder than plain tc....

cheers,
jamal