Re: [PATCH net-next v3 2/2] net: sched: add Flow Queue PIE packet scheduler

["Toke Høiland-Jørgensen" <toke@redhat.com>]

Jakub Kicinski <kuba@kernel.org> writes:

> On Mon, 13 Jan 2020 12:44:38 +0100, Toke Høiland-Jørgensen wrote:
>> Jakub Kicinski <kuba@kernel.org> writes:
>> > On Fri, 10 Jan 2020 11:56:57 +0530, gautamramk@gmail.com wrote:  
>> >> From: "Mohit P. Tahiliani" <tahiliani@nitk.edu.in>
>> >> 
>> >> Principles:
>> >>   - Packets are classified on flows.
>> >>   - This is a Stochastic model (as we use a hash, several flows might
>> >>                                 be hashed on the same slot)
>> >>   - Each flow has a PIE managed queue.
>> >>   - Flows are linked onto two (Round Robin) lists,
>> >>     so that new flows have priority on old ones.
>> >>   - For a given flow, packets are not reordered.
>> >>   - Drops during enqueue only.
>> >>   - ECN capability is off by default.
>> >>   - ECN threshold is at 10% by default.
>> >>   - Uses timestamps to calculate queue delay by default.
>> >> 
>> >> Usage:
>> >> tc qdisc ... fq_pie [ limit PACKETS ] [ flows NUMBER ]
>> >>                     [ alpha NUMBER ] [ beta NUMBER ]
>> >>                     [ target TIME us ] [ tupdate TIME us ]
>> >>                     [ memory_limit BYTES ] [ quantum BYTES ]
>> >>                     [ ecnprob PERCENTAGE ] [ [no]ecn ]
>> >>                     [ [no]bytemode ] [ [no_]dq_rate_estimator ]
>> >> 
>> >> defaults:
>> >>   limit: 10240 packets, flows: 1024
>> >>   alpha: 1/8, beta : 5/4
>> >>   target: 15 ms, tupdate: 15 ms (in jiffies)
>> >>   memory_limit: 32 Mb, quantum: device MTU
>> >>   ecnprob: 10%, ecn: off
>> >>   bytemode: off, dq_rate_estimator: off  
>> >
>> > Some reviews below, but hopefully someone who knows more about qdiscs
>> > will still review :)  
>> 
>> I looked it over, and didn't find anything you hadn't already pointed
>> out below. It's pretty obvious that this started out as a copy of
>> sch_fq_codel. Which is good, because that's pretty solid. And bad,
>> because that means it introduces another almost-identical qdisc without
>> sharing any of the code...
>> 
>> I think it would be worthwhile to try to consolidate things at some
>> point. Either by just merging code from fq_{codel,pie}, but another
>> option would be to express fq_codel and fq_pie using the fq{,_impl}.h
>> includes. Maybe even sch_cake as well, but that may take a bit more
>> work. Not sure if we should require this before merging fq_pie, or just
>> leave it as a possible enhancement for later? WDYT?
>
> Tricky :/ No strong opinion on my side. I'm already a little weary of
> added function calls in the fast path (e.g. pie_drop_early()), but using
> some static inlines wouldn't hurt... Then again since fq_codel doesn't
> use fq{,_impl}.h it indeed seems like a bigger project to clean things
> up.

Yeah, definitely a bigger project; and I do worry about regressions.
Especially since fq{,_impl}.h relies heavily on indirect calls...

> IMHO if this qdisc works and is useful it could probably be merged as
> is. Hopefully we can get an opinion on this from Stephen or others.

OK, fine with me :)

-Toke