Message-ID: <CAA93jw5BiqemeLrb4oDSrJQCrS1HEia7_y1syEfmqh-Fauah5w@mail.gmail.com>
Date:   Wed, 24 Aug 2022 16:55:23 -0700
From:   Dave Taht <dave.taht@...il.com>
To:     Vinicius Costa Gomes <vinicius.gomes@...el.com>
Cc:     Johannes Berg <johannes@...solutions.net>,
        Linux Kernel Network Developers <netdev@...r.kernel.org>,
        Toke Høiland-Jørgensen <toke@...e.dk>,
        Avi Stern <avraham.stern@...el.com>,
        linux-wireless <linux-wireless@...r.kernel.org>,
        Make-Wifi-fast <make-wifi-fast@...ts.bufferbloat.net>
Subject: Re: taprio vs. wireless/mac80211

On Wed, Aug 24, 2022 at 4:36 PM Vinicius Costa Gomes
<vinicius.gomes@...el.com> wrote:
>
> Hi Johannes,
>
> Johannes Berg <johannes@...solutions.net> writes:
>
> > Hi,
> >
> > We're exploring the use of taprio with wireless/mac80211, and in
> > mac80211 today (**) we don't have multiple queues (any more) since all
> > the queuing actually happens in FQ/Codel inside mac80211. We also set
> > IFF_NO_QUEUE, but that of course only really affects the default qdisc,
> > not the fact that you could use it or not.

It would be good if more people understood the packet aggregation
problem deeply,
and lost 8 minutes of their life to this.

https://www.youtube.com/watch?v=Rb-UnHDw02o&t=1745s

Theoretically, in wifi 7, something like single packet taprio-like
scheduling across "DU"s
might become feasible.

> > In mac80211 thus we never back-pressure against the qdiscs, which is why
> > we basically selected a single queue. Also, there's nothing else we do
> > about the queue other than immediately pull a packet from it if
> > available, so it'd basically pure overhead to have real queues there.
> >
> > In a sense, given that we cannot back-pressure against the queues, it
> > feels a bit wrong to even have multiple queues. We also don't benefit in
> > any way from splitting data structures onto multiple CPUs or something
> > since we put it all into the same FQ/Codel anyway.
> >
> >
> > Now, in order to use taprio, you're more or less assuming that you have
> > multiple (equivalent) queues, as far as I can tell.

802.11e's notion of four hardware queues could possibly be utilized
more effectively.

Or you could program those hw queues differently from the default.

> >
> > Obviously we can trivially expose multiple equivalent queues from
> > mac80211, but it feels somewhat wrong if that's just to make taprio be
> > able to do something?
> >
> > Also how many should we have, there's more code to run so in many cases
> > you probably don't want more than one, but if you want to use taprio you
> > need at least two, but who says that's good enough, you might want more
> > classes:
> >
> >         /* taprio imposes that traffic classes map 1:n to tx queues */
> >         if (qopt->num_tc > dev->num_tx_queues) {
> >                 NL_SET_ERR_MSG(extack, "Number of traffic classes is
> > greater than number of HW queues");
> >                 return -EINVAL;
> >         }
> >
> >
> > The way taprio is done almost feels like maybe it shouldn't even care
> > about the number of queues since taprio_dequeue_soft() anyway only
> > queries the sub-qdiscs? I mean, it's scheduling traffic, if you over-
> > subscribe and send more than the link can handle, you've already lost
> > anyway, no?
> >
> > (But then Avi pointed out that the sub qdiscs are initialized per HW
> > queue, so this doesn't really hold either ...)
> >
> >
> > Anyone have recommendations what we should do?
>
> I will need to sleep on this, but at first glance, it seems you are
> showing a limitation of taprio.
>
> Removing that limitation seems possible, but it would add a bit of
> complexity (but not much it seems) to the code, let me write down what I
> am thinking:
>
>  0. right now I can trust that there are more queues than traffic
>  classes, and using the netdev prio->tc->queue map, I can do the
>  scheduling almost entirely on queues. I have to remove this assumption.
>
>  1. with that assumption removed, it means that I can have more traffic
>  classes than queues, and so I have to be able to handle multiple
>  traffic classes mapped to a single queue, i.e. one child qdisc per TC
>  vs. one child per queue that we have today. Enqueueing each packet to
>  the right child qdisc is easy. Dequeueing also is also very similar to
>  what taprio does right now.
>
>  2. it would be great if I knew the context in which each ->dequeue() is
>  called, specifically which queue the ->dequeue() is for, it would
>  reduce the number of children that I would have to check.
>
> After writing this, I got the impression that it's feasible. Anyway,
> will think a bit more about it.
>
> (2) I don't think is possible right now, but I think we can go on
> without it, and leave it as a future optimization.
>
> Does it make sense? Did I understand the problem you are having right?
>
>
> Cheers,
> --
> Vinicius



-- 
FQ World Domination pending: https://blog.cerowrt.org/post/state_of_fq_codel/
Dave Täht CEO, TekLibre, LLC

Powered by blists - more mailing lists