Received-SPF: pass (google.com: best guess record for domain of linux-wireless-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
From:   Toke =?utf-8?Q?H=C3=B8iland-J=C3=B8rgensen?= <toke@redhat.com>
To:     Kan Yan <kyan@google.com>, Yibo Zhao <yiboz@codeaurora.org>
Cc:     Felix Fietkau <nbd@nbd.name>,
        Johannes Berg <johannes@sipsolutions.net>,
        linux-wireless@vger.kernel.org,
        make-wifi-fast@lists.bufferbloat.net,
        John Crispin <john@phrozen.org>,
        Lorenzo Bianconi <lorenzo@kernel.org>,
        linux-wireless-owner@vger.kernel.org
Subject: Re: [PATCH RFC/RFT 4/4] mac80211: Apply Airtime-based Queue Limit (AQL) on packet dequeue
In-Reply-To: <CA+iem5sGkRQy4B=Mz-C90sfqgFmuywa=W7Yjy2CDd8hfppQkvQ@mail.gmail.com>
References: <156889576422.191202.5906619710809654631.stgit@alrua-x1> <156889576869.191202.510507546538322707.stgit@alrua-x1> <08f0ed6e-b746-9689-6dc8-7c0ea705666d@nbd.name> <87wodv19jl.fsf@toke.dk> <b0239b81-e3db-1f10-22cb-ad06536b97a8@nbd.name> <87tv8z13wv.fsf@toke.dk> <CA+iem5uir8GOq2psAVuUParY3MimsGhvdefUopiwHov128CQQA@mail.gmail.com> <87r2421d4f.fsf@toke.dk> <ed46699c32cd3d8d8bcc9bce1375ba04@codeaurora.org> <CA+iem5sGkRQy4B=Mz-C90sfqgFmuywa=W7Yjy2CDd8hfppQkvQ@mail.gmail.com>
Date:   Sun, 29 Sep 2019 21:18:24 +0200
Message-ID: <87muemykqn.fsf@toke.dk>
MIME-Version: 1.0
Content-Type: text/plain; charset=WINDOWS-1252
Content-Transfer-Encoding: quoted-printable
Sender: linux-wireless-owner@vger.kernel.org
Precedence: bulk

Kan Yan <kyan@google.com> writes:

>> No, ath10k would continue to do what it was always doing. Drivers that
>> can report after-the-fact airtime usage per-frame (like ath9k) will
>> continue to do that. In both of those cases, the airtime estimate is
>> only used to throttle the queue, not to schedule for fairness.
> You are right, I didn't realize ath9k reports per frame airtime usage.
>
>> Yeah, I was wondering about that. Makes sense. Why 24ms, exactly?
> The per interface 24 ms queue limit is an empirical number that works
> well for both achieve low latency when there is a lot of stations and
> get high throughput when there is only 1-2 stations.  We could make it
> configurable.

Right. "Found by trial and error" is a fine answer as far as I'm
concerned :)

But yeah, this should probably be configurable, like BQL is.

>> BTW, I think Felix' concern about powersave was in relation to AQL: If
>> we don't handle power save in that, we can end up in a situation where
>>the budget for packets allowed to be queued in the firmware is taken up
>> entirely by stations that are currently in powersave mode; which would
>> throttle the device completely. So we should take that into account for
>> AQL; for the fairness scheduler, stations in powersave are already
>> unscheduled, so that should be fine.
> I think the accounting for the airtime of frames in the power saving
> queue could affect both the fairness scheduler and AQL.
> For chipset with firmware offload, PS handling is mostly done by
> firmware, so host driver's knowledge of PS state could be slightly
> out-of-dated. The power save behavior also make it harder to the
> airtime_weight correct for the fairness scheduler.

Hmm, maybe. I'm not sure how significant this effect would be, but I
guess we'll need to find out :)


> Powersave mode's impact to AQL is much smaller. The lower per station
> queue limit is not impacted by other stations PS behavior, since the
> estimated future airtime is not weighted for other stations and a
> station won't get blocked due to others stations in PS mode.
> Station in PS mode do affects AQL's higher per interface limit, but in
> an inconsequential way. The per interface AQL queue limit is quite
> large (24 ms), hence airtime from packets in PS queue is unlikely to
> have a significant impact on it. Still, it will be better if the
> packet in power saving queue can be taken into account.

I guess the risk is lower when with a 24ms per-iface limit; but with
enough stations I guess it could still happen, no? So we should probably
handle this case...

>> > make it easier to schedule multiple stations, I think it has some meri=
t
>> > that makes it worth trying out. We should probably get the AQL stuff
>> > done first, though, and try the virtual time scheduler on top of that.
>> Agree that we should get the AQL stuff done first since I believe it
>> will help to fix the issue mentioned above.
> That sounds like a good plan. The virtual time scheduler is more
> involved and will take more work to get it right. It make sense to get
> AQL done first.

Cool. Are you going to submit a ported version of your implementation?
Then we can work from the two submissions and see if we can't converge
on something...

-Toke