Re: [PATCH 1/2] mac80211: Implement Airtime-based Queue Limit (AQL) - Toke Høiland-Jørgensen

From: "Toke Høiland-Jørgensen" <toke@redhat.com>
To: Kan Yan <kyan@google.com>
Cc: Johannes Berg <johannes@sipsolutions.net>,
	linux-wireless@vger.kernel.org,
	make-wifi-fast@lists.bufferbloat.net,
	Felix Fietkau <nbd@nbd.name>, Yibo Zhao <yiboz@codeaurora.org>
Subject: Re: [PATCH 1/2] mac80211: Implement Airtime-based Queue Limit (AQL)
Date: Sun, 06 Oct 2019 19:40:41 +0200	[thread overview]
Message-ID: <87pnj9n55y.fsf@toke.dk> (raw)
In-Reply-To: <CA+iem5vJFRxskyHOKf5K73X8aGH965P4hoiCj-wQtK-Z-47pdg@mail.gmail.com>

Kan Yan <kyan@google.com> writes:

> Hi Johannes,
>
> Thanks for help review this patch.  I will fix all style errors.
>
>> > +ieee80211_txq_airtime_check(struct ieee80211_hw *hw, struct ieee80211_txq *txq);
>> Why is it necessary to call this, vs. just returning NULL when an SKB is
>> requested?
> This function is also called by ath10k, from ath10k_mac_tx_can_push(),
> which returns a boolean.
>
>> However, I'm not sure it *shouldn't* actually be per interface (i.e.
>> moving from
>>         local->aql_total_pending_airtime
>> to
>>
>>         sdata->aql_total_pending_airtime
>>
>> because you could have multiple channels etc. involved and then using a
>> single airtime limit across two interfaces that are actually on two
>> different channels (e.g. 2.4 and 5 GHz) doesn't make that much sense.
>>
>> Actually, it does make some sense as long as the NIC is actually
>> channel-hopping ... but that's in the process of changing now, there's
>> going to be hardware really soon (or perhaps already exists) that has
>> real dual-band capabilities...
>
> That's a good point.  I haven't thought about real simultaneous dual
> band chipset and such chipset do exists now. Is RSDB support coming to
> mac80211 soon? Just curious if it will be just virtual interfaces or
> something else. I chose "local" instead of "sdata" thinking about the
> case of several virtual interfaces (AP, STA, MESH) operates in the
> same channel, then the interface total could be a better choice.
>
> I am ok with moving the "aql_total_pending_airtime" into sdata, but
> afraid that's not the most optimal choice for the case of multiple
> virtual interfaces operates in the same channel.
> Maybe we could leave it in "local" for now. What do you think?

I'd lean towards keeping it in 'local' for consistency with all the
other airtime stuff. For now, I think having multiple SSIDs on the same
radio is more common than the reverse (multiple bands on a single
radio).

In particular, the per-group airtime fairness stuff is definitely
designed on the assumption that all BSSes share the same band.

So if and when we start supporting true multi-band devices we'll have to
change these things anyway. So might as well keep everything together so
it all gets fixed :)

-Toke