Re: [Make-wifi-fast] [PATCH RFC v3 2/4] mac80211: Add airtime accounting and scheduling to TXQs

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

Rajkumar Manoharan <rmanohar@codeaurora.org> writes:

> On 2018-09-10 04:13, Toke Høiland-Jørgensen wrote:
>> Johannes Berg <johannes@sipsolutions.net> writes:
>>>> -			       txqi->flags & (1<<IEEE80211_TXQ_STOP) ? "STOP" : "RUN",
>>>> -			       txqi->flags & (1<<IEEE80211_TXQ_AMPDU) ? " AMPDU" : "",
>>>> -			       txqi->flags & (1<<IEEE80211_TXQ_NO_AMSDU) ? " NO-AMSDU" : 
>>>> "");
>>>> +			       txqi->flags & (1 << IEEE80211_TXQ_STOP) ? "STOP" : "RUN",
>>>> +			       txqi->flags & (1 << IEEE80211_TXQ_AMPDU) ? " AMPDU" : "",
>>>> +			       txqi->flags & (1 << IEEE80211_TXQ_NO_AMSDU) ? " NO-AMSDU" 
>>>> : "");
>>> 
>>> consider BIT() instead as a cleanup? :)
>>> 
>>> (or maybe this is just a leftover from merging some other patches?)
>> 
>> Yeah, this is a merging artifact; Rajkumar's patch added another flag,
>> that I removed again. Didn't notice that there was still a whitespace
>> change in this patch...
>> 
> I added the flag based on our last discussion. The driver needs to check
> txq status for each tx_dequeue(). One time txq check is not sufficient
> as it allows the driver to dequeue all frames from txq.
>
> drv_func() {
>        while (ieee80211_airtime_may_transmit(txq) &&
>                hw_has_space() &&
>               (pkt = ieee80211_tx_dequeue(hw, txq)))
>            push_to_hw(pkt);
> }

Yeah, but with airtime only being recorded on TX completion, the odds of
the value changing within that loop are quite low; so it's not going to
work, which is why I removed it.

However, after reading Kan's patches I get where you're coming from; a
check in tx_dequeue() is needed for the BQL-style queue limiting. Will
try to incorporate a version of that in the next series so you can see
what I mean when I say it should be orthogonal; and I'm still not sure
it needs a flag :)

>>>> +bool ieee80211_txq_may_transmit(struct ieee80211_hw *hw,
>>>> +				struct ieee80211_txq *txq)
>>>> +{
>>>> +	struct ieee80211_local *local = hw_to_local(hw);
>>>> +	struct txq_info *txqi = to_txq_info(txq);
>>>> +	bool may_tx = false;
>>>> +
>>>> +	spin_lock_bh(&local->active_txq_lock);
>>>> +
>>>> +	if (ieee80211_txq_check_deficit(local, txqi)) {
>>>> +		may_tx = true;
>>>> +		list_del_init(&txqi->schedule_order);
>>> 
>
> To handle above case, may_transmit should remove the node only
> when it is in list.
>
> if (list_empty(&txqi->schedule_order))
>         list_del_init(&txqi->schedule_order);

I assume you missed a ! in that if, right? :)

> So that it can be used to determine whether txq is running negative.

But still not sure what you mean here?

-Toke