Re: [RFC PATCH 1/1] mac80211: use AQL airtime for expected throughput.

[Felix Fietkau]

On 25.05.22 14:14, Baligh GASMI wrote:
> > > Since the integration of AQL, packet TX airtime estimation is
> > > calculated and counted to be used for the dequeue limit.
> > >
> > > Use this estimated airtime to compute expected throughput for
> > > each station.
> > >
> > > It will be a generic mac80211 implementation. If the driver has
> > > get_expected_throughput implementation, it will be used instead.
> > >
> > > Useful for L2 routing protocols, like B.A.T.M.A.N.
> > >
> > > Signed-off-by: Baligh Gasmi <gasmibal@xxxxxxxxx>
> > > ---
> > >  net/mac80211/driver-ops.h |  2 ++
> > >  net/mac80211/sta_info.h   |  2 ++
> > >  net/mac80211/status.c     | 22 ++++++++++++++++++++++
> > >  net/mac80211/tx.c         |  3 ++-
> > >  4 files changed, 28 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/net/mac80211/driver-ops.h b/net/mac80211/driver-ops.h
> > > index 4e2fc1a08681..4331b79647fa 100644
> > > --- a/net/mac80211/driver-ops.h
> > > +++ b/net/mac80211/driver-ops.h
> > > @@ -1142,6 +1142,8 @@ static inline u32 drv_get_expected_throughput(struct ieee80211_local *local,
> > >       trace_drv_get_expected_throughput(&sta->sta);
> > >       if (local->ops->get_expected_throughput && sta->uploaded)
> > >               ret = local->ops->get_expected_throughput(&local->hw, &sta->sta);
> > > +     else
> > > +             ret = ewma_avg_est_tp_read(&sta->status_stats.avg_est_tp);
> > >       trace_drv_return_u32(local, ret);
> > >
> > >       return ret;
> > > diff --git a/net/mac80211/sta_info.h b/net/mac80211/sta_info.h
> > > index 379fd367197f..fe60be4c671d 100644
> > > --- a/net/mac80211/sta_info.h
> > > +++ b/net/mac80211/sta_info.h
> > > @@ -123,6 +123,7 @@ enum ieee80211_sta_info_flags {
> > >  #define HT_AGG_STATE_STOP_CB         7
> > >  #define HT_AGG_STATE_SENT_ADDBA              8
> > >
> > > +DECLARE_EWMA(avg_est_tp, 8, 16)
> > >  DECLARE_EWMA(avg_signal, 10, 8)
> > >  enum ieee80211_agg_stop_reason {
> > >       AGG_STOP_DECLINED,
> > > @@ -641,6 +642,7 @@ struct sta_info {
> > >               s8 last_ack_signal;
> > >               bool ack_signal_filled;
> > >               struct ewma_avg_signal avg_ack_signal;
> > > +             struct ewma_avg_est_tp avg_est_tp;
> > >       } status_stats;
> > >
> > >       /* Updated from TX path only, no locking requirements */
> > > diff --git a/net/mac80211/status.c b/net/mac80211/status.c
> > > index e81e8a5bb774..647ade3719f5 100644
> > > --- a/net/mac80211/status.c
> > > +++ b/net/mac80211/status.c
> > > @@ -1145,6 +1145,28 @@ void ieee80211_tx_status_ext(struct ieee80211_hw *hw,
> > >                       sta->status_stats.retry_failed++;
> > >               sta->status_stats.retry_count += retry_count;
> > >
> > > +             if (skb && tx_time_est) {
> >
> > Shouldn't this be conditioned on actually being used (i.e., existence of
> > get_expected_throughput op? Also maybe pull it out into its own function
> > to make it clear what it's doing...
>
> It's already the case I think, since the tx_time_est is not-zero only
> when actually an estimated time is set.
> A dedicated function seems good for me, for clarity, yes.
>
> > > +                     /* max average packet size */
> > > +                     size_t pkt_size = skb->len > 1024 ? 1024 : skb->len;
> > > +
> > > +                     if (acked) {
> > > +                             /* ACK packet size */
> > > +                             pkt_size += 14;
> > > +                             /* SIFS x 2 */
> > > +                             tx_time_est += 2 * 2;
> > > +                     }
> > > +
> > > +                     /* Backoff average x retries */
> > > +                     tx_time_est += retry_count ? retry_count * 2 : 2;
> > > +
> > > +                     /* failed tx */
> > > +                     if (!acked && !noack_success)
> > > +                             pkt_size = 0;
> > > +
> > > +                     ewma_avg_est_tp_add(&sta->status_stats.avg_est_tp,
> > > +                                         ((pkt_size * 8) * 1000) / tx_time_est);
> >
> > Could we avoid adding this division in the fast path?
>
> Maybe we can use the do_div() macro for optimization, I don't see how
> we can avoid it.
Here's one idea on how to make it less expensive and improve the average 
of the throughput estimation at the same time:

In sta->status_stats add these fields:

u32 last_tp_update;
u32 pkt_size;
u32 tx_time_est;

And in this part of the code you do something like this:

u32 diff = ((u32)jiffies) - sta->status_stats.last_tp_update;
sta->status_stats.pkt_size += (pkt_size * 8) * 1000;
sta->status_stats.tx_time_est += tx_time_est;
if (diff > HZ / 10) {
	ewma_avg_est_tp_add(&sta->status_stats.avg_est_tp,
			    sta->status_stats.pkt_size /
			    sta->status_stats.tx_time_est);
	sta->status_stats.pkt_size = 0;
	sta->status_stats.tx_time_est = 0;
}

This ensures that you will only get max. 1 update per 100ms interval.

Does that make sense?

- Felix