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

[Baligh GASMI]

> > 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.


>
> > +             }
> > +
> >               if (ieee80211_hw_check(&local->hw, REPORTS_TX_ACK_STATUS)) {
> >                       if (sdata->vif.type == NL80211_IFTYPE_STATION &&
> >                           skb && !(info->flags & IEEE80211_TX_CTL_HW_80211_ENCAP))
> > diff --git a/net/mac80211/tx.c b/net/mac80211/tx.c
> > index b6b20f38de0e..d866a721690d 100644
> > --- a/net/mac80211/tx.c
> > +++ b/net/mac80211/tx.c
> > @@ -3793,7 +3793,8 @@ struct sk_buff *ieee80211_tx_dequeue(struct ieee80211_hw *hw,
> >       IEEE80211_SKB_CB(skb)->control.vif = vif;
> >
> >       if (vif &&
> > -         wiphy_ext_feature_isset(local->hw.wiphy, NL80211_EXT_FEATURE_AQL)) {
> > +         (!local->ops->get_expected_throughput ||
> > +         wiphy_ext_feature_isset(local->hw.wiphy, NL80211_EXT_FEATURE_AQL))) {
>
> This implicitly enables AQL for every driver that doesn't set
> get_expected_throughput, no? That is probably not a good idea...

No, AQL will be disabled if it's already the case, according to the
NL80211_EXT_FEATURE_AQL, only airtime estimation will be activated
with this diff. The function ieee80211_sta_update_pending_airtime() is
already checking for NL80211_EXT_FEATURE_AQL.

>
> -Toke
>

Back to the base idea, what do you think about it, can it be a good
road to take ?
Other factors can be added progressively to better estimate the throughput...