Re: [PATCH] wifi: mac80211: scale the airtime queue limit by the station's weight

[Johannes Berg <johannes@sipsolutions.net>]

On Mon, 2026-08-24 at 16:24 +0200, Julius Bairaktaris wrote:
> The airtime queue limit is the same for every station, so on a driver
> that pushes a whole scheduling selection into the hardware, every station
> is allowed the same airtime in flight and the deficit round robin can
> only decide the order in which they reach that ceiling, not how much of
> the medium each one gets. Setting NL80211_ATTR_AIRTIME_WEIGHT then has
> no effect on the airtime a station receives.
> 
> Measured on an IPQ8074 access point with two stations on one radio, a 1x1
> VHT80 client and a 2x2 HE160 client, both saturated from the access point
> at once, BE aql_txq_limit at 500/1000 us, three interleaved runs per
> weight with the association verified unchanged across every run. Without
> this change the slow station takes 44.4, 43.9 and 44.0 per cent of the
> medium at equal weights, 44.0, 43.5 and 44.4 at 1024:256, and 43.1, 43.4
> and 42.8 at 256:1024. The weight moves the split by less than the spread
> within one setting.
> 
> Scale the per-station limit by the station's weight. A station weighted
> above the default is allowed proportionally more airtime in flight and
> therefore takes proportionally more of the medium. A network that never
> sets a weight is unaffected, since every station keeps the default.
> 
> The same pair then reads 42.1, 43.3 and 42.8 per cent at equal weights,
> 78.0, 78.9 and 77.2 at 1024:256, and 19.6, 18.3 and 20.0 at 256:1024: a
> 4:1 weight produces a 3.6:1 airtime ratio and 1:4 produces 4.2:1. It is
> not free, and over the pair the aggregate is 557 Mbit/s at equal weights,
> 925 when the faster station is favoured and 337 when the slower one is.
> 
> The low limit is consulted without reference to the radio total, so the
> scaled value is bounded by aql_threshold. Without that bound a weight of
> 65535 grants one station 1.3 seconds of standing airtime on the default
> low limit, which no total accounts for. The bound reaches the high limit
> as well, where it changes nothing: that check already requires the radio
> total to sit below the same threshold.
> 
> The bound never reduces a limit below the value configured for it, so an
> access point that raises aql_txq_limit above aql_threshold keeps what it
> asked for at every weight, and scaling a weight up cannot scale a limit
> down.
> 
> That bound is also the ceiling on what a weight can express. At the
> default low limit of 5000 us it takes effect near a weight of 1229, so
> ratios beyond roughly five to one do not resolve: a weight of 65535
> against 256 asks for 256 to 1 and delivers 4.9 to 1.
> 
> This changes nothing where the limit does not bind. A station whose
> pending airtime never approaches the limit is not held back by it at
> either weight, mac80211 holds no backlog to arbitrate, and the share each
> station gets is decided by the hardware's own scheduling.

I can't really follow that. Do you understand the change?

> +/* The airtime a station may keep in flight scales with its weight. The low
> + * limit is consulted unconditionally, so the scaled value is bounded by the
> + * radio-wide threshold rather than by the weight alone, and never below the
> + * limit that was configured.
> + */
> +static u32 ieee80211_aql_sta_limit(struct ieee80211_local *local,
> +				   struct sta_info *sta, u32 limit)
> +{
> +	u32 scaled;
> +
> +	if (sta->airtime_weight == IEEE80211_DEFAULT_AIRTIME_WEIGHT)
> +		return limit;
> +
> +	scaled = mult_frac(limit, sta->airtime_weight,

airtime weight is already used to select when the queues are serviced,
so why should it be used for the queue depth?

johannes