Re: [PATCH ath-next v5 0/9] wifi: ath11k: airtime queue limits, airtime fairness and a driver TXQ scheduler

From: Christopher Griffin

Date: Tue Nov 10 2026 - 02:17:18 EST


(In reply to the v5 cover letter,
https://lore.kernel.org/all/20261018174904.1757161-1-julius@xxxxxxxxxxxxxx/
which my mailbox doesn't hold, so this may not thread.)

Hi Julius,

I tested v5 on an Arista C-360 AP: IPQ8074 (AHB) for 2.4 and 5 GHz,
QCN9074 (PCI) for 6 GHz. The base was OpenWrt main with backports 7.2
and the TX flow control v7 patches 1-3 underneath. Flow control
patches 1 and 2 needed fuzz 1 against OpenWrt's IPQ9574 patch; the
rest applied as posted. I swapped the modules in RAM and ran A/B/A
against the unpatched driver.

2.4 GHz, ch 1 20 MHz, two stations downloading TCP for 60 s each:
an 11g-only station (a FastConnect 7800 with HT/VHT/HE off) and a
2x2 HE20 station (BE200). A wired host pinged the 11g station every
0.2 s:

run | 11g Mbit/s | HE20 Mbit/s | ping avg / p95 ms | ping loss
without (A1) | 1.6 | 64.6 | 5275 / 7710 | 4.7%
with v5 | 1.5 | 52.8 | 375 / 769 | 0%
without (A2) | 1.6 | 68.2 | 5415 / 9953 | 6.7%

On the slow station that's the same bufferbloat cure your cover
letter shows. The codel drops land where they should (385 on the 11g
station, 89 on the HE20 one). Its share of the air stayed at about
5%, which matches what you saw about the firmware serving queued
stations in turn.

6 GHz, ch 37 160 MHz, one 2x2 HE station (FastConnect 7800) alone,
TCP download, 4 streams, 60 s:

run | Mbit/s | TCP retx | loaded ping avg / p95 ms
without, run 1 | 1609 | 2928 | 31.7 / 57.6
without, run 2 | 1622 | 2333 | 17.2 / 27.7
with v5 | 1154 | 13100 | 6.5 / 13.3
with v5, fq memory 16 MB | 1340 (*) | 8077 | 10.4 / 24.3
with v5, fq memory 32 MB | 1398 | 2073 | 15.9 / 33.2
(*) steady part only, after a client-side reassociation

The default run's loss wasn't codel: the station showed no drops or
marks. It was fq_overmemory, 409 in the 60 s. mac80211 gives the TXQs
16 MB only when some band has VHT (ieee80211_txq_setup_flows()), and a
6 GHz-only radio has none, so the QCN9074 radio got 4 MB while the
IPQ8074 5 GHz radio got 16 MB. With AQL the backlog moves from the
firmware rings into mac80211, and at 1.4 to 1.6 Gbit/s 4 MB isn't
enough. At 32 MB there were no overmemory drops. What's left (-13%)
seems to be AQL's own cost at 2-stream HE-MCS 9 to 11. The 4 MB cap
looks worth fixing in mac80211 on its own, for any HE radio without
VHT.

I also tried the series with NL80211_EXT_FEATURE_AQL left unset on
the 6 GHz radio only (aql_enable 0 should behave the same). The radio
then ran 1810 to 1906 Mbit/s for seconds at a time, above the 1615
without the series, but it kept overrunning the TX buffer IDR: about
14000 "failed to transmit frame -28" in two runs of 60 s and 45 s,
each burst followed by a dip to 1000 to 1400 Mbit/s.
ath11k_mac_tx_push_txq() checks the TCL ring for room but not the
IDR, so without the airtime limit nothing bounds the frames in
flight. It may be worth ending the round on a full IDR too.

One note on 4/9: tx_duration there is the user's fes_duration_us, and
it's registered in full for each user of the PPDU. For a DL OFDMA or
MU-MIMO PPDU, that charges every station the whole exchange. My
duplicate of this patch (below) divided it by num_users. I didn't
test MU here (OFDMA and MU-MIMO were off).

That duplicate: I sent "wifi: ath11k: feed mac80211's airtime
fairness" on 11-08 without having found your series. It repeats your
1/9 and 4/9, and I've asked Jeff to drop it.

With the series as posted there were no warnings, crashes or
firmware asserts in any run.

Tested-on: IPQ8074 hw2.0 AHB WLAN.HK.2.9.0.1-02146-QCAHKSWPL_SILICONZ-1
Tested-on: QCN9074 hw1.0 PCI WLAN.HK.2.5.r4-00683 (AP vendor's build)

Tested-by: Christopher Griffin <christophergriffin05@xxxxxxxxx>

The tests and this report were done with Claude, an AI coding
assistant.

Christopher