Re: [PATCH v2 0/2] wifi: ath9k: fix loss of non-aggregate frames
From: Toke Høiland-Jørgensen
Date: Thu Nov 19 2026 - 06:12:17 EST
Toke Høiland-Jørgensen <toke@xxxxxxx> writes:
> Tomas Tonhauser <t.tonhauser@xxxxxxxx> writes:
>
>> On Tue, Nov 17, 2026 at 10:09:58AM +0100, Toke Høiland-Jørgensen wrote:
>>>
>>> So an alternative solution to what appears to be the same problem your
>>> patch 1 is solving appeared:
>>>
>>> https://lore.kernel.org/r/20261110170459.31857-1-dominikengelhardt@xxxxxxxxx
>>>
>>> Could you please check if this approach works in your tests (with or
>>> without patch 2)?
>>>
>>> Thanks!
>>>
>>> -Toke
>>
>> Yes, it works. I tested Dominik's patch instead of patch 1 on the same
>> AP (EAP225 v3, QCA9563, OpenWrt 25.12.5), with patch 2 switched on and
>> off at runtime by a test-only module parameter, alternating runs.
>>
>> With Dominik's patch every filtered frame now reaches mac80211
>> (tx_filtered matches "TXERR Filtered") and the mac80211 retry gets it
>> through:
>>
>> Strong client (-50 dBm), 3x500 pings at 50 ms + 3x25 at 1 s:
>> Dominik only: 0 lost, 105 filtered, 103 via tx_filtered
>> Dominik + patch 2: 0 lost, 111 filtered, retried in ath9k
>>
>> Weak client (-75 dBm), 4x500 pings at 50 ms (1000 bytes),
>> 3x500 at 50 ms (56 bytes), 3x25 at 1 s (1000 bytes):
>> Dominik only: 4 lost, all of them XRETRY
>> Dominik + patch 2: 0 lost
>>
>> So Dominik's patch makes patch 1 unnecessary, and it keeps powersave
>> filtering intact. I'll drop patch 1.
>>
>> Patch 2 still helps with frames that exhaust the hardware retry chain.
>> The link was good during these runs (3 XRETRY in 2000 frames), but in
>> an earlier evening test on the same client it was 130/3000 without the
>> software retry and 1/3000 with it.
>>
>> Dominik's explanation also matches what I see. The strong client got a
>> filtered frame every 1-2.5 s with no XRETRY at all, largely
>> independent of the traffic rate, and both test clients share their
>> block of eight key cache indices with ESP8266 clients that toggle
>> powersave constantly. That answers my open question from the RFC.
>>
>> I think best would be to have my patch 2 on top of Dominik's patch (I
>> can post it now with a dependency note or wait until Dominik's patch
>> is applied).
>
> Yeah, patch 2 is not really a bugfix on its own anyway, so we should
> take that through -next once the bugfixes are merged back.
>
> -Toke
>
> pw-bot: cr
And one more attempt for the bot:
pw-bot: cr
-Toke