[RFC PATCH] wifi: ath12k: service all RX rings in an interrupt group

[Zephyr Madera]

Hi Jeff,

Thanks for clarifying the need for QCN and WCN coverage. I understand
the concern about changing a shared datapath.

I have access to QCN-based hardware, but no WCN7850 device. Could you
clarify the minimum testing you would need for this particular RX-ring
change?

Since the initial submission, I have checked the patch against ath.git
main at 17e4b9ea527b6c9491bed60f646bb2eff7d8db7b. The baseline and
patched x86_64 driver builds pass, along with source-matched
regression fixtures covering multiple rings, a single ring, an empty
mask and shared-budget handling. These are compilation and
host-fixture results, not a substitute for hardware testing.

In particular:

- Which QCN/WCN receive-path differences or failure cases should the
tests exercise?
- What traffic tests, operating modes, duration and instrumentation
would provide sufficient coverage?
- Is there a preferred firmware version or kernel baseline for that testing?
- Could a maintainer or community member with WCN7850 hardware help
test the patch and provide the appropriate testing attribution?
- Can source review continue while that hardware coverage is arranged,
with the limitation stated explicitly?

I can document the testing performed on the hardware available to me
and provide the focused patch and reproduction details. I want to meet
the acceptance criteria without claiming coverage I cannot verify.

Thanks,
John

On Mon, 05 Oct 2026 19:15:52 -0700, Jeff Johnson
<jeff.johnson at oss.qualcomm.com> wrote:
> On 10/5/2026 5:48 PM, John wrote:
> > An RX interrupt group may select multiple receive rings. The Wi-Fi 7
> > handler currently services only the highest set bit, even when the NAPI
> > budget has room for work from another selected ring. Walk all selected
> > bits and retain the shared budget and existing early-exit behavior.
> >
> > On a locally adapted GL-BE3600/QCN6432, group 5 selects bits 2 and 3. A
> > host regression built from the actual RX block completes 3 of 6 queued
> > packets before the fix and 6 of 6 after it. Single-ring, empty-mask and
> > two-poll budget cases pass. Descending priority is unchanged; this does
> > not establish fairness under sustained traffic.
>
> see:
> https://wireless.docs.kernel.org/en/latest/en/users/drivers/ath12k/submittingpatches.html#tested-on-tag
>
> For a datapath change I'd need Tested-on: tags from both WCN and
> QCN chipsets since they have different requirements.
>
> >
> > Signed-off-by: John <dencity7 at gmail.com>
>
> See:
> https://www.kernel.org/doc/html/latest/process/submitting-patches.html#sign-your-work-the-developer-s-certificate-of-origin
>
> >
> > ---
> > RFC validation and limits:
> >
> > Source applicability and the host regression were checked against
> > backports 7.2 plus the 230 wireless patches from perceival/openwrt-flint3
> > commit 2365932733ca8ec3b346621d9cec2eb3df3b2cf3. The new patch applies
> > with zero fuzz. The same source-block regression and exact patch
> > application also pass against official ath.git main at
> > 388d3c42e8aba1cfc3c6d2c52877a46d8cf740db (zero fuzz). No full module
> > build or hardware test was performed against either current revision.
> >
> > Historical tests on Linux 6.18.39 with local backports 7.2-rc4/QCN6432
> > changes passed DHCP, DNS, 32 UDP flows and about 181 seconds of traffic
> > after reboot. These are bounded observations, not throughput or long-term
> > stability qualification. The contribution was developed with AI assistance.
>
> see:
> https://docs.kernel.org/process/coding-assistants.html#attribution
>
> >
> > --- a/drivers/net/wireless/ath/ath12k/wifi7/dp.c
> > +++ b/drivers/net/wireless/ath/ath12k/wifi7/dp.c
> > @@ -52,8 +52,10 @@
> >  			goto done;
> >  	}
> >
> > -	if (dp->hw_params->ring_mask->rx[grp_id]) {
> > -		i = fls(dp->hw_params->ring_mask->rx[grp_id]) - 1;
> > +	ring_mask = dp->hw_params->ring_mask->rx[grp_id];
> > +	while (ring_mask) {
> > +		i = fls(ring_mask) - 1;
> > +		ring_mask &= ~BIT(i);
> >  		work_done = ath12k_wifi7_dp_rx_process(dp, i, napi, budget);
> >  		budget -= work_done;
> >  		tot_work_done += work_done;