wifi: mwifiex: Replace RX workqueues with kthreads

This improves the RX throughput likely by better locality for the work
loads.

We tested this patch on Mediatek MT8173 Chromebooks, and it shows ~80%
Rx throughput improvement on high data rate test cases.

Signed-off-by: Pin-yen Lin <treapking@chromium.org>
---

 .../wireless/marvell/mwifiex/11n_rxreorder.c  |  2 +-
 .../net/wireless/marvell/mwifiex/cfg80211.c   |  2 +-
 drivers/net/wireless/marvell/mwifiex/main.c   | 29 ++++++++-----------
 drivers/net/wireless/marvell/mwifiex/main.h   |  5 ++--
 4 files changed, 17 insertions(+), 21 deletions(-)

Hi,

Thanks Pin-yen for most of the investigation here and for pushing the
patch. With some additional information though, I might suggest *not*
landing this patch at the moment. More details appended:

On Sat, Jun 10, 2023 at 01:41:51AM +0800, Pin-yen Lin wrote:
> I realized that I might have over-simplified the background and the
> impact of this patch...
> 
> The short answer to the question is that the throughput improved from
> 100 mbps to 180 mbps. The test was run on ChromeOS's v5.15 kernel
> fork. More detailed test setting is mentioned in [1].
> 
> However, the throughput of the same test case on our v4.19 kernel is
> 320 mbps. That is, we observed a 320 mbps --> 100 mbps regression when
> we tried to update the kernel version. This patch is more like a
> mitigation of the regression. It improves the throughput, even though
> it is still not as good as the older kernel.
> 
> That being said, this patch does improve the throughput, so we think
> this patch can be landed into the mainline kernel.
> 
> Best regards,
> Pin-yen
> 
> [1]: https://lore.kernel.org/all/ZFvpJb9Dh0FCkLQA@google.com/

I (we?) was optimistic this would be an improvement (or at least, no
worse) due to some of the reasoning at [1]. And, the work here is just a
single work item, queued repeatedly to the same unbound workqueue. So
conceptually, it shouldn't be much different than a kthread_worker,
except for scheduler details -- where again, we'd think this should be
an improvement, as the scheduler would now better track load for the
task (mwifiex RX) in question.

But additional testing on other mwifiex-based systems (RK3399 + PCIE
8997) showed the inverse: some throughput drops on similar benchmarks,
from 110 Mbps to 80 Mbps. (Frankly, both numbers are significantly below
where we might like...)

Considering we still don't have a full explanation for all the
performance differences we've been seeing (on either test platform), and
that at least one of our platforms showed a (smaller) regression, I
think we might want to do more research before committing to this.

Brian

Brian Norris <briannorris@chromium.org> writes:

> Hi,
>
> Thanks Pin-yen for most of the investigation here and for pushing the
> patch. With some additional information though, I might suggest *not*
> landing this patch at the moment. More details appended:
>
> On Sat, Jun 10, 2023 at 01:41:51AM +0800, Pin-yen Lin wrote:
>> I realized that I might have over-simplified the background and the
>> impact of this patch...
>> 
>> The short answer to the question is that the throughput improved from
>> 100 mbps to 180 mbps. The test was run on ChromeOS's v5.15 kernel
>> fork. More detailed test setting is mentioned in [1].
>> 
>> However, the throughput of the same test case on our v4.19 kernel is
>> 320 mbps. That is, we observed a 320 mbps --> 100 mbps regression when
>> we tried to update the kernel version. This patch is more like a
>> mitigation of the regression. It improves the throughput, even though
>> it is still not as good as the older kernel.
>> 
>> That being said, this patch does improve the throughput, so we think
>> this patch can be landed into the mainline kernel.
>> 
>> Best regards,
>> Pin-yen
>> 
>> [1]: https://lore.kernel.org/all/ZFvpJb9Dh0FCkLQA@google.com/
>
> I (we?) was optimistic this would be an improvement (or at least, no
> worse) due to some of the reasoning at [1]. And, the work here is just a
> single work item, queued repeatedly to the same unbound workqueue. So
> conceptually, it shouldn't be much different than a kthread_worker,
> except for scheduler details -- where again, we'd think this should be
> an improvement, as the scheduler would now better track load for the
> task (mwifiex RX) in question.
>
> But additional testing on other mwifiex-based systems (RK3399 + PCIE
> 8997) showed the inverse: some throughput drops on similar benchmarks,
> from 110 Mbps to 80 Mbps. (Frankly, both numbers are significantly below
> where we might like...)
>
> Considering we still don't have a full explanation for all the
> performance differences we've been seeing (on either test platform), and
> that at least one of our platforms showed a (smaller) regression, I
> think we might want to do more research before committing to this.

Yeah, I agree and I'll drop this. This is a really weird problem, I hope
you can get to the bottom of it.