Re: Broadcom WiFi SDIO performance regression after commit "mmc: sdhci: Remove finish_tasklet"

[Adrian Hunter]

On 27/08/20 9:07 am, Dmitry Osipenko wrote:
> Hello!
> 
> I was debugging WiFi performance problems on Acer A500 tablet device
> that has BCM4329 WiFi chip which is connected to NVIDIA Terga20 SoC via
> SDIO and found that the following commit causes a solid 5-10 Mbit/s of
> WiFi throughput regression after 5.2 kernel:

What is that in percentage terms?

> 
> commit c07a48c2651965e84d35cf193dfc0e5f7892d612
> Author: Adrian Hunter <adrian.hunter@xxxxxxxxx>
> Date:   Fri Apr 5 15:40:20 2019 +0300
> 
>     mmc: sdhci: Remove finish_tasklet
> 
>     Remove finish_tasklet. Requests that require DMA-unmapping or
> sdhci_reset
>     are completed either in the IRQ thread or a workqueue if the
> completion is
>     not initiated by the IRQ.
> 
> Reverting the offending commit on top of recent linux-next resolves the
> problem.
> 
> Ulf / Adrian, do you have any ideas what could be done in regards to
> restoring the SDIO performance? Should we just revert the offending commit?
> 

Unfortunately I think we are past the point of returning to the tasklet.

sdhci can complete requests in the irq handler but only if ->pre_req() and
->post_req() are used, which is not supported by SDIO at present.  pre_req
and post_req were introduced to reduce latency for the block driver, so it
seems reasonable perhaps to look at using them in SDIO as well.