Re: [PATCH V2 6/8] brcmfmac: update 43012 F2 watermark setting to fix DMA Error during UDP RX Traffic

[Chi-Hsien Lin <Chi-Hsien.Lin@cypress.com>]

(+Madhan)

On 11/12/2018 5:25, Arend van Spriel wrote:
> On 11/12/2018 8:29 AM, Chi-Hsien Lin wrote:
>> From: Naveen Gupta <naveen.gupta@cypress.com>
>>
>> The number of words that the read FIFO has to contain except
>> the end of frame before sends data back to the host.
>> Max watermark = (512B - 2* (BurstLength))/4 =
>> (512 - 128)/4 = 384/4 = 0x60
>> so if burst length (i.e. BurstLength = 64) is increased,
>> watermark has to be reduced. This is the optimal setting for this chip.
>>
>> Reviewed-by: Arend van Spriel <arend.vanspriel@broadcom.com>
>> Signed-off-by: Naveen Gupta <naveen.gupta@cypress.com>
>> Signed-off-by: Chi-Hsien Lin <chi-hsien.lin@cypress.com>
>> ---
>>  drivers/net/wireless/broadcom/brcm80211/brcmfmac/sdio.c | 12 
>> ++++++++++++
>>  1 file changed, 12 insertions(+)
>>
>> diff --git a/drivers/net/wireless/broadcom/brcm80211/brcmfmac/sdio.c 
>> b/drivers/net/wireless/broadcom/brcm80211/brcmfmac/sdio.c
>> index 7707b0169c21..e1708e297d07 100644
>> --- a/drivers/net/wireless/broadcom/brcm80211/brcmfmac/sdio.c
>> +++ b/drivers/net/wireless/broadcom/brcm80211/brcmfmac/sdio.c
>> @@ -52,6 +52,7 @@
>>  /* watermark expressed in number of words */
>>  #define DEFAULT_F2_WATERMARK    0x8
>>  #define CY_4373_F2_WATERMARK    0x40
>> +#define CY_43012_F2_WATERMARK    0x60
> 
> So basically you increase queuing in firmware rx path. How does this
> affect TCP latency. The DMA error obviously needs fixing, but why go
> from a watermark of 32 bytes to the maximum of 384 bytes. Same question
> for 4373.

This is to answer 2/8 and 6/8 -

For current chips like 4373 and 43012, we are configuring F2 block size 
as 256 and it is recommended to program the F2 watermark (SDIO TX path - 
SDIO Device to host traffic) to at least equal or greater than the block 
size to make sure that at least >= 1 Block size of data is in FIFO 
before sending on the SDIO bus to avoid underflow. We haven't observed 
excess TCP latency with this setting.

Regards,
Chi-hsien Lin

> 
> Regards,
> Arend