Re: MMC performance degradation due to too many requests

[Faiz Abbas]

Hi,

On Wednesday 16 May 2018 07:59 PM, Linus Walleij wrote:
> On Tue, May 15, 2018 at 5:11 PM, Faiz Abbas <faiz_abbas@xxxxxx> wrote:
>> On Tuesday 15 May 2018 06:09 PM, Linus Walleij wrote:
>>> On Mon, May 14, 2018 at 1:30 PM, Faiz Abbas <faiz_abbas@xxxxxx> wrote:
>>>
>>>> Also, I see many times the number of interrupts for mmc1 in 4.14.
>>> (...)
>>>> I also printed out the size of the DMA transfers in the omap_hsmmc
>>>> driver during the dd command. I see that the maximum size of the dma
>>>> transfer has decreased from 1 MB to 256kB between 4.13 and 4.14.
>>>
>>> This seems like my fault for removing the bounce buffer.
>>> (The name bounce buffer is misleading, I suspect the right
>>> term for what is was doing was "merge buffer" or something,
>>> coalescing physically consecutive requests.)
>>>
>>> commit de3ee99b097dd51938276e3af388cd4ad0f2750a
>>> "mmc: Delete bounce buffer handling"
>>> should be the culprit.
>>
>> I don't think so. I checked out to the commit before that and see that
>> the interrupts are still too high.
> 
> OK phew then I'm safe :)
> 
>>> The bounce buffer only makes any performance impact if the
>>> host DMA controller is unable to handle scatter-gather lists
>>> so the number of segments is set down to 1.
>>>
>>> Can you verifty that:
>>> - mmc->max_segs == 1
>>
>> No. It is 64 in v4.14. It was 1024 in v4.13.
> 
> This smells like:
> commit 9442400486a1ec28915e9ab31517a5665ed35705
> "mmc: omap_hsmmc: Reduce max_segs for reliability"
> 

Yes. I was able to bisect the number of interrupts to the same commit as
well.

On top of v4.14,

*Without revert*:
https://pastebin.ubuntu.com/p/VJgvzrCP9g/

avg speed for 100 MB transfer: 3.7 MBps
Number of interrupts per 100 MB transfer: ~2000

*After revert*:
https://pastebin.ubuntu.com/p/ktgTmKVKkN/

avg speed for 100 MB transfer: 4.6 MBps
Number of interrupts per 100 MB transfer: ~500

Also, I'm getting 1M buffers in the omap_hsmmc DMA now.

@Will will it be OK to change the mmc_segs back to 1024?

Thanks,
Faiz