Re: [PATCH v3 6/7] ARM: NOMMU: Set ARM_DMA_MEM_BUFFERABLE for M-class cpus

[Arnd Bergmann]

On Wed, Apr 19, 2017 at 4:10 PM, Vladimir Murzin
<vladimir.murzin@xxxxxxx> wrote:
> On 19/04/17 11:02, Arnd Bergmann wrote:

>> Can you either modify the description to explain why we now need this on
>> all ARMv7M, or add a '|| CPU_V7M' for the 'bool' line to make it optional?
>>
>> Would it be better to leave the default as disabled on CPU_V7M and
>> require users to enable it manually? That way we don't regress the
>> performance of readl/writel on platforms that don't need this.
>>
>
> It is architectural vs implementation differences. Even though existing
> implementations rarely need this sticking with architecture (it permits memory
> re-ordering to happen in many cases) makes code more robust and save some
> debugging time when more sophisticated implementations go wild (see, for
> instance, 8e02676ffa69 "ARM: 8610/1: V7M: Add dsb before jumping in handler
> mode"). We can consider making it user selectable if performance regressions
> are reported.

I would expect the performance difference to be noticeable, even though
MMIO tends to be a rather slow operation, having tons of barriers in drivers
that don't do DMA but transfer data through a series of writel() can be
significant.

      Arnd