Re: [PATCH v13 01/10] iomap: Use correct endian conversion function in mmio_writeXXbe

[Arnd Bergmann]

On Mon, Mar 26, 2018 at 6:21 PM, Logan Gunthorpe <logang@xxxxxxxxxxxx> wrote:
>
>
> On 26/03/18 04:53 AM, Arnd Bergmann wrote:
>> On most architectures, this is not important:
>> - For x86, the stores are aways atomic and no additional barriers
>>   are needed, so the two are the same
>> - For ARM (both 32 and 64-bit), powerpc and many others, we don't
>>   use the generic iowrite() and just fall back to writel() or
>>   writel(swab32()).
>>
>> However, shouldn't we just use the writel(swab32()) logic here as well
>> for the common case rather than risking missing barriers?
>
> Hmm, I don't know... it's complicated?
>
> Doing a bit of digging shows that the existing code was written during a
> time when writel() did not include extra barriers over __raw_writel() in
> any of the common arches.
>
> The commit logs don't seem to provide any guidance as to why this it was
> done this way, but I'd assume it was done to avoid a double swab() call
> on BE arches. Seeing writel() is typically implemented as:
>
> __raw_writel(__cpu_to_le32(value), addr);
>
> Then on BE arches, writel(swab32()) would become:
>
> __raw_writel(swab32(swab32(value)), addr)
>
> Which seems undesirable.

I wouldn't expect it to matter: the byte swap is almost always much
cheaper compared to the actual bus access for the MMIO, and I
would also guess that modern compilers can eliminate the double
byte swap on architectures where writel() is an inline function. Most of
the important architectures use ARCH_USE_BUILTIN_BSWAP, which
guarantees that.

       Arnd