arm64 memcpy_{from|to}io and memset_io

arm64 memcpy_{from|to}io and memset_io

[Radha Mohan]

Hi,
I see that the memcpy_{from|to}io and memset_io are not in an
optimized manner. I guess these are just a copy from
arch/arm/include/asm/io.h where there could be problem with different
implementations.
Do we still need these to be byte write ?
Can we convert them to use a more optimized memcpy ?

We have some drivers, like framebuffer driver using these functions
and end up writing byte-by-byte. This causes a very poor VGA
performance.

Let me know if there are any concerns to convert these to use memcpy.
I can send a patch.

regards,
Radha Mohan

arm64 memcpy_{from|to}io and memset_io

[Arnd Bergmann]

On Tuesday 13 October 2015 23:12:18 Radha Mohan wrote:
> Hi,
> I see that the memcpy_{from|to}io and memset_io are not in an
> optimized manner. I guess these are just a copy from
> arch/arm/include/asm/io.h where there could be problem with different
> implementations.
> Do we still need these to be byte write ?

No.

> Can we convert them to use a more optimized memcpy ?

Yes.

> We have some drivers, like framebuffer driver using these functions
> and end up writing byte-by-byte. This causes a very poor VGA
> performance.
> 
> Let me know if there are any concerns to convert these to use memcpy.
> I can send a patch.

A few things to watch out for:

- you cannot use a static inline to do the job, because gcc might
  replace a plain memcpy() with unaligned pointer dereferences
  that are not allowed on __iomem

- when providing an external implementation of the functions, make sure
  they honor the alignment as well

- I think you need the same barriers that readl/writel have, but only
  at the start/end of the loop, not in the middle.

	Arnd

arm64 memcpy_{from|to}io and memset_io

[Catalin Marinas]

On Tue, Oct 13, 2015 at 11:12:18PM -0700, Radha Mohan wrote:
> I see that the memcpy_{from|to}io and memset_io are not in an
> optimized manner. I guess these are just a copy from
> arch/arm/include/asm/io.h where there could be problem with different
> implementations.

I think you may be looking at an older kernel version. In the latest
mainline, memcpy_*io functions are more optimised in the sense that they
use 64-bit accesses if the alignment permits.

> Do we still need these to be byte write ?

No but see above.

> Can we convert them to use a more optimized memcpy ?

There is a risk to converting them to something like memcpy() as the
latter does not guarantee aligned accesses. Alignment is mandatory for
Device memory access.

> We have some drivers, like framebuffer driver using these functions
> and end up writing byte-by-byte. This causes a very poor VGA
> performance.

You probably have an old kernel version.

-- 
Catalin

arm64 memcpy_{from|to}io and memset_io

[Radha Mohan]

On Wed, Oct 14, 2015 at 9:12 AM, Catalin Marinas
<catalin.marinas@arm.com> wrote:
> On Tue, Oct 13, 2015 at 11:12:18PM -0700, Radha Mohan wrote:
>> I see that the memcpy_{from|to}io and memset_io are not in an
>> optimized manner. I guess these are just a copy from
>> arch/arm/include/asm/io.h where there could be problem with different
>> implementations.
>
> I think you may be looking at an older kernel version. In the latest
> mainline, memcpy_*io functions are more optimised in the sense that they
> use 64-bit accesses if the alignment permits.
>
>> Do we still need these to be byte write ?
>
> No but see above.
>
>> Can we convert them to use a more optimized memcpy ?
>
> There is a risk to converting them to something like memcpy() as the
> latter does not guarantee aligned accesses. Alignment is mandatory for
> Device memory access.
>
>> We have some drivers, like framebuffer driver using these functions
>> and end up writing byte-by-byte. This causes a very poor VGA
>> performance.
>
> You probably have an old kernel version.

Yes, I was alternating between old and new kernels. The newer implementation
 is much better. I will try that. Thanks.

>
> --
> Catalin