Coherent memory allocator

Coherent memory allocator

[Andrew Murray]

Hello,

Whilst trying to reduce boot time on an SH kernel I noticed the
coherent memory allocator (dma_generic_alloc_coherent) in
arch/sh/mm/consistent.c always memset's the memory it allocates to 0 -
regardless of the user's intention (i.e use of __GFP_ZERO flag as
argument).

I would expect this function to not memset it's memory and instead
allow the user to specify if they want the memory zeroed through the
gfp flags. This would ensure memory is only zeroed when required - and
when it is required the __get_free_pages implementation uses the more
efficient clear_page implementation.

However removing the memset may break drivers than rely on zeroed
memory even though they don't request it.

Is my reasoning valid? Shall I submit a patch?

Thanks,

Andrew Murray

Re: Coherent memory allocator

[Conrad Parker]

On 29 July 2010 19:50, Andrew Murray <amurray@mpcdata.com> wrote:
> Hello,
>
> Whilst trying to reduce boot time on an SH kernel I noticed the
> coherent memory allocator (dma_generic_alloc_coherent) in
> arch/sh/mm/consistent.c always memset's the memory it allocates to 0 -
> regardless of the user's intention (i.e use of __GFP_ZERO flag as
> argument).
>
> I would expect this function to not memset it's memory and instead
> allow the user to specify if they want the memory zeroed through the
> gfp flags. This would ensure memory is only zeroed when required - and
> when it is required the __get_free_pages implementation uses the more
> efficient clear_page implementation.
>
> However removing the memset may break drivers than rely on zeroed
> memory even though they don't request it.
>
> Is my reasoning valid? Shall I submit a patch?

I have no idea if your reasoning is valid, but in general if you send
a patch to the list then it makes it easier for anyone who might be
affected to test if their environment breaks ... no need to ask for
permission to send a patch :)

Conrad.