+ arm-move-arm_dma_limit-to-setup_dma_zone.patch added to -mm tree

[Ard Biesheuvel]

On 21 May 2014 08:50, Vladimir Murzin <murzin.v at gmail.com> wrote:
> Hi Ard
>
> On 5/19/14, Ard Biesheuvel <ard.biesheuvel at linaro.org> wrote:
>> Gents,
>>
>> I noticed this patch turning up in -next:
>> http://ozlabs.org/~akpm/mmots/broken-out/arm-move-arm_dma_limit-to-setup_dma_zone.patch
>>
>> which I think is incorrect. More specifically, it replaces a reference
>> to arm_dma_limit, which is defined as either
>>
>> arch/arm/mm/init.c:177:         arm_dma_limit = PHYS_OFFSET +
>> arm_dma_zone_size - 1;
>>
>> or
>>
>> arch/arm/mm/init.c:179:         arm_dma_limit = 0xffffffff;
>>
>> with 'arm_dma_pfn_limit << PAGE_SHIFT', which means the bottom
>> PAGE_SHIFT '1' bytes are lost.
>>
>
> Do not we lost even more with
> dma_contiguous_reserve
>      dma_contiguous_reserve_area
> ...
>         /* Sanitise input arguments */
>         alignment = PAGE_SIZE << max(MAX_ORDER - 1, pageblock_order);
>         base = ALIGN(base, alignment);
>         size = ALIGN(size, alignment);
>         limit &= ~(alignment - 1);
>

That's interesting. It looks like there may be a mismatch here: we
pass an inclusive limit where the callee expects it to be exclusive,
resulting in the top 4 megabytes of the DMA zone to be lost (for
pageblock_order == MAX_ORDER-1). I.e., if arm_dma_zone == SZ_64M, we
will end up calling __memblock_alloc_base() with a limit of 60 megs.

@Russell: care to share your thought on this?

-- 
Ard.