Re: [PATCH V2] net: ethernet: mellanox: correct page conversion

From: Sinan Kaya
Date: Mon Apr 18 2016 - 12:06:58 EST


On 4/18/2016 11:59 AM, David Miller wrote:
> From: okaya@xxxxxxxxxxxxxx
> Date: Mon, 18 Apr 2016 01:06:27 -0400
>
>> On 2016-04-18 00:00, David Miller wrote:
>>> From: Sinan Kaya <okaya@xxxxxxxxxxxxxx>
>>> Date: Sat, 16 Apr 2016 18:23:32 -0400
>>>
>>>> Current code is assuming that the address returned by
>>>> dma_alloc_coherent
>>>> is a logical address. This is not true on ARM/ARM64 systems. This
>>>> patch
>>>> replaces dma_alloc_coherent with dma_map_page API. The address
>>>> returned
>>>> can later by virtually mapped from the CPU side with vmap API.
>>>> Signed-off-by: Sinan Kaya <okaya@xxxxxxxxxxxxxx>
>>> You can't do this.
>>> The DMA map page API gives non-coherent mappings, and thus requires
>>> proper flushing.
>>> So a straight conversion like this is never legitimate.
>>
>> I would agree on proper dma api usage. However, the code is already
>> assuming coherent architecture by mapping the cpu pages as
>> page_kernel.
>>
>> Dma_map_page returns cached buffers and you don't need cache flushes
>> on coherent architecture to make the data visible.
>
> All you are telling me is that there are two bugs instead of one, so now
> both need to be fixed.
>

The removal of vmap also fixes the coherency assumption. It is one fix
for both.

I was thinking of submitting another patch to change the vmap argument
PAGE_KERNEL based on the coherency support of the architecture. I don't need
to do that anymore if the other experiment works.

--
Sinan Kaya
Qualcomm Technologies, Inc. on behalf of Qualcomm Innovation Center, Inc.
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a Linux Foundation Collaborative Project