DMA, CMA, coherence and performance

[Steven Bell]

On Sun, Jan 5, 2014 at 8:02 AM, Peter Teoh <htmldeveloper at gmail.com> wrote:

> I think this discussion should help you:
>
> http://e2e.ti.com/support/embedded/linux/f/354/t/89419.aspx
>
> other failures:
>
> http://stackoverflow.com/questions/14625919/allocating-a-large-dma-buffer
>
> and some guideline here:
>
> https://www.kernel.org/doc/Documentation/DMA-API.txt
>
> https://lkml.org/lkml/2011/3/25/19
>
> As I don't have any specific crashdump or error information, nothing I can
> comment further about your problem.   It is quite difficult to make general
> comment.
>
>
>
> On Fri, Jan 3, 2014 at 7:20 AM, Steven Bell <botsnlinux at gmail.com> wrote:
>
>> Hi,
>>
>> I'm working on a device driver for a video device which continuously
>> reads and writes image frames using DMA. The frames are fairly large, in
>> the range of 2-8MB, and I would like the buffers for them to be contiguous
>> because of my hardware. My understanding is that using the contiguous
>> memory allocator is the current "right way" to get the buffers, and that
>> CMA operates entirely behind the scenes when calls are made to
>> dma_alloc_coherent().
>>
>> However, it seems that for this system, a streaming DMA setup would be
>> more appropriate.  The buffer gets filled with data once, handed to the
>> device, and then isn't touched again until it gets reused with new data.
>> The resources I've read have hinted that streaming DMA has some performance
>> benefits over coherent DMA, so this seems like the way to go.  But I
>> haven't seen any discussion of how to use CMA with streaming DMA (or
>> whether such a thing is even necessary).
>>
>> Does the CMA also work behind get_free_pages, or other kernel memory
>> allocation methods?  Does it matter?  The kernel newbies page on memory
>> allocation (http://kernelnewbies.org/KernelMemoryAllocation) says that
>> get_free_pages up to about 8MB are ok.  Is that a generalization based on
>> typical memory fragmentation, or a guarantee?
>>
>> Thanks,
>> Steven
>>
>> _______________________________________________
>> Kernelnewbies mailing list
>> Kernelnewbies at kernelnewbies.org
>> http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies
>>
>>
>
>
> --
> Regards,
> Peter Teoh
>

Thanks for the reply and the links.  Unfortunately, they don't really
answer my basic question - which is a general conceptual question, not a
specific error or problem that I'm running into.

I want to allocate a contiguous 8MB buffer and use it for streaming DMA.  I
don't need or want a coherent buffer, since I'm using streaming DMA, and so
dma_alloc_coherent() seems like the wrong method to use.  Is that true, or
should I use it anyway?  If not, what method do I use?

The DMA-API and DMA-API-HOWTO documents don't seem to say anything about
allocating memory for streaming DMA operations.  Have I missed something?

Thanks,
Steven
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.kernelnewbies.org/pipermail/kernelnewbies/attachments/20140105/16d1ba0c/attachment.html