Re: [PATCH] pxa_camera: Redesign DMA handling

[Robert Jarzmik <robert.jarzmik@free.fr>]

Guennadi Liakhovetski <g.liakhovetski@gmx.de> writes:

> (moved to the new v4l list)
>
>> The DMA transfers in pxa_camera showed some weaknesses in
>> multiple queued buffers context :
>>  - poll/select problem
>>    The order between list pcdev->capture and DMA chain was
>>    not the same. This creates a discrepancy between video
>>    buffers marked as "done" by the IRQ handler, and the
>>    really finished video buffer.
>
> Could you please describe where and how the order could get wrong?
Sorry, I missed that point in the previous reply.

It's still the same bit of code :
-                       } else {
-                               buf_dma->sg_cpu[nents].ddadr =
-                                       DDADR(pcdev->dma_chans[i]);

That chains the end of the queued buffer to the active buffer (probably the one
running in DMA chain [1]). So we'll get images in the following order:

 <queued_buffer>, then <active_buffer (head of pcdev->capture)>, then <others>.

The desired order is :
 <active_buffer (head of pcdev->capture)>, then <others>, then <queued_buffer>.

>>  - multiple buffers DMA starting
>>    When multiple buffers were queued, the DMA channels were
>
> You mean multiple scatter-gather elements?
No, I mean multiple video buffers.
Multiple scatter-gather elements form one video buffer, and multiple video
buffers are queued to form a list of images (video stream).

Cheers.

--
Robert

[1] I mean _probably_, because the DMA chain can be already ahead of 1 buffer
(we're talking in terms of PXA cycles, which is very hard to watch), while
active pointer is not yet updated.