[PATCH v3 1/3] dma: Support multiple interleaved frames with non-contiguous memory

[Srikanth Thokala]

On Wed, Feb 19, 2014 at 12:33 AM, Jassi Brar <jaswinder.singh at linaro.org> wrote:
> On 18 February 2014 23:16, Srikanth Thokala <sthokal at xilinx.com> wrote:
>> On Tue, Feb 18, 2014 at 10:20 PM, Jassi Brar <jaswinder.singh at linaro.org> wrote:
>>> On 18 February 2014 16:58, Srikanth Thokala <sthokal at xilinx.com> wrote:
>>>> On Mon, Feb 17, 2014 at 3:27 PM, Jassi Brar <jaswinder.singh at linaro.org> wrote:
>>>>> On 15 February 2014 17:30, Srikanth Thokala <sthokal at xilinx.com> wrote:
>>>>>> The current implementation of interleaved DMA API support multiple
>>>>>> frames only when the memory is contiguous by incrementing src_start/
>>>>>> dst_start members of interleaved template.
>>>>>>
>>>>>> But, when the memory is non-contiguous it will restrict slave device
>>>>>> to not submit multiple frames in a batch.  This patch handles this
>>>>>> issue by allowing the slave device to send array of interleaved dma
>>>>>> templates each having a different memory location.
>>>>>>
>>>>> How fragmented could be memory in your case? Is it inefficient to
>>>>> submit separate transfers for each segment/frame?
>>>>> It will help if you could give a typical example (chunk size and gap
>>>>> in bytes) of what you worry about.
>>>>
>>>> With scatter-gather engine feature in the hardware, submitting separate
>>>> transfers for each frame look inefficient. As an example, our DMA engine
>>>> supports up to 16 video frames, with each frame (a typical video frame
>>>> size) being contiguous in memory but frames are scattered into different
>>>> locations. We could not definitely submit frame by frame as it would be
>>>> software overhead (HW interrupting for each frame) resulting in video lags.
>>>>
>>> IIUIC, it is 30fps and one dma interrupt per frame ... it doesn't seem
>>> inefficient at all. Even poor-latency audio would generate a higher
>>> interrupt-rate. So the "inefficiency concern" doesn't seem valid to
>>> me.
>>>
>>> Not to mean we shouldn't strive to reduce the interrupt-rate further.
>>> Another option is to emulate the ring-buffer scheme of ALSA.... which
>>> should be possible since for a session of video playback the frame
>>> buffers' locations wouldn't change.
>>>
>>> Yet another option is to use the full potential of the
>>> interleaved-xfer api as such. It seems you confuse a 'video frame'
>>> with the interleaved-xfer api's 'frame'. They are different.
>>>
>>> Assuming your one video frame is F bytes long and Gk is the gap in
>>> bytes between end of frame [k] and start of frame [k+1] and  Gi != Gj
>>> for i!=j
>>> In the context of interleaved-xfer api, you have just 1 Frame of 16
>>> chunks. Each chunk is Fbytes and the inter-chunk-gap(ICG) is Gk  where
>>> 0<=k<15
>>> So for your use-case .....
>>>   dma_interleaved_template.numf = 1   /* just 1 frame */
>>>   dma_interleaved_template.frame_size = 16  /* containing 16 chunks */
>>>    ...... //other parameters
>>>
>>> You have 3 options to choose from and all should work just as fine.
>>> Otherwise please state your problem in real numbers (video-frames'
>>> size, count & gap in bytes).
>>
>> Initially I interpreted interleaved template the same.  But, Lars corrected me
>> in the subsequent discussion and let me put it here briefly,
>>
>> In the interleaved template, each frame represents a line of size denoted by
>> chunk.size and the stride by icg.  'numf' represent number of frames i.e.
>> number of lines.
>>
>> In video frame context,
>> chunk.size -> hsize
>> chunk.icg -> stride
>> numf -> vsize
>> and frame_size is always 1 as it will have only one chunk in a line.
>>
> But you said in your last post
>   "with each frame (a typical video frame size) being contiguous in memory"
>  ... which is not true from what you write above. Anyways, my first 2
> suggestions still hold.

Yes, each video frame is contiguous and they can be scattered.

>
>> So, the API would not allow to pass multiple frames and we came up with a
>> resolution to pass an array of interleaved template structs to handle this.
>>
> Yeah the API doesn't allow such xfers that don't fall into any
> 'regular expression' of a transfer and also because no controller
> natively supports such xfers -- your controller will break your
> request up into 16 transfers and program them individually, right?

No, it will not program individually.  It has a scatter-gather engine where we
could update the current pointer to the first frame and tail pointer to the last
frame and hardware does the transfer and raise an interrupt when all the frames
are completed (ie. when current reaches tail).

>    BTW if you insist you could still express the 16 video frames as 1
> interleaved-xfer frame with frame_size = (vsize + 1) * 16   ;)
>
> Again, I would suggest you implement ring-buffer type scheme. Say
> prepare 16 interleaved xfer templates and queue them. Upon each
> xfer-done callback (i.e frame rendered), update the data and queue it
> back. It might be much simpler for your actual case. At 30fps, 33ms to
> queue a dma request should _not_ result in any frame-drop.

The driver has similar implementation, where each desc (handling 16 frames)
is pushed to pending queue and then to done queue.  Slave device still can
add desc's to the pending queue and whenever the transfer of 16 frames is
completed we move each desc to done list.

Srikanth

>
> -Jassi
> --
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to majordomo at vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at  http://www.tux.org/lkml/