[PATCH v2] dma: Add Xilinx AXI Video Direct Memory Access Engine driver support

Srikanth Thokala sthokal at xilinx.com
Mon Feb 10 07:51:46 EST 2014 

On Thu, Feb 6, 2014 at 9:23 PM, Lars-Peter Clausen <lars at metafoo.de> wrote:
> On 02/06/2014 02:34 PM, Srikanth Thokala wrote:
>>
>> On Wed, Feb 5, 2014 at 10:00 PM, Lars-Peter Clausen <lars at metafoo.de>
>> wrote:
>>>
>>> On 02/05/2014 05:25 PM, Srikanth Thokala wrote:
>>>>
>>>>
>>>> On Fri, Jan 31, 2014 at 12:21 PM, Srikanth Thokala <sthokal at xilinx.com>
>>>> wrote:
>>>>>
>>>>>
>>>>> Hi Vinod,
>>>>>
>>>>> On Tue, Jan 28, 2014 at 8:43 AM, Vinod Koul <vinod.koul at intel.com>
>>>>> wrote:
>>>>>>
>>>>>>
>>>>>> On Mon, Jan 27, 2014 at 06:42:36PM +0530, Srikanth Thokala wrote:
>>>>>>>
>>>>>>>
>>>>>>> Hi Lars/Vinod,
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> The question here i think would be waht this device supports? Is
>>>>>>>>> the
>>>>>>>>> hardware
>>>>>>>>> capable of doing interleaved transfers, then would make sense.
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>> The hardware does 2D transfers. The parameters for a transfer are
>>>>>>>> height,
>>>>>>>> width and stride. That's only a subset of what interleaved transfers
>>>>>>>> can be
>>>>>>>> (xt->num_frames must be one for 2d transfers). But if I remember
>>>>>>>> correctly
>>>>>>>> there has been some discussion on this in the past and the result of
>>>>>>>> that
>>>>>>>> discussion was that using interleaved transfers for 2D transfers is
>>>>>>>> preferred over adding a custom API for 2D transfers.
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> I went through the prep_interleaved_dma API and I see only one
>>>>>>> descriptor
>>>>>>> is prepared per API call (i.e. per frame).  As our IP supports upto
>>>>>>> 16
>>>>>>> frame
>>>>>>> buffers (can be more in future), isn't it less efficient compared to
>>>>>>> the
>>>>>>> prep_slave_sg where we get a single sg list and can prepare all the
>>>>>>> descriptors
>>>>>>> (of non-contiguous buffers) in one go?  Correct me, if am wrong and
>>>>>>> let
>>>>>>> me
>>>>>>> know your opinions.
>>>>>>
>>>>>>
>>>>>> Well the descriptor maybe one, but that can represent multiple frames,
>>>>>> for
>>>>>> example 16 as in your case. Can you read up the documentation of how
>>>>>> multiple
>>>>>> frames are passed. Pls see include/linux/dmaengine.h
>>>>>>
>>>>>> /**
>>>>>>    * Interleaved Transfer Request
>>>>>>    * ----------------------------
>>>>>>    * A chunk is collection of contiguous bytes to be transfered.
>>>>>>    * The gap(in bytes) between two chunks is called
>>>>>> inter-chunk-gap(ICG).
>>>>>>    * ICGs may or maynot change between chunks.
>>>>>>    * A FRAME is the smallest series of contiguous {chunk,icg} pairs,
>>>>>>    *  that when repeated an integral number of times, specifies the
>>>>>> transfer.
>>>>>>    * A transfer template is specification of a Frame, the number of
>>>>>> times
>>>>>>    *  it is to be repeated and other per-transfer attributes.
>>>>>>    *
>>>>>>    * Practically, a client driver would have ready a template for each
>>>>>>    *  type of transfer it is going to need during its lifetime and
>>>>>>    *  set only 'src_start' and 'dst_start' before submitting the
>>>>>> requests.
>>>>>>    *
>>>>>>    *
>>>>>>    *  |      Frame-1        |       Frame-2       | ~ |
>>>>>> Frame-'numf'  |
>>>>>>    *  |====....==.===...=...|====....==.===...=...| ~
>>>>>> |====....==.===...=...|
>>>>>>    *
>>>>>>    *    ==  Chunk size
>>>>>>    *    ... ICG
>>>>>>    */
>>>>>
>>>>>
>>>>>
>>>>> Yes, it can handle multiple frames specified by 'numf' each of size
>>>>> 'frame_size * sgl[0].size'.
>>>>> But, I see it only works if all the frames' memory is contiguous and
>>>>> in this case we
>>>>> can just increment 'src_start' by the total frame size 'numf' number
>>>>> of times to fill in
>>>>> for each HW descriptor (each frame is one HW descriptor).  So, there
>>>>> is no issue when the
>>>>> memory is contiguous.  If the frames are non contiguous, we have to
>>>>> call this API for each
>>>>> frame (hence for each descriptor), as the src_start for each frame is
>>>>> different.  Is it correct?
>>>>>
>>>>> FYI: This hardware has an inbuilt Scatter-Gather engine.
>>>>>
>>>>
>>>> Ping?
>>>
>>>
>>>
>>> If you want to submit multiple frames at once I think you should look at
>>> how
>>> the current dmaengine API can be extended to allow that. And also provide
>>> an
>>> explanation on how this is superior over submitting them one by one.
>>
>>
>>
>> Sure.  I would start with explaning the current implementation of this
>> driver.
>>
>> Using prep_slave_sg(), we can define multiple segments in a
>> async_tx_descriptor where each frame is defined by a segment (a sg
>> list entry).  So, the slave device could DMA the data (of multiple
>> frames) with a descriptor by calling tx_submit in a transaction i.e.,
>>
>> prep_slave_sg(16)  -> tx_submit(1) -> interrupt  (16 frames)
>>
>> Using interleaved_dma(), we could not divide into segments when we
>> have scattered memory (for the reasons mentioned in above thread).
>> This implies we are restricting the slave device to process frame by
>> frame i.e.,
>>
>> interleaved_dma(1) -> tx_submit(1) -> interrupt -> interleaved_dma(2)
>> -> tx_submit (2) -> interrupt -> ........ tx_submit(16) -> interrupt
>>
>
> The API allows you to create and submit multiple interleaved descriptors
> before you have to issue them.
>
> interleaved_dma(1) -> tx_submit(1) -> interleaved_dma(2) -> tx_submit(2) ->
> ... -> issue_pending() -> interrupt
>
>
>> This implementation makes the hardware to wait until the next frame is
>> submitted.
>>
>> To overcome this, I feel it would be a good option if we could extend
>> interleaved_dma template to modify src_start/dest_start to be a
>> pointer to an array of addresses.  Here, number of addresses will be
>> defined by numf. The other option would be to include scatterlist in
>> the interleaved template. This way we can handle scattered memory
>> using this API.
>
>
> Each "frame" in a interleaved transfer describes a single line in your video
> frame (size = width, icg = stride). numf is the number of lines per video
> frame. So the suggested change does not make that much sense. If you want to
> submit multiple video frames in one batch the best option is in my opinion
> to allow to pass an array of dma_interleaved_template structs instead of a
> single one.

Ok.  I agree with you.  Will fix it in v3.

Srikanth

>
> - Lars
>
>>
>> Srikanth
>>
>>>
>>> - Lars
>>>
>>>
>>> --
>>> 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/
>>
>> --
>> To unsubscribe from this list: send the line "unsubscribe dmaengine" in
>>
>> the body of a message to majordomo at vger.kernel.org
>> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>>
>
> --
> 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/

More information about the linux-arm-kernel mailing list