Re: [PATCH V1] dmaengine: tegra: add dma driver

[Laxman Dewangan <ldewangan@nvidia.com>]

Hi Vinod,

On Monday 23 April 2012 06:47 PM, Laxman Dewangan wrote:
> On Monday 23 April 2012 06:36 PM, Russell King - ARM Linux wrote:
>> On Mon, Apr 23, 2012 at 05:47:03PM +0530, Laxman Dewangan wrote:
>>> Hi Russell,
>>> On Monday 23 April 2012 02:11 PM, Vinod Koul wrote:
>>>> On Fri, 2012-04-20 at 17:46 +0530, Laxman Dewangan wrote:
>>>>> Thanks Vinod for quick review.
>>>> Since I was on vacation, I hadn't noticed Russell has already sent the
>>>> patches for omap dma support.
>>>> http://permalink.gmane.org/gmane.linux.ports.arm.omap/75034
>>>>
>>>> It would be nice if both the efforts are coordinated.
>>>>
>>>> Btw I like the virtual channel support introduced by Russell
>>>>
>>> Can you please point me the virtual channel related change? I am not
>>> able to locate this like search for function vchan_* ().
>>> My driver is also on same line but not used vchan_* and also having
>>> support for cyclic transfer.
>> It's only been posted in RFC form on linux-arm-kernel and linux-omap
>> lists.  The specific patch is:
>>
>> http://lists.arm.linux.org.uk/lurker/message/20120418.101116.082b350f.en.html
>>
>> I wouldn't call it perfected yet, but usable.  It doesn't have any
>> knowledge about cyclic transfers either.
> For simple dma, it is straight to use the virt_chan and it reduce lots
> of code from tegra_dma as most of it moved to the virt_dma.
>
> Some points which I am looking are:
> 1. Extending this for cyclic support:
> In cyclic mode, we need to call callback after period_len but do not
> want to free descriptors. So either I need to add flag on the desctiptor
> to no delet and so when vc->desc_free(vd); is called from callback, it
> will not delete the descriptor.
>
> 2. With very prep call, we are allocating descriptor. Is it is possible
> to allocate some desc in advance and then keep using them. The
> complexity is that if we allocate the desc in advance, we need to
> allocate the desc and sq_req list and maintain the two different lists
> as we dont know the sg_len in advance.
>
> 3. vchan_cookie_complete()  is not possible in the cyclic mode as we
> dont want to call dma_cookie_complete() but just want to do following
> two thing:
>
> list_add_tail(&vd->node,&vc->desc_completed);
> tasklet_schedule(&vc->task);
>
> if we extend this function to bypass dma_cookie_complete(&vd->tx); or
> rather than calling this api, directly call the above apis.
>


I had a  communication with Russell on another change thread and 
understand that some more work need to be done in virt_chan to support 
cyclic one.

I want to have the cyclic dma support in my driver so I want to go on 
following steps:
1. I will post the patch for tegra dma which will not use the virt_dma 
so that my driver will be independent of Russel's change.
2. Once the tegra dma is part of tree, I can move all my dma client to 
use the dmaengine based driver and remove old style dma.
3. Till that time Russell's will have already virt_dma support for 
cyclic one and so I will change tegra_dma.c to use virt_dma.


Let me know your opinion so that I can plan my patch/change accordingly.

Thanks,
Laxman