Re: [PATCHv3 01/12] vb2: add a dev field to use for the default allocation context

[Hans Verkuil <hverkuil@xs4all.nl>]

On 04/24/2016 11:51 PM, Laurent Pinchart wrote:
> Hi Hans,
> 
> On Saturday 23 Apr 2016 12:37:10 Hans Verkuil wrote:
>> On 04/23/2016 02:14 AM, Laurent Pinchart wrote:
>>> On Friday 22 Apr 2016 10:38:08 Hans Verkuil wrote:
>>>> From: Hans Verkuil <hans.verkuil@cisco.com>
>>>>
>>>> The allocation context is nothing more than a per-plane device pointer
>>>> to use when allocating buffers. So just provide a dev pointer in
>>>> vb2_queue for that purpose and drivers can skip
>>>> allocating/releasing/filling in the allocation context unless they
>>>> require different per-plane device pointers as used by some Samsung
>>>> SoCs.
>>>>
>>>> Signed-off-by: Hans Verkuil <hans.verkuil@cisco.com>
>>>> Cc: Laurent Pinchart <Laurent.pinchart@ideasonboard.com>
>>>> Cc: Sakari Ailus <sakari.ailus@iki.fi>
>>>> Cc: Mauro Carvalho Chehab <mchehab@osg.samsung.com>
>>>> Cc: Florian Echtler <floe@butterbrot.org>
>>>> Cc: Federico Vaga <federico.vaga@gmail.com>
>>>> Cc: "Lad, Prabhakar" <prabhakar.csengg@gmail.com>
>>>> Cc: Scott Jiang <scott.jiang.linux@gmail.com>
>>>> Cc: Philipp Zabel <p.zabel@pengutronix.de>
>>>> Cc: Fabien Dessenne <fabien.dessenne@st.com>
>>>> Cc: Benoit Parrot <bparrot@ti.com>
>>>> Cc: Mikhail Ulyanov <mikhail.ulyanov@cogentembedded.com>
>>>> Cc: Guennadi Liakhovetski <g.liakhovetski@gmx.de>
>>>> Cc: Javier Martin <javier.martin@vista-silicon.com>
>>>> Cc: Jonathan Corbet <corbet@lwn.net>
>>>> Cc: Ludovic Desroches <ludovic.desroches@atmel.com>
>>>> Cc: Sergei Shtylyov <sergei.shtylyov@cogentembedded.com>
>>>> Cc: Kyungmin Park <kyungmin.park@samsung.com>
>>>> Cc: Sylwester Nawrocki <s.nawrocki@samsung.com>
>>>> ---
>>>>
>>>>  drivers/media/v4l2-core/videobuf2-core.c | 16 +++++++++-------
>>>>  include/media/videobuf2-core.h           |  3 +++
>>>>  2 files changed, 12 insertions(+), 7 deletions(-)
>>>>
>>>> diff --git a/drivers/media/v4l2-core/videobuf2-core.c
>>>> b/drivers/media/v4l2-core/videobuf2-core.c index 5d016f4..88b5e48 100644
>>>> --- a/drivers/media/v4l2-core/videobuf2-core.c
>>>> +++ b/drivers/media/v4l2-core/videobuf2-core.c
>>>> @@ -206,8 +206,9 @@ static int __vb2_buf_mem_alloc(struct vb2_buffer *vb)
>>>>
>>>>  	for (plane = 0; plane < vb->num_planes; ++plane) {
>>>>  	
>>>>  		unsigned long size = PAGE_ALIGN(vb->planes[plane].length);
>>>>
>>>> -		mem_priv = call_ptr_memop(vb, alloc, q->alloc_ctx[plane],
>>>> -				      size, dma_dir, q->gfp_flags);
>>>> +		mem_priv = call_ptr_memop(vb, alloc,
>>>> +				q->alloc_ctx[plane] ? : &q->dev,
>>>> +				size, dma_dir, q->gfp_flags);
>>>
>>> While the videobuf2-dma-sg allocation context indeed only contains a
>>> pointer to the device, the videobuf2-dma-contig context also contains a
>>> dma_attrs. This patch will break the videobuf2-dma-contig alloc
>>> implementation.
>>
>> Good point. I fixed this in the last patch, but that would mean dma-contig
>> would be broken for the patches in between.
>>
>> I'm moving dma_attrs to struct vb2_queue as the first patch, then the rest
>> will work fine.
> 
> Couldn't a driver require different dma attributes per plane ? Would it make 
> sense to keep the allocation context structure, and use the struct device and 
> dma attributes stored in the queue when no allocation context is provided ?

I kept the dma_attrs part simple for two reasons:

1) No driver in the kernel uses it.
2) I really can't think of any scenario where you get different DMA attrs per plane.
   Perhaps if we make it possible to have a variable number of planes, but all
   that is in the future and I rather take care of it when we actually know what
   we need.

The 'allocation context' idea was simply a bad one: you're stuck with void pointers
(I hate those) and always having to check for ENOMEM when allocating them. When all
you need in almost all cases is just a device pointer.

Regards,

	Hans