[Linaro-mm-sig] [RFC v2] dma-buf: Add buffer sharing framework

[Daniel Vetter]

On Wed, Sep 28, 2011 at 13:12, Hans Verkuil <hverkuil at xs4all.nl> wrote:
> On Wednesday, September 28, 2011 12:52:58 Daniel Vetter wrote:
>>   On get_scatterlist walk the attachment list and grab the dma
>>   requirements of all devices (via some platform/dma specific stuff
>>   attached to struct device). Then walk through a list of dma allocators
>>   your platform provides (page_alloc, cma, ...) until you have one that
>>   satisfies the requirements and alloc the backing storage.
>>
>>   Then create an sg_list out of hit and map_sg it, like above.
>
> Hmm. Here is where I run into problems. Right now the get_scatterlist op
> is set by the driver that exports the buffer. So each driver supporting
> dma_buf would have it's own get_scatterlist implementation, each able to
> walk the attachment list. This feels wrong to me. get_scatterlist is doing
> too much and this functionality should probably be split up somehow.

I think we'll end up with a default implementation that just creates a
dma_buf objects and hides all the magic from drivers. But for special
cases (like when using objects mapped into auto-tiling iommus) and
drivers with special infrastructure (drm drivers) I think we should be
able to override this.

Maybe we should add such a thing as an official function to the api.
The x86 implementation should be fairly simple (array of struct page*
for backing storage), on arm it would deal with all the insanity there
is.

> There is also the problem of how to determine whether all requirements
> are satisfied: that's currently left undefined.

I think that should be a platform/architecture problem. And I'm happy
not to be forced to deal with arm ;-) In general, if anything of that
arm-craziness leaks into the interface, that's bad, so I regard your
question here as a good sign ...

>>   For arm we probably need some helper functions to make this easier.
>>
>> - Sharing of a special buffer object, like a tiled buffer remapping into
>>   omap's TILER. attach would check whether the device can actually access
>>   that special io area (iirc on omap only certain devices can access the
>>   tiler). get_scatterlist would just create a 1-entry sg_list that points
>>   directly to the pyhsical address (in device space) of the remapped
>>   buffer. This is way I want get_scatterlist to return a mapped buffer
>>   (and for api cleanliness).
>
> OK.
>
> I think I would like to see a demo implementation and proper documentation.
> Both help a lot to bring the API into focus and identify what works and what
> doesn't. If it's hard to explain or to implement, then we probably should
> improve that part :-)

Yeah, definitely. I think it'll take a while for the
arm-insanity-induced infrastructure to shape up unfortunately, but
maybe I can hack something together on x86 with i915 as a proof of
concept.

Cheers, Daniel
-- 
Daniel Vetter
daniel.vetter at ffwll.ch - +41 (0) 79 364 57 48 - http://blog.ffwll.ch