[Intel-gfx] [RFC] drm/i915: Add variable gem object size support to i915

[Tvrtko Ursulin]

On 06/25/2014 01:57 PM, Damien Lespiau wrote:
> On Wed, Jun 25, 2014 at 12:46:57PM +0100, Siluvery, Arun wrote:
>> On 25/06/2014 12:14, Damien Lespiau wrote:
>>> On Wed, Jun 25, 2014 at 11:51:33AM +0100, Damien Lespiau wrote:
>>>> (This is not necessarily things one would need to take into account for
>>>> this work, just a few thoughts).
>>>>
>>>> One thing I'm wondering is how fitting the "size" parameter really is
>>>> when talking about inherently 2D buffers.
>>>>
>>>> For instance, let's take a Y-tiled texture with MIPLAYOUT_RIGHT, if we
>>>> want to allocate mip map levels 0 and 1, and use the ioctl "naively" to
>>>> reserve the LOD1 region in one go, we'll end up over allocating the
>>>> space below LOD1 (if I'm not mistaken that is).
>>>>
>>>> This can be mitigated by several calls to this fallocate ioctl, to
>>>> reserve columns of pages (in the case above, columns for the LOD1
>>>> region).
>>>>
>>>> So, how about trying to reduce this ioctl overhead by providing a list
>>>> of (start, length) in the ioctl structure?
>>>
>>> One more thing to factor in is (let's assume one future hardware will
>>> support that):
>>> https://www.opengl.org/registry/specs/ARB/sparse_texture.txt
>>>
>>> So maybe what we really want is to be able to specify region of pages
>>> that could be specified in (x, y, width, height, stride) ? (idea popped
>>> when talking to Neil Roberts (I now have someone working on Mesa in the
>>> office).
>>>
>>
>> Hi Damien,
>>
>> Thank you for your comments and the idea to improve this ioctl.
>> At the moment start, end of a region are expected to be
>> page-aligned; ioctl can be modified to accept a multiple ranges and
>> modify them in one go to reduce the overhead of the ioctl.
>>
>> We can define how we want to specify multiple ranges, if userspace
>> can provide the list as (start, end) pairs kernel can directly use
>> them but what would be the preferred way from the user point of
>> view?
>
> That's a good question to ask a GL team. In the light of sparse
> textures I think the region idea would be better.
>
> We would need to define what the coordinates mean, for instance:
>    - 2D view of the buffer, and the kernel takes care of translating what
>      it means for the underlying pages?
>    - See the buffer object as an array of pages, and those numbers define
>      a region of pages.

This would mean kernel has to know about all possible tiling formats? 
Would that be asking a bit too much (of the kernel)?

How (im)possible would it be to allocate backing store on demand, on 
page by page basis, on write rather than on binding into address space?

Tvrtko