drm_gem_get_pages and proper flushing/coherency

[Oleksandr Andrushchenko]

+ARM mailing list for help and suggestions

Dear ARM community!

Xen hypervizor para-virtualized DRM frontend driver [1] uses shmem
backed pages for display buffers and shares those with the host
driver. Everything works just fine, but in some scenarios I see
artifacts on the screen which are because those shmem pages of the
buffer are not flushed (please see the mail below for more
details from DRM point of view).
For x86 the flushing of the pages can be done with DRM helper [2]
and seem to help. But clflushopt, which is used there, is not defined
for ARM. I had a suggestion to use set_pages_array_*(),
but again it is only defined for x86, not ARM [3].

The implementation/workaround that I have [4] is based on the DMA
approach, e.g. I map shmem pages and it seem to help, but the whole
DMA approach here seems to be an overkill for that.

Now to the question: could anyone on ARM community help understand the
right way to flush those pages, so I don't need to use DMA for that?

Thank you,
Oleksandr

On 11/26/18 2:15 PM, Oleksandr Andrushchenko wrote:
> Hello, all!
>
> My driver (Xen para-virtualized frontend) in some scenarios uses
> drm_gem_get_pages for allocating backing storage for dumb buffers.
> There are use-cases which showed some artifacts on the screen
> (modetest, other) which were worked around by flushing pages of the
> buffer on page flip with drm_clflush_pages. But, the problem here
> is that drm_clflush_pages is not available on ARM platforms (it is a NOP)
> and doing flushes on every page flip seems to be non-optimal.
>
> Other drivers that use drm_gem_get_pages seem to use DMA map/unmap
> on the shmem backed buffer (this is from where drm_gem_get_pages
> allocates the pages) and this is an obvious approach as the buffer needs
> to be shared with real HW for DMA - please correct me if my understanding
> here is wrong.
>
> This is the part I missed in my implementation as I don't really have a
> HW device which needs DMA, but a backend running in a different Xen 
> domain.
>
> Thus, as the buffer is backed with cachable pages the backend may see
>
> artifacts on its side.
>
>
> I am looking for some advices on what would be the best option to
> make sure dumb buffers are not flushed every page flip and still
> the memory remains coherent to the backend. I have implemented a
> DMA map/unmap of the shmem pages on GEM object creation/destruction
> and this does solve the problem, but as the backend is not really
> a DMA device this is a bit misleading.
>
> Is there any other (more?) suitable/preferable way(s) to achieve the 
> same?
>
> Thank you,
> Oleksandr
>
[1] https://elixir.bootlin.com/linux/v4.20-rc7/source/drivers/gpu/drm/xen
[2] 
https://elixir.bootlin.com/linux/v4.20-rc7/source/drivers/gpu/drm/drm_cache.c#L45
[3] 
https://elixir.bootlin.com/linux/v4.20-rc7/source/arch/x86/include/asm/set_memory.h
[4] https://patchwork.kernel.org/patch/10700089/