[Mesa-dev] [PATCH v2 1/2] egl/wayland: Check queryImage return for wl_buffer

[Marek Olšák]

On Mon, Oct 9, 2017 at 5:22 PM, Daniel Stone <daniel at fooishbar.org> wrote:
> Hi Marek,
>
> On 4 October 2017 at 12:38, Marek Olšák <maraeo at gmail.com> wrote:
>> Yeah I don't have an answer to that. I think it's better to pageflip than
>> blit. Can Wayland reallocate the buffer to make it displayable? If not, it
>> may be better to create displayable buffers always.
>
> Not sure what you mean by 'Wayland reallocate'? The buffers are
> (DRI3-style) allocated by the client's EGLSurface, with no
> intermediate server-allocated storage at any point There is no
> explicit hint to the client whether it is being scanned out or not.
> Even if there was, there is no indication of which device the buffer
> might be scanned out on - a device the client might not have access to
> (e.g. a display controller with no render nodes). I've mostly been
> waiting for the allocator to come around before trying to deal with
> this. But even with that, we'd still require the tiling information to
> actually be exposed externally (i.e. through modifiers); currently the
> compositor can't introspect the tiling flags at all, so it has no idea
> that it shouldn't try to scan this particular buffer out.
>
> Which is exactly the problem validateUsage() is supposed to solve. All
> we can do is rely on gbm_bo_import failing (which is best, as it can
> never be used for scanout), or at least the KMS commit fail (worse,
> especially pre-atomic, but still beats displaying garbage?). We don't
> need magic blits, or untiled views, or anything: just failing is fine
> and we'll fall back to composition.

I prefer page flipping over composition. With X/DRI, the DRI state
tracker (or maybe libGL/libEGL) always sets the scanout flag.

Starting with our Raven APU, the scanout and non-scanout tiling mode
is the same for 32bpp, so there won't be any compatibility issue.

Marek