[PATCH] unstable/linux-dmabuf: Add 'direct_display' flag

[Daniel Stone]

Hi,

On Thu, 14 Nov 2019 at 22:44, Scott Anderson <scott at anderso.nz> wrote:
> On 15/11/19 4:04 am, Drew DeVault wrote:
> > I think that there would be value in being able to suggest that a
> > particluar buffer be scanned out for performance reasons. But, as a
> > suggestion, and not a demand, and definitely not for secrecy reasons.
> > wlroots-based compositors would reserve the right to read your buffers
> > whenever we want. If I want to read your buffer for a frame to take a
> > screenshot, it's not going to be the end of the world for performance
> > and I don't want to end up segfaulting because of it.
> >
> > [...]
> >
> > If this protocol change were for the purpose of _suggesting_ scan-out,
> > I'd be on board with it. But this is not that - if it were, we would
> > dispense with any ideas around enforcing it with memory protections.
> > Then we could get an improvement for performance without a regression
> > for usability.
> >
> > Rather, this just seems to be a DRM-enabling change, and the official
> > policy of wlroots will always be to NACK those for the wp and xdg
> > namespaces.
>
> I don't think a hint for "please scan me out" has any value. A Wayland
> compositor that is capable of scanning out will try to scan out
> everybody as often as they can anyway, making the hint meaningless. A
> Wayland compositor that is not capable of scanning out is also going to
> ignore the hint, also making it meaningless.
>
> The only situation I could possibly think of where this could make a
> difference is if you have two surfaces, and the compositor's policy
> would normally be to scan out the focused surface, but chooses to scan
> out e.g. a video player surface with this flag instead. That only works
> if the video player is not occluded.

Right. ST (the SoC vendor) originally requested this a long time ago -
perhaps a few years even. They run Weston on set-top boxes, and would
like to have the underlying video surface hit overlays at the cost of
GPU composition of the subtitle & UI surfaces. According to them, just
doing the EGLImage import of video buffers was a large enough
performance hit - if I remember correctly, mostly fine but occasional
hitches - that they had to be able to bypass it.

I don't really see any overlap with hints here, given that hints are
server -> client, and also don't prevent the buffer from being
imported into the GPU in the first place.

But yes, as said in the MR, the current usecase we have is a system
which (for good reason) distrusts the GPU with protected content.

I'm fine with just putting this in as a Weston-private extension
layered on top of dmabuf.

> But that's also cheapened by the fact that this flag sounds like "free
> performance", and everybody would just set it.

Well, 'free performance but you're very likely to also display nothing
in the general case', which is a less appealing trade-off.

Cheers,
Daniel