Re: [PATCH] drm: add fb max width/height fields to drm_mode_config

From: Ville Syrjälä
Date: Thu Oct 03 2019 - 06:27:23 EST


On Wed, Oct 02, 2019 at 03:55:10PM -0400, Rob Clark wrote:
> On Wed, Oct 2, 2019 at 9:45 AM Ville Syrjälä
> <ville.syrjala@xxxxxxxxxxxxxxx> wrote:
> >
> > On Tue, Oct 01, 2019 at 02:20:55PM -0700, Jeykumar Sankaran wrote:
> > > On 2019-09-30 03:39, Ville Syrjälä wrote:
> > > > On Fri, Sep 27, 2019 at 06:28:51PM -0700, Jeykumar Sankaran wrote:
> > > >> The mode_config max width/height values determine the maximum
> > > >> resolution the pixel reader can handle.
> > > >
> > > > Not according to the docs I "fixed" a while ago.
> > > >
> > > >> But the same values are
> > > >> used to restrict the size of the framebuffer creation. Hardware's
> > > >> with scaling blocks can operate on framebuffers larger/smaller than
> > > >> that of the pixel reader resolutions by scaling them down/up before
> > > >> rendering.
> > > >>
> > > >> This changes adds a separate framebuffer max width/height fields
> > > >> in drm_mode_config to allow vendors to set if they are different
> > > >> than that of the default max resolution values.
> > > >
> > > > If you're going to change the meaning of the old values you need
> > > > to fix the drivers too.
> > > >
> > > > Personally I don't see too much point in this since you most likely
> > > > want to validate all the other timings as well, and so likely need
> > > > some kind of mode_valid implementation anyway. Hence to validate
> > > > modes there's not much benefit of having global min/max values.
> > > >
> > > https://patchwork.kernel.org/patch/10467155/
> > >
> > > I believe you are referring to this patch.
> > >
> > > I am primarily interested in the scaling scenario mentioned here. MSM
> > > and a few other hardware have scaling block that are used both ways:
> > >
> > > 1) Where FB limits are larger than the display limits. Scalar blocks are
> > > used to
> > > downscale the framebuffers and render within display limits.
> > >
> > > In this scenario, with your patch, are you suggesting the drivers
> > > maintain the
> > > display limits locally and use those values in fill_modes() /
> > > mode_valid() to filter
> > > out invalid modes explicitly instead of mode_config.max_width/height?
> > >
> > > 2) Where FB limits are smaller than display limits. Enforced for
> > > performance reasons on low tier hardware.
> > > It reduces the fetch bandwidth and uses post blending scalar block to
> > > scale up the pixel stream
> > > to match the display resolution.
> >
> > As Daniel mentioned in that discussion your typical userspace
> > assumes that it can use a single unscaled framebuffer with any
> > advertised mode. Hence I believe limiting the mode list based
> > on the max framebuffer size is pretty much required unless
> > you want to break existing userspace.
> >
> > In i915 I went a bit further than that recently and now we
> > filter the mode list based on the maximum plane size [1]
> > (which can be less than the max fb size and less than the
> > maximum crtc dimensions). And again that's because userspace
> > assumes that it can just use a single unscaled fullscreen
> > plane to cover the entire crtc.
> >
> > These assumption are also carved into the legacy setcrtc uapi
> > where you can't even specify multiple framebuffers. In theory
> > a driver could internally use multiple planes to overcome some
> > of the limitations, but in i915 at least we don't.
> >
> > [1] https://cgit.freedesktop.org/drm/drm-intel/commit/?id=2d20411e25a3bf3d2914a2219f47ed48dc57aed5
> >
> > >
> > > Any suggestions on how this topology can be handled with a single set of
> > > max/min values?
> > >
> >
> > I think a safe way to relax these rules would be to either:
> > a) Add a client cap by which userspace can inform the kernel
> > it understands there are more complicated limits at play
> > and thus can't assume that everything will just work
> > b) Maybe we could just tie that in with the atomic cap since
> > atomic clients are pretty much required to do the TEST_ONLY
> > dance anyway, so one might hope they have a working fallback
> > strategy. Though I suspect eg. the modesetting ddx wouldn't
> > like this. But we no longer allow atomic with X anyway so
> > that partcular argument may not hold much weight anymore.
>
> What was the conclusion of the hack to not expose atomic to
> modesetting ddx, due to the brokenness of it's atomic use? I guess
> that could also make the modesetting case go away..

I thought it went in? Maybe I'm mistaken.

--
Ville Syrjälä
Intel