Re: [PATCH v3 0/2] Add writeback scaling

From: Daniel Vetter
Date: Tue Apr 23 2019 - 10:02:47 EST


On Tue, Apr 23, 2019 at 3:00 PM Liviu Dudau <Liviu.Dudau@xxxxxxx> wrote:
>
> On Tue, Apr 23, 2019 at 02:24:11PM +0200, Daniel Vetter wrote:
> > On Tue, Apr 23, 2019 at 12:48 PM Ben Davis <Ben.Davis@xxxxxxx> wrote:
> > >
> > > Add support for scaling on writeback. To do this add writeback_dest_w
> > > and writeback_dest_h as writeback connector properties to specify the
> > > desired output dimensions.
> > > Then implement downscaling on writeback for Malidp-550 and Malidp-650
> > > (upscaling on writeback is not supported on these devices).
> > >
> > > v2: Use 0 as default for writeback_w,h and so update range to use 1 as
> > > minimum.
> >
> > Hm I missed that, I don't think that's good, since it prevents
> > userspace from blindly writing back the properties it's read. We've
> > tried hard to avoid that, since we're suggesting compositor can take a
> > snapshot of all kms properties (including the ones they don't
> > understand) and restore that on vt switching. Hence stuff like fence
> > fds returning -1, and accepting -1 as NULL to make this work.
>
> Is that they way it should work, though? I remember a few good years
> back, before KMS, there were lots of issues with X server hanging and
> not restoring the vt mode correctly. Should it not be that on getting
> back control of a DRM master, an application sets it last known good
> mode that it had, before it lost (or gave away) control?

Given that we've recommended both approaches I expect there's a mix of
both approaches out there in the real world, with people just hacking
things up until it works. That's also what we're doing with the
in-kernel fbdev emulation. I'd say defacto this means it's all
ill-defined, but also that we imo should make a best effort to support
either approach.

> > tldr; I think range needs to include 0, and we need make that a
> > special thing, maybe enforced with a
> > drm_connector_state_compute_writeback_dst_h/w, which takes
> > crtc_state->mode.v/hdisplay into account if the writeback_dst_h/w is
> > 0.
>
> Repeating just to make sure we're on the same page: we allow 0 as valid
> range value, but writeback_dst_h/w ultimately gets updated before the
> commit gets passed on to the driver to use crtc_state->mode.v/hdisplay
> so as not to trigger any scaling. Correct?

That would also break stuff, since then you'd scale the next time
v/hdisplay are different. Except if we again clear them back to 0 on
duplicate_state, which I guess would work since writeback is one-shot.
I was thinking more of a helper function, but your approach +
resetting always to 0 on duplicate_state + lots of comments/kerneldoc
should work too. And I guess would be more robust, since driver
authors can't accidentally use the wrong state since it's fixed up.
-Daniel

>
> Best regards,
> Liviu
>
>
> > -Daniel
> >
> > >
> > > v3: Rename properties to specify they are destination width/height.
> > > Make sure the values from the properties are passed to
> > > enable_memwrite rather than the framebuffer dimensions
> > >
> > > Ben Davis (2):
> > > drm: Add writeback_dest_w,h properties
> > > drm/malidp: Enable writeback scaling
> > >
> > > drivers/gpu/drm/arm/malidp_crtc.c | 49 +++++++++++---------
> > > drivers/gpu/drm/arm/malidp_crtc.h | 12 +++++
> > > drivers/gpu/drm/arm/malidp_drv.c | 10 +++--
> > > drivers/gpu/drm/arm/malidp_hw.c | 45 +++++++++++++------
> > > drivers/gpu/drm/arm/malidp_hw.h | 19 ++++++--
> > > drivers/gpu/drm/arm/malidp_mw.c | 94 ++++++++++++++++++++++++++++++---------
> > > drivers/gpu/drm/arm/malidp_regs.h | 1 +
> > > drivers/gpu/drm/drm_atomic_uapi.c | 8 ++++
> > > drivers/gpu/drm/drm_writeback.c | 30 +++++++++++++
> > > include/drm/drm_connector.h | 4 ++
> > > include/drm/drm_mode_config.h | 10 +++++
> > > 11 files changed, 220 insertions(+), 62 deletions(-)
> > > create mode 100644 drivers/gpu/drm/arm/malidp_crtc.h
> > >
> > > --
> > > 2.7.4
> > >
> >
> >
> > --
> > Daniel Vetter
> > Software Engineer, Intel Corporation
> > +41 (0) 79 365 57 48 - http://blog.ffwll.ch
>
> --
> ====================
> | I would like to |
> | fix the world, |
> | but they're not |
> | giving me the |
> \ source code! /
> ---------------
> Â\_(ã)_/Â



--
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch