[Intel-gfx] [PATCH 2/3] drm/radeon: Switch to drm_vblank_on/off

[Daniel Vetter]

On Thu, May 28, 2015 at 04:11:53PM +0900, Michel Dänzer wrote:
> On 27.05.2015 18:41, Daniel Vetter wrote:
> > On Wed, May 27, 2015 at 06:21:24PM +0900, Michel Dänzer wrote:
> >> On 27.05.2015 18:04, Daniel Vetter wrote:
> >>> These should be functionally equivalent to the older per/post modeset
> >>> functions, except that they block out drm_vblank_get right away.
> >>> There's only the clock adjusting code (outside of pageflips) in
> >>> readone which uses drm_vblank_get. But that code doesn't synchronize
> >>> against concurrent modesets and instead handles any such races by
> >>> waiting for the right vblank to arrive with a short timetout.
> >>>
> >>> The longer-term plan here is to switch all kms drivers to
> >>> drm_vblank_on/off so that common code like pending event cleanup can
> >>> be done there, while drm_vblank_pre/post_modeset will be purely
> >>> drm internal for the old UMS ioctl.
> >>>
> >>> Note that the kerneldoc for pre/post_modeset is wrong since as Michel
> >>> Dänzer correctly pointed out it works if only using pre/post_modeset.
> >>> The trouble that lead to this comment is the very old version of
> >>> drm_vblank_off to clear out pending events when disabling a pipe,
> >>> which did seem to wreak havoc with the trick used by pre/post_modeset.
> >>> Michel also expressed dissatisfaction with intel folks pushing new
> >>> interfaces with bogus justifications. I still maintain that having a
> >>> consistent set of vblank behaviour across kms drivers, separate from
> >>> any old UMS functions is a useful goal.
> >>>
> >>> Cc: Michel Dänzer <michel.daenzer at amd.com>
> >>> Signed-off-by: Daniel Vetter <daniel.vetter at intel.com>
> >>
> >> Can you describe at least one tangible benefit this change provides for
> >> the radeon driver?
> >>
> >> Because I'm afraid that this might cause subtle breakage, and since we
> >> don't have any rigorous tests for this like in intel-gpu-tools (yet?),
> >> it might be painful to track it down.
> >>
> >> So, I'd like to have a good reason for taking the risk.
> > 
> > right now at most a bit of code to clean out pending events on modeset
> > disable, for somewhat consistent behaviour with other drivers. But in
> > general it's fairly ill-defined what happens with vblank events.
> 
> Yeah, while that's nice to have, I don't think it makes too much
> difference in practice.
> 
> Anyway, I'm giving this patch a spin, and it does indeed cause userspace
> fallout, at least with DRI3/Present enabled, because the vblank and
> pageflip ioctls now return -EINVAL while the CRTC is off. However, it
> looks like fixing that up might not be too bad, so I'm cautiously
> optimistic for this change. But I'd like some more time for testing and
> fixing userspace.

Hm that's a bummer, since it means the abi userspace sees has already
diverged between drivers :(

And because drm_irq.c is still a giant mess with no clear drm core ->
driver interface there's also no way for drivers to overwrite the normal
behaviour to keep old, driver-specific userspace happy. Adding such a hook
for kms drivers is definitely something I want to do. With all the modeset
entry points and atomic helpers you can do that and so keep old userspace
working even if the more strict semantics of atomic would break it (which
was an important goal with all the super-flexible helper library).

Otoh asking for a vblank event on a dead pipe smells like a userspace bug
and could result in stuck compositors. Not sure what's best here really.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch