[Intel-gfx] [PATCH 1/3] drm/vblank: Introduce drm_crtc_vblank_get_accurate()

[Daniel Vetter]

On Mon, Jul 03, 2017 at 05:26:24PM +0300, Ville Syrjälä wrote:
> On Fri, Jun 30, 2017 at 08:18:19PM +0200, Daniel Vetter wrote:
> > On Fri, Jun 30, 2017 at 5:26 PM, Daniel Vetter <daniel at ffwll.ch> wrote:
> > > On Fri, Jun 30, 2017 at 05:18:41PM +0300, ville.syrjala at linux.intel.com wrote:
> > >> From: Ville Syrjälä <ville.syrjala at linux.intel.com>
> > >>
> > >> Add drm_crtc_vblank_get_accurate() which is the same as
> > >> drm_crtc_vblank_get() except that it will forcefully update the
> > >> the current vblank counter/timestamp value even if the interrupt
> > >> is already enabled.
> > >>
> > >> And we'll switch drm_wait_one_vblank() over to using it to make sure it
> > >> will really wait at least one vblank even if it races with the irq
> > >> handler.
> > >>
> > >> Cc: Daniel Vetter <daniel at ffwll.ch>
> > >> Signed-off-by: Ville Syrjälä <ville.syrjala at linux.intel.com>
> > >
> > > Hm, I just thought of doing an
> > > s/drm_vblank_count/drm_crtc_accurate_vblank_count/ in
> > > drm_crtc_arm_vblank_event.
> > >
> > > What does this buy us above&beyond the other patch? And why isn't
> > > vblank_get accurate always?
> > 
> > This also seems to have the risk of not fixing the arm_vblank_event
> > race if someone puts the vblank_get_accurate outside of the critical
> > section. Then we're back to the same old race. And since that means
> > you need to have a vblank_get_accurate right before the
> > arm_vblank_event in all cases, we might as well just put it into the
> > helper. You wrote in the other thread putting it into arm_vblank_event
> > might be wasteful, but I don't see any scenario where that could be
> > the case ...
> > 
> > Or do I miss something?
> 
> The interrupt most likely will be on already when pipe_update_end() gets
> called since pipe_update_start() will enable it already, and Chris's
> disable_immediate optimization will keep it on until the next interrupt.
> Prior to that optimization the drm_vblank_get() in pipe_update_end()
> would have had to turn on the interrupt and perform the vblank counter
> update, and thus the second update from drm_crtc_accurate_vblank_count()
> would have been wasteful.
> 
> I'm not sure I like relying on the fact that pipe_update_start() already
> turned the interrupt on and it's going to be kept on magically. One
> solution for that would be to hang on to the reference from
> pipe_update_start() until we've armed the event. Then we know the
> vblank_get() won't have to enable the interrupt.

Yeah, relying on some implict vblank_get happening in the right way, far
away from the arm_vblank_event, is the part I don't like. Makes review
harder, and I don't see the gain we'll get in return.

> However, if we start to sample the counter for some other purpose
> (watermarks, fb unpinning etc.) during pipe_update_end() then we'll
> again be faced with another potentially pointless update if we
> decide to use drm_crtc_accurate_vblank_count() again.

I think at that point we should write our own arm_vblank_event and cache
the accurate vblank ts within i915 code. My point is that for a generic
helper, we should aim for safe over fastest option. Having a special
vblank_get that you have to call before you call arm_vblank_event, and
within the critical section, just adds to the list of caveats of that
function. Simply making arm_vblank_event more robust seems like the much
better plan (and we can always fix things in i915 if the overhead is too
much).
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch