On 11.10.12 16:19, Laurent Pinchart wrote:
Hi Rob,
Thanks for the patch.
On Monday 08 October 2012 14:50:39 Rob Clark wrote:
From: Rob Clark rob@ti.com
...
Do you know why some drivers don't call drm_vblank_count_and_time() ? For instance nouveau sets the sequence to 0 and uses do_gettimeofday(), but it looks like it could just call drm_vblank_count_and_time().
At least nouveau could use it. Lucas Stach and me wrote patches for nouveau-kms, and they went through many iterations and missed many kernel merge windows due to slow review until i think both of us got tired of resubmitting with tiny changes. The latest iteration is posted by Lucas on nouveau-devel from 26. April 2012. Not sure if they'd still apply after the nouveau-kms rewrite. I'll probably give them another try once that has landed when i have some spare time.
In principle it's very simple to use drm_vblank_count_and_time(). A driver needs to
1. Call drm_handle_vblank() from its vblank irq handler.
2. Make sure that in the vblank of pageflip completion drm_handle_vblank() is called before drm_vblank_count_and_time(), so the latter picks up updated counts and timestamps.
3. Big bonus for high precision and robustness: Implement the driver->get_vblank_timestamp() hook to provide a precise vblank timestamp. One simple way to do that is like radeon-kms or intel-kms do it: Call back into drm_calc_vbltimestamp_from_scanoutpos() and provide the driver->get_scanout_position() function - a function that returns the current hardware scanline counter. This is precise down to ~ 10 microseconds (at least confirmed by measurements on intel,radeon,nouveau) and robust against delayed vblank irq handling.
-mario