On Wednesday 16 January 2013 13:36:17 Daniel Vetter wrote:
On Wed, Jan 16, 2013 at 11:01 AM, Thierry Reding wrote:
drm_events_release() should be enough to clean up the events, but I suspect the reason why Laurent put that code in was that the drm_crtc private data still has a reference to the event and needs to clear it. Otherwise the next page flip won't be scheduled because .page_flip() would return -EBUSY.
Hm, indeed we seem to have a nice bug in most drivers there :(
That's not the only reason.
drm_events_release() handles vblank events added to the vblank_event_list. That list only contains vblank events added by drm_queue_vblank_event(), only called by drm_wait_vblank(). Page flip events never get there, so they need to be cleaned up manually.
I wrote in the DRM documentation, in the page flipping section,
"FIXME: Could drivers that don't need to wait for rendering to complete just add the event to dev->vblank_event_list and let the DRM core handle everything, as for "normal" vertical blanking events?"
We should investigate that.
However, it seems like {tegra_dc,shmob_drm_crtc}_cancel_page_flip() could both be simplified a lot and just set their event to NULL. Then again, maybe keeping a separate reference isn't all that useful. Maybe the better thing to do here is iterate over the list of pending VBLANK events in *_finish_page_flip() and process each of them? That would allow more than one user-space process to queue page flips.
I think we need a slightly more generally useful solution, since most drivers are currently broken. I've read a bit through the code, but short of refcounting drm events and adding event->file_priv checks at relevent places I don't see a sane solution ... And even that one is rather invasive. Do you have an idea? Imo doing the cleanup in each driver will be rather error-prone, and since usually kms clients wait for flips to complete, also guaranteed to be little tested.