delayed blocker + syncobj + fifo + commit timing #1449
Conversation
| // Expose syncobj protocol if supported by primary GPU | ||
| if supports_syncobj_eventfd(&device_fd) { | ||
| let syncobj_state = DrmSyncobjState::new::<State>(&niri.display_handle, device_fd); | ||
| assert!(self.syncobj_state.replace(syncobj_state).is_none()); |
There was a problem hiding this comment.
Guess this should be removed once the primary node is removed?
| let mut min_next_schedule: Option<Duration> = None; | ||
|
|
||
| for (output, target_presentation_time, refresh_interval) in queued_outputs { | ||
| let next_deadline = self.signal_commit_timing(&output, target_presentation_time); |
There was a problem hiding this comment.
Should this happen at the start of refresh() together with the call to notify_blocker_cleared()? As I understand, this signalling clears blockers for the surfaces that now became ready, which means it should happen earlier, so that the commits are considered by layout.refresh().
| let output = output.clone(); | ||
| self.redraw(backend, &output); | ||
| } | ||
| pub fn queued_outputs(&self) -> Vec<(Output, Duration, Option<Duration>)> { |
There was a problem hiding this comment.
I guess this is not exactly queued_outputs() since it doesn't check if the output is queued for redraw?
| if self.niri.is_queued(&output) { | ||
| self.niri | ||
| .redraw(&mut self.backend, &output, target_presentation_time); | ||
| self.signal_fifo(&output); |
There was a problem hiding this comment.
I'm not sure I entirely understand how fifo is supposed/intended to work. In the protocol:
This means that when the content update containing "set_barrier" was made active at a latching deadline, it will be active for at least one refresh cycle.
What happens if we redraw and signal fifo here, but the redraw doesn't actually submit a buffer (e.g. the damage was outside the output rect), then something else commits right away and we redraw and signal again, before one refresh cycle had passed? Doesn't this go against this line in the protocol?
| if self.niri.is_queued(&output) { | ||
| self.niri | ||
| .redraw(&mut self.backend, &output, target_presentation_time); | ||
| self.signal_fifo(&output); |
There was a problem hiding this comment.
Same question here about clearing commit blockers at the end here, after layout.refresh(). Though this one is trickier because I understand it has to happen after a redraw which itself has to happen at the end.
|
When is it going to be implemented. |
|
How exactly is this a blocker? Many people use niri on NVIDIA just fine. |
|
It just flickers, or so I've heard. NVIDIA doesn’t support implicit sync. |
|
You've heard wrong. The screencast flickers are unrelated to this PR. Nvidia has a workaround for lack of explicit sync since long ago and it works fine. |
3 participants
At this point just a poc, which obviously needs a major cleanup, showcasing and tracking Smithay/smithay#1720