Add wait-for-frame-completion-in-pipewire debug flag for NVIDIA screencasts

4 files changed, 81 insertions, 11 deletions

diff --git a/niri-config/src/lib.rs b/niri-config/src/lib.rs
index ddf1ef00..80bc76d7 100644
--- a/niri-config/src/lib.rs
+++ b/niri-config/src/lib.rs
@@ -2039,6 +2039,8 @@ pub struct DebugConfig {
     #[knuffel(child)]
     pub wait_for_frame_completion_before_queueing: bool,
     #[knuffel(child)]
+    pub wait_for_frame_completion_in_pipewire: bool,
+    #[knuffel(child)]
     pub enable_overlay_planes: bool,
     #[knuffel(child)]
     pub disable_cursor_plane: bool,
@@ -4837,6 +4839,7 @@ mod tests {
                 preview_render: None,
                 dbus_interfaces_in_non_session_instances: false,
                 wait_for_frame_completion_before_queueing: false,
+                wait_for_frame_completion_in_pipewire: false,
                 enable_overlay_planes: false,
                 disable_cursor_plane: false,
                 disable_direct_scanout: false,
diff --git a/src/niri.rs b/src/niri.rs
index 3dc6bf79..e85ec938 100644
--- a/src/niri.rs
+++ b/src/niri.rs
@@ -1630,8 +1630,12 @@ impl State {
 
         match &cast.target {
             CastTarget::Nothing => {
+                let config = self.niri.config.borrow();
+                let wait_for_sync = config.debug.wait_for_frame_completion_in_pipewire;
+                drop(config);
+
                 self.backend.with_primary_renderer(|renderer| {
-                    if cast.dequeue_buffer_and_clear(renderer) {
+                    if cast.dequeue_buffer_and_clear(renderer, wait_for_sync) {
                         cast.last_frame_time = get_monotonic_time();
                     }
                 });
@@ -1671,6 +1675,10 @@ impl State {
                     }
                 }
 
+                let config = self.niri.config.borrow();
+                let wait_for_sync = config.debug.wait_for_frame_completion_in_pipewire;
+                drop(config);
+
                 self.backend.with_primary_renderer(|renderer| {
                     // FIXME: pointer.
                     let elements = mapped
@@ -1678,7 +1686,13 @@ impl State {
                         .rev()
                         .collect::<Vec<_>>();
 
-                    if cast.dequeue_buffer_and_render(renderer, &elements, bbox.size, scale) {
+                    if cast.dequeue_buffer_and_render(
+                        renderer,
+                        &elements,
+                        bbox.size,
+                        scale,
+                        wait_for_sync,
+                    ) {
                         cast.last_frame_time = get_monotonic_time();
                     }
                 });
@@ -4370,6 +4384,10 @@ impl Niri {
 
         let scale = Scale::from(output.current_scale().fractional_scale());
 
+        let config = self.config.borrow();
+        let wait_for_sync = config.debug.wait_for_frame_completion_in_pipewire;
+        drop(config);
+
         let mut elements = None;
         let mut casts_to_stop = vec![];
 
@@ -4401,7 +4419,7 @@ impl Niri {
                 self.render(renderer, output, true, RenderTarget::Screencast)
             });
 
-            if cast.dequeue_buffer_and_render(renderer, elements, size, scale) {
+            if cast.dequeue_buffer_and_render(renderer, elements, size, scale, wait_for_sync) {
                 cast.last_frame_time = target_presentation_time;
             }
         }
@@ -4423,6 +4441,10 @@ impl Niri {
 
         let scale = Scale::from(output.current_scale().fractional_scale());
 
+        let config = self.config.borrow();
+        let wait_for_sync = config.debug.wait_for_frame_completion_in_pipewire;
+        drop(config);
+
         let mut casts_to_stop = vec![];
 
         let mut casts = mem::take(&mut self.casts);
@@ -4461,7 +4483,8 @@ impl Niri {
             // FIXME: pointer.
             let elements: Vec<_> = mapped.render_for_screen_cast(renderer, scale).collect();
 
-            if cast.dequeue_buffer_and_render(renderer, &elements, bbox.size, scale) {
+            if cast.dequeue_buffer_and_render(renderer, &elements, bbox.size, scale, wait_for_sync)
+            {
                 cast.last_frame_time = target_presentation_time;
             }
         }
diff --git a/src/pw_utils.rs b/src/pw_utils.rs
index a518de83..c47fa63f 100644
--- a/src/pw_utils.rs
+++ b/src/pw_utils.rs
@@ -821,6 +821,7 @@ impl Cast {
         elements: &[impl RenderElement<GlesRenderer>],
         size: Size<i32, Physical>,
         scale: Scale<f64>,
+        wait_for_sync: bool,
     ) -> bool {
         let CastState::Ready { damage_tracker, .. } = &mut *self.state.borrow_mut() else {
             error!("cast must be in Ready state to render");
@@ -851,7 +852,7 @@ impl Cast {
         let fd = buffer.datas_mut()[0].as_raw().fd;
         let dmabuf = &self.dmabufs.borrow()[&fd];
 
-        if let Err(err) = render_to_dmabuf(
+        match render_to_dmabuf(
             renderer,
             dmabuf.clone(),
             size,
@@ -859,8 +860,19 @@ impl Cast {
             Transform::Normal,
             elements.iter().rev(),
         ) {
-            warn!("error rendering to dmabuf: {err:?}");
-            return false;
+            Ok(sync_point) => {
+                // FIXME: implement PipeWire explicit sync, and at the very least async wait.
+                if wait_for_sync {
+                    let _span = tracy_client::span!("wait for completion");
+                    if let Err(err) = sync_point.wait() {
+                        warn!("error waiting for pw frame completion: {err:?}");
+                    }
+                }
+            }
+            Err(err) => {
+                warn!("error rendering to dmabuf: {err:?}");
+                return false;
+            }
         }
 
         for (data, (stride, offset)) in
@@ -880,7 +892,11 @@ impl Cast {
         true
     }
 
-    pub fn dequeue_buffer_and_clear(&mut self, renderer: &mut GlesRenderer) -> bool {
+    pub fn dequeue_buffer_and_clear(
+        &mut self,
+        renderer: &mut GlesRenderer,
+        wait_for_sync: bool,
+    ) -> bool {
         // Clear out the damage tracker if we're in Ready state.
         if let CastState::Ready { damage_tracker, .. } = &mut *self.state.borrow_mut() {
             *damage_tracker = None;
@@ -894,9 +910,20 @@ impl Cast {
         let fd = buffer.datas_mut()[0].as_raw().fd;
         let dmabuf = &self.dmabufs.borrow()[&fd];
 
-        if let Err(err) = clear_dmabuf(renderer, dmabuf.clone()) {
-            warn!("error clearing dmabuf: {err:?}");
-            return false;
+        match clear_dmabuf(renderer, dmabuf.clone()) {
+            Ok(sync_point) => {
+                // FIXME: implement PipeWire explicit sync, and at the very least async wait.
+                if wait_for_sync {
+                    let _span = tracy_client::span!("wait for completion");
+                    if let Err(err) = sync_point.wait() {
+                        warn!("error waiting for pw frame completion: {err:?}");
+                    }
+                }
+            }
+            Err(err) => {
+                warn!("error clearing dmabuf: {err:?}");
+                return false;
+            }
         }
 
         for (data, (stride, offset)) in
diff --git a/wiki/Configuration:-Debug-Options.md b/wiki/Configuration:-Debug-Options.md
index e3586b7a..0911ebb1 100644
--- a/wiki/Configuration:-Debug-Options.md
+++ b/wiki/Configuration:-Debug-Options.md
@@ -21,6 +21,7 @@ debug {
     force-pipewire-invalid-modifier
     dbus-interfaces-in-non-session-instances
     wait-for-frame-completion-before-queueing
+    wait-for-frame-completion-in-pipewire
     emulate-zero-presentation-time
     disable-resize-throttling
     disable-transactions
@@ -152,6 +153,22 @@ debug {
 }
 ```
 
+### `wait-for-frame-completion-in-pipewire`
+
+<sup>Since: next release</sup>
+
+Wait until every screencast frame is done rendering before handing it over to PipeWire.
+
+Sometimes helps on NVIDIA to prevent glitched frames when screencasting.
+
+This debug flag will eventually be removed once we handle this properly (via explicit sync in PipeWire).
+
+```kdl
+debug {
+    wait-for-frame-completion-in-pipewire
+}
+```
+
 ### `emulate-zero-presentation-time`
 
 Emulate zero (unknown) presentation time returned from DRM.