clutter/frame-clock: Optimize latency for platforms missing TIMESTAMP_QUERY

Previously if we had no measurements then `compute_max_render_time_us`
would pessimise its answer to ensure triple buffering could be reached:
```
if (frame_clock->state == CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_ONE)
  ret += refresh_interval_us;
```
But that also meant entering triple buffering even when not required.

Now we make `compute_max_render_time_us` more honest and return failure
if the answer isn't known (or is disabled). This in turn allows us to
optimize `calculate_next_update_time_us` for this special case, ensuring
triple buffering can be used, but isn't blindly always used.

This makes a visible difference to the latency when dragging windows in
Xorg, but will also help Wayland sessions on platforms lacking
TIMESTAMP_QUERY such as Raspberry Pi.

Signed-off-by: Mingi Sung <sungmg@saltyming.net>

clutter/clutter/clutter-frame-clock.c

@@ -42,6 +42,15 @@ enum
 
 static guint signals[N_SIGNALS];
 
+typedef enum
+{
+  TRIPLE_BUFFERING_MODE_NEVER,
+  TRIPLE_BUFFERING_MODE_AUTO,
+  TRIPLE_BUFFERING_MODE_ALWAYS,
+} TripleBufferingMode;
+
+static TripleBufferingMode triple_buffering_mode = TRIPLE_BUFFERING_MODE_AUTO;
+
 #define SYNC_DELAY_FALLBACK_FRACTION 0.875f
 
 #define MINIMUM_REFRESH_RATE 30.f
@@ -70,7 +79,10 @@ typedef enum _ClutterFrameClockState
   CLUTTER_FRAME_CLOCK_STATE_IDLE,
   CLUTTER_FRAME_CLOCK_STATE_SCHEDULED,
   CLUTTER_FRAME_CLOCK_STATE_SCHEDULED_NOW,
-  CLUTTER_FRAME_CLOCK_STATE_DISPATCHED,
+  CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_ONE,
+  CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_ONE_AND_SCHEDULED,
+  CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_ONE_AND_SCHEDULED_NOW,
+  CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_TWO,
 } ClutterFrameClockState;
 
 struct _ClutterFrameClock
@@ -91,6 +103,7 @@ struct _ClutterFrameClock
   ClutterFrameClockMode mode;
 
   int64_t last_dispatch_time_us;
+  int64_t prev_last_dispatch_time_us;
   int64_t last_dispatch_lateness_us;
   int64_t last_presentation_time_us;
   int64_t next_update_time_us;
@@ -112,6 +125,9 @@ struct _ClutterFrameClock
   int64_t vblank_duration_us;
   /* Last KMS buffer submission time. */
   int64_t last_flip_time_us;
+  int64_t prev_last_flip_time_us;
+
+  ClutterFrameHint last_flip_hints;
 
   /* Last time we promoted short-term maximum to long-term one */
   int64_t longterm_promotion_us;
@@ -248,10 +264,6 @@ static void
 maybe_update_longterm_max_duration_us (ClutterFrameClock *frame_clock,
                                        ClutterFrameInfo  *frame_info)
 {
-  /* Do not update long-term max if there has been no measurement */
-  if (!frame_clock->shortterm_max_update_duration_us)
-    return;
-
   if ((frame_info->presentation_time - frame_clock->longterm_promotion_us) <
       G_USEC_PER_SEC)
     return;
@@ -278,6 +290,12 @@ void
 clutter_frame_clock_notify_presented (ClutterFrameClock *frame_clock,
                                       ClutterFrameInfo  *frame_info)
 {
+#ifdef CLUTTER_ENABLE_DEBUG
+  const char *debug_state =
+    frame_clock->state == CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_TWO ?
+    "Triple buffering" : "Double buffering";
+#endif
+
   COGL_TRACE_BEGIN_SCOPED (ClutterFrameClockNotifyPresented,
                            "Clutter::FrameClock::presented()");
   COGL_TRACE_DESCRIBE (ClutterFrameClockNotifyPresented,
@@ -367,22 +385,54 @@ clutter_frame_clock_notify_presented (ClutterFrameClock *frame_clock,
 
   frame_clock->got_measurements_last_frame = FALSE;
 
-  if (frame_info->cpu_time_before_buffer_swap_us != 0 &&
-      frame_info->has_valid_gpu_rendering_duration)
+  if ((frame_info->cpu_time_before_buffer_swap_us != 0 &&
+       frame_info->has_valid_gpu_rendering_duration) ||
+      frame_clock->ever_got_measurements)
     {
       int64_t dispatch_to_swap_us, swap_to_rendering_done_us, swap_to_flip_us;
+      int64_t dispatch_time_us = 0, flip_time_us = 0;
 
-      dispatch_to_swap_us =
-        frame_info->cpu_time_before_buffer_swap_us -
-        frame_clock->last_dispatch_time_us;
+      switch (frame_clock->state)
+        {
+        case CLUTTER_FRAME_CLOCK_STATE_INIT:
+        case CLUTTER_FRAME_CLOCK_STATE_IDLE:
+        case CLUTTER_FRAME_CLOCK_STATE_SCHEDULED:
+        case CLUTTER_FRAME_CLOCK_STATE_SCHEDULED_NOW:
+          g_warn_if_reached ();
+          G_GNUC_FALLTHROUGH;
+        case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_ONE:
+        case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_ONE_AND_SCHEDULED:
+        case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_ONE_AND_SCHEDULED_NOW:
+          dispatch_time_us = frame_clock->last_dispatch_time_us;
+          flip_time_us = frame_clock->last_flip_time_us;
+          break;
+        case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_TWO:
+          dispatch_time_us = frame_clock->prev_last_dispatch_time_us;
+          flip_time_us = frame_clock->prev_last_flip_time_us;
+          break;
+        }
+
+      if (frame_info->cpu_time_before_buffer_swap_us == 0)
+        {
+          /* User thread cursor-only updates with no "swap": we do know
+           * the combined time from dispatch to flip at least.
+           */
+          dispatch_to_swap_us = 0;
+          swap_to_flip_us = flip_time_us - dispatch_time_us;
+        }
+      else
+        {
+          dispatch_to_swap_us = frame_info->cpu_time_before_buffer_swap_us -
+                                dispatch_time_us;
+          swap_to_flip_us = flip_time_us -
+                            frame_info->cpu_time_before_buffer_swap_us;
+        }
       swap_to_rendering_done_us =
         frame_info->gpu_rendering_duration_ns / 1000;
-      swap_to_flip_us =
-        frame_clock->last_flip_time_us -
-        frame_info->cpu_time_before_buffer_swap_us;
 
       CLUTTER_NOTE (FRAME_TIMINGS,
-                    "update2dispatch %ld µs, dispatch2swap %ld µs, swap2render %ld µs, swap2flip %ld µs",
+                    "%s: update2dispatch %ld µs, dispatch2swap %ld µs, swap2render %ld µs, swap2flip %ld µs",
+                    debug_state,
                     frame_clock->last_dispatch_lateness_us,
                     dispatch_to_swap_us,
                     swap_to_rendering_done_us,
@@ -393,7 +443,7 @@ clutter_frame_clock_notify_presented (ClutterFrameClock *frame_clock,
                MAX (swap_to_rendering_done_us, swap_to_flip_us) +
                frame_clock->deadline_evasion_us,
                frame_clock->shortterm_max_update_duration_us,
-               frame_clock->refresh_interval_us);
+               2 * frame_clock->refresh_interval_us);
 
       maybe_update_longterm_max_duration_us (frame_clock, frame_info);
 
@@ -402,7 +452,8 @@ clutter_frame_clock_notify_presented (ClutterFrameClock *frame_clock,
     }
   else
     {
-      CLUTTER_NOTE (FRAME_TIMINGS, "update2dispatch %ld µs",
+      CLUTTER_NOTE (FRAME_TIMINGS, "%s: update2dispatch %ld µs",
+                    debug_state,
                     frame_clock->last_dispatch_lateness_us);
     }
 
@@ -420,10 +471,22 @@ clutter_frame_clock_notify_presented (ClutterFrameClock *frame_clock,
     case CLUTTER_FRAME_CLOCK_STATE_SCHEDULED_NOW:
       g_warn_if_reached ();
       break;
-    case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED:
+    case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_ONE:
       frame_clock->state = CLUTTER_FRAME_CLOCK_STATE_IDLE;
       maybe_reschedule_update (frame_clock);
       break;
+    case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_ONE_AND_SCHEDULED:
+      frame_clock->state = CLUTTER_FRAME_CLOCK_STATE_SCHEDULED;
+      maybe_reschedule_update (frame_clock);
+      break;
+    case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_ONE_AND_SCHEDULED_NOW:
+      frame_clock->state = CLUTTER_FRAME_CLOCK_STATE_SCHEDULED_NOW;
+      maybe_reschedule_update (frame_clock);
+      break;
+    case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_TWO:
+      frame_clock->state = CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_ONE;
+      maybe_reschedule_update (frame_clock);
+      break;
     }
 }
 
@@ -441,25 +504,37 @@ clutter_frame_clock_notify_ready (ClutterFrameClock *frame_clock)
     case CLUTTER_FRAME_CLOCK_STATE_SCHEDULED_NOW:
       g_warn_if_reached ();
       break;
-    case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED:
+    case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_ONE:
       frame_clock->state = CLUTTER_FRAME_CLOCK_STATE_IDLE;
       maybe_reschedule_update (frame_clock);
       break;
+    case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_ONE_AND_SCHEDULED:
+      frame_clock->state = CLUTTER_FRAME_CLOCK_STATE_SCHEDULED;
+      maybe_reschedule_update (frame_clock);
+      break;
+    case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_ONE_AND_SCHEDULED_NOW:
+      frame_clock->state = CLUTTER_FRAME_CLOCK_STATE_SCHEDULED_NOW;
+      maybe_reschedule_update (frame_clock);
+      break;
+    case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_TWO:
+      frame_clock->state = CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_ONE;
+      maybe_reschedule_update (frame_clock);
+      break;
     }
 }
 
-static int64_t
-clutter_frame_clock_compute_max_render_time_us (ClutterFrameClock *frame_clock)
+static gboolean
+clutter_frame_clock_compute_max_render_time_us (ClutterFrameClock *frame_clock,
+                                                int64_t           *max_render_time_us)
 {
   int64_t refresh_interval_us;
-  int64_t max_render_time_us;
 
   refresh_interval_us = frame_clock->refresh_interval_us;
 
   if (!frame_clock->ever_got_measurements ||
       G_UNLIKELY (clutter_paint_debug_flags &
                   CLUTTER_DEBUG_DISABLE_DYNAMIC_MAX_RENDER_TIME))
-    return (int64_t) (refresh_interval_us * SYNC_DELAY_FALLBACK_FRACTION);
+    return FALSE;
 
   /* Max render time shows how early the frame clock needs to be dispatched
    * to make it to the predicted next presentation time. It is an estimate of
@@ -473,15 +548,15 @@ clutter_frame_clock_compute_max_render_time_us (ClutterFrameClock *frame_clock)
    * - The duration of vertical blank.
    * - A constant to account for variations in the above estimates.
    */
-  max_render_time_us =
+  *max_render_time_us =
     MAX (frame_clock->longterm_max_update_duration_us,
          frame_clock->shortterm_max_update_duration_us) +
     frame_clock->vblank_duration_us +
     clutter_max_render_time_constant_us;
 
-  max_render_time_us = CLAMP (max_render_time_us, 0, refresh_interval_us);
+  *max_render_time_us = CLAMP (*max_render_time_us, 0, 2 * refresh_interval_us);
 
-  return max_render_time_us;
+  return TRUE;
 }
 
 static void
@@ -496,7 +571,9 @@ calculate_next_update_time_us (ClutterFrameClock *frame_clock,
   int64_t min_render_time_allowed_us;
   int64_t max_render_time_allowed_us;
   int64_t next_presentation_time_us;
+  int64_t next_smooth_presentation_time_us = 0;
   int64_t next_update_time_us;
+  gboolean max_render_time_is_known;
 
   now_us = g_get_monotonic_time ();
 
@@ -516,10 +593,13 @@ calculate_next_update_time_us (ClutterFrameClock *frame_clock,
     }
 
   min_render_time_allowed_us = refresh_interval_us / 2;
-  max_render_time_allowed_us =
-    clutter_frame_clock_compute_max_render_time_us (frame_clock);
 
-  if (min_render_time_allowed_us > max_render_time_allowed_us)
+  max_render_time_is_known =
+    clutter_frame_clock_compute_max_render_time_us (frame_clock,
+                                                    &max_render_time_allowed_us);
+
+  if (max_render_time_is_known &&
+      min_render_time_allowed_us > max_render_time_allowed_us)
     min_render_time_allowed_us = max_render_time_allowed_us;
 
   /*
@@ -540,7 +620,29 @@ calculate_next_update_time_us (ClutterFrameClock *frame_clock,
    *
    */
   last_presentation_time_us = frame_clock->last_presentation_time_us;
-  next_presentation_time_us = last_presentation_time_us + refresh_interval_us;
+  switch (frame_clock->state)
+    {
+    case CLUTTER_FRAME_CLOCK_STATE_INIT:
+    case CLUTTER_FRAME_CLOCK_STATE_IDLE:
+    case CLUTTER_FRAME_CLOCK_STATE_SCHEDULED:
+    case CLUTTER_FRAME_CLOCK_STATE_SCHEDULED_NOW:
+      next_smooth_presentation_time_us = last_presentation_time_us +
+                                         refresh_interval_us;
+      break;
+    case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_ONE:
+    case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_ONE_AND_SCHEDULED:
+    case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_ONE_AND_SCHEDULED_NOW:
+      next_smooth_presentation_time_us = last_presentation_time_us +
+                                         2 * refresh_interval_us;
+      break;
+    case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_TWO:
+      g_warn_if_reached ();  /* quad buffering would be a bug */
+      next_smooth_presentation_time_us = last_presentation_time_us +
+                                         3 * refresh_interval_us;
+      break;
+    }
+
+  next_presentation_time_us = next_smooth_presentation_time_us;
 
   /*
    * However, the last presentation could have happened more than a frame ago.
@@ -607,7 +709,7 @@ calculate_next_update_time_us (ClutterFrameClock *frame_clock,
     }
 
   if (frame_clock->last_presentation_flags & CLUTTER_FRAME_INFO_FLAG_VSYNC &&
-      next_presentation_time_us != last_presentation_time_us + refresh_interval_us)
+      next_presentation_time_us != next_smooth_presentation_time_us)
     {
       /* There was an idle period since the last presentation, so there seems
        * be no constantly updating actor. In this case it's best to start
@@ -619,6 +721,24 @@ calculate_next_update_time_us (ClutterFrameClock *frame_clock,
     }
   else
     {
+      /* If the max render time isn't known then using the current value of
+       * next_presentation_time_us is suboptimal. Targeting always one frame
+       * prior to that we'd lose the ability to scale up to triple buffering
+       * on late presentation. But targeting two frames prior we would be
+       * always triple buffering even when not required.
+       *   So the algorithm for deciding when to scale up to triple buffering
+       * in the absence of render time measurements is to simply target full
+       * frame rate. If we're keeping up then we'll stay double buffering. If
+       * we're not keeping up then this will switch us to triple buffering.
+       */
+      if (!max_render_time_is_known)
+        {
+          max_render_time_allowed_us =
+            (int64_t) (refresh_interval_us * SYNC_DELAY_FALLBACK_FRACTION);
+          next_presentation_time_us =
+            last_presentation_time_us + refresh_interval_us;
+        }
+
       while (next_presentation_time_us - min_render_time_allowed_us < now_us)
         next_presentation_time_us += refresh_interval_us;
 
@@ -650,7 +770,9 @@ calculate_next_variable_update_time_us (ClutterFrameClock *frame_clock,
 
   refresh_interval_us = frame_clock->refresh_interval_us;
 
-  if (frame_clock->last_presentation_time_us == 0)
+  if (frame_clock->last_presentation_time_us == 0 ||
+      !clutter_frame_clock_compute_max_render_time_us (frame_clock,
+                                                       &max_render_time_allowed_us))
     {
       *out_next_update_time_us =
         frame_clock->last_dispatch_time_us ?
@@ -663,9 +785,6 @@ calculate_next_variable_update_time_us (ClutterFrameClock *frame_clock,
       return;
     }
 
-  max_render_time_allowed_us =
-    clutter_frame_clock_compute_max_render_time_us (frame_clock);
-
   last_presentation_time_us = frame_clock->last_presentation_time_us;
   next_presentation_time_us = last_presentation_time_us + refresh_interval_us;
 
@@ -739,7 +858,17 @@ clutter_frame_clock_inhibit (ClutterFrameClock *frame_clock)
           frame_clock->pending_reschedule_now = TRUE;
           frame_clock->state = CLUTTER_FRAME_CLOCK_STATE_IDLE;
           break;
-        case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED:
+        case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_ONE_AND_SCHEDULED:
+          frame_clock->pending_reschedule = TRUE;
+          frame_clock->state = CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_ONE;
+          break;
+        case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_ONE_AND_SCHEDULED_NOW:
+          frame_clock->pending_reschedule = TRUE;
+          frame_clock->pending_reschedule_now = TRUE;
+          frame_clock->state = CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_ONE;
+          break;
+        case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_ONE:
+        case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_TWO:
           break;
         }
 
@@ -758,6 +887,25 @@ clutter_frame_clock_uninhibit (ClutterFrameClock *frame_clock)
     maybe_reschedule_update (frame_clock);
 }
 
+static gboolean
+want_triple_buffering (ClutterFrameClock *frame_clock)
+{
+  switch (triple_buffering_mode)
+    {
+    case TRIPLE_BUFFERING_MODE_NEVER:
+      return FALSE;
+    case TRIPLE_BUFFERING_MODE_AUTO:
+      return frame_clock->mode == CLUTTER_FRAME_CLOCK_MODE_FIXED &&
+             !(frame_clock->last_flip_hints &
+               CLUTTER_FRAME_HINT_DIRECT_SCANOUT_ATTEMPTED);
+    case TRIPLE_BUFFERING_MODE_ALWAYS:
+      return TRUE;
+    }
+
+  g_assert_not_reached ();
+  return FALSE;
+}
+
 void
 clutter_frame_clock_schedule_update_now (ClutterFrameClock *frame_clock)
 {
@@ -775,10 +923,24 @@ clutter_frame_clock_schedule_update_now (ClutterFrameClock *frame_clock)
     case CLUTTER_FRAME_CLOCK_STATE_INIT:
     case CLUTTER_FRAME_CLOCK_STATE_IDLE:
     case CLUTTER_FRAME_CLOCK_STATE_SCHEDULED:
+      frame_clock->state = CLUTTER_FRAME_CLOCK_STATE_SCHEDULED_NOW;
       break;
     case CLUTTER_FRAME_CLOCK_STATE_SCHEDULED_NOW:
+    case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_ONE_AND_SCHEDULED_NOW:
       return;
-    case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED:
+    case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_ONE_AND_SCHEDULED:
+      frame_clock->state =
+        CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_ONE_AND_SCHEDULED_NOW;
+      break;
+    case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_ONE:
+      if (want_triple_buffering (frame_clock))
+        {
+          frame_clock->state =
+            CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_ONE_AND_SCHEDULED_NOW;
+          break;
+        }
+      G_GNUC_FALLTHROUGH;
+    case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_TWO:
       frame_clock->pending_reschedule = TRUE;
       frame_clock->pending_reschedule_now = TRUE;
       return;
@@ -807,13 +969,17 @@ clutter_frame_clock_schedule_update_now (ClutterFrameClock *frame_clock)
 
   frame_clock->next_update_time_us = next_update_time_us;
   g_source_set_ready_time (frame_clock->source, next_update_time_us);
-  frame_clock->state = CLUTTER_FRAME_CLOCK_STATE_SCHEDULED_NOW;
 }
 
 void
 clutter_frame_clock_schedule_update (ClutterFrameClock *frame_clock)
 {
   int64_t next_update_time_us = -1;
+  TripleBufferingMode current_mode = triple_buffering_mode;
+
+  if (current_mode == TRIPLE_BUFFERING_MODE_AUTO &&
+      !want_triple_buffering (frame_clock))
+    current_mode = TRIPLE_BUFFERING_MODE_NEVER;
 
   if (frame_clock->inhibit_count > 0)
     {
@@ -829,11 +995,33 @@ clutter_frame_clock_schedule_update (ClutterFrameClock *frame_clock)
       frame_clock->state = CLUTTER_FRAME_CLOCK_STATE_SCHEDULED;
       return;
     case CLUTTER_FRAME_CLOCK_STATE_IDLE:
+      frame_clock->state = CLUTTER_FRAME_CLOCK_STATE_SCHEDULED;
       break;
     case CLUTTER_FRAME_CLOCK_STATE_SCHEDULED:
     case CLUTTER_FRAME_CLOCK_STATE_SCHEDULED_NOW:
+    case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_ONE_AND_SCHEDULED:
+    case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_ONE_AND_SCHEDULED_NOW:
       return;
-    case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED:
+    case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_ONE:
+      switch (current_mode)
+        {
+        case TRIPLE_BUFFERING_MODE_NEVER:
+          frame_clock->pending_reschedule = TRUE;
+          return;
+        case TRIPLE_BUFFERING_MODE_AUTO:
+          frame_clock->state =
+            CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_ONE_AND_SCHEDULED;
+          break;
+        case TRIPLE_BUFFERING_MODE_ALWAYS:
+          next_update_time_us = g_get_monotonic_time ();
+          frame_clock->next_presentation_time_us = 0;
+          frame_clock->is_next_presentation_time_valid = FALSE;
+          frame_clock->state =
+            CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_ONE_AND_SCHEDULED;
+          goto got_update_time;
+        }
+      break;
+    case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_TWO:
       frame_clock->pending_reschedule = TRUE;
       return;
     }
@@ -858,11 +1046,11 @@ clutter_frame_clock_schedule_update (ClutterFrameClock *frame_clock)
       break;
     }
 
+got_update_time:
   g_warn_if_fail (next_update_time_us != -1);
 
   frame_clock->next_update_time_us = next_update_time_us;
   g_source_set_ready_time (frame_clock->source, next_update_time_us);
-  frame_clock->state = CLUTTER_FRAME_CLOCK_STATE_SCHEDULED;
 }
 
 void
@@ -878,6 +1066,8 @@ clutter_frame_clock_set_mode (ClutterFrameClock     *frame_clock,
     {
     case CLUTTER_FRAME_CLOCK_STATE_INIT:
     case CLUTTER_FRAME_CLOCK_STATE_IDLE:
+    case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_ONE:
+    case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_TWO:
       break;
     case CLUTTER_FRAME_CLOCK_STATE_SCHEDULED:
       frame_clock->pending_reschedule = TRUE;
@@ -888,7 +1078,14 @@ clutter_frame_clock_set_mode (ClutterFrameClock     *frame_clock,
       frame_clock->pending_reschedule_now = TRUE;
       frame_clock->state = CLUTTER_FRAME_CLOCK_STATE_IDLE;
       break;
-    case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED:
+    case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_ONE_AND_SCHEDULED:
+      frame_clock->pending_reschedule = TRUE;
+      frame_clock->state = CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_ONE;
+      break;
+    case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_ONE_AND_SCHEDULED_NOW:
+      frame_clock->pending_reschedule = TRUE;
+      frame_clock->pending_reschedule_now = TRUE;
+      frame_clock->state = CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_ONE;
       break;
     }
 
@@ -945,10 +1142,27 @@ clutter_frame_clock_dispatch (ClutterFrameClock *frame_clock,
     }
 #endif
 
+  frame_clock->prev_last_dispatch_time_us = frame_clock->last_dispatch_time_us;
   frame_clock->last_dispatch_time_us = time_us;
   g_source_set_ready_time (frame_clock->source, -1);
 
-  frame_clock->state = CLUTTER_FRAME_CLOCK_STATE_DISPATCHED;
+  switch (frame_clock->state)
+    {
+    case CLUTTER_FRAME_CLOCK_STATE_INIT:
+    case CLUTTER_FRAME_CLOCK_STATE_IDLE:
+    case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_ONE:
+    case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_TWO:
+      g_warn_if_reached ();
+      return;
+    case CLUTTER_FRAME_CLOCK_STATE_SCHEDULED:
+    case CLUTTER_FRAME_CLOCK_STATE_SCHEDULED_NOW:
+      frame_clock->state = CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_ONE;
+      break;
+    case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_ONE_AND_SCHEDULED:
+    case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_ONE_AND_SCHEDULED_NOW:
+      frame_clock->state = CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_TWO;
+      break;
+    }
 
   frame_count = frame_clock->frame_count++;
 
@@ -979,26 +1193,36 @@ clutter_frame_clock_dispatch (ClutterFrameClock *frame_clock,
   result = iface->frame (frame_clock, frame, frame_clock->listener.user_data);
   COGL_TRACE_END (ClutterFrameClockFrame);
 
-  switch (frame_clock->state)
+  switch (result)
     {
-    case CLUTTER_FRAME_CLOCK_STATE_INIT:
-      g_warn_if_reached ();
+    case CLUTTER_FRAME_RESULT_PENDING_PRESENTED:
       break;
-    case CLUTTER_FRAME_CLOCK_STATE_IDLE:
-      /* Presentation completed synchronously in the above listener */
-    case CLUTTER_FRAME_CLOCK_STATE_SCHEDULED:
-    case CLUTTER_FRAME_CLOCK_STATE_SCHEDULED_NOW:
-      break;
-    case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED:
-      switch (result)
+    case CLUTTER_FRAME_RESULT_IDLE:
+      /* The frame was aborted; nothing to paint/present */
+      switch (frame_clock->state)
         {
-        case CLUTTER_FRAME_RESULT_PENDING_PRESENTED:
+        case CLUTTER_FRAME_CLOCK_STATE_INIT:
+        case CLUTTER_FRAME_CLOCK_STATE_IDLE:
+        case CLUTTER_FRAME_CLOCK_STATE_SCHEDULED:
+        case CLUTTER_FRAME_CLOCK_STATE_SCHEDULED_NOW:
+          g_warn_if_reached ();
           break;
-        case CLUTTER_FRAME_RESULT_IDLE:
-          /* The frame was aborted; nothing to paint/present */
+        case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_ONE:
           frame_clock->state = CLUTTER_FRAME_CLOCK_STATE_IDLE;
           maybe_reschedule_update (frame_clock);
           break;
+        case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_ONE_AND_SCHEDULED:
+          frame_clock->state = CLUTTER_FRAME_CLOCK_STATE_SCHEDULED;
+          maybe_reschedule_update (frame_clock);
+          break;
+        case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_ONE_AND_SCHEDULED_NOW:
+          frame_clock->state = CLUTTER_FRAME_CLOCK_STATE_SCHEDULED_NOW;
+          maybe_reschedule_update (frame_clock);
+          break;
+        case CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_TWO:
+          frame_clock->state = CLUTTER_FRAME_CLOCK_STATE_DISPATCHED_ONE;
+          maybe_reschedule_update (frame_clock);
+          break;
         }
       break;
     }
@@ -1031,21 +1255,31 @@ frame_clock_source_dispatch (GSource     *source,
 }
 
 void
-clutter_frame_clock_record_flip_time (ClutterFrameClock *frame_clock,
-                                      int64_t            flip_time_us)
+clutter_frame_clock_record_flip (ClutterFrameClock *frame_clock,
+                                 int64_t            flip_time_us,
+                                 ClutterFrameHint   hints)
 {
+  frame_clock->prev_last_flip_time_us = frame_clock->last_flip_time_us;
   frame_clock->last_flip_time_us = flip_time_us;
+  frame_clock->last_flip_hints = hints;
 }
 
 GString *
 clutter_frame_clock_get_max_render_time_debug_info (ClutterFrameClock *frame_clock)
 {
+  int64_t max_render_time_us;
   int64_t max_update_duration_us;
   GString *string;
 
-  string = g_string_new (NULL);
-  g_string_append_printf (string, "Max render time: %ld µs",
-                          clutter_frame_clock_compute_max_render_time_us (frame_clock));
+  string = g_string_new ("Max render time: ");
+  if (!clutter_frame_clock_compute_max_render_time_us (frame_clock,
+                                                       &max_render_time_us))
+    {
+      g_string_append (string, "unknown");
+      return string;
+    }
+
+  g_string_append_printf (string, "%ld µs", max_render_time_us);
 
   if (frame_clock->got_measurements_last_frame)
     g_string_append_printf (string, " =");
@@ -1235,6 +1469,15 @@ static void
 clutter_frame_clock_class_init (ClutterFrameClockClass *klass)
 {
   GObjectClass *object_class = G_OBJECT_CLASS (klass);
+  const char *mode_str;
+
+  mode_str = g_getenv ("MUTTER_DEBUG_TRIPLE_BUFFERING");
+  if (!g_strcmp0 (mode_str, "never"))
+    triple_buffering_mode = TRIPLE_BUFFERING_MODE_NEVER;
+  else if (!g_strcmp0 (mode_str, "auto"))
+    triple_buffering_mode = TRIPLE_BUFFERING_MODE_AUTO;
+  else if (!g_strcmp0 (mode_str, "always"))
+    triple_buffering_mode = TRIPLE_BUFFERING_MODE_ALWAYS;
 
   object_class->dispose = clutter_frame_clock_dispose;
 

clutter/clutter/clutter-frame-clock.h

@@ -33,6 +33,12 @@ typedef enum _ClutterFrameResult
   CLUTTER_FRAME_RESULT_IDLE,
 } ClutterFrameResult;
 
+typedef enum _ClutterFrameHint
+{
+  CLUTTER_FRAME_HINT_NONE                     = 0,
+  CLUTTER_FRAME_HINT_DIRECT_SCANOUT_ATTEMPTED = 1 << 0,
+} ClutterFrameHint;
+
 #define CLUTTER_TYPE_FRAME_CLOCK (clutter_frame_clock_get_type ())
 CLUTTER_EXPORT
 G_DECLARE_FINAL_TYPE (ClutterFrameClock, clutter_frame_clock,
@@ -102,8 +108,9 @@ void clutter_frame_clock_remove_timeline (ClutterFrameClock *frame_clock,
 CLUTTER_EXPORT
 float clutter_frame_clock_get_refresh_rate (ClutterFrameClock *frame_clock);
 
-void clutter_frame_clock_record_flip_time (ClutterFrameClock *frame_clock,
-                                           int64_t            flip_time_us);
+void clutter_frame_clock_record_flip (ClutterFrameClock *frame_clock,
+                                      int64_t            flip_time_us,
+                                      ClutterFrameHint   hints);
 
 GString * clutter_frame_clock_get_max_render_time_debug_info (ClutterFrameClock *frame_clock);
 

clutter/clutter/clutter-frame-private.h

@@ -36,6 +36,7 @@ struct _ClutterFrame
 
   gboolean has_result;
   ClutterFrameResult result;
+  ClutterFrameHint hints;
 };
 
 CLUTTER_EXPORT

clutter/clutter/clutter-frame.c

@@ -115,3 +115,16 @@ clutter_frame_set_result (ClutterFrame       *frame,
   frame->result = result;
   frame->has_result = TRUE;
 }
+
+void
+clutter_frame_set_hint (ClutterFrame     *frame,
+                        ClutterFrameHint  hint)
+{
+  frame->hints |= hint;
+}
+
+ClutterFrameHint
+clutter_frame_get_hints (ClutterFrame *frame)
+{
+  return frame->hints;
+}

clutter/clutter/clutter-frame.h

@@ -54,4 +54,11 @@ void clutter_frame_set_result (ClutterFrame       *frame,
 CLUTTER_EXPORT
 gboolean clutter_frame_has_result (ClutterFrame *frame);
 
+CLUTTER_EXPORT
+void clutter_frame_set_hint (ClutterFrame     *frame,
+                             ClutterFrameHint  hint);
+
+CLUTTER_EXPORT
+ClutterFrameHint clutter_frame_get_hints (ClutterFrame *frame);
+
 G_DEFINE_AUTOPTR_CLEANUP_FUNC (ClutterFrame, clutter_frame_unref)

clutter/clutter/clutter-stage-view.c

@@ -1075,14 +1075,21 @@ handle_frame_clock_frame (ClutterFrameClock *frame_clock,
 
       _clutter_stage_window_redraw_view (stage_window, view, frame);
 
-      clutter_frame_clock_record_flip_time (frame_clock,
-                                            g_get_monotonic_time ());
+      clutter_frame_clock_record_flip (frame_clock,
+                                       g_get_monotonic_time (),
+                                       clutter_frame_get_hints (frame));
 
       clutter_stage_emit_after_paint (stage, view, frame);
 
       if (clutter_context_get_show_fps (context))
         end_frame_timing_measurement (view);
     }
+  else
+    {
+      clutter_frame_clock_record_flip (frame_clock,
+                                       g_get_monotonic_time (),
+                                       clutter_frame_get_hints (frame));
+    }
 
   _clutter_stage_window_finish_frame (stage_window, view, frame);
 

src/backends/meta-stage-impl.c

@@ -798,6 +798,8 @@ meta_stage_impl_redraw_view (ClutterStageWindow *stage_window,
     {
       g_autoptr (GError) error = NULL;
 
+      clutter_frame_set_hint (frame, CLUTTER_FRAME_HINT_DIRECT_SCANOUT_ATTEMPTED);
+
       if (meta_stage_impl_scanout_view (stage_impl,
                                         stage_view,
                                         scanout,

src/backends/native/meta-onscreen-native.c

@@ -1514,6 +1514,7 @@ try_post_latest_swap (CoglOnscreen *onscreen)
   COGL_TRACE_SCOPED_ANCHOR (MetaRendererNativePostKmsUpdate);
 
   if (onscreen_native->next_frame == NULL ||
+      onscreen_native->view == NULL ||
       meta_kms_is_shutting_down (kms))
     return;
 

src/tests/native-kms-render.c

@@ -39,6 +39,8 @@
 #include "tests/meta-wayland-test-driver.h"
 #include "tests/meta-wayland-test-utils.h"
 
+#define N_FRAMES_PER_TEST 30
+
 typedef struct
 {
   int number_of_frames_left;
@@ -46,12 +48,15 @@ typedef struct
 
   struct {
     int n_paints;
-    uint32_t fb_id;
+    int n_presentations;
+    int n_direct_scanouts;
+    GList *fb_ids;
   } scanout;
 
   gboolean wait_for_scanout;
 
   struct {
+    int scanouts_attempted;
     gboolean scanout_sabotaged;
     gboolean fallback_painted;
     guint repaint_guard_id;
@@ -101,7 +106,7 @@ meta_test_kms_render_basic (void)
   gulong handler_id;
 
   test = (KmsRenderingTest) {
-    .number_of_frames_left = 10,
+    .number_of_frames_left = N_FRAMES_PER_TEST,
     .loop = g_main_loop_new (NULL, FALSE),
   };
   handler_id = g_signal_connect (stage, "after-update",
@@ -123,7 +128,6 @@ on_scanout_before_update (ClutterStage     *stage,
                           KmsRenderingTest *test)
 {
   test->scanout.n_paints = 0;
-  test->scanout.fb_id = 0;
 }
 
 static void
@@ -135,6 +139,7 @@ on_scanout_before_paint (ClutterStage     *stage,
   CoglScanout *scanout;
   CoglScanoutBuffer *scanout_buffer;
   MetaDrmBuffer *buffer;
+  uint32_t fb_id;
 
   scanout = clutter_stage_view_peek_scanout (stage_view);
   if (!scanout)
@@ -143,8 +148,13 @@ on_scanout_before_paint (ClutterStage     *stage,
   scanout_buffer = cogl_scanout_get_buffer (scanout);
   g_assert_true (META_IS_DRM_BUFFER (scanout_buffer));
   buffer = META_DRM_BUFFER (scanout_buffer);
-  test->scanout.fb_id = meta_drm_buffer_get_fb_id (buffer);
-  g_assert_cmpuint (test->scanout.fb_id, >, 0);
+  fb_id = meta_drm_buffer_get_fb_id (buffer);
+  g_assert_cmpuint (fb_id, >, 0);
+  test->scanout.fb_ids = g_list_append (test->scanout.fb_ids,
+                                        GUINT_TO_POINTER (fb_id));
+
+  /* Triple buffering, but no higher */
+  g_assert_cmpuint (g_list_length (test->scanout.fb_ids), <=, 2);
 }
 
 static void
@@ -173,12 +183,12 @@ on_scanout_presented (ClutterStage     *stage,
   MetaDeviceFile *device_file;
   GError *error = NULL;
   drmModeCrtc *drm_crtc;
+  uint32_t first_fb_id_expected;
 
-  if (test->wait_for_scanout && test->scanout.n_paints > 0)
+  if (test->wait_for_scanout && test->scanout.fb_ids == NULL)
     return;
 
-  if (test->wait_for_scanout && test->scanout.fb_id == 0)
-    return;
+  test->scanout.n_presentations++;
 
   device_pool = meta_backend_native_get_device_pool (backend_native);
 
@@ -197,15 +207,41 @@ on_scanout_presented (ClutterStage     *stage,
   drm_crtc = drmModeGetCrtc (meta_device_file_get_fd (device_file),
                              meta_kms_crtc_get_id (kms_crtc));
   g_assert_nonnull (drm_crtc);
-  if (test->scanout.fb_id == 0)
-    g_assert_cmpuint (drm_crtc->buffer_id, !=, test->scanout.fb_id);
+
+  if (test->scanout.fb_ids)
+    {
+      test->scanout.n_direct_scanouts++;
+      first_fb_id_expected = GPOINTER_TO_UINT (test->scanout.fb_ids->data);
+      test->scanout.fb_ids = g_list_delete_link (test->scanout.fb_ids,
+                                                 test->scanout.fb_ids);
+    }
   else
-    g_assert_cmpuint (drm_crtc->buffer_id, ==, test->scanout.fb_id);
+    {
+      first_fb_id_expected = 0;
+    }
+
+  /* The buffer ID won't match on the first frame because switching from
+   * triple buffered compositing to double buffered direct scanout takes
+   * an extra frame to drain the queue. Thereafter we are in direct scanout
+   * mode and expect the buffer IDs to match.
+   */
+  if (test->scanout.n_presentations > 1)
+    {
+      if (first_fb_id_expected == 0)
+        g_assert_cmpuint (drm_crtc->buffer_id, !=, first_fb_id_expected);
+      else
+        g_assert_cmpuint (drm_crtc->buffer_id, ==, first_fb_id_expected);
+    }
+
   drmModeFreeCrtc (drm_crtc);
 
   meta_device_file_release (device_file);
 
-  g_main_loop_quit (test->loop);
+  test->number_of_frames_left--;
+  if (test->number_of_frames_left <= 0)
+    g_main_loop_quit (test->loop);
+  else
+    clutter_actor_queue_redraw (CLUTTER_ACTOR (stage));
 }
 
 typedef enum
@@ -244,7 +280,9 @@ meta_test_kms_render_client_scanout (void)
   g_assert_nonnull (wayland_test_client);
 
   test = (KmsRenderingTest) {
+    .number_of_frames_left = N_FRAMES_PER_TEST,
     .loop = g_main_loop_new (NULL, FALSE),
+    .scanout = {0},
     .wait_for_scanout = TRUE,
   };
 
@@ -270,7 +308,8 @@ meta_test_kms_render_client_scanout (void)
   clutter_actor_queue_redraw (CLUTTER_ACTOR (stage));
   g_main_loop_run (test.loop);
 
-  g_assert_cmpuint (test.scanout.fb_id, >, 0);
+  g_assert_cmpint (test.scanout.n_presentations, ==, N_FRAMES_PER_TEST);
+  g_assert_cmpint (test.scanout.n_direct_scanouts, ==, N_FRAMES_PER_TEST);
 
   g_debug ("Unmake fullscreen");
   window = meta_find_window_from_title (test_context, "dma-buf-scanout-test");
@@ -292,10 +331,15 @@ meta_test_kms_render_client_scanout (void)
   g_assert_cmpint (buffer_rect.y, ==, 10);
 
   test.wait_for_scanout = FALSE;
+  test.number_of_frames_left = N_FRAMES_PER_TEST;
+  test.scanout.n_presentations = 0;
+  test.scanout.n_direct_scanouts = 0;
+
   clutter_actor_queue_redraw (CLUTTER_ACTOR (stage));
   g_main_loop_run (test.loop);
 
-  g_assert_cmpuint (test.scanout.fb_id, ==, 0);
+  g_assert_cmpint (test.scanout.n_presentations, ==, N_FRAMES_PER_TEST);
+  g_assert_cmpint (test.scanout.n_direct_scanouts, ==, 0);
 
   g_debug ("Moving back to 0, 0");
   meta_window_move_frame (window, TRUE, 0, 0);
@@ -307,10 +351,15 @@ meta_test_kms_render_client_scanout (void)
   g_assert_cmpint (buffer_rect.y, ==, 0);
 
   test.wait_for_scanout = TRUE;
+  test.number_of_frames_left = N_FRAMES_PER_TEST;
+  test.scanout.n_presentations = 0;
+  test.scanout.n_direct_scanouts = 0;
+
   clutter_actor_queue_redraw (CLUTTER_ACTOR (stage));
   g_main_loop_run (test.loop);
 
-  g_assert_cmpuint (test.scanout.fb_id, >, 0);
+  g_assert_cmpint (test.scanout.n_presentations, ==, N_FRAMES_PER_TEST);
+  g_assert_cmpint (test.scanout.n_direct_scanouts, ==, N_FRAMES_PER_TEST);
 
   g_signal_handler_disconnect (stage, before_update_handler_id);
   g_signal_handler_disconnect (stage, before_paint_handler_id);
@@ -364,6 +413,15 @@ on_scanout_fallback_before_paint (ClutterStage     *stage,
   if (!scanout)
     return;
 
+  test->scanout_fallback.scanouts_attempted++;
+
+  /* The first scanout candidate frame will get composited due to triple
+   * buffering draining the queue to drop to double buffering. So don't
+   * sabotage that first frame.
+   */
+  if (test->scanout_fallback.scanouts_attempted < 2)
+    return;
+
   g_assert_false (test->scanout_fallback.scanout_sabotaged);
 
   if (is_atomic_mode_setting (kms_device))
@@ -401,6 +459,15 @@ on_scanout_fallback_paint_view (ClutterStage     *stage,
       g_clear_handle_id (&test->scanout_fallback.repaint_guard_id,
                          g_source_remove);
       test->scanout_fallback.fallback_painted = TRUE;
+      test->scanout_fallback.scanout_sabotaged = FALSE;
+    }
+  else if (test->scanout_fallback.scanouts_attempted == 1)
+    {
+      /* Now that we've seen the first scanout attempt that was inhibited by
+       * triple buffering, try a second frame. The second one should scanout
+       * and will be sabotaged.
+       */
+      clutter_actor_queue_redraw (CLUTTER_ACTOR (stage));
     }
 }
 
@@ -410,11 +477,11 @@ on_scanout_fallback_presented (ClutterStage     *stage,
                                ClutterFrameInfo *frame_info,
                                KmsRenderingTest *test)
 {
-  if (!test->scanout_fallback.scanout_sabotaged)
-    return;
+  if (test->scanout_fallback.fallback_painted)
+    g_main_loop_quit (test->loop);
 
-  g_assert_true (test->scanout_fallback.fallback_painted);
-  g_main_loop_quit (test->loop);
+  test->number_of_frames_left--;
+  g_assert_cmpint (test->number_of_frames_left, >, 0);
 }
 
 static void
@@ -443,6 +510,7 @@ meta_test_kms_render_client_scanout_fallback (void)
   g_assert_nonnull (wayland_test_client);
 
   test = (KmsRenderingTest) {
+    .number_of_frames_left = N_FRAMES_PER_TEST,
     .loop = g_main_loop_new (NULL, FALSE),
   };