mutter-performance-pkgbuild/mr1441.patch

Author: Daniel van Vugt <daniel.van.vugt@canonical.com>
Source: https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1441
Source: https://gitlab.gnome.org/Community/Ubuntu/mutter/-/tree/triple-buffering-v4-46
---

Use triple buffering if and when the previous frame is running late.
This means the next frame will be dispatched on time instead of also starting
late.

It also triggers a GPU clock boost if deemed necessary by the driver.
Although frequency scaling is not required to get a performance gain here
because even a fixed frequency GPU will benefit from not over-sleeping anymore.
If the previous frame is not running late then we stick to double buffering so
there's no latency penalty when the system is able to maintain full frame rate.

---
From 8a4f0f780a66c3b5c77af7ea4729ae8a1ae1352b Mon Sep 17 00:00:00 2001
From: Mingi Sung <sungmg@saltyming.net>
Date: Sun, 23 Jun 2024 20:53:22 +0900
Subject: [PATCH] Merge branch 'triple-buffering-v4-46' into gnome-46

---
 clutter/clutter/clutter-frame-clock.c      | 369 +++++++++++++++++++++++------
 clutter/clutter/clutter-frame-clock.h      |  11 +-
 clutter/clutter/clutter-frame-private.h    |   1 +
 clutter/clutter/clutter-frame.c            |  13 +
 clutter/clutter/clutter-frame.h            |   7 +
 clutter/clutter/clutter-stage-view.c       |  11 +-
 cogl/cogl/cogl-onscreen-private.h          |   5 +-
 cogl/cogl/cogl-onscreen.c                  |   8 +
 src/backends/meta-stage-impl.c             |   2 +
 src/backends/native/meta-kms-impl-device.c |   4 +-
 src/backends/native/meta-kms.c             |   9 +
 src/backends/native/meta-kms.h             |   2 +
 src/backends/native/meta-onscreen-native.c | 368 ++++++++++++++++++++++------
 src/backends/native/meta-onscreen-native.h |   2 +
 src/backends/native/meta-renderer-native.c |  34 ++-
 src/tests/native-kms-render.c              | 106 +++++++--
 16 files changed, 779 insertions(+), 173 deletions(-)

diff --git a/clutter/clutter/clutter-frame-clock.c b/clutter/clutter/clutter-frame-clock.c
index 30a319f60..75e33088c 100644
--- a/clutter/clutter/clutter-frame-clock.c
+++ b/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 0.875
 
 #define MINIMUM_REFRESH_RATE 30.f
@@ -70,8 +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_DISPATCHING,
-  CLUTTER_FRAME_CLOCK_STATE_PENDING_PRESENTED,
+  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
@@ -92,6 +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;
@@ -113,6 +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;
@@ -247,10 +262,6 @@ 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;
@@ -277,6 +288,12 @@ 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,
@@ -366,22 +383,52 @@ 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)
+        {
+          /* Cursor-only updates with no "swap" or "flip" */
+          dispatch_to_swap_us = 0;
+          swap_to_flip_us = 0;
+        }
+      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,
@@ -391,7 +438,7 @@ clutter_frame_clock_notify_presented (ClutterFrameClock *frame_clock,
         CLAMP (frame_clock->last_dispatch_lateness_us + dispatch_to_swap_us +
                MAX (swap_to_rendering_done_us, swap_to_flip_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);
 
@@ -400,7 +447,8 @@ 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);
     }
 
@@ -418,11 +466,22 @@ 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_DISPATCHING:
-    case CLUTTER_FRAME_CLOCK_STATE_PENDING_PRESENTED:
+    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;
     }
 }
 
@@ -440,26 +499,37 @@ 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_DISPATCHING:
-    case CLUTTER_FRAME_CLOCK_STATE_PENDING_PRESENTED:
+    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 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 +543,15 @@ 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 +566,9 @@ 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 +588,13 @@ 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 +615,28 @@ 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:
+      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 +703,7 @@ 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 +715,24 @@ 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 =
+            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 +764,9 @@ 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 +779,6 @@ 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;
 
@@ -739,8 +852,17 @@ 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_DISPATCHING:
-        case CLUTTER_FRAME_CLOCK_STATE_PENDING_PRESENTED:
+        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;
         }
 
@@ -759,6 +881,25 @@ 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)
 {
@@ -776,11 +917,24 @@ 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_DISPATCHING:
-    case CLUTTER_FRAME_CLOCK_STATE_PENDING_PRESENTED:
+    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;
@@ -809,13 +963,17 @@ 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)
     {
@@ -831,12 +989,33 @@ 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_DISPATCHING:
-    case CLUTTER_FRAME_CLOCK_STATE_PENDING_PRESENTED:
+    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;
     }
@@ -861,11 +1040,11 @@ 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
@@ -881,6 +1060,8 @@ 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;
@@ -891,8 +1072,14 @@ 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_DISPATCHING:
-    case CLUTTER_FRAME_CLOCK_STATE_PENDING_PRESENTED:
+    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;
     }
 
@@ -928,7 +1115,7 @@ clutter_frame_clock_dispatch (ClutterFrameClock *frame_clock,
                              frame_clock->refresh_interval_us;
 
   lateness_us = time_us - ideal_dispatch_time_us;
-  if (lateness_us < 0 || lateness_us >= frame_clock->refresh_interval_us)
+  if (lateness_us < 0 || lateness_us >= frame_clock->refresh_interval_us / 4)
     frame_clock->last_dispatch_lateness_us = 0;
   else
     frame_clock->last_dispatch_lateness_us = lateness_us;
@@ -949,10 +1136,27 @@ 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_DISPATCHING;
+  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++;
 
@@ -983,26 +1187,36 @@ 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:
-    case CLUTTER_FRAME_CLOCK_STATE_PENDING_PRESENTED:
-      g_warn_if_reached ();
+    case CLUTTER_FRAME_RESULT_PENDING_PRESENTED:
       break;
-    case CLUTTER_FRAME_CLOCK_STATE_IDLE:
-    case CLUTTER_FRAME_CLOCK_STATE_SCHEDULED:
-    case CLUTTER_FRAME_CLOCK_STATE_SCHEDULED_NOW:
-      break;
-    case CLUTTER_FRAME_CLOCK_STATE_DISPATCHING:
-      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:
-          frame_clock->state = CLUTTER_FRAME_CLOCK_STATE_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:
+        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;
     }
@@ -1035,21 +1249,31 @@ 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, " =");
@@ -1216,8 +1440,6 @@ clutter_frame_clock_dispose (GObject *object)
 {
   ClutterFrameClock *frame_clock = CLUTTER_FRAME_CLOCK (object);
 
-  g_warn_if_fail (frame_clock->state != CLUTTER_FRAME_CLOCK_STATE_DISPATCHING);
-
   if (frame_clock->source)
     {
       g_signal_emit (frame_clock, signals[DESTROY], 0);
@@ -1241,6 +1463,15 @@ 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;
 
diff --git a/clutter/clutter/clutter-frame-clock.h b/clutter/clutter/clutter-frame-clock.h
index a7be5ef31..bfc89bde0 100644
--- a/clutter/clutter/clutter-frame-clock.h
+++ b/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,
@@ -102,7 +108,8 @@ 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);
diff --git a/clutter/clutter/clutter-frame-private.h b/clutter/clutter/clutter-frame-private.h
index ef66b874e..ce140560a 100644
--- a/clutter/clutter/clutter-frame-private.h
+++ b/clutter/clutter/clutter-frame-private.h
@@ -36,6 +36,7 @@ struct _ClutterFrame
 
   gboolean has_result;
   ClutterFrameResult result;
+  ClutterFrameHint hints;
 };
 
 CLUTTER_EXPORT
diff --git a/clutter/clutter/clutter-frame.c b/clutter/clutter/clutter-frame.c
index 7436f9f18..53c289b2c 100644
--- a/clutter/clutter/clutter-frame.c
+++ b/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;
+}
diff --git a/clutter/clutter/clutter-frame.h b/clutter/clutter/clutter-frame.h
index 34f0770bd..c7b3d02ac 100644
--- a/clutter/clutter/clutter-frame.h
+++ b/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)
diff --git a/clutter/clutter/clutter-stage-view.c b/clutter/clutter/clutter-stage-view.c
index f5188e2ac..d53e37785 100644
--- a/clutter/clutter/clutter-stage-view.c
+++ b/clutter/clutter/clutter-stage-view.c
@@ -898,14 +898,21 @@ 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 ())
         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);
 
diff --git a/cogl/cogl/cogl-onscreen-private.h b/cogl/cogl/cogl-onscreen-private.h
index 959a60533..86d8ea2d5 100644
--- a/cogl/cogl/cogl-onscreen-private.h
+++ b/cogl/cogl/cogl-onscreen-private.h
@@ -78,4 +78,7 @@ COGL_EXPORT CoglFrameInfo *
 cogl_onscreen_peek_tail_frame_info (CoglOnscreen *onscreen);
 
 COGL_EXPORT CoglFrameInfo *
-cogl_onscreen_pop_head_frame_info (CoglOnscreen *onscreen);
+cogl_onscreen_pop_head_frame_info  (CoglOnscreen *onscreen);
+
+COGL_EXPORT unsigned int
+cogl_onscreen_count_pending_frames (CoglOnscreen *onscreen);
diff --git a/cogl/cogl/cogl-onscreen.c b/cogl/cogl/cogl-onscreen.c
index afb648bcd..086be7ed7 100644
--- a/cogl/cogl/cogl-onscreen.c
+++ b/cogl/cogl/cogl-onscreen.c
@@ -515,6 +515,14 @@ cogl_onscreen_pop_head_frame_info (CoglOnscreen *onscreen)
   return g_queue_pop_head (&priv->pending_frame_infos);
 }
 
+unsigned int
+cogl_onscreen_count_pending_frames (CoglOnscreen *onscreen)
+{
+  CoglOnscreenPrivate *priv = cogl_onscreen_get_instance_private (onscreen);
+
+  return g_queue_get_length (&priv->pending_frame_infos);
+}
+
 CoglFrameClosure *
 cogl_onscreen_add_frame_callback (CoglOnscreen *onscreen,
                                   CoglFrameCallback callback,
diff --git a/src/backends/meta-stage-impl.c b/src/backends/meta-stage-impl.c
index 7aa24439d..727e1a5f3 100644
--- a/src/backends/meta-stage-impl.c
+++ b/src/backends/meta-stage-impl.c
@@ -774,6 +774,8 @@ 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,
diff --git a/src/backends/native/meta-kms-impl-device.c b/src/backends/native/meta-kms-impl-device.c
index 6758982a6..315ed2717 100644
--- a/src/backends/native/meta-kms-impl-device.c
+++ b/src/backends/native/meta-kms-impl-device.c
@@ -1586,9 +1586,11 @@ meta_kms_impl_device_handle_update (MetaKmsImplDevice *impl_device,
       meta_kms_update_merge_from (crtc_frame->pending_update, update);
       meta_kms_update_free (update);
       update = g_steal_pointer (&crtc_frame->pending_update);
-      disarm_crtc_frame_deadline_timer (crtc_frame);
     }
 
+  if (crtc_frame->deadline.armed)
+    disarm_crtc_frame_deadline_timer (crtc_frame);
+
   meta_kms_device_handle_flush (priv->device, latch_crtc);
 
   feedback = do_process (impl_device, latch_crtc, update, flags);
diff --git a/src/backends/native/meta-kms.c b/src/backends/native/meta-kms.c
index 795008b21..70d1e792c 100644
--- a/src/backends/native/meta-kms.c
+++ b/src/backends/native/meta-kms.c
@@ -63,6 +63,8 @@ struct _MetaKms
   int kernel_thread_inhibit_count;
 
   MetaKmsCursorManager *cursor_manager;
+
+  gboolean shutting_down;
 };
 
 G_DEFINE_TYPE (MetaKms, meta_kms, META_TYPE_THREAD)
@@ -354,6 +356,7 @@ static void
 on_prepare_shutdown (MetaBackend *backend,
                      MetaKms     *kms)
 {
+  kms->shutting_down = TRUE;
   meta_kms_run_impl_task_sync (kms, prepare_shutdown_in_impl, NULL, NULL);
   meta_thread_flush_callbacks (META_THREAD (kms));
 
@@ -408,6 +411,12 @@ meta_kms_new (MetaBackend   *backend,
   return kms;
 }
 
+gboolean
+meta_kms_is_shutting_down (MetaKms *kms)
+{
+  return kms->shutting_down;
+}
+
 static void
 meta_kms_finalize (GObject *object)
 {
diff --git a/src/backends/native/meta-kms.h b/src/backends/native/meta-kms.h
index 743401406..f6b19520b 100644
--- a/src/backends/native/meta-kms.h
+++ b/src/backends/native/meta-kms.h
@@ -60,6 +60,8 @@ MetaKmsDevice * meta_kms_create_device (MetaKms            *kms,
                                         MetaKmsDeviceFlag   flags,
                                         GError            **error);
 
+gboolean meta_kms_is_shutting_down (MetaKms *kms);
+
 MetaKms * meta_kms_new (MetaBackend   *backend,
                         MetaKmsFlags   flags,
                         GError       **error);
diff --git a/src/backends/native/meta-onscreen-native.c b/src/backends/native/meta-onscreen-native.c
index 1a31f04a1..9836663d0 100644
--- a/src/backends/native/meta-onscreen-native.c
+++ b/src/backends/native/meta-onscreen-native.c
@@ -76,7 +76,7 @@ typedef struct _MetaOnscreenNativeSecondaryGpuState
 
   struct {
     MetaDrmBufferDumb *current_dumb_fb;
-    MetaDrmBufferDumb *dumb_fbs[2];
+    MetaDrmBufferDumb *dumb_fbs[3];
   } cpu;
 
   gboolean noted_primary_gpu_copy_ok;
@@ -98,9 +98,13 @@ struct _MetaOnscreenNative
   struct {
     struct gbm_surface *surface;
     MetaDrmBuffer *current_fb;
+    MetaDrmBuffer *posted_fb;
     MetaDrmBuffer *next_fb;
+    MetaDrmBuffer *stalled_fb;
     CoglScanout *current_scanout;
+    CoglScanout *posted_scanout;
     CoglScanout *next_scanout;
+    CoglScanout *stalled_scanout;
   } gbm;
 
 #ifdef HAVE_EGL_DEVICE
@@ -125,6 +129,16 @@ struct _MetaOnscreenNative
   gulong privacy_screen_changed_handler_id;
   gulong color_space_changed_handler_id;
   gulong hdr_metadata_changed_handler_id;
+
+  gboolean needs_flush;
+
+  unsigned int swaps_pending;
+
+  struct {
+    int *rectangles;  /* 4 x n_rectangles */
+    int n_rectangles;
+    ClutterFrame *frame;
+  } next_post;
 };
 
 G_DEFINE_TYPE (MetaOnscreenNative, meta_onscreen_native,
@@ -132,44 +146,42 @@ G_DEFINE_TYPE (MetaOnscreenNative, meta_onscreen_native,
 
 static GQuark blit_source_quark = 0;
 
+static void
+try_post_latest_swap (CoglOnscreen *onscreen);
+
+static void
+post_finish_frame (MetaOnscreenNative *onscreen_native,
+                   MetaKmsUpdate      *kms_update);
+
 static gboolean
 init_secondary_gpu_state (MetaRendererNative  *renderer_native,
                           CoglOnscreen        *onscreen,
                           GError             **error);
 
-static void
-free_current_bo (CoglOnscreen *onscreen)
-{
-  MetaOnscreenNative *onscreen_native = META_ONSCREEN_NATIVE (onscreen);
-
-  g_clear_object (&onscreen_native->gbm.current_fb);
-  g_clear_object (&onscreen_native->gbm.current_scanout);
-}
-
 static void
 meta_onscreen_native_swap_drm_fb (CoglOnscreen *onscreen)
 {
   MetaOnscreenNative *onscreen_native = META_ONSCREEN_NATIVE (onscreen);
 
-  if (!onscreen_native->gbm.next_fb)
+  if (!onscreen_native->gbm.posted_fb)
     return;
 
-  free_current_bo (onscreen);
+  g_set_object (&onscreen_native->gbm.current_fb,
+                onscreen_native->gbm.posted_fb);
+  g_clear_object (&onscreen_native->gbm.posted_fb);
 
-  g_set_object (&onscreen_native->gbm.current_fb, onscreen_native->gbm.next_fb);
-  g_clear_object (&onscreen_native->gbm.next_fb);
   g_set_object (&onscreen_native->gbm.current_scanout,
-                onscreen_native->gbm.next_scanout);
-  g_clear_object (&onscreen_native->gbm.next_scanout);
+                onscreen_native->gbm.posted_scanout);
+  g_clear_object (&onscreen_native->gbm.posted_scanout);
 }
 
 static void
-meta_onscreen_native_clear_next_fb (CoglOnscreen *onscreen)
+meta_onscreen_native_clear_posted_fb (CoglOnscreen *onscreen)
 {
   MetaOnscreenNative *onscreen_native = META_ONSCREEN_NATIVE (onscreen);
 
-  g_clear_object (&onscreen_native->gbm.next_fb);
-  g_clear_object (&onscreen_native->gbm.next_scanout);
+  g_clear_object (&onscreen_native->gbm.posted_fb);
+  g_clear_object (&onscreen_native->gbm.posted_scanout);
 }
 
 static void
@@ -207,7 +219,7 @@ meta_onscreen_native_notify_frame_complete (CoglOnscreen *onscreen)
 
   info = cogl_onscreen_pop_head_frame_info (onscreen);
 
-  g_assert (!cogl_onscreen_peek_head_frame_info (onscreen));
+  g_return_if_fail (info);
 
   _cogl_onscreen_notify_frame_sync (onscreen, info);
   _cogl_onscreen_notify_complete (onscreen, info);
@@ -243,6 +255,7 @@ notify_view_crtc_presented (MetaRendererView *view,
 
   meta_onscreen_native_notify_frame_complete (onscreen);
   meta_onscreen_native_swap_drm_fb (onscreen);
+  try_post_latest_swap (onscreen);
 }
 
 static void
@@ -292,15 +305,13 @@ page_flip_feedback_ready (MetaKmsCrtc *kms_crtc,
   CoglFramebuffer *framebuffer =
     clutter_stage_view_get_onscreen (CLUTTER_STAGE_VIEW (view));
   CoglOnscreen *onscreen = COGL_ONSCREEN (framebuffer);
-  MetaOnscreenNative *onscreen_native = META_ONSCREEN_NATIVE (onscreen);
   CoglFrameInfo *frame_info;
 
   frame_info = cogl_onscreen_peek_head_frame_info (onscreen);
   frame_info->flags |= COGL_FRAME_INFO_FLAG_SYMBOLIC;
 
-  g_warn_if_fail (!onscreen_native->gbm.next_fb);
-
   meta_onscreen_native_notify_frame_complete (onscreen);
+  try_post_latest_swap (onscreen);
 }
 
 static void
@@ -350,7 +361,8 @@ page_flip_feedback_discarded (MetaKmsCrtc  *kms_crtc,
   frame_info->flags |= COGL_FRAME_INFO_FLAG_SYMBOLIC;
 
   meta_onscreen_native_notify_frame_complete (onscreen);
-  meta_onscreen_native_clear_next_fb (onscreen);
+  meta_onscreen_native_clear_posted_fb (onscreen);
+  try_post_latest_swap (onscreen);
 }
 
 static const MetaKmsPageFlipListenerVtable page_flip_listener_vtable = {
@@ -411,18 +423,41 @@ custom_egl_stream_page_flip (gpointer custom_page_flip_data,
 }
 #endif /* HAVE_EGL_DEVICE */
 
-void
-meta_onscreen_native_dummy_power_save_page_flip (CoglOnscreen *onscreen)
+static void
+drop_stalled_swap (CoglOnscreen *onscreen)
 {
   CoglFrameInfo *frame_info;
+  MetaOnscreenNative *onscreen_native = META_ONSCREEN_NATIVE (onscreen);
 
-  meta_onscreen_native_swap_drm_fb (onscreen);
+  /* Remember we can't compare stalled_fb because it's not used by
+   * META_RENDERER_NATIVE_MODE_EGL_DEVICE. So we judge stalled to be whenever
+   * swaps_pending > 1.
+   */
+  if (onscreen_native->swaps_pending <= 1)
+    return;
+
+  onscreen_native->swaps_pending--;
+
+  g_clear_object (&onscreen_native->gbm.stalled_fb);
+  g_clear_object (&onscreen_native->gbm.stalled_scanout);
 
   frame_info = cogl_onscreen_peek_tail_frame_info (onscreen);
   frame_info->flags |= COGL_FRAME_INFO_FLAG_SYMBOLIC;
   meta_onscreen_native_notify_frame_complete (onscreen);
 }
 
+void
+meta_onscreen_native_dummy_power_save_page_flip (CoglOnscreen *onscreen)
+{
+  drop_stalled_swap (onscreen);
+
+  /* If the monitor just woke up and the shell is fully idle (has nothing
+   * more to swap) then we just woke to an indefinitely black screen. Let's
+   * fix that using the last swap (which is never classified as "stalled").
+   */
+  try_post_latest_swap (onscreen);
+}
+
 static void
 apply_transform (MetaCrtcKms            *crtc_kms,
                  MetaKmsPlaneAssignment *kms_plane_assignment,
@@ -521,13 +556,21 @@ meta_onscreen_native_flip_crtc (CoglOnscreen           *onscreen,
   switch (renderer_gpu_data->mode)
     {
     case META_RENDERER_NATIVE_MODE_GBM:
-      buffer = onscreen_native->gbm.next_fb;
+      g_set_object (&onscreen_native->gbm.posted_fb,
+                    onscreen_native->gbm.next_fb);
+      g_clear_object (&onscreen_native->gbm.next_fb);
+
+      buffer = onscreen_native->gbm.posted_fb;
 
-      if (onscreen_native->gbm.next_scanout)
+      g_set_object (&onscreen_native->gbm.posted_scanout,
+                    onscreen_native->gbm.next_scanout);
+      g_clear_object (&onscreen_native->gbm.next_scanout);
+
+      if (onscreen_native->gbm.posted_scanout)
         {
-          cogl_scanout_get_src_rect (onscreen_native->gbm.next_scanout,
+          cogl_scanout_get_src_rect (onscreen_native->gbm.posted_scanout,
                                      &src_rect);
-          cogl_scanout_get_dst_rect (onscreen_native->gbm.next_scanout,
+          cogl_scanout_get_dst_rect (onscreen_native->gbm.posted_scanout,
                                      &dst_rect);
         }
       else
@@ -918,12 +961,17 @@ static MetaDrmBufferDumb *
 secondary_gpu_get_next_dumb_buffer (MetaOnscreenNativeSecondaryGpuState *secondary_gpu_state)
 {
   MetaDrmBufferDumb *current_dumb_fb;
+  const int n_dumb_fbs = G_N_ELEMENTS (secondary_gpu_state->cpu.dumb_fbs);
+  int i;
 
   current_dumb_fb = secondary_gpu_state->cpu.current_dumb_fb;
-  if (current_dumb_fb == secondary_gpu_state->cpu.dumb_fbs[0])
-    return secondary_gpu_state->cpu.dumb_fbs[1];
-  else
-    return secondary_gpu_state->cpu.dumb_fbs[0];
+  for (i = 0; i < n_dumb_fbs; i++)
+    {
+      if (current_dumb_fb == secondary_gpu_state->cpu.dumb_fbs[i])
+        return secondary_gpu_state->cpu.dumb_fbs[(i + 1) % n_dumb_fbs];
+    }
+
+  return secondary_gpu_state->cpu.dumb_fbs[0];
 }
 
 static MetaDrmBuffer *
@@ -1255,10 +1303,17 @@ swap_buffer_result_feedback (const MetaKmsFeedback *kms_feedback,
     g_warning ("Page flip failed: %s", error->message);
 
   frame_info = cogl_onscreen_peek_head_frame_info (onscreen);
-  frame_info->flags |= COGL_FRAME_INFO_FLAG_SYMBOLIC;
 
-  meta_onscreen_native_notify_frame_complete (onscreen);
-  meta_onscreen_native_clear_next_fb (onscreen);
+  /* After resuming from suspend, drop_stalled_swap might have done this
+   * already and emptied the frame_info queue.
+   */
+  if (frame_info)
+    {
+      frame_info->flags |= COGL_FRAME_INFO_FLAG_SYMBOLIC;
+      meta_onscreen_native_notify_frame_complete (onscreen);
+    }
+
+  meta_onscreen_native_clear_posted_fb (onscreen);
 }
 
 static const MetaKmsResultListenerVtable swap_buffer_result_listener_vtable = {
@@ -1279,32 +1334,37 @@ meta_onscreen_native_swap_buffers_with_damage (CoglOnscreen  *onscreen,
   CoglRendererEGL *cogl_renderer_egl = cogl_renderer->winsys;
   MetaRendererNativeGpuData *renderer_gpu_data = cogl_renderer_egl->platform;
   MetaRendererNative *renderer_native = renderer_gpu_data->renderer_native;
-  MetaRenderer *renderer = META_RENDERER (renderer_native);
-  MetaBackend *backend = meta_renderer_get_backend (renderer);
-  MetaMonitorManager *monitor_manager =
-    meta_backend_get_monitor_manager (backend);
   MetaOnscreenNative *onscreen_native = META_ONSCREEN_NATIVE (onscreen);
   MetaOnscreenNativeSecondaryGpuState *secondary_gpu_state;
   MetaGpuKms *render_gpu = onscreen_native->render_gpu;
   MetaDeviceFile *render_device_file;
   ClutterFrame *frame = user_data;
-  MetaFrameNative *frame_native = meta_frame_native_from_frame (frame);
-  MetaKmsUpdate *kms_update;
   CoglOnscreenClass *parent_class;
   gboolean create_timestamp_query = TRUE;
   gboolean egl_context_changed = FALSE;
-  MetaPowerSave power_save_mode;
   g_autoptr (GError) error = NULL;
   MetaDrmBufferFlags buffer_flags;
   MetaDrmBufferGbm *buffer_gbm;
   g_autoptr (MetaDrmBuffer) primary_gpu_fb = NULL;
   g_autoptr (MetaDrmBuffer) secondary_gpu_fb = NULL;
-  MetaKmsCrtc *kms_crtc;
-  MetaKmsDevice *kms_device;
+  size_t rectangles_size;
 
   COGL_TRACE_BEGIN_SCOPED (MetaRendererNativeSwapBuffers,
                            "Meta::OnscreenNative::swap_buffers_with_damage()");
 
+  if (meta_is_topic_enabled (META_DEBUG_KMS))
+    {
+      unsigned int frames_pending =
+        cogl_onscreen_count_pending_frames (onscreen);
+
+      meta_topic (META_DEBUG_KMS,
+                  "Swap buffers: %u frames pending (%s-buffering)",
+                  frames_pending,
+                  frames_pending == 1 ? "double" :
+                  frames_pending == 2 ? "triple" :
+                  "?");
+    }
+
   secondary_gpu_fb =
     update_secondary_gpu_state_pre_swap_buffers (onscreen,
                                                  rectangles,
@@ -1379,7 +1439,17 @@ meta_onscreen_native_swap_buffers_with_damage (CoglOnscreen  *onscreen,
   switch (renderer_gpu_data->mode)
     {
     case META_RENDERER_NATIVE_MODE_GBM:
-      g_warn_if_fail (onscreen_native->gbm.next_fb == NULL);
+      if (onscreen_native->gbm.next_fb != NULL)
+        {
+          g_warn_if_fail (onscreen_native->gbm.stalled_fb == NULL);
+          drop_stalled_swap (onscreen);
+          g_assert (onscreen_native->gbm.stalled_fb == NULL);
+          onscreen_native->gbm.stalled_fb =
+            g_steal_pointer (&onscreen_native->gbm.next_fb);
+          onscreen_native->gbm.stalled_scanout =
+            g_steal_pointer (&onscreen_native->gbm.next_scanout);
+        }
+
       if (onscreen_native->secondary_gpu_state)
         g_set_object (&onscreen_native->gbm.next_fb, secondary_gpu_fb);
       else
@@ -1404,6 +1474,9 @@ meta_onscreen_native_swap_buffers_with_damage (CoglOnscreen  *onscreen,
 #endif
     }
 
+  clutter_frame_set_result (frame,
+                            CLUTTER_FRAME_RESULT_PENDING_PRESENTED);
+
   /*
    * If we changed EGL context, cogl will have the wrong idea about what is
    * current, making it fail to set it when it needs to. Avoid that by making
@@ -1413,12 +1486,78 @@ meta_onscreen_native_swap_buffers_with_damage (CoglOnscreen  *onscreen,
   if (egl_context_changed)
     _cogl_winsys_egl_ensure_current (cogl_display);
 
-  kms_crtc = meta_crtc_kms_get_kms_crtc (META_CRTC_KMS (onscreen_native->crtc));
-  kms_device = meta_kms_crtc_get_device (kms_crtc);
+  rectangles_size = n_rectangles * 4 * sizeof (int);
+  onscreen_native->next_post.rectangles =
+    g_realloc (onscreen_native->next_post.rectangles, rectangles_size);
+  memcpy (onscreen_native->next_post.rectangles, rectangles, rectangles_size);
+  onscreen_native->next_post.n_rectangles = n_rectangles;
+
+  g_clear_pointer (&onscreen_native->next_post.frame, clutter_frame_unref);
+  onscreen_native->next_post.frame = clutter_frame_ref (frame);
+
+  onscreen_native->swaps_pending++;
+  try_post_latest_swap (onscreen);
+}
+
+static void
+try_post_latest_swap (CoglOnscreen *onscreen)
+{
+  CoglFramebuffer *framebuffer = COGL_FRAMEBUFFER (onscreen);
+  CoglContext *cogl_context = cogl_framebuffer_get_context (framebuffer);
+  CoglRenderer *cogl_renderer = cogl_context->display->renderer;
+  CoglRendererEGL *cogl_renderer_egl = cogl_renderer->winsys;
+  MetaRendererNativeGpuData *renderer_gpu_data = cogl_renderer_egl->platform;
+  MetaRendererNative *renderer_native = renderer_gpu_data->renderer_native;
+  MetaRenderer *renderer = META_RENDERER (renderer_native);
+  MetaBackend *backend = meta_renderer_get_backend (renderer);
+  MetaBackendNative *backend_native = META_BACKEND_NATIVE (backend);
+  MetaKms *kms = meta_backend_native_get_kms (backend_native);
+  MetaMonitorManager *monitor_manager =
+    meta_backend_get_monitor_manager (backend);
+  MetaOnscreenNative *onscreen_native = META_ONSCREEN_NATIVE (onscreen);
+  MetaPowerSave power_save_mode;
+  MetaCrtcKms *crtc_kms = META_CRTC_KMS (onscreen_native->crtc);
+  MetaKmsCrtc *kms_crtc = meta_crtc_kms_get_kms_crtc (crtc_kms);
+  MetaKmsDevice *kms_device = meta_kms_crtc_get_device (kms_crtc);
+  MetaKmsUpdate *kms_update;
+  g_autoptr (MetaKmsFeedback) kms_feedback = NULL;
+  g_autoptr (ClutterFrame) frame = NULL;
+  MetaFrameNative *frame_native;
+
+  if (onscreen_native->next_post.frame == NULL ||
+      onscreen_native->view == NULL ||
+      meta_kms_is_shutting_down (kms))
+    return;
 
   power_save_mode = meta_monitor_manager_get_power_save_mode (monitor_manager);
   if (power_save_mode == META_POWER_SAVE_ON)
     {
+      unsigned int frames_pending =
+        cogl_onscreen_count_pending_frames (onscreen);
+      unsigned int posts_pending;
+
+      g_assert (frames_pending >= onscreen_native->swaps_pending);
+      posts_pending = frames_pending - onscreen_native->swaps_pending;
+      if (posts_pending > 0)
+        return;  /* wait for the next frame notification and then try again */
+
+      frame = g_steal_pointer (&onscreen_native->next_post.frame);
+      frame_native = meta_frame_native_from_frame (frame);
+
+      if (onscreen_native->swaps_pending == 0)
+        {
+          if (frame_native)
+            {
+              kms_update = meta_frame_native_steal_kms_update (frame_native);
+              if (kms_update)
+                post_finish_frame (onscreen_native, kms_update);
+            }
+          return;
+        }
+
+      drop_stalled_swap (onscreen);
+      onscreen_native->swaps_pending--;
+
       kms_update = meta_frame_native_ensure_kms_update (frame_native,
                                                         kms_device);
       meta_kms_update_add_result_listener (kms_update,
@@ -1433,15 +1572,13 @@ meta_onscreen_native_swap_buffers_with_damage (CoglOnscreen  *onscreen,
                                       onscreen_native->crtc,
                                       kms_update,
                                       META_KMS_ASSIGN_PLANE_FLAG_NONE,
-                                      rectangles,
-                                      n_rectangles);
+                                      onscreen_native->next_post.rectangles,
+                                      onscreen_native->next_post.n_rectangles);
     }
   else
     {
       meta_renderer_native_queue_power_save_page_flip (renderer_native,
                                                        onscreen);
-      clutter_frame_set_result (frame,
-                                CLUTTER_FRAME_RESULT_PENDING_PRESENTED);
       return;
     }
 
@@ -1461,8 +1598,6 @@ meta_onscreen_native_swap_buffers_with_damage (CoglOnscreen  *onscreen,
           kms_update = meta_frame_native_steal_kms_update (frame_native);
           meta_renderer_native_queue_mode_set_update (renderer_native,
                                                       kms_update);
-          clutter_frame_set_result (frame,
-                                    CLUTTER_FRAME_RESULT_PENDING_PRESENTED);
           return;
         }
       else if (meta_renderer_native_has_pending_mode_set (renderer_native))
@@ -1476,8 +1611,6 @@ meta_onscreen_native_swap_buffers_with_damage (CoglOnscreen  *onscreen,
 
           meta_frame_native_steal_kms_update (frame_native);
           meta_renderer_native_post_mode_set_updates (renderer_native);
-          clutter_frame_set_result (frame,
-                                    CLUTTER_FRAME_RESULT_PENDING_PRESENTED);
           return;
         }
       break;
@@ -1493,8 +1626,6 @@ meta_onscreen_native_swap_buffers_with_damage (CoglOnscreen  *onscreen,
                                                       kms_update);
 
           meta_renderer_native_post_mode_set_updates (renderer_native);
-          clutter_frame_set_result (frame,
-                                    CLUTTER_FRAME_RESULT_PENDING_PRESENTED);
           return;
         }
       break;
@@ -1509,7 +1640,6 @@ meta_onscreen_native_swap_buffers_with_damage (CoglOnscreen  *onscreen,
   kms_update = meta_frame_native_steal_kms_update (frame_native);
   meta_kms_device_post_update (kms_device, kms_update,
                                META_KMS_UPDATE_FLAG_NONE);
-  clutter_frame_set_result (frame, CLUTTER_FRAME_RESULT_PENDING_PRESENTED);
 }
 
 gboolean
@@ -1580,7 +1710,7 @@ scanout_result_feedback (const MetaKmsFeedback *kms_feedback,
 
       g_warning ("Direct scanout page flip failed: %s", error->message);
 
-      cogl_scanout_notify_failed (onscreen_native->gbm.next_scanout,
+      cogl_scanout_notify_failed (onscreen_native->gbm.posted_scanout,
                                   onscreen);
       clutter_stage_view_add_redraw_clip (view, NULL);
       clutter_stage_view_schedule_update_now (view);
@@ -1590,7 +1720,7 @@ scanout_result_feedback (const MetaKmsFeedback *kms_feedback,
   frame_info->flags |= COGL_FRAME_INFO_FLAG_SYMBOLIC;
 
   meta_onscreen_native_notify_frame_complete (onscreen);
-  meta_onscreen_native_clear_next_fb (onscreen);
+  meta_onscreen_native_clear_posted_fb (onscreen);
 }
 
 static const MetaKmsResultListenerVtable scanout_result_listener_vtable = {
@@ -1642,6 +1772,18 @@ meta_onscreen_native_direct_scanout (CoglOnscreen   *onscreen,
       return FALSE;
     }
 
+  /* Our direct scanout frame counts as 1, so more than that means we would
+   * be jumping the queue (and post would fail).
+   */
+  if (cogl_onscreen_count_pending_frames (onscreen) > 1)
+    {
+      g_set_error_literal (error,
+                           COGL_SCANOUT_ERROR,
+                           COGL_SCANOUT_ERROR_INHIBITED,
+                           "Direct scanout is inhibited during triple buffering");
+      return FALSE;
+    }
+
   renderer_gpu_data = meta_renderer_native_get_gpu_data (renderer_native,
                                                          render_gpu);
 
@@ -1757,11 +1899,7 @@ meta_onscreen_native_before_redraw (CoglOnscreen *onscreen,
                                     ClutterFrame *frame)
 {
   MetaOnscreenNative *onscreen_native = META_ONSCREEN_NATIVE (onscreen);
-  MetaCrtcKms *crtc_kms = META_CRTC_KMS (onscreen_native->crtc);
-  MetaKmsCrtc *kms_crtc = meta_crtc_kms_get_kms_crtc (crtc_kms);
 
-  meta_kms_device_await_flush (meta_kms_crtc_get_device (kms_crtc),
-                               kms_crtc);
   maybe_update_frame_sync (onscreen_native, frame);
 }
 
@@ -1877,22 +2015,79 @@ meta_onscreen_native_finish_frame (CoglOnscreen *onscreen,
   MetaKmsDevice *kms_device = meta_kms_crtc_get_device (kms_crtc);
   MetaFrameNative *frame_native = meta_frame_native_from_frame (frame);
   MetaKmsUpdate *kms_update;
+  unsigned int frames_pending = cogl_onscreen_count_pending_frames (onscreen);
+  unsigned int swaps_pending = onscreen_native->swaps_pending;
+  unsigned int posts_pending = frames_pending - swaps_pending;
 
-  kms_update = meta_frame_native_steal_kms_update (frame_native);
-  if (!kms_update)
+  onscreen_native->needs_flush |= meta_kms_device_handle_flush (kms_device,
+                                                                kms_crtc);
+
+  if (!meta_frame_native_has_kms_update (frame_native))
     {
-      if (meta_kms_device_handle_flush (kms_device, kms_crtc))
-        {
-          kms_update = meta_kms_update_new (kms_device);
-          meta_kms_update_set_flushing (kms_update, kms_crtc);
-        }
-      else
+      if (!onscreen_native->needs_flush || posts_pending)
         {
           clutter_frame_set_result (frame, CLUTTER_FRAME_RESULT_IDLE);
           return;
         }
     }
 
+  if (posts_pending && !swaps_pending)
+    {
+      g_return_if_fail (meta_frame_native_has_kms_update (frame_native));
+      g_warn_if_fail (onscreen_native->next_post.frame == NULL);
+
+      g_clear_pointer (&onscreen_native->next_post.frame, clutter_frame_unref);
+      onscreen_native->next_post.frame = clutter_frame_ref (frame);
+      clutter_frame_set_result (frame, CLUTTER_FRAME_RESULT_PENDING_PRESENTED);
+      return;
+    }
+
+  kms_update = meta_frame_native_steal_kms_update (frame_native);
+
+  if (posts_pending && swaps_pending)
+    {
+      MetaFrameNative *older_frame_native;
+      MetaKmsUpdate *older_kms_update;
+
+      g_return_if_fail (kms_update);
+      g_return_if_fail (onscreen_native->next_post.frame != NULL);
+
+      older_frame_native =
+        meta_frame_native_from_frame (onscreen_native->next_post.frame);
+      older_kms_update =
+        meta_frame_native_ensure_kms_update (older_frame_native, kms_device);
+      meta_kms_update_merge_from (older_kms_update, kms_update);
+      meta_kms_update_free (kms_update);
+      clutter_frame_set_result (frame, CLUTTER_FRAME_RESULT_IDLE);
+      return;
+    }
+
+  if (!kms_update)
+    {
+      kms_update = meta_kms_update_new (kms_device);
+      g_warn_if_fail (onscreen_native->needs_flush);
+    }
+
+  if (onscreen_native->needs_flush)
+    {
+      meta_kms_update_set_flushing (kms_update, kms_crtc);
+      onscreen_native->needs_flush = FALSE;
+    }
+
+  post_finish_frame (onscreen_native, kms_update);
+
+  clutter_frame_set_result (frame, CLUTTER_FRAME_RESULT_PENDING_PRESENTED);
+}
+
+static void
+post_finish_frame (MetaOnscreenNative *onscreen_native,
+                   MetaKmsUpdate      *kms_update)
+{
+  MetaCrtc *crtc = onscreen_native->crtc;
+  MetaKmsCrtc *kms_crtc = meta_crtc_kms_get_kms_crtc (META_CRTC_KMS (crtc));
+  MetaKmsDevice *kms_device = meta_kms_crtc_get_device (kms_crtc);
+  g_autoptr (MetaKmsFeedback) kms_feedback = NULL;
+
   meta_kms_update_add_result_listener (kms_update,
                                        &finish_frame_result_listener_vtable,
                                        NULL,
@@ -1915,7 +2110,19 @@ meta_onscreen_native_finish_frame (CoglOnscreen *onscreen,
   meta_kms_update_set_flushing (kms_update, kms_crtc);
   meta_kms_device_post_update (kms_device, kms_update,
                                META_KMS_UPDATE_FLAG_NONE);
-  clutter_frame_set_result (frame, CLUTTER_FRAME_RESULT_PENDING_PRESENTED);
+}
+
+void
+meta_onscreen_native_discard_pending_swaps (CoglOnscreen *onscreen)
+{
+  MetaOnscreenNative *onscreen_native = META_ONSCREEN_NATIVE (onscreen);
+
+  onscreen_native->swaps_pending = 0;
+
+  g_clear_object (&onscreen_native->gbm.stalled_fb);
+  g_clear_object (&onscreen_native->gbm.stalled_scanout);
+  g_clear_object (&onscreen_native->gbm.next_fb);
+  g_clear_object (&onscreen_native->gbm.next_scanout);
 }
 
 static gboolean
@@ -2830,8 +3037,11 @@ meta_onscreen_native_dispose (GObject *object)
     {
     case META_RENDERER_NATIVE_MODE_GBM:
       g_clear_object (&onscreen_native->gbm.next_fb);
+      g_clear_object (&onscreen_native->gbm.posted_fb);
+      g_clear_object (&onscreen_native->gbm.current_fb);
       g_clear_object (&onscreen_native->gbm.next_scanout);
-      free_current_bo (onscreen);
+      g_clear_object (&onscreen_native->gbm.posted_scanout);
+      g_clear_object (&onscreen_native->gbm.current_scanout);
       break;
     case META_RENDERER_NATIVE_MODE_SURFACELESS:
       g_assert_not_reached ();
@@ -2865,6 +3075,10 @@ meta_onscreen_native_dispose (GObject *object)
 
   g_clear_object (&onscreen_native->output);
   g_clear_object (&onscreen_native->crtc);
+
+  g_clear_pointer (&onscreen_native->next_post.rectangles, g_free);
+  g_clear_pointer (&onscreen_native->next_post.frame, clutter_frame_unref);
+  onscreen_native->next_post.n_rectangles = 0;
 }
 
 static void
diff --git a/src/backends/native/meta-onscreen-native.h b/src/backends/native/meta-onscreen-native.h
index 0e1193325..e30357d19 100644
--- a/src/backends/native/meta-onscreen-native.h
+++ b/src/backends/native/meta-onscreen-native.h
@@ -48,6 +48,8 @@ void meta_onscreen_native_dummy_power_save_page_flip (CoglOnscreen *onscreen);
 gboolean meta_onscreen_native_is_buffer_scanout_compatible (CoglOnscreen *onscreen,
                                                             CoglScanout  *scanout);
 
+void meta_onscreen_native_discard_pending_swaps (CoglOnscreen *onscreen);
+
 void meta_onscreen_native_set_view (CoglOnscreen     *onscreen,
                                     MetaRendererView *view);
 
diff --git a/src/backends/native/meta-renderer-native.c b/src/backends/native/meta-renderer-native.c
index aa76d018c..3c22b4e86 100644
--- a/src/backends/native/meta-renderer-native.c
+++ b/src/backends/native/meta-renderer-native.c
@@ -731,12 +731,18 @@ static gboolean
 dummy_power_save_page_flip_cb (gpointer user_data)
 {
   MetaRendererNative *renderer_native = user_data;
+  GList *old_list =
+    g_steal_pointer (&renderer_native->power_save_page_flip_onscreens);
 
-  g_list_foreach (renderer_native->power_save_page_flip_onscreens,
+  g_list_foreach (old_list,
                   (GFunc) meta_onscreen_native_dummy_power_save_page_flip,
                   NULL);
-  g_clear_list (&renderer_native->power_save_page_flip_onscreens,
+  g_clear_list (&old_list,
                 g_object_unref);
+
+  if (renderer_native->power_save_page_flip_onscreens != NULL)
+    return G_SOURCE_CONTINUE;
+
   renderer_native->power_save_page_flip_source_id = 0;
 
   return G_SOURCE_REMOVE;
@@ -748,6 +754,9 @@ meta_renderer_native_queue_power_save_page_flip (MetaRendererNative *renderer_na
 {
   const unsigned int timeout_ms = 100;
 
+  if (g_list_find (renderer_native->power_save_page_flip_onscreens, onscreen))
+    return;
+
   if (!renderer_native->power_save_page_flip_source_id)
     {
       renderer_native->power_save_page_flip_source_id =
@@ -1529,6 +1538,26 @@ detach_onscreens (MetaRenderer *renderer)
     }
 }
 
+static void
+discard_pending_swaps (MetaRenderer *renderer)
+{
+  GList *views = meta_renderer_get_views (renderer);;
+  GList *l;
+
+  for (l = views; l; l = l->next)
+    {
+      ClutterStageView *stage_view = l->data;
+      CoglFramebuffer *fb = clutter_stage_view_get_onscreen (stage_view);
+      CoglOnscreen *onscreen;
+
+      if (!COGL_IS_ONSCREEN (fb))
+        continue;
+
+      onscreen = COGL_ONSCREEN (fb);
+      meta_onscreen_native_discard_pending_swaps (onscreen);
+    }
+}
+
 static void
 meta_renderer_native_rebuild_views (MetaRenderer *renderer)
 {
@@ -1539,6 +1568,7 @@ meta_renderer_native_rebuild_views (MetaRenderer *renderer)
   MetaRendererClass *parent_renderer_class =
     META_RENDERER_CLASS (meta_renderer_native_parent_class);
 
+  discard_pending_swaps (renderer);
   meta_kms_discard_pending_page_flips (kms);
   g_hash_table_remove_all (renderer_native->mode_set_updates);
 
diff --git a/src/tests/native-kms-render.c b/src/tests/native-kms-render.c
index f5ebc23fe..2f870fdc3 100644
--- a/src/tests/native-kms-render.c
+++ b/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),
   };