Simplify the fencing code and differentiate between flushes and

waiting types.
Add a "command_stream_barrier" method to the bo driver.

15 files changed, 537 insertions, 480 deletions

diff --git a/linux-core/drm_bo.c b/linux-core/drm_bo.c
index df10e12b..3b180d15 100644
--- a/linux-core/drm_bo.c
+++ b/linux-core/drm_bo.c
@@ -287,7 +287,7 @@ int drm_bo_wait(struct drm_buffer_object *bo, int lazy, int ignore_signals,
 	DRM_ASSERT_LOCKED(&bo->mutex);
 
 	if (bo->fence) {
-		if (drm_fence_object_signaled(bo->fence, bo->fence_type, 0)) {
+		if (drm_fence_object_signaled(bo->fence, bo->fence_type)) {
 			drm_fence_usage_deref_unlocked(&bo->fence);
 			return 0;
 		}
@@ -354,7 +354,7 @@ static void drm_bo_cleanup_refs(struct drm_buffer_object *bo, int remove_all)
 	DRM_FLAG_MASKED(bo->priv_flags, 0, _DRM_BO_FLAG_UNFENCED);
 
 	if (bo->fence && drm_fence_object_signaled(bo->fence,
-						   bo->fence_type, 0))
+						   bo->fence_type))
 		drm_fence_usage_deref_unlocked(&bo->fence);
 
 	if (bo->fence && remove_all)
@@ -559,7 +559,7 @@ void drm_putback_buffer_objects(struct drm_device *dev)
 
 		list_del_init(&entry->lru);
 		DRM_FLAG_MASKED(entry->priv_flags, 0, _DRM_BO_FLAG_UNFENCED);
-		DRM_WAKEUP(&entry->event_queue);
+		wake_up_all(&entry->event_queue);
 
 		/*
 		 * FIXME: Might want to put back on head of list
@@ -660,7 +660,7 @@ int drm_fence_buffer_objects(struct drm_device *dev,
 			entry->fence_type = entry->new_fence_type;
 			DRM_FLAG_MASKED(entry->priv_flags, 0,
 					_DRM_BO_FLAG_UNFENCED);
-			DRM_WAKEUP(&entry->event_queue);
+			wake_up_all(&entry->event_queue);
 			drm_bo_add_to_lru(entry);
 		}
 		mutex_unlock(&entry->mutex);
@@ -1032,7 +1032,7 @@ static int drm_bo_quick_busy(struct drm_buffer_object *bo)
 
 	BUG_ON(bo->priv_flags & _DRM_BO_FLAG_UNFENCED);
 	if (fence) {
-		if (drm_fence_object_signaled(fence, bo->fence_type, 0)) {
+		if (drm_fence_object_signaled(fence, bo->fence_type)) {
 			drm_fence_usage_deref_unlocked(&bo->fence);
 			return 0;
 		}
@@ -1052,12 +1052,12 @@ static int drm_bo_busy(struct drm_buffer_object *bo)
 
 	BUG_ON(bo->priv_flags & _DRM_BO_FLAG_UNFENCED);
 	if (fence) {
-		if (drm_fence_object_signaled(fence, bo->fence_type, 0)) {
+		if (drm_fence_object_signaled(fence, bo->fence_type)) {
 			drm_fence_usage_deref_unlocked(&bo->fence);
 			return 0;
 		}
 		drm_fence_object_flush(fence, DRM_FENCE_TYPE_EXE);
-		if (drm_fence_object_signaled(fence, bo->fence_type, 0)) {
+		if (drm_fence_object_signaled(fence, bo->fence_type)) {
 			drm_fence_usage_deref_unlocked(&bo->fence);
 			return 0;
 		}
@@ -1249,7 +1249,7 @@ static int drm_buffer_object_map(struct drm_file *file_priv, uint32_t handle,
 	mutex_unlock(&dev->struct_mutex);
 	if (ret) {
 		if (atomic_add_negative(-1, &bo->mapped))
-			DRM_WAKEUP(&bo->event_queue);
+			wake_up_all(&bo->event_queue);
 
 	} else
 		drm_bo_fill_rep_arg(bo, rep);
@@ -1306,7 +1306,7 @@ static void drm_buffer_user_object_unmap(struct drm_file *file_priv,
 	BUG_ON(action != _DRM_REF_TYPE1);
 
 	if (atomic_add_negative(-1, &bo->mapped))
-		DRM_WAKEUP(&bo->event_queue);
+		wake_up_all(&bo->event_queue);
 }
 
 /*
@@ -1364,7 +1364,7 @@ out_unlock:
 		}
 		drm_bo_add_to_lru(bo);
 		if (bo->priv_flags & _DRM_BO_FLAG_UNFENCED) {
-			DRM_WAKEUP(&bo->event_queue);
+			wake_up_all(&bo->event_queue);
 			DRM_FLAG_MASKED(bo->priv_flags, 0,
 					_DRM_BO_FLAG_UNFENCED);
 		}
@@ -1442,13 +1442,21 @@ static int drm_buffer_object_validate(struct drm_buffer_object *bo,
 	 * We're switching command submission mechanism,
 	 * or cannot simply rely on the hardware serializing for us.
 	 *
-	 * Wait for buffer idle.
+	 * Insert a driver-dependant barrier or wait for buffer idle.
 	 */
 
 	if ((fence_class != bo->fence_class) ||
 	    ((ftype ^ bo->fence_type) & bo->fence_type)) {
 
-		ret = drm_bo_wait(bo, 0, 0, no_wait);
+		ret = -EINVAL;
+		if (driver->command_stream_barrier) {
+			ret = driver->command_stream_barrier(bo,
+							     fence_class,
+							     ftype,
+							     no_wait);
+		}
+		if (ret)
+			ret = drm_bo_wait(bo, 0, 0, no_wait);
 
 		if (ret)
 			return ret;
@@ -1539,7 +1547,7 @@ static int drm_buffer_object_validate(struct drm_buffer_object *bo,
 	} else {
 		drm_bo_add_to_lru(bo);
 		if (bo->priv_flags & _DRM_BO_FLAG_UNFENCED) {
-			DRM_WAKEUP(&bo->event_queue);
+			wake_up_all(&bo->event_queue);
 			DRM_FLAG_MASKED(bo->priv_flags, 0,
 					_DRM_BO_FLAG_UNFENCED);
 		}
diff --git a/linux-core/drm_fence.c b/linux-core/drm_fence.c
index 288b4db6..a852c63f 100644
--- a/linux-core/drm_fence.c
+++ b/linux-core/drm_fence.c
@@ -30,6 +30,57 @@
 
 #include "drmP.h"
 
+
+/*
+ * Convenience function to be called by fence::wait methods that
+ * need polling.
+ */
+
+int drm_fence_wait_polling(struct drm_fence_object *fence, int lazy,
+			   int interruptible, uint32_t mask, 
+			   unsigned long end_jiffies)
+{
+	struct drm_device *dev = fence->dev;
+	struct drm_fence_manager *fm = &dev->fm;
+	struct drm_fence_class_manager *fc = &fm->fence_class[fence->fence_class];
+	uint32_t count = 0;
+	int ret;
+
+	DECLARE_WAITQUEUE(entry, current);
+	add_wait_queue(&fc->fence_queue, &entry);
+
+	ret = 0;
+	
+	for (;;) {
+		__set_current_state((interruptible) ? 
+				    TASK_INTERRUPTIBLE :
+				    TASK_UNINTERRUPTIBLE);
+		if (drm_fence_object_signaled(fence, mask))
+			break;
+		if (time_after_eq(jiffies, end_jiffies)) {
+			ret = -EBUSY;
+			break;
+		}
+		if (lazy)
+			schedule_timeout(1);
+		else if ((++count & 0x0F) == 0){
+			__set_current_state(TASK_RUNNING);
+			schedule();
+			__set_current_state((interruptible) ? 
+					    TASK_INTERRUPTIBLE :
+					    TASK_UNINTERRUPTIBLE);
+		}			
+		if (interruptible && signal_pending(current)) {
+			ret = -EAGAIN;
+			break;
+		}
+	}
+	__set_current_state(TASK_RUNNING);
+	remove_wait_queue(&fc->fence_queue, &entry);
+	return ret;
+}
+EXPORT_SYMBOL(drm_fence_wait_polling);
+
 /*
  * Typically called by the IRQ handler.
  */
@@ -39,27 +90,14 @@ void drm_fence_handler(struct drm_device *dev, uint32_t fence_class,
 {
 	int wake = 0;
 	uint32_t diff;
-	uint32_t relevant;
+	uint32_t relevant_type;
+	uint32_t new_type;
 	struct drm_fence_manager *fm = &dev->fm;
 	struct drm_fence_class_manager *fc = &fm->fence_class[fence_class];
 	struct drm_fence_driver *driver = dev->driver->fence_driver;
 	struct list_head *head;
 	struct drm_fence_object *fence, *next;
 	int found = 0;
-	int is_exe = (type & DRM_FENCE_TYPE_EXE);
-	int ge_last_exe;
-
-
-	diff = (sequence - fc->exe_flush_sequence) & driver->sequence_mask;
-
-	if (fc->pending_exe_flush && is_exe && diff < driver->wrap_diff)
-		fc->pending_exe_flush = 0;
-
-	diff = (sequence - fc->last_exe_flush) & driver->sequence_mask;
-	ge_last_exe = diff < driver->wrap_diff;
-
-	if (is_exe && ge_last_exe)
-		fc->last_exe_flush = sequence;
 
 	if (list_empty(&fc->ring))
 		return;
@@ -72,7 +110,7 @@ void drm_fence_handler(struct drm_device *dev, uint32_t fence_class,
 		}
 	}
 
-	fc->pending_flush &= ~type;
+	fc->waiting_types &= ~type;
 	head = (found) ? &fence->ring : &fc->ring;
 
 	list_for_each_entry_safe_reverse(fence, next, head, ring) {
@@ -81,64 +119,60 @@ void drm_fence_handler(struct drm_device *dev, uint32_t fence_class,
 
 		if (error) {
 			fence->error = error;
-			fence->signaled = fence->type;
-			fence->submitted_flush = fence->type;
-			fence->flush_mask = fence->type;
+			fence->signaled_types = fence->type;
 			list_del_init(&fence->ring);
 			wake = 1;
 			break;
 		}
 
-		if (is_exe)
-			type |= fence->native_type;
+		if (type & DRM_FENCE_TYPE_EXE)
+			type |= fence->native_types;
 
-		relevant = type & fence->type;
+		relevant_type = type & fence->type;
+		new_type = (fence->signaled_types | relevant_type) & 
+			~fence->signaled_types;
 
-		if ((fence->signaled | relevant) != fence->signaled) {
-			fence->signaled |= relevant;
-			fence->flush_mask |= relevant;
-			fence->submitted_flush |= relevant;
+		if (new_type) {
+			fence->signaled_types |= new_type;
 			DRM_DEBUG("Fence 0x%08lx signaled 0x%08x\n",
-				  fence->base.hash.key, fence->signaled);
-			wake = 1;
-		}
+				  fence->base.hash.key, fence->signaled_types);
+
+			if (driver->needed_flush)
+				fc->pending_flush |= driver->needed_flush(fence);
 
-		relevant = fence->flush_mask &
-			~(fence->submitted_flush | fence->signaled);
+			if (new_type & fence->waiting_types)
+				wake = 1;
+		}
 
-		fc->pending_flush |= relevant;
-		fence->submitted_flush |= relevant;
+		fc->waiting_types |= fence->waiting_types & ~fence->signaled_types;
 
-		if (!(fence->type & ~fence->signaled)) {
+		if (!(fence->type & ~fence->signaled_types)) {
 			DRM_DEBUG("Fence completely signaled 0x%08lx\n",
 				  fence->base.hash.key);
 			list_del_init(&fence->ring);
 		}
-
 	}
 
 	/*
-	 * Reinstate lost flush flags.
+	 * Reinstate lost waiting types.
 	 */
 
-	if ((fc->pending_flush & type) != type) {
+	if ((fc->waiting_types & type) != type) {
 		head = head->prev;
 		list_for_each_entry(fence, head, ring) {
 			if (&fence->ring == &fc->ring)
 				break;
-			diff = (fc->last_exe_flush - fence->sequence) &
+			diff = (fc->highest_waiting_sequence - fence->sequence) &
 				driver->sequence_mask;
 			if (diff > driver->wrap_diff)
 				break;
-
-			relevant = fence->submitted_flush & ~fence->signaled;
-			fc->pending_flush |= relevant;
+			
+			fc->waiting_types |= fence->waiting_types & ~fence->signaled_types;
 		}
 	}
 
-	if (wake) {
-		DRM_WAKEUP(&fc->fence_queue);
-	}
+	if (wake) 
+		wake_up_all(&fc->fence_queue);
 }
 EXPORT_SYMBOL(drm_fence_handler);
 
@@ -219,41 +253,28 @@ static void drm_fence_object_destroy(struct drm_file *priv,
 	drm_fence_usage_deref_locked(&fence);
 }
 
-int drm_fence_object_signaled(struct drm_fence_object *fence,
-			      uint32_t mask, int poke_flush)
+int drm_fence_object_signaled(struct drm_fence_object *fence, uint32_t mask)
 {
 	unsigned long flags;
 	int signaled;
 	struct drm_device *dev = fence->dev;
 	struct drm_fence_manager *fm = &dev->fm;
 	struct drm_fence_driver *driver = dev->driver->fence_driver;
-
-	if (poke_flush)
-		driver->poke_flush(dev, fence->fence_class);
+	
+	mask &= fence->type;
 	read_lock_irqsave(&fm->lock, flags);
-	signaled =
-	    (fence->type & mask & fence->signaled) == (fence->type & mask);
+	signaled = (mask & fence->signaled_types) == mask;
 	read_unlock_irqrestore(&fm->lock, flags);
-
+	if (!signaled && driver->poll) {
+		write_lock_irqsave(&fm->lock, flags);
+		driver->poll(dev, fence->fence_class, mask);
+		signaled = (mask & fence->signaled_types) == mask;
+		write_unlock_irqrestore(&fm->lock, flags);
+	}
 	return signaled;
 }
 EXPORT_SYMBOL(drm_fence_object_signaled);
 
-static void drm_fence_flush_exe(struct drm_fence_class_manager *fc,
-				struct drm_fence_driver *driver,
-				uint32_t sequence)
-{
-	uint32_t diff;
-
-	if (!fc->pending_exe_flush) {
-		fc->exe_flush_sequence = sequence;
-		fc->pending_exe_flush = 1;
-	} else {
-		diff = (sequence - fc->exe_flush_sequence) & driver->sequence_mask;
-		if (diff < driver->wrap_diff)
-			fc->exe_flush_sequence = sequence;
-	}
-}
 
 int drm_fence_object_flush(struct drm_fence_object *fence,
 			   uint32_t type)
@@ -262,7 +283,10 @@ int drm_fence_object_flush(struct drm_fence_object *fence,
 	struct drm_fence_manager *fm = &dev->fm;
 	struct drm_fence_class_manager *fc = &fm->fence_class[fence->fence_class];
 	struct drm_fence_driver *driver = dev->driver->fence_driver;
-	unsigned long flags;
+	unsigned long irq_flags;
+	uint32_t saved_pending_flush;
+	uint32_t diff;
+	int call_flush;
 
 	if (type & ~fence->type) {
 		DRM_ERROR("Flush trying to extend fence type, "
@@ -270,24 +294,36 @@ int drm_fence_object_flush(struct drm_fence_object *fence,
 		return -EINVAL;
 	}
 
-	write_lock_irqsave(&fm->lock, flags);
-	fence->flush_mask |= type;
-	if ((fence->submitted_flush & fence->signaled)
-	    == fence->submitted_flush) {
-		if ((fence->type & DRM_FENCE_TYPE_EXE) &&
-		    !(fence->submitted_flush & DRM_FENCE_TYPE_EXE)) {
-			drm_fence_flush_exe(fc, driver, fence->sequence);
-			fence->submitted_flush |= DRM_FENCE_TYPE_EXE;
-		} else {
-			fc->pending_flush |= (fence->flush_mask &
-					      ~fence->submitted_flush);
-			fence->submitted_flush = fence->flush_mask;
-		}
-	}
-	write_unlock_irqrestore(&fm->lock, flags);
-	driver->poke_flush(dev, fence->fence_class);
+	write_lock_irqsave(&fm->lock, irq_flags);
+	fence->waiting_types |= type;
+	fc->waiting_types |= fence->waiting_types;
+	diff = (fence->sequence - fc->highest_waiting_sequence) & 
+		driver->sequence_mask;
+
+	if (diff < driver->wrap_diff)
+		fc->highest_waiting_sequence = fence->sequence;
+
+	/*
+	 * fence->waiting_types has changed. Determine whether
+	 * we need to initiate some kind of flush as a result of this.
+	 */
+
+	saved_pending_flush = fc->pending_flush;
+	if (driver->needed_flush) 
+		fc->pending_flush |= driver->needed_flush(fence);
+
+	if (driver->poll)
+		driver->poll(dev, fence->fence_class, fence->waiting_types);
+
+	call_flush = fc->pending_flush;
+	write_unlock_irqrestore(&fm->lock, irq_flags);
+
+	if (call_flush && driver->flush)
+		driver->flush(dev, fence->fence_class);
+
 	return 0;
 }
+EXPORT_SYMBOL(drm_fence_object_flush);
 
 /*
  * Make sure old fence objects are signaled before their fence sequences are
@@ -299,90 +335,52 @@ void drm_fence_flush_old(struct drm_device *dev, uint32_t fence_class,
 {
 	struct drm_fence_manager *fm = &dev->fm;
 	struct drm_fence_class_manager *fc = &fm->fence_class[fence_class];
-	struct drm_fence_driver *driver = dev->driver->fence_driver;
-	uint32_t old_sequence;
-	unsigned long flags;
 	struct drm_fence_object *fence;
+	unsigned long irq_flags;
+	struct drm_fence_driver *driver = dev->driver->fence_driver;
+	int call_flush;
+
 	uint32_t diff;
 
-	write_lock_irqsave(&fm->lock, flags);
-	old_sequence = (sequence - driver->flush_diff) & driver->sequence_mask;
-	diff = (old_sequence - fc->last_exe_flush) & driver->sequence_mask;
+	write_lock_irqsave(&fm->lock, irq_flags);
 
-	if ((diff < driver->wrap_diff) && !fc->pending_exe_flush) {
-		fc->pending_exe_flush = 1;
-		fc->exe_flush_sequence = sequence - (driver->flush_diff / 2);
-	}
-	write_unlock_irqrestore(&fm->lock, flags);
+	list_for_each_entry_reverse(fence, &fc->ring, ring) {
+		diff = (sequence - fence->sequence) & driver->sequence_mask;
+		if (diff <= driver->flush_diff)
+			break;
+	
+		fence->waiting_types = fence->type;
+		fc->waiting_types |= fence->type;
 
-	mutex_lock(&dev->struct_mutex);
-	read_lock_irqsave(&fm->lock, flags);
+		if (driver->needed_flush)
+			fc->pending_flush |= driver->needed_flush(fence);
+	}	
+	
+	if (driver->poll)
+		driver->poll(dev, fence_class, fc->waiting_types);
 
-	if (list_empty(&fc->ring)) {
-		read_unlock_irqrestore(&fm->lock, flags);
-		mutex_unlock(&dev->struct_mutex);
-		return;
-	}
-	fence = drm_fence_reference_locked(list_entry(fc->ring.next, struct drm_fence_object, ring));
-	mutex_unlock(&dev->struct_mutex);
-	diff = (old_sequence - fence->sequence) & driver->sequence_mask;
-	read_unlock_irqrestore(&fm->lock, flags);
-	if (diff < driver->wrap_diff)
-		drm_fence_object_flush(fence, fence->type);
-	drm_fence_usage_deref_unlocked(&fence);
-}
-EXPORT_SYMBOL(drm_fence_flush_old);
+	call_flush = fc->pending_flush;
+	write_unlock_irqrestore(&fm->lock, irq_flags);
 
-static int drm_fence_lazy_wait(struct drm_fence_object *fence,
-			       int ignore_signals,
-			       uint32_t mask)
-{
-	struct drm_device *dev = fence->dev;
-	struct drm_fence_manager *fm = &dev->fm;
-	struct drm_fence_class_manager *fc = &fm->fence_class[fence->fence_class];
-	int signaled;
-	unsigned long _end = jiffies + 3*DRM_HZ;
-	int ret = 0;
+	if (call_flush && driver->flush)
+		driver->flush(dev, fence->fence_class);
+
+	/*
+	 * FIXME: Shold we implement a wait here for really old fences?
+	 */
 
-	do {
-		DRM_WAIT_ON(ret, fc->fence_queue, 3 * DRM_HZ,
-			    (signaled = drm_fence_object_signaled(fence, mask, 1)));
-		if (signaled)
-			return 0;
-		if (time_after_eq(jiffies, _end))
-			break;
-	} while (ret == -EINTR && ignore_signals);
-	if (drm_fence_object_signaled(fence, mask, 0))
-		return 0;
-	if (time_after_eq(jiffies, _end))
-		ret = -EBUSY;
-	if (ret) {
-		if (ret == -EBUSY) {
-			DRM_ERROR("Fence timeout. "
-				  "GPU lockup or fence driver was "
-				  "taken down. %d 0x%08x 0x%02x 0x%02x 0x%02x\n",
-				  fence->fence_class,
-				  fence->sequence,
-				  fence->type,
-				  mask,
-				  fence->signaled);
-			DRM_ERROR("Pending exe flush %d 0x%08x\n",
-				  fc->pending_exe_flush,
-				  fc->exe_flush_sequence);
-		}
-		return ((ret == -EINTR) ? -EAGAIN : ret);
-	}
-	return 0;
 }
+EXPORT_SYMBOL(drm_fence_flush_old);
 
 int drm_fence_object_wait(struct drm_fence_object *fence,
 			  int lazy, int ignore_signals, uint32_t mask)
 {
 	struct drm_device *dev = fence->dev;
 	struct drm_fence_driver *driver = dev->driver->fence_driver;
+	struct drm_fence_manager *fm = &dev->fm;
+	struct drm_fence_class_manager *fc = &fm->fence_class[fence->fence_class];
 	int ret = 0;
-	unsigned long _end;
-	int signaled;
+	unsigned long _end = 3 * DRM_HZ;
 
 	if (mask & ~fence->type) {
 		DRM_ERROR("Wait trying to extend fence type"
@@ -391,58 +389,39 @@ int drm_fence_object_wait(struct drm_fence_object *fence,
 		return -EINVAL;
 	}
 
-	if (drm_fence_object_signaled(fence, mask, 0))
-		return 0;
+	if (driver->wait)
+		return driver->wait(fence, lazy, !ignore_signals, mask);
 
-	_end = jiffies + 3 * DRM_HZ;
 
 	drm_fence_object_flush(fence, mask);
+	if (driver->has_irq(dev, fence->fence_class, mask)) {
+		if (!ignore_signals)
+			ret = wait_event_interruptible_timeout
+				(fc->fence_queue, 
+				 drm_fence_object_signaled(fence, mask), 
+				 3 * DRM_HZ);
+		else 
+			ret = wait_event_timeout
+				(fc->fence_queue, 
+				 drm_fence_object_signaled(fence, mask), 
+				 3 * DRM_HZ);
+
+		if (unlikely(ret == -ERESTARTSYS))
+			return -EAGAIN;
+
+		if (unlikely(ret == 0))
+			return -EBUSY;
 
-	if (lazy && driver->lazy_capable) {
-
-		ret = drm_fence_lazy_wait(fence, ignore_signals, mask);
-		if (ret)
-			return ret;
-
-	} else {
-
-		if (driver->has_irq(dev, fence->fence_class,
-				    DRM_FENCE_TYPE_EXE)) {
-			ret = drm_fence_lazy_wait(fence, ignore_signals,
-						  DRM_FENCE_TYPE_EXE);
-			if (ret)
-				return ret;
-		}
-
-		if (driver->has_irq(dev, fence->fence_class,
-				    mask & ~DRM_FENCE_TYPE_EXE)) {
-			ret = drm_fence_lazy_wait(fence, ignore_signals,
-						  mask);
-			if (ret)
-				return ret;
-		}
-	}
-	if (drm_fence_object_signaled(fence, mask, 0))
 		return 0;
+	}
 
-	/*
-	 * Avoid kernel-space busy-waits.
-	 */
-	if (!ignore_signals)
-		return -EAGAIN;
-
-	do {
-		schedule();
-		signaled = drm_fence_object_signaled(fence, mask, 1);
-	} while (!signaled && !time_after_eq(jiffies, _end));
-
-	if (!signaled)
-		return -EBUSY;
-
-	return 0;
+	return drm_fence_wait_polling(fence, lazy, !ignore_signals, mask,
+				      _end);
 }
 EXPORT_SYMBOL(drm_fence_object_wait);
 
+
+
 int drm_fence_object_emit(struct drm_fence_object *fence, uint32_t fence_flags,
 			  uint32_t fence_class, uint32_t type)
 {
@@ -452,25 +431,24 @@ int drm_fence_object_emit(struct drm_fence_object *fence, uint32_t fence_flags,
 	struct drm_fence_class_manager *fc = &fm->fence_class[fence->fence_class];
 	unsigned long flags;
 	uint32_t sequence;
-	uint32_t native_type;
+	uint32_t native_types;
 	int ret;
 
 	drm_fence_unring(dev, &fence->ring);
 	ret = driver->emit(dev, fence_class, fence_flags, &sequence,
-			   &native_type);
+			   &native_types);
 	if (ret)
 		return ret;
 
 	write_lock_irqsave(&fm->lock, flags);
 	fence->fence_class = fence_class;
 	fence->type = type;
-	fence->flush_mask = 0x00;
-	fence->submitted_flush = 0x00;
-	fence->signaled = 0x00;
+	fence->waiting_types = 0;
+	fence->signaled_types = 0;
 	fence->sequence = sequence;
-	fence->native_type = native_type;
+	fence->native_types = native_types;
 	if (list_empty(&fc->ring))
-		fc->last_exe_flush = sequence - 1;
+		fc->highest_waiting_sequence = sequence - 1;
 	list_add_tail(&fence->ring, &fc->ring);
 	write_unlock_irqrestore(&fm->lock, flags);
 	return 0;
@@ -500,9 +478,8 @@ static int drm_fence_object_init(struct drm_device *dev, uint32_t fence_class,
 	INIT_LIST_HEAD(&fence->base.list);
 	fence->fence_class = fence_class;
 	fence->type = type;
-	fence->flush_mask = 0;
-	fence->submitted_flush = 0;
-	fence->signaled = 0;
+	fence->signaled_types = 0;
+	fence->waiting_types = 0;
 	fence->sequence = 0;
 	fence->dev = dev;
 	write_unlock_irqrestore(&fm->lock, flags);
@@ -598,7 +575,7 @@ void drm_fence_fill_arg(struct drm_fence_object *fence,
 	arg->handle = fence->base.hash.key;
 	arg->fence_class = fence->fence_class;
 	arg->type = fence->type;
-	arg->signaled = fence->signaled;
+	arg->signaled = fence->signaled_types;
 	arg->error = fence->error;
 	arg->sequence = fence->sequence;
 	read_unlock_irqrestore(&fm->lock, irq_flags);
diff --git a/linux-core/drm_objects.h b/linux-core/drm_objects.h
index a2d10b5d..c35d900f 100644
--- a/linux-core/drm_objects.h
+++ b/linux-core/drm_objects.h
@@ -147,12 +147,11 @@ struct drm_fence_object {
 
 	struct list_head ring;
 	int fence_class;
-	uint32_t native_type;
+	uint32_t native_types;
 	uint32_t type;
-	uint32_t signaled;
+	uint32_t signaled_types;
 	uint32_t sequence;
-	uint32_t flush_mask;
-	uint32_t submitted_flush;
+	uint32_t waiting_types;
 	uint32_t error;
 };
 
@@ -162,10 +161,9 @@ struct drm_fence_object {
 struct drm_fence_class_manager {
 	struct list_head ring;
 	uint32_t pending_flush;
+	uint32_t waiting_types;
 	wait_queue_head_t fence_queue;
-	int pending_exe_flush;
-	uint32_t last_exe_flush;
-	uint32_t exe_flush_sequence;
+	uint32_t highest_waiting_sequence;
 };
 
 struct drm_fence_manager {
@@ -177,19 +175,49 @@ struct drm_fence_manager {
 };
 
 struct drm_fence_driver {
+	unsigned long *waiting_jiffies;
 	uint32_t num_classes;
 	uint32_t wrap_diff;
 	uint32_t flush_diff;
 	uint32_t sequence_mask;
-	int lazy_capable;
+
+	/*
+	 * Driver implemented functions:
+	 * has_irq() : 1 if the hardware can update the indicated type_flags using an
+	 * irq handler. 0 if polling is required.
+	 *
+	 * emit() : Emit a sequence number to the command stream.
+	 * Return the sequence number.
+	 *
+	 * flush() : Make sure the flags indicated in fc->pending_flush will eventually
+	 * signal for fc->highest_received_sequence and all preceding sequences.
+	 * Acknowledge by clearing the flags fc->pending_flush.
+	 *
+	 * poll() : Call drm_fence_handler with any new information.
+	 *
+	 * needed_flush() : Given the current state of the fence->type flags and previusly 
+	 * executed or queued flushes, return the type_flags that need flushing.
+	 *
+	 * wait(): Wait for the "mask" flags to signal on a given fence, performing
+	 * whatever's necessary to make this happen.
+	 */
+
 	int (*has_irq) (struct drm_device *dev, uint32_t fence_class,
 			uint32_t flags);
 	int (*emit) (struct drm_device *dev, uint32_t fence_class,
 		     uint32_t flags, uint32_t *breadcrumb,
 		     uint32_t *native_type);
-	void (*poke_flush) (struct drm_device *dev, uint32_t fence_class);
+	void (*flush) (struct drm_device *dev, uint32_t fence_class);
+	void (*poll) (struct drm_device *dev, uint32_t fence_class,
+		uint32_t types);
+	uint32_t (*needed_flush) (struct drm_fence_object *fence);
+	int (*wait) (struct drm_fence_object *fence, int lazy,
+		     int interruptible, uint32_t mask);
 };
 
+extern int drm_fence_wait_polling(struct drm_fence_object *fence, int lazy,
+				  int interruptible, uint32_t mask,
+				  unsigned long end_jiffies);
 extern void drm_fence_handler(struct drm_device *dev, uint32_t fence_class,
 			      uint32_t sequence, uint32_t type,
 			      uint32_t error);
@@ -200,7 +228,7 @@ extern void drm_fence_flush_old(struct drm_device *dev, uint32_t fence_class,
 extern int drm_fence_object_flush(struct drm_fence_object *fence,
 				  uint32_t type);
 extern int drm_fence_object_signaled(struct drm_fence_object *fence,
-				     uint32_t type, int flush);
+				     uint32_t type);
 extern void drm_fence_usage_deref_locked(struct drm_fence_object **fence);
 extern void drm_fence_usage_deref_unlocked(struct drm_fence_object **fence);
 extern struct drm_fence_object *drm_fence_reference_locked(struct drm_fence_object *src);
@@ -576,6 +604,33 @@ struct drm_bo_driver {
 	 * ttm_cache_flush
 	 */
 	void (*ttm_cache_flush)(struct drm_ttm *ttm);
+
+	/*
+	 * command_stream_barrier
+	 *
+	 * @dev: The drm device.
+	 *
+	 * @bo: The buffer object to validate.
+	 *
+	 * @new_fence_class: The new fence class for the buffer object.
+	 *
+	 * @new_fence_type: The new fence type for the buffer object.
+	 *
+	 * @no_wait: whether this should give up and return -EBUSY
+	 * if this operation would require sleeping
+	 *
+	 * Insert a command stream barrier that makes sure that the
+	 * buffer is idle once the commands associated with the
+	 * current validation are starting to execute. If an error
+	 * condition is returned, or the function pointer is NULL,
+	 * the drm core will force buffer idle
+	 * during validation.
+	 */
+
+	int (*command_stream_barrier) (struct drm_buffer_object *bo,
+				       uint32_t new_fence_class,
+				       uint32_t new_fence_type,
+				       int no_wait);				       
 };
 
 /*
diff --git a/linux-core/i915_drv.c b/linux-core/i915_drv.c
index 5f2e6adc..ccc061d3 100644
--- a/linux-core/i915_drv.c
+++ b/linux-core/i915_drv.c
@@ -39,17 +39,9 @@ static struct pci_device_id pciidlist[] = {
 };
 
 #ifdef I915_HAVE_FENCE
-static struct drm_fence_driver i915_fence_driver = {
-	.num_classes = 1,
-	.wrap_diff = (1U << (BREADCRUMB_BITS - 1)),
-	.flush_diff = (1U << (BREADCRUMB_BITS - 2)),
-	.sequence_mask = BREADCRUMB_MASK,
-	.lazy_capable = 1,
-	.emit = i915_fence_emit_sequence,
-	.poke_flush = i915_poke_flush,
-	.has_irq = i915_fence_has_irq,
-};
+extern struct drm_fence_driver i915_fence_driver;
 #endif
+
 #ifdef I915_HAVE_BUFFER
 
 static uint32_t i915_mem_prios[] = {DRM_BO_MEM_PRIV0, DRM_BO_MEM_TT, DRM_BO_MEM_LOCAL};
@@ -67,6 +59,7 @@ static struct drm_bo_driver i915_bo_driver = {
 	.evict_flags = i915_evict_flags,
 	.move = i915_move,
 	.ttm_cache_flush = i915_flush_ttm,
+	.command_stream_barrier = NULL,
 };
 #endif
 
diff --git a/linux-core/i915_fence.c b/linux-core/i915_fence.c
index e3c76df6..8a2e7f13 100644
--- a/linux-core/i915_fence.c
+++ b/linux-core/i915_fence.c
@@ -35,97 +35,121 @@
 #include "i915_drv.h"
 
 /*
- * Implements an intel sync flush operation.
+ * Initiate a sync flush if it's not already pending.
  */
 
-static void i915_perform_flush(struct drm_device *dev)
+static void i915_initiate_rwflush(struct drm_i915_private *dev_priv, 
+				  struct drm_fence_class_manager *fc)
+{
+	if ((fc->pending_flush & DRM_I915_FENCE_TYPE_RW) && 
+	    !dev_priv->flush_pending) {
+		dev_priv->flush_sequence = (uint32_t) READ_BREADCRUMB(dev_priv);
+		dev_priv->flush_flags = fc->pending_flush;
+		dev_priv->saved_flush_status = READ_HWSP(dev_priv, 0);
+		I915_WRITE(I915REG_INSTPM, (1 << 5) | (1 << 21));
+		dev_priv->flush_pending = 1;
+		fc->pending_flush &= ~DRM_I915_FENCE_TYPE_RW;
+	}
+}
+
+static void i915_fence_flush(struct drm_device *dev,
+			     uint32_t fence_class)
+{
+	struct drm_i915_private *dev_priv = 
+		(struct drm_i915_private *) dev->dev_private;
+	struct drm_fence_manager *fm = &dev->fm;
+	struct drm_fence_class_manager *fc = &fm->fence_class[0];
+	unsigned long irq_flags;
+
+	if (unlikely(!dev_priv))
+		return;
+
+	write_lock_irqsave(&fm->lock, irq_flags);
+	i915_initiate_rwflush(dev_priv, fc);
+	write_unlock_irqrestore(&fm->lock, irq_flags);
+}
+
+static void i915_fence_poll(struct drm_device *dev, uint32_t fence_class,
+			    uint32_t waiting_types)
 {
 	drm_i915_private_t *dev_priv = (drm_i915_private_t *) dev->dev_private;
 	struct drm_fence_manager *fm = &dev->fm;
 	struct drm_fence_class_manager *fc = &fm->fence_class[0];
-	struct drm_fence_driver *driver = dev->driver->fence_driver;
 	uint32_t flush_flags = 0;
 	uint32_t flush_sequence = 0;
 	uint32_t i_status;
-	uint32_t diff;
 	uint32_t sequence;
-	int rwflush;
 
-	if (!dev_priv)
+	if (unlikely(!dev_priv))
 		return;
 
-	if (fc->pending_exe_flush) {
-		sequence = READ_BREADCRUMB(dev_priv);
+	/*
+	 * First, report any executed sync flush:
+	 */
+
+	if (dev_priv->flush_pending) {
+		i_status = READ_HWSP(dev_priv, 0);
+		if ((i_status & (1 << 12)) !=
+		    (dev_priv->saved_flush_status & (1 << 12))) {
+			flush_flags = dev_priv->flush_flags;
+			flush_sequence = dev_priv->flush_sequence;
+			dev_priv->flush_pending = 0;
+			drm_fence_handler(dev, 0, flush_sequence, flush_flags, 0);
+		}
+	}		
 
-		/*
-		 * First update fences with the current breadcrumb.
-		 */
+	/*
+	 * Report A new breadcrumb, and adjust IRQs.
+	 */
+
+	if (waiting_types & DRM_FENCE_TYPE_EXE) {
+		sequence = READ_BREADCRUMB(dev_priv);
 
-		diff = (sequence - fc->last_exe_flush) & BREADCRUMB_MASK;
-		if (diff < driver->wrap_diff && diff != 0) {
-			drm_fence_handler(dev, 0, sequence,
+		if (sequence != dev_priv->reported_sequence ||
+		    !dev_priv->reported_sequence_valid) {
+			drm_fence_handler(dev, 0, sequence, 
 					  DRM_FENCE_TYPE_EXE, 0);
+			dev_priv->reported_sequence = sequence;
+			dev_priv->reported_sequence_valid = 1;
 		}
 
-		if (dev_priv->fence_irq_on && !fc->pending_exe_flush) {
+		if (dev_priv->fence_irq_on && !(waiting_types & DRM_FENCE_TYPE_EXE)) {
 			i915_user_irq_off(dev_priv);
 			dev_priv->fence_irq_on = 0;
-		} else if (!dev_priv->fence_irq_on && fc->pending_exe_flush) {
+		} else if (!dev_priv->fence_irq_on && (waiting_types & DRM_FENCE_TYPE_EXE)) {
 			i915_user_irq_on(dev_priv);
 			dev_priv->fence_irq_on = 1;
 		}
 	}
 
-	if (dev_priv->flush_pending) {
-		i_status = READ_HWSP(dev_priv, 0);
-		if ((i_status & (1 << 12)) !=
-		    (dev_priv->saved_flush_status & (1 << 12))) {
-			flush_flags = dev_priv->flush_flags;
-			flush_sequence = dev_priv->flush_sequence;
-			dev_priv->flush_pending = 0;
-			drm_fence_handler(dev, 0, flush_sequence, flush_flags, 0);
-		}
-	}
+	/*
+	 * There may be new RW flushes pending. Start them.
+	 */
+	
+	i915_initiate_rwflush(dev_priv, fc); 
 
-	rwflush = fc->pending_flush & DRM_I915_FENCE_TYPE_RW;
-	if (rwflush && !dev_priv->flush_pending) {
-		dev_priv->flush_sequence = (uint32_t) READ_BREADCRUMB(dev_priv);
-		dev_priv->flush_flags = fc->pending_flush;
-		dev_priv->saved_flush_status = READ_HWSP(dev_priv, 0);
-		I915_WRITE(I915REG_INSTPM, (1 << 5) | (1 << 21));
-		dev_priv->flush_pending = 1;
-		fc->pending_flush &= ~DRM_I915_FENCE_TYPE_RW;
-	}
+	/*
+	 * And possibly, but unlikely, they finish immediately. 
+	 */
 
 	if (dev_priv->flush_pending) {
 		i_status = READ_HWSP(dev_priv, 0);
-		if ((i_status & (1 << 12)) !=
-		    (dev_priv->saved_flush_status & (1 << 12))) {
+		if (unlikely((i_status & (1 << 12)) !=
+		    (dev_priv->saved_flush_status & (1 << 12)))) {
 			flush_flags = dev_priv->flush_flags;
 			flush_sequence = dev_priv->flush_sequence;
 			dev_priv->flush_pending = 0;
 			drm_fence_handler(dev, 0, flush_sequence, flush_flags, 0);
 		}
 	}
-
 }
 
-void i915_poke_flush(struct drm_device *dev, uint32_t class)
-{
-	struct drm_fence_manager *fm = &dev->fm;
-	unsigned long flags;
-
-	write_lock_irqsave(&fm->lock, flags);
-	i915_perform_flush(dev);
-	write_unlock_irqrestore(&fm->lock, flags);
-}
-
-int i915_fence_emit_sequence(struct drm_device *dev, uint32_t class,
+static int i915_fence_emit_sequence(struct drm_device *dev, uint32_t class,
 			     uint32_t flags, uint32_t *sequence,
 			     uint32_t *native_type)
 {
 	drm_i915_private_t *dev_priv = (drm_i915_private_t *) dev->dev_private;
-	if (!dev_priv)
+	if (unlikely(!dev_priv))
 		return -EINVAL;
 
 	i915_emit_irq(dev);
@@ -140,20 +164,130 @@ int i915_fence_emit_sequence(struct drm_device *dev, uint32_t class,
 void i915_fence_handler(struct drm_device *dev)
 {
 	struct drm_fence_manager *fm = &dev->fm;
+	struct drm_fence_class_manager *fc = &fm->fence_class[0];
 
 	write_lock(&fm->lock);
-	i915_perform_flush(dev);
+	i915_fence_poll(dev, 0, fc->waiting_types);
 	write_unlock(&fm->lock);
 }
 
-int i915_fence_has_irq(struct drm_device *dev, uint32_t class, uint32_t flags)
+/*
+ * We need a separate wait function since we need to poll for
+ * sync flushes.
+ */
+
+static int i915_fence_wait(struct drm_fence_object *fence,
+			   int lazy, int interruptible, uint32_t mask)
 {
+	struct drm_device *dev = fence->dev;
+	drm_i915_private_t *dev_priv = (struct drm_i915_private *) dev->dev_private;
+	struct drm_fence_manager *fm = &dev->fm;
+	struct drm_fence_class_manager *fc = &fm->fence_class[0];
+	int ret;
+	unsigned long  _end = jiffies + 3 * DRM_HZ;
+
+	drm_fence_object_flush(fence, mask);
+	if (likely(interruptible))
+		ret = wait_event_interruptible_timeout
+			(fc->fence_queue, drm_fence_object_signaled(fence, DRM_FENCE_TYPE_EXE), 
+			 3 * DRM_HZ);
+	else 
+		ret = wait_event_timeout
+			(fc->fence_queue, drm_fence_object_signaled(fence, DRM_FENCE_TYPE_EXE), 
+			 3 * DRM_HZ);
+
+	if (unlikely(ret == -ERESTARTSYS))
+		return -EAGAIN;
+
+	if (unlikely(ret == 0))
+		return -EBUSY;
+
+	if (likely(mask == DRM_FENCE_TYPE_EXE || 
+		   drm_fence_object_signaled(fence, mask))) 
+		return 0;
+
 	/*
-	 * We have an irq that tells us when we have a new breadcrumb.
+	 * Remove this code snippet when fixed. HWSTAM doesn't let
+	 * flush info through...
 	 */
 
-	if (class == 0 && flags == DRM_FENCE_TYPE_EXE)
-		return 1;
+	if (unlikely(dev_priv && !dev_priv->irq_enabled)) {
+		unsigned long irq_flags;
 
-	return 0;
+		DRM_ERROR("X server disabled IRQs before releasing frame buffer.\n");
+		msleep(100);
+		dev_priv->flush_pending = 0;
+		write_lock_irqsave(&fm->lock, irq_flags);
+		drm_fence_handler(dev, fence->fence_class, 
+				  fence->sequence, fence->type, 0);
+		write_unlock_irqrestore(&fm->lock, irq_flags);
+	}
+
+	/*
+	 * Poll for sync flush completion.
+	 */
+
+	return drm_fence_wait_polling(fence, lazy, interruptible, mask, _end);
+}
+
+static uint32_t i915_fence_needed_flush(struct drm_fence_object *fence)
+{
+	uint32_t flush_flags = fence->waiting_types & 
+		~(DRM_FENCE_TYPE_EXE | fence->signaled_types);
+
+	if (likely(flush_flags == 0 || 
+		   ((flush_flags & ~fence->native_types) == 0) || 
+		   (fence->signaled_types != DRM_FENCE_TYPE_EXE)))
+		return 0;
+	else {
+		struct drm_device *dev = fence->dev;
+		struct drm_i915_private *dev_priv = (struct drm_i915_private *) dev->dev_private;
+		struct drm_fence_driver *driver = dev->driver->fence_driver;
+		
+		if (unlikely(!dev_priv))
+			return 0;
+
+		if (dev_priv->flush_pending) {
+			uint32_t diff = (dev_priv->flush_sequence - fence->sequence) & 
+				driver->sequence_mask;
+
+			if (diff < driver->wrap_diff)
+				return 0;
+		}
+	}
+	return flush_flags;
+}
+
+/*
+ * In the very unlikely event that "poll" is not really called very often
+ * we need the following function to handle sequence wraparounds.
+ */
+
+void i915_invalidate_reported_sequence(struct drm_device *dev)
+{
+	struct drm_i915_private *dev_priv = (struct drm_i915_private *) 
+		dev->dev_private;
+	struct drm_fence_manager *fm = &dev->fm;
+	unsigned long irq_flags;
+
+	if (unlikely(!dev_priv))
+		return;
+	
+	write_lock_irqsave(&fm->lock, irq_flags);
+	dev_priv->reported_sequence_valid = 0;
+	write_unlock_irqrestore(&fm->lock, irq_flags);
 }
+	
+
+struct drm_fence_driver i915_fence_driver = {
+	.num_classes = 1,
+	.wrap_diff = (1U << (BREADCRUMB_BITS - 1)),
+	.flush_diff = (1U << (BREADCRUMB_BITS - 2)),
+	.sequence_mask = BREADCRUMB_MASK,
+	.has_irq = NULL,
+	.emit = i915_fence_emit_sequence,
+	.flush = i915_fence_flush,
+	.poll = i915_fence_poll,
+	.needed_flush = i915_fence_needed_flush,
+	.wait = i915_fence_wait,
+};
diff --git a/linux-core/nouveau_buffer.c b/linux-core/nouveau_buffer.c
index a652bb1d..11549317 100644
--- a/linux-core/nouveau_buffer.c
+++ b/linux-core/nouveau_buffer.c
@@ -294,5 +294,6 @@ struct drm_bo_driver nouveau_bo_driver = {
 	.init_mem_type = nouveau_bo_init_mem_type,
 	.evict_flags = nouveau_bo_evict_flags,
 	.move = nouveau_bo_move,
-	.ttm_cache_flush= nouveau_bo_flush_ttm
+	.ttm_cache_flush= nouveau_bo_flush_ttm,
+	.command_stream_barrier = NULL
 };
diff --git a/linux-core/nouveau_fence.c b/linux-core/nouveau_fence.c
index 4e624a7a..59dcf7d0 100644
--- a/linux-core/nouveau_fence.c
+++ b/linux-core/nouveau_fence.c
@@ -75,7 +75,7 @@ nouveau_fence_emit(struct drm_device *dev, uint32_t class, uint32_t flags,
 }
 
 static void
-nouveau_fence_perform_flush(struct drm_device *dev, uint32_t class)
+nouveau_fence_poll(struct drm_device *dev, uint32_t class, uint32_t waiting_types)
 {
 	struct drm_nouveau_private *dev_priv = dev->dev_private;
 	struct drm_fence_class_manager *fc = &dev->fm.fence_class[class];
@@ -83,42 +83,26 @@ nouveau_fence_perform_flush(struct drm_device *dev, uint32_t class)
 	uint32_t pending_types = 0;
 
 	DRM_DEBUG("class=%d\n", class);
-
-	pending_types = fc->pending_flush |
-			((fc->pending_exe_flush) ? DRM_FENCE_TYPE_EXE : 0);
-	DRM_DEBUG("pending: 0x%08x 0x%08x\n", pending_types,
-					      fc->pending_flush);
+	DRM_DEBUG("pending: 0x%08x 0x%08x\n", waiting_types, fc->waiting_types);
 
 	if (pending_types) {
 		uint32_t sequence = NV_READ(chan->ref_cnt);
 
 		DRM_DEBUG("got 0x%08x\n", sequence);
-		drm_fence_handler(dev, class, sequence, pending_types, 0);
+		drm_fence_handler(dev, class, sequence, waiting_types, 0);
 	}
 }
 
-static void
-nouveau_fence_poke_flush(struct drm_device *dev, uint32_t class)
-{
-	struct drm_fence_manager *fm = &dev->fm;
-	unsigned long flags;
-
-	DRM_DEBUG("class=%d\n", class);
-
-	write_lock_irqsave(&fm->lock, flags);
-	nouveau_fence_perform_flush(dev, class);
-	write_unlock_irqrestore(&fm->lock, flags);
-}
-
 void
 nouveau_fence_handler(struct drm_device *dev, int channel)
 {
 	struct drm_fence_manager *fm = &dev->fm;
+	struct drm_fence_class_manager *fc = &fm->fence_class[channel];
 
 	DRM_DEBUG("class=%d\n", channel);
 
 	write_lock(&fm->lock);
-	nouveau_fence_perform_flush(dev, channel);
+	nouveau_fence_poll(dev, channel, fc->waiting_types);
 	write_unlock(&fm->lock);
 }
 
@@ -127,8 +111,10 @@ struct drm_fence_driver nouveau_fence_driver = {
 	.wrap_diff	= (1 << 30),
 	.flush_diff	= (1 << 29),
 	.sequence_mask	= 0xffffffffU,
-	.lazy_capable	= 1,
 	.has_irq	= nouveau_fence_has_irq,
 	.emit		= nouveau_fence_emit,
-	.poke_flush	= nouveau_fence_poke_flush
+	.flush          = NULL,
+	.poll           = nouveau_fence_poll,
+	.needed_flush   = NULL,
+	.wait           = NULL
 };
diff --git a/linux-core/via_fence.c b/linux-core/via_fence.c
index 9af1bf3b..3a680a32 100644
--- a/linux-core/via_fence.c
+++ b/linux-core/via_fence.c
@@ -38,27 +38,21 @@
  * DRM_VIA_FENCE_TYPE_ACCEL guarantees that all 2D & 3D rendering is complete.
  */
 
-
-static uint32_t via_perform_flush(struct drm_device *dev, uint32_t class)
+static void via_fence_poll(struct drm_device *dev, uint32_t class,
+			   uint32_t waiting_types)
 {
 	drm_via_private_t *dev_priv = (drm_via_private_t *) dev->dev_private;
-	struct drm_fence_class_manager *fc = &dev->fm.fence_class[class];
-	uint32_t pending_flush_types = 0;
 	uint32_t signaled_flush_types = 0;
 	uint32_t status;
 
 	if (class != 0)
-		return 0;
+		return;
 
-	if (!dev_priv)
-		return 0;
+	if (unlikely(!dev_priv))
+		return;
 
 	spin_lock(&dev_priv->fence_lock);
-
-	pending_flush_types = fc->pending_flush |
-		((fc->pending_exe_flush) ? DRM_FENCE_TYPE_EXE : 0);
-
-	if (pending_flush_types) {
+	if (waiting_types) {
 
 		/*
 		 * Take the idlelock. This guarantees that the next time a client tries
@@ -77,7 +71,7 @@ static uint32_t via_perform_flush(struct drm_device *dev, uint32_t class)
 		 * Check if AGP command reader is idle.
 		 */
 
-		if (pending_flush_types & DRM_FENCE_TYPE_EXE)
+		if (waiting_types & DRM_FENCE_TYPE_EXE)
 			if (VIA_READ(0x41C) & 0x80000000)
 				signaled_flush_types |= DRM_FENCE_TYPE_EXE;
 
@@ -85,7 +79,7 @@ static uint32_t via_perform_flush(struct drm_device *dev, uint32_t class)
 		 * Check VRAM command queue empty and 2D + 3D engines idle.
 		 */
 
-		if (pending_flush_types & DRM_VIA_FENCE_TYPE_ACCEL) {
+		if (waiting_types & DRM_VIA_FENCE_TYPE_ACCEL) {
 			status = VIA_READ(VIA_REG_STATUS);
 			if ((status & VIA_VR_QUEUE_BUSY) &&
 			    !(status & (VIA_CMD_RGTR_BUSY | VIA_2D_ENG_BUSY | VIA_3D_ENG_BUSY)))
@@ -93,8 +87,8 @@ static uint32_t via_perform_flush(struct drm_device *dev, uint32_t class)
 		}
 
 		if (signaled_flush_types) {
-			pending_flush_types &= ~signaled_flush_types;
-			if (!pending_flush_types && dev_priv->have_idlelock) {
+			waiting_types &= ~signaled_flush_types;
+			if (!waiting_types && dev_priv->have_idlelock) {
 				drm_idlelock_release(&dev->lock);
 				dev_priv->have_idlelock = 0;
 			}
@@ -105,8 +99,7 @@ static uint32_t via_perform_flush(struct drm_device *dev, uint32_t class)
 
 	spin_unlock(&dev_priv->fence_lock);
 
-	return fc->pending_flush |
-		((fc->pending_exe_flush) ? DRM_FENCE_TYPE_EXE : 0);
+	return;
 }
 
 
@@ -114,8 +107,8 @@ static uint32_t via_perform_flush(struct drm_device *dev, uint32_t class)
  * Emit a fence sequence.
  */
 
-int via_fence_emit_sequence(struct drm_device * dev, uint32_t class, uint32_t flags,
-			     uint32_t * sequence, uint32_t * native_type)
+static int via_fence_emit_sequence(struct drm_device * dev, uint32_t class, uint32_t flags,
+				   uint32_t * sequence, uint32_t * native_type)
 {
 	drm_via_private_t *dev_priv = (drm_via_private_t *) dev->dev_private;
 	int ret = 0;
@@ -150,36 +143,6 @@ int via_fence_emit_sequence(struct drm_device * dev, uint32_t class, uint32_t fl
 }
 
 /**
- * Manual poll (from the fence manager).
- */
-
-void via_poke_flush(struct drm_device * dev, uint32_t class)
-{
-	drm_via_private_t *dev_priv = (drm_via_private_t *) dev->dev_private;
-	struct drm_fence_manager *fm = &dev->fm;
-	unsigned long flags;
-	uint32_t pending_flush;
-
-	if (!dev_priv)
-		return ;
-
-	write_lock_irqsave(&fm->lock, flags);
-	pending_flush = via_perform_flush(dev, class);
-	if (pending_flush)
-		pending_flush = via_perform_flush(dev, class);
-	write_unlock_irqrestore(&fm->lock, flags);
-
-	/*
-	 * Kick the timer if there are more flushes pending.
-	 */
-
-	if (pending_flush && !timer_pending(&dev_priv->fence_timer)) {
-		dev_priv->fence_timer.expires = jiffies + 1;
-		add_timer(&dev_priv->fence_timer);
-	}
-}
-
-/**
  * No irq fence expirations implemented yet.
  * Although both the HQV engines and PCI dmablit engines signal
  * idle with an IRQ, we haven't implemented this yet.
@@ -187,45 +150,20 @@ void via_poke_flush(struct drm_device * dev, uint32_t class)
  * unless the caller wanting to wait for a fence object has indicated a lazy wait.
  */
 
-int via_fence_has_irq(struct drm_device * dev, uint32_t class,
-		      uint32_t flags)
+static int via_fence_has_irq(struct drm_device * dev, uint32_t class,
+			     uint32_t flags)
 {
 	return 0;
 }
 
-/**
- * Regularly call the flush function. This enables lazy waits, so we can
- * set lazy_capable. Lazy waits don't really care when the fence expires,
- * so a timer tick delay should be fine.
- */
-
-void via_fence_timer(unsigned long data)
-{
-	struct drm_device *dev = (struct drm_device *) data;
-	drm_via_private_t *dev_priv = (drm_via_private_t *) dev->dev_private;
-	struct drm_fence_manager *fm = &dev->fm;
-	uint32_t pending_flush;
-	struct drm_fence_class_manager *fc = &dev->fm.fence_class[0];
-
-	if (!dev_priv)
-		return;
-	if (!fm->initialized)
-		goto out_unlock;
-
-	via_poke_flush(dev, 0);
-	pending_flush = fc->pending_flush |
-		((fc->pending_exe_flush) ? DRM_FENCE_TYPE_EXE : 0);
-
-	/*
-	 * disable timer if there are no more flushes pending.
-	 */
-
-	if (!pending_flush && timer_pending(&dev_priv->fence_timer)) {
-		BUG_ON(dev_priv->have_idlelock);
-		del_timer(&dev_priv->fence_timer);
-	}
-	return;
-out_unlock:
-	return;
-
-}
+struct drm_fence_driver via_fence_driver = {
+	.num_classes = 1,
+	.wrap_diff = (1 << 30),
+	.flush_diff = (1 << 20),
+	.sequence_mask = 0xffffffffU,
+	.has_irq = via_fence_has_irq,
+	.emit = via_fence_emit_sequence,
+	.poll = via_fence_poll,
+	.needed_flush = NULL,
+	.wait = NULL
+};
diff --git a/linux-core/xgi_drv.c b/linux-core/xgi_drv.c
index 4f0b4ed0..f0225f89 100644
--- a/linux-core/xgi_drv.c
+++ b/linux-core/xgi_drv.c
@@ -37,16 +37,7 @@ static struct pci_device_id pciidlist[] = {
 	xgi_PCI_IDS
 };
 
-static struct drm_fence_driver xgi_fence_driver = {
-	.num_classes = 1,
-	.wrap_diff = BEGIN_BEGIN_IDENTIFICATION_MASK,
-	.flush_diff = BEGIN_BEGIN_IDENTIFICATION_MASK - 1,
-	.sequence_mask = BEGIN_BEGIN_IDENTIFICATION_MASK,
-	.lazy_capable = 1,
-	.emit = xgi_fence_emit_sequence,
-	.poke_flush = xgi_poke_flush,
-	.has_irq = xgi_fence_has_irq
-};
+extern struct drm_fence_driver xgi_fence_driver;
 
 int xgi_bootstrap(struct drm_device *, void *, struct drm_file *);
 
diff --git a/linux-core/xgi_fence.c b/linux-core/xgi_fence.c
index 9a75581a..63ed29ee 100644
--- a/linux-core/xgi_fence.c
+++ b/linux-core/xgi_fence.c
@@ -30,44 +30,37 @@
 #include "xgi_misc.h"
 #include "xgi_cmdlist.h"
 
-static uint32_t xgi_do_flush(struct drm_device * dev, uint32_t class)
+static void xgi_fence_poll(struct drm_device * dev, uint32_t class, 
+			   uint32_t waiting_types)
 {
 	struct xgi_info * info = dev->dev_private;
-	struct drm_fence_class_manager * fc = &dev->fm.fence_class[class];
-	uint32_t pending_flush_types = 0;
-	uint32_t signaled_flush_types = 0;
+	uint32_t signaled_types = 0;
 
 
 	if ((info == NULL) || (class != 0))
-		return 0;
+		return;
 
 	DRM_SPINLOCK(&info->fence_lock);
 
-	pending_flush_types = fc->pending_flush |
-		((fc->pending_exe_flush) ? DRM_FENCE_TYPE_EXE : 0);
-
-	if (pending_flush_types) {
-		if (pending_flush_types & DRM_FENCE_TYPE_EXE) {
+	if (waiting_types) {
+		if (waiting_types & DRM_FENCE_TYPE_EXE) {
 			const u32 begin_id = le32_to_cpu(DRM_READ32(info->mmio_map,
 							0x2820))
 				& BEGIN_BEGIN_IDENTIFICATION_MASK;
 
 			if (begin_id != info->complete_sequence) {
 				info->complete_sequence = begin_id;
-				signaled_flush_types |= DRM_FENCE_TYPE_EXE;
+				signaled_types |= DRM_FENCE_TYPE_EXE;
 			}
 		}
 
-		if (signaled_flush_types) {
+		if (signaled_types) {
 			drm_fence_handler(dev, 0, info->complete_sequence,
-					  signaled_flush_types, 0);
+					  signaled_types, 0);
 		}
 	}
 
 	DRM_SPINUNLOCK(&info->fence_lock);
-
-	return fc->pending_flush |
-		((fc->pending_exe_flush) ? DRM_FENCE_TYPE_EXE : 0);
 }
 
 
@@ -98,25 +91,13 @@ int xgi_fence_emit_sequence(struct drm_device * dev, uint32_t class,
 }
 
 
-void xgi_poke_flush(struct drm_device * dev, uint32_t class)
-{
-	struct drm_fence_manager * fm = &dev->fm;
-	unsigned long flags;
-
-
-	write_lock_irqsave(&fm->lock, flags);
-	xgi_do_flush(dev, class);
-	write_unlock_irqrestore(&fm->lock, flags);
-}
-
-
 void xgi_fence_handler(struct drm_device * dev)
 {
 	struct drm_fence_manager * fm = &dev->fm;
-
+	struct drm_fence_class_manager *fc = &fm->fence_class[0];
 
 	write_lock(&fm->lock);
-	xgi_do_flush(dev, 0);
+	xgi_fence_poll(dev, 0, fc->waiting_types);
 	write_unlock(&fm->lock);
 }
 
@@ -125,3 +106,17 @@ int xgi_fence_has_irq(struct drm_device *dev, uint32_t class, uint32_t flags)
 {
 	return ((class == 0) && (flags == DRM_FENCE_TYPE_EXE)) ? 1 : 0;
 }
+
+struct drm_fence_driver xgi_fence_driver = {
+	.num_classes = 1,
+	.wrap_diff = BEGIN_BEGIN_IDENTIFICATION_MASK,
+	.flush_diff = BEGIN_BEGIN_IDENTIFICATION_MASK - 1,
+	.sequence_mask = BEGIN_BEGIN_IDENTIFICATION_MASK,
+	.has_irq = xgi_fence_has_irq,
+	.emit = xgi_fence_emit_sequence,
+	.flush = NULL,
+	.poll = xgi_fence_poll,
+	.needed_flush = NULL,
+	.wait = NULL
+};
+
diff --git a/shared-core/i915_dma.c b/shared-core/i915_dma.c
index 287e95ac..15fb8117 100644
--- a/shared-core/i915_dma.c
+++ b/shared-core/i915_dma.c
@@ -422,6 +422,9 @@ void i915_emit_breadcrumb(struct drm_device *dev)
 	RING_LOCALS;
 
 	if (++dev_priv->counter > BREADCRUMB_MASK) {
+#ifdef I915_HAVE_FENCE
+		i915_invalidate_reported_sequence(dev);
+#endif
 		 dev_priv->counter = 1;
 		 DRM_DEBUG("Breadcrumb counter wrapped around\n");
 	}
@@ -1113,7 +1116,7 @@ static int i915_execbuffer(struct drm_device *dev, void *data,
 			fence_arg->handle = fence->base.hash.key;
 			fence_arg->fence_class = fence->fence_class;
 			fence_arg->type = fence->type;
-			fence_arg->signaled = fence->signaled;
+			fence_arg->signaled = fence->signaled_types;
 		}
 	}
 	drm_fence_usage_deref_unlocked(&fence);
diff --git a/shared-core/i915_drv.h b/shared-core/i915_drv.h
index c92758fa..3f6c8066 100644
--- a/shared-core/i915_drv.h
+++ b/shared-core/i915_drv.h
@@ -136,6 +136,8 @@ typedef struct drm_i915_private {
 	uint32_t flush_flags;
 	uint32_t flush_pending;
 	uint32_t saved_flush_status;
+	uint32_t reported_sequence;
+	int reported_sequence_valid;
 #endif
 #ifdef I915_HAVE_BUFFER
 	void *agp_iomap;
@@ -290,15 +292,9 @@ extern void i915_mem_release(struct drm_device * dev,
 			     struct mem_block *heap);
 #ifdef I915_HAVE_FENCE
 /* i915_fence.c */
-
-
 extern void i915_fence_handler(struct drm_device *dev);
-extern int i915_fence_emit_sequence(struct drm_device *dev, uint32_t class,
-				    uint32_t flags,
-				    uint32_t *sequence,
-				    uint32_t *native_type);
-extern void i915_poke_flush(struct drm_device *dev, uint32_t class);
-extern int i915_fence_has_irq(struct drm_device *dev, uint32_t class, uint32_t flags);
+extern void i915_invalidate_reported_sequence(struct drm_device *dev);
+
 #endif
 
 #ifdef I915_HAVE_BUFFER
diff --git a/shared-core/via_drv.c b/shared-core/via_drv.c
index a802e4ae..dd632c3d 100644
--- a/shared-core/via_drv.c
+++ b/shared-core/via_drv.c
@@ -40,17 +40,9 @@ static struct pci_device_id pciidlist[] = {
 
 
 #ifdef VIA_HAVE_FENCE
-static struct drm_fence_driver via_fence_driver = {
-	.num_classes = 1,
-	.wrap_diff = (1 << 30),
-	.flush_diff = (1 << 20),
-	.sequence_mask = 0xffffffffU,
-	.lazy_capable = 1,
-	.emit = via_fence_emit_sequence,
-	.poke_flush = via_poke_flush,
-	.has_irq = via_fence_has_irq,
-};
+extern struct drm_fence_driver via_fence_driver;
 #endif
+
 #ifdef VIA_HAVE_BUFFER
 
 /**
@@ -76,6 +68,8 @@ static struct drm_bo_driver via_bo_driver = {
 	.init_mem_type = via_init_mem_type,
 	.evict_flags = via_evict_flags,
 	.move = NULL,
+	.ttm_cache_flush = NULL,
+	.command_stream_barrier = NULL
 };
 #endif
 
diff --git a/shared-core/via_drv.h b/shared-core/via_drv.h
index 8dd4a727..941a2d77 100644
--- a/shared-core/via_drv.h
+++ b/shared-core/via_drv.h
@@ -196,17 +196,6 @@ extern void via_dmablit_handler(struct drm_device *dev, int engine, int from_irq
 extern void via_init_dmablit(struct drm_device *dev);
 #endif
 
-#ifdef VIA_HAVE_FENCE
-extern void via_fence_timer(unsigned long data);
-extern void via_poke_flush(struct drm_device * dev, uint32_t class);
-extern int via_fence_emit_sequence(struct drm_device * dev, uint32_t class,
-				   uint32_t flags,
-				   uint32_t * sequence,
-				   uint32_t * native_type);
-extern int via_fence_has_irq(struct drm_device * dev, uint32_t class,
-			     uint32_t flags);
-#endif
-
 #ifdef VIA_HAVE_BUFFER
 extern struct drm_ttm_backend *via_create_ttm_backend_entry(struct drm_device *dev);
 extern int via_fence_types(struct drm_buffer_object *bo, uint32_t *fclass,
diff --git a/shared-core/via_map.c b/shared-core/via_map.c
index 11bfa551..54934367 100644
--- a/shared-core/via_map.c
+++ b/shared-core/via_map.c
@@ -69,9 +69,6 @@ static int via_do_init_map(struct drm_device * dev, drm_via_init_t * init)
 	dev_priv->emit_0_sequence = 0;
 	dev_priv->have_idlelock = 0;
 	spin_lock_init(&dev_priv->fence_lock);
-	init_timer(&dev_priv->fence_timer);
-	dev_priv->fence_timer.function = &via_fence_timer;
-	dev_priv->fence_timer.data = (unsigned long) dev;
 #endif /* VIA_HAVE_FENCE */
 	dev->dev_private = (void *)dev_priv;
 #ifdef VIA_HAVE_BUFFER