vmwgfx: Add fence events

Add a way to send DRM events down the gpu fifo by attaching them to
fence objects. This may be useful for Xserver swapbuffer throttling and
page-flip done notifications.

Bump version to 2.2 to signal the availability of the FENCE_EVENT ioctl.

Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Jakob Bornecrantz <jakob@vmware.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_drv.c b/drivers/gpu/drm/vmwgfx/vmwgfx_drv.c
index ddb5abd..5e1994a 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_drv.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_drv.c
@@ -94,6 +94,9 @@
 #define DRM_IOCTL_VMW_FENCE_UNREF				\
 	DRM_IOW(DRM_COMMAND_BASE + DRM_VMW_FENCE_UNREF,		\
 		 struct drm_vmw_fence_arg)
+#define DRM_IOCTL_VMW_FENCE_EVENT				\
+	DRM_IOW(DRM_COMMAND_BASE + DRM_VMW_FENCE_EVENT,		\
+		 struct drm_vmw_fence_event_arg)
 #define DRM_IOCTL_VMW_PRESENT					\
 	DRM_IOW(DRM_COMMAND_BASE + DRM_VMW_PRESENT,		\
 		 struct drm_vmw_present_arg)
@@ -150,6 +153,9 @@
 		      DRM_AUTH | DRM_UNLOCKED),
 	VMW_IOCTL_DEF(VMW_FENCE_UNREF, vmw_fence_obj_unref_ioctl,
 		      DRM_AUTH | DRM_UNLOCKED),
+	VMW_IOCTL_DEF(VMW_FENCE_EVENT,
+		      vmw_fence_event_ioctl,
+		      DRM_AUTH | DRM_UNLOCKED),
 	VMW_IOCTL_DEF(VMW_GET_3D_CAP, vmw_get_cap_3d_ioctl,
 		      DRM_AUTH | DRM_UNLOCKED),
 
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_drv.h b/drivers/gpu/drm/vmwgfx/vmwgfx_drv.h
index 8cce73e..83b2563 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_drv.h
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_drv.h
@@ -40,9 +40,9 @@
 #include "ttm/ttm_module.h"
 #include "vmwgfx_fence.h"
 
-#define VMWGFX_DRIVER_DATE "20110927"
+#define VMWGFX_DRIVER_DATE "20111008"
 #define VMWGFX_DRIVER_MAJOR 2
-#define VMWGFX_DRIVER_MINOR 1
+#define VMWGFX_DRIVER_MINOR 2
 #define VMWGFX_DRIVER_PATCHLEVEL 0
 #define VMWGFX_FILE_PAGE_OFFSET 0x00100000
 #define VMWGFX_FIFO_STATIC_SIZE (1024*1024)
@@ -264,10 +264,12 @@
 	wait_queue_head_t fence_queue;
 	wait_queue_head_t fifo_queue;
 	int fence_queue_waiters; /* Protected by hw_mutex */
+	int goal_queue_waiters; /* Protected by hw_mutex */
 	atomic_t fifo_queue_waiters;
 	uint32_t last_read_seqno;
 	spinlock_t irq_lock;
 	struct vmw_fence_manager *fman;
+	uint32_t irq_mask;
 
 	/*
 	 * Device state
@@ -532,7 +534,13 @@
 				      struct vmw_private *dev_priv,
 				      struct vmw_fence_obj **p_fence,
 				      uint32_t *p_handle);
-
+extern void vmw_execbuf_copy_fence_user(struct vmw_private *dev_priv,
+					struct vmw_fpriv *vmw_fp,
+					int ret,
+					struct drm_vmw_fence_rep __user
+					*user_fence_rep,
+					struct vmw_fence_obj *fence,
+					uint32_t fence_handle);
 
 /**
  * IRQs and wating - vmwgfx_irq.c
@@ -557,6 +565,8 @@
 				struct vmw_fifo_state *fifo_state);
 extern void vmw_seqno_waiter_add(struct vmw_private *dev_priv);
 extern void vmw_seqno_waiter_remove(struct vmw_private *dev_priv);
+extern void vmw_goal_waiter_add(struct vmw_private *dev_priv);
+extern void vmw_goal_waiter_remove(struct vmw_private *dev_priv);
 
 /**
  * Rudimentary fence-like objects currently used only for throttling -
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c b/drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c
index f53b99d3..d4a1d8b 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c
@@ -1057,7 +1057,7 @@
  * object so we wait for it immediately, and then unreference the
  * user-space reference.
  */
-static void
+void
 vmw_execbuf_copy_fence_user(struct vmw_private *dev_priv,
 			    struct vmw_fpriv *vmw_fp,
 			    int ret,
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_fence.c b/drivers/gpu/drm/vmwgfx/vmwgfx_fence.c
index 5f60be7..35d5f61 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_fence.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_fence.c
@@ -34,13 +34,18 @@
 	int num_fence_objects;
 	struct vmw_private *dev_priv;
 	spinlock_t lock;
-	u32 next_seqno;
 	struct list_head fence_list;
 	struct work_struct work;
 	u32 user_fence_size;
 	u32 fence_size;
+	u32 event_fence_action_size;
 	bool fifo_down;
 	struct list_head cleanup_list;
+	uint32_t pending_actions[VMW_ACTION_MAX];
+	struct mutex goal_irq_mutex;
+	bool goal_irq_on; /* Protected by @goal_irq_mutex */
+	bool seqno_valid; /* Protected by @lock, and may not be set to true
+			     without the @goal_irq_mutex held. */
 };
 
 struct vmw_user_fence {
@@ -49,8 +54,51 @@
 };
 
 /**
- * vmw_fence_destroy_locked
+ * struct vmw_event_fence_action - fence action that delivers a drm event.
  *
+ * @e: A struct drm_pending_event that controls the event delivery.
+ * @action: A struct vmw_fence_action to hook up to a fence.
+ * @fence: A referenced pointer to the fence to keep it alive while @action
+ * hangs on it.
+ * @dev: Pointer to a struct drm_device so we can access the event stuff.
+ * @kref: Both @e and @action has destructors, so we need to refcount.
+ * @size: Size accounted for this object.
+ * @tv_sec: If non-null, the variable pointed to will be assigned
+ * current time tv_sec val when the fence signals.
+ * @tv_usec: Must be set if @tv_sec is set, and the variable pointed to will
+ * be assigned the current time tv_usec val when the fence signals.
+ */
+struct vmw_event_fence_action {
+	struct drm_pending_event e;
+	struct vmw_fence_action action;
+	struct vmw_fence_obj *fence;
+	struct drm_device *dev;
+	struct kref kref;
+	uint32_t size;
+	uint32_t *tv_sec;
+	uint32_t *tv_usec;
+};
+
+/**
+ * Note on fencing subsystem usage of irqs:
+ * Typically the vmw_fences_update function is called
+ *
+ * a) When a new fence seqno has been submitted by the fifo code.
+ * b) On-demand when we have waiters. Sleeping waiters will switch on the
+ * ANY_FENCE irq and call vmw_fences_update function each time an ANY_FENCE
+ * irq is received. When the last fence waiter is gone, that IRQ is masked
+ * away.
+ *
+ * In situations where there are no waiters and we don't submit any new fences,
+ * fence objects may not be signaled. This is perfectly OK, since there are
+ * no consumers of the signaled data, but that is NOT ok when there are fence
+ * actions attached to a fence. The fencing subsystem then makes use of the
+ * FENCE_GOAL irq and sets the fence goal seqno to that of the next fence
+ * which has an action attached, and each time vmw_fences_update is called,
+ * the subsystem makes sure the fence goal seqno is updated.
+ *
+ * The fence goal seqno irq is on as long as there are unsignaled fence
+ * objects with actions attached to them.
  */
 
 static void vmw_fence_obj_destroy_locked(struct kref *kref)
@@ -85,24 +133,36 @@
 		container_of(work, struct vmw_fence_manager, work);
 	struct list_head list;
 	struct vmw_fence_action *action, *next_action;
+	bool seqno_valid;
 
 	do {
 		INIT_LIST_HEAD(&list);
+		mutex_lock(&fman->goal_irq_mutex);
+
 		spin_lock_irq(&fman->lock);
 		list_splice_init(&fman->cleanup_list, &list);
+		seqno_valid = fman->seqno_valid;
 		spin_unlock_irq(&fman->lock);
 
+		if (!seqno_valid && fman->goal_irq_on) {
+			fman->goal_irq_on = false;
+			vmw_goal_waiter_remove(fman->dev_priv);
+		}
+		mutex_unlock(&fman->goal_irq_mutex);
+
 		if (list_empty(&list))
 			return;
 
 		/*
 		 * At this point, only we should be able to manipulate the
 		 * list heads of the actions we have on the private list.
+		 * hence fman::lock not held.
 		 */
 
 		list_for_each_entry_safe(action, next_action, &list, head) {
 			list_del_init(&action->head);
-			action->cleanup(action);
+			if (action->cleanup)
+				action->cleanup(action);
 		}
 	} while (1);
 }
@@ -122,6 +182,9 @@
 	fman->fifo_down = true;
 	fman->user_fence_size = ttm_round_pot(sizeof(struct vmw_user_fence));
 	fman->fence_size = ttm_round_pot(sizeof(struct vmw_fence_obj));
+	fman->event_fence_action_size =
+		ttm_round_pot(sizeof(struct vmw_event_fence_action));
+	mutex_init(&fman->goal_irq_mutex);
 
 	return fman;
 }
@@ -214,6 +277,7 @@
 
 	list_for_each_entry_safe(action, next_action, list, head) {
 		list_del_init(&action->head);
+		fman->pending_actions[action->type]--;
 		if (action->seq_passed != NULL)
 			action->seq_passed(action);
 
@@ -222,17 +286,101 @@
 		 * it will be performed by a worker task.
 		 */
 
-		if (action->cleanup != NULL)
-			list_add_tail(&action->head, &fman->cleanup_list);
+		list_add_tail(&action->head, &fman->cleanup_list);
 	}
 }
 
-void vmw_fences_update(struct vmw_fence_manager *fman, u32 seqno)
+/**
+ * vmw_fence_goal_new_locked - Figure out a new device fence goal
+ * seqno if needed.
+ *
+ * @fman: Pointer to a fence manager.
+ * @passed_seqno: The seqno the device currently signals as passed.
+ *
+ * This function should be called with the fence manager lock held.
+ * It is typically called when we have a new passed_seqno, and
+ * we might need to update the fence goal. It checks to see whether
+ * the current fence goal has already passed, and, in that case,
+ * scans through all unsignaled fences to get the next fence object with an
+ * action attached, and sets the seqno of that fence as a new fence goal.
+ *
+ * returns true if the device goal seqno was updated. False otherwise.
+ */
+static bool vmw_fence_goal_new_locked(struct vmw_fence_manager *fman,
+				      u32 passed_seqno)
+{
+	u32 goal_seqno;
+	__le32 __iomem *fifo_mem;
+	struct vmw_fence_obj *fence;
+
+	if (likely(!fman->seqno_valid))
+		return false;
+
+	fifo_mem = fman->dev_priv->mmio_virt;
+	goal_seqno = ioread32(fifo_mem + SVGA_FIFO_FENCE_GOAL);
+	if (likely(passed_seqno - goal_seqno >= VMW_FENCE_WRAP))
+		return false;
+
+	fman->seqno_valid = false;
+	list_for_each_entry(fence, &fman->fence_list, head) {
+		if (!list_empty(&fence->seq_passed_actions)) {
+			fman->seqno_valid = true;
+			iowrite32(fence->seqno,
+				  fifo_mem + SVGA_FIFO_FENCE_GOAL);
+			break;
+		}
+	}
+
+	return true;
+}
+
+
+/**
+ * vmw_fence_goal_check_locked - Replace the device fence goal seqno if
+ * needed.
+ *
+ * @fence: Pointer to a struct vmw_fence_obj the seqno of which should be
+ * considered as a device fence goal.
+ *
+ * This function should be called with the fence manager lock held.
+ * It is typically called when an action has been attached to a fence to
+ * check whether the seqno of that fence should be used for a fence
+ * goal interrupt. This is typically needed if the current fence goal is
+ * invalid, or has a higher seqno than that of the current fence object.
+ *
+ * returns true if the device goal seqno was updated. False otherwise.
+ */
+static bool vmw_fence_goal_check_locked(struct vmw_fence_obj *fence)
+{
+	u32 goal_seqno;
+	__le32 __iomem *fifo_mem;
+
+	if (fence->signaled & DRM_VMW_FENCE_FLAG_EXEC)
+		return false;
+
+	fifo_mem = fence->fman->dev_priv->mmio_virt;
+	goal_seqno = ioread32(fifo_mem + SVGA_FIFO_FENCE_GOAL);
+	if (likely(fence->fman->seqno_valid &&
+		   goal_seqno - fence->seqno < VMW_FENCE_WRAP))
+		return false;
+
+	iowrite32(fence->seqno, fifo_mem + SVGA_FIFO_FENCE_GOAL);
+	fence->fman->seqno_valid = true;
+
+	return true;
+}
+
+void vmw_fences_update(struct vmw_fence_manager *fman)
 {
 	unsigned long flags;
 	struct vmw_fence_obj *fence, *next_fence;
 	struct list_head action_list;
+	bool needs_rerun;
+	uint32_t seqno, new_seqno;
+	__le32 __iomem *fifo_mem = fman->dev_priv->mmio_virt;
 
+	seqno = ioread32(fifo_mem + SVGA_FIFO_FENCE);
+rerun:
 	spin_lock_irqsave(&fman->lock, flags);
 	list_for_each_entry_safe(fence, next_fence, &fman->fence_list, head) {
 		if (seqno - fence->seqno < VMW_FENCE_WRAP) {
@@ -243,14 +391,30 @@
 					 &action_list);
 			vmw_fences_perform_actions(fman, &action_list);
 			wake_up_all(&fence->queue);
-		}
-
+		} else
+			break;
 	}
+
+	needs_rerun = vmw_fence_goal_new_locked(fman, seqno);
+
 	if (!list_empty(&fman->cleanup_list))
 		(void) schedule_work(&fman->work);
 	spin_unlock_irqrestore(&fman->lock, flags);
-}
 
+	/*
+	 * Rerun if the fence goal seqno was updated, and the
+	 * hardware might have raced with that update, so that
+	 * we missed a fence_goal irq.
+	 */
+
+	if (unlikely(needs_rerun)) {
+		new_seqno = ioread32(fifo_mem + SVGA_FIFO_FENCE);
+		if (new_seqno != seqno) {
+			seqno = new_seqno;
+			goto rerun;
+		}
+	}
+}
 
 bool vmw_fence_obj_signaled(struct vmw_fence_obj *fence,
 			    uint32_t flags)
@@ -267,14 +431,8 @@
 	if ((signaled & flags) == flags)
 		return 1;
 
-	if ((signaled & DRM_VMW_FENCE_FLAG_EXEC) == 0) {
-		struct vmw_private *dev_priv = fman->dev_priv;
-		__le32 __iomem *fifo_mem = dev_priv->mmio_virt;
-		u32 seqno;
-
-		seqno = ioread32(fifo_mem + SVGA_FIFO_FENCE);
-		vmw_fences_update(fman, seqno);
-	}
+	if ((signaled & DRM_VMW_FENCE_FLAG_EXEC) == 0)
+		vmw_fences_update(fman);
 
 	spin_lock_irqsave(&fman->lock, irq_flags);
 	signaled = fence->signaled;
@@ -624,3 +782,344 @@
 					 arg->handle,
 					 TTM_REF_USAGE);
 }
+
+/**
+ * vmw_event_fence_action_destroy
+ *
+ * @kref: The struct kref embedded in a struct vmw_event_fence_action.
+ *
+ * The vmw_event_fence_action destructor that may be called either after
+ * the fence action cleanup, or when the event is delivered.
+ * It frees both the vmw_event_fence_action struct and the actual
+ * event structure copied to user-space.
+ */
+static void vmw_event_fence_action_destroy(struct kref *kref)
+{
+	struct vmw_event_fence_action *eaction =
+		container_of(kref, struct vmw_event_fence_action, kref);
+	struct ttm_mem_global *mem_glob =
+		vmw_mem_glob(vmw_priv(eaction->dev));
+	uint32_t size = eaction->size;
+
+	kfree(eaction->e.event);
+	kfree(eaction);
+	ttm_mem_global_free(mem_glob, size);
+}
+
+
+/**
+ * vmw_event_fence_action_delivered
+ *
+ * @e: The struct drm_pending_event embedded in a struct
+ * vmw_event_fence_action.
+ *
+ * The struct drm_pending_event destructor that is called by drm
+ * once the event is delivered. Since we don't know whether this function
+ * will be called before or after the fence action destructor, we
+ * free a refcount and destroy if it becomes zero.
+ */
+static void vmw_event_fence_action_delivered(struct drm_pending_event *e)
+{
+	struct vmw_event_fence_action *eaction =
+		container_of(e, struct vmw_event_fence_action, e);
+
+	kref_put(&eaction->kref, vmw_event_fence_action_destroy);
+}
+
+
+/**
+ * vmw_event_fence_action_seq_passed
+ *
+ * @action: The struct vmw_fence_action embedded in a struct
+ * vmw_event_fence_action.
+ *
+ * This function is called when the seqno of the fence where @action is
+ * attached has passed. It queues the event on the submitter's event list.
+ * This function is always called from atomic context, and may be called
+ * from irq context. It ups a refcount reflecting that we now have two
+ * destructors.
+ */
+static void vmw_event_fence_action_seq_passed(struct vmw_fence_action *action)
+{
+	struct vmw_event_fence_action *eaction =
+		container_of(action, struct vmw_event_fence_action, action);
+	struct drm_device *dev = eaction->dev;
+	struct drm_file *file_priv = eaction->e.file_priv;
+	unsigned long irq_flags;
+
+	kref_get(&eaction->kref);
+	spin_lock_irqsave(&dev->event_lock, irq_flags);
+
+	if (likely(eaction->tv_sec != NULL)) {
+		struct timeval tv;
+
+		do_gettimeofday(&tv);
+		*eaction->tv_sec = tv.tv_sec;
+		*eaction->tv_usec = tv.tv_usec;
+	}
+
+	list_add_tail(&eaction->e.link, &file_priv->event_list);
+	wake_up_all(&file_priv->event_wait);
+	spin_unlock_irqrestore(&dev->event_lock, irq_flags);
+}
+
+/**
+ * vmw_event_fence_action_cleanup
+ *
+ * @action: The struct vmw_fence_action embedded in a struct
+ * vmw_event_fence_action.
+ *
+ * This function is the struct vmw_fence_action destructor. It's typically
+ * called from a workqueue.
+ */
+static void vmw_event_fence_action_cleanup(struct vmw_fence_action *action)
+{
+	struct vmw_event_fence_action *eaction =
+		container_of(action, struct vmw_event_fence_action, action);
+
+	vmw_fence_obj_unreference(&eaction->fence);
+	kref_put(&eaction->kref, vmw_event_fence_action_destroy);
+}
+
+
+/**
+ * vmw_fence_obj_add_action - Add an action to a fence object.
+ *
+ * @fence - The fence object.
+ * @action - The action to add.
+ *
+ * Note that the action callbacks may be executed before this function
+ * returns.
+ */
+void vmw_fence_obj_add_action(struct vmw_fence_obj *fence,
+			      struct vmw_fence_action *action)
+{
+	struct vmw_fence_manager *fman = fence->fman;
+	unsigned long irq_flags;
+	bool run_update = false;
+
+	mutex_lock(&fman->goal_irq_mutex);
+	spin_lock_irqsave(&fman->lock, irq_flags);
+
+	fman->pending_actions[action->type]++;
+	if (fence->signaled & DRM_VMW_FENCE_FLAG_EXEC) {
+		struct list_head action_list;
+
+		INIT_LIST_HEAD(&action_list);
+		list_add_tail(&action->head, &action_list);
+		vmw_fences_perform_actions(fman, &action_list);
+	} else {
+		list_add_tail(&action->head, &fence->seq_passed_actions);
+
+		/*
+		 * This function may set fman::seqno_valid, so it must
+		 * be run with the goal_irq_mutex held.
+		 */
+		run_update = vmw_fence_goal_check_locked(fence);
+	}
+
+	spin_unlock_irqrestore(&fman->lock, irq_flags);
+
+	if (run_update) {
+		if (!fman->goal_irq_on) {
+			fman->goal_irq_on = true;
+			vmw_goal_waiter_add(fman->dev_priv);
+		}
+		vmw_fences_update(fman);
+	}
+	mutex_unlock(&fman->goal_irq_mutex);
+
+}
+
+/**
+ * vmw_event_fence_action_create - Post an event for sending when a fence
+ * object seqno has passed.
+ *
+ * @file_priv: The file connection on which the event should be posted.
+ * @fence: The fence object on which to post the event.
+ * @event: Event to be posted. This event should've been alloced
+ * using k[mz]alloc, and should've been completely initialized.
+ * @interruptible: Interruptible waits if possible.
+ *
+ * As a side effect, the object pointed to by @event may have been
+ * freed when this function returns. If this function returns with
+ * an error code, the caller needs to free that object.
+ */
+
+int vmw_event_fence_action_create(struct drm_file *file_priv,
+				  struct vmw_fence_obj *fence,
+				  struct drm_event *event,
+				  uint32_t *tv_sec,
+				  uint32_t *tv_usec,
+				  bool interruptible)
+{
+	struct vmw_event_fence_action *eaction =
+		kzalloc(sizeof(*eaction), GFP_KERNEL);
+	struct ttm_mem_global *mem_glob =
+		vmw_mem_glob(fence->fman->dev_priv);
+	struct vmw_fence_manager *fman = fence->fman;
+	uint32_t size = fman->event_fence_action_size +
+		ttm_round_pot(event->length);
+	int ret;
+
+	/*
+	 * Account for internal structure size as well as the
+	 * event size itself.
+	 */
+
+	ret = ttm_mem_global_alloc(mem_glob, size, false, interruptible);
+	if (unlikely(ret != 0))
+		return ret;
+
+	eaction = kzalloc(sizeof(*eaction), GFP_KERNEL);
+	if (unlikely(eaction == NULL)) {
+		ttm_mem_global_free(mem_glob, size);
+		return -ENOMEM;
+	}
+
+	eaction->e.event = event;
+	eaction->e.file_priv = file_priv;
+	eaction->e.destroy = vmw_event_fence_action_delivered;
+
+	eaction->action.seq_passed = vmw_event_fence_action_seq_passed;
+	eaction->action.cleanup = vmw_event_fence_action_cleanup;
+	eaction->action.type = VMW_ACTION_EVENT;
+
+	eaction->fence = vmw_fence_obj_reference(fence);
+	eaction->dev = fman->dev_priv->dev;
+	eaction->size = size;
+	eaction->tv_sec = tv_sec;
+	eaction->tv_usec = tv_usec;
+
+	kref_init(&eaction->kref);
+	vmw_fence_obj_add_action(fence, &eaction->action);
+
+	return 0;
+}
+
+int vmw_fence_event_ioctl(struct drm_device *dev, void *data,
+			  struct drm_file *file_priv)
+{
+	struct vmw_private *dev_priv = vmw_priv(dev);
+	struct drm_vmw_fence_event_arg *arg =
+		(struct drm_vmw_fence_event_arg *) data;
+	struct vmw_fence_obj *fence = NULL;
+	struct vmw_fpriv *vmw_fp = vmw_fpriv(file_priv);
+	struct drm_vmw_fence_rep __user *user_fence_rep =
+		(struct drm_vmw_fence_rep __user *)(unsigned long)
+		arg->fence_rep;
+	uint32_t handle;
+	unsigned long irq_flags;
+	struct drm_vmw_event_fence *event;
+	int ret;
+
+	/*
+	 * Look up an existing fence object,
+	 * and if user-space wants a new reference,
+	 * add one.
+	 */
+	if (arg->handle) {
+		struct ttm_base_object *base =
+			ttm_base_object_lookup(vmw_fp->tfile, arg->handle);
+
+		if (unlikely(base == NULL)) {
+			DRM_ERROR("Fence event invalid fence object handle "
+				  "0x%08lx.\n",
+				  (unsigned long)arg->handle);
+			return -EINVAL;
+		}
+		fence = &(container_of(base, struct vmw_user_fence,
+				       base)->fence);
+		(void) vmw_fence_obj_reference(fence);
+
+		if (user_fence_rep != NULL) {
+			bool existed;
+
+			ret = ttm_ref_object_add(vmw_fp->tfile, base,
+						 TTM_REF_USAGE, &existed);
+			if (unlikely(ret != 0)) {
+				DRM_ERROR("Failed to reference a fence "
+					  "object.\n");
+				goto out_no_ref_obj;
+			}
+			handle = base->hash.key;
+		}
+		ttm_base_object_unref(&base);
+	}
+
+	/*
+	 * Create a new fence object.
+	 */
+	if (!fence) {
+		ret = vmw_execbuf_fence_commands(file_priv, dev_priv,
+						 &fence,
+						 (user_fence_rep) ?
+						 &handle : NULL);
+		if (unlikely(ret != 0)) {
+			DRM_ERROR("Fence event failed to create fence.\n");
+			return ret;
+		}
+	}
+
+	BUG_ON(fence == NULL);
+
+	spin_lock_irqsave(&dev->event_lock, irq_flags);
+
+	ret = (file_priv->event_space < sizeof(*event)) ? -EBUSY : 0;
+	if (likely(ret == 0))
+		file_priv->event_space -= sizeof(*event);
+
+	spin_unlock_irqrestore(&dev->event_lock, irq_flags);
+
+	if (unlikely(ret != 0)) {
+		DRM_ERROR("Failed to allocate event space for this file.\n");
+		goto out_no_event_space;
+	}
+
+	event = kzalloc(sizeof(*event), GFP_KERNEL);
+	if (unlikely(event == NULL)) {
+		DRM_ERROR("Failed to allocate an event.\n");
+		goto out_no_event;
+	}
+
+	event->base.type = DRM_VMW_EVENT_FENCE_SIGNALED;
+	event->base.length = sizeof(*event);
+	event->user_data = arg->user_data;
+
+	if (arg->flags & DRM_VMW_FE_FLAG_REQ_TIME)
+		ret = vmw_event_fence_action_create(file_priv, fence,
+						    &event->base,
+						    &event->tv_sec,
+						    &event->tv_usec,
+						    true);
+	else
+		ret = vmw_event_fence_action_create(file_priv, fence,
+						    &event->base,
+						    NULL,
+						    NULL,
+						    true);
+
+	if (unlikely(ret != 0)) {
+		if (ret != -ERESTARTSYS)
+			DRM_ERROR("Failed to attach event to fence.\n");
+		goto out_no_attach;
+	}
+
+	vmw_execbuf_copy_fence_user(dev_priv, vmw_fp, 0, user_fence_rep, fence,
+				    handle);
+	vmw_fence_obj_unreference(&fence);
+	return 0;
+out_no_attach:
+	kfree(event);
+out_no_event:
+	spin_lock_irqsave(&dev->event_lock, irq_flags);
+	file_priv->event_space += sizeof(*event);
+	spin_unlock_irqrestore(&dev->event_lock, irq_flags);
+out_no_event_space:
+	if (user_fence_rep != NULL)
+		ttm_ref_object_base_unref(vmw_fpriv(file_priv)->tfile,
+					  handle, TTM_REF_USAGE);
+out_no_ref_obj:
+	vmw_fence_obj_unreference(&fence);
+	return ret;
+}
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_fence.h b/drivers/gpu/drm/vmwgfx/vmwgfx_fence.h
index 9307406..0854a20 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_fence.h
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_fence.h
@@ -37,8 +37,14 @@
  *
  *
  */
+enum vmw_action_type {
+	VMW_ACTION_EVENT = 0,
+	VMW_ACTION_MAX
+};
+
 struct vmw_fence_action {
 	struct list_head head;
+	enum vmw_action_type type;
 	void (*seq_passed) (struct vmw_fence_action *action);
 	void (*cleanup) (struct vmw_fence_action *action);
 };
@@ -66,8 +72,7 @@
 extern struct vmw_fence_obj *
 vmw_fence_obj_reference(struct vmw_fence_obj *fence);
 
-extern void vmw_fences_update(struct vmw_fence_manager *fman,
-			      u32 sequence);
+extern void vmw_fences_update(struct vmw_fence_manager *fman);
 
 extern bool vmw_fence_obj_signaled(struct vmw_fence_obj *fence,
 				   uint32_t flags);
@@ -102,4 +107,7 @@
 
 extern int vmw_fence_obj_unref_ioctl(struct drm_device *dev, void *data,
 				     struct drm_file *file_priv);
+extern int vmw_fence_event_ioctl(struct drm_device *dev, void *data,
+				 struct drm_file *file_priv);
+
 #endif /* _VMWGFX_FENCE_H_ */
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_fifo.c b/drivers/gpu/drm/vmwgfx/vmwgfx_fifo.c
index 62d6377..03bbc2a 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_fifo.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_fifo.c
@@ -247,9 +247,8 @@
 		spin_lock_irqsave(&dev_priv->irq_lock, irq_flags);
 		outl(SVGA_IRQFLAG_FIFO_PROGRESS,
 		     dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
-		vmw_write(dev_priv, SVGA_REG_IRQMASK,
-			  vmw_read(dev_priv, SVGA_REG_IRQMASK) |
-			  SVGA_IRQFLAG_FIFO_PROGRESS);
+		dev_priv->irq_mask |= SVGA_IRQFLAG_FIFO_PROGRESS;
+		vmw_write(dev_priv, SVGA_REG_IRQMASK, dev_priv->irq_mask);
 		spin_unlock_irqrestore(&dev_priv->irq_lock, irq_flags);
 	}
 	mutex_unlock(&dev_priv->hw_mutex);
@@ -271,9 +270,8 @@
 	mutex_lock(&dev_priv->hw_mutex);
 	if (atomic_dec_and_test(&dev_priv->fifo_queue_waiters)) {
 		spin_lock_irqsave(&dev_priv->irq_lock, irq_flags);
-		vmw_write(dev_priv, SVGA_REG_IRQMASK,
-			  vmw_read(dev_priv, SVGA_REG_IRQMASK) &
-			  ~SVGA_IRQFLAG_FIFO_PROGRESS);
+		dev_priv->irq_mask &= ~SVGA_IRQFLAG_FIFO_PROGRESS;
+		vmw_write(dev_priv, SVGA_REG_IRQMASK, dev_priv->irq_mask);
 		spin_unlock_irqrestore(&dev_priv->irq_lock, irq_flags);
 	}
 	mutex_unlock(&dev_priv->hw_mutex);
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_irq.c b/drivers/gpu/drm/vmwgfx/vmwgfx_irq.c
index a005292..cabc95f 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_irq.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_irq.c
@@ -34,28 +34,30 @@
 {
 	struct drm_device *dev = (struct drm_device *)arg;
 	struct vmw_private *dev_priv = vmw_priv(dev);
-	uint32_t status;
+	uint32_t status, masked_status;
 
 	spin_lock(&dev_priv->irq_lock);
 	status = inl(dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
+	masked_status = status & dev_priv->irq_mask;
 	spin_unlock(&dev_priv->irq_lock);
 
-	if (status & SVGA_IRQFLAG_ANY_FENCE) {
-		__le32 __iomem *fifo_mem = dev_priv->mmio_virt;
-		uint32_t seqno = ioread32(fifo_mem + SVGA_FIFO_FENCE);
+	if (likely(status))
+		outl(status, dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
 
-		vmw_fences_update(dev_priv->fman, seqno);
+	if (!masked_status)
+		return IRQ_NONE;
+
+	if (masked_status & (SVGA_IRQFLAG_ANY_FENCE |
+			     SVGA_IRQFLAG_FENCE_GOAL)) {
+		vmw_fences_update(dev_priv->fman);
 		wake_up_all(&dev_priv->fence_queue);
 	}
-	if (status & SVGA_IRQFLAG_FIFO_PROGRESS)
+
+	if (masked_status & SVGA_IRQFLAG_FIFO_PROGRESS)
 		wake_up_all(&dev_priv->fifo_queue);
 
-	if (likely(status)) {
-		outl(status, dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
-		return IRQ_HANDLED;
-	}
 
-	return IRQ_NONE;
+	return IRQ_HANDLED;
 }
 
 static bool vmw_fifo_idle(struct vmw_private *dev_priv, uint32_t seqno)
@@ -78,7 +80,7 @@
 	if (dev_priv->last_read_seqno != seqno) {
 		dev_priv->last_read_seqno = seqno;
 		vmw_marker_pull(&fifo_state->marker_queue, seqno);
-		vmw_fences_update(dev_priv->fman, seqno);
+		vmw_fences_update(dev_priv->fman);
 	}
 }
 
@@ -189,9 +191,8 @@
 		spin_lock_irqsave(&dev_priv->irq_lock, irq_flags);
 		outl(SVGA_IRQFLAG_ANY_FENCE,
 		     dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
-		vmw_write(dev_priv, SVGA_REG_IRQMASK,
-			  vmw_read(dev_priv, SVGA_REG_IRQMASK) |
-			  SVGA_IRQFLAG_ANY_FENCE);
+		dev_priv->irq_mask |= SVGA_IRQFLAG_ANY_FENCE;
+		vmw_write(dev_priv, SVGA_REG_IRQMASK, dev_priv->irq_mask);
 		spin_unlock_irqrestore(&dev_priv->irq_lock, irq_flags);
 	}
 	mutex_unlock(&dev_priv->hw_mutex);
@@ -204,9 +205,39 @@
 		unsigned long irq_flags;
 
 		spin_lock_irqsave(&dev_priv->irq_lock, irq_flags);
-		vmw_write(dev_priv, SVGA_REG_IRQMASK,
-			  vmw_read(dev_priv, SVGA_REG_IRQMASK) &
-			  ~SVGA_IRQFLAG_ANY_FENCE);
+		dev_priv->irq_mask &= ~SVGA_IRQFLAG_ANY_FENCE;
+		vmw_write(dev_priv, SVGA_REG_IRQMASK, dev_priv->irq_mask);
+		spin_unlock_irqrestore(&dev_priv->irq_lock, irq_flags);
+	}
+	mutex_unlock(&dev_priv->hw_mutex);
+}
+
+
+void vmw_goal_waiter_add(struct vmw_private *dev_priv)
+{
+	mutex_lock(&dev_priv->hw_mutex);
+	if (dev_priv->goal_queue_waiters++ == 0) {
+		unsigned long irq_flags;
+
+		spin_lock_irqsave(&dev_priv->irq_lock, irq_flags);
+		outl(SVGA_IRQFLAG_FENCE_GOAL,
+		     dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
+		dev_priv->irq_mask |= SVGA_IRQFLAG_FENCE_GOAL;
+		vmw_write(dev_priv, SVGA_REG_IRQMASK, dev_priv->irq_mask);
+		spin_unlock_irqrestore(&dev_priv->irq_lock, irq_flags);
+	}
+	mutex_unlock(&dev_priv->hw_mutex);
+}
+
+void vmw_goal_waiter_remove(struct vmw_private *dev_priv)
+{
+	mutex_lock(&dev_priv->hw_mutex);
+	if (--dev_priv->goal_queue_waiters == 0) {
+		unsigned long irq_flags;
+
+		spin_lock_irqsave(&dev_priv->irq_lock, irq_flags);
+		dev_priv->irq_mask &= ~SVGA_IRQFLAG_FENCE_GOAL;
+		vmw_write(dev_priv, SVGA_REG_IRQMASK, dev_priv->irq_mask);
 		spin_unlock_irqrestore(&dev_priv->irq_lock, irq_flags);
 	}
 	mutex_unlock(&dev_priv->hw_mutex);