drm/vmwgfx: Add command buffer support v3

Add command buffer support.
Currently we don't implement preemption or fancy error handling.
Tested with a couple of mesa-demos, compiz/unity and viewperf maya-03.

v2:
- Synchronize with pending work at command buffer manager takedown.
- Add an interface to flush the current command buffer for latency-critical
  command batches and apply it to framebuffer dirtying.

v3:
- Minor fixes of definitions and typos to address reviews.
- Removed new or moved branch predictor hints.
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Sinclair Yeh <syeh@vmware.com>

drivers/gpu/drm/vmwgfx/Makefile

@@ -7,6 +7,6 @@ vmwgfx-y := vmwgfx_execbuf.o vmwgfx_gmr.o vmwgfx_kms.o vmwgfx_drv.o \
 	    vmwgfx_overlay.o vmwgfx_marker.o vmwgfx_gmrid_manager.o \
 	    vmwgfx_fence.o vmwgfx_dmabuf.o vmwgfx_scrn.o vmwgfx_context.o \
 	    vmwgfx_surface.o vmwgfx_prime.o vmwgfx_mob.o vmwgfx_shader.o \
-	    vmwgfx_cmdbuf_res.o \
+	    vmwgfx_cmdbuf_res.o vmwgfx_cmdbuf.o \
 
 obj-$(CONFIG_DRM_VMWGFX) := vmwgfx.o

drivers/gpu/drm/vmwgfx/vmwgfx_buffer.c

@@ -72,6 +72,12 @@ static struct ttm_place mob_placement_flags = {
 	.flags = VMW_PL_FLAG_MOB | TTM_PL_FLAG_CACHED
 };
 
+static struct ttm_place mob_ne_placement_flags = {
+	.fpfn = 0,
+	.lpfn = 0,
+	.flags = VMW_PL_FLAG_MOB | TTM_PL_FLAG_CACHED | TTM_PL_FLAG_NO_EVICT
+};
+
 struct ttm_placement vmw_vram_placement = {
 	.num_placement = 1,
 	.placement = &vram_placement_flags,
@@ -200,6 +206,13 @@ struct ttm_placement vmw_mob_placement = {
 	.busy_placement = &mob_placement_flags
 };
 
+struct ttm_placement vmw_mob_ne_placement = {
+	.num_placement = 1,
+	.num_busy_placement = 1,
+	.placement = &mob_ne_placement_flags,
+	.busy_placement = &mob_ne_placement_flags
+};
+
 struct vmw_ttm_tt {
 	struct ttm_dma_tt dma_ttm;
 	struct vmw_private *dev_priv;

drivers/gpu/drm/vmwgfx/vmwgfx_drv.c

@@ -278,6 +278,8 @@ static void vmw_print_capabilities(uint32_t capabilities)
 		DRM_INFO("  Command Buffers 2.\n");
 	if (capabilities & SVGA_CAP_GBOBJECTS)
 		DRM_INFO("  Guest Backed Resources.\n");
+	if (capabilities & SVGA_CAP_CMD_BUFFERS_3)
+		DRM_INFO("  Command Buffers 3.\n");
 }
 
 /**
@@ -362,6 +364,17 @@ static int vmw_request_device_late(struct vmw_private *dev_priv)
 		}
 	}
 
+	if (dev_priv->cman) {
+		ret = vmw_cmdbuf_set_pool_size(dev_priv->cman,
+					       256*4096, 2*4096);
+		if (ret) {
+			struct vmw_cmdbuf_man *man = dev_priv->cman;
+
+			dev_priv->cman = NULL;
+			vmw_cmdbuf_man_destroy(man);
+		}
+	}
+
 	return 0;
 }
 
@@ -375,6 +388,9 @@ static int vmw_request_device(struct vmw_private *dev_priv)
 		return ret;
 	}
 	vmw_fence_fifo_up(dev_priv->fman);
+	dev_priv->cman = vmw_cmdbuf_man_create(dev_priv);
+	if (IS_ERR(dev_priv->cman))
+		dev_priv->cman = NULL;
 
 	ret = vmw_request_device_late(dev_priv);
 	if (ret)
@@ -387,10 +403,14 @@ static int vmw_request_device(struct vmw_private *dev_priv)
 	return 0;
 
 out_no_query_bo:
+	if (dev_priv->cman)
+		vmw_cmdbuf_remove_pool(dev_priv->cman);
 	if (dev_priv->has_mob) {
 		(void) ttm_bo_evict_mm(&dev_priv->bdev, VMW_PL_MOB);
 		vmw_otables_takedown(dev_priv);
 	}
+	if (dev_priv->cman)
+		vmw_cmdbuf_man_destroy(dev_priv->cman);
 out_no_mob:
 	vmw_fence_fifo_down(dev_priv->fman);
 	vmw_fifo_release(dev_priv, &dev_priv->fifo);
@@ -415,6 +435,9 @@ static void vmw_release_device_early(struct vmw_private *dev_priv)
 	BUG_ON(dev_priv->pinned_bo != NULL);
 
 	ttm_bo_unref(&dev_priv->dummy_query_bo);
+	if (dev_priv->cman)
+		vmw_cmdbuf_remove_pool(dev_priv->cman);
+
 	if (dev_priv->has_mob) {
 		ttm_bo_evict_mm(&dev_priv->bdev, VMW_PL_MOB);
 		vmw_otables_takedown(dev_priv);
@@ -432,6 +455,9 @@ static void vmw_release_device_early(struct vmw_private *dev_priv)
 static void vmw_release_device_late(struct vmw_private *dev_priv)
 {
 	vmw_fence_fifo_down(dev_priv->fman);
+	if (dev_priv->cman)
+		vmw_cmdbuf_man_destroy(dev_priv->cman);
+
 	vmw_fifo_release(dev_priv, &dev_priv->fifo);
 }
 

drivers/gpu/drm/vmwgfx/vmwgfx_drv.h

@@ -453,6 +453,8 @@ struct vmw_private {
 	spinlock_t waiter_lock;
 	int fence_queue_waiters; /* Protected by waiter_lock */
 	int goal_queue_waiters; /* Protected by waiter_lock */
+	int cmdbuf_waiters; /* Protected by irq_lock */
+	int error_waiters; /* Protected by irq_lock */
 	atomic_t fifo_queue_waiters;
 	uint32_t last_read_seqno;
 	spinlock_t irq_lock;
@@ -535,6 +537,8 @@ struct vmw_private {
 	 */
 	struct ttm_buffer_object *otable_bo;
 	struct vmw_otable *otables;
+
+	struct vmw_cmdbuf_man *cman;
 };
 
 static inline struct vmw_surface *vmw_res_to_srf(struct vmw_resource *res)
@@ -729,6 +733,8 @@ extern bool vmw_fifo_have_3d(struct vmw_private *dev_priv);
 extern bool vmw_fifo_have_pitchlock(struct vmw_private *dev_priv);
 extern int vmw_fifo_emit_dummy_query(struct vmw_private *dev_priv,
 				     uint32_t cid);
+extern int vmw_fifo_flush(struct vmw_private *dev_priv,
+			  bool interruptible);
 
 /**
  * TTM glue - vmwgfx_ttm_glue.c
@@ -753,6 +759,7 @@ extern struct ttm_placement vmw_sys_ne_placement;
 extern struct ttm_placement vmw_evictable_placement;
 extern struct ttm_placement vmw_srf_placement;
 extern struct ttm_placement vmw_mob_placement;
+extern struct ttm_placement vmw_mob_ne_placement;
 extern struct ttm_bo_driver vmw_bo_driver;
 extern int vmw_dma_quiescent(struct drm_device *dev);
 extern int vmw_bo_map_dma(struct ttm_buffer_object *bo);
@@ -855,6 +862,10 @@ 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);
+extern void vmw_generic_waiter_add(struct vmw_private *dev_priv, u32 flag,
+				   int *waiter_count);
+extern void vmw_generic_waiter_remove(struct vmw_private *dev_priv,
+				      u32 flag, int *waiter_count);
 
 /**
  * Rudimentary fence-like objects currently used only for throttling -
@@ -1077,6 +1088,35 @@ extern int vmw_cmdbuf_res_remove(struct vmw_cmdbuf_res_manager *man,
 				 struct list_head *list);
 
 
+/*
+ * Command buffer managerment vmwgfx_cmdbuf.c
+ */
+struct vmw_cmdbuf_man;
+struct vmw_cmdbuf_header;
+
+extern struct vmw_cmdbuf_man *
+vmw_cmdbuf_man_create(struct vmw_private *dev_priv);
+extern int vmw_cmdbuf_set_pool_size(struct vmw_cmdbuf_man *man,
+				    size_t size, size_t default_size);
+extern void vmw_cmdbuf_remove_pool(struct vmw_cmdbuf_man *man);
+extern void vmw_cmdbuf_man_destroy(struct vmw_cmdbuf_man *man);
+extern int vmw_cmdbuf_idle(struct vmw_cmdbuf_man *man, bool interruptible,
+			   unsigned long timeout);
+extern void *vmw_cmdbuf_reserve(struct vmw_cmdbuf_man *man, size_t size,
+				int ctx_id, bool interruptible,
+				struct vmw_cmdbuf_header *header);
+extern void vmw_cmdbuf_commit(struct vmw_cmdbuf_man *man, size_t size,
+			      struct vmw_cmdbuf_header *header,
+			      bool flush);
+extern void vmw_cmdbuf_tasklet_schedule(struct vmw_cmdbuf_man *man);
+extern void *vmw_cmdbuf_alloc(struct vmw_cmdbuf_man *man,
+			      size_t size, bool interruptible,
+			      struct vmw_cmdbuf_header **p_header);
+extern void vmw_cmdbuf_header_free(struct vmw_cmdbuf_header *header);
+extern int vmw_cmdbuf_cur_flush(struct vmw_cmdbuf_man *man,
+				bool interruptible);
+
+
 /**
  * Inline helper functions
  */

drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c

@@ -2417,7 +2417,126 @@ vmw_execbuf_copy_fence_user(struct vmw_private *dev_priv,
 	}
 }
 
+/**
+ * vmw_execbuf_submit_fifo - Patch a command batch and submit it using
+ * the fifo.
+ *
+ * @dev_priv: Pointer to a device private structure.
+ * @kernel_commands: Pointer to the unpatched command batch.
+ * @command_size: Size of the unpatched command batch.
+ * @sw_context: Structure holding the relocation lists.
+ *
+ * Side effects: If this function returns 0, then the command batch
+ * pointed to by @kernel_commands will have been modified.
+ */
+static int vmw_execbuf_submit_fifo(struct vmw_private *dev_priv,
+				   void *kernel_commands,
+				   u32 command_size,
+				   struct vmw_sw_context *sw_context)
+{
+	void *cmd = vmw_fifo_reserve(dev_priv, command_size);
+
+	if (!cmd) {
+		DRM_ERROR("Failed reserving fifo space for commands.\n");
+		return -ENOMEM;
+	}
+
+	vmw_apply_relocations(sw_context);
+	memcpy(cmd, kernel_commands, command_size);
+	vmw_resource_relocations_apply(cmd, &sw_context->res_relocations);
+	vmw_resource_relocations_free(&sw_context->res_relocations);
+	vmw_fifo_commit(dev_priv, command_size);
+
+	return 0;
+}
 
+/**
+ * vmw_execbuf_submit_cmdbuf - Patch a command batch and submit it using
+ * the command buffer manager.
+ *
+ * @dev_priv: Pointer to a device private structure.
+ * @header: Opaque handle to the command buffer allocation.
+ * @command_size: Size of the unpatched command batch.
+ * @sw_context: Structure holding the relocation lists.
+ *
+ * Side effects: If this function returns 0, then the command buffer
+ * represented by @header will have been modified.
+ */
+static int vmw_execbuf_submit_cmdbuf(struct vmw_private *dev_priv,
+				     struct vmw_cmdbuf_header *header,
+				     u32 command_size,
+				     struct vmw_sw_context *sw_context)
+{
+	void *cmd = vmw_cmdbuf_reserve(dev_priv->cman, command_size,
+				       SVGA3D_INVALID_ID, false, header);
+
+	vmw_apply_relocations(sw_context);
+	vmw_resource_relocations_apply(cmd, &sw_context->res_relocations);
+	vmw_resource_relocations_free(&sw_context->res_relocations);
+	vmw_cmdbuf_commit(dev_priv->cman, command_size, header, false);
+
+	return 0;
+}
+
+/**
+ * vmw_execbuf_cmdbuf - Prepare, if possible, a user-space command batch for
+ * submission using a command buffer.
+ *
+ * @dev_priv: Pointer to a device private structure.
+ * @user_commands: User-space pointer to the commands to be submitted.
+ * @command_size: Size of the unpatched command batch.
+ * @header: Out parameter returning the opaque pointer to the command buffer.
+ *
+ * This function checks whether we can use the command buffer manager for
+ * submission and if so, creates a command buffer of suitable size and
+ * copies the user data into that buffer.
+ *
+ * On successful return, the function returns a pointer to the data in the
+ * command buffer and *@header is set to non-NULL.
+ * If command buffers could not be used, the function will return the value
+ * of @kernel_commands on function call. That value may be NULL. In that case,
+ * the value of *@header will be set to NULL.
+ * If an error is encountered, the function will return a pointer error value.
+ * If the function is interrupted by a signal while sleeping, it will return
+ * -ERESTARTSYS casted to a pointer error value.
+ */
+void *vmw_execbuf_cmdbuf(struct vmw_private *dev_priv,
+			 void __user *user_commands,
+			 void *kernel_commands,
+			 u32 command_size,
+			 struct vmw_cmdbuf_header **header)
+{
+	size_t cmdbuf_size;
+	int ret;
+
+	*header = NULL;
+	if (!dev_priv->cman || kernel_commands)
+		return kernel_commands;
+
+	if (command_size > SVGA_CB_MAX_SIZE) {
+		DRM_ERROR("Command buffer is too large.\n");
+		return ERR_PTR(-EINVAL);
+	}
+
+	/* If possible, add a little space for fencing. */
+	cmdbuf_size = command_size + 512;
+	cmdbuf_size = min_t(size_t, cmdbuf_size, SVGA_CB_MAX_SIZE);
+	kernel_commands = vmw_cmdbuf_alloc(dev_priv->cman, cmdbuf_size,
+					   true, header);
+	if (IS_ERR(kernel_commands))
+		return kernel_commands;
+
+	ret = copy_from_user(kernel_commands, user_commands,
+			     command_size);
+	if (ret) {
+		DRM_ERROR("Failed copying commands.\n");
+		vmw_cmdbuf_header_free(*header);
+		*header = NULL;
+		return ERR_PTR(-EFAULT);
+	}
+
+	return kernel_commands;
+}
 
 int vmw_execbuf_process(struct drm_file *file_priv,
 			struct vmw_private *dev_priv,
@@ -2432,18 +2551,33 @@ int vmw_execbuf_process(struct drm_file *file_priv,
 	struct vmw_fence_obj *fence = NULL;
 	struct vmw_resource *error_resource;
 	struct list_head resource_list;
+	struct vmw_cmdbuf_header *header;
 	struct ww_acquire_ctx ticket;
 	uint32_t handle;
-	void *cmd;
 	int ret;
 
+     	if (throttle_us) {
+		ret = vmw_wait_lag(dev_priv, &dev_priv->fifo.marker_queue,
+				   throttle_us);
+		
+		if (ret)
+			return ret;
+	}
+	
+	kernel_commands = vmw_execbuf_cmdbuf(dev_priv, user_commands,
+					     kernel_commands, command_size,
+					     &header);
+	if (IS_ERR(kernel_commands))
+		return PTR_ERR(kernel_commands);
+
 	ret = mutex_lock_interruptible(&dev_priv->cmdbuf_mutex);
-	if (unlikely(ret != 0))
-		return -ERESTARTSYS;
+	if (ret) {
+		ret = -ERESTARTSYS;
+		goto out_free_header;
+	}
 
+	sw_context->kernel = false;
 	if (kernel_commands == NULL) {
-		sw_context->kernel = false;
-
 		ret = vmw_resize_cmd_bounce(sw_context, command_size);
 		if (unlikely(ret != 0))
 			goto out_unlock;
@@ -2458,7 +2592,7 @@ int vmw_execbuf_process(struct drm_file *file_priv,
 			goto out_unlock;
 		}
 		kernel_commands = sw_context->cmd_bounce;
-	} else
+	} else if (!header)
 		sw_context->kernel = true;
 
 	sw_context->fp = vmw_fpriv(file_priv);
@@ -2478,7 +2612,6 @@ int vmw_execbuf_process(struct drm_file *file_priv,
 		sw_context->res_ht_initialized = true;
 	}
 	INIT_LIST_HEAD(&sw_context->staged_cmd_res);
-
 	INIT_LIST_HEAD(&resource_list);
 	ret = vmw_cmd_check_all(dev_priv, sw_context, kernel_commands,
 				command_size);
@@ -2502,14 +2635,6 @@ int vmw_execbuf_process(struct drm_file *file_priv,
 	if (unlikely(ret != 0))
 		goto out_err;
 
-	if (throttle_us) {
-		ret = vmw_wait_lag(dev_priv, &dev_priv->fifo.marker_queue,
-				   throttle_us);
-
-		if (unlikely(ret != 0))
-			goto out_err;
-	}
-
 	ret = mutex_lock_interruptible(&dev_priv->binding_mutex);
 	if (unlikely(ret != 0)) {
 		ret = -ERESTARTSYS;
@@ -2522,20 +2647,16 @@ int vmw_execbuf_process(struct drm_file *file_priv,
 			goto out_unlock_binding;
 	}
 
-	cmd = vmw_fifo_reserve(dev_priv, command_size);
-	if (unlikely(cmd == NULL)) {
-		DRM_ERROR("Failed reserving fifo space for commands.\n");
-		ret = -ENOMEM;
-		goto out_unlock_binding;
+	if (!header) {
+		ret = vmw_execbuf_submit_fifo(dev_priv, kernel_commands,
+					      command_size, sw_context);
+	} else {
+		ret = vmw_execbuf_submit_cmdbuf(dev_priv, header, command_size,
+						sw_context);
+		header = NULL;
 	}
-
-	vmw_apply_relocations(sw_context);
-	memcpy(cmd, kernel_commands, command_size);
-
-	vmw_resource_relocations_apply(cmd, &sw_context->res_relocations);
-	vmw_resource_relocations_free(&sw_context->res_relocations);
-
-	vmw_fifo_commit(dev_priv, command_size);
+	if (ret)
+		goto out_unlock_binding;
 
 	vmw_query_bo_switch_commit(dev_priv, sw_context);
 	ret = vmw_execbuf_fence_commands(file_priv, dev_priv,
@@ -2610,6 +2731,9 @@ int vmw_execbuf_process(struct drm_file *file_priv,
 	vmw_resource_list_unreference(&resource_list);
 	if (unlikely(error_resource != NULL))
 		vmw_resource_unreference(&error_resource);
+out_free_header:
+	if (header)
+		vmw_cmdbuf_header_free(header);
 
 	return ret;
 }

drivers/gpu/drm/vmwgfx/vmwgfx_fb.c

@@ -257,6 +257,7 @@ static void vmw_fb_dirty_flush(struct vmw_fb_par *par)
 	cmd->body.width = cpu_to_le32(w);
 	cmd->body.height = cpu_to_le32(h);
 	vmw_fifo_commit(vmw_priv, sizeof(*cmd));
+	vmw_fifo_flush(vmw_priv, false);
 }
 
 static void vmw_fb_dirty_mark(struct vmw_fb_par *par,

drivers/gpu/drm/vmwgfx/vmwgfx_fifo.c

@@ -310,7 +310,8 @@ static int vmw_fifo_wait(struct vmw_private *dev_priv,
  * Returns:
  *   Pointer to the fifo, or null on error (possible hardware hang).
  */
-void *vmw_fifo_reserve(struct vmw_private *dev_priv, uint32_t bytes)
+static void *vmw_local_fifo_reserve(struct vmw_private *dev_priv,
+				    uint32_t bytes)
 {
 	struct vmw_fifo_state *fifo_state = &dev_priv->fifo;
 	__le32 __iomem *fifo_mem = dev_priv->mmio_virt;
@@ -389,9 +390,29 @@ void *vmw_fifo_reserve(struct vmw_private *dev_priv, uint32_t bytes)
 out_err:
 	fifo_state->reserved_size = 0;
 	mutex_unlock(&fifo_state->fifo_mutex);
+
 	return NULL;
 }
 
+void *vmw_fifo_reserve(struct vmw_private *dev_priv, uint32_t bytes)
+{
+	void *ret;
+
+	if (dev_priv->cman)
+		ret = vmw_cmdbuf_reserve(dev_priv->cman, bytes,
+					 SVGA3D_INVALID_ID, false, NULL);
+	else
+		ret = vmw_local_fifo_reserve(dev_priv, bytes);
+	if (IS_ERR_OR_NULL(ret)) {
+		DRM_ERROR("Fifo reserve failure of %u bytes.\n",
+			  (unsigned) bytes);
+		dump_stack();
+		return NULL;
+	}
+
+	return ret;
+}
+
 static void vmw_fifo_res_copy(struct vmw_fifo_state *fifo_state,
 			      __le32 __iomem *fifo_mem,
 			      uint32_t next_cmd,
@@ -434,7 +455,7 @@ static void vmw_fifo_slow_copy(struct vmw_fifo_state *fifo_state,
 	}
 }
 
-void vmw_fifo_commit(struct vmw_private *dev_priv, uint32_t bytes)
+void vmw_local_fifo_commit(struct vmw_private *dev_priv, uint32_t bytes)
 {
 	struct vmw_fifo_state *fifo_state = &dev_priv->fifo;
 	__le32 __iomem *fifo_mem = dev_priv->mmio_virt;
@@ -480,6 +501,46 @@ void vmw_fifo_commit(struct vmw_private *dev_priv, uint32_t bytes)
 	mutex_unlock(&fifo_state->fifo_mutex);
 }
 
+void vmw_fifo_commit(struct vmw_private *dev_priv, uint32_t bytes)
+{
+	if (dev_priv->cman)
+		vmw_cmdbuf_commit(dev_priv->cman, bytes, NULL, false);
+	else
+		vmw_local_fifo_commit(dev_priv, bytes);
+}
+
+
+/**
+ * vmw_fifo_commit_flush - Commit fifo space and flush any buffered commands.
+ *
+ * @dev_priv: Pointer to device private structure.
+ * @bytes: Number of bytes to commit.
+ */
+static void vmw_fifo_commit_flush(struct vmw_private *dev_priv, uint32_t bytes)
+{
+	if (dev_priv->cman)
+		vmw_cmdbuf_commit(dev_priv->cman, bytes, NULL, true);
+	else
+		vmw_local_fifo_commit(dev_priv, bytes);
+}
+
+/**
+ * vmw_fifo_flush - Flush any buffered commands and make sure command processing
+ * starts.
+ *
+ * @dev_priv: Pointer to device private structure.
+ * @interruptible: Whether to wait interruptible if function needs to sleep.
+ */
+int vmw_fifo_flush(struct vmw_private *dev_priv, bool interruptible)
+{
+	might_sleep();
+
+	if (dev_priv->cman)
+		return vmw_cmdbuf_cur_flush(dev_priv->cman, interruptible);
+	else
+		return 0;
+}
+
 int vmw_fifo_send_fence(struct vmw_private *dev_priv, uint32_t *seqno)
 {
 	struct vmw_fifo_state *fifo_state = &dev_priv->fifo;
@@ -517,7 +578,7 @@ int vmw_fifo_send_fence(struct vmw_private *dev_priv, uint32_t *seqno)
 	    ((unsigned long)fm + sizeof(__le32));
 
 	iowrite32(*seqno, &cmd_fence->fence);
-	vmw_fifo_commit(dev_priv, bytes);
+	vmw_fifo_commit_flush(dev_priv, bytes);
 	(void) vmw_marker_push(&fifo_state->marker_queue, *seqno);
 	vmw_update_seqno(dev_priv, fifo_state);
 

drivers/gpu/drm/vmwgfx/vmwgfx_irq.c

@@ -56,6 +56,9 @@ irqreturn_t vmw_irq_handler(int irq, void *arg)
 	if (masked_status & SVGA_IRQFLAG_FIFO_PROGRESS)
 		wake_up_all(&dev_priv->fifo_queue);
 
+	if (masked_status & (SVGA_IRQFLAG_COMMAND_BUFFER |
+			     SVGA_IRQFLAG_ERROR))
+		vmw_cmdbuf_tasklet_schedule(dev_priv->cman);
 
 	return IRQ_HANDLED;
 }
@@ -131,8 +134,16 @@ int vmw_fallback_wait(struct vmw_private *dev_priv,
 	 * Block command submission while waiting for idle.
 	 */
 
-	if (fifo_idle)
+	if (fifo_idle) {
 		down_read(&fifo_state->rwsem);
+		if (dev_priv->cman) {
+			ret = vmw_cmdbuf_idle(dev_priv->cman, interruptible,
+					      10*HZ);
+			if (ret)
+				goto out_err;
+		}
+	}
+
 	signal_seq = atomic_read(&dev_priv->marker_seq);
 	ret = 0;
 
@@ -171,6 +182,7 @@ int vmw_fallback_wait(struct vmw_private *dev_priv,
 		iowrite32(signal_seq, fifo_mem + SVGA_FIFO_FENCE);
 	}
 	wake_up_all(&dev_priv->fence_queue);
+out_err:
 	if (fifo_idle)
 		up_read(&fifo_state->rwsem);
 
@@ -315,3 +327,30 @@ void vmw_irq_uninstall(struct drm_device *dev)
 	status = inl(dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
 	outl(status, dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
 }
+
+void vmw_generic_waiter_add(struct vmw_private *dev_priv,
+			    u32 flag, int *waiter_count)
+{
+	unsigned long irq_flags;
+
+	spin_lock_irqsave(&dev_priv->irq_lock, irq_flags);
+	if ((*waiter_count)++ == 0) {
+		outl(flag, dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
+		dev_priv->irq_mask |= flag;
+		vmw_write(dev_priv, SVGA_REG_IRQMASK, dev_priv->irq_mask);
+	}
+	spin_unlock_irqrestore(&dev_priv->irq_lock, irq_flags);
+}
+
+void vmw_generic_waiter_remove(struct vmw_private *dev_priv,
+			       u32 flag, int *waiter_count)
+{
+	unsigned long irq_flags;
+
+	spin_lock_irqsave(&dev_priv->irq_lock, irq_flags);
+	if (--(*waiter_count) == 0) {
+		dev_priv->irq_mask &= ~flag;
+		vmw_write(dev_priv, SVGA_REG_IRQMASK, dev_priv->irq_mask);
+	}
+	spin_unlock_irqrestore(&dev_priv->irq_lock, irq_flags);
+}

drivers/gpu/drm/vmwgfx/vmwgfx_kms.c

@@ -631,6 +631,7 @@ static int vmw_framebuffer_surface_dirty(struct drm_framebuffer *framebuffer,
 				   flags, color,
 				   clips, num_clips, inc, NULL);
 
+	vmw_fifo_flush(dev_priv, false);
 	ttm_read_unlock(&dev_priv->reservation_sem);
 
 	drm_modeset_unlock_all(dev_priv->dev);
@@ -987,6 +988,7 @@ static int vmw_framebuffer_dmabuf_dirty(struct drm_framebuffer *framebuffer,
 					  clips, num_clips, increment, NULL);
 	}
 
+	vmw_fifo_flush(dev_priv, false);
 	ttm_read_unlock(&dev_priv->reservation_sem);
 
 	drm_modeset_unlock_all(dev_priv->dev);
@@ -1347,6 +1349,8 @@ int vmw_kms_present(struct vmw_private *dev_priv,
 			break;
 	}
 
+	vmw_fifo_flush(dev_priv, false);
+
 	kfree(cmd);
 out_free_tmp:
 	kfree(tmp);