vmwgfx: Fix up query processing

Previously, query results could be placed in any buffer object, but since
we didn't allow pinned buffer objects, query results could be written when
that buffer was evicted, corrupting data in other buffers.

Now, require that buffers holding query results are no more than two pages
large, and allow one single pinned such buffer. When the command submission
code encounters query result structures in other buffers, the queries in the
pinned buffer will be finished using a query barrier for the last hardware
context using the buffer. Also if the command submission code detects
that a new hardware context is used for queries, all queries of the previous
hardware context is also flushed. Currently we use waiting for a no-op
occlusion query as a query barrier for a specific context.

The query buffer is also flushed and unpinned on context destructions,
master drops and before scanout bo placement.

Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Jakob Bornecrantz <jakob@vmware.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

6 files changed, 572 insertions, 22 deletions

diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_dmabuf.c b/drivers/gpu/drm/vmwgfx/vmwgfx_dmabuf.c
index 7f744a82892a..3fa884db08ab 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_dmabuf.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_dmabuf.c
@@ -42,6 +42,7 @@
  * May only be called by the current master since it assumes that the
  * master lock is the current master's lock.
  * This function takes the master's lock in write mode.
+ * Flushes and unpins the query bo to avoid failures.
  *
  * Returns
  *  -ERESTARTSYS if interrupted by a signal.
@@ -59,6 +60,8 @@ int vmw_dmabuf_to_placement(struct vmw_private *dev_priv,
 	if (unlikely(ret != 0))
 		return ret;
 
+	vmw_execbuf_release_pinned_bo(dev_priv, false, 0);
+
 	ret = ttm_bo_reserve(bo, interruptible, false, false, 0);
 	if (unlikely(ret != 0))
 		goto err;
@@ -78,6 +81,7 @@ err:
  * May only be called by the current master since it assumes that the
  * master lock is the current master's lock.
  * This function takes the master's lock in write mode.
+ * Flushes and unpins the query bo if @pin == true to avoid failures.
  *
  * @dev_priv:  Driver private.
  * @buf:  DMA buffer to move.
@@ -100,6 +104,9 @@ int vmw_dmabuf_to_vram_or_gmr(struct vmw_private *dev_priv,
 	if (unlikely(ret != 0))
 		return ret;
 
+	if (pin)
+		vmw_execbuf_release_pinned_bo(dev_priv, false, 0);
+
 	ret = ttm_bo_reserve(bo, interruptible, false, false, 0);
 	if (unlikely(ret != 0))
 		goto err;
@@ -177,6 +184,7 @@ int vmw_dmabuf_to_vram(struct vmw_private *dev_priv,
  * May only be called by the current master since it assumes that the
  * master lock is the current master's lock.
  * This function takes the master's lock in write mode.
+ * Flushes and unpins the query bo if @pin == true to avoid failures.
  *
  * @dev_priv:  Driver private.
  * @buf:  DMA buffer to move.
@@ -205,6 +213,9 @@ int vmw_dmabuf_to_start_of_vram(struct vmw_private *dev_priv,
 	if (unlikely(ret != 0))
 		return ret;
 
+	if (pin)
+		vmw_execbuf_release_pinned_bo(dev_priv, false, 0);
+
 	ret = ttm_bo_reserve(bo, interruptible, false, false, 0);
 	if (unlikely(ret != 0))
 		goto err_unlock;
@@ -276,3 +287,36 @@ void vmw_bo_get_guest_ptr(const struct ttm_buffer_object *bo,
 		ptr->offset = 0;
 	}
 }
+
+
+/**
+ * vmw_bo_pin - Pin or unpin a buffer object without moving it.
+ *
+ * @bo: The buffer object. Must be reserved, and present either in VRAM
+ * or GMR memory.
+ * @pin: Whether to pin or unpin.
+ *
+ */
+void vmw_bo_pin(struct ttm_buffer_object *bo, bool pin)
+{
+	uint32_t pl_flags;
+	struct ttm_placement placement;
+	uint32_t old_mem_type = bo->mem.mem_type;
+	int ret;
+
+	BUG_ON(!atomic_read(&bo->reserved));
+	BUG_ON(old_mem_type != TTM_PL_VRAM &&
+	       old_mem_type != VMW_PL_FLAG_GMR);
+
+	pl_flags = TTM_PL_FLAG_VRAM | VMW_PL_FLAG_GMR | TTM_PL_FLAG_CACHED;
+	if (pin)
+		pl_flags |= TTM_PL_FLAG_NO_EVICT;
+
+	memset(&placement, 0, sizeof(placement));
+	placement.num_placement = 1;
+	placement.placement = &pl_flags;
+
+	ret = ttm_bo_validate(bo, &placement, false, true, true);
+
+	BUG_ON(ret != 0 || bo->mem.mem_type != old_mem_type);
+}
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_drv.c b/drivers/gpu/drm/vmwgfx/vmwgfx_drv.c
index ace4402214c6..7b88104144ca 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_drv.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_drv.c
@@ -213,6 +213,72 @@ static void vmw_print_capabilities(uint32_t capabilities)
 		DRM_INFO("  Screen Object 2.\n");
 }
 
+
+/**
+ * vmw_execbuf_prepare_dummy_query - Initialize a query result structure at
+ * the start of a buffer object.
+ *
+ * @dev_priv: The device private structure.
+ *
+ * This function will idle the buffer using an uninterruptible wait, then
+ * map the first page and initialize a pending occlusion query result structure,
+ * Finally it will unmap the buffer.
+ *
+ * TODO: Since we're only mapping a single page, we should optimize the map
+ * to use kmap_atomic / iomap_atomic.
+ */
+static void vmw_dummy_query_bo_prepare(struct vmw_private *dev_priv)
+{
+	struct ttm_bo_kmap_obj map;
+	volatile SVGA3dQueryResult *result;
+	bool dummy;
+	int ret;
+	struct ttm_bo_device *bdev = &dev_priv->bdev;
+	struct ttm_buffer_object *bo = dev_priv->dummy_query_bo;
+
+	ttm_bo_reserve(bo, false, false, false, 0);
+	spin_lock(&bdev->fence_lock);
+	ret = ttm_bo_wait(bo, false, false, false, TTM_USAGE_READWRITE);
+	spin_unlock(&bdev->fence_lock);
+	if (unlikely(ret != 0))
+		(void) vmw_fallback_wait(dev_priv, false, true, 0, false,
+					 10*HZ);
+
+	ret = ttm_bo_kmap(bo, 0, 1, &map);
+	if (likely(ret == 0)) {
+		result = ttm_kmap_obj_virtual(&map, &dummy);
+		result->totalSize = sizeof(*result);
+		result->state = SVGA3D_QUERYSTATE_PENDING;
+		result->result32 = 0xff;
+		ttm_bo_kunmap(&map);
+	} else
+		DRM_ERROR("Dummy query buffer map failed.\n");
+	ttm_bo_unreserve(bo);
+}
+
+
+/**
+ * vmw_dummy_query_bo_create - create a bo to hold a dummy query result
+ *
+ * @dev_priv: A device private structure.
+ *
+ * This function creates a small buffer object that holds the query
+ * result for dummy queries emitted as query barriers.
+ * No interruptible waits are done within this function.
+ *
+ * Returns an error if bo creation fails.
+ */
+static int vmw_dummy_query_bo_create(struct vmw_private *dev_priv)
+{
+	return ttm_bo_create(&dev_priv->bdev,
+			     PAGE_SIZE,
+			     ttm_bo_type_device,
+			     &vmw_vram_sys_placement,
+			     0, 0, false, NULL,
+			     &dev_priv->dummy_query_bo);
+}
+
+
 static int vmw_request_device(struct vmw_private *dev_priv)
 {
 	int ret;
@@ -223,12 +289,29 @@ static int vmw_request_device(struct vmw_private *dev_priv)
 		return ret;
 	}
 	vmw_fence_fifo_up(dev_priv->fman);
+	ret = vmw_dummy_query_bo_create(dev_priv);
+	if (unlikely(ret != 0))
+		goto out_no_query_bo;
+	vmw_dummy_query_bo_prepare(dev_priv);
 
 	return 0;
+
+out_no_query_bo:
+	vmw_fence_fifo_down(dev_priv->fman);
+	vmw_fifo_release(dev_priv, &dev_priv->fifo);
+	return ret;
 }
 
 static void vmw_release_device(struct vmw_private *dev_priv)
 {
+	/*
+	 * Previous destructions should've released
+	 * the pinned bo.
+	 */
+
+	BUG_ON(dev_priv->pinned_bo != NULL);
+
+	ttm_bo_unref(&dev_priv->dummy_query_bo);
 	vmw_fence_fifo_down(dev_priv->fman);
 	vmw_fifo_release(dev_priv, &dev_priv->fifo);
 }
@@ -794,6 +877,8 @@ static void vmw_master_drop(struct drm_device *dev,
 
 	vmw_fp->locked_master = drm_master_get(file_priv->master);
 	ret = ttm_vt_lock(&vmaster->lock, false, vmw_fp->tfile);
+	vmw_execbuf_release_pinned_bo(dev_priv, false, 0);
+
 	if (unlikely((ret != 0))) {
 		DRM_ERROR("Unable to lock TTM at VT switch.\n");
 		drm_master_put(&vmw_fp->locked_master);
@@ -844,6 +929,7 @@ static int vmwgfx_pm_notifier(struct notifier_block *nb, unsigned long val,
 		 * This empties VRAM and unbinds all GMR bindings.
 		 * Buffer contents is moved to swappable memory.
 		 */
+		vmw_execbuf_release_pinned_bo(dev_priv, false, 0);
 		ttm_bo_swapout_all(&dev_priv->bdev);
 
 		break;
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_drv.h b/drivers/gpu/drm/vmwgfx/vmwgfx_drv.h
index b22b96a491a4..d8d6a8659119 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_drv.h
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_drv.h
@@ -82,6 +82,7 @@ struct vmw_resource {
 	void (*hw_destroy) (struct vmw_resource *res);
 	void (*res_free) (struct vmw_resource *res);
 	bool on_validate_list;
+	struct list_head query_head; /* Protected by the cmdbuf mutex */
 	/* TODO is a generic snooper needed? */
 #if 0
 	void (*snoop)(struct vmw_resource *res,
@@ -142,6 +143,7 @@ struct vmw_sw_context{
 	uint32_t last_cid;
 	bool cid_valid;
 	bool kernel; /**< is the called made from the kernel */
+	struct vmw_resource *cur_ctx;
 	uint32_t last_sid;
 	uint32_t sid_translation;
 	bool sid_valid;
@@ -155,6 +157,11 @@ struct vmw_sw_context{
 	uint32_t cmd_bounce_size;
 	struct vmw_resource *resources[VMWGFX_MAX_VALIDATIONS];
 	uint32_t num_ref_resources;
+	uint32_t fence_flags;
+	struct list_head query_list;
+	struct ttm_buffer_object *cur_query_bo;
+	uint32_t cur_query_cid;
+	bool query_cid_valid;
 };
 
 struct vmw_legacy_display;
@@ -294,6 +301,16 @@ struct vmw_private {
 
 	struct mutex release_mutex;
 	uint32_t num_3d_resources;
+
+	/*
+	 * Query processing. These members
+	 * are protected by the cmdbuf mutex.
+	 */
+
+	struct ttm_buffer_object *dummy_query_bo;
+	struct ttm_buffer_object *pinned_bo;
+	uint32_t query_cid;
+	bool dummy_query_bo_pinned;
 };
 
 static inline struct vmw_private *vmw_priv(struct drm_device *dev)
@@ -418,6 +435,7 @@ extern int vmw_dmabuf_unpin(struct vmw_private *vmw_priv,
 			    bool interruptible);
 extern void vmw_bo_get_guest_ptr(const struct ttm_buffer_object *buf,
 				 SVGAGuestPtr *ptr);
+extern void vmw_bo_pin(struct ttm_buffer_object *bo, bool pin);
 
 /**
  * Misc Ioctl functionality - vmwgfx_ioctl.c
@@ -447,6 +465,8 @@ extern int vmw_fifo_send_fence(struct vmw_private *dev_priv,
 extern void vmw_fifo_ping_host(struct vmw_private *dev_priv, uint32_t reason);
 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);
 
 /**
  * TTM glue - vmwgfx_ttm_glue.c
@@ -485,6 +505,10 @@ extern int vmw_execbuf_process(struct drm_file *file_priv,
 			       struct drm_vmw_fence_rep __user
 			       *user_fence_rep);
 
+extern void
+vmw_execbuf_release_pinned_bo(struct vmw_private *dev_priv,
+			      bool only_on_cid_match, uint32_t cid);
+
 /**
  * IRQs and wating - vmwgfx_irq.c
  */
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c b/drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c
index dea0474f6f3b..efa1d1cc0414 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c
@@ -44,7 +44,6 @@ static int vmw_cmd_ok(struct vmw_private *dev_priv,
 	return 0;
 }
 
-
 static int vmw_resource_to_validate_list(struct vmw_sw_context *sw_context,
 					 struct vmw_resource **p_res)
 {
@@ -68,6 +67,54 @@ out:
 	return ret;
 }
 
+/**
+ * vmw_bo_to_validate_list - add a bo to a validate list
+ *
+ * @sw_context: The software context used for this command submission batch.
+ * @bo: The buffer object to add.
+ * @fence_flags: Fence flags to be or'ed with any other fence flags for
+ * this buffer on this submission batch.
+ * @p_val_node: If non-NULL Will be updated with the validate node number
+ * on return.
+ *
+ * Returns -EINVAL if the limit of number of buffer objects per command
+ * submission is reached.
+ */
+static int vmw_bo_to_validate_list(struct vmw_sw_context *sw_context,
+				   struct ttm_buffer_object *bo,
+				   uint32_t fence_flags,
+				   uint32_t *p_val_node)
+{
+	uint32_t val_node;
+	struct ttm_validate_buffer *val_buf;
+
+	val_node = vmw_dmabuf_validate_node(bo, sw_context->cur_val_buf);
+
+	if (unlikely(val_node >= VMWGFX_MAX_VALIDATIONS)) {
+		DRM_ERROR("Max number of DMA buffers per submission"
+			  " exceeded.\n");
+		return -EINVAL;
+	}
+
+	val_buf = &sw_context->val_bufs[val_node];
+	if (unlikely(val_node == sw_context->cur_val_buf)) {
+		val_buf->new_sync_obj_arg = NULL;
+		val_buf->bo = ttm_bo_reference(bo);
+		val_buf->usage = TTM_USAGE_READWRITE;
+		list_add_tail(&val_buf->head, &sw_context->validate_nodes);
+		++sw_context->cur_val_buf;
+	}
+
+	val_buf->new_sync_obj_arg = (void *)
+		((unsigned long) val_buf->new_sync_obj_arg | fence_flags);
+	sw_context->fence_flags |= fence_flags;
+
+	if (p_val_node)
+		*p_val_node = val_node;
+
+	return 0;
+}
+
 static int vmw_cmd_cid_check(struct vmw_private *dev_priv,
 			     struct vmw_sw_context *sw_context,
 			     SVGA3dCmdHeader *header)
@@ -94,6 +141,7 @@ static int vmw_cmd_cid_check(struct vmw_private *dev_priv,
 
 	sw_context->last_cid = cmd->cid;
 	sw_context->cid_valid = true;
+	sw_context->cur_ctx = ctx;
 	return vmw_resource_to_validate_list(sw_context, &ctx);
 }
 
@@ -114,7 +162,8 @@ static int vmw_cmd_sid_check(struct vmw_private *dev_priv,
 		return 0;
 	}
 
-	ret = vmw_user_surface_lookup_handle(dev_priv, sw_context->tfile,
+	ret = vmw_user_surface_lookup_handle(dev_priv,
+					     sw_context->tfile,
 					     *sid, &srf);
 	if (unlikely(ret != 0)) {
 		DRM_ERROR("Could ot find or use surface 0x%08x "
@@ -225,6 +274,168 @@ static int vmw_cmd_present_check(struct vmw_private *dev_priv,
 	return vmw_cmd_sid_check(dev_priv, sw_context, &cmd->body.sid);
 }
 
+/**
+ * vmw_query_bo_switch_prepare - Prepare to switch pinned buffer for queries.
+ *
+ * @dev_priv: The device private structure.
+ * @cid: The hardware context for the next query.
+ * @new_query_bo: The new buffer holding query results.
+ * @sw_context: The software context used for this command submission.
+ *
+ * This function checks whether @new_query_bo is suitable for holding
+ * query results, and if another buffer currently is pinned for query
+ * results. If so, the function prepares the state of @sw_context for
+ * switching pinned buffers after successful submission of the current
+ * command batch. It also checks whether we're using a new query context.
+ * In that case, it makes sure we emit a query barrier for the old
+ * context before the current query buffer is fenced.
+ */
+static int vmw_query_bo_switch_prepare(struct vmw_private *dev_priv,
+				       uint32_t cid,
+				       struct ttm_buffer_object *new_query_bo,
+				       struct vmw_sw_context *sw_context)
+{
+	int ret;
+	bool add_cid = false;
+	uint32_t cid_to_add;
+
+	if (unlikely(new_query_bo != sw_context->cur_query_bo)) {
+
+		if (unlikely(new_query_bo->num_pages > 4)) {
+			DRM_ERROR("Query buffer too large.\n");
+			return -EINVAL;
+		}
+
+		if (unlikely(sw_context->cur_query_bo != NULL)) {
+			BUG_ON(!sw_context->query_cid_valid);
+			add_cid = true;
+			cid_to_add = sw_context->cur_query_cid;
+			ret = vmw_bo_to_validate_list(sw_context,
+						      sw_context->cur_query_bo,
+						      DRM_VMW_FENCE_FLAG_EXEC,
+						      NULL);
+			if (unlikely(ret != 0))
+				return ret;
+		}
+		sw_context->cur_query_bo = new_query_bo;
+
+		ret = vmw_bo_to_validate_list(sw_context,
+					      dev_priv->dummy_query_bo,
+					      DRM_VMW_FENCE_FLAG_EXEC,
+					      NULL);
+		if (unlikely(ret != 0))
+			return ret;
+
+	}
+
+	if (unlikely(cid != sw_context->cur_query_cid &&
+		     sw_context->query_cid_valid)) {
+		add_cid = true;
+		cid_to_add = sw_context->cur_query_cid;
+	}
+
+	sw_context->cur_query_cid = cid;
+	sw_context->query_cid_valid = true;
+
+	if (add_cid) {
+		struct vmw_resource *ctx = sw_context->cur_ctx;
+
+		if (list_empty(&ctx->query_head))
+			list_add_tail(&ctx->query_head,
+				      &sw_context->query_list);
+		ret = vmw_bo_to_validate_list(sw_context,
+					      dev_priv->dummy_query_bo,
+					      DRM_VMW_FENCE_FLAG_EXEC,
+					      NULL);
+		if (unlikely(ret != 0))
+			return ret;
+	}
+	return 0;
+}
+
+
+/**
+ * vmw_query_bo_switch_commit - Finalize switching pinned query buffer
+ *
+ * @dev_priv: The device private structure.
+ * @sw_context: The software context used for this command submission batch.
+ *
+ * This function will check if we're switching query buffers, and will then,
+ * if no other query waits are issued this command submission batch,
+ * issue a dummy occlusion query wait used as a query barrier. When the fence
+ * object following that query wait has signaled, we are sure that all
+ * preseding queries have finished, and the old query buffer can be unpinned.
+ * However, since both the new query buffer and the old one are fenced with
+ * that fence, we can do an asynchronus unpin now, and be sure that the
+ * old query buffer won't be moved until the fence has signaled.
+ *
+ * As mentioned above, both the new - and old query buffers need to be fenced
+ * using a sequence emitted *after* calling this function.
+ */
+static void vmw_query_bo_switch_commit(struct vmw_private *dev_priv,
+				     struct vmw_sw_context *sw_context)
+{
+
+	struct vmw_resource *ctx, *next_ctx;
+	int ret;
+
+	/*
+	 * The validate list should still hold references to all
+	 * contexts here.
+	 */
+
+	list_for_each_entry_safe(ctx, next_ctx, &sw_context->query_list,
+				 query_head) {
+		list_del_init(&ctx->query_head);
+
+		BUG_ON(!ctx->on_validate_list);
+
+		ret = vmw_fifo_emit_dummy_query(dev_priv, ctx->id);
+
+		if (unlikely(ret != 0))
+			DRM_ERROR("Out of fifo space for dummy query.\n");
+	}
+
+	if (dev_priv->pinned_bo != sw_context->cur_query_bo) {
+		if (dev_priv->pinned_bo) {
+			vmw_bo_pin(dev_priv->pinned_bo, false);
+			ttm_bo_unref(&dev_priv->pinned_bo);
+		}
+
+		vmw_bo_pin(sw_context->cur_query_bo, true);
+
+		/*
+		 * We pin also the dummy_query_bo buffer so that we
+		 * don't need to validate it when emitting
+		 * dummy queries in context destroy paths.
+		 */
+
+		vmw_bo_pin(dev_priv->dummy_query_bo, true);
+		dev_priv->dummy_query_bo_pinned = true;
+
+		dev_priv->query_cid = sw_context->cur_query_cid;
+		dev_priv->pinned_bo =
+			ttm_bo_reference(sw_context->cur_query_bo);
+	}
+}
+
+/**
+ * vmw_query_switch_backoff - clear query barrier list
+ * @sw_context: The sw context used for this submission batch.
+ *
+ * This function is used as part of an error path, where a previously
+ * set up list of query barriers needs to be cleared.
+ *
+ */
+static void vmw_query_switch_backoff(struct vmw_sw_context *sw_context)
+{
+	struct list_head *list, *next;
+
+	list_for_each_safe(list, next, &sw_context->query_list) {
+		list_del_init(list);
+	}
+}
+
 static int vmw_translate_guest_ptr(struct vmw_private *dev_priv,
 				   struct vmw_sw_context *sw_context,
 				   SVGAGuestPtr *ptr,
@@ -234,8 +445,6 @@ static int vmw_translate_guest_ptr(struct vmw_private *dev_priv,
 	struct ttm_buffer_object *bo;
 	uint32_t handle = ptr->gmrId;
 	struct vmw_relocation *reloc;
-	uint32_t cur_validate_node;
-	struct ttm_validate_buffer *val_buf;
 	int ret;
 
 	ret = vmw_user_dmabuf_lookup(sw_context->tfile, handle, &vmw_bo);
@@ -255,23 +464,11 @@ static int vmw_translate_guest_ptr(struct vmw_private *dev_priv,
 	reloc = &sw_context->relocs[sw_context->cur_reloc++];
 	reloc->location = ptr;
 
-	cur_validate_node = vmw_dmabuf_validate_node(bo, sw_context->cur_val_buf);
-	if (unlikely(cur_validate_node >= VMWGFX_MAX_VALIDATIONS)) {
-		DRM_ERROR("Max number of DMA buffers per submission"
-			  " exceeded.\n");
-		ret = -EINVAL;
+	ret = vmw_bo_to_validate_list(sw_context, bo, DRM_VMW_FENCE_FLAG_EXEC,
+				      &reloc->index);
+	if (unlikely(ret != 0))
 		goto out_no_reloc;
-	}
 
-	reloc->index = cur_validate_node;
-	if (unlikely(cur_validate_node == sw_context->cur_val_buf)) {
-		val_buf = &sw_context->val_bufs[cur_validate_node];
-		val_buf->bo = ttm_bo_reference(bo);
-		val_buf->usage = TTM_USAGE_READWRITE;
-		val_buf->new_sync_obj_arg = (void *) DRM_VMW_FENCE_FLAG_EXEC;
-		list_add_tail(&val_buf->head, &sw_context->validate_nodes);
-		++sw_context->cur_val_buf;
-	}
 	*vmw_bo_p = vmw_bo;
 	return 0;
 
@@ -303,8 +500,11 @@ static int vmw_cmd_end_query(struct vmw_private *dev_priv,
 	if (unlikely(ret != 0))
 		return ret;
 
+	ret = vmw_query_bo_switch_prepare(dev_priv, cmd->q.cid,
+					  &vmw_bo->base, sw_context);
+
 	vmw_dmabuf_unreference(&vmw_bo);
-	return 0;
+	return ret;
 }
 
 static int vmw_cmd_wait_query(struct vmw_private *dev_priv,
@@ -317,6 +517,7 @@ static int vmw_cmd_wait_query(struct vmw_private *dev_priv,
 		SVGA3dCmdWaitForQuery q;
 	} *cmd;
 	int ret;
+	struct vmw_resource *ctx;
 
 	cmd = container_of(header, struct vmw_query_cmd, header);
 	ret = vmw_cmd_cid_check(dev_priv, sw_context, header);
@@ -330,6 +531,16 @@ static int vmw_cmd_wait_query(struct vmw_private *dev_priv,
 		return ret;
 
 	vmw_dmabuf_unreference(&vmw_bo);
+
+	/*
+	 * This wait will act as a barrier for previous waits for this
+	 * context.
+	 */
+
+	ctx = sw_context->cur_ctx;
+	if (!list_empty(&ctx->query_head))
+		list_del_init(&ctx->query_head);
+
 	return 0;
 }
 
@@ -687,6 +898,16 @@ static int vmw_validate_single_buffer(struct vmw_private *dev_priv,
 {
 	int ret;
 
+
+	/*
+	 * Don't validate pinned buffers.
+	 */
+
+	if (bo == dev_priv->pinned_bo ||
+	    (bo == dev_priv->dummy_query_bo &&
+	     dev_priv->dummy_query_bo_pinned))
+		return 0;
+
 	/**
 	 * Put BO in VRAM if there is space, otherwise as a GMR.
 	 * If there is no space in VRAM and GMR ids are all used up,
@@ -846,6 +1067,11 @@ int vmw_execbuf_process(struct drm_file *file_priv,
 	sw_context->cur_reloc = 0;
 	sw_context->cur_val_buf = 0;
 	sw_context->num_ref_resources = 0;
+	sw_context->fence_flags = 0;
+	INIT_LIST_HEAD(&sw_context->query_list);
+	sw_context->cur_query_bo = dev_priv->pinned_bo;
+	sw_context->cur_query_cid = dev_priv->query_cid;
+	sw_context->query_cid_valid = (dev_priv->pinned_bo != NULL);
 
 	INIT_LIST_HEAD(&sw_context->validate_nodes);
 
@@ -882,6 +1108,7 @@ int vmw_execbuf_process(struct drm_file *file_priv,
 	memcpy(cmd, kernel_commands, command_size);
 	vmw_fifo_commit(dev_priv, command_size);
 
+	vmw_query_bo_switch_commit(dev_priv, sw_context);
 	ret = vmw_execbuf_fence_commands(file_priv, dev_priv,
 					 &fence,
 					 (user_fence_rep) ? &handle : NULL);
@@ -940,6 +1167,7 @@ int vmw_execbuf_process(struct drm_file *file_priv,
 out_err:
 	vmw_free_relocations(sw_context);
 out_throttle:
+	vmw_query_switch_backoff(sw_context);
 	ttm_eu_backoff_reservation(&sw_context->validate_nodes);
 	vmw_clear_validations(sw_context);
 out_unlock:
@@ -947,6 +1175,113 @@ out_unlock:
 	return ret;
 }
 
+/**
+ * vmw_execbuf_unpin_panic - Idle the fifo and unpin the query buffer.
+ *
+ * @dev_priv: The device private structure.
+ *
+ * This function is called to idle the fifo and unpin the query buffer
+ * if the normal way to do this hits an error, which should typically be
+ * extremely rare.
+ */
+static void vmw_execbuf_unpin_panic(struct vmw_private *dev_priv)
+{
+	DRM_ERROR("Can't unpin query buffer. Trying to recover.\n");
+
+	(void) vmw_fallback_wait(dev_priv, false, true, 0, false, 10*HZ);
+	vmw_bo_pin(dev_priv->pinned_bo, false);
+	vmw_bo_pin(dev_priv->dummy_query_bo, false);
+	dev_priv->dummy_query_bo_pinned = false;
+}
+
+
+/**
+ * vmw_execbuf_release_pinned_bo - Flush queries and unpin the pinned
+ * query bo.
+ *
+ * @dev_priv: The device private structure.
+ * @only_on_cid_match: Only flush and unpin if the current active query cid
+ * matches @cid.
+ * @cid: Optional context id to match.
+ *
+ * This function should be used to unpin the pinned query bo, or
+ * as a query barrier when we need to make sure that all queries have
+ * finished before the next fifo command. (For example on hardware
+ * context destructions where the hardware may otherwise leak unfinished
+ * queries).
+ *
+ * This function does not return any failure codes, but make attempts
+ * to do safe unpinning in case of errors.
+ *
+ * The function will synchronize on the previous query barrier, and will
+ * thus not finish until that barrier has executed.
+ */
+void vmw_execbuf_release_pinned_bo(struct vmw_private *dev_priv,
+				   bool only_on_cid_match, uint32_t cid)
+{
+	int ret = 0;
+	struct list_head validate_list;
+	struct ttm_validate_buffer pinned_val, query_val;
+	struct vmw_fence_obj *fence;
+
+	mutex_lock(&dev_priv->cmdbuf_mutex);
+
+	if (dev_priv->pinned_bo == NULL)
+		goto out_unlock;
+
+	if (only_on_cid_match && cid != dev_priv->query_cid)
+		goto out_unlock;
+
+	INIT_LIST_HEAD(&validate_list);
+
+	pinned_val.new_sync_obj_arg = (void *)(unsigned long)
+		DRM_VMW_FENCE_FLAG_EXEC;
+	pinned_val.bo = ttm_bo_reference(dev_priv->pinned_bo);
+	list_add_tail(&pinned_val.head, &validate_list);
+
+	query_val.new_sync_obj_arg = pinned_val.new_sync_obj_arg;
+	query_val.bo = ttm_bo_reference(dev_priv->dummy_query_bo);
+	list_add_tail(&query_val.head, &validate_list);
+
+	do {
+		ret = ttm_eu_reserve_buffers(&validate_list);
+	} while (ret == -ERESTARTSYS);
+
+	if (unlikely(ret != 0)) {
+		vmw_execbuf_unpin_panic(dev_priv);
+		goto out_no_reserve;
+	}
+
+	ret = vmw_fifo_emit_dummy_query(dev_priv, dev_priv->query_cid);
+	if (unlikely(ret != 0)) {
+		vmw_execbuf_unpin_panic(dev_priv);
+		goto out_no_emit;
+	}
+
+	vmw_bo_pin(dev_priv->pinned_bo, false);
+	vmw_bo_pin(dev_priv->dummy_query_bo, false);
+	dev_priv->dummy_query_bo_pinned = false;
+
+	(void) vmw_execbuf_fence_commands(NULL, dev_priv, &fence, NULL);
+	ttm_eu_fence_buffer_objects(&validate_list, (void *) fence);
+
+	ttm_bo_unref(&query_val.bo);
+	ttm_bo_unref(&pinned_val.bo);
+	ttm_bo_unref(&dev_priv->pinned_bo);
+
+out_unlock:
+	mutex_unlock(&dev_priv->cmdbuf_mutex);
+	return;
+
+out_no_emit:
+	ttm_eu_backoff_reservation(&validate_list);
+out_no_reserve:
+	ttm_bo_unref(&query_val.bo);
+	ttm_bo_unref(&pinned_val.bo);
+	ttm_bo_unref(&dev_priv->pinned_bo);
+	mutex_unlock(&dev_priv->cmdbuf_mutex);
+}
+
 
 int vmw_execbuf_ioctl(struct drm_device *dev, void *data,
 		      struct drm_file *file_priv)
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_fifo.c b/drivers/gpu/drm/vmwgfx/vmwgfx_fifo.c
index d7ed33e732a0..62d6377b8ee8 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_fifo.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_fifo.c
@@ -505,3 +505,60 @@ int vmw_fifo_send_fence(struct vmw_private *dev_priv, uint32_t *seqno)
 out_err:
 	return ret;
 }
+
+/**
+ * vmw_fifo_emit_dummy_query - emits a dummy query to the fifo.
+ *
+ * @dev_priv: The device private structure.
+ * @cid: The hardware context id used for the query.
+ *
+ * This function is used to emit a dummy occlusion query with
+ * no primitives rendered between query begin and query end.
+ * It's used to provide a query barrier, in order to know that when
+ * this query is finished, all preceding queries are also finished.
+ *
+ * A Query results structure should have been initialized at the start
+ * of the dev_priv->dummy_query_bo buffer object. And that buffer object
+ * must also be either reserved or pinned when this function is called.
+ *
+ * Returns -ENOMEM on failure to reserve fifo space.
+ */
+int vmw_fifo_emit_dummy_query(struct vmw_private *dev_priv,
+			      uint32_t cid)
+{
+	/*
+	 * A query wait without a preceding query end will
+	 * actually finish all queries for this cid
+	 * without writing to the query result structure.
+	 */
+
+	struct ttm_buffer_object *bo = dev_priv->dummy_query_bo;
+	struct {
+		SVGA3dCmdHeader header;
+		SVGA3dCmdWaitForQuery body;
+	} *cmd;
+
+	cmd = vmw_fifo_reserve(dev_priv, sizeof(*cmd));
+
+	if (unlikely(cmd == NULL)) {
+		DRM_ERROR("Out of fifo space for dummy query.\n");
+		return -ENOMEM;
+	}
+
+	cmd->header.id = SVGA_3D_CMD_WAIT_FOR_QUERY;
+	cmd->header.size = sizeof(cmd->body);
+	cmd->body.cid = cid;
+	cmd->body.type = SVGA3D_QUERYTYPE_OCCLUSION;
+
+	if (bo->mem.mem_type == TTM_PL_VRAM) {
+		cmd->body.guestResult.gmrId = SVGA_GMR_FRAMEBUFFER;
+		cmd->body.guestResult.offset = bo->offset;
+	} else {
+		cmd->body.guestResult.gmrId = bo->mem.start;
+		cmd->body.guestResult.offset = 0;
+	}
+
+	vmw_fifo_commit(dev_priv, sizeof(*cmd));
+
+	return 0;
+}
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_resource.c b/drivers/gpu/drm/vmwgfx/vmwgfx_resource.c
index c1b6ffd4ce7b..36c9d033220a 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_resource.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_resource.c
@@ -126,7 +126,7 @@ static int vmw_resource_init(struct vmw_private *dev_priv,
 	res->idr = idr;
 	res->avail = false;
 	res->dev_priv = dev_priv;
-
+	INIT_LIST_HEAD(&res->query_head);
 	do {
 		if (unlikely(idr_pre_get(idr, GFP_KERNEL) == 0))
 			return -ENOMEM;
@@ -194,8 +194,12 @@ static void vmw_hw_context_destroy(struct vmw_resource *res)
 	struct {
 		SVGA3dCmdHeader header;
 		SVGA3dCmdDestroyContext body;
-	} *cmd = vmw_fifo_reserve(dev_priv, sizeof(*cmd));
+	} *cmd;
 
+
+	vmw_execbuf_release_pinned_bo(dev_priv, true, res->id);
+
+	cmd = vmw_fifo_reserve(dev_priv, sizeof(*cmd));
 	if (unlikely(cmd == NULL)) {
 		DRM_ERROR("Failed reserving FIFO space for surface "
 			  "destruction.\n");