---

yaml
---
r: 269719
b: refs/heads/master
c: e2fa3a7
h: refs/heads/master
i:
  269717: fcd9183
  269715: e3e941d
  269711: f861bef
v: v3

[refs]

@@ -1,2 +1,2 @@
 ---
-refs/heads/master: e93daed8e2fd5ce3dc98efe9938426127a534ccc
+refs/heads/master: e2fa3a76839ada0d788549607263a036aa654243

trunk/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 @@ int vmw_dmabuf_to_placement(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.
@@ -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);
+}

trunk/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;

trunk/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
  */