drm/vmwgfx: Refactor resource management

Refactor resource management to make it easy to hook up resources
that are backed up by buffers. In particular, resources and their
backing buffers can be evicted and rebound, if supported by the device.
To avoid query deadlocks, the query code is also modified somewhat.

Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Dmitry Torokhov <dtor@vmware.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

6 files changed, 1783 insertions, 811 deletions

diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_dmabuf.c b/drivers/gpu/drm/vmwgfx/vmwgfx_dmabuf.c
index bd78257cba8..655d57f188d 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_dmabuf.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_dmabuf.c
@@ -60,7 +60,7 @@ 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);
+	vmw_execbuf_release_pinned_bo(dev_priv);
 
 	ret = ttm_bo_reserve(bo, interruptible, false, false, 0);
 	if (unlikely(ret != 0))
@@ -105,7 +105,7 @@ int vmw_dmabuf_to_vram_or_gmr(struct vmw_private *dev_priv,
 		return ret;
 
 	if (pin)
-		vmw_execbuf_release_pinned_bo(dev_priv, false, 0);
+		vmw_execbuf_release_pinned_bo(dev_priv);
 
 	ret = ttm_bo_reserve(bo, interruptible, false, false, 0);
 	if (unlikely(ret != 0))
@@ -214,8 +214,7 @@ int vmw_dmabuf_to_start_of_vram(struct vmw_private *dev_priv,
 		return ret;
 
 	if (pin)
-		vmw_execbuf_release_pinned_bo(dev_priv, false, 0);
-
+		vmw_execbuf_release_pinned_bo(dev_priv);
 	ret = ttm_bo_reserve(bo, interruptible, false, false, 0);
 	if (unlikely(ret != 0))
 		goto err_unlock;
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_drv.c b/drivers/gpu/drm/vmwgfx/vmwgfx_drv.c
index 56973cd4173..91581fd5004 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_drv.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_drv.c
@@ -432,6 +432,7 @@ static int vmw_driver_load(struct drm_device *dev, unsigned long chipset)
 	struct vmw_private *dev_priv;
 	int ret;
 	uint32_t svga_id;
+	enum vmw_res_type i;
 
 	dev_priv = kzalloc(sizeof(*dev_priv), GFP_KERNEL);
 	if (unlikely(dev_priv == NULL)) {
@@ -448,15 +449,18 @@ static int vmw_driver_load(struct drm_device *dev, unsigned long chipset)
 	mutex_init(&dev_priv->cmdbuf_mutex);
 	mutex_init(&dev_priv->release_mutex);
 	rwlock_init(&dev_priv->resource_lock);
-	idr_init(&dev_priv->context_idr);
-	idr_init(&dev_priv->surface_idr);
-	idr_init(&dev_priv->stream_idr);
+
+	for (i = vmw_res_context; i < vmw_res_max; ++i) {
+		idr_init(&dev_priv->res_idr[i]);
+		INIT_LIST_HEAD(&dev_priv->res_lru[i]);
+	}
+
 	mutex_init(&dev_priv->init_mutex);
 	init_waitqueue_head(&dev_priv->fence_queue);
 	init_waitqueue_head(&dev_priv->fifo_queue);
 	dev_priv->fence_queue_waiters = 0;
 	atomic_set(&dev_priv->fifo_queue_waiters, 0);
-	INIT_LIST_HEAD(&dev_priv->surface_lru);
+
 	dev_priv->used_memory_size = 0;
 
 	dev_priv->io_start = pci_resource_start(dev->pdev, 0);
@@ -670,9 +674,9 @@ out_err2:
 out_err1:
 	vmw_ttm_global_release(dev_priv);
 out_err0:
-	idr_destroy(&dev_priv->surface_idr);
-	idr_destroy(&dev_priv->context_idr);
-	idr_destroy(&dev_priv->stream_idr);
+	for (i = vmw_res_context; i < vmw_res_max; ++i)
+		idr_destroy(&dev_priv->res_idr[i]);
+
 	kfree(dev_priv);
 	return ret;
 }
@@ -680,9 +684,12 @@ out_err0:
 static int vmw_driver_unload(struct drm_device *dev)
 {
 	struct vmw_private *dev_priv = vmw_priv(dev);
+	enum vmw_res_type i;
 
 	unregister_pm_notifier(&dev_priv->pm_nb);
 
+	if (dev_priv->ctx.res_ht_initialized)
+		drm_ht_remove(&dev_priv->ctx.res_ht);
 	if (dev_priv->ctx.cmd_bounce)
 		vfree(dev_priv->ctx.cmd_bounce);
 	if (dev_priv->enable_fb) {
@@ -709,9 +716,9 @@ static int vmw_driver_unload(struct drm_device *dev)
 	(void)ttm_bo_clean_mm(&dev_priv->bdev, TTM_PL_VRAM);
 	(void)ttm_bo_device_release(&dev_priv->bdev);
 	vmw_ttm_global_release(dev_priv);
-	idr_destroy(&dev_priv->surface_idr);
-	idr_destroy(&dev_priv->context_idr);
-	idr_destroy(&dev_priv->stream_idr);
+
+	for (i = vmw_res_context; i < vmw_res_max; ++i)
+		idr_destroy(&dev_priv->res_idr[i]);
 
 	kfree(dev_priv);
 
@@ -935,7 +942,7 @@ 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);
+	vmw_execbuf_release_pinned_bo(dev_priv);
 
 	if (unlikely((ret != 0))) {
 		DRM_ERROR("Unable to lock TTM at VT switch.\n");
@@ -987,7 +994,8 @@ 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);
+		vmw_execbuf_release_pinned_bo(dev_priv);
+		vmw_resource_evict_all(dev_priv);
 		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 7c6f6e3a3c8..34dce9e2b6d 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_drv.h
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_drv.h
@@ -67,31 +67,46 @@ struct vmw_fpriv {
 
 struct vmw_dma_buffer {
 	struct ttm_buffer_object base;
-	struct list_head validate_list;
-	bool gmr_bound;
-	uint32_t cur_validate_node;
-	bool on_validate_list;
+	struct list_head res_list;
 };
 
+/**
+ * struct vmw_validate_buffer - Carries validation info about buffers.
+ *
+ * @base: Validation info for TTM.
+ * @hash: Hash entry for quick lookup of the TTM buffer object.
+ *
+ * This structure contains also driver private validation info
+ * on top of the info needed by TTM.
+ */
+struct vmw_validate_buffer {
+	struct ttm_validate_buffer base;
+	struct drm_hash_item hash;
+};
+
+struct vmw_res_func;
 struct vmw_resource {
 	struct kref kref;
 	struct vmw_private *dev_priv;
-	struct idr *idr;
 	int id;
-	enum ttm_object_type res_type;
 	bool avail;
-	void (*remove_from_lists) (struct vmw_resource *res);
-	void (*hw_destroy) (struct vmw_resource *res);
+	unsigned long backup_size;
+	bool res_dirty; /* Protected by backup buffer reserved */
+	bool backup_dirty; /* Protected by backup buffer reserved */
+	struct vmw_dma_buffer *backup;
+	unsigned long backup_offset;
+	const struct vmw_res_func *func;
+	struct list_head lru_head; /* Protected by the resource lock */
+	struct list_head mob_head; /* Protected by @backup reserved */
 	void (*res_free) (struct vmw_resource *res);
-	struct list_head validate_head;
-	struct list_head query_head; /* Protected by the cmdbuf mutex */
-	/* TODO is a generic snooper needed? */
-#if 0
-	void (*snoop)(struct vmw_resource *res,
-		      struct ttm_object_file *tfile,
-		      SVGA3dCmdHeader *header);
-	void *snoop_priv;
-#endif
+	void (*hw_destroy) (struct vmw_resource *res);
+};
+
+enum vmw_res_type {
+	vmw_res_context,
+	vmw_res_surface,
+	vmw_res_stream,
+	vmw_res_max
 };
 
 struct vmw_cursor_snooper {
@@ -105,20 +120,18 @@ struct vmw_surface_offset;
 
 struct vmw_surface {
 	struct vmw_resource res;
-	struct list_head lru_head; /* Protected by the resource lock */
 	uint32_t flags;
 	uint32_t format;
 	uint32_t mip_levels[DRM_VMW_MAX_SURFACE_FACES];
+	struct drm_vmw_size base_size;
 	struct drm_vmw_size *sizes;
 	uint32_t num_sizes;
-
 	bool scanout;
-
 	/* TODO so far just a extra pointer */
 	struct vmw_cursor_snooper snooper;
-	struct ttm_buffer_object *backup;
 	struct vmw_surface_offset *offsets;
-	uint32_t backup_size;
+	SVGA3dTextureFilter autogen_filter;
+	uint32_t multisample_count;
 };
 
 struct vmw_marker_queue {
@@ -145,29 +158,46 @@ struct vmw_relocation {
 	uint32_t index;
 };
 
+/**
+ * struct vmw_res_cache_entry - resource information cache entry
+ *
+ * @valid: Whether the entry is valid, which also implies that the execbuf
+ * code holds a reference to the resource, and it's placed on the
+ * validation list.
+ * @handle: User-space handle of a resource.
+ * @res: Non-ref-counted pointer to the resource.
+ *
+ * Used to avoid frequent repeated user-space handle lookups of the
+ * same resource.
+ */
+struct vmw_res_cache_entry {
+	bool valid;
+	uint32_t handle;
+	struct vmw_resource *res;
+	struct vmw_resource_val_node *node;
+};
+
 struct vmw_sw_context{
-	struct ida bo_list;
-	uint32_t last_cid;
-	bool cid_valid;
+	struct drm_open_hash res_ht;
+	bool res_ht_initialized;
 	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;
 	struct ttm_object_file *tfile;
 	struct list_head validate_nodes;
 	struct vmw_relocation relocs[VMWGFX_MAX_RELOCATIONS];
 	uint32_t cur_reloc;
-	struct ttm_validate_buffer val_bufs[VMWGFX_MAX_VALIDATIONS];
+	struct vmw_validate_buffer val_bufs[VMWGFX_MAX_VALIDATIONS];
 	uint32_t cur_val_buf;
 	uint32_t *cmd_bounce;
 	uint32_t cmd_bounce_size;
 	struct list_head resource_list;
 	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 list_head res_relocations;
+	uint32_t *buf_start;
+	struct vmw_res_cache_entry res_cache[vmw_res_max];
+	struct vmw_resource *last_query_ctx;
+	bool needs_post_query_barrier;
+	struct vmw_resource *error_resource;
 };
 
 struct vmw_legacy_display;
@@ -242,10 +272,7 @@ struct vmw_private {
 	 */
 
 	rwlock_t resource_lock;
-	struct idr context_idr;
-	struct idr surface_idr;
-	struct idr stream_idr;
-
+	struct idr res_idr[vmw_res_max];
 	/*
 	 * Block lastclose from racing with firstopen.
 	 */
@@ -320,6 +347,7 @@ struct vmw_private {
 	struct ttm_buffer_object *dummy_query_bo;
 	struct ttm_buffer_object *pinned_bo;
 	uint32_t query_cid;
+	uint32_t query_cid_valid;
 	bool dummy_query_bo_pinned;
 
 	/*
@@ -329,10 +357,15 @@ struct vmw_private {
 	 * protected by the cmdbuf mutex for simplicity.
 	 */
 
-	struct list_head surface_lru;
+	struct list_head res_lru[vmw_res_max];
 	uint32_t used_memory_size;
 };
 
+static inline struct vmw_surface *vmw_res_to_srf(struct vmw_resource *res)
+{
+	return container_of(res, struct vmw_surface, res);
+}
+
 static inline struct vmw_private *vmw_priv(struct drm_device *dev)
 {
 	return (struct vmw_private *)dev->dev_private;
@@ -381,10 +414,16 @@ extern void vmw_gmr_unbind(struct vmw_private *dev_priv, int gmr_id);
 /**
  * Resource utilities - vmwgfx_resource.c
  */
+struct vmw_user_resource_conv;
+extern const struct vmw_user_resource_conv *user_surface_converter;
+extern const struct vmw_user_resource_conv *user_context_converter;
 
 extern struct vmw_resource *vmw_context_alloc(struct vmw_private *dev_priv);
 extern void vmw_resource_unreference(struct vmw_resource **p_res);
 extern struct vmw_resource *vmw_resource_reference(struct vmw_resource *res);
+extern int vmw_resource_validate(struct vmw_resource *res);
+extern int vmw_resource_reserve(struct vmw_resource *res, bool no_backup);
+extern bool vmw_resource_needs_backup(const struct vmw_resource *res);
 extern int vmw_context_destroy_ioctl(struct drm_device *dev, void *data,
 				     struct drm_file *file_priv);
 extern int vmw_context_define_ioctl(struct drm_device *dev, void *data,
@@ -398,14 +437,13 @@ extern int vmw_user_lookup_handle(struct vmw_private *dev_priv,
 				  uint32_t handle,
 				  struct vmw_surface **out_surf,
 				  struct vmw_dma_buffer **out_buf);
+extern int vmw_user_resource_lookup_handle(
+	struct vmw_private *dev_priv,
+	struct ttm_object_file *tfile,
+	uint32_t handle,
+	const struct vmw_user_resource_conv *converter,
+	struct vmw_resource **p_res);
 extern void vmw_surface_res_free(struct vmw_resource *res);
-extern int vmw_surface_init(struct vmw_private *dev_priv,
-			    struct vmw_surface *srf,
-			    void (*res_free) (struct vmw_resource *res));
-extern int vmw_user_surface_lookup_handle(struct vmw_private *dev_priv,
-					  struct ttm_object_file *tfile,
-					  uint32_t handle,
-					  struct vmw_surface **out);
 extern int vmw_surface_destroy_ioctl(struct drm_device *dev, void *data,
 				     struct drm_file *file_priv);
 extern int vmw_surface_define_ioctl(struct drm_device *dev, void *data,
@@ -440,7 +478,14 @@ extern int vmw_user_stream_lookup(struct vmw_private *dev_priv,
 				  struct ttm_object_file *tfile,
 				  uint32_t *inout_id,
 				  struct vmw_resource **out);
-extern void vmw_resource_unreserve(struct list_head *list);
+extern void vmw_resource_unreserve(struct vmw_resource *res,
+				   struct vmw_dma_buffer *new_backup,
+				   unsigned long new_backup_offset);
+extern void vmw_resource_move_notify(struct ttm_buffer_object *bo,
+				     struct ttm_mem_reg *mem);
+extern void vmw_fence_single_bo(struct ttm_buffer_object *bo,
+				struct vmw_fence_obj *fence);
+extern void vmw_resource_evict_all(struct vmw_private *dev_priv);
 
 /**
  * DMA buffer helper routines - vmwgfx_dmabuf.c
@@ -538,10 +583,9 @@ extern int vmw_execbuf_process(struct drm_file *file_priv,
 			       struct drm_vmw_fence_rep __user
 			       *user_fence_rep,
 			       struct vmw_fence_obj **out_fence);
-
-extern void
-vmw_execbuf_release_pinned_bo(struct vmw_private *dev_priv,
-			      bool only_on_cid_match, uint32_t cid);
+extern void __vmw_execbuf_release_pinned_bo(struct vmw_private *dev_priv,
+					    struct vmw_fence_obj *fence);
+extern void vmw_execbuf_release_pinned_bo(struct vmw_private *dev_priv);
 
 extern int vmw_execbuf_fence_commands(struct drm_file *file_priv,
 				      struct vmw_private *dev_priv,
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c b/drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c
index e5775a0db49..534c96703c3 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c
@@ -30,6 +30,181 @@
 #include <drm/ttm/ttm_bo_api.h>
 #include <drm/ttm/ttm_placement.h>
 
+#define VMW_RES_HT_ORDER 12
+
+/**
+ * struct vmw_resource_relocation - Relocation info for resources
+ *
+ * @head: List head for the software context's relocation list.
+ * @res: Non-ref-counted pointer to the resource.
+ * @offset: Offset of 4 byte entries into the command buffer where the
+ * id that needs fixup is located.
+ */
+struct vmw_resource_relocation {
+	struct list_head head;
+	const struct vmw_resource *res;
+	unsigned long offset;
+};
+
+/**
+ * struct vmw_resource_val_node - Validation info for resources
+ *
+ * @head: List head for the software context's resource list.
+ * @hash: Hash entry for quick resouce to val_node lookup.
+ * @res: Ref-counted pointer to the resource.
+ * @switch_backup: Boolean whether to switch backup buffer on unreserve.
+ * @new_backup: Refcounted pointer to the new backup buffer.
+ * @new_backup_offset: New backup buffer offset if @new_backup is non-NUll.
+ * @first_usage: Set to true the first time the resource is referenced in
+ * the command stream.
+ * @no_buffer_needed: Resources do not need to allocate buffer backup on
+ * reservation. The command stream will provide one.
+ */
+struct vmw_resource_val_node {
+	struct list_head head;
+	struct drm_hash_item hash;
+	struct vmw_resource *res;
+	struct vmw_dma_buffer *new_backup;
+	unsigned long new_backup_offset;
+	bool first_usage;
+	bool no_buffer_needed;
+};
+
+/**
+ * vmw_resource_unreserve - unreserve resources previously reserved for
+ * command submission.
+ *
+ * @list_head: list of resources to unreserve.
+ * @backoff: Whether command submission failed.
+ */
+static void vmw_resource_list_unreserve(struct list_head *list,
+					bool backoff)
+{
+	struct vmw_resource_val_node *val;
+
+	list_for_each_entry(val, list, head) {
+		struct vmw_resource *res = val->res;
+		struct vmw_dma_buffer *new_backup =
+			backoff ? NULL : val->new_backup;
+
+		vmw_resource_unreserve(res, new_backup,
+			val->new_backup_offset);
+		vmw_dmabuf_unreference(&val->new_backup);
+	}
+}
+
+
+/**
+ * vmw_resource_val_add - Add a resource to the software context's
+ * resource list if it's not already on it.
+ *
+ * @sw_context: Pointer to the software context.
+ * @res: Pointer to the resource.
+ * @p_node On successful return points to a valid pointer to a
+ * struct vmw_resource_val_node, if non-NULL on entry.
+ */
+static int vmw_resource_val_add(struct vmw_sw_context *sw_context,
+				struct vmw_resource *res,
+				struct vmw_resource_val_node **p_node)
+{
+	struct vmw_resource_val_node *node;
+	struct drm_hash_item *hash;
+	int ret;
+
+	if (likely(drm_ht_find_item(&sw_context->res_ht, (unsigned long) res,
+				    &hash) == 0)) {
+		node = container_of(hash, struct vmw_resource_val_node, hash);
+		node->first_usage = false;
+		if (unlikely(p_node != NULL))
+			*p_node = node;
+		return 0;
+	}
+
+	node = kzalloc(sizeof(*node), GFP_KERNEL);
+	if (unlikely(node == NULL)) {
+		DRM_ERROR("Failed to allocate a resource validation "
+			  "entry.\n");
+		return -ENOMEM;
+	}
+
+	node->hash.key = (unsigned long) res;
+	ret = drm_ht_insert_item(&sw_context->res_ht, &node->hash);
+	if (unlikely(ret != 0)) {
+		DRM_ERROR("Failed to initialize a resource validation "
+			  "entry.\n");
+		kfree(node);
+		return ret;
+	}
+	list_add_tail(&node->head, &sw_context->resource_list);
+	node->res = vmw_resource_reference(res);
+	node->first_usage = true;
+
+	if (unlikely(p_node != NULL))
+		*p_node = node;
+
+	return 0;
+}
+
+/**
+ * vmw_resource_relocation_add - Add a relocation to the relocation list
+ *
+ * @list: Pointer to head of relocation list.
+ * @res: The resource.
+ * @offset: Offset into the command buffer currently being parsed where the
+ * id that needs fixup is located. Granularity is 4 bytes.
+ */
+static int vmw_resource_relocation_add(struct list_head *list,
+				       const struct vmw_resource *res,
+				       unsigned long offset)
+{
+	struct vmw_resource_relocation *rel;
+
+	rel = kmalloc(sizeof(*rel), GFP_KERNEL);
+	if (unlikely(rel == NULL)) {
+		DRM_ERROR("Failed to allocate a resource relocation.\n");
+		return -ENOMEM;
+	}
+
+	rel->res = res;
+	rel->offset = offset;
+	list_add_tail(&rel->head, list);
+
+	return 0;
+}
+
+/**
+ * vmw_resource_relocations_free - Free all relocations on a list
+ *
+ * @list: Pointer to the head of the relocation list.
+ */
+static void vmw_resource_relocations_free(struct list_head *list)
+{
+	struct vmw_resource_relocation *rel, *n;
+
+	list_for_each_entry_safe(rel, n, list, head) {
+		list_del(&rel->head);
+		kfree(rel);
+	}
+}
+
+/**
+ * vmw_resource_relocations_apply - Apply all relocations on a list
+ *
+ * @cb: Pointer to the start of the command buffer bein patch. This need
+ * not be the same buffer as the one being parsed when the relocation
+ * list was built, but the contents must be the same modulo the
+ * resource ids.
+ * @list: Pointer to the head of the relocation list.
+ */
+static void vmw_resource_relocations_apply(uint32_t *cb,
+					   struct list_head *list)
+{
+	struct vmw_resource_relocation *rel;
+
+	list_for_each_entry(rel, list, head)
+		cb[rel->offset] = rel->res->id;
+}
+
 static int vmw_cmd_invalid(struct vmw_private *dev_priv,
 			   struct vmw_sw_context *sw_context,
 			   SVGA3dCmdHeader *header)
@@ -44,25 +219,11 @@ static int vmw_cmd_ok(struct vmw_private *dev_priv,
 	return 0;
 }
 
-static void vmw_resource_to_validate_list(struct vmw_sw_context *sw_context,
-					  struct vmw_resource **p_res)
-{
-	struct vmw_resource *res = *p_res;
-
-	if (list_empty(&res->validate_head)) {
-		list_add_tail(&res->validate_head, &sw_context->resource_list);
-		*p_res = NULL;
-	} else
-		vmw_resource_unreference(p_res);
-}
-
 /**
  * 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.
  *
@@ -74,21 +235,37 @@ static int vmw_bo_to_validate_list(struct vmw_sw_context *sw_context,
 				   uint32_t *p_val_node)
 {
 	uint32_t val_node;
+	struct vmw_validate_buffer *vval_buf;
 	struct ttm_validate_buffer *val_buf;
+	struct drm_hash_item *hash;
+	int ret;
 
-	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)) {
+	if (likely(drm_ht_find_item(&sw_context->res_ht, (unsigned long) bo,
+				    &hash) == 0)) {
+		vval_buf = container_of(hash, struct vmw_validate_buffer,
+					hash);
+		val_buf = &vval_buf->base;
+		val_node = vval_buf - sw_context->val_bufs;
+	} else {
+		val_node = 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;
+		}
+		vval_buf = &sw_context->val_bufs[val_node];
+		vval_buf->hash.key = (unsigned long) bo;
+		ret = drm_ht_insert_item(&sw_context->res_ht, &vval_buf->hash);
+		if (unlikely(ret != 0)) {
+			DRM_ERROR("Failed to initialize a buffer validation "
+				  "entry.\n");
+			return ret;
+		}
+		++sw_context->cur_val_buf;
+		val_buf = &vval_buf->base;
 		val_buf->bo = ttm_bo_reference(bo);
+		val_buf->reserved = false;
 		list_add_tail(&val_buf->head, &sw_context->validate_nodes);
-		++sw_context->cur_val_buf;
 	}
 
 	sw_context->fence_flags |= DRM_VMW_FENCE_FLAG_EXEC;
@@ -99,85 +276,174 @@ static int vmw_bo_to_validate_list(struct vmw_sw_context *sw_context,
 	return 0;
 }
 
-static int vmw_cmd_cid_check(struct vmw_private *dev_priv,
-			     struct vmw_sw_context *sw_context,
-			     SVGA3dCmdHeader *header)
+/**
+ * vmw_resources_reserve - Reserve all resources on the sw_context's
+ * resource list.
+ *
+ * @sw_context: Pointer to the software context.
+ *
+ * Note that since vmware's command submission currently is protected by
+ * the cmdbuf mutex, no fancy deadlock avoidance is required for resources,
+ * since only a single thread at once will attempt this.
+ */
+static int vmw_resources_reserve(struct vmw_sw_context *sw_context)
 {
-	struct vmw_resource *ctx;
-
-	struct vmw_cid_cmd {
-		SVGA3dCmdHeader header;
-		__le32 cid;
-	} *cmd;
+	struct vmw_resource_val_node *val;
 	int ret;
 
-	cmd = container_of(header, struct vmw_cid_cmd, header);
-	if (likely(sw_context->cid_valid && cmd->cid == sw_context->last_cid))
-		return 0;
+	list_for_each_entry(val, &sw_context->resource_list, head) {
+		struct vmw_resource *res = val->res;
 
-	ret = vmw_context_check(dev_priv, sw_context->tfile, cmd->cid,
-				&ctx);
-	if (unlikely(ret != 0)) {
-		DRM_ERROR("Could not find or use context %u\n",
-			  (unsigned) cmd->cid);
-		return ret;
+		ret = vmw_resource_reserve(res, val->no_buffer_needed);
+		if (unlikely(ret != 0))
+			return ret;
+
+		if (res->backup) {
+			struct ttm_buffer_object *bo = &res->backup->base;
+
+			ret = vmw_bo_to_validate_list
+				(sw_context, bo, NULL);
+
+			if (unlikely(ret != 0))
+				return ret;
+		}
 	}
+	return 0;
+}
 
-	sw_context->last_cid = cmd->cid;
-	sw_context->cid_valid = true;
-	sw_context->cur_ctx = ctx;
-	vmw_resource_to_validate_list(sw_context, &ctx);
+/**
+ * vmw_resources_validate - Validate all resources on the sw_context's
+ * resource list.
+ *
+ * @sw_context: Pointer to the software context.
+ *
+ * Before this function is called, all resource backup buffers must have
+ * been validated.
+ */
+static int vmw_resources_validate(struct vmw_sw_context *sw_context)
+{
+	struct vmw_resource_val_node *val;
+	int ret;
+
+	list_for_each_entry(val, &sw_context->resource_list, head) {
+		struct vmw_resource *res = val->res;
 
+		ret = vmw_resource_validate(res);
+		if (unlikely(ret != 0)) {
+			if (ret != -ERESTARTSYS)
+				DRM_ERROR("Failed to validate resource.\n");
+			return ret;
+		}
+	}
 	return 0;
 }
 
-static int vmw_cmd_sid_check(struct vmw_private *dev_priv,
+/**
+ * vmw_cmd_res_check - Check that a resource is present and if so, put it
+ * on the resource validate list unless it's already there.
+ *
+ * @dev_priv: Pointer to a device private structure.
+ * @sw_context: Pointer to the software context.
+ * @res_type: Resource type.
+ * @converter: User-space visisble type specific information.
+ * @id: Pointer to the location in the command buffer currently being
+ * parsed from where the user-space resource id handle is located.
+ */
+static int vmw_cmd_res_check(struct vmw_private *dev_priv,
 			     struct vmw_sw_context *sw_context,
-			     uint32_t *sid)
+			     enum vmw_res_type res_type,
+			     const struct vmw_user_resource_conv *converter,
+			     uint32_t *id,
+			     struct vmw_resource_val_node **p_val)
 {
-	struct vmw_surface *srf;
-	int ret;
+	struct vmw_res_cache_entry *rcache =
+		&sw_context->res_cache[res_type];
 	struct vmw_resource *res;
+	struct vmw_resource_val_node *node;
+	int ret;
 
-	if (*sid == SVGA3D_INVALID_ID)
+	if (*id == SVGA3D_INVALID_ID)
 		return 0;
 
-	if (likely((sw_context->sid_valid  &&
-		      *sid == sw_context->last_sid))) {
-		*sid = sw_context->sid_translation;
-		return 0;
-	}
+	/*
+	 * Fastpath in case of repeated commands referencing the same
+	 * resource
+	 */
 
-	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 "
-			  "address 0x%08lx\n",
-			  (unsigned int) *sid,
-			  (unsigned long) sid);
-		return ret;
+	if (likely(rcache->valid && *id == rcache->handle)) {
+		const struct vmw_resource *res = rcache->res;
+
+		rcache->node->first_usage = false;
+		if (p_val)
+			*p_val = rcache->node;
+
+		return vmw_resource_relocation_add
+			(&sw_context->res_relocations, res,
+			 id - sw_context->buf_start);
 	}
 
-	ret = vmw_surface_validate(dev_priv, srf);
+	ret = vmw_user_resource_lookup_handle(dev_priv,
+					      sw_context->tfile,
+					      *id,
+					      converter,
+					      &res);
 	if (unlikely(ret != 0)) {
-		if (ret != -ERESTARTSYS)
-			DRM_ERROR("Could not validate surface.\n");
-		vmw_surface_unreference(&srf);
+		DRM_ERROR("Could not find or use resource 0x%08x.\n",
+			  (unsigned) *id);
+		dump_stack();
 		return ret;
 	}
 
-	sw_context->last_sid = *sid;
-	sw_context->sid_valid = true;
-	sw_context->sid_translation = srf->res.id;
-	*sid = sw_context->sid_translation;
+	rcache->valid = true;
+	rcache->res = res;
+	rcache->handle = *id;
 
-	res = &srf->res;
-	vmw_resource_to_validate_list(sw_context, &res);
+	ret = vmw_resource_relocation_add(&sw_context->res_relocations,
+					  res,
+					  id - sw_context->buf_start);
+	if (unlikely(ret != 0))
+		goto out_no_reloc;
+
+	ret = vmw_resource_val_add(sw_context, res, &node);
+	if (unlikely(ret != 0))
+		goto out_no_reloc;
 
+	rcache->node = node;
+	if (p_val)
+		*p_val = node;
+	vmw_resource_unreference(&res);
 	return 0;
+
+out_no_reloc:
+	BUG_ON(sw_context->error_resource != NULL);
+	sw_context->error_resource = res;
+
+	return ret;
 }
 
+/**
+ * vmw_cmd_cid_check - Check a command header for valid context information.
+ *
+ * @dev_priv: Pointer to a device private structure.
+ * @sw_context: Pointer to the software context.
+ * @header: A command header with an embedded user-space context handle.
+ *
+ * Convenience function: Call vmw_cmd_res_check with the user-space context
+ * handle embedded in @header.
+ */
+static int vmw_cmd_cid_check(struct vmw_private *dev_priv,
+			     struct vmw_sw_context *sw_context,
+			     SVGA3dCmdHeader *header)
+{
+	struct vmw_cid_cmd {
+		SVGA3dCmdHeader header;
+		__le32 cid;
+	} *cmd;
+
+	cmd = container_of(header, struct vmw_cid_cmd, header);
+	return vmw_cmd_res_check(dev_priv, sw_context, vmw_res_context,
+				 user_context_converter, &cmd->cid, NULL);
+}
 
 static int vmw_cmd_set_render_target_check(struct vmw_private *dev_priv,
 					   struct vmw_sw_context *sw_context,
@@ -194,7 +460,9 @@ static int vmw_cmd_set_render_target_check(struct vmw_private *dev_priv,
 		return ret;
 
 	cmd = container_of(header, struct vmw_sid_cmd, header);
-	ret = vmw_cmd_sid_check(dev_priv, sw_context, &cmd->body.target.sid);
+	ret = vmw_cmd_res_check(dev_priv, sw_context, vmw_res_surface,
+				user_surface_converter,
+				&cmd->body.target.sid, NULL);
 	return ret;
 }
 
@@ -209,10 +477,14 @@ static int vmw_cmd_surface_copy_check(struct vmw_private *dev_priv,
 	int ret;
 
 	cmd = container_of(header, struct vmw_sid_cmd, header);
-	ret = vmw_cmd_sid_check(dev_priv, sw_context, &cmd->body.src.sid);
+	ret = vmw_cmd_res_check(dev_priv, sw_context, vmw_res_surface,
+				user_surface_converter,
+				&cmd->body.src.sid, NULL);
 	if (unlikely(ret != 0))
 		return ret;
-	return vmw_cmd_sid_check(dev_priv, sw_context, &cmd->body.dest.sid);
+	return vmw_cmd_res_check(dev_priv, sw_context, vmw_res_surface,
+				 user_surface_converter,
+				 &cmd->body.dest.sid, NULL);
 }
 
 static int vmw_cmd_stretch_blt_check(struct vmw_private *dev_priv,
@@ -226,10 +498,14 @@ static int vmw_cmd_stretch_blt_check(struct vmw_private *dev_priv,
 	int ret;
 
 	cmd = container_of(header, struct vmw_sid_cmd, header);
-	ret = vmw_cmd_sid_check(dev_priv, sw_context, &cmd->body.src.sid);
+	ret = vmw_cmd_res_check(dev_priv, sw_context, vmw_res_surface,
+				user_surface_converter,
+				&cmd->body.src.sid, NULL);
 	if (unlikely(ret != 0))
 		return ret;
-	return vmw_cmd_sid_check(dev_priv, sw_context, &cmd->body.dest.sid);
+	return vmw_cmd_res_check(dev_priv, sw_context, vmw_res_surface,
+				 user_surface_converter,
+				 &cmd->body.dest.sid, NULL);
 }
 
 static int vmw_cmd_blt_surf_screen_check(struct vmw_private *dev_priv,
@@ -248,7 +524,9 @@ static int vmw_cmd_blt_surf_screen_check(struct vmw_private *dev_priv,
 		return -EPERM;
 	}
 
-	return vmw_cmd_sid_check(dev_priv, sw_context, &cmd->body.srcImage.sid);
+	return vmw_cmd_res_check(dev_priv, sw_context, vmw_res_surface,
+				 user_surface_converter,
+				 &cmd->body.srcImage.sid, NULL);
 }
 
 static int vmw_cmd_present_check(struct vmw_private *dev_priv,
@@ -268,14 +546,15 @@ static int vmw_cmd_present_check(struct vmw_private *dev_priv,
 		return -EPERM;
 	}
 
-	return vmw_cmd_sid_check(dev_priv, sw_context, &cmd->body.sid);
+	return vmw_cmd_res_check(dev_priv, sw_context, vmw_res_surface,
+				 user_surface_converter, &cmd->body.sid,
+				 NULL);
 }
 
 /**
  * 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.
  *
@@ -283,18 +562,18 @@ static int vmw_cmd_present_check(struct vmw_private *dev_priv,
  * 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.
+ * command batch.
  */
 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)
 {
+	struct vmw_res_cache_entry *ctx_entry =
+		&sw_context->res_cache[vmw_res_context];
 	int ret;
-	bool add_cid = false;
-	uint32_t cid_to_add;
+
+	BUG_ON(!ctx_entry->valid);
+	sw_context->last_query_ctx = ctx_entry->res;
 
 	if (unlikely(new_query_bo != sw_context->cur_query_bo)) {
 
@@ -304,9 +583,7 @@ static int vmw_query_bo_switch_prepare(struct vmw_private *dev_priv,
 		}
 
 		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;
+			sw_context->needs_post_query_barrier = true;
 			ret = vmw_bo_to_validate_list(sw_context,
 						      sw_context->cur_query_bo,
 						      NULL);
@@ -323,27 +600,6 @@ static int vmw_query_bo_switch_prepare(struct vmw_private *dev_priv,
 
 	}
 
-	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,
-					      NULL);
-		if (unlikely(ret != 0))
-			return ret;
-	}
 	return 0;
 }
 
@@ -355,10 +611,9 @@ static int vmw_query_bo_switch_prepare(struct vmw_private *dev_priv,
  * @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.
+ * preceding queries have finished, and the old q