drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c

   1 /**************************************************************************
   2  *
   3  * Copyright © 2009 VMware, Inc., Palo Alto, CA., USA
   4  * All Rights Reserved.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the
   8  * "Software"), to deal in the Software without restriction, including
   9  * without limitation the rights to use, copy, modify, merge, publish,
  10  * distribute, sub license, and/or sell copies of the Software, and to
  11  * permit persons to whom the Software is furnished to do so, subject to
  12  * the following conditions:
  13  *
  14  * The above copyright notice and this permission notice (including the
  15  * next paragraph) shall be included in all copies or substantial portions
  16  * of the Software.
  17  *
  18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  19  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  20  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  21  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
  22  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  23  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  24  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  25  *
  26  **************************************************************************/
  27
  28 #include "vmwgfx_drv.h"
  29 #include "vmwgfx_reg.h"
  30 #include <drm/ttm/ttm_bo_api.h>
  31 #include <drm/ttm/ttm_placement.h>
  32
  33 #define VMW_RES_HT_ORDER 12
  34
  35 /**
  36  * struct vmw_resource_relocation - Relocation info for resources
  37  *
  38  * @head: List head for the software context's relocation list.
  39  * @res: Non-ref-counted pointer to the resource.
  40  * @offset: Offset of 4 byte entries into the command buffer where the
  41  * id that needs fixup is located.
  42  */
  43 struct vmw_resource_relocation {
  44         struct list_head head;
  45         const struct vmw_resource *res;
  46         unsigned long offset;
  47 };
  48
  49 /**
  50  * struct vmw_resource_val_node - Validation info for resources
  51  *
  52  * @head: List head for the software context's resource list.
  53  * @hash: Hash entry for quick resouce to val_node lookup.
  54  * @res: Ref-counted pointer to the resource.
  55  * @switch_backup: Boolean whether to switch backup buffer on unreserve.
  56  * @new_backup: Refcounted pointer to the new backup buffer.
  57  * @new_backup_offset: New backup buffer offset if @new_backup is non-NUll.
  58  * @first_usage: Set to true the first time the resource is referenced in
  59  * the command stream.
  60  * @no_buffer_needed: Resources do not need to allocate buffer backup on
  61  * reservation. The command stream will provide one.
  62  */
  63 struct vmw_resource_val_node {
  64         struct list_head head;
  65         struct drm_hash_item hash;
  66         struct vmw_resource *res;
  67         struct vmw_dma_buffer *new_backup;
  68         unsigned long new_backup_offset;
  69         bool first_usage;
  70         bool no_buffer_needed;
  71 };
  72
  73 /**
  74  * vmw_resource_unreserve - unreserve resources previously reserved for
  75  * command submission.
  76  *
  77  * @list_head: list of resources to unreserve.
  78  * @backoff: Whether command submission failed.
  79  */
  80 static void vmw_resource_list_unreserve(struct list_head *list,
  81                                         bool backoff)
  82 {
  83         struct vmw_resource_val_node *val;
  84
  85         list_for_each_entry(val, list, head) {
  86                 struct vmw_resource *res = val->res;
  87                 struct vmw_dma_buffer *new_backup =
  88                         backoff ? NULL : val->new_backup;
  89
  90                 vmw_resource_unreserve(res, new_backup,
  91                         val->new_backup_offset);
  92                 vmw_dmabuf_unreference(&val->new_backup);
  93         }
  94 }
  95
  96
  97 /**
  98  * vmw_resource_val_add - Add a resource to the software context's
  99  * resource list if it's not already on it.
 100  *
 101  * @sw_context: Pointer to the software context.
 102  * @res: Pointer to the resource.
 103  * @p_node On successful return points to a valid pointer to a
 104  * struct vmw_resource_val_node, if non-NULL on entry.
 105  */
 106 static int vmw_resource_val_add(struct vmw_sw_context *sw_context,
 107                                 struct vmw_resource *res,
 108                                 struct vmw_resource_val_node **p_node)
 109 {
 110         struct vmw_resource_val_node *node;
 111         struct drm_hash_item *hash;
 112         int ret;
 113
 114         if (likely(drm_ht_find_item(&sw_context->res_ht, (unsigned long) res,
 115                                     &hash) == 0)) {
 116                 node = container_of(hash, struct vmw_resource_val_node, hash);
 117                 node->first_usage = false;
 118                 if (unlikely(p_node != NULL))
 119                         *p_node = node;
 120                 return 0;
 121         }
 122
 123         node = kzalloc(sizeof(*node), GFP_KERNEL);
 124         if (unlikely(node == NULL)) {
 125                 DRM_ERROR("Failed to allocate a resource validation "
 126                           "entry.\n");
 127                 return -ENOMEM;
 128         }
 129
 130         node->hash.key = (unsigned long) res;
 131         ret = drm_ht_insert_item(&sw_context->res_ht, &node->hash);
 132         if (unlikely(ret != 0)) {
 133                 DRM_ERROR("Failed to initialize a resource validation "
 134                           "entry.\n");
 135                 kfree(node);
 136                 return ret;
 137         }
 138         list_add_tail(&node->head, &sw_context->resource_list);
 139         node->res = vmw_resource_reference(res);
 140         node->first_usage = true;
 141
 142         if (unlikely(p_node != NULL))
 143                 *p_node = node;
 144
 145         return 0;
 146 }
 147
 148 /**
 149  * vmw_resource_relocation_add - Add a relocation to the relocation list
 150  *
 151  * @list: Pointer to head of relocation list.
 152  * @res: The resource.
 153  * @offset: Offset into the command buffer currently being parsed where the
 154  * id that needs fixup is located. Granularity is 4 bytes.
 155  */
 156 static int vmw_resource_relocation_add(struct list_head *list,
 157                                        const struct vmw_resource *res,
 158                                        unsigned long offset)
 159 {
 160         struct vmw_resource_relocation *rel;
 161
 162         rel = kmalloc(sizeof(*rel), GFP_KERNEL);
 163         if (unlikely(rel == NULL)) {
 164                 DRM_ERROR("Failed to allocate a resource relocation.\n");
 165                 return -ENOMEM;
 166         }
 167
 168         rel->res = res;
 169         rel->offset = offset;
 170         list_add_tail(&rel->head, list);
 171
 172         return 0;
 173 }
 174
 175 /**
 176  * vmw_resource_relocations_free - Free all relocations on a list
 177  *
 178  * @list: Pointer to the head of the relocation list.
 179  */
 180 static void vmw_resource_relocations_free(struct list_head *list)
 181 {
 182         struct vmw_resource_relocation *rel, *n;
 183
 184         list_for_each_entry_safe(rel, n, list, head) {
 185                 list_del(&rel->head);
 186                 kfree(rel);
 187         }
 188 }
 189
 190 /**
 191  * vmw_resource_relocations_apply - Apply all relocations on a list
 192  *
 193  * @cb: Pointer to the start of the command buffer bein patch. This need
 194  * not be the same buffer as the one being parsed when the relocation
 195  * list was built, but the contents must be the same modulo the
 196  * resource ids.
 197  * @list: Pointer to the head of the relocation list.
 198  */
 199 static void vmw_resource_relocations_apply(uint32_t *cb,
 200                                            struct list_head *list)
 201 {
 202         struct vmw_resource_relocation *rel;
 203
 204         list_for_each_entry(rel, list, head)
 205                 cb[rel->offset] = rel->res->id;
 206 }
 207
 208 static int vmw_cmd_invalid(struct vmw_private *dev_priv,
 209                            struct vmw_sw_context *sw_context,
 210                            SVGA3dCmdHeader *header)
 211 {
 212         return capable(CAP_SYS_ADMIN) ? : -EINVAL;
 213 }
 214
 215 static int vmw_cmd_ok(struct vmw_private *dev_priv,
 216                       struct vmw_sw_context *sw_context,
 217                       SVGA3dCmdHeader *header)
 218 {
 219         return 0;
 220 }
 221
 222 /**
 223  * vmw_bo_to_validate_list - add a bo to a validate list
 224  *
 225  * @sw_context: The software context used for this command submission batch.
 226  * @bo: The buffer object to add.
 227  * @validate_as_mob: Validate this buffer as a MOB.
 228  * @p_val_node: If non-NULL Will be updated with the validate node number
 229  * on return.
 230  *
 231  * Returns -EINVAL if the limit of number of buffer objects per command
 232  * submission is reached.
 233  */
 234 static int vmw_bo_to_validate_list(struct vmw_sw_context *sw_context,
 235                                    struct ttm_buffer_object *bo,
 236                                    bool validate_as_mob,
 237                                    uint32_t *p_val_node)
 238 {
 239         uint32_t val_node;
 240         struct vmw_validate_buffer *vval_buf;
 241         struct ttm_validate_buffer *val_buf;
 242         struct drm_hash_item *hash;
 243         int ret;
 244
 245         if (likely(drm_ht_find_item(&sw_context->res_ht, (unsigned long) bo,
 246                                     &hash) == 0)) {
 247                 vval_buf = container_of(hash, struct vmw_validate_buffer,
 248                                         hash);
 249                 if (unlikely(vval_buf->validate_as_mob != validate_as_mob)) {
 250                         DRM_ERROR("Inconsistent buffer usage.\n");
 251                         return -EINVAL;
 252                 }
 253                 val_buf = &vval_buf->base;
 254                 val_node = vval_buf - sw_context->val_bufs;
 255         } else {
 256                 val_node = sw_context->cur_val_buf;
 257                 if (unlikely(val_node >= VMWGFX_MAX_VALIDATIONS)) {
 258                         DRM_ERROR("Max number of DMA buffers per submission "
 259                                   "exceeded.\n");
 260                         return -EINVAL;
 261                 }
 262                 vval_buf = &sw_context->val_bufs[val_node];
 263                 vval_buf->hash.key = (unsigned long) bo;
 264                 ret = drm_ht_insert_item(&sw_context->res_ht, &vval_buf->hash);
 265                 if (unlikely(ret != 0)) {
 266                         DRM_ERROR("Failed to initialize a buffer validation "
 267                                   "entry.\n");
 268                         return ret;
 269                 }
 270                 ++sw_context->cur_val_buf;
 271                 val_buf = &vval_buf->base;
 272                 val_buf->bo = ttm_bo_reference(bo);
 273                 val_buf->reserved = false;
 274                 list_add_tail(&val_buf->head, &sw_context->validate_nodes);
 275                 vval_buf->validate_as_mob = validate_as_mob;
 276         }
 277
 278         sw_context->fence_flags |= DRM_VMW_FENCE_FLAG_EXEC;
 279
 280         if (p_val_node)
 281                 *p_val_node = val_node;
 282
 283         return 0;
 284 }
 285
 286 /**
 287  * vmw_resources_reserve - Reserve all resources on the sw_context's
 288  * resource list.
 289  *
 290  * @sw_context: Pointer to the software context.
 291  *
 292  * Note that since vmware's command submission currently is protected by
 293  * the cmdbuf mutex, no fancy deadlock avoidance is required for resources,
 294  * since only a single thread at once will attempt this.
 295  */
 296 static int vmw_resources_reserve(struct vmw_sw_context *sw_context)
 297 {
 298         struct vmw_resource_val_node *val;
 299         int ret;
 300
 301         list_for_each_entry(val, &sw_context->resource_list, head) {
 302                 struct vmw_resource *res = val->res;
 303
 304                 ret = vmw_resource_reserve(res, val->no_buffer_needed);
 305                 if (unlikely(ret != 0))
 306                         return ret;
 307
 308                 if (res->backup) {
 309                         struct ttm_buffer_object *bo = &res->backup->base;
 310
 311                         ret = vmw_bo_to_validate_list
 312                                 (sw_context, bo,
 313                                  vmw_resource_needs_backup(res), NULL);
 314
 315                         if (unlikely(ret != 0))
 316                                 return ret;
 317                 }
 318         }
 319         return 0;
 320 }
 321
 322 /**
 323  * vmw_resources_validate - Validate all resources on the sw_context's
 324  * resource list.
 325  *
 326  * @sw_context: Pointer to the software context.
 327  *
 328  * Before this function is called, all resource backup buffers must have
 329  * been validated.
 330  */
 331 static int vmw_resources_validate(struct vmw_sw_context *sw_context)
 332 {
 333         struct vmw_resource_val_node *val;
 334         int ret;
 335
 336         list_for_each_entry(val, &sw_context->resource_list, head) {
 337                 struct vmw_resource *res = val->res;
 338
 339                 ret = vmw_resource_validate(res);
 340                 if (unlikely(ret != 0)) {
 341                         if (ret != -ERESTARTSYS)
 342                                 DRM_ERROR("Failed to validate resource.\n");
 343                         return ret;
 344                 }
 345         }
 346         return 0;
 347 }
 348
 349 /**
 350  * vmw_cmd_res_check - Check that a resource is present and if so, put it
 351  * on the resource validate list unless it's already there.
 352  *
 353  * @dev_priv: Pointer to a device private structure.
 354  * @sw_context: Pointer to the software context.
 355  * @res_type: Resource type.
 356  * @converter: User-space visisble type specific information.
 357  * @id: Pointer to the location in the command buffer currently being
 358  * parsed from where the user-space resource id handle is located.
 359  */
 360 static int vmw_cmd_res_check(struct vmw_private *dev_priv,
 361                              struct vmw_sw_context *sw_context,
 362                              enum vmw_res_type res_type,
 363                              const struct vmw_user_resource_conv *converter,
 364                              uint32_t *id,
 365                              struct vmw_resource_val_node **p_val)
 366 {
 367         struct vmw_res_cache_entry *rcache =
 368                 &sw_context->res_cache[res_type];
 369         struct vmw_resource *res;
 370         struct vmw_resource_val_node *node;
 371         int ret;
 372
 373         if (*id == SVGA3D_INVALID_ID)
 374                 return 0;
 375
 376         /*
 377          * Fastpath in case of repeated commands referencing the same
 378          * resource
 379          */
 380
 381         if (likely(rcache->valid && *id == rcache->handle)) {
 382                 const struct vmw_resource *res = rcache->res;
 383
 384                 rcache->node->first_usage = false;
 385                 if (p_val)
 386                         *p_val = rcache->node;
 387
 388                 return vmw_resource_relocation_add
 389                         (&sw_context->res_relocations, res,
 390                          id - sw_context->buf_start);
 391         }
 392
 393         ret = vmw_user_resource_lookup_handle(dev_priv,
 394                                               sw_context->tfile,
 395                                               *id,
 396                                               converter,
 397                                               &res);
 398         if (unlikely(ret != 0)) {
 399                 DRM_ERROR("Could not find or use resource 0x%08x.\n",
 400                           (unsigned) *id);
 401                 dump_stack();
 402                 return ret;
 403         }
 404
 405         rcache->valid = true;
 406         rcache->res = res;
 407         rcache->handle = *id;
 408
 409         ret = vmw_resource_relocation_add(&sw_context->res_relocations,
 410                                           res,
 411                                           id - sw_context->buf_start);
 412         if (unlikely(ret != 0))
 413                 goto out_no_reloc;
 414
 415         ret = vmw_resource_val_add(sw_context, res, &node);
 416         if (unlikely(ret != 0))
 417                 goto out_no_reloc;
 418
 419         rcache->node = node;
 420         if (p_val)
 421                 *p_val = node;
 422         vmw_resource_unreference(&res);
 423         return 0;
 424
 425 out_no_reloc:
 426         BUG_ON(sw_context->error_resource != NULL);
 427         sw_context->error_resource = res;
 428
 429         return ret;
 430 }
 431
 432 /**
 433  * vmw_cmd_cid_check - Check a command header for valid context information.
 434  *
 435  * @dev_priv: Pointer to a device private structure.
 436  * @sw_context: Pointer to the software context.
 437  * @header: A command header with an embedded user-space context handle.
 438  *
 439  * Convenience function: Call vmw_cmd_res_check with the user-space context
 440  * handle embedded in @header.
 441  */
 442 static int vmw_cmd_cid_check(struct vmw_private *dev_priv,
 443                              struct vmw_sw_context *sw_context,
 444                              SVGA3dCmdHeader *header)
 445 {
 446         struct vmw_cid_cmd {
 447                 SVGA3dCmdHeader header;
 448                 __le32 cid;
 449         } *cmd;
 450
 451         cmd = container_of(header, struct vmw_cid_cmd, header);
 452         return vmw_cmd_res_check(dev_priv, sw_context, vmw_res_context,
 453                                  user_context_converter, &cmd->cid, NULL);
 454 }
 455
 456 static int vmw_cmd_set_render_target_check(struct vmw_private *dev_priv,
 457                                            struct vmw_sw_context *sw_context,
 458                                            SVGA3dCmdHeader *header)
 459 {
 460         struct vmw_sid_cmd {
 461                 SVGA3dCmdHeader header;
 462                 SVGA3dCmdSetRenderTarget body;
 463         } *cmd;
 464         int ret;
 465
 466         ret = vmw_cmd_cid_check(dev_priv, sw_context, header);
 467         if (unlikely(ret != 0))
 468                 return ret;
 469
 470         cmd = container_of(header, struct vmw_sid_cmd, header);
 471         ret = vmw_cmd_res_check(dev_priv, sw_context, vmw_res_surface,
 472                                 user_surface_converter,
 473                                 &cmd->body.target.sid, NULL);
 474         return ret;
 475 }
 476
 477 static int vmw_cmd_surface_copy_check(struct vmw_private *dev_priv,
 478                                       struct vmw_sw_context *sw_context,
 479                                       SVGA3dCmdHeader *header)
 480 {
 481         struct vmw_sid_cmd {
 482                 SVGA3dCmdHeader header;
 483                 SVGA3dCmdSurfaceCopy body;
 484         } *cmd;
 485         int ret;
 486
 487         cmd = container_of(header, struct vmw_sid_cmd, header);
 488         ret = vmw_cmd_res_check(dev_priv, sw_context, vmw_res_surface,
 489                                 user_surface_converter,
 490                                 &cmd->body.src.sid, NULL);
 491         if (unlikely(ret != 0))
 492                 return ret;
 493         return vmw_cmd_res_check(dev_priv, sw_context, vmw_res_surface,
 494                                  user_surface_converter,
 495                                  &cmd->body.dest.sid, NULL);
 496 }
 497
 498 static int vmw_cmd_stretch_blt_check(struct vmw_private *dev_priv,
 499                                      struct vmw_sw_context *sw_context,
 500                                      SVGA3dCmdHeader *header)
 501 {
 502         struct vmw_sid_cmd {
 503                 SVGA3dCmdHeader header;
 504                 SVGA3dCmdSurfaceStretchBlt body;
 505         } *cmd;
 506         int ret;
 507
 508         cmd = container_of(header, struct vmw_sid_cmd, header);
 509         ret = vmw_cmd_res_check(dev_priv, sw_context, vmw_res_surface,
 510                                 user_surface_converter,
 511                                 &cmd->body.src.sid, NULL);
 512         if (unlikely(ret != 0))
 513                 return ret;
 514         return vmw_cmd_res_check(dev_priv, sw_context, vmw_res_surface,
 515                                  user_surface_converter,
 516                                  &cmd->body.dest.sid, NULL);
 517 }
 518
 519 static int vmw_cmd_blt_surf_screen_check(struct vmw_private *dev_priv,
 520                                          struct vmw_sw_context *sw_context,
 521                                          SVGA3dCmdHeader *header)
 522 {
 523         struct vmw_sid_cmd {
 524                 SVGA3dCmdHeader header;
 525                 SVGA3dCmdBlitSurfaceToScreen body;
 526         } *cmd;
 527
 528         cmd = container_of(header, struct vmw_sid_cmd, header);
 529
 530         if (unlikely(!sw_context->kernel)) {
 531                 DRM_ERROR("Kernel only SVGA3d command: %u.\n", cmd->header.id);
 532                 return -EPERM;
 533         }
 534
 535         return vmw_cmd_res_check(dev_priv, sw_context, vmw_res_surface,
 536                                  user_surface_converter,
 537                                  &cmd->body.srcImage.sid, NULL);
 538 }
 539
 540 static int vmw_cmd_present_check(struct vmw_private *dev_priv,
 541                                  struct vmw_sw_context *sw_context,
 542                                  SVGA3dCmdHeader *header)
 543 {
 544         struct vmw_sid_cmd {
 545                 SVGA3dCmdHeader header;
 546                 SVGA3dCmdPresent body;
 547         } *cmd;
 548
 549
 550         cmd = container_of(header, struct vmw_sid_cmd, header);
 551
 552         if (unlikely(!sw_context->kernel)) {
 553                 DRM_ERROR("Kernel only SVGA3d command: %u.\n", cmd->header.id);
 554                 return -EPERM;
 555         }
 556
 557         return vmw_cmd_res_check(dev_priv, sw_context, vmw_res_surface,
 558                                  user_surface_converter, &cmd->body.sid,
 559                                  NULL);
 560 }
 561
 562 /**
 563  * vmw_query_bo_switch_prepare - Prepare to switch pinned buffer for queries.
 564  *
 565  * @dev_priv: The device private structure.
 566  * @new_query_bo: The new buffer holding query results.
 567  * @sw_context: The software context used for this command submission.
 568  *
 569  * This function checks whether @new_query_bo is suitable for holding
 570  * query results, and if another buffer currently is pinned for query
 571  * results. If so, the function prepares the state of @sw_context for
 572  * switching pinned buffers after successful submission of the current
 573  * command batch.
 574  */
 575 static int vmw_query_bo_switch_prepare(struct vmw_private *dev_priv,
 576                                        struct ttm_buffer_object *new_query_bo,
 577                                        struct vmw_sw_context *sw_context)
 578 {
 579         struct vmw_res_cache_entry *ctx_entry =
 580                 &sw_context->res_cache[vmw_res_context];
 581         int ret;
 582
 583         BUG_ON(!ctx_entry->valid);
 584         sw_context->last_query_ctx = ctx_entry->res;
 585
 586         if (unlikely(new_query_bo != sw_context->cur_query_bo)) {
 587
 588                 if (unlikely(new_query_bo->num_pages > 4)) {
 589                         DRM_ERROR("Query buffer too large.\n");
 590                         return -EINVAL;
 591                 }
 592
 593                 if (unlikely(sw_context->cur_query_bo != NULL)) {
 594                         sw_context->needs_post_query_barrier = true;
 595                         ret = vmw_bo_to_validate_list(sw_context,
 596                                                       sw_context->cur_query_bo,
 597                                                       dev_priv->has_mob, NULL);
 598                         if (unlikely(ret != 0))
 599                                 return ret;
 600                 }
 601                 sw_context->cur_query_bo = new_query_bo;
 602
 603                 ret = vmw_bo_to_validate_list(sw_context,
 604                                               dev_priv->dummy_query_bo,
 605                                               dev_priv->has_mob, NULL);
 606                 if (unlikely(ret != 0))
 607                         return ret;
 608
 609         }
 610
 611         return 0;
 612 }
 613
 614
 615 /**
 616  * vmw_query_bo_switch_commit - Finalize switching pinned query buffer
 617  *
 618  * @dev_priv: The device private structure.
 619  * @sw_context: The software context used for this command submission batch.
 620  *
 621  * This function will check if we're switching query buffers, and will then,
 622  * issue a dummy occlusion query wait used as a query barrier. When the fence
 623  * object following that query wait has signaled, we are sure that all
 624  * preceding queries have finished, and the old query buffer can be unpinned.
 625  * However, since both the new query buffer and the old one are fenced with
 626  * that fence, we can do an asynchronus unpin now, and be sure that the
 627  * old query buffer won't be moved until the fence has signaled.
 628  *
 629  * As mentioned above, both the new - and old query buffers need to be fenced
 630  * using a sequence emitted *after* calling this function.
 631  */
 632 static void vmw_query_bo_switch_commit(struct vmw_private *dev_priv,
 633                                      struct vmw_sw_context *sw_context)
 634 {
 635         /*
 636          * The validate list should still hold references to all
 637          * contexts here.
 638          */
 639
 640         if (sw_context->needs_post_query_barrier) {
 641                 struct vmw_res_cache_entry *ctx_entry =
 642                         &sw_context->res_cache[vmw_res_context];
 643                 struct vmw_resource *ctx;
 644                 int ret;
 645
 646                 BUG_ON(!ctx_entry->valid);
 647                 ctx = ctx_entry->res;
 648
 649                 ret = vmw_fifo_emit_dummy_query(dev_priv, ctx->id);
 650
 651                 if (unlikely(ret != 0))
 652                         DRM_ERROR("Out of fifo space for dummy query.\n");
 653         }
 654
 655         if (dev_priv->pinned_bo != sw_context->cur_query_bo) {
 656                 if (dev_priv->pinned_bo) {
 657                         vmw_bo_pin(dev_priv->pinned_bo, false);
 658                         ttm_bo_unref(&dev_priv->pinned_bo);
 659                 }
 660
 661                 if (!sw_context->needs_post_query_barrier) {
 662                         vmw_bo_pin(sw_context->cur_query_bo, true);
 663
 664                         /*
 665                          * We pin also the dummy_query_bo buffer so that we
 666                          * don't need to validate it when emitting
 667                          * dummy queries in context destroy paths.
 668                          */
 669
 670                         vmw_bo_pin(dev_priv->dummy_query_bo, true);
 671                         dev_priv->dummy_query_bo_pinned = true;
 672
 673                         BUG_ON(sw_context->last_query_ctx == NULL);
 674                         dev_priv->query_cid = sw_context->last_query_ctx->id;
 675                         dev_priv->query_cid_valid = true;
 676                         dev_priv->pinned_bo =
 677                                 ttm_bo_reference(sw_context->cur_query_bo);
 678                 }
 679         }
 680 }
 681
 682 /**
 683  * vmw_translate_guest_pointer - Prepare to translate a user-space buffer
 684  * handle to a valid SVGAGuestPtr
 685  *
 686  * @dev_priv: Pointer to a device private structure.
 687  * @sw_context: The software context used for this command batch validation.
 688  * @ptr: Pointer to the user-space handle to be translated.
 689  * @vmw_bo_p: Points to a location that, on successful return will carry
 690  * a reference-counted pointer to the DMA buffer identified by the
 691  * user-space handle in @id.
 692  *
 693  * This function saves information needed to translate a user-space buffer
 694  * handle to a valid SVGAGuestPtr. The translation does not take place
 695  * immediately, but during a call to vmw_apply_relocations().
 696  * This function builds a relocation list and a list of buffers to validate.
 697  * The former needs to be freed using either vmw_apply_relocations() or
 698  * vmw_free_relocations(). The latter needs to be freed using
 699  * vmw_clear_validations.
 700  */
 701 static int vmw_translate_guest_ptr(struct vmw_private *dev_priv,
 702                                    struct vmw_sw_context *sw_context,
 703                                    SVGAGuestPtr *ptr,
 704                                    struct vmw_dma_buffer **vmw_bo_p)
 705 {
 706         struct vmw_dma_buffer *vmw_bo = NULL;
 707         struct ttm_buffer_object *bo;
 708         uint32_t handle = ptr->gmrId;
 709         struct vmw_relocation *reloc;
 710         int ret;
 711
 712         ret = vmw_user_dmabuf_lookup(sw_context->tfile, handle, &vmw_bo);
 713         if (unlikely(ret != 0)) {
 714                 DRM_ERROR("Could not find or use GMR region.\n");
 715                 return -EINVAL;
 716         }
 717         bo = &vmw_bo->base;
 718
 719         if (unlikely(sw_context->cur_reloc >= VMWGFX_MAX_RELOCATIONS)) {
 720                 DRM_ERROR("Max number relocations per submission"
 721                           " exceeded\n");
 722                 ret = -EINVAL;
 723                 goto out_no_reloc;
 724         }
 725
 726         reloc = &sw_context->relocs[sw_context->cur_reloc++];
 727         reloc->location = ptr;
 728
 729         ret = vmw_bo_to_validate_list(sw_context, bo, false, &reloc->index);
 730         if (unlikely(ret != 0))
 731                 goto out_no_reloc;
 732
 733         *vmw_bo_p = vmw_bo;
 734         return 0;
 735
 736 out_no_reloc:
 737         vmw_dmabuf_unreference(&vmw_bo);
 738         vmw_bo_p = NULL;
 739         return ret;
 740 }
 741
 742 /**
 743  * vmw_cmd_begin_query - validate a  SVGA_3D_CMD_BEGIN_QUERY command.
 744  *
 745  * @dev_priv: Pointer to a device private struct.
 746  * @sw_context: The software context used for this command submission.
 747  * @header: Pointer to the command header in the command stream.
 748  */
 749 static int vmw_cmd_begin_query(struct vmw_private *dev_priv,
 750                                struct vmw_sw_context *sw_context,
 751                                SVGA3dCmdHeader *header)
 752 {
 753         struct vmw_begin_query_cmd {
 754                 SVGA3dCmdHeader header;
 755                 SVGA3dCmdBeginQuery q;
 756         } *cmd;
 757
 758         cmd = container_of(header, struct vmw_begin_query_cmd,
 759                            header);
 760
 761         return vmw_cmd_res_check(dev_priv, sw_context, vmw_res_context,
 762                                  user_context_converter, &cmd->q.cid,
 763                                  NULL);
 764 }
 765
 766 /**
 767  * vmw_cmd_end_query - validate a  SVGA_3D_CMD_END_QUERY command.
 768  *
 769  * @dev_priv: Pointer to a device private struct.
 770  * @sw_context: The software context used for this command submission.
 771  * @header: Pointer to the command header in the command stream.
 772  */
 773 static int vmw_cmd_end_query(struct vmw_private *dev_priv,
 774                              struct vmw_sw_context *sw_context,
 775                              SVGA3dCmdHeader *header)
 776 {
 777         struct vmw_dma_buffer *vmw_bo;
 778         struct vmw_query_cmd {
 779                 SVGA3dCmdHeader header;
 780                 SVGA3dCmdEndQuery q;
 781         } *cmd;
 782         int ret;
 783
 784         cmd = container_of(header, struct vmw_query_cmd, header);
 785         ret = vmw_cmd_cid_check(dev_priv, sw_context, header);
 786         if (unlikely(ret != 0))
 787                 return ret;
 788
 789         ret = vmw_translate_guest_ptr(dev_priv, sw_context,
 790                                       &cmd->q.guestResult,
 791                                       &vmw_bo);
 792         if (unlikely(ret != 0))
 793                 return ret;
 794
 795         ret = vmw_query_bo_switch_prepare(dev_priv, &vmw_bo->base, sw_context);
 796
 797         vmw_dmabuf_unreference(&vmw_bo);
 798         return ret;
 799 }
 800
 801 /*
 802  * vmw_cmd_wait_query - validate a  SVGA_3D_CMD_WAIT_QUERY command.
 803  *
 804  * @dev_priv: Pointer to a device private struct.
 805  * @sw_context: The software context used for this command submission.
 806  * @header: Pointer to the command header in the command stream.
 807  */
 808 static int vmw_cmd_wait_query(struct vmw_private *dev_priv,
 809                               struct vmw_sw_context *sw_context,
 810                               SVGA3dCmdHeader *header)
 811 {
 812         struct vmw_dma_buffer *vmw_bo;
 813         struct vmw_query_cmd {
 814                 SVGA3dCmdHeader header;
 815                 SVGA3dCmdWaitForQuery q;
 816         } *cmd;
 817         int ret;
 818
 819         cmd = container_of(header, struct vmw_query_cmd, header);
 820         ret = vmw_cmd_cid_check(dev_priv, sw_context, header);
 821         if (unlikely(ret != 0))
 822                 return ret;
 823
 824         ret = vmw_translate_guest_ptr(dev_priv, sw_context,
 825                                       &cmd->q.guestResult,
 826                                       &vmw_bo);
 827         if (unlikely(ret != 0))
 828                 return ret;
 829
 830         vmw_dmabuf_unreference(&vmw_bo);
 831         return 0;
 832 }
 833
 834 static int vmw_cmd_dma(struct vmw_private *dev_priv,
 835                        struct vmw_sw_context *sw_context,
 836                        SVGA3dCmdHeader *header)
 837 {
 838         struct vmw_dma_buffer *vmw_bo = NULL;
 839         struct vmw_surface *srf = NULL;
 840         struct vmw_dma_cmd {
 841                 SVGA3dCmdHeader header;
 842                 SVGA3dCmdSurfaceDMA dma;
 843         } *cmd;
 844         int ret;
 845
 846         cmd = container_of(header, struct vmw_dma_cmd, header);
 847         ret = vmw_translate_guest_ptr(dev_priv, sw_context,
 848                                       &cmd->dma.guest.ptr,
 849                                       &vmw_bo);
 850         if (unlikely(ret != 0))
 851                 return ret;
 852
 853         ret = vmw_cmd_res_check(dev_priv, sw_context, vmw_res_surface,
 854                                 user_surface_converter, &cmd->dma.host.sid,
 855                                 NULL);
 856         if (unlikely(ret != 0)) {
 857                 if (unlikely(ret != -ERESTARTSYS))
 858                         DRM_ERROR("could not find surface for DMA.\n");
 859                 goto out_no_surface;
 860         }
 861
 862         srf = vmw_res_to_srf(sw_context->res_cache[vmw_res_surface].res);
 863
 864         vmw_kms_cursor_snoop(srf, sw_context->tfile, &vmw_bo->base, header);
 865
 866 out_no_surface:
 867         vmw_dmabuf_unreference(&vmw_bo);
 868         return ret;
 869 }
 870
 871 static int vmw_cmd_draw(struct vmw_private *dev_priv,
 872                         struct vmw_sw_context *sw_context,
 873                         SVGA3dCmdHeader *header)
 874 {
 875         struct vmw_draw_cmd {
 876                 SVGA3dCmdHeader header;
 877                 SVGA3dCmdDrawPrimitives body;
 878         } *cmd;
 879         SVGA3dVertexDecl *decl = (SVGA3dVertexDecl *)(
 880                 (unsigned long)header + sizeof(*cmd));
 881         SVGA3dPrimitiveRange *range;
 882         uint32_t i;
 883         uint32_t maxnum;
 884         int ret;
 885
 886         ret = vmw_cmd_cid_check(dev_priv, sw_context, header);
 887         if (unlikely(ret != 0))
 888                 return ret;
 889
 890         cmd = container_of(header, struct vmw_draw_cmd, header);
 891         maxnum = (header->size - sizeof(cmd->body)) / sizeof(*decl);
 892
 893         if (unlikely(cmd->body.numVertexDecls > maxnum)) {
 894                 DRM_ERROR("Illegal number of vertex declarations.\n");
 895                 return -EINVAL;
 896         }
 897
 898         for (i = 0; i < cmd->body.numVertexDecls; ++i, ++decl) {
 899                 ret = vmw_cmd_res_check(dev_priv, sw_context, vmw_res_surface,
 900                                         user_surface_converter,
 901                                         &decl->array.surfaceId, NULL);
 902                 if (unlikely(ret != 0))
 903                         return ret;
 904         }
 905
 906         maxnum = (header->size - sizeof(cmd->body) -
 907                   cmd->body.numVertexDecls * sizeof(*decl)) / sizeof(*range);
 908         if (unlikely(cmd->body.numRanges > maxnum)) {
 909                 DRM_ERROR("Illegal number of index ranges.\n");
 910                 return -EINVAL;
 911         }
 912
 913         range = (SVGA3dPrimitiveRange *) decl;
 914         for (i = 0; i < cmd->body.numRanges; ++i, ++range) {
 915                 ret = vmw_cmd_res_check(dev_priv, sw_context, vmw_res_surface,
 916                                         user_surface_converter,
 917                                         &range->indexArray.surfaceId, NULL);
 918                 if (unlikely(ret != 0))
 919                         return ret;
 920         }
 921         return 0;
 922 }
 923
 924
 925 static int vmw_cmd_tex_state(struct vmw_private *dev_priv,
 926                              struct vmw_sw_context *sw_context,
 927                              SVGA3dCmdHeader *header)
 928 {
 929         struct vmw_tex_state_cmd {
 930                 SVGA3dCmdHeader header;
 931                 SVGA3dCmdSetTextureState state;
 932         };
 933
 934         SVGA3dTextureState *last_state = (SVGA3dTextureState *)
 935           ((unsigned long) header + header->size + sizeof(header));
 936         SVGA3dTextureState *cur_state = (SVGA3dTextureState *)
 937                 ((unsigned long) header + sizeof(struct vmw_tex_state_cmd));
 938         int ret;
 939
 940         ret = vmw_cmd_cid_check(dev_priv, sw_context, header);
 941         if (unlikely(ret != 0))
 942                 return ret;
 943
 944         for (; cur_state < last_state; ++cur_state) {
 945                 if (likely(cur_state->name != SVGA3D_TS_BIND_TEXTURE))
 946                         continue;
 947
 948                 ret = vmw_cmd_res_check(dev_priv, sw_context, vmw_res_surface,
 949                                         user_surface_converter,
 950                                         &cur_state->value, NULL);
 951                 if (unlikely(ret != 0))
 952                         return ret;
 953         }
 954
 955         return 0;
 956 }
 957
 958 static int vmw_cmd_check_define_gmrfb(struct vmw_private *dev_priv,
 959                                       struct vmw_sw_context *sw_context,
 960                                       void *buf)
 961 {
 962         struct vmw_dma_buffer *vmw_bo;
 963         int ret;
 964
 965         struct {
 966                 uint32_t header;
 967                 SVGAFifoCmdDefineGMRFB body;
 968         } *cmd = buf;
 969
 970         ret = vmw_translate_guest_ptr(dev_priv, sw_context,
 971                                       &cmd->body.ptr,
 972                                       &vmw_bo);
 973         if (unlikely(ret != 0))
 974                 return ret;
 975
 976         vmw_dmabuf_unreference(&vmw_bo);
 977
 978         return ret;
 979 }
 980
 981 /**
 982  * vmw_cmd_set_shader - Validate an SVGA_3D_CMD_SET_SHADER
 983  * command
 984  *
 985  * @dev_priv: Pointer to a device private struct.
 986  * @sw_context: The software context being used for this batch.
 987  * @header: Pointer to the command header in the command stream.
 988  */
 989 static int vmw_cmd_set_shader(struct vmw_private *dev_priv,
 990                               struct vmw_sw_context *sw_context,
 991                               SVGA3dCmdHeader *header)
 992 {
 993         struct vmw_set_shader_cmd {
 994                 SVGA3dCmdHeader header;
 995                 SVGA3dCmdSetShader body;
 996         } *cmd;
 997         int ret;
 998
 999         cmd = container_of(header, struct vmw_set_shader_cmd,
1000                            header);
1001
1002         ret = vmw_cmd_cid_check(dev_priv, sw_context, header);
1003         if (unlikely(ret != 0))
1004                 return ret;
1005
1006         return 0;
1007 }
1008
1009 static int vmw_cmd_check_not_3d(struct vmw_private *dev_priv,
1010                                 struct vmw_sw_context *sw_context,
1011                                 void *buf, uint32_t *size)
1012 {
1013         uint32_t size_remaining = *size;
1014         uint32_t cmd_id;
1015
1016         cmd_id = le32_to_cpu(((uint32_t *)buf)[0]);
1017         switch (cmd_id) {
1018         case SVGA_CMD_UPDATE:
1019                 *size = sizeof(uint32_t) + sizeof(SVGAFifoCmdUpdate);
1020                 break;
1021         case SVGA_CMD_DEFINE_GMRFB:
1022                 *size = sizeof(uint32_t) + sizeof(SVGAFifoCmdDefineGMRFB);
1023                 break;
1024         case SVGA_CMD_BLIT_GMRFB_TO_SCREEN:
1025                 *size = sizeof(uint32_t) + sizeof(SVGAFifoCmdBlitGMRFBToScreen);
1026                 break;
1027         case SVGA_CMD_BLIT_SCREEN_TO_GMRFB:
1028                 *size = sizeof(uint32_t) + sizeof(SVGAFifoCmdBlitGMRFBToScreen);
1029                 break;
1030         default:
1031                 DRM_ERROR("Unsupported SVGA command: %u.\n", cmd_id);
1032                 return -EINVAL;
1033         }
1034
1035         if (*size > size_remaining) {
1036                 DRM_ERROR("Invalid SVGA command (size mismatch):"
1037                           " %u.\n", cmd_id);
1038                 return -EINVAL;
1039         }
1040
1041         if (unlikely(!sw_context->kernel)) {
1042                 DRM_ERROR("Kernel only SVGA command: %u.\n", cmd_id);
1043                 return -EPERM;
1044         }
1045
1046         if (cmd_id == SVGA_CMD_DEFINE_GMRFB)
1047                 return vmw_cmd_check_define_gmrfb(dev_priv, sw_context, buf);
1048
1049         return 0;
1050 }
1051
1052 typedef int (*vmw_cmd_func) (struct vmw_private *,
1053                              struct vmw_sw_context *,
1054                              SVGA3dCmdHeader *);
1055
1056 #define VMW_CMD_DEF(cmd, func) \
1057         [cmd - SVGA_3D_CMD_BASE] = func
1058
1059 static vmw_cmd_func vmw_cmd_funcs[SVGA_3D_CMD_MAX] = {
1060         VMW_CMD_DEF(SVGA_3D_CMD_SURFACE_DEFINE, &vmw_cmd_invalid),
1061         VMW_CMD_DEF(SVGA_3D_CMD_SURFACE_DESTROY, &vmw_cmd_invalid),
1062         VMW_CMD_DEF(SVGA_3D_CMD_SURFACE_COPY, &vmw_cmd_surface_copy_check),
1063         VMW_CMD_DEF(SVGA_3D_CMD_SURFACE_STRETCHBLT, &vmw_cmd_stretch_blt_check),
1064         VMW_CMD_DEF(SVGA_3D_CMD_SURFACE_DMA, &vmw_cmd_dma),
1065         VMW_CMD_DEF(SVGA_3D_CMD_CONTEXT_DEFINE, &vmw_cmd_invalid),
1066         VMW_CMD_DEF(SVGA_3D_CMD_CONTEXT_DESTROY, &vmw_cmd_invalid),
1067         VMW_CMD_DEF(SVGA_3D_CMD_SETTRANSFORM, &vmw_cmd_cid_check),
1068         VMW_CMD_DEF(SVGA_3D_CMD_SETZRANGE, &vmw_cmd_cid_check),
1069         VMW_CMD_DEF(SVGA_3D_CMD_SETRENDERSTATE, &vmw_cmd_cid_check),
1070         VMW_CMD_DEF(SVGA_3D_CMD_SETRENDERTARGET,
1071                     &vmw_cmd_set_render_target_check),
1072         VMW_CMD_DEF(SVGA_3D_CMD_SETTEXTURESTATE, &vmw_cmd_tex_state),
1073         VMW_CMD_DEF(SVGA_3D_CMD_SETMATERIAL, &vmw_cmd_cid_check),
1074         VMW_CMD_DEF(SVGA_3D_CMD_SETLIGHTDATA, &vmw_cmd_cid_check),
1075         VMW_CMD_DEF(SVGA_3D_CMD_SETLIGHTENABLED, &vmw_cmd_cid_check),
1076         VMW_CMD_DEF(SVGA_3D_CMD_SETVIEWPORT, &vmw_cmd_cid_check),
1077         VMW_CMD_DEF(SVGA_3D_CMD_SETCLIPPLANE, &vmw_cmd_cid_check),
1078         VMW_CMD_DEF(SVGA_3D_CMD_CLEAR, &vmw_cmd_cid_check),
1079         VMW_CMD_DEF(SVGA_3D_CMD_PRESENT, &vmw_cmd_present_check),
1080         VMW_CMD_DEF(SVGA_3D_CMD_SHADER_DEFINE, &vmw_cmd_cid_check),
1081         VMW_CMD_DEF(SVGA_3D_CMD_SHADER_DESTROY, &vmw_cmd_cid_check),
1082         VMW_CMD_DEF(SVGA_3D_CMD_SET_SHADER, &vmw_cmd_set_shader),
1083         VMW_CMD_DEF(SVGA_3D_CMD_SET_SHADER_CONST, &vmw_cmd_cid_check),
1084         VMW_CMD_DEF(SVGA_3D_CMD_DRAW_PRIMITIVES, &vmw_cmd_draw),
1085         VMW_CMD_DEF(SVGA_3D_CMD_SETSCISSORRECT, &vmw_cmd_cid_check),
1086         VMW_CMD_DEF(SVGA_3D_CMD_BEGIN_QUERY, &vmw_cmd_begin_query),
1087         VMW_CMD_DEF(SVGA_3D_CMD_END_QUERY, &vmw_cmd_end_query),
1088         VMW_CMD_DEF(SVGA_3D_CMD_WAIT_FOR_QUERY, &vmw_cmd_wait_query),
1089         VMW_CMD_DEF(SVGA_3D_CMD_PRESENT_READBACK, &vmw_cmd_ok),
1090         VMW_CMD_DEF(SVGA_3D_CMD_BLIT_SURFACE_TO_SCREEN,
1091                     &vmw_cmd_blt_surf_screen_check),
1092         VMW_CMD_DEF(SVGA_3D_CMD_SURFACE_DEFINE_V2, &vmw_cmd_invalid),
1093         VMW_CMD_DEF(SVGA_3D_CMD_GENERATE_MIPMAPS, &vmw_cmd_invalid),
1094         VMW_CMD_DEF(SVGA_3D_CMD_ACTIVATE_SURFACE, &vmw_cmd_invalid),
1095         VMW_CMD_DEF(SVGA_3D_CMD_DEACTIVATE_SURFACE, &vmw_cmd_invalid),
1096 };
1097
1098 static int vmw_cmd_check(struct vmw_private *dev_priv,
1099                          struct vmw_sw_context *sw_context,
1100                          void *buf, uint32_t *size)
1101 {
1102         uint32_t cmd_id;
1103         uint32_t size_remaining = *size;
1104         SVGA3dCmdHeader *header = (SVGA3dCmdHeader *) buf;
1105         int ret;
1106
1107         cmd_id = le32_to_cpu(((uint32_t *)buf)[0]);
1108         /* Handle any none 3D commands */
1109         if (unlikely(cmd_id < SVGA_CMD_MAX))
1110                 return vmw_cmd_check_not_3d(dev_priv, sw_context, buf, size);
1111
1112
1113         cmd_id = le32_to_cpu(header->id);
1114         *size = le32_to_cpu(header->size) + sizeof(SVGA3dCmdHeader);
1115
1116         cmd_id -= SVGA_3D_CMD_BASE;
1117         if (unlikely(*size > size_remaining))
1118                 goto out_err;
1119
1120         if (unlikely(cmd_id >= SVGA_3D_CMD_MAX - SVGA_3D_CMD_BASE))
1121                 goto out_err;
1122
1123         ret = vmw_cmd_funcs[cmd_id](dev_priv, sw_context, header);
1124         if (unlikely(ret != 0))
1125                 goto out_err;
1126
1127         return 0;
1128 out_err:
1129         DRM_ERROR("Illegal / Invalid SVGA3D command: %d\n",
1130                   cmd_id + SVGA_3D_CMD_BASE);
1131         return -EINVAL;
1132 }
1133
1134 static int vmw_cmd_check_all(struct vmw_private *dev_priv,
1135                              struct vmw_sw_context *sw_context,
1136                              void *buf,
1137                              uint32_t size)
1138 {
1139         int32_t cur_size = size;
1140         int ret;
1141
1142         sw_context->buf_start = buf;
1143
1144         while (cur_size > 0) {
1145                 size = cur_size;
1146                 ret = vmw_cmd_check(dev_priv, sw_context, buf, &size);
1147                 if (unlikely(ret != 0))
1148                         return ret;
1149                 buf = (void *)((unsigned long) buf + size);
1150                 cur_size -= size;
1151         }
1152
1153         if (unlikely(cur_size != 0)) {
1154                 DRM_ERROR("Command verifier out of sync.\n");
1155                 return -EINVAL;
1156         }
1157
1158         return 0;
1159 }
1160
1161 static void vmw_free_relocations(struct vmw_sw_context *sw_context)
1162 {
1163         sw_context->cur_reloc = 0;
1164 }
1165
1166 static void vmw_apply_relocations(struct vmw_sw_context *sw_context)
1167 {
1168         uint32_t i;
1169         struct vmw_relocation *reloc;
1170         struct ttm_validate_buffer *validate;
1171         struct ttm_buffer_object *bo;
1172
1173         for (i = 0; i < sw_context->cur_reloc; ++i) {
1174                 reloc = &sw_context->relocs[i];
1175                 validate = &sw_context->val_bufs[reloc->index].base;
1176                 bo = validate->bo;
1177                 switch (bo->mem.mem_type) {
1178                 case TTM_PL_VRAM:
1179                         reloc->location->offset += bo->offset;
1180                         reloc->location->gmrId = SVGA_GMR_FRAMEBUFFER;
1181                         break;
1182                 case VMW_PL_GMR:
1183                         reloc->location->gmrId = bo->mem.start;
1184                         break;
1185                 default:
1186                         BUG();
1187                 }
1188         }
1189         vmw_free_relocations(sw_context);
1190 }
1191
1192 /**
1193  * vmw_resource_list_unrefererence - Free up a resource list and unreference
1194  * all resources referenced by it.
1195  *
1196  * @list: The resource list.
1197  */
1198 static void vmw_resource_list_unreference(struct list_head *list)
1199 {
1200         struct vmw_resource_val_node *val, *val_next;
1201
1202         /*
1203          * Drop references to resources held during command submission.
1204          */
1205
1206         list_for_each_entry_safe(val, val_next, list, head) {
1207                 list_del_init(&val->head);
1208                 vmw_resource_unreference(&val->res);
1209                 kfree(val);
1210         }
1211 }
1212
1213 static void vmw_clear_validations(struct vmw_sw_context *sw_context)
1214 {
1215         struct vmw_validate_buffer *entry, *next;
1216         struct vmw_resource_val_node *val;
1217
1218         /*
1219          * Drop references to DMA buffers held during command submission.
1220          */
1221         list_for_each_entry_safe(entry, next, &sw_context->validate_nodes,
1222                                  base.head) {
1223                 list_del(&entry->base.head);
1224                 ttm_bo_unref(&entry->base.bo);
1225                 (void) drm_ht_remove_item(&sw_context->res_ht, &entry->hash);
1226                 sw_context->cur_val_buf--;
1227         }
1228         BUG_ON(sw_context->cur_val_buf != 0);
1229
1230         list_for_each_entry(val, &sw_context->resource_list, head)
1231                 (void) drm_ht_remove_item(&sw_context->res_ht, &val->hash);
1232 }
1233
1234 static int vmw_validate_single_buffer(struct vmw_private *dev_priv,
1235                                       struct ttm_buffer_object *bo,
1236                                       bool validate_as_mob)
1237 {
1238         int ret;
1239
1240
1241         /*
1242          * Don't validate pinned buffers.
1243          */
1244
1245         if (bo == dev_priv->pinned_bo ||
1246             (bo == dev_priv->dummy_query_bo &&
1247              dev_priv->dummy_query_bo_pinned))
1248                 return 0;
1249
1250         if (validate_as_mob)
1251                 return ttm_bo_validate(bo, &vmw_mob_placement, true, false);
1252
1253         /**
1254          * Put BO in VRAM if there is space, otherwise as a GMR.
1255          * If there is no space in VRAM and GMR ids are all used up,
1256          * start evicting GMRs to make room. If the DMA buffer can't be
1257          * used as a GMR, this will return -ENOMEM.
1258          */
1259
1260         ret = ttm_bo_validate(bo, &vmw_vram_gmr_placement, true, false);
1261         if (likely(ret == 0 || ret == -ERESTARTSYS))
1262                 return ret;
1263
1264         /**
1265          * If that failed, try VRAM again, this time evicting
1266          * previous contents.
1267          */
1268
1269         DRM_INFO("Falling through to VRAM.\n");
1270         ret = ttm_bo_validate(bo, &vmw_vram_placement, true, false);
1271         return ret;
1272 }
1273
1274 static int vmw_validate_buffers(struct vmw_private *dev_priv,
1275                                 struct vmw_sw_context *sw_context)
1276 {
1277         struct vmw_validate_buffer *entry;
1278         int ret;
1279
1280         list_for_each_entry(entry, &sw_context->validate_nodes, base.head) {
1281                 ret = vmw_validate_single_buffer(dev_priv, entry->base.bo,
1282                                                  entry->validate_as_mob);
1283                 if (unlikely(ret != 0))
1284                         return ret;
1285         }
1286         return 0;
1287 }
1288
1289 static int vmw_resize_cmd_bounce(struct vmw_sw_context *sw_context,
1290                                  uint32_t size)
1291 {
1292         if (likely(sw_context->cmd_bounce_size >= size))
1293                 return 0;
1294
1295         if (sw_context->cmd_bounce_size == 0)
1296                 sw_context->cmd_bounce_size = VMWGFX_CMD_BOUNCE_INIT_SIZE;
1297
1298         while (sw_context->cmd_bounce_size < size) {
1299                 sw_context->cmd_bounce_size =
1300                         PAGE_ALIGN(sw_context->cmd_bounce_size +
1301                                    (sw_context->cmd_bounce_size >> 1));
1302         }
1303
1304         if (sw_context->cmd_bounce != NULL)
1305                 vfree(sw_context->cmd_bounce);
1306
1307         sw_context->cmd_bounce = vmalloc(sw_context->cmd_bounce_size);
1308
1309         if (sw_context->cmd_bounce == NULL) {
1310                 DRM_ERROR("Failed to allocate command bounce buffer.\n");
1311                 sw_context->cmd_bounce_size = 0;
1312                 return -ENOMEM;
1313         }
1314
1315         return 0;
1316 }
1317
1318 /**
1319  * vmw_execbuf_fence_commands - create and submit a command stream fence
1320  *
1321  * Creates a fence object and submits a command stream marker.
1322  * If this fails for some reason, We sync the fifo and return NULL.
1323  * It is then safe to fence buffers with a NULL pointer.
1324  *
1325  * If @p_handle is not NULL @file_priv must also not be NULL. Creates
1326  * a userspace handle if @p_handle is not NULL, otherwise not.
1327  */
1328
1329 int vmw_execbuf_fence_commands(struct drm_file *file_priv,
1330                                struct vmw_private *dev_priv,
1331                                struct vmw_fence_obj **p_fence,
1332                                uint32_t *p_handle)
1333 {
1334         uint32_t sequence;
1335         int ret;
1336         bool synced = false;
1337
1338         /* p_handle implies file_priv. */
1339         BUG_ON(p_handle != NULL && file_priv == NULL);
1340
1341         ret = vmw_fifo_send_fence(dev_priv, &sequence);
1342         if (unlikely(ret != 0)) {
1343                 DRM_ERROR("Fence submission error. Syncing.\n");
1344                 synced = true;
1345         }
1346
1347         if (p_handle != NULL)
1348                 ret = vmw_user_fence_create(file_priv, dev_priv->fman,
1349                                             sequence,
1350                                             DRM_VMW_FENCE_FLAG_EXEC,
1351                                             p_fence, p_handle);
1352         else
1353                 ret = vmw_fence_create(dev_priv->fman, sequence,
1354                                        DRM_VMW_FENCE_FLAG_EXEC,
1355                                        p_fence);
1356
1357         if (unlikely(ret != 0 && !synced)) {
1358                 (void) vmw_fallback_wait(dev_priv, false, false,
1359                                          sequence, false,
1360                                          VMW_FENCE_WAIT_TIMEOUT);
1361                 *p_fence = NULL;
1362         }
1363
1364         return 0;
1365 }
1366
1367 /**
1368  * vmw_execbuf_copy_fence_user - copy fence object information to
1369  * user-space.
1370  *
1371  * @dev_priv: Pointer to a vmw_private struct.
1372  * @vmw_fp: Pointer to the struct vmw_fpriv representing the calling file.
1373  * @ret: Return value from fence object creation.
1374  * @user_fence_rep: User space address of a struct drm_vmw_fence_rep to
1375  * which the information should be copied.
1376  * @fence: Pointer to the fenc object.
1377  * @fence_handle: User-space fence handle.
1378  *
1379  * This function copies fence information to user-space. If copying fails,
1380  * The user-space struct drm_vmw_fence_rep::error member is hopefully
1381  * left untouched, and if it's preloaded with an -EFAULT by user-space,
1382  * the error will hopefully be detected.
1383  * Also if copying fails, user-space will be unable to signal the fence
1384  * object so we wait for it immediately, and then unreference the
1385  * user-space reference.
1386  */
1387 void
1388 vmw_execbuf_copy_fence_user(struct vmw_private *dev_priv,
1389                             struct vmw_fpriv *vmw_fp,
1390                             int ret,
1391                             struct drm_vmw_fence_rep __user *user_fence_rep,
1392                             struct vmw_fence_obj *fence,
1393                             uint32_t fence_handle)
1394 {
1395         struct drm_vmw_fence_rep fence_rep;
1396
1397         if (user_fence_rep == NULL)
1398                 return;
1399
1400         memset(&fence_rep, 0, sizeof(fence_rep));
1401
1402         fence_rep.error = ret;
1403         if (ret == 0) {
1404                 BUG_ON(fence == NULL);
1405
1406                 fence_rep.handle = fence_handle;
1407                 fence_rep.seqno = fence->seqno;
1408                 vmw_update_seqno(dev_priv, &dev_priv->fifo);
1409                 fence_rep.passed_seqno = dev_priv->last_read_seqno;
1410         }
1411
1412         /*
1413          * copy_to_user errors will be detected by user space not
1414          * seeing fence_rep::error filled in. Typically
1415          * user-space would have pre-set that member to -EFAULT.
1416          */
1417         ret = copy_to_user(user_fence_rep, &fence_rep,
1418                            sizeof(fence_rep));
1419
1420         /*
1421          * User-space lost the fence object. We need to sync
1422          * and unreference the handle.
1423          */
1424         if (unlikely(ret != 0) && (fence_rep.error == 0)) {
1425                 ttm_ref_object_base_unref(vmw_fp->tfile,
1426                                           fence_handle, TTM_REF_USAGE);
1427                 DRM_ERROR("Fence copy error. Syncing.\n");
1428                 (void) vmw_fence_obj_wait(fence, fence->signal_mask,
1429                                           false, false,
1430                                           VMW_FENCE_WAIT_TIMEOUT);
1431         }
1432 }
1433
1434 int vmw_execbuf_process(struct drm_file *file_priv,
1435                         struct vmw_private *dev_priv,
1436                         void __user *user_commands,
1437                         void *kernel_commands,
1438                         uint32_t command_size,
1439                         uint64_t throttle_us,
1440                         struct drm_vmw_fence_rep __user *user_fence_rep,
1441                         struct vmw_fence_obj **out_fence)
1442 {
1443         struct vmw_sw_context *sw_context = &dev_priv->ctx;
1444         struct vmw_fence_obj *fence = NULL;
1445         struct vmw_resource *error_resource;
1446         struct list_head resource_list;
1447         struct ww_acquire_ctx ticket;
1448         uint32_t handle;
1449         void *cmd;
1450         int ret;
1451
1452         ret = mutex_lock_interruptible(&dev_priv->cmdbuf_mutex);
1453         if (unlikely(ret != 0))
1454                 return -ERESTARTSYS;
1455
1456         if (kernel_commands == NULL) {
1457                 sw_context->kernel = false;
1458
1459                 ret = vmw_resize_cmd_bounce(sw_context, command_size);
1460                 if (unlikely(ret != 0))
1461                         goto out_unlock;
1462
1463
1464                 ret = copy_from_user(sw_context->cmd_bounce,
1465                                      user_commands, command_size);
1466
1467                 if (unlikely(ret != 0)) {
1468                         ret = -EFAULT;
1469                         DRM_ERROR("Failed copying commands.\n");
1470                         goto out_unlock;
1471                 }
1472                 kernel_commands = sw_context->cmd_bounce;
1473         } else
1474                 sw_context->kernel = true;
1475
1476         sw_context->tfile = vmw_fpriv(file_priv)->tfile;
1477         sw_context->cur_reloc = 0;
1478         sw_context->cur_val_buf = 0;
1479         sw_context->fence_flags = 0;
1480         INIT_LIST_HEAD(&sw_context->resource_list);
1481         sw_context->cur_query_bo = dev_priv->pinned_bo;
1482         sw_context->last_query_ctx = NULL;
1483         sw_context->needs_post_query_barrier = false;
1484         memset(sw_context->res_cache, 0, sizeof(sw_context->res_cache));
1485         INIT_LIST_HEAD(&sw_context->validate_nodes);
1486         INIT_LIST_HEAD(&sw_context->res_relocations);
1487         if (!sw_context->res_ht_initialized) {
1488                 ret = drm_ht_create(&sw_context->res_ht, VMW_RES_HT_ORDER);
1489                 if (unlikely(ret != 0))
1490                         goto out_unlock;
1491                 sw_context->res_ht_initialized = true;
1492         }
1493
1494         INIT_LIST_HEAD(&resource_list);
1495         ret = vmw_cmd_check_all(dev_priv, sw_context, kernel_commands,
1496                                 command_size);
1497         if (unlikely(ret != 0))
1498                 goto out_err;
1499
1500         ret = vmw_resources_reserve(sw_context);
1501         if (unlikely(ret != 0))
1502                 goto out_err;
1503
1504         ret = ttm_eu_reserve_buffers(&ticket, &sw_context->validate_nodes);
1505         if (unlikely(ret != 0))
1506                 goto out_err;
1507
1508         ret = vmw_validate_buffers(dev_priv, sw_context);
1509         if (unlikely(ret != 0))
1510                 goto out_err;
1511
1512         ret = vmw_resources_validate(sw_context);
1513         if (unlikely(ret != 0))
1514                 goto out_err;
1515
1516         if (throttle_us) {
1517                 ret = vmw_wait_lag(dev_priv, &dev_priv->fifo.marker_queue,
1518                                    throttle_us);
1519
1520                 if (unlikely(ret != 0))
1521                         goto out_err;
1522         }
1523
1524         cmd = vmw_fifo_reserve(dev_priv, command_size);
1525         if (unlikely(cmd == NULL)) {
1526                 DRM_ERROR("Failed reserving fifo space for commands.\n");
1527                 ret = -ENOMEM;
1528                 goto out_err;
1529         }
1530
1531         vmw_apply_relocations(sw_context);
1532         memcpy(cmd, kernel_commands, command_size);
1533
1534         vmw_resource_relocations_apply(cmd, &sw_context->res_relocations);
1535         vmw_resource_relocations_free(&sw_context->res_relocations);
1536
1537         vmw_fifo_commit(dev_priv, command_size);
1538
1539         vmw_query_bo_switch_commit(dev_priv, sw_context);
1540         ret = vmw_execbuf_fence_commands(file_priv, dev_priv,
1541                                          &fence,
1542                                          (user_fence_rep) ? &handle : NULL);
1543         /*
1544          * This error is harmless, because if fence submission fails,
1545          * vmw_fifo_send_fence will sync. The error will be propagated to
1546          * user-space in @fence_rep
1547          */
1548
1549         if (ret != 0)
1550                 DRM_ERROR("Fence submission error. Syncing.\n");
1551
1552         vmw_resource_list_unreserve(&sw_context->resource_list, false);
1553         ttm_eu_fence_buffer_objects(&ticket, &sw_context->validate_nodes,
1554                                     (void *) fence);
1555
1556         if (unlikely(dev_priv->pinned_bo != NULL &&
1557                      !dev_priv->query_cid_valid))
1558                 __vmw_execbuf_release_pinned_bo(dev_priv, fence);
1559
1560         vmw_clear_validations(sw_context);
1561         vmw_execbuf_copy_fence_user(dev_priv, vmw_fpriv(file_priv), ret,
1562                                     user_fence_rep, fence, handle);
1563
1564         /* Don't unreference when handing fence out */
1565         if (unlikely(out_fence != NULL)) {
1566                 *out_fence = fence;
1567                 fence = NULL;
1568         } else if (likely(fence != NULL)) {
1569                 vmw_fence_obj_unreference(&fence);
1570         }
1571
1572         list_splice_init(&sw_context->resource_list, &resource_list);
1573         mutex_unlock(&dev_priv->cmdbuf_mutex);
1574
1575         /*
1576          * Unreference resources outside of the cmdbuf_mutex to
1577          * avoid deadlocks in resource destruction paths.
1578          */
1579         vmw_resource_list_unreference(&resource_list);
1580
1581         return 0;
1582
1583 out_err:
1584         vmw_resource_relocations_free(&sw_context->res_relocations);
1585         vmw_free_relocations(sw_context);
1586         ttm_eu_backoff_reservation(&ticket, &sw_context->validate_nodes);
1587         vmw_resource_list_unreserve(&sw_context->resource_list, true);
1588         vmw_clear_validations(sw_context);
1589         if (unlikely(dev_priv->pinned_bo != NULL &&
1590                      !dev_priv->query_cid_valid))
1591                 __vmw_execbuf_release_pinned_bo(dev_priv, NULL);
1592 out_unlock:
1593         list_splice_init(&sw_context->resource_list, &resource_list);
1594         error_resource = sw_context->error_resource;
1595         sw_context->error_resource = NULL;
1596         mutex_unlock(&dev_priv->cmdbuf_mutex);
1597
1598         /*
1599          * Unreference resources outside of the cmdbuf_mutex to
1600          * avoid deadlocks in resource destruction paths.
1601          */
1602         vmw_resource_list_unreference(&resource_list);
1603         if (unlikely(error_resource != NULL))
1604                 vmw_resource_unreference(&error_resource);
1605
1606         return ret;
1607 }
1608
1609 /**
1610  * vmw_execbuf_unpin_panic - Idle the fifo and unpin the query buffer.
1611  *
1612  * @dev_priv: The device private structure.
1613  *
1614  * This function is called to idle the fifo and unpin the query buffer
1615  * if the normal way to do this hits an error, which should typically be
1616  * extremely rare.
1617  */
1618 static void vmw_execbuf_unpin_panic(struct vmw_private *dev_priv)
1619 {
1620         DRM_ERROR("Can't unpin query buffer. Trying to recover.\n");
1621
1622         (void) vmw_fallback_wait(dev_priv, false, true, 0, false, 10*HZ);
1623         vmw_bo_pin(dev_priv->pinned_bo, false);
1624         vmw_bo_pin(dev_priv->dummy_query_bo, false);
1625         dev_priv->dummy_query_bo_pinned = false;
1626 }
1627
1628
1629 /**
1630  * __vmw_execbuf_release_pinned_bo - Flush queries and unpin the pinned
1631  * query bo.
1632  *
1633  * @dev_priv: The device private structure.
1634  * @fence: If non-NULL should point to a struct vmw_fence_obj issued
1635  * _after_ a query barrier that flushes all queries touching the current
1636  * buffer pointed to by @dev_priv->pinned_bo
1637  *
1638  * This function should be used to unpin the pinned query bo, or
1639  * as a query barrier when we need to make sure that all queries have
1640  * finished before the next fifo command. (For example on hardware
1641  * context destructions where the hardware may otherwise leak unfinished
1642  * queries).
1643  *
1644  * This function does not return any failure codes, but make attempts
1645  * to do safe unpinning in case of errors.
1646  *
1647  * The function will synchronize on the previous query barrier, and will
1648  * thus not finish until that barrier has executed.
1649  *
1650  * the @dev_priv->cmdbuf_mutex needs to be held by the current thread
1651  * before calling this function.
1652  */
1653 void __vmw_execbuf_release_pinned_bo(struct vmw_private *dev_priv,
1654                                      struct vmw_fence_obj *fence)
1655 {
1656         int ret = 0;
1657         struct list_head validate_list;
1658         struct ttm_validate_buffer pinned_val, query_val;
1659         struct vmw_fence_obj *lfence = NULL;
1660         struct ww_acquire_ctx ticket;
1661
1662         if (dev_priv->pinned_bo == NULL)
1663                 goto out_unlock;
1664
1665         INIT_LIST_HEAD(&validate_list);
1666
1667         pinned_val.bo = ttm_bo_reference(dev_priv->pinned_bo);
1668         list_add_tail(&pinned_val.head, &validate_list);
1669
1670         query_val.bo = ttm_bo_reference(dev_priv->dummy_query_bo);
1671         list_add_tail(&query_val.head, &validate_list);
1672
1673         do {
1674                 ret = ttm_eu_reserve_buffers(&ticket, &validate_list);
1675         } while (ret == -ERESTARTSYS);
1676
1677         if (unlikely(ret != 0)) {
1678                 vmw_execbuf_unpin_panic(dev_priv);
1679                 goto out_no_reserve;
1680         }
1681
1682         if (dev_priv->query_cid_valid) {
1683                 BUG_ON(fence != NULL);
1684                 ret = vmw_fifo_emit_dummy_query(dev_priv, dev_priv->query_cid);
1685                 if (unlikely(ret != 0)) {
1686                         vmw_execbuf_unpin_panic(dev_priv);
1687                         goto out_no_emit;
1688                 }
1689                 dev_priv->query_cid_valid = false;
1690         }
1691
1692         vmw_bo_pin(dev_priv->pinned_bo, false);
1693         vmw_bo_pin(dev_priv->dummy_query_bo, false);
1694         dev_priv->dummy_query_bo_pinned = false;
1695
1696         if (fence == NULL) {
1697                 (void) vmw_execbuf_fence_commands(NULL, dev_priv, &lfence,
1698                                                   NULL);
1699                 fence = lfence;
1700         }
1701         ttm_eu_fence_buffer_objects(&ticket, &validate_list, (void *) fence);
1702         if (lfence != NULL)
1703                 vmw_fence_obj_unreference(&lfence);
1704
1705         ttm_bo_unref(&query_val.bo);
1706         ttm_bo_unref(&pinned_val.bo);
1707         ttm_bo_unref(&dev_priv->pinned_bo);
1708
1709 out_unlock:
1710         return;
1711
1712 out_no_emit:
1713         ttm_eu_backoff_reservation(&ticket, &validate_list);
1714 out_no_reserve:
1715         ttm_bo_unref(&query_val.bo);
1716         ttm_bo_unref(&pinned_val.bo);
1717         ttm_bo_unref(&dev_priv->pinned_bo);
1718 }
1719
1720 /**
1721  * vmw_execbuf_release_pinned_bo - Flush queries and unpin the pinned
1722  * query bo.
1723  *
1724  * @dev_priv: The device private structure.
1725  *
1726  * This function should be used to unpin the pinned query bo, or
1727  * as a query barrier when we need to make sure that all queries have
1728  * finished before the next fifo command. (For example on hardware
1729  * context destructions where the hardware may otherwise leak unfinished
1730  * queries).
1731  *
1732  * This function does not return any failure codes, but make attempts
1733  * to do safe unpinning in case of errors.
1734  *
1735  * The function will synchronize on the previous query barrier, and will
1736  * thus not finish until that barrier has executed.
1737  */
1738 void vmw_execbuf_release_pinned_bo(struct vmw_private *dev_priv)
1739 {
1740         mutex_lock(&dev_priv->cmdbuf_mutex);
1741         if (dev_priv->query_cid_valid)
1742                 __vmw_execbuf_release_pinned_bo(dev_priv, NULL);
1743         mutex_unlock(&dev_priv->cmdbuf_mutex);
1744 }
1745
1746
1747 int vmw_execbuf_ioctl(struct drm_device *dev, void *data,
1748                       struct drm_file *file_priv)
1749 {
1750         struct vmw_private *dev_priv = vmw_priv(dev);
1751         struct drm_vmw_execbuf_arg *arg = (struct drm_vmw_execbuf_arg *)data;
1752         struct vmw_master *vmaster = vmw_master(file_priv->master);
1753         int ret;
1754
1755         /*
1756          * This will allow us to extend the ioctl argument while
1757          * maintaining backwards compatibility:
1758          * We take different code paths depending on the value of
1759          * arg->version.
1760          */
1761
1762         if (unlikely(arg->version != DRM_VMW_EXECBUF_VERSION)) {
1763                 DRM_ERROR("Incorrect execbuf version.\n");
1764                 DRM_ERROR("You're running outdated experimental "
1765                           "vmwgfx user-space drivers.");
1766                 return -EINVAL;
1767         }
1768
1769         ret = ttm_read_lock(&vmaster->lock, true);
1770         if (unlikely(ret != 0))
1771                 return ret;
1772
1773         ret = vmw_execbuf_process(file_priv, dev_priv,
1774                                   (void __user *)(unsigned long)arg->commands,
1775                                   NULL, arg->command_size, arg->throttle_us,
1776                                   (void __user *)(unsigned long)arg->fence_rep,
1777                                   NULL);
1778
1779         if (unlikely(ret != 0))
1780                 goto out_unlock;
1781
1782         vmw_kms_cursor_post_execbuf(dev_priv);
1783
1784 out_unlock:
1785         ttm_read_unlock(&vmaster->lock);
1786         return ret;
1787 }