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  * @p_val_node: If non-NULL Will be updated with the validate node number
 228  * on return.
 229  *
 230  * Returns -EINVAL if the limit of number of buffer objects per command
 231  * submission is reached.
 232  */
 233 static int vmw_bo_to_validate_list(struct vmw_sw_context *sw_context,
 234                                    struct ttm_buffer_object *bo,
 235                                    uint32_t *p_val_node)
 236 {
 237         uint32_t val_node;
 238         struct vmw_validate_buffer *vval_buf;
 239         struct ttm_validate_buffer *val_buf;
 240         struct drm_hash_item *hash;
 241         int ret;
 242
 243         if (likely(drm_ht_find_item(&sw_context->res_ht, (unsigned long) bo,
 244                                     &hash) == 0)) {
 245                 vval_buf = container_of(hash, struct vmw_validate_buffer,
 246                                         hash);
 247                 val_buf = &vval_buf->base;
 248                 val_node = vval_buf - sw_context->val_bufs;
 249         } else {
 250                 val_node = sw_context->cur_val_buf;
 251                 if (unlikely(val_node >= VMWGFX_MAX_VALIDATIONS)) {
 252                         DRM_ERROR("Max number of DMA buffers per submission "
 253                                   "exceeded.\n");
 254                         return -EINVAL;
 255                 }
 256                 vval_buf = &sw_context->val_bufs[val_node];
 257                 vval_buf->hash.key = (unsigned long) bo;
 258                 ret = drm_ht_insert_item(&sw_context->res_ht, &vval_buf->hash);
 259                 if (unlikely(ret != 0)) {
 260                         DRM_ERROR("Failed to initialize a buffer validation "
 261                                   "entry.\n");
 262                         return ret;
 263                 }
 264                 ++sw_context->cur_val_buf;
 265                 val_buf = &vval_buf->base;
 266                 val_buf->bo = ttm_bo_reference(bo);
 267                 val_buf->reserved = false;
 268                 list_add_tail(&val_buf->head, &sw_context->validate_nodes);
 269         }
 270
 271         sw_context->fence_flags |= DRM_VMW_FENCE_FLAG_EXEC;
 272
 273         if (p_val_node)
 274                 *p_val_node = val_node;
 275
 276         return 0;
 277 }
 278
 279 /**
 280  * vmw_resources_reserve - Reserve all resources on the sw_context's
 281  * resource list.
 282  *
 283  * @sw_context: Pointer to the software context.
 284  *
 285  * Note that since vmware's command submission currently is protected by
 286  * the cmdbuf mutex, no fancy deadlock avoidance is required for resources,
 287  * since only a single thread at once will attempt this.
 288  */
 289 static int vmw_resources_reserve(struct vmw_sw_context *sw_context)
 290 {
 291         struct vmw_resource_val_node *val;
 292         int ret;
 293
 294         list_for_each_entry(val, &sw_context->resource_list, head) {
 295                 struct vmw_resource *res = val->res;
 296
 297                 ret = vmw_resource_reserve(res, val->no_buffer_needed);
 298                 if (unlikely(ret != 0))
 299                         return ret;
 300
 301                 if (res->backup) {
 302                         struct ttm_buffer_object *bo = &res->backup->base;
 303
 304                         ret = vmw_bo_to_validate_list
 305                                 (sw_context, bo, NULL);
 306
 307                         if (unlikely(ret != 0))
 308                                 return ret;
 309                 }
 310         }
 311         return 0;
 312 }
 313
 314 /**
 315  * vmw_resources_validate - Validate all resources on the sw_context's
 316  * resource list.
 317  *
 318  * @sw_context: Pointer to the software context.
 319  *
 320  * Before this function is called, all resource backup buffers must have
 321  * been validated.
 322  */
 323 static int vmw_resources_validate(struct vmw_sw_context *sw_context)
 324 {
 325         struct vmw_resource_val_node *val;
 326         int ret;
 327
 328         list_for_each_entry(val, &sw_context->resource_list, head) {
 329                 struct vmw_resource *res = val->res;
 330
 331                 ret = vmw_resource_validate(res);
 332                 if (unlikely(ret != 0)) {
 333                         if (ret != -ERESTARTSYS)
 334                                 DRM_ERROR("Failed to validate resource.\n");
 335                         return ret;
 336                 }
 337         }
 338         return 0;
 339 }
 340
 341 /**
 342  * vmw_cmd_res_check - Check that a resource is present and if so, put it
 343  * on the resource validate list unless it's already there.
 344  *
 345  * @dev_priv: Pointer to a device private structure.
 346  * @sw_context: Pointer to the software context.
 347  * @res_type: Resource type.
 348  * @converter: User-space visisble type specific information.
 349  * @id: Pointer to the location in the command buffer currently being
 350  * parsed from where the user-space resource id handle is located.
 351  */
 352 static int vmw_cmd_res_check(struct vmw_private *dev_priv,
 353                              struct vmw_sw_context *sw_context,
 354                              enum vmw_res_type res_type,
 355                              const struct vmw_user_resource_conv *converter,
 356                              uint32_t *id,
 357                              struct vmw_resource_val_node **p_val)
 358 {
 359         struct vmw_res_cache_entry *rcache =
 360                 &sw_context->res_cache[res_type];
 361         struct vmw_resource *res;
 362         struct vmw_resource_val_node *node;
 363         int ret;
 364
 365         if (*id == SVGA3D_INVALID_ID)
 366                 return 0;
 367
 368         /*
 369          * Fastpath in case of repeated commands referencing the same
 370          * resource
 371          */
 372
 373         if (likely(rcache->valid && *id == rcache->handle)) {
 374                 const struct vmw_resource *res = rcache->res;
 375
 376                 rcache->node->first_usage = false;
 377                 if (p_val)
 378                         *p_val = rcache->node;
 379
 380                 return vmw_resource_relocation_add
 381                         (&sw_context->res_relocations, res,
 382                          id - sw_context->buf_start);
 383         }
 384
 385         ret = vmw_user_resource_lookup_handle(dev_priv,
 386                                               sw_context->tfile,
 387                                               *id,
 388                                               converter,
 389                                               &res);
 390         if (unlikely(ret != 0)) {
 391                 DRM_ERROR("Could not find or use resource 0x%08x.\n",
 392                           (unsigned) *id);
 393                 dump_stack();
 394                 return ret;
 395         }
 396
 397         rcache->valid = true;
 398         rcache->res = res;
 399         rcache->handle = *id;
 400
 401         ret = vmw_resource_relocation_add(&sw_context->res_relocations,
 402                                           res,
 403                                           id - sw_context->buf_start);
 404         if (unlikely(ret != 0))
 405                 goto out_no_reloc;
 406
 407         ret = vmw_resource_val_add(sw_context, res, &node);
 408         if (unlikely(ret != 0))
 409                 goto out_no_reloc;
 410
 411         rcache->node = node;
 412         if (p_val)
 413                 *p_val = node;
 414         vmw_resource_unreference(&res);
 415         return 0;
 416
 417 out_no_reloc:
 418         BUG_ON(sw_context->error_resource != NULL);
 419         sw_context->error_resource = res;
 420
 421         return ret;
 422 }
 423
 424 /**
 425  * vmw_cmd_cid_check - Check a command header for valid context information.
 426  *
 427  * @dev_priv: Pointer to a device private structure.
 428  * @sw_context: Pointer to the software context.
 429  * @header: A command header with an embedded user-space context handle.
 430  *
 431  * Convenience function: Call vmw_cmd_res_check with the user-space context
 432  * handle embedded in @header.
 433  */
 434 static int vmw_cmd_cid_check(struct vmw_private *dev_priv,
 435                              struct vmw_sw_context *sw_context,
 436                              SVGA3dCmdHeader *header)
 437 {
 438         struct vmw_cid_cmd {
 439                 SVGA3dCmdHeader header;
 440                 __le32 cid;
 441         } *cmd;
 442
 443         cmd = container_of(header, struct vmw_cid_cmd, header);
 444         return vmw_cmd_res_check(dev_priv, sw_context, vmw_res_context,
 445                                  user_context_converter, &cmd->cid, NULL);
 446 }
 447
 448 static int vmw_cmd_set_render_target_check(struct vmw_private *dev_priv,
 449                                            struct vmw_sw_context *sw_context,
 450                                            SVGA3dCmdHeader *header)
 451 {
 452         struct vmw_sid_cmd {
 453                 SVGA3dCmdHeader header;
 454                 SVGA3dCmdSetRenderTarget body;
 455         } *cmd;
 456         int ret;
 457
 458         ret = vmw_cmd_cid_check(dev_priv, sw_context, header);
 459         if (unlikely(ret != 0))
 460                 return ret;
 461
 462         cmd = container_of(header, struct vmw_sid_cmd, header);
 463         ret = vmw_cmd_res_check(dev_priv, sw_context, vmw_res_surface,
 464                                 user_surface_converter,
 465                                 &cmd->body.target.sid, NULL);
 466         return ret;
 467 }
 468
 469 static int vmw_cmd_surface_copy_check(struct vmw_private *dev_priv,
 470                                       struct vmw_sw_context *sw_context,
 471                                       SVGA3dCmdHeader *header)
 472 {
 473         struct vmw_sid_cmd {
 474                 SVGA3dCmdHeader header;
 475                 SVGA3dCmdSurfaceCopy body;
 476         } *cmd;
 477         int ret;
 478
 479         cmd = container_of(header, struct vmw_sid_cmd, header);
 480         ret = vmw_cmd_res_check(dev_priv, sw_context, vmw_res_surface,
 481                                 user_surface_converter,
 482                                 &cmd->body.src.sid, NULL);
 483         if (unlikely(ret != 0))
 484                 return ret;
 485         return vmw_cmd_res_check(dev_priv, sw_context, vmw_res_surface,
 486                                  user_surface_converter,
 487                                  &cmd->body.dest.sid, NULL);
 488 }
 489
 490 static int vmw_cmd_stretch_blt_check(struct vmw_private *dev_priv,
 491                                      struct vmw_sw_context *sw_context,
 492                                      SVGA3dCmdHeader *header)
 493 {
 494         struct vmw_sid_cmd {
 495                 SVGA3dCmdHeader header;
 496                 SVGA3dCmdSurfaceStretchBlt body;
 497         } *cmd;
 498         int ret;
 499
 500         cmd = container_of(header, struct vmw_sid_cmd, header);
 501         ret = vmw_cmd_res_check(dev_priv, sw_context, vmw_res_surface,
 502                                 user_surface_converter,
 503                                 &cmd->body.src.sid, NULL);
 504         if (unlikely(ret != 0))
 505                 return ret;
 506         return vmw_cmd_res_check(dev_priv, sw_context, vmw_res_surface,
 507                                  user_surface_converter,
 508                                  &cmd->body.dest.sid, NULL);
 509 }
 510
 511 static int vmw_cmd_blt_surf_screen_check(struct vmw_private *dev_priv,
 512                                          struct vmw_sw_context *sw_context,
 513                                          SVGA3dCmdHeader *header)
 514 {
 515         struct vmw_sid_cmd {
 516                 SVGA3dCmdHeader header;
 517                 SVGA3dCmdBlitSurfaceToScreen body;
 518         } *cmd;
 519
 520         cmd = container_of(header, struct vmw_sid_cmd, header);
 521
 522         if (unlikely(!sw_context->kernel)) {
 523                 DRM_ERROR("Kernel only SVGA3d command: %u.\n", cmd->header.id);
 524                 return -EPERM;
 525         }
 526
 527         return vmw_cmd_res_check(dev_priv, sw_context, vmw_res_surface,
 528                                  user_surface_converter,
 529                                  &cmd->body.srcImage.sid, NULL);
 530 }
 531
 532 static int vmw_cmd_present_check(struct vmw_private *dev_priv,
 533                                  struct vmw_sw_context *sw_context,
 534                                  SVGA3dCmdHeader *header)
 535 {
 536         struct vmw_sid_cmd {
 537                 SVGA3dCmdHeader header;
 538                 SVGA3dCmdPresent body;
 539         } *cmd;
 540
 541
 542         cmd = container_of(header, struct vmw_sid_cmd, header);
 543
 544         if (unlikely(!sw_context->kernel)) {
 545                 DRM_ERROR("Kernel only SVGA3d command: %u.\n", cmd->header.id);
 546                 return -EPERM;
 547         }
 548
 549         return vmw_cmd_res_check(dev_priv, sw_context, vmw_res_surface,
 550                                  user_surface_converter, &cmd->body.sid,
 551                                  NULL);
 552 }
 553
 554 /**
 555  * vmw_query_bo_switch_prepare - Prepare to switch pinned buffer for queries.
 556  *
 557  * @dev_priv: The device private structure.
 558  * @new_query_bo: The new buffer holding query results.
 559  * @sw_context: The software context used for this command submission.
 560  *
 561  * This function checks whether @new_query_bo is suitable for holding
 562  * query results, and if another buffer currently is pinned for query
 563  * results. If so, the function prepares the state of @sw_context for
 564  * switching pinned buffers after successful submission of the current
 565  * command batch.
 566  */
 567 static int vmw_query_bo_switch_prepare(struct vmw_private *dev_priv,
 568                                        struct ttm_buffer_object *new_query_bo,
 569                                        struct vmw_sw_context *sw_context)
 570 {
 571         struct vmw_res_cache_entry *ctx_entry =
 572                 &sw_context->res_cache[vmw_res_context];
 573         int ret;
 574
 575         BUG_ON(!ctx_entry->valid);
 576         sw_context->last_query_ctx = ctx_entry->res;
 577
 578         if (unlikely(new_query_bo != sw_context->cur_query_bo)) {
 579
 580                 if (unlikely(new_query_bo->num_pages > 4)) {
 581                         DRM_ERROR("Query buffer too large.\n");
 582                         return -EINVAL;
 583                 }
 584
 585                 if (unlikely(sw_context->cur_query_bo != NULL)) {
 586                         sw_context->needs_post_query_barrier = true;
 587                         ret = vmw_bo_to_validate_list(sw_context,
 588                                                       sw_context->cur_query_bo,
 589                                                       NULL);
 590                         if (unlikely(ret != 0))
 591                                 return ret;
 592                 }
 593                 sw_context->cur_query_bo = new_query_bo;
 594
 595                 ret = vmw_bo_to_validate_list(sw_context,
 596                                               dev_priv->dummy_query_bo,
 597                                               NULL);
 598                 if (unlikely(ret != 0))
 599                         return ret;
 600
 601         }
 602
 603         return 0;
 604 }
 605
 606
 607 /**
 608  * vmw_query_bo_switch_commit - Finalize switching pinned query buffer
 609  *
 610  * @dev_priv: The device private structure.
 611  * @sw_context: The software context used for this command submission batch.
 612  *
 613  * This function will check if we're switching query buffers, and will then,
 614  * issue a dummy occlusion query wait used as a query barrier. When the fence
 615  * object following that query wait has signaled, we are sure that all
 616  * preceding queries have finished, and the old query buffer can be unpinned.
 617  * However, since both the new query buffer and the old one are fenced with
 618  * that fence, we can do an asynchronus unpin now, and be sure that the
 619  * old query buffer won't be moved until the fence has signaled.
 620  *
 621  * As mentioned above, both the new - and old query buffers need to be fenced
 622  * using a sequence emitted *after* calling this function.
 623  */
 624 static void vmw_query_bo_switch_commit(struct vmw_private *dev_priv,
 625                                      struct vmw_sw_context *sw_context)
 626 {
 627         /*
 628          * The validate list should still hold references to all
 629          * contexts here.
 630          */
 631
 632         if (sw_context->needs_post_query_barrier) {
 633                 struct vmw_res_cache_entry *ctx_entry =
 634                         &sw_context->res_cache[vmw_res_context];
 635                 struct vmw_resource *ctx;
 636                 int ret;
 637
 638                 BUG_ON(!ctx_entry->valid);
 639                 ctx = ctx_entry->res;
 640
 641                 ret = vmw_fifo_emit_dummy_query(dev_priv, ctx->id);
 642
 643                 if (unlikely(ret != 0))
 644                         DRM_ERROR("Out of fifo space for dummy query.\n");
 645         }
 646
 647         if (dev_priv->pinned_bo != sw_context->cur_query_bo) {
 648                 if (dev_priv->pinned_bo) {
 649                         vmw_bo_pin(dev_priv->pinned_bo, false);
 650                         ttm_bo_unref(&dev_priv->pinned_bo);
 651                 }
 652
 653                 if (!sw_context->needs_post_query_barrier) {
 654                         vmw_bo_pin(sw_context->cur_query_bo, true);
 655
 656                         /*
 657                          * We pin also the dummy_query_bo buffer so that we
 658                          * don't need to validate it when emitting
 659                          * dummy queries in context destroy paths.
 660                          */
 661
 662                         vmw_bo_pin(dev_priv->dummy_query_bo, true);
 663                         dev_priv->dummy_query_bo_pinned = true;
 664
 665                         BUG_ON(sw_context->last_query_ctx == NULL);
 666                         dev_priv->query_cid = sw_context->last_query_ctx->id;
 667                         dev_priv->query_cid_valid = true;
 668                         dev_priv->pinned_bo =
 669                                 ttm_bo_reference(sw_context->cur_query_bo);
 670                 }
 671         }
 672 }
 673
 674 /**
 675  * vmw_translate_guest_pointer - Prepare to translate a user-space buffer
 676  * handle to a valid SVGAGuestPtr
 677  *
 678  * @dev_priv: Pointer to a device private structure.
 679  * @sw_context: The software context used for this command batch validation.
 680  * @ptr: Pointer to the user-space handle to be translated.
 681  * @vmw_bo_p: Points to a location that, on successful return will carry
 682  * a reference-counted pointer to the DMA buffer identified by the
 683  * user-space handle in @id.
 684  *
 685  * This function saves information needed to translate a user-space buffer
 686  * handle to a valid SVGAGuestPtr. The translation does not take place
 687  * immediately, but during a call to vmw_apply_relocations().
 688  * This function builds a relocation list and a list of buffers to validate.
 689  * The former needs to be freed using either vmw_apply_relocations() or
 690  * vmw_free_relocations(). The latter needs to be freed using
 691  * vmw_clear_validations.
 692  */
 693 static int vmw_translate_guest_ptr(struct vmw_private *dev_priv,
 694                                    struct vmw_sw_context *sw_context,
 695                                    SVGAGuestPtr *ptr,
 696                                    struct vmw_dma_buffer **vmw_bo_p)
 697 {
 698         struct vmw_dma_buffer *vmw_bo = NULL;
 699         struct ttm_buffer_object *bo;
 700         uint32_t handle = ptr->gmrId;
 701         struct vmw_relocation *reloc;
 702         int ret;
 703
 704         ret = vmw_user_dmabuf_lookup(sw_context->tfile, handle, &vmw_bo);
 705         if (unlikely(ret != 0)) {
 706                 DRM_ERROR("Could not find or use GMR region.\n");
 707                 return -EINVAL;
 708         }
 709         bo = &vmw_bo->base;
 710
 711         if (unlikely(sw_context->cur_reloc >= VMWGFX_MAX_RELOCATIONS)) {
 712                 DRM_ERROR("Max number relocations per submission"
 713                           " exceeded\n");
 714                 ret = -EINVAL;
 715                 goto out_no_reloc;
 716         }
 717
 718         reloc = &sw_context->relocs[sw_context->cur_reloc++];
 719         reloc->location = ptr;
 720
 721         ret = vmw_bo_to_validate_list(sw_context, bo, &reloc->index);
 722         if (unlikely(ret != 0))
 723                 goto out_no_reloc;
 724
 725         *vmw_bo_p = vmw_bo;
 726         return 0;
 727
 728 out_no_reloc:
 729         vmw_dmabuf_unreference(&vmw_bo);
 730         vmw_bo_p = NULL;
 731         return ret;
 732 }
 733
 734 /**
 735  * vmw_cmd_begin_query - validate a  SVGA_3D_CMD_BEGIN_QUERY command.
 736  *
 737  * @dev_priv: Pointer to a device private struct.
 738  * @sw_context: The software context used for this command submission.
 739  * @header: Pointer to the command header in the command stream.
 740  */
 741 static int vmw_cmd_begin_query(struct vmw_private *dev_priv,
 742                                struct vmw_sw_context *sw_context,
 743                                SVGA3dCmdHeader *header)
 744 {
 745         struct vmw_begin_query_cmd {
 746                 SVGA3dCmdHeader header;
 747                 SVGA3dCmdBeginQuery q;
 748         } *cmd;
 749
 750         cmd = container_of(header, struct vmw_begin_query_cmd,
 751                            header);
 752
 753         return vmw_cmd_res_check(dev_priv, sw_context, vmw_res_context,
 754                                  user_context_converter, &cmd->q.cid,
 755                                  NULL);
 756 }
 757
 758 /**
 759  * vmw_cmd_end_query - validate a  SVGA_3D_CMD_END_QUERY command.
 760  *
 761  * @dev_priv: Pointer to a device private struct.
 762  * @sw_context: The software context used for this command submission.
 763  * @header: Pointer to the command header in the command stream.
 764  */
 765 static int vmw_cmd_end_query(struct vmw_private *dev_priv,
 766                              struct vmw_sw_context *sw_context,
 767                              SVGA3dCmdHeader *header)
 768 {
 769         struct vmw_dma_buffer *vmw_bo;
 770         struct vmw_query_cmd {
 771                 SVGA3dCmdHeader header;
 772                 SVGA3dCmdEndQuery q;
 773         } *cmd;
 774         int ret;
 775
 776         cmd = container_of(header, struct vmw_query_cmd, header);
 777         ret = vmw_cmd_cid_check(dev_priv, sw_context, header);
 778         if (unlikely(ret != 0))
 779                 return ret;
 780
 781         ret = vmw_translate_guest_ptr(dev_priv, sw_context,
 782                                       &cmd->q.guestResult,
 783                                       &vmw_bo);
 784         if (unlikely(ret != 0))
 785                 return ret;
 786
 787         ret = vmw_query_bo_switch_prepare(dev_priv, &vmw_bo->base, sw_context);
 788
 789         vmw_dmabuf_unreference(&vmw_bo);
 790         return ret;
 791 }
 792
 793 /*
 794  * vmw_cmd_wait_query - validate a  SVGA_3D_CMD_WAIT_QUERY command.
 795  *
 796  * @dev_priv: Pointer to a device private struct.
 797  * @sw_context: The software context used for this command submission.
 798  * @header: Pointer to the command header in the command stream.
 799  */
 800 static int vmw_cmd_wait_query(struct vmw_private *dev_priv,
 801                               struct vmw_sw_context *sw_context,
 802                               SVGA3dCmdHeader *header)
 803 {
 804         struct vmw_dma_buffer *vmw_bo;
 805         struct vmw_query_cmd {
 806                 SVGA3dCmdHeader header;
 807                 SVGA3dCmdWaitForQuery q;
 808         } *cmd;
 809         int ret;
 810
 811         cmd = container_of(header, struct vmw_query_cmd, header);
 812         ret = vmw_cmd_cid_check(dev_priv, sw_context, header);
 813         if (unlikely(ret != 0))
 814                 return ret;
 815
 816         ret = vmw_translate_guest_ptr(dev_priv, sw_context,
 817                                       &cmd->q.guestResult,
 818                                       &vmw_bo);
 819         if (unlikely(ret != 0))
 820                 return ret;
 821
 822         vmw_dmabuf_unreference(&vmw_bo);
 823         return 0;
 824 }
 825
 826 static int vmw_cmd_dma(struct vmw_private *dev_priv,
 827                        struct vmw_sw_context *sw_context,
 828                        SVGA3dCmdHeader *header)
 829 {
 830         struct vmw_dma_buffer *vmw_bo = NULL;
 831         struct vmw_surface *srf = NULL;
 832         struct vmw_dma_cmd {
 833                 SVGA3dCmdHeader header;
 834                 SVGA3dCmdSurfaceDMA dma;
 835         } *cmd;
 836         int ret;
 837         SVGA3dCmdSurfaceDMASuffix *suffix;
 838         uint32_t bo_size;
 839
 840         cmd = container_of(header, struct vmw_dma_cmd, header);
 841         suffix = (SVGA3dCmdSurfaceDMASuffix *)((unsigned long) &cmd->dma +
 842                                                header->size - sizeof(*suffix));
 843
 844         /* Make sure device and verifier stays in sync. */
 845         if (unlikely(suffix->suffixSize != sizeof(*suffix))) {
 846                 DRM_ERROR("Invalid DMA suffix size.\n");
 847                 return -EINVAL;
 848         }
 849
 850         ret = vmw_translate_guest_ptr(dev_priv, sw_context,
 851                                       &cmd->dma.guest.ptr,
 852                                       &vmw_bo);
 853         if (unlikely(ret != 0))
 854                 return ret;
 855
 856         /* Make sure DMA doesn't cross BO boundaries. */
 857         bo_size = vmw_bo->base.num_pages * PAGE_SIZE;
 858         if (unlikely(cmd->dma.guest.ptr.offset > bo_size)) {
 859                 DRM_ERROR("Invalid DMA offset.\n");
 860                 return -EINVAL;
 861         }
 862
 863         bo_size -= cmd->dma.guest.ptr.offset;
 864         if (unlikely(suffix->maximumOffset > bo_size))
 865                 suffix->maximumOffset = bo_size;
 866
 867         ret = vmw_cmd_res_check(dev_priv, sw_context, vmw_res_surface,
 868                                 user_surface_converter, &cmd->dma.host.sid,
 869                                 NULL);
 870         if (unlikely(ret != 0)) {
 871                 if (unlikely(ret != -ERESTARTSYS))
 872                         DRM_ERROR("could not find surface for DMA.\n");
 873                 goto out_no_surface;
 874         }
 875
 876         srf = vmw_res_to_srf(sw_context->res_cache[vmw_res_surface].res);
 877
 878         vmw_kms_cursor_snoop(srf, sw_context->tfile, &vmw_bo->base, header);
 879
 880 out_no_surface:
 881         vmw_dmabuf_unreference(&vmw_bo);
 882         return ret;
 883 }
 884
 885 static int vmw_cmd_draw(struct vmw_private *dev_priv,
 886                         struct vmw_sw_context *sw_context,
 887                         SVGA3dCmdHeader *header)
 888 {
 889         struct vmw_draw_cmd {
 890                 SVGA3dCmdHeader header;
 891                 SVGA3dCmdDrawPrimitives body;
 892         } *cmd;
 893         SVGA3dVertexDecl *decl = (SVGA3dVertexDecl *)(
 894                 (unsigned long)header + sizeof(*cmd));
 895         SVGA3dPrimitiveRange *range;
 896         uint32_t i;
 897         uint32_t maxnum;
 898         int ret;
 899
 900         ret = vmw_cmd_cid_check(dev_priv, sw_context, header);
 901         if (unlikely(ret != 0))
 902                 return ret;
 903
 904         cmd = container_of(header, struct vmw_draw_cmd, header);
 905         maxnum = (header->size - sizeof(cmd->body)) / sizeof(*decl);
 906
 907         if (unlikely(cmd->body.numVertexDecls > maxnum)) {
 908                 DRM_ERROR("Illegal number of vertex declarations.\n");
 909                 return -EINVAL;
 910         }
 911
 912         for (i = 0; i < cmd->body.numVertexDecls; ++i, ++decl) {
 913                 ret = vmw_cmd_res_check(dev_priv, sw_context, vmw_res_surface,
 914                                         user_surface_converter,
 915                                         &decl->array.surfaceId, NULL);
 916                 if (unlikely(ret != 0))
 917                         return ret;
 918         }
 919
 920         maxnum = (header->size - sizeof(cmd->body) -
 921                   cmd->body.numVertexDecls * sizeof(*decl)) / sizeof(*range);
 922         if (unlikely(cmd->body.numRanges > maxnum)) {
 923                 DRM_ERROR("Illegal number of index ranges.\n");
 924                 return -EINVAL;
 925         }
 926
 927         range = (SVGA3dPrimitiveRange *) decl;
 928         for (i = 0; i < cmd->body.numRanges; ++i, ++range) {
 929                 ret = vmw_cmd_res_check(dev_priv, sw_context, vmw_res_surface,
 930                                         user_surface_converter,
 931                                         &range->indexArray.surfaceId, NULL);
 932                 if (unlikely(ret != 0))
 933                         return ret;
 934         }
 935         return 0;
 936 }
 937
 938
 939 static int vmw_cmd_tex_state(struct vmw_private *dev_priv,
 940                              struct vmw_sw_context *sw_context,
 941                              SVGA3dCmdHeader *header)
 942 {
 943         struct vmw_tex_state_cmd {
 944                 SVGA3dCmdHeader header;
 945                 SVGA3dCmdSetTextureState state;
 946         };
 947
 948         SVGA3dTextureState *last_state = (SVGA3dTextureState *)
 949           ((unsigned long) header + header->size + sizeof(header));
 950         SVGA3dTextureState *cur_state = (SVGA3dTextureState *)
 951                 ((unsigned long) header + sizeof(struct vmw_tex_state_cmd));
 952         int ret;
 953
 954         ret = vmw_cmd_cid_check(dev_priv, sw_context, header);
 955         if (unlikely(ret != 0))
 956                 return ret;
 957
 958         for (; cur_state < last_state; ++cur_state) {
 959                 if (likely(cur_state->name != SVGA3D_TS_BIND_TEXTURE))
 960                         continue;
 961
 962                 ret = vmw_cmd_res_check(dev_priv, sw_context, vmw_res_surface,
 963                                         user_surface_converter,
 964                                         &cur_state->value, NULL);
 965                 if (unlikely(ret != 0))
 966                         return ret;
 967         }
 968
 969         return 0;
 970 }
 971
 972 static int vmw_cmd_check_define_gmrfb(struct vmw_private *dev_priv,
 973                                       struct vmw_sw_context *sw_context,
 974                                       void *buf)
 975 {
 976         struct vmw_dma_buffer *vmw_bo;
 977         int ret;
 978
 979         struct {
 980                 uint32_t header;
 981                 SVGAFifoCmdDefineGMRFB body;
 982         } *cmd = buf;
 983
 984         ret = vmw_translate_guest_ptr(dev_priv, sw_context,
 985                                       &cmd->body.ptr,
 986                                       &vmw_bo);
 987         if (unlikely(ret != 0))
 988                 return ret;
 989
 990         vmw_dmabuf_unreference(&vmw_bo);
 991
 992         return ret;
 993 }
 994
 995 /**
 996  * vmw_cmd_set_shader - Validate an SVGA_3D_CMD_SET_SHADER
 997  * command
 998  *
 999  * @dev_priv: Pointer to a device private struct.
1000  * @sw_context: The software context being used for this batch.
1001  * @header: Pointer to the command header in the command stream.
1002  */
1003 static int vmw_cmd_set_shader(struct vmw_private *dev_priv,
1004                               struct vmw_sw_context *sw_context,
1005                               SVGA3dCmdHeader *header)
1006 {
1007         struct vmw_set_shader_cmd {
1008                 SVGA3dCmdHeader header;
1009                 SVGA3dCmdSetShader body;
1010         } *cmd;
1011         int ret;
1012
1013         cmd = container_of(header, struct vmw_set_shader_cmd,
1014                            header);
1015
1016         ret = vmw_cmd_cid_check(dev_priv, sw_context, header);
1017         if (unlikely(ret != 0))
1018                 return ret;
1019
1020         return 0;
1021 }
1022
1023 static int vmw_cmd_check_not_3d(struct vmw_private *dev_priv,
1024                                 struct vmw_sw_context *sw_context,
1025                                 void *buf, uint32_t *size)
1026 {
1027         uint32_t size_remaining = *size;
1028         uint32_t cmd_id;
1029
1030         cmd_id = le32_to_cpu(((uint32_t *)buf)[0]);
1031         switch (cmd_id) {
1032         case SVGA_CMD_UPDATE:
1033                 *size = sizeof(uint32_t) + sizeof(SVGAFifoCmdUpdate);
1034                 break;
1035         case SVGA_CMD_DEFINE_GMRFB:
1036                 *size = sizeof(uint32_t) + sizeof(SVGAFifoCmdDefineGMRFB);
1037                 break;
1038         case SVGA_CMD_BLIT_GMRFB_TO_SCREEN:
1039                 *size = sizeof(uint32_t) + sizeof(SVGAFifoCmdBlitGMRFBToScreen);
1040                 break;
1041         case SVGA_CMD_BLIT_SCREEN_TO_GMRFB:
1042                 *size = sizeof(uint32_t) + sizeof(SVGAFifoCmdBlitGMRFBToScreen);
1043                 break;
1044         default:
1045                 DRM_ERROR("Unsupported SVGA command: %u.\n", cmd_id);
1046                 return -EINVAL;
1047         }
1048
1049         if (*size > size_remaining) {
1050                 DRM_ERROR("Invalid SVGA command (size mismatch):"
1051                           " %u.\n", cmd_id);
1052                 return -EINVAL;
1053         }
1054
1055         if (unlikely(!sw_context->kernel)) {
1056                 DRM_ERROR("Kernel only SVGA command: %u.\n", cmd_id);
1057                 return -EPERM;
1058         }
1059
1060         if (cmd_id == SVGA_CMD_DEFINE_GMRFB)
1061                 return vmw_cmd_check_define_gmrfb(dev_priv, sw_context, buf);
1062
1063         return 0;
1064 }
1065
1066 typedef int (*vmw_cmd_func) (struct vmw_private *,
1067                              struct vmw_sw_context *,
1068                              SVGA3dCmdHeader *);
1069
1070 #define VMW_CMD_DEF(cmd, func) \
1071         [cmd - SVGA_3D_CMD_BASE] = func
1072
1073 static vmw_cmd_func vmw_cmd_funcs[SVGA_3D_CMD_MAX] = {
1074         VMW_CMD_DEF(SVGA_3D_CMD_SURFACE_DEFINE, &vmw_cmd_invalid),
1075         VMW_CMD_DEF(SVGA_3D_CMD_SURFACE_DESTROY, &vmw_cmd_invalid),
1076         VMW_CMD_DEF(SVGA_3D_CMD_SURFACE_COPY, &vmw_cmd_surface_copy_check),
1077         VMW_CMD_DEF(SVGA_3D_CMD_SURFACE_STRETCHBLT, &vmw_cmd_stretch_blt_check),
1078         VMW_CMD_DEF(SVGA_3D_CMD_SURFACE_DMA, &vmw_cmd_dma),
1079         VMW_CMD_DEF(SVGA_3D_CMD_CONTEXT_DEFINE, &vmw_cmd_invalid),
1080         VMW_CMD_DEF(SVGA_3D_CMD_CONTEXT_DESTROY, &vmw_cmd_invalid),
1081         VMW_CMD_DEF(SVGA_3D_CMD_SETTRANSFORM, &vmw_cmd_cid_check),
1082         VMW_CMD_DEF(SVGA_3D_CMD_SETZRANGE, &vmw_cmd_cid_check),
1083         VMW_CMD_DEF(SVGA_3D_CMD_SETRENDERSTATE, &vmw_cmd_cid_check),
1084         VMW_CMD_DEF(SVGA_3D_CMD_SETRENDERTARGET,
1085                     &vmw_cmd_set_render_target_check),
1086         VMW_CMD_DEF(SVGA_3D_CMD_SETTEXTURESTATE, &vmw_cmd_tex_state),
1087         VMW_CMD_DEF(SVGA_3D_CMD_SETMATERIAL, &vmw_cmd_cid_check),
1088         VMW_CMD_DEF(SVGA_3D_CMD_SETLIGHTDATA, &vmw_cmd_cid_check),
1089         VMW_CMD_DEF(SVGA_3D_CMD_SETLIGHTENABLED, &vmw_cmd_cid_check),
1090         VMW_CMD_DEF(SVGA_3D_CMD_SETVIEWPORT, &vmw_cmd_cid_check),
1091         VMW_CMD_DEF(SVGA_3D_CMD_SETCLIPPLANE, &vmw_cmd_cid_check),
1092         VMW_CMD_DEF(SVGA_3D_CMD_CLEAR, &vmw_cmd_cid_check),
1093         VMW_CMD_DEF(SVGA_3D_CMD_PRESENT, &vmw_cmd_present_check),
1094         VMW_CMD_DEF(SVGA_3D_CMD_SHADER_DEFINE, &vmw_cmd_cid_check),
1095         VMW_CMD_DEF(SVGA_3D_CMD_SHADER_DESTROY, &vmw_cmd_cid_check),
1096         VMW_CMD_DEF(SVGA_3D_CMD_SET_SHADER, &vmw_cmd_set_shader),
1097         VMW_CMD_DEF(SVGA_3D_CMD_SET_SHADER_CONST, &vmw_cmd_cid_check),
1098         VMW_CMD_DEF(SVGA_3D_CMD_DRAW_PRIMITIVES, &vmw_cmd_draw),
1099         VMW_CMD_DEF(SVGA_3D_CMD_SETSCISSORRECT, &vmw_cmd_cid_check),
1100         VMW_CMD_DEF(SVGA_3D_CMD_BEGIN_QUERY, &vmw_cmd_begin_query),
1101         VMW_CMD_DEF(SVGA_3D_CMD_END_QUERY, &vmw_cmd_end_query),
1102         VMW_CMD_DEF(SVGA_3D_CMD_WAIT_FOR_QUERY, &vmw_cmd_wait_query),
1103         VMW_CMD_DEF(SVGA_3D_CMD_PRESENT_READBACK, &vmw_cmd_ok),
1104         VMW_CMD_DEF(SVGA_3D_CMD_BLIT_SURFACE_TO_SCREEN,
1105                     &vmw_cmd_blt_surf_screen_check),
1106         VMW_CMD_DEF(SVGA_3D_CMD_SURFACE_DEFINE_V2, &vmw_cmd_invalid),
1107         VMW_CMD_DEF(SVGA_3D_CMD_GENERATE_MIPMAPS, &vmw_cmd_invalid),
1108         VMW_CMD_DEF(SVGA_3D_CMD_ACTIVATE_SURFACE, &vmw_cmd_invalid),
1109         VMW_CMD_DEF(SVGA_3D_CMD_DEACTIVATE_SURFACE, &vmw_cmd_invalid),
1110 };
1111
1112 static int vmw_cmd_check(struct vmw_private *dev_priv,
1113                          struct vmw_sw_context *sw_context,
1114                          void *buf, uint32_t *size)
1115 {
1116         uint32_t cmd_id;
1117         uint32_t size_remaining = *size;
1118         SVGA3dCmdHeader *header = (SVGA3dCmdHeader *) buf;
1119         int ret;
1120
1121         cmd_id = le32_to_cpu(((uint32_t *)buf)[0]);
1122         /* Handle any none 3D commands */
1123         if (unlikely(cmd_id < SVGA_CMD_MAX))
1124                 return vmw_cmd_check_not_3d(dev_priv, sw_context, buf, size);
1125
1126
1127         cmd_id = le32_to_cpu(header->id);
1128         *size = le32_to_cpu(header->size) + sizeof(SVGA3dCmdHeader);
1129
1130         cmd_id -= SVGA_3D_CMD_BASE;
1131         if (unlikely(*size > size_remaining))
1132                 goto out_err;
1133
1134         if (unlikely(cmd_id >= SVGA_3D_CMD_MAX - SVGA_3D_CMD_BASE))
1135                 goto out_err;
1136
1137         ret = vmw_cmd_funcs[cmd_id](dev_priv, sw_context, header);
1138         if (unlikely(ret != 0))
1139                 goto out_err;
1140
1141         return 0;
1142 out_err:
1143         DRM_ERROR("Illegal / Invalid SVGA3D command: %d\n",
1144                   cmd_id + SVGA_3D_CMD_BASE);
1145         return -EINVAL;
1146 }
1147
1148 static int vmw_cmd_check_all(struct vmw_private *dev_priv,
1149                              struct vmw_sw_context *sw_context,
1150                              void *buf,
1151                              uint32_t size)
1152 {
1153         int32_t cur_size = size;
1154         int ret;
1155
1156         sw_context->buf_start = buf;
1157
1158         while (cur_size > 0) {
1159                 size = cur_size;
1160                 ret = vmw_cmd_check(dev_priv, sw_context, buf, &size);
1161                 if (unlikely(ret != 0))
1162                         return ret;
1163                 buf = (void *)((unsigned long) buf + size);
1164                 cur_size -= size;
1165         }
1166
1167         if (unlikely(cur_size != 0)) {
1168                 DRM_ERROR("Command verifier out of sync.\n");
1169                 return -EINVAL;
1170         }
1171
1172         return 0;
1173 }
1174
1175 static void vmw_free_relocations(struct vmw_sw_context *sw_context)
1176 {
1177         sw_context->cur_reloc = 0;
1178 }
1179
1180 static void vmw_apply_relocations(struct vmw_sw_context *sw_context)
1181 {
1182         uint32_t i;
1183         struct vmw_relocation *reloc;
1184         struct ttm_validate_buffer *validate;
1185         struct ttm_buffer_object *bo;
1186
1187         for (i = 0; i < sw_context->cur_reloc; ++i) {
1188                 reloc = &sw_context->relocs[i];
1189                 validate = &sw_context->val_bufs[reloc->index].base;
1190                 bo = validate->bo;
1191                 switch (bo->mem.mem_type) {
1192                 case TTM_PL_VRAM:
1193                         reloc->location->offset += bo->offset;
1194                         reloc->location->gmrId = SVGA_GMR_FRAMEBUFFER;
1195                         break;
1196                 case VMW_PL_GMR:
1197                         reloc->location->gmrId = bo->mem.start;
1198                         break;
1199                 default:
1200                         BUG();
1201                 }
1202         }
1203         vmw_free_relocations(sw_context);
1204 }
1205
1206 /**
1207  * vmw_resource_list_unrefererence - Free up a resource list and unreference
1208  * all resources referenced by it.
1209  *
1210  * @list: The resource list.
1211  */
1212 static void vmw_resource_list_unreference(struct list_head *list)
1213 {
1214         struct vmw_resource_val_node *val, *val_next;
1215
1216         /*
1217          * Drop references to resources held during command submission.
1218          */
1219
1220         list_for_each_entry_safe(val, val_next, list, head) {
1221                 list_del_init(&val->head);
1222                 vmw_resource_unreference(&val->res);
1223                 kfree(val);
1224         }
1225 }
1226
1227 static void vmw_clear_validations(struct vmw_sw_context *sw_context)
1228 {
1229         struct vmw_validate_buffer *entry, *next;
1230         struct vmw_resource_val_node *val;
1231
1232         /*
1233          * Drop references to DMA buffers held during command submission.
1234          */
1235         list_for_each_entry_safe(entry, next, &sw_context->validate_nodes,
1236                                  base.head) {
1237                 list_del(&entry->base.head);
1238                 ttm_bo_unref(&entry->base.bo);
1239                 (void) drm_ht_remove_item(&sw_context->res_ht, &entry->hash);
1240                 sw_context->cur_val_buf--;
1241         }
1242         BUG_ON(sw_context->cur_val_buf != 0);
1243
1244         list_for_each_entry(val, &sw_context->resource_list, head)
1245                 (void) drm_ht_remove_item(&sw_context->res_ht, &val->hash);
1246 }
1247
1248 static int vmw_validate_single_buffer(struct vmw_private *dev_priv,
1249                                       struct ttm_buffer_object *bo)
1250 {
1251         int ret;
1252
1253
1254         /*
1255          * Don't validate pinned buffers.
1256          */
1257
1258         if (bo == dev_priv->pinned_bo ||
1259             (bo == dev_priv->dummy_query_bo &&
1260              dev_priv->dummy_query_bo_pinned))
1261                 return 0;
1262
1263         /**
1264          * Put BO in VRAM if there is space, otherwise as a GMR.
1265          * If there is no space in VRAM and GMR ids are all used up,
1266          * start evicting GMRs to make room. If the DMA buffer can't be
1267          * used as a GMR, this will return -ENOMEM.
1268          */
1269
1270         ret = ttm_bo_validate(bo, &vmw_vram_gmr_placement, true, false);
1271         if (likely(ret == 0 || ret == -ERESTARTSYS))
1272                 return ret;
1273
1274         /**
1275          * If that failed, try VRAM again, this time evicting
1276          * previous contents.
1277          */
1278
1279         DRM_INFO("Falling through to VRAM.\n");
1280         ret = ttm_bo_validate(bo, &vmw_vram_placement, true, false);
1281         return ret;
1282 }
1283
1284
1285 static int vmw_validate_buffers(struct vmw_private *dev_priv,
1286                                 struct vmw_sw_context *sw_context)
1287 {
1288         struct vmw_validate_buffer *entry;
1289         int ret;
1290
1291         list_for_each_entry(entry, &sw_context->validate_nodes, base.head) {
1292                 ret = vmw_validate_single_buffer(dev_priv, entry->base.bo);
1293                 if (unlikely(ret != 0))
1294                         return ret;
1295         }
1296         return 0;
1297 }
1298
1299 static int vmw_resize_cmd_bounce(struct vmw_sw_context *sw_context,
1300                                  uint32_t size)
1301 {
1302         if (likely(sw_context->cmd_bounce_size >= size))
1303                 return 0;
1304
1305         if (sw_context->cmd_bounce_size == 0)
1306                 sw_context->cmd_bounce_size = VMWGFX_CMD_BOUNCE_INIT_SIZE;
1307
1308         while (sw_context->cmd_bounce_size < size) {
1309                 sw_context->cmd_bounce_size =
1310                         PAGE_ALIGN(sw_context->cmd_bounce_size +
1311                                    (sw_context->cmd_bounce_size >> 1));
1312         }
1313
1314         if (sw_context->cmd_bounce != NULL)
1315                 vfree(sw_context->cmd_bounce);
1316
1317         sw_context->cmd_bounce = vmalloc(sw_context->cmd_bounce_size);
1318
1319         if (sw_context->cmd_bounce == NULL) {
1320                 DRM_ERROR("Failed to allocate command bounce buffer.\n");
1321                 sw_context->cmd_bounce_size = 0;
1322                 return -ENOMEM;
1323         }
1324
1325         return 0;
1326 }
1327
1328 /**
1329  * vmw_execbuf_fence_commands - create and submit a command stream fence
1330  *
1331  * Creates a fence object and submits a command stream marker.
1332  * If this fails for some reason, We sync the fifo and return NULL.
1333  * It is then safe to fence buffers with a NULL pointer.
1334  *
1335  * If @p_handle is not NULL @file_priv must also not be NULL. Creates
1336  * a userspace handle if @p_handle is not NULL, otherwise not.
1337  */
1338
1339 int vmw_execbuf_fence_commands(struct drm_file *file_priv,
1340                                struct vmw_private *dev_priv,
1341                                struct vmw_fence_obj **p_fence,
1342                                uint32_t *p_handle)
1343 {
1344         uint32_t sequence;
1345         int ret;
1346         bool synced = false;
1347
1348         /* p_handle implies file_priv. */
1349         BUG_ON(p_handle != NULL && file_priv == NULL);
1350
1351         ret = vmw_fifo_send_fence(dev_priv, &sequence);
1352         if (unlikely(ret != 0)) {
1353                 DRM_ERROR("Fence submission error. Syncing.\n");
1354                 synced = true;
1355         }
1356
1357         if (p_handle != NULL)
1358                 ret = vmw_user_fence_create(file_priv, dev_priv->fman,
1359                                             sequence,
1360                                             DRM_VMW_FENCE_FLAG_EXEC,
1361                                             p_fence, p_handle);
1362         else
1363                 ret = vmw_fence_create(dev_priv->fman, sequence,
1364                                        DRM_VMW_FENCE_FLAG_EXEC,
1365                                        p_fence);
1366
1367         if (unlikely(ret != 0 && !synced)) {
1368                 (void) vmw_fallback_wait(dev_priv, false, false,
1369                                          sequence, false,
1370                                          VMW_FENCE_WAIT_TIMEOUT);
1371                 *p_fence = NULL;
1372         }
1373
1374         return 0;
1375 }
1376
1377 /**
1378  * vmw_execbuf_copy_fence_user - copy fence object information to
1379  * user-space.
1380  *
1381  * @dev_priv: Pointer to a vmw_private struct.
1382  * @vmw_fp: Pointer to the struct vmw_fpriv representing the calling file.
1383  * @ret: Return value from fence object creation.
1384  * @user_fence_rep: User space address of a struct drm_vmw_fence_rep to
1385  * which the information should be copied.
1386  * @fence: Pointer to the fenc object.
1387  * @fence_handle: User-space fence handle.
1388  *
1389  * This function copies fence information to user-space. If copying fails,
1390  * The user-space struct drm_vmw_fence_rep::error member is hopefully
1391  * left untouched, and if it's preloaded with an -EFAULT by user-space,
1392  * the error will hopefully be detected.
1393  * Also if copying fails, user-space will be unable to signal the fence
1394  * object so we wait for it immediately, and then unreference the
1395  * user-space reference.
1396  */
1397 void
1398 vmw_execbuf_copy_fence_user(struct vmw_private *dev_priv,
1399                             struct vmw_fpriv *vmw_fp,
1400                             int ret,
1401                             struct drm_vmw_fence_rep __user *user_fence_rep,
1402                             struct vmw_fence_obj *fence,
1403                             uint32_t fence_handle)
1404 {
1405         struct drm_vmw_fence_rep fence_rep;
1406
1407         if (user_fence_rep == NULL)
1408                 return;
1409
1410         memset(&fence_rep, 0, sizeof(fence_rep));
1411
1412         fence_rep.error = ret;
1413         if (ret == 0) {
1414                 BUG_ON(fence == NULL);
1415
1416                 fence_rep.handle = fence_handle;
1417                 fence_rep.seqno = fence->seqno;
1418                 vmw_update_seqno(dev_priv, &dev_priv->fifo);
1419                 fence_rep.passed_seqno = dev_priv->last_read_seqno;
1420         }
1421
1422         /*
1423          * copy_to_user errors will be detected by user space not
1424          * seeing fence_rep::error filled in. Typically
1425          * user-space would have pre-set that member to -EFAULT.
1426          */
1427         ret = copy_to_user(user_fence_rep, &fence_rep,
1428                            sizeof(fence_rep));
1429
1430         /*
1431          * User-space lost the fence object. We need to sync
1432          * and unreference the handle.
1433          */
1434         if (unlikely(ret != 0) && (fence_rep.error == 0)) {
1435                 ttm_ref_object_base_unref(vmw_fp->tfile,
1436                                           fence_handle, TTM_REF_USAGE);
1437                 DRM_ERROR("Fence copy error. Syncing.\n");
1438                 (void) vmw_fence_obj_wait(fence, fence->signal_mask,
1439                                           false, false,
1440                                           VMW_FENCE_WAIT_TIMEOUT);
1441         }
1442 }
1443
1444 int vmw_execbuf_process(struct drm_file *file_priv,
1445                         struct vmw_private *dev_priv,
1446                         void __user *user_commands,
1447                         void *kernel_commands,
1448                         uint32_t command_size,
1449                         uint64_t throttle_us,
1450                         struct drm_vmw_fence_rep __user *user_fence_rep,
1451                         struct vmw_fence_obj **out_fence)
1452 {
1453         struct vmw_sw_context *sw_context = &dev_priv->ctx;
1454         struct vmw_fence_obj *fence = NULL;
1455         struct vmw_resource *error_resource;
1456         struct list_head resource_list;
1457         uint32_t handle;
1458         void *cmd;
1459         int ret;
1460
1461         ret = mutex_lock_interruptible(&dev_priv->cmdbuf_mutex);
1462         if (unlikely(ret != 0))
1463                 return -ERESTARTSYS;
1464
1465         if (kernel_commands == NULL) {
1466                 sw_context->kernel = false;
1467
1468                 ret = vmw_resize_cmd_bounce(sw_context, command_size);
1469                 if (unlikely(ret != 0))
1470                         goto out_unlock;
1471
1472
1473                 ret = copy_from_user(sw_context->cmd_bounce,
1474                                      user_commands, command_size);
1475
1476                 if (unlikely(ret != 0)) {
1477                         ret = -EFAULT;
1478                         DRM_ERROR("Failed copying commands.\n");
1479                         goto out_unlock;
1480                 }
1481                 kernel_commands = sw_context->cmd_bounce;
1482         } else
1483                 sw_context->kernel = true;
1484
1485         sw_context->tfile = vmw_fpriv(file_priv)->tfile;
1486         sw_context->cur_reloc = 0;
1487         sw_context->cur_val_buf = 0;
1488         sw_context->fence_flags = 0;
1489         INIT_LIST_HEAD(&sw_context->resource_list);
1490         sw_context->cur_query_bo = dev_priv->pinned_bo;
1491         sw_context->last_query_ctx = NULL;
1492         sw_context->needs_post_query_barrier = false;
1493         memset(sw_context->res_cache, 0, sizeof(sw_context->res_cache));
1494         INIT_LIST_HEAD(&sw_context->validate_nodes);
1495         INIT_LIST_HEAD(&sw_context->res_relocations);
1496         if (!sw_context->res_ht_initialized) {
1497                 ret = drm_ht_create(&sw_context->res_ht, VMW_RES_HT_ORDER);
1498                 if (unlikely(ret != 0))
1499                         goto out_unlock;
1500                 sw_context->res_ht_initialized = true;
1501         }
1502
1503         INIT_LIST_HEAD(&resource_list);
1504         ret = vmw_cmd_check_all(dev_priv, sw_context, kernel_commands,
1505                                 command_size);
1506         if (unlikely(ret != 0))
1507                 goto out_err;
1508
1509         ret = vmw_resources_reserve(sw_context);
1510         if (unlikely(ret != 0))
1511                 goto out_err;
1512
1513         ret = ttm_eu_reserve_buffers(&sw_context->validate_nodes);
1514         if (unlikely(ret != 0))
1515                 goto out_err;
1516
1517         ret = vmw_validate_buffers(dev_priv, sw_context);
1518         if (unlikely(ret != 0))
1519                 goto out_err;
1520
1521         ret = vmw_resources_validate(sw_context);
1522         if (unlikely(ret != 0))
1523                 goto out_err;
1524
1525         if (throttle_us) {
1526                 ret = vmw_wait_lag(dev_priv, &dev_priv->fifo.marker_queue,
1527                                    throttle_us);
1528
1529                 if (unlikely(ret != 0))
1530                         goto out_err;
1531         }
1532
1533         cmd = vmw_fifo_reserve(dev_priv, command_size);
1534         if (unlikely(cmd == NULL)) {
1535                 DRM_ERROR("Failed reserving fifo space for commands.\n");
1536                 ret = -ENOMEM;
1537                 goto out_err;
1538         }
1539
1540         vmw_apply_relocations(sw_context);
1541         memcpy(cmd, kernel_commands, command_size);
1542
1543         vmw_resource_relocations_apply(cmd, &sw_context->res_relocations);
1544         vmw_resource_relocations_free(&sw_context->res_relocations);
1545
1546         vmw_fifo_commit(dev_priv, command_size);
1547
1548         vmw_query_bo_switch_commit(dev_priv, sw_context);
1549         ret = vmw_execbuf_fence_commands(file_priv, dev_priv,
1550                                          &fence,
1551                                          (user_fence_rep) ? &handle : NULL);
1552         /*
1553          * This error is harmless, because if fence submission fails,
1554          * vmw_fifo_send_fence will sync. The error will be propagated to
1555          * user-space in @fence_rep
1556          */
1557
1558         if (ret != 0)
1559                 DRM_ERROR("Fence submission error. Syncing.\n");
1560
1561         vmw_resource_list_unreserve(&sw_context->resource_list, false);
1562         ttm_eu_fence_buffer_objects(&sw_context->validate_nodes,
1563                                     (void *) fence);
1564
1565         if (unlikely(dev_priv->pinned_bo != NULL &&
1566                      !dev_priv->query_cid_valid))
1567                 __vmw_execbuf_release_pinned_bo(dev_priv, fence);
1568
1569         vmw_clear_validations(sw_context);
1570         vmw_execbuf_copy_fence_user(dev_priv, vmw_fpriv(file_priv), ret,
1571                                     user_fence_rep, fence, handle);
1572
1573         /* Don't unreference when handing fence out */
1574         if (unlikely(out_fence != NULL)) {
1575                 *out_fence = fence;
1576                 fence = NULL;
1577         } else if (likely(fence != NULL)) {
1578                 vmw_fence_obj_unreference(&fence);
1579         }
1580
1581         list_splice_init(&sw_context->resource_list, &resource_list);
1582         mutex_unlock(&dev_priv->cmdbuf_mutex);
1583
1584         /*
1585          * Unreference resources outside of the cmdbuf_mutex to
1586          * avoid deadlocks in resource destruction paths.
1587          */
1588         vmw_resource_list_unreference(&resource_list);
1589
1590         return 0;
1591
1592 out_err:
1593         vmw_resource_relocations_free(&sw_context->res_relocations);
1594         vmw_free_relocations(sw_context);
1595         ttm_eu_backoff_reservation(&sw_context->validate_nodes);
1596         vmw_resource_list_unreserve(&sw_context->resource_list, true);
1597         vmw_clear_validations(sw_context);
1598         if (unlikely(dev_priv->pinned_bo != NULL &&
1599                      !dev_priv->query_cid_valid))
1600                 __vmw_execbuf_release_pinned_bo(dev_priv, NULL);
1601 out_unlock:
1602         list_splice_init(&sw_context->resource_list, &resource_list);
1603         error_resource = sw_context->error_resource;
1604         sw_context->error_resource = NULL;
1605         mutex_unlock(&dev_priv->cmdbuf_mutex);
1606
1607         /*
1608          * Unreference resources outside of the cmdbuf_mutex to
1609          * avoid deadlocks in resource destruction paths.
1610          */
1611         vmw_resource_list_unreference(&resource_list);
1612         if (unlikely(error_resource != NULL))
1613                 vmw_resource_unreference(&error_resource);
1614
1615         return ret;
1616 }
1617
1618 /**
1619  * vmw_execbuf_unpin_panic - Idle the fifo and unpin the query buffer.
1620  *
1621  * @dev_priv: The device private structure.
1622  *
1623  * This function is called to idle the fifo and unpin the query buffer
1624  * if the normal way to do this hits an error, which should typically be
1625  * extremely rare.
1626  */
1627 static void vmw_execbuf_unpin_panic(struct vmw_private *dev_priv)
1628 {
1629         DRM_ERROR("Can't unpin query buffer. Trying to recover.\n");
1630
1631         (void) vmw_fallback_wait(dev_priv, false, true, 0, false, 10*HZ);
1632         vmw_bo_pin(dev_priv->pinned_bo, false);
1633         vmw_bo_pin(dev_priv->dummy_query_bo, false);
1634         dev_priv->dummy_query_bo_pinned = false;
1635 }
1636
1637
1638 /**
1639  * __vmw_execbuf_release_pinned_bo - Flush queries and unpin the pinned
1640  * query bo.
1641  *
1642  * @dev_priv: The device private structure.
1643  * @fence: If non-NULL should point to a struct vmw_fence_obj issued
1644  * _after_ a query barrier that flushes all queries touching the current
1645  * buffer pointed to by @dev_priv->pinned_bo
1646  *
1647  * This function should be used to unpin the pinned query bo, or
1648  * as a query barrier when we need to make sure that all queries have
1649  * finished before the next fifo command. (For example on hardware
1650  * context destructions where the hardware may otherwise leak unfinished
1651  * queries).
1652  *
1653  * This function does not return any failure codes, but make attempts
1654  * to do safe unpinning in case of errors.
1655  *
1656  * The function will synchronize on the previous query barrier, and will
1657  * thus not finish until that barrier has executed.
1658  *
1659  * the @dev_priv->cmdbuf_mutex needs to be held by the current thread
1660  * before calling this function.
1661  */
1662 void __vmw_execbuf_release_pinned_bo(struct vmw_private *dev_priv,
1663                                      struct vmw_fence_obj *fence)
1664 {
1665         int ret = 0;
1666         struct list_head validate_list;
1667         struct ttm_validate_buffer pinned_val, query_val;
1668         struct vmw_fence_obj *lfence = NULL;
1669
1670         if (dev_priv->pinned_bo == NULL)
1671                 goto out_unlock;
1672
1673         INIT_LIST_HEAD(&validate_list);
1674
1675         pinned_val.bo = ttm_bo_reference(dev_priv->pinned_bo);
1676         list_add_tail(&pinned_val.head, &validate_list);
1677
1678         query_val.bo = ttm_bo_reference(dev_priv->dummy_query_bo);
1679         list_add_tail(&query_val.head, &validate_list);
1680
1681         do {
1682                 ret = ttm_eu_reserve_buffers(&validate_list);
1683         } while (ret == -ERESTARTSYS);
1684
1685         if (unlikely(ret != 0)) {
1686                 vmw_execbuf_unpin_panic(dev_priv);
1687                 goto out_no_reserve;
1688         }
1689
1690         if (dev_priv->query_cid_valid) {
1691                 BUG_ON(fence != NULL);
1692                 ret = vmw_fifo_emit_dummy_query(dev_priv, dev_priv->query_cid);
1693                 if (unlikely(ret != 0)) {
1694                         vmw_execbuf_unpin_panic(dev_priv);
1695                         goto out_no_emit;
1696                 }
1697                 dev_priv->query_cid_valid = false;
1698         }
1699
1700         vmw_bo_pin(dev_priv->pinned_bo, false);
1701         vmw_bo_pin(dev_priv->dummy_query_bo, false);
1702         dev_priv->dummy_query_bo_pinned = false;
1703
1704         if (fence == NULL) {
1705                 (void) vmw_execbuf_fence_commands(NULL, dev_priv, &lfence,
1706                                                   NULL);
1707                 fence = lfence;
1708         }
1709         ttm_eu_fence_buffer_objects(&validate_list, (void *) fence);
1710         if (lfence != NULL)
1711                 vmw_fence_obj_unreference(&lfence);
1712
1713         ttm_bo_unref(&query_val.bo);
1714         ttm_bo_unref(&pinned_val.bo);
1715         ttm_bo_unref(&dev_priv->pinned_bo);
1716
1717 out_unlock:
1718         return;
1719
1720 out_no_emit:
1721         ttm_eu_backoff_reservation(&validate_list);
1722 out_no_reserve:
1723         ttm_bo_unref(&query_val.bo);
1724         ttm_bo_unref(&pinned_val.bo);
1725         ttm_bo_unref(&dev_priv->pinned_bo);
1726 }
1727
1728 /**
1729  * vmw_execbuf_release_pinned_bo - Flush queries and unpin the pinned
1730  * query bo.
1731  *
1732  * @dev_priv: The device private structure.
1733  *
1734  * This function should be used to unpin the pinned query bo, or
1735  * as a query barrier when we need to make sure that all queries have
1736  * finished before the next fifo command. (For example on hardware
1737  * context destructions where the hardware may otherwise leak unfinished
1738  * queries).
1739  *
1740  * This function does not return any failure codes, but make attempts
1741  * to do safe unpinning in case of errors.
1742  *
1743  * The function will synchronize on the previous query barrier, and will
1744  * thus not finish until that barrier has executed.
1745  */
1746 void vmw_execbuf_release_pinned_bo(struct vmw_private *dev_priv)
1747 {
1748         mutex_lock(&dev_priv->cmdbuf_mutex);
1749         if (dev_priv->query_cid_valid)
1750                 __vmw_execbuf_release_pinned_bo(dev_priv, NULL);
1751         mutex_unlock(&dev_priv->cmdbuf_mutex);
1752 }
1753
1754
1755 int vmw_execbuf_ioctl(struct drm_device *dev, void *data,
1756                       struct drm_file *file_priv)
1757 {
1758         struct vmw_private *dev_priv = vmw_priv(dev);
1759         struct drm_vmw_execbuf_arg *arg = (struct drm_vmw_execbuf_arg *)data;
1760         struct vmw_master *vmaster = vmw_master(file_priv->master);
1761         int ret;
1762
1763         /*
1764          * This will allow us to extend the ioctl argument while
1765          * maintaining backwards compatibility:
1766          * We take different code paths depending on the value of
1767          * arg->version.
1768          */
1769
1770         if (unlikely(arg->version != DRM_VMW_EXECBUF_VERSION)) {
1771                 DRM_ERROR("Incorrect execbuf version.\n");
1772                 DRM_ERROR("You're running outdated experimental "
1773                           "vmwgfx user-space drivers.");
1774                 return -EINVAL;
1775         }
1776
1777         ret = ttm_read_lock(&vmaster->lock, true);
1778         if (unlikely(ret != 0))
1779                 return ret;
1780
1781         ret = vmw_execbuf_process(file_priv, dev_priv,
1782                                   (void __user *)(unsigned long)arg->commands,
1783                                   NULL, arg->command_size, arg->throttle_us,
1784                                   (void __user *)(unsigned long)arg->fence_rep,
1785                                   NULL);
1786
1787         if (unlikely(ret != 0))
1788                 goto out_unlock;
1789
1790         vmw_kms_cursor_post_execbuf(dev_priv);
1791
1792 out_unlock:
1793         ttm_read_unlock(&vmaster->lock);
1794         return ret;
1795 }