drivers/gpu/drm/radeon/radeon_cs.c

   1 /*
   2  * Copyright 2008 Jerome Glisse.
   3  * All Rights Reserved.
   4  *
   5  * Permission is hereby granted, free of charge, to any person obtaining a
   6  * copy of this software and associated documentation files (the "Software"),
   7  * to deal in the Software without restriction, including without limitation
   8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   9  * and/or sell copies of the Software, and to permit persons to whom the
  10  * Software is furnished to do so, subject to the following conditions:
  11  *
  12  * The above copyright notice and this permission notice (including the next
  13  * paragraph) shall be included in all copies or substantial portions of the
  14  * Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  19  * PRECISION INSIGHT AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
  20  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  21  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
  22  * DEALINGS IN THE SOFTWARE.
  23  *
  24  * Authors:
  25  *    Jerome Glisse <glisse@freedesktop.org>
  26  */
  27 #include "drmP.h"
  28 #include "radeon_drm.h"
  29 #include "radeon_reg.h"
  30 #include "radeon.h"
  31
  32 void r100_cs_dump_packet(struct radeon_cs_parser *p,
  33                          struct radeon_cs_packet *pkt);
  34
  35 int radeon_cs_parser_relocs(struct radeon_cs_parser *p)
  36 {
  37         struct drm_device *ddev = p->rdev->ddev;
  38         struct radeon_cs_chunk *chunk;
  39         unsigned i, j;
  40         bool duplicate;
  41
  42         if (p->chunk_relocs_idx == -1) {
  43                 return 0;
  44         }
  45         chunk = &p->chunks[p->chunk_relocs_idx];
  46         /* FIXME: we assume that each relocs use 4 dwords */
  47         p->nrelocs = chunk->length_dw / 4;
  48         p->relocs_ptr = kcalloc(p->nrelocs, sizeof(void *), GFP_KERNEL);
  49         if (p->relocs_ptr == NULL) {
  50                 return -ENOMEM;
  51         }
  52         p->relocs = kcalloc(p->nrelocs, sizeof(struct radeon_cs_reloc), GFP_KERNEL);
  53         if (p->relocs == NULL) {
  54                 return -ENOMEM;
  55         }
  56         for (i = 0; i < p->nrelocs; i++) {
  57                 struct drm_radeon_cs_reloc *r;
  58
  59                 duplicate = false;
  60                 r = (struct drm_radeon_cs_reloc *)&chunk->kdata[i*4];
  61                 for (j = 0; j < i; j++) {
  62                         if (r->handle == p->relocs[j].handle) {
  63                                 p->relocs_ptr[i] = &p->relocs[j];
  64                                 duplicate = true;
  65                                 break;
  66                         }
  67                 }
  68                 if (!duplicate) {
  69                         p->relocs[i].gobj = drm_gem_object_lookup(ddev,
  70                                                                   p->filp,
  71                                                                   r->handle);
  72                         if (p->relocs[i].gobj == NULL) {
  73                                 DRM_ERROR("gem object lookup failed 0x%x\n",
  74                                           r->handle);
  75                                 return -ENOENT;
  76                         }
  77                         p->relocs_ptr[i] = &p->relocs[i];
  78                         p->relocs[i].robj = gem_to_radeon_bo(p->relocs[i].gobj);
  79                         p->relocs[i].lobj.bo = p->relocs[i].robj;
  80                         p->relocs[i].lobj.wdomain = r->write_domain;
  81                         p->relocs[i].lobj.rdomain = r->read_domains;
  82                         p->relocs[i].lobj.tv.bo = &p->relocs[i].robj->tbo;
  83                         p->relocs[i].handle = r->handle;
  84                         p->relocs[i].flags = r->flags;
  85                         radeon_bo_list_add_object(&p->relocs[i].lobj,
  86                                                   &p->validated);
  87
  88                 } else
  89                         p->relocs[i].handle = 0;
  90         }
  91         return radeon_bo_list_validate(&p->validated);
  92 }
  93
  94 static int radeon_cs_get_ring(struct radeon_cs_parser *p, u32 ring, s32 priority)
  95 {
  96         p->priority = priority;
  97
  98         switch (ring) {
  99         default:
 100                 DRM_ERROR("unknown ring id: %d\n", ring);
 101                 return -EINVAL;
 102         case RADEON_CS_RING_GFX:
 103                 p->ring = RADEON_RING_TYPE_GFX_INDEX;
 104                 break;
 105         case RADEON_CS_RING_COMPUTE:
 106                 if (p->rdev->family >= CHIP_TAHITI) {
 107                         if (p->priority > 0)
 108                                 p->ring = CAYMAN_RING_TYPE_CP1_INDEX;
 109                         else
 110                                 p->ring = CAYMAN_RING_TYPE_CP2_INDEX;
 111                 } else
 112                         p->ring = RADEON_RING_TYPE_GFX_INDEX;
 113                 break;
 114         }
 115         return 0;
 116 }
 117
 118 static int radeon_cs_sync_rings(struct radeon_cs_parser *p)
 119 {
 120         bool sync_to_ring[RADEON_NUM_RINGS] = { };
 121         bool need_sync = false;
 122         int i, r;
 123
 124         for (i = 0; i < p->nrelocs; i++) {
 125                 struct radeon_fence *fence;
 126
 127                 if (!p->relocs[i].robj || !p->relocs[i].robj->tbo.sync_obj)
 128                         continue;
 129
 130                 fence = p->relocs[i].robj->tbo.sync_obj;
 131                 if (fence->ring != p->ring && !radeon_fence_signaled(fence)) {
 132                         sync_to_ring[fence->ring] = true;
 133                         need_sync = true;
 134                 }
 135         }
 136
 137         if (!need_sync) {
 138                 return 0;
 139         }
 140
 141         r = radeon_semaphore_create(p->rdev, &p->ib.semaphore);
 142         if (r) {
 143                 return r;
 144         }
 145
 146         return radeon_semaphore_sync_rings(p->rdev, p->ib.semaphore,
 147                                            sync_to_ring, p->ring);
 148 }
 149
 150 /* XXX: note that this is called from the legacy UMS CS ioctl as well */
 151 int radeon_cs_parser_init(struct radeon_cs_parser *p, void *data)
 152 {
 153         struct drm_radeon_cs *cs = data;
 154         uint64_t *chunk_array_ptr;
 155         unsigned size, i;
 156         u32 ring = RADEON_CS_RING_GFX;
 157         s32 priority = 0;
 158
 159         if (!cs->num_chunks) {
 160                 return 0;
 161         }
 162         /* get chunks */
 163         INIT_LIST_HEAD(&p->validated);
 164         p->idx = 0;
 165         p->ib.sa_bo = NULL;
 166         p->ib.semaphore = NULL;
 167         p->const_ib.sa_bo = NULL;
 168         p->const_ib.semaphore = NULL;
 169         p->chunk_ib_idx = -1;
 170         p->chunk_relocs_idx = -1;
 171         p->chunk_flags_idx = -1;
 172         p->chunk_const_ib_idx = -1;
 173         p->chunks_array = kcalloc(cs->num_chunks, sizeof(uint64_t), GFP_KERNEL);
 174         if (p->chunks_array == NULL) {
 175                 return -ENOMEM;
 176         }
 177         chunk_array_ptr = (uint64_t *)(unsigned long)(cs->chunks);
 178         if (DRM_COPY_FROM_USER(p->chunks_array, chunk_array_ptr,
 179                                sizeof(uint64_t)*cs->num_chunks)) {
 180                 return -EFAULT;
 181         }
 182         p->cs_flags = 0;
 183         p->nchunks = cs->num_chunks;
 184         p->chunks = kcalloc(p->nchunks, sizeof(struct radeon_cs_chunk), GFP_KERNEL);
 185         if (p->chunks == NULL) {
 186                 return -ENOMEM;
 187         }
 188         for (i = 0; i < p->nchunks; i++) {
 189                 struct drm_radeon_cs_chunk __user **chunk_ptr = NULL;
 190                 struct drm_radeon_cs_chunk user_chunk;
 191                 uint32_t __user *cdata;
 192
 193                 chunk_ptr = (void __user*)(unsigned long)p->chunks_array[i];
 194                 if (DRM_COPY_FROM_USER(&user_chunk, chunk_ptr,
 195                                        sizeof(struct drm_radeon_cs_chunk))) {
 196                         return -EFAULT;
 197                 }
 198                 p->chunks[i].length_dw = user_chunk.length_dw;
 199                 p->chunks[i].kdata = NULL;
 200                 p->chunks[i].chunk_id = user_chunk.chunk_id;
 201
 202                 if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_RELOCS) {
 203                         p->chunk_relocs_idx = i;
 204                 }
 205                 if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_IB) {
 206                         p->chunk_ib_idx = i;
 207                         /* zero length IB isn't useful */
 208                         if (p->chunks[i].length_dw == 0)
 209                                 return -EINVAL;
 210                 }
 211                 if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_CONST_IB) {
 212                         p->chunk_const_ib_idx = i;
 213                         /* zero length CONST IB isn't useful */
 214                         if (p->chunks[i].length_dw == 0)
 215                                 return -EINVAL;
 216                 }
 217                 if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_FLAGS) {
 218                         p->chunk_flags_idx = i;
 219                         /* zero length flags aren't useful */
 220                         if (p->chunks[i].length_dw == 0)
 221                                 return -EINVAL;
 222                 }
 223
 224                 p->chunks[i].length_dw = user_chunk.length_dw;
 225                 p->chunks[i].user_ptr = (void __user *)(unsigned long)user_chunk.chunk_data;
 226
 227                 cdata = (uint32_t *)(unsigned long)user_chunk.chunk_data;
 228                 if ((p->chunks[i].chunk_id == RADEON_CHUNK_ID_RELOCS) ||
 229                     (p->chunks[i].chunk_id == RADEON_CHUNK_ID_FLAGS)) {
 230                         size = p->chunks[i].length_dw * sizeof(uint32_t);
 231                         p->chunks[i].kdata = kmalloc(size, GFP_KERNEL);
 232                         if (p->chunks[i].kdata == NULL) {
 233                                 return -ENOMEM;
 234                         }
 235                         if (DRM_COPY_FROM_USER(p->chunks[i].kdata,
 236                                                p->chunks[i].user_ptr, size)) {
 237                                 return -EFAULT;
 238                         }
 239                         if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_FLAGS) {
 240                                 p->cs_flags = p->chunks[i].kdata[0];
 241                                 if (p->chunks[i].length_dw > 1)
 242                                         ring = p->chunks[i].kdata[1];
 243                                 if (p->chunks[i].length_dw > 2)
 244                                         priority = (s32)p->chunks[i].kdata[2];
 245                         }
 246                 }
 247         }
 248
 249         /* these are KMS only */
 250         if (p->rdev) {
 251                 if ((p->cs_flags & RADEON_CS_USE_VM) &&
 252                     !p->rdev->vm_manager.enabled) {
 253                         DRM_ERROR("VM not active on asic!\n");
 254                         return -EINVAL;
 255                 }
 256
 257                 /* we only support VM on SI+ */
 258                 if ((p->rdev->family >= CHIP_TAHITI) &&
 259                     ((p->cs_flags & RADEON_CS_USE_VM) == 0)) {
 260                         DRM_ERROR("VM required on SI+!\n");
 261                         return -EINVAL;
 262                 }
 263
 264                 if (radeon_cs_get_ring(p, ring, priority))
 265                         return -EINVAL;
 266         }
 267
 268         /* deal with non-vm */
 269         if ((p->chunk_ib_idx != -1) &&
 270             ((p->cs_flags & RADEON_CS_USE_VM) == 0) &&
 271             (p->chunks[p->chunk_ib_idx].chunk_id == RADEON_CHUNK_ID_IB)) {
 272                 if (p->chunks[p->chunk_ib_idx].length_dw > (16 * 1024)) {
 273                         DRM_ERROR("cs IB too big: %d\n",
 274                                   p->chunks[p->chunk_ib_idx].length_dw);
 275                         return -EINVAL;
 276                 }
 277                 if ((p->rdev->flags & RADEON_IS_AGP)) {
 278                         p->chunks[p->chunk_ib_idx].kpage[0] = kmalloc(PAGE_SIZE, GFP_KERNEL);
 279                         p->chunks[p->chunk_ib_idx].kpage[1] = kmalloc(PAGE_SIZE, GFP_KERNEL);
 280                         if (p->chunks[p->chunk_ib_idx].kpage[0] == NULL ||
 281                             p->chunks[p->chunk_ib_idx].kpage[1] == NULL) {
 282                                 kfree(p->chunks[i].kpage[0]);
 283                                 kfree(p->chunks[i].kpage[1]);
 284                                 return -ENOMEM;
 285                         }
 286                 }
 287                 p->chunks[p->chunk_ib_idx].kpage_idx[0] = -1;
 288                 p->chunks[p->chunk_ib_idx].kpage_idx[1] = -1;
 289                 p->chunks[p->chunk_ib_idx].last_copied_page = -1;
 290                 p->chunks[p->chunk_ib_idx].last_page_index =
 291                         ((p->chunks[p->chunk_ib_idx].length_dw * 4) - 1) / PAGE_SIZE;
 292         }
 293
 294         return 0;
 295 }
 296
 297 /**
 298  * cs_parser_fini() - clean parser states
 299  * @parser:     parser structure holding parsing context.
 300  * @error:      error number
 301  *
 302  * If error is set than unvalidate buffer, otherwise just free memory
 303  * used by parsing context.
 304  **/
 305 static void radeon_cs_parser_fini(struct radeon_cs_parser *parser, int error)
 306 {
 307         unsigned i;
 308
 309         if (!error)
 310                 ttm_eu_fence_buffer_objects(&parser->validated,
 311                                             parser->ib.fence);
 312         else
 313                 ttm_eu_backoff_reservation(&parser->validated);
 314
 315         if (parser->relocs != NULL) {
 316                 for (i = 0; i < parser->nrelocs; i++) {
 317                         if (parser->relocs[i].gobj)
 318                                 drm_gem_object_unreference_unlocked(parser->relocs[i].gobj);
 319                 }
 320         }
 321         kfree(parser->track);
 322         kfree(parser->relocs);
 323         kfree(parser->relocs_ptr);
 324         for (i = 0; i < parser->nchunks; i++) {
 325                 kfree(parser->chunks[i].kdata);
 326                 if ((parser->rdev->flags & RADEON_IS_AGP)) {
 327                         kfree(parser->chunks[i].kpage[0]);
 328                         kfree(parser->chunks[i].kpage[1]);
 329                 }
 330         }
 331         kfree(parser->chunks);
 332         kfree(parser->chunks_array);
 333         radeon_ib_free(parser->rdev, &parser->ib);
 334         radeon_ib_free(parser->rdev, &parser->const_ib);
 335 }
 336
 337 static int radeon_cs_ib_chunk(struct radeon_device *rdev,
 338                               struct radeon_cs_parser *parser)
 339 {
 340         struct radeon_cs_chunk *ib_chunk;
 341         int r;
 342
 343         if (parser->chunk_ib_idx == -1)
 344                 return 0;
 345
 346         if (parser->cs_flags & RADEON_CS_USE_VM)
 347                 return 0;
 348
 349         ib_chunk = &parser->chunks[parser->chunk_ib_idx];
 350         /* Copy the packet into the IB, the parser will read from the
 351          * input memory (cached) and write to the IB (which can be
 352          * uncached).
 353          */
 354         r =  radeon_ib_get(rdev, parser->ring, &parser->ib,
 355                            ib_chunk->length_dw * 4);
 356         if (r) {
 357                 DRM_ERROR("Failed to get ib !\n");
 358                 return r;
 359         }
 360         parser->ib.length_dw = ib_chunk->length_dw;
 361         r = radeon_cs_parse(rdev, parser->ring, parser);
 362         if (r || parser->parser_error) {
 363                 DRM_ERROR("Invalid command stream !\n");
 364                 return r;
 365         }
 366         r = radeon_cs_finish_pages(parser);
 367         if (r) {
 368                 DRM_ERROR("Invalid command stream !\n");
 369                 return r;
 370         }
 371         r = radeon_cs_sync_rings(parser);
 372         if (r) {
 373                 DRM_ERROR("Failed to synchronize rings !\n");
 374         }
 375         parser->ib.vm_id = 0;
 376         r = radeon_ib_schedule(rdev, &parser->ib);
 377         if (r) {
 378                 DRM_ERROR("Failed to schedule IB !\n");
 379         }
 380         return 0;
 381 }
 382
 383 static int radeon_bo_vm_update_pte(struct radeon_cs_parser *parser,
 384                                    struct radeon_vm *vm)
 385 {
 386         struct radeon_bo_list *lobj;
 387         struct radeon_bo *bo;
 388         int r;
 389
 390         list_for_each_entry(lobj, &parser->validated, tv.head) {
 391                 bo = lobj->bo;
 392                 r = radeon_vm_bo_update_pte(parser->rdev, vm, bo, &bo->tbo.mem);
 393                 if (r) {
 394                         return r;
 395                 }
 396         }
 397         return 0;
 398 }
 399
 400 static int radeon_cs_ib_vm_chunk(struct radeon_device *rdev,
 401                                  struct radeon_cs_parser *parser)
 402 {
 403         struct radeon_cs_chunk *ib_chunk;
 404         struct radeon_fpriv *fpriv = parser->filp->driver_priv;
 405         struct radeon_vm *vm = &fpriv->vm;
 406         int r;
 407
 408         if (parser->chunk_ib_idx == -1)
 409                 return 0;
 410
 411         if ((parser->cs_flags & RADEON_CS_USE_VM) == 0)
 412                 return 0;
 413
 414         if ((rdev->family >= CHIP_TAHITI) &&
 415             (parser->chunk_const_ib_idx != -1)) {
 416                 ib_chunk = &parser->chunks[parser->chunk_const_ib_idx];
 417                 if (ib_chunk->length_dw > RADEON_IB_VM_MAX_SIZE) {
 418                         DRM_ERROR("cs IB CONST too big: %d\n", ib_chunk->length_dw);
 419                         return -EINVAL;
 420                 }
 421                 r =  radeon_ib_get(rdev, parser->ring, &parser->const_ib,
 422                                    ib_chunk->length_dw * 4);
 423                 if (r) {
 424                         DRM_ERROR("Failed to get const ib !\n");
 425                         return r;
 426                 }
 427                 parser->const_ib.is_const_ib = true;
 428                 parser->const_ib.length_dw = ib_chunk->length_dw;
 429                 /* Copy the packet into the IB */
 430                 if (DRM_COPY_FROM_USER(parser->const_ib.ptr, ib_chunk->user_ptr,
 431                                        ib_chunk->length_dw * 4)) {
 432                         return -EFAULT;
 433                 }
 434                 r = radeon_ring_ib_parse(rdev, parser->ring, &parser->const_ib);
 435                 if (r) {
 436                         return r;
 437                 }
 438         }
 439
 440         ib_chunk = &parser->chunks[parser->chunk_ib_idx];
 441         if (ib_chunk->length_dw > RADEON_IB_VM_MAX_SIZE) {
 442                 DRM_ERROR("cs IB too big: %d\n", ib_chunk->length_dw);
 443                 return -EINVAL;
 444         }
 445         r =  radeon_ib_get(rdev, parser->ring, &parser->ib,
 446                            ib_chunk->length_dw * 4);
 447         if (r) {
 448                 DRM_ERROR("Failed to get ib !\n");
 449                 return r;
 450         }
 451         parser->ib.length_dw = ib_chunk->length_dw;
 452         /* Copy the packet into the IB */
 453         if (DRM_COPY_FROM_USER(parser->ib.ptr, ib_chunk->user_ptr,
 454                                ib_chunk->length_dw * 4)) {
 455                 return -EFAULT;
 456         }
 457         r = radeon_ring_ib_parse(rdev, parser->ring, &parser->ib);
 458         if (r) {
 459                 return r;
 460         }
 461
 462         mutex_lock(&vm->mutex);
 463         r = radeon_vm_bind(rdev, vm);
 464         if (r) {
 465                 goto out;
 466         }
 467         r = radeon_bo_vm_update_pte(parser, vm);
 468         if (r) {
 469                 goto out;
 470         }
 471         r = radeon_cs_sync_rings(parser);
 472         if (r) {
 473                 DRM_ERROR("Failed to synchronize rings !\n");
 474         }
 475
 476         if ((rdev->family >= CHIP_TAHITI) &&
 477             (parser->chunk_const_ib_idx != -1)) {
 478                 parser->const_ib.vm_id = vm->id;
 479                 /* ib pool is bind at 0 in virtual address space to gpu_addr is the
 480                  * offset inside the pool bo
 481                  */
 482                 parser->const_ib.gpu_addr = parser->const_ib.sa_bo->soffset;
 483                 r = radeon_ib_schedule(rdev, &parser->const_ib);
 484                 if (r)
 485                         goto out;
 486         }
 487
 488         parser->ib.vm_id = vm->id;
 489         /* ib pool is bind at 0 in virtual address space to gpu_addr is the
 490          * offset inside the pool bo
 491          */
 492         parser->ib.gpu_addr = parser->ib.sa_bo->soffset;
 493         parser->ib.is_const_ib = false;
 494         r = radeon_ib_schedule(rdev, &parser->ib);
 495 out:
 496         if (!r) {
 497                 if (vm->fence) {
 498                         radeon_fence_unref(&vm->fence);
 499                 }
 500                 vm->fence = radeon_fence_ref(parser->ib.fence);
 501         }
 502         mutex_unlock(&fpriv->vm.mutex);
 503         return r;
 504 }
 505
 506 static int radeon_cs_handle_lockup(struct radeon_device *rdev, int r)
 507 {
 508         if (r == -EDEADLK) {
 509                 r = radeon_gpu_reset(rdev);
 510                 if (!r)
 511                         r = -EAGAIN;
 512         }
 513         return r;
 514 }
 515
 516 int radeon_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
 517 {
 518         struct radeon_device *rdev = dev->dev_private;
 519         struct radeon_cs_parser parser;
 520         int r;
 521
 522         radeon_mutex_lock(&rdev->cs_mutex);
 523         if (!rdev->accel_working) {
 524                 radeon_mutex_unlock(&rdev->cs_mutex);
 525                 return -EBUSY;
 526         }
 527         /* initialize parser */
 528         memset(&parser, 0, sizeof(struct radeon_cs_parser));
 529         parser.filp = filp;
 530         parser.rdev = rdev;
 531         parser.dev = rdev->dev;
 532         parser.family = rdev->family;
 533         r = radeon_cs_parser_init(&parser, data);
 534         if (r) {
 535                 DRM_ERROR("Failed to initialize parser !\n");
 536                 radeon_cs_parser_fini(&parser, r);
 537                 r = radeon_cs_handle_lockup(rdev, r);
 538                 radeon_mutex_unlock(&rdev->cs_mutex);
 539                 return r;
 540         }
 541         r = radeon_cs_parser_relocs(&parser);
 542         if (r) {
 543                 if (r != -ERESTARTSYS)
 544                         DRM_ERROR("Failed to parse relocation %d!\n", r);
 545                 radeon_cs_parser_fini(&parser, r);
 546                 r = radeon_cs_handle_lockup(rdev, r);
 547                 radeon_mutex_unlock(&rdev->cs_mutex);
 548                 return r;
 549         }
 550         r = radeon_cs_ib_chunk(rdev, &parser);
 551         if (r) {
 552                 goto out;
 553         }
 554         r = radeon_cs_ib_vm_chunk(rdev, &parser);
 555         if (r) {
 556                 goto out;
 557         }
 558 out:
 559         radeon_cs_parser_fini(&parser, r);
 560         r = radeon_cs_handle_lockup(rdev, r);
 561         radeon_mutex_unlock(&rdev->cs_mutex);
 562         return r;
 563 }
 564
 565 int radeon_cs_finish_pages(struct radeon_cs_parser *p)
 566 {
 567         struct radeon_cs_chunk *ibc = &p->chunks[p->chunk_ib_idx];
 568         int i;
 569         int size = PAGE_SIZE;
 570
 571         for (i = ibc->last_copied_page + 1; i <= ibc->last_page_index; i++) {
 572                 if (i == ibc->last_page_index) {
 573                         size = (ibc->length_dw * 4) % PAGE_SIZE;
 574                         if (size == 0)
 575                                 size = PAGE_SIZE;
 576                 }
 577
 578                 if (DRM_COPY_FROM_USER(p->ib.ptr + (i * (PAGE_SIZE/4)),
 579                                        ibc->user_ptr + (i * PAGE_SIZE),
 580                                        size))
 581                         return -EFAULT;
 582         }
 583         return 0;
 584 }
 585
 586 static int radeon_cs_update_pages(struct radeon_cs_parser *p, int pg_idx)
 587 {
 588         int new_page;
 589         struct radeon_cs_chunk *ibc = &p->chunks[p->chunk_ib_idx];
 590         int i;
 591         int size = PAGE_SIZE;
 592         bool copy1 = (p->rdev->flags & RADEON_IS_AGP) ? false : true;
 593
 594         for (i = ibc->last_copied_page + 1; i < pg_idx; i++) {
 595                 if (DRM_COPY_FROM_USER(p->ib.ptr + (i * (PAGE_SIZE/4)),
 596                                        ibc->user_ptr + (i * PAGE_SIZE),
 597                                        PAGE_SIZE)) {
 598                         p->parser_error = -EFAULT;
 599                         return 0;
 600                 }
 601         }
 602
 603         if (pg_idx == ibc->last_page_index) {
 604                 size = (ibc->length_dw * 4) % PAGE_SIZE;
 605                 if (size == 0)
 606                         size = PAGE_SIZE;
 607         }
 608
 609         new_page = ibc->kpage_idx[0] < ibc->kpage_idx[1] ? 0 : 1;
 610         if (copy1)
 611                 ibc->kpage[new_page] = p->ib.ptr + (pg_idx * (PAGE_SIZE / 4));
 612
 613         if (DRM_COPY_FROM_USER(ibc->kpage[new_page],
 614                                ibc->user_ptr + (pg_idx * PAGE_SIZE),
 615                                size)) {
 616                 p->parser_error = -EFAULT;
 617                 return 0;
 618         }
 619
 620         /* copy to IB for non single case */
 621         if (!copy1)
 622                 memcpy((void *)(p->ib.ptr+(pg_idx*(PAGE_SIZE/4))), ibc->kpage[new_page], size);
 623
 624         ibc->last_copied_page = pg_idx;
 625         ibc->kpage_idx[new_page] = pg_idx;
 626
 627         return new_page;
 628 }
 629
 630 u32 radeon_get_ib_value(struct radeon_cs_parser *p, int idx)
 631 {
 632         struct radeon_cs_chunk *ibc = &p->chunks[p->chunk_ib_idx];
 633         u32 pg_idx, pg_offset;
 634         u32 idx_value = 0;
 635         int new_page;
 636
 637         pg_idx = (idx * 4) / PAGE_SIZE;
 638         pg_offset = (idx * 4) % PAGE_SIZE;
 639
 640         if (ibc->kpage_idx[0] == pg_idx)
 641                 return ibc->kpage[0][pg_offset/4];
 642         if (ibc->kpage_idx[1] == pg_idx)
 643                 return ibc->kpage[1][pg_offset/4];
 644
 645         new_page = radeon_cs_update_pages(p, pg_idx);
 646         if (new_page < 0) {
 647                 p->parser_error = new_page;
 648                 return 0;
 649         }
 650
 651         idx_value = ibc->kpage[new_page][pg_offset/4];
 652         return idx_value;
 653 }