drivers/gpu/drm/radeon/radeon_gart.c

   1 /*
   2  * Copyright 2008 Advanced Micro Devices, Inc.
   3  * Copyright 2008 Red Hat Inc.
   4  * Copyright 2009 Jerome Glisse.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the "Software"),
   8  * to deal in the Software without restriction, including without limitation
   9  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  10  * and/or sell copies of the Software, and to permit persons to whom the
  11  * Software is furnished to do so, subject to the following conditions:
  12  *
  13  * The above copyright notice and this permission notice shall be included in
  14  * all copies or substantial portions of the 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  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) 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
  22  * OTHER DEALINGS IN THE SOFTWARE.
  23  *
  24  * Authors: Dave Airlie
  25  *          Alex Deucher
  26  *          Jerome Glisse
  27  */
  28 #include "drmP.h"
  29 #include "radeon_drm.h"
  30 #include "radeon.h"
  31 #include "radeon_reg.h"
  32
  33 /*
  34  * Common GART table functions.
  35  */
  36 int radeon_gart_table_ram_alloc(struct radeon_device *rdev)
  37 {
  38         void *ptr;
  39
  40         ptr = pci_alloc_consistent(rdev->pdev, rdev->gart.table_size,
  41                                    &rdev->gart.table_addr);
  42         if (ptr == NULL) {
  43                 return -ENOMEM;
  44         }
  45 #ifdef CONFIG_X86
  46         if (rdev->family == CHIP_RS400 || rdev->family == CHIP_RS480 ||
  47             rdev->family == CHIP_RS690 || rdev->family == CHIP_RS740) {
  48                 set_memory_uc((unsigned long)ptr,
  49                               rdev->gart.table_size >> PAGE_SHIFT);
  50         }
  51 #endif
  52         rdev->gart.ptr = ptr;
  53         memset((void *)rdev->gart.ptr, 0, rdev->gart.table_size);
  54         return 0;
  55 }
  56
  57 void radeon_gart_table_ram_free(struct radeon_device *rdev)
  58 {
  59         if (rdev->gart.ptr == NULL) {
  60                 return;
  61         }
  62 #ifdef CONFIG_X86
  63         if (rdev->family == CHIP_RS400 || rdev->family == CHIP_RS480 ||
  64             rdev->family == CHIP_RS690 || rdev->family == CHIP_RS740) {
  65                 set_memory_wb((unsigned long)rdev->gart.ptr,
  66                               rdev->gart.table_size >> PAGE_SHIFT);
  67         }
  68 #endif
  69         pci_free_consistent(rdev->pdev, rdev->gart.table_size,
  70                             (void *)rdev->gart.ptr,
  71                             rdev->gart.table_addr);
  72         rdev->gart.ptr = NULL;
  73         rdev->gart.table_addr = 0;
  74 }
  75
  76 int radeon_gart_table_vram_alloc(struct radeon_device *rdev)
  77 {
  78         int r;
  79
  80         if (rdev->gart.robj == NULL) {
  81                 r = radeon_bo_create(rdev, rdev->gart.table_size,
  82                                      PAGE_SIZE, true, RADEON_GEM_DOMAIN_VRAM,
  83                                      NULL, &rdev->gart.robj);
  84                 if (r) {
  85                         return r;
  86                 }
  87         }
  88         return 0;
  89 }
  90
  91 int radeon_gart_table_vram_pin(struct radeon_device *rdev)
  92 {
  93         uint64_t gpu_addr;
  94         int r;
  95
  96         r = radeon_bo_reserve(rdev->gart.robj, false);
  97         if (unlikely(r != 0))
  98                 return r;
  99         r = radeon_bo_pin(rdev->gart.robj,
 100                                 RADEON_GEM_DOMAIN_VRAM, &gpu_addr);
 101         if (r) {
 102                 radeon_bo_unreserve(rdev->gart.robj);
 103                 return r;
 104         }
 105         r = radeon_bo_kmap(rdev->gart.robj, &rdev->gart.ptr);
 106         if (r)
 107                 radeon_bo_unpin(rdev->gart.robj);
 108         radeon_bo_unreserve(rdev->gart.robj);
 109         rdev->gart.table_addr = gpu_addr;
 110         return r;
 111 }
 112
 113 void radeon_gart_table_vram_unpin(struct radeon_device *rdev)
 114 {
 115         int r;
 116
 117         if (rdev->gart.robj == NULL) {
 118                 return;
 119         }
 120         r = radeon_bo_reserve(rdev->gart.robj, false);
 121         if (likely(r == 0)) {
 122                 radeon_bo_kunmap(rdev->gart.robj);
 123                 radeon_bo_unpin(rdev->gart.robj);
 124                 radeon_bo_unreserve(rdev->gart.robj);
 125                 rdev->gart.ptr = NULL;
 126         }
 127 }
 128
 129 void radeon_gart_table_vram_free(struct radeon_device *rdev)
 130 {
 131         if (rdev->gart.robj == NULL) {
 132                 return;
 133         }
 134         radeon_gart_table_vram_unpin(rdev);
 135         radeon_bo_unref(&rdev->gart.robj);
 136 }
 137
 138
 139
 140
 141 /*
 142  * Common gart functions.
 143  */
 144 void radeon_gart_unbind(struct radeon_device *rdev, unsigned offset,
 145                         int pages)
 146 {
 147         unsigned t;
 148         unsigned p;
 149         int i, j;
 150         u64 page_base;
 151
 152         if (!rdev->gart.ready) {
 153                 WARN(1, "trying to unbind memory from uninitialized GART !\n");
 154                 return;
 155         }
 156         t = offset / RADEON_GPU_PAGE_SIZE;
 157         p = t / (PAGE_SIZE / RADEON_GPU_PAGE_SIZE);
 158         for (i = 0; i < pages; i++, p++) {
 159                 if (rdev->gart.pages[p]) {
 160                         rdev->gart.pages[p] = NULL;
 161                         rdev->gart.pages_addr[p] = rdev->dummy_page.addr;
 162                         page_base = rdev->gart.pages_addr[p];
 163                         for (j = 0; j < (PAGE_SIZE / RADEON_GPU_PAGE_SIZE); j++, t++) {
 164                                 if (rdev->gart.ptr) {
 165                                         radeon_gart_set_page(rdev, t, page_base);
 166                                 }
 167                                 page_base += RADEON_GPU_PAGE_SIZE;
 168                         }
 169                 }
 170         }
 171         mb();
 172         radeon_gart_tlb_flush(rdev);
 173 }
 174
 175 int radeon_gart_bind(struct radeon_device *rdev, unsigned offset,
 176                      int pages, struct page **pagelist, dma_addr_t *dma_addr)
 177 {
 178         unsigned t;
 179         unsigned p;
 180         uint64_t page_base;
 181         int i, j;
 182
 183         if (!rdev->gart.ready) {
 184                 WARN(1, "trying to bind memory to uninitialized GART !\n");
 185                 return -EINVAL;
 186         }
 187         t = offset / RADEON_GPU_PAGE_SIZE;
 188         p = t / (PAGE_SIZE / RADEON_GPU_PAGE_SIZE);
 189
 190         for (i = 0; i < pages; i++, p++) {
 191                 rdev->gart.pages_addr[p] = dma_addr[i];
 192                 rdev->gart.pages[p] = pagelist[i];
 193                 if (rdev->gart.ptr) {
 194                         page_base = rdev->gart.pages_addr[p];
 195                         for (j = 0; j < (PAGE_SIZE / RADEON_GPU_PAGE_SIZE); j++, t++) {
 196                                 radeon_gart_set_page(rdev, t, page_base);
 197                                 page_base += RADEON_GPU_PAGE_SIZE;
 198                         }
 199                 }
 200         }
 201         mb();
 202         radeon_gart_tlb_flush(rdev);
 203         return 0;
 204 }
 205
 206 void radeon_gart_restore(struct radeon_device *rdev)
 207 {
 208         int i, j, t;
 209         u64 page_base;
 210
 211         if (!rdev->gart.ptr) {
 212                 return;
 213         }
 214         for (i = 0, t = 0; i < rdev->gart.num_cpu_pages; i++) {
 215                 page_base = rdev->gart.pages_addr[i];
 216                 for (j = 0; j < (PAGE_SIZE / RADEON_GPU_PAGE_SIZE); j++, t++) {
 217                         radeon_gart_set_page(rdev, t, page_base);
 218                         page_base += RADEON_GPU_PAGE_SIZE;
 219                 }
 220         }
 221         mb();
 222         radeon_gart_tlb_flush(rdev);
 223 }
 224
 225 int radeon_gart_init(struct radeon_device *rdev)
 226 {
 227         int r, i;
 228
 229         if (rdev->gart.pages) {
 230                 return 0;
 231         }
 232         /* We need PAGE_SIZE >= RADEON_GPU_PAGE_SIZE */
 233         if (PAGE_SIZE < RADEON_GPU_PAGE_SIZE) {
 234                 DRM_ERROR("Page size is smaller than GPU page size!\n");
 235                 return -EINVAL;
 236         }
 237         r = radeon_dummy_page_init(rdev);
 238         if (r)
 239                 return r;
 240         /* Compute table size */
 241         rdev->gart.num_cpu_pages = rdev->mc.gtt_size / PAGE_SIZE;
 242         rdev->gart.num_gpu_pages = rdev->mc.gtt_size / RADEON_GPU_PAGE_SIZE;
 243         DRM_INFO("GART: num cpu pages %u, num gpu pages %u\n",
 244                  rdev->gart.num_cpu_pages, rdev->gart.num_gpu_pages);
 245         /* Allocate pages table */
 246         rdev->gart.pages = kzalloc(sizeof(void *) * rdev->gart.num_cpu_pages,
 247                                    GFP_KERNEL);
 248         if (rdev->gart.pages == NULL) {
 249                 radeon_gart_fini(rdev);
 250                 return -ENOMEM;
 251         }
 252         rdev->gart.pages_addr = kzalloc(sizeof(dma_addr_t) *
 253                                         rdev->gart.num_cpu_pages, GFP_KERNEL);
 254         if (rdev->gart.pages_addr == NULL) {
 255                 radeon_gart_fini(rdev);
 256                 return -ENOMEM;
 257         }
 258         /* set GART entry to point to the dummy page by default */
 259         for (i = 0; i < rdev->gart.num_cpu_pages; i++) {
 260                 rdev->gart.pages_addr[i] = rdev->dummy_page.addr;
 261         }
 262         return 0;
 263 }
 264
 265 void radeon_gart_fini(struct radeon_device *rdev)
 266 {
 267         if (rdev->gart.pages && rdev->gart.pages_addr && rdev->gart.ready) {
 268                 /* unbind pages */
 269                 radeon_gart_unbind(rdev, 0, rdev->gart.num_cpu_pages);
 270         }
 271         rdev->gart.ready = false;
 272         kfree(rdev->gart.pages);
 273         kfree(rdev->gart.pages_addr);
 274         rdev->gart.pages = NULL;
 275         rdev->gart.pages_addr = NULL;
 276
 277         radeon_dummy_page_fini(rdev);
 278 }
 279
 280 /*
 281  * vm helpers
 282  *
 283  * TODO bind a default page at vm initialization for default address
 284  */
 285 int radeon_vm_manager_init(struct radeon_device *rdev)
 286 {
 287         int r;
 288
 289         rdev->vm_manager.enabled = false;
 290
 291         /* mark first vm as always in use, it's the system one */
 292         /* allocate enough for 2 full VM pts */
 293         r = radeon_sa_bo_manager_init(rdev, &rdev->vm_manager.sa_manager,
 294                                       rdev->vm_manager.max_pfn * 8 * 2,
 295                                       RADEON_GEM_DOMAIN_VRAM);
 296         if (r) {
 297                 dev_err(rdev->dev, "failed to allocate vm bo (%dKB)\n",
 298                         (rdev->vm_manager.max_pfn * 8) >> 10);
 299                 return r;
 300         }
 301
 302         r = rdev->vm_manager.funcs->init(rdev);
 303         if (r == 0)
 304                 rdev->vm_manager.enabled = true;
 305
 306         return r;
 307 }
 308
 309 /* cs mutex must be lock */
 310 static void radeon_vm_unbind_locked(struct radeon_device *rdev,
 311                                     struct radeon_vm *vm)
 312 {
 313         struct radeon_bo_va *bo_va;
 314
 315         if (vm->id == -1) {
 316                 return;
 317         }
 318
 319         /* wait for vm use to end */
 320         if (vm->fence) {
 321                 radeon_fence_wait(vm->fence, false);
 322                 radeon_fence_unref(&vm->fence);
 323         }
 324
 325         /* hw unbind */
 326         rdev->vm_manager.funcs->unbind(rdev, vm);
 327         rdev->vm_manager.use_bitmap &= ~(1 << vm->id);
 328         list_del_init(&vm->list);
 329         vm->id = -1;
 330         radeon_sa_bo_free(rdev, &vm->sa_bo, NULL);
 331         vm->pt = NULL;
 332
 333         list_for_each_entry(bo_va, &vm->va, vm_list) {
 334                 bo_va->valid = false;
 335         }
 336 }
 337
 338 void radeon_vm_manager_fini(struct radeon_device *rdev)
 339 {
 340         if (rdev->vm_manager.sa_manager.bo == NULL)
 341                 return;
 342         radeon_vm_manager_suspend(rdev);
 343         rdev->vm_manager.funcs->fini(rdev);
 344         radeon_sa_bo_manager_fini(rdev, &rdev->vm_manager.sa_manager);
 345         rdev->vm_manager.enabled = false;
 346 }
 347
 348 int radeon_vm_manager_start(struct radeon_device *rdev)
 349 {
 350         if (rdev->vm_manager.sa_manager.bo == NULL) {
 351                 return -EINVAL;
 352         }
 353         return radeon_sa_bo_manager_start(rdev, &rdev->vm_manager.sa_manager);
 354 }
 355
 356 int radeon_vm_manager_suspend(struct radeon_device *rdev)
 357 {
 358         struct radeon_vm *vm, *tmp;
 359
 360         radeon_mutex_lock(&rdev->cs_mutex);
 361         /* unbind all active vm */
 362         list_for_each_entry_safe(vm, tmp, &rdev->vm_manager.lru_vm, list) {
 363                 radeon_vm_unbind_locked(rdev, vm);
 364         }
 365         rdev->vm_manager.funcs->fini(rdev);
 366         radeon_mutex_unlock(&rdev->cs_mutex);
 367         return radeon_sa_bo_manager_suspend(rdev, &rdev->vm_manager.sa_manager);
 368 }
 369
 370 /* cs mutex must be lock */
 371 void radeon_vm_unbind(struct radeon_device *rdev, struct radeon_vm *vm)
 372 {
 373         mutex_lock(&vm->mutex);
 374         radeon_vm_unbind_locked(rdev, vm);
 375         mutex_unlock(&vm->mutex);
 376 }
 377
 378 /* cs mutex must be lock & vm mutex must be lock */
 379 int radeon_vm_bind(struct radeon_device *rdev, struct radeon_vm *vm)
 380 {
 381         struct radeon_vm *vm_evict;
 382         unsigned i;
 383         int id = -1, r;
 384
 385         if (vm == NULL) {
 386                 return -EINVAL;
 387         }
 388
 389         if (vm->id != -1) {
 390                 /* update lru */
 391                 list_del_init(&vm->list);
 392                 list_add_tail(&vm->list, &rdev->vm_manager.lru_vm);
 393                 return 0;
 394         }
 395
 396 retry:
 397         r = radeon_sa_bo_new(rdev, &rdev->vm_manager.sa_manager, &vm->sa_bo,
 398                              RADEON_GPU_PAGE_ALIGN(vm->last_pfn * 8),
 399                              RADEON_GPU_PAGE_SIZE, false);
 400         if (r) {
 401                 if (list_empty(&rdev->vm_manager.lru_vm)) {
 402                         return r;
 403                 }
 404                 vm_evict = list_first_entry(&rdev->vm_manager.lru_vm, struct radeon_vm, list);
 405                 radeon_vm_unbind(rdev, vm_evict);
 406                 goto retry;
 407         }
 408         vm->pt = radeon_sa_bo_cpu_addr(vm->sa_bo);
 409         vm->pt_gpu_addr = radeon_sa_bo_gpu_addr(vm->sa_bo);
 410         memset(vm->pt, 0, RADEON_GPU_PAGE_ALIGN(vm->last_pfn * 8));
 411
 412 retry_id:
 413         /* search for free vm */
 414         for (i = 0; i < rdev->vm_manager.nvm; i++) {
 415                 if (!(rdev->vm_manager.use_bitmap & (1 << i))) {
 416                         id = i;
 417                         break;
 418                 }
 419         }
 420         /* evict vm if necessary */
 421         if (id == -1) {
 422                 vm_evict = list_first_entry(&rdev->vm_manager.lru_vm, struct radeon_vm, list);
 423                 radeon_vm_unbind(rdev, vm_evict);
 424                 goto retry_id;
 425         }
 426
 427         /* do hw bind */
 428         r = rdev->vm_manager.funcs->bind(rdev, vm, id);
 429         if (r) {
 430                 radeon_sa_bo_free(rdev, &vm->sa_bo, NULL);
 431                 return r;
 432         }
 433         rdev->vm_manager.use_bitmap |= 1 << id;
 434         vm->id = id;
 435         list_add_tail(&vm->list, &rdev->vm_manager.lru_vm);
 436         return radeon_vm_bo_update_pte(rdev, vm, rdev->ring_tmp_bo.bo,
 437                                        &rdev->ring_tmp_bo.bo->tbo.mem);
 438 }
 439
 440 /* object have to be reserved */
 441 int radeon_vm_bo_add(struct radeon_device *rdev,
 442                      struct radeon_vm *vm,
 443                      struct radeon_bo *bo,
 444                      uint64_t offset,
 445                      uint32_t flags)
 446 {
 447         struct radeon_bo_va *bo_va, *tmp;
 448         struct list_head *head;
 449         uint64_t size = radeon_bo_size(bo), last_offset = 0;
 450         unsigned last_pfn;
 451
 452         bo_va = kzalloc(sizeof(struct radeon_bo_va), GFP_KERNEL);
 453         if (bo_va == NULL) {
 454                 return -ENOMEM;
 455         }
 456         bo_va->vm = vm;
 457         bo_va->bo = bo;
 458         bo_va->soffset = offset;
 459         bo_va->eoffset = offset + size;
 460         bo_va->flags = flags;
 461         bo_va->valid = false;
 462         INIT_LIST_HEAD(&bo_va->bo_list);
 463         INIT_LIST_HEAD(&bo_va->vm_list);
 464         /* make sure object fit at this offset */
 465         if (bo_va->soffset >= bo_va->eoffset) {
 466                 kfree(bo_va);
 467                 return -EINVAL;
 468         }
 469
 470         last_pfn = bo_va->eoffset / RADEON_GPU_PAGE_SIZE;
 471         if (last_pfn > rdev->vm_manager.max_pfn) {
 472                 kfree(bo_va);
 473                 dev_err(rdev->dev, "va above limit (0x%08X > 0x%08X)\n",
 474                         last_pfn, rdev->vm_manager.max_pfn);
 475                 return -EINVAL;
 476         }
 477
 478         mutex_lock(&vm->mutex);
 479         if (last_pfn > vm->last_pfn) {
 480                 /* release mutex and lock in right order */
 481                 mutex_unlock(&vm->mutex);
 482                 radeon_mutex_lock(&rdev->cs_mutex);
 483                 mutex_lock(&vm->mutex);
 484                 /* and check again */
 485                 if (last_pfn > vm->last_pfn) {
 486                         /* grow va space 32M by 32M */
 487                         unsigned align = ((32 << 20) >> 12) - 1;
 488                         radeon_vm_unbind_locked(rdev, vm);
 489                         vm->last_pfn = (last_pfn + align) & ~align;
 490                 }
 491                 radeon_mutex_unlock(&rdev->cs_mutex);
 492         }
 493         head = &vm->va;
 494         last_offset = 0;
 495         list_for_each_entry(tmp, &vm->va, vm_list) {
 496                 if (bo_va->soffset >= last_offset && bo_va->eoffset < tmp->soffset) {
 497                         /* bo can be added before this one */
 498                         break;
 499                 }
 500                 if (bo_va->soffset >= tmp->soffset && bo_va->soffset < tmp->eoffset) {
 501                         /* bo and tmp overlap, invalid offset */
 502                         dev_err(rdev->dev, "bo %p va 0x%08X conflict with (bo %p 0x%08X 0x%08X)\n",
 503                                 bo, (unsigned)bo_va->soffset, tmp->bo,
 504                                 (unsigned)tmp->soffset, (unsigned)tmp->eoffset);
 505                         kfree(bo_va);
 506                         mutex_unlock(&vm->mutex);
 507                         return -EINVAL;
 508                 }
 509                 last_offset = tmp->eoffset;
 510                 head = &tmp->vm_list;
 511         }
 512         list_add(&bo_va->vm_list, head);
 513         list_add_tail(&bo_va->bo_list, &bo->va);
 514         mutex_unlock(&vm->mutex);
 515         return 0;
 516 }
 517
 518 static u64 radeon_vm_get_addr(struct radeon_device *rdev,
 519                               struct ttm_mem_reg *mem,
 520                               unsigned pfn)
 521 {
 522         u64 addr = 0;
 523
 524         switch (mem->mem_type) {
 525         case TTM_PL_VRAM:
 526                 addr = (mem->start << PAGE_SHIFT);
 527                 addr += pfn * RADEON_GPU_PAGE_SIZE;
 528                 addr += rdev->vm_manager.vram_base_offset;
 529                 break;
 530         case TTM_PL_TT:
 531                 /* offset inside page table */
 532                 addr = mem->start << PAGE_SHIFT;
 533                 addr += pfn * RADEON_GPU_PAGE_SIZE;
 534                 addr = addr >> PAGE_SHIFT;
 535                 /* page table offset */
 536                 addr = rdev->gart.pages_addr[addr];
 537                 /* in case cpu page size != gpu page size*/
 538                 addr += (pfn * RADEON_GPU_PAGE_SIZE) & (~PAGE_MASK);
 539                 break;
 540         default:
 541                 break;
 542         }
 543         return addr;
 544 }
 545
 546 /* object have to be reserved & cs mutex took & vm mutex took */
 547 int radeon_vm_bo_update_pte(struct radeon_device *rdev,
 548                             struct radeon_vm *vm,
 549                             struct radeon_bo *bo,
 550                             struct ttm_mem_reg *mem)
 551 {
 552         struct radeon_bo_va *bo_va;
 553         unsigned ngpu_pages, i;
 554         uint64_t addr = 0, pfn;
 555         uint32_t flags;
 556
 557         /* nothing to do if vm isn't bound */
 558         if (vm->id == -1)
 559                 return 0;
 560
 561         bo_va = radeon_bo_va(bo, vm);
 562         if (bo_va == NULL) {
 563                 dev_err(rdev->dev, "bo %p not in vm %p\n", bo, vm);
 564                 return -EINVAL;
 565         }
 566
 567         if (bo_va->valid)
 568                 return 0;
 569
 570         ngpu_pages = radeon_bo_ngpu_pages(bo);
 571         bo_va->flags &= ~RADEON_VM_PAGE_VALID;
 572         bo_va->flags &= ~RADEON_VM_PAGE_SYSTEM;
 573         if (mem) {
 574                 if (mem->mem_type != TTM_PL_SYSTEM) {
 575                         bo_va->flags |= RADEON_VM_PAGE_VALID;
 576                         bo_va->valid = true;
 577                 }
 578                 if (mem->mem_type == TTM_PL_TT) {
 579                         bo_va->flags |= RADEON_VM_PAGE_SYSTEM;
 580                 }
 581         }
 582         pfn = bo_va->soffset / RADEON_GPU_PAGE_SIZE;
 583         flags = rdev->vm_manager.funcs->page_flags(rdev, bo_va->vm, bo_va->flags);
 584         for (i = 0, addr = 0; i < ngpu_pages; i++) {
 585                 if (mem && bo_va->valid) {
 586                         addr = radeon_vm_get_addr(rdev, mem, i);
 587                 }
 588                 rdev->vm_manager.funcs->set_page(rdev, bo_va->vm, i + pfn, addr, flags);
 589         }
 590         rdev->vm_manager.funcs->tlb_flush(rdev, bo_va->vm);
 591         return 0;
 592 }
 593
 594 /* object have to be reserved */
 595 int radeon_vm_bo_rmv(struct radeon_device *rdev,
 596                      struct radeon_vm *vm,
 597                      struct radeon_bo *bo)
 598 {
 599         struct radeon_bo_va *bo_va;
 600
 601         bo_va = radeon_bo_va(bo, vm);
 602         if (bo_va == NULL)
 603                 return 0;
 604
 605         radeon_mutex_lock(&rdev->cs_mutex);
 606         mutex_lock(&vm->mutex);
 607         radeon_vm_bo_update_pte(rdev, vm, bo, NULL);
 608         radeon_mutex_unlock(&rdev->cs_mutex);
 609         list_del(&bo_va->vm_list);
 610         mutex_unlock(&vm->mutex);
 611         list_del(&bo_va->bo_list);
 612
 613         kfree(bo_va);
 614         return 0;
 615 }
 616
 617 void radeon_vm_bo_invalidate(struct radeon_device *rdev,
 618                              struct radeon_bo *bo)
 619 {
 620         struct radeon_bo_va *bo_va;
 621
 622         BUG_ON(!atomic_read(&bo->tbo.reserved));
 623         list_for_each_entry(bo_va, &bo->va, bo_list) {
 624                 bo_va->valid = false;
 625         }
 626 }
 627
 628 int radeon_vm_init(struct radeon_device *rdev, struct radeon_vm *vm)
 629 {
 630         int r;
 631
 632         vm->id = -1;
 633         vm->fence = NULL;
 634         mutex_init(&vm->mutex);
 635         INIT_LIST_HEAD(&vm->list);
 636         INIT_LIST_HEAD(&vm->va);
 637         /* SI requires equal sized PTs for all VMs, so always set
 638          * last_pfn to max_pfn.  cayman allows variable sized
 639          * pts so we can grow then as needed.  Once we switch
 640          * to two level pts we can unify this again.
 641          */
 642         if (rdev->family >= CHIP_TAHITI)
 643                 vm->last_pfn = rdev->vm_manager.max_pfn;
 644         else
 645                 vm->last_pfn = 0;
 646         /* map the ib pool buffer at 0 in virtual address space, set
 647          * read only
 648          */
 649         r = radeon_vm_bo_add(rdev, vm, rdev->ring_tmp_bo.bo, 0,
 650                              RADEON_VM_PAGE_READABLE | RADEON_VM_PAGE_SNOOPED);
 651         return r;
 652 }
 653
 654 void radeon_vm_fini(struct radeon_device *rdev, struct radeon_vm *vm)
 655 {
 656         struct radeon_bo_va *bo_va, *tmp;
 657         int r;
 658
 659         radeon_mutex_lock(&rdev->cs_mutex);
 660         mutex_lock(&vm->mutex);
 661         radeon_vm_unbind_locked(rdev, vm);
 662         radeon_mutex_unlock(&rdev->cs_mutex);
 663
 664         /* remove all bo */
 665         r = radeon_bo_reserve(rdev->ring_tmp_bo.bo, false);
 666         if (!r) {
 667                 bo_va = radeon_bo_va(rdev->ring_tmp_bo.bo, vm);
 668                 list_del_init(&bo_va->bo_list);
 669                 list_del_init(&bo_va->vm_list);
 670                 radeon_bo_unreserve(rdev->ring_tmp_bo.bo);
 671                 kfree(bo_va);
 672         }
 673         if (!list_empty(&vm->va)) {
 674                 dev_err(rdev->dev, "still active bo inside vm\n");
 675         }
 676         list_for_each_entry_safe(bo_va, tmp, &vm->va, vm_list) {
 677                 list_del_init(&bo_va->vm_list);
 678                 r = radeon_bo_reserve(bo_va->bo, false);
 679                 if (!r) {
 680                         list_del_init(&bo_va->bo_list);
 681                         radeon_bo_unreserve(bo_va->bo);
 682                         kfree(bo_va);
 683                 }
 684         }
 685         mutex_unlock(&vm->mutex);
 686 }