drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c

   1 /*
   2  * Copyright 2009 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
   7  * "Software"), to deal in the Software without restriction, including
   8  * without limitation the rights to use, copy, modify, merge, publish,
   9  * distribute, sub license, and/or sell copies of the Software, and to
  10  * permit persons to whom the Software is furnished to do so, subject to
  11  * the following conditions:
  12  *
  13  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  14  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  15  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  16  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
  17  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  18  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  19  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  20  *
  21  * The above copyright notice and this permission notice (including the
  22  * next paragraph) shall be included in all copies or substantial portions
  23  * of the Software.
  24  *
  25  */
  26 /*
  27  * Authors:
  28  *    Jerome Glisse <glisse@freedesktop.org>
  29  *    Thomas Hellstrom <thomas-at-tungstengraphics-dot-com>
  30  *    Dave Airlie
  31  */
  32 #include <drm/ttm/ttm_bo_api.h>
  33 #include <drm/ttm/ttm_bo_driver.h>
  34 #include <drm/ttm/ttm_placement.h>
  35 #include <drm/ttm/ttm_module.h>
  36 #include <drm/ttm/ttm_page_alloc.h>
  37 #include <drm/drmP.h>
  38 #include <drm/amdgpu_drm.h>
  39 #include <linux/seq_file.h>
  40 #include <linux/slab.h>
  41 #include <linux/swiotlb.h>
  42 #include <linux/swap.h>
  43 #include <linux/pagemap.h>
  44 #include <linux/debugfs.h>
  45 #include "amdgpu.h"
  46 #include "amdgpu_trace.h"
  47 #include "bif/bif_4_1_d.h"
  48
  49 #define DRM_FILE_PAGE_OFFSET (0x100000000ULL >> PAGE_SHIFT)
  50
  51 static int amdgpu_map_buffer(struct ttm_buffer_object *bo,
  52                              struct ttm_mem_reg *mem, unsigned num_pages,
  53                              uint64_t offset, unsigned window,
  54                              struct amdgpu_ring *ring,
  55                              uint64_t *addr);
  56
  57 static int amdgpu_ttm_debugfs_init(struct amdgpu_device *adev);
  58 static void amdgpu_ttm_debugfs_fini(struct amdgpu_device *adev);
  59
  60 /*
  61  * Global memory.
  62  */
  63 static int amdgpu_ttm_mem_global_init(struct drm_global_reference *ref)
  64 {
  65         return ttm_mem_global_init(ref->object);
  66 }
  67
  68 static void amdgpu_ttm_mem_global_release(struct drm_global_reference *ref)
  69 {
  70         ttm_mem_global_release(ref->object);
  71 }
  72
  73 static int amdgpu_ttm_global_init(struct amdgpu_device *adev)
  74 {
  75         struct drm_global_reference *global_ref;
  76         struct amdgpu_ring *ring;
  77         struct amd_sched_rq *rq;
  78         int r;
  79
  80         adev->mman.mem_global_referenced = false;
  81         global_ref = &adev->mman.mem_global_ref;
  82         global_ref->global_type = DRM_GLOBAL_TTM_MEM;
  83         global_ref->size = sizeof(struct ttm_mem_global);
  84         global_ref->init = &amdgpu_ttm_mem_global_init;
  85         global_ref->release = &amdgpu_ttm_mem_global_release;
  86         r = drm_global_item_ref(global_ref);
  87         if (r) {
  88                 DRM_ERROR("Failed setting up TTM memory accounting "
  89                           "subsystem.\n");
  90                 goto error_mem;
  91         }
  92
  93         adev->mman.bo_global_ref.mem_glob =
  94                 adev->mman.mem_global_ref.object;
  95         global_ref = &adev->mman.bo_global_ref.ref;
  96         global_ref->global_type = DRM_GLOBAL_TTM_BO;
  97         global_ref->size = sizeof(struct ttm_bo_global);
  98         global_ref->init = &ttm_bo_global_init;
  99         global_ref->release = &ttm_bo_global_release;
 100         r = drm_global_item_ref(global_ref);
 101         if (r) {
 102                 DRM_ERROR("Failed setting up TTM BO subsystem.\n");
 103                 goto error_bo;
 104         }
 105
 106         mutex_init(&adev->mman.gtt_window_lock);
 107
 108         ring = adev->mman.buffer_funcs_ring;
 109         rq = &ring->sched.sched_rq[AMD_SCHED_PRIORITY_KERNEL];
 110         r = amd_sched_entity_init(&ring->sched, &adev->mman.entity,
 111                                   rq, amdgpu_sched_jobs);
 112         if (r) {
 113                 DRM_ERROR("Failed setting up TTM BO move run queue.\n");
 114                 goto error_entity;
 115         }
 116
 117         adev->mman.mem_global_referenced = true;
 118
 119         return 0;
 120
 121 error_entity:
 122         drm_global_item_unref(&adev->mman.bo_global_ref.ref);
 123 error_bo:
 124         drm_global_item_unref(&adev->mman.mem_global_ref);
 125 error_mem:
 126         return r;
 127 }
 128
 129 static void amdgpu_ttm_global_fini(struct amdgpu_device *adev)
 130 {
 131         if (adev->mman.mem_global_referenced) {
 132                 amd_sched_entity_fini(adev->mman.entity.sched,
 133                                       &adev->mman.entity);
 134                 mutex_destroy(&adev->mman.gtt_window_lock);
 135                 drm_global_item_unref(&adev->mman.bo_global_ref.ref);
 136                 drm_global_item_unref(&adev->mman.mem_global_ref);
 137                 adev->mman.mem_global_referenced = false;
 138         }
 139 }
 140
 141 static int amdgpu_invalidate_caches(struct ttm_bo_device *bdev, uint32_t flags)
 142 {
 143         return 0;
 144 }
 145
 146 static int amdgpu_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
 147                                 struct ttm_mem_type_manager *man)
 148 {
 149         struct amdgpu_device *adev;
 150
 151         adev = amdgpu_ttm_adev(bdev);
 152
 153         switch (type) {
 154         case TTM_PL_SYSTEM:
 155                 /* System memory */
 156                 man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
 157                 man->available_caching = TTM_PL_MASK_CACHING;
 158                 man->default_caching = TTM_PL_FLAG_CACHED;
 159                 break;
 160         case TTM_PL_TT:
 161                 man->func = &amdgpu_gtt_mgr_func;
 162                 man->gpu_offset = adev->mc.gart_start;
 163                 man->available_caching = TTM_PL_MASK_CACHING;
 164                 man->default_caching = TTM_PL_FLAG_CACHED;
 165                 man->flags = TTM_MEMTYPE_FLAG_MAPPABLE | TTM_MEMTYPE_FLAG_CMA;
 166                 break;
 167         case TTM_PL_VRAM:
 168                 /* "On-card" video ram */
 169                 man->func = &amdgpu_vram_mgr_func;
 170                 man->gpu_offset = adev->mc.vram_start;
 171                 man->flags = TTM_MEMTYPE_FLAG_FIXED |
 172                              TTM_MEMTYPE_FLAG_MAPPABLE;
 173                 man->available_caching = TTM_PL_FLAG_UNCACHED | TTM_PL_FLAG_WC;
 174                 man->default_caching = TTM_PL_FLAG_WC;
 175                 break;
 176         case AMDGPU_PL_GDS:
 177         case AMDGPU_PL_GWS:
 178         case AMDGPU_PL_OA:
 179                 /* On-chip GDS memory*/
 180                 man->func = &ttm_bo_manager_func;
 181                 man->gpu_offset = 0;
 182                 man->flags = TTM_MEMTYPE_FLAG_FIXED | TTM_MEMTYPE_FLAG_CMA;
 183                 man->available_caching = TTM_PL_FLAG_UNCACHED;
 184                 man->default_caching = TTM_PL_FLAG_UNCACHED;
 185                 break;
 186         default:
 187                 DRM_ERROR("Unsupported memory type %u\n", (unsigned)type);
 188                 return -EINVAL;
 189         }
 190         return 0;
 191 }
 192
 193 static void amdgpu_evict_flags(struct ttm_buffer_object *bo,
 194                                 struct ttm_placement *placement)
 195 {
 196         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
 197         struct amdgpu_bo *abo;
 198         static const struct ttm_place placements = {
 199                 .fpfn = 0,
 200                 .lpfn = 0,
 201                 .flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM
 202         };
 203
 204         if (!amdgpu_ttm_bo_is_amdgpu_bo(bo)) {
 205                 placement->placement = &placements;
 206                 placement->busy_placement = &placements;
 207                 placement->num_placement = 1;
 208                 placement->num_busy_placement = 1;
 209                 return;
 210         }
 211         abo = container_of(bo, struct amdgpu_bo, tbo);
 212         switch (bo->mem.mem_type) {
 213         case TTM_PL_VRAM:
 214                 if (adev->mman.buffer_funcs &&
 215                     adev->mman.buffer_funcs_ring &&
 216                     adev->mman.buffer_funcs_ring->ready == false) {
 217                         amdgpu_ttm_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_CPU);
 218                 } else if (adev->mc.visible_vram_size < adev->mc.real_vram_size &&
 219                            !(abo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)) {
 220                         unsigned fpfn = adev->mc.visible_vram_size >> PAGE_SHIFT;
 221                         struct drm_mm_node *node = bo->mem.mm_node;
 222                         unsigned long pages_left;
 223
 224                         for (pages_left = bo->mem.num_pages;
 225                              pages_left;
 226                              pages_left -= node->size, node++) {
 227                                 if (node->start < fpfn)
 228                                         break;
 229                         }
 230
 231                         if (!pages_left)
 232                                 goto gtt;
 233
 234                         /* Try evicting to the CPU inaccessible part of VRAM
 235                          * first, but only set GTT as busy placement, so this
 236                          * BO will be evicted to GTT rather than causing other
 237                          * BOs to be evicted from VRAM
 238                          */
 239                         amdgpu_ttm_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM |
 240                                                          AMDGPU_GEM_DOMAIN_GTT);
 241                         abo->placements[0].fpfn = fpfn;
 242                         abo->placements[0].lpfn = 0;
 243                         abo->placement.busy_placement = &abo->placements[1];
 244                         abo->placement.num_busy_placement = 1;
 245                 } else {
 246 gtt:
 247                         amdgpu_ttm_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_GTT);
 248                 }
 249                 break;
 250         case TTM_PL_TT:
 251         default:
 252                 amdgpu_ttm_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_CPU);
 253         }
 254         *placement = abo->placement;
 255 }
 256
 257 static int amdgpu_verify_access(struct ttm_buffer_object *bo, struct file *filp)
 258 {
 259         struct amdgpu_bo *abo = container_of(bo, struct amdgpu_bo, tbo);
 260
 261         if (amdgpu_ttm_tt_get_usermm(bo->ttm))
 262                 return -EPERM;
 263         return drm_vma_node_verify_access(&abo->gem_base.vma_node,
 264                                           filp->private_data);
 265 }
 266
 267 static void amdgpu_move_null(struct ttm_buffer_object *bo,
 268                              struct ttm_mem_reg *new_mem)
 269 {
 270         struct ttm_mem_reg *old_mem = &bo->mem;
 271
 272         BUG_ON(old_mem->mm_node != NULL);
 273         *old_mem = *new_mem;
 274         new_mem->mm_node = NULL;
 275 }
 276
 277 static uint64_t amdgpu_mm_node_addr(struct ttm_buffer_object *bo,
 278                                     struct drm_mm_node *mm_node,
 279                                     struct ttm_mem_reg *mem)
 280 {
 281         uint64_t addr = 0;
 282
 283         if (mem->mem_type != TTM_PL_TT ||
 284             amdgpu_gtt_mgr_is_allocated(mem)) {
 285                 addr = mm_node->start << PAGE_SHIFT;
 286                 addr += bo->bdev->man[mem->mem_type].gpu_offset;
 287         }
 288         return addr;
 289 }
 290
 291 static int amdgpu_move_blit(struct ttm_buffer_object *bo,
 292                             bool evict, bool no_wait_gpu,
 293                             struct ttm_mem_reg *new_mem,
 294                             struct ttm_mem_reg *old_mem)
 295 {
 296         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
 297         struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
 298
 299         struct drm_mm_node *old_mm, *new_mm;
 300         uint64_t old_start, old_size, new_start, new_size;
 301         unsigned long num_pages;
 302         struct dma_fence *fence = NULL;
 303         int r;
 304
 305         BUILD_BUG_ON((PAGE_SIZE % AMDGPU_GPU_PAGE_SIZE) != 0);
 306
 307         if (!ring->ready) {
 308                 DRM_ERROR("Trying to move memory with ring turned off.\n");
 309                 return -EINVAL;
 310         }
 311
 312         old_mm = old_mem->mm_node;
 313         old_size = old_mm->size;
 314         old_start = amdgpu_mm_node_addr(bo, old_mm, old_mem);
 315
 316         new_mm = new_mem->mm_node;
 317         new_size = new_mm->size;
 318         new_start = amdgpu_mm_node_addr(bo, new_mm, new_mem);
 319
 320         num_pages = new_mem->num_pages;
 321         mutex_lock(&adev->mman.gtt_window_lock);
 322         while (num_pages) {
 323                 unsigned long cur_pages = min(min(old_size, new_size),
 324                                               (u64)AMDGPU_GTT_MAX_TRANSFER_SIZE);
 325                 uint64_t from = old_start, to = new_start;
 326                 struct dma_fence *next;
 327
 328                 if (old_mem->mem_type == TTM_PL_TT &&
 329                     !amdgpu_gtt_mgr_is_allocated(old_mem)) {
 330                         r = amdgpu_map_buffer(bo, old_mem, cur_pages,
 331                                               old_start, 0, ring, &from);
 332                         if (r)
 333                                 goto error;
 334                 }
 335
 336                 if (new_mem->mem_type == TTM_PL_TT &&
 337                     !amdgpu_gtt_mgr_is_allocated(new_mem)) {
 338                         r = amdgpu_map_buffer(bo, new_mem, cur_pages,
 339                                               new_start, 1, ring, &to);
 340                         if (r)
 341                                 goto error;
 342                 }
 343
 344                 r = amdgpu_copy_buffer(ring, from, to,
 345                                        cur_pages * PAGE_SIZE,
 346                                        bo->resv, &next, false, true);
 347                 if (r)
 348                         goto error;
 349
 350                 dma_fence_put(fence);
 351                 fence = next;
 352
 353                 num_pages -= cur_pages;
 354                 if (!num_pages)
 355                         break;
 356
 357                 old_size -= cur_pages;
 358                 if (!old_size) {
 359                         old_start = amdgpu_mm_node_addr(bo, ++old_mm, old_mem);
 360                         old_size = old_mm->size;
 361                 } else {
 362                         old_start += cur_pages * PAGE_SIZE;
 363                 }
 364
 365                 new_size -= cur_pages;
 366                 if (!new_size) {
 367                         new_start = amdgpu_mm_node_addr(bo, ++new_mm, new_mem);
 368                         new_size = new_mm->size;
 369                 } else {
 370                         new_start += cur_pages * PAGE_SIZE;
 371                 }
 372         }
 373         mutex_unlock(&adev->mman.gtt_window_lock);
 374
 375         r = ttm_bo_pipeline_move(bo, fence, evict, new_mem);
 376         dma_fence_put(fence);
 377         return r;
 378
 379 error:
 380         mutex_unlock(&adev->mman.gtt_window_lock);
 381
 382         if (fence)
 383                 dma_fence_wait(fence, false);
 384         dma_fence_put(fence);
 385         return r;
 386 }
 387
 388 static int amdgpu_move_vram_ram(struct ttm_buffer_object *bo,
 389                                 bool evict, bool interruptible,
 390                                 bool no_wait_gpu,
 391                                 struct ttm_mem_reg *new_mem)
 392 {
 393         struct amdgpu_device *adev;
 394         struct ttm_mem_reg *old_mem = &bo->mem;
 395         struct ttm_mem_reg tmp_mem;
 396         struct ttm_place placements;
 397         struct ttm_placement placement;
 398         int r;
 399
 400         adev = amdgpu_ttm_adev(bo->bdev);
 401         tmp_mem = *new_mem;
 402         tmp_mem.mm_node = NULL;
 403         placement.num_placement = 1;
 404         placement.placement = &placements;
 405         placement.num_busy_placement = 1;
 406         placement.busy_placement = &placements;
 407         placements.fpfn = 0;
 408         placements.lpfn = 0;
 409         placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
 410         r = ttm_bo_mem_space(bo, &placement, &tmp_mem,
 411                              interruptible, no_wait_gpu);
 412         if (unlikely(r)) {
 413                 return r;
 414         }
 415
 416         r = ttm_tt_set_placement_caching(bo->ttm, tmp_mem.placement);
 417         if (unlikely(r)) {
 418                 goto out_cleanup;
 419         }
 420
 421         r = ttm_tt_bind(bo->ttm, &tmp_mem);
 422         if (unlikely(r)) {
 423                 goto out_cleanup;
 424         }
 425         r = amdgpu_move_blit(bo, true, no_wait_gpu, &tmp_mem, old_mem);
 426         if (unlikely(r)) {
 427                 goto out_cleanup;
 428         }
 429         r = ttm_bo_move_ttm(bo, interruptible, no_wait_gpu, new_mem);
 430 out_cleanup:
 431         ttm_bo_mem_put(bo, &tmp_mem);
 432         return r;
 433 }
 434
 435 static int amdgpu_move_ram_vram(struct ttm_buffer_object *bo,
 436                                 bool evict, bool interruptible,
 437                                 bool no_wait_gpu,
 438                                 struct ttm_mem_reg *new_mem)
 439 {
 440         struct amdgpu_device *adev;
 441         struct ttm_mem_reg *old_mem = &bo->mem;
 442         struct ttm_mem_reg tmp_mem;
 443         struct ttm_placement placement;
 444         struct ttm_place placements;
 445         int r;
 446
 447         adev = amdgpu_ttm_adev(bo->bdev);
 448         tmp_mem = *new_mem;
 449         tmp_mem.mm_node = NULL;
 450         placement.num_placement = 1;
 451         placement.placement = &placements;
 452         placement.num_busy_placement = 1;
 453         placement.busy_placement = &placements;
 454         placements.fpfn = 0;
 455         placements.lpfn = 0;
 456         placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
 457         r = ttm_bo_mem_space(bo, &placement, &tmp_mem,
 458                              interruptible, no_wait_gpu);
 459         if (unlikely(r)) {
 460                 return r;
 461         }
 462         r = ttm_bo_move_ttm(bo, interruptible, no_wait_gpu, &tmp_mem);
 463         if (unlikely(r)) {
 464                 goto out_cleanup;
 465         }
 466         r = amdgpu_move_blit(bo, true, no_wait_gpu, new_mem, old_mem);
 467         if (unlikely(r)) {
 468                 goto out_cleanup;
 469         }
 470 out_cleanup:
 471         ttm_bo_mem_put(bo, &tmp_mem);
 472         return r;
 473 }
 474
 475 static int amdgpu_bo_move(struct ttm_buffer_object *bo,
 476                         bool evict, bool interruptible,
 477                         bool no_wait_gpu,
 478                         struct ttm_mem_reg *new_mem)
 479 {
 480         struct amdgpu_device *adev;
 481         struct amdgpu_bo *abo;
 482         struct ttm_mem_reg *old_mem = &bo->mem;
 483         int r;
 484
 485         /* Can't move a pinned BO */
 486         abo = container_of(bo, struct amdgpu_bo, tbo);
 487         if (WARN_ON_ONCE(abo->pin_count > 0))
 488                 return -EINVAL;
 489
 490         adev = amdgpu_ttm_adev(bo->bdev);
 491
 492         if (old_mem->mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
 493                 amdgpu_move_null(bo, new_mem);
 494                 return 0;
 495         }
 496         if ((old_mem->mem_type == TTM_PL_TT &&
 497              new_mem->mem_type == TTM_PL_SYSTEM) ||
 498             (old_mem->mem_type == TTM_PL_SYSTEM &&
 499              new_mem->mem_type == TTM_PL_TT)) {
 500                 /* bind is enough */
 501                 amdgpu_move_null(bo, new_mem);
 502                 return 0;
 503         }
 504         if (adev->mman.buffer_funcs == NULL ||
 505             adev->mman.buffer_funcs_ring == NULL ||
 506             !adev->mman.buffer_funcs_ring->ready) {
 507                 /* use memcpy */
 508                 goto memcpy;
 509         }
 510
 511         if (old_mem->mem_type == TTM_PL_VRAM &&
 512             new_mem->mem_type == TTM_PL_SYSTEM) {
 513                 r = amdgpu_move_vram_ram(bo, evict, interruptible,
 514                                         no_wait_gpu, new_mem);
 515         } else if (old_mem->mem_type == TTM_PL_SYSTEM &&
 516                    new_mem->mem_type == TTM_PL_VRAM) {
 517                 r = amdgpu_move_ram_vram(bo, evict, interruptible,
 518                                             no_wait_gpu, new_mem);
 519         } else {
 520                 r = amdgpu_move_blit(bo, evict, no_wait_gpu, new_mem, old_mem);
 521         }
 522
 523         if (r) {
 524 memcpy:
 525                 r = ttm_bo_move_memcpy(bo, interruptible, no_wait_gpu, new_mem);
 526                 if (r) {
 527                         return r;
 528                 }
 529         }
 530
 531         if (bo->type == ttm_bo_type_device &&
 532             new_mem->mem_type == TTM_PL_VRAM &&
 533             old_mem->mem_type != TTM_PL_VRAM) {
 534                 /* amdgpu_bo_fault_reserve_notify will re-set this if the CPU
 535                  * accesses the BO after it's moved.
 536                  */
 537                 abo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
 538         }
 539
 540         /* update statistics */
 541         atomic64_add((u64)bo->num_pages << PAGE_SHIFT, &adev->num_bytes_moved);
 542         return 0;
 543 }
 544
 545 static int amdgpu_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
 546 {
 547         struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
 548         struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
 549
 550         mem->bus.addr = NULL;
 551         mem->bus.offset = 0;
 552         mem->bus.size = mem->num_pages << PAGE_SHIFT;
 553         mem->bus.base = 0;
 554         mem->bus.is_iomem = false;
 555         if (!(man->flags & TTM_MEMTYPE_FLAG_MAPPABLE))
 556                 return -EINVAL;
 557         switch (mem->mem_type) {
 558         case TTM_PL_SYSTEM:
 559                 /* system memory */
 560                 return 0;
 561         case TTM_PL_TT:
 562                 break;
 563         case TTM_PL_VRAM:
 564                 mem->bus.offset = mem->start << PAGE_SHIFT;
 565                 /* check if it's visible */
 566                 if ((mem->bus.offset + mem->bus.size) > adev->mc.visible_vram_size)
 567                         return -EINVAL;
 568                 mem->bus.base = adev->mc.aper_base;
 569                 mem->bus.is_iomem = true;
 570                 break;
 571         default:
 572                 return -EINVAL;
 573         }
 574         return 0;
 575 }
 576
 577 static void amdgpu_ttm_io_mem_free(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
 578 {
 579 }
 580
 581 static unsigned long amdgpu_ttm_io_mem_pfn(struct ttm_buffer_object *bo,
 582                                            unsigned long page_offset)
 583 {
 584         struct drm_mm_node *mm = bo->mem.mm_node;
 585         uint64_t size = mm->size;
 586         uint64_t offset = page_offset;
 587
 588         page_offset = do_div(offset, size);
 589         mm += offset;
 590         return (bo->mem.bus.base >> PAGE_SHIFT) + mm->start + page_offset;
 591 }
 592
 593 /*
 594  * TTM backend functions.
 595  */
 596 struct amdgpu_ttm_gup_task_list {
 597         struct list_head        list;
 598         struct task_struct      *task;
 599 };
 600
 601 struct amdgpu_ttm_tt {
 602         struct ttm_dma_tt       ttm;
 603         struct amdgpu_device    *adev;
 604         u64                     offset;
 605         uint64_t                userptr;
 606         struct mm_struct        *usermm;
 607         uint32_t                userflags;
 608         spinlock_t              guptasklock;
 609         struct list_head        guptasks;
 610         atomic_t                mmu_invalidations;
 611         struct list_head        list;
 612 };
 613
 614 int amdgpu_ttm_tt_get_user_pages(struct ttm_tt *ttm, struct page **pages)
 615 {
 616         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 617         unsigned int flags = 0;
 618         unsigned pinned = 0;
 619         int r;
 620
 621         if (!(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY))
 622                 flags |= FOLL_WRITE;
 623
 624         if (gtt->userflags & AMDGPU_GEM_USERPTR_ANONONLY) {
 625                 /* check that we only use anonymous memory
 626                    to prevent problems with writeback */
 627                 unsigned long end = gtt->userptr + ttm->num_pages * PAGE_SIZE;
 628                 struct vm_area_struct *vma;
 629
 630                 vma = find_vma(gtt->usermm, gtt->userptr);
 631                 if (!vma || vma->vm_file || vma->vm_end < end)
 632                         return -EPERM;
 633         }
 634
 635         do {
 636                 unsigned num_pages = ttm->num_pages - pinned;
 637                 uint64_t userptr = gtt->userptr + pinned * PAGE_SIZE;
 638                 struct page **p = pages + pinned;
 639                 struct amdgpu_ttm_gup_task_list guptask;
 640
 641                 guptask.task = current;
 642                 spin_lock(&gtt->guptasklock);
 643                 list_add(&guptask.list, &gtt->guptasks);
 644                 spin_unlock(&gtt->guptasklock);
 645
 646                 r = get_user_pages(userptr, num_pages, flags, p, NULL);
 647
 648                 spin_lock(&gtt->guptasklock);
 649                 list_del(&guptask.list);
 650                 spin_unlock(&gtt->guptasklock);
 651
 652                 if (r < 0)
 653                         goto release_pages;
 654
 655                 pinned += r;
 656
 657         } while (pinned < ttm->num_pages);
 658
 659         return 0;
 660
 661 release_pages:
 662         release_pages(pages, pinned, 0);
 663         return r;
 664 }
 665
 666 static void amdgpu_trace_dma_map(struct ttm_tt *ttm)
 667 {
 668         struct amdgpu_device *adev = amdgpu_ttm_adev(ttm->bdev);
 669         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 670         unsigned i;
 671
 672         if (unlikely(trace_amdgpu_ttm_tt_populate_enabled())) {
 673                 for (i = 0; i < ttm->num_pages; i++) {
 674                         trace_amdgpu_ttm_tt_populate(
 675                                 adev,
 676                                 gtt->ttm.dma_address[i],
 677                                 page_to_phys(ttm->pages[i]));
 678                 }
 679         }
 680 }
 681
 682 static void amdgpu_trace_dma_unmap(struct ttm_tt *ttm)
 683 {
 684         struct amdgpu_device *adev = amdgpu_ttm_adev(ttm->bdev);
 685         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 686         unsigned i;
 687
 688         if (unlikely(trace_amdgpu_ttm_tt_unpopulate_enabled())) {
 689                 for (i = 0; i < ttm->num_pages; i++) {
 690                         trace_amdgpu_ttm_tt_unpopulate(
 691                                 adev,
 692                                 gtt->ttm.dma_address[i],
 693                                 page_to_phys(ttm->pages[i]));
 694                 }
 695         }
 696 }
 697
 698 /* prepare the sg table with the user pages */
 699 static int amdgpu_ttm_tt_pin_userptr(struct ttm_tt *ttm)
 700 {
 701         struct amdgpu_device *adev = amdgpu_ttm_adev(ttm->bdev);
 702         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 703         unsigned nents;
 704         int r;
 705
 706         int write = !(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
 707         enum dma_data_direction direction = write ?
 708                 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
 709
 710         r = sg_alloc_table_from_pages(ttm->sg, ttm->pages, ttm->num_pages, 0,
 711                                       ttm->num_pages << PAGE_SHIFT,
 712                                       GFP_KERNEL);
 713         if (r)
 714                 goto release_sg;
 715
 716         r = -ENOMEM;
 717         nents = dma_map_sg(adev->dev, ttm->sg->sgl, ttm->sg->nents, direction);
 718         if (nents != ttm->sg->nents)
 719                 goto release_sg;
 720
 721         drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
 722                                          gtt->ttm.dma_address, ttm->num_pages);
 723
 724         amdgpu_trace_dma_map(ttm);
 725
 726         return 0;
 727
 728 release_sg:
 729         kfree(ttm->sg);
 730         ttm->sg = NULL;
 731         return r;
 732 }
 733
 734 static void amdgpu_ttm_tt_unpin_userptr(struct ttm_tt *ttm)
 735 {
 736         struct amdgpu_device *adev = amdgpu_ttm_adev(ttm->bdev);
 737         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 738         struct sg_page_iter sg_iter;
 739
 740         int write = !(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
 741         enum dma_data_direction direction = write ?
 742                 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
 743
 744         /* double check that we don't free the table twice */
 745         if (!ttm->sg || !ttm->sg->sgl)
 746                 return;
 747
 748         /* free the sg table and pages again */
 749         dma_unmap_sg(adev->dev, ttm->sg->sgl, ttm->sg->nents, direction);
 750
 751         for_each_sg_page(ttm->sg->sgl, &sg_iter, ttm->sg->nents, 0) {
 752                 struct page *page = sg_page_iter_page(&sg_iter);
 753                 if (!(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY))
 754                         set_page_dirty(page);
 755
 756                 mark_page_accessed(page);
 757                 put_page(page);
 758         }
 759
 760         amdgpu_trace_dma_unmap(ttm);
 761
 762         sg_free_table(ttm->sg);
 763 }
 764
 765 static int amdgpu_ttm_backend_bind(struct ttm_tt *ttm,
 766                                    struct ttm_mem_reg *bo_mem)
 767 {
 768         struct amdgpu_ttm_tt *gtt = (void*)ttm;
 769         uint64_t flags;
 770         int r = 0;
 771
 772         if (gtt->userptr) {
 773                 r = amdgpu_ttm_tt_pin_userptr(ttm);
 774                 if (r) {
 775                         DRM_ERROR("failed to pin userptr\n");
 776                         return r;
 777                 }
 778         }
 779         if (!ttm->num_pages) {
 780                 WARN(1, "nothing to bind %lu pages for mreg %p back %p!\n",
 781                      ttm->num_pages, bo_mem, ttm);
 782         }
 783
 784         if (bo_mem->mem_type == AMDGPU_PL_GDS ||
 785             bo_mem->mem_type == AMDGPU_PL_GWS ||
 786             bo_mem->mem_type == AMDGPU_PL_OA)
 787                 return -EINVAL;
 788
 789         if (!amdgpu_gtt_mgr_is_allocated(bo_mem))
 790                 return 0;
 791
 792         spin_lock(&gtt->adev->gtt_list_lock);
 793         flags = amdgpu_ttm_tt_pte_flags(gtt->adev, ttm, bo_mem);
 794         gtt->offset = (u64)bo_mem->start << PAGE_SHIFT;
 795         r = amdgpu_gart_bind(gtt->adev, gtt->offset, ttm->num_pages,
 796                 ttm->pages, gtt->ttm.dma_address, flags);
 797
 798         if (r) {
 799                 DRM_ERROR("failed to bind %lu pages at 0x%08llX\n",
 800                           ttm->num_pages, gtt->offset);
 801                 goto error_gart_bind;
 802         }
 803
 804         list_add_tail(&gtt->list, &gtt->adev->gtt_list);
 805 error_gart_bind:
 806         spin_unlock(&gtt->adev->gtt_list_lock);
 807         return r;
 808 }
 809
 810 bool amdgpu_ttm_is_bound(struct ttm_tt *ttm)
 811 {
 812         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 813
 814         return gtt && !list_empty(&gtt->list);
 815 }
 816
 817 int amdgpu_ttm_bind(struct ttm_buffer_object *bo, struct ttm_mem_reg *bo_mem)
 818 {
 819         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
 820         struct ttm_tt *ttm = bo->ttm;
 821         struct ttm_mem_reg tmp;
 822
 823         struct ttm_placement placement;
 824         struct ttm_place placements;
 825         int r;
 826
 827         if (!ttm || amdgpu_ttm_is_bound(ttm))
 828                 return 0;
 829
 830         tmp = bo->mem;
 831         tmp.mm_node = NULL;
 832         placement.num_placement = 1;
 833         placement.placement = &placements;
 834         placement.num_busy_placement = 1;
 835         placement.busy_placement = &placements;
 836         placements.fpfn = 0;
 837         placements.lpfn = adev->mc.gart_size >> PAGE_SHIFT;
 838         placements.flags = bo->mem.placement | TTM_PL_FLAG_TT;
 839
 840         r = ttm_bo_mem_space(bo, &placement, &tmp, true, false);
 841         if (unlikely(r))
 842                 return r;
 843
 844         r = ttm_bo_move_ttm(bo, true, false, &tmp);
 845         if (unlikely(r))
 846                 ttm_bo_mem_put(bo, &tmp);
 847         else
 848                 bo->offset = (bo->mem.start << PAGE_SHIFT) +
 849                         bo->bdev->man[bo->mem.mem_type].gpu_offset;
 850
 851         return r;
 852 }
 853
 854 int amdgpu_ttm_recover_gart(struct amdgpu_device *adev)
 855 {
 856         struct amdgpu_ttm_tt *gtt, *tmp;
 857         struct ttm_mem_reg bo_mem;
 858         uint64_t flags;
 859         int r;
 860
 861         bo_mem.mem_type = TTM_PL_TT;
 862         spin_lock(&adev->gtt_list_lock);
 863         list_for_each_entry_safe(gtt, tmp, &adev->gtt_list, list) {
 864                 flags = amdgpu_ttm_tt_pte_flags(gtt->adev, &gtt->ttm.ttm, &bo_mem);
 865                 r = amdgpu_gart_bind(adev, gtt->offset, gtt->ttm.ttm.num_pages,
 866                                      gtt->ttm.ttm.pages, gtt->ttm.dma_address,
 867                                      flags);
 868                 if (r) {
 869                         spin_unlock(&adev->gtt_list_lock);
 870                         DRM_ERROR("failed to bind %lu pages at 0x%08llX\n",
 871                                   gtt->ttm.ttm.num_pages, gtt->offset);
 872                         return r;
 873                 }
 874         }
 875         spin_unlock(&adev->gtt_list_lock);
 876         return 0;
 877 }
 878
 879 static int amdgpu_ttm_backend_unbind(struct ttm_tt *ttm)
 880 {
 881         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 882         int r;
 883
 884         if (gtt->userptr)
 885                 amdgpu_ttm_tt_unpin_userptr(ttm);
 886
 887         if (!amdgpu_ttm_is_bound(ttm))
 888                 return 0;
 889
 890         /* unbind shouldn't be done for GDS/GWS/OA in ttm_bo_clean_mm */
 891         spin_lock(&gtt->adev->gtt_list_lock);
 892         r = amdgpu_gart_unbind(gtt->adev, gtt->offset, ttm->num_pages);
 893         if (r) {
 894                 DRM_ERROR("failed to unbind %lu pages at 0x%08llX\n",
 895                           gtt->ttm.ttm.num_pages, gtt->offset);
 896                 goto error_unbind;
 897         }
 898         list_del_init(&gtt->list);
 899 error_unbind:
 900         spin_unlock(&gtt->adev->gtt_list_lock);
 901         return r;
 902 }
 903
 904 static void amdgpu_ttm_backend_destroy(struct ttm_tt *ttm)
 905 {
 906         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 907
 908         ttm_dma_tt_fini(&gtt->ttm);
 909         kfree(gtt);
 910 }
 911
 912 static struct ttm_backend_func amdgpu_backend_func = {
 913         .bind = &amdgpu_ttm_backend_bind,
 914         .unbind = &amdgpu_ttm_backend_unbind,
 915         .destroy = &amdgpu_ttm_backend_destroy,
 916 };
 917
 918 static struct ttm_tt *amdgpu_ttm_tt_create(struct ttm_bo_device *bdev,
 919                                     unsigned long size, uint32_t page_flags,
 920                                     struct page *dummy_read_page)
 921 {
 922         struct amdgpu_device *adev;
 923         struct amdgpu_ttm_tt *gtt;
 924
 925         adev = amdgpu_ttm_adev(bdev);
 926
 927         gtt = kzalloc(sizeof(struct amdgpu_ttm_tt), GFP_KERNEL);
 928         if (gtt == NULL) {
 929                 return NULL;
 930         }
 931         gtt->ttm.ttm.func = &amdgpu_backend_func;
 932         gtt->adev = adev;
 933         if (ttm_dma_tt_init(&gtt->ttm, bdev, size, page_flags, dummy_read_page)) {
 934                 kfree(gtt);
 935                 return NULL;
 936         }
 937         INIT_LIST_HEAD(&gtt->list);
 938         return &gtt->ttm.ttm;
 939 }
 940
 941 static int amdgpu_ttm_tt_populate(struct ttm_tt *ttm)
 942 {
 943         struct amdgpu_device *adev = amdgpu_ttm_adev(ttm->bdev);
 944         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 945         unsigned i;
 946         int r;
 947         bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
 948
 949         if (ttm->state != tt_unpopulated)
 950                 return 0;
 951
 952         if (gtt && gtt->userptr) {
 953                 ttm->sg = kzalloc(sizeof(struct sg_table), GFP_KERNEL);
 954                 if (!ttm->sg)
 955                         return -ENOMEM;
 956
 957                 ttm->page_flags |= TTM_PAGE_FLAG_SG;
 958                 ttm->state = tt_unbound;
 959                 return 0;
 960         }
 961
 962         if (slave && ttm->sg) {
 963                 drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
 964                                                  gtt->ttm.dma_address, ttm->num_pages);
 965                 ttm->state = tt_unbound;
 966                 r = 0;
 967                 goto trace_mappings;
 968         }
 969
 970 #ifdef CONFIG_SWIOTLB
 971         if (swiotlb_nr_tbl()) {
 972                 r = ttm_dma_populate(&gtt->ttm, adev->dev);
 973                 goto trace_mappings;
 974         }
 975 #endif
 976
 977         r = ttm_pool_populate(ttm);
 978         if (r) {
 979                 return r;
 980         }
 981
 982         for (i = 0; i < ttm->num_pages; i++) {
 983                 gtt->ttm.dma_address[i] = pci_map_page(adev->pdev, ttm->pages[i],
 984                                                        0, PAGE_SIZE,
 985                                                        PCI_DMA_BIDIRECTIONAL);
 986                 if (pci_dma_mapping_error(adev->pdev, gtt->ttm.dma_address[i])) {
 987                         while (i--) {
 988                                 pci_unmap_page(adev->pdev, gtt->ttm.dma_address[i],
 989                                                PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
 990                                 gtt->ttm.dma_address[i] = 0;
 991                         }
 992                         ttm_pool_unpopulate(ttm);
 993                         return -EFAULT;
 994                 }
 995         }
 996
 997         r = 0;
 998 trace_mappings:
 999         if (likely(!r))
1000                 amdgpu_trace_dma_map(ttm);
1001         return r;
1002 }
1003
1004 static void amdgpu_ttm_tt_unpopulate(struct ttm_tt *ttm)
1005 {
1006         struct amdgpu_device *adev;
1007         struct amdgpu_ttm_tt *gtt = (void *)ttm;
1008         unsigned i;
1009         bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
1010
1011         if (gtt && gtt->userptr) {
1012                 kfree(ttm->sg);
1013                 ttm->sg = NULL;
1014                 ttm->page_flags &= ~TTM_PAGE_FLAG_SG;
1015                 return;
1016         }
1017
1018         if (slave)
1019                 return;
1020
1021         adev = amdgpu_ttm_adev(ttm->bdev);
1022
1023         amdgpu_trace_dma_unmap(ttm);
1024
1025 #ifdef CONFIG_SWIOTLB
1026         if (swiotlb_nr_tbl()) {
1027                 ttm_dma_unpopulate(&gtt->ttm, adev->dev);
1028                 return;
1029         }
1030 #endif
1031
1032         for (i = 0; i < ttm->num_pages; i++) {
1033                 if (gtt->ttm.dma_address[i]) {
1034                         pci_unmap_page(adev->pdev, gtt->ttm.dma_address[i],
1035                                        PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
1036                 }
1037         }
1038
1039         ttm_pool_unpopulate(ttm);
1040 }
1041
1042 int amdgpu_ttm_tt_set_userptr(struct ttm_tt *ttm, uint64_t addr,
1043                               uint32_t flags)
1044 {
1045         struct amdgpu_ttm_tt *gtt = (void *)ttm;
1046
1047         if (gtt == NULL)
1048                 return -EINVAL;
1049
1050         gtt->userptr = addr;
1051         gtt->usermm = current->mm;
1052         gtt->userflags = flags;
1053         spin_lock_init(&gtt->guptasklock);
1054         INIT_LIST_HEAD(&gtt->guptasks);
1055         atomic_set(&gtt->mmu_invalidations, 0);
1056
1057         return 0;
1058 }
1059
1060 struct mm_struct *amdgpu_ttm_tt_get_usermm(struct ttm_tt *ttm)
1061 {
1062         struct amdgpu_ttm_tt *gtt = (void *)ttm;
1063
1064         if (gtt == NULL)
1065                 return NULL;
1066
1067         return gtt->usermm;
1068 }
1069
1070 bool amdgpu_ttm_tt_affect_userptr(struct ttm_tt *ttm, unsigned long start,
1071                                   unsigned long end)
1072 {
1073         struct amdgpu_ttm_tt *gtt = (void *)ttm;
1074         struct amdgpu_ttm_gup_task_list *entry;
1075         unsigned long size;
1076
1077         if (gtt == NULL || !gtt->userptr)
1078                 return false;
1079
1080         size = (unsigned long)gtt->ttm.ttm.num_pages * PAGE_SIZE;
1081         if (gtt->userptr > end || gtt->userptr + size <= start)
1082                 return false;
1083
1084         spin_lock(&gtt->guptasklock);
1085         list_for_each_entry(entry, &gtt->guptasks, list) {
1086                 if (entry->task == current) {
1087                         spin_unlock(&gtt->guptasklock);
1088                         return false;
1089                 }
1090         }
1091         spin_unlock(&gtt->guptasklock);
1092
1093         atomic_inc(&gtt->mmu_invalidations);
1094
1095         return true;
1096 }
1097
1098 bool amdgpu_ttm_tt_userptr_invalidated(struct ttm_tt *ttm,
1099                                        int *last_invalidated)
1100 {
1101         struct amdgpu_ttm_tt *gtt = (void *)ttm;
1102         int prev_invalidated = *last_invalidated;
1103
1104         *last_invalidated = atomic_read(&gtt->mmu_invalidations);
1105         return prev_invalidated != *last_invalidated;
1106 }
1107
1108 bool amdgpu_ttm_tt_is_readonly(struct ttm_tt *ttm)
1109 {
1110         struct amdgpu_ttm_tt *gtt = (void *)ttm;
1111
1112         if (gtt == NULL)
1113                 return false;
1114
1115         return !!(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
1116 }
1117
1118 uint64_t amdgpu_ttm_tt_pte_flags(struct amdgpu_device *adev, struct ttm_tt *ttm,
1119                                  struct ttm_mem_reg *mem)
1120 {
1121         uint64_t flags = 0;
1122
1123         if (mem && mem->mem_type != TTM_PL_SYSTEM)
1124                 flags |= AMDGPU_PTE_VALID;
1125
1126         if (mem && mem->mem_type == TTM_PL_TT) {
1127                 flags |= AMDGPU_PTE_SYSTEM;
1128
1129                 if (ttm->caching_state == tt_cached)
1130                         flags |= AMDGPU_PTE_SNOOPED;
1131         }
1132
1133         flags |= adev->gart.gart_pte_flags;
1134         flags |= AMDGPU_PTE_READABLE;
1135
1136         if (!amdgpu_ttm_tt_is_readonly(ttm))
1137                 flags |= AMDGPU_PTE_WRITEABLE;
1138
1139         return flags;
1140 }
1141
1142 static bool amdgpu_ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
1143                                             const struct ttm_place *place)
1144 {
1145         unsigned long num_pages = bo->mem.num_pages;
1146         struct drm_mm_node *node = bo->mem.mm_node;
1147
1148         if (bo->mem.start != AMDGPU_BO_INVALID_OFFSET)
1149                 return ttm_bo_eviction_valuable(bo, place);
1150
1151         switch (bo->mem.mem_type) {
1152         case TTM_PL_TT:
1153                 return true;
1154
1155         case TTM_PL_VRAM:
1156                 /* Check each drm MM node individually */
1157                 while (num_pages) {
1158                         if (place->fpfn < (node->start + node->size) &&
1159                             !(place->lpfn && place->lpfn <= node->start))
1160                                 return true;
1161
1162                         num_pages -= node->size;
1163                         ++node;
1164                 }
1165                 break;
1166
1167         default:
1168                 break;
1169         }
1170
1171         return ttm_bo_eviction_valuable(bo, place);
1172 }
1173
1174 static int amdgpu_ttm_access_memory(struct ttm_buffer_object *bo,
1175                                     unsigned long offset,
1176                                     void *buf, int len, int write)
1177 {
1178         struct amdgpu_bo *abo = container_of(bo, struct amdgpu_bo, tbo);
1179         struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev);
1180         struct drm_mm_node *nodes = abo->tbo.mem.mm_node;
1181         uint32_t value = 0;
1182         int ret = 0;
1183         uint64_t pos;
1184         unsigned long flags;
1185
1186         if (bo->mem.mem_type != TTM_PL_VRAM)
1187                 return -EIO;
1188
1189         while (offset >= (nodes->size << PAGE_SHIFT)) {
1190                 offset -= nodes->size << PAGE_SHIFT;
1191                 ++nodes;
1192         }
1193         pos = (nodes->start << PAGE_SHIFT) + offset;
1194
1195         while (len && pos < adev->mc.mc_vram_size) {
1196                 uint64_t aligned_pos = pos & ~(uint64_t)3;
1197                 uint32_t bytes = 4 - (pos & 3);
1198                 uint32_t shift = (pos & 3) * 8;
1199                 uint32_t mask = 0xffffffff << shift;
1200
1201                 if (len < bytes) {
1202                         mask &= 0xffffffff >> (bytes - len) * 8;
1203                         bytes = len;
1204                 }
1205
1206                 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
1207                 WREG32(mmMM_INDEX, ((uint32_t)aligned_pos) | 0x80000000);
1208                 WREG32(mmMM_INDEX_HI, aligned_pos >> 31);
1209                 if (!write || mask != 0xffffffff)
1210                         value = RREG32(mmMM_DATA);
1211                 if (write) {
1212                         value &= ~mask;
1213                         value |= (*(uint32_t *)buf << shift) & mask;
1214                         WREG32(mmMM_DATA, value);
1215                 }
1216                 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
1217                 if (!write) {
1218                         value = (value & mask) >> shift;
1219                         memcpy(buf, &value, bytes);
1220                 }
1221
1222                 ret += bytes;
1223                 buf = (uint8_t *)buf + bytes;
1224                 pos += bytes;
1225                 len -= bytes;
1226                 if (pos >= (nodes->start + nodes->size) << PAGE_SHIFT) {
1227                         ++nodes;
1228                         pos = (nodes->start << PAGE_SHIFT);
1229                 }
1230         }
1231
1232         return ret;
1233 }
1234
1235 static struct ttm_bo_driver amdgpu_bo_driver = {
1236         .ttm_tt_create = &amdgpu_ttm_tt_create,
1237         .ttm_tt_populate = &amdgpu_ttm_tt_populate,
1238         .ttm_tt_unpopulate = &amdgpu_ttm_tt_unpopulate,
1239         .invalidate_caches = &amdgpu_invalidate_caches,
1240         .init_mem_type = &amdgpu_init_mem_type,
1241         .eviction_valuable = amdgpu_ttm_bo_eviction_valuable,
1242         .evict_flags = &amdgpu_evict_flags,
1243         .move = &amdgpu_bo_move,
1244         .verify_access = &amdgpu_verify_access,
1245         .move_notify = &amdgpu_bo_move_notify,
1246         .fault_reserve_notify = &amdgpu_bo_fault_reserve_notify,
1247         .io_mem_reserve = &amdgpu_ttm_io_mem_reserve,
1248         .io_mem_free = &amdgpu_ttm_io_mem_free,
1249         .io_mem_pfn = amdgpu_ttm_io_mem_pfn,
1250         .access_memory = &amdgpu_ttm_access_memory
1251 };
1252
1253 int amdgpu_ttm_init(struct amdgpu_device *adev)
1254 {
1255         uint64_t gtt_size;
1256         int r;
1257         u64 vis_vram_limit;
1258
1259         r = amdgpu_ttm_global_init(adev);
1260         if (r) {
1261                 return r;
1262         }
1263         /* No others user of address space so set it to 0 */
1264         r = ttm_bo_device_init(&adev->mman.bdev,
1265                                adev->mman.bo_global_ref.ref.object,
1266                                &amdgpu_bo_driver,
1267                                adev->ddev->anon_inode->i_mapping,
1268                                DRM_FILE_PAGE_OFFSET,
1269                                adev->need_dma32);
1270         if (r) {
1271                 DRM_ERROR("failed initializing buffer object driver(%d).\n", r);
1272                 return r;
1273         }
1274         adev->mman.initialized = true;
1275         r = ttm_bo_init_mm(&adev->mman.bdev, TTM_PL_VRAM,
1276                                 adev->mc.real_vram_size >> PAGE_SHIFT);
1277         if (r) {
1278                 DRM_ERROR("Failed initializing VRAM heap.\n");
1279                 return r;
1280         }
1281
1282         /* Reduce size of CPU-visible VRAM if requested */
1283         vis_vram_limit = (u64)amdgpu_vis_vram_limit * 1024 * 1024;
1284         if (amdgpu_vis_vram_limit > 0 &&
1285             vis_vram_limit <= adev->mc.visible_vram_size)
1286                 adev->mc.visible_vram_size = vis_vram_limit;
1287
1288         /* Change the size here instead of the init above so only lpfn is affected */
1289         amdgpu_ttm_set_active_vram_size(adev, adev->mc.visible_vram_size);
1290
1291         r = amdgpu_bo_create_kernel(adev, adev->mc.stolen_size, PAGE_SIZE,
1292                                     AMDGPU_GEM_DOMAIN_VRAM,
1293                                     &adev->stolen_vga_memory,
1294                                     NULL, NULL);
1295         if (r)
1296                 return r;
1297         DRM_INFO("amdgpu: %uM of VRAM memory ready\n",
1298                  (unsigned) (adev->mc.real_vram_size / (1024 * 1024)));
1299
1300         if (amdgpu_gtt_size == -1)
1301                 gtt_size = max((AMDGPU_DEFAULT_GTT_SIZE_MB << 20),
1302                                adev->mc.mc_vram_size);
1303         else
1304                 gtt_size = (uint64_t)amdgpu_gtt_size << 20;
1305         r = ttm_bo_init_mm(&adev->mman.bdev, TTM_PL_TT, gtt_size >> PAGE_SHIFT);
1306         if (r) {
1307                 DRM_ERROR("Failed initializing GTT heap.\n");
1308                 return r;
1309         }
1310         DRM_INFO("amdgpu: %uM of GTT memory ready.\n",
1311                  (unsigned)(gtt_size / (1024 * 1024)));
1312
1313         adev->gds.mem.total_size = adev->gds.mem.total_size << AMDGPU_GDS_SHIFT;
1314         adev->gds.mem.gfx_partition_size = adev->gds.mem.gfx_partition_size << AMDGPU_GDS_SHIFT;
1315         adev->gds.mem.cs_partition_size = adev->gds.mem.cs_partition_size << AMDGPU_GDS_SHIFT;
1316         adev->gds.gws.total_size = adev->gds.gws.total_size << AMDGPU_GWS_SHIFT;
1317         adev->gds.gws.gfx_partition_size = adev->gds.gws.gfx_partition_size << AMDGPU_GWS_SHIFT;
1318         adev->gds.gws.cs_partition_size = adev->gds.gws.cs_partition_size << AMDGPU_GWS_SHIFT;
1319         adev->gds.oa.total_size = adev->gds.oa.total_size << AMDGPU_OA_SHIFT;
1320         adev->gds.oa.gfx_partition_size = adev->gds.oa.gfx_partition_size << AMDGPU_OA_SHIFT;
1321         adev->gds.oa.cs_partition_size = adev->gds.oa.cs_partition_size << AMDGPU_OA_SHIFT;
1322         /* GDS Memory */
1323         if (adev->gds.mem.total_size) {
1324                 r = ttm_bo_init_mm(&adev->mman.bdev, AMDGPU_PL_GDS,
1325                                    adev->gds.mem.total_size >> PAGE_SHIFT);
1326                 if (r) {
1327                         DRM_ERROR("Failed initializing GDS heap.\n");
1328                         return r;
1329                 }
1330         }
1331
1332         /* GWS */
1333         if (adev->gds.gws.total_size) {
1334                 r = ttm_bo_init_mm(&adev->mman.bdev, AMDGPU_PL_GWS,
1335                                    adev->gds.gws.total_size >> PAGE_SHIFT);
1336                 if (r) {
1337                         DRM_ERROR("Failed initializing gws heap.\n");
1338                         return r;
1339                 }
1340         }
1341
1342         /* OA */
1343         if (adev->gds.oa.total_size) {
1344                 r = ttm_bo_init_mm(&adev->mman.bdev, AMDGPU_PL_OA,
1345                                    adev->gds.oa.total_size >> PAGE_SHIFT);
1346                 if (r) {
1347                         DRM_ERROR("Failed initializing oa heap.\n");
1348                         return r;
1349                 }
1350         }
1351
1352         r = amdgpu_ttm_debugfs_init(adev);
1353         if (r) {
1354                 DRM_ERROR("Failed to init debugfs\n");
1355                 return r;
1356         }
1357         return 0;
1358 }
1359
1360 void amdgpu_ttm_fini(struct amdgpu_device *adev)
1361 {
1362         int r;
1363
1364         if (!adev->mman.initialized)
1365                 return;
1366         amdgpu_ttm_debugfs_fini(adev);
1367         if (adev->stolen_vga_memory) {
1368                 r = amdgpu_bo_reserve(adev->stolen_vga_memory, true);
1369                 if (r == 0) {
1370                         amdgpu_bo_unpin(adev->stolen_vga_memory);
1371                         amdgpu_bo_unreserve(adev->stolen_vga_memory);
1372                 }
1373                 amdgpu_bo_unref(&adev->stolen_vga_memory);
1374         }
1375         ttm_bo_clean_mm(&adev->mman.bdev, TTM_PL_VRAM);
1376         ttm_bo_clean_mm(&adev->mman.bdev, TTM_PL_TT);
1377         if (adev->gds.mem.total_size)
1378                 ttm_bo_clean_mm(&adev->mman.bdev, AMDGPU_PL_GDS);
1379         if (adev->gds.gws.total_size)
1380                 ttm_bo_clean_mm(&adev->mman.bdev, AMDGPU_PL_GWS);
1381         if (adev->gds.oa.total_size)
1382                 ttm_bo_clean_mm(&adev->mman.bdev, AMDGPU_PL_OA);
1383         ttm_bo_device_release(&adev->mman.bdev);
1384         amdgpu_gart_fini(adev);
1385         amdgpu_ttm_global_fini(adev);
1386         adev->mman.initialized = false;
1387         DRM_INFO("amdgpu: ttm finalized\n");
1388 }
1389
1390 /* this should only be called at bootup or when userspace
1391  * isn't running */
1392 void amdgpu_ttm_set_active_vram_size(struct amdgpu_device *adev, u64 size)
1393 {
1394         struct ttm_mem_type_manager *man;
1395
1396         if (!adev->mman.initialized)
1397                 return;
1398
1399         man = &adev->mman.bdev.man[TTM_PL_VRAM];
1400         /* this just adjusts TTM size idea, which sets lpfn to the correct value */
1401         man->size = size >> PAGE_SHIFT;
1402 }
1403
1404 int amdgpu_mmap(struct file *filp, struct vm_area_struct *vma)
1405 {
1406         struct drm_file *file_priv;
1407         struct amdgpu_device *adev;
1408
1409         if (unlikely(vma->vm_pgoff < DRM_FILE_PAGE_OFFSET))
1410                 return -EINVAL;
1411
1412         file_priv = filp->private_data;
1413         adev = file_priv->minor->dev->dev_private;
1414         if (adev == NULL)
1415                 return -EINVAL;
1416
1417         return ttm_bo_mmap(filp, vma, &adev->mman.bdev);
1418 }
1419
1420 static int amdgpu_map_buffer(struct ttm_buffer_object *bo,
1421                              struct ttm_mem_reg *mem, unsigned num_pages,
1422                              uint64_t offset, unsigned window,
1423                              struct amdgpu_ring *ring,
1424                              uint64_t *addr)
1425 {
1426         struct amdgpu_ttm_tt *gtt = (void *)bo->ttm;
1427         struct amdgpu_device *adev = ring->adev;
1428         struct ttm_tt *ttm = bo->ttm;
1429         struct amdgpu_job *job;
1430         unsigned num_dw, num_bytes;
1431         dma_addr_t *dma_address;
1432         struct dma_fence *fence;
1433         uint64_t src_addr, dst_addr;
1434         uint64_t flags;
1435         int r;
1436
1437         BUG_ON(adev->mman.buffer_funcs->copy_max_bytes <
1438                AMDGPU_GTT_MAX_TRANSFER_SIZE * 8);
1439
1440         *addr = adev->mc.gart_start;
1441         *addr += (u64)window * AMDGPU_GTT_MAX_TRANSFER_SIZE *
1442                 AMDGPU_GPU_PAGE_SIZE;
1443
1444         num_dw = adev->mman.buffer_funcs->copy_num_dw;
1445         while (num_dw & 0x7)
1446                 num_dw++;
1447
1448         num_bytes = num_pages * 8;
1449
1450         r = amdgpu_job_alloc_with_ib(adev, num_dw * 4 + num_bytes, &job);
1451         if (r)
1452                 return r;
1453
1454         src_addr = num_dw * 4;
1455         src_addr += job->ibs[0].gpu_addr;
1456
1457         dst_addr = adev->gart.table_addr;
1458         dst_addr += window * AMDGPU_GTT_MAX_TRANSFER_SIZE * 8;
1459         amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_addr,
1460                                 dst_addr, num_bytes);
1461
1462         amdgpu_ring_pad_ib(ring, &job->ibs[0]);
1463         WARN_ON(job->ibs[0].length_dw > num_dw);
1464
1465         dma_address = &gtt->ttm.dma_address[offset >> PAGE_SHIFT];
1466         flags = amdgpu_ttm_tt_pte_flags(adev, ttm, mem);
1467         r = amdgpu_gart_map(adev, 0, num_pages, dma_address, flags,
1468                             &job->ibs[0].ptr[num_dw]);
1469         if (r)
1470                 goto error_free;
1471
1472         r = amdgpu_job_submit(job, ring, &adev->mman.entity,
1473                               AMDGPU_FENCE_OWNER_UNDEFINED, &fence);
1474         if (r)
1475                 goto error_free;
1476
1477         dma_fence_put(fence);
1478
1479         return r;
1480
1481 error_free:
1482         amdgpu_job_free(job);
1483         return r;
1484 }
1485
1486 int amdgpu_copy_buffer(struct amdgpu_ring *ring, uint64_t src_offset,
1487                        uint64_t dst_offset, uint32_t byte_count,
1488                        struct reservation_object *resv,
1489                        struct dma_fence **fence, bool direct_submit,
1490                        bool vm_needs_flush)
1491 {
1492         struct amdgpu_device *adev = ring->adev;
1493         struct amdgpu_job *job;
1494
1495         uint32_t max_bytes;
1496         unsigned num_loops, num_dw;
1497         unsigned i;
1498         int r;
1499
1500         max_bytes = adev->mman.buffer_funcs->copy_max_bytes;
1501         num_loops = DIV_ROUND_UP(byte_count, max_bytes);
1502         num_dw = num_loops * adev->mman.buffer_funcs->copy_num_dw;
1503
1504         /* for IB padding */
1505         while (num_dw & 0x7)
1506                 num_dw++;
1507
1508         r = amdgpu_job_alloc_with_ib(adev, num_dw * 4, &job);
1509         if (r)
1510                 return r;
1511
1512         job->vm_needs_flush = vm_needs_flush;
1513         if (resv) {
1514                 r = amdgpu_sync_resv(adev, &job->sync, resv,
1515                                      AMDGPU_FENCE_OWNER_UNDEFINED);
1516                 if (r) {
1517                         DRM_ERROR("sync failed (%d).\n", r);
1518                         goto error_free;
1519                 }
1520         }
1521
1522         for (i = 0; i < num_loops; i++) {
1523                 uint32_t cur_size_in_bytes = min(byte_count, max_bytes);
1524
1525                 amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_offset,
1526                                         dst_offset, cur_size_in_bytes);
1527
1528                 src_offset += cur_size_in_bytes;
1529                 dst_offset += cur_size_in_bytes;
1530                 byte_count -= cur_size_in_bytes;
1531         }
1532
1533         amdgpu_ring_pad_ib(ring, &job->ibs[0]);
1534         WARN_ON(job->ibs[0].length_dw > num_dw);
1535         if (direct_submit) {
1536                 r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs,
1537                                        NULL, fence);
1538                 job->fence = dma_fence_get(*fence);
1539                 if (r)
1540                         DRM_ERROR("Error scheduling IBs (%d)\n", r);
1541                 amdgpu_job_free(job);
1542         } else {
1543                 r = amdgpu_job_submit(job, ring, &adev->mman.entity,
1544                                       AMDGPU_FENCE_OWNER_UNDEFINED, fence);
1545                 if (r)
1546                         goto error_free;
1547         }
1548
1549         return r;
1550
1551 error_free:
1552         amdgpu_job_free(job);
1553         return r;
1554 }
1555
1556 int amdgpu_fill_buffer(struct amdgpu_bo *bo,
1557                        uint64_t src_data,
1558                        struct reservation_object *resv,
1559                        struct dma_fence **fence)
1560 {
1561         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1562         /* max_bytes applies to SDMA_OP_PTEPDE as well as SDMA_OP_CONST_FILL*/
1563         uint32_t max_bytes = adev->mman.buffer_funcs->fill_max_bytes;
1564         struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
1565
1566         struct drm_mm_node *mm_node;
1567         unsigned long num_pages;
1568         unsigned int num_loops, num_dw;
1569
1570         struct amdgpu_job *job;
1571         int r;
1572
1573         if (!ring->ready) {
1574                 DRM_ERROR("Trying to clear memory with ring turned off.\n");
1575                 return -EINVAL;
1576         }
1577
1578         if (bo->tbo.mem.mem_type == TTM_PL_TT) {
1579                 r = amdgpu_ttm_bind(&bo->tbo, &bo->tbo.mem);
1580                 if (r)
1581                         return r;
1582         }
1583
1584         num_pages = bo->tbo.num_pages;
1585         mm_node = bo->tbo.mem.mm_node;
1586         num_loops = 0;
1587         while (num_pages) {
1588                 uint32_t byte_count = mm_node->size << PAGE_SHIFT;
1589
1590                 num_loops += DIV_ROUND_UP(byte_count, max_bytes);
1591                 num_pages -= mm_node->size;
1592                 ++mm_node;
1593         }
1594
1595         /* 10 double words for each SDMA_OP_PTEPDE cmd */
1596         num_dw = num_loops * 10;
1597
1598         /* for IB padding */
1599         num_dw += 64;
1600
1601         r = amdgpu_job_alloc_with_ib(adev, num_dw * 4, &job);
1602         if (r)
1603                 return r;
1604
1605         if (resv) {
1606                 r = amdgpu_sync_resv(adev, &job->sync, resv,
1607                                      AMDGPU_FENCE_OWNER_UNDEFINED);
1608                 if (r) {
1609                         DRM_ERROR("sync failed (%d).\n", r);
1610                         goto error_free;
1611                 }
1612         }
1613
1614         num_pages = bo->tbo.num_pages;
1615         mm_node = bo->tbo.mem.mm_node;
1616
1617         while (num_pages) {
1618                 uint32_t byte_count = mm_node->size << PAGE_SHIFT;
1619                 uint64_t dst_addr;
1620
1621                 WARN_ONCE(byte_count & 0x7, "size should be a multiple of 8");
1622
1623                 dst_addr = amdgpu_mm_node_addr(&bo->tbo, mm_node, &bo->tbo.mem);
1624                 while (byte_count) {
1625                         uint32_t cur_size_in_bytes = min(byte_count, max_bytes);
1626
1627                         amdgpu_vm_set_pte_pde(adev, &job->ibs[0],
1628                                         dst_addr, 0,
1629                                         cur_size_in_bytes >> 3, 0,
1630                                         src_data);
1631
1632                         dst_addr += cur_size_in_bytes;
1633                         byte_count -= cur_size_in_bytes;
1634                 }
1635
1636                 num_pages -= mm_node->size;
1637                 ++mm_node;
1638         }
1639
1640         amdgpu_ring_pad_ib(ring, &job->ibs[0]);
1641         WARN_ON(job->ibs[0].length_dw > num_dw);
1642         r = amdgpu_job_submit(job, ring, &adev->mman.entity,
1643                               AMDGPU_FENCE_OWNER_UNDEFINED, fence);
1644         if (r)
1645                 goto error_free;
1646
1647         return 0;
1648
1649 error_free:
1650         amdgpu_job_free(job);
1651         return r;
1652 }
1653
1654 #if defined(CONFIG_DEBUG_FS)
1655
1656 static int amdgpu_mm_dump_table(struct seq_file *m, void *data)
1657 {
1658         struct drm_info_node *node = (struct drm_info_node *)m->private;
1659         unsigned ttm_pl = *(int *)node->info_ent->data;
1660         struct drm_device *dev = node->minor->dev;
1661         struct amdgpu_device *adev = dev->dev_private;
1662         struct ttm_mem_type_manager *man = &adev->mman.bdev.man[ttm_pl];
1663         struct drm_printer p = drm_seq_file_printer(m);
1664
1665         man->func->debug(man, &p);
1666         return 0;
1667 }
1668
1669 static int ttm_pl_vram = TTM_PL_VRAM;
1670 static int ttm_pl_tt = TTM_PL_TT;
1671
1672 static const struct drm_info_list amdgpu_ttm_debugfs_list[] = {
1673         {"amdgpu_vram_mm", amdgpu_mm_dump_table, 0, &ttm_pl_vram},
1674         {"amdgpu_gtt_mm", amdgpu_mm_dump_table, 0, &ttm_pl_tt},
1675         {"ttm_page_pool", ttm_page_alloc_debugfs, 0, NULL},
1676 #ifdef CONFIG_SWIOTLB
1677         {"ttm_dma_page_pool", ttm_dma_page_alloc_debugfs, 0, NULL}
1678 #endif
1679 };
1680
1681 static ssize_t amdgpu_ttm_vram_read(struct file *f, char __user *buf,
1682                                     size_t size, loff_t *pos)
1683 {
1684         struct amdgpu_device *adev = file_inode(f)->i_private;
1685         ssize_t result = 0;
1686         int r;
1687
1688         if (size & 0x3 || *pos & 0x3)
1689                 return -EINVAL;
1690
1691         if (*pos >= adev->mc.mc_vram_size)
1692                 return -ENXIO;
1693
1694         while (size) {
1695                 unsigned long flags;
1696                 uint32_t value;
1697
1698                 if (*pos >= adev->mc.mc_vram_size)
1699                         return result;
1700
1701                 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
1702                 WREG32(mmMM_INDEX, ((uint32_t)*pos) | 0x80000000);
1703                 WREG32(mmMM_INDEX_HI, *pos >> 31);
1704                 value = RREG32(mmMM_DATA);
1705                 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
1706
1707                 r = put_user(value, (uint32_t *)buf);
1708                 if (r)
1709                         return r;
1710
1711                 result += 4;
1712                 buf += 4;
1713                 *pos += 4;
1714                 size -= 4;
1715         }
1716
1717         return result;
1718 }
1719
1720 static const struct file_operations amdgpu_ttm_vram_fops = {
1721         .owner = THIS_MODULE,
1722         .read = amdgpu_ttm_vram_read,
1723         .llseek = default_llseek
1724 };
1725
1726 #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS
1727
1728 static ssize_t amdgpu_ttm_gtt_read(struct file *f, char __user *buf,
1729                                    size_t size, loff_t *pos)
1730 {
1731         struct amdgpu_device *adev = file_inode(f)->i_private;
1732         ssize_t result = 0;
1733         int r;
1734
1735         while (size) {
1736                 loff_t p = *pos / PAGE_SIZE;
1737                 unsigned off = *pos & ~PAGE_MASK;
1738                 size_t cur_size = min_t(size_t, size, PAGE_SIZE - off);
1739                 struct page *page;
1740                 void *ptr;
1741
1742                 if (p >= adev->gart.num_cpu_pages)
1743                         return result;
1744
1745                 page = adev->gart.pages[p];
1746                 if (page) {
1747                         ptr = kmap(page);
1748                         ptr += off;
1749
1750                         r = copy_to_user(buf, ptr, cur_size);
1751                         kunmap(adev->gart.pages[p]);
1752                 } else
1753                         r = clear_user(buf, cur_size);
1754
1755                 if (r)
1756                         return -EFAULT;
1757
1758                 result += cur_size;
1759                 buf += cur_size;
1760                 *pos += cur_size;
1761                 size -= cur_size;
1762         }
1763
1764         return result;
1765 }
1766
1767 static const struct file_operations amdgpu_ttm_gtt_fops = {
1768         .owner = THIS_MODULE,
1769         .read = amdgpu_ttm_gtt_read,
1770         .llseek = default_llseek
1771 };
1772
1773 #endif
1774
1775 #endif
1776
1777 static int amdgpu_ttm_debugfs_init(struct amdgpu_device *adev)
1778 {
1779 #if defined(CONFIG_DEBUG_FS)
1780         unsigned count;
1781
1782         struct drm_minor *minor = adev->ddev->primary;
1783         struct dentry *ent, *root = minor->debugfs_root;
1784
1785         ent = debugfs_create_file("amdgpu_vram", S_IFREG | S_IRUGO, root,
1786                                   adev, &amdgpu_ttm_vram_fops);
1787         if (IS_ERR(ent))
1788                 return PTR_ERR(ent);
1789         i_size_write(ent->d_inode, adev->mc.mc_vram_size);
1790         adev->mman.vram = ent;
1791
1792 #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS
1793         ent = debugfs_create_file("amdgpu_gtt", S_IFREG | S_IRUGO, root,
1794                                   adev, &amdgpu_ttm_gtt_fops);
1795         if (IS_ERR(ent))
1796                 return PTR_ERR(ent);
1797         i_size_write(ent->d_inode, adev->mc.gart_size);
1798         adev->mman.gtt = ent;
1799
1800 #endif
1801         count = ARRAY_SIZE(amdgpu_ttm_debugfs_list);
1802
1803 #ifdef CONFIG_SWIOTLB
1804         if (!swiotlb_nr_tbl())
1805                 --count;
1806 #endif
1807
1808         return amdgpu_debugfs_add_files(adev, amdgpu_ttm_debugfs_list, count);
1809 #else
1810
1811         return 0;
1812 #endif
1813 }
1814
1815 static void amdgpu_ttm_debugfs_fini(struct amdgpu_device *adev)
1816 {
1817 #if defined(CONFIG_DEBUG_FS)
1818
1819         debugfs_remove(adev->mman.vram);
1820         adev->mman.vram = NULL;
1821
1822 #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS
1823         debugfs_remove(adev->mman.gtt);
1824         adev->mman.gtt = NULL;
1825 #endif
1826
1827 #endif
1828 }