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 <[email protected]>
  29  *    Thomas Hellstrom <thomas-at-tungstengraphics-dot-com>
  30  *    Dave Airlie
  31  */
  32 #include <ttm/ttm_bo_api.h>
  33 #include <ttm/ttm_bo_driver.h>
  34 #include <ttm/ttm_placement.h>
  35 #include <ttm/ttm_module.h>
  36 #include <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 "bif/bif_4_1_d.h"
  47
  48 #define DRM_FILE_PAGE_OFFSET (0x100000000ULL >> PAGE_SHIFT)
  49
  50 static int amdgpu_ttm_debugfs_init(struct amdgpu_device *adev);
  51 static void amdgpu_ttm_debugfs_fini(struct amdgpu_device *adev);
  52
  53 static struct amdgpu_device *amdgpu_get_adev(struct ttm_bo_device *bdev)
  54 {
  55         struct amdgpu_mman *mman;
  56         struct amdgpu_device *adev;
  57
  58         mman = container_of(bdev, struct amdgpu_mman, bdev);
  59         adev = container_of(mman, struct amdgpu_device, mman);
  60         return adev;
  61 }
  62
  63
  64 /*
  65  * Global memory.
  66  */
  67 static int amdgpu_ttm_mem_global_init(struct drm_global_reference *ref)
  68 {
  69         return ttm_mem_global_init(ref->object);
  70 }
  71
  72 static void amdgpu_ttm_mem_global_release(struct drm_global_reference *ref)
  73 {
  74         ttm_mem_global_release(ref->object);
  75 }
  76
  77 static int amdgpu_ttm_global_init(struct amdgpu_device *adev)
  78 {
  79         struct drm_global_reference *global_ref;
  80         struct amdgpu_ring *ring;
  81         struct amd_sched_rq *rq;
  82         int r;
  83
  84         adev->mman.mem_global_referenced = false;
  85         global_ref = &adev->mman.mem_global_ref;
  86         global_ref->global_type = DRM_GLOBAL_TTM_MEM;
  87         global_ref->size = sizeof(struct ttm_mem_global);
  88         global_ref->init = &amdgpu_ttm_mem_global_init;
  89         global_ref->release = &amdgpu_ttm_mem_global_release;
  90         r = drm_global_item_ref(global_ref);
  91         if (r != 0) {
  92                 DRM_ERROR("Failed setting up TTM memory accounting "
  93                           "subsystem.\n");
  94                 return r;
  95         }
  96
  97         adev->mman.bo_global_ref.mem_glob =
  98                 adev->mman.mem_global_ref.object;
  99         global_ref = &adev->mman.bo_global_ref.ref;
 100         global_ref->global_type = DRM_GLOBAL_TTM_BO;
 101         global_ref->size = sizeof(struct ttm_bo_global);
 102         global_ref->init = &ttm_bo_global_init;
 103         global_ref->release = &ttm_bo_global_release;
 104         r = drm_global_item_ref(global_ref);
 105         if (r != 0) {
 106                 DRM_ERROR("Failed setting up TTM BO subsystem.\n");
 107                 drm_global_item_unref(&adev->mman.mem_global_ref);
 108                 return r;
 109         }
 110
 111         ring = adev->mman.buffer_funcs_ring;
 112         rq = &ring->sched.sched_rq[AMD_SCHED_PRIORITY_KERNEL];
 113         r = amd_sched_entity_init(&ring->sched, &adev->mman.entity,
 114                                   rq, amdgpu_sched_jobs);
 115         if (r != 0) {
 116                 DRM_ERROR("Failed setting up TTM BO move run queue.\n");
 117                 drm_global_item_unref(&adev->mman.mem_global_ref);
 118                 drm_global_item_unref(&adev->mman.bo_global_ref.ref);
 119                 return r;
 120         }
 121
 122         adev->mman.mem_global_referenced = true;
 123
 124         return 0;
 125 }
 126
 127 static void amdgpu_ttm_global_fini(struct amdgpu_device *adev)
 128 {
 129         if (adev->mman.mem_global_referenced) {
 130                 amd_sched_entity_fini(adev->mman.entity.sched,
 131                                       &adev->mman.entity);
 132                 drm_global_item_unref(&adev->mman.bo_global_ref.ref);
 133                 drm_global_item_unref(&adev->mman.mem_global_ref);
 134                 adev->mman.mem_global_referenced = false;
 135         }
 136 }
 137
 138 static int amdgpu_invalidate_caches(struct ttm_bo_device *bdev, uint32_t flags)
 139 {
 140         return 0;
 141 }
 142
 143 static int amdgpu_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
 144                                 struct ttm_mem_type_manager *man)
 145 {
 146         struct amdgpu_device *adev;
 147
 148         adev = amdgpu_get_adev(bdev);
 149
 150         switch (type) {
 151         case TTM_PL_SYSTEM:
 152                 /* System memory */
 153                 man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
 154                 man->available_caching = TTM_PL_MASK_CACHING;
 155                 man->default_caching = TTM_PL_FLAG_CACHED;
 156                 break;
 157         case TTM_PL_TT:
 158                 man->func = &ttm_bo_manager_func;
 159                 man->gpu_offset = adev->mc.gtt_start;
 160                 man->available_caching = TTM_PL_MASK_CACHING;
 161                 man->default_caching = TTM_PL_FLAG_CACHED;
 162                 man->flags = TTM_MEMTYPE_FLAG_MAPPABLE | TTM_MEMTYPE_FLAG_CMA;
 163                 break;
 164         case TTM_PL_VRAM:
 165                 /* "On-card" video ram */
 166                 man->func = &ttm_bo_manager_func;
 167                 man->gpu_offset = adev->mc.vram_start;
 168                 man->flags = TTM_MEMTYPE_FLAG_FIXED |
 169                              TTM_MEMTYPE_FLAG_MAPPABLE;
 170                 man->available_caching = TTM_PL_FLAG_UNCACHED | TTM_PL_FLAG_WC;
 171                 man->default_caching = TTM_PL_FLAG_WC;
 172                 break;
 173         case AMDGPU_PL_GDS:
 174         case AMDGPU_PL_GWS:
 175         case AMDGPU_PL_OA:
 176                 /* On-chip GDS memory*/
 177                 man->func = &ttm_bo_manager_func;
 178                 man->gpu_offset = 0;
 179                 man->flags = TTM_MEMTYPE_FLAG_FIXED | TTM_MEMTYPE_FLAG_CMA;
 180                 man->available_caching = TTM_PL_FLAG_UNCACHED;
 181                 man->default_caching = TTM_PL_FLAG_UNCACHED;
 182                 break;
 183         default:
 184                 DRM_ERROR("Unsupported memory type %u\n", (unsigned)type);
 185                 return -EINVAL;
 186         }
 187         return 0;
 188 }
 189
 190 static void amdgpu_evict_flags(struct ttm_buffer_object *bo,
 191                                 struct ttm_placement *placement)
 192 {
 193         struct amdgpu_bo *rbo;
 194         static struct ttm_place placements = {
 195                 .fpfn = 0,
 196                 .lpfn = 0,
 197                 .flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM
 198         };
 199
 200         if (!amdgpu_ttm_bo_is_amdgpu_bo(bo)) {
 201                 placement->placement = &placements;
 202                 placement->busy_placement = &placements;
 203                 placement->num_placement = 1;
 204                 placement->num_busy_placement = 1;
 205                 return;
 206         }
 207         rbo = container_of(bo, struct amdgpu_bo, tbo);
 208         switch (bo->mem.mem_type) {
 209         case TTM_PL_VRAM:
 210                 if (rbo->adev->mman.buffer_funcs_ring->ready == false)
 211                         amdgpu_ttm_placement_from_domain(rbo, AMDGPU_GEM_DOMAIN_CPU);
 212                 else
 213                         amdgpu_ttm_placement_from_domain(rbo, AMDGPU_GEM_DOMAIN_GTT);
 214                 break;
 215         case TTM_PL_TT:
 216         default:
 217                 amdgpu_ttm_placement_from_domain(rbo, AMDGPU_GEM_DOMAIN_CPU);
 218         }
 219         *placement = rbo->placement;
 220 }
 221
 222 static int amdgpu_verify_access(struct ttm_buffer_object *bo, struct file *filp)
 223 {
 224         struct amdgpu_bo *rbo = container_of(bo, struct amdgpu_bo, tbo);
 225
 226         return drm_vma_node_verify_access(&rbo->gem_base.vma_node, filp);
 227 }
 228
 229 static void amdgpu_move_null(struct ttm_buffer_object *bo,
 230                              struct ttm_mem_reg *new_mem)
 231 {
 232         struct ttm_mem_reg *old_mem = &bo->mem;
 233
 234         BUG_ON(old_mem->mm_node != NULL);
 235         *old_mem = *new_mem;
 236         new_mem->mm_node = NULL;
 237 }
 238
 239 static int amdgpu_move_blit(struct ttm_buffer_object *bo,
 240                         bool evict, bool no_wait_gpu,
 241                         struct ttm_mem_reg *new_mem,
 242                         struct ttm_mem_reg *old_mem)
 243 {
 244         struct amdgpu_device *adev;
 245         struct amdgpu_ring *ring;
 246         uint64_t old_start, new_start;
 247         struct fence *fence;
 248         int r;
 249
 250         adev = amdgpu_get_adev(bo->bdev);
 251         ring = adev->mman.buffer_funcs_ring;
 252         old_start = old_mem->start << PAGE_SHIFT;
 253         new_start = new_mem->start << PAGE_SHIFT;
 254
 255         switch (old_mem->mem_type) {
 256         case TTM_PL_VRAM:
 257                 old_start += adev->mc.vram_start;
 258                 break;
 259         case TTM_PL_TT:
 260                 old_start += adev->mc.gtt_start;
 261                 break;
 262         default:
 263                 DRM_ERROR("Unknown placement %d\n", old_mem->mem_type);
 264                 return -EINVAL;
 265         }
 266         switch (new_mem->mem_type) {
 267         case TTM_PL_VRAM:
 268                 new_start += adev->mc.vram_start;
 269                 break;
 270         case TTM_PL_TT:
 271                 new_start += adev->mc.gtt_start;
 272                 break;
 273         default:
 274                 DRM_ERROR("Unknown placement %d\n", old_mem->mem_type);
 275                 return -EINVAL;
 276         }
 277         if (!ring->ready) {
 278                 DRM_ERROR("Trying to move memory with ring turned off.\n");
 279                 return -EINVAL;
 280         }
 281
 282         BUILD_BUG_ON((PAGE_SIZE % AMDGPU_GPU_PAGE_SIZE) != 0);
 283
 284         r = amdgpu_copy_buffer(ring, old_start, new_start,
 285                                new_mem->num_pages * PAGE_SIZE, /* bytes */
 286                                bo->resv, &fence);
 287         /* FIXME: handle copy error */
 288         r = ttm_bo_move_accel_cleanup(bo, fence,
 289                                       evict, no_wait_gpu, new_mem);
 290         fence_put(fence);
 291         return r;
 292 }
 293
 294 static int amdgpu_move_vram_ram(struct ttm_buffer_object *bo,
 295                                 bool evict, bool interruptible,
 296                                 bool no_wait_gpu,
 297                                 struct ttm_mem_reg *new_mem)
 298 {
 299         struct amdgpu_device *adev;
 300         struct ttm_mem_reg *old_mem = &bo->mem;
 301         struct ttm_mem_reg tmp_mem;
 302         struct ttm_place placements;
 303         struct ttm_placement placement;
 304         int r;
 305
 306         adev = amdgpu_get_adev(bo->bdev);
 307         tmp_mem = *new_mem;
 308         tmp_mem.mm_node = NULL;
 309         placement.num_placement = 1;
 310         placement.placement = &placements;
 311         placement.num_busy_placement = 1;
 312         placement.busy_placement = &placements;
 313         placements.fpfn = 0;
 314         placements.lpfn = 0;
 315         placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
 316         r = ttm_bo_mem_space(bo, &placement, &tmp_mem,
 317                              interruptible, no_wait_gpu);
 318         if (unlikely(r)) {
 319                 return r;
 320         }
 321
 322         r = ttm_tt_set_placement_caching(bo->ttm, tmp_mem.placement);
 323         if (unlikely(r)) {
 324                 goto out_cleanup;
 325         }
 326
 327         r = ttm_tt_bind(bo->ttm, &tmp_mem);
 328         if (unlikely(r)) {
 329                 goto out_cleanup;
 330         }
 331         r = amdgpu_move_blit(bo, true, no_wait_gpu, &tmp_mem, old_mem);
 332         if (unlikely(r)) {
 333                 goto out_cleanup;
 334         }
 335         r = ttm_bo_move_ttm(bo, true, no_wait_gpu, new_mem);
 336 out_cleanup:
 337         ttm_bo_mem_put(bo, &tmp_mem);
 338         return r;
 339 }
 340
 341 static int amdgpu_move_ram_vram(struct ttm_buffer_object *bo,
 342                                 bool evict, bool interruptible,
 343                                 bool no_wait_gpu,
 344                                 struct ttm_mem_reg *new_mem)
 345 {
 346         struct amdgpu_device *adev;
 347         struct ttm_mem_reg *old_mem = &bo->mem;
 348         struct ttm_mem_reg tmp_mem;
 349         struct ttm_placement placement;
 350         struct ttm_place placements;
 351         int r;
 352
 353         adev = amdgpu_get_adev(bo->bdev);
 354         tmp_mem = *new_mem;
 355         tmp_mem.mm_node = NULL;
 356         placement.num_placement = 1;
 357         placement.placement = &placements;
 358         placement.num_busy_placement = 1;
 359         placement.busy_placement = &placements;
 360         placements.fpfn = 0;
 361         placements.lpfn = 0;
 362         placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
 363         r = ttm_bo_mem_space(bo, &placement, &tmp_mem,
 364                              interruptible, no_wait_gpu);
 365         if (unlikely(r)) {
 366                 return r;
 367         }
 368         r = ttm_bo_move_ttm(bo, true, no_wait_gpu, &tmp_mem);
 369         if (unlikely(r)) {
 370                 goto out_cleanup;
 371         }
 372         r = amdgpu_move_blit(bo, true, no_wait_gpu, new_mem, old_mem);
 373         if (unlikely(r)) {
 374                 goto out_cleanup;
 375         }
 376 out_cleanup:
 377         ttm_bo_mem_put(bo, &tmp_mem);
 378         return r;
 379 }
 380
 381 static int amdgpu_bo_move(struct ttm_buffer_object *bo,
 382                         bool evict, bool interruptible,
 383                         bool no_wait_gpu,
 384                         struct ttm_mem_reg *new_mem)
 385 {
 386         struct amdgpu_device *adev;
 387         struct amdgpu_bo *abo;
 388         struct ttm_mem_reg *old_mem = &bo->mem;
 389         int r;
 390
 391         /* Can't move a pinned BO */
 392         abo = container_of(bo, struct amdgpu_bo, tbo);
 393         if (WARN_ON_ONCE(abo->pin_count > 0))
 394                 return -EINVAL;
 395
 396         adev = amdgpu_get_adev(bo->bdev);
 397         if (old_mem->mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
 398                 amdgpu_move_null(bo, new_mem);
 399                 return 0;
 400         }
 401         if ((old_mem->mem_type == TTM_PL_TT &&
 402              new_mem->mem_type == TTM_PL_SYSTEM) ||
 403             (old_mem->mem_type == TTM_PL_SYSTEM &&
 404              new_mem->mem_type == TTM_PL_TT)) {
 405                 /* bind is enough */
 406                 amdgpu_move_null(bo, new_mem);
 407                 return 0;
 408         }
 409         if (adev->mman.buffer_funcs == NULL ||
 410             adev->mman.buffer_funcs_ring == NULL ||
 411             !adev->mman.buffer_funcs_ring->ready) {
 412                 /* use memcpy */
 413                 goto memcpy;
 414         }
 415
 416         if (old_mem->mem_type == TTM_PL_VRAM &&
 417             new_mem->mem_type == TTM_PL_SYSTEM) {
 418                 r = amdgpu_move_vram_ram(bo, evict, interruptible,
 419                                         no_wait_gpu, new_mem);
 420         } else if (old_mem->mem_type == TTM_PL_SYSTEM &&
 421                    new_mem->mem_type == TTM_PL_VRAM) {
 422                 r = amdgpu_move_ram_vram(bo, evict, interruptible,
 423                                             no_wait_gpu, new_mem);
 424         } else {
 425                 r = amdgpu_move_blit(bo, evict, no_wait_gpu, new_mem, old_mem);
 426         }
 427
 428         if (r) {
 429 memcpy:
 430                 r = ttm_bo_move_memcpy(bo, evict, no_wait_gpu, new_mem);
 431                 if (r) {
 432                         return r;
 433                 }
 434         }
 435
 436         /* update statistics */
 437         atomic64_add((u64)bo->num_pages << PAGE_SHIFT, &adev->num_bytes_moved);
 438         return 0;
 439 }
 440
 441 static int amdgpu_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
 442 {
 443         struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
 444         struct amdgpu_device *adev = amdgpu_get_adev(bdev);
 445
 446         mem->bus.addr = NULL;
 447         mem->bus.offset = 0;
 448         mem->bus.size = mem->num_pages << PAGE_SHIFT;
 449         mem->bus.base = 0;
 450         mem->bus.is_iomem = false;
 451         if (!(man->flags & TTM_MEMTYPE_FLAG_MAPPABLE))
 452                 return -EINVAL;
 453         switch (mem->mem_type) {
 454         case TTM_PL_SYSTEM:
 455                 /* system memory */
 456                 return 0;
 457         case TTM_PL_TT:
 458                 break;
 459         case TTM_PL_VRAM:
 460                 mem->bus.offset = mem->start << PAGE_SHIFT;
 461                 /* check if it's visible */
 462                 if ((mem->bus.offset + mem->bus.size) > adev->mc.visible_vram_size)
 463                         return -EINVAL;
 464                 mem->bus.base = adev->mc.aper_base;
 465                 mem->bus.is_iomem = true;
 466 #ifdef __alpha__
 467                 /*
 468                  * Alpha: use bus.addr to hold the ioremap() return,
 469                  * so we can modify bus.base below.
 470                  */
 471                 if (mem->placement & TTM_PL_FLAG_WC)
 472                         mem->bus.addr =
 473                                 ioremap_wc(mem->bus.base + mem->bus.offset,
 474                                            mem->bus.size);
 475                 else
 476                         mem->bus.addr =
 477                                 ioremap_nocache(mem->bus.base + mem->bus.offset,
 478                                                 mem->bus.size);
 479
 480                 /*
 481                  * Alpha: Use just the bus offset plus
 482                  * the hose/domain memory base for bus.base.
 483                  * It then can be used to build PTEs for VRAM
 484                  * access, as done in ttm_bo_vm_fault().
 485                  */
 486                 mem->bus.base = (mem->bus.base & 0x0ffffffffUL) +
 487                         adev->ddev->hose->dense_mem_base;
 488 #endif
 489                 break;
 490         default:
 491                 return -EINVAL;
 492         }
 493         return 0;
 494 }
 495
 496 static void amdgpu_ttm_io_mem_free(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
 497 {
 498 }
 499
 500 /*
 501  * TTM backend functions.
 502  */
 503 struct amdgpu_ttm_gup_task_list {
 504         struct list_head        list;
 505         struct task_struct      *task;
 506 };
 507
 508 struct amdgpu_ttm_tt {
 509         struct ttm_dma_tt       ttm;
 510         struct amdgpu_device    *adev;
 511         u64                     offset;
 512         uint64_t                userptr;
 513         struct mm_struct        *usermm;
 514         uint32_t                userflags;
 515         spinlock_t              guptasklock;
 516         struct list_head        guptasks;
 517         atomic_t                mmu_invalidations;
 518 };
 519
 520 int amdgpu_ttm_tt_get_user_pages(struct ttm_tt *ttm, struct page **pages)
 521 {
 522         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 523         int write = !(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
 524         unsigned pinned = 0;
 525         int r;
 526
 527         if (gtt->userflags & AMDGPU_GEM_USERPTR_ANONONLY) {
 528                 /* check that we only use anonymous memory
 529                    to prevent problems with writeback */
 530                 unsigned long end = gtt->userptr + ttm->num_pages * PAGE_SIZE;
 531                 struct vm_area_struct *vma;
 532
 533                 vma = find_vma(gtt->usermm, gtt->userptr);
 534                 if (!vma || vma->vm_file || vma->vm_end < end)
 535                         return -EPERM;
 536         }
 537
 538         do {
 539                 unsigned num_pages = ttm->num_pages - pinned;
 540                 uint64_t userptr = gtt->userptr + pinned * PAGE_SIZE;
 541                 struct page **p = pages + pinned;
 542                 struct amdgpu_ttm_gup_task_list guptask;
 543
 544                 guptask.task = current;
 545                 spin_lock(&gtt->guptasklock);
 546                 list_add(&guptask.list, &gtt->guptasks);
 547                 spin_unlock(&gtt->guptasklock);
 548
 549                 r = get_user_pages(userptr, num_pages, write, 0, p, NULL);
 550
 551                 spin_lock(&gtt->guptasklock);
 552                 list_del(&guptask.list);
 553                 spin_unlock(&gtt->guptasklock);
 554
 555                 if (r < 0)
 556                         goto release_pages;
 557
 558                 pinned += r;
 559
 560         } while (pinned < ttm->num_pages);
 561
 562         return 0;
 563
 564 release_pages:
 565         release_pages(pages, pinned, 0);
 566         return r;
 567 }
 568
 569 /* prepare the sg table with the user pages */
 570 static int amdgpu_ttm_tt_pin_userptr(struct ttm_tt *ttm)
 571 {
 572         struct amdgpu_device *adev = amdgpu_get_adev(ttm->bdev);
 573         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 574         unsigned nents;
 575         int r;
 576
 577         int write = !(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
 578         enum dma_data_direction direction = write ?
 579                 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
 580
 581         r = sg_alloc_table_from_pages(ttm->sg, ttm->pages, ttm->num_pages, 0,
 582                                       ttm->num_pages << PAGE_SHIFT,
 583                                       GFP_KERNEL);
 584         if (r)
 585                 goto release_sg;
 586
 587         r = -ENOMEM;
 588         nents = dma_map_sg(adev->dev, ttm->sg->sgl, ttm->sg->nents, direction);
 589         if (nents != ttm->sg->nents)
 590                 goto release_sg;
 591
 592         drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
 593                                          gtt->ttm.dma_address, ttm->num_pages);
 594
 595         return 0;
 596
 597 release_sg:
 598         kfree(ttm->sg);
 599         return r;
 600 }
 601
 602 static void amdgpu_ttm_tt_unpin_userptr(struct ttm_tt *ttm)
 603 {
 604         struct amdgpu_device *adev = amdgpu_get_adev(ttm->bdev);
 605         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 606         struct sg_page_iter sg_iter;
 607
 608         int write = !(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
 609         enum dma_data_direction direction = write ?
 610                 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
 611
 612         /* double check that we don't free the table twice */
 613         if (!ttm->sg->sgl)
 614                 return;
 615
 616         /* free the sg table and pages again */
 617         dma_unmap_sg(adev->dev, ttm->sg->sgl, ttm->sg->nents, direction);
 618
 619         for_each_sg_page(ttm->sg->sgl, &sg_iter, ttm->sg->nents, 0) {
 620                 struct page *page = sg_page_iter_page(&sg_iter);
 621                 if (!(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY))
 622                         set_page_dirty(page);
 623
 624                 mark_page_accessed(page);
 625                 page_cache_release(page);
 626         }
 627
 628         sg_free_table(ttm->sg);
 629 }
 630
 631 static int amdgpu_ttm_backend_bind(struct ttm_tt *ttm,
 632                                    struct ttm_mem_reg *bo_mem)
 633 {
 634         struct amdgpu_ttm_tt *gtt = (void*)ttm;
 635         uint32_t flags = amdgpu_ttm_tt_pte_flags(gtt->adev, ttm, bo_mem);
 636         int r;
 637
 638         if (gtt->userptr) {
 639                 r = amdgpu_ttm_tt_pin_userptr(ttm);
 640                 if (r) {
 641                         DRM_ERROR("failed to pin userptr\n");
 642                         return r;
 643                 }
 644         }
 645         gtt->offset = (unsigned long)(bo_mem->start << PAGE_SHIFT);
 646         if (!ttm->num_pages) {
 647                 WARN(1, "nothing to bind %lu pages for mreg %p back %p!\n",
 648                      ttm->num_pages, bo_mem, ttm);
 649         }
 650
 651         if (bo_mem->mem_type == AMDGPU_PL_GDS ||
 652             bo_mem->mem_type == AMDGPU_PL_GWS ||
 653             bo_mem->mem_type == AMDGPU_PL_OA)
 654                 return -EINVAL;
 655
 656         r = amdgpu_gart_bind(gtt->adev, gtt->offset, ttm->num_pages,
 657                 ttm->pages, gtt->ttm.dma_address, flags);
 658
 659         if (r) {
 660                 DRM_ERROR("failed to bind %lu pages at 0x%08X\n",
 661                           ttm->num_pages, (unsigned)gtt->offset);
 662                 return r;
 663         }
 664         return 0;
 665 }
 666
 667 static int amdgpu_ttm_backend_unbind(struct ttm_tt *ttm)
 668 {
 669         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 670
 671         /* unbind shouldn't be done for GDS/GWS/OA in ttm_bo_clean_mm */
 672         if (gtt->adev->gart.ready)
 673                 amdgpu_gart_unbind(gtt->adev, gtt->offset, ttm->num_pages);
 674
 675         if (gtt->userptr)
 676                 amdgpu_ttm_tt_unpin_userptr(ttm);
 677
 678         return 0;
 679 }
 680
 681 static void amdgpu_ttm_backend_destroy(struct ttm_tt *ttm)
 682 {
 683         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 684
 685         ttm_dma_tt_fini(&gtt->ttm);
 686         kfree(gtt);
 687 }
 688
 689 static struct ttm_backend_func amdgpu_backend_func = {
 690         .bind = &amdgpu_ttm_backend_bind,
 691         .unbind = &amdgpu_ttm_backend_unbind,
 692         .destroy = &amdgpu_ttm_backend_destroy,
 693 };
 694
 695 static struct ttm_tt *amdgpu_ttm_tt_create(struct ttm_bo_device *bdev,
 696                                     unsigned long size, uint32_t page_flags,
 697                                     struct page *dummy_read_page)
 698 {
 699         struct amdgpu_device *adev;
 700         struct amdgpu_ttm_tt *gtt;
 701
 702         adev = amdgpu_get_adev(bdev);
 703
 704         gtt = kzalloc(sizeof(struct amdgpu_ttm_tt), GFP_KERNEL);
 705         if (gtt == NULL) {
 706                 return NULL;
 707         }
 708         gtt->ttm.ttm.func = &amdgpu_backend_func;
 709         gtt->adev = adev;
 710         if (ttm_dma_tt_init(&gtt->ttm, bdev, size, page_flags, dummy_read_page)) {
 711                 kfree(gtt);
 712                 return NULL;
 713         }
 714         return &gtt->ttm.ttm;
 715 }
 716
 717 static int amdgpu_ttm_tt_populate(struct ttm_tt *ttm)
 718 {
 719         struct amdgpu_device *adev;
 720         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 721         unsigned i;
 722         int r;
 723         bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
 724
 725         if (ttm->state != tt_unpopulated)
 726                 return 0;
 727
 728         if (gtt && gtt->userptr) {
 729                 ttm->sg = kzalloc(sizeof(struct sg_table), GFP_KERNEL);
 730                 if (!ttm->sg)
 731                         return -ENOMEM;
 732
 733                 ttm->page_flags |= TTM_PAGE_FLAG_SG;
 734                 ttm->state = tt_unbound;
 735                 return 0;
 736         }
 737
 738         if (slave && ttm->sg) {
 739                 drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
 740                                                  gtt->ttm.dma_address, ttm->num_pages);
 741                 ttm->state = tt_unbound;
 742                 return 0;
 743         }
 744
 745         adev = amdgpu_get_adev(ttm->bdev);
 746
 747 #ifdef CONFIG_SWIOTLB
 748         if (swiotlb_nr_tbl()) {
 749                 return ttm_dma_populate(&gtt->ttm, adev->dev);
 750         }
 751 #endif
 752
 753         r = ttm_pool_populate(ttm);
 754         if (r) {
 755                 return r;
 756         }
 757
 758         for (i = 0; i < ttm->num_pages; i++) {
 759                 gtt->ttm.dma_address[i] = pci_map_page(adev->pdev, ttm->pages[i],
 760                                                        0, PAGE_SIZE,
 761                                                        PCI_DMA_BIDIRECTIONAL);
 762                 if (pci_dma_mapping_error(adev->pdev, gtt->ttm.dma_address[i])) {
 763                         while (i--) {
 764                                 pci_unmap_page(adev->pdev, gtt->ttm.dma_address[i],
 765                                                PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
 766                                 gtt->ttm.dma_address[i] = 0;
 767                         }
 768                         ttm_pool_unpopulate(ttm);
 769                         return -EFAULT;
 770                 }
 771         }
 772         return 0;
 773 }
 774
 775 static void amdgpu_ttm_tt_unpopulate(struct ttm_tt *ttm)
 776 {
 777         struct amdgpu_device *adev;
 778         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 779         unsigned i;
 780         bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
 781
 782         if (gtt && gtt->userptr) {
 783                 kfree(ttm->sg);
 784                 ttm->page_flags &= ~TTM_PAGE_FLAG_SG;
 785                 return;
 786         }
 787
 788         if (slave)
 789                 return;
 790
 791         adev = amdgpu_get_adev(ttm->bdev);
 792
 793 #ifdef CONFIG_SWIOTLB
 794         if (swiotlb_nr_tbl()) {
 795                 ttm_dma_unpopulate(&gtt->ttm, adev->dev);
 796                 return;
 797         }
 798 #endif
 799
 800         for (i = 0; i < ttm->num_pages; i++) {
 801                 if (gtt->ttm.dma_address[i]) {
 802                         pci_unmap_page(adev->pdev, gtt->ttm.dma_address[i],
 803                                        PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
 804                 }
 805         }
 806
 807         ttm_pool_unpopulate(ttm);
 808 }
 809
 810 int amdgpu_ttm_tt_set_userptr(struct ttm_tt *ttm, uint64_t addr,
 811                               uint32_t flags)
 812 {
 813         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 814
 815         if (gtt == NULL)
 816                 return -EINVAL;
 817
 818         gtt->userptr = addr;
 819         gtt->usermm = current->mm;
 820         gtt->userflags = flags;
 821         spin_lock_init(&gtt->guptasklock);
 822         INIT_LIST_HEAD(&gtt->guptasks);
 823         atomic_set(&gtt->mmu_invalidations, 0);
 824
 825         return 0;
 826 }
 827
 828 struct mm_struct *amdgpu_ttm_tt_get_usermm(struct ttm_tt *ttm)
 829 {
 830         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 831
 832         if (gtt == NULL)
 833                 return NULL;
 834
 835         return gtt->usermm;
 836 }
 837
 838 bool amdgpu_ttm_tt_affect_userptr(struct ttm_tt *ttm, unsigned long start,
 839                                   unsigned long end)
 840 {
 841         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 842         struct amdgpu_ttm_gup_task_list *entry;
 843         unsigned long size;
 844
 845         if (gtt == NULL || !gtt->userptr)
 846                 return false;
 847
 848         size = (unsigned long)gtt->ttm.ttm.num_pages * PAGE_SIZE;
 849         if (gtt->userptr > end || gtt->userptr + size <= start)
 850                 return false;
 851
 852         spin_lock(&gtt->guptasklock);
 853         list_for_each_entry(entry, &gtt->guptasks, list) {
 854                 if (entry->task == current) {
 855                         spin_unlock(&gtt->guptasklock);
 856                         return false;
 857                 }
 858         }
 859         spin_unlock(&gtt->guptasklock);
 860
 861         atomic_inc(&gtt->mmu_invalidations);
 862
 863         return true;
 864 }
 865
 866 bool amdgpu_ttm_tt_userptr_invalidated(struct ttm_tt *ttm,
 867                                        int *last_invalidated)
 868 {
 869         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 870         int prev_invalidated = *last_invalidated;
 871
 872         *last_invalidated = atomic_read(&gtt->mmu_invalidations);
 873         return prev_invalidated != *last_invalidated;
 874 }
 875
 876 bool amdgpu_ttm_tt_is_readonly(struct ttm_tt *ttm)
 877 {
 878         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 879
 880         if (gtt == NULL)
 881                 return false;
 882
 883         return !!(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
 884 }
 885
 886 uint32_t amdgpu_ttm_tt_pte_flags(struct amdgpu_device *adev, struct ttm_tt *ttm,
 887                                  struct ttm_mem_reg *mem)
 888 {
 889         uint32_t flags = 0;
 890
 891         if (mem && mem->mem_type != TTM_PL_SYSTEM)
 892                 flags |= AMDGPU_PTE_VALID;
 893
 894         if (mem && mem->mem_type == TTM_PL_TT) {
 895                 flags |= AMDGPU_PTE_SYSTEM;
 896
 897                 if (ttm->caching_state == tt_cached)
 898                         flags |= AMDGPU_PTE_SNOOPED;
 899         }
 900
 901         if (adev->asic_type >= CHIP_TONGA)
 902                 flags |= AMDGPU_PTE_EXECUTABLE;
 903
 904         flags |= AMDGPU_PTE_READABLE;
 905
 906         if (!amdgpu_ttm_tt_is_readonly(ttm))
 907                 flags |= AMDGPU_PTE_WRITEABLE;
 908
 909         return flags;
 910 }
 911
 912 static struct ttm_bo_driver amdgpu_bo_driver = {
 913         .ttm_tt_create = &amdgpu_ttm_tt_create,
 914         .ttm_tt_populate = &amdgpu_ttm_tt_populate,
 915         .ttm_tt_unpopulate = &amdgpu_ttm_tt_unpopulate,
 916         .invalidate_caches = &amdgpu_invalidate_caches,
 917         .init_mem_type = &amdgpu_init_mem_type,
 918         .evict_flags = &amdgpu_evict_flags,
 919         .move = &amdgpu_bo_move,
 920         .verify_access = &amdgpu_verify_access,
 921         .move_notify = &amdgpu_bo_move_notify,
 922         .fault_reserve_notify = &amdgpu_bo_fault_reserve_notify,
 923         .io_mem_reserve = &amdgpu_ttm_io_mem_reserve,
 924         .io_mem_free = &amdgpu_ttm_io_mem_free,
 925 };
 926
 927 int amdgpu_ttm_init(struct amdgpu_device *adev)
 928 {
 929         int r;
 930
 931         r = amdgpu_ttm_global_init(adev);
 932         if (r) {
 933                 return r;
 934         }
 935         /* No others user of address space so set it to 0 */
 936         r = ttm_bo_device_init(&adev->mman.bdev,
 937                                adev->mman.bo_global_ref.ref.object,
 938                                &amdgpu_bo_driver,
 939                                adev->ddev->anon_inode->i_mapping,
 940                                DRM_FILE_PAGE_OFFSET,
 941                                adev->need_dma32);
 942         if (r) {
 943                 DRM_ERROR("failed initializing buffer object driver(%d).\n", r);
 944                 return r;
 945         }
 946         adev->mman.initialized = true;
 947         r = ttm_bo_init_mm(&adev->mman.bdev, TTM_PL_VRAM,
 948                                 adev->mc.real_vram_size >> PAGE_SHIFT);
 949         if (r) {
 950                 DRM_ERROR("Failed initializing VRAM heap.\n");
 951                 return r;
 952         }
 953         /* Change the size here instead of the init above so only lpfn is affected */
 954         amdgpu_ttm_set_active_vram_size(adev, adev->mc.visible_vram_size);
 955
 956         r = amdgpu_bo_create(adev, 256 * 1024, PAGE_SIZE, true,
 957                              AMDGPU_GEM_DOMAIN_VRAM,
 958                              AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
 959                              NULL, NULL, &adev->stollen_vga_memory);
 960         if (r) {
 961                 return r;
 962         }
 963         r = amdgpu_bo_reserve(adev->stollen_vga_memory, false);
 964         if (r)
 965                 return r;
 966         r = amdgpu_bo_pin(adev->stollen_vga_memory, AMDGPU_GEM_DOMAIN_VRAM, NULL);
 967         amdgpu_bo_unreserve(adev->stollen_vga_memory);
 968         if (r) {
 969                 amdgpu_bo_unref(&adev->stollen_vga_memory);
 970                 return r;
 971         }
 972         DRM_INFO("amdgpu: %uM of VRAM memory ready\n",
 973                  (unsigned) (adev->mc.real_vram_size / (1024 * 1024)));
 974         r = ttm_bo_init_mm(&adev->mman.bdev, TTM_PL_TT,
 975                                 adev->mc.gtt_size >> PAGE_SHIFT);
 976         if (r) {
 977                 DRM_ERROR("Failed initializing GTT heap.\n");
 978                 return r;
 979         }
 980         DRM_INFO("amdgpu: %uM of GTT memory ready.\n",
 981                  (unsigned)(adev->mc.gtt_size / (1024 * 1024)));
 982
 983         adev->gds.mem.total_size = adev->gds.mem.total_size << AMDGPU_GDS_SHIFT;
 984         adev->gds.mem.gfx_partition_size = adev->gds.mem.gfx_partition_size << AMDGPU_GDS_SHIFT;
 985         adev->gds.mem.cs_partition_size = adev->gds.mem.cs_partition_size << AMDGPU_GDS_SHIFT;
 986         adev->gds.gws.total_size = adev->gds.gws.total_size << AMDGPU_GWS_SHIFT;
 987         adev->gds.gws.gfx_partition_size = adev->gds.gws.gfx_partition_size << AMDGPU_GWS_SHIFT;
 988         adev->gds.gws.cs_partition_size = adev->gds.gws.cs_partition_size << AMDGPU_GWS_SHIFT;
 989         adev->gds.oa.total_size = adev->gds.oa.total_size << AMDGPU_OA_SHIFT;
 990         adev->gds.oa.gfx_partition_size = adev->gds.oa.gfx_partition_size << AMDGPU_OA_SHIFT;
 991         adev->gds.oa.cs_partition_size = adev->gds.oa.cs_partition_size << AMDGPU_OA_SHIFT;
 992         /* GDS Memory */
 993         r = ttm_bo_init_mm(&adev->mman.bdev, AMDGPU_PL_GDS,
 994                                 adev->gds.mem.total_size >> PAGE_SHIFT);
 995         if (r) {
 996                 DRM_ERROR("Failed initializing GDS heap.\n");
 997                 return r;
 998         }
 999
1000         /* GWS */
1001         r = ttm_bo_init_mm(&adev->mman.bdev, AMDGPU_PL_GWS,
1002                                 adev->gds.gws.total_size >> PAGE_SHIFT);
1003         if (r) {
1004                 DRM_ERROR("Failed initializing gws heap.\n");
1005                 return r;
1006         }
1007
1008         /* OA */
1009         r = ttm_bo_init_mm(&adev->mman.bdev, AMDGPU_PL_OA,
1010                                 adev->gds.oa.total_size >> PAGE_SHIFT);
1011         if (r) {
1012                 DRM_ERROR("Failed initializing oa heap.\n");
1013                 return r;
1014         }
1015
1016         r = amdgpu_ttm_debugfs_init(adev);
1017         if (r) {
1018                 DRM_ERROR("Failed to init debugfs\n");
1019                 return r;
1020         }
1021         return 0;
1022 }
1023
1024 void amdgpu_ttm_fini(struct amdgpu_device *adev)
1025 {
1026         int r;
1027
1028         if (!adev->mman.initialized)
1029                 return;
1030         amdgpu_ttm_debugfs_fini(adev);
1031         if (adev->stollen_vga_memory) {
1032                 r = amdgpu_bo_reserve(adev->stollen_vga_memory, false);
1033                 if (r == 0) {
1034                         amdgpu_bo_unpin(adev->stollen_vga_memory);
1035                         amdgpu_bo_unreserve(adev->stollen_vga_memory);
1036                 }
1037                 amdgpu_bo_unref(&adev->stollen_vga_memory);
1038         }
1039         ttm_bo_clean_mm(&adev->mman.bdev, TTM_PL_VRAM);
1040         ttm_bo_clean_mm(&adev->mman.bdev, TTM_PL_TT);
1041         ttm_bo_clean_mm(&adev->mman.bdev, AMDGPU_PL_GDS);
1042         ttm_bo_clean_mm(&adev->mman.bdev, AMDGPU_PL_GWS);
1043         ttm_bo_clean_mm(&adev->mman.bdev, AMDGPU_PL_OA);
1044         ttm_bo_device_release(&adev->mman.bdev);
1045         amdgpu_gart_fini(adev);
1046         amdgpu_ttm_global_fini(adev);
1047         adev->mman.initialized = false;
1048         DRM_INFO("amdgpu: ttm finalized\n");
1049 }
1050
1051 /* this should only be called at bootup or when userspace
1052  * isn't running */
1053 void amdgpu_ttm_set_active_vram_size(struct amdgpu_device *adev, u64 size)
1054 {
1055         struct ttm_mem_type_manager *man;
1056
1057         if (!adev->mman.initialized)
1058                 return;
1059
1060         man = &adev->mman.bdev.man[TTM_PL_VRAM];
1061         /* this just adjusts TTM size idea, which sets lpfn to the correct value */
1062         man->size = size >> PAGE_SHIFT;
1063 }
1064
1065 int amdgpu_mmap(struct file *filp, struct vm_area_struct *vma)
1066 {
1067         struct drm_file *file_priv;
1068         struct amdgpu_device *adev;
1069
1070         if (unlikely(vma->vm_pgoff < DRM_FILE_PAGE_OFFSET))
1071                 return -EINVAL;
1072
1073         file_priv = filp->private_data;
1074         adev = file_priv->minor->dev->dev_private;
1075         if (adev == NULL)
1076                 return -EINVAL;
1077
1078         return ttm_bo_mmap(filp, vma, &adev->mman.bdev);
1079 }
1080
1081 int amdgpu_copy_buffer(struct amdgpu_ring *ring,
1082                        uint64_t src_offset,
1083                        uint64_t dst_offset,
1084                        uint32_t byte_count,
1085                        struct reservation_object *resv,
1086                        struct fence **fence)
1087 {
1088         struct amdgpu_device *adev = ring->adev;
1089         struct amdgpu_job *job;
1090
1091         uint32_t max_bytes;
1092         unsigned num_loops, num_dw;
1093         unsigned i;
1094         int r;
1095
1096         max_bytes = adev->mman.buffer_funcs->copy_max_bytes;
1097         num_loops = DIV_ROUND_UP(byte_count, max_bytes);
1098         num_dw = num_loops * adev->mman.buffer_funcs->copy_num_dw;
1099
1100         /* for IB padding */
1101         while (num_dw & 0x7)
1102                 num_dw++;
1103
1104         r = amdgpu_job_alloc_with_ib(adev, num_dw * 4, &job);
1105         if (r)
1106                 return r;
1107
1108         if (resv) {
1109                 r = amdgpu_sync_resv(adev, &job->sync, resv,
1110                                      AMDGPU_FENCE_OWNER_UNDEFINED);
1111                 if (r) {
1112                         DRM_ERROR("sync failed (%d).\n", r);
1113                         goto error_free;
1114                 }
1115         }
1116
1117         for (i = 0; i < num_loops; i++) {
1118                 uint32_t cur_size_in_bytes = min(byte_count, max_bytes);
1119
1120                 amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_offset,
1121                                         dst_offset, cur_size_in_bytes);
1122
1123                 src_offset += cur_size_in_bytes;
1124                 dst_offset += cur_size_in_bytes;
1125                 byte_count -= cur_size_in_bytes;
1126         }
1127
1128         amdgpu_ring_pad_ib(ring, &job->ibs[0]);
1129         WARN_ON(job->ibs[0].length_dw > num_dw);
1130         r = amdgpu_job_submit(job, ring, &adev->mman.entity,
1131                               AMDGPU_FENCE_OWNER_UNDEFINED, fence);
1132         if (r)
1133                 goto error_free;
1134
1135         return 0;
1136
1137 error_free:
1138         amdgpu_job_free(job);
1139         return r;
1140 }
1141
1142 #if defined(CONFIG_DEBUG_FS)
1143
1144 static int amdgpu_mm_dump_table(struct seq_file *m, void *data)
1145 {
1146         struct drm_info_node *node = (struct drm_info_node *)m->private;
1147         unsigned ttm_pl = *(int *)node->info_ent->data;
1148         struct drm_device *dev = node->minor->dev;
1149         struct amdgpu_device *adev = dev->dev_private;
1150         struct drm_mm *mm = (struct drm_mm *)adev->mman.bdev.man[ttm_pl].priv;
1151         int ret;
1152         struct ttm_bo_global *glob = adev->mman.bdev.glob;
1153
1154         spin_lock(&glob->lru_lock);
1155         ret = drm_mm_dump_table(m, mm);
1156         spin_unlock(&glob->lru_lock);
1157         if (ttm_pl == TTM_PL_VRAM)
1158                 seq_printf(m, "man size:%llu pages, ram usage:%lluMB, vis usage:%lluMB\n",
1159                            adev->mman.bdev.man[ttm_pl].size,
1160                            (u64)atomic64_read(&adev->vram_usage) >> 20,
1161                            (u64)atomic64_read(&adev->vram_vis_usage) >> 20);
1162         return ret;
1163 }
1164
1165 static int ttm_pl_vram = TTM_PL_VRAM;
1166 static int ttm_pl_tt = TTM_PL_TT;
1167
1168 static struct drm_info_list amdgpu_ttm_debugfs_list[] = {
1169         {"amdgpu_vram_mm", amdgpu_mm_dump_table, 0, &ttm_pl_vram},
1170         {"amdgpu_gtt_mm", amdgpu_mm_dump_table, 0, &ttm_pl_tt},
1171         {"ttm_page_pool", ttm_page_alloc_debugfs, 0, NULL},
1172 #ifdef CONFIG_SWIOTLB
1173         {"ttm_dma_page_pool", ttm_dma_page_alloc_debugfs, 0, NULL}
1174 #endif
1175 };
1176
1177 static ssize_t amdgpu_ttm_vram_read(struct file *f, char __user *buf,
1178                                     size_t size, loff_t *pos)
1179 {
1180         struct amdgpu_device *adev = f->f_inode->i_private;
1181         ssize_t result = 0;
1182         int r;
1183
1184         if (size & 0x3 || *pos & 0x3)
1185                 return -EINVAL;
1186
1187         while (size) {
1188                 unsigned long flags;
1189                 uint32_t value;
1190
1191                 if (*pos >= adev->mc.mc_vram_size)
1192                         return result;
1193
1194                 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
1195                 WREG32(mmMM_INDEX, ((uint32_t)*pos) | 0x80000000);
1196                 WREG32(mmMM_INDEX_HI, *pos >> 31);
1197                 value = RREG32(mmMM_DATA);
1198                 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
1199
1200                 r = put_user(value, (uint32_t *)buf);
1201                 if (r)
1202                         return r;
1203
1204                 result += 4;
1205                 buf += 4;
1206                 *pos += 4;
1207                 size -= 4;
1208         }
1209
1210         return result;
1211 }
1212
1213 static const struct file_operations amdgpu_ttm_vram_fops = {
1214         .owner = THIS_MODULE,
1215         .read = amdgpu_ttm_vram_read,
1216         .llseek = default_llseek
1217 };
1218
1219 static ssize_t amdgpu_ttm_gtt_read(struct file *f, char __user *buf,
1220                                    size_t size, loff_t *pos)
1221 {
1222         struct amdgpu_device *adev = f->f_inode->i_private;
1223         ssize_t result = 0;
1224         int r;
1225
1226         while (size) {
1227                 loff_t p = *pos / PAGE_SIZE;
1228                 unsigned off = *pos & ~PAGE_MASK;
1229                 size_t cur_size = min_t(size_t, size, PAGE_SIZE - off);
1230                 struct page *page;
1231                 void *ptr;
1232
1233                 if (p >= adev->gart.num_cpu_pages)
1234                         return result;
1235
1236                 page = adev->gart.pages[p];
1237                 if (page) {
1238                         ptr = kmap(page);
1239                         ptr += off;
1240
1241                         r = copy_to_user(buf, ptr, cur_size);
1242                         kunmap(adev->gart.pages[p]);
1243                 } else
1244                         r = clear_user(buf, cur_size);
1245
1246                 if (r)
1247                         return -EFAULT;
1248
1249                 result += cur_size;
1250                 buf += cur_size;
1251                 *pos += cur_size;
1252                 size -= cur_size;
1253         }
1254
1255         return result;
1256 }
1257
1258 static const struct file_operations amdgpu_ttm_gtt_fops = {
1259         .owner = THIS_MODULE,
1260         .read = amdgpu_ttm_gtt_read,
1261         .llseek = default_llseek
1262 };
1263
1264 #endif
1265
1266 static int amdgpu_ttm_debugfs_init(struct amdgpu_device *adev)
1267 {
1268 #if defined(CONFIG_DEBUG_FS)
1269         unsigned count;
1270
1271         struct drm_minor *minor = adev->ddev->primary;
1272         struct dentry *ent, *root = minor->debugfs_root;
1273
1274         ent = debugfs_create_file("amdgpu_vram", S_IFREG | S_IRUGO, root,
1275                                   adev, &amdgpu_ttm_vram_fops);
1276         if (IS_ERR(ent))
1277                 return PTR_ERR(ent);
1278         i_size_write(ent->d_inode, adev->mc.mc_vram_size);
1279         adev->mman.vram = ent;
1280
1281         ent = debugfs_create_file("amdgpu_gtt", S_IFREG | S_IRUGO, root,
1282                                   adev, &amdgpu_ttm_gtt_fops);
1283         if (IS_ERR(ent))
1284                 return PTR_ERR(ent);
1285         i_size_write(ent->d_inode, adev->mc.gtt_size);
1286         adev->mman.gtt = ent;
1287
1288         count = ARRAY_SIZE(amdgpu_ttm_debugfs_list);
1289
1290 #ifdef CONFIG_SWIOTLB
1291         if (!swiotlb_nr_tbl())
1292                 --count;
1293 #endif
1294
1295         return amdgpu_debugfs_add_files(adev, amdgpu_ttm_debugfs_list, count);
1296 #else
1297
1298         return 0;
1299 #endif
1300 }
1301
1302 static void amdgpu_ttm_debugfs_fini(struct amdgpu_device *adev)
1303 {
1304 #if defined(CONFIG_DEBUG_FS)
1305
1306         debugfs_remove(adev->mman.vram);
1307         adev->mman.vram = NULL;
1308
1309         debugfs_remove(adev->mman.gtt);
1310         adev->mman.gtt = NULL;
1311 #endif
1312 }