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 <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 <ttm/ttm_memory.h>
  38 #include <drm/drmP.h>
  39 #include <drm/amdgpu_drm.h>
  40 #include <linux/seq_file.h>
  41 #include <linux/slab.h>
  42 #include <linux/swiotlb.h>
  43 #include <linux/swap.h>
  44 #include <linux/pagemap.h>
  45 #include <linux/debugfs.h>
  46 #include "amdgpu.h"
  47 #include "bif/bif_4_1_d.h"
  48
  49 #define DRM_FILE_PAGE_OFFSET (0x100000000ULL >> PAGE_SHIFT)
  50
  51 static int amdgpu_ttm_debugfs_init(struct amdgpu_device *adev);
  52 static void amdgpu_ttm_debugfs_fini(struct amdgpu_device *adev);
  53
  54 static struct amdgpu_device *amdgpu_get_adev(struct ttm_bo_device *bdev)
  55 {
  56         struct amdgpu_mman *mman;
  57         struct amdgpu_device *adev;
  58
  59         mman = container_of(bdev, struct amdgpu_mman, bdev);
  60         adev = container_of(mman, struct amdgpu_device, mman);
  61         return adev;
  62 }
  63
  64
  65 /*
  66  * Global memory.
  67  */
  68 static int amdgpu_ttm_mem_global_init(struct drm_global_reference *ref)
  69 {
  70         return ttm_mem_global_init(ref->object);
  71 }
  72
  73 static void amdgpu_ttm_mem_global_release(struct drm_global_reference *ref)
  74 {
  75         ttm_mem_global_release(ref->object);
  76 }
  77
  78 int amdgpu_ttm_global_init(struct amdgpu_device *adev)
  79 {
  80         struct drm_global_reference *global_ref;
  81         struct amdgpu_ring *ring;
  82         struct amd_sched_rq *rq;
  83         int r;
  84
  85         adev->mman.mem_global_referenced = false;
  86         global_ref = &adev->mman.mem_global_ref;
  87         global_ref->global_type = DRM_GLOBAL_TTM_MEM;
  88         global_ref->size = sizeof(struct ttm_mem_global);
  89         global_ref->init = &amdgpu_ttm_mem_global_init;
  90         global_ref->release = &amdgpu_ttm_mem_global_release;
  91         r = drm_global_item_ref(global_ref);
  92         if (r) {
  93                 DRM_ERROR("Failed setting up TTM memory accounting "
  94                           "subsystem.\n");
  95                 goto error_mem;
  96         }
  97
  98         adev->mman.bo_global_ref.mem_glob =
  99                 adev->mman.mem_global_ref.object;
 100         global_ref = &adev->mman.bo_global_ref.ref;
 101         global_ref->global_type = DRM_GLOBAL_TTM_BO;
 102         global_ref->size = sizeof(struct ttm_bo_global);
 103         global_ref->init = &ttm_bo_global_init;
 104         global_ref->release = &ttm_bo_global_release;
 105         r = drm_global_item_ref(global_ref);
 106         if (r) {
 107                 DRM_ERROR("Failed setting up TTM BO subsystem.\n");
 108                 goto error_bo;
 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) {
 116                 DRM_ERROR("Failed setting up TTM BO move run queue.\n");
 117                 goto error_entity;
 118         }
 119
 120         adev->mman.mem_global_referenced = true;
 121
 122         return 0;
 123
 124 error_entity:
 125         drm_global_item_unref(&adev->mman.bo_global_ref.ref);
 126 error_bo:
 127         drm_global_item_unref(&adev->mman.mem_global_ref);
 128 error_mem:
 129         return r;
 130 }
 131
 132 static void amdgpu_ttm_global_fini(struct amdgpu_device *adev)
 133 {
 134         if (adev->mman.mem_global_referenced) {
 135                 amd_sched_entity_fini(adev->mman.entity.sched,
 136                                       &adev->mman.entity);
 137                 drm_global_item_unref(&adev->mman.bo_global_ref.ref);
 138                 drm_global_item_unref(&adev->mman.mem_global_ref);
 139                 adev->mman.mem_global_referenced = false;
 140         }
 141 }
 142
 143 static int amdgpu_invalidate_caches(struct ttm_bo_device *bdev, uint32_t flags)
 144 {
 145         return 0;
 146 }
 147
 148 static int amdgpu_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
 149                                 struct ttm_mem_type_manager *man)
 150 {
 151         struct amdgpu_device *adev;
 152
 153         adev = amdgpu_get_adev(bdev);
 154
 155         switch (type) {
 156         case TTM_PL_SYSTEM:
 157                 /* System memory */
 158                 man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
 159                 man->available_caching = TTM_PL_MASK_CACHING;
 160                 man->default_caching = TTM_PL_FLAG_CACHED;
 161                 break;
 162         case TTM_PL_TT:
 163                 man->func = &amdgpu_gtt_mgr_func;
 164                 man->gpu_offset = adev->mc.gtt_start;
 165                 man->available_caching = TTM_PL_MASK_CACHING;
 166                 man->default_caching = TTM_PL_FLAG_CACHED;
 167                 man->flags = TTM_MEMTYPE_FLAG_MAPPABLE | TTM_MEMTYPE_FLAG_CMA;
 168                 break;
 169         case TTM_PL_VRAM:
 170                 /* "On-card" video ram */
 171                 man->func = &ttm_bo_manager_func;
 172                 man->gpu_offset = adev->mc.vram_start;
 173                 man->flags = TTM_MEMTYPE_FLAG_FIXED |
 174                              TTM_MEMTYPE_FLAG_MAPPABLE;
 175                 man->available_caching = TTM_PL_FLAG_UNCACHED | TTM_PL_FLAG_WC;
 176                 man->default_caching = TTM_PL_FLAG_WC;
 177                 break;
 178         case AMDGPU_PL_GDS:
 179         case AMDGPU_PL_GWS:
 180         case AMDGPU_PL_OA:
 181                 /* On-chip GDS memory*/
 182                 man->func = &ttm_bo_manager_func;
 183                 man->gpu_offset = 0;
 184                 man->flags = TTM_MEMTYPE_FLAG_FIXED | TTM_MEMTYPE_FLAG_CMA;
 185                 man->available_caching = TTM_PL_FLAG_UNCACHED;
 186                 man->default_caching = TTM_PL_FLAG_UNCACHED;
 187                 break;
 188         default:
 189                 DRM_ERROR("Unsupported memory type %u\n", (unsigned)type);
 190                 return -EINVAL;
 191         }
 192         return 0;
 193 }
 194
 195 static void amdgpu_evict_flags(struct ttm_buffer_object *bo,
 196                                 struct ttm_placement *placement)
 197 {
 198         struct amdgpu_bo *abo;
 199         static struct ttm_place placements = {
 200                 .fpfn = 0,
 201                 .lpfn = 0,
 202                 .flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM
 203         };
 204         unsigned i;
 205
 206         if (!amdgpu_ttm_bo_is_amdgpu_bo(bo)) {
 207                 placement->placement = &placements;
 208                 placement->busy_placement = &placements;
 209                 placement->num_placement = 1;
 210                 placement->num_busy_placement = 1;
 211                 return;
 212         }
 213         abo = container_of(bo, struct amdgpu_bo, tbo);
 214         switch (bo->mem.mem_type) {
 215         case TTM_PL_VRAM:
 216                 if (abo->adev->mman.buffer_funcs_ring->ready == false) {
 217                         amdgpu_ttm_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_CPU);
 218                 } else {
 219                         amdgpu_ttm_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_GTT);
 220                         for (i = 0; i < abo->placement.num_placement; ++i) {
 221                                 if (!(abo->placements[i].flags &
 222                                       TTM_PL_FLAG_TT))
 223                                         continue;
 224
 225                                 if (abo->placements[i].lpfn)
 226                                         continue;
 227
 228                                 /* set an upper limit to force directly
 229                                  * allocating address space for the BO.
 230                                  */
 231                                 abo->placements[i].lpfn =
 232                                         abo->adev->mc.gtt_size >> PAGE_SHIFT;
 233                         }
 234                 }
 235                 break;
 236         case TTM_PL_TT:
 237         default:
 238                 amdgpu_ttm_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_CPU);
 239         }
 240         *placement = abo->placement;
 241 }
 242
 243 static int amdgpu_verify_access(struct ttm_buffer_object *bo, struct file *filp)
 244 {
 245         struct amdgpu_bo *abo = container_of(bo, struct amdgpu_bo, tbo);
 246
 247         if (amdgpu_ttm_tt_get_usermm(bo->ttm))
 248                 return -EPERM;
 249         return drm_vma_node_verify_access(&abo->gem_base.vma_node,
 250                                           filp->private_data);
 251 }
 252
 253 static void amdgpu_move_null(struct ttm_buffer_object *bo,
 254                              struct ttm_mem_reg *new_mem)
 255 {
 256         struct ttm_mem_reg *old_mem = &bo->mem;
 257
 258         BUG_ON(old_mem->mm_node != NULL);
 259         *old_mem = *new_mem;
 260         new_mem->mm_node = NULL;
 261 }
 262
 263 static int amdgpu_move_blit(struct ttm_buffer_object *bo,
 264                         bool evict, bool no_wait_gpu,
 265                         struct ttm_mem_reg *new_mem,
 266                         struct ttm_mem_reg *old_mem)
 267 {
 268         struct amdgpu_device *adev;
 269         struct amdgpu_ring *ring;
 270         uint64_t old_start, new_start;
 271         struct fence *fence;
 272         int r;
 273
 274         adev = amdgpu_get_adev(bo->bdev);
 275         ring = adev->mman.buffer_funcs_ring;
 276
 277         switch (old_mem->mem_type) {
 278         case TTM_PL_TT:
 279                 r = amdgpu_ttm_bind(bo, old_mem);
 280                 if (r)
 281                         return r;
 282
 283         case TTM_PL_VRAM:
 284                 old_start = (u64)old_mem->start << PAGE_SHIFT;
 285                 old_start += bo->bdev->man[old_mem->mem_type].gpu_offset;
 286                 break;
 287         default:
 288                 DRM_ERROR("Unknown placement %d\n", old_mem->mem_type);
 289                 return -EINVAL;
 290         }
 291         switch (new_mem->mem_type) {
 292         case TTM_PL_TT:
 293                 r = amdgpu_ttm_bind(bo, new_mem);
 294                 if (r)
 295                         return r;
 296
 297         case TTM_PL_VRAM:
 298                 new_start = (u64)new_mem->start << PAGE_SHIFT;
 299                 new_start += bo->bdev->man[new_mem->mem_type].gpu_offset;
 300                 break;
 301         default:
 302                 DRM_ERROR("Unknown placement %d\n", old_mem->mem_type);
 303                 return -EINVAL;
 304         }
 305         if (!ring->ready) {
 306                 DRM_ERROR("Trying to move memory with ring turned off.\n");
 307                 return -EINVAL;
 308         }
 309
 310         BUILD_BUG_ON((PAGE_SIZE % AMDGPU_GPU_PAGE_SIZE) != 0);
 311
 312         r = amdgpu_copy_buffer(ring, old_start, new_start,
 313                                new_mem->num_pages * PAGE_SIZE, /* bytes */
 314                                bo->resv, &fence, false);
 315         if (r)
 316                 return r;
 317
 318         r = ttm_bo_pipeline_move(bo, fence, evict, new_mem);
 319         fence_put(fence);
 320         return r;
 321 }
 322
 323 static int amdgpu_move_vram_ram(struct ttm_buffer_object *bo,
 324                                 bool evict, bool interruptible,
 325                                 bool no_wait_gpu,
 326                                 struct ttm_mem_reg *new_mem)
 327 {
 328         struct amdgpu_device *adev;
 329         struct ttm_mem_reg *old_mem = &bo->mem;
 330         struct ttm_mem_reg tmp_mem;
 331         struct ttm_place placements;
 332         struct ttm_placement placement;
 333         int r;
 334
 335         adev = amdgpu_get_adev(bo->bdev);
 336         tmp_mem = *new_mem;
 337         tmp_mem.mm_node = NULL;
 338         placement.num_placement = 1;
 339         placement.placement = &placements;
 340         placement.num_busy_placement = 1;
 341         placement.busy_placement = &placements;
 342         placements.fpfn = 0;
 343         placements.lpfn = adev->mc.gtt_size >> PAGE_SHIFT;
 344         placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
 345         r = ttm_bo_mem_space(bo, &placement, &tmp_mem,
 346                              interruptible, no_wait_gpu);
 347         if (unlikely(r)) {
 348                 return r;
 349         }
 350
 351         r = ttm_tt_set_placement_caching(bo->ttm, tmp_mem.placement);
 352         if (unlikely(r)) {
 353                 goto out_cleanup;
 354         }
 355
 356         r = ttm_tt_bind(bo->ttm, &tmp_mem);
 357         if (unlikely(r)) {
 358                 goto out_cleanup;
 359         }
 360         r = amdgpu_move_blit(bo, true, no_wait_gpu, &tmp_mem, old_mem);
 361         if (unlikely(r)) {
 362                 goto out_cleanup;
 363         }
 364         r = ttm_bo_move_ttm(bo, interruptible, no_wait_gpu, new_mem);
 365 out_cleanup:
 366         ttm_bo_mem_put(bo, &tmp_mem);
 367         return r;
 368 }
 369
 370 static int amdgpu_move_ram_vram(struct ttm_buffer_object *bo,
 371                                 bool evict, bool interruptible,
 372                                 bool no_wait_gpu,
 373                                 struct ttm_mem_reg *new_mem)
 374 {
 375         struct amdgpu_device *adev;
 376         struct ttm_mem_reg *old_mem = &bo->mem;
 377         struct ttm_mem_reg tmp_mem;
 378         struct ttm_placement placement;
 379         struct ttm_place placements;
 380         int r;
 381
 382         adev = amdgpu_get_adev(bo->bdev);
 383         tmp_mem = *new_mem;
 384         tmp_mem.mm_node = NULL;
 385         placement.num_placement = 1;
 386         placement.placement = &placements;
 387         placement.num_busy_placement = 1;
 388         placement.busy_placement = &placements;
 389         placements.fpfn = 0;
 390         placements.lpfn = adev->mc.gtt_size >> PAGE_SHIFT;
 391         placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
 392         r = ttm_bo_mem_space(bo, &placement, &tmp_mem,
 393                              interruptible, no_wait_gpu);
 394         if (unlikely(r)) {
 395                 return r;
 396         }
 397         r = ttm_bo_move_ttm(bo, interruptible, no_wait_gpu, &tmp_mem);
 398         if (unlikely(r)) {
 399                 goto out_cleanup;
 400         }
 401         r = amdgpu_move_blit(bo, true, no_wait_gpu, new_mem, old_mem);
 402         if (unlikely(r)) {
 403                 goto out_cleanup;
 404         }
 405 out_cleanup:
 406         ttm_bo_mem_put(bo, &tmp_mem);
 407         return r;
 408 }
 409
 410 static int amdgpu_bo_move(struct ttm_buffer_object *bo,
 411                         bool evict, bool interruptible,
 412                         bool no_wait_gpu,
 413                         struct ttm_mem_reg *new_mem)
 414 {
 415         struct amdgpu_device *adev;
 416         struct amdgpu_bo *abo;
 417         struct ttm_mem_reg *old_mem = &bo->mem;
 418         int r;
 419
 420         /* Can't move a pinned BO */
 421         abo = container_of(bo, struct amdgpu_bo, tbo);
 422         if (WARN_ON_ONCE(abo->pin_count > 0))
 423                 return -EINVAL;
 424
 425         adev = amdgpu_get_adev(bo->bdev);
 426
 427         /* remember the eviction */
 428         if (evict)
 429                 atomic64_inc(&adev->num_evictions);
 430
 431         if (old_mem->mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
 432                 amdgpu_move_null(bo, new_mem);
 433                 return 0;
 434         }
 435         if ((old_mem->mem_type == TTM_PL_TT &&
 436              new_mem->mem_type == TTM_PL_SYSTEM) ||
 437             (old_mem->mem_type == TTM_PL_SYSTEM &&
 438              new_mem->mem_type == TTM_PL_TT)) {
 439                 /* bind is enough */
 440                 amdgpu_move_null(bo, new_mem);
 441                 return 0;
 442         }
 443         if (adev->mman.buffer_funcs == NULL ||
 444             adev->mman.buffer_funcs_ring == NULL ||
 445             !adev->mman.buffer_funcs_ring->ready) {
 446                 /* use memcpy */
 447                 goto memcpy;
 448         }
 449
 450         if (old_mem->mem_type == TTM_PL_VRAM &&
 451             new_mem->mem_type == TTM_PL_SYSTEM) {
 452                 r = amdgpu_move_vram_ram(bo, evict, interruptible,
 453                                         no_wait_gpu, new_mem);
 454         } else if (old_mem->mem_type == TTM_PL_SYSTEM &&
 455                    new_mem->mem_type == TTM_PL_VRAM) {
 456                 r = amdgpu_move_ram_vram(bo, evict, interruptible,
 457                                             no_wait_gpu, new_mem);
 458         } else {
 459                 r = amdgpu_move_blit(bo, evict, no_wait_gpu, new_mem, old_mem);
 460         }
 461
 462         if (r) {
 463 memcpy:
 464                 r = ttm_bo_move_memcpy(bo, interruptible, no_wait_gpu, new_mem);
 465                 if (r) {
 466                         return r;
 467                 }
 468         }
 469
 470         /* update statistics */
 471         atomic64_add((u64)bo->num_pages << PAGE_SHIFT, &adev->num_bytes_moved);
 472         return 0;
 473 }
 474
 475 static int amdgpu_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
 476 {
 477         struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
 478         struct amdgpu_device *adev = amdgpu_get_adev(bdev);
 479
 480         mem->bus.addr = NULL;
 481         mem->bus.offset = 0;
 482         mem->bus.size = mem->num_pages << PAGE_SHIFT;
 483         mem->bus.base = 0;
 484         mem->bus.is_iomem = false;
 485         if (!(man->flags & TTM_MEMTYPE_FLAG_MAPPABLE))
 486                 return -EINVAL;
 487         switch (mem->mem_type) {
 488         case TTM_PL_SYSTEM:
 489                 /* system memory */
 490                 return 0;
 491         case TTM_PL_TT:
 492                 break;
 493         case TTM_PL_VRAM:
 494                 mem->bus.offset = mem->start << PAGE_SHIFT;
 495                 /* check if it's visible */
 496                 if ((mem->bus.offset + mem->bus.size) > adev->mc.visible_vram_size)
 497                         return -EINVAL;
 498                 mem->bus.base = adev->mc.aper_base;
 499                 mem->bus.is_iomem = true;
 500 #ifdef __alpha__
 501                 /*
 502                  * Alpha: use bus.addr to hold the ioremap() return,
 503                  * so we can modify bus.base below.
 504                  */
 505                 if (mem->placement & TTM_PL_FLAG_WC)
 506                         mem->bus.addr =
 507                                 ioremap_wc(mem->bus.base + mem->bus.offset,
 508                                            mem->bus.size);
 509                 else
 510                         mem->bus.addr =
 511                                 ioremap_nocache(mem->bus.base + mem->bus.offset,
 512                                                 mem->bus.size);
 513
 514                 /*
 515                  * Alpha: Use just the bus offset plus
 516                  * the hose/domain memory base for bus.base.
 517                  * It then can be used to build PTEs for VRAM
 518                  * access, as done in ttm_bo_vm_fault().
 519                  */
 520                 mem->bus.base = (mem->bus.base & 0x0ffffffffUL) +
 521                         adev->ddev->hose->dense_mem_base;
 522 #endif
 523                 break;
 524         default:
 525                 return -EINVAL;
 526         }
 527         return 0;
 528 }
 529
 530 static void amdgpu_ttm_io_mem_free(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
 531 {
 532 }
 533
 534 /*
 535  * TTM backend functions.
 536  */
 537 struct amdgpu_ttm_gup_task_list {
 538         struct list_head        list;
 539         struct task_struct      *task;
 540 };
 541
 542 struct amdgpu_ttm_tt {
 543         struct ttm_dma_tt       ttm;
 544         struct amdgpu_device    *adev;
 545         u64                     offset;
 546         uint64_t                userptr;
 547         struct mm_struct        *usermm;
 548         uint32_t                userflags;
 549         spinlock_t              guptasklock;
 550         struct list_head        guptasks;
 551         atomic_t                mmu_invalidations;
 552         struct list_head        list;
 553 };
 554
 555 int amdgpu_ttm_tt_get_user_pages(struct ttm_tt *ttm, struct page **pages)
 556 {
 557         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 558         int write = !(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
 559         unsigned pinned = 0;
 560         int r;
 561
 562         if (gtt->userflags & AMDGPU_GEM_USERPTR_ANONONLY) {
 563                 /* check that we only use anonymous memory
 564                    to prevent problems with writeback */
 565                 unsigned long end = gtt->userptr + ttm->num_pages * PAGE_SIZE;
 566                 struct vm_area_struct *vma;
 567
 568                 vma = find_vma(gtt->usermm, gtt->userptr);
 569                 if (!vma || vma->vm_file || vma->vm_end < end)
 570                         return -EPERM;
 571         }
 572
 573         do {
 574                 unsigned num_pages = ttm->num_pages - pinned;
 575                 uint64_t userptr = gtt->userptr + pinned * PAGE_SIZE;
 576                 struct page **p = pages + pinned;
 577                 struct amdgpu_ttm_gup_task_list guptask;
 578
 579                 guptask.task = current;
 580                 spin_lock(&gtt->guptasklock);
 581                 list_add(&guptask.list, &gtt->guptasks);
 582                 spin_unlock(&gtt->guptasklock);
 583
 584                 r = get_user_pages(userptr, num_pages, write, 0, p, NULL);
 585
 586                 spin_lock(&gtt->guptasklock);
 587                 list_del(&guptask.list);
 588                 spin_unlock(&gtt->guptasklock);
 589
 590                 if (r < 0)
 591                         goto release_pages;
 592
 593                 pinned += r;
 594
 595         } while (pinned < ttm->num_pages);
 596
 597         return 0;
 598
 599 release_pages:
 600         release_pages(pages, pinned, 0);
 601         return r;
 602 }
 603
 604 /* prepare the sg table with the user pages */
 605 static int amdgpu_ttm_tt_pin_userptr(struct ttm_tt *ttm)
 606 {
 607         struct amdgpu_device *adev = amdgpu_get_adev(ttm->bdev);
 608         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 609         unsigned nents;
 610         int r;
 611
 612         int write = !(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
 613         enum dma_data_direction direction = write ?
 614                 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
 615
 616         r = sg_alloc_table_from_pages(ttm->sg, ttm->pages, ttm->num_pages, 0,
 617                                       ttm->num_pages << PAGE_SHIFT,
 618                                       GFP_KERNEL);
 619         if (r)
 620                 goto release_sg;
 621
 622         r = -ENOMEM;
 623         nents = dma_map_sg(adev->dev, ttm->sg->sgl, ttm->sg->nents, direction);
 624         if (nents != ttm->sg->nents)
 625                 goto release_sg;
 626
 627         drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
 628                                          gtt->ttm.dma_address, ttm->num_pages);
 629
 630         return 0;
 631
 632 release_sg:
 633         kfree(ttm->sg);
 634         return r;
 635 }
 636
 637 static void amdgpu_ttm_tt_unpin_userptr(struct ttm_tt *ttm)
 638 {
 639         struct amdgpu_device *adev = amdgpu_get_adev(ttm->bdev);
 640         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 641         struct sg_page_iter sg_iter;
 642
 643         int write = !(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
 644         enum dma_data_direction direction = write ?
 645                 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
 646
 647         /* double check that we don't free the table twice */
 648         if (!ttm->sg->sgl)
 649                 return;
 650
 651         /* free the sg table and pages again */
 652         dma_unmap_sg(adev->dev, ttm->sg->sgl, ttm->sg->nents, direction);
 653
 654         for_each_sg_page(ttm->sg->sgl, &sg_iter, ttm->sg->nents, 0) {
 655                 struct page *page = sg_page_iter_page(&sg_iter);
 656                 if (!(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY))
 657                         set_page_dirty(page);
 658
 659                 mark_page_accessed(page);
 660                 put_page(page);
 661         }
 662
 663         sg_free_table(ttm->sg);
 664 }
 665
 666 static int amdgpu_ttm_backend_bind(struct ttm_tt *ttm,
 667                                    struct ttm_mem_reg *bo_mem)
 668 {
 669         struct amdgpu_ttm_tt *gtt = (void*)ttm;
 670         int r;
 671
 672         if (gtt->userptr) {
 673                 r = amdgpu_ttm_tt_pin_userptr(ttm);
 674                 if (r) {
 675                         DRM_ERROR("failed to pin userptr\n");
 676                         return r;
 677                 }
 678         }
 679         if (!ttm->num_pages) {
 680                 WARN(1, "nothing to bind %lu pages for mreg %p back %p!\n",
 681                      ttm->num_pages, bo_mem, ttm);
 682         }
 683
 684         if (bo_mem->mem_type == AMDGPU_PL_GDS ||
 685             bo_mem->mem_type == AMDGPU_PL_GWS ||
 686             bo_mem->mem_type == AMDGPU_PL_OA)
 687                 return -EINVAL;
 688
 689         return 0;
 690 }
 691
 692 bool amdgpu_ttm_is_bound(struct ttm_tt *ttm)
 693 {
 694         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 695
 696         return gtt && !list_empty(&gtt->list);
 697 }
 698
 699 int amdgpu_ttm_bind(struct ttm_buffer_object *bo, struct ttm_mem_reg *bo_mem)
 700 {
 701         struct ttm_tt *ttm = bo->ttm;
 702         struct amdgpu_ttm_tt *gtt = (void *)bo->ttm;
 703         uint32_t flags;
 704         int r;
 705
 706         if (!ttm || amdgpu_ttm_is_bound(ttm))
 707                 return 0;
 708
 709         r = amdgpu_gtt_mgr_alloc(&bo->bdev->man[TTM_PL_TT], bo,
 710                                  NULL, bo_mem);
 711         if (r) {
 712                 DRM_ERROR("Failed to allocate GTT address space (%d)\n", r);
 713                 return r;
 714         }
 715
 716         flags = amdgpu_ttm_tt_pte_flags(gtt->adev, ttm, bo_mem);
 717         gtt->offset = (u64)bo_mem->start << PAGE_SHIFT;
 718         r = amdgpu_gart_bind(gtt->adev, gtt->offset, ttm->num_pages,
 719                 ttm->pages, gtt->ttm.dma_address, flags);
 720
 721         if (r) {
 722                 DRM_ERROR("failed to bind %lu pages at 0x%08llX\n",
 723                           ttm->num_pages, gtt->offset);
 724                 return r;
 725         }
 726         spin_lock(&gtt->adev->gtt_list_lock);
 727         list_add_tail(&gtt->list, &gtt->adev->gtt_list);
 728         spin_unlock(&gtt->adev->gtt_list_lock);
 729         return 0;
 730 }
 731
 732 int amdgpu_ttm_recover_gart(struct amdgpu_device *adev)
 733 {
 734         struct amdgpu_ttm_tt *gtt, *tmp;
 735         struct ttm_mem_reg bo_mem;
 736         uint32_t flags;
 737         int r;
 738
 739         bo_mem.mem_type = TTM_PL_TT;
 740         spin_lock(&adev->gtt_list_lock);
 741         list_for_each_entry_safe(gtt, tmp, &adev->gtt_list, list) {
 742                 flags = amdgpu_ttm_tt_pte_flags(gtt->adev, &gtt->ttm.ttm, &bo_mem);
 743                 r = amdgpu_gart_bind(adev, gtt->offset, gtt->ttm.ttm.num_pages,
 744                                      gtt->ttm.ttm.pages, gtt->ttm.dma_address,
 745                                      flags);
 746                 if (r) {
 747                         spin_unlock(&adev->gtt_list_lock);
 748                         DRM_ERROR("failed to bind %lu pages at 0x%08llX\n",
 749                                   gtt->ttm.ttm.num_pages, gtt->offset);
 750                         return r;
 751                 }
 752         }
 753         spin_unlock(&adev->gtt_list_lock);
 754         return 0;
 755 }
 756
 757 static int amdgpu_ttm_backend_unbind(struct ttm_tt *ttm)
 758 {
 759         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 760
 761         if (gtt->userptr)
 762                 amdgpu_ttm_tt_unpin_userptr(ttm);
 763
 764         if (!amdgpu_ttm_is_bound(ttm))
 765                 return 0;
 766
 767         /* unbind shouldn't be done for GDS/GWS/OA in ttm_bo_clean_mm */
 768         if (gtt->adev->gart.ready)
 769                 amdgpu_gart_unbind(gtt->adev, gtt->offset, ttm->num_pages);
 770
 771         spin_lock(&gtt->adev->gtt_list_lock);
 772         list_del_init(&gtt->list);
 773         spin_unlock(&gtt->adev->gtt_list_lock);
 774
 775         return 0;
 776 }
 777
 778 static void amdgpu_ttm_backend_destroy(struct ttm_tt *ttm)
 779 {
 780         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 781
 782         ttm_dma_tt_fini(&gtt->ttm);
 783         kfree(gtt);
 784 }
 785
 786 static struct ttm_backend_func amdgpu_backend_func = {
 787         .bind = &amdgpu_ttm_backend_bind,
 788         .unbind = &amdgpu_ttm_backend_unbind,
 789         .destroy = &amdgpu_ttm_backend_destroy,
 790 };
 791
 792 static struct ttm_tt *amdgpu_ttm_tt_create(struct ttm_bo_device *bdev,
 793                                     unsigned long size, uint32_t page_flags,
 794                                     struct page *dummy_read_page)
 795 {
 796         struct amdgpu_device *adev;
 797         struct amdgpu_ttm_tt *gtt;
 798
 799         adev = amdgpu_get_adev(bdev);
 800
 801         gtt = kzalloc(sizeof(struct amdgpu_ttm_tt), GFP_KERNEL);
 802         if (gtt == NULL) {
 803                 return NULL;
 804         }
 805         gtt->ttm.ttm.func = &amdgpu_backend_func;
 806         gtt->adev = adev;
 807         if (ttm_dma_tt_init(&gtt->ttm, bdev, size, page_flags, dummy_read_page)) {
 808                 kfree(gtt);
 809                 return NULL;
 810         }
 811         INIT_LIST_HEAD(&gtt->list);
 812         return &gtt->ttm.ttm;
 813 }
 814
 815 static int amdgpu_ttm_tt_populate(struct ttm_tt *ttm)
 816 {
 817         struct amdgpu_device *adev;
 818         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 819         unsigned i;
 820         int r;
 821         bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
 822
 823         if (ttm->state != tt_unpopulated)
 824                 return 0;
 825
 826         if (gtt && gtt->userptr) {
 827                 ttm->sg = kzalloc(sizeof(struct sg_table), GFP_KERNEL);
 828                 if (!ttm->sg)
 829                         return -ENOMEM;
 830
 831                 ttm->page_flags |= TTM_PAGE_FLAG_SG;
 832                 ttm->state = tt_unbound;
 833                 return 0;
 834         }
 835
 836         if (slave && ttm->sg) {
 837                 drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
 838                                                  gtt->ttm.dma_address, ttm->num_pages);
 839                 ttm->state = tt_unbound;
 840                 return 0;
 841         }
 842
 843         adev = amdgpu_get_adev(ttm->bdev);
 844
 845 #ifdef CONFIG_SWIOTLB
 846         if (swiotlb_nr_tbl()) {
 847                 return ttm_dma_populate(&gtt->ttm, adev->dev);
 848         }
 849 #endif
 850
 851         r = ttm_pool_populate(ttm);
 852         if (r) {
 853                 return r;
 854         }
 855
 856         for (i = 0; i < ttm->num_pages; i++) {
 857                 gtt->ttm.dma_address[i] = pci_map_page(adev->pdev, ttm->pages[i],
 858                                                        0, PAGE_SIZE,
 859                                                        PCI_DMA_BIDIRECTIONAL);
 860                 if (pci_dma_mapping_error(adev->pdev, gtt->ttm.dma_address[i])) {
 861                         while (i--) {
 862                                 pci_unmap_page(adev->pdev, gtt->ttm.dma_address[i],
 863                                                PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
 864                                 gtt->ttm.dma_address[i] = 0;
 865                         }
 866                         ttm_pool_unpopulate(ttm);
 867                         return -EFAULT;
 868                 }
 869         }
 870         return 0;
 871 }
 872
 873 static void amdgpu_ttm_tt_unpopulate(struct ttm_tt *ttm)
 874 {
 875         struct amdgpu_device *adev;
 876         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 877         unsigned i;
 878         bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
 879
 880         if (gtt && gtt->userptr) {
 881                 kfree(ttm->sg);
 882                 ttm->page_flags &= ~TTM_PAGE_FLAG_SG;
 883                 return;
 884         }
 885
 886         if (slave)
 887                 return;
 888
 889         adev = amdgpu_get_adev(ttm->bdev);
 890
 891 #ifdef CONFIG_SWIOTLB
 892         if (swiotlb_nr_tbl()) {
 893                 ttm_dma_unpopulate(&gtt->ttm, adev->dev);
 894                 return;
 895         }
 896 #endif
 897
 898         for (i = 0; i < ttm->num_pages; i++) {
 899                 if (gtt->ttm.dma_address[i]) {
 900                         pci_unmap_page(adev->pdev, gtt->ttm.dma_address[i],
 901                                        PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
 902                 }
 903         }
 904
 905         ttm_pool_unpopulate(ttm);
 906 }
 907
 908 int amdgpu_ttm_tt_set_userptr(struct ttm_tt *ttm, uint64_t addr,
 909                               uint32_t flags)
 910 {
 911         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 912
 913         if (gtt == NULL)
 914                 return -EINVAL;
 915
 916         gtt->userptr = addr;
 917         gtt->usermm = current->mm;
 918         gtt->userflags = flags;
 919         spin_lock_init(&gtt->guptasklock);
 920         INIT_LIST_HEAD(&gtt->guptasks);
 921         atomic_set(&gtt->mmu_invalidations, 0);
 922
 923         return 0;
 924 }
 925
 926 struct mm_struct *amdgpu_ttm_tt_get_usermm(struct ttm_tt *ttm)
 927 {
 928         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 929
 930         if (gtt == NULL)
 931                 return NULL;
 932
 933         return gtt->usermm;
 934 }
 935
 936 bool amdgpu_ttm_tt_affect_userptr(struct ttm_tt *ttm, unsigned long start,
 937                                   unsigned long end)
 938 {
 939         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 940         struct amdgpu_ttm_gup_task_list *entry;
 941         unsigned long size;
 942
 943         if (gtt == NULL || !gtt->userptr)
 944                 return false;
 945
 946         size = (unsigned long)gtt->ttm.ttm.num_pages * PAGE_SIZE;
 947         if (gtt->userptr > end || gtt->userptr + size <= start)
 948                 return false;
 949
 950         spin_lock(&gtt->guptasklock);
 951         list_for_each_entry(entry, &gtt->guptasks, list) {
 952                 if (entry->task == current) {
 953                         spin_unlock(&gtt->guptasklock);
 954                         return false;
 955                 }
 956         }
 957         spin_unlock(&gtt->guptasklock);
 958
 959         atomic_inc(&gtt->mmu_invalidations);
 960
 961         return true;
 962 }
 963
 964 bool amdgpu_ttm_tt_userptr_invalidated(struct ttm_tt *ttm,
 965                                        int *last_invalidated)
 966 {
 967         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 968         int prev_invalidated = *last_invalidated;
 969
 970         *last_invalidated = atomic_read(&gtt->mmu_invalidations);
 971         return prev_invalidated != *last_invalidated;
 972 }
 973
 974 bool amdgpu_ttm_tt_is_readonly(struct ttm_tt *ttm)
 975 {
 976         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 977
 978         if (gtt == NULL)
 979                 return false;
 980
 981         return !!(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
 982 }
 983
 984 uint32_t amdgpu_ttm_tt_pte_flags(struct amdgpu_device *adev, struct ttm_tt *ttm,
 985                                  struct ttm_mem_reg *mem)
 986 {
 987         uint32_t flags = 0;
 988
 989         if (mem && mem->mem_type != TTM_PL_SYSTEM)
 990                 flags |= AMDGPU_PTE_VALID;
 991
 992         if (mem && mem->mem_type == TTM_PL_TT) {
 993                 flags |= AMDGPU_PTE_SYSTEM;
 994
 995                 if (ttm->caching_state == tt_cached)
 996                         flags |= AMDGPU_PTE_SNOOPED;
 997         }
 998
 999         if (adev->asic_type >= CHIP_TONGA)
1000                 flags |= AMDGPU_PTE_EXECUTABLE;
1001
1002         flags |= AMDGPU_PTE_READABLE;
1003
1004         if (!amdgpu_ttm_tt_is_readonly(ttm))
1005                 flags |= AMDGPU_PTE_WRITEABLE;
1006
1007         return flags;
1008 }
1009
1010 static void amdgpu_ttm_lru_removal(struct ttm_buffer_object *tbo)
1011 {
1012         struct amdgpu_device *adev = amdgpu_get_adev(tbo->bdev);
1013         unsigned i, j;
1014
1015         for (i = 0; i < AMDGPU_TTM_LRU_SIZE; ++i) {
1016                 struct amdgpu_mman_lru *lru = &adev->mman.log2_size[i];
1017
1018                 for (j = 0; j < TTM_NUM_MEM_TYPES; ++j)
1019                         if (&tbo->lru == lru->lru[j])
1020                                 lru->lru[j] = tbo->lru.prev;
1021
1022                 if (&tbo->swap == lru->swap_lru)
1023                         lru->swap_lru = tbo->swap.prev;
1024         }
1025 }
1026
1027 static struct amdgpu_mman_lru *amdgpu_ttm_lru(struct ttm_buffer_object *tbo)
1028 {
1029         struct amdgpu_device *adev = amdgpu_get_adev(tbo->bdev);
1030         unsigned log2_size = min(ilog2(tbo->num_pages),
1031                                  AMDGPU_TTM_LRU_SIZE - 1);
1032
1033         return &adev->mman.log2_size[log2_size];
1034 }
1035
1036 static struct list_head *amdgpu_ttm_lru_tail(struct ttm_buffer_object *tbo)
1037 {
1038         struct amdgpu_mman_lru *lru = amdgpu_ttm_lru(tbo);
1039         struct list_head *res = lru->lru[tbo->mem.mem_type];
1040
1041         lru->lru[tbo->mem.mem_type] = &tbo->lru;
1042         while ((++lru)->lru[tbo->mem.mem_type] == res)
1043                 lru->lru[tbo->mem.mem_type] = &tbo->lru;
1044
1045         return res;
1046 }
1047
1048 static struct list_head *amdgpu_ttm_swap_lru_tail(struct ttm_buffer_object *tbo)
1049 {
1050         struct amdgpu_mman_lru *lru = amdgpu_ttm_lru(tbo);
1051         struct list_head *res = lru->swap_lru;
1052
1053         lru->swap_lru = &tbo->swap;
1054         while ((++lru)->swap_lru == res)
1055                 lru->swap_lru = &tbo->swap;
1056
1057         return res;
1058 }
1059
1060 static struct ttm_bo_driver amdgpu_bo_driver = {
1061         .ttm_tt_create = &amdgpu_ttm_tt_create,
1062         .ttm_tt_populate = &amdgpu_ttm_tt_populate,
1063         .ttm_tt_unpopulate = &amdgpu_ttm_tt_unpopulate,
1064         .invalidate_caches = &amdgpu_invalidate_caches,
1065         .init_mem_type = &amdgpu_init_mem_type,
1066         .evict_flags = &amdgpu_evict_flags,
1067         .move = &amdgpu_bo_move,
1068         .verify_access = &amdgpu_verify_access,
1069         .move_notify = &amdgpu_bo_move_notify,
1070         .fault_reserve_notify = &amdgpu_bo_fault_reserve_notify,
1071         .io_mem_reserve = &amdgpu_ttm_io_mem_reserve,
1072         .io_mem_free = &amdgpu_ttm_io_mem_free,
1073         .lru_removal = &amdgpu_ttm_lru_removal,
1074         .lru_tail = &amdgpu_ttm_lru_tail,
1075         .swap_lru_tail = &amdgpu_ttm_swap_lru_tail,
1076 };
1077
1078 int amdgpu_ttm_init(struct amdgpu_device *adev)
1079 {
1080         unsigned i, j;
1081         int r;
1082
1083         /* No others user of address space so set it to 0 */
1084         r = ttm_bo_device_init(&adev->mman.bdev,
1085                                adev->mman.bo_global_ref.ref.object,
1086                                &amdgpu_bo_driver,
1087                                adev->ddev->anon_inode->i_mapping,
1088                                DRM_FILE_PAGE_OFFSET,
1089                                adev->need_dma32);
1090         if (r) {
1091                 DRM_ERROR("failed initializing buffer object driver(%d).\n", r);
1092                 return r;
1093         }
1094
1095         for (i = 0; i < AMDGPU_TTM_LRU_SIZE; ++i) {
1096                 struct amdgpu_mman_lru *lru = &adev->mman.log2_size[i];
1097
1098                 for (j = 0; j < TTM_NUM_MEM_TYPES; ++j)
1099                         lru->lru[j] = &adev->mman.bdev.man[j].lru;
1100                 lru->swap_lru = &adev->mman.bdev.glob->swap_lru;
1101         }
1102
1103         for (j = 0; j < TTM_NUM_MEM_TYPES; ++j)
1104                 adev->mman.guard.lru[j] = NULL;
1105         adev->mman.guard.swap_lru = NULL;
1106
1107         adev->mman.initialized = true;
1108         r = ttm_bo_init_mm(&adev->mman.bdev, TTM_PL_VRAM,
1109                                 adev->mc.real_vram_size >> PAGE_SHIFT);
1110         if (r) {
1111                 DRM_ERROR("Failed initializing VRAM heap.\n");
1112                 return r;
1113         }
1114         /* Change the size here instead of the init above so only lpfn is affected */
1115         amdgpu_ttm_set_active_vram_size(adev, adev->mc.visible_vram_size);
1116
1117         r = amdgpu_bo_create(adev, 256 * 1024, PAGE_SIZE, true,
1118                              AMDGPU_GEM_DOMAIN_VRAM,
1119                              AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
1120                              NULL, NULL, &adev->stollen_vga_memory);
1121         if (r) {
1122                 return r;
1123         }
1124         r = amdgpu_bo_reserve(adev->stollen_vga_memory, false);
1125         if (r)
1126                 return r;
1127         r = amdgpu_bo_pin(adev->stollen_vga_memory, AMDGPU_GEM_DOMAIN_VRAM, NULL);
1128         amdgpu_bo_unreserve(adev->stollen_vga_memory);
1129         if (r) {
1130                 amdgpu_bo_unref(&adev->stollen_vga_memory);
1131                 return r;
1132         }
1133         DRM_INFO("amdgpu: %uM of VRAM memory ready\n",
1134                  (unsigned) (adev->mc.real_vram_size / (1024 * 1024)));
1135         r = ttm_bo_init_mm(&adev->mman.bdev, TTM_PL_TT,
1136                                 adev->mc.gtt_size >> PAGE_SHIFT);
1137         if (r) {
1138                 DRM_ERROR("Failed initializing GTT heap.\n");
1139                 return r;
1140         }
1141         DRM_INFO("amdgpu: %uM of GTT memory ready.\n",
1142                  (unsigned)(adev->mc.gtt_size / (1024 * 1024)));
1143
1144         adev->gds.mem.total_size = adev->gds.mem.total_size << AMDGPU_GDS_SHIFT;
1145         adev->gds.mem.gfx_partition_size = adev->gds.mem.gfx_partition_size << AMDGPU_GDS_SHIFT;
1146         adev->gds.mem.cs_partition_size = adev->gds.mem.cs_partition_size << AMDGPU_GDS_SHIFT;
1147         adev->gds.gws.total_size = adev->gds.gws.total_size << AMDGPU_GWS_SHIFT;
1148         adev->gds.gws.gfx_partition_size = adev->gds.gws.gfx_partition_size << AMDGPU_GWS_SHIFT;
1149         adev->gds.gws.cs_partition_size = adev->gds.gws.cs_partition_size << AMDGPU_GWS_SHIFT;
1150         adev->gds.oa.total_size = adev->gds.oa.total_size << AMDGPU_OA_SHIFT;
1151         adev->gds.oa.gfx_partition_size = adev->gds.oa.gfx_partition_size << AMDGPU_OA_SHIFT;
1152         adev->gds.oa.cs_partition_size = adev->gds.oa.cs_partition_size << AMDGPU_OA_SHIFT;
1153         /* GDS Memory */
1154         r = ttm_bo_init_mm(&adev->mman.bdev, AMDGPU_PL_GDS,
1155                                 adev->gds.mem.total_size >> PAGE_SHIFT);
1156         if (r) {
1157                 DRM_ERROR("Failed initializing GDS heap.\n");
1158                 return r;
1159         }
1160
1161         /* GWS */
1162         r = ttm_bo_init_mm(&adev->mman.bdev, AMDGPU_PL_GWS,
1163                                 adev->gds.gws.total_size >> PAGE_SHIFT);
1164         if (r) {
1165                 DRM_ERROR("Failed initializing gws heap.\n");
1166                 return r;
1167         }
1168
1169         /* OA */
1170         r = ttm_bo_init_mm(&adev->mman.bdev, AMDGPU_PL_OA,
1171                                 adev->gds.oa.total_size >> PAGE_SHIFT);
1172         if (r) {
1173                 DRM_ERROR("Failed initializing oa heap.\n");
1174                 return r;
1175         }
1176
1177         r = amdgpu_ttm_debugfs_init(adev);
1178         if (r) {
1179                 DRM_ERROR("Failed to init debugfs\n");
1180                 return r;
1181         }
1182         return 0;
1183 }
1184
1185 void amdgpu_ttm_fini(struct amdgpu_device *adev)
1186 {
1187         int r;
1188
1189         if (!adev->mman.initialized)
1190                 return;
1191         amdgpu_ttm_debugfs_fini(adev);
1192         if (adev->stollen_vga_memory) {
1193                 r = amdgpu_bo_reserve(adev->stollen_vga_memory, false);
1194                 if (r == 0) {
1195                         amdgpu_bo_unpin(adev->stollen_vga_memory);
1196                         amdgpu_bo_unreserve(adev->stollen_vga_memory);
1197                 }
1198                 amdgpu_bo_unref(&adev->stollen_vga_memory);
1199         }
1200         ttm_bo_clean_mm(&adev->mman.bdev, TTM_PL_VRAM);
1201         ttm_bo_clean_mm(&adev->mman.bdev, TTM_PL_TT);
1202         ttm_bo_clean_mm(&adev->mman.bdev, AMDGPU_PL_GDS);
1203         ttm_bo_clean_mm(&adev->mman.bdev, AMDGPU_PL_GWS);
1204         ttm_bo_clean_mm(&adev->mman.bdev, AMDGPU_PL_OA);
1205         ttm_bo_device_release(&adev->mman.bdev);
1206         amdgpu_gart_fini(adev);
1207         amdgpu_ttm_global_fini(adev);
1208         adev->mman.initialized = false;
1209         DRM_INFO("amdgpu: ttm finalized\n");
1210 }
1211
1212 /* this should only be called at bootup or when userspace
1213  * isn't running */
1214 void amdgpu_ttm_set_active_vram_size(struct amdgpu_device *adev, u64 size)
1215 {
1216         struct ttm_mem_type_manager *man;
1217
1218         if (!adev->mman.initialized)
1219                 return;
1220
1221         man = &adev->mman.bdev.man[TTM_PL_VRAM];
1222         /* this just adjusts TTM size idea, which sets lpfn to the correct value */
1223         man->size = size >> PAGE_SHIFT;
1224 }
1225
1226 int amdgpu_mmap(struct file *filp, struct vm_area_struct *vma)
1227 {
1228         struct drm_file *file_priv;
1229         struct amdgpu_device *adev;
1230
1231         if (unlikely(vma->vm_pgoff < DRM_FILE_PAGE_OFFSET))
1232                 return -EINVAL;
1233
1234         file_priv = filp->private_data;
1235         adev = file_priv->minor->dev->dev_private;
1236         if (adev == NULL)
1237                 return -EINVAL;
1238
1239         return ttm_bo_mmap(filp, vma, &adev->mman.bdev);
1240 }
1241
1242 int amdgpu_copy_buffer(struct amdgpu_ring *ring,
1243                        uint64_t src_offset,
1244                        uint64_t dst_offset,
1245                        uint32_t byte_count,
1246                        struct reservation_object *resv,
1247                        struct fence **fence, bool direct_submit)
1248 {
1249         struct amdgpu_device *adev = ring->adev;
1250         struct amdgpu_job *job;
1251
1252         uint32_t max_bytes;
1253         unsigned num_loops, num_dw;
1254         unsigned i;
1255         int r;
1256
1257         max_bytes = adev->mman.buffer_funcs->copy_max_bytes;
1258         num_loops = DIV_ROUND_UP(byte_count, max_bytes);
1259         num_dw = num_loops * adev->mman.buffer_funcs->copy_num_dw;
1260
1261         /* for IB padding */
1262         while (num_dw & 0x7)
1263                 num_dw++;
1264
1265         r = amdgpu_job_alloc_with_ib(adev, num_dw * 4, &job);
1266         if (r)
1267                 return r;
1268
1269         if (resv) {
1270                 r = amdgpu_sync_resv(adev, &job->sync, resv,
1271                                      AMDGPU_FENCE_OWNER_UNDEFINED);
1272                 if (r) {
1273                         DRM_ERROR("sync failed (%d).\n", r);
1274                         goto error_free;
1275                 }
1276         }
1277
1278         for (i = 0; i < num_loops; i++) {
1279                 uint32_t cur_size_in_bytes = min(byte_count, max_bytes);
1280
1281                 amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_offset,
1282                                         dst_offset, cur_size_in_bytes);
1283
1284                 src_offset += cur_size_in_bytes;
1285                 dst_offset += cur_size_in_bytes;
1286                 byte_count -= cur_size_in_bytes;
1287         }
1288
1289         amdgpu_ring_pad_ib(ring, &job->ibs[0]);
1290         WARN_ON(job->ibs[0].length_dw > num_dw);
1291         if (direct_submit) {
1292                 r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs,
1293                                        NULL, NULL, fence);
1294                 job->fence = fence_get(*fence);
1295                 if (r)
1296                         DRM_ERROR("Error scheduling IBs (%d)\n", r);
1297                 amdgpu_job_free(job);
1298         } else {
1299                 r = amdgpu_job_submit(job, ring, &adev->mman.entity,
1300                                       AMDGPU_FENCE_OWNER_UNDEFINED, fence);
1301                 if (r)
1302                         goto error_free;
1303         }
1304
1305         return r;
1306
1307 error_free:
1308         amdgpu_job_free(job);
1309         return r;
1310 }
1311
1312 int amdgpu_fill_buffer(struct amdgpu_bo *bo,
1313                 uint32_t src_data,
1314                 struct reservation_object *resv,
1315                 struct fence **fence)
1316 {
1317         struct amdgpu_device *adev = bo->adev;
1318         struct amdgpu_job *job;
1319         struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
1320
1321         uint32_t max_bytes, byte_count;
1322         uint64_t dst_offset;
1323         unsigned int num_loops, num_dw;
1324         unsigned int i;
1325         int r;
1326
1327         byte_count = bo->tbo.num_pages << PAGE_SHIFT;
1328         max_bytes = adev->mman.buffer_funcs->fill_max_bytes;
1329         num_loops = DIV_ROUND_UP(byte_count, max_bytes);
1330         num_dw = num_loops * adev->mman.buffer_funcs->fill_num_dw;
1331
1332         /* for IB padding */
1333         while (num_dw & 0x7)
1334                 num_dw++;
1335
1336         r = amdgpu_job_alloc_with_ib(adev, num_dw * 4, &job);
1337         if (r)
1338                 return r;
1339
1340         if (resv) {
1341                 r = amdgpu_sync_resv(adev, &job->sync, resv,
1342                                 AMDGPU_FENCE_OWNER_UNDEFINED);
1343                 if (r) {
1344                         DRM_ERROR("sync failed (%d).\n", r);
1345                         goto error_free;
1346                 }
1347         }
1348
1349         dst_offset = bo->tbo.mem.start << PAGE_SHIFT;
1350         for (i = 0; i < num_loops; i++) {
1351                 uint32_t cur_size_in_bytes = min(byte_count, max_bytes);
1352
1353                 amdgpu_emit_fill_buffer(adev, &job->ibs[0], src_data,
1354                                 dst_offset, cur_size_in_bytes);
1355
1356                 dst_offset += cur_size_in_bytes;
1357                 byte_count -= cur_size_in_bytes;
1358         }
1359
1360         amdgpu_ring_pad_ib(ring, &job->ibs[0]);
1361         WARN_ON(job->ibs[0].length_dw > num_dw);
1362         r = amdgpu_job_submit(job, ring, &adev->mman.entity,
1363                         AMDGPU_FENCE_OWNER_UNDEFINED, fence);
1364         if (r)
1365                 goto error_free;
1366
1367         return 0;
1368
1369 error_free:
1370         amdgpu_job_free(job);
1371         return r;
1372 }
1373
1374 #if defined(CONFIG_DEBUG_FS)
1375
1376 static int amdgpu_mm_dump_table(struct seq_file *m, void *data)
1377 {
1378         struct drm_info_node *node = (struct drm_info_node *)m->private;
1379         unsigned ttm_pl = *(int *)node->info_ent->data;
1380         struct drm_device *dev = node->minor->dev;
1381         struct amdgpu_device *adev = dev->dev_private;
1382         struct drm_mm *mm = (struct drm_mm *)adev->mman.bdev.man[ttm_pl].priv;
1383         int ret;
1384         struct ttm_bo_global *glob = adev->mman.bdev.glob;
1385
1386         spin_lock(&glob->lru_lock);
1387         ret = drm_mm_dump_table(m, mm);
1388         spin_unlock(&glob->lru_lock);
1389         if (ttm_pl == TTM_PL_VRAM)
1390                 seq_printf(m, "man size:%llu pages, ram usage:%lluMB, vis usage:%lluMB\n",
1391                            adev->mman.bdev.man[ttm_pl].size,
1392                            (u64)atomic64_read(&adev->vram_usage) >> 20,
1393                            (u64)atomic64_read(&adev->vram_vis_usage) >> 20);
1394         return ret;
1395 }
1396
1397 static int ttm_pl_vram = TTM_PL_VRAM;
1398 static int ttm_pl_tt = TTM_PL_TT;
1399
1400 static const struct drm_info_list amdgpu_ttm_debugfs_list[] = {
1401         {"amdgpu_vram_mm", amdgpu_mm_dump_table, 0, &ttm_pl_vram},
1402         {"amdgpu_gtt_mm", amdgpu_mm_dump_table, 0, &ttm_pl_tt},
1403         {"ttm_page_pool", ttm_page_alloc_debugfs, 0, NULL},
1404 #ifdef CONFIG_SWIOTLB
1405         {"ttm_dma_page_pool", ttm_dma_page_alloc_debugfs, 0, NULL}
1406 #endif
1407 };
1408
1409 static ssize_t amdgpu_ttm_vram_read(struct file *f, char __user *buf,
1410                                     size_t size, loff_t *pos)
1411 {
1412         struct amdgpu_device *adev = f->f_inode->i_private;
1413         ssize_t result = 0;
1414         int r;
1415
1416         if (size & 0x3 || *pos & 0x3)
1417                 return -EINVAL;
1418
1419         while (size) {
1420                 unsigned long flags;
1421                 uint32_t value;
1422
1423                 if (*pos >= adev->mc.mc_vram_size)
1424                         return result;
1425
1426                 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
1427                 WREG32(mmMM_INDEX, ((uint32_t)*pos) | 0x80000000);
1428                 WREG32(mmMM_INDEX_HI, *pos >> 31);
1429                 value = RREG32(mmMM_DATA);
1430                 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
1431
1432                 r = put_user(value, (uint32_t *)buf);
1433                 if (r)
1434                         return r;
1435
1436                 result += 4;
1437                 buf += 4;
1438                 *pos += 4;
1439                 size -= 4;
1440         }
1441
1442         return result;
1443 }
1444
1445 static const struct file_operations amdgpu_ttm_vram_fops = {
1446         .owner = THIS_MODULE,
1447         .read = amdgpu_ttm_vram_read,
1448         .llseek = default_llseek
1449 };
1450
1451 #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS
1452
1453 static ssize_t amdgpu_ttm_gtt_read(struct file *f, char __user *buf,
1454                                    size_t size, loff_t *pos)
1455 {
1456         struct amdgpu_device *adev = f->f_inode->i_private;
1457         ssize_t result = 0;
1458         int r;
1459
1460         while (size) {
1461                 loff_t p = *pos / PAGE_SIZE;
1462                 unsigned off = *pos & ~PAGE_MASK;
1463                 size_t cur_size = min_t(size_t, size, PAGE_SIZE - off);
1464                 struct page *page;
1465                 void *ptr;
1466
1467                 if (p >= adev->gart.num_cpu_pages)
1468                         return result;
1469
1470                 page = adev->gart.pages[p];
1471                 if (page) {
1472                         ptr = kmap(page);
1473                         ptr += off;
1474
1475                         r = copy_to_user(buf, ptr, cur_size);
1476                         kunmap(adev->gart.pages[p]);
1477                 } else
1478                         r = clear_user(buf, cur_size);
1479
1480                 if (r)
1481                         return -EFAULT;
1482
1483                 result += cur_size;
1484                 buf += cur_size;
1485                 *pos += cur_size;
1486                 size -= cur_size;
1487         }
1488
1489         return result;
1490 }
1491
1492 static const struct file_operations amdgpu_ttm_gtt_fops = {
1493         .owner = THIS_MODULE,
1494         .read = amdgpu_ttm_gtt_read,
1495         .llseek = default_llseek
1496 };
1497
1498 #endif
1499
1500 #endif
1501
1502 static int amdgpu_ttm_debugfs_init(struct amdgpu_device *adev)
1503 {
1504 #if defined(CONFIG_DEBUG_FS)
1505         unsigned count;
1506
1507         struct drm_minor *minor = adev->ddev->primary;
1508         struct dentry *ent, *root = minor->debugfs_root;
1509
1510         ent = debugfs_create_file("amdgpu_vram", S_IFREG | S_IRUGO, root,
1511                                   adev, &amdgpu_ttm_vram_fops);
1512         if (IS_ERR(ent))
1513                 return PTR_ERR(ent);
1514         i_size_write(ent->d_inode, adev->mc.mc_vram_size);
1515         adev->mman.vram = ent;
1516
1517 #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS
1518         ent = debugfs_create_file("amdgpu_gtt", S_IFREG | S_IRUGO, root,
1519                                   adev, &amdgpu_ttm_gtt_fops);
1520         if (IS_ERR(ent))
1521                 return PTR_ERR(ent);
1522         i_size_write(ent->d_inode, adev->mc.gtt_size);
1523         adev->mman.gtt = ent;
1524
1525 #endif
1526         count = ARRAY_SIZE(amdgpu_ttm_debugfs_list);
1527
1528 #ifdef CONFIG_SWIOTLB
1529         if (!swiotlb_nr_tbl())
1530                 --count;
1531 #endif
1532
1533         return amdgpu_debugfs_add_files(adev, amdgpu_ttm_debugfs_list, count);
1534 #else
1535
1536         return 0;
1537 #endif
1538 }
1539
1540 static void amdgpu_ttm_debugfs_fini(struct amdgpu_device *adev)
1541 {
1542 #if defined(CONFIG_DEBUG_FS)
1543
1544         debugfs_remove(adev->mman.vram);
1545         adev->mman.vram = NULL;
1546
1547 #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS
1548         debugfs_remove(adev->mman.gtt);
1549         adev->mman.gtt = NULL;
1550 #endif
1551
1552 #endif
1553 }
1554
1555 u64 amdgpu_ttm_get_gtt_mem_size(struct amdgpu_device *adev)
1556 {
1557         return ttm_get_kernel_zone_memory_size(adev->mman.mem_global_ref.object);
1558 }