drivers/gpu/drm/amd/amdgpu/amdgpu_object.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 <linux/list.h>
  33 #include <linux/slab.h>
  34 #include <drm/drmP.h>
  35 #include <drm/amdgpu_drm.h>
  36 #include <drm/drm_cache.h>
  37 #include "amdgpu.h"
  38 #include "amdgpu_trace.h"
  39
  40 static void amdgpu_ttm_bo_destroy(struct ttm_buffer_object *tbo)
  41 {
  42         struct amdgpu_device *adev = amdgpu_ttm_adev(tbo->bdev);
  43         struct amdgpu_bo *bo;
  44
  45         bo = container_of(tbo, struct amdgpu_bo, tbo);
  46
  47         amdgpu_bo_kunmap(bo);
  48
  49         drm_gem_object_release(&bo->gem_base);
  50         amdgpu_bo_unref(&bo->parent);
  51         if (!list_empty(&bo->shadow_list)) {
  52                 mutex_lock(&adev->shadow_list_lock);
  53                 list_del_init(&bo->shadow_list);
  54                 mutex_unlock(&adev->shadow_list_lock);
  55         }
  56         kfree(bo->metadata);
  57         kfree(bo);
  58 }
  59
  60 bool amdgpu_ttm_bo_is_amdgpu_bo(struct ttm_buffer_object *bo)
  61 {
  62         if (bo->destroy == &amdgpu_ttm_bo_destroy)
  63                 return true;
  64         return false;
  65 }
  66
  67 static void amdgpu_ttm_placement_init(struct amdgpu_device *adev,
  68                                       struct ttm_placement *placement,
  69                                       struct ttm_place *places,
  70                                       u32 domain, u64 flags)
  71 {
  72         u32 c = 0;
  73
  74         if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
  75                 unsigned visible_pfn = adev->mc.visible_vram_size >> PAGE_SHIFT;
  76
  77                 places[c].fpfn = 0;
  78                 places[c].lpfn = 0;
  79                 places[c].flags = TTM_PL_FLAG_WC | TTM_PL_FLAG_UNCACHED |
  80                         TTM_PL_FLAG_VRAM;
  81
  82                 if (flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
  83                         places[c].lpfn = visible_pfn;
  84                 else
  85                         places[c].flags |= TTM_PL_FLAG_TOPDOWN;
  86
  87                 if (flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)
  88                         places[c].flags |= TTM_PL_FLAG_CONTIGUOUS;
  89                 c++;
  90         }
  91
  92         if (domain & AMDGPU_GEM_DOMAIN_GTT) {
  93                 places[c].fpfn = 0;
  94                 if (flags & AMDGPU_GEM_CREATE_SHADOW)
  95                         places[c].lpfn = adev->mc.gart_size >> PAGE_SHIFT;
  96                 else
  97                         places[c].lpfn = 0;
  98                 places[c].flags = TTM_PL_FLAG_TT;
  99                 if (flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
 100                         places[c].flags |= TTM_PL_FLAG_WC |
 101                                 TTM_PL_FLAG_UNCACHED;
 102                 else
 103                         places[c].flags |= TTM_PL_FLAG_CACHED;
 104                 c++;
 105         }
 106
 107         if (domain & AMDGPU_GEM_DOMAIN_CPU) {
 108                 places[c].fpfn = 0;
 109                 places[c].lpfn = 0;
 110                 places[c].flags = TTM_PL_FLAG_SYSTEM;
 111                 if (flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
 112                         places[c].flags |= TTM_PL_FLAG_WC |
 113                                 TTM_PL_FLAG_UNCACHED;
 114                 else
 115                         places[c].flags |= TTM_PL_FLAG_CACHED;
 116                 c++;
 117         }
 118
 119         if (domain & AMDGPU_GEM_DOMAIN_GDS) {
 120                 places[c].fpfn = 0;
 121                 places[c].lpfn = 0;
 122                 places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_GDS;
 123                 c++;
 124         }
 125
 126         if (domain & AMDGPU_GEM_DOMAIN_GWS) {
 127                 places[c].fpfn = 0;
 128                 places[c].lpfn = 0;
 129                 places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_GWS;
 130                 c++;
 131         }
 132
 133         if (domain & AMDGPU_GEM_DOMAIN_OA) {
 134                 places[c].fpfn = 0;
 135                 places[c].lpfn = 0;
 136                 places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_OA;
 137                 c++;
 138         }
 139
 140         if (!c) {
 141                 places[c].fpfn = 0;
 142                 places[c].lpfn = 0;
 143                 places[c].flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM;
 144                 c++;
 145         }
 146
 147         placement->num_placement = c;
 148         placement->placement = places;
 149
 150         placement->num_busy_placement = c;
 151         placement->busy_placement = places;
 152 }
 153
 154 void amdgpu_ttm_placement_from_domain(struct amdgpu_bo *abo, u32 domain)
 155 {
 156         struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev);
 157
 158         amdgpu_ttm_placement_init(adev, &abo->placement, abo->placements,
 159                                   domain, abo->flags);
 160 }
 161
 162 static void amdgpu_fill_placement_to_bo(struct amdgpu_bo *bo,
 163                                         struct ttm_placement *placement)
 164 {
 165         BUG_ON(placement->num_placement > (AMDGPU_GEM_DOMAIN_MAX + 1));
 166
 167         memcpy(bo->placements, placement->placement,
 168                placement->num_placement * sizeof(struct ttm_place));
 169         bo->placement.num_placement = placement->num_placement;
 170         bo->placement.num_busy_placement = placement->num_busy_placement;
 171         bo->placement.placement = bo->placements;
 172         bo->placement.busy_placement = bo->placements;
 173 }
 174
 175 /**
 176  * amdgpu_bo_create_reserved - create reserved BO for kernel use
 177  *
 178  * @adev: amdgpu device object
 179  * @size: size for the new BO
 180  * @align: alignment for the new BO
 181  * @domain: where to place it
 182  * @bo_ptr: resulting BO
 183  * @gpu_addr: GPU addr of the pinned BO
 184  * @cpu_addr: optional CPU address mapping
 185  *
 186  * Allocates and pins a BO for kernel internal use, and returns it still
 187  * reserved.
 188  *
 189  * Returns 0 on success, negative error code otherwise.
 190  */
 191 int amdgpu_bo_create_reserved(struct amdgpu_device *adev,
 192                               unsigned long size, int align,
 193                               u32 domain, struct amdgpu_bo **bo_ptr,
 194                               u64 *gpu_addr, void **cpu_addr)
 195 {
 196         bool free = false;
 197         int r;
 198
 199         if (!*bo_ptr) {
 200                 r = amdgpu_bo_create(adev, size, align, true, domain,
 201                                      AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
 202                                      AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS,
 203                                      NULL, NULL, 0, bo_ptr);
 204                 if (r) {
 205                         dev_err(adev->dev, "(%d) failed to allocate kernel bo\n",
 206                                 r);
 207                         return r;
 208                 }
 209                 free = true;
 210         }
 211
 212         r = amdgpu_bo_reserve(*bo_ptr, false);
 213         if (r) {
 214                 dev_err(adev->dev, "(%d) failed to reserve kernel bo\n", r);
 215                 goto error_free;
 216         }
 217
 218         r = amdgpu_bo_pin(*bo_ptr, domain, gpu_addr);
 219         if (r) {
 220                 dev_err(adev->dev, "(%d) kernel bo pin failed\n", r);
 221                 goto error_unreserve;
 222         }
 223
 224         if (cpu_addr) {
 225                 r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
 226                 if (r) {
 227                         dev_err(adev->dev, "(%d) kernel bo map failed\n", r);
 228                         goto error_unreserve;
 229                 }
 230         }
 231
 232         return 0;
 233
 234 error_unreserve:
 235         amdgpu_bo_unreserve(*bo_ptr);
 236
 237 error_free:
 238         if (free)
 239                 amdgpu_bo_unref(bo_ptr);
 240
 241         return r;
 242 }
 243
 244 /**
 245  * amdgpu_bo_create_kernel - create BO for kernel use
 246  *
 247  * @adev: amdgpu device object
 248  * @size: size for the new BO
 249  * @align: alignment for the new BO
 250  * @domain: where to place it
 251  * @bo_ptr: resulting BO
 252  * @gpu_addr: GPU addr of the pinned BO
 253  * @cpu_addr: optional CPU address mapping
 254  *
 255  * Allocates and pins a BO for kernel internal use.
 256  *
 257  * Returns 0 on success, negative error code otherwise.
 258  */
 259 int amdgpu_bo_create_kernel(struct amdgpu_device *adev,
 260                             unsigned long size, int align,
 261                             u32 domain, struct amdgpu_bo **bo_ptr,
 262                             u64 *gpu_addr, void **cpu_addr)
 263 {
 264         int r;
 265
 266         r = amdgpu_bo_create_reserved(adev, size, align, domain, bo_ptr,
 267                                       gpu_addr, cpu_addr);
 268
 269         if (r)
 270                 return r;
 271
 272         amdgpu_bo_unreserve(*bo_ptr);
 273
 274         return 0;
 275 }
 276
 277 /**
 278  * amdgpu_bo_free_kernel - free BO for kernel use
 279  *
 280  * @bo: amdgpu BO to free
 281  *
 282  * unmaps and unpin a BO for kernel internal use.
 283  */
 284 void amdgpu_bo_free_kernel(struct amdgpu_bo **bo, u64 *gpu_addr,
 285                            void **cpu_addr)
 286 {
 287         if (*bo == NULL)
 288                 return;
 289
 290         if (likely(amdgpu_bo_reserve(*bo, true) == 0)) {
 291                 if (cpu_addr)
 292                         amdgpu_bo_kunmap(*bo);
 293
 294                 amdgpu_bo_unpin(*bo);
 295                 amdgpu_bo_unreserve(*bo);
 296         }
 297         amdgpu_bo_unref(bo);
 298
 299         if (gpu_addr)
 300                 *gpu_addr = 0;
 301
 302         if (cpu_addr)
 303                 *cpu_addr = NULL;
 304 }
 305
 306 int amdgpu_bo_create_restricted(struct amdgpu_device *adev,
 307                                 unsigned long size, int byte_align,
 308                                 bool kernel, u32 domain, u64 flags,
 309                                 struct sg_table *sg,
 310                                 struct ttm_placement *placement,
 311                                 struct reservation_object *resv,
 312                                 uint64_t init_value,
 313                                 struct amdgpu_bo **bo_ptr)
 314 {
 315         struct amdgpu_bo *bo;
 316         enum ttm_bo_type type;
 317         unsigned long page_align;
 318         u64 initial_bytes_moved, bytes_moved;
 319         size_t acc_size;
 320         int r;
 321
 322         page_align = roundup(byte_align, PAGE_SIZE) >> PAGE_SHIFT;
 323         size = ALIGN(size, PAGE_SIZE);
 324
 325         if (kernel) {
 326                 type = ttm_bo_type_kernel;
 327         } else if (sg) {
 328                 type = ttm_bo_type_sg;
 329         } else {
 330                 type = ttm_bo_type_device;
 331         }
 332         *bo_ptr = NULL;
 333
 334         acc_size = ttm_bo_dma_acc_size(&adev->mman.bdev, size,
 335                                        sizeof(struct amdgpu_bo));
 336
 337         bo = kzalloc(sizeof(struct amdgpu_bo), GFP_KERNEL);
 338         if (bo == NULL)
 339                 return -ENOMEM;
 340         r = drm_gem_object_init(adev->ddev, &bo->gem_base, size);
 341         if (unlikely(r)) {
 342                 kfree(bo);
 343                 return r;
 344         }
 345         INIT_LIST_HEAD(&bo->shadow_list);
 346         INIT_LIST_HEAD(&bo->va);
 347         bo->preferred_domains = domain & (AMDGPU_GEM_DOMAIN_VRAM |
 348                                          AMDGPU_GEM_DOMAIN_GTT |
 349                                          AMDGPU_GEM_DOMAIN_CPU |
 350                                          AMDGPU_GEM_DOMAIN_GDS |
 351                                          AMDGPU_GEM_DOMAIN_GWS |
 352                                          AMDGPU_GEM_DOMAIN_OA);
 353         bo->allowed_domains = bo->preferred_domains;
 354         if (!kernel && bo->allowed_domains == AMDGPU_GEM_DOMAIN_VRAM)
 355                 bo->allowed_domains |= AMDGPU_GEM_DOMAIN_GTT;
 356
 357         bo->flags = flags;
 358
 359 #ifdef CONFIG_X86_32
 360         /* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
 361          * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
 362          */
 363         bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
 364 #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
 365         /* Don't try to enable write-combining when it can't work, or things
 366          * may be slow
 367          * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
 368          */
 369
 370 #ifndef CONFIG_COMPILE_TEST
 371 #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
 372          thanks to write-combining
 373 #endif
 374
 375         if (bo->flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
 376                 DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
 377                               "better performance thanks to write-combining\n");
 378         bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
 379 #else
 380         /* For architectures that don't support WC memory,
 381          * mask out the WC flag from the BO
 382          */
 383         if (!drm_arch_can_wc_memory())
 384                 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
 385 #endif
 386
 387         amdgpu_fill_placement_to_bo(bo, placement);
 388         /* Kernel allocation are uninterruptible */
 389
 390         initial_bytes_moved = atomic64_read(&adev->num_bytes_moved);
 391         r = ttm_bo_init_reserved(&adev->mman.bdev, &bo->tbo, size, type,
 392                                  &bo->placement, page_align, !kernel, NULL,
 393                                  acc_size, sg, resv, &amdgpu_ttm_bo_destroy);
 394         bytes_moved = atomic64_read(&adev->num_bytes_moved) -
 395                       initial_bytes_moved;
 396         if (adev->mc.visible_vram_size < adev->mc.real_vram_size &&
 397             bo->tbo.mem.mem_type == TTM_PL_VRAM &&
 398             bo->tbo.mem.start < adev->mc.visible_vram_size >> PAGE_SHIFT)
 399                 amdgpu_cs_report_moved_bytes(adev, bytes_moved, bytes_moved);
 400         else
 401                 amdgpu_cs_report_moved_bytes(adev, bytes_moved, 0);
 402
 403         if (unlikely(r != 0))
 404                 return r;
 405
 406         if (kernel)
 407                 bo->tbo.priority = 1;
 408
 409         if (flags & AMDGPU_GEM_CREATE_VRAM_CLEARED &&
 410             bo->tbo.mem.placement & TTM_PL_FLAG_VRAM) {
 411                 struct dma_fence *fence;
 412
 413                 r = amdgpu_fill_buffer(bo, init_value, bo->tbo.resv, &fence);
 414                 if (unlikely(r))
 415                         goto fail_unreserve;
 416
 417                 amdgpu_bo_fence(bo, fence, false);
 418                 dma_fence_put(bo->tbo.moving);
 419                 bo->tbo.moving = dma_fence_get(fence);
 420                 dma_fence_put(fence);
 421         }
 422         if (!resv)
 423                 amdgpu_bo_unreserve(bo);
 424         *bo_ptr = bo;
 425
 426         trace_amdgpu_bo_create(bo);
 427
 428         /* Treat CPU_ACCESS_REQUIRED only as a hint if given by UMD */
 429         if (type == ttm_bo_type_device)
 430                 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
 431
 432         return 0;
 433
 434 fail_unreserve:
 435         if (!resv)
 436                 ww_mutex_unlock(&bo->tbo.resv->lock);
 437         amdgpu_bo_unref(&bo);
 438         return r;
 439 }
 440
 441 static int amdgpu_bo_create_shadow(struct amdgpu_device *adev,
 442                                    unsigned long size, int byte_align,
 443                                    struct amdgpu_bo *bo)
 444 {
 445         struct ttm_placement placement = {0};
 446         struct ttm_place placements[AMDGPU_GEM_DOMAIN_MAX + 1];
 447         int r;
 448
 449         if (bo->shadow)
 450                 return 0;
 451
 452         memset(&placements, 0, sizeof(placements));
 453         amdgpu_ttm_placement_init(adev, &placement, placements,
 454                                   AMDGPU_GEM_DOMAIN_GTT,
 455                                   AMDGPU_GEM_CREATE_CPU_GTT_USWC |
 456                                   AMDGPU_GEM_CREATE_SHADOW);
 457
 458         r = amdgpu_bo_create_restricted(adev, size, byte_align, true,
 459                                         AMDGPU_GEM_DOMAIN_GTT,
 460                                         AMDGPU_GEM_CREATE_CPU_GTT_USWC |
 461                                         AMDGPU_GEM_CREATE_SHADOW,
 462                                         NULL, &placement,
 463                                         bo->tbo.resv,
 464                                         0,
 465                                         &bo->shadow);
 466         if (!r) {
 467                 bo->shadow->parent = amdgpu_bo_ref(bo);
 468                 mutex_lock(&adev->shadow_list_lock);
 469                 list_add_tail(&bo->shadow_list, &adev->shadow_list);
 470                 mutex_unlock(&adev->shadow_list_lock);
 471         }
 472
 473         return r;
 474 }
 475
 476 /* init_value will only take effect when flags contains
 477  * AMDGPU_GEM_CREATE_VRAM_CLEARED.
 478  */
 479 int amdgpu_bo_create(struct amdgpu_device *adev,
 480                      unsigned long size, int byte_align,
 481                      bool kernel, u32 domain, u64 flags,
 482                      struct sg_table *sg,
 483                      struct reservation_object *resv,
 484                      uint64_t init_value,
 485                      struct amdgpu_bo **bo_ptr)
 486 {
 487         struct ttm_placement placement = {0};
 488         struct ttm_place placements[AMDGPU_GEM_DOMAIN_MAX + 1];
 489         uint64_t parent_flags = flags & ~AMDGPU_GEM_CREATE_SHADOW;
 490         int r;
 491
 492         memset(&placements, 0, sizeof(placements));
 493         amdgpu_ttm_placement_init(adev, &placement, placements,
 494                                   domain, parent_flags);
 495
 496         r = amdgpu_bo_create_restricted(adev, size, byte_align, kernel, domain,
 497                                         parent_flags, sg, &placement, resv,
 498                                         init_value, bo_ptr);
 499         if (r)
 500                 return r;
 501
 502         if ((flags & AMDGPU_GEM_CREATE_SHADOW) && amdgpu_need_backup(adev)) {
 503                 if (!resv)
 504                         WARN_ON(reservation_object_lock((*bo_ptr)->tbo.resv,
 505                                                         NULL));
 506
 507                 r = amdgpu_bo_create_shadow(adev, size, byte_align, (*bo_ptr));
 508
 509                 if (!resv)
 510                         reservation_object_unlock((*bo_ptr)->tbo.resv);
 511
 512                 if (r)
 513                         amdgpu_bo_unref(bo_ptr);
 514         }
 515
 516         return r;
 517 }
 518
 519 int amdgpu_bo_backup_to_shadow(struct amdgpu_device *adev,
 520                                struct amdgpu_ring *ring,
 521                                struct amdgpu_bo *bo,
 522                                struct reservation_object *resv,
 523                                struct dma_fence **fence,
 524                                bool direct)
 525
 526 {
 527         struct amdgpu_bo *shadow = bo->shadow;
 528         uint64_t bo_addr, shadow_addr;
 529         int r;
 530
 531         if (!shadow)
 532                 return -EINVAL;
 533
 534         bo_addr = amdgpu_bo_gpu_offset(bo);
 535         shadow_addr = amdgpu_bo_gpu_offset(bo->shadow);
 536
 537         r = reservation_object_reserve_shared(bo->tbo.resv);
 538         if (r)
 539                 goto err;
 540
 541         r = amdgpu_copy_buffer(ring, bo_addr, shadow_addr,
 542                                amdgpu_bo_size(bo), resv, fence,
 543                                direct, false);
 544         if (!r)
 545                 amdgpu_bo_fence(bo, *fence, true);
 546
 547 err:
 548         return r;
 549 }
 550
 551 int amdgpu_bo_validate(struct amdgpu_bo *bo)
 552 {
 553         uint32_t domain;
 554         int r;
 555
 556         if (bo->pin_count)
 557                 return 0;
 558
 559         domain = bo->preferred_domains;
 560
 561 retry:
 562         amdgpu_ttm_placement_from_domain(bo, domain);
 563         r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
 564         if (unlikely(r == -ENOMEM) && domain != bo->allowed_domains) {
 565                 domain = bo->allowed_domains;
 566                 goto retry;
 567         }
 568
 569         return r;
 570 }
 571
 572 int amdgpu_bo_restore_from_shadow(struct amdgpu_device *adev,
 573                                   struct amdgpu_ring *ring,
 574                                   struct amdgpu_bo *bo,
 575                                   struct reservation_object *resv,
 576                                   struct dma_fence **fence,
 577                                   bool direct)
 578
 579 {
 580         struct amdgpu_bo *shadow = bo->shadow;
 581         uint64_t bo_addr, shadow_addr;
 582         int r;
 583
 584         if (!shadow)
 585                 return -EINVAL;
 586
 587         bo_addr = amdgpu_bo_gpu_offset(bo);
 588         shadow_addr = amdgpu_bo_gpu_offset(bo->shadow);
 589
 590         r = reservation_object_reserve_shared(bo->tbo.resv);
 591         if (r)
 592                 goto err;
 593
 594         r = amdgpu_copy_buffer(ring, shadow_addr, bo_addr,
 595                                amdgpu_bo_size(bo), resv, fence,
 596                                direct, false);
 597         if (!r)
 598                 amdgpu_bo_fence(bo, *fence, true);
 599
 600 err:
 601         return r;
 602 }
 603
 604 int amdgpu_bo_kmap(struct amdgpu_bo *bo, void **ptr)
 605 {
 606         void *kptr;
 607         long r;
 608
 609         if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
 610                 return -EPERM;
 611
 612         kptr = amdgpu_bo_kptr(bo);
 613         if (kptr) {
 614                 if (ptr)
 615                         *ptr = kptr;
 616                 return 0;
 617         }
 618
 619         r = reservation_object_wait_timeout_rcu(bo->tbo.resv, false, false,
 620                                                 MAX_SCHEDULE_TIMEOUT);
 621         if (r < 0)
 622                 return r;
 623
 624         r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.num_pages, &bo->kmap);
 625         if (r)
 626                 return r;
 627
 628         if (ptr)
 629                 *ptr = amdgpu_bo_kptr(bo);
 630
 631         return 0;
 632 }
 633
 634 void *amdgpu_bo_kptr(struct amdgpu_bo *bo)
 635 {
 636         bool is_iomem;
 637
 638         return ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
 639 }
 640
 641 void amdgpu_bo_kunmap(struct amdgpu_bo *bo)
 642 {
 643         if (bo->kmap.bo)
 644                 ttm_bo_kunmap(&bo->kmap);
 645 }
 646
 647 struct amdgpu_bo *amdgpu_bo_ref(struct amdgpu_bo *bo)
 648 {
 649         if (bo == NULL)
 650                 return NULL;
 651
 652         ttm_bo_reference(&bo->tbo);
 653         return bo;
 654 }
 655
 656 void amdgpu_bo_unref(struct amdgpu_bo **bo)
 657 {
 658         struct ttm_buffer_object *tbo;
 659
 660         if ((*bo) == NULL)
 661                 return;
 662
 663         tbo = &((*bo)->tbo);
 664         ttm_bo_unref(&tbo);
 665         if (tbo == NULL)
 666                 *bo = NULL;
 667 }
 668
 669 int amdgpu_bo_pin_restricted(struct amdgpu_bo *bo, u32 domain,
 670                              u64 min_offset, u64 max_offset,
 671                              u64 *gpu_addr)
 672 {
 673         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
 674         int r, i;
 675         unsigned fpfn, lpfn;
 676
 677         if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm))
 678                 return -EPERM;
 679
 680         if (WARN_ON_ONCE(min_offset > max_offset))
 681                 return -EINVAL;
 682
 683         /* A shared bo cannot be migrated to VRAM */
 684         if (bo->prime_shared_count && (domain == AMDGPU_GEM_DOMAIN_VRAM))
 685                 return -EINVAL;
 686
 687         if (bo->pin_count) {
 688                 uint32_t mem_type = bo->tbo.mem.mem_type;
 689
 690                 if (domain != amdgpu_mem_type_to_domain(mem_type))
 691                         return -EINVAL;
 692
 693                 bo->pin_count++;
 694                 if (gpu_addr)
 695                         *gpu_addr = amdgpu_bo_gpu_offset(bo);
 696
 697                 if (max_offset != 0) {
 698                         u64 domain_start = bo->tbo.bdev->man[mem_type].gpu_offset;
 699                         WARN_ON_ONCE(max_offset <
 700                                      (amdgpu_bo_gpu_offset(bo) - domain_start));
 701                 }
 702
 703                 return 0;
 704         }
 705
 706         bo->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
 707         amdgpu_ttm_placement_from_domain(bo, domain);
 708         for (i = 0; i < bo->placement.num_placement; i++) {
 709                 /* force to pin into visible video ram */
 710                 if ((bo->placements[i].flags & TTM_PL_FLAG_VRAM) &&
 711                     !(bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS) &&
 712                     (!max_offset || max_offset >
 713                      adev->mc.visible_vram_size)) {
 714                         if (WARN_ON_ONCE(min_offset >
 715                                          adev->mc.visible_vram_size))
 716                                 return -EINVAL;
 717                         fpfn = min_offset >> PAGE_SHIFT;
 718                         lpfn = adev->mc.visible_vram_size >> PAGE_SHIFT;
 719                 } else {
 720                         fpfn = min_offset >> PAGE_SHIFT;
 721                         lpfn = max_offset >> PAGE_SHIFT;
 722                 }
 723                 if (fpfn > bo->placements[i].fpfn)
 724                         bo->placements[i].fpfn = fpfn;
 725                 if (!bo->placements[i].lpfn ||
 726                     (lpfn && lpfn < bo->placements[i].lpfn))
 727                         bo->placements[i].lpfn = lpfn;
 728                 bo->placements[i].flags |= TTM_PL_FLAG_NO_EVICT;
 729         }
 730
 731         r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
 732         if (unlikely(r)) {
 733                 dev_err(adev->dev, "%p pin failed\n", bo);
 734                 goto error;
 735         }
 736
 737         bo->pin_count = 1;
 738         if (gpu_addr != NULL) {
 739                 r = amdgpu_ttm_bind(&bo->tbo, &bo->tbo.mem);
 740                 if (unlikely(r)) {
 741                         dev_err(adev->dev, "%p bind failed\n", bo);
 742                         goto error;
 743                 }
 744                 *gpu_addr = amdgpu_bo_gpu_offset(bo);
 745         }
 746         if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
 747                 adev->vram_pin_size += amdgpu_bo_size(bo);
 748                 if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
 749                         adev->invisible_pin_size += amdgpu_bo_size(bo);
 750         } else if (domain == AMDGPU_GEM_DOMAIN_GTT) {
 751                 adev->gart_pin_size += amdgpu_bo_size(bo);
 752         }
 753
 754 error:
 755         return r;
 756 }
 757
 758 int amdgpu_bo_pin(struct amdgpu_bo *bo, u32 domain, u64 *gpu_addr)
 759 {
 760         return amdgpu_bo_pin_restricted(bo, domain, 0, 0, gpu_addr);
 761 }
 762
 763 int amdgpu_bo_unpin(struct amdgpu_bo *bo)
 764 {
 765         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
 766         int r, i;
 767
 768         if (!bo->pin_count) {
 769                 dev_warn(adev->dev, "%p unpin not necessary\n", bo);
 770                 return 0;
 771         }
 772         bo->pin_count--;
 773         if (bo->pin_count)
 774                 return 0;
 775         for (i = 0; i < bo->placement.num_placement; i++) {
 776                 bo->placements[i].lpfn = 0;
 777                 bo->placements[i].flags &= ~TTM_PL_FLAG_NO_EVICT;
 778         }
 779         r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
 780         if (unlikely(r)) {
 781                 dev_err(adev->dev, "%p validate failed for unpin\n", bo);
 782                 goto error;
 783         }
 784
 785         if (bo->tbo.mem.mem_type == TTM_PL_VRAM) {
 786                 adev->vram_pin_size -= amdgpu_bo_size(bo);
 787                 if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
 788                         adev->invisible_pin_size -= amdgpu_bo_size(bo);
 789         } else if (bo->tbo.mem.mem_type == TTM_PL_TT) {
 790                 adev->gart_pin_size -= amdgpu_bo_size(bo);
 791         }
 792
 793 error:
 794         return r;
 795 }
 796
 797 int amdgpu_bo_evict_vram(struct amdgpu_device *adev)
 798 {
 799         /* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */
 800         if (0 && (adev->flags & AMD_IS_APU)) {
 801                 /* Useless to evict on IGP chips */
 802                 return 0;
 803         }
 804         return ttm_bo_evict_mm(&adev->mman.bdev, TTM_PL_VRAM);
 805 }
 806
 807 static const char *amdgpu_vram_names[] = {
 808         "UNKNOWN",
 809         "GDDR1",
 810         "DDR2",
 811         "GDDR3",
 812         "GDDR4",
 813         "GDDR5",
 814         "HBM",
 815         "DDR3"
 816 };
 817
 818 int amdgpu_bo_init(struct amdgpu_device *adev)
 819 {
 820         /* reserve PAT memory space to WC for VRAM */
 821         arch_io_reserve_memtype_wc(adev->mc.aper_base,
 822                                    adev->mc.aper_size);
 823
 824         /* Add an MTRR for the VRAM */
 825         adev->mc.vram_mtrr = arch_phys_wc_add(adev->mc.aper_base,
 826                                               adev->mc.aper_size);
 827         DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
 828                 adev->mc.mc_vram_size >> 20,
 829                 (unsigned long long)adev->mc.aper_size >> 20);
 830         DRM_INFO("RAM width %dbits %s\n",
 831                  adev->mc.vram_width, amdgpu_vram_names[adev->mc.vram_type]);
 832         return amdgpu_ttm_init(adev);
 833 }
 834
 835 void amdgpu_bo_fini(struct amdgpu_device *adev)
 836 {
 837         amdgpu_ttm_fini(adev);
 838         arch_phys_wc_del(adev->mc.vram_mtrr);
 839         arch_io_free_memtype_wc(adev->mc.aper_base, adev->mc.aper_size);
 840 }
 841
 842 int amdgpu_bo_fbdev_mmap(struct amdgpu_bo *bo,
 843                              struct vm_area_struct *vma)
 844 {
 845         return ttm_fbdev_mmap(vma, &bo->tbo);
 846 }
 847
 848 int amdgpu_bo_set_tiling_flags(struct amdgpu_bo *bo, u64 tiling_flags)
 849 {
 850         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
 851
 852         if (adev->family <= AMDGPU_FAMILY_CZ &&
 853             AMDGPU_TILING_GET(tiling_flags, TILE_SPLIT) > 6)
 854                 return -EINVAL;
 855
 856         bo->tiling_flags = tiling_flags;
 857         return 0;
 858 }
 859
 860 void amdgpu_bo_get_tiling_flags(struct amdgpu_bo *bo, u64 *tiling_flags)
 861 {
 862         lockdep_assert_held(&bo->tbo.resv->lock.base);
 863
 864         if (tiling_flags)
 865                 *tiling_flags = bo->tiling_flags;
 866 }
 867
 868 int amdgpu_bo_set_metadata (struct amdgpu_bo *bo, void *metadata,
 869                             uint32_t metadata_size, uint64_t flags)
 870 {
 871         void *buffer;
 872
 873         if (!metadata_size) {
 874                 if (bo->metadata_size) {
 875                         kfree(bo->metadata);
 876                         bo->metadata = NULL;
 877                         bo->metadata_size = 0;
 878                 }
 879                 return 0;
 880         }
 881
 882         if (metadata == NULL)
 883                 return -EINVAL;
 884
 885         buffer = kmemdup(metadata, metadata_size, GFP_KERNEL);
 886         if (buffer == NULL)
 887                 return -ENOMEM;
 888
 889         kfree(bo->metadata);
 890         bo->metadata_flags = flags;
 891         bo->metadata = buffer;
 892         bo->metadata_size = metadata_size;
 893
 894         return 0;
 895 }
 896
 897 int amdgpu_bo_get_metadata(struct amdgpu_bo *bo, void *buffer,
 898                            size_t buffer_size, uint32_t *metadata_size,
 899                            uint64_t *flags)
 900 {
 901         if (!buffer && !metadata_size)
 902                 return -EINVAL;
 903
 904         if (buffer) {
 905                 if (buffer_size < bo->metadata_size)
 906                         return -EINVAL;
 907
 908                 if (bo->metadata_size)
 909                         memcpy(buffer, bo->metadata, bo->metadata_size);
 910         }
 911
 912         if (metadata_size)
 913                 *metadata_size = bo->metadata_size;
 914         if (flags)
 915                 *flags = bo->metadata_flags;
 916
 917         return 0;
 918 }
 919
 920 void amdgpu_bo_move_notify(struct ttm_buffer_object *bo,
 921                            bool evict,
 922                            struct ttm_mem_reg *new_mem)
 923 {
 924         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
 925         struct amdgpu_bo *abo;
 926         struct ttm_mem_reg *old_mem = &bo->mem;
 927
 928         if (!amdgpu_ttm_bo_is_amdgpu_bo(bo))
 929                 return;
 930
 931         abo = container_of(bo, struct amdgpu_bo, tbo);
 932         amdgpu_vm_bo_invalidate(adev, abo);
 933
 934         amdgpu_bo_kunmap(abo);
 935
 936         /* remember the eviction */
 937         if (evict)
 938                 atomic64_inc(&adev->num_evictions);
 939
 940         /* update statistics */
 941         if (!new_mem)
 942                 return;
 943
 944         /* move_notify is called before move happens */
 945         trace_amdgpu_ttm_bo_move(abo, new_mem->mem_type, old_mem->mem_type);
 946 }
 947
 948 int amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
 949 {
 950         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
 951         struct amdgpu_bo *abo;
 952         unsigned long offset, size;
 953         int r;
 954
 955         if (!amdgpu_ttm_bo_is_amdgpu_bo(bo))
 956                 return 0;
 957
 958         abo = container_of(bo, struct amdgpu_bo, tbo);
 959
 960         /* Remember that this BO was accessed by the CPU */
 961         abo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
 962
 963         if (bo->mem.mem_type != TTM_PL_VRAM)
 964                 return 0;
 965
 966         size = bo->mem.num_pages << PAGE_SHIFT;
 967         offset = bo->mem.start << PAGE_SHIFT;
 968         if ((offset + size) <= adev->mc.visible_vram_size)
 969                 return 0;
 970
 971         /* Can't move a pinned BO to visible VRAM */
 972         if (abo->pin_count > 0)
 973                 return -EINVAL;
 974
 975         /* hurrah the memory is not visible ! */
 976         atomic64_inc(&adev->num_vram_cpu_page_faults);
 977         amdgpu_ttm_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM |
 978                                          AMDGPU_GEM_DOMAIN_GTT);
 979
 980         /* Avoid costly evictions; only set GTT as a busy placement */
 981         abo->placement.num_busy_placement = 1;
 982         abo->placement.busy_placement = &abo->placements[1];
 983
 984         r = ttm_bo_validate(bo, &abo->placement, false, false);
 985         if (unlikely(r != 0))
 986                 return r;
 987
 988         offset = bo->mem.start << PAGE_SHIFT;
 989         /* this should never happen */
 990         if (bo->mem.mem_type == TTM_PL_VRAM &&
 991             (offset + size) > adev->mc.visible_vram_size)
 992                 return -EINVAL;
 993
 994         return 0;
 995 }
 996
 997 /**
 998  * amdgpu_bo_fence - add fence to buffer object
 999  *
1000  * @bo: buffer object in question
1001  * @fence: fence to add
1002  * @shared: true if fence should be added shared
1003  *
1004  */
1005 void amdgpu_bo_fence(struct amdgpu_bo *bo, struct dma_fence *fence,
1006                      bool shared)
1007 {
1008         struct reservation_object *resv = bo->tbo.resv;
1009
1010         if (shared)
1011                 reservation_object_add_shared_fence(resv, fence);
1012         else
1013                 reservation_object_add_excl_fence(resv, fence);
1014 }
1015
1016 /**
1017  * amdgpu_bo_gpu_offset - return GPU offset of bo
1018  * @bo: amdgpu object for which we query the offset
1019  *
1020  * Returns current GPU offset of the object.
1021  *
1022  * Note: object should either be pinned or reserved when calling this
1023  * function, it might be useful to add check for this for debugging.
1024  */
1025 u64 amdgpu_bo_gpu_offset(struct amdgpu_bo *bo)
1026 {
1027         WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_SYSTEM);
1028         WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_TT &&
1029                      !amdgpu_ttm_is_bound(bo->tbo.ttm));
1030         WARN_ON_ONCE(!ww_mutex_is_locked(&bo->tbo.resv->lock) &&
1031                      !bo->pin_count);
1032         WARN_ON_ONCE(bo->tbo.mem.start == AMDGPU_BO_INVALID_OFFSET);
1033         WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_VRAM &&
1034                      !(bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS));
1035
1036         return bo->tbo.offset;
1037 }