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

   1 /*
   2  * Copyright 2009 Jerome Glisse.
   3  * All Rights Reserved.
   4  *
   5  * Permission is hereby granted, free of charge, to any person obtaining a
   6  * copy of this software and associated documentation files (the
   7  * "Software"), to deal in the Software without restriction, including
   8  * without limitation the rights to use, copy, modify, merge, publish,
   9  * distribute, sub license, and/or sell copies of the Software, and to
  10  * permit persons to whom the Software is furnished to do so, subject to
  11  * the following conditions:
  12  *
  13  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  14  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  15  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  16  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
  17  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  18  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  19  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  20  *
  21  * The above copyright notice and this permission notice (including the
  22  * next paragraph) shall be included in all copies or substantial portions
  23  * of the Software.
  24  *
  25  */
  26 /*
  27  * Authors:
  28  *    Jerome Glisse <[email protected]>
  29  *    Thomas Hellstrom <thomas-at-tungstengraphics-dot-com>
  30  *    Dave Airlie
  31  */
  32 #include <drm/ttm/ttm_bo_api.h>
  33 #include <drm/ttm/ttm_bo_driver.h>
  34 #include <drm/ttm/ttm_placement.h>
  35 #include <drm/ttm/ttm_module.h>
  36 #include <drm/ttm/ttm_page_alloc.h>
  37 #include <drm/drmP.h>
  38 #include <drm/amdgpu_drm.h>
  39 #include <linux/seq_file.h>
  40 #include <linux/slab.h>
  41 #include <linux/swiotlb.h>
  42 #include <linux/swap.h>
  43 #include <linux/pagemap.h>
  44 #include <linux/debugfs.h>
  45 #include <linux/iommu.h>
  46 #include "amdgpu.h"
  47 #include "amdgpu_object.h"
  48 #include "amdgpu_trace.h"
  49 #include "amdgpu_amdkfd.h"
  50 #include "bif/bif_4_1_d.h"
  51
  52 #define DRM_FILE_PAGE_OFFSET (0x100000000ULL >> PAGE_SHIFT)
  53
  54 static int amdgpu_map_buffer(struct ttm_buffer_object *bo,
  55                              struct ttm_mem_reg *mem, unsigned num_pages,
  56                              uint64_t offset, unsigned window,
  57                              struct amdgpu_ring *ring,
  58                              uint64_t *addr);
  59
  60 static int amdgpu_ttm_debugfs_init(struct amdgpu_device *adev);
  61 static void amdgpu_ttm_debugfs_fini(struct amdgpu_device *adev);
  62
  63 /*
  64  * Global memory.
  65  */
  66 static int amdgpu_ttm_mem_global_init(struct drm_global_reference *ref)
  67 {
  68         return ttm_mem_global_init(ref->object);
  69 }
  70
  71 static void amdgpu_ttm_mem_global_release(struct drm_global_reference *ref)
  72 {
  73         ttm_mem_global_release(ref->object);
  74 }
  75
  76 static int amdgpu_ttm_global_init(struct amdgpu_device *adev)
  77 {
  78         struct drm_global_reference *global_ref;
  79         struct amdgpu_ring *ring;
  80         struct drm_sched_rq *rq;
  81         int r;
  82
  83         adev->mman.mem_global_referenced = false;
  84         global_ref = &adev->mman.mem_global_ref;
  85         global_ref->global_type = DRM_GLOBAL_TTM_MEM;
  86         global_ref->size = sizeof(struct ttm_mem_global);
  87         global_ref->init = &amdgpu_ttm_mem_global_init;
  88         global_ref->release = &amdgpu_ttm_mem_global_release;
  89         r = drm_global_item_ref(global_ref);
  90         if (r) {
  91                 DRM_ERROR("Failed setting up TTM memory accounting "
  92                           "subsystem.\n");
  93                 goto error_mem;
  94         }
  95
  96         adev->mman.bo_global_ref.mem_glob =
  97                 adev->mman.mem_global_ref.object;
  98         global_ref = &adev->mman.bo_global_ref.ref;
  99         global_ref->global_type = DRM_GLOBAL_TTM_BO;
 100         global_ref->size = sizeof(struct ttm_bo_global);
 101         global_ref->init = &ttm_bo_global_init;
 102         global_ref->release = &ttm_bo_global_release;
 103         r = drm_global_item_ref(global_ref);
 104         if (r) {
 105                 DRM_ERROR("Failed setting up TTM BO subsystem.\n");
 106                 goto error_bo;
 107         }
 108
 109         mutex_init(&adev->mman.gtt_window_lock);
 110
 111         ring = adev->mman.buffer_funcs_ring;
 112         rq = &ring->sched.sched_rq[DRM_SCHED_PRIORITY_KERNEL];
 113         r = drm_sched_entity_init(&ring->sched, &adev->mman.entity,
 114                                   rq, amdgpu_sched_jobs, NULL);
 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                 drm_sched_entity_fini(adev->mman.entity.sched,
 136                                       &adev->mman.entity);
 137                 mutex_destroy(&adev->mman.gtt_window_lock);
 138                 drm_global_item_unref(&adev->mman.bo_global_ref.ref);
 139                 drm_global_item_unref(&adev->mman.mem_global_ref);
 140                 adev->mman.mem_global_referenced = false;
 141         }
 142 }
 143
 144 static int amdgpu_invalidate_caches(struct ttm_bo_device *bdev, uint32_t flags)
 145 {
 146         return 0;
 147 }
 148
 149 static int amdgpu_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
 150                                 struct ttm_mem_type_manager *man)
 151 {
 152         struct amdgpu_device *adev;
 153
 154         adev = amdgpu_ttm_adev(bdev);
 155
 156         switch (type) {
 157         case TTM_PL_SYSTEM:
 158                 /* System memory */
 159                 man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
 160                 man->available_caching = TTM_PL_MASK_CACHING;
 161                 man->default_caching = TTM_PL_FLAG_CACHED;
 162                 break;
 163         case TTM_PL_TT:
 164                 man->func = &amdgpu_gtt_mgr_func;
 165                 man->gpu_offset = adev->gmc.gart_start;
 166                 man->available_caching = TTM_PL_MASK_CACHING;
 167                 man->default_caching = TTM_PL_FLAG_CACHED;
 168                 man->flags = TTM_MEMTYPE_FLAG_MAPPABLE | TTM_MEMTYPE_FLAG_CMA;
 169                 break;
 170         case TTM_PL_VRAM:
 171                 /* "On-card" video ram */
 172                 man->func = &amdgpu_vram_mgr_func;
 173                 man->gpu_offset = adev->gmc.vram_start;
 174                 man->flags = TTM_MEMTYPE_FLAG_FIXED |
 175                              TTM_MEMTYPE_FLAG_MAPPABLE;
 176                 man->available_caching = TTM_PL_FLAG_UNCACHED | TTM_PL_FLAG_WC;
 177                 man->default_caching = TTM_PL_FLAG_WC;
 178                 break;
 179         case AMDGPU_PL_GDS:
 180         case AMDGPU_PL_GWS:
 181         case AMDGPU_PL_OA:
 182                 /* On-chip GDS memory*/
 183                 man->func = &ttm_bo_manager_func;
 184                 man->gpu_offset = 0;
 185                 man->flags = TTM_MEMTYPE_FLAG_FIXED | TTM_MEMTYPE_FLAG_CMA;
 186                 man->available_caching = TTM_PL_FLAG_UNCACHED;
 187                 man->default_caching = TTM_PL_FLAG_UNCACHED;
 188                 break;
 189         default:
 190                 DRM_ERROR("Unsupported memory type %u\n", (unsigned)type);
 191                 return -EINVAL;
 192         }
 193         return 0;
 194 }
 195
 196 static void amdgpu_evict_flags(struct ttm_buffer_object *bo,
 197                                 struct ttm_placement *placement)
 198 {
 199         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
 200         struct amdgpu_bo *abo;
 201         static const struct ttm_place placements = {
 202                 .fpfn = 0,
 203                 .lpfn = 0,
 204                 .flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM
 205         };
 206
 207         if (bo->type == ttm_bo_type_sg) {
 208                 placement->num_placement = 0;
 209                 placement->num_busy_placement = 0;
 210                 return;
 211         }
 212
 213         if (!amdgpu_ttm_bo_is_amdgpu_bo(bo)) {
 214                 placement->placement = &placements;
 215                 placement->busy_placement = &placements;
 216                 placement->num_placement = 1;
 217                 placement->num_busy_placement = 1;
 218                 return;
 219         }
 220         abo = ttm_to_amdgpu_bo(bo);
 221         switch (bo->mem.mem_type) {
 222         case TTM_PL_VRAM:
 223                 if (!adev->mman.buffer_funcs_enabled) {
 224                         amdgpu_ttm_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_CPU);
 225                 } else if (adev->gmc.visible_vram_size < adev->gmc.real_vram_size &&
 226                            !(abo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED) &&
 227                            amdgpu_bo_in_cpu_visible_vram(abo)) {
 228
 229                         /* Try evicting to the CPU inaccessible part of VRAM
 230                          * first, but only set GTT as busy placement, so this
 231                          * BO will be evicted to GTT rather than causing other
 232                          * BOs to be evicted from VRAM
 233                          */
 234                         amdgpu_ttm_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM |
 235                                                          AMDGPU_GEM_DOMAIN_GTT);
 236                         abo->placements[0].fpfn = adev->gmc.visible_vram_size >> PAGE_SHIFT;
 237                         abo->placements[0].lpfn = 0;
 238                         abo->placement.busy_placement = &abo->placements[1];
 239                         abo->placement.num_busy_placement = 1;
 240                 } else {
 241                         amdgpu_ttm_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_GTT);
 242                 }
 243                 break;
 244         case TTM_PL_TT:
 245         default:
 246                 amdgpu_ttm_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_CPU);
 247         }
 248         *placement = abo->placement;
 249 }
 250
 251 static int amdgpu_verify_access(struct ttm_buffer_object *bo, struct file *filp)
 252 {
 253         struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
 254
 255         /*
 256          * Don't verify access for KFD BOs. They don't have a GEM
 257          * object associated with them.
 258          */
 259         if (abo->kfd_bo)
 260                 return 0;
 261
 262         if (amdgpu_ttm_tt_get_usermm(bo->ttm))
 263                 return -EPERM;
 264         return drm_vma_node_verify_access(&abo->gem_base.vma_node,
 265                                           filp->private_data);
 266 }
 267
 268 static void amdgpu_move_null(struct ttm_buffer_object *bo,
 269                              struct ttm_mem_reg *new_mem)
 270 {
 271         struct ttm_mem_reg *old_mem = &bo->mem;
 272
 273         BUG_ON(old_mem->mm_node != NULL);
 274         *old_mem = *new_mem;
 275         new_mem->mm_node = NULL;
 276 }
 277
 278 static uint64_t amdgpu_mm_node_addr(struct ttm_buffer_object *bo,
 279                                     struct drm_mm_node *mm_node,
 280                                     struct ttm_mem_reg *mem)
 281 {
 282         uint64_t addr = 0;
 283
 284         if (mem->mem_type != TTM_PL_TT || amdgpu_gtt_mgr_has_gart_addr(mem)) {
 285                 addr = mm_node->start << PAGE_SHIFT;
 286                 addr += bo->bdev->man[mem->mem_type].gpu_offset;
 287         }
 288         return addr;
 289 }
 290
 291 /**
 292  * amdgpu_find_mm_node - Helper function finds the drm_mm_node
 293  *  corresponding to @offset. It also modifies the offset to be
 294  *  within the drm_mm_node returned
 295  */
 296 static struct drm_mm_node *amdgpu_find_mm_node(struct ttm_mem_reg *mem,
 297                                                unsigned long *offset)
 298 {
 299         struct drm_mm_node *mm_node = mem->mm_node;
 300
 301         while (*offset >= (mm_node->size << PAGE_SHIFT)) {
 302                 *offset -= (mm_node->size << PAGE_SHIFT);
 303                 ++mm_node;
 304         }
 305         return mm_node;
 306 }
 307
 308 /**
 309  * amdgpu_copy_ttm_mem_to_mem - Helper function for copy
 310  *
 311  * The function copies @size bytes from {src->mem + src->offset} to
 312  * {dst->mem + dst->offset}. src->bo and dst->bo could be same BO for a
 313  * move and different for a BO to BO copy.
 314  *
 315  * @f: Returns the last fence if multiple jobs are submitted.
 316  */
 317 int amdgpu_ttm_copy_mem_to_mem(struct amdgpu_device *adev,
 318                                struct amdgpu_copy_mem *src,
 319                                struct amdgpu_copy_mem *dst,
 320                                uint64_t size,
 321                                struct reservation_object *resv,
 322                                struct dma_fence **f)
 323 {
 324         struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
 325         struct drm_mm_node *src_mm, *dst_mm;
 326         uint64_t src_node_start, dst_node_start, src_node_size,
 327                  dst_node_size, src_page_offset, dst_page_offset;
 328         struct dma_fence *fence = NULL;
 329         int r = 0;
 330         const uint64_t GTT_MAX_BYTES = (AMDGPU_GTT_MAX_TRANSFER_SIZE *
 331                                         AMDGPU_GPU_PAGE_SIZE);
 332
 333         if (!adev->mman.buffer_funcs_enabled) {
 334                 DRM_ERROR("Trying to move memory with ring turned off.\n");
 335                 return -EINVAL;
 336         }
 337
 338         src_mm = amdgpu_find_mm_node(src->mem, &src->offset);
 339         src_node_start = amdgpu_mm_node_addr(src->bo, src_mm, src->mem) +
 340                                              src->offset;
 341         src_node_size = (src_mm->size << PAGE_SHIFT) - src->offset;
 342         src_page_offset = src_node_start & (PAGE_SIZE - 1);
 343
 344         dst_mm = amdgpu_find_mm_node(dst->mem, &dst->offset);
 345         dst_node_start = amdgpu_mm_node_addr(dst->bo, dst_mm, dst->mem) +
 346                                              dst->offset;
 347         dst_node_size = (dst_mm->size << PAGE_SHIFT) - dst->offset;
 348         dst_page_offset = dst_node_start & (PAGE_SIZE - 1);
 349
 350         mutex_lock(&adev->mman.gtt_window_lock);
 351
 352         while (size) {
 353                 unsigned long cur_size;
 354                 uint64_t from = src_node_start, to = dst_node_start;
 355                 struct dma_fence *next;
 356
 357                 /* Copy size cannot exceed GTT_MAX_BYTES. So if src or dst
 358                  * begins at an offset, then adjust the size accordingly
 359                  */
 360                 cur_size = min3(min(src_node_size, dst_node_size), size,
 361                                 GTT_MAX_BYTES);
 362                 if (cur_size + src_page_offset > GTT_MAX_BYTES ||
 363                     cur_size + dst_page_offset > GTT_MAX_BYTES)
 364                         cur_size -= max(src_page_offset, dst_page_offset);
 365
 366                 /* Map only what needs to be accessed. Map src to window 0 and
 367                  * dst to window 1
 368                  */
 369                 if (src->mem->mem_type == TTM_PL_TT &&
 370                     !amdgpu_gtt_mgr_has_gart_addr(src->mem)) {
 371                         r = amdgpu_map_buffer(src->bo, src->mem,
 372                                         PFN_UP(cur_size + src_page_offset),
 373                                         src_node_start, 0, ring,
 374                                         &from);
 375                         if (r)
 376                                 goto error;
 377                         /* Adjust the offset because amdgpu_map_buffer returns
 378                          * start of mapped page
 379                          */
 380                         from += src_page_offset;
 381                 }
 382
 383                 if (dst->mem->mem_type == TTM_PL_TT &&
 384                     !amdgpu_gtt_mgr_has_gart_addr(dst->mem)) {
 385                         r = amdgpu_map_buffer(dst->bo, dst->mem,
 386                                         PFN_UP(cur_size + dst_page_offset),
 387                                         dst_node_start, 1, ring,
 388                                         &to);
 389                         if (r)
 390                                 goto error;
 391                         to += dst_page_offset;
 392                 }
 393
 394                 r = amdgpu_copy_buffer(ring, from, to, cur_size,
 395                                        resv, &next, false, true);
 396                 if (r)
 397                         goto error;
 398
 399                 dma_fence_put(fence);
 400                 fence = next;
 401
 402                 size -= cur_size;
 403                 if (!size)
 404                         break;
 405
 406                 src_node_size -= cur_size;
 407                 if (!src_node_size) {
 408                         src_node_start = amdgpu_mm_node_addr(src->bo, ++src_mm,
 409                                                              src->mem);
 410                         src_node_size = (src_mm->size << PAGE_SHIFT);
 411                 } else {
 412                         src_node_start += cur_size;
 413                         src_page_offset = src_node_start & (PAGE_SIZE - 1);
 414                 }
 415                 dst_node_size -= cur_size;
 416                 if (!dst_node_size) {
 417                         dst_node_start = amdgpu_mm_node_addr(dst->bo, ++dst_mm,
 418                                                              dst->mem);
 419                         dst_node_size = (dst_mm->size << PAGE_SHIFT);
 420                 } else {
 421                         dst_node_start += cur_size;
 422                         dst_page_offset = dst_node_start & (PAGE_SIZE - 1);
 423                 }
 424         }
 425 error:
 426         mutex_unlock(&adev->mman.gtt_window_lock);
 427         if (f)
 428                 *f = dma_fence_get(fence);
 429         dma_fence_put(fence);
 430         return r;
 431 }
 432
 433
 434 static int amdgpu_move_blit(struct ttm_buffer_object *bo,
 435                             bool evict, bool no_wait_gpu,
 436                             struct ttm_mem_reg *new_mem,
 437                             struct ttm_mem_reg *old_mem)
 438 {
 439         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
 440         struct amdgpu_copy_mem src, dst;
 441         struct dma_fence *fence = NULL;
 442         int r;
 443
 444         src.bo = bo;
 445         dst.bo = bo;
 446         src.mem = old_mem;
 447         dst.mem = new_mem;
 448         src.offset = 0;
 449         dst.offset = 0;
 450
 451         r = amdgpu_ttm_copy_mem_to_mem(adev, &src, &dst,
 452                                        new_mem->num_pages << PAGE_SHIFT,
 453                                        bo->resv, &fence);
 454         if (r)
 455                 goto error;
 456
 457         r = ttm_bo_pipeline_move(bo, fence, evict, new_mem);
 458         dma_fence_put(fence);
 459         return r;
 460
 461 error:
 462         if (fence)
 463                 dma_fence_wait(fence, false);
 464         dma_fence_put(fence);
 465         return r;
 466 }
 467
 468 static int amdgpu_move_vram_ram(struct ttm_buffer_object *bo, bool evict,
 469                                 struct ttm_operation_ctx *ctx,
 470                                 struct ttm_mem_reg *new_mem)
 471 {
 472         struct amdgpu_device *adev;
 473         struct ttm_mem_reg *old_mem = &bo->mem;
 474         struct ttm_mem_reg tmp_mem;
 475         struct ttm_place placements;
 476         struct ttm_placement placement;
 477         int r;
 478
 479         adev = amdgpu_ttm_adev(bo->bdev);
 480         tmp_mem = *new_mem;
 481         tmp_mem.mm_node = NULL;
 482         placement.num_placement = 1;
 483         placement.placement = &placements;
 484         placement.num_busy_placement = 1;
 485         placement.busy_placement = &placements;
 486         placements.fpfn = 0;
 487         placements.lpfn = 0;
 488         placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
 489         r = ttm_bo_mem_space(bo, &placement, &tmp_mem, ctx);
 490         if (unlikely(r)) {
 491                 return r;
 492         }
 493
 494         r = ttm_tt_set_placement_caching(bo->ttm, tmp_mem.placement);
 495         if (unlikely(r)) {
 496                 goto out_cleanup;
 497         }
 498
 499         r = ttm_tt_bind(bo->ttm, &tmp_mem, ctx);
 500         if (unlikely(r)) {
 501                 goto out_cleanup;
 502         }
 503         r = amdgpu_move_blit(bo, true, ctx->no_wait_gpu, &tmp_mem, old_mem);
 504         if (unlikely(r)) {
 505                 goto out_cleanup;
 506         }
 507         r = ttm_bo_move_ttm(bo, ctx, new_mem);
 508 out_cleanup:
 509         ttm_bo_mem_put(bo, &tmp_mem);
 510         return r;
 511 }
 512
 513 static int amdgpu_move_ram_vram(struct ttm_buffer_object *bo, bool evict,
 514                                 struct ttm_operation_ctx *ctx,
 515                                 struct ttm_mem_reg *new_mem)
 516 {
 517         struct amdgpu_device *adev;
 518         struct ttm_mem_reg *old_mem = &bo->mem;
 519         struct ttm_mem_reg tmp_mem;
 520         struct ttm_placement placement;
 521         struct ttm_place placements;
 522         int r;
 523
 524         adev = amdgpu_ttm_adev(bo->bdev);
 525         tmp_mem = *new_mem;
 526         tmp_mem.mm_node = NULL;
 527         placement.num_placement = 1;
 528         placement.placement = &placements;
 529         placement.num_busy_placement = 1;
 530         placement.busy_placement = &placements;
 531         placements.fpfn = 0;
 532         placements.lpfn = 0;
 533         placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
 534         r = ttm_bo_mem_space(bo, &placement, &tmp_mem, ctx);
 535         if (unlikely(r)) {
 536                 return r;
 537         }
 538         r = ttm_bo_move_ttm(bo, ctx, &tmp_mem);
 539         if (unlikely(r)) {
 540                 goto out_cleanup;
 541         }
 542         r = amdgpu_move_blit(bo, true, ctx->no_wait_gpu, new_mem, old_mem);
 543         if (unlikely(r)) {
 544                 goto out_cleanup;
 545         }
 546 out_cleanup:
 547         ttm_bo_mem_put(bo, &tmp_mem);
 548         return r;
 549 }
 550
 551 static int amdgpu_bo_move(struct ttm_buffer_object *bo, bool evict,
 552                           struct ttm_operation_ctx *ctx,
 553                           struct ttm_mem_reg *new_mem)
 554 {
 555         struct amdgpu_device *adev;
 556         struct amdgpu_bo *abo;
 557         struct ttm_mem_reg *old_mem = &bo->mem;
 558         int r;
 559
 560         /* Can't move a pinned BO */
 561         abo = ttm_to_amdgpu_bo(bo);
 562         if (WARN_ON_ONCE(abo->pin_count > 0))
 563                 return -EINVAL;
 564
 565         adev = amdgpu_ttm_adev(bo->bdev);
 566
 567         if (old_mem->mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
 568                 amdgpu_move_null(bo, new_mem);
 569                 return 0;
 570         }
 571         if ((old_mem->mem_type == TTM_PL_TT &&
 572              new_mem->mem_type == TTM_PL_SYSTEM) ||
 573             (old_mem->mem_type == TTM_PL_SYSTEM &&
 574              new_mem->mem_type == TTM_PL_TT)) {
 575                 /* bind is enough */
 576                 amdgpu_move_null(bo, new_mem);
 577                 return 0;
 578         }
 579
 580         if (!adev->mman.buffer_funcs_enabled)
 581                 goto memcpy;
 582
 583         if (old_mem->mem_type == TTM_PL_VRAM &&
 584             new_mem->mem_type == TTM_PL_SYSTEM) {
 585                 r = amdgpu_move_vram_ram(bo, evict, ctx, new_mem);
 586         } else if (old_mem->mem_type == TTM_PL_SYSTEM &&
 587                    new_mem->mem_type == TTM_PL_VRAM) {
 588                 r = amdgpu_move_ram_vram(bo, evict, ctx, new_mem);
 589         } else {
 590                 r = amdgpu_move_blit(bo, evict, ctx->no_wait_gpu,
 591                                      new_mem, old_mem);
 592         }
 593
 594         if (r) {
 595 memcpy:
 596                 r = ttm_bo_move_memcpy(bo, ctx, new_mem);
 597                 if (r) {
 598                         return r;
 599                 }
 600         }
 601
 602         if (bo->type == ttm_bo_type_device &&
 603             new_mem->mem_type == TTM_PL_VRAM &&
 604             old_mem->mem_type != TTM_PL_VRAM) {
 605                 /* amdgpu_bo_fault_reserve_notify will re-set this if the CPU
 606                  * accesses the BO after it's moved.
 607                  */
 608                 abo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
 609         }
 610
 611         /* update statistics */
 612         atomic64_add((u64)bo->num_pages << PAGE_SHIFT, &adev->num_bytes_moved);
 613         return 0;
 614 }
 615
 616 static int amdgpu_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
 617 {
 618         struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
 619         struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
 620         struct drm_mm_node *mm_node = mem->mm_node;
 621
 622         mem->bus.addr = NULL;
 623         mem->bus.offset = 0;
 624         mem->bus.size = mem->num_pages << PAGE_SHIFT;
 625         mem->bus.base = 0;
 626         mem->bus.is_iomem = false;
 627         if (!(man->flags & TTM_MEMTYPE_FLAG_MAPPABLE))
 628                 return -EINVAL;
 629         switch (mem->mem_type) {
 630         case TTM_PL_SYSTEM:
 631                 /* system memory */
 632                 return 0;
 633         case TTM_PL_TT:
 634                 break;
 635         case TTM_PL_VRAM:
 636                 mem->bus.offset = mem->start << PAGE_SHIFT;
 637                 /* check if it's visible */
 638                 if ((mem->bus.offset + mem->bus.size) > adev->gmc.visible_vram_size)
 639                         return -EINVAL;
 640                 /* Only physically contiguous buffers apply. In a contiguous
 641                  * buffer, size of the first mm_node would match the number of
 642                  * pages in ttm_mem_reg.
 643                  */
 644                 if (adev->mman.aper_base_kaddr &&
 645                     (mm_node->size == mem->num_pages))
 646                         mem->bus.addr = (u8 *)adev->mman.aper_base_kaddr +
 647                                         mem->bus.offset;
 648
 649                 mem->bus.base = adev->gmc.aper_base;
 650                 mem->bus.is_iomem = true;
 651                 break;
 652         default:
 653                 return -EINVAL;
 654         }
 655         return 0;
 656 }
 657
 658 static void amdgpu_ttm_io_mem_free(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
 659 {
 660 }
 661
 662 static unsigned long amdgpu_ttm_io_mem_pfn(struct ttm_buffer_object *bo,
 663                                            unsigned long page_offset)
 664 {
 665         struct drm_mm_node *mm;
 666         unsigned long offset = (page_offset << PAGE_SHIFT);
 667
 668         mm = amdgpu_find_mm_node(&bo->mem, &offset);
 669         return (bo->mem.bus.base >> PAGE_SHIFT) + mm->start +
 670                 (offset >> PAGE_SHIFT);
 671 }
 672
 673 /*
 674  * TTM backend functions.
 675  */
 676 struct amdgpu_ttm_gup_task_list {
 677         struct list_head        list;
 678         struct task_struct      *task;
 679 };
 680
 681 struct amdgpu_ttm_tt {
 682         struct ttm_dma_tt       ttm;
 683         u64                     offset;
 684         uint64_t                userptr;
 685         struct mm_struct        *usermm;
 686         uint32_t                userflags;
 687         spinlock_t              guptasklock;
 688         struct list_head        guptasks;
 689         atomic_t                mmu_invalidations;
 690         uint32_t                last_set_pages;
 691 };
 692
 693 int amdgpu_ttm_tt_get_user_pages(struct ttm_tt *ttm, struct page **pages)
 694 {
 695         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 696         unsigned int flags = 0;
 697         unsigned pinned = 0;
 698         int r;
 699
 700         if (!(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY))
 701                 flags |= FOLL_WRITE;
 702
 703         down_read(&current->mm->mmap_sem);
 704
 705         if (gtt->userflags & AMDGPU_GEM_USERPTR_ANONONLY) {
 706                 /* check that we only use anonymous memory
 707                    to prevent problems with writeback */
 708                 unsigned long end = gtt->userptr + ttm->num_pages * PAGE_SIZE;
 709                 struct vm_area_struct *vma;
 710
 711                 vma = find_vma(gtt->usermm, gtt->userptr);
 712                 if (!vma || vma->vm_file || vma->vm_end < end) {
 713                         up_read(&current->mm->mmap_sem);
 714                         return -EPERM;
 715                 }
 716         }
 717
 718         do {
 719                 unsigned num_pages = ttm->num_pages - pinned;
 720                 uint64_t userptr = gtt->userptr + pinned * PAGE_SIZE;
 721                 struct page **p = pages + pinned;
 722                 struct amdgpu_ttm_gup_task_list guptask;
 723
 724                 guptask.task = current;
 725                 spin_lock(&gtt->guptasklock);
 726                 list_add(&guptask.list, &gtt->guptasks);
 727                 spin_unlock(&gtt->guptasklock);
 728
 729                 r = get_user_pages(userptr, num_pages, flags, p, NULL);
 730
 731                 spin_lock(&gtt->guptasklock);
 732                 list_del(&guptask.list);
 733                 spin_unlock(&gtt->guptasklock);
 734
 735                 if (r < 0)
 736                         goto release_pages;
 737
 738                 pinned += r;
 739
 740         } while (pinned < ttm->num_pages);
 741
 742         up_read(&current->mm->mmap_sem);
 743         return 0;
 744
 745 release_pages:
 746         release_pages(pages, pinned);
 747         up_read(&current->mm->mmap_sem);
 748         return r;
 749 }
 750
 751 void amdgpu_ttm_tt_set_user_pages(struct ttm_tt *ttm, struct page **pages)
 752 {
 753         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 754         unsigned i;
 755
 756         gtt->last_set_pages = atomic_read(&gtt->mmu_invalidations);
 757         for (i = 0; i < ttm->num_pages; ++i) {
 758                 if (ttm->pages[i])
 759                         put_page(ttm->pages[i]);
 760
 761                 ttm->pages[i] = pages ? pages[i] : NULL;
 762         }
 763 }
 764
 765 void amdgpu_ttm_tt_mark_user_pages(struct ttm_tt *ttm)
 766 {
 767         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 768         unsigned i;
 769
 770         for (i = 0; i < ttm->num_pages; ++i) {
 771                 struct page *page = ttm->pages[i];
 772
 773                 if (!page)
 774                         continue;
 775
 776                 if (!(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY))
 777                         set_page_dirty(page);
 778
 779                 mark_page_accessed(page);
 780         }
 781 }
 782
 783 /* prepare the sg table with the user pages */
 784 static int amdgpu_ttm_tt_pin_userptr(struct ttm_tt *ttm)
 785 {
 786         struct amdgpu_device *adev = amdgpu_ttm_adev(ttm->bdev);
 787         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 788         unsigned nents;
 789         int r;
 790
 791         int write = !(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
 792         enum dma_data_direction direction = write ?
 793                 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
 794
 795         r = sg_alloc_table_from_pages(ttm->sg, ttm->pages, ttm->num_pages, 0,
 796                                       ttm->num_pages << PAGE_SHIFT,
 797                                       GFP_KERNEL);
 798         if (r)
 799                 goto release_sg;
 800
 801         r = -ENOMEM;
 802         nents = dma_map_sg(adev->dev, ttm->sg->sgl, ttm->sg->nents, direction);
 803         if (nents != ttm->sg->nents)
 804                 goto release_sg;
 805
 806         drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
 807                                          gtt->ttm.dma_address, ttm->num_pages);
 808
 809         return 0;
 810
 811 release_sg:
 812         kfree(ttm->sg);
 813         return r;
 814 }
 815
 816 static void amdgpu_ttm_tt_unpin_userptr(struct ttm_tt *ttm)
 817 {
 818         struct amdgpu_device *adev = amdgpu_ttm_adev(ttm->bdev);
 819         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 820
 821         int write = !(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
 822         enum dma_data_direction direction = write ?
 823                 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
 824
 825         /* double check that we don't free the table twice */
 826         if (!ttm->sg->sgl)
 827                 return;
 828
 829         /* free the sg table and pages again */
 830         dma_unmap_sg(adev->dev, ttm->sg->sgl, ttm->sg->nents, direction);
 831
 832         amdgpu_ttm_tt_mark_user_pages(ttm);
 833
 834         sg_free_table(ttm->sg);
 835 }
 836
 837 int amdgpu_ttm_gart_bind(struct amdgpu_device *adev,
 838                                 struct ttm_buffer_object *tbo,
 839                                 uint64_t flags)
 840 {
 841         struct amdgpu_bo *abo = ttm_to_amdgpu_bo(tbo);
 842         struct ttm_tt *ttm = tbo->ttm;
 843         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 844         int r;
 845
 846         if (abo->flags & AMDGPU_GEM_CREATE_MQD_GFX9) {
 847                 uint64_t page_idx = 1;
 848
 849                 r = amdgpu_gart_bind(adev, gtt->offset, page_idx,
 850                                 ttm->pages, gtt->ttm.dma_address, flags);
 851                 if (r)
 852                         goto gart_bind_fail;
 853
 854                 /* Patch mtype of the second part BO */
 855                 flags &=  ~AMDGPU_PTE_MTYPE_MASK;
 856                 flags |= AMDGPU_PTE_MTYPE(AMDGPU_MTYPE_NC);
 857
 858                 r = amdgpu_gart_bind(adev,
 859                                 gtt->offset + (page_idx << PAGE_SHIFT),
 860                                 ttm->num_pages - page_idx,
 861                                 &ttm->pages[page_idx],
 862                                 &(gtt->ttm.dma_address[page_idx]), flags);
 863         } else {
 864                 r = amdgpu_gart_bind(adev, gtt->offset, ttm->num_pages,
 865                                      ttm->pages, gtt->ttm.dma_address, flags);
 866         }
 867
 868 gart_bind_fail:
 869         if (r)
 870                 DRM_ERROR("failed to bind %lu pages at 0x%08llX\n",
 871                           ttm->num_pages, gtt->offset);
 872
 873         return r;
 874 }
 875
 876 static int amdgpu_ttm_backend_bind(struct ttm_tt *ttm,
 877                                    struct ttm_mem_reg *bo_mem)
 878 {
 879         struct amdgpu_device *adev = amdgpu_ttm_adev(ttm->bdev);
 880         struct amdgpu_ttm_tt *gtt = (void*)ttm;
 881         uint64_t flags;
 882         int r = 0;
 883
 884         if (gtt->userptr) {
 885                 r = amdgpu_ttm_tt_pin_userptr(ttm);
 886                 if (r) {
 887                         DRM_ERROR("failed to pin userptr\n");
 888                         return r;
 889                 }
 890         }
 891         if (!ttm->num_pages) {
 892                 WARN(1, "nothing to bind %lu pages for mreg %p back %p!\n",
 893                      ttm->num_pages, bo_mem, ttm);
 894         }
 895
 896         if (bo_mem->mem_type == AMDGPU_PL_GDS ||
 897             bo_mem->mem_type == AMDGPU_PL_GWS ||
 898             bo_mem->mem_type == AMDGPU_PL_OA)
 899                 return -EINVAL;
 900
 901         if (!amdgpu_gtt_mgr_has_gart_addr(bo_mem)) {
 902                 gtt->offset = AMDGPU_BO_INVALID_OFFSET;
 903                 return 0;
 904         }
 905
 906         flags = amdgpu_ttm_tt_pte_flags(adev, ttm, bo_mem);
 907         gtt->offset = (u64)bo_mem->start << PAGE_SHIFT;
 908         r = amdgpu_gart_bind(adev, gtt->offset, ttm->num_pages,
 909                 ttm->pages, gtt->ttm.dma_address, flags);
 910
 911         if (r)
 912                 DRM_ERROR("failed to bind %lu pages at 0x%08llX\n",
 913                           ttm->num_pages, gtt->offset);
 914         return r;
 915 }
 916
 917 int amdgpu_ttm_alloc_gart(struct ttm_buffer_object *bo)
 918 {
 919         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
 920         struct ttm_operation_ctx ctx = { false, false };
 921         struct amdgpu_ttm_tt *gtt = (void*)bo->ttm;
 922         struct ttm_mem_reg tmp;
 923         struct ttm_placement placement;
 924         struct ttm_place placements;
 925         uint64_t flags;
 926         int r;
 927
 928         if (bo->mem.mem_type != TTM_PL_TT ||
 929             amdgpu_gtt_mgr_has_gart_addr(&bo->mem))
 930                 return 0;
 931
 932         tmp = bo->mem;
 933         tmp.mm_node = NULL;
 934         placement.num_placement = 1;
 935         placement.placement = &placements;
 936         placement.num_busy_placement = 1;
 937         placement.busy_placement = &placements;
 938         placements.fpfn = 0;
 939         placements.lpfn = adev->gmc.gart_size >> PAGE_SHIFT;
 940         placements.flags = (bo->mem.placement & ~TTM_PL_MASK_MEM) |
 941                 TTM_PL_FLAG_TT;
 942
 943         r = ttm_bo_mem_space(bo, &placement, &tmp, &ctx);
 944         if (unlikely(r))
 945                 return r;
 946
 947         flags = amdgpu_ttm_tt_pte_flags(adev, bo->ttm, &tmp);
 948         gtt->offset = (u64)tmp.start << PAGE_SHIFT;
 949         r = amdgpu_ttm_gart_bind(adev, bo, flags);
 950         if (unlikely(r)) {
 951                 ttm_bo_mem_put(bo, &tmp);
 952                 return r;
 953         }
 954
 955         ttm_bo_mem_put(bo, &bo->mem);
 956         bo->mem = tmp;
 957         bo->offset = (bo->mem.start << PAGE_SHIFT) +
 958                 bo->bdev->man[bo->mem.mem_type].gpu_offset;
 959
 960         return 0;
 961 }
 962
 963 int amdgpu_ttm_recover_gart(struct ttm_buffer_object *tbo)
 964 {
 965         struct amdgpu_device *adev = amdgpu_ttm_adev(tbo->bdev);
 966         uint64_t flags;
 967         int r;
 968
 969         if (!tbo->ttm)
 970                 return 0;
 971
 972         flags = amdgpu_ttm_tt_pte_flags(adev, tbo->ttm, &tbo->mem);
 973         r = amdgpu_ttm_gart_bind(adev, tbo, flags);
 974
 975         return r;
 976 }
 977
 978 static int amdgpu_ttm_backend_unbind(struct ttm_tt *ttm)
 979 {
 980         struct amdgpu_device *adev = amdgpu_ttm_adev(ttm->bdev);
 981         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 982         int r;
 983
 984         if (gtt->userptr)
 985                 amdgpu_ttm_tt_unpin_userptr(ttm);
 986
 987         if (gtt->offset == AMDGPU_BO_INVALID_OFFSET)
 988                 return 0;
 989
 990         /* unbind shouldn't be done for GDS/GWS/OA in ttm_bo_clean_mm */
 991         r = amdgpu_gart_unbind(adev, gtt->offset, ttm->num_pages);
 992         if (r)
 993                 DRM_ERROR("failed to unbind %lu pages at 0x%08llX\n",
 994                           gtt->ttm.ttm.num_pages, gtt->offset);
 995         return r;
 996 }
 997
 998 static void amdgpu_ttm_backend_destroy(struct ttm_tt *ttm)
 999 {
1000         struct amdgpu_ttm_tt *gtt = (void *)ttm;
1001
1002         ttm_dma_tt_fini(&gtt->ttm);
1003         kfree(gtt);
1004 }
1005
1006 static struct ttm_backend_func amdgpu_backend_func = {
1007         .bind = &amdgpu_ttm_backend_bind,
1008         .unbind = &amdgpu_ttm_backend_unbind,
1009         .destroy = &amdgpu_ttm_backend_destroy,
1010 };
1011
1012 static struct ttm_tt *amdgpu_ttm_tt_create(struct ttm_buffer_object *bo,
1013                                            uint32_t page_flags)
1014 {
1015         struct amdgpu_device *adev;
1016         struct amdgpu_ttm_tt *gtt;
1017
1018         adev = amdgpu_ttm_adev(bo->bdev);
1019
1020         gtt = kzalloc(sizeof(struct amdgpu_ttm_tt), GFP_KERNEL);
1021         if (gtt == NULL) {
1022                 return NULL;
1023         }
1024         gtt->ttm.ttm.func = &amdgpu_backend_func;
1025         if (ttm_sg_tt_init(&gtt->ttm, bo, page_flags)) {
1026                 kfree(gtt);
1027                 return NULL;
1028         }
1029         return &gtt->ttm.ttm;
1030 }
1031
1032 static int amdgpu_ttm_tt_populate(struct ttm_tt *ttm,
1033                         struct ttm_operation_ctx *ctx)
1034 {
1035         struct amdgpu_device *adev = amdgpu_ttm_adev(ttm->bdev);
1036         struct amdgpu_ttm_tt *gtt = (void *)ttm;
1037         bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
1038
1039         if (gtt && gtt->userptr) {
1040                 ttm->sg = kzalloc(sizeof(struct sg_table), GFP_KERNEL);
1041                 if (!ttm->sg)
1042                         return -ENOMEM;
1043
1044                 ttm->page_flags |= TTM_PAGE_FLAG_SG;
1045                 ttm->state = tt_unbound;
1046                 return 0;
1047         }
1048
1049         if (slave && ttm->sg) {
1050                 drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
1051                                                  gtt->ttm.dma_address,
1052                                                  ttm->num_pages);
1053                 ttm->state = tt_unbound;
1054                 return 0;
1055         }
1056
1057 #ifdef CONFIG_SWIOTLB
1058         if (adev->need_swiotlb && swiotlb_nr_tbl()) {
1059                 return ttm_dma_populate(&gtt->ttm, adev->dev, ctx);
1060         }
1061 #endif
1062
1063         return ttm_populate_and_map_pages(adev->dev, &gtt->ttm, ctx);
1064 }
1065
1066 static void amdgpu_ttm_tt_unpopulate(struct ttm_tt *ttm)
1067 {
1068         struct amdgpu_device *adev;
1069         struct amdgpu_ttm_tt *gtt = (void *)ttm;
1070         bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
1071
1072         if (gtt && gtt->userptr) {
1073                 amdgpu_ttm_tt_set_user_pages(ttm, NULL);
1074                 kfree(ttm->sg);
1075                 ttm->page_flags &= ~TTM_PAGE_FLAG_SG;
1076                 return;
1077         }
1078
1079         if (slave)
1080                 return;
1081
1082         adev = amdgpu_ttm_adev(ttm->bdev);
1083
1084 #ifdef CONFIG_SWIOTLB
1085         if (adev->need_swiotlb && swiotlb_nr_tbl()) {
1086                 ttm_dma_unpopulate(&gtt->ttm, adev->dev);
1087                 return;
1088         }
1089 #endif
1090
1091         ttm_unmap_and_unpopulate_pages(adev->dev, &gtt->ttm);
1092 }
1093
1094 int amdgpu_ttm_tt_set_userptr(struct ttm_tt *ttm, uint64_t addr,
1095                               uint32_t flags)
1096 {
1097         struct amdgpu_ttm_tt *gtt = (void *)ttm;
1098
1099         if (gtt == NULL)
1100                 return -EINVAL;
1101
1102         gtt->userptr = addr;
1103         gtt->usermm = current->mm;
1104         gtt->userflags = flags;
1105         spin_lock_init(&gtt->guptasklock);
1106         INIT_LIST_HEAD(&gtt->guptasks);
1107         atomic_set(&gtt->mmu_invalidations, 0);
1108         gtt->last_set_pages = 0;
1109
1110         return 0;
1111 }
1112
1113 struct mm_struct *amdgpu_ttm_tt_get_usermm(struct ttm_tt *ttm)
1114 {
1115         struct amdgpu_ttm_tt *gtt = (void *)ttm;
1116
1117         if (gtt == NULL)
1118                 return NULL;
1119
1120         return gtt->usermm;
1121 }
1122
1123 bool amdgpu_ttm_tt_affect_userptr(struct ttm_tt *ttm, unsigned long start,
1124                                   unsigned long end)
1125 {
1126         struct amdgpu_ttm_tt *gtt = (void *)ttm;
1127         struct amdgpu_ttm_gup_task_list *entry;
1128         unsigned long size;
1129
1130         if (gtt == NULL || !gtt->userptr)
1131                 return false;
1132
1133         size = (unsigned long)gtt->ttm.ttm.num_pages * PAGE_SIZE;
1134         if (gtt->userptr > end || gtt->userptr + size <= start)
1135                 return false;
1136
1137         spin_lock(&gtt->guptasklock);
1138         list_for_each_entry(entry, &gtt->guptasks, list) {
1139                 if (entry->task == current) {
1140                         spin_unlock(&gtt->guptasklock);
1141                         return false;
1142                 }
1143         }
1144         spin_unlock(&gtt->guptasklock);
1145
1146         atomic_inc(&gtt->mmu_invalidations);
1147
1148         return true;
1149 }
1150
1151 bool amdgpu_ttm_tt_userptr_invalidated(struct ttm_tt *ttm,
1152                                        int *last_invalidated)
1153 {
1154         struct amdgpu_ttm_tt *gtt = (void *)ttm;
1155         int prev_invalidated = *last_invalidated;
1156
1157         *last_invalidated = atomic_read(&gtt->mmu_invalidations);
1158         return prev_invalidated != *last_invalidated;
1159 }
1160
1161 bool amdgpu_ttm_tt_userptr_needs_pages(struct ttm_tt *ttm)
1162 {
1163         struct amdgpu_ttm_tt *gtt = (void *)ttm;
1164
1165         if (gtt == NULL || !gtt->userptr)
1166                 return false;
1167
1168         return atomic_read(&gtt->mmu_invalidations) != gtt->last_set_pages;
1169 }
1170
1171 bool amdgpu_ttm_tt_is_readonly(struct ttm_tt *ttm)
1172 {
1173         struct amdgpu_ttm_tt *gtt = (void *)ttm;
1174
1175         if (gtt == NULL)
1176                 return false;
1177
1178         return !!(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
1179 }
1180
1181 uint64_t amdgpu_ttm_tt_pte_flags(struct amdgpu_device *adev, struct ttm_tt *ttm,
1182                                  struct ttm_mem_reg *mem)
1183 {
1184         uint64_t flags = 0;
1185
1186         if (mem && mem->mem_type != TTM_PL_SYSTEM)
1187                 flags |= AMDGPU_PTE_VALID;
1188
1189         if (mem && mem->mem_type == TTM_PL_TT) {
1190                 flags |= AMDGPU_PTE_SYSTEM;
1191
1192                 if (ttm->caching_state == tt_cached)
1193                         flags |= AMDGPU_PTE_SNOOPED;
1194         }
1195
1196         flags |= adev->gart.gart_pte_flags;
1197         flags |= AMDGPU_PTE_READABLE;
1198
1199         if (!amdgpu_ttm_tt_is_readonly(ttm))
1200                 flags |= AMDGPU_PTE_WRITEABLE;
1201
1202         return flags;
1203 }
1204
1205 static bool amdgpu_ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
1206                                             const struct ttm_place *place)
1207 {
1208         unsigned long num_pages = bo->mem.num_pages;
1209         struct drm_mm_node *node = bo->mem.mm_node;
1210         struct reservation_object_list *flist;
1211         struct dma_fence *f;
1212         int i;
1213
1214         /* If bo is a KFD BO, check if the bo belongs to the current process.
1215          * If true, then return false as any KFD process needs all its BOs to
1216          * be resident to run successfully
1217          */
1218         flist = reservation_object_get_list(bo->resv);
1219         if (flist) {
1220                 for (i = 0; i < flist->shared_count; ++i) {
1221                         f = rcu_dereference_protected(flist->shared[i],
1222                                 reservation_object_held(bo->resv));
1223                         if (amdkfd_fence_check_mm(f, current->mm))
1224                                 return false;
1225                 }
1226         }
1227
1228         switch (bo->mem.mem_type) {
1229         case TTM_PL_TT:
1230                 return true;
1231
1232         case TTM_PL_VRAM:
1233                 /* Check each drm MM node individually */
1234                 while (num_pages) {
1235                         if (place->fpfn < (node->start + node->size) &&
1236                             !(place->lpfn && place->lpfn <= node->start))
1237                                 return true;
1238
1239                         num_pages -= node->size;
1240                         ++node;
1241                 }
1242                 return false;
1243
1244         default:
1245                 break;
1246         }
1247
1248         return ttm_bo_eviction_valuable(bo, place);
1249 }
1250
1251 static int amdgpu_ttm_access_memory(struct ttm_buffer_object *bo,
1252                                     unsigned long offset,
1253                                     void *buf, int len, int write)
1254 {
1255         struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
1256         struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev);
1257         struct drm_mm_node *nodes;
1258         uint32_t value = 0;
1259         int ret = 0;
1260         uint64_t pos;
1261         unsigned long flags;
1262
1263         if (bo->mem.mem_type != TTM_PL_VRAM)
1264                 return -EIO;
1265
1266         nodes = amdgpu_find_mm_node(&abo->tbo.mem, &offset);
1267         pos = (nodes->start << PAGE_SHIFT) + offset;
1268
1269         while (len && pos < adev->gmc.mc_vram_size) {
1270                 uint64_t aligned_pos = pos & ~(uint64_t)3;
1271                 uint32_t bytes = 4 - (pos & 3);
1272                 uint32_t shift = (pos & 3) * 8;
1273                 uint32_t mask = 0xffffffff << shift;
1274
1275                 if (len < bytes) {
1276                         mask &= 0xffffffff >> (bytes - len) * 8;
1277                         bytes = len;
1278                 }
1279
1280                 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
1281                 WREG32_NO_KIQ(mmMM_INDEX, ((uint32_t)aligned_pos) | 0x80000000);
1282                 WREG32_NO_KIQ(mmMM_INDEX_HI, aligned_pos >> 31);
1283                 if (!write || mask != 0xffffffff)
1284                         value = RREG32_NO_KIQ(mmMM_DATA);
1285                 if (write) {
1286                         value &= ~mask;
1287                         value |= (*(uint32_t *)buf << shift) & mask;
1288                         WREG32_NO_KIQ(mmMM_DATA, value);
1289                 }
1290                 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
1291                 if (!write) {
1292                         value = (value & mask) >> shift;
1293                         memcpy(buf, &value, bytes);
1294                 }
1295
1296                 ret += bytes;
1297                 buf = (uint8_t *)buf + bytes;
1298                 pos += bytes;
1299                 len -= bytes;
1300                 if (pos >= (nodes->start + nodes->size) << PAGE_SHIFT) {
1301                         ++nodes;
1302                         pos = (nodes->start << PAGE_SHIFT);
1303                 }
1304         }
1305
1306         return ret;
1307 }
1308
1309 static struct ttm_bo_driver amdgpu_bo_driver = {
1310         .ttm_tt_create = &amdgpu_ttm_tt_create,
1311         .ttm_tt_populate = &amdgpu_ttm_tt_populate,
1312         .ttm_tt_unpopulate = &amdgpu_ttm_tt_unpopulate,
1313         .invalidate_caches = &amdgpu_invalidate_caches,
1314         .init_mem_type = &amdgpu_init_mem_type,
1315         .eviction_valuable = amdgpu_ttm_bo_eviction_valuable,
1316         .evict_flags = &amdgpu_evict_flags,
1317         .move = &amdgpu_bo_move,
1318         .verify_access = &amdgpu_verify_access,
1319         .move_notify = &amdgpu_bo_move_notify,
1320         .fault_reserve_notify = &amdgpu_bo_fault_reserve_notify,
1321         .io_mem_reserve = &amdgpu_ttm_io_mem_reserve,
1322         .io_mem_free = &amdgpu_ttm_io_mem_free,
1323         .io_mem_pfn = amdgpu_ttm_io_mem_pfn,
1324         .access_memory = &amdgpu_ttm_access_memory
1325 };
1326
1327 /*
1328  * Firmware Reservation functions
1329  */
1330 /**
1331  * amdgpu_ttm_fw_reserve_vram_fini - free fw reserved vram
1332  *
1333  * @adev: amdgpu_device pointer
1334  *
1335  * free fw reserved vram if it has been reserved.
1336  */
1337 static void amdgpu_ttm_fw_reserve_vram_fini(struct amdgpu_device *adev)
1338 {
1339         amdgpu_bo_free_kernel(&adev->fw_vram_usage.reserved_bo,
1340                 NULL, &adev->fw_vram_usage.va);
1341 }
1342
1343 /**
1344  * amdgpu_ttm_fw_reserve_vram_init - create bo vram reservation from fw
1345  *
1346  * @adev: amdgpu_device pointer
1347  *
1348  * create bo vram reservation from fw.
1349  */
1350 static int amdgpu_ttm_fw_reserve_vram_init(struct amdgpu_device *adev)
1351 {
1352         struct ttm_operation_ctx ctx = { false, false };
1353         struct amdgpu_bo_param bp;
1354         int r = 0;
1355         int i;
1356         u64 vram_size = adev->gmc.visible_vram_size;
1357         u64 offset = adev->fw_vram_usage.start_offset;
1358         u64 size = adev->fw_vram_usage.size;
1359         struct amdgpu_bo *bo;
1360
1361         memset(&bp, 0, sizeof(bp));
1362         bp.size = adev->fw_vram_usage.size;
1363         bp.byte_align = PAGE_SIZE;
1364         bp.domain = AMDGPU_GEM_DOMAIN_VRAM;
1365         bp.flags = AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
1366                 AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
1367         bp.type = ttm_bo_type_kernel;
1368         bp.resv = NULL;
1369         adev->fw_vram_usage.va = NULL;
1370         adev->fw_vram_usage.reserved_bo = NULL;
1371
1372         if (adev->fw_vram_usage.size > 0 &&
1373                 adev->fw_vram_usage.size <= vram_size) {
1374
1375                 r = amdgpu_bo_create(adev, &bp,
1376                                      &adev->fw_vram_usage.reserved_bo);
1377                 if (r)
1378                         goto error_create;
1379
1380                 r = amdgpu_bo_reserve(adev->fw_vram_usage.reserved_bo, false);
1381                 if (r)
1382                         goto error_reserve;
1383
1384                 /* remove the original mem node and create a new one at the
1385                  * request position
1386                  */
1387                 bo = adev->fw_vram_usage.reserved_bo;
1388                 offset = ALIGN(offset, PAGE_SIZE);
1389                 for (i = 0; i < bo->placement.num_placement; ++i) {
1390                         bo->placements[i].fpfn = offset >> PAGE_SHIFT;
1391                         bo->placements[i].lpfn = (offset + size) >> PAGE_SHIFT;
1392                 }
1393
1394                 ttm_bo_mem_put(&bo->tbo, &bo->tbo.mem);
1395                 r = ttm_bo_mem_space(&bo->tbo, &bo->placement,
1396                                      &bo->tbo.mem, &ctx);
1397                 if (r)
1398                         goto error_pin;
1399
1400                 r = amdgpu_bo_pin_restricted(adev->fw_vram_usage.reserved_bo,
1401                         AMDGPU_GEM_DOMAIN_VRAM,
1402                         adev->fw_vram_usage.start_offset,
1403                         (adev->fw_vram_usage.start_offset +
1404                         adev->fw_vram_usage.size), NULL);
1405                 if (r)
1406                         goto error_pin;
1407                 r = amdgpu_bo_kmap(adev->fw_vram_usage.reserved_bo,
1408                         &adev->fw_vram_usage.va);
1409                 if (r)
1410                         goto error_kmap;
1411
1412                 amdgpu_bo_unreserve(adev->fw_vram_usage.reserved_bo);
1413         }
1414         return r;
1415
1416 error_kmap:
1417         amdgpu_bo_unpin(adev->fw_vram_usage.reserved_bo);
1418 error_pin:
1419         amdgpu_bo_unreserve(adev->fw_vram_usage.reserved_bo);
1420 error_reserve:
1421         amdgpu_bo_unref(&adev->fw_vram_usage.reserved_bo);
1422 error_create:
1423         adev->fw_vram_usage.va = NULL;
1424         adev->fw_vram_usage.reserved_bo = NULL;
1425         return r;
1426 }
1427
1428 int amdgpu_ttm_init(struct amdgpu_device *adev)
1429 {
1430         uint64_t gtt_size;
1431         int r;
1432         u64 vis_vram_limit;
1433
1434         r = amdgpu_ttm_global_init(adev);
1435         if (r) {
1436                 return r;
1437         }
1438         /* No others user of address space so set it to 0 */
1439         r = ttm_bo_device_init(&adev->mman.bdev,
1440                                adev->mman.bo_global_ref.ref.object,
1441                                &amdgpu_bo_driver,
1442                                adev->ddev->anon_inode->i_mapping,
1443                                DRM_FILE_PAGE_OFFSET,
1444                                adev->need_dma32);
1445         if (r) {
1446                 DRM_ERROR("failed initializing buffer object driver(%d).\n", r);
1447                 return r;
1448         }
1449         adev->mman.initialized = true;
1450
1451         /* We opt to avoid OOM on system pages allocations */
1452         adev->mman.bdev.no_retry = true;
1453
1454         r = ttm_bo_init_mm(&adev->mman.bdev, TTM_PL_VRAM,
1455                                 adev->gmc.real_vram_size >> PAGE_SHIFT);
1456         if (r) {
1457                 DRM_ERROR("Failed initializing VRAM heap.\n");
1458                 return r;
1459         }
1460
1461         /* Reduce size of CPU-visible VRAM if requested */
1462         vis_vram_limit = (u64)amdgpu_vis_vram_limit * 1024 * 1024;
1463         if (amdgpu_vis_vram_limit > 0 &&
1464             vis_vram_limit <= adev->gmc.visible_vram_size)
1465                 adev->gmc.visible_vram_size = vis_vram_limit;
1466
1467         /* Change the size here instead of the init above so only lpfn is affected */
1468         amdgpu_ttm_set_buffer_funcs_status(adev, false);
1469 #ifdef CONFIG_64BIT
1470         adev->mman.aper_base_kaddr = ioremap_wc(adev->gmc.aper_base,
1471                                                 adev->gmc.visible_vram_size);
1472 #endif
1473
1474         /*
1475          *The reserved vram for firmware must be pinned to the specified
1476          *place on the VRAM, so reserve it early.
1477          */
1478         r = amdgpu_ttm_fw_reserve_vram_init(adev);
1479         if (r) {
1480                 return r;
1481         }
1482
1483         if (adev->gmc.stolen_size) {
1484                 r = amdgpu_bo_create_kernel(adev, adev->gmc.stolen_size, PAGE_SIZE,
1485                                             AMDGPU_GEM_DOMAIN_VRAM,
1486                                             &adev->stolen_vga_memory,
1487                                             NULL, NULL);
1488                 if (r)
1489                         return r;
1490         }
1491         DRM_INFO("amdgpu: %uM of VRAM memory ready\n",
1492                  (unsigned) (adev->gmc.real_vram_size / (1024 * 1024)));
1493
1494         if (amdgpu_gtt_size == -1) {
1495                 struct sysinfo si;
1496
1497                 si_meminfo(&si);
1498                 gtt_size = min(max((AMDGPU_DEFAULT_GTT_SIZE_MB << 20),
1499                                adev->gmc.mc_vram_size),
1500                                ((uint64_t)si.totalram * si.mem_unit * 3/4));
1501         }
1502         else
1503                 gtt_size = (uint64_t)amdgpu_gtt_size << 20;
1504         r = ttm_bo_init_mm(&adev->mman.bdev, TTM_PL_TT, gtt_size >> PAGE_SHIFT);
1505         if (r) {
1506                 DRM_ERROR("Failed initializing GTT heap.\n");
1507                 return r;
1508         }
1509         DRM_INFO("amdgpu: %uM of GTT memory ready.\n",
1510                  (unsigned)(gtt_size / (1024 * 1024)));
1511
1512         adev->gds.mem.total_size = adev->gds.mem.total_size << AMDGPU_GDS_SHIFT;
1513         adev->gds.mem.gfx_partition_size = adev->gds.mem.gfx_partition_size << AMDGPU_GDS_SHIFT;
1514         adev->gds.mem.cs_partition_size = adev->gds.mem.cs_partition_size << AMDGPU_GDS_SHIFT;
1515         adev->gds.gws.total_size = adev->gds.gws.total_size << AMDGPU_GWS_SHIFT;
1516         adev->gds.gws.gfx_partition_size = adev->gds.gws.gfx_partition_size << AMDGPU_GWS_SHIFT;
1517         adev->gds.gws.cs_partition_size = adev->gds.gws.cs_partition_size << AMDGPU_GWS_SHIFT;
1518         adev->gds.oa.total_size = adev->gds.oa.total_size << AMDGPU_OA_SHIFT;
1519         adev->gds.oa.gfx_partition_size = adev->gds.oa.gfx_partition_size << AMDGPU_OA_SHIFT;
1520         adev->gds.oa.cs_partition_size = adev->gds.oa.cs_partition_size << AMDGPU_OA_SHIFT;
1521         /* GDS Memory */
1522         if (adev->gds.mem.total_size) {
1523                 r = ttm_bo_init_mm(&adev->mman.bdev, AMDGPU_PL_GDS,
1524                                    adev->gds.mem.total_size >> PAGE_SHIFT);
1525                 if (r) {
1526                         DRM_ERROR("Failed initializing GDS heap.\n");
1527                         return r;
1528                 }
1529         }
1530
1531         /* GWS */
1532         if (adev->gds.gws.total_size) {
1533                 r = ttm_bo_init_mm(&adev->mman.bdev, AMDGPU_PL_GWS,
1534                                    adev->gds.gws.total_size >> PAGE_SHIFT);
1535                 if (r) {
1536                         DRM_ERROR("Failed initializing gws heap.\n");
1537                         return r;
1538                 }
1539         }
1540
1541         /* OA */
1542         if (adev->gds.oa.total_size) {
1543                 r = ttm_bo_init_mm(&adev->mman.bdev, AMDGPU_PL_OA,
1544                                    adev->gds.oa.total_size >> PAGE_SHIFT);
1545                 if (r) {
1546                         DRM_ERROR("Failed initializing oa heap.\n");
1547                         return r;
1548                 }
1549         }
1550
1551         r = amdgpu_ttm_debugfs_init(adev);
1552         if (r) {
1553                 DRM_ERROR("Failed to init debugfs\n");
1554                 return r;
1555         }
1556         return 0;
1557 }
1558
1559 void amdgpu_ttm_late_init(struct amdgpu_device *adev)
1560 {
1561         amdgpu_bo_free_kernel(&adev->stolen_vga_memory, NULL, NULL);
1562 }
1563
1564 void amdgpu_ttm_fini(struct amdgpu_device *adev)
1565 {
1566         if (!adev->mman.initialized)
1567                 return;
1568
1569         amdgpu_ttm_debugfs_fini(adev);
1570         amdgpu_ttm_fw_reserve_vram_fini(adev);
1571         if (adev->mman.aper_base_kaddr)
1572                 iounmap(adev->mman.aper_base_kaddr);
1573         adev->mman.aper_base_kaddr = NULL;
1574
1575         ttm_bo_clean_mm(&adev->mman.bdev, TTM_PL_VRAM);
1576         ttm_bo_clean_mm(&adev->mman.bdev, TTM_PL_TT);
1577         if (adev->gds.mem.total_size)
1578                 ttm_bo_clean_mm(&adev->mman.bdev, AMDGPU_PL_GDS);
1579         if (adev->gds.gws.total_size)
1580                 ttm_bo_clean_mm(&adev->mman.bdev, AMDGPU_PL_GWS);
1581         if (adev->gds.oa.total_size)
1582                 ttm_bo_clean_mm(&adev->mman.bdev, AMDGPU_PL_OA);
1583         ttm_bo_device_release(&adev->mman.bdev);
1584         amdgpu_ttm_global_fini(adev);
1585         adev->mman.initialized = false;
1586         DRM_INFO("amdgpu: ttm finalized\n");
1587 }
1588
1589 /**
1590  * amdgpu_ttm_set_buffer_funcs_status - enable/disable use of buffer functions
1591  *
1592  * @adev: amdgpu_device pointer
1593  * @enable: true when we can use buffer functions.
1594  *
1595  * Enable/disable use of buffer functions during suspend/resume. This should
1596  * only be called at bootup or when userspace isn't running.
1597  */
1598 void amdgpu_ttm_set_buffer_funcs_status(struct amdgpu_device *adev, bool enable)
1599 {
1600         struct ttm_mem_type_manager *man = &adev->mman.bdev.man[TTM_PL_VRAM];
1601         uint64_t size;
1602
1603         if (!adev->mman.initialized || adev->in_gpu_reset)
1604                 return;
1605
1606         /* this just adjusts TTM size idea, which sets lpfn to the correct value */
1607         if (enable)
1608                 size = adev->gmc.real_vram_size;
1609         else
1610                 size = adev->gmc.visible_vram_size;
1611         man->size = size >> PAGE_SHIFT;
1612         adev->mman.buffer_funcs_enabled = enable;
1613 }
1614
1615 int amdgpu_mmap(struct file *filp, struct vm_area_struct *vma)
1616 {
1617         struct drm_file *file_priv;
1618         struct amdgpu_device *adev;
1619
1620         if (unlikely(vma->vm_pgoff < DRM_FILE_PAGE_OFFSET))
1621                 return -EINVAL;
1622
1623         file_priv = filp->private_data;
1624         adev = file_priv->minor->dev->dev_private;
1625         if (adev == NULL)
1626                 return -EINVAL;
1627
1628         return ttm_bo_mmap(filp, vma, &adev->mman.bdev);
1629 }
1630
1631 static int amdgpu_map_buffer(struct ttm_buffer_object *bo,
1632                              struct ttm_mem_reg *mem, unsigned num_pages,
1633                              uint64_t offset, unsigned window,
1634                              struct amdgpu_ring *ring,
1635                              uint64_t *addr)
1636 {
1637         struct amdgpu_ttm_tt *gtt = (void *)bo->ttm;
1638         struct amdgpu_device *adev = ring->adev;
1639         struct ttm_tt *ttm = bo->ttm;
1640         struct amdgpu_job *job;
1641         unsigned num_dw, num_bytes;
1642         dma_addr_t *dma_address;
1643         struct dma_fence *fence;
1644         uint64_t src_addr, dst_addr;
1645         uint64_t flags;
1646         int r;
1647
1648         BUG_ON(adev->mman.buffer_funcs->copy_max_bytes <
1649                AMDGPU_GTT_MAX_TRANSFER_SIZE * 8);
1650
1651         *addr = adev->gmc.gart_start;
1652         *addr += (u64)window * AMDGPU_GTT_MAX_TRANSFER_SIZE *
1653                 AMDGPU_GPU_PAGE_SIZE;
1654
1655         num_dw = adev->mman.buffer_funcs->copy_num_dw;
1656         while (num_dw & 0x7)
1657                 num_dw++;
1658
1659         num_bytes = num_pages * 8;
1660
1661         r = amdgpu_job_alloc_with_ib(adev, num_dw * 4 + num_bytes, &job);
1662         if (r)
1663                 return r;
1664
1665         src_addr = num_dw * 4;
1666         src_addr += job->ibs[0].gpu_addr;
1667
1668         dst_addr = adev->gart.table_addr;
1669         dst_addr += window * AMDGPU_GTT_MAX_TRANSFER_SIZE * 8;
1670         amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_addr,
1671                                 dst_addr, num_bytes);
1672
1673         amdgpu_ring_pad_ib(ring, &job->ibs[0]);
1674         WARN_ON(job->ibs[0].length_dw > num_dw);
1675
1676         dma_address = &gtt->ttm.dma_address[offset >> PAGE_SHIFT];
1677         flags = amdgpu_ttm_tt_pte_flags(adev, ttm, mem);
1678         r = amdgpu_gart_map(adev, 0, num_pages, dma_address, flags,
1679                             &job->ibs[0].ptr[num_dw]);
1680         if (r)
1681                 goto error_free;
1682
1683         r = amdgpu_job_submit(job, ring, &adev->mman.entity,
1684                               AMDGPU_FENCE_OWNER_UNDEFINED, &fence);
1685         if (r)
1686                 goto error_free;
1687
1688         dma_fence_put(fence);
1689
1690         return r;
1691
1692 error_free:
1693         amdgpu_job_free(job);
1694         return r;
1695 }
1696
1697 int amdgpu_copy_buffer(struct amdgpu_ring *ring, uint64_t src_offset,
1698                        uint64_t dst_offset, uint32_t byte_count,
1699                        struct reservation_object *resv,
1700                        struct dma_fence **fence, bool direct_submit,
1701                        bool vm_needs_flush)
1702 {
1703         struct amdgpu_device *adev = ring->adev;
1704         struct amdgpu_job *job;
1705
1706         uint32_t max_bytes;
1707         unsigned num_loops, num_dw;
1708         unsigned i;
1709         int r;
1710
1711         if (direct_submit && !ring->ready) {
1712                 DRM_ERROR("Trying to move memory with ring turned off.\n");
1713                 return -EINVAL;
1714         }
1715
1716         max_bytes = adev->mman.buffer_funcs->copy_max_bytes;
1717         num_loops = DIV_ROUND_UP(byte_count, max_bytes);
1718         num_dw = num_loops * adev->mman.buffer_funcs->copy_num_dw;
1719
1720         /* for IB padding */
1721         while (num_dw & 0x7)
1722                 num_dw++;
1723
1724         r = amdgpu_job_alloc_with_ib(adev, num_dw * 4, &job);
1725         if (r)
1726                 return r;
1727
1728         job->vm_needs_flush = vm_needs_flush;
1729         if (resv) {
1730                 r = amdgpu_sync_resv(adev, &job->sync, resv,
1731                                      AMDGPU_FENCE_OWNER_UNDEFINED,
1732                                      false);
1733                 if (r) {
1734                         DRM_ERROR("sync failed (%d).\n", r);
1735                         goto error_free;
1736                 }
1737         }
1738
1739         for (i = 0; i < num_loops; i++) {
1740                 uint32_t cur_size_in_bytes = min(byte_count, max_bytes);
1741
1742                 amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_offset,
1743                                         dst_offset, cur_size_in_bytes);
1744
1745                 src_offset += cur_size_in_bytes;
1746                 dst_offset += cur_size_in_bytes;
1747                 byte_count -= cur_size_in_bytes;
1748         }
1749
1750         amdgpu_ring_pad_ib(ring, &job->ibs[0]);
1751         WARN_ON(job->ibs[0].length_dw > num_dw);
1752         if (direct_submit) {
1753                 r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs,
1754                                        NULL, fence);
1755                 job->fence = dma_fence_get(*fence);
1756                 if (r)
1757                         DRM_ERROR("Error scheduling IBs (%d)\n", r);
1758                 amdgpu_job_free(job);
1759         } else {
1760                 r = amdgpu_job_submit(job, ring, &adev->mman.entity,
1761                                       AMDGPU_FENCE_OWNER_UNDEFINED, fence);
1762                 if (r)
1763                         goto error_free;
1764         }
1765
1766         return r;
1767
1768 error_free:
1769         amdgpu_job_free(job);
1770         return r;
1771 }
1772
1773 int amdgpu_fill_buffer(struct amdgpu_bo *bo,
1774                        uint32_t src_data,
1775                        struct reservation_object *resv,
1776                        struct dma_fence **fence)
1777 {
1778         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1779         uint32_t max_bytes = adev->mman.buffer_funcs->fill_max_bytes;
1780         struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
1781
1782         struct drm_mm_node *mm_node;
1783         unsigned long num_pages;
1784         unsigned int num_loops, num_dw;
1785
1786         struct amdgpu_job *job;
1787         int r;
1788
1789         if (!adev->mman.buffer_funcs_enabled) {
1790                 DRM_ERROR("Trying to clear memory with ring turned off.\n");
1791                 return -EINVAL;
1792         }
1793
1794         if (bo->tbo.mem.mem_type == TTM_PL_TT) {
1795                 r = amdgpu_ttm_alloc_gart(&bo->tbo);
1796                 if (r)
1797                         return r;
1798         }
1799
1800         num_pages = bo->tbo.num_pages;
1801         mm_node = bo->tbo.mem.mm_node;
1802         num_loops = 0;
1803         while (num_pages) {
1804                 uint32_t byte_count = mm_node->size << PAGE_SHIFT;
1805
1806                 num_loops += DIV_ROUND_UP(byte_count, max_bytes);
1807                 num_pages -= mm_node->size;
1808                 ++mm_node;
1809         }
1810         num_dw = num_loops * adev->mman.buffer_funcs->fill_num_dw;
1811
1812         /* for IB padding */
1813         num_dw += 64;
1814
1815         r = amdgpu_job_alloc_with_ib(adev, num_dw * 4, &job);
1816         if (r)
1817                 return r;
1818
1819         if (resv) {
1820                 r = amdgpu_sync_resv(adev, &job->sync, resv,
1821                                      AMDGPU_FENCE_OWNER_UNDEFINED, false);
1822                 if (r) {
1823                         DRM_ERROR("sync failed (%d).\n", r);
1824                         goto error_free;
1825                 }
1826         }
1827
1828         num_pages = bo->tbo.num_pages;
1829         mm_node = bo->tbo.mem.mm_node;
1830
1831         while (num_pages) {
1832                 uint32_t byte_count = mm_node->size << PAGE_SHIFT;
1833                 uint64_t dst_addr;
1834
1835                 dst_addr = amdgpu_mm_node_addr(&bo->tbo, mm_node, &bo->tbo.mem);
1836                 while (byte_count) {
1837                         uint32_t cur_size_in_bytes = min(byte_count, max_bytes);
1838
1839                         amdgpu_emit_fill_buffer(adev, &job->ibs[0], src_data,
1840                                                 dst_addr, cur_size_in_bytes);
1841
1842                         dst_addr += cur_size_in_bytes;
1843                         byte_count -= cur_size_in_bytes;
1844                 }
1845
1846                 num_pages -= mm_node->size;
1847                 ++mm_node;
1848         }
1849
1850         amdgpu_ring_pad_ib(ring, &job->ibs[0]);
1851         WARN_ON(job->ibs[0].length_dw > num_dw);
1852         r = amdgpu_job_submit(job, ring, &adev->mman.entity,
1853                               AMDGPU_FENCE_OWNER_UNDEFINED, fence);
1854         if (r)
1855                 goto error_free;
1856
1857         return 0;
1858
1859 error_free:
1860         amdgpu_job_free(job);
1861         return r;
1862 }
1863
1864 #if defined(CONFIG_DEBUG_FS)
1865
1866 static int amdgpu_mm_dump_table(struct seq_file *m, void *data)
1867 {
1868         struct drm_info_node *node = (struct drm_info_node *)m->private;
1869         unsigned ttm_pl = *(int *)node->info_ent->data;
1870         struct drm_device *dev = node->minor->dev;
1871         struct amdgpu_device *adev = dev->dev_private;
1872         struct ttm_mem_type_manager *man = &adev->mman.bdev.man[ttm_pl];
1873         struct drm_printer p = drm_seq_file_printer(m);
1874
1875         man->func->debug(man, &p);
1876         return 0;
1877 }
1878
1879 static int ttm_pl_vram = TTM_PL_VRAM;
1880 static int ttm_pl_tt = TTM_PL_TT;
1881
1882 static const struct drm_info_list amdgpu_ttm_debugfs_list[] = {
1883         {"amdgpu_vram_mm", amdgpu_mm_dump_table, 0, &ttm_pl_vram},
1884         {"amdgpu_gtt_mm", amdgpu_mm_dump_table, 0, &ttm_pl_tt},
1885         {"ttm_page_pool", ttm_page_alloc_debugfs, 0, NULL},
1886 #ifdef CONFIG_SWIOTLB
1887         {"ttm_dma_page_pool", ttm_dma_page_alloc_debugfs, 0, NULL}
1888 #endif
1889 };
1890
1891 static ssize_t amdgpu_ttm_vram_read(struct file *f, char __user *buf,
1892                                     size_t size, loff_t *pos)
1893 {
1894         struct amdgpu_device *adev = file_inode(f)->i_private;
1895         ssize_t result = 0;
1896         int r;
1897
1898         if (size & 0x3 || *pos & 0x3)
1899                 return -EINVAL;
1900
1901         if (*pos >= adev->gmc.mc_vram_size)
1902                 return -ENXIO;
1903
1904         while (size) {
1905                 unsigned long flags;
1906                 uint32_t value;
1907
1908                 if (*pos >= adev->gmc.mc_vram_size)
1909                         return result;
1910
1911                 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
1912                 WREG32_NO_KIQ(mmMM_INDEX, ((uint32_t)*pos) | 0x80000000);
1913                 WREG32_NO_KIQ(mmMM_INDEX_HI, *pos >> 31);
1914                 value = RREG32_NO_KIQ(mmMM_DATA);
1915                 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
1916
1917                 r = put_user(value, (uint32_t *)buf);
1918                 if (r)
1919                         return r;
1920
1921                 result += 4;
1922                 buf += 4;
1923                 *pos += 4;
1924                 size -= 4;
1925         }
1926
1927         return result;
1928 }
1929
1930 static ssize_t amdgpu_ttm_vram_write(struct file *f, const char __user *buf,
1931                                     size_t size, loff_t *pos)
1932 {
1933         struct amdgpu_device *adev = file_inode(f)->i_private;
1934         ssize_t result = 0;
1935         int r;
1936
1937         if (size & 0x3 || *pos & 0x3)
1938                 return -EINVAL;
1939
1940         if (*pos >= adev->gmc.mc_vram_size)
1941                 return -ENXIO;
1942
1943         while (size) {
1944                 unsigned long flags;
1945                 uint32_t value;
1946
1947                 if (*pos >= adev->gmc.mc_vram_size)
1948                         return result;
1949
1950                 r = get_user(value, (uint32_t *)buf);
1951                 if (r)
1952                         return r;
1953
1954                 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
1955                 WREG32_NO_KIQ(mmMM_INDEX, ((uint32_t)*pos) | 0x80000000);
1956                 WREG32_NO_KIQ(mmMM_INDEX_HI, *pos >> 31);
1957                 WREG32_NO_KIQ(mmMM_DATA, value);
1958                 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
1959
1960                 result += 4;
1961                 buf += 4;
1962                 *pos += 4;
1963                 size -= 4;
1964         }
1965
1966         return result;
1967 }
1968
1969 static const struct file_operations amdgpu_ttm_vram_fops = {
1970         .owner = THIS_MODULE,
1971         .read = amdgpu_ttm_vram_read,
1972         .write = amdgpu_ttm_vram_write,
1973         .llseek = default_llseek,
1974 };
1975
1976 #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS
1977
1978 static ssize_t amdgpu_ttm_gtt_read(struct file *f, char __user *buf,
1979                                    size_t size, loff_t *pos)
1980 {
1981         struct amdgpu_device *adev = file_inode(f)->i_private;
1982         ssize_t result = 0;
1983         int r;
1984
1985         while (size) {
1986                 loff_t p = *pos / PAGE_SIZE;
1987                 unsigned off = *pos & ~PAGE_MASK;
1988                 size_t cur_size = min_t(size_t, size, PAGE_SIZE - off);
1989                 struct page *page;
1990                 void *ptr;
1991
1992                 if (p >= adev->gart.num_cpu_pages)
1993                         return result;
1994
1995                 page = adev->gart.pages[p];
1996                 if (page) {
1997                         ptr = kmap(page);
1998                         ptr += off;
1999
2000                         r = copy_to_user(buf, ptr, cur_size);
2001                         kunmap(adev->gart.pages[p]);
2002                 } else
2003                         r = clear_user(buf, cur_size);
2004
2005                 if (r)
2006                         return -EFAULT;
2007
2008                 result += cur_size;
2009                 buf += cur_size;
2010                 *pos += cur_size;
2011                 size -= cur_size;
2012         }
2013
2014         return result;
2015 }
2016
2017 static const struct file_operations amdgpu_ttm_gtt_fops = {
2018         .owner = THIS_MODULE,
2019         .read = amdgpu_ttm_gtt_read,
2020         .llseek = default_llseek
2021 };
2022
2023 #endif
2024
2025 static ssize_t amdgpu_iomem_read(struct file *f, char __user *buf,
2026                                  size_t size, loff_t *pos)
2027 {
2028         struct amdgpu_device *adev = file_inode(f)->i_private;
2029         struct iommu_domain *dom;
2030         ssize_t result = 0;
2031         int r;
2032
2033         dom = iommu_get_domain_for_dev(adev->dev);
2034
2035         while (size) {
2036                 phys_addr_t addr = *pos & PAGE_MASK;
2037                 loff_t off = *pos & ~PAGE_MASK;
2038                 size_t bytes = PAGE_SIZE - off;
2039                 unsigned long pfn;
2040                 struct page *p;
2041                 void *ptr;
2042
2043                 bytes = bytes < size ? bytes : size;
2044
2045                 addr = dom ? iommu_iova_to_phys(dom, addr) : addr;
2046
2047                 pfn = addr >> PAGE_SHIFT;
2048                 if (!pfn_valid(pfn))
2049                         return -EPERM;
2050
2051                 p = pfn_to_page(pfn);
2052                 if (p->mapping != adev->mman.bdev.dev_mapping)
2053                         return -EPERM;
2054
2055                 ptr = kmap(p);
2056                 r = copy_to_user(buf, ptr + off, bytes);
2057                 kunmap(p);
2058                 if (r)
2059                         return -EFAULT;
2060
2061                 size -= bytes;
2062                 *pos += bytes;
2063                 result += bytes;
2064         }
2065
2066         return result;
2067 }
2068
2069 static ssize_t amdgpu_iomem_write(struct file *f, const char __user *buf,
2070                                  size_t size, loff_t *pos)
2071 {
2072         struct amdgpu_device *adev = file_inode(f)->i_private;
2073         struct iommu_domain *dom;
2074         ssize_t result = 0;
2075         int r;
2076
2077         dom = iommu_get_domain_for_dev(adev->dev);
2078
2079         while (size) {
2080                 phys_addr_t addr = *pos & PAGE_MASK;
2081                 loff_t off = *pos & ~PAGE_MASK;
2082                 size_t bytes = PAGE_SIZE - off;
2083                 unsigned long pfn;
2084                 struct page *p;
2085                 void *ptr;
2086
2087                 bytes = bytes < size ? bytes : size;
2088
2089                 addr = dom ? iommu_iova_to_phys(dom, addr) : addr;
2090
2091                 pfn = addr >> PAGE_SHIFT;
2092                 if (!pfn_valid(pfn))
2093                         return -EPERM;
2094
2095                 p = pfn_to_page(pfn);
2096                 if (p->mapping != adev->mman.bdev.dev_mapping)
2097                         return -EPERM;
2098
2099                 ptr = kmap(p);
2100                 r = copy_from_user(ptr + off, buf, bytes);
2101                 kunmap(p);
2102                 if (r)
2103                         return -EFAULT;
2104
2105                 size -= bytes;
2106                 *pos += bytes;
2107                 result += bytes;
2108         }
2109
2110         return result;
2111 }
2112
2113 static const struct file_operations amdgpu_ttm_iomem_fops = {
2114         .owner = THIS_MODULE,
2115         .read = amdgpu_iomem_read,
2116         .write = amdgpu_iomem_write,
2117         .llseek = default_llseek
2118 };
2119
2120 static const struct {
2121         char *name;
2122         const struct file_operations *fops;
2123         int domain;
2124 } ttm_debugfs_entries[] = {
2125         { "amdgpu_vram", &amdgpu_ttm_vram_fops, TTM_PL_VRAM },
2126 #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS
2127         { "amdgpu_gtt", &amdgpu_ttm_gtt_fops, TTM_PL_TT },
2128 #endif
2129         { "amdgpu_iomem", &amdgpu_ttm_iomem_fops, TTM_PL_SYSTEM },
2130 };
2131
2132 #endif
2133
2134 static int amdgpu_ttm_debugfs_init(struct amdgpu_device *adev)
2135 {
2136 #if defined(CONFIG_DEBUG_FS)
2137         unsigned count;
2138
2139         struct drm_minor *minor = adev->ddev->primary;
2140         struct dentry *ent, *root = minor->debugfs_root;
2141
2142         for (count = 0; count < ARRAY_SIZE(ttm_debugfs_entries); count++) {
2143                 ent = debugfs_create_file(
2144                                 ttm_debugfs_entries[count].name,
2145                                 S_IFREG | S_IRUGO, root,
2146                                 adev,
2147                                 ttm_debugfs_entries[count].fops);
2148                 if (IS_ERR(ent))
2149                         return PTR_ERR(ent);
2150                 if (ttm_debugfs_entries[count].domain == TTM_PL_VRAM)
2151                         i_size_write(ent->d_inode, adev->gmc.mc_vram_size);
2152                 else if (ttm_debugfs_entries[count].domain == TTM_PL_TT)
2153                         i_size_write(ent->d_inode, adev->gmc.gart_size);
2154                 adev->mman.debugfs_entries[count] = ent;
2155         }
2156
2157         count = ARRAY_SIZE(amdgpu_ttm_debugfs_list);
2158
2159 #ifdef CONFIG_SWIOTLB
2160         if (!(adev->need_swiotlb && swiotlb_nr_tbl()))
2161                 --count;
2162 #endif
2163
2164         return amdgpu_debugfs_add_files(adev, amdgpu_ttm_debugfs_list, count);
2165 #else
2166         return 0;
2167 #endif
2168 }
2169
2170 static void amdgpu_ttm_debugfs_fini(struct amdgpu_device *adev)
2171 {
2172 #if defined(CONFIG_DEBUG_FS)
2173         unsigned i;
2174
2175         for (i = 0; i < ARRAY_SIZE(ttm_debugfs_entries); i++)
2176                 debugfs_remove(adev->mman.debugfs_entries[i]);
2177 #endif
2178 }