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