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

   1 /*
   2  * Copyright 2009 Jerome Glisse.
   3  * All Rights Reserved.
   4  *
   5  * Permission is hereby granted, free of charge, to any person obtaining a
   6  * copy of this software and associated documentation files (the
   7  * "Software"), to deal in the Software without restriction, including
   8  * without limitation the rights to use, copy, modify, merge, publish,
   9  * distribute, sub license, and/or sell copies of the Software, and to
  10  * permit persons to whom the Software is furnished to do so, subject to
  11  * the following conditions:
  12  *
  13  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  14  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  15  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  16  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
  17  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  18  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  19  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  20  *
  21  * The above copyright notice and this permission notice (including the
  22  * next paragraph) shall be included in all copies or substantial portions
  23  * of the Software.
  24  *
  25  */
  26 /*
  27  * Authors:
  28  *    Jerome Glisse <glisse@freedesktop.org>
  29  *    Thomas Hellstrom <thomas-at-tungstengraphics-dot-com>
  30  *    Dave Airlie
  31  */
  32
  33 #include <linux/dma-mapping.h>
  34 #include <linux/iommu.h>
  35 #include <linux/hmm.h>
  36 #include <linux/pagemap.h>
  37 #include <linux/sched/task.h>
  38 #include <linux/sched/mm.h>
  39 #include <linux/seq_file.h>
  40 #include <linux/slab.h>
  41 #include <linux/swap.h>
  42 #include <linux/swiotlb.h>
  43 #include <linux/dma-buf.h>
  44 #include <linux/sizes.h>
  45
  46 #include <drm/ttm/ttm_bo_api.h>
  47 #include <drm/ttm/ttm_bo_driver.h>
  48 #include <drm/ttm/ttm_placement.h>
  49
  50 #include <drm/drm_debugfs.h>
  51 #include <drm/amdgpu_drm.h>
  52
  53 #include "amdgpu.h"
  54 #include "amdgpu_object.h"
  55 #include "amdgpu_trace.h"
  56 #include "amdgpu_amdkfd.h"
  57 #include "amdgpu_sdma.h"
  58 #include "amdgpu_ras.h"
  59 #include "amdgpu_atomfirmware.h"
  60 #include "bif/bif_4_1_d.h"
  61
  62 #define AMDGPU_TTM_VRAM_MAX_DW_READ     (size_t)128
  63
  64 static int amdgpu_ttm_backend_bind(struct ttm_bo_device *bdev,
  65                                    struct ttm_tt *ttm,
  66                                    struct ttm_resource *bo_mem);
  67 static void amdgpu_ttm_backend_unbind(struct ttm_bo_device *bdev,
  68                                       struct ttm_tt *ttm);
  69
  70 static int amdgpu_ttm_init_on_chip(struct amdgpu_device *adev,
  71                                     unsigned int type,
  72                                     uint64_t size_in_page)
  73 {
  74         return ttm_range_man_init(&adev->mman.bdev, type,
  75                                   false, size_in_page);
  76 }
  77
  78 /**
  79  * amdgpu_evict_flags - Compute placement flags
  80  *
  81  * @bo: The buffer object to evict
  82  * @placement: Possible destination(s) for evicted BO
  83  *
  84  * Fill in placement data when ttm_bo_evict() is called
  85  */
  86 static void amdgpu_evict_flags(struct ttm_buffer_object *bo,
  87                                 struct ttm_placement *placement)
  88 {
  89         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
  90         struct amdgpu_bo *abo;
  91         static const struct ttm_place placements = {
  92                 .fpfn = 0,
  93                 .lpfn = 0,
  94                 .mem_type = TTM_PL_SYSTEM,
  95                 .flags = 0
  96         };
  97
  98         /* Don't handle scatter gather BOs */
  99         if (bo->type == ttm_bo_type_sg) {
 100                 placement->num_placement = 0;
 101                 placement->num_busy_placement = 0;
 102                 return;
 103         }
 104
 105         /* Object isn't an AMDGPU object so ignore */
 106         if (!amdgpu_bo_is_amdgpu_bo(bo)) {
 107                 placement->placement = &placements;
 108                 placement->busy_placement = &placements;
 109                 placement->num_placement = 1;
 110                 placement->num_busy_placement = 1;
 111                 return;
 112         }
 113
 114         abo = ttm_to_amdgpu_bo(bo);
 115         switch (bo->mem.mem_type) {
 116         case AMDGPU_PL_GDS:
 117         case AMDGPU_PL_GWS:
 118         case AMDGPU_PL_OA:
 119                 placement->num_placement = 0;
 120                 placement->num_busy_placement = 0;
 121                 return;
 122
 123         case TTM_PL_VRAM:
 124                 if (!adev->mman.buffer_funcs_enabled) {
 125                         /* Move to system memory */
 126                         amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_CPU);
 127                 } else if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
 128                            !(abo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED) &&
 129                            amdgpu_bo_in_cpu_visible_vram(abo)) {
 130
 131                         /* Try evicting to the CPU inaccessible part of VRAM
 132                          * first, but only set GTT as busy placement, so this
 133                          * BO will be evicted to GTT rather than causing other
 134                          * BOs to be evicted from VRAM
 135                          */
 136                         amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM |
 137                                                          AMDGPU_GEM_DOMAIN_GTT);
 138                         abo->placements[0].fpfn = adev->gmc.visible_vram_size >> PAGE_SHIFT;
 139                         abo->placements[0].lpfn = 0;
 140                         abo->placement.busy_placement = &abo->placements[1];
 141                         abo->placement.num_busy_placement = 1;
 142                 } else {
 143                         /* Move to GTT memory */
 144                         amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_GTT);
 145                 }
 146                 break;
 147         case TTM_PL_TT:
 148         default:
 149                 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_CPU);
 150                 break;
 151         }
 152         *placement = abo->placement;
 153 }
 154
 155 /**
 156  * amdgpu_verify_access - Verify access for a mmap call
 157  *
 158  * @bo: The buffer object to map
 159  * @filp: The file pointer from the process performing the mmap
 160  *
 161  * This is called by ttm_bo_mmap() to verify whether a process
 162  * has the right to mmap a BO to their process space.
 163  */
 164 static int amdgpu_verify_access(struct ttm_buffer_object *bo, struct file *filp)
 165 {
 166         struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
 167
 168         /*
 169          * Don't verify access for KFD BOs. They don't have a GEM
 170          * object associated with them.
 171          */
 172         if (abo->kfd_bo)
 173                 return 0;
 174
 175         if (amdgpu_ttm_tt_get_usermm(bo->ttm))
 176                 return -EPERM;
 177         return drm_vma_node_verify_access(&abo->tbo.base.vma_node,
 178                                           filp->private_data);
 179 }
 180
 181 /**
 182  * amdgpu_mm_node_addr - Compute the GPU relative offset of a GTT buffer.
 183  *
 184  * @bo: The bo to assign the memory to.
 185  * @mm_node: Memory manager node for drm allocator.
 186  * @mem: The region where the bo resides.
 187  *
 188  */
 189 static uint64_t amdgpu_mm_node_addr(struct ttm_buffer_object *bo,
 190                                     struct drm_mm_node *mm_node,
 191                                     struct ttm_resource *mem)
 192 {
 193         uint64_t addr = 0;
 194
 195         if (mm_node->start != AMDGPU_BO_INVALID_OFFSET) {
 196                 addr = mm_node->start << PAGE_SHIFT;
 197                 addr += amdgpu_ttm_domain_start(amdgpu_ttm_adev(bo->bdev),
 198                                                 mem->mem_type);
 199         }
 200         return addr;
 201 }
 202
 203 /**
 204  * amdgpu_find_mm_node - Helper function finds the drm_mm_node corresponding to
 205  * @offset. It also modifies the offset to be within the drm_mm_node returned
 206  *
 207  * @mem: The region where the bo resides.
 208  * @offset: The offset that drm_mm_node is used for finding.
 209  *
 210  */
 211 static struct drm_mm_node *amdgpu_find_mm_node(struct ttm_resource *mem,
 212                                                uint64_t *offset)
 213 {
 214         struct drm_mm_node *mm_node = mem->mm_node;
 215
 216         while (*offset >= (mm_node->size << PAGE_SHIFT)) {
 217                 *offset -= (mm_node->size << PAGE_SHIFT);
 218                 ++mm_node;
 219         }
 220         return mm_node;
 221 }
 222
 223 /**
 224  * amdgpu_ttm_map_buffer - Map memory into the GART windows
 225  * @bo: buffer object to map
 226  * @mem: memory object to map
 227  * @mm_node: drm_mm node object to map
 228  * @num_pages: number of pages to map
 229  * @offset: offset into @mm_node where to start
 230  * @window: which GART window to use
 231  * @ring: DMA ring to use for the copy
 232  * @tmz: if we should setup a TMZ enabled mapping
 233  * @addr: resulting address inside the MC address space
 234  *
 235  * Setup one of the GART windows to access a specific piece of memory or return
 236  * the physical address for local memory.
 237  */
 238 static int amdgpu_ttm_map_buffer(struct ttm_buffer_object *bo,
 239                                  struct ttm_resource *mem,
 240                                  struct drm_mm_node *mm_node,
 241                                  unsigned num_pages, uint64_t offset,
 242                                  unsigned window, struct amdgpu_ring *ring,
 243                                  bool tmz, uint64_t *addr)
 244 {
 245         struct amdgpu_device *adev = ring->adev;
 246         struct amdgpu_job *job;
 247         unsigned num_dw, num_bytes;
 248         struct dma_fence *fence;
 249         uint64_t src_addr, dst_addr;
 250         void *cpu_addr;
 251         uint64_t flags;
 252         unsigned int i;
 253         int r;
 254
 255         BUG_ON(adev->mman.buffer_funcs->copy_max_bytes <
 256                AMDGPU_GTT_MAX_TRANSFER_SIZE * 8);
 257
 258         /* Map only what can't be accessed directly */
 259         if (!tmz && mem->start != AMDGPU_BO_INVALID_OFFSET) {
 260                 *addr = amdgpu_mm_node_addr(bo, mm_node, mem) + offset;
 261                 return 0;
 262         }
 263
 264         *addr = adev->gmc.gart_start;
 265         *addr += (u64)window * AMDGPU_GTT_MAX_TRANSFER_SIZE *
 266                 AMDGPU_GPU_PAGE_SIZE;
 267         *addr += offset & ~PAGE_MASK;
 268
 269         num_dw = ALIGN(adev->mman.buffer_funcs->copy_num_dw, 8);
 270         num_bytes = num_pages * 8;
 271
 272         r = amdgpu_job_alloc_with_ib(adev, num_dw * 4 + num_bytes,
 273                                      AMDGPU_IB_POOL_DELAYED, &job);
 274         if (r)
 275                 return r;
 276
 277         src_addr = num_dw * 4;
 278         src_addr += job->ibs[0].gpu_addr;
 279
 280         dst_addr = amdgpu_bo_gpu_offset(adev->gart.bo);
 281         dst_addr += window * AMDGPU_GTT_MAX_TRANSFER_SIZE * 8;
 282         amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_addr,
 283                                 dst_addr, num_bytes, false);
 284
 285         amdgpu_ring_pad_ib(ring, &job->ibs[0]);
 286         WARN_ON(job->ibs[0].length_dw > num_dw);
 287
 288         flags = amdgpu_ttm_tt_pte_flags(adev, bo->ttm, mem);
 289         if (tmz)
 290                 flags |= AMDGPU_PTE_TMZ;
 291
 292         cpu_addr = &job->ibs[0].ptr[num_dw];
 293
 294         if (mem->mem_type == TTM_PL_TT) {
 295                 dma_addr_t *dma_address;
 296
 297                 dma_address = &bo->ttm->dma_address[offset >> PAGE_SHIFT];
 298                 r = amdgpu_gart_map(adev, 0, num_pages, dma_address, flags,
 299                                     cpu_addr);
 300                 if (r)
 301                         goto error_free;
 302         } else {
 303                 dma_addr_t dma_address;
 304
 305                 dma_address = (mm_node->start << PAGE_SHIFT) + offset;
 306                 dma_address += adev->vm_manager.vram_base_offset;
 307
 308                 for (i = 0; i < num_pages; ++i) {
 309                         r = amdgpu_gart_map(adev, i << PAGE_SHIFT, 1,
 310                                             &dma_address, flags, cpu_addr);
 311                         if (r)
 312                                 goto error_free;
 313
 314                         dma_address += PAGE_SIZE;
 315                 }
 316         }
 317
 318         r = amdgpu_job_submit(job, &adev->mman.entity,
 319                               AMDGPU_FENCE_OWNER_UNDEFINED, &fence);
 320         if (r)
 321                 goto error_free;
 322
 323         dma_fence_put(fence);
 324
 325         return r;
 326
 327 error_free:
 328         amdgpu_job_free(job);
 329         return r;
 330 }
 331
 332 /**
 333  * amdgpu_copy_ttm_mem_to_mem - Helper function for copy
 334  * @adev: amdgpu device
 335  * @src: buffer/address where to read from
 336  * @dst: buffer/address where to write to
 337  * @size: number of bytes to copy
 338  * @tmz: if a secure copy should be used
 339  * @resv: resv object to sync to
 340  * @f: Returns the last fence if multiple jobs are submitted.
 341  *
 342  * The function copies @size bytes from {src->mem + src->offset} to
 343  * {dst->mem + dst->offset}. src->bo and dst->bo could be same BO for a
 344  * move and different for a BO to BO copy.
 345  *
 346  */
 347 int amdgpu_ttm_copy_mem_to_mem(struct amdgpu_device *adev,
 348                                const struct amdgpu_copy_mem *src,
 349                                const struct amdgpu_copy_mem *dst,
 350                                uint64_t size, bool tmz,
 351                                struct dma_resv *resv,
 352                                struct dma_fence **f)
 353 {
 354         const uint32_t GTT_MAX_BYTES = (AMDGPU_GTT_MAX_TRANSFER_SIZE *
 355                                         AMDGPU_GPU_PAGE_SIZE);
 356
 357         uint64_t src_node_size, dst_node_size, src_offset, dst_offset;
 358         struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
 359         struct drm_mm_node *src_mm, *dst_mm;
 360         struct dma_fence *fence = NULL;
 361         int r = 0;
 362
 363         if (!adev->mman.buffer_funcs_enabled) {
 364                 DRM_ERROR("Trying to move memory with ring turned off.\n");
 365                 return -EINVAL;
 366         }
 367
 368         src_offset = src->offset;
 369         if (src->mem->mm_node) {
 370                 src_mm = amdgpu_find_mm_node(src->mem, &src_offset);
 371                 src_node_size = (src_mm->size << PAGE_SHIFT) - src_offset;
 372         } else {
 373                 src_mm = NULL;
 374                 src_node_size = ULLONG_MAX;
 375         }
 376
 377         dst_offset = dst->offset;
 378         if (dst->mem->mm_node) {
 379                 dst_mm = amdgpu_find_mm_node(dst->mem, &dst_offset);
 380                 dst_node_size = (dst_mm->size << PAGE_SHIFT) - dst_offset;
 381         } else {
 382                 dst_mm = NULL;
 383                 dst_node_size = ULLONG_MAX;
 384         }
 385
 386         mutex_lock(&adev->mman.gtt_window_lock);
 387
 388         while (size) {
 389                 uint32_t src_page_offset = src_offset & ~PAGE_MASK;
 390                 uint32_t dst_page_offset = dst_offset & ~PAGE_MASK;
 391                 struct dma_fence *next;
 392                 uint32_t cur_size;
 393                 uint64_t from, to;
 394
 395                 /* Copy size cannot exceed GTT_MAX_BYTES. So if src or dst
 396                  * begins at an offset, then adjust the size accordingly
 397                  */
 398                 cur_size = max(src_page_offset, dst_page_offset);
 399                 cur_size = min(min3(src_node_size, dst_node_size, size),
 400                                (uint64_t)(GTT_MAX_BYTES - cur_size));
 401
 402                 /* Map src to window 0 and dst to window 1. */
 403                 r = amdgpu_ttm_map_buffer(src->bo, src->mem, src_mm,
 404                                           PFN_UP(cur_size + src_page_offset),
 405                                           src_offset, 0, ring, tmz, &from);
 406                 if (r)
 407                         goto error;
 408
 409                 r = amdgpu_ttm_map_buffer(dst->bo, dst->mem, dst_mm,
 410                                           PFN_UP(cur_size + dst_page_offset),
 411                                           dst_offset, 1, ring, tmz, &to);
 412                 if (r)
 413                         goto error;
 414
 415                 r = amdgpu_copy_buffer(ring, from, to, cur_size,
 416                                        resv, &next, false, true, tmz);
 417                 if (r)
 418                         goto error;
 419
 420                 dma_fence_put(fence);
 421                 fence = next;
 422
 423                 size -= cur_size;
 424                 if (!size)
 425                         break;
 426
 427                 src_node_size -= cur_size;
 428                 if (!src_node_size) {
 429                         ++src_mm;
 430                         src_node_size = src_mm->size << PAGE_SHIFT;
 431                         src_offset = 0;
 432                 } else {
 433                         src_offset += cur_size;
 434                 }
 435
 436                 dst_node_size -= cur_size;
 437                 if (!dst_node_size) {
 438                         ++dst_mm;
 439                         dst_node_size = dst_mm->size << PAGE_SHIFT;
 440                         dst_offset = 0;
 441                 } else {
 442                         dst_offset += cur_size;
 443                 }
 444         }
 445 error:
 446         mutex_unlock(&adev->mman.gtt_window_lock);
 447         if (f)
 448                 *f = dma_fence_get(fence);
 449         dma_fence_put(fence);
 450         return r;
 451 }
 452
 453 /*
 454  * amdgpu_move_blit - Copy an entire buffer to another buffer
 455  *
 456  * This is a helper called by amdgpu_bo_move() and amdgpu_move_vram_ram() to
 457  * help move buffers to and from VRAM.
 458  */
 459 static int amdgpu_move_blit(struct ttm_buffer_object *bo,
 460                             bool evict,
 461                             struct ttm_resource *new_mem,
 462                             struct ttm_resource *old_mem)
 463 {
 464         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
 465         struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
 466         struct amdgpu_copy_mem src, dst;
 467         struct dma_fence *fence = NULL;
 468         int r;
 469
 470         src.bo = bo;
 471         dst.bo = bo;
 472         src.mem = old_mem;
 473         dst.mem = new_mem;
 474         src.offset = 0;
 475         dst.offset = 0;
 476
 477         r = amdgpu_ttm_copy_mem_to_mem(adev, &src, &dst,
 478                                        new_mem->num_pages << PAGE_SHIFT,
 479                                        amdgpu_bo_encrypted(abo),
 480                                        bo->base.resv, &fence);
 481         if (r)
 482                 goto error;
 483
 484         /* clear the space being freed */
 485         if (old_mem->mem_type == TTM_PL_VRAM &&
 486             (abo->flags & AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE)) {
 487                 struct dma_fence *wipe_fence = NULL;
 488
 489                 r = amdgpu_fill_buffer(ttm_to_amdgpu_bo(bo), AMDGPU_POISON,
 490                                        NULL, &wipe_fence);
 491                 if (r) {
 492                         goto error;
 493                 } else if (wipe_fence) {
 494                         dma_fence_put(fence);
 495                         fence = wipe_fence;
 496                 }
 497         }
 498
 499         /* Always block for VM page tables before committing the new location */
 500         if (bo->type == ttm_bo_type_kernel)
 501                 r = ttm_bo_move_accel_cleanup(bo, fence, true, false, new_mem);
 502         else
 503                 r = ttm_bo_move_accel_cleanup(bo, fence, evict, true, new_mem);
 504         dma_fence_put(fence);
 505         return r;
 506
 507 error:
 508         if (fence)
 509                 dma_fence_wait(fence, false);
 510         dma_fence_put(fence);
 511         return r;
 512 }
 513
 514 /*
 515  * amdgpu_mem_visible - Check that memory can be accessed by ttm_bo_move_memcpy
 516  *
 517  * Called by amdgpu_bo_move()
 518  */
 519 static bool amdgpu_mem_visible(struct amdgpu_device *adev,
 520                                struct ttm_resource *mem)
 521 {
 522         struct drm_mm_node *nodes = mem->mm_node;
 523
 524         if (mem->mem_type == TTM_PL_SYSTEM ||
 525             mem->mem_type == TTM_PL_TT)
 526                 return true;
 527         if (mem->mem_type != TTM_PL_VRAM)
 528                 return false;
 529
 530         /* ttm_resource_ioremap only supports contiguous memory */
 531         if (nodes->size != mem->num_pages)
 532                 return false;
 533
 534         return ((nodes->start + nodes->size) << PAGE_SHIFT)
 535                 <= adev->gmc.visible_vram_size;
 536 }
 537
 538 /*
 539  * amdgpu_bo_move - Move a buffer object to a new memory location
 540  *
 541  * Called by ttm_bo_handle_move_mem()
 542  */
 543 static int amdgpu_bo_move(struct ttm_buffer_object *bo, bool evict,
 544                           struct ttm_operation_ctx *ctx,
 545                           struct ttm_resource *new_mem,
 546                           struct ttm_place *hop)
 547 {
 548         struct amdgpu_device *adev;
 549         struct amdgpu_bo *abo;
 550         struct ttm_resource *old_mem = &bo->mem;
 551         int r;
 552
 553         if (new_mem->mem_type == TTM_PL_TT) {
 554                 r = amdgpu_ttm_backend_bind(bo->bdev, bo->ttm, new_mem);
 555                 if (r)
 556                         return r;
 557         }
 558
 559         /* Can't move a pinned BO */
 560         abo = ttm_to_amdgpu_bo(bo);
 561         if (WARN_ON_ONCE(abo->tbo.pin_count > 0))
 562                 return -EINVAL;
 563
 564         adev = amdgpu_ttm_adev(bo->bdev);
 565
 566         if (old_mem->mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
 567                 ttm_bo_move_null(bo, new_mem);
 568                 goto out;
 569         }
 570         if (old_mem->mem_type == TTM_PL_SYSTEM &&
 571             new_mem->mem_type == TTM_PL_TT) {
 572                 ttm_bo_move_null(bo, new_mem);
 573                 goto out;
 574         }
 575         if (old_mem->mem_type == TTM_PL_TT &&
 576             new_mem->mem_type == TTM_PL_SYSTEM) {
 577                 r = ttm_bo_wait_ctx(bo, ctx);
 578                 if (r)
 579                         return r;
 580
 581                 amdgpu_ttm_backend_unbind(bo->bdev, bo->ttm);
 582                 ttm_resource_free(bo, &bo->mem);
 583                 ttm_bo_assign_mem(bo, new_mem);
 584                 goto out;
 585         }
 586
 587         if (old_mem->mem_type == AMDGPU_PL_GDS ||
 588             old_mem->mem_type == AMDGPU_PL_GWS ||
 589             old_mem->mem_type == AMDGPU_PL_OA ||
 590             new_mem->mem_type == AMDGPU_PL_GDS ||
 591             new_mem->mem_type == AMDGPU_PL_GWS ||
 592             new_mem->mem_type == AMDGPU_PL_OA) {
 593                 /* Nothing to save here */
 594                 ttm_bo_move_null(bo, new_mem);
 595                 goto out;
 596         }
 597
 598         if (adev->mman.buffer_funcs_enabled) {
 599                 if (((old_mem->mem_type == TTM_PL_SYSTEM &&
 600                       new_mem->mem_type == TTM_PL_VRAM) ||
 601                      (old_mem->mem_type == TTM_PL_VRAM &&
 602                       new_mem->mem_type == TTM_PL_SYSTEM))) {
 603                         hop->fpfn = 0;
 604                         hop->lpfn = 0;
 605                         hop->mem_type = TTM_PL_TT;
 606                         hop->flags = 0;
 607                         return -EMULTIHOP;
 608                 }
 609
 610                 r = amdgpu_move_blit(bo, evict, new_mem, old_mem);
 611         } else {
 612                 r = -ENODEV;
 613         }
 614
 615         if (r) {
 616                 /* Check that all memory is CPU accessible */
 617                 if (!amdgpu_mem_visible(adev, old_mem) ||
 618                     !amdgpu_mem_visible(adev, new_mem)) {
 619                         pr_err("Move buffer fallback to memcpy unavailable\n");
 620                         return r;
 621                 }
 622
 623                 r = ttm_bo_move_memcpy(bo, ctx, new_mem);
 624                 if (r)
 625                         return r;
 626         }
 627
 628         if (bo->type == ttm_bo_type_device &&
 629             new_mem->mem_type == TTM_PL_VRAM &&
 630             old_mem->mem_type != TTM_PL_VRAM) {
 631                 /* amdgpu_bo_fault_reserve_notify will re-set this if the CPU
 632                  * accesses the BO after it's moved.
 633                  */
 634                 abo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
 635         }
 636
 637 out:
 638         /* update statistics */
 639         atomic64_add(bo->base.size, &adev->num_bytes_moved);
 640         amdgpu_bo_move_notify(bo, evict, new_mem);
 641         return 0;
 642 }
 643
 644 /*
 645  * amdgpu_ttm_io_mem_reserve - Reserve a block of memory during a fault
 646  *
 647  * Called by ttm_mem_io_reserve() ultimately via ttm_bo_vm_fault()
 648  */
 649 static int amdgpu_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_resource *mem)
 650 {
 651         struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
 652         struct drm_mm_node *mm_node = mem->mm_node;
 653         size_t bus_size = (size_t)mem->num_pages << PAGE_SHIFT;
 654
 655         switch (mem->mem_type) {
 656         case TTM_PL_SYSTEM:
 657                 /* system memory */
 658                 return 0;
 659         case TTM_PL_TT:
 660                 break;
 661         case TTM_PL_VRAM:
 662                 mem->bus.offset = mem->start << PAGE_SHIFT;
 663                 /* check if it's visible */
 664                 if ((mem->bus.offset + bus_size) > adev->gmc.visible_vram_size)
 665                         return -EINVAL;
 666                 /* Only physically contiguous buffers apply. In a contiguous
 667                  * buffer, size of the first mm_node would match the number of
 668                  * pages in ttm_resource.
 669                  */
 670                 if (adev->mman.aper_base_kaddr &&
 671                     (mm_node->size == mem->num_pages))
 672                         mem->bus.addr = (u8 *)adev->mman.aper_base_kaddr +
 673                                         mem->bus.offset;
 674
 675                 mem->bus.offset += adev->gmc.aper_base;
 676                 mem->bus.is_iomem = true;
 677                 mem->bus.caching = ttm_write_combined;
 678                 break;
 679         default:
 680                 return -EINVAL;
 681         }
 682         return 0;
 683 }
 684
 685 static unsigned long amdgpu_ttm_io_mem_pfn(struct ttm_buffer_object *bo,
 686                                            unsigned long page_offset)
 687 {
 688         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
 689         uint64_t offset = (page_offset << PAGE_SHIFT);
 690         struct drm_mm_node *mm;
 691
 692         mm = amdgpu_find_mm_node(&bo->mem, &offset);
 693         offset += adev->gmc.aper_base;
 694         return mm->start + (offset >> PAGE_SHIFT);
 695 }
 696
 697 /**
 698  * amdgpu_ttm_domain_start - Returns GPU start address
 699  * @adev: amdgpu device object
 700  * @type: type of the memory
 701  *
 702  * Returns:
 703  * GPU start address of a memory domain
 704  */
 705
 706 uint64_t amdgpu_ttm_domain_start(struct amdgpu_device *adev, uint32_t type)
 707 {
 708         switch (type) {
 709         case TTM_PL_TT:
 710                 return adev->gmc.gart_start;
 711         case TTM_PL_VRAM:
 712                 return adev->gmc.vram_start;
 713         }
 714
 715         return 0;
 716 }
 717
 718 /*
 719  * TTM backend functions.
 720  */
 721 struct amdgpu_ttm_tt {
 722         struct ttm_tt   ttm;
 723         struct drm_gem_object   *gobj;
 724         u64                     offset;
 725         uint64_t                userptr;
 726         struct task_struct      *usertask;
 727         uint32_t                userflags;
 728         bool                    bound;
 729 #if IS_ENABLED(CONFIG_DRM_AMDGPU_USERPTR)
 730         struct hmm_range        *range;
 731 #endif
 732 };
 733
 734 #ifdef CONFIG_DRM_AMDGPU_USERPTR
 735 /*
 736  * amdgpu_ttm_tt_get_user_pages - get device accessible pages that back user
 737  * memory and start HMM tracking CPU page table update
 738  *
 739  * Calling function must call amdgpu_ttm_tt_userptr_range_done() once and only
 740  * once afterwards to stop HMM tracking
 741  */
 742 int amdgpu_ttm_tt_get_user_pages(struct amdgpu_bo *bo, struct page **pages)
 743 {
 744         struct ttm_tt *ttm = bo->tbo.ttm;
 745         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 746         unsigned long start = gtt->userptr;
 747         struct vm_area_struct *vma;
 748         struct hmm_range *range;
 749         unsigned long timeout;
 750         struct mm_struct *mm;
 751         unsigned long i;
 752         int r = 0;
 753
 754         mm = bo->notifier.mm;
 755         if (unlikely(!mm)) {
 756                 DRM_DEBUG_DRIVER("BO is not registered?\n");
 757                 return -EFAULT;
 758         }
 759
 760         /* Another get_user_pages is running at the same time?? */
 761         if (WARN_ON(gtt->range))
 762                 return -EFAULT;
 763
 764         if (!mmget_not_zero(mm)) /* Happens during process shutdown */
 765                 return -ESRCH;
 766
 767         range = kzalloc(sizeof(*range), GFP_KERNEL);
 768         if (unlikely(!range)) {
 769                 r = -ENOMEM;
 770                 goto out;
 771         }
 772         range->notifier = &bo->notifier;
 773         range->start = bo->notifier.interval_tree.start;
 774         range->end = bo->notifier.interval_tree.last + 1;
 775         range->default_flags = HMM_PFN_REQ_FAULT;
 776         if (!amdgpu_ttm_tt_is_readonly(ttm))
 777                 range->default_flags |= HMM_PFN_REQ_WRITE;
 778
 779         range->hmm_pfns = kvmalloc_array(ttm->num_pages,
 780                                          sizeof(*range->hmm_pfns), GFP_KERNEL);
 781         if (unlikely(!range->hmm_pfns)) {
 782                 r = -ENOMEM;
 783                 goto out_free_ranges;
 784         }
 785
 786         mmap_read_lock(mm);
 787         vma = find_vma(mm, start);
 788         if (unlikely(!vma || start < vma->vm_start)) {
 789                 r = -EFAULT;
 790                 goto out_unlock;
 791         }
 792         if (unlikely((gtt->userflags & AMDGPU_GEM_USERPTR_ANONONLY) &&
 793                 vma->vm_file)) {
 794                 r = -EPERM;
 795                 goto out_unlock;
 796         }
 797         mmap_read_unlock(mm);
 798         timeout = jiffies + msecs_to_jiffies(HMM_RANGE_DEFAULT_TIMEOUT);
 799
 800 retry:
 801         range->notifier_seq = mmu_interval_read_begin(&bo->notifier);
 802
 803         mmap_read_lock(mm);
 804         r = hmm_range_fault(range);
 805         mmap_read_unlock(mm);
 806         if (unlikely(r)) {
 807                 /*
 808                  * FIXME: This timeout should encompass the retry from
 809                  * mmu_interval_read_retry() as well.
 810                  */
 811                 if (r == -EBUSY && !time_after(jiffies, timeout))
 812                         goto retry;
 813                 goto out_free_pfns;
 814         }
 815
 816         /*
 817          * Due to default_flags, all pages are HMM_PFN_VALID or
 818          * hmm_range_fault() fails. FIXME: The pages cannot be touched outside
 819          * the notifier_lock, and mmu_interval_read_retry() must be done first.
 820          */
 821         for (i = 0; i < ttm->num_pages; i++)
 822                 pages[i] = hmm_pfn_to_page(range->hmm_pfns[i]);
 823
 824         gtt->range = range;
 825         mmput(mm);
 826
 827         return 0;
 828
 829 out_unlock:
 830         mmap_read_unlock(mm);
 831 out_free_pfns:
 832         kvfree(range->hmm_pfns);
 833 out_free_ranges:
 834         kfree(range);
 835 out:
 836         mmput(mm);
 837         return r;
 838 }
 839
 840 /*
 841  * amdgpu_ttm_tt_userptr_range_done - stop HMM track the CPU page table change
 842  * Check if the pages backing this ttm range have been invalidated
 843  *
 844  * Returns: true if pages are still valid
 845  */
 846 bool amdgpu_ttm_tt_get_user_pages_done(struct ttm_tt *ttm)
 847 {
 848         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 849         bool r = false;
 850
 851         if (!gtt || !gtt->userptr)
 852                 return false;
 853
 854         DRM_DEBUG_DRIVER("user_pages_done 0x%llx pages 0x%x\n",
 855                 gtt->userptr, ttm->num_pages);
 856
 857         WARN_ONCE(!gtt->range || !gtt->range->hmm_pfns,
 858                 "No user pages to check\n");
 859
 860         if (gtt->range) {
 861                 /*
 862                  * FIXME: Must always hold notifier_lock for this, and must
 863                  * not ignore the return code.
 864                  */
 865                 r = mmu_interval_read_retry(gtt->range->notifier,
 866                                          gtt->range->notifier_seq);
 867                 kvfree(gtt->range->hmm_pfns);
 868                 kfree(gtt->range);
 869                 gtt->range = NULL;
 870         }
 871
 872         return !r;
 873 }
 874 #endif
 875
 876 /*
 877  * amdgpu_ttm_tt_set_user_pages - Copy pages in, putting old pages as necessary.
 878  *
 879  * Called by amdgpu_cs_list_validate(). This creates the page list
 880  * that backs user memory and will ultimately be mapped into the device
 881  * address space.
 882  */
 883 void amdgpu_ttm_tt_set_user_pages(struct ttm_tt *ttm, struct page **pages)
 884 {
 885         unsigned long i;
 886
 887         for (i = 0; i < ttm->num_pages; ++i)
 888                 ttm->pages[i] = pages ? pages[i] : NULL;
 889 }
 890
 891 /*
 892  * amdgpu_ttm_tt_pin_userptr - prepare the sg table with the user pages
 893  *
 894  * Called by amdgpu_ttm_backend_bind()
 895  **/
 896 static int amdgpu_ttm_tt_pin_userptr(struct ttm_bo_device *bdev,
 897                                      struct ttm_tt *ttm)
 898 {
 899         struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
 900         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 901         int r;
 902
 903         int write = !(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
 904         enum dma_data_direction direction = write ?
 905                 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
 906
 907         /* Allocate an SG array and squash pages into it */
 908         r = sg_alloc_table_from_pages(ttm->sg, ttm->pages, ttm->num_pages, 0,
 909                                       ttm->num_pages << PAGE_SHIFT,
 910                                       GFP_KERNEL);
 911         if (r)
 912                 goto release_sg;
 913
 914         /* Map SG to device */
 915         r = dma_map_sgtable(adev->dev, ttm->sg, direction, 0);
 916         if (r)
 917                 goto release_sg;
 918
 919         /* convert SG to linear array of pages and dma addresses */
 920         drm_prime_sg_to_dma_addr_array(ttm->sg, gtt->ttm.dma_address,
 921                                        ttm->num_pages);
 922
 923         return 0;
 924
 925 release_sg:
 926         kfree(ttm->sg);
 927         ttm->sg = NULL;
 928         return r;
 929 }
 930
 931 /*
 932  * amdgpu_ttm_tt_unpin_userptr - Unpin and unmap userptr pages
 933  */
 934 static void amdgpu_ttm_tt_unpin_userptr(struct ttm_bo_device *bdev,
 935                                         struct ttm_tt *ttm)
 936 {
 937         struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
 938         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 939
 940         int write = !(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
 941         enum dma_data_direction direction = write ?
 942                 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
 943
 944         /* double check that we don't free the table twice */
 945         if (!ttm->sg->sgl)
 946                 return;
 947
 948         /* unmap the pages mapped to the device */
 949         dma_unmap_sgtable(adev->dev, ttm->sg, direction, 0);
 950         sg_free_table(ttm->sg);
 951
 952 #if IS_ENABLED(CONFIG_DRM_AMDGPU_USERPTR)
 953         if (gtt->range) {
 954                 unsigned long i;
 955
 956                 for (i = 0; i < ttm->num_pages; i++) {
 957                         if (ttm->pages[i] !=
 958                             hmm_pfn_to_page(gtt->range->hmm_pfns[i]))
 959                                 break;
 960                 }
 961
 962                 WARN((i == ttm->num_pages), "Missing get_user_page_done\n");
 963         }
 964 #endif
 965 }
 966
 967 static int amdgpu_ttm_gart_bind(struct amdgpu_device *adev,
 968                                 struct ttm_buffer_object *tbo,
 969                                 uint64_t flags)
 970 {
 971         struct amdgpu_bo *abo = ttm_to_amdgpu_bo(tbo);
 972         struct ttm_tt *ttm = tbo->ttm;
 973         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 974         int r;
 975
 976         if (amdgpu_bo_encrypted(abo))
 977                 flags |= AMDGPU_PTE_TMZ;
 978
 979         if (abo->flags & AMDGPU_GEM_CREATE_CP_MQD_GFX9) {
 980                 uint64_t page_idx = 1;
 981
 982                 r = amdgpu_gart_bind(adev, gtt->offset, page_idx,
 983                                 ttm->pages, gtt->ttm.dma_address, flags);
 984                 if (r)
 985                         goto gart_bind_fail;
 986
 987                 /* The memory type of the first page defaults to UC. Now
 988                  * modify the memory type to NC from the second page of
 989                  * the BO onward.
 990                  */
 991                 flags &= ~AMDGPU_PTE_MTYPE_VG10_MASK;
 992                 flags |= AMDGPU_PTE_MTYPE_VG10(AMDGPU_MTYPE_NC);
 993
 994                 r = amdgpu_gart_bind(adev,
 995                                 gtt->offset + (page_idx << PAGE_SHIFT),
 996                                 ttm->num_pages - page_idx,
 997                                 &ttm->pages[page_idx],
 998                                 &(gtt->ttm.dma_address[page_idx]), flags);
 999         } else {
1000                 r = amdgpu_gart_bind(adev, gtt->offset, ttm->num_pages,
1001                                      ttm->pages, gtt->ttm.dma_address, flags);
1002         }
1003
1004 gart_bind_fail:
1005         if (r)
1006                 DRM_ERROR("failed to bind %u pages at 0x%08llX\n",
1007                           ttm->num_pages, gtt->offset);
1008
1009         return r;
1010 }
1011
1012 /*
1013  * amdgpu_ttm_backend_bind - Bind GTT memory
1014  *
1015  * Called by ttm_tt_bind() on behalf of ttm_bo_handle_move_mem().
1016  * This handles binding GTT memory to the device address space.
1017  */
1018 static int amdgpu_ttm_backend_bind(struct ttm_bo_device *bdev,
1019                                    struct ttm_tt *ttm,
1020                                    struct ttm_resource *bo_mem)
1021 {
1022         struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
1023         struct amdgpu_ttm_tt *gtt = (void*)ttm;
1024         uint64_t flags;
1025         int r = 0;
1026
1027         if (!bo_mem)
1028                 return -EINVAL;
1029
1030         if (gtt->bound)
1031                 return 0;
1032
1033         if (gtt->userptr) {
1034                 r = amdgpu_ttm_tt_pin_userptr(bdev, ttm);
1035                 if (r) {
1036                         DRM_ERROR("failed to pin userptr\n");
1037                         return r;
1038                 }
1039         }
1040         if (!ttm->num_pages) {
1041                 WARN(1, "nothing to bind %u pages for mreg %p back %p!\n",
1042                      ttm->num_pages, bo_mem, ttm);
1043         }
1044
1045         if (bo_mem->mem_type == AMDGPU_PL_GDS ||
1046             bo_mem->mem_type == AMDGPU_PL_GWS ||
1047             bo_mem->mem_type == AMDGPU_PL_OA)
1048                 return -EINVAL;
1049
1050         if (!amdgpu_gtt_mgr_has_gart_addr(bo_mem)) {
1051                 gtt->offset = AMDGPU_BO_INVALID_OFFSET;
1052                 return 0;
1053         }
1054
1055         /* compute PTE flags relevant to this BO memory */
1056         flags = amdgpu_ttm_tt_pte_flags(adev, ttm, bo_mem);
1057
1058         /* bind pages into GART page tables */
1059         gtt->offset = (u64)bo_mem->start << PAGE_SHIFT;
1060         r = amdgpu_gart_bind(adev, gtt->offset, ttm->num_pages,
1061                 ttm->pages, gtt->ttm.dma_address, flags);
1062
1063         if (r)
1064                 DRM_ERROR("failed to bind %u pages at 0x%08llX\n",
1065                           ttm->num_pages, gtt->offset);
1066         gtt->bound = true;
1067         return r;
1068 }
1069
1070 /*
1071  * amdgpu_ttm_alloc_gart - Make sure buffer object is accessible either
1072  * through AGP or GART aperture.
1073  *
1074  * If bo is accessible through AGP aperture, then use AGP aperture
1075  * to access bo; otherwise allocate logical space in GART aperture
1076  * and map bo to GART aperture.
1077  */
1078 int amdgpu_ttm_alloc_gart(struct ttm_buffer_object *bo)
1079 {
1080         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1081         struct ttm_operation_ctx ctx = { false, false };
1082         struct amdgpu_ttm_tt *gtt = (void *)bo->ttm;
1083         struct ttm_resource tmp;
1084         struct ttm_placement placement;
1085         struct ttm_place placements;
1086         uint64_t addr, flags;
1087         int r;
1088
1089         if (bo->mem.start != AMDGPU_BO_INVALID_OFFSET)
1090                 return 0;
1091
1092         addr = amdgpu_gmc_agp_addr(bo);
1093         if (addr != AMDGPU_BO_INVALID_OFFSET) {
1094                 bo->mem.start = addr >> PAGE_SHIFT;
1095         } else {
1096
1097                 /* allocate GART space */
1098                 tmp = bo->mem;
1099                 tmp.mm_node = NULL;
1100                 placement.num_placement = 1;
1101                 placement.placement = &placements;
1102                 placement.num_busy_placement = 1;
1103                 placement.busy_placement = &placements;
1104                 placements.fpfn = 0;
1105                 placements.lpfn = adev->gmc.gart_size >> PAGE_SHIFT;
1106                 placements.mem_type = TTM_PL_TT;
1107                 placements.flags = bo->mem.placement;
1108
1109                 r = ttm_bo_mem_space(bo, &placement, &tmp, &ctx);
1110                 if (unlikely(r))
1111                         return r;
1112
1113                 /* compute PTE flags for this buffer object */
1114                 flags = amdgpu_ttm_tt_pte_flags(adev, bo->ttm, &tmp);
1115
1116                 /* Bind pages */
1117                 gtt->offset = (u64)tmp.start << PAGE_SHIFT;
1118                 r = amdgpu_ttm_gart_bind(adev, bo, flags);
1119                 if (unlikely(r)) {
1120                         ttm_resource_free(bo, &tmp);
1121                         return r;
1122                 }
1123
1124                 ttm_resource_free(bo, &bo->mem);
1125                 bo->mem = tmp;
1126         }
1127
1128         return 0;
1129 }
1130
1131 /*
1132  * amdgpu_ttm_recover_gart - Rebind GTT pages
1133  *
1134  * Called by amdgpu_gtt_mgr_recover() from amdgpu_device_reset() to
1135  * rebind GTT pages during a GPU reset.
1136  */
1137 int amdgpu_ttm_recover_gart(struct ttm_buffer_object *tbo)
1138 {
1139         struct amdgpu_device *adev = amdgpu_ttm_adev(tbo->bdev);
1140         uint64_t flags;
1141         int r;
1142
1143         if (!tbo->ttm)
1144                 return 0;
1145
1146         flags = amdgpu_ttm_tt_pte_flags(adev, tbo->ttm, &tbo->mem);
1147         r = amdgpu_ttm_gart_bind(adev, tbo, flags);
1148
1149         return r;
1150 }
1151
1152 /*
1153  * amdgpu_ttm_backend_unbind - Unbind GTT mapped pages
1154  *
1155  * Called by ttm_tt_unbind() on behalf of ttm_bo_move_ttm() and
1156  * ttm_tt_destroy().
1157  */
1158 static void amdgpu_ttm_backend_unbind(struct ttm_bo_device *bdev,
1159                                       struct ttm_tt *ttm)
1160 {
1161         struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
1162         struct amdgpu_ttm_tt *gtt = (void *)ttm;
1163         int r;
1164
1165         if (!gtt->bound)
1166                 return;
1167
1168         /* if the pages have userptr pinning then clear that first */
1169         if (gtt->userptr)
1170                 amdgpu_ttm_tt_unpin_userptr(bdev, ttm);
1171
1172         if (gtt->offset == AMDGPU_BO_INVALID_OFFSET)
1173                 return;
1174
1175         /* unbind shouldn't be done for GDS/GWS/OA in ttm_bo_clean_mm */
1176         r = amdgpu_gart_unbind(adev, gtt->offset, ttm->num_pages);
1177         if (r)
1178                 DRM_ERROR("failed to unbind %u pages at 0x%08llX\n",
1179                           gtt->ttm.num_pages, gtt->offset);
1180         gtt->bound = false;
1181 }
1182
1183 static void amdgpu_ttm_backend_destroy(struct ttm_bo_device *bdev,
1184                                        struct ttm_tt *ttm)
1185 {
1186         struct amdgpu_ttm_tt *gtt = (void *)ttm;
1187
1188         amdgpu_ttm_backend_unbind(bdev, ttm);
1189         ttm_tt_destroy_common(bdev, ttm);
1190         if (gtt->usertask)
1191                 put_task_struct(gtt->usertask);
1192
1193         ttm_tt_fini(&gtt->ttm);
1194         kfree(gtt);
1195 }
1196
1197 /**
1198  * amdgpu_ttm_tt_create - Create a ttm_tt object for a given BO
1199  *
1200  * @bo: The buffer object to create a GTT ttm_tt object around
1201  * @page_flags: Page flags to be added to the ttm_tt object
1202  *
1203  * Called by ttm_tt_create().
1204  */
1205 static struct ttm_tt *amdgpu_ttm_tt_create(struct ttm_buffer_object *bo,
1206                                            uint32_t page_flags)
1207 {
1208         struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
1209         struct amdgpu_ttm_tt *gtt;
1210         enum ttm_caching caching;
1211
1212         gtt = kzalloc(sizeof(struct amdgpu_ttm_tt), GFP_KERNEL);
1213         if (gtt == NULL) {
1214                 return NULL;
1215         }
1216         gtt->gobj = &bo->base;
1217
1218         if (abo->flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
1219                 caching = ttm_write_combined;
1220         else
1221                 caching = ttm_cached;
1222
1223         /* allocate space for the uninitialized page entries */
1224         if (ttm_sg_tt_init(&gtt->ttm, bo, page_flags, caching)) {
1225                 kfree(gtt);
1226                 return NULL;
1227         }
1228         return &gtt->ttm;
1229 }
1230
1231 /*
1232  * amdgpu_ttm_tt_populate - Map GTT pages visible to the device
1233  *
1234  * Map the pages of a ttm_tt object to an address space visible
1235  * to the underlying device.
1236  */
1237 static int amdgpu_ttm_tt_populate(struct ttm_bo_device *bdev,
1238                                   struct ttm_tt *ttm,
1239                                   struct ttm_operation_ctx *ctx)
1240 {
1241         struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
1242         struct amdgpu_ttm_tt *gtt = (void *)ttm;
1243
1244         /* user pages are bound by amdgpu_ttm_tt_pin_userptr() */
1245         if (gtt && gtt->userptr) {
1246                 ttm->sg = kzalloc(sizeof(struct sg_table), GFP_KERNEL);
1247                 if (!ttm->sg)
1248                         return -ENOMEM;
1249
1250                 ttm->page_flags |= TTM_PAGE_FLAG_SG;
1251                 return 0;
1252         }
1253
1254         if (ttm->page_flags & TTM_PAGE_FLAG_SG) {
1255                 if (!ttm->sg) {
1256                         struct dma_buf_attachment *attach;
1257                         struct sg_table *sgt;
1258
1259                         attach = gtt->gobj->import_attach;
1260                         sgt = dma_buf_map_attachment(attach, DMA_BIDIRECTIONAL);
1261                         if (IS_ERR(sgt))
1262                                 return PTR_ERR(sgt);
1263
1264                         ttm->sg = sgt;
1265                 }
1266
1267                 drm_prime_sg_to_dma_addr_array(ttm->sg, gtt->ttm.dma_address,
1268                                                ttm->num_pages);
1269                 return 0;
1270         }
1271
1272         return ttm_pool_alloc(&adev->mman.bdev.pool, ttm, ctx);
1273 }
1274
1275 /*
1276  * amdgpu_ttm_tt_unpopulate - unmap GTT pages and unpopulate page arrays
1277  *
1278  * Unmaps pages of a ttm_tt object from the device address space and
1279  * unpopulates the page array backing it.
1280  */
1281 static void amdgpu_ttm_tt_unpopulate(struct ttm_bo_device *bdev,
1282                                      struct ttm_tt *ttm)
1283 {
1284         struct amdgpu_ttm_tt *gtt = (void *)ttm;
1285         struct amdgpu_device *adev;
1286
1287         if (gtt && gtt->userptr) {
1288                 amdgpu_ttm_tt_set_user_pages(ttm, NULL);
1289                 kfree(ttm->sg);
1290                 ttm->page_flags &= ~TTM_PAGE_FLAG_SG;
1291                 return;
1292         }
1293
1294         if (ttm->sg && gtt->gobj->import_attach) {
1295                 struct dma_buf_attachment *attach;
1296
1297                 attach = gtt->gobj->import_attach;
1298                 dma_buf_unmap_attachment(attach, ttm->sg, DMA_BIDIRECTIONAL);
1299                 ttm->sg = NULL;
1300                 return;
1301         }
1302
1303         if (ttm->page_flags & TTM_PAGE_FLAG_SG)
1304                 return;
1305
1306         adev = amdgpu_ttm_adev(bdev);
1307         return ttm_pool_free(&adev->mman.bdev.pool, ttm);
1308 }
1309
1310 /**
1311  * amdgpu_ttm_tt_set_userptr - Initialize userptr GTT ttm_tt for the current
1312  * task
1313  *
1314  * @bo: The ttm_buffer_object to bind this userptr to
1315  * @addr:  The address in the current tasks VM space to use
1316  * @flags: Requirements of userptr object.
1317  *
1318  * Called by amdgpu_gem_userptr_ioctl() to bind userptr pages
1319  * to current task
1320  */
1321 int amdgpu_ttm_tt_set_userptr(struct ttm_buffer_object *bo,
1322                               uint64_t addr, uint32_t flags)
1323 {
1324         struct amdgpu_ttm_tt *gtt;
1325
1326         if (!bo->ttm) {
1327                 /* TODO: We want a separate TTM object type for userptrs */
1328                 bo->ttm = amdgpu_ttm_tt_create(bo, 0);
1329                 if (bo->ttm == NULL)
1330                         return -ENOMEM;
1331         }
1332
1333         gtt = (void *)bo->ttm;
1334         gtt->userptr = addr;
1335         gtt->userflags = flags;
1336
1337         if (gtt->usertask)
1338                 put_task_struct(gtt->usertask);
1339         gtt->usertask = current->group_leader;
1340         get_task_struct(gtt->usertask);
1341
1342         return 0;
1343 }
1344
1345 /*
1346  * amdgpu_ttm_tt_get_usermm - Return memory manager for ttm_tt object
1347  */
1348 struct mm_struct *amdgpu_ttm_tt_get_usermm(struct ttm_tt *ttm)
1349 {
1350         struct amdgpu_ttm_tt *gtt = (void *)ttm;
1351
1352         if (gtt == NULL)
1353                 return NULL;
1354
1355         if (gtt->usertask == NULL)
1356                 return NULL;
1357
1358         return gtt->usertask->mm;
1359 }
1360
1361 /*
1362  * amdgpu_ttm_tt_affect_userptr - Determine if a ttm_tt object lays inside an
1363  * address range for the current task.
1364  *
1365  */
1366 bool amdgpu_ttm_tt_affect_userptr(struct ttm_tt *ttm, unsigned long start,
1367                                   unsigned long end)
1368 {
1369         struct amdgpu_ttm_tt *gtt = (void *)ttm;
1370         unsigned long size;
1371
1372         if (gtt == NULL || !gtt->userptr)
1373                 return false;
1374
1375         /* Return false if no part of the ttm_tt object lies within
1376          * the range
1377          */
1378         size = (unsigned long)gtt->ttm.num_pages * PAGE_SIZE;
1379         if (gtt->userptr > end || gtt->userptr + size <= start)
1380                 return false;
1381
1382         return true;
1383 }
1384
1385 /*
1386  * amdgpu_ttm_tt_is_userptr - Have the pages backing by userptr?
1387  */
1388 bool amdgpu_ttm_tt_is_userptr(struct ttm_tt *ttm)
1389 {
1390         struct amdgpu_ttm_tt *gtt = (void *)ttm;
1391
1392         if (gtt == NULL || !gtt->userptr)
1393                 return false;
1394
1395         return true;
1396 }
1397
1398 /*
1399  * amdgpu_ttm_tt_is_readonly - Is the ttm_tt object read only?
1400  */
1401 bool amdgpu_ttm_tt_is_readonly(struct ttm_tt *ttm)
1402 {
1403         struct amdgpu_ttm_tt *gtt = (void *)ttm;
1404
1405         if (gtt == NULL)
1406                 return false;
1407
1408         return !!(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
1409 }
1410
1411 /**
1412  * amdgpu_ttm_tt_pde_flags - Compute PDE flags for ttm_tt object
1413  *
1414  * @ttm: The ttm_tt object to compute the flags for
1415  * @mem: The memory registry backing this ttm_tt object
1416  *
1417  * Figure out the flags to use for a VM PDE (Page Directory Entry).
1418  */
1419 uint64_t amdgpu_ttm_tt_pde_flags(struct ttm_tt *ttm, struct ttm_resource *mem)
1420 {
1421         uint64_t flags = 0;
1422
1423         if (mem && mem->mem_type != TTM_PL_SYSTEM)
1424                 flags |= AMDGPU_PTE_VALID;
1425
1426         if (mem && mem->mem_type == TTM_PL_TT) {
1427                 flags |= AMDGPU_PTE_SYSTEM;
1428
1429                 if (ttm->caching == ttm_cached)
1430                         flags |= AMDGPU_PTE_SNOOPED;
1431         }
1432
1433         return flags;
1434 }
1435
1436 /**
1437  * amdgpu_ttm_tt_pte_flags - Compute PTE flags for ttm_tt object
1438  *
1439  * @adev: amdgpu_device pointer
1440  * @ttm: The ttm_tt object to compute the flags for
1441  * @mem: The memory registry backing this ttm_tt object
1442  *
1443  * Figure out the flags to use for a VM PTE (Page Table Entry).
1444  */
1445 uint64_t amdgpu_ttm_tt_pte_flags(struct amdgpu_device *adev, struct ttm_tt *ttm,
1446                                  struct ttm_resource *mem)
1447 {
1448         uint64_t flags = amdgpu_ttm_tt_pde_flags(ttm, mem);
1449
1450         flags |= adev->gart.gart_pte_flags;
1451         flags |= AMDGPU_PTE_READABLE;
1452
1453         if (!amdgpu_ttm_tt_is_readonly(ttm))
1454                 flags |= AMDGPU_PTE_WRITEABLE;
1455
1456         return flags;
1457 }
1458
1459 /*
1460  * amdgpu_ttm_bo_eviction_valuable - Check to see if we can evict a buffer
1461  * object.
1462  *
1463  * Return true if eviction is sensible. Called by ttm_mem_evict_first() on
1464  * behalf of ttm_bo_mem_force_space() which tries to evict buffer objects until
1465  * it can find space for a new object and by ttm_bo_force_list_clean() which is
1466  * used to clean out a memory space.
1467  */
1468 static bool amdgpu_ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
1469                                             const struct ttm_place *place)
1470 {
1471         unsigned long num_pages = bo->mem.num_pages;
1472         struct drm_mm_node *node = bo->mem.mm_node;
1473         struct dma_resv_list *flist;
1474         struct dma_fence *f;
1475         int i;
1476
1477         if (bo->type == ttm_bo_type_kernel &&
1478             !amdgpu_vm_evictable(ttm_to_amdgpu_bo(bo)))
1479                 return false;
1480
1481         /* If bo is a KFD BO, check if the bo belongs to the current process.
1482          * If true, then return false as any KFD process needs all its BOs to
1483          * be resident to run successfully
1484          */
1485         flist = dma_resv_get_list(bo->base.resv);
1486         if (flist) {
1487                 for (i = 0; i < flist->shared_count; ++i) {
1488                         f = rcu_dereference_protected(flist->shared[i],
1489                                 dma_resv_held(bo->base.resv));
1490                         if (amdkfd_fence_check_mm(f, current->mm))
1491                                 return false;
1492                 }
1493         }
1494
1495         switch (bo->mem.mem_type) {
1496         case TTM_PL_TT:
1497                 if (amdgpu_bo_is_amdgpu_bo(bo) &&
1498                     amdgpu_bo_encrypted(ttm_to_amdgpu_bo(bo)))
1499                         return false;
1500                 return true;
1501
1502         case TTM_PL_VRAM:
1503                 /* Check each drm MM node individually */
1504                 while (num_pages) {
1505                         if (place->fpfn < (node->start + node->size) &&
1506                             !(place->lpfn && place->lpfn <= node->start))
1507                                 return true;
1508
1509                         num_pages -= node->size;
1510                         ++node;
1511                 }
1512                 return false;
1513
1514         default:
1515                 break;
1516         }
1517
1518         return ttm_bo_eviction_valuable(bo, place);
1519 }
1520
1521 /**
1522  * amdgpu_ttm_access_memory - Read or Write memory that backs a buffer object.
1523  *
1524  * @bo:  The buffer object to read/write
1525  * @offset:  Offset into buffer object
1526  * @buf:  Secondary buffer to write/read from
1527  * @len: Length in bytes of access
1528  * @write:  true if writing
1529  *
1530  * This is used to access VRAM that backs a buffer object via MMIO
1531  * access for debugging purposes.
1532  */
1533 static int amdgpu_ttm_access_memory(struct ttm_buffer_object *bo,
1534                                     unsigned long offset,
1535                                     void *buf, int len, int write)
1536 {
1537         struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
1538         struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev);
1539         struct drm_mm_node *nodes;
1540         uint32_t value = 0;
1541         int ret = 0;
1542         uint64_t pos;
1543         unsigned long flags;
1544
1545         if (bo->mem.mem_type != TTM_PL_VRAM)
1546                 return -EIO;
1547
1548         pos = offset;
1549         nodes = amdgpu_find_mm_node(&abo->tbo.mem, &pos);
1550         pos += (nodes->start << PAGE_SHIFT);
1551
1552         while (len && pos < adev->gmc.mc_vram_size) {
1553                 uint64_t aligned_pos = pos & ~(uint64_t)3;
1554                 uint64_t bytes = 4 - (pos & 3);
1555                 uint32_t shift = (pos & 3) * 8;
1556                 uint32_t mask = 0xffffffff << shift;
1557
1558                 if (len < bytes) {
1559                         mask &= 0xffffffff >> (bytes - len) * 8;
1560                         bytes = len;
1561                 }
1562
1563                 if (mask != 0xffffffff) {
1564                         spin_lock_irqsave(&adev->mmio_idx_lock, flags);
1565                         WREG32_NO_KIQ(mmMM_INDEX, ((uint32_t)aligned_pos) | 0x80000000);
1566                         WREG32_NO_KIQ(mmMM_INDEX_HI, aligned_pos >> 31);
1567                         if (!write || mask != 0xffffffff)
1568                                 value = RREG32_NO_KIQ(mmMM_DATA);
1569                         if (write) {
1570                                 value &= ~mask;
1571                                 value |= (*(uint32_t *)buf << shift) & mask;
1572                                 WREG32_NO_KIQ(mmMM_DATA, value);
1573                         }
1574                         spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
1575                         if (!write) {
1576                                 value = (value & mask) >> shift;
1577                                 memcpy(buf, &value, bytes);
1578                         }
1579                 } else {
1580                         bytes = (nodes->start + nodes->size) << PAGE_SHIFT;
1581                         bytes = min(bytes - pos, (uint64_t)len & ~0x3ull);
1582
1583                         amdgpu_device_vram_access(adev, pos, (uint32_t *)buf,
1584                                                   bytes, write);
1585                 }
1586
1587                 ret += bytes;
1588                 buf = (uint8_t *)buf + bytes;
1589                 pos += bytes;
1590                 len -= bytes;
1591                 if (pos >= (nodes->start + nodes->size) << PAGE_SHIFT) {
1592                         ++nodes;
1593                         pos = (nodes->start << PAGE_SHIFT);
1594                 }
1595         }
1596
1597         return ret;
1598 }
1599
1600 static void
1601 amdgpu_bo_delete_mem_notify(struct ttm_buffer_object *bo)
1602 {
1603         amdgpu_bo_move_notify(bo, false, NULL);
1604 }
1605
1606 static struct ttm_bo_driver amdgpu_bo_driver = {
1607         .ttm_tt_create = &amdgpu_ttm_tt_create,
1608         .ttm_tt_populate = &amdgpu_ttm_tt_populate,
1609         .ttm_tt_unpopulate = &amdgpu_ttm_tt_unpopulate,
1610         .ttm_tt_destroy = &amdgpu_ttm_backend_destroy,
1611         .eviction_valuable = amdgpu_ttm_bo_eviction_valuable,
1612         .evict_flags = &amdgpu_evict_flags,
1613         .move = &amdgpu_bo_move,
1614         .verify_access = &amdgpu_verify_access,
1615         .delete_mem_notify = &amdgpu_bo_delete_mem_notify,
1616         .release_notify = &amdgpu_bo_release_notify,
1617         .io_mem_reserve = &amdgpu_ttm_io_mem_reserve,
1618         .io_mem_pfn = amdgpu_ttm_io_mem_pfn,
1619         .access_memory = &amdgpu_ttm_access_memory,
1620         .del_from_lru_notify = &amdgpu_vm_del_from_lru_notify
1621 };
1622
1623 /*
1624  * Firmware Reservation functions
1625  */
1626 /**
1627  * amdgpu_ttm_fw_reserve_vram_fini - free fw reserved vram
1628  *
1629  * @adev: amdgpu_device pointer
1630  *
1631  * free fw reserved vram if it has been reserved.
1632  */
1633 static void amdgpu_ttm_fw_reserve_vram_fini(struct amdgpu_device *adev)
1634 {
1635         amdgpu_bo_free_kernel(&adev->mman.fw_vram_usage_reserved_bo,
1636                 NULL, &adev->mman.fw_vram_usage_va);
1637 }
1638
1639 /**
1640  * amdgpu_ttm_fw_reserve_vram_init - create bo vram reservation from fw
1641  *
1642  * @adev: amdgpu_device pointer
1643  *
1644  * create bo vram reservation from fw.
1645  */
1646 static int amdgpu_ttm_fw_reserve_vram_init(struct amdgpu_device *adev)
1647 {
1648         uint64_t vram_size = adev->gmc.visible_vram_size;
1649
1650         adev->mman.fw_vram_usage_va = NULL;
1651         adev->mman.fw_vram_usage_reserved_bo = NULL;
1652
1653         if (adev->mman.fw_vram_usage_size == 0 ||
1654             adev->mman.fw_vram_usage_size > vram_size)
1655                 return 0;
1656
1657         return amdgpu_bo_create_kernel_at(adev,
1658                                           adev->mman.fw_vram_usage_start_offset,
1659                                           adev->mman.fw_vram_usage_size,
1660                                           AMDGPU_GEM_DOMAIN_VRAM,
1661                                           &adev->mman.fw_vram_usage_reserved_bo,
1662                                           &adev->mman.fw_vram_usage_va);
1663 }
1664
1665 /*
1666  * Memoy training reservation functions
1667  */
1668
1669 /**
1670  * amdgpu_ttm_training_reserve_vram_fini - free memory training reserved vram
1671  *
1672  * @adev: amdgpu_device pointer
1673  *
1674  * free memory training reserved vram if it has been reserved.
1675  */
1676 static int amdgpu_ttm_training_reserve_vram_fini(struct amdgpu_device *adev)
1677 {
1678         struct psp_memory_training_context *ctx = &adev->psp.mem_train_ctx;
1679
1680         ctx->init = PSP_MEM_TRAIN_NOT_SUPPORT;
1681         amdgpu_bo_free_kernel(&ctx->c2p_bo, NULL, NULL);
1682         ctx->c2p_bo = NULL;
1683
1684         return 0;
1685 }
1686
1687 static void amdgpu_ttm_training_data_block_init(struct amdgpu_device *adev)
1688 {
1689         struct psp_memory_training_context *ctx = &adev->psp.mem_train_ctx;
1690
1691         memset(ctx, 0, sizeof(*ctx));
1692
1693         ctx->c2p_train_data_offset =
1694                 ALIGN((adev->gmc.mc_vram_size - adev->mman.discovery_tmr_size - SZ_1M), SZ_1M);
1695         ctx->p2c_train_data_offset =
1696                 (adev->gmc.mc_vram_size - GDDR6_MEM_TRAINING_OFFSET);
1697         ctx->train_data_size =
1698                 GDDR6_MEM_TRAINING_DATA_SIZE_IN_BYTES;
1699
1700         DRM_DEBUG("train_data_size:%llx,p2c_train_data_offset:%llx,c2p_train_data_offset:%llx.\n",
1701                         ctx->train_data_size,
1702                         ctx->p2c_train_data_offset,
1703                         ctx->c2p_train_data_offset);
1704 }
1705
1706 /*
1707  * reserve TMR memory at the top of VRAM which holds
1708  * IP Discovery data and is protected by PSP.
1709  */
1710 static int amdgpu_ttm_reserve_tmr(struct amdgpu_device *adev)
1711 {
1712         int ret;
1713         struct psp_memory_training_context *ctx = &adev->psp.mem_train_ctx;
1714         bool mem_train_support = false;
1715
1716         if (!amdgpu_sriov_vf(adev)) {
1717                 ret = amdgpu_mem_train_support(adev);
1718                 if (ret == 1)
1719                         mem_train_support = true;
1720                 else if (ret == -1)
1721                         return -EINVAL;
1722                 else
1723                         DRM_DEBUG("memory training does not support!\n");
1724         }
1725
1726         /*
1727          * Query reserved tmr size through atom firmwareinfo for Sienna_Cichlid and onwards for all
1728          * the use cases (IP discovery/G6 memory training/profiling/diagnostic data.etc)
1729          *
1730          * Otherwise, fallback to legacy approach to check and reserve tmr block for ip
1731          * discovery data and G6 memory training data respectively
1732          */
1733         adev->mman.discovery_tmr_size =
1734                 amdgpu_atomfirmware_get_fw_reserved_fb_size(adev);
1735         if (!adev->mman.discovery_tmr_size)
1736                 adev->mman.discovery_tmr_size = DISCOVERY_TMR_OFFSET;
1737
1738         if (mem_train_support) {
1739                 /* reserve vram for mem train according to TMR location */
1740                 amdgpu_ttm_training_data_block_init(adev);
1741                 ret = amdgpu_bo_create_kernel_at(adev,
1742                                          ctx->c2p_train_data_offset,
1743                                          ctx->train_data_size,
1744                                          AMDGPU_GEM_DOMAIN_VRAM,
1745                                          &ctx->c2p_bo,
1746                                          NULL);
1747                 if (ret) {
1748                         DRM_ERROR("alloc c2p_bo failed(%d)!\n", ret);
1749                         amdgpu_ttm_training_reserve_vram_fini(adev);
1750                         return ret;
1751                 }
1752                 ctx->init = PSP_MEM_TRAIN_RESERVE_SUCCESS;
1753         }
1754
1755         ret = amdgpu_bo_create_kernel_at(adev,
1756                                 adev->gmc.real_vram_size - adev->mman.discovery_tmr_size,
1757                                 adev->mman.discovery_tmr_size,
1758                                 AMDGPU_GEM_DOMAIN_VRAM,
1759                                 &adev->mman.discovery_memory,
1760                                 NULL);
1761         if (ret) {
1762                 DRM_ERROR("alloc tmr failed(%d)!\n", ret);
1763                 amdgpu_bo_free_kernel(&adev->mman.discovery_memory, NULL, NULL);
1764                 return ret;
1765         }
1766
1767         return 0;
1768 }
1769
1770 /*
1771  * amdgpu_ttm_init - Init the memory management (ttm) as well as various
1772  * gtt/vram related fields.
1773  *
1774  * This initializes all of the memory space pools that the TTM layer
1775  * will need such as the GTT space (system memory mapped to the device),
1776  * VRAM (on-board memory), and on-chip memories (GDS, GWS, OA) which
1777  * can be mapped per VMID.
1778  */
1779 int amdgpu_ttm_init(struct amdgpu_device *adev)
1780 {
1781         uint64_t gtt_size;
1782         int r;
1783         u64 vis_vram_limit;
1784
1785         mutex_init(&adev->mman.gtt_window_lock);
1786
1787         /* No others user of address space so set it to 0 */
1788         r = ttm_bo_device_init(&adev->mman.bdev, &amdgpu_bo_driver, adev->dev,
1789                                adev_to_drm(adev)->anon_inode->i_mapping,
1790                                adev_to_drm(adev)->vma_offset_manager,
1791                                adev->need_swiotlb,
1792                                dma_addressing_limited(adev->dev));
1793         if (r) {
1794                 DRM_ERROR("failed initializing buffer object driver(%d).\n", r);
1795                 return r;
1796         }
1797         adev->mman.initialized = true;
1798
1799         /* Initialize VRAM pool with all of VRAM divided into pages */
1800         r = amdgpu_vram_mgr_init(adev);
1801         if (r) {
1802                 DRM_ERROR("Failed initializing VRAM heap.\n");
1803                 return r;
1804         }
1805
1806         /* Reduce size of CPU-visible VRAM if requested */
1807         vis_vram_limit = (u64)amdgpu_vis_vram_limit * 1024 * 1024;
1808         if (amdgpu_vis_vram_limit > 0 &&
1809             vis_vram_limit <= adev->gmc.visible_vram_size)
1810                 adev->gmc.visible_vram_size = vis_vram_limit;
1811
1812         /* Change the size here instead of the init above so only lpfn is affected */
1813         amdgpu_ttm_set_buffer_funcs_status(adev, false);
1814 #ifdef CONFIG_64BIT
1815         adev->mman.aper_base_kaddr = ioremap_wc(adev->gmc.aper_base,
1816                                                 adev->gmc.visible_vram_size);
1817 #endif
1818
1819         /*
1820          *The reserved vram for firmware must be pinned to the specified
1821          *place on the VRAM, so reserve it early.
1822          */
1823         r = amdgpu_ttm_fw_reserve_vram_init(adev);
1824         if (r) {
1825                 return r;
1826         }
1827
1828         /*
1829          * only NAVI10 and onwards ASIC support for IP discovery.
1830          * If IP discovery enabled, a block of memory should be
1831          * reserved for IP discovey.
1832          */
1833         if (adev->mman.discovery_bin) {
1834                 r = amdgpu_ttm_reserve_tmr(adev);
1835                 if (r)
1836                         return r;
1837         }
1838
1839         /* allocate memory as required for VGA
1840          * This is used for VGA emulation and pre-OS scanout buffers to
1841          * avoid display artifacts while transitioning between pre-OS
1842          * and driver.  */
1843         r = amdgpu_bo_create_kernel_at(adev, 0, adev->mman.stolen_vga_size,
1844                                        AMDGPU_GEM_DOMAIN_VRAM,
1845                                        &adev->mman.stolen_vga_memory,
1846                                        NULL);
1847         if (r)
1848                 return r;
1849         r = amdgpu_bo_create_kernel_at(adev, adev->mman.stolen_vga_size,
1850                                        adev->mman.stolen_extended_size,
1851                                        AMDGPU_GEM_DOMAIN_VRAM,
1852                                        &adev->mman.stolen_extended_memory,
1853                                        NULL);
1854         if (r)
1855                 return r;
1856
1857         DRM_INFO("amdgpu: %uM of VRAM memory ready\n",
1858                  (unsigned) (adev->gmc.real_vram_size / (1024 * 1024)));
1859
1860         /* Compute GTT size, either bsaed on 3/4th the size of RAM size
1861          * or whatever the user passed on module init */
1862         if (amdgpu_gtt_size == -1) {
1863                 struct sysinfo si;
1864
1865                 si_meminfo(&si);
1866                 gtt_size = min(max((AMDGPU_DEFAULT_GTT_SIZE_MB << 20),
1867                                adev->gmc.mc_vram_size),
1868                                ((uint64_t)si.totalram * si.mem_unit * 3/4));
1869         }
1870         else
1871                 gtt_size = (uint64_t)amdgpu_gtt_size << 20;
1872
1873         /* Initialize GTT memory pool */
1874         r = amdgpu_gtt_mgr_init(adev, gtt_size);
1875         if (r) {
1876                 DRM_ERROR("Failed initializing GTT heap.\n");
1877                 return r;
1878         }
1879         DRM_INFO("amdgpu: %uM of GTT memory ready.\n",
1880                  (unsigned)(gtt_size / (1024 * 1024)));
1881
1882         /* Initialize various on-chip memory pools */
1883         r = amdgpu_ttm_init_on_chip(adev, AMDGPU_PL_GDS, adev->gds.gds_size);
1884         if (r) {
1885                 DRM_ERROR("Failed initializing GDS heap.\n");
1886                 return r;
1887         }
1888
1889         r = amdgpu_ttm_init_on_chip(adev, AMDGPU_PL_GWS, adev->gds.gws_size);
1890         if (r) {
1891                 DRM_ERROR("Failed initializing gws heap.\n");
1892                 return r;
1893         }
1894
1895         r = amdgpu_ttm_init_on_chip(adev, AMDGPU_PL_OA, adev->gds.oa_size);
1896         if (r) {
1897                 DRM_ERROR("Failed initializing oa heap.\n");
1898                 return r;
1899         }
1900
1901         return 0;
1902 }
1903
1904 /*
1905  * amdgpu_ttm_fini - De-initialize the TTM memory pools
1906  */
1907 void amdgpu_ttm_fini(struct amdgpu_device *adev)
1908 {
1909         if (!adev->mman.initialized)
1910                 return;
1911
1912         amdgpu_ttm_training_reserve_vram_fini(adev);
1913         /* return the stolen vga memory back to VRAM */
1914         amdgpu_bo_free_kernel(&adev->mman.stolen_vga_memory, NULL, NULL);
1915         amdgpu_bo_free_kernel(&adev->mman.stolen_extended_memory, NULL, NULL);
1916         /* return the IP Discovery TMR memory back to VRAM */
1917         amdgpu_bo_free_kernel(&adev->mman.discovery_memory, NULL, NULL);
1918         amdgpu_ttm_fw_reserve_vram_fini(adev);
1919
1920         if (adev->mman.aper_base_kaddr)
1921                 iounmap(adev->mman.aper_base_kaddr);
1922         adev->mman.aper_base_kaddr = NULL;
1923
1924         amdgpu_vram_mgr_fini(adev);
1925         amdgpu_gtt_mgr_fini(adev);
1926         ttm_range_man_fini(&adev->mman.bdev, AMDGPU_PL_GDS);
1927         ttm_range_man_fini(&adev->mman.bdev, AMDGPU_PL_GWS);
1928         ttm_range_man_fini(&adev->mman.bdev, AMDGPU_PL_OA);
1929         ttm_bo_device_release(&adev->mman.bdev);
1930         adev->mman.initialized = false;
1931         DRM_INFO("amdgpu: ttm finalized\n");
1932 }
1933
1934 /**
1935  * amdgpu_ttm_set_buffer_funcs_status - enable/disable use of buffer functions
1936  *
1937  * @adev: amdgpu_device pointer
1938  * @enable: true when we can use buffer functions.
1939  *
1940  * Enable/disable use of buffer functions during suspend/resume. This should
1941  * only be called at bootup or when userspace isn't running.
1942  */
1943 void amdgpu_ttm_set_buffer_funcs_status(struct amdgpu_device *adev, bool enable)
1944 {
1945         struct ttm_resource_manager *man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM);
1946         uint64_t size;
1947         int r;
1948
1949         if (!adev->mman.initialized || amdgpu_in_reset(adev) ||
1950             adev->mman.buffer_funcs_enabled == enable)
1951                 return;
1952
1953         if (enable) {
1954                 struct amdgpu_ring *ring;
1955                 struct drm_gpu_scheduler *sched;
1956
1957                 ring = adev->mman.buffer_funcs_ring;
1958                 sched = &ring->sched;
1959                 r = drm_sched_entity_init(&adev->mman.entity,
1960                                           DRM_SCHED_PRIORITY_KERNEL, &sched,
1961                                           1, NULL);
1962                 if (r) {
1963                         DRM_ERROR("Failed setting up TTM BO move entity (%d)\n",
1964                                   r);
1965                         return;
1966                 }
1967         } else {
1968                 drm_sched_entity_destroy(&adev->mman.entity);
1969                 dma_fence_put(man->move);
1970                 man->move = NULL;
1971         }
1972
1973         /* this just adjusts TTM size idea, which sets lpfn to the correct value */
1974         if (enable)
1975                 size = adev->gmc.real_vram_size;
1976         else
1977                 size = adev->gmc.visible_vram_size;
1978         man->size = size >> PAGE_SHIFT;
1979         adev->mman.buffer_funcs_enabled = enable;
1980 }
1981
1982 static vm_fault_t amdgpu_ttm_fault(struct vm_fault *vmf)
1983 {
1984         struct ttm_buffer_object *bo = vmf->vma->vm_private_data;
1985         vm_fault_t ret;
1986
1987         ret = ttm_bo_vm_reserve(bo, vmf);
1988         if (ret)
1989                 return ret;
1990
1991         ret = amdgpu_bo_fault_reserve_notify(bo);
1992         if (ret)
1993                 goto unlock;
1994
1995         ret = ttm_bo_vm_fault_reserved(vmf, vmf->vma->vm_page_prot,
1996                                        TTM_BO_VM_NUM_PREFAULT, 1);
1997         if (ret == VM_FAULT_RETRY && !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT))
1998                 return ret;
1999
2000 unlock:
2001         dma_resv_unlock(bo->base.resv);
2002         return ret;
2003 }
2004
2005 static struct vm_operations_struct amdgpu_ttm_vm_ops = {
2006         .fault = amdgpu_ttm_fault,
2007         .open = ttm_bo_vm_open,
2008         .close = ttm_bo_vm_close,
2009         .access = ttm_bo_vm_access
2010 };
2011
2012 int amdgpu_mmap(struct file *filp, struct vm_area_struct *vma)
2013 {
2014         struct drm_file *file_priv = filp->private_data;
2015         struct amdgpu_device *adev = drm_to_adev(file_priv->minor->dev);
2016         int r;
2017
2018         r = ttm_bo_mmap(filp, vma, &adev->mman.bdev);
2019         if (unlikely(r != 0))
2020                 return r;
2021
2022         vma->vm_ops = &amdgpu_ttm_vm_ops;
2023         return 0;
2024 }
2025
2026 int amdgpu_copy_buffer(struct amdgpu_ring *ring, uint64_t src_offset,
2027                        uint64_t dst_offset, uint32_t byte_count,
2028                        struct dma_resv *resv,
2029                        struct dma_fence **fence, bool direct_submit,
2030                        bool vm_needs_flush, bool tmz)
2031 {
2032         enum amdgpu_ib_pool_type pool = direct_submit ? AMDGPU_IB_POOL_DIRECT :
2033                 AMDGPU_IB_POOL_DELAYED;
2034         struct amdgpu_device *adev = ring->adev;
2035         struct amdgpu_job *job;
2036
2037         uint32_t max_bytes;
2038         unsigned num_loops, num_dw;
2039         unsigned i;
2040         int r;
2041
2042         if (direct_submit && !ring->sched.ready) {
2043                 DRM_ERROR("Trying to move memory with ring turned off.\n");
2044                 return -EINVAL;
2045         }
2046
2047         max_bytes = adev->mman.buffer_funcs->copy_max_bytes;
2048         num_loops = DIV_ROUND_UP(byte_count, max_bytes);
2049         num_dw = ALIGN(num_loops * adev->mman.buffer_funcs->copy_num_dw, 8);
2050
2051         r = amdgpu_job_alloc_with_ib(adev, num_dw * 4, pool, &job);
2052         if (r)
2053                 return r;
2054
2055         if (vm_needs_flush) {
2056                 job->vm_pd_addr = amdgpu_gmc_pd_addr(adev->gart.bo);
2057                 job->vm_needs_flush = true;
2058         }
2059         if (resv) {
2060                 r = amdgpu_sync_resv(adev, &job->sync, resv,
2061                                      AMDGPU_SYNC_ALWAYS,
2062                                      AMDGPU_FENCE_OWNER_UNDEFINED);
2063                 if (r) {
2064                         DRM_ERROR("sync failed (%d).\n", r);
2065                         goto error_free;
2066                 }
2067         }
2068
2069         for (i = 0; i < num_loops; i++) {
2070                 uint32_t cur_size_in_bytes = min(byte_count, max_bytes);
2071
2072                 amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_offset,
2073                                         dst_offset, cur_size_in_bytes, tmz);
2074
2075                 src_offset += cur_size_in_bytes;
2076                 dst_offset += cur_size_in_bytes;
2077                 byte_count -= cur_size_in_bytes;
2078         }
2079
2080         amdgpu_ring_pad_ib(ring, &job->ibs[0]);
2081         WARN_ON(job->ibs[0].length_dw > num_dw);
2082         if (direct_submit)
2083                 r = amdgpu_job_submit_direct(job, ring, fence);
2084         else
2085                 r = amdgpu_job_submit(job, &adev->mman.entity,
2086                                       AMDGPU_FENCE_OWNER_UNDEFINED, fence);
2087         if (r)
2088                 goto error_free;
2089
2090         return r;
2091
2092 error_free:
2093         amdgpu_job_free(job);
2094         DRM_ERROR("Error scheduling IBs (%d)\n", r);
2095         return r;
2096 }
2097
2098 int amdgpu_fill_buffer(struct amdgpu_bo *bo,
2099                        uint32_t src_data,
2100                        struct dma_resv *resv,
2101                        struct dma_fence **fence)
2102 {
2103         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
2104         uint32_t max_bytes = adev->mman.buffer_funcs->fill_max_bytes;
2105         struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
2106
2107         struct drm_mm_node *mm_node;
2108         unsigned long num_pages;
2109         unsigned int num_loops, num_dw;
2110
2111         struct amdgpu_job *job;
2112         int r;
2113
2114         if (!adev->mman.buffer_funcs_enabled) {
2115                 DRM_ERROR("Trying to clear memory with ring turned off.\n");
2116                 return -EINVAL;
2117         }
2118
2119         if (bo->tbo.mem.mem_type == TTM_PL_TT) {
2120                 r = amdgpu_ttm_alloc_gart(&bo->tbo);
2121                 if (r)
2122                         return r;
2123         }
2124
2125         num_pages = bo->tbo.mem.num_pages;
2126         mm_node = bo->tbo.mem.mm_node;
2127         num_loops = 0;
2128         while (num_pages) {
2129                 uint64_t byte_count = mm_node->size << PAGE_SHIFT;
2130
2131                 num_loops += DIV_ROUND_UP_ULL(byte_count, max_bytes);
2132                 num_pages -= mm_node->size;
2133                 ++mm_node;
2134         }
2135         num_dw = num_loops * adev->mman.buffer_funcs->fill_num_dw;
2136
2137         /* for IB padding */
2138         num_dw += 64;
2139
2140         r = amdgpu_job_alloc_with_ib(adev, num_dw * 4, AMDGPU_IB_POOL_DELAYED,
2141                                      &job);
2142         if (r)
2143                 return r;
2144
2145         if (resv) {
2146                 r = amdgpu_sync_resv(adev, &job->sync, resv,
2147                                      AMDGPU_SYNC_ALWAYS,
2148                                      AMDGPU_FENCE_OWNER_UNDEFINED);
2149                 if (r) {
2150                         DRM_ERROR("sync failed (%d).\n", r);
2151                         goto error_free;
2152                 }
2153         }
2154
2155         num_pages = bo->tbo.mem.num_pages;
2156         mm_node = bo->tbo.mem.mm_node;
2157
2158         while (num_pages) {
2159                 uint64_t byte_count = mm_node->size << PAGE_SHIFT;
2160                 uint64_t dst_addr;
2161
2162                 dst_addr = amdgpu_mm_node_addr(&bo->tbo, mm_node, &bo->tbo.mem);
2163                 while (byte_count) {
2164                         uint32_t cur_size_in_bytes = min_t(uint64_t, byte_count,
2165                                                            max_bytes);
2166
2167                         amdgpu_emit_fill_buffer(adev, &job->ibs[0], src_data,
2168                                                 dst_addr, cur_size_in_bytes);
2169
2170                         dst_addr += cur_size_in_bytes;
2171                         byte_count -= cur_size_in_bytes;
2172                 }
2173
2174                 num_pages -= mm_node->size;
2175                 ++mm_node;
2176         }
2177
2178         amdgpu_ring_pad_ib(ring, &job->ibs[0]);
2179         WARN_ON(job->ibs[0].length_dw > num_dw);
2180         r = amdgpu_job_submit(job, &adev->mman.entity,
2181                               AMDGPU_FENCE_OWNER_UNDEFINED, fence);
2182         if (r)
2183                 goto error_free;
2184
2185         return 0;
2186
2187 error_free:
2188         amdgpu_job_free(job);
2189         return r;
2190 }
2191
2192 #if defined(CONFIG_DEBUG_FS)
2193
2194 static int amdgpu_mm_dump_table(struct seq_file *m, void *data)
2195 {
2196         struct drm_info_node *node = (struct drm_info_node *)m->private;
2197         unsigned ttm_pl = (uintptr_t)node->info_ent->data;
2198         struct drm_device *dev = node->minor->dev;
2199         struct amdgpu_device *adev = drm_to_adev(dev);
2200         struct ttm_resource_manager *man = ttm_manager_type(&adev->mman.bdev, ttm_pl);
2201         struct drm_printer p = drm_seq_file_printer(m);
2202
2203         man->func->debug(man, &p);
2204         return 0;
2205 }
2206
2207 static int amdgpu_ttm_pool_debugfs(struct seq_file *m, void *data)
2208 {
2209         struct drm_info_node *node = (struct drm_info_node *)m->private;
2210         struct drm_device *dev = node->minor->dev;
2211         struct amdgpu_device *adev = drm_to_adev(dev);
2212
2213         return ttm_pool_debugfs(&adev->mman.bdev.pool, m);
2214 }
2215
2216 static const struct drm_info_list amdgpu_ttm_debugfs_list[] = {
2217         {"amdgpu_vram_mm", amdgpu_mm_dump_table, 0, (void *)TTM_PL_VRAM},
2218         {"amdgpu_gtt_mm", amdgpu_mm_dump_table, 0, (void *)TTM_PL_TT},
2219         {"amdgpu_gds_mm", amdgpu_mm_dump_table, 0, (void *)AMDGPU_PL_GDS},
2220         {"amdgpu_gws_mm", amdgpu_mm_dump_table, 0, (void *)AMDGPU_PL_GWS},
2221         {"amdgpu_oa_mm", amdgpu_mm_dump_table, 0, (void *)AMDGPU_PL_OA},
2222         {"ttm_page_pool", amdgpu_ttm_pool_debugfs, 0, NULL},
2223 };
2224
2225 /*
2226  * amdgpu_ttm_vram_read - Linear read access to VRAM
2227  *
2228  * Accesses VRAM via MMIO for debugging purposes.
2229  */
2230 static ssize_t amdgpu_ttm_vram_read(struct file *f, char __user *buf,
2231                                     size_t size, loff_t *pos)
2232 {
2233         struct amdgpu_device *adev = file_inode(f)->i_private;
2234         ssize_t result = 0;
2235
2236         if (size & 0x3 || *pos & 0x3)
2237                 return -EINVAL;
2238
2239         if (*pos >= adev->gmc.mc_vram_size)
2240                 return -ENXIO;
2241
2242         size = min(size, (size_t)(adev->gmc.mc_vram_size - *pos));
2243         while (size) {
2244                 size_t bytes = min(size, AMDGPU_TTM_VRAM_MAX_DW_READ * 4);
2245                 uint32_t value[AMDGPU_TTM_VRAM_MAX_DW_READ];
2246
2247                 amdgpu_device_vram_access(adev, *pos, value, bytes, false);
2248                 if (copy_to_user(buf, value, bytes))
2249                         return -EFAULT;
2250
2251                 result += bytes;
2252                 buf += bytes;
2253                 *pos += bytes;
2254                 size -= bytes;
2255         }
2256
2257         return result;
2258 }
2259
2260 /*
2261  * amdgpu_ttm_vram_write - Linear write access to VRAM
2262  *
2263  * Accesses VRAM via MMIO for debugging purposes.
2264  */
2265 static ssize_t amdgpu_ttm_vram_write(struct file *f, const char __user *buf,
2266                                     size_t size, loff_t *pos)
2267 {
2268         struct amdgpu_device *adev = file_inode(f)->i_private;
2269         ssize_t result = 0;
2270         int r;
2271
2272         if (size & 0x3 || *pos & 0x3)
2273                 return -EINVAL;
2274
2275         if (*pos >= adev->gmc.mc_vram_size)
2276                 return -ENXIO;
2277
2278         while (size) {
2279                 unsigned long flags;
2280                 uint32_t value;
2281
2282                 if (*pos >= adev->gmc.mc_vram_size)
2283                         return result;
2284
2285                 r = get_user(value, (uint32_t *)buf);
2286                 if (r)
2287                         return r;
2288
2289                 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
2290                 WREG32_NO_KIQ(mmMM_INDEX, ((uint32_t)*pos) | 0x80000000);
2291                 WREG32_NO_KIQ(mmMM_INDEX_HI, *pos >> 31);
2292                 WREG32_NO_KIQ(mmMM_DATA, value);
2293                 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
2294
2295                 result += 4;
2296                 buf += 4;
2297                 *pos += 4;
2298                 size -= 4;
2299         }
2300
2301         return result;
2302 }
2303
2304 static const struct file_operations amdgpu_ttm_vram_fops = {
2305         .owner = THIS_MODULE,
2306         .read = amdgpu_ttm_vram_read,
2307         .write = amdgpu_ttm_vram_write,
2308         .llseek = default_llseek,
2309 };
2310
2311 #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS
2312
2313 /*
2314  * amdgpu_ttm_gtt_read - Linear read access to GTT memory
2315  */
2316 static ssize_t amdgpu_ttm_gtt_read(struct file *f, char __user *buf,
2317                                    size_t size, loff_t *pos)
2318 {
2319         struct amdgpu_device *adev = file_inode(f)->i_private;
2320         ssize_t result = 0;
2321         int r;
2322
2323         while (size) {
2324                 loff_t p = *pos / PAGE_SIZE;
2325                 unsigned off = *pos & ~PAGE_MASK;
2326                 size_t cur_size = min_t(size_t, size, PAGE_SIZE - off);
2327                 struct page *page;
2328                 void *ptr;
2329
2330                 if (p >= adev->gart.num_cpu_pages)
2331                         return result;
2332
2333                 page = adev->gart.pages[p];
2334                 if (page) {
2335                         ptr = kmap(page);
2336                         ptr += off;
2337
2338                         r = copy_to_user(buf, ptr, cur_size);
2339                         kunmap(adev->gart.pages[p]);
2340                 } else
2341                         r = clear_user(buf, cur_size);
2342
2343                 if (r)
2344                         return -EFAULT;
2345
2346                 result += cur_size;
2347                 buf += cur_size;
2348                 *pos += cur_size;
2349                 size -= cur_size;
2350         }
2351
2352         return result;
2353 }
2354
2355 static const struct file_operations amdgpu_ttm_gtt_fops = {
2356         .owner = THIS_MODULE,
2357         .read = amdgpu_ttm_gtt_read,
2358         .llseek = default_llseek
2359 };
2360
2361 #endif
2362
2363 /*
2364  * amdgpu_iomem_read - Virtual read access to GPU mapped memory
2365  *
2366  * This function is used to read memory that has been mapped to the
2367  * GPU and the known addresses are not physical addresses but instead
2368  * bus addresses (e.g., what you'd put in an IB or ring buffer).
2369  */
2370 static ssize_t amdgpu_iomem_read(struct file *f, char __user *buf,
2371                                  size_t size, loff_t *pos)
2372 {
2373         struct amdgpu_device *adev = file_inode(f)->i_private;
2374         struct iommu_domain *dom;
2375         ssize_t result = 0;
2376         int r;
2377
2378         /* retrieve the IOMMU domain if any for this device */
2379         dom = iommu_get_domain_for_dev(adev->dev);
2380
2381         while (size) {
2382                 phys_addr_t addr = *pos & PAGE_MASK;
2383                 loff_t off = *pos & ~PAGE_MASK;
2384                 size_t bytes = PAGE_SIZE - off;
2385                 unsigned long pfn;
2386                 struct page *p;
2387                 void *ptr;
2388
2389                 bytes = bytes < size ? bytes : size;
2390
2391                 /* Translate the bus address to a physical address.  If
2392                  * the domain is NULL it means there is no IOMMU active
2393                  * and the address translation is the identity
2394                  */
2395                 addr = dom ? iommu_iova_to_phys(dom, addr) : addr;
2396
2397                 pfn = addr >> PAGE_SHIFT;
2398                 if (!pfn_valid(pfn))
2399                         return -EPERM;
2400
2401                 p = pfn_to_page(pfn);
2402                 if (p->mapping != adev->mman.bdev.dev_mapping)
2403                         return -EPERM;
2404
2405                 ptr = kmap(p);
2406                 r = copy_to_user(buf, ptr + off, bytes);
2407                 kunmap(p);
2408                 if (r)
2409                         return -EFAULT;
2410
2411                 size -= bytes;
2412                 *pos += bytes;
2413                 result += bytes;
2414         }
2415
2416         return result;
2417 }
2418
2419 /*
2420  * amdgpu_iomem_write - Virtual write access to GPU mapped memory
2421  *
2422  * This function is used to write memory that has been mapped to the
2423  * GPU and the known addresses are not physical addresses but instead
2424  * bus addresses (e.g., what you'd put in an IB or ring buffer).
2425  */
2426 static ssize_t amdgpu_iomem_write(struct file *f, const char __user *buf,
2427                                  size_t size, loff_t *pos)
2428 {
2429         struct amdgpu_device *adev = file_inode(f)->i_private;
2430         struct iommu_domain *dom;
2431         ssize_t result = 0;
2432         int r;
2433
2434         dom = iommu_get_domain_for_dev(adev->dev);
2435
2436         while (size) {
2437                 phys_addr_t addr = *pos & PAGE_MASK;
2438                 loff_t off = *pos & ~PAGE_MASK;
2439                 size_t bytes = PAGE_SIZE - off;
2440                 unsigned long pfn;
2441                 struct page *p;
2442                 void *ptr;
2443
2444                 bytes = bytes < size ? bytes : size;
2445
2446                 addr = dom ? iommu_iova_to_phys(dom, addr) : addr;
2447
2448                 pfn = addr >> PAGE_SHIFT;
2449                 if (!pfn_valid(pfn))
2450                         return -EPERM;
2451
2452                 p = pfn_to_page(pfn);
2453                 if (p->mapping != adev->mman.bdev.dev_mapping)
2454                         return -EPERM;
2455
2456                 ptr = kmap(p);
2457                 r = copy_from_user(ptr + off, buf, bytes);
2458                 kunmap(p);
2459                 if (r)
2460                         return -EFAULT;
2461
2462                 size -= bytes;
2463                 *pos += bytes;
2464                 result += bytes;
2465         }
2466
2467         return result;
2468 }
2469
2470 static const struct file_operations amdgpu_ttm_iomem_fops = {
2471         .owner = THIS_MODULE,
2472         .read = amdgpu_iomem_read,
2473         .write = amdgpu_iomem_write,
2474         .llseek = default_llseek
2475 };
2476
2477 static const struct {
2478         char *name;
2479         const struct file_operations *fops;
2480         int domain;
2481 } ttm_debugfs_entries[] = {
2482         { "amdgpu_vram", &amdgpu_ttm_vram_fops, TTM_PL_VRAM },
2483 #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS
2484         { "amdgpu_gtt", &amdgpu_ttm_gtt_fops, TTM_PL_TT },
2485 #endif
2486         { "amdgpu_iomem", &amdgpu_ttm_iomem_fops, TTM_PL_SYSTEM },
2487 };
2488
2489 #endif
2490
2491 int amdgpu_ttm_debugfs_init(struct amdgpu_device *adev)
2492 {
2493 #if defined(CONFIG_DEBUG_FS)
2494         unsigned count;
2495
2496         struct drm_minor *minor = adev_to_drm(adev)->primary;
2497         struct dentry *ent, *root = minor->debugfs_root;
2498
2499         for (count = 0; count < ARRAY_SIZE(ttm_debugfs_entries); count++) {
2500                 ent = debugfs_create_file(
2501                                 ttm_debugfs_entries[count].name,
2502                                 S_IFREG | S_IRUGO, root,
2503                                 adev,
2504                                 ttm_debugfs_entries[count].fops);
2505                 if (IS_ERR(ent))
2506                         return PTR_ERR(ent);
2507                 if (ttm_debugfs_entries[count].domain == TTM_PL_VRAM)
2508                         i_size_write(ent->d_inode, adev->gmc.mc_vram_size);
2509                 else if (ttm_debugfs_entries[count].domain == TTM_PL_TT)
2510                         i_size_write(ent->d_inode, adev->gmc.gart_size);
2511                 adev->mman.debugfs_entries[count] = ent;
2512         }
2513
2514         count = ARRAY_SIZE(amdgpu_ttm_debugfs_list);
2515         return amdgpu_debugfs_add_files(adev, amdgpu_ttm_debugfs_list, count);
2516 #else
2517         return 0;
2518 #endif
2519 }