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 #include <drm/ttm/ttm_module.h>
  50 #include <drm/ttm/ttm_page_alloc.h>
  51
  52 #include <drm/drm_debugfs.h>
  53 #include <drm/amdgpu_drm.h>
  54
  55 #include "amdgpu.h"
  56 #include "amdgpu_object.h"
  57 #include "amdgpu_trace.h"
  58 #include "amdgpu_amdkfd.h"
  59 #include "amdgpu_sdma.h"
  60 #include "amdgpu_ras.h"
  61 #include "amdgpu_atomfirmware.h"
  62 #include "bif/bif_4_1_d.h"
  63
  64 #define AMDGPU_TTM_VRAM_MAX_DW_READ     (size_t)128
  65
  66
  67 /**
  68  * amdgpu_init_mem_type - Initialize a memory manager for a specific type of
  69  * memory request.
  70  *
  71  * @bdev: The TTM BO device object (contains a reference to amdgpu_device)
  72  * @type: The type of memory requested
  73  * @man: The memory type manager for each domain
  74  *
  75  * This is called by ttm_bo_init_mm() when a buffer object is being
  76  * initialized.
  77  */
  78 static int amdgpu_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
  79                                 struct ttm_mem_type_manager *man)
  80 {
  81         struct amdgpu_device *adev;
  82
  83         adev = amdgpu_ttm_adev(bdev);
  84
  85         switch (type) {
  86         case TTM_PL_SYSTEM:
  87                 /* System memory */
  88                 man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
  89                 man->available_caching = TTM_PL_MASK_CACHING;
  90                 man->default_caching = TTM_PL_FLAG_CACHED;
  91                 break;
  92         case TTM_PL_TT:
  93                 /* GTT memory  */
  94                 man->func = &amdgpu_gtt_mgr_func;
  95                 man->available_caching = TTM_PL_MASK_CACHING;
  96                 man->default_caching = TTM_PL_FLAG_CACHED;
  97                 man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
  98                 break;
  99         case TTM_PL_VRAM:
 100                 /* "On-card" video ram */
 101                 man->func = &amdgpu_vram_mgr_func;
 102                 man->flags = TTM_MEMTYPE_FLAG_FIXED |
 103                              TTM_MEMTYPE_FLAG_MAPPABLE;
 104                 man->available_caching = TTM_PL_FLAG_UNCACHED | TTM_PL_FLAG_WC;
 105                 man->default_caching = TTM_PL_FLAG_WC;
 106                 break;
 107         case AMDGPU_PL_GDS:
 108         case AMDGPU_PL_GWS:
 109         case AMDGPU_PL_OA:
 110                 /* On-chip GDS memory*/
 111                 man->func = &ttm_bo_manager_func;
 112                 man->flags = TTM_MEMTYPE_FLAG_FIXED;
 113                 man->available_caching = TTM_PL_FLAG_UNCACHED;
 114                 man->default_caching = TTM_PL_FLAG_UNCACHED;
 115                 break;
 116         default:
 117                 DRM_ERROR("Unsupported memory type %u\n", (unsigned)type);
 118                 return -EINVAL;
 119         }
 120         return 0;
 121 }
 122
 123 /**
 124  * amdgpu_evict_flags - Compute placement flags
 125  *
 126  * @bo: The buffer object to evict
 127  * @placement: Possible destination(s) for evicted BO
 128  *
 129  * Fill in placement data when ttm_bo_evict() is called
 130  */
 131 static void amdgpu_evict_flags(struct ttm_buffer_object *bo,
 132                                 struct ttm_placement *placement)
 133 {
 134         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
 135         struct amdgpu_bo *abo;
 136         static const struct ttm_place placements = {
 137                 .fpfn = 0,
 138                 .lpfn = 0,
 139                 .flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM
 140         };
 141
 142         /* Don't handle scatter gather BOs */
 143         if (bo->type == ttm_bo_type_sg) {
 144                 placement->num_placement = 0;
 145                 placement->num_busy_placement = 0;
 146                 return;
 147         }
 148
 149         /* Object isn't an AMDGPU object so ignore */
 150         if (!amdgpu_bo_is_amdgpu_bo(bo)) {
 151                 placement->placement = &placements;
 152                 placement->busy_placement = &placements;
 153                 placement->num_placement = 1;
 154                 placement->num_busy_placement = 1;
 155                 return;
 156         }
 157
 158         abo = ttm_to_amdgpu_bo(bo);
 159         switch (bo->mem.mem_type) {
 160         case AMDGPU_PL_GDS:
 161         case AMDGPU_PL_GWS:
 162         case AMDGPU_PL_OA:
 163                 placement->num_placement = 0;
 164                 placement->num_busy_placement = 0;
 165                 return;
 166
 167         case TTM_PL_VRAM:
 168                 if (!adev->mman.buffer_funcs_enabled) {
 169                         /* Move to system memory */
 170                         amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_CPU);
 171                 } else if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
 172                            !(abo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED) &&
 173                            amdgpu_bo_in_cpu_visible_vram(abo)) {
 174
 175                         /* Try evicting to the CPU inaccessible part of VRAM
 176                          * first, but only set GTT as busy placement, so this
 177                          * BO will be evicted to GTT rather than causing other
 178                          * BOs to be evicted from VRAM
 179                          */
 180                         amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM |
 181                                                          AMDGPU_GEM_DOMAIN_GTT);
 182                         abo->placements[0].fpfn = adev->gmc.visible_vram_size >> PAGE_SHIFT;
 183                         abo->placements[0].lpfn = 0;
 184                         abo->placement.busy_placement = &abo->placements[1];
 185                         abo->placement.num_busy_placement = 1;
 186                 } else {
 187                         /* Move to GTT memory */
 188                         amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_GTT);
 189                 }
 190                 break;
 191         case TTM_PL_TT:
 192         default:
 193                 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_CPU);
 194                 break;
 195         }
 196         *placement = abo->placement;
 197 }
 198
 199 /**
 200  * amdgpu_verify_access - Verify access for a mmap call
 201  *
 202  * @bo: The buffer object to map
 203  * @filp: The file pointer from the process performing the mmap
 204  *
 205  * This is called by ttm_bo_mmap() to verify whether a process
 206  * has the right to mmap a BO to their process space.
 207  */
 208 static int amdgpu_verify_access(struct ttm_buffer_object *bo, struct file *filp)
 209 {
 210         struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
 211
 212         /*
 213          * Don't verify access for KFD BOs. They don't have a GEM
 214          * object associated with them.
 215          */
 216         if (abo->kfd_bo)
 217                 return 0;
 218
 219         if (amdgpu_ttm_tt_get_usermm(bo->ttm))
 220                 return -EPERM;
 221         return drm_vma_node_verify_access(&abo->tbo.base.vma_node,
 222                                           filp->private_data);
 223 }
 224
 225 /**
 226  * amdgpu_move_null - Register memory for a buffer object
 227  *
 228  * @bo: The bo to assign the memory to
 229  * @new_mem: The memory to be assigned.
 230  *
 231  * Assign the memory from new_mem to the memory of the buffer object bo.
 232  */
 233 static void amdgpu_move_null(struct ttm_buffer_object *bo,
 234                              struct ttm_mem_reg *new_mem)
 235 {
 236         struct ttm_mem_reg *old_mem = &bo->mem;
 237
 238         BUG_ON(old_mem->mm_node != NULL);
 239         *old_mem = *new_mem;
 240         new_mem->mm_node = NULL;
 241 }
 242
 243 /**
 244  * amdgpu_mm_node_addr - Compute the GPU relative offset of a GTT buffer.
 245  *
 246  * @bo: The bo to assign the memory to.
 247  * @mm_node: Memory manager node for drm allocator.
 248  * @mem: The region where the bo resides.
 249  *
 250  */
 251 static uint64_t amdgpu_mm_node_addr(struct ttm_buffer_object *bo,
 252                                     struct drm_mm_node *mm_node,
 253                                     struct ttm_mem_reg *mem)
 254 {
 255         uint64_t addr = 0;
 256
 257         if (mm_node->start != AMDGPU_BO_INVALID_OFFSET) {
 258                 addr = mm_node->start << PAGE_SHIFT;
 259                 addr += amdgpu_ttm_domain_start(amdgpu_ttm_adev(bo->bdev),
 260                                                 mem->mem_type);
 261         }
 262         return addr;
 263 }
 264
 265 /**
 266  * amdgpu_find_mm_node - Helper function finds the drm_mm_node corresponding to
 267  * @offset. It also modifies the offset to be within the drm_mm_node returned
 268  *
 269  * @mem: The region where the bo resides.
 270  * @offset: The offset that drm_mm_node is used for finding.
 271  *
 272  */
 273 static struct drm_mm_node *amdgpu_find_mm_node(struct ttm_mem_reg *mem,
 274                                                uint64_t *offset)
 275 {
 276         struct drm_mm_node *mm_node = mem->mm_node;
 277
 278         while (*offset >= (mm_node->size << PAGE_SHIFT)) {
 279                 *offset -= (mm_node->size << PAGE_SHIFT);
 280                 ++mm_node;
 281         }
 282         return mm_node;
 283 }
 284
 285 /**
 286  * amdgpu_ttm_map_buffer - Map memory into the GART windows
 287  * @bo: buffer object to map
 288  * @mem: memory object to map
 289  * @mm_node: drm_mm node object to map
 290  * @num_pages: number of pages to map
 291  * @offset: offset into @mm_node where to start
 292  * @window: which GART window to use
 293  * @ring: DMA ring to use for the copy
 294  * @tmz: if we should setup a TMZ enabled mapping
 295  * @addr: resulting address inside the MC address space
 296  *
 297  * Setup one of the GART windows to access a specific piece of memory or return
 298  * the physical address for local memory.
 299  */
 300 static int amdgpu_ttm_map_buffer(struct ttm_buffer_object *bo,
 301                                  struct ttm_mem_reg *mem,
 302                                  struct drm_mm_node *mm_node,
 303                                  unsigned num_pages, uint64_t offset,
 304                                  unsigned window, struct amdgpu_ring *ring,
 305                                  bool tmz, uint64_t *addr)
 306 {
 307         struct amdgpu_device *adev = ring->adev;
 308         struct amdgpu_job *job;
 309         unsigned num_dw, num_bytes;
 310         struct dma_fence *fence;
 311         uint64_t src_addr, dst_addr;
 312         void *cpu_addr;
 313         uint64_t flags;
 314         unsigned int i;
 315         int r;
 316
 317         BUG_ON(adev->mman.buffer_funcs->copy_max_bytes <
 318                AMDGPU_GTT_MAX_TRANSFER_SIZE * 8);
 319
 320         /* Map only what can't be accessed directly */
 321         if (!tmz && mem->start != AMDGPU_BO_INVALID_OFFSET) {
 322                 *addr = amdgpu_mm_node_addr(bo, mm_node, mem) + offset;
 323                 return 0;
 324         }
 325
 326         *addr = adev->gmc.gart_start;
 327         *addr += (u64)window * AMDGPU_GTT_MAX_TRANSFER_SIZE *
 328                 AMDGPU_GPU_PAGE_SIZE;
 329         *addr += offset & ~PAGE_MASK;
 330
 331         num_dw = ALIGN(adev->mman.buffer_funcs->copy_num_dw, 8);
 332         num_bytes = num_pages * 8;
 333
 334         r = amdgpu_job_alloc_with_ib(adev, num_dw * 4 + num_bytes,
 335                                      AMDGPU_IB_POOL_DELAYED, &job);
 336         if (r)
 337                 return r;
 338
 339         src_addr = num_dw * 4;
 340         src_addr += job->ibs[0].gpu_addr;
 341
 342         dst_addr = amdgpu_bo_gpu_offset(adev->gart.bo);
 343         dst_addr += window * AMDGPU_GTT_MAX_TRANSFER_SIZE * 8;
 344         amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_addr,
 345                                 dst_addr, num_bytes, false);
 346
 347         amdgpu_ring_pad_ib(ring, &job->ibs[0]);
 348         WARN_ON(job->ibs[0].length_dw > num_dw);
 349
 350         flags = amdgpu_ttm_tt_pte_flags(adev, bo->ttm, mem);
 351         if (tmz)
 352                 flags |= AMDGPU_PTE_TMZ;
 353
 354         cpu_addr = &job->ibs[0].ptr[num_dw];
 355
 356         if (mem->mem_type == TTM_PL_TT) {
 357                 struct ttm_dma_tt *dma;
 358                 dma_addr_t *dma_address;
 359
 360                 dma = container_of(bo->ttm, struct ttm_dma_tt, ttm);
 361                 dma_address = &dma->dma_address[offset >> PAGE_SHIFT];
 362                 r = amdgpu_gart_map(adev, 0, num_pages, dma_address, flags,
 363                                     cpu_addr);
 364                 if (r)
 365                         goto error_free;
 366         } else {
 367                 dma_addr_t dma_address;
 368
 369                 dma_address = (mm_node->start << PAGE_SHIFT) + offset;
 370                 dma_address += adev->vm_manager.vram_base_offset;
 371
 372                 for (i = 0; i < num_pages; ++i) {
 373                         r = amdgpu_gart_map(adev, i << PAGE_SHIFT, 1,
 374                                             &dma_address, flags, cpu_addr);
 375                         if (r)
 376                                 goto error_free;
 377
 378                         dma_address += PAGE_SIZE;
 379                 }
 380         }
 381
 382         r = amdgpu_job_submit(job, &adev->mman.entity,
 383                               AMDGPU_FENCE_OWNER_UNDEFINED, &fence);
 384         if (r)
 385                 goto error_free;
 386
 387         dma_fence_put(fence);
 388
 389         return r;
 390
 391 error_free:
 392         amdgpu_job_free(job);
 393         return r;
 394 }
 395
 396 /**
 397  * amdgpu_copy_ttm_mem_to_mem - Helper function for copy
 398  * @adev: amdgpu device
 399  * @src: buffer/address where to read from
 400  * @dst: buffer/address where to write to
 401  * @size: number of bytes to copy
 402  * @tmz: if a secure copy should be used
 403  * @resv: resv object to sync to
 404  * @f: Returns the last fence if multiple jobs are submitted.
 405  *
 406  * The function copies @size bytes from {src->mem + src->offset} to
 407  * {dst->mem + dst->offset}. src->bo and dst->bo could be same BO for a
 408  * move and different for a BO to BO copy.
 409  *
 410  */
 411 int amdgpu_ttm_copy_mem_to_mem(struct amdgpu_device *adev,
 412                                const struct amdgpu_copy_mem *src,
 413                                const struct amdgpu_copy_mem *dst,
 414                                uint64_t size, bool tmz,
 415                                struct dma_resv *resv,
 416                                struct dma_fence **f)
 417 {
 418         const uint32_t GTT_MAX_BYTES = (AMDGPU_GTT_MAX_TRANSFER_SIZE *
 419                                         AMDGPU_GPU_PAGE_SIZE);
 420
 421         uint64_t src_node_size, dst_node_size, src_offset, dst_offset;
 422         struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
 423         struct drm_mm_node *src_mm, *dst_mm;
 424         struct dma_fence *fence = NULL;
 425         int r = 0;
 426
 427         if (!adev->mman.buffer_funcs_enabled) {
 428                 DRM_ERROR("Trying to move memory with ring turned off.\n");
 429                 return -EINVAL;
 430         }
 431
 432         src_offset = src->offset;
 433         if (src->mem->mm_node) {
 434                 src_mm = amdgpu_find_mm_node(src->mem, &src_offset);
 435                 src_node_size = (src_mm->size << PAGE_SHIFT) - src_offset;
 436         } else {
 437                 src_mm = NULL;
 438                 src_node_size = ULLONG_MAX;
 439         }
 440
 441         dst_offset = dst->offset;
 442         if (dst->mem->mm_node) {
 443                 dst_mm = amdgpu_find_mm_node(dst->mem, &dst_offset);
 444                 dst_node_size = (dst_mm->size << PAGE_SHIFT) - dst_offset;
 445         } else {
 446                 dst_mm = NULL;
 447                 dst_node_size = ULLONG_MAX;
 448         }
 449
 450         mutex_lock(&adev->mman.gtt_window_lock);
 451
 452         while (size) {
 453                 uint32_t src_page_offset = src_offset & ~PAGE_MASK;
 454                 uint32_t dst_page_offset = dst_offset & ~PAGE_MASK;
 455                 struct dma_fence *next;
 456                 uint32_t cur_size;
 457                 uint64_t from, to;
 458
 459                 /* Copy size cannot exceed GTT_MAX_BYTES. So if src or dst
 460                  * begins at an offset, then adjust the size accordingly
 461                  */
 462                 cur_size = max(src_page_offset, dst_page_offset);
 463                 cur_size = min(min3(src_node_size, dst_node_size, size),
 464                                (uint64_t)(GTT_MAX_BYTES - cur_size));
 465
 466                 /* Map src to window 0 and dst to window 1. */
 467                 r = amdgpu_ttm_map_buffer(src->bo, src->mem, src_mm,
 468                                           PFN_UP(cur_size + src_page_offset),
 469                                           src_offset, 0, ring, tmz, &from);
 470                 if (r)
 471                         goto error;
 472
 473                 r = amdgpu_ttm_map_buffer(dst->bo, dst->mem, dst_mm,
 474                                           PFN_UP(cur_size + dst_page_offset),
 475                                           dst_offset, 1, ring, tmz, &to);
 476                 if (r)
 477                         goto error;
 478
 479                 r = amdgpu_copy_buffer(ring, from, to, cur_size,
 480                                        resv, &next, false, true, tmz);
 481                 if (r)
 482                         goto error;
 483
 484                 dma_fence_put(fence);
 485                 fence = next;
 486
 487                 size -= cur_size;
 488                 if (!size)
 489                         break;
 490
 491                 src_node_size -= cur_size;
 492                 if (!src_node_size) {
 493                         ++src_mm;
 494                         src_node_size = src_mm->size << PAGE_SHIFT;
 495                         src_offset = 0;
 496                 } else {
 497                         src_offset += cur_size;
 498                 }
 499
 500                 dst_node_size -= cur_size;
 501                 if (!dst_node_size) {
 502                         ++dst_mm;
 503                         dst_node_size = dst_mm->size << PAGE_SHIFT;
 504                         dst_offset = 0;
 505                 } else {
 506                         dst_offset += cur_size;
 507                 }
 508         }
 509 error:
 510         mutex_unlock(&adev->mman.gtt_window_lock);
 511         if (f)
 512                 *f = dma_fence_get(fence);
 513         dma_fence_put(fence);
 514         return r;
 515 }
 516
 517 /**
 518  * amdgpu_move_blit - Copy an entire buffer to another buffer
 519  *
 520  * This is a helper called by amdgpu_bo_move() and amdgpu_move_vram_ram() to
 521  * help move buffers to and from VRAM.
 522  */
 523 static int amdgpu_move_blit(struct ttm_buffer_object *bo,
 524                             bool evict, bool no_wait_gpu,
 525                             struct ttm_mem_reg *new_mem,
 526                             struct ttm_mem_reg *old_mem)
 527 {
 528         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
 529         struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
 530         struct amdgpu_copy_mem src, dst;
 531         struct dma_fence *fence = NULL;
 532         int r;
 533
 534         src.bo = bo;
 535         dst.bo = bo;
 536         src.mem = old_mem;
 537         dst.mem = new_mem;
 538         src.offset = 0;
 539         dst.offset = 0;
 540
 541         r = amdgpu_ttm_copy_mem_to_mem(adev, &src, &dst,
 542                                        new_mem->num_pages << PAGE_SHIFT,
 543                                        amdgpu_bo_encrypted(abo),
 544                                        bo->base.resv, &fence);
 545         if (r)
 546                 goto error;
 547
 548         /* clear the space being freed */
 549         if (old_mem->mem_type == TTM_PL_VRAM &&
 550             (abo->flags & AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE)) {
 551                 struct dma_fence *wipe_fence = NULL;
 552
 553                 r = amdgpu_fill_buffer(ttm_to_amdgpu_bo(bo), AMDGPU_POISON,
 554                                        NULL, &wipe_fence);
 555                 if (r) {
 556                         goto error;
 557                 } else if (wipe_fence) {
 558                         dma_fence_put(fence);
 559                         fence = wipe_fence;
 560                 }
 561         }
 562
 563         /* Always block for VM page tables before committing the new location */
 564         if (bo->type == ttm_bo_type_kernel)
 565                 r = ttm_bo_move_accel_cleanup(bo, fence, true, new_mem);
 566         else
 567                 r = ttm_bo_pipeline_move(bo, fence, evict, new_mem);
 568         dma_fence_put(fence);
 569         return r;
 570
 571 error:
 572         if (fence)
 573                 dma_fence_wait(fence, false);
 574         dma_fence_put(fence);
 575         return r;
 576 }
 577
 578 /**
 579  * amdgpu_move_vram_ram - Copy VRAM buffer to RAM buffer
 580  *
 581  * Called by amdgpu_bo_move().
 582  */
 583 static int amdgpu_move_vram_ram(struct ttm_buffer_object *bo, bool evict,
 584                                 struct ttm_operation_ctx *ctx,
 585                                 struct ttm_mem_reg *new_mem)
 586 {
 587         struct ttm_mem_reg *old_mem = &bo->mem;
 588         struct ttm_mem_reg tmp_mem;
 589         struct ttm_place placements;
 590         struct ttm_placement placement;
 591         int r;
 592
 593         /* create space/pages for new_mem in GTT space */
 594         tmp_mem = *new_mem;
 595         tmp_mem.mm_node = NULL;
 596         placement.num_placement = 1;
 597         placement.placement = &placements;
 598         placement.num_busy_placement = 1;
 599         placement.busy_placement = &placements;
 600         placements.fpfn = 0;
 601         placements.lpfn = 0;
 602         placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
 603         r = ttm_bo_mem_space(bo, &placement, &tmp_mem, ctx);
 604         if (unlikely(r)) {
 605                 pr_err("Failed to find GTT space for blit from VRAM\n");
 606                 return r;
 607         }
 608
 609         /* set caching flags */
 610         r = ttm_tt_set_placement_caching(bo->ttm, tmp_mem.placement);
 611         if (unlikely(r)) {
 612                 goto out_cleanup;
 613         }
 614
 615         /* Bind the memory to the GTT space */
 616         r = ttm_tt_bind(bo->ttm, &tmp_mem, ctx);
 617         if (unlikely(r)) {
 618                 goto out_cleanup;
 619         }
 620
 621         /* blit VRAM to GTT */
 622         r = amdgpu_move_blit(bo, evict, ctx->no_wait_gpu, &tmp_mem, old_mem);
 623         if (unlikely(r)) {
 624                 goto out_cleanup;
 625         }
 626
 627         /* move BO (in tmp_mem) to new_mem */
 628         r = ttm_bo_move_ttm(bo, ctx, new_mem);
 629 out_cleanup:
 630         ttm_bo_mem_put(bo, &tmp_mem);
 631         return r;
 632 }
 633
 634 /**
 635  * amdgpu_move_ram_vram - Copy buffer from RAM to VRAM
 636  *
 637  * Called by amdgpu_bo_move().
 638  */
 639 static int amdgpu_move_ram_vram(struct ttm_buffer_object *bo, bool evict,
 640                                 struct ttm_operation_ctx *ctx,
 641                                 struct ttm_mem_reg *new_mem)
 642 {
 643         struct ttm_mem_reg *old_mem = &bo->mem;
 644         struct ttm_mem_reg tmp_mem;
 645         struct ttm_placement placement;
 646         struct ttm_place placements;
 647         int r;
 648
 649         /* make space in GTT for old_mem buffer */
 650         tmp_mem = *new_mem;
 651         tmp_mem.mm_node = NULL;
 652         placement.num_placement = 1;
 653         placement.placement = &placements;
 654         placement.num_busy_placement = 1;
 655         placement.busy_placement = &placements;
 656         placements.fpfn = 0;
 657         placements.lpfn = 0;
 658         placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
 659         r = ttm_bo_mem_space(bo, &placement, &tmp_mem, ctx);
 660         if (unlikely(r)) {
 661                 pr_err("Failed to find GTT space for blit to VRAM\n");
 662                 return r;
 663         }
 664
 665         /* move/bind old memory to GTT space */
 666         r = ttm_bo_move_ttm(bo, ctx, &tmp_mem);
 667         if (unlikely(r)) {
 668                 goto out_cleanup;
 669         }
 670
 671         /* copy to VRAM */
 672         r = amdgpu_move_blit(bo, evict, ctx->no_wait_gpu, new_mem, old_mem);
 673         if (unlikely(r)) {
 674                 goto out_cleanup;
 675         }
 676 out_cleanup:
 677         ttm_bo_mem_put(bo, &tmp_mem);
 678         return r;
 679 }
 680
 681 /**
 682  * amdgpu_mem_visible - Check that memory can be accessed by ttm_bo_move_memcpy
 683  *
 684  * Called by amdgpu_bo_move()
 685  */
 686 static bool amdgpu_mem_visible(struct amdgpu_device *adev,
 687                                struct ttm_mem_reg *mem)
 688 {
 689         struct drm_mm_node *nodes = mem->mm_node;
 690
 691         if (mem->mem_type == TTM_PL_SYSTEM ||
 692             mem->mem_type == TTM_PL_TT)
 693                 return true;
 694         if (mem->mem_type != TTM_PL_VRAM)
 695                 return false;
 696
 697         /* ttm_mem_reg_ioremap only supports contiguous memory */
 698         if (nodes->size != mem->num_pages)
 699                 return false;
 700
 701         return ((nodes->start + nodes->size) << PAGE_SHIFT)
 702                 <= adev->gmc.visible_vram_size;
 703 }
 704
 705 /**
 706  * amdgpu_bo_move - Move a buffer object to a new memory location
 707  *
 708  * Called by ttm_bo_handle_move_mem()
 709  */
 710 static int amdgpu_bo_move(struct ttm_buffer_object *bo, bool evict,
 711                           struct ttm_operation_ctx *ctx,
 712                           struct ttm_mem_reg *new_mem)
 713 {
 714         struct amdgpu_device *adev;
 715         struct amdgpu_bo *abo;
 716         struct ttm_mem_reg *old_mem = &bo->mem;
 717         int r;
 718
 719         /* Can't move a pinned BO */
 720         abo = ttm_to_amdgpu_bo(bo);
 721         if (WARN_ON_ONCE(abo->pin_count > 0))
 722                 return -EINVAL;
 723
 724         adev = amdgpu_ttm_adev(bo->bdev);
 725
 726         if (old_mem->mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
 727                 amdgpu_move_null(bo, new_mem);
 728                 return 0;
 729         }
 730         if ((old_mem->mem_type == TTM_PL_TT &&
 731              new_mem->mem_type == TTM_PL_SYSTEM) ||
 732             (old_mem->mem_type == TTM_PL_SYSTEM &&
 733              new_mem->mem_type == TTM_PL_TT)) {
 734                 /* bind is enough */
 735                 amdgpu_move_null(bo, new_mem);
 736                 return 0;
 737         }
 738         if (old_mem->mem_type == AMDGPU_PL_GDS ||
 739             old_mem->mem_type == AMDGPU_PL_GWS ||
 740             old_mem->mem_type == AMDGPU_PL_OA ||
 741             new_mem->mem_type == AMDGPU_PL_GDS ||
 742             new_mem->mem_type == AMDGPU_PL_GWS ||
 743             new_mem->mem_type == AMDGPU_PL_OA) {
 744                 /* Nothing to save here */
 745                 amdgpu_move_null(bo, new_mem);
 746                 return 0;
 747         }
 748
 749         if (!adev->mman.buffer_funcs_enabled) {
 750                 r = -ENODEV;
 751                 goto memcpy;
 752         }
 753
 754         if (old_mem->mem_type == TTM_PL_VRAM &&
 755             new_mem->mem_type == TTM_PL_SYSTEM) {
 756                 r = amdgpu_move_vram_ram(bo, evict, ctx, new_mem);
 757         } else if (old_mem->mem_type == TTM_PL_SYSTEM &&
 758                    new_mem->mem_type == TTM_PL_VRAM) {
 759                 r = amdgpu_move_ram_vram(bo, evict, ctx, new_mem);
 760         } else {
 761                 r = amdgpu_move_blit(bo, evict, ctx->no_wait_gpu,
 762                                      new_mem, old_mem);
 763         }
 764
 765         if (r) {
 766 memcpy:
 767                 /* Check that all memory is CPU accessible */
 768                 if (!amdgpu_mem_visible(adev, old_mem) ||
 769                     !amdgpu_mem_visible(adev, new_mem)) {
 770                         pr_err("Move buffer fallback to memcpy unavailable\n");
 771                         return r;
 772                 }
 773
 774                 r = ttm_bo_move_memcpy(bo, ctx, new_mem);
 775                 if (r)
 776                         return r;
 777         }
 778
 779         if (bo->type == ttm_bo_type_device &&
 780             new_mem->mem_type == TTM_PL_VRAM &&
 781             old_mem->mem_type != TTM_PL_VRAM) {
 782                 /* amdgpu_bo_fault_reserve_notify will re-set this if the CPU
 783                  * accesses the BO after it's moved.
 784                  */
 785                 abo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
 786         }
 787
 788         /* update statistics */
 789         atomic64_add((u64)bo->num_pages << PAGE_SHIFT, &adev->num_bytes_moved);
 790         return 0;
 791 }
 792
 793 /**
 794  * amdgpu_ttm_io_mem_reserve - Reserve a block of memory during a fault
 795  *
 796  * Called by ttm_mem_io_reserve() ultimately via ttm_bo_vm_fault()
 797  */
 798 static int amdgpu_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
 799 {
 800         struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
 801         struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
 802         struct drm_mm_node *mm_node = mem->mm_node;
 803
 804         mem->bus.addr = NULL;
 805         mem->bus.offset = 0;
 806         mem->bus.size = mem->num_pages << PAGE_SHIFT;
 807         mem->bus.base = 0;
 808         mem->bus.is_iomem = false;
 809         if (!(man->flags & TTM_MEMTYPE_FLAG_MAPPABLE))
 810                 return -EINVAL;
 811         switch (mem->mem_type) {
 812         case TTM_PL_SYSTEM:
 813                 /* system memory */
 814                 return 0;
 815         case TTM_PL_TT:
 816                 break;
 817         case TTM_PL_VRAM:
 818                 mem->bus.offset = mem->start << PAGE_SHIFT;
 819                 /* check if it's visible */
 820                 if ((mem->bus.offset + mem->bus.size) > adev->gmc.visible_vram_size)
 821                         return -EINVAL;
 822                 /* Only physically contiguous buffers apply. In a contiguous
 823                  * buffer, size of the first mm_node would match the number of
 824                  * pages in ttm_mem_reg.
 825                  */
 826                 if (adev->mman.aper_base_kaddr &&
 827                     (mm_node->size == mem->num_pages))
 828                         mem->bus.addr = (u8 *)adev->mman.aper_base_kaddr +
 829                                         mem->bus.offset;
 830
 831                 mem->bus.base = adev->gmc.aper_base;
 832                 mem->bus.is_iomem = true;
 833                 break;
 834         default:
 835                 return -EINVAL;
 836         }
 837         return 0;
 838 }
 839
 840 static unsigned long amdgpu_ttm_io_mem_pfn(struct ttm_buffer_object *bo,
 841                                            unsigned long page_offset)
 842 {
 843         uint64_t offset = (page_offset << PAGE_SHIFT);
 844         struct drm_mm_node *mm;
 845
 846         mm = amdgpu_find_mm_node(&bo->mem, &offset);
 847         return (bo->mem.bus.base >> PAGE_SHIFT) + mm->start +
 848                 (offset >> PAGE_SHIFT);
 849 }
 850
 851 /**
 852  * amdgpu_ttm_domain_start - Returns GPU start address
 853  * @adev: amdgpu device object
 854  * @type: type of the memory
 855  *
 856  * Returns:
 857  * GPU start address of a memory domain
 858  */
 859
 860 uint64_t amdgpu_ttm_domain_start(struct amdgpu_device *adev, uint32_t type)
 861 {
 862         switch (type) {
 863         case TTM_PL_TT:
 864                 return adev->gmc.gart_start;
 865         case TTM_PL_VRAM:
 866                 return adev->gmc.vram_start;
 867         }
 868
 869         return 0;
 870 }
 871
 872 /*
 873  * TTM backend functions.
 874  */
 875 struct amdgpu_ttm_tt {
 876         struct ttm_dma_tt       ttm;
 877         struct drm_gem_object   *gobj;
 878         u64                     offset;
 879         uint64_t                userptr;
 880         struct task_struct      *usertask;
 881         uint32_t                userflags;
 882 #if IS_ENABLED(CONFIG_DRM_AMDGPU_USERPTR)
 883         struct hmm_range        *range;
 884 #endif
 885 };
 886
 887 #ifdef CONFIG_DRM_AMDGPU_USERPTR
 888 /**
 889  * amdgpu_ttm_tt_get_user_pages - get device accessible pages that back user
 890  * memory and start HMM tracking CPU page table update
 891  *
 892  * Calling function must call amdgpu_ttm_tt_userptr_range_done() once and only
 893  * once afterwards to stop HMM tracking
 894  */
 895 int amdgpu_ttm_tt_get_user_pages(struct amdgpu_bo *bo, struct page **pages)
 896 {
 897         struct ttm_tt *ttm = bo->tbo.ttm;
 898         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 899         unsigned long start = gtt->userptr;
 900         struct vm_area_struct *vma;
 901         struct hmm_range *range;
 902         unsigned long timeout;
 903         struct mm_struct *mm;
 904         unsigned long i;
 905         int r = 0;
 906
 907         mm = bo->notifier.mm;
 908         if (unlikely(!mm)) {
 909                 DRM_DEBUG_DRIVER("BO is not registered?\n");
 910                 return -EFAULT;
 911         }
 912
 913         /* Another get_user_pages is running at the same time?? */
 914         if (WARN_ON(gtt->range))
 915                 return -EFAULT;
 916
 917         if (!mmget_not_zero(mm)) /* Happens during process shutdown */
 918                 return -ESRCH;
 919
 920         range = kzalloc(sizeof(*range), GFP_KERNEL);
 921         if (unlikely(!range)) {
 922                 r = -ENOMEM;
 923                 goto out;
 924         }
 925         range->notifier = &bo->notifier;
 926         range->start = bo->notifier.interval_tree.start;
 927         range->end = bo->notifier.interval_tree.last + 1;
 928         range->default_flags = HMM_PFN_REQ_FAULT;
 929         if (!amdgpu_ttm_tt_is_readonly(ttm))
 930                 range->default_flags |= HMM_PFN_REQ_WRITE;
 931
 932         range->hmm_pfns = kvmalloc_array(ttm->num_pages,
 933                                          sizeof(*range->hmm_pfns), GFP_KERNEL);
 934         if (unlikely(!range->hmm_pfns)) {
 935                 r = -ENOMEM;
 936                 goto out_free_ranges;
 937         }
 938
 939         mmap_read_lock(mm);
 940         vma = find_vma(mm, start);
 941         if (unlikely(!vma || start < vma->vm_start)) {
 942                 r = -EFAULT;
 943                 goto out_unlock;
 944         }
 945         if (unlikely((gtt->userflags & AMDGPU_GEM_USERPTR_ANONONLY) &&
 946                 vma->vm_file)) {
 947                 r = -EPERM;
 948                 goto out_unlock;
 949         }
 950         mmap_read_unlock(mm);
 951         timeout = jiffies + msecs_to_jiffies(HMM_RANGE_DEFAULT_TIMEOUT);
 952
 953 retry:
 954         range->notifier_seq = mmu_interval_read_begin(&bo->notifier);
 955
 956         mmap_read_lock(mm);
 957         r = hmm_range_fault(range);
 958         mmap_read_unlock(mm);
 959         if (unlikely(r)) {
 960                 /*
 961                  * FIXME: This timeout should encompass the retry from
 962                  * mmu_interval_read_retry() as well.
 963                  */
 964                 if (r == -EBUSY && !time_after(jiffies, timeout))
 965                         goto retry;
 966                 goto out_free_pfns;
 967         }
 968
 969         /*
 970          * Due to default_flags, all pages are HMM_PFN_VALID or
 971          * hmm_range_fault() fails. FIXME: The pages cannot be touched outside
 972          * the notifier_lock, and mmu_interval_read_retry() must be done first.
 973          */
 974         for (i = 0; i < ttm->num_pages; i++)
 975                 pages[i] = hmm_pfn_to_page(range->hmm_pfns[i]);
 976
 977         gtt->range = range;
 978         mmput(mm);
 979
 980         return 0;
 981
 982 out_unlock:
 983         mmap_read_unlock(mm);
 984 out_free_pfns:
 985         kvfree(range->hmm_pfns);
 986 out_free_ranges:
 987         kfree(range);
 988 out:
 989         mmput(mm);
 990         return r;
 991 }
 992
 993 /**
 994  * amdgpu_ttm_tt_userptr_range_done - stop HMM track the CPU page table change
 995  * Check if the pages backing this ttm range have been invalidated
 996  *
 997  * Returns: true if pages are still valid
 998  */
 999 bool amdgpu_ttm_tt_get_user_pages_done(struct ttm_tt *ttm)
1000 {
1001         struct amdgpu_ttm_tt *gtt = (void *)ttm;
1002         bool r = false;
1003
1004         if (!gtt || !gtt->userptr)
1005                 return false;
1006
1007         DRM_DEBUG_DRIVER("user_pages_done 0x%llx pages 0x%lx\n",
1008                 gtt->userptr, ttm->num_pages);
1009
1010         WARN_ONCE(!gtt->range || !gtt->range->hmm_pfns,
1011                 "No user pages to check\n");
1012
1013         if (gtt->range) {
1014                 /*
1015                  * FIXME: Must always hold notifier_lock for this, and must
1016                  * not ignore the return code.
1017                  */
1018                 r = mmu_interval_read_retry(gtt->range->notifier,
1019                                          gtt->range->notifier_seq);
1020                 kvfree(gtt->range->hmm_pfns);
1021                 kfree(gtt->range);
1022                 gtt->range = NULL;
1023         }
1024
1025         return !r;
1026 }
1027 #endif
1028
1029 /**
1030  * amdgpu_ttm_tt_set_user_pages - Copy pages in, putting old pages as necessary.
1031  *
1032  * Called by amdgpu_cs_list_validate(). This creates the page list
1033  * that backs user memory and will ultimately be mapped into the device
1034  * address space.
1035  */
1036 void amdgpu_ttm_tt_set_user_pages(struct ttm_tt *ttm, struct page **pages)
1037 {
1038         unsigned long i;
1039
1040         for (i = 0; i < ttm->num_pages; ++i)
1041                 ttm->pages[i] = pages ? pages[i] : NULL;
1042 }
1043
1044 /**
1045  * amdgpu_ttm_tt_pin_userptr -  prepare the sg table with the user pages
1046  *
1047  * Called by amdgpu_ttm_backend_bind()
1048  **/
1049 static int amdgpu_ttm_tt_pin_userptr(struct ttm_tt *ttm)
1050 {
1051         struct amdgpu_device *adev = amdgpu_ttm_adev(ttm->bdev);
1052         struct amdgpu_ttm_tt *gtt = (void *)ttm;
1053         int r;
1054
1055         int write = !(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
1056         enum dma_data_direction direction = write ?
1057                 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
1058
1059         /* Allocate an SG array and squash pages into it */
1060         r = sg_alloc_table_from_pages(ttm->sg, ttm->pages, ttm->num_pages, 0,
1061                                       ttm->num_pages << PAGE_SHIFT,
1062                                       GFP_KERNEL);
1063         if (r)
1064                 goto release_sg;
1065
1066         /* Map SG to device */
1067         r = dma_map_sgtable(adev->dev, ttm->sg, direction, 0);
1068         if (r)
1069                 goto release_sg;
1070
1071         /* convert SG to linear array of pages and dma addresses */
1072         drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
1073                                          gtt->ttm.dma_address, ttm->num_pages);
1074
1075         return 0;
1076
1077 release_sg:
1078         kfree(ttm->sg);
1079         ttm->sg = NULL;
1080         return r;
1081 }
1082
1083 /**
1084  * amdgpu_ttm_tt_unpin_userptr - Unpin and unmap userptr pages
1085  */
1086 static void amdgpu_ttm_tt_unpin_userptr(struct ttm_tt *ttm)
1087 {
1088         struct amdgpu_device *adev = amdgpu_ttm_adev(ttm->bdev);
1089         struct amdgpu_ttm_tt *gtt = (void *)ttm;
1090
1091         int write = !(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
1092         enum dma_data_direction direction = write ?
1093                 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
1094
1095         /* double check that we don't free the table twice */
1096         if (!ttm->sg->sgl)
1097                 return;
1098
1099         /* unmap the pages mapped to the device */
1100         dma_unmap_sgtable(adev->dev, ttm->sg, direction, 0);
1101         sg_free_table(ttm->sg);
1102
1103 #if IS_ENABLED(CONFIG_DRM_AMDGPU_USERPTR)
1104         if (gtt->range) {
1105                 unsigned long i;
1106
1107                 for (i = 0; i < ttm->num_pages; i++) {
1108                         if (ttm->pages[i] !=
1109                             hmm_pfn_to_page(gtt->range->hmm_pfns[i]))
1110                                 break;
1111                 }
1112
1113                 WARN((i == ttm->num_pages), "Missing get_user_page_done\n");
1114         }
1115 #endif
1116 }
1117
1118 static int amdgpu_ttm_gart_bind(struct amdgpu_device *adev,
1119                                 struct ttm_buffer_object *tbo,
1120                                 uint64_t flags)
1121 {
1122         struct amdgpu_bo *abo = ttm_to_amdgpu_bo(tbo);
1123         struct ttm_tt *ttm = tbo->ttm;
1124         struct amdgpu_ttm_tt *gtt = (void *)ttm;
1125         int r;
1126
1127         if (amdgpu_bo_encrypted(abo))
1128                 flags |= AMDGPU_PTE_TMZ;
1129
1130         if (abo->flags & AMDGPU_GEM_CREATE_CP_MQD_GFX9) {
1131                 uint64_t page_idx = 1;
1132
1133                 r = amdgpu_gart_bind(adev, gtt->offset, page_idx,
1134                                 ttm->pages, gtt->ttm.dma_address, flags);
1135                 if (r)
1136                         goto gart_bind_fail;
1137
1138                 /* The memory type of the first page defaults to UC. Now
1139                  * modify the memory type to NC from the second page of
1140                  * the BO onward.
1141                  */
1142                 flags &= ~AMDGPU_PTE_MTYPE_VG10_MASK;
1143                 flags |= AMDGPU_PTE_MTYPE_VG10(AMDGPU_MTYPE_NC);
1144
1145                 r = amdgpu_gart_bind(adev,
1146                                 gtt->offset + (page_idx << PAGE_SHIFT),
1147                                 ttm->num_pages - page_idx,
1148                                 &ttm->pages[page_idx],
1149                                 &(gtt->ttm.dma_address[page_idx]), flags);
1150         } else {
1151                 r = amdgpu_gart_bind(adev, gtt->offset, ttm->num_pages,
1152                                      ttm->pages, gtt->ttm.dma_address, flags);
1153         }
1154
1155 gart_bind_fail:
1156         if (r)
1157                 DRM_ERROR("failed to bind %lu pages at 0x%08llX\n",
1158                           ttm->num_pages, gtt->offset);
1159
1160         return r;
1161 }
1162
1163 /**
1164  * amdgpu_ttm_backend_bind - Bind GTT memory
1165  *
1166  * Called by ttm_tt_bind() on behalf of ttm_bo_handle_move_mem().
1167  * This handles binding GTT memory to the device address space.
1168  */
1169 static int amdgpu_ttm_backend_bind(struct ttm_tt *ttm,
1170                                    struct ttm_mem_reg *bo_mem)
1171 {
1172         struct amdgpu_device *adev = amdgpu_ttm_adev(ttm->bdev);
1173         struct amdgpu_ttm_tt *gtt = (void*)ttm;
1174         uint64_t flags;
1175         int r = 0;
1176
1177         if (gtt->userptr) {
1178                 r = amdgpu_ttm_tt_pin_userptr(ttm);
1179                 if (r) {
1180                         DRM_ERROR("failed to pin userptr\n");
1181                         return r;
1182                 }
1183         }
1184         if (!ttm->num_pages) {
1185                 WARN(1, "nothing to bind %lu pages for mreg %p back %p!\n",
1186                      ttm->num_pages, bo_mem, ttm);
1187         }
1188
1189         if (bo_mem->mem_type == AMDGPU_PL_GDS ||
1190             bo_mem->mem_type == AMDGPU_PL_GWS ||
1191             bo_mem->mem_type == AMDGPU_PL_OA)
1192                 return -EINVAL;
1193
1194         if (!amdgpu_gtt_mgr_has_gart_addr(bo_mem)) {
1195                 gtt->offset = AMDGPU_BO_INVALID_OFFSET;
1196                 return 0;
1197         }
1198
1199         /* compute PTE flags relevant to this BO memory */
1200         flags = amdgpu_ttm_tt_pte_flags(adev, ttm, bo_mem);
1201
1202         /* bind pages into GART page tables */
1203         gtt->offset = (u64)bo_mem->start << PAGE_SHIFT;
1204         r = amdgpu_gart_bind(adev, gtt->offset, ttm->num_pages,
1205                 ttm->pages, gtt->ttm.dma_address, flags);
1206
1207         if (r)
1208                 DRM_ERROR("failed to bind %lu pages at 0x%08llX\n",
1209                           ttm->num_pages, gtt->offset);
1210         return r;
1211 }
1212
1213 /**
1214  * amdgpu_ttm_alloc_gart - Allocate GART memory for buffer object
1215  */
1216 int amdgpu_ttm_alloc_gart(struct ttm_buffer_object *bo)
1217 {
1218         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1219         struct ttm_operation_ctx ctx = { false, false };
1220         struct amdgpu_ttm_tt *gtt = (void*)bo->ttm;
1221         struct ttm_mem_reg tmp;
1222         struct ttm_placement placement;
1223         struct ttm_place placements;
1224         uint64_t addr, flags;
1225         int r;
1226
1227         if (bo->mem.start != AMDGPU_BO_INVALID_OFFSET)
1228                 return 0;
1229
1230         addr = amdgpu_gmc_agp_addr(bo);
1231         if (addr != AMDGPU_BO_INVALID_OFFSET) {
1232                 bo->mem.start = addr >> PAGE_SHIFT;
1233         } else {
1234
1235                 /* allocate GART space */
1236                 tmp = bo->mem;
1237                 tmp.mm_node = NULL;
1238                 placement.num_placement = 1;
1239                 placement.placement = &placements;
1240                 placement.num_busy_placement = 1;
1241                 placement.busy_placement = &placements;
1242                 placements.fpfn = 0;
1243                 placements.lpfn = adev->gmc.gart_size >> PAGE_SHIFT;
1244                 placements.flags = (bo->mem.placement & ~TTM_PL_MASK_MEM) |
1245                         TTM_PL_FLAG_TT;
1246
1247                 r = ttm_bo_mem_space(bo, &placement, &tmp, &ctx);
1248                 if (unlikely(r))
1249                         return r;
1250
1251                 /* compute PTE flags for this buffer object */
1252                 flags = amdgpu_ttm_tt_pte_flags(adev, bo->ttm, &tmp);
1253
1254                 /* Bind pages */
1255                 gtt->offset = (u64)tmp.start << PAGE_SHIFT;
1256                 r = amdgpu_ttm_gart_bind(adev, bo, flags);
1257                 if (unlikely(r)) {
1258                         ttm_bo_mem_put(bo, &tmp);
1259                         return r;
1260                 }
1261
1262                 ttm_bo_mem_put(bo, &bo->mem);
1263                 bo->mem = tmp;
1264         }
1265
1266         return 0;
1267 }
1268
1269 /**
1270  * amdgpu_ttm_recover_gart - Rebind GTT pages
1271  *
1272  * Called by amdgpu_gtt_mgr_recover() from amdgpu_device_reset() to
1273  * rebind GTT pages during a GPU reset.
1274  */
1275 int amdgpu_ttm_recover_gart(struct ttm_buffer_object *tbo)
1276 {
1277         struct amdgpu_device *adev = amdgpu_ttm_adev(tbo->bdev);
1278         uint64_t flags;
1279         int r;
1280
1281         if (!tbo->ttm)
1282                 return 0;
1283
1284         flags = amdgpu_ttm_tt_pte_flags(adev, tbo->ttm, &tbo->mem);
1285         r = amdgpu_ttm_gart_bind(adev, tbo, flags);
1286
1287         return r;
1288 }
1289
1290 /**
1291  * amdgpu_ttm_backend_unbind - Unbind GTT mapped pages
1292  *
1293  * Called by ttm_tt_unbind() on behalf of ttm_bo_move_ttm() and
1294  * ttm_tt_destroy().
1295  */
1296 static void amdgpu_ttm_backend_unbind(struct ttm_tt *ttm)
1297 {
1298         struct amdgpu_device *adev = amdgpu_ttm_adev(ttm->bdev);
1299         struct amdgpu_ttm_tt *gtt = (void *)ttm;
1300         int r;
1301
1302         /* if the pages have userptr pinning then clear that first */
1303         if (gtt->userptr)
1304                 amdgpu_ttm_tt_unpin_userptr(ttm);
1305
1306         if (gtt->offset == AMDGPU_BO_INVALID_OFFSET)
1307                 return;
1308
1309         /* unbind shouldn't be done for GDS/GWS/OA in ttm_bo_clean_mm */
1310         r = amdgpu_gart_unbind(adev, gtt->offset, ttm->num_pages);
1311         if (r)
1312                 DRM_ERROR("failed to unbind %lu pages at 0x%08llX\n",
1313                           gtt->ttm.ttm.num_pages, gtt->offset);
1314 }
1315
1316 static void amdgpu_ttm_backend_destroy(struct ttm_tt *ttm)
1317 {
1318         struct amdgpu_ttm_tt *gtt = (void *)ttm;
1319
1320         if (gtt->usertask)
1321                 put_task_struct(gtt->usertask);
1322
1323         ttm_dma_tt_fini(&gtt->ttm);
1324         kfree(gtt);
1325 }
1326
1327 static struct ttm_backend_func amdgpu_backend_func = {
1328         .bind = &amdgpu_ttm_backend_bind,
1329         .unbind = &amdgpu_ttm_backend_unbind,
1330         .destroy = &amdgpu_ttm_backend_destroy,
1331 };
1332
1333 /**
1334  * amdgpu_ttm_tt_create - Create a ttm_tt object for a given BO
1335  *
1336  * @bo: The buffer object to create a GTT ttm_tt object around
1337  *
1338  * Called by ttm_tt_create().
1339  */
1340 static struct ttm_tt *amdgpu_ttm_tt_create(struct ttm_buffer_object *bo,
1341                                            uint32_t page_flags)
1342 {
1343         struct amdgpu_ttm_tt *gtt;
1344
1345         gtt = kzalloc(sizeof(struct amdgpu_ttm_tt), GFP_KERNEL);
1346         if (gtt == NULL) {
1347                 return NULL;
1348         }
1349         gtt->ttm.ttm.func = &amdgpu_backend_func;
1350         gtt->gobj = &bo->base;
1351
1352         /* allocate space for the uninitialized page entries */
1353         if (ttm_sg_tt_init(&gtt->ttm, bo, page_flags)) {
1354                 kfree(gtt);
1355                 return NULL;
1356         }
1357         return &gtt->ttm.ttm;
1358 }
1359
1360 /**
1361  * amdgpu_ttm_tt_populate - Map GTT pages visible to the device
1362  *
1363  * Map the pages of a ttm_tt object to an address space visible
1364  * to the underlying device.
1365  */
1366 static int amdgpu_ttm_tt_populate(struct ttm_tt *ttm,
1367                         struct ttm_operation_ctx *ctx)
1368 {
1369         struct amdgpu_device *adev = amdgpu_ttm_adev(ttm->bdev);
1370         struct amdgpu_ttm_tt *gtt = (void *)ttm;
1371
1372         /* user pages are bound by amdgpu_ttm_tt_pin_userptr() */
1373         if (gtt && gtt->userptr) {
1374                 ttm->sg = kzalloc(sizeof(struct sg_table), GFP_KERNEL);
1375                 if (!ttm->sg)
1376                         return -ENOMEM;
1377
1378                 ttm->page_flags |= TTM_PAGE_FLAG_SG;
1379                 ttm->state = tt_unbound;
1380                 return 0;
1381         }
1382
1383         if (ttm->page_flags & TTM_PAGE_FLAG_SG) {
1384                 if (!ttm->sg) {
1385                         struct dma_buf_attachment *attach;
1386                         struct sg_table *sgt;
1387
1388                         attach = gtt->gobj->import_attach;
1389                         sgt = dma_buf_map_attachment(attach, DMA_BIDIRECTIONAL);
1390                         if (IS_ERR(sgt))
1391                                 return PTR_ERR(sgt);
1392
1393                         ttm->sg = sgt;
1394                 }
1395
1396                 drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
1397                                                  gtt->ttm.dma_address,
1398                                                  ttm->num_pages);
1399                 ttm->state = tt_unbound;
1400                 return 0;
1401         }
1402
1403 #ifdef CONFIG_SWIOTLB
1404         if (adev->need_swiotlb && swiotlb_nr_tbl()) {
1405                 return ttm_dma_populate(&gtt->ttm, adev->dev, ctx);
1406         }
1407 #endif
1408
1409         /* fall back to generic helper to populate the page array
1410          * and map them to the device */
1411         return ttm_populate_and_map_pages(adev->dev, &gtt->ttm, ctx);
1412 }
1413
1414 /**
1415  * amdgpu_ttm_tt_unpopulate - unmap GTT pages and unpopulate page arrays
1416  *
1417  * Unmaps pages of a ttm_tt object from the device address space and
1418  * unpopulates the page array backing it.
1419  */
1420 static void amdgpu_ttm_tt_unpopulate(struct ttm_tt *ttm)
1421 {
1422         struct amdgpu_ttm_tt *gtt = (void *)ttm;
1423         struct amdgpu_device *adev;
1424
1425         if (gtt && gtt->userptr) {
1426                 amdgpu_ttm_tt_set_user_pages(ttm, NULL);
1427                 kfree(ttm->sg);
1428                 ttm->page_flags &= ~TTM_PAGE_FLAG_SG;
1429                 return;
1430         }
1431
1432         if (ttm->sg && gtt->gobj->import_attach) {
1433                 struct dma_buf_attachment *attach;
1434
1435                 attach = gtt->gobj->import_attach;
1436                 dma_buf_unmap_attachment(attach, ttm->sg, DMA_BIDIRECTIONAL);
1437                 ttm->sg = NULL;
1438                 return;
1439         }
1440
1441         if (ttm->page_flags & TTM_PAGE_FLAG_SG)
1442                 return;
1443
1444         adev = amdgpu_ttm_adev(ttm->bdev);
1445
1446 #ifdef CONFIG_SWIOTLB
1447         if (adev->need_swiotlb && swiotlb_nr_tbl()) {
1448                 ttm_dma_unpopulate(&gtt->ttm, adev->dev);
1449                 return;
1450         }
1451 #endif
1452
1453         /* fall back to generic helper to unmap and unpopulate array */
1454         ttm_unmap_and_unpopulate_pages(adev->dev, &gtt->ttm);
1455 }
1456
1457 /**
1458  * amdgpu_ttm_tt_set_userptr - Initialize userptr GTT ttm_tt for the current
1459  * task
1460  *
1461  * @ttm: The ttm_tt object to bind this userptr object to
1462  * @addr:  The address in the current tasks VM space to use
1463  * @flags: Requirements of userptr object.
1464  *
1465  * Called by amdgpu_gem_userptr_ioctl() to bind userptr pages
1466  * to current task
1467  */
1468 int amdgpu_ttm_tt_set_userptr(struct ttm_tt *ttm, uint64_t addr,
1469                               uint32_t flags)
1470 {
1471         struct amdgpu_ttm_tt *gtt = (void *)ttm;
1472
1473         if (gtt == NULL)
1474                 return -EINVAL;
1475
1476         gtt->userptr = addr;
1477         gtt->userflags = flags;
1478
1479         if (gtt->usertask)
1480                 put_task_struct(gtt->usertask);
1481         gtt->usertask = current->group_leader;
1482         get_task_struct(gtt->usertask);
1483
1484         return 0;
1485 }
1486
1487 /**
1488  * amdgpu_ttm_tt_get_usermm - Return memory manager for ttm_tt object
1489  */
1490 struct mm_struct *amdgpu_ttm_tt_get_usermm(struct ttm_tt *ttm)
1491 {
1492         struct amdgpu_ttm_tt *gtt = (void *)ttm;
1493
1494         if (gtt == NULL)
1495                 return NULL;
1496
1497         if (gtt->usertask == NULL)
1498                 return NULL;
1499
1500         return gtt->usertask->mm;
1501 }
1502
1503 /**
1504  * amdgpu_ttm_tt_affect_userptr - Determine if a ttm_tt object lays inside an
1505  * address range for the current task.
1506  *
1507  */
1508 bool amdgpu_ttm_tt_affect_userptr(struct ttm_tt *ttm, unsigned long start,
1509                                   unsigned long end)
1510 {
1511         struct amdgpu_ttm_tt *gtt = (void *)ttm;
1512         unsigned long size;
1513
1514         if (gtt == NULL || !gtt->userptr)
1515                 return false;
1516
1517         /* Return false if no part of the ttm_tt object lies within
1518          * the range
1519          */
1520         size = (unsigned long)gtt->ttm.ttm.num_pages * PAGE_SIZE;
1521         if (gtt->userptr > end || gtt->userptr + size <= start)
1522                 return false;
1523
1524         return true;
1525 }
1526
1527 /**
1528  * amdgpu_ttm_tt_is_userptr - Have the pages backing by userptr?
1529  */
1530 bool amdgpu_ttm_tt_is_userptr(struct ttm_tt *ttm)
1531 {
1532         struct amdgpu_ttm_tt *gtt = (void *)ttm;
1533
1534         if (gtt == NULL || !gtt->userptr)
1535                 return false;
1536
1537         return true;
1538 }
1539
1540 /**
1541  * amdgpu_ttm_tt_is_readonly - Is the ttm_tt object read only?
1542  */
1543 bool amdgpu_ttm_tt_is_readonly(struct ttm_tt *ttm)
1544 {
1545         struct amdgpu_ttm_tt *gtt = (void *)ttm;
1546
1547         if (gtt == NULL)
1548                 return false;
1549
1550         return !!(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
1551 }
1552
1553 /**
1554  * amdgpu_ttm_tt_pde_flags - Compute PDE flags for ttm_tt object
1555  *
1556  * @ttm: The ttm_tt object to compute the flags for
1557  * @mem: The memory registry backing this ttm_tt object
1558  *
1559  * Figure out the flags to use for a VM PDE (Page Directory Entry).
1560  */
1561 uint64_t amdgpu_ttm_tt_pde_flags(struct ttm_tt *ttm, struct ttm_mem_reg *mem)
1562 {
1563         uint64_t flags = 0;
1564
1565         if (mem && mem->mem_type != TTM_PL_SYSTEM)
1566                 flags |= AMDGPU_PTE_VALID;
1567
1568         if (mem && mem->mem_type == TTM_PL_TT) {
1569                 flags |= AMDGPU_PTE_SYSTEM;
1570
1571                 if (ttm->caching_state == tt_cached)
1572                         flags |= AMDGPU_PTE_SNOOPED;
1573         }
1574
1575         return flags;
1576 }
1577
1578 /**
1579  * amdgpu_ttm_tt_pte_flags - Compute PTE flags for ttm_tt object
1580  *
1581  * @ttm: The ttm_tt object to compute the flags for
1582  * @mem: The memory registry backing this ttm_tt object
1583
1584  * Figure out the flags to use for a VM PTE (Page Table Entry).
1585  */
1586 uint64_t amdgpu_ttm_tt_pte_flags(struct amdgpu_device *adev, struct ttm_tt *ttm,
1587                                  struct ttm_mem_reg *mem)
1588 {
1589         uint64_t flags = amdgpu_ttm_tt_pde_flags(ttm, mem);
1590
1591         flags |= adev->gart.gart_pte_flags;
1592         flags |= AMDGPU_PTE_READABLE;
1593
1594         if (!amdgpu_ttm_tt_is_readonly(ttm))
1595                 flags |= AMDGPU_PTE_WRITEABLE;
1596
1597         return flags;
1598 }
1599
1600 /**
1601  * amdgpu_ttm_bo_eviction_valuable - Check to see if we can evict a buffer
1602  * object.
1603  *
1604  * Return true if eviction is sensible. Called by ttm_mem_evict_first() on
1605  * behalf of ttm_bo_mem_force_space() which tries to evict buffer objects until
1606  * it can find space for a new object and by ttm_bo_force_list_clean() which is
1607  * used to clean out a memory space.
1608  */
1609 static bool amdgpu_ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
1610                                             const struct ttm_place *place)
1611 {
1612         unsigned long num_pages = bo->mem.num_pages;
1613         struct drm_mm_node *node = bo->mem.mm_node;
1614         struct dma_resv_list *flist;
1615         struct dma_fence *f;
1616         int i;
1617
1618         if (bo->type == ttm_bo_type_kernel &&
1619             !amdgpu_vm_evictable(ttm_to_amdgpu_bo(bo)))
1620                 return false;
1621
1622         /* If bo is a KFD BO, check if the bo belongs to the current process.
1623          * If true, then return false as any KFD process needs all its BOs to
1624          * be resident to run successfully
1625          */
1626         flist = dma_resv_get_list(bo->base.resv);
1627         if (flist) {
1628                 for (i = 0; i < flist->shared_count; ++i) {
1629                         f = rcu_dereference_protected(flist->shared[i],
1630                                 dma_resv_held(bo->base.resv));
1631                         if (amdkfd_fence_check_mm(f, current->mm))
1632                                 return false;
1633                 }
1634         }
1635
1636         switch (bo->mem.mem_type) {
1637         case TTM_PL_TT:
1638                 if (amdgpu_bo_is_amdgpu_bo(bo) &&
1639                     amdgpu_bo_encrypted(ttm_to_amdgpu_bo(bo)))
1640                         return false;
1641                 return true;
1642
1643         case TTM_PL_VRAM:
1644                 /* Check each drm MM node individually */
1645                 while (num_pages) {
1646                         if (place->fpfn < (node->start + node->size) &&
1647                             !(place->lpfn && place->lpfn <= node->start))
1648                                 return true;
1649
1650                         num_pages -= node->size;
1651                         ++node;
1652                 }
1653                 return false;
1654
1655         default:
1656                 break;
1657         }
1658
1659         return ttm_bo_eviction_valuable(bo, place);
1660 }
1661
1662 /**
1663  * amdgpu_ttm_access_memory - Read or Write memory that backs a buffer object.
1664  *
1665  * @bo:  The buffer object to read/write
1666  * @offset:  Offset into buffer object
1667  * @buf:  Secondary buffer to write/read from
1668  * @len: Length in bytes of access
1669  * @write:  true if writing
1670  *
1671  * This is used to access VRAM that backs a buffer object via MMIO
1672  * access for debugging purposes.
1673  */
1674 static int amdgpu_ttm_access_memory(struct ttm_buffer_object *bo,
1675                                     unsigned long offset,
1676                                     void *buf, int len, int write)
1677 {
1678         struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
1679         struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev);
1680         struct drm_mm_node *nodes;
1681         uint32_t value = 0;
1682         int ret = 0;
1683         uint64_t pos;
1684         unsigned long flags;
1685
1686         if (bo->mem.mem_type != TTM_PL_VRAM)
1687                 return -EIO;
1688
1689         pos = offset;
1690         nodes = amdgpu_find_mm_node(&abo->tbo.mem, &pos);
1691         pos += (nodes->start << PAGE_SHIFT);
1692
1693         while (len && pos < adev->gmc.mc_vram_size) {
1694                 uint64_t aligned_pos = pos & ~(uint64_t)3;
1695                 uint64_t bytes = 4 - (pos & 3);
1696                 uint32_t shift = (pos & 3) * 8;
1697                 uint32_t mask = 0xffffffff << shift;
1698
1699                 if (len < bytes) {
1700                         mask &= 0xffffffff >> (bytes - len) * 8;
1701                         bytes = len;
1702                 }
1703
1704                 if (mask != 0xffffffff) {
1705                         spin_lock_irqsave(&adev->mmio_idx_lock, flags);
1706                         WREG32_NO_KIQ(mmMM_INDEX, ((uint32_t)aligned_pos) | 0x80000000);
1707                         WREG32_NO_KIQ(mmMM_INDEX_HI, aligned_pos >> 31);
1708                         if (!write || mask != 0xffffffff)
1709                                 value = RREG32_NO_KIQ(mmMM_DATA);
1710                         if (write) {
1711                                 value &= ~mask;
1712                                 value |= (*(uint32_t *)buf << shift) & mask;
1713                                 WREG32_NO_KIQ(mmMM_DATA, value);
1714                         }
1715                         spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
1716                         if (!write) {
1717                                 value = (value & mask) >> shift;
1718                                 memcpy(buf, &value, bytes);
1719                         }
1720                 } else {
1721                         bytes = (nodes->start + nodes->size) << PAGE_SHIFT;
1722                         bytes = min(bytes - pos, (uint64_t)len & ~0x3ull);
1723
1724                         amdgpu_device_vram_access(adev, pos, (uint32_t *)buf,
1725                                                   bytes, write);
1726                 }
1727
1728                 ret += bytes;
1729                 buf = (uint8_t *)buf + bytes;
1730                 pos += bytes;
1731                 len -= bytes;
1732                 if (pos >= (nodes->start + nodes->size) << PAGE_SHIFT) {
1733                         ++nodes;
1734                         pos = (nodes->start << PAGE_SHIFT);
1735                 }
1736         }
1737
1738         return ret;
1739 }
1740
1741 static struct ttm_bo_driver amdgpu_bo_driver = {
1742         .ttm_tt_create = &amdgpu_ttm_tt_create,
1743         .ttm_tt_populate = &amdgpu_ttm_tt_populate,
1744         .ttm_tt_unpopulate = &amdgpu_ttm_tt_unpopulate,
1745         .init_mem_type = &amdgpu_init_mem_type,
1746         .eviction_valuable = amdgpu_ttm_bo_eviction_valuable,
1747         .evict_flags = &amdgpu_evict_flags,
1748         .move = &amdgpu_bo_move,
1749         .verify_access = &amdgpu_verify_access,
1750         .move_notify = &amdgpu_bo_move_notify,
1751         .release_notify = &amdgpu_bo_release_notify,
1752         .fault_reserve_notify = &amdgpu_bo_fault_reserve_notify,
1753         .io_mem_reserve = &amdgpu_ttm_io_mem_reserve,
1754         .io_mem_pfn = amdgpu_ttm_io_mem_pfn,
1755         .access_memory = &amdgpu_ttm_access_memory,
1756         .del_from_lru_notify = &amdgpu_vm_del_from_lru_notify
1757 };
1758
1759 /*
1760  * Firmware Reservation functions
1761  */
1762 /**
1763  * amdgpu_ttm_fw_reserve_vram_fini - free fw reserved vram
1764  *
1765  * @adev: amdgpu_device pointer
1766  *
1767  * free fw reserved vram if it has been reserved.
1768  */
1769 static void amdgpu_ttm_fw_reserve_vram_fini(struct amdgpu_device *adev)
1770 {
1771         amdgpu_bo_free_kernel(&adev->fw_vram_usage.reserved_bo,
1772                 NULL, &adev->fw_vram_usage.va);
1773 }
1774
1775 /**
1776  * amdgpu_ttm_fw_reserve_vram_init - create bo vram reservation from fw
1777  *
1778  * @adev: amdgpu_device pointer
1779  *
1780  * create bo vram reservation from fw.
1781  */
1782 static int amdgpu_ttm_fw_reserve_vram_init(struct amdgpu_device *adev)
1783 {
1784         uint64_t vram_size = adev->gmc.visible_vram_size;
1785
1786         adev->fw_vram_usage.va = NULL;
1787         adev->fw_vram_usage.reserved_bo = NULL;
1788
1789         if (adev->fw_vram_usage.size == 0 ||
1790             adev->fw_vram_usage.size > vram_size)
1791                 return 0;
1792
1793         return amdgpu_bo_create_kernel_at(adev,
1794                                           adev->fw_vram_usage.start_offset,
1795                                           adev->fw_vram_usage.size,
1796                                           AMDGPU_GEM_DOMAIN_VRAM,
1797                                           &adev->fw_vram_usage.reserved_bo,
1798                                           &adev->fw_vram_usage.va);
1799 }
1800
1801 /*
1802  * Memoy training reservation functions
1803  */
1804
1805 /**
1806  * amdgpu_ttm_training_reserve_vram_fini - free memory training reserved vram
1807  *
1808  * @adev: amdgpu_device pointer
1809  *
1810  * free memory training reserved vram if it has been reserved.
1811  */
1812 static int amdgpu_ttm_training_reserve_vram_fini(struct amdgpu_device *adev)
1813 {
1814         struct psp_memory_training_context *ctx = &adev->psp.mem_train_ctx;
1815
1816         ctx->init = PSP_MEM_TRAIN_NOT_SUPPORT;
1817         amdgpu_bo_free_kernel(&ctx->c2p_bo, NULL, NULL);
1818         ctx->c2p_bo = NULL;
1819
1820         return 0;
1821 }
1822
1823 static void amdgpu_ttm_training_data_block_init(struct amdgpu_device *adev)
1824 {
1825         struct psp_memory_training_context *ctx = &adev->psp.mem_train_ctx;
1826
1827         memset(ctx, 0, sizeof(*ctx));
1828
1829         ctx->c2p_train_data_offset =
1830                 ALIGN((adev->gmc.mc_vram_size - adev->discovery_tmr_size - SZ_1M), SZ_1M);
1831         ctx->p2c_train_data_offset =
1832                 (adev->gmc.mc_vram_size - GDDR6_MEM_TRAINING_OFFSET);
1833         ctx->train_data_size =
1834                 GDDR6_MEM_TRAINING_DATA_SIZE_IN_BYTES;
1835
1836         DRM_DEBUG("train_data_size:%llx,p2c_train_data_offset:%llx,c2p_train_data_offset:%llx.\n",
1837                         ctx->train_data_size,
1838                         ctx->p2c_train_data_offset,
1839                         ctx->c2p_train_data_offset);
1840 }
1841
1842 /*
1843  * reserve TMR memory at the top of VRAM which holds
1844  * IP Discovery data and is protected by PSP.
1845  */
1846 static int amdgpu_ttm_reserve_tmr(struct amdgpu_device *adev)
1847 {
1848         int ret;
1849         struct psp_memory_training_context *ctx = &adev->psp.mem_train_ctx;
1850         bool mem_train_support = false;
1851
1852         if (!amdgpu_sriov_vf(adev)) {
1853                 ret = amdgpu_mem_train_support(adev);
1854                 if (ret == 1)
1855                         mem_train_support = true;
1856                 else if (ret == -1)
1857                         return -EINVAL;
1858                 else
1859                         DRM_DEBUG("memory training does not support!\n");
1860         }
1861
1862         /*
1863          * Query reserved tmr size through atom firmwareinfo for Sienna_Cichlid and onwards for all
1864          * the use cases (IP discovery/G6 memory training/profiling/diagnostic data.etc)
1865          *
1866          * Otherwise, fallback to legacy approach to check and reserve tmr block for ip
1867          * discovery data and G6 memory training data respectively
1868          */
1869         adev->discovery_tmr_size =
1870                 amdgpu_atomfirmware_get_fw_reserved_fb_size(adev);
1871         if (!adev->discovery_tmr_size)
1872                 adev->discovery_tmr_size = DISCOVERY_TMR_OFFSET;
1873
1874         if (mem_train_support) {
1875                 /* reserve vram for mem train according to TMR location */
1876                 amdgpu_ttm_training_data_block_init(adev);
1877                 ret = amdgpu_bo_create_kernel_at(adev,
1878                                          ctx->c2p_train_data_offset,
1879                                          ctx->train_data_size,
1880                                          AMDGPU_GEM_DOMAIN_VRAM,
1881                                          &ctx->c2p_bo,
1882                                          NULL);
1883                 if (ret) {
1884                         DRM_ERROR("alloc c2p_bo failed(%d)!\n", ret);
1885                         amdgpu_ttm_training_reserve_vram_fini(adev);
1886                         return ret;
1887                 }
1888                 ctx->init = PSP_MEM_TRAIN_RESERVE_SUCCESS;
1889         }
1890
1891         ret = amdgpu_bo_create_kernel_at(adev,
1892                                 adev->gmc.real_vram_size - adev->discovery_tmr_size,
1893                                 adev->discovery_tmr_size,
1894                                 AMDGPU_GEM_DOMAIN_VRAM,
1895                                 &adev->discovery_memory,
1896                                 NULL);
1897         if (ret) {
1898                 DRM_ERROR("alloc tmr failed(%d)!\n", ret);
1899                 amdgpu_bo_free_kernel(&adev->discovery_memory, NULL, NULL);
1900                 return ret;
1901         }
1902
1903         return 0;
1904 }
1905
1906 /**
1907  * amdgpu_ttm_init - Init the memory management (ttm) as well as various
1908  * gtt/vram related fields.
1909  *
1910  * This initializes all of the memory space pools that the TTM layer
1911  * will need such as the GTT space (system memory mapped to the device),
1912  * VRAM (on-board memory), and on-chip memories (GDS, GWS, OA) which
1913  * can be mapped per VMID.
1914  */
1915 int amdgpu_ttm_init(struct amdgpu_device *adev)
1916 {
1917         uint64_t gtt_size;
1918         int r;
1919         u64 vis_vram_limit;
1920         void *stolen_vga_buf;
1921
1922         mutex_init(&adev->mman.gtt_window_lock);
1923
1924         /* No others user of address space so set it to 0 */
1925         r = ttm_bo_device_init(&adev->mman.bdev,
1926                                &amdgpu_bo_driver,
1927                                adev->ddev->anon_inode->i_mapping,
1928                                adev->ddev->vma_offset_manager,
1929                                dma_addressing_limited(adev->dev));
1930         if (r) {
1931                 DRM_ERROR("failed initializing buffer object driver(%d).\n", r);
1932                 return r;
1933         }
1934         adev->mman.initialized = true;
1935
1936         /* We opt to avoid OOM on system pages allocations */
1937         adev->mman.bdev.no_retry = true;
1938
1939         /* Initialize VRAM pool with all of VRAM divided into pages */
1940         r = ttm_bo_init_mm(&adev->mman.bdev, TTM_PL_VRAM,
1941                                 adev->gmc.real_vram_size >> PAGE_SHIFT);
1942         if (r) {
1943                 DRM_ERROR("Failed initializing VRAM heap.\n");
1944                 return r;
1945         }
1946
1947         /* Reduce size of CPU-visible VRAM if requested */
1948         vis_vram_limit = (u64)amdgpu_vis_vram_limit * 1024 * 1024;
1949         if (amdgpu_vis_vram_limit > 0 &&
1950             vis_vram_limit <= adev->gmc.visible_vram_size)
1951                 adev->gmc.visible_vram_size = vis_vram_limit;
1952
1953         /* Change the size here instead of the init above so only lpfn is affected */
1954         amdgpu_ttm_set_buffer_funcs_status(adev, false);
1955 #ifdef CONFIG_64BIT
1956         adev->mman.aper_base_kaddr = ioremap_wc(adev->gmc.aper_base,
1957                                                 adev->gmc.visible_vram_size);
1958 #endif
1959
1960         /*
1961          *The reserved vram for firmware must be pinned to the specified
1962          *place on the VRAM, so reserve it early.
1963          */
1964         r = amdgpu_ttm_fw_reserve_vram_init(adev);
1965         if (r) {
1966                 return r;
1967         }
1968
1969         /*
1970          * only NAVI10 and onwards ASIC support for IP discovery.
1971          * If IP discovery enabled, a block of memory should be
1972          * reserved for IP discovey.
1973          */
1974         if (adev->discovery_bin) {
1975                 r = amdgpu_ttm_reserve_tmr(adev);
1976                 if (r)
1977                         return r;
1978         }
1979
1980         /* allocate memory as required for VGA
1981          * This is used for VGA emulation and pre-OS scanout buffers to
1982          * avoid display artifacts while transitioning between pre-OS
1983          * and driver.  */
1984         r = amdgpu_bo_create_kernel(adev, adev->gmc.stolen_size, PAGE_SIZE,
1985                                     AMDGPU_GEM_DOMAIN_VRAM,
1986                                     &adev->stolen_vga_memory,
1987                                     NULL, &stolen_vga_buf);
1988         if (r)
1989                 return r;
1990
1991         DRM_INFO("amdgpu: %uM of VRAM memory ready\n",
1992                  (unsigned) (adev->gmc.real_vram_size / (1024 * 1024)));
1993
1994         /* Compute GTT size, either bsaed on 3/4th the size of RAM size
1995          * or whatever the user passed on module init */
1996         if (amdgpu_gtt_size == -1) {
1997                 struct sysinfo si;
1998
1999                 si_meminfo(&si);
2000                 gtt_size = min(max((AMDGPU_DEFAULT_GTT_SIZE_MB << 20),
2001                                adev->gmc.mc_vram_size),
2002                                ((uint64_t)si.totalram * si.mem_unit * 3/4));
2003         }
2004         else
2005                 gtt_size = (uint64_t)amdgpu_gtt_size << 20;
2006
2007         /* Initialize GTT memory pool */
2008         r = ttm_bo_init_mm(&adev->mman.bdev, TTM_PL_TT, gtt_size >> PAGE_SHIFT);
2009         if (r) {
2010                 DRM_ERROR("Failed initializing GTT heap.\n");
2011                 return r;
2012         }
2013         DRM_INFO("amdgpu: %uM of GTT memory ready.\n",
2014                  (unsigned)(gtt_size / (1024 * 1024)));
2015
2016         /* Initialize various on-chip memory pools */
2017         r = ttm_bo_init_mm(&adev->mman.bdev, AMDGPU_PL_GDS,
2018                            adev->gds.gds_size);
2019         if (r) {
2020                 DRM_ERROR("Failed initializing GDS heap.\n");
2021                 return r;
2022         }
2023
2024         r = ttm_bo_init_mm(&adev->mman.bdev, AMDGPU_PL_GWS,
2025                            adev->gds.gws_size);
2026         if (r) {
2027                 DRM_ERROR("Failed initializing gws heap.\n");
2028                 return r;
2029         }
2030
2031         r = ttm_bo_init_mm(&adev->mman.bdev, AMDGPU_PL_OA,
2032                            adev->gds.oa_size);
2033         if (r) {
2034                 DRM_ERROR("Failed initializing oa heap.\n");
2035                 return r;
2036         }
2037
2038         return 0;
2039 }
2040
2041 /**
2042  * amdgpu_ttm_late_init - Handle any late initialization for amdgpu_ttm
2043  */
2044 void amdgpu_ttm_late_init(struct amdgpu_device *adev)
2045 {
2046         void *stolen_vga_buf;
2047         /* return the VGA stolen memory (if any) back to VRAM */
2048         amdgpu_bo_free_kernel(&adev->stolen_vga_memory, NULL, &stolen_vga_buf);
2049 }
2050
2051 /**
2052  * amdgpu_ttm_fini - De-initialize the TTM memory pools
2053  */
2054 void amdgpu_ttm_fini(struct amdgpu_device *adev)
2055 {
2056         if (!adev->mman.initialized)
2057                 return;
2058
2059         amdgpu_ttm_training_reserve_vram_fini(adev);
2060         /* return the IP Discovery TMR memory back to VRAM */
2061         amdgpu_bo_free_kernel(&adev->discovery_memory, NULL, NULL);
2062         amdgpu_ttm_fw_reserve_vram_fini(adev);
2063
2064         if (adev->mman.aper_base_kaddr)
2065                 iounmap(adev->mman.aper_base_kaddr);
2066         adev->mman.aper_base_kaddr = NULL;
2067
2068         ttm_bo_clean_mm(&adev->mman.bdev, TTM_PL_VRAM);
2069         ttm_bo_clean_mm(&adev->mman.bdev, TTM_PL_TT);
2070         ttm_bo_clean_mm(&adev->mman.bdev, AMDGPU_PL_GDS);
2071         ttm_bo_clean_mm(&adev->mman.bdev, AMDGPU_PL_GWS);
2072         ttm_bo_clean_mm(&adev->mman.bdev, AMDGPU_PL_OA);
2073         ttm_bo_device_release(&adev->mman.bdev);
2074         adev->mman.initialized = false;
2075         DRM_INFO("amdgpu: ttm finalized\n");
2076 }
2077
2078 /**
2079  * amdgpu_ttm_set_buffer_funcs_status - enable/disable use of buffer functions
2080  *
2081  * @adev: amdgpu_device pointer
2082  * @enable: true when we can use buffer functions.
2083  *
2084  * Enable/disable use of buffer functions during suspend/resume. This should
2085  * only be called at bootup or when userspace isn't running.
2086  */
2087 void amdgpu_ttm_set_buffer_funcs_status(struct amdgpu_device *adev, bool enable)
2088 {
2089         struct ttm_mem_type_manager *man = &adev->mman.bdev.man[TTM_PL_VRAM];
2090         uint64_t size;
2091         int r;
2092
2093         if (!adev->mman.initialized || adev->in_gpu_reset ||
2094             adev->mman.buffer_funcs_enabled == enable)
2095                 return;
2096
2097         if (enable) {
2098                 struct amdgpu_ring *ring;
2099                 struct drm_gpu_scheduler *sched;
2100
2101                 ring = adev->mman.buffer_funcs_ring;
2102                 sched = &ring->sched;
2103                 r = drm_sched_entity_init(&adev->mman.entity,
2104                                           DRM_SCHED_PRIORITY_KERNEL, &sched,
2105                                           1, NULL);
2106                 if (r) {
2107                         DRM_ERROR("Failed setting up TTM BO move entity (%d)\n",
2108                                   r);
2109                         return;
2110                 }
2111         } else {
2112                 drm_sched_entity_destroy(&adev->mman.entity);
2113                 dma_fence_put(man->move);
2114                 man->move = NULL;
2115         }
2116
2117         /* this just adjusts TTM size idea, which sets lpfn to the correct value */
2118         if (enable)
2119                 size = adev->gmc.real_vram_size;
2120         else
2121                 size = adev->gmc.visible_vram_size;
2122         man->size = size >> PAGE_SHIFT;
2123         adev->mman.buffer_funcs_enabled = enable;
2124 }
2125
2126 int amdgpu_mmap(struct file *filp, struct vm_area_struct *vma)
2127 {
2128         struct drm_file *file_priv = filp->private_data;
2129         struct amdgpu_device *adev = file_priv->minor->dev->dev_private;
2130
2131         if (adev == NULL)
2132                 return -EINVAL;
2133
2134         return ttm_bo_mmap(filp, vma, &adev->mman.bdev);
2135 }
2136
2137 int amdgpu_copy_buffer(struct amdgpu_ring *ring, uint64_t src_offset,
2138                        uint64_t dst_offset, uint32_t byte_count,
2139                        struct dma_resv *resv,
2140                        struct dma_fence **fence, bool direct_submit,
2141                        bool vm_needs_flush, bool tmz)
2142 {
2143         enum amdgpu_ib_pool_type pool = direct_submit ? AMDGPU_IB_POOL_DIRECT :
2144                 AMDGPU_IB_POOL_DELAYED;
2145         struct amdgpu_device *adev = ring->adev;
2146         struct amdgpu_job *job;
2147
2148         uint32_t max_bytes;
2149         unsigned num_loops, num_dw;
2150         unsigned i;
2151         int r;
2152
2153         if (direct_submit && !ring->sched.ready) {
2154                 DRM_ERROR("Trying to move memory with ring turned off.\n");
2155                 return -EINVAL;
2156         }
2157
2158         max_bytes = adev->mman.buffer_funcs->copy_max_bytes;
2159         num_loops = DIV_ROUND_UP(byte_count, max_bytes);
2160         num_dw = ALIGN(num_loops * adev->mman.buffer_funcs->copy_num_dw, 8);
2161
2162         r = amdgpu_job_alloc_with_ib(adev, num_dw * 4, pool, &job);
2163         if (r)
2164                 return r;
2165
2166         if (vm_needs_flush) {
2167                 job->vm_pd_addr = amdgpu_gmc_pd_addr(adev->gart.bo);
2168                 job->vm_needs_flush = true;
2169         }
2170         if (resv) {
2171                 r = amdgpu_sync_resv(adev, &job->sync, resv,
2172                                      AMDGPU_SYNC_ALWAYS,
2173                                      AMDGPU_FENCE_OWNER_UNDEFINED);
2174                 if (r) {
2175                         DRM_ERROR("sync failed (%d).\n", r);
2176                         goto error_free;
2177                 }
2178         }
2179
2180         for (i = 0; i < num_loops; i++) {
2181                 uint32_t cur_size_in_bytes = min(byte_count, max_bytes);
2182
2183                 amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_offset,
2184                                         dst_offset, cur_size_in_bytes, tmz);
2185
2186                 src_offset += cur_size_in_bytes;
2187                 dst_offset += cur_size_in_bytes;
2188                 byte_count -= cur_size_in_bytes;
2189         }
2190
2191         amdgpu_ring_pad_ib(ring, &job->ibs[0]);
2192         WARN_ON(job->ibs[0].length_dw > num_dw);
2193         if (direct_submit)
2194                 r = amdgpu_job_submit_direct(job, ring, fence);
2195         else
2196                 r = amdgpu_job_submit(job, &adev->mman.entity,
2197                                       AMDGPU_FENCE_OWNER_UNDEFINED, fence);
2198         if (r)
2199                 goto error_free;
2200
2201         return r;
2202
2203 error_free:
2204         amdgpu_job_free(job);
2205         DRM_ERROR("Error scheduling IBs (%d)\n", r);
2206         return r;
2207 }
2208
2209 int amdgpu_fill_buffer(struct amdgpu_bo *bo,
2210                        uint32_t src_data,
2211                        struct dma_resv *resv,
2212                        struct dma_fence **fence)
2213 {
2214         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
2215         uint32_t max_bytes = adev->mman.buffer_funcs->fill_max_bytes;
2216         struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
2217
2218         struct drm_mm_node *mm_node;
2219         unsigned long num_pages;
2220         unsigned int num_loops, num_dw;
2221
2222         struct amdgpu_job *job;
2223         int r;
2224
2225         if (!adev->mman.buffer_funcs_enabled) {
2226                 DRM_ERROR("Trying to clear memory with ring turned off.\n");
2227                 return -EINVAL;
2228         }
2229
2230         if (bo->tbo.mem.mem_type == TTM_PL_TT) {
2231                 r = amdgpu_ttm_alloc_gart(&bo->tbo);
2232                 if (r)
2233                         return r;
2234         }
2235
2236         num_pages = bo->tbo.num_pages;
2237         mm_node = bo->tbo.mem.mm_node;
2238         num_loops = 0;
2239         while (num_pages) {
2240                 uint64_t byte_count = mm_node->size << PAGE_SHIFT;
2241
2242                 num_loops += DIV_ROUND_UP_ULL(byte_count, max_bytes);
2243                 num_pages -= mm_node->size;
2244                 ++mm_node;
2245         }
2246         num_dw = num_loops * adev->mman.buffer_funcs->fill_num_dw;
2247
2248         /* for IB padding */
2249         num_dw += 64;
2250
2251         r = amdgpu_job_alloc_with_ib(adev, num_dw * 4, AMDGPU_IB_POOL_DELAYED,
2252                                      &job);
2253         if (r)
2254                 return r;
2255
2256         if (resv) {
2257                 r = amdgpu_sync_resv(adev, &job->sync, resv,
2258                                      AMDGPU_SYNC_ALWAYS,
2259                                      AMDGPU_FENCE_OWNER_UNDEFINED);
2260                 if (r) {
2261                         DRM_ERROR("sync failed (%d).\n", r);
2262                         goto error_free;
2263                 }
2264         }
2265
2266         num_pages = bo->tbo.num_pages;
2267         mm_node = bo->tbo.mem.mm_node;
2268
2269         while (num_pages) {
2270                 uint64_t byte_count = mm_node->size << PAGE_SHIFT;
2271                 uint64_t dst_addr;
2272
2273                 dst_addr = amdgpu_mm_node_addr(&bo->tbo, mm_node, &bo->tbo.mem);
2274                 while (byte_count) {
2275                         uint32_t cur_size_in_bytes = min_t(uint64_t, byte_count,
2276                                                            max_bytes);
2277
2278                         amdgpu_emit_fill_buffer(adev, &job->ibs[0], src_data,
2279                                                 dst_addr, cur_size_in_bytes);
2280
2281                         dst_addr += cur_size_in_bytes;
2282                         byte_count -= cur_size_in_bytes;
2283                 }
2284
2285                 num_pages -= mm_node->size;
2286                 ++mm_node;
2287         }
2288
2289         amdgpu_ring_pad_ib(ring, &job->ibs[0]);
2290         WARN_ON(job->ibs[0].length_dw > num_dw);
2291         r = amdgpu_job_submit(job, &adev->mman.entity,
2292                               AMDGPU_FENCE_OWNER_UNDEFINED, fence);
2293         if (r)
2294                 goto error_free;
2295
2296         return 0;
2297
2298 error_free:
2299         amdgpu_job_free(job);
2300         return r;
2301 }
2302
2303 #if defined(CONFIG_DEBUG_FS)
2304
2305 static int amdgpu_mm_dump_table(struct seq_file *m, void *data)
2306 {
2307         struct drm_info_node *node = (struct drm_info_node *)m->private;
2308         unsigned ttm_pl = (uintptr_t)node->info_ent->data;
2309         struct drm_device *dev = node->minor->dev;
2310         struct amdgpu_device *adev = dev->dev_private;
2311         struct ttm_mem_type_manager *man = &adev->mman.bdev.man[ttm_pl];
2312         struct drm_printer p = drm_seq_file_printer(m);
2313
2314         man->func->debug(man, &p);
2315         return 0;
2316 }
2317
2318 static const struct drm_info_list amdgpu_ttm_debugfs_list[] = {
2319         {"amdgpu_vram_mm", amdgpu_mm_dump_table, 0, (void *)TTM_PL_VRAM},
2320         {"amdgpu_gtt_mm", amdgpu_mm_dump_table, 0, (void *)TTM_PL_TT},
2321         {"amdgpu_gds_mm", amdgpu_mm_dump_table, 0, (void *)AMDGPU_PL_GDS},
2322         {"amdgpu_gws_mm", amdgpu_mm_dump_table, 0, (void *)AMDGPU_PL_GWS},
2323         {"amdgpu_oa_mm", amdgpu_mm_dump_table, 0, (void *)AMDGPU_PL_OA},
2324         {"ttm_page_pool", ttm_page_alloc_debugfs, 0, NULL},
2325 #ifdef CONFIG_SWIOTLB
2326         {"ttm_dma_page_pool", ttm_dma_page_alloc_debugfs, 0, NULL}
2327 #endif
2328 };
2329
2330 /**
2331  * amdgpu_ttm_vram_read - Linear read access to VRAM
2332  *
2333  * Accesses VRAM via MMIO for debugging purposes.
2334  */
2335 static ssize_t amdgpu_ttm_vram_read(struct file *f, char __user *buf,
2336                                     size_t size, loff_t *pos)
2337 {
2338         struct amdgpu_device *adev = file_inode(f)->i_private;
2339         ssize_t result = 0;
2340
2341         if (size & 0x3 || *pos & 0x3)
2342                 return -EINVAL;
2343
2344         if (*pos >= adev->gmc.mc_vram_size)
2345                 return -ENXIO;
2346
2347         size = min(size, (size_t)(adev->gmc.mc_vram_size - *pos));
2348         while (size) {
2349                 size_t bytes = min(size, AMDGPU_TTM_VRAM_MAX_DW_READ * 4);
2350                 uint32_t value[AMDGPU_TTM_VRAM_MAX_DW_READ];
2351
2352                 amdgpu_device_vram_access(adev, *pos, value, bytes, false);
2353                 if (copy_to_user(buf, value, bytes))
2354                         return -EFAULT;
2355
2356                 result += bytes;
2357                 buf += bytes;
2358                 *pos += bytes;
2359                 size -= bytes;
2360         }
2361
2362         return result;
2363 }
2364
2365 /**
2366  * amdgpu_ttm_vram_write - Linear write access to VRAM
2367  *
2368  * Accesses VRAM via MMIO for debugging purposes.
2369  */
2370 static ssize_t amdgpu_ttm_vram_write(struct file *f, const char __user *buf,
2371                                     size_t size, loff_t *pos)
2372 {
2373         struct amdgpu_device *adev = file_inode(f)->i_private;
2374         ssize_t result = 0;
2375         int r;
2376
2377         if (size & 0x3 || *pos & 0x3)
2378                 return -EINVAL;
2379
2380         if (*pos >= adev->gmc.mc_vram_size)
2381                 return -ENXIO;
2382
2383         while (size) {
2384                 unsigned long flags;
2385                 uint32_t value;
2386
2387                 if (*pos >= adev->gmc.mc_vram_size)
2388                         return result;
2389
2390                 r = get_user(value, (uint32_t *)buf);
2391                 if (r)
2392                         return r;
2393
2394                 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
2395                 WREG32_NO_KIQ(mmMM_INDEX, ((uint32_t)*pos) | 0x80000000);
2396                 WREG32_NO_KIQ(mmMM_INDEX_HI, *pos >> 31);
2397                 WREG32_NO_KIQ(mmMM_DATA, value);
2398                 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
2399
2400                 result += 4;
2401                 buf += 4;
2402                 *pos += 4;
2403                 size -= 4;
2404         }
2405
2406         return result;
2407 }
2408
2409 static const struct file_operations amdgpu_ttm_vram_fops = {
2410         .owner = THIS_MODULE,
2411         .read = amdgpu_ttm_vram_read,
2412         .write = amdgpu_ttm_vram_write,
2413         .llseek = default_llseek,
2414 };
2415
2416 #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS
2417
2418 /**
2419  * amdgpu_ttm_gtt_read - Linear read access to GTT memory
2420  */
2421 static ssize_t amdgpu_ttm_gtt_read(struct file *f, char __user *buf,
2422                                    size_t size, loff_t *pos)
2423 {
2424         struct amdgpu_device *adev = file_inode(f)->i_private;
2425         ssize_t result = 0;
2426         int r;
2427
2428         while (size) {
2429                 loff_t p = *pos / PAGE_SIZE;
2430                 unsigned off = *pos & ~PAGE_MASK;
2431                 size_t cur_size = min_t(size_t, size, PAGE_SIZE - off);
2432                 struct page *page;
2433                 void *ptr;
2434
2435                 if (p >= adev->gart.num_cpu_pages)
2436                         return result;
2437
2438                 page = adev->gart.pages[p];
2439                 if (page) {
2440                         ptr = kmap(page);
2441                         ptr += off;
2442
2443                         r = copy_to_user(buf, ptr, cur_size);
2444                         kunmap(adev->gart.pages[p]);
2445                 } else
2446                         r = clear_user(buf, cur_size);
2447
2448                 if (r)
2449                         return -EFAULT;
2450
2451                 result += cur_size;
2452                 buf += cur_size;
2453                 *pos += cur_size;
2454                 size -= cur_size;
2455         }
2456
2457         return result;
2458 }
2459
2460 static const struct file_operations amdgpu_ttm_gtt_fops = {
2461         .owner = THIS_MODULE,
2462         .read = amdgpu_ttm_gtt_read,
2463         .llseek = default_llseek
2464 };
2465
2466 #endif
2467
2468 /**
2469  * amdgpu_iomem_read - Virtual read access to GPU mapped memory
2470  *
2471  * This function is used to read memory that has been mapped to the
2472  * GPU and the known addresses are not physical addresses but instead
2473  * bus addresses (e.g., what you'd put in an IB or ring buffer).
2474  */
2475 static ssize_t amdgpu_iomem_read(struct file *f, char __user *buf,
2476                                  size_t size, loff_t *pos)
2477 {
2478         struct amdgpu_device *adev = file_inode(f)->i_private;
2479         struct iommu_domain *dom;
2480         ssize_t result = 0;
2481         int r;
2482
2483         /* retrieve the IOMMU domain if any for this device */
2484         dom = iommu_get_domain_for_dev(adev->dev);
2485
2486         while (size) {
2487                 phys_addr_t addr = *pos & PAGE_MASK;
2488                 loff_t off = *pos & ~PAGE_MASK;
2489                 size_t bytes = PAGE_SIZE - off;
2490                 unsigned long pfn;
2491                 struct page *p;
2492                 void *ptr;
2493
2494                 bytes = bytes < size ? bytes : size;
2495
2496                 /* Translate the bus address to a physical address.  If
2497                  * the domain is NULL it means there is no IOMMU active
2498                  * and the address translation is the identity
2499                  */
2500                 addr = dom ? iommu_iova_to_phys(dom, addr) : addr;
2501
2502                 pfn = addr >> PAGE_SHIFT;
2503                 if (!pfn_valid(pfn))
2504                         return -EPERM;
2505
2506                 p = pfn_to_page(pfn);
2507                 if (p->mapping != adev->mman.bdev.dev_mapping)
2508                         return -EPERM;
2509
2510                 ptr = kmap(p);
2511                 r = copy_to_user(buf, ptr + off, bytes);
2512                 kunmap(p);
2513                 if (r)
2514                         return -EFAULT;
2515
2516                 size -= bytes;
2517                 *pos += bytes;
2518                 result += bytes;
2519         }
2520
2521         return result;
2522 }
2523
2524 /**
2525  * amdgpu_iomem_write - Virtual write access to GPU mapped memory
2526  *
2527  * This function is used to write memory that has been mapped to the
2528  * GPU and the known addresses are not physical addresses but instead
2529  * bus addresses (e.g., what you'd put in an IB or ring buffer).
2530  */
2531 static ssize_t amdgpu_iomem_write(struct file *f, const char __user *buf,
2532                                  size_t size, loff_t *pos)
2533 {
2534         struct amdgpu_device *adev = file_inode(f)->i_private;
2535         struct iommu_domain *dom;
2536         ssize_t result = 0;
2537         int r;
2538
2539         dom = iommu_get_domain_for_dev(adev->dev);
2540
2541         while (size) {
2542                 phys_addr_t addr = *pos & PAGE_MASK;
2543                 loff_t off = *pos & ~PAGE_MASK;
2544                 size_t bytes = PAGE_SIZE - off;
2545                 unsigned long pfn;
2546                 struct page *p;
2547                 void *ptr;
2548
2549                 bytes = bytes < size ? bytes : size;
2550
2551                 addr = dom ? iommu_iova_to_phys(dom, addr) : addr;
2552
2553                 pfn = addr >> PAGE_SHIFT;
2554                 if (!pfn_valid(pfn))
2555                         return -EPERM;
2556
2557                 p = pfn_to_page(pfn);
2558                 if (p->mapping != adev->mman.bdev.dev_mapping)
2559                         return -EPERM;
2560
2561                 ptr = kmap(p);
2562                 r = copy_from_user(ptr + off, buf, bytes);
2563                 kunmap(p);
2564                 if (r)
2565                         return -EFAULT;
2566
2567                 size -= bytes;
2568                 *pos += bytes;
2569                 result += bytes;
2570         }
2571
2572         return result;
2573 }
2574
2575 static const struct file_operations amdgpu_ttm_iomem_fops = {
2576         .owner = THIS_MODULE,
2577         .read = amdgpu_iomem_read,
2578         .write = amdgpu_iomem_write,
2579         .llseek = default_llseek
2580 };
2581
2582 static const struct {
2583         char *name;
2584         const struct file_operations *fops;
2585         int domain;
2586 } ttm_debugfs_entries[] = {
2587         { "amdgpu_vram", &amdgpu_ttm_vram_fops, TTM_PL_VRAM },
2588 #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS
2589         { "amdgpu_gtt", &amdgpu_ttm_gtt_fops, TTM_PL_TT },
2590 #endif
2591         { "amdgpu_iomem", &amdgpu_ttm_iomem_fops, TTM_PL_SYSTEM },
2592 };
2593
2594 #endif
2595
2596 int amdgpu_ttm_debugfs_init(struct amdgpu_device *adev)
2597 {
2598 #if defined(CONFIG_DEBUG_FS)
2599         unsigned count;
2600
2601         struct drm_minor *minor = adev->ddev->primary;
2602         struct dentry *ent, *root = minor->debugfs_root;
2603
2604         for (count = 0; count < ARRAY_SIZE(ttm_debugfs_entries); count++) {
2605                 ent = debugfs_create_file(
2606                                 ttm_debugfs_entries[count].name,
2607                                 S_IFREG | S_IRUGO, root,
2608                                 adev,
2609                                 ttm_debugfs_entries[count].fops);
2610                 if (IS_ERR(ent))
2611                         return PTR_ERR(ent);
2612                 if (ttm_debugfs_entries[count].domain == TTM_PL_VRAM)
2613                         i_size_write(ent->d_inode, adev->gmc.mc_vram_size);
2614                 else if (ttm_debugfs_entries[count].domain == TTM_PL_TT)
2615                         i_size_write(ent->d_inode, adev->gmc.gart_size);
2616                 adev->mman.debugfs_entries[count] = ent;
2617         }
2618
2619         count = ARRAY_SIZE(amdgpu_ttm_debugfs_list);
2620
2621 #ifdef CONFIG_SWIOTLB
2622         if (!(adev->need_swiotlb && swiotlb_nr_tbl()))
2623                 --count;
2624 #endif
2625
2626         return amdgpu_debugfs_add_files(adev, amdgpu_ttm_debugfs_list, count);
2627 #else
2628         return 0;
2629 #endif
2630 }