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/pagemap.h>
  36 #include <linux/sched/task.h>
  37 #include <linux/sched/mm.h>
  38 #include <linux/seq_file.h>
  39 #include <linux/slab.h>
  40 #include <linux/swap.h>
  41 #include <linux/dma-buf.h>
  42 #include <linux/sizes.h>
  43 #include <linux/module.h>
  44
  45 #include <drm/drm_drv.h>
  46 #include <drm/ttm/ttm_bo.h>
  47 #include <drm/ttm/ttm_placement.h>
  48 #include <drm/ttm/ttm_range_manager.h>
  49 #include <drm/ttm/ttm_tt.h>
  50
  51 #include <drm/amdgpu_drm.h>
  52
  53 #include "amdgpu.h"
  54 #include "amdgpu_object.h"
  55 #include "amdgpu_trace.h"
  56 #include "amdgpu_amdkfd.h"
  57 #include "amdgpu_sdma.h"
  58 #include "amdgpu_ras.h"
  59 #include "amdgpu_hmm.h"
  60 #include "amdgpu_atomfirmware.h"
  61 #include "amdgpu_res_cursor.h"
  62 #include "bif/bif_4_1_d.h"
  63
  64 MODULE_IMPORT_NS("DMA_BUF");
  65
  66 #define AMDGPU_TTM_VRAM_MAX_DW_READ     ((size_t)128)
  67
  68 static int amdgpu_ttm_backend_bind(struct ttm_device *bdev,
  69                                    struct ttm_tt *ttm,
  70                                    struct ttm_resource *bo_mem);
  71 static void amdgpu_ttm_backend_unbind(struct ttm_device *bdev,
  72                                       struct ttm_tt *ttm);
  73
  74 static int amdgpu_ttm_init_on_chip(struct amdgpu_device *adev,
  75                                     unsigned int type,
  76                                     uint64_t size_in_page)
  77 {
  78         return ttm_range_man_init(&adev->mman.bdev, type,
  79                                   false, size_in_page);
  80 }
  81
  82 /**
  83  * amdgpu_evict_flags - Compute placement flags
  84  *
  85  * @bo: The buffer object to evict
  86  * @placement: Possible destination(s) for evicted BO
  87  *
  88  * Fill in placement data when ttm_bo_evict() is called
  89  */
  90 static void amdgpu_evict_flags(struct ttm_buffer_object *bo,
  91                                 struct ttm_placement *placement)
  92 {
  93         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
  94         struct amdgpu_bo *abo;
  95         static const struct ttm_place placements = {
  96                 .fpfn = 0,
  97                 .lpfn = 0,
  98                 .mem_type = TTM_PL_SYSTEM,
  99                 .flags = 0
 100         };
 101
 102         /* Don't handle scatter gather BOs */
 103         if (bo->type == ttm_bo_type_sg) {
 104                 placement->num_placement = 0;
 105                 return;
 106         }
 107
 108         /* Object isn't an AMDGPU object so ignore */
 109         if (!amdgpu_bo_is_amdgpu_bo(bo)) {
 110                 placement->placement = &placements;
 111                 placement->num_placement = 1;
 112                 return;
 113         }
 114
 115         abo = ttm_to_amdgpu_bo(bo);
 116         if (abo->flags & AMDGPU_GEM_CREATE_DISCARDABLE) {
 117                 placement->num_placement = 0;
 118                 return;
 119         }
 120
 121         switch (bo->resource->mem_type) {
 122         case AMDGPU_PL_GDS:
 123         case AMDGPU_PL_GWS:
 124         case AMDGPU_PL_OA:
 125         case AMDGPU_PL_DOORBELL:
 126                 placement->num_placement = 0;
 127                 return;
 128
 129         case TTM_PL_VRAM:
 130                 if (!adev->mman.buffer_funcs_enabled) {
 131                         /* Move to system memory */
 132                         amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_CPU);
 133
 134                 } else if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
 135                            !(abo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED) &&
 136                            amdgpu_res_cpu_visible(adev, bo->resource)) {
 137
 138                         /* Try evicting to the CPU inaccessible part of VRAM
 139                          * first, but only set GTT as busy placement, so this
 140                          * BO will be evicted to GTT rather than causing other
 141                          * BOs to be evicted from VRAM
 142                          */
 143                         amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM |
 144                                                         AMDGPU_GEM_DOMAIN_GTT |
 145                                                         AMDGPU_GEM_DOMAIN_CPU);
 146                         abo->placements[0].fpfn = adev->gmc.visible_vram_size >> PAGE_SHIFT;
 147                         abo->placements[0].lpfn = 0;
 148                         abo->placements[0].flags |= TTM_PL_FLAG_DESIRED;
 149                 } else {
 150                         /* Move to GTT memory */
 151                         amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_GTT |
 152                                                         AMDGPU_GEM_DOMAIN_CPU);
 153                 }
 154                 break;
 155         case TTM_PL_TT:
 156         case AMDGPU_PL_PREEMPT:
 157         default:
 158                 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_CPU);
 159                 break;
 160         }
 161         *placement = abo->placement;
 162 }
 163
 164 /**
 165  * amdgpu_ttm_map_buffer - Map memory into the GART windows
 166  * @bo: buffer object to map
 167  * @mem: memory object to map
 168  * @mm_cur: range to map
 169  * @window: which GART window to use
 170  * @ring: DMA ring to use for the copy
 171  * @tmz: if we should setup a TMZ enabled mapping
 172  * @size: in number of bytes to map, out number of bytes mapped
 173  * @addr: resulting address inside the MC address space
 174  *
 175  * Setup one of the GART windows to access a specific piece of memory or return
 176  * the physical address for local memory.
 177  */
 178 static int amdgpu_ttm_map_buffer(struct ttm_buffer_object *bo,
 179                                  struct ttm_resource *mem,
 180                                  struct amdgpu_res_cursor *mm_cur,
 181                                  unsigned int window, struct amdgpu_ring *ring,
 182                                  bool tmz, uint64_t *size, uint64_t *addr)
 183 {
 184         struct amdgpu_device *adev = ring->adev;
 185         unsigned int offset, num_pages, num_dw, num_bytes;
 186         uint64_t src_addr, dst_addr;
 187         struct amdgpu_job *job;
 188         void *cpu_addr;
 189         uint64_t flags;
 190         unsigned int i;
 191         int r;
 192
 193         BUG_ON(adev->mman.buffer_funcs->copy_max_bytes <
 194                AMDGPU_GTT_MAX_TRANSFER_SIZE * 8);
 195
 196         if (WARN_ON(mem->mem_type == AMDGPU_PL_PREEMPT))
 197                 return -EINVAL;
 198
 199         /* Map only what can't be accessed directly */
 200         if (!tmz && mem->start != AMDGPU_BO_INVALID_OFFSET) {
 201                 *addr = amdgpu_ttm_domain_start(adev, mem->mem_type) +
 202                         mm_cur->start;
 203                 return 0;
 204         }
 205
 206
 207         /*
 208          * If start begins at an offset inside the page, then adjust the size
 209          * and addr accordingly
 210          */
 211         offset = mm_cur->start & ~PAGE_MASK;
 212
 213         num_pages = PFN_UP(*size + offset);
 214         num_pages = min_t(uint32_t, num_pages, AMDGPU_GTT_MAX_TRANSFER_SIZE);
 215
 216         *size = min(*size, (uint64_t)num_pages * PAGE_SIZE - offset);
 217
 218         *addr = adev->gmc.gart_start;
 219         *addr += (u64)window * AMDGPU_GTT_MAX_TRANSFER_SIZE *
 220                 AMDGPU_GPU_PAGE_SIZE;
 221         *addr += offset;
 222
 223         num_dw = ALIGN(adev->mman.buffer_funcs->copy_num_dw, 8);
 224         num_bytes = num_pages * 8 * AMDGPU_GPU_PAGES_IN_CPU_PAGE;
 225
 226         r = amdgpu_job_alloc_with_ib(adev, &adev->mman.high_pr,
 227                                      AMDGPU_FENCE_OWNER_UNDEFINED,
 228                                      num_dw * 4 + num_bytes,
 229                                      AMDGPU_IB_POOL_DELAYED, &job);
 230         if (r)
 231                 return r;
 232
 233         src_addr = num_dw * 4;
 234         src_addr += job->ibs[0].gpu_addr;
 235
 236         dst_addr = amdgpu_bo_gpu_offset(adev->gart.bo);
 237         dst_addr += window * AMDGPU_GTT_MAX_TRANSFER_SIZE * 8;
 238         amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_addr,
 239                                 dst_addr, num_bytes, 0);
 240
 241         amdgpu_ring_pad_ib(ring, &job->ibs[0]);
 242         WARN_ON(job->ibs[0].length_dw > num_dw);
 243
 244         flags = amdgpu_ttm_tt_pte_flags(adev, bo->ttm, mem);
 245         if (tmz)
 246                 flags |= AMDGPU_PTE_TMZ;
 247
 248         cpu_addr = &job->ibs[0].ptr[num_dw];
 249
 250         if (mem->mem_type == TTM_PL_TT) {
 251                 dma_addr_t *dma_addr;
 252
 253                 dma_addr = &bo->ttm->dma_address[mm_cur->start >> PAGE_SHIFT];
 254                 amdgpu_gart_map(adev, 0, num_pages, dma_addr, flags, cpu_addr);
 255         } else {
 256                 dma_addr_t dma_address;
 257
 258                 dma_address = mm_cur->start;
 259                 dma_address += adev->vm_manager.vram_base_offset;
 260
 261                 for (i = 0; i < num_pages; ++i) {
 262                         amdgpu_gart_map(adev, i << PAGE_SHIFT, 1, &dma_address,
 263                                         flags, cpu_addr);
 264                         dma_address += PAGE_SIZE;
 265                 }
 266         }
 267
 268         dma_fence_put(amdgpu_job_submit(job));
 269         return 0;
 270 }
 271
 272 /**
 273  * amdgpu_ttm_copy_mem_to_mem - Helper function for copy
 274  * @adev: amdgpu device
 275  * @src: buffer/address where to read from
 276  * @dst: buffer/address where to write to
 277  * @size: number of bytes to copy
 278  * @tmz: if a secure copy should be used
 279  * @resv: resv object to sync to
 280  * @f: Returns the last fence if multiple jobs are submitted.
 281  *
 282  * The function copies @size bytes from {src->mem + src->offset} to
 283  * {dst->mem + dst->offset}. src->bo and dst->bo could be same BO for a
 284  * move and different for a BO to BO copy.
 285  *
 286  */
 287 int amdgpu_ttm_copy_mem_to_mem(struct amdgpu_device *adev,
 288                                const struct amdgpu_copy_mem *src,
 289                                const struct amdgpu_copy_mem *dst,
 290                                uint64_t size, bool tmz,
 291                                struct dma_resv *resv,
 292                                struct dma_fence **f)
 293 {
 294         struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
 295         struct amdgpu_res_cursor src_mm, dst_mm;
 296         struct dma_fence *fence = NULL;
 297         int r = 0;
 298         uint32_t copy_flags = 0;
 299         struct amdgpu_bo *abo_src, *abo_dst;
 300
 301         if (!adev->mman.buffer_funcs_enabled) {
 302                 DRM_ERROR("Trying to move memory with ring turned off.\n");
 303                 return -EINVAL;
 304         }
 305
 306         amdgpu_res_first(src->mem, src->offset, size, &src_mm);
 307         amdgpu_res_first(dst->mem, dst->offset, size, &dst_mm);
 308
 309         mutex_lock(&adev->mman.gtt_window_lock);
 310         while (src_mm.remaining) {
 311                 uint64_t from, to, cur_size, tiling_flags;
 312                 uint32_t num_type, data_format, max_com, write_compress_disable;
 313                 struct dma_fence *next;
 314
 315                 /* Never copy more than 256MiB at once to avoid a timeout */
 316                 cur_size = min3(src_mm.size, dst_mm.size, 256ULL << 20);
 317
 318                 /* Map src to window 0 and dst to window 1. */
 319                 r = amdgpu_ttm_map_buffer(src->bo, src->mem, &src_mm,
 320                                           0, ring, tmz, &cur_size, &from);
 321                 if (r)
 322                         goto error;
 323
 324                 r = amdgpu_ttm_map_buffer(dst->bo, dst->mem, &dst_mm,
 325                                           1, ring, tmz, &cur_size, &to);
 326                 if (r)
 327                         goto error;
 328
 329                 abo_src = ttm_to_amdgpu_bo(src->bo);
 330                 abo_dst = ttm_to_amdgpu_bo(dst->bo);
 331                 if (tmz)
 332                         copy_flags |= AMDGPU_COPY_FLAGS_TMZ;
 333                 if ((abo_src->flags & AMDGPU_GEM_CREATE_GFX12_DCC) &&
 334                     (abo_src->tbo.resource->mem_type == TTM_PL_VRAM))
 335                         copy_flags |= AMDGPU_COPY_FLAGS_READ_DECOMPRESSED;
 336                 if ((abo_dst->flags & AMDGPU_GEM_CREATE_GFX12_DCC) &&
 337                     (dst->mem->mem_type == TTM_PL_VRAM)) {
 338                         copy_flags |= AMDGPU_COPY_FLAGS_WRITE_COMPRESSED;
 339                         amdgpu_bo_get_tiling_flags(abo_dst, &tiling_flags);
 340                         max_com = AMDGPU_TILING_GET(tiling_flags, GFX12_DCC_MAX_COMPRESSED_BLOCK);
 341                         num_type = AMDGPU_TILING_GET(tiling_flags, GFX12_DCC_NUMBER_TYPE);
 342                         data_format = AMDGPU_TILING_GET(tiling_flags, GFX12_DCC_DATA_FORMAT);
 343                         write_compress_disable =
 344                                 AMDGPU_TILING_GET(tiling_flags, GFX12_DCC_WRITE_COMPRESS_DISABLE);
 345                         copy_flags |= (AMDGPU_COPY_FLAGS_SET(MAX_COMPRESSED, max_com) |
 346                                        AMDGPU_COPY_FLAGS_SET(NUMBER_TYPE, num_type) |
 347                                        AMDGPU_COPY_FLAGS_SET(DATA_FORMAT, data_format) |
 348                                        AMDGPU_COPY_FLAGS_SET(WRITE_COMPRESS_DISABLE,
 349                                                              write_compress_disable));
 350                 }
 351
 352                 r = amdgpu_copy_buffer(ring, from, to, cur_size, resv,
 353                                        &next, false, true, copy_flags);
 354                 if (r)
 355                         goto error;
 356
 357                 dma_fence_put(fence);
 358                 fence = next;
 359
 360                 amdgpu_res_next(&src_mm, cur_size);
 361                 amdgpu_res_next(&dst_mm, cur_size);
 362         }
 363 error:
 364         mutex_unlock(&adev->mman.gtt_window_lock);
 365         if (f)
 366                 *f = dma_fence_get(fence);
 367         dma_fence_put(fence);
 368         return r;
 369 }
 370
 371 /*
 372  * amdgpu_move_blit - Copy an entire buffer to another buffer
 373  *
 374  * This is a helper called by amdgpu_bo_move() and amdgpu_move_vram_ram() to
 375  * help move buffers to and from VRAM.
 376  */
 377 static int amdgpu_move_blit(struct ttm_buffer_object *bo,
 378                             bool evict,
 379                             struct ttm_resource *new_mem,
 380                             struct ttm_resource *old_mem)
 381 {
 382         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
 383         struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
 384         struct amdgpu_copy_mem src, dst;
 385         struct dma_fence *fence = NULL;
 386         int r;
 387
 388         src.bo = bo;
 389         dst.bo = bo;
 390         src.mem = old_mem;
 391         dst.mem = new_mem;
 392         src.offset = 0;
 393         dst.offset = 0;
 394
 395         r = amdgpu_ttm_copy_mem_to_mem(adev, &src, &dst,
 396                                        new_mem->size,
 397                                        amdgpu_bo_encrypted(abo),
 398                                        bo->base.resv, &fence);
 399         if (r)
 400                 goto error;
 401
 402         /* clear the space being freed */
 403         if (old_mem->mem_type == TTM_PL_VRAM &&
 404             (abo->flags & AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE)) {
 405                 struct dma_fence *wipe_fence = NULL;
 406
 407                 r = amdgpu_fill_buffer(abo, 0, NULL, &wipe_fence,
 408                                        false);
 409                 if (r) {
 410                         goto error;
 411                 } else if (wipe_fence) {
 412                         amdgpu_vram_mgr_set_cleared(bo->resource);
 413                         dma_fence_put(fence);
 414                         fence = wipe_fence;
 415                 }
 416         }
 417
 418         /* Always block for VM page tables before committing the new location */
 419         if (bo->type == ttm_bo_type_kernel)
 420                 r = ttm_bo_move_accel_cleanup(bo, fence, true, false, new_mem);
 421         else
 422                 r = ttm_bo_move_accel_cleanup(bo, fence, evict, true, new_mem);
 423         dma_fence_put(fence);
 424         return r;
 425
 426 error:
 427         if (fence)
 428                 dma_fence_wait(fence, false);
 429         dma_fence_put(fence);
 430         return r;
 431 }
 432
 433 /**
 434  * amdgpu_res_cpu_visible - Check that resource can be accessed by CPU
 435  * @adev: amdgpu device
 436  * @res: the resource to check
 437  *
 438  * Returns: true if the full resource is CPU visible, false otherwise.
 439  */
 440 bool amdgpu_res_cpu_visible(struct amdgpu_device *adev,
 441                             struct ttm_resource *res)
 442 {
 443         struct amdgpu_res_cursor cursor;
 444
 445         if (!res)
 446                 return false;
 447
 448         if (res->mem_type == TTM_PL_SYSTEM || res->mem_type == TTM_PL_TT ||
 449             res->mem_type == AMDGPU_PL_PREEMPT || res->mem_type == AMDGPU_PL_DOORBELL)
 450                 return true;
 451
 452         if (res->mem_type != TTM_PL_VRAM)
 453                 return false;
 454
 455         amdgpu_res_first(res, 0, res->size, &cursor);
 456         while (cursor.remaining) {
 457                 if ((cursor.start + cursor.size) > adev->gmc.visible_vram_size)
 458                         return false;
 459                 amdgpu_res_next(&cursor, cursor.size);
 460         }
 461
 462         return true;
 463 }
 464
 465 /*
 466  * amdgpu_res_copyable - Check that memory can be accessed by ttm_bo_move_memcpy
 467  *
 468  * Called by amdgpu_bo_move()
 469  */
 470 static bool amdgpu_res_copyable(struct amdgpu_device *adev,
 471                                 struct ttm_resource *mem)
 472 {
 473         if (!amdgpu_res_cpu_visible(adev, mem))
 474                 return false;
 475
 476         /* ttm_resource_ioremap only supports contiguous memory */
 477         if (mem->mem_type == TTM_PL_VRAM &&
 478             !(mem->placement & TTM_PL_FLAG_CONTIGUOUS))
 479                 return false;
 480
 481         return true;
 482 }
 483
 484 /*
 485  * amdgpu_bo_move - Move a buffer object to a new memory location
 486  *
 487  * Called by ttm_bo_handle_move_mem()
 488  */
 489 static int amdgpu_bo_move(struct ttm_buffer_object *bo, bool evict,
 490                           struct ttm_operation_ctx *ctx,
 491                           struct ttm_resource *new_mem,
 492                           struct ttm_place *hop)
 493 {
 494         struct amdgpu_device *adev;
 495         struct amdgpu_bo *abo;
 496         struct ttm_resource *old_mem = bo->resource;
 497         int r;
 498
 499         if (new_mem->mem_type == TTM_PL_TT ||
 500             new_mem->mem_type == AMDGPU_PL_PREEMPT) {
 501                 r = amdgpu_ttm_backend_bind(bo->bdev, bo->ttm, new_mem);
 502                 if (r)
 503                         return r;
 504         }
 505
 506         abo = ttm_to_amdgpu_bo(bo);
 507         adev = amdgpu_ttm_adev(bo->bdev);
 508
 509         if (!old_mem || (old_mem->mem_type == TTM_PL_SYSTEM &&
 510                          bo->ttm == NULL)) {
 511                 amdgpu_bo_move_notify(bo, evict, new_mem);
 512                 ttm_bo_move_null(bo, new_mem);
 513                 return 0;
 514         }
 515         if (old_mem->mem_type == TTM_PL_SYSTEM &&
 516             (new_mem->mem_type == TTM_PL_TT ||
 517              new_mem->mem_type == AMDGPU_PL_PREEMPT)) {
 518                 amdgpu_bo_move_notify(bo, evict, new_mem);
 519                 ttm_bo_move_null(bo, new_mem);
 520                 return 0;
 521         }
 522         if ((old_mem->mem_type == TTM_PL_TT ||
 523              old_mem->mem_type == AMDGPU_PL_PREEMPT) &&
 524             new_mem->mem_type == TTM_PL_SYSTEM) {
 525                 r = ttm_bo_wait_ctx(bo, ctx);
 526                 if (r)
 527                         return r;
 528
 529                 amdgpu_ttm_backend_unbind(bo->bdev, bo->ttm);
 530                 amdgpu_bo_move_notify(bo, evict, new_mem);
 531                 ttm_resource_free(bo, &bo->resource);
 532                 ttm_bo_assign_mem(bo, new_mem);
 533                 return 0;
 534         }
 535
 536         if (old_mem->mem_type == AMDGPU_PL_GDS ||
 537             old_mem->mem_type == AMDGPU_PL_GWS ||
 538             old_mem->mem_type == AMDGPU_PL_OA ||
 539             old_mem->mem_type == AMDGPU_PL_DOORBELL ||
 540             new_mem->mem_type == AMDGPU_PL_GDS ||
 541             new_mem->mem_type == AMDGPU_PL_GWS ||
 542             new_mem->mem_type == AMDGPU_PL_OA ||
 543             new_mem->mem_type == AMDGPU_PL_DOORBELL) {
 544                 /* Nothing to save here */
 545                 amdgpu_bo_move_notify(bo, evict, new_mem);
 546                 ttm_bo_move_null(bo, new_mem);
 547                 return 0;
 548         }
 549
 550         if (bo->type == ttm_bo_type_device &&
 551             new_mem->mem_type == TTM_PL_VRAM &&
 552             old_mem->mem_type != TTM_PL_VRAM) {
 553                 /* amdgpu_bo_fault_reserve_notify will re-set this if the CPU
 554                  * accesses the BO after it's moved.
 555                  */
 556                 abo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
 557         }
 558
 559         if (adev->mman.buffer_funcs_enabled &&
 560             ((old_mem->mem_type == TTM_PL_SYSTEM &&
 561               new_mem->mem_type == TTM_PL_VRAM) ||
 562              (old_mem->mem_type == TTM_PL_VRAM &&
 563               new_mem->mem_type == TTM_PL_SYSTEM))) {
 564                 hop->fpfn = 0;
 565                 hop->lpfn = 0;
 566                 hop->mem_type = TTM_PL_TT;
 567                 hop->flags = TTM_PL_FLAG_TEMPORARY;
 568                 return -EMULTIHOP;
 569         }
 570
 571         amdgpu_bo_move_notify(bo, evict, new_mem);
 572         if (adev->mman.buffer_funcs_enabled)
 573                 r = amdgpu_move_blit(bo, evict, new_mem, old_mem);
 574         else
 575                 r = -ENODEV;
 576
 577         if (r) {
 578                 /* Check that all memory is CPU accessible */
 579                 if (!amdgpu_res_copyable(adev, old_mem) ||
 580                     !amdgpu_res_copyable(adev, new_mem)) {
 581                         pr_err("Move buffer fallback to memcpy unavailable\n");
 582                         return r;
 583                 }
 584
 585                 r = ttm_bo_move_memcpy(bo, ctx, new_mem);
 586                 if (r)
 587                         return r;
 588         }
 589
 590         /* update statistics after the move */
 591         if (evict)
 592                 atomic64_inc(&adev->num_evictions);
 593         atomic64_add(bo->base.size, &adev->num_bytes_moved);
 594         return 0;
 595 }
 596
 597 /*
 598  * amdgpu_ttm_io_mem_reserve - Reserve a block of memory during a fault
 599  *
 600  * Called by ttm_mem_io_reserve() ultimately via ttm_bo_vm_fault()
 601  */
 602 static int amdgpu_ttm_io_mem_reserve(struct ttm_device *bdev,
 603                                      struct ttm_resource *mem)
 604 {
 605         struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
 606
 607         switch (mem->mem_type) {
 608         case TTM_PL_SYSTEM:
 609                 /* system memory */
 610                 return 0;
 611         case TTM_PL_TT:
 612         case AMDGPU_PL_PREEMPT:
 613                 break;
 614         case TTM_PL_VRAM:
 615                 mem->bus.offset = mem->start << PAGE_SHIFT;
 616
 617                 if (adev->mman.aper_base_kaddr &&
 618                     mem->placement & TTM_PL_FLAG_CONTIGUOUS)
 619                         mem->bus.addr = (u8 *)adev->mman.aper_base_kaddr +
 620                                         mem->bus.offset;
 621
 622                 mem->bus.offset += adev->gmc.aper_base;
 623                 mem->bus.is_iomem = true;
 624                 break;
 625         case AMDGPU_PL_DOORBELL:
 626                 mem->bus.offset = mem->start << PAGE_SHIFT;
 627                 mem->bus.offset += adev->doorbell.base;
 628                 mem->bus.is_iomem = true;
 629                 mem->bus.caching = ttm_uncached;
 630                 break;
 631         default:
 632                 return -EINVAL;
 633         }
 634         return 0;
 635 }
 636
 637 static unsigned long amdgpu_ttm_io_mem_pfn(struct ttm_buffer_object *bo,
 638                                            unsigned long page_offset)
 639 {
 640         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
 641         struct amdgpu_res_cursor cursor;
 642
 643         amdgpu_res_first(bo->resource, (u64)page_offset << PAGE_SHIFT, 0,
 644                          &cursor);
 645
 646         if (bo->resource->mem_type == AMDGPU_PL_DOORBELL)
 647                 return ((uint64_t)(adev->doorbell.base + cursor.start)) >> PAGE_SHIFT;
 648
 649         return (adev->gmc.aper_base + cursor.start) >> PAGE_SHIFT;
 650 }
 651
 652 /**
 653  * amdgpu_ttm_domain_start - Returns GPU start address
 654  * @adev: amdgpu device object
 655  * @type: type of the memory
 656  *
 657  * Returns:
 658  * GPU start address of a memory domain
 659  */
 660
 661 uint64_t amdgpu_ttm_domain_start(struct amdgpu_device *adev, uint32_t type)
 662 {
 663         switch (type) {
 664         case TTM_PL_TT:
 665                 return adev->gmc.gart_start;
 666         case TTM_PL_VRAM:
 667                 return adev->gmc.vram_start;
 668         }
 669
 670         return 0;
 671 }
 672
 673 /*
 674  * TTM backend functions.
 675  */
 676 struct amdgpu_ttm_tt {
 677         struct ttm_tt   ttm;
 678         struct drm_gem_object   *gobj;
 679         u64                     offset;
 680         uint64_t                userptr;
 681         struct task_struct      *usertask;
 682         uint32_t                userflags;
 683         bool                    bound;
 684         int32_t                 pool_id;
 685 };
 686
 687 #define ttm_to_amdgpu_ttm_tt(ptr)       container_of(ptr, struct amdgpu_ttm_tt, ttm)
 688
 689 #ifdef CONFIG_DRM_AMDGPU_USERPTR
 690 /*
 691  * amdgpu_ttm_tt_get_user_pages - get device accessible pages that back user
 692  * memory and start HMM tracking CPU page table update
 693  *
 694  * Calling function must call amdgpu_ttm_tt_userptr_range_done() once and only
 695  * once afterwards to stop HMM tracking
 696  */
 697 int amdgpu_ttm_tt_get_user_pages(struct amdgpu_bo *bo, struct page **pages,
 698                                  struct hmm_range **range)
 699 {
 700         struct ttm_tt *ttm = bo->tbo.ttm;
 701         struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
 702         unsigned long start = gtt->userptr;
 703         struct vm_area_struct *vma;
 704         struct mm_struct *mm;
 705         bool readonly;
 706         int r = 0;
 707
 708         /* Make sure get_user_pages_done() can cleanup gracefully */
 709         *range = NULL;
 710
 711         mm = bo->notifier.mm;
 712         if (unlikely(!mm)) {
 713                 DRM_DEBUG_DRIVER("BO is not registered?\n");
 714                 return -EFAULT;
 715         }
 716
 717         if (!mmget_not_zero(mm)) /* Happens during process shutdown */
 718                 return -ESRCH;
 719
 720         mmap_read_lock(mm);
 721         vma = vma_lookup(mm, start);
 722         if (unlikely(!vma)) {
 723                 r = -EFAULT;
 724                 goto out_unlock;
 725         }
 726         if (unlikely((gtt->userflags & AMDGPU_GEM_USERPTR_ANONONLY) &&
 727                 vma->vm_file)) {
 728                 r = -EPERM;
 729                 goto out_unlock;
 730         }
 731
 732         readonly = amdgpu_ttm_tt_is_readonly(ttm);
 733         r = amdgpu_hmm_range_get_pages(&bo->notifier, start, ttm->num_pages,
 734                                        readonly, NULL, pages, range);
 735 out_unlock:
 736         mmap_read_unlock(mm);
 737         if (r)
 738                 pr_debug("failed %d to get user pages 0x%lx\n", r, start);
 739
 740         mmput(mm);
 741
 742         return r;
 743 }
 744
 745 /* amdgpu_ttm_tt_discard_user_pages - Discard range and pfn array allocations
 746  */
 747 void amdgpu_ttm_tt_discard_user_pages(struct ttm_tt *ttm,
 748                                       struct hmm_range *range)
 749 {
 750         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 751
 752         if (gtt && gtt->userptr && range)
 753                 amdgpu_hmm_range_get_pages_done(range);
 754 }
 755
 756 /*
 757  * amdgpu_ttm_tt_get_user_pages_done - stop HMM track the CPU page table change
 758  * Check if the pages backing this ttm range have been invalidated
 759  *
 760  * Returns: true if pages are still valid
 761  */
 762 bool amdgpu_ttm_tt_get_user_pages_done(struct ttm_tt *ttm,
 763                                        struct hmm_range *range)
 764 {
 765         struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
 766
 767         if (!gtt || !gtt->userptr || !range)
 768                 return false;
 769
 770         DRM_DEBUG_DRIVER("user_pages_done 0x%llx pages 0x%x\n",
 771                 gtt->userptr, ttm->num_pages);
 772
 773         WARN_ONCE(!range->hmm_pfns, "No user pages to check\n");
 774
 775         return !amdgpu_hmm_range_get_pages_done(range);
 776 }
 777 #endif
 778
 779 /*
 780  * amdgpu_ttm_tt_set_user_pages - Copy pages in, putting old pages as necessary.
 781  *
 782  * Called by amdgpu_cs_list_validate(). This creates the page list
 783  * that backs user memory and will ultimately be mapped into the device
 784  * address space.
 785  */
 786 void amdgpu_ttm_tt_set_user_pages(struct ttm_tt *ttm, struct page **pages)
 787 {
 788         unsigned long i;
 789
 790         for (i = 0; i < ttm->num_pages; ++i)
 791                 ttm->pages[i] = pages ? pages[i] : NULL;
 792 }
 793
 794 /*
 795  * amdgpu_ttm_tt_pin_userptr - prepare the sg table with the user pages
 796  *
 797  * Called by amdgpu_ttm_backend_bind()
 798  **/
 799 static int amdgpu_ttm_tt_pin_userptr(struct ttm_device *bdev,
 800                                      struct ttm_tt *ttm)
 801 {
 802         struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
 803         struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
 804         int write = !(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
 805         enum dma_data_direction direction = write ?
 806                 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
 807         int r;
 808
 809         /* Allocate an SG array and squash pages into it */
 810         r = sg_alloc_table_from_pages(ttm->sg, ttm->pages, ttm->num_pages, 0,
 811                                       (u64)ttm->num_pages << PAGE_SHIFT,
 812                                       GFP_KERNEL);
 813         if (r)
 814                 goto release_sg;
 815
 816         /* Map SG to device */
 817         r = dma_map_sgtable(adev->dev, ttm->sg, direction, 0);
 818         if (r)
 819                 goto release_sg_table;
 820
 821         /* convert SG to linear array of pages and dma addresses */
 822         drm_prime_sg_to_dma_addr_array(ttm->sg, gtt->ttm.dma_address,
 823                                        ttm->num_pages);
 824
 825         return 0;
 826
 827 release_sg_table:
 828         sg_free_table(ttm->sg);
 829 release_sg:
 830         kfree(ttm->sg);
 831         ttm->sg = NULL;
 832         return r;
 833 }
 834
 835 /*
 836  * amdgpu_ttm_tt_unpin_userptr - Unpin and unmap userptr pages
 837  */
 838 static void amdgpu_ttm_tt_unpin_userptr(struct ttm_device *bdev,
 839                                         struct ttm_tt *ttm)
 840 {
 841         struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
 842         struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
 843         int write = !(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
 844         enum dma_data_direction direction = write ?
 845                 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
 846
 847         /* double check that we don't free the table twice */
 848         if (!ttm->sg || !ttm->sg->sgl)
 849                 return;
 850
 851         /* unmap the pages mapped to the device */
 852         dma_unmap_sgtable(adev->dev, ttm->sg, direction, 0);
 853         sg_free_table(ttm->sg);
 854 }
 855
 856 /*
 857  * total_pages is constructed as MQD0+CtrlStack0 + MQD1+CtrlStack1 + ...
 858  * MQDn+CtrlStackn where n is the number of XCCs per partition.
 859  * pages_per_xcc is the size of one MQD+CtrlStack. The first page is MQD
 860  * and uses memory type default, UC. The rest of pages_per_xcc are
 861  * Ctrl stack and modify their memory type to NC.
 862  */
 863 static void amdgpu_ttm_gart_bind_gfx9_mqd(struct amdgpu_device *adev,
 864                                 struct ttm_tt *ttm, uint64_t flags)
 865 {
 866         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 867         uint64_t total_pages = ttm->num_pages;
 868         int num_xcc = max(1U, adev->gfx.num_xcc_per_xcp);
 869         uint64_t page_idx, pages_per_xcc;
 870         int i;
 871         uint64_t ctrl_flags = AMDGPU_PTE_MTYPE_VG10(flags, AMDGPU_MTYPE_NC);
 872
 873         pages_per_xcc = total_pages;
 874         do_div(pages_per_xcc, num_xcc);
 875
 876         for (i = 0, page_idx = 0; i < num_xcc; i++, page_idx += pages_per_xcc) {
 877                 /* MQD page: use default flags */
 878                 amdgpu_gart_bind(adev,
 879                                 gtt->offset + (page_idx << PAGE_SHIFT),
 880                                 1, &gtt->ttm.dma_address[page_idx], flags);
 881                 /*
 882                  * Ctrl pages - modify the memory type to NC (ctrl_flags) from
 883                  * the second page of the BO onward.
 884                  */
 885                 amdgpu_gart_bind(adev,
 886                                 gtt->offset + ((page_idx + 1) << PAGE_SHIFT),
 887                                 pages_per_xcc - 1,
 888                                 &gtt->ttm.dma_address[page_idx + 1],
 889                                 ctrl_flags);
 890         }
 891 }
 892
 893 static void amdgpu_ttm_gart_bind(struct amdgpu_device *adev,
 894                                  struct ttm_buffer_object *tbo,
 895                                  uint64_t flags)
 896 {
 897         struct amdgpu_bo *abo = ttm_to_amdgpu_bo(tbo);
 898         struct ttm_tt *ttm = tbo->ttm;
 899         struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
 900
 901         if (amdgpu_bo_encrypted(abo))
 902                 flags |= AMDGPU_PTE_TMZ;
 903
 904         if (abo->flags & AMDGPU_GEM_CREATE_CP_MQD_GFX9) {
 905                 amdgpu_ttm_gart_bind_gfx9_mqd(adev, ttm, flags);
 906         } else {
 907                 amdgpu_gart_bind(adev, gtt->offset, ttm->num_pages,
 908                                  gtt->ttm.dma_address, flags);
 909         }
 910         gtt->bound = true;
 911 }
 912
 913 /*
 914  * amdgpu_ttm_backend_bind - Bind GTT memory
 915  *
 916  * Called by ttm_tt_bind() on behalf of ttm_bo_handle_move_mem().
 917  * This handles binding GTT memory to the device address space.
 918  */
 919 static int amdgpu_ttm_backend_bind(struct ttm_device *bdev,
 920                                    struct ttm_tt *ttm,
 921                                    struct ttm_resource *bo_mem)
 922 {
 923         struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
 924         struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
 925         uint64_t flags;
 926         int r;
 927
 928         if (!bo_mem)
 929                 return -EINVAL;
 930
 931         if (gtt->bound)
 932                 return 0;
 933
 934         if (gtt->userptr) {
 935                 r = amdgpu_ttm_tt_pin_userptr(bdev, ttm);
 936                 if (r) {
 937                         DRM_ERROR("failed to pin userptr\n");
 938                         return r;
 939                 }
 940         } else if (ttm->page_flags & TTM_TT_FLAG_EXTERNAL) {
 941                 if (!ttm->sg) {
 942                         struct dma_buf_attachment *attach;
 943                         struct sg_table *sgt;
 944
 945                         attach = gtt->gobj->import_attach;
 946                         sgt = dma_buf_map_attachment(attach, DMA_BIDIRECTIONAL);
 947                         if (IS_ERR(sgt))
 948                                 return PTR_ERR(sgt);
 949
 950                         ttm->sg = sgt;
 951                 }
 952
 953                 drm_prime_sg_to_dma_addr_array(ttm->sg, gtt->ttm.dma_address,
 954                                                ttm->num_pages);
 955         }
 956
 957         if (!ttm->num_pages) {
 958                 WARN(1, "nothing to bind %u pages for mreg %p back %p!\n",
 959                      ttm->num_pages, bo_mem, ttm);
 960         }
 961
 962         if (bo_mem->mem_type != TTM_PL_TT ||
 963             !amdgpu_gtt_mgr_has_gart_addr(bo_mem)) {
 964                 gtt->offset = AMDGPU_BO_INVALID_OFFSET;
 965                 return 0;
 966         }
 967
 968         /* compute PTE flags relevant to this BO memory */
 969         flags = amdgpu_ttm_tt_pte_flags(adev, ttm, bo_mem);
 970
 971         /* bind pages into GART page tables */
 972         gtt->offset = (u64)bo_mem->start << PAGE_SHIFT;
 973         amdgpu_gart_bind(adev, gtt->offset, ttm->num_pages,
 974                          gtt->ttm.dma_address, flags);
 975         gtt->bound = true;
 976         return 0;
 977 }
 978
 979 /*
 980  * amdgpu_ttm_alloc_gart - Make sure buffer object is accessible either
 981  * through AGP or GART aperture.
 982  *
 983  * If bo is accessible through AGP aperture, then use AGP aperture
 984  * to access bo; otherwise allocate logical space in GART aperture
 985  * and map bo to GART aperture.
 986  */
 987 int amdgpu_ttm_alloc_gart(struct ttm_buffer_object *bo)
 988 {
 989         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
 990         struct ttm_operation_ctx ctx = { false, false };
 991         struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(bo->ttm);
 992         struct ttm_placement placement;
 993         struct ttm_place placements;
 994         struct ttm_resource *tmp;
 995         uint64_t addr, flags;
 996         int r;
 997
 998         if (bo->resource->start != AMDGPU_BO_INVALID_OFFSET)
 999                 return 0;
1000
1001         addr = amdgpu_gmc_agp_addr(bo);
1002         if (addr != AMDGPU_BO_INVALID_OFFSET)
1003                 return 0;
1004
1005         /* allocate GART space */
1006         placement.num_placement = 1;
1007         placement.placement = &placements;
1008         placements.fpfn = 0;
1009         placements.lpfn = adev->gmc.gart_size >> PAGE_SHIFT;
1010         placements.mem_type = TTM_PL_TT;
1011         placements.flags = bo->resource->placement;
1012
1013         r = ttm_bo_mem_space(bo, &placement, &tmp, &ctx);
1014         if (unlikely(r))
1015                 return r;
1016
1017         /* compute PTE flags for this buffer object */
1018         flags = amdgpu_ttm_tt_pte_flags(adev, bo->ttm, tmp);
1019
1020         /* Bind pages */
1021         gtt->offset = (u64)tmp->start << PAGE_SHIFT;
1022         amdgpu_ttm_gart_bind(adev, bo, flags);
1023         amdgpu_gart_invalidate_tlb(adev);
1024         ttm_resource_free(bo, &bo->resource);
1025         ttm_bo_assign_mem(bo, tmp);
1026
1027         return 0;
1028 }
1029
1030 /*
1031  * amdgpu_ttm_recover_gart - Rebind GTT pages
1032  *
1033  * Called by amdgpu_gtt_mgr_recover() from amdgpu_device_reset() to
1034  * rebind GTT pages during a GPU reset.
1035  */
1036 void amdgpu_ttm_recover_gart(struct ttm_buffer_object *tbo)
1037 {
1038         struct amdgpu_device *adev = amdgpu_ttm_adev(tbo->bdev);
1039         uint64_t flags;
1040
1041         if (!tbo->ttm)
1042                 return;
1043
1044         flags = amdgpu_ttm_tt_pte_flags(adev, tbo->ttm, tbo->resource);
1045         amdgpu_ttm_gart_bind(adev, tbo, flags);
1046 }
1047
1048 /*
1049  * amdgpu_ttm_backend_unbind - Unbind GTT mapped pages
1050  *
1051  * Called by ttm_tt_unbind() on behalf of ttm_bo_move_ttm() and
1052  * ttm_tt_destroy().
1053  */
1054 static void amdgpu_ttm_backend_unbind(struct ttm_device *bdev,
1055                                       struct ttm_tt *ttm)
1056 {
1057         struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
1058         struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
1059
1060         /* if the pages have userptr pinning then clear that first */
1061         if (gtt->userptr) {
1062                 amdgpu_ttm_tt_unpin_userptr(bdev, ttm);
1063         } else if (ttm->sg && gtt->gobj->import_attach) {
1064                 struct dma_buf_attachment *attach;
1065
1066                 attach = gtt->gobj->import_attach;
1067                 dma_buf_unmap_attachment(attach, ttm->sg, DMA_BIDIRECTIONAL);
1068                 ttm->sg = NULL;
1069         }
1070
1071         if (!gtt->bound)
1072                 return;
1073
1074         if (gtt->offset == AMDGPU_BO_INVALID_OFFSET)
1075                 return;
1076
1077         /* unbind shouldn't be done for GDS/GWS/OA in ttm_bo_clean_mm */
1078         amdgpu_gart_unbind(adev, gtt->offset, ttm->num_pages);
1079         gtt->bound = false;
1080 }
1081
1082 static void amdgpu_ttm_backend_destroy(struct ttm_device *bdev,
1083                                        struct ttm_tt *ttm)
1084 {
1085         struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
1086
1087         if (gtt->usertask)
1088                 put_task_struct(gtt->usertask);
1089
1090         ttm_tt_fini(&gtt->ttm);
1091         kfree(gtt);
1092 }
1093
1094 /**
1095  * amdgpu_ttm_tt_create - Create a ttm_tt object for a given BO
1096  *
1097  * @bo: The buffer object to create a GTT ttm_tt object around
1098  * @page_flags: Page flags to be added to the ttm_tt object
1099  *
1100  * Called by ttm_tt_create().
1101  */
1102 static struct ttm_tt *amdgpu_ttm_tt_create(struct ttm_buffer_object *bo,
1103                                            uint32_t page_flags)
1104 {
1105         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1106         struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
1107         struct amdgpu_ttm_tt *gtt;
1108         enum ttm_caching caching;
1109
1110         gtt = kzalloc(sizeof(struct amdgpu_ttm_tt), GFP_KERNEL);
1111         if (!gtt)
1112                 return NULL;
1113
1114         gtt->gobj = &bo->base;
1115         if (adev->gmc.mem_partitions && abo->xcp_id >= 0)
1116                 gtt->pool_id = KFD_XCP_MEM_ID(adev, abo->xcp_id);
1117         else
1118                 gtt->pool_id = abo->xcp_id;
1119
1120         if (abo->flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
1121                 caching = ttm_write_combined;
1122         else
1123                 caching = ttm_cached;
1124
1125         /* allocate space for the uninitialized page entries */
1126         if (ttm_sg_tt_init(&gtt->ttm, bo, page_flags, caching)) {
1127                 kfree(gtt);
1128                 return NULL;
1129         }
1130         return &gtt->ttm;
1131 }
1132
1133 /*
1134  * amdgpu_ttm_tt_populate - Map GTT pages visible to the device
1135  *
1136  * Map the pages of a ttm_tt object to an address space visible
1137  * to the underlying device.
1138  */
1139 static int amdgpu_ttm_tt_populate(struct ttm_device *bdev,
1140                                   struct ttm_tt *ttm,
1141                                   struct ttm_operation_ctx *ctx)
1142 {
1143         struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
1144         struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
1145         struct ttm_pool *pool;
1146         pgoff_t i;
1147         int ret;
1148
1149         /* user pages are bound by amdgpu_ttm_tt_pin_userptr() */
1150         if (gtt->userptr) {
1151                 ttm->sg = kzalloc(sizeof(struct sg_table), GFP_KERNEL);
1152                 if (!ttm->sg)
1153                         return -ENOMEM;
1154                 return 0;
1155         }
1156
1157         if (ttm->page_flags & TTM_TT_FLAG_EXTERNAL)
1158                 return 0;
1159
1160         if (adev->mman.ttm_pools && gtt->pool_id >= 0)
1161                 pool = &adev->mman.ttm_pools[gtt->pool_id];
1162         else
1163                 pool = &adev->mman.bdev.pool;
1164         ret = ttm_pool_alloc(pool, ttm, ctx);
1165         if (ret)
1166                 return ret;
1167
1168         for (i = 0; i < ttm->num_pages; ++i)
1169                 ttm->pages[i]->mapping = bdev->dev_mapping;
1170
1171         return 0;
1172 }
1173
1174 /*
1175  * amdgpu_ttm_tt_unpopulate - unmap GTT pages and unpopulate page arrays
1176  *
1177  * Unmaps pages of a ttm_tt object from the device address space and
1178  * unpopulates the page array backing it.
1179  */
1180 static void amdgpu_ttm_tt_unpopulate(struct ttm_device *bdev,
1181                                      struct ttm_tt *ttm)
1182 {
1183         struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
1184         struct amdgpu_device *adev;
1185         struct ttm_pool *pool;
1186         pgoff_t i;
1187
1188         amdgpu_ttm_backend_unbind(bdev, ttm);
1189
1190         if (gtt->userptr) {
1191                 amdgpu_ttm_tt_set_user_pages(ttm, NULL);
1192                 kfree(ttm->sg);
1193                 ttm->sg = NULL;
1194                 return;
1195         }
1196
1197         if (ttm->page_flags & TTM_TT_FLAG_EXTERNAL)
1198                 return;
1199
1200         for (i = 0; i < ttm->num_pages; ++i)
1201                 ttm->pages[i]->mapping = NULL;
1202
1203         adev = amdgpu_ttm_adev(bdev);
1204
1205         if (adev->mman.ttm_pools && gtt->pool_id >= 0)
1206                 pool = &adev->mman.ttm_pools[gtt->pool_id];
1207         else
1208                 pool = &adev->mman.bdev.pool;
1209
1210         return ttm_pool_free(pool, ttm);
1211 }
1212
1213 /**
1214  * amdgpu_ttm_tt_get_userptr - Return the userptr GTT ttm_tt for the current
1215  * task
1216  *
1217  * @tbo: The ttm_buffer_object that contains the userptr
1218  * @user_addr:  The returned value
1219  */
1220 int amdgpu_ttm_tt_get_userptr(const struct ttm_buffer_object *tbo,
1221                               uint64_t *user_addr)
1222 {
1223         struct amdgpu_ttm_tt *gtt;
1224
1225         if (!tbo->ttm)
1226                 return -EINVAL;
1227
1228         gtt = (void *)tbo->ttm;
1229         *user_addr = gtt->userptr;
1230         return 0;
1231 }
1232
1233 /**
1234  * amdgpu_ttm_tt_set_userptr - Initialize userptr GTT ttm_tt for the current
1235  * task
1236  *
1237  * @bo: The ttm_buffer_object to bind this userptr to
1238  * @addr:  The address in the current tasks VM space to use
1239  * @flags: Requirements of userptr object.
1240  *
1241  * Called by amdgpu_gem_userptr_ioctl() and kfd_ioctl_alloc_memory_of_gpu() to
1242  * bind userptr pages to current task and by kfd_ioctl_acquire_vm() to
1243  * initialize GPU VM for a KFD process.
1244  */
1245 int amdgpu_ttm_tt_set_userptr(struct ttm_buffer_object *bo,
1246                               uint64_t addr, uint32_t flags)
1247 {
1248         struct amdgpu_ttm_tt *gtt;
1249
1250         if (!bo->ttm) {
1251                 /* TODO: We want a separate TTM object type for userptrs */
1252                 bo->ttm = amdgpu_ttm_tt_create(bo, 0);
1253                 if (bo->ttm == NULL)
1254                         return -ENOMEM;
1255         }
1256
1257         /* Set TTM_TT_FLAG_EXTERNAL before populate but after create. */
1258         bo->ttm->page_flags |= TTM_TT_FLAG_EXTERNAL;
1259
1260         gtt = ttm_to_amdgpu_ttm_tt(bo->ttm);
1261         gtt->userptr = addr;
1262         gtt->userflags = flags;
1263
1264         if (gtt->usertask)
1265                 put_task_struct(gtt->usertask);
1266         gtt->usertask = current->group_leader;
1267         get_task_struct(gtt->usertask);
1268
1269         return 0;
1270 }
1271
1272 /*
1273  * amdgpu_ttm_tt_get_usermm - Return memory manager for ttm_tt object
1274  */
1275 struct mm_struct *amdgpu_ttm_tt_get_usermm(struct ttm_tt *ttm)
1276 {
1277         struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
1278
1279         if (gtt == NULL)
1280                 return NULL;
1281
1282         if (gtt->usertask == NULL)
1283                 return NULL;
1284
1285         return gtt->usertask->mm;
1286 }
1287
1288 /*
1289  * amdgpu_ttm_tt_affect_userptr - Determine if a ttm_tt object lays inside an
1290  * address range for the current task.
1291  *
1292  */
1293 bool amdgpu_ttm_tt_affect_userptr(struct ttm_tt *ttm, unsigned long start,
1294                                   unsigned long end, unsigned long *userptr)
1295 {
1296         struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
1297         unsigned long size;
1298
1299         if (gtt == NULL || !gtt->userptr)
1300                 return false;
1301
1302         /* Return false if no part of the ttm_tt object lies within
1303          * the range
1304          */
1305         size = (unsigned long)gtt->ttm.num_pages * PAGE_SIZE;
1306         if (gtt->userptr > end || gtt->userptr + size <= start)
1307                 return false;
1308
1309         if (userptr)
1310                 *userptr = gtt->userptr;
1311         return true;
1312 }
1313
1314 /*
1315  * amdgpu_ttm_tt_is_userptr - Have the pages backing by userptr?
1316  */
1317 bool amdgpu_ttm_tt_is_userptr(struct ttm_tt *ttm)
1318 {
1319         struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
1320
1321         if (gtt == NULL || !gtt->userptr)
1322                 return false;
1323
1324         return true;
1325 }
1326
1327 /*
1328  * amdgpu_ttm_tt_is_readonly - Is the ttm_tt object read only?
1329  */
1330 bool amdgpu_ttm_tt_is_readonly(struct ttm_tt *ttm)
1331 {
1332         struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
1333
1334         if (gtt == NULL)
1335                 return false;
1336
1337         return !!(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
1338 }
1339
1340 /**
1341  * amdgpu_ttm_tt_pde_flags - Compute PDE flags for ttm_tt object
1342  *
1343  * @ttm: The ttm_tt object to compute the flags for
1344  * @mem: The memory registry backing this ttm_tt object
1345  *
1346  * Figure out the flags to use for a VM PDE (Page Directory Entry).
1347  */
1348 uint64_t amdgpu_ttm_tt_pde_flags(struct ttm_tt *ttm, struct ttm_resource *mem)
1349 {
1350         uint64_t flags = 0;
1351
1352         if (mem && mem->mem_type != TTM_PL_SYSTEM)
1353                 flags |= AMDGPU_PTE_VALID;
1354
1355         if (mem && (mem->mem_type == TTM_PL_TT ||
1356                     mem->mem_type == AMDGPU_PL_DOORBELL ||
1357                     mem->mem_type == AMDGPU_PL_PREEMPT)) {
1358                 flags |= AMDGPU_PTE_SYSTEM;
1359
1360                 if (ttm->caching == ttm_cached)
1361                         flags |= AMDGPU_PTE_SNOOPED;
1362         }
1363
1364         if (mem && mem->mem_type == TTM_PL_VRAM &&
1365                         mem->bus.caching == ttm_cached)
1366                 flags |= AMDGPU_PTE_SNOOPED;
1367
1368         return flags;
1369 }
1370
1371 /**
1372  * amdgpu_ttm_tt_pte_flags - Compute PTE flags for ttm_tt object
1373  *
1374  * @adev: amdgpu_device pointer
1375  * @ttm: The ttm_tt object to compute the flags for
1376  * @mem: The memory registry backing this ttm_tt object
1377  *
1378  * Figure out the flags to use for a VM PTE (Page Table Entry).
1379  */
1380 uint64_t amdgpu_ttm_tt_pte_flags(struct amdgpu_device *adev, struct ttm_tt *ttm,
1381                                  struct ttm_resource *mem)
1382 {
1383         uint64_t flags = amdgpu_ttm_tt_pde_flags(ttm, mem);
1384
1385         flags |= adev->gart.gart_pte_flags;
1386         flags |= AMDGPU_PTE_READABLE;
1387
1388         if (!amdgpu_ttm_tt_is_readonly(ttm))
1389                 flags |= AMDGPU_PTE_WRITEABLE;
1390
1391         return flags;
1392 }
1393
1394 /*
1395  * amdgpu_ttm_bo_eviction_valuable - Check to see if we can evict a buffer
1396  * object.
1397  *
1398  * Return true if eviction is sensible. Called by ttm_mem_evict_first() on
1399  * behalf of ttm_bo_mem_force_space() which tries to evict buffer objects until
1400  * it can find space for a new object and by ttm_bo_force_list_clean() which is
1401  * used to clean out a memory space.
1402  */
1403 static bool amdgpu_ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
1404                                             const struct ttm_place *place)
1405 {
1406         struct dma_resv_iter resv_cursor;
1407         struct dma_fence *f;
1408
1409         if (!amdgpu_bo_is_amdgpu_bo(bo))
1410                 return ttm_bo_eviction_valuable(bo, place);
1411
1412         /* Swapout? */
1413         if (bo->resource->mem_type == TTM_PL_SYSTEM)
1414                 return true;
1415
1416         if (bo->type == ttm_bo_type_kernel &&
1417             !amdgpu_vm_evictable(ttm_to_amdgpu_bo(bo)))
1418                 return false;
1419
1420         /* If bo is a KFD BO, check if the bo belongs to the current process.
1421          * If true, then return false as any KFD process needs all its BOs to
1422          * be resident to run successfully
1423          */
1424         dma_resv_for_each_fence(&resv_cursor, bo->base.resv,
1425                                 DMA_RESV_USAGE_BOOKKEEP, f) {
1426                 if (amdkfd_fence_check_mm(f, current->mm) &&
1427                     !(place->flags & TTM_PL_FLAG_CONTIGUOUS))
1428                         return false;
1429         }
1430
1431         /* Preemptible BOs don't own system resources managed by the
1432          * driver (pages, VRAM, GART space). They point to resources
1433          * owned by someone else (e.g. pageable memory in user mode
1434          * or a DMABuf). They are used in a preemptible context so we
1435          * can guarantee no deadlocks and good QoS in case of MMU
1436          * notifiers or DMABuf move notifiers from the resource owner.
1437          */
1438         if (bo->resource->mem_type == AMDGPU_PL_PREEMPT)
1439                 return false;
1440
1441         if (bo->resource->mem_type == TTM_PL_TT &&
1442             amdgpu_bo_encrypted(ttm_to_amdgpu_bo(bo)))
1443                 return false;
1444
1445         return ttm_bo_eviction_valuable(bo, place);
1446 }
1447
1448 static void amdgpu_ttm_vram_mm_access(struct amdgpu_device *adev, loff_t pos,
1449                                       void *buf, size_t size, bool write)
1450 {
1451         while (size) {
1452                 uint64_t aligned_pos = ALIGN_DOWN(pos, 4);
1453                 uint64_t bytes = 4 - (pos & 0x3);
1454                 uint32_t shift = (pos & 0x3) * 8;
1455                 uint32_t mask = 0xffffffff << shift;
1456                 uint32_t value = 0;
1457
1458                 if (size < bytes) {
1459                         mask &= 0xffffffff >> (bytes - size) * 8;
1460                         bytes = size;
1461                 }
1462
1463                 if (mask != 0xffffffff) {
1464                         amdgpu_device_mm_access(adev, aligned_pos, &value, 4, false);
1465                         if (write) {
1466                                 value &= ~mask;
1467                                 value |= (*(uint32_t *)buf << shift) & mask;
1468                                 amdgpu_device_mm_access(adev, aligned_pos, &value, 4, true);
1469                         } else {
1470                                 value = (value & mask) >> shift;
1471                                 memcpy(buf, &value, bytes);
1472                         }
1473                 } else {
1474                         amdgpu_device_mm_access(adev, aligned_pos, buf, 4, write);
1475                 }
1476
1477                 pos += bytes;
1478                 buf += bytes;
1479                 size -= bytes;
1480         }
1481 }
1482
1483 static int amdgpu_ttm_access_memory_sdma(struct ttm_buffer_object *bo,
1484                                         unsigned long offset, void *buf,
1485                                         int len, int write)
1486 {
1487         struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
1488         struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev);
1489         struct amdgpu_res_cursor src_mm;
1490         struct amdgpu_job *job;
1491         struct dma_fence *fence;
1492         uint64_t src_addr, dst_addr;
1493         unsigned int num_dw;
1494         int r, idx;
1495
1496         if (len != PAGE_SIZE)
1497                 return -EINVAL;
1498
1499         if (!adev->mman.sdma_access_ptr)
1500                 return -EACCES;
1501
1502         if (!drm_dev_enter(adev_to_drm(adev), &idx))
1503                 return -ENODEV;
1504
1505         if (write)
1506                 memcpy(adev->mman.sdma_access_ptr, buf, len);
1507
1508         num_dw = ALIGN(adev->mman.buffer_funcs->copy_num_dw, 8);
1509         r = amdgpu_job_alloc_with_ib(adev, &adev->mman.high_pr,
1510                                      AMDGPU_FENCE_OWNER_UNDEFINED,
1511                                      num_dw * 4, AMDGPU_IB_POOL_DELAYED,
1512                                      &job);
1513         if (r)
1514                 goto out;
1515
1516         amdgpu_res_first(abo->tbo.resource, offset, len, &src_mm);
1517         src_addr = amdgpu_ttm_domain_start(adev, bo->resource->mem_type) +
1518                 src_mm.start;
1519         dst_addr = amdgpu_bo_gpu_offset(adev->mman.sdma_access_bo);
1520         if (write)
1521                 swap(src_addr, dst_addr);
1522
1523         amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_addr, dst_addr,
1524                                 PAGE_SIZE, 0);
1525
1526         amdgpu_ring_pad_ib(adev->mman.buffer_funcs_ring, &job->ibs[0]);
1527         WARN_ON(job->ibs[0].length_dw > num_dw);
1528
1529         fence = amdgpu_job_submit(job);
1530
1531         if (!dma_fence_wait_timeout(fence, false, adev->sdma_timeout))
1532                 r = -ETIMEDOUT;
1533         dma_fence_put(fence);
1534
1535         if (!(r || write))
1536                 memcpy(buf, adev->mman.sdma_access_ptr, len);
1537 out:
1538         drm_dev_exit(idx);
1539         return r;
1540 }
1541
1542 /**
1543  * amdgpu_ttm_access_memory - Read or Write memory that backs a buffer object.
1544  *
1545  * @bo:  The buffer object to read/write
1546  * @offset:  Offset into buffer object
1547  * @buf:  Secondary buffer to write/read from
1548  * @len: Length in bytes of access
1549  * @write:  true if writing
1550  *
1551  * This is used to access VRAM that backs a buffer object via MMIO
1552  * access for debugging purposes.
1553  */
1554 static int amdgpu_ttm_access_memory(struct ttm_buffer_object *bo,
1555                                     unsigned long offset, void *buf, int len,
1556                                     int write)
1557 {
1558         struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
1559         struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev);
1560         struct amdgpu_res_cursor cursor;
1561         int ret = 0;
1562
1563         if (bo->resource->mem_type != TTM_PL_VRAM)
1564                 return -EIO;
1565
1566         if (amdgpu_device_has_timeouts_enabled(adev) &&
1567                         !amdgpu_ttm_access_memory_sdma(bo, offset, buf, len, write))
1568                 return len;
1569
1570         amdgpu_res_first(bo->resource, offset, len, &cursor);
1571         while (cursor.remaining) {
1572                 size_t count, size = cursor.size;
1573                 loff_t pos = cursor.start;
1574
1575                 count = amdgpu_device_aper_access(adev, pos, buf, size, write);
1576                 size -= count;
1577                 if (size) {
1578                         /* using MM to access rest vram and handle un-aligned address */
1579                         pos += count;
1580                         buf += count;
1581                         amdgpu_ttm_vram_mm_access(adev, pos, buf, size, write);
1582                 }
1583
1584                 ret += cursor.size;
1585                 buf += cursor.size;
1586                 amdgpu_res_next(&cursor, cursor.size);
1587         }
1588
1589         return ret;
1590 }
1591
1592 static void
1593 amdgpu_bo_delete_mem_notify(struct ttm_buffer_object *bo)
1594 {
1595         amdgpu_bo_move_notify(bo, false, NULL);
1596 }
1597
1598 static struct ttm_device_funcs amdgpu_bo_driver = {
1599         .ttm_tt_create = &amdgpu_ttm_tt_create,
1600         .ttm_tt_populate = &amdgpu_ttm_tt_populate,
1601         .ttm_tt_unpopulate = &amdgpu_ttm_tt_unpopulate,
1602         .ttm_tt_destroy = &amdgpu_ttm_backend_destroy,
1603         .eviction_valuable = amdgpu_ttm_bo_eviction_valuable,
1604         .evict_flags = &amdgpu_evict_flags,
1605         .move = &amdgpu_bo_move,
1606         .delete_mem_notify = &amdgpu_bo_delete_mem_notify,
1607         .release_notify = &amdgpu_bo_release_notify,
1608         .io_mem_reserve = &amdgpu_ttm_io_mem_reserve,
1609         .io_mem_pfn = amdgpu_ttm_io_mem_pfn,
1610         .access_memory = &amdgpu_ttm_access_memory,
1611 };
1612
1613 /*
1614  * Firmware Reservation functions
1615  */
1616 /**
1617  * amdgpu_ttm_fw_reserve_vram_fini - free fw reserved vram
1618  *
1619  * @adev: amdgpu_device pointer
1620  *
1621  * free fw reserved vram if it has been reserved.
1622  */
1623 static void amdgpu_ttm_fw_reserve_vram_fini(struct amdgpu_device *adev)
1624 {
1625         amdgpu_bo_free_kernel(&adev->mman.fw_vram_usage_reserved_bo,
1626                 NULL, &adev->mman.fw_vram_usage_va);
1627 }
1628
1629 /*
1630  * Driver Reservation functions
1631  */
1632 /**
1633  * amdgpu_ttm_drv_reserve_vram_fini - free drv reserved vram
1634  *
1635  * @adev: amdgpu_device pointer
1636  *
1637  * free drv reserved vram if it has been reserved.
1638  */
1639 static void amdgpu_ttm_drv_reserve_vram_fini(struct amdgpu_device *adev)
1640 {
1641         amdgpu_bo_free_kernel(&adev->mman.drv_vram_usage_reserved_bo,
1642                                                   NULL,
1643                                                   &adev->mman.drv_vram_usage_va);
1644 }
1645
1646 /**
1647  * amdgpu_ttm_fw_reserve_vram_init - create bo vram reservation from fw
1648  *
1649  * @adev: amdgpu_device pointer
1650  *
1651  * create bo vram reservation from fw.
1652  */
1653 static int amdgpu_ttm_fw_reserve_vram_init(struct amdgpu_device *adev)
1654 {
1655         uint64_t vram_size = adev->gmc.visible_vram_size;
1656
1657         adev->mman.fw_vram_usage_va = NULL;
1658         adev->mman.fw_vram_usage_reserved_bo = NULL;
1659
1660         if (adev->mman.fw_vram_usage_size == 0 ||
1661             adev->mman.fw_vram_usage_size > vram_size)
1662                 return 0;
1663
1664         return amdgpu_bo_create_kernel_at(adev,
1665                                           adev->mman.fw_vram_usage_start_offset,
1666                                           adev->mman.fw_vram_usage_size,
1667                                           &adev->mman.fw_vram_usage_reserved_bo,
1668                                           &adev->mman.fw_vram_usage_va);
1669 }
1670
1671 /**
1672  * amdgpu_ttm_drv_reserve_vram_init - create bo vram reservation from driver
1673  *
1674  * @adev: amdgpu_device pointer
1675  *
1676  * create bo vram reservation from drv.
1677  */
1678 static int amdgpu_ttm_drv_reserve_vram_init(struct amdgpu_device *adev)
1679 {
1680         u64 vram_size = adev->gmc.visible_vram_size;
1681
1682         adev->mman.drv_vram_usage_va = NULL;
1683         adev->mman.drv_vram_usage_reserved_bo = NULL;
1684
1685         if (adev->mman.drv_vram_usage_size == 0 ||
1686             adev->mman.drv_vram_usage_size > vram_size)
1687                 return 0;
1688
1689         return amdgpu_bo_create_kernel_at(adev,
1690                                           adev->mman.drv_vram_usage_start_offset,
1691                                           adev->mman.drv_vram_usage_size,
1692                                           &adev->mman.drv_vram_usage_reserved_bo,
1693                                           &adev->mman.drv_vram_usage_va);
1694 }
1695
1696 /*
1697  * Memoy training reservation functions
1698  */
1699
1700 /**
1701  * amdgpu_ttm_training_reserve_vram_fini - free memory training reserved vram
1702  *
1703  * @adev: amdgpu_device pointer
1704  *
1705  * free memory training reserved vram if it has been reserved.
1706  */
1707 static int amdgpu_ttm_training_reserve_vram_fini(struct amdgpu_device *adev)
1708 {
1709         struct psp_memory_training_context *ctx = &adev->psp.mem_train_ctx;
1710
1711         ctx->init = PSP_MEM_TRAIN_NOT_SUPPORT;
1712         amdgpu_bo_free_kernel(&ctx->c2p_bo, NULL, NULL);
1713         ctx->c2p_bo = NULL;
1714
1715         return 0;
1716 }
1717
1718 static void amdgpu_ttm_training_data_block_init(struct amdgpu_device *adev,
1719                                                 uint32_t reserve_size)
1720 {
1721         struct psp_memory_training_context *ctx = &adev->psp.mem_train_ctx;
1722
1723         memset(ctx, 0, sizeof(*ctx));
1724
1725         ctx->c2p_train_data_offset =
1726                 ALIGN((adev->gmc.mc_vram_size - reserve_size - SZ_1M), SZ_1M);
1727         ctx->p2c_train_data_offset =
1728                 (adev->gmc.mc_vram_size - GDDR6_MEM_TRAINING_OFFSET);
1729         ctx->train_data_size =
1730                 GDDR6_MEM_TRAINING_DATA_SIZE_IN_BYTES;
1731
1732         DRM_DEBUG("train_data_size:%llx,p2c_train_data_offset:%llx,c2p_train_data_offset:%llx.\n",
1733                         ctx->train_data_size,
1734                         ctx->p2c_train_data_offset,
1735                         ctx->c2p_train_data_offset);
1736 }
1737
1738 /*
1739  * reserve TMR memory at the top of VRAM which holds
1740  * IP Discovery data and is protected by PSP.
1741  */
1742 static int amdgpu_ttm_reserve_tmr(struct amdgpu_device *adev)
1743 {
1744         struct psp_memory_training_context *ctx = &adev->psp.mem_train_ctx;
1745         bool mem_train_support = false;
1746         uint32_t reserve_size = 0;
1747         int ret;
1748
1749         if (adev->bios && !amdgpu_sriov_vf(adev)) {
1750                 if (amdgpu_atomfirmware_mem_training_supported(adev))
1751                         mem_train_support = true;
1752                 else
1753                         DRM_DEBUG("memory training does not support!\n");
1754         }
1755
1756         /*
1757          * Query reserved tmr size through atom firmwareinfo for Sienna_Cichlid and onwards for all
1758          * the use cases (IP discovery/G6 memory training/profiling/diagnostic data.etc)
1759          *
1760          * Otherwise, fallback to legacy approach to check and reserve tmr block for ip
1761          * discovery data and G6 memory training data respectively
1762          */
1763         if (adev->bios)
1764                 reserve_size =
1765                         amdgpu_atomfirmware_get_fw_reserved_fb_size(adev);
1766
1767         if (!adev->bios &&
1768             (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 3) ||
1769              amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 4) ||
1770              amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 5, 0)))
1771                 reserve_size = max(reserve_size, (uint32_t)280 << 20);
1772         else if (!reserve_size)
1773                 reserve_size = DISCOVERY_TMR_OFFSET;
1774
1775         if (mem_train_support) {
1776                 /* reserve vram for mem train according to TMR location */
1777                 amdgpu_ttm_training_data_block_init(adev, reserve_size);
1778                 ret = amdgpu_bo_create_kernel_at(adev,
1779                                                  ctx->c2p_train_data_offset,
1780                                                  ctx->train_data_size,
1781                                                  &ctx->c2p_bo,
1782                                                  NULL);
1783                 if (ret) {
1784                         DRM_ERROR("alloc c2p_bo failed(%d)!\n", ret);
1785                         amdgpu_ttm_training_reserve_vram_fini(adev);
1786                         return ret;
1787                 }
1788                 ctx->init = PSP_MEM_TRAIN_RESERVE_SUCCESS;
1789         }
1790
1791         if (!adev->gmc.is_app_apu) {
1792                 ret = amdgpu_bo_create_kernel_at(
1793                         adev, adev->gmc.real_vram_size - reserve_size,
1794                         reserve_size, &adev->mman.fw_reserved_memory, NULL);
1795                 if (ret) {
1796                         DRM_ERROR("alloc tmr failed(%d)!\n", ret);
1797                         amdgpu_bo_free_kernel(&adev->mman.fw_reserved_memory,
1798                                               NULL, NULL);
1799                         return ret;
1800                 }
1801         } else {
1802                 DRM_DEBUG_DRIVER("backdoor fw loading path for PSP TMR, no reservation needed\n");
1803         }
1804
1805         return 0;
1806 }
1807
1808 static int amdgpu_ttm_pools_init(struct amdgpu_device *adev)
1809 {
1810         int i;
1811
1812         if (!adev->gmc.is_app_apu || !adev->gmc.num_mem_partitions)
1813                 return 0;
1814
1815         adev->mman.ttm_pools = kcalloc(adev->gmc.num_mem_partitions,
1816                                        sizeof(*adev->mman.ttm_pools),
1817                                        GFP_KERNEL);
1818         if (!adev->mman.ttm_pools)
1819                 return -ENOMEM;
1820
1821         for (i = 0; i < adev->gmc.num_mem_partitions; i++) {
1822                 ttm_pool_init(&adev->mman.ttm_pools[i], adev->dev,
1823                               adev->gmc.mem_partitions[i].numa.node,
1824                               false, false);
1825         }
1826         return 0;
1827 }
1828
1829 static void amdgpu_ttm_pools_fini(struct amdgpu_device *adev)
1830 {
1831         int i;
1832
1833         if (!adev->gmc.is_app_apu || !adev->mman.ttm_pools)
1834                 return;
1835
1836         for (i = 0; i < adev->gmc.num_mem_partitions; i++)
1837                 ttm_pool_fini(&adev->mman.ttm_pools[i]);
1838
1839         kfree(adev->mman.ttm_pools);
1840         adev->mman.ttm_pools = NULL;
1841 }
1842
1843 /*
1844  * amdgpu_ttm_init - Init the memory management (ttm) as well as various
1845  * gtt/vram related fields.
1846  *
1847  * This initializes all of the memory space pools that the TTM layer
1848  * will need such as the GTT space (system memory mapped to the device),
1849  * VRAM (on-board memory), and on-chip memories (GDS, GWS, OA) which
1850  * can be mapped per VMID.
1851  */
1852 int amdgpu_ttm_init(struct amdgpu_device *adev)
1853 {
1854         uint64_t gtt_size;
1855         int r;
1856
1857         mutex_init(&adev->mman.gtt_window_lock);
1858
1859         dma_set_max_seg_size(adev->dev, UINT_MAX);
1860         /* No others user of address space so set it to 0 */
1861         r = ttm_device_init(&adev->mman.bdev, &amdgpu_bo_driver, adev->dev,
1862                                adev_to_drm(adev)->anon_inode->i_mapping,
1863                                adev_to_drm(adev)->vma_offset_manager,
1864                                adev->need_swiotlb,
1865                                dma_addressing_limited(adev->dev));
1866         if (r) {
1867                 DRM_ERROR("failed initializing buffer object driver(%d).\n", r);
1868                 return r;
1869         }
1870
1871         r = amdgpu_ttm_pools_init(adev);
1872         if (r) {
1873                 DRM_ERROR("failed to init ttm pools(%d).\n", r);
1874                 return r;
1875         }
1876         adev->mman.initialized = true;
1877
1878         /* Initialize VRAM pool with all of VRAM divided into pages */
1879         r = amdgpu_vram_mgr_init(adev);
1880         if (r) {
1881                 DRM_ERROR("Failed initializing VRAM heap.\n");
1882                 return r;
1883         }
1884
1885         /* Change the size here instead of the init above so only lpfn is affected */
1886         amdgpu_ttm_set_buffer_funcs_status(adev, false);
1887 #ifdef CONFIG_64BIT
1888 #ifdef CONFIG_X86
1889         if (adev->gmc.xgmi.connected_to_cpu)
1890                 adev->mman.aper_base_kaddr = ioremap_cache(adev->gmc.aper_base,
1891                                 adev->gmc.visible_vram_size);
1892
1893         else if (adev->gmc.is_app_apu)
1894                 DRM_DEBUG_DRIVER(
1895                         "No need to ioremap when real vram size is 0\n");
1896         else
1897 #endif
1898                 adev->mman.aper_base_kaddr = ioremap_wc(adev->gmc.aper_base,
1899                                 adev->gmc.visible_vram_size);
1900 #endif
1901
1902         /*
1903          *The reserved vram for firmware must be pinned to the specified
1904          *place on the VRAM, so reserve it early.
1905          */
1906         r = amdgpu_ttm_fw_reserve_vram_init(adev);
1907         if (r)
1908                 return r;
1909
1910         /*
1911          *The reserved vram for driver must be pinned to the specified
1912          *place on the VRAM, so reserve it early.
1913          */
1914         r = amdgpu_ttm_drv_reserve_vram_init(adev);
1915         if (r)
1916                 return r;
1917
1918         /*
1919          * only NAVI10 and onwards ASIC support for IP discovery.
1920          * If IP discovery enabled, a block of memory should be
1921          * reserved for IP discovey.
1922          */
1923         if (adev->mman.discovery_bin) {
1924                 r = amdgpu_ttm_reserve_tmr(adev);
1925                 if (r)
1926                         return r;
1927         }
1928
1929         /* allocate memory as required for VGA
1930          * This is used for VGA emulation and pre-OS scanout buffers to
1931          * avoid display artifacts while transitioning between pre-OS
1932          * and driver.
1933          */
1934         if (!adev->gmc.is_app_apu) {
1935                 r = amdgpu_bo_create_kernel_at(adev, 0,
1936                                                adev->mman.stolen_vga_size,
1937                                                &adev->mman.stolen_vga_memory,
1938                                                NULL);
1939                 if (r)
1940                         return r;
1941
1942                 r = amdgpu_bo_create_kernel_at(adev, adev->mman.stolen_vga_size,
1943                                                adev->mman.stolen_extended_size,
1944                                                &adev->mman.stolen_extended_memory,
1945                                                NULL);
1946
1947                 if (r)
1948                         return r;
1949
1950                 r = amdgpu_bo_create_kernel_at(adev,
1951                                                adev->mman.stolen_reserved_offset,
1952                                                adev->mman.stolen_reserved_size,
1953                                                &adev->mman.stolen_reserved_memory,
1954                                                NULL);
1955                 if (r)
1956                         return r;
1957         } else {
1958                 DRM_DEBUG_DRIVER("Skipped stolen memory reservation\n");
1959         }
1960
1961         DRM_INFO("amdgpu: %uM of VRAM memory ready\n",
1962                  (unsigned int)(adev->gmc.real_vram_size / (1024 * 1024)));
1963
1964         /* Compute GTT size, either based on TTM limit
1965          * or whatever the user passed on module init.
1966          */
1967         gtt_size = ttm_tt_pages_limit() << PAGE_SHIFT;
1968         if (amdgpu_gtt_size != -1) {
1969                 uint64_t configured_size = (uint64_t)amdgpu_gtt_size << 20;
1970
1971                 drm_warn(&adev->ddev,
1972                         "Configuring gttsize via module parameter is deprecated, please use ttm.pages_limit\n");
1973                 if (gtt_size != configured_size)
1974                         drm_warn(&adev->ddev,
1975                                 "GTT size has been set as %llu but TTM size has been set as %llu, this is unusual\n",
1976                                 configured_size, gtt_size);
1977
1978                 gtt_size = configured_size;
1979         }
1980
1981         /* Initialize GTT memory pool */
1982         r = amdgpu_gtt_mgr_init(adev, gtt_size);
1983         if (r) {
1984                 DRM_ERROR("Failed initializing GTT heap.\n");
1985                 return r;
1986         }
1987         DRM_INFO("amdgpu: %uM of GTT memory ready.\n",
1988                  (unsigned int)(gtt_size / (1024 * 1024)));
1989
1990         if (adev->flags & AMD_IS_APU) {
1991                 if (adev->gmc.real_vram_size < gtt_size)
1992                         adev->apu_prefer_gtt = true;
1993         }
1994
1995         /* Initialize doorbell pool on PCI BAR */
1996         r = amdgpu_ttm_init_on_chip(adev, AMDGPU_PL_DOORBELL, adev->doorbell.size / PAGE_SIZE);
1997         if (r) {
1998                 DRM_ERROR("Failed initializing doorbell heap.\n");
1999                 return r;
2000         }
2001
2002         /* Create a boorbell page for kernel usages */
2003         r = amdgpu_doorbell_create_kernel_doorbells(adev);
2004         if (r) {
2005                 DRM_ERROR("Failed to initialize kernel doorbells.\n");
2006                 return r;
2007         }
2008
2009         /* Initialize preemptible memory pool */
2010         r = amdgpu_preempt_mgr_init(adev);
2011         if (r) {
2012                 DRM_ERROR("Failed initializing PREEMPT heap.\n");
2013                 return r;
2014         }
2015
2016         /* Initialize various on-chip memory pools */
2017         r = amdgpu_ttm_init_on_chip(adev, AMDGPU_PL_GDS, adev->gds.gds_size);
2018         if (r) {
2019                 DRM_ERROR("Failed initializing GDS heap.\n");
2020                 return r;
2021         }
2022
2023         r = amdgpu_ttm_init_on_chip(adev, AMDGPU_PL_GWS, adev->gds.gws_size);
2024         if (r) {
2025                 DRM_ERROR("Failed initializing gws heap.\n");
2026                 return r;
2027         }
2028
2029         r = amdgpu_ttm_init_on_chip(adev, AMDGPU_PL_OA, adev->gds.oa_size);
2030         if (r) {
2031                 DRM_ERROR("Failed initializing oa heap.\n");
2032                 return r;
2033         }
2034         if (amdgpu_bo_create_kernel(adev, PAGE_SIZE, PAGE_SIZE,
2035                                 AMDGPU_GEM_DOMAIN_GTT,
2036                                 &adev->mman.sdma_access_bo, NULL,
2037                                 &adev->mman.sdma_access_ptr))
2038                 DRM_WARN("Debug VRAM access will use slowpath MM access\n");
2039
2040         return 0;
2041 }
2042
2043 /*
2044  * amdgpu_ttm_fini - De-initialize the TTM memory pools
2045  */
2046 void amdgpu_ttm_fini(struct amdgpu_device *adev)
2047 {
2048         int idx;
2049
2050         if (!adev->mman.initialized)
2051                 return;
2052
2053         amdgpu_ttm_pools_fini(adev);
2054
2055         amdgpu_ttm_training_reserve_vram_fini(adev);
2056         /* return the stolen vga memory back to VRAM */
2057         if (!adev->gmc.is_app_apu) {
2058                 amdgpu_bo_free_kernel(&adev->mman.stolen_vga_memory, NULL, NULL);
2059                 amdgpu_bo_free_kernel(&adev->mman.stolen_extended_memory, NULL, NULL);
2060                 /* return the FW reserved memory back to VRAM */
2061                 amdgpu_bo_free_kernel(&adev->mman.fw_reserved_memory, NULL,
2062                                       NULL);
2063                 if (adev->mman.stolen_reserved_size)
2064                         amdgpu_bo_free_kernel(&adev->mman.stolen_reserved_memory,
2065                                               NULL, NULL);
2066         }
2067         amdgpu_bo_free_kernel(&adev->mman.sdma_access_bo, NULL,
2068                                         &adev->mman.sdma_access_ptr);
2069         amdgpu_ttm_fw_reserve_vram_fini(adev);
2070         amdgpu_ttm_drv_reserve_vram_fini(adev);
2071
2072         if (drm_dev_enter(adev_to_drm(adev), &idx)) {
2073
2074                 if (adev->mman.aper_base_kaddr)
2075                         iounmap(adev->mman.aper_base_kaddr);
2076                 adev->mman.aper_base_kaddr = NULL;
2077
2078                 drm_dev_exit(idx);
2079         }
2080
2081         amdgpu_vram_mgr_fini(adev);
2082         amdgpu_gtt_mgr_fini(adev);
2083         amdgpu_preempt_mgr_fini(adev);
2084         amdgpu_doorbell_fini(adev);
2085
2086         ttm_range_man_fini(&adev->mman.bdev, AMDGPU_PL_GDS);
2087         ttm_range_man_fini(&adev->mman.bdev, AMDGPU_PL_GWS);
2088         ttm_range_man_fini(&adev->mman.bdev, AMDGPU_PL_OA);
2089         ttm_range_man_fini(&adev->mman.bdev, AMDGPU_PL_DOORBELL);
2090         ttm_device_fini(&adev->mman.bdev);
2091         adev->mman.initialized = false;
2092         DRM_INFO("amdgpu: ttm finalized\n");
2093 }
2094
2095 /**
2096  * amdgpu_ttm_set_buffer_funcs_status - enable/disable use of buffer functions
2097  *
2098  * @adev: amdgpu_device pointer
2099  * @enable: true when we can use buffer functions.
2100  *
2101  * Enable/disable use of buffer functions during suspend/resume. This should
2102  * only be called at bootup or when userspace isn't running.
2103  */
2104 void amdgpu_ttm_set_buffer_funcs_status(struct amdgpu_device *adev, bool enable)
2105 {
2106         struct ttm_resource_manager *man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM);
2107         uint64_t size;
2108         int r;
2109
2110         if (!adev->mman.initialized || amdgpu_in_reset(adev) ||
2111             adev->mman.buffer_funcs_enabled == enable || adev->gmc.is_app_apu)
2112                 return;
2113
2114         if (enable) {
2115                 struct amdgpu_ring *ring;
2116                 struct drm_gpu_scheduler *sched;
2117
2118                 ring = adev->mman.buffer_funcs_ring;
2119                 sched = &ring->sched;
2120                 r = drm_sched_entity_init(&adev->mman.high_pr,
2121                                           DRM_SCHED_PRIORITY_KERNEL, &sched,
2122                                           1, NULL);
2123                 if (r) {
2124                         DRM_ERROR("Failed setting up TTM BO move entity (%d)\n",
2125                                   r);
2126                         return;
2127                 }
2128
2129                 r = drm_sched_entity_init(&adev->mman.low_pr,
2130                                           DRM_SCHED_PRIORITY_NORMAL, &sched,
2131                                           1, NULL);
2132                 if (r) {
2133                         DRM_ERROR("Failed setting up TTM BO move entity (%d)\n",
2134                                   r);
2135                         goto error_free_entity;
2136                 }
2137         } else {
2138                 drm_sched_entity_destroy(&adev->mman.high_pr);
2139                 drm_sched_entity_destroy(&adev->mman.low_pr);
2140                 dma_fence_put(man->move);
2141                 man->move = NULL;
2142         }
2143
2144         /* this just adjusts TTM size idea, which sets lpfn to the correct value */
2145         if (enable)
2146                 size = adev->gmc.real_vram_size;
2147         else
2148                 size = adev->gmc.visible_vram_size;
2149         man->size = size;
2150         adev->mman.buffer_funcs_enabled = enable;
2151
2152         return;
2153
2154 error_free_entity:
2155         drm_sched_entity_destroy(&adev->mman.high_pr);
2156 }
2157
2158 static int amdgpu_ttm_prepare_job(struct amdgpu_device *adev,
2159                                   bool direct_submit,
2160                                   unsigned int num_dw,
2161                                   struct dma_resv *resv,
2162                                   bool vm_needs_flush,
2163                                   struct amdgpu_job **job,
2164                                   bool delayed)
2165 {
2166         enum amdgpu_ib_pool_type pool = direct_submit ?
2167                 AMDGPU_IB_POOL_DIRECT :
2168                 AMDGPU_IB_POOL_DELAYED;
2169         int r;
2170         struct drm_sched_entity *entity = delayed ? &adev->mman.low_pr :
2171                                                     &adev->mman.high_pr;
2172         r = amdgpu_job_alloc_with_ib(adev, entity,
2173                                      AMDGPU_FENCE_OWNER_UNDEFINED,
2174                                      num_dw * 4, pool, job);
2175         if (r)
2176                 return r;
2177
2178         if (vm_needs_flush) {
2179                 (*job)->vm_pd_addr = amdgpu_gmc_pd_addr(adev->gmc.pdb0_bo ?
2180                                                         adev->gmc.pdb0_bo :
2181                                                         adev->gart.bo);
2182                 (*job)->vm_needs_flush = true;
2183         }
2184         if (!resv)
2185                 return 0;
2186
2187         return drm_sched_job_add_resv_dependencies(&(*job)->base, resv,
2188                                                    DMA_RESV_USAGE_BOOKKEEP);
2189 }
2190
2191 int amdgpu_copy_buffer(struct amdgpu_ring *ring, uint64_t src_offset,
2192                        uint64_t dst_offset, uint32_t byte_count,
2193                        struct dma_resv *resv,
2194                        struct dma_fence **fence, bool direct_submit,
2195                        bool vm_needs_flush, uint32_t copy_flags)
2196 {
2197         struct amdgpu_device *adev = ring->adev;
2198         unsigned int num_loops, num_dw;
2199         struct amdgpu_job *job;
2200         uint32_t max_bytes;
2201         unsigned int i;
2202         int r;
2203
2204         if (!direct_submit && !ring->sched.ready) {
2205                 DRM_ERROR("Trying to move memory with ring turned off.\n");
2206                 return -EINVAL;
2207         }
2208
2209         max_bytes = adev->mman.buffer_funcs->copy_max_bytes;
2210         num_loops = DIV_ROUND_UP(byte_count, max_bytes);
2211         num_dw = ALIGN(num_loops * adev->mman.buffer_funcs->copy_num_dw, 8);
2212         r = amdgpu_ttm_prepare_job(adev, direct_submit, num_dw,
2213                                    resv, vm_needs_flush, &job, false);
2214         if (r)
2215                 return r;
2216
2217         for (i = 0; i < num_loops; i++) {
2218                 uint32_t cur_size_in_bytes = min(byte_count, max_bytes);
2219
2220                 amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_offset,
2221                                         dst_offset, cur_size_in_bytes, copy_flags);
2222                 src_offset += cur_size_in_bytes;
2223                 dst_offset += cur_size_in_bytes;
2224                 byte_count -= cur_size_in_bytes;
2225         }
2226
2227         amdgpu_ring_pad_ib(ring, &job->ibs[0]);
2228         WARN_ON(job->ibs[0].length_dw > num_dw);
2229         if (direct_submit)
2230                 r = amdgpu_job_submit_direct(job, ring, fence);
2231         else
2232                 *fence = amdgpu_job_submit(job);
2233         if (r)
2234                 goto error_free;
2235
2236         return r;
2237
2238 error_free:
2239         amdgpu_job_free(job);
2240         DRM_ERROR("Error scheduling IBs (%d)\n", r);
2241         return r;
2242 }
2243
2244 static int amdgpu_ttm_fill_mem(struct amdgpu_ring *ring, uint32_t src_data,
2245                                uint64_t dst_addr, uint32_t byte_count,
2246                                struct dma_resv *resv,
2247                                struct dma_fence **fence,
2248                                bool vm_needs_flush, bool delayed)
2249 {
2250         struct amdgpu_device *adev = ring->adev;
2251         unsigned int num_loops, num_dw;
2252         struct amdgpu_job *job;
2253         uint32_t max_bytes;
2254         unsigned int i;
2255         int r;
2256
2257         max_bytes = adev->mman.buffer_funcs->fill_max_bytes;
2258         num_loops = DIV_ROUND_UP_ULL(byte_count, max_bytes);
2259         num_dw = ALIGN(num_loops * adev->mman.buffer_funcs->fill_num_dw, 8);
2260         r = amdgpu_ttm_prepare_job(adev, false, num_dw, resv, vm_needs_flush,
2261                                    &job, delayed);
2262         if (r)
2263                 return r;
2264
2265         for (i = 0; i < num_loops; i++) {
2266                 uint32_t cur_size = min(byte_count, max_bytes);
2267
2268                 amdgpu_emit_fill_buffer(adev, &job->ibs[0], src_data, dst_addr,
2269                                         cur_size);
2270
2271                 dst_addr += cur_size;
2272                 byte_count -= cur_size;
2273         }
2274
2275         amdgpu_ring_pad_ib(ring, &job->ibs[0]);
2276         WARN_ON(job->ibs[0].length_dw > num_dw);
2277         *fence = amdgpu_job_submit(job);
2278         return 0;
2279 }
2280
2281 /**
2282  * amdgpu_ttm_clear_buffer - clear memory buffers
2283  * @bo: amdgpu buffer object
2284  * @resv: reservation object
2285  * @fence: dma_fence associated with the operation
2286  *
2287  * Clear the memory buffer resource.
2288  *
2289  * Returns:
2290  * 0 for success or a negative error code on failure.
2291  */
2292 int amdgpu_ttm_clear_buffer(struct amdgpu_bo *bo,
2293                             struct dma_resv *resv,
2294                             struct dma_fence **fence)
2295 {
2296         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
2297         struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
2298         struct amdgpu_res_cursor cursor;
2299         u64 addr;
2300         int r = 0;
2301
2302         if (!adev->mman.buffer_funcs_enabled)
2303                 return -EINVAL;
2304
2305         if (!fence)
2306                 return -EINVAL;
2307
2308         *fence = dma_fence_get_stub();
2309
2310         amdgpu_res_first(bo->tbo.resource, 0, amdgpu_bo_size(bo), &cursor);
2311
2312         mutex_lock(&adev->mman.gtt_window_lock);
2313         while (cursor.remaining) {
2314                 struct dma_fence *next = NULL;
2315                 u64 size;
2316
2317                 if (amdgpu_res_cleared(&cursor)) {
2318                         amdgpu_res_next(&cursor, cursor.size);
2319                         continue;
2320                 }
2321
2322                 /* Never clear more than 256MiB at once to avoid timeouts */
2323                 size = min(cursor.size, 256ULL << 20);
2324
2325                 r = amdgpu_ttm_map_buffer(&bo->tbo, bo->tbo.resource, &cursor,
2326                                           1, ring, false, &size, &addr);
2327                 if (r)
2328                         goto err;
2329
2330                 r = amdgpu_ttm_fill_mem(ring, 0, addr, size, resv,
2331                                         &next, true, true);
2332                 if (r)
2333                         goto err;
2334
2335                 dma_fence_put(*fence);
2336                 *fence = next;
2337
2338                 amdgpu_res_next(&cursor, size);
2339         }
2340 err:
2341         mutex_unlock(&adev->mman.gtt_window_lock);
2342
2343         return r;
2344 }
2345
2346 int amdgpu_fill_buffer(struct amdgpu_bo *bo,
2347                         uint32_t src_data,
2348                         struct dma_resv *resv,
2349                         struct dma_fence **f,
2350                         bool delayed)
2351 {
2352         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
2353         struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
2354         struct dma_fence *fence = NULL;
2355         struct amdgpu_res_cursor dst;
2356         int r;
2357
2358         if (!adev->mman.buffer_funcs_enabled) {
2359                 DRM_ERROR("Trying to clear memory with ring turned off.\n");
2360                 return -EINVAL;
2361         }
2362
2363         amdgpu_res_first(bo->tbo.resource, 0, amdgpu_bo_size(bo), &dst);
2364
2365         mutex_lock(&adev->mman.gtt_window_lock);
2366         while (dst.remaining) {
2367                 struct dma_fence *next;
2368                 uint64_t cur_size, to;
2369
2370                 /* Never fill more than 256MiB at once to avoid timeouts */
2371                 cur_size = min(dst.size, 256ULL << 20);
2372
2373                 r = amdgpu_ttm_map_buffer(&bo->tbo, bo->tbo.resource, &dst,
2374                                           1, ring, false, &cur_size, &to);
2375                 if (r)
2376                         goto error;
2377
2378                 r = amdgpu_ttm_fill_mem(ring, src_data, to, cur_size, resv,
2379                                         &next, true, delayed);
2380                 if (r)
2381                         goto error;
2382
2383                 dma_fence_put(fence);
2384                 fence = next;
2385
2386                 amdgpu_res_next(&dst, cur_size);
2387         }
2388 error:
2389         mutex_unlock(&adev->mman.gtt_window_lock);
2390         if (f)
2391                 *f = dma_fence_get(fence);
2392         dma_fence_put(fence);
2393         return r;
2394 }
2395
2396 /**
2397  * amdgpu_ttm_evict_resources - evict memory buffers
2398  * @adev: amdgpu device object
2399  * @mem_type: evicted BO's memory type
2400  *
2401  * Evicts all @mem_type buffers on the lru list of the memory type.
2402  *
2403  * Returns:
2404  * 0 for success or a negative error code on failure.
2405  */
2406 int amdgpu_ttm_evict_resources(struct amdgpu_device *adev, int mem_type)
2407 {
2408         struct ttm_resource_manager *man;
2409
2410         switch (mem_type) {
2411         case TTM_PL_VRAM:
2412         case TTM_PL_TT:
2413         case AMDGPU_PL_GWS:
2414         case AMDGPU_PL_GDS:
2415         case AMDGPU_PL_OA:
2416                 man = ttm_manager_type(&adev->mman.bdev, mem_type);
2417                 break;
2418         default:
2419                 DRM_ERROR("Trying to evict invalid memory type\n");
2420                 return -EINVAL;
2421         }
2422
2423         return ttm_resource_manager_evict_all(&adev->mman.bdev, man);
2424 }
2425
2426 #if defined(CONFIG_DEBUG_FS)
2427
2428 static int amdgpu_ttm_page_pool_show(struct seq_file *m, void *unused)
2429 {
2430         struct amdgpu_device *adev = m->private;
2431
2432         return ttm_pool_debugfs(&adev->mman.bdev.pool, m);
2433 }
2434
2435 DEFINE_SHOW_ATTRIBUTE(amdgpu_ttm_page_pool);
2436
2437 /*
2438  * amdgpu_ttm_vram_read - Linear read access to VRAM
2439  *
2440  * Accesses VRAM via MMIO for debugging purposes.
2441  */
2442 static ssize_t amdgpu_ttm_vram_read(struct file *f, char __user *buf,
2443                                     size_t size, loff_t *pos)
2444 {
2445         struct amdgpu_device *adev = file_inode(f)->i_private;
2446         ssize_t result = 0;
2447
2448         if (size & 0x3 || *pos & 0x3)
2449                 return -EINVAL;
2450
2451         if (*pos >= adev->gmc.mc_vram_size)
2452                 return -ENXIO;
2453
2454         size = min(size, (size_t)(adev->gmc.mc_vram_size - *pos));
2455         while (size) {
2456                 size_t bytes = min(size, AMDGPU_TTM_VRAM_MAX_DW_READ * 4);
2457                 uint32_t value[AMDGPU_TTM_VRAM_MAX_DW_READ];
2458
2459                 amdgpu_device_vram_access(adev, *pos, value, bytes, false);
2460                 if (copy_to_user(buf, value, bytes))
2461                         return -EFAULT;
2462
2463                 result += bytes;
2464                 buf += bytes;
2465                 *pos += bytes;
2466                 size -= bytes;
2467         }
2468
2469         return result;
2470 }
2471
2472 /*
2473  * amdgpu_ttm_vram_write - Linear write access to VRAM
2474  *
2475  * Accesses VRAM via MMIO for debugging purposes.
2476  */
2477 static ssize_t amdgpu_ttm_vram_write(struct file *f, const char __user *buf,
2478                                     size_t size, loff_t *pos)
2479 {
2480         struct amdgpu_device *adev = file_inode(f)->i_private;
2481         ssize_t result = 0;
2482         int r;
2483
2484         if (size & 0x3 || *pos & 0x3)
2485                 return -EINVAL;
2486
2487         if (*pos >= adev->gmc.mc_vram_size)
2488                 return -ENXIO;
2489
2490         while (size) {
2491                 uint32_t value;
2492
2493                 if (*pos >= adev->gmc.mc_vram_size)
2494                         return result;
2495
2496                 r = get_user(value, (uint32_t *)buf);
2497                 if (r)
2498                         return r;
2499
2500                 amdgpu_device_mm_access(adev, *pos, &value, 4, true);
2501
2502                 result += 4;
2503                 buf += 4;
2504                 *pos += 4;
2505                 size -= 4;
2506         }
2507
2508         return result;
2509 }
2510
2511 static const struct file_operations amdgpu_ttm_vram_fops = {
2512         .owner = THIS_MODULE,
2513         .read = amdgpu_ttm_vram_read,
2514         .write = amdgpu_ttm_vram_write,
2515         .llseek = default_llseek,
2516 };
2517
2518 /*
2519  * amdgpu_iomem_read - Virtual read access to GPU mapped memory
2520  *
2521  * This function is used to read memory that has been mapped to the
2522  * GPU and the known addresses are not physical addresses but instead
2523  * bus addresses (e.g., what you'd put in an IB or ring buffer).
2524  */
2525 static ssize_t amdgpu_iomem_read(struct file *f, char __user *buf,
2526                                  size_t size, loff_t *pos)
2527 {
2528         struct amdgpu_device *adev = file_inode(f)->i_private;
2529         struct iommu_domain *dom;
2530         ssize_t result = 0;
2531         int r;
2532
2533         /* retrieve the IOMMU domain if any for this device */
2534         dom = iommu_get_domain_for_dev(adev->dev);
2535
2536         while (size) {
2537                 phys_addr_t addr = *pos & PAGE_MASK;
2538                 loff_t off = *pos & ~PAGE_MASK;
2539                 size_t bytes = PAGE_SIZE - off;
2540                 unsigned long pfn;
2541                 struct page *p;
2542                 void *ptr;
2543
2544                 bytes = min(bytes, size);
2545
2546                 /* Translate the bus address to a physical address.  If
2547                  * the domain is NULL it means there is no IOMMU active
2548                  * and the address translation is the identity
2549                  */
2550                 addr = dom ? iommu_iova_to_phys(dom, addr) : addr;
2551
2552                 pfn = addr >> PAGE_SHIFT;
2553                 if (!pfn_valid(pfn))
2554                         return -EPERM;
2555
2556                 p = pfn_to_page(pfn);
2557                 if (p->mapping != adev->mman.bdev.dev_mapping)
2558                         return -EPERM;
2559
2560                 ptr = kmap_local_page(p);
2561                 r = copy_to_user(buf, ptr + off, bytes);
2562                 kunmap_local(ptr);
2563                 if (r)
2564                         return -EFAULT;
2565
2566                 size -= bytes;
2567                 *pos += bytes;
2568                 result += bytes;
2569         }
2570
2571         return result;
2572 }
2573
2574 /*
2575  * amdgpu_iomem_write - Virtual write access to GPU mapped memory
2576  *
2577  * This function is used to write memory that has been mapped to the
2578  * GPU and the known addresses are not physical addresses but instead
2579  * bus addresses (e.g., what you'd put in an IB or ring buffer).
2580  */
2581 static ssize_t amdgpu_iomem_write(struct file *f, const char __user *buf,
2582                                  size_t size, loff_t *pos)
2583 {
2584         struct amdgpu_device *adev = file_inode(f)->i_private;
2585         struct iommu_domain *dom;
2586         ssize_t result = 0;
2587         int r;
2588
2589         dom = iommu_get_domain_for_dev(adev->dev);
2590
2591         while (size) {
2592                 phys_addr_t addr = *pos & PAGE_MASK;
2593                 loff_t off = *pos & ~PAGE_MASK;
2594                 size_t bytes = PAGE_SIZE - off;
2595                 unsigned long pfn;
2596                 struct page *p;
2597                 void *ptr;
2598
2599                 bytes = min(bytes, size);
2600
2601                 addr = dom ? iommu_iova_to_phys(dom, addr) : addr;
2602
2603                 pfn = addr >> PAGE_SHIFT;
2604                 if (!pfn_valid(pfn))
2605                         return -EPERM;
2606
2607                 p = pfn_to_page(pfn);
2608                 if (p->mapping != adev->mman.bdev.dev_mapping)
2609                         return -EPERM;
2610
2611                 ptr = kmap_local_page(p);
2612                 r = copy_from_user(ptr + off, buf, bytes);
2613                 kunmap_local(ptr);
2614                 if (r)
2615                         return -EFAULT;
2616
2617                 size -= bytes;
2618                 *pos += bytes;
2619                 result += bytes;
2620         }
2621
2622         return result;
2623 }
2624
2625 static const struct file_operations amdgpu_ttm_iomem_fops = {
2626         .owner = THIS_MODULE,
2627         .read = amdgpu_iomem_read,
2628         .write = amdgpu_iomem_write,
2629         .llseek = default_llseek
2630 };
2631
2632 #endif
2633
2634 void amdgpu_ttm_debugfs_init(struct amdgpu_device *adev)
2635 {
2636 #if defined(CONFIG_DEBUG_FS)
2637         struct drm_minor *minor = adev_to_drm(adev)->primary;
2638         struct dentry *root = minor->debugfs_root;
2639
2640         debugfs_create_file_size("amdgpu_vram", 0444, root, adev,
2641                                  &amdgpu_ttm_vram_fops, adev->gmc.mc_vram_size);
2642         debugfs_create_file("amdgpu_iomem", 0444, root, adev,
2643                             &amdgpu_ttm_iomem_fops);
2644         debugfs_create_file("ttm_page_pool", 0444, root, adev,
2645                             &amdgpu_ttm_page_pool_fops);
2646         ttm_resource_manager_create_debugfs(ttm_manager_type(&adev->mman.bdev,
2647                                                              TTM_PL_VRAM),
2648                                             root, "amdgpu_vram_mm");
2649         ttm_resource_manager_create_debugfs(ttm_manager_type(&adev->mman.bdev,
2650                                                              TTM_PL_TT),
2651                                             root, "amdgpu_gtt_mm");
2652         ttm_resource_manager_create_debugfs(ttm_manager_type(&adev->mman.bdev,
2653                                                              AMDGPU_PL_GDS),
2654                                             root, "amdgpu_gds_mm");
2655         ttm_resource_manager_create_debugfs(ttm_manager_type(&adev->mman.bdev,
2656                                                              AMDGPU_PL_GWS),
2657                                             root, "amdgpu_gws_mm");
2658         ttm_resource_manager_create_debugfs(ttm_manager_type(&adev->mman.bdev,
2659                                                              AMDGPU_PL_OA),
2660                                             root, "amdgpu_oa_mm");
2661
2662 #endif
2663 }