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_bo_in_cpu_visible_vram(abo)) {
 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, false);
 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
 299         if (!adev->mman.buffer_funcs_enabled) {
 300                 DRM_ERROR("Trying to move memory with ring turned off.\n");
 301                 return -EINVAL;
 302         }
 303
 304         amdgpu_res_first(src->mem, src->offset, size, &src_mm);
 305         amdgpu_res_first(dst->mem, dst->offset, size, &dst_mm);
 306
 307         mutex_lock(&adev->mman.gtt_window_lock);
 308         while (src_mm.remaining) {
 309                 uint64_t from, to, cur_size;
 310                 struct dma_fence *next;
 311
 312                 /* Never copy more than 256MiB at once to avoid a timeout */
 313                 cur_size = min3(src_mm.size, dst_mm.size, 256ULL << 20);
 314
 315                 /* Map src to window 0 and dst to window 1. */
 316                 r = amdgpu_ttm_map_buffer(src->bo, src->mem, &src_mm,
 317                                           0, ring, tmz, &cur_size, &from);
 318                 if (r)
 319                         goto error;
 320
 321                 r = amdgpu_ttm_map_buffer(dst->bo, dst->mem, &dst_mm,
 322                                           1, ring, tmz, &cur_size, &to);
 323                 if (r)
 324                         goto error;
 325
 326                 r = amdgpu_copy_buffer(ring, from, to, cur_size,
 327                                        resv, &next, false, true, tmz);
 328                 if (r)
 329                         goto error;
 330
 331                 dma_fence_put(fence);
 332                 fence = next;
 333
 334                 amdgpu_res_next(&src_mm, cur_size);
 335                 amdgpu_res_next(&dst_mm, cur_size);
 336         }
 337 error:
 338         mutex_unlock(&adev->mman.gtt_window_lock);
 339         if (f)
 340                 *f = dma_fence_get(fence);
 341         dma_fence_put(fence);
 342         return r;
 343 }
 344
 345 /*
 346  * amdgpu_move_blit - Copy an entire buffer to another buffer
 347  *
 348  * This is a helper called by amdgpu_bo_move() and amdgpu_move_vram_ram() to
 349  * help move buffers to and from VRAM.
 350  */
 351 static int amdgpu_move_blit(struct ttm_buffer_object *bo,
 352                             bool evict,
 353                             struct ttm_resource *new_mem,
 354                             struct ttm_resource *old_mem)
 355 {
 356         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
 357         struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
 358         struct amdgpu_copy_mem src, dst;
 359         struct dma_fence *fence = NULL;
 360         int r;
 361
 362         src.bo = bo;
 363         dst.bo = bo;
 364         src.mem = old_mem;
 365         dst.mem = new_mem;
 366         src.offset = 0;
 367         dst.offset = 0;
 368
 369         r = amdgpu_ttm_copy_mem_to_mem(adev, &src, &dst,
 370                                        new_mem->size,
 371                                        amdgpu_bo_encrypted(abo),
 372                                        bo->base.resv, &fence);
 373         if (r)
 374                 goto error;
 375
 376         /* clear the space being freed */
 377         if (old_mem->mem_type == TTM_PL_VRAM &&
 378             (abo->flags & AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE)) {
 379                 struct dma_fence *wipe_fence = NULL;
 380
 381                 r = amdgpu_fill_buffer(abo, AMDGPU_POISON, NULL, &wipe_fence,
 382                                         false);
 383                 if (r) {
 384                         goto error;
 385                 } else if (wipe_fence) {
 386                         dma_fence_put(fence);
 387                         fence = wipe_fence;
 388                 }
 389         }
 390
 391         /* Always block for VM page tables before committing the new location */
 392         if (bo->type == ttm_bo_type_kernel)
 393                 r = ttm_bo_move_accel_cleanup(bo, fence, true, false, new_mem);
 394         else
 395                 r = ttm_bo_move_accel_cleanup(bo, fence, evict, true, new_mem);
 396         dma_fence_put(fence);
 397         return r;
 398
 399 error:
 400         if (fence)
 401                 dma_fence_wait(fence, false);
 402         dma_fence_put(fence);
 403         return r;
 404 }
 405
 406 /*
 407  * amdgpu_mem_visible - Check that memory can be accessed by ttm_bo_move_memcpy
 408  *
 409  * Called by amdgpu_bo_move()
 410  */
 411 static bool amdgpu_mem_visible(struct amdgpu_device *adev,
 412                                struct ttm_resource *mem)
 413 {
 414         u64 mem_size = (u64)mem->size;
 415         struct amdgpu_res_cursor cursor;
 416         u64 end;
 417
 418         if (mem->mem_type == TTM_PL_SYSTEM ||
 419             mem->mem_type == TTM_PL_TT)
 420                 return true;
 421         if (mem->mem_type != TTM_PL_VRAM)
 422                 return false;
 423
 424         amdgpu_res_first(mem, 0, mem_size, &cursor);
 425         end = cursor.start + cursor.size;
 426         while (cursor.remaining) {
 427                 amdgpu_res_next(&cursor, cursor.size);
 428
 429                 if (!cursor.remaining)
 430                         break;
 431
 432                 /* ttm_resource_ioremap only supports contiguous memory */
 433                 if (end != cursor.start)
 434                         return false;
 435
 436                 end = cursor.start + cursor.size;
 437         }
 438
 439         return end <= adev->gmc.visible_vram_size;
 440 }
 441
 442 /*
 443  * amdgpu_bo_move - Move a buffer object to a new memory location
 444  *
 445  * Called by ttm_bo_handle_move_mem()
 446  */
 447 static int amdgpu_bo_move(struct ttm_buffer_object *bo, bool evict,
 448                           struct ttm_operation_ctx *ctx,
 449                           struct ttm_resource *new_mem,
 450                           struct ttm_place *hop)
 451 {
 452         struct amdgpu_device *adev;
 453         struct amdgpu_bo *abo;
 454         struct ttm_resource *old_mem = bo->resource;
 455         int r;
 456
 457         if (new_mem->mem_type == TTM_PL_TT ||
 458             new_mem->mem_type == AMDGPU_PL_PREEMPT) {
 459                 r = amdgpu_ttm_backend_bind(bo->bdev, bo->ttm, new_mem);
 460                 if (r)
 461                         return r;
 462         }
 463
 464         abo = ttm_to_amdgpu_bo(bo);
 465         adev = amdgpu_ttm_adev(bo->bdev);
 466
 467         if (!old_mem || (old_mem->mem_type == TTM_PL_SYSTEM &&
 468                          bo->ttm == NULL)) {
 469                 ttm_bo_move_null(bo, new_mem);
 470                 goto out;
 471         }
 472         if (old_mem->mem_type == TTM_PL_SYSTEM &&
 473             (new_mem->mem_type == TTM_PL_TT ||
 474              new_mem->mem_type == AMDGPU_PL_PREEMPT)) {
 475                 ttm_bo_move_null(bo, new_mem);
 476                 goto out;
 477         }
 478         if ((old_mem->mem_type == TTM_PL_TT ||
 479              old_mem->mem_type == AMDGPU_PL_PREEMPT) &&
 480             new_mem->mem_type == TTM_PL_SYSTEM) {
 481                 r = ttm_bo_wait_ctx(bo, ctx);
 482                 if (r)
 483                         return r;
 484
 485                 amdgpu_ttm_backend_unbind(bo->bdev, bo->ttm);
 486                 ttm_resource_free(bo, &bo->resource);
 487                 ttm_bo_assign_mem(bo, new_mem);
 488                 goto out;
 489         }
 490
 491         if (old_mem->mem_type == AMDGPU_PL_GDS ||
 492             old_mem->mem_type == AMDGPU_PL_GWS ||
 493             old_mem->mem_type == AMDGPU_PL_OA ||
 494             old_mem->mem_type == AMDGPU_PL_DOORBELL ||
 495             new_mem->mem_type == AMDGPU_PL_GDS ||
 496             new_mem->mem_type == AMDGPU_PL_GWS ||
 497             new_mem->mem_type == AMDGPU_PL_OA ||
 498             new_mem->mem_type == AMDGPU_PL_DOORBELL) {
 499                 /* Nothing to save here */
 500                 ttm_bo_move_null(bo, new_mem);
 501                 goto out;
 502         }
 503
 504         if (bo->type == ttm_bo_type_device &&
 505             new_mem->mem_type == TTM_PL_VRAM &&
 506             old_mem->mem_type != TTM_PL_VRAM) {
 507                 /* amdgpu_bo_fault_reserve_notify will re-set this if the CPU
 508                  * accesses the BO after it's moved.
 509                  */
 510                 abo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
 511         }
 512
 513         if (adev->mman.buffer_funcs_enabled) {
 514                 if (((old_mem->mem_type == TTM_PL_SYSTEM &&
 515                       new_mem->mem_type == TTM_PL_VRAM) ||
 516                      (old_mem->mem_type == TTM_PL_VRAM &&
 517                       new_mem->mem_type == TTM_PL_SYSTEM))) {
 518                         hop->fpfn = 0;
 519                         hop->lpfn = 0;
 520                         hop->mem_type = TTM_PL_TT;
 521                         hop->flags = TTM_PL_FLAG_TEMPORARY;
 522                         return -EMULTIHOP;
 523                 }
 524
 525                 r = amdgpu_move_blit(bo, evict, new_mem, old_mem);
 526         } else {
 527                 r = -ENODEV;
 528         }
 529
 530         if (r) {
 531                 /* Check that all memory is CPU accessible */
 532                 if (!amdgpu_mem_visible(adev, old_mem) ||
 533                     !amdgpu_mem_visible(adev, new_mem)) {
 534                         pr_err("Move buffer fallback to memcpy unavailable\n");
 535                         return r;
 536                 }
 537
 538                 r = ttm_bo_move_memcpy(bo, ctx, new_mem);
 539                 if (r)
 540                         return r;
 541         }
 542
 543         trace_amdgpu_bo_move(abo, new_mem->mem_type, old_mem->mem_type);
 544 out:
 545         /* update statistics */
 546         atomic64_add(bo->base.size, &adev->num_bytes_moved);
 547         amdgpu_bo_move_notify(bo, evict);
 548         return 0;
 549 }
 550
 551 /*
 552  * amdgpu_ttm_io_mem_reserve - Reserve a block of memory during a fault
 553  *
 554  * Called by ttm_mem_io_reserve() ultimately via ttm_bo_vm_fault()
 555  */
 556 static int amdgpu_ttm_io_mem_reserve(struct ttm_device *bdev,
 557                                      struct ttm_resource *mem)
 558 {
 559         struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
 560         size_t bus_size = (size_t)mem->size;
 561
 562         switch (mem->mem_type) {
 563         case TTM_PL_SYSTEM:
 564                 /* system memory */
 565                 return 0;
 566         case TTM_PL_TT:
 567         case AMDGPU_PL_PREEMPT:
 568                 break;
 569         case TTM_PL_VRAM:
 570                 mem->bus.offset = mem->start << PAGE_SHIFT;
 571                 /* check if it's visible */
 572                 if ((mem->bus.offset + bus_size) > adev->gmc.visible_vram_size)
 573                         return -EINVAL;
 574
 575                 if (adev->mman.aper_base_kaddr &&
 576                     mem->placement & TTM_PL_FLAG_CONTIGUOUS)
 577                         mem->bus.addr = (u8 *)adev->mman.aper_base_kaddr +
 578                                         mem->bus.offset;
 579
 580                 mem->bus.offset += adev->gmc.aper_base;
 581                 mem->bus.is_iomem = true;
 582                 break;
 583         case AMDGPU_PL_DOORBELL:
 584                 mem->bus.offset = mem->start << PAGE_SHIFT;
 585                 mem->bus.offset += adev->doorbell.base;
 586                 mem->bus.is_iomem = true;
 587                 mem->bus.caching = ttm_uncached;
 588                 break;
 589         default:
 590                 return -EINVAL;
 591         }
 592         return 0;
 593 }
 594
 595 static unsigned long amdgpu_ttm_io_mem_pfn(struct ttm_buffer_object *bo,
 596                                            unsigned long page_offset)
 597 {
 598         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
 599         struct amdgpu_res_cursor cursor;
 600
 601         amdgpu_res_first(bo->resource, (u64)page_offset << PAGE_SHIFT, 0,
 602                          &cursor);
 603
 604         if (bo->resource->mem_type == AMDGPU_PL_DOORBELL)
 605                 return ((uint64_t)(adev->doorbell.base + cursor.start)) >> PAGE_SHIFT;
 606
 607         return (adev->gmc.aper_base + cursor.start) >> PAGE_SHIFT;
 608 }
 609
 610 /**
 611  * amdgpu_ttm_domain_start - Returns GPU start address
 612  * @adev: amdgpu device object
 613  * @type: type of the memory
 614  *
 615  * Returns:
 616  * GPU start address of a memory domain
 617  */
 618
 619 uint64_t amdgpu_ttm_domain_start(struct amdgpu_device *adev, uint32_t type)
 620 {
 621         switch (type) {
 622         case TTM_PL_TT:
 623                 return adev->gmc.gart_start;
 624         case TTM_PL_VRAM:
 625                 return adev->gmc.vram_start;
 626         }
 627
 628         return 0;
 629 }
 630
 631 /*
 632  * TTM backend functions.
 633  */
 634 struct amdgpu_ttm_tt {
 635         struct ttm_tt   ttm;
 636         struct drm_gem_object   *gobj;
 637         u64                     offset;
 638         uint64_t                userptr;
 639         struct task_struct      *usertask;
 640         uint32_t                userflags;
 641         bool                    bound;
 642         int32_t                 pool_id;
 643 };
 644
 645 #define ttm_to_amdgpu_ttm_tt(ptr)       container_of(ptr, struct amdgpu_ttm_tt, ttm)
 646
 647 #ifdef CONFIG_DRM_AMDGPU_USERPTR
 648 /*
 649  * amdgpu_ttm_tt_get_user_pages - get device accessible pages that back user
 650  * memory and start HMM tracking CPU page table update
 651  *
 652  * Calling function must call amdgpu_ttm_tt_userptr_range_done() once and only
 653  * once afterwards to stop HMM tracking
 654  */
 655 int amdgpu_ttm_tt_get_user_pages(struct amdgpu_bo *bo, struct page **pages,
 656                                  struct hmm_range **range)
 657 {
 658         struct ttm_tt *ttm = bo->tbo.ttm;
 659         struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
 660         unsigned long start = gtt->userptr;
 661         struct vm_area_struct *vma;
 662         struct mm_struct *mm;
 663         bool readonly;
 664         int r = 0;
 665
 666         /* Make sure get_user_pages_done() can cleanup gracefully */
 667         *range = NULL;
 668
 669         mm = bo->notifier.mm;
 670         if (unlikely(!mm)) {
 671                 DRM_DEBUG_DRIVER("BO is not registered?\n");
 672                 return -EFAULT;
 673         }
 674
 675         if (!mmget_not_zero(mm)) /* Happens during process shutdown */
 676                 return -ESRCH;
 677
 678         mmap_read_lock(mm);
 679         vma = vma_lookup(mm, start);
 680         if (unlikely(!vma)) {
 681                 r = -EFAULT;
 682                 goto out_unlock;
 683         }
 684         if (unlikely((gtt->userflags & AMDGPU_GEM_USERPTR_ANONONLY) &&
 685                 vma->vm_file)) {
 686                 r = -EPERM;
 687                 goto out_unlock;
 688         }
 689
 690         readonly = amdgpu_ttm_tt_is_readonly(ttm);
 691         r = amdgpu_hmm_range_get_pages(&bo->notifier, start, ttm->num_pages,
 692                                        readonly, NULL, pages, range);
 693 out_unlock:
 694         mmap_read_unlock(mm);
 695         if (r)
 696                 pr_debug("failed %d to get user pages 0x%lx\n", r, start);
 697
 698         mmput(mm);
 699
 700         return r;
 701 }
 702
 703 /* amdgpu_ttm_tt_discard_user_pages - Discard range and pfn array allocations
 704  */
 705 void amdgpu_ttm_tt_discard_user_pages(struct ttm_tt *ttm,
 706                                       struct hmm_range *range)
 707 {
 708         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 709
 710         if (gtt && gtt->userptr && range)
 711                 amdgpu_hmm_range_get_pages_done(range);
 712 }
 713
 714 /*
 715  * amdgpu_ttm_tt_get_user_pages_done - stop HMM track the CPU page table change
 716  * Check if the pages backing this ttm range have been invalidated
 717  *
 718  * Returns: true if pages are still valid
 719  */
 720 bool amdgpu_ttm_tt_get_user_pages_done(struct ttm_tt *ttm,
 721                                        struct hmm_range *range)
 722 {
 723         struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
 724
 725         if (!gtt || !gtt->userptr || !range)
 726                 return false;
 727
 728         DRM_DEBUG_DRIVER("user_pages_done 0x%llx pages 0x%x\n",
 729                 gtt->userptr, ttm->num_pages);
 730
 731         WARN_ONCE(!range->hmm_pfns, "No user pages to check\n");
 732
 733         return !amdgpu_hmm_range_get_pages_done(range);
 734 }
 735 #endif
 736
 737 /*
 738  * amdgpu_ttm_tt_set_user_pages - Copy pages in, putting old pages as necessary.
 739  *
 740  * Called by amdgpu_cs_list_validate(). This creates the page list
 741  * that backs user memory and will ultimately be mapped into the device
 742  * address space.
 743  */
 744 void amdgpu_ttm_tt_set_user_pages(struct ttm_tt *ttm, struct page **pages)
 745 {
 746         unsigned long i;
 747
 748         for (i = 0; i < ttm->num_pages; ++i)
 749                 ttm->pages[i] = pages ? pages[i] : NULL;
 750 }
 751
 752 /*
 753  * amdgpu_ttm_tt_pin_userptr - prepare the sg table with the user pages
 754  *
 755  * Called by amdgpu_ttm_backend_bind()
 756  **/
 757 static int amdgpu_ttm_tt_pin_userptr(struct ttm_device *bdev,
 758                                      struct ttm_tt *ttm)
 759 {
 760         struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
 761         struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
 762         int write = !(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
 763         enum dma_data_direction direction = write ?
 764                 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
 765         int r;
 766
 767         /* Allocate an SG array and squash pages into it */
 768         r = sg_alloc_table_from_pages(ttm->sg, ttm->pages, ttm->num_pages, 0,
 769                                       (u64)ttm->num_pages << PAGE_SHIFT,
 770                                       GFP_KERNEL);
 771         if (r)
 772                 goto release_sg;
 773
 774         /* Map SG to device */
 775         r = dma_map_sgtable(adev->dev, ttm->sg, direction, 0);
 776         if (r)
 777                 goto release_sg;
 778
 779         /* convert SG to linear array of pages and dma addresses */
 780         drm_prime_sg_to_dma_addr_array(ttm->sg, gtt->ttm.dma_address,
 781                                        ttm->num_pages);
 782
 783         return 0;
 784
 785 release_sg:
 786         kfree(ttm->sg);
 787         ttm->sg = NULL;
 788         return r;
 789 }
 790
 791 /*
 792  * amdgpu_ttm_tt_unpin_userptr - Unpin and unmap userptr pages
 793  */
 794 static void amdgpu_ttm_tt_unpin_userptr(struct ttm_device *bdev,
 795                                         struct ttm_tt *ttm)
 796 {
 797         struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
 798         struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
 799         int write = !(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
 800         enum dma_data_direction direction = write ?
 801                 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
 802
 803         /* double check that we don't free the table twice */
 804         if (!ttm->sg || !ttm->sg->sgl)
 805                 return;
 806
 807         /* unmap the pages mapped to the device */
 808         dma_unmap_sgtable(adev->dev, ttm->sg, direction, 0);
 809         sg_free_table(ttm->sg);
 810 }
 811
 812 /*
 813  * total_pages is constructed as MQD0+CtrlStack0 + MQD1+CtrlStack1 + ...
 814  * MQDn+CtrlStackn where n is the number of XCCs per partition.
 815  * pages_per_xcc is the size of one MQD+CtrlStack. The first page is MQD
 816  * and uses memory type default, UC. The rest of pages_per_xcc are
 817  * Ctrl stack and modify their memory type to NC.
 818  */
 819 static void amdgpu_ttm_gart_bind_gfx9_mqd(struct amdgpu_device *adev,
 820                                 struct ttm_tt *ttm, uint64_t flags)
 821 {
 822         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 823         uint64_t total_pages = ttm->num_pages;
 824         int num_xcc = max(1U, adev->gfx.num_xcc_per_xcp);
 825         uint64_t page_idx, pages_per_xcc;
 826         int i;
 827         uint64_t ctrl_flags = (flags & ~AMDGPU_PTE_MTYPE_VG10_MASK) |
 828                         AMDGPU_PTE_MTYPE_VG10(AMDGPU_MTYPE_NC);
 829
 830         pages_per_xcc = total_pages;
 831         do_div(pages_per_xcc, num_xcc);
 832
 833         for (i = 0, page_idx = 0; i < num_xcc; i++, page_idx += pages_per_xcc) {
 834                 /* MQD page: use default flags */
 835                 amdgpu_gart_bind(adev,
 836                                 gtt->offset + (page_idx << PAGE_SHIFT),
 837                                 1, &gtt->ttm.dma_address[page_idx], flags);
 838                 /*
 839                  * Ctrl pages - modify the memory type to NC (ctrl_flags) from
 840                  * the second page of the BO onward.
 841                  */
 842                 amdgpu_gart_bind(adev,
 843                                 gtt->offset + ((page_idx + 1) << PAGE_SHIFT),
 844                                 pages_per_xcc - 1,
 845                                 &gtt->ttm.dma_address[page_idx + 1],
 846                                 ctrl_flags);
 847         }
 848 }
 849
 850 static void amdgpu_ttm_gart_bind(struct amdgpu_device *adev,
 851                                  struct ttm_buffer_object *tbo,
 852                                  uint64_t flags)
 853 {
 854         struct amdgpu_bo *abo = ttm_to_amdgpu_bo(tbo);
 855         struct ttm_tt *ttm = tbo->ttm;
 856         struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
 857
 858         if (amdgpu_bo_encrypted(abo))
 859                 flags |= AMDGPU_PTE_TMZ;
 860
 861         if (abo->flags & AMDGPU_GEM_CREATE_CP_MQD_GFX9) {
 862                 amdgpu_ttm_gart_bind_gfx9_mqd(adev, ttm, flags);
 863         } else {
 864                 amdgpu_gart_bind(adev, gtt->offset, ttm->num_pages,
 865                                  gtt->ttm.dma_address, flags);
 866         }
 867 }
 868
 869 /*
 870  * amdgpu_ttm_backend_bind - Bind GTT memory
 871  *
 872  * Called by ttm_tt_bind() on behalf of ttm_bo_handle_move_mem().
 873  * This handles binding GTT memory to the device address space.
 874  */
 875 static int amdgpu_ttm_backend_bind(struct ttm_device *bdev,
 876                                    struct ttm_tt *ttm,
 877                                    struct ttm_resource *bo_mem)
 878 {
 879         struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
 880         struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
 881         uint64_t flags;
 882         int r;
 883
 884         if (!bo_mem)
 885                 return -EINVAL;
 886
 887         if (gtt->bound)
 888                 return 0;
 889
 890         if (gtt->userptr) {
 891                 r = amdgpu_ttm_tt_pin_userptr(bdev, ttm);
 892                 if (r) {
 893                         DRM_ERROR("failed to pin userptr\n");
 894                         return r;
 895                 }
 896         } else if (ttm->page_flags & TTM_TT_FLAG_EXTERNAL) {
 897                 if (!ttm->sg) {
 898                         struct dma_buf_attachment *attach;
 899                         struct sg_table *sgt;
 900
 901                         attach = gtt->gobj->import_attach;
 902                         sgt = dma_buf_map_attachment(attach, DMA_BIDIRECTIONAL);
 903                         if (IS_ERR(sgt))
 904                                 return PTR_ERR(sgt);
 905
 906                         ttm->sg = sgt;
 907                 }
 908
 909                 drm_prime_sg_to_dma_addr_array(ttm->sg, gtt->ttm.dma_address,
 910                                                ttm->num_pages);
 911         }
 912
 913         if (!ttm->num_pages) {
 914                 WARN(1, "nothing to bind %u pages for mreg %p back %p!\n",
 915                      ttm->num_pages, bo_mem, ttm);
 916         }
 917
 918         if (bo_mem->mem_type != TTM_PL_TT ||
 919             !amdgpu_gtt_mgr_has_gart_addr(bo_mem)) {
 920                 gtt->offset = AMDGPU_BO_INVALID_OFFSET;
 921                 return 0;
 922         }
 923
 924         /* compute PTE flags relevant to this BO memory */
 925         flags = amdgpu_ttm_tt_pte_flags(adev, ttm, bo_mem);
 926
 927         /* bind pages into GART page tables */
 928         gtt->offset = (u64)bo_mem->start << PAGE_SHIFT;
 929         amdgpu_gart_bind(adev, gtt->offset, ttm->num_pages,
 930                          gtt->ttm.dma_address, flags);
 931         gtt->bound = true;
 932         return 0;
 933 }
 934
 935 /*
 936  * amdgpu_ttm_alloc_gart - Make sure buffer object is accessible either
 937  * through AGP or GART aperture.
 938  *
 939  * If bo is accessible through AGP aperture, then use AGP aperture
 940  * to access bo; otherwise allocate logical space in GART aperture
 941  * and map bo to GART aperture.
 942  */
 943 int amdgpu_ttm_alloc_gart(struct ttm_buffer_object *bo)
 944 {
 945         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
 946         struct ttm_operation_ctx ctx = { false, false };
 947         struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(bo->ttm);
 948         struct ttm_placement placement;
 949         struct ttm_place placements;
 950         struct ttm_resource *tmp;
 951         uint64_t addr, flags;
 952         int r;
 953
 954         if (bo->resource->start != AMDGPU_BO_INVALID_OFFSET)
 955                 return 0;
 956
 957         addr = amdgpu_gmc_agp_addr(bo);
 958         if (addr != AMDGPU_BO_INVALID_OFFSET)
 959                 return 0;
 960
 961         /* allocate GART space */
 962         placement.num_placement = 1;
 963         placement.placement = &placements;
 964         placements.fpfn = 0;
 965         placements.lpfn = adev->gmc.gart_size >> PAGE_SHIFT;
 966         placements.mem_type = TTM_PL_TT;
 967         placements.flags = bo->resource->placement;
 968
 969         r = ttm_bo_mem_space(bo, &placement, &tmp, &ctx);
 970         if (unlikely(r))
 971                 return r;
 972
 973         /* compute PTE flags for this buffer object */
 974         flags = amdgpu_ttm_tt_pte_flags(adev, bo->ttm, tmp);
 975
 976         /* Bind pages */
 977         gtt->offset = (u64)tmp->start << PAGE_SHIFT;
 978         amdgpu_ttm_gart_bind(adev, bo, flags);
 979         amdgpu_gart_invalidate_tlb(adev);
 980         ttm_resource_free(bo, &bo->resource);
 981         ttm_bo_assign_mem(bo, tmp);
 982
 983         return 0;
 984 }
 985
 986 /*
 987  * amdgpu_ttm_recover_gart - Rebind GTT pages
 988  *
 989  * Called by amdgpu_gtt_mgr_recover() from amdgpu_device_reset() to
 990  * rebind GTT pages during a GPU reset.
 991  */
 992 void amdgpu_ttm_recover_gart(struct ttm_buffer_object *tbo)
 993 {
 994         struct amdgpu_device *adev = amdgpu_ttm_adev(tbo->bdev);
 995         uint64_t flags;
 996
 997         if (!tbo->ttm)
 998                 return;
 999
1000         flags = amdgpu_ttm_tt_pte_flags(adev, tbo->ttm, tbo->resource);
1001         amdgpu_ttm_gart_bind(adev, tbo, flags);
1002 }
1003
1004 /*
1005  * amdgpu_ttm_backend_unbind - Unbind GTT mapped pages
1006  *
1007  * Called by ttm_tt_unbind() on behalf of ttm_bo_move_ttm() and
1008  * ttm_tt_destroy().
1009  */
1010 static void amdgpu_ttm_backend_unbind(struct ttm_device *bdev,
1011                                       struct ttm_tt *ttm)
1012 {
1013         struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
1014         struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
1015
1016         /* if the pages have userptr pinning then clear that first */
1017         if (gtt->userptr) {
1018                 amdgpu_ttm_tt_unpin_userptr(bdev, ttm);
1019         } else if (ttm->sg && gtt->gobj->import_attach) {
1020                 struct dma_buf_attachment *attach;
1021
1022                 attach = gtt->gobj->import_attach;
1023                 dma_buf_unmap_attachment(attach, ttm->sg, DMA_BIDIRECTIONAL);
1024                 ttm->sg = NULL;
1025         }
1026
1027         if (!gtt->bound)
1028                 return;
1029
1030         if (gtt->offset == AMDGPU_BO_INVALID_OFFSET)
1031                 return;
1032
1033         /* unbind shouldn't be done for GDS/GWS/OA in ttm_bo_clean_mm */
1034         amdgpu_gart_unbind(adev, gtt->offset, ttm->num_pages);
1035         gtt->bound = false;
1036 }
1037
1038 static void amdgpu_ttm_backend_destroy(struct ttm_device *bdev,
1039                                        struct ttm_tt *ttm)
1040 {
1041         struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
1042
1043         if (gtt->usertask)
1044                 put_task_struct(gtt->usertask);
1045
1046         ttm_tt_fini(&gtt->ttm);
1047         kfree(gtt);
1048 }
1049
1050 /**
1051  * amdgpu_ttm_tt_create - Create a ttm_tt object for a given BO
1052  *
1053  * @bo: The buffer object to create a GTT ttm_tt object around
1054  * @page_flags: Page flags to be added to the ttm_tt object
1055  *
1056  * Called by ttm_tt_create().
1057  */
1058 static struct ttm_tt *amdgpu_ttm_tt_create(struct ttm_buffer_object *bo,
1059                                            uint32_t page_flags)
1060 {
1061         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1062         struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
1063         struct amdgpu_ttm_tt *gtt;
1064         enum ttm_caching caching;
1065
1066         gtt = kzalloc(sizeof(struct amdgpu_ttm_tt), GFP_KERNEL);
1067         if (!gtt)
1068                 return NULL;
1069
1070         gtt->gobj = &bo->base;
1071         if (adev->gmc.mem_partitions && abo->xcp_id >= 0)
1072                 gtt->pool_id = KFD_XCP_MEM_ID(adev, abo->xcp_id);
1073         else
1074                 gtt->pool_id = abo->xcp_id;
1075
1076         if (abo->flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
1077                 caching = ttm_write_combined;
1078         else
1079                 caching = ttm_cached;
1080
1081         /* allocate space for the uninitialized page entries */
1082         if (ttm_sg_tt_init(&gtt->ttm, bo, page_flags, caching)) {
1083                 kfree(gtt);
1084                 return NULL;
1085         }
1086         return &gtt->ttm;
1087 }
1088
1089 /*
1090  * amdgpu_ttm_tt_populate - Map GTT pages visible to the device
1091  *
1092  * Map the pages of a ttm_tt object to an address space visible
1093  * to the underlying device.
1094  */
1095 static int amdgpu_ttm_tt_populate(struct ttm_device *bdev,
1096                                   struct ttm_tt *ttm,
1097                                   struct ttm_operation_ctx *ctx)
1098 {
1099         struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
1100         struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
1101         struct ttm_pool *pool;
1102         pgoff_t i;
1103         int ret;
1104
1105         /* user pages are bound by amdgpu_ttm_tt_pin_userptr() */
1106         if (gtt->userptr) {
1107                 ttm->sg = kzalloc(sizeof(struct sg_table), GFP_KERNEL);
1108                 if (!ttm->sg)
1109                         return -ENOMEM;
1110                 return 0;
1111         }
1112
1113         if (ttm->page_flags & TTM_TT_FLAG_EXTERNAL)
1114                 return 0;
1115
1116         if (adev->mman.ttm_pools && gtt->pool_id >= 0)
1117                 pool = &adev->mman.ttm_pools[gtt->pool_id];
1118         else
1119                 pool = &adev->mman.bdev.pool;
1120         ret = ttm_pool_alloc(pool, ttm, ctx);
1121         if (ret)
1122                 return ret;
1123
1124         for (i = 0; i < ttm->num_pages; ++i)
1125                 ttm->pages[i]->mapping = bdev->dev_mapping;
1126
1127         return 0;
1128 }
1129
1130 /*
1131  * amdgpu_ttm_tt_unpopulate - unmap GTT pages and unpopulate page arrays
1132  *
1133  * Unmaps pages of a ttm_tt object from the device address space and
1134  * unpopulates the page array backing it.
1135  */
1136 static void amdgpu_ttm_tt_unpopulate(struct ttm_device *bdev,
1137                                      struct ttm_tt *ttm)
1138 {
1139         struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
1140         struct amdgpu_device *adev;
1141         struct ttm_pool *pool;
1142         pgoff_t i;
1143
1144         amdgpu_ttm_backend_unbind(bdev, ttm);
1145
1146         if (gtt->userptr) {
1147                 amdgpu_ttm_tt_set_user_pages(ttm, NULL);
1148                 kfree(ttm->sg);
1149                 ttm->sg = NULL;
1150                 return;
1151         }
1152
1153         if (ttm->page_flags & TTM_TT_FLAG_EXTERNAL)
1154                 return;
1155
1156         for (i = 0; i < ttm->num_pages; ++i)
1157                 ttm->pages[i]->mapping = NULL;
1158
1159         adev = amdgpu_ttm_adev(bdev);
1160
1161         if (adev->mman.ttm_pools && gtt->pool_id >= 0)
1162                 pool = &adev->mman.ttm_pools[gtt->pool_id];
1163         else
1164                 pool = &adev->mman.bdev.pool;
1165
1166         return ttm_pool_free(pool, ttm);
1167 }
1168
1169 /**
1170  * amdgpu_ttm_tt_get_userptr - Return the userptr GTT ttm_tt for the current
1171  * task
1172  *
1173  * @tbo: The ttm_buffer_object that contains the userptr
1174  * @user_addr:  The returned value
1175  */
1176 int amdgpu_ttm_tt_get_userptr(const struct ttm_buffer_object *tbo,
1177                               uint64_t *user_addr)
1178 {
1179         struct amdgpu_ttm_tt *gtt;
1180
1181         if (!tbo->ttm)
1182                 return -EINVAL;
1183
1184         gtt = (void *)tbo->ttm;
1185         *user_addr = gtt->userptr;
1186         return 0;
1187 }
1188
1189 /**
1190  * amdgpu_ttm_tt_set_userptr - Initialize userptr GTT ttm_tt for the current
1191  * task
1192  *
1193  * @bo: The ttm_buffer_object to bind this userptr to
1194  * @addr:  The address in the current tasks VM space to use
1195  * @flags: Requirements of userptr object.
1196  *
1197  * Called by amdgpu_gem_userptr_ioctl() and kfd_ioctl_alloc_memory_of_gpu() to
1198  * bind userptr pages to current task and by kfd_ioctl_acquire_vm() to
1199  * initialize GPU VM for a KFD process.
1200  */
1201 int amdgpu_ttm_tt_set_userptr(struct ttm_buffer_object *bo,
1202                               uint64_t addr, uint32_t flags)
1203 {
1204         struct amdgpu_ttm_tt *gtt;
1205
1206         if (!bo->ttm) {
1207                 /* TODO: We want a separate TTM object type for userptrs */
1208                 bo->ttm = amdgpu_ttm_tt_create(bo, 0);
1209                 if (bo->ttm == NULL)
1210                         return -ENOMEM;
1211         }
1212
1213         /* Set TTM_TT_FLAG_EXTERNAL before populate but after create. */
1214         bo->ttm->page_flags |= TTM_TT_FLAG_EXTERNAL;
1215
1216         gtt = ttm_to_amdgpu_ttm_tt(bo->ttm);
1217         gtt->userptr = addr;
1218         gtt->userflags = flags;
1219
1220         if (gtt->usertask)
1221                 put_task_struct(gtt->usertask);
1222         gtt->usertask = current->group_leader;
1223         get_task_struct(gtt->usertask);
1224
1225         return 0;
1226 }
1227
1228 /*
1229  * amdgpu_ttm_tt_get_usermm - Return memory manager for ttm_tt object
1230  */
1231 struct mm_struct *amdgpu_ttm_tt_get_usermm(struct ttm_tt *ttm)
1232 {
1233         struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
1234
1235         if (gtt == NULL)
1236                 return NULL;
1237
1238         if (gtt->usertask == NULL)
1239                 return NULL;
1240
1241         return gtt->usertask->mm;
1242 }
1243
1244 /*
1245  * amdgpu_ttm_tt_affect_userptr - Determine if a ttm_tt object lays inside an
1246  * address range for the current task.
1247  *
1248  */
1249 bool amdgpu_ttm_tt_affect_userptr(struct ttm_tt *ttm, unsigned long start,
1250                                   unsigned long end, unsigned long *userptr)
1251 {
1252         struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
1253         unsigned long size;
1254
1255         if (gtt == NULL || !gtt->userptr)
1256                 return false;
1257
1258         /* Return false if no part of the ttm_tt object lies within
1259          * the range
1260          */
1261         size = (unsigned long)gtt->ttm.num_pages * PAGE_SIZE;
1262         if (gtt->userptr > end || gtt->userptr + size <= start)
1263                 return false;
1264
1265         if (userptr)
1266                 *userptr = gtt->userptr;
1267         return true;
1268 }
1269
1270 /*
1271  * amdgpu_ttm_tt_is_userptr - Have the pages backing by userptr?
1272  */
1273 bool amdgpu_ttm_tt_is_userptr(struct ttm_tt *ttm)
1274 {
1275         struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
1276
1277         if (gtt == NULL || !gtt->userptr)
1278                 return false;
1279
1280         return true;
1281 }
1282
1283 /*
1284  * amdgpu_ttm_tt_is_readonly - Is the ttm_tt object read only?
1285  */
1286 bool amdgpu_ttm_tt_is_readonly(struct ttm_tt *ttm)
1287 {
1288         struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
1289
1290         if (gtt == NULL)
1291                 return false;
1292
1293         return !!(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
1294 }
1295
1296 /**
1297  * amdgpu_ttm_tt_pde_flags - Compute PDE flags for ttm_tt object
1298  *
1299  * @ttm: The ttm_tt object to compute the flags for
1300  * @mem: The memory registry backing this ttm_tt object
1301  *
1302  * Figure out the flags to use for a VM PDE (Page Directory Entry).
1303  */
1304 uint64_t amdgpu_ttm_tt_pde_flags(struct ttm_tt *ttm, struct ttm_resource *mem)
1305 {
1306         uint64_t flags = 0;
1307
1308         if (mem && mem->mem_type != TTM_PL_SYSTEM)
1309                 flags |= AMDGPU_PTE_VALID;
1310
1311         if (mem && (mem->mem_type == TTM_PL_TT ||
1312                     mem->mem_type == AMDGPU_PL_DOORBELL ||
1313                     mem->mem_type == AMDGPU_PL_PREEMPT)) {
1314                 flags |= AMDGPU_PTE_SYSTEM;
1315
1316                 if (ttm->caching == ttm_cached)
1317                         flags |= AMDGPU_PTE_SNOOPED;
1318         }
1319
1320         if (mem && mem->mem_type == TTM_PL_VRAM &&
1321                         mem->bus.caching == ttm_cached)
1322                 flags |= AMDGPU_PTE_SNOOPED;
1323
1324         return flags;
1325 }
1326
1327 /**
1328  * amdgpu_ttm_tt_pte_flags - Compute PTE flags for ttm_tt object
1329  *
1330  * @adev: amdgpu_device pointer
1331  * @ttm: The ttm_tt object to compute the flags for
1332  * @mem: The memory registry backing this ttm_tt object
1333  *
1334  * Figure out the flags to use for a VM PTE (Page Table Entry).
1335  */
1336 uint64_t amdgpu_ttm_tt_pte_flags(struct amdgpu_device *adev, struct ttm_tt *ttm,
1337                                  struct ttm_resource *mem)
1338 {
1339         uint64_t flags = amdgpu_ttm_tt_pde_flags(ttm, mem);
1340
1341         flags |= adev->gart.gart_pte_flags;
1342         flags |= AMDGPU_PTE_READABLE;
1343
1344         if (!amdgpu_ttm_tt_is_readonly(ttm))
1345                 flags |= AMDGPU_PTE_WRITEABLE;
1346
1347         return flags;
1348 }
1349
1350 /*
1351  * amdgpu_ttm_bo_eviction_valuable - Check to see if we can evict a buffer
1352  * object.
1353  *
1354  * Return true if eviction is sensible. Called by ttm_mem_evict_first() on
1355  * behalf of ttm_bo_mem_force_space() which tries to evict buffer objects until
1356  * it can find space for a new object and by ttm_bo_force_list_clean() which is
1357  * used to clean out a memory space.
1358  */
1359 static bool amdgpu_ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
1360                                             const struct ttm_place *place)
1361 {
1362         struct dma_resv_iter resv_cursor;
1363         struct dma_fence *f;
1364
1365         if (!amdgpu_bo_is_amdgpu_bo(bo))
1366                 return ttm_bo_eviction_valuable(bo, place);
1367
1368         /* Swapout? */
1369         if (bo->resource->mem_type == TTM_PL_SYSTEM)
1370                 return true;
1371
1372         if (bo->type == ttm_bo_type_kernel &&
1373             !amdgpu_vm_evictable(ttm_to_amdgpu_bo(bo)))
1374                 return false;
1375
1376         /* If bo is a KFD BO, check if the bo belongs to the current process.
1377          * If true, then return false as any KFD process needs all its BOs to
1378          * be resident to run successfully
1379          */
1380         dma_resv_for_each_fence(&resv_cursor, bo->base.resv,
1381                                 DMA_RESV_USAGE_BOOKKEEP, f) {
1382                 if (amdkfd_fence_check_mm(f, current->mm))
1383                         return false;
1384         }
1385
1386         /* Preemptible BOs don't own system resources managed by the
1387          * driver (pages, VRAM, GART space). They point to resources
1388          * owned by someone else (e.g. pageable memory in user mode
1389          * or a DMABuf). They are used in a preemptible context so we
1390          * can guarantee no deadlocks and good QoS in case of MMU
1391          * notifiers or DMABuf move notifiers from the resource owner.
1392          */
1393         if (bo->resource->mem_type == AMDGPU_PL_PREEMPT)
1394                 return false;
1395
1396         if (bo->resource->mem_type == TTM_PL_TT &&
1397             amdgpu_bo_encrypted(ttm_to_amdgpu_bo(bo)))
1398                 return false;
1399
1400         return ttm_bo_eviction_valuable(bo, place);
1401 }
1402
1403 static void amdgpu_ttm_vram_mm_access(struct amdgpu_device *adev, loff_t pos,
1404                                       void *buf, size_t size, bool write)
1405 {
1406         while (size) {
1407                 uint64_t aligned_pos = ALIGN_DOWN(pos, 4);
1408                 uint64_t bytes = 4 - (pos & 0x3);
1409                 uint32_t shift = (pos & 0x3) * 8;
1410                 uint32_t mask = 0xffffffff << shift;
1411                 uint32_t value = 0;
1412
1413                 if (size < bytes) {
1414                         mask &= 0xffffffff >> (bytes - size) * 8;
1415                         bytes = size;
1416                 }
1417
1418                 if (mask != 0xffffffff) {
1419                         amdgpu_device_mm_access(adev, aligned_pos, &value, 4, false);
1420                         if (write) {
1421                                 value &= ~mask;
1422                                 value |= (*(uint32_t *)buf << shift) & mask;
1423                                 amdgpu_device_mm_access(adev, aligned_pos, &value, 4, true);
1424                         } else {
1425                                 value = (value & mask) >> shift;
1426                                 memcpy(buf, &value, bytes);
1427                         }
1428                 } else {
1429                         amdgpu_device_mm_access(adev, aligned_pos, buf, 4, write);
1430                 }
1431
1432                 pos += bytes;
1433                 buf += bytes;
1434                 size -= bytes;
1435         }
1436 }
1437
1438 static int amdgpu_ttm_access_memory_sdma(struct ttm_buffer_object *bo,
1439                                         unsigned long offset, void *buf,
1440                                         int len, int write)
1441 {
1442         struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
1443         struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev);
1444         struct amdgpu_res_cursor src_mm;
1445         struct amdgpu_job *job;
1446         struct dma_fence *fence;
1447         uint64_t src_addr, dst_addr;
1448         unsigned int num_dw;
1449         int r, idx;
1450
1451         if (len != PAGE_SIZE)
1452                 return -EINVAL;
1453
1454         if (!adev->mman.sdma_access_ptr)
1455                 return -EACCES;
1456
1457         if (!drm_dev_enter(adev_to_drm(adev), &idx))
1458                 return -ENODEV;
1459
1460         if (write)
1461                 memcpy(adev->mman.sdma_access_ptr, buf, len);
1462
1463         num_dw = ALIGN(adev->mman.buffer_funcs->copy_num_dw, 8);
1464         r = amdgpu_job_alloc_with_ib(adev, &adev->mman.high_pr,
1465                                      AMDGPU_FENCE_OWNER_UNDEFINED,
1466                                      num_dw * 4, AMDGPU_IB_POOL_DELAYED,
1467                                      &job);
1468         if (r)
1469                 goto out;
1470
1471         amdgpu_res_first(abo->tbo.resource, offset, len, &src_mm);
1472         src_addr = amdgpu_ttm_domain_start(adev, bo->resource->mem_type) +
1473                 src_mm.start;
1474         dst_addr = amdgpu_bo_gpu_offset(adev->mman.sdma_access_bo);
1475         if (write)
1476                 swap(src_addr, dst_addr);
1477
1478         amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_addr, dst_addr,
1479                                 PAGE_SIZE, false);
1480
1481         amdgpu_ring_pad_ib(adev->mman.buffer_funcs_ring, &job->ibs[0]);
1482         WARN_ON(job->ibs[0].length_dw > num_dw);
1483
1484         fence = amdgpu_job_submit(job);
1485
1486         if (!dma_fence_wait_timeout(fence, false, adev->sdma_timeout))
1487                 r = -ETIMEDOUT;
1488         dma_fence_put(fence);
1489
1490         if (!(r || write))
1491                 memcpy(buf, adev->mman.sdma_access_ptr, len);
1492 out:
1493         drm_dev_exit(idx);
1494         return r;
1495 }
1496
1497 /**
1498  * amdgpu_ttm_access_memory - Read or Write memory that backs a buffer object.
1499  *
1500  * @bo:  The buffer object to read/write
1501  * @offset:  Offset into buffer object
1502  * @buf:  Secondary buffer to write/read from
1503  * @len: Length in bytes of access
1504  * @write:  true if writing
1505  *
1506  * This is used to access VRAM that backs a buffer object via MMIO
1507  * access for debugging purposes.
1508  */
1509 static int amdgpu_ttm_access_memory(struct ttm_buffer_object *bo,
1510                                     unsigned long offset, void *buf, int len,
1511                                     int write)
1512 {
1513         struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
1514         struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev);
1515         struct amdgpu_res_cursor cursor;
1516         int ret = 0;
1517
1518         if (bo->resource->mem_type != TTM_PL_VRAM)
1519                 return -EIO;
1520
1521         if (amdgpu_device_has_timeouts_enabled(adev) &&
1522                         !amdgpu_ttm_access_memory_sdma(bo, offset, buf, len, write))
1523                 return len;
1524
1525         amdgpu_res_first(bo->resource, offset, len, &cursor);
1526         while (cursor.remaining) {
1527                 size_t count, size = cursor.size;
1528                 loff_t pos = cursor.start;
1529
1530                 count = amdgpu_device_aper_access(adev, pos, buf, size, write);
1531                 size -= count;
1532                 if (size) {
1533                         /* using MM to access rest vram and handle un-aligned address */
1534                         pos += count;
1535                         buf += count;
1536                         amdgpu_ttm_vram_mm_access(adev, pos, buf, size, write);
1537                 }
1538
1539                 ret += cursor.size;
1540                 buf += cursor.size;
1541                 amdgpu_res_next(&cursor, cursor.size);
1542         }
1543
1544         return ret;
1545 }
1546
1547 static void
1548 amdgpu_bo_delete_mem_notify(struct ttm_buffer_object *bo)
1549 {
1550         amdgpu_bo_move_notify(bo, false);
1551 }
1552
1553 static struct ttm_device_funcs amdgpu_bo_driver = {
1554         .ttm_tt_create = &amdgpu_ttm_tt_create,
1555         .ttm_tt_populate = &amdgpu_ttm_tt_populate,
1556         .ttm_tt_unpopulate = &amdgpu_ttm_tt_unpopulate,
1557         .ttm_tt_destroy = &amdgpu_ttm_backend_destroy,
1558         .eviction_valuable = amdgpu_ttm_bo_eviction_valuable,
1559         .evict_flags = &amdgpu_evict_flags,
1560         .move = &amdgpu_bo_move,
1561         .delete_mem_notify = &amdgpu_bo_delete_mem_notify,
1562         .release_notify = &amdgpu_bo_release_notify,
1563         .io_mem_reserve = &amdgpu_ttm_io_mem_reserve,
1564         .io_mem_pfn = amdgpu_ttm_io_mem_pfn,
1565         .access_memory = &amdgpu_ttm_access_memory,
1566 };
1567
1568 /*
1569  * Firmware Reservation functions
1570  */
1571 /**
1572  * amdgpu_ttm_fw_reserve_vram_fini - free fw reserved vram
1573  *
1574  * @adev: amdgpu_device pointer
1575  *
1576  * free fw reserved vram if it has been reserved.
1577  */
1578 static void amdgpu_ttm_fw_reserve_vram_fini(struct amdgpu_device *adev)
1579 {
1580         amdgpu_bo_free_kernel(&adev->mman.fw_vram_usage_reserved_bo,
1581                 NULL, &adev->mman.fw_vram_usage_va);
1582 }
1583
1584 /*
1585  * Driver Reservation functions
1586  */
1587 /**
1588  * amdgpu_ttm_drv_reserve_vram_fini - free drv reserved vram
1589  *
1590  * @adev: amdgpu_device pointer
1591  *
1592  * free drv reserved vram if it has been reserved.
1593  */
1594 static void amdgpu_ttm_drv_reserve_vram_fini(struct amdgpu_device *adev)
1595 {
1596         amdgpu_bo_free_kernel(&adev->mman.drv_vram_usage_reserved_bo,
1597                                                   NULL,
1598                                                   &adev->mman.drv_vram_usage_va);
1599 }
1600
1601 /**
1602  * amdgpu_ttm_fw_reserve_vram_init - create bo vram reservation from fw
1603  *
1604  * @adev: amdgpu_device pointer
1605  *
1606  * create bo vram reservation from fw.
1607  */
1608 static int amdgpu_ttm_fw_reserve_vram_init(struct amdgpu_device *adev)
1609 {
1610         uint64_t vram_size = adev->gmc.visible_vram_size;
1611
1612         adev->mman.fw_vram_usage_va = NULL;
1613         adev->mman.fw_vram_usage_reserved_bo = NULL;
1614
1615         if (adev->mman.fw_vram_usage_size == 0 ||
1616             adev->mman.fw_vram_usage_size > vram_size)
1617                 return 0;
1618
1619         return amdgpu_bo_create_kernel_at(adev,
1620                                           adev->mman.fw_vram_usage_start_offset,
1621                                           adev->mman.fw_vram_usage_size,
1622                                           &adev->mman.fw_vram_usage_reserved_bo,
1623                                           &adev->mman.fw_vram_usage_va);
1624 }
1625
1626 /**
1627  * amdgpu_ttm_drv_reserve_vram_init - create bo vram reservation from driver
1628  *
1629  * @adev: amdgpu_device pointer
1630  *
1631  * create bo vram reservation from drv.
1632  */
1633 static int amdgpu_ttm_drv_reserve_vram_init(struct amdgpu_device *adev)
1634 {
1635         u64 vram_size = adev->gmc.visible_vram_size;
1636
1637         adev->mman.drv_vram_usage_va = NULL;
1638         adev->mman.drv_vram_usage_reserved_bo = NULL;
1639
1640         if (adev->mman.drv_vram_usage_size == 0 ||
1641             adev->mman.drv_vram_usage_size > vram_size)
1642                 return 0;
1643
1644         return amdgpu_bo_create_kernel_at(adev,
1645                                           adev->mman.drv_vram_usage_start_offset,
1646                                           adev->mman.drv_vram_usage_size,
1647                                           &adev->mman.drv_vram_usage_reserved_bo,
1648                                           &adev->mman.drv_vram_usage_va);
1649 }
1650
1651 /*
1652  * Memoy training reservation functions
1653  */
1654
1655 /**
1656  * amdgpu_ttm_training_reserve_vram_fini - free memory training reserved vram
1657  *
1658  * @adev: amdgpu_device pointer
1659  *
1660  * free memory training reserved vram if it has been reserved.
1661  */
1662 static int amdgpu_ttm_training_reserve_vram_fini(struct amdgpu_device *adev)
1663 {
1664         struct psp_memory_training_context *ctx = &adev->psp.mem_train_ctx;
1665
1666         ctx->init = PSP_MEM_TRAIN_NOT_SUPPORT;
1667         amdgpu_bo_free_kernel(&ctx->c2p_bo, NULL, NULL);
1668         ctx->c2p_bo = NULL;
1669
1670         return 0;
1671 }
1672
1673 static void amdgpu_ttm_training_data_block_init(struct amdgpu_device *adev,
1674                                                 uint32_t reserve_size)
1675 {
1676         struct psp_memory_training_context *ctx = &adev->psp.mem_train_ctx;
1677
1678         memset(ctx, 0, sizeof(*ctx));
1679
1680         ctx->c2p_train_data_offset =
1681                 ALIGN((adev->gmc.mc_vram_size - reserve_size - SZ_1M), SZ_1M);
1682         ctx->p2c_train_data_offset =
1683                 (adev->gmc.mc_vram_size - GDDR6_MEM_TRAINING_OFFSET);
1684         ctx->train_data_size =
1685                 GDDR6_MEM_TRAINING_DATA_SIZE_IN_BYTES;
1686
1687         DRM_DEBUG("train_data_size:%llx,p2c_train_data_offset:%llx,c2p_train_data_offset:%llx.\n",
1688                         ctx->train_data_size,
1689                         ctx->p2c_train_data_offset,
1690                         ctx->c2p_train_data_offset);
1691 }
1692
1693 /*
1694  * reserve TMR memory at the top of VRAM which holds
1695  * IP Discovery data and is protected by PSP.
1696  */
1697 static int amdgpu_ttm_reserve_tmr(struct amdgpu_device *adev)
1698 {
1699         struct psp_memory_training_context *ctx = &adev->psp.mem_train_ctx;
1700         bool mem_train_support = false;
1701         uint32_t reserve_size = 0;
1702         int ret;
1703
1704         if (adev->bios && !amdgpu_sriov_vf(adev)) {
1705                 if (amdgpu_atomfirmware_mem_training_supported(adev))
1706                         mem_train_support = true;
1707                 else
1708                         DRM_DEBUG("memory training does not support!\n");
1709         }
1710
1711         /*
1712          * Query reserved tmr size through atom firmwareinfo for Sienna_Cichlid and onwards for all
1713          * the use cases (IP discovery/G6 memory training/profiling/diagnostic data.etc)
1714          *
1715          * Otherwise, fallback to legacy approach to check and reserve tmr block for ip
1716          * discovery data and G6 memory training data respectively
1717          */
1718         if (adev->bios)
1719                 reserve_size =
1720                         amdgpu_atomfirmware_get_fw_reserved_fb_size(adev);
1721
1722         if (!adev->bios &&
1723             amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 3))
1724                 reserve_size = max(reserve_size, (uint32_t)280 << 20);
1725         else if (!reserve_size)
1726                 reserve_size = DISCOVERY_TMR_OFFSET;
1727
1728         if (mem_train_support) {
1729                 /* reserve vram for mem train according to TMR location */
1730                 amdgpu_ttm_training_data_block_init(adev, reserve_size);
1731                 ret = amdgpu_bo_create_kernel_at(adev,
1732                                                  ctx->c2p_train_data_offset,
1733                                                  ctx->train_data_size,
1734                                                  &ctx->c2p_bo,
1735                                                  NULL);
1736                 if (ret) {
1737                         DRM_ERROR("alloc c2p_bo failed(%d)!\n", ret);
1738                         amdgpu_ttm_training_reserve_vram_fini(adev);
1739                         return ret;
1740                 }
1741                 ctx->init = PSP_MEM_TRAIN_RESERVE_SUCCESS;
1742         }
1743
1744         if (!adev->gmc.is_app_apu) {
1745                 ret = amdgpu_bo_create_kernel_at(
1746                         adev, adev->gmc.real_vram_size - reserve_size,
1747                         reserve_size, &adev->mman.fw_reserved_memory, NULL);
1748                 if (ret) {
1749                         DRM_ERROR("alloc tmr failed(%d)!\n", ret);
1750                         amdgpu_bo_free_kernel(&adev->mman.fw_reserved_memory,
1751                                               NULL, NULL);
1752                         return ret;
1753                 }
1754         } else {
1755                 DRM_DEBUG_DRIVER("backdoor fw loading path for PSP TMR, no reservation needed\n");
1756         }
1757
1758         return 0;
1759 }
1760
1761 static int amdgpu_ttm_pools_init(struct amdgpu_device *adev)
1762 {
1763         int i;
1764
1765         if (!adev->gmc.is_app_apu || !adev->gmc.num_mem_partitions)
1766                 return 0;
1767
1768         adev->mman.ttm_pools = kcalloc(adev->gmc.num_mem_partitions,
1769                                        sizeof(*adev->mman.ttm_pools),
1770                                        GFP_KERNEL);
1771         if (!adev->mman.ttm_pools)
1772                 return -ENOMEM;
1773
1774         for (i = 0; i < adev->gmc.num_mem_partitions; i++) {
1775                 ttm_pool_init(&adev->mman.ttm_pools[i], adev->dev,
1776                               adev->gmc.mem_partitions[i].numa.node,
1777                               false, false);
1778         }
1779         return 0;
1780 }
1781
1782 static void amdgpu_ttm_pools_fini(struct amdgpu_device *adev)
1783 {
1784         int i;
1785
1786         if (!adev->gmc.is_app_apu || !adev->mman.ttm_pools)
1787                 return;
1788
1789         for (i = 0; i < adev->gmc.num_mem_partitions; i++)
1790                 ttm_pool_fini(&adev->mman.ttm_pools[i]);
1791
1792         kfree(adev->mman.ttm_pools);
1793         adev->mman.ttm_pools = NULL;
1794 }
1795
1796 /*
1797  * amdgpu_ttm_init - Init the memory management (ttm) as well as various
1798  * gtt/vram related fields.
1799  *
1800  * This initializes all of the memory space pools that the TTM layer
1801  * will need such as the GTT space (system memory mapped to the device),
1802  * VRAM (on-board memory), and on-chip memories (GDS, GWS, OA) which
1803  * can be mapped per VMID.
1804  */
1805 int amdgpu_ttm_init(struct amdgpu_device *adev)
1806 {
1807         uint64_t gtt_size;
1808         int r;
1809
1810         mutex_init(&adev->mman.gtt_window_lock);
1811
1812         /* No others user of address space so set it to 0 */
1813         r = ttm_device_init(&adev->mman.bdev, &amdgpu_bo_driver, adev->dev,
1814                                adev_to_drm(adev)->anon_inode->i_mapping,
1815                                adev_to_drm(adev)->vma_offset_manager,
1816                                adev->need_swiotlb,
1817                                dma_addressing_limited(adev->dev));
1818         if (r) {
1819                 DRM_ERROR("failed initializing buffer object driver(%d).\n", r);
1820                 return r;
1821         }
1822
1823         r = amdgpu_ttm_pools_init(adev);
1824         if (r) {
1825                 DRM_ERROR("failed to init ttm pools(%d).\n", r);
1826                 return r;
1827         }
1828         adev->mman.initialized = true;
1829
1830         /* Initialize VRAM pool with all of VRAM divided into pages */
1831         r = amdgpu_vram_mgr_init(adev);
1832         if (r) {
1833                 DRM_ERROR("Failed initializing VRAM heap.\n");
1834                 return r;
1835         }
1836
1837         /* Change the size here instead of the init above so only lpfn is affected */
1838         amdgpu_ttm_set_buffer_funcs_status(adev, false);
1839 #ifdef CONFIG_64BIT
1840 #ifdef CONFIG_X86
1841         if (adev->gmc.xgmi.connected_to_cpu)
1842                 adev->mman.aper_base_kaddr = ioremap_cache(adev->gmc.aper_base,
1843                                 adev->gmc.visible_vram_size);
1844
1845         else if (adev->gmc.is_app_apu)
1846                 DRM_DEBUG_DRIVER(
1847                         "No need to ioremap when real vram size is 0\n");
1848         else
1849 #endif
1850                 adev->mman.aper_base_kaddr = ioremap_wc(adev->gmc.aper_base,
1851                                 adev->gmc.visible_vram_size);
1852 #endif
1853
1854         /*
1855          *The reserved vram for firmware must be pinned to the specified
1856          *place on the VRAM, so reserve it early.
1857          */
1858         r = amdgpu_ttm_fw_reserve_vram_init(adev);
1859         if (r)
1860                 return r;
1861
1862         /*
1863          *The reserved vram for driver must be pinned to the specified
1864          *place on the VRAM, so reserve it early.
1865          */
1866         r = amdgpu_ttm_drv_reserve_vram_init(adev);
1867         if (r)
1868                 return r;
1869
1870         /*
1871          * only NAVI10 and onwards ASIC support for IP discovery.
1872          * If IP discovery enabled, a block of memory should be
1873          * reserved for IP discovey.
1874          */
1875         if (adev->mman.discovery_bin) {
1876                 r = amdgpu_ttm_reserve_tmr(adev);
1877                 if (r)
1878                         return r;
1879         }
1880
1881         /* allocate memory as required for VGA
1882          * This is used for VGA emulation and pre-OS scanout buffers to
1883          * avoid display artifacts while transitioning between pre-OS
1884          * and driver.
1885          */
1886         if (!adev->gmc.is_app_apu) {
1887                 r = amdgpu_bo_create_kernel_at(adev, 0,
1888                                                adev->mman.stolen_vga_size,
1889                                                &adev->mman.stolen_vga_memory,
1890                                                NULL);
1891                 if (r)
1892                         return r;
1893
1894                 r = amdgpu_bo_create_kernel_at(adev, adev->mman.stolen_vga_size,
1895                                                adev->mman.stolen_extended_size,
1896                                                &adev->mman.stolen_extended_memory,
1897                                                NULL);
1898
1899                 if (r)
1900                         return r;
1901
1902                 r = amdgpu_bo_create_kernel_at(adev,
1903                                                adev->mman.stolen_reserved_offset,
1904                                                adev->mman.stolen_reserved_size,
1905                                                &adev->mman.stolen_reserved_memory,
1906                                                NULL);
1907                 if (r)
1908                         return r;
1909         } else {
1910                 DRM_DEBUG_DRIVER("Skipped stolen memory reservation\n");
1911         }
1912
1913         DRM_INFO("amdgpu: %uM of VRAM memory ready\n",
1914                  (unsigned int)(adev->gmc.real_vram_size / (1024 * 1024)));
1915
1916         /* Compute GTT size, either based on TTM limit
1917          * or whatever the user passed on module init.
1918          */
1919         if (amdgpu_gtt_size == -1)
1920                 gtt_size = ttm_tt_pages_limit() << PAGE_SHIFT;
1921         else
1922                 gtt_size = (uint64_t)amdgpu_gtt_size << 20;
1923
1924         /* Initialize GTT memory pool */
1925         r = amdgpu_gtt_mgr_init(adev, gtt_size);
1926         if (r) {
1927                 DRM_ERROR("Failed initializing GTT heap.\n");
1928                 return r;
1929         }
1930         DRM_INFO("amdgpu: %uM of GTT memory ready.\n",
1931                  (unsigned int)(gtt_size / (1024 * 1024)));
1932
1933         /* Initiailize doorbell pool on PCI BAR */
1934         r = amdgpu_ttm_init_on_chip(adev, AMDGPU_PL_DOORBELL, adev->doorbell.size / PAGE_SIZE);
1935         if (r) {
1936                 DRM_ERROR("Failed initializing doorbell heap.\n");
1937                 return r;
1938         }
1939
1940         /* Create a boorbell page for kernel usages */
1941         r = amdgpu_doorbell_create_kernel_doorbells(adev);
1942         if (r) {
1943                 DRM_ERROR("Failed to initialize kernel doorbells.\n");
1944                 return r;
1945         }
1946
1947         /* Initialize preemptible memory pool */
1948         r = amdgpu_preempt_mgr_init(adev);
1949         if (r) {
1950                 DRM_ERROR("Failed initializing PREEMPT heap.\n");
1951                 return r;
1952         }
1953
1954         /* Initialize various on-chip memory pools */
1955         r = amdgpu_ttm_init_on_chip(adev, AMDGPU_PL_GDS, adev->gds.gds_size);
1956         if (r) {
1957                 DRM_ERROR("Failed initializing GDS heap.\n");
1958                 return r;
1959         }
1960
1961         r = amdgpu_ttm_init_on_chip(adev, AMDGPU_PL_GWS, adev->gds.gws_size);
1962         if (r) {
1963                 DRM_ERROR("Failed initializing gws heap.\n");
1964                 return r;
1965         }
1966
1967         r = amdgpu_ttm_init_on_chip(adev, AMDGPU_PL_OA, adev->gds.oa_size);
1968         if (r) {
1969                 DRM_ERROR("Failed initializing oa heap.\n");
1970                 return r;
1971         }
1972         if (amdgpu_bo_create_kernel(adev, PAGE_SIZE, PAGE_SIZE,
1973                                 AMDGPU_GEM_DOMAIN_GTT,
1974                                 &adev->mman.sdma_access_bo, NULL,
1975                                 &adev->mman.sdma_access_ptr))
1976                 DRM_WARN("Debug VRAM access will use slowpath MM access\n");
1977
1978         return 0;
1979 }
1980
1981 /*
1982  * amdgpu_ttm_fini - De-initialize the TTM memory pools
1983  */
1984 void amdgpu_ttm_fini(struct amdgpu_device *adev)
1985 {
1986         int idx;
1987
1988         if (!adev->mman.initialized)
1989                 return;
1990
1991         amdgpu_ttm_pools_fini(adev);
1992
1993         amdgpu_ttm_training_reserve_vram_fini(adev);
1994         /* return the stolen vga memory back to VRAM */
1995         if (!adev->gmc.is_app_apu) {
1996                 amdgpu_bo_free_kernel(&adev->mman.stolen_vga_memory, NULL, NULL);
1997                 amdgpu_bo_free_kernel(&adev->mman.stolen_extended_memory, NULL, NULL);
1998                 /* return the FW reserved memory back to VRAM */
1999                 amdgpu_bo_free_kernel(&adev->mman.fw_reserved_memory, NULL,
2000                                       NULL);
2001                 if (adev->mman.stolen_reserved_size)
2002                         amdgpu_bo_free_kernel(&adev->mman.stolen_reserved_memory,
2003                                               NULL, NULL);
2004         }
2005         amdgpu_bo_free_kernel(&adev->mman.sdma_access_bo, NULL,
2006                                         &adev->mman.sdma_access_ptr);
2007         amdgpu_ttm_fw_reserve_vram_fini(adev);
2008         amdgpu_ttm_drv_reserve_vram_fini(adev);
2009
2010         if (drm_dev_enter(adev_to_drm(adev), &idx)) {
2011
2012                 if (adev->mman.aper_base_kaddr)
2013                         iounmap(adev->mman.aper_base_kaddr);
2014                 adev->mman.aper_base_kaddr = NULL;
2015
2016                 drm_dev_exit(idx);
2017         }
2018
2019         amdgpu_vram_mgr_fini(adev);
2020         amdgpu_gtt_mgr_fini(adev);
2021         amdgpu_preempt_mgr_fini(adev);
2022         ttm_range_man_fini(&adev->mman.bdev, AMDGPU_PL_GDS);
2023         ttm_range_man_fini(&adev->mman.bdev, AMDGPU_PL_GWS);
2024         ttm_range_man_fini(&adev->mman.bdev, AMDGPU_PL_OA);
2025         ttm_device_fini(&adev->mman.bdev);
2026         adev->mman.initialized = false;
2027         DRM_INFO("amdgpu: ttm finalized\n");
2028 }
2029
2030 /**
2031  * amdgpu_ttm_set_buffer_funcs_status - enable/disable use of buffer functions
2032  *
2033  * @adev: amdgpu_device pointer
2034  * @enable: true when we can use buffer functions.
2035  *
2036  * Enable/disable use of buffer functions during suspend/resume. This should
2037  * only be called at bootup or when userspace isn't running.
2038  */
2039 void amdgpu_ttm_set_buffer_funcs_status(struct amdgpu_device *adev, bool enable)
2040 {
2041         struct ttm_resource_manager *man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM);
2042         uint64_t size;
2043         int r;
2044
2045         if (!adev->mman.initialized || amdgpu_in_reset(adev) ||
2046             adev->mman.buffer_funcs_enabled == enable || adev->gmc.is_app_apu)
2047                 return;
2048
2049         if (enable) {
2050                 struct amdgpu_ring *ring;
2051                 struct drm_gpu_scheduler *sched;
2052
2053                 ring = adev->mman.buffer_funcs_ring;
2054                 sched = &ring->sched;
2055                 r = drm_sched_entity_init(&adev->mman.high_pr,
2056                                           DRM_SCHED_PRIORITY_KERNEL, &sched,
2057                                           1, NULL);
2058                 if (r) {
2059                         DRM_ERROR("Failed setting up TTM BO move entity (%d)\n",
2060                                   r);
2061                         return;
2062                 }
2063
2064                 r = drm_sched_entity_init(&adev->mman.low_pr,
2065                                           DRM_SCHED_PRIORITY_NORMAL, &sched,
2066                                           1, NULL);
2067                 if (r) {
2068                         DRM_ERROR("Failed setting up TTM BO move entity (%d)\n",
2069                                   r);
2070                         goto error_free_entity;
2071                 }
2072         } else {
2073                 drm_sched_entity_destroy(&adev->mman.high_pr);
2074                 drm_sched_entity_destroy(&adev->mman.low_pr);
2075                 dma_fence_put(man->move);
2076                 man->move = NULL;
2077         }
2078
2079         /* this just adjusts TTM size idea, which sets lpfn to the correct value */
2080         if (enable)
2081                 size = adev->gmc.real_vram_size;
2082         else
2083                 size = adev->gmc.visible_vram_size;
2084         man->size = size;
2085         adev->mman.buffer_funcs_enabled = enable;
2086
2087         return;
2088
2089 error_free_entity:
2090         drm_sched_entity_destroy(&adev->mman.high_pr);
2091 }
2092
2093 static int amdgpu_ttm_prepare_job(struct amdgpu_device *adev,
2094                                   bool direct_submit,
2095                                   unsigned int num_dw,
2096                                   struct dma_resv *resv,
2097                                   bool vm_needs_flush,
2098                                   struct amdgpu_job **job,
2099                                   bool delayed)
2100 {
2101         enum amdgpu_ib_pool_type pool = direct_submit ?
2102                 AMDGPU_IB_POOL_DIRECT :
2103                 AMDGPU_IB_POOL_DELAYED;
2104         int r;
2105         struct drm_sched_entity *entity = delayed ? &adev->mman.low_pr :
2106                                                     &adev->mman.high_pr;
2107         r = amdgpu_job_alloc_with_ib(adev, entity,
2108                                      AMDGPU_FENCE_OWNER_UNDEFINED,
2109                                      num_dw * 4, pool, job);
2110         if (r)
2111                 return r;
2112
2113         if (vm_needs_flush) {
2114                 (*job)->vm_pd_addr = amdgpu_gmc_pd_addr(adev->gmc.pdb0_bo ?
2115                                                         adev->gmc.pdb0_bo :
2116                                                         adev->gart.bo);
2117                 (*job)->vm_needs_flush = true;
2118         }
2119         if (!resv)
2120                 return 0;
2121
2122         return drm_sched_job_add_resv_dependencies(&(*job)->base, resv,
2123                                                    DMA_RESV_USAGE_BOOKKEEP);
2124 }
2125
2126 int amdgpu_copy_buffer(struct amdgpu_ring *ring, uint64_t src_offset,
2127                        uint64_t dst_offset, uint32_t byte_count,
2128                        struct dma_resv *resv,
2129                        struct dma_fence **fence, bool direct_submit,
2130                        bool vm_needs_flush, bool tmz)
2131 {
2132         struct amdgpu_device *adev = ring->adev;
2133         unsigned int num_loops, num_dw;
2134         struct amdgpu_job *job;
2135         uint32_t max_bytes;
2136         unsigned int i;
2137         int r;
2138
2139         if (!direct_submit && !ring->sched.ready) {
2140                 DRM_ERROR("Trying to move memory with ring turned off.\n");
2141                 return -EINVAL;
2142         }
2143
2144         max_bytes = adev->mman.buffer_funcs->copy_max_bytes;
2145         num_loops = DIV_ROUND_UP(byte_count, max_bytes);
2146         num_dw = ALIGN(num_loops * adev->mman.buffer_funcs->copy_num_dw, 8);
2147         r = amdgpu_ttm_prepare_job(adev, direct_submit, num_dw,
2148                                    resv, vm_needs_flush, &job, false);
2149         if (r)
2150                 return r;
2151
2152         for (i = 0; i < num_loops; i++) {
2153                 uint32_t cur_size_in_bytes = min(byte_count, max_bytes);
2154
2155                 amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_offset,
2156                                         dst_offset, cur_size_in_bytes, tmz);
2157
2158                 src_offset += cur_size_in_bytes;
2159                 dst_offset += cur_size_in_bytes;
2160                 byte_count -= cur_size_in_bytes;
2161         }
2162
2163         amdgpu_ring_pad_ib(ring, &job->ibs[0]);
2164         WARN_ON(job->ibs[0].length_dw > num_dw);
2165         if (direct_submit)
2166                 r = amdgpu_job_submit_direct(job, ring, fence);
2167         else
2168                 *fence = amdgpu_job_submit(job);
2169         if (r)
2170                 goto error_free;
2171
2172         return r;
2173
2174 error_free:
2175         amdgpu_job_free(job);
2176         DRM_ERROR("Error scheduling IBs (%d)\n", r);
2177         return r;
2178 }
2179
2180 static int amdgpu_ttm_fill_mem(struct amdgpu_ring *ring, uint32_t src_data,
2181                                uint64_t dst_addr, uint32_t byte_count,
2182                                struct dma_resv *resv,
2183                                struct dma_fence **fence,
2184                                bool vm_needs_flush, bool delayed)
2185 {
2186         struct amdgpu_device *adev = ring->adev;
2187         unsigned int num_loops, num_dw;
2188         struct amdgpu_job *job;
2189         uint32_t max_bytes;
2190         unsigned int i;
2191         int r;
2192
2193         max_bytes = adev->mman.buffer_funcs->fill_max_bytes;
2194         num_loops = DIV_ROUND_UP_ULL(byte_count, max_bytes);
2195         num_dw = ALIGN(num_loops * adev->mman.buffer_funcs->fill_num_dw, 8);
2196         r = amdgpu_ttm_prepare_job(adev, false, num_dw, resv, vm_needs_flush,
2197                                    &job, delayed);
2198         if (r)
2199                 return r;
2200
2201         for (i = 0; i < num_loops; i++) {
2202                 uint32_t cur_size = min(byte_count, max_bytes);
2203
2204                 amdgpu_emit_fill_buffer(adev, &job->ibs[0], src_data, dst_addr,
2205                                         cur_size);
2206
2207                 dst_addr += cur_size;
2208                 byte_count -= cur_size;
2209         }
2210
2211         amdgpu_ring_pad_ib(ring, &job->ibs[0]);
2212         WARN_ON(job->ibs[0].length_dw > num_dw);
2213         *fence = amdgpu_job_submit(job);
2214         return 0;
2215 }
2216
2217 int amdgpu_fill_buffer(struct amdgpu_bo *bo,
2218                         uint32_t src_data,
2219                         struct dma_resv *resv,
2220                         struct dma_fence **f,
2221                         bool delayed)
2222 {
2223         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
2224         struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
2225         struct dma_fence *fence = NULL;
2226         struct amdgpu_res_cursor dst;
2227         int r;
2228
2229         if (!adev->mman.buffer_funcs_enabled) {
2230                 DRM_ERROR("Trying to clear memory with ring turned off.\n");
2231                 return -EINVAL;
2232         }
2233
2234         amdgpu_res_first(bo->tbo.resource, 0, amdgpu_bo_size(bo), &dst);
2235
2236         mutex_lock(&adev->mman.gtt_window_lock);
2237         while (dst.remaining) {
2238                 struct dma_fence *next;
2239                 uint64_t cur_size, to;
2240
2241                 /* Never fill more than 256MiB at once to avoid timeouts */
2242                 cur_size = min(dst.size, 256ULL << 20);
2243
2244                 r = amdgpu_ttm_map_buffer(&bo->tbo, bo->tbo.resource, &dst,
2245                                           1, ring, false, &cur_size, &to);
2246                 if (r)
2247                         goto error;
2248
2249                 r = amdgpu_ttm_fill_mem(ring, src_data, to, cur_size, resv,
2250                                         &next, true, delayed);
2251                 if (r)
2252                         goto error;
2253
2254                 dma_fence_put(fence);
2255                 fence = next;
2256
2257                 amdgpu_res_next(&dst, cur_size);
2258         }
2259 error:
2260         mutex_unlock(&adev->mman.gtt_window_lock);
2261         if (f)
2262                 *f = dma_fence_get(fence);
2263         dma_fence_put(fence);
2264         return r;
2265 }
2266
2267 /**
2268  * amdgpu_ttm_evict_resources - evict memory buffers
2269  * @adev: amdgpu device object
2270  * @mem_type: evicted BO's memory type
2271  *
2272  * Evicts all @mem_type buffers on the lru list of the memory type.
2273  *
2274  * Returns:
2275  * 0 for success or a negative error code on failure.
2276  */
2277 int amdgpu_ttm_evict_resources(struct amdgpu_device *adev, int mem_type)
2278 {
2279         struct ttm_resource_manager *man;
2280
2281         switch (mem_type) {
2282         case TTM_PL_VRAM:
2283         case TTM_PL_TT:
2284         case AMDGPU_PL_GWS:
2285         case AMDGPU_PL_GDS:
2286         case AMDGPU_PL_OA:
2287                 man = ttm_manager_type(&adev->mman.bdev, mem_type);
2288                 break;
2289         default:
2290                 DRM_ERROR("Trying to evict invalid memory type\n");
2291                 return -EINVAL;
2292         }
2293
2294         return ttm_resource_manager_evict_all(&adev->mman.bdev, man);
2295 }
2296
2297 #if defined(CONFIG_DEBUG_FS)
2298
2299 static int amdgpu_ttm_page_pool_show(struct seq_file *m, void *unused)
2300 {
2301         struct amdgpu_device *adev = m->private;
2302
2303         return ttm_pool_debugfs(&adev->mman.bdev.pool, m);
2304 }
2305
2306 DEFINE_SHOW_ATTRIBUTE(amdgpu_ttm_page_pool);
2307
2308 /*
2309  * amdgpu_ttm_vram_read - Linear read access to VRAM
2310  *
2311  * Accesses VRAM via MMIO for debugging purposes.
2312  */
2313 static ssize_t amdgpu_ttm_vram_read(struct file *f, char __user *buf,
2314                                     size_t size, loff_t *pos)
2315 {
2316         struct amdgpu_device *adev = file_inode(f)->i_private;
2317         ssize_t result = 0;
2318
2319         if (size & 0x3 || *pos & 0x3)
2320                 return -EINVAL;
2321
2322         if (*pos >= adev->gmc.mc_vram_size)
2323                 return -ENXIO;
2324
2325         size = min(size, (size_t)(adev->gmc.mc_vram_size - *pos));
2326         while (size) {
2327                 size_t bytes = min(size, AMDGPU_TTM_VRAM_MAX_DW_READ * 4);
2328                 uint32_t value[AMDGPU_TTM_VRAM_MAX_DW_READ];
2329
2330                 amdgpu_device_vram_access(adev, *pos, value, bytes, false);
2331                 if (copy_to_user(buf, value, bytes))
2332                         return -EFAULT;
2333
2334                 result += bytes;
2335                 buf += bytes;
2336                 *pos += bytes;
2337                 size -= bytes;
2338         }
2339
2340         return result;
2341 }
2342
2343 /*
2344  * amdgpu_ttm_vram_write - Linear write access to VRAM
2345  *
2346  * Accesses VRAM via MMIO for debugging purposes.
2347  */
2348 static ssize_t amdgpu_ttm_vram_write(struct file *f, const char __user *buf,
2349                                     size_t size, loff_t *pos)
2350 {
2351         struct amdgpu_device *adev = file_inode(f)->i_private;
2352         ssize_t result = 0;
2353         int r;
2354
2355         if (size & 0x3 || *pos & 0x3)
2356                 return -EINVAL;
2357
2358         if (*pos >= adev->gmc.mc_vram_size)
2359                 return -ENXIO;
2360
2361         while (size) {
2362                 uint32_t value;
2363
2364                 if (*pos >= adev->gmc.mc_vram_size)
2365                         return result;
2366
2367                 r = get_user(value, (uint32_t *)buf);
2368                 if (r)
2369                         return r;
2370
2371                 amdgpu_device_mm_access(adev, *pos, &value, 4, true);
2372
2373                 result += 4;
2374                 buf += 4;
2375                 *pos += 4;
2376                 size -= 4;
2377         }
2378
2379         return result;
2380 }
2381
2382 static const struct file_operations amdgpu_ttm_vram_fops = {
2383         .owner = THIS_MODULE,
2384         .read = amdgpu_ttm_vram_read,
2385         .write = amdgpu_ttm_vram_write,
2386         .llseek = default_llseek,
2387 };
2388
2389 /*
2390  * amdgpu_iomem_read - Virtual read access to GPU mapped memory
2391  *
2392  * This function is used to read memory that has been mapped to the
2393  * GPU and the known addresses are not physical addresses but instead
2394  * bus addresses (e.g., what you'd put in an IB or ring buffer).
2395  */
2396 static ssize_t amdgpu_iomem_read(struct file *f, char __user *buf,
2397                                  size_t size, loff_t *pos)
2398 {
2399         struct amdgpu_device *adev = file_inode(f)->i_private;
2400         struct iommu_domain *dom;
2401         ssize_t result = 0;
2402         int r;
2403
2404         /* retrieve the IOMMU domain if any for this device */
2405         dom = iommu_get_domain_for_dev(adev->dev);
2406
2407         while (size) {
2408                 phys_addr_t addr = *pos & PAGE_MASK;
2409                 loff_t off = *pos & ~PAGE_MASK;
2410                 size_t bytes = PAGE_SIZE - off;
2411                 unsigned long pfn;
2412                 struct page *p;
2413                 void *ptr;
2414
2415                 bytes = min(bytes, size);
2416
2417                 /* Translate the bus address to a physical address.  If
2418                  * the domain is NULL it means there is no IOMMU active
2419                  * and the address translation is the identity
2420                  */
2421                 addr = dom ? iommu_iova_to_phys(dom, addr) : addr;
2422
2423                 pfn = addr >> PAGE_SHIFT;
2424                 if (!pfn_valid(pfn))
2425                         return -EPERM;
2426
2427                 p = pfn_to_page(pfn);
2428                 if (p->mapping != adev->mman.bdev.dev_mapping)
2429                         return -EPERM;
2430
2431                 ptr = kmap_local_page(p);
2432                 r = copy_to_user(buf, ptr + off, bytes);
2433                 kunmap_local(ptr);
2434                 if (r)
2435                         return -EFAULT;
2436
2437                 size -= bytes;
2438                 *pos += bytes;
2439                 result += bytes;
2440         }
2441
2442         return result;
2443 }
2444
2445 /*
2446  * amdgpu_iomem_write - Virtual write access to GPU mapped memory
2447  *
2448  * This function is used to write memory that has been mapped to the
2449  * GPU and the known addresses are not physical addresses but instead
2450  * bus addresses (e.g., what you'd put in an IB or ring buffer).
2451  */
2452 static ssize_t amdgpu_iomem_write(struct file *f, const char __user *buf,
2453                                  size_t size, loff_t *pos)
2454 {
2455         struct amdgpu_device *adev = file_inode(f)->i_private;
2456         struct iommu_domain *dom;
2457         ssize_t result = 0;
2458         int r;
2459
2460         dom = iommu_get_domain_for_dev(adev->dev);
2461
2462         while (size) {
2463                 phys_addr_t addr = *pos & PAGE_MASK;
2464                 loff_t off = *pos & ~PAGE_MASK;
2465                 size_t bytes = PAGE_SIZE - off;
2466                 unsigned long pfn;
2467                 struct page *p;
2468                 void *ptr;
2469
2470                 bytes = min(bytes, size);
2471
2472                 addr = dom ? iommu_iova_to_phys(dom, addr) : addr;
2473
2474                 pfn = addr >> PAGE_SHIFT;
2475                 if (!pfn_valid(pfn))
2476                         return -EPERM;
2477
2478                 p = pfn_to_page(pfn);
2479                 if (p->mapping != adev->mman.bdev.dev_mapping)
2480                         return -EPERM;
2481
2482                 ptr = kmap_local_page(p);
2483                 r = copy_from_user(ptr + off, buf, bytes);
2484                 kunmap_local(ptr);
2485                 if (r)
2486                         return -EFAULT;
2487
2488                 size -= bytes;
2489                 *pos += bytes;
2490                 result += bytes;
2491         }
2492
2493         return result;
2494 }
2495
2496 static const struct file_operations amdgpu_ttm_iomem_fops = {
2497         .owner = THIS_MODULE,
2498         .read = amdgpu_iomem_read,
2499         .write = amdgpu_iomem_write,
2500         .llseek = default_llseek
2501 };
2502
2503 #endif
2504
2505 void amdgpu_ttm_debugfs_init(struct amdgpu_device *adev)
2506 {
2507 #if defined(CONFIG_DEBUG_FS)
2508         struct drm_minor *minor = adev_to_drm(adev)->primary;
2509         struct dentry *root = minor->debugfs_root;
2510
2511         debugfs_create_file_size("amdgpu_vram", 0444, root, adev,
2512                                  &amdgpu_ttm_vram_fops, adev->gmc.mc_vram_size);
2513         debugfs_create_file("amdgpu_iomem", 0444, root, adev,
2514                             &amdgpu_ttm_iomem_fops);
2515         debugfs_create_file("ttm_page_pool", 0444, root, adev,
2516                             &amdgpu_ttm_page_pool_fops);
2517         ttm_resource_manager_create_debugfs(ttm_manager_type(&adev->mman.bdev,
2518                                                              TTM_PL_VRAM),
2519                                             root, "amdgpu_vram_mm");
2520         ttm_resource_manager_create_debugfs(ttm_manager_type(&adev->mman.bdev,
2521                                                              TTM_PL_TT),
2522                                             root, "amdgpu_gtt_mm");
2523         ttm_resource_manager_create_debugfs(ttm_manager_type(&adev->mman.bdev,
2524                                                              AMDGPU_PL_GDS),
2525                                             root, "amdgpu_gds_mm");
2526         ttm_resource_manager_create_debugfs(ttm_manager_type(&adev->mman.bdev,
2527                                                              AMDGPU_PL_GWS),
2528                                             root, "amdgpu_gws_mm");
2529         ttm_resource_manager_create_debugfs(ttm_manager_type(&adev->mman.bdev,
2530                                                              AMDGPU_PL_OA),
2531                                             root, "amdgpu_oa_mm");
2532
2533 #endif
2534 }