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

   1 /*
   2  * Copyright 2009 Jerome Glisse.
   3  * All Rights Reserved.
   4  *
   5  * Permission is hereby granted, free of charge, to any person obtaining a
   6  * copy of this software and associated documentation files (the
   7  * "Software"), to deal in the Software without restriction, including
   8  * without limitation the rights to use, copy, modify, merge, publish,
   9  * distribute, sub license, and/or sell copies of the Software, and to
  10  * permit persons to whom the Software is furnished to do so, subject to
  11  * the following conditions:
  12  *
  13  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  14  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  15  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  16  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
  17  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  18  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  19  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  20  *
  21  * The above copyright notice and this permission notice (including the
  22  * next paragraph) shall be included in all copies or substantial portions
  23  * of the Software.
  24  *
  25  */
  26 /*
  27  * Authors:
  28  *    Jerome Glisse <[email protected]>
  29  *    Thomas Hellstrom <thomas-at-tungstengraphics-dot-com>
  30  *    Dave Airlie
  31  */
  32
  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/swiotlb.h>
  42 #include <linux/dma-buf.h>
  43 #include <linux/sizes.h>
  44 #include <linux/module.h>
  45
  46 #include <drm/ttm/ttm_bo_api.h>
  47 #include <drm/ttm/ttm_bo_driver.h>
  48 #include <drm/ttm/ttm_placement.h>
  49 #include <drm/ttm/ttm_range_manager.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_atomfirmware.h"
  60 #include "amdgpu_res_cursor.h"
  61 #include "bif/bif_4_1_d.h"
  62
  63 MODULE_IMPORT_NS(DMA_BUF);
  64
  65 #define AMDGPU_TTM_VRAM_MAX_DW_READ     (size_t)128
  66
  67 static int amdgpu_ttm_backend_bind(struct ttm_device *bdev,
  68                                    struct ttm_tt *ttm,
  69                                    struct ttm_resource *bo_mem);
  70 static void amdgpu_ttm_backend_unbind(struct ttm_device *bdev,
  71                                       struct ttm_tt *ttm);
  72
  73 static int amdgpu_ttm_init_on_chip(struct amdgpu_device *adev,
  74                                     unsigned int type,
  75                                     uint64_t size_in_page)
  76 {
  77         return ttm_range_man_init(&adev->mman.bdev, type,
  78                                   false, size_in_page);
  79 }
  80
  81 /**
  82  * amdgpu_evict_flags - Compute placement flags
  83  *
  84  * @bo: The buffer object to evict
  85  * @placement: Possible destination(s) for evicted BO
  86  *
  87  * Fill in placement data when ttm_bo_evict() is called
  88  */
  89 static void amdgpu_evict_flags(struct ttm_buffer_object *bo,
  90                                 struct ttm_placement *placement)
  91 {
  92         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
  93         struct amdgpu_bo *abo;
  94         static const struct ttm_place placements = {
  95                 .fpfn = 0,
  96                 .lpfn = 0,
  97                 .mem_type = TTM_PL_SYSTEM,
  98                 .flags = 0
  99         };
 100
 101         /* Don't handle scatter gather BOs */
 102         if (bo->type == ttm_bo_type_sg) {
 103                 placement->num_placement = 0;
 104                 placement->num_busy_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->busy_placement = &placements;
 112                 placement->num_placement = 1;
 113                 placement->num_busy_placement = 1;
 114                 return;
 115         }
 116
 117         abo = ttm_to_amdgpu_bo(bo);
 118         if (abo->flags & AMDGPU_AMDKFD_CREATE_SVM_BO) {
 119                 struct dma_fence *fence;
 120                 struct dma_resv *resv = &bo->base._resv;
 121
 122                 rcu_read_lock();
 123                 fence = rcu_dereference(resv->fence_excl);
 124                 if (fence && !fence->ops->signaled)
 125                         dma_fence_enable_sw_signaling(fence);
 126
 127                 placement->num_placement = 0;
 128                 placement->num_busy_placement = 0;
 129                 rcu_read_unlock();
 130                 return;
 131         }
 132
 133         switch (bo->resource->mem_type) {
 134         case AMDGPU_PL_GDS:
 135         case AMDGPU_PL_GWS:
 136         case AMDGPU_PL_OA:
 137                 placement->num_placement = 0;
 138                 placement->num_busy_placement = 0;
 139                 return;
 140
 141         case TTM_PL_VRAM:
 142                 if (!adev->mman.buffer_funcs_enabled) {
 143                         /* Move to system memory */
 144                         amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_CPU);
 145                 } else if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
 146                            !(abo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED) &&
 147                            amdgpu_bo_in_cpu_visible_vram(abo)) {
 148
 149                         /* Try evicting to the CPU inaccessible part of VRAM
 150                          * first, but only set GTT as busy placement, so this
 151                          * BO will be evicted to GTT rather than causing other
 152                          * BOs to be evicted from VRAM
 153                          */
 154                         amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM |
 155                                                         AMDGPU_GEM_DOMAIN_GTT |
 156                                                         AMDGPU_GEM_DOMAIN_CPU);
 157                         abo->placements[0].fpfn = adev->gmc.visible_vram_size >> PAGE_SHIFT;
 158                         abo->placements[0].lpfn = 0;
 159                         abo->placement.busy_placement = &abo->placements[1];
 160                         abo->placement.num_busy_placement = 1;
 161                 } else {
 162                         /* Move to GTT memory */
 163                         amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_GTT |
 164                                                         AMDGPU_GEM_DOMAIN_CPU);
 165                 }
 166                 break;
 167         case TTM_PL_TT:
 168         case AMDGPU_PL_PREEMPT:
 169         default:
 170                 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_CPU);
 171                 break;
 172         }
 173         *placement = abo->placement;
 174 }
 175
 176 /**
 177  * amdgpu_ttm_map_buffer - Map memory into the GART windows
 178  * @bo: buffer object to map
 179  * @mem: memory object to map
 180  * @mm_cur: range to map
 181  * @num_pages: number of pages to map
 182  * @window: which GART window to use
 183  * @ring: DMA ring to use for the copy
 184  * @tmz: if we should setup a TMZ enabled mapping
 185  * @addr: resulting address inside the MC address space
 186  *
 187  * Setup one of the GART windows to access a specific piece of memory or return
 188  * the physical address for local memory.
 189  */
 190 static int amdgpu_ttm_map_buffer(struct ttm_buffer_object *bo,
 191                                  struct ttm_resource *mem,
 192                                  struct amdgpu_res_cursor *mm_cur,
 193                                  unsigned num_pages, unsigned window,
 194                                  struct amdgpu_ring *ring, bool tmz,
 195                                  uint64_t *addr)
 196 {
 197         struct amdgpu_device *adev = ring->adev;
 198         struct amdgpu_job *job;
 199         unsigned num_dw, num_bytes;
 200         struct dma_fence *fence;
 201         uint64_t src_addr, dst_addr;
 202         void *cpu_addr;
 203         uint64_t flags;
 204         unsigned int i;
 205         int r;
 206
 207         BUG_ON(adev->mman.buffer_funcs->copy_max_bytes <
 208                AMDGPU_GTT_MAX_TRANSFER_SIZE * 8);
 209         BUG_ON(mem->mem_type == AMDGPU_PL_PREEMPT);
 210
 211         /* Map only what can't be accessed directly */
 212         if (!tmz && mem->start != AMDGPU_BO_INVALID_OFFSET) {
 213                 *addr = amdgpu_ttm_domain_start(adev, mem->mem_type) +
 214                         mm_cur->start;
 215                 return 0;
 216         }
 217
 218         *addr = adev->gmc.gart_start;
 219         *addr += (u64)window * AMDGPU_GTT_MAX_TRANSFER_SIZE *
 220                 AMDGPU_GPU_PAGE_SIZE;
 221         *addr += mm_cur->start & ~PAGE_MASK;
 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, num_dw * 4 + num_bytes,
 227                                      AMDGPU_IB_POOL_DELAYED, &job);
 228         if (r)
 229                 return r;
 230
 231         src_addr = num_dw * 4;
 232         src_addr += job->ibs[0].gpu_addr;
 233
 234         dst_addr = amdgpu_bo_gpu_offset(adev->gart.bo);
 235         dst_addr += window * AMDGPU_GTT_MAX_TRANSFER_SIZE * 8;
 236         amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_addr,
 237                                 dst_addr, num_bytes, false);
 238
 239         amdgpu_ring_pad_ib(ring, &job->ibs[0]);
 240         WARN_ON(job->ibs[0].length_dw > num_dw);
 241
 242         flags = amdgpu_ttm_tt_pte_flags(adev, bo->ttm, mem);
 243         if (tmz)
 244                 flags |= AMDGPU_PTE_TMZ;
 245
 246         cpu_addr = &job->ibs[0].ptr[num_dw];
 247
 248         if (mem->mem_type == TTM_PL_TT) {
 249                 dma_addr_t *dma_addr;
 250
 251                 dma_addr = &bo->ttm->dma_address[mm_cur->start >> PAGE_SHIFT];
 252                 r = amdgpu_gart_map(adev, 0, num_pages, dma_addr, flags,
 253                                     cpu_addr);
 254                 if (r)
 255                         goto error_free;
 256         } else {
 257                 dma_addr_t dma_address;
 258
 259                 dma_address = mm_cur->start;
 260                 dma_address += adev->vm_manager.vram_base_offset;
 261
 262                 for (i = 0; i < num_pages; ++i) {
 263                         r = amdgpu_gart_map(adev, i << PAGE_SHIFT, 1,
 264                                             &dma_address, flags, cpu_addr);
 265                         if (r)
 266                                 goto error_free;
 267
 268                         dma_address += PAGE_SIZE;
 269                 }
 270         }
 271
 272         r = amdgpu_job_submit(job, &adev->mman.entity,
 273                               AMDGPU_FENCE_OWNER_UNDEFINED, &fence);
 274         if (r)
 275                 goto error_free;
 276
 277         dma_fence_put(fence);
 278
 279         return r;
 280
 281 error_free:
 282         amdgpu_job_free(job);
 283         return r;
 284 }
 285
 286 /**
 287  * amdgpu_ttm_copy_mem_to_mem - Helper function for copy
 288  * @adev: amdgpu device
 289  * @src: buffer/address where to read from
 290  * @dst: buffer/address where to write to
 291  * @size: number of bytes to copy
 292  * @tmz: if a secure copy should be used
 293  * @resv: resv object to sync to
 294  * @f: Returns the last fence if multiple jobs are submitted.
 295  *
 296  * The function copies @size bytes from {src->mem + src->offset} to
 297  * {dst->mem + dst->offset}. src->bo and dst->bo could be same BO for a
 298  * move and different for a BO to BO copy.
 299  *
 300  */
 301 int amdgpu_ttm_copy_mem_to_mem(struct amdgpu_device *adev,
 302                                const struct amdgpu_copy_mem *src,
 303                                const struct amdgpu_copy_mem *dst,
 304                                uint64_t size, bool tmz,
 305                                struct dma_resv *resv,
 306                                struct dma_fence **f)
 307 {
 308         const uint32_t GTT_MAX_BYTES = (AMDGPU_GTT_MAX_TRANSFER_SIZE *
 309                                         AMDGPU_GPU_PAGE_SIZE);
 310
 311         struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
 312         struct amdgpu_res_cursor src_mm, dst_mm;
 313         struct dma_fence *fence = NULL;
 314         int r = 0;
 315
 316         if (!adev->mman.buffer_funcs_enabled) {
 317                 DRM_ERROR("Trying to move memory with ring turned off.\n");
 318                 return -EINVAL;
 319         }
 320
 321         amdgpu_res_first(src->mem, src->offset, size, &src_mm);
 322         amdgpu_res_first(dst->mem, dst->offset, size, &dst_mm);
 323
 324         mutex_lock(&adev->mman.gtt_window_lock);
 325         while (src_mm.remaining) {
 326                 uint32_t src_page_offset = src_mm.start & ~PAGE_MASK;
 327                 uint32_t dst_page_offset = dst_mm.start & ~PAGE_MASK;
 328                 struct dma_fence *next;
 329                 uint32_t cur_size;
 330                 uint64_t from, to;
 331
 332                 /* Copy size cannot exceed GTT_MAX_BYTES. So if src or dst
 333                  * begins at an offset, then adjust the size accordingly
 334                  */
 335                 cur_size = max(src_page_offset, dst_page_offset);
 336                 cur_size = min(min3(src_mm.size, dst_mm.size, size),
 337                                (uint64_t)(GTT_MAX_BYTES - cur_size));
 338
 339                 /* Map src to window 0 and dst to window 1. */
 340                 r = amdgpu_ttm_map_buffer(src->bo, src->mem, &src_mm,
 341                                           PFN_UP(cur_size + src_page_offset),
 342                                           0, ring, tmz, &from);
 343                 if (r)
 344                         goto error;
 345
 346                 r = amdgpu_ttm_map_buffer(dst->bo, dst->mem, &dst_mm,
 347                                           PFN_UP(cur_size + dst_page_offset),
 348                                           1, ring, tmz, &to);
 349                 if (r)
 350                         goto error;
 351
 352                 r = amdgpu_copy_buffer(ring, from, to, cur_size,
 353                                        resv, &next, false, true, tmz);
 354                 if (r)
 355                         goto error;
 356
 357                 dma_fence_put(fence);
 358                 fence = next;
 359
 360                 amdgpu_res_next(&src_mm, cur_size);
 361                 amdgpu_res_next(&dst_mm, cur_size);
 362         }
 363 error:
 364         mutex_unlock(&adev->mman.gtt_window_lock);
 365         if (f)
 366                 *f = dma_fence_get(fence);
 367         dma_fence_put(fence);
 368         return r;
 369 }
 370
 371 /*
 372  * amdgpu_move_blit - Copy an entire buffer to another buffer
 373  *
 374  * This is a helper called by amdgpu_bo_move() and amdgpu_move_vram_ram() to
 375  * help move buffers to and from VRAM.
 376  */
 377 static int amdgpu_move_blit(struct ttm_buffer_object *bo,
 378                             bool evict,
 379                             struct ttm_resource *new_mem,
 380                             struct ttm_resource *old_mem)
 381 {
 382         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
 383         struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
 384         struct amdgpu_copy_mem src, dst;
 385         struct dma_fence *fence = NULL;
 386         int r;
 387
 388         src.bo = bo;
 389         dst.bo = bo;
 390         src.mem = old_mem;
 391         dst.mem = new_mem;
 392         src.offset = 0;
 393         dst.offset = 0;
 394
 395         r = amdgpu_ttm_copy_mem_to_mem(adev, &src, &dst,
 396                                        new_mem->num_pages << PAGE_SHIFT,
 397                                        amdgpu_bo_encrypted(abo),
 398                                        bo->base.resv, &fence);
 399         if (r)
 400                 goto error;
 401
 402         /* clear the space being freed */
 403         if (old_mem->mem_type == TTM_PL_VRAM &&
 404             (abo->flags & AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE)) {
 405                 struct dma_fence *wipe_fence = NULL;
 406
 407                 r = amdgpu_fill_buffer(ttm_to_amdgpu_bo(bo), AMDGPU_POISON,
 408                                        NULL, &wipe_fence);
 409                 if (r) {
 410                         goto error;
 411                 } else if (wipe_fence) {
 412                         dma_fence_put(fence);
 413                         fence = wipe_fence;
 414                 }
 415         }
 416
 417         /* Always block for VM page tables before committing the new location */
 418         if (bo->type == ttm_bo_type_kernel)
 419                 r = ttm_bo_move_accel_cleanup(bo, fence, true, false, new_mem);
 420         else
 421                 r = ttm_bo_move_accel_cleanup(bo, fence, evict, true, new_mem);
 422         dma_fence_put(fence);
 423         return r;
 424
 425 error:
 426         if (fence)
 427                 dma_fence_wait(fence, false);
 428         dma_fence_put(fence);
 429         return r;
 430 }
 431
 432 /*
 433  * amdgpu_mem_visible - Check that memory can be accessed by ttm_bo_move_memcpy
 434  *
 435  * Called by amdgpu_bo_move()
 436  */
 437 static bool amdgpu_mem_visible(struct amdgpu_device *adev,
 438                                struct ttm_resource *mem)
 439 {
 440         uint64_t mem_size = (u64)mem->num_pages << PAGE_SHIFT;
 441         struct amdgpu_res_cursor cursor;
 442
 443         if (mem->mem_type == TTM_PL_SYSTEM ||
 444             mem->mem_type == TTM_PL_TT)
 445                 return true;
 446         if (mem->mem_type != TTM_PL_VRAM)
 447                 return false;
 448
 449         amdgpu_res_first(mem, 0, mem_size, &cursor);
 450
 451         /* ttm_resource_ioremap only supports contiguous memory */
 452         if (cursor.size != mem_size)
 453                 return false;
 454
 455         return cursor.start + cursor.size <= adev->gmc.visible_vram_size;
 456 }
 457
 458 /*
 459  * amdgpu_bo_move - Move a buffer object to a new memory location
 460  *
 461  * Called by ttm_bo_handle_move_mem()
 462  */
 463 static int amdgpu_bo_move(struct ttm_buffer_object *bo, bool evict,
 464                           struct ttm_operation_ctx *ctx,
 465                           struct ttm_resource *new_mem,
 466                           struct ttm_place *hop)
 467 {
 468         struct amdgpu_device *adev;
 469         struct amdgpu_bo *abo;
 470         struct ttm_resource *old_mem = bo->resource;
 471         int r;
 472
 473         if (new_mem->mem_type == TTM_PL_TT ||
 474             new_mem->mem_type == AMDGPU_PL_PREEMPT) {
 475                 r = amdgpu_ttm_backend_bind(bo->bdev, bo->ttm, new_mem);
 476                 if (r)
 477                         return r;
 478         }
 479
 480         /* Can't move a pinned BO */
 481         abo = ttm_to_amdgpu_bo(bo);
 482         if (WARN_ON_ONCE(abo->tbo.pin_count > 0))
 483                 return -EINVAL;
 484
 485         adev = amdgpu_ttm_adev(bo->bdev);
 486
 487         if (old_mem->mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
 488                 ttm_bo_move_null(bo, new_mem);
 489                 goto out;
 490         }
 491         if (old_mem->mem_type == TTM_PL_SYSTEM &&
 492             (new_mem->mem_type == TTM_PL_TT ||
 493              new_mem->mem_type == AMDGPU_PL_PREEMPT)) {
 494                 ttm_bo_move_null(bo, new_mem);
 495                 goto out;
 496         }
 497         if ((old_mem->mem_type == TTM_PL_TT ||
 498              old_mem->mem_type == AMDGPU_PL_PREEMPT) &&
 499             new_mem->mem_type == TTM_PL_SYSTEM) {
 500                 r = ttm_bo_wait_ctx(bo, ctx);
 501                 if (r)
 502                         return r;
 503
 504                 amdgpu_ttm_backend_unbind(bo->bdev, bo->ttm);
 505                 ttm_resource_free(bo, &bo->resource);
 506                 ttm_bo_assign_mem(bo, new_mem);
 507                 goto out;
 508         }
 509
 510         if (old_mem->mem_type == AMDGPU_PL_GDS ||
 511             old_mem->mem_type == AMDGPU_PL_GWS ||
 512             old_mem->mem_type == AMDGPU_PL_OA ||
 513             new_mem->mem_type == AMDGPU_PL_GDS ||
 514             new_mem->mem_type == AMDGPU_PL_GWS ||
 515             new_mem->mem_type == AMDGPU_PL_OA) {
 516                 /* Nothing to save here */
 517                 ttm_bo_move_null(bo, new_mem);
 518                 goto out;
 519         }
 520
 521         if (bo->type == ttm_bo_type_device &&
 522             new_mem->mem_type == TTM_PL_VRAM &&
 523             old_mem->mem_type != TTM_PL_VRAM) {
 524                 /* amdgpu_bo_fault_reserve_notify will re-set this if the CPU
 525                  * accesses the BO after it's moved.
 526                  */
 527                 abo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
 528         }
 529
 530         if (adev->mman.buffer_funcs_enabled) {
 531                 if (((old_mem->mem_type == TTM_PL_SYSTEM &&
 532                       new_mem->mem_type == TTM_PL_VRAM) ||
 533                      (old_mem->mem_type == TTM_PL_VRAM &&
 534                       new_mem->mem_type == TTM_PL_SYSTEM))) {
 535                         hop->fpfn = 0;
 536                         hop->lpfn = 0;
 537                         hop->mem_type = TTM_PL_TT;
 538                         hop->flags = TTM_PL_FLAG_TEMPORARY;
 539                         return -EMULTIHOP;
 540                 }
 541
 542                 r = amdgpu_move_blit(bo, evict, new_mem, old_mem);
 543         } else {
 544                 r = -ENODEV;
 545         }
 546
 547         if (r) {
 548                 /* Check that all memory is CPU accessible */
 549                 if (!amdgpu_mem_visible(adev, old_mem) ||
 550                     !amdgpu_mem_visible(adev, new_mem)) {
 551                         pr_err("Move buffer fallback to memcpy unavailable\n");
 552                         return r;
 553                 }
 554
 555                 r = ttm_bo_move_memcpy(bo, ctx, new_mem);
 556                 if (r)
 557                         return r;
 558         }
 559
 560 out:
 561         /* update statistics */
 562         atomic64_add(bo->base.size, &adev->num_bytes_moved);
 563         amdgpu_bo_move_notify(bo, evict, new_mem);
 564         return 0;
 565 }
 566
 567 /*
 568  * amdgpu_ttm_io_mem_reserve - Reserve a block of memory during a fault
 569  *
 570  * Called by ttm_mem_io_reserve() ultimately via ttm_bo_vm_fault()
 571  */
 572 static int amdgpu_ttm_io_mem_reserve(struct ttm_device *bdev,
 573                                      struct ttm_resource *mem)
 574 {
 575         struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
 576         size_t bus_size = (size_t)mem->num_pages << PAGE_SHIFT;
 577
 578         switch (mem->mem_type) {
 579         case TTM_PL_SYSTEM:
 580                 /* system memory */
 581                 return 0;
 582         case TTM_PL_TT:
 583         case AMDGPU_PL_PREEMPT:
 584                 break;
 585         case TTM_PL_VRAM:
 586                 mem->bus.offset = mem->start << PAGE_SHIFT;
 587                 /* check if it's visible */
 588                 if ((mem->bus.offset + bus_size) > adev->gmc.visible_vram_size)
 589                         return -EINVAL;
 590
 591                 if (adev->mman.aper_base_kaddr &&
 592                     mem->placement & TTM_PL_FLAG_CONTIGUOUS)
 593                         mem->bus.addr = (u8 *)adev->mman.aper_base_kaddr +
 594                                         mem->bus.offset;
 595
 596                 mem->bus.offset += adev->gmc.aper_base;
 597                 mem->bus.is_iomem = true;
 598                 break;
 599         default:
 600                 return -EINVAL;
 601         }
 602         return 0;
 603 }
 604
 605 static unsigned long amdgpu_ttm_io_mem_pfn(struct ttm_buffer_object *bo,
 606                                            unsigned long page_offset)
 607 {
 608         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
 609         struct amdgpu_res_cursor cursor;
 610
 611         amdgpu_res_first(bo->resource, (u64)page_offset << PAGE_SHIFT, 0,
 612                          &cursor);
 613         return (adev->gmc.aper_base + cursor.start) >> PAGE_SHIFT;
 614 }
 615
 616 /**
 617  * amdgpu_ttm_domain_start - Returns GPU start address
 618  * @adev: amdgpu device object
 619  * @type: type of the memory
 620  *
 621  * Returns:
 622  * GPU start address of a memory domain
 623  */
 624
 625 uint64_t amdgpu_ttm_domain_start(struct amdgpu_device *adev, uint32_t type)
 626 {
 627         switch (type) {
 628         case TTM_PL_TT:
 629                 return adev->gmc.gart_start;
 630         case TTM_PL_VRAM:
 631                 return adev->gmc.vram_start;
 632         }
 633
 634         return 0;
 635 }
 636
 637 /*
 638  * TTM backend functions.
 639  */
 640 struct amdgpu_ttm_tt {
 641         struct ttm_tt   ttm;
 642         struct drm_gem_object   *gobj;
 643         u64                     offset;
 644         uint64_t                userptr;
 645         struct task_struct      *usertask;
 646         uint32_t                userflags;
 647         bool                    bound;
 648 #if IS_ENABLED(CONFIG_DRM_AMDGPU_USERPTR)
 649         struct hmm_range        *range;
 650 #endif
 651 };
 652
 653 #ifdef CONFIG_DRM_AMDGPU_USERPTR
 654 /*
 655  * amdgpu_ttm_tt_get_user_pages - get device accessible pages that back user
 656  * memory and start HMM tracking CPU page table update
 657  *
 658  * Calling function must call amdgpu_ttm_tt_userptr_range_done() once and only
 659  * once afterwards to stop HMM tracking
 660  */
 661 int amdgpu_ttm_tt_get_user_pages(struct amdgpu_bo *bo, struct page **pages)
 662 {
 663         struct ttm_tt *ttm = bo->tbo.ttm;
 664         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 665         unsigned long start = gtt->userptr;
 666         struct vm_area_struct *vma;
 667         struct mm_struct *mm;
 668         bool readonly;
 669         int r = 0;
 670
 671         mm = bo->notifier.mm;
 672         if (unlikely(!mm)) {
 673                 DRM_DEBUG_DRIVER("BO is not registered?\n");
 674                 return -EFAULT;
 675         }
 676
 677         /* Another get_user_pages is running at the same time?? */
 678         if (WARN_ON(gtt->range))
 679                 return -EFAULT;
 680
 681         if (!mmget_not_zero(mm)) /* Happens during process shutdown */
 682                 return -ESRCH;
 683
 684         mmap_read_lock(mm);
 685         vma = vma_lookup(mm, start);
 686         if (unlikely(!vma)) {
 687                 r = -EFAULT;
 688                 goto out_unlock;
 689         }
 690         if (unlikely((gtt->userflags & AMDGPU_GEM_USERPTR_ANONONLY) &&
 691                 vma->vm_file)) {
 692                 r = -EPERM;
 693                 goto out_unlock;
 694         }
 695
 696         readonly = amdgpu_ttm_tt_is_readonly(ttm);
 697         r = amdgpu_hmm_range_get_pages(&bo->notifier, mm, pages, start,
 698                                        ttm->num_pages, &gtt->range, readonly,
 699                                        true, NULL);
 700 out_unlock:
 701         mmap_read_unlock(mm);
 702         if (r)
 703                 pr_debug("failed %d to get user pages 0x%lx\n", r, start);
 704
 705         mmput(mm);
 706
 707         return r;
 708 }
 709
 710 /*
 711  * amdgpu_ttm_tt_userptr_range_done - stop HMM track the CPU page table change
 712  * Check if the pages backing this ttm range have been invalidated
 713  *
 714  * Returns: true if pages are still valid
 715  */
 716 bool amdgpu_ttm_tt_get_user_pages_done(struct ttm_tt *ttm)
 717 {
 718         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 719         bool r = false;
 720
 721         if (!gtt || !gtt->userptr)
 722                 return false;
 723
 724         DRM_DEBUG_DRIVER("user_pages_done 0x%llx pages 0x%x\n",
 725                 gtt->userptr, ttm->num_pages);
 726
 727         WARN_ONCE(!gtt->range || !gtt->range->hmm_pfns,
 728                 "No user pages to check\n");
 729
 730         if (gtt->range) {
 731                 /*
 732                  * FIXME: Must always hold notifier_lock for this, and must
 733                  * not ignore the return code.
 734                  */
 735                 r = amdgpu_hmm_range_get_pages_done(gtt->range);
 736                 gtt->range = NULL;
 737         }
 738
 739         return !r;
 740 }
 741 #endif
 742
 743 /*
 744  * amdgpu_ttm_tt_set_user_pages - Copy pages in, putting old pages as necessary.
 745  *
 746  * Called by amdgpu_cs_list_validate(). This creates the page list
 747  * that backs user memory and will ultimately be mapped into the device
 748  * address space.
 749  */
 750 void amdgpu_ttm_tt_set_user_pages(struct ttm_tt *ttm, struct page **pages)
 751 {
 752         unsigned long i;
 753
 754         for (i = 0; i < ttm->num_pages; ++i)
 755                 ttm->pages[i] = pages ? pages[i] : NULL;
 756 }
 757
 758 /*
 759  * amdgpu_ttm_tt_pin_userptr - prepare the sg table with the user pages
 760  *
 761  * Called by amdgpu_ttm_backend_bind()
 762  **/
 763 static int amdgpu_ttm_tt_pin_userptr(struct ttm_device *bdev,
 764                                      struct ttm_tt *ttm)
 765 {
 766         struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
 767         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 768         int write = !(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
 769         enum dma_data_direction direction = write ?
 770                 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
 771         int r;
 772
 773         /* Allocate an SG array and squash pages into it */
 774         r = sg_alloc_table_from_pages(ttm->sg, ttm->pages, ttm->num_pages, 0,
 775                                       (u64)ttm->num_pages << PAGE_SHIFT,
 776                                       GFP_KERNEL);
 777         if (r)
 778                 goto release_sg;
 779
 780         /* Map SG to device */
 781         r = dma_map_sgtable(adev->dev, ttm->sg, direction, 0);
 782         if (r)
 783                 goto release_sg;
 784
 785         /* convert SG to linear array of pages and dma addresses */
 786         drm_prime_sg_to_dma_addr_array(ttm->sg, gtt->ttm.dma_address,
 787                                        ttm->num_pages);
 788
 789         return 0;
 790
 791 release_sg:
 792         kfree(ttm->sg);
 793         ttm->sg = NULL;
 794         return r;
 795 }
 796
 797 /*
 798  * amdgpu_ttm_tt_unpin_userptr - Unpin and unmap userptr pages
 799  */
 800 static void amdgpu_ttm_tt_unpin_userptr(struct ttm_device *bdev,
 801                                         struct ttm_tt *ttm)
 802 {
 803         struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
 804         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 805         int write = !(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
 806         enum dma_data_direction direction = write ?
 807                 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
 808
 809         /* double check that we don't free the table twice */
 810         if (!ttm->sg || !ttm->sg->sgl)
 811                 return;
 812
 813         /* unmap the pages mapped to the device */
 814         dma_unmap_sgtable(adev->dev, ttm->sg, direction, 0);
 815         sg_free_table(ttm->sg);
 816
 817 #if IS_ENABLED(CONFIG_DRM_AMDGPU_USERPTR)
 818         if (gtt->range) {
 819                 unsigned long i;
 820
 821                 for (i = 0; i < ttm->num_pages; i++) {
 822                         if (ttm->pages[i] !=
 823                             hmm_pfn_to_page(gtt->range->hmm_pfns[i]))
 824                                 break;
 825                 }
 826
 827                 WARN((i == ttm->num_pages), "Missing get_user_page_done\n");
 828         }
 829 #endif
 830 }
 831
 832 static int amdgpu_ttm_gart_bind(struct amdgpu_device *adev,
 833                                 struct ttm_buffer_object *tbo,
 834                                 uint64_t flags)
 835 {
 836         struct amdgpu_bo *abo = ttm_to_amdgpu_bo(tbo);
 837         struct ttm_tt *ttm = tbo->ttm;
 838         struct amdgpu_ttm_tt *gtt = (void *)ttm;
 839         int r;
 840
 841         if (amdgpu_bo_encrypted(abo))
 842                 flags |= AMDGPU_PTE_TMZ;
 843
 844         if (abo->flags & AMDGPU_GEM_CREATE_CP_MQD_GFX9) {
 845                 uint64_t page_idx = 1;
 846
 847                 r = amdgpu_gart_bind(adev, gtt->offset, page_idx,
 848                                 gtt->ttm.dma_address, flags);
 849                 if (r)
 850                         goto gart_bind_fail;
 851
 852                 /* The memory type of the first page defaults to UC. Now
 853                  * modify the memory type to NC from the second page of
 854                  * the BO onward.
 855                  */
 856                 flags &= ~AMDGPU_PTE_MTYPE_VG10_MASK;
 857                 flags |= AMDGPU_PTE_MTYPE_VG10(AMDGPU_MTYPE_NC);
 858
 859                 r = amdgpu_gart_bind(adev,
 860                                 gtt->offset + (page_idx << PAGE_SHIFT),
 861                                 ttm->num_pages - page_idx,
 862                                 &(gtt->ttm.dma_address[page_idx]), flags);
 863         } else {
 864                 r = amdgpu_gart_bind(adev, gtt->offset, ttm->num_pages,
 865                                      gtt->ttm.dma_address, flags);
 866         }
 867
 868 gart_bind_fail:
 869         if (r)
 870                 DRM_ERROR("failed to bind %u pages at 0x%08llX\n",
 871                           ttm->num_pages, gtt->offset);
 872
 873         return r;
 874 }
 875
 876 /*
 877  * amdgpu_ttm_backend_bind - Bind GTT memory
 878  *
 879  * Called by ttm_tt_bind() on behalf of ttm_bo_handle_move_mem().
 880  * This handles binding GTT memory to the device address space.
 881  */
 882 static int amdgpu_ttm_backend_bind(struct ttm_device *bdev,
 883                                    struct ttm_tt *ttm,
 884                                    struct ttm_resource *bo_mem)
 885 {
 886         struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
 887         struct amdgpu_ttm_tt *gtt = (void*)ttm;
 888         uint64_t flags;
 889         int r = 0;
 890
 891         if (!bo_mem)
 892                 return -EINVAL;
 893
 894         if (gtt->bound)
 895                 return 0;
 896
 897         if (gtt->userptr) {
 898                 r = amdgpu_ttm_tt_pin_userptr(bdev, ttm);
 899                 if (r) {
 900                         DRM_ERROR("failed to pin userptr\n");
 901                         return r;
 902                 }
 903         } else if (ttm->page_flags & TTM_TT_FLAG_EXTERNAL) {
 904                 if (!ttm->sg) {
 905                         struct dma_buf_attachment *attach;
 906                         struct sg_table *sgt;
 907
 908                         attach = gtt->gobj->import_attach;
 909                         sgt = dma_buf_map_attachment(attach, DMA_BIDIRECTIONAL);
 910                         if (IS_ERR(sgt))
 911                                 return PTR_ERR(sgt);
 912
 913                         ttm->sg = sgt;
 914                 }
 915
 916                 drm_prime_sg_to_dma_addr_array(ttm->sg, gtt->ttm.dma_address,
 917                                                ttm->num_pages);
 918         }
 919
 920         if (!ttm->num_pages) {
 921                 WARN(1, "nothing to bind %u pages for mreg %p back %p!\n",
 922                      ttm->num_pages, bo_mem, ttm);
 923         }
 924
 925         if (bo_mem->mem_type == AMDGPU_PL_GDS ||
 926             bo_mem->mem_type == AMDGPU_PL_GWS ||
 927             bo_mem->mem_type == AMDGPU_PL_OA)
 928                 return -EINVAL;
 929
 930         if (bo_mem->mem_type != TTM_PL_TT ||
 931             !amdgpu_gtt_mgr_has_gart_addr(bo_mem)) {
 932                 gtt->offset = AMDGPU_BO_INVALID_OFFSET;
 933                 return 0;
 934         }
 935
 936         /* compute PTE flags relevant to this BO memory */
 937         flags = amdgpu_ttm_tt_pte_flags(adev, ttm, bo_mem);
 938
 939         /* bind pages into GART page tables */
 940         gtt->offset = (u64)bo_mem->start << PAGE_SHIFT;
 941         r = amdgpu_gart_bind(adev, gtt->offset, ttm->num_pages,
 942                 gtt->ttm.dma_address, flags);
 943
 944         if (r)
 945                 DRM_ERROR("failed to bind %u pages at 0x%08llX\n",
 946                           ttm->num_pages, gtt->offset);
 947         gtt->bound = true;
 948         return r;
 949 }
 950
 951 /*
 952  * amdgpu_ttm_alloc_gart - Make sure buffer object is accessible either
 953  * through AGP or GART aperture.
 954  *
 955  * If bo is accessible through AGP aperture, then use AGP aperture
 956  * to access bo; otherwise allocate logical space in GART aperture
 957  * and map bo to GART aperture.
 958  */
 959 int amdgpu_ttm_alloc_gart(struct ttm_buffer_object *bo)
 960 {
 961         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
 962         struct ttm_operation_ctx ctx = { false, false };
 963         struct amdgpu_ttm_tt *gtt = (void *)bo->ttm;
 964         struct ttm_placement placement;
 965         struct ttm_place placements;
 966         struct ttm_resource *tmp;
 967         uint64_t addr, flags;
 968         int r;
 969
 970         if (bo->resource->start != AMDGPU_BO_INVALID_OFFSET)
 971                 return 0;
 972
 973         addr = amdgpu_gmc_agp_addr(bo);
 974         if (addr != AMDGPU_BO_INVALID_OFFSET) {
 975                 bo->resource->start = addr >> PAGE_SHIFT;
 976                 return 0;
 977         }
 978
 979         /* allocate GART space */
 980         placement.num_placement = 1;
 981         placement.placement = &placements;
 982         placement.num_busy_placement = 1;
 983         placement.busy_placement = &placements;
 984         placements.fpfn = 0;
 985         placements.lpfn = adev->gmc.gart_size >> PAGE_SHIFT;
 986         placements.mem_type = TTM_PL_TT;
 987         placements.flags = bo->resource->placement;
 988
 989         r = ttm_bo_mem_space(bo, &placement, &tmp, &ctx);
 990         if (unlikely(r))
 991                 return r;
 992
 993         /* compute PTE flags for this buffer object */
 994         flags = amdgpu_ttm_tt_pte_flags(adev, bo->ttm, tmp);
 995
 996         /* Bind pages */
 997         gtt->offset = (u64)tmp->start << PAGE_SHIFT;
 998         r = amdgpu_ttm_gart_bind(adev, bo, flags);
 999         if (unlikely(r)) {
1000                 ttm_resource_free(bo, &tmp);
1001                 return r;
1002         }
1003
1004         amdgpu_gart_invalidate_tlb(adev);
1005         ttm_resource_free(bo, &bo->resource);
1006         ttm_bo_assign_mem(bo, tmp);
1007
1008         return 0;
1009 }
1010
1011 /*
1012  * amdgpu_ttm_recover_gart - Rebind GTT pages
1013  *
1014  * Called by amdgpu_gtt_mgr_recover() from amdgpu_device_reset() to
1015  * rebind GTT pages during a GPU reset.
1016  */
1017 int amdgpu_ttm_recover_gart(struct ttm_buffer_object *tbo)
1018 {
1019         struct amdgpu_device *adev = amdgpu_ttm_adev(tbo->bdev);
1020         uint64_t flags;
1021         int r;
1022
1023         if (!tbo->ttm)
1024                 return 0;
1025
1026         flags = amdgpu_ttm_tt_pte_flags(adev, tbo->ttm, tbo->resource);
1027         r = amdgpu_ttm_gart_bind(adev, tbo, flags);
1028
1029         return r;
1030 }
1031
1032 /*
1033  * amdgpu_ttm_backend_unbind - Unbind GTT mapped pages
1034  *
1035  * Called by ttm_tt_unbind() on behalf of ttm_bo_move_ttm() and
1036  * ttm_tt_destroy().
1037  */
1038 static void amdgpu_ttm_backend_unbind(struct ttm_device *bdev,
1039                                       struct ttm_tt *ttm)
1040 {
1041         struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
1042         struct amdgpu_ttm_tt *gtt = (void *)ttm;
1043         int r;
1044
1045         /* if the pages have userptr pinning then clear that first */
1046         if (gtt->userptr) {
1047                 amdgpu_ttm_tt_unpin_userptr(bdev, ttm);
1048         } else if (ttm->sg && gtt->gobj->import_attach) {
1049                 struct dma_buf_attachment *attach;
1050
1051                 attach = gtt->gobj->import_attach;
1052                 dma_buf_unmap_attachment(attach, ttm->sg, DMA_BIDIRECTIONAL);
1053                 ttm->sg = NULL;
1054         }
1055
1056         if (!gtt->bound)
1057                 return;
1058
1059         if (gtt->offset == AMDGPU_BO_INVALID_OFFSET)
1060                 return;
1061
1062         /* unbind shouldn't be done for GDS/GWS/OA in ttm_bo_clean_mm */
1063         r = amdgpu_gart_unbind(adev, gtt->offset, ttm->num_pages);
1064         if (r)
1065                 DRM_ERROR("failed to unbind %u pages at 0x%08llX\n",
1066                           gtt->ttm.num_pages, gtt->offset);
1067         gtt->bound = false;
1068 }
1069
1070 static void amdgpu_ttm_backend_destroy(struct ttm_device *bdev,
1071                                        struct ttm_tt *ttm)
1072 {
1073         struct amdgpu_ttm_tt *gtt = (void *)ttm;
1074
1075         if (gtt->usertask)
1076                 put_task_struct(gtt->usertask);
1077
1078         ttm_tt_fini(&gtt->ttm);
1079         kfree(gtt);
1080 }
1081
1082 /**
1083  * amdgpu_ttm_tt_create - Create a ttm_tt object for a given BO
1084  *
1085  * @bo: The buffer object to create a GTT ttm_tt object around
1086  * @page_flags: Page flags to be added to the ttm_tt object
1087  *
1088  * Called by ttm_tt_create().
1089  */
1090 static struct ttm_tt *amdgpu_ttm_tt_create(struct ttm_buffer_object *bo,
1091                                            uint32_t page_flags)
1092 {
1093         struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
1094         struct amdgpu_ttm_tt *gtt;
1095         enum ttm_caching caching;
1096
1097         gtt = kzalloc(sizeof(struct amdgpu_ttm_tt), GFP_KERNEL);
1098         if (gtt == NULL) {
1099                 return NULL;
1100         }
1101         gtt->gobj = &bo->base;
1102
1103         if (abo->flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
1104                 caching = ttm_write_combined;
1105         else
1106                 caching = ttm_cached;
1107
1108         /* allocate space for the uninitialized page entries */
1109         if (ttm_sg_tt_init(&gtt->ttm, bo, page_flags, caching)) {
1110                 kfree(gtt);
1111                 return NULL;
1112         }
1113         return &gtt->ttm;
1114 }
1115
1116 /*
1117  * amdgpu_ttm_tt_populate - Map GTT pages visible to the device
1118  *
1119  * Map the pages of a ttm_tt object to an address space visible
1120  * to the underlying device.
1121  */
1122 static int amdgpu_ttm_tt_populate(struct ttm_device *bdev,
1123                                   struct ttm_tt *ttm,
1124                                   struct ttm_operation_ctx *ctx)
1125 {
1126         struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
1127         struct amdgpu_ttm_tt *gtt = (void *)ttm;
1128         pgoff_t i;
1129         int ret;
1130
1131         /* user pages are bound by amdgpu_ttm_tt_pin_userptr() */
1132         if (gtt->userptr) {
1133                 ttm->sg = kzalloc(sizeof(struct sg_table), GFP_KERNEL);
1134                 if (!ttm->sg)
1135                         return -ENOMEM;
1136                 return 0;
1137         }
1138
1139         if (ttm->page_flags & TTM_TT_FLAG_EXTERNAL)
1140                 return 0;
1141
1142         ret = ttm_pool_alloc(&adev->mman.bdev.pool, ttm, ctx);
1143         if (ret)
1144                 return ret;
1145
1146         for (i = 0; i < ttm->num_pages; ++i)
1147                 ttm->pages[i]->mapping = bdev->dev_mapping;
1148
1149         return 0;
1150 }
1151
1152 /*
1153  * amdgpu_ttm_tt_unpopulate - unmap GTT pages and unpopulate page arrays
1154  *
1155  * Unmaps pages of a ttm_tt object from the device address space and
1156  * unpopulates the page array backing it.
1157  */
1158 static void amdgpu_ttm_tt_unpopulate(struct ttm_device *bdev,
1159                                      struct ttm_tt *ttm)
1160 {
1161         struct amdgpu_ttm_tt *gtt = (void *)ttm;
1162         struct amdgpu_device *adev;
1163         pgoff_t i;
1164
1165         amdgpu_ttm_backend_unbind(bdev, ttm);
1166
1167         if (gtt->userptr) {
1168                 amdgpu_ttm_tt_set_user_pages(ttm, NULL);
1169                 kfree(ttm->sg);
1170                 ttm->sg = NULL;
1171                 return;
1172         }
1173
1174         if (ttm->page_flags & TTM_TT_FLAG_EXTERNAL)
1175                 return;
1176
1177         for (i = 0; i < ttm->num_pages; ++i)
1178                 ttm->pages[i]->mapping = NULL;
1179
1180         adev = amdgpu_ttm_adev(bdev);
1181         return ttm_pool_free(&adev->mman.bdev.pool, ttm);
1182 }
1183
1184 /**
1185  * amdgpu_ttm_tt_set_userptr - Initialize userptr GTT ttm_tt for the current
1186  * task
1187  *
1188  * @bo: The ttm_buffer_object to bind this userptr to
1189  * @addr:  The address in the current tasks VM space to use
1190  * @flags: Requirements of userptr object.
1191  *
1192  * Called by amdgpu_gem_userptr_ioctl() to bind userptr pages
1193  * to current task
1194  */
1195 int amdgpu_ttm_tt_set_userptr(struct ttm_buffer_object *bo,
1196                               uint64_t addr, uint32_t flags)
1197 {
1198         struct amdgpu_ttm_tt *gtt;
1199
1200         if (!bo->ttm) {
1201                 /* TODO: We want a separate TTM object type for userptrs */
1202                 bo->ttm = amdgpu_ttm_tt_create(bo, 0);
1203                 if (bo->ttm == NULL)
1204                         return -ENOMEM;
1205         }
1206
1207         /* Set TTM_TT_FLAG_EXTERNAL before populate but after create. */
1208         bo->ttm->page_flags |= TTM_TT_FLAG_EXTERNAL;
1209
1210         gtt = (void *)bo->ttm;
1211         gtt->userptr = addr;
1212         gtt->userflags = flags;
1213
1214         if (gtt->usertask)
1215                 put_task_struct(gtt->usertask);
1216         gtt->usertask = current->group_leader;
1217         get_task_struct(gtt->usertask);
1218
1219         return 0;
1220 }
1221
1222 /*
1223  * amdgpu_ttm_tt_get_usermm - Return memory manager for ttm_tt object
1224  */
1225 struct mm_struct *amdgpu_ttm_tt_get_usermm(struct ttm_tt *ttm)
1226 {
1227         struct amdgpu_ttm_tt *gtt = (void *)ttm;
1228
1229         if (gtt == NULL)
1230                 return NULL;
1231
1232         if (gtt->usertask == NULL)
1233                 return NULL;
1234
1235         return gtt->usertask->mm;
1236 }
1237
1238 /*
1239  * amdgpu_ttm_tt_affect_userptr - Determine if a ttm_tt object lays inside an
1240  * address range for the current task.
1241  *
1242  */
1243 bool amdgpu_ttm_tt_affect_userptr(struct ttm_tt *ttm, unsigned long start,
1244                                   unsigned long end, unsigned long *userptr)
1245 {
1246         struct amdgpu_ttm_tt *gtt = (void *)ttm;
1247         unsigned long size;
1248
1249         if (gtt == NULL || !gtt->userptr)
1250                 return false;
1251
1252         /* Return false if no part of the ttm_tt object lies within
1253          * the range
1254          */
1255         size = (unsigned long)gtt->ttm.num_pages * PAGE_SIZE;
1256         if (gtt->userptr > end || gtt->userptr + size <= start)
1257                 return false;
1258
1259         if (userptr)
1260                 *userptr = gtt->userptr;
1261         return true;
1262 }
1263
1264 /*
1265  * amdgpu_ttm_tt_is_userptr - Have the pages backing by userptr?
1266  */
1267 bool amdgpu_ttm_tt_is_userptr(struct ttm_tt *ttm)
1268 {
1269         struct amdgpu_ttm_tt *gtt = (void *)ttm;
1270
1271         if (gtt == NULL || !gtt->userptr)
1272                 return false;
1273
1274         return true;
1275 }
1276
1277 /*
1278  * amdgpu_ttm_tt_is_readonly - Is the ttm_tt object read only?
1279  */
1280 bool amdgpu_ttm_tt_is_readonly(struct ttm_tt *ttm)
1281 {
1282         struct amdgpu_ttm_tt *gtt = (void *)ttm;
1283
1284         if (gtt == NULL)
1285                 return false;
1286
1287         return !!(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
1288 }
1289
1290 /**
1291  * amdgpu_ttm_tt_pde_flags - Compute PDE flags for ttm_tt object
1292  *
1293  * @ttm: The ttm_tt object to compute the flags for
1294  * @mem: The memory registry backing this ttm_tt object
1295  *
1296  * Figure out the flags to use for a VM PDE (Page Directory Entry).
1297  */
1298 uint64_t amdgpu_ttm_tt_pde_flags(struct ttm_tt *ttm, struct ttm_resource *mem)
1299 {
1300         uint64_t flags = 0;
1301
1302         if (mem && mem->mem_type != TTM_PL_SYSTEM)
1303                 flags |= AMDGPU_PTE_VALID;
1304
1305         if (mem && (mem->mem_type == TTM_PL_TT ||
1306                     mem->mem_type == AMDGPU_PL_PREEMPT)) {
1307                 flags |= AMDGPU_PTE_SYSTEM;
1308
1309                 if (ttm->caching == ttm_cached)
1310                         flags |= AMDGPU_PTE_SNOOPED;
1311         }
1312
1313         if (mem && mem->mem_type == TTM_PL_VRAM &&
1314                         mem->bus.caching == ttm_cached)
1315                 flags |= AMDGPU_PTE_SNOOPED;
1316
1317         return flags;
1318 }
1319
1320 /**
1321  * amdgpu_ttm_tt_pte_flags - Compute PTE flags for ttm_tt object
1322  *
1323  * @adev: amdgpu_device pointer
1324  * @ttm: The ttm_tt object to compute the flags for
1325  * @mem: The memory registry backing this ttm_tt object
1326  *
1327  * Figure out the flags to use for a VM PTE (Page Table Entry).
1328  */
1329 uint64_t amdgpu_ttm_tt_pte_flags(struct amdgpu_device *adev, struct ttm_tt *ttm,
1330                                  struct ttm_resource *mem)
1331 {
1332         uint64_t flags = amdgpu_ttm_tt_pde_flags(ttm, mem);
1333
1334         flags |= adev->gart.gart_pte_flags;
1335         flags |= AMDGPU_PTE_READABLE;
1336
1337         if (!amdgpu_ttm_tt_is_readonly(ttm))
1338                 flags |= AMDGPU_PTE_WRITEABLE;
1339
1340         return flags;
1341 }
1342
1343 /*
1344  * amdgpu_ttm_bo_eviction_valuable - Check to see if we can evict a buffer
1345  * object.
1346  *
1347  * Return true if eviction is sensible. Called by ttm_mem_evict_first() on
1348  * behalf of ttm_bo_mem_force_space() which tries to evict buffer objects until
1349  * it can find space for a new object and by ttm_bo_force_list_clean() which is
1350  * used to clean out a memory space.
1351  */
1352 static bool amdgpu_ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
1353                                             const struct ttm_place *place)
1354 {
1355         unsigned long num_pages = bo->resource->num_pages;
1356         struct amdgpu_res_cursor cursor;
1357         struct dma_resv_list *flist;
1358         struct dma_fence *f;
1359         int i;
1360
1361         /* Swapout? */
1362         if (bo->resource->mem_type == TTM_PL_SYSTEM)
1363                 return true;
1364
1365         if (bo->type == ttm_bo_type_kernel &&
1366             !amdgpu_vm_evictable(ttm_to_amdgpu_bo(bo)))
1367                 return false;
1368
1369         /* If bo is a KFD BO, check if the bo belongs to the current process.
1370          * If true, then return false as any KFD process needs all its BOs to
1371          * be resident to run successfully
1372          */
1373         flist = dma_resv_shared_list(bo->base.resv);
1374         if (flist) {
1375                 for (i = 0; i < flist->shared_count; ++i) {
1376                         f = rcu_dereference_protected(flist->shared[i],
1377                                 dma_resv_held(bo->base.resv));
1378                         if (amdkfd_fence_check_mm(f, current->mm))
1379                                 return false;
1380                 }
1381         }
1382
1383         switch (bo->resource->mem_type) {
1384         case AMDGPU_PL_PREEMPT:
1385                 /* Preemptible BOs don't own system resources managed by the
1386                  * driver (pages, VRAM, GART space). They point to resources
1387                  * owned by someone else (e.g. pageable memory in user mode
1388                  * or a DMABuf). They are used in a preemptible context so we
1389                  * can guarantee no deadlocks and good QoS in case of MMU
1390                  * notifiers or DMABuf move notifiers from the resource owner.
1391                  */
1392                 return false;
1393         case TTM_PL_TT:
1394                 if (amdgpu_bo_is_amdgpu_bo(bo) &&
1395                     amdgpu_bo_encrypted(ttm_to_amdgpu_bo(bo)))
1396                         return false;
1397                 return true;
1398
1399         case TTM_PL_VRAM:
1400                 /* Check each drm MM node individually */
1401                 amdgpu_res_first(bo->resource, 0, (u64)num_pages << PAGE_SHIFT,
1402                                  &cursor);
1403                 while (cursor.remaining) {
1404                         if (place->fpfn < PFN_DOWN(cursor.start + cursor.size)
1405                             && !(place->lpfn &&
1406                                  place->lpfn <= PFN_DOWN(cursor.start)))
1407                                 return true;
1408
1409                         amdgpu_res_next(&cursor, cursor.size);
1410                 }
1411                 return false;
1412
1413         default:
1414                 break;
1415         }
1416
1417         return ttm_bo_eviction_valuable(bo, place);
1418 }
1419
1420 static void amdgpu_ttm_vram_mm_access(struct amdgpu_device *adev, loff_t pos,
1421                                       void *buf, size_t size, bool write)
1422 {
1423         while (size) {
1424                 uint64_t aligned_pos = ALIGN_DOWN(pos, 4);
1425                 uint64_t bytes = 4 - (pos & 0x3);
1426                 uint32_t shift = (pos & 0x3) * 8;
1427                 uint32_t mask = 0xffffffff << shift;
1428                 uint32_t value = 0;
1429
1430                 if (size < bytes) {
1431                         mask &= 0xffffffff >> (bytes - size) * 8;
1432                         bytes = size;
1433                 }
1434
1435                 if (mask != 0xffffffff) {
1436                         amdgpu_device_mm_access(adev, aligned_pos, &value, 4, false);
1437                         if (write) {
1438                                 value &= ~mask;
1439                                 value |= (*(uint32_t *)buf << shift) & mask;
1440                                 amdgpu_device_mm_access(adev, aligned_pos, &value, 4, true);
1441                         } else {
1442                                 value = (value & mask) >> shift;
1443                                 memcpy(buf, &value, bytes);
1444                         }
1445                 } else {
1446                         amdgpu_device_mm_access(adev, aligned_pos, buf, 4, write);
1447                 }
1448
1449                 pos += bytes;
1450                 buf += bytes;
1451                 size -= bytes;
1452         }
1453 }
1454
1455 /**
1456  * amdgpu_ttm_access_memory - Read or Write memory that backs a buffer object.
1457  *
1458  * @bo:  The buffer object to read/write
1459  * @offset:  Offset into buffer object
1460  * @buf:  Secondary buffer to write/read from
1461  * @len: Length in bytes of access
1462  * @write:  true if writing
1463  *
1464  * This is used to access VRAM that backs a buffer object via MMIO
1465  * access for debugging purposes.
1466  */
1467 static int amdgpu_ttm_access_memory(struct ttm_buffer_object *bo,
1468                                     unsigned long offset, void *buf, int len,
1469                                     int write)
1470 {
1471         struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
1472         struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev);
1473         struct amdgpu_res_cursor cursor;
1474         int ret = 0;
1475
1476         if (bo->resource->mem_type != TTM_PL_VRAM)
1477                 return -EIO;
1478
1479         amdgpu_res_first(bo->resource, offset, len, &cursor);
1480         while (cursor.remaining) {
1481                 size_t count, size = cursor.size;
1482                 loff_t pos = cursor.start;
1483
1484                 count = amdgpu_device_aper_access(adev, pos, buf, size, write);
1485                 size -= count;
1486                 if (size) {
1487                         /* using MM to access rest vram and handle un-aligned address */
1488                         pos += count;
1489                         buf += count;
1490                         amdgpu_ttm_vram_mm_access(adev, pos, buf, size, write);
1491                 }
1492
1493                 ret += cursor.size;
1494                 buf += cursor.size;
1495                 amdgpu_res_next(&cursor, cursor.size);
1496         }
1497
1498         return ret;
1499 }
1500
1501 static void
1502 amdgpu_bo_delete_mem_notify(struct ttm_buffer_object *bo)
1503 {
1504         amdgpu_bo_move_notify(bo, false, NULL);
1505 }
1506
1507 static struct ttm_device_funcs amdgpu_bo_driver = {
1508         .ttm_tt_create = &amdgpu_ttm_tt_create,
1509         .ttm_tt_populate = &amdgpu_ttm_tt_populate,
1510         .ttm_tt_unpopulate = &amdgpu_ttm_tt_unpopulate,
1511         .ttm_tt_destroy = &amdgpu_ttm_backend_destroy,
1512         .eviction_valuable = amdgpu_ttm_bo_eviction_valuable,
1513         .evict_flags = &amdgpu_evict_flags,
1514         .move = &amdgpu_bo_move,
1515         .delete_mem_notify = &amdgpu_bo_delete_mem_notify,
1516         .release_notify = &amdgpu_bo_release_notify,
1517         .io_mem_reserve = &amdgpu_ttm_io_mem_reserve,
1518         .io_mem_pfn = amdgpu_ttm_io_mem_pfn,
1519         .access_memory = &amdgpu_ttm_access_memory,
1520         .del_from_lru_notify = &amdgpu_vm_del_from_lru_notify
1521 };
1522
1523 /*
1524  * Firmware Reservation functions
1525  */
1526 /**
1527  * amdgpu_ttm_fw_reserve_vram_fini - free fw reserved vram
1528  *
1529  * @adev: amdgpu_device pointer
1530  *
1531  * free fw reserved vram if it has been reserved.
1532  */
1533 static void amdgpu_ttm_fw_reserve_vram_fini(struct amdgpu_device *adev)
1534 {
1535         amdgpu_bo_free_kernel(&adev->mman.fw_vram_usage_reserved_bo,
1536                 NULL, &adev->mman.fw_vram_usage_va);
1537 }
1538
1539 /**
1540  * amdgpu_ttm_fw_reserve_vram_init - create bo vram reservation from fw
1541  *
1542  * @adev: amdgpu_device pointer
1543  *
1544  * create bo vram reservation from fw.
1545  */
1546 static int amdgpu_ttm_fw_reserve_vram_init(struct amdgpu_device *adev)
1547 {
1548         uint64_t vram_size = adev->gmc.visible_vram_size;
1549
1550         adev->mman.fw_vram_usage_va = NULL;
1551         adev->mman.fw_vram_usage_reserved_bo = NULL;
1552
1553         if (adev->mman.fw_vram_usage_size == 0 ||
1554             adev->mman.fw_vram_usage_size > vram_size)
1555                 return 0;
1556
1557         return amdgpu_bo_create_kernel_at(adev,
1558                                           adev->mman.fw_vram_usage_start_offset,
1559                                           adev->mman.fw_vram_usage_size,
1560                                           AMDGPU_GEM_DOMAIN_VRAM,
1561                                           &adev->mman.fw_vram_usage_reserved_bo,
1562                                           &adev->mman.fw_vram_usage_va);
1563 }
1564
1565 /*
1566  * Memoy training reservation functions
1567  */
1568
1569 /**
1570  * amdgpu_ttm_training_reserve_vram_fini - free memory training reserved vram
1571  *
1572  * @adev: amdgpu_device pointer
1573  *
1574  * free memory training reserved vram if it has been reserved.
1575  */
1576 static int amdgpu_ttm_training_reserve_vram_fini(struct amdgpu_device *adev)
1577 {
1578         struct psp_memory_training_context *ctx = &adev->psp.mem_train_ctx;
1579
1580         ctx->init = PSP_MEM_TRAIN_NOT_SUPPORT;
1581         amdgpu_bo_free_kernel(&ctx->c2p_bo, NULL, NULL);
1582         ctx->c2p_bo = NULL;
1583
1584         return 0;
1585 }
1586
1587 static void amdgpu_ttm_training_data_block_init(struct amdgpu_device *adev)
1588 {
1589         struct psp_memory_training_context *ctx = &adev->psp.mem_train_ctx;
1590
1591         memset(ctx, 0, sizeof(*ctx));
1592
1593         ctx->c2p_train_data_offset =
1594                 ALIGN((adev->gmc.mc_vram_size - adev->mman.discovery_tmr_size - SZ_1M), SZ_1M);
1595         ctx->p2c_train_data_offset =
1596                 (adev->gmc.mc_vram_size - GDDR6_MEM_TRAINING_OFFSET);
1597         ctx->train_data_size =
1598                 GDDR6_MEM_TRAINING_DATA_SIZE_IN_BYTES;
1599
1600         DRM_DEBUG("train_data_size:%llx,p2c_train_data_offset:%llx,c2p_train_data_offset:%llx.\n",
1601                         ctx->train_data_size,
1602                         ctx->p2c_train_data_offset,
1603                         ctx->c2p_train_data_offset);
1604 }
1605
1606 /*
1607  * reserve TMR memory at the top of VRAM which holds
1608  * IP Discovery data and is protected by PSP.
1609  */
1610 static int amdgpu_ttm_reserve_tmr(struct amdgpu_device *adev)
1611 {
1612         int ret;
1613         struct psp_memory_training_context *ctx = &adev->psp.mem_train_ctx;
1614         bool mem_train_support = false;
1615
1616         if (!amdgpu_sriov_vf(adev)) {
1617                 if (amdgpu_atomfirmware_mem_training_supported(adev))
1618                         mem_train_support = true;
1619                 else
1620                         DRM_DEBUG("memory training does not support!\n");
1621         }
1622
1623         /*
1624          * Query reserved tmr size through atom firmwareinfo for Sienna_Cichlid and onwards for all
1625          * the use cases (IP discovery/G6 memory training/profiling/diagnostic data.etc)
1626          *
1627          * Otherwise, fallback to legacy approach to check and reserve tmr block for ip
1628          * discovery data and G6 memory training data respectively
1629          */
1630         adev->mman.discovery_tmr_size =
1631                 amdgpu_atomfirmware_get_fw_reserved_fb_size(adev);
1632         if (!adev->mman.discovery_tmr_size)
1633                 adev->mman.discovery_tmr_size = DISCOVERY_TMR_OFFSET;
1634
1635         if (mem_train_support) {
1636                 /* reserve vram for mem train according to TMR location */
1637                 amdgpu_ttm_training_data_block_init(adev);
1638                 ret = amdgpu_bo_create_kernel_at(adev,
1639                                          ctx->c2p_train_data_offset,
1640                                          ctx->train_data_size,
1641                                          AMDGPU_GEM_DOMAIN_VRAM,
1642                                          &ctx->c2p_bo,
1643                                          NULL);
1644                 if (ret) {
1645                         DRM_ERROR("alloc c2p_bo failed(%d)!\n", ret);
1646                         amdgpu_ttm_training_reserve_vram_fini(adev);
1647                         return ret;
1648                 }
1649                 ctx->init = PSP_MEM_TRAIN_RESERVE_SUCCESS;
1650         }
1651
1652         ret = amdgpu_bo_create_kernel_at(adev,
1653                                 adev->gmc.real_vram_size - adev->mman.discovery_tmr_size,
1654                                 adev->mman.discovery_tmr_size,
1655                                 AMDGPU_GEM_DOMAIN_VRAM,
1656                                 &adev->mman.discovery_memory,
1657                                 NULL);
1658         if (ret) {
1659                 DRM_ERROR("alloc tmr failed(%d)!\n", ret);
1660                 amdgpu_bo_free_kernel(&adev->mman.discovery_memory, NULL, NULL);
1661                 return ret;
1662         }
1663
1664         return 0;
1665 }
1666
1667 /*
1668  * amdgpu_ttm_init - Init the memory management (ttm) as well as various
1669  * gtt/vram related fields.
1670  *
1671  * This initializes all of the memory space pools that the TTM layer
1672  * will need such as the GTT space (system memory mapped to the device),
1673  * VRAM (on-board memory), and on-chip memories (GDS, GWS, OA) which
1674  * can be mapped per VMID.
1675  */
1676 int amdgpu_ttm_init(struct amdgpu_device *adev)
1677 {
1678         uint64_t gtt_size;
1679         int r;
1680         u64 vis_vram_limit;
1681
1682         mutex_init(&adev->mman.gtt_window_lock);
1683
1684         /* No others user of address space so set it to 0 */
1685         r = ttm_device_init(&adev->mman.bdev, &amdgpu_bo_driver, adev->dev,
1686                                adev_to_drm(adev)->anon_inode->i_mapping,
1687                                adev_to_drm(adev)->vma_offset_manager,
1688                                adev->need_swiotlb,
1689                                dma_addressing_limited(adev->dev));
1690         if (r) {
1691                 DRM_ERROR("failed initializing buffer object driver(%d).\n", r);
1692                 return r;
1693         }
1694         adev->mman.initialized = true;
1695
1696         /* Initialize VRAM pool with all of VRAM divided into pages */
1697         r = amdgpu_vram_mgr_init(adev);
1698         if (r) {
1699                 DRM_ERROR("Failed initializing VRAM heap.\n");
1700                 return r;
1701         }
1702
1703         /* Reduce size of CPU-visible VRAM if requested */
1704         vis_vram_limit = (u64)amdgpu_vis_vram_limit * 1024 * 1024;
1705         if (amdgpu_vis_vram_limit > 0 &&
1706             vis_vram_limit <= adev->gmc.visible_vram_size)
1707                 adev->gmc.visible_vram_size = vis_vram_limit;
1708
1709         /* Change the size here instead of the init above so only lpfn is affected */
1710         amdgpu_ttm_set_buffer_funcs_status(adev, false);
1711 #ifdef CONFIG_64BIT
1712 #ifdef CONFIG_X86
1713         if (adev->gmc.xgmi.connected_to_cpu)
1714                 adev->mman.aper_base_kaddr = ioremap_cache(adev->gmc.aper_base,
1715                                 adev->gmc.visible_vram_size);
1716
1717         else
1718 #endif
1719                 adev->mman.aper_base_kaddr = ioremap_wc(adev->gmc.aper_base,
1720                                 adev->gmc.visible_vram_size);
1721 #endif
1722
1723         /*
1724          *The reserved vram for firmware must be pinned to the specified
1725          *place on the VRAM, so reserve it early.
1726          */
1727         r = amdgpu_ttm_fw_reserve_vram_init(adev);
1728         if (r) {
1729                 return r;
1730         }
1731
1732         /*
1733          * only NAVI10 and onwards ASIC support for IP discovery.
1734          * If IP discovery enabled, a block of memory should be
1735          * reserved for IP discovey.
1736          */
1737         if (adev->mman.discovery_bin) {
1738                 r = amdgpu_ttm_reserve_tmr(adev);
1739                 if (r)
1740                         return r;
1741         }
1742
1743         /* allocate memory as required for VGA
1744          * This is used for VGA emulation and pre-OS scanout buffers to
1745          * avoid display artifacts while transitioning between pre-OS
1746          * and driver.  */
1747         r = amdgpu_bo_create_kernel_at(adev, 0, adev->mman.stolen_vga_size,
1748                                        AMDGPU_GEM_DOMAIN_VRAM,
1749                                        &adev->mman.stolen_vga_memory,
1750                                        NULL);
1751         if (r)
1752                 return r;
1753         r = amdgpu_bo_create_kernel_at(adev, adev->mman.stolen_vga_size,
1754                                        adev->mman.stolen_extended_size,
1755                                        AMDGPU_GEM_DOMAIN_VRAM,
1756                                        &adev->mman.stolen_extended_memory,
1757                                        NULL);
1758         if (r)
1759                 return r;
1760         r = amdgpu_bo_create_kernel_at(adev, adev->mman.stolen_reserved_offset,
1761                                        adev->mman.stolen_reserved_size,
1762                                        AMDGPU_GEM_DOMAIN_VRAM,
1763                                        &adev->mman.stolen_reserved_memory,
1764                                        NULL);
1765         if (r)
1766                 return r;
1767
1768         DRM_INFO("amdgpu: %uM of VRAM memory ready\n",
1769                  (unsigned) (adev->gmc.real_vram_size / (1024 * 1024)));
1770
1771         /* Compute GTT size, either bsaed on 3/4th the size of RAM size
1772          * or whatever the user passed on module init */
1773         if (amdgpu_gtt_size == -1) {
1774                 struct sysinfo si;
1775
1776                 si_meminfo(&si);
1777                 gtt_size = min(max((AMDGPU_DEFAULT_GTT_SIZE_MB << 20),
1778                                adev->gmc.mc_vram_size),
1779                                ((uint64_t)si.totalram * si.mem_unit * 3/4));
1780         }
1781         else
1782                 gtt_size = (uint64_t)amdgpu_gtt_size << 20;
1783
1784         /* Initialize GTT memory pool */
1785         r = amdgpu_gtt_mgr_init(adev, gtt_size);
1786         if (r) {
1787                 DRM_ERROR("Failed initializing GTT heap.\n");
1788                 return r;
1789         }
1790         DRM_INFO("amdgpu: %uM of GTT memory ready.\n",
1791                  (unsigned)(gtt_size / (1024 * 1024)));
1792
1793         /* Initialize preemptible memory pool */
1794         r = amdgpu_preempt_mgr_init(adev);
1795         if (r) {
1796                 DRM_ERROR("Failed initializing PREEMPT heap.\n");
1797                 return r;
1798         }
1799
1800         /* Initialize various on-chip memory pools */
1801         r = amdgpu_ttm_init_on_chip(adev, AMDGPU_PL_GDS, adev->gds.gds_size);
1802         if (r) {
1803                 DRM_ERROR("Failed initializing GDS heap.\n");
1804                 return r;
1805         }
1806
1807         r = amdgpu_ttm_init_on_chip(adev, AMDGPU_PL_GWS, adev->gds.gws_size);
1808         if (r) {
1809                 DRM_ERROR("Failed initializing gws heap.\n");
1810                 return r;
1811         }
1812
1813         r = amdgpu_ttm_init_on_chip(adev, AMDGPU_PL_OA, adev->gds.oa_size);
1814         if (r) {
1815                 DRM_ERROR("Failed initializing oa heap.\n");
1816                 return r;
1817         }
1818
1819         return 0;
1820 }
1821
1822 /*
1823  * amdgpu_ttm_fini - De-initialize the TTM memory pools
1824  */
1825 void amdgpu_ttm_fini(struct amdgpu_device *adev)
1826 {
1827         if (!adev->mman.initialized)
1828                 return;
1829
1830         amdgpu_ttm_training_reserve_vram_fini(adev);
1831         /* return the stolen vga memory back to VRAM */
1832         amdgpu_bo_free_kernel(&adev->mman.stolen_vga_memory, NULL, NULL);
1833         amdgpu_bo_free_kernel(&adev->mman.stolen_extended_memory, NULL, NULL);
1834         /* return the IP Discovery TMR memory back to VRAM */
1835         amdgpu_bo_free_kernel(&adev->mman.discovery_memory, NULL, NULL);
1836         if (adev->mman.stolen_reserved_size)
1837                 amdgpu_bo_free_kernel(&adev->mman.stolen_reserved_memory,
1838                                       NULL, NULL);
1839         amdgpu_ttm_fw_reserve_vram_fini(adev);
1840
1841         amdgpu_vram_mgr_fini(adev);
1842         amdgpu_gtt_mgr_fini(adev);
1843         amdgpu_preempt_mgr_fini(adev);
1844         ttm_range_man_fini(&adev->mman.bdev, AMDGPU_PL_GDS);
1845         ttm_range_man_fini(&adev->mman.bdev, AMDGPU_PL_GWS);
1846         ttm_range_man_fini(&adev->mman.bdev, AMDGPU_PL_OA);
1847         ttm_device_fini(&adev->mman.bdev);
1848         adev->mman.initialized = false;
1849         DRM_INFO("amdgpu: ttm finalized\n");
1850 }
1851
1852 /**
1853  * amdgpu_ttm_set_buffer_funcs_status - enable/disable use of buffer functions
1854  *
1855  * @adev: amdgpu_device pointer
1856  * @enable: true when we can use buffer functions.
1857  *
1858  * Enable/disable use of buffer functions during suspend/resume. This should
1859  * only be called at bootup or when userspace isn't running.
1860  */
1861 void amdgpu_ttm_set_buffer_funcs_status(struct amdgpu_device *adev, bool enable)
1862 {
1863         struct ttm_resource_manager *man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM);
1864         uint64_t size;
1865         int r;
1866
1867         if (!adev->mman.initialized || amdgpu_in_reset(adev) ||
1868             adev->mman.buffer_funcs_enabled == enable)
1869                 return;
1870
1871         if (enable) {
1872                 struct amdgpu_ring *ring;
1873                 struct drm_gpu_scheduler *sched;
1874
1875                 ring = adev->mman.buffer_funcs_ring;
1876                 sched = &ring->sched;
1877                 r = drm_sched_entity_init(&adev->mman.entity,
1878                                           DRM_SCHED_PRIORITY_KERNEL, &sched,
1879                                           1, NULL);
1880                 if (r) {
1881                         DRM_ERROR("Failed setting up TTM BO move entity (%d)\n",
1882                                   r);
1883                         return;
1884                 }
1885         } else {
1886                 drm_sched_entity_destroy(&adev->mman.entity);
1887                 dma_fence_put(man->move);
1888                 man->move = NULL;
1889         }
1890
1891         /* this just adjusts TTM size idea, which sets lpfn to the correct value */
1892         if (enable)
1893                 size = adev->gmc.real_vram_size;
1894         else
1895                 size = adev->gmc.visible_vram_size;
1896         man->size = size >> PAGE_SHIFT;
1897         adev->mman.buffer_funcs_enabled = enable;
1898 }
1899
1900 int amdgpu_copy_buffer(struct amdgpu_ring *ring, uint64_t src_offset,
1901                        uint64_t dst_offset, uint32_t byte_count,
1902                        struct dma_resv *resv,
1903                        struct dma_fence **fence, bool direct_submit,
1904                        bool vm_needs_flush, bool tmz)
1905 {
1906         enum amdgpu_ib_pool_type pool = direct_submit ? AMDGPU_IB_POOL_DIRECT :
1907                 AMDGPU_IB_POOL_DELAYED;
1908         struct amdgpu_device *adev = ring->adev;
1909         struct amdgpu_job *job;
1910
1911         uint32_t max_bytes;
1912         unsigned num_loops, num_dw;
1913         unsigned i;
1914         int r;
1915
1916         if (direct_submit && !ring->sched.ready) {
1917                 DRM_ERROR("Trying to move memory with ring turned off.\n");
1918                 return -EINVAL;
1919         }
1920
1921         max_bytes = adev->mman.buffer_funcs->copy_max_bytes;
1922         num_loops = DIV_ROUND_UP(byte_count, max_bytes);
1923         num_dw = ALIGN(num_loops * adev->mman.buffer_funcs->copy_num_dw, 8);
1924
1925         r = amdgpu_job_alloc_with_ib(adev, num_dw * 4, pool, &job);
1926         if (r)
1927                 return r;
1928
1929         if (vm_needs_flush) {
1930                 job->vm_pd_addr = amdgpu_gmc_pd_addr(adev->gmc.pdb0_bo ?
1931                                         adev->gmc.pdb0_bo : adev->gart.bo);
1932                 job->vm_needs_flush = true;
1933         }
1934         if (resv) {
1935                 r = amdgpu_sync_resv(adev, &job->sync, resv,
1936                                      AMDGPU_SYNC_ALWAYS,
1937                                      AMDGPU_FENCE_OWNER_UNDEFINED);
1938                 if (r) {
1939                         DRM_ERROR("sync failed (%d).\n", r);
1940                         goto error_free;
1941                 }
1942         }
1943
1944         for (i = 0; i < num_loops; i++) {
1945                 uint32_t cur_size_in_bytes = min(byte_count, max_bytes);
1946
1947                 amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_offset,
1948                                         dst_offset, cur_size_in_bytes, tmz);
1949
1950                 src_offset += cur_size_in_bytes;
1951                 dst_offset += cur_size_in_bytes;
1952                 byte_count -= cur_size_in_bytes;
1953         }
1954
1955         amdgpu_ring_pad_ib(ring, &job->ibs[0]);
1956         WARN_ON(job->ibs[0].length_dw > num_dw);
1957         if (direct_submit)
1958                 r = amdgpu_job_submit_direct(job, ring, fence);
1959         else
1960                 r = amdgpu_job_submit(job, &adev->mman.entity,
1961                                       AMDGPU_FENCE_OWNER_UNDEFINED, fence);
1962         if (r)
1963                 goto error_free;
1964
1965         return r;
1966
1967 error_free:
1968         amdgpu_job_free(job);
1969         DRM_ERROR("Error scheduling IBs (%d)\n", r);
1970         return r;
1971 }
1972
1973 int amdgpu_fill_buffer(struct amdgpu_bo *bo,
1974                        uint32_t src_data,
1975                        struct dma_resv *resv,
1976                        struct dma_fence **fence)
1977 {
1978         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1979         uint32_t max_bytes = adev->mman.buffer_funcs->fill_max_bytes;
1980         struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
1981
1982         struct amdgpu_res_cursor cursor;
1983         unsigned int num_loops, num_dw;
1984         uint64_t num_bytes;
1985
1986         struct amdgpu_job *job;
1987         int r;
1988
1989         if (!adev->mman.buffer_funcs_enabled) {
1990                 DRM_ERROR("Trying to clear memory with ring turned off.\n");
1991                 return -EINVAL;
1992         }
1993
1994         if (bo->tbo.resource->mem_type == AMDGPU_PL_PREEMPT) {
1995                 DRM_ERROR("Trying to clear preemptible memory.\n");
1996                 return -EINVAL;
1997         }
1998
1999         if (bo->tbo.resource->mem_type == TTM_PL_TT) {
2000                 r = amdgpu_ttm_alloc_gart(&bo->tbo);
2001                 if (r)
2002                         return r;
2003         }
2004
2005         num_bytes = bo->tbo.resource->num_pages << PAGE_SHIFT;
2006         num_loops = 0;
2007
2008         amdgpu_res_first(bo->tbo.resource, 0, num_bytes, &cursor);
2009         while (cursor.remaining) {
2010                 num_loops += DIV_ROUND_UP_ULL(cursor.size, max_bytes);
2011                 amdgpu_res_next(&cursor, cursor.size);
2012         }
2013         num_dw = num_loops * adev->mman.buffer_funcs->fill_num_dw;
2014
2015         /* for IB padding */
2016         num_dw += 64;
2017
2018         r = amdgpu_job_alloc_with_ib(adev, num_dw * 4, AMDGPU_IB_POOL_DELAYED,
2019                                      &job);
2020         if (r)
2021                 return r;
2022
2023         if (resv) {
2024                 r = amdgpu_sync_resv(adev, &job->sync, resv,
2025                                      AMDGPU_SYNC_ALWAYS,
2026                                      AMDGPU_FENCE_OWNER_UNDEFINED);
2027                 if (r) {
2028                         DRM_ERROR("sync failed (%d).\n", r);
2029                         goto error_free;
2030                 }
2031         }
2032
2033         amdgpu_res_first(bo->tbo.resource, 0, num_bytes, &cursor);
2034         while (cursor.remaining) {
2035                 uint32_t cur_size = min_t(uint64_t, cursor.size, max_bytes);
2036                 uint64_t dst_addr = cursor.start;
2037
2038                 dst_addr += amdgpu_ttm_domain_start(adev,
2039                                                     bo->tbo.resource->mem_type);
2040                 amdgpu_emit_fill_buffer(adev, &job->ibs[0], src_data, dst_addr,
2041                                         cur_size);
2042
2043                 amdgpu_res_next(&cursor, cur_size);
2044         }
2045
2046         amdgpu_ring_pad_ib(ring, &job->ibs[0]);
2047         WARN_ON(job->ibs[0].length_dw > num_dw);
2048         r = amdgpu_job_submit(job, &adev->mman.entity,
2049                               AMDGPU_FENCE_OWNER_UNDEFINED, fence);
2050         if (r)
2051                 goto error_free;
2052
2053         return 0;
2054
2055 error_free:
2056         amdgpu_job_free(job);
2057         return r;
2058 }
2059
2060 /**
2061  * amdgpu_ttm_evict_resources - evict memory buffers
2062  * @adev: amdgpu device object
2063  * @mem_type: evicted BO's memory type
2064  *
2065  * Evicts all @mem_type buffers on the lru list of the memory type.
2066  *
2067  * Returns:
2068  * 0 for success or a negative error code on failure.
2069  */
2070 int amdgpu_ttm_evict_resources(struct amdgpu_device *adev, int mem_type)
2071 {
2072         struct ttm_resource_manager *man;
2073
2074         switch (mem_type) {
2075         case TTM_PL_VRAM:
2076         case TTM_PL_TT:
2077         case AMDGPU_PL_GWS:
2078         case AMDGPU_PL_GDS:
2079         case AMDGPU_PL_OA:
2080                 man = ttm_manager_type(&adev->mman.bdev, mem_type);
2081                 break;
2082         default:
2083                 DRM_ERROR("Trying to evict invalid memory type\n");
2084                 return -EINVAL;
2085         }
2086
2087         return ttm_resource_manager_evict_all(&adev->mman.bdev, man);
2088 }
2089
2090 #if defined(CONFIG_DEBUG_FS)
2091
2092 static int amdgpu_mm_vram_table_show(struct seq_file *m, void *unused)
2093 {
2094         struct amdgpu_device *adev = (struct amdgpu_device *)m->private;
2095         struct ttm_resource_manager *man = ttm_manager_type(&adev->mman.bdev,
2096                                                             TTM_PL_VRAM);
2097         struct drm_printer p = drm_seq_file_printer(m);
2098
2099         man->func->debug(man, &p);
2100         return 0;
2101 }
2102
2103 static int amdgpu_ttm_page_pool_show(struct seq_file *m, void *unused)
2104 {
2105         struct amdgpu_device *adev = (struct amdgpu_device *)m->private;
2106
2107         return ttm_pool_debugfs(&adev->mman.bdev.pool, m);
2108 }
2109
2110 static int amdgpu_mm_tt_table_show(struct seq_file *m, void *unused)
2111 {
2112         struct amdgpu_device *adev = (struct amdgpu_device *)m->private;
2113         struct ttm_resource_manager *man = ttm_manager_type(&adev->mman.bdev,
2114                                                             TTM_PL_TT);
2115         struct drm_printer p = drm_seq_file_printer(m);
2116
2117         man->func->debug(man, &p);
2118         return 0;
2119 }
2120
2121 static int amdgpu_mm_gds_table_show(struct seq_file *m, void *unused)
2122 {
2123         struct amdgpu_device *adev = (struct amdgpu_device *)m->private;
2124         struct ttm_resource_manager *man = ttm_manager_type(&adev->mman.bdev,
2125                                                             AMDGPU_PL_GDS);
2126         struct drm_printer p = drm_seq_file_printer(m);
2127
2128         man->func->debug(man, &p);
2129         return 0;
2130 }
2131
2132 static int amdgpu_mm_gws_table_show(struct seq_file *m, void *unused)
2133 {
2134         struct amdgpu_device *adev = (struct amdgpu_device *)m->private;
2135         struct ttm_resource_manager *man = ttm_manager_type(&adev->mman.bdev,
2136                                                             AMDGPU_PL_GWS);
2137         struct drm_printer p = drm_seq_file_printer(m);
2138
2139         man->func->debug(man, &p);
2140         return 0;
2141 }
2142
2143 static int amdgpu_mm_oa_table_show(struct seq_file *m, void *unused)
2144 {
2145         struct amdgpu_device *adev = (struct amdgpu_device *)m->private;
2146         struct ttm_resource_manager *man = ttm_manager_type(&adev->mman.bdev,
2147                                                             AMDGPU_PL_OA);
2148         struct drm_printer p = drm_seq_file_printer(m);
2149
2150         man->func->debug(man, &p);
2151         return 0;
2152 }
2153
2154 DEFINE_SHOW_ATTRIBUTE(amdgpu_mm_vram_table);
2155 DEFINE_SHOW_ATTRIBUTE(amdgpu_mm_tt_table);
2156 DEFINE_SHOW_ATTRIBUTE(amdgpu_mm_gds_table);
2157 DEFINE_SHOW_ATTRIBUTE(amdgpu_mm_gws_table);
2158 DEFINE_SHOW_ATTRIBUTE(amdgpu_mm_oa_table);
2159 DEFINE_SHOW_ATTRIBUTE(amdgpu_ttm_page_pool);
2160
2161 /*
2162  * amdgpu_ttm_vram_read - Linear read access to VRAM
2163  *
2164  * Accesses VRAM via MMIO for debugging purposes.
2165  */
2166 static ssize_t amdgpu_ttm_vram_read(struct file *f, char __user *buf,
2167                                     size_t size, loff_t *pos)
2168 {
2169         struct amdgpu_device *adev = file_inode(f)->i_private;
2170         ssize_t result = 0;
2171
2172         if (size & 0x3 || *pos & 0x3)
2173                 return -EINVAL;
2174
2175         if (*pos >= adev->gmc.mc_vram_size)
2176                 return -ENXIO;
2177
2178         size = min(size, (size_t)(adev->gmc.mc_vram_size - *pos));
2179         while (size) {
2180                 size_t bytes = min(size, AMDGPU_TTM_VRAM_MAX_DW_READ * 4);
2181                 uint32_t value[AMDGPU_TTM_VRAM_MAX_DW_READ];
2182
2183                 amdgpu_device_vram_access(adev, *pos, value, bytes, false);
2184                 if (copy_to_user(buf, value, bytes))
2185                         return -EFAULT;
2186
2187                 result += bytes;
2188                 buf += bytes;
2189                 *pos += bytes;
2190                 size -= bytes;
2191         }
2192
2193         return result;
2194 }
2195
2196 /*
2197  * amdgpu_ttm_vram_write - Linear write access to VRAM
2198  *
2199  * Accesses VRAM via MMIO for debugging purposes.
2200  */
2201 static ssize_t amdgpu_ttm_vram_write(struct file *f, const char __user *buf,
2202                                     size_t size, loff_t *pos)
2203 {
2204         struct amdgpu_device *adev = file_inode(f)->i_private;
2205         ssize_t result = 0;
2206         int r;
2207
2208         if (size & 0x3 || *pos & 0x3)
2209                 return -EINVAL;
2210
2211         if (*pos >= adev->gmc.mc_vram_size)
2212                 return -ENXIO;
2213
2214         while (size) {
2215                 uint32_t value;
2216
2217                 if (*pos >= adev->gmc.mc_vram_size)
2218                         return result;
2219
2220                 r = get_user(value, (uint32_t *)buf);
2221                 if (r)
2222                         return r;
2223
2224                 amdgpu_device_mm_access(adev, *pos, &value, 4, true);
2225
2226                 result += 4;
2227                 buf += 4;
2228                 *pos += 4;
2229                 size -= 4;
2230         }
2231
2232         return result;
2233 }
2234
2235 static const struct file_operations amdgpu_ttm_vram_fops = {
2236         .owner = THIS_MODULE,
2237         .read = amdgpu_ttm_vram_read,
2238         .write = amdgpu_ttm_vram_write,
2239         .llseek = default_llseek,
2240 };
2241
2242 /*
2243  * amdgpu_iomem_read - Virtual read access to GPU mapped memory
2244  *
2245  * This function is used to read memory that has been mapped to the
2246  * GPU and the known addresses are not physical addresses but instead
2247  * bus addresses (e.g., what you'd put in an IB or ring buffer).
2248  */
2249 static ssize_t amdgpu_iomem_read(struct file *f, char __user *buf,
2250                                  size_t size, loff_t *pos)
2251 {
2252         struct amdgpu_device *adev = file_inode(f)->i_private;
2253         struct iommu_domain *dom;
2254         ssize_t result = 0;
2255         int r;
2256
2257         /* retrieve the IOMMU domain if any for this device */
2258         dom = iommu_get_domain_for_dev(adev->dev);
2259
2260         while (size) {
2261                 phys_addr_t addr = *pos & PAGE_MASK;
2262                 loff_t off = *pos & ~PAGE_MASK;
2263                 size_t bytes = PAGE_SIZE - off;
2264                 unsigned long pfn;
2265                 struct page *p;
2266                 void *ptr;
2267
2268                 bytes = bytes < size ? bytes : size;
2269
2270                 /* Translate the bus address to a physical address.  If
2271                  * the domain is NULL it means there is no IOMMU active
2272                  * and the address translation is the identity
2273                  */
2274                 addr = dom ? iommu_iova_to_phys(dom, addr) : addr;
2275
2276                 pfn = addr >> PAGE_SHIFT;
2277                 if (!pfn_valid(pfn))
2278                         return -EPERM;
2279
2280                 p = pfn_to_page(pfn);
2281                 if (p->mapping != adev->mman.bdev.dev_mapping)
2282                         return -EPERM;
2283
2284                 ptr = kmap(p);
2285                 r = copy_to_user(buf, ptr + off, bytes);
2286                 kunmap(p);
2287                 if (r)
2288                         return -EFAULT;
2289
2290                 size -= bytes;
2291                 *pos += bytes;
2292                 result += bytes;
2293         }
2294
2295         return result;
2296 }
2297
2298 /*
2299  * amdgpu_iomem_write - Virtual write access to GPU mapped memory
2300  *
2301  * This function is used to write memory that has been mapped to the
2302  * GPU and the known addresses are not physical addresses but instead
2303  * bus addresses (e.g., what you'd put in an IB or ring buffer).
2304  */
2305 static ssize_t amdgpu_iomem_write(struct file *f, const char __user *buf,
2306                                  size_t size, loff_t *pos)
2307 {
2308         struct amdgpu_device *adev = file_inode(f)->i_private;
2309         struct iommu_domain *dom;
2310         ssize_t result = 0;
2311         int r;
2312
2313         dom = iommu_get_domain_for_dev(adev->dev);
2314
2315         while (size) {
2316                 phys_addr_t addr = *pos & PAGE_MASK;
2317                 loff_t off = *pos & ~PAGE_MASK;
2318                 size_t bytes = PAGE_SIZE - off;
2319                 unsigned long pfn;
2320                 struct page *p;
2321                 void *ptr;
2322
2323                 bytes = bytes < size ? bytes : size;
2324
2325                 addr = dom ? iommu_iova_to_phys(dom, addr) : addr;
2326
2327                 pfn = addr >> PAGE_SHIFT;
2328                 if (!pfn_valid(pfn))
2329                         return -EPERM;
2330
2331                 p = pfn_to_page(pfn);
2332                 if (p->mapping != adev->mman.bdev.dev_mapping)
2333                         return -EPERM;
2334
2335                 ptr = kmap(p);
2336                 r = copy_from_user(ptr + off, buf, bytes);
2337                 kunmap(p);
2338                 if (r)
2339                         return -EFAULT;
2340
2341                 size -= bytes;
2342                 *pos += bytes;
2343                 result += bytes;
2344         }
2345
2346         return result;
2347 }
2348
2349 static const struct file_operations amdgpu_ttm_iomem_fops = {
2350         .owner = THIS_MODULE,
2351         .read = amdgpu_iomem_read,
2352         .write = amdgpu_iomem_write,
2353         .llseek = default_llseek
2354 };
2355
2356 #endif
2357
2358 void amdgpu_ttm_debugfs_init(struct amdgpu_device *adev)
2359 {
2360 #if defined(CONFIG_DEBUG_FS)
2361         struct drm_minor *minor = adev_to_drm(adev)->primary;
2362         struct dentry *root = minor->debugfs_root;
2363
2364         debugfs_create_file_size("amdgpu_vram", 0444, root, adev,
2365                                  &amdgpu_ttm_vram_fops, adev->gmc.mc_vram_size);
2366         debugfs_create_file("amdgpu_iomem", 0444, root, adev,
2367                             &amdgpu_ttm_iomem_fops);
2368         debugfs_create_file("amdgpu_vram_mm", 0444, root, adev,
2369                             &amdgpu_mm_vram_table_fops);
2370         debugfs_create_file("amdgpu_gtt_mm", 0444, root, adev,
2371                             &amdgpu_mm_tt_table_fops);
2372         debugfs_create_file("amdgpu_gds_mm", 0444, root, adev,
2373                             &amdgpu_mm_gds_table_fops);
2374         debugfs_create_file("amdgpu_gws_mm", 0444, root, adev,
2375                             &amdgpu_mm_gws_table_fops);
2376         debugfs_create_file("amdgpu_oa_mm", 0444, root, adev,
2377                             &amdgpu_mm_oa_table_fops);
2378         debugfs_create_file("ttm_page_pool", 0444, root, adev,
2379                             &amdgpu_ttm_page_pool_fops);
2380 #endif
2381 }