drivers/gpu/drm/radeon/radeon_ttm.c

   1 /*
   2  * Copyright 2009 Jerome Glisse.
   3  * All Rights Reserved.
   4  *
   5  * Permission is hereby granted, free of charge, to any person obtaining a
   6  * copy of this software and associated documentation files (the
   7  * "Software"), to deal in the Software without restriction, including
   8  * without limitation the rights to use, copy, modify, merge, publish,
   9  * distribute, sub license, and/or sell copies of the Software, and to
  10  * permit persons to whom the Software is furnished to do so, subject to
  11  * the following conditions:
  12  *
  13  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  14  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  15  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  16  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
  17  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  18  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  19  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  20  *
  21  * The above copyright notice and this permission notice (including the
  22  * next paragraph) shall be included in all copies or substantial portions
  23  * of the Software.
  24  *
  25  */
  26 /*
  27  * Authors:
  28  *    Jerome Glisse <glisse@freedesktop.org>
  29  *    Thomas Hellstrom <thomas-at-tungstengraphics-dot-com>
  30  *    Dave Airlie
  31  */
  32
  33 #include <linux/dma-mapping.h>
  34 #include <linux/pagemap.h>
  35 #include <linux/pci.h>
  36 #include <linux/seq_file.h>
  37 #include <linux/slab.h>
  38 #include <linux/swap.h>
  39 #include <linux/swiotlb.h>
  40
  41 #include <drm/drm_agpsupport.h>
  42 #include <drm/drm_debugfs.h>
  43 #include <drm/drm_device.h>
  44 #include <drm/drm_file.h>
  45 #include <drm/drm_prime.h>
  46 #include <drm/radeon_drm.h>
  47 #include <drm/ttm/ttm_bo_api.h>
  48 #include <drm/ttm/ttm_bo_driver.h>
  49 #include <drm/ttm/ttm_module.h>
  50 #include <drm/ttm/ttm_placement.h>
  51
  52 #include "radeon_reg.h"
  53 #include "radeon.h"
  54 #include "radeon_ttm.h"
  55
  56 static int radeon_ttm_debugfs_init(struct radeon_device *rdev);
  57 static void radeon_ttm_debugfs_fini(struct radeon_device *rdev);
  58
  59 static int radeon_ttm_tt_bind(struct ttm_bo_device *bdev,
  60                               struct ttm_tt *ttm,
  61                               struct ttm_resource *bo_mem);
  62 static void radeon_ttm_tt_unbind(struct ttm_bo_device *bdev,
  63                                  struct ttm_tt *ttm);
  64
  65 struct radeon_device *radeon_get_rdev(struct ttm_bo_device *bdev)
  66 {
  67         struct radeon_mman *mman;
  68         struct radeon_device *rdev;
  69
  70         mman = container_of(bdev, struct radeon_mman, bdev);
  71         rdev = container_of(mman, struct radeon_device, mman);
  72         return rdev;
  73 }
  74
  75 static int radeon_ttm_init_vram(struct radeon_device *rdev)
  76 {
  77         return ttm_range_man_init(&rdev->mman.bdev, TTM_PL_VRAM,
  78                                   false, rdev->mc.real_vram_size >> PAGE_SHIFT);
  79 }
  80
  81 static int radeon_ttm_init_gtt(struct radeon_device *rdev)
  82 {
  83         return ttm_range_man_init(&rdev->mman.bdev, TTM_PL_TT,
  84                                   true, rdev->mc.gtt_size >> PAGE_SHIFT);
  85 }
  86
  87 static void radeon_evict_flags(struct ttm_buffer_object *bo,
  88                                 struct ttm_placement *placement)
  89 {
  90         static const struct ttm_place placements = {
  91                 .fpfn = 0,
  92                 .lpfn = 0,
  93                 .mem_type = TTM_PL_SYSTEM,
  94                 .flags = 0
  95         };
  96
  97         struct radeon_bo *rbo;
  98
  99         if (!radeon_ttm_bo_is_radeon_bo(bo)) {
 100                 placement->placement = &placements;
 101                 placement->busy_placement = &placements;
 102                 placement->num_placement = 1;
 103                 placement->num_busy_placement = 1;
 104                 return;
 105         }
 106         rbo = container_of(bo, struct radeon_bo, tbo);
 107         switch (bo->mem.mem_type) {
 108         case TTM_PL_VRAM:
 109                 if (rbo->rdev->ring[radeon_copy_ring_index(rbo->rdev)].ready == false)
 110                         radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_CPU);
 111                 else if (rbo->rdev->mc.visible_vram_size < rbo->rdev->mc.real_vram_size &&
 112                          bo->mem.start < (rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT)) {
 113                         unsigned fpfn = rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
 114                         int i;
 115
 116                         /* Try evicting to the CPU inaccessible part of VRAM
 117                          * first, but only set GTT as busy placement, so this
 118                          * BO will be evicted to GTT rather than causing other
 119                          * BOs to be evicted from VRAM
 120                          */
 121                         radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_VRAM |
 122                                                          RADEON_GEM_DOMAIN_GTT);
 123                         rbo->placement.num_busy_placement = 0;
 124                         for (i = 0; i < rbo->placement.num_placement; i++) {
 125                                 if (rbo->placements[i].mem_type == TTM_PL_VRAM) {
 126                                         if (rbo->placements[i].fpfn < fpfn)
 127                                                 rbo->placements[i].fpfn = fpfn;
 128                                 } else {
 129                                         rbo->placement.busy_placement =
 130                                                 &rbo->placements[i];
 131                                         rbo->placement.num_busy_placement = 1;
 132                                 }
 133                         }
 134                 } else
 135                         radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_GTT);
 136                 break;
 137         case TTM_PL_TT:
 138         default:
 139                 radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_CPU);
 140         }
 141         *placement = rbo->placement;
 142 }
 143
 144 static int radeon_verify_access(struct ttm_buffer_object *bo, struct file *filp)
 145 {
 146         struct radeon_bo *rbo = container_of(bo, struct radeon_bo, tbo);
 147         struct radeon_device *rdev = radeon_get_rdev(bo->bdev);
 148
 149         if (radeon_ttm_tt_has_userptr(rdev, bo->ttm))
 150                 return -EPERM;
 151         return drm_vma_node_verify_access(&rbo->tbo.base.vma_node,
 152                                           filp->private_data);
 153 }
 154
 155 static int radeon_move_blit(struct ttm_buffer_object *bo,
 156                         bool evict,
 157                         struct ttm_resource *new_mem,
 158                         struct ttm_resource *old_mem)
 159 {
 160         struct radeon_device *rdev;
 161         uint64_t old_start, new_start;
 162         struct radeon_fence *fence;
 163         unsigned num_pages;
 164         int r, ridx;
 165
 166         rdev = radeon_get_rdev(bo->bdev);
 167         ridx = radeon_copy_ring_index(rdev);
 168         old_start = (u64)old_mem->start << PAGE_SHIFT;
 169         new_start = (u64)new_mem->start << PAGE_SHIFT;
 170
 171         switch (old_mem->mem_type) {
 172         case TTM_PL_VRAM:
 173                 old_start += rdev->mc.vram_start;
 174                 break;
 175         case TTM_PL_TT:
 176                 old_start += rdev->mc.gtt_start;
 177                 break;
 178         default:
 179                 DRM_ERROR("Unknown placement %d\n", old_mem->mem_type);
 180                 return -EINVAL;
 181         }
 182         switch (new_mem->mem_type) {
 183         case TTM_PL_VRAM:
 184                 new_start += rdev->mc.vram_start;
 185                 break;
 186         case TTM_PL_TT:
 187                 new_start += rdev->mc.gtt_start;
 188                 break;
 189         default:
 190                 DRM_ERROR("Unknown placement %d\n", old_mem->mem_type);
 191                 return -EINVAL;
 192         }
 193         if (!rdev->ring[ridx].ready) {
 194                 DRM_ERROR("Trying to move memory with ring turned off.\n");
 195                 return -EINVAL;
 196         }
 197
 198         BUILD_BUG_ON((PAGE_SIZE % RADEON_GPU_PAGE_SIZE) != 0);
 199
 200         num_pages = new_mem->num_pages * (PAGE_SIZE / RADEON_GPU_PAGE_SIZE);
 201         fence = radeon_copy(rdev, old_start, new_start, num_pages, bo->base.resv);
 202         if (IS_ERR(fence))
 203                 return PTR_ERR(fence);
 204
 205         r = ttm_bo_move_accel_cleanup(bo, &fence->base, evict, false, new_mem);
 206         radeon_fence_unref(&fence);
 207         return r;
 208 }
 209
 210 static int radeon_bo_move(struct ttm_buffer_object *bo, bool evict,
 211                           struct ttm_operation_ctx *ctx,
 212                           struct ttm_resource *new_mem,
 213                           struct ttm_place *hop)
 214 {
 215         struct radeon_device *rdev;
 216         struct radeon_bo *rbo;
 217         struct ttm_resource *old_mem = &bo->mem;
 218         int r;
 219
 220         if (new_mem->mem_type == TTM_PL_TT) {
 221                 r = radeon_ttm_tt_bind(bo->bdev, bo->ttm, new_mem);
 222                 if (r)
 223                         return r;
 224         }
 225
 226         r = ttm_bo_wait_ctx(bo, ctx);
 227         if (r)
 228                 return r;
 229
 230         /* Can't move a pinned BO */
 231         rbo = container_of(bo, struct radeon_bo, tbo);
 232         if (WARN_ON_ONCE(rbo->tbo.pin_count > 0))
 233                 return -EINVAL;
 234
 235         rdev = radeon_get_rdev(bo->bdev);
 236         if (old_mem->mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
 237                 ttm_bo_move_null(bo, new_mem);
 238                 goto out;
 239         }
 240         if (old_mem->mem_type == TTM_PL_SYSTEM &&
 241             new_mem->mem_type == TTM_PL_TT) {
 242                 ttm_bo_move_null(bo, new_mem);
 243                 goto out;
 244         }
 245
 246         if (old_mem->mem_type == TTM_PL_TT &&
 247             new_mem->mem_type == TTM_PL_SYSTEM) {
 248                 radeon_ttm_tt_unbind(bo->bdev, bo->ttm);
 249                 ttm_resource_free(bo, &bo->mem);
 250                 ttm_bo_assign_mem(bo, new_mem);
 251                 goto out;
 252         }
 253         if (rdev->ring[radeon_copy_ring_index(rdev)].ready &&
 254             rdev->asic->copy.copy != NULL) {
 255                 if ((old_mem->mem_type == TTM_PL_SYSTEM &&
 256                      new_mem->mem_type == TTM_PL_VRAM) ||
 257                     (old_mem->mem_type == TTM_PL_VRAM &&
 258                      new_mem->mem_type == TTM_PL_SYSTEM)) {
 259                         hop->fpfn = 0;
 260                         hop->lpfn = 0;
 261                         hop->mem_type = TTM_PL_TT;
 262                         hop->flags = 0;
 263                         return -EMULTIHOP;
 264                 }
 265
 266                 r = radeon_move_blit(bo, evict, new_mem, old_mem);
 267         } else {
 268                 r = -ENODEV;
 269         }
 270
 271         if (r) {
 272                 r = ttm_bo_move_memcpy(bo, ctx, new_mem);
 273                 if (r)
 274                         return r;
 275         }
 276
 277 out:
 278         /* update statistics */
 279         atomic64_add((u64)bo->num_pages << PAGE_SHIFT, &rdev->num_bytes_moved);
 280         radeon_bo_move_notify(bo, evict, new_mem);
 281         return 0;
 282 }
 283
 284 static int radeon_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_resource *mem)
 285 {
 286         struct radeon_device *rdev = radeon_get_rdev(bdev);
 287         size_t bus_size = (size_t)mem->num_pages << PAGE_SHIFT;
 288
 289         switch (mem->mem_type) {
 290         case TTM_PL_SYSTEM:
 291                 /* system memory */
 292                 return 0;
 293         case TTM_PL_TT:
 294 #if IS_ENABLED(CONFIG_AGP)
 295                 if (rdev->flags & RADEON_IS_AGP) {
 296                         /* RADEON_IS_AGP is set only if AGP is active */
 297                         mem->bus.offset = (mem->start << PAGE_SHIFT) +
 298                                 rdev->mc.agp_base;
 299                         mem->bus.is_iomem = !rdev->ddev->agp->cant_use_aperture;
 300                         mem->bus.caching = ttm_write_combined;
 301                 }
 302 #endif
 303                 break;
 304         case TTM_PL_VRAM:
 305                 mem->bus.offset = mem->start << PAGE_SHIFT;
 306                 /* check if it's visible */
 307                 if ((mem->bus.offset + bus_size) > rdev->mc.visible_vram_size)
 308                         return -EINVAL;
 309                 mem->bus.offset += rdev->mc.aper_base;
 310                 mem->bus.is_iomem = true;
 311                 mem->bus.caching = ttm_write_combined;
 312 #ifdef __alpha__
 313                 /*
 314                  * Alpha: use bus.addr to hold the ioremap() return,
 315                  * so we can modify bus.base below.
 316                  */
 317                 mem->bus.addr = ioremap_wc(mem->bus.offset, bus_size);
 318                 if (!mem->bus.addr)
 319                         return -ENOMEM;
 320
 321                 /*
 322                  * Alpha: Use just the bus offset plus
 323                  * the hose/domain memory base for bus.base.
 324                  * It then can be used to build PTEs for VRAM
 325                  * access, as done in ttm_bo_vm_fault().
 326                  */
 327                 mem->bus.offset = (mem->bus.offset & 0x0ffffffffUL) +
 328                         rdev->ddev->hose->dense_mem_base;
 329 #endif
 330                 break;
 331         default:
 332                 return -EINVAL;
 333         }
 334         return 0;
 335 }
 336
 337 /*
 338  * TTM backend functions.
 339  */
 340 struct radeon_ttm_tt {
 341         struct ttm_tt           ttm;
 342         u64                             offset;
 343
 344         uint64_t                        userptr;
 345         struct mm_struct                *usermm;
 346         uint32_t                        userflags;
 347         bool bound;
 348 };
 349
 350 /* prepare the sg table with the user pages */
 351 static int radeon_ttm_tt_pin_userptr(struct ttm_bo_device *bdev, struct ttm_tt *ttm)
 352 {
 353         struct radeon_device *rdev = radeon_get_rdev(bdev);
 354         struct radeon_ttm_tt *gtt = (void *)ttm;
 355         unsigned pinned = 0;
 356         int r;
 357
 358         int write = !(gtt->userflags & RADEON_GEM_USERPTR_READONLY);
 359         enum dma_data_direction direction = write ?
 360                 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
 361
 362         if (current->mm != gtt->usermm)
 363                 return -EPERM;
 364
 365         if (gtt->userflags & RADEON_GEM_USERPTR_ANONONLY) {
 366                 /* check that we only pin down anonymous memory
 367                    to prevent problems with writeback */
 368                 unsigned long end = gtt->userptr + ttm->num_pages * PAGE_SIZE;
 369                 struct vm_area_struct *vma;
 370                 vma = find_vma(gtt->usermm, gtt->userptr);
 371                 if (!vma || vma->vm_file || vma->vm_end < end)
 372                         return -EPERM;
 373         }
 374
 375         do {
 376                 unsigned num_pages = ttm->num_pages - pinned;
 377                 uint64_t userptr = gtt->userptr + pinned * PAGE_SIZE;
 378                 struct page **pages = ttm->pages + pinned;
 379
 380                 r = get_user_pages(userptr, num_pages, write ? FOLL_WRITE : 0,
 381                                    pages, NULL);
 382                 if (r < 0)
 383                         goto release_pages;
 384
 385                 pinned += r;
 386
 387         } while (pinned < ttm->num_pages);
 388
 389         r = sg_alloc_table_from_pages(ttm->sg, ttm->pages, ttm->num_pages, 0,
 390                                       ttm->num_pages << PAGE_SHIFT,
 391                                       GFP_KERNEL);
 392         if (r)
 393                 goto release_sg;
 394
 395         r = dma_map_sgtable(rdev->dev, ttm->sg, direction, 0);
 396         if (r)
 397                 goto release_sg;
 398
 399         drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
 400                                          gtt->ttm.dma_address, ttm->num_pages);
 401
 402         return 0;
 403
 404 release_sg:
 405         kfree(ttm->sg);
 406
 407 release_pages:
 408         release_pages(ttm->pages, pinned);
 409         return r;
 410 }
 411
 412 static void radeon_ttm_tt_unpin_userptr(struct ttm_bo_device *bdev, struct ttm_tt *ttm)
 413 {
 414         struct radeon_device *rdev = radeon_get_rdev(bdev);
 415         struct radeon_ttm_tt *gtt = (void *)ttm;
 416         struct sg_page_iter sg_iter;
 417
 418         int write = !(gtt->userflags & RADEON_GEM_USERPTR_READONLY);
 419         enum dma_data_direction direction = write ?
 420                 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
 421
 422         /* double check that we don't free the table twice */
 423         if (!ttm->sg->sgl)
 424                 return;
 425
 426         /* free the sg table and pages again */
 427         dma_unmap_sgtable(rdev->dev, ttm->sg, direction, 0);
 428
 429         for_each_sgtable_page(ttm->sg, &sg_iter, 0) {
 430                 struct page *page = sg_page_iter_page(&sg_iter);
 431                 if (!(gtt->userflags & RADEON_GEM_USERPTR_READONLY))
 432                         set_page_dirty(page);
 433
 434                 mark_page_accessed(page);
 435                 put_page(page);
 436         }
 437
 438         sg_free_table(ttm->sg);
 439 }
 440
 441 static bool radeon_ttm_backend_is_bound(struct ttm_tt *ttm)
 442 {
 443         struct radeon_ttm_tt *gtt = (void*)ttm;
 444
 445         return (gtt->bound);
 446 }
 447
 448 static int radeon_ttm_backend_bind(struct ttm_bo_device *bdev,
 449                                    struct ttm_tt *ttm,
 450                                    struct ttm_resource *bo_mem)
 451 {
 452         struct radeon_ttm_tt *gtt = (void*)ttm;
 453         struct radeon_device *rdev = radeon_get_rdev(bdev);
 454         uint32_t flags = RADEON_GART_PAGE_VALID | RADEON_GART_PAGE_READ |
 455                 RADEON_GART_PAGE_WRITE;
 456         int r;
 457
 458         if (gtt->bound)
 459                 return 0;
 460
 461         if (gtt->userptr) {
 462                 radeon_ttm_tt_pin_userptr(bdev, ttm);
 463                 flags &= ~RADEON_GART_PAGE_WRITE;
 464         }
 465
 466         gtt->offset = (unsigned long)(bo_mem->start << PAGE_SHIFT);
 467         if (!ttm->num_pages) {
 468                 WARN(1, "nothing to bind %u pages for mreg %p back %p!\n",
 469                      ttm->num_pages, bo_mem, ttm);
 470         }
 471         if (ttm->caching == ttm_cached)
 472                 flags |= RADEON_GART_PAGE_SNOOP;
 473         r = radeon_gart_bind(rdev, gtt->offset, ttm->num_pages,
 474                              ttm->pages, gtt->ttm.dma_address, flags);
 475         if (r) {
 476                 DRM_ERROR("failed to bind %u pages at 0x%08X\n",
 477                           ttm->num_pages, (unsigned)gtt->offset);
 478                 return r;
 479         }
 480         gtt->bound = true;
 481         return 0;
 482 }
 483
 484 static void radeon_ttm_backend_unbind(struct ttm_bo_device *bdev, struct ttm_tt *ttm)
 485 {
 486         struct radeon_ttm_tt *gtt = (void *)ttm;
 487         struct radeon_device *rdev = radeon_get_rdev(bdev);
 488
 489         if (!gtt->bound)
 490                 return;
 491
 492         radeon_gart_unbind(rdev, gtt->offset, ttm->num_pages);
 493
 494         if (gtt->userptr)
 495                 radeon_ttm_tt_unpin_userptr(bdev, ttm);
 496         gtt->bound = false;
 497 }
 498
 499 static void radeon_ttm_backend_destroy(struct ttm_bo_device *bdev, struct ttm_tt *ttm)
 500 {
 501         struct radeon_ttm_tt *gtt = (void *)ttm;
 502
 503         radeon_ttm_backend_unbind(bdev, ttm);
 504         ttm_tt_destroy_common(bdev, ttm);
 505
 506         ttm_tt_fini(&gtt->ttm);
 507         kfree(gtt);
 508 }
 509
 510 static struct ttm_tt *radeon_ttm_tt_create(struct ttm_buffer_object *bo,
 511                                            uint32_t page_flags)
 512 {
 513         struct radeon_ttm_tt *gtt;
 514         enum ttm_caching caching;
 515         struct radeon_bo *rbo;
 516 #if IS_ENABLED(CONFIG_AGP)
 517         struct radeon_device *rdev = radeon_get_rdev(bo->bdev);
 518
 519         if (rdev->flags & RADEON_IS_AGP) {
 520                 return ttm_agp_tt_create(bo, rdev->ddev->agp->bridge,
 521                                          page_flags);
 522         }
 523 #endif
 524         rbo = container_of(bo, struct radeon_bo, tbo);
 525
 526         gtt = kzalloc(sizeof(struct radeon_ttm_tt), GFP_KERNEL);
 527         if (gtt == NULL) {
 528                 return NULL;
 529         }
 530
 531         if (rbo->flags & RADEON_GEM_GTT_UC)
 532                 caching = ttm_uncached;
 533         else if (rbo->flags & RADEON_GEM_GTT_WC)
 534                 caching = ttm_write_combined;
 535         else
 536                 caching = ttm_cached;
 537
 538         if (ttm_dma_tt_init(&gtt->ttm, bo, page_flags, caching)) {
 539                 kfree(gtt);
 540                 return NULL;
 541         }
 542         return &gtt->ttm;
 543 }
 544
 545 static struct radeon_ttm_tt *radeon_ttm_tt_to_gtt(struct radeon_device *rdev,
 546                                                   struct ttm_tt *ttm)
 547 {
 548 #if IS_ENABLED(CONFIG_AGP)
 549         if (rdev->flags & RADEON_IS_AGP)
 550                 return NULL;
 551 #endif
 552
 553         if (!ttm)
 554                 return NULL;
 555         return container_of(ttm, struct radeon_ttm_tt, ttm);
 556 }
 557
 558 static int radeon_ttm_tt_populate(struct ttm_bo_device *bdev,
 559                                   struct ttm_tt *ttm,
 560                                   struct ttm_operation_ctx *ctx)
 561 {
 562         struct radeon_device *rdev = radeon_get_rdev(bdev);
 563         struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(rdev, ttm);
 564         bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
 565
 566         if (gtt && gtt->userptr) {
 567                 ttm->sg = kzalloc(sizeof(struct sg_table), GFP_KERNEL);
 568                 if (!ttm->sg)
 569                         return -ENOMEM;
 570
 571                 ttm->page_flags |= TTM_PAGE_FLAG_SG;
 572                 return 0;
 573         }
 574
 575         if (slave && ttm->sg) {
 576                 drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
 577                                                  gtt->ttm.dma_address, ttm->num_pages);
 578                 return 0;
 579         }
 580
 581         return ttm_pool_alloc(&rdev->mman.bdev.pool, ttm, ctx);
 582 }
 583
 584 static void radeon_ttm_tt_unpopulate(struct ttm_bo_device *bdev, struct ttm_tt *ttm)
 585 {
 586         struct radeon_device *rdev = radeon_get_rdev(bdev);
 587         struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(rdev, ttm);
 588         bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
 589
 590         if (gtt && gtt->userptr) {
 591                 kfree(ttm->sg);
 592                 ttm->page_flags &= ~TTM_PAGE_FLAG_SG;
 593                 return;
 594         }
 595
 596         if (slave)
 597                 return;
 598
 599         return ttm_pool_free(&rdev->mman.bdev.pool, ttm);
 600 }
 601
 602 int radeon_ttm_tt_set_userptr(struct radeon_device *rdev,
 603                               struct ttm_tt *ttm, uint64_t addr,
 604                               uint32_t flags)
 605 {
 606         struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(rdev, ttm);
 607
 608         if (gtt == NULL)
 609                 return -EINVAL;
 610
 611         gtt->userptr = addr;
 612         gtt->usermm = current->mm;
 613         gtt->userflags = flags;
 614         return 0;
 615 }
 616
 617 bool radeon_ttm_tt_is_bound(struct ttm_bo_device *bdev,
 618                             struct ttm_tt *ttm)
 619 {
 620 #if IS_ENABLED(CONFIG_AGP)
 621         struct radeon_device *rdev = radeon_get_rdev(bdev);
 622         if (rdev->flags & RADEON_IS_AGP)
 623                 return ttm_agp_is_bound(ttm);
 624 #endif
 625         return radeon_ttm_backend_is_bound(ttm);
 626 }
 627
 628 static int radeon_ttm_tt_bind(struct ttm_bo_device *bdev,
 629                               struct ttm_tt *ttm,
 630                               struct ttm_resource *bo_mem)
 631 {
 632 #if IS_ENABLED(CONFIG_AGP)
 633         struct radeon_device *rdev = radeon_get_rdev(bdev);
 634 #endif
 635
 636         if (!bo_mem)
 637                 return -EINVAL;
 638 #if IS_ENABLED(CONFIG_AGP)
 639         if (rdev->flags & RADEON_IS_AGP)
 640                 return ttm_agp_bind(ttm, bo_mem);
 641 #endif
 642
 643         return radeon_ttm_backend_bind(bdev, ttm, bo_mem);
 644 }
 645
 646 static void radeon_ttm_tt_unbind(struct ttm_bo_device *bdev,
 647                                  struct ttm_tt *ttm)
 648 {
 649 #if IS_ENABLED(CONFIG_AGP)
 650         struct radeon_device *rdev = radeon_get_rdev(bdev);
 651
 652         if (rdev->flags & RADEON_IS_AGP) {
 653                 ttm_agp_unbind(ttm);
 654                 return;
 655         }
 656 #endif
 657         radeon_ttm_backend_unbind(bdev, ttm);
 658 }
 659
 660 static void radeon_ttm_tt_destroy(struct ttm_bo_device *bdev,
 661                                   struct ttm_tt *ttm)
 662 {
 663 #if IS_ENABLED(CONFIG_AGP)
 664         struct radeon_device *rdev = radeon_get_rdev(bdev);
 665
 666         if (rdev->flags & RADEON_IS_AGP) {
 667                 ttm_agp_unbind(ttm);
 668                 ttm_tt_destroy_common(bdev, ttm);
 669                 ttm_agp_destroy(ttm);
 670                 return;
 671         }
 672 #endif
 673         radeon_ttm_backend_destroy(bdev, ttm);
 674 }
 675
 676 bool radeon_ttm_tt_has_userptr(struct radeon_device *rdev,
 677                                struct ttm_tt *ttm)
 678 {
 679         struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(rdev, ttm);
 680
 681         if (gtt == NULL)
 682                 return false;
 683
 684         return !!gtt->userptr;
 685 }
 686
 687 bool radeon_ttm_tt_is_readonly(struct radeon_device *rdev,
 688                                struct ttm_tt *ttm)
 689 {
 690         struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(rdev, ttm);
 691
 692         if (gtt == NULL)
 693                 return false;
 694
 695         return !!(gtt->userflags & RADEON_GEM_USERPTR_READONLY);
 696 }
 697
 698 static void
 699 radeon_bo_delete_mem_notify(struct ttm_buffer_object *bo)
 700 {
 701         radeon_bo_move_notify(bo, false, NULL);
 702 }
 703
 704 static struct ttm_bo_driver radeon_bo_driver = {
 705         .ttm_tt_create = &radeon_ttm_tt_create,
 706         .ttm_tt_populate = &radeon_ttm_tt_populate,
 707         .ttm_tt_unpopulate = &radeon_ttm_tt_unpopulate,
 708         .ttm_tt_destroy = &radeon_ttm_tt_destroy,
 709         .eviction_valuable = ttm_bo_eviction_valuable,
 710         .evict_flags = &radeon_evict_flags,
 711         .move = &radeon_bo_move,
 712         .verify_access = &radeon_verify_access,
 713         .delete_mem_notify = &radeon_bo_delete_mem_notify,
 714         .io_mem_reserve = &radeon_ttm_io_mem_reserve,
 715 };
 716
 717 int radeon_ttm_init(struct radeon_device *rdev)
 718 {
 719         int r;
 720
 721         /* No others user of address space so set it to 0 */
 722         r = ttm_bo_device_init(&rdev->mman.bdev, &radeon_bo_driver, rdev->dev,
 723                                rdev->ddev->anon_inode->i_mapping,
 724                                rdev->ddev->vma_offset_manager,
 725                                rdev->need_swiotlb,
 726                                dma_addressing_limited(&rdev->pdev->dev));
 727         if (r) {
 728                 DRM_ERROR("failed initializing buffer object driver(%d).\n", r);
 729                 return r;
 730         }
 731         rdev->mman.initialized = true;
 732
 733         ttm_pool_init(&rdev->mman.bdev.pool, rdev->dev, rdev->need_swiotlb,
 734                       dma_addressing_limited(&rdev->pdev->dev));
 735
 736         r = radeon_ttm_init_vram(rdev);
 737         if (r) {
 738                 DRM_ERROR("Failed initializing VRAM heap.\n");
 739                 return r;
 740         }
 741         /* Change the size here instead of the init above so only lpfn is affected */
 742         radeon_ttm_set_active_vram_size(rdev, rdev->mc.visible_vram_size);
 743
 744         r = radeon_bo_create(rdev, 256 * 1024, PAGE_SIZE, true,
 745                              RADEON_GEM_DOMAIN_VRAM, 0, NULL,
 746                              NULL, &rdev->stolen_vga_memory);
 747         if (r) {
 748                 return r;
 749         }
 750         r = radeon_bo_reserve(rdev->stolen_vga_memory, false);
 751         if (r)
 752                 return r;
 753         r = radeon_bo_pin(rdev->stolen_vga_memory, RADEON_GEM_DOMAIN_VRAM, NULL);
 754         radeon_bo_unreserve(rdev->stolen_vga_memory);
 755         if (r) {
 756                 radeon_bo_unref(&rdev->stolen_vga_memory);
 757                 return r;
 758         }
 759         DRM_INFO("radeon: %uM of VRAM memory ready\n",
 760                  (unsigned) (rdev->mc.real_vram_size / (1024 * 1024)));
 761
 762         r = radeon_ttm_init_gtt(rdev);
 763         if (r) {
 764                 DRM_ERROR("Failed initializing GTT heap.\n");
 765                 return r;
 766         }
 767         DRM_INFO("radeon: %uM of GTT memory ready.\n",
 768                  (unsigned)(rdev->mc.gtt_size / (1024 * 1024)));
 769
 770         r = radeon_ttm_debugfs_init(rdev);
 771         if (r) {
 772                 DRM_ERROR("Failed to init debugfs\n");
 773                 return r;
 774         }
 775         return 0;
 776 }
 777
 778 void radeon_ttm_fini(struct radeon_device *rdev)
 779 {
 780         int r;
 781
 782         if (!rdev->mman.initialized)
 783                 return;
 784         radeon_ttm_debugfs_fini(rdev);
 785         if (rdev->stolen_vga_memory) {
 786                 r = radeon_bo_reserve(rdev->stolen_vga_memory, false);
 787                 if (r == 0) {
 788                         radeon_bo_unpin(rdev->stolen_vga_memory);
 789                         radeon_bo_unreserve(rdev->stolen_vga_memory);
 790                 }
 791                 radeon_bo_unref(&rdev->stolen_vga_memory);
 792         }
 793         ttm_range_man_fini(&rdev->mman.bdev, TTM_PL_VRAM);
 794         ttm_range_man_fini(&rdev->mman.bdev, TTM_PL_TT);
 795         ttm_bo_device_release(&rdev->mman.bdev);
 796         radeon_gart_fini(rdev);
 797         rdev->mman.initialized = false;
 798         DRM_INFO("radeon: ttm finalized\n");
 799 }
 800
 801 /* this should only be called at bootup or when userspace
 802  * isn't running */
 803 void radeon_ttm_set_active_vram_size(struct radeon_device *rdev, u64 size)
 804 {
 805         struct ttm_resource_manager *man;
 806
 807         if (!rdev->mman.initialized)
 808                 return;
 809
 810         man = ttm_manager_type(&rdev->mman.bdev, TTM_PL_VRAM);
 811         /* this just adjusts TTM size idea, which sets lpfn to the correct value */
 812         man->size = size >> PAGE_SHIFT;
 813 }
 814
 815 static vm_fault_t radeon_ttm_fault(struct vm_fault *vmf)
 816 {
 817         struct ttm_buffer_object *bo = vmf->vma->vm_private_data;
 818         struct radeon_device *rdev = radeon_get_rdev(bo->bdev);
 819         vm_fault_t ret;
 820
 821         down_read(&rdev->pm.mclk_lock);
 822
 823         ret = ttm_bo_vm_reserve(bo, vmf);
 824         if (ret)
 825                 goto unlock_mclk;
 826
 827         ret = radeon_bo_fault_reserve_notify(bo);
 828         if (ret)
 829                 goto unlock_resv;
 830
 831         ret = ttm_bo_vm_fault_reserved(vmf, vmf->vma->vm_page_prot,
 832                                        TTM_BO_VM_NUM_PREFAULT, 1);
 833         if (ret == VM_FAULT_RETRY && !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT))
 834                 goto unlock_mclk;
 835
 836 unlock_resv:
 837         dma_resv_unlock(bo->base.resv);
 838
 839 unlock_mclk:
 840         up_read(&rdev->pm.mclk_lock);
 841         return ret;
 842 }
 843
 844 static struct vm_operations_struct radeon_ttm_vm_ops = {
 845         .fault = radeon_ttm_fault,
 846         .open = ttm_bo_vm_open,
 847         .close = ttm_bo_vm_close,
 848         .access = ttm_bo_vm_access
 849 };
 850
 851 int radeon_mmap(struct file *filp, struct vm_area_struct *vma)
 852 {
 853         int r;
 854         struct drm_file *file_priv = filp->private_data;
 855         struct radeon_device *rdev = file_priv->minor->dev->dev_private;
 856
 857         if (rdev == NULL)
 858                 return -EINVAL;
 859
 860         r = ttm_bo_mmap(filp, vma, &rdev->mman.bdev);
 861         if (unlikely(r != 0))
 862                 return r;
 863
 864         vma->vm_ops = &radeon_ttm_vm_ops;
 865         return 0;
 866 }
 867
 868 #if defined(CONFIG_DEBUG_FS)
 869
 870 static int radeon_mm_dump_table(struct seq_file *m, void *data)
 871 {
 872         struct drm_info_node *node = (struct drm_info_node *)m->private;
 873         unsigned ttm_pl = *(int*)node->info_ent->data;
 874         struct drm_device *dev = node->minor->dev;
 875         struct radeon_device *rdev = dev->dev_private;
 876         struct ttm_resource_manager *man = ttm_manager_type(&rdev->mman.bdev, ttm_pl);
 877         struct drm_printer p = drm_seq_file_printer(m);
 878
 879         man->func->debug(man, &p);
 880         return 0;
 881 }
 882
 883 static int radeon_ttm_pool_debugfs(struct seq_file *m, void *data)
 884 {
 885         struct drm_info_node *node = (struct drm_info_node *)m->private;
 886         struct drm_device *dev = node->minor->dev;
 887         struct radeon_device *rdev = dev->dev_private;
 888
 889         return ttm_pool_debugfs(&rdev->mman.bdev.pool, m);
 890 }
 891
 892 static int ttm_pl_vram = TTM_PL_VRAM;
 893 static int ttm_pl_tt = TTM_PL_TT;
 894
 895 static struct drm_info_list radeon_ttm_debugfs_list[] = {
 896         {"radeon_vram_mm", radeon_mm_dump_table, 0, &ttm_pl_vram},
 897         {"radeon_gtt_mm", radeon_mm_dump_table, 0, &ttm_pl_tt},
 898         {"ttm_page_pool", radeon_ttm_pool_debugfs, 0, NULL}
 899 };
 900
 901 static int radeon_ttm_vram_open(struct inode *inode, struct file *filep)
 902 {
 903         struct radeon_device *rdev = inode->i_private;
 904         i_size_write(inode, rdev->mc.mc_vram_size);
 905         filep->private_data = inode->i_private;
 906         return 0;
 907 }
 908
 909 static ssize_t radeon_ttm_vram_read(struct file *f, char __user *buf,
 910                                     size_t size, loff_t *pos)
 911 {
 912         struct radeon_device *rdev = f->private_data;
 913         ssize_t result = 0;
 914         int r;
 915
 916         if (size & 0x3 || *pos & 0x3)
 917                 return -EINVAL;
 918
 919         while (size) {
 920                 unsigned long flags;
 921                 uint32_t value;
 922
 923                 if (*pos >= rdev->mc.mc_vram_size)
 924                         return result;
 925
 926                 spin_lock_irqsave(&rdev->mmio_idx_lock, flags);
 927                 WREG32(RADEON_MM_INDEX, ((uint32_t)*pos) | 0x80000000);
 928                 if (rdev->family >= CHIP_CEDAR)
 929                         WREG32(EVERGREEN_MM_INDEX_HI, *pos >> 31);
 930                 value = RREG32(RADEON_MM_DATA);
 931                 spin_unlock_irqrestore(&rdev->mmio_idx_lock, flags);
 932
 933                 r = put_user(value, (uint32_t *)buf);
 934                 if (r)
 935                         return r;
 936
 937                 result += 4;
 938                 buf += 4;
 939                 *pos += 4;
 940                 size -= 4;
 941         }
 942
 943         return result;
 944 }
 945
 946 static const struct file_operations radeon_ttm_vram_fops = {
 947         .owner = THIS_MODULE,
 948         .open = radeon_ttm_vram_open,
 949         .read = radeon_ttm_vram_read,
 950         .llseek = default_llseek
 951 };
 952
 953 static int radeon_ttm_gtt_open(struct inode *inode, struct file *filep)
 954 {
 955         struct radeon_device *rdev = inode->i_private;
 956         i_size_write(inode, rdev->mc.gtt_size);
 957         filep->private_data = inode->i_private;
 958         return 0;
 959 }
 960
 961 static ssize_t radeon_ttm_gtt_read(struct file *f, char __user *buf,
 962                                    size_t size, loff_t *pos)
 963 {
 964         struct radeon_device *rdev = f->private_data;
 965         ssize_t result = 0;
 966         int r;
 967
 968         while (size) {
 969                 loff_t p = *pos / PAGE_SIZE;
 970                 unsigned off = *pos & ~PAGE_MASK;
 971                 size_t cur_size = min_t(size_t, size, PAGE_SIZE - off);
 972                 struct page *page;
 973                 void *ptr;
 974
 975                 if (p >= rdev->gart.num_cpu_pages)
 976                         return result;
 977
 978                 page = rdev->gart.pages[p];
 979                 if (page) {
 980                         ptr = kmap(page);
 981                         ptr += off;
 982
 983                         r = copy_to_user(buf, ptr, cur_size);
 984                         kunmap(rdev->gart.pages[p]);
 985                 } else
 986                         r = clear_user(buf, cur_size);
 987
 988                 if (r)
 989                         return -EFAULT;
 990
 991                 result += cur_size;
 992                 buf += cur_size;
 993                 *pos += cur_size;
 994                 size -= cur_size;
 995         }
 996
 997         return result;
 998 }
 999
1000 static const struct file_operations radeon_ttm_gtt_fops = {
1001         .owner = THIS_MODULE,
1002         .open = radeon_ttm_gtt_open,
1003         .read = radeon_ttm_gtt_read,
1004         .llseek = default_llseek
1005 };
1006
1007 #endif
1008
1009 static int radeon_ttm_debugfs_init(struct radeon_device *rdev)
1010 {
1011 #if defined(CONFIG_DEBUG_FS)
1012         unsigned count;
1013
1014         struct drm_minor *minor = rdev->ddev->primary;
1015         struct dentry *root = minor->debugfs_root;
1016
1017         rdev->mman.vram = debugfs_create_file("radeon_vram", S_IFREG | S_IRUGO,
1018                                               root, rdev,
1019                                               &radeon_ttm_vram_fops);
1020
1021         rdev->mman.gtt = debugfs_create_file("radeon_gtt", S_IFREG | S_IRUGO,
1022                                              root, rdev, &radeon_ttm_gtt_fops);
1023
1024         count = ARRAY_SIZE(radeon_ttm_debugfs_list);
1025
1026         return radeon_debugfs_add_files(rdev, radeon_ttm_debugfs_list, count);
1027 #else
1028
1029         return 0;
1030 #endif
1031 }
1032
1033 static void radeon_ttm_debugfs_fini(struct radeon_device *rdev)
1034 {
1035 #if defined(CONFIG_DEBUG_FS)
1036
1037         debugfs_remove(rdev->mman.vram);
1038         rdev->mman.vram = NULL;
1039
1040         debugfs_remove(rdev->mman.gtt);
1041         rdev->mman.gtt = NULL;
1042 #endif
1043 }