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 #include <ttm/ttm_bo_api.h>
  33 #include <ttm/ttm_bo_driver.h>
  34 #include <ttm/ttm_placement.h>
  35 #include <ttm/ttm_module.h>
  36 #include <ttm/ttm_page_alloc.h>
  37 #include <drm/drmP.h>
  38 #include <drm/radeon_drm.h>
  39 #include <linux/seq_file.h>
  40 #include <linux/slab.h>
  41 #include "radeon_reg.h"
  42 #include "radeon.h"
  43
  44 #define DRM_FILE_PAGE_OFFSET (0x100000000ULL >> PAGE_SHIFT)
  45
  46 static int radeon_ttm_debugfs_init(struct radeon_device *rdev);
  47
  48 static struct radeon_device *radeon_get_rdev(struct ttm_bo_device *bdev)
  49 {
  50         struct radeon_mman *mman;
  51         struct radeon_device *rdev;
  52
  53         mman = container_of(bdev, struct radeon_mman, bdev);
  54         rdev = container_of(mman, struct radeon_device, mman);
  55         return rdev;
  56 }
  57
  58
  59 /*
  60  * Global memory.
  61  */
  62 static int radeon_ttm_mem_global_init(struct drm_global_reference *ref)
  63 {
  64         return ttm_mem_global_init(ref->object);
  65 }
  66
  67 static void radeon_ttm_mem_global_release(struct drm_global_reference *ref)
  68 {
  69         ttm_mem_global_release(ref->object);
  70 }
  71
  72 static int radeon_ttm_global_init(struct radeon_device *rdev)
  73 {
  74         struct drm_global_reference *global_ref;
  75         int r;
  76
  77         rdev->mman.mem_global_referenced = false;
  78         global_ref = &rdev->mman.mem_global_ref;
  79         global_ref->global_type = DRM_GLOBAL_TTM_MEM;
  80         global_ref->size = sizeof(struct ttm_mem_global);
  81         global_ref->init = &radeon_ttm_mem_global_init;
  82         global_ref->release = &radeon_ttm_mem_global_release;
  83         r = drm_global_item_ref(global_ref);
  84         if (r != 0) {
  85                 DRM_ERROR("Failed setting up TTM memory accounting "
  86                           "subsystem.\n");
  87                 return r;
  88         }
  89
  90         rdev->mman.bo_global_ref.mem_glob =
  91                 rdev->mman.mem_global_ref.object;
  92         global_ref = &rdev->mman.bo_global_ref.ref;
  93         global_ref->global_type = DRM_GLOBAL_TTM_BO;
  94         global_ref->size = sizeof(struct ttm_bo_global);
  95         global_ref->init = &ttm_bo_global_init;
  96         global_ref->release = &ttm_bo_global_release;
  97         r = drm_global_item_ref(global_ref);
  98         if (r != 0) {
  99                 DRM_ERROR("Failed setting up TTM BO subsystem.\n");
 100                 drm_global_item_unref(&rdev->mman.mem_global_ref);
 101                 return r;
 102         }
 103
 104         rdev->mman.mem_global_referenced = true;
 105         return 0;
 106 }
 107
 108 static void radeon_ttm_global_fini(struct radeon_device *rdev)
 109 {
 110         if (rdev->mman.mem_global_referenced) {
 111                 drm_global_item_unref(&rdev->mman.bo_global_ref.ref);
 112                 drm_global_item_unref(&rdev->mman.mem_global_ref);
 113                 rdev->mman.mem_global_referenced = false;
 114         }
 115 }
 116
 117 static int radeon_invalidate_caches(struct ttm_bo_device *bdev, uint32_t flags)
 118 {
 119         return 0;
 120 }
 121
 122 static int radeon_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
 123                                 struct ttm_mem_type_manager *man)
 124 {
 125         struct radeon_device *rdev;
 126
 127         rdev = radeon_get_rdev(bdev);
 128
 129         switch (type) {
 130         case TTM_PL_SYSTEM:
 131                 /* System memory */
 132                 man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
 133                 man->available_caching = TTM_PL_MASK_CACHING;
 134                 man->default_caching = TTM_PL_FLAG_CACHED;
 135                 break;
 136         case TTM_PL_TT:
 137                 man->func = &ttm_bo_manager_func;
 138                 man->gpu_offset = rdev->mc.gtt_start;
 139                 man->available_caching = TTM_PL_MASK_CACHING;
 140                 man->default_caching = TTM_PL_FLAG_CACHED;
 141                 man->flags = TTM_MEMTYPE_FLAG_MAPPABLE | TTM_MEMTYPE_FLAG_CMA;
 142 #if __OS_HAS_AGP
 143                 if (rdev->flags & RADEON_IS_AGP) {
 144                         if (!(drm_core_has_AGP(rdev->ddev) && rdev->ddev->agp)) {
 145                                 DRM_ERROR("AGP is not enabled for memory type %u\n",
 146                                           (unsigned)type);
 147                                 return -EINVAL;
 148                         }
 149                         if (!rdev->ddev->agp->cant_use_aperture)
 150                                 man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
 151                         man->available_caching = TTM_PL_FLAG_UNCACHED |
 152                                                  TTM_PL_FLAG_WC;
 153                         man->default_caching = TTM_PL_FLAG_WC;
 154                 }
 155 #endif
 156                 break;
 157         case TTM_PL_VRAM:
 158                 /* "On-card" video ram */
 159                 man->func = &ttm_bo_manager_func;
 160                 man->gpu_offset = rdev->mc.vram_start;
 161                 man->flags = TTM_MEMTYPE_FLAG_FIXED |
 162                              TTM_MEMTYPE_FLAG_MAPPABLE;
 163                 man->available_caching = TTM_PL_FLAG_UNCACHED | TTM_PL_FLAG_WC;
 164                 man->default_caching = TTM_PL_FLAG_WC;
 165                 break;
 166         default:
 167                 DRM_ERROR("Unsupported memory type %u\n", (unsigned)type);
 168                 return -EINVAL;
 169         }
 170         return 0;
 171 }
 172
 173 static void radeon_evict_flags(struct ttm_buffer_object *bo,
 174                                 struct ttm_placement *placement)
 175 {
 176         struct radeon_bo *rbo;
 177         static u32 placements = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM;
 178
 179         if (!radeon_ttm_bo_is_radeon_bo(bo)) {
 180                 placement->fpfn = 0;
 181                 placement->lpfn = 0;
 182                 placement->placement = &placements;
 183                 placement->busy_placement = &placements;
 184                 placement->num_placement = 1;
 185                 placement->num_busy_placement = 1;
 186                 return;
 187         }
 188         rbo = container_of(bo, struct radeon_bo, tbo);
 189         switch (bo->mem.mem_type) {
 190         case TTM_PL_VRAM:
 191                 if (rbo->rdev->cp[RADEON_RING_TYPE_GFX_INDEX].ready == false)
 192                         radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_CPU);
 193                 else
 194                         radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_GTT);
 195                 break;
 196         case TTM_PL_TT:
 197         default:
 198                 radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_CPU);
 199         }
 200         *placement = rbo->placement;
 201 }
 202
 203 static int radeon_verify_access(struct ttm_buffer_object *bo, struct file *filp)
 204 {
 205         return 0;
 206 }
 207
 208 static void radeon_move_null(struct ttm_buffer_object *bo,
 209                              struct ttm_mem_reg *new_mem)
 210 {
 211         struct ttm_mem_reg *old_mem = &bo->mem;
 212
 213         BUG_ON(old_mem->mm_node != NULL);
 214         *old_mem = *new_mem;
 215         new_mem->mm_node = NULL;
 216 }
 217
 218 static int radeon_move_blit(struct ttm_buffer_object *bo,
 219                         bool evict, int no_wait_reserve, bool no_wait_gpu,
 220                         struct ttm_mem_reg *new_mem,
 221                         struct ttm_mem_reg *old_mem)
 222 {
 223         struct radeon_device *rdev;
 224         uint64_t old_start, new_start;
 225         struct radeon_fence *fence;
 226         int r;
 227
 228         rdev = radeon_get_rdev(bo->bdev);
 229         r = radeon_fence_create(rdev, &fence, RADEON_RING_TYPE_GFX_INDEX);
 230         if (unlikely(r)) {
 231                 return r;
 232         }
 233         old_start = old_mem->start << PAGE_SHIFT;
 234         new_start = new_mem->start << PAGE_SHIFT;
 235
 236         switch (old_mem->mem_type) {
 237         case TTM_PL_VRAM:
 238                 old_start += rdev->mc.vram_start;
 239                 break;
 240         case TTM_PL_TT:
 241                 old_start += rdev->mc.gtt_start;
 242                 break;
 243         default:
 244                 DRM_ERROR("Unknown placement %d\n", old_mem->mem_type);
 245                 return -EINVAL;
 246         }
 247         switch (new_mem->mem_type) {
 248         case TTM_PL_VRAM:
 249                 new_start += rdev->mc.vram_start;
 250                 break;
 251         case TTM_PL_TT:
 252                 new_start += rdev->mc.gtt_start;
 253                 break;
 254         default:
 255                 DRM_ERROR("Unknown placement %d\n", old_mem->mem_type);
 256                 return -EINVAL;
 257         }
 258         if (!rdev->cp[RADEON_RING_TYPE_GFX_INDEX].ready) {
 259                 DRM_ERROR("Trying to move memory with CP turned off.\n");
 260                 return -EINVAL;
 261         }
 262
 263         BUILD_BUG_ON((PAGE_SIZE % RADEON_GPU_PAGE_SIZE) != 0);
 264
 265         r = radeon_copy(rdev, old_start, new_start,
 266                         new_mem->num_pages * (PAGE_SIZE / RADEON_GPU_PAGE_SIZE), /* GPU pages */
 267                         fence);
 268         /* FIXME: handle copy error */
 269         r = ttm_bo_move_accel_cleanup(bo, (void *)fence, NULL,
 270                                       evict, no_wait_reserve, no_wait_gpu, new_mem);
 271         radeon_fence_unref(&fence);
 272         return r;
 273 }
 274
 275 static int radeon_move_vram_ram(struct ttm_buffer_object *bo,
 276                                 bool evict, bool interruptible,
 277                                 bool no_wait_reserve, bool no_wait_gpu,
 278                                 struct ttm_mem_reg *new_mem)
 279 {
 280         struct radeon_device *rdev;
 281         struct ttm_mem_reg *old_mem = &bo->mem;
 282         struct ttm_mem_reg tmp_mem;
 283         u32 placements;
 284         struct ttm_placement placement;
 285         int r;
 286
 287         rdev = radeon_get_rdev(bo->bdev);
 288         tmp_mem = *new_mem;
 289         tmp_mem.mm_node = NULL;
 290         placement.fpfn = 0;
 291         placement.lpfn = 0;
 292         placement.num_placement = 1;
 293         placement.placement = &placements;
 294         placement.num_busy_placement = 1;
 295         placement.busy_placement = &placements;
 296         placements = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
 297         r = ttm_bo_mem_space(bo, &placement, &tmp_mem,
 298                              interruptible, no_wait_reserve, no_wait_gpu);
 299         if (unlikely(r)) {
 300                 return r;
 301         }
 302
 303         r = ttm_tt_set_placement_caching(bo->ttm, tmp_mem.placement);
 304         if (unlikely(r)) {
 305                 goto out_cleanup;
 306         }
 307
 308         r = ttm_tt_bind(bo->ttm, &tmp_mem);
 309         if (unlikely(r)) {
 310                 goto out_cleanup;
 311         }
 312         r = radeon_move_blit(bo, true, no_wait_reserve, no_wait_gpu, &tmp_mem, old_mem);
 313         if (unlikely(r)) {
 314                 goto out_cleanup;
 315         }
 316         r = ttm_bo_move_ttm(bo, true, no_wait_reserve, no_wait_gpu, new_mem);
 317 out_cleanup:
 318         ttm_bo_mem_put(bo, &tmp_mem);
 319         return r;
 320 }
 321
 322 static int radeon_move_ram_vram(struct ttm_buffer_object *bo,
 323                                 bool evict, bool interruptible,
 324                                 bool no_wait_reserve, bool no_wait_gpu,
 325                                 struct ttm_mem_reg *new_mem)
 326 {
 327         struct radeon_device *rdev;
 328         struct ttm_mem_reg *old_mem = &bo->mem;
 329         struct ttm_mem_reg tmp_mem;
 330         struct ttm_placement placement;
 331         u32 placements;
 332         int r;
 333
 334         rdev = radeon_get_rdev(bo->bdev);
 335         tmp_mem = *new_mem;
 336         tmp_mem.mm_node = NULL;
 337         placement.fpfn = 0;
 338         placement.lpfn = 0;
 339         placement.num_placement = 1;
 340         placement.placement = &placements;
 341         placement.num_busy_placement = 1;
 342         placement.busy_placement = &placements;
 343         placements = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
 344         r = ttm_bo_mem_space(bo, &placement, &tmp_mem, interruptible, no_wait_reserve, no_wait_gpu);
 345         if (unlikely(r)) {
 346                 return r;
 347         }
 348         r = ttm_bo_move_ttm(bo, true, no_wait_reserve, no_wait_gpu, &tmp_mem);
 349         if (unlikely(r)) {
 350                 goto out_cleanup;
 351         }
 352         r = radeon_move_blit(bo, true, no_wait_reserve, no_wait_gpu, new_mem, old_mem);
 353         if (unlikely(r)) {
 354                 goto out_cleanup;
 355         }
 356 out_cleanup:
 357         ttm_bo_mem_put(bo, &tmp_mem);
 358         return r;
 359 }
 360
 361 static int radeon_bo_move(struct ttm_buffer_object *bo,
 362                         bool evict, bool interruptible,
 363                         bool no_wait_reserve, bool no_wait_gpu,
 364                         struct ttm_mem_reg *new_mem)
 365 {
 366         struct radeon_device *rdev;
 367         struct ttm_mem_reg *old_mem = &bo->mem;
 368         int r;
 369
 370         rdev = radeon_get_rdev(bo->bdev);
 371         if (old_mem->mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
 372                 radeon_move_null(bo, new_mem);
 373                 return 0;
 374         }
 375         if ((old_mem->mem_type == TTM_PL_TT &&
 376              new_mem->mem_type == TTM_PL_SYSTEM) ||
 377             (old_mem->mem_type == TTM_PL_SYSTEM &&
 378              new_mem->mem_type == TTM_PL_TT)) {
 379                 /* bind is enough */
 380                 radeon_move_null(bo, new_mem);
 381                 return 0;
 382         }
 383         if (!rdev->cp[RADEON_RING_TYPE_GFX_INDEX].ready || rdev->asic->copy == NULL) {
 384                 /* use memcpy */
 385                 goto memcpy;
 386         }
 387
 388         if (old_mem->mem_type == TTM_PL_VRAM &&
 389             new_mem->mem_type == TTM_PL_SYSTEM) {
 390                 r = radeon_move_vram_ram(bo, evict, interruptible,
 391                                         no_wait_reserve, no_wait_gpu, new_mem);
 392         } else if (old_mem->mem_type == TTM_PL_SYSTEM &&
 393                    new_mem->mem_type == TTM_PL_VRAM) {
 394                 r = radeon_move_ram_vram(bo, evict, interruptible,
 395                                             no_wait_reserve, no_wait_gpu, new_mem);
 396         } else {
 397                 r = radeon_move_blit(bo, evict, no_wait_reserve, no_wait_gpu, new_mem, old_mem);
 398         }
 399
 400         if (r) {
 401 memcpy:
 402                 r = ttm_bo_move_memcpy(bo, evict, no_wait_reserve, no_wait_gpu, new_mem);
 403         }
 404         return r;
 405 }
 406
 407 static int radeon_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
 408 {
 409         struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
 410         struct radeon_device *rdev = radeon_get_rdev(bdev);
 411
 412         mem->bus.addr = NULL;
 413         mem->bus.offset = 0;
 414         mem->bus.size = mem->num_pages << PAGE_SHIFT;
 415         mem->bus.base = 0;
 416         mem->bus.is_iomem = false;
 417         if (!(man->flags & TTM_MEMTYPE_FLAG_MAPPABLE))
 418                 return -EINVAL;
 419         switch (mem->mem_type) {
 420         case TTM_PL_SYSTEM:
 421                 /* system memory */
 422                 return 0;
 423         case TTM_PL_TT:
 424 #if __OS_HAS_AGP
 425                 if (rdev->flags & RADEON_IS_AGP) {
 426                         /* RADEON_IS_AGP is set only if AGP is active */
 427                         mem->bus.offset = mem->start << PAGE_SHIFT;
 428                         mem->bus.base = rdev->mc.agp_base;
 429                         mem->bus.is_iomem = !rdev->ddev->agp->cant_use_aperture;
 430                 }
 431 #endif
 432                 break;
 433         case TTM_PL_VRAM:
 434                 mem->bus.offset = mem->start << PAGE_SHIFT;
 435                 /* check if it's visible */
 436                 if ((mem->bus.offset + mem->bus.size) > rdev->mc.visible_vram_size)
 437                         return -EINVAL;
 438                 mem->bus.base = rdev->mc.aper_base;
 439                 mem->bus.is_iomem = true;
 440 #ifdef __alpha__
 441                 /*
 442                  * Alpha: use bus.addr to hold the ioremap() return,
 443                  * so we can modify bus.base below.
 444                  */
 445                 if (mem->placement & TTM_PL_FLAG_WC)
 446                         mem->bus.addr =
 447                                 ioremap_wc(mem->bus.base + mem->bus.offset,
 448                                            mem->bus.size);
 449                 else
 450                         mem->bus.addr =
 451                                 ioremap_nocache(mem->bus.base + mem->bus.offset,
 452                                                 mem->bus.size);
 453
 454                 /*
 455                  * Alpha: Use just the bus offset plus
 456                  * the hose/domain memory base for bus.base.
 457                  * It then can be used to build PTEs for VRAM
 458                  * access, as done in ttm_bo_vm_fault().
 459                  */
 460                 mem->bus.base = (mem->bus.base & 0x0ffffffffUL) +
 461                         rdev->ddev->hose->dense_mem_base;
 462 #endif
 463                 break;
 464         default:
 465                 return -EINVAL;
 466         }
 467         return 0;
 468 }
 469
 470 static void radeon_ttm_io_mem_free(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
 471 {
 472 }
 473
 474 static int radeon_sync_obj_wait(void *sync_obj, void *sync_arg,
 475                                 bool lazy, bool interruptible)
 476 {
 477         return radeon_fence_wait((struct radeon_fence *)sync_obj, interruptible);
 478 }
 479
 480 static int radeon_sync_obj_flush(void *sync_obj, void *sync_arg)
 481 {
 482         return 0;
 483 }
 484
 485 static void radeon_sync_obj_unref(void **sync_obj)
 486 {
 487         radeon_fence_unref((struct radeon_fence **)sync_obj);
 488 }
 489
 490 static void *radeon_sync_obj_ref(void *sync_obj)
 491 {
 492         return radeon_fence_ref((struct radeon_fence *)sync_obj);
 493 }
 494
 495 static bool radeon_sync_obj_signaled(void *sync_obj, void *sync_arg)
 496 {
 497         return radeon_fence_signaled((struct radeon_fence *)sync_obj);
 498 }
 499
 500 /*
 501  * TTM backend functions.
 502  */
 503 struct radeon_ttm_tt {
 504         struct ttm_dma_tt               ttm;
 505         struct radeon_device            *rdev;
 506         u64                             offset;
 507 };
 508
 509 static int radeon_ttm_backend_bind(struct ttm_tt *ttm,
 510                                    struct ttm_mem_reg *bo_mem)
 511 {
 512         struct radeon_ttm_tt *gtt = (void*)ttm;
 513         int r;
 514
 515         gtt->offset = (unsigned long)(bo_mem->start << PAGE_SHIFT);
 516         if (!ttm->num_pages) {
 517                 WARN(1, "nothing to bind %lu pages for mreg %p back %p!\n",
 518                      ttm->num_pages, bo_mem, ttm);
 519         }
 520         r = radeon_gart_bind(gtt->rdev, gtt->offset,
 521                              ttm->num_pages, ttm->pages, gtt->ttm.dma_address);
 522         if (r) {
 523                 DRM_ERROR("failed to bind %lu pages at 0x%08X\n",
 524                           ttm->num_pages, (unsigned)gtt->offset);
 525                 return r;
 526         }
 527         return 0;
 528 }
 529
 530 static int radeon_ttm_backend_unbind(struct ttm_tt *ttm)
 531 {
 532         struct radeon_ttm_tt *gtt = (void *)ttm;
 533
 534         radeon_gart_unbind(gtt->rdev, gtt->offset, ttm->num_pages);
 535         return 0;
 536 }
 537
 538 static void radeon_ttm_backend_destroy(struct ttm_tt *ttm)
 539 {
 540         struct radeon_ttm_tt *gtt = (void *)ttm;
 541
 542         ttm_dma_tt_fini(&gtt->ttm);
 543         kfree(gtt);
 544 }
 545
 546 static struct ttm_backend_func radeon_backend_func = {
 547         .bind = &radeon_ttm_backend_bind,
 548         .unbind = &radeon_ttm_backend_unbind,
 549         .destroy = &radeon_ttm_backend_destroy,
 550 };
 551
 552 struct ttm_tt *radeon_ttm_tt_create(struct ttm_bo_device *bdev,
 553                                     unsigned long size, uint32_t page_flags,
 554                                     struct page *dummy_read_page)
 555 {
 556         struct radeon_device *rdev;
 557         struct radeon_ttm_tt *gtt;
 558
 559         rdev = radeon_get_rdev(bdev);
 560 #if __OS_HAS_AGP
 561         if (rdev->flags & RADEON_IS_AGP) {
 562                 return ttm_agp_tt_create(bdev, rdev->ddev->agp->bridge,
 563                                          size, page_flags, dummy_read_page);
 564         }
 565 #endif
 566
 567         gtt = kzalloc(sizeof(struct radeon_ttm_tt), GFP_KERNEL);
 568         if (gtt == NULL) {
 569                 return NULL;
 570         }
 571         gtt->ttm.ttm.func = &radeon_backend_func;
 572         gtt->rdev = rdev;
 573         if (ttm_dma_tt_init(&gtt->ttm, bdev, size, page_flags, dummy_read_page)) {
 574                 kfree(gtt);
 575                 return NULL;
 576         }
 577         return &gtt->ttm.ttm;
 578 }
 579
 580 static int radeon_ttm_tt_populate(struct ttm_tt *ttm)
 581 {
 582         struct radeon_device *rdev;
 583         struct radeon_ttm_tt *gtt = (void *)ttm;
 584         unsigned i;
 585         int r;
 586
 587         if (ttm->state != tt_unpopulated)
 588                 return 0;
 589
 590         rdev = radeon_get_rdev(ttm->bdev);
 591
 592 #ifdef CONFIG_SWIOTLB
 593         if (swiotlb_nr_tbl()) {
 594                 return ttm_dma_populate(&gtt->ttm, rdev->dev);
 595         }
 596 #endif
 597
 598         r = ttm_pool_populate(ttm);
 599         if (r) {
 600                 return r;
 601         }
 602
 603         for (i = 0; i < ttm->num_pages; i++) {
 604                 gtt->ttm.dma_address[i] = pci_map_page(rdev->pdev, ttm->pages[i],
 605                                                        0, PAGE_SIZE,
 606                                                        PCI_DMA_BIDIRECTIONAL);
 607                 if (pci_dma_mapping_error(rdev->pdev, gtt->ttm.dma_address[i])) {
 608                         while (--i) {
 609                                 pci_unmap_page(rdev->pdev, gtt->ttm.dma_address[i],
 610                                                PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
 611                                 gtt->ttm.dma_address[i] = 0;
 612                         }
 613                         ttm_pool_unpopulate(ttm);
 614                         return -EFAULT;
 615                 }
 616         }
 617         return 0;
 618 }
 619
 620 static void radeon_ttm_tt_unpopulate(struct ttm_tt *ttm)
 621 {
 622         struct radeon_device *rdev;
 623         struct radeon_ttm_tt *gtt = (void *)ttm;
 624         unsigned i;
 625
 626         rdev = radeon_get_rdev(ttm->bdev);
 627
 628 #ifdef CONFIG_SWIOTLB
 629         if (swiotlb_nr_tbl()) {
 630                 ttm_dma_unpopulate(&gtt->ttm, rdev->dev);
 631                 return;
 632         }
 633 #endif
 634
 635         for (i = 0; i < ttm->num_pages; i++) {
 636                 if (gtt->ttm.dma_address[i]) {
 637                         pci_unmap_page(rdev->pdev, gtt->ttm.dma_address[i],
 638                                        PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
 639                 }
 640         }
 641
 642         ttm_pool_unpopulate(ttm);
 643 }
 644
 645 static struct ttm_bo_driver radeon_bo_driver = {
 646         .ttm_tt_create = &radeon_ttm_tt_create,
 647         .ttm_tt_populate = &radeon_ttm_tt_populate,
 648         .ttm_tt_unpopulate = &radeon_ttm_tt_unpopulate,
 649         .invalidate_caches = &radeon_invalidate_caches,
 650         .init_mem_type = &radeon_init_mem_type,
 651         .evict_flags = &radeon_evict_flags,
 652         .move = &radeon_bo_move,
 653         .verify_access = &radeon_verify_access,
 654         .sync_obj_signaled = &radeon_sync_obj_signaled,
 655         .sync_obj_wait = &radeon_sync_obj_wait,
 656         .sync_obj_flush = &radeon_sync_obj_flush,
 657         .sync_obj_unref = &radeon_sync_obj_unref,
 658         .sync_obj_ref = &radeon_sync_obj_ref,
 659         .move_notify = &radeon_bo_move_notify,
 660         .fault_reserve_notify = &radeon_bo_fault_reserve_notify,
 661         .io_mem_reserve = &radeon_ttm_io_mem_reserve,
 662         .io_mem_free = &radeon_ttm_io_mem_free,
 663 };
 664
 665 int radeon_ttm_init(struct radeon_device *rdev)
 666 {
 667         int r;
 668
 669         r = radeon_ttm_global_init(rdev);
 670         if (r) {
 671                 return r;
 672         }
 673         /* No others user of address space so set it to 0 */
 674         r = ttm_bo_device_init(&rdev->mman.bdev,
 675                                rdev->mman.bo_global_ref.ref.object,
 676                                &radeon_bo_driver, DRM_FILE_PAGE_OFFSET,
 677                                rdev->need_dma32);
 678         if (r) {
 679                 DRM_ERROR("failed initializing buffer object driver(%d).\n", r);
 680                 return r;
 681         }
 682         rdev->mman.initialized = true;
 683         r = ttm_bo_init_mm(&rdev->mman.bdev, TTM_PL_VRAM,
 684                                 rdev->mc.real_vram_size >> PAGE_SHIFT);
 685         if (r) {
 686                 DRM_ERROR("Failed initializing VRAM heap.\n");
 687                 return r;
 688         }
 689         r = radeon_bo_create(rdev, 256 * 1024, PAGE_SIZE, true,
 690                                 RADEON_GEM_DOMAIN_VRAM,
 691                                 &rdev->stollen_vga_memory);
 692         if (r) {
 693                 return r;
 694         }
 695         r = radeon_bo_reserve(rdev->stollen_vga_memory, false);
 696         if (r)
 697                 return r;
 698         r = radeon_bo_pin(rdev->stollen_vga_memory, RADEON_GEM_DOMAIN_VRAM, NULL);
 699         radeon_bo_unreserve(rdev->stollen_vga_memory);
 700         if (r) {
 701                 radeon_bo_unref(&rdev->stollen_vga_memory);
 702                 return r;
 703         }
 704         DRM_INFO("radeon: %uM of VRAM memory ready\n",
 705                  (unsigned)rdev->mc.real_vram_size / (1024 * 1024));
 706         r = ttm_bo_init_mm(&rdev->mman.bdev, TTM_PL_TT,
 707                                 rdev->mc.gtt_size >> PAGE_SHIFT);
 708         if (r) {
 709                 DRM_ERROR("Failed initializing GTT heap.\n");
 710                 return r;
 711         }
 712         DRM_INFO("radeon: %uM of GTT memory ready.\n",
 713                  (unsigned)(rdev->mc.gtt_size / (1024 * 1024)));
 714         if (unlikely(rdev->mman.bdev.dev_mapping == NULL)) {
 715                 rdev->mman.bdev.dev_mapping = rdev->ddev->dev_mapping;
 716         }
 717
 718         r = radeon_ttm_debugfs_init(rdev);
 719         if (r) {
 720                 DRM_ERROR("Failed to init debugfs\n");
 721                 return r;
 722         }
 723         return 0;
 724 }
 725
 726 void radeon_ttm_fini(struct radeon_device *rdev)
 727 {
 728         int r;
 729
 730         if (!rdev->mman.initialized)
 731                 return;
 732         if (rdev->stollen_vga_memory) {
 733                 r = radeon_bo_reserve(rdev->stollen_vga_memory, false);
 734                 if (r == 0) {
 735                         radeon_bo_unpin(rdev->stollen_vga_memory);
 736                         radeon_bo_unreserve(rdev->stollen_vga_memory);
 737                 }
 738                 radeon_bo_unref(&rdev->stollen_vga_memory);
 739         }
 740         ttm_bo_clean_mm(&rdev->mman.bdev, TTM_PL_VRAM);
 741         ttm_bo_clean_mm(&rdev->mman.bdev, TTM_PL_TT);
 742         ttm_bo_device_release(&rdev->mman.bdev);
 743         radeon_gart_fini(rdev);
 744         radeon_ttm_global_fini(rdev);
 745         rdev->mman.initialized = false;
 746         DRM_INFO("radeon: ttm finalized\n");
 747 }
 748
 749 /* this should only be called at bootup or when userspace
 750  * isn't running */
 751 void radeon_ttm_set_active_vram_size(struct radeon_device *rdev, u64 size)
 752 {
 753         struct ttm_mem_type_manager *man;
 754
 755         if (!rdev->mman.initialized)
 756                 return;
 757
 758         man = &rdev->mman.bdev.man[TTM_PL_VRAM];
 759         /* this just adjusts TTM size idea, which sets lpfn to the correct value */
 760         man->size = size >> PAGE_SHIFT;
 761 }
 762
 763 static struct vm_operations_struct radeon_ttm_vm_ops;
 764 static const struct vm_operations_struct *ttm_vm_ops = NULL;
 765
 766 static int radeon_ttm_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 767 {
 768         struct ttm_buffer_object *bo;
 769         struct radeon_device *rdev;
 770         int r;
 771
 772         bo = (struct ttm_buffer_object *)vma->vm_private_data;
 773         if (bo == NULL) {
 774                 return VM_FAULT_NOPAGE;
 775         }
 776         rdev = radeon_get_rdev(bo->bdev);
 777         mutex_lock(&rdev->vram_mutex);
 778         r = ttm_vm_ops->fault(vma, vmf);
 779         mutex_unlock(&rdev->vram_mutex);
 780         return r;
 781 }
 782
 783 int radeon_mmap(struct file *filp, struct vm_area_struct *vma)
 784 {
 785         struct drm_file *file_priv;
 786         struct radeon_device *rdev;
 787         int r;
 788
 789         if (unlikely(vma->vm_pgoff < DRM_FILE_PAGE_OFFSET)) {
 790                 return drm_mmap(filp, vma);
 791         }
 792
 793         file_priv = filp->private_data;
 794         rdev = file_priv->minor->dev->dev_private;
 795         if (rdev == NULL) {
 796                 return -EINVAL;
 797         }
 798         r = ttm_bo_mmap(filp, vma, &rdev->mman.bdev);
 799         if (unlikely(r != 0)) {
 800                 return r;
 801         }
 802         if (unlikely(ttm_vm_ops == NULL)) {
 803                 ttm_vm_ops = vma->vm_ops;
 804                 radeon_ttm_vm_ops = *ttm_vm_ops;
 805                 radeon_ttm_vm_ops.fault = &radeon_ttm_fault;
 806         }
 807         vma->vm_ops = &radeon_ttm_vm_ops;
 808         return 0;
 809 }
 810
 811
 812 #define RADEON_DEBUGFS_MEM_TYPES 2
 813
 814 #if defined(CONFIG_DEBUG_FS)
 815 static int radeon_mm_dump_table(struct seq_file *m, void *data)
 816 {
 817         struct drm_info_node *node = (struct drm_info_node *)m->private;
 818         struct drm_mm *mm = (struct drm_mm *)node->info_ent->data;
 819         struct drm_device *dev = node->minor->dev;
 820         struct radeon_device *rdev = dev->dev_private;
 821         int ret;
 822         struct ttm_bo_global *glob = rdev->mman.bdev.glob;
 823
 824         spin_lock(&glob->lru_lock);
 825         ret = drm_mm_dump_table(m, mm);
 826         spin_unlock(&glob->lru_lock);
 827         return ret;
 828 }
 829 #endif
 830
 831 static int radeon_ttm_debugfs_init(struct radeon_device *rdev)
 832 {
 833 #if defined(CONFIG_DEBUG_FS)
 834         static struct drm_info_list radeon_mem_types_list[RADEON_DEBUGFS_MEM_TYPES+2];
 835         static char radeon_mem_types_names[RADEON_DEBUGFS_MEM_TYPES+2][32];
 836         unsigned i;
 837
 838         for (i = 0; i < RADEON_DEBUGFS_MEM_TYPES; i++) {
 839                 if (i == 0)
 840                         sprintf(radeon_mem_types_names[i], "radeon_vram_mm");
 841                 else
 842                         sprintf(radeon_mem_types_names[i], "radeon_gtt_mm");
 843                 radeon_mem_types_list[i].name = radeon_mem_types_names[i];
 844                 radeon_mem_types_list[i].show = &radeon_mm_dump_table;
 845                 radeon_mem_types_list[i].driver_features = 0;
 846                 if (i == 0)
 847                         radeon_mem_types_list[i].data = rdev->mman.bdev.man[TTM_PL_VRAM].priv;
 848                 else
 849                         radeon_mem_types_list[i].data = rdev->mman.bdev.man[TTM_PL_TT].priv;
 850
 851         }
 852         /* Add ttm page pool to debugfs */
 853         sprintf(radeon_mem_types_names[i], "ttm_page_pool");
 854         radeon_mem_types_list[i].name = radeon_mem_types_names[i];
 855         radeon_mem_types_list[i].show = &ttm_page_alloc_debugfs;
 856         radeon_mem_types_list[i].driver_features = 0;
 857         radeon_mem_types_list[i++].data = NULL;
 858 #ifdef CONFIG_SWIOTLB
 859         if (swiotlb_nr_tbl()) {
 860                 sprintf(radeon_mem_types_names[i], "ttm_dma_page_pool");
 861                 radeon_mem_types_list[i].name = radeon_mem_types_names[i];
 862                 radeon_mem_types_list[i].show = &ttm_dma_page_alloc_debugfs;
 863                 radeon_mem_types_list[i].driver_features = 0;
 864                 radeon_mem_types_list[i++].data = NULL;
 865         }
 866 #endif
 867         return radeon_debugfs_add_files(rdev, radeon_mem_types_list, i);
 868
 869 #endif
 870         return 0;
 871 }