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 <linux/swiotlb.h>
  42 #include <linux/swap.h>
  43 #include <linux/pagemap.h>
  44 #include <linux/debugfs.h>
  45 #include "radeon_reg.h"
  46 #include "radeon.h"
  47
  48 #define DRM_FILE_PAGE_OFFSET (0x100000000ULL >> PAGE_SHIFT)
  49
  50 static int radeon_ttm_debugfs_init(struct radeon_device *rdev);
  51 static void radeon_ttm_debugfs_fini(struct radeon_device *rdev);
  52
  53 static struct radeon_device *radeon_get_rdev(struct ttm_bo_device *bdev)
  54 {
  55         struct radeon_mman *mman;
  56         struct radeon_device *rdev;
  57
  58         mman = container_of(bdev, struct radeon_mman, bdev);
  59         rdev = container_of(mman, struct radeon_device, mman);
  60         return rdev;
  61 }
  62
  63
  64 /*
  65  * Global memory.
  66  */
  67 static int radeon_ttm_mem_global_init(struct drm_global_reference *ref)
  68 {
  69         return ttm_mem_global_init(ref->object);
  70 }
  71
  72 static void radeon_ttm_mem_global_release(struct drm_global_reference *ref)
  73 {
  74         ttm_mem_global_release(ref->object);
  75 }
  76
  77 static int radeon_ttm_global_init(struct radeon_device *rdev)
  78 {
  79         struct drm_global_reference *global_ref;
  80         int r;
  81
  82         rdev->mman.mem_global_referenced = false;
  83         global_ref = &rdev->mman.mem_global_ref;
  84         global_ref->global_type = DRM_GLOBAL_TTM_MEM;
  85         global_ref->size = sizeof(struct ttm_mem_global);
  86         global_ref->init = &radeon_ttm_mem_global_init;
  87         global_ref->release = &radeon_ttm_mem_global_release;
  88         r = drm_global_item_ref(global_ref);
  89         if (r != 0) {
  90                 DRM_ERROR("Failed setting up TTM memory accounting "
  91                           "subsystem.\n");
  92                 return r;
  93         }
  94
  95         rdev->mman.bo_global_ref.mem_glob =
  96                 rdev->mman.mem_global_ref.object;
  97         global_ref = &rdev->mman.bo_global_ref.ref;
  98         global_ref->global_type = DRM_GLOBAL_TTM_BO;
  99         global_ref->size = sizeof(struct ttm_bo_global);
 100         global_ref->init = &ttm_bo_global_init;
 101         global_ref->release = &ttm_bo_global_release;
 102         r = drm_global_item_ref(global_ref);
 103         if (r != 0) {
 104                 DRM_ERROR("Failed setting up TTM BO subsystem.\n");
 105                 drm_global_item_unref(&rdev->mman.mem_global_ref);
 106                 return r;
 107         }
 108
 109         rdev->mman.mem_global_referenced = true;
 110         return 0;
 111 }
 112
 113 static void radeon_ttm_global_fini(struct radeon_device *rdev)
 114 {
 115         if (rdev->mman.mem_global_referenced) {
 116                 drm_global_item_unref(&rdev->mman.bo_global_ref.ref);
 117                 drm_global_item_unref(&rdev->mman.mem_global_ref);
 118                 rdev->mman.mem_global_referenced = false;
 119         }
 120 }
 121
 122 static int radeon_invalidate_caches(struct ttm_bo_device *bdev, uint32_t flags)
 123 {
 124         return 0;
 125 }
 126
 127 static int radeon_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
 128                                 struct ttm_mem_type_manager *man)
 129 {
 130         struct radeon_device *rdev;
 131
 132         rdev = radeon_get_rdev(bdev);
 133
 134         switch (type) {
 135         case TTM_PL_SYSTEM:
 136                 /* System memory */
 137                 man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
 138                 man->available_caching = TTM_PL_MASK_CACHING;
 139                 man->default_caching = TTM_PL_FLAG_CACHED;
 140                 break;
 141         case TTM_PL_TT:
 142                 man->func = &ttm_bo_manager_func;
 143                 man->gpu_offset = rdev->mc.gtt_start;
 144                 man->available_caching = TTM_PL_MASK_CACHING;
 145                 man->default_caching = TTM_PL_FLAG_CACHED;
 146                 man->flags = TTM_MEMTYPE_FLAG_MAPPABLE | TTM_MEMTYPE_FLAG_CMA;
 147 #if __OS_HAS_AGP
 148                 if (rdev->flags & RADEON_IS_AGP) {
 149                         if (!rdev->ddev->agp) {
 150                                 DRM_ERROR("AGP is not enabled for memory type %u\n",
 151                                           (unsigned)type);
 152                                 return -EINVAL;
 153                         }
 154                         if (!rdev->ddev->agp->cant_use_aperture)
 155                                 man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
 156                         man->available_caching = TTM_PL_FLAG_UNCACHED |
 157                                                  TTM_PL_FLAG_WC;
 158                         man->default_caching = TTM_PL_FLAG_WC;
 159                 }
 160 #endif
 161                 break;
 162         case TTM_PL_VRAM:
 163                 /* "On-card" video ram */
 164                 man->func = &ttm_bo_manager_func;
 165                 man->gpu_offset = rdev->mc.vram_start;
 166                 man->flags = TTM_MEMTYPE_FLAG_FIXED |
 167                              TTM_MEMTYPE_FLAG_MAPPABLE;
 168                 man->available_caching = TTM_PL_FLAG_UNCACHED | TTM_PL_FLAG_WC;
 169                 man->default_caching = TTM_PL_FLAG_WC;
 170                 break;
 171         default:
 172                 DRM_ERROR("Unsupported memory type %u\n", (unsigned)type);
 173                 return -EINVAL;
 174         }
 175         return 0;
 176 }
 177
 178 static void radeon_evict_flags(struct ttm_buffer_object *bo,
 179                                 struct ttm_placement *placement)
 180 {
 181         static struct ttm_place placements = {
 182                 .fpfn = 0,
 183                 .lpfn = 0,
 184                 .flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM
 185         };
 186
 187         struct radeon_bo *rbo;
 188
 189         if (!radeon_ttm_bo_is_radeon_bo(bo)) {
 190                 placement->placement = &placements;
 191                 placement->busy_placement = &placements;
 192                 placement->num_placement = 1;
 193                 placement->num_busy_placement = 1;
 194                 return;
 195         }
 196         rbo = container_of(bo, struct radeon_bo, tbo);
 197         switch (bo->mem.mem_type) {
 198         case TTM_PL_VRAM:
 199                 if (rbo->rdev->ring[RADEON_RING_TYPE_GFX_INDEX].ready == false)
 200                         radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_CPU);
 201                 else
 202                         radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_GTT);
 203                 break;
 204         case TTM_PL_TT:
 205         default:
 206                 radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_CPU);
 207         }
 208         *placement = rbo->placement;
 209 }
 210
 211 static int radeon_verify_access(struct ttm_buffer_object *bo, struct file *filp)
 212 {
 213         struct radeon_bo *rbo = container_of(bo, struct radeon_bo, tbo);
 214
 215         return drm_vma_node_verify_access(&rbo->gem_base.vma_node, filp);
 216 }
 217
 218 static void radeon_move_null(struct ttm_buffer_object *bo,
 219                              struct ttm_mem_reg *new_mem)
 220 {
 221         struct ttm_mem_reg *old_mem = &bo->mem;
 222
 223         BUG_ON(old_mem->mm_node != NULL);
 224         *old_mem = *new_mem;
 225         new_mem->mm_node = NULL;
 226 }
 227
 228 static int radeon_move_blit(struct ttm_buffer_object *bo,
 229                         bool evict, bool no_wait_gpu,
 230                         struct ttm_mem_reg *new_mem,
 231                         struct ttm_mem_reg *old_mem)
 232 {
 233         struct radeon_device *rdev;
 234         uint64_t old_start, new_start;
 235         struct radeon_fence *fence;
 236         int r, ridx;
 237
 238         rdev = radeon_get_rdev(bo->bdev);
 239         ridx = radeon_copy_ring_index(rdev);
 240         old_start = old_mem->start << PAGE_SHIFT;
 241         new_start = new_mem->start << PAGE_SHIFT;
 242
 243         switch (old_mem->mem_type) {
 244         case TTM_PL_VRAM:
 245                 old_start += rdev->mc.vram_start;
 246                 break;
 247         case TTM_PL_TT:
 248                 old_start += rdev->mc.gtt_start;
 249                 break;
 250         default:
 251                 DRM_ERROR("Unknown placement %d\n", old_mem->mem_type);
 252                 return -EINVAL;
 253         }
 254         switch (new_mem->mem_type) {
 255         case TTM_PL_VRAM:
 256                 new_start += rdev->mc.vram_start;
 257                 break;
 258         case TTM_PL_TT:
 259                 new_start += rdev->mc.gtt_start;
 260                 break;
 261         default:
 262                 DRM_ERROR("Unknown placement %d\n", old_mem->mem_type);
 263                 return -EINVAL;
 264         }
 265         if (!rdev->ring[ridx].ready) {
 266                 DRM_ERROR("Trying to move memory with ring turned off.\n");
 267                 return -EINVAL;
 268         }
 269
 270         BUILD_BUG_ON((PAGE_SIZE % RADEON_GPU_PAGE_SIZE) != 0);
 271
 272         /* sync other rings */
 273         fence = bo->sync_obj;
 274         r = radeon_copy(rdev, old_start, new_start,
 275                         new_mem->num_pages * (PAGE_SIZE / RADEON_GPU_PAGE_SIZE), /* GPU pages */
 276                         &fence);
 277         /* FIXME: handle copy error */
 278         r = ttm_bo_move_accel_cleanup(bo, (void *)fence,
 279                                       evict, no_wait_gpu, new_mem);
 280         radeon_fence_unref(&fence);
 281         return r;
 282 }
 283
 284 static int radeon_move_vram_ram(struct ttm_buffer_object *bo,
 285                                 bool evict, bool interruptible,
 286                                 bool no_wait_gpu,
 287                                 struct ttm_mem_reg *new_mem)
 288 {
 289         struct radeon_device *rdev;
 290         struct ttm_mem_reg *old_mem = &bo->mem;
 291         struct ttm_mem_reg tmp_mem;
 292         struct ttm_place placements;
 293         struct ttm_placement placement;
 294         int r;
 295
 296         rdev = radeon_get_rdev(bo->bdev);
 297         tmp_mem = *new_mem;
 298         tmp_mem.mm_node = NULL;
 299         placement.num_placement = 1;
 300         placement.placement = &placements;
 301         placement.num_busy_placement = 1;
 302         placement.busy_placement = &placements;
 303         placements.fpfn = 0;
 304         placements.lpfn = 0;
 305         placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
 306         r = ttm_bo_mem_space(bo, &placement, &tmp_mem,
 307                              interruptible, no_wait_gpu);
 308         if (unlikely(r)) {
 309                 return r;
 310         }
 311
 312         r = ttm_tt_set_placement_caching(bo->ttm, tmp_mem.placement);
 313         if (unlikely(r)) {
 314                 goto out_cleanup;
 315         }
 316
 317         r = ttm_tt_bind(bo->ttm, &tmp_mem);
 318         if (unlikely(r)) {
 319                 goto out_cleanup;
 320         }
 321         r = radeon_move_blit(bo, true, no_wait_gpu, &tmp_mem, old_mem);
 322         if (unlikely(r)) {
 323                 goto out_cleanup;
 324         }
 325         r = ttm_bo_move_ttm(bo, true, no_wait_gpu, new_mem);
 326 out_cleanup:
 327         ttm_bo_mem_put(bo, &tmp_mem);
 328         return r;
 329 }
 330
 331 static int radeon_move_ram_vram(struct ttm_buffer_object *bo,
 332                                 bool evict, bool interruptible,
 333                                 bool no_wait_gpu,
 334                                 struct ttm_mem_reg *new_mem)
 335 {
 336         struct radeon_device *rdev;
 337         struct ttm_mem_reg *old_mem = &bo->mem;
 338         struct ttm_mem_reg tmp_mem;
 339         struct ttm_placement placement;
 340         struct ttm_place placements;
 341         int r;
 342
 343         rdev = radeon_get_rdev(bo->bdev);
 344         tmp_mem = *new_mem;
 345         tmp_mem.mm_node = NULL;
 346         placement.num_placement = 1;
 347         placement.placement = &placements;
 348         placement.num_busy_placement = 1;
 349         placement.busy_placement = &placements;
 350         placements.fpfn = 0;
 351         placements.lpfn = 0;
 352         placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
 353         r = ttm_bo_mem_space(bo, &placement, &tmp_mem,
 354                              interruptible, no_wait_gpu);
 355         if (unlikely(r)) {
 356                 return r;
 357         }
 358         r = ttm_bo_move_ttm(bo, true, no_wait_gpu, &tmp_mem);
 359         if (unlikely(r)) {
 360                 goto out_cleanup;
 361         }
 362         r = radeon_move_blit(bo, true, no_wait_gpu, new_mem, old_mem);
 363         if (unlikely(r)) {
 364                 goto out_cleanup;
 365         }
 366 out_cleanup:
 367         ttm_bo_mem_put(bo, &tmp_mem);
 368         return r;
 369 }
 370
 371 static int radeon_bo_move(struct ttm_buffer_object *bo,
 372                         bool evict, bool interruptible,
 373                         bool no_wait_gpu,
 374                         struct ttm_mem_reg *new_mem)
 375 {
 376         struct radeon_device *rdev;
 377         struct ttm_mem_reg *old_mem = &bo->mem;
 378         int r;
 379
 380         rdev = radeon_get_rdev(bo->bdev);
 381         if (old_mem->mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
 382                 radeon_move_null(bo, new_mem);
 383                 return 0;
 384         }
 385         if ((old_mem->mem_type == TTM_PL_TT &&
 386              new_mem->mem_type == TTM_PL_SYSTEM) ||
 387             (old_mem->mem_type == TTM_PL_SYSTEM &&
 388              new_mem->mem_type == TTM_PL_TT)) {
 389                 /* bind is enough */
 390                 radeon_move_null(bo, new_mem);
 391                 return 0;
 392         }
 393         if (!rdev->ring[radeon_copy_ring_index(rdev)].ready ||
 394             rdev->asic->copy.copy == NULL) {
 395                 /* use memcpy */
 396                 goto memcpy;
 397         }
 398
 399         if (old_mem->mem_type == TTM_PL_VRAM &&
 400             new_mem->mem_type == TTM_PL_SYSTEM) {
 401                 r = radeon_move_vram_ram(bo, evict, interruptible,
 402                                         no_wait_gpu, new_mem);
 403         } else if (old_mem->mem_type == TTM_PL_SYSTEM &&
 404                    new_mem->mem_type == TTM_PL_VRAM) {
 405                 r = radeon_move_ram_vram(bo, evict, interruptible,
 406                                             no_wait_gpu, new_mem);
 407         } else {
 408                 r = radeon_move_blit(bo, evict, no_wait_gpu, new_mem, old_mem);
 409         }
 410
 411         if (r) {
 412 memcpy:
 413                 r = ttm_bo_move_memcpy(bo, evict, no_wait_gpu, new_mem);
 414                 if (r) {
 415                         return r;
 416                 }
 417         }
 418
 419         /* update statistics */
 420         atomic64_add((u64)bo->num_pages << PAGE_SHIFT, &rdev->num_bytes_moved);
 421         return 0;
 422 }
 423
 424 static int radeon_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
 425 {
 426         struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
 427         struct radeon_device *rdev = radeon_get_rdev(bdev);
 428
 429         mem->bus.addr = NULL;
 430         mem->bus.offset = 0;
 431         mem->bus.size = mem->num_pages << PAGE_SHIFT;
 432         mem->bus.base = 0;
 433         mem->bus.is_iomem = false;
 434         if (!(man->flags & TTM_MEMTYPE_FLAG_MAPPABLE))
 435                 return -EINVAL;
 436         switch (mem->mem_type) {
 437         case TTM_PL_SYSTEM:
 438                 /* system memory */
 439                 return 0;
 440         case TTM_PL_TT:
 441 #if __OS_HAS_AGP
 442                 if (rdev->flags & RADEON_IS_AGP) {
 443                         /* RADEON_IS_AGP is set only if AGP is active */
 444                         mem->bus.offset = mem->start << PAGE_SHIFT;
 445                         mem->bus.base = rdev->mc.agp_base;
 446                         mem->bus.is_iomem = !rdev->ddev->agp->cant_use_aperture;
 447                 }
 448 #endif
 449                 break;
 450         case TTM_PL_VRAM:
 451                 mem->bus.offset = mem->start << PAGE_SHIFT;
 452                 /* check if it's visible */
 453                 if ((mem->bus.offset + mem->bus.size) > rdev->mc.visible_vram_size)
 454                         return -EINVAL;
 455                 mem->bus.base = rdev->mc.aper_base;
 456                 mem->bus.is_iomem = true;
 457 #ifdef __alpha__
 458                 /*
 459                  * Alpha: use bus.addr to hold the ioremap() return,
 460                  * so we can modify bus.base below.
 461                  */
 462                 if (mem->placement & TTM_PL_FLAG_WC)
 463                         mem->bus.addr =
 464                                 ioremap_wc(mem->bus.base + mem->bus.offset,
 465                                            mem->bus.size);
 466                 else
 467                         mem->bus.addr =
 468                                 ioremap_nocache(mem->bus.base + mem->bus.offset,
 469                                                 mem->bus.size);
 470
 471                 /*
 472                  * Alpha: Use just the bus offset plus
 473                  * the hose/domain memory base for bus.base.
 474                  * It then can be used to build PTEs for VRAM
 475                  * access, as done in ttm_bo_vm_fault().
 476                  */
 477                 mem->bus.base = (mem->bus.base & 0x0ffffffffUL) +
 478                         rdev->ddev->hose->dense_mem_base;
 479 #endif
 480                 break;
 481         default:
 482                 return -EINVAL;
 483         }
 484         return 0;
 485 }
 486
 487 static void radeon_ttm_io_mem_free(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
 488 {
 489 }
 490
 491 static int radeon_sync_obj_wait(void *sync_obj, bool lazy, bool interruptible)
 492 {
 493         return radeon_fence_wait((struct radeon_fence *)sync_obj, interruptible);
 494 }
 495
 496 static int radeon_sync_obj_flush(void *sync_obj)
 497 {
 498         return 0;
 499 }
 500
 501 static void radeon_sync_obj_unref(void **sync_obj)
 502 {
 503         radeon_fence_unref((struct radeon_fence **)sync_obj);
 504 }
 505
 506 static void *radeon_sync_obj_ref(void *sync_obj)
 507 {
 508         return radeon_fence_ref((struct radeon_fence *)sync_obj);
 509 }
 510
 511 static bool radeon_sync_obj_signaled(void *sync_obj)
 512 {
 513         return radeon_fence_signaled((struct radeon_fence *)sync_obj);
 514 }
 515
 516 /*
 517  * TTM backend functions.
 518  */
 519 struct radeon_ttm_tt {
 520         struct ttm_dma_tt               ttm;
 521         struct radeon_device            *rdev;
 522         u64                             offset;
 523
 524         uint64_t                        userptr;
 525         struct mm_struct                *usermm;
 526         uint32_t                        userflags;
 527 };
 528
 529 /* prepare the sg table with the user pages */
 530 static int radeon_ttm_tt_pin_userptr(struct ttm_tt *ttm)
 531 {
 532         struct radeon_device *rdev = radeon_get_rdev(ttm->bdev);
 533         struct radeon_ttm_tt *gtt = (void *)ttm;
 534         unsigned pinned = 0, nents;
 535         int r;
 536
 537         int write = !(gtt->userflags & RADEON_GEM_USERPTR_READONLY);
 538         enum dma_data_direction direction = write ?
 539                 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
 540
 541         if (current->mm != gtt->usermm)
 542                 return -EPERM;
 543
 544         if (gtt->userflags & RADEON_GEM_USERPTR_ANONONLY) {
 545                 /* check that we only pin down anonymous memory
 546                    to prevent problems with writeback */
 547                 unsigned long end = gtt->userptr + ttm->num_pages * PAGE_SIZE;
 548                 struct vm_area_struct *vma;
 549                 vma = find_vma(gtt->usermm, gtt->userptr);
 550                 if (!vma || vma->vm_file || vma->vm_end < end)
 551                         return -EPERM;
 552         }
 553
 554         do {
 555                 unsigned num_pages = ttm->num_pages - pinned;
 556                 uint64_t userptr = gtt->userptr + pinned * PAGE_SIZE;
 557                 struct page **pages = ttm->pages + pinned;
 558
 559                 r = get_user_pages(current, current->mm, userptr, num_pages,
 560                                    write, 0, pages, NULL);
 561                 if (r < 0)
 562                         goto release_pages;
 563
 564                 pinned += r;
 565
 566         } while (pinned < ttm->num_pages);
 567
 568         r = sg_alloc_table_from_pages(ttm->sg, ttm->pages, ttm->num_pages, 0,
 569                                       ttm->num_pages << PAGE_SHIFT,
 570                                       GFP_KERNEL);
 571         if (r)
 572                 goto release_sg;
 573
 574         r = -ENOMEM;
 575         nents = dma_map_sg(rdev->dev, ttm->sg->sgl, ttm->sg->nents, direction);
 576         if (nents != ttm->sg->nents)
 577                 goto release_sg;
 578
 579         drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
 580                                          gtt->ttm.dma_address, ttm->num_pages);
 581
 582         return 0;
 583
 584 release_sg:
 585         kfree(ttm->sg);
 586
 587 release_pages:
 588         release_pages(ttm->pages, pinned, 0);
 589         return r;
 590 }
 591
 592 static void radeon_ttm_tt_unpin_userptr(struct ttm_tt *ttm)
 593 {
 594         struct radeon_device *rdev = radeon_get_rdev(ttm->bdev);
 595         struct radeon_ttm_tt *gtt = (void *)ttm;
 596         struct scatterlist *sg;
 597         int i;
 598
 599         int write = !(gtt->userflags & RADEON_GEM_USERPTR_READONLY);
 600         enum dma_data_direction direction = write ?
 601                 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
 602
 603         /* free the sg table and pages again */
 604         dma_unmap_sg(rdev->dev, ttm->sg->sgl, ttm->sg->nents, direction);
 605
 606         for_each_sg(ttm->sg->sgl, sg, ttm->sg->nents, i) {
 607                 struct page *page = sg_page(sg);
 608
 609                 if (!(gtt->userflags & RADEON_GEM_USERPTR_READONLY))
 610                         set_page_dirty(page);
 611
 612                 mark_page_accessed(page);
 613                 page_cache_release(page);
 614         }
 615
 616         sg_free_table(ttm->sg);
 617 }
 618
 619 static int radeon_ttm_backend_bind(struct ttm_tt *ttm,
 620                                    struct ttm_mem_reg *bo_mem)
 621 {
 622         struct radeon_ttm_tt *gtt = (void*)ttm;
 623         uint32_t flags = RADEON_GART_PAGE_VALID | RADEON_GART_PAGE_READ |
 624                 RADEON_GART_PAGE_WRITE;
 625         int r;
 626
 627         if (gtt->userptr) {
 628                 radeon_ttm_tt_pin_userptr(ttm);
 629                 flags &= ~RADEON_GART_PAGE_WRITE;
 630         }
 631
 632         gtt->offset = (unsigned long)(bo_mem->start << PAGE_SHIFT);
 633         if (!ttm->num_pages) {
 634                 WARN(1, "nothing to bind %lu pages for mreg %p back %p!\n",
 635                      ttm->num_pages, bo_mem, ttm);
 636         }
 637         if (ttm->caching_state == tt_cached)
 638                 flags |= RADEON_GART_PAGE_SNOOP;
 639         r = radeon_gart_bind(gtt->rdev, gtt->offset, ttm->num_pages,
 640                              ttm->pages, gtt->ttm.dma_address, flags);
 641         if (r) {
 642                 DRM_ERROR("failed to bind %lu pages at 0x%08X\n",
 643                           ttm->num_pages, (unsigned)gtt->offset);
 644                 return r;
 645         }
 646         return 0;
 647 }
 648
 649 static int radeon_ttm_backend_unbind(struct ttm_tt *ttm)
 650 {
 651         struct radeon_ttm_tt *gtt = (void *)ttm;
 652
 653         radeon_gart_unbind(gtt->rdev, gtt->offset, ttm->num_pages);
 654
 655         if (gtt->userptr)
 656                 radeon_ttm_tt_unpin_userptr(ttm);
 657
 658         return 0;
 659 }
 660
 661 static void radeon_ttm_backend_destroy(struct ttm_tt *ttm)
 662 {
 663         struct radeon_ttm_tt *gtt = (void *)ttm;
 664
 665         ttm_dma_tt_fini(&gtt->ttm);
 666         kfree(gtt);
 667 }
 668
 669 static struct ttm_backend_func radeon_backend_func = {
 670         .bind = &radeon_ttm_backend_bind,
 671         .unbind = &radeon_ttm_backend_unbind,
 672         .destroy = &radeon_ttm_backend_destroy,
 673 };
 674
 675 static struct ttm_tt *radeon_ttm_tt_create(struct ttm_bo_device *bdev,
 676                                     unsigned long size, uint32_t page_flags,
 677                                     struct page *dummy_read_page)
 678 {
 679         struct radeon_device *rdev;
 680         struct radeon_ttm_tt *gtt;
 681
 682         rdev = radeon_get_rdev(bdev);
 683 #if __OS_HAS_AGP
 684         if (rdev->flags & RADEON_IS_AGP) {
 685                 return ttm_agp_tt_create(bdev, rdev->ddev->agp->bridge,
 686                                          size, page_flags, dummy_read_page);
 687         }
 688 #endif
 689
 690         gtt = kzalloc(sizeof(struct radeon_ttm_tt), GFP_KERNEL);
 691         if (gtt == NULL) {
 692                 return NULL;
 693         }
 694         gtt->ttm.ttm.func = &radeon_backend_func;
 695         gtt->rdev = rdev;
 696         if (ttm_dma_tt_init(&gtt->ttm, bdev, size, page_flags, dummy_read_page)) {
 697                 kfree(gtt);
 698                 return NULL;
 699         }
 700         return &gtt->ttm.ttm;
 701 }
 702
 703 static int radeon_ttm_tt_populate(struct ttm_tt *ttm)
 704 {
 705         struct radeon_device *rdev;
 706         struct radeon_ttm_tt *gtt = (void *)ttm;
 707         unsigned i;
 708         int r;
 709         bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
 710
 711         if (ttm->state != tt_unpopulated)
 712                 return 0;
 713
 714         if (gtt->userptr) {
 715                 ttm->sg = kcalloc(1, sizeof(struct sg_table), GFP_KERNEL);
 716                 if (!ttm->sg)
 717                         return -ENOMEM;
 718
 719                 ttm->page_flags |= TTM_PAGE_FLAG_SG;
 720                 ttm->state = tt_unbound;
 721                 return 0;
 722         }
 723
 724         if (slave && ttm->sg) {
 725                 drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
 726                                                  gtt->ttm.dma_address, ttm->num_pages);
 727                 ttm->state = tt_unbound;
 728                 return 0;
 729         }
 730
 731         rdev = radeon_get_rdev(ttm->bdev);
 732 #if __OS_HAS_AGP
 733         if (rdev->flags & RADEON_IS_AGP) {
 734                 return ttm_agp_tt_populate(ttm);
 735         }
 736 #endif
 737
 738 #ifdef CONFIG_SWIOTLB
 739         if (swiotlb_nr_tbl()) {
 740                 return ttm_dma_populate(&gtt->ttm, rdev->dev);
 741         }
 742 #endif
 743
 744         r = ttm_pool_populate(ttm);
 745         if (r) {
 746                 return r;
 747         }
 748
 749         for (i = 0; i < ttm->num_pages; i++) {
 750                 gtt->ttm.dma_address[i] = pci_map_page(rdev->pdev, ttm->pages[i],
 751                                                        0, PAGE_SIZE,
 752                                                        PCI_DMA_BIDIRECTIONAL);
 753                 if (pci_dma_mapping_error(rdev->pdev, gtt->ttm.dma_address[i])) {
 754                         while (--i) {
 755                                 pci_unmap_page(rdev->pdev, gtt->ttm.dma_address[i],
 756                                                PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
 757                                 gtt->ttm.dma_address[i] = 0;
 758                         }
 759                         ttm_pool_unpopulate(ttm);
 760                         return -EFAULT;
 761                 }
 762         }
 763         return 0;
 764 }
 765
 766 static void radeon_ttm_tt_unpopulate(struct ttm_tt *ttm)
 767 {
 768         struct radeon_device *rdev;
 769         struct radeon_ttm_tt *gtt = (void *)ttm;
 770         unsigned i;
 771         bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
 772
 773         if (gtt->userptr) {
 774                 kfree(ttm->sg);
 775                 ttm->page_flags &= ~TTM_PAGE_FLAG_SG;
 776                 return;
 777         }
 778
 779         if (slave)
 780                 return;
 781
 782         rdev = radeon_get_rdev(ttm->bdev);
 783 #if __OS_HAS_AGP
 784         if (rdev->flags & RADEON_IS_AGP) {
 785                 ttm_agp_tt_unpopulate(ttm);
 786                 return;
 787         }
 788 #endif
 789
 790 #ifdef CONFIG_SWIOTLB
 791         if (swiotlb_nr_tbl()) {
 792                 ttm_dma_unpopulate(&gtt->ttm, rdev->dev);
 793                 return;
 794         }
 795 #endif
 796
 797         for (i = 0; i < ttm->num_pages; i++) {
 798                 if (gtt->ttm.dma_address[i]) {
 799                         pci_unmap_page(rdev->pdev, gtt->ttm.dma_address[i],
 800                                        PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
 801                 }
 802         }
 803
 804         ttm_pool_unpopulate(ttm);
 805 }
 806
 807 int radeon_ttm_tt_set_userptr(struct ttm_tt *ttm, uint64_t addr,
 808                               uint32_t flags)
 809 {
 810         struct radeon_ttm_tt *gtt = (void *)ttm;
 811
 812         if (gtt == NULL)
 813                 return -EINVAL;
 814
 815         gtt->userptr = addr;
 816         gtt->usermm = current->mm;
 817         gtt->userflags = flags;
 818         return 0;
 819 }
 820
 821 bool radeon_ttm_tt_has_userptr(struct ttm_tt *ttm)
 822 {
 823         struct radeon_ttm_tt *gtt = (void *)ttm;
 824
 825         if (gtt == NULL)
 826                 return false;
 827
 828         return !!gtt->userptr;
 829 }
 830
 831 bool radeon_ttm_tt_is_readonly(struct ttm_tt *ttm)
 832 {
 833         struct radeon_ttm_tt *gtt = (void *)ttm;
 834
 835         if (gtt == NULL)
 836                 return false;
 837
 838         return !!(gtt->userflags & RADEON_GEM_USERPTR_READONLY);
 839 }
 840
 841 static struct ttm_bo_driver radeon_bo_driver = {
 842         .ttm_tt_create = &radeon_ttm_tt_create,
 843         .ttm_tt_populate = &radeon_ttm_tt_populate,
 844         .ttm_tt_unpopulate = &radeon_ttm_tt_unpopulate,
 845         .invalidate_caches = &radeon_invalidate_caches,
 846         .init_mem_type = &radeon_init_mem_type,
 847         .evict_flags = &radeon_evict_flags,
 848         .move = &radeon_bo_move,
 849         .verify_access = &radeon_verify_access,
 850         .sync_obj_signaled = &radeon_sync_obj_signaled,
 851         .sync_obj_wait = &radeon_sync_obj_wait,
 852         .sync_obj_flush = &radeon_sync_obj_flush,
 853         .sync_obj_unref = &radeon_sync_obj_unref,
 854         .sync_obj_ref = &radeon_sync_obj_ref,
 855         .move_notify = &radeon_bo_move_notify,
 856         .fault_reserve_notify = &radeon_bo_fault_reserve_notify,
 857         .io_mem_reserve = &radeon_ttm_io_mem_reserve,
 858         .io_mem_free = &radeon_ttm_io_mem_free,
 859 };
 860
 861 int radeon_ttm_init(struct radeon_device *rdev)
 862 {
 863         int r;
 864
 865         r = radeon_ttm_global_init(rdev);
 866         if (r) {
 867                 return r;
 868         }
 869         /* No others user of address space so set it to 0 */
 870         r = ttm_bo_device_init(&rdev->mman.bdev,
 871                                rdev->mman.bo_global_ref.ref.object,
 872                                &radeon_bo_driver,
 873                                rdev->ddev->anon_inode->i_mapping,
 874                                DRM_FILE_PAGE_OFFSET,
 875                                rdev->need_dma32);
 876         if (r) {
 877                 DRM_ERROR("failed initializing buffer object driver(%d).\n", r);
 878                 return r;
 879         }
 880         rdev->mman.initialized = true;
 881         r = ttm_bo_init_mm(&rdev->mman.bdev, TTM_PL_VRAM,
 882                                 rdev->mc.real_vram_size >> PAGE_SHIFT);
 883         if (r) {
 884                 DRM_ERROR("Failed initializing VRAM heap.\n");
 885                 return r;
 886         }
 887         /* Change the size here instead of the init above so only lpfn is affected */
 888         radeon_ttm_set_active_vram_size(rdev, rdev->mc.visible_vram_size);
 889
 890         r = radeon_bo_create(rdev, 256 * 1024, PAGE_SIZE, true,
 891                              RADEON_GEM_DOMAIN_VRAM, 0,
 892                              NULL, &rdev->stollen_vga_memory);
 893         if (r) {
 894                 return r;
 895         }
 896         r = radeon_bo_reserve(rdev->stollen_vga_memory, false);
 897         if (r)
 898                 return r;
 899         r = radeon_bo_pin(rdev->stollen_vga_memory, RADEON_GEM_DOMAIN_VRAM, NULL);
 900         radeon_bo_unreserve(rdev->stollen_vga_memory);
 901         if (r) {
 902                 radeon_bo_unref(&rdev->stollen_vga_memory);
 903                 return r;
 904         }
 905         DRM_INFO("radeon: %uM of VRAM memory ready\n",
 906                  (unsigned) (rdev->mc.real_vram_size / (1024 * 1024)));
 907         r = ttm_bo_init_mm(&rdev->mman.bdev, TTM_PL_TT,
 908                                 rdev->mc.gtt_size >> PAGE_SHIFT);
 909         if (r) {
 910                 DRM_ERROR("Failed initializing GTT heap.\n");
 911                 return r;
 912         }
 913         DRM_INFO("radeon: %uM of GTT memory ready.\n",
 914                  (unsigned)(rdev->mc.gtt_size / (1024 * 1024)));
 915
 916         r = radeon_ttm_debugfs_init(rdev);
 917         if (r) {
 918                 DRM_ERROR("Failed to init debugfs\n");
 919                 return r;
 920         }
 921         return 0;
 922 }
 923
 924 void radeon_ttm_fini(struct radeon_device *rdev)
 925 {
 926         int r;
 927
 928         if (!rdev->mman.initialized)
 929                 return;
 930         radeon_ttm_debugfs_fini(rdev);
 931         if (rdev->stollen_vga_memory) {
 932                 r = radeon_bo_reserve(rdev->stollen_vga_memory, false);
 933                 if (r == 0) {
 934                         radeon_bo_unpin(rdev->stollen_vga_memory);
 935                         radeon_bo_unreserve(rdev->stollen_vga_memory);
 936                 }
 937                 radeon_bo_unref(&rdev->stollen_vga_memory);
 938         }
 939         ttm_bo_clean_mm(&rdev->mman.bdev, TTM_PL_VRAM);
 940         ttm_bo_clean_mm(&rdev->mman.bdev, TTM_PL_TT);
 941         ttm_bo_device_release(&rdev->mman.bdev);
 942         radeon_gart_fini(rdev);
 943         radeon_ttm_global_fini(rdev);
 944         rdev->mman.initialized = false;
 945         DRM_INFO("radeon: ttm finalized\n");
 946 }
 947
 948 /* this should only be called at bootup or when userspace
 949  * isn't running */
 950 void radeon_ttm_set_active_vram_size(struct radeon_device *rdev, u64 size)
 951 {
 952         struct ttm_mem_type_manager *man;
 953
 954         if (!rdev->mman.initialized)
 955                 return;
 956
 957         man = &rdev->mman.bdev.man[TTM_PL_VRAM];
 958         /* this just adjusts TTM size idea, which sets lpfn to the correct value */
 959         man->size = size >> PAGE_SHIFT;
 960 }
 961
 962 static struct vm_operations_struct radeon_ttm_vm_ops;
 963 static const struct vm_operations_struct *ttm_vm_ops = NULL;
 964
 965 static int radeon_ttm_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 966 {
 967         struct ttm_buffer_object *bo;
 968         struct radeon_device *rdev;
 969         int r;
 970
 971         bo = (struct ttm_buffer_object *)vma->vm_private_data;
 972         if (bo == NULL) {
 973                 return VM_FAULT_NOPAGE;
 974         }
 975         rdev = radeon_get_rdev(bo->bdev);
 976         down_read(&rdev->pm.mclk_lock);
 977         r = ttm_vm_ops->fault(vma, vmf);
 978         up_read(&rdev->pm.mclk_lock);
 979         return r;
 980 }
 981
 982 int radeon_mmap(struct file *filp, struct vm_area_struct *vma)
 983 {
 984         struct drm_file *file_priv;
 985         struct radeon_device *rdev;
 986         int r;
 987
 988         if (unlikely(vma->vm_pgoff < DRM_FILE_PAGE_OFFSET)) {
 989                 return drm_mmap(filp, vma);
 990         }
 991
 992         file_priv = filp->private_data;
 993         rdev = file_priv->minor->dev->dev_private;
 994         if (rdev == NULL) {
 995                 return -EINVAL;
 996         }
 997         r = ttm_bo_mmap(filp, vma, &rdev->mman.bdev);
 998         if (unlikely(r != 0)) {
 999                 return r;
1000         }
1001         if (unlikely(ttm_vm_ops == NULL)) {
1002                 ttm_vm_ops = vma->vm_ops;
1003                 radeon_ttm_vm_ops = *ttm_vm_ops;
1004                 radeon_ttm_vm_ops.fault = &radeon_ttm_fault;
1005         }
1006         vma->vm_ops = &radeon_ttm_vm_ops;
1007         return 0;
1008 }
1009
1010 #if defined(CONFIG_DEBUG_FS)
1011
1012 static int radeon_mm_dump_table(struct seq_file *m, void *data)
1013 {
1014         struct drm_info_node *node = (struct drm_info_node *)m->private;
1015         unsigned ttm_pl = *(int *)node->info_ent->data;
1016         struct drm_device *dev = node->minor->dev;
1017         struct radeon_device *rdev = dev->dev_private;
1018         struct drm_mm *mm = (struct drm_mm *)rdev->mman.bdev.man[ttm_pl].priv;
1019         int ret;
1020         struct ttm_bo_global *glob = rdev->mman.bdev.glob;
1021
1022         spin_lock(&glob->lru_lock);
1023         ret = drm_mm_dump_table(m, mm);
1024         spin_unlock(&glob->lru_lock);
1025         return ret;
1026 }
1027
1028 static int ttm_pl_vram = TTM_PL_VRAM;
1029 static int ttm_pl_tt = TTM_PL_TT;
1030
1031 static struct drm_info_list radeon_ttm_debugfs_list[] = {
1032         {"radeon_vram_mm", radeon_mm_dump_table, 0, &ttm_pl_vram},
1033         {"radeon_gtt_mm", radeon_mm_dump_table, 0, &ttm_pl_tt},
1034         {"ttm_page_pool", ttm_page_alloc_debugfs, 0, NULL},
1035 #ifdef CONFIG_SWIOTLB
1036         {"ttm_dma_page_pool", ttm_dma_page_alloc_debugfs, 0, NULL}
1037 #endif
1038 };
1039
1040 static int radeon_ttm_vram_open(struct inode *inode, struct file *filep)
1041 {
1042         struct radeon_device *rdev = inode->i_private;
1043         i_size_write(inode, rdev->mc.mc_vram_size);
1044         filep->private_data = inode->i_private;
1045         return 0;
1046 }
1047
1048 static ssize_t radeon_ttm_vram_read(struct file *f, char __user *buf,
1049                                     size_t size, loff_t *pos)
1050 {
1051         struct radeon_device *rdev = f->private_data;
1052         ssize_t result = 0;
1053         int r;
1054
1055         if (size & 0x3 || *pos & 0x3)
1056                 return -EINVAL;
1057
1058         while (size) {
1059                 unsigned long flags;
1060                 uint32_t value;
1061
1062                 if (*pos >= rdev->mc.mc_vram_size)
1063                         return result;
1064
1065                 spin_lock_irqsave(&rdev->mmio_idx_lock, flags);
1066                 WREG32(RADEON_MM_INDEX, ((uint32_t)*pos) | 0x80000000);
1067                 if (rdev->family >= CHIP_CEDAR)
1068                         WREG32(EVERGREEN_MM_INDEX_HI, *pos >> 31);
1069                 value = RREG32(RADEON_MM_DATA);
1070                 spin_unlock_irqrestore(&rdev->mmio_idx_lock, flags);
1071
1072                 r = put_user(value, (uint32_t *)buf);
1073                 if (r)
1074                         return r;
1075
1076                 result += 4;
1077                 buf += 4;
1078                 *pos += 4;
1079                 size -= 4;
1080         }
1081
1082         return result;
1083 }
1084
1085 static const struct file_operations radeon_ttm_vram_fops = {
1086         .owner = THIS_MODULE,
1087         .open = radeon_ttm_vram_open,
1088         .read = radeon_ttm_vram_read,
1089         .llseek = default_llseek
1090 };
1091
1092 static int radeon_ttm_gtt_open(struct inode *inode, struct file *filep)
1093 {
1094         struct radeon_device *rdev = inode->i_private;
1095         i_size_write(inode, rdev->mc.gtt_size);
1096         filep->private_data = inode->i_private;
1097         return 0;
1098 }
1099
1100 static ssize_t radeon_ttm_gtt_read(struct file *f, char __user *buf,
1101                                    size_t size, loff_t *pos)
1102 {
1103         struct radeon_device *rdev = f->private_data;
1104         ssize_t result = 0;
1105         int r;
1106
1107         while (size) {
1108                 loff_t p = *pos / PAGE_SIZE;
1109                 unsigned off = *pos & ~PAGE_MASK;
1110                 size_t cur_size = min_t(size_t, size, PAGE_SIZE - off);
1111                 struct page *page;
1112                 void *ptr;
1113
1114                 if (p >= rdev->gart.num_cpu_pages)
1115                         return result;
1116
1117                 page = rdev->gart.pages[p];
1118                 if (page) {
1119                         ptr = kmap(page);
1120                         ptr += off;
1121
1122                         r = copy_to_user(buf, ptr, cur_size);
1123                         kunmap(rdev->gart.pages[p]);
1124                 } else
1125                         r = clear_user(buf, cur_size);
1126
1127                 if (r)
1128                         return -EFAULT;
1129
1130                 result += cur_size;
1131                 buf += cur_size;
1132                 *pos += cur_size;
1133                 size -= cur_size;
1134         }
1135
1136         return result;
1137 }
1138
1139 static const struct file_operations radeon_ttm_gtt_fops = {
1140         .owner = THIS_MODULE,
1141         .open = radeon_ttm_gtt_open,
1142         .read = radeon_ttm_gtt_read,
1143         .llseek = default_llseek
1144 };
1145
1146 #endif
1147
1148 static int radeon_ttm_debugfs_init(struct radeon_device *rdev)
1149 {
1150 #if defined(CONFIG_DEBUG_FS)
1151         unsigned count;
1152
1153         struct drm_minor *minor = rdev->ddev->primary;
1154         struct dentry *ent, *root = minor->debugfs_root;
1155
1156         ent = debugfs_create_file("radeon_vram", S_IFREG | S_IRUGO, root,
1157                                   rdev, &radeon_ttm_vram_fops);
1158         if (IS_ERR(ent))
1159                 return PTR_ERR(ent);
1160         rdev->mman.vram = ent;
1161
1162         ent = debugfs_create_file("radeon_gtt", S_IFREG | S_IRUGO, root,
1163                                   rdev, &radeon_ttm_gtt_fops);
1164         if (IS_ERR(ent))
1165                 return PTR_ERR(ent);
1166         rdev->mman.gtt = ent;
1167
1168         count = ARRAY_SIZE(radeon_ttm_debugfs_list);
1169
1170 #ifdef CONFIG_SWIOTLB
1171         if (!swiotlb_nr_tbl())
1172                 --count;
1173 #endif
1174
1175         return radeon_debugfs_add_files(rdev, radeon_ttm_debugfs_list, count);
1176 #else
1177
1178         return 0;
1179 #endif
1180 }
1181
1182 static void radeon_ttm_debugfs_fini(struct radeon_device *rdev)
1183 {
1184 #if defined(CONFIG_DEBUG_FS)
1185
1186         debugfs_remove(rdev->mman.vram);
1187         rdev->mman.vram = NULL;
1188
1189         debugfs_remove(rdev->mman.gtt);
1190         rdev->mman.gtt = NULL;
1191 #endif
1192 }