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