drivers/gpu/drm/radeon/radeon_ttm.c

   1 /*
   2  * Copyright 2009 Jerome Glisse.
   3  * All Rights Reserved.
   4  *
   5  * Permission is hereby granted, free of charge, to any person obtaining a
   6  * copy of this software and associated documentation files (the
   7  * "Software"), to deal in the Software without restriction, including
   8  * without limitation the rights to use, copy, modify, merge, publish,
   9  * distribute, sub license, and/or sell copies of the Software, and to
  10  * permit persons to whom the Software is furnished to do so, subject to
  11  * the following conditions:
  12  *
  13  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  14  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  15  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  16  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
  17  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  18  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  19  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  20  *
  21  * The above copyright notice and this permission notice (including the
  22  * next paragraph) shall be included in all copies or substantial portions
  23  * of the Software.
  24  *
  25  */
  26 /*
  27  * Authors:
  28  *    Jerome Glisse <glisse@freedesktop.org>
  29  *    Thomas Hellstrom <thomas-at-tungstengraphics-dot-com>
  30  *    Dave Airlie
  31  */
  32
  33 #include <linux/dma-mapping.h>
  34 #include <linux/pagemap.h>
  35 #include <linux/pci.h>
  36 #include <linux/seq_file.h>
  37 #include <linux/slab.h>
  38 #include <linux/swap.h>
  39 #include <linux/swiotlb.h>
  40
  41 #include <drm/drm_agpsupport.h>
  42 #include <drm/drm_device.h>
  43 #include <drm/drm_file.h>
  44 #include <drm/drm_prime.h>
  45 #include <drm/radeon_drm.h>
  46 #include <drm/ttm/ttm_bo_api.h>
  47 #include <drm/ttm/ttm_bo_driver.h>
  48 #include <drm/ttm/ttm_placement.h>
  49
  50 #include "radeon_reg.h"
  51 #include "radeon.h"
  52 #include "radeon_ttm.h"
  53
  54 static void radeon_ttm_debugfs_init(struct radeon_device *rdev);
  55
  56 static int radeon_ttm_tt_bind(struct ttm_bo_device *bdev,
  57                               struct ttm_tt *ttm,
  58                               struct ttm_resource *bo_mem);
  59 static void radeon_ttm_tt_unbind(struct ttm_bo_device *bdev,
  60                                  struct ttm_tt *ttm);
  61
  62 struct radeon_device *radeon_get_rdev(struct ttm_bo_device *bdev)
  63 {
  64         struct radeon_mman *mman;
  65         struct radeon_device *rdev;
  66
  67         mman = container_of(bdev, struct radeon_mman, bdev);
  68         rdev = container_of(mman, struct radeon_device, mman);
  69         return rdev;
  70 }
  71
  72 static int radeon_ttm_init_vram(struct radeon_device *rdev)
  73 {
  74         return ttm_range_man_init(&rdev->mman.bdev, TTM_PL_VRAM,
  75                                   false, rdev->mc.real_vram_size >> PAGE_SHIFT);
  76 }
  77
  78 static int radeon_ttm_init_gtt(struct radeon_device *rdev)
  79 {
  80         return ttm_range_man_init(&rdev->mman.bdev, TTM_PL_TT,
  81                                   true, rdev->mc.gtt_size >> PAGE_SHIFT);
  82 }
  83
  84 static void radeon_evict_flags(struct ttm_buffer_object *bo,
  85                                 struct ttm_placement *placement)
  86 {
  87         static const struct ttm_place placements = {
  88                 .fpfn = 0,
  89                 .lpfn = 0,
  90                 .mem_type = TTM_PL_SYSTEM,
  91                 .flags = 0
  92         };
  93
  94         struct radeon_bo *rbo;
  95
  96         if (!radeon_ttm_bo_is_radeon_bo(bo)) {
  97                 placement->placement = &placements;
  98                 placement->busy_placement = &placements;
  99                 placement->num_placement = 1;
 100                 placement->num_busy_placement = 1;
 101                 return;
 102         }
 103         rbo = container_of(bo, struct radeon_bo, tbo);
 104         switch (bo->mem.mem_type) {
 105         case TTM_PL_VRAM:
 106                 if (rbo->rdev->ring[radeon_copy_ring_index(rbo->rdev)].ready == false)
 107                         radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_CPU);
 108                 else if (rbo->rdev->mc.visible_vram_size < rbo->rdev->mc.real_vram_size &&
 109                          bo->mem.start < (rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT)) {
 110                         unsigned fpfn = rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
 111                         int i;
 112
 113                         /* Try evicting to the CPU inaccessible part of VRAM
 114                          * first, but only set GTT as busy placement, so this
 115                          * BO will be evicted to GTT rather than causing other
 116                          * BOs to be evicted from VRAM
 117                          */
 118                         radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_VRAM |
 119                                                          RADEON_GEM_DOMAIN_GTT);
 120                         rbo->placement.num_busy_placement = 0;
 121                         for (i = 0; i < rbo->placement.num_placement; i++) {
 122                                 if (rbo->placements[i].mem_type == TTM_PL_VRAM) {
 123                                         if (rbo->placements[i].fpfn < fpfn)
 124                                                 rbo->placements[i].fpfn = fpfn;
 125                                 } else {
 126                                         rbo->placement.busy_placement =
 127                                                 &rbo->placements[i];
 128                                         rbo->placement.num_busy_placement = 1;
 129                                 }
 130                         }
 131                 } else
 132                         radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_GTT);
 133                 break;
 134         case TTM_PL_TT:
 135         default:
 136                 radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_CPU);
 137         }
 138         *placement = rbo->placement;
 139 }
 140
 141 static int radeon_verify_access(struct ttm_buffer_object *bo, struct file *filp)
 142 {
 143         struct radeon_bo *rbo = container_of(bo, struct radeon_bo, tbo);
 144         struct radeon_device *rdev = radeon_get_rdev(bo->bdev);
 145
 146         if (radeon_ttm_tt_has_userptr(rdev, bo->ttm))
 147                 return -EPERM;
 148         return drm_vma_node_verify_access(&rbo->tbo.base.vma_node,
 149                                           filp->private_data);
 150 }
 151
 152 static int radeon_move_blit(struct ttm_buffer_object *bo,
 153                         bool evict,
 154                         struct ttm_resource *new_mem,
 155                         struct ttm_resource *old_mem)
 156 {
 157         struct radeon_device *rdev;
 158         uint64_t old_start, new_start;
 159         struct radeon_fence *fence;
 160         unsigned num_pages;
 161         int r, ridx;
 162
 163         rdev = radeon_get_rdev(bo->bdev);
 164         ridx = radeon_copy_ring_index(rdev);
 165         old_start = (u64)old_mem->start << PAGE_SHIFT;
 166         new_start = (u64)new_mem->start << PAGE_SHIFT;
 167
 168         switch (old_mem->mem_type) {
 169         case TTM_PL_VRAM:
 170                 old_start += rdev->mc.vram_start;
 171                 break;
 172         case TTM_PL_TT:
 173                 old_start += rdev->mc.gtt_start;
 174                 break;
 175         default:
 176                 DRM_ERROR("Unknown placement %d\n", old_mem->mem_type);
 177                 return -EINVAL;
 178         }
 179         switch (new_mem->mem_type) {
 180         case TTM_PL_VRAM:
 181                 new_start += rdev->mc.vram_start;
 182                 break;
 183         case TTM_PL_TT:
 184                 new_start += rdev->mc.gtt_start;
 185                 break;
 186         default:
 187                 DRM_ERROR("Unknown placement %d\n", old_mem->mem_type);
 188                 return -EINVAL;
 189         }
 190         if (!rdev->ring[ridx].ready) {
 191                 DRM_ERROR("Trying to move memory with ring turned off.\n");
 192                 return -EINVAL;
 193         }
 194
 195         BUILD_BUG_ON((PAGE_SIZE % RADEON_GPU_PAGE_SIZE) != 0);
 196
 197         num_pages = new_mem->num_pages * (PAGE_SIZE / RADEON_GPU_PAGE_SIZE);
 198         fence = radeon_copy(rdev, old_start, new_start, num_pages, bo->base.resv);
 199         if (IS_ERR(fence))
 200                 return PTR_ERR(fence);
 201
 202         r = ttm_bo_move_accel_cleanup(bo, &fence->base, evict, false, new_mem);
 203         radeon_fence_unref(&fence);
 204         return r;
 205 }
 206
 207 static int radeon_bo_move(struct ttm_buffer_object *bo, bool evict,
 208                           struct ttm_operation_ctx *ctx,
 209                           struct ttm_resource *new_mem,
 210                           struct ttm_place *hop)
 211 {
 212         struct radeon_device *rdev;
 213         struct radeon_bo *rbo;
 214         struct ttm_resource *old_mem = &bo->mem;
 215         int r;
 216
 217         if (new_mem->mem_type == TTM_PL_TT) {
 218                 r = radeon_ttm_tt_bind(bo->bdev, bo->ttm, new_mem);
 219                 if (r)
 220                         return r;
 221         }
 222
 223         r = ttm_bo_wait_ctx(bo, ctx);
 224         if (r)
 225                 return r;
 226
 227         /* Can't move a pinned BO */
 228         rbo = container_of(bo, struct radeon_bo, tbo);
 229         if (WARN_ON_ONCE(rbo->tbo.pin_count > 0))
 230                 return -EINVAL;
 231
 232         rdev = radeon_get_rdev(bo->bdev);
 233         if (old_mem->mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
 234                 ttm_bo_move_null(bo, new_mem);
 235                 goto out;
 236         }
 237         if (old_mem->mem_type == TTM_PL_SYSTEM &&
 238             new_mem->mem_type == TTM_PL_TT) {
 239                 ttm_bo_move_null(bo, new_mem);
 240                 goto out;
 241         }
 242
 243         if (old_mem->mem_type == TTM_PL_TT &&
 244             new_mem->mem_type == TTM_PL_SYSTEM) {
 245                 radeon_ttm_tt_unbind(bo->bdev, bo->ttm);
 246                 ttm_resource_free(bo, &bo->mem);
 247                 ttm_bo_assign_mem(bo, new_mem);
 248                 goto out;
 249         }
 250         if (rdev->ring[radeon_copy_ring_index(rdev)].ready &&
 251             rdev->asic->copy.copy != NULL) {
 252                 if ((old_mem->mem_type == TTM_PL_SYSTEM &&
 253                      new_mem->mem_type == TTM_PL_VRAM) ||
 254                     (old_mem->mem_type == TTM_PL_VRAM &&
 255                      new_mem->mem_type == TTM_PL_SYSTEM)) {
 256                         hop->fpfn = 0;
 257                         hop->lpfn = 0;
 258                         hop->mem_type = TTM_PL_TT;
 259                         hop->flags = 0;
 260                         return -EMULTIHOP;
 261                 }
 262
 263                 r = radeon_move_blit(bo, evict, new_mem, old_mem);
 264         } else {
 265                 r = -ENODEV;
 266         }
 267
 268         if (r) {
 269                 r = ttm_bo_move_memcpy(bo, ctx, new_mem);
 270                 if (r)
 271                         return r;
 272         }
 273
 274 out:
 275         /* update statistics */
 276         atomic64_add(bo->base.size, &rdev->num_bytes_moved);
 277         radeon_bo_move_notify(bo, evict, new_mem);
 278         return 0;
 279 }
 280
 281 static int radeon_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_resource *mem)
 282 {
 283         struct radeon_device *rdev = radeon_get_rdev(bdev);
 284         size_t bus_size = (size_t)mem->num_pages << PAGE_SHIFT;
 285
 286         switch (mem->mem_type) {
 287         case TTM_PL_SYSTEM:
 288                 /* system memory */
 289                 return 0;
 290         case TTM_PL_TT:
 291 #if IS_ENABLED(CONFIG_AGP)
 292                 if (rdev->flags & RADEON_IS_AGP) {
 293                         /* RADEON_IS_AGP is set only if AGP is active */
 294                         mem->bus.offset = (mem->start << PAGE_SHIFT) +
 295                                 rdev->mc.agp_base;
 296                         mem->bus.is_iomem = !rdev->ddev->agp->cant_use_aperture;
 297                         mem->bus.caching = ttm_write_combined;
 298                 }
 299 #endif
 300                 break;
 301         case TTM_PL_VRAM:
 302                 mem->bus.offset = mem->start << PAGE_SHIFT;
 303                 /* check if it's visible */
 304                 if ((mem->bus.offset + bus_size) > rdev->mc.visible_vram_size)
 305                         return -EINVAL;
 306                 mem->bus.offset += rdev->mc.aper_base;
 307                 mem->bus.is_iomem = true;
 308                 mem->bus.caching = ttm_write_combined;
 309 #ifdef __alpha__
 310                 /*
 311                  * Alpha: use bus.addr to hold the ioremap() return,
 312                  * so we can modify bus.base below.
 313                  */
 314                 mem->bus.addr = ioremap_wc(mem->bus.offset, bus_size);
 315                 if (!mem->bus.addr)
 316                         return -ENOMEM;
 317
 318                 /*
 319                  * Alpha: Use just the bus offset plus
 320                  * the hose/domain memory base for bus.base.
 321                  * It then can be used to build PTEs for VRAM
 322                  * access, as done in ttm_bo_vm_fault().
 323                  */
 324                 mem->bus.offset = (mem->bus.offset & 0x0ffffffffUL) +
 325                         rdev->hose->dense_mem_base;
 326 #endif
 327                 break;
 328         default:
 329                 return -EINVAL;
 330         }
 331         return 0;
 332 }
 333
 334 /*
 335  * TTM backend functions.
 336  */
 337 struct radeon_ttm_tt {
 338         struct ttm_tt           ttm;
 339         u64                             offset;
 340
 341         uint64_t                        userptr;
 342         struct mm_struct                *usermm;
 343         uint32_t                        userflags;
 344         bool bound;
 345 };
 346
 347 /* prepare the sg table with the user pages */
 348 static int radeon_ttm_tt_pin_userptr(struct ttm_bo_device *bdev, struct ttm_tt *ttm)
 349 {
 350         struct radeon_device *rdev = radeon_get_rdev(bdev);
 351         struct radeon_ttm_tt *gtt = (void *)ttm;
 352         unsigned pinned = 0;
 353         int r;
 354
 355         int write = !(gtt->userflags & RADEON_GEM_USERPTR_READONLY);
 356         enum dma_data_direction direction = write ?
 357                 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
 358
 359         if (current->mm != gtt->usermm)
 360                 return -EPERM;
 361
 362         if (gtt->userflags & RADEON_GEM_USERPTR_ANONONLY) {
 363                 /* check that we only pin down anonymous memory
 364                    to prevent problems with writeback */
 365                 unsigned long end = gtt->userptr + ttm->num_pages * PAGE_SIZE;
 366                 struct vm_area_struct *vma;
 367                 vma = find_vma(gtt->usermm, gtt->userptr);
 368                 if (!vma || vma->vm_file || vma->vm_end < end)
 369                         return -EPERM;
 370         }
 371
 372         do {
 373                 unsigned num_pages = ttm->num_pages - pinned;
 374                 uint64_t userptr = gtt->userptr + pinned * PAGE_SIZE;
 375                 struct page **pages = ttm->pages + pinned;
 376
 377                 r = get_user_pages(userptr, num_pages, write ? FOLL_WRITE : 0,
 378                                    pages, NULL);
 379                 if (r < 0)
 380                         goto release_pages;
 381
 382                 pinned += r;
 383
 384         } while (pinned < ttm->num_pages);
 385
 386         r = sg_alloc_table_from_pages(ttm->sg, ttm->pages, ttm->num_pages, 0,
 387                                       ttm->num_pages << PAGE_SHIFT,
 388                                       GFP_KERNEL);
 389         if (r)
 390                 goto release_sg;
 391
 392         r = dma_map_sgtable(rdev->dev, ttm->sg, direction, 0);
 393         if (r)
 394                 goto release_sg;
 395
 396         drm_prime_sg_to_dma_addr_array(ttm->sg, gtt->ttm.dma_address,
 397                                        ttm->num_pages);
 398
 399         return 0;
 400
 401 release_sg:
 402         kfree(ttm->sg);
 403
 404 release_pages:
 405         release_pages(ttm->pages, pinned);
 406         return r;
 407 }
 408
 409 static void radeon_ttm_tt_unpin_userptr(struct ttm_bo_device *bdev, struct ttm_tt *ttm)
 410 {
 411         struct radeon_device *rdev = radeon_get_rdev(bdev);
 412         struct radeon_ttm_tt *gtt = (void *)ttm;
 413         struct sg_page_iter sg_iter;
 414
 415         int write = !(gtt->userflags & RADEON_GEM_USERPTR_READONLY);
 416         enum dma_data_direction direction = write ?
 417                 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
 418
 419         /* double check that we don't free the table twice */
 420         if (!ttm->sg->sgl)
 421                 return;
 422
 423         /* free the sg table and pages again */
 424         dma_unmap_sgtable(rdev->dev, ttm->sg, direction, 0);
 425
 426         for_each_sgtable_page(ttm->sg, &sg_iter, 0) {
 427                 struct page *page = sg_page_iter_page(&sg_iter);
 428                 if (!(gtt->userflags & RADEON_GEM_USERPTR_READONLY))
 429                         set_page_dirty(page);
 430
 431                 mark_page_accessed(page);
 432                 put_page(page);
 433         }
 434
 435         sg_free_table(ttm->sg);
 436 }
 437
 438 static bool radeon_ttm_backend_is_bound(struct ttm_tt *ttm)
 439 {
 440         struct radeon_ttm_tt *gtt = (void*)ttm;
 441
 442         return (gtt->bound);
 443 }
 444
 445 static int radeon_ttm_backend_bind(struct ttm_bo_device *bdev,
 446                                    struct ttm_tt *ttm,
 447                                    struct ttm_resource *bo_mem)
 448 {
 449         struct radeon_ttm_tt *gtt = (void*)ttm;
 450         struct radeon_device *rdev = radeon_get_rdev(bdev);
 451         uint32_t flags = RADEON_GART_PAGE_VALID | RADEON_GART_PAGE_READ |
 452                 RADEON_GART_PAGE_WRITE;
 453         int r;
 454
 455         if (gtt->bound)
 456                 return 0;
 457
 458         if (gtt->userptr) {
 459                 radeon_ttm_tt_pin_userptr(bdev, ttm);
 460                 flags &= ~RADEON_GART_PAGE_WRITE;
 461         }
 462
 463         gtt->offset = (unsigned long)(bo_mem->start << PAGE_SHIFT);
 464         if (!ttm->num_pages) {
 465                 WARN(1, "nothing to bind %u pages for mreg %p back %p!\n",
 466                      ttm->num_pages, bo_mem, ttm);
 467         }
 468         if (ttm->caching == ttm_cached)
 469                 flags |= RADEON_GART_PAGE_SNOOP;
 470         r = radeon_gart_bind(rdev, gtt->offset, ttm->num_pages,
 471                              ttm->pages, gtt->ttm.dma_address, flags);
 472         if (r) {
 473                 DRM_ERROR("failed to bind %u pages at 0x%08X\n",
 474                           ttm->num_pages, (unsigned)gtt->offset);
 475                 return r;
 476         }
 477         gtt->bound = true;
 478         return 0;
 479 }
 480
 481 static void radeon_ttm_backend_unbind(struct ttm_bo_device *bdev, struct ttm_tt *ttm)
 482 {
 483         struct radeon_ttm_tt *gtt = (void *)ttm;
 484         struct radeon_device *rdev = radeon_get_rdev(bdev);
 485
 486         if (!gtt->bound)
 487                 return;
 488
 489         radeon_gart_unbind(rdev, gtt->offset, ttm->num_pages);
 490
 491         if (gtt->userptr)
 492                 radeon_ttm_tt_unpin_userptr(bdev, ttm);
 493         gtt->bound = false;
 494 }
 495
 496 static void radeon_ttm_backend_destroy(struct ttm_bo_device *bdev, struct ttm_tt *ttm)
 497 {
 498         struct radeon_ttm_tt *gtt = (void *)ttm;
 499
 500         radeon_ttm_backend_unbind(bdev, ttm);
 501         ttm_tt_destroy_common(bdev, ttm);
 502
 503         ttm_tt_fini(&gtt->ttm);
 504         kfree(gtt);
 505 }
 506
 507 static struct ttm_tt *radeon_ttm_tt_create(struct ttm_buffer_object *bo,
 508                                            uint32_t page_flags)
 509 {
 510         struct radeon_ttm_tt *gtt;
 511         enum ttm_caching caching;
 512         struct radeon_bo *rbo;
 513 #if IS_ENABLED(CONFIG_AGP)
 514         struct radeon_device *rdev = radeon_get_rdev(bo->bdev);
 515
 516         if (rdev->flags & RADEON_IS_AGP) {
 517                 return ttm_agp_tt_create(bo, rdev->ddev->agp->bridge,
 518                                          page_flags);
 519         }
 520 #endif
 521         rbo = container_of(bo, struct radeon_bo, tbo);
 522
 523         gtt = kzalloc(sizeof(struct radeon_ttm_tt), GFP_KERNEL);
 524         if (gtt == NULL) {
 525                 return NULL;
 526         }
 527
 528         if (rbo->flags & RADEON_GEM_GTT_UC)
 529                 caching = ttm_uncached;
 530         else if (rbo->flags & RADEON_GEM_GTT_WC)
 531                 caching = ttm_write_combined;
 532         else
 533                 caching = ttm_cached;
 534
 535         if (ttm_sg_tt_init(&gtt->ttm, bo, page_flags, caching)) {
 536                 kfree(gtt);
 537                 return NULL;
 538         }
 539         return &gtt->ttm;
 540 }
 541
 542 static struct radeon_ttm_tt *radeon_ttm_tt_to_gtt(struct radeon_device *rdev,
 543                                                   struct ttm_tt *ttm)
 544 {
 545 #if IS_ENABLED(CONFIG_AGP)
 546         if (rdev->flags & RADEON_IS_AGP)
 547                 return NULL;
 548 #endif
 549
 550         if (!ttm)
 551                 return NULL;
 552         return container_of(ttm, struct radeon_ttm_tt, ttm);
 553 }
 554
 555 static int radeon_ttm_tt_populate(struct ttm_bo_device *bdev,
 556                                   struct ttm_tt *ttm,
 557                                   struct ttm_operation_ctx *ctx)
 558 {
 559         struct radeon_device *rdev = radeon_get_rdev(bdev);
 560         struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(rdev, ttm);
 561         bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
 562
 563         if (gtt && gtt->userptr) {
 564                 ttm->sg = kzalloc(sizeof(struct sg_table), GFP_KERNEL);
 565                 if (!ttm->sg)
 566                         return -ENOMEM;
 567
 568                 ttm->page_flags |= TTM_PAGE_FLAG_SG;
 569                 return 0;
 570         }
 571
 572         if (slave && ttm->sg) {
 573                 drm_prime_sg_to_dma_addr_array(ttm->sg, gtt->ttm.dma_address,
 574                                                ttm->num_pages);
 575                 return 0;
 576         }
 577
 578         return ttm_pool_alloc(&rdev->mman.bdev.pool, ttm, ctx);
 579 }
 580
 581 static void radeon_ttm_tt_unpopulate(struct ttm_bo_device *bdev, struct ttm_tt *ttm)
 582 {
 583         struct radeon_device *rdev = radeon_get_rdev(bdev);
 584         struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(rdev, ttm);
 585         bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
 586
 587         if (gtt && gtt->userptr) {
 588                 kfree(ttm->sg);
 589                 ttm->page_flags &= ~TTM_PAGE_FLAG_SG;
 590                 return;
 591         }
 592
 593         if (slave)
 594                 return;
 595
 596         return ttm_pool_free(&rdev->mman.bdev.pool, ttm);
 597 }
 598
 599 int radeon_ttm_tt_set_userptr(struct radeon_device *rdev,
 600                               struct ttm_tt *ttm, uint64_t addr,
 601                               uint32_t flags)
 602 {
 603         struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(rdev, ttm);
 604
 605         if (gtt == NULL)
 606                 return -EINVAL;
 607
 608         gtt->userptr = addr;
 609         gtt->usermm = current->mm;
 610         gtt->userflags = flags;
 611         return 0;
 612 }
 613
 614 bool radeon_ttm_tt_is_bound(struct ttm_bo_device *bdev,
 615                             struct ttm_tt *ttm)
 616 {
 617 #if IS_ENABLED(CONFIG_AGP)
 618         struct radeon_device *rdev = radeon_get_rdev(bdev);
 619         if (rdev->flags & RADEON_IS_AGP)
 620                 return ttm_agp_is_bound(ttm);
 621 #endif
 622         return radeon_ttm_backend_is_bound(ttm);
 623 }
 624
 625 static int radeon_ttm_tt_bind(struct ttm_bo_device *bdev,
 626                               struct ttm_tt *ttm,
 627                               struct ttm_resource *bo_mem)
 628 {
 629 #if IS_ENABLED(CONFIG_AGP)
 630         struct radeon_device *rdev = radeon_get_rdev(bdev);
 631 #endif
 632
 633         if (!bo_mem)
 634                 return -EINVAL;
 635 #if IS_ENABLED(CONFIG_AGP)
 636         if (rdev->flags & RADEON_IS_AGP)
 637                 return ttm_agp_bind(ttm, bo_mem);
 638 #endif
 639
 640         return radeon_ttm_backend_bind(bdev, ttm, bo_mem);
 641 }
 642
 643 static void radeon_ttm_tt_unbind(struct ttm_bo_device *bdev,
 644                                  struct ttm_tt *ttm)
 645 {
 646 #if IS_ENABLED(CONFIG_AGP)
 647         struct radeon_device *rdev = radeon_get_rdev(bdev);
 648
 649         if (rdev->flags & RADEON_IS_AGP) {
 650                 ttm_agp_unbind(ttm);
 651                 return;
 652         }
 653 #endif
 654         radeon_ttm_backend_unbind(bdev, ttm);
 655 }
 656
 657 static void radeon_ttm_tt_destroy(struct ttm_bo_device *bdev,
 658                                   struct ttm_tt *ttm)
 659 {
 660 #if IS_ENABLED(CONFIG_AGP)
 661         struct radeon_device *rdev = radeon_get_rdev(bdev);
 662
 663         if (rdev->flags & RADEON_IS_AGP) {
 664                 ttm_agp_unbind(ttm);
 665                 ttm_tt_destroy_common(bdev, ttm);
 666                 ttm_agp_destroy(ttm);
 667                 return;
 668         }
 669 #endif
 670         radeon_ttm_backend_destroy(bdev, ttm);
 671 }
 672
 673 bool radeon_ttm_tt_has_userptr(struct radeon_device *rdev,
 674                                struct ttm_tt *ttm)
 675 {
 676         struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(rdev, ttm);
 677
 678         if (gtt == NULL)
 679                 return false;
 680
 681         return !!gtt->userptr;
 682 }
 683
 684 bool radeon_ttm_tt_is_readonly(struct radeon_device *rdev,
 685                                struct ttm_tt *ttm)
 686 {
 687         struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(rdev, ttm);
 688
 689         if (gtt == NULL)
 690                 return false;
 691
 692         return !!(gtt->userflags & RADEON_GEM_USERPTR_READONLY);
 693 }
 694
 695 static void
 696 radeon_bo_delete_mem_notify(struct ttm_buffer_object *bo)
 697 {
 698         radeon_bo_move_notify(bo, false, NULL);
 699 }
 700
 701 static struct ttm_bo_driver radeon_bo_driver = {
 702         .ttm_tt_create = &radeon_ttm_tt_create,
 703         .ttm_tt_populate = &radeon_ttm_tt_populate,
 704         .ttm_tt_unpopulate = &radeon_ttm_tt_unpopulate,
 705         .ttm_tt_destroy = &radeon_ttm_tt_destroy,
 706         .eviction_valuable = ttm_bo_eviction_valuable,
 707         .evict_flags = &radeon_evict_flags,
 708         .move = &radeon_bo_move,
 709         .verify_access = &radeon_verify_access,
 710         .delete_mem_notify = &radeon_bo_delete_mem_notify,
 711         .io_mem_reserve = &radeon_ttm_io_mem_reserve,
 712 };
 713
 714 int radeon_ttm_init(struct radeon_device *rdev)
 715 {
 716         int r;
 717
 718         /* No others user of address space so set it to 0 */
 719         r = ttm_bo_device_init(&rdev->mman.bdev, &radeon_bo_driver, rdev->dev,
 720                                rdev->ddev->anon_inode->i_mapping,
 721                                rdev->ddev->vma_offset_manager,
 722                                rdev->need_swiotlb,
 723                                dma_addressing_limited(&rdev->pdev->dev));
 724         if (r) {
 725                 DRM_ERROR("failed initializing buffer object driver(%d).\n", r);
 726                 return r;
 727         }
 728         rdev->mman.initialized = true;
 729
 730         r = radeon_ttm_init_vram(rdev);
 731         if (r) {
 732                 DRM_ERROR("Failed initializing VRAM heap.\n");
 733                 return r;
 734         }
 735         /* Change the size here instead of the init above so only lpfn is affected */
 736         radeon_ttm_set_active_vram_size(rdev, rdev->mc.visible_vram_size);
 737
 738         r = radeon_bo_create(rdev, 256 * 1024, PAGE_SIZE, true,
 739                              RADEON_GEM_DOMAIN_VRAM, 0, NULL,
 740                              NULL, &rdev->stolen_vga_memory);
 741         if (r) {
 742                 return r;
 743         }
 744         r = radeon_bo_reserve(rdev->stolen_vga_memory, false);
 745         if (r)
 746                 return r;
 747         r = radeon_bo_pin(rdev->stolen_vga_memory, RADEON_GEM_DOMAIN_VRAM, NULL);
 748         radeon_bo_unreserve(rdev->stolen_vga_memory);
 749         if (r) {
 750                 radeon_bo_unref(&rdev->stolen_vga_memory);
 751                 return r;
 752         }
 753         DRM_INFO("radeon: %uM of VRAM memory ready\n",
 754                  (unsigned) (rdev->mc.real_vram_size / (1024 * 1024)));
 755
 756         r = radeon_ttm_init_gtt(rdev);
 757         if (r) {
 758                 DRM_ERROR("Failed initializing GTT heap.\n");
 759                 return r;
 760         }
 761         DRM_INFO("radeon: %uM of GTT memory ready.\n",
 762                  (unsigned)(rdev->mc.gtt_size / (1024 * 1024)));
 763
 764         radeon_ttm_debugfs_init(rdev);
 765
 766         return 0;
 767 }
 768
 769 void radeon_ttm_fini(struct radeon_device *rdev)
 770 {
 771         int r;
 772
 773         if (!rdev->mman.initialized)
 774                 return;
 775
 776         if (rdev->stolen_vga_memory) {
 777                 r = radeon_bo_reserve(rdev->stolen_vga_memory, false);
 778                 if (r == 0) {
 779                         radeon_bo_unpin(rdev->stolen_vga_memory);
 780                         radeon_bo_unreserve(rdev->stolen_vga_memory);
 781                 }
 782                 radeon_bo_unref(&rdev->stolen_vga_memory);
 783         }
 784         ttm_range_man_fini(&rdev->mman.bdev, TTM_PL_VRAM);
 785         ttm_range_man_fini(&rdev->mman.bdev, TTM_PL_TT);
 786         ttm_bo_device_release(&rdev->mman.bdev);
 787         radeon_gart_fini(rdev);
 788         rdev->mman.initialized = false;
 789         DRM_INFO("radeon: ttm finalized\n");
 790 }
 791
 792 /* this should only be called at bootup or when userspace
 793  * isn't running */
 794 void radeon_ttm_set_active_vram_size(struct radeon_device *rdev, u64 size)
 795 {
 796         struct ttm_resource_manager *man;
 797
 798         if (!rdev->mman.initialized)
 799                 return;
 800
 801         man = ttm_manager_type(&rdev->mman.bdev, TTM_PL_VRAM);
 802         /* this just adjusts TTM size idea, which sets lpfn to the correct value */
 803         man->size = size >> PAGE_SHIFT;
 804 }
 805
 806 static vm_fault_t radeon_ttm_fault(struct vm_fault *vmf)
 807 {
 808         struct ttm_buffer_object *bo = vmf->vma->vm_private_data;
 809         struct radeon_device *rdev = radeon_get_rdev(bo->bdev);
 810         vm_fault_t ret;
 811
 812         down_read(&rdev->pm.mclk_lock);
 813
 814         ret = ttm_bo_vm_reserve(bo, vmf);
 815         if (ret)
 816                 goto unlock_mclk;
 817
 818         ret = radeon_bo_fault_reserve_notify(bo);
 819         if (ret)
 820                 goto unlock_resv;
 821
 822         ret = ttm_bo_vm_fault_reserved(vmf, vmf->vma->vm_page_prot,
 823                                        TTM_BO_VM_NUM_PREFAULT, 1);
 824         if (ret == VM_FAULT_RETRY && !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT))
 825                 goto unlock_mclk;
 826
 827 unlock_resv:
 828         dma_resv_unlock(bo->base.resv);
 829
 830 unlock_mclk:
 831         up_read(&rdev->pm.mclk_lock);
 832         return ret;
 833 }
 834
 835 static struct vm_operations_struct radeon_ttm_vm_ops = {
 836         .fault = radeon_ttm_fault,
 837         .open = ttm_bo_vm_open,
 838         .close = ttm_bo_vm_close,
 839         .access = ttm_bo_vm_access
 840 };
 841
 842 int radeon_mmap(struct file *filp, struct vm_area_struct *vma)
 843 {
 844         int r;
 845         struct drm_file *file_priv = filp->private_data;
 846         struct radeon_device *rdev = file_priv->minor->dev->dev_private;
 847
 848         if (rdev == NULL)
 849                 return -EINVAL;
 850
 851         r = ttm_bo_mmap(filp, vma, &rdev->mman.bdev);
 852         if (unlikely(r != 0))
 853                 return r;
 854
 855         vma->vm_ops = &radeon_ttm_vm_ops;
 856         return 0;
 857 }
 858
 859 #if defined(CONFIG_DEBUG_FS)
 860
 861 static int radeon_mm_vram_dump_table_show(struct seq_file *m, void *unused)
 862 {
 863         struct radeon_device *rdev = (struct radeon_device *)m->private;
 864         struct ttm_resource_manager *man = ttm_manager_type(&rdev->mman.bdev,
 865                                                             TTM_PL_VRAM);
 866         struct drm_printer p = drm_seq_file_printer(m);
 867
 868         man->func->debug(man, &p);
 869         return 0;
 870 }
 871
 872 static int radeon_ttm_page_pool_show(struct seq_file *m, void *data)
 873 {
 874         struct radeon_device *rdev = (struct radeon_device *)m->private;
 875
 876         return ttm_pool_debugfs(&rdev->mman.bdev.pool, m);
 877 }
 878
 879 static int radeon_mm_gtt_dump_table_show(struct seq_file *m, void *unused)
 880 {
 881         struct radeon_device *rdev = (struct radeon_device *)m->private;
 882         struct ttm_resource_manager *man = ttm_manager_type(&rdev->mman.bdev,
 883                                                             TTM_PL_TT);
 884         struct drm_printer p = drm_seq_file_printer(m);
 885
 886         man->func->debug(man, &p);
 887         return 0;
 888 }
 889
 890 DEFINE_SHOW_ATTRIBUTE(radeon_mm_vram_dump_table);
 891 DEFINE_SHOW_ATTRIBUTE(radeon_mm_gtt_dump_table);
 892 DEFINE_SHOW_ATTRIBUTE(radeon_ttm_page_pool);
 893
 894 static int radeon_ttm_vram_open(struct inode *inode, struct file *filep)
 895 {
 896         struct radeon_device *rdev = inode->i_private;
 897         i_size_write(inode, rdev->mc.mc_vram_size);
 898         filep->private_data = inode->i_private;
 899         return 0;
 900 }
 901
 902 static ssize_t radeon_ttm_vram_read(struct file *f, char __user *buf,
 903                                     size_t size, loff_t *pos)
 904 {
 905         struct radeon_device *rdev = f->private_data;
 906         ssize_t result = 0;
 907         int r;
 908
 909         if (size & 0x3 || *pos & 0x3)
 910                 return -EINVAL;
 911
 912         while (size) {
 913                 unsigned long flags;
 914                 uint32_t value;
 915
 916                 if (*pos >= rdev->mc.mc_vram_size)
 917                         return result;
 918
 919                 spin_lock_irqsave(&rdev->mmio_idx_lock, flags);
 920                 WREG32(RADEON_MM_INDEX, ((uint32_t)*pos) | 0x80000000);
 921                 if (rdev->family >= CHIP_CEDAR)
 922                         WREG32(EVERGREEN_MM_INDEX_HI, *pos >> 31);
 923                 value = RREG32(RADEON_MM_DATA);
 924                 spin_unlock_irqrestore(&rdev->mmio_idx_lock, flags);
 925
 926                 r = put_user(value, (uint32_t __user *)buf);
 927                 if (r)
 928                         return r;
 929
 930                 result += 4;
 931                 buf += 4;
 932                 *pos += 4;
 933                 size -= 4;
 934         }
 935
 936         return result;
 937 }
 938
 939 static const struct file_operations radeon_ttm_vram_fops = {
 940         .owner = THIS_MODULE,
 941         .open = radeon_ttm_vram_open,
 942         .read = radeon_ttm_vram_read,
 943         .llseek = default_llseek
 944 };
 945
 946 static int radeon_ttm_gtt_open(struct inode *inode, struct file *filep)
 947 {
 948         struct radeon_device *rdev = inode->i_private;
 949         i_size_write(inode, rdev->mc.gtt_size);
 950         filep->private_data = inode->i_private;
 951         return 0;
 952 }
 953
 954 static ssize_t radeon_ttm_gtt_read(struct file *f, char __user *buf,
 955                                    size_t size, loff_t *pos)
 956 {
 957         struct radeon_device *rdev = f->private_data;
 958         ssize_t result = 0;
 959         int r;
 960
 961         while (size) {
 962                 loff_t p = *pos / PAGE_SIZE;
 963                 unsigned off = *pos & ~PAGE_MASK;
 964                 size_t cur_size = min_t(size_t, size, PAGE_SIZE - off);
 965                 struct page *page;
 966                 void *ptr;
 967
 968                 if (p >= rdev->gart.num_cpu_pages)
 969                         return result;
 970
 971                 page = rdev->gart.pages[p];
 972                 if (page) {
 973                         ptr = kmap(page);
 974                         ptr += off;
 975
 976                         r = copy_to_user(buf, ptr, cur_size);
 977                         kunmap(rdev->gart.pages[p]);
 978                 } else
 979                         r = clear_user(buf, cur_size);
 980
 981                 if (r)
 982                         return -EFAULT;
 983
 984                 result += cur_size;
 985                 buf += cur_size;
 986                 *pos += cur_size;
 987                 size -= cur_size;
 988         }
 989
 990         return result;
 991 }
 992
 993 static const struct file_operations radeon_ttm_gtt_fops = {
 994         .owner = THIS_MODULE,
 995         .open = radeon_ttm_gtt_open,
 996         .read = radeon_ttm_gtt_read,
 997         .llseek = default_llseek
 998 };
 999
1000 #endif
1001
1002 static void radeon_ttm_debugfs_init(struct radeon_device *rdev)
1003 {
1004 #if defined(CONFIG_DEBUG_FS)
1005         struct drm_minor *minor = rdev->ddev->primary;
1006         struct dentry *root = minor->debugfs_root;
1007
1008         debugfs_create_file("radeon_vram", 0444, root, rdev,
1009                             &radeon_ttm_vram_fops);
1010
1011         debugfs_create_file("radeon_gtt", 0444, root, rdev,
1012                             &radeon_ttm_gtt_fops);
1013
1014         debugfs_create_file("radeon_vram_mm", 0444, root, rdev,
1015                             &radeon_mm_vram_dump_table_fops);
1016         debugfs_create_file("radeon_gtt_mm", 0444, root, rdev,
1017                             &radeon_mm_gtt_dump_table_fops);
1018         debugfs_create_file("ttm_page_pool", 0444, root, rdev,
1019                             &radeon_ttm_page_pool_fops);
1020 #endif
1021 }