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