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