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