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