sys/dev/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  * $FreeBSD: head/sys/dev/drm2/radeon/radeon_ttm.c 254885 2013-08-25 19:37:15Z dumbbell $
  33  */
  34
  35 #include <drm/ttm/ttm_bo_api.h>
  36 #include <drm/ttm/ttm_bo_driver.h>
  37 #include <drm/ttm/ttm_placement.h>
  38 #include <drm/ttm/ttm_module.h>
  39 #include <drm/ttm/ttm_page_alloc.h>
  40 #include <drm/drmP.h>
  41 #include <uapi_drm/radeon_drm.h>
  42 #include "radeon_reg.h"
  43 #include "radeon.h"
  44
  45 #define DRM_FILE_PAGE_OFFSET (0x100000000ULL >> PAGE_SHIFT)
  46
  47 static int radeon_ttm_debugfs_init(struct radeon_device *rdev);
  48
  49 static struct radeon_device *radeon_get_rdev(struct ttm_bo_device *bdev)
  50 {
  51         struct radeon_mman *mman;
  52         struct radeon_device *rdev;
  53
  54         mman = container_of(bdev, struct radeon_mman, bdev);
  55         rdev = container_of(mman, struct radeon_device, mman);
  56         return rdev;
  57 }
  58
  59
  60 /*
  61  * Global memory.
  62  */
  63 static int radeon_ttm_mem_global_init(struct drm_global_reference *ref)
  64 {
  65         return ttm_mem_global_init(ref->object);
  66 }
  67
  68 static void radeon_ttm_mem_global_release(struct drm_global_reference *ref)
  69 {
  70         ttm_mem_global_release(ref->object);
  71 }
  72
  73 static int radeon_ttm_global_init(struct radeon_device *rdev)
  74 {
  75         struct drm_global_reference *global_ref;
  76         int r;
  77
  78         rdev->mman.mem_global_referenced = false;
  79         global_ref = &rdev->mman.mem_global_ref;
  80         global_ref->global_type = DRM_GLOBAL_TTM_MEM;
  81         global_ref->size = sizeof(struct ttm_mem_global);
  82         global_ref->init = &radeon_ttm_mem_global_init;
  83         global_ref->release = &radeon_ttm_mem_global_release;
  84         r = drm_global_item_ref(global_ref);
  85         if (r != 0) {
  86                 DRM_ERROR("Failed setting up TTM memory accounting "
  87                           "subsystem.\n");
  88                 return r;
  89         }
  90
  91         rdev->mman.bo_global_ref.mem_glob =
  92                 rdev->mman.mem_global_ref.object;
  93         global_ref = &rdev->mman.bo_global_ref.ref;
  94         global_ref->global_type = DRM_GLOBAL_TTM_BO;
  95         global_ref->size = sizeof(struct ttm_bo_global);
  96         global_ref->init = &ttm_bo_global_init;
  97         global_ref->release = &ttm_bo_global_release;
  98         r = drm_global_item_ref(global_ref);
  99         if (r != 0) {
 100                 DRM_ERROR("Failed setting up TTM BO subsystem.\n");
 101                 drm_global_item_unref(&rdev->mman.mem_global_ref);
 102                 return r;
 103         }
 104
 105         rdev->mman.mem_global_referenced = true;
 106         return 0;
 107 }
 108
 109 static void radeon_ttm_global_fini(struct radeon_device *rdev)
 110 {
 111         if (rdev->mman.mem_global_referenced) {
 112                 drm_global_item_unref(&rdev->mman.bo_global_ref.ref);
 113                 drm_global_item_unref(&rdev->mman.mem_global_ref);
 114                 rdev->mman.mem_global_referenced = false;
 115         }
 116 }
 117
 118 static int radeon_invalidate_caches(struct ttm_bo_device *bdev, uint32_t flags)
 119 {
 120         return 0;
 121 }
 122
 123 static int radeon_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
 124                                 struct ttm_mem_type_manager *man)
 125 {
 126         struct radeon_device *rdev;
 127
 128         rdev = radeon_get_rdev(bdev);
 129
 130         switch (type) {
 131         case TTM_PL_SYSTEM:
 132                 /* System memory */
 133                 man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
 134                 man->available_caching = TTM_PL_MASK_CACHING;
 135                 man->default_caching = TTM_PL_FLAG_CACHED;
 136                 break;
 137         case TTM_PL_TT:
 138                 man->func = &ttm_bo_manager_func;
 139                 man->gpu_offset = rdev->mc.gtt_start;
 140                 man->available_caching = TTM_PL_MASK_CACHING;
 141                 man->default_caching = TTM_PL_FLAG_CACHED;
 142                 man->flags = TTM_MEMTYPE_FLAG_MAPPABLE | TTM_MEMTYPE_FLAG_CMA;
 143 #if __OS_HAS_AGP
 144                 if (rdev->flags & RADEON_IS_AGP) {
 145                         if (!(drm_core_has_AGP(rdev->ddev) && rdev->ddev->agp)) {
 146                                 DRM_ERROR("AGP is not enabled for memory type %u\n",
 147                                           (unsigned)type);
 148                                 return -EINVAL;
 149                         }
 150                         if (!rdev->ddev->agp->cant_use_aperture)
 151                                 man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
 152                         man->available_caching = TTM_PL_FLAG_UNCACHED |
 153                                                  TTM_PL_FLAG_WC;
 154                         man->default_caching = TTM_PL_FLAG_WC;
 155                 }
 156 #endif
 157                 break;
 158         case TTM_PL_VRAM:
 159                 /* "On-card" video ram */
 160                 man->func = &ttm_bo_manager_func;
 161                 man->gpu_offset = rdev->mc.vram_start;
 162                 man->flags = TTM_MEMTYPE_FLAG_FIXED |
 163                              TTM_MEMTYPE_FLAG_MAPPABLE;
 164                 man->available_caching = TTM_PL_FLAG_UNCACHED | TTM_PL_FLAG_WC;
 165                 man->default_caching = TTM_PL_FLAG_WC;
 166                 break;
 167         default:
 168                 DRM_ERROR("Unsupported memory type %u\n", (unsigned)type);
 169                 return -EINVAL;
 170         }
 171         return 0;
 172 }
 173
 174 static void radeon_evict_flags(struct ttm_buffer_object *bo,
 175                                 struct ttm_placement *placement)
 176 {
 177         struct radeon_bo *rbo;
 178         static u32 placements = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM;
 179
 180         if (!radeon_ttm_bo_is_radeon_bo(bo)) {
 181                 placement->fpfn = 0;
 182                 placement->lpfn = 0;
 183                 placement->placement = &placements;
 184                 placement->busy_placement = &placements;
 185                 placement->num_placement = 1;
 186                 placement->num_busy_placement = 1;
 187                 return;
 188         }
 189         rbo = container_of(bo, struct radeon_bo, tbo);
 190         switch (bo->mem.mem_type) {
 191         case TTM_PL_VRAM:
 192                 if (rbo->rdev->ring[RADEON_RING_TYPE_GFX_INDEX].ready == false)
 193                         radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_CPU);
 194                 else
 195                         radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_GTT);
 196                 break;
 197         case TTM_PL_TT:
 198         default:
 199                 radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_CPU);
 200         }
 201         *placement = rbo->placement;
 202 }
 203
 204 static int radeon_verify_access(struct ttm_buffer_object *bo)
 205 {
 206         return 0;
 207 }
 208
 209 static void radeon_move_null(struct ttm_buffer_object *bo,
 210                              struct ttm_mem_reg *new_mem)
 211 {
 212         struct ttm_mem_reg *old_mem = &bo->mem;
 213
 214         KASSERT(old_mem->mm_node == NULL, ("old_mem->mm_node != NULL"));
 215         *old_mem = *new_mem;
 216         new_mem->mm_node = NULL;
 217 }
 218
 219 static int radeon_move_blit(struct ttm_buffer_object *bo,
 220                         bool evict, bool no_wait_gpu,
 221                         struct ttm_mem_reg *new_mem,
 222                         struct ttm_mem_reg *old_mem)
 223 {
 224         struct radeon_device *rdev;
 225         uint64_t old_start, new_start;
 226         struct radeon_fence *fence;
 227         int r, ridx;
 228
 229         rdev = radeon_get_rdev(bo->bdev);
 230         ridx = radeon_copy_ring_index(rdev);
 231         old_start = old_mem->start << PAGE_SHIFT;
 232         new_start = new_mem->start << PAGE_SHIFT;
 233
 234         switch (old_mem->mem_type) {
 235         case TTM_PL_VRAM:
 236                 old_start += rdev->mc.vram_start;
 237                 break;
 238         case TTM_PL_TT:
 239                 old_start += rdev->mc.gtt_start;
 240                 break;
 241         default:
 242                 DRM_ERROR("Unknown placement %d\n", old_mem->mem_type);
 243                 return -EINVAL;
 244         }
 245         switch (new_mem->mem_type) {
 246         case TTM_PL_VRAM:
 247                 new_start += rdev->mc.vram_start;
 248                 break;
 249         case TTM_PL_TT:
 250                 new_start += rdev->mc.gtt_start;
 251                 break;
 252         default:
 253                 DRM_ERROR("Unknown placement %d\n", old_mem->mem_type);
 254                 return -EINVAL;
 255         }
 256         if (!rdev->ring[ridx].ready) {
 257                 DRM_ERROR("Trying to move memory with ring turned off.\n");
 258                 return -EINVAL;
 259         }
 260
 261         CTASSERT((PAGE_SIZE % RADEON_GPU_PAGE_SIZE) == 0);
 262
 263         /* sync other rings */
 264         fence = bo->sync_obj;
 265         r = radeon_copy(rdev, old_start, new_start,
 266                         new_mem->num_pages * (PAGE_SIZE / RADEON_GPU_PAGE_SIZE), /* GPU pages */
 267                         &fence);
 268         /* FIXME: handle copy error */
 269         r = ttm_bo_move_accel_cleanup(bo, (void *)fence,
 270                                       evict, no_wait_gpu, new_mem);
 271         radeon_fence_unref(&fence);
 272         return r;
 273 }
 274
 275 static int radeon_move_vram_ram(struct ttm_buffer_object *bo,
 276                                 bool evict, bool interruptible,
 277                                 bool no_wait_gpu,
 278                                 struct ttm_mem_reg *new_mem)
 279 {
 280         struct radeon_device *rdev;
 281         struct ttm_mem_reg *old_mem = &bo->mem;
 282         struct ttm_mem_reg tmp_mem;
 283         u32 placements;
 284         struct ttm_placement placement;
 285         int r;
 286
 287         rdev = radeon_get_rdev(bo->bdev);
 288         tmp_mem = *new_mem;
 289         tmp_mem.mm_node = NULL;
 290         placement.fpfn = 0;
 291         placement.lpfn = 0;
 292         placement.num_placement = 1;
 293         placement.placement = &placements;
 294         placement.num_busy_placement = 1;
 295         placement.busy_placement = &placements;
 296         placements = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
 297         r = ttm_bo_mem_space(bo, &placement, &tmp_mem,
 298                              interruptible, no_wait_gpu);
 299         if (unlikely(r)) {
 300                 return r;
 301         }
 302
 303         r = ttm_tt_set_placement_caching(bo->ttm, tmp_mem.placement);
 304         if (unlikely(r)) {
 305                 goto out_cleanup;
 306         }
 307
 308         r = ttm_tt_bind(bo->ttm, &tmp_mem);
 309         if (unlikely(r)) {
 310                 goto out_cleanup;
 311         }
 312         r = radeon_move_blit(bo, true, no_wait_gpu, &tmp_mem, old_mem);
 313         if (unlikely(r)) {
 314                 goto out_cleanup;
 315         }
 316         r = ttm_bo_move_ttm(bo, true, no_wait_gpu, new_mem);
 317 out_cleanup:
 318         ttm_bo_mem_put(bo, &tmp_mem);
 319         return r;
 320 }
 321
 322 static int radeon_move_ram_vram(struct ttm_buffer_object *bo,
 323                                 bool evict, bool interruptible,
 324                                 bool no_wait_gpu,
 325                                 struct ttm_mem_reg *new_mem)
 326 {
 327         struct radeon_device *rdev;
 328         struct ttm_mem_reg *old_mem = &bo->mem;
 329         struct ttm_mem_reg tmp_mem;
 330         struct ttm_placement placement;
 331         u32 placements;
 332         int r;
 333
 334         rdev = radeon_get_rdev(bo->bdev);
 335         tmp_mem = *new_mem;
 336         tmp_mem.mm_node = NULL;
 337         placement.fpfn = 0;
 338         placement.lpfn = 0;
 339         placement.num_placement = 1;
 340         placement.placement = &placements;
 341         placement.num_busy_placement = 1;
 342         placement.busy_placement = &placements;
 343         placements = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
 344         r = ttm_bo_mem_space(bo, &placement, &tmp_mem,
 345                              interruptible, no_wait_gpu);
 346         if (unlikely(r)) {
 347                 return r;
 348         }
 349         r = ttm_bo_move_ttm(bo, true, no_wait_gpu, &tmp_mem);
 350         if (unlikely(r)) {
 351                 goto out_cleanup;
 352         }
 353         r = radeon_move_blit(bo, true, no_wait_gpu, new_mem, old_mem);
 354         if (unlikely(r)) {
 355                 goto out_cleanup;
 356         }
 357 out_cleanup:
 358         ttm_bo_mem_put(bo, &tmp_mem);
 359         return r;
 360 }
 361
 362 static int radeon_bo_move(struct ttm_buffer_object *bo,
 363                         bool evict, bool interruptible,
 364                         bool no_wait_gpu,
 365                         struct ttm_mem_reg *new_mem)
 366 {
 367         struct radeon_device *rdev;
 368         struct ttm_mem_reg *old_mem = &bo->mem;
 369         int r;
 370
 371         rdev = radeon_get_rdev(bo->bdev);
 372         if (old_mem->mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
 373                 radeon_move_null(bo, new_mem);
 374                 return 0;
 375         }
 376         if ((old_mem->mem_type == TTM_PL_TT &&
 377              new_mem->mem_type == TTM_PL_SYSTEM) ||
 378             (old_mem->mem_type == TTM_PL_SYSTEM &&
 379              new_mem->mem_type == TTM_PL_TT)) {
 380                 /* bind is enough */
 381                 radeon_move_null(bo, new_mem);
 382                 return 0;
 383         }
 384         if (!rdev->ring[radeon_copy_ring_index(rdev)].ready ||
 385             rdev->asic->copy.copy == NULL) {
 386                 /* use memcpy */
 387                 goto memcpy;
 388         }
 389
 390         if (old_mem->mem_type == TTM_PL_VRAM &&
 391             new_mem->mem_type == TTM_PL_SYSTEM) {
 392                 r = radeon_move_vram_ram(bo, evict, interruptible,
 393                                         no_wait_gpu, new_mem);
 394         } else if (old_mem->mem_type == TTM_PL_SYSTEM &&
 395                    new_mem->mem_type == TTM_PL_VRAM) {
 396                 r = radeon_move_ram_vram(bo, evict, interruptible,
 397                                             no_wait_gpu, new_mem);
 398         } else {
 399                 r = radeon_move_blit(bo, evict, no_wait_gpu, new_mem, old_mem);
 400         }
 401
 402         if (r) {
 403 memcpy:
 404                 r = ttm_bo_move_memcpy(bo, evict, no_wait_gpu, new_mem);
 405         }
 406         return r;
 407 }
 408
 409 static int radeon_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
 410 {
 411         struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
 412         struct radeon_device *rdev = radeon_get_rdev(bdev);
 413
 414         mem->bus.addr = NULL;
 415         mem->bus.offset = 0;
 416         mem->bus.size = mem->num_pages << PAGE_SHIFT;
 417         mem->bus.base = 0;
 418         mem->bus.is_iomem = false;
 419         if (!(man->flags & TTM_MEMTYPE_FLAG_MAPPABLE))
 420                 return -EINVAL;
 421         switch (mem->mem_type) {
 422         case TTM_PL_SYSTEM:
 423                 /* system memory */
 424                 return 0;
 425         case TTM_PL_TT:
 426 #if __OS_HAS_AGP
 427                 if (rdev->flags & RADEON_IS_AGP) {
 428                         /* RADEON_IS_AGP is set only if AGP is active */
 429                         mem->bus.offset = mem->start << PAGE_SHIFT;
 430                         mem->bus.base = rdev->mc.agp_base;
 431                         mem->bus.is_iomem = !rdev->ddev->agp->cant_use_aperture;
 432                 }
 433 #endif
 434                 break;
 435         case TTM_PL_VRAM:
 436                 mem->bus.offset = mem->start << PAGE_SHIFT;
 437                 /* check if it's visible */
 438                 if ((mem->bus.offset + mem->bus.size) > rdev->mc.visible_vram_size)
 439                         return -EINVAL;
 440                 mem->bus.base = rdev->mc.aper_base;
 441                 mem->bus.is_iomem = true;
 442 #ifdef __alpha__
 443                 /*
 444                  * Alpha: use bus.addr to hold the ioremap() return,
 445                  * so we can modify bus.base below.
 446                  */
 447                 if (mem->placement & TTM_PL_FLAG_WC)
 448                         mem->bus.addr =
 449                                 ioremap_wc(mem->bus.base + mem->bus.offset,
 450                                            mem->bus.size);
 451                 else
 452                         mem->bus.addr =
 453                                 ioremap_nocache(mem->bus.base + mem->bus.offset,
 454                                                 mem->bus.size);
 455
 456                 /*
 457                  * Alpha: Use just the bus offset plus
 458                  * the hose/domain memory base for bus.base.
 459                  * It then can be used to build PTEs for VRAM
 460                  * access, as done in ttm_bo_vm_fault().
 461                  */
 462                 mem->bus.base = (mem->bus.base & 0x0ffffffffUL) +
 463                         rdev->ddev->hose->dense_mem_base;
 464 #endif
 465                 break;
 466         default:
 467                 return -EINVAL;
 468         }
 469         return 0;
 470 }
 471
 472 static void radeon_ttm_io_mem_free(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
 473 {
 474 }
 475
 476 static int radeon_sync_obj_wait(void *sync_obj, bool lazy, bool interruptible)
 477 {
 478         return radeon_fence_wait((struct radeon_fence *)sync_obj, interruptible);
 479 }
 480
 481 static int radeon_sync_obj_flush(void *sync_obj)
 482 {
 483         return 0;
 484 }
 485
 486 static void radeon_sync_obj_unref(void **sync_obj)
 487 {
 488         radeon_fence_unref((struct radeon_fence **)sync_obj);
 489 }
 490
 491 static void *radeon_sync_obj_ref(void *sync_obj)
 492 {
 493         return radeon_fence_ref((struct radeon_fence *)sync_obj);
 494 }
 495
 496 static bool radeon_sync_obj_signaled(void *sync_obj)
 497 {
 498         return radeon_fence_signaled((struct radeon_fence *)sync_obj);
 499 }
 500
 501 /*
 502  * TTM backend functions.
 503  */
 504 struct radeon_ttm_tt {
 505         struct ttm_dma_tt               ttm;
 506         struct radeon_device            *rdev;
 507         u64                             offset;
 508 };
 509
 510 static int radeon_ttm_backend_bind(struct ttm_tt *ttm,
 511                                    struct ttm_mem_reg *bo_mem)
 512 {
 513         struct radeon_ttm_tt *gtt = (void*)ttm;
 514         int r;
 515
 516         gtt->offset = (unsigned long)(bo_mem->start << PAGE_SHIFT);
 517         if (!ttm->num_pages) {
 518                 DRM_ERROR("nothing to bind %lu pages for mreg %p back %p!\n",
 519                      ttm->num_pages, bo_mem, ttm);
 520         }
 521         r = radeon_gart_bind(gtt->rdev, gtt->offset,
 522                              ttm->num_pages, ttm->pages, gtt->ttm.dma_address);
 523         if (r) {
 524                 DRM_ERROR("failed to bind %lu pages at 0x%08X\n",
 525                           ttm->num_pages, (unsigned)gtt->offset);
 526                 return r;
 527         }
 528         return 0;
 529 }
 530
 531 static int radeon_ttm_backend_unbind(struct ttm_tt *ttm)
 532 {
 533         struct radeon_ttm_tt *gtt = (void *)ttm;
 534
 535         radeon_gart_unbind(gtt->rdev, gtt->offset, ttm->num_pages);
 536         return 0;
 537 }
 538
 539 static void radeon_ttm_backend_destroy(struct ttm_tt *ttm)
 540 {
 541         struct radeon_ttm_tt *gtt = (void *)ttm;
 542
 543         ttm_dma_tt_fini(&gtt->ttm);
 544         drm_free(gtt, M_DRM);
 545 }
 546
 547 static struct ttm_backend_func radeon_backend_func = {
 548         .bind = &radeon_ttm_backend_bind,
 549         .unbind = &radeon_ttm_backend_unbind,
 550         .destroy = &radeon_ttm_backend_destroy,
 551 };
 552
 553 static struct ttm_tt *radeon_ttm_tt_create(struct ttm_bo_device *bdev,
 554                                     unsigned long size, uint32_t page_flags,
 555                                     vm_page_t dummy_read_page)
 556 {
 557         struct radeon_device *rdev;
 558         struct radeon_ttm_tt *gtt;
 559
 560         rdev = radeon_get_rdev(bdev);
 561 #if __OS_HAS_AGP
 562 #ifdef DUMBBELL_WIP
 563         if (rdev->flags & RADEON_IS_AGP) {
 564                 return ttm_agp_tt_create(bdev, rdev->ddev->agp->agpdev,
 565                                          size, page_flags, dummy_read_page);
 566         }
 567 #endif /* DUMBBELL_WIP */
 568 #endif
 569
 570         gtt = kmalloc(sizeof(struct radeon_ttm_tt), M_DRM,
 571                       M_WAITOK | M_ZERO);
 572         if (gtt == NULL) {
 573                 return NULL;
 574         }
 575         gtt->ttm.ttm.func = &radeon_backend_func;
 576         gtt->rdev = rdev;
 577         if (ttm_dma_tt_init(&gtt->ttm, bdev, size, page_flags, dummy_read_page)) {
 578                 drm_free(gtt, M_DRM);
 579                 return NULL;
 580         }
 581         return &gtt->ttm.ttm;
 582 }
 583
 584 static int radeon_ttm_tt_populate(struct ttm_tt *ttm)
 585 {
 586         struct radeon_device *rdev;
 587         struct radeon_ttm_tt *gtt = (void *)ttm;
 588         unsigned i;
 589         int r;
 590 #ifdef DUMBBELL_WIP
 591         bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
 592 #endif /* DUMBBELL_WIP */
 593
 594         if (ttm->state != tt_unpopulated)
 595                 return 0;
 596
 597 #ifdef DUMBBELL_WIP
 598         /*
 599          * Maybe unneeded on FreeBSD.
 600          *   -- dumbbell@
 601          */
 602         if (slave && ttm->sg) {
 603                 drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
 604                                                  gtt->ttm.dma_address, ttm->num_pages);
 605                 ttm->state = tt_unbound;
 606                 return 0;
 607         }
 608 #endif /* DUMBBELL_WIP */
 609
 610         rdev = radeon_get_rdev(ttm->bdev);
 611 #if __OS_HAS_AGP
 612 #ifdef DUMBBELL_WIP
 613         if (rdev->flags & RADEON_IS_AGP) {
 614                 return ttm_agp_tt_populate(ttm);
 615         }
 616 #endif /* DUMBBELL_WIP */
 617 #endif
 618
 619 #ifdef CONFIG_SWIOTLB
 620         if (swiotlb_nr_tbl()) {
 621                 return ttm_dma_populate(&gtt->ttm, rdev->dev);
 622         }
 623 #endif
 624
 625         r = ttm_pool_populate(ttm);
 626         if (r) {
 627                 return r;
 628         }
 629
 630         for (i = 0; i < ttm->num_pages; i++) {
 631                 gtt->ttm.dma_address[i] = VM_PAGE_TO_PHYS(ttm->pages[i]);
 632 #ifdef DUMBBELL_WIP
 633                 gtt->ttm.dma_address[i] = pci_map_page(rdev->pdev, ttm->pages[i],
 634                                                        0, PAGE_SIZE,
 635                                                        PCI_DMA_BIDIRECTIONAL);
 636                 if (pci_dma_mapping_error(rdev->pdev, gtt->ttm.dma_address[i])) {
 637                         while (--i) {
 638                                 pci_unmap_page(rdev->pdev, gtt->ttm.dma_address[i],
 639                                                PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
 640                                 gtt->ttm.dma_address[i] = 0;
 641                         }
 642                         ttm_pool_unpopulate(ttm);
 643                         return -EFAULT;
 644                 }
 645 #endif /* DUMBBELL_WIP */
 646         }
 647         return 0;
 648 }
 649
 650 static void radeon_ttm_tt_unpopulate(struct ttm_tt *ttm)
 651 {
 652         struct radeon_device *rdev;
 653         struct radeon_ttm_tt *gtt = (void *)ttm;
 654         unsigned i;
 655         bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
 656
 657         if (slave)
 658                 return;
 659
 660         rdev = radeon_get_rdev(ttm->bdev);
 661 #if __OS_HAS_AGP
 662 #ifdef DUMBBELL_WIP
 663         if (rdev->flags & RADEON_IS_AGP) {
 664                 ttm_agp_tt_unpopulate(ttm);
 665                 return;
 666         }
 667 #endif /* DUMBBELL_WIP */
 668 #endif
 669
 670 #ifdef CONFIG_SWIOTLB
 671         if (swiotlb_nr_tbl()) {
 672                 ttm_dma_unpopulate(&gtt->ttm, rdev->dev);
 673                 return;
 674         }
 675 #endif
 676
 677         for (i = 0; i < ttm->num_pages; i++) {
 678                 if (gtt->ttm.dma_address[i]) {
 679                         gtt->ttm.dma_address[i] = 0;
 680 #ifdef DUMBBELL_WIP
 681                         pci_unmap_page(rdev->pdev, gtt->ttm.dma_address[i],
 682                                        PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
 683 #endif /* DUMBBELL_WIP */
 684                 }
 685         }
 686
 687         ttm_pool_unpopulate(ttm);
 688 }
 689
 690 static struct ttm_bo_driver radeon_bo_driver = {
 691         .ttm_tt_create = &radeon_ttm_tt_create,
 692         .ttm_tt_populate = &radeon_ttm_tt_populate,
 693         .ttm_tt_unpopulate = &radeon_ttm_tt_unpopulate,
 694         .invalidate_caches = &radeon_invalidate_caches,
 695         .init_mem_type = &radeon_init_mem_type,
 696         .evict_flags = &radeon_evict_flags,
 697         .move = &radeon_bo_move,
 698         .verify_access = &radeon_verify_access,
 699         .sync_obj_signaled = &radeon_sync_obj_signaled,
 700         .sync_obj_wait = &radeon_sync_obj_wait,
 701         .sync_obj_flush = &radeon_sync_obj_flush,
 702         .sync_obj_unref = &radeon_sync_obj_unref,
 703         .sync_obj_ref = &radeon_sync_obj_ref,
 704         .move_notify = &radeon_bo_move_notify,
 705         .fault_reserve_notify = &radeon_bo_fault_reserve_notify,
 706         .io_mem_reserve = &radeon_ttm_io_mem_reserve,
 707         .io_mem_free = &radeon_ttm_io_mem_free,
 708 };
 709
 710 int radeon_ttm_init(struct radeon_device *rdev)
 711 {
 712         int r, r2;
 713
 714         r = radeon_ttm_global_init(rdev);
 715         if (r) {
 716                 return r;
 717         }
 718         /* No others user of address space so set it to 0 */
 719         r = ttm_bo_device_init(&rdev->mman.bdev,
 720                                rdev->mman.bo_global_ref.ref.object,
 721                                &radeon_bo_driver, DRM_FILE_PAGE_OFFSET,
 722                                rdev->need_dma32);
 723         if (r) {
 724                 DRM_ERROR("failed initializing buffer object driver(%d).\n", r);
 725                 return r;
 726         }
 727         rdev->mman.initialized = true;
 728         rdev->ddev->drm_ttm_bdev = &rdev->mman.bdev;
 729         r = ttm_bo_init_mm(&rdev->mman.bdev, TTM_PL_VRAM,
 730                                 rdev->mc.real_vram_size >> PAGE_SHIFT);
 731         if (r) {
 732                 DRM_ERROR("Failed initializing VRAM heap.\n");
 733                 return r;
 734         }
 735         r = radeon_bo_create(rdev, 256 * 1024, PAGE_SIZE, true,
 736                              RADEON_GEM_DOMAIN_VRAM,
 737                              NULL, &rdev->stollen_vga_memory);
 738         if (r) {
 739                 return r;
 740         }
 741         r = radeon_bo_reserve(rdev->stollen_vga_memory, false);
 742         if (r) {
 743                 radeon_bo_unref(&rdev->stollen_vga_memory);
 744                 return r;
 745         }
 746         r = radeon_bo_pin(rdev->stollen_vga_memory, RADEON_GEM_DOMAIN_VRAM, NULL);
 747         radeon_bo_unreserve(rdev->stollen_vga_memory);
 748         if (r) {
 749                 radeon_bo_unref(&rdev->stollen_vga_memory);
 750                 return r;
 751         }
 752         DRM_INFO("radeon: %uM of VRAM memory ready\n",
 753                  (unsigned)rdev->mc.real_vram_size / (1024 * 1024));
 754         r = ttm_bo_init_mm(&rdev->mman.bdev, TTM_PL_TT,
 755                                 rdev->mc.gtt_size >> PAGE_SHIFT);
 756         if (r) {
 757                 DRM_ERROR("Failed initializing GTT heap.\n");
 758                 r2 = radeon_bo_reserve(rdev->stollen_vga_memory, false);
 759                 if (likely(r2 == 0)) {
 760                         radeon_bo_unpin(rdev->stollen_vga_memory);
 761                         radeon_bo_unreserve(rdev->stollen_vga_memory);
 762                 }
 763                 radeon_bo_unref(&rdev->stollen_vga_memory);
 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                 r2 = radeon_bo_reserve(rdev->stollen_vga_memory, false);
 773                 if (likely(r2 == 0)) {
 774                         radeon_bo_unpin(rdev->stollen_vga_memory);
 775                         radeon_bo_unreserve(rdev->stollen_vga_memory);
 776                 }
 777                 radeon_bo_unref(&rdev->stollen_vga_memory);
 778                 return r;
 779         }
 780         return 0;
 781 }
 782
 783 void radeon_ttm_fini(struct radeon_device *rdev)
 784 {
 785         int r;
 786
 787         if (!rdev->mman.initialized)
 788                 return;
 789         if (rdev->stollen_vga_memory) {
 790                 r = radeon_bo_reserve(rdev->stollen_vga_memory, false);
 791                 if (r == 0) {
 792                         radeon_bo_unpin(rdev->stollen_vga_memory);
 793                         radeon_bo_unreserve(rdev->stollen_vga_memory);
 794                 }
 795                 radeon_bo_unref(&rdev->stollen_vga_memory);
 796         }
 797         ttm_bo_clean_mm(&rdev->mman.bdev, TTM_PL_VRAM);
 798         ttm_bo_clean_mm(&rdev->mman.bdev, TTM_PL_TT);
 799         ttm_bo_device_release(&rdev->mman.bdev);
 800         radeon_gart_fini(rdev);
 801         radeon_ttm_global_fini(rdev);
 802         rdev->mman.initialized = false;
 803         DRM_INFO("radeon: ttm finalized\n");
 804 }
 805
 806 /* this should only be called at bootup or when userspace
 807  * isn't running */
 808 void radeon_ttm_set_active_vram_size(struct radeon_device *rdev, u64 size)
 809 {
 810         struct ttm_mem_type_manager *man;
 811
 812         if (!rdev->mman.initialized)
 813                 return;
 814
 815         man = &rdev->mman.bdev.man[TTM_PL_VRAM];
 816         /* this just adjusts TTM size idea, which sets lpfn to the correct value */
 817         man->size = size >> PAGE_SHIFT;
 818 }
 819
 820 #ifdef DUMBBELL_WIP
 821 static struct vm_operations_struct radeon_ttm_vm_ops;
 822 static const struct vm_operations_struct *ttm_vm_ops = NULL;
 823
 824 static int radeon_ttm_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 825 {
 826         struct ttm_buffer_object *bo;
 827         struct radeon_device *rdev;
 828         int r;
 829
 830         bo = (struct ttm_buffer_object *)vma->vm_private_data;
 831         if (bo == NULL) {
 832                 return VM_FAULT_NOPAGE;
 833         }
 834         rdev = radeon_get_rdev(bo->bdev);
 835         lockmgr(&rdev->pm.mclk_lock, LK_SHARED);
 836         r = ttm_vm_ops->fault(vma, vmf);
 837         lockmgr(&rdev->pm.mclk_lock, LK_RELEASE);
 838         return r;
 839 }
 840
 841 int radeon_mmap(struct file *filp, struct vm_area_struct *vma)
 842 {
 843         struct drm_file *file_priv;
 844         struct radeon_device *rdev;
 845         int r;
 846
 847         if (unlikely(vma->vm_pgoff < DRM_FILE_PAGE_OFFSET)) {
 848                 return drm_mmap(filp, vma);
 849         }
 850
 851         file_priv = filp->private_data;
 852         rdev = file_priv->minor->dev->dev_private;
 853         if (rdev == NULL) {
 854                 return -EINVAL;
 855         }
 856         r = ttm_bo_mmap(filp, vma, &rdev->mman.bdev);
 857         if (unlikely(r != 0)) {
 858                 return r;
 859         }
 860         if (unlikely(ttm_vm_ops == NULL)) {
 861                 ttm_vm_ops = vma->vm_ops;
 862                 radeon_ttm_vm_ops = *ttm_vm_ops;
 863                 radeon_ttm_vm_ops.fault = &radeon_ttm_fault;
 864         }
 865         vma->vm_ops = &radeon_ttm_vm_ops;
 866         return 0;
 867 }
 868 #endif /* DUMBBELL_WIP */
 869
 870
 871 #define RADEON_DEBUGFS_MEM_TYPES 2
 872
 873 #if defined(CONFIG_DEBUG_FS)
 874 static int radeon_mm_dump_table(struct seq_file *m, void *data)
 875 {
 876         struct drm_info_node *node = (struct drm_info_node *)m->private;
 877         struct drm_mm *mm = (struct drm_mm *)node->info_ent->data;
 878         struct drm_device *dev = node->minor->dev;
 879         struct radeon_device *rdev = dev->dev_private;
 880         int ret;
 881         struct ttm_bo_global *glob = rdev->mman.bdev.glob;
 882
 883         spin_lock(&glob->lru_lock);
 884         ret = drm_mm_dump_table(m, mm);
 885         spin_unlock(&glob->lru_lock);
 886         return ret;
 887 }
 888 #endif
 889
 890 static int radeon_ttm_debugfs_init(struct radeon_device *rdev)
 891 {
 892 #if defined(CONFIG_DEBUG_FS)
 893         static struct drm_info_list radeon_mem_types_list[RADEON_DEBUGFS_MEM_TYPES+2];
 894         static char radeon_mem_types_names[RADEON_DEBUGFS_MEM_TYPES+2][32];
 895         unsigned i;
 896
 897         for (i = 0; i < RADEON_DEBUGFS_MEM_TYPES; i++) {
 898                 if (i == 0)
 899                         sprintf(radeon_mem_types_names[i], "radeon_vram_mm");
 900                 else
 901                         sprintf(radeon_mem_types_names[i], "radeon_gtt_mm");
 902                 radeon_mem_types_list[i].name = radeon_mem_types_names[i];
 903                 radeon_mem_types_list[i].show = &radeon_mm_dump_table;
 904                 radeon_mem_types_list[i].driver_features = 0;
 905                 if (i == 0)
 906                         radeon_mem_types_list[i].data = rdev->mman.bdev.man[TTM_PL_VRAM].priv;
 907                 else
 908                         radeon_mem_types_list[i].data = rdev->mman.bdev.man[TTM_PL_TT].priv;
 909
 910         }
 911         /* Add ttm page pool to debugfs */
 912         sprintf(radeon_mem_types_names[i], "ttm_page_pool");
 913         radeon_mem_types_list[i].name = radeon_mem_types_names[i];
 914         radeon_mem_types_list[i].show = &ttm_page_alloc_debugfs;
 915         radeon_mem_types_list[i].driver_features = 0;
 916         radeon_mem_types_list[i++].data = NULL;
 917 #ifdef CONFIG_SWIOTLB
 918         if (swiotlb_nr_tbl()) {
 919                 sprintf(radeon_mem_types_names[i], "ttm_dma_page_pool");
 920                 radeon_mem_types_list[i].name = radeon_mem_types_names[i];
 921                 radeon_mem_types_list[i].show = &ttm_dma_page_alloc_debugfs;
 922                 radeon_mem_types_list[i].driver_features = 0;
 923                 radeon_mem_types_list[i++].data = NULL;
 924         }
 925 #endif
 926         return radeon_debugfs_add_files(rdev, radeon_mem_types_list, i);
 927
 928 #endif
 929         return 0;
 930 }