drivers/gpu/drm/radeon/radeon_object.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 #include <linux/list.h>
  33 #include <linux/slab.h>
  34 #include <drm/drmP.h>
  35 #include <drm/radeon_drm.h>
  36 #include <drm/drm_cache.h>
  37 #include "radeon.h"
  38 #include "radeon_trace.h"
  39
  40
  41 int radeon_ttm_init(struct radeon_device *rdev);
  42 void radeon_ttm_fini(struct radeon_device *rdev);
  43 static void radeon_bo_clear_surface_reg(struct radeon_bo *bo);
  44
  45 /*
  46  * To exclude mutual BO access we rely on bo_reserve exclusion, as all
  47  * function are calling it.
  48  */
  49
  50 static void radeon_update_memory_usage(struct radeon_bo *bo,
  51                                        unsigned mem_type, int sign)
  52 {
  53         struct radeon_device *rdev = bo->rdev;
  54         u64 size = (u64)bo->tbo.num_pages << PAGE_SHIFT;
  55
  56         switch (mem_type) {
  57         case TTM_PL_TT:
  58                 if (sign > 0)
  59                         atomic64_add(size, &rdev->gtt_usage);
  60                 else
  61                         atomic64_sub(size, &rdev->gtt_usage);
  62                 break;
  63         case TTM_PL_VRAM:
  64                 if (sign > 0)
  65                         atomic64_add(size, &rdev->vram_usage);
  66                 else
  67                         atomic64_sub(size, &rdev->vram_usage);
  68                 break;
  69         }
  70 }
  71
  72 static void radeon_ttm_bo_destroy(struct ttm_buffer_object *tbo)
  73 {
  74         struct radeon_bo *bo;
  75
  76         bo = container_of(tbo, struct radeon_bo, tbo);
  77
  78         radeon_update_memory_usage(bo, bo->tbo.mem.mem_type, -1);
  79
  80         mutex_lock(&bo->rdev->gem.mutex);
  81         list_del_init(&bo->list);
  82         mutex_unlock(&bo->rdev->gem.mutex);
  83         radeon_bo_clear_surface_reg(bo);
  84         WARN_ON(!list_empty(&bo->va));
  85         drm_gem_object_release(&bo->gem_base);
  86         kfree(bo);
  87 }
  88
  89 bool radeon_ttm_bo_is_radeon_bo(struct ttm_buffer_object *bo)
  90 {
  91         if (bo->destroy == &radeon_ttm_bo_destroy)
  92                 return true;
  93         return false;
  94 }
  95
  96 void radeon_ttm_placement_from_domain(struct radeon_bo *rbo, u32 domain)
  97 {
  98         u32 c = 0, i;
  99
 100         rbo->placement.placement = rbo->placements;
 101         rbo->placement.busy_placement = rbo->placements;
 102         if (domain & RADEON_GEM_DOMAIN_VRAM) {
 103                 /* Try placing BOs which don't need CPU access outside of the
 104                  * CPU accessible part of VRAM
 105                  */
 106                 if ((rbo->flags & RADEON_GEM_NO_CPU_ACCESS) &&
 107                     rbo->rdev->mc.visible_vram_size < rbo->rdev->mc.real_vram_size) {
 108                         rbo->placements[c].fpfn =
 109                                 rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
 110                         rbo->placements[c++].flags = TTM_PL_FLAG_WC |
 111                                                      TTM_PL_FLAG_UNCACHED |
 112                                                      TTM_PL_FLAG_VRAM;
 113                 }
 114
 115                 rbo->placements[c].fpfn = 0;
 116                 rbo->placements[c++].flags = TTM_PL_FLAG_WC |
 117                                              TTM_PL_FLAG_UNCACHED |
 118                                              TTM_PL_FLAG_VRAM;
 119         }
 120
 121         if (domain & RADEON_GEM_DOMAIN_GTT) {
 122                 if (rbo->flags & RADEON_GEM_GTT_UC) {
 123                         rbo->placements[c].fpfn = 0;
 124                         rbo->placements[c++].flags = TTM_PL_FLAG_UNCACHED |
 125                                 TTM_PL_FLAG_TT;
 126
 127                 } else if ((rbo->flags & RADEON_GEM_GTT_WC) ||
 128                            (rbo->rdev->flags & RADEON_IS_AGP)) {
 129                         rbo->placements[c].fpfn = 0;
 130                         rbo->placements[c++].flags = TTM_PL_FLAG_WC |
 131                                 TTM_PL_FLAG_UNCACHED |
 132                                 TTM_PL_FLAG_TT;
 133                 } else {
 134                         rbo->placements[c].fpfn = 0;
 135                         rbo->placements[c++].flags = TTM_PL_FLAG_CACHED |
 136                                                      TTM_PL_FLAG_TT;
 137                 }
 138         }
 139
 140         if (domain & RADEON_GEM_DOMAIN_CPU) {
 141                 if (rbo->flags & RADEON_GEM_GTT_UC) {
 142                         rbo->placements[c].fpfn = 0;
 143                         rbo->placements[c++].flags = TTM_PL_FLAG_UNCACHED |
 144                                 TTM_PL_FLAG_SYSTEM;
 145
 146                 } else if ((rbo->flags & RADEON_GEM_GTT_WC) ||
 147                     rbo->rdev->flags & RADEON_IS_AGP) {
 148                         rbo->placements[c].fpfn = 0;
 149                         rbo->placements[c++].flags = TTM_PL_FLAG_WC |
 150                                 TTM_PL_FLAG_UNCACHED |
 151                                 TTM_PL_FLAG_SYSTEM;
 152                 } else {
 153                         rbo->placements[c].fpfn = 0;
 154                         rbo->placements[c++].flags = TTM_PL_FLAG_CACHED |
 155                                                      TTM_PL_FLAG_SYSTEM;
 156                 }
 157         }
 158         if (!c) {
 159                 rbo->placements[c].fpfn = 0;
 160                 rbo->placements[c++].flags = TTM_PL_MASK_CACHING |
 161                                              TTM_PL_FLAG_SYSTEM;
 162         }
 163
 164         rbo->placement.num_placement = c;
 165         rbo->placement.num_busy_placement = c;
 166
 167         for (i = 0; i < c; ++i) {
 168                 if ((rbo->flags & RADEON_GEM_CPU_ACCESS) &&
 169                     (rbo->placements[i].flags & TTM_PL_FLAG_VRAM) &&
 170                     !rbo->placements[i].fpfn)
 171                         rbo->placements[i].lpfn =
 172                                 rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
 173                 else
 174                         rbo->placements[i].lpfn = 0;
 175         }
 176 }
 177
 178 int radeon_bo_create(struct radeon_device *rdev,
 179                      unsigned long size, int byte_align, bool kernel,
 180                      u32 domain, u32 flags, struct sg_table *sg,
 181                      struct reservation_object *resv,
 182                      struct radeon_bo **bo_ptr)
 183 {
 184         struct radeon_bo *bo;
 185         enum ttm_bo_type type;
 186         unsigned long page_align = roundup(byte_align, PAGE_SIZE) >> PAGE_SHIFT;
 187         size_t acc_size;
 188         int r;
 189
 190         size = ALIGN(size, PAGE_SIZE);
 191
 192         if (kernel) {
 193                 type = ttm_bo_type_kernel;
 194         } else if (sg) {
 195                 type = ttm_bo_type_sg;
 196         } else {
 197                 type = ttm_bo_type_device;
 198         }
 199         *bo_ptr = NULL;
 200
 201         acc_size = ttm_bo_dma_acc_size(&rdev->mman.bdev, size,
 202                                        sizeof(struct radeon_bo));
 203
 204         bo = kzalloc(sizeof(struct radeon_bo), GFP_KERNEL);
 205         if (bo == NULL)
 206                 return -ENOMEM;
 207         r = drm_gem_object_init(rdev->ddev, &bo->gem_base, size);
 208         if (unlikely(r)) {
 209                 kfree(bo);
 210                 return r;
 211         }
 212         bo->rdev = rdev;
 213         bo->surface_reg = -1;
 214         INIT_LIST_HEAD(&bo->list);
 215         INIT_LIST_HEAD(&bo->va);
 216         bo->initial_domain = domain & (RADEON_GEM_DOMAIN_VRAM |
 217                                        RADEON_GEM_DOMAIN_GTT |
 218                                        RADEON_GEM_DOMAIN_CPU);
 219
 220         bo->flags = flags;
 221         /* PCI GART is always snooped */
 222         if (!(rdev->flags & RADEON_IS_PCIE))
 223                 bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
 224
 225         /* Write-combined CPU mappings of GTT cause GPU hangs with RV6xx
 226          * See https://bugs.freedesktop.org/show_bug.cgi?id=91268
 227          */
 228         if (rdev->family >= CHIP_RV610 && rdev->family <= CHIP_RV635)
 229                 bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
 230
 231 #ifdef CONFIG_X86_32
 232         /* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
 233          * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
 234          */
 235         bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
 236 #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
 237         /* Don't try to enable write-combining when it can't work, or things
 238          * may be slow
 239          * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
 240          */
 241 #ifndef CONFIG_COMPILE_TEST
 242 #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
 243          thanks to write-combining
 244 #endif
 245
 246         if (bo->flags & RADEON_GEM_GTT_WC)
 247                 DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
 248                               "better performance thanks to write-combining\n");
 249         bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
 250 #else
 251         /* For architectures that don't support WC memory,
 252          * mask out the WC flag from the BO
 253          */
 254         if (!drm_arch_can_wc_memory())
 255                 bo->flags &= ~RADEON_GEM_GTT_WC;
 256 #endif
 257
 258         radeon_ttm_placement_from_domain(bo, domain);
 259         /* Kernel allocation are uninterruptible */
 260         down_read(&rdev->pm.mclk_lock);
 261         r = ttm_bo_init(&rdev->mman.bdev, &bo->tbo, size, type,
 262                         &bo->placement, page_align, !kernel, NULL,
 263                         acc_size, sg, resv, &radeon_ttm_bo_destroy);
 264         up_read(&rdev->pm.mclk_lock);
 265         if (unlikely(r != 0)) {
 266                 return r;
 267         }
 268         *bo_ptr = bo;
 269
 270         trace_radeon_bo_create(bo);
 271
 272         return 0;
 273 }
 274
 275 int radeon_bo_kmap(struct radeon_bo *bo, void **ptr)
 276 {
 277         bool is_iomem;
 278         int r;
 279
 280         if (bo->kptr) {
 281                 if (ptr) {
 282                         *ptr = bo->kptr;
 283                 }
 284                 return 0;
 285         }
 286         r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.num_pages, &bo->kmap);
 287         if (r) {
 288                 return r;
 289         }
 290         bo->kptr = ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
 291         if (ptr) {
 292                 *ptr = bo->kptr;
 293         }
 294         radeon_bo_check_tiling(bo, 0, 0);
 295         return 0;
 296 }
 297
 298 void radeon_bo_kunmap(struct radeon_bo *bo)
 299 {
 300         if (bo->kptr == NULL)
 301                 return;
 302         bo->kptr = NULL;
 303         radeon_bo_check_tiling(bo, 0, 0);
 304         ttm_bo_kunmap(&bo->kmap);
 305 }
 306
 307 struct radeon_bo *radeon_bo_ref(struct radeon_bo *bo)
 308 {
 309         if (bo == NULL)
 310                 return NULL;
 311
 312         ttm_bo_reference(&bo->tbo);
 313         return bo;
 314 }
 315
 316 void radeon_bo_unref(struct radeon_bo **bo)
 317 {
 318         struct ttm_buffer_object *tbo;
 319         struct radeon_device *rdev;
 320
 321         if ((*bo) == NULL)
 322                 return;
 323         rdev = (*bo)->rdev;
 324         tbo = &((*bo)->tbo);
 325         ttm_bo_unref(&tbo);
 326         if (tbo == NULL)
 327                 *bo = NULL;
 328 }
 329
 330 int radeon_bo_pin_restricted(struct radeon_bo *bo, u32 domain, u64 max_offset,
 331                              u64 *gpu_addr)
 332 {
 333         int r, i;
 334
 335         if (radeon_ttm_tt_has_userptr(bo->tbo.ttm))
 336                 return -EPERM;
 337
 338         if (bo->pin_count) {
 339                 bo->pin_count++;
 340                 if (gpu_addr)
 341                         *gpu_addr = radeon_bo_gpu_offset(bo);
 342
 343                 if (max_offset != 0) {
 344                         u64 domain_start;
 345
 346                         if (domain == RADEON_GEM_DOMAIN_VRAM)
 347                                 domain_start = bo->rdev->mc.vram_start;
 348                         else
 349                                 domain_start = bo->rdev->mc.gtt_start;
 350                         WARN_ON_ONCE(max_offset <
 351                                      (radeon_bo_gpu_offset(bo) - domain_start));
 352                 }
 353
 354                 return 0;
 355         }
 356         radeon_ttm_placement_from_domain(bo, domain);
 357         for (i = 0; i < bo->placement.num_placement; i++) {
 358                 /* force to pin into visible video ram */
 359                 if ((bo->placements[i].flags & TTM_PL_FLAG_VRAM) &&
 360                     !(bo->flags & RADEON_GEM_NO_CPU_ACCESS) &&
 361                     (!max_offset || max_offset > bo->rdev->mc.visible_vram_size))
 362                         bo->placements[i].lpfn =
 363                                 bo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
 364                 else
 365                         bo->placements[i].lpfn = max_offset >> PAGE_SHIFT;
 366
 367                 bo->placements[i].flags |= TTM_PL_FLAG_NO_EVICT;
 368         }
 369
 370         r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
 371         if (likely(r == 0)) {
 372                 bo->pin_count = 1;
 373                 if (gpu_addr != NULL)
 374                         *gpu_addr = radeon_bo_gpu_offset(bo);
 375                 if (domain == RADEON_GEM_DOMAIN_VRAM)
 376                         bo->rdev->vram_pin_size += radeon_bo_size(bo);
 377                 else
 378                         bo->rdev->gart_pin_size += radeon_bo_size(bo);
 379         } else {
 380                 dev_err(bo->rdev->dev, "%p pin failed\n", bo);
 381         }
 382         return r;
 383 }
 384
 385 int radeon_bo_pin(struct radeon_bo *bo, u32 domain, u64 *gpu_addr)
 386 {
 387         return radeon_bo_pin_restricted(bo, domain, 0, gpu_addr);
 388 }
 389
 390 int radeon_bo_unpin(struct radeon_bo *bo)
 391 {
 392         int r, i;
 393
 394         if (!bo->pin_count) {
 395                 dev_warn(bo->rdev->dev, "%p unpin not necessary\n", bo);
 396                 return 0;
 397         }
 398         bo->pin_count--;
 399         if (bo->pin_count)
 400                 return 0;
 401         for (i = 0; i < bo->placement.num_placement; i++) {
 402                 bo->placements[i].lpfn = 0;
 403                 bo->placements[i].flags &= ~TTM_PL_FLAG_NO_EVICT;
 404         }
 405         r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
 406         if (likely(r == 0)) {
 407                 if (bo->tbo.mem.mem_type == TTM_PL_VRAM)
 408                         bo->rdev->vram_pin_size -= radeon_bo_size(bo);
 409                 else
 410                         bo->rdev->gart_pin_size -= radeon_bo_size(bo);
 411         } else {
 412                 dev_err(bo->rdev->dev, "%p validate failed for unpin\n", bo);
 413         }
 414         return r;
 415 }
 416
 417 int radeon_bo_evict_vram(struct radeon_device *rdev)
 418 {
 419         /* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */
 420         if (0 && (rdev->flags & RADEON_IS_IGP)) {
 421                 if (rdev->mc.igp_sideport_enabled == false)
 422                         /* Useless to evict on IGP chips */
 423                         return 0;
 424         }
 425         return ttm_bo_evict_mm(&rdev->mman.bdev, TTM_PL_VRAM);
 426 }
 427
 428 void radeon_bo_force_delete(struct radeon_device *rdev)
 429 {
 430         struct radeon_bo *bo, *n;
 431
 432         if (list_empty(&rdev->gem.objects)) {
 433                 return;
 434         }
 435         dev_err(rdev->dev, "Userspace still has active objects !\n");
 436         list_for_each_entry_safe(bo, n, &rdev->gem.objects, list) {
 437                 dev_err(rdev->dev, "%p %p %lu %lu force free\n",
 438                         &bo->gem_base, bo, (unsigned long)bo->gem_base.size,
 439                         *((unsigned long *)&bo->gem_base.refcount));
 440                 mutex_lock(&bo->rdev->gem.mutex);
 441                 list_del_init(&bo->list);
 442                 mutex_unlock(&bo->rdev->gem.mutex);
 443                 /* this should unref the ttm bo */
 444                 drm_gem_object_unreference_unlocked(&bo->gem_base);
 445         }
 446 }
 447
 448 int radeon_bo_init(struct radeon_device *rdev)
 449 {
 450         /* reserve PAT memory space to WC for VRAM */
 451         arch_io_reserve_memtype_wc(rdev->mc.aper_base,
 452                                    rdev->mc.aper_size);
 453
 454         /* Add an MTRR for the VRAM */
 455         if (!rdev->fastfb_working) {
 456                 rdev->mc.vram_mtrr = arch_phys_wc_add(rdev->mc.aper_base,
 457                                                       rdev->mc.aper_size);
 458         }
 459         DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
 460                 rdev->mc.mc_vram_size >> 20,
 461                 (unsigned long long)rdev->mc.aper_size >> 20);
 462         DRM_INFO("RAM width %dbits %cDR\n",
 463                         rdev->mc.vram_width, rdev->mc.vram_is_ddr ? 'D' : 'S');
 464         return radeon_ttm_init(rdev);
 465 }
 466
 467 void radeon_bo_fini(struct radeon_device *rdev)
 468 {
 469         radeon_ttm_fini(rdev);
 470         arch_phys_wc_del(rdev->mc.vram_mtrr);
 471         arch_io_free_memtype_wc(rdev->mc.aper_base, rdev->mc.aper_size);
 472 }
 473
 474 /* Returns how many bytes TTM can move per IB.
 475  */
 476 static u64 radeon_bo_get_threshold_for_moves(struct radeon_device *rdev)
 477 {
 478         u64 real_vram_size = rdev->mc.real_vram_size;
 479         u64 vram_usage = atomic64_read(&rdev->vram_usage);
 480
 481         /* This function is based on the current VRAM usage.
 482          *
 483          * - If all of VRAM is free, allow relocating the number of bytes that
 484          *   is equal to 1/4 of the size of VRAM for this IB.
 485
 486          * - If more than one half of VRAM is occupied, only allow relocating
 487          *   1 MB of data for this IB.
 488          *
 489          * - From 0 to one half of used VRAM, the threshold decreases
 490          *   linearly.
 491          *         __________________
 492          * 1/4 of -|\               |
 493          * VRAM    | \              |
 494          *         |  \             |
 495          *         |   \            |
 496          *         |    \           |
 497          *         |     \          |
 498          *         |      \         |
 499          *         |       \________|1 MB
 500          *         |----------------|
 501          *    VRAM 0 %             100 %
 502          *         used            used
 503          *
 504          * Note: It's a threshold, not a limit. The threshold must be crossed
 505          * for buffer relocations to stop, so any buffer of an arbitrary size
 506          * can be moved as long as the threshold isn't crossed before
 507          * the relocation takes place. We don't want to disable buffer
 508          * relocations completely.
 509          *
 510          * The idea is that buffers should be placed in VRAM at creation time
 511          * and TTM should only do a minimum number of relocations during
 512          * command submission. In practice, you need to submit at least
 513          * a dozen IBs to move all buffers to VRAM if they are in GTT.
 514          *
 515          * Also, things can get pretty crazy under memory pressure and actual
 516          * VRAM usage can change a lot, so playing safe even at 50% does
 517          * consistently increase performance.
 518          */
 519
 520         u64 half_vram = real_vram_size >> 1;
 521         u64 half_free_vram = vram_usage >= half_vram ? 0 : half_vram - vram_usage;
 522         u64 bytes_moved_threshold = half_free_vram >> 1;
 523         return max(bytes_moved_threshold, 1024*1024ull);
 524 }
 525
 526 int radeon_bo_list_validate(struct radeon_device *rdev,
 527                             struct ww_acquire_ctx *ticket,
 528                             struct list_head *head, int ring)
 529 {
 530         struct radeon_bo_list *lobj;
 531         struct list_head duplicates;
 532         int r;
 533         u64 bytes_moved = 0, initial_bytes_moved;
 534         u64 bytes_moved_threshold = radeon_bo_get_threshold_for_moves(rdev);
 535
 536         INIT_LIST_HEAD(&duplicates);
 537         r = ttm_eu_reserve_buffers(ticket, head, true, &duplicates);
 538         if (unlikely(r != 0)) {
 539                 return r;
 540         }
 541
 542         list_for_each_entry(lobj, head, tv.head) {
 543                 struct radeon_bo *bo = lobj->robj;
 544                 if (!bo->pin_count) {
 545                         u32 domain = lobj->prefered_domains;
 546                         u32 allowed = lobj->allowed_domains;
 547                         u32 current_domain =
 548                                 radeon_mem_type_to_domain(bo->tbo.mem.mem_type);
 549
 550                         /* Check if this buffer will be moved and don't move it
 551                          * if we have moved too many buffers for this IB already.
 552                          *
 553                          * Note that this allows moving at least one buffer of
 554                          * any size, because it doesn't take the current "bo"
 555                          * into account. We don't want to disallow buffer moves
 556                          * completely.
 557                          */
 558                         if ((allowed & current_domain) != 0 &&
 559                             (domain & current_domain) == 0 && /* will be moved */
 560                             bytes_moved > bytes_moved_threshold) {
 561                                 /* don't move it */
 562                                 domain = current_domain;
 563                         }
 564
 565                 retry:
 566                         radeon_ttm_placement_from_domain(bo, domain);
 567                         if (ring == R600_RING_TYPE_UVD_INDEX)
 568                                 radeon_uvd_force_into_uvd_segment(bo, allowed);
 569
 570                         initial_bytes_moved = atomic64_read(&rdev->num_bytes_moved);
 571                         r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
 572                         bytes_moved += atomic64_read(&rdev->num_bytes_moved) -
 573                                        initial_bytes_moved;
 574
 575                         if (unlikely(r)) {
 576                                 if (r != -ERESTARTSYS &&
 577                                     domain != lobj->allowed_domains) {
 578                                         domain = lobj->allowed_domains;
 579                                         goto retry;
 580                                 }
 581                                 ttm_eu_backoff_reservation(ticket, head);
 582                                 return r;
 583                         }
 584                 }
 585                 lobj->gpu_offset = radeon_bo_gpu_offset(bo);
 586                 lobj->tiling_flags = bo->tiling_flags;
 587         }
 588
 589         list_for_each_entry(lobj, &duplicates, tv.head) {
 590                 lobj->gpu_offset = radeon_bo_gpu_offset(lobj->robj);
 591                 lobj->tiling_flags = lobj->robj->tiling_flags;
 592         }
 593
 594         return 0;
 595 }
 596
 597 int radeon_bo_get_surface_reg(struct radeon_bo *bo)
 598 {
 599         struct radeon_device *rdev = bo->rdev;
 600         struct radeon_surface_reg *reg;
 601         struct radeon_bo *old_object;
 602         int steal;
 603         int i;
 604
 605         lockdep_assert_held(&bo->tbo.resv->lock.base);
 606
 607         if (!bo->tiling_flags)
 608                 return 0;
 609
 610         if (bo->surface_reg >= 0) {
 611                 reg = &rdev->surface_regs[bo->surface_reg];
 612                 i = bo->surface_reg;
 613                 goto out;
 614         }
 615
 616         steal = -1;
 617         for (i = 0; i < RADEON_GEM_MAX_SURFACES; i++) {
 618
 619                 reg = &rdev->surface_regs[i];
 620                 if (!reg->bo)
 621                         break;
 622
 623                 old_object = reg->bo;
 624                 if (old_object->pin_count == 0)
 625                         steal = i;
 626         }
 627
 628         /* if we are all out */
 629         if (i == RADEON_GEM_MAX_SURFACES) {
 630                 if (steal == -1)
 631                         return -ENOMEM;
 632                 /* find someone with a surface reg and nuke their BO */
 633                 reg = &rdev->surface_regs[steal];
 634                 old_object = reg->bo;
 635                 /* blow away the mapping */
 636                 DRM_DEBUG("stealing surface reg %d from %p\n", steal, old_object);
 637                 ttm_bo_unmap_virtual(&old_object->tbo);
 638                 old_object->surface_reg = -1;
 639                 i = steal;
 640         }
 641
 642         bo->surface_reg = i;
 643         reg->bo = bo;
 644
 645 out:
 646         radeon_set_surface_reg(rdev, i, bo->tiling_flags, bo->pitch,
 647                                bo->tbo.mem.start << PAGE_SHIFT,
 648                                bo->tbo.num_pages << PAGE_SHIFT);
 649         return 0;
 650 }
 651
 652 static void radeon_bo_clear_surface_reg(struct radeon_bo *bo)
 653 {
 654         struct radeon_device *rdev = bo->rdev;
 655         struct radeon_surface_reg *reg;
 656
 657         if (bo->surface_reg == -1)
 658                 return;
 659
 660         reg = &rdev->surface_regs[bo->surface_reg];
 661         radeon_clear_surface_reg(rdev, bo->surface_reg);
 662
 663         reg->bo = NULL;
 664         bo->surface_reg = -1;
 665 }
 666
 667 int radeon_bo_set_tiling_flags(struct radeon_bo *bo,
 668                                 uint32_t tiling_flags, uint32_t pitch)
 669 {
 670         struct radeon_device *rdev = bo->rdev;
 671         int r;
 672
 673         if (rdev->family >= CHIP_CEDAR) {
 674                 unsigned bankw, bankh, mtaspect, tilesplit, stilesplit;
 675
 676                 bankw = (tiling_flags >> RADEON_TILING_EG_BANKW_SHIFT) & RADEON_TILING_EG_BANKW_MASK;
 677                 bankh = (tiling_flags >> RADEON_TILING_EG_BANKH_SHIFT) & RADEON_TILING_EG_BANKH_MASK;
 678                 mtaspect = (tiling_flags >> RADEON_TILING_EG_MACRO_TILE_ASPECT_SHIFT) & RADEON_TILING_EG_MACRO_TILE_ASPECT_MASK;
 679                 tilesplit = (tiling_flags >> RADEON_TILING_EG_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_TILE_SPLIT_MASK;
 680                 stilesplit = (tiling_flags >> RADEON_TILING_EG_STENCIL_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_STENCIL_TILE_SPLIT_MASK;
 681                 switch (bankw) {
 682                 case 0:
 683                 case 1:
 684                 case 2:
 685                 case 4:
 686                 case 8:
 687                         break;
 688                 default:
 689                         return -EINVAL;
 690                 }
 691                 switch (bankh) {
 692                 case 0:
 693                 case 1:
 694                 case 2:
 695                 case 4:
 696                 case 8:
 697                         break;
 698                 default:
 699                         return -EINVAL;
 700                 }
 701                 switch (mtaspect) {
 702                 case 0:
 703                 case 1:
 704                 case 2:
 705                 case 4:
 706                 case 8:
 707                         break;
 708                 default:
 709                         return -EINVAL;
 710                 }
 711                 if (tilesplit > 6) {
 712                         return -EINVAL;
 713                 }
 714                 if (stilesplit > 6) {
 715                         return -EINVAL;
 716                 }
 717         }
 718         r = radeon_bo_reserve(bo, false);
 719         if (unlikely(r != 0))
 720                 return r;
 721         bo->tiling_flags = tiling_flags;
 722         bo->pitch = pitch;
 723         radeon_bo_unreserve(bo);
 724         return 0;
 725 }
 726
 727 void radeon_bo_get_tiling_flags(struct radeon_bo *bo,
 728                                 uint32_t *tiling_flags,
 729                                 uint32_t *pitch)
 730 {
 731         lockdep_assert_held(&bo->tbo.resv->lock.base);
 732
 733         if (tiling_flags)
 734                 *tiling_flags = bo->tiling_flags;
 735         if (pitch)
 736                 *pitch = bo->pitch;
 737 }
 738
 739 int radeon_bo_check_tiling(struct radeon_bo *bo, bool has_moved,
 740                                 bool force_drop)
 741 {
 742         if (!force_drop)
 743                 lockdep_assert_held(&bo->tbo.resv->lock.base);
 744
 745         if (!(bo->tiling_flags & RADEON_TILING_SURFACE))
 746                 return 0;
 747
 748         if (force_drop) {
 749                 radeon_bo_clear_surface_reg(bo);
 750                 return 0;
 751         }
 752
 753         if (bo->tbo.mem.mem_type != TTM_PL_VRAM) {
 754                 if (!has_moved)
 755                         return 0;
 756
 757                 if (bo->surface_reg >= 0)
 758                         radeon_bo_clear_surface_reg(bo);
 759                 return 0;
 760         }
 761
 762         if ((bo->surface_reg >= 0) && !has_moved)
 763                 return 0;
 764
 765         return radeon_bo_get_surface_reg(bo);
 766 }
 767
 768 void radeon_bo_move_notify(struct ttm_buffer_object *bo,
 769                            struct ttm_mem_reg *new_mem)
 770 {
 771         struct radeon_bo *rbo;
 772
 773         if (!radeon_ttm_bo_is_radeon_bo(bo))
 774                 return;
 775
 776         rbo = container_of(bo, struct radeon_bo, tbo);
 777         radeon_bo_check_tiling(rbo, 0, 1);
 778         radeon_vm_bo_invalidate(rbo->rdev, rbo);
 779
 780         /* update statistics */
 781         if (!new_mem)
 782                 return;
 783
 784         radeon_update_memory_usage(rbo, bo->mem.mem_type, -1);
 785         radeon_update_memory_usage(rbo, new_mem->mem_type, 1);
 786 }
 787
 788 int radeon_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
 789 {
 790         struct radeon_device *rdev;
 791         struct radeon_bo *rbo;
 792         unsigned long offset, size, lpfn;
 793         int i, r;
 794
 795         if (!radeon_ttm_bo_is_radeon_bo(bo))
 796                 return 0;
 797         rbo = container_of(bo, struct radeon_bo, tbo);
 798         radeon_bo_check_tiling(rbo, 0, 0);
 799         rdev = rbo->rdev;
 800         if (bo->mem.mem_type != TTM_PL_VRAM)
 801                 return 0;
 802
 803         size = bo->mem.num_pages << PAGE_SHIFT;
 804         offset = bo->mem.start << PAGE_SHIFT;
 805         if ((offset + size) <= rdev->mc.visible_vram_size)
 806                 return 0;
 807
 808         /* Can't move a pinned BO to visible VRAM */
 809         if (rbo->pin_count > 0)
 810                 return -EINVAL;
 811
 812         /* hurrah the memory is not visible ! */
 813         radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_VRAM);
 814         lpfn =  rdev->mc.visible_vram_size >> PAGE_SHIFT;
 815         for (i = 0; i < rbo->placement.num_placement; i++) {
 816                 /* Force into visible VRAM */
 817                 if ((rbo->placements[i].flags & TTM_PL_FLAG_VRAM) &&
 818                     (!rbo->placements[i].lpfn || rbo->placements[i].lpfn > lpfn))
 819                         rbo->placements[i].lpfn = lpfn;
 820         }
 821         r = ttm_bo_validate(bo, &rbo->placement, false, false);
 822         if (unlikely(r == -ENOMEM)) {
 823                 radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_GTT);
 824                 return ttm_bo_validate(bo, &rbo->placement, false, false);
 825         } else if (unlikely(r != 0)) {
 826                 return r;
 827         }
 828
 829         offset = bo->mem.start << PAGE_SHIFT;
 830         /* this should never happen */
 831         if ((offset + size) > rdev->mc.visible_vram_size)
 832                 return -EINVAL;
 833
 834         return 0;
 835 }
 836
 837 int radeon_bo_wait(struct radeon_bo *bo, u32 *mem_type, bool no_wait)
 838 {
 839         int r;
 840
 841         r = ttm_bo_reserve(&bo->tbo, true, no_wait, NULL);
 842         if (unlikely(r != 0))
 843                 return r;
 844         if (mem_type)
 845                 *mem_type = bo->tbo.mem.mem_type;
 846
 847         r = ttm_bo_wait(&bo->tbo, true, no_wait);
 848         ttm_bo_unreserve(&bo->tbo);
 849         return r;
 850 }
 851
 852 /**
 853  * radeon_bo_fence - add fence to buffer object
 854  *
 855  * @bo: buffer object in question
 856  * @fence: fence to add
 857  * @shared: true if fence should be added shared
 858  *
 859  */
 860 void radeon_bo_fence(struct radeon_bo *bo, struct radeon_fence *fence,
 861                      bool shared)
 862 {
 863         struct reservation_object *resv = bo->tbo.resv;
 864
 865         if (shared)
 866                 reservation_object_add_shared_fence(resv, &fence->base);
 867         else
 868                 reservation_object_add_excl_fence(resv, &fence->base);
 869 }