drivers/gpu/drm/ttm/ttm_bo.c

   1 /**************************************************************************
   2  *
   3  * Copyright (c) 2006-2009 VMware, Inc., Palo Alto, CA., USA
   4  * All Rights Reserved.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the
   8  * "Software"), to deal in the Software without restriction, including
   9  * without limitation the rights to use, copy, modify, merge, publish,
  10  * distribute, sub license, and/or sell copies of the Software, and to
  11  * permit persons to whom the Software is furnished to do so, subject to
  12  * the following conditions:
  13  *
  14  * The above copyright notice and this permission notice (including the
  15  * next paragraph) shall be included in all copies or substantial portions
  16  * of the Software.
  17  *
  18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  19  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  20  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  21  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
  22  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  23  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  24  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  25  *
  26  **************************************************************************/
  27 /*
  28  * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com>
  29  */
  30
  31 #define pr_fmt(fmt) "[TTM] " fmt
  32
  33 #include <drm/ttm/ttm_module.h>
  34 #include <drm/ttm/ttm_bo_driver.h>
  35 #include <drm/ttm/ttm_placement.h>
  36 #include <linux/jiffies.h>
  37 #include <linux/slab.h>
  38 #include <linux/sched.h>
  39 #include <linux/mm.h>
  40 #include <linux/file.h>
  41 #include <linux/module.h>
  42 #include <linux/atomic.h>
  43
  44 #define TTM_ASSERT_LOCKED(param)
  45 #define TTM_DEBUG(fmt, arg...)
  46 #define TTM_BO_HASH_ORDER 13
  47
  48 static int ttm_bo_setup_vm(struct ttm_buffer_object *bo);
  49 static int ttm_bo_swapout(struct ttm_mem_shrink *shrink);
  50 static void ttm_bo_global_kobj_release(struct kobject *kobj);
  51
  52 static struct attribute ttm_bo_count = {
  53         .name = "bo_count",
  54         .mode = S_IRUGO
  55 };
  56
  57 static inline int ttm_mem_type_from_flags(uint32_t flags, uint32_t *mem_type)
  58 {
  59         int i;
  60
  61         for (i = 0; i <= TTM_PL_PRIV5; i++)
  62                 if (flags & (1 << i)) {
  63                         *mem_type = i;
  64                         return 0;
  65                 }
  66         return -EINVAL;
  67 }
  68
  69 static void ttm_mem_type_debug(struct ttm_bo_device *bdev, int mem_type)
  70 {
  71         struct ttm_mem_type_manager *man = &bdev->man[mem_type];
  72
  73         pr_err("    has_type: %d\n", man->has_type);
  74         pr_err("    use_type: %d\n", man->use_type);
  75         pr_err("    flags: 0x%08X\n", man->flags);
  76         pr_err("    gpu_offset: 0x%08lX\n", man->gpu_offset);
  77         pr_err("    size: %llu\n", man->size);
  78         pr_err("    available_caching: 0x%08X\n", man->available_caching);
  79         pr_err("    default_caching: 0x%08X\n", man->default_caching);
  80         if (mem_type != TTM_PL_SYSTEM)
  81                 (*man->func->debug)(man, TTM_PFX);
  82 }
  83
  84 static void ttm_bo_mem_space_debug(struct ttm_buffer_object *bo,
  85                                         struct ttm_placement *placement)
  86 {
  87         int i, ret, mem_type;
  88
  89         pr_err("No space for %p (%lu pages, %luK, %luM)\n",
  90                bo, bo->mem.num_pages, bo->mem.size >> 10,
  91                bo->mem.size >> 20);
  92         for (i = 0; i < placement->num_placement; i++) {
  93                 ret = ttm_mem_type_from_flags(placement->placement[i],
  94                                                 &mem_type);
  95                 if (ret)
  96                         return;
  97                 pr_err("  placement[%d]=0x%08X (%d)\n",
  98                        i, placement->placement[i], mem_type);
  99                 ttm_mem_type_debug(bo->bdev, mem_type);
 100         }
 101 }
 102
 103 static ssize_t ttm_bo_global_show(struct kobject *kobj,
 104                                   struct attribute *attr,
 105                                   char *buffer)
 106 {
 107         struct ttm_bo_global *glob =
 108                 container_of(kobj, struct ttm_bo_global, kobj);
 109
 110         return snprintf(buffer, PAGE_SIZE, "%lu\n",
 111                         (unsigned long) atomic_read(&glob->bo_count));
 112 }
 113
 114 static struct attribute *ttm_bo_global_attrs[] = {
 115         &ttm_bo_count,
 116         NULL
 117 };
 118
 119 static const struct sysfs_ops ttm_bo_global_ops = {
 120         .show = &ttm_bo_global_show
 121 };
 122
 123 static struct kobj_type ttm_bo_glob_kobj_type  = {
 124         .release = &ttm_bo_global_kobj_release,
 125         .sysfs_ops = &ttm_bo_global_ops,
 126         .default_attrs = ttm_bo_global_attrs
 127 };
 128
 129
 130 static inline uint32_t ttm_bo_type_flags(unsigned type)
 131 {
 132         return 1 << (type);
 133 }
 134
 135 static void ttm_bo_release_list(struct kref *list_kref)
 136 {
 137         struct ttm_buffer_object *bo =
 138             container_of(list_kref, struct ttm_buffer_object, list_kref);
 139         struct ttm_bo_device *bdev = bo->bdev;
 140         size_t acc_size = bo->acc_size;
 141
 142         BUG_ON(atomic_read(&bo->list_kref.refcount));
 143         BUG_ON(atomic_read(&bo->kref.refcount));
 144         BUG_ON(atomic_read(&bo->cpu_writers));
 145         BUG_ON(bo->sync_obj != NULL);
 146         BUG_ON(bo->mem.mm_node != NULL);
 147         BUG_ON(!list_empty(&bo->lru));
 148         BUG_ON(!list_empty(&bo->ddestroy));
 149
 150         if (bo->ttm)
 151                 ttm_tt_destroy(bo->ttm);
 152         atomic_dec(&bo->glob->bo_count);
 153         if (bo->destroy)
 154                 bo->destroy(bo);
 155         else {
 156                 kfree(bo);
 157         }
 158         ttm_mem_global_free(bdev->glob->mem_glob, acc_size);
 159 }
 160
 161 int ttm_bo_wait_unreserved(struct ttm_buffer_object *bo, bool interruptible)
 162 {
 163         if (interruptible) {
 164                 return wait_event_interruptible(bo->event_queue,
 165                                                !ttm_bo_is_reserved(bo));
 166         } else {
 167                 wait_event(bo->event_queue, !ttm_bo_is_reserved(bo));
 168                 return 0;
 169         }
 170 }
 171 EXPORT_SYMBOL(ttm_bo_wait_unreserved);
 172
 173 void ttm_bo_add_to_lru(struct ttm_buffer_object *bo)
 174 {
 175         struct ttm_bo_device *bdev = bo->bdev;
 176         struct ttm_mem_type_manager *man;
 177
 178         BUG_ON(!ttm_bo_is_reserved(bo));
 179
 180         if (!(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
 181
 182                 BUG_ON(!list_empty(&bo->lru));
 183
 184                 man = &bdev->man[bo->mem.mem_type];
 185                 list_add_tail(&bo->lru, &man->lru);
 186                 kref_get(&bo->list_kref);
 187
 188                 if (bo->ttm != NULL) {
 189                         list_add_tail(&bo->swap, &bo->glob->swap_lru);
 190                         kref_get(&bo->list_kref);
 191                 }
 192         }
 193 }
 194
 195 int ttm_bo_del_from_lru(struct ttm_buffer_object *bo)
 196 {
 197         int put_count = 0;
 198
 199         if (!list_empty(&bo->swap)) {
 200                 list_del_init(&bo->swap);
 201                 ++put_count;
 202         }
 203         if (!list_empty(&bo->lru)) {
 204                 list_del_init(&bo->lru);
 205                 ++put_count;
 206         }
 207
 208         /*
 209          * TODO: Add a driver hook to delete from
 210          * driver-specific LRU's here.
 211          */
 212
 213         return put_count;
 214 }
 215
 216 int ttm_bo_reserve_locked(struct ttm_buffer_object *bo,
 217                           bool interruptible,
 218                           bool no_wait, bool use_sequence, uint32_t sequence)
 219 {
 220         struct ttm_bo_global *glob = bo->glob;
 221         int ret;
 222
 223         while (unlikely(atomic_read(&bo->reserved) != 0)) {
 224                 /**
 225                  * Deadlock avoidance for multi-bo reserving.
 226                  */
 227                 if (use_sequence && bo->seq_valid) {
 228                         /**
 229                          * We've already reserved this one.
 230                          */
 231                         if (unlikely(sequence == bo->val_seq))
 232                                 return -EDEADLK;
 233                         /**
 234                          * Already reserved by a thread that will not back
 235                          * off for us. We need to back off.
 236                          */
 237                         if (unlikely(sequence - bo->val_seq < (1 << 31)))
 238                                 return -EAGAIN;
 239                 }
 240
 241                 if (no_wait)
 242                         return -EBUSY;
 243
 244                 spin_unlock(&glob->lru_lock);
 245                 ret = ttm_bo_wait_unreserved(bo, interruptible);
 246                 spin_lock(&glob->lru_lock);
 247
 248                 if (unlikely(ret))
 249                         return ret;
 250         }
 251
 252         atomic_set(&bo->reserved, 1);
 253         if (use_sequence) {
 254                 /**
 255                  * Wake up waiters that may need to recheck for deadlock,
 256                  * if we decreased the sequence number.
 257                  */
 258                 if (unlikely((bo->val_seq - sequence < (1 << 31))
 259                              || !bo->seq_valid))
 260                         wake_up_all(&bo->event_queue);
 261
 262                 bo->val_seq = sequence;
 263                 bo->seq_valid = true;
 264         } else {
 265                 bo->seq_valid = false;
 266         }
 267
 268         return 0;
 269 }
 270 EXPORT_SYMBOL(ttm_bo_reserve);
 271
 272 static void ttm_bo_ref_bug(struct kref *list_kref)
 273 {
 274         BUG();
 275 }
 276
 277 void ttm_bo_list_ref_sub(struct ttm_buffer_object *bo, int count,
 278                          bool never_free)
 279 {
 280         kref_sub(&bo->list_kref, count,
 281                  (never_free) ? ttm_bo_ref_bug : ttm_bo_release_list);
 282 }
 283
 284 int ttm_bo_reserve(struct ttm_buffer_object *bo,
 285                    bool interruptible,
 286                    bool no_wait, bool use_sequence, uint32_t sequence)
 287 {
 288         struct ttm_bo_global *glob = bo->glob;
 289         int put_count = 0;
 290         int ret;
 291
 292         spin_lock(&glob->lru_lock);
 293         ret = ttm_bo_reserve_locked(bo, interruptible, no_wait, use_sequence,
 294                                     sequence);
 295         if (likely(ret == 0))
 296                 put_count = ttm_bo_del_from_lru(bo);
 297         spin_unlock(&glob->lru_lock);
 298
 299         ttm_bo_list_ref_sub(bo, put_count, true);
 300
 301         return ret;
 302 }
 303
 304 void ttm_bo_unreserve_locked(struct ttm_buffer_object *bo)
 305 {
 306         ttm_bo_add_to_lru(bo);
 307         atomic_set(&bo->reserved, 0);
 308         wake_up_all(&bo->event_queue);
 309 }
 310
 311 void ttm_bo_unreserve(struct ttm_buffer_object *bo)
 312 {
 313         struct ttm_bo_global *glob = bo->glob;
 314
 315         spin_lock(&glob->lru_lock);
 316         ttm_bo_unreserve_locked(bo);
 317         spin_unlock(&glob->lru_lock);
 318 }
 319 EXPORT_SYMBOL(ttm_bo_unreserve);
 320
 321 /*
 322  * Call bo->mutex locked.
 323  */
 324 static int ttm_bo_add_ttm(struct ttm_buffer_object *bo, bool zero_alloc)
 325 {
 326         struct ttm_bo_device *bdev = bo->bdev;
 327         struct ttm_bo_global *glob = bo->glob;
 328         int ret = 0;
 329         uint32_t page_flags = 0;
 330
 331         TTM_ASSERT_LOCKED(&bo->mutex);
 332         bo->ttm = NULL;
 333
 334         if (bdev->need_dma32)
 335                 page_flags |= TTM_PAGE_FLAG_DMA32;
 336
 337         switch (bo->type) {
 338         case ttm_bo_type_device:
 339                 if (zero_alloc)
 340                         page_flags |= TTM_PAGE_FLAG_ZERO_ALLOC;
 341         case ttm_bo_type_kernel:
 342                 bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
 343                                                       page_flags, glob->dummy_read_page);
 344                 if (unlikely(bo->ttm == NULL))
 345                         ret = -ENOMEM;
 346                 break;
 347         case ttm_bo_type_sg:
 348                 bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
 349                                                       page_flags | TTM_PAGE_FLAG_SG,
 350                                                       glob->dummy_read_page);
 351                 if (unlikely(bo->ttm == NULL)) {
 352                         ret = -ENOMEM;
 353                         break;
 354                 }
 355                 bo->ttm->sg = bo->sg;
 356                 break;
 357         default:
 358                 pr_err("Illegal buffer object type\n");
 359                 ret = -EINVAL;
 360                 break;
 361         }
 362
 363         return ret;
 364 }
 365
 366 static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
 367                                   struct ttm_mem_reg *mem,
 368                                   bool evict, bool interruptible,
 369                                   bool no_wait_gpu)
 370 {
 371         struct ttm_bo_device *bdev = bo->bdev;
 372         bool old_is_pci = ttm_mem_reg_is_pci(bdev, &bo->mem);
 373         bool new_is_pci = ttm_mem_reg_is_pci(bdev, mem);
 374         struct ttm_mem_type_manager *old_man = &bdev->man[bo->mem.mem_type];
 375         struct ttm_mem_type_manager *new_man = &bdev->man[mem->mem_type];
 376         int ret = 0;
 377
 378         if (old_is_pci || new_is_pci ||
 379             ((mem->placement & bo->mem.placement & TTM_PL_MASK_CACHING) == 0)) {
 380                 ret = ttm_mem_io_lock(old_man, true);
 381                 if (unlikely(ret != 0))
 382                         goto out_err;
 383                 ttm_bo_unmap_virtual_locked(bo);
 384                 ttm_mem_io_unlock(old_man);
 385         }
 386
 387         /*
 388          * Create and bind a ttm if required.
 389          */
 390
 391         if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
 392                 if (bo->ttm == NULL) {
 393                         bool zero = !(old_man->flags & TTM_MEMTYPE_FLAG_FIXED);
 394                         ret = ttm_bo_add_ttm(bo, zero);
 395                         if (ret)
 396                                 goto out_err;
 397                 }
 398
 399                 ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement);
 400                 if (ret)
 401                         goto out_err;
 402
 403                 if (mem->mem_type != TTM_PL_SYSTEM) {
 404                         ret = ttm_tt_bind(bo->ttm, mem);
 405                         if (ret)
 406                                 goto out_err;
 407                 }
 408
 409                 if (bo->mem.mem_type == TTM_PL_SYSTEM) {
 410                         if (bdev->driver->move_notify)
 411                                 bdev->driver->move_notify(bo, mem);
 412                         bo->mem = *mem;
 413                         mem->mm_node = NULL;
 414                         goto moved;
 415                 }
 416         }
 417
 418         if (bdev->driver->move_notify)
 419                 bdev->driver->move_notify(bo, mem);
 420
 421         if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) &&
 422             !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED))
 423                 ret = ttm_bo_move_ttm(bo, evict, no_wait_gpu, mem);
 424         else if (bdev->driver->move)
 425                 ret = bdev->driver->move(bo, evict, interruptible,
 426                                          no_wait_gpu, mem);
 427         else
 428                 ret = ttm_bo_move_memcpy(bo, evict, no_wait_gpu, mem);
 429
 430         if (ret) {
 431                 if (bdev->driver->move_notify) {
 432                         struct ttm_mem_reg tmp_mem = *mem;
 433                         *mem = bo->mem;
 434                         bo->mem = tmp_mem;
 435                         bdev->driver->move_notify(bo, mem);
 436                         bo->mem = *mem;
 437                 }
 438
 439                 goto out_err;
 440         }
 441
 442 moved:
 443         if (bo->evicted) {
 444                 ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement);
 445                 if (ret)
 446                         pr_err("Can not flush read caches\n");
 447                 bo->evicted = false;
 448         }
 449
 450         if (bo->mem.mm_node) {
 451                 bo->offset = (bo->mem.start << PAGE_SHIFT) +
 452                     bdev->man[bo->mem.mem_type].gpu_offset;
 453                 bo->cur_placement = bo->mem.placement;
 454         } else
 455                 bo->offset = 0;
 456
 457         return 0;
 458
 459 out_err:
 460         new_man = &bdev->man[bo->mem.mem_type];
 461         if ((new_man->flags & TTM_MEMTYPE_FLAG_FIXED) && bo->ttm) {
 462                 ttm_tt_unbind(bo->ttm);
 463                 ttm_tt_destroy(bo->ttm);
 464                 bo->ttm = NULL;
 465         }
 466
 467         return ret;
 468 }
 469
 470 /**
 471  * Call bo::reserved.
 472  * Will release GPU memory type usage on destruction.
 473  * This is the place to put in driver specific hooks to release
 474  * driver private resources.
 475  * Will release the bo::reserved lock.
 476  */
 477
 478 static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
 479 {
 480         if (bo->bdev->driver->move_notify)
 481                 bo->bdev->driver->move_notify(bo, NULL);
 482
 483         if (bo->ttm) {
 484                 ttm_tt_unbind(bo->ttm);
 485                 ttm_tt_destroy(bo->ttm);
 486                 bo->ttm = NULL;
 487         }
 488         ttm_bo_mem_put(bo, &bo->mem);
 489
 490         atomic_set(&bo->reserved, 0);
 491         wake_up_all(&bo->event_queue);
 492
 493         /*
 494          * Since the final reference to this bo may not be dropped by
 495          * the current task we have to put a memory barrier here to make
 496          * sure the changes done in this function are always visible.
 497          *
 498          * This function only needs protection against the final kref_put.
 499          */
 500         smp_mb__before_atomic_dec();
 501 }
 502
 503 static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo)
 504 {
 505         struct ttm_bo_device *bdev = bo->bdev;
 506         struct ttm_bo_global *glob = bo->glob;
 507         struct ttm_bo_driver *driver = bdev->driver;
 508         void *sync_obj = NULL;
 509         int put_count;
 510         int ret;
 511
 512         spin_lock(&glob->lru_lock);
 513         ret = ttm_bo_reserve_locked(bo, false, true, false, 0);
 514
 515         spin_lock(&bdev->fence_lock);
 516         (void) ttm_bo_wait(bo, false, false, true);
 517         if (!ret && !bo->sync_obj) {
 518                 spin_unlock(&bdev->fence_lock);
 519                 put_count = ttm_bo_del_from_lru(bo);
 520
 521                 spin_unlock(&glob->lru_lock);
 522                 ttm_bo_cleanup_memtype_use(bo);
 523
 524                 ttm_bo_list_ref_sub(bo, put_count, true);
 525
 526                 return;
 527         }
 528         if (bo->sync_obj)
 529                 sync_obj = driver->sync_obj_ref(bo->sync_obj);
 530         spin_unlock(&bdev->fence_lock);
 531
 532         if (!ret) {
 533                 atomic_set(&bo->reserved, 0);
 534                 wake_up_all(&bo->event_queue);
 535         }
 536
 537         kref_get(&bo->list_kref);
 538         list_add_tail(&bo->ddestroy, &bdev->ddestroy);
 539         spin_unlock(&glob->lru_lock);
 540
 541         if (sync_obj) {
 542                 driver->sync_obj_flush(sync_obj);
 543                 driver->sync_obj_unref(&sync_obj);
 544         }
 545         schedule_delayed_work(&bdev->wq,
 546                               ((HZ / 100) < 1) ? 1 : HZ / 100);
 547 }
 548
 549 /**
 550  * function ttm_bo_cleanup_refs_and_unlock
 551  * If bo idle, remove from delayed- and lru lists, and unref.
 552  * If not idle, do nothing.
 553  *
 554  * Must be called with lru_lock and reservation held, this function
 555  * will drop both before returning.
 556  *
 557  * @interruptible         Any sleeps should occur interruptibly.
 558  * @no_wait_gpu           Never wait for gpu. Return -EBUSY instead.
 559  */
 560
 561 static int ttm_bo_cleanup_refs_and_unlock(struct ttm_buffer_object *bo,
 562                                           bool interruptible,
 563                                           bool no_wait_gpu)
 564 {
 565         struct ttm_bo_device *bdev = bo->bdev;
 566         struct ttm_bo_driver *driver = bdev->driver;
 567         struct ttm_bo_global *glob = bo->glob;
 568         int put_count;
 569         int ret;
 570
 571         spin_lock(&bdev->fence_lock);
 572         ret = ttm_bo_wait(bo, false, false, true);
 573
 574         if (ret && !no_wait_gpu) {
 575                 void *sync_obj;
 576
 577                 /*
 578                  * Take a reference to the fence and unreserve,
 579                  * at this point the buffer should be dead, so
 580                  * no new sync objects can be attached.
 581                  */
 582                 sync_obj = driver->sync_obj_ref(&bo->sync_obj);
 583                 spin_unlock(&bdev->fence_lock);
 584
 585                 atomic_set(&bo->reserved, 0);
 586                 wake_up_all(&bo->event_queue);
 587                 spin_unlock(&glob->lru_lock);
 588
 589                 ret = driver->sync_obj_wait(sync_obj, false, interruptible);
 590                 driver->sync_obj_unref(&sync_obj);
 591                 if (ret)
 592                         return ret;
 593
 594                 /*
 595                  * remove sync_obj with ttm_bo_wait, the wait should be
 596                  * finished, and no new wait object should have been added.
 597                  */
 598                 spin_lock(&bdev->fence_lock);
 599                 ret = ttm_bo_wait(bo, false, false, true);
 600                 WARN_ON(ret);
 601                 spin_unlock(&bdev->fence_lock);
 602                 if (ret)
 603                         return ret;
 604
 605                 spin_lock(&glob->lru_lock);
 606                 ret = ttm_bo_reserve_locked(bo, false, true, false, 0);
 607
 608                 /*
 609                  * We raced, and lost, someone else holds the reservation now,
 610                  * and is probably busy in ttm_bo_cleanup_memtype_use.
 611                  *
 612                  * Even if it's not the case, because we finished waiting any
 613                  * delayed destruction would succeed, so just return success
 614                  * here.
 615                  */
 616                 if (ret) {
 617                         spin_unlock(&glob->lru_lock);
 618                         return 0;
 619                 }
 620         } else
 621                 spin_unlock(&bdev->fence_lock);
 622
 623         if (ret || unlikely(list_empty(&bo->ddestroy))) {
 624                 atomic_set(&bo->reserved, 0);
 625                 wake_up_all(&bo->event_queue);
 626                 spin_unlock(&glob->lru_lock);
 627                 return ret;
 628         }
 629
 630         put_count = ttm_bo_del_from_lru(bo);
 631         list_del_init(&bo->ddestroy);
 632         ++put_count;
 633
 634         spin_unlock(&glob->lru_lock);
 635         ttm_bo_cleanup_memtype_use(bo);
 636
 637         ttm_bo_list_ref_sub(bo, put_count, true);
 638
 639         return 0;
 640 }
 641
 642 /**
 643  * Traverse the delayed list, and call ttm_bo_cleanup_refs on all
 644  * encountered buffers.
 645  */
 646
 647 static int ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all)
 648 {
 649         struct ttm_bo_global *glob = bdev->glob;
 650         struct ttm_buffer_object *entry = NULL;
 651         int ret = 0;
 652
 653         spin_lock(&glob->lru_lock);
 654         if (list_empty(&bdev->ddestroy))
 655                 goto out_unlock;
 656
 657         entry = list_first_entry(&bdev->ddestroy,
 658                 struct ttm_buffer_object, ddestroy);
 659         kref_get(&entry->list_kref);
 660
 661         for (;;) {
 662                 struct ttm_buffer_object *nentry = NULL;
 663
 664                 if (entry->ddestroy.next != &bdev->ddestroy) {
 665                         nentry = list_first_entry(&entry->ddestroy,
 666                                 struct ttm_buffer_object, ddestroy);
 667                         kref_get(&nentry->list_kref);
 668                 }
 669
 670                 ret = ttm_bo_reserve_locked(entry, false, !remove_all, false, 0);
 671                 if (!ret)
 672                         ret = ttm_bo_cleanup_refs_and_unlock(entry, false,
 673                                                              !remove_all);
 674                 else
 675                         spin_unlock(&glob->lru_lock);
 676
 677                 kref_put(&entry->list_kref, ttm_bo_release_list);
 678                 entry = nentry;
 679
 680                 if (ret || !entry)
 681                         goto out;
 682
 683                 spin_lock(&glob->lru_lock);
 684                 if (list_empty(&entry->ddestroy))
 685                         break;
 686         }
 687
 688 out_unlock:
 689         spin_unlock(&glob->lru_lock);
 690 out:
 691         if (entry)
 692                 kref_put(&entry->list_kref, ttm_bo_release_list);
 693         return ret;
 694 }
 695
 696 static void ttm_bo_delayed_workqueue(struct work_struct *work)
 697 {
 698         struct ttm_bo_device *bdev =
 699             container_of(work, struct ttm_bo_device, wq.work);
 700
 701         if (ttm_bo_delayed_delete(bdev, false)) {
 702                 schedule_delayed_work(&bdev->wq,
 703                                       ((HZ / 100) < 1) ? 1 : HZ / 100);
 704         }
 705 }
 706
 707 static void ttm_bo_release(struct kref *kref)
 708 {
 709         struct ttm_buffer_object *bo =
 710             container_of(kref, struct ttm_buffer_object, kref);
 711         struct ttm_bo_device *bdev = bo->bdev;
 712         struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
 713
 714         write_lock(&bdev->vm_lock);
 715         if (likely(bo->vm_node != NULL)) {
 716                 rb_erase(&bo->vm_rb, &bdev->addr_space_rb);
 717                 drm_mm_put_block(bo->vm_node);
 718                 bo->vm_node = NULL;
 719         }
 720         write_unlock(&bdev->vm_lock);
 721         ttm_mem_io_lock(man, false);
 722         ttm_mem_io_free_vm(bo);
 723         ttm_mem_io_unlock(man);
 724         ttm_bo_cleanup_refs_or_queue(bo);
 725         kref_put(&bo->list_kref, ttm_bo_release_list);
 726 }
 727
 728 void ttm_bo_unref(struct ttm_buffer_object **p_bo)
 729 {
 730         struct ttm_buffer_object *bo = *p_bo;
 731
 732         *p_bo = NULL;
 733         kref_put(&bo->kref, ttm_bo_release);
 734 }
 735 EXPORT_SYMBOL(ttm_bo_unref);
 736
 737 int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev)
 738 {
 739         return cancel_delayed_work_sync(&bdev->wq);
 740 }
 741 EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue);
 742
 743 void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched)
 744 {
 745         if (resched)
 746                 schedule_delayed_work(&bdev->wq,
 747                                       ((HZ / 100) < 1) ? 1 : HZ / 100);
 748 }
 749 EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue);
 750
 751 static int ttm_bo_evict(struct ttm_buffer_object *bo, bool interruptible,
 752                         bool no_wait_gpu)
 753 {
 754         struct ttm_bo_device *bdev = bo->bdev;
 755         struct ttm_mem_reg evict_mem;
 756         struct ttm_placement placement;
 757         int ret = 0;
 758
 759         spin_lock(&bdev->fence_lock);
 760         ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
 761         spin_unlock(&bdev->fence_lock);
 762
 763         if (unlikely(ret != 0)) {
 764                 if (ret != -ERESTARTSYS) {
 765                         pr_err("Failed to expire sync object before buffer eviction\n");
 766                 }
 767                 goto out;
 768         }
 769
 770         BUG_ON(!ttm_bo_is_reserved(bo));
 771
 772         evict_mem = bo->mem;
 773         evict_mem.mm_node = NULL;
 774         evict_mem.bus.io_reserved_vm = false;
 775         evict_mem.bus.io_reserved_count = 0;
 776
 777         placement.fpfn = 0;
 778         placement.lpfn = 0;
 779         placement.num_placement = 0;
 780         placement.num_busy_placement = 0;
 781         bdev->driver->evict_flags(bo, &placement);
 782         ret = ttm_bo_mem_space(bo, &placement, &evict_mem, interruptible,
 783                                 no_wait_gpu);
 784         if (ret) {
 785                 if (ret != -ERESTARTSYS) {
 786                         pr_err("Failed to find memory space for buffer 0x%p eviction\n",
 787                                bo);
 788                         ttm_bo_mem_space_debug(bo, &placement);
 789                 }
 790                 goto out;
 791         }
 792
 793         ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, interruptible,
 794                                      no_wait_gpu);
 795         if (ret) {
 796                 if (ret != -ERESTARTSYS)
 797                         pr_err("Buffer eviction failed\n");
 798                 ttm_bo_mem_put(bo, &evict_mem);
 799                 goto out;
 800         }
 801         bo->evicted = true;
 802 out:
 803         return ret;
 804 }
 805
 806 static int ttm_mem_evict_first(struct ttm_bo_device *bdev,
 807                                 uint32_t mem_type,
 808                                 bool interruptible,
 809                                 bool no_wait_gpu)
 810 {
 811         struct ttm_bo_global *glob = bdev->glob;
 812         struct ttm_mem_type_manager *man = &bdev->man[mem_type];
 813         struct ttm_buffer_object *bo;
 814         int ret = -EBUSY, put_count;
 815
 816         spin_lock(&glob->lru_lock);
 817         list_for_each_entry(bo, &man->lru, lru) {
 818                 ret = ttm_bo_reserve_locked(bo, false, true, false, 0);
 819                 if (!ret)
 820                         break;
 821         }
 822
 823         if (ret) {
 824                 spin_unlock(&glob->lru_lock);
 825                 return ret;
 826         }
 827
 828         kref_get(&bo->list_kref);
 829
 830         if (!list_empty(&bo->ddestroy)) {
 831                 ret = ttm_bo_cleanup_refs_and_unlock(bo, interruptible,
 832                                                      no_wait_gpu);
 833                 kref_put(&bo->list_kref, ttm_bo_release_list);
 834                 return ret;
 835         }
 836
 837         put_count = ttm_bo_del_from_lru(bo);
 838         spin_unlock(&glob->lru_lock);
 839
 840         BUG_ON(ret != 0);
 841
 842         ttm_bo_list_ref_sub(bo, put_count, true);
 843
 844         ret = ttm_bo_evict(bo, interruptible, no_wait_gpu);
 845         ttm_bo_unreserve(bo);
 846
 847         kref_put(&bo->list_kref, ttm_bo_release_list);
 848         return ret;
 849 }
 850
 851 void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem)
 852 {
 853         struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type];
 854
 855         if (mem->mm_node)
 856                 (*man->func->put_node)(man, mem);
 857 }
 858 EXPORT_SYMBOL(ttm_bo_mem_put);
 859
 860 /**
 861  * Repeatedly evict memory from the LRU for @mem_type until we create enough
 862  * space, or we've evicted everything and there isn't enough space.
 863  */
 864 static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
 865                                         uint32_t mem_type,
 866                                         struct ttm_placement *placement,
 867                                         struct ttm_mem_reg *mem,
 868                                         bool interruptible,
 869                                         bool no_wait_gpu)
 870 {
 871         struct ttm_bo_device *bdev = bo->bdev;
 872         struct ttm_mem_type_manager *man = &bdev->man[mem_type];
 873         int ret;
 874
 875         do {
 876                 ret = (*man->func->get_node)(man, bo, placement, mem);
 877                 if (unlikely(ret != 0))
 878                         return ret;
 879                 if (mem->mm_node)
 880                         break;
 881                 ret = ttm_mem_evict_first(bdev, mem_type,
 882                                           interruptible, no_wait_gpu);
 883                 if (unlikely(ret != 0))
 884                         return ret;
 885         } while (1);
 886         if (mem->mm_node == NULL)
 887                 return -ENOMEM;
 888         mem->mem_type = mem_type;
 889         return 0;
 890 }
 891
 892 static uint32_t ttm_bo_select_caching(struct ttm_mem_type_manager *man,
 893                                       uint32_t cur_placement,
 894                                       uint32_t proposed_placement)
 895 {
 896         uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING;
 897         uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING;
 898
 899         /**
 900          * Keep current caching if possible.
 901          */
 902
 903         if ((cur_placement & caching) != 0)
 904                 result |= (cur_placement & caching);
 905         else if ((man->default_caching & caching) != 0)
 906                 result |= man->default_caching;
 907         else if ((TTM_PL_FLAG_CACHED & caching) != 0)
 908                 result |= TTM_PL_FLAG_CACHED;
 909         else if ((TTM_PL_FLAG_WC & caching) != 0)
 910                 result |= TTM_PL_FLAG_WC;
 911         else if ((TTM_PL_FLAG_UNCACHED & caching) != 0)
 912                 result |= TTM_PL_FLAG_UNCACHED;
 913
 914         return result;
 915 }
 916
 917 static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man,
 918                                  uint32_t mem_type,
 919                                  uint32_t proposed_placement,
 920                                  uint32_t *masked_placement)
 921 {
 922         uint32_t cur_flags = ttm_bo_type_flags(mem_type);
 923
 924         if ((cur_flags & proposed_placement & TTM_PL_MASK_MEM) == 0)
 925                 return false;
 926
 927         if ((proposed_placement & man->available_caching) == 0)
 928                 return false;
 929
 930         cur_flags |= (proposed_placement & man->available_caching);
 931
 932         *masked_placement = cur_flags;
 933         return true;
 934 }
 935
 936 /**
 937  * Creates space for memory region @mem according to its type.
 938  *
 939  * This function first searches for free space in compatible memory types in
 940  * the priority order defined by the driver.  If free space isn't found, then
 941  * ttm_bo_mem_force_space is attempted in priority order to evict and find
 942  * space.
 943  */
 944 int ttm_bo_mem_space(struct ttm_buffer_object *bo,
 945                         struct ttm_placement *placement,
 946                         struct ttm_mem_reg *mem,
 947                         bool interruptible,
 948                         bool no_wait_gpu)
 949 {
 950         struct ttm_bo_device *bdev = bo->bdev;
 951         struct ttm_mem_type_manager *man;
 952         uint32_t mem_type = TTM_PL_SYSTEM;
 953         uint32_t cur_flags = 0;
 954         bool type_found = false;
 955         bool type_ok = false;
 956         bool has_erestartsys = false;
 957         int i, ret;
 958
 959         mem->mm_node = NULL;
 960         for (i = 0; i < placement->num_placement; ++i) {
 961                 ret = ttm_mem_type_from_flags(placement->placement[i],
 962                                                 &mem_type);
 963                 if (ret)
 964                         return ret;
 965                 man = &bdev->man[mem_type];
 966
 967                 type_ok = ttm_bo_mt_compatible(man,
 968                                                 mem_type,
 969                                                 placement->placement[i],
 970                                                 &cur_flags);
 971
 972                 if (!type_ok)
 973                         continue;
 974
 975                 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
 976                                                   cur_flags);
 977                 /*
 978                  * Use the access and other non-mapping-related flag bits from
 979                  * the memory placement flags to the current flags
 980                  */
 981                 ttm_flag_masked(&cur_flags, placement->placement[i],
 982                                 ~TTM_PL_MASK_MEMTYPE);
 983
 984                 if (mem_type == TTM_PL_SYSTEM)
 985                         break;
 986
 987                 if (man->has_type && man->use_type) {
 988                         type_found = true;
 989                         ret = (*man->func->get_node)(man, bo, placement, mem);
 990                         if (unlikely(ret))
 991                                 return ret;
 992                 }
 993                 if (mem->mm_node)
 994                         break;
 995         }
 996
 997         if ((type_ok && (mem_type == TTM_PL_SYSTEM)) || mem->mm_node) {
 998                 mem->mem_type = mem_type;
 999                 mem->placement = cur_flags;
1000                 return 0;
1001         }
1002
1003         if (!type_found)
1004                 return -EINVAL;
1005
1006         for (i = 0; i < placement->num_busy_placement; ++i) {
1007                 ret = ttm_mem_type_from_flags(placement->busy_placement[i],
1008                                                 &mem_type);
1009                 if (ret)
1010                         return ret;
1011                 man = &bdev->man[mem_type];
1012                 if (!man->has_type)
1013                         continue;
1014                 if (!ttm_bo_mt_compatible(man,
1015                                                 mem_type,
1016                                                 placement->busy_placement[i],
1017                                                 &cur_flags))
1018                         continue;
1019
1020                 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
1021                                                   cur_flags);
1022                 /*
1023                  * Use the access and other non-mapping-related flag bits from
1024                  * the memory placement flags to the current flags
1025                  */
1026                 ttm_flag_masked(&cur_flags, placement->busy_placement[i],
1027                                 ~TTM_PL_MASK_MEMTYPE);
1028
1029
1030                 if (mem_type == TTM_PL_SYSTEM) {
1031                         mem->mem_type = mem_type;
1032                         mem->placement = cur_flags;
1033                         mem->mm_node = NULL;
1034                         return 0;
1035                 }
1036
1037                 ret = ttm_bo_mem_force_space(bo, mem_type, placement, mem,
1038                                                 interruptible, no_wait_gpu);
1039                 if (ret == 0 && mem->mm_node) {
1040                         mem->placement = cur_flags;
1041                         return 0;
1042                 }
1043                 if (ret == -ERESTARTSYS)
1044                         has_erestartsys = true;
1045         }
1046         ret = (has_erestartsys) ? -ERESTARTSYS : -ENOMEM;
1047         return ret;
1048 }
1049 EXPORT_SYMBOL(ttm_bo_mem_space);
1050
1051 int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
1052                         struct ttm_placement *placement,
1053                         bool interruptible,
1054                         bool no_wait_gpu)
1055 {
1056         int ret = 0;
1057         struct ttm_mem_reg mem;
1058         struct ttm_bo_device *bdev = bo->bdev;
1059
1060         BUG_ON(!ttm_bo_is_reserved(bo));
1061
1062         /*
1063          * FIXME: It's possible to pipeline buffer moves.
1064          * Have the driver move function wait for idle when necessary,
1065          * instead of doing it here.
1066          */
1067         spin_lock(&bdev->fence_lock);
1068         ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
1069         spin_unlock(&bdev->fence_lock);
1070         if (ret)
1071                 return ret;
1072         mem.num_pages = bo->num_pages;
1073         mem.size = mem.num_pages << PAGE_SHIFT;
1074         mem.page_alignment = bo->mem.page_alignment;
1075         mem.bus.io_reserved_vm = false;
1076         mem.bus.io_reserved_count = 0;
1077         /*
1078          * Determine where to move the buffer.
1079          */
1080         ret = ttm_bo_mem_space(bo, placement, &mem,
1081                                interruptible, no_wait_gpu);
1082         if (ret)
1083                 goto out_unlock;
1084         ret = ttm_bo_handle_move_mem(bo, &mem, false,
1085                                      interruptible, no_wait_gpu);
1086 out_unlock:
1087         if (ret && mem.mm_node)
1088                 ttm_bo_mem_put(bo, &mem);
1089         return ret;
1090 }
1091
1092 static int ttm_bo_mem_compat(struct ttm_placement *placement,
1093                              struct ttm_mem_reg *mem)
1094 {
1095         int i;
1096
1097         if (mem->mm_node && placement->lpfn != 0 &&
1098             (mem->start < placement->fpfn ||
1099              mem->start + mem->num_pages > placement->lpfn))
1100                 return -1;
1101
1102         for (i = 0; i < placement->num_placement; i++) {
1103                 if ((placement->placement[i] & mem->placement &
1104                         TTM_PL_MASK_CACHING) &&
1105                         (placement->placement[i] & mem->placement &
1106                         TTM_PL_MASK_MEM))
1107                         return i;
1108         }
1109         return -1;
1110 }
1111
1112 int ttm_bo_validate(struct ttm_buffer_object *bo,
1113                         struct ttm_placement *placement,
1114                         bool interruptible,
1115                         bool no_wait_gpu)
1116 {
1117         int ret;
1118
1119         BUG_ON(!ttm_bo_is_reserved(bo));
1120         /* Check that range is valid */
1121         if (placement->lpfn || placement->fpfn)
1122                 if (placement->fpfn > placement->lpfn ||
1123                         (placement->lpfn - placement->fpfn) < bo->num_pages)
1124                         return -EINVAL;
1125         /*
1126          * Check whether we need to move buffer.
1127          */
1128         ret = ttm_bo_mem_compat(placement, &bo->mem);
1129         if (ret < 0) {
1130                 ret = ttm_bo_move_buffer(bo, placement, interruptible,
1131                                          no_wait_gpu);
1132                 if (ret)
1133                         return ret;
1134         } else {
1135                 /*
1136                  * Use the access and other non-mapping-related flag bits from
1137                  * the compatible memory placement flags to the active flags
1138                  */
1139                 ttm_flag_masked(&bo->mem.placement, placement->placement[ret],
1140                                 ~TTM_PL_MASK_MEMTYPE);
1141         }
1142         /*
1143          * We might need to add a TTM.
1144          */
1145         if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
1146                 ret = ttm_bo_add_ttm(bo, true);
1147                 if (ret)
1148                         return ret;
1149         }
1150         return 0;
1151 }
1152 EXPORT_SYMBOL(ttm_bo_validate);
1153
1154 int ttm_bo_check_placement(struct ttm_buffer_object *bo,
1155                                 struct ttm_placement *placement)
1156 {
1157         BUG_ON((placement->fpfn || placement->lpfn) &&
1158                (bo->mem.num_pages > (placement->lpfn - placement->fpfn)));
1159
1160         return 0;
1161 }
1162
1163 int ttm_bo_init(struct ttm_bo_device *bdev,
1164                 struct ttm_buffer_object *bo,
1165                 unsigned long size,
1166                 enum ttm_bo_type type,
1167                 struct ttm_placement *placement,
1168                 uint32_t page_alignment,
1169                 bool interruptible,
1170                 struct file *persistent_swap_storage,
1171                 size_t acc_size,
1172                 struct sg_table *sg,
1173                 void (*destroy) (struct ttm_buffer_object *))
1174 {
1175         int ret = 0;
1176         unsigned long num_pages;
1177         struct ttm_mem_global *mem_glob = bdev->glob->mem_glob;
1178
1179         ret = ttm_mem_global_alloc(mem_glob, acc_size, false, false);
1180         if (ret) {
1181                 pr_err("Out of kernel memory\n");
1182                 if (destroy)
1183                         (*destroy)(bo);
1184                 else
1185                         kfree(bo);
1186                 return -ENOMEM;
1187         }
1188
1189         num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1190         if (num_pages == 0) {
1191                 pr_err("Illegal buffer object size\n");
1192                 if (destroy)
1193                         (*destroy)(bo);
1194                 else
1195                         kfree(bo);
1196                 ttm_mem_global_free(mem_glob, acc_size);
1197                 return -EINVAL;
1198         }
1199         bo->destroy = destroy;
1200
1201         kref_init(&bo->kref);
1202         kref_init(&bo->list_kref);
1203         atomic_set(&bo->cpu_writers, 0);
1204         atomic_set(&bo->reserved, 1);
1205         init_waitqueue_head(&bo->event_queue);
1206         INIT_LIST_HEAD(&bo->lru);
1207         INIT_LIST_HEAD(&bo->ddestroy);
1208         INIT_LIST_HEAD(&bo->swap);
1209         INIT_LIST_HEAD(&bo->io_reserve_lru);
1210         bo->bdev = bdev;
1211         bo->glob = bdev->glob;
1212         bo->type = type;
1213         bo->num_pages = num_pages;
1214         bo->mem.size = num_pages << PAGE_SHIFT;
1215         bo->mem.mem_type = TTM_PL_SYSTEM;
1216         bo->mem.num_pages = bo->num_pages;
1217         bo->mem.mm_node = NULL;
1218         bo->mem.page_alignment = page_alignment;
1219         bo->mem.bus.io_reserved_vm = false;
1220         bo->mem.bus.io_reserved_count = 0;
1221         bo->priv_flags = 0;
1222         bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED);
1223         bo->seq_valid = false;
1224         bo->persistent_swap_storage = persistent_swap_storage;
1225         bo->acc_size = acc_size;
1226         bo->sg = sg;
1227         atomic_inc(&bo->glob->bo_count);
1228
1229         ret = ttm_bo_check_placement(bo, placement);
1230         if (unlikely(ret != 0))
1231                 goto out_err;
1232
1233         /*
1234          * For ttm_bo_type_device buffers, allocate
1235          * address space from the device.
1236          */
1237         if (bo->type == ttm_bo_type_device ||
1238             bo->type == ttm_bo_type_sg) {
1239                 ret = ttm_bo_setup_vm(bo);
1240                 if (ret)
1241                         goto out_err;
1242         }
1243
1244         ret = ttm_bo_validate(bo, placement, interruptible, false);
1245         if (ret)
1246                 goto out_err;
1247
1248         ttm_bo_unreserve(bo);
1249         return 0;
1250
1251 out_err:
1252         ttm_bo_unreserve(bo);
1253         ttm_bo_unref(&bo);
1254
1255         return ret;
1256 }
1257 EXPORT_SYMBOL(ttm_bo_init);
1258
1259 size_t ttm_bo_acc_size(struct ttm_bo_device *bdev,
1260                        unsigned long bo_size,
1261                        unsigned struct_size)
1262 {
1263         unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1264         size_t size = 0;
1265
1266         size += ttm_round_pot(struct_size);
1267         size += PAGE_ALIGN(npages * sizeof(void *));
1268         size += ttm_round_pot(sizeof(struct ttm_tt));
1269         return size;
1270 }
1271 EXPORT_SYMBOL(ttm_bo_acc_size);
1272
1273 size_t ttm_bo_dma_acc_size(struct ttm_bo_device *bdev,
1274                            unsigned long bo_size,
1275                            unsigned struct_size)
1276 {
1277         unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1278         size_t size = 0;
1279
1280         size += ttm_round_pot(struct_size);
1281         size += PAGE_ALIGN(npages * sizeof(void *));
1282         size += PAGE_ALIGN(npages * sizeof(dma_addr_t));
1283         size += ttm_round_pot(sizeof(struct ttm_dma_tt));
1284         return size;
1285 }
1286 EXPORT_SYMBOL(ttm_bo_dma_acc_size);
1287
1288 int ttm_bo_create(struct ttm_bo_device *bdev,
1289                         unsigned long size,
1290                         enum ttm_bo_type type,
1291                         struct ttm_placement *placement,
1292                         uint32_t page_alignment,
1293                         bool interruptible,
1294                         struct file *persistent_swap_storage,
1295                         struct ttm_buffer_object **p_bo)
1296 {
1297         struct ttm_buffer_object *bo;
1298         size_t acc_size;
1299         int ret;
1300
1301         bo = kzalloc(sizeof(*bo), GFP_KERNEL);
1302         if (unlikely(bo == NULL))
1303                 return -ENOMEM;
1304
1305         acc_size = ttm_bo_acc_size(bdev, size, sizeof(struct ttm_buffer_object));
1306         ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment,
1307                           interruptible, persistent_swap_storage, acc_size,
1308                           NULL, NULL);
1309         if (likely(ret == 0))
1310                 *p_bo = bo;
1311
1312         return ret;
1313 }
1314 EXPORT_SYMBOL(ttm_bo_create);
1315
1316 static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev,
1317                                         unsigned mem_type, bool allow_errors)
1318 {
1319         struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1320         struct ttm_bo_global *glob = bdev->glob;
1321         int ret;
1322
1323         /*
1324          * Can't use standard list traversal since we're unlocking.
1325          */
1326
1327         spin_lock(&glob->lru_lock);
1328         while (!list_empty(&man->lru)) {
1329                 spin_unlock(&glob->lru_lock);
1330                 ret = ttm_mem_evict_first(bdev, mem_type, false, false);
1331                 if (ret) {
1332                         if (allow_errors) {
1333                                 return ret;
1334                         } else {
1335                                 pr_err("Cleanup eviction failed\n");
1336                         }
1337                 }
1338                 spin_lock(&glob->lru_lock);
1339         }
1340         spin_unlock(&glob->lru_lock);
1341         return 0;
1342 }
1343
1344 int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1345 {
1346         struct ttm_mem_type_manager *man;
1347         int ret = -EINVAL;
1348
1349         if (mem_type >= TTM_NUM_MEM_TYPES) {
1350                 pr_err("Illegal memory type %d\n", mem_type);
1351                 return ret;
1352         }
1353         man = &bdev->man[mem_type];
1354
1355         if (!man->has_type) {
1356                 pr_err("Trying to take down uninitialized memory manager type %u\n",
1357                        mem_type);
1358                 return ret;
1359         }
1360
1361         man->use_type = false;
1362         man->has_type = false;
1363
1364         ret = 0;
1365         if (mem_type > 0) {
1366                 ttm_bo_force_list_clean(bdev, mem_type, false);
1367
1368                 ret = (*man->func->takedown)(man);
1369         }
1370
1371         return ret;
1372 }
1373 EXPORT_SYMBOL(ttm_bo_clean_mm);
1374
1375 int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1376 {
1377         struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1378
1379         if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) {
1380                 pr_err("Illegal memory manager memory type %u\n", mem_type);
1381                 return -EINVAL;
1382         }
1383
1384         if (!man->has_type) {
1385                 pr_err("Memory type %u has not been initialized\n", mem_type);
1386                 return 0;
1387         }
1388
1389         return ttm_bo_force_list_clean(bdev, mem_type, true);
1390 }
1391 EXPORT_SYMBOL(ttm_bo_evict_mm);
1392
1393 int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type,
1394                         unsigned long p_size)
1395 {
1396         int ret = -EINVAL;
1397         struct ttm_mem_type_manager *man;
1398
1399         BUG_ON(type >= TTM_NUM_MEM_TYPES);
1400         man = &bdev->man[type];
1401         BUG_ON(man->has_type);
1402         man->io_reserve_fastpath = true;
1403         man->use_io_reserve_lru = false;
1404         mutex_init(&man->io_reserve_mutex);
1405         INIT_LIST_HEAD(&man->io_reserve_lru);
1406
1407         ret = bdev->driver->init_mem_type(bdev, type, man);
1408         if (ret)
1409                 return ret;
1410         man->bdev = bdev;
1411
1412         ret = 0;
1413         if (type != TTM_PL_SYSTEM) {
1414                 ret = (*man->func->init)(man, p_size);
1415                 if (ret)
1416                         return ret;
1417         }
1418         man->has_type = true;
1419         man->use_type = true;
1420         man->size = p_size;
1421
1422         INIT_LIST_HEAD(&man->lru);
1423
1424         return 0;
1425 }
1426 EXPORT_SYMBOL(ttm_bo_init_mm);
1427
1428 static void ttm_bo_global_kobj_release(struct kobject *kobj)
1429 {
1430         struct ttm_bo_global *glob =
1431                 container_of(kobj, struct ttm_bo_global, kobj);
1432
1433         ttm_mem_unregister_shrink(glob->mem_glob, &glob->shrink);
1434         __free_page(glob->dummy_read_page);
1435         kfree(glob);
1436 }
1437
1438 void ttm_bo_global_release(struct drm_global_reference *ref)
1439 {
1440         struct ttm_bo_global *glob = ref->object;
1441
1442         kobject_del(&glob->kobj);
1443         kobject_put(&glob->kobj);
1444 }
1445 EXPORT_SYMBOL(ttm_bo_global_release);
1446
1447 int ttm_bo_global_init(struct drm_global_reference *ref)
1448 {
1449         struct ttm_bo_global_ref *bo_ref =
1450                 container_of(ref, struct ttm_bo_global_ref, ref);
1451         struct ttm_bo_global *glob = ref->object;
1452         int ret;
1453
1454         mutex_init(&glob->device_list_mutex);
1455         spin_lock_init(&glob->lru_lock);
1456         glob->mem_glob = bo_ref->mem_glob;
1457         glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32);
1458
1459         if (unlikely(glob->dummy_read_page == NULL)) {
1460                 ret = -ENOMEM;
1461                 goto out_no_drp;
1462         }
1463
1464         INIT_LIST_HEAD(&glob->swap_lru);
1465         INIT_LIST_HEAD(&glob->device_list);
1466
1467         ttm_mem_init_shrink(&glob->shrink, ttm_bo_swapout);
1468         ret = ttm_mem_register_shrink(glob->mem_glob, &glob->shrink);
1469         if (unlikely(ret != 0)) {
1470                 pr_err("Could not register buffer object swapout\n");
1471                 goto out_no_shrink;
1472         }
1473
1474         atomic_set(&glob->bo_count, 0);
1475
1476         ret = kobject_init_and_add(
1477                 &glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects");
1478         if (unlikely(ret != 0))
1479                 kobject_put(&glob->kobj);
1480         return ret;
1481 out_no_shrink:
1482         __free_page(glob->dummy_read_page);
1483 out_no_drp:
1484         kfree(glob);
1485         return ret;
1486 }
1487 EXPORT_SYMBOL(ttm_bo_global_init);
1488
1489
1490 int ttm_bo_device_release(struct ttm_bo_device *bdev)
1491 {
1492         int ret = 0;
1493         unsigned i = TTM_NUM_MEM_TYPES;
1494         struct ttm_mem_type_manager *man;
1495         struct ttm_bo_global *glob = bdev->glob;
1496
1497         while (i--) {
1498                 man = &bdev->man[i];
1499                 if (man->has_type) {
1500                         man->use_type = false;
1501                         if ((i != TTM_PL_SYSTEM) && ttm_bo_clean_mm(bdev, i)) {
1502                                 ret = -EBUSY;
1503                                 pr_err("DRM memory manager type %d is not clean\n",
1504                                        i);
1505                         }
1506                         man->has_type = false;
1507                 }
1508         }
1509
1510         mutex_lock(&glob->device_list_mutex);
1511         list_del(&bdev->device_list);
1512         mutex_unlock(&glob->device_list_mutex);
1513
1514         cancel_delayed_work_sync(&bdev->wq);
1515
1516         while (ttm_bo_delayed_delete(bdev, true))
1517                 ;
1518
1519         spin_lock(&glob->lru_lock);
1520         if (list_empty(&bdev->ddestroy))
1521                 TTM_DEBUG("Delayed destroy list was clean\n");
1522
1523         if (list_empty(&bdev->man[0].lru))
1524                 TTM_DEBUG("Swap list was clean\n");
1525         spin_unlock(&glob->lru_lock);
1526
1527         BUG_ON(!drm_mm_clean(&bdev->addr_space_mm));
1528         write_lock(&bdev->vm_lock);
1529         drm_mm_takedown(&bdev->addr_space_mm);
1530         write_unlock(&bdev->vm_lock);
1531
1532         return ret;
1533 }
1534 EXPORT_SYMBOL(ttm_bo_device_release);
1535
1536 int ttm_bo_device_init(struct ttm_bo_device *bdev,
1537                        struct ttm_bo_global *glob,
1538                        struct ttm_bo_driver *driver,
1539                        uint64_t file_page_offset,
1540                        bool need_dma32)
1541 {
1542         int ret = -EINVAL;
1543
1544         rwlock_init(&bdev->vm_lock);
1545         bdev->driver = driver;
1546
1547         memset(bdev->man, 0, sizeof(bdev->man));
1548
1549         /*
1550          * Initialize the system memory buffer type.
1551          * Other types need to be driver / IOCTL initialized.
1552          */
1553         ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0);
1554         if (unlikely(ret != 0))
1555                 goto out_no_sys;
1556
1557         bdev->addr_space_rb = RB_ROOT;
1558         ret = drm_mm_init(&bdev->addr_space_mm, file_page_offset, 0x10000000);
1559         if (unlikely(ret != 0))
1560                 goto out_no_addr_mm;
1561
1562         INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue);
1563         INIT_LIST_HEAD(&bdev->ddestroy);
1564         bdev->dev_mapping = NULL;
1565         bdev->glob = glob;
1566         bdev->need_dma32 = need_dma32;
1567         bdev->val_seq = 0;
1568         spin_lock_init(&bdev->fence_lock);
1569         mutex_lock(&glob->device_list_mutex);
1570         list_add_tail(&bdev->device_list, &glob->device_list);
1571         mutex_unlock(&glob->device_list_mutex);
1572
1573         return 0;
1574 out_no_addr_mm:
1575         ttm_bo_clean_mm(bdev, 0);
1576 out_no_sys:
1577         return ret;
1578 }
1579 EXPORT_SYMBOL(ttm_bo_device_init);
1580
1581 /*
1582  * buffer object vm functions.
1583  */
1584
1585 bool ttm_mem_reg_is_pci(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
1586 {
1587         struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
1588
1589         if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
1590                 if (mem->mem_type == TTM_PL_SYSTEM)
1591                         return false;
1592
1593                 if (man->flags & TTM_MEMTYPE_FLAG_CMA)
1594                         return false;
1595
1596                 if (mem->placement & TTM_PL_FLAG_CACHED)
1597                         return false;
1598         }
1599         return true;
1600 }
1601
1602 void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo)
1603 {
1604         struct ttm_bo_device *bdev = bo->bdev;
1605         loff_t offset = (loff_t) bo->addr_space_offset;
1606         loff_t holelen = ((loff_t) bo->mem.num_pages) << PAGE_SHIFT;
1607
1608         if (!bdev->dev_mapping)
1609                 return;
1610         unmap_mapping_range(bdev->dev_mapping, offset, holelen, 1);
1611         ttm_mem_io_free_vm(bo);
1612 }
1613
1614 void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
1615 {
1616         struct ttm_bo_device *bdev = bo->bdev;
1617         struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
1618
1619         ttm_mem_io_lock(man, false);
1620         ttm_bo_unmap_virtual_locked(bo);
1621         ttm_mem_io_unlock(man);
1622 }
1623
1624
1625 EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1626
1627 static void ttm_bo_vm_insert_rb(struct ttm_buffer_object *bo)
1628 {
1629         struct ttm_bo_device *bdev = bo->bdev;
1630         struct rb_node **cur = &bdev->addr_space_rb.rb_node;
1631         struct rb_node *parent = NULL;
1632         struct ttm_buffer_object *cur_bo;
1633         unsigned long offset = bo->vm_node->start;
1634         unsigned long cur_offset;
1635
1636         while (*cur) {
1637                 parent = *cur;
1638                 cur_bo = rb_entry(parent, struct ttm_buffer_object, vm_rb);
1639                 cur_offset = cur_bo->vm_node->start;
1640                 if (offset < cur_offset)
1641                         cur = &parent->rb_left;
1642                 else if (offset > cur_offset)
1643                         cur = &parent->rb_right;
1644                 else
1645                         BUG();
1646         }
1647
1648         rb_link_node(&bo->vm_rb, parent, cur);
1649         rb_insert_color(&bo->vm_rb, &bdev->addr_space_rb);
1650 }
1651
1652 /**
1653  * ttm_bo_setup_vm:
1654  *
1655  * @bo: the buffer to allocate address space for
1656  *
1657  * Allocate address space in the drm device so that applications
1658  * can mmap the buffer and access the contents. This only
1659  * applies to ttm_bo_type_device objects as others are not
1660  * placed in the drm device address space.
1661  */
1662
1663 static int ttm_bo_setup_vm(struct ttm_buffer_object *bo)
1664 {
1665         struct ttm_bo_device *bdev = bo->bdev;
1666         int ret;
1667
1668 retry_pre_get:
1669         ret = drm_mm_pre_get(&bdev->addr_space_mm);
1670         if (unlikely(ret != 0))
1671                 return ret;
1672
1673         write_lock(&bdev->vm_lock);
1674         bo->vm_node = drm_mm_search_free(&bdev->addr_space_mm,
1675                                          bo->mem.num_pages, 0, 0);
1676
1677         if (unlikely(bo->vm_node == NULL)) {
1678                 ret = -ENOMEM;
1679                 goto out_unlock;
1680         }
1681
1682         bo->vm_node = drm_mm_get_block_atomic(bo->vm_node,
1683                                               bo->mem.num_pages, 0);
1684
1685         if (unlikely(bo->vm_node == NULL)) {
1686                 write_unlock(&bdev->vm_lock);
1687                 goto retry_pre_get;
1688         }
1689
1690         ttm_bo_vm_insert_rb(bo);
1691         write_unlock(&bdev->vm_lock);
1692         bo->addr_space_offset = ((uint64_t) bo->vm_node->start) << PAGE_SHIFT;
1693
1694         return 0;
1695 out_unlock:
1696         write_unlock(&bdev->vm_lock);
1697         return ret;
1698 }
1699
1700 int ttm_bo_wait(struct ttm_buffer_object *bo,
1701                 bool lazy, bool interruptible, bool no_wait)
1702 {
1703         struct ttm_bo_driver *driver = bo->bdev->driver;
1704         struct ttm_bo_device *bdev = bo->bdev;
1705         void *sync_obj;
1706         int ret = 0;
1707
1708         if (likely(bo->sync_obj == NULL))
1709                 return 0;
1710
1711         while (bo->sync_obj) {
1712
1713                 if (driver->sync_obj_signaled(bo->sync_obj)) {
1714                         void *tmp_obj = bo->sync_obj;
1715                         bo->sync_obj = NULL;
1716                         clear_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags);
1717                         spin_unlock(&bdev->fence_lock);
1718                         driver->sync_obj_unref(&tmp_obj);
1719                         spin_lock(&bdev->fence_lock);
1720                         continue;
1721                 }
1722
1723                 if (no_wait)
1724                         return -EBUSY;
1725
1726                 sync_obj = driver->sync_obj_ref(bo->sync_obj);
1727                 spin_unlock(&bdev->fence_lock);
1728                 ret = driver->sync_obj_wait(sync_obj,
1729                                             lazy, interruptible);
1730                 if (unlikely(ret != 0)) {
1731                         driver->sync_obj_unref(&sync_obj);
1732                         spin_lock(&bdev->fence_lock);
1733                         return ret;
1734                 }
1735                 spin_lock(&bdev->fence_lock);
1736                 if (likely(bo->sync_obj == sync_obj)) {
1737                         void *tmp_obj = bo->sync_obj;
1738                         bo->sync_obj = NULL;
1739                         clear_bit(TTM_BO_PRIV_FLAG_MOVING,
1740                                   &bo->priv_flags);
1741                         spin_unlock(&bdev->fence_lock);
1742                         driver->sync_obj_unref(&sync_obj);
1743                         driver->sync_obj_unref(&tmp_obj);
1744                         spin_lock(&bdev->fence_lock);
1745                 } else {
1746                         spin_unlock(&bdev->fence_lock);
1747                         driver->sync_obj_unref(&sync_obj);
1748                         spin_lock(&bdev->fence_lock);
1749                 }
1750         }
1751         return 0;
1752 }
1753 EXPORT_SYMBOL(ttm_bo_wait);
1754
1755 int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait)
1756 {
1757         struct ttm_bo_device *bdev = bo->bdev;
1758         int ret = 0;
1759
1760         /*
1761          * Using ttm_bo_reserve makes sure the lru lists are updated.
1762          */
1763
1764         ret = ttm_bo_reserve(bo, true, no_wait, false, 0);
1765         if (unlikely(ret != 0))
1766                 return ret;
1767         spin_lock(&bdev->fence_lock);
1768         ret = ttm_bo_wait(bo, false, true, no_wait);
1769         spin_unlock(&bdev->fence_lock);
1770         if (likely(ret == 0))
1771                 atomic_inc(&bo->cpu_writers);
1772         ttm_bo_unreserve(bo);
1773         return ret;
1774 }
1775 EXPORT_SYMBOL(ttm_bo_synccpu_write_grab);
1776
1777 void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo)
1778 {
1779         atomic_dec(&bo->cpu_writers);
1780 }
1781 EXPORT_SYMBOL(ttm_bo_synccpu_write_release);
1782
1783 /**
1784  * A buffer object shrink method that tries to swap out the first
1785  * buffer object on the bo_global::swap_lru list.
1786  */
1787
1788 static int ttm_bo_swapout(struct ttm_mem_shrink *shrink)
1789 {
1790         struct ttm_bo_global *glob =
1791             container_of(shrink, struct ttm_bo_global, shrink);
1792         struct ttm_buffer_object *bo;
1793         int ret = -EBUSY;
1794         int put_count;
1795         uint32_t swap_placement = (TTM_PL_FLAG_CACHED | TTM_PL_FLAG_SYSTEM);
1796
1797         spin_lock(&glob->lru_lock);
1798         list_for_each_entry(bo, &glob->swap_lru, swap) {
1799                 ret = ttm_bo_reserve_locked(bo, false, true, false, 0);
1800                 if (!ret)
1801                         break;
1802         }
1803
1804         if (ret) {
1805                 spin_unlock(&glob->lru_lock);
1806                 return ret;
1807         }
1808
1809         kref_get(&bo->list_kref);
1810
1811         if (!list_empty(&bo->ddestroy)) {
1812                 ret = ttm_bo_cleanup_refs_and_unlock(bo, false, false);
1813                 kref_put(&bo->list_kref, ttm_bo_release_list);
1814                 return ret;
1815         }
1816
1817         put_count = ttm_bo_del_from_lru(bo);
1818         spin_unlock(&glob->lru_lock);
1819
1820         ttm_bo_list_ref_sub(bo, put_count, true);
1821
1822         /**
1823          * Wait for GPU, then move to system cached.
1824          */
1825
1826         spin_lock(&bo->bdev->fence_lock);
1827         ret = ttm_bo_wait(bo, false, false, false);
1828         spin_unlock(&bo->bdev->fence_lock);
1829
1830         if (unlikely(ret != 0))
1831                 goto out;
1832
1833         if ((bo->mem.placement & swap_placement) != swap_placement) {
1834                 struct ttm_mem_reg evict_mem;
1835
1836                 evict_mem = bo->mem;
1837                 evict_mem.mm_node = NULL;
1838                 evict_mem.placement = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED;
1839                 evict_mem.mem_type = TTM_PL_SYSTEM;
1840
1841                 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true,
1842                                              false, false);
1843                 if (unlikely(ret != 0))
1844                         goto out;
1845         }
1846
1847         ttm_bo_unmap_virtual(bo);
1848
1849         /**
1850          * Swap out. Buffer will be swapped in again as soon as
1851          * anyone tries to access a ttm page.
1852          */
1853
1854         if (bo->bdev->driver->swap_notify)
1855                 bo->bdev->driver->swap_notify(bo);
1856
1857         ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage);
1858 out:
1859
1860         /**
1861          *
1862          * Unreserve without putting on LRU to avoid swapping out an
1863          * already swapped buffer.
1864          */
1865
1866         atomic_set(&bo->reserved, 0);
1867         wake_up_all(&bo->event_queue);
1868         kref_put(&bo->list_kref, ttm_bo_release_list);
1869         return ret;
1870 }
1871
1872 void ttm_bo_swapout_all(struct ttm_bo_device *bdev)
1873 {
1874         while (ttm_bo_swapout(&bdev->glob->shrink) == 0)
1875                 ;
1876 }
1877 EXPORT_SYMBOL(ttm_bo_swapout_all);