sys/dev/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  * $FreeBSD: head/sys/dev/drm2/ttm/ttm_bo.c 248060 2013-03-08 18:11:02Z dumbbell $
  31  */
  32
  33 #include <dev/drm/drmP.h>
  34 #include <dev/drm/ttm/ttm_module.h>
  35 #include <dev/drm/ttm/ttm_bo_driver.h>
  36 #include <dev/drm/ttm/ttm_placement.h>
  37
  38 #define TTM_ASSERT_LOCKED(param)
  39 #define TTM_DEBUG(fmt, arg...)
  40 #define TTM_BO_HASH_ORDER 13
  41
  42 static int ttm_bo_setup_vm(struct ttm_buffer_object *bo);
  43 static int ttm_bo_swapout(struct ttm_mem_shrink *shrink);
  44 static void ttm_bo_global_kobj_release(struct ttm_bo_global *glob);
  45
  46 MALLOC_DEFINE(M_TTM_BO, "ttm_bo", "TTM Buffer Objects");
  47
  48 static inline int ttm_mem_type_from_flags(uint32_t flags, uint32_t *mem_type)
  49 {
  50         int i;
  51
  52         for (i = 0; i <= TTM_PL_PRIV5; i++)
  53                 if (flags & (1 << i)) {
  54                         *mem_type = i;
  55                         return 0;
  56                 }
  57         return -EINVAL;
  58 }
  59
  60 static void ttm_mem_type_debug(struct ttm_bo_device *bdev, int mem_type)
  61 {
  62         struct ttm_mem_type_manager *man = &bdev->man[mem_type];
  63
  64         kprintf("    has_type: %d\n", man->has_type);
  65         kprintf("    use_type: %d\n", man->use_type);
  66         kprintf("    flags: 0x%08X\n", man->flags);
  67         kprintf("    gpu_offset: 0x%08lX\n", man->gpu_offset);
  68         kprintf("    size: %ju\n", (uintmax_t)man->size);
  69         kprintf("    available_caching: 0x%08X\n", man->available_caching);
  70         kprintf("    default_caching: 0x%08X\n", man->default_caching);
  71         if (mem_type != TTM_PL_SYSTEM)
  72                 (*man->func->debug)(man, TTM_PFX);
  73 }
  74
  75 static void ttm_bo_mem_space_debug(struct ttm_buffer_object *bo,
  76                                         struct ttm_placement *placement)
  77 {
  78         int i, ret, mem_type;
  79
  80         kprintf("No space for %p (%lu pages, %luK, %luM)\n",
  81                bo, bo->mem.num_pages, bo->mem.size >> 10,
  82                bo->mem.size >> 20);
  83         for (i = 0; i < placement->num_placement; i++) {
  84                 ret = ttm_mem_type_from_flags(placement->placement[i],
  85                                                 &mem_type);
  86                 if (ret)
  87                         return;
  88                 kprintf("  placement[%d]=0x%08X (%d)\n",
  89                        i, placement->placement[i], mem_type);
  90                 ttm_mem_type_debug(bo->bdev, mem_type);
  91         }
  92 }
  93
  94 #if 0
  95 static ssize_t ttm_bo_global_show(struct ttm_bo_global *glob,
  96     char *buffer)
  97 {
  98
  99         return snprintf(buffer, PAGE_SIZE, "%lu\n",
 100                         (unsigned long) atomic_read(&glob->bo_count));
 101 }
 102 #endif
 103
 104 static inline uint32_t ttm_bo_type_flags(unsigned type)
 105 {
 106         return 1 << (type);
 107 }
 108
 109 static void ttm_bo_release_list(struct ttm_buffer_object *bo)
 110 {
 111         struct ttm_bo_device *bdev = bo->bdev;
 112         size_t acc_size = bo->acc_size;
 113
 114         KKASSERT(atomic_read(&bo->list_kref) == 0);
 115         KKASSERT(atomic_read(&bo->kref) == 0);
 116         KKASSERT(atomic_read(&bo->cpu_writers) == 0);
 117         KKASSERT(bo->sync_obj == NULL);
 118         KKASSERT(bo->mem.mm_node == NULL);
 119         KKASSERT(list_empty(&bo->lru));
 120         KKASSERT(list_empty(&bo->ddestroy));
 121
 122         if (bo->ttm)
 123                 ttm_tt_destroy(bo->ttm);
 124         atomic_dec(&bo->glob->bo_count);
 125         if (bo->destroy)
 126                 bo->destroy(bo);
 127         else {
 128                 drm_free(bo, M_TTM_BO);
 129         }
 130         ttm_mem_global_free(bdev->glob->mem_glob, acc_size);
 131 }
 132
 133 int
 134 ttm_bo_wait_unreserved_locked(struct ttm_buffer_object *bo, bool interruptible)
 135 {
 136         const char *wmsg;
 137         int flags, ret;
 138
 139         ret = 0;
 140         if (interruptible) {
 141                 flags = PCATCH;
 142                 wmsg = "ttbowi";
 143         } else {
 144                 flags = 0;
 145                 wmsg = "ttbowu";
 146         }
 147         while (!ttm_bo_is_reserved(bo)) {
 148                 ret = -lksleep(bo, &bo->glob->lru_lock, 0, wmsg, 0);
 149                 if (ret != 0)
 150                         break;
 151         }
 152         return (ret);
 153 }
 154
 155 void ttm_bo_add_to_lru(struct ttm_buffer_object *bo)
 156 {
 157         struct ttm_bo_device *bdev = bo->bdev;
 158         struct ttm_mem_type_manager *man;
 159
 160         KKASSERT(ttm_bo_is_reserved(bo));
 161
 162         if (!(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
 163
 164                 KKASSERT(list_empty(&bo->lru));
 165
 166                 man = &bdev->man[bo->mem.mem_type];
 167                 list_add_tail(&bo->lru, &man->lru);
 168                 refcount_acquire(&bo->list_kref);
 169
 170                 if (bo->ttm != NULL) {
 171                         list_add_tail(&bo->swap, &bo->glob->swap_lru);
 172                         refcount_acquire(&bo->list_kref);
 173                 }
 174         }
 175 }
 176
 177 int ttm_bo_del_from_lru(struct ttm_buffer_object *bo)
 178 {
 179         int put_count = 0;
 180
 181         if (!list_empty(&bo->swap)) {
 182                 list_del_init(&bo->swap);
 183                 ++put_count;
 184         }
 185         if (!list_empty(&bo->lru)) {
 186                 list_del_init(&bo->lru);
 187                 ++put_count;
 188         }
 189
 190         /*
 191          * TODO: Add a driver hook to delete from
 192          * driver-specific LRU's here.
 193          */
 194
 195         return put_count;
 196 }
 197
 198 int ttm_bo_reserve_locked(struct ttm_buffer_object *bo,
 199                           bool interruptible,
 200                           bool no_wait, bool use_sequence, uint32_t sequence)
 201 {
 202         int ret;
 203
 204         while (unlikely(atomic_read(&bo->reserved) != 0)) {
 205                 /**
 206                  * Deadlock avoidance for multi-bo reserving.
 207                  */
 208                 if (use_sequence && bo->seq_valid) {
 209                         /**
 210                          * We've already reserved this one.
 211                          */
 212                         if (unlikely(sequence == bo->val_seq))
 213                                 return -EDEADLK;
 214                         /**
 215                          * Already reserved by a thread that will not back
 216                          * off for us. We need to back off.
 217                          */
 218                         if (unlikely(sequence - bo->val_seq < (1 << 31)))
 219                                 return -EAGAIN;
 220                 }
 221
 222                 if (no_wait)
 223                         return -EBUSY;
 224
 225                 ret = ttm_bo_wait_unreserved_locked(bo, interruptible);
 226                 if (unlikely(ret))
 227                         return ret;
 228         }
 229
 230         atomic_set(&bo->reserved, 1);
 231         if (use_sequence) {
 232                 /**
 233                  * Wake up waiters that may need to recheck for deadlock,
 234                  * if we decreased the sequence number.
 235                  */
 236                 if (unlikely((bo->val_seq - sequence < (1 << 31))
 237                              || !bo->seq_valid))
 238                         wakeup(bo);
 239
 240                 bo->val_seq = sequence;
 241                 bo->seq_valid = true;
 242         } else {
 243                 bo->seq_valid = false;
 244         }
 245
 246         return 0;
 247 }
 248
 249 void ttm_bo_list_ref_sub(struct ttm_buffer_object *bo, int count,
 250                          bool never_free)
 251 {
 252         u_int old;
 253
 254         old = atomic_fetchadd_int(&bo->list_kref, -count);
 255         if (old <= count) {
 256                 if (never_free)
 257                         panic("ttm_bo_ref_buf");
 258                 ttm_bo_release_list(bo);
 259         }
 260 }
 261
 262 int ttm_bo_reserve(struct ttm_buffer_object *bo,
 263                    bool interruptible,
 264                    bool no_wait, bool use_sequence, uint32_t sequence)
 265 {
 266         struct ttm_bo_global *glob = bo->glob;
 267         int put_count = 0;
 268         int ret;
 269
 270         lockmgr(&glob->lru_lock, LK_EXCLUSIVE);
 271         ret = ttm_bo_reserve_locked(bo, interruptible, no_wait, use_sequence,
 272                                     sequence);
 273         if (likely(ret == 0))
 274                 put_count = ttm_bo_del_from_lru(bo);
 275         lockmgr(&glob->lru_lock, LK_RELEASE);
 276
 277         ttm_bo_list_ref_sub(bo, put_count, true);
 278
 279         return ret;
 280 }
 281
 282 void ttm_bo_unreserve_locked(struct ttm_buffer_object *bo)
 283 {
 284         ttm_bo_add_to_lru(bo);
 285         atomic_set(&bo->reserved, 0);
 286         wakeup(bo);
 287 }
 288
 289 void ttm_bo_unreserve(struct ttm_buffer_object *bo)
 290 {
 291         struct ttm_bo_global *glob = bo->glob;
 292
 293         lockmgr(&glob->lru_lock, LK_EXCLUSIVE);
 294         ttm_bo_unreserve_locked(bo);
 295         lockmgr(&glob->lru_lock, LK_RELEASE);
 296 }
 297
 298 /*
 299  * Call bo->mutex locked.
 300  */
 301 static int ttm_bo_add_ttm(struct ttm_buffer_object *bo, bool zero_alloc)
 302 {
 303         struct ttm_bo_device *bdev = bo->bdev;
 304         struct ttm_bo_global *glob = bo->glob;
 305         int ret = 0;
 306         uint32_t page_flags = 0;
 307
 308         TTM_ASSERT_LOCKED(&bo->mutex);
 309         bo->ttm = NULL;
 310
 311         if (bdev->need_dma32)
 312                 page_flags |= TTM_PAGE_FLAG_DMA32;
 313
 314         switch (bo->type) {
 315         case ttm_bo_type_device:
 316                 if (zero_alloc)
 317                         page_flags |= TTM_PAGE_FLAG_ZERO_ALLOC;
 318         case ttm_bo_type_kernel:
 319                 bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
 320                                                       page_flags, glob->dummy_read_page);
 321                 if (unlikely(bo->ttm == NULL))
 322                         ret = -ENOMEM;
 323                 break;
 324         case ttm_bo_type_sg:
 325                 bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
 326                                                       page_flags | TTM_PAGE_FLAG_SG,
 327                                                       glob->dummy_read_page);
 328                 if (unlikely(bo->ttm == NULL)) {
 329                         ret = -ENOMEM;
 330                         break;
 331                 }
 332                 bo->ttm->sg = bo->sg;
 333                 break;
 334         default:
 335                 kprintf("[TTM] Illegal buffer object type\n");
 336                 ret = -EINVAL;
 337                 break;
 338         }
 339
 340         return ret;
 341 }
 342
 343 static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
 344                                   struct ttm_mem_reg *mem,
 345                                   bool evict, bool interruptible,
 346                                   bool no_wait_gpu)
 347 {
 348         struct ttm_bo_device *bdev = bo->bdev;
 349         bool old_is_pci = ttm_mem_reg_is_pci(bdev, &bo->mem);
 350         bool new_is_pci = ttm_mem_reg_is_pci(bdev, mem);
 351         struct ttm_mem_type_manager *old_man = &bdev->man[bo->mem.mem_type];
 352         struct ttm_mem_type_manager *new_man = &bdev->man[mem->mem_type];
 353         int ret = 0;
 354
 355         if (old_is_pci || new_is_pci ||
 356             ((mem->placement & bo->mem.placement & TTM_PL_MASK_CACHING) == 0)) {
 357                 ret = ttm_mem_io_lock(old_man, true);
 358                 if (unlikely(ret != 0))
 359                         goto out_err;
 360                 ttm_bo_unmap_virtual_locked(bo);
 361                 ttm_mem_io_unlock(old_man);
 362         }
 363
 364         /*
 365          * Create and bind a ttm if required.
 366          */
 367
 368         if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
 369                 if (bo->ttm == NULL) {
 370                         bool zero = !(old_man->flags & TTM_MEMTYPE_FLAG_FIXED);
 371                         ret = ttm_bo_add_ttm(bo, zero);
 372                         if (ret)
 373                                 goto out_err;
 374                 }
 375
 376                 ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement);
 377                 if (ret)
 378                         goto out_err;
 379
 380                 if (mem->mem_type != TTM_PL_SYSTEM) {
 381                         ret = ttm_tt_bind(bo->ttm, mem);
 382                         if (ret)
 383                                 goto out_err;
 384                 }
 385
 386                 if (bo->mem.mem_type == TTM_PL_SYSTEM) {
 387                         if (bdev->driver->move_notify)
 388                                 bdev->driver->move_notify(bo, mem);
 389                         bo->mem = *mem;
 390                         mem->mm_node = NULL;
 391                         goto moved;
 392                 }
 393         }
 394
 395         if (bdev->driver->move_notify)
 396                 bdev->driver->move_notify(bo, mem);
 397
 398         if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) &&
 399             !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED))
 400                 ret = ttm_bo_move_ttm(bo, evict, no_wait_gpu, mem);
 401         else if (bdev->driver->move)
 402                 ret = bdev->driver->move(bo, evict, interruptible,
 403                                          no_wait_gpu, mem);
 404         else
 405                 ret = ttm_bo_move_memcpy(bo, evict, no_wait_gpu, mem);
 406
 407         if (ret) {
 408                 if (bdev->driver->move_notify) {
 409                         struct ttm_mem_reg tmp_mem = *mem;
 410                         *mem = bo->mem;
 411                         bo->mem = tmp_mem;
 412                         bdev->driver->move_notify(bo, mem);
 413                         bo->mem = *mem;
 414                 }
 415
 416                 goto out_err;
 417         }
 418
 419 moved:
 420         if (bo->evicted) {
 421                 ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement);
 422                 if (ret)
 423                         kprintf("[TTM] Can not flush read caches\n");
 424                 bo->evicted = false;
 425         }
 426
 427         if (bo->mem.mm_node) {
 428                 bo->offset = (bo->mem.start << PAGE_SHIFT) +
 429                     bdev->man[bo->mem.mem_type].gpu_offset;
 430                 bo->cur_placement = bo->mem.placement;
 431         } else
 432                 bo->offset = 0;
 433
 434         return 0;
 435
 436 out_err:
 437         new_man = &bdev->man[bo->mem.mem_type];
 438         if ((new_man->flags & TTM_MEMTYPE_FLAG_FIXED) && bo->ttm) {
 439                 ttm_tt_unbind(bo->ttm);
 440                 ttm_tt_destroy(bo->ttm);
 441                 bo->ttm = NULL;
 442         }
 443
 444         return ret;
 445 }
 446
 447 /**
 448  * Call bo::reserved.
 449  * Will release GPU memory type usage on destruction.
 450  * This is the place to put in driver specific hooks to release
 451  * driver private resources.
 452  * Will release the bo::reserved lock.
 453  */
 454
 455 static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
 456 {
 457         if (bo->bdev->driver->move_notify)
 458                 bo->bdev->driver->move_notify(bo, NULL);
 459
 460         if (bo->ttm) {
 461                 ttm_tt_unbind(bo->ttm);
 462                 ttm_tt_destroy(bo->ttm);
 463                 bo->ttm = NULL;
 464         }
 465         ttm_bo_mem_put(bo, &bo->mem);
 466
 467         atomic_set(&bo->reserved, 0);
 468         wakeup(&bo);
 469
 470         /*
 471          * Since the final reference to this bo may not be dropped by
 472          * the current task we have to put a memory barrier here to make
 473          * sure the changes done in this function are always visible.
 474          *
 475          * This function only needs protection against the final kref_put.
 476          */
 477         cpu_mfence();
 478 }
 479
 480 static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo)
 481 {
 482         struct ttm_bo_device *bdev = bo->bdev;
 483         struct ttm_bo_global *glob = bo->glob;
 484         struct ttm_bo_driver *driver = bdev->driver;
 485         void *sync_obj = NULL;
 486         int put_count;
 487         int ret;
 488
 489         lockmgr(&glob->lru_lock, LK_EXCLUSIVE);
 490         ret = ttm_bo_reserve_locked(bo, false, true, false, 0);
 491
 492         lockmgr(&bdev->fence_lock, LK_EXCLUSIVE);
 493         (void) ttm_bo_wait(bo, false, false, true);
 494         if (!ret && !bo->sync_obj) {
 495                 lockmgr(&bdev->fence_lock, LK_RELEASE);
 496                 put_count = ttm_bo_del_from_lru(bo);
 497
 498                 lockmgr(&glob->lru_lock, LK_RELEASE);
 499                 ttm_bo_cleanup_memtype_use(bo);
 500
 501                 ttm_bo_list_ref_sub(bo, put_count, true);
 502
 503                 return;
 504         }
 505         if (bo->sync_obj)
 506                 sync_obj = driver->sync_obj_ref(bo->sync_obj);
 507         lockmgr(&bdev->fence_lock, LK_RELEASE);
 508
 509         if (!ret) {
 510                 atomic_set(&bo->reserved, 0);
 511                 wakeup(bo);
 512         }
 513
 514         refcount_acquire(&bo->list_kref);
 515         list_add_tail(&bo->ddestroy, &bdev->ddestroy);
 516         lockmgr(&glob->lru_lock, LK_RELEASE);
 517
 518         if (sync_obj) {
 519                 driver->sync_obj_flush(sync_obj);
 520                 driver->sync_obj_unref(&sync_obj);
 521         }
 522         taskqueue_enqueue_timeout(taskqueue_thread[mycpuid], &bdev->wq,
 523             ((hz / 100) < 1) ? 1 : hz / 100);
 524 }
 525
 526 /**
 527  * function ttm_bo_cleanup_refs_and_unlock
 528  * If bo idle, remove from delayed- and lru lists, and unref.
 529  * If not idle, do nothing.
 530  *
 531  * Must be called with lru_lock and reservation held, this function
 532  * will drop both before returning.
 533  *
 534  * @interruptible         Any sleeps should occur interruptibly.
 535  * @no_wait_gpu           Never wait for gpu. Return -EBUSY instead.
 536  */
 537
 538 static int ttm_bo_cleanup_refs_and_unlock(struct ttm_buffer_object *bo,
 539                                           bool interruptible,
 540                                           bool no_wait_gpu)
 541 {
 542         struct ttm_bo_device *bdev = bo->bdev;
 543         struct ttm_bo_driver *driver = bdev->driver;
 544         struct ttm_bo_global *glob = bo->glob;
 545         int put_count;
 546         int ret;
 547
 548         lockmgr(&bdev->fence_lock, LK_EXCLUSIVE);
 549         ret = ttm_bo_wait(bo, false, false, true);
 550
 551         if (ret && !no_wait_gpu) {
 552                 void *sync_obj;
 553
 554                 /*
 555                  * Take a reference to the fence and unreserve,
 556                  * at this point the buffer should be dead, so
 557                  * no new sync objects can be attached.
 558                  */
 559                 sync_obj = driver->sync_obj_ref(bo->sync_obj);
 560                 lockmgr(&bdev->fence_lock, LK_RELEASE);
 561
 562                 atomic_set(&bo->reserved, 0);
 563                 wakeup(bo);
 564                 lockmgr(&glob->lru_lock, LK_RELEASE);
 565
 566                 ret = driver->sync_obj_wait(sync_obj, false, interruptible);
 567                 driver->sync_obj_unref(&sync_obj);
 568                 if (ret)
 569                         return ret;
 570
 571                 /*
 572                  * remove sync_obj with ttm_bo_wait, the wait should be
 573                  * finished, and no new wait object should have been added.
 574                  */
 575                 lockmgr(&bdev->fence_lock, LK_EXCLUSIVE);
 576                 ret = ttm_bo_wait(bo, false, false, true);
 577                 lockmgr(&bdev->fence_lock, LK_RELEASE);
 578                 if (ret)
 579                         return ret;
 580
 581                 lockmgr(&glob->lru_lock, LK_EXCLUSIVE);
 582                 ret = ttm_bo_reserve_locked(bo, false, true, false, 0);
 583
 584                 /*
 585                  * We raced, and lost, someone else holds the reservation now,
 586                  * and is probably busy in ttm_bo_cleanup_memtype_use.
 587                  *
 588                  * Even if it's not the case, because we finished waiting any
 589                  * delayed destruction would succeed, so just return success
 590                  * here.
 591                  */
 592                 if (ret) {
 593                         lockmgr(&glob->lru_lock, LK_RELEASE);
 594                         return 0;
 595                 }
 596         } else
 597                 lockmgr(&bdev->fence_lock, LK_RELEASE);
 598
 599         if (ret || unlikely(list_empty(&bo->ddestroy))) {
 600                 atomic_set(&bo->reserved, 0);
 601                 wakeup(bo);
 602                 lockmgr(&glob->lru_lock, LK_RELEASE);
 603                 return ret;
 604         }
 605
 606         put_count = ttm_bo_del_from_lru(bo);
 607         list_del_init(&bo->ddestroy);
 608         ++put_count;
 609
 610         lockmgr(&glob->lru_lock, LK_RELEASE);
 611         ttm_bo_cleanup_memtype_use(bo);
 612
 613         ttm_bo_list_ref_sub(bo, put_count, true);
 614
 615         return 0;
 616 }
 617
 618 /**
 619  * Traverse the delayed list, and call ttm_bo_cleanup_refs on all
 620  * encountered buffers.
 621  */
 622
 623 static int ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all)
 624 {
 625         struct ttm_bo_global *glob = bdev->glob;
 626         struct ttm_buffer_object *entry = NULL;
 627         int ret = 0;
 628
 629         lockmgr(&glob->lru_lock, LK_EXCLUSIVE);
 630         if (list_empty(&bdev->ddestroy))
 631                 goto out_unlock;
 632
 633         entry = list_first_entry(&bdev->ddestroy,
 634                 struct ttm_buffer_object, ddestroy);
 635         refcount_acquire(&entry->list_kref);
 636
 637         for (;;) {
 638                 struct ttm_buffer_object *nentry = NULL;
 639
 640                 if (entry->ddestroy.next != &bdev->ddestroy) {
 641                         nentry = list_first_entry(&entry->ddestroy,
 642                                 struct ttm_buffer_object, ddestroy);
 643                         refcount_acquire(&nentry->list_kref);
 644                 }
 645
 646                 ret = ttm_bo_reserve_locked(entry, false, !remove_all, false, 0);
 647                 if (!ret)
 648                         ret = ttm_bo_cleanup_refs_and_unlock(entry, false,
 649                                                              !remove_all);
 650                 else
 651                         lockmgr(&glob->lru_lock, LK_RELEASE);
 652
 653                 if (refcount_release(&entry->list_kref))
 654                         ttm_bo_release_list(entry);
 655                 entry = nentry;
 656
 657                 if (ret || !entry)
 658                         goto out;
 659
 660                 lockmgr(&glob->lru_lock, LK_EXCLUSIVE);
 661                 if (list_empty(&entry->ddestroy))
 662                         break;
 663         }
 664
 665 out_unlock:
 666         lockmgr(&glob->lru_lock, LK_RELEASE);
 667 out:
 668         if (entry && refcount_release(&entry->list_kref))
 669                 ttm_bo_release_list(entry);
 670         return ret;
 671 }
 672
 673 static void ttm_bo_delayed_workqueue(void *arg, int pending __unused)
 674 {
 675         struct ttm_bo_device *bdev = arg;
 676
 677         if (ttm_bo_delayed_delete(bdev, false)) {
 678                 taskqueue_enqueue_timeout(taskqueue_thread[mycpuid], &bdev->wq,
 679                     ((hz / 100) < 1) ? 1 : hz / 100);
 680         }
 681 }
 682
 683 static void ttm_bo_release(struct ttm_buffer_object *bo)
 684 {
 685         struct ttm_bo_device *bdev = bo->bdev;
 686         struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
 687
 688         lockmgr(&bdev->vm_lock, LK_EXCLUSIVE);
 689         if (likely(bo->vm_node != NULL)) {
 690                 RB_REMOVE(ttm_bo_device_buffer_objects,
 691                     &bdev->addr_space_rb, bo);
 692                 drm_mm_put_block(bo->vm_node);
 693                 bo->vm_node = NULL;
 694         }
 695         lockmgr(&bdev->vm_lock, LK_RELEASE);
 696         ttm_mem_io_lock(man, false);
 697         ttm_mem_io_free_vm(bo);
 698         ttm_mem_io_unlock(man);
 699         ttm_bo_cleanup_refs_or_queue(bo);
 700         if (refcount_release(&bo->list_kref))
 701                 ttm_bo_release_list(bo);
 702 }
 703
 704 void ttm_bo_unref(struct ttm_buffer_object **p_bo)
 705 {
 706         struct ttm_buffer_object *bo = *p_bo;
 707
 708         *p_bo = NULL;
 709         if (refcount_release(&bo->kref))
 710                 ttm_bo_release(bo);
 711 }
 712
 713 int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev)
 714 {
 715         int pending;
 716
 717         taskqueue_cancel_timeout(taskqueue_thread[mycpuid], &bdev->wq, &pending);
 718         if (pending)
 719                 taskqueue_drain_timeout(taskqueue_thread[mycpuid], &bdev->wq);
 720         return (pending);
 721 }
 722
 723 void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched)
 724 {
 725         if (resched) {
 726                 taskqueue_enqueue_timeout(taskqueue_thread[mycpuid], &bdev->wq,
 727                     ((hz / 100) < 1) ? 1 : hz / 100);
 728         }
 729 }
 730
 731 static int ttm_bo_evict(struct ttm_buffer_object *bo, bool interruptible,
 732                         bool no_wait_gpu)
 733 {
 734         struct ttm_bo_device *bdev = bo->bdev;
 735         struct ttm_mem_reg evict_mem;
 736         struct ttm_placement placement;
 737         int ret = 0;
 738
 739         lockmgr(&bdev->fence_lock, LK_EXCLUSIVE);
 740         ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
 741         lockmgr(&bdev->fence_lock, LK_RELEASE);
 742
 743         if (unlikely(ret != 0)) {
 744                 if (ret != -ERESTART) {
 745                         kprintf("[TTM] Failed to expire sync object before buffer eviction\n");
 746                 }
 747                 goto out;
 748         }
 749
 750         KKASSERT(ttm_bo_is_reserved(bo));
 751
 752         evict_mem = bo->mem;
 753         evict_mem.mm_node = NULL;
 754         evict_mem.bus.io_reserved_vm = false;
 755         evict_mem.bus.io_reserved_count = 0;
 756
 757         placement.fpfn = 0;
 758         placement.lpfn = 0;
 759         placement.num_placement = 0;
 760         placement.num_busy_placement = 0;
 761         bdev->driver->evict_flags(bo, &placement);
 762         ret = ttm_bo_mem_space(bo, &placement, &evict_mem, interruptible,
 763                                 no_wait_gpu);
 764         if (ret) {
 765                 if (ret != -ERESTART) {
 766                         kprintf("[TTM] Failed to find memory space for buffer 0x%p eviction\n",
 767                                bo);
 768                         ttm_bo_mem_space_debug(bo, &placement);
 769                 }
 770                 goto out;
 771         }
 772
 773         ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, interruptible,
 774                                      no_wait_gpu);
 775         if (ret) {
 776                 if (ret != -ERESTART)
 777                         kprintf("[TTM] Buffer eviction failed\n");
 778                 ttm_bo_mem_put(bo, &evict_mem);
 779                 goto out;
 780         }
 781         bo->evicted = true;
 782 out:
 783         return ret;
 784 }
 785
 786 static int ttm_mem_evict_first(struct ttm_bo_device *bdev,
 787                                 uint32_t mem_type,
 788                                 bool interruptible,
 789                                 bool no_wait_gpu)
 790 {
 791         struct ttm_bo_global *glob = bdev->glob;
 792         struct ttm_mem_type_manager *man = &bdev->man[mem_type];
 793         struct ttm_buffer_object *bo;
 794         int ret = -EBUSY, put_count;
 795
 796         lockmgr(&glob->lru_lock, LK_EXCLUSIVE);
 797         list_for_each_entry(bo, &man->lru, lru) {
 798                 ret = ttm_bo_reserve_locked(bo, false, true, false, 0);
 799                 if (!ret)
 800                         break;
 801         }
 802
 803         if (ret) {
 804                 lockmgr(&glob->lru_lock, LK_RELEASE);
 805                 return ret;
 806         }
 807
 808         refcount_acquire(&bo->list_kref);
 809
 810         if (!list_empty(&bo->ddestroy)) {
 811                 ret = ttm_bo_cleanup_refs_and_unlock(bo, interruptible,
 812                                                      no_wait_gpu);
 813                 if (refcount_release(&bo->list_kref))
 814                         ttm_bo_release_list(bo);
 815                 return ret;
 816         }
 817
 818         put_count = ttm_bo_del_from_lru(bo);
 819         lockmgr(&glob->lru_lock, LK_RELEASE);
 820
 821         KKASSERT(ret == 0);
 822
 823         ttm_bo_list_ref_sub(bo, put_count, true);
 824
 825         ret = ttm_bo_evict(bo, interruptible, no_wait_gpu);
 826         ttm_bo_unreserve(bo);
 827
 828         if (refcount_release(&bo->list_kref))
 829                 ttm_bo_release_list(bo);
 830         return ret;
 831 }
 832
 833 void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem)
 834 {
 835         struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type];
 836
 837         if (mem->mm_node)
 838                 (*man->func->put_node)(man, mem);
 839 }
 840
 841 /**
 842  * Repeatedly evict memory from the LRU for @mem_type until we create enough
 843  * space, or we've evicted everything and there isn't enough space.
 844  */
 845 static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
 846                                         uint32_t mem_type,
 847                                         struct ttm_placement *placement,
 848                                         struct ttm_mem_reg *mem,
 849                                         bool interruptible,
 850                                         bool no_wait_gpu)
 851 {
 852         struct ttm_bo_device *bdev = bo->bdev;
 853         struct ttm_mem_type_manager *man = &bdev->man[mem_type];
 854         int ret;
 855
 856         do {
 857                 ret = (*man->func->get_node)(man, bo, placement, mem);
 858                 if (unlikely(ret != 0))
 859                         return ret;
 860                 if (mem->mm_node)
 861                         break;
 862                 ret = ttm_mem_evict_first(bdev, mem_type,
 863                                           interruptible, no_wait_gpu);
 864                 if (unlikely(ret != 0))
 865                         return ret;
 866         } while (1);
 867         if (mem->mm_node == NULL)
 868                 return -ENOMEM;
 869         mem->mem_type = mem_type;
 870         return 0;
 871 }
 872
 873 static uint32_t ttm_bo_select_caching(struct ttm_mem_type_manager *man,
 874                                       uint32_t cur_placement,
 875                                       uint32_t proposed_placement)
 876 {
 877         uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING;
 878         uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING;
 879
 880         /**
 881          * Keep current caching if possible.
 882          */
 883
 884         if ((cur_placement & caching) != 0)
 885                 result |= (cur_placement & caching);
 886         else if ((man->default_caching & caching) != 0)
 887                 result |= man->default_caching;
 888         else if ((TTM_PL_FLAG_CACHED & caching) != 0)
 889                 result |= TTM_PL_FLAG_CACHED;
 890         else if ((TTM_PL_FLAG_WC & caching) != 0)
 891                 result |= TTM_PL_FLAG_WC;
 892         else if ((TTM_PL_FLAG_UNCACHED & caching) != 0)
 893                 result |= TTM_PL_FLAG_UNCACHED;
 894
 895         return result;
 896 }
 897
 898 static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man,
 899                                  uint32_t mem_type,
 900                                  uint32_t proposed_placement,
 901                                  uint32_t *masked_placement)
 902 {
 903         uint32_t cur_flags = ttm_bo_type_flags(mem_type);
 904
 905         if ((cur_flags & proposed_placement & TTM_PL_MASK_MEM) == 0)
 906                 return false;
 907
 908         if ((proposed_placement & man->available_caching) == 0)
 909                 return false;
 910
 911         cur_flags |= (proposed_placement & man->available_caching);
 912
 913         *masked_placement = cur_flags;
 914         return true;
 915 }
 916
 917 /**
 918  * Creates space for memory region @mem according to its type.
 919  *
 920  * This function first searches for free space in compatible memory types in
 921  * the priority order defined by the driver.  If free space isn't found, then
 922  * ttm_bo_mem_force_space is attempted in priority order to evict and find
 923  * space.
 924  */
 925 int ttm_bo_mem_space(struct ttm_buffer_object *bo,
 926                         struct ttm_placement *placement,
 927                         struct ttm_mem_reg *mem,
 928                         bool interruptible,
 929                         bool no_wait_gpu)
 930 {
 931         struct ttm_bo_device *bdev = bo->bdev;
 932         struct ttm_mem_type_manager *man;
 933         uint32_t mem_type = TTM_PL_SYSTEM;
 934         uint32_t cur_flags = 0;
 935         bool type_found = false;
 936         bool type_ok = false;
 937         bool has_erestartsys = false;
 938         int i, ret;
 939
 940         mem->mm_node = NULL;
 941         for (i = 0; i < placement->num_placement; ++i) {
 942                 ret = ttm_mem_type_from_flags(placement->placement[i],
 943                                                 &mem_type);
 944                 if (ret)
 945                         return ret;
 946                 man = &bdev->man[mem_type];
 947
 948                 type_ok = ttm_bo_mt_compatible(man,
 949                                                 mem_type,
 950                                                 placement->placement[i],
 951                                                 &cur_flags);
 952
 953                 if (!type_ok)
 954                         continue;
 955
 956                 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
 957                                                   cur_flags);
 958                 /*
 959                  * Use the access and other non-mapping-related flag bits from
 960                  * the memory placement flags to the current flags
 961                  */
 962                 ttm_flag_masked(&cur_flags, placement->placement[i],
 963                                 ~TTM_PL_MASK_MEMTYPE);
 964
 965                 if (mem_type == TTM_PL_SYSTEM)
 966                         break;
 967
 968                 if (man->has_type && man->use_type) {
 969                         type_found = true;
 970                         ret = (*man->func->get_node)(man, bo, placement, mem);
 971                         if (unlikely(ret))
 972                                 return ret;
 973                 }
 974                 if (mem->mm_node)
 975                         break;
 976         }
 977
 978         if ((type_ok && (mem_type == TTM_PL_SYSTEM)) || mem->mm_node) {
 979                 mem->mem_type = mem_type;
 980                 mem->placement = cur_flags;
 981                 return 0;
 982         }
 983
 984         if (!type_found)
 985                 return -EINVAL;
 986
 987         for (i = 0; i < placement->num_busy_placement; ++i) {
 988                 ret = ttm_mem_type_from_flags(placement->busy_placement[i],
 989                                                 &mem_type);
 990                 if (ret)
 991                         return ret;
 992                 man = &bdev->man[mem_type];
 993                 if (!man->has_type)
 994                         continue;
 995                 if (!ttm_bo_mt_compatible(man,
 996                                                 mem_type,
 997                                                 placement->busy_placement[i],
 998                                                 &cur_flags))
 999                         continue;
1000
1001                 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
1002                                                   cur_flags);
1003                 /*
1004                  * Use the access and other non-mapping-related flag bits from
1005                  * the memory placement flags to the current flags
1006                  */
1007                 ttm_flag_masked(&cur_flags, placement->busy_placement[i],
1008                                 ~TTM_PL_MASK_MEMTYPE);
1009
1010
1011                 if (mem_type == TTM_PL_SYSTEM) {
1012                         mem->mem_type = mem_type;
1013                         mem->placement = cur_flags;
1014                         mem->mm_node = NULL;
1015                         return 0;
1016                 }
1017
1018                 ret = ttm_bo_mem_force_space(bo, mem_type, placement, mem,
1019                                                 interruptible, no_wait_gpu);
1020                 if (ret == 0 && mem->mm_node) {
1021                         mem->placement = cur_flags;
1022                         return 0;
1023                 }
1024                 if (ret == -ERESTART)
1025                         has_erestartsys = true;
1026         }
1027         ret = (has_erestartsys) ? -ERESTART : -ENOMEM;
1028         return ret;
1029 }
1030
1031 static
1032 int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
1033                         struct ttm_placement *placement,
1034                         bool interruptible,
1035                         bool no_wait_gpu)
1036 {
1037         int ret = 0;
1038         struct ttm_mem_reg mem;
1039         struct ttm_bo_device *bdev = bo->bdev;
1040
1041         KKASSERT(ttm_bo_is_reserved(bo));
1042
1043         /*
1044          * FIXME: It's possible to pipeline buffer moves.
1045          * Have the driver move function wait for idle when necessary,
1046          * instead of doing it here.
1047          */
1048         lockmgr(&bdev->fence_lock, LK_EXCLUSIVE);
1049         ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
1050         lockmgr(&bdev->fence_lock, LK_RELEASE);
1051         if (ret)
1052                 return ret;
1053         mem.num_pages = bo->num_pages;
1054         mem.size = mem.num_pages << PAGE_SHIFT;
1055         mem.page_alignment = bo->mem.page_alignment;
1056         mem.bus.io_reserved_vm = false;
1057         mem.bus.io_reserved_count = 0;
1058         /*
1059          * Determine where to move the buffer.
1060          */
1061         ret = ttm_bo_mem_space(bo, placement, &mem,
1062                                interruptible, no_wait_gpu);
1063         if (ret)
1064                 goto out_unlock;
1065         ret = ttm_bo_handle_move_mem(bo, &mem, false,
1066                                      interruptible, no_wait_gpu);
1067 out_unlock:
1068         if (ret && mem.mm_node)
1069                 ttm_bo_mem_put(bo, &mem);
1070         return ret;
1071 }
1072
1073 static int ttm_bo_mem_compat(struct ttm_placement *placement,
1074                              struct ttm_mem_reg *mem)
1075 {
1076         int i;
1077
1078         if (mem->mm_node && placement->lpfn != 0 &&
1079             (mem->start < placement->fpfn ||
1080              mem->start + mem->num_pages > placement->lpfn))
1081                 return -1;
1082
1083         for (i = 0; i < placement->num_placement; i++) {
1084                 if ((placement->placement[i] & mem->placement &
1085                         TTM_PL_MASK_CACHING) &&
1086                         (placement->placement[i] & mem->placement &
1087                         TTM_PL_MASK_MEM))
1088                         return i;
1089         }
1090         return -1;
1091 }
1092
1093 int ttm_bo_validate(struct ttm_buffer_object *bo,
1094                         struct ttm_placement *placement,
1095                         bool interruptible,
1096                         bool no_wait_gpu)
1097 {
1098         int ret;
1099
1100         KKASSERT(ttm_bo_is_reserved(bo));
1101         /* Check that range is valid */
1102         if (placement->lpfn || placement->fpfn)
1103                 if (placement->fpfn > placement->lpfn ||
1104                         (placement->lpfn - placement->fpfn) < bo->num_pages)
1105                         return -EINVAL;
1106         /*
1107          * Check whether we need to move buffer.
1108          */
1109         ret = ttm_bo_mem_compat(placement, &bo->mem);
1110         if (ret < 0) {
1111                 ret = ttm_bo_move_buffer(bo, placement, interruptible,
1112                                          no_wait_gpu);
1113                 if (ret)
1114                         return ret;
1115         } else {
1116                 /*
1117                  * Use the access and other non-mapping-related flag bits from
1118                  * the compatible memory placement flags to the active flags
1119                  */
1120                 ttm_flag_masked(&bo->mem.placement, placement->placement[ret],
1121                                 ~TTM_PL_MASK_MEMTYPE);
1122         }
1123         /*
1124          * We might need to add a TTM.
1125          */
1126         if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
1127                 ret = ttm_bo_add_ttm(bo, true);
1128                 if (ret)
1129                         return ret;
1130         }
1131         return 0;
1132 }
1133
1134 int ttm_bo_check_placement(struct ttm_buffer_object *bo,
1135                                 struct ttm_placement *placement)
1136 {
1137         KKASSERT(!((placement->fpfn || placement->lpfn) &&
1138             (bo->mem.num_pages > (placement->lpfn - placement->fpfn))));
1139
1140         return 0;
1141 }
1142
1143 int ttm_bo_init(struct ttm_bo_device *bdev,
1144                 struct ttm_buffer_object *bo,
1145                 unsigned long size,
1146                 enum ttm_bo_type type,
1147                 struct ttm_placement *placement,
1148                 uint32_t page_alignment,
1149                 bool interruptible,
1150                 struct vm_object *persistent_swap_storage,
1151                 size_t acc_size,
1152                 struct sg_table *sg,
1153                 void (*destroy) (struct ttm_buffer_object *))
1154 {
1155         int ret = 0;
1156         unsigned long num_pages;
1157         struct ttm_mem_global *mem_glob = bdev->glob->mem_glob;
1158
1159         ret = ttm_mem_global_alloc(mem_glob, acc_size, false, false);
1160         if (ret) {
1161                 kprintf("[TTM] Out of kernel memory\n");
1162                 if (destroy)
1163                         (*destroy)(bo);
1164                 else
1165                         drm_free(bo, M_TTM_BO);
1166                 return -ENOMEM;
1167         }
1168
1169         num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1170         if (num_pages == 0) {
1171                 kprintf("[TTM] Illegal buffer object size\n");
1172                 if (destroy)
1173                         (*destroy)(bo);
1174                 else
1175                         drm_free(bo, M_TTM_BO);
1176                 ttm_mem_global_free(mem_glob, acc_size);
1177                 return -EINVAL;
1178         }
1179         bo->destroy = destroy;
1180
1181         refcount_init(&bo->kref, 1);
1182         refcount_init(&bo->list_kref, 1);
1183         atomic_set(&bo->cpu_writers, 0);
1184         atomic_set(&bo->reserved, 1);
1185         INIT_LIST_HEAD(&bo->lru);
1186         INIT_LIST_HEAD(&bo->ddestroy);
1187         INIT_LIST_HEAD(&bo->swap);
1188         INIT_LIST_HEAD(&bo->io_reserve_lru);
1189         bo->bdev = bdev;
1190         bo->glob = bdev->glob;
1191         bo->type = type;
1192         bo->num_pages = num_pages;
1193         bo->mem.size = num_pages << PAGE_SHIFT;
1194         bo->mem.mem_type = TTM_PL_SYSTEM;
1195         bo->mem.num_pages = bo->num_pages;
1196         bo->mem.mm_node = NULL;
1197         bo->mem.page_alignment = page_alignment;
1198         bo->mem.bus.io_reserved_vm = false;
1199         bo->mem.bus.io_reserved_count = 0;
1200         bo->priv_flags = 0;
1201         bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED);
1202         bo->seq_valid = false;
1203         bo->persistent_swap_storage = persistent_swap_storage;
1204         bo->acc_size = acc_size;
1205         bo->sg = sg;
1206         atomic_inc(&bo->glob->bo_count);
1207
1208         ret = ttm_bo_check_placement(bo, placement);
1209         if (unlikely(ret != 0))
1210                 goto out_err;
1211
1212         /*
1213          * For ttm_bo_type_device buffers, allocate
1214          * address space from the device.
1215          */
1216         if (bo->type == ttm_bo_type_device ||
1217             bo->type == ttm_bo_type_sg) {
1218                 ret = ttm_bo_setup_vm(bo);
1219                 if (ret)
1220                         goto out_err;
1221         }
1222
1223         ret = ttm_bo_validate(bo, placement, interruptible, false);
1224         if (ret)
1225                 goto out_err;
1226
1227         ttm_bo_unreserve(bo);
1228         return 0;
1229
1230 out_err:
1231         ttm_bo_unreserve(bo);
1232         ttm_bo_unref(&bo);
1233
1234         return ret;
1235 }
1236
1237 size_t ttm_bo_acc_size(struct ttm_bo_device *bdev,
1238                        unsigned long bo_size,
1239                        unsigned struct_size)
1240 {
1241         unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1242         size_t size = 0;
1243
1244         size += ttm_round_pot(struct_size);
1245         size += PAGE_ALIGN(npages * sizeof(void *));
1246         size += ttm_round_pot(sizeof(struct ttm_tt));
1247         return size;
1248 }
1249
1250 size_t ttm_bo_dma_acc_size(struct ttm_bo_device *bdev,
1251                            unsigned long bo_size,
1252                            unsigned struct_size)
1253 {
1254         unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1255         size_t size = 0;
1256
1257         size += ttm_round_pot(struct_size);
1258         size += PAGE_ALIGN(npages * sizeof(void *));
1259         size += PAGE_ALIGN(npages * sizeof(dma_addr_t));
1260         size += ttm_round_pot(sizeof(struct ttm_dma_tt));
1261         return size;
1262 }
1263
1264 int ttm_bo_create(struct ttm_bo_device *bdev,
1265                         unsigned long size,
1266                         enum ttm_bo_type type,
1267                         struct ttm_placement *placement,
1268                         uint32_t page_alignment,
1269                         bool interruptible,
1270                         struct vm_object *persistent_swap_storage,
1271                         struct ttm_buffer_object **p_bo)
1272 {
1273         struct ttm_buffer_object *bo;
1274         size_t acc_size;
1275         int ret;
1276
1277         bo = kmalloc(sizeof(*bo), M_TTM_BO, M_WAITOK | M_ZERO);
1278         acc_size = ttm_bo_acc_size(bdev, size, sizeof(struct ttm_buffer_object));
1279         ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment,
1280                           interruptible, persistent_swap_storage, acc_size,
1281                           NULL, NULL);
1282         if (likely(ret == 0))
1283                 *p_bo = bo;
1284
1285         return ret;
1286 }
1287
1288 static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev,
1289                                         unsigned mem_type, bool allow_errors)
1290 {
1291         struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1292         struct ttm_bo_global *glob = bdev->glob;
1293         int ret;
1294
1295         /*
1296          * Can't use standard list traversal since we're unlocking.
1297          */
1298
1299         lockmgr(&glob->lru_lock, LK_EXCLUSIVE);
1300         while (!list_empty(&man->lru)) {
1301                 lockmgr(&glob->lru_lock, LK_RELEASE);
1302                 ret = ttm_mem_evict_first(bdev, mem_type, false, false);
1303                 if (ret) {
1304                         if (allow_errors) {
1305                                 return ret;
1306                         } else {
1307                                 kprintf("[TTM] Cleanup eviction failed\n");
1308                         }
1309                 }
1310                 lockmgr(&glob->lru_lock, LK_EXCLUSIVE);
1311         }
1312         lockmgr(&glob->lru_lock, LK_RELEASE);
1313         return 0;
1314 }
1315
1316 int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1317 {
1318         struct ttm_mem_type_manager *man;
1319         int ret = -EINVAL;
1320
1321         if (mem_type >= TTM_NUM_MEM_TYPES) {
1322                 kprintf("[TTM] Illegal memory type %d\n", mem_type);
1323                 return ret;
1324         }
1325         man = &bdev->man[mem_type];
1326
1327         if (!man->has_type) {
1328                 kprintf("[TTM] Trying to take down uninitialized memory manager type %u\n",
1329                        mem_type);
1330                 return ret;
1331         }
1332
1333         man->use_type = false;
1334         man->has_type = false;
1335
1336         ret = 0;
1337         if (mem_type > 0) {
1338                 ttm_bo_force_list_clean(bdev, mem_type, false);
1339
1340                 ret = (*man->func->takedown)(man);
1341         }
1342
1343         return ret;
1344 }
1345
1346 int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1347 {
1348         struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1349
1350         if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) {
1351                 kprintf("[TTM] Illegal memory manager memory type %u\n", mem_type);
1352                 return -EINVAL;
1353         }
1354
1355         if (!man->has_type) {
1356                 kprintf("[TTM] Memory type %u has not been initialized\n", mem_type);
1357                 return 0;
1358         }
1359
1360         return ttm_bo_force_list_clean(bdev, mem_type, true);
1361 }
1362
1363 int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type,
1364                         unsigned long p_size)
1365 {
1366         int ret = -EINVAL;
1367         struct ttm_mem_type_manager *man;
1368
1369         KKASSERT(type < TTM_NUM_MEM_TYPES);
1370         man = &bdev->man[type];
1371         KKASSERT(!man->has_type);
1372         man->io_reserve_fastpath = true;
1373         man->use_io_reserve_lru = false;
1374         lockinit(&man->io_reserve_mutex, "ttmman", 0, LK_CANRECURSE);
1375         INIT_LIST_HEAD(&man->io_reserve_lru);
1376
1377         ret = bdev->driver->init_mem_type(bdev, type, man);
1378         if (ret)
1379                 return ret;
1380         man->bdev = bdev;
1381
1382         ret = 0;
1383         if (type != TTM_PL_SYSTEM) {
1384                 ret = (*man->func->init)(man, p_size);
1385                 if (ret)
1386                         return ret;
1387         }
1388         man->has_type = true;
1389         man->use_type = true;
1390         man->size = p_size;
1391
1392         INIT_LIST_HEAD(&man->lru);
1393
1394         return 0;
1395 }
1396
1397 static void ttm_bo_global_kobj_release(struct ttm_bo_global *glob)
1398 {
1399
1400         ttm_mem_unregister_shrink(glob->mem_glob, &glob->shrink);
1401         vm_page_free(glob->dummy_read_page);
1402 }
1403
1404 void ttm_bo_global_release(struct drm_global_reference *ref)
1405 {
1406         struct ttm_bo_global *glob = ref->object;
1407
1408         if (refcount_release(&glob->kobj_ref))
1409                 ttm_bo_global_kobj_release(glob);
1410 }
1411
1412 int ttm_bo_global_init(struct drm_global_reference *ref)
1413 {
1414         struct ttm_bo_global_ref *bo_ref =
1415                 container_of(ref, struct ttm_bo_global_ref, ref);
1416         struct ttm_bo_global *glob = ref->object;
1417         int ret;
1418
1419         lockinit(&glob->device_list_mutex, "ttmdlm", 0, LK_CANRECURSE);
1420         lockinit(&glob->lru_lock, "ttmlru", 0, LK_CANRECURSE);
1421         glob->mem_glob = bo_ref->mem_glob;
1422         glob->dummy_read_page = vm_page_alloc_contig(
1423             0, VM_MAX_ADDRESS, PAGE_SIZE, 0, 1*PAGE_SIZE, VM_MEMATTR_UNCACHEABLE);
1424
1425         if (unlikely(glob->dummy_read_page == NULL)) {
1426                 ret = -ENOMEM;
1427                 goto out_no_drp;
1428         }
1429
1430         INIT_LIST_HEAD(&glob->swap_lru);
1431         INIT_LIST_HEAD(&glob->device_list);
1432
1433         ttm_mem_init_shrink(&glob->shrink, ttm_bo_swapout);
1434         ret = ttm_mem_register_shrink(glob->mem_glob, &glob->shrink);
1435         if (unlikely(ret != 0)) {
1436                 kprintf("[TTM] Could not register buffer object swapout\n");
1437                 goto out_no_shrink;
1438         }
1439
1440         atomic_set(&glob->bo_count, 0);
1441
1442         refcount_init(&glob->kobj_ref, 1);
1443         return (0);
1444
1445 out_no_shrink:
1446         vm_page_free(glob->dummy_read_page);
1447 out_no_drp:
1448         drm_free(glob, M_DRM_GLOBAL);
1449         return ret;
1450 }
1451
1452 int ttm_bo_device_release(struct ttm_bo_device *bdev)
1453 {
1454         int ret = 0;
1455         unsigned i = TTM_NUM_MEM_TYPES;
1456         struct ttm_mem_type_manager *man;
1457         struct ttm_bo_global *glob = bdev->glob;
1458
1459         while (i--) {
1460                 man = &bdev->man[i];
1461                 if (man->has_type) {
1462                         man->use_type = false;
1463                         if ((i != TTM_PL_SYSTEM) && ttm_bo_clean_mm(bdev, i)) {
1464                                 ret = -EBUSY;
1465                                 kprintf("[TTM] DRM memory manager type %d is not clean\n",
1466                                        i);
1467                         }
1468                         man->has_type = false;
1469                 }
1470         }
1471
1472         lockmgr(&glob->device_list_mutex, LK_EXCLUSIVE);
1473         list_del(&bdev->device_list);
1474         lockmgr(&glob->device_list_mutex, LK_RELEASE);
1475
1476         if (taskqueue_cancel_timeout(taskqueue_thread[mycpuid], &bdev->wq, NULL))
1477                 taskqueue_drain_timeout(taskqueue_thread[mycpuid], &bdev->wq);
1478
1479         while (ttm_bo_delayed_delete(bdev, true))
1480                 ;
1481
1482         lockmgr(&glob->lru_lock, LK_EXCLUSIVE);
1483         if (list_empty(&bdev->ddestroy))
1484                 TTM_DEBUG("Delayed destroy list was clean\n");
1485
1486         if (list_empty(&bdev->man[0].lru))
1487                 TTM_DEBUG("Swap list was clean\n");
1488         lockmgr(&glob->lru_lock, LK_RELEASE);
1489
1490         KKASSERT(drm_mm_clean(&bdev->addr_space_mm));
1491         lockmgr(&bdev->vm_lock, LK_EXCLUSIVE);
1492         drm_mm_takedown(&bdev->addr_space_mm);
1493         lockmgr(&bdev->vm_lock, LK_RELEASE);
1494
1495         return ret;
1496 }
1497
1498 int ttm_bo_device_init(struct ttm_bo_device *bdev,
1499                        struct ttm_bo_global *glob,
1500                        struct ttm_bo_driver *driver,
1501                        uint64_t file_page_offset,
1502                        bool need_dma32)
1503 {
1504         int ret = -EINVAL;
1505
1506         lockinit(&bdev->vm_lock, "ttmvml", 0, LK_CANRECURSE);
1507         bdev->driver = driver;
1508
1509         memset(bdev->man, 0, sizeof(bdev->man));
1510
1511         /*
1512          * Initialize the system memory buffer type.
1513          * Other types need to be driver / IOCTL initialized.
1514          */
1515         ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0);
1516         if (unlikely(ret != 0))
1517                 goto out_no_sys;
1518
1519         RB_INIT(&bdev->addr_space_rb);
1520         ret = drm_mm_init(&bdev->addr_space_mm, file_page_offset, 0x10000000);
1521         if (unlikely(ret != 0))
1522                 goto out_no_addr_mm;
1523
1524         TIMEOUT_TASK_INIT(taskqueue_thread[mycpuid], &bdev->wq, 0,
1525             ttm_bo_delayed_workqueue, bdev);
1526         INIT_LIST_HEAD(&bdev->ddestroy);
1527         bdev->dev_mapping = NULL;
1528         bdev->glob = glob;
1529         bdev->need_dma32 = need_dma32;
1530         bdev->val_seq = 0;
1531         lockinit(&bdev->fence_lock, "ttmfence", 0, LK_CANRECURSE);
1532         lockmgr(&glob->device_list_mutex, LK_EXCLUSIVE);
1533         list_add_tail(&bdev->device_list, &glob->device_list);
1534         lockmgr(&glob->device_list_mutex, LK_RELEASE);
1535
1536         return 0;
1537 out_no_addr_mm:
1538         ttm_bo_clean_mm(bdev, 0);
1539 out_no_sys:
1540         return ret;
1541 }
1542
1543 /*
1544  * buffer object vm functions.
1545  */
1546
1547 bool ttm_mem_reg_is_pci(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
1548 {
1549         struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
1550
1551         if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
1552                 if (mem->mem_type == TTM_PL_SYSTEM)
1553                         return false;
1554
1555                 if (man->flags & TTM_MEMTYPE_FLAG_CMA)
1556                         return false;
1557
1558                 if (mem->placement & TTM_PL_FLAG_CACHED)
1559                         return false;
1560         }
1561         return true;
1562 }
1563
1564 void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo)
1565 {
1566         struct ttm_bo_device *bdev = bo->bdev;
1567         /* off_t offset = (off_t)bo->addr_space_offset;XXXKIB */
1568         /* off_t holelen = ((off_t)bo->mem.num_pages) << PAGE_SHIFT;XXXKIB */
1569
1570         if (!bdev->dev_mapping)
1571                 return;
1572         /* unmap_mapping_range(bdev->dev_mapping, offset, holelen, 1); XXXKIB */
1573         ttm_mem_io_free_vm(bo);
1574 }
1575
1576 void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
1577 {
1578         struct ttm_bo_device *bdev = bo->bdev;
1579         struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
1580
1581         ttm_mem_io_lock(man, false);
1582         ttm_bo_unmap_virtual_locked(bo);
1583         ttm_mem_io_unlock(man);
1584 }
1585
1586 static void ttm_bo_vm_insert_rb(struct ttm_buffer_object *bo)
1587 {
1588         struct ttm_bo_device *bdev = bo->bdev;
1589
1590         /* The caller acquired bdev->vm_lock. */
1591         RB_INSERT(ttm_bo_device_buffer_objects, &bdev->addr_space_rb, bo);
1592 }
1593
1594 /**
1595  * ttm_bo_setup_vm:
1596  *
1597  * @bo: the buffer to allocate address space for
1598  *
1599  * Allocate address space in the drm device so that applications
1600  * can mmap the buffer and access the contents. This only
1601  * applies to ttm_bo_type_device objects as others are not
1602  * placed in the drm device address space.
1603  */
1604
1605 static int ttm_bo_setup_vm(struct ttm_buffer_object *bo)
1606 {
1607         struct ttm_bo_device *bdev = bo->bdev;
1608         int ret;
1609
1610 retry_pre_get:
1611         ret = drm_mm_pre_get(&bdev->addr_space_mm);
1612         if (unlikely(ret != 0))
1613                 return ret;
1614
1615         lockmgr(&bdev->vm_lock, LK_EXCLUSIVE);
1616         bo->vm_node = drm_mm_search_free(&bdev->addr_space_mm,
1617                                          bo->mem.num_pages, 0, 0);
1618
1619         if (unlikely(bo->vm_node == NULL)) {
1620                 ret = -ENOMEM;
1621                 goto out_unlock;
1622         }
1623
1624         bo->vm_node = drm_mm_get_block_atomic(bo->vm_node,
1625                                               bo->mem.num_pages, 0);
1626
1627         if (unlikely(bo->vm_node == NULL)) {
1628                 lockmgr(&bdev->vm_lock, LK_RELEASE);
1629                 goto retry_pre_get;
1630         }
1631
1632         ttm_bo_vm_insert_rb(bo);
1633         lockmgr(&bdev->vm_lock, LK_RELEASE);
1634         bo->addr_space_offset = ((uint64_t) bo->vm_node->start) << PAGE_SHIFT;
1635
1636         return 0;
1637 out_unlock:
1638         lockmgr(&bdev->vm_lock, LK_RELEASE);
1639         return ret;
1640 }
1641
1642 int ttm_bo_wait(struct ttm_buffer_object *bo,
1643                 bool lazy, bool interruptible, bool no_wait)
1644 {
1645         struct ttm_bo_driver *driver = bo->bdev->driver;
1646         struct ttm_bo_device *bdev = bo->bdev;
1647         void *sync_obj;
1648         int ret = 0;
1649
1650         if (likely(bo->sync_obj == NULL))
1651                 return 0;
1652
1653         while (bo->sync_obj) {
1654
1655                 if (driver->sync_obj_signaled(bo->sync_obj)) {
1656                         void *tmp_obj = bo->sync_obj;
1657                         bo->sync_obj = NULL;
1658                         clear_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags);
1659                         lockmgr(&bdev->fence_lock, LK_RELEASE);
1660                         driver->sync_obj_unref(&tmp_obj);
1661                         lockmgr(&bdev->fence_lock, LK_EXCLUSIVE);
1662                         continue;
1663                 }
1664
1665                 if (no_wait)
1666                         return -EBUSY;
1667
1668                 sync_obj = driver->sync_obj_ref(bo->sync_obj);
1669                 lockmgr(&bdev->fence_lock, LK_RELEASE);
1670                 ret = driver->sync_obj_wait(sync_obj,
1671                                             lazy, interruptible);
1672                 if (unlikely(ret != 0)) {
1673                         driver->sync_obj_unref(&sync_obj);
1674                         lockmgr(&bdev->fence_lock, LK_EXCLUSIVE);
1675                         return ret;
1676                 }
1677                 lockmgr(&bdev->fence_lock, LK_EXCLUSIVE);
1678                 if (likely(bo->sync_obj == sync_obj)) {
1679                         void *tmp_obj = bo->sync_obj;
1680                         bo->sync_obj = NULL;
1681                         clear_bit(TTM_BO_PRIV_FLAG_MOVING,
1682                                   &bo->priv_flags);
1683                         lockmgr(&bdev->fence_lock, LK_RELEASE);
1684                         driver->sync_obj_unref(&sync_obj);
1685                         driver->sync_obj_unref(&tmp_obj);
1686                         lockmgr(&bdev->fence_lock, LK_EXCLUSIVE);
1687                 } else {
1688                         lockmgr(&bdev->fence_lock, LK_RELEASE);
1689                         driver->sync_obj_unref(&sync_obj);
1690                         lockmgr(&bdev->fence_lock, LK_EXCLUSIVE);
1691                 }
1692         }
1693         return 0;
1694 }
1695
1696 int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait)
1697 {
1698         struct ttm_bo_device *bdev = bo->bdev;
1699         int ret = 0;
1700
1701         /*
1702          * Using ttm_bo_reserve makes sure the lru lists are updated.
1703          */
1704
1705         ret = ttm_bo_reserve(bo, true, no_wait, false, 0);
1706         if (unlikely(ret != 0))
1707                 return ret;
1708         lockmgr(&bdev->fence_lock, LK_EXCLUSIVE);
1709         ret = ttm_bo_wait(bo, false, true, no_wait);
1710         lockmgr(&bdev->fence_lock, LK_RELEASE);
1711         if (likely(ret == 0))
1712                 atomic_inc(&bo->cpu_writers);
1713         ttm_bo_unreserve(bo);
1714         return ret;
1715 }
1716
1717 void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo)
1718 {
1719         atomic_dec(&bo->cpu_writers);
1720 }
1721
1722 /**
1723  * A buffer object shrink method that tries to swap out the first
1724  * buffer object on the bo_global::swap_lru list.
1725  */
1726
1727 static int ttm_bo_swapout(struct ttm_mem_shrink *shrink)
1728 {
1729         struct ttm_bo_global *glob =
1730             container_of(shrink, struct ttm_bo_global, shrink);
1731         struct ttm_buffer_object *bo;
1732         int ret = -EBUSY;
1733         int put_count;
1734         uint32_t swap_placement = (TTM_PL_FLAG_CACHED | TTM_PL_FLAG_SYSTEM);
1735
1736         lockmgr(&glob->lru_lock, LK_EXCLUSIVE);
1737         list_for_each_entry(bo, &glob->swap_lru, swap) {
1738                 ret = ttm_bo_reserve_locked(bo, false, true, false, 0);
1739                 if (!ret)
1740                         break;
1741         }
1742
1743         if (ret) {
1744                 lockmgr(&glob->lru_lock, LK_RELEASE);
1745                 return ret;
1746         }
1747
1748         refcount_acquire(&bo->list_kref);
1749
1750         if (!list_empty(&bo->ddestroy)) {
1751                 ret = ttm_bo_cleanup_refs_and_unlock(bo, false, false);
1752                 if (refcount_release(&bo->list_kref))
1753                         ttm_bo_release_list(bo);
1754                 return ret;
1755         }
1756
1757         put_count = ttm_bo_del_from_lru(bo);
1758         lockmgr(&glob->lru_lock, LK_RELEASE);
1759
1760         ttm_bo_list_ref_sub(bo, put_count, true);
1761
1762         /**
1763          * Wait for GPU, then move to system cached.
1764          */
1765
1766         lockmgr(&bo->bdev->fence_lock, LK_EXCLUSIVE);
1767         ret = ttm_bo_wait(bo, false, false, false);
1768         lockmgr(&bo->bdev->fence_lock, LK_RELEASE);
1769
1770         if (unlikely(ret != 0))
1771                 goto out;
1772
1773         if ((bo->mem.placement & swap_placement) != swap_placement) {
1774                 struct ttm_mem_reg evict_mem;
1775
1776                 evict_mem = bo->mem;
1777                 evict_mem.mm_node = NULL;
1778                 evict_mem.placement = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED;
1779                 evict_mem.mem_type = TTM_PL_SYSTEM;
1780
1781                 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true,
1782                                              false, false);
1783                 if (unlikely(ret != 0))
1784                         goto out;
1785         }
1786
1787         ttm_bo_unmap_virtual(bo);
1788
1789         /**
1790          * Swap out. Buffer will be swapped in again as soon as
1791          * anyone tries to access a ttm page.
1792          */
1793
1794         if (bo->bdev->driver->swap_notify)
1795                 bo->bdev->driver->swap_notify(bo);
1796
1797         ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage);
1798 out:
1799
1800         /**
1801          *
1802          * Unreserve without putting on LRU to avoid swapping out an
1803          * already swapped buffer.
1804          */
1805
1806         atomic_set(&bo->reserved, 0);
1807         wakeup(bo);
1808         if (refcount_release(&bo->list_kref))
1809                 ttm_bo_release_list(bo);
1810         return ret;
1811 }
1812
1813 void ttm_bo_swapout_all(struct ttm_bo_device *bdev)
1814 {
1815         while (ttm_bo_swapout(&bdev->glob->shrink) == 0)
1816                 ;
1817 }