1 /* SPDX-License-Identifier: GPL-2.0 OR MIT */ 2 /************************************************************************** 3 * 4 * Copyright (c) 2006-2009 VMware, Inc., Palo Alto, CA., USA 5 * All Rights Reserved. 6 * 7 * Permission is hereby granted, free of charge, to any person obtaining a 8 * copy of this software and associated documentation files (the 9 * "Software"), to deal in the Software without restriction, including 10 * without limitation the rights to use, copy, modify, merge, publish, 11 * distribute, sub license, and/or sell copies of the Software, and to 12 * permit persons to whom the Software is furnished to do so, subject to 13 * the following conditions: 14 * 15 * The above copyright notice and this permission notice (including the 16 * next paragraph) shall be included in all copies or substantial portions 17 * of the Software. 18 * 19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 20 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 21 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL 22 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, 23 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR 24 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE 25 * USE OR OTHER DEALINGS IN THE SOFTWARE. 26 * 27 **************************************************************************/ 28 /* 29 * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com> 30 */ 31 32 #define pr_fmt(fmt) "[TTM] " fmt 33 34 #include <drm/ttm/ttm_module.h> 35 #include <drm/ttm/ttm_bo_driver.h> 36 #include <drm/ttm/ttm_placement.h> 37 #include <linux/jiffies.h> 38 #include <linux/slab.h> 39 #include <linux/sched.h> 40 #include <linux/mm.h> 41 #include <linux/file.h> 42 #include <linux/module.h> 43 #include <linux/atomic.h> 44 #include <linux/reservation.h> 45 46 static void ttm_bo_global_kobj_release(struct kobject *kobj); 47 48 /** 49 * ttm_global_mutex - protecting the global BO state 50 */ 51 DEFINE_MUTEX(ttm_global_mutex); 52 unsigned ttm_bo_glob_use_count; 53 struct ttm_bo_global ttm_bo_glob; 54 55 static struct attribute ttm_bo_count = { 56 .name = "bo_count", 57 .mode = S_IRUGO 58 }; 59 60 /* default destructor */ 61 static void ttm_bo_default_destroy(struct ttm_buffer_object *bo) 62 { 63 kfree(bo); 64 } 65 66 static inline int ttm_mem_type_from_place(const struct ttm_place *place, 67 uint32_t *mem_type) 68 { 69 int pos; 70 71 pos = ffs(place->flags & TTM_PL_MASK_MEM); 72 if (unlikely(!pos)) 73 return -EINVAL; 74 75 *mem_type = pos - 1; 76 return 0; 77 } 78 79 static void ttm_mem_type_debug(struct ttm_bo_device *bdev, struct drm_printer *p, 80 int mem_type) 81 { 82 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 83 84 drm_printf(p, " has_type: %d\n", man->has_type); 85 drm_printf(p, " use_type: %d\n", man->use_type); 86 drm_printf(p, " flags: 0x%08X\n", man->flags); 87 drm_printf(p, " gpu_offset: 0x%08llX\n", man->gpu_offset); 88 drm_printf(p, " size: %llu\n", man->size); 89 drm_printf(p, " available_caching: 0x%08X\n", man->available_caching); 90 drm_printf(p, " default_caching: 0x%08X\n", man->default_caching); 91 if (mem_type != TTM_PL_SYSTEM) 92 (*man->func->debug)(man, p); 93 } 94 95 static void ttm_bo_mem_space_debug(struct ttm_buffer_object *bo, 96 struct ttm_placement *placement) 97 { 98 struct drm_printer p = drm_debug_printer(TTM_PFX); 99 int i, ret, mem_type; 100 101 drm_printf(&p, "No space for %p (%lu pages, %luK, %luM)\n", 102 bo, bo->mem.num_pages, bo->mem.size >> 10, 103 bo->mem.size >> 20); 104 for (i = 0; i < placement->num_placement; i++) { 105 ret = ttm_mem_type_from_place(&placement->placement[i], 106 &mem_type); 107 if (ret) 108 return; 109 drm_printf(&p, " placement[%d]=0x%08X (%d)\n", 110 i, placement->placement[i].flags, mem_type); 111 ttm_mem_type_debug(bo->bdev, &p, mem_type); 112 } 113 } 114 115 static ssize_t ttm_bo_global_show(struct kobject *kobj, 116 struct attribute *attr, 117 char *buffer) 118 { 119 struct ttm_bo_global *glob = 120 container_of(kobj, struct ttm_bo_global, kobj); 121 122 return snprintf(buffer, PAGE_SIZE, "%d\n", 123 atomic_read(&glob->bo_count)); 124 } 125 126 static struct attribute *ttm_bo_global_attrs[] = { 127 &ttm_bo_count, 128 NULL 129 }; 130 131 static const struct sysfs_ops ttm_bo_global_ops = { 132 .show = &ttm_bo_global_show 133 }; 134 135 static struct kobj_type ttm_bo_glob_kobj_type = { 136 .release = &ttm_bo_global_kobj_release, 137 .sysfs_ops = &ttm_bo_global_ops, 138 .default_attrs = ttm_bo_global_attrs 139 }; 140 141 142 static inline uint32_t ttm_bo_type_flags(unsigned type) 143 { 144 return 1 << (type); 145 } 146 147 static void ttm_bo_release_list(struct kref *list_kref) 148 { 149 struct ttm_buffer_object *bo = 150 container_of(list_kref, struct ttm_buffer_object, list_kref); 151 struct ttm_bo_device *bdev = bo->bdev; 152 size_t acc_size = bo->acc_size; 153 154 BUG_ON(kref_read(&bo->list_kref)); 155 BUG_ON(kref_read(&bo->kref)); 156 BUG_ON(atomic_read(&bo->cpu_writers)); 157 BUG_ON(bo->mem.mm_node != NULL); 158 BUG_ON(!list_empty(&bo->lru)); 159 BUG_ON(!list_empty(&bo->ddestroy)); 160 ttm_tt_destroy(bo->ttm); 161 atomic_dec(&bo->bdev->glob->bo_count); 162 dma_fence_put(bo->moving); 163 reservation_object_fini(&bo->ttm_resv); 164 mutex_destroy(&bo->wu_mutex); 165 bo->destroy(bo); 166 ttm_mem_global_free(bdev->glob->mem_glob, acc_size); 167 } 168 169 static void ttm_bo_add_mem_to_lru(struct ttm_buffer_object *bo, 170 struct ttm_mem_reg *mem) 171 { 172 struct ttm_bo_device *bdev = bo->bdev; 173 struct ttm_mem_type_manager *man; 174 175 reservation_object_assert_held(bo->resv); 176 177 if (!list_empty(&bo->lru)) 178 return; 179 180 if (mem->placement & TTM_PL_FLAG_NO_EVICT) 181 return; 182 183 man = &bdev->man[mem->mem_type]; 184 list_add_tail(&bo->lru, &man->lru[bo->priority]); 185 kref_get(&bo->list_kref); 186 187 if (bo->ttm && !(bo->ttm->page_flags & 188 (TTM_PAGE_FLAG_SG | TTM_PAGE_FLAG_SWAPPED))) { 189 list_add_tail(&bo->swap, &bdev->glob->swap_lru[bo->priority]); 190 kref_get(&bo->list_kref); 191 } 192 } 193 194 void ttm_bo_add_to_lru(struct ttm_buffer_object *bo) 195 { 196 ttm_bo_add_mem_to_lru(bo, &bo->mem); 197 } 198 EXPORT_SYMBOL(ttm_bo_add_to_lru); 199 200 static void ttm_bo_ref_bug(struct kref *list_kref) 201 { 202 BUG(); 203 } 204 205 void ttm_bo_del_from_lru(struct ttm_buffer_object *bo) 206 { 207 struct ttm_bo_device *bdev = bo->bdev; 208 bool notify = false; 209 210 if (!list_empty(&bo->swap)) { 211 list_del_init(&bo->swap); 212 kref_put(&bo->list_kref, ttm_bo_ref_bug); 213 notify = true; 214 } 215 if (!list_empty(&bo->lru)) { 216 list_del_init(&bo->lru); 217 kref_put(&bo->list_kref, ttm_bo_ref_bug); 218 notify = true; 219 } 220 221 if (notify && bdev->driver->del_from_lru_notify) 222 bdev->driver->del_from_lru_notify(bo); 223 } 224 225 void ttm_bo_del_sub_from_lru(struct ttm_buffer_object *bo) 226 { 227 struct ttm_bo_global *glob = bo->bdev->glob; 228 229 spin_lock(&glob->lru_lock); 230 ttm_bo_del_from_lru(bo); 231 spin_unlock(&glob->lru_lock); 232 } 233 EXPORT_SYMBOL(ttm_bo_del_sub_from_lru); 234 235 static void ttm_bo_bulk_move_set_pos(struct ttm_lru_bulk_move_pos *pos, 236 struct ttm_buffer_object *bo) 237 { 238 if (!pos->first) 239 pos->first = bo; 240 pos->last = bo; 241 } 242 243 void ttm_bo_move_to_lru_tail(struct ttm_buffer_object *bo, 244 struct ttm_lru_bulk_move *bulk) 245 { 246 reservation_object_assert_held(bo->resv); 247 248 ttm_bo_del_from_lru(bo); 249 ttm_bo_add_to_lru(bo); 250 251 if (bulk && !(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) { 252 switch (bo->mem.mem_type) { 253 case TTM_PL_TT: 254 ttm_bo_bulk_move_set_pos(&bulk->tt[bo->priority], bo); 255 break; 256 257 case TTM_PL_VRAM: 258 ttm_bo_bulk_move_set_pos(&bulk->vram[bo->priority], bo); 259 break; 260 } 261 if (bo->ttm && !(bo->ttm->page_flags & 262 (TTM_PAGE_FLAG_SG | TTM_PAGE_FLAG_SWAPPED))) 263 ttm_bo_bulk_move_set_pos(&bulk->swap[bo->priority], bo); 264 } 265 } 266 EXPORT_SYMBOL(ttm_bo_move_to_lru_tail); 267 268 void ttm_bo_bulk_move_lru_tail(struct ttm_lru_bulk_move *bulk) 269 { 270 unsigned i; 271 272 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) { 273 struct ttm_lru_bulk_move_pos *pos = &bulk->tt[i]; 274 struct ttm_mem_type_manager *man; 275 276 if (!pos->first) 277 continue; 278 279 reservation_object_assert_held(pos->first->resv); 280 reservation_object_assert_held(pos->last->resv); 281 282 man = &pos->first->bdev->man[TTM_PL_TT]; 283 list_bulk_move_tail(&man->lru[i], &pos->first->lru, 284 &pos->last->lru); 285 } 286 287 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) { 288 struct ttm_lru_bulk_move_pos *pos = &bulk->vram[i]; 289 struct ttm_mem_type_manager *man; 290 291 if (!pos->first) 292 continue; 293 294 reservation_object_assert_held(pos->first->resv); 295 reservation_object_assert_held(pos->last->resv); 296 297 man = &pos->first->bdev->man[TTM_PL_VRAM]; 298 list_bulk_move_tail(&man->lru[i], &pos->first->lru, 299 &pos->last->lru); 300 } 301 302 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) { 303 struct ttm_lru_bulk_move_pos *pos = &bulk->swap[i]; 304 struct list_head *lru; 305 306 if (!pos->first) 307 continue; 308 309 reservation_object_assert_held(pos->first->resv); 310 reservation_object_assert_held(pos->last->resv); 311 312 lru = &pos->first->bdev->glob->swap_lru[i]; 313 list_bulk_move_tail(lru, &pos->first->swap, &pos->last->swap); 314 } 315 } 316 EXPORT_SYMBOL(ttm_bo_bulk_move_lru_tail); 317 318 static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo, 319 struct ttm_mem_reg *mem, bool evict, 320 struct ttm_operation_ctx *ctx) 321 { 322 struct ttm_bo_device *bdev = bo->bdev; 323 bool old_is_pci = ttm_mem_reg_is_pci(bdev, &bo->mem); 324 bool new_is_pci = ttm_mem_reg_is_pci(bdev, mem); 325 struct ttm_mem_type_manager *old_man = &bdev->man[bo->mem.mem_type]; 326 struct ttm_mem_type_manager *new_man = &bdev->man[mem->mem_type]; 327 int ret = 0; 328 329 if (old_is_pci || new_is_pci || 330 ((mem->placement & bo->mem.placement & TTM_PL_MASK_CACHING) == 0)) { 331 ret = ttm_mem_io_lock(old_man, true); 332 if (unlikely(ret != 0)) 333 goto out_err; 334 ttm_bo_unmap_virtual_locked(bo); 335 ttm_mem_io_unlock(old_man); 336 } 337 338 /* 339 * Create and bind a ttm if required. 340 */ 341 342 if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) { 343 if (bo->ttm == NULL) { 344 bool zero = !(old_man->flags & TTM_MEMTYPE_FLAG_FIXED); 345 ret = ttm_tt_create(bo, zero); 346 if (ret) 347 goto out_err; 348 } 349 350 ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement); 351 if (ret) 352 goto out_err; 353 354 if (mem->mem_type != TTM_PL_SYSTEM) { 355 ret = ttm_tt_bind(bo->ttm, mem, ctx); 356 if (ret) 357 goto out_err; 358 } 359 360 if (bo->mem.mem_type == TTM_PL_SYSTEM) { 361 if (bdev->driver->move_notify) 362 bdev->driver->move_notify(bo, evict, mem); 363 bo->mem = *mem; 364 mem->mm_node = NULL; 365 goto moved; 366 } 367 } 368 369 if (bdev->driver->move_notify) 370 bdev->driver->move_notify(bo, evict, mem); 371 372 if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) && 373 !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) 374 ret = ttm_bo_move_ttm(bo, ctx, mem); 375 else if (bdev->driver->move) 376 ret = bdev->driver->move(bo, evict, ctx, mem); 377 else 378 ret = ttm_bo_move_memcpy(bo, ctx, mem); 379 380 if (ret) { 381 if (bdev->driver->move_notify) { 382 swap(*mem, bo->mem); 383 bdev->driver->move_notify(bo, false, mem); 384 swap(*mem, bo->mem); 385 } 386 387 goto out_err; 388 } 389 390 moved: 391 if (bo->evicted) { 392 if (bdev->driver->invalidate_caches) { 393 ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement); 394 if (ret) 395 pr_err("Can not flush read caches\n"); 396 } 397 bo->evicted = false; 398 } 399 400 if (bo->mem.mm_node) 401 bo->offset = (bo->mem.start << PAGE_SHIFT) + 402 bdev->man[bo->mem.mem_type].gpu_offset; 403 else 404 bo->offset = 0; 405 406 ctx->bytes_moved += bo->num_pages << PAGE_SHIFT; 407 return 0; 408 409 out_err: 410 new_man = &bdev->man[bo->mem.mem_type]; 411 if (new_man->flags & TTM_MEMTYPE_FLAG_FIXED) { 412 ttm_tt_destroy(bo->ttm); 413 bo->ttm = NULL; 414 } 415 416 return ret; 417 } 418 419 /** 420 * Call bo::reserved. 421 * Will release GPU memory type usage on destruction. 422 * This is the place to put in driver specific hooks to release 423 * driver private resources. 424 * Will release the bo::reserved lock. 425 */ 426 427 static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo) 428 { 429 if (bo->bdev->driver->move_notify) 430 bo->bdev->driver->move_notify(bo, false, NULL); 431 432 ttm_tt_destroy(bo->ttm); 433 bo->ttm = NULL; 434 ttm_bo_mem_put(bo, &bo->mem); 435 } 436 437 static int ttm_bo_individualize_resv(struct ttm_buffer_object *bo) 438 { 439 int r; 440 441 if (bo->resv == &bo->ttm_resv) 442 return 0; 443 444 BUG_ON(!reservation_object_trylock(&bo->ttm_resv)); 445 446 r = reservation_object_copy_fences(&bo->ttm_resv, bo->resv); 447 if (r) 448 reservation_object_unlock(&bo->ttm_resv); 449 450 return r; 451 } 452 453 static void ttm_bo_flush_all_fences(struct ttm_buffer_object *bo) 454 { 455 struct reservation_object_list *fobj; 456 struct dma_fence *fence; 457 int i; 458 459 fobj = reservation_object_get_list(&bo->ttm_resv); 460 fence = reservation_object_get_excl(&bo->ttm_resv); 461 if (fence && !fence->ops->signaled) 462 dma_fence_enable_sw_signaling(fence); 463 464 for (i = 0; fobj && i < fobj->shared_count; ++i) { 465 fence = rcu_dereference_protected(fobj->shared[i], 466 reservation_object_held(bo->resv)); 467 468 if (!fence->ops->signaled) 469 dma_fence_enable_sw_signaling(fence); 470 } 471 } 472 473 static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo) 474 { 475 struct ttm_bo_device *bdev = bo->bdev; 476 struct ttm_bo_global *glob = bdev->glob; 477 int ret; 478 479 ret = ttm_bo_individualize_resv(bo); 480 if (ret) { 481 /* Last resort, if we fail to allocate memory for the 482 * fences block for the BO to become idle 483 */ 484 reservation_object_wait_timeout_rcu(bo->resv, true, false, 485 30 * HZ); 486 spin_lock(&glob->lru_lock); 487 goto error; 488 } 489 490 spin_lock(&glob->lru_lock); 491 ret = reservation_object_trylock(bo->resv) ? 0 : -EBUSY; 492 if (!ret) { 493 if (reservation_object_test_signaled_rcu(&bo->ttm_resv, true)) { 494 ttm_bo_del_from_lru(bo); 495 spin_unlock(&glob->lru_lock); 496 if (bo->resv != &bo->ttm_resv) 497 reservation_object_unlock(&bo->ttm_resv); 498 499 ttm_bo_cleanup_memtype_use(bo); 500 reservation_object_unlock(bo->resv); 501 return; 502 } 503 504 ttm_bo_flush_all_fences(bo); 505 506 /* 507 * Make NO_EVICT bos immediately available to 508 * shrinkers, now that they are queued for 509 * destruction. 510 */ 511 if (bo->mem.placement & TTM_PL_FLAG_NO_EVICT) { 512 bo->mem.placement &= ~TTM_PL_FLAG_NO_EVICT; 513 ttm_bo_add_to_lru(bo); 514 } 515 516 reservation_object_unlock(bo->resv); 517 } 518 if (bo->resv != &bo->ttm_resv) 519 reservation_object_unlock(&bo->ttm_resv); 520 521 error: 522 kref_get(&bo->list_kref); 523 list_add_tail(&bo->ddestroy, &bdev->ddestroy); 524 spin_unlock(&glob->lru_lock); 525 526 schedule_delayed_work(&bdev->wq, 527 ((HZ / 100) < 1) ? 1 : HZ / 100); 528 } 529 530 /** 531 * function ttm_bo_cleanup_refs 532 * If bo idle, remove from delayed- and lru lists, and unref. 533 * If not idle, do nothing. 534 * 535 * Must be called with lru_lock and reservation held, this function 536 * will drop the lru lock and optionally the reservation lock before returning. 537 * 538 * @interruptible Any sleeps should occur interruptibly. 539 * @no_wait_gpu Never wait for gpu. Return -EBUSY instead. 540 * @unlock_resv Unlock the reservation lock as well. 541 */ 542 543 static int ttm_bo_cleanup_refs(struct ttm_buffer_object *bo, 544 bool interruptible, bool no_wait_gpu, 545 bool unlock_resv) 546 { 547 struct ttm_bo_global *glob = bo->bdev->glob; 548 struct reservation_object *resv; 549 int ret; 550 551 if (unlikely(list_empty(&bo->ddestroy))) 552 resv = bo->resv; 553 else 554 resv = &bo->ttm_resv; 555 556 if (reservation_object_test_signaled_rcu(resv, true)) 557 ret = 0; 558 else 559 ret = -EBUSY; 560 561 if (ret && !no_wait_gpu) { 562 long lret; 563 564 if (unlock_resv) 565 reservation_object_unlock(bo->resv); 566 spin_unlock(&glob->lru_lock); 567 568 lret = reservation_object_wait_timeout_rcu(resv, true, 569 interruptible, 570 30 * HZ); 571 572 if (lret < 0) 573 return lret; 574 else if (lret == 0) 575 return -EBUSY; 576 577 spin_lock(&glob->lru_lock); 578 if (unlock_resv && !reservation_object_trylock(bo->resv)) { 579 /* 580 * We raced, and lost, someone else holds the reservation now, 581 * and is probably busy in ttm_bo_cleanup_memtype_use. 582 * 583 * Even if it's not the case, because we finished waiting any 584 * delayed destruction would succeed, so just return success 585 * here. 586 */ 587 spin_unlock(&glob->lru_lock); 588 return 0; 589 } 590 ret = 0; 591 } 592 593 if (ret || unlikely(list_empty(&bo->ddestroy))) { 594 if (unlock_resv) 595 reservation_object_unlock(bo->resv); 596 spin_unlock(&glob->lru_lock); 597 return ret; 598 } 599 600 ttm_bo_del_from_lru(bo); 601 list_del_init(&bo->ddestroy); 602 kref_put(&bo->list_kref, ttm_bo_ref_bug); 603 604 spin_unlock(&glob->lru_lock); 605 ttm_bo_cleanup_memtype_use(bo); 606 607 if (unlock_resv) 608 reservation_object_unlock(bo->resv); 609 610 return 0; 611 } 612 613 /** 614 * Traverse the delayed list, and call ttm_bo_cleanup_refs on all 615 * encountered buffers. 616 */ 617 static bool ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all) 618 { 619 struct ttm_bo_global *glob = bdev->glob; 620 struct list_head removed; 621 bool empty; 622 623 INIT_LIST_HEAD(&removed); 624 625 spin_lock(&glob->lru_lock); 626 while (!list_empty(&bdev->ddestroy)) { 627 struct ttm_buffer_object *bo; 628 629 bo = list_first_entry(&bdev->ddestroy, struct ttm_buffer_object, 630 ddestroy); 631 kref_get(&bo->list_kref); 632 list_move_tail(&bo->ddestroy, &removed); 633 634 if (remove_all || bo->resv != &bo->ttm_resv) { 635 spin_unlock(&glob->lru_lock); 636 reservation_object_lock(bo->resv, NULL); 637 638 spin_lock(&glob->lru_lock); 639 ttm_bo_cleanup_refs(bo, false, !remove_all, true); 640 641 } else if (reservation_object_trylock(bo->resv)) { 642 ttm_bo_cleanup_refs(bo, false, !remove_all, true); 643 } else { 644 spin_unlock(&glob->lru_lock); 645 } 646 647 kref_put(&bo->list_kref, ttm_bo_release_list); 648 spin_lock(&glob->lru_lock); 649 } 650 list_splice_tail(&removed, &bdev->ddestroy); 651 empty = list_empty(&bdev->ddestroy); 652 spin_unlock(&glob->lru_lock); 653 654 return empty; 655 } 656 657 static void ttm_bo_delayed_workqueue(struct work_struct *work) 658 { 659 struct ttm_bo_device *bdev = 660 container_of(work, struct ttm_bo_device, wq.work); 661 662 if (!ttm_bo_delayed_delete(bdev, false)) 663 schedule_delayed_work(&bdev->wq, 664 ((HZ / 100) < 1) ? 1 : HZ / 100); 665 } 666 667 static void ttm_bo_release(struct kref *kref) 668 { 669 struct ttm_buffer_object *bo = 670 container_of(kref, struct ttm_buffer_object, kref); 671 struct ttm_bo_device *bdev = bo->bdev; 672 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type]; 673 674 drm_vma_offset_remove(&bdev->vma_manager, &bo->vma_node); 675 ttm_mem_io_lock(man, false); 676 ttm_mem_io_free_vm(bo); 677 ttm_mem_io_unlock(man); 678 ttm_bo_cleanup_refs_or_queue(bo); 679 kref_put(&bo->list_kref, ttm_bo_release_list); 680 } 681 682 void ttm_bo_put(struct ttm_buffer_object *bo) 683 { 684 kref_put(&bo->kref, ttm_bo_release); 685 } 686 EXPORT_SYMBOL(ttm_bo_put); 687 688 int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev) 689 { 690 return cancel_delayed_work_sync(&bdev->wq); 691 } 692 EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue); 693 694 void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched) 695 { 696 if (resched) 697 schedule_delayed_work(&bdev->wq, 698 ((HZ / 100) < 1) ? 1 : HZ / 100); 699 } 700 EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue); 701 702 static int ttm_bo_evict(struct ttm_buffer_object *bo, 703 struct ttm_operation_ctx *ctx) 704 { 705 struct ttm_bo_device *bdev = bo->bdev; 706 struct ttm_mem_reg evict_mem; 707 struct ttm_placement placement; 708 int ret = 0; 709 710 reservation_object_assert_held(bo->resv); 711 712 placement.num_placement = 0; 713 placement.num_busy_placement = 0; 714 bdev->driver->evict_flags(bo, &placement); 715 716 if (!placement.num_placement && !placement.num_busy_placement) { 717 ret = ttm_bo_pipeline_gutting(bo); 718 if (ret) 719 return ret; 720 721 return ttm_tt_create(bo, false); 722 } 723 724 evict_mem = bo->mem; 725 evict_mem.mm_node = NULL; 726 evict_mem.bus.io_reserved_vm = false; 727 evict_mem.bus.io_reserved_count = 0; 728 729 ret = ttm_bo_mem_space(bo, &placement, &evict_mem, ctx); 730 if (ret) { 731 if (ret != -ERESTARTSYS) { 732 pr_err("Failed to find memory space for buffer 0x%p eviction\n", 733 bo); 734 ttm_bo_mem_space_debug(bo, &placement); 735 } 736 goto out; 737 } 738 739 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, ctx); 740 if (unlikely(ret)) { 741 if (ret != -ERESTARTSYS) 742 pr_err("Buffer eviction failed\n"); 743 ttm_bo_mem_put(bo, &evict_mem); 744 goto out; 745 } 746 bo->evicted = true; 747 out: 748 return ret; 749 } 750 751 bool ttm_bo_eviction_valuable(struct ttm_buffer_object *bo, 752 const struct ttm_place *place) 753 { 754 /* Don't evict this BO if it's outside of the 755 * requested placement range 756 */ 757 if (place->fpfn >= (bo->mem.start + bo->mem.size) || 758 (place->lpfn && place->lpfn <= bo->mem.start)) 759 return false; 760 761 return true; 762 } 763 EXPORT_SYMBOL(ttm_bo_eviction_valuable); 764 765 /** 766 * Check the target bo is allowable to be evicted or swapout, including cases: 767 * 768 * a. if share same reservation object with ctx->resv, have assumption 769 * reservation objects should already be locked, so not lock again and 770 * return true directly when either the opreation allow_reserved_eviction 771 * or the target bo already is in delayed free list; 772 * 773 * b. Otherwise, trylock it. 774 */ 775 static bool ttm_bo_evict_swapout_allowable(struct ttm_buffer_object *bo, 776 struct ttm_operation_ctx *ctx, bool *locked, bool *busy) 777 { 778 bool ret = false; 779 780 if (bo->resv == ctx->resv) { 781 reservation_object_assert_held(bo->resv); 782 if (ctx->flags & TTM_OPT_FLAG_ALLOW_RES_EVICT 783 || !list_empty(&bo->ddestroy)) 784 ret = true; 785 *locked = false; 786 if (busy) 787 *busy = false; 788 } else { 789 ret = reservation_object_trylock(bo->resv); 790 *locked = ret; 791 if (busy) 792 *busy = !ret; 793 } 794 795 return ret; 796 } 797 798 /** 799 * ttm_mem_evict_wait_busy - wait for a busy BO to become available 800 * 801 * @busy_bo: BO which couldn't be locked with trylock 802 * @ctx: operation context 803 * @ticket: acquire ticket 804 * 805 * Try to lock a busy buffer object to avoid failing eviction. 806 */ 807 static int ttm_mem_evict_wait_busy(struct ttm_buffer_object *busy_bo, 808 struct ttm_operation_ctx *ctx, 809 struct ww_acquire_ctx *ticket) 810 { 811 int r; 812 813 if (!busy_bo || !ticket) 814 return -EBUSY; 815 816 if (ctx->interruptible) 817 r = reservation_object_lock_interruptible(busy_bo->resv, 818 ticket); 819 else 820 r = reservation_object_lock(busy_bo->resv, ticket); 821 822 /* 823 * TODO: It would be better to keep the BO locked until allocation is at 824 * least tried one more time, but that would mean a much larger rework 825 * of TTM. 826 */ 827 if (!r) 828 reservation_object_unlock(busy_bo->resv); 829 830 return r == -EDEADLK ? -EBUSY : r; 831 } 832 833 static int ttm_mem_evict_first(struct ttm_bo_device *bdev, 834 uint32_t mem_type, 835 const struct ttm_place *place, 836 struct ttm_operation_ctx *ctx, 837 struct ww_acquire_ctx *ticket) 838 { 839 struct ttm_buffer_object *bo = NULL, *busy_bo = NULL; 840 struct ttm_bo_global *glob = bdev->glob; 841 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 842 bool locked = false; 843 unsigned i; 844 int ret; 845 846 spin_lock(&glob->lru_lock); 847 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) { 848 list_for_each_entry(bo, &man->lru[i], lru) { 849 bool busy; 850 851 if (!ttm_bo_evict_swapout_allowable(bo, ctx, &locked, 852 &busy)) { 853 if (busy && !busy_bo && 854 bo->resv->lock.ctx != ticket) 855 busy_bo = bo; 856 continue; 857 } 858 859 if (place && !bdev->driver->eviction_valuable(bo, 860 place)) { 861 if (locked) 862 reservation_object_unlock(bo->resv); 863 continue; 864 } 865 break; 866 } 867 868 /* If the inner loop terminated early, we have our candidate */ 869 if (&bo->lru != &man->lru[i]) 870 break; 871 872 bo = NULL; 873 } 874 875 if (!bo) { 876 if (busy_bo) 877 ttm_bo_get(busy_bo); 878 spin_unlock(&glob->lru_lock); 879 ret = ttm_mem_evict_wait_busy(busy_bo, ctx, ticket); 880 if (busy_bo) 881 ttm_bo_put(busy_bo); 882 return ret; 883 } 884 885 kref_get(&bo->list_kref); 886 887 if (!list_empty(&bo->ddestroy)) { 888 ret = ttm_bo_cleanup_refs(bo, ctx->interruptible, 889 ctx->no_wait_gpu, locked); 890 kref_put(&bo->list_kref, ttm_bo_release_list); 891 return ret; 892 } 893 894 ttm_bo_del_from_lru(bo); 895 spin_unlock(&glob->lru_lock); 896 897 ret = ttm_bo_evict(bo, ctx); 898 if (locked) { 899 ttm_bo_unreserve(bo); 900 } else { 901 spin_lock(&glob->lru_lock); 902 ttm_bo_add_to_lru(bo); 903 spin_unlock(&glob->lru_lock); 904 } 905 906 kref_put(&bo->list_kref, ttm_bo_release_list); 907 return ret; 908 } 909 910 void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem) 911 { 912 struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type]; 913 914 if (mem->mm_node) 915 (*man->func->put_node)(man, mem); 916 } 917 EXPORT_SYMBOL(ttm_bo_mem_put); 918 919 /** 920 * Add the last move fence to the BO and reserve a new shared slot. 921 */ 922 static int ttm_bo_add_move_fence(struct ttm_buffer_object *bo, 923 struct ttm_mem_type_manager *man, 924 struct ttm_mem_reg *mem) 925 { 926 struct dma_fence *fence; 927 int ret; 928 929 spin_lock(&man->move_lock); 930 fence = dma_fence_get(man->move); 931 spin_unlock(&man->move_lock); 932 933 if (fence) { 934 reservation_object_add_shared_fence(bo->resv, fence); 935 936 ret = reservation_object_reserve_shared(bo->resv, 1); 937 if (unlikely(ret)) { 938 dma_fence_put(fence); 939 return ret; 940 } 941 942 dma_fence_put(bo->moving); 943 bo->moving = fence; 944 } 945 946 return 0; 947 } 948 949 /** 950 * Repeatedly evict memory from the LRU for @mem_type until we create enough 951 * space, or we've evicted everything and there isn't enough space. 952 */ 953 static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo, 954 const struct ttm_place *place, 955 struct ttm_mem_reg *mem, 956 struct ttm_operation_ctx *ctx) 957 { 958 struct ttm_bo_device *bdev = bo->bdev; 959 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type]; 960 int ret; 961 962 do { 963 ret = (*man->func->get_node)(man, bo, place, mem); 964 if (unlikely(ret != 0)) 965 return ret; 966 if (mem->mm_node) 967 break; 968 ret = ttm_mem_evict_first(bdev, mem->mem_type, place, ctx, 969 bo->resv->lock.ctx); 970 if (unlikely(ret != 0)) 971 return ret; 972 } while (1); 973 974 return ttm_bo_add_move_fence(bo, man, mem); 975 } 976 977 static uint32_t ttm_bo_select_caching(struct ttm_mem_type_manager *man, 978 uint32_t cur_placement, 979 uint32_t proposed_placement) 980 { 981 uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING; 982 uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING; 983 984 /** 985 * Keep current caching if possible. 986 */ 987 988 if ((cur_placement & caching) != 0) 989 result |= (cur_placement & caching); 990 else if ((man->default_caching & caching) != 0) 991 result |= man->default_caching; 992 else if ((TTM_PL_FLAG_CACHED & caching) != 0) 993 result |= TTM_PL_FLAG_CACHED; 994 else if ((TTM_PL_FLAG_WC & caching) != 0) 995 result |= TTM_PL_FLAG_WC; 996 else if ((TTM_PL_FLAG_UNCACHED & caching) != 0) 997 result |= TTM_PL_FLAG_UNCACHED; 998 999 return result; 1000 } 1001 1002 static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man, 1003 uint32_t mem_type, 1004 const struct ttm_place *place, 1005 uint32_t *masked_placement) 1006 { 1007 uint32_t cur_flags = ttm_bo_type_flags(mem_type); 1008 1009 if ((cur_flags & place->flags & TTM_PL_MASK_MEM) == 0) 1010 return false; 1011 1012 if ((place->flags & man->available_caching) == 0) 1013 return false; 1014 1015 cur_flags |= (place->flags & man->available_caching); 1016 1017 *masked_placement = cur_flags; 1018 return true; 1019 } 1020 1021 /** 1022 * ttm_bo_mem_placement - check if placement is compatible 1023 * @bo: BO to find memory for 1024 * @place: where to search 1025 * @mem: the memory object to fill in 1026 * @ctx: operation context 1027 * 1028 * Check if placement is compatible and fill in mem structure. 1029 * Returns -EBUSY if placement won't work or negative error code. 1030 * 0 when placement can be used. 1031 */ 1032 static int ttm_bo_mem_placement(struct ttm_buffer_object *bo, 1033 const struct ttm_place *place, 1034 struct ttm_mem_reg *mem, 1035 struct ttm_operation_ctx *ctx) 1036 { 1037 struct ttm_bo_device *bdev = bo->bdev; 1038 uint32_t mem_type = TTM_PL_SYSTEM; 1039 struct ttm_mem_type_manager *man; 1040 uint32_t cur_flags = 0; 1041 int ret; 1042 1043 ret = ttm_mem_type_from_place(place, &mem_type); 1044 if (ret) 1045 return ret; 1046 1047 man = &bdev->man[mem_type]; 1048 if (!man->has_type || !man->use_type) 1049 return -EBUSY; 1050 1051 if (!ttm_bo_mt_compatible(man, mem_type, place, &cur_flags)) 1052 return -EBUSY; 1053 1054 cur_flags = ttm_bo_select_caching(man, bo->mem.placement, cur_flags); 1055 /* 1056 * Use the access and other non-mapping-related flag bits from 1057 * the memory placement flags to the current flags 1058 */ 1059 ttm_flag_masked(&cur_flags, place->flags, ~TTM_PL_MASK_MEMTYPE); 1060 1061 mem->mem_type = mem_type; 1062 mem->placement = cur_flags; 1063 1064 if (bo->mem.mem_type < mem_type && !list_empty(&bo->lru)) { 1065 spin_lock(&bo->bdev->glob->lru_lock); 1066 ttm_bo_del_from_lru(bo); 1067 ttm_bo_add_mem_to_lru(bo, mem); 1068 spin_unlock(&bo->bdev->glob->lru_lock); 1069 } 1070 1071 return 0; 1072 } 1073 1074 /** 1075 * Creates space for memory region @mem according to its type. 1076 * 1077 * This function first searches for free space in compatible memory types in 1078 * the priority order defined by the driver. If free space isn't found, then 1079 * ttm_bo_mem_force_space is attempted in priority order to evict and find 1080 * space. 1081 */ 1082 int ttm_bo_mem_space(struct ttm_buffer_object *bo, 1083 struct ttm_placement *placement, 1084 struct ttm_mem_reg *mem, 1085 struct ttm_operation_ctx *ctx) 1086 { 1087 struct ttm_bo_device *bdev = bo->bdev; 1088 bool type_found = false; 1089 int i, ret; 1090 1091 ret = reservation_object_reserve_shared(bo->resv, 1); 1092 if (unlikely(ret)) 1093 return ret; 1094 1095 mem->mm_node = NULL; 1096 for (i = 0; i < placement->num_placement; ++i) { 1097 const struct ttm_place *place = &placement->placement[i]; 1098 struct ttm_mem_type_manager *man; 1099 1100 ret = ttm_bo_mem_placement(bo, place, mem, ctx); 1101 if (ret == -EBUSY) 1102 continue; 1103 if (ret) 1104 goto error; 1105 1106 type_found = true; 1107 mem->mm_node = NULL; 1108 if (mem->mem_type == TTM_PL_SYSTEM) 1109 return 0; 1110 1111 man = &bdev->man[mem->mem_type]; 1112 ret = (*man->func->get_node)(man, bo, place, mem); 1113 if (unlikely(ret)) 1114 goto error; 1115 1116 if (mem->mm_node) { 1117 ret = ttm_bo_add_move_fence(bo, man, mem); 1118 if (unlikely(ret)) { 1119 (*man->func->put_node)(man, mem); 1120 goto error; 1121 } 1122 return 0; 1123 } 1124 } 1125 1126 for (i = 0; i < placement->num_busy_placement; ++i) { 1127 const struct ttm_place *place = &placement->busy_placement[i]; 1128 1129 ret = ttm_bo_mem_placement(bo, place, mem, ctx); 1130 if (ret == -EBUSY) 1131 continue; 1132 if (ret) 1133 goto error; 1134 1135 type_found = true; 1136 mem->mm_node = NULL; 1137 if (mem->mem_type == TTM_PL_SYSTEM) 1138 return 0; 1139 1140 ret = ttm_bo_mem_force_space(bo, place, mem, ctx); 1141 if (ret == 0 && mem->mm_node) 1142 return 0; 1143 1144 if (ret && ret != -EBUSY) 1145 goto error; 1146 } 1147 1148 ret = -ENOMEM; 1149 if (!type_found) { 1150 pr_err(TTM_PFX "No compatible memory type found\n"); 1151 ret = -EINVAL; 1152 } 1153 1154 error: 1155 if (bo->mem.mem_type == TTM_PL_SYSTEM && !list_empty(&bo->lru)) { 1156 spin_lock(&bo->bdev->glob->lru_lock); 1157 ttm_bo_move_to_lru_tail(bo, NULL); 1158 spin_unlock(&bo->bdev->glob->lru_lock); 1159 } 1160 1161 return ret; 1162 } 1163 EXPORT_SYMBOL(ttm_bo_mem_space); 1164 1165 static int ttm_bo_move_buffer(struct ttm_buffer_object *bo, 1166 struct ttm_placement *placement, 1167 struct ttm_operation_ctx *ctx) 1168 { 1169 int ret = 0; 1170 struct ttm_mem_reg mem; 1171 1172 reservation_object_assert_held(bo->resv); 1173 1174 mem.num_pages = bo->num_pages; 1175 mem.size = mem.num_pages << PAGE_SHIFT; 1176 mem.page_alignment = bo->mem.page_alignment; 1177 mem.bus.io_reserved_vm = false; 1178 mem.bus.io_reserved_count = 0; 1179 /* 1180 * Determine where to move the buffer. 1181 */ 1182 ret = ttm_bo_mem_space(bo, placement, &mem, ctx); 1183 if (ret) 1184 goto out_unlock; 1185 ret = ttm_bo_handle_move_mem(bo, &mem, false, ctx); 1186 out_unlock: 1187 if (ret && mem.mm_node) 1188 ttm_bo_mem_put(bo, &mem); 1189 return ret; 1190 } 1191 1192 static bool ttm_bo_places_compat(const struct ttm_place *places, 1193 unsigned num_placement, 1194 struct ttm_mem_reg *mem, 1195 uint32_t *new_flags) 1196 { 1197 unsigned i; 1198 1199 for (i = 0; i < num_placement; i++) { 1200 const struct ttm_place *heap = &places[i]; 1201 1202 if (mem->mm_node && (mem->start < heap->fpfn || 1203 (heap->lpfn != 0 && (mem->start + mem->num_pages) > heap->lpfn))) 1204 continue; 1205 1206 *new_flags = heap->flags; 1207 if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) && 1208 (*new_flags & mem->placement & TTM_PL_MASK_MEM) && 1209 (!(*new_flags & TTM_PL_FLAG_CONTIGUOUS) || 1210 (mem->placement & TTM_PL_FLAG_CONTIGUOUS))) 1211 return true; 1212 } 1213 return false; 1214 } 1215 1216 bool ttm_bo_mem_compat(struct ttm_placement *placement, 1217 struct ttm_mem_reg *mem, 1218 uint32_t *new_flags) 1219 { 1220 if (ttm_bo_places_compat(placement->placement, placement->num_placement, 1221 mem, new_flags)) 1222 return true; 1223 1224 if ((placement->busy_placement != placement->placement || 1225 placement->num_busy_placement > placement->num_placement) && 1226 ttm_bo_places_compat(placement->busy_placement, 1227 placement->num_busy_placement, 1228 mem, new_flags)) 1229 return true; 1230 1231 return false; 1232 } 1233 EXPORT_SYMBOL(ttm_bo_mem_compat); 1234 1235 int ttm_bo_validate(struct ttm_buffer_object *bo, 1236 struct ttm_placement *placement, 1237 struct ttm_operation_ctx *ctx) 1238 { 1239 int ret; 1240 uint32_t new_flags; 1241 1242 reservation_object_assert_held(bo->resv); 1243 /* 1244 * Check whether we need to move buffer. 1245 */ 1246 if (!ttm_bo_mem_compat(placement, &bo->mem, &new_flags)) { 1247 ret = ttm_bo_move_buffer(bo, placement, ctx); 1248 if (ret) 1249 return ret; 1250 } else { 1251 /* 1252 * Use the access and other non-mapping-related flag bits from 1253 * the compatible memory placement flags to the active flags 1254 */ 1255 ttm_flag_masked(&bo->mem.placement, new_flags, 1256 ~TTM_PL_MASK_MEMTYPE); 1257 } 1258 /* 1259 * We might need to add a TTM. 1260 */ 1261 if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) { 1262 ret = ttm_tt_create(bo, true); 1263 if (ret) 1264 return ret; 1265 } 1266 return 0; 1267 } 1268 EXPORT_SYMBOL(ttm_bo_validate); 1269 1270 int ttm_bo_init_reserved(struct ttm_bo_device *bdev, 1271 struct ttm_buffer_object *bo, 1272 unsigned long size, 1273 enum ttm_bo_type type, 1274 struct ttm_placement *placement, 1275 uint32_t page_alignment, 1276 struct ttm_operation_ctx *ctx, 1277 size_t acc_size, 1278 struct sg_table *sg, 1279 struct reservation_object *resv, 1280 void (*destroy) (struct ttm_buffer_object *)) 1281 { 1282 int ret = 0; 1283 unsigned long num_pages; 1284 struct ttm_mem_global *mem_glob = bdev->glob->mem_glob; 1285 bool locked; 1286 1287 ret = ttm_mem_global_alloc(mem_glob, acc_size, ctx); 1288 if (ret) { 1289 pr_err("Out of kernel memory\n"); 1290 if (destroy) 1291 (*destroy)(bo); 1292 else 1293 kfree(bo); 1294 return -ENOMEM; 1295 } 1296 1297 num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT; 1298 if (num_pages == 0) { 1299 pr_err("Illegal buffer object size\n"); 1300 if (destroy) 1301 (*destroy)(bo); 1302 else 1303 kfree(bo); 1304 ttm_mem_global_free(mem_glob, acc_size); 1305 return -EINVAL; 1306 } 1307 bo->destroy = destroy ? destroy : ttm_bo_default_destroy; 1308 1309 kref_init(&bo->kref); 1310 kref_init(&bo->list_kref); 1311 atomic_set(&bo->cpu_writers, 0); 1312 INIT_LIST_HEAD(&bo->lru); 1313 INIT_LIST_HEAD(&bo->ddestroy); 1314 INIT_LIST_HEAD(&bo->swap); 1315 INIT_LIST_HEAD(&bo->io_reserve_lru); 1316 mutex_init(&bo->wu_mutex); 1317 bo->bdev = bdev; 1318 bo->type = type; 1319 bo->num_pages = num_pages; 1320 bo->mem.size = num_pages << PAGE_SHIFT; 1321 bo->mem.mem_type = TTM_PL_SYSTEM; 1322 bo->mem.num_pages = bo->num_pages; 1323 bo->mem.mm_node = NULL; 1324 bo->mem.page_alignment = page_alignment; 1325 bo->mem.bus.io_reserved_vm = false; 1326 bo->mem.bus.io_reserved_count = 0; 1327 bo->moving = NULL; 1328 bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED); 1329 bo->acc_size = acc_size; 1330 bo->sg = sg; 1331 if (resv) { 1332 bo->resv = resv; 1333 reservation_object_assert_held(bo->resv); 1334 } else { 1335 bo->resv = &bo->ttm_resv; 1336 } 1337 reservation_object_init(&bo->ttm_resv); 1338 atomic_inc(&bo->bdev->glob->bo_count); 1339 drm_vma_node_reset(&bo->vma_node); 1340 1341 /* 1342 * For ttm_bo_type_device buffers, allocate 1343 * address space from the device. 1344 */ 1345 if (bo->type == ttm_bo_type_device || 1346 bo->type == ttm_bo_type_sg) 1347 ret = drm_vma_offset_add(&bdev->vma_manager, &bo->vma_node, 1348 bo->mem.num_pages); 1349 1350 /* passed reservation objects should already be locked, 1351 * since otherwise lockdep will be angered in radeon. 1352 */ 1353 if (!resv) { 1354 locked = reservation_object_trylock(bo->resv); 1355 WARN_ON(!locked); 1356 } 1357 1358 if (likely(!ret)) 1359 ret = ttm_bo_validate(bo, placement, ctx); 1360 1361 if (unlikely(ret)) { 1362 if (!resv) 1363 ttm_bo_unreserve(bo); 1364 1365 ttm_bo_put(bo); 1366 return ret; 1367 } 1368 1369 if (resv && !(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) { 1370 spin_lock(&bdev->glob->lru_lock); 1371 ttm_bo_add_to_lru(bo); 1372 spin_unlock(&bdev->glob->lru_lock); 1373 } 1374 1375 return ret; 1376 } 1377 EXPORT_SYMBOL(ttm_bo_init_reserved); 1378 1379 int ttm_bo_init(struct ttm_bo_device *bdev, 1380 struct ttm_buffer_object *bo, 1381 unsigned long size, 1382 enum ttm_bo_type type, 1383 struct ttm_placement *placement, 1384 uint32_t page_alignment, 1385 bool interruptible, 1386 size_t acc_size, 1387 struct sg_table *sg, 1388 struct reservation_object *resv, 1389 void (*destroy) (struct ttm_buffer_object *)) 1390 { 1391 struct ttm_operation_ctx ctx = { interruptible, false }; 1392 int ret; 1393 1394 ret = ttm_bo_init_reserved(bdev, bo, size, type, placement, 1395 page_alignment, &ctx, acc_size, 1396 sg, resv, destroy); 1397 if (ret) 1398 return ret; 1399 1400 if (!resv) 1401 ttm_bo_unreserve(bo); 1402 1403 return 0; 1404 } 1405 EXPORT_SYMBOL(ttm_bo_init); 1406 1407 size_t ttm_bo_acc_size(struct ttm_bo_device *bdev, 1408 unsigned long bo_size, 1409 unsigned struct_size) 1410 { 1411 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT; 1412 size_t size = 0; 1413 1414 size += ttm_round_pot(struct_size); 1415 size += ttm_round_pot(npages * sizeof(void *)); 1416 size += ttm_round_pot(sizeof(struct ttm_tt)); 1417 return size; 1418 } 1419 EXPORT_SYMBOL(ttm_bo_acc_size); 1420 1421 size_t ttm_bo_dma_acc_size(struct ttm_bo_device *bdev, 1422 unsigned long bo_size, 1423 unsigned struct_size) 1424 { 1425 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT; 1426 size_t size = 0; 1427 1428 size += ttm_round_pot(struct_size); 1429 size += ttm_round_pot(npages * (2*sizeof(void *) + sizeof(dma_addr_t))); 1430 size += ttm_round_pot(sizeof(struct ttm_dma_tt)); 1431 return size; 1432 } 1433 EXPORT_SYMBOL(ttm_bo_dma_acc_size); 1434 1435 int ttm_bo_create(struct ttm_bo_device *bdev, 1436 unsigned long size, 1437 enum ttm_bo_type type, 1438 struct ttm_placement *placement, 1439 uint32_t page_alignment, 1440 bool interruptible, 1441 struct ttm_buffer_object **p_bo) 1442 { 1443 struct ttm_buffer_object *bo; 1444 size_t acc_size; 1445 int ret; 1446 1447 bo = kzalloc(sizeof(*bo), GFP_KERNEL); 1448 if (unlikely(bo == NULL)) 1449 return -ENOMEM; 1450 1451 acc_size = ttm_bo_acc_size(bdev, size, sizeof(struct ttm_buffer_object)); 1452 ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment, 1453 interruptible, acc_size, 1454 NULL, NULL, NULL); 1455 if (likely(ret == 0)) 1456 *p_bo = bo; 1457 1458 return ret; 1459 } 1460 EXPORT_SYMBOL(ttm_bo_create); 1461 1462 static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev, 1463 unsigned mem_type) 1464 { 1465 struct ttm_operation_ctx ctx = { 1466 .interruptible = false, 1467 .no_wait_gpu = false, 1468 .flags = TTM_OPT_FLAG_FORCE_ALLOC 1469 }; 1470 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 1471 struct ttm_bo_global *glob = bdev->glob; 1472 struct dma_fence *fence; 1473 int ret; 1474 unsigned i; 1475 1476 /* 1477 * Can't use standard list traversal since we're unlocking. 1478 */ 1479 1480 spin_lock(&glob->lru_lock); 1481 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) { 1482 while (!list_empty(&man->lru[i])) { 1483 spin_unlock(&glob->lru_lock); 1484 ret = ttm_mem_evict_first(bdev, mem_type, NULL, &ctx, 1485 NULL); 1486 if (ret) 1487 return ret; 1488 spin_lock(&glob->lru_lock); 1489 } 1490 } 1491 spin_unlock(&glob->lru_lock); 1492 1493 spin_lock(&man->move_lock); 1494 fence = dma_fence_get(man->move); 1495 spin_unlock(&man->move_lock); 1496 1497 if (fence) { 1498 ret = dma_fence_wait(fence, false); 1499 dma_fence_put(fence); 1500 if (ret) 1501 return ret; 1502 } 1503 1504 return 0; 1505 } 1506 1507 int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type) 1508 { 1509 struct ttm_mem_type_manager *man; 1510 int ret = -EINVAL; 1511 1512 if (mem_type >= TTM_NUM_MEM_TYPES) { 1513 pr_err("Illegal memory type %d\n", mem_type); 1514 return ret; 1515 } 1516 man = &bdev->man[mem_type]; 1517 1518 if (!man->has_type) { 1519 pr_err("Trying to take down uninitialized memory manager type %u\n", 1520 mem_type); 1521 return ret; 1522 } 1523 1524 man->use_type = false; 1525 man->has_type = false; 1526 1527 ret = 0; 1528 if (mem_type > 0) { 1529 ret = ttm_bo_force_list_clean(bdev, mem_type); 1530 if (ret) { 1531 pr_err("Cleanup eviction failed\n"); 1532 return ret; 1533 } 1534 1535 ret = (*man->func->takedown)(man); 1536 } 1537 1538 dma_fence_put(man->move); 1539 man->move = NULL; 1540 1541 return ret; 1542 } 1543 EXPORT_SYMBOL(ttm_bo_clean_mm); 1544 1545 int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type) 1546 { 1547 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 1548 1549 if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) { 1550 pr_err("Illegal memory manager memory type %u\n", mem_type); 1551 return -EINVAL; 1552 } 1553 1554 if (!man->has_type) { 1555 pr_err("Memory type %u has not been initialized\n", mem_type); 1556 return 0; 1557 } 1558 1559 return ttm_bo_force_list_clean(bdev, mem_type); 1560 } 1561 EXPORT_SYMBOL(ttm_bo_evict_mm); 1562 1563 int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type, 1564 unsigned long p_size) 1565 { 1566 int ret; 1567 struct ttm_mem_type_manager *man; 1568 unsigned i; 1569 1570 BUG_ON(type >= TTM_NUM_MEM_TYPES); 1571 man = &bdev->man[type]; 1572 BUG_ON(man->has_type); 1573 man->io_reserve_fastpath = true; 1574 man->use_io_reserve_lru = false; 1575 mutex_init(&man->io_reserve_mutex); 1576 spin_lock_init(&man->move_lock); 1577 INIT_LIST_HEAD(&man->io_reserve_lru); 1578 1579 ret = bdev->driver->init_mem_type(bdev, type, man); 1580 if (ret) 1581 return ret; 1582 man->bdev = bdev; 1583 1584 if (type != TTM_PL_SYSTEM) { 1585 ret = (*man->func->init)(man, p_size); 1586 if (ret) 1587 return ret; 1588 } 1589 man->has_type = true; 1590 man->use_type = true; 1591 man->size = p_size; 1592 1593 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) 1594 INIT_LIST_HEAD(&man->lru[i]); 1595 man->move = NULL; 1596 1597 return 0; 1598 } 1599 EXPORT_SYMBOL(ttm_bo_init_mm); 1600 1601 static void ttm_bo_global_kobj_release(struct kobject *kobj) 1602 { 1603 struct ttm_bo_global *glob = 1604 container_of(kobj, struct ttm_bo_global, kobj); 1605 1606 __free_page(glob->dummy_read_page); 1607 } 1608 1609 static void ttm_bo_global_release(void) 1610 { 1611 struct ttm_bo_global *glob = &ttm_bo_glob; 1612 1613 mutex_lock(&ttm_global_mutex); 1614 if (--ttm_bo_glob_use_count > 0) 1615 goto out; 1616 1617 kobject_del(&glob->kobj); 1618 kobject_put(&glob->kobj); 1619 ttm_mem_global_release(&ttm_mem_glob); 1620 memset(glob, 0, sizeof(*glob)); 1621 out: 1622 mutex_unlock(&ttm_global_mutex); 1623 } 1624 1625 static int ttm_bo_global_init(void) 1626 { 1627 struct ttm_bo_global *glob = &ttm_bo_glob; 1628 int ret = 0; 1629 unsigned i; 1630 1631 mutex_lock(&ttm_global_mutex); 1632 if (++ttm_bo_glob_use_count > 1) 1633 goto out; 1634 1635 ret = ttm_mem_global_init(&ttm_mem_glob); 1636 if (ret) 1637 goto out; 1638 1639 spin_lock_init(&glob->lru_lock); 1640 glob->mem_glob = &ttm_mem_glob; 1641 glob->mem_glob->bo_glob = glob; 1642 glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32); 1643 1644 if (unlikely(glob->dummy_read_page == NULL)) { 1645 ret = -ENOMEM; 1646 goto out; 1647 } 1648 1649 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) 1650 INIT_LIST_HEAD(&glob->swap_lru[i]); 1651 INIT_LIST_HEAD(&glob->device_list); 1652 atomic_set(&glob->bo_count, 0); 1653 1654 ret = kobject_init_and_add( 1655 &glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects"); 1656 if (unlikely(ret != 0)) 1657 kobject_put(&glob->kobj); 1658 out: 1659 mutex_unlock(&ttm_global_mutex); 1660 return ret; 1661 } 1662 1663 int ttm_bo_device_release(struct ttm_bo_device *bdev) 1664 { 1665 int ret = 0; 1666 unsigned i = TTM_NUM_MEM_TYPES; 1667 struct ttm_mem_type_manager *man; 1668 struct ttm_bo_global *glob = bdev->glob; 1669 1670 while (i--) { 1671 man = &bdev->man[i]; 1672 if (man->has_type) { 1673 man->use_type = false; 1674 if ((i != TTM_PL_SYSTEM) && ttm_bo_clean_mm(bdev, i)) { 1675 ret = -EBUSY; 1676 pr_err("DRM memory manager type %d is not clean\n", 1677 i); 1678 } 1679 man->has_type = false; 1680 } 1681 } 1682 1683 mutex_lock(&ttm_global_mutex); 1684 list_del(&bdev->device_list); 1685 mutex_unlock(&ttm_global_mutex); 1686 1687 cancel_delayed_work_sync(&bdev->wq); 1688 1689 if (ttm_bo_delayed_delete(bdev, true)) 1690 pr_debug("Delayed destroy list was clean\n"); 1691 1692 spin_lock(&glob->lru_lock); 1693 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) 1694 if (list_empty(&bdev->man[0].lru[0])) 1695 pr_debug("Swap list %d was clean\n", i); 1696 spin_unlock(&glob->lru_lock); 1697 1698 drm_vma_offset_manager_destroy(&bdev->vma_manager); 1699 1700 if (!ret) 1701 ttm_bo_global_release(); 1702 1703 return ret; 1704 } 1705 EXPORT_SYMBOL(ttm_bo_device_release); 1706 1707 int ttm_bo_device_init(struct ttm_bo_device *bdev, 1708 struct ttm_bo_driver *driver, 1709 struct address_space *mapping, 1710 bool need_dma32) 1711 { 1712 struct ttm_bo_global *glob = &ttm_bo_glob; 1713 int ret; 1714 1715 ret = ttm_bo_global_init(); 1716 if (ret) 1717 return ret; 1718 1719 bdev->driver = driver; 1720 1721 memset(bdev->man, 0, sizeof(bdev->man)); 1722 1723 /* 1724 * Initialize the system memory buffer type. 1725 * Other types need to be driver / IOCTL initialized. 1726 */ 1727 ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0); 1728 if (unlikely(ret != 0)) 1729 goto out_no_sys; 1730 1731 drm_vma_offset_manager_init(&bdev->vma_manager, 1732 DRM_FILE_PAGE_OFFSET_START, 1733 DRM_FILE_PAGE_OFFSET_SIZE); 1734 INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue); 1735 INIT_LIST_HEAD(&bdev->ddestroy); 1736 bdev->dev_mapping = mapping; 1737 bdev->glob = glob; 1738 bdev->need_dma32 = need_dma32; 1739 mutex_lock(&ttm_global_mutex); 1740 list_add_tail(&bdev->device_list, &glob->device_list); 1741 mutex_unlock(&ttm_global_mutex); 1742 1743 return 0; 1744 out_no_sys: 1745 ttm_bo_global_release(); 1746 return ret; 1747 } 1748 EXPORT_SYMBOL(ttm_bo_device_init); 1749 1750 /* 1751 * buffer object vm functions. 1752 */ 1753 1754 bool ttm_mem_reg_is_pci(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem) 1755 { 1756 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type]; 1757 1758 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) { 1759 if (mem->mem_type == TTM_PL_SYSTEM) 1760 return false; 1761 1762 if (man->flags & TTM_MEMTYPE_FLAG_CMA) 1763 return false; 1764 1765 if (mem->placement & TTM_PL_FLAG_CACHED) 1766 return false; 1767 } 1768 return true; 1769 } 1770 1771 void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo) 1772 { 1773 struct ttm_bo_device *bdev = bo->bdev; 1774 1775 drm_vma_node_unmap(&bo->vma_node, bdev->dev_mapping); 1776 ttm_mem_io_free_vm(bo); 1777 } 1778 1779 void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo) 1780 { 1781 struct ttm_bo_device *bdev = bo->bdev; 1782 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type]; 1783 1784 ttm_mem_io_lock(man, false); 1785 ttm_bo_unmap_virtual_locked(bo); 1786 ttm_mem_io_unlock(man); 1787 } 1788 1789 1790 EXPORT_SYMBOL(ttm_bo_unmap_virtual); 1791 1792 int ttm_bo_wait(struct ttm_buffer_object *bo, 1793 bool interruptible, bool no_wait) 1794 { 1795 long timeout = 15 * HZ; 1796 1797 if (no_wait) { 1798 if (reservation_object_test_signaled_rcu(bo->resv, true)) 1799 return 0; 1800 else 1801 return -EBUSY; 1802 } 1803 1804 timeout = reservation_object_wait_timeout_rcu(bo->resv, true, 1805 interruptible, timeout); 1806 if (timeout < 0) 1807 return timeout; 1808 1809 if (timeout == 0) 1810 return -EBUSY; 1811 1812 reservation_object_add_excl_fence(bo->resv, NULL); 1813 return 0; 1814 } 1815 EXPORT_SYMBOL(ttm_bo_wait); 1816 1817 int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait) 1818 { 1819 int ret = 0; 1820 1821 /* 1822 * Using ttm_bo_reserve makes sure the lru lists are updated. 1823 */ 1824 1825 ret = ttm_bo_reserve(bo, true, no_wait, NULL); 1826 if (unlikely(ret != 0)) 1827 return ret; 1828 ret = ttm_bo_wait(bo, true, no_wait); 1829 if (likely(ret == 0)) 1830 atomic_inc(&bo->cpu_writers); 1831 ttm_bo_unreserve(bo); 1832 return ret; 1833 } 1834 EXPORT_SYMBOL(ttm_bo_synccpu_write_grab); 1835 1836 void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo) 1837 { 1838 atomic_dec(&bo->cpu_writers); 1839 } 1840 EXPORT_SYMBOL(ttm_bo_synccpu_write_release); 1841 1842 /** 1843 * A buffer object shrink method that tries to swap out the first 1844 * buffer object on the bo_global::swap_lru list. 1845 */ 1846 int ttm_bo_swapout(struct ttm_bo_global *glob, struct ttm_operation_ctx *ctx) 1847 { 1848 struct ttm_buffer_object *bo; 1849 int ret = -EBUSY; 1850 bool locked; 1851 unsigned i; 1852 1853 spin_lock(&glob->lru_lock); 1854 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) { 1855 list_for_each_entry(bo, &glob->swap_lru[i], swap) { 1856 if (ttm_bo_evict_swapout_allowable(bo, ctx, &locked, 1857 NULL)) { 1858 ret = 0; 1859 break; 1860 } 1861 } 1862 if (!ret) 1863 break; 1864 } 1865 1866 if (ret) { 1867 spin_unlock(&glob->lru_lock); 1868 return ret; 1869 } 1870 1871 kref_get(&bo->list_kref); 1872 1873 if (!list_empty(&bo->ddestroy)) { 1874 ret = ttm_bo_cleanup_refs(bo, false, false, locked); 1875 kref_put(&bo->list_kref, ttm_bo_release_list); 1876 return ret; 1877 } 1878 1879 ttm_bo_del_from_lru(bo); 1880 spin_unlock(&glob->lru_lock); 1881 1882 /** 1883 * Move to system cached 1884 */ 1885 1886 if (bo->mem.mem_type != TTM_PL_SYSTEM || 1887 bo->ttm->caching_state != tt_cached) { 1888 struct ttm_operation_ctx ctx = { false, false }; 1889 struct ttm_mem_reg evict_mem; 1890 1891 evict_mem = bo->mem; 1892 evict_mem.mm_node = NULL; 1893 evict_mem.placement = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED; 1894 evict_mem.mem_type = TTM_PL_SYSTEM; 1895 1896 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, &ctx); 1897 if (unlikely(ret != 0)) 1898 goto out; 1899 } 1900 1901 /** 1902 * Make sure BO is idle. 1903 */ 1904 1905 ret = ttm_bo_wait(bo, false, false); 1906 if (unlikely(ret != 0)) 1907 goto out; 1908 1909 ttm_bo_unmap_virtual(bo); 1910 1911 /** 1912 * Swap out. Buffer will be swapped in again as soon as 1913 * anyone tries to access a ttm page. 1914 */ 1915 1916 if (bo->bdev->driver->swap_notify) 1917 bo->bdev->driver->swap_notify(bo); 1918 1919 ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage); 1920 out: 1921 1922 /** 1923 * 1924 * Unreserve without putting on LRU to avoid swapping out an 1925 * already swapped buffer. 1926 */ 1927 if (locked) 1928 reservation_object_unlock(bo->resv); 1929 kref_put(&bo->list_kref, ttm_bo_release_list); 1930 return ret; 1931 } 1932 EXPORT_SYMBOL(ttm_bo_swapout); 1933 1934 void ttm_bo_swapout_all(struct ttm_bo_device *bdev) 1935 { 1936 struct ttm_operation_ctx ctx = { 1937 .interruptible = false, 1938 .no_wait_gpu = false 1939 }; 1940 1941 while (ttm_bo_swapout(bdev->glob, &ctx) == 0) 1942 ; 1943 } 1944 EXPORT_SYMBOL(ttm_bo_swapout_all); 1945 1946 /** 1947 * ttm_bo_wait_unreserved - interruptible wait for a buffer object to become 1948 * unreserved 1949 * 1950 * @bo: Pointer to buffer 1951 */ 1952 int ttm_bo_wait_unreserved(struct ttm_buffer_object *bo) 1953 { 1954 int ret; 1955 1956 /* 1957 * In the absense of a wait_unlocked API, 1958 * Use the bo::wu_mutex to avoid triggering livelocks due to 1959 * concurrent use of this function. Note that this use of 1960 * bo::wu_mutex can go away if we change locking order to 1961 * mmap_sem -> bo::reserve. 1962 */ 1963 ret = mutex_lock_interruptible(&bo->wu_mutex); 1964 if (unlikely(ret != 0)) 1965 return -ERESTARTSYS; 1966 if (!ww_mutex_is_locked(&bo->resv->lock)) 1967 goto out_unlock; 1968 ret = reservation_object_lock_interruptible(bo->resv, NULL); 1969 if (ret == -EINTR) 1970 ret = -ERESTARTSYS; 1971 if (unlikely(ret != 0)) 1972 goto out_unlock; 1973 reservation_object_unlock(bo->resv); 1974 1975 out_unlock: 1976 mutex_unlock(&bo->wu_mutex); 1977 return ret; 1978 } 1979