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 static struct attribute ttm_bo_count = { 49 .name = "bo_count", 50 .mode = S_IRUGO 51 }; 52 53 /* default destructor */ 54 static void ttm_bo_default_destroy(struct ttm_buffer_object *bo) 55 { 56 kfree(bo); 57 } 58 59 static inline int ttm_mem_type_from_place(const struct ttm_place *place, 60 uint32_t *mem_type) 61 { 62 int pos; 63 64 pos = ffs(place->flags & TTM_PL_MASK_MEM); 65 if (unlikely(!pos)) 66 return -EINVAL; 67 68 *mem_type = pos - 1; 69 return 0; 70 } 71 72 static void ttm_mem_type_debug(struct ttm_bo_device *bdev, int mem_type) 73 { 74 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 75 struct drm_printer p = drm_debug_printer(TTM_PFX); 76 77 pr_err(" has_type: %d\n", man->has_type); 78 pr_err(" use_type: %d\n", man->use_type); 79 pr_err(" flags: 0x%08X\n", man->flags); 80 pr_err(" gpu_offset: 0x%08llX\n", man->gpu_offset); 81 pr_err(" size: %llu\n", man->size); 82 pr_err(" available_caching: 0x%08X\n", man->available_caching); 83 pr_err(" default_caching: 0x%08X\n", man->default_caching); 84 if (mem_type != TTM_PL_SYSTEM) 85 (*man->func->debug)(man, &p); 86 } 87 88 static void ttm_bo_mem_space_debug(struct ttm_buffer_object *bo, 89 struct ttm_placement *placement) 90 { 91 int i, ret, mem_type; 92 93 pr_err("No space for %p (%lu pages, %luK, %luM)\n", 94 bo, bo->mem.num_pages, bo->mem.size >> 10, 95 bo->mem.size >> 20); 96 for (i = 0; i < placement->num_placement; i++) { 97 ret = ttm_mem_type_from_place(&placement->placement[i], 98 &mem_type); 99 if (ret) 100 return; 101 pr_err(" placement[%d]=0x%08X (%d)\n", 102 i, placement->placement[i].flags, mem_type); 103 ttm_mem_type_debug(bo->bdev, mem_type); 104 } 105 } 106 107 static ssize_t ttm_bo_global_show(struct kobject *kobj, 108 struct attribute *attr, 109 char *buffer) 110 { 111 struct ttm_bo_global *glob = 112 container_of(kobj, struct ttm_bo_global, kobj); 113 114 return snprintf(buffer, PAGE_SIZE, "%d\n", 115 atomic_read(&glob->bo_count)); 116 } 117 118 static struct attribute *ttm_bo_global_attrs[] = { 119 &ttm_bo_count, 120 NULL 121 }; 122 123 static const struct sysfs_ops ttm_bo_global_ops = { 124 .show = &ttm_bo_global_show 125 }; 126 127 static struct kobj_type ttm_bo_glob_kobj_type = { 128 .release = &ttm_bo_global_kobj_release, 129 .sysfs_ops = &ttm_bo_global_ops, 130 .default_attrs = ttm_bo_global_attrs 131 }; 132 133 134 static inline uint32_t ttm_bo_type_flags(unsigned type) 135 { 136 return 1 << (type); 137 } 138 139 static void ttm_bo_release_list(struct kref *list_kref) 140 { 141 struct ttm_buffer_object *bo = 142 container_of(list_kref, struct ttm_buffer_object, list_kref); 143 struct ttm_bo_device *bdev = bo->bdev; 144 size_t acc_size = bo->acc_size; 145 146 BUG_ON(kref_read(&bo->list_kref)); 147 BUG_ON(kref_read(&bo->kref)); 148 BUG_ON(atomic_read(&bo->cpu_writers)); 149 BUG_ON(bo->mem.mm_node != NULL); 150 BUG_ON(!list_empty(&bo->lru)); 151 BUG_ON(!list_empty(&bo->ddestroy)); 152 ttm_tt_destroy(bo->ttm); 153 atomic_dec(&bo->bdev->glob->bo_count); 154 dma_fence_put(bo->moving); 155 reservation_object_fini(&bo->ttm_resv); 156 mutex_destroy(&bo->wu_mutex); 157 bo->destroy(bo); 158 ttm_mem_global_free(bdev->glob->mem_glob, acc_size); 159 } 160 161 void ttm_bo_add_to_lru(struct ttm_buffer_object *bo) 162 { 163 struct ttm_bo_device *bdev = bo->bdev; 164 struct ttm_mem_type_manager *man; 165 166 reservation_object_assert_held(bo->resv); 167 168 if (!(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) { 169 BUG_ON(!list_empty(&bo->lru)); 170 171 man = &bdev->man[bo->mem.mem_type]; 172 list_add_tail(&bo->lru, &man->lru[bo->priority]); 173 kref_get(&bo->list_kref); 174 175 if (bo->ttm && !(bo->ttm->page_flags & 176 (TTM_PAGE_FLAG_SG | TTM_PAGE_FLAG_SWAPPED))) { 177 list_add_tail(&bo->swap, 178 &bdev->glob->swap_lru[bo->priority]); 179 kref_get(&bo->list_kref); 180 } 181 } 182 } 183 EXPORT_SYMBOL(ttm_bo_add_to_lru); 184 185 static void ttm_bo_ref_bug(struct kref *list_kref) 186 { 187 BUG(); 188 } 189 190 void ttm_bo_del_from_lru(struct ttm_buffer_object *bo) 191 { 192 if (!list_empty(&bo->swap)) { 193 list_del_init(&bo->swap); 194 kref_put(&bo->list_kref, ttm_bo_ref_bug); 195 } 196 if (!list_empty(&bo->lru)) { 197 list_del_init(&bo->lru); 198 kref_put(&bo->list_kref, ttm_bo_ref_bug); 199 } 200 201 /* 202 * TODO: Add a driver hook to delete from 203 * driver-specific LRU's here. 204 */ 205 } 206 207 void ttm_bo_del_sub_from_lru(struct ttm_buffer_object *bo) 208 { 209 struct ttm_bo_global *glob = bo->bdev->glob; 210 211 spin_lock(&glob->lru_lock); 212 ttm_bo_del_from_lru(bo); 213 spin_unlock(&glob->lru_lock); 214 } 215 EXPORT_SYMBOL(ttm_bo_del_sub_from_lru); 216 217 static void ttm_bo_bulk_move_set_pos(struct ttm_lru_bulk_move_pos *pos, 218 struct ttm_buffer_object *bo) 219 { 220 if (!pos->first) 221 pos->first = bo; 222 pos->last = bo; 223 } 224 225 void ttm_bo_move_to_lru_tail(struct ttm_buffer_object *bo, 226 struct ttm_lru_bulk_move *bulk) 227 { 228 reservation_object_assert_held(bo->resv); 229 230 ttm_bo_del_from_lru(bo); 231 ttm_bo_add_to_lru(bo); 232 233 if (bulk && !(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) { 234 switch (bo->mem.mem_type) { 235 case TTM_PL_TT: 236 ttm_bo_bulk_move_set_pos(&bulk->tt[bo->priority], bo); 237 break; 238 239 case TTM_PL_VRAM: 240 ttm_bo_bulk_move_set_pos(&bulk->vram[bo->priority], bo); 241 break; 242 } 243 if (bo->ttm && !(bo->ttm->page_flags & 244 (TTM_PAGE_FLAG_SG | TTM_PAGE_FLAG_SWAPPED))) 245 ttm_bo_bulk_move_set_pos(&bulk->swap[bo->priority], bo); 246 } 247 } 248 EXPORT_SYMBOL(ttm_bo_move_to_lru_tail); 249 250 static void ttm_list_move_bulk_tail(struct list_head *list, 251 struct list_head *first, 252 struct list_head *last) 253 { 254 first->prev->next = last->next; 255 last->next->prev = first->prev; 256 257 list->prev->next = first; 258 first->prev = list->prev; 259 260 last->next = list; 261 list->prev = last; 262 } 263 264 void ttm_bo_bulk_move_lru_tail(struct ttm_lru_bulk_move *bulk) 265 { 266 unsigned i; 267 268 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) { 269 struct ttm_lru_bulk_move_pos *pos = &bulk->tt[i]; 270 struct ttm_mem_type_manager *man; 271 272 if (!pos->first) 273 continue; 274 275 reservation_object_assert_held(pos->first->resv); 276 reservation_object_assert_held(pos->last->resv); 277 278 man = &pos->first->bdev->man[TTM_PL_TT]; 279 ttm_list_move_bulk_tail(&man->lru[i], &pos->first->lru, 280 &pos->last->lru); 281 } 282 283 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) { 284 struct ttm_lru_bulk_move_pos *pos = &bulk->vram[i]; 285 struct ttm_mem_type_manager *man; 286 287 if (!pos->first) 288 continue; 289 290 reservation_object_assert_held(pos->first->resv); 291 reservation_object_assert_held(pos->last->resv); 292 293 man = &pos->first->bdev->man[TTM_PL_VRAM]; 294 ttm_list_move_bulk_tail(&man->lru[i], &pos->first->lru, 295 &pos->last->lru); 296 } 297 298 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) { 299 struct ttm_lru_bulk_move_pos *pos = &bulk->swap[i]; 300 struct list_head *lru; 301 302 if (!pos->first) 303 continue; 304 305 reservation_object_assert_held(pos->first->resv); 306 reservation_object_assert_held(pos->last->resv); 307 308 lru = &pos->first->bdev->glob->swap_lru[i]; 309 ttm_list_move_bulk_tail(lru, &pos->first->swap, 310 &pos->last->swap); 311 } 312 } 313 EXPORT_SYMBOL(ttm_bo_bulk_move_lru_tail); 314 315 static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo, 316 struct ttm_mem_reg *mem, bool evict, 317 struct ttm_operation_ctx *ctx) 318 { 319 struct ttm_bo_device *bdev = bo->bdev; 320 bool old_is_pci = ttm_mem_reg_is_pci(bdev, &bo->mem); 321 bool new_is_pci = ttm_mem_reg_is_pci(bdev, mem); 322 struct ttm_mem_type_manager *old_man = &bdev->man[bo->mem.mem_type]; 323 struct ttm_mem_type_manager *new_man = &bdev->man[mem->mem_type]; 324 int ret = 0; 325 326 if (old_is_pci || new_is_pci || 327 ((mem->placement & bo->mem.placement & TTM_PL_MASK_CACHING) == 0)) { 328 ret = ttm_mem_io_lock(old_man, true); 329 if (unlikely(ret != 0)) 330 goto out_err; 331 ttm_bo_unmap_virtual_locked(bo); 332 ttm_mem_io_unlock(old_man); 333 } 334 335 /* 336 * Create and bind a ttm if required. 337 */ 338 339 if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) { 340 if (bo->ttm == NULL) { 341 bool zero = !(old_man->flags & TTM_MEMTYPE_FLAG_FIXED); 342 ret = ttm_tt_create(bo, zero); 343 if (ret) 344 goto out_err; 345 } 346 347 ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement); 348 if (ret) 349 goto out_err; 350 351 if (mem->mem_type != TTM_PL_SYSTEM) { 352 ret = ttm_tt_bind(bo->ttm, mem, ctx); 353 if (ret) 354 goto out_err; 355 } 356 357 if (bo->mem.mem_type == TTM_PL_SYSTEM) { 358 if (bdev->driver->move_notify) 359 bdev->driver->move_notify(bo, evict, mem); 360 bo->mem = *mem; 361 mem->mm_node = NULL; 362 goto moved; 363 } 364 } 365 366 if (bdev->driver->move_notify) 367 bdev->driver->move_notify(bo, evict, mem); 368 369 if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) && 370 !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) 371 ret = ttm_bo_move_ttm(bo, ctx, mem); 372 else if (bdev->driver->move) 373 ret = bdev->driver->move(bo, evict, ctx, mem); 374 else 375 ret = ttm_bo_move_memcpy(bo, ctx, mem); 376 377 if (ret) { 378 if (bdev->driver->move_notify) { 379 swap(*mem, bo->mem); 380 bdev->driver->move_notify(bo, false, mem); 381 swap(*mem, bo->mem); 382 } 383 384 goto out_err; 385 } 386 387 moved: 388 if (bo->evicted) { 389 if (bdev->driver->invalidate_caches) { 390 ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement); 391 if (ret) 392 pr_err("Can not flush read caches\n"); 393 } 394 bo->evicted = false; 395 } 396 397 if (bo->mem.mm_node) 398 bo->offset = (bo->mem.start << PAGE_SHIFT) + 399 bdev->man[bo->mem.mem_type].gpu_offset; 400 else 401 bo->offset = 0; 402 403 ctx->bytes_moved += bo->num_pages << PAGE_SHIFT; 404 return 0; 405 406 out_err: 407 new_man = &bdev->man[bo->mem.mem_type]; 408 if (new_man->flags & TTM_MEMTYPE_FLAG_FIXED) { 409 ttm_tt_destroy(bo->ttm); 410 bo->ttm = NULL; 411 } 412 413 return ret; 414 } 415 416 /** 417 * Call bo::reserved. 418 * Will release GPU memory type usage on destruction. 419 * This is the place to put in driver specific hooks to release 420 * driver private resources. 421 * Will release the bo::reserved lock. 422 */ 423 424 static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo) 425 { 426 if (bo->bdev->driver->move_notify) 427 bo->bdev->driver->move_notify(bo, false, NULL); 428 429 ttm_tt_destroy(bo->ttm); 430 bo->ttm = NULL; 431 ttm_bo_mem_put(bo, &bo->mem); 432 } 433 434 static int ttm_bo_individualize_resv(struct ttm_buffer_object *bo) 435 { 436 int r; 437 438 if (bo->resv == &bo->ttm_resv) 439 return 0; 440 441 BUG_ON(!reservation_object_trylock(&bo->ttm_resv)); 442 443 r = reservation_object_copy_fences(&bo->ttm_resv, bo->resv); 444 if (r) 445 reservation_object_unlock(&bo->ttm_resv); 446 447 return r; 448 } 449 450 static void ttm_bo_flush_all_fences(struct ttm_buffer_object *bo) 451 { 452 struct reservation_object_list *fobj; 453 struct dma_fence *fence; 454 int i; 455 456 fobj = reservation_object_get_list(&bo->ttm_resv); 457 fence = reservation_object_get_excl(&bo->ttm_resv); 458 if (fence && !fence->ops->signaled) 459 dma_fence_enable_sw_signaling(fence); 460 461 for (i = 0; fobj && i < fobj->shared_count; ++i) { 462 fence = rcu_dereference_protected(fobj->shared[i], 463 reservation_object_held(bo->resv)); 464 465 if (!fence->ops->signaled) 466 dma_fence_enable_sw_signaling(fence); 467 } 468 } 469 470 static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo) 471 { 472 struct ttm_bo_device *bdev = bo->bdev; 473 struct ttm_bo_global *glob = bdev->glob; 474 int ret; 475 476 ret = ttm_bo_individualize_resv(bo); 477 if (ret) { 478 /* Last resort, if we fail to allocate memory for the 479 * fences block for the BO to become idle 480 */ 481 reservation_object_wait_timeout_rcu(bo->resv, true, false, 482 30 * HZ); 483 spin_lock(&glob->lru_lock); 484 goto error; 485 } 486 487 spin_lock(&glob->lru_lock); 488 ret = reservation_object_trylock(bo->resv) ? 0 : -EBUSY; 489 if (!ret) { 490 if (reservation_object_test_signaled_rcu(&bo->ttm_resv, true)) { 491 ttm_bo_del_from_lru(bo); 492 spin_unlock(&glob->lru_lock); 493 if (bo->resv != &bo->ttm_resv) 494 reservation_object_unlock(&bo->ttm_resv); 495 496 ttm_bo_cleanup_memtype_use(bo); 497 reservation_object_unlock(bo->resv); 498 return; 499 } 500 501 ttm_bo_flush_all_fences(bo); 502 503 /* 504 * Make NO_EVICT bos immediately available to 505 * shrinkers, now that they are queued for 506 * destruction. 507 */ 508 if (bo->mem.placement & TTM_PL_FLAG_NO_EVICT) { 509 bo->mem.placement &= ~TTM_PL_FLAG_NO_EVICT; 510 ttm_bo_add_to_lru(bo); 511 } 512 513 reservation_object_unlock(bo->resv); 514 } 515 if (bo->resv != &bo->ttm_resv) 516 reservation_object_unlock(&bo->ttm_resv); 517 518 error: 519 kref_get(&bo->list_kref); 520 list_add_tail(&bo->ddestroy, &bdev->ddestroy); 521 spin_unlock(&glob->lru_lock); 522 523 schedule_delayed_work(&bdev->wq, 524 ((HZ / 100) < 1) ? 1 : HZ / 100); 525 } 526 527 /** 528 * function ttm_bo_cleanup_refs 529 * If bo idle, remove from delayed- and lru lists, and unref. 530 * If not idle, do nothing. 531 * 532 * Must be called with lru_lock and reservation held, this function 533 * will drop the lru lock and optionally the reservation lock before returning. 534 * 535 * @interruptible Any sleeps should occur interruptibly. 536 * @no_wait_gpu Never wait for gpu. Return -EBUSY instead. 537 * @unlock_resv Unlock the reservation lock as well. 538 */ 539 540 static int ttm_bo_cleanup_refs(struct ttm_buffer_object *bo, 541 bool interruptible, bool no_wait_gpu, 542 bool unlock_resv) 543 { 544 struct ttm_bo_global *glob = bo->bdev->glob; 545 struct reservation_object *resv; 546 int ret; 547 548 if (unlikely(list_empty(&bo->ddestroy))) 549 resv = bo->resv; 550 else 551 resv = &bo->ttm_resv; 552 553 if (reservation_object_test_signaled_rcu(resv, true)) 554 ret = 0; 555 else 556 ret = -EBUSY; 557 558 if (ret && !no_wait_gpu) { 559 long lret; 560 561 if (unlock_resv) 562 reservation_object_unlock(bo->resv); 563 spin_unlock(&glob->lru_lock); 564 565 lret = reservation_object_wait_timeout_rcu(resv, true, 566 interruptible, 567 30 * HZ); 568 569 if (lret < 0) 570 return lret; 571 else if (lret == 0) 572 return -EBUSY; 573 574 spin_lock(&glob->lru_lock); 575 if (unlock_resv && !reservation_object_trylock(bo->resv)) { 576 /* 577 * We raced, and lost, someone else holds the reservation now, 578 * and is probably busy in ttm_bo_cleanup_memtype_use. 579 * 580 * Even if it's not the case, because we finished waiting any 581 * delayed destruction would succeed, so just return success 582 * here. 583 */ 584 spin_unlock(&glob->lru_lock); 585 return 0; 586 } 587 ret = 0; 588 } 589 590 if (ret || unlikely(list_empty(&bo->ddestroy))) { 591 if (unlock_resv) 592 reservation_object_unlock(bo->resv); 593 spin_unlock(&glob->lru_lock); 594 return ret; 595 } 596 597 ttm_bo_del_from_lru(bo); 598 list_del_init(&bo->ddestroy); 599 kref_put(&bo->list_kref, ttm_bo_ref_bug); 600 601 spin_unlock(&glob->lru_lock); 602 ttm_bo_cleanup_memtype_use(bo); 603 604 if (unlock_resv) 605 reservation_object_unlock(bo->resv); 606 607 return 0; 608 } 609 610 /** 611 * Traverse the delayed list, and call ttm_bo_cleanup_refs on all 612 * encountered buffers. 613 */ 614 static bool ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all) 615 { 616 struct ttm_bo_global *glob = bdev->glob; 617 struct list_head removed; 618 bool empty; 619 620 INIT_LIST_HEAD(&removed); 621 622 spin_lock(&glob->lru_lock); 623 while (!list_empty(&bdev->ddestroy)) { 624 struct ttm_buffer_object *bo; 625 626 bo = list_first_entry(&bdev->ddestroy, struct ttm_buffer_object, 627 ddestroy); 628 kref_get(&bo->list_kref); 629 list_move_tail(&bo->ddestroy, &removed); 630 631 if (remove_all || bo->resv != &bo->ttm_resv) { 632 spin_unlock(&glob->lru_lock); 633 reservation_object_lock(bo->resv, NULL); 634 635 spin_lock(&glob->lru_lock); 636 ttm_bo_cleanup_refs(bo, false, !remove_all, true); 637 638 } else if (reservation_object_trylock(bo->resv)) { 639 ttm_bo_cleanup_refs(bo, false, !remove_all, true); 640 } else { 641 spin_unlock(&glob->lru_lock); 642 } 643 644 kref_put(&bo->list_kref, ttm_bo_release_list); 645 spin_lock(&glob->lru_lock); 646 } 647 list_splice_tail(&removed, &bdev->ddestroy); 648 empty = list_empty(&bdev->ddestroy); 649 spin_unlock(&glob->lru_lock); 650 651 return empty; 652 } 653 654 static void ttm_bo_delayed_workqueue(struct work_struct *work) 655 { 656 struct ttm_bo_device *bdev = 657 container_of(work, struct ttm_bo_device, wq.work); 658 659 if (!ttm_bo_delayed_delete(bdev, false)) 660 schedule_delayed_work(&bdev->wq, 661 ((HZ / 100) < 1) ? 1 : HZ / 100); 662 } 663 664 static void ttm_bo_release(struct kref *kref) 665 { 666 struct ttm_buffer_object *bo = 667 container_of(kref, struct ttm_buffer_object, kref); 668 struct ttm_bo_device *bdev = bo->bdev; 669 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type]; 670 671 drm_vma_offset_remove(&bdev->vma_manager, &bo->vma_node); 672 ttm_mem_io_lock(man, false); 673 ttm_mem_io_free_vm(bo); 674 ttm_mem_io_unlock(man); 675 ttm_bo_cleanup_refs_or_queue(bo); 676 kref_put(&bo->list_kref, ttm_bo_release_list); 677 } 678 679 void ttm_bo_put(struct ttm_buffer_object *bo) 680 { 681 kref_put(&bo->kref, ttm_bo_release); 682 } 683 EXPORT_SYMBOL(ttm_bo_put); 684 685 void ttm_bo_unref(struct ttm_buffer_object **p_bo) 686 { 687 struct ttm_buffer_object *bo = *p_bo; 688 689 *p_bo = NULL; 690 ttm_bo_put(bo); 691 } 692 EXPORT_SYMBOL(ttm_bo_unref); 693 694 int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev) 695 { 696 return cancel_delayed_work_sync(&bdev->wq); 697 } 698 EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue); 699 700 void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched) 701 { 702 if (resched) 703 schedule_delayed_work(&bdev->wq, 704 ((HZ / 100) < 1) ? 1 : HZ / 100); 705 } 706 EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue); 707 708 static int ttm_bo_evict(struct ttm_buffer_object *bo, 709 struct ttm_operation_ctx *ctx) 710 { 711 struct ttm_bo_device *bdev = bo->bdev; 712 struct ttm_mem_reg evict_mem; 713 struct ttm_placement placement; 714 int ret = 0; 715 716 reservation_object_assert_held(bo->resv); 717 718 placement.num_placement = 0; 719 placement.num_busy_placement = 0; 720 bdev->driver->evict_flags(bo, &placement); 721 722 if (!placement.num_placement && !placement.num_busy_placement) { 723 ret = ttm_bo_pipeline_gutting(bo); 724 if (ret) 725 return ret; 726 727 return ttm_tt_create(bo, false); 728 } 729 730 evict_mem = bo->mem; 731 evict_mem.mm_node = NULL; 732 evict_mem.bus.io_reserved_vm = false; 733 evict_mem.bus.io_reserved_count = 0; 734 735 ret = ttm_bo_mem_space(bo, &placement, &evict_mem, ctx); 736 if (ret) { 737 if (ret != -ERESTARTSYS) { 738 pr_err("Failed to find memory space for buffer 0x%p eviction\n", 739 bo); 740 ttm_bo_mem_space_debug(bo, &placement); 741 } 742 goto out; 743 } 744 745 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, ctx); 746 if (unlikely(ret)) { 747 if (ret != -ERESTARTSYS) 748 pr_err("Buffer eviction failed\n"); 749 ttm_bo_mem_put(bo, &evict_mem); 750 goto out; 751 } 752 bo->evicted = true; 753 out: 754 return ret; 755 } 756 757 bool ttm_bo_eviction_valuable(struct ttm_buffer_object *bo, 758 const struct ttm_place *place) 759 { 760 /* Don't evict this BO if it's outside of the 761 * requested placement range 762 */ 763 if (place->fpfn >= (bo->mem.start + bo->mem.size) || 764 (place->lpfn && place->lpfn <= bo->mem.start)) 765 return false; 766 767 return true; 768 } 769 EXPORT_SYMBOL(ttm_bo_eviction_valuable); 770 771 /** 772 * Check the target bo is allowable to be evicted or swapout, including cases: 773 * 774 * a. if share same reservation object with ctx->resv, have assumption 775 * reservation objects should already be locked, so not lock again and 776 * return true directly when either the opreation allow_reserved_eviction 777 * or the target bo already is in delayed free list; 778 * 779 * b. Otherwise, trylock it. 780 */ 781 static bool ttm_bo_evict_swapout_allowable(struct ttm_buffer_object *bo, 782 struct ttm_operation_ctx *ctx, bool *locked) 783 { 784 bool ret = false; 785 786 *locked = false; 787 if (bo->resv == ctx->resv) { 788 reservation_object_assert_held(bo->resv); 789 if (ctx->flags & TTM_OPT_FLAG_ALLOW_RES_EVICT 790 || !list_empty(&bo->ddestroy)) 791 ret = true; 792 } else { 793 *locked = reservation_object_trylock(bo->resv); 794 ret = *locked; 795 } 796 797 return ret; 798 } 799 800 static int ttm_mem_evict_first(struct ttm_bo_device *bdev, 801 uint32_t mem_type, 802 const struct ttm_place *place, 803 struct ttm_operation_ctx *ctx) 804 { 805 struct ttm_bo_global *glob = bdev->glob; 806 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 807 struct ttm_buffer_object *bo = NULL; 808 bool locked = false; 809 unsigned i; 810 int ret; 811 812 spin_lock(&glob->lru_lock); 813 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) { 814 list_for_each_entry(bo, &man->lru[i], lru) { 815 if (!ttm_bo_evict_swapout_allowable(bo, ctx, &locked)) 816 continue; 817 818 if (place && !bdev->driver->eviction_valuable(bo, 819 place)) { 820 if (locked) 821 reservation_object_unlock(bo->resv); 822 continue; 823 } 824 break; 825 } 826 827 /* If the inner loop terminated early, we have our candidate */ 828 if (&bo->lru != &man->lru[i]) 829 break; 830 831 bo = NULL; 832 } 833 834 if (!bo) { 835 spin_unlock(&glob->lru_lock); 836 return -EBUSY; 837 } 838 839 kref_get(&bo->list_kref); 840 841 if (!list_empty(&bo->ddestroy)) { 842 ret = ttm_bo_cleanup_refs(bo, ctx->interruptible, 843 ctx->no_wait_gpu, locked); 844 kref_put(&bo->list_kref, ttm_bo_release_list); 845 return ret; 846 } 847 848 ttm_bo_del_from_lru(bo); 849 spin_unlock(&glob->lru_lock); 850 851 ret = ttm_bo_evict(bo, ctx); 852 if (locked) { 853 ttm_bo_unreserve(bo); 854 } else { 855 spin_lock(&glob->lru_lock); 856 ttm_bo_add_to_lru(bo); 857 spin_unlock(&glob->lru_lock); 858 } 859 860 kref_put(&bo->list_kref, ttm_bo_release_list); 861 return ret; 862 } 863 864 void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem) 865 { 866 struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type]; 867 868 if (mem->mm_node) 869 (*man->func->put_node)(man, mem); 870 } 871 EXPORT_SYMBOL(ttm_bo_mem_put); 872 873 /** 874 * Add the last move fence to the BO and reserve a new shared slot. 875 */ 876 static int ttm_bo_add_move_fence(struct ttm_buffer_object *bo, 877 struct ttm_mem_type_manager *man, 878 struct ttm_mem_reg *mem) 879 { 880 struct dma_fence *fence; 881 int ret; 882 883 spin_lock(&man->move_lock); 884 fence = dma_fence_get(man->move); 885 spin_unlock(&man->move_lock); 886 887 if (fence) { 888 reservation_object_add_shared_fence(bo->resv, fence); 889 890 ret = reservation_object_reserve_shared(bo->resv); 891 if (unlikely(ret)) 892 return ret; 893 894 dma_fence_put(bo->moving); 895 bo->moving = fence; 896 } 897 898 return 0; 899 } 900 901 /** 902 * Repeatedly evict memory from the LRU for @mem_type until we create enough 903 * space, or we've evicted everything and there isn't enough space. 904 */ 905 static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo, 906 uint32_t mem_type, 907 const struct ttm_place *place, 908 struct ttm_mem_reg *mem, 909 struct ttm_operation_ctx *ctx) 910 { 911 struct ttm_bo_device *bdev = bo->bdev; 912 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 913 int ret; 914 915 do { 916 ret = (*man->func->get_node)(man, bo, place, mem); 917 if (unlikely(ret != 0)) 918 return ret; 919 if (mem->mm_node) 920 break; 921 ret = ttm_mem_evict_first(bdev, mem_type, place, ctx); 922 if (unlikely(ret != 0)) 923 return ret; 924 } while (1); 925 mem->mem_type = mem_type; 926 return ttm_bo_add_move_fence(bo, man, mem); 927 } 928 929 static uint32_t ttm_bo_select_caching(struct ttm_mem_type_manager *man, 930 uint32_t cur_placement, 931 uint32_t proposed_placement) 932 { 933 uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING; 934 uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING; 935 936 /** 937 * Keep current caching if possible. 938 */ 939 940 if ((cur_placement & caching) != 0) 941 result |= (cur_placement & caching); 942 else if ((man->default_caching & caching) != 0) 943 result |= man->default_caching; 944 else if ((TTM_PL_FLAG_CACHED & caching) != 0) 945 result |= TTM_PL_FLAG_CACHED; 946 else if ((TTM_PL_FLAG_WC & caching) != 0) 947 result |= TTM_PL_FLAG_WC; 948 else if ((TTM_PL_FLAG_UNCACHED & caching) != 0) 949 result |= TTM_PL_FLAG_UNCACHED; 950 951 return result; 952 } 953 954 static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man, 955 uint32_t mem_type, 956 const struct ttm_place *place, 957 uint32_t *masked_placement) 958 { 959 uint32_t cur_flags = ttm_bo_type_flags(mem_type); 960 961 if ((cur_flags & place->flags & TTM_PL_MASK_MEM) == 0) 962 return false; 963 964 if ((place->flags & man->available_caching) == 0) 965 return false; 966 967 cur_flags |= (place->flags & man->available_caching); 968 969 *masked_placement = cur_flags; 970 return true; 971 } 972 973 /** 974 * Creates space for memory region @mem according to its type. 975 * 976 * This function first searches for free space in compatible memory types in 977 * the priority order defined by the driver. If free space isn't found, then 978 * ttm_bo_mem_force_space is attempted in priority order to evict and find 979 * space. 980 */ 981 int ttm_bo_mem_space(struct ttm_buffer_object *bo, 982 struct ttm_placement *placement, 983 struct ttm_mem_reg *mem, 984 struct ttm_operation_ctx *ctx) 985 { 986 struct ttm_bo_device *bdev = bo->bdev; 987 struct ttm_mem_type_manager *man; 988 uint32_t mem_type = TTM_PL_SYSTEM; 989 uint32_t cur_flags = 0; 990 bool type_found = false; 991 bool type_ok = false; 992 bool has_erestartsys = false; 993 int i, ret; 994 995 ret = reservation_object_reserve_shared(bo->resv); 996 if (unlikely(ret)) 997 return ret; 998 999 mem->mm_node = NULL; 1000 for (i = 0; i < placement->num_placement; ++i) { 1001 const struct ttm_place *place = &placement->placement[i]; 1002 1003 ret = ttm_mem_type_from_place(place, &mem_type); 1004 if (ret) 1005 return ret; 1006 man = &bdev->man[mem_type]; 1007 if (!man->has_type || !man->use_type) 1008 continue; 1009 1010 type_ok = ttm_bo_mt_compatible(man, mem_type, place, 1011 &cur_flags); 1012 1013 if (!type_ok) 1014 continue; 1015 1016 type_found = true; 1017 cur_flags = ttm_bo_select_caching(man, bo->mem.placement, 1018 cur_flags); 1019 /* 1020 * Use the access and other non-mapping-related flag bits from 1021 * the memory placement flags to the current flags 1022 */ 1023 ttm_flag_masked(&cur_flags, place->flags, 1024 ~TTM_PL_MASK_MEMTYPE); 1025 1026 if (mem_type == TTM_PL_SYSTEM) 1027 break; 1028 1029 ret = (*man->func->get_node)(man, bo, place, mem); 1030 if (unlikely(ret)) 1031 return ret; 1032 1033 if (mem->mm_node) { 1034 ret = ttm_bo_add_move_fence(bo, man, mem); 1035 if (unlikely(ret)) { 1036 (*man->func->put_node)(man, mem); 1037 return ret; 1038 } 1039 break; 1040 } 1041 } 1042 1043 if ((type_ok && (mem_type == TTM_PL_SYSTEM)) || mem->mm_node) { 1044 mem->mem_type = mem_type; 1045 mem->placement = cur_flags; 1046 return 0; 1047 } 1048 1049 for (i = 0; i < placement->num_busy_placement; ++i) { 1050 const struct ttm_place *place = &placement->busy_placement[i]; 1051 1052 ret = ttm_mem_type_from_place(place, &mem_type); 1053 if (ret) 1054 return ret; 1055 man = &bdev->man[mem_type]; 1056 if (!man->has_type || !man->use_type) 1057 continue; 1058 if (!ttm_bo_mt_compatible(man, mem_type, place, &cur_flags)) 1059 continue; 1060 1061 type_found = true; 1062 cur_flags = ttm_bo_select_caching(man, bo->mem.placement, 1063 cur_flags); 1064 /* 1065 * Use the access and other non-mapping-related flag bits from 1066 * the memory placement flags to the current flags 1067 */ 1068 ttm_flag_masked(&cur_flags, place->flags, 1069 ~TTM_PL_MASK_MEMTYPE); 1070 1071 if (mem_type == TTM_PL_SYSTEM) { 1072 mem->mem_type = mem_type; 1073 mem->placement = cur_flags; 1074 mem->mm_node = NULL; 1075 return 0; 1076 } 1077 1078 ret = ttm_bo_mem_force_space(bo, mem_type, place, mem, ctx); 1079 if (ret == 0 && mem->mm_node) { 1080 mem->placement = cur_flags; 1081 return 0; 1082 } 1083 if (ret == -ERESTARTSYS) 1084 has_erestartsys = true; 1085 } 1086 1087 if (!type_found) { 1088 pr_err(TTM_PFX "No compatible memory type found\n"); 1089 return -EINVAL; 1090 } 1091 1092 return (has_erestartsys) ? -ERESTARTSYS : -ENOMEM; 1093 } 1094 EXPORT_SYMBOL(ttm_bo_mem_space); 1095 1096 static int ttm_bo_move_buffer(struct ttm_buffer_object *bo, 1097 struct ttm_placement *placement, 1098 struct ttm_operation_ctx *ctx) 1099 { 1100 int ret = 0; 1101 struct ttm_mem_reg mem; 1102 1103 reservation_object_assert_held(bo->resv); 1104 1105 mem.num_pages = bo->num_pages; 1106 mem.size = mem.num_pages << PAGE_SHIFT; 1107 mem.page_alignment = bo->mem.page_alignment; 1108 mem.bus.io_reserved_vm = false; 1109 mem.bus.io_reserved_count = 0; 1110 /* 1111 * Determine where to move the buffer. 1112 */ 1113 ret = ttm_bo_mem_space(bo, placement, &mem, ctx); 1114 if (ret) 1115 goto out_unlock; 1116 ret = ttm_bo_handle_move_mem(bo, &mem, false, ctx); 1117 out_unlock: 1118 if (ret && mem.mm_node) 1119 ttm_bo_mem_put(bo, &mem); 1120 return ret; 1121 } 1122 1123 static bool ttm_bo_places_compat(const struct ttm_place *places, 1124 unsigned num_placement, 1125 struct ttm_mem_reg *mem, 1126 uint32_t *new_flags) 1127 { 1128 unsigned i; 1129 1130 for (i = 0; i < num_placement; i++) { 1131 const struct ttm_place *heap = &places[i]; 1132 1133 if (mem->mm_node && (mem->start < heap->fpfn || 1134 (heap->lpfn != 0 && (mem->start + mem->num_pages) > heap->lpfn))) 1135 continue; 1136 1137 *new_flags = heap->flags; 1138 if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) && 1139 (*new_flags & mem->placement & TTM_PL_MASK_MEM) && 1140 (!(*new_flags & TTM_PL_FLAG_CONTIGUOUS) || 1141 (mem->placement & TTM_PL_FLAG_CONTIGUOUS))) 1142 return true; 1143 } 1144 return false; 1145 } 1146 1147 bool ttm_bo_mem_compat(struct ttm_placement *placement, 1148 struct ttm_mem_reg *mem, 1149 uint32_t *new_flags) 1150 { 1151 if (ttm_bo_places_compat(placement->placement, placement->num_placement, 1152 mem, new_flags)) 1153 return true; 1154 1155 if ((placement->busy_placement != placement->placement || 1156 placement->num_busy_placement > placement->num_placement) && 1157 ttm_bo_places_compat(placement->busy_placement, 1158 placement->num_busy_placement, 1159 mem, new_flags)) 1160 return true; 1161 1162 return false; 1163 } 1164 EXPORT_SYMBOL(ttm_bo_mem_compat); 1165 1166 int ttm_bo_validate(struct ttm_buffer_object *bo, 1167 struct ttm_placement *placement, 1168 struct ttm_operation_ctx *ctx) 1169 { 1170 int ret; 1171 uint32_t new_flags; 1172 1173 reservation_object_assert_held(bo->resv); 1174 /* 1175 * Check whether we need to move buffer. 1176 */ 1177 if (!ttm_bo_mem_compat(placement, &bo->mem, &new_flags)) { 1178 ret = ttm_bo_move_buffer(bo, placement, ctx); 1179 if (ret) 1180 return ret; 1181 } else { 1182 /* 1183 * Use the access and other non-mapping-related flag bits from 1184 * the compatible memory placement flags to the active flags 1185 */ 1186 ttm_flag_masked(&bo->mem.placement, new_flags, 1187 ~TTM_PL_MASK_MEMTYPE); 1188 } 1189 /* 1190 * We might need to add a TTM. 1191 */ 1192 if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) { 1193 ret = ttm_tt_create(bo, true); 1194 if (ret) 1195 return ret; 1196 } 1197 return 0; 1198 } 1199 EXPORT_SYMBOL(ttm_bo_validate); 1200 1201 int ttm_bo_init_reserved(struct ttm_bo_device *bdev, 1202 struct ttm_buffer_object *bo, 1203 unsigned long size, 1204 enum ttm_bo_type type, 1205 struct ttm_placement *placement, 1206 uint32_t page_alignment, 1207 struct ttm_operation_ctx *ctx, 1208 size_t acc_size, 1209 struct sg_table *sg, 1210 struct reservation_object *resv, 1211 void (*destroy) (struct ttm_buffer_object *)) 1212 { 1213 int ret = 0; 1214 unsigned long num_pages; 1215 struct ttm_mem_global *mem_glob = bdev->glob->mem_glob; 1216 bool locked; 1217 1218 ret = ttm_mem_global_alloc(mem_glob, acc_size, ctx); 1219 if (ret) { 1220 pr_err("Out of kernel memory\n"); 1221 if (destroy) 1222 (*destroy)(bo); 1223 else 1224 kfree(bo); 1225 return -ENOMEM; 1226 } 1227 1228 num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT; 1229 if (num_pages == 0) { 1230 pr_err("Illegal buffer object size\n"); 1231 if (destroy) 1232 (*destroy)(bo); 1233 else 1234 kfree(bo); 1235 ttm_mem_global_free(mem_glob, acc_size); 1236 return -EINVAL; 1237 } 1238 bo->destroy = destroy ? destroy : ttm_bo_default_destroy; 1239 1240 kref_init(&bo->kref); 1241 kref_init(&bo->list_kref); 1242 atomic_set(&bo->cpu_writers, 0); 1243 INIT_LIST_HEAD(&bo->lru); 1244 INIT_LIST_HEAD(&bo->ddestroy); 1245 INIT_LIST_HEAD(&bo->swap); 1246 INIT_LIST_HEAD(&bo->io_reserve_lru); 1247 mutex_init(&bo->wu_mutex); 1248 bo->bdev = bdev; 1249 bo->type = type; 1250 bo->num_pages = num_pages; 1251 bo->mem.size = num_pages << PAGE_SHIFT; 1252 bo->mem.mem_type = TTM_PL_SYSTEM; 1253 bo->mem.num_pages = bo->num_pages; 1254 bo->mem.mm_node = NULL; 1255 bo->mem.page_alignment = page_alignment; 1256 bo->mem.bus.io_reserved_vm = false; 1257 bo->mem.bus.io_reserved_count = 0; 1258 bo->moving = NULL; 1259 bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED); 1260 bo->acc_size = acc_size; 1261 bo->sg = sg; 1262 if (resv) { 1263 bo->resv = resv; 1264 reservation_object_assert_held(bo->resv); 1265 } else { 1266 bo->resv = &bo->ttm_resv; 1267 } 1268 reservation_object_init(&bo->ttm_resv); 1269 atomic_inc(&bo->bdev->glob->bo_count); 1270 drm_vma_node_reset(&bo->vma_node); 1271 1272 /* 1273 * For ttm_bo_type_device buffers, allocate 1274 * address space from the device. 1275 */ 1276 if (bo->type == ttm_bo_type_device || 1277 bo->type == ttm_bo_type_sg) 1278 ret = drm_vma_offset_add(&bdev->vma_manager, &bo->vma_node, 1279 bo->mem.num_pages); 1280 1281 /* passed reservation objects should already be locked, 1282 * since otherwise lockdep will be angered in radeon. 1283 */ 1284 if (!resv) { 1285 locked = reservation_object_trylock(bo->resv); 1286 WARN_ON(!locked); 1287 } 1288 1289 if (likely(!ret)) 1290 ret = ttm_bo_validate(bo, placement, ctx); 1291 1292 if (unlikely(ret)) { 1293 if (!resv) 1294 ttm_bo_unreserve(bo); 1295 1296 ttm_bo_put(bo); 1297 return ret; 1298 } 1299 1300 if (resv && !(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) { 1301 spin_lock(&bdev->glob->lru_lock); 1302 ttm_bo_add_to_lru(bo); 1303 spin_unlock(&bdev->glob->lru_lock); 1304 } 1305 1306 return ret; 1307 } 1308 EXPORT_SYMBOL(ttm_bo_init_reserved); 1309 1310 int ttm_bo_init(struct ttm_bo_device *bdev, 1311 struct ttm_buffer_object *bo, 1312 unsigned long size, 1313 enum ttm_bo_type type, 1314 struct ttm_placement *placement, 1315 uint32_t page_alignment, 1316 bool interruptible, 1317 size_t acc_size, 1318 struct sg_table *sg, 1319 struct reservation_object *resv, 1320 void (*destroy) (struct ttm_buffer_object *)) 1321 { 1322 struct ttm_operation_ctx ctx = { interruptible, false }; 1323 int ret; 1324 1325 ret = ttm_bo_init_reserved(bdev, bo, size, type, placement, 1326 page_alignment, &ctx, acc_size, 1327 sg, resv, destroy); 1328 if (ret) 1329 return ret; 1330 1331 if (!resv) 1332 ttm_bo_unreserve(bo); 1333 1334 return 0; 1335 } 1336 EXPORT_SYMBOL(ttm_bo_init); 1337 1338 size_t ttm_bo_acc_size(struct ttm_bo_device *bdev, 1339 unsigned long bo_size, 1340 unsigned struct_size) 1341 { 1342 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT; 1343 size_t size = 0; 1344 1345 size += ttm_round_pot(struct_size); 1346 size += ttm_round_pot(npages * sizeof(void *)); 1347 size += ttm_round_pot(sizeof(struct ttm_tt)); 1348 return size; 1349 } 1350 EXPORT_SYMBOL(ttm_bo_acc_size); 1351 1352 size_t ttm_bo_dma_acc_size(struct ttm_bo_device *bdev, 1353 unsigned long bo_size, 1354 unsigned struct_size) 1355 { 1356 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT; 1357 size_t size = 0; 1358 1359 size += ttm_round_pot(struct_size); 1360 size += ttm_round_pot(npages * (2*sizeof(void *) + sizeof(dma_addr_t))); 1361 size += ttm_round_pot(sizeof(struct ttm_dma_tt)); 1362 return size; 1363 } 1364 EXPORT_SYMBOL(ttm_bo_dma_acc_size); 1365 1366 int ttm_bo_create(struct ttm_bo_device *bdev, 1367 unsigned long size, 1368 enum ttm_bo_type type, 1369 struct ttm_placement *placement, 1370 uint32_t page_alignment, 1371 bool interruptible, 1372 struct ttm_buffer_object **p_bo) 1373 { 1374 struct ttm_buffer_object *bo; 1375 size_t acc_size; 1376 int ret; 1377 1378 bo = kzalloc(sizeof(*bo), GFP_KERNEL); 1379 if (unlikely(bo == NULL)) 1380 return -ENOMEM; 1381 1382 acc_size = ttm_bo_acc_size(bdev, size, sizeof(struct ttm_buffer_object)); 1383 ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment, 1384 interruptible, acc_size, 1385 NULL, NULL, NULL); 1386 if (likely(ret == 0)) 1387 *p_bo = bo; 1388 1389 return ret; 1390 } 1391 EXPORT_SYMBOL(ttm_bo_create); 1392 1393 static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev, 1394 unsigned mem_type) 1395 { 1396 struct ttm_operation_ctx ctx = { 1397 .interruptible = false, 1398 .no_wait_gpu = false, 1399 .flags = TTM_OPT_FLAG_FORCE_ALLOC 1400 }; 1401 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 1402 struct ttm_bo_global *glob = bdev->glob; 1403 struct dma_fence *fence; 1404 int ret; 1405 unsigned i; 1406 1407 /* 1408 * Can't use standard list traversal since we're unlocking. 1409 */ 1410 1411 spin_lock(&glob->lru_lock); 1412 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) { 1413 while (!list_empty(&man->lru[i])) { 1414 spin_unlock(&glob->lru_lock); 1415 ret = ttm_mem_evict_first(bdev, mem_type, NULL, &ctx); 1416 if (ret) 1417 return ret; 1418 spin_lock(&glob->lru_lock); 1419 } 1420 } 1421 spin_unlock(&glob->lru_lock); 1422 1423 spin_lock(&man->move_lock); 1424 fence = dma_fence_get(man->move); 1425 spin_unlock(&man->move_lock); 1426 1427 if (fence) { 1428 ret = dma_fence_wait(fence, false); 1429 dma_fence_put(fence); 1430 if (ret) 1431 return ret; 1432 } 1433 1434 return 0; 1435 } 1436 1437 int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type) 1438 { 1439 struct ttm_mem_type_manager *man; 1440 int ret = -EINVAL; 1441 1442 if (mem_type >= TTM_NUM_MEM_TYPES) { 1443 pr_err("Illegal memory type %d\n", mem_type); 1444 return ret; 1445 } 1446 man = &bdev->man[mem_type]; 1447 1448 if (!man->has_type) { 1449 pr_err("Trying to take down uninitialized memory manager type %u\n", 1450 mem_type); 1451 return ret; 1452 } 1453 1454 man->use_type = false; 1455 man->has_type = false; 1456 1457 ret = 0; 1458 if (mem_type > 0) { 1459 ret = ttm_bo_force_list_clean(bdev, mem_type); 1460 if (ret) { 1461 pr_err("Cleanup eviction failed\n"); 1462 return ret; 1463 } 1464 1465 ret = (*man->func->takedown)(man); 1466 } 1467 1468 dma_fence_put(man->move); 1469 man->move = NULL; 1470 1471 return ret; 1472 } 1473 EXPORT_SYMBOL(ttm_bo_clean_mm); 1474 1475 int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type) 1476 { 1477 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 1478 1479 if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) { 1480 pr_err("Illegal memory manager memory type %u\n", mem_type); 1481 return -EINVAL; 1482 } 1483 1484 if (!man->has_type) { 1485 pr_err("Memory type %u has not been initialized\n", mem_type); 1486 return 0; 1487 } 1488 1489 return ttm_bo_force_list_clean(bdev, mem_type); 1490 } 1491 EXPORT_SYMBOL(ttm_bo_evict_mm); 1492 1493 int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type, 1494 unsigned long p_size) 1495 { 1496 int ret; 1497 struct ttm_mem_type_manager *man; 1498 unsigned i; 1499 1500 BUG_ON(type >= TTM_NUM_MEM_TYPES); 1501 man = &bdev->man[type]; 1502 BUG_ON(man->has_type); 1503 man->io_reserve_fastpath = true; 1504 man->use_io_reserve_lru = false; 1505 mutex_init(&man->io_reserve_mutex); 1506 spin_lock_init(&man->move_lock); 1507 INIT_LIST_HEAD(&man->io_reserve_lru); 1508 1509 ret = bdev->driver->init_mem_type(bdev, type, man); 1510 if (ret) 1511 return ret; 1512 man->bdev = bdev; 1513 1514 if (type != TTM_PL_SYSTEM) { 1515 ret = (*man->func->init)(man, p_size); 1516 if (ret) 1517 return ret; 1518 } 1519 man->has_type = true; 1520 man->use_type = true; 1521 man->size = p_size; 1522 1523 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) 1524 INIT_LIST_HEAD(&man->lru[i]); 1525 man->move = NULL; 1526 1527 return 0; 1528 } 1529 EXPORT_SYMBOL(ttm_bo_init_mm); 1530 1531 static void ttm_bo_global_kobj_release(struct kobject *kobj) 1532 { 1533 struct ttm_bo_global *glob = 1534 container_of(kobj, struct ttm_bo_global, kobj); 1535 1536 __free_page(glob->dummy_read_page); 1537 kfree(glob); 1538 } 1539 1540 void ttm_bo_global_release(struct drm_global_reference *ref) 1541 { 1542 struct ttm_bo_global *glob = ref->object; 1543 1544 kobject_del(&glob->kobj); 1545 kobject_put(&glob->kobj); 1546 } 1547 EXPORT_SYMBOL(ttm_bo_global_release); 1548 1549 int ttm_bo_global_init(struct drm_global_reference *ref) 1550 { 1551 struct ttm_bo_global_ref *bo_ref = 1552 container_of(ref, struct ttm_bo_global_ref, ref); 1553 struct ttm_bo_global *glob = ref->object; 1554 int ret; 1555 unsigned i; 1556 1557 mutex_init(&glob->device_list_mutex); 1558 spin_lock_init(&glob->lru_lock); 1559 glob->mem_glob = bo_ref->mem_glob; 1560 glob->mem_glob->bo_glob = glob; 1561 glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32); 1562 1563 if (unlikely(glob->dummy_read_page == NULL)) { 1564 ret = -ENOMEM; 1565 goto out_no_drp; 1566 } 1567 1568 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) 1569 INIT_LIST_HEAD(&glob->swap_lru[i]); 1570 INIT_LIST_HEAD(&glob->device_list); 1571 atomic_set(&glob->bo_count, 0); 1572 1573 ret = kobject_init_and_add( 1574 &glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects"); 1575 if (unlikely(ret != 0)) 1576 kobject_put(&glob->kobj); 1577 return ret; 1578 out_no_drp: 1579 kfree(glob); 1580 return ret; 1581 } 1582 EXPORT_SYMBOL(ttm_bo_global_init); 1583 1584 1585 int ttm_bo_device_release(struct ttm_bo_device *bdev) 1586 { 1587 int ret = 0; 1588 unsigned i = TTM_NUM_MEM_TYPES; 1589 struct ttm_mem_type_manager *man; 1590 struct ttm_bo_global *glob = bdev->glob; 1591 1592 while (i--) { 1593 man = &bdev->man[i]; 1594 if (man->has_type) { 1595 man->use_type = false; 1596 if ((i != TTM_PL_SYSTEM) && ttm_bo_clean_mm(bdev, i)) { 1597 ret = -EBUSY; 1598 pr_err("DRM memory manager type %d is not clean\n", 1599 i); 1600 } 1601 man->has_type = false; 1602 } 1603 } 1604 1605 mutex_lock(&glob->device_list_mutex); 1606 list_del(&bdev->device_list); 1607 mutex_unlock(&glob->device_list_mutex); 1608 1609 cancel_delayed_work_sync(&bdev->wq); 1610 1611 if (ttm_bo_delayed_delete(bdev, true)) 1612 pr_debug("Delayed destroy list was clean\n"); 1613 1614 spin_lock(&glob->lru_lock); 1615 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) 1616 if (list_empty(&bdev->man[0].lru[0])) 1617 pr_debug("Swap list %d was clean\n", i); 1618 spin_unlock(&glob->lru_lock); 1619 1620 drm_vma_offset_manager_destroy(&bdev->vma_manager); 1621 1622 return ret; 1623 } 1624 EXPORT_SYMBOL(ttm_bo_device_release); 1625 1626 int ttm_bo_device_init(struct ttm_bo_device *bdev, 1627 struct ttm_bo_global *glob, 1628 struct ttm_bo_driver *driver, 1629 struct address_space *mapping, 1630 uint64_t file_page_offset, 1631 bool need_dma32) 1632 { 1633 int ret = -EINVAL; 1634 1635 bdev->driver = driver; 1636 1637 memset(bdev->man, 0, sizeof(bdev->man)); 1638 1639 /* 1640 * Initialize the system memory buffer type. 1641 * Other types need to be driver / IOCTL initialized. 1642 */ 1643 ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0); 1644 if (unlikely(ret != 0)) 1645 goto out_no_sys; 1646 1647 drm_vma_offset_manager_init(&bdev->vma_manager, file_page_offset, 1648 0x10000000); 1649 INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue); 1650 INIT_LIST_HEAD(&bdev->ddestroy); 1651 bdev->dev_mapping = mapping; 1652 bdev->glob = glob; 1653 bdev->need_dma32 = need_dma32; 1654 mutex_lock(&glob->device_list_mutex); 1655 list_add_tail(&bdev->device_list, &glob->device_list); 1656 mutex_unlock(&glob->device_list_mutex); 1657 1658 return 0; 1659 out_no_sys: 1660 return ret; 1661 } 1662 EXPORT_SYMBOL(ttm_bo_device_init); 1663 1664 /* 1665 * buffer object vm functions. 1666 */ 1667 1668 bool ttm_mem_reg_is_pci(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem) 1669 { 1670 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type]; 1671 1672 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) { 1673 if (mem->mem_type == TTM_PL_SYSTEM) 1674 return false; 1675 1676 if (man->flags & TTM_MEMTYPE_FLAG_CMA) 1677 return false; 1678 1679 if (mem->placement & TTM_PL_FLAG_CACHED) 1680 return false; 1681 } 1682 return true; 1683 } 1684 1685 void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo) 1686 { 1687 struct ttm_bo_device *bdev = bo->bdev; 1688 1689 drm_vma_node_unmap(&bo->vma_node, bdev->dev_mapping); 1690 ttm_mem_io_free_vm(bo); 1691 } 1692 1693 void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo) 1694 { 1695 struct ttm_bo_device *bdev = bo->bdev; 1696 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type]; 1697 1698 ttm_mem_io_lock(man, false); 1699 ttm_bo_unmap_virtual_locked(bo); 1700 ttm_mem_io_unlock(man); 1701 } 1702 1703 1704 EXPORT_SYMBOL(ttm_bo_unmap_virtual); 1705 1706 int ttm_bo_wait(struct ttm_buffer_object *bo, 1707 bool interruptible, bool no_wait) 1708 { 1709 long timeout = 15 * HZ; 1710 1711 if (no_wait) { 1712 if (reservation_object_test_signaled_rcu(bo->resv, true)) 1713 return 0; 1714 else 1715 return -EBUSY; 1716 } 1717 1718 timeout = reservation_object_wait_timeout_rcu(bo->resv, true, 1719 interruptible, timeout); 1720 if (timeout < 0) 1721 return timeout; 1722 1723 if (timeout == 0) 1724 return -EBUSY; 1725 1726 reservation_object_add_excl_fence(bo->resv, NULL); 1727 return 0; 1728 } 1729 EXPORT_SYMBOL(ttm_bo_wait); 1730 1731 int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait) 1732 { 1733 int ret = 0; 1734 1735 /* 1736 * Using ttm_bo_reserve makes sure the lru lists are updated. 1737 */ 1738 1739 ret = ttm_bo_reserve(bo, true, no_wait, NULL); 1740 if (unlikely(ret != 0)) 1741 return ret; 1742 ret = ttm_bo_wait(bo, true, no_wait); 1743 if (likely(ret == 0)) 1744 atomic_inc(&bo->cpu_writers); 1745 ttm_bo_unreserve(bo); 1746 return ret; 1747 } 1748 EXPORT_SYMBOL(ttm_bo_synccpu_write_grab); 1749 1750 void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo) 1751 { 1752 atomic_dec(&bo->cpu_writers); 1753 } 1754 EXPORT_SYMBOL(ttm_bo_synccpu_write_release); 1755 1756 /** 1757 * A buffer object shrink method that tries to swap out the first 1758 * buffer object on the bo_global::swap_lru list. 1759 */ 1760 int ttm_bo_swapout(struct ttm_bo_global *glob, struct ttm_operation_ctx *ctx) 1761 { 1762 struct ttm_buffer_object *bo; 1763 int ret = -EBUSY; 1764 bool locked; 1765 unsigned i; 1766 1767 spin_lock(&glob->lru_lock); 1768 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) { 1769 list_for_each_entry(bo, &glob->swap_lru[i], swap) { 1770 if (ttm_bo_evict_swapout_allowable(bo, ctx, &locked)) { 1771 ret = 0; 1772 break; 1773 } 1774 } 1775 if (!ret) 1776 break; 1777 } 1778 1779 if (ret) { 1780 spin_unlock(&glob->lru_lock); 1781 return ret; 1782 } 1783 1784 kref_get(&bo->list_kref); 1785 1786 if (!list_empty(&bo->ddestroy)) { 1787 ret = ttm_bo_cleanup_refs(bo, false, false, locked); 1788 kref_put(&bo->list_kref, ttm_bo_release_list); 1789 return ret; 1790 } 1791 1792 ttm_bo_del_from_lru(bo); 1793 spin_unlock(&glob->lru_lock); 1794 1795 /** 1796 * Move to system cached 1797 */ 1798 1799 if (bo->mem.mem_type != TTM_PL_SYSTEM || 1800 bo->ttm->caching_state != tt_cached) { 1801 struct ttm_operation_ctx ctx = { false, false }; 1802 struct ttm_mem_reg evict_mem; 1803 1804 evict_mem = bo->mem; 1805 evict_mem.mm_node = NULL; 1806 evict_mem.placement = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED; 1807 evict_mem.mem_type = TTM_PL_SYSTEM; 1808 1809 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, &ctx); 1810 if (unlikely(ret != 0)) 1811 goto out; 1812 } 1813 1814 /** 1815 * Make sure BO is idle. 1816 */ 1817 1818 ret = ttm_bo_wait(bo, false, false); 1819 if (unlikely(ret != 0)) 1820 goto out; 1821 1822 ttm_bo_unmap_virtual(bo); 1823 1824 /** 1825 * Swap out. Buffer will be swapped in again as soon as 1826 * anyone tries to access a ttm page. 1827 */ 1828 1829 if (bo->bdev->driver->swap_notify) 1830 bo->bdev->driver->swap_notify(bo); 1831 1832 ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage); 1833 out: 1834 1835 /** 1836 * 1837 * Unreserve without putting on LRU to avoid swapping out an 1838 * already swapped buffer. 1839 */ 1840 if (locked) 1841 reservation_object_unlock(bo->resv); 1842 kref_put(&bo->list_kref, ttm_bo_release_list); 1843 return ret; 1844 } 1845 EXPORT_SYMBOL(ttm_bo_swapout); 1846 1847 void ttm_bo_swapout_all(struct ttm_bo_device *bdev) 1848 { 1849 struct ttm_operation_ctx ctx = { 1850 .interruptible = false, 1851 .no_wait_gpu = false 1852 }; 1853 1854 while (ttm_bo_swapout(bdev->glob, &ctx) == 0) 1855 ; 1856 } 1857 EXPORT_SYMBOL(ttm_bo_swapout_all); 1858 1859 /** 1860 * ttm_bo_wait_unreserved - interruptible wait for a buffer object to become 1861 * unreserved 1862 * 1863 * @bo: Pointer to buffer 1864 */ 1865 int ttm_bo_wait_unreserved(struct ttm_buffer_object *bo) 1866 { 1867 int ret; 1868 1869 /* 1870 * In the absense of a wait_unlocked API, 1871 * Use the bo::wu_mutex to avoid triggering livelocks due to 1872 * concurrent use of this function. Note that this use of 1873 * bo::wu_mutex can go away if we change locking order to 1874 * mmap_sem -> bo::reserve. 1875 */ 1876 ret = mutex_lock_interruptible(&bo->wu_mutex); 1877 if (unlikely(ret != 0)) 1878 return -ERESTARTSYS; 1879 if (!ww_mutex_is_locked(&bo->resv->lock)) 1880 goto out_unlock; 1881 ret = reservation_object_lock_interruptible(bo->resv, NULL); 1882 if (ret == -EINTR) 1883 ret = -ERESTARTSYS; 1884 if (unlikely(ret != 0)) 1885 goto out_unlock; 1886 reservation_object_unlock(bo->resv); 1887 1888 out_unlock: 1889 mutex_unlock(&bo->wu_mutex); 1890 return ret; 1891 } 1892