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