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