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, no_wait_gpu, mem); 358 else if (bdev->driver->move) 359 ret = bdev->driver->move(bo, evict, interruptible, 360 no_wait_gpu, mem); 361 else 362 ret = ttm_bo_move_memcpy(bo, evict, interruptible, 363 no_wait_gpu, mem); 364 365 if (ret) { 366 if (bdev->driver->move_notify) { 367 struct ttm_mem_reg tmp_mem = *mem; 368 *mem = bo->mem; 369 bo->mem = tmp_mem; 370 bdev->driver->move_notify(bo, mem); 371 bo->mem = *mem; 372 *mem = tmp_mem; 373 } 374 375 goto out_err; 376 } 377 378 moved: 379 if (bo->evicted) { 380 if (bdev->driver->invalidate_caches) { 381 ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement); 382 if (ret) 383 pr_err("Can not flush read caches\n"); 384 } 385 bo->evicted = false; 386 } 387 388 if (bo->mem.mm_node) { 389 bo->offset = (bo->mem.start << PAGE_SHIFT) + 390 bdev->man[bo->mem.mem_type].gpu_offset; 391 bo->cur_placement = bo->mem.placement; 392 } else 393 bo->offset = 0; 394 395 return 0; 396 397 out_err: 398 new_man = &bdev->man[bo->mem.mem_type]; 399 if (new_man->flags & TTM_MEMTYPE_FLAG_FIXED) { 400 ttm_tt_destroy(bo->ttm); 401 bo->ttm = NULL; 402 } 403 404 return ret; 405 } 406 407 /** 408 * Call bo::reserved. 409 * Will release GPU memory type usage on destruction. 410 * This is the place to put in driver specific hooks to release 411 * driver private resources. 412 * Will release the bo::reserved lock. 413 */ 414 415 static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo) 416 { 417 if (bo->bdev->driver->move_notify) 418 bo->bdev->driver->move_notify(bo, NULL); 419 420 ttm_tt_destroy(bo->ttm); 421 bo->ttm = NULL; 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, NULL); 456 457 if (!ret) { 458 if (!ttm_bo_wait(bo, 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, 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, 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, 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, NULL); 599 if (remove_all && ret) { 600 spin_unlock(&glob->lru_lock); 601 ret = __ttm_bo_reserve(entry, false, false, NULL); 602 spin_lock(&glob->lru_lock); 603 } 604 605 if (!ret) 606 ret = ttm_bo_cleanup_refs_and_unlock(entry, false, 607 !remove_all); 608 else 609 spin_unlock(&glob->lru_lock); 610 611 kref_put(&entry->list_kref, ttm_bo_release_list); 612 entry = nentry; 613 614 if (ret || !entry) 615 goto out; 616 617 spin_lock(&glob->lru_lock); 618 if (list_empty(&entry->ddestroy)) 619 break; 620 } 621 622 out_unlock: 623 spin_unlock(&glob->lru_lock); 624 out: 625 if (entry) 626 kref_put(&entry->list_kref, ttm_bo_release_list); 627 return ret; 628 } 629 630 static void ttm_bo_delayed_workqueue(struct work_struct *work) 631 { 632 struct ttm_bo_device *bdev = 633 container_of(work, struct ttm_bo_device, wq.work); 634 635 if (ttm_bo_delayed_delete(bdev, false)) { 636 schedule_delayed_work(&bdev->wq, 637 ((HZ / 100) < 1) ? 1 : HZ / 100); 638 } 639 } 640 641 static void ttm_bo_release(struct kref *kref) 642 { 643 struct ttm_buffer_object *bo = 644 container_of(kref, struct ttm_buffer_object, kref); 645 struct ttm_bo_device *bdev = bo->bdev; 646 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type]; 647 648 drm_vma_offset_remove(&bdev->vma_manager, &bo->vma_node); 649 ttm_mem_io_lock(man, false); 650 ttm_mem_io_free_vm(bo); 651 ttm_mem_io_unlock(man); 652 ttm_bo_cleanup_refs_or_queue(bo); 653 kref_put(&bo->list_kref, ttm_bo_release_list); 654 } 655 656 void ttm_bo_unref(struct ttm_buffer_object **p_bo) 657 { 658 struct ttm_buffer_object *bo = *p_bo; 659 660 *p_bo = NULL; 661 kref_put(&bo->kref, ttm_bo_release); 662 } 663 EXPORT_SYMBOL(ttm_bo_unref); 664 665 int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev) 666 { 667 return cancel_delayed_work_sync(&bdev->wq); 668 } 669 EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue); 670 671 void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched) 672 { 673 if (resched) 674 schedule_delayed_work(&bdev->wq, 675 ((HZ / 100) < 1) ? 1 : HZ / 100); 676 } 677 EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue); 678 679 static int ttm_bo_evict(struct ttm_buffer_object *bo, bool interruptible, 680 bool no_wait_gpu) 681 { 682 struct ttm_bo_device *bdev = bo->bdev; 683 struct ttm_mem_reg evict_mem; 684 struct ttm_placement placement; 685 int ret = 0; 686 687 lockdep_assert_held(&bo->resv->lock.base); 688 689 evict_mem = bo->mem; 690 evict_mem.mm_node = NULL; 691 evict_mem.bus.io_reserved_vm = false; 692 evict_mem.bus.io_reserved_count = 0; 693 694 placement.num_placement = 0; 695 placement.num_busy_placement = 0; 696 bdev->driver->evict_flags(bo, &placement); 697 ret = ttm_bo_mem_space(bo, &placement, &evict_mem, interruptible, 698 no_wait_gpu); 699 if (ret) { 700 if (ret != -ERESTARTSYS) { 701 pr_err("Failed to find memory space for buffer 0x%p eviction\n", 702 bo); 703 ttm_bo_mem_space_debug(bo, &placement); 704 } 705 goto out; 706 } 707 708 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, interruptible, 709 no_wait_gpu); 710 if (unlikely(ret)) { 711 if (ret != -ERESTARTSYS) 712 pr_err("Buffer eviction failed\n"); 713 ttm_bo_mem_put(bo, &evict_mem); 714 goto out; 715 } 716 bo->evicted = true; 717 out: 718 return ret; 719 } 720 721 static int ttm_mem_evict_first(struct ttm_bo_device *bdev, 722 uint32_t mem_type, 723 const struct ttm_place *place, 724 bool interruptible, 725 bool no_wait_gpu) 726 { 727 struct ttm_bo_global *glob = bdev->glob; 728 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 729 struct ttm_buffer_object *bo; 730 int ret = -EBUSY, put_count; 731 732 spin_lock(&glob->lru_lock); 733 list_for_each_entry(bo, &man->lru, lru) { 734 ret = __ttm_bo_reserve(bo, false, true, NULL); 735 if (!ret) { 736 if (place && (place->fpfn || place->lpfn)) { 737 /* Don't evict this BO if it's outside of the 738 * requested placement range 739 */ 740 if (place->fpfn >= (bo->mem.start + bo->mem.size) || 741 (place->lpfn && place->lpfn <= bo->mem.start)) { 742 __ttm_bo_unreserve(bo); 743 ret = -EBUSY; 744 continue; 745 } 746 } 747 748 break; 749 } 750 } 751 752 if (ret) { 753 spin_unlock(&glob->lru_lock); 754 return ret; 755 } 756 757 kref_get(&bo->list_kref); 758 759 if (!list_empty(&bo->ddestroy)) { 760 ret = ttm_bo_cleanup_refs_and_unlock(bo, interruptible, 761 no_wait_gpu); 762 kref_put(&bo->list_kref, ttm_bo_release_list); 763 return ret; 764 } 765 766 put_count = ttm_bo_del_from_lru(bo); 767 spin_unlock(&glob->lru_lock); 768 769 BUG_ON(ret != 0); 770 771 ttm_bo_list_ref_sub(bo, put_count, true); 772 773 ret = ttm_bo_evict(bo, interruptible, no_wait_gpu); 774 ttm_bo_unreserve(bo); 775 776 kref_put(&bo->list_kref, ttm_bo_release_list); 777 return ret; 778 } 779 780 void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem) 781 { 782 struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type]; 783 784 if (mem->mm_node) 785 (*man->func->put_node)(man, mem); 786 } 787 EXPORT_SYMBOL(ttm_bo_mem_put); 788 789 /** 790 * Add the last move fence to the BO and reserve a new shared slot. 791 */ 792 static int ttm_bo_add_move_fence(struct ttm_buffer_object *bo, 793 struct ttm_mem_type_manager *man, 794 struct ttm_mem_reg *mem) 795 { 796 struct fence *fence; 797 int ret; 798 799 spin_lock(&man->move_lock); 800 fence = fence_get(man->move); 801 spin_unlock(&man->move_lock); 802 803 if (fence) { 804 reservation_object_add_shared_fence(bo->resv, fence); 805 806 ret = reservation_object_reserve_shared(bo->resv); 807 if (unlikely(ret)) 808 return ret; 809 810 fence_put(bo->moving); 811 bo->moving = fence; 812 } 813 814 return 0; 815 } 816 817 /** 818 * Repeatedly evict memory from the LRU for @mem_type until we create enough 819 * space, or we've evicted everything and there isn't enough space. 820 */ 821 static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo, 822 uint32_t mem_type, 823 const struct ttm_place *place, 824 struct ttm_mem_reg *mem, 825 bool interruptible, 826 bool no_wait_gpu) 827 { 828 struct ttm_bo_device *bdev = bo->bdev; 829 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 830 int ret; 831 832 do { 833 ret = (*man->func->get_node)(man, bo, place, mem); 834 if (unlikely(ret != 0)) 835 return ret; 836 if (mem->mm_node) 837 break; 838 ret = ttm_mem_evict_first(bdev, mem_type, place, 839 interruptible, no_wait_gpu); 840 if (unlikely(ret != 0)) 841 return ret; 842 } while (1); 843 mem->mem_type = mem_type; 844 return ttm_bo_add_move_fence(bo, man, mem); 845 } 846 847 static uint32_t ttm_bo_select_caching(struct ttm_mem_type_manager *man, 848 uint32_t cur_placement, 849 uint32_t proposed_placement) 850 { 851 uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING; 852 uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING; 853 854 /** 855 * Keep current caching if possible. 856 */ 857 858 if ((cur_placement & caching) != 0) 859 result |= (cur_placement & caching); 860 else if ((man->default_caching & caching) != 0) 861 result |= man->default_caching; 862 else if ((TTM_PL_FLAG_CACHED & caching) != 0) 863 result |= TTM_PL_FLAG_CACHED; 864 else if ((TTM_PL_FLAG_WC & caching) != 0) 865 result |= TTM_PL_FLAG_WC; 866 else if ((TTM_PL_FLAG_UNCACHED & caching) != 0) 867 result |= TTM_PL_FLAG_UNCACHED; 868 869 return result; 870 } 871 872 static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man, 873 uint32_t mem_type, 874 const struct ttm_place *place, 875 uint32_t *masked_placement) 876 { 877 uint32_t cur_flags = ttm_bo_type_flags(mem_type); 878 879 if ((cur_flags & place->flags & TTM_PL_MASK_MEM) == 0) 880 return false; 881 882 if ((place->flags & man->available_caching) == 0) 883 return false; 884 885 cur_flags |= (place->flags & man->available_caching); 886 887 *masked_placement = cur_flags; 888 return true; 889 } 890 891 /** 892 * Creates space for memory region @mem according to its type. 893 * 894 * This function first searches for free space in compatible memory types in 895 * the priority order defined by the driver. If free space isn't found, then 896 * ttm_bo_mem_force_space is attempted in priority order to evict and find 897 * space. 898 */ 899 int ttm_bo_mem_space(struct ttm_buffer_object *bo, 900 struct ttm_placement *placement, 901 struct ttm_mem_reg *mem, 902 bool interruptible, 903 bool no_wait_gpu) 904 { 905 struct ttm_bo_device *bdev = bo->bdev; 906 struct ttm_mem_type_manager *man; 907 uint32_t mem_type = TTM_PL_SYSTEM; 908 uint32_t cur_flags = 0; 909 bool type_found = false; 910 bool type_ok = false; 911 bool has_erestartsys = false; 912 int i, ret; 913 914 ret = reservation_object_reserve_shared(bo->resv); 915 if (unlikely(ret)) 916 return ret; 917 918 mem->mm_node = NULL; 919 for (i = 0; i < placement->num_placement; ++i) { 920 const struct ttm_place *place = &placement->placement[i]; 921 922 ret = ttm_mem_type_from_place(place, &mem_type); 923 if (ret) 924 return ret; 925 man = &bdev->man[mem_type]; 926 if (!man->has_type || !man->use_type) 927 continue; 928 929 type_ok = ttm_bo_mt_compatible(man, mem_type, place, 930 &cur_flags); 931 932 if (!type_ok) 933 continue; 934 935 type_found = true; 936 cur_flags = ttm_bo_select_caching(man, bo->mem.placement, 937 cur_flags); 938 /* 939 * Use the access and other non-mapping-related flag bits from 940 * the memory placement flags to the current flags 941 */ 942 ttm_flag_masked(&cur_flags, place->flags, 943 ~TTM_PL_MASK_MEMTYPE); 944 945 if (mem_type == TTM_PL_SYSTEM) 946 break; 947 948 ret = (*man->func->get_node)(man, bo, place, mem); 949 if (unlikely(ret)) 950 return ret; 951 952 if (mem->mm_node) { 953 ret = ttm_bo_add_move_fence(bo, man, mem); 954 if (unlikely(ret)) { 955 (*man->func->put_node)(man, mem); 956 return ret; 957 } 958 break; 959 } 960 } 961 962 if ((type_ok && (mem_type == TTM_PL_SYSTEM)) || mem->mm_node) { 963 mem->mem_type = mem_type; 964 mem->placement = cur_flags; 965 return 0; 966 } 967 968 for (i = 0; i < placement->num_busy_placement; ++i) { 969 const struct ttm_place *place = &placement->busy_placement[i]; 970 971 ret = ttm_mem_type_from_place(place, &mem_type); 972 if (ret) 973 return ret; 974 man = &bdev->man[mem_type]; 975 if (!man->has_type || !man->use_type) 976 continue; 977 if (!ttm_bo_mt_compatible(man, mem_type, place, &cur_flags)) 978 continue; 979 980 type_found = true; 981 cur_flags = ttm_bo_select_caching(man, bo->mem.placement, 982 cur_flags); 983 /* 984 * Use the access and other non-mapping-related flag bits from 985 * the memory placement flags to the current flags 986 */ 987 ttm_flag_masked(&cur_flags, place->flags, 988 ~TTM_PL_MASK_MEMTYPE); 989 990 if (mem_type == TTM_PL_SYSTEM) { 991 mem->mem_type = mem_type; 992 mem->placement = cur_flags; 993 mem->mm_node = NULL; 994 return 0; 995 } 996 997 ret = ttm_bo_mem_force_space(bo, mem_type, place, mem, 998 interruptible, no_wait_gpu); 999 if (ret == 0 && mem->mm_node) { 1000 mem->placement = cur_flags; 1001 return 0; 1002 } 1003 if (ret == -ERESTARTSYS) 1004 has_erestartsys = true; 1005 } 1006 1007 if (!type_found) { 1008 printk(KERN_ERR TTM_PFX "No compatible memory type found.\n"); 1009 return -EINVAL; 1010 } 1011 1012 return (has_erestartsys) ? -ERESTARTSYS : -ENOMEM; 1013 } 1014 EXPORT_SYMBOL(ttm_bo_mem_space); 1015 1016 static int ttm_bo_move_buffer(struct ttm_buffer_object *bo, 1017 struct ttm_placement *placement, 1018 bool interruptible, 1019 bool no_wait_gpu) 1020 { 1021 int ret = 0; 1022 struct ttm_mem_reg mem; 1023 1024 lockdep_assert_held(&bo->resv->lock.base); 1025 1026 mem.num_pages = bo->num_pages; 1027 mem.size = mem.num_pages << PAGE_SHIFT; 1028 mem.page_alignment = bo->mem.page_alignment; 1029 mem.bus.io_reserved_vm = false; 1030 mem.bus.io_reserved_count = 0; 1031 /* 1032 * Determine where to move the buffer. 1033 */ 1034 ret = ttm_bo_mem_space(bo, placement, &mem, 1035 interruptible, no_wait_gpu); 1036 if (ret) 1037 goto out_unlock; 1038 ret = ttm_bo_handle_move_mem(bo, &mem, false, 1039 interruptible, no_wait_gpu); 1040 out_unlock: 1041 if (ret && mem.mm_node) 1042 ttm_bo_mem_put(bo, &mem); 1043 return ret; 1044 } 1045 1046 bool ttm_bo_mem_compat(struct ttm_placement *placement, 1047 struct ttm_mem_reg *mem, 1048 uint32_t *new_flags) 1049 { 1050 int i; 1051 1052 for (i = 0; i < placement->num_placement; i++) { 1053 const struct ttm_place *heap = &placement->placement[i]; 1054 if (mem->mm_node && 1055 (mem->start < heap->fpfn || 1056 (heap->lpfn != 0 && (mem->start + mem->num_pages) > heap->lpfn))) 1057 continue; 1058 1059 *new_flags = heap->flags; 1060 if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) && 1061 (*new_flags & mem->placement & TTM_PL_MASK_MEM)) 1062 return true; 1063 } 1064 1065 for (i = 0; i < placement->num_busy_placement; i++) { 1066 const struct ttm_place *heap = &placement->busy_placement[i]; 1067 if (mem->mm_node && 1068 (mem->start < heap->fpfn || 1069 (heap->lpfn != 0 && (mem->start + mem->num_pages) > heap->lpfn))) 1070 continue; 1071 1072 *new_flags = heap->flags; 1073 if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) && 1074 (*new_flags & mem->placement & TTM_PL_MASK_MEM)) 1075 return true; 1076 } 1077 1078 return false; 1079 } 1080 EXPORT_SYMBOL(ttm_bo_mem_compat); 1081 1082 int ttm_bo_validate(struct ttm_buffer_object *bo, 1083 struct ttm_placement *placement, 1084 bool interruptible, 1085 bool no_wait_gpu) 1086 { 1087 int ret; 1088 uint32_t new_flags; 1089 1090 lockdep_assert_held(&bo->resv->lock.base); 1091 /* 1092 * Check whether we need to move buffer. 1093 */ 1094 if (!ttm_bo_mem_compat(placement, &bo->mem, &new_flags)) { 1095 ret = ttm_bo_move_buffer(bo, placement, interruptible, 1096 no_wait_gpu); 1097 if (ret) 1098 return ret; 1099 } else { 1100 /* 1101 * Use the access and other non-mapping-related flag bits from 1102 * the compatible memory placement flags to the active flags 1103 */ 1104 ttm_flag_masked(&bo->mem.placement, new_flags, 1105 ~TTM_PL_MASK_MEMTYPE); 1106 } 1107 /* 1108 * We might need to add a TTM. 1109 */ 1110 if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) { 1111 ret = ttm_bo_add_ttm(bo, true); 1112 if (ret) 1113 return ret; 1114 } 1115 return 0; 1116 } 1117 EXPORT_SYMBOL(ttm_bo_validate); 1118 1119 int ttm_bo_init(struct ttm_bo_device *bdev, 1120 struct ttm_buffer_object *bo, 1121 unsigned long size, 1122 enum ttm_bo_type type, 1123 struct ttm_placement *placement, 1124 uint32_t page_alignment, 1125 bool interruptible, 1126 struct file *persistent_swap_storage, 1127 size_t acc_size, 1128 struct sg_table *sg, 1129 struct reservation_object *resv, 1130 void (*destroy) (struct ttm_buffer_object *)) 1131 { 1132 int ret = 0; 1133 unsigned long num_pages; 1134 struct ttm_mem_global *mem_glob = bdev->glob->mem_glob; 1135 bool locked; 1136 1137 ret = ttm_mem_global_alloc(mem_glob, acc_size, false, false); 1138 if (ret) { 1139 pr_err("Out of kernel memory\n"); 1140 if (destroy) 1141 (*destroy)(bo); 1142 else 1143 kfree(bo); 1144 return -ENOMEM; 1145 } 1146 1147 num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT; 1148 if (num_pages == 0) { 1149 pr_err("Illegal buffer object size\n"); 1150 if (destroy) 1151 (*destroy)(bo); 1152 else 1153 kfree(bo); 1154 ttm_mem_global_free(mem_glob, acc_size); 1155 return -EINVAL; 1156 } 1157 bo->destroy = destroy; 1158 1159 kref_init(&bo->kref); 1160 kref_init(&bo->list_kref); 1161 atomic_set(&bo->cpu_writers, 0); 1162 INIT_LIST_HEAD(&bo->lru); 1163 INIT_LIST_HEAD(&bo->ddestroy); 1164 INIT_LIST_HEAD(&bo->swap); 1165 INIT_LIST_HEAD(&bo->io_reserve_lru); 1166 mutex_init(&bo->wu_mutex); 1167 bo->bdev = bdev; 1168 bo->glob = bdev->glob; 1169 bo->type = type; 1170 bo->num_pages = num_pages; 1171 bo->mem.size = num_pages << PAGE_SHIFT; 1172 bo->mem.mem_type = TTM_PL_SYSTEM; 1173 bo->mem.num_pages = bo->num_pages; 1174 bo->mem.mm_node = NULL; 1175 bo->mem.page_alignment = page_alignment; 1176 bo->mem.bus.io_reserved_vm = false; 1177 bo->mem.bus.io_reserved_count = 0; 1178 bo->moving = NULL; 1179 bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED); 1180 bo->persistent_swap_storage = persistent_swap_storage; 1181 bo->acc_size = acc_size; 1182 bo->sg = sg; 1183 if (resv) { 1184 bo->resv = resv; 1185 lockdep_assert_held(&bo->resv->lock.base); 1186 } else { 1187 bo->resv = &bo->ttm_resv; 1188 reservation_object_init(&bo->ttm_resv); 1189 } 1190 atomic_inc(&bo->glob->bo_count); 1191 drm_vma_node_reset(&bo->vma_node); 1192 1193 /* 1194 * For ttm_bo_type_device buffers, allocate 1195 * address space from the device. 1196 */ 1197 if (bo->type == ttm_bo_type_device || 1198 bo->type == ttm_bo_type_sg) 1199 ret = drm_vma_offset_add(&bdev->vma_manager, &bo->vma_node, 1200 bo->mem.num_pages); 1201 1202 /* passed reservation objects should already be locked, 1203 * since otherwise lockdep will be angered in radeon. 1204 */ 1205 if (!resv) { 1206 locked = ww_mutex_trylock(&bo->resv->lock); 1207 WARN_ON(!locked); 1208 } 1209 1210 if (likely(!ret)) 1211 ret = ttm_bo_validate(bo, placement, interruptible, false); 1212 1213 if (!resv) { 1214 ttm_bo_unreserve(bo); 1215 1216 } else if (!(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) { 1217 spin_lock(&bo->glob->lru_lock); 1218 ttm_bo_add_to_lru(bo); 1219 spin_unlock(&bo->glob->lru_lock); 1220 } 1221 1222 if (unlikely(ret)) 1223 ttm_bo_unref(&bo); 1224 1225 return ret; 1226 } 1227 EXPORT_SYMBOL(ttm_bo_init); 1228 1229 size_t ttm_bo_acc_size(struct ttm_bo_device *bdev, 1230 unsigned long bo_size, 1231 unsigned struct_size) 1232 { 1233 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT; 1234 size_t size = 0; 1235 1236 size += ttm_round_pot(struct_size); 1237 size += ttm_round_pot(npages * sizeof(void *)); 1238 size += ttm_round_pot(sizeof(struct ttm_tt)); 1239 return size; 1240 } 1241 EXPORT_SYMBOL(ttm_bo_acc_size); 1242 1243 size_t ttm_bo_dma_acc_size(struct ttm_bo_device *bdev, 1244 unsigned long bo_size, 1245 unsigned struct_size) 1246 { 1247 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT; 1248 size_t size = 0; 1249 1250 size += ttm_round_pot(struct_size); 1251 size += ttm_round_pot(npages * (2*sizeof(void *) + sizeof(dma_addr_t))); 1252 size += ttm_round_pot(sizeof(struct ttm_dma_tt)); 1253 return size; 1254 } 1255 EXPORT_SYMBOL(ttm_bo_dma_acc_size); 1256 1257 int ttm_bo_create(struct ttm_bo_device *bdev, 1258 unsigned long size, 1259 enum ttm_bo_type type, 1260 struct ttm_placement *placement, 1261 uint32_t page_alignment, 1262 bool interruptible, 1263 struct file *persistent_swap_storage, 1264 struct ttm_buffer_object **p_bo) 1265 { 1266 struct ttm_buffer_object *bo; 1267 size_t acc_size; 1268 int ret; 1269 1270 bo = kzalloc(sizeof(*bo), GFP_KERNEL); 1271 if (unlikely(bo == NULL)) 1272 return -ENOMEM; 1273 1274 acc_size = ttm_bo_acc_size(bdev, size, sizeof(struct ttm_buffer_object)); 1275 ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment, 1276 interruptible, persistent_swap_storage, acc_size, 1277 NULL, NULL, NULL); 1278 if (likely(ret == 0)) 1279 *p_bo = bo; 1280 1281 return ret; 1282 } 1283 EXPORT_SYMBOL(ttm_bo_create); 1284 1285 static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev, 1286 unsigned mem_type, bool allow_errors) 1287 { 1288 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 1289 struct ttm_bo_global *glob = bdev->glob; 1290 struct fence *fence; 1291 int ret; 1292 1293 /* 1294 * Can't use standard list traversal since we're unlocking. 1295 */ 1296 1297 spin_lock(&glob->lru_lock); 1298 while (!list_empty(&man->lru)) { 1299 spin_unlock(&glob->lru_lock); 1300 ret = ttm_mem_evict_first(bdev, mem_type, NULL, false, false); 1301 if (ret) { 1302 if (allow_errors) { 1303 return ret; 1304 } else { 1305 pr_err("Cleanup eviction failed\n"); 1306 } 1307 } 1308 spin_lock(&glob->lru_lock); 1309 } 1310 spin_unlock(&glob->lru_lock); 1311 1312 spin_lock(&man->move_lock); 1313 fence = fence_get(man->move); 1314 spin_unlock(&man->move_lock); 1315 1316 if (fence) { 1317 ret = fence_wait(fence, false); 1318 fence_put(fence); 1319 if (ret) { 1320 if (allow_errors) { 1321 return ret; 1322 } else { 1323 pr_err("Cleanup eviction failed\n"); 1324 } 1325 } 1326 } 1327 1328 return 0; 1329 } 1330 1331 int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type) 1332 { 1333 struct ttm_mem_type_manager *man; 1334 int ret = -EINVAL; 1335 1336 if (mem_type >= TTM_NUM_MEM_TYPES) { 1337 pr_err("Illegal memory type %d\n", mem_type); 1338 return ret; 1339 } 1340 man = &bdev->man[mem_type]; 1341 1342 if (!man->has_type) { 1343 pr_err("Trying to take down uninitialized memory manager type %u\n", 1344 mem_type); 1345 return ret; 1346 } 1347 fence_put(man->move); 1348 1349 man->use_type = false; 1350 man->has_type = false; 1351 1352 ret = 0; 1353 if (mem_type > 0) { 1354 ttm_bo_force_list_clean(bdev, mem_type, false); 1355 1356 ret = (*man->func->takedown)(man); 1357 } 1358 1359 return ret; 1360 } 1361 EXPORT_SYMBOL(ttm_bo_clean_mm); 1362 1363 int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type) 1364 { 1365 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 1366 1367 if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) { 1368 pr_err("Illegal memory manager memory type %u\n", mem_type); 1369 return -EINVAL; 1370 } 1371 1372 if (!man->has_type) { 1373 pr_err("Memory type %u has not been initialized\n", mem_type); 1374 return 0; 1375 } 1376 1377 return ttm_bo_force_list_clean(bdev, mem_type, true); 1378 } 1379 EXPORT_SYMBOL(ttm_bo_evict_mm); 1380 1381 int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type, 1382 unsigned long p_size) 1383 { 1384 int ret = -EINVAL; 1385 struct ttm_mem_type_manager *man; 1386 1387 BUG_ON(type >= TTM_NUM_MEM_TYPES); 1388 man = &bdev->man[type]; 1389 BUG_ON(man->has_type); 1390 man->io_reserve_fastpath = true; 1391 man->use_io_reserve_lru = false; 1392 mutex_init(&man->io_reserve_mutex); 1393 spin_lock_init(&man->move_lock); 1394 INIT_LIST_HEAD(&man->io_reserve_lru); 1395 1396 ret = bdev->driver->init_mem_type(bdev, type, man); 1397 if (ret) 1398 return ret; 1399 man->bdev = bdev; 1400 1401 ret = 0; 1402 if (type != TTM_PL_SYSTEM) { 1403 ret = (*man->func->init)(man, p_size); 1404 if (ret) 1405 return ret; 1406 } 1407 man->has_type = true; 1408 man->use_type = true; 1409 man->size = p_size; 1410 1411 INIT_LIST_HEAD(&man->lru); 1412 man->move = NULL; 1413 1414 return 0; 1415 } 1416 EXPORT_SYMBOL(ttm_bo_init_mm); 1417 1418 static void ttm_bo_global_kobj_release(struct kobject *kobj) 1419 { 1420 struct ttm_bo_global *glob = 1421 container_of(kobj, struct ttm_bo_global, kobj); 1422 1423 ttm_mem_unregister_shrink(glob->mem_glob, &glob->shrink); 1424 __free_page(glob->dummy_read_page); 1425 kfree(glob); 1426 } 1427 1428 void ttm_bo_global_release(struct drm_global_reference *ref) 1429 { 1430 struct ttm_bo_global *glob = ref->object; 1431 1432 kobject_del(&glob->kobj); 1433 kobject_put(&glob->kobj); 1434 } 1435 EXPORT_SYMBOL(ttm_bo_global_release); 1436 1437 int ttm_bo_global_init(struct drm_global_reference *ref) 1438 { 1439 struct ttm_bo_global_ref *bo_ref = 1440 container_of(ref, struct ttm_bo_global_ref, ref); 1441 struct ttm_bo_global *glob = ref->object; 1442 int ret; 1443 1444 mutex_init(&glob->device_list_mutex); 1445 spin_lock_init(&glob->lru_lock); 1446 glob->mem_glob = bo_ref->mem_glob; 1447 glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32); 1448 1449 if (unlikely(glob->dummy_read_page == NULL)) { 1450 ret = -ENOMEM; 1451 goto out_no_drp; 1452 } 1453 1454 INIT_LIST_HEAD(&glob->swap_lru); 1455 INIT_LIST_HEAD(&glob->device_list); 1456 1457 ttm_mem_init_shrink(&glob->shrink, ttm_bo_swapout); 1458 ret = ttm_mem_register_shrink(glob->mem_glob, &glob->shrink); 1459 if (unlikely(ret != 0)) { 1460 pr_err("Could not register buffer object swapout\n"); 1461 goto out_no_shrink; 1462 } 1463 1464 atomic_set(&glob->bo_count, 0); 1465 1466 ret = kobject_init_and_add( 1467 &glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects"); 1468 if (unlikely(ret != 0)) 1469 kobject_put(&glob->kobj); 1470 return ret; 1471 out_no_shrink: 1472 __free_page(glob->dummy_read_page); 1473 out_no_drp: 1474 kfree(glob); 1475 return ret; 1476 } 1477 EXPORT_SYMBOL(ttm_bo_global_init); 1478 1479 1480 int ttm_bo_device_release(struct ttm_bo_device *bdev) 1481 { 1482 int ret = 0; 1483 unsigned i = TTM_NUM_MEM_TYPES; 1484 struct ttm_mem_type_manager *man; 1485 struct ttm_bo_global *glob = bdev->glob; 1486 1487 while (i--) { 1488 man = &bdev->man[i]; 1489 if (man->has_type) { 1490 man->use_type = false; 1491 if ((i != TTM_PL_SYSTEM) && ttm_bo_clean_mm(bdev, i)) { 1492 ret = -EBUSY; 1493 pr_err("DRM memory manager type %d is not clean\n", 1494 i); 1495 } 1496 man->has_type = false; 1497 } 1498 } 1499 1500 mutex_lock(&glob->device_list_mutex); 1501 list_del(&bdev->device_list); 1502 mutex_unlock(&glob->device_list_mutex); 1503 1504 cancel_delayed_work_sync(&bdev->wq); 1505 1506 while (ttm_bo_delayed_delete(bdev, true)) 1507 ; 1508 1509 spin_lock(&glob->lru_lock); 1510 if (list_empty(&bdev->ddestroy)) 1511 TTM_DEBUG("Delayed destroy list was clean\n"); 1512 1513 if (list_empty(&bdev->man[0].lru)) 1514 TTM_DEBUG("Swap list was clean\n"); 1515 spin_unlock(&glob->lru_lock); 1516 1517 drm_vma_offset_manager_destroy(&bdev->vma_manager); 1518 1519 return ret; 1520 } 1521 EXPORT_SYMBOL(ttm_bo_device_release); 1522 1523 int ttm_bo_device_init(struct ttm_bo_device *bdev, 1524 struct ttm_bo_global *glob, 1525 struct ttm_bo_driver *driver, 1526 struct address_space *mapping, 1527 uint64_t file_page_offset, 1528 bool need_dma32) 1529 { 1530 int ret = -EINVAL; 1531 1532 bdev->driver = driver; 1533 1534 memset(bdev->man, 0, sizeof(bdev->man)); 1535 1536 /* 1537 * Initialize the system memory buffer type. 1538 * Other types need to be driver / IOCTL initialized. 1539 */ 1540 ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0); 1541 if (unlikely(ret != 0)) 1542 goto out_no_sys; 1543 1544 drm_vma_offset_manager_init(&bdev->vma_manager, file_page_offset, 1545 0x10000000); 1546 INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue); 1547 INIT_LIST_HEAD(&bdev->ddestroy); 1548 bdev->dev_mapping = mapping; 1549 bdev->glob = glob; 1550 bdev->need_dma32 = need_dma32; 1551 mutex_lock(&glob->device_list_mutex); 1552 list_add_tail(&bdev->device_list, &glob->device_list); 1553 mutex_unlock(&glob->device_list_mutex); 1554 1555 return 0; 1556 out_no_sys: 1557 return ret; 1558 } 1559 EXPORT_SYMBOL(ttm_bo_device_init); 1560 1561 /* 1562 * buffer object vm functions. 1563 */ 1564 1565 bool ttm_mem_reg_is_pci(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem) 1566 { 1567 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type]; 1568 1569 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) { 1570 if (mem->mem_type == TTM_PL_SYSTEM) 1571 return false; 1572 1573 if (man->flags & TTM_MEMTYPE_FLAG_CMA) 1574 return false; 1575 1576 if (mem->placement & TTM_PL_FLAG_CACHED) 1577 return false; 1578 } 1579 return true; 1580 } 1581 1582 void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo) 1583 { 1584 struct ttm_bo_device *bdev = bo->bdev; 1585 1586 drm_vma_node_unmap(&bo->vma_node, bdev->dev_mapping); 1587 ttm_mem_io_free_vm(bo); 1588 } 1589 1590 void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo) 1591 { 1592 struct ttm_bo_device *bdev = bo->bdev; 1593 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type]; 1594 1595 ttm_mem_io_lock(man, false); 1596 ttm_bo_unmap_virtual_locked(bo); 1597 ttm_mem_io_unlock(man); 1598 } 1599 1600 1601 EXPORT_SYMBOL(ttm_bo_unmap_virtual); 1602 1603 int ttm_bo_wait(struct ttm_buffer_object *bo, 1604 bool interruptible, bool no_wait) 1605 { 1606 long timeout = no_wait ? 0 : 15 * HZ; 1607 1608 timeout = reservation_object_wait_timeout_rcu(bo->resv, true, 1609 interruptible, timeout); 1610 if (timeout < 0) 1611 return timeout; 1612 1613 if (timeout == 0) 1614 return -EBUSY; 1615 1616 reservation_object_add_excl_fence(bo->resv, NULL); 1617 return 0; 1618 } 1619 EXPORT_SYMBOL(ttm_bo_wait); 1620 1621 int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait) 1622 { 1623 int ret = 0; 1624 1625 /* 1626 * Using ttm_bo_reserve makes sure the lru lists are updated. 1627 */ 1628 1629 ret = ttm_bo_reserve(bo, true, no_wait, NULL); 1630 if (unlikely(ret != 0)) 1631 return ret; 1632 ret = ttm_bo_wait(bo, true, no_wait); 1633 if (likely(ret == 0)) 1634 atomic_inc(&bo->cpu_writers); 1635 ttm_bo_unreserve(bo); 1636 return ret; 1637 } 1638 EXPORT_SYMBOL(ttm_bo_synccpu_write_grab); 1639 1640 void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo) 1641 { 1642 atomic_dec(&bo->cpu_writers); 1643 } 1644 EXPORT_SYMBOL(ttm_bo_synccpu_write_release); 1645 1646 /** 1647 * A buffer object shrink method that tries to swap out the first 1648 * buffer object on the bo_global::swap_lru list. 1649 */ 1650 1651 static int ttm_bo_swapout(struct ttm_mem_shrink *shrink) 1652 { 1653 struct ttm_bo_global *glob = 1654 container_of(shrink, struct ttm_bo_global, shrink); 1655 struct ttm_buffer_object *bo; 1656 int ret = -EBUSY; 1657 int put_count; 1658 uint32_t swap_placement = (TTM_PL_FLAG_CACHED | TTM_PL_FLAG_SYSTEM); 1659 1660 spin_lock(&glob->lru_lock); 1661 list_for_each_entry(bo, &glob->swap_lru, swap) { 1662 ret = __ttm_bo_reserve(bo, false, true, NULL); 1663 if (!ret) 1664 break; 1665 } 1666 1667 if (ret) { 1668 spin_unlock(&glob->lru_lock); 1669 return ret; 1670 } 1671 1672 kref_get(&bo->list_kref); 1673 1674 if (!list_empty(&bo->ddestroy)) { 1675 ret = ttm_bo_cleanup_refs_and_unlock(bo, false, false); 1676 kref_put(&bo->list_kref, ttm_bo_release_list); 1677 return ret; 1678 } 1679 1680 put_count = ttm_bo_del_from_lru(bo); 1681 spin_unlock(&glob->lru_lock); 1682 1683 ttm_bo_list_ref_sub(bo, put_count, true); 1684 1685 /** 1686 * Move to system cached 1687 */ 1688 1689 if ((bo->mem.placement & swap_placement) != swap_placement) { 1690 struct ttm_mem_reg evict_mem; 1691 1692 evict_mem = bo->mem; 1693 evict_mem.mm_node = NULL; 1694 evict_mem.placement = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED; 1695 evict_mem.mem_type = TTM_PL_SYSTEM; 1696 1697 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, 1698 false, false); 1699 if (unlikely(ret != 0)) 1700 goto out; 1701 } 1702 1703 /** 1704 * Make sure BO is idle. 1705 */ 1706 1707 ret = ttm_bo_wait(bo, false, false); 1708 if (unlikely(ret != 0)) 1709 goto out; 1710 1711 ttm_bo_unmap_virtual(bo); 1712 1713 /** 1714 * Swap out. Buffer will be swapped in again as soon as 1715 * anyone tries to access a ttm page. 1716 */ 1717 1718 if (bo->bdev->driver->swap_notify) 1719 bo->bdev->driver->swap_notify(bo); 1720 1721 ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage); 1722 out: 1723 1724 /** 1725 * 1726 * Unreserve without putting on LRU to avoid swapping out an 1727 * already swapped buffer. 1728 */ 1729 1730 __ttm_bo_unreserve(bo); 1731 kref_put(&bo->list_kref, ttm_bo_release_list); 1732 return ret; 1733 } 1734 1735 void ttm_bo_swapout_all(struct ttm_bo_device *bdev) 1736 { 1737 while (ttm_bo_swapout(&bdev->glob->shrink) == 0) 1738 ; 1739 } 1740 EXPORT_SYMBOL(ttm_bo_swapout_all); 1741 1742 /** 1743 * ttm_bo_wait_unreserved - interruptible wait for a buffer object to become 1744 * unreserved 1745 * 1746 * @bo: Pointer to buffer 1747 */ 1748 int ttm_bo_wait_unreserved(struct ttm_buffer_object *bo) 1749 { 1750 int ret; 1751 1752 /* 1753 * In the absense of a wait_unlocked API, 1754 * Use the bo::wu_mutex to avoid triggering livelocks due to 1755 * concurrent use of this function. Note that this use of 1756 * bo::wu_mutex can go away if we change locking order to 1757 * mmap_sem -> bo::reserve. 1758 */ 1759 ret = mutex_lock_interruptible(&bo->wu_mutex); 1760 if (unlikely(ret != 0)) 1761 return -ERESTARTSYS; 1762 if (!ww_mutex_is_locked(&bo->resv->lock)) 1763 goto out_unlock; 1764 ret = __ttm_bo_reserve(bo, true, false, NULL); 1765 if (unlikely(ret != 0)) 1766 goto out_unlock; 1767 __ttm_bo_unreserve(bo); 1768 1769 out_unlock: 1770 mutex_unlock(&bo->wu_mutex); 1771 return ret; 1772 } 1773