1 /* $OpenBSD: ttm_bo.c,v 1.19 2015/10/23 08:21:58 kettenis Exp $ */ 2 /************************************************************************** 3 * 4 * Copyright (c) 2006-2009 VMware, Inc., Palo Alto, CA., USA 5 * All Rights Reserved. 6 * 7 * Permission is hereby granted, free of charge, to any person obtaining a 8 * copy of this software and associated documentation files (the 9 * "Software"), to deal in the Software without restriction, including 10 * without limitation the rights to use, copy, modify, merge, publish, 11 * distribute, sub license, and/or sell copies of the Software, and to 12 * permit persons to whom the Software is furnished to do so, subject to 13 * the following conditions: 14 * 15 * The above copyright notice and this permission notice (including the 16 * next paragraph) shall be included in all copies or substantial portions 17 * of the Software. 18 * 19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 20 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 21 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL 22 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, 23 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR 24 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE 25 * USE OR OTHER DEALINGS IN THE SOFTWARE. 26 * 27 **************************************************************************/ 28 /* 29 * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com> 30 */ 31 32 #define pr_fmt(fmt) "[TTM] " fmt 33 34 #include <dev/pci/drm/ttm/ttm_module.h> 35 #include <dev/pci/drm/ttm/ttm_bo_driver.h> 36 #include <dev/pci/drm/ttm/ttm_placement.h> 37 38 #define TTM_ASSERT_LOCKED(param) 39 #define TTM_DEBUG(fmt, arg...) 40 #define TTM_BO_HASH_ORDER 13 41 42 static int ttm_bo_setup_vm(struct ttm_buffer_object *bo); 43 static int ttm_bo_swapout(struct ttm_mem_shrink *shrink); 44 static void ttm_bo_global_kobj_release(struct kobject *kobj); 45 46 int ttm_bo_move_buffer(struct ttm_buffer_object *, struct ttm_placement *, 47 bool, bool); 48 49 #ifdef notyet 50 static struct attribute ttm_bo_count = { 51 .name = "bo_count", 52 .mode = S_IRUGO 53 }; 54 #endif 55 56 struct kobject * 57 ttm_get_kobj(void) 58 { 59 return (NULL); 60 } 61 62 static inline int ttm_mem_type_from_flags(uint32_t flags, uint32_t *mem_type) 63 { 64 int i; 65 66 for (i = 0; i <= TTM_PL_PRIV5; i++) 67 if (flags & (1 << i)) { 68 *mem_type = i; 69 return 0; 70 } 71 return -EINVAL; 72 } 73 74 static void ttm_mem_type_debug(struct ttm_bo_device *bdev, int mem_type) 75 { 76 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 77 78 pr_err(" has_type: %d\n", man->has_type); 79 pr_err(" use_type: %d\n", man->use_type); 80 pr_err(" flags: 0x%08X\n", man->flags); 81 pr_err(" gpu_offset: 0x%08lX\n", man->gpu_offset); 82 pr_err(" size: %llu\n", man->size); 83 pr_err(" available_caching: 0x%08X\n", man->available_caching); 84 pr_err(" default_caching: 0x%08X\n", man->default_caching); 85 if (mem_type != TTM_PL_SYSTEM) 86 (*man->func->debug)(man, TTM_PFX); 87 } 88 89 static void ttm_bo_mem_space_debug(struct ttm_buffer_object *bo, 90 struct ttm_placement *placement) 91 { 92 int i, ret, mem_type; 93 94 pr_err("No space for %p (%lu pages, %luK, %luM)\n", 95 bo, bo->mem.num_pages, bo->mem.size >> 10, 96 bo->mem.size >> 20); 97 for (i = 0; i < placement->num_placement; i++) { 98 ret = ttm_mem_type_from_flags(placement->placement[i], 99 &mem_type); 100 if (ret) 101 return; 102 pr_err(" placement[%d]=0x%08X (%d)\n", 103 i, placement->placement[i], mem_type); 104 ttm_mem_type_debug(bo->bdev, mem_type); 105 } 106 } 107 108 #ifdef notyet 109 static ssize_t ttm_bo_global_show(struct kobject *kobj, 110 struct attribute *attr, 111 char *buffer) 112 { 113 struct ttm_bo_global *glob = 114 container_of(kobj, struct ttm_bo_global, kobj); 115 116 return snprintf(buffer, PAGE_SIZE, "%lu\n", 117 (unsigned long) atomic_read(&glob->bo_count)); 118 } 119 120 static struct attribute *ttm_bo_global_attrs[] = { 121 &ttm_bo_count, 122 NULL 123 }; 124 125 static const struct sysfs_ops ttm_bo_global_ops = { 126 .show = &ttm_bo_global_show 127 }; 128 #endif 129 130 static struct kobj_type ttm_bo_glob_kobj_type = { 131 .release = &ttm_bo_global_kobj_release, 132 #ifdef __linux__ 133 .sysfs_ops = &ttm_bo_global_ops, 134 .default_attrs = ttm_bo_global_attrs 135 #endif 136 }; 137 138 static inline uint32_t ttm_bo_type_flags(unsigned type) 139 { 140 return 1 << (type); 141 } 142 143 static void ttm_bo_release_list(struct kref *list_kref) 144 { 145 struct ttm_buffer_object *bo = 146 container_of(list_kref, struct ttm_buffer_object, list_kref); 147 struct ttm_bo_device *bdev = bo->bdev; 148 size_t acc_size = bo->acc_size; 149 150 BUG_ON(atomic_read(&bo->list_kref.refcount)); 151 BUG_ON(atomic_read(&bo->kref.refcount)); 152 BUG_ON(atomic_read(&bo->cpu_writers)); 153 BUG_ON(bo->sync_obj != NULL); 154 BUG_ON(bo->mem.mm_node != NULL); 155 BUG_ON(!list_empty(&bo->lru)); 156 BUG_ON(!list_empty(&bo->ddestroy)); 157 158 if (bo->ttm) 159 ttm_tt_destroy(bo->ttm); 160 atomic_dec(&bo->glob->bo_count); 161 if (bo->destroy) 162 bo->destroy(bo); 163 else { 164 kfree(bo); 165 } 166 ttm_mem_global_free(bdev->glob->mem_glob, acc_size); 167 } 168 169 int ttm_bo_wait_unreserved(struct ttm_buffer_object *bo, bool interruptible) 170 { 171 int ret = 0; 172 173 while (ret == 0) { 174 if (!ttm_bo_is_reserved(bo)) 175 break; 176 ret = -tsleep(&bo->event_queue, 177 PZERO | (interruptible ? PCATCH : 0), "ttmwt", 0); 178 179 } 180 181 return (ret); 182 } 183 EXPORT_SYMBOL(ttm_bo_wait_unreserved); 184 185 void ttm_bo_add_to_lru(struct ttm_buffer_object *bo) 186 { 187 struct ttm_bo_device *bdev = bo->bdev; 188 struct ttm_mem_type_manager *man; 189 190 BUG_ON(!ttm_bo_is_reserved(bo)); 191 192 if (!(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) { 193 194 BUG_ON(!list_empty(&bo->lru)); 195 196 man = &bdev->man[bo->mem.mem_type]; 197 list_add_tail(&bo->lru, &man->lru); 198 kref_get(&bo->list_kref); 199 200 if (bo->ttm != NULL) { 201 list_add_tail(&bo->swap, &bo->glob->swap_lru); 202 kref_get(&bo->list_kref); 203 } 204 } 205 } 206 207 int ttm_bo_del_from_lru(struct ttm_buffer_object *bo) 208 { 209 int put_count = 0; 210 211 if (!list_empty(&bo->swap)) { 212 list_del_init(&bo->swap); 213 ++put_count; 214 } 215 if (!list_empty(&bo->lru)) { 216 list_del_init(&bo->lru); 217 ++put_count; 218 } 219 220 /* 221 * TODO: Add a driver hook to delete from 222 * driver-specific LRU's here. 223 */ 224 225 return put_count; 226 } 227 228 int ttm_bo_reserve_locked(struct ttm_buffer_object *bo, 229 bool interruptible, 230 bool no_wait, bool use_sequence, uint32_t sequence) 231 { 232 struct ttm_bo_global *glob = bo->glob; 233 int ret; 234 235 while (unlikely(atomic_read(&bo->reserved) != 0)) { 236 /** 237 * Deadlock avoidance for multi-bo reserving. 238 */ 239 if (use_sequence && bo->seq_valid) { 240 /** 241 * We've already reserved this one. 242 */ 243 if (unlikely(sequence == bo->val_seq)) 244 return -EDEADLK; 245 /** 246 * Already reserved by a thread that will not back 247 * off for us. We need to back off. 248 */ 249 if (unlikely(sequence - bo->val_seq < (1 << 31))) 250 return -EAGAIN; 251 } 252 253 if (no_wait) 254 return -EBUSY; 255 256 spin_unlock(&glob->lru_lock); 257 ret = ttm_bo_wait_unreserved(bo, interruptible); 258 spin_lock(&glob->lru_lock); 259 260 if (unlikely(ret)) 261 return ret; 262 } 263 264 atomic_set(&bo->reserved, 1); 265 if (use_sequence) { 266 /** 267 * Wake up waiters that may need to recheck for deadlock, 268 * if we decreased the sequence number. 269 */ 270 if (unlikely((bo->val_seq - sequence < (1 << 31)) 271 || !bo->seq_valid)) 272 wake_up_all(&bo->event_queue); 273 274 bo->val_seq = sequence; 275 bo->seq_valid = true; 276 } else { 277 bo->seq_valid = false; 278 } 279 280 return 0; 281 } 282 EXPORT_SYMBOL(ttm_bo_reserve); 283 284 static void ttm_bo_ref_bug(struct kref *list_kref) 285 { 286 BUG(); 287 } 288 289 void ttm_bo_list_ref_sub(struct ttm_buffer_object *bo, int count, 290 bool never_free) 291 { 292 kref_sub(&bo->list_kref, count, 293 (never_free) ? ttm_bo_ref_bug : ttm_bo_release_list); 294 } 295 296 int ttm_bo_reserve(struct ttm_buffer_object *bo, 297 bool interruptible, 298 bool no_wait, bool use_sequence, uint32_t sequence) 299 { 300 struct ttm_bo_global *glob = bo->glob; 301 int put_count = 0; 302 int ret; 303 304 spin_lock(&glob->lru_lock); 305 ret = ttm_bo_reserve_locked(bo, interruptible, no_wait, use_sequence, 306 sequence); 307 if (likely(ret == 0)) 308 put_count = ttm_bo_del_from_lru(bo); 309 spin_unlock(&glob->lru_lock); 310 311 ttm_bo_list_ref_sub(bo, put_count, true); 312 313 return ret; 314 } 315 316 void ttm_bo_unreserve_locked(struct ttm_buffer_object *bo) 317 { 318 ttm_bo_add_to_lru(bo); 319 atomic_set(&bo->reserved, 0); 320 wake_up_all(&bo->event_queue); 321 } 322 323 void ttm_bo_unreserve(struct ttm_buffer_object *bo) 324 { 325 struct ttm_bo_global *glob = bo->glob; 326 327 spin_lock(&glob->lru_lock); 328 ttm_bo_unreserve_locked(bo); 329 spin_unlock(&glob->lru_lock); 330 } 331 EXPORT_SYMBOL(ttm_bo_unreserve); 332 333 /* 334 * Call bo->mutex locked. 335 */ 336 static int ttm_bo_add_ttm(struct ttm_buffer_object *bo, bool zero_alloc) 337 { 338 struct ttm_bo_device *bdev = bo->bdev; 339 struct ttm_bo_global *glob = bo->glob; 340 int ret = 0; 341 uint32_t page_flags = 0; 342 343 #ifdef notyet 344 rw_assert_wrlock(&bo->mutex); 345 #endif 346 bo->ttm = NULL; 347 348 if (bdev->need_dma32) 349 page_flags |= TTM_PAGE_FLAG_DMA32; 350 351 switch (bo->type) { 352 case ttm_bo_type_device: 353 if (zero_alloc) 354 page_flags |= TTM_PAGE_FLAG_ZERO_ALLOC; 355 case ttm_bo_type_kernel: 356 bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT, 357 page_flags, glob->dummy_read_page); 358 if (unlikely(bo->ttm == NULL)) 359 ret = -ENOMEM; 360 break; 361 case ttm_bo_type_sg: 362 bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT, 363 page_flags | TTM_PAGE_FLAG_SG, 364 glob->dummy_read_page); 365 if (unlikely(bo->ttm == NULL)) { 366 ret = -ENOMEM; 367 break; 368 } 369 bo->ttm->sg = bo->sg; 370 break; 371 default: 372 pr_err("Illegal buffer object type\n"); 373 ret = -EINVAL; 374 break; 375 } 376 377 return ret; 378 } 379 380 static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo, 381 struct ttm_mem_reg *mem, 382 bool evict, bool interruptible, 383 bool no_wait_gpu) 384 { 385 struct ttm_bo_device *bdev = bo->bdev; 386 bool old_is_pci = ttm_mem_reg_is_pci(bdev, &bo->mem); 387 bool new_is_pci = ttm_mem_reg_is_pci(bdev, mem); 388 struct ttm_mem_type_manager *old_man = &bdev->man[bo->mem.mem_type]; 389 struct ttm_mem_type_manager *new_man = &bdev->man[mem->mem_type]; 390 int ret = 0; 391 392 if (old_is_pci || new_is_pci || 393 ((mem->placement & bo->mem.placement & TTM_PL_MASK_CACHING) == 0)) { 394 ret = ttm_mem_io_lock(old_man, true); 395 if (unlikely(ret != 0)) 396 goto out_err; 397 ttm_bo_unmap_virtual_locked(bo); 398 ttm_mem_io_unlock(old_man); 399 } 400 401 /* 402 * Create and bind a ttm if required. 403 */ 404 405 if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) { 406 if (bo->ttm == NULL) { 407 bool zero = !(old_man->flags & TTM_MEMTYPE_FLAG_FIXED); 408 ret = ttm_bo_add_ttm(bo, zero); 409 if (ret) 410 goto out_err; 411 } 412 413 ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement); 414 if (ret) 415 goto out_err; 416 417 if (mem->mem_type != TTM_PL_SYSTEM) { 418 ret = ttm_tt_bind(bo->ttm, mem); 419 if (ret) 420 goto out_err; 421 } 422 423 if (bo->mem.mem_type == TTM_PL_SYSTEM) { 424 if (bdev->driver->move_notify) 425 bdev->driver->move_notify(bo, mem); 426 bo->mem = *mem; 427 mem->mm_node = NULL; 428 goto moved; 429 } 430 } 431 432 if (bdev->driver->move_notify) 433 bdev->driver->move_notify(bo, mem); 434 435 if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) && 436 !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) 437 ret = ttm_bo_move_ttm(bo, evict, no_wait_gpu, mem); 438 else if (bdev->driver->move) 439 ret = bdev->driver->move(bo, evict, interruptible, 440 no_wait_gpu, mem); 441 else 442 ret = ttm_bo_move_memcpy(bo, evict, no_wait_gpu, mem); 443 444 if (ret) { 445 if (bdev->driver->move_notify) { 446 struct ttm_mem_reg tmp_mem = *mem; 447 *mem = bo->mem; 448 bo->mem = tmp_mem; 449 bdev->driver->move_notify(bo, mem); 450 bo->mem = *mem; 451 *mem = tmp_mem; 452 } 453 454 goto out_err; 455 } 456 457 moved: 458 if (bo->evicted) { 459 if (bdev->driver->invalidate_caches) { 460 ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement); 461 if (ret) 462 pr_err("Can not flush read caches\n"); 463 } 464 bo->evicted = false; 465 } 466 467 if (bo->mem.mm_node) { 468 bo->offset = (bo->mem.start << PAGE_SHIFT) + 469 bdev->man[bo->mem.mem_type].gpu_offset; 470 bo->cur_placement = bo->mem.placement; 471 } else 472 bo->offset = 0; 473 474 return 0; 475 476 out_err: 477 new_man = &bdev->man[bo->mem.mem_type]; 478 if ((new_man->flags & TTM_MEMTYPE_FLAG_FIXED) && bo->ttm) { 479 ttm_tt_unbind(bo->ttm); 480 ttm_tt_destroy(bo->ttm); 481 bo->ttm = NULL; 482 } 483 484 return ret; 485 } 486 487 /** 488 * Call bo::reserved. 489 * Will release GPU memory type usage on destruction. 490 * This is the place to put in driver specific hooks to release 491 * driver private resources. 492 * Will release the bo::reserved lock. 493 */ 494 495 static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo) 496 { 497 if (bo->bdev->driver->move_notify) 498 bo->bdev->driver->move_notify(bo, NULL); 499 500 if (bo->ttm) { 501 ttm_tt_unbind(bo->ttm); 502 ttm_tt_destroy(bo->ttm); 503 bo->ttm = NULL; 504 } 505 ttm_bo_mem_put(bo, &bo->mem); 506 507 atomic_set(&bo->reserved, 0); 508 wake_up_all(&bo->event_queue); 509 510 /* 511 * Since the final reference to this bo may not be dropped by 512 * the current task we have to put a memory barrier here to make 513 * sure the changes done in this function are always visible. 514 * 515 * This function only needs protection against the final kref_put. 516 */ 517 smp_mb__before_atomic_dec(); 518 } 519 520 static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo) 521 { 522 struct ttm_bo_device *bdev = bo->bdev; 523 struct ttm_bo_global *glob = bo->glob; 524 struct ttm_bo_driver *driver = bdev->driver; 525 void *sync_obj = NULL; 526 int put_count; 527 int ret; 528 529 spin_lock(&glob->lru_lock); 530 ret = ttm_bo_reserve_locked(bo, false, true, false, 0); 531 532 spin_lock(&bdev->fence_lock); 533 (void) ttm_bo_wait(bo, false, false, true); 534 if (!ret && !bo->sync_obj) { 535 spin_unlock(&bdev->fence_lock); 536 put_count = ttm_bo_del_from_lru(bo); 537 538 spin_unlock(&glob->lru_lock); 539 ttm_bo_cleanup_memtype_use(bo); 540 541 ttm_bo_list_ref_sub(bo, put_count, true); 542 543 return; 544 } 545 if (bo->sync_obj) 546 sync_obj = driver->sync_obj_ref(bo->sync_obj); 547 spin_unlock(&bdev->fence_lock); 548 549 if (!ret) { 550 atomic_set(&bo->reserved, 0); 551 wake_up_all(&bo->event_queue); 552 } 553 554 kref_get(&bo->list_kref); 555 list_add_tail(&bo->ddestroy, &bdev->ddestroy); 556 spin_unlock(&glob->lru_lock); 557 558 if (sync_obj) { 559 driver->sync_obj_flush(sync_obj); 560 driver->sync_obj_unref(&sync_obj); 561 } 562 schedule_delayed_work(&bdev->wq, 563 ((HZ / 100) < 1) ? 1 : HZ / 100); 564 } 565 566 /** 567 * function ttm_bo_cleanup_refs_and_unlock 568 * If bo idle, remove from delayed- and lru lists, and unref. 569 * If not idle, do nothing. 570 * 571 * Must be called with lru_lock and reservation held, this function 572 * will drop both before returning. 573 * 574 * @interruptible Any sleeps should occur interruptibly. 575 * @no_wait_gpu Never wait for gpu. Return -EBUSY instead. 576 */ 577 578 static int ttm_bo_cleanup_refs_and_unlock(struct ttm_buffer_object *bo, 579 bool interruptible, 580 bool no_wait_gpu) 581 { 582 struct ttm_bo_device *bdev = bo->bdev; 583 struct ttm_bo_driver *driver = bdev->driver; 584 struct ttm_bo_global *glob = bo->glob; 585 int put_count; 586 int ret; 587 588 spin_lock(&bdev->fence_lock); 589 ret = ttm_bo_wait(bo, false, false, true); 590 591 if (ret && !no_wait_gpu) { 592 void *sync_obj; 593 594 /* 595 * Take a reference to the fence and unreserve, 596 * at this point the buffer should be dead, so 597 * no new sync objects can be attached. 598 */ 599 sync_obj = driver->sync_obj_ref(bo->sync_obj); 600 spin_unlock(&bdev->fence_lock); 601 602 atomic_set(&bo->reserved, 0); 603 wake_up_all(&bo->event_queue); 604 spin_unlock(&glob->lru_lock); 605 606 ret = driver->sync_obj_wait(sync_obj, false, interruptible); 607 driver->sync_obj_unref(&sync_obj); 608 if (ret) 609 return ret; 610 611 /* 612 * remove sync_obj with ttm_bo_wait, the wait should be 613 * finished, and no new wait object should have been added. 614 */ 615 spin_lock(&bdev->fence_lock); 616 ret = ttm_bo_wait(bo, false, false, true); 617 WARN_ON(ret); 618 spin_unlock(&bdev->fence_lock); 619 if (ret) 620 return ret; 621 622 spin_lock(&glob->lru_lock); 623 ret = ttm_bo_reserve_locked(bo, false, true, false, 0); 624 625 /* 626 * We raced, and lost, someone else holds the reservation now, 627 * and is probably busy in ttm_bo_cleanup_memtype_use. 628 * 629 * Even if it's not the case, because we finished waiting any 630 * delayed destruction would succeed, so just return success 631 * here. 632 */ 633 if (ret) { 634 spin_unlock(&glob->lru_lock); 635 return 0; 636 } 637 } else 638 spin_unlock(&bdev->fence_lock); 639 640 if (ret || unlikely(list_empty(&bo->ddestroy))) { 641 atomic_set(&bo->reserved, 0); 642 wake_up_all(&bo->event_queue); 643 spin_unlock(&glob->lru_lock); 644 return ret; 645 } 646 647 put_count = ttm_bo_del_from_lru(bo); 648 list_del_init(&bo->ddestroy); 649 ++put_count; 650 651 spin_unlock(&glob->lru_lock); 652 ttm_bo_cleanup_memtype_use(bo); 653 654 ttm_bo_list_ref_sub(bo, put_count, true); 655 656 return 0; 657 } 658 659 /** 660 * Traverse the delayed list, and call ttm_bo_cleanup_refs on all 661 * encountered buffers. 662 */ 663 664 static int ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all) 665 { 666 struct ttm_bo_global *glob = bdev->glob; 667 struct ttm_buffer_object *entry = NULL; 668 int ret = 0; 669 670 spin_lock(&glob->lru_lock); 671 if (list_empty(&bdev->ddestroy)) 672 goto out_unlock; 673 674 entry = list_first_entry(&bdev->ddestroy, 675 struct ttm_buffer_object, ddestroy); 676 kref_get(&entry->list_kref); 677 678 for (;;) { 679 struct ttm_buffer_object *nentry = NULL; 680 681 if (entry->ddestroy.next != &bdev->ddestroy) { 682 nentry = list_first_entry(&entry->ddestroy, 683 struct ttm_buffer_object, ddestroy); 684 kref_get(&nentry->list_kref); 685 } 686 687 ret = ttm_bo_reserve_locked(entry, false, !remove_all, false, 0); 688 if (!ret) 689 ret = ttm_bo_cleanup_refs_and_unlock(entry, false, 690 !remove_all); 691 else 692 spin_unlock(&glob->lru_lock); 693 694 kref_put(&entry->list_kref, ttm_bo_release_list); 695 entry = nentry; 696 697 if (ret || !entry) 698 goto out; 699 700 spin_lock(&glob->lru_lock); 701 if (list_empty(&entry->ddestroy)) 702 break; 703 } 704 705 out_unlock: 706 spin_unlock(&glob->lru_lock); 707 out: 708 if (entry) 709 kref_put(&entry->list_kref, ttm_bo_release_list); 710 return ret; 711 } 712 713 static void ttm_bo_delayed_workqueue(struct work_struct *work) 714 { 715 struct ttm_bo_device *bdev = 716 container_of(work, struct ttm_bo_device, wq.work); 717 718 if (ttm_bo_delayed_delete(bdev, false)) { 719 schedule_delayed_work(&bdev->wq, 720 ((HZ / 100) < 1) ? 1 : HZ / 100); 721 } 722 } 723 724 static void ttm_bo_release(struct kref *kref) 725 { 726 struct ttm_buffer_object *bo = 727 container_of(kref, struct ttm_buffer_object, kref); 728 struct ttm_bo_device *bdev = bo->bdev; 729 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type]; 730 731 write_lock(&bdev->vm_lock); 732 if (likely(bo->vm_node != NULL)) { 733 RB_REMOVE(ttm_bo_device_buffer_objects, 734 &bdev->addr_space_rb, bo); 735 drm_mm_put_block(bo->vm_node); 736 bo->vm_node = NULL; 737 } 738 write_unlock(&bdev->vm_lock); 739 ttm_mem_io_lock(man, false); 740 ttm_mem_io_free_vm(bo); 741 ttm_mem_io_unlock(man); 742 ttm_bo_cleanup_refs_or_queue(bo); 743 kref_put(&bo->list_kref, ttm_bo_release_list); 744 } 745 746 void ttm_bo_unref(struct ttm_buffer_object **p_bo) 747 { 748 struct ttm_buffer_object *bo = *p_bo; 749 750 *p_bo = NULL; 751 kref_put(&bo->kref, ttm_bo_release); 752 } 753 EXPORT_SYMBOL(ttm_bo_unref); 754 755 int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev) 756 { 757 return cancel_delayed_work_sync(&bdev->wq); 758 } 759 EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue); 760 761 void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched) 762 { 763 if (resched) 764 schedule_delayed_work(&bdev->wq, 765 ((HZ / 100) < 1) ? 1 : HZ / 100); 766 } 767 EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue); 768 769 static int ttm_bo_evict(struct ttm_buffer_object *bo, bool interruptible, 770 bool no_wait_gpu) 771 { 772 struct ttm_bo_device *bdev = bo->bdev; 773 struct ttm_mem_reg evict_mem; 774 struct ttm_placement placement; 775 int ret = 0; 776 777 spin_lock(&bdev->fence_lock); 778 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu); 779 spin_unlock(&bdev->fence_lock); 780 781 if (unlikely(ret != 0)) { 782 if (ret != -ERESTARTSYS) { 783 pr_err("Failed to expire sync object before buffer eviction\n"); 784 } 785 goto out; 786 } 787 788 BUG_ON(!ttm_bo_is_reserved(bo)); 789 790 evict_mem = bo->mem; 791 evict_mem.mm_node = NULL; 792 evict_mem.bus.io_reserved_vm = false; 793 evict_mem.bus.io_reserved_count = 0; 794 795 placement.fpfn = 0; 796 placement.lpfn = 0; 797 placement.num_placement = 0; 798 placement.num_busy_placement = 0; 799 bdev->driver->evict_flags(bo, &placement); 800 ret = ttm_bo_mem_space(bo, &placement, &evict_mem, interruptible, 801 no_wait_gpu); 802 if (ret) { 803 if (ret != -ERESTARTSYS) { 804 pr_err("Failed to find memory space for buffer 0x%p eviction\n", 805 bo); 806 ttm_bo_mem_space_debug(bo, &placement); 807 } 808 goto out; 809 } 810 811 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, interruptible, 812 no_wait_gpu); 813 if (ret) { 814 if (ret != -ERESTARTSYS) 815 pr_err("Buffer eviction failed\n"); 816 ttm_bo_mem_put(bo, &evict_mem); 817 goto out; 818 } 819 bo->evicted = true; 820 out: 821 return ret; 822 } 823 824 static int ttm_mem_evict_first(struct ttm_bo_device *bdev, 825 uint32_t mem_type, 826 bool interruptible, 827 bool no_wait_gpu) 828 { 829 struct ttm_bo_global *glob = bdev->glob; 830 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 831 struct ttm_buffer_object *bo; 832 int ret = -EBUSY, put_count; 833 834 spin_lock(&glob->lru_lock); 835 list_for_each_entry(bo, &man->lru, lru) { 836 ret = ttm_bo_reserve_locked(bo, false, true, false, 0); 837 if (!ret) 838 break; 839 } 840 841 if (ret) { 842 spin_unlock(&glob->lru_lock); 843 return ret; 844 } 845 846 kref_get(&bo->list_kref); 847 848 if (!list_empty(&bo->ddestroy)) { 849 ret = ttm_bo_cleanup_refs_and_unlock(bo, interruptible, 850 no_wait_gpu); 851 kref_put(&bo->list_kref, ttm_bo_release_list); 852 return ret; 853 } 854 855 put_count = ttm_bo_del_from_lru(bo); 856 spin_unlock(&glob->lru_lock); 857 858 BUG_ON(ret != 0); 859 860 ttm_bo_list_ref_sub(bo, put_count, true); 861 862 ret = ttm_bo_evict(bo, interruptible, no_wait_gpu); 863 ttm_bo_unreserve(bo); 864 865 kref_put(&bo->list_kref, ttm_bo_release_list); 866 return ret; 867 } 868 869 void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem) 870 { 871 struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type]; 872 873 if (mem->mm_node) 874 (*man->func->put_node)(man, mem); 875 } 876 EXPORT_SYMBOL(ttm_bo_mem_put); 877 878 /** 879 * Repeatedly evict memory from the LRU for @mem_type until we create enough 880 * space, or we've evicted everything and there isn't enough space. 881 */ 882 static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo, 883 uint32_t mem_type, 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 = &bdev->man[mem_type]; 891 int ret; 892 893 do { 894 ret = (*man->func->get_node)(man, bo, placement, mem); 895 if (unlikely(ret != 0)) 896 return ret; 897 if (mem->mm_node) 898 break; 899 ret = ttm_mem_evict_first(bdev, mem_type, 900 interruptible, no_wait_gpu); 901 if (unlikely(ret != 0)) 902 return ret; 903 } while (1); 904 if (mem->mm_node == NULL) 905 return -ENOMEM; 906 mem->mem_type = mem_type; 907 return 0; 908 } 909 910 static uint32_t ttm_bo_select_caching(struct ttm_mem_type_manager *man, 911 uint32_t cur_placement, 912 uint32_t proposed_placement) 913 { 914 uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING; 915 uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING; 916 917 /** 918 * Keep current caching if possible. 919 */ 920 921 if ((cur_placement & caching) != 0) 922 result |= (cur_placement & caching); 923 else if ((man->default_caching & caching) != 0) 924 result |= man->default_caching; 925 else if ((TTM_PL_FLAG_CACHED & caching) != 0) 926 result |= TTM_PL_FLAG_CACHED; 927 else if ((TTM_PL_FLAG_WC & caching) != 0) 928 result |= TTM_PL_FLAG_WC; 929 else if ((TTM_PL_FLAG_UNCACHED & caching) != 0) 930 result |= TTM_PL_FLAG_UNCACHED; 931 932 return result; 933 } 934 935 static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man, 936 uint32_t mem_type, 937 uint32_t proposed_placement, 938 uint32_t *masked_placement) 939 { 940 uint32_t cur_flags = ttm_bo_type_flags(mem_type); 941 942 if ((cur_flags & proposed_placement & TTM_PL_MASK_MEM) == 0) 943 return false; 944 945 if ((proposed_placement & man->available_caching) == 0) 946 return false; 947 948 cur_flags |= (proposed_placement & man->available_caching); 949 950 *masked_placement = cur_flags; 951 return true; 952 } 953 954 /** 955 * Creates space for memory region @mem according to its type. 956 * 957 * This function first searches for free space in compatible memory types in 958 * the priority order defined by the driver. If free space isn't found, then 959 * ttm_bo_mem_force_space is attempted in priority order to evict and find 960 * space. 961 */ 962 int ttm_bo_mem_space(struct ttm_buffer_object *bo, 963 struct ttm_placement *placement, 964 struct ttm_mem_reg *mem, 965 bool interruptible, 966 bool no_wait_gpu) 967 { 968 struct ttm_bo_device *bdev = bo->bdev; 969 struct ttm_mem_type_manager *man; 970 uint32_t mem_type = TTM_PL_SYSTEM; 971 uint32_t cur_flags = 0; 972 bool type_found = false; 973 bool type_ok = false; 974 bool has_erestartsys = false; 975 int i, ret; 976 977 mem->mm_node = NULL; 978 for (i = 0; i < placement->num_placement; ++i) { 979 ret = ttm_mem_type_from_flags(placement->placement[i], 980 &mem_type); 981 if (ret) 982 return ret; 983 man = &bdev->man[mem_type]; 984 985 type_ok = ttm_bo_mt_compatible(man, 986 mem_type, 987 placement->placement[i], 988 &cur_flags); 989 990 if (!type_ok) 991 continue; 992 993 cur_flags = ttm_bo_select_caching(man, bo->mem.placement, 994 cur_flags); 995 /* 996 * Use the access and other non-mapping-related flag bits from 997 * the memory placement flags to the current flags 998 */ 999 ttm_flag_masked(&cur_flags, placement->placement[i], 1000 ~TTM_PL_MASK_MEMTYPE); 1001 1002 if (mem_type == TTM_PL_SYSTEM) 1003 break; 1004 1005 if (man->has_type && man->use_type) { 1006 type_found = true; 1007 ret = (*man->func->get_node)(man, bo, placement, mem); 1008 if (unlikely(ret)) 1009 return ret; 1010 } 1011 if (mem->mm_node) 1012 break; 1013 } 1014 1015 if ((type_ok && (mem_type == TTM_PL_SYSTEM)) || mem->mm_node) { 1016 mem->mem_type = mem_type; 1017 mem->placement = cur_flags; 1018 return 0; 1019 } 1020 1021 if (!type_found) 1022 return -EINVAL; 1023 1024 for (i = 0; i < placement->num_busy_placement; ++i) { 1025 ret = ttm_mem_type_from_flags(placement->busy_placement[i], 1026 &mem_type); 1027 if (ret) 1028 return ret; 1029 man = &bdev->man[mem_type]; 1030 if (!man->has_type) 1031 continue; 1032 if (!ttm_bo_mt_compatible(man, 1033 mem_type, 1034 placement->busy_placement[i], 1035 &cur_flags)) 1036 continue; 1037 1038 cur_flags = ttm_bo_select_caching(man, bo->mem.placement, 1039 cur_flags); 1040 /* 1041 * Use the access and other non-mapping-related flag bits from 1042 * the memory placement flags to the current flags 1043 */ 1044 ttm_flag_masked(&cur_flags, placement->busy_placement[i], 1045 ~TTM_PL_MASK_MEMTYPE); 1046 1047 1048 if (mem_type == TTM_PL_SYSTEM) { 1049 mem->mem_type = mem_type; 1050 mem->placement = cur_flags; 1051 mem->mm_node = NULL; 1052 return 0; 1053 } 1054 1055 ret = ttm_bo_mem_force_space(bo, mem_type, placement, mem, 1056 interruptible, no_wait_gpu); 1057 if (ret == 0 && mem->mm_node) { 1058 mem->placement = cur_flags; 1059 return 0; 1060 } 1061 if (ret == -ERESTART) 1062 has_erestartsys = true; 1063 } 1064 ret = (has_erestartsys) ? -ERESTART: -ENOMEM; 1065 return ret; 1066 } 1067 EXPORT_SYMBOL(ttm_bo_mem_space); 1068 1069 int ttm_bo_move_buffer(struct ttm_buffer_object *bo, 1070 struct ttm_placement *placement, 1071 bool interruptible, 1072 bool no_wait_gpu) 1073 { 1074 int ret = 0; 1075 struct ttm_mem_reg mem; 1076 struct ttm_bo_device *bdev = bo->bdev; 1077 1078 BUG_ON(!ttm_bo_is_reserved(bo)); 1079 1080 /* 1081 * FIXME: It's possible to pipeline buffer moves. 1082 * Have the driver move function wait for idle when necessary, 1083 * instead of doing it here. 1084 */ 1085 spin_lock(&bdev->fence_lock); 1086 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu); 1087 spin_unlock(&bdev->fence_lock); 1088 if (ret) 1089 return ret; 1090 mem.num_pages = bo->num_pages; 1091 mem.size = mem.num_pages << PAGE_SHIFT; 1092 mem.page_alignment = bo->mem.page_alignment; 1093 mem.bus.io_reserved_vm = false; 1094 mem.bus.io_reserved_count = 0; 1095 /* 1096 * Determine where to move the buffer. 1097 */ 1098 ret = ttm_bo_mem_space(bo, placement, &mem, 1099 interruptible, no_wait_gpu); 1100 if (ret) 1101 goto out_unlock; 1102 ret = ttm_bo_handle_move_mem(bo, &mem, false, 1103 interruptible, no_wait_gpu); 1104 out_unlock: 1105 if (ret && mem.mm_node) 1106 ttm_bo_mem_put(bo, &mem); 1107 return ret; 1108 } 1109 1110 static bool ttm_bo_mem_compat(struct ttm_placement *placement, 1111 struct ttm_mem_reg *mem, 1112 uint32_t *new_flags) 1113 { 1114 int i; 1115 1116 if (mem->mm_node && placement->lpfn != 0 && 1117 (mem->start < placement->fpfn || 1118 mem->start + mem->num_pages > placement->lpfn)) 1119 return false; 1120 1121 for (i = 0; i < placement->num_placement; i++) { 1122 *new_flags = placement->placement[i]; 1123 if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) && 1124 (*new_flags & mem->placement & TTM_PL_MASK_MEM)) 1125 return true; 1126 } 1127 1128 for (i = 0; i < placement->num_busy_placement; i++) { 1129 *new_flags = placement->busy_placement[i]; 1130 if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) && 1131 (*new_flags & mem->placement & TTM_PL_MASK_MEM)) 1132 return true; 1133 } 1134 1135 return false; 1136 } 1137 1138 int ttm_bo_validate(struct ttm_buffer_object *bo, 1139 struct ttm_placement *placement, 1140 bool interruptible, 1141 bool no_wait_gpu) 1142 { 1143 int ret; 1144 uint32_t new_flags; 1145 1146 BUG_ON(!ttm_bo_is_reserved(bo)); 1147 /* Check that range is valid */ 1148 if (placement->lpfn || placement->fpfn) 1149 if (placement->fpfn > placement->lpfn || 1150 (placement->lpfn - placement->fpfn) < bo->num_pages) 1151 return -EINVAL; 1152 /* 1153 * Check whether we need to move buffer. 1154 */ 1155 if (!ttm_bo_mem_compat(placement, &bo->mem, &new_flags)) { 1156 ret = ttm_bo_move_buffer(bo, placement, interruptible, 1157 no_wait_gpu); 1158 if (ret) 1159 return ret; 1160 } else { 1161 /* 1162 * Use the access and other non-mapping-related flag bits from 1163 * the compatible memory placement flags to the active flags 1164 */ 1165 ttm_flag_masked(&bo->mem.placement, new_flags, 1166 ~TTM_PL_MASK_MEMTYPE); 1167 } 1168 /* 1169 * We might need to add a TTM. 1170 */ 1171 if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) { 1172 ret = ttm_bo_add_ttm(bo, true); 1173 if (ret) 1174 return ret; 1175 } 1176 return 0; 1177 } 1178 EXPORT_SYMBOL(ttm_bo_validate); 1179 1180 int ttm_bo_check_placement(struct ttm_buffer_object *bo, 1181 struct ttm_placement *placement) 1182 { 1183 BUG_ON((placement->fpfn || placement->lpfn) && 1184 (bo->mem.num_pages > (placement->lpfn - placement->fpfn))); 1185 1186 return 0; 1187 } 1188 1189 int ttm_bo_init(struct ttm_bo_device *bdev, 1190 struct ttm_buffer_object *bo, 1191 unsigned long size, 1192 enum ttm_bo_type type, 1193 struct ttm_placement *placement, 1194 uint32_t page_alignment, 1195 bool interruptible, 1196 struct uvm_object *persistent_swap_storage, 1197 size_t acc_size, 1198 struct sg_table *sg, 1199 void (*destroy) (struct ttm_buffer_object *)) 1200 { 1201 int ret = 0; 1202 unsigned long num_pages; 1203 struct ttm_mem_global *mem_glob = bdev->glob->mem_glob; 1204 1205 ret = ttm_mem_global_alloc(mem_glob, acc_size, false, false); 1206 if (ret) { 1207 pr_err("Out of kernel memory\n"); 1208 if (destroy) 1209 (*destroy)(bo); 1210 else 1211 kfree(bo); 1212 return -ENOMEM; 1213 } 1214 1215 num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT; 1216 if (num_pages == 0) { 1217 pr_err("Illegal buffer object size\n"); 1218 if (destroy) 1219 (*destroy)(bo); 1220 else 1221 kfree(bo); 1222 ttm_mem_global_free(mem_glob, acc_size); 1223 return -EINVAL; 1224 } 1225 bo->destroy = destroy; 1226 1227 uvm_objinit(&bo->uobj, NULL, 0); 1228 kref_init(&bo->kref); 1229 kref_init(&bo->list_kref); 1230 atomic_set(&bo->cpu_writers, 0); 1231 atomic_set(&bo->reserved, 1); 1232 init_waitqueue_head(&bo->event_queue); 1233 INIT_LIST_HEAD(&bo->lru); 1234 INIT_LIST_HEAD(&bo->ddestroy); 1235 INIT_LIST_HEAD(&bo->swap); 1236 INIT_LIST_HEAD(&bo->io_reserve_lru); 1237 bo->bdev = bdev; 1238 bo->glob = bdev->glob; 1239 bo->type = type; 1240 bo->num_pages = num_pages; 1241 bo->mem.size = num_pages << PAGE_SHIFT; 1242 bo->mem.mem_type = TTM_PL_SYSTEM; 1243 bo->mem.num_pages = bo->num_pages; 1244 bo->mem.mm_node = NULL; 1245 bo->mem.page_alignment = page_alignment; 1246 bo->mem.bus.io_reserved_vm = false; 1247 bo->mem.bus.io_reserved_count = 0; 1248 bo->priv_flags = 0; 1249 bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED); 1250 bo->seq_valid = false; 1251 bo->persistent_swap_storage = persistent_swap_storage; 1252 bo->acc_size = acc_size; 1253 bo->sg = sg; 1254 atomic_inc(&bo->glob->bo_count); 1255 1256 ret = ttm_bo_check_placement(bo, placement); 1257 if (unlikely(ret != 0)) 1258 goto out_err; 1259 1260 /* 1261 * For ttm_bo_type_device buffers, allocate 1262 * address space from the device. 1263 */ 1264 if (bo->type == ttm_bo_type_device || 1265 bo->type == ttm_bo_type_sg) { 1266 ret = ttm_bo_setup_vm(bo); 1267 if (ret) 1268 goto out_err; 1269 } 1270 1271 ret = ttm_bo_validate(bo, placement, interruptible, false); 1272 if (ret) 1273 goto out_err; 1274 1275 ttm_bo_unreserve(bo); 1276 return 0; 1277 1278 out_err: 1279 ttm_bo_unreserve(bo); 1280 ttm_bo_unref(&bo); 1281 1282 return ret; 1283 } 1284 EXPORT_SYMBOL(ttm_bo_init); 1285 1286 size_t ttm_bo_acc_size(struct ttm_bo_device *bdev, 1287 unsigned long bo_size, 1288 unsigned struct_size) 1289 { 1290 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT; 1291 size_t size = 0; 1292 1293 size += ttm_round_pot(struct_size); 1294 size += PAGE_ALIGN(npages * sizeof(void *)); 1295 size += ttm_round_pot(sizeof(struct ttm_tt)); 1296 return size; 1297 } 1298 EXPORT_SYMBOL(ttm_bo_acc_size); 1299 1300 size_t ttm_bo_dma_acc_size(struct ttm_bo_device *bdev, 1301 unsigned long bo_size, 1302 unsigned struct_size) 1303 { 1304 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT; 1305 size_t size = 0; 1306 1307 size += ttm_round_pot(struct_size); 1308 size += PAGE_ALIGN(npages * sizeof(void *)); 1309 size += PAGE_ALIGN(npages * sizeof(bus_addr_t)); 1310 size += ttm_round_pot(sizeof(struct ttm_dma_tt)); 1311 return size; 1312 } 1313 EXPORT_SYMBOL(ttm_bo_dma_acc_size); 1314 1315 int ttm_bo_create(struct ttm_bo_device *bdev, 1316 unsigned long size, 1317 enum ttm_bo_type type, 1318 struct ttm_placement *placement, 1319 uint32_t page_alignment, 1320 bool interruptible, 1321 struct uvm_object *persistent_swap_storage, 1322 struct ttm_buffer_object **p_bo) 1323 { 1324 struct ttm_buffer_object *bo; 1325 size_t acc_size; 1326 int ret; 1327 1328 bo = kzalloc(sizeof(*bo), GFP_KERNEL); 1329 if (unlikely(bo == NULL)) 1330 return -ENOMEM; 1331 1332 acc_size = ttm_bo_acc_size(bdev, size, sizeof(struct ttm_buffer_object)); 1333 ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment, 1334 interruptible, persistent_swap_storage, acc_size, 1335 NULL, NULL); 1336 if (likely(ret == 0)) 1337 *p_bo = bo; 1338 1339 return ret; 1340 } 1341 EXPORT_SYMBOL(ttm_bo_create); 1342 1343 static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev, 1344 unsigned mem_type, bool allow_errors) 1345 { 1346 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 1347 struct ttm_bo_global *glob = bdev->glob; 1348 int ret; 1349 1350 /* 1351 * Can't use standard list traversal since we're unlocking. 1352 */ 1353 1354 spin_lock(&glob->lru_lock); 1355 while (!list_empty(&man->lru)) { 1356 spin_unlock(&glob->lru_lock); 1357 ret = ttm_mem_evict_first(bdev, mem_type, false, false); 1358 if (ret) { 1359 if (allow_errors) { 1360 return ret; 1361 } else { 1362 pr_err("Cleanup eviction failed\n"); 1363 } 1364 } 1365 spin_lock(&glob->lru_lock); 1366 } 1367 spin_unlock(&glob->lru_lock); 1368 return 0; 1369 } 1370 1371 int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type) 1372 { 1373 struct ttm_mem_type_manager *man; 1374 int ret = -EINVAL; 1375 1376 if (mem_type >= TTM_NUM_MEM_TYPES) { 1377 pr_err("Illegal memory type %d\n", mem_type); 1378 return ret; 1379 } 1380 man = &bdev->man[mem_type]; 1381 1382 if (!man->has_type) { 1383 pr_err("Trying to take down uninitialized memory manager type %u\n", 1384 mem_type); 1385 return ret; 1386 } 1387 1388 man->use_type = false; 1389 man->has_type = false; 1390 1391 ret = 0; 1392 if (mem_type > 0) { 1393 ttm_bo_force_list_clean(bdev, mem_type, false); 1394 1395 ret = (*man->func->takedown)(man); 1396 } 1397 1398 return ret; 1399 } 1400 EXPORT_SYMBOL(ttm_bo_clean_mm); 1401 1402 int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type) 1403 { 1404 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 1405 1406 if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) { 1407 pr_err("Illegal memory manager memory type %u\n", mem_type); 1408 return -EINVAL; 1409 } 1410 1411 if (!man->has_type) { 1412 pr_err("Memory type %u has not been initialized\n", mem_type); 1413 return 0; 1414 } 1415 1416 return ttm_bo_force_list_clean(bdev, mem_type, true); 1417 } 1418 EXPORT_SYMBOL(ttm_bo_evict_mm); 1419 1420 int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type, 1421 unsigned long p_size) 1422 { 1423 int ret = -EINVAL; 1424 struct ttm_mem_type_manager *man; 1425 1426 BUG_ON(type >= TTM_NUM_MEM_TYPES); 1427 man = &bdev->man[type]; 1428 BUG_ON(man->has_type); 1429 man->io_reserve_fastpath = true; 1430 man->use_io_reserve_lru = false; 1431 rw_init(&man->io_reserve_mutex, "ttm_iores"); 1432 INIT_LIST_HEAD(&man->io_reserve_lru); 1433 1434 ret = bdev->driver->init_mem_type(bdev, type, man); 1435 if (ret) 1436 return ret; 1437 man->bdev = bdev; 1438 1439 ret = 0; 1440 if (type != TTM_PL_SYSTEM) { 1441 ret = (*man->func->init)(man, p_size); 1442 if (ret) 1443 return ret; 1444 } 1445 man->has_type = true; 1446 man->use_type = true; 1447 man->size = p_size; 1448 1449 INIT_LIST_HEAD(&man->lru); 1450 1451 return 0; 1452 } 1453 EXPORT_SYMBOL(ttm_bo_init_mm); 1454 1455 static void ttm_bo_global_kobj_release(struct kobject *kobj) 1456 { 1457 struct ttm_bo_global *glob = 1458 container_of(kobj, struct ttm_bo_global, kobj); 1459 1460 ttm_mem_unregister_shrink(glob->mem_glob, &glob->shrink); 1461 km_free(glob->dummy_read_page, PAGE_SIZE, &kv_any, &kp_dma_zero); 1462 kfree(glob); 1463 } 1464 1465 void ttm_bo_global_release(struct drm_global_reference *ref) 1466 { 1467 struct ttm_bo_global *glob = ref->object; 1468 1469 kobject_del(&glob->kobj); 1470 kobject_put(&glob->kobj); 1471 } 1472 EXPORT_SYMBOL(ttm_bo_global_release); 1473 1474 int ttm_bo_global_init(struct drm_global_reference *ref) 1475 { 1476 struct ttm_bo_global_ref *bo_ref = 1477 container_of(ref, struct ttm_bo_global_ref, ref); 1478 struct ttm_bo_global *glob = ref->object; 1479 int ret; 1480 1481 rw_init(&glob->device_list_mutex, "ttm_devlist"); 1482 mtx_init(&glob->lru_lock, IPL_NONE); 1483 glob->mem_glob = bo_ref->mem_glob; 1484 glob->dummy_read_page = km_alloc(PAGE_SIZE, &kv_any, &kp_dma_zero, 1485 &kd_waitok); 1486 1487 if (unlikely(glob->dummy_read_page == NULL)) { 1488 ret = -ENOMEM; 1489 goto out_no_drp; 1490 } 1491 1492 INIT_LIST_HEAD(&glob->swap_lru); 1493 INIT_LIST_HEAD(&glob->device_list); 1494 1495 ttm_mem_init_shrink(&glob->shrink, ttm_bo_swapout); 1496 ret = ttm_mem_register_shrink(glob->mem_glob, &glob->shrink); 1497 if (unlikely(ret != 0)) { 1498 pr_err("Could not register buffer object swapout\n"); 1499 goto out_no_shrink; 1500 } 1501 1502 atomic_set(&glob->bo_count, 0); 1503 1504 ret = kobject_init_and_add( 1505 &glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects"); 1506 if (unlikely(ret != 0)) 1507 kobject_put(&glob->kobj); 1508 return ret; 1509 out_no_shrink: 1510 km_free(glob->dummy_read_page, PAGE_SIZE, &kv_any, &kp_dma_zero); 1511 out_no_drp: 1512 kfree(glob); 1513 return ret; 1514 } 1515 EXPORT_SYMBOL(ttm_bo_global_init); 1516 1517 1518 int ttm_bo_device_release(struct ttm_bo_device *bdev) 1519 { 1520 int ret = 0; 1521 unsigned i = TTM_NUM_MEM_TYPES; 1522 struct ttm_mem_type_manager *man; 1523 struct ttm_bo_global *glob = bdev->glob; 1524 1525 while (i--) { 1526 man = &bdev->man[i]; 1527 if (man->has_type) { 1528 man->use_type = false; 1529 if ((i != TTM_PL_SYSTEM) && ttm_bo_clean_mm(bdev, i)) { 1530 ret = -EBUSY; 1531 pr_err("DRM memory manager type %d is not clean\n", 1532 i); 1533 } 1534 man->has_type = false; 1535 } 1536 } 1537 1538 mutex_lock(&glob->device_list_mutex); 1539 list_del(&bdev->device_list); 1540 mutex_unlock(&glob->device_list_mutex); 1541 1542 cancel_delayed_work_sync(&bdev->wq); 1543 1544 while (ttm_bo_delayed_delete(bdev, true)) 1545 ; 1546 1547 spin_lock(&glob->lru_lock); 1548 if (list_empty(&bdev->ddestroy)) 1549 TTM_DEBUG("Delayed destroy list was clean\n"); 1550 1551 if (list_empty(&bdev->man[0].lru)) 1552 TTM_DEBUG("Swap list was clean\n"); 1553 spin_unlock(&glob->lru_lock); 1554 1555 BUG_ON(!drm_mm_clean(&bdev->addr_space_mm)); 1556 write_lock(&bdev->vm_lock); 1557 drm_mm_takedown(&bdev->addr_space_mm); 1558 write_unlock(&bdev->vm_lock); 1559 1560 return ret; 1561 } 1562 EXPORT_SYMBOL(ttm_bo_device_release); 1563 1564 int ttm_bo_device_init(struct ttm_bo_device *bdev, 1565 struct ttm_bo_global *glob, 1566 struct ttm_bo_driver *driver, 1567 uint64_t file_page_offset, 1568 bool need_dma32) 1569 { 1570 int ret = -EINVAL; 1571 1572 rw_init(&bdev->vm_lock, "ttmvm"); 1573 bdev->driver = driver; 1574 1575 memset(bdev->man, 0, sizeof(bdev->man)); 1576 1577 /* 1578 * Initialize the system memory buffer type. 1579 * Other types need to be driver / IOCTL initialized. 1580 */ 1581 ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0); 1582 if (unlikely(ret != 0)) 1583 goto out_no_sys; 1584 1585 RB_INIT(&bdev->addr_space_rb); 1586 drm_mm_init(&bdev->addr_space_mm, file_page_offset, 0x10000000); 1587 1588 INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue); 1589 INIT_LIST_HEAD(&bdev->ddestroy); 1590 bdev->dev_mapping = NULL; 1591 bdev->glob = glob; 1592 bdev->need_dma32 = need_dma32; 1593 bdev->val_seq = 0; 1594 mtx_init(&bdev->fence_lock, IPL_NONE); 1595 mutex_lock(&glob->device_list_mutex); 1596 list_add_tail(&bdev->device_list, &glob->device_list); 1597 mutex_unlock(&glob->device_list_mutex); 1598 1599 return 0; 1600 out_no_sys: 1601 return ret; 1602 } 1603 EXPORT_SYMBOL(ttm_bo_device_init); 1604 1605 /* 1606 * buffer object vm functions. 1607 */ 1608 1609 bool ttm_mem_reg_is_pci(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem) 1610 { 1611 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type]; 1612 1613 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) { 1614 if (mem->mem_type == TTM_PL_SYSTEM) 1615 return false; 1616 1617 if (man->flags & TTM_MEMTYPE_FLAG_CMA) 1618 return false; 1619 1620 if (mem->placement & TTM_PL_FLAG_CACHED) 1621 return false; 1622 } 1623 return true; 1624 } 1625 1626 void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo) 1627 { 1628 struct ttm_tt *ttm = bo->ttm; 1629 struct vm_page *page; 1630 bus_addr_t addr; 1631 paddr_t paddr; 1632 int i; 1633 1634 if (bo->mem.bus.is_iomem) { 1635 for (i = 0; i < bo->mem.num_pages; ++i) { 1636 addr = bo->mem.bus.base + bo->mem.bus.offset; 1637 paddr = bus_space_mmap(bo->bdev->memt, addr, 1638 i << PAGE_SHIFT, 0, 0); 1639 page = PHYS_TO_VM_PAGE(paddr); 1640 if (unlikely(page == NULL)) 1641 continue; 1642 pmap_page_protect(page, PROT_NONE); 1643 } 1644 } else if (ttm) { 1645 for (i = 0; i < ttm->num_pages; ++i) { 1646 page = ttm->pages[i]; 1647 if (unlikely(page == NULL)) 1648 continue; 1649 pmap_page_protect(page, PROT_NONE); 1650 } 1651 } 1652 ttm_mem_io_free_vm(bo); 1653 } 1654 1655 void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo) 1656 { 1657 struct ttm_bo_device *bdev = bo->bdev; 1658 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type]; 1659 1660 ttm_mem_io_lock(man, false); 1661 ttm_bo_unmap_virtual_locked(bo); 1662 ttm_mem_io_unlock(man); 1663 } 1664 1665 1666 EXPORT_SYMBOL(ttm_bo_unmap_virtual); 1667 1668 static void ttm_bo_vm_insert_rb(struct ttm_buffer_object *bo) 1669 { 1670 struct ttm_bo_device *bdev = bo->bdev; 1671 1672 /* The caller acquired bdev->vm_lock. */ 1673 RB_INSERT(ttm_bo_device_buffer_objects, &bdev->addr_space_rb, bo); 1674 } 1675 1676 /** 1677 * ttm_bo_setup_vm: 1678 * 1679 * @bo: the buffer to allocate address space for 1680 * 1681 * Allocate address space in the drm device so that applications 1682 * can mmap the buffer and access the contents. This only 1683 * applies to ttm_bo_type_device objects as others are not 1684 * placed in the drm device address space. 1685 */ 1686 1687 static int ttm_bo_setup_vm(struct ttm_buffer_object *bo) 1688 { 1689 struct ttm_bo_device *bdev = bo->bdev; 1690 int ret; 1691 1692 retry_pre_get: 1693 ret = drm_mm_pre_get(&bdev->addr_space_mm); 1694 if (unlikely(ret != 0)) 1695 return ret; 1696 1697 write_lock(&bdev->vm_lock); 1698 bo->vm_node = drm_mm_search_free(&bdev->addr_space_mm, 1699 bo->mem.num_pages, 0, 0); 1700 1701 if (unlikely(bo->vm_node == NULL)) { 1702 ret = -ENOMEM; 1703 goto out_unlock; 1704 } 1705 1706 bo->vm_node = drm_mm_get_block_atomic(bo->vm_node, 1707 bo->mem.num_pages, 0); 1708 1709 if (unlikely(bo->vm_node == NULL)) { 1710 write_unlock(&bdev->vm_lock); 1711 goto retry_pre_get; 1712 } 1713 1714 ttm_bo_vm_insert_rb(bo); 1715 write_unlock(&bdev->vm_lock); 1716 bo->addr_space_offset = ((uint64_t) bo->vm_node->start) << PAGE_SHIFT; 1717 1718 return 0; 1719 out_unlock: 1720 write_unlock(&bdev->vm_lock); 1721 return ret; 1722 } 1723 1724 int ttm_bo_wait(struct ttm_buffer_object *bo, 1725 bool lazy, bool interruptible, bool no_wait) 1726 { 1727 struct ttm_bo_driver *driver = bo->bdev->driver; 1728 struct ttm_bo_device *bdev = bo->bdev; 1729 void *sync_obj; 1730 int ret = 0; 1731 1732 if (likely(bo->sync_obj == NULL)) 1733 return 0; 1734 1735 while (bo->sync_obj) { 1736 1737 if (driver->sync_obj_signaled(bo->sync_obj)) { 1738 void *tmp_obj = bo->sync_obj; 1739 bo->sync_obj = NULL; 1740 clear_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags); 1741 spin_unlock(&bdev->fence_lock); 1742 driver->sync_obj_unref(&tmp_obj); 1743 spin_lock(&bdev->fence_lock); 1744 continue; 1745 } 1746 1747 if (no_wait) 1748 return -EBUSY; 1749 1750 sync_obj = driver->sync_obj_ref(bo->sync_obj); 1751 spin_unlock(&bdev->fence_lock); 1752 ret = driver->sync_obj_wait(sync_obj, 1753 lazy, interruptible); 1754 if (unlikely(ret != 0)) { 1755 driver->sync_obj_unref(&sync_obj); 1756 spin_lock(&bdev->fence_lock); 1757 return ret; 1758 } 1759 spin_lock(&bdev->fence_lock); 1760 if (likely(bo->sync_obj == sync_obj)) { 1761 void *tmp_obj = bo->sync_obj; 1762 bo->sync_obj = NULL; 1763 clear_bit(TTM_BO_PRIV_FLAG_MOVING, 1764 &bo->priv_flags); 1765 spin_unlock(&bdev->fence_lock); 1766 driver->sync_obj_unref(&sync_obj); 1767 driver->sync_obj_unref(&tmp_obj); 1768 spin_lock(&bdev->fence_lock); 1769 } else { 1770 spin_unlock(&bdev->fence_lock); 1771 driver->sync_obj_unref(&sync_obj); 1772 spin_lock(&bdev->fence_lock); 1773 } 1774 } 1775 return 0; 1776 } 1777 EXPORT_SYMBOL(ttm_bo_wait); 1778 1779 int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait) 1780 { 1781 struct ttm_bo_device *bdev = bo->bdev; 1782 int ret = 0; 1783 1784 /* 1785 * Using ttm_bo_reserve makes sure the lru lists are updated. 1786 */ 1787 1788 ret = ttm_bo_reserve(bo, true, no_wait, false, 0); 1789 if (unlikely(ret != 0)) 1790 return ret; 1791 spin_lock(&bdev->fence_lock); 1792 ret = ttm_bo_wait(bo, false, true, no_wait); 1793 spin_unlock(&bdev->fence_lock); 1794 if (likely(ret == 0)) 1795 atomic_inc(&bo->cpu_writers); 1796 ttm_bo_unreserve(bo); 1797 return ret; 1798 } 1799 EXPORT_SYMBOL(ttm_bo_synccpu_write_grab); 1800 1801 void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo) 1802 { 1803 atomic_dec(&bo->cpu_writers); 1804 } 1805 EXPORT_SYMBOL(ttm_bo_synccpu_write_release); 1806 1807 /** 1808 * A buffer object shrink method that tries to swap out the first 1809 * buffer object on the bo_global::swap_lru list. 1810 */ 1811 1812 static int ttm_bo_swapout(struct ttm_mem_shrink *shrink) 1813 { 1814 struct ttm_bo_global *glob = 1815 container_of(shrink, struct ttm_bo_global, shrink); 1816 struct ttm_buffer_object *bo; 1817 int ret = -EBUSY; 1818 int put_count; 1819 uint32_t swap_placement = (TTM_PL_FLAG_CACHED | TTM_PL_FLAG_SYSTEM); 1820 1821 spin_lock(&glob->lru_lock); 1822 list_for_each_entry(bo, &glob->swap_lru, swap) { 1823 ret = ttm_bo_reserve_locked(bo, false, true, false, 0); 1824 if (!ret) 1825 break; 1826 } 1827 1828 if (ret) { 1829 spin_unlock(&glob->lru_lock); 1830 return ret; 1831 } 1832 1833 kref_get(&bo->list_kref); 1834 1835 if (!list_empty(&bo->ddestroy)) { 1836 ret = ttm_bo_cleanup_refs_and_unlock(bo, false, false); 1837 kref_put(&bo->list_kref, ttm_bo_release_list); 1838 return ret; 1839 } 1840 1841 put_count = ttm_bo_del_from_lru(bo); 1842 spin_unlock(&glob->lru_lock); 1843 1844 ttm_bo_list_ref_sub(bo, put_count, true); 1845 1846 /** 1847 * Wait for GPU, then move to system cached. 1848 */ 1849 1850 spin_lock(&bo->bdev->fence_lock); 1851 ret = ttm_bo_wait(bo, false, false, false); 1852 spin_unlock(&bo->bdev->fence_lock); 1853 1854 if (unlikely(ret != 0)) 1855 goto out; 1856 1857 if ((bo->mem.placement & swap_placement) != swap_placement) { 1858 struct ttm_mem_reg evict_mem; 1859 1860 evict_mem = bo->mem; 1861 evict_mem.mm_node = NULL; 1862 evict_mem.placement = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED; 1863 evict_mem.mem_type = TTM_PL_SYSTEM; 1864 1865 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, 1866 false, false); 1867 if (unlikely(ret != 0)) 1868 goto out; 1869 } 1870 1871 ttm_bo_unmap_virtual(bo); 1872 1873 /** 1874 * Swap out. Buffer will be swapped in again as soon as 1875 * anyone tries to access a ttm page. 1876 */ 1877 1878 if (bo->bdev->driver->swap_notify) 1879 bo->bdev->driver->swap_notify(bo); 1880 1881 ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage); 1882 out: 1883 1884 /** 1885 * 1886 * Unreserve without putting on LRU to avoid swapping out an 1887 * already swapped buffer. 1888 */ 1889 1890 atomic_set(&bo->reserved, 0); 1891 wake_up_all(&bo->event_queue); 1892 kref_put(&bo->list_kref, ttm_bo_release_list); 1893 return ret; 1894 } 1895 1896 void ttm_bo_swapout_all(struct ttm_bo_device *bdev) 1897 { 1898 while (ttm_bo_swapout(&bdev->glob->shrink) == 0) 1899 ; 1900 } 1901 EXPORT_SYMBOL(ttm_bo_swapout_all); 1902