1 /* $OpenBSD: ttm_bo.c,v 1.20 2017/06/04 14:02:24 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 drm_vma_offset_remove(&bdev->vma_manager, &bo->vma_node); 732 ttm_mem_io_lock(man, false); 733 ttm_mem_io_free_vm(bo); 734 ttm_mem_io_unlock(man); 735 ttm_bo_cleanup_refs_or_queue(bo); 736 kref_put(&bo->list_kref, ttm_bo_release_list); 737 } 738 739 void ttm_bo_unref(struct ttm_buffer_object **p_bo) 740 { 741 struct ttm_buffer_object *bo = *p_bo; 742 743 *p_bo = NULL; 744 kref_put(&bo->kref, ttm_bo_release); 745 } 746 EXPORT_SYMBOL(ttm_bo_unref); 747 748 int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev) 749 { 750 return cancel_delayed_work_sync(&bdev->wq); 751 } 752 EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue); 753 754 void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched) 755 { 756 if (resched) 757 schedule_delayed_work(&bdev->wq, 758 ((HZ / 100) < 1) ? 1 : HZ / 100); 759 } 760 EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue); 761 762 static int ttm_bo_evict(struct ttm_buffer_object *bo, bool interruptible, 763 bool no_wait_gpu) 764 { 765 struct ttm_bo_device *bdev = bo->bdev; 766 struct ttm_mem_reg evict_mem; 767 struct ttm_placement placement; 768 int ret = 0; 769 770 spin_lock(&bdev->fence_lock); 771 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu); 772 spin_unlock(&bdev->fence_lock); 773 774 if (unlikely(ret != 0)) { 775 if (ret != -ERESTARTSYS) { 776 pr_err("Failed to expire sync object before buffer eviction\n"); 777 } 778 goto out; 779 } 780 781 BUG_ON(!ttm_bo_is_reserved(bo)); 782 783 evict_mem = bo->mem; 784 evict_mem.mm_node = NULL; 785 evict_mem.bus.io_reserved_vm = false; 786 evict_mem.bus.io_reserved_count = 0; 787 788 placement.fpfn = 0; 789 placement.lpfn = 0; 790 placement.num_placement = 0; 791 placement.num_busy_placement = 0; 792 bdev->driver->evict_flags(bo, &placement); 793 ret = ttm_bo_mem_space(bo, &placement, &evict_mem, interruptible, 794 no_wait_gpu); 795 if (ret) { 796 if (ret != -ERESTARTSYS) { 797 pr_err("Failed to find memory space for buffer 0x%p eviction\n", 798 bo); 799 ttm_bo_mem_space_debug(bo, &placement); 800 } 801 goto out; 802 } 803 804 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, interruptible, 805 no_wait_gpu); 806 if (ret) { 807 if (ret != -ERESTARTSYS) 808 pr_err("Buffer eviction failed\n"); 809 ttm_bo_mem_put(bo, &evict_mem); 810 goto out; 811 } 812 bo->evicted = true; 813 out: 814 return ret; 815 } 816 817 static int ttm_mem_evict_first(struct ttm_bo_device *bdev, 818 uint32_t mem_type, 819 bool interruptible, 820 bool no_wait_gpu) 821 { 822 struct ttm_bo_global *glob = bdev->glob; 823 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 824 struct ttm_buffer_object *bo; 825 int ret = -EBUSY, put_count; 826 827 spin_lock(&glob->lru_lock); 828 list_for_each_entry(bo, &man->lru, lru) { 829 ret = ttm_bo_reserve_locked(bo, false, true, false, 0); 830 if (!ret) 831 break; 832 } 833 834 if (ret) { 835 spin_unlock(&glob->lru_lock); 836 return ret; 837 } 838 839 kref_get(&bo->list_kref); 840 841 if (!list_empty(&bo->ddestroy)) { 842 ret = ttm_bo_cleanup_refs_and_unlock(bo, interruptible, 843 no_wait_gpu); 844 kref_put(&bo->list_kref, ttm_bo_release_list); 845 return ret; 846 } 847 848 put_count = ttm_bo_del_from_lru(bo); 849 spin_unlock(&glob->lru_lock); 850 851 BUG_ON(ret != 0); 852 853 ttm_bo_list_ref_sub(bo, put_count, true); 854 855 ret = ttm_bo_evict(bo, interruptible, no_wait_gpu); 856 ttm_bo_unreserve(bo); 857 858 kref_put(&bo->list_kref, ttm_bo_release_list); 859 return ret; 860 } 861 862 void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem) 863 { 864 struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type]; 865 866 if (mem->mm_node) 867 (*man->func->put_node)(man, mem); 868 } 869 EXPORT_SYMBOL(ttm_bo_mem_put); 870 871 /** 872 * Repeatedly evict memory from the LRU for @mem_type until we create enough 873 * space, or we've evicted everything and there isn't enough space. 874 */ 875 static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo, 876 uint32_t mem_type, 877 struct ttm_placement *placement, 878 struct ttm_mem_reg *mem, 879 bool interruptible, 880 bool no_wait_gpu) 881 { 882 struct ttm_bo_device *bdev = bo->bdev; 883 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 884 int ret; 885 886 do { 887 ret = (*man->func->get_node)(man, bo, placement, mem); 888 if (unlikely(ret != 0)) 889 return ret; 890 if (mem->mm_node) 891 break; 892 ret = ttm_mem_evict_first(bdev, mem_type, 893 interruptible, no_wait_gpu); 894 if (unlikely(ret != 0)) 895 return ret; 896 } while (1); 897 if (mem->mm_node == NULL) 898 return -ENOMEM; 899 mem->mem_type = mem_type; 900 return 0; 901 } 902 903 static uint32_t ttm_bo_select_caching(struct ttm_mem_type_manager *man, 904 uint32_t cur_placement, 905 uint32_t proposed_placement) 906 { 907 uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING; 908 uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING; 909 910 /** 911 * Keep current caching if possible. 912 */ 913 914 if ((cur_placement & caching) != 0) 915 result |= (cur_placement & caching); 916 else if ((man->default_caching & caching) != 0) 917 result |= man->default_caching; 918 else if ((TTM_PL_FLAG_CACHED & caching) != 0) 919 result |= TTM_PL_FLAG_CACHED; 920 else if ((TTM_PL_FLAG_WC & caching) != 0) 921 result |= TTM_PL_FLAG_WC; 922 else if ((TTM_PL_FLAG_UNCACHED & caching) != 0) 923 result |= TTM_PL_FLAG_UNCACHED; 924 925 return result; 926 } 927 928 static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man, 929 uint32_t mem_type, 930 uint32_t proposed_placement, 931 uint32_t *masked_placement) 932 { 933 uint32_t cur_flags = ttm_bo_type_flags(mem_type); 934 935 if ((cur_flags & proposed_placement & TTM_PL_MASK_MEM) == 0) 936 return false; 937 938 if ((proposed_placement & man->available_caching) == 0) 939 return false; 940 941 cur_flags |= (proposed_placement & man->available_caching); 942 943 *masked_placement = cur_flags; 944 return true; 945 } 946 947 /** 948 * Creates space for memory region @mem according to its type. 949 * 950 * This function first searches for free space in compatible memory types in 951 * the priority order defined by the driver. If free space isn't found, then 952 * ttm_bo_mem_force_space is attempted in priority order to evict and find 953 * space. 954 */ 955 int ttm_bo_mem_space(struct ttm_buffer_object *bo, 956 struct ttm_placement *placement, 957 struct ttm_mem_reg *mem, 958 bool interruptible, 959 bool no_wait_gpu) 960 { 961 struct ttm_bo_device *bdev = bo->bdev; 962 struct ttm_mem_type_manager *man; 963 uint32_t mem_type = TTM_PL_SYSTEM; 964 uint32_t cur_flags = 0; 965 bool type_found = false; 966 bool type_ok = false; 967 bool has_erestartsys = false; 968 int i, ret; 969 970 mem->mm_node = NULL; 971 for (i = 0; i < placement->num_placement; ++i) { 972 ret = ttm_mem_type_from_flags(placement->placement[i], 973 &mem_type); 974 if (ret) 975 return ret; 976 man = &bdev->man[mem_type]; 977 978 type_ok = ttm_bo_mt_compatible(man, 979 mem_type, 980 placement->placement[i], 981 &cur_flags); 982 983 if (!type_ok) 984 continue; 985 986 cur_flags = ttm_bo_select_caching(man, bo->mem.placement, 987 cur_flags); 988 /* 989 * Use the access and other non-mapping-related flag bits from 990 * the memory placement flags to the current flags 991 */ 992 ttm_flag_masked(&cur_flags, placement->placement[i], 993 ~TTM_PL_MASK_MEMTYPE); 994 995 if (mem_type == TTM_PL_SYSTEM) 996 break; 997 998 if (man->has_type && man->use_type) { 999 type_found = true; 1000 ret = (*man->func->get_node)(man, bo, placement, mem); 1001 if (unlikely(ret)) 1002 return ret; 1003 } 1004 if (mem->mm_node) 1005 break; 1006 } 1007 1008 if ((type_ok && (mem_type == TTM_PL_SYSTEM)) || mem->mm_node) { 1009 mem->mem_type = mem_type; 1010 mem->placement = cur_flags; 1011 return 0; 1012 } 1013 1014 if (!type_found) 1015 return -EINVAL; 1016 1017 for (i = 0; i < placement->num_busy_placement; ++i) { 1018 ret = ttm_mem_type_from_flags(placement->busy_placement[i], 1019 &mem_type); 1020 if (ret) 1021 return ret; 1022 man = &bdev->man[mem_type]; 1023 if (!man->has_type) 1024 continue; 1025 if (!ttm_bo_mt_compatible(man, 1026 mem_type, 1027 placement->busy_placement[i], 1028 &cur_flags)) 1029 continue; 1030 1031 cur_flags = ttm_bo_select_caching(man, bo->mem.placement, 1032 cur_flags); 1033 /* 1034 * Use the access and other non-mapping-related flag bits from 1035 * the memory placement flags to the current flags 1036 */ 1037 ttm_flag_masked(&cur_flags, placement->busy_placement[i], 1038 ~TTM_PL_MASK_MEMTYPE); 1039 1040 1041 if (mem_type == TTM_PL_SYSTEM) { 1042 mem->mem_type = mem_type; 1043 mem->placement = cur_flags; 1044 mem->mm_node = NULL; 1045 return 0; 1046 } 1047 1048 ret = ttm_bo_mem_force_space(bo, mem_type, placement, mem, 1049 interruptible, no_wait_gpu); 1050 if (ret == 0 && mem->mm_node) { 1051 mem->placement = cur_flags; 1052 return 0; 1053 } 1054 if (ret == -ERESTART) 1055 has_erestartsys = true; 1056 } 1057 ret = (has_erestartsys) ? -ERESTART: -ENOMEM; 1058 return ret; 1059 } 1060 EXPORT_SYMBOL(ttm_bo_mem_space); 1061 1062 int ttm_bo_move_buffer(struct ttm_buffer_object *bo, 1063 struct ttm_placement *placement, 1064 bool interruptible, 1065 bool no_wait_gpu) 1066 { 1067 int ret = 0; 1068 struct ttm_mem_reg mem; 1069 struct ttm_bo_device *bdev = bo->bdev; 1070 1071 BUG_ON(!ttm_bo_is_reserved(bo)); 1072 1073 /* 1074 * FIXME: It's possible to pipeline buffer moves. 1075 * Have the driver move function wait for idle when necessary, 1076 * instead of doing it here. 1077 */ 1078 spin_lock(&bdev->fence_lock); 1079 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu); 1080 spin_unlock(&bdev->fence_lock); 1081 if (ret) 1082 return ret; 1083 mem.num_pages = bo->num_pages; 1084 mem.size = mem.num_pages << PAGE_SHIFT; 1085 mem.page_alignment = bo->mem.page_alignment; 1086 mem.bus.io_reserved_vm = false; 1087 mem.bus.io_reserved_count = 0; 1088 /* 1089 * Determine where to move the buffer. 1090 */ 1091 ret = ttm_bo_mem_space(bo, placement, &mem, 1092 interruptible, no_wait_gpu); 1093 if (ret) 1094 goto out_unlock; 1095 ret = ttm_bo_handle_move_mem(bo, &mem, false, 1096 interruptible, no_wait_gpu); 1097 out_unlock: 1098 if (ret && mem.mm_node) 1099 ttm_bo_mem_put(bo, &mem); 1100 return ret; 1101 } 1102 1103 static bool ttm_bo_mem_compat(struct ttm_placement *placement, 1104 struct ttm_mem_reg *mem, 1105 uint32_t *new_flags) 1106 { 1107 int i; 1108 1109 if (mem->mm_node && placement->lpfn != 0 && 1110 (mem->start < placement->fpfn || 1111 mem->start + mem->num_pages > placement->lpfn)) 1112 return false; 1113 1114 for (i = 0; i < placement->num_placement; i++) { 1115 *new_flags = placement->placement[i]; 1116 if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) && 1117 (*new_flags & mem->placement & TTM_PL_MASK_MEM)) 1118 return true; 1119 } 1120 1121 for (i = 0; i < placement->num_busy_placement; i++) { 1122 *new_flags = placement->busy_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 return false; 1129 } 1130 1131 int ttm_bo_validate(struct ttm_buffer_object *bo, 1132 struct ttm_placement *placement, 1133 bool interruptible, 1134 bool no_wait_gpu) 1135 { 1136 int ret; 1137 uint32_t new_flags; 1138 1139 BUG_ON(!ttm_bo_is_reserved(bo)); 1140 /* Check that range is valid */ 1141 if (placement->lpfn || placement->fpfn) 1142 if (placement->fpfn > placement->lpfn || 1143 (placement->lpfn - placement->fpfn) < bo->num_pages) 1144 return -EINVAL; 1145 /* 1146 * Check whether we need to move buffer. 1147 */ 1148 if (!ttm_bo_mem_compat(placement, &bo->mem, &new_flags)) { 1149 ret = ttm_bo_move_buffer(bo, placement, interruptible, 1150 no_wait_gpu); 1151 if (ret) 1152 return ret; 1153 } else { 1154 /* 1155 * Use the access and other non-mapping-related flag bits from 1156 * the compatible memory placement flags to the active flags 1157 */ 1158 ttm_flag_masked(&bo->mem.placement, new_flags, 1159 ~TTM_PL_MASK_MEMTYPE); 1160 } 1161 /* 1162 * We might need to add a TTM. 1163 */ 1164 if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) { 1165 ret = ttm_bo_add_ttm(bo, true); 1166 if (ret) 1167 return ret; 1168 } 1169 return 0; 1170 } 1171 EXPORT_SYMBOL(ttm_bo_validate); 1172 1173 int ttm_bo_check_placement(struct ttm_buffer_object *bo, 1174 struct ttm_placement *placement) 1175 { 1176 BUG_ON((placement->fpfn || placement->lpfn) && 1177 (bo->mem.num_pages > (placement->lpfn - placement->fpfn))); 1178 1179 return 0; 1180 } 1181 1182 int ttm_bo_init(struct ttm_bo_device *bdev, 1183 struct ttm_buffer_object *bo, 1184 unsigned long size, 1185 enum ttm_bo_type type, 1186 struct ttm_placement *placement, 1187 uint32_t page_alignment, 1188 bool interruptible, 1189 struct uvm_object *persistent_swap_storage, 1190 size_t acc_size, 1191 struct sg_table *sg, 1192 void (*destroy) (struct ttm_buffer_object *)) 1193 { 1194 int ret = 0; 1195 unsigned long num_pages; 1196 struct ttm_mem_global *mem_glob = bdev->glob->mem_glob; 1197 1198 ret = ttm_mem_global_alloc(mem_glob, acc_size, false, false); 1199 if (ret) { 1200 pr_err("Out of kernel memory\n"); 1201 if (destroy) 1202 (*destroy)(bo); 1203 else 1204 kfree(bo); 1205 return -ENOMEM; 1206 } 1207 1208 num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT; 1209 if (num_pages == 0) { 1210 pr_err("Illegal buffer object size\n"); 1211 if (destroy) 1212 (*destroy)(bo); 1213 else 1214 kfree(bo); 1215 ttm_mem_global_free(mem_glob, acc_size); 1216 return -EINVAL; 1217 } 1218 bo->destroy = destroy; 1219 1220 uvm_objinit(&bo->uobj, NULL, 0); 1221 kref_init(&bo->kref); 1222 kref_init(&bo->list_kref); 1223 atomic_set(&bo->cpu_writers, 0); 1224 atomic_set(&bo->reserved, 1); 1225 init_waitqueue_head(&bo->event_queue); 1226 INIT_LIST_HEAD(&bo->lru); 1227 INIT_LIST_HEAD(&bo->ddestroy); 1228 INIT_LIST_HEAD(&bo->swap); 1229 INIT_LIST_HEAD(&bo->io_reserve_lru); 1230 bo->bdev = bdev; 1231 bo->glob = bdev->glob; 1232 bo->type = type; 1233 bo->num_pages = num_pages; 1234 bo->mem.size = num_pages << PAGE_SHIFT; 1235 bo->mem.mem_type = TTM_PL_SYSTEM; 1236 bo->mem.num_pages = bo->num_pages; 1237 bo->mem.mm_node = NULL; 1238 bo->mem.page_alignment = page_alignment; 1239 bo->mem.bus.io_reserved_vm = false; 1240 bo->mem.bus.io_reserved_count = 0; 1241 bo->priv_flags = 0; 1242 bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED); 1243 bo->seq_valid = false; 1244 bo->persistent_swap_storage = persistent_swap_storage; 1245 bo->acc_size = acc_size; 1246 bo->sg = sg; 1247 atomic_inc(&bo->glob->bo_count); 1248 drm_vma_node_reset(&bo->vma_node); 1249 1250 ret = ttm_bo_check_placement(bo, placement); 1251 if (unlikely(ret != 0)) 1252 goto out_err; 1253 1254 /* 1255 * For ttm_bo_type_device buffers, allocate 1256 * address space from the device. 1257 */ 1258 if (bo->type == ttm_bo_type_device || 1259 bo->type == ttm_bo_type_sg) { 1260 ret = ttm_bo_setup_vm(bo); 1261 if (ret) 1262 goto out_err; 1263 } 1264 1265 ret = ttm_bo_validate(bo, placement, interruptible, false); 1266 if (ret) 1267 goto out_err; 1268 1269 ttm_bo_unreserve(bo); 1270 return 0; 1271 1272 out_err: 1273 ttm_bo_unreserve(bo); 1274 ttm_bo_unref(&bo); 1275 1276 return ret; 1277 } 1278 EXPORT_SYMBOL(ttm_bo_init); 1279 1280 size_t ttm_bo_acc_size(struct ttm_bo_device *bdev, 1281 unsigned long bo_size, 1282 unsigned struct_size) 1283 { 1284 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT; 1285 size_t size = 0; 1286 1287 size += ttm_round_pot(struct_size); 1288 size += PAGE_ALIGN(npages * sizeof(void *)); 1289 size += ttm_round_pot(sizeof(struct ttm_tt)); 1290 return size; 1291 } 1292 EXPORT_SYMBOL(ttm_bo_acc_size); 1293 1294 size_t ttm_bo_dma_acc_size(struct ttm_bo_device *bdev, 1295 unsigned long bo_size, 1296 unsigned struct_size) 1297 { 1298 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT; 1299 size_t size = 0; 1300 1301 size += ttm_round_pot(struct_size); 1302 size += PAGE_ALIGN(npages * sizeof(void *)); 1303 size += PAGE_ALIGN(npages * sizeof(bus_addr_t)); 1304 size += ttm_round_pot(sizeof(struct ttm_dma_tt)); 1305 return size; 1306 } 1307 EXPORT_SYMBOL(ttm_bo_dma_acc_size); 1308 1309 int ttm_bo_create(struct ttm_bo_device *bdev, 1310 unsigned long size, 1311 enum ttm_bo_type type, 1312 struct ttm_placement *placement, 1313 uint32_t page_alignment, 1314 bool interruptible, 1315 struct uvm_object *persistent_swap_storage, 1316 struct ttm_buffer_object **p_bo) 1317 { 1318 struct ttm_buffer_object *bo; 1319 size_t acc_size; 1320 int ret; 1321 1322 bo = kzalloc(sizeof(*bo), GFP_KERNEL); 1323 if (unlikely(bo == NULL)) 1324 return -ENOMEM; 1325 1326 acc_size = ttm_bo_acc_size(bdev, size, sizeof(struct ttm_buffer_object)); 1327 ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment, 1328 interruptible, persistent_swap_storage, acc_size, 1329 NULL, NULL); 1330 if (likely(ret == 0)) 1331 *p_bo = bo; 1332 1333 return ret; 1334 } 1335 EXPORT_SYMBOL(ttm_bo_create); 1336 1337 static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev, 1338 unsigned mem_type, bool allow_errors) 1339 { 1340 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 1341 struct ttm_bo_global *glob = bdev->glob; 1342 int ret; 1343 1344 /* 1345 * Can't use standard list traversal since we're unlocking. 1346 */ 1347 1348 spin_lock(&glob->lru_lock); 1349 while (!list_empty(&man->lru)) { 1350 spin_unlock(&glob->lru_lock); 1351 ret = ttm_mem_evict_first(bdev, mem_type, false, false); 1352 if (ret) { 1353 if (allow_errors) { 1354 return ret; 1355 } else { 1356 pr_err("Cleanup eviction failed\n"); 1357 } 1358 } 1359 spin_lock(&glob->lru_lock); 1360 } 1361 spin_unlock(&glob->lru_lock); 1362 return 0; 1363 } 1364 1365 int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type) 1366 { 1367 struct ttm_mem_type_manager *man; 1368 int ret = -EINVAL; 1369 1370 if (mem_type >= TTM_NUM_MEM_TYPES) { 1371 pr_err("Illegal memory type %d\n", mem_type); 1372 return ret; 1373 } 1374 man = &bdev->man[mem_type]; 1375 1376 if (!man->has_type) { 1377 pr_err("Trying to take down uninitialized memory manager type %u\n", 1378 mem_type); 1379 return ret; 1380 } 1381 1382 man->use_type = false; 1383 man->has_type = false; 1384 1385 ret = 0; 1386 if (mem_type > 0) { 1387 ttm_bo_force_list_clean(bdev, mem_type, false); 1388 1389 ret = (*man->func->takedown)(man); 1390 } 1391 1392 return ret; 1393 } 1394 EXPORT_SYMBOL(ttm_bo_clean_mm); 1395 1396 int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type) 1397 { 1398 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 1399 1400 if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) { 1401 pr_err("Illegal memory manager memory type %u\n", mem_type); 1402 return -EINVAL; 1403 } 1404 1405 if (!man->has_type) { 1406 pr_err("Memory type %u has not been initialized\n", mem_type); 1407 return 0; 1408 } 1409 1410 return ttm_bo_force_list_clean(bdev, mem_type, true); 1411 } 1412 EXPORT_SYMBOL(ttm_bo_evict_mm); 1413 1414 int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type, 1415 unsigned long p_size) 1416 { 1417 int ret = -EINVAL; 1418 struct ttm_mem_type_manager *man; 1419 1420 BUG_ON(type >= TTM_NUM_MEM_TYPES); 1421 man = &bdev->man[type]; 1422 BUG_ON(man->has_type); 1423 man->io_reserve_fastpath = true; 1424 man->use_io_reserve_lru = false; 1425 rw_init(&man->io_reserve_mutex, "ttm_iores"); 1426 INIT_LIST_HEAD(&man->io_reserve_lru); 1427 1428 ret = bdev->driver->init_mem_type(bdev, type, man); 1429 if (ret) 1430 return ret; 1431 man->bdev = bdev; 1432 1433 ret = 0; 1434 if (type != TTM_PL_SYSTEM) { 1435 ret = (*man->func->init)(man, p_size); 1436 if (ret) 1437 return ret; 1438 } 1439 man->has_type = true; 1440 man->use_type = true; 1441 man->size = p_size; 1442 1443 INIT_LIST_HEAD(&man->lru); 1444 1445 return 0; 1446 } 1447 EXPORT_SYMBOL(ttm_bo_init_mm); 1448 1449 static void ttm_bo_global_kobj_release(struct kobject *kobj) 1450 { 1451 struct ttm_bo_global *glob = 1452 container_of(kobj, struct ttm_bo_global, kobj); 1453 1454 ttm_mem_unregister_shrink(glob->mem_glob, &glob->shrink); 1455 km_free(glob->dummy_read_page, PAGE_SIZE, &kv_any, &kp_dma_zero); 1456 kfree(glob); 1457 } 1458 1459 void ttm_bo_global_release(struct drm_global_reference *ref) 1460 { 1461 struct ttm_bo_global *glob = ref->object; 1462 1463 kobject_del(&glob->kobj); 1464 kobject_put(&glob->kobj); 1465 } 1466 EXPORT_SYMBOL(ttm_bo_global_release); 1467 1468 int ttm_bo_global_init(struct drm_global_reference *ref) 1469 { 1470 struct ttm_bo_global_ref *bo_ref = 1471 container_of(ref, struct ttm_bo_global_ref, ref); 1472 struct ttm_bo_global *glob = ref->object; 1473 int ret; 1474 1475 rw_init(&glob->device_list_mutex, "ttm_devlist"); 1476 mtx_init(&glob->lru_lock, IPL_NONE); 1477 glob->mem_glob = bo_ref->mem_glob; 1478 glob->dummy_read_page = km_alloc(PAGE_SIZE, &kv_any, &kp_dma_zero, 1479 &kd_waitok); 1480 1481 if (unlikely(glob->dummy_read_page == NULL)) { 1482 ret = -ENOMEM; 1483 goto out_no_drp; 1484 } 1485 1486 INIT_LIST_HEAD(&glob->swap_lru); 1487 INIT_LIST_HEAD(&glob->device_list); 1488 1489 ttm_mem_init_shrink(&glob->shrink, ttm_bo_swapout); 1490 ret = ttm_mem_register_shrink(glob->mem_glob, &glob->shrink); 1491 if (unlikely(ret != 0)) { 1492 pr_err("Could not register buffer object swapout\n"); 1493 goto out_no_shrink; 1494 } 1495 1496 atomic_set(&glob->bo_count, 0); 1497 1498 ret = kobject_init_and_add( 1499 &glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects"); 1500 if (unlikely(ret != 0)) 1501 kobject_put(&glob->kobj); 1502 return ret; 1503 out_no_shrink: 1504 km_free(glob->dummy_read_page, PAGE_SIZE, &kv_any, &kp_dma_zero); 1505 out_no_drp: 1506 kfree(glob); 1507 return ret; 1508 } 1509 EXPORT_SYMBOL(ttm_bo_global_init); 1510 1511 1512 int ttm_bo_device_release(struct ttm_bo_device *bdev) 1513 { 1514 int ret = 0; 1515 unsigned i = TTM_NUM_MEM_TYPES; 1516 struct ttm_mem_type_manager *man; 1517 struct ttm_bo_global *glob = bdev->glob; 1518 1519 while (i--) { 1520 man = &bdev->man[i]; 1521 if (man->has_type) { 1522 man->use_type = false; 1523 if ((i != TTM_PL_SYSTEM) && ttm_bo_clean_mm(bdev, i)) { 1524 ret = -EBUSY; 1525 pr_err("DRM memory manager type %d is not clean\n", 1526 i); 1527 } 1528 man->has_type = false; 1529 } 1530 } 1531 1532 mutex_lock(&glob->device_list_mutex); 1533 list_del(&bdev->device_list); 1534 mutex_unlock(&glob->device_list_mutex); 1535 1536 cancel_delayed_work_sync(&bdev->wq); 1537 1538 while (ttm_bo_delayed_delete(bdev, true)) 1539 ; 1540 1541 spin_lock(&glob->lru_lock); 1542 if (list_empty(&bdev->ddestroy)) 1543 TTM_DEBUG("Delayed destroy list was clean\n"); 1544 1545 if (list_empty(&bdev->man[0].lru)) 1546 TTM_DEBUG("Swap list was clean\n"); 1547 spin_unlock(&glob->lru_lock); 1548 1549 drm_vma_offset_manager_destroy(&bdev->vma_manager); 1550 1551 return ret; 1552 } 1553 EXPORT_SYMBOL(ttm_bo_device_release); 1554 1555 int ttm_bo_device_init(struct ttm_bo_device *bdev, 1556 struct ttm_bo_global *glob, 1557 struct ttm_bo_driver *driver, 1558 uint64_t file_page_offset, 1559 bool need_dma32) 1560 { 1561 int ret = -EINVAL; 1562 1563 bdev->driver = driver; 1564 1565 memset(bdev->man, 0, sizeof(bdev->man)); 1566 1567 /* 1568 * Initialize the system memory buffer type. 1569 * Other types need to be driver / IOCTL initialized. 1570 */ 1571 ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0); 1572 if (unlikely(ret != 0)) 1573 goto out_no_sys; 1574 1575 drm_vma_offset_manager_init(&bdev->vma_manager, file_page_offset, 1576 0x10000000); 1577 INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue); 1578 INIT_LIST_HEAD(&bdev->ddestroy); 1579 bdev->dev_mapping = NULL; 1580 bdev->glob = glob; 1581 bdev->need_dma32 = need_dma32; 1582 bdev->val_seq = 0; 1583 mtx_init(&bdev->fence_lock, IPL_NONE); 1584 mutex_lock(&glob->device_list_mutex); 1585 list_add_tail(&bdev->device_list, &glob->device_list); 1586 mutex_unlock(&glob->device_list_mutex); 1587 1588 return 0; 1589 out_no_sys: 1590 return ret; 1591 } 1592 EXPORT_SYMBOL(ttm_bo_device_init); 1593 1594 /* 1595 * buffer object vm functions. 1596 */ 1597 1598 bool ttm_mem_reg_is_pci(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem) 1599 { 1600 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type]; 1601 1602 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) { 1603 if (mem->mem_type == TTM_PL_SYSTEM) 1604 return false; 1605 1606 if (man->flags & TTM_MEMTYPE_FLAG_CMA) 1607 return false; 1608 1609 if (mem->placement & TTM_PL_FLAG_CACHED) 1610 return false; 1611 } 1612 return true; 1613 } 1614 1615 void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo) 1616 { 1617 struct ttm_tt *ttm = bo->ttm; 1618 struct vm_page *page; 1619 bus_addr_t addr; 1620 paddr_t paddr; 1621 int i; 1622 1623 if (bo->mem.bus.is_iomem) { 1624 for (i = 0; i < bo->mem.num_pages; ++i) { 1625 addr = bo->mem.bus.base + bo->mem.bus.offset; 1626 paddr = bus_space_mmap(bo->bdev->memt, addr, 1627 i << PAGE_SHIFT, 0, 0); 1628 page = PHYS_TO_VM_PAGE(paddr); 1629 if (unlikely(page == NULL)) 1630 continue; 1631 pmap_page_protect(page, PROT_NONE); 1632 } 1633 } else if (ttm) { 1634 for (i = 0; i < ttm->num_pages; ++i) { 1635 page = ttm->pages[i]; 1636 if (unlikely(page == NULL)) 1637 continue; 1638 pmap_page_protect(page, PROT_NONE); 1639 } 1640 } 1641 ttm_mem_io_free_vm(bo); 1642 } 1643 1644 void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo) 1645 { 1646 struct ttm_bo_device *bdev = bo->bdev; 1647 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type]; 1648 1649 ttm_mem_io_lock(man, false); 1650 ttm_bo_unmap_virtual_locked(bo); 1651 ttm_mem_io_unlock(man); 1652 } 1653 1654 1655 EXPORT_SYMBOL(ttm_bo_unmap_virtual); 1656 1657 /** 1658 * ttm_bo_setup_vm: 1659 * 1660 * @bo: the buffer to allocate address space for 1661 * 1662 * Allocate address space in the drm device so that applications 1663 * can mmap the buffer and access the contents. This only 1664 * applies to ttm_bo_type_device objects as others are not 1665 * placed in the drm device address space. 1666 */ 1667 1668 static int ttm_bo_setup_vm(struct ttm_buffer_object *bo) 1669 { 1670 struct ttm_bo_device *bdev = bo->bdev; 1671 1672 return drm_vma_offset_add(&bdev->vma_manager, &bo->vma_node, 1673 bo->mem.num_pages); 1674 } 1675 1676 int ttm_bo_wait(struct ttm_buffer_object *bo, 1677 bool lazy, bool interruptible, bool no_wait) 1678 { 1679 struct ttm_bo_driver *driver = bo->bdev->driver; 1680 struct ttm_bo_device *bdev = bo->bdev; 1681 void *sync_obj; 1682 int ret = 0; 1683 1684 if (likely(bo->sync_obj == NULL)) 1685 return 0; 1686 1687 while (bo->sync_obj) { 1688 1689 if (driver->sync_obj_signaled(bo->sync_obj)) { 1690 void *tmp_obj = bo->sync_obj; 1691 bo->sync_obj = NULL; 1692 clear_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags); 1693 spin_unlock(&bdev->fence_lock); 1694 driver->sync_obj_unref(&tmp_obj); 1695 spin_lock(&bdev->fence_lock); 1696 continue; 1697 } 1698 1699 if (no_wait) 1700 return -EBUSY; 1701 1702 sync_obj = driver->sync_obj_ref(bo->sync_obj); 1703 spin_unlock(&bdev->fence_lock); 1704 ret = driver->sync_obj_wait(sync_obj, 1705 lazy, interruptible); 1706 if (unlikely(ret != 0)) { 1707 driver->sync_obj_unref(&sync_obj); 1708 spin_lock(&bdev->fence_lock); 1709 return ret; 1710 } 1711 spin_lock(&bdev->fence_lock); 1712 if (likely(bo->sync_obj == sync_obj)) { 1713 void *tmp_obj = bo->sync_obj; 1714 bo->sync_obj = NULL; 1715 clear_bit(TTM_BO_PRIV_FLAG_MOVING, 1716 &bo->priv_flags); 1717 spin_unlock(&bdev->fence_lock); 1718 driver->sync_obj_unref(&sync_obj); 1719 driver->sync_obj_unref(&tmp_obj); 1720 spin_lock(&bdev->fence_lock); 1721 } else { 1722 spin_unlock(&bdev->fence_lock); 1723 driver->sync_obj_unref(&sync_obj); 1724 spin_lock(&bdev->fence_lock); 1725 } 1726 } 1727 return 0; 1728 } 1729 EXPORT_SYMBOL(ttm_bo_wait); 1730 1731 int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait) 1732 { 1733 struct ttm_bo_device *bdev = bo->bdev; 1734 int ret = 0; 1735 1736 /* 1737 * Using ttm_bo_reserve makes sure the lru lists are updated. 1738 */ 1739 1740 ret = ttm_bo_reserve(bo, true, no_wait, false, 0); 1741 if (unlikely(ret != 0)) 1742 return ret; 1743 spin_lock(&bdev->fence_lock); 1744 ret = ttm_bo_wait(bo, false, true, no_wait); 1745 spin_unlock(&bdev->fence_lock); 1746 if (likely(ret == 0)) 1747 atomic_inc(&bo->cpu_writers); 1748 ttm_bo_unreserve(bo); 1749 return ret; 1750 } 1751 EXPORT_SYMBOL(ttm_bo_synccpu_write_grab); 1752 1753 void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo) 1754 { 1755 atomic_dec(&bo->cpu_writers); 1756 } 1757 EXPORT_SYMBOL(ttm_bo_synccpu_write_release); 1758 1759 /** 1760 * A buffer object shrink method that tries to swap out the first 1761 * buffer object on the bo_global::swap_lru list. 1762 */ 1763 1764 static int ttm_bo_swapout(struct ttm_mem_shrink *shrink) 1765 { 1766 struct ttm_bo_global *glob = 1767 container_of(shrink, struct ttm_bo_global, shrink); 1768 struct ttm_buffer_object *bo; 1769 int ret = -EBUSY; 1770 int put_count; 1771 uint32_t swap_placement = (TTM_PL_FLAG_CACHED | TTM_PL_FLAG_SYSTEM); 1772 1773 spin_lock(&glob->lru_lock); 1774 list_for_each_entry(bo, &glob->swap_lru, swap) { 1775 ret = ttm_bo_reserve_locked(bo, false, true, false, 0); 1776 if (!ret) 1777 break; 1778 } 1779 1780 if (ret) { 1781 spin_unlock(&glob->lru_lock); 1782 return ret; 1783 } 1784 1785 kref_get(&bo->list_kref); 1786 1787 if (!list_empty(&bo->ddestroy)) { 1788 ret = ttm_bo_cleanup_refs_and_unlock(bo, false, false); 1789 kref_put(&bo->list_kref, ttm_bo_release_list); 1790 return ret; 1791 } 1792 1793 put_count = ttm_bo_del_from_lru(bo); 1794 spin_unlock(&glob->lru_lock); 1795 1796 ttm_bo_list_ref_sub(bo, put_count, true); 1797 1798 /** 1799 * Wait for GPU, then move to system cached. 1800 */ 1801 1802 spin_lock(&bo->bdev->fence_lock); 1803 ret = ttm_bo_wait(bo, false, false, false); 1804 spin_unlock(&bo->bdev->fence_lock); 1805 1806 if (unlikely(ret != 0)) 1807 goto out; 1808 1809 if ((bo->mem.placement & swap_placement) != swap_placement) { 1810 struct ttm_mem_reg evict_mem; 1811 1812 evict_mem = bo->mem; 1813 evict_mem.mm_node = NULL; 1814 evict_mem.placement = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED; 1815 evict_mem.mem_type = TTM_PL_SYSTEM; 1816 1817 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, 1818 false, false); 1819 if (unlikely(ret != 0)) 1820 goto out; 1821 } 1822 1823 ttm_bo_unmap_virtual(bo); 1824 1825 /** 1826 * Swap out. Buffer will be swapped in again as soon as 1827 * anyone tries to access a ttm page. 1828 */ 1829 1830 if (bo->bdev->driver->swap_notify) 1831 bo->bdev->driver->swap_notify(bo); 1832 1833 ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage); 1834 out: 1835 1836 /** 1837 * 1838 * Unreserve without putting on LRU to avoid swapping out an 1839 * already swapped buffer. 1840 */ 1841 1842 atomic_set(&bo->reserved, 0); 1843 wake_up_all(&bo->event_queue); 1844 kref_put(&bo->list_kref, ttm_bo_release_list); 1845 return ret; 1846 } 1847 1848 void ttm_bo_swapout_all(struct ttm_bo_device *bdev) 1849 { 1850 while (ttm_bo_swapout(&bdev->glob->shrink) == 0) 1851 ; 1852 } 1853 EXPORT_SYMBOL(ttm_bo_swapout_all); 1854