1/************************************************************************** 2 * 3 * Copyright (c) 2006-2009 VMware, Inc., Palo Alto, CA., USA 4 * All Rights Reserved. 5 * 6 * Permission is hereby granted, free of charge, to any person obtaining a 7 * copy of this software and associated documentation files (the 8 * "Software"), to deal in the Software without restriction, including 9 * without limitation the rights to use, copy, modify, merge, publish, 10 * distribute, sub license, and/or sell copies of the Software, and to 11 * permit persons to whom the Software is furnished to do so, subject to 12 * the following conditions: 13 * 14 * The above copyright notice and this permission notice (including the 15 * next paragraph) shall be included in all copies or substantial portions 16 * of the Software. 17 * 18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 19 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 20 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL 21 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, 22 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR 23 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE 24 * USE OR OTHER DEALINGS IN THE SOFTWARE. 25 * 26 **************************************************************************/ 27/* 28 * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com> 29 */ 30 31#include <sys/cdefs.h> 32__FBSDID("$FreeBSD$"); 33 34#include <dev/drm2/drmP.h> 35#include <dev/drm2/ttm/ttm_module.h> 36#include <dev/drm2/ttm/ttm_bo_driver.h> 37#include <dev/drm2/ttm/ttm_placement.h> 38 39#define TTM_ASSERT_LOCKED(param) 40#define TTM_DEBUG(fmt, arg...) 41#define TTM_BO_HASH_ORDER 13 42 43static int ttm_bo_setup_vm(struct ttm_buffer_object *bo); 44static int ttm_bo_swapout(struct ttm_mem_shrink *shrink); 45static void ttm_bo_global_kobj_release(struct ttm_bo_global *glob); 46 47MALLOC_DEFINE(M_TTM_BO, "ttm_bo", "TTM Buffer Objects"); 48 49static inline int ttm_mem_type_from_flags(uint32_t flags, uint32_t *mem_type) 50{ 51 int i; 52 53 for (i = 0; i <= TTM_PL_PRIV5; i++) 54 if (flags & (1 << i)) { 55 *mem_type = i; 56 return 0; 57 } 58 return -EINVAL; 59} 60 61static void ttm_mem_type_debug(struct ttm_bo_device *bdev, int mem_type) 62{ 63 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 64 65 printf(" has_type: %d\n", man->has_type); 66 printf(" use_type: %d\n", man->use_type); 67 printf(" flags: 0x%08X\n", man->flags); 68 printf(" gpu_offset: 0x%08lX\n", man->gpu_offset); 69 printf(" size: %ju\n", (uintmax_t)man->size); 70 printf(" available_caching: 0x%08X\n", man->available_caching); 71 printf(" default_caching: 0x%08X\n", man->default_caching); 72 if (mem_type != TTM_PL_SYSTEM) 73 (*man->func->debug)(man, TTM_PFX); 74} 75 76static void ttm_bo_mem_space_debug(struct ttm_buffer_object *bo, 77 struct ttm_placement *placement) 78{ 79 int i, ret, mem_type; 80 81 printf("No space for %p (%lu pages, %luK, %luM)\n", 82 bo, bo->mem.num_pages, bo->mem.size >> 10, 83 bo->mem.size >> 20); 84 for (i = 0; i < placement->num_placement; i++) { 85 ret = ttm_mem_type_from_flags(placement->placement[i], 86 &mem_type); 87 if (ret) 88 return; 89 printf(" placement[%d]=0x%08X (%d)\n", 90 i, placement->placement[i], mem_type); 91 ttm_mem_type_debug(bo->bdev, mem_type); 92 } 93} 94 95#if 0 96static ssize_t ttm_bo_global_show(struct ttm_bo_global *glob, 97 char *buffer) 98{ 99 100 return snprintf(buffer, PAGE_SIZE, "%lu\n", 101 (unsigned long) atomic_read(&glob->bo_count)); 102} 103#endif 104 105static inline uint32_t ttm_bo_type_flags(unsigned type) 106{ 107 return 1 << (type); 108} 109 110static void ttm_bo_release_list(struct ttm_buffer_object *bo) 111{ 112 struct ttm_bo_device *bdev = bo->bdev; 113 size_t acc_size = bo->acc_size; 114 115 MPASS(atomic_read(&bo->list_kref) == 0); 116 MPASS(atomic_read(&bo->kref) == 0); 117 MPASS(atomic_read(&bo->cpu_writers) == 0); 118 MPASS(bo->sync_obj == NULL); 119 MPASS(bo->mem.mm_node == NULL); 120 MPASS(list_empty(&bo->lru)); 121 MPASS(list_empty(&bo->ddestroy)); 122 123 if (bo->ttm) 124 ttm_tt_destroy(bo->ttm); 125 atomic_dec(&bo->glob->bo_count); 126 if (bo->destroy) 127 bo->destroy(bo); 128 else { 129 free(bo, M_TTM_BO); 130 } 131 ttm_mem_global_free(bdev->glob->mem_glob, acc_size); 132} 133 134static int 135ttm_bo_wait_unreserved_locked(struct ttm_buffer_object *bo, bool interruptible) 136{ 137 const char *wmsg; 138 int flags, ret; 139 140 ret = 0; 141 if (interruptible) { 142 flags = PCATCH; 143 wmsg = "ttbowi"; 144 } else { 145 flags = 0; 146 wmsg = "ttbowu"; 147 } 148 while (ttm_bo_is_reserved(bo)) { 149 ret = -msleep(bo, &bo->glob->lru_lock, flags, wmsg, 0); 150 if (ret == -EINTR) 151 ret = -ERESTARTSYS; 152 if (ret != 0) 153 break; 154 } 155 return (ret); 156} 157 158void ttm_bo_add_to_lru(struct ttm_buffer_object *bo) 159{ 160 struct ttm_bo_device *bdev = bo->bdev; 161 struct ttm_mem_type_manager *man; 162 163 MPASS(ttm_bo_is_reserved(bo)); 164 165 if (!(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) { 166 167 MPASS(list_empty(&bo->lru)); 168 169 man = &bdev->man[bo->mem.mem_type]; 170 list_add_tail(&bo->lru, &man->lru); 171 refcount_acquire(&bo->list_kref); 172 173 if (bo->ttm != NULL) { 174 list_add_tail(&bo->swap, &bo->glob->swap_lru); 175 refcount_acquire(&bo->list_kref); 176 } 177 } 178} 179 180int ttm_bo_del_from_lru(struct ttm_buffer_object *bo) 181{ 182 int put_count = 0; 183 184 if (!list_empty(&bo->swap)) { 185 list_del_init(&bo->swap); 186 ++put_count; 187 } 188 if (!list_empty(&bo->lru)) { 189 list_del_init(&bo->lru); 190 ++put_count; 191 } 192 193 /* 194 * TODO: Add a driver hook to delete from 195 * driver-specific LRU's here. 196 */ 197 198 return put_count; 199} 200 201int ttm_bo_reserve_nolru(struct ttm_buffer_object *bo, 202 bool interruptible, 203 bool no_wait, bool use_sequence, uint32_t sequence) 204{ 205 int ret; 206 207 while (unlikely(atomic_xchg(&bo->reserved, 1) != 0)) { 208 /** 209 * Deadlock avoidance for multi-bo reserving. 210 */ 211 if (use_sequence && bo->seq_valid) { 212 /** 213 * We've already reserved this one. 214 */ 215 if (unlikely(sequence == bo->val_seq)) 216 return -EDEADLK; 217 /** 218 * Already reserved by a thread that will not back 219 * off for us. We need to back off. 220 */ 221 if (unlikely(sequence - bo->val_seq < (1U << 31))) 222 return -EAGAIN; 223 } 224 225 if (no_wait) 226 return -EBUSY; 227 228 ret = ttm_bo_wait_unreserved_locked(bo, interruptible); 229 230 if (unlikely(ret)) 231 return ret; 232 } 233 234 if (use_sequence) { 235 bool wake_up = false; 236 /** 237 * Wake up waiters that may need to recheck for deadlock, 238 * if we decreased the sequence number. 239 */ 240 if (unlikely((bo->val_seq - sequence < (1U << 31)) 241 || !bo->seq_valid)) 242 wake_up = true; 243 244 /* 245 * In the worst case with memory ordering these values can be 246 * seen in the wrong order. However since we call wake_up_all 247 * in that case, this will hopefully not pose a problem, 248 * and the worst case would only cause someone to accidentally 249 * hit -EAGAIN in ttm_bo_reserve when they see old value of 250 * val_seq. However this would only happen if seq_valid was 251 * written before val_seq was, and just means some slightly 252 * increased cpu usage 253 */ 254 bo->val_seq = sequence; 255 bo->seq_valid = true; 256 if (wake_up) 257 wakeup(bo); 258 } else { 259 bo->seq_valid = false; 260 } 261 262 return 0; 263} 264 265void ttm_bo_list_ref_sub(struct ttm_buffer_object *bo, int count, 266 bool never_free) 267{ 268 u_int old; 269 270 old = atomic_fetchadd_int(&bo->list_kref, -count); 271 if (old <= count) { 272 if (never_free) 273 panic("ttm_bo_ref_buf"); 274 ttm_bo_release_list(bo); 275 } 276} 277 278int ttm_bo_reserve(struct ttm_buffer_object *bo, 279 bool interruptible, 280 bool no_wait, bool use_sequence, uint32_t sequence) 281{ 282 struct ttm_bo_global *glob = bo->glob; 283 int put_count = 0; 284 int ret; 285 286 mtx_lock(&bo->glob->lru_lock); 287 ret = ttm_bo_reserve_nolru(bo, interruptible, no_wait, use_sequence, 288 sequence); 289 if (likely(ret == 0)) { 290 put_count = ttm_bo_del_from_lru(bo); 291 mtx_unlock(&glob->lru_lock); 292 ttm_bo_list_ref_sub(bo, put_count, true); 293 } else 294 mtx_unlock(&bo->glob->lru_lock); 295 296 return ret; 297} 298 299int ttm_bo_reserve_slowpath_nolru(struct ttm_buffer_object *bo, 300 bool interruptible, uint32_t sequence) 301{ 302 bool wake_up = false; 303 int ret; 304 305 while (unlikely(atomic_xchg(&bo->reserved, 1) != 0)) { 306 if (bo->seq_valid && sequence == bo->val_seq) { 307 DRM_ERROR( 308 "%s: bo->seq_valid && sequence == bo->val_seq", 309 __func__); 310 } 311 312 ret = ttm_bo_wait_unreserved_locked(bo, interruptible); 313 314 if (unlikely(ret)) 315 return ret; 316 } 317 318 if ((bo->val_seq - sequence < (1U << 31)) || !bo->seq_valid) 319 wake_up = true; 320 321 /** 322 * Wake up waiters that may need to recheck for deadlock, 323 * if we decreased the sequence number. 324 */ 325 bo->val_seq = sequence; 326 bo->seq_valid = true; 327 if (wake_up) 328 wakeup(bo); 329 330 return 0; 331} 332 333int ttm_bo_reserve_slowpath(struct ttm_buffer_object *bo, 334 bool interruptible, uint32_t sequence) 335{ 336 struct ttm_bo_global *glob = bo->glob; 337 int put_count, ret; 338 339 mtx_lock(&glob->lru_lock); 340 ret = ttm_bo_reserve_slowpath_nolru(bo, interruptible, sequence); 341 if (likely(!ret)) { 342 put_count = ttm_bo_del_from_lru(bo); 343 mtx_unlock(&glob->lru_lock); 344 ttm_bo_list_ref_sub(bo, put_count, true); 345 } else 346 mtx_unlock(&glob->lru_lock); 347 return ret; 348} 349 350void ttm_bo_unreserve_locked(struct ttm_buffer_object *bo) 351{ 352 ttm_bo_add_to_lru(bo); 353 atomic_set(&bo->reserved, 0); 354 wakeup(bo); 355} 356 357void ttm_bo_unreserve(struct ttm_buffer_object *bo) 358{ 359 struct ttm_bo_global *glob = bo->glob; 360 361 mtx_lock(&glob->lru_lock); 362 ttm_bo_unreserve_locked(bo); 363 mtx_unlock(&glob->lru_lock); 364} 365 366/* 367 * Call bo->mutex locked. 368 */ 369static int ttm_bo_add_ttm(struct ttm_buffer_object *bo, bool zero_alloc) 370{ 371 struct ttm_bo_device *bdev = bo->bdev; 372 struct ttm_bo_global *glob = bo->glob; 373 int ret = 0; 374 uint32_t page_flags = 0; 375 376 TTM_ASSERT_LOCKED(&bo->mutex); 377 bo->ttm = NULL; 378 379 if (bdev->need_dma32) 380 page_flags |= TTM_PAGE_FLAG_DMA32; 381 382 switch (bo->type) { 383 case ttm_bo_type_device: 384 if (zero_alloc) 385 page_flags |= TTM_PAGE_FLAG_ZERO_ALLOC; 386 case ttm_bo_type_kernel: 387 bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT, 388 page_flags, glob->dummy_read_page); 389 if (unlikely(bo->ttm == NULL)) 390 ret = -ENOMEM; 391 break; 392 case ttm_bo_type_sg: 393 bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT, 394 page_flags | TTM_PAGE_FLAG_SG, 395 glob->dummy_read_page); 396 if (unlikely(bo->ttm == NULL)) { 397 ret = -ENOMEM; 398 break; 399 } 400 bo->ttm->sg = bo->sg; 401 break; 402 default: 403 printf("[TTM] Illegal buffer object type\n"); 404 ret = -EINVAL; 405 break; 406 } 407 408 return ret; 409} 410 411static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo, 412 struct ttm_mem_reg *mem, 413 bool evict, bool interruptible, 414 bool no_wait_gpu) 415{ 416 struct ttm_bo_device *bdev = bo->bdev; 417 bool old_is_pci = ttm_mem_reg_is_pci(bdev, &bo->mem); 418 bool new_is_pci = ttm_mem_reg_is_pci(bdev, mem); 419 struct ttm_mem_type_manager *old_man = &bdev->man[bo->mem.mem_type]; 420 struct ttm_mem_type_manager *new_man = &bdev->man[mem->mem_type]; 421 int ret = 0; 422 423 if (old_is_pci || new_is_pci || 424 ((mem->placement & bo->mem.placement & TTM_PL_MASK_CACHING) == 0)) { 425 ret = ttm_mem_io_lock(old_man, true); 426 if (unlikely(ret != 0)) 427 goto out_err; 428 ttm_bo_unmap_virtual_locked(bo); 429 ttm_mem_io_unlock(old_man); 430 } 431 432 /* 433 * Create and bind a ttm if required. 434 */ 435 436 if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) { 437 if (bo->ttm == NULL) { 438 bool zero = !(old_man->flags & TTM_MEMTYPE_FLAG_FIXED); 439 ret = ttm_bo_add_ttm(bo, zero); 440 if (ret) 441 goto out_err; 442 } 443 444 ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement); 445 if (ret) 446 goto out_err; 447 448 if (mem->mem_type != TTM_PL_SYSTEM) { 449 ret = ttm_tt_bind(bo->ttm, mem); 450 if (ret) 451 goto out_err; 452 } 453 454 if (bo->mem.mem_type == TTM_PL_SYSTEM) { 455 if (bdev->driver->move_notify) 456 bdev->driver->move_notify(bo, mem); 457 bo->mem = *mem; 458 mem->mm_node = NULL; 459 goto moved; 460 } 461 } 462 463 if (bdev->driver->move_notify) 464 bdev->driver->move_notify(bo, mem); 465 466 if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) && 467 !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) 468 ret = ttm_bo_move_ttm(bo, evict, no_wait_gpu, mem); 469 else if (bdev->driver->move) 470 ret = bdev->driver->move(bo, evict, interruptible, 471 no_wait_gpu, mem); 472 else 473 ret = ttm_bo_move_memcpy(bo, evict, no_wait_gpu, mem); 474 475 if (ret) { 476 if (bdev->driver->move_notify) { 477 struct ttm_mem_reg tmp_mem = *mem; 478 *mem = bo->mem; 479 bo->mem = tmp_mem; 480 bdev->driver->move_notify(bo, mem); 481 bo->mem = *mem; 482 *mem = tmp_mem; 483 } 484 485 goto out_err; 486 } 487 488moved: 489 if (bo->evicted) { 490 ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement); 491 if (ret) 492 printf("[TTM] Can not flush read caches\n"); 493 bo->evicted = false; 494 } 495 496 if (bo->mem.mm_node) { 497 bo->offset = (bo->mem.start << PAGE_SHIFT) + 498 bdev->man[bo->mem.mem_type].gpu_offset; 499 bo->cur_placement = bo->mem.placement; 500 } else 501 bo->offset = 0; 502 503 return 0; 504 505out_err: 506 new_man = &bdev->man[bo->mem.mem_type]; 507 if ((new_man->flags & TTM_MEMTYPE_FLAG_FIXED) && bo->ttm) { 508 ttm_tt_unbind(bo->ttm); 509 ttm_tt_destroy(bo->ttm); 510 bo->ttm = NULL; 511 } 512 513 return ret; 514} 515 516/** 517 * Call bo::reserved. 518 * Will release GPU memory type usage on destruction. 519 * This is the place to put in driver specific hooks to release 520 * driver private resources. 521 * Will release the bo::reserved lock. 522 */ 523 524static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo) 525{ 526 if (bo->bdev->driver->move_notify) 527 bo->bdev->driver->move_notify(bo, NULL); 528 529 if (bo->ttm) { 530 ttm_tt_unbind(bo->ttm); 531 ttm_tt_destroy(bo->ttm); 532 bo->ttm = NULL; 533 } 534 ttm_bo_mem_put(bo, &bo->mem); 535 536 atomic_set(&bo->reserved, 0); 537 wakeup(&bo); 538 539 /* 540 * Since the final reference to this bo may not be dropped by 541 * the current task we have to put a memory barrier here to make 542 * sure the changes done in this function are always visible. 543 * 544 * This function only needs protection against the final kref_put. 545 */ 546 mb(); 547} 548 549static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo) 550{ 551 struct ttm_bo_device *bdev = bo->bdev; 552 struct ttm_bo_global *glob = bo->glob; 553 struct ttm_bo_driver *driver = bdev->driver; 554 void *sync_obj = NULL; 555 int put_count; 556 int ret; 557 558 mtx_lock(&glob->lru_lock); 559 ret = ttm_bo_reserve_nolru(bo, false, true, false, 0); 560 561 mtx_lock(&bdev->fence_lock); 562 (void) ttm_bo_wait(bo, false, false, true); 563 if (!ret && !bo->sync_obj) { 564 mtx_unlock(&bdev->fence_lock); 565 put_count = ttm_bo_del_from_lru(bo); 566 567 mtx_unlock(&glob->lru_lock); 568 ttm_bo_cleanup_memtype_use(bo); 569 570 ttm_bo_list_ref_sub(bo, put_count, true); 571 572 return; 573 } 574 if (bo->sync_obj) 575 sync_obj = driver->sync_obj_ref(bo->sync_obj); 576 mtx_unlock(&bdev->fence_lock); 577 578 if (!ret) { 579 atomic_set(&bo->reserved, 0); 580 wakeup(bo); 581 } 582 583 refcount_acquire(&bo->list_kref); 584 list_add_tail(&bo->ddestroy, &bdev->ddestroy); 585 mtx_unlock(&glob->lru_lock); 586 587 if (sync_obj) { 588 driver->sync_obj_flush(sync_obj); 589 driver->sync_obj_unref(&sync_obj); 590 } 591 taskqueue_enqueue_timeout(taskqueue_thread, &bdev->wq, 592 ((hz / 100) < 1) ? 1 : hz / 100); 593} 594 595/** 596 * function ttm_bo_cleanup_refs_and_unlock 597 * If bo idle, remove from delayed- and lru lists, and unref. 598 * If not idle, do nothing. 599 * 600 * Must be called with lru_lock and reservation held, this function 601 * will drop both before returning. 602 * 603 * @interruptible Any sleeps should occur interruptibly. 604 * @no_wait_gpu Never wait for gpu. Return -EBUSY instead. 605 */ 606 607static int ttm_bo_cleanup_refs_and_unlock(struct ttm_buffer_object *bo, 608 bool interruptible, 609 bool no_wait_gpu) 610{ 611 struct ttm_bo_device *bdev = bo->bdev; 612 struct ttm_bo_driver *driver = bdev->driver; 613 struct ttm_bo_global *glob = bo->glob; 614 int put_count; 615 int ret; 616 617 mtx_lock(&bdev->fence_lock); 618 ret = ttm_bo_wait(bo, false, false, true); 619 620 if (ret && !no_wait_gpu) { 621 void *sync_obj; 622 623 /* 624 * Take a reference to the fence and unreserve, 625 * at this point the buffer should be dead, so 626 * no new sync objects can be attached. 627 */ 628 sync_obj = driver->sync_obj_ref(bo->sync_obj); 629 mtx_unlock(&bdev->fence_lock); 630 631 atomic_set(&bo->reserved, 0); 632 wakeup(bo); 633 mtx_unlock(&glob->lru_lock); 634 635 ret = driver->sync_obj_wait(sync_obj, false, interruptible); 636 driver->sync_obj_unref(&sync_obj); 637 if (ret) 638 return ret; 639 640 /* 641 * remove sync_obj with ttm_bo_wait, the wait should be 642 * finished, and no new wait object should have been added. 643 */ 644 mtx_lock(&bdev->fence_lock); 645 ret = ttm_bo_wait(bo, false, false, true); 646 mtx_unlock(&bdev->fence_lock); 647 if (ret) 648 return ret; 649 650 mtx_lock(&glob->lru_lock); 651 ret = ttm_bo_reserve_nolru(bo, false, true, false, 0); 652 653 /* 654 * We raced, and lost, someone else holds the reservation now, 655 * and is probably busy in ttm_bo_cleanup_memtype_use. 656 * 657 * Even if it's not the case, because we finished waiting any 658 * delayed destruction would succeed, so just return success 659 * here. 660 */ 661 if (ret) { 662 mtx_unlock(&glob->lru_lock); 663 return 0; 664 } 665 } else 666 mtx_unlock(&bdev->fence_lock); 667 668 if (ret || unlikely(list_empty(&bo->ddestroy))) { 669 atomic_set(&bo->reserved, 0); 670 wakeup(bo); 671 mtx_unlock(&glob->lru_lock); 672 return ret; 673 } 674 675 put_count = ttm_bo_del_from_lru(bo); 676 list_del_init(&bo->ddestroy); 677 ++put_count; 678 679 mtx_unlock(&glob->lru_lock); 680 ttm_bo_cleanup_memtype_use(bo); 681 682 ttm_bo_list_ref_sub(bo, put_count, true); 683 684 return 0; 685} 686 687/** 688 * Traverse the delayed list, and call ttm_bo_cleanup_refs on all 689 * encountered buffers. 690 */ 691 692static int ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all) 693{ 694 struct ttm_bo_global *glob = bdev->glob; 695 struct ttm_buffer_object *entry = NULL; 696 int ret = 0; 697 698 mtx_lock(&glob->lru_lock); 699 if (list_empty(&bdev->ddestroy)) 700 goto out_unlock; 701 702 entry = list_first_entry(&bdev->ddestroy, 703 struct ttm_buffer_object, ddestroy); 704 refcount_acquire(&entry->list_kref); 705 706 for (;;) { 707 struct ttm_buffer_object *nentry = NULL; 708 709 if (entry->ddestroy.next != &bdev->ddestroy) { 710 nentry = list_first_entry(&entry->ddestroy, 711 struct ttm_buffer_object, ddestroy); 712 refcount_acquire(&nentry->list_kref); 713 } 714 715 ret = ttm_bo_reserve_nolru(entry, false, true, false, 0); 716 if (remove_all && ret) { 717 ret = ttm_bo_reserve_nolru(entry, false, false, 718 false, 0); 719 } 720 721 if (!ret) 722 ret = ttm_bo_cleanup_refs_and_unlock(entry, false, 723 !remove_all); 724 else 725 mtx_unlock(&glob->lru_lock); 726 727 if (refcount_release(&entry->list_kref)) 728 ttm_bo_release_list(entry); 729 entry = nentry; 730 731 if (ret || !entry) 732 goto out; 733 734 mtx_lock(&glob->lru_lock); 735 if (list_empty(&entry->ddestroy)) 736 break; 737 } 738 739out_unlock: 740 mtx_unlock(&glob->lru_lock); 741out: 742 if (entry && refcount_release(&entry->list_kref)) 743 ttm_bo_release_list(entry); 744 return ret; 745} 746 747static void ttm_bo_delayed_workqueue(void *arg, int pending __unused) 748{ 749 struct ttm_bo_device *bdev = arg; 750 751 if (ttm_bo_delayed_delete(bdev, false)) { 752 taskqueue_enqueue_timeout(taskqueue_thread, &bdev->wq, 753 ((hz / 100) < 1) ? 1 : hz / 100); 754 } 755} 756 757static void ttm_bo_release(struct ttm_buffer_object *bo) 758{ 759 struct ttm_bo_device *bdev = bo->bdev; 760 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type]; 761 762 rw_wlock(&bdev->vm_lock); 763 if (likely(bo->vm_node != NULL)) { 764 RB_REMOVE(ttm_bo_device_buffer_objects, 765 &bdev->addr_space_rb, bo); 766 drm_mm_put_block(bo->vm_node); 767 bo->vm_node = NULL; 768 } 769 rw_wunlock(&bdev->vm_lock); 770 ttm_mem_io_lock(man, false); 771 ttm_mem_io_free_vm(bo); 772 ttm_mem_io_unlock(man); 773 ttm_bo_cleanup_refs_or_queue(bo); 774 if (refcount_release(&bo->list_kref)) 775 ttm_bo_release_list(bo); 776} 777 778void ttm_bo_unref(struct ttm_buffer_object **p_bo) 779{ 780 struct ttm_buffer_object *bo = *p_bo; 781 782 *p_bo = NULL; 783 if (refcount_release(&bo->kref)) 784 ttm_bo_release(bo); 785} 786 787int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev) 788{ 789 int pending; 790 791 taskqueue_cancel_timeout(taskqueue_thread, &bdev->wq, &pending); 792 if (pending) 793 taskqueue_drain_timeout(taskqueue_thread, &bdev->wq); 794 return (pending); 795} 796 797void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched) 798{ 799 if (resched) { 800 taskqueue_enqueue_timeout(taskqueue_thread, &bdev->wq, 801 ((hz / 100) < 1) ? 1 : hz / 100); 802 } 803} 804 805static int ttm_bo_evict(struct ttm_buffer_object *bo, bool interruptible, 806 bool no_wait_gpu) 807{ 808 struct ttm_bo_device *bdev = bo->bdev; 809 struct ttm_mem_reg evict_mem; 810 struct ttm_placement placement; 811 int ret = 0; 812 813 mtx_lock(&bdev->fence_lock); 814 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu); 815 mtx_unlock(&bdev->fence_lock); 816 817 if (unlikely(ret != 0)) { 818 if (ret != -ERESTART) { 819 printf("[TTM] Failed to expire sync object before buffer eviction\n"); 820 } 821 goto out; 822 } 823 824 MPASS(ttm_bo_is_reserved(bo)); 825 826 evict_mem = bo->mem; 827 evict_mem.mm_node = NULL; 828 evict_mem.bus.io_reserved_vm = false; 829 evict_mem.bus.io_reserved_count = 0; 830 831 placement.fpfn = 0; 832 placement.lpfn = 0; 833 placement.num_placement = 0; 834 placement.num_busy_placement = 0; 835 bdev->driver->evict_flags(bo, &placement); 836 ret = ttm_bo_mem_space(bo, &placement, &evict_mem, interruptible, 837 no_wait_gpu); 838 if (ret) { 839 if (ret != -ERESTART) { 840 printf("[TTM] Failed to find memory space for buffer 0x%p eviction\n", 841 bo); 842 ttm_bo_mem_space_debug(bo, &placement); 843 } 844 goto out; 845 } 846 847 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, interruptible, 848 no_wait_gpu); 849 if (ret) { 850 if (ret != -ERESTART) 851 printf("[TTM] Buffer eviction failed\n"); 852 ttm_bo_mem_put(bo, &evict_mem); 853 goto out; 854 } 855 bo->evicted = true; 856out: 857 return ret; 858} 859 860static int ttm_mem_evict_first(struct ttm_bo_device *bdev, 861 uint32_t mem_type, 862 bool interruptible, 863 bool no_wait_gpu) 864{ 865 struct ttm_bo_global *glob = bdev->glob; 866 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 867 struct ttm_buffer_object *bo; 868 int ret = -EBUSY, put_count; 869 870 mtx_lock(&glob->lru_lock); 871 list_for_each_entry(bo, &man->lru, lru) { 872 ret = ttm_bo_reserve_nolru(bo, false, true, false, 0); 873 if (!ret) 874 break; 875 } 876 877 if (ret) { 878 mtx_unlock(&glob->lru_lock); 879 return ret; 880 } 881 882 refcount_acquire(&bo->list_kref); 883 884 if (!list_empty(&bo->ddestroy)) { 885 ret = ttm_bo_cleanup_refs_and_unlock(bo, interruptible, 886 no_wait_gpu); 887 if (refcount_release(&bo->list_kref)) 888 ttm_bo_release_list(bo); 889 return ret; 890 } 891 892 put_count = ttm_bo_del_from_lru(bo); 893 mtx_unlock(&glob->lru_lock); 894 895 MPASS(ret == 0); 896 897 ttm_bo_list_ref_sub(bo, put_count, true); 898 899 ret = ttm_bo_evict(bo, interruptible, no_wait_gpu); 900 ttm_bo_unreserve(bo); 901 902 if (refcount_release(&bo->list_kref)) 903 ttm_bo_release_list(bo); 904 return ret; 905} 906 907void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem) 908{ 909 struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type]; 910 911 if (mem->mm_node) 912 (*man->func->put_node)(man, mem); 913} 914 915/** 916 * Repeatedly evict memory from the LRU for @mem_type until we create enough 917 * space, or we've evicted everything and there isn't enough space. 918 */ 919static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo, 920 uint32_t mem_type, 921 struct ttm_placement *placement, 922 struct ttm_mem_reg *mem, 923 bool interruptible, 924 bool no_wait_gpu) 925{ 926 struct ttm_bo_device *bdev = bo->bdev; 927 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 928 int ret; 929 930 do { 931 ret = (*man->func->get_node)(man, bo, placement, mem); 932 if (unlikely(ret != 0)) 933 return ret; 934 if (mem->mm_node) 935 break; 936 ret = ttm_mem_evict_first(bdev, mem_type, 937 interruptible, no_wait_gpu); 938 if (unlikely(ret != 0)) 939 return ret; 940 } while (1); 941 if (mem->mm_node == NULL) 942 return -ENOMEM; 943 mem->mem_type = mem_type; 944 return 0; 945} 946 947static uint32_t ttm_bo_select_caching(struct ttm_mem_type_manager *man, 948 uint32_t cur_placement, 949 uint32_t proposed_placement) 950{ 951 uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING; 952 uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING; 953 954 /** 955 * Keep current caching if possible. 956 */ 957 958 if ((cur_placement & caching) != 0) 959 result |= (cur_placement & caching); 960 else if ((man->default_caching & caching) != 0) 961 result |= man->default_caching; 962 else if ((TTM_PL_FLAG_CACHED & caching) != 0) 963 result |= TTM_PL_FLAG_CACHED; 964 else if ((TTM_PL_FLAG_WC & caching) != 0) 965 result |= TTM_PL_FLAG_WC; 966 else if ((TTM_PL_FLAG_UNCACHED & caching) != 0) 967 result |= TTM_PL_FLAG_UNCACHED; 968 969 return result; 970} 971 972static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man, 973 uint32_t mem_type, 974 uint32_t proposed_placement, 975 uint32_t *masked_placement) 976{ 977 uint32_t cur_flags = ttm_bo_type_flags(mem_type); 978 979 if ((cur_flags & proposed_placement & TTM_PL_MASK_MEM) == 0) 980 return false; 981 982 if ((proposed_placement & man->available_caching) == 0) 983 return false; 984 985 cur_flags |= (proposed_placement & man->available_caching); 986 987 *masked_placement = cur_flags; 988 return true; 989} 990 991/** 992 * Creates space for memory region @mem according to its type. 993 * 994 * This function first searches for free space in compatible memory types in 995 * the priority order defined by the driver. If free space isn't found, then 996 * ttm_bo_mem_force_space is attempted in priority order to evict and find 997 * space. 998 */ 999int ttm_bo_mem_space(struct ttm_buffer_object *bo, 1000 struct ttm_placement *placement, 1001 struct ttm_mem_reg *mem, 1002 bool interruptible, 1003 bool no_wait_gpu) 1004{ 1005 struct ttm_bo_device *bdev = bo->bdev; 1006 struct ttm_mem_type_manager *man; 1007 uint32_t mem_type = TTM_PL_SYSTEM; 1008 uint32_t cur_flags = 0; 1009 bool type_found = false; 1010 bool type_ok = false; 1011 bool has_erestartsys = false; 1012 int i, ret; 1013 1014 mem->mm_node = NULL; 1015 for (i = 0; i < placement->num_placement; ++i) { 1016 ret = ttm_mem_type_from_flags(placement->placement[i], 1017 &mem_type); 1018 if (ret) 1019 return ret; 1020 man = &bdev->man[mem_type]; 1021 1022 type_ok = ttm_bo_mt_compatible(man, 1023 mem_type, 1024 placement->placement[i], 1025 &cur_flags); 1026 1027 if (!type_ok) 1028 continue; 1029 1030 cur_flags = ttm_bo_select_caching(man, bo->mem.placement, 1031 cur_flags); 1032 /* 1033 * Use the access and other non-mapping-related flag bits from 1034 * the memory placement flags to the current flags 1035 */ 1036 ttm_flag_masked(&cur_flags, placement->placement[i], 1037 ~TTM_PL_MASK_MEMTYPE); 1038 1039 if (mem_type == TTM_PL_SYSTEM) 1040 break; 1041 1042 if (man->has_type && man->use_type) { 1043 type_found = true; 1044 ret = (*man->func->get_node)(man, bo, placement, mem); 1045 if (unlikely(ret)) 1046 return ret; 1047 } 1048 if (mem->mm_node) 1049 break; 1050 } 1051 1052 if ((type_ok && (mem_type == TTM_PL_SYSTEM)) || mem->mm_node) { 1053 mem->mem_type = mem_type; 1054 mem->placement = cur_flags; 1055 return 0; 1056 } 1057 1058 if (!type_found) 1059 return -EINVAL; 1060 1061 for (i = 0; i < placement->num_busy_placement; ++i) { 1062 ret = ttm_mem_type_from_flags(placement->busy_placement[i], 1063 &mem_type); 1064 if (ret) 1065 return ret; 1066 man = &bdev->man[mem_type]; 1067 if (!man->has_type) 1068 continue; 1069 if (!ttm_bo_mt_compatible(man, 1070 mem_type, 1071 placement->busy_placement[i], 1072 &cur_flags)) 1073 continue; 1074 1075 cur_flags = ttm_bo_select_caching(man, bo->mem.placement, 1076 cur_flags); 1077 /* 1078 * Use the access and other non-mapping-related flag bits from 1079 * the memory placement flags to the current flags 1080 */ 1081 ttm_flag_masked(&cur_flags, placement->busy_placement[i], 1082 ~TTM_PL_MASK_MEMTYPE); 1083 1084 1085 if (mem_type == TTM_PL_SYSTEM) { 1086 mem->mem_type = mem_type; 1087 mem->placement = cur_flags; 1088 mem->mm_node = NULL; 1089 return 0; 1090 } 1091 1092 ret = ttm_bo_mem_force_space(bo, mem_type, placement, mem, 1093 interruptible, no_wait_gpu); 1094 if (ret == 0 && mem->mm_node) { 1095 mem->placement = cur_flags; 1096 return 0; 1097 } 1098 if (ret == -ERESTART) 1099 has_erestartsys = true; 1100 } 1101 ret = (has_erestartsys) ? -ERESTART : -ENOMEM; 1102 return ret; 1103} 1104 1105static 1106int ttm_bo_move_buffer(struct ttm_buffer_object *bo, 1107 struct ttm_placement *placement, 1108 bool interruptible, 1109 bool no_wait_gpu) 1110{ 1111 int ret = 0; 1112 struct ttm_mem_reg mem; 1113 struct ttm_bo_device *bdev = bo->bdev; 1114 1115 MPASS(ttm_bo_is_reserved(bo)); 1116 1117 /* 1118 * FIXME: It's possible to pipeline buffer moves. 1119 * Have the driver move function wait for idle when necessary, 1120 * instead of doing it here. 1121 */ 1122 mtx_lock(&bdev->fence_lock); 1123 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu); 1124 mtx_unlock(&bdev->fence_lock); 1125 if (ret) 1126 return ret; 1127 mem.num_pages = bo->num_pages; 1128 mem.size = mem.num_pages << PAGE_SHIFT; 1129 mem.page_alignment = bo->mem.page_alignment; 1130 mem.bus.io_reserved_vm = false; 1131 mem.bus.io_reserved_count = 0; 1132 /* 1133 * Determine where to move the buffer. 1134 */ 1135 ret = ttm_bo_mem_space(bo, placement, &mem, 1136 interruptible, no_wait_gpu); 1137 if (ret) 1138 goto out_unlock; 1139 ret = ttm_bo_handle_move_mem(bo, &mem, false, 1140 interruptible, no_wait_gpu); 1141out_unlock: 1142 if (ret && mem.mm_node) 1143 ttm_bo_mem_put(bo, &mem); 1144 return ret; 1145} 1146 1147static int ttm_bo_mem_compat(struct ttm_placement *placement, 1148 struct ttm_mem_reg *mem) 1149{ 1150 int i; 1151 1152 if (mem->mm_node && placement->lpfn != 0 && 1153 (mem->start < placement->fpfn || 1154 mem->start + mem->num_pages > placement->lpfn)) 1155 return -1; 1156 1157 for (i = 0; i < placement->num_placement; i++) { 1158 if ((placement->placement[i] & mem->placement & 1159 TTM_PL_MASK_CACHING) && 1160 (placement->placement[i] & mem->placement & 1161 TTM_PL_MASK_MEM)) 1162 return i; 1163 } 1164 return -1; 1165} 1166 1167int ttm_bo_validate(struct ttm_buffer_object *bo, 1168 struct ttm_placement *placement, 1169 bool interruptible, 1170 bool no_wait_gpu) 1171{ 1172 int ret; 1173 1174 MPASS(ttm_bo_is_reserved(bo)); 1175 /* Check that range is valid */ 1176 if (placement->lpfn || placement->fpfn) 1177 if (placement->fpfn > placement->lpfn || 1178 (placement->lpfn - placement->fpfn) < bo->num_pages) 1179 return -EINVAL; 1180 /* 1181 * Check whether we need to move buffer. 1182 */ 1183 ret = ttm_bo_mem_compat(placement, &bo->mem); 1184 if (ret < 0) { 1185 ret = ttm_bo_move_buffer(bo, placement, interruptible, 1186 no_wait_gpu); 1187 if (ret) 1188 return ret; 1189 } else { 1190 /* 1191 * Use the access and other non-mapping-related flag bits from 1192 * the compatible memory placement flags to the active flags 1193 */ 1194 ttm_flag_masked(&bo->mem.placement, placement->placement[ret], 1195 ~TTM_PL_MASK_MEMTYPE); 1196 } 1197 /* 1198 * We might need to add a TTM. 1199 */ 1200 if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) { 1201 ret = ttm_bo_add_ttm(bo, true); 1202 if (ret) 1203 return ret; 1204 } 1205 return 0; 1206} 1207 1208int ttm_bo_check_placement(struct ttm_buffer_object *bo, 1209 struct ttm_placement *placement) 1210{ 1211 MPASS(!((placement->fpfn || placement->lpfn) && 1212 (bo->mem.num_pages > (placement->lpfn - placement->fpfn)))); 1213 1214 return 0; 1215} 1216 1217int ttm_bo_init(struct ttm_bo_device *bdev, 1218 struct ttm_buffer_object *bo, 1219 unsigned long size, 1220 enum ttm_bo_type type, 1221 struct ttm_placement *placement, 1222 uint32_t page_alignment, 1223 bool interruptible, 1224 struct vm_object *persistent_swap_storage, 1225 size_t acc_size, 1226 struct sg_table *sg, 1227 void (*destroy) (struct ttm_buffer_object *)) 1228{ 1229 int ret = 0; 1230 unsigned long num_pages; 1231 struct ttm_mem_global *mem_glob = bdev->glob->mem_glob; 1232 1233 ret = ttm_mem_global_alloc(mem_glob, acc_size, false, false); 1234 if (ret) { 1235 printf("[TTM] Out of kernel memory\n"); 1236 if (destroy) 1237 (*destroy)(bo); 1238 else 1239 free(bo, M_TTM_BO); 1240 return -ENOMEM; 1241 } 1242 1243 num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT; 1244 if (num_pages == 0) { 1245 printf("[TTM] Illegal buffer object size\n"); 1246 if (destroy) 1247 (*destroy)(bo); 1248 else 1249 free(bo, M_TTM_BO); 1250 ttm_mem_global_free(mem_glob, acc_size); 1251 return -EINVAL; 1252 } 1253 bo->destroy = destroy; 1254 1255 refcount_init(&bo->kref, 1); 1256 refcount_init(&bo->list_kref, 1); 1257 atomic_set(&bo->cpu_writers, 0); 1258 atomic_set(&bo->reserved, 1); 1259 INIT_LIST_HEAD(&bo->lru); 1260 INIT_LIST_HEAD(&bo->ddestroy); 1261 INIT_LIST_HEAD(&bo->swap); 1262 INIT_LIST_HEAD(&bo->io_reserve_lru); 1263 bo->bdev = bdev; 1264 bo->glob = bdev->glob; 1265 bo->type = type; 1266 bo->num_pages = num_pages; 1267 bo->mem.size = num_pages << PAGE_SHIFT; 1268 bo->mem.mem_type = TTM_PL_SYSTEM; 1269 bo->mem.num_pages = bo->num_pages; 1270 bo->mem.mm_node = NULL; 1271 bo->mem.page_alignment = page_alignment; 1272 bo->mem.bus.io_reserved_vm = false; 1273 bo->mem.bus.io_reserved_count = 0; 1274 bo->priv_flags = 0; 1275 bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED); 1276 bo->seq_valid = false; 1277 bo->persistent_swap_storage = persistent_swap_storage; 1278 bo->acc_size = acc_size; 1279 bo->sg = sg; 1280 atomic_inc(&bo->glob->bo_count); 1281 1282 ret = ttm_bo_check_placement(bo, placement); 1283 if (unlikely(ret != 0)) 1284 goto out_err; 1285 1286 /* 1287 * For ttm_bo_type_device buffers, allocate 1288 * address space from the device. 1289 */ 1290 if (bo->type == ttm_bo_type_device || 1291 bo->type == ttm_bo_type_sg) { 1292 ret = ttm_bo_setup_vm(bo); 1293 if (ret) 1294 goto out_err; 1295 } 1296 1297 ret = ttm_bo_validate(bo, placement, interruptible, false); 1298 if (ret) 1299 goto out_err; 1300 1301 ttm_bo_unreserve(bo); 1302 return 0; 1303 1304out_err: 1305 ttm_bo_unreserve(bo); 1306 ttm_bo_unref(&bo); 1307 1308 return ret; 1309} 1310 1311size_t ttm_bo_acc_size(struct ttm_bo_device *bdev, 1312 unsigned long bo_size, 1313 unsigned struct_size) 1314{ 1315 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT; 1316 size_t size = 0; 1317 1318 size += ttm_round_pot(struct_size); 1319 size += PAGE_ALIGN(npages * sizeof(void *)); 1320 size += ttm_round_pot(sizeof(struct ttm_tt)); 1321 return size; 1322} 1323 1324size_t ttm_bo_dma_acc_size(struct ttm_bo_device *bdev, 1325 unsigned long bo_size, 1326 unsigned struct_size) 1327{ 1328 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT; 1329 size_t size = 0; 1330 1331 size += ttm_round_pot(struct_size); 1332 size += PAGE_ALIGN(npages * sizeof(void *)); 1333 size += PAGE_ALIGN(npages * sizeof(dma_addr_t)); 1334 size += ttm_round_pot(sizeof(struct ttm_dma_tt)); 1335 return size; 1336} 1337 1338int ttm_bo_create(struct ttm_bo_device *bdev, 1339 unsigned long size, 1340 enum ttm_bo_type type, 1341 struct ttm_placement *placement, 1342 uint32_t page_alignment, 1343 bool interruptible, 1344 struct vm_object *persistent_swap_storage, 1345 struct ttm_buffer_object **p_bo) 1346{ 1347 struct ttm_buffer_object *bo; 1348 size_t acc_size; 1349 int ret; 1350 1351 bo = malloc(sizeof(*bo), M_TTM_BO, M_WAITOK | M_ZERO); 1352 acc_size = ttm_bo_acc_size(bdev, size, sizeof(struct ttm_buffer_object)); 1353 ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment, 1354 interruptible, persistent_swap_storage, acc_size, 1355 NULL, NULL); 1356 if (likely(ret == 0)) 1357 *p_bo = bo; 1358 1359 return ret; 1360} 1361 1362static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev, 1363 unsigned mem_type, bool allow_errors) 1364{ 1365 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 1366 struct ttm_bo_global *glob = bdev->glob; 1367 int ret; 1368 1369 /* 1370 * Can't use standard list traversal since we're unlocking. 1371 */ 1372 1373 mtx_lock(&glob->lru_lock); 1374 while (!list_empty(&man->lru)) { 1375 mtx_unlock(&glob->lru_lock); 1376 ret = ttm_mem_evict_first(bdev, mem_type, false, false); 1377 if (ret) { 1378 if (allow_errors) { 1379 return ret; 1380 } else { 1381 printf("[TTM] Cleanup eviction failed\n"); 1382 } 1383 } 1384 mtx_lock(&glob->lru_lock); 1385 } 1386 mtx_unlock(&glob->lru_lock); 1387 return 0; 1388} 1389 1390int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type) 1391{ 1392 struct ttm_mem_type_manager *man; 1393 int ret = -EINVAL; 1394 1395 if (mem_type >= TTM_NUM_MEM_TYPES) { 1396 printf("[TTM] Illegal memory type %d\n", mem_type); 1397 return ret; 1398 } 1399 man = &bdev->man[mem_type]; 1400 1401 if (!man->has_type) { 1402 printf("[TTM] Trying to take down uninitialized memory manager type %u\n", 1403 mem_type); 1404 return ret; 1405 } 1406 1407 man->use_type = false; 1408 man->has_type = false; 1409 1410 ret = 0; 1411 if (mem_type > 0) { 1412 ttm_bo_force_list_clean(bdev, mem_type, false); 1413 1414 ret = (*man->func->takedown)(man); 1415 } 1416 1417 return ret; 1418} 1419 1420int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type) 1421{ 1422 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 1423 1424 if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) { 1425 printf("[TTM] Illegal memory manager memory type %u\n", mem_type); 1426 return -EINVAL; 1427 } 1428 1429 if (!man->has_type) { 1430 printf("[TTM] Memory type %u has not been initialized\n", mem_type); 1431 return 0; 1432 } 1433 1434 return ttm_bo_force_list_clean(bdev, mem_type, true); 1435} 1436 1437int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type, 1438 unsigned long p_size) 1439{ 1440 int ret = -EINVAL; 1441 struct ttm_mem_type_manager *man; 1442 1443 MPASS(type < TTM_NUM_MEM_TYPES); 1444 man = &bdev->man[type]; 1445 MPASS(!man->has_type); 1446 man->io_reserve_fastpath = true; 1447 man->use_io_reserve_lru = false; 1448 sx_init(&man->io_reserve_mutex, "ttmman"); 1449 INIT_LIST_HEAD(&man->io_reserve_lru); 1450 1451 ret = bdev->driver->init_mem_type(bdev, type, man); 1452 if (ret) 1453 return ret; 1454 man->bdev = bdev; 1455 1456 ret = 0; 1457 if (type != TTM_PL_SYSTEM) { 1458 ret = (*man->func->init)(man, p_size); 1459 if (ret) 1460 return ret; 1461 } 1462 man->has_type = true; 1463 man->use_type = true; 1464 man->size = p_size; 1465 1466 INIT_LIST_HEAD(&man->lru); 1467 1468 return 0; 1469} 1470 1471static void ttm_bo_global_kobj_release(struct ttm_bo_global *glob) 1472{ 1473 1474 ttm_mem_unregister_shrink(glob->mem_glob, &glob->shrink); 1475 vm_page_free(glob->dummy_read_page); 1476} 1477 1478void ttm_bo_global_release(struct drm_global_reference *ref) 1479{ 1480 struct ttm_bo_global *glob = ref->object; 1481 1482 if (refcount_release(&glob->kobj_ref)) 1483 ttm_bo_global_kobj_release(glob); 1484} 1485 1486int ttm_bo_global_init(struct drm_global_reference *ref) 1487{ 1488 struct ttm_bo_global_ref *bo_ref = 1489 container_of(ref, struct ttm_bo_global_ref, ref); 1490 struct ttm_bo_global *glob = ref->object; 1491 int ret; 1492 1493 sx_init(&glob->device_list_mutex, "ttmdlm"); 1494 mtx_init(&glob->lru_lock, "ttmlru", NULL, MTX_DEF); 1495 glob->mem_glob = bo_ref->mem_glob; 1496 glob->dummy_read_page = vm_page_alloc_contig(NULL, 0, 1497 VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ, 1498 1, 0, VM_MAX_ADDRESS, PAGE_SIZE, 0, VM_MEMATTR_UNCACHEABLE); 1499 1500 if (unlikely(glob->dummy_read_page == NULL)) { 1501 ret = -ENOMEM; 1502 goto out_no_drp; 1503 } 1504 1505 INIT_LIST_HEAD(&glob->swap_lru); 1506 INIT_LIST_HEAD(&glob->device_list); 1507 1508 ttm_mem_init_shrink(&glob->shrink, ttm_bo_swapout); 1509 ret = ttm_mem_register_shrink(glob->mem_glob, &glob->shrink); 1510 if (unlikely(ret != 0)) { 1511 printf("[TTM] Could not register buffer object swapout\n"); 1512 goto out_no_shrink; 1513 } 1514 1515 atomic_set(&glob->bo_count, 0); 1516 1517 refcount_init(&glob->kobj_ref, 1); 1518 return (0); 1519 1520out_no_shrink: 1521 vm_page_free(glob->dummy_read_page); 1522out_no_drp: 1523 free(glob, M_DRM_GLOBAL); 1524 return ret; 1525} 1526 1527int ttm_bo_device_release(struct ttm_bo_device *bdev) 1528{ 1529 int ret = 0; 1530 unsigned i = TTM_NUM_MEM_TYPES; 1531 struct ttm_mem_type_manager *man; 1532 struct ttm_bo_global *glob = bdev->glob; 1533 1534 while (i--) { 1535 man = &bdev->man[i]; 1536 if (man->has_type) { 1537 man->use_type = false; 1538 if ((i != TTM_PL_SYSTEM) && ttm_bo_clean_mm(bdev, i)) { 1539 ret = -EBUSY; 1540 printf("[TTM] DRM memory manager type %d is not clean\n", 1541 i); 1542 } 1543 man->has_type = false; 1544 } 1545 } 1546 1547 sx_xlock(&glob->device_list_mutex); 1548 list_del(&bdev->device_list); 1549 sx_xunlock(&glob->device_list_mutex); 1550 1551 if (taskqueue_cancel_timeout(taskqueue_thread, &bdev->wq, NULL)) 1552 taskqueue_drain_timeout(taskqueue_thread, &bdev->wq); 1553 1554 while (ttm_bo_delayed_delete(bdev, true)) 1555 ; 1556 1557 mtx_lock(&glob->lru_lock); 1558 if (list_empty(&bdev->ddestroy)) 1559 TTM_DEBUG("Delayed destroy list was clean\n"); 1560 1561 if (list_empty(&bdev->man[0].lru)) 1562 TTM_DEBUG("Swap list was clean\n"); 1563 mtx_unlock(&glob->lru_lock); 1564 1565 MPASS(drm_mm_clean(&bdev->addr_space_mm)); 1566 rw_wlock(&bdev->vm_lock); 1567 drm_mm_takedown(&bdev->addr_space_mm); 1568 rw_wunlock(&bdev->vm_lock); 1569 1570 return ret; 1571} 1572 1573int ttm_bo_device_init(struct ttm_bo_device *bdev, 1574 struct ttm_bo_global *glob, 1575 struct ttm_bo_driver *driver, 1576 uint64_t file_page_offset, 1577 bool need_dma32) 1578{ 1579 int ret = -EINVAL; 1580 1581 rw_init(&bdev->vm_lock, "ttmvml"); 1582 bdev->driver = driver; 1583 1584 memset(bdev->man, 0, sizeof(bdev->man)); 1585 1586 /* 1587 * Initialize the system memory buffer type. 1588 * Other types need to be driver / IOCTL initialized. 1589 */ 1590 ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0); 1591 if (unlikely(ret != 0)) 1592 goto out_no_sys; 1593 1594 RB_INIT(&bdev->addr_space_rb); 1595 ret = drm_mm_init(&bdev->addr_space_mm, file_page_offset, 0x10000000); 1596 if (unlikely(ret != 0)) 1597 goto out_no_addr_mm; 1598 1599 TIMEOUT_TASK_INIT(taskqueue_thread, &bdev->wq, 0, 1600 ttm_bo_delayed_workqueue, bdev); 1601 INIT_LIST_HEAD(&bdev->ddestroy); 1602 bdev->dev_mapping = NULL; 1603 bdev->glob = glob; 1604 bdev->need_dma32 = need_dma32; 1605 bdev->val_seq = 0; 1606 mtx_init(&bdev->fence_lock, "ttmfence", NULL, MTX_DEF); 1607 sx_xlock(&glob->device_list_mutex); 1608 list_add_tail(&bdev->device_list, &glob->device_list); 1609 sx_xunlock(&glob->device_list_mutex); 1610 1611 return 0; 1612out_no_addr_mm: 1613 ttm_bo_clean_mm(bdev, 0); 1614out_no_sys: 1615 return ret; 1616} 1617 1618/* 1619 * buffer object vm functions. 1620 */ 1621 1622bool ttm_mem_reg_is_pci(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem) 1623{ 1624 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type]; 1625 1626 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) { 1627 if (mem->mem_type == TTM_PL_SYSTEM) 1628 return false; 1629 1630 if (man->flags & TTM_MEMTYPE_FLAG_CMA) 1631 return false; 1632 1633 if (mem->placement & TTM_PL_FLAG_CACHED) 1634 return false; 1635 } 1636 return true; 1637} 1638 1639void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo) 1640{ 1641 1642 ttm_bo_release_mmap(bo); 1643 ttm_mem_io_free_vm(bo); 1644} 1645 1646void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo) 1647{ 1648 struct ttm_bo_device *bdev = bo->bdev; 1649 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type]; 1650 1651 ttm_mem_io_lock(man, false); 1652 ttm_bo_unmap_virtual_locked(bo); 1653 ttm_mem_io_unlock(man); 1654} 1655 1656static void ttm_bo_vm_insert_rb(struct ttm_buffer_object *bo) 1657{ 1658 struct ttm_bo_device *bdev = bo->bdev; 1659 1660 /* The caller acquired bdev->vm_lock. */ 1661 RB_INSERT(ttm_bo_device_buffer_objects, &bdev->addr_space_rb, bo); 1662} 1663 1664/** 1665 * ttm_bo_setup_vm: 1666 * 1667 * @bo: the buffer to allocate address space for 1668 * 1669 * Allocate address space in the drm device so that applications 1670 * can mmap the buffer and access the contents. This only 1671 * applies to ttm_bo_type_device objects as others are not 1672 * placed in the drm device address space. 1673 */ 1674 1675static int ttm_bo_setup_vm(struct ttm_buffer_object *bo) 1676{ 1677 struct ttm_bo_device *bdev = bo->bdev; 1678 int ret; 1679 1680retry_pre_get: 1681 ret = drm_mm_pre_get(&bdev->addr_space_mm); 1682 if (unlikely(ret != 0)) 1683 return ret; 1684 1685 rw_wlock(&bdev->vm_lock); 1686 bo->vm_node = drm_mm_search_free(&bdev->addr_space_mm, 1687 bo->mem.num_pages, 0, 0); 1688 1689 if (unlikely(bo->vm_node == NULL)) { 1690 ret = -ENOMEM; 1691 goto out_unlock; 1692 } 1693 1694 bo->vm_node = drm_mm_get_block_atomic(bo->vm_node, 1695 bo->mem.num_pages, 0); 1696 1697 if (unlikely(bo->vm_node == NULL)) { 1698 rw_wunlock(&bdev->vm_lock); 1699 goto retry_pre_get; 1700 } 1701 1702 ttm_bo_vm_insert_rb(bo); 1703 rw_wunlock(&bdev->vm_lock); 1704 bo->addr_space_offset = ((uint64_t) bo->vm_node->start) << PAGE_SHIFT; 1705 1706 return 0; 1707out_unlock: 1708 rw_wunlock(&bdev->vm_lock); 1709 return ret; 1710} 1711 1712int ttm_bo_wait(struct ttm_buffer_object *bo, 1713 bool lazy, bool interruptible, bool no_wait) 1714{ 1715 struct ttm_bo_driver *driver = bo->bdev->driver; 1716 struct ttm_bo_device *bdev = bo->bdev; 1717 void *sync_obj; 1718 int ret = 0; 1719 1720 if (likely(bo->sync_obj == NULL)) 1721 return 0; 1722 1723 while (bo->sync_obj) { 1724 1725 if (driver->sync_obj_signaled(bo->sync_obj)) { 1726 void *tmp_obj = bo->sync_obj; 1727 bo->sync_obj = NULL; 1728 clear_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags); 1729 mtx_unlock(&bdev->fence_lock); 1730 driver->sync_obj_unref(&tmp_obj); 1731 mtx_lock(&bdev->fence_lock); 1732 continue; 1733 } 1734 1735 if (no_wait) 1736 return -EBUSY; 1737 1738 sync_obj = driver->sync_obj_ref(bo->sync_obj); 1739 mtx_unlock(&bdev->fence_lock); 1740 ret = driver->sync_obj_wait(sync_obj, 1741 lazy, interruptible); 1742 if (unlikely(ret != 0)) { 1743 driver->sync_obj_unref(&sync_obj); 1744 mtx_lock(&bdev->fence_lock); 1745 return ret; 1746 } 1747 mtx_lock(&bdev->fence_lock); 1748 if (likely(bo->sync_obj == sync_obj)) { 1749 void *tmp_obj = bo->sync_obj; 1750 bo->sync_obj = NULL; 1751 clear_bit(TTM_BO_PRIV_FLAG_MOVING, 1752 &bo->priv_flags); 1753 mtx_unlock(&bdev->fence_lock); 1754 driver->sync_obj_unref(&sync_obj); 1755 driver->sync_obj_unref(&tmp_obj); 1756 mtx_lock(&bdev->fence_lock); 1757 } else { 1758 mtx_unlock(&bdev->fence_lock); 1759 driver->sync_obj_unref(&sync_obj); 1760 mtx_lock(&bdev->fence_lock); 1761 } 1762 } 1763 return 0; 1764} 1765 1766int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait) 1767{ 1768 struct ttm_bo_device *bdev = bo->bdev; 1769 int ret = 0; 1770 1771 /* 1772 * Using ttm_bo_reserve makes sure the lru lists are updated. 1773 */ 1774 1775 ret = ttm_bo_reserve(bo, true, no_wait, false, 0); 1776 if (unlikely(ret != 0)) 1777 return ret; 1778 mtx_lock(&bdev->fence_lock); 1779 ret = ttm_bo_wait(bo, false, true, no_wait); 1780 mtx_unlock(&bdev->fence_lock); 1781 if (likely(ret == 0)) 1782 atomic_inc(&bo->cpu_writers); 1783 ttm_bo_unreserve(bo); 1784 return ret; 1785} 1786 1787void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo) 1788{ 1789 atomic_dec(&bo->cpu_writers); 1790} 1791 1792/** 1793 * A buffer object shrink method that tries to swap out the first 1794 * buffer object on the bo_global::swap_lru list. 1795 */ 1796 1797static int ttm_bo_swapout(struct ttm_mem_shrink *shrink) 1798{ 1799 struct ttm_bo_global *glob = 1800 container_of(shrink, struct ttm_bo_global, shrink); 1801 struct ttm_buffer_object *bo; 1802 int ret = -EBUSY; 1803 int put_count; 1804 uint32_t swap_placement = (TTM_PL_FLAG_CACHED | TTM_PL_FLAG_SYSTEM); 1805 1806 mtx_lock(&glob->lru_lock); 1807 list_for_each_entry(bo, &glob->swap_lru, swap) { 1808 ret = ttm_bo_reserve_nolru(bo, false, true, false, 0); 1809 if (!ret) 1810 break; 1811 } 1812 1813 if (ret) { 1814 mtx_unlock(&glob->lru_lock); 1815 return ret; 1816 } 1817 1818 refcount_acquire(&bo->list_kref); 1819 1820 if (!list_empty(&bo->ddestroy)) { 1821 ret = ttm_bo_cleanup_refs_and_unlock(bo, false, false); 1822 if (refcount_release(&bo->list_kref)) 1823 ttm_bo_release_list(bo); 1824 return ret; 1825 } 1826 1827 put_count = ttm_bo_del_from_lru(bo); 1828 mtx_unlock(&glob->lru_lock); 1829 1830 ttm_bo_list_ref_sub(bo, put_count, true); 1831 1832 /** 1833 * Wait for GPU, then move to system cached. 1834 */ 1835 1836 mtx_lock(&bo->bdev->fence_lock); 1837 ret = ttm_bo_wait(bo, false, false, false); 1838 mtx_unlock(&bo->bdev->fence_lock); 1839 1840 if (unlikely(ret != 0)) 1841 goto out; 1842 1843 if ((bo->mem.placement & swap_placement) != swap_placement) { 1844 struct ttm_mem_reg evict_mem; 1845 1846 evict_mem = bo->mem; 1847 evict_mem.mm_node = NULL; 1848 evict_mem.placement = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED; 1849 evict_mem.mem_type = TTM_PL_SYSTEM; 1850 1851 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, 1852 false, false); 1853 if (unlikely(ret != 0)) 1854 goto out; 1855 } 1856 1857 ttm_bo_unmap_virtual(bo); 1858 1859 /** 1860 * Swap out. Buffer will be swapped in again as soon as 1861 * anyone tries to access a ttm page. 1862 */ 1863 1864 if (bo->bdev->driver->swap_notify) 1865 bo->bdev->driver->swap_notify(bo); 1866 1867 ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage); 1868out: 1869 1870 /** 1871 * 1872 * Unreserve without putting on LRU to avoid swapping out an 1873 * already swapped buffer. 1874 */ 1875 1876 atomic_set(&bo->reserved, 0); 1877 wakeup(bo); 1878 if (refcount_release(&bo->list_kref)) 1879 ttm_bo_release_list(bo); 1880 return ret; 1881} 1882 1883void ttm_bo_swapout_all(struct ttm_bo_device *bdev) 1884{ 1885 while (ttm_bo_swapout(&bdev->glob->shrink) == 0) 1886 ; 1887} 1888