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#include <vm/vm_pageout.h> 39 40#define TTM_ASSERT_LOCKED(param) 41#define TTM_DEBUG(fmt, arg...) 42#define TTM_BO_HASH_ORDER 13 43 44static int ttm_bo_setup_vm(struct ttm_buffer_object *bo); 45static int ttm_bo_swapout(struct ttm_mem_shrink *shrink); 46static void ttm_bo_global_kobj_release(struct ttm_bo_global *glob); 47 48MALLOC_DEFINE(M_TTM_BO, "ttm_bo", "TTM Buffer Objects"); 49 50static inline int ttm_mem_type_from_flags(uint32_t flags, uint32_t *mem_type) 51{ 52 int i; 53 54 for (i = 0; i <= TTM_PL_PRIV5; i++) 55 if (flags & (1 << i)) { 56 *mem_type = i; 57 return 0; 58 } 59 return -EINVAL; 60} 61 62static void ttm_mem_type_debug(struct ttm_bo_device *bdev, int mem_type) 63{ 64 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 65 66 printf(" has_type: %d\n", man->has_type); 67 printf(" use_type: %d\n", man->use_type); 68 printf(" flags: 0x%08X\n", man->flags); 69 printf(" gpu_offset: 0x%08lX\n", man->gpu_offset); 70 printf(" size: %ju\n", (uintmax_t)man->size); 71 printf(" available_caching: 0x%08X\n", man->available_caching); 72 printf(" default_caching: 0x%08X\n", man->default_caching); 73 if (mem_type != TTM_PL_SYSTEM) 74 (*man->func->debug)(man, TTM_PFX); 75} 76 77static void ttm_bo_mem_space_debug(struct ttm_buffer_object *bo, 78 struct ttm_placement *placement) 79{ 80 int i, ret, mem_type; 81 82 printf("No space for %p (%lu pages, %luK, %luM)\n", 83 bo, bo->mem.num_pages, bo->mem.size >> 10, 84 bo->mem.size >> 20); 85 for (i = 0; i < placement->num_placement; i++) { 86 ret = ttm_mem_type_from_flags(placement->placement[i], 87 &mem_type); 88 if (ret) 89 return; 90 printf(" placement[%d]=0x%08X (%d)\n", 91 i, placement->placement[i], mem_type); 92 ttm_mem_type_debug(bo->bdev, mem_type); 93 } 94} 95 96#if 0 97static ssize_t ttm_bo_global_show(struct ttm_bo_global *glob, 98 char *buffer) 99{ 100 101 return snprintf(buffer, PAGE_SIZE, "%lu\n", 102 (unsigned long) atomic_read(&glob->bo_count)); 103} 104#endif 105 106static inline uint32_t ttm_bo_type_flags(unsigned type) 107{ 108 return 1 << (type); 109} 110 111static void ttm_bo_release_list(struct ttm_buffer_object *bo) 112{ 113 struct ttm_bo_device *bdev = bo->bdev; 114 size_t acc_size = bo->acc_size; 115 116 MPASS(atomic_read(&bo->list_kref) == 0); 117 MPASS(atomic_read(&bo->kref) == 0); 118 MPASS(atomic_read(&bo->cpu_writers) == 0); 119 MPASS(bo->sync_obj == NULL); 120 MPASS(bo->mem.mm_node == NULL); 121 MPASS(list_empty(&bo->lru)); 122 MPASS(list_empty(&bo->ddestroy)); 123 124 if (bo->ttm) 125 ttm_tt_destroy(bo->ttm); 126 atomic_dec(&bo->glob->bo_count); 127 if (bo->destroy) 128 bo->destroy(bo); 129 else { 130 free(bo, M_TTM_BO); 131 } 132 ttm_mem_global_free(bdev->glob->mem_glob, acc_size); 133} 134 135static int 136ttm_bo_wait_unreserved_locked(struct ttm_buffer_object *bo, bool interruptible) 137{ 138 const char *wmsg; 139 int flags, ret; 140 141 ret = 0; 142 if (interruptible) { 143 flags = PCATCH; 144 wmsg = "ttbowi"; 145 } else { 146 flags = 0; 147 wmsg = "ttbowu"; 148 } 149 while (ttm_bo_is_reserved(bo)) { 150 ret = -msleep(bo, &bo->glob->lru_lock, flags, wmsg, 0); 151 if (ret == -EINTR || ret == -ERESTART) 152 ret = -ERESTARTSYS; 153 if (ret != 0) 154 break; 155 } 156 return (ret); 157} 158 159void ttm_bo_add_to_lru(struct ttm_buffer_object *bo) 160{ 161 struct ttm_bo_device *bdev = bo->bdev; 162 struct ttm_mem_type_manager *man; 163 164 MPASS(ttm_bo_is_reserved(bo)); 165 166 if (!(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) { 167 168 MPASS(list_empty(&bo->lru)); 169 170 man = &bdev->man[bo->mem.mem_type]; 171 list_add_tail(&bo->lru, &man->lru); 172 refcount_acquire(&bo->list_kref); 173 174 if (bo->ttm != NULL) { 175 list_add_tail(&bo->swap, &bo->glob->swap_lru); 176 refcount_acquire(&bo->list_kref); 177 } 178 } 179} 180 181int ttm_bo_del_from_lru(struct ttm_buffer_object *bo) 182{ 183 int put_count = 0; 184 185 if (!list_empty(&bo->swap)) { 186 list_del_init(&bo->swap); 187 ++put_count; 188 } 189 if (!list_empty(&bo->lru)) { 190 list_del_init(&bo->lru); 191 ++put_count; 192 } 193 194 /* 195 * TODO: Add a driver hook to delete from 196 * driver-specific LRU's here. 197 */ 198 199 return put_count; 200} 201 202int ttm_bo_reserve_nolru(struct ttm_buffer_object *bo, 203 bool interruptible, 204 bool no_wait, bool use_sequence, uint32_t sequence) 205{ 206 int ret; 207 208 while (unlikely(atomic_xchg(&bo->reserved, 1) != 0)) { 209 /** 210 * Deadlock avoidance for multi-bo reserving. 211 */ 212 if (use_sequence && bo->seq_valid) { 213 /** 214 * We've already reserved this one. 215 */ 216 if (unlikely(sequence == bo->val_seq)) 217 return -EDEADLK; 218 /** 219 * Already reserved by a thread that will not back 220 * off for us. We need to back off. 221 */ 222 if (unlikely(sequence - bo->val_seq < (1U << 31))) 223 return -EAGAIN; 224 } 225 226 if (no_wait) 227 return -EBUSY; 228 229 ret = ttm_bo_wait_unreserved_locked(bo, interruptible); 230 231 if (unlikely(ret)) 232 return ret; 233 } 234 235 if (use_sequence) { 236 bool wake_up = false; 237 /** 238 * Wake up waiters that may need to recheck for deadlock, 239 * if we decreased the sequence number. 240 */ 241 if (unlikely((bo->val_seq - sequence < (1U << 31)) 242 || !bo->seq_valid)) 243 wake_up = true; 244 245 /* 246 * In the worst case with memory ordering these values can be 247 * seen in the wrong order. However since we call wake_up_all 248 * in that case, this will hopefully not pose a problem, 249 * and the worst case would only cause someone to accidentally 250 * hit -EAGAIN in ttm_bo_reserve when they see old value of 251 * val_seq. However this would only happen if seq_valid was 252 * written before val_seq was, and just means some slightly 253 * increased cpu usage 254 */ 255 bo->val_seq = sequence; 256 bo->seq_valid = true; 257 if (wake_up) 258 wakeup(bo); 259 } else { 260 bo->seq_valid = false; 261 } 262 263 return 0; 264} 265 266void ttm_bo_list_ref_sub(struct ttm_buffer_object *bo, int count, 267 bool never_free) 268{ 269 u_int old; 270 271 old = atomic_fetchadd_int(&bo->list_kref, -count); 272 if (old <= count) { 273 if (never_free) 274 panic("ttm_bo_ref_buf"); 275 ttm_bo_release_list(bo); 276 } 277} 278 279int ttm_bo_reserve(struct ttm_buffer_object *bo, 280 bool interruptible, 281 bool no_wait, bool use_sequence, uint32_t sequence) 282{ 283 struct ttm_bo_global *glob = bo->glob; 284 int put_count = 0; 285 int ret; 286 287 mtx_lock(&bo->glob->lru_lock); 288 ret = ttm_bo_reserve_nolru(bo, interruptible, no_wait, use_sequence, 289 sequence); 290 if (likely(ret == 0)) { 291 put_count = ttm_bo_del_from_lru(bo); 292 mtx_unlock(&glob->lru_lock); 293 ttm_bo_list_ref_sub(bo, put_count, true); 294 } else 295 mtx_unlock(&bo->glob->lru_lock); 296 297 return ret; 298} 299 300int ttm_bo_reserve_slowpath_nolru(struct ttm_buffer_object *bo, 301 bool interruptible, uint32_t sequence) 302{ 303 bool wake_up = false; 304 int ret; 305 306 while (unlikely(atomic_xchg(&bo->reserved, 1) != 0)) { 307 if (bo->seq_valid && sequence == bo->val_seq) { 308 DRM_ERROR( 309 "%s: bo->seq_valid && sequence == bo->val_seq", 310 __func__); 311 } 312 313 ret = ttm_bo_wait_unreserved_locked(bo, interruptible); 314 315 if (unlikely(ret)) 316 return ret; 317 } 318 319 if ((bo->val_seq - sequence < (1U << 31)) || !bo->seq_valid) 320 wake_up = true; 321 322 /** 323 * Wake up waiters that may need to recheck for deadlock, 324 * if we decreased the sequence number. 325 */ 326 bo->val_seq = sequence; 327 bo->seq_valid = true; 328 if (wake_up) 329 wakeup(bo); 330 331 return 0; 332} 333 334int ttm_bo_reserve_slowpath(struct ttm_buffer_object *bo, 335 bool interruptible, uint32_t sequence) 336{ 337 struct ttm_bo_global *glob = bo->glob; 338 int put_count, ret; 339 340 mtx_lock(&glob->lru_lock); 341 ret = ttm_bo_reserve_slowpath_nolru(bo, interruptible, sequence); 342 if (likely(!ret)) { 343 put_count = ttm_bo_del_from_lru(bo); 344 mtx_unlock(&glob->lru_lock); 345 ttm_bo_list_ref_sub(bo, put_count, true); 346 } else 347 mtx_unlock(&glob->lru_lock); 348 return ret; 349} 350 351void ttm_bo_unreserve_locked(struct ttm_buffer_object *bo) 352{ 353 ttm_bo_add_to_lru(bo); 354 atomic_set(&bo->reserved, 0); 355 wakeup(bo); 356} 357 358void ttm_bo_unreserve(struct ttm_buffer_object *bo) 359{ 360 struct ttm_bo_global *glob = bo->glob; 361 362 mtx_lock(&glob->lru_lock); 363 ttm_bo_unreserve_locked(bo); 364 mtx_unlock(&glob->lru_lock); 365} 366 367/* 368 * Call bo->mutex locked. 369 */ 370static int ttm_bo_add_ttm(struct ttm_buffer_object *bo, bool zero_alloc) 371{ 372 struct ttm_bo_device *bdev = bo->bdev; 373 struct ttm_bo_global *glob = bo->glob; 374 int ret = 0; 375 uint32_t page_flags = 0; 376 377 TTM_ASSERT_LOCKED(&bo->mutex); 378 bo->ttm = NULL; 379 380 if (bdev->need_dma32) 381 page_flags |= TTM_PAGE_FLAG_DMA32; 382 383 switch (bo->type) { 384 case ttm_bo_type_device: 385 if (zero_alloc) 386 page_flags |= TTM_PAGE_FLAG_ZERO_ALLOC; 387 case ttm_bo_type_kernel: 388 bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT, 389 page_flags, glob->dummy_read_page); 390 if (unlikely(bo->ttm == NULL)) 391 ret = -ENOMEM; 392 break; 393 case ttm_bo_type_sg: 394 bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT, 395 page_flags | TTM_PAGE_FLAG_SG, 396 glob->dummy_read_page); 397 if (unlikely(bo->ttm == NULL)) { 398 ret = -ENOMEM; 399 break; 400 } 401 bo->ttm->sg = bo->sg; 402 break; 403 default: 404 printf("[TTM] Illegal buffer object type\n"); 405 ret = -EINVAL; 406 break; 407 } 408 409 return ret; 410} 411 412static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo, 413 struct ttm_mem_reg *mem, 414 bool evict, bool interruptible, 415 bool no_wait_gpu) 416{ 417 struct ttm_bo_device *bdev = bo->bdev; 418 bool old_is_pci = ttm_mem_reg_is_pci(bdev, &bo->mem); 419 bool new_is_pci = ttm_mem_reg_is_pci(bdev, mem); 420 struct ttm_mem_type_manager *old_man = &bdev->man[bo->mem.mem_type]; 421 struct ttm_mem_type_manager *new_man = &bdev->man[mem->mem_type]; 422 int ret = 0; 423 424 if (old_is_pci || new_is_pci || 425 ((mem->placement & bo->mem.placement & TTM_PL_MASK_CACHING) == 0)) { 426 ret = ttm_mem_io_lock(old_man, true); 427 if (unlikely(ret != 0)) 428 goto out_err; 429 ttm_bo_unmap_virtual_locked(bo); 430 ttm_mem_io_unlock(old_man); 431 } 432 433 /* 434 * Create and bind a ttm if required. 435 */ 436 437 if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) { 438 if (bo->ttm == NULL) { 439 bool zero = !(old_man->flags & TTM_MEMTYPE_FLAG_FIXED); 440 ret = ttm_bo_add_ttm(bo, zero); 441 if (ret) 442 goto out_err; 443 } 444 445 ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement); 446 if (ret) 447 goto out_err; 448 449 if (mem->mem_type != TTM_PL_SYSTEM) { 450 ret = ttm_tt_bind(bo->ttm, mem); 451 if (ret) 452 goto out_err; 453 } 454 455 if (bo->mem.mem_type == TTM_PL_SYSTEM) { 456 if (bdev->driver->move_notify) 457 bdev->driver->move_notify(bo, mem); 458 bo->mem = *mem; 459 mem->mm_node = NULL; 460 goto moved; 461 } 462 } 463 464 if (bdev->driver->move_notify) 465 bdev->driver->move_notify(bo, mem); 466 467 if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) && 468 !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) 469 ret = ttm_bo_move_ttm(bo, evict, no_wait_gpu, mem); 470 else if (bdev->driver->move) 471 ret = bdev->driver->move(bo, evict, interruptible, 472 no_wait_gpu, mem); 473 else 474 ret = ttm_bo_move_memcpy(bo, evict, no_wait_gpu, mem); 475 476 if (ret) { 477 if (bdev->driver->move_notify) { 478 struct ttm_mem_reg tmp_mem = *mem; 479 *mem = bo->mem; 480 bo->mem = tmp_mem; 481 bdev->driver->move_notify(bo, mem); 482 bo->mem = *mem; 483 *mem = tmp_mem; 484 } 485 486 goto out_err; 487 } 488 489moved: 490 if (bo->evicted) { 491 ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement); 492 if (ret) 493 printf("[TTM] Can not flush read caches\n"); 494 bo->evicted = false; 495 } 496 497 if (bo->mem.mm_node) { 498 bo->offset = (bo->mem.start << PAGE_SHIFT) + 499 bdev->man[bo->mem.mem_type].gpu_offset; 500 bo->cur_placement = bo->mem.placement; 501 } else 502 bo->offset = 0; 503 504 return 0; 505 506out_err: 507 new_man = &bdev->man[bo->mem.mem_type]; 508 if ((new_man->flags & TTM_MEMTYPE_FLAG_FIXED) && bo->ttm) { 509 ttm_tt_unbind(bo->ttm); 510 ttm_tt_destroy(bo->ttm); 511 bo->ttm = NULL; 512 } 513 514 return ret; 515} 516 517/** 518 * Call bo::reserved. 519 * Will release GPU memory type usage on destruction. 520 * This is the place to put in driver specific hooks to release 521 * driver private resources. 522 * Will release the bo::reserved lock. 523 */ 524 525static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo) 526{ 527 if (bo->bdev->driver->move_notify) 528 bo->bdev->driver->move_notify(bo, NULL); 529 530 if (bo->ttm) { 531 ttm_tt_unbind(bo->ttm); 532 ttm_tt_destroy(bo->ttm); 533 bo->ttm = NULL; 534 } 535 ttm_bo_mem_put(bo, &bo->mem); 536 537 atomic_set(&bo->reserved, 0); 538 wakeup(&bo); 539 540 /* 541 * Since the final reference to this bo may not be dropped by 542 * the current task we have to put a memory barrier here to make 543 * sure the changes done in this function are always visible. 544 * 545 * This function only needs protection against the final kref_put. 546 */ 547 mb(); 548} 549 550static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo) 551{ 552 struct ttm_bo_device *bdev = bo->bdev; 553 struct ttm_bo_global *glob = bo->glob; 554 struct ttm_bo_driver *driver = bdev->driver; 555 void *sync_obj = NULL; 556 int put_count; 557 int ret; 558 559 mtx_lock(&glob->lru_lock); 560 ret = ttm_bo_reserve_nolru(bo, false, true, false, 0); 561 562 mtx_lock(&bdev->fence_lock); 563 (void) ttm_bo_wait(bo, false, false, true); 564 if (!ret && !bo->sync_obj) { 565 mtx_unlock(&bdev->fence_lock); 566 put_count = ttm_bo_del_from_lru(bo); 567 568 mtx_unlock(&glob->lru_lock); 569 ttm_bo_cleanup_memtype_use(bo); 570 571 ttm_bo_list_ref_sub(bo, put_count, true); 572 573 return; 574 } 575 if (bo->sync_obj) 576 sync_obj = driver->sync_obj_ref(bo->sync_obj); 577 mtx_unlock(&bdev->fence_lock); 578 579 if (!ret) { 580 atomic_set(&bo->reserved, 0); 581 wakeup(bo); 582 } 583 584 refcount_acquire(&bo->list_kref); 585 list_add_tail(&bo->ddestroy, &bdev->ddestroy); 586 mtx_unlock(&glob->lru_lock); 587 588 if (sync_obj) { 589 driver->sync_obj_flush(sync_obj); 590 driver->sync_obj_unref(&sync_obj); 591 } 592 taskqueue_enqueue_timeout(taskqueue_thread, &bdev->wq, 593 ((hz / 100) < 1) ? 1 : hz / 100); 594} 595 596/** 597 * function ttm_bo_cleanup_refs_and_unlock 598 * If bo idle, remove from delayed- and lru lists, and unref. 599 * If not idle, do nothing. 600 * 601 * Must be called with lru_lock and reservation held, this function 602 * will drop both before returning. 603 * 604 * @interruptible Any sleeps should occur interruptibly. 605 * @no_wait_gpu Never wait for gpu. Return -EBUSY instead. 606 */ 607 608static int ttm_bo_cleanup_refs_and_unlock(struct ttm_buffer_object *bo, 609 bool interruptible, 610 bool no_wait_gpu) 611{ 612 struct ttm_bo_device *bdev = bo->bdev; 613 struct ttm_bo_driver *driver = bdev->driver; 614 struct ttm_bo_global *glob = bo->glob; 615 int put_count; 616 int ret; 617 618 mtx_lock(&bdev->fence_lock); 619 ret = ttm_bo_wait(bo, false, false, true); 620 621 if (ret && !no_wait_gpu) { 622 void *sync_obj; 623 624 /* 625 * Take a reference to the fence and unreserve, 626 * at this point the buffer should be dead, so 627 * no new sync objects can be attached. 628 */ 629 sync_obj = driver->sync_obj_ref(bo->sync_obj); 630 mtx_unlock(&bdev->fence_lock); 631 632 atomic_set(&bo->reserved, 0); 633 wakeup(bo); 634 mtx_unlock(&glob->lru_lock); 635 636 ret = driver->sync_obj_wait(sync_obj, false, interruptible); 637 driver->sync_obj_unref(&sync_obj); 638 if (ret) 639 return ret; 640 641 /* 642 * remove sync_obj with ttm_bo_wait, the wait should be 643 * finished, and no new wait object should have been added. 644 */ 645 mtx_lock(&bdev->fence_lock); 646 ret = ttm_bo_wait(bo, false, false, true); 647 mtx_unlock(&bdev->fence_lock); 648 if (ret) 649 return ret; 650 651 mtx_lock(&glob->lru_lock); 652 ret = ttm_bo_reserve_nolru(bo, false, true, false, 0); 653 654 /* 655 * We raced, and lost, someone else holds the reservation now, 656 * and is probably busy in ttm_bo_cleanup_memtype_use. 657 * 658 * Even if it's not the case, because we finished waiting any 659 * delayed destruction would succeed, so just return success 660 * here. 661 */ 662 if (ret) { 663 mtx_unlock(&glob->lru_lock); 664 return 0; 665 } 666 } else 667 mtx_unlock(&bdev->fence_lock); 668 669 if (ret || unlikely(list_empty(&bo->ddestroy))) { 670 atomic_set(&bo->reserved, 0); 671 wakeup(bo); 672 mtx_unlock(&glob->lru_lock); 673 return ret; 674 } 675 676 put_count = ttm_bo_del_from_lru(bo); 677 list_del_init(&bo->ddestroy); 678 ++put_count; 679 680 mtx_unlock(&glob->lru_lock); 681 ttm_bo_cleanup_memtype_use(bo); 682 683 ttm_bo_list_ref_sub(bo, put_count, true); 684 685 return 0; 686} 687 688/** 689 * Traverse the delayed list, and call ttm_bo_cleanup_refs on all 690 * encountered buffers. 691 */ 692 693static int ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all) 694{ 695 struct ttm_bo_global *glob = bdev->glob; 696 struct ttm_buffer_object *entry = NULL; 697 int ret = 0; 698 699 mtx_lock(&glob->lru_lock); 700 if (list_empty(&bdev->ddestroy)) 701 goto out_unlock; 702 703 entry = list_first_entry(&bdev->ddestroy, 704 struct ttm_buffer_object, ddestroy); 705 refcount_acquire(&entry->list_kref); 706 707 for (;;) { 708 struct ttm_buffer_object *nentry = NULL; 709 710 if (entry->ddestroy.next != &bdev->ddestroy) { 711 nentry = list_first_entry(&entry->ddestroy, 712 struct ttm_buffer_object, ddestroy); 713 refcount_acquire(&nentry->list_kref); 714 } 715 716 ret = ttm_bo_reserve_nolru(entry, false, true, false, 0); 717 if (remove_all && ret) { 718 ret = ttm_bo_reserve_nolru(entry, false, false, 719 false, 0); 720 } 721 722 if (!ret) 723 ret = ttm_bo_cleanup_refs_and_unlock(entry, false, 724 !remove_all); 725 else 726 mtx_unlock(&glob->lru_lock); 727 728 if (refcount_release(&entry->list_kref)) 729 ttm_bo_release_list(entry); 730 entry = nentry; 731 732 if (ret || !entry) 733 goto out; 734 735 mtx_lock(&glob->lru_lock); 736 if (list_empty(&entry->ddestroy)) 737 break; 738 } 739 740out_unlock: 741 mtx_unlock(&glob->lru_lock); 742out: 743 if (entry && refcount_release(&entry->list_kref)) 744 ttm_bo_release_list(entry); 745 return ret; 746} 747 748static void ttm_bo_delayed_workqueue(void *arg, int pending __unused) 749{ 750 struct ttm_bo_device *bdev = arg; 751 752 if (ttm_bo_delayed_delete(bdev, false)) { 753 taskqueue_enqueue_timeout(taskqueue_thread, &bdev->wq, 754 ((hz / 100) < 1) ? 1 : hz / 100); 755 } 756} 757 758static void ttm_bo_release(struct ttm_buffer_object *bo) 759{ 760 struct ttm_bo_device *bdev = bo->bdev; 761 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type]; 762 763 rw_wlock(&bdev->vm_lock); 764 if (likely(bo->vm_node != NULL)) { 765 RB_REMOVE(ttm_bo_device_buffer_objects, 766 &bdev->addr_space_rb, bo); 767 drm_mm_put_block(bo->vm_node); 768 bo->vm_node = NULL; 769 } 770 rw_wunlock(&bdev->vm_lock); 771 ttm_mem_io_lock(man, false); 772 ttm_mem_io_free_vm(bo); 773 ttm_mem_io_unlock(man); 774 ttm_bo_cleanup_refs_or_queue(bo); 775 if (refcount_release(&bo->list_kref)) 776 ttm_bo_release_list(bo); 777} 778 779void ttm_bo_unref(struct ttm_buffer_object **p_bo) 780{ 781 struct ttm_buffer_object *bo = *p_bo; 782 783 *p_bo = NULL; 784 if (refcount_release(&bo->kref)) 785 ttm_bo_release(bo); 786} 787 788int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev) 789{ 790 int pending; 791 792 if (taskqueue_cancel_timeout(taskqueue_thread, &bdev->wq, &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 != -ERESTARTSYS) { 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 != -ERESTARTSYS) { 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 != -ERESTARTSYS) 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 == -ERESTARTSYS) 1099 has_erestartsys = true; 1100 } 1101 ret = (has_erestartsys) ? -ERESTARTSYS : -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 req, ret; 1492 int tries; 1493 1494 sx_init(&glob->device_list_mutex, "ttmdlm"); 1495 mtx_init(&glob->lru_lock, "ttmlru", NULL, MTX_DEF); 1496 glob->mem_glob = bo_ref->mem_glob; 1497 req = VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ; 1498 tries = 0; 1499retry: 1500 glob->dummy_read_page = vm_page_alloc_contig(NULL, 0, req, 1501 1, 0, VM_MAX_ADDRESS, PAGE_SIZE, 0, VM_MEMATTR_UNCACHEABLE); 1502 1503 if (unlikely(glob->dummy_read_page == NULL)) { 1504 if (tries < 1 && vm_page_reclaim_contig(req, 1, 1505 0, VM_MAX_ADDRESS, PAGE_SIZE, 0)) { 1506 tries++; 1507 goto retry; 1508 } 1509 ret = -ENOMEM; 1510 goto out_no_drp; 1511 } 1512 1513 INIT_LIST_HEAD(&glob->swap_lru); 1514 INIT_LIST_HEAD(&glob->device_list); 1515 1516 ttm_mem_init_shrink(&glob->shrink, ttm_bo_swapout); 1517 ret = ttm_mem_register_shrink(glob->mem_glob, &glob->shrink); 1518 if (unlikely(ret != 0)) { 1519 printf("[TTM] Could not register buffer object swapout\n"); 1520 goto out_no_shrink; 1521 } 1522 1523 atomic_set(&glob->bo_count, 0); 1524 1525 refcount_init(&glob->kobj_ref, 1); 1526 return (0); 1527 1528out_no_shrink: 1529 vm_page_free(glob->dummy_read_page); 1530out_no_drp: 1531 free(glob, M_DRM_GLOBAL); 1532 return ret; 1533} 1534 1535int ttm_bo_device_release(struct ttm_bo_device *bdev) 1536{ 1537 int ret = 0; 1538 unsigned i = TTM_NUM_MEM_TYPES; 1539 struct ttm_mem_type_manager *man; 1540 struct ttm_bo_global *glob = bdev->glob; 1541 1542 while (i--) { 1543 man = &bdev->man[i]; 1544 if (man->has_type) { 1545 man->use_type = false; 1546 if ((i != TTM_PL_SYSTEM) && ttm_bo_clean_mm(bdev, i)) { 1547 ret = -EBUSY; 1548 printf("[TTM] DRM memory manager type %d is not clean\n", 1549 i); 1550 } 1551 man->has_type = false; 1552 } 1553 } 1554 1555 sx_xlock(&glob->device_list_mutex); 1556 list_del(&bdev->device_list); 1557 sx_xunlock(&glob->device_list_mutex); 1558 1559 if (taskqueue_cancel_timeout(taskqueue_thread, &bdev->wq, NULL)) 1560 taskqueue_drain_timeout(taskqueue_thread, &bdev->wq); 1561 1562 while (ttm_bo_delayed_delete(bdev, true)) 1563 ; 1564 1565 mtx_lock(&glob->lru_lock); 1566 if (list_empty(&bdev->ddestroy)) 1567 TTM_DEBUG("Delayed destroy list was clean\n"); 1568 1569 if (list_empty(&bdev->man[0].lru)) 1570 TTM_DEBUG("Swap list was clean\n"); 1571 mtx_unlock(&glob->lru_lock); 1572 1573 MPASS(drm_mm_clean(&bdev->addr_space_mm)); 1574 rw_wlock(&bdev->vm_lock); 1575 drm_mm_takedown(&bdev->addr_space_mm); 1576 rw_wunlock(&bdev->vm_lock); 1577 1578 return ret; 1579} 1580 1581int ttm_bo_device_init(struct ttm_bo_device *bdev, 1582 struct ttm_bo_global *glob, 1583 struct ttm_bo_driver *driver, 1584 uint64_t file_page_offset, 1585 bool need_dma32) 1586{ 1587 int ret = -EINVAL; 1588 1589 rw_init(&bdev->vm_lock, "ttmvml"); 1590 bdev->driver = driver; 1591 1592 memset(bdev->man, 0, sizeof(bdev->man)); 1593 1594 /* 1595 * Initialize the system memory buffer type. 1596 * Other types need to be driver / IOCTL initialized. 1597 */ 1598 ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0); 1599 if (unlikely(ret != 0)) 1600 goto out_no_sys; 1601 1602 RB_INIT(&bdev->addr_space_rb); 1603 ret = drm_mm_init(&bdev->addr_space_mm, file_page_offset, 0x10000000); 1604 if (unlikely(ret != 0)) 1605 goto out_no_addr_mm; 1606 1607 TIMEOUT_TASK_INIT(taskqueue_thread, &bdev->wq, 0, 1608 ttm_bo_delayed_workqueue, bdev); 1609 INIT_LIST_HEAD(&bdev->ddestroy); 1610 bdev->dev_mapping = NULL; 1611 bdev->glob = glob; 1612 bdev->need_dma32 = need_dma32; 1613 bdev->val_seq = 0; 1614 mtx_init(&bdev->fence_lock, "ttmfence", NULL, MTX_DEF); 1615 sx_xlock(&glob->device_list_mutex); 1616 list_add_tail(&bdev->device_list, &glob->device_list); 1617 sx_xunlock(&glob->device_list_mutex); 1618 1619 return 0; 1620out_no_addr_mm: 1621 ttm_bo_clean_mm(bdev, 0); 1622out_no_sys: 1623 return ret; 1624} 1625 1626/* 1627 * buffer object vm functions. 1628 */ 1629 1630bool ttm_mem_reg_is_pci(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem) 1631{ 1632 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type]; 1633 1634 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) { 1635 if (mem->mem_type == TTM_PL_SYSTEM) 1636 return false; 1637 1638 if (man->flags & TTM_MEMTYPE_FLAG_CMA) 1639 return false; 1640 1641 if (mem->placement & TTM_PL_FLAG_CACHED) 1642 return false; 1643 } 1644 return true; 1645} 1646 1647void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo) 1648{ 1649 1650 ttm_bo_release_mmap(bo); 1651 ttm_mem_io_free_vm(bo); 1652} 1653 1654void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo) 1655{ 1656 struct ttm_bo_device *bdev = bo->bdev; 1657 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type]; 1658 1659 ttm_mem_io_lock(man, false); 1660 ttm_bo_unmap_virtual_locked(bo); 1661 ttm_mem_io_unlock(man); 1662} 1663 1664static void ttm_bo_vm_insert_rb(struct ttm_buffer_object *bo) 1665{ 1666 struct ttm_bo_device *bdev = bo->bdev; 1667 1668 /* The caller acquired bdev->vm_lock. */ 1669 RB_INSERT(ttm_bo_device_buffer_objects, &bdev->addr_space_rb, bo); 1670} 1671 1672/** 1673 * ttm_bo_setup_vm: 1674 * 1675 * @bo: the buffer to allocate address space for 1676 * 1677 * Allocate address space in the drm device so that applications 1678 * can mmap the buffer and access the contents. This only 1679 * applies to ttm_bo_type_device objects as others are not 1680 * placed in the drm device address space. 1681 */ 1682 1683static int ttm_bo_setup_vm(struct ttm_buffer_object *bo) 1684{ 1685 struct ttm_bo_device *bdev = bo->bdev; 1686 int ret; 1687 1688retry_pre_get: 1689 ret = drm_mm_pre_get(&bdev->addr_space_mm); 1690 if (unlikely(ret != 0)) 1691 return ret; 1692 1693 rw_wlock(&bdev->vm_lock); 1694 bo->vm_node = drm_mm_search_free(&bdev->addr_space_mm, 1695 bo->mem.num_pages, 0, 0); 1696 1697 if (unlikely(bo->vm_node == NULL)) { 1698 ret = -ENOMEM; 1699 goto out_unlock; 1700 } 1701 1702 bo->vm_node = drm_mm_get_block_atomic(bo->vm_node, 1703 bo->mem.num_pages, 0); 1704 1705 if (unlikely(bo->vm_node == NULL)) { 1706 rw_wunlock(&bdev->vm_lock); 1707 goto retry_pre_get; 1708 } 1709 1710 ttm_bo_vm_insert_rb(bo); 1711 rw_wunlock(&bdev->vm_lock); 1712 bo->addr_space_offset = ((uint64_t) bo->vm_node->start) << PAGE_SHIFT; 1713 1714 return 0; 1715out_unlock: 1716 rw_wunlock(&bdev->vm_lock); 1717 return ret; 1718} 1719 1720int ttm_bo_wait(struct ttm_buffer_object *bo, 1721 bool lazy, bool interruptible, bool no_wait) 1722{ 1723 struct ttm_bo_driver *driver = bo->bdev->driver; 1724 struct ttm_bo_device *bdev = bo->bdev; 1725 void *sync_obj; 1726 int ret = 0; 1727 1728 if (likely(bo->sync_obj == NULL)) 1729 return 0; 1730 1731 while (bo->sync_obj) { 1732 1733 if (driver->sync_obj_signaled(bo->sync_obj)) { 1734 void *tmp_obj = bo->sync_obj; 1735 bo->sync_obj = NULL; 1736 clear_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags); 1737 mtx_unlock(&bdev->fence_lock); 1738 driver->sync_obj_unref(&tmp_obj); 1739 mtx_lock(&bdev->fence_lock); 1740 continue; 1741 } 1742 1743 if (no_wait) 1744 return -EBUSY; 1745 1746 sync_obj = driver->sync_obj_ref(bo->sync_obj); 1747 mtx_unlock(&bdev->fence_lock); 1748 ret = driver->sync_obj_wait(sync_obj, 1749 lazy, interruptible); 1750 if (unlikely(ret != 0)) { 1751 driver->sync_obj_unref(&sync_obj); 1752 mtx_lock(&bdev->fence_lock); 1753 return ret; 1754 } 1755 mtx_lock(&bdev->fence_lock); 1756 if (likely(bo->sync_obj == sync_obj)) { 1757 void *tmp_obj = bo->sync_obj; 1758 bo->sync_obj = NULL; 1759 clear_bit(TTM_BO_PRIV_FLAG_MOVING, 1760 &bo->priv_flags); 1761 mtx_unlock(&bdev->fence_lock); 1762 driver->sync_obj_unref(&sync_obj); 1763 driver->sync_obj_unref(&tmp_obj); 1764 mtx_lock(&bdev->fence_lock); 1765 } else { 1766 mtx_unlock(&bdev->fence_lock); 1767 driver->sync_obj_unref(&sync_obj); 1768 mtx_lock(&bdev->fence_lock); 1769 } 1770 } 1771 return 0; 1772} 1773 1774int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait) 1775{ 1776 struct ttm_bo_device *bdev = bo->bdev; 1777 int ret = 0; 1778 1779 /* 1780 * Using ttm_bo_reserve makes sure the lru lists are updated. 1781 */ 1782 1783 ret = ttm_bo_reserve(bo, true, no_wait, false, 0); 1784 if (unlikely(ret != 0)) 1785 return ret; 1786 mtx_lock(&bdev->fence_lock); 1787 ret = ttm_bo_wait(bo, false, true, no_wait); 1788 mtx_unlock(&bdev->fence_lock); 1789 if (likely(ret == 0)) 1790 atomic_inc(&bo->cpu_writers); 1791 ttm_bo_unreserve(bo); 1792 return ret; 1793} 1794 1795void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo) 1796{ 1797 atomic_dec(&bo->cpu_writers); 1798} 1799 1800/** 1801 * A buffer object shrink method that tries to swap out the first 1802 * buffer object on the bo_global::swap_lru list. 1803 */ 1804 1805static int ttm_bo_swapout(struct ttm_mem_shrink *shrink) 1806{ 1807 struct ttm_bo_global *glob = 1808 container_of(shrink, struct ttm_bo_global, shrink); 1809 struct ttm_buffer_object *bo; 1810 int ret = -EBUSY; 1811 int put_count; 1812 uint32_t swap_placement = (TTM_PL_FLAG_CACHED | TTM_PL_FLAG_SYSTEM); 1813 1814 mtx_lock(&glob->lru_lock); 1815 list_for_each_entry(bo, &glob->swap_lru, swap) { 1816 ret = ttm_bo_reserve_nolru(bo, false, true, false, 0); 1817 if (!ret) 1818 break; 1819 } 1820 1821 if (ret) { 1822 mtx_unlock(&glob->lru_lock); 1823 return ret; 1824 } 1825 1826 refcount_acquire(&bo->list_kref); 1827 1828 if (!list_empty(&bo->ddestroy)) { 1829 ret = ttm_bo_cleanup_refs_and_unlock(bo, false, false); 1830 if (refcount_release(&bo->list_kref)) 1831 ttm_bo_release_list(bo); 1832 return ret; 1833 } 1834 1835 put_count = ttm_bo_del_from_lru(bo); 1836 mtx_unlock(&glob->lru_lock); 1837 1838 ttm_bo_list_ref_sub(bo, put_count, true); 1839 1840 /** 1841 * Wait for GPU, then move to system cached. 1842 */ 1843 1844 mtx_lock(&bo->bdev->fence_lock); 1845 ret = ttm_bo_wait(bo, false, false, false); 1846 mtx_unlock(&bo->bdev->fence_lock); 1847 1848 if (unlikely(ret != 0)) 1849 goto out; 1850 1851 if ((bo->mem.placement & swap_placement) != swap_placement) { 1852 struct ttm_mem_reg evict_mem; 1853 1854 evict_mem = bo->mem; 1855 evict_mem.mm_node = NULL; 1856 evict_mem.placement = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED; 1857 evict_mem.mem_type = TTM_PL_SYSTEM; 1858 1859 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, 1860 false, false); 1861 if (unlikely(ret != 0)) 1862 goto out; 1863 } 1864 1865 ttm_bo_unmap_virtual(bo); 1866 1867 /** 1868 * Swap out. Buffer will be swapped in again as soon as 1869 * anyone tries to access a ttm page. 1870 */ 1871 1872 if (bo->bdev->driver->swap_notify) 1873 bo->bdev->driver->swap_notify(bo); 1874 1875 ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage); 1876out: 1877 1878 /** 1879 * 1880 * Unreserve without putting on LRU to avoid swapping out an 1881 * already swapped buffer. 1882 */ 1883 1884 atomic_set(&bo->reserved, 0); 1885 wakeup(bo); 1886 if (refcount_release(&bo->list_kref)) 1887 ttm_bo_release_list(bo); 1888 return ret; 1889} 1890 1891void ttm_bo_swapout_all(struct ttm_bo_device *bdev) 1892{ 1893 while (ttm_bo_swapout(&bdev->glob->shrink) == 0) 1894 ; 1895} 1896