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#include <dev/drm2/drmP.h> 33#include <dev/drm2/ttm/ttm_module.h> 34#include <dev/drm2/ttm/ttm_bo_driver.h> 35#include <dev/drm2/ttm/ttm_placement.h> 36#include <vm/vm_pageout.h> 37 38#define TTM_ASSERT_LOCKED(param) 39#define TTM_DEBUG(fmt, arg...) 40#define TTM_BO_HASH_ORDER 13 41 42static int ttm_bo_setup_vm(struct ttm_buffer_object *bo); 43static int ttm_bo_swapout(struct ttm_mem_shrink *shrink); 44static void ttm_bo_global_kobj_release(struct ttm_bo_global *glob); 45 46MALLOC_DEFINE(M_TTM_BO, "ttm_bo", "TTM Buffer Objects"); 47 48static inline int ttm_mem_type_from_flags(uint32_t flags, uint32_t *mem_type) 49{ 50 int i; 51 52 for (i = 0; i <= TTM_PL_PRIV5; i++) 53 if (flags & (1 << i)) { 54 *mem_type = i; 55 return 0; 56 } 57 return -EINVAL; 58} 59 60static void ttm_mem_type_debug(struct ttm_bo_device *bdev, int mem_type) 61{ 62 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 63 64 printf(" has_type: %d\n", man->has_type); 65 printf(" use_type: %d\n", man->use_type); 66 printf(" flags: 0x%08X\n", man->flags); 67 printf(" gpu_offset: 0x%08lX\n", man->gpu_offset); 68 printf(" size: %ju\n", (uintmax_t)man->size); 69 printf(" available_caching: 0x%08X\n", man->available_caching); 70 printf(" default_caching: 0x%08X\n", man->default_caching); 71 if (mem_type != TTM_PL_SYSTEM) 72 (*man->func->debug)(man, TTM_PFX); 73} 74 75static void ttm_bo_mem_space_debug(struct ttm_buffer_object *bo, 76 struct ttm_placement *placement) 77{ 78 int i, ret, mem_type; 79 80 printf("No space for %p (%lu pages, %luK, %luM)\n", 81 bo, bo->mem.num_pages, bo->mem.size >> 10, 82 bo->mem.size >> 20); 83 for (i = 0; i < placement->num_placement; i++) { 84 ret = ttm_mem_type_from_flags(placement->placement[i], 85 &mem_type); 86 if (ret) 87 return; 88 printf(" placement[%d]=0x%08X (%d)\n", 89 i, placement->placement[i], mem_type); 90 ttm_mem_type_debug(bo->bdev, mem_type); 91 } 92} 93 94#if 0 95static ssize_t ttm_bo_global_show(struct ttm_bo_global *glob, 96 char *buffer) 97{ 98 99 return snprintf(buffer, PAGE_SIZE, "%lu\n", 100 (unsigned long) atomic_read(&glob->bo_count)); 101} 102#endif 103 104static inline uint32_t ttm_bo_type_flags(unsigned type) 105{ 106 return 1 << (type); 107} 108 109static void ttm_bo_release_list(struct ttm_buffer_object *bo) 110{ 111 struct ttm_bo_device *bdev = bo->bdev; 112 size_t acc_size = bo->acc_size; 113 114 MPASS(atomic_read(&bo->list_kref) == 0); 115 MPASS(atomic_read(&bo->kref) == 0); 116 MPASS(atomic_read(&bo->cpu_writers) == 0); 117 MPASS(bo->sync_obj == NULL); 118 MPASS(bo->mem.mm_node == NULL); 119 MPASS(list_empty(&bo->lru)); 120 MPASS(list_empty(&bo->ddestroy)); 121 122 if (bo->ttm) 123 ttm_tt_destroy(bo->ttm); 124 atomic_dec(&bo->glob->bo_count); 125 if (bo->destroy) 126 bo->destroy(bo); 127 else { 128 free(bo, M_TTM_BO); 129 } 130 ttm_mem_global_free(bdev->glob->mem_glob, acc_size); 131} 132 133static int 134ttm_bo_wait_unreserved_locked(struct ttm_buffer_object *bo, bool interruptible) 135{ 136 const char *wmsg; 137 int flags, ret; 138 139 ret = 0; 140 if (interruptible) { 141 flags = PCATCH; 142 wmsg = "ttbowi"; 143 } else { 144 flags = 0; 145 wmsg = "ttbowu"; 146 } 147 while (ttm_bo_is_reserved(bo)) { 148 ret = -msleep(bo, &bo->glob->lru_lock, flags, wmsg, 0); 149 if (ret == -EINTR || ret == -ERESTART) 150 ret = -ERESTARTSYS; 151 if (ret != 0) 152 break; 153 } 154 return (ret); 155} 156 157void ttm_bo_add_to_lru(struct ttm_buffer_object *bo) 158{ 159 struct ttm_bo_device *bdev = bo->bdev; 160 struct ttm_mem_type_manager *man; 161 162 MPASS(ttm_bo_is_reserved(bo)); 163 164 if (!(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) { 165 166 MPASS(list_empty(&bo->lru)); 167 168 man = &bdev->man[bo->mem.mem_type]; 169 list_add_tail(&bo->lru, &man->lru); 170 refcount_acquire(&bo->list_kref); 171 172 if (bo->ttm != NULL) { 173 list_add_tail(&bo->swap, &bo->glob->swap_lru); 174 refcount_acquire(&bo->list_kref); 175 } 176 } 177} 178 179int ttm_bo_del_from_lru(struct ttm_buffer_object *bo) 180{ 181 int put_count = 0; 182 183 if (!list_empty(&bo->swap)) { 184 list_del_init(&bo->swap); 185 ++put_count; 186 } 187 if (!list_empty(&bo->lru)) { 188 list_del_init(&bo->lru); 189 ++put_count; 190 } 191 192 /* 193 * TODO: Add a driver hook to delete from 194 * driver-specific LRU's here. 195 */ 196 197 return put_count; 198} 199 200int ttm_bo_reserve_nolru(struct ttm_buffer_object *bo, 201 bool interruptible, 202 bool no_wait, bool use_sequence, uint32_t sequence) 203{ 204 int ret; 205 206 while (unlikely(atomic_xchg(&bo->reserved, 1) != 0)) { 207 /** 208 * Deadlock avoidance for multi-bo reserving. 209 */ 210 if (use_sequence && bo->seq_valid) { 211 /** 212 * We've already reserved this one. 213 */ 214 if (unlikely(sequence == bo->val_seq)) 215 return -EDEADLK; 216 /** 217 * Already reserved by a thread that will not back 218 * off for us. We need to back off. 219 */ 220 if (unlikely(sequence - bo->val_seq < (1U << 31))) 221 return -EAGAIN; 222 } 223 224 if (no_wait) 225 return -EBUSY; 226 227 ret = ttm_bo_wait_unreserved_locked(bo, interruptible); 228 229 if (unlikely(ret)) 230 return ret; 231 } 232 233 if (use_sequence) { 234 bool wake_up = false; 235 /** 236 * Wake up waiters that may need to recheck for deadlock, 237 * if we decreased the sequence number. 238 */ 239 if (unlikely((bo->val_seq - sequence < (1U << 31)) 240 || !bo->seq_valid)) 241 wake_up = true; 242 243 /* 244 * In the worst case with memory ordering these values can be 245 * seen in the wrong order. However since we call wake_up_all 246 * in that case, this will hopefully not pose a problem, 247 * and the worst case would only cause someone to accidentally 248 * hit -EAGAIN in ttm_bo_reserve when they see old value of 249 * val_seq. However this would only happen if seq_valid was 250 * written before val_seq was, and just means some slightly 251 * increased cpu usage 252 */ 253 bo->val_seq = sequence; 254 bo->seq_valid = true; 255 if (wake_up) 256 wakeup(bo); 257 } else { 258 bo->seq_valid = false; 259 } 260 261 return 0; 262} 263 264void ttm_bo_list_ref_sub(struct ttm_buffer_object *bo, int count, 265 bool never_free) 266{ 267 u_int old; 268 269 old = atomic_fetchadd_int(&bo->list_kref, -count); 270 if (old <= count) { 271 if (never_free) 272 panic("ttm_bo_ref_buf"); 273 ttm_bo_release_list(bo); 274 } 275} 276 277int ttm_bo_reserve(struct ttm_buffer_object *bo, 278 bool interruptible, 279 bool no_wait, bool use_sequence, uint32_t sequence) 280{ 281 struct ttm_bo_global *glob = bo->glob; 282 int put_count = 0; 283 int ret; 284 285 mtx_lock(&bo->glob->lru_lock); 286 ret = ttm_bo_reserve_nolru(bo, interruptible, no_wait, use_sequence, 287 sequence); 288 if (likely(ret == 0)) { 289 put_count = ttm_bo_del_from_lru(bo); 290 mtx_unlock(&glob->lru_lock); 291 ttm_bo_list_ref_sub(bo, put_count, true); 292 } else 293 mtx_unlock(&bo->glob->lru_lock); 294 295 return ret; 296} 297 298int ttm_bo_reserve_slowpath_nolru(struct ttm_buffer_object *bo, 299 bool interruptible, uint32_t sequence) 300{ 301 bool wake_up = false; 302 int ret; 303 304 while (unlikely(atomic_xchg(&bo->reserved, 1) != 0)) { 305 if (bo->seq_valid && sequence == bo->val_seq) { 306 DRM_ERROR( 307 "%s: bo->seq_valid && sequence == bo->val_seq", 308 __func__); 309 } 310 311 ret = ttm_bo_wait_unreserved_locked(bo, interruptible); 312 313 if (unlikely(ret)) 314 return ret; 315 } 316 317 if ((bo->val_seq - sequence < (1U << 31)) || !bo->seq_valid) 318 wake_up = true; 319 320 /** 321 * Wake up waiters that may need to recheck for deadlock, 322 * if we decreased the sequence number. 323 */ 324 bo->val_seq = sequence; 325 bo->seq_valid = true; 326 if (wake_up) 327 wakeup(bo); 328 329 return 0; 330} 331 332int ttm_bo_reserve_slowpath(struct ttm_buffer_object *bo, 333 bool interruptible, uint32_t sequence) 334{ 335 struct ttm_bo_global *glob = bo->glob; 336 int put_count, ret; 337 338 mtx_lock(&glob->lru_lock); 339 ret = ttm_bo_reserve_slowpath_nolru(bo, interruptible, sequence); 340 if (likely(!ret)) { 341 put_count = ttm_bo_del_from_lru(bo); 342 mtx_unlock(&glob->lru_lock); 343 ttm_bo_list_ref_sub(bo, put_count, true); 344 } else 345 mtx_unlock(&glob->lru_lock); 346 return ret; 347} 348 349void ttm_bo_unreserve_locked(struct ttm_buffer_object *bo) 350{ 351 ttm_bo_add_to_lru(bo); 352 atomic_set(&bo->reserved, 0); 353 wakeup(bo); 354} 355 356void ttm_bo_unreserve(struct ttm_buffer_object *bo) 357{ 358 struct ttm_bo_global *glob = bo->glob; 359 360 mtx_lock(&glob->lru_lock); 361 ttm_bo_unreserve_locked(bo); 362 mtx_unlock(&glob->lru_lock); 363} 364 365/* 366 * Call bo->mutex locked. 367 */ 368static int ttm_bo_add_ttm(struct ttm_buffer_object *bo, bool zero_alloc) 369{ 370 struct ttm_bo_device *bdev = bo->bdev; 371 struct ttm_bo_global *glob = bo->glob; 372 int ret = 0; 373 uint32_t page_flags = 0; 374 375 TTM_ASSERT_LOCKED(&bo->mutex); 376 bo->ttm = NULL; 377 378 if (bdev->need_dma32) 379 page_flags |= TTM_PAGE_FLAG_DMA32; 380 381 switch (bo->type) { 382 case ttm_bo_type_device: 383 if (zero_alloc) 384 page_flags |= TTM_PAGE_FLAG_ZERO_ALLOC; 385 case ttm_bo_type_kernel: 386 bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT, 387 page_flags, glob->dummy_read_page); 388 if (unlikely(bo->ttm == NULL)) 389 ret = -ENOMEM; 390 break; 391 case ttm_bo_type_sg: 392 bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT, 393 page_flags | TTM_PAGE_FLAG_SG, 394 glob->dummy_read_page); 395 if (unlikely(bo->ttm == NULL)) { 396 ret = -ENOMEM; 397 break; 398 } 399 bo->ttm->sg = bo->sg; 400 break; 401 default: 402 printf("[TTM] Illegal buffer object type\n"); 403 ret = -EINVAL; 404 break; 405 } 406 407 return ret; 408} 409 410static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo, 411 struct ttm_mem_reg *mem, 412 bool evict, bool interruptible, 413 bool no_wait_gpu) 414{ 415 struct ttm_bo_device *bdev = bo->bdev; 416 bool old_is_pci = ttm_mem_reg_is_pci(bdev, &bo->mem); 417 bool new_is_pci = ttm_mem_reg_is_pci(bdev, mem); 418 struct ttm_mem_type_manager *old_man = &bdev->man[bo->mem.mem_type]; 419 struct ttm_mem_type_manager *new_man = &bdev->man[mem->mem_type]; 420 int ret = 0; 421 422 if (old_is_pci || new_is_pci || 423 ((mem->placement & bo->mem.placement & TTM_PL_MASK_CACHING) == 0)) { 424 ret = ttm_mem_io_lock(old_man, true); 425 if (unlikely(ret != 0)) 426 goto out_err; 427 ttm_bo_unmap_virtual_locked(bo); 428 ttm_mem_io_unlock(old_man); 429 } 430 431 /* 432 * Create and bind a ttm if required. 433 */ 434 435 if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) { 436 if (bo->ttm == NULL) { 437 bool zero = !(old_man->flags & TTM_MEMTYPE_FLAG_FIXED); 438 ret = ttm_bo_add_ttm(bo, zero); 439 if (ret) 440 goto out_err; 441 } 442 443 ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement); 444 if (ret) 445 goto out_err; 446 447 if (mem->mem_type != TTM_PL_SYSTEM) { 448 ret = ttm_tt_bind(bo->ttm, mem); 449 if (ret) 450 goto out_err; 451 } 452 453 if (bo->mem.mem_type == TTM_PL_SYSTEM) { 454 if (bdev->driver->move_notify) 455 bdev->driver->move_notify(bo, mem); 456 bo->mem = *mem; 457 mem->mm_node = NULL; 458 goto moved; 459 } 460 } 461 462 if (bdev->driver->move_notify) 463 bdev->driver->move_notify(bo, mem); 464 465 if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) && 466 !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) 467 ret = ttm_bo_move_ttm(bo, evict, no_wait_gpu, mem); 468 else if (bdev->driver->move) 469 ret = bdev->driver->move(bo, evict, interruptible, 470 no_wait_gpu, mem); 471 else 472 ret = ttm_bo_move_memcpy(bo, evict, no_wait_gpu, mem); 473 474 if (ret) { 475 if (bdev->driver->move_notify) { 476 struct ttm_mem_reg tmp_mem = *mem; 477 *mem = bo->mem; 478 bo->mem = tmp_mem; 479 bdev->driver->move_notify(bo, mem); 480 bo->mem = *mem; 481 *mem = tmp_mem; 482 } 483 484 goto out_err; 485 } 486 487moved: 488 if (bo->evicted) { 489 ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement); 490 if (ret) 491 printf("[TTM] Can not flush read caches\n"); 492 bo->evicted = false; 493 } 494 495 if (bo->mem.mm_node) { 496 bo->offset = (bo->mem.start << PAGE_SHIFT) + 497 bdev->man[bo->mem.mem_type].gpu_offset; 498 bo->cur_placement = bo->mem.placement; 499 } else 500 bo->offset = 0; 501 502 return 0; 503 504out_err: 505 new_man = &bdev->man[bo->mem.mem_type]; 506 if ((new_man->flags & TTM_MEMTYPE_FLAG_FIXED) && bo->ttm) { 507 ttm_tt_unbind(bo->ttm); 508 ttm_tt_destroy(bo->ttm); 509 bo->ttm = NULL; 510 } 511 512 return ret; 513} 514 515/** 516 * Call bo::reserved. 517 * Will release GPU memory type usage on destruction. 518 * This is the place to put in driver specific hooks to release 519 * driver private resources. 520 * Will release the bo::reserved lock. 521 */ 522 523static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo) 524{ 525 if (bo->bdev->driver->move_notify) 526 bo->bdev->driver->move_notify(bo, NULL); 527 528 if (bo->ttm) { 529 ttm_tt_unbind(bo->ttm); 530 ttm_tt_destroy(bo->ttm); 531 bo->ttm = NULL; 532 } 533 ttm_bo_mem_put(bo, &bo->mem); 534 535 atomic_set(&bo->reserved, 0); 536 wakeup(&bo); 537 538 /* 539 * Since the final reference to this bo may not be dropped by 540 * the current task we have to put a memory barrier here to make 541 * sure the changes done in this function are always visible. 542 * 543 * This function only needs protection against the final kref_put. 544 */ 545 mb(); 546} 547 548static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo) 549{ 550 struct ttm_bo_device *bdev = bo->bdev; 551 struct ttm_bo_global *glob = bo->glob; 552 struct ttm_bo_driver *driver = bdev->driver; 553 void *sync_obj = NULL; 554 int put_count; 555 int ret; 556 557 mtx_lock(&glob->lru_lock); 558 ret = ttm_bo_reserve_nolru(bo, false, true, false, 0); 559 560 mtx_lock(&bdev->fence_lock); 561 (void) ttm_bo_wait(bo, false, false, true); 562 if (!ret && !bo->sync_obj) { 563 mtx_unlock(&bdev->fence_lock); 564 put_count = ttm_bo_del_from_lru(bo); 565 566 mtx_unlock(&glob->lru_lock); 567 ttm_bo_cleanup_memtype_use(bo); 568 569 ttm_bo_list_ref_sub(bo, put_count, true); 570 571 return; 572 } 573 if (bo->sync_obj) 574 sync_obj = driver->sync_obj_ref(bo->sync_obj); 575 mtx_unlock(&bdev->fence_lock); 576 577 if (!ret) { 578 atomic_set(&bo->reserved, 0); 579 wakeup(bo); 580 } 581 582 refcount_acquire(&bo->list_kref); 583 list_add_tail(&bo->ddestroy, &bdev->ddestroy); 584 mtx_unlock(&glob->lru_lock); 585 586 if (sync_obj) { 587 driver->sync_obj_flush(sync_obj); 588 driver->sync_obj_unref(&sync_obj); 589 } 590 taskqueue_enqueue_timeout(taskqueue_thread, &bdev->wq, 591 ((hz / 100) < 1) ? 1 : hz / 100); 592} 593 594/** 595 * function ttm_bo_cleanup_refs_and_unlock 596 * If bo idle, remove from delayed- and lru lists, and unref. 597 * If not idle, do nothing. 598 * 599 * Must be called with lru_lock and reservation held, this function 600 * will drop both before returning. 601 * 602 * @interruptible Any sleeps should occur interruptibly. 603 * @no_wait_gpu Never wait for gpu. Return -EBUSY instead. 604 */ 605 606static int ttm_bo_cleanup_refs_and_unlock(struct ttm_buffer_object *bo, 607 bool interruptible, 608 bool no_wait_gpu) 609{ 610 struct ttm_bo_device *bdev = bo->bdev; 611 struct ttm_bo_driver *driver = bdev->driver; 612 struct ttm_bo_global *glob = bo->glob; 613 int put_count; 614 int ret; 615 616 mtx_lock(&bdev->fence_lock); 617 ret = ttm_bo_wait(bo, false, false, true); 618 619 if (ret && !no_wait_gpu) { 620 void *sync_obj; 621 622 /* 623 * Take a reference to the fence and unreserve, 624 * at this point the buffer should be dead, so 625 * no new sync objects can be attached. 626 */ 627 sync_obj = driver->sync_obj_ref(bo->sync_obj); 628 mtx_unlock(&bdev->fence_lock); 629 630 atomic_set(&bo->reserved, 0); 631 wakeup(bo); 632 mtx_unlock(&glob->lru_lock); 633 634 ret = driver->sync_obj_wait(sync_obj, false, interruptible); 635 driver->sync_obj_unref(&sync_obj); 636 if (ret) 637 return ret; 638 639 /* 640 * remove sync_obj with ttm_bo_wait, the wait should be 641 * finished, and no new wait object should have been added. 642 */ 643 mtx_lock(&bdev->fence_lock); 644 ret = ttm_bo_wait(bo, false, false, true); 645 mtx_unlock(&bdev->fence_lock); 646 if (ret) 647 return ret; 648 649 mtx_lock(&glob->lru_lock); 650 ret = ttm_bo_reserve_nolru(bo, false, true, false, 0); 651 652 /* 653 * We raced, and lost, someone else holds the reservation now, 654 * and is probably busy in ttm_bo_cleanup_memtype_use. 655 * 656 * Even if it's not the case, because we finished waiting any 657 * delayed destruction would succeed, so just return success 658 * here. 659 */ 660 if (ret) { 661 mtx_unlock(&glob->lru_lock); 662 return 0; 663 } 664 } else 665 mtx_unlock(&bdev->fence_lock); 666 667 if (ret || unlikely(list_empty(&bo->ddestroy))) { 668 atomic_set(&bo->reserved, 0); 669 wakeup(bo); 670 mtx_unlock(&glob->lru_lock); 671 return ret; 672 } 673 674 put_count = ttm_bo_del_from_lru(bo); 675 list_del_init(&bo->ddestroy); 676 ++put_count; 677 678 mtx_unlock(&glob->lru_lock); 679 ttm_bo_cleanup_memtype_use(bo); 680 681 ttm_bo_list_ref_sub(bo, put_count, true); 682 683 return 0; 684} 685 686/** 687 * Traverse the delayed list, and call ttm_bo_cleanup_refs on all 688 * encountered buffers. 689 */ 690 691static int ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all) 692{ 693 struct ttm_bo_global *glob = bdev->glob; 694 struct ttm_buffer_object *entry = NULL; 695 int ret = 0; 696 697 mtx_lock(&glob->lru_lock); 698 if (list_empty(&bdev->ddestroy)) 699 goto out_unlock; 700 701 entry = list_first_entry(&bdev->ddestroy, 702 struct ttm_buffer_object, ddestroy); 703 refcount_acquire(&entry->list_kref); 704 705 for (;;) { 706 struct ttm_buffer_object *nentry = NULL; 707 708 if (entry->ddestroy.next != &bdev->ddestroy) { 709 nentry = list_first_entry(&entry->ddestroy, 710 struct ttm_buffer_object, ddestroy); 711 refcount_acquire(&nentry->list_kref); 712 } 713 714 ret = ttm_bo_reserve_nolru(entry, false, true, false, 0); 715 if (remove_all && ret) { 716 ret = ttm_bo_reserve_nolru(entry, false, false, 717 false, 0); 718 } 719 720 if (!ret) 721 ret = ttm_bo_cleanup_refs_and_unlock(entry, false, 722 !remove_all); 723 else 724 mtx_unlock(&glob->lru_lock); 725 726 if (refcount_release(&entry->list_kref)) 727 ttm_bo_release_list(entry); 728 entry = nentry; 729 730 if (ret || !entry) 731 goto out; 732 733 mtx_lock(&glob->lru_lock); 734 if (list_empty(&entry->ddestroy)) 735 break; 736 } 737 738out_unlock: 739 mtx_unlock(&glob->lru_lock); 740out: 741 if (entry && refcount_release(&entry->list_kref)) 742 ttm_bo_release_list(entry); 743 return ret; 744} 745 746static void ttm_bo_delayed_workqueue(void *arg, int pending __unused) 747{ 748 struct ttm_bo_device *bdev = arg; 749 750 if (ttm_bo_delayed_delete(bdev, false)) { 751 taskqueue_enqueue_timeout(taskqueue_thread, &bdev->wq, 752 ((hz / 100) < 1) ? 1 : hz / 100); 753 } 754} 755 756static void ttm_bo_release(struct ttm_buffer_object *bo) 757{ 758 struct ttm_bo_device *bdev = bo->bdev; 759 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type]; 760 761 rw_wlock(&bdev->vm_lock); 762 if (likely(bo->vm_node != NULL)) { 763 RB_REMOVE(ttm_bo_device_buffer_objects, 764 &bdev->addr_space_rb, bo); 765 drm_mm_put_block(bo->vm_node); 766 bo->vm_node = NULL; 767 } 768 rw_wunlock(&bdev->vm_lock); 769 ttm_mem_io_lock(man, false); 770 ttm_mem_io_free_vm(bo); 771 ttm_mem_io_unlock(man); 772 ttm_bo_cleanup_refs_or_queue(bo); 773 if (refcount_release(&bo->list_kref)) 774 ttm_bo_release_list(bo); 775} 776 777void ttm_bo_unref(struct ttm_buffer_object **p_bo) 778{ 779 struct ttm_buffer_object *bo = *p_bo; 780 781 *p_bo = NULL; 782 if (refcount_release(&bo->kref)) 783 ttm_bo_release(bo); 784} 785 786int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev) 787{ 788 int pending; 789 790 if (taskqueue_cancel_timeout(taskqueue_thread, &bdev->wq, &pending)) 791 taskqueue_drain_timeout(taskqueue_thread, &bdev->wq); 792 return (pending); 793} 794 795void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched) 796{ 797 if (resched) { 798 taskqueue_enqueue_timeout(taskqueue_thread, &bdev->wq, 799 ((hz / 100) < 1) ? 1 : hz / 100); 800 } 801} 802 803static int ttm_bo_evict(struct ttm_buffer_object *bo, bool interruptible, 804 bool no_wait_gpu) 805{ 806 struct ttm_bo_device *bdev = bo->bdev; 807 struct ttm_mem_reg evict_mem; 808 struct ttm_placement placement; 809 int ret = 0; 810 811 mtx_lock(&bdev->fence_lock); 812 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu); 813 mtx_unlock(&bdev->fence_lock); 814 815 if (unlikely(ret != 0)) { 816 if (ret != -ERESTARTSYS) { 817 printf("[TTM] Failed to expire sync object before buffer eviction\n"); 818 } 819 goto out; 820 } 821 822 MPASS(ttm_bo_is_reserved(bo)); 823 824 evict_mem = bo->mem; 825 evict_mem.mm_node = NULL; 826 evict_mem.bus.io_reserved_vm = false; 827 evict_mem.bus.io_reserved_count = 0; 828 829 placement.fpfn = 0; 830 placement.lpfn = 0; 831 placement.num_placement = 0; 832 placement.num_busy_placement = 0; 833 bdev->driver->evict_flags(bo, &placement); 834 ret = ttm_bo_mem_space(bo, &placement, &evict_mem, interruptible, 835 no_wait_gpu); 836 if (ret) { 837 if (ret != -ERESTARTSYS) { 838 printf("[TTM] Failed to find memory space for buffer 0x%p eviction\n", 839 bo); 840 ttm_bo_mem_space_debug(bo, &placement); 841 } 842 goto out; 843 } 844 845 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, interruptible, 846 no_wait_gpu); 847 if (ret) { 848 if (ret != -ERESTARTSYS) 849 printf("[TTM] Buffer eviction failed\n"); 850 ttm_bo_mem_put(bo, &evict_mem); 851 goto out; 852 } 853 bo->evicted = true; 854out: 855 return ret; 856} 857 858static int ttm_mem_evict_first(struct ttm_bo_device *bdev, 859 uint32_t mem_type, 860 bool interruptible, 861 bool no_wait_gpu) 862{ 863 struct ttm_bo_global *glob = bdev->glob; 864 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 865 struct ttm_buffer_object *bo; 866 int ret = -EBUSY, put_count; 867 868 mtx_lock(&glob->lru_lock); 869 list_for_each_entry(bo, &man->lru, lru) { 870 ret = ttm_bo_reserve_nolru(bo, false, true, false, 0); 871 if (!ret) 872 break; 873 } 874 875 if (ret) { 876 mtx_unlock(&glob->lru_lock); 877 return ret; 878 } 879 880 refcount_acquire(&bo->list_kref); 881 882 if (!list_empty(&bo->ddestroy)) { 883 ret = ttm_bo_cleanup_refs_and_unlock(bo, interruptible, 884 no_wait_gpu); 885 if (refcount_release(&bo->list_kref)) 886 ttm_bo_release_list(bo); 887 return ret; 888 } 889 890 put_count = ttm_bo_del_from_lru(bo); 891 mtx_unlock(&glob->lru_lock); 892 893 MPASS(ret == 0); 894 895 ttm_bo_list_ref_sub(bo, put_count, true); 896 897 ret = ttm_bo_evict(bo, interruptible, no_wait_gpu); 898 ttm_bo_unreserve(bo); 899 900 if (refcount_release(&bo->list_kref)) 901 ttm_bo_release_list(bo); 902 return ret; 903} 904 905void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem) 906{ 907 struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type]; 908 909 if (mem->mm_node) 910 (*man->func->put_node)(man, mem); 911} 912 913/** 914 * Repeatedly evict memory from the LRU for @mem_type until we create enough 915 * space, or we've evicted everything and there isn't enough space. 916 */ 917static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo, 918 uint32_t mem_type, 919 struct ttm_placement *placement, 920 struct ttm_mem_reg *mem, 921 bool interruptible, 922 bool no_wait_gpu) 923{ 924 struct ttm_bo_device *bdev = bo->bdev; 925 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 926 int ret; 927 928 do { 929 ret = (*man->func->get_node)(man, bo, placement, mem); 930 if (unlikely(ret != 0)) 931 return ret; 932 if (mem->mm_node) 933 break; 934 ret = ttm_mem_evict_first(bdev, mem_type, 935 interruptible, no_wait_gpu); 936 if (unlikely(ret != 0)) 937 return ret; 938 } while (1); 939 if (mem->mm_node == NULL) 940 return -ENOMEM; 941 mem->mem_type = mem_type; 942 return 0; 943} 944 945static uint32_t ttm_bo_select_caching(struct ttm_mem_type_manager *man, 946 uint32_t cur_placement, 947 uint32_t proposed_placement) 948{ 949 uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING; 950 uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING; 951 952 /** 953 * Keep current caching if possible. 954 */ 955 956 if ((cur_placement & caching) != 0) 957 result |= (cur_placement & caching); 958 else if ((man->default_caching & caching) != 0) 959 result |= man->default_caching; 960 else if ((TTM_PL_FLAG_CACHED & caching) != 0) 961 result |= TTM_PL_FLAG_CACHED; 962 else if ((TTM_PL_FLAG_WC & caching) != 0) 963 result |= TTM_PL_FLAG_WC; 964 else if ((TTM_PL_FLAG_UNCACHED & caching) != 0) 965 result |= TTM_PL_FLAG_UNCACHED; 966 967 return result; 968} 969 970static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man, 971 uint32_t mem_type, 972 uint32_t proposed_placement, 973 uint32_t *masked_placement) 974{ 975 uint32_t cur_flags = ttm_bo_type_flags(mem_type); 976 977 if ((cur_flags & proposed_placement & TTM_PL_MASK_MEM) == 0) 978 return false; 979 980 if ((proposed_placement & man->available_caching) == 0) 981 return false; 982 983 cur_flags |= (proposed_placement & man->available_caching); 984 985 *masked_placement = cur_flags; 986 return true; 987} 988 989/** 990 * Creates space for memory region @mem according to its type. 991 * 992 * This function first searches for free space in compatible memory types in 993 * the priority order defined by the driver. If free space isn't found, then 994 * ttm_bo_mem_force_space is attempted in priority order to evict and find 995 * space. 996 */ 997int ttm_bo_mem_space(struct ttm_buffer_object *bo, 998 struct ttm_placement *placement, 999 struct ttm_mem_reg *mem, 1000 bool interruptible, 1001 bool no_wait_gpu) 1002{ 1003 struct ttm_bo_device *bdev = bo->bdev; 1004 struct ttm_mem_type_manager *man; 1005 uint32_t mem_type = TTM_PL_SYSTEM; 1006 uint32_t cur_flags = 0; 1007 bool type_found = false; 1008 bool type_ok = false; 1009 bool has_erestartsys = false; 1010 int i, ret; 1011 1012 mem->mm_node = NULL; 1013 for (i = 0; i < placement->num_placement; ++i) { 1014 ret = ttm_mem_type_from_flags(placement->placement[i], 1015 &mem_type); 1016 if (ret) 1017 return ret; 1018 man = &bdev->man[mem_type]; 1019 1020 type_ok = ttm_bo_mt_compatible(man, 1021 mem_type, 1022 placement->placement[i], 1023 &cur_flags); 1024 1025 if (!type_ok) 1026 continue; 1027 1028 cur_flags = ttm_bo_select_caching(man, bo->mem.placement, 1029 cur_flags); 1030 /* 1031 * Use the access and other non-mapping-related flag bits from 1032 * the memory placement flags to the current flags 1033 */ 1034 ttm_flag_masked(&cur_flags, placement->placement[i], 1035 ~TTM_PL_MASK_MEMTYPE); 1036 1037 if (mem_type == TTM_PL_SYSTEM) 1038 break; 1039 1040 if (man->has_type && man->use_type) { 1041 type_found = true; 1042 ret = (*man->func->get_node)(man, bo, placement, mem); 1043 if (unlikely(ret)) 1044 return ret; 1045 } 1046 if (mem->mm_node) 1047 break; 1048 } 1049 1050 if ((type_ok && (mem_type == TTM_PL_SYSTEM)) || mem->mm_node) { 1051 mem->mem_type = mem_type; 1052 mem->placement = cur_flags; 1053 return 0; 1054 } 1055 1056 if (!type_found) 1057 return -EINVAL; 1058 1059 for (i = 0; i < placement->num_busy_placement; ++i) { 1060 ret = ttm_mem_type_from_flags(placement->busy_placement[i], 1061 &mem_type); 1062 if (ret) 1063 return ret; 1064 man = &bdev->man[mem_type]; 1065 if (!man->has_type) 1066 continue; 1067 if (!ttm_bo_mt_compatible(man, 1068 mem_type, 1069 placement->busy_placement[i], 1070 &cur_flags)) 1071 continue; 1072 1073 cur_flags = ttm_bo_select_caching(man, bo->mem.placement, 1074 cur_flags); 1075 /* 1076 * Use the access and other non-mapping-related flag bits from 1077 * the memory placement flags to the current flags 1078 */ 1079 ttm_flag_masked(&cur_flags, placement->busy_placement[i], 1080 ~TTM_PL_MASK_MEMTYPE); 1081 1082 1083 if (mem_type == TTM_PL_SYSTEM) { 1084 mem->mem_type = mem_type; 1085 mem->placement = cur_flags; 1086 mem->mm_node = NULL; 1087 return 0; 1088 } 1089 1090 ret = ttm_bo_mem_force_space(bo, mem_type, placement, mem, 1091 interruptible, no_wait_gpu); 1092 if (ret == 0 && mem->mm_node) { 1093 mem->placement = cur_flags; 1094 return 0; 1095 } 1096 if (ret == -ERESTARTSYS) 1097 has_erestartsys = true; 1098 } 1099 ret = (has_erestartsys) ? -ERESTARTSYS : -ENOMEM; 1100 return ret; 1101} 1102 1103static 1104int ttm_bo_move_buffer(struct ttm_buffer_object *bo, 1105 struct ttm_placement *placement, 1106 bool interruptible, 1107 bool no_wait_gpu) 1108{ 1109 int ret = 0; 1110 struct ttm_mem_reg mem; 1111 struct ttm_bo_device *bdev = bo->bdev; 1112 1113 MPASS(ttm_bo_is_reserved(bo)); 1114 1115 /* 1116 * FIXME: It's possible to pipeline buffer moves. 1117 * Have the driver move function wait for idle when necessary, 1118 * instead of doing it here. 1119 */ 1120 mtx_lock(&bdev->fence_lock); 1121 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu); 1122 mtx_unlock(&bdev->fence_lock); 1123 if (ret) 1124 return ret; 1125 mem.num_pages = bo->num_pages; 1126 mem.size = mem.num_pages << PAGE_SHIFT; 1127 mem.page_alignment = bo->mem.page_alignment; 1128 mem.bus.io_reserved_vm = false; 1129 mem.bus.io_reserved_count = 0; 1130 /* 1131 * Determine where to move the buffer. 1132 */ 1133 ret = ttm_bo_mem_space(bo, placement, &mem, 1134 interruptible, no_wait_gpu); 1135 if (ret) 1136 goto out_unlock; 1137 ret = ttm_bo_handle_move_mem(bo, &mem, false, 1138 interruptible, no_wait_gpu); 1139out_unlock: 1140 if (ret && mem.mm_node) 1141 ttm_bo_mem_put(bo, &mem); 1142 return ret; 1143} 1144 1145static int ttm_bo_mem_compat(struct ttm_placement *placement, 1146 struct ttm_mem_reg *mem) 1147{ 1148 int i; 1149 1150 if (mem->mm_node && placement->lpfn != 0 && 1151 (mem->start < placement->fpfn || 1152 mem->start + mem->num_pages > placement->lpfn)) 1153 return -1; 1154 1155 for (i = 0; i < placement->num_placement; i++) { 1156 if ((placement->placement[i] & mem->placement & 1157 TTM_PL_MASK_CACHING) && 1158 (placement->placement[i] & mem->placement & 1159 TTM_PL_MASK_MEM)) 1160 return i; 1161 } 1162 return -1; 1163} 1164 1165int ttm_bo_validate(struct ttm_buffer_object *bo, 1166 struct ttm_placement *placement, 1167 bool interruptible, 1168 bool no_wait_gpu) 1169{ 1170 int ret; 1171 1172 MPASS(ttm_bo_is_reserved(bo)); 1173 /* Check that range is valid */ 1174 if (placement->lpfn || placement->fpfn) 1175 if (placement->fpfn > placement->lpfn || 1176 (placement->lpfn - placement->fpfn) < bo->num_pages) 1177 return -EINVAL; 1178 /* 1179 * Check whether we need to move buffer. 1180 */ 1181 ret = ttm_bo_mem_compat(placement, &bo->mem); 1182 if (ret < 0) { 1183 ret = ttm_bo_move_buffer(bo, placement, interruptible, 1184 no_wait_gpu); 1185 if (ret) 1186 return ret; 1187 } else { 1188 /* 1189 * Use the access and other non-mapping-related flag bits from 1190 * the compatible memory placement flags to the active flags 1191 */ 1192 ttm_flag_masked(&bo->mem.placement, placement->placement[ret], 1193 ~TTM_PL_MASK_MEMTYPE); 1194 } 1195 /* 1196 * We might need to add a TTM. 1197 */ 1198 if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) { 1199 ret = ttm_bo_add_ttm(bo, true); 1200 if (ret) 1201 return ret; 1202 } 1203 return 0; 1204} 1205 1206int ttm_bo_check_placement(struct ttm_buffer_object *bo, 1207 struct ttm_placement *placement) 1208{ 1209 MPASS(!((placement->fpfn || placement->lpfn) && 1210 (bo->mem.num_pages > (placement->lpfn - placement->fpfn)))); 1211 1212 return 0; 1213} 1214 1215int ttm_bo_init(struct ttm_bo_device *bdev, 1216 struct ttm_buffer_object *bo, 1217 unsigned long size, 1218 enum ttm_bo_type type, 1219 struct ttm_placement *placement, 1220 uint32_t page_alignment, 1221 bool interruptible, 1222 struct vm_object *persistent_swap_storage, 1223 size_t acc_size, 1224 struct sg_table *sg, 1225 void (*destroy) (struct ttm_buffer_object *)) 1226{ 1227 int ret = 0; 1228 unsigned long num_pages; 1229 struct ttm_mem_global *mem_glob = bdev->glob->mem_glob; 1230 1231 ret = ttm_mem_global_alloc(mem_glob, acc_size, false, false); 1232 if (ret) { 1233 printf("[TTM] Out of kernel memory\n"); 1234 if (destroy) 1235 (*destroy)(bo); 1236 else 1237 free(bo, M_TTM_BO); 1238 return -ENOMEM; 1239 } 1240 1241 num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT; 1242 if (num_pages == 0) { 1243 printf("[TTM] Illegal buffer object size\n"); 1244 if (destroy) 1245 (*destroy)(bo); 1246 else 1247 free(bo, M_TTM_BO); 1248 ttm_mem_global_free(mem_glob, acc_size); 1249 return -EINVAL; 1250 } 1251 bo->destroy = destroy; 1252 1253 refcount_init(&bo->kref, 1); 1254 refcount_init(&bo->list_kref, 1); 1255 atomic_set(&bo->cpu_writers, 0); 1256 atomic_set(&bo->reserved, 1); 1257 INIT_LIST_HEAD(&bo->lru); 1258 INIT_LIST_HEAD(&bo->ddestroy); 1259 INIT_LIST_HEAD(&bo->swap); 1260 INIT_LIST_HEAD(&bo->io_reserve_lru); 1261 bo->bdev = bdev; 1262 bo->glob = bdev->glob; 1263 bo->type = type; 1264 bo->num_pages = num_pages; 1265 bo->mem.size = num_pages << PAGE_SHIFT; 1266 bo->mem.mem_type = TTM_PL_SYSTEM; 1267 bo->mem.num_pages = bo->num_pages; 1268 bo->mem.mm_node = NULL; 1269 bo->mem.page_alignment = page_alignment; 1270 bo->mem.bus.io_reserved_vm = false; 1271 bo->mem.bus.io_reserved_count = 0; 1272 bo->priv_flags = 0; 1273 bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED); 1274 bo->seq_valid = false; 1275 bo->persistent_swap_storage = persistent_swap_storage; 1276 bo->acc_size = acc_size; 1277 bo->sg = sg; 1278 atomic_inc(&bo->glob->bo_count); 1279 1280 ret = ttm_bo_check_placement(bo, placement); 1281 if (unlikely(ret != 0)) 1282 goto out_err; 1283 1284 /* 1285 * For ttm_bo_type_device buffers, allocate 1286 * address space from the device. 1287 */ 1288 if (bo->type == ttm_bo_type_device || 1289 bo->type == ttm_bo_type_sg) { 1290 ret = ttm_bo_setup_vm(bo); 1291 if (ret) 1292 goto out_err; 1293 } 1294 1295 ret = ttm_bo_validate(bo, placement, interruptible, false); 1296 if (ret) 1297 goto out_err; 1298 1299 ttm_bo_unreserve(bo); 1300 return 0; 1301 1302out_err: 1303 ttm_bo_unreserve(bo); 1304 ttm_bo_unref(&bo); 1305 1306 return ret; 1307} 1308 1309size_t ttm_bo_acc_size(struct ttm_bo_device *bdev, 1310 unsigned long bo_size, 1311 unsigned struct_size) 1312{ 1313 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT; 1314 size_t size = 0; 1315 1316 size += ttm_round_pot(struct_size); 1317 size += PAGE_ALIGN(npages * sizeof(void *)); 1318 size += ttm_round_pot(sizeof(struct ttm_tt)); 1319 return size; 1320} 1321 1322size_t ttm_bo_dma_acc_size(struct ttm_bo_device *bdev, 1323 unsigned long bo_size, 1324 unsigned struct_size) 1325{ 1326 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT; 1327 size_t size = 0; 1328 1329 size += ttm_round_pot(struct_size); 1330 size += PAGE_ALIGN(npages * sizeof(void *)); 1331 size += PAGE_ALIGN(npages * sizeof(dma_addr_t)); 1332 size += ttm_round_pot(sizeof(struct ttm_dma_tt)); 1333 return size; 1334} 1335 1336int ttm_bo_create(struct ttm_bo_device *bdev, 1337 unsigned long size, 1338 enum ttm_bo_type type, 1339 struct ttm_placement *placement, 1340 uint32_t page_alignment, 1341 bool interruptible, 1342 struct vm_object *persistent_swap_storage, 1343 struct ttm_buffer_object **p_bo) 1344{ 1345 struct ttm_buffer_object *bo; 1346 size_t acc_size; 1347 int ret; 1348 1349 bo = malloc(sizeof(*bo), M_TTM_BO, M_WAITOK | M_ZERO); 1350 acc_size = ttm_bo_acc_size(bdev, size, sizeof(struct ttm_buffer_object)); 1351 ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment, 1352 interruptible, persistent_swap_storage, acc_size, 1353 NULL, NULL); 1354 if (likely(ret == 0)) 1355 *p_bo = bo; 1356 1357 return ret; 1358} 1359 1360static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev, 1361 unsigned mem_type, bool allow_errors) 1362{ 1363 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 1364 struct ttm_bo_global *glob = bdev->glob; 1365 int ret; 1366 1367 /* 1368 * Can't use standard list traversal since we're unlocking. 1369 */ 1370 1371 mtx_lock(&glob->lru_lock); 1372 while (!list_empty(&man->lru)) { 1373 mtx_unlock(&glob->lru_lock); 1374 ret = ttm_mem_evict_first(bdev, mem_type, false, false); 1375 if (ret) { 1376 if (allow_errors) { 1377 return ret; 1378 } else { 1379 printf("[TTM] Cleanup eviction failed\n"); 1380 } 1381 } 1382 mtx_lock(&glob->lru_lock); 1383 } 1384 mtx_unlock(&glob->lru_lock); 1385 return 0; 1386} 1387 1388int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type) 1389{ 1390 struct ttm_mem_type_manager *man; 1391 int ret = -EINVAL; 1392 1393 if (mem_type >= TTM_NUM_MEM_TYPES) { 1394 printf("[TTM] Illegal memory type %d\n", mem_type); 1395 return ret; 1396 } 1397 man = &bdev->man[mem_type]; 1398 1399 if (!man->has_type) { 1400 printf("[TTM] Trying to take down uninitialized memory manager type %u\n", 1401 mem_type); 1402 return ret; 1403 } 1404 1405 man->use_type = false; 1406 man->has_type = false; 1407 1408 ret = 0; 1409 if (mem_type > 0) { 1410 ttm_bo_force_list_clean(bdev, mem_type, false); 1411 1412 ret = (*man->func->takedown)(man); 1413 } 1414 1415 return ret; 1416} 1417 1418int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type) 1419{ 1420 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 1421 1422 if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) { 1423 printf("[TTM] Illegal memory manager memory type %u\n", mem_type); 1424 return -EINVAL; 1425 } 1426 1427 if (!man->has_type) { 1428 printf("[TTM] Memory type %u has not been initialized\n", mem_type); 1429 return 0; 1430 } 1431 1432 return ttm_bo_force_list_clean(bdev, mem_type, true); 1433} 1434 1435int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type, 1436 unsigned long p_size) 1437{ 1438 int ret = -EINVAL; 1439 struct ttm_mem_type_manager *man; 1440 1441 MPASS(type < TTM_NUM_MEM_TYPES); 1442 man = &bdev->man[type]; 1443 MPASS(!man->has_type); 1444 man->io_reserve_fastpath = true; 1445 man->use_io_reserve_lru = false; 1446 sx_init(&man->io_reserve_mutex, "ttmman"); 1447 INIT_LIST_HEAD(&man->io_reserve_lru); 1448 1449 ret = bdev->driver->init_mem_type(bdev, type, man); 1450 if (ret) 1451 return ret; 1452 man->bdev = bdev; 1453 1454 ret = 0; 1455 if (type != TTM_PL_SYSTEM) { 1456 ret = (*man->func->init)(man, p_size); 1457 if (ret) 1458 return ret; 1459 } 1460 man->has_type = true; 1461 man->use_type = true; 1462 man->size = p_size; 1463 1464 INIT_LIST_HEAD(&man->lru); 1465 1466 return 0; 1467} 1468 1469static void ttm_bo_global_kobj_release(struct ttm_bo_global *glob) 1470{ 1471 1472 ttm_mem_unregister_shrink(glob->mem_glob, &glob->shrink); 1473 vm_page_free(glob->dummy_read_page); 1474} 1475 1476void ttm_bo_global_release(struct drm_global_reference *ref) 1477{ 1478 struct ttm_bo_global *glob = ref->object; 1479 1480 if (refcount_release(&glob->kobj_ref)) 1481 ttm_bo_global_kobj_release(glob); 1482} 1483 1484int ttm_bo_global_init(struct drm_global_reference *ref) 1485{ 1486 struct ttm_bo_global_ref *bo_ref = 1487 container_of(ref, struct ttm_bo_global_ref, ref); 1488 struct ttm_bo_global *glob = ref->object; 1489 int ret; 1490 int tries; 1491 1492 sx_init(&glob->device_list_mutex, "ttmdlm"); 1493 mtx_init(&glob->lru_lock, "ttmlru", NULL, MTX_DEF); 1494 glob->mem_glob = bo_ref->mem_glob; 1495 tries = 0; 1496retry: 1497 glob->dummy_read_page = vm_page_alloc_noobj_contig(0, 1, 0, 1498 VM_MAX_ADDRESS, PAGE_SIZE, 0, VM_MEMATTR_UNCACHEABLE); 1499 1500 if (unlikely(glob->dummy_read_page == NULL)) { 1501 if (tries < 1 && (vm_page_reclaim_contig(0, 1, 0, 1502 VM_MAX_ADDRESS, PAGE_SIZE, 0) == 0)) { 1503 tries++; 1504 goto retry; 1505 } 1506 ret = -ENOMEM; 1507 goto out_no_drp; 1508 } 1509 1510 INIT_LIST_HEAD(&glob->swap_lru); 1511 INIT_LIST_HEAD(&glob->device_list); 1512 1513 ttm_mem_init_shrink(&glob->shrink, ttm_bo_swapout); 1514 ret = ttm_mem_register_shrink(glob->mem_glob, &glob->shrink); 1515 if (unlikely(ret != 0)) { 1516 printf("[TTM] Could not register buffer object swapout\n"); 1517 goto out_no_shrink; 1518 } 1519 1520 atomic_set(&glob->bo_count, 0); 1521 1522 refcount_init(&glob->kobj_ref, 1); 1523 return (0); 1524 1525out_no_shrink: 1526 vm_page_free(glob->dummy_read_page); 1527out_no_drp: 1528 free(glob, M_DRM_GLOBAL); 1529 return ret; 1530} 1531 1532int ttm_bo_device_release(struct ttm_bo_device *bdev) 1533{ 1534 int ret = 0; 1535 unsigned i = TTM_NUM_MEM_TYPES; 1536 struct ttm_mem_type_manager *man; 1537 struct ttm_bo_global *glob = bdev->glob; 1538 1539 while (i--) { 1540 man = &bdev->man[i]; 1541 if (man->has_type) { 1542 man->use_type = false; 1543 if ((i != TTM_PL_SYSTEM) && ttm_bo_clean_mm(bdev, i)) { 1544 ret = -EBUSY; 1545 printf("[TTM] DRM memory manager type %d is not clean\n", 1546 i); 1547 } 1548 man->has_type = false; 1549 } 1550 } 1551 1552 sx_xlock(&glob->device_list_mutex); 1553 list_del(&bdev->device_list); 1554 sx_xunlock(&glob->device_list_mutex); 1555 1556 if (taskqueue_cancel_timeout(taskqueue_thread, &bdev->wq, NULL)) 1557 taskqueue_drain_timeout(taskqueue_thread, &bdev->wq); 1558 1559 while (ttm_bo_delayed_delete(bdev, true)) 1560 ; 1561 1562 mtx_lock(&glob->lru_lock); 1563 if (list_empty(&bdev->ddestroy)) 1564 TTM_DEBUG("Delayed destroy list was clean\n"); 1565 1566 if (list_empty(&bdev->man[0].lru)) 1567 TTM_DEBUG("Swap list was clean\n"); 1568 mtx_unlock(&glob->lru_lock); 1569 1570 MPASS(drm_mm_clean(&bdev->addr_space_mm)); 1571 rw_wlock(&bdev->vm_lock); 1572 drm_mm_takedown(&bdev->addr_space_mm); 1573 rw_wunlock(&bdev->vm_lock); 1574 1575 return ret; 1576} 1577 1578int ttm_bo_device_init(struct ttm_bo_device *bdev, 1579 struct ttm_bo_global *glob, 1580 struct ttm_bo_driver *driver, 1581 uint64_t file_page_offset, 1582 bool need_dma32) 1583{ 1584 int ret = -EINVAL; 1585 1586 rw_init(&bdev->vm_lock, "ttmvml"); 1587 bdev->driver = driver; 1588 1589 memset(bdev->man, 0, sizeof(bdev->man)); 1590 1591 /* 1592 * Initialize the system memory buffer type. 1593 * Other types need to be driver / IOCTL initialized. 1594 */ 1595 ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0); 1596 if (unlikely(ret != 0)) 1597 goto out_no_sys; 1598 1599 RB_INIT(&bdev->addr_space_rb); 1600 ret = drm_mm_init(&bdev->addr_space_mm, file_page_offset, 0x10000000); 1601 if (unlikely(ret != 0)) 1602 goto out_no_addr_mm; 1603 1604 TIMEOUT_TASK_INIT(taskqueue_thread, &bdev->wq, 0, 1605 ttm_bo_delayed_workqueue, bdev); 1606 INIT_LIST_HEAD(&bdev->ddestroy); 1607 bdev->dev_mapping = NULL; 1608 bdev->glob = glob; 1609 bdev->need_dma32 = need_dma32; 1610 bdev->val_seq = 0; 1611 mtx_init(&bdev->fence_lock, "ttmfence", NULL, MTX_DEF); 1612 sx_xlock(&glob->device_list_mutex); 1613 list_add_tail(&bdev->device_list, &glob->device_list); 1614 sx_xunlock(&glob->device_list_mutex); 1615 1616 return 0; 1617out_no_addr_mm: 1618 ttm_bo_clean_mm(bdev, 0); 1619out_no_sys: 1620 return ret; 1621} 1622 1623/* 1624 * buffer object vm functions. 1625 */ 1626 1627bool ttm_mem_reg_is_pci(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem) 1628{ 1629 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type]; 1630 1631 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) { 1632 if (mem->mem_type == TTM_PL_SYSTEM) 1633 return false; 1634 1635 if (man->flags & TTM_MEMTYPE_FLAG_CMA) 1636 return false; 1637 1638 if (mem->placement & TTM_PL_FLAG_CACHED) 1639 return false; 1640 } 1641 return true; 1642} 1643 1644void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo) 1645{ 1646 1647 ttm_bo_release_mmap(bo); 1648 ttm_mem_io_free_vm(bo); 1649} 1650 1651void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo) 1652{ 1653 struct ttm_bo_device *bdev = bo->bdev; 1654 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type]; 1655 1656 ttm_mem_io_lock(man, false); 1657 ttm_bo_unmap_virtual_locked(bo); 1658 ttm_mem_io_unlock(man); 1659} 1660 1661static void ttm_bo_vm_insert_rb(struct ttm_buffer_object *bo) 1662{ 1663 struct ttm_bo_device *bdev = bo->bdev; 1664 1665 /* The caller acquired bdev->vm_lock. */ 1666 RB_INSERT(ttm_bo_device_buffer_objects, &bdev->addr_space_rb, bo); 1667} 1668 1669/** 1670 * ttm_bo_setup_vm: 1671 * 1672 * @bo: the buffer to allocate address space for 1673 * 1674 * Allocate address space in the drm device so that applications 1675 * can mmap the buffer and access the contents. This only 1676 * applies to ttm_bo_type_device objects as others are not 1677 * placed in the drm device address space. 1678 */ 1679 1680static int ttm_bo_setup_vm(struct ttm_buffer_object *bo) 1681{ 1682 struct ttm_bo_device *bdev = bo->bdev; 1683 int ret; 1684 1685retry_pre_get: 1686 ret = drm_mm_pre_get(&bdev->addr_space_mm); 1687 if (unlikely(ret != 0)) 1688 return ret; 1689 1690 rw_wlock(&bdev->vm_lock); 1691 bo->vm_node = drm_mm_search_free(&bdev->addr_space_mm, 1692 bo->mem.num_pages, 0, 0); 1693 1694 if (unlikely(bo->vm_node == NULL)) { 1695 ret = -ENOMEM; 1696 goto out_unlock; 1697 } 1698 1699 bo->vm_node = drm_mm_get_block_atomic(bo->vm_node, 1700 bo->mem.num_pages, 0); 1701 1702 if (unlikely(bo->vm_node == NULL)) { 1703 rw_wunlock(&bdev->vm_lock); 1704 goto retry_pre_get; 1705 } 1706 1707 ttm_bo_vm_insert_rb(bo); 1708 rw_wunlock(&bdev->vm_lock); 1709 bo->addr_space_offset = ((uint64_t) bo->vm_node->start) << PAGE_SHIFT; 1710 1711 return 0; 1712out_unlock: 1713 rw_wunlock(&bdev->vm_lock); 1714 return ret; 1715} 1716 1717int ttm_bo_wait(struct ttm_buffer_object *bo, 1718 bool lazy, bool interruptible, bool no_wait) 1719{ 1720 struct ttm_bo_driver *driver = bo->bdev->driver; 1721 struct ttm_bo_device *bdev = bo->bdev; 1722 void *sync_obj; 1723 int ret = 0; 1724 1725 if (likely(bo->sync_obj == NULL)) 1726 return 0; 1727 1728 while (bo->sync_obj) { 1729 1730 if (driver->sync_obj_signaled(bo->sync_obj)) { 1731 void *tmp_obj = bo->sync_obj; 1732 bo->sync_obj = NULL; 1733 clear_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags); 1734 mtx_unlock(&bdev->fence_lock); 1735 driver->sync_obj_unref(&tmp_obj); 1736 mtx_lock(&bdev->fence_lock); 1737 continue; 1738 } 1739 1740 if (no_wait) 1741 return -EBUSY; 1742 1743 sync_obj = driver->sync_obj_ref(bo->sync_obj); 1744 mtx_unlock(&bdev->fence_lock); 1745 ret = driver->sync_obj_wait(sync_obj, 1746 lazy, interruptible); 1747 if (unlikely(ret != 0)) { 1748 driver->sync_obj_unref(&sync_obj); 1749 mtx_lock(&bdev->fence_lock); 1750 return ret; 1751 } 1752 mtx_lock(&bdev->fence_lock); 1753 if (likely(bo->sync_obj == sync_obj)) { 1754 void *tmp_obj = bo->sync_obj; 1755 bo->sync_obj = NULL; 1756 clear_bit(TTM_BO_PRIV_FLAG_MOVING, 1757 &bo->priv_flags); 1758 mtx_unlock(&bdev->fence_lock); 1759 driver->sync_obj_unref(&sync_obj); 1760 driver->sync_obj_unref(&tmp_obj); 1761 mtx_lock(&bdev->fence_lock); 1762 } else { 1763 mtx_unlock(&bdev->fence_lock); 1764 driver->sync_obj_unref(&sync_obj); 1765 mtx_lock(&bdev->fence_lock); 1766 } 1767 } 1768 return 0; 1769} 1770 1771int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait) 1772{ 1773 struct ttm_bo_device *bdev = bo->bdev; 1774 int ret = 0; 1775 1776 /* 1777 * Using ttm_bo_reserve makes sure the lru lists are updated. 1778 */ 1779 1780 ret = ttm_bo_reserve(bo, true, no_wait, false, 0); 1781 if (unlikely(ret != 0)) 1782 return ret; 1783 mtx_lock(&bdev->fence_lock); 1784 ret = ttm_bo_wait(bo, false, true, no_wait); 1785 mtx_unlock(&bdev->fence_lock); 1786 if (likely(ret == 0)) 1787 atomic_inc(&bo->cpu_writers); 1788 ttm_bo_unreserve(bo); 1789 return ret; 1790} 1791 1792void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo) 1793{ 1794 atomic_dec(&bo->cpu_writers); 1795} 1796 1797/** 1798 * A buffer object shrink method that tries to swap out the first 1799 * buffer object on the bo_global::swap_lru list. 1800 */ 1801 1802static int ttm_bo_swapout(struct ttm_mem_shrink *shrink) 1803{ 1804 struct ttm_bo_global *glob = 1805 container_of(shrink, struct ttm_bo_global, shrink); 1806 struct ttm_buffer_object *bo; 1807 int ret = -EBUSY; 1808 int put_count; 1809 uint32_t swap_placement = (TTM_PL_FLAG_CACHED | TTM_PL_FLAG_SYSTEM); 1810 1811 mtx_lock(&glob->lru_lock); 1812 list_for_each_entry(bo, &glob->swap_lru, swap) { 1813 ret = ttm_bo_reserve_nolru(bo, false, true, false, 0); 1814 if (!ret) 1815 break; 1816 } 1817 1818 if (ret) { 1819 mtx_unlock(&glob->lru_lock); 1820 return ret; 1821 } 1822 1823 refcount_acquire(&bo->list_kref); 1824 1825 if (!list_empty(&bo->ddestroy)) { 1826 ret = ttm_bo_cleanup_refs_and_unlock(bo, false, false); 1827 if (refcount_release(&bo->list_kref)) 1828 ttm_bo_release_list(bo); 1829 return ret; 1830 } 1831 1832 put_count = ttm_bo_del_from_lru(bo); 1833 mtx_unlock(&glob->lru_lock); 1834 1835 ttm_bo_list_ref_sub(bo, put_count, true); 1836 1837 /** 1838 * Wait for GPU, then move to system cached. 1839 */ 1840 1841 mtx_lock(&bo->bdev->fence_lock); 1842 ret = ttm_bo_wait(bo, false, false, false); 1843 mtx_unlock(&bo->bdev->fence_lock); 1844 1845 if (unlikely(ret != 0)) 1846 goto out; 1847 1848 if ((bo->mem.placement & swap_placement) != swap_placement) { 1849 struct ttm_mem_reg evict_mem; 1850 1851 evict_mem = bo->mem; 1852 evict_mem.mm_node = NULL; 1853 evict_mem.placement = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED; 1854 evict_mem.mem_type = TTM_PL_SYSTEM; 1855 1856 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, 1857 false, false); 1858 if (unlikely(ret != 0)) 1859 goto out; 1860 } 1861 1862 ttm_bo_unmap_virtual(bo); 1863 1864 /** 1865 * Swap out. Buffer will be swapped in again as soon as 1866 * anyone tries to access a ttm page. 1867 */ 1868 1869 if (bo->bdev->driver->swap_notify) 1870 bo->bdev->driver->swap_notify(bo); 1871 1872 ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage); 1873out: 1874 1875 /** 1876 * 1877 * Unreserve without putting on LRU to avoid swapping out an 1878 * already swapped buffer. 1879 */ 1880 1881 atomic_set(&bo->reserved, 0); 1882 wakeup(bo); 1883 if (refcount_release(&bo->list_kref)) 1884 ttm_bo_release_list(bo); 1885 return ret; 1886} 1887 1888void ttm_bo_swapout_all(struct ttm_bo_device *bdev) 1889{ 1890 while (ttm_bo_swapout(&bdev->glob->shrink) == 0) 1891 ; 1892} 1893