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 #define pr_fmt(fmt) "[TTM] " fmt 32 33 #include <drm/ttm/ttm_module.h> 34 #include <drm/ttm/ttm_bo_driver.h> 35 #include <drm/ttm/ttm_placement.h> 36 #include <linux/atomic.h> 37 #include <linux/errno.h> 38 #include <linux/export.h> 39 #include <linux/wait.h> 40 41 #define TTM_ASSERT_LOCKED(param) 42 #define TTM_DEBUG(fmt, arg...) 43 #define TTM_BO_HASH_ORDER 13 44 45 static int ttm_bo_setup_vm(struct ttm_buffer_object *bo); 46 static int ttm_bo_swapout(struct ttm_mem_shrink *shrink); 47 static void ttm_bo_global_kobj_release(struct ttm_bo_global *glob); 48 49 static inline int ttm_mem_type_from_place(const struct ttm_place *place, 50 uint32_t *mem_type) 51 { 52 int i; 53 54 for (i = 0; i <= TTM_PL_PRIV5; i++) 55 if (place->flags & (1 << i)) { 56 *mem_type = i; 57 return 0; 58 } 59 return -EINVAL; 60 } 61 62 static 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 pr_err(" has_type: %d\n", man->has_type); 67 pr_err(" use_type: %d\n", man->use_type); 68 pr_err(" flags: 0x%08X\n", man->flags); 69 pr_err(" gpu_offset: 0x%08lX\n", man->gpu_offset); 70 pr_err(" size: %ju\n", (uintmax_t)man->size); 71 pr_err(" available_caching: 0x%08X\n", man->available_caching); 72 pr_err(" default_caching: 0x%08X\n", man->default_caching); 73 if (mem_type != TTM_PL_SYSTEM) 74 (*man->func->debug)(man, TTM_PFX); 75 } 76 77 static 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 pr_err("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_place(&placement->placement[i], 87 &mem_type); 88 if (ret) 89 return; 90 pr_err(" placement[%d]=0x%08X (%d)\n", 91 i, placement->placement[i].flags, mem_type); 92 ttm_mem_type_debug(bo->bdev, mem_type); 93 } 94 } 95 96 #if 0 97 static 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 105 106 static inline uint32_t ttm_bo_type_flags(unsigned type) 107 { 108 return 1 << (type); 109 } 110 111 static void ttm_bo_release_list(struct kref *list_kref) 112 { 113 struct ttm_buffer_object *bo = 114 container_of(list_kref, struct ttm_buffer_object, list_kref); 115 struct ttm_bo_device *bdev = bo->bdev; 116 size_t acc_size = bo->acc_size; 117 118 BUG_ON(atomic_read(&bo->list_kref.refcount)); 119 BUG_ON(atomic_read(&bo->kref.refcount)); 120 BUG_ON(atomic_read(&bo->cpu_writers)); 121 BUG_ON(bo->sync_obj != NULL); 122 BUG_ON(bo->mem.mm_node != NULL); 123 BUG_ON(!list_empty(&bo->lru)); 124 BUG_ON(!list_empty(&bo->ddestroy)); 125 126 if (bo->ttm) 127 ttm_tt_destroy(bo->ttm); 128 atomic_dec(&bo->glob->bo_count); 129 if (bo->destroy) 130 bo->destroy(bo); 131 else { 132 kfree(bo); 133 } 134 ttm_mem_global_free(bdev->glob->mem_glob, acc_size); 135 } 136 137 static int ttm_bo_wait_unreserved(struct ttm_buffer_object *bo, 138 bool interruptible) 139 { 140 if (interruptible) { 141 return wait_event_interruptible(bo->event_queue, 142 !ttm_bo_is_reserved(bo)); 143 } else { 144 wait_event(bo->event_queue, !ttm_bo_is_reserved(bo)); 145 return 0; 146 } 147 } 148 149 void ttm_bo_add_to_lru(struct ttm_buffer_object *bo) 150 { 151 struct ttm_bo_device *bdev = bo->bdev; 152 struct ttm_mem_type_manager *man; 153 154 BUG_ON(!ttm_bo_is_reserved(bo)); 155 156 if (!(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) { 157 158 BUG_ON(!list_empty(&bo->lru)); 159 160 man = &bdev->man[bo->mem.mem_type]; 161 list_add_tail(&bo->lru, &man->lru); 162 kref_get(&bo->list_kref); 163 164 if (bo->ttm != NULL) { 165 list_add_tail(&bo->swap, &bo->glob->swap_lru); 166 kref_get(&bo->list_kref); 167 } 168 } 169 } 170 171 int ttm_bo_del_from_lru(struct ttm_buffer_object *bo) 172 { 173 int put_count = 0; 174 175 if (!list_empty(&bo->swap)) { 176 list_del_init(&bo->swap); 177 ++put_count; 178 } 179 if (!list_empty(&bo->lru)) { 180 list_del_init(&bo->lru); 181 ++put_count; 182 } 183 184 /* 185 * TODO: Add a driver hook to delete from 186 * driver-specific LRU's here. 187 */ 188 189 return put_count; 190 } 191 192 int ttm_bo_reserve_nolru(struct ttm_buffer_object *bo, 193 bool interruptible, 194 bool no_wait, bool use_ticket, 195 struct ww_acquire_ctx *ticket) 196 { 197 int ret; 198 199 while (unlikely(atomic_xchg(&bo->reserved, 1) != 0)) { 200 /** 201 * Deadlock avoidance for multi-bo reserving. 202 */ 203 if (use_ticket && bo->seq_valid) { 204 /** 205 * We've already reserved this one. 206 */ 207 if (unlikely(ticket->stamp == bo->val_seq)) 208 return -EDEADLK; 209 /** 210 * Already reserved by a thread that will not back 211 * off for us. We need to back off. 212 */ 213 if (unlikely(ticket->stamp - bo->val_seq <= LONG_MAX)) 214 return -EAGAIN; 215 } 216 217 if (no_wait) 218 return -EBUSY; 219 220 ret = ttm_bo_wait_unreserved(bo, interruptible); 221 222 if (unlikely(ret)) 223 return ret; 224 } 225 226 if (use_ticket) { 227 bool wake_up = false; 228 229 /** 230 * Wake up waiters that may need to recheck for deadlock, 231 * if we decreased the sequence number. 232 */ 233 if (unlikely((bo->val_seq - ticket->stamp <= LONG_MAX) 234 || !bo->seq_valid)) 235 wake_up = true; 236 237 /* 238 * In the worst case with memory ordering these values can be 239 * seen in the wrong order. However since we call wake_up_all 240 * in that case, this will hopefully not pose a problem, 241 * and the worst case would only cause someone to accidentally 242 * hit -EAGAIN in ttm_bo_reserve when they see old value of 243 * val_seq. However this would only happen if seq_valid was 244 * written before val_seq was, and just means some slightly 245 * increased cpu usage 246 */ 247 bo->val_seq = ticket->stamp; 248 bo->seq_valid = true; 249 if (wake_up) 250 wake_up_all(&bo->event_queue); 251 } else { 252 bo->seq_valid = false; 253 } 254 255 return 0; 256 } 257 EXPORT_SYMBOL(ttm_bo_reserve); 258 259 static void ttm_bo_ref_bug(struct kref *list_kref) 260 { 261 BUG(); 262 } 263 264 void ttm_bo_list_ref_sub(struct ttm_buffer_object *bo, int count, 265 bool never_free) 266 { 267 kref_sub(&bo->list_kref, count, 268 (never_free) ? ttm_bo_ref_bug : ttm_bo_release_list); 269 } 270 271 int ttm_bo_reserve(struct ttm_buffer_object *bo, 272 bool interruptible, 273 bool no_wait, bool use_ticket, 274 struct ww_acquire_ctx *ticket) 275 { 276 struct ttm_bo_global *glob = bo->glob; 277 int put_count = 0; 278 int ret; 279 280 ret = ttm_bo_reserve_nolru(bo, interruptible, no_wait, use_ticket, 281 ticket); 282 if (likely(ret == 0)) { 283 lockmgr(&glob->lru_lock, LK_EXCLUSIVE); 284 put_count = ttm_bo_del_from_lru(bo); 285 lockmgr(&glob->lru_lock, LK_RELEASE); 286 ttm_bo_list_ref_sub(bo, put_count, true); 287 } 288 289 return ret; 290 } 291 292 int ttm_bo_reserve_slowpath_nolru(struct ttm_buffer_object *bo, 293 bool interruptible, 294 struct ww_acquire_ctx *ticket) 295 { 296 bool wake_up = false; 297 int ret; 298 299 while (unlikely(atomic_xchg(&bo->reserved, 1) != 0)) { 300 WARN_ON(bo->seq_valid && ticket->stamp == bo->val_seq); 301 302 ret = ttm_bo_wait_unreserved(bo, interruptible); 303 304 if (unlikely(ret)) 305 return ret; 306 } 307 308 if (bo->val_seq - ticket->stamp < LONG_MAX || !bo->seq_valid) 309 wake_up = true; 310 311 /** 312 * Wake up waiters that may need to recheck for deadlock, 313 * if we decreased the sequence number. 314 */ 315 bo->val_seq = ticket->stamp; 316 bo->seq_valid = true; 317 if (wake_up) 318 wake_up_all(&bo->event_queue); 319 320 return 0; 321 } 322 323 int ttm_bo_reserve_slowpath(struct ttm_buffer_object *bo, 324 bool interruptible, struct ww_acquire_ctx *ticket) 325 { 326 struct ttm_bo_global *glob = bo->glob; 327 int put_count, ret; 328 329 ret = ttm_bo_reserve_slowpath_nolru(bo, interruptible, ticket); 330 if (likely(!ret)) { 331 lockmgr(&glob->lru_lock, LK_EXCLUSIVE); 332 put_count = ttm_bo_del_from_lru(bo); 333 lockmgr(&glob->lru_lock, LK_RELEASE); 334 ttm_bo_list_ref_sub(bo, put_count, true); 335 } 336 return ret; 337 } 338 EXPORT_SYMBOL(ttm_bo_reserve_slowpath); 339 340 /* 341 * Must interlock with event_queue to avoid race against 342 * wait_event_common() which can cause wait_event_common() 343 * to become stuck. 344 */ 345 static void 346 ttm_bo_unreserve_core(struct ttm_buffer_object *bo) 347 { 348 lockmgr(&bo->event_queue.lock, LK_EXCLUSIVE); 349 atomic_set(&bo->reserved, 0); 350 lockmgr(&bo->event_queue.lock, LK_RELEASE); 351 wake_up_all(&bo->event_queue); 352 } 353 354 void ttm_bo_unreserve_ticket_locked(struct ttm_buffer_object *bo, struct ww_acquire_ctx *ticket) 355 { 356 ttm_bo_add_to_lru(bo); 357 ttm_bo_unreserve_core(bo); 358 } 359 360 void ttm_bo_unreserve(struct ttm_buffer_object *bo) 361 { 362 struct ttm_bo_global *glob = bo->glob; 363 364 lockmgr(&glob->lru_lock, LK_EXCLUSIVE); 365 ttm_bo_unreserve_ticket_locked(bo, NULL); 366 lockmgr(&glob->lru_lock, LK_RELEASE); 367 } 368 EXPORT_SYMBOL(ttm_bo_unreserve); 369 370 void ttm_bo_unreserve_ticket(struct ttm_buffer_object *bo, struct ww_acquire_ctx *ticket) 371 { 372 struct ttm_bo_global *glob = bo->glob; 373 374 lockmgr(&glob->lru_lock, LK_EXCLUSIVE); 375 ttm_bo_unreserve_ticket_locked(bo, ticket); 376 lockmgr(&glob->lru_lock, LK_RELEASE); 377 } 378 EXPORT_SYMBOL(ttm_bo_unreserve_ticket); 379 380 /* 381 * Call bo->mutex locked. 382 */ 383 static int ttm_bo_add_ttm(struct ttm_buffer_object *bo, bool zero_alloc) 384 { 385 struct ttm_bo_device *bdev = bo->bdev; 386 struct ttm_bo_global *glob = bo->glob; 387 int ret = 0; 388 uint32_t page_flags = 0; 389 390 TTM_ASSERT_LOCKED(&bo->mutex); 391 bo->ttm = NULL; 392 393 if (bdev->need_dma32) 394 page_flags |= TTM_PAGE_FLAG_DMA32; 395 396 switch (bo->type) { 397 case ttm_bo_type_device: 398 if (zero_alloc) 399 page_flags |= TTM_PAGE_FLAG_ZERO_ALLOC; 400 case ttm_bo_type_kernel: 401 bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT, 402 page_flags, glob->dummy_read_page); 403 if (unlikely(bo->ttm == NULL)) 404 ret = -ENOMEM; 405 break; 406 case ttm_bo_type_sg: 407 bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT, 408 page_flags | TTM_PAGE_FLAG_SG, 409 glob->dummy_read_page); 410 if (unlikely(bo->ttm == NULL)) { 411 ret = -ENOMEM; 412 break; 413 } 414 bo->ttm->sg = bo->sg; 415 break; 416 default: 417 pr_err("Illegal buffer object type\n"); 418 ret = -EINVAL; 419 break; 420 } 421 422 return ret; 423 } 424 425 static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo, 426 struct ttm_mem_reg *mem, 427 bool evict, bool interruptible, 428 bool no_wait_gpu) 429 { 430 struct ttm_bo_device *bdev = bo->bdev; 431 bool old_is_pci = ttm_mem_reg_is_pci(bdev, &bo->mem); 432 bool new_is_pci = ttm_mem_reg_is_pci(bdev, mem); 433 struct ttm_mem_type_manager *old_man = &bdev->man[bo->mem.mem_type]; 434 struct ttm_mem_type_manager *new_man = &bdev->man[mem->mem_type]; 435 int ret = 0; 436 437 if (old_is_pci || new_is_pci || 438 ((mem->placement & bo->mem.placement & TTM_PL_MASK_CACHING) == 0)) { 439 ret = ttm_mem_io_lock(old_man, true); 440 if (unlikely(ret != 0)) 441 goto out_err; 442 ttm_bo_unmap_virtual_locked(bo); 443 ttm_mem_io_unlock(old_man); 444 } 445 446 /* 447 * Create and bind a ttm if required. 448 */ 449 450 if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) { 451 if (bo->ttm == NULL) { 452 bool zero = !(old_man->flags & TTM_MEMTYPE_FLAG_FIXED); 453 ret = ttm_bo_add_ttm(bo, zero); 454 if (ret) 455 goto out_err; 456 } 457 458 ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement); 459 if (ret) 460 goto out_err; 461 462 if (mem->mem_type != TTM_PL_SYSTEM) { 463 ret = ttm_tt_bind(bo->ttm, mem); 464 if (ret) 465 goto out_err; 466 } 467 468 if (bo->mem.mem_type == TTM_PL_SYSTEM) { 469 if (bdev->driver->move_notify) 470 bdev->driver->move_notify(bo, mem); 471 bo->mem = *mem; 472 mem->mm_node = NULL; 473 goto moved; 474 } 475 } 476 477 if (bdev->driver->move_notify) 478 bdev->driver->move_notify(bo, mem); 479 480 if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) && 481 !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) 482 ret = ttm_bo_move_ttm(bo, evict, no_wait_gpu, mem); 483 else if (bdev->driver->move) 484 ret = bdev->driver->move(bo, evict, interruptible, 485 no_wait_gpu, mem); 486 else 487 ret = ttm_bo_move_memcpy(bo, evict, no_wait_gpu, mem); 488 489 if (ret) { 490 if (bdev->driver->move_notify) { 491 struct ttm_mem_reg tmp_mem = *mem; 492 *mem = bo->mem; 493 bo->mem = tmp_mem; 494 bdev->driver->move_notify(bo, mem); 495 bo->mem = *mem; 496 *mem = tmp_mem; 497 } 498 499 goto out_err; 500 } 501 502 moved: 503 if (bo->evicted) { 504 ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement); 505 if (ret) 506 pr_err("Can not flush read caches\n"); 507 bo->evicted = false; 508 } 509 510 if (bo->mem.mm_node) { 511 bo->offset = (bo->mem.start << PAGE_SHIFT) + 512 bdev->man[bo->mem.mem_type].gpu_offset; 513 bo->cur_placement = bo->mem.placement; 514 } else 515 bo->offset = 0; 516 517 return 0; 518 519 out_err: 520 new_man = &bdev->man[bo->mem.mem_type]; 521 if ((new_man->flags & TTM_MEMTYPE_FLAG_FIXED) && bo->ttm) { 522 ttm_tt_unbind(bo->ttm); 523 ttm_tt_destroy(bo->ttm); 524 bo->ttm = NULL; 525 } 526 527 return ret; 528 } 529 530 /** 531 * Call bo::reserved. 532 * Will release GPU memory type usage on destruction. 533 * This is the place to put in driver specific hooks to release 534 * driver private resources. 535 * Will release the bo::reserved lock. 536 */ 537 538 static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo) 539 { 540 if (bo->bdev->driver->move_notify) 541 bo->bdev->driver->move_notify(bo, NULL); 542 543 if (bo->ttm) { 544 ttm_tt_unbind(bo->ttm); 545 ttm_tt_destroy(bo->ttm); 546 bo->ttm = NULL; 547 } 548 ttm_bo_mem_put(bo, &bo->mem); 549 ttm_bo_unreserve_core(bo); 550 551 /* 552 * Since the final reference to this bo may not be dropped by 553 * the current task we have to put a memory barrier here to make 554 * sure the changes done in this function are always visible. 555 * 556 * This function only needs protection against the final kref_put. 557 */ 558 cpu_mfence(); 559 } 560 561 static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo) 562 { 563 struct ttm_bo_device *bdev = bo->bdev; 564 struct ttm_bo_global *glob = bo->glob; 565 struct ttm_bo_driver *driver = bdev->driver; 566 void *sync_obj = NULL; 567 int put_count; 568 int ret; 569 570 lockmgr(&glob->lru_lock, LK_EXCLUSIVE); 571 ret = ttm_bo_reserve_nolru(bo, false, true, false, 0); 572 573 lockmgr(&bdev->fence_lock, LK_EXCLUSIVE); 574 (void) ttm_bo_wait(bo, false, false, true); 575 if (!ret && !bo->sync_obj) { 576 lockmgr(&bdev->fence_lock, LK_RELEASE); 577 put_count = ttm_bo_del_from_lru(bo); 578 579 lockmgr(&glob->lru_lock, LK_RELEASE); 580 ttm_bo_cleanup_memtype_use(bo); 581 582 ttm_bo_list_ref_sub(bo, put_count, true); 583 584 return; 585 } 586 if (bo->sync_obj) 587 sync_obj = driver->sync_obj_ref(bo->sync_obj); 588 lockmgr(&bdev->fence_lock, LK_RELEASE); 589 590 if (!ret) { 591 592 /* 593 * Make NO_EVICT bos immediately available to 594 * shrinkers, now that they are queued for 595 * destruction. 596 */ 597 if (bo->mem.placement & TTM_PL_FLAG_NO_EVICT) { 598 bo->mem.placement &= ~TTM_PL_FLAG_NO_EVICT; 599 ttm_bo_add_to_lru(bo); 600 } 601 602 ttm_bo_unreserve_core(bo); 603 } 604 605 kref_get(&bo->list_kref); 606 list_add_tail(&bo->ddestroy, &bdev->ddestroy); 607 lockmgr(&glob->lru_lock, LK_RELEASE); 608 609 if (sync_obj) { 610 driver->sync_obj_flush(sync_obj); 611 driver->sync_obj_unref(&sync_obj); 612 } 613 schedule_delayed_work(&bdev->wq, 614 ((hz / 100) < 1) ? 1 : hz / 100); 615 } 616 617 /** 618 * function ttm_bo_cleanup_refs_and_unlock 619 * If bo idle, remove from delayed- and lru lists, and unref. 620 * If not idle, do nothing. 621 * 622 * Must be called with lru_lock and reservation held, this function 623 * will drop both before returning. 624 * 625 * @interruptible Any sleeps should occur interruptibly. 626 * @no_wait_gpu Never wait for gpu. Return -EBUSY instead. 627 */ 628 629 static int ttm_bo_cleanup_refs_and_unlock(struct ttm_buffer_object *bo, 630 bool interruptible, 631 bool no_wait_gpu) 632 { 633 struct ttm_bo_device *bdev = bo->bdev; 634 struct ttm_bo_driver *driver = bdev->driver; 635 struct ttm_bo_global *glob = bo->glob; 636 int put_count; 637 int ret; 638 639 lockmgr(&bdev->fence_lock, LK_EXCLUSIVE); 640 ret = ttm_bo_wait(bo, false, false, true); 641 642 if (ret && !no_wait_gpu) { 643 void *sync_obj; 644 645 /* 646 * Take a reference to the fence and unreserve, 647 * at this point the buffer should be dead, so 648 * no new sync objects can be attached. 649 */ 650 sync_obj = driver->sync_obj_ref(bo->sync_obj); 651 lockmgr(&bdev->fence_lock, LK_RELEASE); 652 653 ttm_bo_unreserve_core(bo); 654 lockmgr(&glob->lru_lock, LK_RELEASE); 655 656 ret = driver->sync_obj_wait(sync_obj, false, interruptible); 657 driver->sync_obj_unref(&sync_obj); 658 if (ret) 659 return ret; 660 661 /* 662 * remove sync_obj with ttm_bo_wait, the wait should be 663 * finished, and no new wait object should have been added. 664 */ 665 lockmgr(&bdev->fence_lock, LK_EXCLUSIVE); 666 ret = ttm_bo_wait(bo, false, false, true); 667 WARN_ON(ret); 668 lockmgr(&bdev->fence_lock, LK_RELEASE); 669 if (ret) 670 return ret; 671 672 lockmgr(&glob->lru_lock, LK_EXCLUSIVE); 673 ret = ttm_bo_reserve_nolru(bo, false, true, false, 0); 674 675 /* 676 * We raced, and lost, someone else holds the reservation now, 677 * and is probably busy in ttm_bo_cleanup_memtype_use. 678 * 679 * Even if it's not the case, because we finished waiting any 680 * delayed destruction would succeed, so just return success 681 * here. 682 */ 683 if (ret) { 684 lockmgr(&glob->lru_lock, LK_RELEASE); 685 return 0; 686 } 687 } else 688 lockmgr(&bdev->fence_lock, LK_RELEASE); 689 690 if (ret || unlikely(list_empty(&bo->ddestroy))) { 691 ttm_bo_unreserve_core(bo); 692 lockmgr(&glob->lru_lock, LK_RELEASE); 693 return ret; 694 } 695 696 put_count = ttm_bo_del_from_lru(bo); 697 list_del_init(&bo->ddestroy); 698 ++put_count; 699 700 lockmgr(&glob->lru_lock, LK_RELEASE); 701 ttm_bo_cleanup_memtype_use(bo); 702 703 ttm_bo_list_ref_sub(bo, put_count, true); 704 705 return 0; 706 } 707 708 /** 709 * Traverse the delayed list, and call ttm_bo_cleanup_refs on all 710 * encountered buffers. 711 */ 712 713 static int ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all) 714 { 715 struct ttm_bo_global *glob = bdev->glob; 716 struct ttm_buffer_object *entry = NULL; 717 int ret = 0; 718 719 lockmgr(&glob->lru_lock, LK_EXCLUSIVE); 720 if (list_empty(&bdev->ddestroy)) 721 goto out_unlock; 722 723 entry = list_first_entry(&bdev->ddestroy, 724 struct ttm_buffer_object, ddestroy); 725 kref_get(&entry->list_kref); 726 727 for (;;) { 728 struct ttm_buffer_object *nentry = NULL; 729 730 if (entry->ddestroy.next != &bdev->ddestroy) { 731 nentry = list_first_entry(&entry->ddestroy, 732 struct ttm_buffer_object, ddestroy); 733 kref_get(&nentry->list_kref); 734 } 735 736 ret = ttm_bo_reserve_nolru(entry, false, true, false, 0); 737 if (remove_all && ret) { 738 lockmgr(&glob->lru_lock, LK_RELEASE); 739 ret = ttm_bo_reserve_nolru(entry, false, false, 740 false, 0); 741 lockmgr(&glob->lru_lock, LK_EXCLUSIVE); 742 } 743 744 if (!ret) 745 ret = ttm_bo_cleanup_refs_and_unlock(entry, false, 746 !remove_all); 747 else 748 lockmgr(&glob->lru_lock, LK_RELEASE); 749 750 kref_put(&entry->list_kref, ttm_bo_release_list); 751 entry = nentry; 752 753 if (ret || !entry) 754 goto out; 755 756 lockmgr(&glob->lru_lock, LK_EXCLUSIVE); 757 if (list_empty(&entry->ddestroy)) 758 break; 759 } 760 761 out_unlock: 762 lockmgr(&glob->lru_lock, LK_RELEASE); 763 out: 764 if (entry) 765 kref_put(&entry->list_kref, ttm_bo_release_list); 766 return ret; 767 } 768 769 static void ttm_bo_delayed_workqueue(struct work_struct *work) 770 { 771 struct ttm_bo_device *bdev = 772 container_of(work, struct ttm_bo_device, wq.work); 773 774 if (ttm_bo_delayed_delete(bdev, false)) { 775 schedule_delayed_work(&bdev->wq, 776 ((hz / 100) < 1) ? 1 : hz / 100); 777 } 778 } 779 780 /* 781 * NOTE: bdev->vm_lock already held on call, this function release it. 782 */ 783 static void ttm_bo_release(struct kref *kref) 784 { 785 struct ttm_buffer_object *bo = 786 container_of(kref, struct ttm_buffer_object, kref); 787 struct ttm_bo_device *bdev = bo->bdev; 788 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type]; 789 int release_active; 790 791 if (atomic_read(&bo->kref.refcount) > 0) { 792 lockmgr(&bdev->vm_lock, LK_RELEASE); 793 return; 794 } 795 if (likely(bo->vm_node != NULL)) { 796 RB_REMOVE(ttm_bo_device_buffer_objects, 797 &bdev->addr_space_rb, bo); 798 drm_mm_put_block(bo->vm_node); 799 bo->vm_node = NULL; 800 } 801 802 /* 803 * Should we clean up our implied list_kref? Because ttm_bo_release() 804 * can be called reentrantly due to races (this may not be true any 805 * more with the lock management changes in the deref), it is possible 806 * to get here twice, but there's only one list_kref ref to drop and 807 * in the other path 'bo' can be kfree()d by another thread the 808 * instant we release our lock. 809 */ 810 release_active = test_bit(TTM_BO_PRIV_FLAG_ACTIVE, &bo->priv_flags); 811 if (release_active) { 812 clear_bit(TTM_BO_PRIV_FLAG_ACTIVE, &bo->priv_flags); 813 lockmgr(&bdev->vm_lock, LK_RELEASE); 814 ttm_mem_io_lock(man, false); 815 ttm_mem_io_free_vm(bo); 816 ttm_mem_io_unlock(man); 817 ttm_bo_cleanup_refs_or_queue(bo); 818 kref_put(&bo->list_kref, ttm_bo_release_list); 819 } else { 820 lockmgr(&bdev->vm_lock, LK_RELEASE); 821 } 822 } 823 824 void ttm_bo_unref(struct ttm_buffer_object **p_bo) 825 { 826 struct ttm_buffer_object *bo = *p_bo; 827 struct ttm_bo_device *bdev = bo->bdev; 828 829 *p_bo = NULL; 830 lockmgr(&bdev->vm_lock, LK_EXCLUSIVE); 831 if (kref_put(&bo->kref, ttm_bo_release) == 0) 832 lockmgr(&bdev->vm_lock, LK_RELEASE); 833 } 834 EXPORT_SYMBOL(ttm_bo_unref); 835 836 int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev) 837 { 838 return cancel_delayed_work_sync(&bdev->wq); 839 } 840 EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue); 841 842 void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched) 843 { 844 if (resched) 845 schedule_delayed_work(&bdev->wq, 846 ((hz / 100) < 1) ? 1 : hz / 100); 847 } 848 EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue); 849 850 static int ttm_bo_evict(struct ttm_buffer_object *bo, bool interruptible, 851 bool no_wait_gpu) 852 { 853 struct ttm_bo_device *bdev = bo->bdev; 854 struct ttm_mem_reg evict_mem; 855 struct ttm_placement placement; 856 int ret = 0; 857 858 lockmgr(&bdev->fence_lock, LK_EXCLUSIVE); 859 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu); 860 lockmgr(&bdev->fence_lock, LK_RELEASE); 861 862 if (unlikely(ret != 0)) { 863 if (ret != -ERESTARTSYS) { 864 pr_err("Failed to expire sync object before buffer eviction\n"); 865 } 866 goto out; 867 } 868 869 BUG_ON(!ttm_bo_is_reserved(bo)); 870 871 evict_mem = bo->mem; 872 evict_mem.mm_node = NULL; 873 evict_mem.bus.io_reserved_vm = false; 874 evict_mem.bus.io_reserved_count = 0; 875 876 placement.num_placement = 0; 877 placement.num_busy_placement = 0; 878 bdev->driver->evict_flags(bo, &placement); 879 ret = ttm_bo_mem_space(bo, &placement, &evict_mem, interruptible, 880 no_wait_gpu); 881 if (ret) { 882 if (ret != -ERESTARTSYS) { 883 pr_err("Failed to find memory space for buffer 0x%p eviction\n", 884 bo); 885 ttm_bo_mem_space_debug(bo, &placement); 886 } 887 goto out; 888 } 889 890 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, interruptible, 891 no_wait_gpu); 892 if (ret) { 893 if (ret != -ERESTARTSYS) 894 pr_err("Buffer eviction failed\n"); 895 ttm_bo_mem_put(bo, &evict_mem); 896 goto out; 897 } 898 bo->evicted = true; 899 out: 900 return ret; 901 } 902 903 static int ttm_mem_evict_first(struct ttm_bo_device *bdev, 904 uint32_t mem_type, 905 bool interruptible, 906 bool no_wait_gpu) 907 { 908 struct ttm_bo_global *glob = bdev->glob; 909 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 910 struct ttm_buffer_object *bo; 911 int ret = -EBUSY, put_count; 912 913 lockmgr(&glob->lru_lock, LK_EXCLUSIVE); 914 list_for_each_entry(bo, &man->lru, lru) { 915 ret = ttm_bo_reserve_nolru(bo, false, true, false, 0); 916 if (!ret) 917 break; 918 } 919 920 if (ret) { 921 lockmgr(&glob->lru_lock, LK_RELEASE); 922 return ret; 923 } 924 925 kref_get(&bo->list_kref); 926 927 if (!list_empty(&bo->ddestroy)) { 928 ret = ttm_bo_cleanup_refs_and_unlock(bo, interruptible, 929 no_wait_gpu); 930 kref_put(&bo->list_kref, ttm_bo_release_list); 931 return ret; 932 } 933 934 put_count = ttm_bo_del_from_lru(bo); 935 lockmgr(&glob->lru_lock, LK_RELEASE); 936 937 BUG_ON(ret != 0); 938 939 ttm_bo_list_ref_sub(bo, put_count, true); 940 941 ret = ttm_bo_evict(bo, interruptible, no_wait_gpu); 942 ttm_bo_unreserve(bo); 943 944 kref_put(&bo->list_kref, ttm_bo_release_list); 945 return ret; 946 } 947 948 void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem) 949 { 950 struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type]; 951 952 if (mem->mm_node) 953 (*man->func->put_node)(man, mem); 954 } 955 EXPORT_SYMBOL(ttm_bo_mem_put); 956 957 /** 958 * Repeatedly evict memory from the LRU for @mem_type until we create enough 959 * space, or we've evicted everything and there isn't enough space. 960 */ 961 static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo, 962 uint32_t mem_type, 963 const struct ttm_place *place, 964 struct ttm_mem_reg *mem, 965 bool interruptible, 966 bool no_wait_gpu) 967 { 968 struct ttm_bo_device *bdev = bo->bdev; 969 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 970 int ret; 971 972 do { 973 ret = (*man->func->get_node)(man, bo, place, mem); 974 if (unlikely(ret != 0)) 975 return ret; 976 if (mem->mm_node) 977 break; 978 ret = ttm_mem_evict_first(bdev, mem_type, 979 interruptible, no_wait_gpu); 980 if (unlikely(ret != 0)) 981 return ret; 982 } while (1); 983 if (mem->mm_node == NULL) 984 return -ENOMEM; 985 mem->mem_type = mem_type; 986 return 0; 987 } 988 989 static uint32_t ttm_bo_select_caching(struct ttm_mem_type_manager *man, 990 uint32_t cur_placement, 991 uint32_t proposed_placement) 992 { 993 uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING; 994 uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING; 995 996 /** 997 * Keep current caching if possible. 998 */ 999 1000 if ((cur_placement & caching) != 0) 1001 result |= (cur_placement & caching); 1002 else if ((man->default_caching & caching) != 0) 1003 result |= man->default_caching; 1004 else if ((TTM_PL_FLAG_CACHED & caching) != 0) 1005 result |= TTM_PL_FLAG_CACHED; 1006 else if ((TTM_PL_FLAG_WC & caching) != 0) 1007 result |= TTM_PL_FLAG_WC; 1008 else if ((TTM_PL_FLAG_UNCACHED & caching) != 0) 1009 result |= TTM_PL_FLAG_UNCACHED; 1010 1011 return result; 1012 } 1013 1014 static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man, 1015 uint32_t mem_type, 1016 const struct ttm_place *place, 1017 uint32_t *masked_placement) 1018 { 1019 uint32_t cur_flags = ttm_bo_type_flags(mem_type); 1020 1021 if ((cur_flags & place->flags & TTM_PL_MASK_MEM) == 0) 1022 return false; 1023 1024 if ((place->flags & man->available_caching) == 0) 1025 return false; 1026 1027 cur_flags |= (place->flags & man->available_caching); 1028 1029 *masked_placement = cur_flags; 1030 return true; 1031 } 1032 1033 /** 1034 * Creates space for memory region @mem according to its type. 1035 * 1036 * This function first searches for free space in compatible memory types in 1037 * the priority order defined by the driver. If free space isn't found, then 1038 * ttm_bo_mem_force_space is attempted in priority order to evict and find 1039 * space. 1040 */ 1041 int ttm_bo_mem_space(struct ttm_buffer_object *bo, 1042 struct ttm_placement *placement, 1043 struct ttm_mem_reg *mem, 1044 bool interruptible, 1045 bool no_wait_gpu) 1046 { 1047 struct ttm_bo_device *bdev = bo->bdev; 1048 struct ttm_mem_type_manager *man; 1049 uint32_t mem_type = TTM_PL_SYSTEM; 1050 uint32_t cur_flags = 0; 1051 bool type_found = false; 1052 bool type_ok = false; 1053 bool has_erestartsys = false; 1054 int i, ret; 1055 1056 mem->mm_node = NULL; 1057 for (i = 0; i < placement->num_placement; ++i) { 1058 const struct ttm_place *place = &placement->placement[i]; 1059 1060 ret = ttm_mem_type_from_place(place, &mem_type); 1061 if (ret) 1062 return ret; 1063 man = &bdev->man[mem_type]; 1064 1065 type_ok = ttm_bo_mt_compatible(man, mem_type, place, 1066 &cur_flags); 1067 1068 if (!type_ok) 1069 continue; 1070 1071 cur_flags = ttm_bo_select_caching(man, bo->mem.placement, 1072 cur_flags); 1073 /* 1074 * Use the access and other non-mapping-related flag bits from 1075 * the memory placement flags to the current flags 1076 */ 1077 ttm_flag_masked(&cur_flags, place->flags, 1078 ~TTM_PL_MASK_MEMTYPE); 1079 1080 if (mem_type == TTM_PL_SYSTEM) 1081 break; 1082 1083 if (man->has_type && man->use_type) { 1084 type_found = true; 1085 ret = (*man->func->get_node)(man, bo, place, mem); 1086 if (unlikely(ret)) 1087 return ret; 1088 } 1089 if (mem->mm_node) 1090 break; 1091 } 1092 1093 if ((type_ok && (mem_type == TTM_PL_SYSTEM)) || mem->mm_node) { 1094 mem->mem_type = mem_type; 1095 mem->placement = cur_flags; 1096 return 0; 1097 } 1098 1099 if (!type_found) 1100 return -EINVAL; 1101 1102 for (i = 0; i < placement->num_busy_placement; ++i) { 1103 const struct ttm_place *place = &placement->busy_placement[i]; 1104 1105 ret = ttm_mem_type_from_place(place, &mem_type); 1106 if (ret) 1107 return ret; 1108 man = &bdev->man[mem_type]; 1109 if (!man->has_type) 1110 continue; 1111 if (!ttm_bo_mt_compatible(man, mem_type, place, &cur_flags)) 1112 continue; 1113 1114 cur_flags = ttm_bo_select_caching(man, bo->mem.placement, 1115 cur_flags); 1116 /* 1117 * Use the access and other non-mapping-related flag bits from 1118 * the memory placement flags to the current flags 1119 */ 1120 ttm_flag_masked(&cur_flags, place->flags, 1121 ~TTM_PL_MASK_MEMTYPE); 1122 1123 if (mem_type == TTM_PL_SYSTEM) { 1124 mem->mem_type = mem_type; 1125 mem->placement = cur_flags; 1126 mem->mm_node = NULL; 1127 return 0; 1128 } 1129 1130 ret = ttm_bo_mem_force_space(bo, mem_type, place, mem, 1131 interruptible, no_wait_gpu); 1132 if (ret == 0 && mem->mm_node) { 1133 mem->placement = cur_flags; 1134 return 0; 1135 } 1136 if (ret == -ERESTARTSYS) 1137 has_erestartsys = true; 1138 } 1139 ret = (has_erestartsys) ? -ERESTARTSYS : -ENOMEM; 1140 return ret; 1141 } 1142 EXPORT_SYMBOL(ttm_bo_mem_space); 1143 1144 static int ttm_bo_move_buffer(struct ttm_buffer_object *bo, 1145 struct ttm_placement *placement, 1146 bool interruptible, 1147 bool no_wait_gpu) 1148 { 1149 int ret = 0; 1150 struct ttm_mem_reg mem; 1151 struct ttm_bo_device *bdev = bo->bdev; 1152 1153 BUG_ON(!ttm_bo_is_reserved(bo)); 1154 1155 /* 1156 * FIXME: It's possible to pipeline buffer moves. 1157 * Have the driver move function wait for idle when necessary, 1158 * instead of doing it here. 1159 */ 1160 lockmgr(&bdev->fence_lock, LK_EXCLUSIVE); 1161 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu); 1162 lockmgr(&bdev->fence_lock, LK_RELEASE); 1163 if (ret) 1164 return ret; 1165 mem.num_pages = bo->num_pages; 1166 mem.size = mem.num_pages << PAGE_SHIFT; 1167 mem.page_alignment = bo->mem.page_alignment; 1168 mem.bus.io_reserved_vm = false; 1169 mem.bus.io_reserved_count = 0; 1170 /* 1171 * Determine where to move the buffer. 1172 */ 1173 ret = ttm_bo_mem_space(bo, placement, &mem, 1174 interruptible, no_wait_gpu); 1175 if (ret) 1176 goto out_unlock; 1177 ret = ttm_bo_handle_move_mem(bo, &mem, false, 1178 interruptible, no_wait_gpu); 1179 out_unlock: 1180 if (ret && mem.mm_node) 1181 ttm_bo_mem_put(bo, &mem); 1182 return ret; 1183 } 1184 1185 static bool ttm_bo_mem_compat(struct ttm_placement *placement, 1186 struct ttm_mem_reg *mem, 1187 uint32_t *new_flags) 1188 { 1189 int i; 1190 1191 for (i = 0; i < placement->num_placement; i++) { 1192 const struct ttm_place *heap = &placement->placement[i]; 1193 if (mem->mm_node && 1194 (mem->start < heap->fpfn || 1195 (heap->lpfn != 0 && (mem->start + mem->num_pages) > heap->lpfn))) 1196 continue; 1197 1198 *new_flags = heap->flags; 1199 if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) && 1200 (*new_flags & mem->placement & TTM_PL_MASK_MEM)) 1201 return true; 1202 } 1203 1204 for (i = 0; i < placement->num_busy_placement; i++) { 1205 const struct ttm_place *heap = &placement->busy_placement[i]; 1206 if (mem->mm_node && 1207 (mem->start < heap->fpfn || 1208 (heap->lpfn != 0 && (mem->start + mem->num_pages) > heap->lpfn))) 1209 continue; 1210 1211 *new_flags = heap->flags; 1212 if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) && 1213 (*new_flags & mem->placement & TTM_PL_MASK_MEM)) 1214 return true; 1215 } 1216 1217 return false; 1218 } 1219 1220 int ttm_bo_validate(struct ttm_buffer_object *bo, 1221 struct ttm_placement *placement, 1222 bool interruptible, 1223 bool no_wait_gpu) 1224 { 1225 int ret; 1226 uint32_t new_flags; 1227 1228 BUG_ON(!ttm_bo_is_reserved(bo)); 1229 /* 1230 * Check whether we need to move buffer. 1231 */ 1232 if (!ttm_bo_mem_compat(placement, &bo->mem, &new_flags)) { 1233 ret = ttm_bo_move_buffer(bo, placement, interruptible, 1234 no_wait_gpu); 1235 if (ret) 1236 return ret; 1237 } else { 1238 /* 1239 * Use the access and other non-mapping-related flag bits from 1240 * the compatible memory placement flags to the active flags 1241 */ 1242 ttm_flag_masked(&bo->mem.placement, new_flags, 1243 ~TTM_PL_MASK_MEMTYPE); 1244 } 1245 /* 1246 * We might need to add a TTM. 1247 */ 1248 if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) { 1249 ret = ttm_bo_add_ttm(bo, true); 1250 if (ret) 1251 return ret; 1252 } 1253 return 0; 1254 } 1255 EXPORT_SYMBOL(ttm_bo_validate); 1256 1257 int ttm_bo_init(struct ttm_bo_device *bdev, 1258 struct ttm_buffer_object *bo, 1259 unsigned long size, 1260 enum ttm_bo_type type, 1261 struct ttm_placement *placement, 1262 uint32_t page_alignment, 1263 bool interruptible, 1264 struct vm_object *persistent_swap_storage, 1265 size_t acc_size, 1266 struct sg_table *sg, 1267 void (*destroy) (struct ttm_buffer_object *)) 1268 { 1269 int ret = 0; 1270 unsigned long num_pages; 1271 struct ttm_mem_global *mem_glob = bdev->glob->mem_glob; 1272 1273 ret = ttm_mem_global_alloc(mem_glob, acc_size, false, false); 1274 if (ret) { 1275 pr_err("Out of kernel memory\n"); 1276 if (destroy) 1277 (*destroy)(bo); 1278 else 1279 kfree(bo); 1280 return -ENOMEM; 1281 } 1282 1283 num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT; 1284 if (num_pages == 0) { 1285 pr_err("Illegal buffer object size\n"); 1286 if (destroy) 1287 (*destroy)(bo); 1288 else 1289 kfree(bo); 1290 ttm_mem_global_free(mem_glob, acc_size); 1291 return -EINVAL; 1292 } 1293 bo->destroy = destroy; 1294 1295 kref_init(&bo->kref); 1296 kref_init(&bo->list_kref); 1297 atomic_set(&bo->cpu_writers, 0); 1298 atomic_set(&bo->reserved, 1); 1299 init_waitqueue_head(&bo->event_queue); 1300 INIT_LIST_HEAD(&bo->lru); 1301 INIT_LIST_HEAD(&bo->ddestroy); 1302 INIT_LIST_HEAD(&bo->swap); 1303 INIT_LIST_HEAD(&bo->io_reserve_lru); 1304 /*bzero(&bo->vm_rb, sizeof(bo->vm_rb));*/ 1305 bo->bdev = bdev; 1306 bo->glob = bdev->glob; 1307 bo->type = type; 1308 bo->num_pages = num_pages; 1309 bo->mem.size = num_pages << PAGE_SHIFT; 1310 bo->mem.mem_type = TTM_PL_SYSTEM; 1311 bo->mem.num_pages = bo->num_pages; 1312 bo->mem.mm_node = NULL; 1313 bo->mem.page_alignment = page_alignment; 1314 bo->mem.bus.io_reserved_vm = false; 1315 bo->mem.bus.io_reserved_count = 0; 1316 bo->priv_flags = 0; 1317 bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED); 1318 bo->seq_valid = false; 1319 bo->persistent_swap_storage = persistent_swap_storage; 1320 bo->acc_size = acc_size; 1321 bo->sg = sg; 1322 atomic_inc(&bo->glob->bo_count); 1323 1324 /* 1325 * Mirror ref from kref_init() for list_kref. 1326 */ 1327 set_bit(TTM_BO_PRIV_FLAG_ACTIVE, &bo->priv_flags); 1328 1329 /* 1330 * For ttm_bo_type_device buffers, allocate 1331 * address space from the device. 1332 */ 1333 if (bo->type == ttm_bo_type_device || 1334 bo->type == ttm_bo_type_sg) { 1335 ret = ttm_bo_setup_vm(bo); 1336 if (ret) 1337 goto out_err; 1338 } 1339 1340 ret = ttm_bo_validate(bo, placement, interruptible, false); 1341 if (ret) 1342 goto out_err; 1343 1344 ttm_bo_unreserve(bo); 1345 return 0; 1346 1347 out_err: 1348 ttm_bo_unreserve(bo); 1349 ttm_bo_unref(&bo); 1350 1351 return ret; 1352 } 1353 EXPORT_SYMBOL(ttm_bo_init); 1354 1355 size_t ttm_bo_acc_size(struct ttm_bo_device *bdev, 1356 unsigned long bo_size, 1357 unsigned struct_size) 1358 { 1359 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT; 1360 size_t size = 0; 1361 1362 size += ttm_round_pot(struct_size); 1363 size += PAGE_ALIGN(npages * sizeof(void *)); 1364 size += ttm_round_pot(sizeof(struct ttm_tt)); 1365 return size; 1366 } 1367 EXPORT_SYMBOL(ttm_bo_acc_size); 1368 1369 size_t ttm_bo_dma_acc_size(struct ttm_bo_device *bdev, 1370 unsigned long bo_size, 1371 unsigned struct_size) 1372 { 1373 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT; 1374 size_t size = 0; 1375 1376 size += ttm_round_pot(struct_size); 1377 size += PAGE_ALIGN(npages * sizeof(void *)); 1378 size += PAGE_ALIGN(npages * sizeof(dma_addr_t)); 1379 size += ttm_round_pot(sizeof(struct ttm_dma_tt)); 1380 return size; 1381 } 1382 EXPORT_SYMBOL(ttm_bo_dma_acc_size); 1383 1384 int ttm_bo_create(struct ttm_bo_device *bdev, 1385 unsigned long size, 1386 enum ttm_bo_type type, 1387 struct ttm_placement *placement, 1388 uint32_t page_alignment, 1389 bool interruptible, 1390 struct vm_object *persistent_swap_storage, 1391 struct ttm_buffer_object **p_bo) 1392 { 1393 struct ttm_buffer_object *bo; 1394 size_t acc_size; 1395 int ret; 1396 1397 *p_bo = NULL; 1398 bo = kzalloc(sizeof(*bo), GFP_KERNEL); 1399 if (unlikely(bo == NULL)) 1400 return -ENOMEM; 1401 1402 acc_size = ttm_bo_acc_size(bdev, size, sizeof(struct ttm_buffer_object)); 1403 ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment, 1404 interruptible, persistent_swap_storage, acc_size, 1405 NULL, NULL); 1406 if (likely(ret == 0)) 1407 *p_bo = bo; 1408 1409 return ret; 1410 } 1411 EXPORT_SYMBOL(ttm_bo_create); 1412 1413 static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev, 1414 unsigned mem_type, bool allow_errors) 1415 { 1416 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 1417 struct ttm_bo_global *glob = bdev->glob; 1418 int ret; 1419 1420 /* 1421 * Can't use standard list traversal since we're unlocking. 1422 */ 1423 1424 lockmgr(&glob->lru_lock, LK_EXCLUSIVE); 1425 while (!list_empty(&man->lru)) { 1426 lockmgr(&glob->lru_lock, LK_RELEASE); 1427 ret = ttm_mem_evict_first(bdev, mem_type, false, false); 1428 if (ret) { 1429 if (allow_errors) { 1430 return ret; 1431 } else { 1432 pr_err("Cleanup eviction failed\n"); 1433 } 1434 } 1435 lockmgr(&glob->lru_lock, LK_EXCLUSIVE); 1436 } 1437 lockmgr(&glob->lru_lock, LK_RELEASE); 1438 return 0; 1439 } 1440 1441 int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type) 1442 { 1443 struct ttm_mem_type_manager *man; 1444 int ret = -EINVAL; 1445 1446 if (mem_type >= TTM_NUM_MEM_TYPES) { 1447 pr_err("Illegal memory type %d\n", mem_type); 1448 return ret; 1449 } 1450 man = &bdev->man[mem_type]; 1451 1452 if (!man->has_type) { 1453 pr_err("Trying to take down uninitialized memory manager type %u\n", 1454 mem_type); 1455 return ret; 1456 } 1457 1458 man->use_type = false; 1459 man->has_type = false; 1460 1461 ret = 0; 1462 if (mem_type > 0) { 1463 ttm_bo_force_list_clean(bdev, mem_type, false); 1464 1465 ret = (*man->func->takedown)(man); 1466 } 1467 1468 return ret; 1469 } 1470 EXPORT_SYMBOL(ttm_bo_clean_mm); 1471 1472 int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type) 1473 { 1474 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 1475 1476 if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) { 1477 pr_err("Illegal memory manager memory type %u\n", mem_type); 1478 return -EINVAL; 1479 } 1480 1481 if (!man->has_type) { 1482 pr_err("Memory type %u has not been initialized\n", mem_type); 1483 return 0; 1484 } 1485 1486 return ttm_bo_force_list_clean(bdev, mem_type, true); 1487 } 1488 EXPORT_SYMBOL(ttm_bo_evict_mm); 1489 1490 int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type, 1491 unsigned long p_size) 1492 { 1493 int ret = -EINVAL; 1494 struct ttm_mem_type_manager *man; 1495 1496 BUG_ON(type >= TTM_NUM_MEM_TYPES); 1497 man = &bdev->man[type]; 1498 BUG_ON(man->has_type); 1499 man->io_reserve_fastpath = true; 1500 man->use_io_reserve_lru = false; 1501 lockinit(&man->io_reserve_mutex, "ttmman", 0, LK_CANRECURSE); 1502 INIT_LIST_HEAD(&man->io_reserve_lru); 1503 1504 ret = bdev->driver->init_mem_type(bdev, type, man); 1505 if (ret) 1506 return ret; 1507 man->bdev = bdev; 1508 1509 ret = 0; 1510 if (type != TTM_PL_SYSTEM) { 1511 ret = (*man->func->init)(man, p_size); 1512 if (ret) 1513 return ret; 1514 } 1515 man->has_type = true; 1516 man->use_type = true; 1517 man->size = p_size; 1518 1519 INIT_LIST_HEAD(&man->lru); 1520 1521 return 0; 1522 } 1523 EXPORT_SYMBOL(ttm_bo_init_mm); 1524 1525 static void ttm_bo_global_kobj_release(struct ttm_bo_global *glob) 1526 { 1527 ttm_mem_unregister_shrink(glob->mem_glob, &glob->shrink); 1528 vm_page_free_contig(glob->dummy_read_page, PAGE_SIZE); 1529 glob->dummy_read_page = NULL; 1530 /* 1531 vm_page_free(glob->dummy_read_page); 1532 */ 1533 } 1534 1535 void ttm_bo_global_release(struct drm_global_reference *ref) 1536 { 1537 struct ttm_bo_global *glob = ref->object; 1538 1539 if (refcount_release(&glob->kobj_ref)) 1540 ttm_bo_global_kobj_release(glob); 1541 } 1542 EXPORT_SYMBOL(ttm_bo_global_release); 1543 1544 int ttm_bo_global_init(struct drm_global_reference *ref) 1545 { 1546 struct ttm_bo_global_ref *bo_ref = 1547 container_of(ref, struct ttm_bo_global_ref, ref); 1548 struct ttm_bo_global *glob = ref->object; 1549 int ret; 1550 1551 lockinit(&glob->device_list_mutex, "ttmdlm", 0, LK_CANRECURSE); 1552 lockinit(&glob->lru_lock, "ttmlru", 0, LK_CANRECURSE); 1553 glob->mem_glob = bo_ref->mem_glob; 1554 glob->dummy_read_page = vm_page_alloc_contig( 1555 0, VM_MAX_ADDRESS, PAGE_SIZE, 0, 1*PAGE_SIZE, VM_MEMATTR_UNCACHEABLE); 1556 1557 if (unlikely(glob->dummy_read_page == NULL)) { 1558 ret = -ENOMEM; 1559 goto out_no_drp; 1560 } 1561 1562 INIT_LIST_HEAD(&glob->swap_lru); 1563 INIT_LIST_HEAD(&glob->device_list); 1564 1565 ttm_mem_init_shrink(&glob->shrink, ttm_bo_swapout); 1566 ret = ttm_mem_register_shrink(glob->mem_glob, &glob->shrink); 1567 if (unlikely(ret != 0)) { 1568 pr_err("Could not register buffer object swapout\n"); 1569 goto out_no_shrink; 1570 } 1571 1572 atomic_set(&glob->bo_count, 0); 1573 1574 refcount_init(&glob->kobj_ref, 1); 1575 return (0); 1576 1577 out_no_shrink: 1578 vm_page_free_contig(glob->dummy_read_page, PAGE_SIZE); 1579 glob->dummy_read_page = NULL; 1580 /* 1581 vm_page_free(glob->dummy_read_page); 1582 */ 1583 out_no_drp: 1584 kfree(glob); 1585 return ret; 1586 } 1587 EXPORT_SYMBOL(ttm_bo_global_init); 1588 1589 1590 int ttm_bo_device_release(struct ttm_bo_device *bdev) 1591 { 1592 int ret = 0; 1593 unsigned i = TTM_NUM_MEM_TYPES; 1594 struct ttm_mem_type_manager *man; 1595 struct ttm_bo_global *glob = bdev->glob; 1596 1597 while (i--) { 1598 man = &bdev->man[i]; 1599 if (man->has_type) { 1600 man->use_type = false; 1601 if ((i != TTM_PL_SYSTEM) && ttm_bo_clean_mm(bdev, i)) { 1602 ret = -EBUSY; 1603 pr_err("DRM memory manager type %d is not clean\n", 1604 i); 1605 } 1606 man->has_type = false; 1607 } 1608 } 1609 1610 lockmgr(&glob->device_list_mutex, LK_EXCLUSIVE); 1611 list_del(&bdev->device_list); 1612 lockmgr(&glob->device_list_mutex, LK_RELEASE); 1613 1614 cancel_delayed_work_sync(&bdev->wq); 1615 1616 while (ttm_bo_delayed_delete(bdev, true)) 1617 ; 1618 1619 lockmgr(&glob->lru_lock, LK_EXCLUSIVE); 1620 if (list_empty(&bdev->ddestroy)) 1621 TTM_DEBUG("Delayed destroy list was clean\n"); 1622 1623 if (list_empty(&bdev->man[0].lru)) 1624 TTM_DEBUG("Swap list was clean\n"); 1625 lockmgr(&glob->lru_lock, LK_RELEASE); 1626 1627 BUG_ON(!drm_mm_clean(&bdev->addr_space_mm)); 1628 lockmgr(&bdev->vm_lock, LK_EXCLUSIVE); 1629 drm_mm_takedown(&bdev->addr_space_mm); 1630 lockmgr(&bdev->vm_lock, LK_RELEASE); 1631 1632 return ret; 1633 } 1634 EXPORT_SYMBOL(ttm_bo_device_release); 1635 1636 int ttm_bo_device_init(struct ttm_bo_device *bdev, 1637 struct ttm_bo_global *glob, 1638 struct ttm_bo_driver *driver, 1639 uint64_t file_page_offset, 1640 bool need_dma32) 1641 { 1642 int ret = -EINVAL; 1643 1644 lockinit(&bdev->vm_lock, "ttmvml", 0, LK_CANRECURSE); 1645 bdev->driver = driver; 1646 1647 memset(bdev->man, 0, sizeof(bdev->man)); 1648 1649 /* 1650 * Initialize the system memory buffer type. 1651 * Other types need to be driver / IOCTL initialized. 1652 */ 1653 ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0); 1654 if (unlikely(ret != 0)) 1655 goto out_no_sys; 1656 1657 RB_INIT(&bdev->addr_space_rb); 1658 drm_mm_init(&bdev->addr_space_mm, file_page_offset, 0x10000000); 1659 1660 INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue); 1661 INIT_LIST_HEAD(&bdev->ddestroy); 1662 bdev->dev_mapping = NULL; 1663 bdev->glob = glob; 1664 bdev->need_dma32 = need_dma32; 1665 bdev->val_seq = 0; 1666 lockinit(&bdev->fence_lock, "ttmfence", 0, LK_CANRECURSE); 1667 lockmgr(&glob->device_list_mutex, LK_EXCLUSIVE); 1668 list_add_tail(&bdev->device_list, &glob->device_list); 1669 lockmgr(&glob->device_list_mutex, LK_RELEASE); 1670 1671 return 0; 1672 out_no_sys: 1673 return ret; 1674 } 1675 EXPORT_SYMBOL(ttm_bo_device_init); 1676 1677 /* 1678 * buffer object vm functions. 1679 */ 1680 1681 bool ttm_mem_reg_is_pci(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem) 1682 { 1683 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type]; 1684 1685 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) { 1686 if (mem->mem_type == TTM_PL_SYSTEM) 1687 return false; 1688 1689 if (man->flags & TTM_MEMTYPE_FLAG_CMA) 1690 return false; 1691 1692 if (mem->placement & TTM_PL_FLAG_CACHED) 1693 return false; 1694 } 1695 return true; 1696 } 1697 1698 void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo) 1699 { 1700 1701 ttm_bo_release_mmap(bo); 1702 ttm_mem_io_free_vm(bo); 1703 } 1704 1705 void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo) 1706 { 1707 struct ttm_bo_device *bdev = bo->bdev; 1708 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type]; 1709 1710 ttm_mem_io_lock(man, false); 1711 ttm_bo_unmap_virtual_locked(bo); 1712 ttm_mem_io_unlock(man); 1713 } 1714 1715 1716 EXPORT_SYMBOL(ttm_bo_unmap_virtual); 1717 1718 static void ttm_bo_vm_insert_rb(struct ttm_buffer_object *bo) 1719 { 1720 struct ttm_bo_device *bdev = bo->bdev; 1721 1722 /* The caller acquired bdev->vm_lock. */ 1723 RB_INSERT(ttm_bo_device_buffer_objects, &bdev->addr_space_rb, bo); 1724 } 1725 1726 /** 1727 * ttm_bo_setup_vm: 1728 * 1729 * @bo: the buffer to allocate address space for 1730 * 1731 * Allocate address space in the drm device so that applications 1732 * can mmap the buffer and access the contents. This only 1733 * applies to ttm_bo_type_device objects as others are not 1734 * placed in the drm device address space. 1735 */ 1736 1737 static int ttm_bo_setup_vm(struct ttm_buffer_object *bo) 1738 { 1739 struct ttm_bo_device *bdev = bo->bdev; 1740 int ret; 1741 1742 retry_pre_get: 1743 ret = drm_mm_pre_get(&bdev->addr_space_mm); 1744 if (unlikely(ret != 0)) 1745 return ret; 1746 1747 lockmgr(&bdev->vm_lock, LK_EXCLUSIVE); 1748 bo->vm_node = drm_mm_search_free(&bdev->addr_space_mm, 1749 bo->mem.num_pages, 0, 0); 1750 1751 if (unlikely(bo->vm_node == NULL)) { 1752 ret = -ENOMEM; 1753 goto out_unlock; 1754 } 1755 1756 bo->vm_node = drm_mm_get_block_atomic(bo->vm_node, 1757 bo->mem.num_pages, 0); 1758 1759 if (unlikely(bo->vm_node == NULL)) { 1760 lockmgr(&bdev->vm_lock, LK_RELEASE); 1761 goto retry_pre_get; 1762 } 1763 1764 ttm_bo_vm_insert_rb(bo); 1765 lockmgr(&bdev->vm_lock, LK_RELEASE); 1766 bo->addr_space_offset = ((uint64_t) bo->vm_node->start) << PAGE_SHIFT; 1767 1768 return 0; 1769 out_unlock: 1770 lockmgr(&bdev->vm_lock, LK_RELEASE); 1771 return ret; 1772 } 1773 1774 int ttm_bo_wait(struct ttm_buffer_object *bo, 1775 bool lazy, bool interruptible, bool no_wait) 1776 { 1777 struct ttm_bo_driver *driver = bo->bdev->driver; 1778 struct ttm_bo_device *bdev = bo->bdev; 1779 void *sync_obj; 1780 int ret = 0; 1781 1782 if (likely(bo->sync_obj == NULL)) 1783 return 0; 1784 1785 while (bo->sync_obj) { 1786 1787 if (driver->sync_obj_signaled(bo->sync_obj)) { 1788 void *tmp_obj = bo->sync_obj; 1789 bo->sync_obj = NULL; 1790 clear_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags); 1791 lockmgr(&bdev->fence_lock, LK_RELEASE); 1792 driver->sync_obj_unref(&tmp_obj); 1793 lockmgr(&bdev->fence_lock, LK_EXCLUSIVE); 1794 continue; 1795 } 1796 1797 if (no_wait) 1798 return -EBUSY; 1799 1800 sync_obj = driver->sync_obj_ref(bo->sync_obj); 1801 lockmgr(&bdev->fence_lock, LK_RELEASE); 1802 ret = driver->sync_obj_wait(sync_obj, 1803 lazy, interruptible); 1804 if (unlikely(ret != 0)) { 1805 driver->sync_obj_unref(&sync_obj); 1806 lockmgr(&bdev->fence_lock, LK_EXCLUSIVE); 1807 return ret; 1808 } 1809 lockmgr(&bdev->fence_lock, LK_EXCLUSIVE); 1810 if (likely(bo->sync_obj == sync_obj)) { 1811 void *tmp_obj = bo->sync_obj; 1812 bo->sync_obj = NULL; 1813 clear_bit(TTM_BO_PRIV_FLAG_MOVING, 1814 &bo->priv_flags); 1815 lockmgr(&bdev->fence_lock, LK_RELEASE); 1816 driver->sync_obj_unref(&sync_obj); 1817 driver->sync_obj_unref(&tmp_obj); 1818 lockmgr(&bdev->fence_lock, LK_EXCLUSIVE); 1819 } else { 1820 lockmgr(&bdev->fence_lock, LK_RELEASE); 1821 driver->sync_obj_unref(&sync_obj); 1822 lockmgr(&bdev->fence_lock, LK_EXCLUSIVE); 1823 } 1824 } 1825 return 0; 1826 } 1827 EXPORT_SYMBOL(ttm_bo_wait); 1828 1829 int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait) 1830 { 1831 struct ttm_bo_device *bdev = bo->bdev; 1832 int ret = 0; 1833 1834 /* 1835 * Using ttm_bo_reserve makes sure the lru lists are updated. 1836 */ 1837 1838 ret = ttm_bo_reserve(bo, true, no_wait, false, 0); 1839 if (unlikely(ret != 0)) 1840 return ret; 1841 lockmgr(&bdev->fence_lock, LK_EXCLUSIVE); 1842 ret = ttm_bo_wait(bo, false, true, no_wait); 1843 lockmgr(&bdev->fence_lock, LK_RELEASE); 1844 if (likely(ret == 0)) 1845 atomic_inc(&bo->cpu_writers); 1846 ttm_bo_unreserve(bo); 1847 return ret; 1848 } 1849 EXPORT_SYMBOL(ttm_bo_synccpu_write_grab); 1850 1851 void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo) 1852 { 1853 atomic_dec(&bo->cpu_writers); 1854 } 1855 EXPORT_SYMBOL(ttm_bo_synccpu_write_release); 1856 1857 /** 1858 * A buffer object shrink method that tries to swap out the first 1859 * buffer object on the bo_global::swap_lru list. 1860 */ 1861 1862 static int ttm_bo_swapout(struct ttm_mem_shrink *shrink) 1863 { 1864 struct ttm_bo_global *glob = 1865 container_of(shrink, struct ttm_bo_global, shrink); 1866 struct ttm_buffer_object *bo; 1867 int ret = -EBUSY; 1868 int put_count; 1869 uint32_t swap_placement = (TTM_PL_FLAG_CACHED | TTM_PL_FLAG_SYSTEM); 1870 1871 lockmgr(&glob->lru_lock, LK_EXCLUSIVE); 1872 list_for_each_entry(bo, &glob->swap_lru, swap) { 1873 ret = ttm_bo_reserve_nolru(bo, false, true, false, 0); 1874 if (!ret) 1875 break; 1876 } 1877 1878 if (ret) { 1879 lockmgr(&glob->lru_lock, LK_RELEASE); 1880 return ret; 1881 } 1882 1883 kref_get(&bo->list_kref); 1884 1885 if (!list_empty(&bo->ddestroy)) { 1886 ret = ttm_bo_cleanup_refs_and_unlock(bo, false, false); 1887 kref_put(&bo->list_kref, ttm_bo_release_list); 1888 return ret; 1889 } 1890 1891 put_count = ttm_bo_del_from_lru(bo); 1892 lockmgr(&glob->lru_lock, LK_RELEASE); 1893 1894 ttm_bo_list_ref_sub(bo, put_count, true); 1895 1896 /** 1897 * Wait for GPU, then move to system cached. 1898 */ 1899 1900 lockmgr(&bo->bdev->fence_lock, LK_EXCLUSIVE); 1901 ret = ttm_bo_wait(bo, false, false, false); 1902 lockmgr(&bo->bdev->fence_lock, LK_RELEASE); 1903 1904 if (unlikely(ret != 0)) 1905 goto out; 1906 1907 if ((bo->mem.placement & swap_placement) != swap_placement) { 1908 struct ttm_mem_reg evict_mem; 1909 1910 evict_mem = bo->mem; 1911 evict_mem.mm_node = NULL; 1912 evict_mem.placement = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED; 1913 evict_mem.mem_type = TTM_PL_SYSTEM; 1914 1915 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, 1916 false, false); 1917 if (unlikely(ret != 0)) 1918 goto out; 1919 } 1920 1921 ttm_bo_unmap_virtual(bo); 1922 1923 /** 1924 * Swap out. Buffer will be swapped in again as soon as 1925 * anyone tries to access a ttm page. 1926 */ 1927 1928 if (bo->bdev->driver->swap_notify) 1929 bo->bdev->driver->swap_notify(bo); 1930 1931 ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage); 1932 out: 1933 1934 /** 1935 * 1936 * Unreserve without putting on LRU to avoid swapping out an 1937 * already swapped buffer. 1938 */ 1939 1940 ttm_bo_unreserve_core(bo); 1941 kref_put(&bo->list_kref, ttm_bo_release_list); 1942 return ret; 1943 } 1944 1945 void ttm_bo_swapout_all(struct ttm_bo_device *bdev) 1946 { 1947 while (ttm_bo_swapout(&bdev->glob->shrink) == 0) 1948 ; 1949 } 1950 EXPORT_SYMBOL(ttm_bo_swapout_all); 1951