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/jiffies.h> 37 #include <linux/slab.h> 38 #include <linux/sched.h> 39 #include <linux/mm.h> 40 #include <linux/file.h> 41 #include <linux/module.h> 42 #include <linux/atomic.h> 43 44 #define TTM_ASSERT_LOCKED(param) 45 #define TTM_DEBUG(fmt, arg...) 46 #define TTM_BO_HASH_ORDER 13 47 48 static int ttm_bo_setup_vm(struct ttm_buffer_object *bo); 49 static int ttm_bo_swapout(struct ttm_mem_shrink *shrink); 50 static void ttm_bo_global_kobj_release(struct kobject *kobj); 51 52 static struct attribute ttm_bo_count = { 53 .name = "bo_count", 54 .mode = S_IRUGO 55 }; 56 57 static inline int ttm_mem_type_from_flags(uint32_t flags, uint32_t *mem_type) 58 { 59 int i; 60 61 for (i = 0; i <= TTM_PL_PRIV5; i++) 62 if (flags & (1 << i)) { 63 *mem_type = i; 64 return 0; 65 } 66 return -EINVAL; 67 } 68 69 static void ttm_mem_type_debug(struct ttm_bo_device *bdev, int mem_type) 70 { 71 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 72 73 pr_err(" has_type: %d\n", man->has_type); 74 pr_err(" use_type: %d\n", man->use_type); 75 pr_err(" flags: 0x%08X\n", man->flags); 76 pr_err(" gpu_offset: 0x%08lX\n", man->gpu_offset); 77 pr_err(" size: %llu\n", man->size); 78 pr_err(" available_caching: 0x%08X\n", man->available_caching); 79 pr_err(" default_caching: 0x%08X\n", man->default_caching); 80 if (mem_type != TTM_PL_SYSTEM) 81 (*man->func->debug)(man, TTM_PFX); 82 } 83 84 static void ttm_bo_mem_space_debug(struct ttm_buffer_object *bo, 85 struct ttm_placement *placement) 86 { 87 int i, ret, mem_type; 88 89 pr_err("No space for %p (%lu pages, %luK, %luM)\n", 90 bo, bo->mem.num_pages, bo->mem.size >> 10, 91 bo->mem.size >> 20); 92 for (i = 0; i < placement->num_placement; i++) { 93 ret = ttm_mem_type_from_flags(placement->placement[i], 94 &mem_type); 95 if (ret) 96 return; 97 pr_err(" placement[%d]=0x%08X (%d)\n", 98 i, placement->placement[i], mem_type); 99 ttm_mem_type_debug(bo->bdev, mem_type); 100 } 101 } 102 103 static ssize_t ttm_bo_global_show(struct kobject *kobj, 104 struct attribute *attr, 105 char *buffer) 106 { 107 struct ttm_bo_global *glob = 108 container_of(kobj, struct ttm_bo_global, kobj); 109 110 return snprintf(buffer, PAGE_SIZE, "%lu\n", 111 (unsigned long) atomic_read(&glob->bo_count)); 112 } 113 114 static struct attribute *ttm_bo_global_attrs[] = { 115 &ttm_bo_count, 116 NULL 117 }; 118 119 static const struct sysfs_ops ttm_bo_global_ops = { 120 .show = &ttm_bo_global_show 121 }; 122 123 static struct kobj_type ttm_bo_glob_kobj_type = { 124 .release = &ttm_bo_global_kobj_release, 125 .sysfs_ops = &ttm_bo_global_ops, 126 .default_attrs = ttm_bo_global_attrs 127 }; 128 129 130 static inline uint32_t ttm_bo_type_flags(unsigned type) 131 { 132 return 1 << (type); 133 } 134 135 static void ttm_bo_release_list(struct kref *list_kref) 136 { 137 struct ttm_buffer_object *bo = 138 container_of(list_kref, struct ttm_buffer_object, list_kref); 139 struct ttm_bo_device *bdev = bo->bdev; 140 size_t acc_size = bo->acc_size; 141 142 BUG_ON(atomic_read(&bo->list_kref.refcount)); 143 BUG_ON(atomic_read(&bo->kref.refcount)); 144 BUG_ON(atomic_read(&bo->cpu_writers)); 145 BUG_ON(bo->sync_obj != NULL); 146 BUG_ON(bo->mem.mm_node != NULL); 147 BUG_ON(!list_empty(&bo->lru)); 148 BUG_ON(!list_empty(&bo->ddestroy)); 149 150 if (bo->ttm) 151 ttm_tt_destroy(bo->ttm); 152 atomic_dec(&bo->glob->bo_count); 153 if (bo->destroy) 154 bo->destroy(bo); 155 else { 156 kfree(bo); 157 } 158 ttm_mem_global_free(bdev->glob->mem_glob, acc_size); 159 } 160 161 int ttm_bo_wait_unreserved(struct ttm_buffer_object *bo, bool interruptible) 162 { 163 if (interruptible) { 164 return wait_event_interruptible(bo->event_queue, 165 !ttm_bo_is_reserved(bo)); 166 } else { 167 wait_event(bo->event_queue, !ttm_bo_is_reserved(bo)); 168 return 0; 169 } 170 } 171 EXPORT_SYMBOL(ttm_bo_wait_unreserved); 172 173 void ttm_bo_add_to_lru(struct ttm_buffer_object *bo) 174 { 175 struct ttm_bo_device *bdev = bo->bdev; 176 struct ttm_mem_type_manager *man; 177 178 BUG_ON(!ttm_bo_is_reserved(bo)); 179 180 if (!(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) { 181 182 BUG_ON(!list_empty(&bo->lru)); 183 184 man = &bdev->man[bo->mem.mem_type]; 185 list_add_tail(&bo->lru, &man->lru); 186 kref_get(&bo->list_kref); 187 188 if (bo->ttm != NULL) { 189 list_add_tail(&bo->swap, &bo->glob->swap_lru); 190 kref_get(&bo->list_kref); 191 } 192 } 193 } 194 195 int ttm_bo_del_from_lru(struct ttm_buffer_object *bo) 196 { 197 int put_count = 0; 198 199 if (!list_empty(&bo->swap)) { 200 list_del_init(&bo->swap); 201 ++put_count; 202 } 203 if (!list_empty(&bo->lru)) { 204 list_del_init(&bo->lru); 205 ++put_count; 206 } 207 208 /* 209 * TODO: Add a driver hook to delete from 210 * driver-specific LRU's here. 211 */ 212 213 return put_count; 214 } 215 216 int ttm_bo_reserve_locked(struct ttm_buffer_object *bo, 217 bool interruptible, 218 bool no_wait, bool use_sequence, uint32_t sequence) 219 { 220 struct ttm_bo_global *glob = bo->glob; 221 int ret; 222 223 while (unlikely(atomic_read(&bo->reserved) != 0)) { 224 /** 225 * Deadlock avoidance for multi-bo reserving. 226 */ 227 if (use_sequence && bo->seq_valid) { 228 /** 229 * We've already reserved this one. 230 */ 231 if (unlikely(sequence == bo->val_seq)) 232 return -EDEADLK; 233 /** 234 * Already reserved by a thread that will not back 235 * off for us. We need to back off. 236 */ 237 if (unlikely(sequence - bo->val_seq < (1 << 31))) 238 return -EAGAIN; 239 } 240 241 if (no_wait) 242 return -EBUSY; 243 244 spin_unlock(&glob->lru_lock); 245 ret = ttm_bo_wait_unreserved(bo, interruptible); 246 spin_lock(&glob->lru_lock); 247 248 if (unlikely(ret)) 249 return ret; 250 } 251 252 atomic_set(&bo->reserved, 1); 253 if (use_sequence) { 254 /** 255 * Wake up waiters that may need to recheck for deadlock, 256 * if we decreased the sequence number. 257 */ 258 if (unlikely((bo->val_seq - sequence < (1 << 31)) 259 || !bo->seq_valid)) 260 wake_up_all(&bo->event_queue); 261 262 bo->val_seq = sequence; 263 bo->seq_valid = true; 264 } else { 265 bo->seq_valid = false; 266 } 267 268 return 0; 269 } 270 EXPORT_SYMBOL(ttm_bo_reserve); 271 272 static void ttm_bo_ref_bug(struct kref *list_kref) 273 { 274 BUG(); 275 } 276 277 void ttm_bo_list_ref_sub(struct ttm_buffer_object *bo, int count, 278 bool never_free) 279 { 280 kref_sub(&bo->list_kref, count, 281 (never_free) ? ttm_bo_ref_bug : ttm_bo_release_list); 282 } 283 284 int ttm_bo_reserve(struct ttm_buffer_object *bo, 285 bool interruptible, 286 bool no_wait, bool use_sequence, uint32_t sequence) 287 { 288 struct ttm_bo_global *glob = bo->glob; 289 int put_count = 0; 290 int ret; 291 292 spin_lock(&glob->lru_lock); 293 ret = ttm_bo_reserve_locked(bo, interruptible, no_wait, use_sequence, 294 sequence); 295 if (likely(ret == 0)) 296 put_count = ttm_bo_del_from_lru(bo); 297 spin_unlock(&glob->lru_lock); 298 299 ttm_bo_list_ref_sub(bo, put_count, true); 300 301 return ret; 302 } 303 304 void ttm_bo_unreserve_locked(struct ttm_buffer_object *bo) 305 { 306 ttm_bo_add_to_lru(bo); 307 atomic_set(&bo->reserved, 0); 308 wake_up_all(&bo->event_queue); 309 } 310 311 void ttm_bo_unreserve(struct ttm_buffer_object *bo) 312 { 313 struct ttm_bo_global *glob = bo->glob; 314 315 spin_lock(&glob->lru_lock); 316 ttm_bo_unreserve_locked(bo); 317 spin_unlock(&glob->lru_lock); 318 } 319 EXPORT_SYMBOL(ttm_bo_unreserve); 320 321 /* 322 * Call bo->mutex locked. 323 */ 324 static int ttm_bo_add_ttm(struct ttm_buffer_object *bo, bool zero_alloc) 325 { 326 struct ttm_bo_device *bdev = bo->bdev; 327 struct ttm_bo_global *glob = bo->glob; 328 int ret = 0; 329 uint32_t page_flags = 0; 330 331 TTM_ASSERT_LOCKED(&bo->mutex); 332 bo->ttm = NULL; 333 334 if (bdev->need_dma32) 335 page_flags |= TTM_PAGE_FLAG_DMA32; 336 337 switch (bo->type) { 338 case ttm_bo_type_device: 339 if (zero_alloc) 340 page_flags |= TTM_PAGE_FLAG_ZERO_ALLOC; 341 case ttm_bo_type_kernel: 342 bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT, 343 page_flags, glob->dummy_read_page); 344 if (unlikely(bo->ttm == NULL)) 345 ret = -ENOMEM; 346 break; 347 case ttm_bo_type_sg: 348 bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT, 349 page_flags | TTM_PAGE_FLAG_SG, 350 glob->dummy_read_page); 351 if (unlikely(bo->ttm == NULL)) { 352 ret = -ENOMEM; 353 break; 354 } 355 bo->ttm->sg = bo->sg; 356 break; 357 default: 358 pr_err("Illegal buffer object type\n"); 359 ret = -EINVAL; 360 break; 361 } 362 363 return ret; 364 } 365 366 static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo, 367 struct ttm_mem_reg *mem, 368 bool evict, bool interruptible, 369 bool no_wait_gpu) 370 { 371 struct ttm_bo_device *bdev = bo->bdev; 372 bool old_is_pci = ttm_mem_reg_is_pci(bdev, &bo->mem); 373 bool new_is_pci = ttm_mem_reg_is_pci(bdev, mem); 374 struct ttm_mem_type_manager *old_man = &bdev->man[bo->mem.mem_type]; 375 struct ttm_mem_type_manager *new_man = &bdev->man[mem->mem_type]; 376 int ret = 0; 377 378 if (old_is_pci || new_is_pci || 379 ((mem->placement & bo->mem.placement & TTM_PL_MASK_CACHING) == 0)) { 380 ret = ttm_mem_io_lock(old_man, true); 381 if (unlikely(ret != 0)) 382 goto out_err; 383 ttm_bo_unmap_virtual_locked(bo); 384 ttm_mem_io_unlock(old_man); 385 } 386 387 /* 388 * Create and bind a ttm if required. 389 */ 390 391 if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) { 392 if (bo->ttm == NULL) { 393 bool zero = !(old_man->flags & TTM_MEMTYPE_FLAG_FIXED); 394 ret = ttm_bo_add_ttm(bo, zero); 395 if (ret) 396 goto out_err; 397 } 398 399 ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement); 400 if (ret) 401 goto out_err; 402 403 if (mem->mem_type != TTM_PL_SYSTEM) { 404 ret = ttm_tt_bind(bo->ttm, mem); 405 if (ret) 406 goto out_err; 407 } 408 409 if (bo->mem.mem_type == TTM_PL_SYSTEM) { 410 if (bdev->driver->move_notify) 411 bdev->driver->move_notify(bo, mem); 412 bo->mem = *mem; 413 mem->mm_node = NULL; 414 goto moved; 415 } 416 } 417 418 if (bdev->driver->move_notify) 419 bdev->driver->move_notify(bo, mem); 420 421 if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) && 422 !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) 423 ret = ttm_bo_move_ttm(bo, evict, no_wait_gpu, mem); 424 else if (bdev->driver->move) 425 ret = bdev->driver->move(bo, evict, interruptible, 426 no_wait_gpu, mem); 427 else 428 ret = ttm_bo_move_memcpy(bo, evict, no_wait_gpu, mem); 429 430 if (ret) { 431 if (bdev->driver->move_notify) { 432 struct ttm_mem_reg tmp_mem = *mem; 433 *mem = bo->mem; 434 bo->mem = tmp_mem; 435 bdev->driver->move_notify(bo, mem); 436 bo->mem = *mem; 437 *mem = tmp_mem; 438 } 439 440 goto out_err; 441 } 442 443 moved: 444 if (bo->evicted) { 445 ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement); 446 if (ret) 447 pr_err("Can not flush read caches\n"); 448 bo->evicted = false; 449 } 450 451 if (bo->mem.mm_node) { 452 bo->offset = (bo->mem.start << PAGE_SHIFT) + 453 bdev->man[bo->mem.mem_type].gpu_offset; 454 bo->cur_placement = bo->mem.placement; 455 } else 456 bo->offset = 0; 457 458 return 0; 459 460 out_err: 461 new_man = &bdev->man[bo->mem.mem_type]; 462 if ((new_man->flags & TTM_MEMTYPE_FLAG_FIXED) && bo->ttm) { 463 ttm_tt_unbind(bo->ttm); 464 ttm_tt_destroy(bo->ttm); 465 bo->ttm = NULL; 466 } 467 468 return ret; 469 } 470 471 /** 472 * Call bo::reserved. 473 * Will release GPU memory type usage on destruction. 474 * This is the place to put in driver specific hooks to release 475 * driver private resources. 476 * Will release the bo::reserved lock. 477 */ 478 479 static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo) 480 { 481 if (bo->bdev->driver->move_notify) 482 bo->bdev->driver->move_notify(bo, NULL); 483 484 if (bo->ttm) { 485 ttm_tt_unbind(bo->ttm); 486 ttm_tt_destroy(bo->ttm); 487 bo->ttm = NULL; 488 } 489 ttm_bo_mem_put(bo, &bo->mem); 490 491 atomic_set(&bo->reserved, 0); 492 wake_up_all(&bo->event_queue); 493 494 /* 495 * Since the final reference to this bo may not be dropped by 496 * the current task we have to put a memory barrier here to make 497 * sure the changes done in this function are always visible. 498 * 499 * This function only needs protection against the final kref_put. 500 */ 501 smp_mb__before_atomic_dec(); 502 } 503 504 static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo) 505 { 506 struct ttm_bo_device *bdev = bo->bdev; 507 struct ttm_bo_global *glob = bo->glob; 508 struct ttm_bo_driver *driver = bdev->driver; 509 void *sync_obj = NULL; 510 int put_count; 511 int ret; 512 513 spin_lock(&glob->lru_lock); 514 ret = ttm_bo_reserve_locked(bo, false, true, false, 0); 515 516 spin_lock(&bdev->fence_lock); 517 (void) ttm_bo_wait(bo, false, false, true); 518 if (!ret && !bo->sync_obj) { 519 spin_unlock(&bdev->fence_lock); 520 put_count = ttm_bo_del_from_lru(bo); 521 522 spin_unlock(&glob->lru_lock); 523 ttm_bo_cleanup_memtype_use(bo); 524 525 ttm_bo_list_ref_sub(bo, put_count, true); 526 527 return; 528 } 529 if (bo->sync_obj) 530 sync_obj = driver->sync_obj_ref(bo->sync_obj); 531 spin_unlock(&bdev->fence_lock); 532 533 if (!ret) { 534 atomic_set(&bo->reserved, 0); 535 wake_up_all(&bo->event_queue); 536 } 537 538 kref_get(&bo->list_kref); 539 list_add_tail(&bo->ddestroy, &bdev->ddestroy); 540 spin_unlock(&glob->lru_lock); 541 542 if (sync_obj) { 543 driver->sync_obj_flush(sync_obj); 544 driver->sync_obj_unref(&sync_obj); 545 } 546 schedule_delayed_work(&bdev->wq, 547 ((HZ / 100) < 1) ? 1 : HZ / 100); 548 } 549 550 /** 551 * function ttm_bo_cleanup_refs_and_unlock 552 * If bo idle, remove from delayed- and lru lists, and unref. 553 * If not idle, do nothing. 554 * 555 * Must be called with lru_lock and reservation held, this function 556 * will drop both before returning. 557 * 558 * @interruptible Any sleeps should occur interruptibly. 559 * @no_wait_gpu Never wait for gpu. Return -EBUSY instead. 560 */ 561 562 static int ttm_bo_cleanup_refs_and_unlock(struct ttm_buffer_object *bo, 563 bool interruptible, 564 bool no_wait_gpu) 565 { 566 struct ttm_bo_device *bdev = bo->bdev; 567 struct ttm_bo_driver *driver = bdev->driver; 568 struct ttm_bo_global *glob = bo->glob; 569 int put_count; 570 int ret; 571 572 spin_lock(&bdev->fence_lock); 573 ret = ttm_bo_wait(bo, false, false, true); 574 575 if (ret && !no_wait_gpu) { 576 void *sync_obj; 577 578 /* 579 * Take a reference to the fence and unreserve, 580 * at this point the buffer should be dead, so 581 * no new sync objects can be attached. 582 */ 583 sync_obj = driver->sync_obj_ref(bo->sync_obj); 584 spin_unlock(&bdev->fence_lock); 585 586 atomic_set(&bo->reserved, 0); 587 wake_up_all(&bo->event_queue); 588 spin_unlock(&glob->lru_lock); 589 590 ret = driver->sync_obj_wait(sync_obj, false, interruptible); 591 driver->sync_obj_unref(&sync_obj); 592 if (ret) 593 return ret; 594 595 /* 596 * remove sync_obj with ttm_bo_wait, the wait should be 597 * finished, and no new wait object should have been added. 598 */ 599 spin_lock(&bdev->fence_lock); 600 ret = ttm_bo_wait(bo, false, false, true); 601 WARN_ON(ret); 602 spin_unlock(&bdev->fence_lock); 603 if (ret) 604 return ret; 605 606 spin_lock(&glob->lru_lock); 607 ret = ttm_bo_reserve_locked(bo, false, true, false, 0); 608 609 /* 610 * We raced, and lost, someone else holds the reservation now, 611 * and is probably busy in ttm_bo_cleanup_memtype_use. 612 * 613 * Even if it's not the case, because we finished waiting any 614 * delayed destruction would succeed, so just return success 615 * here. 616 */ 617 if (ret) { 618 spin_unlock(&glob->lru_lock); 619 return 0; 620 } 621 } else 622 spin_unlock(&bdev->fence_lock); 623 624 if (ret || unlikely(list_empty(&bo->ddestroy))) { 625 atomic_set(&bo->reserved, 0); 626 wake_up_all(&bo->event_queue); 627 spin_unlock(&glob->lru_lock); 628 return ret; 629 } 630 631 put_count = ttm_bo_del_from_lru(bo); 632 list_del_init(&bo->ddestroy); 633 ++put_count; 634 635 spin_unlock(&glob->lru_lock); 636 ttm_bo_cleanup_memtype_use(bo); 637 638 ttm_bo_list_ref_sub(bo, put_count, true); 639 640 return 0; 641 } 642 643 /** 644 * Traverse the delayed list, and call ttm_bo_cleanup_refs on all 645 * encountered buffers. 646 */ 647 648 static int ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all) 649 { 650 struct ttm_bo_global *glob = bdev->glob; 651 struct ttm_buffer_object *entry = NULL; 652 int ret = 0; 653 654 spin_lock(&glob->lru_lock); 655 if (list_empty(&bdev->ddestroy)) 656 goto out_unlock; 657 658 entry = list_first_entry(&bdev->ddestroy, 659 struct ttm_buffer_object, ddestroy); 660 kref_get(&entry->list_kref); 661 662 for (;;) { 663 struct ttm_buffer_object *nentry = NULL; 664 665 if (entry->ddestroy.next != &bdev->ddestroy) { 666 nentry = list_first_entry(&entry->ddestroy, 667 struct ttm_buffer_object, ddestroy); 668 kref_get(&nentry->list_kref); 669 } 670 671 ret = ttm_bo_reserve_locked(entry, false, !remove_all, false, 0); 672 if (!ret) 673 ret = ttm_bo_cleanup_refs_and_unlock(entry, false, 674 !remove_all); 675 else 676 spin_unlock(&glob->lru_lock); 677 678 kref_put(&entry->list_kref, ttm_bo_release_list); 679 entry = nentry; 680 681 if (ret || !entry) 682 goto out; 683 684 spin_lock(&glob->lru_lock); 685 if (list_empty(&entry->ddestroy)) 686 break; 687 } 688 689 out_unlock: 690 spin_unlock(&glob->lru_lock); 691 out: 692 if (entry) 693 kref_put(&entry->list_kref, ttm_bo_release_list); 694 return ret; 695 } 696 697 static void ttm_bo_delayed_workqueue(struct work_struct *work) 698 { 699 struct ttm_bo_device *bdev = 700 container_of(work, struct ttm_bo_device, wq.work); 701 702 if (ttm_bo_delayed_delete(bdev, false)) { 703 schedule_delayed_work(&bdev->wq, 704 ((HZ / 100) < 1) ? 1 : HZ / 100); 705 } 706 } 707 708 static void ttm_bo_release(struct kref *kref) 709 { 710 struct ttm_buffer_object *bo = 711 container_of(kref, struct ttm_buffer_object, kref); 712 struct ttm_bo_device *bdev = bo->bdev; 713 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type]; 714 715 write_lock(&bdev->vm_lock); 716 if (likely(bo->vm_node != NULL)) { 717 rb_erase(&bo->vm_rb, &bdev->addr_space_rb); 718 drm_mm_put_block(bo->vm_node); 719 bo->vm_node = NULL; 720 } 721 write_unlock(&bdev->vm_lock); 722 ttm_mem_io_lock(man, false); 723 ttm_mem_io_free_vm(bo); 724 ttm_mem_io_unlock(man); 725 ttm_bo_cleanup_refs_or_queue(bo); 726 kref_put(&bo->list_kref, ttm_bo_release_list); 727 } 728 729 void ttm_bo_unref(struct ttm_buffer_object **p_bo) 730 { 731 struct ttm_buffer_object *bo = *p_bo; 732 733 *p_bo = NULL; 734 kref_put(&bo->kref, ttm_bo_release); 735 } 736 EXPORT_SYMBOL(ttm_bo_unref); 737 738 int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev) 739 { 740 return cancel_delayed_work_sync(&bdev->wq); 741 } 742 EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue); 743 744 void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched) 745 { 746 if (resched) 747 schedule_delayed_work(&bdev->wq, 748 ((HZ / 100) < 1) ? 1 : HZ / 100); 749 } 750 EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue); 751 752 static int ttm_bo_evict(struct ttm_buffer_object *bo, bool interruptible, 753 bool no_wait_gpu) 754 { 755 struct ttm_bo_device *bdev = bo->bdev; 756 struct ttm_mem_reg evict_mem; 757 struct ttm_placement placement; 758 int ret = 0; 759 760 spin_lock(&bdev->fence_lock); 761 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu); 762 spin_unlock(&bdev->fence_lock); 763 764 if (unlikely(ret != 0)) { 765 if (ret != -ERESTARTSYS) { 766 pr_err("Failed to expire sync object before buffer eviction\n"); 767 } 768 goto out; 769 } 770 771 BUG_ON(!ttm_bo_is_reserved(bo)); 772 773 evict_mem = bo->mem; 774 evict_mem.mm_node = NULL; 775 evict_mem.bus.io_reserved_vm = false; 776 evict_mem.bus.io_reserved_count = 0; 777 778 placement.fpfn = 0; 779 placement.lpfn = 0; 780 placement.num_placement = 0; 781 placement.num_busy_placement = 0; 782 bdev->driver->evict_flags(bo, &placement); 783 ret = ttm_bo_mem_space(bo, &placement, &evict_mem, interruptible, 784 no_wait_gpu); 785 if (ret) { 786 if (ret != -ERESTARTSYS) { 787 pr_err("Failed to find memory space for buffer 0x%p eviction\n", 788 bo); 789 ttm_bo_mem_space_debug(bo, &placement); 790 } 791 goto out; 792 } 793 794 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, interruptible, 795 no_wait_gpu); 796 if (ret) { 797 if (ret != -ERESTARTSYS) 798 pr_err("Buffer eviction failed\n"); 799 ttm_bo_mem_put(bo, &evict_mem); 800 goto out; 801 } 802 bo->evicted = true; 803 out: 804 return ret; 805 } 806 807 static int ttm_mem_evict_first(struct ttm_bo_device *bdev, 808 uint32_t mem_type, 809 bool interruptible, 810 bool no_wait_gpu) 811 { 812 struct ttm_bo_global *glob = bdev->glob; 813 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 814 struct ttm_buffer_object *bo; 815 int ret = -EBUSY, put_count; 816 817 spin_lock(&glob->lru_lock); 818 list_for_each_entry(bo, &man->lru, lru) { 819 ret = ttm_bo_reserve_locked(bo, false, true, false, 0); 820 if (!ret) 821 break; 822 } 823 824 if (ret) { 825 spin_unlock(&glob->lru_lock); 826 return ret; 827 } 828 829 kref_get(&bo->list_kref); 830 831 if (!list_empty(&bo->ddestroy)) { 832 ret = ttm_bo_cleanup_refs_and_unlock(bo, interruptible, 833 no_wait_gpu); 834 kref_put(&bo->list_kref, ttm_bo_release_list); 835 return ret; 836 } 837 838 put_count = ttm_bo_del_from_lru(bo); 839 spin_unlock(&glob->lru_lock); 840 841 BUG_ON(ret != 0); 842 843 ttm_bo_list_ref_sub(bo, put_count, true); 844 845 ret = ttm_bo_evict(bo, interruptible, no_wait_gpu); 846 ttm_bo_unreserve(bo); 847 848 kref_put(&bo->list_kref, ttm_bo_release_list); 849 return ret; 850 } 851 852 void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem) 853 { 854 struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type]; 855 856 if (mem->mm_node) 857 (*man->func->put_node)(man, mem); 858 } 859 EXPORT_SYMBOL(ttm_bo_mem_put); 860 861 /** 862 * Repeatedly evict memory from the LRU for @mem_type until we create enough 863 * space, or we've evicted everything and there isn't enough space. 864 */ 865 static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo, 866 uint32_t mem_type, 867 struct ttm_placement *placement, 868 struct ttm_mem_reg *mem, 869 bool interruptible, 870 bool no_wait_gpu) 871 { 872 struct ttm_bo_device *bdev = bo->bdev; 873 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 874 int ret; 875 876 do { 877 ret = (*man->func->get_node)(man, bo, placement, mem); 878 if (unlikely(ret != 0)) 879 return ret; 880 if (mem->mm_node) 881 break; 882 ret = ttm_mem_evict_first(bdev, mem_type, 883 interruptible, no_wait_gpu); 884 if (unlikely(ret != 0)) 885 return ret; 886 } while (1); 887 if (mem->mm_node == NULL) 888 return -ENOMEM; 889 mem->mem_type = mem_type; 890 return 0; 891 } 892 893 static uint32_t ttm_bo_select_caching(struct ttm_mem_type_manager *man, 894 uint32_t cur_placement, 895 uint32_t proposed_placement) 896 { 897 uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING; 898 uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING; 899 900 /** 901 * Keep current caching if possible. 902 */ 903 904 if ((cur_placement & caching) != 0) 905 result |= (cur_placement & caching); 906 else if ((man->default_caching & caching) != 0) 907 result |= man->default_caching; 908 else if ((TTM_PL_FLAG_CACHED & caching) != 0) 909 result |= TTM_PL_FLAG_CACHED; 910 else if ((TTM_PL_FLAG_WC & caching) != 0) 911 result |= TTM_PL_FLAG_WC; 912 else if ((TTM_PL_FLAG_UNCACHED & caching) != 0) 913 result |= TTM_PL_FLAG_UNCACHED; 914 915 return result; 916 } 917 918 static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man, 919 uint32_t mem_type, 920 uint32_t proposed_placement, 921 uint32_t *masked_placement) 922 { 923 uint32_t cur_flags = ttm_bo_type_flags(mem_type); 924 925 if ((cur_flags & proposed_placement & TTM_PL_MASK_MEM) == 0) 926 return false; 927 928 if ((proposed_placement & man->available_caching) == 0) 929 return false; 930 931 cur_flags |= (proposed_placement & man->available_caching); 932 933 *masked_placement = cur_flags; 934 return true; 935 } 936 937 /** 938 * Creates space for memory region @mem according to its type. 939 * 940 * This function first searches for free space in compatible memory types in 941 * the priority order defined by the driver. If free space isn't found, then 942 * ttm_bo_mem_force_space is attempted in priority order to evict and find 943 * space. 944 */ 945 int ttm_bo_mem_space(struct ttm_buffer_object *bo, 946 struct ttm_placement *placement, 947 struct ttm_mem_reg *mem, 948 bool interruptible, 949 bool no_wait_gpu) 950 { 951 struct ttm_bo_device *bdev = bo->bdev; 952 struct ttm_mem_type_manager *man; 953 uint32_t mem_type = TTM_PL_SYSTEM; 954 uint32_t cur_flags = 0; 955 bool type_found = false; 956 bool type_ok = false; 957 bool has_erestartsys = false; 958 int i, ret; 959 960 mem->mm_node = NULL; 961 for (i = 0; i < placement->num_placement; ++i) { 962 ret = ttm_mem_type_from_flags(placement->placement[i], 963 &mem_type); 964 if (ret) 965 return ret; 966 man = &bdev->man[mem_type]; 967 968 type_ok = ttm_bo_mt_compatible(man, 969 mem_type, 970 placement->placement[i], 971 &cur_flags); 972 973 if (!type_ok) 974 continue; 975 976 cur_flags = ttm_bo_select_caching(man, bo->mem.placement, 977 cur_flags); 978 /* 979 * Use the access and other non-mapping-related flag bits from 980 * the memory placement flags to the current flags 981 */ 982 ttm_flag_masked(&cur_flags, placement->placement[i], 983 ~TTM_PL_MASK_MEMTYPE); 984 985 if (mem_type == TTM_PL_SYSTEM) 986 break; 987 988 if (man->has_type && man->use_type) { 989 type_found = true; 990 ret = (*man->func->get_node)(man, bo, placement, mem); 991 if (unlikely(ret)) 992 return ret; 993 } 994 if (mem->mm_node) 995 break; 996 } 997 998 if ((type_ok && (mem_type == TTM_PL_SYSTEM)) || mem->mm_node) { 999 mem->mem_type = mem_type; 1000 mem->placement = cur_flags; 1001 return 0; 1002 } 1003 1004 if (!type_found) 1005 return -EINVAL; 1006 1007 for (i = 0; i < placement->num_busy_placement; ++i) { 1008 ret = ttm_mem_type_from_flags(placement->busy_placement[i], 1009 &mem_type); 1010 if (ret) 1011 return ret; 1012 man = &bdev->man[mem_type]; 1013 if (!man->has_type) 1014 continue; 1015 if (!ttm_bo_mt_compatible(man, 1016 mem_type, 1017 placement->busy_placement[i], 1018 &cur_flags)) 1019 continue; 1020 1021 cur_flags = ttm_bo_select_caching(man, bo->mem.placement, 1022 cur_flags); 1023 /* 1024 * Use the access and other non-mapping-related flag bits from 1025 * the memory placement flags to the current flags 1026 */ 1027 ttm_flag_masked(&cur_flags, placement->busy_placement[i], 1028 ~TTM_PL_MASK_MEMTYPE); 1029 1030 1031 if (mem_type == TTM_PL_SYSTEM) { 1032 mem->mem_type = mem_type; 1033 mem->placement = cur_flags; 1034 mem->mm_node = NULL; 1035 return 0; 1036 } 1037 1038 ret = ttm_bo_mem_force_space(bo, mem_type, placement, mem, 1039 interruptible, no_wait_gpu); 1040 if (ret == 0 && mem->mm_node) { 1041 mem->placement = cur_flags; 1042 return 0; 1043 } 1044 if (ret == -ERESTARTSYS) 1045 has_erestartsys = true; 1046 } 1047 ret = (has_erestartsys) ? -ERESTARTSYS : -ENOMEM; 1048 return ret; 1049 } 1050 EXPORT_SYMBOL(ttm_bo_mem_space); 1051 1052 int ttm_bo_move_buffer(struct ttm_buffer_object *bo, 1053 struct ttm_placement *placement, 1054 bool interruptible, 1055 bool no_wait_gpu) 1056 { 1057 int ret = 0; 1058 struct ttm_mem_reg mem; 1059 struct ttm_bo_device *bdev = bo->bdev; 1060 1061 BUG_ON(!ttm_bo_is_reserved(bo)); 1062 1063 /* 1064 * FIXME: It's possible to pipeline buffer moves. 1065 * Have the driver move function wait for idle when necessary, 1066 * instead of doing it here. 1067 */ 1068 spin_lock(&bdev->fence_lock); 1069 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu); 1070 spin_unlock(&bdev->fence_lock); 1071 if (ret) 1072 return ret; 1073 mem.num_pages = bo->num_pages; 1074 mem.size = mem.num_pages << PAGE_SHIFT; 1075 mem.page_alignment = bo->mem.page_alignment; 1076 mem.bus.io_reserved_vm = false; 1077 mem.bus.io_reserved_count = 0; 1078 /* 1079 * Determine where to move the buffer. 1080 */ 1081 ret = ttm_bo_mem_space(bo, placement, &mem, 1082 interruptible, no_wait_gpu); 1083 if (ret) 1084 goto out_unlock; 1085 ret = ttm_bo_handle_move_mem(bo, &mem, false, 1086 interruptible, no_wait_gpu); 1087 out_unlock: 1088 if (ret && mem.mm_node) 1089 ttm_bo_mem_put(bo, &mem); 1090 return ret; 1091 } 1092 1093 static int ttm_bo_mem_compat(struct ttm_placement *placement, 1094 struct ttm_mem_reg *mem) 1095 { 1096 int i; 1097 1098 if (mem->mm_node && placement->lpfn != 0 && 1099 (mem->start < placement->fpfn || 1100 mem->start + mem->num_pages > placement->lpfn)) 1101 return -1; 1102 1103 for (i = 0; i < placement->num_placement; i++) { 1104 if ((placement->placement[i] & mem->placement & 1105 TTM_PL_MASK_CACHING) && 1106 (placement->placement[i] & mem->placement & 1107 TTM_PL_MASK_MEM)) 1108 return i; 1109 } 1110 return -1; 1111 } 1112 1113 int ttm_bo_validate(struct ttm_buffer_object *bo, 1114 struct ttm_placement *placement, 1115 bool interruptible, 1116 bool no_wait_gpu) 1117 { 1118 int ret; 1119 1120 BUG_ON(!ttm_bo_is_reserved(bo)); 1121 /* Check that range is valid */ 1122 if (placement->lpfn || placement->fpfn) 1123 if (placement->fpfn > placement->lpfn || 1124 (placement->lpfn - placement->fpfn) < bo->num_pages) 1125 return -EINVAL; 1126 /* 1127 * Check whether we need to move buffer. 1128 */ 1129 ret = ttm_bo_mem_compat(placement, &bo->mem); 1130 if (ret < 0) { 1131 ret = ttm_bo_move_buffer(bo, placement, interruptible, 1132 no_wait_gpu); 1133 if (ret) 1134 return ret; 1135 } else { 1136 /* 1137 * Use the access and other non-mapping-related flag bits from 1138 * the compatible memory placement flags to the active flags 1139 */ 1140 ttm_flag_masked(&bo->mem.placement, placement->placement[ret], 1141 ~TTM_PL_MASK_MEMTYPE); 1142 } 1143 /* 1144 * We might need to add a TTM. 1145 */ 1146 if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) { 1147 ret = ttm_bo_add_ttm(bo, true); 1148 if (ret) 1149 return ret; 1150 } 1151 return 0; 1152 } 1153 EXPORT_SYMBOL(ttm_bo_validate); 1154 1155 int ttm_bo_check_placement(struct ttm_buffer_object *bo, 1156 struct ttm_placement *placement) 1157 { 1158 BUG_ON((placement->fpfn || placement->lpfn) && 1159 (bo->mem.num_pages > (placement->lpfn - placement->fpfn))); 1160 1161 return 0; 1162 } 1163 1164 int ttm_bo_init(struct ttm_bo_device *bdev, 1165 struct ttm_buffer_object *bo, 1166 unsigned long size, 1167 enum ttm_bo_type type, 1168 struct ttm_placement *placement, 1169 uint32_t page_alignment, 1170 bool interruptible, 1171 struct file *persistent_swap_storage, 1172 size_t acc_size, 1173 struct sg_table *sg, 1174 void (*destroy) (struct ttm_buffer_object *)) 1175 { 1176 int ret = 0; 1177 unsigned long num_pages; 1178 struct ttm_mem_global *mem_glob = bdev->glob->mem_glob; 1179 1180 ret = ttm_mem_global_alloc(mem_glob, acc_size, false, false); 1181 if (ret) { 1182 pr_err("Out of kernel memory\n"); 1183 if (destroy) 1184 (*destroy)(bo); 1185 else 1186 kfree(bo); 1187 return -ENOMEM; 1188 } 1189 1190 num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT; 1191 if (num_pages == 0) { 1192 pr_err("Illegal buffer object size\n"); 1193 if (destroy) 1194 (*destroy)(bo); 1195 else 1196 kfree(bo); 1197 ttm_mem_global_free(mem_glob, acc_size); 1198 return -EINVAL; 1199 } 1200 bo->destroy = destroy; 1201 1202 kref_init(&bo->kref); 1203 kref_init(&bo->list_kref); 1204 atomic_set(&bo->cpu_writers, 0); 1205 atomic_set(&bo->reserved, 1); 1206 init_waitqueue_head(&bo->event_queue); 1207 INIT_LIST_HEAD(&bo->lru); 1208 INIT_LIST_HEAD(&bo->ddestroy); 1209 INIT_LIST_HEAD(&bo->swap); 1210 INIT_LIST_HEAD(&bo->io_reserve_lru); 1211 bo->bdev = bdev; 1212 bo->glob = bdev->glob; 1213 bo->type = type; 1214 bo->num_pages = num_pages; 1215 bo->mem.size = num_pages << PAGE_SHIFT; 1216 bo->mem.mem_type = TTM_PL_SYSTEM; 1217 bo->mem.num_pages = bo->num_pages; 1218 bo->mem.mm_node = NULL; 1219 bo->mem.page_alignment = page_alignment; 1220 bo->mem.bus.io_reserved_vm = false; 1221 bo->mem.bus.io_reserved_count = 0; 1222 bo->priv_flags = 0; 1223 bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED); 1224 bo->seq_valid = false; 1225 bo->persistent_swap_storage = persistent_swap_storage; 1226 bo->acc_size = acc_size; 1227 bo->sg = sg; 1228 atomic_inc(&bo->glob->bo_count); 1229 1230 ret = ttm_bo_check_placement(bo, placement); 1231 if (unlikely(ret != 0)) 1232 goto out_err; 1233 1234 /* 1235 * For ttm_bo_type_device buffers, allocate 1236 * address space from the device. 1237 */ 1238 if (bo->type == ttm_bo_type_device || 1239 bo->type == ttm_bo_type_sg) { 1240 ret = ttm_bo_setup_vm(bo); 1241 if (ret) 1242 goto out_err; 1243 } 1244 1245 ret = ttm_bo_validate(bo, placement, interruptible, false); 1246 if (ret) 1247 goto out_err; 1248 1249 ttm_bo_unreserve(bo); 1250 return 0; 1251 1252 out_err: 1253 ttm_bo_unreserve(bo); 1254 ttm_bo_unref(&bo); 1255 1256 return ret; 1257 } 1258 EXPORT_SYMBOL(ttm_bo_init); 1259 1260 size_t ttm_bo_acc_size(struct ttm_bo_device *bdev, 1261 unsigned long bo_size, 1262 unsigned struct_size) 1263 { 1264 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT; 1265 size_t size = 0; 1266 1267 size += ttm_round_pot(struct_size); 1268 size += PAGE_ALIGN(npages * sizeof(void *)); 1269 size += ttm_round_pot(sizeof(struct ttm_tt)); 1270 return size; 1271 } 1272 EXPORT_SYMBOL(ttm_bo_acc_size); 1273 1274 size_t ttm_bo_dma_acc_size(struct ttm_bo_device *bdev, 1275 unsigned long bo_size, 1276 unsigned struct_size) 1277 { 1278 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT; 1279 size_t size = 0; 1280 1281 size += ttm_round_pot(struct_size); 1282 size += PAGE_ALIGN(npages * sizeof(void *)); 1283 size += PAGE_ALIGN(npages * sizeof(dma_addr_t)); 1284 size += ttm_round_pot(sizeof(struct ttm_dma_tt)); 1285 return size; 1286 } 1287 EXPORT_SYMBOL(ttm_bo_dma_acc_size); 1288 1289 int ttm_bo_create(struct ttm_bo_device *bdev, 1290 unsigned long size, 1291 enum ttm_bo_type type, 1292 struct ttm_placement *placement, 1293 uint32_t page_alignment, 1294 bool interruptible, 1295 struct file *persistent_swap_storage, 1296 struct ttm_buffer_object **p_bo) 1297 { 1298 struct ttm_buffer_object *bo; 1299 size_t acc_size; 1300 int ret; 1301 1302 bo = kzalloc(sizeof(*bo), GFP_KERNEL); 1303 if (unlikely(bo == NULL)) 1304 return -ENOMEM; 1305 1306 acc_size = ttm_bo_acc_size(bdev, size, sizeof(struct ttm_buffer_object)); 1307 ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment, 1308 interruptible, persistent_swap_storage, acc_size, 1309 NULL, NULL); 1310 if (likely(ret == 0)) 1311 *p_bo = bo; 1312 1313 return ret; 1314 } 1315 EXPORT_SYMBOL(ttm_bo_create); 1316 1317 static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev, 1318 unsigned mem_type, bool allow_errors) 1319 { 1320 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 1321 struct ttm_bo_global *glob = bdev->glob; 1322 int ret; 1323 1324 /* 1325 * Can't use standard list traversal since we're unlocking. 1326 */ 1327 1328 spin_lock(&glob->lru_lock); 1329 while (!list_empty(&man->lru)) { 1330 spin_unlock(&glob->lru_lock); 1331 ret = ttm_mem_evict_first(bdev, mem_type, false, false); 1332 if (ret) { 1333 if (allow_errors) { 1334 return ret; 1335 } else { 1336 pr_err("Cleanup eviction failed\n"); 1337 } 1338 } 1339 spin_lock(&glob->lru_lock); 1340 } 1341 spin_unlock(&glob->lru_lock); 1342 return 0; 1343 } 1344 1345 int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type) 1346 { 1347 struct ttm_mem_type_manager *man; 1348 int ret = -EINVAL; 1349 1350 if (mem_type >= TTM_NUM_MEM_TYPES) { 1351 pr_err("Illegal memory type %d\n", mem_type); 1352 return ret; 1353 } 1354 man = &bdev->man[mem_type]; 1355 1356 if (!man->has_type) { 1357 pr_err("Trying to take down uninitialized memory manager type %u\n", 1358 mem_type); 1359 return ret; 1360 } 1361 1362 man->use_type = false; 1363 man->has_type = false; 1364 1365 ret = 0; 1366 if (mem_type > 0) { 1367 ttm_bo_force_list_clean(bdev, mem_type, false); 1368 1369 ret = (*man->func->takedown)(man); 1370 } 1371 1372 return ret; 1373 } 1374 EXPORT_SYMBOL(ttm_bo_clean_mm); 1375 1376 int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type) 1377 { 1378 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 1379 1380 if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) { 1381 pr_err("Illegal memory manager memory type %u\n", mem_type); 1382 return -EINVAL; 1383 } 1384 1385 if (!man->has_type) { 1386 pr_err("Memory type %u has not been initialized\n", mem_type); 1387 return 0; 1388 } 1389 1390 return ttm_bo_force_list_clean(bdev, mem_type, true); 1391 } 1392 EXPORT_SYMBOL(ttm_bo_evict_mm); 1393 1394 int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type, 1395 unsigned long p_size) 1396 { 1397 int ret = -EINVAL; 1398 struct ttm_mem_type_manager *man; 1399 1400 BUG_ON(type >= TTM_NUM_MEM_TYPES); 1401 man = &bdev->man[type]; 1402 BUG_ON(man->has_type); 1403 man->io_reserve_fastpath = true; 1404 man->use_io_reserve_lru = false; 1405 mutex_init(&man->io_reserve_mutex); 1406 INIT_LIST_HEAD(&man->io_reserve_lru); 1407 1408 ret = bdev->driver->init_mem_type(bdev, type, man); 1409 if (ret) 1410 return ret; 1411 man->bdev = bdev; 1412 1413 ret = 0; 1414 if (type != TTM_PL_SYSTEM) { 1415 ret = (*man->func->init)(man, p_size); 1416 if (ret) 1417 return ret; 1418 } 1419 man->has_type = true; 1420 man->use_type = true; 1421 man->size = p_size; 1422 1423 INIT_LIST_HEAD(&man->lru); 1424 1425 return 0; 1426 } 1427 EXPORT_SYMBOL(ttm_bo_init_mm); 1428 1429 static void ttm_bo_global_kobj_release(struct kobject *kobj) 1430 { 1431 struct ttm_bo_global *glob = 1432 container_of(kobj, struct ttm_bo_global, kobj); 1433 1434 ttm_mem_unregister_shrink(glob->mem_glob, &glob->shrink); 1435 __free_page(glob->dummy_read_page); 1436 kfree(glob); 1437 } 1438 1439 void ttm_bo_global_release(struct drm_global_reference *ref) 1440 { 1441 struct ttm_bo_global *glob = ref->object; 1442 1443 kobject_del(&glob->kobj); 1444 kobject_put(&glob->kobj); 1445 } 1446 EXPORT_SYMBOL(ttm_bo_global_release); 1447 1448 int ttm_bo_global_init(struct drm_global_reference *ref) 1449 { 1450 struct ttm_bo_global_ref *bo_ref = 1451 container_of(ref, struct ttm_bo_global_ref, ref); 1452 struct ttm_bo_global *glob = ref->object; 1453 int ret; 1454 1455 mutex_init(&glob->device_list_mutex); 1456 spin_lock_init(&glob->lru_lock); 1457 glob->mem_glob = bo_ref->mem_glob; 1458 glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32); 1459 1460 if (unlikely(glob->dummy_read_page == NULL)) { 1461 ret = -ENOMEM; 1462 goto out_no_drp; 1463 } 1464 1465 INIT_LIST_HEAD(&glob->swap_lru); 1466 INIT_LIST_HEAD(&glob->device_list); 1467 1468 ttm_mem_init_shrink(&glob->shrink, ttm_bo_swapout); 1469 ret = ttm_mem_register_shrink(glob->mem_glob, &glob->shrink); 1470 if (unlikely(ret != 0)) { 1471 pr_err("Could not register buffer object swapout\n"); 1472 goto out_no_shrink; 1473 } 1474 1475 atomic_set(&glob->bo_count, 0); 1476 1477 ret = kobject_init_and_add( 1478 &glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects"); 1479 if (unlikely(ret != 0)) 1480 kobject_put(&glob->kobj); 1481 return ret; 1482 out_no_shrink: 1483 __free_page(glob->dummy_read_page); 1484 out_no_drp: 1485 kfree(glob); 1486 return ret; 1487 } 1488 EXPORT_SYMBOL(ttm_bo_global_init); 1489 1490 1491 int ttm_bo_device_release(struct ttm_bo_device *bdev) 1492 { 1493 int ret = 0; 1494 unsigned i = TTM_NUM_MEM_TYPES; 1495 struct ttm_mem_type_manager *man; 1496 struct ttm_bo_global *glob = bdev->glob; 1497 1498 while (i--) { 1499 man = &bdev->man[i]; 1500 if (man->has_type) { 1501 man->use_type = false; 1502 if ((i != TTM_PL_SYSTEM) && ttm_bo_clean_mm(bdev, i)) { 1503 ret = -EBUSY; 1504 pr_err("DRM memory manager type %d is not clean\n", 1505 i); 1506 } 1507 man->has_type = false; 1508 } 1509 } 1510 1511 mutex_lock(&glob->device_list_mutex); 1512 list_del(&bdev->device_list); 1513 mutex_unlock(&glob->device_list_mutex); 1514 1515 cancel_delayed_work_sync(&bdev->wq); 1516 1517 while (ttm_bo_delayed_delete(bdev, true)) 1518 ; 1519 1520 spin_lock(&glob->lru_lock); 1521 if (list_empty(&bdev->ddestroy)) 1522 TTM_DEBUG("Delayed destroy list was clean\n"); 1523 1524 if (list_empty(&bdev->man[0].lru)) 1525 TTM_DEBUG("Swap list was clean\n"); 1526 spin_unlock(&glob->lru_lock); 1527 1528 BUG_ON(!drm_mm_clean(&bdev->addr_space_mm)); 1529 write_lock(&bdev->vm_lock); 1530 drm_mm_takedown(&bdev->addr_space_mm); 1531 write_unlock(&bdev->vm_lock); 1532 1533 return ret; 1534 } 1535 EXPORT_SYMBOL(ttm_bo_device_release); 1536 1537 int ttm_bo_device_init(struct ttm_bo_device *bdev, 1538 struct ttm_bo_global *glob, 1539 struct ttm_bo_driver *driver, 1540 uint64_t file_page_offset, 1541 bool need_dma32) 1542 { 1543 int ret = -EINVAL; 1544 1545 rwlock_init(&bdev->vm_lock); 1546 bdev->driver = driver; 1547 1548 memset(bdev->man, 0, sizeof(bdev->man)); 1549 1550 /* 1551 * Initialize the system memory buffer type. 1552 * Other types need to be driver / IOCTL initialized. 1553 */ 1554 ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0); 1555 if (unlikely(ret != 0)) 1556 goto out_no_sys; 1557 1558 bdev->addr_space_rb = RB_ROOT; 1559 ret = drm_mm_init(&bdev->addr_space_mm, file_page_offset, 0x10000000); 1560 if (unlikely(ret != 0)) 1561 goto out_no_addr_mm; 1562 1563 INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue); 1564 INIT_LIST_HEAD(&bdev->ddestroy); 1565 bdev->dev_mapping = NULL; 1566 bdev->glob = glob; 1567 bdev->need_dma32 = need_dma32; 1568 bdev->val_seq = 0; 1569 spin_lock_init(&bdev->fence_lock); 1570 mutex_lock(&glob->device_list_mutex); 1571 list_add_tail(&bdev->device_list, &glob->device_list); 1572 mutex_unlock(&glob->device_list_mutex); 1573 1574 return 0; 1575 out_no_addr_mm: 1576 ttm_bo_clean_mm(bdev, 0); 1577 out_no_sys: 1578 return ret; 1579 } 1580 EXPORT_SYMBOL(ttm_bo_device_init); 1581 1582 /* 1583 * buffer object vm functions. 1584 */ 1585 1586 bool ttm_mem_reg_is_pci(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem) 1587 { 1588 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type]; 1589 1590 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) { 1591 if (mem->mem_type == TTM_PL_SYSTEM) 1592 return false; 1593 1594 if (man->flags & TTM_MEMTYPE_FLAG_CMA) 1595 return false; 1596 1597 if (mem->placement & TTM_PL_FLAG_CACHED) 1598 return false; 1599 } 1600 return true; 1601 } 1602 1603 void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo) 1604 { 1605 struct ttm_bo_device *bdev = bo->bdev; 1606 loff_t offset = (loff_t) bo->addr_space_offset; 1607 loff_t holelen = ((loff_t) bo->mem.num_pages) << PAGE_SHIFT; 1608 1609 if (!bdev->dev_mapping) 1610 return; 1611 unmap_mapping_range(bdev->dev_mapping, offset, holelen, 1); 1612 ttm_mem_io_free_vm(bo); 1613 } 1614 1615 void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo) 1616 { 1617 struct ttm_bo_device *bdev = bo->bdev; 1618 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type]; 1619 1620 ttm_mem_io_lock(man, false); 1621 ttm_bo_unmap_virtual_locked(bo); 1622 ttm_mem_io_unlock(man); 1623 } 1624 1625 1626 EXPORT_SYMBOL(ttm_bo_unmap_virtual); 1627 1628 static void ttm_bo_vm_insert_rb(struct ttm_buffer_object *bo) 1629 { 1630 struct ttm_bo_device *bdev = bo->bdev; 1631 struct rb_node **cur = &bdev->addr_space_rb.rb_node; 1632 struct rb_node *parent = NULL; 1633 struct ttm_buffer_object *cur_bo; 1634 unsigned long offset = bo->vm_node->start; 1635 unsigned long cur_offset; 1636 1637 while (*cur) { 1638 parent = *cur; 1639 cur_bo = rb_entry(parent, struct ttm_buffer_object, vm_rb); 1640 cur_offset = cur_bo->vm_node->start; 1641 if (offset < cur_offset) 1642 cur = &parent->rb_left; 1643 else if (offset > cur_offset) 1644 cur = &parent->rb_right; 1645 else 1646 BUG(); 1647 } 1648 1649 rb_link_node(&bo->vm_rb, parent, cur); 1650 rb_insert_color(&bo->vm_rb, &bdev->addr_space_rb); 1651 } 1652 1653 /** 1654 * ttm_bo_setup_vm: 1655 * 1656 * @bo: the buffer to allocate address space for 1657 * 1658 * Allocate address space in the drm device so that applications 1659 * can mmap the buffer and access the contents. This only 1660 * applies to ttm_bo_type_device objects as others are not 1661 * placed in the drm device address space. 1662 */ 1663 1664 static int ttm_bo_setup_vm(struct ttm_buffer_object *bo) 1665 { 1666 struct ttm_bo_device *bdev = bo->bdev; 1667 int ret; 1668 1669 retry_pre_get: 1670 ret = drm_mm_pre_get(&bdev->addr_space_mm); 1671 if (unlikely(ret != 0)) 1672 return ret; 1673 1674 write_lock(&bdev->vm_lock); 1675 bo->vm_node = drm_mm_search_free(&bdev->addr_space_mm, 1676 bo->mem.num_pages, 0, 0); 1677 1678 if (unlikely(bo->vm_node == NULL)) { 1679 ret = -ENOMEM; 1680 goto out_unlock; 1681 } 1682 1683 bo->vm_node = drm_mm_get_block_atomic(bo->vm_node, 1684 bo->mem.num_pages, 0); 1685 1686 if (unlikely(bo->vm_node == NULL)) { 1687 write_unlock(&bdev->vm_lock); 1688 goto retry_pre_get; 1689 } 1690 1691 ttm_bo_vm_insert_rb(bo); 1692 write_unlock(&bdev->vm_lock); 1693 bo->addr_space_offset = ((uint64_t) bo->vm_node->start) << PAGE_SHIFT; 1694 1695 return 0; 1696 out_unlock: 1697 write_unlock(&bdev->vm_lock); 1698 return ret; 1699 } 1700 1701 int ttm_bo_wait(struct ttm_buffer_object *bo, 1702 bool lazy, bool interruptible, bool no_wait) 1703 { 1704 struct ttm_bo_driver *driver = bo->bdev->driver; 1705 struct ttm_bo_device *bdev = bo->bdev; 1706 void *sync_obj; 1707 int ret = 0; 1708 1709 if (likely(bo->sync_obj == NULL)) 1710 return 0; 1711 1712 while (bo->sync_obj) { 1713 1714 if (driver->sync_obj_signaled(bo->sync_obj)) { 1715 void *tmp_obj = bo->sync_obj; 1716 bo->sync_obj = NULL; 1717 clear_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags); 1718 spin_unlock(&bdev->fence_lock); 1719 driver->sync_obj_unref(&tmp_obj); 1720 spin_lock(&bdev->fence_lock); 1721 continue; 1722 } 1723 1724 if (no_wait) 1725 return -EBUSY; 1726 1727 sync_obj = driver->sync_obj_ref(bo->sync_obj); 1728 spin_unlock(&bdev->fence_lock); 1729 ret = driver->sync_obj_wait(sync_obj, 1730 lazy, interruptible); 1731 if (unlikely(ret != 0)) { 1732 driver->sync_obj_unref(&sync_obj); 1733 spin_lock(&bdev->fence_lock); 1734 return ret; 1735 } 1736 spin_lock(&bdev->fence_lock); 1737 if (likely(bo->sync_obj == sync_obj)) { 1738 void *tmp_obj = bo->sync_obj; 1739 bo->sync_obj = NULL; 1740 clear_bit(TTM_BO_PRIV_FLAG_MOVING, 1741 &bo->priv_flags); 1742 spin_unlock(&bdev->fence_lock); 1743 driver->sync_obj_unref(&sync_obj); 1744 driver->sync_obj_unref(&tmp_obj); 1745 spin_lock(&bdev->fence_lock); 1746 } else { 1747 spin_unlock(&bdev->fence_lock); 1748 driver->sync_obj_unref(&sync_obj); 1749 spin_lock(&bdev->fence_lock); 1750 } 1751 } 1752 return 0; 1753 } 1754 EXPORT_SYMBOL(ttm_bo_wait); 1755 1756 int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait) 1757 { 1758 struct ttm_bo_device *bdev = bo->bdev; 1759 int ret = 0; 1760 1761 /* 1762 * Using ttm_bo_reserve makes sure the lru lists are updated. 1763 */ 1764 1765 ret = ttm_bo_reserve(bo, true, no_wait, false, 0); 1766 if (unlikely(ret != 0)) 1767 return ret; 1768 spin_lock(&bdev->fence_lock); 1769 ret = ttm_bo_wait(bo, false, true, no_wait); 1770 spin_unlock(&bdev->fence_lock); 1771 if (likely(ret == 0)) 1772 atomic_inc(&bo->cpu_writers); 1773 ttm_bo_unreserve(bo); 1774 return ret; 1775 } 1776 EXPORT_SYMBOL(ttm_bo_synccpu_write_grab); 1777 1778 void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo) 1779 { 1780 atomic_dec(&bo->cpu_writers); 1781 } 1782 EXPORT_SYMBOL(ttm_bo_synccpu_write_release); 1783 1784 /** 1785 * A buffer object shrink method that tries to swap out the first 1786 * buffer object on the bo_global::swap_lru list. 1787 */ 1788 1789 static int ttm_bo_swapout(struct ttm_mem_shrink *shrink) 1790 { 1791 struct ttm_bo_global *glob = 1792 container_of(shrink, struct ttm_bo_global, shrink); 1793 struct ttm_buffer_object *bo; 1794 int ret = -EBUSY; 1795 int put_count; 1796 uint32_t swap_placement = (TTM_PL_FLAG_CACHED | TTM_PL_FLAG_SYSTEM); 1797 1798 spin_lock(&glob->lru_lock); 1799 list_for_each_entry(bo, &glob->swap_lru, swap) { 1800 ret = ttm_bo_reserve_locked(bo, false, true, false, 0); 1801 if (!ret) 1802 break; 1803 } 1804 1805 if (ret) { 1806 spin_unlock(&glob->lru_lock); 1807 return ret; 1808 } 1809 1810 kref_get(&bo->list_kref); 1811 1812 if (!list_empty(&bo->ddestroy)) { 1813 ret = ttm_bo_cleanup_refs_and_unlock(bo, false, false); 1814 kref_put(&bo->list_kref, ttm_bo_release_list); 1815 return ret; 1816 } 1817 1818 put_count = ttm_bo_del_from_lru(bo); 1819 spin_unlock(&glob->lru_lock); 1820 1821 ttm_bo_list_ref_sub(bo, put_count, true); 1822 1823 /** 1824 * Wait for GPU, then move to system cached. 1825 */ 1826 1827 spin_lock(&bo->bdev->fence_lock); 1828 ret = ttm_bo_wait(bo, false, false, false); 1829 spin_unlock(&bo->bdev->fence_lock); 1830 1831 if (unlikely(ret != 0)) 1832 goto out; 1833 1834 if ((bo->mem.placement & swap_placement) != swap_placement) { 1835 struct ttm_mem_reg evict_mem; 1836 1837 evict_mem = bo->mem; 1838 evict_mem.mm_node = NULL; 1839 evict_mem.placement = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED; 1840 evict_mem.mem_type = TTM_PL_SYSTEM; 1841 1842 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, 1843 false, false); 1844 if (unlikely(ret != 0)) 1845 goto out; 1846 } 1847 1848 ttm_bo_unmap_virtual(bo); 1849 1850 /** 1851 * Swap out. Buffer will be swapped in again as soon as 1852 * anyone tries to access a ttm page. 1853 */ 1854 1855 if (bo->bdev->driver->swap_notify) 1856 bo->bdev->driver->swap_notify(bo); 1857 1858 ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage); 1859 out: 1860 1861 /** 1862 * 1863 * Unreserve without putting on LRU to avoid swapping out an 1864 * already swapped buffer. 1865 */ 1866 1867 atomic_set(&bo->reserved, 0); 1868 wake_up_all(&bo->event_queue); 1869 kref_put(&bo->list_kref, ttm_bo_release_list); 1870 return ret; 1871 } 1872 1873 void ttm_bo_swapout_all(struct ttm_bo_device *bdev) 1874 { 1875 while (ttm_bo_swapout(&bdev->glob->shrink) == 0) 1876 ; 1877 } 1878 EXPORT_SYMBOL(ttm_bo_swapout_all); 1879