1 /* 2 * Copyright (C) 2007 Oracle. All rights reserved. 3 * 4 * This program is free software; you can redistribute it and/or 5 * modify it under the terms of the GNU General Public 6 * License v2 as published by the Free Software Foundation. 7 * 8 * This program is distributed in the hope that it will be useful, 9 * but WITHOUT ANY WARRANTY; without even the implied warranty of 10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU 11 * General Public License for more details. 12 * 13 * You should have received a copy of the GNU General Public 14 * License along with this program; if not, write to the 15 * Free Software Foundation, Inc., 59 Temple Place - Suite 330, 16 * Boston, MA 021110-1307, USA. 17 */ 18 #include <linux/sched.h> 19 #include <linux/sched/signal.h> 20 #include <linux/pagemap.h> 21 #include <linux/writeback.h> 22 #include <linux/blkdev.h> 23 #include <linux/sort.h> 24 #include <linux/rcupdate.h> 25 #include <linux/kthread.h> 26 #include <linux/slab.h> 27 #include <linux/ratelimit.h> 28 #include <linux/percpu_counter.h> 29 #include <linux/lockdep.h> 30 #include "hash.h" 31 #include "tree-log.h" 32 #include "disk-io.h" 33 #include "print-tree.h" 34 #include "volumes.h" 35 #include "raid56.h" 36 #include "locking.h" 37 #include "free-space-cache.h" 38 #include "free-space-tree.h" 39 #include "math.h" 40 #include "sysfs.h" 41 #include "qgroup.h" 42 #include "ref-verify.h" 43 44 #undef SCRAMBLE_DELAYED_REFS 45 46 /* 47 * control flags for do_chunk_alloc's force field 48 * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk 49 * if we really need one. 50 * 51 * CHUNK_ALLOC_LIMITED means to only try and allocate one 52 * if we have very few chunks already allocated. This is 53 * used as part of the clustering code to help make sure 54 * we have a good pool of storage to cluster in, without 55 * filling the FS with empty chunks 56 * 57 * CHUNK_ALLOC_FORCE means it must try to allocate one 58 * 59 */ 60 enum { 61 CHUNK_ALLOC_NO_FORCE = 0, 62 CHUNK_ALLOC_LIMITED = 1, 63 CHUNK_ALLOC_FORCE = 2, 64 }; 65 66 static int __btrfs_free_extent(struct btrfs_trans_handle *trans, 67 struct btrfs_fs_info *fs_info, 68 struct btrfs_delayed_ref_node *node, u64 parent, 69 u64 root_objectid, u64 owner_objectid, 70 u64 owner_offset, int refs_to_drop, 71 struct btrfs_delayed_extent_op *extra_op); 72 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op, 73 struct extent_buffer *leaf, 74 struct btrfs_extent_item *ei); 75 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans, 76 struct btrfs_fs_info *fs_info, 77 u64 parent, u64 root_objectid, 78 u64 flags, u64 owner, u64 offset, 79 struct btrfs_key *ins, int ref_mod); 80 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans, 81 struct btrfs_fs_info *fs_info, 82 u64 parent, u64 root_objectid, 83 u64 flags, struct btrfs_disk_key *key, 84 int level, struct btrfs_key *ins); 85 static int do_chunk_alloc(struct btrfs_trans_handle *trans, 86 struct btrfs_fs_info *fs_info, u64 flags, 87 int force); 88 static int find_next_key(struct btrfs_path *path, int level, 89 struct btrfs_key *key); 90 static void dump_space_info(struct btrfs_fs_info *fs_info, 91 struct btrfs_space_info *info, u64 bytes, 92 int dump_block_groups); 93 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv, 94 u64 num_bytes); 95 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info, 96 struct btrfs_space_info *space_info, 97 u64 num_bytes); 98 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info, 99 struct btrfs_space_info *space_info, 100 u64 num_bytes); 101 102 static noinline int 103 block_group_cache_done(struct btrfs_block_group_cache *cache) 104 { 105 smp_mb(); 106 return cache->cached == BTRFS_CACHE_FINISHED || 107 cache->cached == BTRFS_CACHE_ERROR; 108 } 109 110 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits) 111 { 112 return (cache->flags & bits) == bits; 113 } 114 115 void btrfs_get_block_group(struct btrfs_block_group_cache *cache) 116 { 117 atomic_inc(&cache->count); 118 } 119 120 void btrfs_put_block_group(struct btrfs_block_group_cache *cache) 121 { 122 if (atomic_dec_and_test(&cache->count)) { 123 WARN_ON(cache->pinned > 0); 124 WARN_ON(cache->reserved > 0); 125 126 /* 127 * If not empty, someone is still holding mutex of 128 * full_stripe_lock, which can only be released by caller. 129 * And it will definitely cause use-after-free when caller 130 * tries to release full stripe lock. 131 * 132 * No better way to resolve, but only to warn. 133 */ 134 WARN_ON(!RB_EMPTY_ROOT(&cache->full_stripe_locks_root.root)); 135 kfree(cache->free_space_ctl); 136 kfree(cache); 137 } 138 } 139 140 /* 141 * this adds the block group to the fs_info rb tree for the block group 142 * cache 143 */ 144 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info, 145 struct btrfs_block_group_cache *block_group) 146 { 147 struct rb_node **p; 148 struct rb_node *parent = NULL; 149 struct btrfs_block_group_cache *cache; 150 151 spin_lock(&info->block_group_cache_lock); 152 p = &info->block_group_cache_tree.rb_node; 153 154 while (*p) { 155 parent = *p; 156 cache = rb_entry(parent, struct btrfs_block_group_cache, 157 cache_node); 158 if (block_group->key.objectid < cache->key.objectid) { 159 p = &(*p)->rb_left; 160 } else if (block_group->key.objectid > cache->key.objectid) { 161 p = &(*p)->rb_right; 162 } else { 163 spin_unlock(&info->block_group_cache_lock); 164 return -EEXIST; 165 } 166 } 167 168 rb_link_node(&block_group->cache_node, parent, p); 169 rb_insert_color(&block_group->cache_node, 170 &info->block_group_cache_tree); 171 172 if (info->first_logical_byte > block_group->key.objectid) 173 info->first_logical_byte = block_group->key.objectid; 174 175 spin_unlock(&info->block_group_cache_lock); 176 177 return 0; 178 } 179 180 /* 181 * This will return the block group at or after bytenr if contains is 0, else 182 * it will return the block group that contains the bytenr 183 */ 184 static struct btrfs_block_group_cache * 185 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr, 186 int contains) 187 { 188 struct btrfs_block_group_cache *cache, *ret = NULL; 189 struct rb_node *n; 190 u64 end, start; 191 192 spin_lock(&info->block_group_cache_lock); 193 n = info->block_group_cache_tree.rb_node; 194 195 while (n) { 196 cache = rb_entry(n, struct btrfs_block_group_cache, 197 cache_node); 198 end = cache->key.objectid + cache->key.offset - 1; 199 start = cache->key.objectid; 200 201 if (bytenr < start) { 202 if (!contains && (!ret || start < ret->key.objectid)) 203 ret = cache; 204 n = n->rb_left; 205 } else if (bytenr > start) { 206 if (contains && bytenr <= end) { 207 ret = cache; 208 break; 209 } 210 n = n->rb_right; 211 } else { 212 ret = cache; 213 break; 214 } 215 } 216 if (ret) { 217 btrfs_get_block_group(ret); 218 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid) 219 info->first_logical_byte = ret->key.objectid; 220 } 221 spin_unlock(&info->block_group_cache_lock); 222 223 return ret; 224 } 225 226 static int add_excluded_extent(struct btrfs_fs_info *fs_info, 227 u64 start, u64 num_bytes) 228 { 229 u64 end = start + num_bytes - 1; 230 set_extent_bits(&fs_info->freed_extents[0], 231 start, end, EXTENT_UPTODATE); 232 set_extent_bits(&fs_info->freed_extents[1], 233 start, end, EXTENT_UPTODATE); 234 return 0; 235 } 236 237 static void free_excluded_extents(struct btrfs_fs_info *fs_info, 238 struct btrfs_block_group_cache *cache) 239 { 240 u64 start, end; 241 242 start = cache->key.objectid; 243 end = start + cache->key.offset - 1; 244 245 clear_extent_bits(&fs_info->freed_extents[0], 246 start, end, EXTENT_UPTODATE); 247 clear_extent_bits(&fs_info->freed_extents[1], 248 start, end, EXTENT_UPTODATE); 249 } 250 251 static int exclude_super_stripes(struct btrfs_fs_info *fs_info, 252 struct btrfs_block_group_cache *cache) 253 { 254 u64 bytenr; 255 u64 *logical; 256 int stripe_len; 257 int i, nr, ret; 258 259 if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) { 260 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid; 261 cache->bytes_super += stripe_len; 262 ret = add_excluded_extent(fs_info, cache->key.objectid, 263 stripe_len); 264 if (ret) 265 return ret; 266 } 267 268 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) { 269 bytenr = btrfs_sb_offset(i); 270 ret = btrfs_rmap_block(fs_info, cache->key.objectid, 271 bytenr, 0, &logical, &nr, &stripe_len); 272 if (ret) 273 return ret; 274 275 while (nr--) { 276 u64 start, len; 277 278 if (logical[nr] > cache->key.objectid + 279 cache->key.offset) 280 continue; 281 282 if (logical[nr] + stripe_len <= cache->key.objectid) 283 continue; 284 285 start = logical[nr]; 286 if (start < cache->key.objectid) { 287 start = cache->key.objectid; 288 len = (logical[nr] + stripe_len) - start; 289 } else { 290 len = min_t(u64, stripe_len, 291 cache->key.objectid + 292 cache->key.offset - start); 293 } 294 295 cache->bytes_super += len; 296 ret = add_excluded_extent(fs_info, start, len); 297 if (ret) { 298 kfree(logical); 299 return ret; 300 } 301 } 302 303 kfree(logical); 304 } 305 return 0; 306 } 307 308 static struct btrfs_caching_control * 309 get_caching_control(struct btrfs_block_group_cache *cache) 310 { 311 struct btrfs_caching_control *ctl; 312 313 spin_lock(&cache->lock); 314 if (!cache->caching_ctl) { 315 spin_unlock(&cache->lock); 316 return NULL; 317 } 318 319 ctl = cache->caching_ctl; 320 refcount_inc(&ctl->count); 321 spin_unlock(&cache->lock); 322 return ctl; 323 } 324 325 static void put_caching_control(struct btrfs_caching_control *ctl) 326 { 327 if (refcount_dec_and_test(&ctl->count)) 328 kfree(ctl); 329 } 330 331 #ifdef CONFIG_BTRFS_DEBUG 332 static void fragment_free_space(struct btrfs_block_group_cache *block_group) 333 { 334 struct btrfs_fs_info *fs_info = block_group->fs_info; 335 u64 start = block_group->key.objectid; 336 u64 len = block_group->key.offset; 337 u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ? 338 fs_info->nodesize : fs_info->sectorsize; 339 u64 step = chunk << 1; 340 341 while (len > chunk) { 342 btrfs_remove_free_space(block_group, start, chunk); 343 start += step; 344 if (len < step) 345 len = 0; 346 else 347 len -= step; 348 } 349 } 350 #endif 351 352 /* 353 * this is only called by cache_block_group, since we could have freed extents 354 * we need to check the pinned_extents for any extents that can't be used yet 355 * since their free space will be released as soon as the transaction commits. 356 */ 357 u64 add_new_free_space(struct btrfs_block_group_cache *block_group, 358 struct btrfs_fs_info *info, u64 start, u64 end) 359 { 360 u64 extent_start, extent_end, size, total_added = 0; 361 int ret; 362 363 while (start < end) { 364 ret = find_first_extent_bit(info->pinned_extents, start, 365 &extent_start, &extent_end, 366 EXTENT_DIRTY | EXTENT_UPTODATE, 367 NULL); 368 if (ret) 369 break; 370 371 if (extent_start <= start) { 372 start = extent_end + 1; 373 } else if (extent_start > start && extent_start < end) { 374 size = extent_start - start; 375 total_added += size; 376 ret = btrfs_add_free_space(block_group, start, 377 size); 378 BUG_ON(ret); /* -ENOMEM or logic error */ 379 start = extent_end + 1; 380 } else { 381 break; 382 } 383 } 384 385 if (start < end) { 386 size = end - start; 387 total_added += size; 388 ret = btrfs_add_free_space(block_group, start, size); 389 BUG_ON(ret); /* -ENOMEM or logic error */ 390 } 391 392 return total_added; 393 } 394 395 static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl) 396 { 397 struct btrfs_block_group_cache *block_group = caching_ctl->block_group; 398 struct btrfs_fs_info *fs_info = block_group->fs_info; 399 struct btrfs_root *extent_root = fs_info->extent_root; 400 struct btrfs_path *path; 401 struct extent_buffer *leaf; 402 struct btrfs_key key; 403 u64 total_found = 0; 404 u64 last = 0; 405 u32 nritems; 406 int ret; 407 bool wakeup = true; 408 409 path = btrfs_alloc_path(); 410 if (!path) 411 return -ENOMEM; 412 413 last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET); 414 415 #ifdef CONFIG_BTRFS_DEBUG 416 /* 417 * If we're fragmenting we don't want to make anybody think we can 418 * allocate from this block group until we've had a chance to fragment 419 * the free space. 420 */ 421 if (btrfs_should_fragment_free_space(block_group)) 422 wakeup = false; 423 #endif 424 /* 425 * We don't want to deadlock with somebody trying to allocate a new 426 * extent for the extent root while also trying to search the extent 427 * root to add free space. So we skip locking and search the commit 428 * root, since its read-only 429 */ 430 path->skip_locking = 1; 431 path->search_commit_root = 1; 432 path->reada = READA_FORWARD; 433 434 key.objectid = last; 435 key.offset = 0; 436 key.type = BTRFS_EXTENT_ITEM_KEY; 437 438 next: 439 ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0); 440 if (ret < 0) 441 goto out; 442 443 leaf = path->nodes[0]; 444 nritems = btrfs_header_nritems(leaf); 445 446 while (1) { 447 if (btrfs_fs_closing(fs_info) > 1) { 448 last = (u64)-1; 449 break; 450 } 451 452 if (path->slots[0] < nritems) { 453 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 454 } else { 455 ret = find_next_key(path, 0, &key); 456 if (ret) 457 break; 458 459 if (need_resched() || 460 rwsem_is_contended(&fs_info->commit_root_sem)) { 461 if (wakeup) 462 caching_ctl->progress = last; 463 btrfs_release_path(path); 464 up_read(&fs_info->commit_root_sem); 465 mutex_unlock(&caching_ctl->mutex); 466 cond_resched(); 467 mutex_lock(&caching_ctl->mutex); 468 down_read(&fs_info->commit_root_sem); 469 goto next; 470 } 471 472 ret = btrfs_next_leaf(extent_root, path); 473 if (ret < 0) 474 goto out; 475 if (ret) 476 break; 477 leaf = path->nodes[0]; 478 nritems = btrfs_header_nritems(leaf); 479 continue; 480 } 481 482 if (key.objectid < last) { 483 key.objectid = last; 484 key.offset = 0; 485 key.type = BTRFS_EXTENT_ITEM_KEY; 486 487 if (wakeup) 488 caching_ctl->progress = last; 489 btrfs_release_path(path); 490 goto next; 491 } 492 493 if (key.objectid < block_group->key.objectid) { 494 path->slots[0]++; 495 continue; 496 } 497 498 if (key.objectid >= block_group->key.objectid + 499 block_group->key.offset) 500 break; 501 502 if (key.type == BTRFS_EXTENT_ITEM_KEY || 503 key.type == BTRFS_METADATA_ITEM_KEY) { 504 total_found += add_new_free_space(block_group, 505 fs_info, last, 506 key.objectid); 507 if (key.type == BTRFS_METADATA_ITEM_KEY) 508 last = key.objectid + 509 fs_info->nodesize; 510 else 511 last = key.objectid + key.offset; 512 513 if (total_found > CACHING_CTL_WAKE_UP) { 514 total_found = 0; 515 if (wakeup) 516 wake_up(&caching_ctl->wait); 517 } 518 } 519 path->slots[0]++; 520 } 521 ret = 0; 522 523 total_found += add_new_free_space(block_group, fs_info, last, 524 block_group->key.objectid + 525 block_group->key.offset); 526 caching_ctl->progress = (u64)-1; 527 528 out: 529 btrfs_free_path(path); 530 return ret; 531 } 532 533 static noinline void caching_thread(struct btrfs_work *work) 534 { 535 struct btrfs_block_group_cache *block_group; 536 struct btrfs_fs_info *fs_info; 537 struct btrfs_caching_control *caching_ctl; 538 struct btrfs_root *extent_root; 539 int ret; 540 541 caching_ctl = container_of(work, struct btrfs_caching_control, work); 542 block_group = caching_ctl->block_group; 543 fs_info = block_group->fs_info; 544 extent_root = fs_info->extent_root; 545 546 mutex_lock(&caching_ctl->mutex); 547 down_read(&fs_info->commit_root_sem); 548 549 if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE)) 550 ret = load_free_space_tree(caching_ctl); 551 else 552 ret = load_extent_tree_free(caching_ctl); 553 554 spin_lock(&block_group->lock); 555 block_group->caching_ctl = NULL; 556 block_group->cached = ret ? BTRFS_CACHE_ERROR : BTRFS_CACHE_FINISHED; 557 spin_unlock(&block_group->lock); 558 559 #ifdef CONFIG_BTRFS_DEBUG 560 if (btrfs_should_fragment_free_space(block_group)) { 561 u64 bytes_used; 562 563 spin_lock(&block_group->space_info->lock); 564 spin_lock(&block_group->lock); 565 bytes_used = block_group->key.offset - 566 btrfs_block_group_used(&block_group->item); 567 block_group->space_info->bytes_used += bytes_used >> 1; 568 spin_unlock(&block_group->lock); 569 spin_unlock(&block_group->space_info->lock); 570 fragment_free_space(block_group); 571 } 572 #endif 573 574 caching_ctl->progress = (u64)-1; 575 576 up_read(&fs_info->commit_root_sem); 577 free_excluded_extents(fs_info, block_group); 578 mutex_unlock(&caching_ctl->mutex); 579 580 wake_up(&caching_ctl->wait); 581 582 put_caching_control(caching_ctl); 583 btrfs_put_block_group(block_group); 584 } 585 586 static int cache_block_group(struct btrfs_block_group_cache *cache, 587 int load_cache_only) 588 { 589 DEFINE_WAIT(wait); 590 struct btrfs_fs_info *fs_info = cache->fs_info; 591 struct btrfs_caching_control *caching_ctl; 592 int ret = 0; 593 594 caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS); 595 if (!caching_ctl) 596 return -ENOMEM; 597 598 INIT_LIST_HEAD(&caching_ctl->list); 599 mutex_init(&caching_ctl->mutex); 600 init_waitqueue_head(&caching_ctl->wait); 601 caching_ctl->block_group = cache; 602 caching_ctl->progress = cache->key.objectid; 603 refcount_set(&caching_ctl->count, 1); 604 btrfs_init_work(&caching_ctl->work, btrfs_cache_helper, 605 caching_thread, NULL, NULL); 606 607 spin_lock(&cache->lock); 608 /* 609 * This should be a rare occasion, but this could happen I think in the 610 * case where one thread starts to load the space cache info, and then 611 * some other thread starts a transaction commit which tries to do an 612 * allocation while the other thread is still loading the space cache 613 * info. The previous loop should have kept us from choosing this block 614 * group, but if we've moved to the state where we will wait on caching 615 * block groups we need to first check if we're doing a fast load here, 616 * so we can wait for it to finish, otherwise we could end up allocating 617 * from a block group who's cache gets evicted for one reason or 618 * another. 619 */ 620 while (cache->cached == BTRFS_CACHE_FAST) { 621 struct btrfs_caching_control *ctl; 622 623 ctl = cache->caching_ctl; 624 refcount_inc(&ctl->count); 625 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE); 626 spin_unlock(&cache->lock); 627 628 schedule(); 629 630 finish_wait(&ctl->wait, &wait); 631 put_caching_control(ctl); 632 spin_lock(&cache->lock); 633 } 634 635 if (cache->cached != BTRFS_CACHE_NO) { 636 spin_unlock(&cache->lock); 637 kfree(caching_ctl); 638 return 0; 639 } 640 WARN_ON(cache->caching_ctl); 641 cache->caching_ctl = caching_ctl; 642 cache->cached = BTRFS_CACHE_FAST; 643 spin_unlock(&cache->lock); 644 645 if (btrfs_test_opt(fs_info, SPACE_CACHE)) { 646 mutex_lock(&caching_ctl->mutex); 647 ret = load_free_space_cache(fs_info, cache); 648 649 spin_lock(&cache->lock); 650 if (ret == 1) { 651 cache->caching_ctl = NULL; 652 cache->cached = BTRFS_CACHE_FINISHED; 653 cache->last_byte_to_unpin = (u64)-1; 654 caching_ctl->progress = (u64)-1; 655 } else { 656 if (load_cache_only) { 657 cache->caching_ctl = NULL; 658 cache->cached = BTRFS_CACHE_NO; 659 } else { 660 cache->cached = BTRFS_CACHE_STARTED; 661 cache->has_caching_ctl = 1; 662 } 663 } 664 spin_unlock(&cache->lock); 665 #ifdef CONFIG_BTRFS_DEBUG 666 if (ret == 1 && 667 btrfs_should_fragment_free_space(cache)) { 668 u64 bytes_used; 669 670 spin_lock(&cache->space_info->lock); 671 spin_lock(&cache->lock); 672 bytes_used = cache->key.offset - 673 btrfs_block_group_used(&cache->item); 674 cache->space_info->bytes_used += bytes_used >> 1; 675 spin_unlock(&cache->lock); 676 spin_unlock(&cache->space_info->lock); 677 fragment_free_space(cache); 678 } 679 #endif 680 mutex_unlock(&caching_ctl->mutex); 681 682 wake_up(&caching_ctl->wait); 683 if (ret == 1) { 684 put_caching_control(caching_ctl); 685 free_excluded_extents(fs_info, cache); 686 return 0; 687 } 688 } else { 689 /* 690 * We're either using the free space tree or no caching at all. 691 * Set cached to the appropriate value and wakeup any waiters. 692 */ 693 spin_lock(&cache->lock); 694 if (load_cache_only) { 695 cache->caching_ctl = NULL; 696 cache->cached = BTRFS_CACHE_NO; 697 } else { 698 cache->cached = BTRFS_CACHE_STARTED; 699 cache->has_caching_ctl = 1; 700 } 701 spin_unlock(&cache->lock); 702 wake_up(&caching_ctl->wait); 703 } 704 705 if (load_cache_only) { 706 put_caching_control(caching_ctl); 707 return 0; 708 } 709 710 down_write(&fs_info->commit_root_sem); 711 refcount_inc(&caching_ctl->count); 712 list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups); 713 up_write(&fs_info->commit_root_sem); 714 715 btrfs_get_block_group(cache); 716 717 btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work); 718 719 return ret; 720 } 721 722 /* 723 * return the block group that starts at or after bytenr 724 */ 725 static struct btrfs_block_group_cache * 726 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr) 727 { 728 return block_group_cache_tree_search(info, bytenr, 0); 729 } 730 731 /* 732 * return the block group that contains the given bytenr 733 */ 734 struct btrfs_block_group_cache *btrfs_lookup_block_group( 735 struct btrfs_fs_info *info, 736 u64 bytenr) 737 { 738 return block_group_cache_tree_search(info, bytenr, 1); 739 } 740 741 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info, 742 u64 flags) 743 { 744 struct list_head *head = &info->space_info; 745 struct btrfs_space_info *found; 746 747 flags &= BTRFS_BLOCK_GROUP_TYPE_MASK; 748 749 rcu_read_lock(); 750 list_for_each_entry_rcu(found, head, list) { 751 if (found->flags & flags) { 752 rcu_read_unlock(); 753 return found; 754 } 755 } 756 rcu_read_unlock(); 757 return NULL; 758 } 759 760 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, s64 num_bytes, 761 u64 owner, u64 root_objectid) 762 { 763 struct btrfs_space_info *space_info; 764 u64 flags; 765 766 if (owner < BTRFS_FIRST_FREE_OBJECTID) { 767 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID) 768 flags = BTRFS_BLOCK_GROUP_SYSTEM; 769 else 770 flags = BTRFS_BLOCK_GROUP_METADATA; 771 } else { 772 flags = BTRFS_BLOCK_GROUP_DATA; 773 } 774 775 space_info = __find_space_info(fs_info, flags); 776 ASSERT(space_info); 777 percpu_counter_add(&space_info->total_bytes_pinned, num_bytes); 778 } 779 780 /* 781 * after adding space to the filesystem, we need to clear the full flags 782 * on all the space infos. 783 */ 784 void btrfs_clear_space_info_full(struct btrfs_fs_info *info) 785 { 786 struct list_head *head = &info->space_info; 787 struct btrfs_space_info *found; 788 789 rcu_read_lock(); 790 list_for_each_entry_rcu(found, head, list) 791 found->full = 0; 792 rcu_read_unlock(); 793 } 794 795 /* simple helper to search for an existing data extent at a given offset */ 796 int btrfs_lookup_data_extent(struct btrfs_fs_info *fs_info, u64 start, u64 len) 797 { 798 int ret; 799 struct btrfs_key key; 800 struct btrfs_path *path; 801 802 path = btrfs_alloc_path(); 803 if (!path) 804 return -ENOMEM; 805 806 key.objectid = start; 807 key.offset = len; 808 key.type = BTRFS_EXTENT_ITEM_KEY; 809 ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0); 810 btrfs_free_path(path); 811 return ret; 812 } 813 814 /* 815 * helper function to lookup reference count and flags of a tree block. 816 * 817 * the head node for delayed ref is used to store the sum of all the 818 * reference count modifications queued up in the rbtree. the head 819 * node may also store the extent flags to set. This way you can check 820 * to see what the reference count and extent flags would be if all of 821 * the delayed refs are not processed. 822 */ 823 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans, 824 struct btrfs_fs_info *fs_info, u64 bytenr, 825 u64 offset, int metadata, u64 *refs, u64 *flags) 826 { 827 struct btrfs_delayed_ref_head *head; 828 struct btrfs_delayed_ref_root *delayed_refs; 829 struct btrfs_path *path; 830 struct btrfs_extent_item *ei; 831 struct extent_buffer *leaf; 832 struct btrfs_key key; 833 u32 item_size; 834 u64 num_refs; 835 u64 extent_flags; 836 int ret; 837 838 /* 839 * If we don't have skinny metadata, don't bother doing anything 840 * different 841 */ 842 if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA)) { 843 offset = fs_info->nodesize; 844 metadata = 0; 845 } 846 847 path = btrfs_alloc_path(); 848 if (!path) 849 return -ENOMEM; 850 851 if (!trans) { 852 path->skip_locking = 1; 853 path->search_commit_root = 1; 854 } 855 856 search_again: 857 key.objectid = bytenr; 858 key.offset = offset; 859 if (metadata) 860 key.type = BTRFS_METADATA_ITEM_KEY; 861 else 862 key.type = BTRFS_EXTENT_ITEM_KEY; 863 864 ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0); 865 if (ret < 0) 866 goto out_free; 867 868 if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) { 869 if (path->slots[0]) { 870 path->slots[0]--; 871 btrfs_item_key_to_cpu(path->nodes[0], &key, 872 path->slots[0]); 873 if (key.objectid == bytenr && 874 key.type == BTRFS_EXTENT_ITEM_KEY && 875 key.offset == fs_info->nodesize) 876 ret = 0; 877 } 878 } 879 880 if (ret == 0) { 881 leaf = path->nodes[0]; 882 item_size = btrfs_item_size_nr(leaf, path->slots[0]); 883 if (item_size >= sizeof(*ei)) { 884 ei = btrfs_item_ptr(leaf, path->slots[0], 885 struct btrfs_extent_item); 886 num_refs = btrfs_extent_refs(leaf, ei); 887 extent_flags = btrfs_extent_flags(leaf, ei); 888 } else { 889 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 890 struct btrfs_extent_item_v0 *ei0; 891 BUG_ON(item_size != sizeof(*ei0)); 892 ei0 = btrfs_item_ptr(leaf, path->slots[0], 893 struct btrfs_extent_item_v0); 894 num_refs = btrfs_extent_refs_v0(leaf, ei0); 895 /* FIXME: this isn't correct for data */ 896 extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF; 897 #else 898 BUG(); 899 #endif 900 } 901 BUG_ON(num_refs == 0); 902 } else { 903 num_refs = 0; 904 extent_flags = 0; 905 ret = 0; 906 } 907 908 if (!trans) 909 goto out; 910 911 delayed_refs = &trans->transaction->delayed_refs; 912 spin_lock(&delayed_refs->lock); 913 head = btrfs_find_delayed_ref_head(delayed_refs, bytenr); 914 if (head) { 915 if (!mutex_trylock(&head->mutex)) { 916 refcount_inc(&head->refs); 917 spin_unlock(&delayed_refs->lock); 918 919 btrfs_release_path(path); 920 921 /* 922 * Mutex was contended, block until it's released and try 923 * again 924 */ 925 mutex_lock(&head->mutex); 926 mutex_unlock(&head->mutex); 927 btrfs_put_delayed_ref_head(head); 928 goto search_again; 929 } 930 spin_lock(&head->lock); 931 if (head->extent_op && head->extent_op->update_flags) 932 extent_flags |= head->extent_op->flags_to_set; 933 else 934 BUG_ON(num_refs == 0); 935 936 num_refs += head->ref_mod; 937 spin_unlock(&head->lock); 938 mutex_unlock(&head->mutex); 939 } 940 spin_unlock(&delayed_refs->lock); 941 out: 942 WARN_ON(num_refs == 0); 943 if (refs) 944 *refs = num_refs; 945 if (flags) 946 *flags = extent_flags; 947 out_free: 948 btrfs_free_path(path); 949 return ret; 950 } 951 952 /* 953 * Back reference rules. Back refs have three main goals: 954 * 955 * 1) differentiate between all holders of references to an extent so that 956 * when a reference is dropped we can make sure it was a valid reference 957 * before freeing the extent. 958 * 959 * 2) Provide enough information to quickly find the holders of an extent 960 * if we notice a given block is corrupted or bad. 961 * 962 * 3) Make it easy to migrate blocks for FS shrinking or storage pool 963 * maintenance. This is actually the same as #2, but with a slightly 964 * different use case. 965 * 966 * There are two kinds of back refs. The implicit back refs is optimized 967 * for pointers in non-shared tree blocks. For a given pointer in a block, 968 * back refs of this kind provide information about the block's owner tree 969 * and the pointer's key. These information allow us to find the block by 970 * b-tree searching. The full back refs is for pointers in tree blocks not 971 * referenced by their owner trees. The location of tree block is recorded 972 * in the back refs. Actually the full back refs is generic, and can be 973 * used in all cases the implicit back refs is used. The major shortcoming 974 * of the full back refs is its overhead. Every time a tree block gets 975 * COWed, we have to update back refs entry for all pointers in it. 976 * 977 * For a newly allocated tree block, we use implicit back refs for 978 * pointers in it. This means most tree related operations only involve 979 * implicit back refs. For a tree block created in old transaction, the 980 * only way to drop a reference to it is COW it. So we can detect the 981 * event that tree block loses its owner tree's reference and do the 982 * back refs conversion. 983 * 984 * When a tree block is COWed through a tree, there are four cases: 985 * 986 * The reference count of the block is one and the tree is the block's 987 * owner tree. Nothing to do in this case. 988 * 989 * The reference count of the block is one and the tree is not the 990 * block's owner tree. In this case, full back refs is used for pointers 991 * in the block. Remove these full back refs, add implicit back refs for 992 * every pointers in the new block. 993 * 994 * The reference count of the block is greater than one and the tree is 995 * the block's owner tree. In this case, implicit back refs is used for 996 * pointers in the block. Add full back refs for every pointers in the 997 * block, increase lower level extents' reference counts. The original 998 * implicit back refs are entailed to the new block. 999 * 1000 * The reference count of the block is greater than one and the tree is 1001 * not the block's owner tree. Add implicit back refs for every pointer in 1002 * the new block, increase lower level extents' reference count. 1003 * 1004 * Back Reference Key composing: 1005 * 1006 * The key objectid corresponds to the first byte in the extent, 1007 * The key type is used to differentiate between types of back refs. 1008 * There are different meanings of the key offset for different types 1009 * of back refs. 1010 * 1011 * File extents can be referenced by: 1012 * 1013 * - multiple snapshots, subvolumes, or different generations in one subvol 1014 * - different files inside a single subvolume 1015 * - different offsets inside a file (bookend extents in file.c) 1016 * 1017 * The extent ref structure for the implicit back refs has fields for: 1018 * 1019 * - Objectid of the subvolume root 1020 * - objectid of the file holding the reference 1021 * - original offset in the file 1022 * - how many bookend extents 1023 * 1024 * The key offset for the implicit back refs is hash of the first 1025 * three fields. 1026 * 1027 * The extent ref structure for the full back refs has field for: 1028 * 1029 * - number of pointers in the tree leaf 1030 * 1031 * The key offset for the implicit back refs is the first byte of 1032 * the tree leaf 1033 * 1034 * When a file extent is allocated, The implicit back refs is used. 1035 * the fields are filled in: 1036 * 1037 * (root_key.objectid, inode objectid, offset in file, 1) 1038 * 1039 * When a file extent is removed file truncation, we find the 1040 * corresponding implicit back refs and check the following fields: 1041 * 1042 * (btrfs_header_owner(leaf), inode objectid, offset in file) 1043 * 1044 * Btree extents can be referenced by: 1045 * 1046 * - Different subvolumes 1047 * 1048 * Both the implicit back refs and the full back refs for tree blocks 1049 * only consist of key. The key offset for the implicit back refs is 1050 * objectid of block's owner tree. The key offset for the full back refs 1051 * is the first byte of parent block. 1052 * 1053 * When implicit back refs is used, information about the lowest key and 1054 * level of the tree block are required. These information are stored in 1055 * tree block info structure. 1056 */ 1057 1058 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 1059 static int convert_extent_item_v0(struct btrfs_trans_handle *trans, 1060 struct btrfs_fs_info *fs_info, 1061 struct btrfs_path *path, 1062 u64 owner, u32 extra_size) 1063 { 1064 struct btrfs_root *root = fs_info->extent_root; 1065 struct btrfs_extent_item *item; 1066 struct btrfs_extent_item_v0 *ei0; 1067 struct btrfs_extent_ref_v0 *ref0; 1068 struct btrfs_tree_block_info *bi; 1069 struct extent_buffer *leaf; 1070 struct btrfs_key key; 1071 struct btrfs_key found_key; 1072 u32 new_size = sizeof(*item); 1073 u64 refs; 1074 int ret; 1075 1076 leaf = path->nodes[0]; 1077 BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0)); 1078 1079 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 1080 ei0 = btrfs_item_ptr(leaf, path->slots[0], 1081 struct btrfs_extent_item_v0); 1082 refs = btrfs_extent_refs_v0(leaf, ei0); 1083 1084 if (owner == (u64)-1) { 1085 while (1) { 1086 if (path->slots[0] >= btrfs_header_nritems(leaf)) { 1087 ret = btrfs_next_leaf(root, path); 1088 if (ret < 0) 1089 return ret; 1090 BUG_ON(ret > 0); /* Corruption */ 1091 leaf = path->nodes[0]; 1092 } 1093 btrfs_item_key_to_cpu(leaf, &found_key, 1094 path->slots[0]); 1095 BUG_ON(key.objectid != found_key.objectid); 1096 if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) { 1097 path->slots[0]++; 1098 continue; 1099 } 1100 ref0 = btrfs_item_ptr(leaf, path->slots[0], 1101 struct btrfs_extent_ref_v0); 1102 owner = btrfs_ref_objectid_v0(leaf, ref0); 1103 break; 1104 } 1105 } 1106 btrfs_release_path(path); 1107 1108 if (owner < BTRFS_FIRST_FREE_OBJECTID) 1109 new_size += sizeof(*bi); 1110 1111 new_size -= sizeof(*ei0); 1112 ret = btrfs_search_slot(trans, root, &key, path, 1113 new_size + extra_size, 1); 1114 if (ret < 0) 1115 return ret; 1116 BUG_ON(ret); /* Corruption */ 1117 1118 btrfs_extend_item(fs_info, path, new_size); 1119 1120 leaf = path->nodes[0]; 1121 item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item); 1122 btrfs_set_extent_refs(leaf, item, refs); 1123 /* FIXME: get real generation */ 1124 btrfs_set_extent_generation(leaf, item, 0); 1125 if (owner < BTRFS_FIRST_FREE_OBJECTID) { 1126 btrfs_set_extent_flags(leaf, item, 1127 BTRFS_EXTENT_FLAG_TREE_BLOCK | 1128 BTRFS_BLOCK_FLAG_FULL_BACKREF); 1129 bi = (struct btrfs_tree_block_info *)(item + 1); 1130 /* FIXME: get first key of the block */ 1131 memzero_extent_buffer(leaf, (unsigned long)bi, sizeof(*bi)); 1132 btrfs_set_tree_block_level(leaf, bi, (int)owner); 1133 } else { 1134 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA); 1135 } 1136 btrfs_mark_buffer_dirty(leaf); 1137 return 0; 1138 } 1139 #endif 1140 1141 /* 1142 * is_data == BTRFS_REF_TYPE_BLOCK, tree block type is required, 1143 * is_data == BTRFS_REF_TYPE_DATA, data type is requried, 1144 * is_data == BTRFS_REF_TYPE_ANY, either type is OK. 1145 */ 1146 int btrfs_get_extent_inline_ref_type(const struct extent_buffer *eb, 1147 struct btrfs_extent_inline_ref *iref, 1148 enum btrfs_inline_ref_type is_data) 1149 { 1150 int type = btrfs_extent_inline_ref_type(eb, iref); 1151 u64 offset = btrfs_extent_inline_ref_offset(eb, iref); 1152 1153 if (type == BTRFS_TREE_BLOCK_REF_KEY || 1154 type == BTRFS_SHARED_BLOCK_REF_KEY || 1155 type == BTRFS_SHARED_DATA_REF_KEY || 1156 type == BTRFS_EXTENT_DATA_REF_KEY) { 1157 if (is_data == BTRFS_REF_TYPE_BLOCK) { 1158 if (type == BTRFS_TREE_BLOCK_REF_KEY) 1159 return type; 1160 if (type == BTRFS_SHARED_BLOCK_REF_KEY) { 1161 ASSERT(eb->fs_info); 1162 /* 1163 * Every shared one has parent tree 1164 * block, which must be aligned to 1165 * nodesize. 1166 */ 1167 if (offset && 1168 IS_ALIGNED(offset, eb->fs_info->nodesize)) 1169 return type; 1170 } 1171 } else if (is_data == BTRFS_REF_TYPE_DATA) { 1172 if (type == BTRFS_EXTENT_DATA_REF_KEY) 1173 return type; 1174 if (type == BTRFS_SHARED_DATA_REF_KEY) { 1175 ASSERT(eb->fs_info); 1176 /* 1177 * Every shared one has parent tree 1178 * block, which must be aligned to 1179 * nodesize. 1180 */ 1181 if (offset && 1182 IS_ALIGNED(offset, eb->fs_info->nodesize)) 1183 return type; 1184 } 1185 } else { 1186 ASSERT(is_data == BTRFS_REF_TYPE_ANY); 1187 return type; 1188 } 1189 } 1190 1191 btrfs_print_leaf((struct extent_buffer *)eb); 1192 btrfs_err(eb->fs_info, "eb %llu invalid extent inline ref type %d", 1193 eb->start, type); 1194 WARN_ON(1); 1195 1196 return BTRFS_REF_TYPE_INVALID; 1197 } 1198 1199 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset) 1200 { 1201 u32 high_crc = ~(u32)0; 1202 u32 low_crc = ~(u32)0; 1203 __le64 lenum; 1204 1205 lenum = cpu_to_le64(root_objectid); 1206 high_crc = btrfs_crc32c(high_crc, &lenum, sizeof(lenum)); 1207 lenum = cpu_to_le64(owner); 1208 low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum)); 1209 lenum = cpu_to_le64(offset); 1210 low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum)); 1211 1212 return ((u64)high_crc << 31) ^ (u64)low_crc; 1213 } 1214 1215 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf, 1216 struct btrfs_extent_data_ref *ref) 1217 { 1218 return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref), 1219 btrfs_extent_data_ref_objectid(leaf, ref), 1220 btrfs_extent_data_ref_offset(leaf, ref)); 1221 } 1222 1223 static int match_extent_data_ref(struct extent_buffer *leaf, 1224 struct btrfs_extent_data_ref *ref, 1225 u64 root_objectid, u64 owner, u64 offset) 1226 { 1227 if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid || 1228 btrfs_extent_data_ref_objectid(leaf, ref) != owner || 1229 btrfs_extent_data_ref_offset(leaf, ref) != offset) 1230 return 0; 1231 return 1; 1232 } 1233 1234 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans, 1235 struct btrfs_fs_info *fs_info, 1236 struct btrfs_path *path, 1237 u64 bytenr, u64 parent, 1238 u64 root_objectid, 1239 u64 owner, u64 offset) 1240 { 1241 struct btrfs_root *root = fs_info->extent_root; 1242 struct btrfs_key key; 1243 struct btrfs_extent_data_ref *ref; 1244 struct extent_buffer *leaf; 1245 u32 nritems; 1246 int ret; 1247 int recow; 1248 int err = -ENOENT; 1249 1250 key.objectid = bytenr; 1251 if (parent) { 1252 key.type = BTRFS_SHARED_DATA_REF_KEY; 1253 key.offset = parent; 1254 } else { 1255 key.type = BTRFS_EXTENT_DATA_REF_KEY; 1256 key.offset = hash_extent_data_ref(root_objectid, 1257 owner, offset); 1258 } 1259 again: 1260 recow = 0; 1261 ret = btrfs_search_slot(trans, root, &key, path, -1, 1); 1262 if (ret < 0) { 1263 err = ret; 1264 goto fail; 1265 } 1266 1267 if (parent) { 1268 if (!ret) 1269 return 0; 1270 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 1271 key.type = BTRFS_EXTENT_REF_V0_KEY; 1272 btrfs_release_path(path); 1273 ret = btrfs_search_slot(trans, root, &key, path, -1, 1); 1274 if (ret < 0) { 1275 err = ret; 1276 goto fail; 1277 } 1278 if (!ret) 1279 return 0; 1280 #endif 1281 goto fail; 1282 } 1283 1284 leaf = path->nodes[0]; 1285 nritems = btrfs_header_nritems(leaf); 1286 while (1) { 1287 if (path->slots[0] >= nritems) { 1288 ret = btrfs_next_leaf(root, path); 1289 if (ret < 0) 1290 err = ret; 1291 if (ret) 1292 goto fail; 1293 1294 leaf = path->nodes[0]; 1295 nritems = btrfs_header_nritems(leaf); 1296 recow = 1; 1297 } 1298 1299 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 1300 if (key.objectid != bytenr || 1301 key.type != BTRFS_EXTENT_DATA_REF_KEY) 1302 goto fail; 1303 1304 ref = btrfs_item_ptr(leaf, path->slots[0], 1305 struct btrfs_extent_data_ref); 1306 1307 if (match_extent_data_ref(leaf, ref, root_objectid, 1308 owner, offset)) { 1309 if (recow) { 1310 btrfs_release_path(path); 1311 goto again; 1312 } 1313 err = 0; 1314 break; 1315 } 1316 path->slots[0]++; 1317 } 1318 fail: 1319 return err; 1320 } 1321 1322 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans, 1323 struct btrfs_fs_info *fs_info, 1324 struct btrfs_path *path, 1325 u64 bytenr, u64 parent, 1326 u64 root_objectid, u64 owner, 1327 u64 offset, int refs_to_add) 1328 { 1329 struct btrfs_root *root = fs_info->extent_root; 1330 struct btrfs_key key; 1331 struct extent_buffer *leaf; 1332 u32 size; 1333 u32 num_refs; 1334 int ret; 1335 1336 key.objectid = bytenr; 1337 if (parent) { 1338 key.type = BTRFS_SHARED_DATA_REF_KEY; 1339 key.offset = parent; 1340 size = sizeof(struct btrfs_shared_data_ref); 1341 } else { 1342 key.type = BTRFS_EXTENT_DATA_REF_KEY; 1343 key.offset = hash_extent_data_ref(root_objectid, 1344 owner, offset); 1345 size = sizeof(struct btrfs_extent_data_ref); 1346 } 1347 1348 ret = btrfs_insert_empty_item(trans, root, path, &key, size); 1349 if (ret && ret != -EEXIST) 1350 goto fail; 1351 1352 leaf = path->nodes[0]; 1353 if (parent) { 1354 struct btrfs_shared_data_ref *ref; 1355 ref = btrfs_item_ptr(leaf, path->slots[0], 1356 struct btrfs_shared_data_ref); 1357 if (ret == 0) { 1358 btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add); 1359 } else { 1360 num_refs = btrfs_shared_data_ref_count(leaf, ref); 1361 num_refs += refs_to_add; 1362 btrfs_set_shared_data_ref_count(leaf, ref, num_refs); 1363 } 1364 } else { 1365 struct btrfs_extent_data_ref *ref; 1366 while (ret == -EEXIST) { 1367 ref = btrfs_item_ptr(leaf, path->slots[0], 1368 struct btrfs_extent_data_ref); 1369 if (match_extent_data_ref(leaf, ref, root_objectid, 1370 owner, offset)) 1371 break; 1372 btrfs_release_path(path); 1373 key.offset++; 1374 ret = btrfs_insert_empty_item(trans, root, path, &key, 1375 size); 1376 if (ret && ret != -EEXIST) 1377 goto fail; 1378 1379 leaf = path->nodes[0]; 1380 } 1381 ref = btrfs_item_ptr(leaf, path->slots[0], 1382 struct btrfs_extent_data_ref); 1383 if (ret == 0) { 1384 btrfs_set_extent_data_ref_root(leaf, ref, 1385 root_objectid); 1386 btrfs_set_extent_data_ref_objectid(leaf, ref, owner); 1387 btrfs_set_extent_data_ref_offset(leaf, ref, offset); 1388 btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add); 1389 } else { 1390 num_refs = btrfs_extent_data_ref_count(leaf, ref); 1391 num_refs += refs_to_add; 1392 btrfs_set_extent_data_ref_count(leaf, ref, num_refs); 1393 } 1394 } 1395 btrfs_mark_buffer_dirty(leaf); 1396 ret = 0; 1397 fail: 1398 btrfs_release_path(path); 1399 return ret; 1400 } 1401 1402 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans, 1403 struct btrfs_fs_info *fs_info, 1404 struct btrfs_path *path, 1405 int refs_to_drop, int *last_ref) 1406 { 1407 struct btrfs_key key; 1408 struct btrfs_extent_data_ref *ref1 = NULL; 1409 struct btrfs_shared_data_ref *ref2 = NULL; 1410 struct extent_buffer *leaf; 1411 u32 num_refs = 0; 1412 int ret = 0; 1413 1414 leaf = path->nodes[0]; 1415 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 1416 1417 if (key.type == BTRFS_EXTENT_DATA_REF_KEY) { 1418 ref1 = btrfs_item_ptr(leaf, path->slots[0], 1419 struct btrfs_extent_data_ref); 1420 num_refs = btrfs_extent_data_ref_count(leaf, ref1); 1421 } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) { 1422 ref2 = btrfs_item_ptr(leaf, path->slots[0], 1423 struct btrfs_shared_data_ref); 1424 num_refs = btrfs_shared_data_ref_count(leaf, ref2); 1425 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 1426 } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) { 1427 struct btrfs_extent_ref_v0 *ref0; 1428 ref0 = btrfs_item_ptr(leaf, path->slots[0], 1429 struct btrfs_extent_ref_v0); 1430 num_refs = btrfs_ref_count_v0(leaf, ref0); 1431 #endif 1432 } else { 1433 BUG(); 1434 } 1435 1436 BUG_ON(num_refs < refs_to_drop); 1437 num_refs -= refs_to_drop; 1438 1439 if (num_refs == 0) { 1440 ret = btrfs_del_item(trans, fs_info->extent_root, path); 1441 *last_ref = 1; 1442 } else { 1443 if (key.type == BTRFS_EXTENT_DATA_REF_KEY) 1444 btrfs_set_extent_data_ref_count(leaf, ref1, num_refs); 1445 else if (key.type == BTRFS_SHARED_DATA_REF_KEY) 1446 btrfs_set_shared_data_ref_count(leaf, ref2, num_refs); 1447 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 1448 else { 1449 struct btrfs_extent_ref_v0 *ref0; 1450 ref0 = btrfs_item_ptr(leaf, path->slots[0], 1451 struct btrfs_extent_ref_v0); 1452 btrfs_set_ref_count_v0(leaf, ref0, num_refs); 1453 } 1454 #endif 1455 btrfs_mark_buffer_dirty(leaf); 1456 } 1457 return ret; 1458 } 1459 1460 static noinline u32 extent_data_ref_count(struct btrfs_path *path, 1461 struct btrfs_extent_inline_ref *iref) 1462 { 1463 struct btrfs_key key; 1464 struct extent_buffer *leaf; 1465 struct btrfs_extent_data_ref *ref1; 1466 struct btrfs_shared_data_ref *ref2; 1467 u32 num_refs = 0; 1468 int type; 1469 1470 leaf = path->nodes[0]; 1471 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 1472 if (iref) { 1473 /* 1474 * If type is invalid, we should have bailed out earlier than 1475 * this call. 1476 */ 1477 type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA); 1478 ASSERT(type != BTRFS_REF_TYPE_INVALID); 1479 if (type == BTRFS_EXTENT_DATA_REF_KEY) { 1480 ref1 = (struct btrfs_extent_data_ref *)(&iref->offset); 1481 num_refs = btrfs_extent_data_ref_count(leaf, ref1); 1482 } else { 1483 ref2 = (struct btrfs_shared_data_ref *)(iref + 1); 1484 num_refs = btrfs_shared_data_ref_count(leaf, ref2); 1485 } 1486 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) { 1487 ref1 = btrfs_item_ptr(leaf, path->slots[0], 1488 struct btrfs_extent_data_ref); 1489 num_refs = btrfs_extent_data_ref_count(leaf, ref1); 1490 } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) { 1491 ref2 = btrfs_item_ptr(leaf, path->slots[0], 1492 struct btrfs_shared_data_ref); 1493 num_refs = btrfs_shared_data_ref_count(leaf, ref2); 1494 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 1495 } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) { 1496 struct btrfs_extent_ref_v0 *ref0; 1497 ref0 = btrfs_item_ptr(leaf, path->slots[0], 1498 struct btrfs_extent_ref_v0); 1499 num_refs = btrfs_ref_count_v0(leaf, ref0); 1500 #endif 1501 } else { 1502 WARN_ON(1); 1503 } 1504 return num_refs; 1505 } 1506 1507 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans, 1508 struct btrfs_fs_info *fs_info, 1509 struct btrfs_path *path, 1510 u64 bytenr, u64 parent, 1511 u64 root_objectid) 1512 { 1513 struct btrfs_root *root = fs_info->extent_root; 1514 struct btrfs_key key; 1515 int ret; 1516 1517 key.objectid = bytenr; 1518 if (parent) { 1519 key.type = BTRFS_SHARED_BLOCK_REF_KEY; 1520 key.offset = parent; 1521 } else { 1522 key.type = BTRFS_TREE_BLOCK_REF_KEY; 1523 key.offset = root_objectid; 1524 } 1525 1526 ret = btrfs_search_slot(trans, root, &key, path, -1, 1); 1527 if (ret > 0) 1528 ret = -ENOENT; 1529 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 1530 if (ret == -ENOENT && parent) { 1531 btrfs_release_path(path); 1532 key.type = BTRFS_EXTENT_REF_V0_KEY; 1533 ret = btrfs_search_slot(trans, root, &key, path, -1, 1); 1534 if (ret > 0) 1535 ret = -ENOENT; 1536 } 1537 #endif 1538 return ret; 1539 } 1540 1541 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans, 1542 struct btrfs_fs_info *fs_info, 1543 struct btrfs_path *path, 1544 u64 bytenr, u64 parent, 1545 u64 root_objectid) 1546 { 1547 struct btrfs_key key; 1548 int ret; 1549 1550 key.objectid = bytenr; 1551 if (parent) { 1552 key.type = BTRFS_SHARED_BLOCK_REF_KEY; 1553 key.offset = parent; 1554 } else { 1555 key.type = BTRFS_TREE_BLOCK_REF_KEY; 1556 key.offset = root_objectid; 1557 } 1558 1559 ret = btrfs_insert_empty_item(trans, fs_info->extent_root, 1560 path, &key, 0); 1561 btrfs_release_path(path); 1562 return ret; 1563 } 1564 1565 static inline int extent_ref_type(u64 parent, u64 owner) 1566 { 1567 int type; 1568 if (owner < BTRFS_FIRST_FREE_OBJECTID) { 1569 if (parent > 0) 1570 type = BTRFS_SHARED_BLOCK_REF_KEY; 1571 else 1572 type = BTRFS_TREE_BLOCK_REF_KEY; 1573 } else { 1574 if (parent > 0) 1575 type = BTRFS_SHARED_DATA_REF_KEY; 1576 else 1577 type = BTRFS_EXTENT_DATA_REF_KEY; 1578 } 1579 return type; 1580 } 1581 1582 static int find_next_key(struct btrfs_path *path, int level, 1583 struct btrfs_key *key) 1584 1585 { 1586 for (; level < BTRFS_MAX_LEVEL; level++) { 1587 if (!path->nodes[level]) 1588 break; 1589 if (path->slots[level] + 1 >= 1590 btrfs_header_nritems(path->nodes[level])) 1591 continue; 1592 if (level == 0) 1593 btrfs_item_key_to_cpu(path->nodes[level], key, 1594 path->slots[level] + 1); 1595 else 1596 btrfs_node_key_to_cpu(path->nodes[level], key, 1597 path->slots[level] + 1); 1598 return 0; 1599 } 1600 return 1; 1601 } 1602 1603 /* 1604 * look for inline back ref. if back ref is found, *ref_ret is set 1605 * to the address of inline back ref, and 0 is returned. 1606 * 1607 * if back ref isn't found, *ref_ret is set to the address where it 1608 * should be inserted, and -ENOENT is returned. 1609 * 1610 * if insert is true and there are too many inline back refs, the path 1611 * points to the extent item, and -EAGAIN is returned. 1612 * 1613 * NOTE: inline back refs are ordered in the same way that back ref 1614 * items in the tree are ordered. 1615 */ 1616 static noinline_for_stack 1617 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans, 1618 struct btrfs_fs_info *fs_info, 1619 struct btrfs_path *path, 1620 struct btrfs_extent_inline_ref **ref_ret, 1621 u64 bytenr, u64 num_bytes, 1622 u64 parent, u64 root_objectid, 1623 u64 owner, u64 offset, int insert) 1624 { 1625 struct btrfs_root *root = fs_info->extent_root; 1626 struct btrfs_key key; 1627 struct extent_buffer *leaf; 1628 struct btrfs_extent_item *ei; 1629 struct btrfs_extent_inline_ref *iref; 1630 u64 flags; 1631 u64 item_size; 1632 unsigned long ptr; 1633 unsigned long end; 1634 int extra_size; 1635 int type; 1636 int want; 1637 int ret; 1638 int err = 0; 1639 bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA); 1640 int needed; 1641 1642 key.objectid = bytenr; 1643 key.type = BTRFS_EXTENT_ITEM_KEY; 1644 key.offset = num_bytes; 1645 1646 want = extent_ref_type(parent, owner); 1647 if (insert) { 1648 extra_size = btrfs_extent_inline_ref_size(want); 1649 path->keep_locks = 1; 1650 } else 1651 extra_size = -1; 1652 1653 /* 1654 * Owner is our parent level, so we can just add one to get the level 1655 * for the block we are interested in. 1656 */ 1657 if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) { 1658 key.type = BTRFS_METADATA_ITEM_KEY; 1659 key.offset = owner; 1660 } 1661 1662 again: 1663 ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1); 1664 if (ret < 0) { 1665 err = ret; 1666 goto out; 1667 } 1668 1669 /* 1670 * We may be a newly converted file system which still has the old fat 1671 * extent entries for metadata, so try and see if we have one of those. 1672 */ 1673 if (ret > 0 && skinny_metadata) { 1674 skinny_metadata = false; 1675 if (path->slots[0]) { 1676 path->slots[0]--; 1677 btrfs_item_key_to_cpu(path->nodes[0], &key, 1678 path->slots[0]); 1679 if (key.objectid == bytenr && 1680 key.type == BTRFS_EXTENT_ITEM_KEY && 1681 key.offset == num_bytes) 1682 ret = 0; 1683 } 1684 if (ret) { 1685 key.objectid = bytenr; 1686 key.type = BTRFS_EXTENT_ITEM_KEY; 1687 key.offset = num_bytes; 1688 btrfs_release_path(path); 1689 goto again; 1690 } 1691 } 1692 1693 if (ret && !insert) { 1694 err = -ENOENT; 1695 goto out; 1696 } else if (WARN_ON(ret)) { 1697 err = -EIO; 1698 goto out; 1699 } 1700 1701 leaf = path->nodes[0]; 1702 item_size = btrfs_item_size_nr(leaf, path->slots[0]); 1703 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 1704 if (item_size < sizeof(*ei)) { 1705 if (!insert) { 1706 err = -ENOENT; 1707 goto out; 1708 } 1709 ret = convert_extent_item_v0(trans, fs_info, path, owner, 1710 extra_size); 1711 if (ret < 0) { 1712 err = ret; 1713 goto out; 1714 } 1715 leaf = path->nodes[0]; 1716 item_size = btrfs_item_size_nr(leaf, path->slots[0]); 1717 } 1718 #endif 1719 BUG_ON(item_size < sizeof(*ei)); 1720 1721 ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item); 1722 flags = btrfs_extent_flags(leaf, ei); 1723 1724 ptr = (unsigned long)(ei + 1); 1725 end = (unsigned long)ei + item_size; 1726 1727 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) { 1728 ptr += sizeof(struct btrfs_tree_block_info); 1729 BUG_ON(ptr > end); 1730 } 1731 1732 if (owner >= BTRFS_FIRST_FREE_OBJECTID) 1733 needed = BTRFS_REF_TYPE_DATA; 1734 else 1735 needed = BTRFS_REF_TYPE_BLOCK; 1736 1737 err = -ENOENT; 1738 while (1) { 1739 if (ptr >= end) { 1740 WARN_ON(ptr > end); 1741 break; 1742 } 1743 iref = (struct btrfs_extent_inline_ref *)ptr; 1744 type = btrfs_get_extent_inline_ref_type(leaf, iref, needed); 1745 if (type == BTRFS_REF_TYPE_INVALID) { 1746 err = -EINVAL; 1747 goto out; 1748 } 1749 1750 if (want < type) 1751 break; 1752 if (want > type) { 1753 ptr += btrfs_extent_inline_ref_size(type); 1754 continue; 1755 } 1756 1757 if (type == BTRFS_EXTENT_DATA_REF_KEY) { 1758 struct btrfs_extent_data_ref *dref; 1759 dref = (struct btrfs_extent_data_ref *)(&iref->offset); 1760 if (match_extent_data_ref(leaf, dref, root_objectid, 1761 owner, offset)) { 1762 err = 0; 1763 break; 1764 } 1765 if (hash_extent_data_ref_item(leaf, dref) < 1766 hash_extent_data_ref(root_objectid, owner, offset)) 1767 break; 1768 } else { 1769 u64 ref_offset; 1770 ref_offset = btrfs_extent_inline_ref_offset(leaf, iref); 1771 if (parent > 0) { 1772 if (parent == ref_offset) { 1773 err = 0; 1774 break; 1775 } 1776 if (ref_offset < parent) 1777 break; 1778 } else { 1779 if (root_objectid == ref_offset) { 1780 err = 0; 1781 break; 1782 } 1783 if (ref_offset < root_objectid) 1784 break; 1785 } 1786 } 1787 ptr += btrfs_extent_inline_ref_size(type); 1788 } 1789 if (err == -ENOENT && insert) { 1790 if (item_size + extra_size >= 1791 BTRFS_MAX_EXTENT_ITEM_SIZE(root)) { 1792 err = -EAGAIN; 1793 goto out; 1794 } 1795 /* 1796 * To add new inline back ref, we have to make sure 1797 * there is no corresponding back ref item. 1798 * For simplicity, we just do not add new inline back 1799 * ref if there is any kind of item for this block 1800 */ 1801 if (find_next_key(path, 0, &key) == 0 && 1802 key.objectid == bytenr && 1803 key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) { 1804 err = -EAGAIN; 1805 goto out; 1806 } 1807 } 1808 *ref_ret = (struct btrfs_extent_inline_ref *)ptr; 1809 out: 1810 if (insert) { 1811 path->keep_locks = 0; 1812 btrfs_unlock_up_safe(path, 1); 1813 } 1814 return err; 1815 } 1816 1817 /* 1818 * helper to add new inline back ref 1819 */ 1820 static noinline_for_stack 1821 void setup_inline_extent_backref(struct btrfs_fs_info *fs_info, 1822 struct btrfs_path *path, 1823 struct btrfs_extent_inline_ref *iref, 1824 u64 parent, u64 root_objectid, 1825 u64 owner, u64 offset, int refs_to_add, 1826 struct btrfs_delayed_extent_op *extent_op) 1827 { 1828 struct extent_buffer *leaf; 1829 struct btrfs_extent_item *ei; 1830 unsigned long ptr; 1831 unsigned long end; 1832 unsigned long item_offset; 1833 u64 refs; 1834 int size; 1835 int type; 1836 1837 leaf = path->nodes[0]; 1838 ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item); 1839 item_offset = (unsigned long)iref - (unsigned long)ei; 1840 1841 type = extent_ref_type(parent, owner); 1842 size = btrfs_extent_inline_ref_size(type); 1843 1844 btrfs_extend_item(fs_info, path, size); 1845 1846 ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item); 1847 refs = btrfs_extent_refs(leaf, ei); 1848 refs += refs_to_add; 1849 btrfs_set_extent_refs(leaf, ei, refs); 1850 if (extent_op) 1851 __run_delayed_extent_op(extent_op, leaf, ei); 1852 1853 ptr = (unsigned long)ei + item_offset; 1854 end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]); 1855 if (ptr < end - size) 1856 memmove_extent_buffer(leaf, ptr + size, ptr, 1857 end - size - ptr); 1858 1859 iref = (struct btrfs_extent_inline_ref *)ptr; 1860 btrfs_set_extent_inline_ref_type(leaf, iref, type); 1861 if (type == BTRFS_EXTENT_DATA_REF_KEY) { 1862 struct btrfs_extent_data_ref *dref; 1863 dref = (struct btrfs_extent_data_ref *)(&iref->offset); 1864 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid); 1865 btrfs_set_extent_data_ref_objectid(leaf, dref, owner); 1866 btrfs_set_extent_data_ref_offset(leaf, dref, offset); 1867 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add); 1868 } else if (type == BTRFS_SHARED_DATA_REF_KEY) { 1869 struct btrfs_shared_data_ref *sref; 1870 sref = (struct btrfs_shared_data_ref *)(iref + 1); 1871 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add); 1872 btrfs_set_extent_inline_ref_offset(leaf, iref, parent); 1873 } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) { 1874 btrfs_set_extent_inline_ref_offset(leaf, iref, parent); 1875 } else { 1876 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid); 1877 } 1878 btrfs_mark_buffer_dirty(leaf); 1879 } 1880 1881 static int lookup_extent_backref(struct btrfs_trans_handle *trans, 1882 struct btrfs_fs_info *fs_info, 1883 struct btrfs_path *path, 1884 struct btrfs_extent_inline_ref **ref_ret, 1885 u64 bytenr, u64 num_bytes, u64 parent, 1886 u64 root_objectid, u64 owner, u64 offset) 1887 { 1888 int ret; 1889 1890 ret = lookup_inline_extent_backref(trans, fs_info, path, ref_ret, 1891 bytenr, num_bytes, parent, 1892 root_objectid, owner, offset, 0); 1893 if (ret != -ENOENT) 1894 return ret; 1895 1896 btrfs_release_path(path); 1897 *ref_ret = NULL; 1898 1899 if (owner < BTRFS_FIRST_FREE_OBJECTID) { 1900 ret = lookup_tree_block_ref(trans, fs_info, path, bytenr, 1901 parent, root_objectid); 1902 } else { 1903 ret = lookup_extent_data_ref(trans, fs_info, path, bytenr, 1904 parent, root_objectid, owner, 1905 offset); 1906 } 1907 return ret; 1908 } 1909 1910 /* 1911 * helper to update/remove inline back ref 1912 */ 1913 static noinline_for_stack 1914 void update_inline_extent_backref(struct btrfs_fs_info *fs_info, 1915 struct btrfs_path *path, 1916 struct btrfs_extent_inline_ref *iref, 1917 int refs_to_mod, 1918 struct btrfs_delayed_extent_op *extent_op, 1919 int *last_ref) 1920 { 1921 struct extent_buffer *leaf; 1922 struct btrfs_extent_item *ei; 1923 struct btrfs_extent_data_ref *dref = NULL; 1924 struct btrfs_shared_data_ref *sref = NULL; 1925 unsigned long ptr; 1926 unsigned long end; 1927 u32 item_size; 1928 int size; 1929 int type; 1930 u64 refs; 1931 1932 leaf = path->nodes[0]; 1933 ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item); 1934 refs = btrfs_extent_refs(leaf, ei); 1935 WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0); 1936 refs += refs_to_mod; 1937 btrfs_set_extent_refs(leaf, ei, refs); 1938 if (extent_op) 1939 __run_delayed_extent_op(extent_op, leaf, ei); 1940 1941 /* 1942 * If type is invalid, we should have bailed out after 1943 * lookup_inline_extent_backref(). 1944 */ 1945 type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_ANY); 1946 ASSERT(type != BTRFS_REF_TYPE_INVALID); 1947 1948 if (type == BTRFS_EXTENT_DATA_REF_KEY) { 1949 dref = (struct btrfs_extent_data_ref *)(&iref->offset); 1950 refs = btrfs_extent_data_ref_count(leaf, dref); 1951 } else if (type == BTRFS_SHARED_DATA_REF_KEY) { 1952 sref = (struct btrfs_shared_data_ref *)(iref + 1); 1953 refs = btrfs_shared_data_ref_count(leaf, sref); 1954 } else { 1955 refs = 1; 1956 BUG_ON(refs_to_mod != -1); 1957 } 1958 1959 BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod); 1960 refs += refs_to_mod; 1961 1962 if (refs > 0) { 1963 if (type == BTRFS_EXTENT_DATA_REF_KEY) 1964 btrfs_set_extent_data_ref_count(leaf, dref, refs); 1965 else 1966 btrfs_set_shared_data_ref_count(leaf, sref, refs); 1967 } else { 1968 *last_ref = 1; 1969 size = btrfs_extent_inline_ref_size(type); 1970 item_size = btrfs_item_size_nr(leaf, path->slots[0]); 1971 ptr = (unsigned long)iref; 1972 end = (unsigned long)ei + item_size; 1973 if (ptr + size < end) 1974 memmove_extent_buffer(leaf, ptr, ptr + size, 1975 end - ptr - size); 1976 item_size -= size; 1977 btrfs_truncate_item(fs_info, path, item_size, 1); 1978 } 1979 btrfs_mark_buffer_dirty(leaf); 1980 } 1981 1982 static noinline_for_stack 1983 int insert_inline_extent_backref(struct btrfs_trans_handle *trans, 1984 struct btrfs_fs_info *fs_info, 1985 struct btrfs_path *path, 1986 u64 bytenr, u64 num_bytes, u64 parent, 1987 u64 root_objectid, u64 owner, 1988 u64 offset, int refs_to_add, 1989 struct btrfs_delayed_extent_op *extent_op) 1990 { 1991 struct btrfs_extent_inline_ref *iref; 1992 int ret; 1993 1994 ret = lookup_inline_extent_backref(trans, fs_info, path, &iref, 1995 bytenr, num_bytes, parent, 1996 root_objectid, owner, offset, 1); 1997 if (ret == 0) { 1998 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID); 1999 update_inline_extent_backref(fs_info, path, iref, 2000 refs_to_add, extent_op, NULL); 2001 } else if (ret == -ENOENT) { 2002 setup_inline_extent_backref(fs_info, path, iref, parent, 2003 root_objectid, owner, offset, 2004 refs_to_add, extent_op); 2005 ret = 0; 2006 } 2007 return ret; 2008 } 2009 2010 static int insert_extent_backref(struct btrfs_trans_handle *trans, 2011 struct btrfs_fs_info *fs_info, 2012 struct btrfs_path *path, 2013 u64 bytenr, u64 parent, u64 root_objectid, 2014 u64 owner, u64 offset, int refs_to_add) 2015 { 2016 int ret; 2017 if (owner < BTRFS_FIRST_FREE_OBJECTID) { 2018 BUG_ON(refs_to_add != 1); 2019 ret = insert_tree_block_ref(trans, fs_info, path, bytenr, 2020 parent, root_objectid); 2021 } else { 2022 ret = insert_extent_data_ref(trans, fs_info, path, bytenr, 2023 parent, root_objectid, 2024 owner, offset, refs_to_add); 2025 } 2026 return ret; 2027 } 2028 2029 static int remove_extent_backref(struct btrfs_trans_handle *trans, 2030 struct btrfs_fs_info *fs_info, 2031 struct btrfs_path *path, 2032 struct btrfs_extent_inline_ref *iref, 2033 int refs_to_drop, int is_data, int *last_ref) 2034 { 2035 int ret = 0; 2036 2037 BUG_ON(!is_data && refs_to_drop != 1); 2038 if (iref) { 2039 update_inline_extent_backref(fs_info, path, iref, 2040 -refs_to_drop, NULL, last_ref); 2041 } else if (is_data) { 2042 ret = remove_extent_data_ref(trans, fs_info, path, refs_to_drop, 2043 last_ref); 2044 } else { 2045 *last_ref = 1; 2046 ret = btrfs_del_item(trans, fs_info->extent_root, path); 2047 } 2048 return ret; 2049 } 2050 2051 #define in_range(b, first, len) ((b) >= (first) && (b) < (first) + (len)) 2052 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len, 2053 u64 *discarded_bytes) 2054 { 2055 int j, ret = 0; 2056 u64 bytes_left, end; 2057 u64 aligned_start = ALIGN(start, 1 << 9); 2058 2059 if (WARN_ON(start != aligned_start)) { 2060 len -= aligned_start - start; 2061 len = round_down(len, 1 << 9); 2062 start = aligned_start; 2063 } 2064 2065 *discarded_bytes = 0; 2066 2067 if (!len) 2068 return 0; 2069 2070 end = start + len; 2071 bytes_left = len; 2072 2073 /* Skip any superblocks on this device. */ 2074 for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) { 2075 u64 sb_start = btrfs_sb_offset(j); 2076 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE; 2077 u64 size = sb_start - start; 2078 2079 if (!in_range(sb_start, start, bytes_left) && 2080 !in_range(sb_end, start, bytes_left) && 2081 !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE)) 2082 continue; 2083 2084 /* 2085 * Superblock spans beginning of range. Adjust start and 2086 * try again. 2087 */ 2088 if (sb_start <= start) { 2089 start += sb_end - start; 2090 if (start > end) { 2091 bytes_left = 0; 2092 break; 2093 } 2094 bytes_left = end - start; 2095 continue; 2096 } 2097 2098 if (size) { 2099 ret = blkdev_issue_discard(bdev, start >> 9, size >> 9, 2100 GFP_NOFS, 0); 2101 if (!ret) 2102 *discarded_bytes += size; 2103 else if (ret != -EOPNOTSUPP) 2104 return ret; 2105 } 2106 2107 start = sb_end; 2108 if (start > end) { 2109 bytes_left = 0; 2110 break; 2111 } 2112 bytes_left = end - start; 2113 } 2114 2115 if (bytes_left) { 2116 ret = blkdev_issue_discard(bdev, start >> 9, bytes_left >> 9, 2117 GFP_NOFS, 0); 2118 if (!ret) 2119 *discarded_bytes += bytes_left; 2120 } 2121 return ret; 2122 } 2123 2124 int btrfs_discard_extent(struct btrfs_fs_info *fs_info, u64 bytenr, 2125 u64 num_bytes, u64 *actual_bytes) 2126 { 2127 int ret; 2128 u64 discarded_bytes = 0; 2129 struct btrfs_bio *bbio = NULL; 2130 2131 2132 /* 2133 * Avoid races with device replace and make sure our bbio has devices 2134 * associated to its stripes that don't go away while we are discarding. 2135 */ 2136 btrfs_bio_counter_inc_blocked(fs_info); 2137 /* Tell the block device(s) that the sectors can be discarded */ 2138 ret = btrfs_map_block(fs_info, BTRFS_MAP_DISCARD, bytenr, &num_bytes, 2139 &bbio, 0); 2140 /* Error condition is -ENOMEM */ 2141 if (!ret) { 2142 struct btrfs_bio_stripe *stripe = bbio->stripes; 2143 int i; 2144 2145 2146 for (i = 0; i < bbio->num_stripes; i++, stripe++) { 2147 u64 bytes; 2148 struct request_queue *req_q; 2149 2150 req_q = bdev_get_queue(stripe->dev->bdev); 2151 if (!blk_queue_discard(req_q)) 2152 continue; 2153 2154 ret = btrfs_issue_discard(stripe->dev->bdev, 2155 stripe->physical, 2156 stripe->length, 2157 &bytes); 2158 if (!ret) 2159 discarded_bytes += bytes; 2160 else if (ret != -EOPNOTSUPP) 2161 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */ 2162 2163 /* 2164 * Just in case we get back EOPNOTSUPP for some reason, 2165 * just ignore the return value so we don't screw up 2166 * people calling discard_extent. 2167 */ 2168 ret = 0; 2169 } 2170 btrfs_put_bbio(bbio); 2171 } 2172 btrfs_bio_counter_dec(fs_info); 2173 2174 if (actual_bytes) 2175 *actual_bytes = discarded_bytes; 2176 2177 2178 if (ret == -EOPNOTSUPP) 2179 ret = 0; 2180 return ret; 2181 } 2182 2183 /* Can return -ENOMEM */ 2184 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans, 2185 struct btrfs_root *root, 2186 u64 bytenr, u64 num_bytes, u64 parent, 2187 u64 root_objectid, u64 owner, u64 offset) 2188 { 2189 struct btrfs_fs_info *fs_info = root->fs_info; 2190 int old_ref_mod, new_ref_mod; 2191 int ret; 2192 2193 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID && 2194 root_objectid == BTRFS_TREE_LOG_OBJECTID); 2195 2196 btrfs_ref_tree_mod(root, bytenr, num_bytes, parent, root_objectid, 2197 owner, offset, BTRFS_ADD_DELAYED_REF); 2198 2199 if (owner < BTRFS_FIRST_FREE_OBJECTID) { 2200 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr, 2201 num_bytes, parent, 2202 root_objectid, (int)owner, 2203 BTRFS_ADD_DELAYED_REF, NULL, 2204 &old_ref_mod, &new_ref_mod); 2205 } else { 2206 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr, 2207 num_bytes, parent, 2208 root_objectid, owner, offset, 2209 0, BTRFS_ADD_DELAYED_REF, 2210 &old_ref_mod, &new_ref_mod); 2211 } 2212 2213 if (ret == 0 && old_ref_mod < 0 && new_ref_mod >= 0) 2214 add_pinned_bytes(fs_info, -num_bytes, owner, root_objectid); 2215 2216 return ret; 2217 } 2218 2219 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans, 2220 struct btrfs_fs_info *fs_info, 2221 struct btrfs_delayed_ref_node *node, 2222 u64 parent, u64 root_objectid, 2223 u64 owner, u64 offset, int refs_to_add, 2224 struct btrfs_delayed_extent_op *extent_op) 2225 { 2226 struct btrfs_path *path; 2227 struct extent_buffer *leaf; 2228 struct btrfs_extent_item *item; 2229 struct btrfs_key key; 2230 u64 bytenr = node->bytenr; 2231 u64 num_bytes = node->num_bytes; 2232 u64 refs; 2233 int ret; 2234 2235 path = btrfs_alloc_path(); 2236 if (!path) 2237 return -ENOMEM; 2238 2239 path->reada = READA_FORWARD; 2240 path->leave_spinning = 1; 2241 /* this will setup the path even if it fails to insert the back ref */ 2242 ret = insert_inline_extent_backref(trans, fs_info, path, bytenr, 2243 num_bytes, parent, root_objectid, 2244 owner, offset, 2245 refs_to_add, extent_op); 2246 if ((ret < 0 && ret != -EAGAIN) || !ret) 2247 goto out; 2248 2249 /* 2250 * Ok we had -EAGAIN which means we didn't have space to insert and 2251 * inline extent ref, so just update the reference count and add a 2252 * normal backref. 2253 */ 2254 leaf = path->nodes[0]; 2255 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 2256 item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item); 2257 refs = btrfs_extent_refs(leaf, item); 2258 btrfs_set_extent_refs(leaf, item, refs + refs_to_add); 2259 if (extent_op) 2260 __run_delayed_extent_op(extent_op, leaf, item); 2261 2262 btrfs_mark_buffer_dirty(leaf); 2263 btrfs_release_path(path); 2264 2265 path->reada = READA_FORWARD; 2266 path->leave_spinning = 1; 2267 /* now insert the actual backref */ 2268 ret = insert_extent_backref(trans, fs_info, path, bytenr, parent, 2269 root_objectid, owner, offset, refs_to_add); 2270 if (ret) 2271 btrfs_abort_transaction(trans, ret); 2272 out: 2273 btrfs_free_path(path); 2274 return ret; 2275 } 2276 2277 static int run_delayed_data_ref(struct btrfs_trans_handle *trans, 2278 struct btrfs_fs_info *fs_info, 2279 struct btrfs_delayed_ref_node *node, 2280 struct btrfs_delayed_extent_op *extent_op, 2281 int insert_reserved) 2282 { 2283 int ret = 0; 2284 struct btrfs_delayed_data_ref *ref; 2285 struct btrfs_key ins; 2286 u64 parent = 0; 2287 u64 ref_root = 0; 2288 u64 flags = 0; 2289 2290 ins.objectid = node->bytenr; 2291 ins.offset = node->num_bytes; 2292 ins.type = BTRFS_EXTENT_ITEM_KEY; 2293 2294 ref = btrfs_delayed_node_to_data_ref(node); 2295 trace_run_delayed_data_ref(fs_info, node, ref, node->action); 2296 2297 if (node->type == BTRFS_SHARED_DATA_REF_KEY) 2298 parent = ref->parent; 2299 ref_root = ref->root; 2300 2301 if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) { 2302 if (extent_op) 2303 flags |= extent_op->flags_to_set; 2304 ret = alloc_reserved_file_extent(trans, fs_info, 2305 parent, ref_root, flags, 2306 ref->objectid, ref->offset, 2307 &ins, node->ref_mod); 2308 } else if (node->action == BTRFS_ADD_DELAYED_REF) { 2309 ret = __btrfs_inc_extent_ref(trans, fs_info, node, parent, 2310 ref_root, ref->objectid, 2311 ref->offset, node->ref_mod, 2312 extent_op); 2313 } else if (node->action == BTRFS_DROP_DELAYED_REF) { 2314 ret = __btrfs_free_extent(trans, fs_info, node, parent, 2315 ref_root, ref->objectid, 2316 ref->offset, node->ref_mod, 2317 extent_op); 2318 } else { 2319 BUG(); 2320 } 2321 return ret; 2322 } 2323 2324 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op, 2325 struct extent_buffer *leaf, 2326 struct btrfs_extent_item *ei) 2327 { 2328 u64 flags = btrfs_extent_flags(leaf, ei); 2329 if (extent_op->update_flags) { 2330 flags |= extent_op->flags_to_set; 2331 btrfs_set_extent_flags(leaf, ei, flags); 2332 } 2333 2334 if (extent_op->update_key) { 2335 struct btrfs_tree_block_info *bi; 2336 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)); 2337 bi = (struct btrfs_tree_block_info *)(ei + 1); 2338 btrfs_set_tree_block_key(leaf, bi, &extent_op->key); 2339 } 2340 } 2341 2342 static int run_delayed_extent_op(struct btrfs_trans_handle *trans, 2343 struct btrfs_fs_info *fs_info, 2344 struct btrfs_delayed_ref_head *head, 2345 struct btrfs_delayed_extent_op *extent_op) 2346 { 2347 struct btrfs_key key; 2348 struct btrfs_path *path; 2349 struct btrfs_extent_item *ei; 2350 struct extent_buffer *leaf; 2351 u32 item_size; 2352 int ret; 2353 int err = 0; 2354 int metadata = !extent_op->is_data; 2355 2356 if (trans->aborted) 2357 return 0; 2358 2359 if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA)) 2360 metadata = 0; 2361 2362 path = btrfs_alloc_path(); 2363 if (!path) 2364 return -ENOMEM; 2365 2366 key.objectid = head->bytenr; 2367 2368 if (metadata) { 2369 key.type = BTRFS_METADATA_ITEM_KEY; 2370 key.offset = extent_op->level; 2371 } else { 2372 key.type = BTRFS_EXTENT_ITEM_KEY; 2373 key.offset = head->num_bytes; 2374 } 2375 2376 again: 2377 path->reada = READA_FORWARD; 2378 path->leave_spinning = 1; 2379 ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 1); 2380 if (ret < 0) { 2381 err = ret; 2382 goto out; 2383 } 2384 if (ret > 0) { 2385 if (metadata) { 2386 if (path->slots[0] > 0) { 2387 path->slots[0]--; 2388 btrfs_item_key_to_cpu(path->nodes[0], &key, 2389 path->slots[0]); 2390 if (key.objectid == head->bytenr && 2391 key.type == BTRFS_EXTENT_ITEM_KEY && 2392 key.offset == head->num_bytes) 2393 ret = 0; 2394 } 2395 if (ret > 0) { 2396 btrfs_release_path(path); 2397 metadata = 0; 2398 2399 key.objectid = head->bytenr; 2400 key.offset = head->num_bytes; 2401 key.type = BTRFS_EXTENT_ITEM_KEY; 2402 goto again; 2403 } 2404 } else { 2405 err = -EIO; 2406 goto out; 2407 } 2408 } 2409 2410 leaf = path->nodes[0]; 2411 item_size = btrfs_item_size_nr(leaf, path->slots[0]); 2412 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 2413 if (item_size < sizeof(*ei)) { 2414 ret = convert_extent_item_v0(trans, fs_info, path, (u64)-1, 0); 2415 if (ret < 0) { 2416 err = ret; 2417 goto out; 2418 } 2419 leaf = path->nodes[0]; 2420 item_size = btrfs_item_size_nr(leaf, path->slots[0]); 2421 } 2422 #endif 2423 BUG_ON(item_size < sizeof(*ei)); 2424 ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item); 2425 __run_delayed_extent_op(extent_op, leaf, ei); 2426 2427 btrfs_mark_buffer_dirty(leaf); 2428 out: 2429 btrfs_free_path(path); 2430 return err; 2431 } 2432 2433 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans, 2434 struct btrfs_fs_info *fs_info, 2435 struct btrfs_delayed_ref_node *node, 2436 struct btrfs_delayed_extent_op *extent_op, 2437 int insert_reserved) 2438 { 2439 int ret = 0; 2440 struct btrfs_delayed_tree_ref *ref; 2441 struct btrfs_key ins; 2442 u64 parent = 0; 2443 u64 ref_root = 0; 2444 bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA); 2445 2446 ref = btrfs_delayed_node_to_tree_ref(node); 2447 trace_run_delayed_tree_ref(fs_info, node, ref, node->action); 2448 2449 if (node->type == BTRFS_SHARED_BLOCK_REF_KEY) 2450 parent = ref->parent; 2451 ref_root = ref->root; 2452 2453 ins.objectid = node->bytenr; 2454 if (skinny_metadata) { 2455 ins.offset = ref->level; 2456 ins.type = BTRFS_METADATA_ITEM_KEY; 2457 } else { 2458 ins.offset = node->num_bytes; 2459 ins.type = BTRFS_EXTENT_ITEM_KEY; 2460 } 2461 2462 if (node->ref_mod != 1) { 2463 btrfs_err(fs_info, 2464 "btree block(%llu) has %d references rather than 1: action %d ref_root %llu parent %llu", 2465 node->bytenr, node->ref_mod, node->action, ref_root, 2466 parent); 2467 return -EIO; 2468 } 2469 if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) { 2470 BUG_ON(!extent_op || !extent_op->update_flags); 2471 ret = alloc_reserved_tree_block(trans, fs_info, 2472 parent, ref_root, 2473 extent_op->flags_to_set, 2474 &extent_op->key, 2475 ref->level, &ins); 2476 } else if (node->action == BTRFS_ADD_DELAYED_REF) { 2477 ret = __btrfs_inc_extent_ref(trans, fs_info, node, 2478 parent, ref_root, 2479 ref->level, 0, 1, 2480 extent_op); 2481 } else if (node->action == BTRFS_DROP_DELAYED_REF) { 2482 ret = __btrfs_free_extent(trans, fs_info, node, 2483 parent, ref_root, 2484 ref->level, 0, 1, extent_op); 2485 } else { 2486 BUG(); 2487 } 2488 return ret; 2489 } 2490 2491 /* helper function to actually process a single delayed ref entry */ 2492 static int run_one_delayed_ref(struct btrfs_trans_handle *trans, 2493 struct btrfs_fs_info *fs_info, 2494 struct btrfs_delayed_ref_node *node, 2495 struct btrfs_delayed_extent_op *extent_op, 2496 int insert_reserved) 2497 { 2498 int ret = 0; 2499 2500 if (trans->aborted) { 2501 if (insert_reserved) 2502 btrfs_pin_extent(fs_info, node->bytenr, 2503 node->num_bytes, 1); 2504 return 0; 2505 } 2506 2507 if (node->type == BTRFS_TREE_BLOCK_REF_KEY || 2508 node->type == BTRFS_SHARED_BLOCK_REF_KEY) 2509 ret = run_delayed_tree_ref(trans, fs_info, node, extent_op, 2510 insert_reserved); 2511 else if (node->type == BTRFS_EXTENT_DATA_REF_KEY || 2512 node->type == BTRFS_SHARED_DATA_REF_KEY) 2513 ret = run_delayed_data_ref(trans, fs_info, node, extent_op, 2514 insert_reserved); 2515 else 2516 BUG(); 2517 return ret; 2518 } 2519 2520 static inline struct btrfs_delayed_ref_node * 2521 select_delayed_ref(struct btrfs_delayed_ref_head *head) 2522 { 2523 struct btrfs_delayed_ref_node *ref; 2524 2525 if (RB_EMPTY_ROOT(&head->ref_tree)) 2526 return NULL; 2527 2528 /* 2529 * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first. 2530 * This is to prevent a ref count from going down to zero, which deletes 2531 * the extent item from the extent tree, when there still are references 2532 * to add, which would fail because they would not find the extent item. 2533 */ 2534 if (!list_empty(&head->ref_add_list)) 2535 return list_first_entry(&head->ref_add_list, 2536 struct btrfs_delayed_ref_node, add_list); 2537 2538 ref = rb_entry(rb_first(&head->ref_tree), 2539 struct btrfs_delayed_ref_node, ref_node); 2540 ASSERT(list_empty(&ref->add_list)); 2541 return ref; 2542 } 2543 2544 static void unselect_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs, 2545 struct btrfs_delayed_ref_head *head) 2546 { 2547 spin_lock(&delayed_refs->lock); 2548 head->processing = 0; 2549 delayed_refs->num_heads_ready++; 2550 spin_unlock(&delayed_refs->lock); 2551 btrfs_delayed_ref_unlock(head); 2552 } 2553 2554 static int cleanup_extent_op(struct btrfs_trans_handle *trans, 2555 struct btrfs_fs_info *fs_info, 2556 struct btrfs_delayed_ref_head *head) 2557 { 2558 struct btrfs_delayed_extent_op *extent_op = head->extent_op; 2559 int ret; 2560 2561 if (!extent_op) 2562 return 0; 2563 head->extent_op = NULL; 2564 if (head->must_insert_reserved) { 2565 btrfs_free_delayed_extent_op(extent_op); 2566 return 0; 2567 } 2568 spin_unlock(&head->lock); 2569 ret = run_delayed_extent_op(trans, fs_info, head, extent_op); 2570 btrfs_free_delayed_extent_op(extent_op); 2571 return ret ? ret : 1; 2572 } 2573 2574 static int cleanup_ref_head(struct btrfs_trans_handle *trans, 2575 struct btrfs_fs_info *fs_info, 2576 struct btrfs_delayed_ref_head *head) 2577 { 2578 struct btrfs_delayed_ref_root *delayed_refs; 2579 int ret; 2580 2581 delayed_refs = &trans->transaction->delayed_refs; 2582 2583 ret = cleanup_extent_op(trans, fs_info, head); 2584 if (ret < 0) { 2585 unselect_delayed_ref_head(delayed_refs, head); 2586 btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret); 2587 return ret; 2588 } else if (ret) { 2589 return ret; 2590 } 2591 2592 /* 2593 * Need to drop our head ref lock and re-acquire the delayed ref lock 2594 * and then re-check to make sure nobody got added. 2595 */ 2596 spin_unlock(&head->lock); 2597 spin_lock(&delayed_refs->lock); 2598 spin_lock(&head->lock); 2599 if (!RB_EMPTY_ROOT(&head->ref_tree) || head->extent_op) { 2600 spin_unlock(&head->lock); 2601 spin_unlock(&delayed_refs->lock); 2602 return 1; 2603 } 2604 delayed_refs->num_heads--; 2605 rb_erase(&head->href_node, &delayed_refs->href_root); 2606 RB_CLEAR_NODE(&head->href_node); 2607 spin_unlock(&delayed_refs->lock); 2608 spin_unlock(&head->lock); 2609 atomic_dec(&delayed_refs->num_entries); 2610 2611 trace_run_delayed_ref_head(fs_info, head, 0); 2612 2613 if (head->total_ref_mod < 0) { 2614 struct btrfs_block_group_cache *cache; 2615 2616 cache = btrfs_lookup_block_group(fs_info, head->bytenr); 2617 ASSERT(cache); 2618 percpu_counter_add(&cache->space_info->total_bytes_pinned, 2619 -head->num_bytes); 2620 btrfs_put_block_group(cache); 2621 2622 if (head->is_data) { 2623 spin_lock(&delayed_refs->lock); 2624 delayed_refs->pending_csums -= head->num_bytes; 2625 spin_unlock(&delayed_refs->lock); 2626 } 2627 } 2628 2629 if (head->must_insert_reserved) { 2630 btrfs_pin_extent(fs_info, head->bytenr, 2631 head->num_bytes, 1); 2632 if (head->is_data) { 2633 ret = btrfs_del_csums(trans, fs_info, head->bytenr, 2634 head->num_bytes); 2635 } 2636 } 2637 2638 /* Also free its reserved qgroup space */ 2639 btrfs_qgroup_free_delayed_ref(fs_info, head->qgroup_ref_root, 2640 head->qgroup_reserved); 2641 btrfs_delayed_ref_unlock(head); 2642 btrfs_put_delayed_ref_head(head); 2643 return 0; 2644 } 2645 2646 /* 2647 * Returns 0 on success or if called with an already aborted transaction. 2648 * Returns -ENOMEM or -EIO on failure and will abort the transaction. 2649 */ 2650 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans, 2651 struct btrfs_fs_info *fs_info, 2652 unsigned long nr) 2653 { 2654 struct btrfs_delayed_ref_root *delayed_refs; 2655 struct btrfs_delayed_ref_node *ref; 2656 struct btrfs_delayed_ref_head *locked_ref = NULL; 2657 struct btrfs_delayed_extent_op *extent_op; 2658 ktime_t start = ktime_get(); 2659 int ret; 2660 unsigned long count = 0; 2661 unsigned long actual_count = 0; 2662 int must_insert_reserved = 0; 2663 2664 delayed_refs = &trans->transaction->delayed_refs; 2665 while (1) { 2666 if (!locked_ref) { 2667 if (count >= nr) 2668 break; 2669 2670 spin_lock(&delayed_refs->lock); 2671 locked_ref = btrfs_select_ref_head(trans); 2672 if (!locked_ref) { 2673 spin_unlock(&delayed_refs->lock); 2674 break; 2675 } 2676 2677 /* grab the lock that says we are going to process 2678 * all the refs for this head */ 2679 ret = btrfs_delayed_ref_lock(trans, locked_ref); 2680 spin_unlock(&delayed_refs->lock); 2681 /* 2682 * we may have dropped the spin lock to get the head 2683 * mutex lock, and that might have given someone else 2684 * time to free the head. If that's true, it has been 2685 * removed from our list and we can move on. 2686 */ 2687 if (ret == -EAGAIN) { 2688 locked_ref = NULL; 2689 count++; 2690 continue; 2691 } 2692 } 2693 2694 /* 2695 * We need to try and merge add/drops of the same ref since we 2696 * can run into issues with relocate dropping the implicit ref 2697 * and then it being added back again before the drop can 2698 * finish. If we merged anything we need to re-loop so we can 2699 * get a good ref. 2700 * Or we can get node references of the same type that weren't 2701 * merged when created due to bumps in the tree mod seq, and 2702 * we need to merge them to prevent adding an inline extent 2703 * backref before dropping it (triggering a BUG_ON at 2704 * insert_inline_extent_backref()). 2705 */ 2706 spin_lock(&locked_ref->lock); 2707 btrfs_merge_delayed_refs(trans, fs_info, delayed_refs, 2708 locked_ref); 2709 2710 /* 2711 * locked_ref is the head node, so we have to go one 2712 * node back for any delayed ref updates 2713 */ 2714 ref = select_delayed_ref(locked_ref); 2715 2716 if (ref && ref->seq && 2717 btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) { 2718 spin_unlock(&locked_ref->lock); 2719 unselect_delayed_ref_head(delayed_refs, locked_ref); 2720 locked_ref = NULL; 2721 cond_resched(); 2722 count++; 2723 continue; 2724 } 2725 2726 /* 2727 * We're done processing refs in this ref_head, clean everything 2728 * up and move on to the next ref_head. 2729 */ 2730 if (!ref) { 2731 ret = cleanup_ref_head(trans, fs_info, locked_ref); 2732 if (ret > 0 ) { 2733 /* We dropped our lock, we need to loop. */ 2734 ret = 0; 2735 continue; 2736 } else if (ret) { 2737 return ret; 2738 } 2739 locked_ref = NULL; 2740 count++; 2741 continue; 2742 } 2743 2744 actual_count++; 2745 ref->in_tree = 0; 2746 rb_erase(&ref->ref_node, &locked_ref->ref_tree); 2747 RB_CLEAR_NODE(&ref->ref_node); 2748 if (!list_empty(&ref->add_list)) 2749 list_del(&ref->add_list); 2750 /* 2751 * When we play the delayed ref, also correct the ref_mod on 2752 * head 2753 */ 2754 switch (ref->action) { 2755 case BTRFS_ADD_DELAYED_REF: 2756 case BTRFS_ADD_DELAYED_EXTENT: 2757 locked_ref->ref_mod -= ref->ref_mod; 2758 break; 2759 case BTRFS_DROP_DELAYED_REF: 2760 locked_ref->ref_mod += ref->ref_mod; 2761 break; 2762 default: 2763 WARN_ON(1); 2764 } 2765 atomic_dec(&delayed_refs->num_entries); 2766 2767 /* 2768 * Record the must-insert_reserved flag before we drop the spin 2769 * lock. 2770 */ 2771 must_insert_reserved = locked_ref->must_insert_reserved; 2772 locked_ref->must_insert_reserved = 0; 2773 2774 extent_op = locked_ref->extent_op; 2775 locked_ref->extent_op = NULL; 2776 spin_unlock(&locked_ref->lock); 2777 2778 ret = run_one_delayed_ref(trans, fs_info, ref, extent_op, 2779 must_insert_reserved); 2780 2781 btrfs_free_delayed_extent_op(extent_op); 2782 if (ret) { 2783 unselect_delayed_ref_head(delayed_refs, locked_ref); 2784 btrfs_put_delayed_ref(ref); 2785 btrfs_debug(fs_info, "run_one_delayed_ref returned %d", 2786 ret); 2787 return ret; 2788 } 2789 2790 btrfs_put_delayed_ref(ref); 2791 count++; 2792 cond_resched(); 2793 } 2794 2795 /* 2796 * We don't want to include ref heads since we can have empty ref heads 2797 * and those will drastically skew our runtime down since we just do 2798 * accounting, no actual extent tree updates. 2799 */ 2800 if (actual_count > 0) { 2801 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start)); 2802 u64 avg; 2803 2804 /* 2805 * We weigh the current average higher than our current runtime 2806 * to avoid large swings in the average. 2807 */ 2808 spin_lock(&delayed_refs->lock); 2809 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime; 2810 fs_info->avg_delayed_ref_runtime = avg >> 2; /* div by 4 */ 2811 spin_unlock(&delayed_refs->lock); 2812 } 2813 return 0; 2814 } 2815 2816 #ifdef SCRAMBLE_DELAYED_REFS 2817 /* 2818 * Normally delayed refs get processed in ascending bytenr order. This 2819 * correlates in most cases to the order added. To expose dependencies on this 2820 * order, we start to process the tree in the middle instead of the beginning 2821 */ 2822 static u64 find_middle(struct rb_root *root) 2823 { 2824 struct rb_node *n = root->rb_node; 2825 struct btrfs_delayed_ref_node *entry; 2826 int alt = 1; 2827 u64 middle; 2828 u64 first = 0, last = 0; 2829 2830 n = rb_first(root); 2831 if (n) { 2832 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node); 2833 first = entry->bytenr; 2834 } 2835 n = rb_last(root); 2836 if (n) { 2837 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node); 2838 last = entry->bytenr; 2839 } 2840 n = root->rb_node; 2841 2842 while (n) { 2843 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node); 2844 WARN_ON(!entry->in_tree); 2845 2846 middle = entry->bytenr; 2847 2848 if (alt) 2849 n = n->rb_left; 2850 else 2851 n = n->rb_right; 2852 2853 alt = 1 - alt; 2854 } 2855 return middle; 2856 } 2857 #endif 2858 2859 static inline u64 heads_to_leaves(struct btrfs_fs_info *fs_info, u64 heads) 2860 { 2861 u64 num_bytes; 2862 2863 num_bytes = heads * (sizeof(struct btrfs_extent_item) + 2864 sizeof(struct btrfs_extent_inline_ref)); 2865 if (!btrfs_fs_incompat(fs_info, SKINNY_METADATA)) 2866 num_bytes += heads * sizeof(struct btrfs_tree_block_info); 2867 2868 /* 2869 * We don't ever fill up leaves all the way so multiply by 2 just to be 2870 * closer to what we're really going to want to use. 2871 */ 2872 return div_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(fs_info)); 2873 } 2874 2875 /* 2876 * Takes the number of bytes to be csumm'ed and figures out how many leaves it 2877 * would require to store the csums for that many bytes. 2878 */ 2879 u64 btrfs_csum_bytes_to_leaves(struct btrfs_fs_info *fs_info, u64 csum_bytes) 2880 { 2881 u64 csum_size; 2882 u64 num_csums_per_leaf; 2883 u64 num_csums; 2884 2885 csum_size = BTRFS_MAX_ITEM_SIZE(fs_info); 2886 num_csums_per_leaf = div64_u64(csum_size, 2887 (u64)btrfs_super_csum_size(fs_info->super_copy)); 2888 num_csums = div64_u64(csum_bytes, fs_info->sectorsize); 2889 num_csums += num_csums_per_leaf - 1; 2890 num_csums = div64_u64(num_csums, num_csums_per_leaf); 2891 return num_csums; 2892 } 2893 2894 int btrfs_check_space_for_delayed_refs(struct btrfs_trans_handle *trans, 2895 struct btrfs_fs_info *fs_info) 2896 { 2897 struct btrfs_block_rsv *global_rsv; 2898 u64 num_heads = trans->transaction->delayed_refs.num_heads_ready; 2899 u64 csum_bytes = trans->transaction->delayed_refs.pending_csums; 2900 unsigned int num_dirty_bgs = trans->transaction->num_dirty_bgs; 2901 u64 num_bytes, num_dirty_bgs_bytes; 2902 int ret = 0; 2903 2904 num_bytes = btrfs_calc_trans_metadata_size(fs_info, 1); 2905 num_heads = heads_to_leaves(fs_info, num_heads); 2906 if (num_heads > 1) 2907 num_bytes += (num_heads - 1) * fs_info->nodesize; 2908 num_bytes <<= 1; 2909 num_bytes += btrfs_csum_bytes_to_leaves(fs_info, csum_bytes) * 2910 fs_info->nodesize; 2911 num_dirty_bgs_bytes = btrfs_calc_trans_metadata_size(fs_info, 2912 num_dirty_bgs); 2913 global_rsv = &fs_info->global_block_rsv; 2914 2915 /* 2916 * If we can't allocate any more chunks lets make sure we have _lots_ of 2917 * wiggle room since running delayed refs can create more delayed refs. 2918 */ 2919 if (global_rsv->space_info->full) { 2920 num_dirty_bgs_bytes <<= 1; 2921 num_bytes <<= 1; 2922 } 2923 2924 spin_lock(&global_rsv->lock); 2925 if (global_rsv->reserved <= num_bytes + num_dirty_bgs_bytes) 2926 ret = 1; 2927 spin_unlock(&global_rsv->lock); 2928 return ret; 2929 } 2930 2931 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans, 2932 struct btrfs_fs_info *fs_info) 2933 { 2934 u64 num_entries = 2935 atomic_read(&trans->transaction->delayed_refs.num_entries); 2936 u64 avg_runtime; 2937 u64 val; 2938 2939 smp_mb(); 2940 avg_runtime = fs_info->avg_delayed_ref_runtime; 2941 val = num_entries * avg_runtime; 2942 if (val >= NSEC_PER_SEC) 2943 return 1; 2944 if (val >= NSEC_PER_SEC / 2) 2945 return 2; 2946 2947 return btrfs_check_space_for_delayed_refs(trans, fs_info); 2948 } 2949 2950 struct async_delayed_refs { 2951 struct btrfs_root *root; 2952 u64 transid; 2953 int count; 2954 int error; 2955 int sync; 2956 struct completion wait; 2957 struct btrfs_work work; 2958 }; 2959 2960 static inline struct async_delayed_refs * 2961 to_async_delayed_refs(struct btrfs_work *work) 2962 { 2963 return container_of(work, struct async_delayed_refs, work); 2964 } 2965 2966 static void delayed_ref_async_start(struct btrfs_work *work) 2967 { 2968 struct async_delayed_refs *async = to_async_delayed_refs(work); 2969 struct btrfs_trans_handle *trans; 2970 struct btrfs_fs_info *fs_info = async->root->fs_info; 2971 int ret; 2972 2973 /* if the commit is already started, we don't need to wait here */ 2974 if (btrfs_transaction_blocked(fs_info)) 2975 goto done; 2976 2977 trans = btrfs_join_transaction(async->root); 2978 if (IS_ERR(trans)) { 2979 async->error = PTR_ERR(trans); 2980 goto done; 2981 } 2982 2983 /* 2984 * trans->sync means that when we call end_transaction, we won't 2985 * wait on delayed refs 2986 */ 2987 trans->sync = true; 2988 2989 /* Don't bother flushing if we got into a different transaction */ 2990 if (trans->transid > async->transid) 2991 goto end; 2992 2993 ret = btrfs_run_delayed_refs(trans, fs_info, async->count); 2994 if (ret) 2995 async->error = ret; 2996 end: 2997 ret = btrfs_end_transaction(trans); 2998 if (ret && !async->error) 2999 async->error = ret; 3000 done: 3001 if (async->sync) 3002 complete(&async->wait); 3003 else 3004 kfree(async); 3005 } 3006 3007 int btrfs_async_run_delayed_refs(struct btrfs_fs_info *fs_info, 3008 unsigned long count, u64 transid, int wait) 3009 { 3010 struct async_delayed_refs *async; 3011 int ret; 3012 3013 async = kmalloc(sizeof(*async), GFP_NOFS); 3014 if (!async) 3015 return -ENOMEM; 3016 3017 async->root = fs_info->tree_root; 3018 async->count = count; 3019 async->error = 0; 3020 async->transid = transid; 3021 if (wait) 3022 async->sync = 1; 3023 else 3024 async->sync = 0; 3025 init_completion(&async->wait); 3026 3027 btrfs_init_work(&async->work, btrfs_extent_refs_helper, 3028 delayed_ref_async_start, NULL, NULL); 3029 3030 btrfs_queue_work(fs_info->extent_workers, &async->work); 3031 3032 if (wait) { 3033 wait_for_completion(&async->wait); 3034 ret = async->error; 3035 kfree(async); 3036 return ret; 3037 } 3038 return 0; 3039 } 3040 3041 /* 3042 * this starts processing the delayed reference count updates and 3043 * extent insertions we have queued up so far. count can be 3044 * 0, which means to process everything in the tree at the start 3045 * of the run (but not newly added entries), or it can be some target 3046 * number you'd like to process. 3047 * 3048 * Returns 0 on success or if called with an aborted transaction 3049 * Returns <0 on error and aborts the transaction 3050 */ 3051 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans, 3052 struct btrfs_fs_info *fs_info, unsigned long count) 3053 { 3054 struct rb_node *node; 3055 struct btrfs_delayed_ref_root *delayed_refs; 3056 struct btrfs_delayed_ref_head *head; 3057 int ret; 3058 int run_all = count == (unsigned long)-1; 3059 bool can_flush_pending_bgs = trans->can_flush_pending_bgs; 3060 3061 /* We'll clean this up in btrfs_cleanup_transaction */ 3062 if (trans->aborted) 3063 return 0; 3064 3065 if (test_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags)) 3066 return 0; 3067 3068 delayed_refs = &trans->transaction->delayed_refs; 3069 if (count == 0) 3070 count = atomic_read(&delayed_refs->num_entries) * 2; 3071 3072 again: 3073 #ifdef SCRAMBLE_DELAYED_REFS 3074 delayed_refs->run_delayed_start = find_middle(&delayed_refs->root); 3075 #endif 3076 trans->can_flush_pending_bgs = false; 3077 ret = __btrfs_run_delayed_refs(trans, fs_info, count); 3078 if (ret < 0) { 3079 btrfs_abort_transaction(trans, ret); 3080 return ret; 3081 } 3082 3083 if (run_all) { 3084 if (!list_empty(&trans->new_bgs)) 3085 btrfs_create_pending_block_groups(trans, fs_info); 3086 3087 spin_lock(&delayed_refs->lock); 3088 node = rb_first(&delayed_refs->href_root); 3089 if (!node) { 3090 spin_unlock(&delayed_refs->lock); 3091 goto out; 3092 } 3093 head = rb_entry(node, struct btrfs_delayed_ref_head, 3094 href_node); 3095 refcount_inc(&head->refs); 3096 spin_unlock(&delayed_refs->lock); 3097 3098 /* Mutex was contended, block until it's released and retry. */ 3099 mutex_lock(&head->mutex); 3100 mutex_unlock(&head->mutex); 3101 3102 btrfs_put_delayed_ref_head(head); 3103 cond_resched(); 3104 goto again; 3105 } 3106 out: 3107 trans->can_flush_pending_bgs = can_flush_pending_bgs; 3108 return 0; 3109 } 3110 3111 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans, 3112 struct btrfs_fs_info *fs_info, 3113 u64 bytenr, u64 num_bytes, u64 flags, 3114 int level, int is_data) 3115 { 3116 struct btrfs_delayed_extent_op *extent_op; 3117 int ret; 3118 3119 extent_op = btrfs_alloc_delayed_extent_op(); 3120 if (!extent_op) 3121 return -ENOMEM; 3122 3123 extent_op->flags_to_set = flags; 3124 extent_op->update_flags = true; 3125 extent_op->update_key = false; 3126 extent_op->is_data = is_data ? true : false; 3127 extent_op->level = level; 3128 3129 ret = btrfs_add_delayed_extent_op(fs_info, trans, bytenr, 3130 num_bytes, extent_op); 3131 if (ret) 3132 btrfs_free_delayed_extent_op(extent_op); 3133 return ret; 3134 } 3135 3136 static noinline int check_delayed_ref(struct btrfs_root *root, 3137 struct btrfs_path *path, 3138 u64 objectid, u64 offset, u64 bytenr) 3139 { 3140 struct btrfs_delayed_ref_head *head; 3141 struct btrfs_delayed_ref_node *ref; 3142 struct btrfs_delayed_data_ref *data_ref; 3143 struct btrfs_delayed_ref_root *delayed_refs; 3144 struct btrfs_transaction *cur_trans; 3145 struct rb_node *node; 3146 int ret = 0; 3147 3148 cur_trans = root->fs_info->running_transaction; 3149 if (!cur_trans) 3150 return 0; 3151 3152 delayed_refs = &cur_trans->delayed_refs; 3153 spin_lock(&delayed_refs->lock); 3154 head = btrfs_find_delayed_ref_head(delayed_refs, bytenr); 3155 if (!head) { 3156 spin_unlock(&delayed_refs->lock); 3157 return 0; 3158 } 3159 3160 if (!mutex_trylock(&head->mutex)) { 3161 refcount_inc(&head->refs); 3162 spin_unlock(&delayed_refs->lock); 3163 3164 btrfs_release_path(path); 3165 3166 /* 3167 * Mutex was contended, block until it's released and let 3168 * caller try again 3169 */ 3170 mutex_lock(&head->mutex); 3171 mutex_unlock(&head->mutex); 3172 btrfs_put_delayed_ref_head(head); 3173 return -EAGAIN; 3174 } 3175 spin_unlock(&delayed_refs->lock); 3176 3177 spin_lock(&head->lock); 3178 /* 3179 * XXX: We should replace this with a proper search function in the 3180 * future. 3181 */ 3182 for (node = rb_first(&head->ref_tree); node; node = rb_next(node)) { 3183 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node); 3184 /* If it's a shared ref we know a cross reference exists */ 3185 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) { 3186 ret = 1; 3187 break; 3188 } 3189 3190 data_ref = btrfs_delayed_node_to_data_ref(ref); 3191 3192 /* 3193 * If our ref doesn't match the one we're currently looking at 3194 * then we have a cross reference. 3195 */ 3196 if (data_ref->root != root->root_key.objectid || 3197 data_ref->objectid != objectid || 3198 data_ref->offset != offset) { 3199 ret = 1; 3200 break; 3201 } 3202 } 3203 spin_unlock(&head->lock); 3204 mutex_unlock(&head->mutex); 3205 return ret; 3206 } 3207 3208 static noinline int check_committed_ref(struct btrfs_root *root, 3209 struct btrfs_path *path, 3210 u64 objectid, u64 offset, u64 bytenr) 3211 { 3212 struct btrfs_fs_info *fs_info = root->fs_info; 3213 struct btrfs_root *extent_root = fs_info->extent_root; 3214 struct extent_buffer *leaf; 3215 struct btrfs_extent_data_ref *ref; 3216 struct btrfs_extent_inline_ref *iref; 3217 struct btrfs_extent_item *ei; 3218 struct btrfs_key key; 3219 u32 item_size; 3220 int type; 3221 int ret; 3222 3223 key.objectid = bytenr; 3224 key.offset = (u64)-1; 3225 key.type = BTRFS_EXTENT_ITEM_KEY; 3226 3227 ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0); 3228 if (ret < 0) 3229 goto out; 3230 BUG_ON(ret == 0); /* Corruption */ 3231 3232 ret = -ENOENT; 3233 if (path->slots[0] == 0) 3234 goto out; 3235 3236 path->slots[0]--; 3237 leaf = path->nodes[0]; 3238 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 3239 3240 if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY) 3241 goto out; 3242 3243 ret = 1; 3244 item_size = btrfs_item_size_nr(leaf, path->slots[0]); 3245 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 3246 if (item_size < sizeof(*ei)) { 3247 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0)); 3248 goto out; 3249 } 3250 #endif 3251 ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item); 3252 3253 if (item_size != sizeof(*ei) + 3254 btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY)) 3255 goto out; 3256 3257 if (btrfs_extent_generation(leaf, ei) <= 3258 btrfs_root_last_snapshot(&root->root_item)) 3259 goto out; 3260 3261 iref = (struct btrfs_extent_inline_ref *)(ei + 1); 3262 3263 type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA); 3264 if (type != BTRFS_EXTENT_DATA_REF_KEY) 3265 goto out; 3266 3267 ref = (struct btrfs_extent_data_ref *)(&iref->offset); 3268 if (btrfs_extent_refs(leaf, ei) != 3269 btrfs_extent_data_ref_count(leaf, ref) || 3270 btrfs_extent_data_ref_root(leaf, ref) != 3271 root->root_key.objectid || 3272 btrfs_extent_data_ref_objectid(leaf, ref) != objectid || 3273 btrfs_extent_data_ref_offset(leaf, ref) != offset) 3274 goto out; 3275 3276 ret = 0; 3277 out: 3278 return ret; 3279 } 3280 3281 int btrfs_cross_ref_exist(struct btrfs_root *root, u64 objectid, u64 offset, 3282 u64 bytenr) 3283 { 3284 struct btrfs_path *path; 3285 int ret; 3286 int ret2; 3287 3288 path = btrfs_alloc_path(); 3289 if (!path) 3290 return -ENOENT; 3291 3292 do { 3293 ret = check_committed_ref(root, path, objectid, 3294 offset, bytenr); 3295 if (ret && ret != -ENOENT) 3296 goto out; 3297 3298 ret2 = check_delayed_ref(root, path, objectid, 3299 offset, bytenr); 3300 } while (ret2 == -EAGAIN); 3301 3302 if (ret2 && ret2 != -ENOENT) { 3303 ret = ret2; 3304 goto out; 3305 } 3306 3307 if (ret != -ENOENT || ret2 != -ENOENT) 3308 ret = 0; 3309 out: 3310 btrfs_free_path(path); 3311 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID) 3312 WARN_ON(ret > 0); 3313 return ret; 3314 } 3315 3316 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans, 3317 struct btrfs_root *root, 3318 struct extent_buffer *buf, 3319 int full_backref, int inc) 3320 { 3321 struct btrfs_fs_info *fs_info = root->fs_info; 3322 u64 bytenr; 3323 u64 num_bytes; 3324 u64 parent; 3325 u64 ref_root; 3326 u32 nritems; 3327 struct btrfs_key key; 3328 struct btrfs_file_extent_item *fi; 3329 int i; 3330 int level; 3331 int ret = 0; 3332 int (*process_func)(struct btrfs_trans_handle *, 3333 struct btrfs_root *, 3334 u64, u64, u64, u64, u64, u64); 3335 3336 3337 if (btrfs_is_testing(fs_info)) 3338 return 0; 3339 3340 ref_root = btrfs_header_owner(buf); 3341 nritems = btrfs_header_nritems(buf); 3342 level = btrfs_header_level(buf); 3343 3344 if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0) 3345 return 0; 3346 3347 if (inc) 3348 process_func = btrfs_inc_extent_ref; 3349 else 3350 process_func = btrfs_free_extent; 3351 3352 if (full_backref) 3353 parent = buf->start; 3354 else 3355 parent = 0; 3356 3357 for (i = 0; i < nritems; i++) { 3358 if (level == 0) { 3359 btrfs_item_key_to_cpu(buf, &key, i); 3360 if (key.type != BTRFS_EXTENT_DATA_KEY) 3361 continue; 3362 fi = btrfs_item_ptr(buf, i, 3363 struct btrfs_file_extent_item); 3364 if (btrfs_file_extent_type(buf, fi) == 3365 BTRFS_FILE_EXTENT_INLINE) 3366 continue; 3367 bytenr = btrfs_file_extent_disk_bytenr(buf, fi); 3368 if (bytenr == 0) 3369 continue; 3370 3371 num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi); 3372 key.offset -= btrfs_file_extent_offset(buf, fi); 3373 ret = process_func(trans, root, bytenr, num_bytes, 3374 parent, ref_root, key.objectid, 3375 key.offset); 3376 if (ret) 3377 goto fail; 3378 } else { 3379 bytenr = btrfs_node_blockptr(buf, i); 3380 num_bytes = fs_info->nodesize; 3381 ret = process_func(trans, root, bytenr, num_bytes, 3382 parent, ref_root, level - 1, 0); 3383 if (ret) 3384 goto fail; 3385 } 3386 } 3387 return 0; 3388 fail: 3389 return ret; 3390 } 3391 3392 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root, 3393 struct extent_buffer *buf, int full_backref) 3394 { 3395 return __btrfs_mod_ref(trans, root, buf, full_backref, 1); 3396 } 3397 3398 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root, 3399 struct extent_buffer *buf, int full_backref) 3400 { 3401 return __btrfs_mod_ref(trans, root, buf, full_backref, 0); 3402 } 3403 3404 static int write_one_cache_group(struct btrfs_trans_handle *trans, 3405 struct btrfs_fs_info *fs_info, 3406 struct btrfs_path *path, 3407 struct btrfs_block_group_cache *cache) 3408 { 3409 int ret; 3410 struct btrfs_root *extent_root = fs_info->extent_root; 3411 unsigned long bi; 3412 struct extent_buffer *leaf; 3413 3414 ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1); 3415 if (ret) { 3416 if (ret > 0) 3417 ret = -ENOENT; 3418 goto fail; 3419 } 3420 3421 leaf = path->nodes[0]; 3422 bi = btrfs_item_ptr_offset(leaf, path->slots[0]); 3423 write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item)); 3424 btrfs_mark_buffer_dirty(leaf); 3425 fail: 3426 btrfs_release_path(path); 3427 return ret; 3428 3429 } 3430 3431 static struct btrfs_block_group_cache * 3432 next_block_group(struct btrfs_fs_info *fs_info, 3433 struct btrfs_block_group_cache *cache) 3434 { 3435 struct rb_node *node; 3436 3437 spin_lock(&fs_info->block_group_cache_lock); 3438 3439 /* If our block group was removed, we need a full search. */ 3440 if (RB_EMPTY_NODE(&cache->cache_node)) { 3441 const u64 next_bytenr = cache->key.objectid + cache->key.offset; 3442 3443 spin_unlock(&fs_info->block_group_cache_lock); 3444 btrfs_put_block_group(cache); 3445 cache = btrfs_lookup_first_block_group(fs_info, next_bytenr); return cache; 3446 } 3447 node = rb_next(&cache->cache_node); 3448 btrfs_put_block_group(cache); 3449 if (node) { 3450 cache = rb_entry(node, struct btrfs_block_group_cache, 3451 cache_node); 3452 btrfs_get_block_group(cache); 3453 } else 3454 cache = NULL; 3455 spin_unlock(&fs_info->block_group_cache_lock); 3456 return cache; 3457 } 3458 3459 static int cache_save_setup(struct btrfs_block_group_cache *block_group, 3460 struct btrfs_trans_handle *trans, 3461 struct btrfs_path *path) 3462 { 3463 struct btrfs_fs_info *fs_info = block_group->fs_info; 3464 struct btrfs_root *root = fs_info->tree_root; 3465 struct inode *inode = NULL; 3466 struct extent_changeset *data_reserved = NULL; 3467 u64 alloc_hint = 0; 3468 int dcs = BTRFS_DC_ERROR; 3469 u64 num_pages = 0; 3470 int retries = 0; 3471 int ret = 0; 3472 3473 /* 3474 * If this block group is smaller than 100 megs don't bother caching the 3475 * block group. 3476 */ 3477 if (block_group->key.offset < (100 * SZ_1M)) { 3478 spin_lock(&block_group->lock); 3479 block_group->disk_cache_state = BTRFS_DC_WRITTEN; 3480 spin_unlock(&block_group->lock); 3481 return 0; 3482 } 3483 3484 if (trans->aborted) 3485 return 0; 3486 again: 3487 inode = lookup_free_space_inode(fs_info, block_group, path); 3488 if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) { 3489 ret = PTR_ERR(inode); 3490 btrfs_release_path(path); 3491 goto out; 3492 } 3493 3494 if (IS_ERR(inode)) { 3495 BUG_ON(retries); 3496 retries++; 3497 3498 if (block_group->ro) 3499 goto out_free; 3500 3501 ret = create_free_space_inode(fs_info, trans, block_group, 3502 path); 3503 if (ret) 3504 goto out_free; 3505 goto again; 3506 } 3507 3508 /* 3509 * We want to set the generation to 0, that way if anything goes wrong 3510 * from here on out we know not to trust this cache when we load up next 3511 * time. 3512 */ 3513 BTRFS_I(inode)->generation = 0; 3514 ret = btrfs_update_inode(trans, root, inode); 3515 if (ret) { 3516 /* 3517 * So theoretically we could recover from this, simply set the 3518 * super cache generation to 0 so we know to invalidate the 3519 * cache, but then we'd have to keep track of the block groups 3520 * that fail this way so we know we _have_ to reset this cache 3521 * before the next commit or risk reading stale cache. So to 3522 * limit our exposure to horrible edge cases lets just abort the 3523 * transaction, this only happens in really bad situations 3524 * anyway. 3525 */ 3526 btrfs_abort_transaction(trans, ret); 3527 goto out_put; 3528 } 3529 WARN_ON(ret); 3530 3531 /* We've already setup this transaction, go ahead and exit */ 3532 if (block_group->cache_generation == trans->transid && 3533 i_size_read(inode)) { 3534 dcs = BTRFS_DC_SETUP; 3535 goto out_put; 3536 } 3537 3538 if (i_size_read(inode) > 0) { 3539 ret = btrfs_check_trunc_cache_free_space(fs_info, 3540 &fs_info->global_block_rsv); 3541 if (ret) 3542 goto out_put; 3543 3544 ret = btrfs_truncate_free_space_cache(trans, NULL, inode); 3545 if (ret) 3546 goto out_put; 3547 } 3548 3549 spin_lock(&block_group->lock); 3550 if (block_group->cached != BTRFS_CACHE_FINISHED || 3551 !btrfs_test_opt(fs_info, SPACE_CACHE)) { 3552 /* 3553 * don't bother trying to write stuff out _if_ 3554 * a) we're not cached, 3555 * b) we're with nospace_cache mount option, 3556 * c) we're with v2 space_cache (FREE_SPACE_TREE). 3557 */ 3558 dcs = BTRFS_DC_WRITTEN; 3559 spin_unlock(&block_group->lock); 3560 goto out_put; 3561 } 3562 spin_unlock(&block_group->lock); 3563 3564 /* 3565 * We hit an ENOSPC when setting up the cache in this transaction, just 3566 * skip doing the setup, we've already cleared the cache so we're safe. 3567 */ 3568 if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) { 3569 ret = -ENOSPC; 3570 goto out_put; 3571 } 3572 3573 /* 3574 * Try to preallocate enough space based on how big the block group is. 3575 * Keep in mind this has to include any pinned space which could end up 3576 * taking up quite a bit since it's not folded into the other space 3577 * cache. 3578 */ 3579 num_pages = div_u64(block_group->key.offset, SZ_256M); 3580 if (!num_pages) 3581 num_pages = 1; 3582 3583 num_pages *= 16; 3584 num_pages *= PAGE_SIZE; 3585 3586 ret = btrfs_check_data_free_space(inode, &data_reserved, 0, num_pages); 3587 if (ret) 3588 goto out_put; 3589 3590 ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages, 3591 num_pages, num_pages, 3592 &alloc_hint); 3593 /* 3594 * Our cache requires contiguous chunks so that we don't modify a bunch 3595 * of metadata or split extents when writing the cache out, which means 3596 * we can enospc if we are heavily fragmented in addition to just normal 3597 * out of space conditions. So if we hit this just skip setting up any 3598 * other block groups for this transaction, maybe we'll unpin enough 3599 * space the next time around. 3600 */ 3601 if (!ret) 3602 dcs = BTRFS_DC_SETUP; 3603 else if (ret == -ENOSPC) 3604 set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags); 3605 3606 out_put: 3607 iput(inode); 3608 out_free: 3609 btrfs_release_path(path); 3610 out: 3611 spin_lock(&block_group->lock); 3612 if (!ret && dcs == BTRFS_DC_SETUP) 3613 block_group->cache_generation = trans->transid; 3614 block_group->disk_cache_state = dcs; 3615 spin_unlock(&block_group->lock); 3616 3617 extent_changeset_free(data_reserved); 3618 return ret; 3619 } 3620 3621 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans, 3622 struct btrfs_fs_info *fs_info) 3623 { 3624 struct btrfs_block_group_cache *cache, *tmp; 3625 struct btrfs_transaction *cur_trans = trans->transaction; 3626 struct btrfs_path *path; 3627 3628 if (list_empty(&cur_trans->dirty_bgs) || 3629 !btrfs_test_opt(fs_info, SPACE_CACHE)) 3630 return 0; 3631 3632 path = btrfs_alloc_path(); 3633 if (!path) 3634 return -ENOMEM; 3635 3636 /* Could add new block groups, use _safe just in case */ 3637 list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs, 3638 dirty_list) { 3639 if (cache->disk_cache_state == BTRFS_DC_CLEAR) 3640 cache_save_setup(cache, trans, path); 3641 } 3642 3643 btrfs_free_path(path); 3644 return 0; 3645 } 3646 3647 /* 3648 * transaction commit does final block group cache writeback during a 3649 * critical section where nothing is allowed to change the FS. This is 3650 * required in order for the cache to actually match the block group, 3651 * but can introduce a lot of latency into the commit. 3652 * 3653 * So, btrfs_start_dirty_block_groups is here to kick off block group 3654 * cache IO. There's a chance we'll have to redo some of it if the 3655 * block group changes again during the commit, but it greatly reduces 3656 * the commit latency by getting rid of the easy block groups while 3657 * we're still allowing others to join the commit. 3658 */ 3659 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans, 3660 struct btrfs_fs_info *fs_info) 3661 { 3662 struct btrfs_block_group_cache *cache; 3663 struct btrfs_transaction *cur_trans = trans->transaction; 3664 int ret = 0; 3665 int should_put; 3666 struct btrfs_path *path = NULL; 3667 LIST_HEAD(dirty); 3668 struct list_head *io = &cur_trans->io_bgs; 3669 int num_started = 0; 3670 int loops = 0; 3671 3672 spin_lock(&cur_trans->dirty_bgs_lock); 3673 if (list_empty(&cur_trans->dirty_bgs)) { 3674 spin_unlock(&cur_trans->dirty_bgs_lock); 3675 return 0; 3676 } 3677 list_splice_init(&cur_trans->dirty_bgs, &dirty); 3678 spin_unlock(&cur_trans->dirty_bgs_lock); 3679 3680 again: 3681 /* 3682 * make sure all the block groups on our dirty list actually 3683 * exist 3684 */ 3685 btrfs_create_pending_block_groups(trans, fs_info); 3686 3687 if (!path) { 3688 path = btrfs_alloc_path(); 3689 if (!path) 3690 return -ENOMEM; 3691 } 3692 3693 /* 3694 * cache_write_mutex is here only to save us from balance or automatic 3695 * removal of empty block groups deleting this block group while we are 3696 * writing out the cache 3697 */ 3698 mutex_lock(&trans->transaction->cache_write_mutex); 3699 while (!list_empty(&dirty)) { 3700 cache = list_first_entry(&dirty, 3701 struct btrfs_block_group_cache, 3702 dirty_list); 3703 /* 3704 * this can happen if something re-dirties a block 3705 * group that is already under IO. Just wait for it to 3706 * finish and then do it all again 3707 */ 3708 if (!list_empty(&cache->io_list)) { 3709 list_del_init(&cache->io_list); 3710 btrfs_wait_cache_io(trans, cache, path); 3711 btrfs_put_block_group(cache); 3712 } 3713 3714 3715 /* 3716 * btrfs_wait_cache_io uses the cache->dirty_list to decide 3717 * if it should update the cache_state. Don't delete 3718 * until after we wait. 3719 * 3720 * Since we're not running in the commit critical section 3721 * we need the dirty_bgs_lock to protect from update_block_group 3722 */ 3723 spin_lock(&cur_trans->dirty_bgs_lock); 3724 list_del_init(&cache->dirty_list); 3725 spin_unlock(&cur_trans->dirty_bgs_lock); 3726 3727 should_put = 1; 3728 3729 cache_save_setup(cache, trans, path); 3730 3731 if (cache->disk_cache_state == BTRFS_DC_SETUP) { 3732 cache->io_ctl.inode = NULL; 3733 ret = btrfs_write_out_cache(fs_info, trans, 3734 cache, path); 3735 if (ret == 0 && cache->io_ctl.inode) { 3736 num_started++; 3737 should_put = 0; 3738 3739 /* 3740 * the cache_write_mutex is protecting 3741 * the io_list 3742 */ 3743 list_add_tail(&cache->io_list, io); 3744 } else { 3745 /* 3746 * if we failed to write the cache, the 3747 * generation will be bad and life goes on 3748 */ 3749 ret = 0; 3750 } 3751 } 3752 if (!ret) { 3753 ret = write_one_cache_group(trans, fs_info, 3754 path, cache); 3755 /* 3756 * Our block group might still be attached to the list 3757 * of new block groups in the transaction handle of some 3758 * other task (struct btrfs_trans_handle->new_bgs). This 3759 * means its block group item isn't yet in the extent 3760 * tree. If this happens ignore the error, as we will 3761 * try again later in the critical section of the 3762 * transaction commit. 3763 */ 3764 if (ret == -ENOENT) { 3765 ret = 0; 3766 spin_lock(&cur_trans->dirty_bgs_lock); 3767 if (list_empty(&cache->dirty_list)) { 3768 list_add_tail(&cache->dirty_list, 3769 &cur_trans->dirty_bgs); 3770 btrfs_get_block_group(cache); 3771 } 3772 spin_unlock(&cur_trans->dirty_bgs_lock); 3773 } else if (ret) { 3774 btrfs_abort_transaction(trans, ret); 3775 } 3776 } 3777 3778 /* if its not on the io list, we need to put the block group */ 3779 if (should_put) 3780 btrfs_put_block_group(cache); 3781 3782 if (ret) 3783 break; 3784 3785 /* 3786 * Avoid blocking other tasks for too long. It might even save 3787 * us from writing caches for block groups that are going to be 3788 * removed. 3789 */ 3790 mutex_unlock(&trans->transaction->cache_write_mutex); 3791 mutex_lock(&trans->transaction->cache_write_mutex); 3792 } 3793 mutex_unlock(&trans->transaction->cache_write_mutex); 3794 3795 /* 3796 * go through delayed refs for all the stuff we've just kicked off 3797 * and then loop back (just once) 3798 */ 3799 ret = btrfs_run_delayed_refs(trans, fs_info, 0); 3800 if (!ret && loops == 0) { 3801 loops++; 3802 spin_lock(&cur_trans->dirty_bgs_lock); 3803 list_splice_init(&cur_trans->dirty_bgs, &dirty); 3804 /* 3805 * dirty_bgs_lock protects us from concurrent block group 3806 * deletes too (not just cache_write_mutex). 3807 */ 3808 if (!list_empty(&dirty)) { 3809 spin_unlock(&cur_trans->dirty_bgs_lock); 3810 goto again; 3811 } 3812 spin_unlock(&cur_trans->dirty_bgs_lock); 3813 } else if (ret < 0) { 3814 btrfs_cleanup_dirty_bgs(cur_trans, fs_info); 3815 } 3816 3817 btrfs_free_path(path); 3818 return ret; 3819 } 3820 3821 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans, 3822 struct btrfs_fs_info *fs_info) 3823 { 3824 struct btrfs_block_group_cache *cache; 3825 struct btrfs_transaction *cur_trans = trans->transaction; 3826 int ret = 0; 3827 int should_put; 3828 struct btrfs_path *path; 3829 struct list_head *io = &cur_trans->io_bgs; 3830 int num_started = 0; 3831 3832 path = btrfs_alloc_path(); 3833 if (!path) 3834 return -ENOMEM; 3835 3836 /* 3837 * Even though we are in the critical section of the transaction commit, 3838 * we can still have concurrent tasks adding elements to this 3839 * transaction's list of dirty block groups. These tasks correspond to 3840 * endio free space workers started when writeback finishes for a 3841 * space cache, which run inode.c:btrfs_finish_ordered_io(), and can 3842 * allocate new block groups as a result of COWing nodes of the root 3843 * tree when updating the free space inode. The writeback for the space 3844 * caches is triggered by an earlier call to 3845 * btrfs_start_dirty_block_groups() and iterations of the following 3846 * loop. 3847 * Also we want to do the cache_save_setup first and then run the 3848 * delayed refs to make sure we have the best chance at doing this all 3849 * in one shot. 3850 */ 3851 spin_lock(&cur_trans->dirty_bgs_lock); 3852 while (!list_empty(&cur_trans->dirty_bgs)) { 3853 cache = list_first_entry(&cur_trans->dirty_bgs, 3854 struct btrfs_block_group_cache, 3855 dirty_list); 3856 3857 /* 3858 * this can happen if cache_save_setup re-dirties a block 3859 * group that is already under IO. Just wait for it to 3860 * finish and then do it all again 3861 */ 3862 if (!list_empty(&cache->io_list)) { 3863 spin_unlock(&cur_trans->dirty_bgs_lock); 3864 list_del_init(&cache->io_list); 3865 btrfs_wait_cache_io(trans, cache, path); 3866 btrfs_put_block_group(cache); 3867 spin_lock(&cur_trans->dirty_bgs_lock); 3868 } 3869 3870 /* 3871 * don't remove from the dirty list until after we've waited 3872 * on any pending IO 3873 */ 3874 list_del_init(&cache->dirty_list); 3875 spin_unlock(&cur_trans->dirty_bgs_lock); 3876 should_put = 1; 3877 3878 cache_save_setup(cache, trans, path); 3879 3880 if (!ret) 3881 ret = btrfs_run_delayed_refs(trans, fs_info, 3882 (unsigned long) -1); 3883 3884 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) { 3885 cache->io_ctl.inode = NULL; 3886 ret = btrfs_write_out_cache(fs_info, trans, 3887 cache, path); 3888 if (ret == 0 && cache->io_ctl.inode) { 3889 num_started++; 3890 should_put = 0; 3891 list_add_tail(&cache->io_list, io); 3892 } else { 3893 /* 3894 * if we failed to write the cache, the 3895 * generation will be bad and life goes on 3896 */ 3897 ret = 0; 3898 } 3899 } 3900 if (!ret) { 3901 ret = write_one_cache_group(trans, fs_info, 3902 path, cache); 3903 /* 3904 * One of the free space endio workers might have 3905 * created a new block group while updating a free space 3906 * cache's inode (at inode.c:btrfs_finish_ordered_io()) 3907 * and hasn't released its transaction handle yet, in 3908 * which case the new block group is still attached to 3909 * its transaction handle and its creation has not 3910 * finished yet (no block group item in the extent tree 3911 * yet, etc). If this is the case, wait for all free 3912 * space endio workers to finish and retry. This is a 3913 * a very rare case so no need for a more efficient and 3914 * complex approach. 3915 */ 3916 if (ret == -ENOENT) { 3917 wait_event(cur_trans->writer_wait, 3918 atomic_read(&cur_trans->num_writers) == 1); 3919 ret = write_one_cache_group(trans, fs_info, 3920 path, cache); 3921 } 3922 if (ret) 3923 btrfs_abort_transaction(trans, ret); 3924 } 3925 3926 /* if its not on the io list, we need to put the block group */ 3927 if (should_put) 3928 btrfs_put_block_group(cache); 3929 spin_lock(&cur_trans->dirty_bgs_lock); 3930 } 3931 spin_unlock(&cur_trans->dirty_bgs_lock); 3932 3933 while (!list_empty(io)) { 3934 cache = list_first_entry(io, struct btrfs_block_group_cache, 3935 io_list); 3936 list_del_init(&cache->io_list); 3937 btrfs_wait_cache_io(trans, cache, path); 3938 btrfs_put_block_group(cache); 3939 } 3940 3941 btrfs_free_path(path); 3942 return ret; 3943 } 3944 3945 int btrfs_extent_readonly(struct btrfs_fs_info *fs_info, u64 bytenr) 3946 { 3947 struct btrfs_block_group_cache *block_group; 3948 int readonly = 0; 3949 3950 block_group = btrfs_lookup_block_group(fs_info, bytenr); 3951 if (!block_group || block_group->ro) 3952 readonly = 1; 3953 if (block_group) 3954 btrfs_put_block_group(block_group); 3955 return readonly; 3956 } 3957 3958 bool btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr) 3959 { 3960 struct btrfs_block_group_cache *bg; 3961 bool ret = true; 3962 3963 bg = btrfs_lookup_block_group(fs_info, bytenr); 3964 if (!bg) 3965 return false; 3966 3967 spin_lock(&bg->lock); 3968 if (bg->ro) 3969 ret = false; 3970 else 3971 atomic_inc(&bg->nocow_writers); 3972 spin_unlock(&bg->lock); 3973 3974 /* no put on block group, done by btrfs_dec_nocow_writers */ 3975 if (!ret) 3976 btrfs_put_block_group(bg); 3977 3978 return ret; 3979 3980 } 3981 3982 void btrfs_dec_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr) 3983 { 3984 struct btrfs_block_group_cache *bg; 3985 3986 bg = btrfs_lookup_block_group(fs_info, bytenr); 3987 ASSERT(bg); 3988 if (atomic_dec_and_test(&bg->nocow_writers)) 3989 wake_up_atomic_t(&bg->nocow_writers); 3990 /* 3991 * Once for our lookup and once for the lookup done by a previous call 3992 * to btrfs_inc_nocow_writers() 3993 */ 3994 btrfs_put_block_group(bg); 3995 btrfs_put_block_group(bg); 3996 } 3997 3998 void btrfs_wait_nocow_writers(struct btrfs_block_group_cache *bg) 3999 { 4000 wait_on_atomic_t(&bg->nocow_writers, atomic_t_wait, 4001 TASK_UNINTERRUPTIBLE); 4002 } 4003 4004 static const char *alloc_name(u64 flags) 4005 { 4006 switch (flags) { 4007 case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA: 4008 return "mixed"; 4009 case BTRFS_BLOCK_GROUP_METADATA: 4010 return "metadata"; 4011 case BTRFS_BLOCK_GROUP_DATA: 4012 return "data"; 4013 case BTRFS_BLOCK_GROUP_SYSTEM: 4014 return "system"; 4015 default: 4016 WARN_ON(1); 4017 return "invalid-combination"; 4018 }; 4019 } 4020 4021 static int create_space_info(struct btrfs_fs_info *info, u64 flags, 4022 struct btrfs_space_info **new) 4023 { 4024 4025 struct btrfs_space_info *space_info; 4026 int i; 4027 int ret; 4028 4029 space_info = kzalloc(sizeof(*space_info), GFP_NOFS); 4030 if (!space_info) 4031 return -ENOMEM; 4032 4033 ret = percpu_counter_init(&space_info->total_bytes_pinned, 0, 4034 GFP_KERNEL); 4035 if (ret) { 4036 kfree(space_info); 4037 return ret; 4038 } 4039 4040 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) 4041 INIT_LIST_HEAD(&space_info->block_groups[i]); 4042 init_rwsem(&space_info->groups_sem); 4043 spin_lock_init(&space_info->lock); 4044 space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK; 4045 space_info->force_alloc = CHUNK_ALLOC_NO_FORCE; 4046 init_waitqueue_head(&space_info->wait); 4047 INIT_LIST_HEAD(&space_info->ro_bgs); 4048 INIT_LIST_HEAD(&space_info->tickets); 4049 INIT_LIST_HEAD(&space_info->priority_tickets); 4050 4051 ret = kobject_init_and_add(&space_info->kobj, &space_info_ktype, 4052 info->space_info_kobj, "%s", 4053 alloc_name(space_info->flags)); 4054 if (ret) { 4055 percpu_counter_destroy(&space_info->total_bytes_pinned); 4056 kfree(space_info); 4057 return ret; 4058 } 4059 4060 *new = space_info; 4061 list_add_rcu(&space_info->list, &info->space_info); 4062 if (flags & BTRFS_BLOCK_GROUP_DATA) 4063 info->data_sinfo = space_info; 4064 4065 return ret; 4066 } 4067 4068 static void update_space_info(struct btrfs_fs_info *info, u64 flags, 4069 u64 total_bytes, u64 bytes_used, 4070 u64 bytes_readonly, 4071 struct btrfs_space_info **space_info) 4072 { 4073 struct btrfs_space_info *found; 4074 int factor; 4075 4076 if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 | 4077 BTRFS_BLOCK_GROUP_RAID10)) 4078 factor = 2; 4079 else 4080 factor = 1; 4081 4082 found = __find_space_info(info, flags); 4083 ASSERT(found); 4084 spin_lock(&found->lock); 4085 found->total_bytes += total_bytes; 4086 found->disk_total += total_bytes * factor; 4087 found->bytes_used += bytes_used; 4088 found->disk_used += bytes_used * factor; 4089 found->bytes_readonly += bytes_readonly; 4090 if (total_bytes > 0) 4091 found->full = 0; 4092 space_info_add_new_bytes(info, found, total_bytes - 4093 bytes_used - bytes_readonly); 4094 spin_unlock(&found->lock); 4095 *space_info = found; 4096 } 4097 4098 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags) 4099 { 4100 u64 extra_flags = chunk_to_extended(flags) & 4101 BTRFS_EXTENDED_PROFILE_MASK; 4102 4103 write_seqlock(&fs_info->profiles_lock); 4104 if (flags & BTRFS_BLOCK_GROUP_DATA) 4105 fs_info->avail_data_alloc_bits |= extra_flags; 4106 if (flags & BTRFS_BLOCK_GROUP_METADATA) 4107 fs_info->avail_metadata_alloc_bits |= extra_flags; 4108 if (flags & BTRFS_BLOCK_GROUP_SYSTEM) 4109 fs_info->avail_system_alloc_bits |= extra_flags; 4110 write_sequnlock(&fs_info->profiles_lock); 4111 } 4112 4113 /* 4114 * returns target flags in extended format or 0 if restripe for this 4115 * chunk_type is not in progress 4116 * 4117 * should be called with either volume_mutex or balance_lock held 4118 */ 4119 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags) 4120 { 4121 struct btrfs_balance_control *bctl = fs_info->balance_ctl; 4122 u64 target = 0; 4123 4124 if (!bctl) 4125 return 0; 4126 4127 if (flags & BTRFS_BLOCK_GROUP_DATA && 4128 bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) { 4129 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target; 4130 } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM && 4131 bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) { 4132 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target; 4133 } else if (flags & BTRFS_BLOCK_GROUP_METADATA && 4134 bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) { 4135 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target; 4136 } 4137 4138 return target; 4139 } 4140 4141 /* 4142 * @flags: available profiles in extended format (see ctree.h) 4143 * 4144 * Returns reduced profile in chunk format. If profile changing is in 4145 * progress (either running or paused) picks the target profile (if it's 4146 * already available), otherwise falls back to plain reducing. 4147 */ 4148 static u64 btrfs_reduce_alloc_profile(struct btrfs_fs_info *fs_info, u64 flags) 4149 { 4150 u64 num_devices = fs_info->fs_devices->rw_devices; 4151 u64 target; 4152 u64 raid_type; 4153 u64 allowed = 0; 4154 4155 /* 4156 * see if restripe for this chunk_type is in progress, if so 4157 * try to reduce to the target profile 4158 */ 4159 spin_lock(&fs_info->balance_lock); 4160 target = get_restripe_target(fs_info, flags); 4161 if (target) { 4162 /* pick target profile only if it's already available */ 4163 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) { 4164 spin_unlock(&fs_info->balance_lock); 4165 return extended_to_chunk(target); 4166 } 4167 } 4168 spin_unlock(&fs_info->balance_lock); 4169 4170 /* First, mask out the RAID levels which aren't possible */ 4171 for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) { 4172 if (num_devices >= btrfs_raid_array[raid_type].devs_min) 4173 allowed |= btrfs_raid_group[raid_type]; 4174 } 4175 allowed &= flags; 4176 4177 if (allowed & BTRFS_BLOCK_GROUP_RAID6) 4178 allowed = BTRFS_BLOCK_GROUP_RAID6; 4179 else if (allowed & BTRFS_BLOCK_GROUP_RAID5) 4180 allowed = BTRFS_BLOCK_GROUP_RAID5; 4181 else if (allowed & BTRFS_BLOCK_GROUP_RAID10) 4182 allowed = BTRFS_BLOCK_GROUP_RAID10; 4183 else if (allowed & BTRFS_BLOCK_GROUP_RAID1) 4184 allowed = BTRFS_BLOCK_GROUP_RAID1; 4185 else if (allowed & BTRFS_BLOCK_GROUP_RAID0) 4186 allowed = BTRFS_BLOCK_GROUP_RAID0; 4187 4188 flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK; 4189 4190 return extended_to_chunk(flags | allowed); 4191 } 4192 4193 static u64 get_alloc_profile(struct btrfs_fs_info *fs_info, u64 orig_flags) 4194 { 4195 unsigned seq; 4196 u64 flags; 4197 4198 do { 4199 flags = orig_flags; 4200 seq = read_seqbegin(&fs_info->profiles_lock); 4201 4202 if (flags & BTRFS_BLOCK_GROUP_DATA) 4203 flags |= fs_info->avail_data_alloc_bits; 4204 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM) 4205 flags |= fs_info->avail_system_alloc_bits; 4206 else if (flags & BTRFS_BLOCK_GROUP_METADATA) 4207 flags |= fs_info->avail_metadata_alloc_bits; 4208 } while (read_seqretry(&fs_info->profiles_lock, seq)); 4209 4210 return btrfs_reduce_alloc_profile(fs_info, flags); 4211 } 4212 4213 static u64 get_alloc_profile_by_root(struct btrfs_root *root, int data) 4214 { 4215 struct btrfs_fs_info *fs_info = root->fs_info; 4216 u64 flags; 4217 u64 ret; 4218 4219 if (data) 4220 flags = BTRFS_BLOCK_GROUP_DATA; 4221 else if (root == fs_info->chunk_root) 4222 flags = BTRFS_BLOCK_GROUP_SYSTEM; 4223 else 4224 flags = BTRFS_BLOCK_GROUP_METADATA; 4225 4226 ret = get_alloc_profile(fs_info, flags); 4227 return ret; 4228 } 4229 4230 u64 btrfs_data_alloc_profile(struct btrfs_fs_info *fs_info) 4231 { 4232 return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_DATA); 4233 } 4234 4235 u64 btrfs_metadata_alloc_profile(struct btrfs_fs_info *fs_info) 4236 { 4237 return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_METADATA); 4238 } 4239 4240 u64 btrfs_system_alloc_profile(struct btrfs_fs_info *fs_info) 4241 { 4242 return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_SYSTEM); 4243 } 4244 4245 static u64 btrfs_space_info_used(struct btrfs_space_info *s_info, 4246 bool may_use_included) 4247 { 4248 ASSERT(s_info); 4249 return s_info->bytes_used + s_info->bytes_reserved + 4250 s_info->bytes_pinned + s_info->bytes_readonly + 4251 (may_use_included ? s_info->bytes_may_use : 0); 4252 } 4253 4254 int btrfs_alloc_data_chunk_ondemand(struct btrfs_inode *inode, u64 bytes) 4255 { 4256 struct btrfs_root *root = inode->root; 4257 struct btrfs_fs_info *fs_info = root->fs_info; 4258 struct btrfs_space_info *data_sinfo = fs_info->data_sinfo; 4259 u64 used; 4260 int ret = 0; 4261 int need_commit = 2; 4262 int have_pinned_space; 4263 4264 /* make sure bytes are sectorsize aligned */ 4265 bytes = ALIGN(bytes, fs_info->sectorsize); 4266 4267 if (btrfs_is_free_space_inode(inode)) { 4268 need_commit = 0; 4269 ASSERT(current->journal_info); 4270 } 4271 4272 again: 4273 /* make sure we have enough space to handle the data first */ 4274 spin_lock(&data_sinfo->lock); 4275 used = btrfs_space_info_used(data_sinfo, true); 4276 4277 if (used + bytes > data_sinfo->total_bytes) { 4278 struct btrfs_trans_handle *trans; 4279 4280 /* 4281 * if we don't have enough free bytes in this space then we need 4282 * to alloc a new chunk. 4283 */ 4284 if (!data_sinfo->full) { 4285 u64 alloc_target; 4286 4287 data_sinfo->force_alloc = CHUNK_ALLOC_FORCE; 4288 spin_unlock(&data_sinfo->lock); 4289 4290 alloc_target = btrfs_data_alloc_profile(fs_info); 4291 /* 4292 * It is ugly that we don't call nolock join 4293 * transaction for the free space inode case here. 4294 * But it is safe because we only do the data space 4295 * reservation for the free space cache in the 4296 * transaction context, the common join transaction 4297 * just increase the counter of the current transaction 4298 * handler, doesn't try to acquire the trans_lock of 4299 * the fs. 4300 */ 4301 trans = btrfs_join_transaction(root); 4302 if (IS_ERR(trans)) 4303 return PTR_ERR(trans); 4304 4305 ret = do_chunk_alloc(trans, fs_info, alloc_target, 4306 CHUNK_ALLOC_NO_FORCE); 4307 btrfs_end_transaction(trans); 4308 if (ret < 0) { 4309 if (ret != -ENOSPC) 4310 return ret; 4311 else { 4312 have_pinned_space = 1; 4313 goto commit_trans; 4314 } 4315 } 4316 4317 goto again; 4318 } 4319 4320 /* 4321 * If we don't have enough pinned space to deal with this 4322 * allocation, and no removed chunk in current transaction, 4323 * don't bother committing the transaction. 4324 */ 4325 have_pinned_space = percpu_counter_compare( 4326 &data_sinfo->total_bytes_pinned, 4327 used + bytes - data_sinfo->total_bytes); 4328 spin_unlock(&data_sinfo->lock); 4329 4330 /* commit the current transaction and try again */ 4331 commit_trans: 4332 if (need_commit && 4333 !atomic_read(&fs_info->open_ioctl_trans)) { 4334 need_commit--; 4335 4336 if (need_commit > 0) { 4337 btrfs_start_delalloc_roots(fs_info, 0, -1); 4338 btrfs_wait_ordered_roots(fs_info, U64_MAX, 0, 4339 (u64)-1); 4340 } 4341 4342 trans = btrfs_join_transaction(root); 4343 if (IS_ERR(trans)) 4344 return PTR_ERR(trans); 4345 if (have_pinned_space >= 0 || 4346 test_bit(BTRFS_TRANS_HAVE_FREE_BGS, 4347 &trans->transaction->flags) || 4348 need_commit > 0) { 4349 ret = btrfs_commit_transaction(trans); 4350 if (ret) 4351 return ret; 4352 /* 4353 * The cleaner kthread might still be doing iput 4354 * operations. Wait for it to finish so that 4355 * more space is released. 4356 */ 4357 mutex_lock(&fs_info->cleaner_delayed_iput_mutex); 4358 mutex_unlock(&fs_info->cleaner_delayed_iput_mutex); 4359 goto again; 4360 } else { 4361 btrfs_end_transaction(trans); 4362 } 4363 } 4364 4365 trace_btrfs_space_reservation(fs_info, 4366 "space_info:enospc", 4367 data_sinfo->flags, bytes, 1); 4368 return -ENOSPC; 4369 } 4370 data_sinfo->bytes_may_use += bytes; 4371 trace_btrfs_space_reservation(fs_info, "space_info", 4372 data_sinfo->flags, bytes, 1); 4373 spin_unlock(&data_sinfo->lock); 4374 4375 return ret; 4376 } 4377 4378 int btrfs_check_data_free_space(struct inode *inode, 4379 struct extent_changeset **reserved, u64 start, u64 len) 4380 { 4381 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); 4382 int ret; 4383 4384 /* align the range */ 4385 len = round_up(start + len, fs_info->sectorsize) - 4386 round_down(start, fs_info->sectorsize); 4387 start = round_down(start, fs_info->sectorsize); 4388 4389 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode), len); 4390 if (ret < 0) 4391 return ret; 4392 4393 /* Use new btrfs_qgroup_reserve_data to reserve precious data space. */ 4394 ret = btrfs_qgroup_reserve_data(inode, reserved, start, len); 4395 if (ret < 0) 4396 btrfs_free_reserved_data_space_noquota(inode, start, len); 4397 else 4398 ret = 0; 4399 return ret; 4400 } 4401 4402 /* 4403 * Called if we need to clear a data reservation for this inode 4404 * Normally in a error case. 4405 * 4406 * This one will *NOT* use accurate qgroup reserved space API, just for case 4407 * which we can't sleep and is sure it won't affect qgroup reserved space. 4408 * Like clear_bit_hook(). 4409 */ 4410 void btrfs_free_reserved_data_space_noquota(struct inode *inode, u64 start, 4411 u64 len) 4412 { 4413 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); 4414 struct btrfs_space_info *data_sinfo; 4415 4416 /* Make sure the range is aligned to sectorsize */ 4417 len = round_up(start + len, fs_info->sectorsize) - 4418 round_down(start, fs_info->sectorsize); 4419 start = round_down(start, fs_info->sectorsize); 4420 4421 data_sinfo = fs_info->data_sinfo; 4422 spin_lock(&data_sinfo->lock); 4423 if (WARN_ON(data_sinfo->bytes_may_use < len)) 4424 data_sinfo->bytes_may_use = 0; 4425 else 4426 data_sinfo->bytes_may_use -= len; 4427 trace_btrfs_space_reservation(fs_info, "space_info", 4428 data_sinfo->flags, len, 0); 4429 spin_unlock(&data_sinfo->lock); 4430 } 4431 4432 /* 4433 * Called if we need to clear a data reservation for this inode 4434 * Normally in a error case. 4435 * 4436 * This one will handle the per-inode data rsv map for accurate reserved 4437 * space framework. 4438 */ 4439 void btrfs_free_reserved_data_space(struct inode *inode, 4440 struct extent_changeset *reserved, u64 start, u64 len) 4441 { 4442 struct btrfs_root *root = BTRFS_I(inode)->root; 4443 4444 /* Make sure the range is aligned to sectorsize */ 4445 len = round_up(start + len, root->fs_info->sectorsize) - 4446 round_down(start, root->fs_info->sectorsize); 4447 start = round_down(start, root->fs_info->sectorsize); 4448 4449 btrfs_free_reserved_data_space_noquota(inode, start, len); 4450 btrfs_qgroup_free_data(inode, reserved, start, len); 4451 } 4452 4453 static void force_metadata_allocation(struct btrfs_fs_info *info) 4454 { 4455 struct list_head *head = &info->space_info; 4456 struct btrfs_space_info *found; 4457 4458 rcu_read_lock(); 4459 list_for_each_entry_rcu(found, head, list) { 4460 if (found->flags & BTRFS_BLOCK_GROUP_METADATA) 4461 found->force_alloc = CHUNK_ALLOC_FORCE; 4462 } 4463 rcu_read_unlock(); 4464 } 4465 4466 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global) 4467 { 4468 return (global->size << 1); 4469 } 4470 4471 static int should_alloc_chunk(struct btrfs_fs_info *fs_info, 4472 struct btrfs_space_info *sinfo, int force) 4473 { 4474 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv; 4475 u64 bytes_used = btrfs_space_info_used(sinfo, false); 4476 u64 thresh; 4477 4478 if (force == CHUNK_ALLOC_FORCE) 4479 return 1; 4480 4481 /* 4482 * We need to take into account the global rsv because for all intents 4483 * and purposes it's used space. Don't worry about locking the 4484 * global_rsv, it doesn't change except when the transaction commits. 4485 */ 4486 if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA) 4487 bytes_used += calc_global_rsv_need_space(global_rsv); 4488 4489 /* 4490 * in limited mode, we want to have some free space up to 4491 * about 1% of the FS size. 4492 */ 4493 if (force == CHUNK_ALLOC_LIMITED) { 4494 thresh = btrfs_super_total_bytes(fs_info->super_copy); 4495 thresh = max_t(u64, SZ_64M, div_factor_fine(thresh, 1)); 4496 4497 if (sinfo->total_bytes - bytes_used < thresh) 4498 return 1; 4499 } 4500 4501 if (bytes_used + SZ_2M < div_factor(sinfo->total_bytes, 8)) 4502 return 0; 4503 return 1; 4504 } 4505 4506 static u64 get_profile_num_devs(struct btrfs_fs_info *fs_info, u64 type) 4507 { 4508 u64 num_dev; 4509 4510 if (type & (BTRFS_BLOCK_GROUP_RAID10 | 4511 BTRFS_BLOCK_GROUP_RAID0 | 4512 BTRFS_BLOCK_GROUP_RAID5 | 4513 BTRFS_BLOCK_GROUP_RAID6)) 4514 num_dev = fs_info->fs_devices->rw_devices; 4515 else if (type & BTRFS_BLOCK_GROUP_RAID1) 4516 num_dev = 2; 4517 else 4518 num_dev = 1; /* DUP or single */ 4519 4520 return num_dev; 4521 } 4522 4523 /* 4524 * If @is_allocation is true, reserve space in the system space info necessary 4525 * for allocating a chunk, otherwise if it's false, reserve space necessary for 4526 * removing a chunk. 4527 */ 4528 void check_system_chunk(struct btrfs_trans_handle *trans, 4529 struct btrfs_fs_info *fs_info, u64 type) 4530 { 4531 struct btrfs_space_info *info; 4532 u64 left; 4533 u64 thresh; 4534 int ret = 0; 4535 u64 num_devs; 4536 4537 /* 4538 * Needed because we can end up allocating a system chunk and for an 4539 * atomic and race free space reservation in the chunk block reserve. 4540 */ 4541 ASSERT(mutex_is_locked(&fs_info->chunk_mutex)); 4542 4543 info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM); 4544 spin_lock(&info->lock); 4545 left = info->total_bytes - btrfs_space_info_used(info, true); 4546 spin_unlock(&info->lock); 4547 4548 num_devs = get_profile_num_devs(fs_info, type); 4549 4550 /* num_devs device items to update and 1 chunk item to add or remove */ 4551 thresh = btrfs_calc_trunc_metadata_size(fs_info, num_devs) + 4552 btrfs_calc_trans_metadata_size(fs_info, 1); 4553 4554 if (left < thresh && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) { 4555 btrfs_info(fs_info, "left=%llu, need=%llu, flags=%llu", 4556 left, thresh, type); 4557 dump_space_info(fs_info, info, 0, 0); 4558 } 4559 4560 if (left < thresh) { 4561 u64 flags = btrfs_system_alloc_profile(fs_info); 4562 4563 /* 4564 * Ignore failure to create system chunk. We might end up not 4565 * needing it, as we might not need to COW all nodes/leafs from 4566 * the paths we visit in the chunk tree (they were already COWed 4567 * or created in the current transaction for example). 4568 */ 4569 ret = btrfs_alloc_chunk(trans, fs_info, flags); 4570 } 4571 4572 if (!ret) { 4573 ret = btrfs_block_rsv_add(fs_info->chunk_root, 4574 &fs_info->chunk_block_rsv, 4575 thresh, BTRFS_RESERVE_NO_FLUSH); 4576 if (!ret) 4577 trans->chunk_bytes_reserved += thresh; 4578 } 4579 } 4580 4581 /* 4582 * If force is CHUNK_ALLOC_FORCE: 4583 * - return 1 if it successfully allocates a chunk, 4584 * - return errors including -ENOSPC otherwise. 4585 * If force is NOT CHUNK_ALLOC_FORCE: 4586 * - return 0 if it doesn't need to allocate a new chunk, 4587 * - return 1 if it successfully allocates a chunk, 4588 * - return errors including -ENOSPC otherwise. 4589 */ 4590 static int do_chunk_alloc(struct btrfs_trans_handle *trans, 4591 struct btrfs_fs_info *fs_info, u64 flags, int force) 4592 { 4593 struct btrfs_space_info *space_info; 4594 int wait_for_alloc = 0; 4595 int ret = 0; 4596 4597 /* Don't re-enter if we're already allocating a chunk */ 4598 if (trans->allocating_chunk) 4599 return -ENOSPC; 4600 4601 space_info = __find_space_info(fs_info, flags); 4602 if (!space_info) { 4603 ret = create_space_info(fs_info, flags, &space_info); 4604 if (ret) 4605 return ret; 4606 } 4607 4608 again: 4609 spin_lock(&space_info->lock); 4610 if (force < space_info->force_alloc) 4611 force = space_info->force_alloc; 4612 if (space_info->full) { 4613 if (should_alloc_chunk(fs_info, space_info, force)) 4614 ret = -ENOSPC; 4615 else 4616 ret = 0; 4617 spin_unlock(&space_info->lock); 4618 return ret; 4619 } 4620 4621 if (!should_alloc_chunk(fs_info, space_info, force)) { 4622 spin_unlock(&space_info->lock); 4623 return 0; 4624 } else if (space_info->chunk_alloc) { 4625 wait_for_alloc = 1; 4626 } else { 4627 space_info->chunk_alloc = 1; 4628 } 4629 4630 spin_unlock(&space_info->lock); 4631 4632 mutex_lock(&fs_info->chunk_mutex); 4633 4634 /* 4635 * The chunk_mutex is held throughout the entirety of a chunk 4636 * allocation, so once we've acquired the chunk_mutex we know that the 4637 * other guy is done and we need to recheck and see if we should 4638 * allocate. 4639 */ 4640 if (wait_for_alloc) { 4641 mutex_unlock(&fs_info->chunk_mutex); 4642 wait_for_alloc = 0; 4643 goto again; 4644 } 4645 4646 trans->allocating_chunk = true; 4647 4648 /* 4649 * If we have mixed data/metadata chunks we want to make sure we keep 4650 * allocating mixed chunks instead of individual chunks. 4651 */ 4652 if (btrfs_mixed_space_info(space_info)) 4653 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA); 4654 4655 /* 4656 * if we're doing a data chunk, go ahead and make sure that 4657 * we keep a reasonable number of metadata chunks allocated in the 4658 * FS as well. 4659 */ 4660 if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) { 4661 fs_info->data_chunk_allocations++; 4662 if (!(fs_info->data_chunk_allocations % 4663 fs_info->metadata_ratio)) 4664 force_metadata_allocation(fs_info); 4665 } 4666 4667 /* 4668 * Check if we have enough space in SYSTEM chunk because we may need 4669 * to update devices. 4670 */ 4671 check_system_chunk(trans, fs_info, flags); 4672 4673 ret = btrfs_alloc_chunk(trans, fs_info, flags); 4674 trans->allocating_chunk = false; 4675 4676 spin_lock(&space_info->lock); 4677 if (ret < 0 && ret != -ENOSPC) 4678 goto out; 4679 if (ret) 4680 space_info->full = 1; 4681 else 4682 ret = 1; 4683 4684 space_info->force_alloc = CHUNK_ALLOC_NO_FORCE; 4685 out: 4686 space_info->chunk_alloc = 0; 4687 spin_unlock(&space_info->lock); 4688 mutex_unlock(&fs_info->chunk_mutex); 4689 /* 4690 * When we allocate a new chunk we reserve space in the chunk block 4691 * reserve to make sure we can COW nodes/leafs in the chunk tree or 4692 * add new nodes/leafs to it if we end up needing to do it when 4693 * inserting the chunk item and updating device items as part of the 4694 * second phase of chunk allocation, performed by 4695 * btrfs_finish_chunk_alloc(). So make sure we don't accumulate a 4696 * large number of new block groups to create in our transaction 4697 * handle's new_bgs list to avoid exhausting the chunk block reserve 4698 * in extreme cases - like having a single transaction create many new 4699 * block groups when starting to write out the free space caches of all 4700 * the block groups that were made dirty during the lifetime of the 4701 * transaction. 4702 */ 4703 if (trans->can_flush_pending_bgs && 4704 trans->chunk_bytes_reserved >= (u64)SZ_2M) { 4705 btrfs_create_pending_block_groups(trans, fs_info); 4706 btrfs_trans_release_chunk_metadata(trans); 4707 } 4708 return ret; 4709 } 4710 4711 static int can_overcommit(struct btrfs_fs_info *fs_info, 4712 struct btrfs_space_info *space_info, u64 bytes, 4713 enum btrfs_reserve_flush_enum flush, 4714 bool system_chunk) 4715 { 4716 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv; 4717 u64 profile; 4718 u64 space_size; 4719 u64 avail; 4720 u64 used; 4721 4722 /* Don't overcommit when in mixed mode. */ 4723 if (space_info->flags & BTRFS_BLOCK_GROUP_DATA) 4724 return 0; 4725 4726 if (system_chunk) 4727 profile = btrfs_system_alloc_profile(fs_info); 4728 else 4729 profile = btrfs_metadata_alloc_profile(fs_info); 4730 4731 used = btrfs_space_info_used(space_info, false); 4732 4733 /* 4734 * We only want to allow over committing if we have lots of actual space 4735 * free, but if we don't have enough space to handle the global reserve 4736 * space then we could end up having a real enospc problem when trying 4737 * to allocate a chunk or some other such important allocation. 4738 */ 4739 spin_lock(&global_rsv->lock); 4740 space_size = calc_global_rsv_need_space(global_rsv); 4741 spin_unlock(&global_rsv->lock); 4742 if (used + space_size >= space_info->total_bytes) 4743 return 0; 4744 4745 used += space_info->bytes_may_use; 4746 4747 avail = atomic64_read(&fs_info->free_chunk_space); 4748 4749 /* 4750 * If we have dup, raid1 or raid10 then only half of the free 4751 * space is actually useable. For raid56, the space info used 4752 * doesn't include the parity drive, so we don't have to 4753 * change the math 4754 */ 4755 if (profile & (BTRFS_BLOCK_GROUP_DUP | 4756 BTRFS_BLOCK_GROUP_RAID1 | 4757 BTRFS_BLOCK_GROUP_RAID10)) 4758 avail >>= 1; 4759 4760 /* 4761 * If we aren't flushing all things, let us overcommit up to 4762 * 1/2th of the space. If we can flush, don't let us overcommit 4763 * too much, let it overcommit up to 1/8 of the space. 4764 */ 4765 if (flush == BTRFS_RESERVE_FLUSH_ALL) 4766 avail >>= 3; 4767 else 4768 avail >>= 1; 4769 4770 if (used + bytes < space_info->total_bytes + avail) 4771 return 1; 4772 return 0; 4773 } 4774 4775 static void btrfs_writeback_inodes_sb_nr(struct btrfs_fs_info *fs_info, 4776 unsigned long nr_pages, int nr_items) 4777 { 4778 struct super_block *sb = fs_info->sb; 4779 4780 if (down_read_trylock(&sb->s_umount)) { 4781 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE); 4782 up_read(&sb->s_umount); 4783 } else { 4784 /* 4785 * We needn't worry the filesystem going from r/w to r/o though 4786 * we don't acquire ->s_umount mutex, because the filesystem 4787 * should guarantee the delalloc inodes list be empty after 4788 * the filesystem is readonly(all dirty pages are written to 4789 * the disk). 4790 */ 4791 btrfs_start_delalloc_roots(fs_info, 0, nr_items); 4792 if (!current->journal_info) 4793 btrfs_wait_ordered_roots(fs_info, nr_items, 0, (u64)-1); 4794 } 4795 } 4796 4797 static inline u64 calc_reclaim_items_nr(struct btrfs_fs_info *fs_info, 4798 u64 to_reclaim) 4799 { 4800 u64 bytes; 4801 u64 nr; 4802 4803 bytes = btrfs_calc_trans_metadata_size(fs_info, 1); 4804 nr = div64_u64(to_reclaim, bytes); 4805 if (!nr) 4806 nr = 1; 4807 return nr; 4808 } 4809 4810 #define EXTENT_SIZE_PER_ITEM SZ_256K 4811 4812 /* 4813 * shrink metadata reservation for delalloc 4814 */ 4815 static void shrink_delalloc(struct btrfs_fs_info *fs_info, u64 to_reclaim, 4816 u64 orig, bool wait_ordered) 4817 { 4818 struct btrfs_space_info *space_info; 4819 struct btrfs_trans_handle *trans; 4820 u64 delalloc_bytes; 4821 u64 max_reclaim; 4822 u64 items; 4823 long time_left; 4824 unsigned long nr_pages; 4825 int loops; 4826 enum btrfs_reserve_flush_enum flush; 4827 4828 /* Calc the number of the pages we need flush for space reservation */ 4829 items = calc_reclaim_items_nr(fs_info, to_reclaim); 4830 to_reclaim = items * EXTENT_SIZE_PER_ITEM; 4831 4832 trans = (struct btrfs_trans_handle *)current->journal_info; 4833 space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA); 4834 4835 delalloc_bytes = percpu_counter_sum_positive( 4836 &fs_info->delalloc_bytes); 4837 if (delalloc_bytes == 0) { 4838 if (trans) 4839 return; 4840 if (wait_ordered) 4841 btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1); 4842 return; 4843 } 4844 4845 loops = 0; 4846 while (delalloc_bytes && loops < 3) { 4847 max_reclaim = min(delalloc_bytes, to_reclaim); 4848 nr_pages = max_reclaim >> PAGE_SHIFT; 4849 btrfs_writeback_inodes_sb_nr(fs_info, nr_pages, items); 4850 /* 4851 * We need to wait for the async pages to actually start before 4852 * we do anything. 4853 */ 4854 max_reclaim = atomic_read(&fs_info->async_delalloc_pages); 4855 if (!max_reclaim) 4856 goto skip_async; 4857 4858 if (max_reclaim <= nr_pages) 4859 max_reclaim = 0; 4860 else 4861 max_reclaim -= nr_pages; 4862 4863 wait_event(fs_info->async_submit_wait, 4864 atomic_read(&fs_info->async_delalloc_pages) <= 4865 (int)max_reclaim); 4866 skip_async: 4867 if (!trans) 4868 flush = BTRFS_RESERVE_FLUSH_ALL; 4869 else 4870 flush = BTRFS_RESERVE_NO_FLUSH; 4871 spin_lock(&space_info->lock); 4872 if (list_empty(&space_info->tickets) && 4873 list_empty(&space_info->priority_tickets)) { 4874 spin_unlock(&space_info->lock); 4875 break; 4876 } 4877 spin_unlock(&space_info->lock); 4878 4879 loops++; 4880 if (wait_ordered && !trans) { 4881 btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1); 4882 } else { 4883 time_left = schedule_timeout_killable(1); 4884 if (time_left) 4885 break; 4886 } 4887 delalloc_bytes = percpu_counter_sum_positive( 4888 &fs_info->delalloc_bytes); 4889 } 4890 } 4891 4892 struct reserve_ticket { 4893 u64 bytes; 4894 int error; 4895 struct list_head list; 4896 wait_queue_head_t wait; 4897 }; 4898 4899 /** 4900 * maybe_commit_transaction - possibly commit the transaction if its ok to 4901 * @root - the root we're allocating for 4902 * @bytes - the number of bytes we want to reserve 4903 * @force - force the commit 4904 * 4905 * This will check to make sure that committing the transaction will actually 4906 * get us somewhere and then commit the transaction if it does. Otherwise it 4907 * will return -ENOSPC. 4908 */ 4909 static int may_commit_transaction(struct btrfs_fs_info *fs_info, 4910 struct btrfs_space_info *space_info) 4911 { 4912 struct reserve_ticket *ticket = NULL; 4913 struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_block_rsv; 4914 struct btrfs_trans_handle *trans; 4915 u64 bytes; 4916 4917 trans = (struct btrfs_trans_handle *)current->journal_info; 4918 if (trans) 4919 return -EAGAIN; 4920 4921 spin_lock(&space_info->lock); 4922 if (!list_empty(&space_info->priority_tickets)) 4923 ticket = list_first_entry(&space_info->priority_tickets, 4924 struct reserve_ticket, list); 4925 else if (!list_empty(&space_info->tickets)) 4926 ticket = list_first_entry(&space_info->tickets, 4927 struct reserve_ticket, list); 4928 bytes = (ticket) ? ticket->bytes : 0; 4929 spin_unlock(&space_info->lock); 4930 4931 if (!bytes) 4932 return 0; 4933 4934 /* See if there is enough pinned space to make this reservation */ 4935 if (percpu_counter_compare(&space_info->total_bytes_pinned, 4936 bytes) >= 0) 4937 goto commit; 4938 4939 /* 4940 * See if there is some space in the delayed insertion reservation for 4941 * this reservation. 4942 */ 4943 if (space_info != delayed_rsv->space_info) 4944 return -ENOSPC; 4945 4946 spin_lock(&delayed_rsv->lock); 4947 if (delayed_rsv->size > bytes) 4948 bytes = 0; 4949 else 4950 bytes -= delayed_rsv->size; 4951 spin_unlock(&delayed_rsv->lock); 4952 4953 if (percpu_counter_compare(&space_info->total_bytes_pinned, 4954 bytes) < 0) { 4955 return -ENOSPC; 4956 } 4957 4958 commit: 4959 trans = btrfs_join_transaction(fs_info->extent_root); 4960 if (IS_ERR(trans)) 4961 return -ENOSPC; 4962 4963 return btrfs_commit_transaction(trans); 4964 } 4965 4966 /* 4967 * Try to flush some data based on policy set by @state. This is only advisory 4968 * and may fail for various reasons. The caller is supposed to examine the 4969 * state of @space_info to detect the outcome. 4970 */ 4971 static void flush_space(struct btrfs_fs_info *fs_info, 4972 struct btrfs_space_info *space_info, u64 num_bytes, 4973 int state) 4974 { 4975 struct btrfs_root *root = fs_info->extent_root; 4976 struct btrfs_trans_handle *trans; 4977 int nr; 4978 int ret = 0; 4979 4980 switch (state) { 4981 case FLUSH_DELAYED_ITEMS_NR: 4982 case FLUSH_DELAYED_ITEMS: 4983 if (state == FLUSH_DELAYED_ITEMS_NR) 4984 nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2; 4985 else 4986 nr = -1; 4987 4988 trans = btrfs_join_transaction(root); 4989 if (IS_ERR(trans)) { 4990 ret = PTR_ERR(trans); 4991 break; 4992 } 4993 ret = btrfs_run_delayed_items_nr(trans, fs_info, nr); 4994 btrfs_end_transaction(trans); 4995 break; 4996 case FLUSH_DELALLOC: 4997 case FLUSH_DELALLOC_WAIT: 4998 shrink_delalloc(fs_info, num_bytes * 2, num_bytes, 4999 state == FLUSH_DELALLOC_WAIT); 5000 break; 5001 case ALLOC_CHUNK: 5002 trans = btrfs_join_transaction(root); 5003 if (IS_ERR(trans)) { 5004 ret = PTR_ERR(trans); 5005 break; 5006 } 5007 ret = do_chunk_alloc(trans, fs_info, 5008 btrfs_metadata_alloc_profile(fs_info), 5009 CHUNK_ALLOC_NO_FORCE); 5010 btrfs_end_transaction(trans); 5011 if (ret > 0 || ret == -ENOSPC) 5012 ret = 0; 5013 break; 5014 case COMMIT_TRANS: 5015 ret = may_commit_transaction(fs_info, space_info); 5016 break; 5017 default: 5018 ret = -ENOSPC; 5019 break; 5020 } 5021 5022 trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state, 5023 ret); 5024 return; 5025 } 5026 5027 static inline u64 5028 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info, 5029 struct btrfs_space_info *space_info, 5030 bool system_chunk) 5031 { 5032 struct reserve_ticket *ticket; 5033 u64 used; 5034 u64 expected; 5035 u64 to_reclaim = 0; 5036 5037 list_for_each_entry(ticket, &space_info->tickets, list) 5038 to_reclaim += ticket->bytes; 5039 list_for_each_entry(ticket, &space_info->priority_tickets, list) 5040 to_reclaim += ticket->bytes; 5041 if (to_reclaim) 5042 return to_reclaim; 5043 5044 to_reclaim = min_t(u64, num_online_cpus() * SZ_1M, SZ_16M); 5045 if (can_overcommit(fs_info, space_info, to_reclaim, 5046 BTRFS_RESERVE_FLUSH_ALL, system_chunk)) 5047 return 0; 5048 5049 used = btrfs_space_info_used(space_info, true); 5050 5051 if (can_overcommit(fs_info, space_info, SZ_1M, 5052 BTRFS_RESERVE_FLUSH_ALL, system_chunk)) 5053 expected = div_factor_fine(space_info->total_bytes, 95); 5054 else 5055 expected = div_factor_fine(space_info->total_bytes, 90); 5056 5057 if (used > expected) 5058 to_reclaim = used - expected; 5059 else 5060 to_reclaim = 0; 5061 to_reclaim = min(to_reclaim, space_info->bytes_may_use + 5062 space_info->bytes_reserved); 5063 return to_reclaim; 5064 } 5065 5066 static inline int need_do_async_reclaim(struct btrfs_fs_info *fs_info, 5067 struct btrfs_space_info *space_info, 5068 u64 used, bool system_chunk) 5069 { 5070 u64 thresh = div_factor_fine(space_info->total_bytes, 98); 5071 5072 /* If we're just plain full then async reclaim just slows us down. */ 5073 if ((space_info->bytes_used + space_info->bytes_reserved) >= thresh) 5074 return 0; 5075 5076 if (!btrfs_calc_reclaim_metadata_size(fs_info, space_info, 5077 system_chunk)) 5078 return 0; 5079 5080 return (used >= thresh && !btrfs_fs_closing(fs_info) && 5081 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state)); 5082 } 5083 5084 static void wake_all_tickets(struct list_head *head) 5085 { 5086 struct reserve_ticket *ticket; 5087 5088 while (!list_empty(head)) { 5089 ticket = list_first_entry(head, struct reserve_ticket, list); 5090 list_del_init(&ticket->list); 5091 ticket->error = -ENOSPC; 5092 wake_up(&ticket->wait); 5093 } 5094 } 5095 5096 /* 5097 * This is for normal flushers, we can wait all goddamned day if we want to. We 5098 * will loop and continuously try to flush as long as we are making progress. 5099 * We count progress as clearing off tickets each time we have to loop. 5100 */ 5101 static void btrfs_async_reclaim_metadata_space(struct work_struct *work) 5102 { 5103 struct btrfs_fs_info *fs_info; 5104 struct btrfs_space_info *space_info; 5105 u64 to_reclaim; 5106 int flush_state; 5107 int commit_cycles = 0; 5108 u64 last_tickets_id; 5109 5110 fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work); 5111 space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA); 5112 5113 spin_lock(&space_info->lock); 5114 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info, 5115 false); 5116 if (!to_reclaim) { 5117 space_info->flush = 0; 5118 spin_unlock(&space_info->lock); 5119 return; 5120 } 5121 last_tickets_id = space_info->tickets_id; 5122 spin_unlock(&space_info->lock); 5123 5124 flush_state = FLUSH_DELAYED_ITEMS_NR; 5125 do { 5126 flush_space(fs_info, space_info, to_reclaim, flush_state); 5127 spin_lock(&space_info->lock); 5128 if (list_empty(&space_info->tickets)) { 5129 space_info->flush = 0; 5130 spin_unlock(&space_info->lock); 5131 return; 5132 } 5133 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, 5134 space_info, 5135 false); 5136 if (last_tickets_id == space_info->tickets_id) { 5137 flush_state++; 5138 } else { 5139 last_tickets_id = space_info->tickets_id; 5140 flush_state = FLUSH_DELAYED_ITEMS_NR; 5141 if (commit_cycles) 5142 commit_cycles--; 5143 } 5144 5145 if (flush_state > COMMIT_TRANS) { 5146 commit_cycles++; 5147 if (commit_cycles > 2) { 5148 wake_all_tickets(&space_info->tickets); 5149 space_info->flush = 0; 5150 } else { 5151 flush_state = FLUSH_DELAYED_ITEMS_NR; 5152 } 5153 } 5154 spin_unlock(&space_info->lock); 5155 } while (flush_state <= COMMIT_TRANS); 5156 } 5157 5158 void btrfs_init_async_reclaim_work(struct work_struct *work) 5159 { 5160 INIT_WORK(work, btrfs_async_reclaim_metadata_space); 5161 } 5162 5163 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info, 5164 struct btrfs_space_info *space_info, 5165 struct reserve_ticket *ticket) 5166 { 5167 u64 to_reclaim; 5168 int flush_state = FLUSH_DELAYED_ITEMS_NR; 5169 5170 spin_lock(&space_info->lock); 5171 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info, 5172 false); 5173 if (!to_reclaim) { 5174 spin_unlock(&space_info->lock); 5175 return; 5176 } 5177 spin_unlock(&space_info->lock); 5178 5179 do { 5180 flush_space(fs_info, space_info, to_reclaim, flush_state); 5181 flush_state++; 5182 spin_lock(&space_info->lock); 5183 if (ticket->bytes == 0) { 5184 spin_unlock(&space_info->lock); 5185 return; 5186 } 5187 spin_unlock(&space_info->lock); 5188 5189 /* 5190 * Priority flushers can't wait on delalloc without 5191 * deadlocking. 5192 */ 5193 if (flush_state == FLUSH_DELALLOC || 5194 flush_state == FLUSH_DELALLOC_WAIT) 5195 flush_state = ALLOC_CHUNK; 5196 } while (flush_state < COMMIT_TRANS); 5197 } 5198 5199 static int wait_reserve_ticket(struct btrfs_fs_info *fs_info, 5200 struct btrfs_space_info *space_info, 5201 struct reserve_ticket *ticket, u64 orig_bytes) 5202 5203 { 5204 DEFINE_WAIT(wait); 5205 int ret = 0; 5206 5207 spin_lock(&space_info->lock); 5208 while (ticket->bytes > 0 && ticket->error == 0) { 5209 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE); 5210 if (ret) { 5211 ret = -EINTR; 5212 break; 5213 } 5214 spin_unlock(&space_info->lock); 5215 5216 schedule(); 5217 5218 finish_wait(&ticket->wait, &wait); 5219 spin_lock(&space_info->lock); 5220 } 5221 if (!ret) 5222 ret = ticket->error; 5223 if (!list_empty(&ticket->list)) 5224 list_del_init(&ticket->list); 5225 if (ticket->bytes && ticket->bytes < orig_bytes) { 5226 u64 num_bytes = orig_bytes - ticket->bytes; 5227 space_info->bytes_may_use -= num_bytes; 5228 trace_btrfs_space_reservation(fs_info, "space_info", 5229 space_info->flags, num_bytes, 0); 5230 } 5231 spin_unlock(&space_info->lock); 5232 5233 return ret; 5234 } 5235 5236 /** 5237 * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space 5238 * @root - the root we're allocating for 5239 * @space_info - the space info we want to allocate from 5240 * @orig_bytes - the number of bytes we want 5241 * @flush - whether or not we can flush to make our reservation 5242 * 5243 * This will reserve orig_bytes number of bytes from the space info associated 5244 * with the block_rsv. If there is not enough space it will make an attempt to 5245 * flush out space to make room. It will do this by flushing delalloc if 5246 * possible or committing the transaction. If flush is 0 then no attempts to 5247 * regain reservations will be made and this will fail if there is not enough 5248 * space already. 5249 */ 5250 static int __reserve_metadata_bytes(struct btrfs_fs_info *fs_info, 5251 struct btrfs_space_info *space_info, 5252 u64 orig_bytes, 5253 enum btrfs_reserve_flush_enum flush, 5254 bool system_chunk) 5255 { 5256 struct reserve_ticket ticket; 5257 u64 used; 5258 int ret = 0; 5259 5260 ASSERT(orig_bytes); 5261 ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL); 5262 5263 spin_lock(&space_info->lock); 5264 ret = -ENOSPC; 5265 used = btrfs_space_info_used(space_info, true); 5266 5267 /* 5268 * If we have enough space then hooray, make our reservation and carry 5269 * on. If not see if we can overcommit, and if we can, hooray carry on. 5270 * If not things get more complicated. 5271 */ 5272 if (used + orig_bytes <= space_info->total_bytes) { 5273 space_info->bytes_may_use += orig_bytes; 5274 trace_btrfs_space_reservation(fs_info, "space_info", 5275 space_info->flags, orig_bytes, 1); 5276 ret = 0; 5277 } else if (can_overcommit(fs_info, space_info, orig_bytes, flush, 5278 system_chunk)) { 5279 space_info->bytes_may_use += orig_bytes; 5280 trace_btrfs_space_reservation(fs_info, "space_info", 5281 space_info->flags, orig_bytes, 1); 5282 ret = 0; 5283 } 5284 5285 /* 5286 * If we couldn't make a reservation then setup our reservation ticket 5287 * and kick the async worker if it's not already running. 5288 * 5289 * If we are a priority flusher then we just need to add our ticket to 5290 * the list and we will do our own flushing further down. 5291 */ 5292 if (ret && flush != BTRFS_RESERVE_NO_FLUSH) { 5293 ticket.bytes = orig_bytes; 5294 ticket.error = 0; 5295 init_waitqueue_head(&ticket.wait); 5296 if (flush == BTRFS_RESERVE_FLUSH_ALL) { 5297 list_add_tail(&ticket.list, &space_info->tickets); 5298 if (!space_info->flush) { 5299 space_info->flush = 1; 5300 trace_btrfs_trigger_flush(fs_info, 5301 space_info->flags, 5302 orig_bytes, flush, 5303 "enospc"); 5304 queue_work(system_unbound_wq, 5305 &fs_info->async_reclaim_work); 5306 } 5307 } else { 5308 list_add_tail(&ticket.list, 5309 &space_info->priority_tickets); 5310 } 5311 } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) { 5312 used += orig_bytes; 5313 /* 5314 * We will do the space reservation dance during log replay, 5315 * which means we won't have fs_info->fs_root set, so don't do 5316 * the async reclaim as we will panic. 5317 */ 5318 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) && 5319 need_do_async_reclaim(fs_info, space_info, 5320 used, system_chunk) && 5321 !work_busy(&fs_info->async_reclaim_work)) { 5322 trace_btrfs_trigger_flush(fs_info, space_info->flags, 5323 orig_bytes, flush, "preempt"); 5324 queue_work(system_unbound_wq, 5325 &fs_info->async_reclaim_work); 5326 } 5327 } 5328 spin_unlock(&space_info->lock); 5329 if (!ret || flush == BTRFS_RESERVE_NO_FLUSH) 5330 return ret; 5331 5332 if (flush == BTRFS_RESERVE_FLUSH_ALL) 5333 return wait_reserve_ticket(fs_info, space_info, &ticket, 5334 orig_bytes); 5335 5336 ret = 0; 5337 priority_reclaim_metadata_space(fs_info, space_info, &ticket); 5338 spin_lock(&space_info->lock); 5339 if (ticket.bytes) { 5340 if (ticket.bytes < orig_bytes) { 5341 u64 num_bytes = orig_bytes - ticket.bytes; 5342 space_info->bytes_may_use -= num_bytes; 5343 trace_btrfs_space_reservation(fs_info, "space_info", 5344 space_info->flags, 5345 num_bytes, 0); 5346 5347 } 5348 list_del_init(&ticket.list); 5349 ret = -ENOSPC; 5350 } 5351 spin_unlock(&space_info->lock); 5352 ASSERT(list_empty(&ticket.list)); 5353 return ret; 5354 } 5355 5356 /** 5357 * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space 5358 * @root - the root we're allocating for 5359 * @block_rsv - the block_rsv we're allocating for 5360 * @orig_bytes - the number of bytes we want 5361 * @flush - whether or not we can flush to make our reservation 5362 * 5363 * This will reserve orgi_bytes number of bytes from the space info associated 5364 * with the block_rsv. If there is not enough space it will make an attempt to 5365 * flush out space to make room. It will do this by flushing delalloc if 5366 * possible or committing the transaction. If flush is 0 then no attempts to 5367 * regain reservations will be made and this will fail if there is not enough 5368 * space already. 5369 */ 5370 static int reserve_metadata_bytes(struct btrfs_root *root, 5371 struct btrfs_block_rsv *block_rsv, 5372 u64 orig_bytes, 5373 enum btrfs_reserve_flush_enum flush) 5374 { 5375 struct btrfs_fs_info *fs_info = root->fs_info; 5376 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv; 5377 int ret; 5378 bool system_chunk = (root == fs_info->chunk_root); 5379 5380 ret = __reserve_metadata_bytes(fs_info, block_rsv->space_info, 5381 orig_bytes, flush, system_chunk); 5382 if (ret == -ENOSPC && 5383 unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) { 5384 if (block_rsv != global_rsv && 5385 !block_rsv_use_bytes(global_rsv, orig_bytes)) 5386 ret = 0; 5387 } 5388 if (ret == -ENOSPC) 5389 trace_btrfs_space_reservation(fs_info, "space_info:enospc", 5390 block_rsv->space_info->flags, 5391 orig_bytes, 1); 5392 return ret; 5393 } 5394 5395 static struct btrfs_block_rsv *get_block_rsv( 5396 const struct btrfs_trans_handle *trans, 5397 const struct btrfs_root *root) 5398 { 5399 struct btrfs_fs_info *fs_info = root->fs_info; 5400 struct btrfs_block_rsv *block_rsv = NULL; 5401 5402 if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) || 5403 (root == fs_info->csum_root && trans->adding_csums) || 5404 (root == fs_info->uuid_root)) 5405 block_rsv = trans->block_rsv; 5406 5407 if (!block_rsv) 5408 block_rsv = root->block_rsv; 5409 5410 if (!block_rsv) 5411 block_rsv = &fs_info->empty_block_rsv; 5412 5413 return block_rsv; 5414 } 5415 5416 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv, 5417 u64 num_bytes) 5418 { 5419 int ret = -ENOSPC; 5420 spin_lock(&block_rsv->lock); 5421 if (block_rsv->reserved >= num_bytes) { 5422 block_rsv->reserved -= num_bytes; 5423 if (block_rsv->reserved < block_rsv->size) 5424 block_rsv->full = 0; 5425 ret = 0; 5426 } 5427 spin_unlock(&block_rsv->lock); 5428 return ret; 5429 } 5430 5431 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv, 5432 u64 num_bytes, int update_size) 5433 { 5434 spin_lock(&block_rsv->lock); 5435 block_rsv->reserved += num_bytes; 5436 if (update_size) 5437 block_rsv->size += num_bytes; 5438 else if (block_rsv->reserved >= block_rsv->size) 5439 block_rsv->full = 1; 5440 spin_unlock(&block_rsv->lock); 5441 } 5442 5443 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info, 5444 struct btrfs_block_rsv *dest, u64 num_bytes, 5445 int min_factor) 5446 { 5447 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv; 5448 u64 min_bytes; 5449 5450 if (global_rsv->space_info != dest->space_info) 5451 return -ENOSPC; 5452 5453 spin_lock(&global_rsv->lock); 5454 min_bytes = div_factor(global_rsv->size, min_factor); 5455 if (global_rsv->reserved < min_bytes + num_bytes) { 5456 spin_unlock(&global_rsv->lock); 5457 return -ENOSPC; 5458 } 5459 global_rsv->reserved -= num_bytes; 5460 if (global_rsv->reserved < global_rsv->size) 5461 global_rsv->full = 0; 5462 spin_unlock(&global_rsv->lock); 5463 5464 block_rsv_add_bytes(dest, num_bytes, 1); 5465 return 0; 5466 } 5467 5468 /* 5469 * This is for space we already have accounted in space_info->bytes_may_use, so 5470 * basically when we're returning space from block_rsv's. 5471 */ 5472 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info, 5473 struct btrfs_space_info *space_info, 5474 u64 num_bytes) 5475 { 5476 struct reserve_ticket *ticket; 5477 struct list_head *head; 5478 u64 used; 5479 enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH; 5480 bool check_overcommit = false; 5481 5482 spin_lock(&space_info->lock); 5483 head = &space_info->priority_tickets; 5484 5485 /* 5486 * If we are over our limit then we need to check and see if we can 5487 * overcommit, and if we can't then we just need to free up our space 5488 * and not satisfy any requests. 5489 */ 5490 used = btrfs_space_info_used(space_info, true); 5491 if (used - num_bytes >= space_info->total_bytes) 5492 check_overcommit = true; 5493 again: 5494 while (!list_empty(head) && num_bytes) { 5495 ticket = list_first_entry(head, struct reserve_ticket, 5496 list); 5497 /* 5498 * We use 0 bytes because this space is already reserved, so 5499 * adding the ticket space would be a double count. 5500 */ 5501 if (check_overcommit && 5502 !can_overcommit(fs_info, space_info, 0, flush, false)) 5503 break; 5504 if (num_bytes >= ticket->bytes) { 5505 list_del_init(&ticket->list); 5506 num_bytes -= ticket->bytes; 5507 ticket->bytes = 0; 5508 space_info->tickets_id++; 5509 wake_up(&ticket->wait); 5510 } else { 5511 ticket->bytes -= num_bytes; 5512 num_bytes = 0; 5513 } 5514 } 5515 5516 if (num_bytes && head == &space_info->priority_tickets) { 5517 head = &space_info->tickets; 5518 flush = BTRFS_RESERVE_FLUSH_ALL; 5519 goto again; 5520 } 5521 space_info->bytes_may_use -= num_bytes; 5522 trace_btrfs_space_reservation(fs_info, "space_info", 5523 space_info->flags, num_bytes, 0); 5524 spin_unlock(&space_info->lock); 5525 } 5526 5527 /* 5528 * This is for newly allocated space that isn't accounted in 5529 * space_info->bytes_may_use yet. So if we allocate a chunk or unpin an extent 5530 * we use this helper. 5531 */ 5532 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info, 5533 struct btrfs_space_info *space_info, 5534 u64 num_bytes) 5535 { 5536 struct reserve_ticket *ticket; 5537 struct list_head *head = &space_info->priority_tickets; 5538 5539 again: 5540 while (!list_empty(head) && num_bytes) { 5541 ticket = list_first_entry(head, struct reserve_ticket, 5542 list); 5543 if (num_bytes >= ticket->bytes) { 5544 trace_btrfs_space_reservation(fs_info, "space_info", 5545 space_info->flags, 5546 ticket->bytes, 1); 5547 list_del_init(&ticket->list); 5548 num_bytes -= ticket->bytes; 5549 space_info->bytes_may_use += ticket->bytes; 5550 ticket->bytes = 0; 5551 space_info->tickets_id++; 5552 wake_up(&ticket->wait); 5553 } else { 5554 trace_btrfs_space_reservation(fs_info, "space_info", 5555 space_info->flags, 5556 num_bytes, 1); 5557 space_info->bytes_may_use += num_bytes; 5558 ticket->bytes -= num_bytes; 5559 num_bytes = 0; 5560 } 5561 } 5562 5563 if (num_bytes && head == &space_info->priority_tickets) { 5564 head = &space_info->tickets; 5565 goto again; 5566 } 5567 } 5568 5569 static u64 block_rsv_release_bytes(struct btrfs_fs_info *fs_info, 5570 struct btrfs_block_rsv *block_rsv, 5571 struct btrfs_block_rsv *dest, u64 num_bytes) 5572 { 5573 struct btrfs_space_info *space_info = block_rsv->space_info; 5574 u64 ret; 5575 5576 spin_lock(&block_rsv->lock); 5577 if (num_bytes == (u64)-1) 5578 num_bytes = block_rsv->size; 5579 block_rsv->size -= num_bytes; 5580 if (block_rsv->reserved >= block_rsv->size) { 5581 num_bytes = block_rsv->reserved - block_rsv->size; 5582 block_rsv->reserved = block_rsv->size; 5583 block_rsv->full = 1; 5584 } else { 5585 num_bytes = 0; 5586 } 5587 spin_unlock(&block_rsv->lock); 5588 5589 ret = num_bytes; 5590 if (num_bytes > 0) { 5591 if (dest) { 5592 spin_lock(&dest->lock); 5593 if (!dest->full) { 5594 u64 bytes_to_add; 5595 5596 bytes_to_add = dest->size - dest->reserved; 5597 bytes_to_add = min(num_bytes, bytes_to_add); 5598 dest->reserved += bytes_to_add; 5599 if (dest->reserved >= dest->size) 5600 dest->full = 1; 5601 num_bytes -= bytes_to_add; 5602 } 5603 spin_unlock(&dest->lock); 5604 } 5605 if (num_bytes) 5606 space_info_add_old_bytes(fs_info, space_info, 5607 num_bytes); 5608 } 5609 return ret; 5610 } 5611 5612 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src, 5613 struct btrfs_block_rsv *dst, u64 num_bytes, 5614 int update_size) 5615 { 5616 int ret; 5617 5618 ret = block_rsv_use_bytes(src, num_bytes); 5619 if (ret) 5620 return ret; 5621 5622 block_rsv_add_bytes(dst, num_bytes, update_size); 5623 return 0; 5624 } 5625 5626 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type) 5627 { 5628 memset(rsv, 0, sizeof(*rsv)); 5629 spin_lock_init(&rsv->lock); 5630 rsv->type = type; 5631 } 5632 5633 void btrfs_init_metadata_block_rsv(struct btrfs_fs_info *fs_info, 5634 struct btrfs_block_rsv *rsv, 5635 unsigned short type) 5636 { 5637 btrfs_init_block_rsv(rsv, type); 5638 rsv->space_info = __find_space_info(fs_info, 5639 BTRFS_BLOCK_GROUP_METADATA); 5640 } 5641 5642 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_fs_info *fs_info, 5643 unsigned short type) 5644 { 5645 struct btrfs_block_rsv *block_rsv; 5646 5647 block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS); 5648 if (!block_rsv) 5649 return NULL; 5650 5651 btrfs_init_metadata_block_rsv(fs_info, block_rsv, type); 5652 return block_rsv; 5653 } 5654 5655 void btrfs_free_block_rsv(struct btrfs_fs_info *fs_info, 5656 struct btrfs_block_rsv *rsv) 5657 { 5658 if (!rsv) 5659 return; 5660 btrfs_block_rsv_release(fs_info, rsv, (u64)-1); 5661 kfree(rsv); 5662 } 5663 5664 void __btrfs_free_block_rsv(struct btrfs_block_rsv *rsv) 5665 { 5666 kfree(rsv); 5667 } 5668 5669 int btrfs_block_rsv_add(struct btrfs_root *root, 5670 struct btrfs_block_rsv *block_rsv, u64 num_bytes, 5671 enum btrfs_reserve_flush_enum flush) 5672 { 5673 int ret; 5674 5675 if (num_bytes == 0) 5676 return 0; 5677 5678 ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush); 5679 if (!ret) { 5680 block_rsv_add_bytes(block_rsv, num_bytes, 1); 5681 return 0; 5682 } 5683 5684 return ret; 5685 } 5686 5687 int btrfs_block_rsv_check(struct btrfs_block_rsv *block_rsv, int min_factor) 5688 { 5689 u64 num_bytes = 0; 5690 int ret = -ENOSPC; 5691 5692 if (!block_rsv) 5693 return 0; 5694 5695 spin_lock(&block_rsv->lock); 5696 num_bytes = div_factor(block_rsv->size, min_factor); 5697 if (block_rsv->reserved >= num_bytes) 5698 ret = 0; 5699 spin_unlock(&block_rsv->lock); 5700 5701 return ret; 5702 } 5703 5704 int btrfs_block_rsv_refill(struct btrfs_root *root, 5705 struct btrfs_block_rsv *block_rsv, u64 min_reserved, 5706 enum btrfs_reserve_flush_enum flush) 5707 { 5708 u64 num_bytes = 0; 5709 int ret = -ENOSPC; 5710 5711 if (!block_rsv) 5712 return 0; 5713 5714 spin_lock(&block_rsv->lock); 5715 num_bytes = min_reserved; 5716 if (block_rsv->reserved >= num_bytes) 5717 ret = 0; 5718 else 5719 num_bytes -= block_rsv->reserved; 5720 spin_unlock(&block_rsv->lock); 5721 5722 if (!ret) 5723 return 0; 5724 5725 ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush); 5726 if (!ret) { 5727 block_rsv_add_bytes(block_rsv, num_bytes, 0); 5728 return 0; 5729 } 5730 5731 return ret; 5732 } 5733 5734 /** 5735 * btrfs_inode_rsv_refill - refill the inode block rsv. 5736 * @inode - the inode we are refilling. 5737 * @flush - the flusing restriction. 5738 * 5739 * Essentially the same as btrfs_block_rsv_refill, except it uses the 5740 * block_rsv->size as the minimum size. We'll either refill the missing amount 5741 * or return if we already have enough space. This will also handle the resreve 5742 * tracepoint for the reserved amount. 5743 */ 5744 static int btrfs_inode_rsv_refill(struct btrfs_inode *inode, 5745 enum btrfs_reserve_flush_enum flush) 5746 { 5747 struct btrfs_root *root = inode->root; 5748 struct btrfs_block_rsv *block_rsv = &inode->block_rsv; 5749 u64 num_bytes = 0; 5750 int ret = -ENOSPC; 5751 5752 spin_lock(&block_rsv->lock); 5753 if (block_rsv->reserved < block_rsv->size) 5754 num_bytes = block_rsv->size - block_rsv->reserved; 5755 spin_unlock(&block_rsv->lock); 5756 5757 if (num_bytes == 0) 5758 return 0; 5759 5760 ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush); 5761 if (!ret) { 5762 block_rsv_add_bytes(block_rsv, num_bytes, 0); 5763 trace_btrfs_space_reservation(root->fs_info, "delalloc", 5764 btrfs_ino(inode), num_bytes, 1); 5765 } 5766 return ret; 5767 } 5768 5769 /** 5770 * btrfs_inode_rsv_release - release any excessive reservation. 5771 * @inode - the inode we need to release from. 5772 * 5773 * This is the same as btrfs_block_rsv_release, except that it handles the 5774 * tracepoint for the reservation. 5775 */ 5776 static void btrfs_inode_rsv_release(struct btrfs_inode *inode) 5777 { 5778 struct btrfs_fs_info *fs_info = inode->root->fs_info; 5779 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv; 5780 struct btrfs_block_rsv *block_rsv = &inode->block_rsv; 5781 u64 released = 0; 5782 5783 /* 5784 * Since we statically set the block_rsv->size we just want to say we 5785 * are releasing 0 bytes, and then we'll just get the reservation over 5786 * the size free'd. 5787 */ 5788 released = block_rsv_release_bytes(fs_info, block_rsv, global_rsv, 0); 5789 if (released > 0) 5790 trace_btrfs_space_reservation(fs_info, "delalloc", 5791 btrfs_ino(inode), released, 0); 5792 } 5793 5794 void btrfs_block_rsv_release(struct btrfs_fs_info *fs_info, 5795 struct btrfs_block_rsv *block_rsv, 5796 u64 num_bytes) 5797 { 5798 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv; 5799 5800 if (global_rsv == block_rsv || 5801 block_rsv->space_info != global_rsv->space_info) 5802 global_rsv = NULL; 5803 block_rsv_release_bytes(fs_info, block_rsv, global_rsv, num_bytes); 5804 } 5805 5806 static void update_global_block_rsv(struct btrfs_fs_info *fs_info) 5807 { 5808 struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv; 5809 struct btrfs_space_info *sinfo = block_rsv->space_info; 5810 u64 num_bytes; 5811 5812 /* 5813 * The global block rsv is based on the size of the extent tree, the 5814 * checksum tree and the root tree. If the fs is empty we want to set 5815 * it to a minimal amount for safety. 5816 */ 5817 num_bytes = btrfs_root_used(&fs_info->extent_root->root_item) + 5818 btrfs_root_used(&fs_info->csum_root->root_item) + 5819 btrfs_root_used(&fs_info->tree_root->root_item); 5820 num_bytes = max_t(u64, num_bytes, SZ_16M); 5821 5822 spin_lock(&sinfo->lock); 5823 spin_lock(&block_rsv->lock); 5824 5825 block_rsv->size = min_t(u64, num_bytes, SZ_512M); 5826 5827 if (block_rsv->reserved < block_rsv->size) { 5828 num_bytes = btrfs_space_info_used(sinfo, true); 5829 if (sinfo->total_bytes > num_bytes) { 5830 num_bytes = sinfo->total_bytes - num_bytes; 5831 num_bytes = min(num_bytes, 5832 block_rsv->size - block_rsv->reserved); 5833 block_rsv->reserved += num_bytes; 5834 sinfo->bytes_may_use += num_bytes; 5835 trace_btrfs_space_reservation(fs_info, "space_info", 5836 sinfo->flags, num_bytes, 5837 1); 5838 } 5839 } else if (block_rsv->reserved > block_rsv->size) { 5840 num_bytes = block_rsv->reserved - block_rsv->size; 5841 sinfo->bytes_may_use -= num_bytes; 5842 trace_btrfs_space_reservation(fs_info, "space_info", 5843 sinfo->flags, num_bytes, 0); 5844 block_rsv->reserved = block_rsv->size; 5845 } 5846 5847 if (block_rsv->reserved == block_rsv->size) 5848 block_rsv->full = 1; 5849 else 5850 block_rsv->full = 0; 5851 5852 spin_unlock(&block_rsv->lock); 5853 spin_unlock(&sinfo->lock); 5854 } 5855 5856 static void init_global_block_rsv(struct btrfs_fs_info *fs_info) 5857 { 5858 struct btrfs_space_info *space_info; 5859 5860 space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM); 5861 fs_info->chunk_block_rsv.space_info = space_info; 5862 5863 space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA); 5864 fs_info->global_block_rsv.space_info = space_info; 5865 fs_info->trans_block_rsv.space_info = space_info; 5866 fs_info->empty_block_rsv.space_info = space_info; 5867 fs_info->delayed_block_rsv.space_info = space_info; 5868 5869 fs_info->extent_root->block_rsv = &fs_info->global_block_rsv; 5870 fs_info->csum_root->block_rsv = &fs_info->global_block_rsv; 5871 fs_info->dev_root->block_rsv = &fs_info->global_block_rsv; 5872 fs_info->tree_root->block_rsv = &fs_info->global_block_rsv; 5873 if (fs_info->quota_root) 5874 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv; 5875 fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv; 5876 5877 update_global_block_rsv(fs_info); 5878 } 5879 5880 static void release_global_block_rsv(struct btrfs_fs_info *fs_info) 5881 { 5882 block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL, 5883 (u64)-1); 5884 WARN_ON(fs_info->trans_block_rsv.size > 0); 5885 WARN_ON(fs_info->trans_block_rsv.reserved > 0); 5886 WARN_ON(fs_info->chunk_block_rsv.size > 0); 5887 WARN_ON(fs_info->chunk_block_rsv.reserved > 0); 5888 WARN_ON(fs_info->delayed_block_rsv.size > 0); 5889 WARN_ON(fs_info->delayed_block_rsv.reserved > 0); 5890 } 5891 5892 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans, 5893 struct btrfs_fs_info *fs_info) 5894 { 5895 if (!trans->block_rsv) { 5896 ASSERT(!trans->bytes_reserved); 5897 return; 5898 } 5899 5900 if (!trans->bytes_reserved) 5901 return; 5902 5903 ASSERT(trans->block_rsv == &fs_info->trans_block_rsv); 5904 trace_btrfs_space_reservation(fs_info, "transaction", 5905 trans->transid, trans->bytes_reserved, 0); 5906 btrfs_block_rsv_release(fs_info, trans->block_rsv, 5907 trans->bytes_reserved); 5908 trans->bytes_reserved = 0; 5909 } 5910 5911 /* 5912 * To be called after all the new block groups attached to the transaction 5913 * handle have been created (btrfs_create_pending_block_groups()). 5914 */ 5915 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans) 5916 { 5917 struct btrfs_fs_info *fs_info = trans->fs_info; 5918 5919 if (!trans->chunk_bytes_reserved) 5920 return; 5921 5922 WARN_ON_ONCE(!list_empty(&trans->new_bgs)); 5923 5924 block_rsv_release_bytes(fs_info, &fs_info->chunk_block_rsv, NULL, 5925 trans->chunk_bytes_reserved); 5926 trans->chunk_bytes_reserved = 0; 5927 } 5928 5929 /* Can only return 0 or -ENOSPC */ 5930 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans, 5931 struct btrfs_inode *inode) 5932 { 5933 struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb); 5934 struct btrfs_root *root = inode->root; 5935 /* 5936 * We always use trans->block_rsv here as we will have reserved space 5937 * for our orphan when starting the transaction, using get_block_rsv() 5938 * here will sometimes make us choose the wrong block rsv as we could be 5939 * doing a reloc inode for a non refcounted root. 5940 */ 5941 struct btrfs_block_rsv *src_rsv = trans->block_rsv; 5942 struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv; 5943 5944 /* 5945 * We need to hold space in order to delete our orphan item once we've 5946 * added it, so this takes the reservation so we can release it later 5947 * when we are truly done with the orphan item. 5948 */ 5949 u64 num_bytes = btrfs_calc_trans_metadata_size(fs_info, 1); 5950 5951 trace_btrfs_space_reservation(fs_info, "orphan", btrfs_ino(inode), 5952 num_bytes, 1); 5953 return btrfs_block_rsv_migrate(src_rsv, dst_rsv, num_bytes, 1); 5954 } 5955 5956 void btrfs_orphan_release_metadata(struct btrfs_inode *inode) 5957 { 5958 struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb); 5959 struct btrfs_root *root = inode->root; 5960 u64 num_bytes = btrfs_calc_trans_metadata_size(fs_info, 1); 5961 5962 trace_btrfs_space_reservation(fs_info, "orphan", btrfs_ino(inode), 5963 num_bytes, 0); 5964 btrfs_block_rsv_release(fs_info, root->orphan_block_rsv, num_bytes); 5965 } 5966 5967 /* 5968 * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation 5969 * root: the root of the parent directory 5970 * rsv: block reservation 5971 * items: the number of items that we need do reservation 5972 * qgroup_reserved: used to return the reserved size in qgroup 5973 * 5974 * This function is used to reserve the space for snapshot/subvolume 5975 * creation and deletion. Those operations are different with the 5976 * common file/directory operations, they change two fs/file trees 5977 * and root tree, the number of items that the qgroup reserves is 5978 * different with the free space reservation. So we can not use 5979 * the space reservation mechanism in start_transaction(). 5980 */ 5981 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root, 5982 struct btrfs_block_rsv *rsv, 5983 int items, 5984 u64 *qgroup_reserved, 5985 bool use_global_rsv) 5986 { 5987 u64 num_bytes; 5988 int ret; 5989 struct btrfs_fs_info *fs_info = root->fs_info; 5990 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv; 5991 5992 if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)) { 5993 /* One for parent inode, two for dir entries */ 5994 num_bytes = 3 * fs_info->nodesize; 5995 ret = btrfs_qgroup_reserve_meta(root, num_bytes, true); 5996 if (ret) 5997 return ret; 5998 } else { 5999 num_bytes = 0; 6000 } 6001 6002 *qgroup_reserved = num_bytes; 6003 6004 num_bytes = btrfs_calc_trans_metadata_size(fs_info, items); 6005 rsv->space_info = __find_space_info(fs_info, 6006 BTRFS_BLOCK_GROUP_METADATA); 6007 ret = btrfs_block_rsv_add(root, rsv, num_bytes, 6008 BTRFS_RESERVE_FLUSH_ALL); 6009 6010 if (ret == -ENOSPC && use_global_rsv) 6011 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes, 1); 6012 6013 if (ret && *qgroup_reserved) 6014 btrfs_qgroup_free_meta(root, *qgroup_reserved); 6015 6016 return ret; 6017 } 6018 6019 void btrfs_subvolume_release_metadata(struct btrfs_fs_info *fs_info, 6020 struct btrfs_block_rsv *rsv) 6021 { 6022 btrfs_block_rsv_release(fs_info, rsv, (u64)-1); 6023 } 6024 6025 static void btrfs_calculate_inode_block_rsv_size(struct btrfs_fs_info *fs_info, 6026 struct btrfs_inode *inode) 6027 { 6028 struct btrfs_block_rsv *block_rsv = &inode->block_rsv; 6029 u64 reserve_size = 0; 6030 u64 csum_leaves; 6031 unsigned outstanding_extents; 6032 6033 lockdep_assert_held(&inode->lock); 6034 outstanding_extents = inode->outstanding_extents; 6035 if (outstanding_extents) 6036 reserve_size = btrfs_calc_trans_metadata_size(fs_info, 6037 outstanding_extents + 1); 6038 csum_leaves = btrfs_csum_bytes_to_leaves(fs_info, 6039 inode->csum_bytes); 6040 reserve_size += btrfs_calc_trans_metadata_size(fs_info, 6041 csum_leaves); 6042 6043 spin_lock(&block_rsv->lock); 6044 block_rsv->size = reserve_size; 6045 spin_unlock(&block_rsv->lock); 6046 } 6047 6048 int btrfs_delalloc_reserve_metadata(struct btrfs_inode *inode, u64 num_bytes) 6049 { 6050 struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb); 6051 struct btrfs_root *root = inode->root; 6052 unsigned nr_extents; 6053 enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL; 6054 int ret = 0; 6055 bool delalloc_lock = true; 6056 6057 /* If we are a free space inode we need to not flush since we will be in 6058 * the middle of a transaction commit. We also don't need the delalloc 6059 * mutex since we won't race with anybody. We need this mostly to make 6060 * lockdep shut its filthy mouth. 6061 * 6062 * If we have a transaction open (can happen if we call truncate_block 6063 * from truncate), then we need FLUSH_LIMIT so we don't deadlock. 6064 */ 6065 if (btrfs_is_free_space_inode(inode)) { 6066 flush = BTRFS_RESERVE_NO_FLUSH; 6067 delalloc_lock = false; 6068 } else if (current->journal_info) { 6069 flush = BTRFS_RESERVE_FLUSH_LIMIT; 6070 } 6071 6072 if (flush != BTRFS_RESERVE_NO_FLUSH && 6073 btrfs_transaction_in_commit(fs_info)) 6074 schedule_timeout(1); 6075 6076 if (delalloc_lock) 6077 mutex_lock(&inode->delalloc_mutex); 6078 6079 num_bytes = ALIGN(num_bytes, fs_info->sectorsize); 6080 6081 /* Add our new extents and calculate the new rsv size. */ 6082 spin_lock(&inode->lock); 6083 nr_extents = count_max_extents(num_bytes); 6084 btrfs_mod_outstanding_extents(inode, nr_extents); 6085 inode->csum_bytes += num_bytes; 6086 btrfs_calculate_inode_block_rsv_size(fs_info, inode); 6087 spin_unlock(&inode->lock); 6088 6089 if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)) { 6090 ret = btrfs_qgroup_reserve_meta(root, 6091 nr_extents * fs_info->nodesize, true); 6092 if (ret) 6093 goto out_fail; 6094 } 6095 6096 ret = btrfs_inode_rsv_refill(inode, flush); 6097 if (unlikely(ret)) { 6098 btrfs_qgroup_free_meta(root, 6099 nr_extents * fs_info->nodesize); 6100 goto out_fail; 6101 } 6102 6103 if (delalloc_lock) 6104 mutex_unlock(&inode->delalloc_mutex); 6105 return 0; 6106 6107 out_fail: 6108 spin_lock(&inode->lock); 6109 nr_extents = count_max_extents(num_bytes); 6110 btrfs_mod_outstanding_extents(inode, -nr_extents); 6111 inode->csum_bytes -= num_bytes; 6112 btrfs_calculate_inode_block_rsv_size(fs_info, inode); 6113 spin_unlock(&inode->lock); 6114 6115 btrfs_inode_rsv_release(inode); 6116 if (delalloc_lock) 6117 mutex_unlock(&inode->delalloc_mutex); 6118 return ret; 6119 } 6120 6121 /** 6122 * btrfs_delalloc_release_metadata - release a metadata reservation for an inode 6123 * @inode: the inode to release the reservation for. 6124 * @num_bytes: the number of bytes we are releasing. 6125 * 6126 * This will release the metadata reservation for an inode. This can be called 6127 * once we complete IO for a given set of bytes to release their metadata 6128 * reservations, or on error for the same reason. 6129 */ 6130 void btrfs_delalloc_release_metadata(struct btrfs_inode *inode, u64 num_bytes) 6131 { 6132 struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb); 6133 6134 num_bytes = ALIGN(num_bytes, fs_info->sectorsize); 6135 spin_lock(&inode->lock); 6136 inode->csum_bytes -= num_bytes; 6137 btrfs_calculate_inode_block_rsv_size(fs_info, inode); 6138 spin_unlock(&inode->lock); 6139 6140 if (btrfs_is_testing(fs_info)) 6141 return; 6142 6143 btrfs_inode_rsv_release(inode); 6144 } 6145 6146 /** 6147 * btrfs_delalloc_release_extents - release our outstanding_extents 6148 * @inode: the inode to balance the reservation for. 6149 * @num_bytes: the number of bytes we originally reserved with 6150 * 6151 * When we reserve space we increase outstanding_extents for the extents we may 6152 * add. Once we've set the range as delalloc or created our ordered extents we 6153 * have outstanding_extents to track the real usage, so we use this to free our 6154 * temporarily tracked outstanding_extents. This _must_ be used in conjunction 6155 * with btrfs_delalloc_reserve_metadata. 6156 */ 6157 void btrfs_delalloc_release_extents(struct btrfs_inode *inode, u64 num_bytes) 6158 { 6159 struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb); 6160 unsigned num_extents; 6161 6162 spin_lock(&inode->lock); 6163 num_extents = count_max_extents(num_bytes); 6164 btrfs_mod_outstanding_extents(inode, -num_extents); 6165 btrfs_calculate_inode_block_rsv_size(fs_info, inode); 6166 spin_unlock(&inode->lock); 6167 6168 if (btrfs_is_testing(fs_info)) 6169 return; 6170 6171 btrfs_inode_rsv_release(inode); 6172 } 6173 6174 /** 6175 * btrfs_delalloc_reserve_space - reserve data and metadata space for 6176 * delalloc 6177 * @inode: inode we're writing to 6178 * @start: start range we are writing to 6179 * @len: how long the range we are writing to 6180 * @reserved: mandatory parameter, record actually reserved qgroup ranges of 6181 * current reservation. 6182 * 6183 * This will do the following things 6184 * 6185 * o reserve space in data space info for num bytes 6186 * and reserve precious corresponding qgroup space 6187 * (Done in check_data_free_space) 6188 * 6189 * o reserve space for metadata space, based on the number of outstanding 6190 * extents and how much csums will be needed 6191 * also reserve metadata space in a per root over-reserve method. 6192 * o add to the inodes->delalloc_bytes 6193 * o add it to the fs_info's delalloc inodes list. 6194 * (Above 3 all done in delalloc_reserve_metadata) 6195 * 6196 * Return 0 for success 6197 * Return <0 for error(-ENOSPC or -EQUOT) 6198 */ 6199 int btrfs_delalloc_reserve_space(struct inode *inode, 6200 struct extent_changeset **reserved, u64 start, u64 len) 6201 { 6202 int ret; 6203 6204 ret = btrfs_check_data_free_space(inode, reserved, start, len); 6205 if (ret < 0) 6206 return ret; 6207 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode), len); 6208 if (ret < 0) 6209 btrfs_free_reserved_data_space(inode, *reserved, start, len); 6210 return ret; 6211 } 6212 6213 /** 6214 * btrfs_delalloc_release_space - release data and metadata space for delalloc 6215 * @inode: inode we're releasing space for 6216 * @start: start position of the space already reserved 6217 * @len: the len of the space already reserved 6218 * @release_bytes: the len of the space we consumed or didn't use 6219 * 6220 * This function will release the metadata space that was not used and will 6221 * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes 6222 * list if there are no delalloc bytes left. 6223 * Also it will handle the qgroup reserved space. 6224 */ 6225 void btrfs_delalloc_release_space(struct inode *inode, 6226 struct extent_changeset *reserved, 6227 u64 start, u64 len) 6228 { 6229 btrfs_delalloc_release_metadata(BTRFS_I(inode), len); 6230 btrfs_free_reserved_data_space(inode, reserved, start, len); 6231 } 6232 6233 static int update_block_group(struct btrfs_trans_handle *trans, 6234 struct btrfs_fs_info *info, u64 bytenr, 6235 u64 num_bytes, int alloc) 6236 { 6237 struct btrfs_block_group_cache *cache = NULL; 6238 u64 total = num_bytes; 6239 u64 old_val; 6240 u64 byte_in_group; 6241 int factor; 6242 6243 /* block accounting for super block */ 6244 spin_lock(&info->delalloc_root_lock); 6245 old_val = btrfs_super_bytes_used(info->super_copy); 6246 if (alloc) 6247 old_val += num_bytes; 6248 else 6249 old_val -= num_bytes; 6250 btrfs_set_super_bytes_used(info->super_copy, old_val); 6251 spin_unlock(&info->delalloc_root_lock); 6252 6253 while (total) { 6254 cache = btrfs_lookup_block_group(info, bytenr); 6255 if (!cache) 6256 return -ENOENT; 6257 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP | 6258 BTRFS_BLOCK_GROUP_RAID1 | 6259 BTRFS_BLOCK_GROUP_RAID10)) 6260 factor = 2; 6261 else 6262 factor = 1; 6263 /* 6264 * If this block group has free space cache written out, we 6265 * need to make sure to load it if we are removing space. This 6266 * is because we need the unpinning stage to actually add the 6267 * space back to the block group, otherwise we will leak space. 6268 */ 6269 if (!alloc && cache->cached == BTRFS_CACHE_NO) 6270 cache_block_group(cache, 1); 6271 6272 byte_in_group = bytenr - cache->key.objectid; 6273 WARN_ON(byte_in_group > cache->key.offset); 6274 6275 spin_lock(&cache->space_info->lock); 6276 spin_lock(&cache->lock); 6277 6278 if (btrfs_test_opt(info, SPACE_CACHE) && 6279 cache->disk_cache_state < BTRFS_DC_CLEAR) 6280 cache->disk_cache_state = BTRFS_DC_CLEAR; 6281 6282 old_val = btrfs_block_group_used(&cache->item); 6283 num_bytes = min(total, cache->key.offset - byte_in_group); 6284 if (alloc) { 6285 old_val += num_bytes; 6286 btrfs_set_block_group_used(&cache->item, old_val); 6287 cache->reserved -= num_bytes; 6288 cache->space_info->bytes_reserved -= num_bytes; 6289 cache->space_info->bytes_used += num_bytes; 6290 cache->space_info->disk_used += num_bytes * factor; 6291 spin_unlock(&cache->lock); 6292 spin_unlock(&cache->space_info->lock); 6293 } else { 6294 old_val -= num_bytes; 6295 btrfs_set_block_group_used(&cache->item, old_val); 6296 cache->pinned += num_bytes; 6297 cache->space_info->bytes_pinned += num_bytes; 6298 cache->space_info->bytes_used -= num_bytes; 6299 cache->space_info->disk_used -= num_bytes * factor; 6300 spin_unlock(&cache->lock); 6301 spin_unlock(&cache->space_info->lock); 6302 6303 trace_btrfs_space_reservation(info, "pinned", 6304 cache->space_info->flags, 6305 num_bytes, 1); 6306 percpu_counter_add(&cache->space_info->total_bytes_pinned, 6307 num_bytes); 6308 set_extent_dirty(info->pinned_extents, 6309 bytenr, bytenr + num_bytes - 1, 6310 GFP_NOFS | __GFP_NOFAIL); 6311 } 6312 6313 spin_lock(&trans->transaction->dirty_bgs_lock); 6314 if (list_empty(&cache->dirty_list)) { 6315 list_add_tail(&cache->dirty_list, 6316 &trans->transaction->dirty_bgs); 6317 trans->transaction->num_dirty_bgs++; 6318 btrfs_get_block_group(cache); 6319 } 6320 spin_unlock(&trans->transaction->dirty_bgs_lock); 6321 6322 /* 6323 * No longer have used bytes in this block group, queue it for 6324 * deletion. We do this after adding the block group to the 6325 * dirty list to avoid races between cleaner kthread and space 6326 * cache writeout. 6327 */ 6328 if (!alloc && old_val == 0) { 6329 spin_lock(&info->unused_bgs_lock); 6330 if (list_empty(&cache->bg_list)) { 6331 btrfs_get_block_group(cache); 6332 list_add_tail(&cache->bg_list, 6333 &info->unused_bgs); 6334 } 6335 spin_unlock(&info->unused_bgs_lock); 6336 } 6337 6338 btrfs_put_block_group(cache); 6339 total -= num_bytes; 6340 bytenr += num_bytes; 6341 } 6342 return 0; 6343 } 6344 6345 static u64 first_logical_byte(struct btrfs_fs_info *fs_info, u64 search_start) 6346 { 6347 struct btrfs_block_group_cache *cache; 6348 u64 bytenr; 6349 6350 spin_lock(&fs_info->block_group_cache_lock); 6351 bytenr = fs_info->first_logical_byte; 6352 spin_unlock(&fs_info->block_group_cache_lock); 6353 6354 if (bytenr < (u64)-1) 6355 return bytenr; 6356 6357 cache = btrfs_lookup_first_block_group(fs_info, search_start); 6358 if (!cache) 6359 return 0; 6360 6361 bytenr = cache->key.objectid; 6362 btrfs_put_block_group(cache); 6363 6364 return bytenr; 6365 } 6366 6367 static int pin_down_extent(struct btrfs_fs_info *fs_info, 6368 struct btrfs_block_group_cache *cache, 6369 u64 bytenr, u64 num_bytes, int reserved) 6370 { 6371 spin_lock(&cache->space_info->lock); 6372 spin_lock(&cache->lock); 6373 cache->pinned += num_bytes; 6374 cache->space_info->bytes_pinned += num_bytes; 6375 if (reserved) { 6376 cache->reserved -= num_bytes; 6377 cache->space_info->bytes_reserved -= num_bytes; 6378 } 6379 spin_unlock(&cache->lock); 6380 spin_unlock(&cache->space_info->lock); 6381 6382 trace_btrfs_space_reservation(fs_info, "pinned", 6383 cache->space_info->flags, num_bytes, 1); 6384 percpu_counter_add(&cache->space_info->total_bytes_pinned, num_bytes); 6385 set_extent_dirty(fs_info->pinned_extents, bytenr, 6386 bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL); 6387 return 0; 6388 } 6389 6390 /* 6391 * this function must be called within transaction 6392 */ 6393 int btrfs_pin_extent(struct btrfs_fs_info *fs_info, 6394 u64 bytenr, u64 num_bytes, int reserved) 6395 { 6396 struct btrfs_block_group_cache *cache; 6397 6398 cache = btrfs_lookup_block_group(fs_info, bytenr); 6399 BUG_ON(!cache); /* Logic error */ 6400 6401 pin_down_extent(fs_info, cache, bytenr, num_bytes, reserved); 6402 6403 btrfs_put_block_group(cache); 6404 return 0; 6405 } 6406 6407 /* 6408 * this function must be called within transaction 6409 */ 6410 int btrfs_pin_extent_for_log_replay(struct btrfs_fs_info *fs_info, 6411 u64 bytenr, u64 num_bytes) 6412 { 6413 struct btrfs_block_group_cache *cache; 6414 int ret; 6415 6416 cache = btrfs_lookup_block_group(fs_info, bytenr); 6417 if (!cache) 6418 return -EINVAL; 6419 6420 /* 6421 * pull in the free space cache (if any) so that our pin 6422 * removes the free space from the cache. We have load_only set 6423 * to one because the slow code to read in the free extents does check 6424 * the pinned extents. 6425 */ 6426 cache_block_group(cache, 1); 6427 6428 pin_down_extent(fs_info, cache, bytenr, num_bytes, 0); 6429 6430 /* remove us from the free space cache (if we're there at all) */ 6431 ret = btrfs_remove_free_space(cache, bytenr, num_bytes); 6432 btrfs_put_block_group(cache); 6433 return ret; 6434 } 6435 6436 static int __exclude_logged_extent(struct btrfs_fs_info *fs_info, 6437 u64 start, u64 num_bytes) 6438 { 6439 int ret; 6440 struct btrfs_block_group_cache *block_group; 6441 struct btrfs_caching_control *caching_ctl; 6442 6443 block_group = btrfs_lookup_block_group(fs_info, start); 6444 if (!block_group) 6445 return -EINVAL; 6446 6447 cache_block_group(block_group, 0); 6448 caching_ctl = get_caching_control(block_group); 6449 6450 if (!caching_ctl) { 6451 /* Logic error */ 6452 BUG_ON(!block_group_cache_done(block_group)); 6453 ret = btrfs_remove_free_space(block_group, start, num_bytes); 6454 } else { 6455 mutex_lock(&caching_ctl->mutex); 6456 6457 if (start >= caching_ctl->progress) { 6458 ret = add_excluded_extent(fs_info, start, num_bytes); 6459 } else if (start + num_bytes <= caching_ctl->progress) { 6460 ret = btrfs_remove_free_space(block_group, 6461 start, num_bytes); 6462 } else { 6463 num_bytes = caching_ctl->progress - start; 6464 ret = btrfs_remove_free_space(block_group, 6465 start, num_bytes); 6466 if (ret) 6467 goto out_lock; 6468 6469 num_bytes = (start + num_bytes) - 6470 caching_ctl->progress; 6471 start = caching_ctl->progress; 6472 ret = add_excluded_extent(fs_info, start, num_bytes); 6473 } 6474 out_lock: 6475 mutex_unlock(&caching_ctl->mutex); 6476 put_caching_control(caching_ctl); 6477 } 6478 btrfs_put_block_group(block_group); 6479 return ret; 6480 } 6481 6482 int btrfs_exclude_logged_extents(struct btrfs_fs_info *fs_info, 6483 struct extent_buffer *eb) 6484 { 6485 struct btrfs_file_extent_item *item; 6486 struct btrfs_key key; 6487 int found_type; 6488 int i; 6489 6490 if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS)) 6491 return 0; 6492 6493 for (i = 0; i < btrfs_header_nritems(eb); i++) { 6494 btrfs_item_key_to_cpu(eb, &key, i); 6495 if (key.type != BTRFS_EXTENT_DATA_KEY) 6496 continue; 6497 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item); 6498 found_type = btrfs_file_extent_type(eb, item); 6499 if (found_type == BTRFS_FILE_EXTENT_INLINE) 6500 continue; 6501 if (btrfs_file_extent_disk_bytenr(eb, item) == 0) 6502 continue; 6503 key.objectid = btrfs_file_extent_disk_bytenr(eb, item); 6504 key.offset = btrfs_file_extent_disk_num_bytes(eb, item); 6505 __exclude_logged_extent(fs_info, key.objectid, key.offset); 6506 } 6507 6508 return 0; 6509 } 6510 6511 static void 6512 btrfs_inc_block_group_reservations(struct btrfs_block_group_cache *bg) 6513 { 6514 atomic_inc(&bg->reservations); 6515 } 6516 6517 void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info, 6518 const u64 start) 6519 { 6520 struct btrfs_block_group_cache *bg; 6521 6522 bg = btrfs_lookup_block_group(fs_info, start); 6523 ASSERT(bg); 6524 if (atomic_dec_and_test(&bg->reservations)) 6525 wake_up_atomic_t(&bg->reservations); 6526 btrfs_put_block_group(bg); 6527 } 6528 6529 void btrfs_wait_block_group_reservations(struct btrfs_block_group_cache *bg) 6530 { 6531 struct btrfs_space_info *space_info = bg->space_info; 6532 6533 ASSERT(bg->ro); 6534 6535 if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA)) 6536 return; 6537 6538 /* 6539 * Our block group is read only but before we set it to read only, 6540 * some task might have had allocated an extent from it already, but it 6541 * has not yet created a respective ordered extent (and added it to a 6542 * root's list of ordered extents). 6543 * Therefore wait for any task currently allocating extents, since the 6544 * block group's reservations counter is incremented while a read lock 6545 * on the groups' semaphore is held and decremented after releasing 6546 * the read access on that semaphore and creating the ordered extent. 6547 */ 6548 down_write(&space_info->groups_sem); 6549 up_write(&space_info->groups_sem); 6550 6551 wait_on_atomic_t(&bg->reservations, atomic_t_wait, 6552 TASK_UNINTERRUPTIBLE); 6553 } 6554 6555 /** 6556 * btrfs_add_reserved_bytes - update the block_group and space info counters 6557 * @cache: The cache we are manipulating 6558 * @ram_bytes: The number of bytes of file content, and will be same to 6559 * @num_bytes except for the compress path. 6560 * @num_bytes: The number of bytes in question 6561 * @delalloc: The blocks are allocated for the delalloc write 6562 * 6563 * This is called by the allocator when it reserves space. If this is a 6564 * reservation and the block group has become read only we cannot make the 6565 * reservation and return -EAGAIN, otherwise this function always succeeds. 6566 */ 6567 static int btrfs_add_reserved_bytes(struct btrfs_block_group_cache *cache, 6568 u64 ram_bytes, u64 num_bytes, int delalloc) 6569 { 6570 struct btrfs_space_info *space_info = cache->space_info; 6571 int ret = 0; 6572 6573 spin_lock(&space_info->lock); 6574 spin_lock(&cache->lock); 6575 if (cache->ro) { 6576 ret = -EAGAIN; 6577 } else { 6578 cache->reserved += num_bytes; 6579 space_info->bytes_reserved += num_bytes; 6580 6581 trace_btrfs_space_reservation(cache->fs_info, 6582 "space_info", space_info->flags, 6583 ram_bytes, 0); 6584 space_info->bytes_may_use -= ram_bytes; 6585 if (delalloc) 6586 cache->delalloc_bytes += num_bytes; 6587 } 6588 spin_unlock(&cache->lock); 6589 spin_unlock(&space_info->lock); 6590 return ret; 6591 } 6592 6593 /** 6594 * btrfs_free_reserved_bytes - update the block_group and space info counters 6595 * @cache: The cache we are manipulating 6596 * @num_bytes: The number of bytes in question 6597 * @delalloc: The blocks are allocated for the delalloc write 6598 * 6599 * This is called by somebody who is freeing space that was never actually used 6600 * on disk. For example if you reserve some space for a new leaf in transaction 6601 * A and before transaction A commits you free that leaf, you call this with 6602 * reserve set to 0 in order to clear the reservation. 6603 */ 6604 6605 static int btrfs_free_reserved_bytes(struct btrfs_block_group_cache *cache, 6606 u64 num_bytes, int delalloc) 6607 { 6608 struct btrfs_space_info *space_info = cache->space_info; 6609 int ret = 0; 6610 6611 spin_lock(&space_info->lock); 6612 spin_lock(&cache->lock); 6613 if (cache->ro) 6614 space_info->bytes_readonly += num_bytes; 6615 cache->reserved -= num_bytes; 6616 space_info->bytes_reserved -= num_bytes; 6617 6618 if (delalloc) 6619 cache->delalloc_bytes -= num_bytes; 6620 spin_unlock(&cache->lock); 6621 spin_unlock(&space_info->lock); 6622 return ret; 6623 } 6624 void btrfs_prepare_extent_commit(struct btrfs_fs_info *fs_info) 6625 { 6626 struct btrfs_caching_control *next; 6627 struct btrfs_caching_control *caching_ctl; 6628 struct btrfs_block_group_cache *cache; 6629 6630 down_write(&fs_info->commit_root_sem); 6631 6632 list_for_each_entry_safe(caching_ctl, next, 6633 &fs_info->caching_block_groups, list) { 6634 cache = caching_ctl->block_group; 6635 if (block_group_cache_done(cache)) { 6636 cache->last_byte_to_unpin = (u64)-1; 6637 list_del_init(&caching_ctl->list); 6638 put_caching_control(caching_ctl); 6639 } else { 6640 cache->last_byte_to_unpin = caching_ctl->progress; 6641 } 6642 } 6643 6644 if (fs_info->pinned_extents == &fs_info->freed_extents[0]) 6645 fs_info->pinned_extents = &fs_info->freed_extents[1]; 6646 else 6647 fs_info->pinned_extents = &fs_info->freed_extents[0]; 6648 6649 up_write(&fs_info->commit_root_sem); 6650 6651 update_global_block_rsv(fs_info); 6652 } 6653 6654 /* 6655 * Returns the free cluster for the given space info and sets empty_cluster to 6656 * what it should be based on the mount options. 6657 */ 6658 static struct btrfs_free_cluster * 6659 fetch_cluster_info(struct btrfs_fs_info *fs_info, 6660 struct btrfs_space_info *space_info, u64 *empty_cluster) 6661 { 6662 struct btrfs_free_cluster *ret = NULL; 6663 6664 *empty_cluster = 0; 6665 if (btrfs_mixed_space_info(space_info)) 6666 return ret; 6667 6668 if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) { 6669 ret = &fs_info->meta_alloc_cluster; 6670 if (btrfs_test_opt(fs_info, SSD)) 6671 *empty_cluster = SZ_2M; 6672 else 6673 *empty_cluster = SZ_64K; 6674 } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) && 6675 btrfs_test_opt(fs_info, SSD_SPREAD)) { 6676 *empty_cluster = SZ_2M; 6677 ret = &fs_info->data_alloc_cluster; 6678 } 6679 6680 return ret; 6681 } 6682 6683 static int unpin_extent_range(struct btrfs_fs_info *fs_info, 6684 u64 start, u64 end, 6685 const bool return_free_space) 6686 { 6687 struct btrfs_block_group_cache *cache = NULL; 6688 struct btrfs_space_info *space_info; 6689 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv; 6690 struct btrfs_free_cluster *cluster = NULL; 6691 u64 len; 6692 u64 total_unpinned = 0; 6693 u64 empty_cluster = 0; 6694 bool readonly; 6695 6696 while (start <= end) { 6697 readonly = false; 6698 if (!cache || 6699 start >= cache->key.objectid + cache->key.offset) { 6700 if (cache) 6701 btrfs_put_block_group(cache); 6702 total_unpinned = 0; 6703 cache = btrfs_lookup_block_group(fs_info, start); 6704 BUG_ON(!cache); /* Logic error */ 6705 6706 cluster = fetch_cluster_info(fs_info, 6707 cache->space_info, 6708 &empty_cluster); 6709 empty_cluster <<= 1; 6710 } 6711 6712 len = cache->key.objectid + cache->key.offset - start; 6713 len = min(len, end + 1 - start); 6714 6715 if (start < cache->last_byte_to_unpin) { 6716 len = min(len, cache->last_byte_to_unpin - start); 6717 if (return_free_space) 6718 btrfs_add_free_space(cache, start, len); 6719 } 6720 6721 start += len; 6722 total_unpinned += len; 6723 space_info = cache->space_info; 6724 6725 /* 6726 * If this space cluster has been marked as fragmented and we've 6727 * unpinned enough in this block group to potentially allow a 6728 * cluster to be created inside of it go ahead and clear the 6729 * fragmented check. 6730 */ 6731 if (cluster && cluster->fragmented && 6732 total_unpinned > empty_cluster) { 6733 spin_lock(&cluster->lock); 6734 cluster->fragmented = 0; 6735 spin_unlock(&cluster->lock); 6736 } 6737 6738 spin_lock(&space_info->lock); 6739 spin_lock(&cache->lock); 6740 cache->pinned -= len; 6741 space_info->bytes_pinned -= len; 6742 6743 trace_btrfs_space_reservation(fs_info, "pinned", 6744 space_info->flags, len, 0); 6745 space_info->max_extent_size = 0; 6746 percpu_counter_add(&space_info->total_bytes_pinned, -len); 6747 if (cache->ro) { 6748 space_info->bytes_readonly += len; 6749 readonly = true; 6750 } 6751 spin_unlock(&cache->lock); 6752 if (!readonly && return_free_space && 6753 global_rsv->space_info == space_info) { 6754 u64 to_add = len; 6755 6756 spin_lock(&global_rsv->lock); 6757 if (!global_rsv->full) { 6758 to_add = min(len, global_rsv->size - 6759 global_rsv->reserved); 6760 global_rsv->reserved += to_add; 6761 space_info->bytes_may_use += to_add; 6762 if (global_rsv->reserved >= global_rsv->size) 6763 global_rsv->full = 1; 6764 trace_btrfs_space_reservation(fs_info, 6765 "space_info", 6766 space_info->flags, 6767 to_add, 1); 6768 len -= to_add; 6769 } 6770 spin_unlock(&global_rsv->lock); 6771 /* Add to any tickets we may have */ 6772 if (len) 6773 space_info_add_new_bytes(fs_info, space_info, 6774 len); 6775 } 6776 spin_unlock(&space_info->lock); 6777 } 6778 6779 if (cache) 6780 btrfs_put_block_group(cache); 6781 return 0; 6782 } 6783 6784 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans, 6785 struct btrfs_fs_info *fs_info) 6786 { 6787 struct btrfs_block_group_cache *block_group, *tmp; 6788 struct list_head *deleted_bgs; 6789 struct extent_io_tree *unpin; 6790 u64 start; 6791 u64 end; 6792 int ret; 6793 6794 if (fs_info->pinned_extents == &fs_info->freed_extents[0]) 6795 unpin = &fs_info->freed_extents[1]; 6796 else 6797 unpin = &fs_info->freed_extents[0]; 6798 6799 while (!trans->aborted) { 6800 mutex_lock(&fs_info->unused_bg_unpin_mutex); 6801 ret = find_first_extent_bit(unpin, 0, &start, &end, 6802 EXTENT_DIRTY, NULL); 6803 if (ret) { 6804 mutex_unlock(&fs_info->unused_bg_unpin_mutex); 6805 break; 6806 } 6807 6808 if (btrfs_test_opt(fs_info, DISCARD)) 6809 ret = btrfs_discard_extent(fs_info, start, 6810 end + 1 - start, NULL); 6811 6812 clear_extent_dirty(unpin, start, end); 6813 unpin_extent_range(fs_info, start, end, true); 6814 mutex_unlock(&fs_info->unused_bg_unpin_mutex); 6815 cond_resched(); 6816 } 6817 6818 /* 6819 * Transaction is finished. We don't need the lock anymore. We 6820 * do need to clean up the block groups in case of a transaction 6821 * abort. 6822 */ 6823 deleted_bgs = &trans->transaction->deleted_bgs; 6824 list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) { 6825 u64 trimmed = 0; 6826 6827 ret = -EROFS; 6828 if (!trans->aborted) 6829 ret = btrfs_discard_extent(fs_info, 6830 block_group->key.objectid, 6831 block_group->key.offset, 6832 &trimmed); 6833 6834 list_del_init(&block_group->bg_list); 6835 btrfs_put_block_group_trimming(block_group); 6836 btrfs_put_block_group(block_group); 6837 6838 if (ret) { 6839 const char *errstr = btrfs_decode_error(ret); 6840 btrfs_warn(fs_info, 6841 "discard failed while removing blockgroup: errno=%d %s", 6842 ret, errstr); 6843 } 6844 } 6845 6846 return 0; 6847 } 6848 6849 static int __btrfs_free_extent(struct btrfs_trans_handle *trans, 6850 struct btrfs_fs_info *info, 6851 struct btrfs_delayed_ref_node *node, u64 parent, 6852 u64 root_objectid, u64 owner_objectid, 6853 u64 owner_offset, int refs_to_drop, 6854 struct btrfs_delayed_extent_op *extent_op) 6855 { 6856 struct btrfs_key key; 6857 struct btrfs_path *path; 6858 struct btrfs_root *extent_root = info->extent_root; 6859 struct extent_buffer *leaf; 6860 struct btrfs_extent_item *ei; 6861 struct btrfs_extent_inline_ref *iref; 6862 int ret; 6863 int is_data; 6864 int extent_slot = 0; 6865 int found_extent = 0; 6866 int num_to_del = 1; 6867 u32 item_size; 6868 u64 refs; 6869 u64 bytenr = node->bytenr; 6870 u64 num_bytes = node->num_bytes; 6871 int last_ref = 0; 6872 bool skinny_metadata = btrfs_fs_incompat(info, SKINNY_METADATA); 6873 6874 path = btrfs_alloc_path(); 6875 if (!path) 6876 return -ENOMEM; 6877 6878 path->reada = READA_FORWARD; 6879 path->leave_spinning = 1; 6880 6881 is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID; 6882 BUG_ON(!is_data && refs_to_drop != 1); 6883 6884 if (is_data) 6885 skinny_metadata = false; 6886 6887 ret = lookup_extent_backref(trans, info, path, &iref, 6888 bytenr, num_bytes, parent, 6889 root_objectid, owner_objectid, 6890 owner_offset); 6891 if (ret == 0) { 6892 extent_slot = path->slots[0]; 6893 while (extent_slot >= 0) { 6894 btrfs_item_key_to_cpu(path->nodes[0], &key, 6895 extent_slot); 6896 if (key.objectid != bytenr) 6897 break; 6898 if (key.type == BTRFS_EXTENT_ITEM_KEY && 6899 key.offset == num_bytes) { 6900 found_extent = 1; 6901 break; 6902 } 6903 if (key.type == BTRFS_METADATA_ITEM_KEY && 6904 key.offset == owner_objectid) { 6905 found_extent = 1; 6906 break; 6907 } 6908 if (path->slots[0] - extent_slot > 5) 6909 break; 6910 extent_slot--; 6911 } 6912 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 6913 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot); 6914 if (found_extent && item_size < sizeof(*ei)) 6915 found_extent = 0; 6916 #endif 6917 if (!found_extent) { 6918 BUG_ON(iref); 6919 ret = remove_extent_backref(trans, info, path, NULL, 6920 refs_to_drop, 6921 is_data, &last_ref); 6922 if (ret) { 6923 btrfs_abort_transaction(trans, ret); 6924 goto out; 6925 } 6926 btrfs_release_path(path); 6927 path->leave_spinning = 1; 6928 6929 key.objectid = bytenr; 6930 key.type = BTRFS_EXTENT_ITEM_KEY; 6931 key.offset = num_bytes; 6932 6933 if (!is_data && skinny_metadata) { 6934 key.type = BTRFS_METADATA_ITEM_KEY; 6935 key.offset = owner_objectid; 6936 } 6937 6938 ret = btrfs_search_slot(trans, extent_root, 6939 &key, path, -1, 1); 6940 if (ret > 0 && skinny_metadata && path->slots[0]) { 6941 /* 6942 * Couldn't find our skinny metadata item, 6943 * see if we have ye olde extent item. 6944 */ 6945 path->slots[0]--; 6946 btrfs_item_key_to_cpu(path->nodes[0], &key, 6947 path->slots[0]); 6948 if (key.objectid == bytenr && 6949 key.type == BTRFS_EXTENT_ITEM_KEY && 6950 key.offset == num_bytes) 6951 ret = 0; 6952 } 6953 6954 if (ret > 0 && skinny_metadata) { 6955 skinny_metadata = false; 6956 key.objectid = bytenr; 6957 key.type = BTRFS_EXTENT_ITEM_KEY; 6958 key.offset = num_bytes; 6959 btrfs_release_path(path); 6960 ret = btrfs_search_slot(trans, extent_root, 6961 &key, path, -1, 1); 6962 } 6963 6964 if (ret) { 6965 btrfs_err(info, 6966 "umm, got %d back from search, was looking for %llu", 6967 ret, bytenr); 6968 if (ret > 0) 6969 btrfs_print_leaf(path->nodes[0]); 6970 } 6971 if (ret < 0) { 6972 btrfs_abort_transaction(trans, ret); 6973 goto out; 6974 } 6975 extent_slot = path->slots[0]; 6976 } 6977 } else if (WARN_ON(ret == -ENOENT)) { 6978 btrfs_print_leaf(path->nodes[0]); 6979 btrfs_err(info, 6980 "unable to find ref byte nr %llu parent %llu root %llu owner %llu offset %llu", 6981 bytenr, parent, root_objectid, owner_objectid, 6982 owner_offset); 6983 btrfs_abort_transaction(trans, ret); 6984 goto out; 6985 } else { 6986 btrfs_abort_transaction(trans, ret); 6987 goto out; 6988 } 6989 6990 leaf = path->nodes[0]; 6991 item_size = btrfs_item_size_nr(leaf, extent_slot); 6992 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 6993 if (item_size < sizeof(*ei)) { 6994 BUG_ON(found_extent || extent_slot != path->slots[0]); 6995 ret = convert_extent_item_v0(trans, info, path, owner_objectid, 6996 0); 6997 if (ret < 0) { 6998 btrfs_abort_transaction(trans, ret); 6999 goto out; 7000 } 7001 7002 btrfs_release_path(path); 7003 path->leave_spinning = 1; 7004 7005 key.objectid = bytenr; 7006 key.type = BTRFS_EXTENT_ITEM_KEY; 7007 key.offset = num_bytes; 7008 7009 ret = btrfs_search_slot(trans, extent_root, &key, path, 7010 -1, 1); 7011 if (ret) { 7012 btrfs_err(info, 7013 "umm, got %d back from search, was looking for %llu", 7014 ret, bytenr); 7015 btrfs_print_leaf(path->nodes[0]); 7016 } 7017 if (ret < 0) { 7018 btrfs_abort_transaction(trans, ret); 7019 goto out; 7020 } 7021 7022 extent_slot = path->slots[0]; 7023 leaf = path->nodes[0]; 7024 item_size = btrfs_item_size_nr(leaf, extent_slot); 7025 } 7026 #endif 7027 BUG_ON(item_size < sizeof(*ei)); 7028 ei = btrfs_item_ptr(leaf, extent_slot, 7029 struct btrfs_extent_item); 7030 if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID && 7031 key.type == BTRFS_EXTENT_ITEM_KEY) { 7032 struct btrfs_tree_block_info *bi; 7033 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi)); 7034 bi = (struct btrfs_tree_block_info *)(ei + 1); 7035 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi)); 7036 } 7037 7038 refs = btrfs_extent_refs(leaf, ei); 7039 if (refs < refs_to_drop) { 7040 btrfs_err(info, 7041 "trying to drop %d refs but we only have %Lu for bytenr %Lu", 7042 refs_to_drop, refs, bytenr); 7043 ret = -EINVAL; 7044 btrfs_abort_transaction(trans, ret); 7045 goto out; 7046 } 7047 refs -= refs_to_drop; 7048 7049 if (refs > 0) { 7050 if (extent_op) 7051 __run_delayed_extent_op(extent_op, leaf, ei); 7052 /* 7053 * In the case of inline back ref, reference count will 7054 * be updated by remove_extent_backref 7055 */ 7056 if (iref) { 7057 BUG_ON(!found_extent); 7058 } else { 7059 btrfs_set_extent_refs(leaf, ei, refs); 7060 btrfs_mark_buffer_dirty(leaf); 7061 } 7062 if (found_extent) { 7063 ret = remove_extent_backref(trans, info, path, 7064 iref, refs_to_drop, 7065 is_data, &last_ref); 7066 if (ret) { 7067 btrfs_abort_transaction(trans, ret); 7068 goto out; 7069 } 7070 } 7071 } else { 7072 if (found_extent) { 7073 BUG_ON(is_data && refs_to_drop != 7074 extent_data_ref_count(path, iref)); 7075 if (iref) { 7076 BUG_ON(path->slots[0] != extent_slot); 7077 } else { 7078 BUG_ON(path->slots[0] != extent_slot + 1); 7079 path->slots[0] = extent_slot; 7080 num_to_del = 2; 7081 } 7082 } 7083 7084 last_ref = 1; 7085 ret = btrfs_del_items(trans, extent_root, path, path->slots[0], 7086 num_to_del); 7087 if (ret) { 7088 btrfs_abort_transaction(trans, ret); 7089 goto out; 7090 } 7091 btrfs_release_path(path); 7092 7093 if (is_data) { 7094 ret = btrfs_del_csums(trans, info, bytenr, num_bytes); 7095 if (ret) { 7096 btrfs_abort_transaction(trans, ret); 7097 goto out; 7098 } 7099 } 7100 7101 ret = add_to_free_space_tree(trans, info, bytenr, num_bytes); 7102 if (ret) { 7103 btrfs_abort_transaction(trans, ret); 7104 goto out; 7105 } 7106 7107 ret = update_block_group(trans, info, bytenr, num_bytes, 0); 7108 if (ret) { 7109 btrfs_abort_transaction(trans, ret); 7110 goto out; 7111 } 7112 } 7113 btrfs_release_path(path); 7114 7115 out: 7116 btrfs_free_path(path); 7117 return ret; 7118 } 7119 7120 /* 7121 * when we free an block, it is possible (and likely) that we free the last 7122 * delayed ref for that extent as well. This searches the delayed ref tree for 7123 * a given extent, and if there are no other delayed refs to be processed, it 7124 * removes it from the tree. 7125 */ 7126 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans, 7127 u64 bytenr) 7128 { 7129 struct btrfs_delayed_ref_head *head; 7130 struct btrfs_delayed_ref_root *delayed_refs; 7131 int ret = 0; 7132 7133 delayed_refs = &trans->transaction->delayed_refs; 7134 spin_lock(&delayed_refs->lock); 7135 head = btrfs_find_delayed_ref_head(delayed_refs, bytenr); 7136 if (!head) 7137 goto out_delayed_unlock; 7138 7139 spin_lock(&head->lock); 7140 if (!RB_EMPTY_ROOT(&head->ref_tree)) 7141 goto out; 7142 7143 if (head->extent_op) { 7144 if (!head->must_insert_reserved) 7145 goto out; 7146 btrfs_free_delayed_extent_op(head->extent_op); 7147 head->extent_op = NULL; 7148 } 7149 7150 /* 7151 * waiting for the lock here would deadlock. If someone else has it 7152 * locked they are already in the process of dropping it anyway 7153 */ 7154 if (!mutex_trylock(&head->mutex)) 7155 goto out; 7156 7157 /* 7158 * at this point we have a head with no other entries. Go 7159 * ahead and process it. 7160 */ 7161 rb_erase(&head->href_node, &delayed_refs->href_root); 7162 RB_CLEAR_NODE(&head->href_node); 7163 atomic_dec(&delayed_refs->num_entries); 7164 7165 /* 7166 * we don't take a ref on the node because we're removing it from the 7167 * tree, so we just steal the ref the tree was holding. 7168 */ 7169 delayed_refs->num_heads--; 7170 if (head->processing == 0) 7171 delayed_refs->num_heads_ready--; 7172 head->processing = 0; 7173 spin_unlock(&head->lock); 7174 spin_unlock(&delayed_refs->lock); 7175 7176 BUG_ON(head->extent_op); 7177 if (head->must_insert_reserved) 7178 ret = 1; 7179 7180 mutex_unlock(&head->mutex); 7181 btrfs_put_delayed_ref_head(head); 7182 return ret; 7183 out: 7184 spin_unlock(&head->lock); 7185 7186 out_delayed_unlock: 7187 spin_unlock(&delayed_refs->lock); 7188 return 0; 7189 } 7190 7191 void btrfs_free_tree_block(struct btrfs_trans_handle *trans, 7192 struct btrfs_root *root, 7193 struct extent_buffer *buf, 7194 u64 parent, int last_ref) 7195 { 7196 struct btrfs_fs_info *fs_info = root->fs_info; 7197 int pin = 1; 7198 int ret; 7199 7200 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) { 7201 int old_ref_mod, new_ref_mod; 7202 7203 btrfs_ref_tree_mod(root, buf->start, buf->len, parent, 7204 root->root_key.objectid, 7205 btrfs_header_level(buf), 0, 7206 BTRFS_DROP_DELAYED_REF); 7207 ret = btrfs_add_delayed_tree_ref(fs_info, trans, buf->start, 7208 buf->len, parent, 7209 root->root_key.objectid, 7210 btrfs_header_level(buf), 7211 BTRFS_DROP_DELAYED_REF, NULL, 7212 &old_ref_mod, &new_ref_mod); 7213 BUG_ON(ret); /* -ENOMEM */ 7214 pin = old_ref_mod >= 0 && new_ref_mod < 0; 7215 } 7216 7217 if (last_ref && btrfs_header_generation(buf) == trans->transid) { 7218 struct btrfs_block_group_cache *cache; 7219 7220 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) { 7221 ret = check_ref_cleanup(trans, buf->start); 7222 if (!ret) 7223 goto out; 7224 } 7225 7226 pin = 0; 7227 cache = btrfs_lookup_block_group(fs_info, buf->start); 7228 7229 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) { 7230 pin_down_extent(fs_info, cache, buf->start, 7231 buf->len, 1); 7232 btrfs_put_block_group(cache); 7233 goto out; 7234 } 7235 7236 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags)); 7237 7238 btrfs_add_free_space(cache, buf->start, buf->len); 7239 btrfs_free_reserved_bytes(cache, buf->len, 0); 7240 btrfs_put_block_group(cache); 7241 trace_btrfs_reserved_extent_free(fs_info, buf->start, buf->len); 7242 } 7243 out: 7244 if (pin) 7245 add_pinned_bytes(fs_info, buf->len, btrfs_header_level(buf), 7246 root->root_key.objectid); 7247 7248 if (last_ref) { 7249 /* 7250 * Deleting the buffer, clear the corrupt flag since it doesn't 7251 * matter anymore. 7252 */ 7253 clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags); 7254 } 7255 } 7256 7257 /* Can return -ENOMEM */ 7258 int btrfs_free_extent(struct btrfs_trans_handle *trans, 7259 struct btrfs_root *root, 7260 u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid, 7261 u64 owner, u64 offset) 7262 { 7263 struct btrfs_fs_info *fs_info = root->fs_info; 7264 int old_ref_mod, new_ref_mod; 7265 int ret; 7266 7267 if (btrfs_is_testing(fs_info)) 7268 return 0; 7269 7270 if (root_objectid != BTRFS_TREE_LOG_OBJECTID) 7271 btrfs_ref_tree_mod(root, bytenr, num_bytes, parent, 7272 root_objectid, owner, offset, 7273 BTRFS_DROP_DELAYED_REF); 7274 7275 /* 7276 * tree log blocks never actually go into the extent allocation 7277 * tree, just update pinning info and exit early. 7278 */ 7279 if (root_objectid == BTRFS_TREE_LOG_OBJECTID) { 7280 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID); 7281 /* unlocks the pinned mutex */ 7282 btrfs_pin_extent(fs_info, bytenr, num_bytes, 1); 7283 old_ref_mod = new_ref_mod = 0; 7284 ret = 0; 7285 } else if (owner < BTRFS_FIRST_FREE_OBJECTID) { 7286 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr, 7287 num_bytes, parent, 7288 root_objectid, (int)owner, 7289 BTRFS_DROP_DELAYED_REF, NULL, 7290 &old_ref_mod, &new_ref_mod); 7291 } else { 7292 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr, 7293 num_bytes, parent, 7294 root_objectid, owner, offset, 7295 0, BTRFS_DROP_DELAYED_REF, 7296 &old_ref_mod, &new_ref_mod); 7297 } 7298 7299 if (ret == 0 && old_ref_mod >= 0 && new_ref_mod < 0) 7300 add_pinned_bytes(fs_info, num_bytes, owner, root_objectid); 7301 7302 return ret; 7303 } 7304 7305 /* 7306 * when we wait for progress in the block group caching, its because 7307 * our allocation attempt failed at least once. So, we must sleep 7308 * and let some progress happen before we try again. 7309 * 7310 * This function will sleep at least once waiting for new free space to 7311 * show up, and then it will check the block group free space numbers 7312 * for our min num_bytes. Another option is to have it go ahead 7313 * and look in the rbtree for a free extent of a given size, but this 7314 * is a good start. 7315 * 7316 * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using 7317 * any of the information in this block group. 7318 */ 7319 static noinline void 7320 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache, 7321 u64 num_bytes) 7322 { 7323 struct btrfs_caching_control *caching_ctl; 7324 7325 caching_ctl = get_caching_control(cache); 7326 if (!caching_ctl) 7327 return; 7328 7329 wait_event(caching_ctl->wait, block_group_cache_done(cache) || 7330 (cache->free_space_ctl->free_space >= num_bytes)); 7331 7332 put_caching_control(caching_ctl); 7333 } 7334 7335 static noinline int 7336 wait_block_group_cache_done(struct btrfs_block_group_cache *cache) 7337 { 7338 struct btrfs_caching_control *caching_ctl; 7339 int ret = 0; 7340 7341 caching_ctl = get_caching_control(cache); 7342 if (!caching_ctl) 7343 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0; 7344 7345 wait_event(caching_ctl->wait, block_group_cache_done(cache)); 7346 if (cache->cached == BTRFS_CACHE_ERROR) 7347 ret = -EIO; 7348 put_caching_control(caching_ctl); 7349 return ret; 7350 } 7351 7352 int __get_raid_index(u64 flags) 7353 { 7354 if (flags & BTRFS_BLOCK_GROUP_RAID10) 7355 return BTRFS_RAID_RAID10; 7356 else if (flags & BTRFS_BLOCK_GROUP_RAID1) 7357 return BTRFS_RAID_RAID1; 7358 else if (flags & BTRFS_BLOCK_GROUP_DUP) 7359 return BTRFS_RAID_DUP; 7360 else if (flags & BTRFS_BLOCK_GROUP_RAID0) 7361 return BTRFS_RAID_RAID0; 7362 else if (flags & BTRFS_BLOCK_GROUP_RAID5) 7363 return BTRFS_RAID_RAID5; 7364 else if (flags & BTRFS_BLOCK_GROUP_RAID6) 7365 return BTRFS_RAID_RAID6; 7366 7367 return BTRFS_RAID_SINGLE; /* BTRFS_BLOCK_GROUP_SINGLE */ 7368 } 7369 7370 int get_block_group_index(struct btrfs_block_group_cache *cache) 7371 { 7372 return __get_raid_index(cache->flags); 7373 } 7374 7375 static const char *btrfs_raid_type_names[BTRFS_NR_RAID_TYPES] = { 7376 [BTRFS_RAID_RAID10] = "raid10", 7377 [BTRFS_RAID_RAID1] = "raid1", 7378 [BTRFS_RAID_DUP] = "dup", 7379 [BTRFS_RAID_RAID0] = "raid0", 7380 [BTRFS_RAID_SINGLE] = "single", 7381 [BTRFS_RAID_RAID5] = "raid5", 7382 [BTRFS_RAID_RAID6] = "raid6", 7383 }; 7384 7385 static const char *get_raid_name(enum btrfs_raid_types type) 7386 { 7387 if (type >= BTRFS_NR_RAID_TYPES) 7388 return NULL; 7389 7390 return btrfs_raid_type_names[type]; 7391 } 7392 7393 enum btrfs_loop_type { 7394 LOOP_CACHING_NOWAIT = 0, 7395 LOOP_CACHING_WAIT = 1, 7396 LOOP_ALLOC_CHUNK = 2, 7397 LOOP_NO_EMPTY_SIZE = 3, 7398 }; 7399 7400 static inline void 7401 btrfs_lock_block_group(struct btrfs_block_group_cache *cache, 7402 int delalloc) 7403 { 7404 if (delalloc) 7405 down_read(&cache->data_rwsem); 7406 } 7407 7408 static inline void 7409 btrfs_grab_block_group(struct btrfs_block_group_cache *cache, 7410 int delalloc) 7411 { 7412 btrfs_get_block_group(cache); 7413 if (delalloc) 7414 down_read(&cache->data_rwsem); 7415 } 7416 7417 static struct btrfs_block_group_cache * 7418 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group, 7419 struct btrfs_free_cluster *cluster, 7420 int delalloc) 7421 { 7422 struct btrfs_block_group_cache *used_bg = NULL; 7423 7424 spin_lock(&cluster->refill_lock); 7425 while (1) { 7426 used_bg = cluster->block_group; 7427 if (!used_bg) 7428 return NULL; 7429 7430 if (used_bg == block_group) 7431 return used_bg; 7432 7433 btrfs_get_block_group(used_bg); 7434 7435 if (!delalloc) 7436 return used_bg; 7437 7438 if (down_read_trylock(&used_bg->data_rwsem)) 7439 return used_bg; 7440 7441 spin_unlock(&cluster->refill_lock); 7442 7443 /* We should only have one-level nested. */ 7444 down_read_nested(&used_bg->data_rwsem, SINGLE_DEPTH_NESTING); 7445 7446 spin_lock(&cluster->refill_lock); 7447 if (used_bg == cluster->block_group) 7448 return used_bg; 7449 7450 up_read(&used_bg->data_rwsem); 7451 btrfs_put_block_group(used_bg); 7452 } 7453 } 7454 7455 static inline void 7456 btrfs_release_block_group(struct btrfs_block_group_cache *cache, 7457 int delalloc) 7458 { 7459 if (delalloc) 7460 up_read(&cache->data_rwsem); 7461 btrfs_put_block_group(cache); 7462 } 7463 7464 /* 7465 * walks the btree of allocated extents and find a hole of a given size. 7466 * The key ins is changed to record the hole: 7467 * ins->objectid == start position 7468 * ins->flags = BTRFS_EXTENT_ITEM_KEY 7469 * ins->offset == the size of the hole. 7470 * Any available blocks before search_start are skipped. 7471 * 7472 * If there is no suitable free space, we will record the max size of 7473 * the free space extent currently. 7474 */ 7475 static noinline int find_free_extent(struct btrfs_fs_info *fs_info, 7476 u64 ram_bytes, u64 num_bytes, u64 empty_size, 7477 u64 hint_byte, struct btrfs_key *ins, 7478 u64 flags, int delalloc) 7479 { 7480 int ret = 0; 7481 struct btrfs_root *root = fs_info->extent_root; 7482 struct btrfs_free_cluster *last_ptr = NULL; 7483 struct btrfs_block_group_cache *block_group = NULL; 7484 u64 search_start = 0; 7485 u64 max_extent_size = 0; 7486 u64 empty_cluster = 0; 7487 struct btrfs_space_info *space_info; 7488 int loop = 0; 7489 int index = __get_raid_index(flags); 7490 bool failed_cluster_refill = false; 7491 bool failed_alloc = false; 7492 bool use_cluster = true; 7493 bool have_caching_bg = false; 7494 bool orig_have_caching_bg = false; 7495 bool full_search = false; 7496 7497 WARN_ON(num_bytes < fs_info->sectorsize); 7498 ins->type = BTRFS_EXTENT_ITEM_KEY; 7499 ins->objectid = 0; 7500 ins->offset = 0; 7501 7502 trace_find_free_extent(fs_info, num_bytes, empty_size, flags); 7503 7504 space_info = __find_space_info(fs_info, flags); 7505 if (!space_info) { 7506 btrfs_err(fs_info, "No space info for %llu", flags); 7507 return -ENOSPC; 7508 } 7509 7510 /* 7511 * If our free space is heavily fragmented we may not be able to make 7512 * big contiguous allocations, so instead of doing the expensive search 7513 * for free space, simply return ENOSPC with our max_extent_size so we 7514 * can go ahead and search for a more manageable chunk. 7515 * 7516 * If our max_extent_size is large enough for our allocation simply 7517 * disable clustering since we will likely not be able to find enough 7518 * space to create a cluster and induce latency trying. 7519 */ 7520 if (unlikely(space_info->max_extent_size)) { 7521 spin_lock(&space_info->lock); 7522 if (space_info->max_extent_size && 7523 num_bytes > space_info->max_extent_size) { 7524 ins->offset = space_info->max_extent_size; 7525 spin_unlock(&space_info->lock); 7526 return -ENOSPC; 7527 } else if (space_info->max_extent_size) { 7528 use_cluster = false; 7529 } 7530 spin_unlock(&space_info->lock); 7531 } 7532 7533 last_ptr = fetch_cluster_info(fs_info, space_info, &empty_cluster); 7534 if (last_ptr) { 7535 spin_lock(&last_ptr->lock); 7536 if (last_ptr->block_group) 7537 hint_byte = last_ptr->window_start; 7538 if (last_ptr->fragmented) { 7539 /* 7540 * We still set window_start so we can keep track of the 7541 * last place we found an allocation to try and save 7542 * some time. 7543 */ 7544 hint_byte = last_ptr->window_start; 7545 use_cluster = false; 7546 } 7547 spin_unlock(&last_ptr->lock); 7548 } 7549 7550 search_start = max(search_start, first_logical_byte(fs_info, 0)); 7551 search_start = max(search_start, hint_byte); 7552 if (search_start == hint_byte) { 7553 block_group = btrfs_lookup_block_group(fs_info, search_start); 7554 /* 7555 * we don't want to use the block group if it doesn't match our 7556 * allocation bits, or if its not cached. 7557 * 7558 * However if we are re-searching with an ideal block group 7559 * picked out then we don't care that the block group is cached. 7560 */ 7561 if (block_group && block_group_bits(block_group, flags) && 7562 block_group->cached != BTRFS_CACHE_NO) { 7563 down_read(&space_info->groups_sem); 7564 if (list_empty(&block_group->list) || 7565 block_group->ro) { 7566 /* 7567 * someone is removing this block group, 7568 * we can't jump into the have_block_group 7569 * target because our list pointers are not 7570 * valid 7571 */ 7572 btrfs_put_block_group(block_group); 7573 up_read(&space_info->groups_sem); 7574 } else { 7575 index = get_block_group_index(block_group); 7576 btrfs_lock_block_group(block_group, delalloc); 7577 goto have_block_group; 7578 } 7579 } else if (block_group) { 7580 btrfs_put_block_group(block_group); 7581 } 7582 } 7583 search: 7584 have_caching_bg = false; 7585 if (index == 0 || index == __get_raid_index(flags)) 7586 full_search = true; 7587 down_read(&space_info->groups_sem); 7588 list_for_each_entry(block_group, &space_info->block_groups[index], 7589 list) { 7590 u64 offset; 7591 int cached; 7592 7593 /* If the block group is read-only, we can skip it entirely. */ 7594 if (unlikely(block_group->ro)) 7595 continue; 7596 7597 btrfs_grab_block_group(block_group, delalloc); 7598 search_start = block_group->key.objectid; 7599 7600 /* 7601 * this can happen if we end up cycling through all the 7602 * raid types, but we want to make sure we only allocate 7603 * for the proper type. 7604 */ 7605 if (!block_group_bits(block_group, flags)) { 7606 u64 extra = BTRFS_BLOCK_GROUP_DUP | 7607 BTRFS_BLOCK_GROUP_RAID1 | 7608 BTRFS_BLOCK_GROUP_RAID5 | 7609 BTRFS_BLOCK_GROUP_RAID6 | 7610 BTRFS_BLOCK_GROUP_RAID10; 7611 7612 /* 7613 * if they asked for extra copies and this block group 7614 * doesn't provide them, bail. This does allow us to 7615 * fill raid0 from raid1. 7616 */ 7617 if ((flags & extra) && !(block_group->flags & extra)) 7618 goto loop; 7619 } 7620 7621 have_block_group: 7622 cached = block_group_cache_done(block_group); 7623 if (unlikely(!cached)) { 7624 have_caching_bg = true; 7625 ret = cache_block_group(block_group, 0); 7626 BUG_ON(ret < 0); 7627 ret = 0; 7628 } 7629 7630 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR)) 7631 goto loop; 7632 7633 /* 7634 * Ok we want to try and use the cluster allocator, so 7635 * lets look there 7636 */ 7637 if (last_ptr && use_cluster) { 7638 struct btrfs_block_group_cache *used_block_group; 7639 unsigned long aligned_cluster; 7640 /* 7641 * the refill lock keeps out other 7642 * people trying to start a new cluster 7643 */ 7644 used_block_group = btrfs_lock_cluster(block_group, 7645 last_ptr, 7646 delalloc); 7647 if (!used_block_group) 7648 goto refill_cluster; 7649 7650 if (used_block_group != block_group && 7651 (used_block_group->ro || 7652 !block_group_bits(used_block_group, flags))) 7653 goto release_cluster; 7654 7655 offset = btrfs_alloc_from_cluster(used_block_group, 7656 last_ptr, 7657 num_bytes, 7658 used_block_group->key.objectid, 7659 &max_extent_size); 7660 if (offset) { 7661 /* we have a block, we're done */ 7662 spin_unlock(&last_ptr->refill_lock); 7663 trace_btrfs_reserve_extent_cluster(fs_info, 7664 used_block_group, 7665 search_start, num_bytes); 7666 if (used_block_group != block_group) { 7667 btrfs_release_block_group(block_group, 7668 delalloc); 7669 block_group = used_block_group; 7670 } 7671 goto checks; 7672 } 7673 7674 WARN_ON(last_ptr->block_group != used_block_group); 7675 release_cluster: 7676 /* If we are on LOOP_NO_EMPTY_SIZE, we can't 7677 * set up a new clusters, so lets just skip it 7678 * and let the allocator find whatever block 7679 * it can find. If we reach this point, we 7680 * will have tried the cluster allocator 7681 * plenty of times and not have found 7682 * anything, so we are likely way too 7683 * fragmented for the clustering stuff to find 7684 * anything. 7685 * 7686 * However, if the cluster is taken from the 7687 * current block group, release the cluster 7688 * first, so that we stand a better chance of 7689 * succeeding in the unclustered 7690 * allocation. */ 7691 if (loop >= LOOP_NO_EMPTY_SIZE && 7692 used_block_group != block_group) { 7693 spin_unlock(&last_ptr->refill_lock); 7694 btrfs_release_block_group(used_block_group, 7695 delalloc); 7696 goto unclustered_alloc; 7697 } 7698 7699 /* 7700 * this cluster didn't work out, free it and 7701 * start over 7702 */ 7703 btrfs_return_cluster_to_free_space(NULL, last_ptr); 7704 7705 if (used_block_group != block_group) 7706 btrfs_release_block_group(used_block_group, 7707 delalloc); 7708 refill_cluster: 7709 if (loop >= LOOP_NO_EMPTY_SIZE) { 7710 spin_unlock(&last_ptr->refill_lock); 7711 goto unclustered_alloc; 7712 } 7713 7714 aligned_cluster = max_t(unsigned long, 7715 empty_cluster + empty_size, 7716 block_group->full_stripe_len); 7717 7718 /* allocate a cluster in this block group */ 7719 ret = btrfs_find_space_cluster(fs_info, block_group, 7720 last_ptr, search_start, 7721 num_bytes, 7722 aligned_cluster); 7723 if (ret == 0) { 7724 /* 7725 * now pull our allocation out of this 7726 * cluster 7727 */ 7728 offset = btrfs_alloc_from_cluster(block_group, 7729 last_ptr, 7730 num_bytes, 7731 search_start, 7732 &max_extent_size); 7733 if (offset) { 7734 /* we found one, proceed */ 7735 spin_unlock(&last_ptr->refill_lock); 7736 trace_btrfs_reserve_extent_cluster(fs_info, 7737 block_group, search_start, 7738 num_bytes); 7739 goto checks; 7740 } 7741 } else if (!cached && loop > LOOP_CACHING_NOWAIT 7742 && !failed_cluster_refill) { 7743 spin_unlock(&last_ptr->refill_lock); 7744 7745 failed_cluster_refill = true; 7746 wait_block_group_cache_progress(block_group, 7747 num_bytes + empty_cluster + empty_size); 7748 goto have_block_group; 7749 } 7750 7751 /* 7752 * at this point we either didn't find a cluster 7753 * or we weren't able to allocate a block from our 7754 * cluster. Free the cluster we've been trying 7755 * to use, and go to the next block group 7756 */ 7757 btrfs_return_cluster_to_free_space(NULL, last_ptr); 7758 spin_unlock(&last_ptr->refill_lock); 7759 goto loop; 7760 } 7761 7762 unclustered_alloc: 7763 /* 7764 * We are doing an unclustered alloc, set the fragmented flag so 7765 * we don't bother trying to setup a cluster again until we get 7766 * more space. 7767 */ 7768 if (unlikely(last_ptr)) { 7769 spin_lock(&last_ptr->lock); 7770 last_ptr->fragmented = 1; 7771 spin_unlock(&last_ptr->lock); 7772 } 7773 if (cached) { 7774 struct btrfs_free_space_ctl *ctl = 7775 block_group->free_space_ctl; 7776 7777 spin_lock(&ctl->tree_lock); 7778 if (ctl->free_space < 7779 num_bytes + empty_cluster + empty_size) { 7780 if (ctl->free_space > max_extent_size) 7781 max_extent_size = ctl->free_space; 7782 spin_unlock(&ctl->tree_lock); 7783 goto loop; 7784 } 7785 spin_unlock(&ctl->tree_lock); 7786 } 7787 7788 offset = btrfs_find_space_for_alloc(block_group, search_start, 7789 num_bytes, empty_size, 7790 &max_extent_size); 7791 /* 7792 * If we didn't find a chunk, and we haven't failed on this 7793 * block group before, and this block group is in the middle of 7794 * caching and we are ok with waiting, then go ahead and wait 7795 * for progress to be made, and set failed_alloc to true. 7796 * 7797 * If failed_alloc is true then we've already waited on this 7798 * block group once and should move on to the next block group. 7799 */ 7800 if (!offset && !failed_alloc && !cached && 7801 loop > LOOP_CACHING_NOWAIT) { 7802 wait_block_group_cache_progress(block_group, 7803 num_bytes + empty_size); 7804 failed_alloc = true; 7805 goto have_block_group; 7806 } else if (!offset) { 7807 goto loop; 7808 } 7809 checks: 7810 search_start = ALIGN(offset, fs_info->stripesize); 7811 7812 /* move on to the next group */ 7813 if (search_start + num_bytes > 7814 block_group->key.objectid + block_group->key.offset) { 7815 btrfs_add_free_space(block_group, offset, num_bytes); 7816 goto loop; 7817 } 7818 7819 if (offset < search_start) 7820 btrfs_add_free_space(block_group, offset, 7821 search_start - offset); 7822 BUG_ON(offset > search_start); 7823 7824 ret = btrfs_add_reserved_bytes(block_group, ram_bytes, 7825 num_bytes, delalloc); 7826 if (ret == -EAGAIN) { 7827 btrfs_add_free_space(block_group, offset, num_bytes); 7828 goto loop; 7829 } 7830 btrfs_inc_block_group_reservations(block_group); 7831 7832 /* we are all good, lets return */ 7833 ins->objectid = search_start; 7834 ins->offset = num_bytes; 7835 7836 trace_btrfs_reserve_extent(fs_info, block_group, 7837 search_start, num_bytes); 7838 btrfs_release_block_group(block_group, delalloc); 7839 break; 7840 loop: 7841 failed_cluster_refill = false; 7842 failed_alloc = false; 7843 BUG_ON(index != get_block_group_index(block_group)); 7844 btrfs_release_block_group(block_group, delalloc); 7845 cond_resched(); 7846 } 7847 up_read(&space_info->groups_sem); 7848 7849 if ((loop == LOOP_CACHING_NOWAIT) && have_caching_bg 7850 && !orig_have_caching_bg) 7851 orig_have_caching_bg = true; 7852 7853 if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg) 7854 goto search; 7855 7856 if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES) 7857 goto search; 7858 7859 /* 7860 * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking 7861 * caching kthreads as we move along 7862 * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching 7863 * LOOP_ALLOC_CHUNK, force a chunk allocation and try again 7864 * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try 7865 * again 7866 */ 7867 if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) { 7868 index = 0; 7869 if (loop == LOOP_CACHING_NOWAIT) { 7870 /* 7871 * We want to skip the LOOP_CACHING_WAIT step if we 7872 * don't have any uncached bgs and we've already done a 7873 * full search through. 7874 */ 7875 if (orig_have_caching_bg || !full_search) 7876 loop = LOOP_CACHING_WAIT; 7877 else 7878 loop = LOOP_ALLOC_CHUNK; 7879 } else { 7880 loop++; 7881 } 7882 7883 if (loop == LOOP_ALLOC_CHUNK) { 7884 struct btrfs_trans_handle *trans; 7885 int exist = 0; 7886 7887 trans = current->journal_info; 7888 if (trans) 7889 exist = 1; 7890 else 7891 trans = btrfs_join_transaction(root); 7892 7893 if (IS_ERR(trans)) { 7894 ret = PTR_ERR(trans); 7895 goto out; 7896 } 7897 7898 ret = do_chunk_alloc(trans, fs_info, flags, 7899 CHUNK_ALLOC_FORCE); 7900 7901 /* 7902 * If we can't allocate a new chunk we've already looped 7903 * through at least once, move on to the NO_EMPTY_SIZE 7904 * case. 7905 */ 7906 if (ret == -ENOSPC) 7907 loop = LOOP_NO_EMPTY_SIZE; 7908 7909 /* 7910 * Do not bail out on ENOSPC since we 7911 * can do more things. 7912 */ 7913 if (ret < 0 && ret != -ENOSPC) 7914 btrfs_abort_transaction(trans, ret); 7915 else 7916 ret = 0; 7917 if (!exist) 7918 btrfs_end_transaction(trans); 7919 if (ret) 7920 goto out; 7921 } 7922 7923 if (loop == LOOP_NO_EMPTY_SIZE) { 7924 /* 7925 * Don't loop again if we already have no empty_size and 7926 * no empty_cluster. 7927 */ 7928 if (empty_size == 0 && 7929 empty_cluster == 0) { 7930 ret = -ENOSPC; 7931 goto out; 7932 } 7933 empty_size = 0; 7934 empty_cluster = 0; 7935 } 7936 7937 goto search; 7938 } else if (!ins->objectid) { 7939 ret = -ENOSPC; 7940 } else if (ins->objectid) { 7941 if (!use_cluster && last_ptr) { 7942 spin_lock(&last_ptr->lock); 7943 last_ptr->window_start = ins->objectid; 7944 spin_unlock(&last_ptr->lock); 7945 } 7946 ret = 0; 7947 } 7948 out: 7949 if (ret == -ENOSPC) { 7950 spin_lock(&space_info->lock); 7951 space_info->max_extent_size = max_extent_size; 7952 spin_unlock(&space_info->lock); 7953 ins->offset = max_extent_size; 7954 } 7955 return ret; 7956 } 7957 7958 static void dump_space_info(struct btrfs_fs_info *fs_info, 7959 struct btrfs_space_info *info, u64 bytes, 7960 int dump_block_groups) 7961 { 7962 struct btrfs_block_group_cache *cache; 7963 int index = 0; 7964 7965 spin_lock(&info->lock); 7966 btrfs_info(fs_info, "space_info %llu has %llu free, is %sfull", 7967 info->flags, 7968 info->total_bytes - btrfs_space_info_used(info, true), 7969 info->full ? "" : "not "); 7970 btrfs_info(fs_info, 7971 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu", 7972 info->total_bytes, info->bytes_used, info->bytes_pinned, 7973 info->bytes_reserved, info->bytes_may_use, 7974 info->bytes_readonly); 7975 spin_unlock(&info->lock); 7976 7977 if (!dump_block_groups) 7978 return; 7979 7980 down_read(&info->groups_sem); 7981 again: 7982 list_for_each_entry(cache, &info->block_groups[index], list) { 7983 spin_lock(&cache->lock); 7984 btrfs_info(fs_info, 7985 "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s", 7986 cache->key.objectid, cache->key.offset, 7987 btrfs_block_group_used(&cache->item), cache->pinned, 7988 cache->reserved, cache->ro ? "[readonly]" : ""); 7989 btrfs_dump_free_space(cache, bytes); 7990 spin_unlock(&cache->lock); 7991 } 7992 if (++index < BTRFS_NR_RAID_TYPES) 7993 goto again; 7994 up_read(&info->groups_sem); 7995 } 7996 7997 int btrfs_reserve_extent(struct btrfs_root *root, u64 ram_bytes, 7998 u64 num_bytes, u64 min_alloc_size, 7999 u64 empty_size, u64 hint_byte, 8000 struct btrfs_key *ins, int is_data, int delalloc) 8001 { 8002 struct btrfs_fs_info *fs_info = root->fs_info; 8003 bool final_tried = num_bytes == min_alloc_size; 8004 u64 flags; 8005 int ret; 8006 8007 flags = get_alloc_profile_by_root(root, is_data); 8008 again: 8009 WARN_ON(num_bytes < fs_info->sectorsize); 8010 ret = find_free_extent(fs_info, ram_bytes, num_bytes, empty_size, 8011 hint_byte, ins, flags, delalloc); 8012 if (!ret && !is_data) { 8013 btrfs_dec_block_group_reservations(fs_info, ins->objectid); 8014 } else if (ret == -ENOSPC) { 8015 if (!final_tried && ins->offset) { 8016 num_bytes = min(num_bytes >> 1, ins->offset); 8017 num_bytes = round_down(num_bytes, 8018 fs_info->sectorsize); 8019 num_bytes = max(num_bytes, min_alloc_size); 8020 ram_bytes = num_bytes; 8021 if (num_bytes == min_alloc_size) 8022 final_tried = true; 8023 goto again; 8024 } else if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) { 8025 struct btrfs_space_info *sinfo; 8026 8027 sinfo = __find_space_info(fs_info, flags); 8028 btrfs_err(fs_info, 8029 "allocation failed flags %llu, wanted %llu", 8030 flags, num_bytes); 8031 if (sinfo) 8032 dump_space_info(fs_info, sinfo, num_bytes, 1); 8033 } 8034 } 8035 8036 return ret; 8037 } 8038 8039 static int __btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info, 8040 u64 start, u64 len, 8041 int pin, int delalloc) 8042 { 8043 struct btrfs_block_group_cache *cache; 8044 int ret = 0; 8045 8046 cache = btrfs_lookup_block_group(fs_info, start); 8047 if (!cache) { 8048 btrfs_err(fs_info, "Unable to find block group for %llu", 8049 start); 8050 return -ENOSPC; 8051 } 8052 8053 if (pin) 8054 pin_down_extent(fs_info, cache, start, len, 1); 8055 else { 8056 if (btrfs_test_opt(fs_info, DISCARD)) 8057 ret = btrfs_discard_extent(fs_info, start, len, NULL); 8058 btrfs_add_free_space(cache, start, len); 8059 btrfs_free_reserved_bytes(cache, len, delalloc); 8060 trace_btrfs_reserved_extent_free(fs_info, start, len); 8061 } 8062 8063 btrfs_put_block_group(cache); 8064 return ret; 8065 } 8066 8067 int btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info, 8068 u64 start, u64 len, int delalloc) 8069 { 8070 return __btrfs_free_reserved_extent(fs_info, start, len, 0, delalloc); 8071 } 8072 8073 int btrfs_free_and_pin_reserved_extent(struct btrfs_fs_info *fs_info, 8074 u64 start, u64 len) 8075 { 8076 return __btrfs_free_reserved_extent(fs_info, start, len, 1, 0); 8077 } 8078 8079 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans, 8080 struct btrfs_fs_info *fs_info, 8081 u64 parent, u64 root_objectid, 8082 u64 flags, u64 owner, u64 offset, 8083 struct btrfs_key *ins, int ref_mod) 8084 { 8085 int ret; 8086 struct btrfs_extent_item *extent_item; 8087 struct btrfs_extent_inline_ref *iref; 8088 struct btrfs_path *path; 8089 struct extent_buffer *leaf; 8090 int type; 8091 u32 size; 8092 8093 if (parent > 0) 8094 type = BTRFS_SHARED_DATA_REF_KEY; 8095 else 8096 type = BTRFS_EXTENT_DATA_REF_KEY; 8097 8098 size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type); 8099 8100 path = btrfs_alloc_path(); 8101 if (!path) 8102 return -ENOMEM; 8103 8104 path->leave_spinning = 1; 8105 ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path, 8106 ins, size); 8107 if (ret) { 8108 btrfs_free_path(path); 8109 return ret; 8110 } 8111 8112 leaf = path->nodes[0]; 8113 extent_item = btrfs_item_ptr(leaf, path->slots[0], 8114 struct btrfs_extent_item); 8115 btrfs_set_extent_refs(leaf, extent_item, ref_mod); 8116 btrfs_set_extent_generation(leaf, extent_item, trans->transid); 8117 btrfs_set_extent_flags(leaf, extent_item, 8118 flags | BTRFS_EXTENT_FLAG_DATA); 8119 8120 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1); 8121 btrfs_set_extent_inline_ref_type(leaf, iref, type); 8122 if (parent > 0) { 8123 struct btrfs_shared_data_ref *ref; 8124 ref = (struct btrfs_shared_data_ref *)(iref + 1); 8125 btrfs_set_extent_inline_ref_offset(leaf, iref, parent); 8126 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod); 8127 } else { 8128 struct btrfs_extent_data_ref *ref; 8129 ref = (struct btrfs_extent_data_ref *)(&iref->offset); 8130 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid); 8131 btrfs_set_extent_data_ref_objectid(leaf, ref, owner); 8132 btrfs_set_extent_data_ref_offset(leaf, ref, offset); 8133 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod); 8134 } 8135 8136 btrfs_mark_buffer_dirty(path->nodes[0]); 8137 btrfs_free_path(path); 8138 8139 ret = remove_from_free_space_tree(trans, fs_info, ins->objectid, 8140 ins->offset); 8141 if (ret) 8142 return ret; 8143 8144 ret = update_block_group(trans, fs_info, ins->objectid, ins->offset, 1); 8145 if (ret) { /* -ENOENT, logic error */ 8146 btrfs_err(fs_info, "update block group failed for %llu %llu", 8147 ins->objectid, ins->offset); 8148 BUG(); 8149 } 8150 trace_btrfs_reserved_extent_alloc(fs_info, ins->objectid, ins->offset); 8151 return ret; 8152 } 8153 8154 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans, 8155 struct btrfs_fs_info *fs_info, 8156 u64 parent, u64 root_objectid, 8157 u64 flags, struct btrfs_disk_key *key, 8158 int level, struct btrfs_key *ins) 8159 { 8160 int ret; 8161 struct btrfs_extent_item *extent_item; 8162 struct btrfs_tree_block_info *block_info; 8163 struct btrfs_extent_inline_ref *iref; 8164 struct btrfs_path *path; 8165 struct extent_buffer *leaf; 8166 u32 size = sizeof(*extent_item) + sizeof(*iref); 8167 u64 num_bytes = ins->offset; 8168 bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA); 8169 8170 if (!skinny_metadata) 8171 size += sizeof(*block_info); 8172 8173 path = btrfs_alloc_path(); 8174 if (!path) { 8175 btrfs_free_and_pin_reserved_extent(fs_info, ins->objectid, 8176 fs_info->nodesize); 8177 return -ENOMEM; 8178 } 8179 8180 path->leave_spinning = 1; 8181 ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path, 8182 ins, size); 8183 if (ret) { 8184 btrfs_free_path(path); 8185 btrfs_free_and_pin_reserved_extent(fs_info, ins->objectid, 8186 fs_info->nodesize); 8187 return ret; 8188 } 8189 8190 leaf = path->nodes[0]; 8191 extent_item = btrfs_item_ptr(leaf, path->slots[0], 8192 struct btrfs_extent_item); 8193 btrfs_set_extent_refs(leaf, extent_item, 1); 8194 btrfs_set_extent_generation(leaf, extent_item, trans->transid); 8195 btrfs_set_extent_flags(leaf, extent_item, 8196 flags | BTRFS_EXTENT_FLAG_TREE_BLOCK); 8197 8198 if (skinny_metadata) { 8199 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1); 8200 num_bytes = fs_info->nodesize; 8201 } else { 8202 block_info = (struct btrfs_tree_block_info *)(extent_item + 1); 8203 btrfs_set_tree_block_key(leaf, block_info, key); 8204 btrfs_set_tree_block_level(leaf, block_info, level); 8205 iref = (struct btrfs_extent_inline_ref *)(block_info + 1); 8206 } 8207 8208 if (parent > 0) { 8209 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)); 8210 btrfs_set_extent_inline_ref_type(leaf, iref, 8211 BTRFS_SHARED_BLOCK_REF_KEY); 8212 btrfs_set_extent_inline_ref_offset(leaf, iref, parent); 8213 } else { 8214 btrfs_set_extent_inline_ref_type(leaf, iref, 8215 BTRFS_TREE_BLOCK_REF_KEY); 8216 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid); 8217 } 8218 8219 btrfs_mark_buffer_dirty(leaf); 8220 btrfs_free_path(path); 8221 8222 ret = remove_from_free_space_tree(trans, fs_info, ins->objectid, 8223 num_bytes); 8224 if (ret) 8225 return ret; 8226 8227 ret = update_block_group(trans, fs_info, ins->objectid, 8228 fs_info->nodesize, 1); 8229 if (ret) { /* -ENOENT, logic error */ 8230 btrfs_err(fs_info, "update block group failed for %llu %llu", 8231 ins->objectid, ins->offset); 8232 BUG(); 8233 } 8234 8235 trace_btrfs_reserved_extent_alloc(fs_info, ins->objectid, 8236 fs_info->nodesize); 8237 return ret; 8238 } 8239 8240 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans, 8241 struct btrfs_root *root, u64 owner, 8242 u64 offset, u64 ram_bytes, 8243 struct btrfs_key *ins) 8244 { 8245 struct btrfs_fs_info *fs_info = root->fs_info; 8246 int ret; 8247 8248 BUG_ON(root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID); 8249 8250 btrfs_ref_tree_mod(root, ins->objectid, ins->offset, 0, 8251 root->root_key.objectid, owner, offset, 8252 BTRFS_ADD_DELAYED_EXTENT); 8253 8254 ret = btrfs_add_delayed_data_ref(fs_info, trans, ins->objectid, 8255 ins->offset, 0, 8256 root->root_key.objectid, owner, 8257 offset, ram_bytes, 8258 BTRFS_ADD_DELAYED_EXTENT, NULL, NULL); 8259 return ret; 8260 } 8261 8262 /* 8263 * this is used by the tree logging recovery code. It records that 8264 * an extent has been allocated and makes sure to clear the free 8265 * space cache bits as well 8266 */ 8267 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans, 8268 struct btrfs_fs_info *fs_info, 8269 u64 root_objectid, u64 owner, u64 offset, 8270 struct btrfs_key *ins) 8271 { 8272 int ret; 8273 struct btrfs_block_group_cache *block_group; 8274 struct btrfs_space_info *space_info; 8275 8276 /* 8277 * Mixed block groups will exclude before processing the log so we only 8278 * need to do the exclude dance if this fs isn't mixed. 8279 */ 8280 if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS)) { 8281 ret = __exclude_logged_extent(fs_info, ins->objectid, 8282 ins->offset); 8283 if (ret) 8284 return ret; 8285 } 8286 8287 block_group = btrfs_lookup_block_group(fs_info, ins->objectid); 8288 if (!block_group) 8289 return -EINVAL; 8290 8291 space_info = block_group->space_info; 8292 spin_lock(&space_info->lock); 8293 spin_lock(&block_group->lock); 8294 space_info->bytes_reserved += ins->offset; 8295 block_group->reserved += ins->offset; 8296 spin_unlock(&block_group->lock); 8297 spin_unlock(&space_info->lock); 8298 8299 ret = alloc_reserved_file_extent(trans, fs_info, 0, root_objectid, 8300 0, owner, offset, ins, 1); 8301 btrfs_put_block_group(block_group); 8302 return ret; 8303 } 8304 8305 static struct extent_buffer * 8306 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root, 8307 u64 bytenr, int level) 8308 { 8309 struct btrfs_fs_info *fs_info = root->fs_info; 8310 struct extent_buffer *buf; 8311 8312 buf = btrfs_find_create_tree_block(fs_info, bytenr); 8313 if (IS_ERR(buf)) 8314 return buf; 8315 8316 btrfs_set_header_generation(buf, trans->transid); 8317 btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level); 8318 btrfs_tree_lock(buf); 8319 clean_tree_block(fs_info, buf); 8320 clear_bit(EXTENT_BUFFER_STALE, &buf->bflags); 8321 8322 btrfs_set_lock_blocking(buf); 8323 set_extent_buffer_uptodate(buf); 8324 8325 if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) { 8326 buf->log_index = root->log_transid % 2; 8327 /* 8328 * we allow two log transactions at a time, use different 8329 * EXENT bit to differentiate dirty pages. 8330 */ 8331 if (buf->log_index == 0) 8332 set_extent_dirty(&root->dirty_log_pages, buf->start, 8333 buf->start + buf->len - 1, GFP_NOFS); 8334 else 8335 set_extent_new(&root->dirty_log_pages, buf->start, 8336 buf->start + buf->len - 1); 8337 } else { 8338 buf->log_index = -1; 8339 set_extent_dirty(&trans->transaction->dirty_pages, buf->start, 8340 buf->start + buf->len - 1, GFP_NOFS); 8341 } 8342 trans->dirty = true; 8343 /* this returns a buffer locked for blocking */ 8344 return buf; 8345 } 8346 8347 static struct btrfs_block_rsv * 8348 use_block_rsv(struct btrfs_trans_handle *trans, 8349 struct btrfs_root *root, u32 blocksize) 8350 { 8351 struct btrfs_fs_info *fs_info = root->fs_info; 8352 struct btrfs_block_rsv *block_rsv; 8353 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv; 8354 int ret; 8355 bool global_updated = false; 8356 8357 block_rsv = get_block_rsv(trans, root); 8358 8359 if (unlikely(block_rsv->size == 0)) 8360 goto try_reserve; 8361 again: 8362 ret = block_rsv_use_bytes(block_rsv, blocksize); 8363 if (!ret) 8364 return block_rsv; 8365 8366 if (block_rsv->failfast) 8367 return ERR_PTR(ret); 8368 8369 if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) { 8370 global_updated = true; 8371 update_global_block_rsv(fs_info); 8372 goto again; 8373 } 8374 8375 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) { 8376 static DEFINE_RATELIMIT_STATE(_rs, 8377 DEFAULT_RATELIMIT_INTERVAL * 10, 8378 /*DEFAULT_RATELIMIT_BURST*/ 1); 8379 if (__ratelimit(&_rs)) 8380 WARN(1, KERN_DEBUG 8381 "BTRFS: block rsv returned %d\n", ret); 8382 } 8383 try_reserve: 8384 ret = reserve_metadata_bytes(root, block_rsv, blocksize, 8385 BTRFS_RESERVE_NO_FLUSH); 8386 if (!ret) 8387 return block_rsv; 8388 /* 8389 * If we couldn't reserve metadata bytes try and use some from 8390 * the global reserve if its space type is the same as the global 8391 * reservation. 8392 */ 8393 if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL && 8394 block_rsv->space_info == global_rsv->space_info) { 8395 ret = block_rsv_use_bytes(global_rsv, blocksize); 8396 if (!ret) 8397 return global_rsv; 8398 } 8399 return ERR_PTR(ret); 8400 } 8401 8402 static void unuse_block_rsv(struct btrfs_fs_info *fs_info, 8403 struct btrfs_block_rsv *block_rsv, u32 blocksize) 8404 { 8405 block_rsv_add_bytes(block_rsv, blocksize, 0); 8406 block_rsv_release_bytes(fs_info, block_rsv, NULL, 0); 8407 } 8408 8409 /* 8410 * finds a free extent and does all the dirty work required for allocation 8411 * returns the tree buffer or an ERR_PTR on error. 8412 */ 8413 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans, 8414 struct btrfs_root *root, 8415 u64 parent, u64 root_objectid, 8416 const struct btrfs_disk_key *key, 8417 int level, u64 hint, 8418 u64 empty_size) 8419 { 8420 struct btrfs_fs_info *fs_info = root->fs_info; 8421 struct btrfs_key ins; 8422 struct btrfs_block_rsv *block_rsv; 8423 struct extent_buffer *buf; 8424 struct btrfs_delayed_extent_op *extent_op; 8425 u64 flags = 0; 8426 int ret; 8427 u32 blocksize = fs_info->nodesize; 8428 bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA); 8429 8430 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS 8431 if (btrfs_is_testing(fs_info)) { 8432 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr, 8433 level); 8434 if (!IS_ERR(buf)) 8435 root->alloc_bytenr += blocksize; 8436 return buf; 8437 } 8438 #endif 8439 8440 block_rsv = use_block_rsv(trans, root, blocksize); 8441 if (IS_ERR(block_rsv)) 8442 return ERR_CAST(block_rsv); 8443 8444 ret = btrfs_reserve_extent(root, blocksize, blocksize, blocksize, 8445 empty_size, hint, &ins, 0, 0); 8446 if (ret) 8447 goto out_unuse; 8448 8449 buf = btrfs_init_new_buffer(trans, root, ins.objectid, level); 8450 if (IS_ERR(buf)) { 8451 ret = PTR_ERR(buf); 8452 goto out_free_reserved; 8453 } 8454 8455 if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) { 8456 if (parent == 0) 8457 parent = ins.objectid; 8458 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF; 8459 } else 8460 BUG_ON(parent > 0); 8461 8462 if (root_objectid != BTRFS_TREE_LOG_OBJECTID) { 8463 extent_op = btrfs_alloc_delayed_extent_op(); 8464 if (!extent_op) { 8465 ret = -ENOMEM; 8466 goto out_free_buf; 8467 } 8468 if (key) 8469 memcpy(&extent_op->key, key, sizeof(extent_op->key)); 8470 else 8471 memset(&extent_op->key, 0, sizeof(extent_op->key)); 8472 extent_op->flags_to_set = flags; 8473 extent_op->update_key = skinny_metadata ? false : true; 8474 extent_op->update_flags = true; 8475 extent_op->is_data = false; 8476 extent_op->level = level; 8477 8478 btrfs_ref_tree_mod(root, ins.objectid, ins.offset, parent, 8479 root_objectid, level, 0, 8480 BTRFS_ADD_DELAYED_EXTENT); 8481 ret = btrfs_add_delayed_tree_ref(fs_info, trans, ins.objectid, 8482 ins.offset, parent, 8483 root_objectid, level, 8484 BTRFS_ADD_DELAYED_EXTENT, 8485 extent_op, NULL, NULL); 8486 if (ret) 8487 goto out_free_delayed; 8488 } 8489 return buf; 8490 8491 out_free_delayed: 8492 btrfs_free_delayed_extent_op(extent_op); 8493 out_free_buf: 8494 free_extent_buffer(buf); 8495 out_free_reserved: 8496 btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 0); 8497 out_unuse: 8498 unuse_block_rsv(fs_info, block_rsv, blocksize); 8499 return ERR_PTR(ret); 8500 } 8501 8502 struct walk_control { 8503 u64 refs[BTRFS_MAX_LEVEL]; 8504 u64 flags[BTRFS_MAX_LEVEL]; 8505 struct btrfs_key update_progress; 8506 int stage; 8507 int level; 8508 int shared_level; 8509 int update_ref; 8510 int keep_locks; 8511 int reada_slot; 8512 int reada_count; 8513 int for_reloc; 8514 }; 8515 8516 #define DROP_REFERENCE 1 8517 #define UPDATE_BACKREF 2 8518 8519 static noinline void reada_walk_down(struct btrfs_trans_handle *trans, 8520 struct btrfs_root *root, 8521 struct walk_control *wc, 8522 struct btrfs_path *path) 8523 { 8524 struct btrfs_fs_info *fs_info = root->fs_info; 8525 u64 bytenr; 8526 u64 generation; 8527 u64 refs; 8528 u64 flags; 8529 u32 nritems; 8530 struct btrfs_key key; 8531 struct extent_buffer *eb; 8532 int ret; 8533 int slot; 8534 int nread = 0; 8535 8536 if (path->slots[wc->level] < wc->reada_slot) { 8537 wc->reada_count = wc->reada_count * 2 / 3; 8538 wc->reada_count = max(wc->reada_count, 2); 8539 } else { 8540 wc->reada_count = wc->reada_count * 3 / 2; 8541 wc->reada_count = min_t(int, wc->reada_count, 8542 BTRFS_NODEPTRS_PER_BLOCK(fs_info)); 8543 } 8544 8545 eb = path->nodes[wc->level]; 8546 nritems = btrfs_header_nritems(eb); 8547 8548 for (slot = path->slots[wc->level]; slot < nritems; slot++) { 8549 if (nread >= wc->reada_count) 8550 break; 8551 8552 cond_resched(); 8553 bytenr = btrfs_node_blockptr(eb, slot); 8554 generation = btrfs_node_ptr_generation(eb, slot); 8555 8556 if (slot == path->slots[wc->level]) 8557 goto reada; 8558 8559 if (wc->stage == UPDATE_BACKREF && 8560 generation <= root->root_key.offset) 8561 continue; 8562 8563 /* We don't lock the tree block, it's OK to be racy here */ 8564 ret = btrfs_lookup_extent_info(trans, fs_info, bytenr, 8565 wc->level - 1, 1, &refs, 8566 &flags); 8567 /* We don't care about errors in readahead. */ 8568 if (ret < 0) 8569 continue; 8570 BUG_ON(refs == 0); 8571 8572 if (wc->stage == DROP_REFERENCE) { 8573 if (refs == 1) 8574 goto reada; 8575 8576 if (wc->level == 1 && 8577 (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) 8578 continue; 8579 if (!wc->update_ref || 8580 generation <= root->root_key.offset) 8581 continue; 8582 btrfs_node_key_to_cpu(eb, &key, slot); 8583 ret = btrfs_comp_cpu_keys(&key, 8584 &wc->update_progress); 8585 if (ret < 0) 8586 continue; 8587 } else { 8588 if (wc->level == 1 && 8589 (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) 8590 continue; 8591 } 8592 reada: 8593 readahead_tree_block(fs_info, bytenr); 8594 nread++; 8595 } 8596 wc->reada_slot = slot; 8597 } 8598 8599 /* 8600 * helper to process tree block while walking down the tree. 8601 * 8602 * when wc->stage == UPDATE_BACKREF, this function updates 8603 * back refs for pointers in the block. 8604 * 8605 * NOTE: return value 1 means we should stop walking down. 8606 */ 8607 static noinline int walk_down_proc(struct btrfs_trans_handle *trans, 8608 struct btrfs_root *root, 8609 struct btrfs_path *path, 8610 struct walk_control *wc, int lookup_info) 8611 { 8612 struct btrfs_fs_info *fs_info = root->fs_info; 8613 int level = wc->level; 8614 struct extent_buffer *eb = path->nodes[level]; 8615 u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF; 8616 int ret; 8617 8618 if (wc->stage == UPDATE_BACKREF && 8619 btrfs_header_owner(eb) != root->root_key.objectid) 8620 return 1; 8621 8622 /* 8623 * when reference count of tree block is 1, it won't increase 8624 * again. once full backref flag is set, we never clear it. 8625 */ 8626 if (lookup_info && 8627 ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) || 8628 (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) { 8629 BUG_ON(!path->locks[level]); 8630 ret = btrfs_lookup_extent_info(trans, fs_info, 8631 eb->start, level, 1, 8632 &wc->refs[level], 8633 &wc->flags[level]); 8634 BUG_ON(ret == -ENOMEM); 8635 if (ret) 8636 return ret; 8637 BUG_ON(wc->refs[level] == 0); 8638 } 8639 8640 if (wc->stage == DROP_REFERENCE) { 8641 if (wc->refs[level] > 1) 8642 return 1; 8643 8644 if (path->locks[level] && !wc->keep_locks) { 8645 btrfs_tree_unlock_rw(eb, path->locks[level]); 8646 path->locks[level] = 0; 8647 } 8648 return 0; 8649 } 8650 8651 /* wc->stage == UPDATE_BACKREF */ 8652 if (!(wc->flags[level] & flag)) { 8653 BUG_ON(!path->locks[level]); 8654 ret = btrfs_inc_ref(trans, root, eb, 1); 8655 BUG_ON(ret); /* -ENOMEM */ 8656 ret = btrfs_dec_ref(trans, root, eb, 0); 8657 BUG_ON(ret); /* -ENOMEM */ 8658 ret = btrfs_set_disk_extent_flags(trans, fs_info, eb->start, 8659 eb->len, flag, 8660 btrfs_header_level(eb), 0); 8661 BUG_ON(ret); /* -ENOMEM */ 8662 wc->flags[level] |= flag; 8663 } 8664 8665 /* 8666 * the block is shared by multiple trees, so it's not good to 8667 * keep the tree lock 8668 */ 8669 if (path->locks[level] && level > 0) { 8670 btrfs_tree_unlock_rw(eb, path->locks[level]); 8671 path->locks[level] = 0; 8672 } 8673 return 0; 8674 } 8675 8676 /* 8677 * helper to process tree block pointer. 8678 * 8679 * when wc->stage == DROP_REFERENCE, this function checks 8680 * reference count of the block pointed to. if the block 8681 * is shared and we need update back refs for the subtree 8682 * rooted at the block, this function changes wc->stage to 8683 * UPDATE_BACKREF. if the block is shared and there is no 8684 * need to update back, this function drops the reference 8685 * to the block. 8686 * 8687 * NOTE: return value 1 means we should stop walking down. 8688 */ 8689 static noinline int do_walk_down(struct btrfs_trans_handle *trans, 8690 struct btrfs_root *root, 8691 struct btrfs_path *path, 8692 struct walk_control *wc, int *lookup_info) 8693 { 8694 struct btrfs_fs_info *fs_info = root->fs_info; 8695 u64 bytenr; 8696 u64 generation; 8697 u64 parent; 8698 u32 blocksize; 8699 struct btrfs_key key; 8700 struct extent_buffer *next; 8701 int level = wc->level; 8702 int reada = 0; 8703 int ret = 0; 8704 bool need_account = false; 8705 8706 generation = btrfs_node_ptr_generation(path->nodes[level], 8707 path->slots[level]); 8708 /* 8709 * if the lower level block was created before the snapshot 8710 * was created, we know there is no need to update back refs 8711 * for the subtree 8712 */ 8713 if (wc->stage == UPDATE_BACKREF && 8714 generation <= root->root_key.offset) { 8715 *lookup_info = 1; 8716 return 1; 8717 } 8718 8719 bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]); 8720 blocksize = fs_info->nodesize; 8721 8722 next = find_extent_buffer(fs_info, bytenr); 8723 if (!next) { 8724 next = btrfs_find_create_tree_block(fs_info, bytenr); 8725 if (IS_ERR(next)) 8726 return PTR_ERR(next); 8727 8728 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next, 8729 level - 1); 8730 reada = 1; 8731 } 8732 btrfs_tree_lock(next); 8733 btrfs_set_lock_blocking(next); 8734 8735 ret = btrfs_lookup_extent_info(trans, fs_info, bytenr, level - 1, 1, 8736 &wc->refs[level - 1], 8737 &wc->flags[level - 1]); 8738 if (ret < 0) 8739 goto out_unlock; 8740 8741 if (unlikely(wc->refs[level - 1] == 0)) { 8742 btrfs_err(fs_info, "Missing references."); 8743 ret = -EIO; 8744 goto out_unlock; 8745 } 8746 *lookup_info = 0; 8747 8748 if (wc->stage == DROP_REFERENCE) { 8749 if (wc->refs[level - 1] > 1) { 8750 need_account = true; 8751 if (level == 1 && 8752 (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF)) 8753 goto skip; 8754 8755 if (!wc->update_ref || 8756 generation <= root->root_key.offset) 8757 goto skip; 8758 8759 btrfs_node_key_to_cpu(path->nodes[level], &key, 8760 path->slots[level]); 8761 ret = btrfs_comp_cpu_keys(&key, &wc->update_progress); 8762 if (ret < 0) 8763 goto skip; 8764 8765 wc->stage = UPDATE_BACKREF; 8766 wc->shared_level = level - 1; 8767 } 8768 } else { 8769 if (level == 1 && 8770 (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF)) 8771 goto skip; 8772 } 8773 8774 if (!btrfs_buffer_uptodate(next, generation, 0)) { 8775 btrfs_tree_unlock(next); 8776 free_extent_buffer(next); 8777 next = NULL; 8778 *lookup_info = 1; 8779 } 8780 8781 if (!next) { 8782 if (reada && level == 1) 8783 reada_walk_down(trans, root, wc, path); 8784 next = read_tree_block(fs_info, bytenr, generation); 8785 if (IS_ERR(next)) { 8786 return PTR_ERR(next); 8787 } else if (!extent_buffer_uptodate(next)) { 8788 free_extent_buffer(next); 8789 return -EIO; 8790 } 8791 btrfs_tree_lock(next); 8792 btrfs_set_lock_blocking(next); 8793 } 8794 8795 level--; 8796 ASSERT(level == btrfs_header_level(next)); 8797 if (level != btrfs_header_level(next)) { 8798 btrfs_err(root->fs_info, "mismatched level"); 8799 ret = -EIO; 8800 goto out_unlock; 8801 } 8802 path->nodes[level] = next; 8803 path->slots[level] = 0; 8804 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING; 8805 wc->level = level; 8806 if (wc->level == 1) 8807 wc->reada_slot = 0; 8808 return 0; 8809 skip: 8810 wc->refs[level - 1] = 0; 8811 wc->flags[level - 1] = 0; 8812 if (wc->stage == DROP_REFERENCE) { 8813 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) { 8814 parent = path->nodes[level]->start; 8815 } else { 8816 ASSERT(root->root_key.objectid == 8817 btrfs_header_owner(path->nodes[level])); 8818 if (root->root_key.objectid != 8819 btrfs_header_owner(path->nodes[level])) { 8820 btrfs_err(root->fs_info, 8821 "mismatched block owner"); 8822 ret = -EIO; 8823 goto out_unlock; 8824 } 8825 parent = 0; 8826 } 8827 8828 if (need_account) { 8829 ret = btrfs_qgroup_trace_subtree(trans, root, next, 8830 generation, level - 1); 8831 if (ret) { 8832 btrfs_err_rl(fs_info, 8833 "Error %d accounting shared subtree. Quota is out of sync, rescan required.", 8834 ret); 8835 } 8836 } 8837 ret = btrfs_free_extent(trans, root, bytenr, blocksize, 8838 parent, root->root_key.objectid, 8839 level - 1, 0); 8840 if (ret) 8841 goto out_unlock; 8842 } 8843 8844 *lookup_info = 1; 8845 ret = 1; 8846 8847 out_unlock: 8848 btrfs_tree_unlock(next); 8849 free_extent_buffer(next); 8850 8851 return ret; 8852 } 8853 8854 /* 8855 * helper to process tree block while walking up the tree. 8856 * 8857 * when wc->stage == DROP_REFERENCE, this function drops 8858 * reference count on the block. 8859 * 8860 * when wc->stage == UPDATE_BACKREF, this function changes 8861 * wc->stage back to DROP_REFERENCE if we changed wc->stage 8862 * to UPDATE_BACKREF previously while processing the block. 8863 * 8864 * NOTE: return value 1 means we should stop walking up. 8865 */ 8866 static noinline int walk_up_proc(struct btrfs_trans_handle *trans, 8867 struct btrfs_root *root, 8868 struct btrfs_path *path, 8869 struct walk_control *wc) 8870 { 8871 struct btrfs_fs_info *fs_info = root->fs_info; 8872 int ret; 8873 int level = wc->level; 8874 struct extent_buffer *eb = path->nodes[level]; 8875 u64 parent = 0; 8876 8877 if (wc->stage == UPDATE_BACKREF) { 8878 BUG_ON(wc->shared_level < level); 8879 if (level < wc->shared_level) 8880 goto out; 8881 8882 ret = find_next_key(path, level + 1, &wc->update_progress); 8883 if (ret > 0) 8884 wc->update_ref = 0; 8885 8886 wc->stage = DROP_REFERENCE; 8887 wc->shared_level = -1; 8888 path->slots[level] = 0; 8889 8890 /* 8891 * check reference count again if the block isn't locked. 8892 * we should start walking down the tree again if reference 8893 * count is one. 8894 */ 8895 if (!path->locks[level]) { 8896 BUG_ON(level == 0); 8897 btrfs_tree_lock(eb); 8898 btrfs_set_lock_blocking(eb); 8899 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING; 8900 8901 ret = btrfs_lookup_extent_info(trans, fs_info, 8902 eb->start, level, 1, 8903 &wc->refs[level], 8904 &wc->flags[level]); 8905 if (ret < 0) { 8906 btrfs_tree_unlock_rw(eb, path->locks[level]); 8907 path->locks[level] = 0; 8908 return ret; 8909 } 8910 BUG_ON(wc->refs[level] == 0); 8911 if (wc->refs[level] == 1) { 8912 btrfs_tree_unlock_rw(eb, path->locks[level]); 8913 path->locks[level] = 0; 8914 return 1; 8915 } 8916 } 8917 } 8918 8919 /* wc->stage == DROP_REFERENCE */ 8920 BUG_ON(wc->refs[level] > 1 && !path->locks[level]); 8921 8922 if (wc->refs[level] == 1) { 8923 if (level == 0) { 8924 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) 8925 ret = btrfs_dec_ref(trans, root, eb, 1); 8926 else 8927 ret = btrfs_dec_ref(trans, root, eb, 0); 8928 BUG_ON(ret); /* -ENOMEM */ 8929 ret = btrfs_qgroup_trace_leaf_items(trans, fs_info, eb); 8930 if (ret) { 8931 btrfs_err_rl(fs_info, 8932 "error %d accounting leaf items. Quota is out of sync, rescan required.", 8933 ret); 8934 } 8935 } 8936 /* make block locked assertion in clean_tree_block happy */ 8937 if (!path->locks[level] && 8938 btrfs_header_generation(eb) == trans->transid) { 8939 btrfs_tree_lock(eb); 8940 btrfs_set_lock_blocking(eb); 8941 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING; 8942 } 8943 clean_tree_block(fs_info, eb); 8944 } 8945 8946 if (eb == root->node) { 8947 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) 8948 parent = eb->start; 8949 else 8950 BUG_ON(root->root_key.objectid != 8951 btrfs_header_owner(eb)); 8952 } else { 8953 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF) 8954 parent = path->nodes[level + 1]->start; 8955 else 8956 BUG_ON(root->root_key.objectid != 8957 btrfs_header_owner(path->nodes[level + 1])); 8958 } 8959 8960 btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1); 8961 out: 8962 wc->refs[level] = 0; 8963 wc->flags[level] = 0; 8964 return 0; 8965 } 8966 8967 static noinline int walk_down_tree(struct btrfs_trans_handle *trans, 8968 struct btrfs_root *root, 8969 struct btrfs_path *path, 8970 struct walk_control *wc) 8971 { 8972 int level = wc->level; 8973 int lookup_info = 1; 8974 int ret; 8975 8976 while (level >= 0) { 8977 ret = walk_down_proc(trans, root, path, wc, lookup_info); 8978 if (ret > 0) 8979 break; 8980 8981 if (level == 0) 8982 break; 8983 8984 if (path->slots[level] >= 8985 btrfs_header_nritems(path->nodes[level])) 8986 break; 8987 8988 ret = do_walk_down(trans, root, path, wc, &lookup_info); 8989 if (ret > 0) { 8990 path->slots[level]++; 8991 continue; 8992 } else if (ret < 0) 8993 return ret; 8994 level = wc->level; 8995 } 8996 return 0; 8997 } 8998 8999 static noinline int walk_up_tree(struct btrfs_trans_handle *trans, 9000 struct btrfs_root *root, 9001 struct btrfs_path *path, 9002 struct walk_control *wc, int max_level) 9003 { 9004 int level = wc->level; 9005 int ret; 9006 9007 path->slots[level] = btrfs_header_nritems(path->nodes[level]); 9008 while (level < max_level && path->nodes[level]) { 9009 wc->level = level; 9010 if (path->slots[level] + 1 < 9011 btrfs_header_nritems(path->nodes[level])) { 9012 path->slots[level]++; 9013 return 0; 9014 } else { 9015 ret = walk_up_proc(trans, root, path, wc); 9016 if (ret > 0) 9017 return 0; 9018 9019 if (path->locks[level]) { 9020 btrfs_tree_unlock_rw(path->nodes[level], 9021 path->locks[level]); 9022 path->locks[level] = 0; 9023 } 9024 free_extent_buffer(path->nodes[level]); 9025 path->nodes[level] = NULL; 9026 level++; 9027 } 9028 } 9029 return 1; 9030 } 9031 9032 /* 9033 * drop a subvolume tree. 9034 * 9035 * this function traverses the tree freeing any blocks that only 9036 * referenced by the tree. 9037 * 9038 * when a shared tree block is found. this function decreases its 9039 * reference count by one. if update_ref is true, this function 9040 * also make sure backrefs for the shared block and all lower level 9041 * blocks are properly updated. 9042 * 9043 * If called with for_reloc == 0, may exit early with -EAGAIN 9044 */ 9045 int btrfs_drop_snapshot(struct btrfs_root *root, 9046 struct btrfs_block_rsv *block_rsv, int update_ref, 9047 int for_reloc) 9048 { 9049 struct btrfs_fs_info *fs_info = root->fs_info; 9050 struct btrfs_path *path; 9051 struct btrfs_trans_handle *trans; 9052 struct btrfs_root *tree_root = fs_info->tree_root; 9053 struct btrfs_root_item *root_item = &root->root_item; 9054 struct walk_control *wc; 9055 struct btrfs_key key; 9056 int err = 0; 9057 int ret; 9058 int level; 9059 bool root_dropped = false; 9060 9061 btrfs_debug(fs_info, "Drop subvolume %llu", root->objectid); 9062 9063 path = btrfs_alloc_path(); 9064 if (!path) { 9065 err = -ENOMEM; 9066 goto out; 9067 } 9068 9069 wc = kzalloc(sizeof(*wc), GFP_NOFS); 9070 if (!wc) { 9071 btrfs_free_path(path); 9072 err = -ENOMEM; 9073 goto out; 9074 } 9075 9076 trans = btrfs_start_transaction(tree_root, 0); 9077 if (IS_ERR(trans)) { 9078 err = PTR_ERR(trans); 9079 goto out_free; 9080 } 9081 9082 if (block_rsv) 9083 trans->block_rsv = block_rsv; 9084 9085 if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) { 9086 level = btrfs_header_level(root->node); 9087 path->nodes[level] = btrfs_lock_root_node(root); 9088 btrfs_set_lock_blocking(path->nodes[level]); 9089 path->slots[level] = 0; 9090 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING; 9091 memset(&wc->update_progress, 0, 9092 sizeof(wc->update_progress)); 9093 } else { 9094 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress); 9095 memcpy(&wc->update_progress, &key, 9096 sizeof(wc->update_progress)); 9097 9098 level = root_item->drop_level; 9099 BUG_ON(level == 0); 9100 path->lowest_level = level; 9101 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0); 9102 path->lowest_level = 0; 9103 if (ret < 0) { 9104 err = ret; 9105 goto out_end_trans; 9106 } 9107 WARN_ON(ret > 0); 9108 9109 /* 9110 * unlock our path, this is safe because only this 9111 * function is allowed to delete this snapshot 9112 */ 9113 btrfs_unlock_up_safe(path, 0); 9114 9115 level = btrfs_header_level(root->node); 9116 while (1) { 9117 btrfs_tree_lock(path->nodes[level]); 9118 btrfs_set_lock_blocking(path->nodes[level]); 9119 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING; 9120 9121 ret = btrfs_lookup_extent_info(trans, fs_info, 9122 path->nodes[level]->start, 9123 level, 1, &wc->refs[level], 9124 &wc->flags[level]); 9125 if (ret < 0) { 9126 err = ret; 9127 goto out_end_trans; 9128 } 9129 BUG_ON(wc->refs[level] == 0); 9130 9131 if (level == root_item->drop_level) 9132 break; 9133 9134 btrfs_tree_unlock(path->nodes[level]); 9135 path->locks[level] = 0; 9136 WARN_ON(wc->refs[level] != 1); 9137 level--; 9138 } 9139 } 9140 9141 wc->level = level; 9142 wc->shared_level = -1; 9143 wc->stage = DROP_REFERENCE; 9144 wc->update_ref = update_ref; 9145 wc->keep_locks = 0; 9146 wc->for_reloc = for_reloc; 9147 wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info); 9148 9149 while (1) { 9150 9151 ret = walk_down_tree(trans, root, path, wc); 9152 if (ret < 0) { 9153 err = ret; 9154 break; 9155 } 9156 9157 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL); 9158 if (ret < 0) { 9159 err = ret; 9160 break; 9161 } 9162 9163 if (ret > 0) { 9164 BUG_ON(wc->stage != DROP_REFERENCE); 9165 break; 9166 } 9167 9168 if (wc->stage == DROP_REFERENCE) { 9169 level = wc->level; 9170 btrfs_node_key(path->nodes[level], 9171 &root_item->drop_progress, 9172 path->slots[level]); 9173 root_item->drop_level = level; 9174 } 9175 9176 BUG_ON(wc->level == 0); 9177 if (btrfs_should_end_transaction(trans) || 9178 (!for_reloc && btrfs_need_cleaner_sleep(fs_info))) { 9179 ret = btrfs_update_root(trans, tree_root, 9180 &root->root_key, 9181 root_item); 9182 if (ret) { 9183 btrfs_abort_transaction(trans, ret); 9184 err = ret; 9185 goto out_end_trans; 9186 } 9187 9188 btrfs_end_transaction_throttle(trans); 9189 if (!for_reloc && btrfs_need_cleaner_sleep(fs_info)) { 9190 btrfs_debug(fs_info, 9191 "drop snapshot early exit"); 9192 err = -EAGAIN; 9193 goto out_free; 9194 } 9195 9196 trans = btrfs_start_transaction(tree_root, 0); 9197 if (IS_ERR(trans)) { 9198 err = PTR_ERR(trans); 9199 goto out_free; 9200 } 9201 if (block_rsv) 9202 trans->block_rsv = block_rsv; 9203 } 9204 } 9205 btrfs_release_path(path); 9206 if (err) 9207 goto out_end_trans; 9208 9209 ret = btrfs_del_root(trans, fs_info, &root->root_key); 9210 if (ret) { 9211 btrfs_abort_transaction(trans, ret); 9212 err = ret; 9213 goto out_end_trans; 9214 } 9215 9216 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) { 9217 ret = btrfs_find_root(tree_root, &root->root_key, path, 9218 NULL, NULL); 9219 if (ret < 0) { 9220 btrfs_abort_transaction(trans, ret); 9221 err = ret; 9222 goto out_end_trans; 9223 } else if (ret > 0) { 9224 /* if we fail to delete the orphan item this time 9225 * around, it'll get picked up the next time. 9226 * 9227 * The most common failure here is just -ENOENT. 9228 */ 9229 btrfs_del_orphan_item(trans, tree_root, 9230 root->root_key.objectid); 9231 } 9232 } 9233 9234 if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) { 9235 btrfs_add_dropped_root(trans, root); 9236 } else { 9237 free_extent_buffer(root->node); 9238 free_extent_buffer(root->commit_root); 9239 btrfs_put_fs_root(root); 9240 } 9241 root_dropped = true; 9242 out_end_trans: 9243 btrfs_end_transaction_throttle(trans); 9244 out_free: 9245 kfree(wc); 9246 btrfs_free_path(path); 9247 out: 9248 /* 9249 * So if we need to stop dropping the snapshot for whatever reason we 9250 * need to make sure to add it back to the dead root list so that we 9251 * keep trying to do the work later. This also cleans up roots if we 9252 * don't have it in the radix (like when we recover after a power fail 9253 * or unmount) so we don't leak memory. 9254 */ 9255 if (!for_reloc && !root_dropped) 9256 btrfs_add_dead_root(root); 9257 if (err && err != -EAGAIN) 9258 btrfs_handle_fs_error(fs_info, err, NULL); 9259 return err; 9260 } 9261 9262 /* 9263 * drop subtree rooted at tree block 'node'. 9264 * 9265 * NOTE: this function will unlock and release tree block 'node' 9266 * only used by relocation code 9267 */ 9268 int btrfs_drop_subtree(struct btrfs_trans_handle *trans, 9269 struct btrfs_root *root, 9270 struct extent_buffer *node, 9271 struct extent_buffer *parent) 9272 { 9273 struct btrfs_fs_info *fs_info = root->fs_info; 9274 struct btrfs_path *path; 9275 struct walk_control *wc; 9276 int level; 9277 int parent_level; 9278 int ret = 0; 9279 int wret; 9280 9281 BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID); 9282 9283 path = btrfs_alloc_path(); 9284 if (!path) 9285 return -ENOMEM; 9286 9287 wc = kzalloc(sizeof(*wc), GFP_NOFS); 9288 if (!wc) { 9289 btrfs_free_path(path); 9290 return -ENOMEM; 9291 } 9292 9293 btrfs_assert_tree_locked(parent); 9294 parent_level = btrfs_header_level(parent); 9295 extent_buffer_get(parent); 9296 path->nodes[parent_level] = parent; 9297 path->slots[parent_level] = btrfs_header_nritems(parent); 9298 9299 btrfs_assert_tree_locked(node); 9300 level = btrfs_header_level(node); 9301 path->nodes[level] = node; 9302 path->slots[level] = 0; 9303 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING; 9304 9305 wc->refs[parent_level] = 1; 9306 wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF; 9307 wc->level = level; 9308 wc->shared_level = -1; 9309 wc->stage = DROP_REFERENCE; 9310 wc->update_ref = 0; 9311 wc->keep_locks = 1; 9312 wc->for_reloc = 1; 9313 wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info); 9314 9315 while (1) { 9316 wret = walk_down_tree(trans, root, path, wc); 9317 if (wret < 0) { 9318 ret = wret; 9319 break; 9320 } 9321 9322 wret = walk_up_tree(trans, root, path, wc, parent_level); 9323 if (wret < 0) 9324 ret = wret; 9325 if (wret != 0) 9326 break; 9327 } 9328 9329 kfree(wc); 9330 btrfs_free_path(path); 9331 return ret; 9332 } 9333 9334 static u64 update_block_group_flags(struct btrfs_fs_info *fs_info, u64 flags) 9335 { 9336 u64 num_devices; 9337 u64 stripped; 9338 9339 /* 9340 * if restripe for this chunk_type is on pick target profile and 9341 * return, otherwise do the usual balance 9342 */ 9343 stripped = get_restripe_target(fs_info, flags); 9344 if (stripped) 9345 return extended_to_chunk(stripped); 9346 9347 num_devices = fs_info->fs_devices->rw_devices; 9348 9349 stripped = BTRFS_BLOCK_GROUP_RAID0 | 9350 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 | 9351 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10; 9352 9353 if (num_devices == 1) { 9354 stripped |= BTRFS_BLOCK_GROUP_DUP; 9355 stripped = flags & ~stripped; 9356 9357 /* turn raid0 into single device chunks */ 9358 if (flags & BTRFS_BLOCK_GROUP_RAID0) 9359 return stripped; 9360 9361 /* turn mirroring into duplication */ 9362 if (flags & (BTRFS_BLOCK_GROUP_RAID1 | 9363 BTRFS_BLOCK_GROUP_RAID10)) 9364 return stripped | BTRFS_BLOCK_GROUP_DUP; 9365 } else { 9366 /* they already had raid on here, just return */ 9367 if (flags & stripped) 9368 return flags; 9369 9370 stripped |= BTRFS_BLOCK_GROUP_DUP; 9371 stripped = flags & ~stripped; 9372 9373 /* switch duplicated blocks with raid1 */ 9374 if (flags & BTRFS_BLOCK_GROUP_DUP) 9375 return stripped | BTRFS_BLOCK_GROUP_RAID1; 9376 9377 /* this is drive concat, leave it alone */ 9378 } 9379 9380 return flags; 9381 } 9382 9383 static int inc_block_group_ro(struct btrfs_block_group_cache *cache, int force) 9384 { 9385 struct btrfs_space_info *sinfo = cache->space_info; 9386 u64 num_bytes; 9387 u64 min_allocable_bytes; 9388 int ret = -ENOSPC; 9389 9390 /* 9391 * We need some metadata space and system metadata space for 9392 * allocating chunks in some corner cases until we force to set 9393 * it to be readonly. 9394 */ 9395 if ((sinfo->flags & 9396 (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) && 9397 !force) 9398 min_allocable_bytes = SZ_1M; 9399 else 9400 min_allocable_bytes = 0; 9401 9402 spin_lock(&sinfo->lock); 9403 spin_lock(&cache->lock); 9404 9405 if (cache->ro) { 9406 cache->ro++; 9407 ret = 0; 9408 goto out; 9409 } 9410 9411 num_bytes = cache->key.offset - cache->reserved - cache->pinned - 9412 cache->bytes_super - btrfs_block_group_used(&cache->item); 9413 9414 if (btrfs_space_info_used(sinfo, true) + num_bytes + 9415 min_allocable_bytes <= sinfo->total_bytes) { 9416 sinfo->bytes_readonly += num_bytes; 9417 cache->ro++; 9418 list_add_tail(&cache->ro_list, &sinfo->ro_bgs); 9419 ret = 0; 9420 } 9421 out: 9422 spin_unlock(&cache->lock); 9423 spin_unlock(&sinfo->lock); 9424 return ret; 9425 } 9426 9427 int btrfs_inc_block_group_ro(struct btrfs_fs_info *fs_info, 9428 struct btrfs_block_group_cache *cache) 9429 9430 { 9431 struct btrfs_trans_handle *trans; 9432 u64 alloc_flags; 9433 int ret; 9434 9435 again: 9436 trans = btrfs_join_transaction(fs_info->extent_root); 9437 if (IS_ERR(trans)) 9438 return PTR_ERR(trans); 9439 9440 /* 9441 * we're not allowed to set block groups readonly after the dirty 9442 * block groups cache has started writing. If it already started, 9443 * back off and let this transaction commit 9444 */ 9445 mutex_lock(&fs_info->ro_block_group_mutex); 9446 if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) { 9447 u64 transid = trans->transid; 9448 9449 mutex_unlock(&fs_info->ro_block_group_mutex); 9450 btrfs_end_transaction(trans); 9451 9452 ret = btrfs_wait_for_commit(fs_info, transid); 9453 if (ret) 9454 return ret; 9455 goto again; 9456 } 9457 9458 /* 9459 * if we are changing raid levels, try to allocate a corresponding 9460 * block group with the new raid level. 9461 */ 9462 alloc_flags = update_block_group_flags(fs_info, cache->flags); 9463 if (alloc_flags != cache->flags) { 9464 ret = do_chunk_alloc(trans, fs_info, alloc_flags, 9465 CHUNK_ALLOC_FORCE); 9466 /* 9467 * ENOSPC is allowed here, we may have enough space 9468 * already allocated at the new raid level to 9469 * carry on 9470 */ 9471 if (ret == -ENOSPC) 9472 ret = 0; 9473 if (ret < 0) 9474 goto out; 9475 } 9476 9477 ret = inc_block_group_ro(cache, 0); 9478 if (!ret) 9479 goto out; 9480 alloc_flags = get_alloc_profile(fs_info, cache->space_info->flags); 9481 ret = do_chunk_alloc(trans, fs_info, alloc_flags, 9482 CHUNK_ALLOC_FORCE); 9483 if (ret < 0) 9484 goto out; 9485 ret = inc_block_group_ro(cache, 0); 9486 out: 9487 if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) { 9488 alloc_flags = update_block_group_flags(fs_info, cache->flags); 9489 mutex_lock(&fs_info->chunk_mutex); 9490 check_system_chunk(trans, fs_info, alloc_flags); 9491 mutex_unlock(&fs_info->chunk_mutex); 9492 } 9493 mutex_unlock(&fs_info->ro_block_group_mutex); 9494 9495 btrfs_end_transaction(trans); 9496 return ret; 9497 } 9498 9499 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans, 9500 struct btrfs_fs_info *fs_info, u64 type) 9501 { 9502 u64 alloc_flags = get_alloc_profile(fs_info, type); 9503 9504 return do_chunk_alloc(trans, fs_info, alloc_flags, CHUNK_ALLOC_FORCE); 9505 } 9506 9507 /* 9508 * helper to account the unused space of all the readonly block group in the 9509 * space_info. takes mirrors into account. 9510 */ 9511 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo) 9512 { 9513 struct btrfs_block_group_cache *block_group; 9514 u64 free_bytes = 0; 9515 int factor; 9516 9517 /* It's df, we don't care if it's racy */ 9518 if (list_empty(&sinfo->ro_bgs)) 9519 return 0; 9520 9521 spin_lock(&sinfo->lock); 9522 list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) { 9523 spin_lock(&block_group->lock); 9524 9525 if (!block_group->ro) { 9526 spin_unlock(&block_group->lock); 9527 continue; 9528 } 9529 9530 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 | 9531 BTRFS_BLOCK_GROUP_RAID10 | 9532 BTRFS_BLOCK_GROUP_DUP)) 9533 factor = 2; 9534 else 9535 factor = 1; 9536 9537 free_bytes += (block_group->key.offset - 9538 btrfs_block_group_used(&block_group->item)) * 9539 factor; 9540 9541 spin_unlock(&block_group->lock); 9542 } 9543 spin_unlock(&sinfo->lock); 9544 9545 return free_bytes; 9546 } 9547 9548 void btrfs_dec_block_group_ro(struct btrfs_block_group_cache *cache) 9549 { 9550 struct btrfs_space_info *sinfo = cache->space_info; 9551 u64 num_bytes; 9552 9553 BUG_ON(!cache->ro); 9554 9555 spin_lock(&sinfo->lock); 9556 spin_lock(&cache->lock); 9557 if (!--cache->ro) { 9558 num_bytes = cache->key.offset - cache->reserved - 9559 cache->pinned - cache->bytes_super - 9560 btrfs_block_group_used(&cache->item); 9561 sinfo->bytes_readonly -= num_bytes; 9562 list_del_init(&cache->ro_list); 9563 } 9564 spin_unlock(&cache->lock); 9565 spin_unlock(&sinfo->lock); 9566 } 9567 9568 /* 9569 * checks to see if its even possible to relocate this block group. 9570 * 9571 * @return - -1 if it's not a good idea to relocate this block group, 0 if its 9572 * ok to go ahead and try. 9573 */ 9574 int btrfs_can_relocate(struct btrfs_fs_info *fs_info, u64 bytenr) 9575 { 9576 struct btrfs_root *root = fs_info->extent_root; 9577 struct btrfs_block_group_cache *block_group; 9578 struct btrfs_space_info *space_info; 9579 struct btrfs_fs_devices *fs_devices = fs_info->fs_devices; 9580 struct btrfs_device *device; 9581 struct btrfs_trans_handle *trans; 9582 u64 min_free; 9583 u64 dev_min = 1; 9584 u64 dev_nr = 0; 9585 u64 target; 9586 int debug; 9587 int index; 9588 int full = 0; 9589 int ret = 0; 9590 9591 debug = btrfs_test_opt(fs_info, ENOSPC_DEBUG); 9592 9593 block_group = btrfs_lookup_block_group(fs_info, bytenr); 9594 9595 /* odd, couldn't find the block group, leave it alone */ 9596 if (!block_group) { 9597 if (debug) 9598 btrfs_warn(fs_info, 9599 "can't find block group for bytenr %llu", 9600 bytenr); 9601 return -1; 9602 } 9603 9604 min_free = btrfs_block_group_used(&block_group->item); 9605 9606 /* no bytes used, we're good */ 9607 if (!min_free) 9608 goto out; 9609 9610 space_info = block_group->space_info; 9611 spin_lock(&space_info->lock); 9612 9613 full = space_info->full; 9614 9615 /* 9616 * if this is the last block group we have in this space, we can't 9617 * relocate it unless we're able to allocate a new chunk below. 9618 * 9619 * Otherwise, we need to make sure we have room in the space to handle 9620 * all of the extents from this block group. If we can, we're good 9621 */ 9622 if ((space_info->total_bytes != block_group->key.offset) && 9623 (btrfs_space_info_used(space_info, false) + min_free < 9624 space_info->total_bytes)) { 9625 spin_unlock(&space_info->lock); 9626 goto out; 9627 } 9628 spin_unlock(&space_info->lock); 9629 9630 /* 9631 * ok we don't have enough space, but maybe we have free space on our 9632 * devices to allocate new chunks for relocation, so loop through our 9633 * alloc devices and guess if we have enough space. if this block 9634 * group is going to be restriped, run checks against the target 9635 * profile instead of the current one. 9636 */ 9637 ret = -1; 9638 9639 /* 9640 * index: 9641 * 0: raid10 9642 * 1: raid1 9643 * 2: dup 9644 * 3: raid0 9645 * 4: single 9646 */ 9647 target = get_restripe_target(fs_info, block_group->flags); 9648 if (target) { 9649 index = __get_raid_index(extended_to_chunk(target)); 9650 } else { 9651 /* 9652 * this is just a balance, so if we were marked as full 9653 * we know there is no space for a new chunk 9654 */ 9655 if (full) { 9656 if (debug) 9657 btrfs_warn(fs_info, 9658 "no space to alloc new chunk for block group %llu", 9659 block_group->key.objectid); 9660 goto out; 9661 } 9662 9663 index = get_block_group_index(block_group); 9664 } 9665 9666 if (index == BTRFS_RAID_RAID10) { 9667 dev_min = 4; 9668 /* Divide by 2 */ 9669 min_free >>= 1; 9670 } else if (index == BTRFS_RAID_RAID1) { 9671 dev_min = 2; 9672 } else if (index == BTRFS_RAID_DUP) { 9673 /* Multiply by 2 */ 9674 min_free <<= 1; 9675 } else if (index == BTRFS_RAID_RAID0) { 9676 dev_min = fs_devices->rw_devices; 9677 min_free = div64_u64(min_free, dev_min); 9678 } 9679 9680 /* We need to do this so that we can look at pending chunks */ 9681 trans = btrfs_join_transaction(root); 9682 if (IS_ERR(trans)) { 9683 ret = PTR_ERR(trans); 9684 goto out; 9685 } 9686 9687 mutex_lock(&fs_info->chunk_mutex); 9688 list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) { 9689 u64 dev_offset; 9690 9691 /* 9692 * check to make sure we can actually find a chunk with enough 9693 * space to fit our block group in. 9694 */ 9695 if (device->total_bytes > device->bytes_used + min_free && 9696 !test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state)) { 9697 ret = find_free_dev_extent(trans, device, min_free, 9698 &dev_offset, NULL); 9699 if (!ret) 9700 dev_nr++; 9701 9702 if (dev_nr >= dev_min) 9703 break; 9704 9705 ret = -1; 9706 } 9707 } 9708 if (debug && ret == -1) 9709 btrfs_warn(fs_info, 9710 "no space to allocate a new chunk for block group %llu", 9711 block_group->key.objectid); 9712 mutex_unlock(&fs_info->chunk_mutex); 9713 btrfs_end_transaction(trans); 9714 out: 9715 btrfs_put_block_group(block_group); 9716 return ret; 9717 } 9718 9719 static int find_first_block_group(struct btrfs_fs_info *fs_info, 9720 struct btrfs_path *path, 9721 struct btrfs_key *key) 9722 { 9723 struct btrfs_root *root = fs_info->extent_root; 9724 int ret = 0; 9725 struct btrfs_key found_key; 9726 struct extent_buffer *leaf; 9727 int slot; 9728 9729 ret = btrfs_search_slot(NULL, root, key, path, 0, 0); 9730 if (ret < 0) 9731 goto out; 9732 9733 while (1) { 9734 slot = path->slots[0]; 9735 leaf = path->nodes[0]; 9736 if (slot >= btrfs_header_nritems(leaf)) { 9737 ret = btrfs_next_leaf(root, path); 9738 if (ret == 0) 9739 continue; 9740 if (ret < 0) 9741 goto out; 9742 break; 9743 } 9744 btrfs_item_key_to_cpu(leaf, &found_key, slot); 9745 9746 if (found_key.objectid >= key->objectid && 9747 found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) { 9748 struct extent_map_tree *em_tree; 9749 struct extent_map *em; 9750 9751 em_tree = &root->fs_info->mapping_tree.map_tree; 9752 read_lock(&em_tree->lock); 9753 em = lookup_extent_mapping(em_tree, found_key.objectid, 9754 found_key.offset); 9755 read_unlock(&em_tree->lock); 9756 if (!em) { 9757 btrfs_err(fs_info, 9758 "logical %llu len %llu found bg but no related chunk", 9759 found_key.objectid, found_key.offset); 9760 ret = -ENOENT; 9761 } else { 9762 ret = 0; 9763 } 9764 free_extent_map(em); 9765 goto out; 9766 } 9767 path->slots[0]++; 9768 } 9769 out: 9770 return ret; 9771 } 9772 9773 void btrfs_put_block_group_cache(struct btrfs_fs_info *info) 9774 { 9775 struct btrfs_block_group_cache *block_group; 9776 u64 last = 0; 9777 9778 while (1) { 9779 struct inode *inode; 9780 9781 block_group = btrfs_lookup_first_block_group(info, last); 9782 while (block_group) { 9783 spin_lock(&block_group->lock); 9784 if (block_group->iref) 9785 break; 9786 spin_unlock(&block_group->lock); 9787 block_group = next_block_group(info, block_group); 9788 } 9789 if (!block_group) { 9790 if (last == 0) 9791 break; 9792 last = 0; 9793 continue; 9794 } 9795 9796 inode = block_group->inode; 9797 block_group->iref = 0; 9798 block_group->inode = NULL; 9799 spin_unlock(&block_group->lock); 9800 ASSERT(block_group->io_ctl.inode == NULL); 9801 iput(inode); 9802 last = block_group->key.objectid + block_group->key.offset; 9803 btrfs_put_block_group(block_group); 9804 } 9805 } 9806 9807 /* 9808 * Must be called only after stopping all workers, since we could have block 9809 * group caching kthreads running, and therefore they could race with us if we 9810 * freed the block groups before stopping them. 9811 */ 9812 int btrfs_free_block_groups(struct btrfs_fs_info *info) 9813 { 9814 struct btrfs_block_group_cache *block_group; 9815 struct btrfs_space_info *space_info; 9816 struct btrfs_caching_control *caching_ctl; 9817 struct rb_node *n; 9818 9819 down_write(&info->commit_root_sem); 9820 while (!list_empty(&info->caching_block_groups)) { 9821 caching_ctl = list_entry(info->caching_block_groups.next, 9822 struct btrfs_caching_control, list); 9823 list_del(&caching_ctl->list); 9824 put_caching_control(caching_ctl); 9825 } 9826 up_write(&info->commit_root_sem); 9827 9828 spin_lock(&info->unused_bgs_lock); 9829 while (!list_empty(&info->unused_bgs)) { 9830 block_group = list_first_entry(&info->unused_bgs, 9831 struct btrfs_block_group_cache, 9832 bg_list); 9833 list_del_init(&block_group->bg_list); 9834 btrfs_put_block_group(block_group); 9835 } 9836 spin_unlock(&info->unused_bgs_lock); 9837 9838 spin_lock(&info->block_group_cache_lock); 9839 while ((n = rb_last(&info->block_group_cache_tree)) != NULL) { 9840 block_group = rb_entry(n, struct btrfs_block_group_cache, 9841 cache_node); 9842 rb_erase(&block_group->cache_node, 9843 &info->block_group_cache_tree); 9844 RB_CLEAR_NODE(&block_group->cache_node); 9845 spin_unlock(&info->block_group_cache_lock); 9846 9847 down_write(&block_group->space_info->groups_sem); 9848 list_del(&block_group->list); 9849 up_write(&block_group->space_info->groups_sem); 9850 9851 /* 9852 * We haven't cached this block group, which means we could 9853 * possibly have excluded extents on this block group. 9854 */ 9855 if (block_group->cached == BTRFS_CACHE_NO || 9856 block_group->cached == BTRFS_CACHE_ERROR) 9857 free_excluded_extents(info, block_group); 9858 9859 btrfs_remove_free_space_cache(block_group); 9860 ASSERT(block_group->cached != BTRFS_CACHE_STARTED); 9861 ASSERT(list_empty(&block_group->dirty_list)); 9862 ASSERT(list_empty(&block_group->io_list)); 9863 ASSERT(list_empty(&block_group->bg_list)); 9864 ASSERT(atomic_read(&block_group->count) == 1); 9865 btrfs_put_block_group(block_group); 9866 9867 spin_lock(&info->block_group_cache_lock); 9868 } 9869 spin_unlock(&info->block_group_cache_lock); 9870 9871 /* now that all the block groups are freed, go through and 9872 * free all the space_info structs. This is only called during 9873 * the final stages of unmount, and so we know nobody is 9874 * using them. We call synchronize_rcu() once before we start, 9875 * just to be on the safe side. 9876 */ 9877 synchronize_rcu(); 9878 9879 release_global_block_rsv(info); 9880 9881 while (!list_empty(&info->space_info)) { 9882 int i; 9883 9884 space_info = list_entry(info->space_info.next, 9885 struct btrfs_space_info, 9886 list); 9887 9888 /* 9889 * Do not hide this behind enospc_debug, this is actually 9890 * important and indicates a real bug if this happens. 9891 */ 9892 if (WARN_ON(space_info->bytes_pinned > 0 || 9893 space_info->bytes_reserved > 0 || 9894 space_info->bytes_may_use > 0)) 9895 dump_space_info(info, space_info, 0, 0); 9896 list_del(&space_info->list); 9897 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) { 9898 struct kobject *kobj; 9899 kobj = space_info->block_group_kobjs[i]; 9900 space_info->block_group_kobjs[i] = NULL; 9901 if (kobj) { 9902 kobject_del(kobj); 9903 kobject_put(kobj); 9904 } 9905 } 9906 kobject_del(&space_info->kobj); 9907 kobject_put(&space_info->kobj); 9908 } 9909 return 0; 9910 } 9911 9912 static void link_block_group(struct btrfs_block_group_cache *cache) 9913 { 9914 struct btrfs_space_info *space_info = cache->space_info; 9915 int index = get_block_group_index(cache); 9916 bool first = false; 9917 9918 down_write(&space_info->groups_sem); 9919 if (list_empty(&space_info->block_groups[index])) 9920 first = true; 9921 list_add_tail(&cache->list, &space_info->block_groups[index]); 9922 up_write(&space_info->groups_sem); 9923 9924 if (first) { 9925 struct raid_kobject *rkobj; 9926 int ret; 9927 9928 rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS); 9929 if (!rkobj) 9930 goto out_err; 9931 rkobj->raid_type = index; 9932 kobject_init(&rkobj->kobj, &btrfs_raid_ktype); 9933 ret = kobject_add(&rkobj->kobj, &space_info->kobj, 9934 "%s", get_raid_name(index)); 9935 if (ret) { 9936 kobject_put(&rkobj->kobj); 9937 goto out_err; 9938 } 9939 space_info->block_group_kobjs[index] = &rkobj->kobj; 9940 } 9941 9942 return; 9943 out_err: 9944 btrfs_warn(cache->fs_info, 9945 "failed to add kobject for block cache, ignoring"); 9946 } 9947 9948 static struct btrfs_block_group_cache * 9949 btrfs_create_block_group_cache(struct btrfs_fs_info *fs_info, 9950 u64 start, u64 size) 9951 { 9952 struct btrfs_block_group_cache *cache; 9953 9954 cache = kzalloc(sizeof(*cache), GFP_NOFS); 9955 if (!cache) 9956 return NULL; 9957 9958 cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl), 9959 GFP_NOFS); 9960 if (!cache->free_space_ctl) { 9961 kfree(cache); 9962 return NULL; 9963 } 9964 9965 cache->key.objectid = start; 9966 cache->key.offset = size; 9967 cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY; 9968 9969 cache->fs_info = fs_info; 9970 cache->full_stripe_len = btrfs_full_stripe_len(fs_info, start); 9971 set_free_space_tree_thresholds(cache); 9972 9973 atomic_set(&cache->count, 1); 9974 spin_lock_init(&cache->lock); 9975 init_rwsem(&cache->data_rwsem); 9976 INIT_LIST_HEAD(&cache->list); 9977 INIT_LIST_HEAD(&cache->cluster_list); 9978 INIT_LIST_HEAD(&cache->bg_list); 9979 INIT_LIST_HEAD(&cache->ro_list); 9980 INIT_LIST_HEAD(&cache->dirty_list); 9981 INIT_LIST_HEAD(&cache->io_list); 9982 btrfs_init_free_space_ctl(cache); 9983 atomic_set(&cache->trimming, 0); 9984 mutex_init(&cache->free_space_lock); 9985 btrfs_init_full_stripe_locks_tree(&cache->full_stripe_locks_root); 9986 9987 return cache; 9988 } 9989 9990 int btrfs_read_block_groups(struct btrfs_fs_info *info) 9991 { 9992 struct btrfs_path *path; 9993 int ret; 9994 struct btrfs_block_group_cache *cache; 9995 struct btrfs_space_info *space_info; 9996 struct btrfs_key key; 9997 struct btrfs_key found_key; 9998 struct extent_buffer *leaf; 9999 int need_clear = 0; 10000 u64 cache_gen; 10001 u64 feature; 10002 int mixed; 10003 10004 feature = btrfs_super_incompat_flags(info->super_copy); 10005 mixed = !!(feature & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS); 10006 10007 key.objectid = 0; 10008 key.offset = 0; 10009 key.type = BTRFS_BLOCK_GROUP_ITEM_KEY; 10010 path = btrfs_alloc_path(); 10011 if (!path) 10012 return -ENOMEM; 10013 path->reada = READA_FORWARD; 10014 10015 cache_gen = btrfs_super_cache_generation(info->super_copy); 10016 if (btrfs_test_opt(info, SPACE_CACHE) && 10017 btrfs_super_generation(info->super_copy) != cache_gen) 10018 need_clear = 1; 10019 if (btrfs_test_opt(info, CLEAR_CACHE)) 10020 need_clear = 1; 10021 10022 while (1) { 10023 ret = find_first_block_group(info, path, &key); 10024 if (ret > 0) 10025 break; 10026 if (ret != 0) 10027 goto error; 10028 10029 leaf = path->nodes[0]; 10030 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]); 10031 10032 cache = btrfs_create_block_group_cache(info, found_key.objectid, 10033 found_key.offset); 10034 if (!cache) { 10035 ret = -ENOMEM; 10036 goto error; 10037 } 10038 10039 if (need_clear) { 10040 /* 10041 * When we mount with old space cache, we need to 10042 * set BTRFS_DC_CLEAR and set dirty flag. 10043 * 10044 * a) Setting 'BTRFS_DC_CLEAR' makes sure that we 10045 * truncate the old free space cache inode and 10046 * setup a new one. 10047 * b) Setting 'dirty flag' makes sure that we flush 10048 * the new space cache info onto disk. 10049 */ 10050 if (btrfs_test_opt(info, SPACE_CACHE)) 10051 cache->disk_cache_state = BTRFS_DC_CLEAR; 10052 } 10053 10054 read_extent_buffer(leaf, &cache->item, 10055 btrfs_item_ptr_offset(leaf, path->slots[0]), 10056 sizeof(cache->item)); 10057 cache->flags = btrfs_block_group_flags(&cache->item); 10058 if (!mixed && 10059 ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) && 10060 (cache->flags & BTRFS_BLOCK_GROUP_DATA))) { 10061 btrfs_err(info, 10062 "bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups", 10063 cache->key.objectid); 10064 ret = -EINVAL; 10065 goto error; 10066 } 10067 10068 key.objectid = found_key.objectid + found_key.offset; 10069 btrfs_release_path(path); 10070 10071 /* 10072 * We need to exclude the super stripes now so that the space 10073 * info has super bytes accounted for, otherwise we'll think 10074 * we have more space than we actually do. 10075 */ 10076 ret = exclude_super_stripes(info, cache); 10077 if (ret) { 10078 /* 10079 * We may have excluded something, so call this just in 10080 * case. 10081 */ 10082 free_excluded_extents(info, cache); 10083 btrfs_put_block_group(cache); 10084 goto error; 10085 } 10086 10087 /* 10088 * check for two cases, either we are full, and therefore 10089 * don't need to bother with the caching work since we won't 10090 * find any space, or we are empty, and we can just add all 10091 * the space in and be done with it. This saves us _alot_ of 10092 * time, particularly in the full case. 10093 */ 10094 if (found_key.offset == btrfs_block_group_used(&cache->item)) { 10095 cache->last_byte_to_unpin = (u64)-1; 10096 cache->cached = BTRFS_CACHE_FINISHED; 10097 free_excluded_extents(info, cache); 10098 } else if (btrfs_block_group_used(&cache->item) == 0) { 10099 cache->last_byte_to_unpin = (u64)-1; 10100 cache->cached = BTRFS_CACHE_FINISHED; 10101 add_new_free_space(cache, info, 10102 found_key.objectid, 10103 found_key.objectid + 10104 found_key.offset); 10105 free_excluded_extents(info, cache); 10106 } 10107 10108 ret = btrfs_add_block_group_cache(info, cache); 10109 if (ret) { 10110 btrfs_remove_free_space_cache(cache); 10111 btrfs_put_block_group(cache); 10112 goto error; 10113 } 10114 10115 trace_btrfs_add_block_group(info, cache, 0); 10116 update_space_info(info, cache->flags, found_key.offset, 10117 btrfs_block_group_used(&cache->item), 10118 cache->bytes_super, &space_info); 10119 10120 cache->space_info = space_info; 10121 10122 link_block_group(cache); 10123 10124 set_avail_alloc_bits(info, cache->flags); 10125 if (btrfs_chunk_readonly(info, cache->key.objectid)) { 10126 inc_block_group_ro(cache, 1); 10127 } else if (btrfs_block_group_used(&cache->item) == 0) { 10128 spin_lock(&info->unused_bgs_lock); 10129 /* Should always be true but just in case. */ 10130 if (list_empty(&cache->bg_list)) { 10131 btrfs_get_block_group(cache); 10132 list_add_tail(&cache->bg_list, 10133 &info->unused_bgs); 10134 } 10135 spin_unlock(&info->unused_bgs_lock); 10136 } 10137 } 10138 10139 list_for_each_entry_rcu(space_info, &info->space_info, list) { 10140 if (!(get_alloc_profile(info, space_info->flags) & 10141 (BTRFS_BLOCK_GROUP_RAID10 | 10142 BTRFS_BLOCK_GROUP_RAID1 | 10143 BTRFS_BLOCK_GROUP_RAID5 | 10144 BTRFS_BLOCK_GROUP_RAID6 | 10145 BTRFS_BLOCK_GROUP_DUP))) 10146 continue; 10147 /* 10148 * avoid allocating from un-mirrored block group if there are 10149 * mirrored block groups. 10150 */ 10151 list_for_each_entry(cache, 10152 &space_info->block_groups[BTRFS_RAID_RAID0], 10153 list) 10154 inc_block_group_ro(cache, 1); 10155 list_for_each_entry(cache, 10156 &space_info->block_groups[BTRFS_RAID_SINGLE], 10157 list) 10158 inc_block_group_ro(cache, 1); 10159 } 10160 10161 init_global_block_rsv(info); 10162 ret = 0; 10163 error: 10164 btrfs_free_path(path); 10165 return ret; 10166 } 10167 10168 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans, 10169 struct btrfs_fs_info *fs_info) 10170 { 10171 struct btrfs_block_group_cache *block_group, *tmp; 10172 struct btrfs_root *extent_root = fs_info->extent_root; 10173 struct btrfs_block_group_item item; 10174 struct btrfs_key key; 10175 int ret = 0; 10176 bool can_flush_pending_bgs = trans->can_flush_pending_bgs; 10177 10178 trans->can_flush_pending_bgs = false; 10179 list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) { 10180 if (ret) 10181 goto next; 10182 10183 spin_lock(&block_group->lock); 10184 memcpy(&item, &block_group->item, sizeof(item)); 10185 memcpy(&key, &block_group->key, sizeof(key)); 10186 spin_unlock(&block_group->lock); 10187 10188 ret = btrfs_insert_item(trans, extent_root, &key, &item, 10189 sizeof(item)); 10190 if (ret) 10191 btrfs_abort_transaction(trans, ret); 10192 ret = btrfs_finish_chunk_alloc(trans, fs_info, key.objectid, 10193 key.offset); 10194 if (ret) 10195 btrfs_abort_transaction(trans, ret); 10196 add_block_group_free_space(trans, fs_info, block_group); 10197 /* already aborted the transaction if it failed. */ 10198 next: 10199 list_del_init(&block_group->bg_list); 10200 } 10201 trans->can_flush_pending_bgs = can_flush_pending_bgs; 10202 } 10203 10204 int btrfs_make_block_group(struct btrfs_trans_handle *trans, 10205 struct btrfs_fs_info *fs_info, u64 bytes_used, 10206 u64 type, u64 chunk_offset, u64 size) 10207 { 10208 struct btrfs_block_group_cache *cache; 10209 int ret; 10210 10211 btrfs_set_log_full_commit(fs_info, trans); 10212 10213 cache = btrfs_create_block_group_cache(fs_info, chunk_offset, size); 10214 if (!cache) 10215 return -ENOMEM; 10216 10217 btrfs_set_block_group_used(&cache->item, bytes_used); 10218 btrfs_set_block_group_chunk_objectid(&cache->item, 10219 BTRFS_FIRST_CHUNK_TREE_OBJECTID); 10220 btrfs_set_block_group_flags(&cache->item, type); 10221 10222 cache->flags = type; 10223 cache->last_byte_to_unpin = (u64)-1; 10224 cache->cached = BTRFS_CACHE_FINISHED; 10225 cache->needs_free_space = 1; 10226 ret = exclude_super_stripes(fs_info, cache); 10227 if (ret) { 10228 /* 10229 * We may have excluded something, so call this just in 10230 * case. 10231 */ 10232 free_excluded_extents(fs_info, cache); 10233 btrfs_put_block_group(cache); 10234 return ret; 10235 } 10236 10237 add_new_free_space(cache, fs_info, chunk_offset, chunk_offset + size); 10238 10239 free_excluded_extents(fs_info, cache); 10240 10241 #ifdef CONFIG_BTRFS_DEBUG 10242 if (btrfs_should_fragment_free_space(cache)) { 10243 u64 new_bytes_used = size - bytes_used; 10244 10245 bytes_used += new_bytes_used >> 1; 10246 fragment_free_space(cache); 10247 } 10248 #endif 10249 /* 10250 * Ensure the corresponding space_info object is created and 10251 * assigned to our block group. We want our bg to be added to the rbtree 10252 * with its ->space_info set. 10253 */ 10254 cache->space_info = __find_space_info(fs_info, cache->flags); 10255 if (!cache->space_info) { 10256 ret = create_space_info(fs_info, cache->flags, 10257 &cache->space_info); 10258 if (ret) { 10259 btrfs_remove_free_space_cache(cache); 10260 btrfs_put_block_group(cache); 10261 return ret; 10262 } 10263 } 10264 10265 ret = btrfs_add_block_group_cache(fs_info, cache); 10266 if (ret) { 10267 btrfs_remove_free_space_cache(cache); 10268 btrfs_put_block_group(cache); 10269 return ret; 10270 } 10271 10272 /* 10273 * Now that our block group has its ->space_info set and is inserted in 10274 * the rbtree, update the space info's counters. 10275 */ 10276 trace_btrfs_add_block_group(fs_info, cache, 1); 10277 update_space_info(fs_info, cache->flags, size, bytes_used, 10278 cache->bytes_super, &cache->space_info); 10279 update_global_block_rsv(fs_info); 10280 10281 link_block_group(cache); 10282 10283 list_add_tail(&cache->bg_list, &trans->new_bgs); 10284 10285 set_avail_alloc_bits(fs_info, type); 10286 return 0; 10287 } 10288 10289 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags) 10290 { 10291 u64 extra_flags = chunk_to_extended(flags) & 10292 BTRFS_EXTENDED_PROFILE_MASK; 10293 10294 write_seqlock(&fs_info->profiles_lock); 10295 if (flags & BTRFS_BLOCK_GROUP_DATA) 10296 fs_info->avail_data_alloc_bits &= ~extra_flags; 10297 if (flags & BTRFS_BLOCK_GROUP_METADATA) 10298 fs_info->avail_metadata_alloc_bits &= ~extra_flags; 10299 if (flags & BTRFS_BLOCK_GROUP_SYSTEM) 10300 fs_info->avail_system_alloc_bits &= ~extra_flags; 10301 write_sequnlock(&fs_info->profiles_lock); 10302 } 10303 10304 int btrfs_remove_block_group(struct btrfs_trans_handle *trans, 10305 struct btrfs_fs_info *fs_info, u64 group_start, 10306 struct extent_map *em) 10307 { 10308 struct btrfs_root *root = fs_info->extent_root; 10309 struct btrfs_path *path; 10310 struct btrfs_block_group_cache *block_group; 10311 struct btrfs_free_cluster *cluster; 10312 struct btrfs_root *tree_root = fs_info->tree_root; 10313 struct btrfs_key key; 10314 struct inode *inode; 10315 struct kobject *kobj = NULL; 10316 int ret; 10317 int index; 10318 int factor; 10319 struct btrfs_caching_control *caching_ctl = NULL; 10320 bool remove_em; 10321 10322 block_group = btrfs_lookup_block_group(fs_info, group_start); 10323 BUG_ON(!block_group); 10324 BUG_ON(!block_group->ro); 10325 10326 /* 10327 * Free the reserved super bytes from this block group before 10328 * remove it. 10329 */ 10330 free_excluded_extents(fs_info, block_group); 10331 btrfs_free_ref_tree_range(fs_info, block_group->key.objectid, 10332 block_group->key.offset); 10333 10334 memcpy(&key, &block_group->key, sizeof(key)); 10335 index = get_block_group_index(block_group); 10336 if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP | 10337 BTRFS_BLOCK_GROUP_RAID1 | 10338 BTRFS_BLOCK_GROUP_RAID10)) 10339 factor = 2; 10340 else 10341 factor = 1; 10342 10343 /* make sure this block group isn't part of an allocation cluster */ 10344 cluster = &fs_info->data_alloc_cluster; 10345 spin_lock(&cluster->refill_lock); 10346 btrfs_return_cluster_to_free_space(block_group, cluster); 10347 spin_unlock(&cluster->refill_lock); 10348 10349 /* 10350 * make sure this block group isn't part of a metadata 10351 * allocation cluster 10352 */ 10353 cluster = &fs_info->meta_alloc_cluster; 10354 spin_lock(&cluster->refill_lock); 10355 btrfs_return_cluster_to_free_space(block_group, cluster); 10356 spin_unlock(&cluster->refill_lock); 10357 10358 path = btrfs_alloc_path(); 10359 if (!path) { 10360 ret = -ENOMEM; 10361 goto out; 10362 } 10363 10364 /* 10365 * get the inode first so any iput calls done for the io_list 10366 * aren't the final iput (no unlinks allowed now) 10367 */ 10368 inode = lookup_free_space_inode(fs_info, block_group, path); 10369 10370 mutex_lock(&trans->transaction->cache_write_mutex); 10371 /* 10372 * make sure our free spache cache IO is done before remove the 10373 * free space inode 10374 */ 10375 spin_lock(&trans->transaction->dirty_bgs_lock); 10376 if (!list_empty(&block_group->io_list)) { 10377 list_del_init(&block_group->io_list); 10378 10379 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode); 10380 10381 spin_unlock(&trans->transaction->dirty_bgs_lock); 10382 btrfs_wait_cache_io(trans, block_group, path); 10383 btrfs_put_block_group(block_group); 10384 spin_lock(&trans->transaction->dirty_bgs_lock); 10385 } 10386 10387 if (!list_empty(&block_group->dirty_list)) { 10388 list_del_init(&block_group->dirty_list); 10389 btrfs_put_block_group(block_group); 10390 } 10391 spin_unlock(&trans->transaction->dirty_bgs_lock); 10392 mutex_unlock(&trans->transaction->cache_write_mutex); 10393 10394 if (!IS_ERR(inode)) { 10395 ret = btrfs_orphan_add(trans, BTRFS_I(inode)); 10396 if (ret) { 10397 btrfs_add_delayed_iput(inode); 10398 goto out; 10399 } 10400 clear_nlink(inode); 10401 /* One for the block groups ref */ 10402 spin_lock(&block_group->lock); 10403 if (block_group->iref) { 10404 block_group->iref = 0; 10405 block_group->inode = NULL; 10406 spin_unlock(&block_group->lock); 10407 iput(inode); 10408 } else { 10409 spin_unlock(&block_group->lock); 10410 } 10411 /* One for our lookup ref */ 10412 btrfs_add_delayed_iput(inode); 10413 } 10414 10415 key.objectid = BTRFS_FREE_SPACE_OBJECTID; 10416 key.offset = block_group->key.objectid; 10417 key.type = 0; 10418 10419 ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1); 10420 if (ret < 0) 10421 goto out; 10422 if (ret > 0) 10423 btrfs_release_path(path); 10424 if (ret == 0) { 10425 ret = btrfs_del_item(trans, tree_root, path); 10426 if (ret) 10427 goto out; 10428 btrfs_release_path(path); 10429 } 10430 10431 spin_lock(&fs_info->block_group_cache_lock); 10432 rb_erase(&block_group->cache_node, 10433 &fs_info->block_group_cache_tree); 10434 RB_CLEAR_NODE(&block_group->cache_node); 10435 10436 if (fs_info->first_logical_byte == block_group->key.objectid) 10437 fs_info->first_logical_byte = (u64)-1; 10438 spin_unlock(&fs_info->block_group_cache_lock); 10439 10440 down_write(&block_group->space_info->groups_sem); 10441 /* 10442 * we must use list_del_init so people can check to see if they 10443 * are still on the list after taking the semaphore 10444 */ 10445 list_del_init(&block_group->list); 10446 if (list_empty(&block_group->space_info->block_groups[index])) { 10447 kobj = block_group->space_info->block_group_kobjs[index]; 10448 block_group->space_info->block_group_kobjs[index] = NULL; 10449 clear_avail_alloc_bits(fs_info, block_group->flags); 10450 } 10451 up_write(&block_group->space_info->groups_sem); 10452 if (kobj) { 10453 kobject_del(kobj); 10454 kobject_put(kobj); 10455 } 10456 10457 if (block_group->has_caching_ctl) 10458 caching_ctl = get_caching_control(block_group); 10459 if (block_group->cached == BTRFS_CACHE_STARTED) 10460 wait_block_group_cache_done(block_group); 10461 if (block_group->has_caching_ctl) { 10462 down_write(&fs_info->commit_root_sem); 10463 if (!caching_ctl) { 10464 struct btrfs_caching_control *ctl; 10465 10466 list_for_each_entry(ctl, 10467 &fs_info->caching_block_groups, list) 10468 if (ctl->block_group == block_group) { 10469 caching_ctl = ctl; 10470 refcount_inc(&caching_ctl->count); 10471 break; 10472 } 10473 } 10474 if (caching_ctl) 10475 list_del_init(&caching_ctl->list); 10476 up_write(&fs_info->commit_root_sem); 10477 if (caching_ctl) { 10478 /* Once for the caching bgs list and once for us. */ 10479 put_caching_control(caching_ctl); 10480 put_caching_control(caching_ctl); 10481 } 10482 } 10483 10484 spin_lock(&trans->transaction->dirty_bgs_lock); 10485 if (!list_empty(&block_group->dirty_list)) { 10486 WARN_ON(1); 10487 } 10488 if (!list_empty(&block_group->io_list)) { 10489 WARN_ON(1); 10490 } 10491 spin_unlock(&trans->transaction->dirty_bgs_lock); 10492 btrfs_remove_free_space_cache(block_group); 10493 10494 spin_lock(&block_group->space_info->lock); 10495 list_del_init(&block_group->ro_list); 10496 10497 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) { 10498 WARN_ON(block_group->space_info->total_bytes 10499 < block_group->key.offset); 10500 WARN_ON(block_group->space_info->bytes_readonly 10501 < block_group->key.offset); 10502 WARN_ON(block_group->space_info->disk_total 10503 < block_group->key.offset * factor); 10504 } 10505 block_group->space_info->total_bytes -= block_group->key.offset; 10506 block_group->space_info->bytes_readonly -= block_group->key.offset; 10507 block_group->space_info->disk_total -= block_group->key.offset * factor; 10508 10509 spin_unlock(&block_group->space_info->lock); 10510 10511 memcpy(&key, &block_group->key, sizeof(key)); 10512 10513 mutex_lock(&fs_info->chunk_mutex); 10514 if (!list_empty(&em->list)) { 10515 /* We're in the transaction->pending_chunks list. */ 10516 free_extent_map(em); 10517 } 10518 spin_lock(&block_group->lock); 10519 block_group->removed = 1; 10520 /* 10521 * At this point trimming can't start on this block group, because we 10522 * removed the block group from the tree fs_info->block_group_cache_tree 10523 * so no one can't find it anymore and even if someone already got this 10524 * block group before we removed it from the rbtree, they have already 10525 * incremented block_group->trimming - if they didn't, they won't find 10526 * any free space entries because we already removed them all when we 10527 * called btrfs_remove_free_space_cache(). 10528 * 10529 * And we must not remove the extent map from the fs_info->mapping_tree 10530 * to prevent the same logical address range and physical device space 10531 * ranges from being reused for a new block group. This is because our 10532 * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is 10533 * completely transactionless, so while it is trimming a range the 10534 * currently running transaction might finish and a new one start, 10535 * allowing for new block groups to be created that can reuse the same 10536 * physical device locations unless we take this special care. 10537 * 10538 * There may also be an implicit trim operation if the file system 10539 * is mounted with -odiscard. The same protections must remain 10540 * in place until the extents have been discarded completely when 10541 * the transaction commit has completed. 10542 */ 10543 remove_em = (atomic_read(&block_group->trimming) == 0); 10544 /* 10545 * Make sure a trimmer task always sees the em in the pinned_chunks list 10546 * if it sees block_group->removed == 1 (needs to lock block_group->lock 10547 * before checking block_group->removed). 10548 */ 10549 if (!remove_em) { 10550 /* 10551 * Our em might be in trans->transaction->pending_chunks which 10552 * is protected by fs_info->chunk_mutex ([lock|unlock]_chunks), 10553 * and so is the fs_info->pinned_chunks list. 10554 * 10555 * So at this point we must be holding the chunk_mutex to avoid 10556 * any races with chunk allocation (more specifically at 10557 * volumes.c:contains_pending_extent()), to ensure it always 10558 * sees the em, either in the pending_chunks list or in the 10559 * pinned_chunks list. 10560 */ 10561 list_move_tail(&em->list, &fs_info->pinned_chunks); 10562 } 10563 spin_unlock(&block_group->lock); 10564 10565 if (remove_em) { 10566 struct extent_map_tree *em_tree; 10567 10568 em_tree = &fs_info->mapping_tree.map_tree; 10569 write_lock(&em_tree->lock); 10570 /* 10571 * The em might be in the pending_chunks list, so make sure the 10572 * chunk mutex is locked, since remove_extent_mapping() will 10573 * delete us from that list. 10574 */ 10575 remove_extent_mapping(em_tree, em); 10576 write_unlock(&em_tree->lock); 10577 /* once for the tree */ 10578 free_extent_map(em); 10579 } 10580 10581 mutex_unlock(&fs_info->chunk_mutex); 10582 10583 ret = remove_block_group_free_space(trans, fs_info, block_group); 10584 if (ret) 10585 goto out; 10586 10587 btrfs_put_block_group(block_group); 10588 btrfs_put_block_group(block_group); 10589 10590 ret = btrfs_search_slot(trans, root, &key, path, -1, 1); 10591 if (ret > 0) 10592 ret = -EIO; 10593 if (ret < 0) 10594 goto out; 10595 10596 ret = btrfs_del_item(trans, root, path); 10597 out: 10598 btrfs_free_path(path); 10599 return ret; 10600 } 10601 10602 struct btrfs_trans_handle * 10603 btrfs_start_trans_remove_block_group(struct btrfs_fs_info *fs_info, 10604 const u64 chunk_offset) 10605 { 10606 struct extent_map_tree *em_tree = &fs_info->mapping_tree.map_tree; 10607 struct extent_map *em; 10608 struct map_lookup *map; 10609 unsigned int num_items; 10610 10611 read_lock(&em_tree->lock); 10612 em = lookup_extent_mapping(em_tree, chunk_offset, 1); 10613 read_unlock(&em_tree->lock); 10614 ASSERT(em && em->start == chunk_offset); 10615 10616 /* 10617 * We need to reserve 3 + N units from the metadata space info in order 10618 * to remove a block group (done at btrfs_remove_chunk() and at 10619 * btrfs_remove_block_group()), which are used for: 10620 * 10621 * 1 unit for adding the free space inode's orphan (located in the tree 10622 * of tree roots). 10623 * 1 unit for deleting the block group item (located in the extent 10624 * tree). 10625 * 1 unit for deleting the free space item (located in tree of tree 10626 * roots). 10627 * N units for deleting N device extent items corresponding to each 10628 * stripe (located in the device tree). 10629 * 10630 * In order to remove a block group we also need to reserve units in the 10631 * system space info in order to update the chunk tree (update one or 10632 * more device items and remove one chunk item), but this is done at 10633 * btrfs_remove_chunk() through a call to check_system_chunk(). 10634 */ 10635 map = em->map_lookup; 10636 num_items = 3 + map->num_stripes; 10637 free_extent_map(em); 10638 10639 return btrfs_start_transaction_fallback_global_rsv(fs_info->extent_root, 10640 num_items, 1); 10641 } 10642 10643 /* 10644 * Process the unused_bgs list and remove any that don't have any allocated 10645 * space inside of them. 10646 */ 10647 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info) 10648 { 10649 struct btrfs_block_group_cache *block_group; 10650 struct btrfs_space_info *space_info; 10651 struct btrfs_trans_handle *trans; 10652 int ret = 0; 10653 10654 if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags)) 10655 return; 10656 10657 spin_lock(&fs_info->unused_bgs_lock); 10658 while (!list_empty(&fs_info->unused_bgs)) { 10659 u64 start, end; 10660 int trimming; 10661 10662 block_group = list_first_entry(&fs_info->unused_bgs, 10663 struct btrfs_block_group_cache, 10664 bg_list); 10665 list_del_init(&block_group->bg_list); 10666 10667 space_info = block_group->space_info; 10668 10669 if (ret || btrfs_mixed_space_info(space_info)) { 10670 btrfs_put_block_group(block_group); 10671 continue; 10672 } 10673 spin_unlock(&fs_info->unused_bgs_lock); 10674 10675 mutex_lock(&fs_info->delete_unused_bgs_mutex); 10676 10677 /* Don't want to race with allocators so take the groups_sem */ 10678 down_write(&space_info->groups_sem); 10679 spin_lock(&block_group->lock); 10680 if (block_group->reserved || 10681 btrfs_block_group_used(&block_group->item) || 10682 block_group->ro || 10683 list_is_singular(&block_group->list)) { 10684 /* 10685 * We want to bail if we made new allocations or have 10686 * outstanding allocations in this block group. We do 10687 * the ro check in case balance is currently acting on 10688 * this block group. 10689 */ 10690 spin_unlock(&block_group->lock); 10691 up_write(&space_info->groups_sem); 10692 goto next; 10693 } 10694 spin_unlock(&block_group->lock); 10695 10696 /* We don't want to force the issue, only flip if it's ok. */ 10697 ret = inc_block_group_ro(block_group, 0); 10698 up_write(&space_info->groups_sem); 10699 if (ret < 0) { 10700 ret = 0; 10701 goto next; 10702 } 10703 10704 /* 10705 * Want to do this before we do anything else so we can recover 10706 * properly if we fail to join the transaction. 10707 */ 10708 trans = btrfs_start_trans_remove_block_group(fs_info, 10709 block_group->key.objectid); 10710 if (IS_ERR(trans)) { 10711 btrfs_dec_block_group_ro(block_group); 10712 ret = PTR_ERR(trans); 10713 goto next; 10714 } 10715 10716 /* 10717 * We could have pending pinned extents for this block group, 10718 * just delete them, we don't care about them anymore. 10719 */ 10720 start = block_group->key.objectid; 10721 end = start + block_group->key.offset - 1; 10722 /* 10723 * Hold the unused_bg_unpin_mutex lock to avoid racing with 10724 * btrfs_finish_extent_commit(). If we are at transaction N, 10725 * another task might be running finish_extent_commit() for the 10726 * previous transaction N - 1, and have seen a range belonging 10727 * to the block group in freed_extents[] before we were able to 10728 * clear the whole block group range from freed_extents[]. This 10729 * means that task can lookup for the block group after we 10730 * unpinned it from freed_extents[] and removed it, leading to 10731 * a BUG_ON() at btrfs_unpin_extent_range(). 10732 */ 10733 mutex_lock(&fs_info->unused_bg_unpin_mutex); 10734 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end, 10735 EXTENT_DIRTY); 10736 if (ret) { 10737 mutex_unlock(&fs_info->unused_bg_unpin_mutex); 10738 btrfs_dec_block_group_ro(block_group); 10739 goto end_trans; 10740 } 10741 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end, 10742 EXTENT_DIRTY); 10743 if (ret) { 10744 mutex_unlock(&fs_info->unused_bg_unpin_mutex); 10745 btrfs_dec_block_group_ro(block_group); 10746 goto end_trans; 10747 } 10748 mutex_unlock(&fs_info->unused_bg_unpin_mutex); 10749 10750 /* Reset pinned so btrfs_put_block_group doesn't complain */ 10751 spin_lock(&space_info->lock); 10752 spin_lock(&block_group->lock); 10753 10754 space_info->bytes_pinned -= block_group->pinned; 10755 space_info->bytes_readonly += block_group->pinned; 10756 percpu_counter_add(&space_info->total_bytes_pinned, 10757 -block_group->pinned); 10758 block_group->pinned = 0; 10759 10760 spin_unlock(&block_group->lock); 10761 spin_unlock(&space_info->lock); 10762 10763 /* DISCARD can flip during remount */ 10764 trimming = btrfs_test_opt(fs_info, DISCARD); 10765 10766 /* Implicit trim during transaction commit. */ 10767 if (trimming) 10768 btrfs_get_block_group_trimming(block_group); 10769 10770 /* 10771 * Btrfs_remove_chunk will abort the transaction if things go 10772 * horribly wrong. 10773 */ 10774 ret = btrfs_remove_chunk(trans, fs_info, 10775 block_group->key.objectid); 10776 10777 if (ret) { 10778 if (trimming) 10779 btrfs_put_block_group_trimming(block_group); 10780 goto end_trans; 10781 } 10782 10783 /* 10784 * If we're not mounted with -odiscard, we can just forget 10785 * about this block group. Otherwise we'll need to wait 10786 * until transaction commit to do the actual discard. 10787 */ 10788 if (trimming) { 10789 spin_lock(&fs_info->unused_bgs_lock); 10790 /* 10791 * A concurrent scrub might have added us to the list 10792 * fs_info->unused_bgs, so use a list_move operation 10793 * to add the block group to the deleted_bgs list. 10794 */ 10795 list_move(&block_group->bg_list, 10796 &trans->transaction->deleted_bgs); 10797 spin_unlock(&fs_info->unused_bgs_lock); 10798 btrfs_get_block_group(block_group); 10799 } 10800 end_trans: 10801 btrfs_end_transaction(trans); 10802 next: 10803 mutex_unlock(&fs_info->delete_unused_bgs_mutex); 10804 btrfs_put_block_group(block_group); 10805 spin_lock(&fs_info->unused_bgs_lock); 10806 } 10807 spin_unlock(&fs_info->unused_bgs_lock); 10808 } 10809 10810 int btrfs_init_space_info(struct btrfs_fs_info *fs_info) 10811 { 10812 struct btrfs_space_info *space_info; 10813 struct btrfs_super_block *disk_super; 10814 u64 features; 10815 u64 flags; 10816 int mixed = 0; 10817 int ret; 10818 10819 disk_super = fs_info->super_copy; 10820 if (!btrfs_super_root(disk_super)) 10821 return -EINVAL; 10822 10823 features = btrfs_super_incompat_flags(disk_super); 10824 if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS) 10825 mixed = 1; 10826 10827 flags = BTRFS_BLOCK_GROUP_SYSTEM; 10828 ret = create_space_info(fs_info, flags, &space_info); 10829 if (ret) 10830 goto out; 10831 10832 if (mixed) { 10833 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA; 10834 ret = create_space_info(fs_info, flags, &space_info); 10835 } else { 10836 flags = BTRFS_BLOCK_GROUP_METADATA; 10837 ret = create_space_info(fs_info, flags, &space_info); 10838 if (ret) 10839 goto out; 10840 10841 flags = BTRFS_BLOCK_GROUP_DATA; 10842 ret = create_space_info(fs_info, flags, &space_info); 10843 } 10844 out: 10845 return ret; 10846 } 10847 10848 int btrfs_error_unpin_extent_range(struct btrfs_fs_info *fs_info, 10849 u64 start, u64 end) 10850 { 10851 return unpin_extent_range(fs_info, start, end, false); 10852 } 10853 10854 /* 10855 * It used to be that old block groups would be left around forever. 10856 * Iterating over them would be enough to trim unused space. Since we 10857 * now automatically remove them, we also need to iterate over unallocated 10858 * space. 10859 * 10860 * We don't want a transaction for this since the discard may take a 10861 * substantial amount of time. We don't require that a transaction be 10862 * running, but we do need to take a running transaction into account 10863 * to ensure that we're not discarding chunks that were released in 10864 * the current transaction. 10865 * 10866 * Holding the chunks lock will prevent other threads from allocating 10867 * or releasing chunks, but it won't prevent a running transaction 10868 * from committing and releasing the memory that the pending chunks 10869 * list head uses. For that, we need to take a reference to the 10870 * transaction. 10871 */ 10872 static int btrfs_trim_free_extents(struct btrfs_device *device, 10873 u64 minlen, u64 *trimmed) 10874 { 10875 u64 start = 0, len = 0; 10876 int ret; 10877 10878 *trimmed = 0; 10879 10880 /* Not writeable = nothing to do. */ 10881 if (!test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state)) 10882 return 0; 10883 10884 /* No free space = nothing to do. */ 10885 if (device->total_bytes <= device->bytes_used) 10886 return 0; 10887 10888 ret = 0; 10889 10890 while (1) { 10891 struct btrfs_fs_info *fs_info = device->fs_info; 10892 struct btrfs_transaction *trans; 10893 u64 bytes; 10894 10895 ret = mutex_lock_interruptible(&fs_info->chunk_mutex); 10896 if (ret) 10897 return ret; 10898 10899 down_read(&fs_info->commit_root_sem); 10900 10901 spin_lock(&fs_info->trans_lock); 10902 trans = fs_info->running_transaction; 10903 if (trans) 10904 refcount_inc(&trans->use_count); 10905 spin_unlock(&fs_info->trans_lock); 10906 10907 ret = find_free_dev_extent_start(trans, device, minlen, start, 10908 &start, &len); 10909 if (trans) 10910 btrfs_put_transaction(trans); 10911 10912 if (ret) { 10913 up_read(&fs_info->commit_root_sem); 10914 mutex_unlock(&fs_info->chunk_mutex); 10915 if (ret == -ENOSPC) 10916 ret = 0; 10917 break; 10918 } 10919 10920 ret = btrfs_issue_discard(device->bdev, start, len, &bytes); 10921 up_read(&fs_info->commit_root_sem); 10922 mutex_unlock(&fs_info->chunk_mutex); 10923 10924 if (ret) 10925 break; 10926 10927 start += len; 10928 *trimmed += bytes; 10929 10930 if (fatal_signal_pending(current)) { 10931 ret = -ERESTARTSYS; 10932 break; 10933 } 10934 10935 cond_resched(); 10936 } 10937 10938 return ret; 10939 } 10940 10941 int btrfs_trim_fs(struct btrfs_fs_info *fs_info, struct fstrim_range *range) 10942 { 10943 struct btrfs_block_group_cache *cache = NULL; 10944 struct btrfs_device *device; 10945 struct list_head *devices; 10946 u64 group_trimmed; 10947 u64 start; 10948 u64 end; 10949 u64 trimmed = 0; 10950 u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy); 10951 int ret = 0; 10952 10953 /* 10954 * try to trim all FS space, our block group may start from non-zero. 10955 */ 10956 if (range->len == total_bytes) 10957 cache = btrfs_lookup_first_block_group(fs_info, range->start); 10958 else 10959 cache = btrfs_lookup_block_group(fs_info, range->start); 10960 10961 while (cache) { 10962 if (cache->key.objectid >= (range->start + range->len)) { 10963 btrfs_put_block_group(cache); 10964 break; 10965 } 10966 10967 start = max(range->start, cache->key.objectid); 10968 end = min(range->start + range->len, 10969 cache->key.objectid + cache->key.offset); 10970 10971 if (end - start >= range->minlen) { 10972 if (!block_group_cache_done(cache)) { 10973 ret = cache_block_group(cache, 0); 10974 if (ret) { 10975 btrfs_put_block_group(cache); 10976 break; 10977 } 10978 ret = wait_block_group_cache_done(cache); 10979 if (ret) { 10980 btrfs_put_block_group(cache); 10981 break; 10982 } 10983 } 10984 ret = btrfs_trim_block_group(cache, 10985 &group_trimmed, 10986 start, 10987 end, 10988 range->minlen); 10989 10990 trimmed += group_trimmed; 10991 if (ret) { 10992 btrfs_put_block_group(cache); 10993 break; 10994 } 10995 } 10996 10997 cache = next_block_group(fs_info, cache); 10998 } 10999 11000 mutex_lock(&fs_info->fs_devices->device_list_mutex); 11001 devices = &fs_info->fs_devices->alloc_list; 11002 list_for_each_entry(device, devices, dev_alloc_list) { 11003 ret = btrfs_trim_free_extents(device, range->minlen, 11004 &group_trimmed); 11005 if (ret) 11006 break; 11007 11008 trimmed += group_trimmed; 11009 } 11010 mutex_unlock(&fs_info->fs_devices->device_list_mutex); 11011 11012 range->len = trimmed; 11013 return ret; 11014 } 11015 11016 /* 11017 * btrfs_{start,end}_write_no_snapshotting() are similar to 11018 * mnt_{want,drop}_write(), they are used to prevent some tasks from writing 11019 * data into the page cache through nocow before the subvolume is snapshoted, 11020 * but flush the data into disk after the snapshot creation, or to prevent 11021 * operations while snapshotting is ongoing and that cause the snapshot to be 11022 * inconsistent (writes followed by expanding truncates for example). 11023 */ 11024 void btrfs_end_write_no_snapshotting(struct btrfs_root *root) 11025 { 11026 percpu_counter_dec(&root->subv_writers->counter); 11027 /* 11028 * Make sure counter is updated before we wake up waiters. 11029 */ 11030 smp_mb(); 11031 if (waitqueue_active(&root->subv_writers->wait)) 11032 wake_up(&root->subv_writers->wait); 11033 } 11034 11035 int btrfs_start_write_no_snapshotting(struct btrfs_root *root) 11036 { 11037 if (atomic_read(&root->will_be_snapshotted)) 11038 return 0; 11039 11040 percpu_counter_inc(&root->subv_writers->counter); 11041 /* 11042 * Make sure counter is updated before we check for snapshot creation. 11043 */ 11044 smp_mb(); 11045 if (atomic_read(&root->will_be_snapshotted)) { 11046 btrfs_end_write_no_snapshotting(root); 11047 return 0; 11048 } 11049 return 1; 11050 } 11051 11052 void btrfs_wait_for_snapshot_creation(struct btrfs_root *root) 11053 { 11054 while (true) { 11055 int ret; 11056 11057 ret = btrfs_start_write_no_snapshotting(root); 11058 if (ret) 11059 break; 11060 wait_on_atomic_t(&root->will_be_snapshotted, atomic_t_wait, 11061 TASK_UNINTERRUPTIBLE); 11062 } 11063 } 11064