1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Copyright (C) 2007 Oracle. All rights reserved. 4 */ 5 6 #include <linux/fs.h> 7 #include <linux/slab.h> 8 #include <linux/sched.h> 9 #include <linux/sched/mm.h> 10 #include <linux/writeback.h> 11 #include <linux/pagemap.h> 12 #include <linux/blkdev.h> 13 #include <linux/uuid.h> 14 #include <linux/timekeeping.h> 15 #include "misc.h" 16 #include "ctree.h" 17 #include "disk-io.h" 18 #include "transaction.h" 19 #include "locking.h" 20 #include "tree-log.h" 21 #include "volumes.h" 22 #include "dev-replace.h" 23 #include "qgroup.h" 24 #include "block-group.h" 25 #include "space-info.h" 26 #include "fs.h" 27 #include "accessors.h" 28 #include "extent-tree.h" 29 #include "root-tree.h" 30 #include "dir-item.h" 31 #include "uuid-tree.h" 32 #include "ioctl.h" 33 #include "relocation.h" 34 #include "scrub.h" 35 #include "ordered-data.h" 36 #include "delayed-inode.h" 37 38 static struct kmem_cache *btrfs_trans_handle_cachep; 39 40 /* 41 * Transaction states and transitions 42 * 43 * No running transaction (fs tree blocks are not modified) 44 * | 45 * | To next stage: 46 * | Call start_transaction() variants. Except btrfs_join_transaction_nostart(). 47 * V 48 * Transaction N [[TRANS_STATE_RUNNING]] 49 * | 50 * | New trans handles can be attached to transaction N by calling all 51 * | start_transaction() variants. 52 * | 53 * | To next stage: 54 * | Call btrfs_commit_transaction() on any trans handle attached to 55 * | transaction N 56 * V 57 * Transaction N [[TRANS_STATE_COMMIT_PREP]] 58 * | 59 * | If there are simultaneous calls to btrfs_commit_transaction() one will win 60 * | the race and the rest will wait for the winner to commit the transaction. 61 * | 62 * | The winner will wait for previous running transaction to completely finish 63 * | if there is one. 64 * | 65 * Transaction N [[TRANS_STATE_COMMIT_START]] 66 * | 67 * | Then one of the following happens: 68 * | - Wait for all other trans handle holders to release. 69 * | The btrfs_commit_transaction() caller will do the commit work. 70 * | - Wait for current transaction to be committed by others. 71 * | Other btrfs_commit_transaction() caller will do the commit work. 72 * | 73 * | At this stage, only btrfs_join_transaction*() variants can attach 74 * | to this running transaction. 75 * | All other variants will wait for current one to finish and attach to 76 * | transaction N+1. 77 * | 78 * | To next stage: 79 * | Caller is chosen to commit transaction N, and all other trans handle 80 * | haven been released. 81 * V 82 * Transaction N [[TRANS_STATE_COMMIT_DOING]] 83 * | 84 * | The heavy lifting transaction work is started. 85 * | From running delayed refs (modifying extent tree) to creating pending 86 * | snapshots, running qgroups. 87 * | In short, modify supporting trees to reflect modifications of subvolume 88 * | trees. 89 * | 90 * | At this stage, all start_transaction() calls will wait for this 91 * | transaction to finish and attach to transaction N+1. 92 * | 93 * | To next stage: 94 * | Until all supporting trees are updated. 95 * V 96 * Transaction N [[TRANS_STATE_UNBLOCKED]] 97 * | Transaction N+1 98 * | All needed trees are modified, thus we only [[TRANS_STATE_RUNNING]] 99 * | need to write them back to disk and update | 100 * | super blocks. | 101 * | | 102 * | At this stage, new transaction is allowed to | 103 * | start. | 104 * | All new start_transaction() calls will be | 105 * | attached to transid N+1. | 106 * | | 107 * | To next stage: | 108 * | Until all tree blocks and super blocks are | 109 * | written to block devices | 110 * V | 111 * Transaction N [[TRANS_STATE_COMPLETED]] V 112 * All tree blocks and super blocks are written. Transaction N+1 113 * This transaction is finished and all its [[TRANS_STATE_COMMIT_START]] 114 * data structures will be cleaned up. | Life goes on 115 */ 116 static const unsigned int btrfs_blocked_trans_types[TRANS_STATE_MAX] = { 117 [TRANS_STATE_RUNNING] = 0U, 118 [TRANS_STATE_COMMIT_PREP] = 0U, 119 [TRANS_STATE_COMMIT_START] = (__TRANS_START | __TRANS_ATTACH), 120 [TRANS_STATE_COMMIT_DOING] = (__TRANS_START | 121 __TRANS_ATTACH | 122 __TRANS_JOIN | 123 __TRANS_JOIN_NOSTART), 124 [TRANS_STATE_UNBLOCKED] = (__TRANS_START | 125 __TRANS_ATTACH | 126 __TRANS_JOIN | 127 __TRANS_JOIN_NOLOCK | 128 __TRANS_JOIN_NOSTART), 129 [TRANS_STATE_SUPER_COMMITTED] = (__TRANS_START | 130 __TRANS_ATTACH | 131 __TRANS_JOIN | 132 __TRANS_JOIN_NOLOCK | 133 __TRANS_JOIN_NOSTART), 134 [TRANS_STATE_COMPLETED] = (__TRANS_START | 135 __TRANS_ATTACH | 136 __TRANS_JOIN | 137 __TRANS_JOIN_NOLOCK | 138 __TRANS_JOIN_NOSTART), 139 }; 140 141 void btrfs_put_transaction(struct btrfs_transaction *transaction) 142 { 143 if (refcount_dec_and_test(&transaction->use_count)) { 144 BUG_ON(!list_empty(&transaction->list)); 145 WARN_ON(!xa_empty(&transaction->delayed_refs.head_refs)); 146 WARN_ON(!xa_empty(&transaction->delayed_refs.dirty_extents)); 147 if (transaction->delayed_refs.pending_csums) 148 btrfs_err(transaction->fs_info, 149 "pending csums is %llu", 150 transaction->delayed_refs.pending_csums); 151 /* 152 * If any block groups are found in ->deleted_bgs then it's 153 * because the transaction was aborted and a commit did not 154 * happen (things failed before writing the new superblock 155 * and calling btrfs_finish_extent_commit()), so we can not 156 * discard the physical locations of the block groups. 157 */ 158 while (!list_empty(&transaction->deleted_bgs)) { 159 struct btrfs_block_group *cache; 160 161 cache = list_first_entry(&transaction->deleted_bgs, 162 struct btrfs_block_group, 163 bg_list); 164 /* 165 * Not strictly necessary to lock, as no other task will be using a 166 * block_group on the deleted_bgs list during a transaction abort. 167 */ 168 spin_lock(&transaction->fs_info->unused_bgs_lock); 169 list_del_init(&cache->bg_list); 170 spin_unlock(&transaction->fs_info->unused_bgs_lock); 171 btrfs_unfreeze_block_group(cache); 172 btrfs_put_block_group(cache); 173 } 174 WARN_ON(!list_empty(&transaction->dev_update_list)); 175 kfree(transaction); 176 } 177 } 178 179 static noinline void switch_commit_roots(struct btrfs_trans_handle *trans) 180 { 181 struct btrfs_transaction *cur_trans = trans->transaction; 182 struct btrfs_fs_info *fs_info = trans->fs_info; 183 struct btrfs_root *root, *tmp; 184 185 /* 186 * At this point no one can be using this transaction to modify any tree 187 * and no one can start another transaction to modify any tree either. 188 */ 189 ASSERT(cur_trans->state == TRANS_STATE_COMMIT_DOING, 190 "cur_trans->state=%d", cur_trans->state); 191 192 down_write(&fs_info->commit_root_sem); 193 194 if (test_bit(BTRFS_FS_RELOC_RUNNING, &fs_info->flags)) 195 fs_info->last_reloc_trans = trans->transid; 196 197 list_for_each_entry_safe(root, tmp, &cur_trans->switch_commits, 198 dirty_list) { 199 list_del_init(&root->dirty_list); 200 free_extent_buffer(root->commit_root); 201 root->commit_root = btrfs_root_node(root); 202 btrfs_extent_io_tree_release(&root->dirty_log_pages); 203 btrfs_qgroup_clean_swapped_blocks(root); 204 } 205 206 /* We can free old roots now. */ 207 spin_lock(&cur_trans->dropped_roots_lock); 208 while (!list_empty(&cur_trans->dropped_roots)) { 209 root = list_first_entry(&cur_trans->dropped_roots, 210 struct btrfs_root, root_list); 211 list_del_init(&root->root_list); 212 spin_unlock(&cur_trans->dropped_roots_lock); 213 btrfs_free_log(trans, root); 214 btrfs_drop_and_free_fs_root(fs_info, root); 215 spin_lock(&cur_trans->dropped_roots_lock); 216 } 217 spin_unlock(&cur_trans->dropped_roots_lock); 218 219 up_write(&fs_info->commit_root_sem); 220 } 221 222 static inline void extwriter_counter_inc(struct btrfs_transaction *trans, 223 unsigned int type) 224 { 225 if (type & TRANS_EXTWRITERS) 226 atomic_inc(&trans->num_extwriters); 227 } 228 229 static inline void extwriter_counter_dec(struct btrfs_transaction *trans, 230 unsigned int type) 231 { 232 if (type & TRANS_EXTWRITERS) 233 atomic_dec(&trans->num_extwriters); 234 } 235 236 static inline void extwriter_counter_init(struct btrfs_transaction *trans, 237 unsigned int type) 238 { 239 atomic_set(&trans->num_extwriters, ((type & TRANS_EXTWRITERS) ? 1 : 0)); 240 } 241 242 static inline int extwriter_counter_read(struct btrfs_transaction *trans) 243 { 244 return atomic_read(&trans->num_extwriters); 245 } 246 247 /* 248 * To be called after doing the chunk btree updates right after allocating a new 249 * chunk (after btrfs_chunk_alloc_add_chunk_item() is called), when removing a 250 * chunk after all chunk btree updates and after finishing the second phase of 251 * chunk allocation (btrfs_create_pending_block_groups()) in case some block 252 * group had its chunk item insertion delayed to the second phase. 253 */ 254 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans) 255 { 256 struct btrfs_fs_info *fs_info = trans->fs_info; 257 258 if (!trans->chunk_bytes_reserved) 259 return; 260 261 btrfs_block_rsv_release(fs_info, &fs_info->chunk_block_rsv, 262 trans->chunk_bytes_reserved, NULL); 263 trans->chunk_bytes_reserved = 0; 264 } 265 266 /* 267 * either allocate a new transaction or hop into the existing one 268 */ 269 static noinline int join_transaction(struct btrfs_fs_info *fs_info, 270 unsigned int type) 271 { 272 struct btrfs_transaction *cur_trans; 273 274 spin_lock(&fs_info->trans_lock); 275 loop: 276 /* The file system has been taken offline. No new transactions. */ 277 if (BTRFS_FS_ERROR(fs_info)) { 278 spin_unlock(&fs_info->trans_lock); 279 return -EROFS; 280 } 281 282 cur_trans = fs_info->running_transaction; 283 if (cur_trans) { 284 if (TRANS_ABORTED(cur_trans)) { 285 const int abort_error = cur_trans->aborted; 286 287 spin_unlock(&fs_info->trans_lock); 288 return abort_error; 289 } 290 if (btrfs_blocked_trans_types[cur_trans->state] & type) { 291 spin_unlock(&fs_info->trans_lock); 292 return -EBUSY; 293 } 294 refcount_inc(&cur_trans->use_count); 295 atomic_inc(&cur_trans->num_writers); 296 extwriter_counter_inc(cur_trans, type); 297 spin_unlock(&fs_info->trans_lock); 298 btrfs_lockdep_acquire(fs_info, btrfs_trans_num_writers); 299 btrfs_lockdep_acquire(fs_info, btrfs_trans_num_extwriters); 300 return 0; 301 } 302 spin_unlock(&fs_info->trans_lock); 303 304 /* 305 * If we are ATTACH or TRANS_JOIN_NOSTART, we just want to catch the 306 * current transaction, and commit it. If there is no transaction, just 307 * return ENOENT. 308 */ 309 if (type == TRANS_ATTACH || type == TRANS_JOIN_NOSTART) 310 return -ENOENT; 311 312 /* 313 * JOIN_NOLOCK only happens during the transaction commit, so 314 * it is impossible that ->running_transaction is NULL 315 */ 316 BUG_ON(type == TRANS_JOIN_NOLOCK); 317 318 cur_trans = kmalloc_obj(*cur_trans, GFP_NOFS); 319 if (!cur_trans) 320 return -ENOMEM; 321 322 btrfs_lockdep_acquire(fs_info, btrfs_trans_num_writers); 323 btrfs_lockdep_acquire(fs_info, btrfs_trans_num_extwriters); 324 325 spin_lock(&fs_info->trans_lock); 326 if (fs_info->running_transaction) { 327 /* 328 * someone started a transaction after we unlocked. Make sure 329 * to redo the checks above 330 */ 331 btrfs_lockdep_release(fs_info, btrfs_trans_num_extwriters); 332 btrfs_lockdep_release(fs_info, btrfs_trans_num_writers); 333 kfree(cur_trans); 334 goto loop; 335 } else if (BTRFS_FS_ERROR(fs_info)) { 336 spin_unlock(&fs_info->trans_lock); 337 btrfs_lockdep_release(fs_info, btrfs_trans_num_extwriters); 338 btrfs_lockdep_release(fs_info, btrfs_trans_num_writers); 339 kfree(cur_trans); 340 return -EROFS; 341 } 342 343 cur_trans->fs_info = fs_info; 344 atomic_set(&cur_trans->pending_ordered, 0); 345 init_waitqueue_head(&cur_trans->pending_wait); 346 atomic_set(&cur_trans->num_writers, 1); 347 extwriter_counter_init(cur_trans, type); 348 init_waitqueue_head(&cur_trans->writer_wait); 349 init_waitqueue_head(&cur_trans->commit_wait); 350 cur_trans->state = TRANS_STATE_RUNNING; 351 /* 352 * One for this trans handle, one so it will live on until we 353 * commit the transaction. 354 */ 355 refcount_set(&cur_trans->use_count, 2); 356 cur_trans->flags = 0; 357 cur_trans->start_time = ktime_get_seconds(); 358 359 memset(&cur_trans->delayed_refs, 0, sizeof(cur_trans->delayed_refs)); 360 361 xa_init(&cur_trans->delayed_refs.head_refs); 362 xa_init(&cur_trans->delayed_refs.dirty_extents); 363 364 /* 365 * although the tree mod log is per file system and not per transaction, 366 * the log must never go across transaction boundaries. 367 */ 368 smp_mb(); 369 if (!list_empty(&fs_info->tree_mod_seq_list)) 370 WARN(1, KERN_ERR "BTRFS: tree_mod_seq_list not empty when creating a fresh transaction\n"); 371 if (!RB_EMPTY_ROOT(&fs_info->tree_mod_log)) 372 WARN(1, KERN_ERR "BTRFS: tree_mod_log rb tree not empty when creating a fresh transaction\n"); 373 atomic64_set(&fs_info->tree_mod_seq, 0); 374 375 spin_lock_init(&cur_trans->delayed_refs.lock); 376 377 INIT_LIST_HEAD(&cur_trans->pending_snapshots); 378 INIT_LIST_HEAD(&cur_trans->dev_update_list); 379 INIT_LIST_HEAD(&cur_trans->switch_commits); 380 INIT_LIST_HEAD(&cur_trans->dirty_bgs); 381 INIT_LIST_HEAD(&cur_trans->io_bgs); 382 INIT_LIST_HEAD(&cur_trans->dropped_roots); 383 mutex_init(&cur_trans->cache_write_mutex); 384 spin_lock_init(&cur_trans->dirty_bgs_lock); 385 INIT_LIST_HEAD(&cur_trans->deleted_bgs); 386 spin_lock_init(&cur_trans->dropped_roots_lock); 387 list_add_tail(&cur_trans->list, &fs_info->trans_list); 388 btrfs_extent_io_tree_init(fs_info, &cur_trans->dirty_pages, 389 IO_TREE_TRANS_DIRTY_PAGES); 390 btrfs_extent_io_tree_init(fs_info, &cur_trans->pinned_extents, 391 IO_TREE_FS_PINNED_EXTENTS); 392 btrfs_set_fs_generation(fs_info, fs_info->generation + 1); 393 cur_trans->transid = fs_info->generation; 394 fs_info->running_transaction = cur_trans; 395 cur_trans->aborted = 0; 396 spin_unlock(&fs_info->trans_lock); 397 398 return 0; 399 } 400 401 /* 402 * This does all the record keeping required to make sure that a shareable root 403 * is properly recorded in a given transaction. This is required to make sure 404 * the old root from before we joined the transaction is deleted when the 405 * transaction commits. 406 */ 407 static int record_root_in_trans(struct btrfs_trans_handle *trans, 408 struct btrfs_root *root, 409 bool force) 410 { 411 struct btrfs_fs_info *fs_info = root->fs_info; 412 int ret = 0; 413 414 if ((test_bit(BTRFS_ROOT_SHAREABLE, &root->state) && 415 btrfs_get_root_last_trans(root) < trans->transid) || force) { 416 WARN_ON(!force && root->commit_root != root->node); 417 418 /* 419 * see below for IN_TRANS_SETUP usage rules 420 * we have the reloc mutex held now, so there 421 * is only one writer in this function 422 */ 423 set_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state); 424 425 /* make sure readers find IN_TRANS_SETUP before 426 * they find our root->last_trans update 427 */ 428 smp_wmb(); 429 430 spin_lock(&fs_info->fs_roots_radix_lock); 431 if (btrfs_get_root_last_trans(root) == trans->transid && !force) { 432 spin_unlock(&fs_info->fs_roots_radix_lock); 433 return 0; 434 } 435 radix_tree_tag_set(&fs_info->fs_roots_radix, 436 (unsigned long)btrfs_root_id(root), 437 BTRFS_ROOT_TRANS_TAG); 438 spin_unlock(&fs_info->fs_roots_radix_lock); 439 btrfs_set_root_last_trans(root, trans->transid); 440 441 /* this is pretty tricky. We don't want to 442 * take the relocation lock in btrfs_record_root_in_trans 443 * unless we're really doing the first setup for this root in 444 * this transaction. 445 * 446 * Normally we'd use root->last_trans as a flag to decide 447 * if we want to take the expensive mutex. 448 * 449 * But, we have to set root->last_trans before we 450 * init the relocation root, otherwise, we trip over warnings 451 * in ctree.c. The solution used here is to flag ourselves 452 * with root IN_TRANS_SETUP. When this is 1, we're still 453 * fixing up the reloc trees and everyone must wait. 454 * 455 * When this is zero, they can trust root->last_trans and fly 456 * through btrfs_record_root_in_trans without having to take the 457 * lock. smp_wmb() makes sure that all the writes above are 458 * done before we pop in the zero below 459 */ 460 ret = btrfs_init_reloc_root(trans, root); 461 smp_mb__before_atomic(); 462 clear_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state); 463 } 464 return ret; 465 } 466 467 468 void btrfs_add_dropped_root(struct btrfs_trans_handle *trans, 469 struct btrfs_root *root) 470 { 471 struct btrfs_fs_info *fs_info = root->fs_info; 472 struct btrfs_transaction *cur_trans = trans->transaction; 473 474 /* Add ourselves to the transaction dropped list */ 475 spin_lock(&cur_trans->dropped_roots_lock); 476 list_add_tail(&root->root_list, &cur_trans->dropped_roots); 477 spin_unlock(&cur_trans->dropped_roots_lock); 478 479 /* Make sure we don't try to update the root at commit time */ 480 spin_lock(&fs_info->fs_roots_radix_lock); 481 radix_tree_tag_clear(&fs_info->fs_roots_radix, 482 (unsigned long)btrfs_root_id(root), 483 BTRFS_ROOT_TRANS_TAG); 484 spin_unlock(&fs_info->fs_roots_radix_lock); 485 } 486 487 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans, 488 struct btrfs_root *root) 489 { 490 struct btrfs_fs_info *fs_info = root->fs_info; 491 int ret; 492 493 if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) 494 return 0; 495 496 /* 497 * see record_root_in_trans for comments about IN_TRANS_SETUP usage 498 * and barriers 499 */ 500 smp_rmb(); 501 if (btrfs_get_root_last_trans(root) == trans->transid && 502 !test_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state)) 503 return 0; 504 505 mutex_lock(&fs_info->reloc_mutex); 506 ret = record_root_in_trans(trans, root, 0); 507 mutex_unlock(&fs_info->reloc_mutex); 508 509 return ret; 510 } 511 512 static inline int is_transaction_blocked(struct btrfs_transaction *trans) 513 { 514 return (trans->state >= TRANS_STATE_COMMIT_START && 515 trans->state < TRANS_STATE_UNBLOCKED && 516 !TRANS_ABORTED(trans)); 517 } 518 519 /* wait for commit against the current transaction to become unblocked 520 * when this is done, it is safe to start a new transaction, but the current 521 * transaction might not be fully on disk. 522 */ 523 static void wait_current_trans(struct btrfs_fs_info *fs_info, unsigned int type) 524 { 525 struct btrfs_transaction *cur_trans; 526 527 spin_lock(&fs_info->trans_lock); 528 cur_trans = fs_info->running_transaction; 529 if (cur_trans && is_transaction_blocked(cur_trans) && 530 (btrfs_blocked_trans_types[cur_trans->state] & type)) { 531 refcount_inc(&cur_trans->use_count); 532 spin_unlock(&fs_info->trans_lock); 533 534 btrfs_might_wait_for_state(fs_info, BTRFS_LOCKDEP_TRANS_UNBLOCKED); 535 wait_event(fs_info->transaction_wait, 536 cur_trans->state >= TRANS_STATE_UNBLOCKED || 537 TRANS_ABORTED(cur_trans)); 538 btrfs_put_transaction(cur_trans); 539 } else { 540 spin_unlock(&fs_info->trans_lock); 541 } 542 } 543 544 static bool may_wait_transaction(struct btrfs_fs_info *fs_info, int type) 545 { 546 if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags)) 547 return false; 548 549 if (type == TRANS_START) 550 return true; 551 552 return false; 553 } 554 555 static inline bool need_reserve_reloc_root(struct btrfs_root *root) 556 { 557 struct btrfs_fs_info *fs_info = root->fs_info; 558 559 if (!fs_info->reloc_ctl || 560 !test_bit(BTRFS_ROOT_SHAREABLE, &root->state) || 561 btrfs_root_id(root) == BTRFS_TREE_RELOC_OBJECTID || 562 root->reloc_root) 563 return false; 564 565 return true; 566 } 567 568 static int btrfs_reserve_trans_metadata(struct btrfs_fs_info *fs_info, 569 enum btrfs_reserve_flush_enum flush, 570 u64 num_bytes, 571 u64 *delayed_refs_bytes) 572 { 573 struct btrfs_space_info *si = fs_info->trans_block_rsv.space_info; 574 u64 bytes = num_bytes + *delayed_refs_bytes; 575 int ret; 576 577 /* 578 * We want to reserve all the bytes we may need all at once, so we only 579 * do 1 enospc flushing cycle per transaction start. 580 */ 581 ret = btrfs_reserve_metadata_bytes(si, bytes, flush); 582 583 /* 584 * If we are an emergency flush, which can steal from the global block 585 * reserve, then attempt to not reserve space for the delayed refs, as 586 * we will consume space for them from the global block reserve. 587 */ 588 if (ret && flush == BTRFS_RESERVE_FLUSH_ALL_STEAL) { 589 bytes -= *delayed_refs_bytes; 590 *delayed_refs_bytes = 0; 591 ret = btrfs_reserve_metadata_bytes(si, bytes, flush); 592 } 593 594 return ret; 595 } 596 597 static struct btrfs_trans_handle * 598 start_transaction(struct btrfs_root *root, unsigned int num_items, 599 unsigned int type, enum btrfs_reserve_flush_enum flush, 600 bool enforce_qgroups) 601 { 602 struct btrfs_fs_info *fs_info = root->fs_info; 603 struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv; 604 struct btrfs_block_rsv *trans_rsv = &fs_info->trans_block_rsv; 605 struct btrfs_trans_handle *h; 606 struct btrfs_transaction *cur_trans; 607 u64 num_bytes = 0; 608 u64 qgroup_reserved = 0; 609 u64 delayed_refs_bytes = 0; 610 bool reloc_reserved = false; 611 bool do_chunk_alloc = false; 612 int ret; 613 614 if (BTRFS_FS_ERROR(fs_info)) 615 return ERR_PTR(-EROFS); 616 617 if (current->journal_info) { 618 WARN_ON(type & TRANS_EXTWRITERS); 619 h = current->journal_info; 620 refcount_inc(&h->use_count); 621 WARN_ON(refcount_read(&h->use_count) > 2); 622 h->orig_rsv = h->block_rsv; 623 h->block_rsv = NULL; 624 goto got_it; 625 } 626 627 /* 628 * Do the reservation before we join the transaction so we can do all 629 * the appropriate flushing if need be. 630 */ 631 if (num_items && root != fs_info->chunk_root) { 632 qgroup_reserved = num_items * fs_info->nodesize; 633 /* 634 * Use prealloc for now, as there might be a currently running 635 * transaction that could free this reserved space prematurely 636 * by committing. 637 */ 638 ret = btrfs_qgroup_reserve_meta_prealloc(root, qgroup_reserved, 639 enforce_qgroups, false); 640 if (ret) 641 return ERR_PTR(ret); 642 643 num_bytes = btrfs_calc_insert_metadata_size(fs_info, num_items); 644 /* 645 * If we plan to insert/update/delete "num_items" from a btree, 646 * we will also generate delayed refs for extent buffers in the 647 * respective btree paths, so reserve space for the delayed refs 648 * that will be generated by the caller as it modifies btrees. 649 * Try to reserve them to avoid excessive use of the global 650 * block reserve. 651 */ 652 delayed_refs_bytes = btrfs_calc_delayed_ref_bytes(fs_info, num_items); 653 654 /* 655 * Do the reservation for the relocation root creation 656 */ 657 if (need_reserve_reloc_root(root)) { 658 num_bytes += fs_info->nodesize; 659 reloc_reserved = true; 660 } 661 662 ret = btrfs_reserve_trans_metadata(fs_info, flush, num_bytes, 663 &delayed_refs_bytes); 664 if (ret) 665 goto reserve_fail; 666 667 btrfs_block_rsv_add_bytes(trans_rsv, num_bytes, true); 668 669 if (trans_rsv->space_info->force_alloc) 670 do_chunk_alloc = true; 671 } else if (num_items == 0 && flush == BTRFS_RESERVE_FLUSH_ALL && 672 !btrfs_block_rsv_full(delayed_refs_rsv)) { 673 /* 674 * Some people call with btrfs_start_transaction(root, 0) 675 * because they can be throttled, but have some other mechanism 676 * for reserving space. We still want these guys to refill the 677 * delayed block_rsv so just add 1 items worth of reservation 678 * here. 679 */ 680 ret = btrfs_delayed_refs_rsv_refill(fs_info, flush); 681 if (ret) 682 goto reserve_fail; 683 } 684 again: 685 h = kmem_cache_zalloc(btrfs_trans_handle_cachep, GFP_NOFS); 686 if (!h) { 687 ret = -ENOMEM; 688 goto alloc_fail; 689 } 690 691 /* 692 * If we are JOIN_NOLOCK we're already committing a transaction and 693 * waiting on this guy, so we don't need to do the sb_start_intwrite 694 * because we're already holding a ref. We need this because we could 695 * have raced in and did an fsync() on a file which can kick a commit 696 * and then we deadlock with somebody doing a freeze. 697 * 698 * If we are ATTACH, it means we just want to catch the current 699 * transaction and commit it, so we needn't do sb_start_intwrite(). 700 */ 701 if (type & __TRANS_FREEZABLE) 702 sb_start_intwrite(fs_info->sb); 703 704 if (may_wait_transaction(fs_info, type)) 705 wait_current_trans(fs_info, type); 706 707 do { 708 ret = join_transaction(fs_info, type); 709 if (ret == -EBUSY) { 710 wait_current_trans(fs_info, type); 711 if (unlikely(type == TRANS_ATTACH || 712 type == TRANS_JOIN_NOSTART)) 713 ret = -ENOENT; 714 } 715 } while (ret == -EBUSY); 716 717 if (ret < 0) 718 goto join_fail; 719 720 cur_trans = fs_info->running_transaction; 721 722 h->transid = cur_trans->transid; 723 h->transaction = cur_trans; 724 refcount_set(&h->use_count, 1); 725 h->fs_info = root->fs_info; 726 727 h->type = type; 728 INIT_LIST_HEAD(&h->new_bgs); 729 btrfs_init_metadata_block_rsv(fs_info, &h->delayed_rsv, BTRFS_BLOCK_RSV_DELREFS); 730 731 smp_mb(); 732 if (cur_trans->state >= TRANS_STATE_COMMIT_START && 733 may_wait_transaction(fs_info, type)) { 734 current->journal_info = h; 735 btrfs_commit_transaction(h); 736 goto again; 737 } 738 739 if (num_bytes) { 740 trace_btrfs_space_reservation(fs_info, "transaction", 741 h->transid, num_bytes, 1); 742 h->block_rsv = trans_rsv; 743 h->bytes_reserved = num_bytes; 744 if (delayed_refs_bytes > 0) { 745 trace_btrfs_space_reservation(fs_info, 746 "local_delayed_refs_rsv", 747 h->transid, 748 delayed_refs_bytes, 1); 749 h->delayed_refs_bytes_reserved = delayed_refs_bytes; 750 btrfs_block_rsv_add_bytes(&h->delayed_rsv, delayed_refs_bytes, true); 751 delayed_refs_bytes = 0; 752 } 753 h->reloc_reserved = reloc_reserved; 754 } 755 756 got_it: 757 if (!current->journal_info) 758 current->journal_info = h; 759 760 /* 761 * If the space_info is marked ALLOC_FORCE then we'll get upgraded to 762 * ALLOC_FORCE the first run through, and then we won't allocate for 763 * anybody else who races in later. We don't care about the return 764 * value here. 765 */ 766 if (do_chunk_alloc && num_bytes) { 767 struct btrfs_space_info *space_info = h->block_rsv->space_info; 768 u64 flags = space_info->flags; 769 770 btrfs_chunk_alloc(h, space_info, btrfs_get_alloc_profile(fs_info, flags), 771 CHUNK_ALLOC_NO_FORCE); 772 } 773 774 /* 775 * btrfs_record_root_in_trans() needs to alloc new extents, and may 776 * call btrfs_join_transaction() while we're also starting a 777 * transaction. 778 * 779 * Thus it need to be called after current->journal_info initialized, 780 * or we can deadlock. 781 */ 782 ret = btrfs_record_root_in_trans(h, root); 783 if (ret) { 784 /* 785 * The transaction handle is fully initialized and linked with 786 * other structures so it needs to be ended in case of errors, 787 * not just freed. 788 */ 789 btrfs_end_transaction(h); 790 goto reserve_fail; 791 } 792 /* 793 * Now that we have found a transaction to be a part of, convert the 794 * qgroup reservation from prealloc to pertrans. A different transaction 795 * can't race in and free our pertrans out from under us. 796 */ 797 if (qgroup_reserved) 798 btrfs_qgroup_convert_reserved_meta(root, qgroup_reserved); 799 800 return h; 801 802 join_fail: 803 if (type & __TRANS_FREEZABLE) 804 sb_end_intwrite(fs_info->sb); 805 kmem_cache_free(btrfs_trans_handle_cachep, h); 806 alloc_fail: 807 if (num_bytes) 808 btrfs_block_rsv_release(fs_info, trans_rsv, num_bytes, NULL); 809 if (delayed_refs_bytes) 810 btrfs_space_info_free_bytes_may_use(trans_rsv->space_info, delayed_refs_bytes); 811 reserve_fail: 812 btrfs_qgroup_free_meta_prealloc(root, qgroup_reserved); 813 return ERR_PTR(ret); 814 } 815 816 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root, 817 unsigned int num_items) 818 { 819 return start_transaction(root, num_items, TRANS_START, 820 BTRFS_RESERVE_FLUSH_ALL, true); 821 } 822 823 struct btrfs_trans_handle *btrfs_start_transaction_fallback_global_rsv( 824 struct btrfs_root *root, 825 unsigned int num_items) 826 { 827 return start_transaction(root, num_items, TRANS_START, 828 BTRFS_RESERVE_FLUSH_ALL_STEAL, false); 829 } 830 831 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root) 832 { 833 return start_transaction(root, 0, TRANS_JOIN, BTRFS_RESERVE_NO_FLUSH, 834 true); 835 } 836 837 struct btrfs_trans_handle *btrfs_join_transaction_spacecache(struct btrfs_root *root) 838 { 839 return start_transaction(root, 0, TRANS_JOIN_NOLOCK, 840 BTRFS_RESERVE_NO_FLUSH, true); 841 } 842 843 /* 844 * Similar to regular join but it never starts a transaction when none is 845 * running or when there's a running one at a state >= TRANS_STATE_UNBLOCKED. 846 * This is similar to btrfs_attach_transaction() but it allows the join to 847 * happen if the transaction commit already started but it's not yet in the 848 * "doing" phase (the state is < TRANS_STATE_COMMIT_DOING). 849 */ 850 struct btrfs_trans_handle *btrfs_join_transaction_nostart(struct btrfs_root *root) 851 { 852 return start_transaction(root, 0, TRANS_JOIN_NOSTART, 853 BTRFS_RESERVE_NO_FLUSH, true); 854 } 855 856 /* 857 * Catch the running transaction. 858 * 859 * It is used when we want to commit the current the transaction, but 860 * don't want to start a new one. 861 * 862 * Note: If this function return -ENOENT, it just means there is no 863 * running transaction. But it is possible that the inactive transaction 864 * is still in the memory, not fully on disk. If you hope there is no 865 * inactive transaction in the fs when -ENOENT is returned, you should 866 * invoke 867 * btrfs_attach_transaction_barrier() 868 */ 869 struct btrfs_trans_handle *btrfs_attach_transaction(struct btrfs_root *root) 870 { 871 return start_transaction(root, 0, TRANS_ATTACH, 872 BTRFS_RESERVE_NO_FLUSH, true); 873 } 874 875 /* 876 * Catch the running transaction. 877 * 878 * It is similar to the above function, the difference is this one 879 * will wait for all the inactive transactions until they fully 880 * complete. 881 */ 882 struct btrfs_trans_handle * 883 btrfs_attach_transaction_barrier(struct btrfs_root *root) 884 { 885 struct btrfs_trans_handle *trans; 886 887 trans = start_transaction(root, 0, TRANS_ATTACH, 888 BTRFS_RESERVE_NO_FLUSH, true); 889 if (trans == ERR_PTR(-ENOENT)) { 890 int ret; 891 892 ret = btrfs_wait_for_commit(root->fs_info, 0); 893 if (ret) 894 return ERR_PTR(ret); 895 } 896 897 return trans; 898 } 899 900 /* Wait for a transaction commit to reach at least the given state. */ 901 static noinline void wait_for_commit(struct btrfs_transaction *commit, 902 const enum btrfs_trans_state min_state) 903 { 904 struct btrfs_fs_info *fs_info = commit->fs_info; 905 u64 transid = commit->transid; 906 bool put = false; 907 908 /* 909 * At the moment this function is called with min_state either being 910 * TRANS_STATE_COMPLETED or TRANS_STATE_SUPER_COMMITTED. 911 */ 912 if (min_state == TRANS_STATE_COMPLETED) 913 btrfs_might_wait_for_state(fs_info, BTRFS_LOCKDEP_TRANS_COMPLETED); 914 else 915 btrfs_might_wait_for_state(fs_info, BTRFS_LOCKDEP_TRANS_SUPER_COMMITTED); 916 917 while (1) { 918 wait_event(commit->commit_wait, commit->state >= min_state); 919 if (put) 920 btrfs_put_transaction(commit); 921 922 if (min_state < TRANS_STATE_COMPLETED) 923 break; 924 925 /* 926 * A transaction isn't really completed until all of the 927 * previous transactions are completed, but with fsync we can 928 * end up with SUPER_COMMITTED transactions before a COMPLETED 929 * transaction. Wait for those. 930 */ 931 932 spin_lock(&fs_info->trans_lock); 933 commit = list_first_entry_or_null(&fs_info->trans_list, 934 struct btrfs_transaction, 935 list); 936 if (!commit || commit->transid > transid) { 937 spin_unlock(&fs_info->trans_lock); 938 break; 939 } 940 refcount_inc(&commit->use_count); 941 put = true; 942 spin_unlock(&fs_info->trans_lock); 943 } 944 } 945 946 int btrfs_wait_for_commit(struct btrfs_fs_info *fs_info, u64 transid) 947 { 948 struct btrfs_transaction *cur_trans = NULL, *t; 949 int ret = 0; 950 951 if (transid) { 952 if (transid <= btrfs_get_last_trans_committed(fs_info)) 953 return 0; 954 955 /* find specified transaction */ 956 spin_lock(&fs_info->trans_lock); 957 list_for_each_entry(t, &fs_info->trans_list, list) { 958 if (t->transid == transid) { 959 cur_trans = t; 960 refcount_inc(&cur_trans->use_count); 961 ret = 0; 962 break; 963 } 964 if (t->transid > transid) { 965 ret = 0; 966 break; 967 } 968 } 969 spin_unlock(&fs_info->trans_lock); 970 971 /* 972 * The specified transaction doesn't exist, or we 973 * raced with btrfs_commit_transaction 974 */ 975 if (!cur_trans) { 976 if (transid > btrfs_get_last_trans_committed(fs_info)) 977 ret = -EINVAL; 978 return ret; 979 } 980 } else { 981 /* find newest transaction that is committing | committed */ 982 spin_lock(&fs_info->trans_lock); 983 list_for_each_entry_reverse(t, &fs_info->trans_list, 984 list) { 985 if (t->state >= TRANS_STATE_COMMIT_START) { 986 if (t->state == TRANS_STATE_COMPLETED) 987 break; 988 cur_trans = t; 989 refcount_inc(&cur_trans->use_count); 990 break; 991 } 992 } 993 spin_unlock(&fs_info->trans_lock); 994 /* Nothing committing or committed. */ 995 if (!cur_trans) 996 return ret; 997 } 998 999 wait_for_commit(cur_trans, TRANS_STATE_COMPLETED); 1000 ret = cur_trans->aborted; 1001 btrfs_put_transaction(cur_trans); 1002 1003 return ret; 1004 } 1005 1006 void btrfs_throttle(struct btrfs_fs_info *fs_info) 1007 { 1008 wait_current_trans(fs_info, TRANS_START); 1009 } 1010 1011 bool btrfs_should_end_transaction(struct btrfs_trans_handle *trans) 1012 { 1013 struct btrfs_transaction *cur_trans = trans->transaction; 1014 1015 if (cur_trans->state >= TRANS_STATE_COMMIT_START || 1016 test_bit(BTRFS_DELAYED_REFS_FLUSHING, &cur_trans->delayed_refs.flags)) 1017 return true; 1018 1019 if (btrfs_check_space_for_delayed_refs(trans->fs_info)) 1020 return true; 1021 1022 return !!btrfs_block_rsv_check(&trans->fs_info->global_block_rsv, 50); 1023 } 1024 1025 static void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans) 1026 1027 { 1028 struct btrfs_fs_info *fs_info = trans->fs_info; 1029 1030 if (!trans->block_rsv) { 1031 ASSERT(trans->bytes_reserved == 0, 1032 "trans->bytes_reserved=%llu", trans->bytes_reserved); 1033 ASSERT(trans->delayed_refs_bytes_reserved == 0, 1034 "trans->delayed_refs_bytes_reserved=%llu", 1035 trans->delayed_refs_bytes_reserved); 1036 return; 1037 } 1038 1039 if (!trans->bytes_reserved) { 1040 ASSERT(trans->delayed_refs_bytes_reserved == 0, 1041 "trans->delayed_refs_bytes_reserved=%llu", 1042 trans->delayed_refs_bytes_reserved); 1043 return; 1044 } 1045 1046 ASSERT(trans->block_rsv == &fs_info->trans_block_rsv); 1047 trace_btrfs_space_reservation(fs_info, "transaction", 1048 trans->transid, trans->bytes_reserved, 0); 1049 btrfs_block_rsv_release(fs_info, trans->block_rsv, 1050 trans->bytes_reserved, NULL); 1051 trans->bytes_reserved = 0; 1052 1053 if (!trans->delayed_refs_bytes_reserved) 1054 return; 1055 1056 trace_btrfs_space_reservation(fs_info, "local_delayed_refs_rsv", 1057 trans->transid, 1058 trans->delayed_refs_bytes_reserved, 0); 1059 btrfs_block_rsv_release(fs_info, &trans->delayed_rsv, 1060 trans->delayed_refs_bytes_reserved, NULL); 1061 trans->delayed_refs_bytes_reserved = 0; 1062 } 1063 1064 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans, 1065 int throttle) 1066 { 1067 struct btrfs_fs_info *info = trans->fs_info; 1068 struct btrfs_transaction *cur_trans = trans->transaction; 1069 int ret = 0; 1070 1071 if (refcount_read(&trans->use_count) > 1) { 1072 refcount_dec(&trans->use_count); 1073 trans->block_rsv = trans->orig_rsv; 1074 return 0; 1075 } 1076 1077 btrfs_trans_release_metadata(trans); 1078 trans->block_rsv = NULL; 1079 1080 btrfs_create_pending_block_groups(trans); 1081 1082 btrfs_trans_release_chunk_metadata(trans); 1083 1084 if (trans->type & __TRANS_FREEZABLE) 1085 sb_end_intwrite(info->sb); 1086 1087 WARN_ON(cur_trans != info->running_transaction); 1088 WARN_ON(atomic_read(&cur_trans->num_writers) < 1); 1089 atomic_dec(&cur_trans->num_writers); 1090 extwriter_counter_dec(cur_trans, trans->type); 1091 1092 cond_wake_up(&cur_trans->writer_wait); 1093 1094 btrfs_lockdep_release(info, btrfs_trans_num_extwriters); 1095 btrfs_lockdep_release(info, btrfs_trans_num_writers); 1096 1097 btrfs_put_transaction(cur_trans); 1098 1099 if (current->journal_info == trans) 1100 current->journal_info = NULL; 1101 1102 if (throttle) 1103 btrfs_run_delayed_iputs(info); 1104 1105 if (TRANS_ABORTED(trans) || BTRFS_FS_ERROR(info)) { 1106 wake_up_process(info->transaction_kthread); 1107 if (TRANS_ABORTED(trans)) 1108 ret = trans->aborted; 1109 else 1110 ret = -EROFS; 1111 } 1112 1113 kmem_cache_free(btrfs_trans_handle_cachep, trans); 1114 return ret; 1115 } 1116 1117 int btrfs_end_transaction(struct btrfs_trans_handle *trans) 1118 { 1119 return __btrfs_end_transaction(trans, 0); 1120 } 1121 1122 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans) 1123 { 1124 return __btrfs_end_transaction(trans, 1); 1125 } 1126 1127 /* 1128 * when btree blocks are allocated, they have some corresponding bits set for 1129 * them in one of two extent_io trees. This is used to make sure all of 1130 * those extents are sent to disk but does not wait on them 1131 */ 1132 int btrfs_write_marked_extents(struct btrfs_fs_info *fs_info, 1133 struct extent_io_tree *dirty_pages, int mark) 1134 { 1135 int ret = 0; 1136 struct address_space *mapping = fs_info->btree_inode->i_mapping; 1137 struct extent_state *cached_state = NULL; 1138 u64 start = 0; 1139 u64 end; 1140 1141 while (btrfs_find_first_extent_bit(dirty_pages, start, &start, &end, 1142 mark, &cached_state)) { 1143 bool wait_writeback = false; 1144 1145 ret = btrfs_convert_extent_bit(dirty_pages, start, end, 1146 EXTENT_NEED_WAIT, 1147 mark, &cached_state); 1148 /* 1149 * convert_extent_bit can return -ENOMEM, which is most of the 1150 * time a temporary error. So when it happens, ignore the error 1151 * and wait for writeback of this range to finish - because we 1152 * failed to set the bit EXTENT_NEED_WAIT for the range, a call 1153 * to __btrfs_wait_marked_extents() would not know that 1154 * writeback for this range started and therefore wouldn't 1155 * wait for it to finish - we don't want to commit a 1156 * superblock that points to btree nodes/leafs for which 1157 * writeback hasn't finished yet (and without errors). 1158 * We cleanup any entries left in the io tree when committing 1159 * the transaction (through extent_io_tree_release()). 1160 */ 1161 if (ret == -ENOMEM) { 1162 ret = 0; 1163 wait_writeback = true; 1164 } 1165 if (!ret) 1166 ret = filemap_fdatawrite_range(mapping, start, end); 1167 if (!ret && wait_writeback) 1168 btrfs_btree_wait_writeback_range(fs_info, start, end); 1169 btrfs_free_extent_state(cached_state); 1170 if (ret) 1171 break; 1172 cached_state = NULL; 1173 cond_resched(); 1174 start = end + 1; 1175 } 1176 return ret; 1177 } 1178 1179 /* 1180 * when btree blocks are allocated, they have some corresponding bits set for 1181 * them in one of two extent_io trees. This is used to make sure all of 1182 * those extents are on disk for transaction or log commit. We wait 1183 * on all the pages and clear them from the dirty pages state tree 1184 */ 1185 static int __btrfs_wait_marked_extents(struct btrfs_fs_info *fs_info, 1186 struct extent_io_tree *dirty_pages) 1187 { 1188 struct extent_state *cached_state = NULL; 1189 u64 start = 0; 1190 u64 end; 1191 int ret = 0; 1192 1193 while (btrfs_find_first_extent_bit(dirty_pages, start, &start, &end, 1194 EXTENT_NEED_WAIT, &cached_state)) { 1195 /* 1196 * Ignore -ENOMEM errors returned by clear_extent_bit(). 1197 * When committing the transaction, we'll remove any entries 1198 * left in the io tree. For a log commit, we don't remove them 1199 * after committing the log because the tree can be accessed 1200 * concurrently - we do it only at transaction commit time when 1201 * it's safe to do it (through extent_io_tree_release()). 1202 */ 1203 ret = btrfs_clear_extent_bit(dirty_pages, start, end, 1204 EXTENT_NEED_WAIT, &cached_state); 1205 if (ret == -ENOMEM) 1206 ret = 0; 1207 if (!ret) 1208 btrfs_btree_wait_writeback_range(fs_info, start, end); 1209 btrfs_free_extent_state(cached_state); 1210 if (ret) 1211 break; 1212 cached_state = NULL; 1213 cond_resched(); 1214 start = end + 1; 1215 } 1216 return ret; 1217 } 1218 1219 static int btrfs_wait_extents(struct btrfs_fs_info *fs_info, 1220 struct extent_io_tree *dirty_pages) 1221 { 1222 bool errors = false; 1223 int ret; 1224 1225 ret = __btrfs_wait_marked_extents(fs_info, dirty_pages); 1226 if (test_and_clear_bit(BTRFS_FS_BTREE_ERR, &fs_info->flags)) 1227 errors = true; 1228 1229 if (errors && !ret) 1230 ret = -EIO; 1231 return ret; 1232 } 1233 1234 int btrfs_wait_tree_log_extents(struct btrfs_root *log_root, int mark) 1235 { 1236 struct btrfs_fs_info *fs_info = log_root->fs_info; 1237 struct extent_io_tree *dirty_pages = &log_root->dirty_log_pages; 1238 bool errors = false; 1239 int ret; 1240 1241 ASSERT(btrfs_root_id(log_root) == BTRFS_TREE_LOG_OBJECTID, 1242 "root_id(log_root)=%llu", btrfs_root_id(log_root)); 1243 1244 ret = __btrfs_wait_marked_extents(fs_info, dirty_pages); 1245 if ((mark & EXTENT_DIRTY_LOG1) && 1246 test_and_clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags)) 1247 errors = true; 1248 1249 if ((mark & EXTENT_DIRTY_LOG2) && 1250 test_and_clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags)) 1251 errors = true; 1252 1253 if (errors && !ret) 1254 ret = -EIO; 1255 return ret; 1256 } 1257 1258 /* 1259 * When btree blocks are allocated the corresponding extents are marked dirty. 1260 * This function ensures such extents are persisted on disk for transaction or 1261 * log commit. 1262 * 1263 * @trans: transaction whose dirty pages we'd like to write 1264 */ 1265 static int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans) 1266 { 1267 int ret; 1268 int ret2; 1269 struct extent_io_tree *dirty_pages = &trans->transaction->dirty_pages; 1270 struct btrfs_fs_info *fs_info = trans->fs_info; 1271 struct blk_plug plug; 1272 1273 blk_start_plug(&plug); 1274 ret = btrfs_write_marked_extents(fs_info, dirty_pages, EXTENT_DIRTY); 1275 blk_finish_plug(&plug); 1276 ret2 = btrfs_wait_extents(fs_info, dirty_pages); 1277 1278 btrfs_extent_io_tree_release(&trans->transaction->dirty_pages); 1279 1280 if (ret) 1281 return ret; 1282 else if (ret2) 1283 return ret2; 1284 else 1285 return 0; 1286 } 1287 1288 /* 1289 * this is used to update the root pointer in the tree of tree roots. 1290 * 1291 * But, in the case of the extent allocation tree, updating the root 1292 * pointer may allocate blocks which may change the root of the extent 1293 * allocation tree. 1294 * 1295 * So, this loops and repeats and makes sure the cowonly root didn't 1296 * change while the root pointer was being updated in the metadata. 1297 */ 1298 static int update_cowonly_root(struct btrfs_trans_handle *trans, 1299 struct btrfs_root *root) 1300 { 1301 int ret; 1302 u64 old_root_bytenr; 1303 u64 old_root_used; 1304 struct btrfs_fs_info *fs_info = root->fs_info; 1305 struct btrfs_root *tree_root = fs_info->tree_root; 1306 1307 old_root_used = btrfs_root_used(&root->root_item); 1308 1309 while (1) { 1310 old_root_bytenr = btrfs_root_bytenr(&root->root_item); 1311 if (old_root_bytenr == root->node->start && 1312 old_root_used == btrfs_root_used(&root->root_item)) 1313 break; 1314 1315 btrfs_set_root_node(&root->root_item, root->node); 1316 ret = btrfs_update_root(trans, tree_root, 1317 &root->root_key, 1318 &root->root_item); 1319 if (ret) 1320 return ret; 1321 1322 old_root_used = btrfs_root_used(&root->root_item); 1323 } 1324 1325 return 0; 1326 } 1327 1328 /* 1329 * update all the cowonly tree roots on disk 1330 * 1331 * The error handling in this function may not be obvious. Any of the 1332 * failures will cause the file system to go offline. We still need 1333 * to clean up the delayed refs. 1334 */ 1335 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans) 1336 { 1337 struct btrfs_fs_info *fs_info = trans->fs_info; 1338 struct list_head *dirty_bgs = &trans->transaction->dirty_bgs; 1339 struct list_head *io_bgs = &trans->transaction->io_bgs; 1340 struct extent_buffer *eb; 1341 int ret; 1342 1343 /* 1344 * At this point no one can be using this transaction to modify any tree 1345 * and no one can start another transaction to modify any tree either. 1346 */ 1347 ASSERT(trans->transaction->state == TRANS_STATE_COMMIT_DOING, 1348 "trans->transaction->state=%d", trans->transaction->state); 1349 1350 eb = btrfs_lock_root_node(fs_info->tree_root); 1351 ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL, 1352 0, &eb, BTRFS_NESTING_COW); 1353 btrfs_tree_unlock(eb); 1354 free_extent_buffer(eb); 1355 1356 if (ret) 1357 return ret; 1358 1359 ret = btrfs_run_dev_stats(trans); 1360 if (ret) 1361 return ret; 1362 ret = btrfs_run_dev_replace(trans); 1363 if (ret) 1364 return ret; 1365 ret = btrfs_run_qgroups(trans); 1366 if (ret) 1367 return ret; 1368 1369 ret = btrfs_setup_space_cache(trans); 1370 if (ret) 1371 return ret; 1372 1373 again: 1374 while (!list_empty(&fs_info->dirty_cowonly_roots)) { 1375 struct btrfs_root *root; 1376 1377 root = list_first_entry(&fs_info->dirty_cowonly_roots, 1378 struct btrfs_root, dirty_list); 1379 clear_bit(BTRFS_ROOT_DIRTY, &root->state); 1380 list_move_tail(&root->dirty_list, 1381 &trans->transaction->switch_commits); 1382 1383 ret = update_cowonly_root(trans, root); 1384 if (ret) 1385 return ret; 1386 } 1387 1388 /* Now flush any delayed refs generated by updating all of the roots */ 1389 ret = btrfs_run_delayed_refs(trans, U64_MAX); 1390 if (ret) 1391 return ret; 1392 1393 while (!list_empty(dirty_bgs) || !list_empty(io_bgs)) { 1394 ret = btrfs_write_dirty_block_groups(trans); 1395 if (ret) 1396 return ret; 1397 1398 /* 1399 * We're writing the dirty block groups, which could generate 1400 * delayed refs, which could generate more dirty block groups, 1401 * so we want to keep this flushing in this loop to make sure 1402 * everything gets run. 1403 */ 1404 ret = btrfs_run_delayed_refs(trans, U64_MAX); 1405 if (ret) 1406 return ret; 1407 } 1408 1409 if (!list_empty(&fs_info->dirty_cowonly_roots)) 1410 goto again; 1411 1412 /* Update dev-replace pointer once everything is committed */ 1413 fs_info->dev_replace.committed_cursor_left = 1414 fs_info->dev_replace.cursor_left_last_write_of_item; 1415 1416 return 0; 1417 } 1418 1419 /* 1420 * If we had a pending drop we need to see if there are any others left in our 1421 * dead roots list, and if not clear our bit and wake any waiters. 1422 */ 1423 void btrfs_maybe_wake_unfinished_drop(struct btrfs_fs_info *fs_info) 1424 { 1425 /* 1426 * We put the drop in progress roots at the front of the list, so if the 1427 * first entry doesn't have UNFINISHED_DROP set we can wake everybody 1428 * up. 1429 */ 1430 spin_lock(&fs_info->trans_lock); 1431 if (!list_empty(&fs_info->dead_roots)) { 1432 struct btrfs_root *root = list_first_entry(&fs_info->dead_roots, 1433 struct btrfs_root, 1434 root_list); 1435 if (test_bit(BTRFS_ROOT_UNFINISHED_DROP, &root->state)) { 1436 spin_unlock(&fs_info->trans_lock); 1437 return; 1438 } 1439 } 1440 spin_unlock(&fs_info->trans_lock); 1441 1442 btrfs_wake_unfinished_drop(fs_info); 1443 } 1444 1445 /* 1446 * dead roots are old snapshots that need to be deleted. This allocates 1447 * a dirty root struct and adds it into the list of dead roots that need to 1448 * be deleted 1449 */ 1450 void btrfs_add_dead_root(struct btrfs_root *root) 1451 { 1452 struct btrfs_fs_info *fs_info = root->fs_info; 1453 1454 spin_lock(&fs_info->trans_lock); 1455 if (list_empty(&root->root_list)) { 1456 btrfs_grab_root(root); 1457 1458 /* We want to process the partially complete drops first. */ 1459 if (test_bit(BTRFS_ROOT_UNFINISHED_DROP, &root->state)) 1460 list_add(&root->root_list, &fs_info->dead_roots); 1461 else 1462 list_add_tail(&root->root_list, &fs_info->dead_roots); 1463 } 1464 spin_unlock(&fs_info->trans_lock); 1465 } 1466 1467 /* 1468 * Update each subvolume root and its relocation root, if it exists, in the tree 1469 * of tree roots. Also free log roots if they exist. 1470 */ 1471 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans) 1472 { 1473 struct btrfs_fs_info *fs_info = trans->fs_info; 1474 struct btrfs_root *gang[8]; 1475 int i; 1476 int ret; 1477 1478 /* 1479 * At this point no one can be using this transaction to modify any tree 1480 * and no one can start another transaction to modify any tree either. 1481 */ 1482 ASSERT(trans->transaction->state == TRANS_STATE_COMMIT_DOING, 1483 "trans->transaction->state=%d", trans->transaction->state); 1484 1485 spin_lock(&fs_info->fs_roots_radix_lock); 1486 while (1) { 1487 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix, 1488 (void **)gang, 0, 1489 ARRAY_SIZE(gang), 1490 BTRFS_ROOT_TRANS_TAG); 1491 if (ret == 0) 1492 break; 1493 for (i = 0; i < ret; i++) { 1494 struct btrfs_root *root = gang[i]; 1495 int ret2; 1496 1497 /* 1498 * At this point we can neither have tasks logging inodes 1499 * from a root nor trying to commit a log tree. 1500 */ 1501 ASSERT(atomic_read(&root->log_writers) == 0, 1502 "atomic_read(&root->log_writers)=%d", 1503 atomic_read(&root->log_writers)); 1504 ASSERT(atomic_read(&root->log_commit[0]) == 0, 1505 "atomic_read(&root->log_commit[0])=%d", 1506 atomic_read(&root->log_commit[0])); 1507 ASSERT(atomic_read(&root->log_commit[1]) == 0, 1508 "atomic_read(&root->log_commit[1])=%d", 1509 atomic_read(&root->log_commit[1])); 1510 1511 radix_tree_tag_clear(&fs_info->fs_roots_radix, 1512 (unsigned long)btrfs_root_id(root), 1513 BTRFS_ROOT_TRANS_TAG); 1514 btrfs_qgroup_free_meta_all_pertrans(root); 1515 spin_unlock(&fs_info->fs_roots_radix_lock); 1516 1517 btrfs_free_log(trans, root); 1518 ret2 = btrfs_update_reloc_root(trans, root); 1519 if (unlikely(ret2)) 1520 return ret2; 1521 1522 /* see comments in should_cow_block() */ 1523 clear_bit(BTRFS_ROOT_FORCE_COW, &root->state); 1524 smp_mb__after_atomic(); 1525 1526 if (root->commit_root != root->node) { 1527 list_add_tail(&root->dirty_list, 1528 &trans->transaction->switch_commits); 1529 btrfs_set_root_node(&root->root_item, 1530 root->node); 1531 } 1532 1533 ret2 = btrfs_update_root(trans, fs_info->tree_root, 1534 &root->root_key, 1535 &root->root_item); 1536 if (unlikely(ret2)) 1537 return ret2; 1538 spin_lock(&fs_info->fs_roots_radix_lock); 1539 } 1540 } 1541 spin_unlock(&fs_info->fs_roots_radix_lock); 1542 return 0; 1543 } 1544 1545 /* 1546 * Do all special snapshot related qgroup dirty hack. 1547 * 1548 * Will do all needed qgroup inherit and dirty hack like switch commit 1549 * roots inside one transaction and write all btree into disk, to make 1550 * qgroup works. 1551 */ 1552 static int qgroup_account_snapshot(struct btrfs_trans_handle *trans, 1553 struct btrfs_root *src, 1554 struct btrfs_root *parent, 1555 struct btrfs_qgroup_inherit *inherit, 1556 u64 dst_objectid) 1557 { 1558 struct btrfs_fs_info *fs_info = src->fs_info; 1559 int ret; 1560 1561 /* 1562 * Save some performance in the case that qgroups are not enabled. If 1563 * this check races with the ioctl, rescan will kick in anyway. 1564 */ 1565 if (!btrfs_qgroup_full_accounting(fs_info)) 1566 return 0; 1567 1568 /* 1569 * Ensure dirty @src will be committed. Or, after coming 1570 * commit_fs_roots() and switch_commit_roots(), any dirty but not 1571 * recorded root will never be updated again, causing an outdated root 1572 * item. 1573 */ 1574 ret = record_root_in_trans(trans, src, 1); 1575 if (ret) 1576 return ret; 1577 1578 /* 1579 * btrfs_qgroup_inherit relies on a consistent view of the usage for the 1580 * src root, so we must run the delayed refs here. 1581 * 1582 * However this isn't particularly fool proof, because there's no 1583 * synchronization keeping us from changing the tree after this point 1584 * before we do the qgroup_inherit, or even from making changes while 1585 * we're doing the qgroup_inherit. But that's a problem for the future, 1586 * for now flush the delayed refs to narrow the race window where the 1587 * qgroup counters could end up wrong. 1588 */ 1589 ret = btrfs_run_delayed_refs(trans, U64_MAX); 1590 if (unlikely(ret)) { 1591 btrfs_abort_transaction(trans, ret); 1592 return ret; 1593 } 1594 1595 ret = commit_fs_roots(trans); 1596 if (ret) 1597 goto out; 1598 ret = btrfs_qgroup_account_extents(trans); 1599 if (ret < 0) 1600 goto out; 1601 1602 /* Now qgroup are all updated, we can inherit it to new qgroups */ 1603 ret = btrfs_qgroup_inherit(trans, btrfs_root_id(src), dst_objectid, 1604 btrfs_root_id(parent), inherit); 1605 if (ret < 0) 1606 goto out; 1607 1608 /* 1609 * Now we do a simplified commit transaction, which will: 1610 * 1) commit all subvolume and extent tree 1611 * To ensure all subvolume and extent tree have a valid 1612 * commit_root to accounting later insert_dir_item() 1613 * 2) write all btree blocks onto disk 1614 * This is to make sure later btree modification will be cowed 1615 * Or commit_root can be populated and cause wrong qgroup numbers 1616 * In this simplified commit, we don't really care about other trees 1617 * like chunk and root tree, as they won't affect qgroup. 1618 * And we don't write super to avoid half committed status. 1619 */ 1620 ret = commit_cowonly_roots(trans); 1621 if (ret) 1622 goto out; 1623 switch_commit_roots(trans); 1624 ret = btrfs_write_and_wait_transaction(trans); 1625 if (unlikely(ret)) 1626 btrfs_err(fs_info, 1627 "error while writing out transaction during qgroup snapshot accounting: %d", ret); 1628 1629 out: 1630 /* 1631 * Force parent root to be updated, as we recorded it before so its 1632 * last_trans == cur_transid. 1633 * Or it won't be committed again onto disk after later 1634 * insert_dir_item() 1635 */ 1636 if (!ret) 1637 ret = record_root_in_trans(trans, parent, 1); 1638 return ret; 1639 } 1640 1641 /* 1642 * new snapshots need to be created at a very specific time in the 1643 * transaction commit. This does the actual creation. 1644 * 1645 * Note: 1646 * If the error which may affect the commitment of the current transaction 1647 * happens, we should return the error number. If the error which just affect 1648 * the creation of the pending snapshots, just return 0. 1649 */ 1650 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans, 1651 struct btrfs_pending_snapshot *pending) 1652 { 1653 1654 struct btrfs_fs_info *fs_info = trans->fs_info; 1655 struct btrfs_key key; 1656 struct btrfs_root_item *new_root_item; 1657 struct btrfs_root *tree_root = fs_info->tree_root; 1658 struct btrfs_root *root = pending->root; 1659 struct btrfs_root *parent_root; 1660 struct btrfs_block_rsv *rsv; 1661 struct btrfs_inode *parent_inode = pending->dir; 1662 BTRFS_PATH_AUTO_FREE(path); 1663 struct btrfs_dir_item *dir_item; 1664 struct extent_buffer *tmp; 1665 struct extent_buffer *old; 1666 struct timespec64 cur_time; 1667 int ret = 0; 1668 u64 to_reserve = 0; 1669 u64 index = 0; 1670 u64 objectid; 1671 u64 root_flags; 1672 unsigned int nofs_flags; 1673 struct fscrypt_name fname; 1674 1675 ASSERT(pending->path); 1676 path = pending->path; 1677 1678 ASSERT(pending->root_item); 1679 new_root_item = pending->root_item; 1680 1681 /* 1682 * We're inside a transaction and must make sure that any potential 1683 * allocations with GFP_KERNEL in fscrypt won't recurse back to 1684 * filesystem. 1685 */ 1686 nofs_flags = memalloc_nofs_save(); 1687 pending->error = fscrypt_setup_filename(&parent_inode->vfs_inode, 1688 &pending->dentry->d_name, 0, 1689 &fname); 1690 memalloc_nofs_restore(nofs_flags); 1691 if (unlikely(pending->error)) 1692 goto free_pending; 1693 1694 pending->error = btrfs_get_free_objectid(tree_root, &objectid); 1695 if (unlikely(pending->error)) 1696 goto free_fname; 1697 1698 /* 1699 * Make qgroup to skip current new snapshot's qgroupid, as it is 1700 * accounted by later btrfs_qgroup_inherit(). 1701 */ 1702 btrfs_set_skip_qgroup(trans, objectid); 1703 1704 btrfs_reloc_pre_snapshot(pending, &to_reserve); 1705 1706 if (to_reserve > 0) { 1707 pending->error = btrfs_block_rsv_add(fs_info, 1708 &pending->block_rsv, 1709 to_reserve, 1710 BTRFS_RESERVE_NO_FLUSH); 1711 if (unlikely(pending->error)) 1712 goto clear_skip_qgroup; 1713 } 1714 1715 rsv = trans->block_rsv; 1716 trans->block_rsv = &pending->block_rsv; 1717 trans->bytes_reserved = trans->block_rsv->reserved; 1718 trace_btrfs_space_reservation(fs_info, "transaction", 1719 trans->transid, 1720 trans->bytes_reserved, 1); 1721 parent_root = parent_inode->root; 1722 ret = record_root_in_trans(trans, parent_root, 0); 1723 if (unlikely(ret)) 1724 goto fail; 1725 cur_time = current_time(&parent_inode->vfs_inode); 1726 1727 /* 1728 * insert the directory item 1729 */ 1730 ret = btrfs_set_inode_index(parent_inode, &index); 1731 if (unlikely(ret)) { 1732 btrfs_abort_transaction(trans, ret); 1733 goto fail; 1734 } 1735 1736 /* check if there is a file/dir which has the same name. */ 1737 dir_item = btrfs_lookup_dir_item(NULL, parent_root, path, 1738 btrfs_ino(parent_inode), 1739 &fname.disk_name, 0); 1740 if (unlikely(dir_item != NULL && !IS_ERR(dir_item))) { 1741 pending->error = -EEXIST; 1742 goto dir_item_existed; 1743 } else if (IS_ERR(dir_item)) { 1744 ret = PTR_ERR(dir_item); 1745 btrfs_abort_transaction(trans, ret); 1746 goto fail; 1747 } 1748 btrfs_release_path(path); 1749 1750 ret = btrfs_create_qgroup(trans, objectid); 1751 if (ret && ret != -EEXIST) { 1752 if (unlikely(ret != -ENOTCONN || btrfs_qgroup_enabled(fs_info))) { 1753 btrfs_abort_transaction(trans, ret); 1754 goto fail; 1755 } 1756 } 1757 1758 /* 1759 * pull in the delayed directory update 1760 * and the delayed inode item 1761 * otherwise we corrupt the FS during 1762 * snapshot 1763 */ 1764 ret = btrfs_run_delayed_items(trans); 1765 if (unlikely(ret)) { 1766 btrfs_abort_transaction(trans, ret); 1767 goto fail; 1768 } 1769 1770 ret = record_root_in_trans(trans, root, 0); 1771 if (unlikely(ret)) { 1772 btrfs_abort_transaction(trans, ret); 1773 goto fail; 1774 } 1775 btrfs_set_root_last_snapshot(&root->root_item, trans->transid); 1776 memcpy(new_root_item, &root->root_item, sizeof(*new_root_item)); 1777 btrfs_check_and_init_root_item(new_root_item); 1778 1779 root_flags = btrfs_root_flags(new_root_item); 1780 if (pending->readonly) 1781 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY; 1782 else 1783 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY; 1784 btrfs_set_root_flags(new_root_item, root_flags); 1785 1786 btrfs_set_root_generation_v2(new_root_item, 1787 trans->transid); 1788 generate_random_guid(new_root_item->uuid); 1789 memcpy(new_root_item->parent_uuid, root->root_item.uuid, 1790 BTRFS_UUID_SIZE); 1791 if (!(root_flags & BTRFS_ROOT_SUBVOL_RDONLY)) { 1792 memset(new_root_item->received_uuid, 0, 1793 sizeof(new_root_item->received_uuid)); 1794 memset(&new_root_item->stime, 0, sizeof(new_root_item->stime)); 1795 memset(&new_root_item->rtime, 0, sizeof(new_root_item->rtime)); 1796 btrfs_set_root_stransid(new_root_item, 0); 1797 btrfs_set_root_rtransid(new_root_item, 0); 1798 } 1799 btrfs_set_stack_timespec_sec(&new_root_item->otime, cur_time.tv_sec); 1800 btrfs_set_stack_timespec_nsec(&new_root_item->otime, cur_time.tv_nsec); 1801 btrfs_set_root_otransid(new_root_item, trans->transid); 1802 1803 old = btrfs_lock_root_node(root); 1804 ret = btrfs_cow_block(trans, root, old, NULL, 0, &old, 1805 BTRFS_NESTING_COW); 1806 if (unlikely(ret)) { 1807 btrfs_tree_unlock(old); 1808 free_extent_buffer(old); 1809 btrfs_abort_transaction(trans, ret); 1810 goto fail; 1811 } 1812 1813 ret = btrfs_copy_root(trans, root, old, &tmp, objectid); 1814 /* clean up in any case */ 1815 btrfs_tree_unlock(old); 1816 free_extent_buffer(old); 1817 if (unlikely(ret)) { 1818 btrfs_abort_transaction(trans, ret); 1819 goto fail; 1820 } 1821 /* see comments in should_cow_block() */ 1822 set_bit(BTRFS_ROOT_FORCE_COW, &root->state); 1823 smp_mb__after_atomic(); 1824 1825 btrfs_set_root_node(new_root_item, tmp); 1826 /* record when the snapshot was created in key.offset */ 1827 key.objectid = objectid; 1828 key.type = BTRFS_ROOT_ITEM_KEY; 1829 key.offset = trans->transid; 1830 ret = btrfs_insert_root(trans, tree_root, &key, new_root_item); 1831 btrfs_tree_unlock(tmp); 1832 free_extent_buffer(tmp); 1833 if (unlikely(ret)) { 1834 btrfs_abort_transaction(trans, ret); 1835 goto fail; 1836 } 1837 1838 /* 1839 * insert root back/forward references 1840 */ 1841 ret = btrfs_add_root_ref(trans, objectid, 1842 btrfs_root_id(parent_root), 1843 btrfs_ino(parent_inode), index, 1844 &fname.disk_name); 1845 if (unlikely(ret)) { 1846 btrfs_abort_transaction(trans, ret); 1847 goto fail; 1848 } 1849 1850 key.offset = (u64)-1; 1851 pending->snap = btrfs_get_new_fs_root(fs_info, objectid, &pending->anon_dev); 1852 if (IS_ERR(pending->snap)) { 1853 ret = PTR_ERR(pending->snap); 1854 pending->snap = NULL; 1855 btrfs_abort_transaction(trans, ret); 1856 goto fail; 1857 } 1858 1859 ret = btrfs_reloc_post_snapshot(trans, pending); 1860 if (unlikely(ret)) { 1861 btrfs_abort_transaction(trans, ret); 1862 goto fail; 1863 } 1864 1865 /* 1866 * Do special qgroup accounting for snapshot, as we do some qgroup 1867 * snapshot hack to do fast snapshot. 1868 * To co-operate with that hack, we do hack again. 1869 * Or snapshot will be greatly slowed down by a subtree qgroup rescan 1870 */ 1871 if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_FULL) 1872 ret = qgroup_account_snapshot(trans, root, parent_root, 1873 pending->inherit, objectid); 1874 else if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_SIMPLE) 1875 ret = btrfs_qgroup_inherit(trans, btrfs_root_id(root), objectid, 1876 btrfs_root_id(parent_root), pending->inherit); 1877 if (unlikely(ret < 0)) 1878 goto fail; 1879 1880 ret = btrfs_insert_dir_item(trans, &fname.disk_name, 1881 parent_inode, &key, BTRFS_FT_DIR, 1882 index); 1883 if (unlikely(ret)) { 1884 btrfs_abort_transaction(trans, ret); 1885 goto fail; 1886 } 1887 1888 btrfs_i_size_write(parent_inode, parent_inode->vfs_inode.i_size + 1889 fname.disk_name.len * 2); 1890 inode_set_mtime_to_ts(&parent_inode->vfs_inode, 1891 inode_set_ctime_current(&parent_inode->vfs_inode)); 1892 ret = btrfs_update_inode_fallback(trans, parent_inode); 1893 if (unlikely(ret)) { 1894 btrfs_abort_transaction(trans, ret); 1895 goto fail; 1896 } 1897 ret = btrfs_uuid_tree_add(trans, new_root_item->uuid, 1898 BTRFS_UUID_KEY_SUBVOL, 1899 objectid); 1900 if (unlikely(ret)) { 1901 btrfs_abort_transaction(trans, ret); 1902 goto fail; 1903 } 1904 if (!btrfs_is_empty_uuid(new_root_item->received_uuid)) { 1905 ret = btrfs_uuid_tree_add(trans, new_root_item->received_uuid, 1906 BTRFS_UUID_KEY_RECEIVED_SUBVOL, 1907 objectid); 1908 /* 1909 * We are creating of lot of snapshots of the same root that was 1910 * received (has a received UUID) and reached a leaf's limit for 1911 * an item. We can safely ignore this and avoid a transaction 1912 * abort. A deletion of this snapshot will still work since we 1913 * ignore if an item with a BTRFS_UUID_KEY_RECEIVED_SUBVOL key 1914 * is missing (see btrfs_delete_subvolume()). Send/receive will 1915 * work too since it peeks the first root id from the existing 1916 * item (it could peek any), and in case it's missing it 1917 * falls back to search by BTRFS_UUID_KEY_SUBVOL keys. 1918 * Creation of a snapshot does not require CAP_SYS_ADMIN, so 1919 * we don't want users triggering transaction aborts, either 1920 * intentionally or not. 1921 */ 1922 if (ret == -EOVERFLOW) 1923 ret = 0; 1924 if (unlikely(ret && ret != -EEXIST)) { 1925 btrfs_abort_transaction(trans, ret); 1926 goto fail; 1927 } 1928 } 1929 1930 fail: 1931 pending->error = ret; 1932 dir_item_existed: 1933 trans->block_rsv = rsv; 1934 trans->bytes_reserved = 0; 1935 clear_skip_qgroup: 1936 btrfs_clear_skip_qgroup(trans); 1937 free_fname: 1938 fscrypt_free_filename(&fname); 1939 free_pending: 1940 kfree(new_root_item); 1941 pending->root_item = NULL; 1942 pending->path = NULL; 1943 1944 return ret; 1945 } 1946 1947 /* 1948 * create all the snapshots we've scheduled for creation 1949 */ 1950 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans) 1951 { 1952 struct btrfs_pending_snapshot *pending, *next; 1953 struct list_head *head = &trans->transaction->pending_snapshots; 1954 int ret = 0; 1955 1956 list_for_each_entry_safe(pending, next, head, list) { 1957 list_del(&pending->list); 1958 ret = create_pending_snapshot(trans, pending); 1959 if (unlikely(ret)) 1960 break; 1961 } 1962 return ret; 1963 } 1964 1965 static void update_super_roots(struct btrfs_fs_info *fs_info) 1966 { 1967 struct btrfs_root_item *root_item; 1968 struct btrfs_super_block *super; 1969 1970 super = fs_info->super_copy; 1971 1972 root_item = &fs_info->chunk_root->root_item; 1973 super->chunk_root = root_item->bytenr; 1974 super->chunk_root_generation = root_item->generation; 1975 super->chunk_root_level = root_item->level; 1976 1977 root_item = &fs_info->tree_root->root_item; 1978 super->root = root_item->bytenr; 1979 super->generation = root_item->generation; 1980 super->root_level = root_item->level; 1981 if (btrfs_test_opt(fs_info, SPACE_CACHE)) 1982 super->cache_generation = root_item->generation; 1983 else if (test_bit(BTRFS_FS_CLEANUP_SPACE_CACHE_V1, &fs_info->flags)) 1984 super->cache_generation = 0; 1985 if (test_bit(BTRFS_FS_UPDATE_UUID_TREE_GEN, &fs_info->flags)) 1986 super->uuid_tree_generation = root_item->generation; 1987 1988 if (btrfs_fs_incompat(fs_info, REMAP_TREE)) { 1989 root_item = &fs_info->remap_root->root_item; 1990 super->remap_root = root_item->bytenr; 1991 super->remap_root_generation = root_item->generation; 1992 super->remap_root_level = root_item->level; 1993 } 1994 } 1995 1996 int btrfs_transaction_blocked(struct btrfs_fs_info *info) 1997 { 1998 struct btrfs_transaction *trans; 1999 int ret = 0; 2000 2001 spin_lock(&info->trans_lock); 2002 trans = info->running_transaction; 2003 if (trans) 2004 ret = is_transaction_blocked(trans); 2005 spin_unlock(&info->trans_lock); 2006 return ret; 2007 } 2008 2009 void btrfs_commit_transaction_async(struct btrfs_trans_handle *trans) 2010 { 2011 struct btrfs_fs_info *fs_info = trans->fs_info; 2012 struct btrfs_transaction *cur_trans; 2013 2014 /* Kick the transaction kthread. */ 2015 set_bit(BTRFS_FS_COMMIT_TRANS, &fs_info->flags); 2016 wake_up_process(fs_info->transaction_kthread); 2017 2018 /* take transaction reference */ 2019 cur_trans = trans->transaction; 2020 refcount_inc(&cur_trans->use_count); 2021 2022 btrfs_end_transaction(trans); 2023 2024 /* 2025 * Wait for the current transaction commit to start and block 2026 * subsequent transaction joins 2027 */ 2028 btrfs_might_wait_for_state(fs_info, BTRFS_LOCKDEP_TRANS_COMMIT_PREP); 2029 wait_event(fs_info->transaction_blocked_wait, 2030 cur_trans->state >= TRANS_STATE_COMMIT_START || 2031 TRANS_ABORTED(cur_trans)); 2032 btrfs_put_transaction(cur_trans); 2033 } 2034 2035 /* 2036 * If there is a running transaction commit it or if it's already committing, 2037 * wait for its commit to complete. Does not start and commit a new transaction 2038 * if there isn't any running. 2039 */ 2040 int btrfs_commit_current_transaction(struct btrfs_root *root) 2041 { 2042 struct btrfs_trans_handle *trans; 2043 2044 trans = btrfs_attach_transaction_barrier(root); 2045 if (IS_ERR(trans)) { 2046 int ret = PTR_ERR(trans); 2047 2048 return (ret == -ENOENT) ? 0 : ret; 2049 } 2050 2051 return btrfs_commit_transaction(trans); 2052 } 2053 2054 static void cleanup_transaction(struct btrfs_trans_handle *trans, int err) 2055 { 2056 struct btrfs_fs_info *fs_info = trans->fs_info; 2057 struct btrfs_transaction *cur_trans = trans->transaction; 2058 2059 WARN_ON(refcount_read(&trans->use_count) > 1); 2060 2061 btrfs_abort_transaction(trans, err); 2062 2063 spin_lock(&fs_info->trans_lock); 2064 2065 /* 2066 * If the transaction is removed from the list, it means this 2067 * transaction has been committed successfully, so it is impossible 2068 * to call the cleanup function. 2069 */ 2070 BUG_ON(list_empty(&cur_trans->list)); 2071 2072 if (cur_trans == fs_info->running_transaction) { 2073 cur_trans->state = TRANS_STATE_COMMIT_DOING; 2074 spin_unlock(&fs_info->trans_lock); 2075 2076 /* 2077 * The thread has already released the lockdep map as reader 2078 * already in btrfs_commit_transaction(). 2079 */ 2080 btrfs_might_wait_for_event(fs_info, btrfs_trans_num_writers); 2081 wait_event(cur_trans->writer_wait, 2082 atomic_read(&cur_trans->num_writers) == 1); 2083 2084 spin_lock(&fs_info->trans_lock); 2085 } 2086 2087 /* 2088 * Now that we know no one else is still using the transaction we can 2089 * remove the transaction from the list of transactions. This avoids 2090 * the transaction kthread from cleaning up the transaction while some 2091 * other task is still using it, which could result in a use-after-free 2092 * on things like log trees, as it forces the transaction kthread to 2093 * wait for this transaction to be cleaned up by us. 2094 */ 2095 list_del_init(&cur_trans->list); 2096 2097 spin_unlock(&fs_info->trans_lock); 2098 2099 btrfs_cleanup_one_transaction(trans->transaction); 2100 2101 spin_lock(&fs_info->trans_lock); 2102 if (cur_trans == fs_info->running_transaction) 2103 fs_info->running_transaction = NULL; 2104 spin_unlock(&fs_info->trans_lock); 2105 2106 if (trans->type & __TRANS_FREEZABLE) 2107 sb_end_intwrite(fs_info->sb); 2108 btrfs_put_transaction(cur_trans); 2109 btrfs_put_transaction(cur_trans); 2110 2111 trace_btrfs_transaction_commit(fs_info); 2112 2113 if (current->journal_info == trans) 2114 current->journal_info = NULL; 2115 2116 /* 2117 * If relocation is running, we can't cancel scrub because that will 2118 * result in a deadlock. Before relocating a block group, relocation 2119 * pauses scrub, then starts and commits a transaction before unpausing 2120 * scrub. If the transaction commit is being done by the relocation 2121 * task or triggered by another task and the relocation task is waiting 2122 * for the commit, and we end up here due to an error in the commit 2123 * path, then calling btrfs_scrub_cancel() will deadlock, as we are 2124 * asking for scrub to stop while having it asked to be paused higher 2125 * above in relocation code. 2126 */ 2127 if (!test_bit(BTRFS_FS_RELOC_RUNNING, &fs_info->flags)) 2128 btrfs_scrub_cancel(fs_info); 2129 2130 kmem_cache_free(btrfs_trans_handle_cachep, trans); 2131 } 2132 2133 /* 2134 * Release reserved delayed ref space of all pending block groups of the 2135 * transaction and remove them from the list 2136 */ 2137 static void btrfs_cleanup_pending_block_groups(struct btrfs_trans_handle *trans) 2138 { 2139 struct btrfs_fs_info *fs_info = trans->fs_info; 2140 struct btrfs_block_group *block_group, *tmp; 2141 2142 list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) { 2143 btrfs_dec_delayed_refs_rsv_bg_inserts(fs_info); 2144 /* 2145 * Not strictly necessary to lock, as no other task will be using a 2146 * block_group on the new_bgs list during a transaction abort. 2147 */ 2148 spin_lock(&fs_info->unused_bgs_lock); 2149 list_del_init(&block_group->bg_list); 2150 btrfs_put_block_group(block_group); 2151 spin_unlock(&fs_info->unused_bgs_lock); 2152 } 2153 } 2154 2155 static inline int btrfs_start_delalloc_flush(struct btrfs_fs_info *fs_info) 2156 { 2157 /* 2158 * We use try_to_writeback_inodes_sb() here because if we used 2159 * btrfs_start_delalloc_roots we would deadlock with fs freeze. 2160 * Currently are holding the fs freeze lock, if we do an async flush 2161 * we'll do btrfs_join_transaction() and deadlock because we need to 2162 * wait for the fs freeze lock. Using the direct flushing we benefit 2163 * from already being in a transaction and our join_transaction doesn't 2164 * have to re-take the fs freeze lock. 2165 * 2166 * Note that try_to_writeback_inodes_sb() will only trigger writeback 2167 * if it can read lock sb->s_umount. It will always be able to lock it, 2168 * except when the filesystem is being unmounted or being frozen, but in 2169 * those cases sync_filesystem() is called, which results in calling 2170 * writeback_inodes_sb() while holding a write lock on sb->s_umount. 2171 * Note that we don't call writeback_inodes_sb() directly, because it 2172 * will emit a warning if sb->s_umount is not locked. 2173 */ 2174 if (btrfs_test_opt(fs_info, FLUSHONCOMMIT)) 2175 try_to_writeback_inodes_sb(fs_info->sb, WB_REASON_SYNC); 2176 return 0; 2177 } 2178 2179 static inline void btrfs_wait_delalloc_flush(struct btrfs_fs_info *fs_info) 2180 { 2181 if (btrfs_test_opt(fs_info, FLUSHONCOMMIT)) 2182 btrfs_wait_ordered_roots(fs_info, U64_MAX, NULL); 2183 } 2184 2185 /* 2186 * Add a pending snapshot associated with the given transaction handle to the 2187 * respective handle. This must be called after the transaction commit started 2188 * and while holding fs_info->trans_lock. 2189 * This serves to guarantee a caller of btrfs_commit_transaction() that it can 2190 * safely free the pending snapshot pointer in case btrfs_commit_transaction() 2191 * returns an error. 2192 */ 2193 static void add_pending_snapshot(struct btrfs_trans_handle *trans) 2194 { 2195 struct btrfs_transaction *cur_trans = trans->transaction; 2196 2197 if (!trans->pending_snapshot) 2198 return; 2199 2200 lockdep_assert_held(&trans->fs_info->trans_lock); 2201 ASSERT(cur_trans->state >= TRANS_STATE_COMMIT_PREP, 2202 "cur_trans->state=%d", cur_trans->state); 2203 2204 list_add(&trans->pending_snapshot->list, &cur_trans->pending_snapshots); 2205 } 2206 2207 static void update_commit_stats(struct btrfs_fs_info *fs_info) 2208 { 2209 ktime_t now = ktime_get_ns(); 2210 ktime_t interval = now - fs_info->commit_stats.critical_section_start_time; 2211 2212 ASSERT(fs_info->commit_stats.critical_section_start_time); 2213 2214 fs_info->commit_stats.commit_count++; 2215 fs_info->commit_stats.last_commit_dur = interval; 2216 fs_info->commit_stats.max_commit_dur = 2217 max_t(u64, fs_info->commit_stats.max_commit_dur, interval); 2218 fs_info->commit_stats.total_commit_dur += interval; 2219 fs_info->commit_stats.critical_section_start_time = 0; 2220 } 2221 2222 int btrfs_commit_transaction(struct btrfs_trans_handle *trans) 2223 { 2224 struct btrfs_fs_info *fs_info = trans->fs_info; 2225 struct btrfs_transaction *cur_trans = trans->transaction; 2226 struct btrfs_transaction *prev_trans = NULL; 2227 int ret; 2228 2229 ASSERT(refcount_read(&trans->use_count) == 1, 2230 "refcount_read(&trans->use_count)=%d", refcount_read(&trans->use_count)); 2231 btrfs_trans_state_lockdep_acquire(fs_info, BTRFS_LOCKDEP_TRANS_COMMIT_PREP); 2232 2233 clear_bit(BTRFS_FS_NEED_TRANS_COMMIT, &fs_info->flags); 2234 2235 /* Stop the commit early if ->aborted is set */ 2236 if (TRANS_ABORTED(cur_trans)) { 2237 ret = cur_trans->aborted; 2238 goto lockdep_trans_commit_start_release; 2239 } 2240 2241 btrfs_trans_release_metadata(trans); 2242 trans->block_rsv = NULL; 2243 2244 /* 2245 * We only want one transaction commit doing the flushing so we do not 2246 * waste a bunch of time on lock contention on the extent root node. 2247 */ 2248 if (!test_and_set_bit(BTRFS_DELAYED_REFS_FLUSHING, 2249 &cur_trans->delayed_refs.flags)) { 2250 /* 2251 * Make a pass through all the delayed refs we have so far. 2252 * Any running threads may add more while we are here. 2253 */ 2254 ret = btrfs_run_delayed_refs(trans, 0); 2255 if (ret) 2256 goto lockdep_trans_commit_start_release; 2257 } 2258 2259 btrfs_create_pending_block_groups(trans); 2260 2261 if (!test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &cur_trans->flags)) { 2262 int run_it = 0; 2263 2264 /* this mutex is also taken before trying to set 2265 * block groups readonly. We need to make sure 2266 * that nobody has set a block group readonly 2267 * after a extents from that block group have been 2268 * allocated for cache files. btrfs_set_block_group_ro 2269 * will wait for the transaction to commit if it 2270 * finds BTRFS_TRANS_DIRTY_BG_RUN set. 2271 * 2272 * The BTRFS_TRANS_DIRTY_BG_RUN flag is also used to make sure 2273 * only one process starts all the block group IO. It wouldn't 2274 * hurt to have more than one go through, but there's no 2275 * real advantage to it either. 2276 */ 2277 mutex_lock(&fs_info->ro_block_group_mutex); 2278 if (!test_and_set_bit(BTRFS_TRANS_DIRTY_BG_RUN, 2279 &cur_trans->flags)) 2280 run_it = 1; 2281 mutex_unlock(&fs_info->ro_block_group_mutex); 2282 2283 if (run_it) { 2284 ret = btrfs_start_dirty_block_groups(trans); 2285 if (unlikely(ret)) 2286 goto lockdep_trans_commit_start_release; 2287 } 2288 } 2289 2290 spin_lock(&fs_info->trans_lock); 2291 if (cur_trans->state >= TRANS_STATE_COMMIT_PREP) { 2292 enum btrfs_trans_state want_state = TRANS_STATE_COMPLETED; 2293 2294 add_pending_snapshot(trans); 2295 2296 spin_unlock(&fs_info->trans_lock); 2297 refcount_inc(&cur_trans->use_count); 2298 2299 if (trans->in_fsync) 2300 want_state = TRANS_STATE_SUPER_COMMITTED; 2301 2302 btrfs_trans_state_lockdep_release(fs_info, 2303 BTRFS_LOCKDEP_TRANS_COMMIT_PREP); 2304 ret = btrfs_end_transaction(trans); 2305 wait_for_commit(cur_trans, want_state); 2306 2307 if (TRANS_ABORTED(cur_trans)) 2308 ret = cur_trans->aborted; 2309 2310 btrfs_put_transaction(cur_trans); 2311 2312 return ret; 2313 } 2314 2315 cur_trans->state = TRANS_STATE_COMMIT_PREP; 2316 wake_up(&fs_info->transaction_blocked_wait); 2317 btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_COMMIT_PREP); 2318 2319 if (!list_is_first(&cur_trans->list, &fs_info->trans_list)) { 2320 enum btrfs_trans_state want_state = TRANS_STATE_COMPLETED; 2321 2322 if (trans->in_fsync) 2323 want_state = TRANS_STATE_SUPER_COMMITTED; 2324 2325 prev_trans = list_prev_entry(cur_trans, list); 2326 if (prev_trans->state < want_state) { 2327 refcount_inc(&prev_trans->use_count); 2328 spin_unlock(&fs_info->trans_lock); 2329 2330 wait_for_commit(prev_trans, want_state); 2331 2332 ret = READ_ONCE(prev_trans->aborted); 2333 2334 btrfs_put_transaction(prev_trans); 2335 if (unlikely(ret)) 2336 goto lockdep_release; 2337 spin_lock(&fs_info->trans_lock); 2338 } 2339 } else { 2340 /* 2341 * The previous transaction was aborted and was already removed 2342 * from the list of transactions at fs_info->trans_list. So we 2343 * abort to prevent writing a new superblock that reflects a 2344 * corrupt state (pointing to trees with unwritten nodes/leafs). 2345 */ 2346 if (BTRFS_FS_ERROR(fs_info)) { 2347 spin_unlock(&fs_info->trans_lock); 2348 ret = -EROFS; 2349 goto lockdep_release; 2350 } 2351 } 2352 2353 cur_trans->state = TRANS_STATE_COMMIT_START; 2354 wake_up(&fs_info->transaction_blocked_wait); 2355 spin_unlock(&fs_info->trans_lock); 2356 2357 /* 2358 * Get the time spent on the work done by the commit thread and not 2359 * the time spent waiting on a previous commit 2360 */ 2361 fs_info->commit_stats.critical_section_start_time = ktime_get_ns(); 2362 extwriter_counter_dec(cur_trans, trans->type); 2363 2364 ret = btrfs_start_delalloc_flush(fs_info); 2365 if (unlikely(ret)) 2366 goto lockdep_release; 2367 2368 ret = btrfs_run_delayed_items(trans); 2369 if (unlikely(ret)) 2370 goto lockdep_release; 2371 2372 /* 2373 * The thread has started/joined the transaction thus it holds the 2374 * lockdep map as a reader. It has to release it before acquiring the 2375 * lockdep map as a writer. 2376 */ 2377 btrfs_lockdep_release(fs_info, btrfs_trans_num_extwriters); 2378 btrfs_might_wait_for_event(fs_info, btrfs_trans_num_extwriters); 2379 wait_event(cur_trans->writer_wait, 2380 extwriter_counter_read(cur_trans) == 0); 2381 2382 /* some pending stuffs might be added after the previous flush. */ 2383 ret = btrfs_run_delayed_items(trans); 2384 if (unlikely(ret)) { 2385 btrfs_lockdep_release(fs_info, btrfs_trans_num_writers); 2386 goto cleanup_transaction; 2387 } 2388 2389 btrfs_wait_delalloc_flush(fs_info); 2390 2391 /* 2392 * Wait for all ordered extents started by a fast fsync that joined this 2393 * transaction. Otherwise if this transaction commits before the ordered 2394 * extents complete we lose logged data after a power failure. 2395 */ 2396 btrfs_might_wait_for_event(fs_info, btrfs_trans_pending_ordered); 2397 wait_event(cur_trans->pending_wait, 2398 atomic_read(&cur_trans->pending_ordered) == 0); 2399 2400 btrfs_scrub_pause(fs_info); 2401 /* 2402 * Ok now we need to make sure to block out any other joins while we 2403 * commit the transaction. We could have started a join before setting 2404 * COMMIT_DOING so make sure to wait for num_writers to == 1 again. 2405 */ 2406 spin_lock(&fs_info->trans_lock); 2407 add_pending_snapshot(trans); 2408 cur_trans->state = TRANS_STATE_COMMIT_DOING; 2409 spin_unlock(&fs_info->trans_lock); 2410 2411 /* 2412 * The thread has started/joined the transaction thus it holds the 2413 * lockdep map as a reader. It has to release it before acquiring the 2414 * lockdep map as a writer. 2415 */ 2416 btrfs_lockdep_release(fs_info, btrfs_trans_num_writers); 2417 btrfs_might_wait_for_event(fs_info, btrfs_trans_num_writers); 2418 wait_event(cur_trans->writer_wait, 2419 atomic_read(&cur_trans->num_writers) == 1); 2420 2421 /* 2422 * Make lockdep happy by acquiring the state locks after 2423 * btrfs_trans_num_writers is released. If we acquired the state locks 2424 * before releasing the btrfs_trans_num_writers lock then lockdep would 2425 * complain because we did not follow the reverse order unlocking rule. 2426 */ 2427 btrfs_trans_state_lockdep_acquire(fs_info, BTRFS_LOCKDEP_TRANS_COMPLETED); 2428 btrfs_trans_state_lockdep_acquire(fs_info, BTRFS_LOCKDEP_TRANS_SUPER_COMMITTED); 2429 btrfs_trans_state_lockdep_acquire(fs_info, BTRFS_LOCKDEP_TRANS_UNBLOCKED); 2430 2431 /* 2432 * We've started the commit, clear the flag in case we were triggered to 2433 * do an async commit but somebody else started before the transaction 2434 * kthread could do the work. 2435 */ 2436 clear_bit(BTRFS_FS_COMMIT_TRANS, &fs_info->flags); 2437 2438 if (TRANS_ABORTED(cur_trans)) { 2439 ret = cur_trans->aborted; 2440 btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_UNBLOCKED); 2441 goto scrub_continue; 2442 } 2443 /* 2444 * the reloc mutex makes sure that we stop 2445 * the balancing code from coming in and moving 2446 * extents around in the middle of the commit 2447 */ 2448 mutex_lock(&fs_info->reloc_mutex); 2449 2450 /* 2451 * We needn't worry about the delayed items because we will 2452 * deal with them in create_pending_snapshot(), which is the 2453 * core function of the snapshot creation. 2454 */ 2455 ret = create_pending_snapshots(trans); 2456 if (unlikely(ret)) 2457 goto unlock_reloc; 2458 2459 /* 2460 * We insert the dir indexes of the snapshots and update the inode 2461 * of the snapshots' parents after the snapshot creation, so there 2462 * are some delayed items which are not dealt with. Now deal with 2463 * them. 2464 * 2465 * We needn't worry that this operation will corrupt the snapshots, 2466 * because all the tree which are snapshotted will be forced to COW 2467 * the nodes and leaves. 2468 */ 2469 ret = btrfs_run_delayed_items(trans); 2470 if (unlikely(ret)) 2471 goto unlock_reloc; 2472 2473 ret = btrfs_run_delayed_refs(trans, U64_MAX); 2474 if (unlikely(ret)) 2475 goto unlock_reloc; 2476 2477 /* 2478 * make sure none of the code above managed to slip in a 2479 * delayed item 2480 */ 2481 btrfs_assert_delayed_root_empty(fs_info); 2482 2483 WARN_ON(cur_trans != trans->transaction); 2484 2485 ret = commit_fs_roots(trans); 2486 if (unlikely(ret)) 2487 goto unlock_reloc; 2488 2489 /* commit_fs_roots gets rid of all the tree log roots, it is now 2490 * safe to free the root of tree log roots 2491 */ 2492 btrfs_free_log_root_tree(trans, fs_info); 2493 2494 /* 2495 * Since fs roots are all committed, we can get a quite accurate 2496 * new_roots. So let's do quota accounting. 2497 */ 2498 ret = btrfs_qgroup_account_extents(trans); 2499 if (unlikely(ret < 0)) 2500 goto unlock_reloc; 2501 2502 ret = commit_cowonly_roots(trans); 2503 if (unlikely(ret)) 2504 goto unlock_reloc; 2505 2506 /* 2507 * The tasks which save the space cache and inode cache may also 2508 * update ->aborted, check it. 2509 */ 2510 if (TRANS_ABORTED(cur_trans)) { 2511 ret = cur_trans->aborted; 2512 goto unlock_reloc; 2513 } 2514 2515 cur_trans = fs_info->running_transaction; 2516 2517 btrfs_set_root_node(&fs_info->tree_root->root_item, 2518 fs_info->tree_root->node); 2519 list_add_tail(&fs_info->tree_root->dirty_list, 2520 &cur_trans->switch_commits); 2521 2522 btrfs_set_root_node(&fs_info->chunk_root->root_item, 2523 fs_info->chunk_root->node); 2524 list_add_tail(&fs_info->chunk_root->dirty_list, 2525 &cur_trans->switch_commits); 2526 2527 switch_commit_roots(trans); 2528 2529 ASSERT(list_empty(&cur_trans->dirty_bgs)); 2530 ASSERT(list_empty(&cur_trans->io_bgs)); 2531 update_super_roots(fs_info); 2532 2533 btrfs_set_super_log_root(fs_info->super_copy, 0); 2534 btrfs_set_super_log_root_level(fs_info->super_copy, 0); 2535 memcpy(fs_info->super_for_commit, fs_info->super_copy, 2536 sizeof(*fs_info->super_copy)); 2537 2538 btrfs_commit_device_sizes(cur_trans); 2539 2540 clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags); 2541 clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags); 2542 2543 btrfs_trans_release_chunk_metadata(trans); 2544 2545 /* 2546 * Before changing the transaction state to TRANS_STATE_UNBLOCKED and 2547 * setting fs_info->running_transaction to NULL, lock tree_log_mutex to 2548 * make sure that before we commit our superblock, no other task can 2549 * start a new transaction and commit a log tree before we commit our 2550 * superblock. Anyone trying to commit a log tree locks this mutex before 2551 * writing its superblock. 2552 */ 2553 mutex_lock(&fs_info->tree_log_mutex); 2554 2555 spin_lock(&fs_info->trans_lock); 2556 cur_trans->state = TRANS_STATE_UNBLOCKED; 2557 fs_info->running_transaction = NULL; 2558 spin_unlock(&fs_info->trans_lock); 2559 mutex_unlock(&fs_info->reloc_mutex); 2560 2561 wake_up(&fs_info->transaction_wait); 2562 btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_UNBLOCKED); 2563 2564 /* If we have features changed, wake up the cleaner to update sysfs. */ 2565 if (test_bit(BTRFS_FS_FEATURE_CHANGED, &fs_info->flags) && 2566 fs_info->cleaner_kthread) 2567 wake_up_process(fs_info->cleaner_kthread); 2568 2569 ret = btrfs_write_and_wait_transaction(trans); 2570 if (unlikely(ret)) { 2571 btrfs_err(fs_info, "error while writing out transaction: %d", ret); 2572 mutex_unlock(&fs_info->tree_log_mutex); 2573 goto scrub_continue; 2574 } 2575 2576 ret = write_all_supers(fs_info, 0); 2577 /* 2578 * the super is written, we can safely allow the tree-loggers 2579 * to go about their business 2580 */ 2581 mutex_unlock(&fs_info->tree_log_mutex); 2582 if (unlikely(ret)) 2583 goto scrub_continue; 2584 2585 update_commit_stats(fs_info); 2586 /* 2587 * We needn't acquire the lock here because there is no other task 2588 * which can change it. 2589 */ 2590 cur_trans->state = TRANS_STATE_SUPER_COMMITTED; 2591 wake_up(&cur_trans->commit_wait); 2592 btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_SUPER_COMMITTED); 2593 2594 ret = btrfs_finish_extent_commit(trans); 2595 if (unlikely(ret)) 2596 goto scrub_continue; 2597 2598 if (test_bit(BTRFS_TRANS_HAVE_FREE_BGS, &cur_trans->flags)) 2599 btrfs_clear_space_info_full(fs_info); 2600 2601 btrfs_set_last_trans_committed(fs_info, cur_trans->transid); 2602 /* 2603 * We needn't acquire the lock here because there is no other task 2604 * which can change it. 2605 */ 2606 cur_trans->state = TRANS_STATE_COMPLETED; 2607 wake_up(&cur_trans->commit_wait); 2608 btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_COMPLETED); 2609 2610 spin_lock(&fs_info->trans_lock); 2611 list_del_init(&cur_trans->list); 2612 spin_unlock(&fs_info->trans_lock); 2613 2614 btrfs_put_transaction(cur_trans); 2615 btrfs_put_transaction(cur_trans); 2616 2617 if (trans->type & __TRANS_FREEZABLE) 2618 sb_end_intwrite(fs_info->sb); 2619 2620 trace_btrfs_transaction_commit(fs_info); 2621 2622 btrfs_scrub_continue(fs_info); 2623 2624 if (current->journal_info == trans) 2625 current->journal_info = NULL; 2626 2627 kmem_cache_free(btrfs_trans_handle_cachep, trans); 2628 2629 return ret; 2630 2631 unlock_reloc: 2632 mutex_unlock(&fs_info->reloc_mutex); 2633 btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_UNBLOCKED); 2634 scrub_continue: 2635 btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_SUPER_COMMITTED); 2636 btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_COMPLETED); 2637 btrfs_scrub_continue(fs_info); 2638 cleanup_transaction: 2639 btrfs_trans_release_metadata(trans); 2640 btrfs_cleanup_pending_block_groups(trans); 2641 btrfs_trans_release_chunk_metadata(trans); 2642 trans->block_rsv = NULL; 2643 btrfs_warn(fs_info, "Skipping commit of aborted transaction."); 2644 if (current->journal_info == trans) 2645 current->journal_info = NULL; 2646 cleanup_transaction(trans, ret); 2647 2648 return ret; 2649 2650 lockdep_release: 2651 btrfs_lockdep_release(fs_info, btrfs_trans_num_extwriters); 2652 btrfs_lockdep_release(fs_info, btrfs_trans_num_writers); 2653 goto cleanup_transaction; 2654 2655 lockdep_trans_commit_start_release: 2656 btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_COMMIT_PREP); 2657 btrfs_end_transaction(trans); 2658 return ret; 2659 } 2660 2661 /* 2662 * return < 0 if error 2663 * 0 if there are no more dead_roots at the time of call 2664 * 1 there are more to be processed, call me again 2665 * 2666 * The return value indicates there are certainly more snapshots to delete, but 2667 * if there comes a new one during processing, it may return 0. We don't mind, 2668 * because btrfs_commit_super will poke cleaner thread and it will process it a 2669 * few seconds later. 2670 */ 2671 int btrfs_clean_one_deleted_snapshot(struct btrfs_fs_info *fs_info) 2672 { 2673 struct btrfs_root *root; 2674 int ret; 2675 2676 spin_lock(&fs_info->trans_lock); 2677 if (list_empty(&fs_info->dead_roots)) { 2678 spin_unlock(&fs_info->trans_lock); 2679 return 0; 2680 } 2681 root = list_first_entry(&fs_info->dead_roots, 2682 struct btrfs_root, root_list); 2683 list_del_init(&root->root_list); 2684 spin_unlock(&fs_info->trans_lock); 2685 2686 btrfs_debug(fs_info, "cleaner removing %llu", btrfs_root_id(root)); 2687 2688 btrfs_kill_all_delayed_nodes(root); 2689 2690 if (btrfs_header_backref_rev(root->node) < 2691 BTRFS_MIXED_BACKREF_REV) 2692 ret = btrfs_drop_snapshot(root, false, false); 2693 else 2694 ret = btrfs_drop_snapshot(root, true, false); 2695 2696 btrfs_put_root(root); 2697 return (ret < 0) ? 0 : 1; 2698 } 2699 2700 /* 2701 * We only mark the transaction aborted and then set the file system read-only. 2702 * This will prevent new transactions from starting or trying to join this 2703 * one. 2704 * 2705 * This means that error recovery at the call site is limited to freeing 2706 * any local memory allocations and passing the error code up without 2707 * further cleanup. The transaction should complete as it normally would 2708 * in the call path but will return -EIO. 2709 * 2710 * We'll complete the cleanup in btrfs_end_transaction and 2711 * btrfs_commit_transaction. 2712 */ 2713 void __cold __btrfs_abort_transaction(struct btrfs_trans_handle *trans, 2714 const char *function, 2715 unsigned int line, int error, bool first_hit) 2716 { 2717 struct btrfs_fs_info *fs_info = trans->fs_info; 2718 2719 WRITE_ONCE(trans->aborted, error); 2720 WRITE_ONCE(trans->transaction->aborted, error); 2721 if (first_hit && error == -ENOSPC) 2722 btrfs_dump_space_info_for_trans_abort(fs_info); 2723 /* Wake up anybody who may be waiting on this transaction */ 2724 wake_up(&fs_info->transaction_wait); 2725 wake_up(&fs_info->transaction_blocked_wait); 2726 __btrfs_handle_fs_error(fs_info, function, line, error, NULL); 2727 } 2728 2729 int __init btrfs_transaction_init(void) 2730 { 2731 btrfs_trans_handle_cachep = KMEM_CACHE(btrfs_trans_handle, SLAB_TEMPORARY); 2732 if (!btrfs_trans_handle_cachep) 2733 return -ENOMEM; 2734 return 0; 2735 } 2736 2737 void __cold btrfs_transaction_exit(void) 2738 { 2739 kmem_cache_destroy(btrfs_trans_handle_cachep); 2740 } 2741