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 36 static struct kmem_cache *btrfs_trans_handle_cachep; 37 38 /* 39 * Transaction states and transitions 40 * 41 * No running transaction (fs tree blocks are not modified) 42 * | 43 * | To next stage: 44 * | Call start_transaction() variants. Except btrfs_join_transaction_nostart(). 45 * V 46 * Transaction N [[TRANS_STATE_RUNNING]] 47 * | 48 * | New trans handles can be attached to transaction N by calling all 49 * | start_transaction() variants. 50 * | 51 * | To next stage: 52 * | Call btrfs_commit_transaction() on any trans handle attached to 53 * | transaction N 54 * V 55 * Transaction N [[TRANS_STATE_COMMIT_PREP]] 56 * | 57 * | If there are simultaneous calls to btrfs_commit_transaction() one will win 58 * | the race and the rest will wait for the winner to commit the transaction. 59 * | 60 * | The winner will wait for previous running transaction to completely finish 61 * | if there is one. 62 * | 63 * Transaction N [[TRANS_STATE_COMMIT_START]] 64 * | 65 * | Then one of the following happens: 66 * | - Wait for all other trans handle holders to release. 67 * | The btrfs_commit_transaction() caller will do the commit work. 68 * | - Wait for current transaction to be committed by others. 69 * | Other btrfs_commit_transaction() caller will do the commit work. 70 * | 71 * | At this stage, only btrfs_join_transaction*() variants can attach 72 * | to this running transaction. 73 * | All other variants will wait for current one to finish and attach to 74 * | transaction N+1. 75 * | 76 * | To next stage: 77 * | Caller is chosen to commit transaction N, and all other trans handle 78 * | haven been released. 79 * V 80 * Transaction N [[TRANS_STATE_COMMIT_DOING]] 81 * | 82 * | The heavy lifting transaction work is started. 83 * | From running delayed refs (modifying extent tree) to creating pending 84 * | snapshots, running qgroups. 85 * | In short, modify supporting trees to reflect modifications of subvolume 86 * | trees. 87 * | 88 * | At this stage, all start_transaction() calls will wait for this 89 * | transaction to finish and attach to transaction N+1. 90 * | 91 * | To next stage: 92 * | Until all supporting trees are updated. 93 * V 94 * Transaction N [[TRANS_STATE_UNBLOCKED]] 95 * | Transaction N+1 96 * | All needed trees are modified, thus we only [[TRANS_STATE_RUNNING]] 97 * | need to write them back to disk and update | 98 * | super blocks. | 99 * | | 100 * | At this stage, new transaction is allowed to | 101 * | start. | 102 * | All new start_transaction() calls will be | 103 * | attached to transid N+1. | 104 * | | 105 * | To next stage: | 106 * | Until all tree blocks are super blocks are | 107 * | written to block devices | 108 * V | 109 * Transaction N [[TRANS_STATE_COMPLETED]] V 110 * All tree blocks and super blocks are written. Transaction N+1 111 * This transaction is finished and all its [[TRANS_STATE_COMMIT_START]] 112 * data structures will be cleaned up. | Life goes on 113 */ 114 static const unsigned int btrfs_blocked_trans_types[TRANS_STATE_MAX] = { 115 [TRANS_STATE_RUNNING] = 0U, 116 [TRANS_STATE_COMMIT_PREP] = 0U, 117 [TRANS_STATE_COMMIT_START] = (__TRANS_START | __TRANS_ATTACH), 118 [TRANS_STATE_COMMIT_DOING] = (__TRANS_START | 119 __TRANS_ATTACH | 120 __TRANS_JOIN | 121 __TRANS_JOIN_NOSTART), 122 [TRANS_STATE_UNBLOCKED] = (__TRANS_START | 123 __TRANS_ATTACH | 124 __TRANS_JOIN | 125 __TRANS_JOIN_NOLOCK | 126 __TRANS_JOIN_NOSTART), 127 [TRANS_STATE_SUPER_COMMITTED] = (__TRANS_START | 128 __TRANS_ATTACH | 129 __TRANS_JOIN | 130 __TRANS_JOIN_NOLOCK | 131 __TRANS_JOIN_NOSTART), 132 [TRANS_STATE_COMPLETED] = (__TRANS_START | 133 __TRANS_ATTACH | 134 __TRANS_JOIN | 135 __TRANS_JOIN_NOLOCK | 136 __TRANS_JOIN_NOSTART), 137 }; 138 139 void btrfs_put_transaction(struct btrfs_transaction *transaction) 140 { 141 WARN_ON(refcount_read(&transaction->use_count) == 0); 142 if (refcount_dec_and_test(&transaction->use_count)) { 143 BUG_ON(!list_empty(&transaction->list)); 144 WARN_ON(!xa_empty(&transaction->delayed_refs.head_refs)); 145 WARN_ON(!xa_empty(&transaction->delayed_refs.dirty_extents)); 146 if (transaction->delayed_refs.pending_csums) 147 btrfs_err(transaction->fs_info, 148 "pending csums is %llu", 149 transaction->delayed_refs.pending_csums); 150 /* 151 * If any block groups are found in ->deleted_bgs then it's 152 * because the transaction was aborted and a commit did not 153 * happen (things failed before writing the new superblock 154 * and calling btrfs_finish_extent_commit()), so we can not 155 * discard the physical locations of the block groups. 156 */ 157 while (!list_empty(&transaction->deleted_bgs)) { 158 struct btrfs_block_group *cache; 159 160 cache = list_first_entry(&transaction->deleted_bgs, 161 struct btrfs_block_group, 162 bg_list); 163 /* 164 * Not strictly necessary to lock, as no other task will be using a 165 * block_group on the deleted_bgs list during a transaction abort. 166 */ 167 spin_lock(&transaction->fs_info->unused_bgs_lock); 168 list_del_init(&cache->bg_list); 169 spin_unlock(&transaction->fs_info->unused_bgs_lock); 170 btrfs_unfreeze_block_group(cache); 171 btrfs_put_block_group(cache); 172 } 173 WARN_ON(!list_empty(&transaction->dev_update_list)); 174 kfree(transaction); 175 } 176 } 177 178 static noinline void switch_commit_roots(struct btrfs_trans_handle *trans) 179 { 180 struct btrfs_transaction *cur_trans = trans->transaction; 181 struct btrfs_fs_info *fs_info = trans->fs_info; 182 struct btrfs_root *root, *tmp; 183 184 /* 185 * At this point no one can be using this transaction to modify any tree 186 * and no one can start another transaction to modify any tree either. 187 */ 188 ASSERT(cur_trans->state == TRANS_STATE_COMMIT_DOING); 189 190 down_write(&fs_info->commit_root_sem); 191 192 if (test_bit(BTRFS_FS_RELOC_RUNNING, &fs_info->flags)) 193 fs_info->last_reloc_trans = trans->transid; 194 195 list_for_each_entry_safe(root, tmp, &cur_trans->switch_commits, 196 dirty_list) { 197 list_del_init(&root->dirty_list); 198 free_extent_buffer(root->commit_root); 199 root->commit_root = btrfs_root_node(root); 200 extent_io_tree_release(&root->dirty_log_pages); 201 btrfs_qgroup_clean_swapped_blocks(root); 202 } 203 204 /* We can free old roots now. */ 205 spin_lock(&cur_trans->dropped_roots_lock); 206 while (!list_empty(&cur_trans->dropped_roots)) { 207 root = list_first_entry(&cur_trans->dropped_roots, 208 struct btrfs_root, root_list); 209 list_del_init(&root->root_list); 210 spin_unlock(&cur_trans->dropped_roots_lock); 211 btrfs_free_log(trans, root); 212 btrfs_drop_and_free_fs_root(fs_info, root); 213 spin_lock(&cur_trans->dropped_roots_lock); 214 } 215 spin_unlock(&cur_trans->dropped_roots_lock); 216 217 up_write(&fs_info->commit_root_sem); 218 } 219 220 static inline void extwriter_counter_inc(struct btrfs_transaction *trans, 221 unsigned int type) 222 { 223 if (type & TRANS_EXTWRITERS) 224 atomic_inc(&trans->num_extwriters); 225 } 226 227 static inline void extwriter_counter_dec(struct btrfs_transaction *trans, 228 unsigned int type) 229 { 230 if (type & TRANS_EXTWRITERS) 231 atomic_dec(&trans->num_extwriters); 232 } 233 234 static inline void extwriter_counter_init(struct btrfs_transaction *trans, 235 unsigned int type) 236 { 237 atomic_set(&trans->num_extwriters, ((type & TRANS_EXTWRITERS) ? 1 : 0)); 238 } 239 240 static inline int extwriter_counter_read(struct btrfs_transaction *trans) 241 { 242 return atomic_read(&trans->num_extwriters); 243 } 244 245 /* 246 * To be called after doing the chunk btree updates right after allocating a new 247 * chunk (after btrfs_chunk_alloc_add_chunk_item() is called), when removing a 248 * chunk after all chunk btree updates and after finishing the second phase of 249 * chunk allocation (btrfs_create_pending_block_groups()) in case some block 250 * group had its chunk item insertion delayed to the second phase. 251 */ 252 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans) 253 { 254 struct btrfs_fs_info *fs_info = trans->fs_info; 255 256 if (!trans->chunk_bytes_reserved) 257 return; 258 259 btrfs_block_rsv_release(fs_info, &fs_info->chunk_block_rsv, 260 trans->chunk_bytes_reserved, NULL); 261 trans->chunk_bytes_reserved = 0; 262 } 263 264 /* 265 * either allocate a new transaction or hop into the existing one 266 */ 267 static noinline int join_transaction(struct btrfs_fs_info *fs_info, 268 unsigned int type) 269 { 270 struct btrfs_transaction *cur_trans; 271 272 spin_lock(&fs_info->trans_lock); 273 loop: 274 /* The file system has been taken offline. No new transactions. */ 275 if (BTRFS_FS_ERROR(fs_info)) { 276 spin_unlock(&fs_info->trans_lock); 277 return -EROFS; 278 } 279 280 cur_trans = fs_info->running_transaction; 281 if (cur_trans) { 282 if (TRANS_ABORTED(cur_trans)) { 283 const int abort_error = cur_trans->aborted; 284 285 spin_unlock(&fs_info->trans_lock); 286 return abort_error; 287 } 288 if (btrfs_blocked_trans_types[cur_trans->state] & type) { 289 spin_unlock(&fs_info->trans_lock); 290 return -EBUSY; 291 } 292 refcount_inc(&cur_trans->use_count); 293 atomic_inc(&cur_trans->num_writers); 294 extwriter_counter_inc(cur_trans, type); 295 spin_unlock(&fs_info->trans_lock); 296 btrfs_lockdep_acquire(fs_info, btrfs_trans_num_writers); 297 btrfs_lockdep_acquire(fs_info, btrfs_trans_num_extwriters); 298 return 0; 299 } 300 spin_unlock(&fs_info->trans_lock); 301 302 /* 303 * If we are ATTACH or TRANS_JOIN_NOSTART, we just want to catch the 304 * current transaction, and commit it. If there is no transaction, just 305 * return ENOENT. 306 */ 307 if (type == TRANS_ATTACH || type == TRANS_JOIN_NOSTART) 308 return -ENOENT; 309 310 /* 311 * JOIN_NOLOCK only happens during the transaction commit, so 312 * it is impossible that ->running_transaction is NULL 313 */ 314 BUG_ON(type == TRANS_JOIN_NOLOCK); 315 316 cur_trans = kmalloc(sizeof(*cur_trans), GFP_NOFS); 317 if (!cur_trans) 318 return -ENOMEM; 319 320 btrfs_lockdep_acquire(fs_info, btrfs_trans_num_writers); 321 btrfs_lockdep_acquire(fs_info, btrfs_trans_num_extwriters); 322 323 spin_lock(&fs_info->trans_lock); 324 if (fs_info->running_transaction) { 325 /* 326 * someone started a transaction after we unlocked. Make sure 327 * to redo the checks above 328 */ 329 btrfs_lockdep_release(fs_info, btrfs_trans_num_extwriters); 330 btrfs_lockdep_release(fs_info, btrfs_trans_num_writers); 331 kfree(cur_trans); 332 goto loop; 333 } else if (BTRFS_FS_ERROR(fs_info)) { 334 spin_unlock(&fs_info->trans_lock); 335 btrfs_lockdep_release(fs_info, btrfs_trans_num_extwriters); 336 btrfs_lockdep_release(fs_info, btrfs_trans_num_writers); 337 kfree(cur_trans); 338 return -EROFS; 339 } 340 341 cur_trans->fs_info = fs_info; 342 atomic_set(&cur_trans->pending_ordered, 0); 343 init_waitqueue_head(&cur_trans->pending_wait); 344 atomic_set(&cur_trans->num_writers, 1); 345 extwriter_counter_init(cur_trans, type); 346 init_waitqueue_head(&cur_trans->writer_wait); 347 init_waitqueue_head(&cur_trans->commit_wait); 348 cur_trans->state = TRANS_STATE_RUNNING; 349 /* 350 * One for this trans handle, one so it will live on until we 351 * commit the transaction. 352 */ 353 refcount_set(&cur_trans->use_count, 2); 354 cur_trans->flags = 0; 355 cur_trans->start_time = ktime_get_seconds(); 356 357 memset(&cur_trans->delayed_refs, 0, sizeof(cur_trans->delayed_refs)); 358 359 xa_init(&cur_trans->delayed_refs.head_refs); 360 xa_init(&cur_trans->delayed_refs.dirty_extents); 361 362 /* 363 * although the tree mod log is per file system and not per transaction, 364 * the log must never go across transaction boundaries. 365 */ 366 smp_mb(); 367 if (!list_empty(&fs_info->tree_mod_seq_list)) 368 WARN(1, KERN_ERR "BTRFS: tree_mod_seq_list not empty when creating a fresh transaction\n"); 369 if (!RB_EMPTY_ROOT(&fs_info->tree_mod_log)) 370 WARN(1, KERN_ERR "BTRFS: tree_mod_log rb tree not empty when creating a fresh transaction\n"); 371 atomic64_set(&fs_info->tree_mod_seq, 0); 372 373 spin_lock_init(&cur_trans->delayed_refs.lock); 374 375 INIT_LIST_HEAD(&cur_trans->pending_snapshots); 376 INIT_LIST_HEAD(&cur_trans->dev_update_list); 377 INIT_LIST_HEAD(&cur_trans->switch_commits); 378 INIT_LIST_HEAD(&cur_trans->dirty_bgs); 379 INIT_LIST_HEAD(&cur_trans->io_bgs); 380 INIT_LIST_HEAD(&cur_trans->dropped_roots); 381 mutex_init(&cur_trans->cache_write_mutex); 382 spin_lock_init(&cur_trans->dirty_bgs_lock); 383 INIT_LIST_HEAD(&cur_trans->deleted_bgs); 384 spin_lock_init(&cur_trans->dropped_roots_lock); 385 list_add_tail(&cur_trans->list, &fs_info->trans_list); 386 extent_io_tree_init(fs_info, &cur_trans->dirty_pages, 387 IO_TREE_TRANS_DIRTY_PAGES); 388 extent_io_tree_init(fs_info, &cur_trans->pinned_extents, 389 IO_TREE_FS_PINNED_EXTENTS); 390 btrfs_set_fs_generation(fs_info, fs_info->generation + 1); 391 cur_trans->transid = fs_info->generation; 392 fs_info->running_transaction = cur_trans; 393 cur_trans->aborted = 0; 394 spin_unlock(&fs_info->trans_lock); 395 396 return 0; 397 } 398 399 /* 400 * This does all the record keeping required to make sure that a shareable root 401 * is properly recorded in a given transaction. This is required to make sure 402 * the old root from before we joined the transaction is deleted when the 403 * transaction commits. 404 */ 405 static int record_root_in_trans(struct btrfs_trans_handle *trans, 406 struct btrfs_root *root, 407 int force) 408 { 409 struct btrfs_fs_info *fs_info = root->fs_info; 410 int ret = 0; 411 412 if ((test_bit(BTRFS_ROOT_SHAREABLE, &root->state) && 413 btrfs_get_root_last_trans(root) < trans->transid) || force) { 414 WARN_ON(!force && root->commit_root != root->node); 415 416 /* 417 * see below for IN_TRANS_SETUP usage rules 418 * we have the reloc mutex held now, so there 419 * is only one writer in this function 420 */ 421 set_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state); 422 423 /* make sure readers find IN_TRANS_SETUP before 424 * they find our root->last_trans update 425 */ 426 smp_wmb(); 427 428 spin_lock(&fs_info->fs_roots_radix_lock); 429 if (btrfs_get_root_last_trans(root) == trans->transid && !force) { 430 spin_unlock(&fs_info->fs_roots_radix_lock); 431 return 0; 432 } 433 radix_tree_tag_set(&fs_info->fs_roots_radix, 434 (unsigned long)btrfs_root_id(root), 435 BTRFS_ROOT_TRANS_TAG); 436 spin_unlock(&fs_info->fs_roots_radix_lock); 437 btrfs_set_root_last_trans(root, trans->transid); 438 439 /* this is pretty tricky. We don't want to 440 * take the relocation lock in btrfs_record_root_in_trans 441 * unless we're really doing the first setup for this root in 442 * this transaction. 443 * 444 * Normally we'd use root->last_trans as a flag to decide 445 * if we want to take the expensive mutex. 446 * 447 * But, we have to set root->last_trans before we 448 * init the relocation root, otherwise, we trip over warnings 449 * in ctree.c. The solution used here is to flag ourselves 450 * with root IN_TRANS_SETUP. When this is 1, we're still 451 * fixing up the reloc trees and everyone must wait. 452 * 453 * When this is zero, they can trust root->last_trans and fly 454 * through btrfs_record_root_in_trans without having to take the 455 * lock. smp_wmb() makes sure that all the writes above are 456 * done before we pop in the zero below 457 */ 458 ret = btrfs_init_reloc_root(trans, root); 459 smp_mb__before_atomic(); 460 clear_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state); 461 } 462 return ret; 463 } 464 465 466 void btrfs_add_dropped_root(struct btrfs_trans_handle *trans, 467 struct btrfs_root *root) 468 { 469 struct btrfs_fs_info *fs_info = root->fs_info; 470 struct btrfs_transaction *cur_trans = trans->transaction; 471 472 /* Add ourselves to the transaction dropped list */ 473 spin_lock(&cur_trans->dropped_roots_lock); 474 list_add_tail(&root->root_list, &cur_trans->dropped_roots); 475 spin_unlock(&cur_trans->dropped_roots_lock); 476 477 /* Make sure we don't try to update the root at commit time */ 478 spin_lock(&fs_info->fs_roots_radix_lock); 479 radix_tree_tag_clear(&fs_info->fs_roots_radix, 480 (unsigned long)btrfs_root_id(root), 481 BTRFS_ROOT_TRANS_TAG); 482 spin_unlock(&fs_info->fs_roots_radix_lock); 483 } 484 485 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans, 486 struct btrfs_root *root) 487 { 488 struct btrfs_fs_info *fs_info = root->fs_info; 489 int ret; 490 491 if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) 492 return 0; 493 494 /* 495 * see record_root_in_trans for comments about IN_TRANS_SETUP usage 496 * and barriers 497 */ 498 smp_rmb(); 499 if (btrfs_get_root_last_trans(root) == trans->transid && 500 !test_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state)) 501 return 0; 502 503 mutex_lock(&fs_info->reloc_mutex); 504 ret = record_root_in_trans(trans, root, 0); 505 mutex_unlock(&fs_info->reloc_mutex); 506 507 return ret; 508 } 509 510 static inline int is_transaction_blocked(struct btrfs_transaction *trans) 511 { 512 return (trans->state >= TRANS_STATE_COMMIT_START && 513 trans->state < TRANS_STATE_UNBLOCKED && 514 !TRANS_ABORTED(trans)); 515 } 516 517 /* wait for commit against the current transaction to become unblocked 518 * when this is done, it is safe to start a new transaction, but the current 519 * transaction might not be fully on disk. 520 */ 521 static void wait_current_trans(struct btrfs_fs_info *fs_info) 522 { 523 struct btrfs_transaction *cur_trans; 524 525 spin_lock(&fs_info->trans_lock); 526 cur_trans = fs_info->running_transaction; 527 if (cur_trans && is_transaction_blocked(cur_trans)) { 528 refcount_inc(&cur_trans->use_count); 529 spin_unlock(&fs_info->trans_lock); 530 531 btrfs_might_wait_for_state(fs_info, BTRFS_LOCKDEP_TRANS_UNBLOCKED); 532 wait_event(fs_info->transaction_wait, 533 cur_trans->state >= TRANS_STATE_UNBLOCKED || 534 TRANS_ABORTED(cur_trans)); 535 btrfs_put_transaction(cur_trans); 536 } else { 537 spin_unlock(&fs_info->trans_lock); 538 } 539 } 540 541 static int may_wait_transaction(struct btrfs_fs_info *fs_info, int type) 542 { 543 if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags)) 544 return 0; 545 546 if (type == TRANS_START) 547 return 1; 548 549 return 0; 550 } 551 552 static inline bool need_reserve_reloc_root(struct btrfs_root *root) 553 { 554 struct btrfs_fs_info *fs_info = root->fs_info; 555 556 if (!fs_info->reloc_ctl || 557 !test_bit(BTRFS_ROOT_SHAREABLE, &root->state) || 558 btrfs_root_id(root) == BTRFS_TREE_RELOC_OBJECTID || 559 root->reloc_root) 560 return false; 561 562 return true; 563 } 564 565 static int btrfs_reserve_trans_metadata(struct btrfs_fs_info *fs_info, 566 enum btrfs_reserve_flush_enum flush, 567 u64 num_bytes, 568 u64 *delayed_refs_bytes) 569 { 570 struct btrfs_space_info *si = fs_info->trans_block_rsv.space_info; 571 u64 bytes = num_bytes + *delayed_refs_bytes; 572 int ret; 573 574 /* 575 * We want to reserve all the bytes we may need all at once, so we only 576 * do 1 enospc flushing cycle per transaction start. 577 */ 578 ret = btrfs_reserve_metadata_bytes(fs_info, si, bytes, flush); 579 580 /* 581 * If we are an emergency flush, which can steal from the global block 582 * reserve, then attempt to not reserve space for the delayed refs, as 583 * we will consume space for them from the global block reserve. 584 */ 585 if (ret && flush == BTRFS_RESERVE_FLUSH_ALL_STEAL) { 586 bytes -= *delayed_refs_bytes; 587 *delayed_refs_bytes = 0; 588 ret = btrfs_reserve_metadata_bytes(fs_info, si, bytes, flush); 589 } 590 591 return ret; 592 } 593 594 static struct btrfs_trans_handle * 595 start_transaction(struct btrfs_root *root, unsigned int num_items, 596 unsigned int type, enum btrfs_reserve_flush_enum flush, 597 bool enforce_qgroups) 598 { 599 struct btrfs_fs_info *fs_info = root->fs_info; 600 struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv; 601 struct btrfs_block_rsv *trans_rsv = &fs_info->trans_block_rsv; 602 struct btrfs_trans_handle *h; 603 struct btrfs_transaction *cur_trans; 604 u64 num_bytes = 0; 605 u64 qgroup_reserved = 0; 606 u64 delayed_refs_bytes = 0; 607 bool reloc_reserved = false; 608 bool do_chunk_alloc = false; 609 int ret; 610 611 if (BTRFS_FS_ERROR(fs_info)) 612 return ERR_PTR(-EROFS); 613 614 if (current->journal_info) { 615 WARN_ON(type & TRANS_EXTWRITERS); 616 h = current->journal_info; 617 refcount_inc(&h->use_count); 618 WARN_ON(refcount_read(&h->use_count) > 2); 619 h->orig_rsv = h->block_rsv; 620 h->block_rsv = NULL; 621 goto got_it; 622 } 623 624 /* 625 * Do the reservation before we join the transaction so we can do all 626 * the appropriate flushing if need be. 627 */ 628 if (num_items && root != fs_info->chunk_root) { 629 qgroup_reserved = num_items * fs_info->nodesize; 630 /* 631 * Use prealloc for now, as there might be a currently running 632 * transaction that could free this reserved space prematurely 633 * by committing. 634 */ 635 ret = btrfs_qgroup_reserve_meta_prealloc(root, qgroup_reserved, 636 enforce_qgroups, false); 637 if (ret) 638 return ERR_PTR(ret); 639 640 num_bytes = btrfs_calc_insert_metadata_size(fs_info, num_items); 641 /* 642 * If we plan to insert/update/delete "num_items" from a btree, 643 * we will also generate delayed refs for extent buffers in the 644 * respective btree paths, so reserve space for the delayed refs 645 * that will be generated by the caller as it modifies btrees. 646 * Try to reserve them to avoid excessive use of the global 647 * block reserve. 648 */ 649 delayed_refs_bytes = btrfs_calc_delayed_ref_bytes(fs_info, num_items); 650 651 /* 652 * Do the reservation for the relocation root creation 653 */ 654 if (need_reserve_reloc_root(root)) { 655 num_bytes += fs_info->nodesize; 656 reloc_reserved = true; 657 } 658 659 ret = btrfs_reserve_trans_metadata(fs_info, flush, num_bytes, 660 &delayed_refs_bytes); 661 if (ret) 662 goto reserve_fail; 663 664 btrfs_block_rsv_add_bytes(trans_rsv, num_bytes, true); 665 666 if (trans_rsv->space_info->force_alloc) 667 do_chunk_alloc = true; 668 } else if (num_items == 0 && flush == BTRFS_RESERVE_FLUSH_ALL && 669 !btrfs_block_rsv_full(delayed_refs_rsv)) { 670 /* 671 * Some people call with btrfs_start_transaction(root, 0) 672 * because they can be throttled, but have some other mechanism 673 * for reserving space. We still want these guys to refill the 674 * delayed block_rsv so just add 1 items worth of reservation 675 * here. 676 */ 677 ret = btrfs_delayed_refs_rsv_refill(fs_info, flush); 678 if (ret) 679 goto reserve_fail; 680 } 681 again: 682 h = kmem_cache_zalloc(btrfs_trans_handle_cachep, GFP_NOFS); 683 if (!h) { 684 ret = -ENOMEM; 685 goto alloc_fail; 686 } 687 688 /* 689 * If we are JOIN_NOLOCK we're already committing a transaction and 690 * waiting on this guy, so we don't need to do the sb_start_intwrite 691 * because we're already holding a ref. We need this because we could 692 * have raced in and did an fsync() on a file which can kick a commit 693 * and then we deadlock with somebody doing a freeze. 694 * 695 * If we are ATTACH, it means we just want to catch the current 696 * transaction and commit it, so we needn't do sb_start_intwrite(). 697 */ 698 if (type & __TRANS_FREEZABLE) 699 sb_start_intwrite(fs_info->sb); 700 701 if (may_wait_transaction(fs_info, type)) 702 wait_current_trans(fs_info); 703 704 do { 705 ret = join_transaction(fs_info, type); 706 if (ret == -EBUSY) { 707 wait_current_trans(fs_info); 708 if (unlikely(type == TRANS_ATTACH || 709 type == TRANS_JOIN_NOSTART)) 710 ret = -ENOENT; 711 } 712 } while (ret == -EBUSY); 713 714 if (ret < 0) 715 goto join_fail; 716 717 cur_trans = fs_info->running_transaction; 718 719 h->transid = cur_trans->transid; 720 h->transaction = cur_trans; 721 refcount_set(&h->use_count, 1); 722 h->fs_info = root->fs_info; 723 724 h->type = type; 725 INIT_LIST_HEAD(&h->new_bgs); 726 btrfs_init_metadata_block_rsv(fs_info, &h->delayed_rsv, BTRFS_BLOCK_RSV_DELOPS); 727 728 smp_mb(); 729 if (cur_trans->state >= TRANS_STATE_COMMIT_START && 730 may_wait_transaction(fs_info, type)) { 731 current->journal_info = h; 732 btrfs_commit_transaction(h); 733 goto again; 734 } 735 736 if (num_bytes) { 737 trace_btrfs_space_reservation(fs_info, "transaction", 738 h->transid, num_bytes, 1); 739 h->block_rsv = trans_rsv; 740 h->bytes_reserved = num_bytes; 741 if (delayed_refs_bytes > 0) { 742 trace_btrfs_space_reservation(fs_info, 743 "local_delayed_refs_rsv", 744 h->transid, 745 delayed_refs_bytes, 1); 746 h->delayed_refs_bytes_reserved = delayed_refs_bytes; 747 btrfs_block_rsv_add_bytes(&h->delayed_rsv, delayed_refs_bytes, true); 748 delayed_refs_bytes = 0; 749 } 750 h->reloc_reserved = reloc_reserved; 751 } 752 753 got_it: 754 if (!current->journal_info) 755 current->journal_info = h; 756 757 /* 758 * If the space_info is marked ALLOC_FORCE then we'll get upgraded to 759 * ALLOC_FORCE the first run through, and then we won't allocate for 760 * anybody else who races in later. We don't care about the return 761 * value here. 762 */ 763 if (do_chunk_alloc && num_bytes) { 764 u64 flags = h->block_rsv->space_info->flags; 765 766 btrfs_chunk_alloc(h, btrfs_get_alloc_profile(fs_info, flags), 767 CHUNK_ALLOC_NO_FORCE); 768 } 769 770 /* 771 * btrfs_record_root_in_trans() needs to alloc new extents, and may 772 * call btrfs_join_transaction() while we're also starting a 773 * transaction. 774 * 775 * Thus it need to be called after current->journal_info initialized, 776 * or we can deadlock. 777 */ 778 ret = btrfs_record_root_in_trans(h, root); 779 if (ret) { 780 /* 781 * The transaction handle is fully initialized and linked with 782 * other structures so it needs to be ended in case of errors, 783 * not just freed. 784 */ 785 btrfs_end_transaction(h); 786 goto reserve_fail; 787 } 788 /* 789 * Now that we have found a transaction to be a part of, convert the 790 * qgroup reservation from prealloc to pertrans. A different transaction 791 * can't race in and free our pertrans out from under us. 792 */ 793 if (qgroup_reserved) 794 btrfs_qgroup_convert_reserved_meta(root, qgroup_reserved); 795 796 return h; 797 798 join_fail: 799 if (type & __TRANS_FREEZABLE) 800 sb_end_intwrite(fs_info->sb); 801 kmem_cache_free(btrfs_trans_handle_cachep, h); 802 alloc_fail: 803 if (num_bytes) 804 btrfs_block_rsv_release(fs_info, trans_rsv, num_bytes, NULL); 805 if (delayed_refs_bytes) 806 btrfs_space_info_free_bytes_may_use(trans_rsv->space_info, delayed_refs_bytes); 807 reserve_fail: 808 btrfs_qgroup_free_meta_prealloc(root, qgroup_reserved); 809 return ERR_PTR(ret); 810 } 811 812 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root, 813 unsigned int num_items) 814 { 815 return start_transaction(root, num_items, TRANS_START, 816 BTRFS_RESERVE_FLUSH_ALL, true); 817 } 818 819 struct btrfs_trans_handle *btrfs_start_transaction_fallback_global_rsv( 820 struct btrfs_root *root, 821 unsigned int num_items) 822 { 823 return start_transaction(root, num_items, TRANS_START, 824 BTRFS_RESERVE_FLUSH_ALL_STEAL, false); 825 } 826 827 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root) 828 { 829 return start_transaction(root, 0, TRANS_JOIN, BTRFS_RESERVE_NO_FLUSH, 830 true); 831 } 832 833 struct btrfs_trans_handle *btrfs_join_transaction_spacecache(struct btrfs_root *root) 834 { 835 return start_transaction(root, 0, TRANS_JOIN_NOLOCK, 836 BTRFS_RESERVE_NO_FLUSH, true); 837 } 838 839 /* 840 * Similar to regular join but it never starts a transaction when none is 841 * running or when there's a running one at a state >= TRANS_STATE_UNBLOCKED. 842 * This is similar to btrfs_attach_transaction() but it allows the join to 843 * happen if the transaction commit already started but it's not yet in the 844 * "doing" phase (the state is < TRANS_STATE_COMMIT_DOING). 845 */ 846 struct btrfs_trans_handle *btrfs_join_transaction_nostart(struct btrfs_root *root) 847 { 848 return start_transaction(root, 0, TRANS_JOIN_NOSTART, 849 BTRFS_RESERVE_NO_FLUSH, true); 850 } 851 852 /* 853 * Catch the running transaction. 854 * 855 * It is used when we want to commit the current the transaction, but 856 * don't want to start a new one. 857 * 858 * Note: If this function return -ENOENT, it just means there is no 859 * running transaction. But it is possible that the inactive transaction 860 * is still in the memory, not fully on disk. If you hope there is no 861 * inactive transaction in the fs when -ENOENT is returned, you should 862 * invoke 863 * btrfs_attach_transaction_barrier() 864 */ 865 struct btrfs_trans_handle *btrfs_attach_transaction(struct btrfs_root *root) 866 { 867 return start_transaction(root, 0, TRANS_ATTACH, 868 BTRFS_RESERVE_NO_FLUSH, true); 869 } 870 871 /* 872 * Catch the running transaction. 873 * 874 * It is similar to the above function, the difference is this one 875 * will wait for all the inactive transactions until they fully 876 * complete. 877 */ 878 struct btrfs_trans_handle * 879 btrfs_attach_transaction_barrier(struct btrfs_root *root) 880 { 881 struct btrfs_trans_handle *trans; 882 883 trans = start_transaction(root, 0, TRANS_ATTACH, 884 BTRFS_RESERVE_NO_FLUSH, true); 885 if (trans == ERR_PTR(-ENOENT)) { 886 int ret; 887 888 ret = btrfs_wait_for_commit(root->fs_info, 0); 889 if (ret) 890 return ERR_PTR(ret); 891 } 892 893 return trans; 894 } 895 896 /* Wait for a transaction commit to reach at least the given state. */ 897 static noinline void wait_for_commit(struct btrfs_transaction *commit, 898 const enum btrfs_trans_state min_state) 899 { 900 struct btrfs_fs_info *fs_info = commit->fs_info; 901 u64 transid = commit->transid; 902 bool put = false; 903 904 /* 905 * At the moment this function is called with min_state either being 906 * TRANS_STATE_COMPLETED or TRANS_STATE_SUPER_COMMITTED. 907 */ 908 if (min_state == TRANS_STATE_COMPLETED) 909 btrfs_might_wait_for_state(fs_info, BTRFS_LOCKDEP_TRANS_COMPLETED); 910 else 911 btrfs_might_wait_for_state(fs_info, BTRFS_LOCKDEP_TRANS_SUPER_COMMITTED); 912 913 while (1) { 914 wait_event(commit->commit_wait, commit->state >= min_state); 915 if (put) 916 btrfs_put_transaction(commit); 917 918 if (min_state < TRANS_STATE_COMPLETED) 919 break; 920 921 /* 922 * A transaction isn't really completed until all of the 923 * previous transactions are completed, but with fsync we can 924 * end up with SUPER_COMMITTED transactions before a COMPLETED 925 * transaction. Wait for those. 926 */ 927 928 spin_lock(&fs_info->trans_lock); 929 commit = list_first_entry_or_null(&fs_info->trans_list, 930 struct btrfs_transaction, 931 list); 932 if (!commit || commit->transid > transid) { 933 spin_unlock(&fs_info->trans_lock); 934 break; 935 } 936 refcount_inc(&commit->use_count); 937 put = true; 938 spin_unlock(&fs_info->trans_lock); 939 } 940 } 941 942 int btrfs_wait_for_commit(struct btrfs_fs_info *fs_info, u64 transid) 943 { 944 struct btrfs_transaction *cur_trans = NULL, *t; 945 int ret = 0; 946 947 if (transid) { 948 if (transid <= btrfs_get_last_trans_committed(fs_info)) 949 goto out; 950 951 /* find specified transaction */ 952 spin_lock(&fs_info->trans_lock); 953 list_for_each_entry(t, &fs_info->trans_list, list) { 954 if (t->transid == transid) { 955 cur_trans = t; 956 refcount_inc(&cur_trans->use_count); 957 ret = 0; 958 break; 959 } 960 if (t->transid > transid) { 961 ret = 0; 962 break; 963 } 964 } 965 spin_unlock(&fs_info->trans_lock); 966 967 /* 968 * The specified transaction doesn't exist, or we 969 * raced with btrfs_commit_transaction 970 */ 971 if (!cur_trans) { 972 if (transid > btrfs_get_last_trans_committed(fs_info)) 973 ret = -EINVAL; 974 goto out; 975 } 976 } else { 977 /* find newest transaction that is committing | committed */ 978 spin_lock(&fs_info->trans_lock); 979 list_for_each_entry_reverse(t, &fs_info->trans_list, 980 list) { 981 if (t->state >= TRANS_STATE_COMMIT_START) { 982 if (t->state == TRANS_STATE_COMPLETED) 983 break; 984 cur_trans = t; 985 refcount_inc(&cur_trans->use_count); 986 break; 987 } 988 } 989 spin_unlock(&fs_info->trans_lock); 990 if (!cur_trans) 991 goto out; /* nothing committing|committed */ 992 } 993 994 wait_for_commit(cur_trans, TRANS_STATE_COMPLETED); 995 ret = cur_trans->aborted; 996 btrfs_put_transaction(cur_trans); 997 out: 998 return ret; 999 } 1000 1001 void btrfs_throttle(struct btrfs_fs_info *fs_info) 1002 { 1003 wait_current_trans(fs_info); 1004 } 1005 1006 bool btrfs_should_end_transaction(struct btrfs_trans_handle *trans) 1007 { 1008 struct btrfs_transaction *cur_trans = trans->transaction; 1009 1010 if (cur_trans->state >= TRANS_STATE_COMMIT_START || 1011 test_bit(BTRFS_DELAYED_REFS_FLUSHING, &cur_trans->delayed_refs.flags)) 1012 return true; 1013 1014 if (btrfs_check_space_for_delayed_refs(trans->fs_info)) 1015 return true; 1016 1017 return !!btrfs_block_rsv_check(&trans->fs_info->global_block_rsv, 50); 1018 } 1019 1020 static void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans) 1021 1022 { 1023 struct btrfs_fs_info *fs_info = trans->fs_info; 1024 1025 if (!trans->block_rsv) { 1026 ASSERT(!trans->bytes_reserved); 1027 ASSERT(!trans->delayed_refs_bytes_reserved); 1028 return; 1029 } 1030 1031 if (!trans->bytes_reserved) { 1032 ASSERT(!trans->delayed_refs_bytes_reserved); 1033 return; 1034 } 1035 1036 ASSERT(trans->block_rsv == &fs_info->trans_block_rsv); 1037 trace_btrfs_space_reservation(fs_info, "transaction", 1038 trans->transid, trans->bytes_reserved, 0); 1039 btrfs_block_rsv_release(fs_info, trans->block_rsv, 1040 trans->bytes_reserved, NULL); 1041 trans->bytes_reserved = 0; 1042 1043 if (!trans->delayed_refs_bytes_reserved) 1044 return; 1045 1046 trace_btrfs_space_reservation(fs_info, "local_delayed_refs_rsv", 1047 trans->transid, 1048 trans->delayed_refs_bytes_reserved, 0); 1049 btrfs_block_rsv_release(fs_info, &trans->delayed_rsv, 1050 trans->delayed_refs_bytes_reserved, NULL); 1051 trans->delayed_refs_bytes_reserved = 0; 1052 } 1053 1054 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans, 1055 int throttle) 1056 { 1057 struct btrfs_fs_info *info = trans->fs_info; 1058 struct btrfs_transaction *cur_trans = trans->transaction; 1059 int ret = 0; 1060 1061 if (refcount_read(&trans->use_count) > 1) { 1062 refcount_dec(&trans->use_count); 1063 trans->block_rsv = trans->orig_rsv; 1064 return 0; 1065 } 1066 1067 btrfs_trans_release_metadata(trans); 1068 trans->block_rsv = NULL; 1069 1070 btrfs_create_pending_block_groups(trans); 1071 1072 btrfs_trans_release_chunk_metadata(trans); 1073 1074 if (trans->type & __TRANS_FREEZABLE) 1075 sb_end_intwrite(info->sb); 1076 1077 WARN_ON(cur_trans != info->running_transaction); 1078 WARN_ON(atomic_read(&cur_trans->num_writers) < 1); 1079 atomic_dec(&cur_trans->num_writers); 1080 extwriter_counter_dec(cur_trans, trans->type); 1081 1082 cond_wake_up(&cur_trans->writer_wait); 1083 1084 btrfs_lockdep_release(info, btrfs_trans_num_extwriters); 1085 btrfs_lockdep_release(info, btrfs_trans_num_writers); 1086 1087 btrfs_put_transaction(cur_trans); 1088 1089 if (current->journal_info == trans) 1090 current->journal_info = NULL; 1091 1092 if (throttle) 1093 btrfs_run_delayed_iputs(info); 1094 1095 if (TRANS_ABORTED(trans) || BTRFS_FS_ERROR(info)) { 1096 wake_up_process(info->transaction_kthread); 1097 if (TRANS_ABORTED(trans)) 1098 ret = trans->aborted; 1099 else 1100 ret = -EROFS; 1101 } 1102 1103 kmem_cache_free(btrfs_trans_handle_cachep, trans); 1104 return ret; 1105 } 1106 1107 int btrfs_end_transaction(struct btrfs_trans_handle *trans) 1108 { 1109 return __btrfs_end_transaction(trans, 0); 1110 } 1111 1112 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans) 1113 { 1114 return __btrfs_end_transaction(trans, 1); 1115 } 1116 1117 /* 1118 * when btree blocks are allocated, they have some corresponding bits set for 1119 * them in one of two extent_io trees. This is used to make sure all of 1120 * those extents are sent to disk but does not wait on them 1121 */ 1122 int btrfs_write_marked_extents(struct btrfs_fs_info *fs_info, 1123 struct extent_io_tree *dirty_pages, int mark) 1124 { 1125 int ret = 0; 1126 struct address_space *mapping = fs_info->btree_inode->i_mapping; 1127 struct extent_state *cached_state = NULL; 1128 u64 start = 0; 1129 u64 end; 1130 1131 while (find_first_extent_bit(dirty_pages, start, &start, &end, 1132 mark, &cached_state)) { 1133 bool wait_writeback = false; 1134 1135 ret = convert_extent_bit(dirty_pages, start, end, 1136 EXTENT_NEED_WAIT, 1137 mark, &cached_state); 1138 /* 1139 * convert_extent_bit can return -ENOMEM, which is most of the 1140 * time a temporary error. So when it happens, ignore the error 1141 * and wait for writeback of this range to finish - because we 1142 * failed to set the bit EXTENT_NEED_WAIT for the range, a call 1143 * to __btrfs_wait_marked_extents() would not know that 1144 * writeback for this range started and therefore wouldn't 1145 * wait for it to finish - we don't want to commit a 1146 * superblock that points to btree nodes/leafs for which 1147 * writeback hasn't finished yet (and without errors). 1148 * We cleanup any entries left in the io tree when committing 1149 * the transaction (through extent_io_tree_release()). 1150 */ 1151 if (ret == -ENOMEM) { 1152 ret = 0; 1153 wait_writeback = true; 1154 } 1155 if (!ret) 1156 ret = filemap_fdatawrite_range(mapping, start, end); 1157 if (!ret && wait_writeback) 1158 ret = filemap_fdatawait_range(mapping, start, end); 1159 free_extent_state(cached_state); 1160 if (ret) 1161 break; 1162 cached_state = NULL; 1163 cond_resched(); 1164 start = end + 1; 1165 } 1166 return ret; 1167 } 1168 1169 /* 1170 * when btree blocks are allocated, they have some corresponding bits set for 1171 * them in one of two extent_io trees. This is used to make sure all of 1172 * those extents are on disk for transaction or log commit. We wait 1173 * on all the pages and clear them from the dirty pages state tree 1174 */ 1175 static int __btrfs_wait_marked_extents(struct btrfs_fs_info *fs_info, 1176 struct extent_io_tree *dirty_pages) 1177 { 1178 struct address_space *mapping = fs_info->btree_inode->i_mapping; 1179 struct extent_state *cached_state = NULL; 1180 u64 start = 0; 1181 u64 end; 1182 int ret = 0; 1183 1184 while (find_first_extent_bit(dirty_pages, start, &start, &end, 1185 EXTENT_NEED_WAIT, &cached_state)) { 1186 /* 1187 * Ignore -ENOMEM errors returned by clear_extent_bit(). 1188 * When committing the transaction, we'll remove any entries 1189 * left in the io tree. For a log commit, we don't remove them 1190 * after committing the log because the tree can be accessed 1191 * concurrently - we do it only at transaction commit time when 1192 * it's safe to do it (through extent_io_tree_release()). 1193 */ 1194 ret = clear_extent_bit(dirty_pages, start, end, 1195 EXTENT_NEED_WAIT, &cached_state); 1196 if (ret == -ENOMEM) 1197 ret = 0; 1198 if (!ret) 1199 ret = filemap_fdatawait_range(mapping, start, end); 1200 free_extent_state(cached_state); 1201 if (ret) 1202 break; 1203 cached_state = NULL; 1204 cond_resched(); 1205 start = end + 1; 1206 } 1207 return ret; 1208 } 1209 1210 static int btrfs_wait_extents(struct btrfs_fs_info *fs_info, 1211 struct extent_io_tree *dirty_pages) 1212 { 1213 bool errors = false; 1214 int err; 1215 1216 err = __btrfs_wait_marked_extents(fs_info, dirty_pages); 1217 if (test_and_clear_bit(BTRFS_FS_BTREE_ERR, &fs_info->flags)) 1218 errors = true; 1219 1220 if (errors && !err) 1221 err = -EIO; 1222 return err; 1223 } 1224 1225 int btrfs_wait_tree_log_extents(struct btrfs_root *log_root, int mark) 1226 { 1227 struct btrfs_fs_info *fs_info = log_root->fs_info; 1228 struct extent_io_tree *dirty_pages = &log_root->dirty_log_pages; 1229 bool errors = false; 1230 int err; 1231 1232 ASSERT(btrfs_root_id(log_root) == BTRFS_TREE_LOG_OBJECTID); 1233 1234 err = __btrfs_wait_marked_extents(fs_info, dirty_pages); 1235 if ((mark & EXTENT_DIRTY) && 1236 test_and_clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags)) 1237 errors = true; 1238 1239 if ((mark & EXTENT_NEW) && 1240 test_and_clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags)) 1241 errors = true; 1242 1243 if (errors && !err) 1244 err = -EIO; 1245 return err; 1246 } 1247 1248 /* 1249 * When btree blocks are allocated the corresponding extents are marked dirty. 1250 * This function ensures such extents are persisted on disk for transaction or 1251 * log commit. 1252 * 1253 * @trans: transaction whose dirty pages we'd like to write 1254 */ 1255 static int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans) 1256 { 1257 int ret; 1258 int ret2; 1259 struct extent_io_tree *dirty_pages = &trans->transaction->dirty_pages; 1260 struct btrfs_fs_info *fs_info = trans->fs_info; 1261 struct blk_plug plug; 1262 1263 blk_start_plug(&plug); 1264 ret = btrfs_write_marked_extents(fs_info, dirty_pages, EXTENT_DIRTY); 1265 blk_finish_plug(&plug); 1266 ret2 = btrfs_wait_extents(fs_info, dirty_pages); 1267 1268 extent_io_tree_release(&trans->transaction->dirty_pages); 1269 1270 if (ret) 1271 return ret; 1272 else if (ret2) 1273 return ret2; 1274 else 1275 return 0; 1276 } 1277 1278 /* 1279 * this is used to update the root pointer in the tree of tree roots. 1280 * 1281 * But, in the case of the extent allocation tree, updating the root 1282 * pointer may allocate blocks which may change the root of the extent 1283 * allocation tree. 1284 * 1285 * So, this loops and repeats and makes sure the cowonly root didn't 1286 * change while the root pointer was being updated in the metadata. 1287 */ 1288 static int update_cowonly_root(struct btrfs_trans_handle *trans, 1289 struct btrfs_root *root) 1290 { 1291 int ret; 1292 u64 old_root_bytenr; 1293 u64 old_root_used; 1294 struct btrfs_fs_info *fs_info = root->fs_info; 1295 struct btrfs_root *tree_root = fs_info->tree_root; 1296 1297 old_root_used = btrfs_root_used(&root->root_item); 1298 1299 while (1) { 1300 old_root_bytenr = btrfs_root_bytenr(&root->root_item); 1301 if (old_root_bytenr == root->node->start && 1302 old_root_used == btrfs_root_used(&root->root_item)) 1303 break; 1304 1305 btrfs_set_root_node(&root->root_item, root->node); 1306 ret = btrfs_update_root(trans, tree_root, 1307 &root->root_key, 1308 &root->root_item); 1309 if (ret) 1310 return ret; 1311 1312 old_root_used = btrfs_root_used(&root->root_item); 1313 } 1314 1315 return 0; 1316 } 1317 1318 /* 1319 * update all the cowonly tree roots on disk 1320 * 1321 * The error handling in this function may not be obvious. Any of the 1322 * failures will cause the file system to go offline. We still need 1323 * to clean up the delayed refs. 1324 */ 1325 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans) 1326 { 1327 struct btrfs_fs_info *fs_info = trans->fs_info; 1328 struct list_head *dirty_bgs = &trans->transaction->dirty_bgs; 1329 struct list_head *io_bgs = &trans->transaction->io_bgs; 1330 struct list_head *next; 1331 struct extent_buffer *eb; 1332 int ret; 1333 1334 /* 1335 * At this point no one can be using this transaction to modify any tree 1336 * and no one can start another transaction to modify any tree either. 1337 */ 1338 ASSERT(trans->transaction->state == TRANS_STATE_COMMIT_DOING); 1339 1340 eb = btrfs_lock_root_node(fs_info->tree_root); 1341 ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL, 1342 0, &eb, BTRFS_NESTING_COW); 1343 btrfs_tree_unlock(eb); 1344 free_extent_buffer(eb); 1345 1346 if (ret) 1347 return ret; 1348 1349 ret = btrfs_run_dev_stats(trans); 1350 if (ret) 1351 return ret; 1352 ret = btrfs_run_dev_replace(trans); 1353 if (ret) 1354 return ret; 1355 ret = btrfs_run_qgroups(trans); 1356 if (ret) 1357 return ret; 1358 1359 ret = btrfs_setup_space_cache(trans); 1360 if (ret) 1361 return ret; 1362 1363 again: 1364 while (!list_empty(&fs_info->dirty_cowonly_roots)) { 1365 struct btrfs_root *root; 1366 next = fs_info->dirty_cowonly_roots.next; 1367 list_del_init(next); 1368 root = list_entry(next, struct btrfs_root, dirty_list); 1369 clear_bit(BTRFS_ROOT_DIRTY, &root->state); 1370 1371 list_add_tail(&root->dirty_list, 1372 &trans->transaction->switch_commits); 1373 ret = update_cowonly_root(trans, root); 1374 if (ret) 1375 return ret; 1376 } 1377 1378 /* Now flush any delayed refs generated by updating all of the roots */ 1379 ret = btrfs_run_delayed_refs(trans, U64_MAX); 1380 if (ret) 1381 return ret; 1382 1383 while (!list_empty(dirty_bgs) || !list_empty(io_bgs)) { 1384 ret = btrfs_write_dirty_block_groups(trans); 1385 if (ret) 1386 return ret; 1387 1388 /* 1389 * We're writing the dirty block groups, which could generate 1390 * delayed refs, which could generate more dirty block groups, 1391 * so we want to keep this flushing in this loop to make sure 1392 * everything gets run. 1393 */ 1394 ret = btrfs_run_delayed_refs(trans, U64_MAX); 1395 if (ret) 1396 return ret; 1397 } 1398 1399 if (!list_empty(&fs_info->dirty_cowonly_roots)) 1400 goto again; 1401 1402 /* Update dev-replace pointer once everything is committed */ 1403 fs_info->dev_replace.committed_cursor_left = 1404 fs_info->dev_replace.cursor_left_last_write_of_item; 1405 1406 return 0; 1407 } 1408 1409 /* 1410 * If we had a pending drop we need to see if there are any others left in our 1411 * dead roots list, and if not clear our bit and wake any waiters. 1412 */ 1413 void btrfs_maybe_wake_unfinished_drop(struct btrfs_fs_info *fs_info) 1414 { 1415 /* 1416 * We put the drop in progress roots at the front of the list, so if the 1417 * first entry doesn't have UNFINISHED_DROP set we can wake everybody 1418 * up. 1419 */ 1420 spin_lock(&fs_info->trans_lock); 1421 if (!list_empty(&fs_info->dead_roots)) { 1422 struct btrfs_root *root = list_first_entry(&fs_info->dead_roots, 1423 struct btrfs_root, 1424 root_list); 1425 if (test_bit(BTRFS_ROOT_UNFINISHED_DROP, &root->state)) { 1426 spin_unlock(&fs_info->trans_lock); 1427 return; 1428 } 1429 } 1430 spin_unlock(&fs_info->trans_lock); 1431 1432 btrfs_wake_unfinished_drop(fs_info); 1433 } 1434 1435 /* 1436 * dead roots are old snapshots that need to be deleted. This allocates 1437 * a dirty root struct and adds it into the list of dead roots that need to 1438 * be deleted 1439 */ 1440 void btrfs_add_dead_root(struct btrfs_root *root) 1441 { 1442 struct btrfs_fs_info *fs_info = root->fs_info; 1443 1444 spin_lock(&fs_info->trans_lock); 1445 if (list_empty(&root->root_list)) { 1446 btrfs_grab_root(root); 1447 1448 /* We want to process the partially complete drops first. */ 1449 if (test_bit(BTRFS_ROOT_UNFINISHED_DROP, &root->state)) 1450 list_add(&root->root_list, &fs_info->dead_roots); 1451 else 1452 list_add_tail(&root->root_list, &fs_info->dead_roots); 1453 } 1454 spin_unlock(&fs_info->trans_lock); 1455 } 1456 1457 /* 1458 * Update each subvolume root and its relocation root, if it exists, in the tree 1459 * of tree roots. Also free log roots if they exist. 1460 */ 1461 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans) 1462 { 1463 struct btrfs_fs_info *fs_info = trans->fs_info; 1464 struct btrfs_root *gang[8]; 1465 int i; 1466 int ret; 1467 1468 /* 1469 * At this point no one can be using this transaction to modify any tree 1470 * and no one can start another transaction to modify any tree either. 1471 */ 1472 ASSERT(trans->transaction->state == TRANS_STATE_COMMIT_DOING); 1473 1474 spin_lock(&fs_info->fs_roots_radix_lock); 1475 while (1) { 1476 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix, 1477 (void **)gang, 0, 1478 ARRAY_SIZE(gang), 1479 BTRFS_ROOT_TRANS_TAG); 1480 if (ret == 0) 1481 break; 1482 for (i = 0; i < ret; i++) { 1483 struct btrfs_root *root = gang[i]; 1484 int ret2; 1485 1486 /* 1487 * At this point we can neither have tasks logging inodes 1488 * from a root nor trying to commit a log tree. 1489 */ 1490 ASSERT(atomic_read(&root->log_writers) == 0); 1491 ASSERT(atomic_read(&root->log_commit[0]) == 0); 1492 ASSERT(atomic_read(&root->log_commit[1]) == 0); 1493 1494 radix_tree_tag_clear(&fs_info->fs_roots_radix, 1495 (unsigned long)btrfs_root_id(root), 1496 BTRFS_ROOT_TRANS_TAG); 1497 btrfs_qgroup_free_meta_all_pertrans(root); 1498 spin_unlock(&fs_info->fs_roots_radix_lock); 1499 1500 btrfs_free_log(trans, root); 1501 ret2 = btrfs_update_reloc_root(trans, root); 1502 if (ret2) 1503 return ret2; 1504 1505 /* see comments in should_cow_block() */ 1506 clear_bit(BTRFS_ROOT_FORCE_COW, &root->state); 1507 smp_mb__after_atomic(); 1508 1509 if (root->commit_root != root->node) { 1510 list_add_tail(&root->dirty_list, 1511 &trans->transaction->switch_commits); 1512 btrfs_set_root_node(&root->root_item, 1513 root->node); 1514 } 1515 1516 ret2 = btrfs_update_root(trans, fs_info->tree_root, 1517 &root->root_key, 1518 &root->root_item); 1519 if (ret2) 1520 return ret2; 1521 spin_lock(&fs_info->fs_roots_radix_lock); 1522 } 1523 } 1524 spin_unlock(&fs_info->fs_roots_radix_lock); 1525 return 0; 1526 } 1527 1528 /* 1529 * Do all special snapshot related qgroup dirty hack. 1530 * 1531 * Will do all needed qgroup inherit and dirty hack like switch commit 1532 * roots inside one transaction and write all btree into disk, to make 1533 * qgroup works. 1534 */ 1535 static int qgroup_account_snapshot(struct btrfs_trans_handle *trans, 1536 struct btrfs_root *src, 1537 struct btrfs_root *parent, 1538 struct btrfs_qgroup_inherit *inherit, 1539 u64 dst_objectid) 1540 { 1541 struct btrfs_fs_info *fs_info = src->fs_info; 1542 int ret; 1543 1544 /* 1545 * Save some performance in the case that qgroups are not enabled. If 1546 * this check races with the ioctl, rescan will kick in anyway. 1547 */ 1548 if (!btrfs_qgroup_full_accounting(fs_info)) 1549 return 0; 1550 1551 /* 1552 * Ensure dirty @src will be committed. Or, after coming 1553 * commit_fs_roots() and switch_commit_roots(), any dirty but not 1554 * recorded root will never be updated again, causing an outdated root 1555 * item. 1556 */ 1557 ret = record_root_in_trans(trans, src, 1); 1558 if (ret) 1559 return ret; 1560 1561 /* 1562 * btrfs_qgroup_inherit relies on a consistent view of the usage for the 1563 * src root, so we must run the delayed refs here. 1564 * 1565 * However this isn't particularly fool proof, because there's no 1566 * synchronization keeping us from changing the tree after this point 1567 * before we do the qgroup_inherit, or even from making changes while 1568 * we're doing the qgroup_inherit. But that's a problem for the future, 1569 * for now flush the delayed refs to narrow the race window where the 1570 * qgroup counters could end up wrong. 1571 */ 1572 ret = btrfs_run_delayed_refs(trans, U64_MAX); 1573 if (ret) { 1574 btrfs_abort_transaction(trans, ret); 1575 return ret; 1576 } 1577 1578 ret = commit_fs_roots(trans); 1579 if (ret) 1580 goto out; 1581 ret = btrfs_qgroup_account_extents(trans); 1582 if (ret < 0) 1583 goto out; 1584 1585 /* Now qgroup are all updated, we can inherit it to new qgroups */ 1586 ret = btrfs_qgroup_inherit(trans, btrfs_root_id(src), dst_objectid, 1587 btrfs_root_id(parent), inherit); 1588 if (ret < 0) 1589 goto out; 1590 1591 /* 1592 * Now we do a simplified commit transaction, which will: 1593 * 1) commit all subvolume and extent tree 1594 * To ensure all subvolume and extent tree have a valid 1595 * commit_root to accounting later insert_dir_item() 1596 * 2) write all btree blocks onto disk 1597 * This is to make sure later btree modification will be cowed 1598 * Or commit_root can be populated and cause wrong qgroup numbers 1599 * In this simplified commit, we don't really care about other trees 1600 * like chunk and root tree, as they won't affect qgroup. 1601 * And we don't write super to avoid half committed status. 1602 */ 1603 ret = commit_cowonly_roots(trans); 1604 if (ret) 1605 goto out; 1606 switch_commit_roots(trans); 1607 ret = btrfs_write_and_wait_transaction(trans); 1608 if (ret) 1609 btrfs_handle_fs_error(fs_info, ret, 1610 "Error while writing out transaction for qgroup"); 1611 1612 out: 1613 /* 1614 * Force parent root to be updated, as we recorded it before so its 1615 * last_trans == cur_transid. 1616 * Or it won't be committed again onto disk after later 1617 * insert_dir_item() 1618 */ 1619 if (!ret) 1620 ret = record_root_in_trans(trans, parent, 1); 1621 return ret; 1622 } 1623 1624 /* 1625 * new snapshots need to be created at a very specific time in the 1626 * transaction commit. This does the actual creation. 1627 * 1628 * Note: 1629 * If the error which may affect the commitment of the current transaction 1630 * happens, we should return the error number. If the error which just affect 1631 * the creation of the pending snapshots, just return 0. 1632 */ 1633 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans, 1634 struct btrfs_pending_snapshot *pending) 1635 { 1636 1637 struct btrfs_fs_info *fs_info = trans->fs_info; 1638 struct btrfs_key key; 1639 struct btrfs_root_item *new_root_item; 1640 struct btrfs_root *tree_root = fs_info->tree_root; 1641 struct btrfs_root *root = pending->root; 1642 struct btrfs_root *parent_root; 1643 struct btrfs_block_rsv *rsv; 1644 struct btrfs_inode *parent_inode = pending->dir; 1645 struct btrfs_path *path; 1646 struct btrfs_dir_item *dir_item; 1647 struct extent_buffer *tmp; 1648 struct extent_buffer *old; 1649 struct timespec64 cur_time; 1650 int ret = 0; 1651 u64 to_reserve = 0; 1652 u64 index = 0; 1653 u64 objectid; 1654 u64 root_flags; 1655 unsigned int nofs_flags; 1656 struct fscrypt_name fname; 1657 1658 ASSERT(pending->path); 1659 path = pending->path; 1660 1661 ASSERT(pending->root_item); 1662 new_root_item = pending->root_item; 1663 1664 /* 1665 * We're inside a transaction and must make sure that any potential 1666 * allocations with GFP_KERNEL in fscrypt won't recurse back to 1667 * filesystem. 1668 */ 1669 nofs_flags = memalloc_nofs_save(); 1670 pending->error = fscrypt_setup_filename(&parent_inode->vfs_inode, 1671 &pending->dentry->d_name, 0, 1672 &fname); 1673 memalloc_nofs_restore(nofs_flags); 1674 if (pending->error) 1675 goto free_pending; 1676 1677 pending->error = btrfs_get_free_objectid(tree_root, &objectid); 1678 if (pending->error) 1679 goto free_fname; 1680 1681 /* 1682 * Make qgroup to skip current new snapshot's qgroupid, as it is 1683 * accounted by later btrfs_qgroup_inherit(). 1684 */ 1685 btrfs_set_skip_qgroup(trans, objectid); 1686 1687 btrfs_reloc_pre_snapshot(pending, &to_reserve); 1688 1689 if (to_reserve > 0) { 1690 pending->error = btrfs_block_rsv_add(fs_info, 1691 &pending->block_rsv, 1692 to_reserve, 1693 BTRFS_RESERVE_NO_FLUSH); 1694 if (pending->error) 1695 goto clear_skip_qgroup; 1696 } 1697 1698 key.objectid = objectid; 1699 key.type = BTRFS_ROOT_ITEM_KEY; 1700 key.offset = (u64)-1; 1701 1702 rsv = trans->block_rsv; 1703 trans->block_rsv = &pending->block_rsv; 1704 trans->bytes_reserved = trans->block_rsv->reserved; 1705 trace_btrfs_space_reservation(fs_info, "transaction", 1706 trans->transid, 1707 trans->bytes_reserved, 1); 1708 parent_root = parent_inode->root; 1709 ret = record_root_in_trans(trans, parent_root, 0); 1710 if (ret) 1711 goto fail; 1712 cur_time = current_time(&parent_inode->vfs_inode); 1713 1714 /* 1715 * insert the directory item 1716 */ 1717 ret = btrfs_set_inode_index(parent_inode, &index); 1718 if (ret) { 1719 btrfs_abort_transaction(trans, ret); 1720 goto fail; 1721 } 1722 1723 /* check if there is a file/dir which has the same name. */ 1724 dir_item = btrfs_lookup_dir_item(NULL, parent_root, path, 1725 btrfs_ino(parent_inode), 1726 &fname.disk_name, 0); 1727 if (dir_item != NULL && !IS_ERR(dir_item)) { 1728 pending->error = -EEXIST; 1729 goto dir_item_existed; 1730 } else if (IS_ERR(dir_item)) { 1731 ret = PTR_ERR(dir_item); 1732 btrfs_abort_transaction(trans, ret); 1733 goto fail; 1734 } 1735 btrfs_release_path(path); 1736 1737 ret = btrfs_create_qgroup(trans, objectid); 1738 if (ret && ret != -EEXIST) { 1739 btrfs_abort_transaction(trans, ret); 1740 goto fail; 1741 } 1742 1743 /* 1744 * pull in the delayed directory update 1745 * and the delayed inode item 1746 * otherwise we corrupt the FS during 1747 * snapshot 1748 */ 1749 ret = btrfs_run_delayed_items(trans); 1750 if (ret) { /* Transaction aborted */ 1751 btrfs_abort_transaction(trans, ret); 1752 goto fail; 1753 } 1754 1755 ret = record_root_in_trans(trans, root, 0); 1756 if (ret) { 1757 btrfs_abort_transaction(trans, ret); 1758 goto fail; 1759 } 1760 btrfs_set_root_last_snapshot(&root->root_item, trans->transid); 1761 memcpy(new_root_item, &root->root_item, sizeof(*new_root_item)); 1762 btrfs_check_and_init_root_item(new_root_item); 1763 1764 root_flags = btrfs_root_flags(new_root_item); 1765 if (pending->readonly) 1766 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY; 1767 else 1768 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY; 1769 btrfs_set_root_flags(new_root_item, root_flags); 1770 1771 btrfs_set_root_generation_v2(new_root_item, 1772 trans->transid); 1773 generate_random_guid(new_root_item->uuid); 1774 memcpy(new_root_item->parent_uuid, root->root_item.uuid, 1775 BTRFS_UUID_SIZE); 1776 if (!(root_flags & BTRFS_ROOT_SUBVOL_RDONLY)) { 1777 memset(new_root_item->received_uuid, 0, 1778 sizeof(new_root_item->received_uuid)); 1779 memset(&new_root_item->stime, 0, sizeof(new_root_item->stime)); 1780 memset(&new_root_item->rtime, 0, sizeof(new_root_item->rtime)); 1781 btrfs_set_root_stransid(new_root_item, 0); 1782 btrfs_set_root_rtransid(new_root_item, 0); 1783 } 1784 btrfs_set_stack_timespec_sec(&new_root_item->otime, cur_time.tv_sec); 1785 btrfs_set_stack_timespec_nsec(&new_root_item->otime, cur_time.tv_nsec); 1786 btrfs_set_root_otransid(new_root_item, trans->transid); 1787 1788 old = btrfs_lock_root_node(root); 1789 ret = btrfs_cow_block(trans, root, old, NULL, 0, &old, 1790 BTRFS_NESTING_COW); 1791 if (ret) { 1792 btrfs_tree_unlock(old); 1793 free_extent_buffer(old); 1794 btrfs_abort_transaction(trans, ret); 1795 goto fail; 1796 } 1797 1798 ret = btrfs_copy_root(trans, root, old, &tmp, objectid); 1799 /* clean up in any case */ 1800 btrfs_tree_unlock(old); 1801 free_extent_buffer(old); 1802 if (ret) { 1803 btrfs_abort_transaction(trans, ret); 1804 goto fail; 1805 } 1806 /* see comments in should_cow_block() */ 1807 set_bit(BTRFS_ROOT_FORCE_COW, &root->state); 1808 smp_wmb(); 1809 1810 btrfs_set_root_node(new_root_item, tmp); 1811 /* record when the snapshot was created in key.offset */ 1812 key.offset = trans->transid; 1813 ret = btrfs_insert_root(trans, tree_root, &key, new_root_item); 1814 btrfs_tree_unlock(tmp); 1815 free_extent_buffer(tmp); 1816 if (ret) { 1817 btrfs_abort_transaction(trans, ret); 1818 goto fail; 1819 } 1820 1821 /* 1822 * insert root back/forward references 1823 */ 1824 ret = btrfs_add_root_ref(trans, objectid, 1825 btrfs_root_id(parent_root), 1826 btrfs_ino(parent_inode), index, 1827 &fname.disk_name); 1828 if (ret) { 1829 btrfs_abort_transaction(trans, ret); 1830 goto fail; 1831 } 1832 1833 key.offset = (u64)-1; 1834 pending->snap = btrfs_get_new_fs_root(fs_info, objectid, &pending->anon_dev); 1835 if (IS_ERR(pending->snap)) { 1836 ret = PTR_ERR(pending->snap); 1837 pending->snap = NULL; 1838 btrfs_abort_transaction(trans, ret); 1839 goto fail; 1840 } 1841 1842 ret = btrfs_reloc_post_snapshot(trans, pending); 1843 if (ret) { 1844 btrfs_abort_transaction(trans, ret); 1845 goto fail; 1846 } 1847 1848 /* 1849 * Do special qgroup accounting for snapshot, as we do some qgroup 1850 * snapshot hack to do fast snapshot. 1851 * To co-operate with that hack, we do hack again. 1852 * Or snapshot will be greatly slowed down by a subtree qgroup rescan 1853 */ 1854 if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_FULL) 1855 ret = qgroup_account_snapshot(trans, root, parent_root, 1856 pending->inherit, objectid); 1857 else if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_SIMPLE) 1858 ret = btrfs_qgroup_inherit(trans, btrfs_root_id(root), objectid, 1859 btrfs_root_id(parent_root), pending->inherit); 1860 if (ret < 0) 1861 goto fail; 1862 1863 ret = btrfs_insert_dir_item(trans, &fname.disk_name, 1864 parent_inode, &key, BTRFS_FT_DIR, 1865 index); 1866 if (ret) { 1867 btrfs_abort_transaction(trans, ret); 1868 goto fail; 1869 } 1870 1871 btrfs_i_size_write(parent_inode, parent_inode->vfs_inode.i_size + 1872 fname.disk_name.len * 2); 1873 inode_set_mtime_to_ts(&parent_inode->vfs_inode, 1874 inode_set_ctime_current(&parent_inode->vfs_inode)); 1875 ret = btrfs_update_inode_fallback(trans, parent_inode); 1876 if (ret) { 1877 btrfs_abort_transaction(trans, ret); 1878 goto fail; 1879 } 1880 ret = btrfs_uuid_tree_add(trans, new_root_item->uuid, 1881 BTRFS_UUID_KEY_SUBVOL, 1882 objectid); 1883 if (ret) { 1884 btrfs_abort_transaction(trans, ret); 1885 goto fail; 1886 } 1887 if (!btrfs_is_empty_uuid(new_root_item->received_uuid)) { 1888 ret = btrfs_uuid_tree_add(trans, new_root_item->received_uuid, 1889 BTRFS_UUID_KEY_RECEIVED_SUBVOL, 1890 objectid); 1891 if (ret && ret != -EEXIST) { 1892 btrfs_abort_transaction(trans, ret); 1893 goto fail; 1894 } 1895 } 1896 1897 fail: 1898 pending->error = ret; 1899 dir_item_existed: 1900 trans->block_rsv = rsv; 1901 trans->bytes_reserved = 0; 1902 clear_skip_qgroup: 1903 btrfs_clear_skip_qgroup(trans); 1904 free_fname: 1905 fscrypt_free_filename(&fname); 1906 free_pending: 1907 kfree(new_root_item); 1908 pending->root_item = NULL; 1909 btrfs_free_path(path); 1910 pending->path = NULL; 1911 1912 return ret; 1913 } 1914 1915 /* 1916 * create all the snapshots we've scheduled for creation 1917 */ 1918 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans) 1919 { 1920 struct btrfs_pending_snapshot *pending, *next; 1921 struct list_head *head = &trans->transaction->pending_snapshots; 1922 int ret = 0; 1923 1924 list_for_each_entry_safe(pending, next, head, list) { 1925 list_del(&pending->list); 1926 ret = create_pending_snapshot(trans, pending); 1927 if (ret) 1928 break; 1929 } 1930 return ret; 1931 } 1932 1933 static void update_super_roots(struct btrfs_fs_info *fs_info) 1934 { 1935 struct btrfs_root_item *root_item; 1936 struct btrfs_super_block *super; 1937 1938 super = fs_info->super_copy; 1939 1940 root_item = &fs_info->chunk_root->root_item; 1941 super->chunk_root = root_item->bytenr; 1942 super->chunk_root_generation = root_item->generation; 1943 super->chunk_root_level = root_item->level; 1944 1945 root_item = &fs_info->tree_root->root_item; 1946 super->root = root_item->bytenr; 1947 super->generation = root_item->generation; 1948 super->root_level = root_item->level; 1949 if (btrfs_test_opt(fs_info, SPACE_CACHE)) 1950 super->cache_generation = root_item->generation; 1951 else if (test_bit(BTRFS_FS_CLEANUP_SPACE_CACHE_V1, &fs_info->flags)) 1952 super->cache_generation = 0; 1953 if (test_bit(BTRFS_FS_UPDATE_UUID_TREE_GEN, &fs_info->flags)) 1954 super->uuid_tree_generation = root_item->generation; 1955 } 1956 1957 int btrfs_transaction_blocked(struct btrfs_fs_info *info) 1958 { 1959 struct btrfs_transaction *trans; 1960 int ret = 0; 1961 1962 spin_lock(&info->trans_lock); 1963 trans = info->running_transaction; 1964 if (trans) 1965 ret = is_transaction_blocked(trans); 1966 spin_unlock(&info->trans_lock); 1967 return ret; 1968 } 1969 1970 void btrfs_commit_transaction_async(struct btrfs_trans_handle *trans) 1971 { 1972 struct btrfs_fs_info *fs_info = trans->fs_info; 1973 struct btrfs_transaction *cur_trans; 1974 1975 /* Kick the transaction kthread. */ 1976 set_bit(BTRFS_FS_COMMIT_TRANS, &fs_info->flags); 1977 wake_up_process(fs_info->transaction_kthread); 1978 1979 /* take transaction reference */ 1980 cur_trans = trans->transaction; 1981 refcount_inc(&cur_trans->use_count); 1982 1983 btrfs_end_transaction(trans); 1984 1985 /* 1986 * Wait for the current transaction commit to start and block 1987 * subsequent transaction joins 1988 */ 1989 btrfs_might_wait_for_state(fs_info, BTRFS_LOCKDEP_TRANS_COMMIT_PREP); 1990 wait_event(fs_info->transaction_blocked_wait, 1991 cur_trans->state >= TRANS_STATE_COMMIT_START || 1992 TRANS_ABORTED(cur_trans)); 1993 btrfs_put_transaction(cur_trans); 1994 } 1995 1996 /* 1997 * If there is a running transaction commit it or if it's already committing, 1998 * wait for its commit to complete. Does not start and commit a new transaction 1999 * if there isn't any running. 2000 */ 2001 int btrfs_commit_current_transaction(struct btrfs_root *root) 2002 { 2003 struct btrfs_trans_handle *trans; 2004 2005 trans = btrfs_attach_transaction_barrier(root); 2006 if (IS_ERR(trans)) { 2007 int ret = PTR_ERR(trans); 2008 2009 return (ret == -ENOENT) ? 0 : ret; 2010 } 2011 2012 return btrfs_commit_transaction(trans); 2013 } 2014 2015 static void cleanup_transaction(struct btrfs_trans_handle *trans, int err) 2016 { 2017 struct btrfs_fs_info *fs_info = trans->fs_info; 2018 struct btrfs_transaction *cur_trans = trans->transaction; 2019 2020 WARN_ON(refcount_read(&trans->use_count) > 1); 2021 2022 btrfs_abort_transaction(trans, err); 2023 2024 spin_lock(&fs_info->trans_lock); 2025 2026 /* 2027 * If the transaction is removed from the list, it means this 2028 * transaction has been committed successfully, so it is impossible 2029 * to call the cleanup function. 2030 */ 2031 BUG_ON(list_empty(&cur_trans->list)); 2032 2033 if (cur_trans == fs_info->running_transaction) { 2034 cur_trans->state = TRANS_STATE_COMMIT_DOING; 2035 spin_unlock(&fs_info->trans_lock); 2036 2037 /* 2038 * The thread has already released the lockdep map as reader 2039 * already in btrfs_commit_transaction(). 2040 */ 2041 btrfs_might_wait_for_event(fs_info, btrfs_trans_num_writers); 2042 wait_event(cur_trans->writer_wait, 2043 atomic_read(&cur_trans->num_writers) == 1); 2044 2045 spin_lock(&fs_info->trans_lock); 2046 } 2047 2048 /* 2049 * Now that we know no one else is still using the transaction we can 2050 * remove the transaction from the list of transactions. This avoids 2051 * the transaction kthread from cleaning up the transaction while some 2052 * other task is still using it, which could result in a use-after-free 2053 * on things like log trees, as it forces the transaction kthread to 2054 * wait for this transaction to be cleaned up by us. 2055 */ 2056 list_del_init(&cur_trans->list); 2057 2058 spin_unlock(&fs_info->trans_lock); 2059 2060 btrfs_cleanup_one_transaction(trans->transaction); 2061 2062 spin_lock(&fs_info->trans_lock); 2063 if (cur_trans == fs_info->running_transaction) 2064 fs_info->running_transaction = NULL; 2065 spin_unlock(&fs_info->trans_lock); 2066 2067 if (trans->type & __TRANS_FREEZABLE) 2068 sb_end_intwrite(fs_info->sb); 2069 btrfs_put_transaction(cur_trans); 2070 btrfs_put_transaction(cur_trans); 2071 2072 trace_btrfs_transaction_commit(fs_info); 2073 2074 if (current->journal_info == trans) 2075 current->journal_info = NULL; 2076 2077 /* 2078 * If relocation is running, we can't cancel scrub because that will 2079 * result in a deadlock. Before relocating a block group, relocation 2080 * pauses scrub, then starts and commits a transaction before unpausing 2081 * scrub. If the transaction commit is being done by the relocation 2082 * task or triggered by another task and the relocation task is waiting 2083 * for the commit, and we end up here due to an error in the commit 2084 * path, then calling btrfs_scrub_cancel() will deadlock, as we are 2085 * asking for scrub to stop while having it asked to be paused higher 2086 * above in relocation code. 2087 */ 2088 if (!test_bit(BTRFS_FS_RELOC_RUNNING, &fs_info->flags)) 2089 btrfs_scrub_cancel(fs_info); 2090 2091 kmem_cache_free(btrfs_trans_handle_cachep, trans); 2092 } 2093 2094 /* 2095 * Release reserved delayed ref space of all pending block groups of the 2096 * transaction and remove them from the list 2097 */ 2098 static void btrfs_cleanup_pending_block_groups(struct btrfs_trans_handle *trans) 2099 { 2100 struct btrfs_fs_info *fs_info = trans->fs_info; 2101 struct btrfs_block_group *block_group, *tmp; 2102 2103 list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) { 2104 btrfs_dec_delayed_refs_rsv_bg_inserts(fs_info); 2105 /* 2106 * Not strictly necessary to lock, as no other task will be using a 2107 * block_group on the new_bgs list during a transaction abort. 2108 */ 2109 spin_lock(&fs_info->unused_bgs_lock); 2110 list_del_init(&block_group->bg_list); 2111 btrfs_put_block_group(block_group); 2112 spin_unlock(&fs_info->unused_bgs_lock); 2113 } 2114 } 2115 2116 static inline int btrfs_start_delalloc_flush(struct btrfs_fs_info *fs_info) 2117 { 2118 /* 2119 * We use try_to_writeback_inodes_sb() here because if we used 2120 * btrfs_start_delalloc_roots we would deadlock with fs freeze. 2121 * Currently are holding the fs freeze lock, if we do an async flush 2122 * we'll do btrfs_join_transaction() and deadlock because we need to 2123 * wait for the fs freeze lock. Using the direct flushing we benefit 2124 * from already being in a transaction and our join_transaction doesn't 2125 * have to re-take the fs freeze lock. 2126 * 2127 * Note that try_to_writeback_inodes_sb() will only trigger writeback 2128 * if it can read lock sb->s_umount. It will always be able to lock it, 2129 * except when the filesystem is being unmounted or being frozen, but in 2130 * those cases sync_filesystem() is called, which results in calling 2131 * writeback_inodes_sb() while holding a write lock on sb->s_umount. 2132 * Note that we don't call writeback_inodes_sb() directly, because it 2133 * will emit a warning if sb->s_umount is not locked. 2134 */ 2135 if (btrfs_test_opt(fs_info, FLUSHONCOMMIT)) 2136 try_to_writeback_inodes_sb(fs_info->sb, WB_REASON_SYNC); 2137 return 0; 2138 } 2139 2140 static inline void btrfs_wait_delalloc_flush(struct btrfs_fs_info *fs_info) 2141 { 2142 if (btrfs_test_opt(fs_info, FLUSHONCOMMIT)) 2143 btrfs_wait_ordered_roots(fs_info, U64_MAX, NULL); 2144 } 2145 2146 /* 2147 * Add a pending snapshot associated with the given transaction handle to the 2148 * respective handle. This must be called after the transaction commit started 2149 * and while holding fs_info->trans_lock. 2150 * This serves to guarantee a caller of btrfs_commit_transaction() that it can 2151 * safely free the pending snapshot pointer in case btrfs_commit_transaction() 2152 * returns an error. 2153 */ 2154 static void add_pending_snapshot(struct btrfs_trans_handle *trans) 2155 { 2156 struct btrfs_transaction *cur_trans = trans->transaction; 2157 2158 if (!trans->pending_snapshot) 2159 return; 2160 2161 lockdep_assert_held(&trans->fs_info->trans_lock); 2162 ASSERT(cur_trans->state >= TRANS_STATE_COMMIT_PREP); 2163 2164 list_add(&trans->pending_snapshot->list, &cur_trans->pending_snapshots); 2165 } 2166 2167 static void update_commit_stats(struct btrfs_fs_info *fs_info, ktime_t interval) 2168 { 2169 fs_info->commit_stats.commit_count++; 2170 fs_info->commit_stats.last_commit_dur = interval; 2171 fs_info->commit_stats.max_commit_dur = 2172 max_t(u64, fs_info->commit_stats.max_commit_dur, interval); 2173 fs_info->commit_stats.total_commit_dur += interval; 2174 } 2175 2176 int btrfs_commit_transaction(struct btrfs_trans_handle *trans) 2177 { 2178 struct btrfs_fs_info *fs_info = trans->fs_info; 2179 struct btrfs_transaction *cur_trans = trans->transaction; 2180 struct btrfs_transaction *prev_trans = NULL; 2181 int ret; 2182 ktime_t start_time; 2183 ktime_t interval; 2184 2185 ASSERT(refcount_read(&trans->use_count) == 1); 2186 btrfs_trans_state_lockdep_acquire(fs_info, BTRFS_LOCKDEP_TRANS_COMMIT_PREP); 2187 2188 clear_bit(BTRFS_FS_NEED_TRANS_COMMIT, &fs_info->flags); 2189 2190 /* Stop the commit early if ->aborted is set */ 2191 if (TRANS_ABORTED(cur_trans)) { 2192 ret = cur_trans->aborted; 2193 goto lockdep_trans_commit_start_release; 2194 } 2195 2196 btrfs_trans_release_metadata(trans); 2197 trans->block_rsv = NULL; 2198 2199 /* 2200 * We only want one transaction commit doing the flushing so we do not 2201 * waste a bunch of time on lock contention on the extent root node. 2202 */ 2203 if (!test_and_set_bit(BTRFS_DELAYED_REFS_FLUSHING, 2204 &cur_trans->delayed_refs.flags)) { 2205 /* 2206 * Make a pass through all the delayed refs we have so far. 2207 * Any running threads may add more while we are here. 2208 */ 2209 ret = btrfs_run_delayed_refs(trans, 0); 2210 if (ret) 2211 goto lockdep_trans_commit_start_release; 2212 } 2213 2214 btrfs_create_pending_block_groups(trans); 2215 2216 if (!test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &cur_trans->flags)) { 2217 int run_it = 0; 2218 2219 /* this mutex is also taken before trying to set 2220 * block groups readonly. We need to make sure 2221 * that nobody has set a block group readonly 2222 * after a extents from that block group have been 2223 * allocated for cache files. btrfs_set_block_group_ro 2224 * will wait for the transaction to commit if it 2225 * finds BTRFS_TRANS_DIRTY_BG_RUN set. 2226 * 2227 * The BTRFS_TRANS_DIRTY_BG_RUN flag is also used to make sure 2228 * only one process starts all the block group IO. It wouldn't 2229 * hurt to have more than one go through, but there's no 2230 * real advantage to it either. 2231 */ 2232 mutex_lock(&fs_info->ro_block_group_mutex); 2233 if (!test_and_set_bit(BTRFS_TRANS_DIRTY_BG_RUN, 2234 &cur_trans->flags)) 2235 run_it = 1; 2236 mutex_unlock(&fs_info->ro_block_group_mutex); 2237 2238 if (run_it) { 2239 ret = btrfs_start_dirty_block_groups(trans); 2240 if (ret) 2241 goto lockdep_trans_commit_start_release; 2242 } 2243 } 2244 2245 spin_lock(&fs_info->trans_lock); 2246 if (cur_trans->state >= TRANS_STATE_COMMIT_PREP) { 2247 enum btrfs_trans_state want_state = TRANS_STATE_COMPLETED; 2248 2249 add_pending_snapshot(trans); 2250 2251 spin_unlock(&fs_info->trans_lock); 2252 refcount_inc(&cur_trans->use_count); 2253 2254 if (trans->in_fsync) 2255 want_state = TRANS_STATE_SUPER_COMMITTED; 2256 2257 btrfs_trans_state_lockdep_release(fs_info, 2258 BTRFS_LOCKDEP_TRANS_COMMIT_PREP); 2259 ret = btrfs_end_transaction(trans); 2260 wait_for_commit(cur_trans, want_state); 2261 2262 if (TRANS_ABORTED(cur_trans)) 2263 ret = cur_trans->aborted; 2264 2265 btrfs_put_transaction(cur_trans); 2266 2267 return ret; 2268 } 2269 2270 cur_trans->state = TRANS_STATE_COMMIT_PREP; 2271 wake_up(&fs_info->transaction_blocked_wait); 2272 btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_COMMIT_PREP); 2273 2274 if (cur_trans->list.prev != &fs_info->trans_list) { 2275 enum btrfs_trans_state want_state = TRANS_STATE_COMPLETED; 2276 2277 if (trans->in_fsync) 2278 want_state = TRANS_STATE_SUPER_COMMITTED; 2279 2280 prev_trans = list_entry(cur_trans->list.prev, 2281 struct btrfs_transaction, list); 2282 if (prev_trans->state < want_state) { 2283 refcount_inc(&prev_trans->use_count); 2284 spin_unlock(&fs_info->trans_lock); 2285 2286 wait_for_commit(prev_trans, want_state); 2287 2288 ret = READ_ONCE(prev_trans->aborted); 2289 2290 btrfs_put_transaction(prev_trans); 2291 if (ret) 2292 goto lockdep_release; 2293 spin_lock(&fs_info->trans_lock); 2294 } 2295 } else { 2296 /* 2297 * The previous transaction was aborted and was already removed 2298 * from the list of transactions at fs_info->trans_list. So we 2299 * abort to prevent writing a new superblock that reflects a 2300 * corrupt state (pointing to trees with unwritten nodes/leafs). 2301 */ 2302 if (BTRFS_FS_ERROR(fs_info)) { 2303 spin_unlock(&fs_info->trans_lock); 2304 ret = -EROFS; 2305 goto lockdep_release; 2306 } 2307 } 2308 2309 cur_trans->state = TRANS_STATE_COMMIT_START; 2310 wake_up(&fs_info->transaction_blocked_wait); 2311 spin_unlock(&fs_info->trans_lock); 2312 2313 /* 2314 * Get the time spent on the work done by the commit thread and not 2315 * the time spent waiting on a previous commit 2316 */ 2317 start_time = ktime_get_ns(); 2318 2319 extwriter_counter_dec(cur_trans, trans->type); 2320 2321 ret = btrfs_start_delalloc_flush(fs_info); 2322 if (ret) 2323 goto lockdep_release; 2324 2325 ret = btrfs_run_delayed_items(trans); 2326 if (ret) 2327 goto lockdep_release; 2328 2329 /* 2330 * The thread has started/joined the transaction thus it holds the 2331 * lockdep map as a reader. It has to release it before acquiring the 2332 * lockdep map as a writer. 2333 */ 2334 btrfs_lockdep_release(fs_info, btrfs_trans_num_extwriters); 2335 btrfs_might_wait_for_event(fs_info, btrfs_trans_num_extwriters); 2336 wait_event(cur_trans->writer_wait, 2337 extwriter_counter_read(cur_trans) == 0); 2338 2339 /* some pending stuffs might be added after the previous flush. */ 2340 ret = btrfs_run_delayed_items(trans); 2341 if (ret) { 2342 btrfs_lockdep_release(fs_info, btrfs_trans_num_writers); 2343 goto cleanup_transaction; 2344 } 2345 2346 btrfs_wait_delalloc_flush(fs_info); 2347 2348 /* 2349 * Wait for all ordered extents started by a fast fsync that joined this 2350 * transaction. Otherwise if this transaction commits before the ordered 2351 * extents complete we lose logged data after a power failure. 2352 */ 2353 btrfs_might_wait_for_event(fs_info, btrfs_trans_pending_ordered); 2354 wait_event(cur_trans->pending_wait, 2355 atomic_read(&cur_trans->pending_ordered) == 0); 2356 2357 btrfs_scrub_pause(fs_info); 2358 /* 2359 * Ok now we need to make sure to block out any other joins while we 2360 * commit the transaction. We could have started a join before setting 2361 * COMMIT_DOING so make sure to wait for num_writers to == 1 again. 2362 */ 2363 spin_lock(&fs_info->trans_lock); 2364 add_pending_snapshot(trans); 2365 cur_trans->state = TRANS_STATE_COMMIT_DOING; 2366 spin_unlock(&fs_info->trans_lock); 2367 2368 /* 2369 * The thread has started/joined the transaction thus it holds the 2370 * lockdep map as a reader. It has to release it before acquiring the 2371 * lockdep map as a writer. 2372 */ 2373 btrfs_lockdep_release(fs_info, btrfs_trans_num_writers); 2374 btrfs_might_wait_for_event(fs_info, btrfs_trans_num_writers); 2375 wait_event(cur_trans->writer_wait, 2376 atomic_read(&cur_trans->num_writers) == 1); 2377 2378 /* 2379 * Make lockdep happy by acquiring the state locks after 2380 * btrfs_trans_num_writers is released. If we acquired the state locks 2381 * before releasing the btrfs_trans_num_writers lock then lockdep would 2382 * complain because we did not follow the reverse order unlocking rule. 2383 */ 2384 btrfs_trans_state_lockdep_acquire(fs_info, BTRFS_LOCKDEP_TRANS_COMPLETED); 2385 btrfs_trans_state_lockdep_acquire(fs_info, BTRFS_LOCKDEP_TRANS_SUPER_COMMITTED); 2386 btrfs_trans_state_lockdep_acquire(fs_info, BTRFS_LOCKDEP_TRANS_UNBLOCKED); 2387 2388 /* 2389 * We've started the commit, clear the flag in case we were triggered to 2390 * do an async commit but somebody else started before the transaction 2391 * kthread could do the work. 2392 */ 2393 clear_bit(BTRFS_FS_COMMIT_TRANS, &fs_info->flags); 2394 2395 if (TRANS_ABORTED(cur_trans)) { 2396 ret = cur_trans->aborted; 2397 btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_UNBLOCKED); 2398 goto scrub_continue; 2399 } 2400 /* 2401 * the reloc mutex makes sure that we stop 2402 * the balancing code from coming in and moving 2403 * extents around in the middle of the commit 2404 */ 2405 mutex_lock(&fs_info->reloc_mutex); 2406 2407 /* 2408 * We needn't worry about the delayed items because we will 2409 * deal with them in create_pending_snapshot(), which is the 2410 * core function of the snapshot creation. 2411 */ 2412 ret = create_pending_snapshots(trans); 2413 if (ret) 2414 goto unlock_reloc; 2415 2416 /* 2417 * We insert the dir indexes of the snapshots and update the inode 2418 * of the snapshots' parents after the snapshot creation, so there 2419 * are some delayed items which are not dealt with. Now deal with 2420 * them. 2421 * 2422 * We needn't worry that this operation will corrupt the snapshots, 2423 * because all the tree which are snapshoted will be forced to COW 2424 * the nodes and leaves. 2425 */ 2426 ret = btrfs_run_delayed_items(trans); 2427 if (ret) 2428 goto unlock_reloc; 2429 2430 ret = btrfs_run_delayed_refs(trans, U64_MAX); 2431 if (ret) 2432 goto unlock_reloc; 2433 2434 /* 2435 * make sure none of the code above managed to slip in a 2436 * delayed item 2437 */ 2438 btrfs_assert_delayed_root_empty(fs_info); 2439 2440 WARN_ON(cur_trans != trans->transaction); 2441 2442 ret = commit_fs_roots(trans); 2443 if (ret) 2444 goto unlock_reloc; 2445 2446 /* commit_fs_roots gets rid of all the tree log roots, it is now 2447 * safe to free the root of tree log roots 2448 */ 2449 btrfs_free_log_root_tree(trans, fs_info); 2450 2451 /* 2452 * Since fs roots are all committed, we can get a quite accurate 2453 * new_roots. So let's do quota accounting. 2454 */ 2455 ret = btrfs_qgroup_account_extents(trans); 2456 if (ret < 0) 2457 goto unlock_reloc; 2458 2459 ret = commit_cowonly_roots(trans); 2460 if (ret) 2461 goto unlock_reloc; 2462 2463 /* 2464 * The tasks which save the space cache and inode cache may also 2465 * update ->aborted, check it. 2466 */ 2467 if (TRANS_ABORTED(cur_trans)) { 2468 ret = cur_trans->aborted; 2469 goto unlock_reloc; 2470 } 2471 2472 cur_trans = fs_info->running_transaction; 2473 2474 btrfs_set_root_node(&fs_info->tree_root->root_item, 2475 fs_info->tree_root->node); 2476 list_add_tail(&fs_info->tree_root->dirty_list, 2477 &cur_trans->switch_commits); 2478 2479 btrfs_set_root_node(&fs_info->chunk_root->root_item, 2480 fs_info->chunk_root->node); 2481 list_add_tail(&fs_info->chunk_root->dirty_list, 2482 &cur_trans->switch_commits); 2483 2484 if (btrfs_fs_incompat(fs_info, EXTENT_TREE_V2)) { 2485 btrfs_set_root_node(&fs_info->block_group_root->root_item, 2486 fs_info->block_group_root->node); 2487 list_add_tail(&fs_info->block_group_root->dirty_list, 2488 &cur_trans->switch_commits); 2489 } 2490 2491 switch_commit_roots(trans); 2492 2493 ASSERT(list_empty(&cur_trans->dirty_bgs)); 2494 ASSERT(list_empty(&cur_trans->io_bgs)); 2495 update_super_roots(fs_info); 2496 2497 btrfs_set_super_log_root(fs_info->super_copy, 0); 2498 btrfs_set_super_log_root_level(fs_info->super_copy, 0); 2499 memcpy(fs_info->super_for_commit, fs_info->super_copy, 2500 sizeof(*fs_info->super_copy)); 2501 2502 btrfs_commit_device_sizes(cur_trans); 2503 2504 clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags); 2505 clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags); 2506 2507 btrfs_trans_release_chunk_metadata(trans); 2508 2509 /* 2510 * Before changing the transaction state to TRANS_STATE_UNBLOCKED and 2511 * setting fs_info->running_transaction to NULL, lock tree_log_mutex to 2512 * make sure that before we commit our superblock, no other task can 2513 * start a new transaction and commit a log tree before we commit our 2514 * superblock. Anyone trying to commit a log tree locks this mutex before 2515 * writing its superblock. 2516 */ 2517 mutex_lock(&fs_info->tree_log_mutex); 2518 2519 spin_lock(&fs_info->trans_lock); 2520 cur_trans->state = TRANS_STATE_UNBLOCKED; 2521 fs_info->running_transaction = NULL; 2522 spin_unlock(&fs_info->trans_lock); 2523 mutex_unlock(&fs_info->reloc_mutex); 2524 2525 wake_up(&fs_info->transaction_wait); 2526 btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_UNBLOCKED); 2527 2528 /* If we have features changed, wake up the cleaner to update sysfs. */ 2529 if (test_bit(BTRFS_FS_FEATURE_CHANGED, &fs_info->flags) && 2530 fs_info->cleaner_kthread) 2531 wake_up_process(fs_info->cleaner_kthread); 2532 2533 ret = btrfs_write_and_wait_transaction(trans); 2534 if (ret) { 2535 btrfs_handle_fs_error(fs_info, ret, 2536 "Error while writing out transaction"); 2537 mutex_unlock(&fs_info->tree_log_mutex); 2538 goto scrub_continue; 2539 } 2540 2541 ret = write_all_supers(fs_info, 0); 2542 /* 2543 * the super is written, we can safely allow the tree-loggers 2544 * to go about their business 2545 */ 2546 mutex_unlock(&fs_info->tree_log_mutex); 2547 if (ret) 2548 goto scrub_continue; 2549 2550 /* 2551 * We needn't acquire the lock here because there is no other task 2552 * which can change it. 2553 */ 2554 cur_trans->state = TRANS_STATE_SUPER_COMMITTED; 2555 wake_up(&cur_trans->commit_wait); 2556 btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_SUPER_COMMITTED); 2557 2558 btrfs_finish_extent_commit(trans); 2559 2560 if (test_bit(BTRFS_TRANS_HAVE_FREE_BGS, &cur_trans->flags)) 2561 btrfs_clear_space_info_full(fs_info); 2562 2563 btrfs_set_last_trans_committed(fs_info, cur_trans->transid); 2564 /* 2565 * We needn't acquire the lock here because there is no other task 2566 * which can change it. 2567 */ 2568 cur_trans->state = TRANS_STATE_COMPLETED; 2569 wake_up(&cur_trans->commit_wait); 2570 btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_COMPLETED); 2571 2572 spin_lock(&fs_info->trans_lock); 2573 list_del_init(&cur_trans->list); 2574 spin_unlock(&fs_info->trans_lock); 2575 2576 btrfs_put_transaction(cur_trans); 2577 btrfs_put_transaction(cur_trans); 2578 2579 if (trans->type & __TRANS_FREEZABLE) 2580 sb_end_intwrite(fs_info->sb); 2581 2582 trace_btrfs_transaction_commit(fs_info); 2583 2584 interval = ktime_get_ns() - start_time; 2585 2586 btrfs_scrub_continue(fs_info); 2587 2588 if (current->journal_info == trans) 2589 current->journal_info = NULL; 2590 2591 kmem_cache_free(btrfs_trans_handle_cachep, trans); 2592 2593 update_commit_stats(fs_info, interval); 2594 2595 return ret; 2596 2597 unlock_reloc: 2598 mutex_unlock(&fs_info->reloc_mutex); 2599 btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_UNBLOCKED); 2600 scrub_continue: 2601 btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_SUPER_COMMITTED); 2602 btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_COMPLETED); 2603 btrfs_scrub_continue(fs_info); 2604 cleanup_transaction: 2605 btrfs_trans_release_metadata(trans); 2606 btrfs_cleanup_pending_block_groups(trans); 2607 btrfs_trans_release_chunk_metadata(trans); 2608 trans->block_rsv = NULL; 2609 btrfs_warn(fs_info, "Skipping commit of aborted transaction."); 2610 if (current->journal_info == trans) 2611 current->journal_info = NULL; 2612 cleanup_transaction(trans, ret); 2613 2614 return ret; 2615 2616 lockdep_release: 2617 btrfs_lockdep_release(fs_info, btrfs_trans_num_extwriters); 2618 btrfs_lockdep_release(fs_info, btrfs_trans_num_writers); 2619 goto cleanup_transaction; 2620 2621 lockdep_trans_commit_start_release: 2622 btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_COMMIT_PREP); 2623 btrfs_end_transaction(trans); 2624 return ret; 2625 } 2626 2627 /* 2628 * return < 0 if error 2629 * 0 if there are no more dead_roots at the time of call 2630 * 1 there are more to be processed, call me again 2631 * 2632 * The return value indicates there are certainly more snapshots to delete, but 2633 * if there comes a new one during processing, it may return 0. We don't mind, 2634 * because btrfs_commit_super will poke cleaner thread and it will process it a 2635 * few seconds later. 2636 */ 2637 int btrfs_clean_one_deleted_snapshot(struct btrfs_fs_info *fs_info) 2638 { 2639 struct btrfs_root *root; 2640 int ret; 2641 2642 spin_lock(&fs_info->trans_lock); 2643 if (list_empty(&fs_info->dead_roots)) { 2644 spin_unlock(&fs_info->trans_lock); 2645 return 0; 2646 } 2647 root = list_first_entry(&fs_info->dead_roots, 2648 struct btrfs_root, root_list); 2649 list_del_init(&root->root_list); 2650 spin_unlock(&fs_info->trans_lock); 2651 2652 btrfs_debug(fs_info, "cleaner removing %llu", btrfs_root_id(root)); 2653 2654 btrfs_kill_all_delayed_nodes(root); 2655 2656 if (btrfs_header_backref_rev(root->node) < 2657 BTRFS_MIXED_BACKREF_REV) 2658 ret = btrfs_drop_snapshot(root, 0, 0); 2659 else 2660 ret = btrfs_drop_snapshot(root, 1, 0); 2661 2662 btrfs_put_root(root); 2663 return (ret < 0) ? 0 : 1; 2664 } 2665 2666 /* 2667 * We only mark the transaction aborted and then set the file system read-only. 2668 * This will prevent new transactions from starting or trying to join this 2669 * one. 2670 * 2671 * This means that error recovery at the call site is limited to freeing 2672 * any local memory allocations and passing the error code up without 2673 * further cleanup. The transaction should complete as it normally would 2674 * in the call path but will return -EIO. 2675 * 2676 * We'll complete the cleanup in btrfs_end_transaction and 2677 * btrfs_commit_transaction. 2678 */ 2679 void __cold __btrfs_abort_transaction(struct btrfs_trans_handle *trans, 2680 const char *function, 2681 unsigned int line, int error, bool first_hit) 2682 { 2683 struct btrfs_fs_info *fs_info = trans->fs_info; 2684 2685 WRITE_ONCE(trans->aborted, error); 2686 WRITE_ONCE(trans->transaction->aborted, error); 2687 if (first_hit && error == -ENOSPC) 2688 btrfs_dump_space_info_for_trans_abort(fs_info); 2689 /* Wake up anybody who may be waiting on this transaction */ 2690 wake_up(&fs_info->transaction_wait); 2691 wake_up(&fs_info->transaction_blocked_wait); 2692 __btrfs_handle_fs_error(fs_info, function, line, error, NULL); 2693 } 2694 2695 int __init btrfs_transaction_init(void) 2696 { 2697 btrfs_trans_handle_cachep = KMEM_CACHE(btrfs_trans_handle, SLAB_TEMPORARY); 2698 if (!btrfs_trans_handle_cachep) 2699 return -ENOMEM; 2700 return 0; 2701 } 2702 2703 void __cold btrfs_transaction_exit(void) 2704 { 2705 kmem_cache_destroy(btrfs_trans_handle_cachep); 2706 } 2707