Loading...
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 "zoned.h"
27#include "fs.h"
28#include "accessors.h"
29#include "extent-tree.h"
30#include "root-tree.h"
31#include "defrag.h"
32#include "dir-item.h"
33#include "uuid-tree.h"
34#include "ioctl.h"
35#include "relocation.h"
36#include "scrub.h"
37
38static struct kmem_cache *btrfs_trans_handle_cachep;
39
40#define BTRFS_ROOT_TRANS_TAG 0
41
42/*
43 * Transaction states and transitions
44 *
45 * No running transaction (fs tree blocks are not modified)
46 * |
47 * | To next stage:
48 * | Call start_transaction() variants. Except btrfs_join_transaction_nostart().
49 * V
50 * Transaction N [[TRANS_STATE_RUNNING]]
51 * |
52 * | New trans handles can be attached to transaction N by calling all
53 * | start_transaction() variants.
54 * |
55 * | To next stage:
56 * | Call btrfs_commit_transaction() on any trans handle attached to
57 * | transaction N
58 * V
59 * Transaction N [[TRANS_STATE_COMMIT_START]]
60 * |
61 * | Will wait for previous running transaction to completely finish if there
62 * | is one
63 * |
64 * | Then one of the following happes:
65 * | - Wait for all other trans handle holders to release.
66 * | The btrfs_commit_transaction() caller will do the commit work.
67 * | - Wait for current transaction to be committed by others.
68 * | Other btrfs_commit_transaction() caller will do the commit work.
69 * |
70 * | At this stage, only btrfs_join_transaction*() variants can attach
71 * | to this running transaction.
72 * | All other variants will wait for current one to finish and attach to
73 * | transaction N+1.
74 * |
75 * | To next stage:
76 * | Caller is chosen to commit transaction N, and all other trans handle
77 * | haven been released.
78 * V
79 * Transaction N [[TRANS_STATE_COMMIT_DOING]]
80 * |
81 * | The heavy lifting transaction work is started.
82 * | From running delayed refs (modifying extent tree) to creating pending
83 * | snapshots, running qgroups.
84 * | In short, modify supporting trees to reflect modifications of subvolume
85 * | trees.
86 * |
87 * | At this stage, all start_transaction() calls will wait for this
88 * | transaction to finish and attach to transaction N+1.
89 * |
90 * | To next stage:
91 * | Until all supporting trees are updated.
92 * V
93 * Transaction N [[TRANS_STATE_UNBLOCKED]]
94 * | Transaction N+1
95 * | All needed trees are modified, thus we only [[TRANS_STATE_RUNNING]]
96 * | need to write them back to disk and update |
97 * | super blocks. |
98 * | |
99 * | At this stage, new transaction is allowed to |
100 * | start. |
101 * | All new start_transaction() calls will be |
102 * | attached to transid N+1. |
103 * | |
104 * | To next stage: |
105 * | Until all tree blocks are super blocks are |
106 * | written to block devices |
107 * V |
108 * Transaction N [[TRANS_STATE_COMPLETED]] V
109 * All tree blocks and super blocks are written. Transaction N+1
110 * This transaction is finished and all its [[TRANS_STATE_COMMIT_START]]
111 * data structures will be cleaned up. | Life goes on
112 */
113static const unsigned int btrfs_blocked_trans_types[TRANS_STATE_MAX] = {
114 [TRANS_STATE_RUNNING] = 0U,
115 [TRANS_STATE_COMMIT_START] = (__TRANS_START | __TRANS_ATTACH),
116 [TRANS_STATE_COMMIT_DOING] = (__TRANS_START |
117 __TRANS_ATTACH |
118 __TRANS_JOIN |
119 __TRANS_JOIN_NOSTART),
120 [TRANS_STATE_UNBLOCKED] = (__TRANS_START |
121 __TRANS_ATTACH |
122 __TRANS_JOIN |
123 __TRANS_JOIN_NOLOCK |
124 __TRANS_JOIN_NOSTART),
125 [TRANS_STATE_SUPER_COMMITTED] = (__TRANS_START |
126 __TRANS_ATTACH |
127 __TRANS_JOIN |
128 __TRANS_JOIN_NOLOCK |
129 __TRANS_JOIN_NOSTART),
130 [TRANS_STATE_COMPLETED] = (__TRANS_START |
131 __TRANS_ATTACH |
132 __TRANS_JOIN |
133 __TRANS_JOIN_NOLOCK |
134 __TRANS_JOIN_NOSTART),
135};
136
137void btrfs_put_transaction(struct btrfs_transaction *transaction)
138{
139 WARN_ON(refcount_read(&transaction->use_count) == 0);
140 if (refcount_dec_and_test(&transaction->use_count)) {
141 BUG_ON(!list_empty(&transaction->list));
142 WARN_ON(!RB_EMPTY_ROOT(
143 &transaction->delayed_refs.href_root.rb_root));
144 WARN_ON(!RB_EMPTY_ROOT(
145 &transaction->delayed_refs.dirty_extent_root));
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 list_del_init(&cache->bg_list);
164 btrfs_unfreeze_block_group(cache);
165 btrfs_put_block_group(cache);
166 }
167 WARN_ON(!list_empty(&transaction->dev_update_list));
168 kfree(transaction);
169 }
170}
171
172static noinline void switch_commit_roots(struct btrfs_trans_handle *trans)
173{
174 struct btrfs_transaction *cur_trans = trans->transaction;
175 struct btrfs_fs_info *fs_info = trans->fs_info;
176 struct btrfs_root *root, *tmp;
177
178 /*
179 * At this point no one can be using this transaction to modify any tree
180 * and no one can start another transaction to modify any tree either.
181 */
182 ASSERT(cur_trans->state == TRANS_STATE_COMMIT_DOING);
183
184 down_write(&fs_info->commit_root_sem);
185
186 if (test_bit(BTRFS_FS_RELOC_RUNNING, &fs_info->flags))
187 fs_info->last_reloc_trans = trans->transid;
188
189 list_for_each_entry_safe(root, tmp, &cur_trans->switch_commits,
190 dirty_list) {
191 list_del_init(&root->dirty_list);
192 free_extent_buffer(root->commit_root);
193 root->commit_root = btrfs_root_node(root);
194 extent_io_tree_release(&root->dirty_log_pages);
195 btrfs_qgroup_clean_swapped_blocks(root);
196 }
197
198 /* We can free old roots now. */
199 spin_lock(&cur_trans->dropped_roots_lock);
200 while (!list_empty(&cur_trans->dropped_roots)) {
201 root = list_first_entry(&cur_trans->dropped_roots,
202 struct btrfs_root, root_list);
203 list_del_init(&root->root_list);
204 spin_unlock(&cur_trans->dropped_roots_lock);
205 btrfs_free_log(trans, root);
206 btrfs_drop_and_free_fs_root(fs_info, root);
207 spin_lock(&cur_trans->dropped_roots_lock);
208 }
209 spin_unlock(&cur_trans->dropped_roots_lock);
210
211 up_write(&fs_info->commit_root_sem);
212}
213
214static inline void extwriter_counter_inc(struct btrfs_transaction *trans,
215 unsigned int type)
216{
217 if (type & TRANS_EXTWRITERS)
218 atomic_inc(&trans->num_extwriters);
219}
220
221static inline void extwriter_counter_dec(struct btrfs_transaction *trans,
222 unsigned int type)
223{
224 if (type & TRANS_EXTWRITERS)
225 atomic_dec(&trans->num_extwriters);
226}
227
228static inline void extwriter_counter_init(struct btrfs_transaction *trans,
229 unsigned int type)
230{
231 atomic_set(&trans->num_extwriters, ((type & TRANS_EXTWRITERS) ? 1 : 0));
232}
233
234static inline int extwriter_counter_read(struct btrfs_transaction *trans)
235{
236 return atomic_read(&trans->num_extwriters);
237}
238
239/*
240 * To be called after doing the chunk btree updates right after allocating a new
241 * chunk (after btrfs_chunk_alloc_add_chunk_item() is called), when removing a
242 * chunk after all chunk btree updates and after finishing the second phase of
243 * chunk allocation (btrfs_create_pending_block_groups()) in case some block
244 * group had its chunk item insertion delayed to the second phase.
245 */
246void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
247{
248 struct btrfs_fs_info *fs_info = trans->fs_info;
249
250 if (!trans->chunk_bytes_reserved)
251 return;
252
253 btrfs_block_rsv_release(fs_info, &fs_info->chunk_block_rsv,
254 trans->chunk_bytes_reserved, NULL);
255 trans->chunk_bytes_reserved = 0;
256}
257
258/*
259 * either allocate a new transaction or hop into the existing one
260 */
261static noinline int join_transaction(struct btrfs_fs_info *fs_info,
262 unsigned int type)
263{
264 struct btrfs_transaction *cur_trans;
265
266 spin_lock(&fs_info->trans_lock);
267loop:
268 /* The file system has been taken offline. No new transactions. */
269 if (BTRFS_FS_ERROR(fs_info)) {
270 spin_unlock(&fs_info->trans_lock);
271 return -EROFS;
272 }
273
274 cur_trans = fs_info->running_transaction;
275 if (cur_trans) {
276 if (TRANS_ABORTED(cur_trans)) {
277 spin_unlock(&fs_info->trans_lock);
278 return cur_trans->aborted;
279 }
280 if (btrfs_blocked_trans_types[cur_trans->state] & type) {
281 spin_unlock(&fs_info->trans_lock);
282 return -EBUSY;
283 }
284 refcount_inc(&cur_trans->use_count);
285 atomic_inc(&cur_trans->num_writers);
286 extwriter_counter_inc(cur_trans, type);
287 spin_unlock(&fs_info->trans_lock);
288 btrfs_lockdep_acquire(fs_info, btrfs_trans_num_writers);
289 btrfs_lockdep_acquire(fs_info, btrfs_trans_num_extwriters);
290 return 0;
291 }
292 spin_unlock(&fs_info->trans_lock);
293
294 /*
295 * If we are ATTACH, we just want to catch the current transaction,
296 * and commit it. If there is no transaction, just return ENOENT.
297 */
298 if (type == TRANS_ATTACH)
299 return -ENOENT;
300
301 /*
302 * JOIN_NOLOCK only happens during the transaction commit, so
303 * it is impossible that ->running_transaction is NULL
304 */
305 BUG_ON(type == TRANS_JOIN_NOLOCK);
306
307 cur_trans = kmalloc(sizeof(*cur_trans), GFP_NOFS);
308 if (!cur_trans)
309 return -ENOMEM;
310
311 btrfs_lockdep_acquire(fs_info, btrfs_trans_num_writers);
312 btrfs_lockdep_acquire(fs_info, btrfs_trans_num_extwriters);
313
314 spin_lock(&fs_info->trans_lock);
315 if (fs_info->running_transaction) {
316 /*
317 * someone started a transaction after we unlocked. Make sure
318 * to redo the checks above
319 */
320 btrfs_lockdep_release(fs_info, btrfs_trans_num_extwriters);
321 btrfs_lockdep_release(fs_info, btrfs_trans_num_writers);
322 kfree(cur_trans);
323 goto loop;
324 } else if (BTRFS_FS_ERROR(fs_info)) {
325 spin_unlock(&fs_info->trans_lock);
326 btrfs_lockdep_release(fs_info, btrfs_trans_num_extwriters);
327 btrfs_lockdep_release(fs_info, btrfs_trans_num_writers);
328 kfree(cur_trans);
329 return -EROFS;
330 }
331
332 cur_trans->fs_info = fs_info;
333 atomic_set(&cur_trans->pending_ordered, 0);
334 init_waitqueue_head(&cur_trans->pending_wait);
335 atomic_set(&cur_trans->num_writers, 1);
336 extwriter_counter_init(cur_trans, type);
337 init_waitqueue_head(&cur_trans->writer_wait);
338 init_waitqueue_head(&cur_trans->commit_wait);
339 cur_trans->state = TRANS_STATE_RUNNING;
340 /*
341 * One for this trans handle, one so it will live on until we
342 * commit the transaction.
343 */
344 refcount_set(&cur_trans->use_count, 2);
345 cur_trans->flags = 0;
346 cur_trans->start_time = ktime_get_seconds();
347
348 memset(&cur_trans->delayed_refs, 0, sizeof(cur_trans->delayed_refs));
349
350 cur_trans->delayed_refs.href_root = RB_ROOT_CACHED;
351 cur_trans->delayed_refs.dirty_extent_root = RB_ROOT;
352 atomic_set(&cur_trans->delayed_refs.num_entries, 0);
353
354 /*
355 * although the tree mod log is per file system and not per transaction,
356 * the log must never go across transaction boundaries.
357 */
358 smp_mb();
359 if (!list_empty(&fs_info->tree_mod_seq_list))
360 WARN(1, KERN_ERR "BTRFS: tree_mod_seq_list not empty when creating a fresh transaction\n");
361 if (!RB_EMPTY_ROOT(&fs_info->tree_mod_log))
362 WARN(1, KERN_ERR "BTRFS: tree_mod_log rb tree not empty when creating a fresh transaction\n");
363 atomic64_set(&fs_info->tree_mod_seq, 0);
364
365 spin_lock_init(&cur_trans->delayed_refs.lock);
366
367 INIT_LIST_HEAD(&cur_trans->pending_snapshots);
368 INIT_LIST_HEAD(&cur_trans->dev_update_list);
369 INIT_LIST_HEAD(&cur_trans->switch_commits);
370 INIT_LIST_HEAD(&cur_trans->dirty_bgs);
371 INIT_LIST_HEAD(&cur_trans->io_bgs);
372 INIT_LIST_HEAD(&cur_trans->dropped_roots);
373 mutex_init(&cur_trans->cache_write_mutex);
374 spin_lock_init(&cur_trans->dirty_bgs_lock);
375 INIT_LIST_HEAD(&cur_trans->deleted_bgs);
376 spin_lock_init(&cur_trans->dropped_roots_lock);
377 INIT_LIST_HEAD(&cur_trans->releasing_ebs);
378 spin_lock_init(&cur_trans->releasing_ebs_lock);
379 list_add_tail(&cur_trans->list, &fs_info->trans_list);
380 extent_io_tree_init(fs_info, &cur_trans->dirty_pages,
381 IO_TREE_TRANS_DIRTY_PAGES);
382 extent_io_tree_init(fs_info, &cur_trans->pinned_extents,
383 IO_TREE_FS_PINNED_EXTENTS);
384 fs_info->generation++;
385 cur_trans->transid = fs_info->generation;
386 fs_info->running_transaction = cur_trans;
387 cur_trans->aborted = 0;
388 spin_unlock(&fs_info->trans_lock);
389
390 return 0;
391}
392
393/*
394 * This does all the record keeping required to make sure that a shareable root
395 * is properly recorded in a given transaction. This is required to make sure
396 * the old root from before we joined the transaction is deleted when the
397 * transaction commits.
398 */
399static int record_root_in_trans(struct btrfs_trans_handle *trans,
400 struct btrfs_root *root,
401 int force)
402{
403 struct btrfs_fs_info *fs_info = root->fs_info;
404 int ret = 0;
405
406 if ((test_bit(BTRFS_ROOT_SHAREABLE, &root->state) &&
407 root->last_trans < trans->transid) || force) {
408 WARN_ON(!force && root->commit_root != root->node);
409
410 /*
411 * see below for IN_TRANS_SETUP usage rules
412 * we have the reloc mutex held now, so there
413 * is only one writer in this function
414 */
415 set_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
416
417 /* make sure readers find IN_TRANS_SETUP before
418 * they find our root->last_trans update
419 */
420 smp_wmb();
421
422 spin_lock(&fs_info->fs_roots_radix_lock);
423 if (root->last_trans == trans->transid && !force) {
424 spin_unlock(&fs_info->fs_roots_radix_lock);
425 return 0;
426 }
427 radix_tree_tag_set(&fs_info->fs_roots_radix,
428 (unsigned long)root->root_key.objectid,
429 BTRFS_ROOT_TRANS_TAG);
430 spin_unlock(&fs_info->fs_roots_radix_lock);
431 root->last_trans = trans->transid;
432
433 /* this is pretty tricky. We don't want to
434 * take the relocation lock in btrfs_record_root_in_trans
435 * unless we're really doing the first setup for this root in
436 * this transaction.
437 *
438 * Normally we'd use root->last_trans as a flag to decide
439 * if we want to take the expensive mutex.
440 *
441 * But, we have to set root->last_trans before we
442 * init the relocation root, otherwise, we trip over warnings
443 * in ctree.c. The solution used here is to flag ourselves
444 * with root IN_TRANS_SETUP. When this is 1, we're still
445 * fixing up the reloc trees and everyone must wait.
446 *
447 * When this is zero, they can trust root->last_trans and fly
448 * through btrfs_record_root_in_trans without having to take the
449 * lock. smp_wmb() makes sure that all the writes above are
450 * done before we pop in the zero below
451 */
452 ret = btrfs_init_reloc_root(trans, root);
453 smp_mb__before_atomic();
454 clear_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
455 }
456 return ret;
457}
458
459
460void btrfs_add_dropped_root(struct btrfs_trans_handle *trans,
461 struct btrfs_root *root)
462{
463 struct btrfs_fs_info *fs_info = root->fs_info;
464 struct btrfs_transaction *cur_trans = trans->transaction;
465
466 /* Add ourselves to the transaction dropped list */
467 spin_lock(&cur_trans->dropped_roots_lock);
468 list_add_tail(&root->root_list, &cur_trans->dropped_roots);
469 spin_unlock(&cur_trans->dropped_roots_lock);
470
471 /* Make sure we don't try to update the root at commit time */
472 spin_lock(&fs_info->fs_roots_radix_lock);
473 radix_tree_tag_clear(&fs_info->fs_roots_radix,
474 (unsigned long)root->root_key.objectid,
475 BTRFS_ROOT_TRANS_TAG);
476 spin_unlock(&fs_info->fs_roots_radix_lock);
477}
478
479int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
480 struct btrfs_root *root)
481{
482 struct btrfs_fs_info *fs_info = root->fs_info;
483 int ret;
484
485 if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
486 return 0;
487
488 /*
489 * see record_root_in_trans for comments about IN_TRANS_SETUP usage
490 * and barriers
491 */
492 smp_rmb();
493 if (root->last_trans == trans->transid &&
494 !test_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state))
495 return 0;
496
497 mutex_lock(&fs_info->reloc_mutex);
498 ret = record_root_in_trans(trans, root, 0);
499 mutex_unlock(&fs_info->reloc_mutex);
500
501 return ret;
502}
503
504static inline int is_transaction_blocked(struct btrfs_transaction *trans)
505{
506 return (trans->state >= TRANS_STATE_COMMIT_START &&
507 trans->state < TRANS_STATE_UNBLOCKED &&
508 !TRANS_ABORTED(trans));
509}
510
511/* wait for commit against the current transaction to become unblocked
512 * when this is done, it is safe to start a new transaction, but the current
513 * transaction might not be fully on disk.
514 */
515static void wait_current_trans(struct btrfs_fs_info *fs_info)
516{
517 struct btrfs_transaction *cur_trans;
518
519 spin_lock(&fs_info->trans_lock);
520 cur_trans = fs_info->running_transaction;
521 if (cur_trans && is_transaction_blocked(cur_trans)) {
522 refcount_inc(&cur_trans->use_count);
523 spin_unlock(&fs_info->trans_lock);
524
525 btrfs_might_wait_for_state(fs_info, BTRFS_LOCKDEP_TRANS_UNBLOCKED);
526 wait_event(fs_info->transaction_wait,
527 cur_trans->state >= TRANS_STATE_UNBLOCKED ||
528 TRANS_ABORTED(cur_trans));
529 btrfs_put_transaction(cur_trans);
530 } else {
531 spin_unlock(&fs_info->trans_lock);
532 }
533}
534
535static int may_wait_transaction(struct btrfs_fs_info *fs_info, int type)
536{
537 if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
538 return 0;
539
540 if (type == TRANS_START)
541 return 1;
542
543 return 0;
544}
545
546static inline bool need_reserve_reloc_root(struct btrfs_root *root)
547{
548 struct btrfs_fs_info *fs_info = root->fs_info;
549
550 if (!fs_info->reloc_ctl ||
551 !test_bit(BTRFS_ROOT_SHAREABLE, &root->state) ||
552 root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
553 root->reloc_root)
554 return false;
555
556 return true;
557}
558
559static struct btrfs_trans_handle *
560start_transaction(struct btrfs_root *root, unsigned int num_items,
561 unsigned int type, enum btrfs_reserve_flush_enum flush,
562 bool enforce_qgroups)
563{
564 struct btrfs_fs_info *fs_info = root->fs_info;
565 struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
566 struct btrfs_trans_handle *h;
567 struct btrfs_transaction *cur_trans;
568 u64 num_bytes = 0;
569 u64 qgroup_reserved = 0;
570 bool reloc_reserved = false;
571 bool do_chunk_alloc = false;
572 int ret;
573
574 if (BTRFS_FS_ERROR(fs_info))
575 return ERR_PTR(-EROFS);
576
577 if (current->journal_info) {
578 WARN_ON(type & TRANS_EXTWRITERS);
579 h = current->journal_info;
580 refcount_inc(&h->use_count);
581 WARN_ON(refcount_read(&h->use_count) > 2);
582 h->orig_rsv = h->block_rsv;
583 h->block_rsv = NULL;
584 goto got_it;
585 }
586
587 /*
588 * Do the reservation before we join the transaction so we can do all
589 * the appropriate flushing if need be.
590 */
591 if (num_items && root != fs_info->chunk_root) {
592 struct btrfs_block_rsv *rsv = &fs_info->trans_block_rsv;
593 u64 delayed_refs_bytes = 0;
594
595 qgroup_reserved = num_items * fs_info->nodesize;
596 ret = btrfs_qgroup_reserve_meta_pertrans(root, qgroup_reserved,
597 enforce_qgroups);
598 if (ret)
599 return ERR_PTR(ret);
600
601 /*
602 * We want to reserve all the bytes we may need all at once, so
603 * we only do 1 enospc flushing cycle per transaction start. We
604 * accomplish this by simply assuming we'll do 2 x num_items
605 * worth of delayed refs updates in this trans handle, and
606 * refill that amount for whatever is missing in the reserve.
607 */
608 num_bytes = btrfs_calc_insert_metadata_size(fs_info, num_items);
609 if (flush == BTRFS_RESERVE_FLUSH_ALL &&
610 btrfs_block_rsv_full(delayed_refs_rsv) == 0) {
611 delayed_refs_bytes = num_bytes;
612 num_bytes <<= 1;
613 }
614
615 /*
616 * Do the reservation for the relocation root creation
617 */
618 if (need_reserve_reloc_root(root)) {
619 num_bytes += fs_info->nodesize;
620 reloc_reserved = true;
621 }
622
623 ret = btrfs_block_rsv_add(fs_info, rsv, num_bytes, flush);
624 if (ret)
625 goto reserve_fail;
626 if (delayed_refs_bytes) {
627 btrfs_migrate_to_delayed_refs_rsv(fs_info, rsv,
628 delayed_refs_bytes);
629 num_bytes -= delayed_refs_bytes;
630 }
631
632 if (rsv->space_info->force_alloc)
633 do_chunk_alloc = true;
634 } else if (num_items == 0 && flush == BTRFS_RESERVE_FLUSH_ALL &&
635 !btrfs_block_rsv_full(delayed_refs_rsv)) {
636 /*
637 * Some people call with btrfs_start_transaction(root, 0)
638 * because they can be throttled, but have some other mechanism
639 * for reserving space. We still want these guys to refill the
640 * delayed block_rsv so just add 1 items worth of reservation
641 * here.
642 */
643 ret = btrfs_delayed_refs_rsv_refill(fs_info, flush);
644 if (ret)
645 goto reserve_fail;
646 }
647again:
648 h = kmem_cache_zalloc(btrfs_trans_handle_cachep, GFP_NOFS);
649 if (!h) {
650 ret = -ENOMEM;
651 goto alloc_fail;
652 }
653
654 /*
655 * If we are JOIN_NOLOCK we're already committing a transaction and
656 * waiting on this guy, so we don't need to do the sb_start_intwrite
657 * because we're already holding a ref. We need this because we could
658 * have raced in and did an fsync() on a file which can kick a commit
659 * and then we deadlock with somebody doing a freeze.
660 *
661 * If we are ATTACH, it means we just want to catch the current
662 * transaction and commit it, so we needn't do sb_start_intwrite().
663 */
664 if (type & __TRANS_FREEZABLE)
665 sb_start_intwrite(fs_info->sb);
666
667 if (may_wait_transaction(fs_info, type))
668 wait_current_trans(fs_info);
669
670 do {
671 ret = join_transaction(fs_info, type);
672 if (ret == -EBUSY) {
673 wait_current_trans(fs_info);
674 if (unlikely(type == TRANS_ATTACH ||
675 type == TRANS_JOIN_NOSTART))
676 ret = -ENOENT;
677 }
678 } while (ret == -EBUSY);
679
680 if (ret < 0)
681 goto join_fail;
682
683 cur_trans = fs_info->running_transaction;
684
685 h->transid = cur_trans->transid;
686 h->transaction = cur_trans;
687 refcount_set(&h->use_count, 1);
688 h->fs_info = root->fs_info;
689
690 h->type = type;
691 INIT_LIST_HEAD(&h->new_bgs);
692
693 smp_mb();
694 if (cur_trans->state >= TRANS_STATE_COMMIT_START &&
695 may_wait_transaction(fs_info, type)) {
696 current->journal_info = h;
697 btrfs_commit_transaction(h);
698 goto again;
699 }
700
701 if (num_bytes) {
702 trace_btrfs_space_reservation(fs_info, "transaction",
703 h->transid, num_bytes, 1);
704 h->block_rsv = &fs_info->trans_block_rsv;
705 h->bytes_reserved = num_bytes;
706 h->reloc_reserved = reloc_reserved;
707 }
708
709got_it:
710 if (!current->journal_info)
711 current->journal_info = h;
712
713 /*
714 * If the space_info is marked ALLOC_FORCE then we'll get upgraded to
715 * ALLOC_FORCE the first run through, and then we won't allocate for
716 * anybody else who races in later. We don't care about the return
717 * value here.
718 */
719 if (do_chunk_alloc && num_bytes) {
720 u64 flags = h->block_rsv->space_info->flags;
721
722 btrfs_chunk_alloc(h, btrfs_get_alloc_profile(fs_info, flags),
723 CHUNK_ALLOC_NO_FORCE);
724 }
725
726 /*
727 * btrfs_record_root_in_trans() needs to alloc new extents, and may
728 * call btrfs_join_transaction() while we're also starting a
729 * transaction.
730 *
731 * Thus it need to be called after current->journal_info initialized,
732 * or we can deadlock.
733 */
734 ret = btrfs_record_root_in_trans(h, root);
735 if (ret) {
736 /*
737 * The transaction handle is fully initialized and linked with
738 * other structures so it needs to be ended in case of errors,
739 * not just freed.
740 */
741 btrfs_end_transaction(h);
742 return ERR_PTR(ret);
743 }
744
745 return h;
746
747join_fail:
748 if (type & __TRANS_FREEZABLE)
749 sb_end_intwrite(fs_info->sb);
750 kmem_cache_free(btrfs_trans_handle_cachep, h);
751alloc_fail:
752 if (num_bytes)
753 btrfs_block_rsv_release(fs_info, &fs_info->trans_block_rsv,
754 num_bytes, NULL);
755reserve_fail:
756 btrfs_qgroup_free_meta_pertrans(root, qgroup_reserved);
757 return ERR_PTR(ret);
758}
759
760struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
761 unsigned int num_items)
762{
763 return start_transaction(root, num_items, TRANS_START,
764 BTRFS_RESERVE_FLUSH_ALL, true);
765}
766
767struct btrfs_trans_handle *btrfs_start_transaction_fallback_global_rsv(
768 struct btrfs_root *root,
769 unsigned int num_items)
770{
771 return start_transaction(root, num_items, TRANS_START,
772 BTRFS_RESERVE_FLUSH_ALL_STEAL, false);
773}
774
775struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
776{
777 return start_transaction(root, 0, TRANS_JOIN, BTRFS_RESERVE_NO_FLUSH,
778 true);
779}
780
781struct btrfs_trans_handle *btrfs_join_transaction_spacecache(struct btrfs_root *root)
782{
783 return start_transaction(root, 0, TRANS_JOIN_NOLOCK,
784 BTRFS_RESERVE_NO_FLUSH, true);
785}
786
787/*
788 * Similar to regular join but it never starts a transaction when none is
789 * running or after waiting for the current one to finish.
790 */
791struct btrfs_trans_handle *btrfs_join_transaction_nostart(struct btrfs_root *root)
792{
793 return start_transaction(root, 0, TRANS_JOIN_NOSTART,
794 BTRFS_RESERVE_NO_FLUSH, true);
795}
796
797/*
798 * btrfs_attach_transaction() - catch the running transaction
799 *
800 * It is used when we want to commit the current the transaction, but
801 * don't want to start a new one.
802 *
803 * Note: If this function return -ENOENT, it just means there is no
804 * running transaction. But it is possible that the inactive transaction
805 * is still in the memory, not fully on disk. If you hope there is no
806 * inactive transaction in the fs when -ENOENT is returned, you should
807 * invoke
808 * btrfs_attach_transaction_barrier()
809 */
810struct btrfs_trans_handle *btrfs_attach_transaction(struct btrfs_root *root)
811{
812 return start_transaction(root, 0, TRANS_ATTACH,
813 BTRFS_RESERVE_NO_FLUSH, true);
814}
815
816/*
817 * btrfs_attach_transaction_barrier() - catch the running transaction
818 *
819 * It is similar to the above function, the difference is this one
820 * will wait for all the inactive transactions until they fully
821 * complete.
822 */
823struct btrfs_trans_handle *
824btrfs_attach_transaction_barrier(struct btrfs_root *root)
825{
826 struct btrfs_trans_handle *trans;
827
828 trans = start_transaction(root, 0, TRANS_ATTACH,
829 BTRFS_RESERVE_NO_FLUSH, true);
830 if (trans == ERR_PTR(-ENOENT))
831 btrfs_wait_for_commit(root->fs_info, 0);
832
833 return trans;
834}
835
836/* Wait for a transaction commit to reach at least the given state. */
837static noinline void wait_for_commit(struct btrfs_transaction *commit,
838 const enum btrfs_trans_state min_state)
839{
840 struct btrfs_fs_info *fs_info = commit->fs_info;
841 u64 transid = commit->transid;
842 bool put = false;
843
844 /*
845 * At the moment this function is called with min_state either being
846 * TRANS_STATE_COMPLETED or TRANS_STATE_SUPER_COMMITTED.
847 */
848 if (min_state == TRANS_STATE_COMPLETED)
849 btrfs_might_wait_for_state(fs_info, BTRFS_LOCKDEP_TRANS_COMPLETED);
850 else
851 btrfs_might_wait_for_state(fs_info, BTRFS_LOCKDEP_TRANS_SUPER_COMMITTED);
852
853 while (1) {
854 wait_event(commit->commit_wait, commit->state >= min_state);
855 if (put)
856 btrfs_put_transaction(commit);
857
858 if (min_state < TRANS_STATE_COMPLETED)
859 break;
860
861 /*
862 * A transaction isn't really completed until all of the
863 * previous transactions are completed, but with fsync we can
864 * end up with SUPER_COMMITTED transactions before a COMPLETED
865 * transaction. Wait for those.
866 */
867
868 spin_lock(&fs_info->trans_lock);
869 commit = list_first_entry_or_null(&fs_info->trans_list,
870 struct btrfs_transaction,
871 list);
872 if (!commit || commit->transid > transid) {
873 spin_unlock(&fs_info->trans_lock);
874 break;
875 }
876 refcount_inc(&commit->use_count);
877 put = true;
878 spin_unlock(&fs_info->trans_lock);
879 }
880}
881
882int btrfs_wait_for_commit(struct btrfs_fs_info *fs_info, u64 transid)
883{
884 struct btrfs_transaction *cur_trans = NULL, *t;
885 int ret = 0;
886
887 if (transid) {
888 if (transid <= fs_info->last_trans_committed)
889 goto out;
890
891 /* find specified transaction */
892 spin_lock(&fs_info->trans_lock);
893 list_for_each_entry(t, &fs_info->trans_list, list) {
894 if (t->transid == transid) {
895 cur_trans = t;
896 refcount_inc(&cur_trans->use_count);
897 ret = 0;
898 break;
899 }
900 if (t->transid > transid) {
901 ret = 0;
902 break;
903 }
904 }
905 spin_unlock(&fs_info->trans_lock);
906
907 /*
908 * The specified transaction doesn't exist, or we
909 * raced with btrfs_commit_transaction
910 */
911 if (!cur_trans) {
912 if (transid > fs_info->last_trans_committed)
913 ret = -EINVAL;
914 goto out;
915 }
916 } else {
917 /* find newest transaction that is committing | committed */
918 spin_lock(&fs_info->trans_lock);
919 list_for_each_entry_reverse(t, &fs_info->trans_list,
920 list) {
921 if (t->state >= TRANS_STATE_COMMIT_START) {
922 if (t->state == TRANS_STATE_COMPLETED)
923 break;
924 cur_trans = t;
925 refcount_inc(&cur_trans->use_count);
926 break;
927 }
928 }
929 spin_unlock(&fs_info->trans_lock);
930 if (!cur_trans)
931 goto out; /* nothing committing|committed */
932 }
933
934 wait_for_commit(cur_trans, TRANS_STATE_COMPLETED);
935 btrfs_put_transaction(cur_trans);
936out:
937 return ret;
938}
939
940void btrfs_throttle(struct btrfs_fs_info *fs_info)
941{
942 wait_current_trans(fs_info);
943}
944
945static bool should_end_transaction(struct btrfs_trans_handle *trans)
946{
947 struct btrfs_fs_info *fs_info = trans->fs_info;
948
949 if (btrfs_check_space_for_delayed_refs(fs_info))
950 return true;
951
952 return !!btrfs_block_rsv_check(&fs_info->global_block_rsv, 50);
953}
954
955bool btrfs_should_end_transaction(struct btrfs_trans_handle *trans)
956{
957 struct btrfs_transaction *cur_trans = trans->transaction;
958
959 if (cur_trans->state >= TRANS_STATE_COMMIT_START ||
960 test_bit(BTRFS_DELAYED_REFS_FLUSHING, &cur_trans->delayed_refs.flags))
961 return true;
962
963 return should_end_transaction(trans);
964}
965
966static void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans)
967
968{
969 struct btrfs_fs_info *fs_info = trans->fs_info;
970
971 if (!trans->block_rsv) {
972 ASSERT(!trans->bytes_reserved);
973 return;
974 }
975
976 if (!trans->bytes_reserved)
977 return;
978
979 ASSERT(trans->block_rsv == &fs_info->trans_block_rsv);
980 trace_btrfs_space_reservation(fs_info, "transaction",
981 trans->transid, trans->bytes_reserved, 0);
982 btrfs_block_rsv_release(fs_info, trans->block_rsv,
983 trans->bytes_reserved, NULL);
984 trans->bytes_reserved = 0;
985}
986
987static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
988 int throttle)
989{
990 struct btrfs_fs_info *info = trans->fs_info;
991 struct btrfs_transaction *cur_trans = trans->transaction;
992 int err = 0;
993
994 if (refcount_read(&trans->use_count) > 1) {
995 refcount_dec(&trans->use_count);
996 trans->block_rsv = trans->orig_rsv;
997 return 0;
998 }
999
1000 btrfs_trans_release_metadata(trans);
1001 trans->block_rsv = NULL;
1002
1003 btrfs_create_pending_block_groups(trans);
1004
1005 btrfs_trans_release_chunk_metadata(trans);
1006
1007 if (trans->type & __TRANS_FREEZABLE)
1008 sb_end_intwrite(info->sb);
1009
1010 WARN_ON(cur_trans != info->running_transaction);
1011 WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
1012 atomic_dec(&cur_trans->num_writers);
1013 extwriter_counter_dec(cur_trans, trans->type);
1014
1015 cond_wake_up(&cur_trans->writer_wait);
1016
1017 btrfs_lockdep_release(info, btrfs_trans_num_extwriters);
1018 btrfs_lockdep_release(info, btrfs_trans_num_writers);
1019
1020 btrfs_put_transaction(cur_trans);
1021
1022 if (current->journal_info == trans)
1023 current->journal_info = NULL;
1024
1025 if (throttle)
1026 btrfs_run_delayed_iputs(info);
1027
1028 if (TRANS_ABORTED(trans) || BTRFS_FS_ERROR(info)) {
1029 wake_up_process(info->transaction_kthread);
1030 if (TRANS_ABORTED(trans))
1031 err = trans->aborted;
1032 else
1033 err = -EROFS;
1034 }
1035
1036 kmem_cache_free(btrfs_trans_handle_cachep, trans);
1037 return err;
1038}
1039
1040int btrfs_end_transaction(struct btrfs_trans_handle *trans)
1041{
1042 return __btrfs_end_transaction(trans, 0);
1043}
1044
1045int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans)
1046{
1047 return __btrfs_end_transaction(trans, 1);
1048}
1049
1050/*
1051 * when btree blocks are allocated, they have some corresponding bits set for
1052 * them in one of two extent_io trees. This is used to make sure all of
1053 * those extents are sent to disk but does not wait on them
1054 */
1055int btrfs_write_marked_extents(struct btrfs_fs_info *fs_info,
1056 struct extent_io_tree *dirty_pages, int mark)
1057{
1058 int err = 0;
1059 int werr = 0;
1060 struct address_space *mapping = fs_info->btree_inode->i_mapping;
1061 struct extent_state *cached_state = NULL;
1062 u64 start = 0;
1063 u64 end;
1064
1065 atomic_inc(&BTRFS_I(fs_info->btree_inode)->sync_writers);
1066 while (!find_first_extent_bit(dirty_pages, start, &start, &end,
1067 mark, &cached_state)) {
1068 bool wait_writeback = false;
1069
1070 err = convert_extent_bit(dirty_pages, start, end,
1071 EXTENT_NEED_WAIT,
1072 mark, &cached_state);
1073 /*
1074 * convert_extent_bit can return -ENOMEM, which is most of the
1075 * time a temporary error. So when it happens, ignore the error
1076 * and wait for writeback of this range to finish - because we
1077 * failed to set the bit EXTENT_NEED_WAIT for the range, a call
1078 * to __btrfs_wait_marked_extents() would not know that
1079 * writeback for this range started and therefore wouldn't
1080 * wait for it to finish - we don't want to commit a
1081 * superblock that points to btree nodes/leafs for which
1082 * writeback hasn't finished yet (and without errors).
1083 * We cleanup any entries left in the io tree when committing
1084 * the transaction (through extent_io_tree_release()).
1085 */
1086 if (err == -ENOMEM) {
1087 err = 0;
1088 wait_writeback = true;
1089 }
1090 if (!err)
1091 err = filemap_fdatawrite_range(mapping, start, end);
1092 if (err)
1093 werr = err;
1094 else if (wait_writeback)
1095 werr = filemap_fdatawait_range(mapping, start, end);
1096 free_extent_state(cached_state);
1097 cached_state = NULL;
1098 cond_resched();
1099 start = end + 1;
1100 }
1101 atomic_dec(&BTRFS_I(fs_info->btree_inode)->sync_writers);
1102 return werr;
1103}
1104
1105/*
1106 * when btree blocks are allocated, they have some corresponding bits set for
1107 * them in one of two extent_io trees. This is used to make sure all of
1108 * those extents are on disk for transaction or log commit. We wait
1109 * on all the pages and clear them from the dirty pages state tree
1110 */
1111static int __btrfs_wait_marked_extents(struct btrfs_fs_info *fs_info,
1112 struct extent_io_tree *dirty_pages)
1113{
1114 int err = 0;
1115 int werr = 0;
1116 struct address_space *mapping = fs_info->btree_inode->i_mapping;
1117 struct extent_state *cached_state = NULL;
1118 u64 start = 0;
1119 u64 end;
1120
1121 while (!find_first_extent_bit(dirty_pages, start, &start, &end,
1122 EXTENT_NEED_WAIT, &cached_state)) {
1123 /*
1124 * Ignore -ENOMEM errors returned by clear_extent_bit().
1125 * When committing the transaction, we'll remove any entries
1126 * left in the io tree. For a log commit, we don't remove them
1127 * after committing the log because the tree can be accessed
1128 * concurrently - we do it only at transaction commit time when
1129 * it's safe to do it (through extent_io_tree_release()).
1130 */
1131 err = clear_extent_bit(dirty_pages, start, end,
1132 EXTENT_NEED_WAIT, &cached_state);
1133 if (err == -ENOMEM)
1134 err = 0;
1135 if (!err)
1136 err = filemap_fdatawait_range(mapping, start, end);
1137 if (err)
1138 werr = err;
1139 free_extent_state(cached_state);
1140 cached_state = NULL;
1141 cond_resched();
1142 start = end + 1;
1143 }
1144 if (err)
1145 werr = err;
1146 return werr;
1147}
1148
1149static int btrfs_wait_extents(struct btrfs_fs_info *fs_info,
1150 struct extent_io_tree *dirty_pages)
1151{
1152 bool errors = false;
1153 int err;
1154
1155 err = __btrfs_wait_marked_extents(fs_info, dirty_pages);
1156 if (test_and_clear_bit(BTRFS_FS_BTREE_ERR, &fs_info->flags))
1157 errors = true;
1158
1159 if (errors && !err)
1160 err = -EIO;
1161 return err;
1162}
1163
1164int btrfs_wait_tree_log_extents(struct btrfs_root *log_root, int mark)
1165{
1166 struct btrfs_fs_info *fs_info = log_root->fs_info;
1167 struct extent_io_tree *dirty_pages = &log_root->dirty_log_pages;
1168 bool errors = false;
1169 int err;
1170
1171 ASSERT(log_root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
1172
1173 err = __btrfs_wait_marked_extents(fs_info, dirty_pages);
1174 if ((mark & EXTENT_DIRTY) &&
1175 test_and_clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags))
1176 errors = true;
1177
1178 if ((mark & EXTENT_NEW) &&
1179 test_and_clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags))
1180 errors = true;
1181
1182 if (errors && !err)
1183 err = -EIO;
1184 return err;
1185}
1186
1187/*
1188 * When btree blocks are allocated the corresponding extents are marked dirty.
1189 * This function ensures such extents are persisted on disk for transaction or
1190 * log commit.
1191 *
1192 * @trans: transaction whose dirty pages we'd like to write
1193 */
1194static int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans)
1195{
1196 int ret;
1197 int ret2;
1198 struct extent_io_tree *dirty_pages = &trans->transaction->dirty_pages;
1199 struct btrfs_fs_info *fs_info = trans->fs_info;
1200 struct blk_plug plug;
1201
1202 blk_start_plug(&plug);
1203 ret = btrfs_write_marked_extents(fs_info, dirty_pages, EXTENT_DIRTY);
1204 blk_finish_plug(&plug);
1205 ret2 = btrfs_wait_extents(fs_info, dirty_pages);
1206
1207 extent_io_tree_release(&trans->transaction->dirty_pages);
1208
1209 if (ret)
1210 return ret;
1211 else if (ret2)
1212 return ret2;
1213 else
1214 return 0;
1215}
1216
1217/*
1218 * this is used to update the root pointer in the tree of tree roots.
1219 *
1220 * But, in the case of the extent allocation tree, updating the root
1221 * pointer may allocate blocks which may change the root of the extent
1222 * allocation tree.
1223 *
1224 * So, this loops and repeats and makes sure the cowonly root didn't
1225 * change while the root pointer was being updated in the metadata.
1226 */
1227static int update_cowonly_root(struct btrfs_trans_handle *trans,
1228 struct btrfs_root *root)
1229{
1230 int ret;
1231 u64 old_root_bytenr;
1232 u64 old_root_used;
1233 struct btrfs_fs_info *fs_info = root->fs_info;
1234 struct btrfs_root *tree_root = fs_info->tree_root;
1235
1236 old_root_used = btrfs_root_used(&root->root_item);
1237
1238 while (1) {
1239 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
1240 if (old_root_bytenr == root->node->start &&
1241 old_root_used == btrfs_root_used(&root->root_item))
1242 break;
1243
1244 btrfs_set_root_node(&root->root_item, root->node);
1245 ret = btrfs_update_root(trans, tree_root,
1246 &root->root_key,
1247 &root->root_item);
1248 if (ret)
1249 return ret;
1250
1251 old_root_used = btrfs_root_used(&root->root_item);
1252 }
1253
1254 return 0;
1255}
1256
1257/*
1258 * update all the cowonly tree roots on disk
1259 *
1260 * The error handling in this function may not be obvious. Any of the
1261 * failures will cause the file system to go offline. We still need
1262 * to clean up the delayed refs.
1263 */
1264static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans)
1265{
1266 struct btrfs_fs_info *fs_info = trans->fs_info;
1267 struct list_head *dirty_bgs = &trans->transaction->dirty_bgs;
1268 struct list_head *io_bgs = &trans->transaction->io_bgs;
1269 struct list_head *next;
1270 struct extent_buffer *eb;
1271 int ret;
1272
1273 /*
1274 * At this point no one can be using this transaction to modify any tree
1275 * and no one can start another transaction to modify any tree either.
1276 */
1277 ASSERT(trans->transaction->state == TRANS_STATE_COMMIT_DOING);
1278
1279 eb = btrfs_lock_root_node(fs_info->tree_root);
1280 ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL,
1281 0, &eb, BTRFS_NESTING_COW);
1282 btrfs_tree_unlock(eb);
1283 free_extent_buffer(eb);
1284
1285 if (ret)
1286 return ret;
1287
1288 ret = btrfs_run_dev_stats(trans);
1289 if (ret)
1290 return ret;
1291 ret = btrfs_run_dev_replace(trans);
1292 if (ret)
1293 return ret;
1294 ret = btrfs_run_qgroups(trans);
1295 if (ret)
1296 return ret;
1297
1298 ret = btrfs_setup_space_cache(trans);
1299 if (ret)
1300 return ret;
1301
1302again:
1303 while (!list_empty(&fs_info->dirty_cowonly_roots)) {
1304 struct btrfs_root *root;
1305 next = fs_info->dirty_cowonly_roots.next;
1306 list_del_init(next);
1307 root = list_entry(next, struct btrfs_root, dirty_list);
1308 clear_bit(BTRFS_ROOT_DIRTY, &root->state);
1309
1310 list_add_tail(&root->dirty_list,
1311 &trans->transaction->switch_commits);
1312 ret = update_cowonly_root(trans, root);
1313 if (ret)
1314 return ret;
1315 }
1316
1317 /* Now flush any delayed refs generated by updating all of the roots */
1318 ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
1319 if (ret)
1320 return ret;
1321
1322 while (!list_empty(dirty_bgs) || !list_empty(io_bgs)) {
1323 ret = btrfs_write_dirty_block_groups(trans);
1324 if (ret)
1325 return ret;
1326
1327 /*
1328 * We're writing the dirty block groups, which could generate
1329 * delayed refs, which could generate more dirty block groups,
1330 * so we want to keep this flushing in this loop to make sure
1331 * everything gets run.
1332 */
1333 ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
1334 if (ret)
1335 return ret;
1336 }
1337
1338 if (!list_empty(&fs_info->dirty_cowonly_roots))
1339 goto again;
1340
1341 /* Update dev-replace pointer once everything is committed */
1342 fs_info->dev_replace.committed_cursor_left =
1343 fs_info->dev_replace.cursor_left_last_write_of_item;
1344
1345 return 0;
1346}
1347
1348/*
1349 * If we had a pending drop we need to see if there are any others left in our
1350 * dead roots list, and if not clear our bit and wake any waiters.
1351 */
1352void btrfs_maybe_wake_unfinished_drop(struct btrfs_fs_info *fs_info)
1353{
1354 /*
1355 * We put the drop in progress roots at the front of the list, so if the
1356 * first entry doesn't have UNFINISHED_DROP set we can wake everybody
1357 * up.
1358 */
1359 spin_lock(&fs_info->trans_lock);
1360 if (!list_empty(&fs_info->dead_roots)) {
1361 struct btrfs_root *root = list_first_entry(&fs_info->dead_roots,
1362 struct btrfs_root,
1363 root_list);
1364 if (test_bit(BTRFS_ROOT_UNFINISHED_DROP, &root->state)) {
1365 spin_unlock(&fs_info->trans_lock);
1366 return;
1367 }
1368 }
1369 spin_unlock(&fs_info->trans_lock);
1370
1371 btrfs_wake_unfinished_drop(fs_info);
1372}
1373
1374/*
1375 * dead roots are old snapshots that need to be deleted. This allocates
1376 * a dirty root struct and adds it into the list of dead roots that need to
1377 * be deleted
1378 */
1379void btrfs_add_dead_root(struct btrfs_root *root)
1380{
1381 struct btrfs_fs_info *fs_info = root->fs_info;
1382
1383 spin_lock(&fs_info->trans_lock);
1384 if (list_empty(&root->root_list)) {
1385 btrfs_grab_root(root);
1386
1387 /* We want to process the partially complete drops first. */
1388 if (test_bit(BTRFS_ROOT_UNFINISHED_DROP, &root->state))
1389 list_add(&root->root_list, &fs_info->dead_roots);
1390 else
1391 list_add_tail(&root->root_list, &fs_info->dead_roots);
1392 }
1393 spin_unlock(&fs_info->trans_lock);
1394}
1395
1396/*
1397 * Update each subvolume root and its relocation root, if it exists, in the tree
1398 * of tree roots. Also free log roots if they exist.
1399 */
1400static noinline int commit_fs_roots(struct btrfs_trans_handle *trans)
1401{
1402 struct btrfs_fs_info *fs_info = trans->fs_info;
1403 struct btrfs_root *gang[8];
1404 int i;
1405 int ret;
1406
1407 /*
1408 * At this point no one can be using this transaction to modify any tree
1409 * and no one can start another transaction to modify any tree either.
1410 */
1411 ASSERT(trans->transaction->state == TRANS_STATE_COMMIT_DOING);
1412
1413 spin_lock(&fs_info->fs_roots_radix_lock);
1414 while (1) {
1415 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
1416 (void **)gang, 0,
1417 ARRAY_SIZE(gang),
1418 BTRFS_ROOT_TRANS_TAG);
1419 if (ret == 0)
1420 break;
1421 for (i = 0; i < ret; i++) {
1422 struct btrfs_root *root = gang[i];
1423 int ret2;
1424
1425 /*
1426 * At this point we can neither have tasks logging inodes
1427 * from a root nor trying to commit a log tree.
1428 */
1429 ASSERT(atomic_read(&root->log_writers) == 0);
1430 ASSERT(atomic_read(&root->log_commit[0]) == 0);
1431 ASSERT(atomic_read(&root->log_commit[1]) == 0);
1432
1433 radix_tree_tag_clear(&fs_info->fs_roots_radix,
1434 (unsigned long)root->root_key.objectid,
1435 BTRFS_ROOT_TRANS_TAG);
1436 spin_unlock(&fs_info->fs_roots_radix_lock);
1437
1438 btrfs_free_log(trans, root);
1439 ret2 = btrfs_update_reloc_root(trans, root);
1440 if (ret2)
1441 return ret2;
1442
1443 /* see comments in should_cow_block() */
1444 clear_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1445 smp_mb__after_atomic();
1446
1447 if (root->commit_root != root->node) {
1448 list_add_tail(&root->dirty_list,
1449 &trans->transaction->switch_commits);
1450 btrfs_set_root_node(&root->root_item,
1451 root->node);
1452 }
1453
1454 ret2 = btrfs_update_root(trans, fs_info->tree_root,
1455 &root->root_key,
1456 &root->root_item);
1457 if (ret2)
1458 return ret2;
1459 spin_lock(&fs_info->fs_roots_radix_lock);
1460 btrfs_qgroup_free_meta_all_pertrans(root);
1461 }
1462 }
1463 spin_unlock(&fs_info->fs_roots_radix_lock);
1464 return 0;
1465}
1466
1467/*
1468 * defrag a given btree.
1469 * Every leaf in the btree is read and defragged.
1470 */
1471int btrfs_defrag_root(struct btrfs_root *root)
1472{
1473 struct btrfs_fs_info *info = root->fs_info;
1474 struct btrfs_trans_handle *trans;
1475 int ret;
1476
1477 if (test_and_set_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state))
1478 return 0;
1479
1480 while (1) {
1481 trans = btrfs_start_transaction(root, 0);
1482 if (IS_ERR(trans)) {
1483 ret = PTR_ERR(trans);
1484 break;
1485 }
1486
1487 ret = btrfs_defrag_leaves(trans, root);
1488
1489 btrfs_end_transaction(trans);
1490 btrfs_btree_balance_dirty(info);
1491 cond_resched();
1492
1493 if (btrfs_fs_closing(info) || ret != -EAGAIN)
1494 break;
1495
1496 if (btrfs_defrag_cancelled(info)) {
1497 btrfs_debug(info, "defrag_root cancelled");
1498 ret = -EAGAIN;
1499 break;
1500 }
1501 }
1502 clear_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state);
1503 return ret;
1504}
1505
1506/*
1507 * Do all special snapshot related qgroup dirty hack.
1508 *
1509 * Will do all needed qgroup inherit and dirty hack like switch commit
1510 * roots inside one transaction and write all btree into disk, to make
1511 * qgroup works.
1512 */
1513static int qgroup_account_snapshot(struct btrfs_trans_handle *trans,
1514 struct btrfs_root *src,
1515 struct btrfs_root *parent,
1516 struct btrfs_qgroup_inherit *inherit,
1517 u64 dst_objectid)
1518{
1519 struct btrfs_fs_info *fs_info = src->fs_info;
1520 int ret;
1521
1522 /*
1523 * Save some performance in the case that qgroups are not
1524 * enabled. If this check races with the ioctl, rescan will
1525 * kick in anyway.
1526 */
1527 if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
1528 return 0;
1529
1530 /*
1531 * Ensure dirty @src will be committed. Or, after coming
1532 * commit_fs_roots() and switch_commit_roots(), any dirty but not
1533 * recorded root will never be updated again, causing an outdated root
1534 * item.
1535 */
1536 ret = record_root_in_trans(trans, src, 1);
1537 if (ret)
1538 return ret;
1539
1540 /*
1541 * btrfs_qgroup_inherit relies on a consistent view of the usage for the
1542 * src root, so we must run the delayed refs here.
1543 *
1544 * However this isn't particularly fool proof, because there's no
1545 * synchronization keeping us from changing the tree after this point
1546 * before we do the qgroup_inherit, or even from making changes while
1547 * we're doing the qgroup_inherit. But that's a problem for the future,
1548 * for now flush the delayed refs to narrow the race window where the
1549 * qgroup counters could end up wrong.
1550 */
1551 ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
1552 if (ret) {
1553 btrfs_abort_transaction(trans, ret);
1554 return ret;
1555 }
1556
1557 ret = commit_fs_roots(trans);
1558 if (ret)
1559 goto out;
1560 ret = btrfs_qgroup_account_extents(trans);
1561 if (ret < 0)
1562 goto out;
1563
1564 /* Now qgroup are all updated, we can inherit it to new qgroups */
1565 ret = btrfs_qgroup_inherit(trans, src->root_key.objectid, dst_objectid,
1566 inherit);
1567 if (ret < 0)
1568 goto out;
1569
1570 /*
1571 * Now we do a simplified commit transaction, which will:
1572 * 1) commit all subvolume and extent tree
1573 * To ensure all subvolume and extent tree have a valid
1574 * commit_root to accounting later insert_dir_item()
1575 * 2) write all btree blocks onto disk
1576 * This is to make sure later btree modification will be cowed
1577 * Or commit_root can be populated and cause wrong qgroup numbers
1578 * In this simplified commit, we don't really care about other trees
1579 * like chunk and root tree, as they won't affect qgroup.
1580 * And we don't write super to avoid half committed status.
1581 */
1582 ret = commit_cowonly_roots(trans);
1583 if (ret)
1584 goto out;
1585 switch_commit_roots(trans);
1586 ret = btrfs_write_and_wait_transaction(trans);
1587 if (ret)
1588 btrfs_handle_fs_error(fs_info, ret,
1589 "Error while writing out transaction for qgroup");
1590
1591out:
1592 /*
1593 * Force parent root to be updated, as we recorded it before so its
1594 * last_trans == cur_transid.
1595 * Or it won't be committed again onto disk after later
1596 * insert_dir_item()
1597 */
1598 if (!ret)
1599 ret = record_root_in_trans(trans, parent, 1);
1600 return ret;
1601}
1602
1603/*
1604 * new snapshots need to be created at a very specific time in the
1605 * transaction commit. This does the actual creation.
1606 *
1607 * Note:
1608 * If the error which may affect the commitment of the current transaction
1609 * happens, we should return the error number. If the error which just affect
1610 * the creation of the pending snapshots, just return 0.
1611 */
1612static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
1613 struct btrfs_pending_snapshot *pending)
1614{
1615
1616 struct btrfs_fs_info *fs_info = trans->fs_info;
1617 struct btrfs_key key;
1618 struct btrfs_root_item *new_root_item;
1619 struct btrfs_root *tree_root = fs_info->tree_root;
1620 struct btrfs_root *root = pending->root;
1621 struct btrfs_root *parent_root;
1622 struct btrfs_block_rsv *rsv;
1623 struct inode *parent_inode = pending->dir;
1624 struct btrfs_path *path;
1625 struct btrfs_dir_item *dir_item;
1626 struct extent_buffer *tmp;
1627 struct extent_buffer *old;
1628 struct timespec64 cur_time;
1629 int ret = 0;
1630 u64 to_reserve = 0;
1631 u64 index = 0;
1632 u64 objectid;
1633 u64 root_flags;
1634 unsigned int nofs_flags;
1635 struct fscrypt_name fname;
1636
1637 ASSERT(pending->path);
1638 path = pending->path;
1639
1640 ASSERT(pending->root_item);
1641 new_root_item = pending->root_item;
1642
1643 /*
1644 * We're inside a transaction and must make sure that any potential
1645 * allocations with GFP_KERNEL in fscrypt won't recurse back to
1646 * filesystem.
1647 */
1648 nofs_flags = memalloc_nofs_save();
1649 pending->error = fscrypt_setup_filename(parent_inode,
1650 &pending->dentry->d_name, 0,
1651 &fname);
1652 memalloc_nofs_restore(nofs_flags);
1653 if (pending->error)
1654 goto free_pending;
1655
1656 pending->error = btrfs_get_free_objectid(tree_root, &objectid);
1657 if (pending->error)
1658 goto free_fname;
1659
1660 /*
1661 * Make qgroup to skip current new snapshot's qgroupid, as it is
1662 * accounted by later btrfs_qgroup_inherit().
1663 */
1664 btrfs_set_skip_qgroup(trans, objectid);
1665
1666 btrfs_reloc_pre_snapshot(pending, &to_reserve);
1667
1668 if (to_reserve > 0) {
1669 pending->error = btrfs_block_rsv_add(fs_info,
1670 &pending->block_rsv,
1671 to_reserve,
1672 BTRFS_RESERVE_NO_FLUSH);
1673 if (pending->error)
1674 goto clear_skip_qgroup;
1675 }
1676
1677 key.objectid = objectid;
1678 key.offset = (u64)-1;
1679 key.type = BTRFS_ROOT_ITEM_KEY;
1680
1681 rsv = trans->block_rsv;
1682 trans->block_rsv = &pending->block_rsv;
1683 trans->bytes_reserved = trans->block_rsv->reserved;
1684 trace_btrfs_space_reservation(fs_info, "transaction",
1685 trans->transid,
1686 trans->bytes_reserved, 1);
1687 parent_root = BTRFS_I(parent_inode)->root;
1688 ret = record_root_in_trans(trans, parent_root, 0);
1689 if (ret)
1690 goto fail;
1691 cur_time = current_time(parent_inode);
1692
1693 /*
1694 * insert the directory item
1695 */
1696 ret = btrfs_set_inode_index(BTRFS_I(parent_inode), &index);
1697 BUG_ON(ret); /* -ENOMEM */
1698
1699 /* check if there is a file/dir which has the same name. */
1700 dir_item = btrfs_lookup_dir_item(NULL, parent_root, path,
1701 btrfs_ino(BTRFS_I(parent_inode)),
1702 &fname.disk_name, 0);
1703 if (dir_item != NULL && !IS_ERR(dir_item)) {
1704 pending->error = -EEXIST;
1705 goto dir_item_existed;
1706 } else if (IS_ERR(dir_item)) {
1707 ret = PTR_ERR(dir_item);
1708 btrfs_abort_transaction(trans, ret);
1709 goto fail;
1710 }
1711 btrfs_release_path(path);
1712
1713 /*
1714 * pull in the delayed directory update
1715 * and the delayed inode item
1716 * otherwise we corrupt the FS during
1717 * snapshot
1718 */
1719 ret = btrfs_run_delayed_items(trans);
1720 if (ret) { /* Transaction aborted */
1721 btrfs_abort_transaction(trans, ret);
1722 goto fail;
1723 }
1724
1725 ret = record_root_in_trans(trans, root, 0);
1726 if (ret) {
1727 btrfs_abort_transaction(trans, ret);
1728 goto fail;
1729 }
1730 btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
1731 memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
1732 btrfs_check_and_init_root_item(new_root_item);
1733
1734 root_flags = btrfs_root_flags(new_root_item);
1735 if (pending->readonly)
1736 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
1737 else
1738 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
1739 btrfs_set_root_flags(new_root_item, root_flags);
1740
1741 btrfs_set_root_generation_v2(new_root_item,
1742 trans->transid);
1743 generate_random_guid(new_root_item->uuid);
1744 memcpy(new_root_item->parent_uuid, root->root_item.uuid,
1745 BTRFS_UUID_SIZE);
1746 if (!(root_flags & BTRFS_ROOT_SUBVOL_RDONLY)) {
1747 memset(new_root_item->received_uuid, 0,
1748 sizeof(new_root_item->received_uuid));
1749 memset(&new_root_item->stime, 0, sizeof(new_root_item->stime));
1750 memset(&new_root_item->rtime, 0, sizeof(new_root_item->rtime));
1751 btrfs_set_root_stransid(new_root_item, 0);
1752 btrfs_set_root_rtransid(new_root_item, 0);
1753 }
1754 btrfs_set_stack_timespec_sec(&new_root_item->otime, cur_time.tv_sec);
1755 btrfs_set_stack_timespec_nsec(&new_root_item->otime, cur_time.tv_nsec);
1756 btrfs_set_root_otransid(new_root_item, trans->transid);
1757
1758 old = btrfs_lock_root_node(root);
1759 ret = btrfs_cow_block(trans, root, old, NULL, 0, &old,
1760 BTRFS_NESTING_COW);
1761 if (ret) {
1762 btrfs_tree_unlock(old);
1763 free_extent_buffer(old);
1764 btrfs_abort_transaction(trans, ret);
1765 goto fail;
1766 }
1767
1768 ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1769 /* clean up in any case */
1770 btrfs_tree_unlock(old);
1771 free_extent_buffer(old);
1772 if (ret) {
1773 btrfs_abort_transaction(trans, ret);
1774 goto fail;
1775 }
1776 /* see comments in should_cow_block() */
1777 set_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1778 smp_wmb();
1779
1780 btrfs_set_root_node(new_root_item, tmp);
1781 /* record when the snapshot was created in key.offset */
1782 key.offset = trans->transid;
1783 ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1784 btrfs_tree_unlock(tmp);
1785 free_extent_buffer(tmp);
1786 if (ret) {
1787 btrfs_abort_transaction(trans, ret);
1788 goto fail;
1789 }
1790
1791 /*
1792 * insert root back/forward references
1793 */
1794 ret = btrfs_add_root_ref(trans, objectid,
1795 parent_root->root_key.objectid,
1796 btrfs_ino(BTRFS_I(parent_inode)), index,
1797 &fname.disk_name);
1798 if (ret) {
1799 btrfs_abort_transaction(trans, ret);
1800 goto fail;
1801 }
1802
1803 key.offset = (u64)-1;
1804 pending->snap = btrfs_get_new_fs_root(fs_info, objectid, pending->anon_dev);
1805 if (IS_ERR(pending->snap)) {
1806 ret = PTR_ERR(pending->snap);
1807 pending->snap = NULL;
1808 btrfs_abort_transaction(trans, ret);
1809 goto fail;
1810 }
1811
1812 ret = btrfs_reloc_post_snapshot(trans, pending);
1813 if (ret) {
1814 btrfs_abort_transaction(trans, ret);
1815 goto fail;
1816 }
1817
1818 /*
1819 * Do special qgroup accounting for snapshot, as we do some qgroup
1820 * snapshot hack to do fast snapshot.
1821 * To co-operate with that hack, we do hack again.
1822 * Or snapshot will be greatly slowed down by a subtree qgroup rescan
1823 */
1824 ret = qgroup_account_snapshot(trans, root, parent_root,
1825 pending->inherit, objectid);
1826 if (ret < 0)
1827 goto fail;
1828
1829 ret = btrfs_insert_dir_item(trans, &fname.disk_name,
1830 BTRFS_I(parent_inode), &key, BTRFS_FT_DIR,
1831 index);
1832 /* We have check then name at the beginning, so it is impossible. */
1833 BUG_ON(ret == -EEXIST || ret == -EOVERFLOW);
1834 if (ret) {
1835 btrfs_abort_transaction(trans, ret);
1836 goto fail;
1837 }
1838
1839 btrfs_i_size_write(BTRFS_I(parent_inode), parent_inode->i_size +
1840 fname.disk_name.len * 2);
1841 parent_inode->i_mtime = current_time(parent_inode);
1842 parent_inode->i_ctime = parent_inode->i_mtime;
1843 ret = btrfs_update_inode_fallback(trans, parent_root, BTRFS_I(parent_inode));
1844 if (ret) {
1845 btrfs_abort_transaction(trans, ret);
1846 goto fail;
1847 }
1848 ret = btrfs_uuid_tree_add(trans, new_root_item->uuid,
1849 BTRFS_UUID_KEY_SUBVOL,
1850 objectid);
1851 if (ret) {
1852 btrfs_abort_transaction(trans, ret);
1853 goto fail;
1854 }
1855 if (!btrfs_is_empty_uuid(new_root_item->received_uuid)) {
1856 ret = btrfs_uuid_tree_add(trans, new_root_item->received_uuid,
1857 BTRFS_UUID_KEY_RECEIVED_SUBVOL,
1858 objectid);
1859 if (ret && ret != -EEXIST) {
1860 btrfs_abort_transaction(trans, ret);
1861 goto fail;
1862 }
1863 }
1864
1865fail:
1866 pending->error = ret;
1867dir_item_existed:
1868 trans->block_rsv = rsv;
1869 trans->bytes_reserved = 0;
1870clear_skip_qgroup:
1871 btrfs_clear_skip_qgroup(trans);
1872free_fname:
1873 fscrypt_free_filename(&fname);
1874free_pending:
1875 kfree(new_root_item);
1876 pending->root_item = NULL;
1877 btrfs_free_path(path);
1878 pending->path = NULL;
1879
1880 return ret;
1881}
1882
1883/*
1884 * create all the snapshots we've scheduled for creation
1885 */
1886static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans)
1887{
1888 struct btrfs_pending_snapshot *pending, *next;
1889 struct list_head *head = &trans->transaction->pending_snapshots;
1890 int ret = 0;
1891
1892 list_for_each_entry_safe(pending, next, head, list) {
1893 list_del(&pending->list);
1894 ret = create_pending_snapshot(trans, pending);
1895 if (ret)
1896 break;
1897 }
1898 return ret;
1899}
1900
1901static void update_super_roots(struct btrfs_fs_info *fs_info)
1902{
1903 struct btrfs_root_item *root_item;
1904 struct btrfs_super_block *super;
1905
1906 super = fs_info->super_copy;
1907
1908 root_item = &fs_info->chunk_root->root_item;
1909 super->chunk_root = root_item->bytenr;
1910 super->chunk_root_generation = root_item->generation;
1911 super->chunk_root_level = root_item->level;
1912
1913 root_item = &fs_info->tree_root->root_item;
1914 super->root = root_item->bytenr;
1915 super->generation = root_item->generation;
1916 super->root_level = root_item->level;
1917 if (btrfs_test_opt(fs_info, SPACE_CACHE))
1918 super->cache_generation = root_item->generation;
1919 else if (test_bit(BTRFS_FS_CLEANUP_SPACE_CACHE_V1, &fs_info->flags))
1920 super->cache_generation = 0;
1921 if (test_bit(BTRFS_FS_UPDATE_UUID_TREE_GEN, &fs_info->flags))
1922 super->uuid_tree_generation = root_item->generation;
1923}
1924
1925int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1926{
1927 struct btrfs_transaction *trans;
1928 int ret = 0;
1929
1930 spin_lock(&info->trans_lock);
1931 trans = info->running_transaction;
1932 if (trans)
1933 ret = (trans->state >= TRANS_STATE_COMMIT_START);
1934 spin_unlock(&info->trans_lock);
1935 return ret;
1936}
1937
1938int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1939{
1940 struct btrfs_transaction *trans;
1941 int ret = 0;
1942
1943 spin_lock(&info->trans_lock);
1944 trans = info->running_transaction;
1945 if (trans)
1946 ret = is_transaction_blocked(trans);
1947 spin_unlock(&info->trans_lock);
1948 return ret;
1949}
1950
1951void btrfs_commit_transaction_async(struct btrfs_trans_handle *trans)
1952{
1953 struct btrfs_fs_info *fs_info = trans->fs_info;
1954 struct btrfs_transaction *cur_trans;
1955
1956 /* Kick the transaction kthread. */
1957 set_bit(BTRFS_FS_COMMIT_TRANS, &fs_info->flags);
1958 wake_up_process(fs_info->transaction_kthread);
1959
1960 /* take transaction reference */
1961 cur_trans = trans->transaction;
1962 refcount_inc(&cur_trans->use_count);
1963
1964 btrfs_end_transaction(trans);
1965
1966 /*
1967 * Wait for the current transaction commit to start and block
1968 * subsequent transaction joins
1969 */
1970 btrfs_might_wait_for_state(fs_info, BTRFS_LOCKDEP_TRANS_COMMIT_START);
1971 wait_event(fs_info->transaction_blocked_wait,
1972 cur_trans->state >= TRANS_STATE_COMMIT_START ||
1973 TRANS_ABORTED(cur_trans));
1974 btrfs_put_transaction(cur_trans);
1975}
1976
1977static void cleanup_transaction(struct btrfs_trans_handle *trans, int err)
1978{
1979 struct btrfs_fs_info *fs_info = trans->fs_info;
1980 struct btrfs_transaction *cur_trans = trans->transaction;
1981
1982 WARN_ON(refcount_read(&trans->use_count) > 1);
1983
1984 btrfs_abort_transaction(trans, err);
1985
1986 spin_lock(&fs_info->trans_lock);
1987
1988 /*
1989 * If the transaction is removed from the list, it means this
1990 * transaction has been committed successfully, so it is impossible
1991 * to call the cleanup function.
1992 */
1993 BUG_ON(list_empty(&cur_trans->list));
1994
1995 if (cur_trans == fs_info->running_transaction) {
1996 cur_trans->state = TRANS_STATE_COMMIT_DOING;
1997 spin_unlock(&fs_info->trans_lock);
1998
1999 /*
2000 * The thread has already released the lockdep map as reader
2001 * already in btrfs_commit_transaction().
2002 */
2003 btrfs_might_wait_for_event(fs_info, btrfs_trans_num_writers);
2004 wait_event(cur_trans->writer_wait,
2005 atomic_read(&cur_trans->num_writers) == 1);
2006
2007 spin_lock(&fs_info->trans_lock);
2008 }
2009
2010 /*
2011 * Now that we know no one else is still using the transaction we can
2012 * remove the transaction from the list of transactions. This avoids
2013 * the transaction kthread from cleaning up the transaction while some
2014 * other task is still using it, which could result in a use-after-free
2015 * on things like log trees, as it forces the transaction kthread to
2016 * wait for this transaction to be cleaned up by us.
2017 */
2018 list_del_init(&cur_trans->list);
2019
2020 spin_unlock(&fs_info->trans_lock);
2021
2022 btrfs_cleanup_one_transaction(trans->transaction, fs_info);
2023
2024 spin_lock(&fs_info->trans_lock);
2025 if (cur_trans == fs_info->running_transaction)
2026 fs_info->running_transaction = NULL;
2027 spin_unlock(&fs_info->trans_lock);
2028
2029 if (trans->type & __TRANS_FREEZABLE)
2030 sb_end_intwrite(fs_info->sb);
2031 btrfs_put_transaction(cur_trans);
2032 btrfs_put_transaction(cur_trans);
2033
2034 trace_btrfs_transaction_commit(fs_info);
2035
2036 if (current->journal_info == trans)
2037 current->journal_info = NULL;
2038 btrfs_scrub_cancel(fs_info);
2039
2040 kmem_cache_free(btrfs_trans_handle_cachep, trans);
2041}
2042
2043/*
2044 * Release reserved delayed ref space of all pending block groups of the
2045 * transaction and remove them from the list
2046 */
2047static void btrfs_cleanup_pending_block_groups(struct btrfs_trans_handle *trans)
2048{
2049 struct btrfs_fs_info *fs_info = trans->fs_info;
2050 struct btrfs_block_group *block_group, *tmp;
2051
2052 list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
2053 btrfs_delayed_refs_rsv_release(fs_info, 1);
2054 list_del_init(&block_group->bg_list);
2055 }
2056}
2057
2058static inline int btrfs_start_delalloc_flush(struct btrfs_fs_info *fs_info)
2059{
2060 /*
2061 * We use try_to_writeback_inodes_sb() here because if we used
2062 * btrfs_start_delalloc_roots we would deadlock with fs freeze.
2063 * Currently are holding the fs freeze lock, if we do an async flush
2064 * we'll do btrfs_join_transaction() and deadlock because we need to
2065 * wait for the fs freeze lock. Using the direct flushing we benefit
2066 * from already being in a transaction and our join_transaction doesn't
2067 * have to re-take the fs freeze lock.
2068 *
2069 * Note that try_to_writeback_inodes_sb() will only trigger writeback
2070 * if it can read lock sb->s_umount. It will always be able to lock it,
2071 * except when the filesystem is being unmounted or being frozen, but in
2072 * those cases sync_filesystem() is called, which results in calling
2073 * writeback_inodes_sb() while holding a write lock on sb->s_umount.
2074 * Note that we don't call writeback_inodes_sb() directly, because it
2075 * will emit a warning if sb->s_umount is not locked.
2076 */
2077 if (btrfs_test_opt(fs_info, FLUSHONCOMMIT))
2078 try_to_writeback_inodes_sb(fs_info->sb, WB_REASON_SYNC);
2079 return 0;
2080}
2081
2082static inline void btrfs_wait_delalloc_flush(struct btrfs_fs_info *fs_info)
2083{
2084 if (btrfs_test_opt(fs_info, FLUSHONCOMMIT))
2085 btrfs_wait_ordered_roots(fs_info, U64_MAX, 0, (u64)-1);
2086}
2087
2088/*
2089 * Add a pending snapshot associated with the given transaction handle to the
2090 * respective handle. This must be called after the transaction commit started
2091 * and while holding fs_info->trans_lock.
2092 * This serves to guarantee a caller of btrfs_commit_transaction() that it can
2093 * safely free the pending snapshot pointer in case btrfs_commit_transaction()
2094 * returns an error.
2095 */
2096static void add_pending_snapshot(struct btrfs_trans_handle *trans)
2097{
2098 struct btrfs_transaction *cur_trans = trans->transaction;
2099
2100 if (!trans->pending_snapshot)
2101 return;
2102
2103 lockdep_assert_held(&trans->fs_info->trans_lock);
2104 ASSERT(cur_trans->state >= TRANS_STATE_COMMIT_START);
2105
2106 list_add(&trans->pending_snapshot->list, &cur_trans->pending_snapshots);
2107}
2108
2109static void update_commit_stats(struct btrfs_fs_info *fs_info, ktime_t interval)
2110{
2111 fs_info->commit_stats.commit_count++;
2112 fs_info->commit_stats.last_commit_dur = interval;
2113 fs_info->commit_stats.max_commit_dur =
2114 max_t(u64, fs_info->commit_stats.max_commit_dur, interval);
2115 fs_info->commit_stats.total_commit_dur += interval;
2116}
2117
2118int btrfs_commit_transaction(struct btrfs_trans_handle *trans)
2119{
2120 struct btrfs_fs_info *fs_info = trans->fs_info;
2121 struct btrfs_transaction *cur_trans = trans->transaction;
2122 struct btrfs_transaction *prev_trans = NULL;
2123 int ret;
2124 ktime_t start_time;
2125 ktime_t interval;
2126
2127 ASSERT(refcount_read(&trans->use_count) == 1);
2128 btrfs_trans_state_lockdep_acquire(fs_info, BTRFS_LOCKDEP_TRANS_COMMIT_START);
2129
2130 clear_bit(BTRFS_FS_NEED_TRANS_COMMIT, &fs_info->flags);
2131
2132 /* Stop the commit early if ->aborted is set */
2133 if (TRANS_ABORTED(cur_trans)) {
2134 ret = cur_trans->aborted;
2135 goto lockdep_trans_commit_start_release;
2136 }
2137
2138 btrfs_trans_release_metadata(trans);
2139 trans->block_rsv = NULL;
2140
2141 /*
2142 * We only want one transaction commit doing the flushing so we do not
2143 * waste a bunch of time on lock contention on the extent root node.
2144 */
2145 if (!test_and_set_bit(BTRFS_DELAYED_REFS_FLUSHING,
2146 &cur_trans->delayed_refs.flags)) {
2147 /*
2148 * Make a pass through all the delayed refs we have so far.
2149 * Any running threads may add more while we are here.
2150 */
2151 ret = btrfs_run_delayed_refs(trans, 0);
2152 if (ret)
2153 goto lockdep_trans_commit_start_release;
2154 }
2155
2156 btrfs_create_pending_block_groups(trans);
2157
2158 if (!test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &cur_trans->flags)) {
2159 int run_it = 0;
2160
2161 /* this mutex is also taken before trying to set
2162 * block groups readonly. We need to make sure
2163 * that nobody has set a block group readonly
2164 * after a extents from that block group have been
2165 * allocated for cache files. btrfs_set_block_group_ro
2166 * will wait for the transaction to commit if it
2167 * finds BTRFS_TRANS_DIRTY_BG_RUN set.
2168 *
2169 * The BTRFS_TRANS_DIRTY_BG_RUN flag is also used to make sure
2170 * only one process starts all the block group IO. It wouldn't
2171 * hurt to have more than one go through, but there's no
2172 * real advantage to it either.
2173 */
2174 mutex_lock(&fs_info->ro_block_group_mutex);
2175 if (!test_and_set_bit(BTRFS_TRANS_DIRTY_BG_RUN,
2176 &cur_trans->flags))
2177 run_it = 1;
2178 mutex_unlock(&fs_info->ro_block_group_mutex);
2179
2180 if (run_it) {
2181 ret = btrfs_start_dirty_block_groups(trans);
2182 if (ret)
2183 goto lockdep_trans_commit_start_release;
2184 }
2185 }
2186
2187 spin_lock(&fs_info->trans_lock);
2188 if (cur_trans->state >= TRANS_STATE_COMMIT_START) {
2189 enum btrfs_trans_state want_state = TRANS_STATE_COMPLETED;
2190
2191 add_pending_snapshot(trans);
2192
2193 spin_unlock(&fs_info->trans_lock);
2194 refcount_inc(&cur_trans->use_count);
2195
2196 if (trans->in_fsync)
2197 want_state = TRANS_STATE_SUPER_COMMITTED;
2198
2199 btrfs_trans_state_lockdep_release(fs_info,
2200 BTRFS_LOCKDEP_TRANS_COMMIT_START);
2201 ret = btrfs_end_transaction(trans);
2202 wait_for_commit(cur_trans, want_state);
2203
2204 if (TRANS_ABORTED(cur_trans))
2205 ret = cur_trans->aborted;
2206
2207 btrfs_put_transaction(cur_trans);
2208
2209 return ret;
2210 }
2211
2212 cur_trans->state = TRANS_STATE_COMMIT_START;
2213 wake_up(&fs_info->transaction_blocked_wait);
2214 btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_COMMIT_START);
2215
2216 if (cur_trans->list.prev != &fs_info->trans_list) {
2217 enum btrfs_trans_state want_state = TRANS_STATE_COMPLETED;
2218
2219 if (trans->in_fsync)
2220 want_state = TRANS_STATE_SUPER_COMMITTED;
2221
2222 prev_trans = list_entry(cur_trans->list.prev,
2223 struct btrfs_transaction, list);
2224 if (prev_trans->state < want_state) {
2225 refcount_inc(&prev_trans->use_count);
2226 spin_unlock(&fs_info->trans_lock);
2227
2228 wait_for_commit(prev_trans, want_state);
2229
2230 ret = READ_ONCE(prev_trans->aborted);
2231
2232 btrfs_put_transaction(prev_trans);
2233 if (ret)
2234 goto lockdep_release;
2235 } else {
2236 spin_unlock(&fs_info->trans_lock);
2237 }
2238 } else {
2239 spin_unlock(&fs_info->trans_lock);
2240 /*
2241 * The previous transaction was aborted and was already removed
2242 * from the list of transactions at fs_info->trans_list. So we
2243 * abort to prevent writing a new superblock that reflects a
2244 * corrupt state (pointing to trees with unwritten nodes/leafs).
2245 */
2246 if (BTRFS_FS_ERROR(fs_info)) {
2247 ret = -EROFS;
2248 goto lockdep_release;
2249 }
2250 }
2251
2252 /*
2253 * Get the time spent on the work done by the commit thread and not
2254 * the time spent waiting on a previous commit
2255 */
2256 start_time = ktime_get_ns();
2257
2258 extwriter_counter_dec(cur_trans, trans->type);
2259
2260 ret = btrfs_start_delalloc_flush(fs_info);
2261 if (ret)
2262 goto lockdep_release;
2263
2264 ret = btrfs_run_delayed_items(trans);
2265 if (ret)
2266 goto lockdep_release;
2267
2268 /*
2269 * The thread has started/joined the transaction thus it holds the
2270 * lockdep map as a reader. It has to release it before acquiring the
2271 * lockdep map as a writer.
2272 */
2273 btrfs_lockdep_release(fs_info, btrfs_trans_num_extwriters);
2274 btrfs_might_wait_for_event(fs_info, btrfs_trans_num_extwriters);
2275 wait_event(cur_trans->writer_wait,
2276 extwriter_counter_read(cur_trans) == 0);
2277
2278 /* some pending stuffs might be added after the previous flush. */
2279 ret = btrfs_run_delayed_items(trans);
2280 if (ret) {
2281 btrfs_lockdep_release(fs_info, btrfs_trans_num_writers);
2282 goto cleanup_transaction;
2283 }
2284
2285 btrfs_wait_delalloc_flush(fs_info);
2286
2287 /*
2288 * Wait for all ordered extents started by a fast fsync that joined this
2289 * transaction. Otherwise if this transaction commits before the ordered
2290 * extents complete we lose logged data after a power failure.
2291 */
2292 btrfs_might_wait_for_event(fs_info, btrfs_trans_pending_ordered);
2293 wait_event(cur_trans->pending_wait,
2294 atomic_read(&cur_trans->pending_ordered) == 0);
2295
2296 btrfs_scrub_pause(fs_info);
2297 /*
2298 * Ok now we need to make sure to block out any other joins while we
2299 * commit the transaction. We could have started a join before setting
2300 * COMMIT_DOING so make sure to wait for num_writers to == 1 again.
2301 */
2302 spin_lock(&fs_info->trans_lock);
2303 add_pending_snapshot(trans);
2304 cur_trans->state = TRANS_STATE_COMMIT_DOING;
2305 spin_unlock(&fs_info->trans_lock);
2306
2307 /*
2308 * The thread has started/joined the transaction thus it holds the
2309 * lockdep map as a reader. It has to release it before acquiring the
2310 * lockdep map as a writer.
2311 */
2312 btrfs_lockdep_release(fs_info, btrfs_trans_num_writers);
2313 btrfs_might_wait_for_event(fs_info, btrfs_trans_num_writers);
2314 wait_event(cur_trans->writer_wait,
2315 atomic_read(&cur_trans->num_writers) == 1);
2316
2317 /*
2318 * Make lockdep happy by acquiring the state locks after
2319 * btrfs_trans_num_writers is released. If we acquired the state locks
2320 * before releasing the btrfs_trans_num_writers lock then lockdep would
2321 * complain because we did not follow the reverse order unlocking rule.
2322 */
2323 btrfs_trans_state_lockdep_acquire(fs_info, BTRFS_LOCKDEP_TRANS_COMPLETED);
2324 btrfs_trans_state_lockdep_acquire(fs_info, BTRFS_LOCKDEP_TRANS_SUPER_COMMITTED);
2325 btrfs_trans_state_lockdep_acquire(fs_info, BTRFS_LOCKDEP_TRANS_UNBLOCKED);
2326
2327 /*
2328 * We've started the commit, clear the flag in case we were triggered to
2329 * do an async commit but somebody else started before the transaction
2330 * kthread could do the work.
2331 */
2332 clear_bit(BTRFS_FS_COMMIT_TRANS, &fs_info->flags);
2333
2334 if (TRANS_ABORTED(cur_trans)) {
2335 ret = cur_trans->aborted;
2336 btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_UNBLOCKED);
2337 goto scrub_continue;
2338 }
2339 /*
2340 * the reloc mutex makes sure that we stop
2341 * the balancing code from coming in and moving
2342 * extents around in the middle of the commit
2343 */
2344 mutex_lock(&fs_info->reloc_mutex);
2345
2346 /*
2347 * We needn't worry about the delayed items because we will
2348 * deal with them in create_pending_snapshot(), which is the
2349 * core function of the snapshot creation.
2350 */
2351 ret = create_pending_snapshots(trans);
2352 if (ret)
2353 goto unlock_reloc;
2354
2355 /*
2356 * We insert the dir indexes of the snapshots and update the inode
2357 * of the snapshots' parents after the snapshot creation, so there
2358 * are some delayed items which are not dealt with. Now deal with
2359 * them.
2360 *
2361 * We needn't worry that this operation will corrupt the snapshots,
2362 * because all the tree which are snapshoted will be forced to COW
2363 * the nodes and leaves.
2364 */
2365 ret = btrfs_run_delayed_items(trans);
2366 if (ret)
2367 goto unlock_reloc;
2368
2369 ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
2370 if (ret)
2371 goto unlock_reloc;
2372
2373 /*
2374 * make sure none of the code above managed to slip in a
2375 * delayed item
2376 */
2377 btrfs_assert_delayed_root_empty(fs_info);
2378
2379 WARN_ON(cur_trans != trans->transaction);
2380
2381 ret = commit_fs_roots(trans);
2382 if (ret)
2383 goto unlock_reloc;
2384
2385 /* commit_fs_roots gets rid of all the tree log roots, it is now
2386 * safe to free the root of tree log roots
2387 */
2388 btrfs_free_log_root_tree(trans, fs_info);
2389
2390 /*
2391 * Since fs roots are all committed, we can get a quite accurate
2392 * new_roots. So let's do quota accounting.
2393 */
2394 ret = btrfs_qgroup_account_extents(trans);
2395 if (ret < 0)
2396 goto unlock_reloc;
2397
2398 ret = commit_cowonly_roots(trans);
2399 if (ret)
2400 goto unlock_reloc;
2401
2402 /*
2403 * The tasks which save the space cache and inode cache may also
2404 * update ->aborted, check it.
2405 */
2406 if (TRANS_ABORTED(cur_trans)) {
2407 ret = cur_trans->aborted;
2408 goto unlock_reloc;
2409 }
2410
2411 cur_trans = fs_info->running_transaction;
2412
2413 btrfs_set_root_node(&fs_info->tree_root->root_item,
2414 fs_info->tree_root->node);
2415 list_add_tail(&fs_info->tree_root->dirty_list,
2416 &cur_trans->switch_commits);
2417
2418 btrfs_set_root_node(&fs_info->chunk_root->root_item,
2419 fs_info->chunk_root->node);
2420 list_add_tail(&fs_info->chunk_root->dirty_list,
2421 &cur_trans->switch_commits);
2422
2423 if (btrfs_fs_incompat(fs_info, EXTENT_TREE_V2)) {
2424 btrfs_set_root_node(&fs_info->block_group_root->root_item,
2425 fs_info->block_group_root->node);
2426 list_add_tail(&fs_info->block_group_root->dirty_list,
2427 &cur_trans->switch_commits);
2428 }
2429
2430 switch_commit_roots(trans);
2431
2432 ASSERT(list_empty(&cur_trans->dirty_bgs));
2433 ASSERT(list_empty(&cur_trans->io_bgs));
2434 update_super_roots(fs_info);
2435
2436 btrfs_set_super_log_root(fs_info->super_copy, 0);
2437 btrfs_set_super_log_root_level(fs_info->super_copy, 0);
2438 memcpy(fs_info->super_for_commit, fs_info->super_copy,
2439 sizeof(*fs_info->super_copy));
2440
2441 btrfs_commit_device_sizes(cur_trans);
2442
2443 clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags);
2444 clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags);
2445
2446 btrfs_trans_release_chunk_metadata(trans);
2447
2448 /*
2449 * Before changing the transaction state to TRANS_STATE_UNBLOCKED and
2450 * setting fs_info->running_transaction to NULL, lock tree_log_mutex to
2451 * make sure that before we commit our superblock, no other task can
2452 * start a new transaction and commit a log tree before we commit our
2453 * superblock. Anyone trying to commit a log tree locks this mutex before
2454 * writing its superblock.
2455 */
2456 mutex_lock(&fs_info->tree_log_mutex);
2457
2458 spin_lock(&fs_info->trans_lock);
2459 cur_trans->state = TRANS_STATE_UNBLOCKED;
2460 fs_info->running_transaction = NULL;
2461 spin_unlock(&fs_info->trans_lock);
2462 mutex_unlock(&fs_info->reloc_mutex);
2463
2464 wake_up(&fs_info->transaction_wait);
2465 btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_UNBLOCKED);
2466
2467 ret = btrfs_write_and_wait_transaction(trans);
2468 if (ret) {
2469 btrfs_handle_fs_error(fs_info, ret,
2470 "Error while writing out transaction");
2471 mutex_unlock(&fs_info->tree_log_mutex);
2472 goto scrub_continue;
2473 }
2474
2475 /*
2476 * At this point, we should have written all the tree blocks allocated
2477 * in this transaction. So it's now safe to free the redirtyied extent
2478 * buffers.
2479 */
2480 btrfs_free_redirty_list(cur_trans);
2481
2482 ret = write_all_supers(fs_info, 0);
2483 /*
2484 * the super is written, we can safely allow the tree-loggers
2485 * to go about their business
2486 */
2487 mutex_unlock(&fs_info->tree_log_mutex);
2488 if (ret)
2489 goto scrub_continue;
2490
2491 /*
2492 * We needn't acquire the lock here because there is no other task
2493 * which can change it.
2494 */
2495 cur_trans->state = TRANS_STATE_SUPER_COMMITTED;
2496 wake_up(&cur_trans->commit_wait);
2497 btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_SUPER_COMMITTED);
2498
2499 btrfs_finish_extent_commit(trans);
2500
2501 if (test_bit(BTRFS_TRANS_HAVE_FREE_BGS, &cur_trans->flags))
2502 btrfs_clear_space_info_full(fs_info);
2503
2504 fs_info->last_trans_committed = cur_trans->transid;
2505 /*
2506 * We needn't acquire the lock here because there is no other task
2507 * which can change it.
2508 */
2509 cur_trans->state = TRANS_STATE_COMPLETED;
2510 wake_up(&cur_trans->commit_wait);
2511 btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_COMPLETED);
2512
2513 spin_lock(&fs_info->trans_lock);
2514 list_del_init(&cur_trans->list);
2515 spin_unlock(&fs_info->trans_lock);
2516
2517 btrfs_put_transaction(cur_trans);
2518 btrfs_put_transaction(cur_trans);
2519
2520 if (trans->type & __TRANS_FREEZABLE)
2521 sb_end_intwrite(fs_info->sb);
2522
2523 trace_btrfs_transaction_commit(fs_info);
2524
2525 interval = ktime_get_ns() - start_time;
2526
2527 btrfs_scrub_continue(fs_info);
2528
2529 if (current->journal_info == trans)
2530 current->journal_info = NULL;
2531
2532 kmem_cache_free(btrfs_trans_handle_cachep, trans);
2533
2534 update_commit_stats(fs_info, interval);
2535
2536 return ret;
2537
2538unlock_reloc:
2539 mutex_unlock(&fs_info->reloc_mutex);
2540 btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_UNBLOCKED);
2541scrub_continue:
2542 btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_SUPER_COMMITTED);
2543 btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_COMPLETED);
2544 btrfs_scrub_continue(fs_info);
2545cleanup_transaction:
2546 btrfs_trans_release_metadata(trans);
2547 btrfs_cleanup_pending_block_groups(trans);
2548 btrfs_trans_release_chunk_metadata(trans);
2549 trans->block_rsv = NULL;
2550 btrfs_warn(fs_info, "Skipping commit of aborted transaction.");
2551 if (current->journal_info == trans)
2552 current->journal_info = NULL;
2553 cleanup_transaction(trans, ret);
2554
2555 return ret;
2556
2557lockdep_release:
2558 btrfs_lockdep_release(fs_info, btrfs_trans_num_extwriters);
2559 btrfs_lockdep_release(fs_info, btrfs_trans_num_writers);
2560 goto cleanup_transaction;
2561
2562lockdep_trans_commit_start_release:
2563 btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_COMMIT_START);
2564 btrfs_end_transaction(trans);
2565 return ret;
2566}
2567
2568/*
2569 * return < 0 if error
2570 * 0 if there are no more dead_roots at the time of call
2571 * 1 there are more to be processed, call me again
2572 *
2573 * The return value indicates there are certainly more snapshots to delete, but
2574 * if there comes a new one during processing, it may return 0. We don't mind,
2575 * because btrfs_commit_super will poke cleaner thread and it will process it a
2576 * few seconds later.
2577 */
2578int btrfs_clean_one_deleted_snapshot(struct btrfs_fs_info *fs_info)
2579{
2580 struct btrfs_root *root;
2581 int ret;
2582
2583 spin_lock(&fs_info->trans_lock);
2584 if (list_empty(&fs_info->dead_roots)) {
2585 spin_unlock(&fs_info->trans_lock);
2586 return 0;
2587 }
2588 root = list_first_entry(&fs_info->dead_roots,
2589 struct btrfs_root, root_list);
2590 list_del_init(&root->root_list);
2591 spin_unlock(&fs_info->trans_lock);
2592
2593 btrfs_debug(fs_info, "cleaner removing %llu", root->root_key.objectid);
2594
2595 btrfs_kill_all_delayed_nodes(root);
2596
2597 if (btrfs_header_backref_rev(root->node) <
2598 BTRFS_MIXED_BACKREF_REV)
2599 ret = btrfs_drop_snapshot(root, 0, 0);
2600 else
2601 ret = btrfs_drop_snapshot(root, 1, 0);
2602
2603 btrfs_put_root(root);
2604 return (ret < 0) ? 0 : 1;
2605}
2606
2607int __init btrfs_transaction_init(void)
2608{
2609 btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
2610 sizeof(struct btrfs_trans_handle), 0,
2611 SLAB_TEMPORARY | SLAB_MEM_SPREAD, NULL);
2612 if (!btrfs_trans_handle_cachep)
2613 return -ENOMEM;
2614 return 0;
2615}
2616
2617void __cold btrfs_transaction_exit(void)
2618{
2619 kmem_cache_destroy(btrfs_trans_handle_cachep);
2620}
1/*
2 * Copyright (C) 2007 Oracle. All rights reserved.
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public
6 * License v2 as published by the Free Software Foundation.
7 *
8 * This program is distributed in the hope that it will be useful,
9 * but WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
11 * General Public License for more details.
12 *
13 * You should have received a copy of the GNU General Public
14 * License along with this program; if not, write to the
15 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16 * Boston, MA 021110-1307, USA.
17 */
18
19#include <linux/fs.h>
20#include <linux/slab.h>
21#include <linux/sched.h>
22#include <linux/writeback.h>
23#include <linux/pagemap.h>
24#include <linux/blkdev.h>
25#include <linux/uuid.h>
26#include "ctree.h"
27#include "disk-io.h"
28#include "transaction.h"
29#include "locking.h"
30#include "tree-log.h"
31#include "inode-map.h"
32#include "volumes.h"
33#include "dev-replace.h"
34#include "qgroup.h"
35
36#define BTRFS_ROOT_TRANS_TAG 0
37
38static const unsigned int btrfs_blocked_trans_types[TRANS_STATE_MAX] = {
39 [TRANS_STATE_RUNNING] = 0U,
40 [TRANS_STATE_BLOCKED] = (__TRANS_USERSPACE |
41 __TRANS_START),
42 [TRANS_STATE_COMMIT_START] = (__TRANS_USERSPACE |
43 __TRANS_START |
44 __TRANS_ATTACH),
45 [TRANS_STATE_COMMIT_DOING] = (__TRANS_USERSPACE |
46 __TRANS_START |
47 __TRANS_ATTACH |
48 __TRANS_JOIN),
49 [TRANS_STATE_UNBLOCKED] = (__TRANS_USERSPACE |
50 __TRANS_START |
51 __TRANS_ATTACH |
52 __TRANS_JOIN |
53 __TRANS_JOIN_NOLOCK),
54 [TRANS_STATE_COMPLETED] = (__TRANS_USERSPACE |
55 __TRANS_START |
56 __TRANS_ATTACH |
57 __TRANS_JOIN |
58 __TRANS_JOIN_NOLOCK),
59};
60
61void btrfs_put_transaction(struct btrfs_transaction *transaction)
62{
63 WARN_ON(atomic_read(&transaction->use_count) == 0);
64 if (atomic_dec_and_test(&transaction->use_count)) {
65 BUG_ON(!list_empty(&transaction->list));
66 WARN_ON(!RB_EMPTY_ROOT(&transaction->delayed_refs.href_root));
67 if (transaction->delayed_refs.pending_csums)
68 btrfs_err(transaction->fs_info,
69 "pending csums is %llu",
70 transaction->delayed_refs.pending_csums);
71 while (!list_empty(&transaction->pending_chunks)) {
72 struct extent_map *em;
73
74 em = list_first_entry(&transaction->pending_chunks,
75 struct extent_map, list);
76 list_del_init(&em->list);
77 free_extent_map(em);
78 }
79 /*
80 * If any block groups are found in ->deleted_bgs then it's
81 * because the transaction was aborted and a commit did not
82 * happen (things failed before writing the new superblock
83 * and calling btrfs_finish_extent_commit()), so we can not
84 * discard the physical locations of the block groups.
85 */
86 while (!list_empty(&transaction->deleted_bgs)) {
87 struct btrfs_block_group_cache *cache;
88
89 cache = list_first_entry(&transaction->deleted_bgs,
90 struct btrfs_block_group_cache,
91 bg_list);
92 list_del_init(&cache->bg_list);
93 btrfs_put_block_group_trimming(cache);
94 btrfs_put_block_group(cache);
95 }
96 kmem_cache_free(btrfs_transaction_cachep, transaction);
97 }
98}
99
100static void clear_btree_io_tree(struct extent_io_tree *tree)
101{
102 spin_lock(&tree->lock);
103 /*
104 * Do a single barrier for the waitqueue_active check here, the state
105 * of the waitqueue should not change once clear_btree_io_tree is
106 * called.
107 */
108 smp_mb();
109 while (!RB_EMPTY_ROOT(&tree->state)) {
110 struct rb_node *node;
111 struct extent_state *state;
112
113 node = rb_first(&tree->state);
114 state = rb_entry(node, struct extent_state, rb_node);
115 rb_erase(&state->rb_node, &tree->state);
116 RB_CLEAR_NODE(&state->rb_node);
117 /*
118 * btree io trees aren't supposed to have tasks waiting for
119 * changes in the flags of extent states ever.
120 */
121 ASSERT(!waitqueue_active(&state->wq));
122 free_extent_state(state);
123
124 cond_resched_lock(&tree->lock);
125 }
126 spin_unlock(&tree->lock);
127}
128
129static noinline void switch_commit_roots(struct btrfs_transaction *trans,
130 struct btrfs_fs_info *fs_info)
131{
132 struct btrfs_root *root, *tmp;
133
134 down_write(&fs_info->commit_root_sem);
135 list_for_each_entry_safe(root, tmp, &trans->switch_commits,
136 dirty_list) {
137 list_del_init(&root->dirty_list);
138 free_extent_buffer(root->commit_root);
139 root->commit_root = btrfs_root_node(root);
140 if (is_fstree(root->objectid))
141 btrfs_unpin_free_ino(root);
142 clear_btree_io_tree(&root->dirty_log_pages);
143 }
144
145 /* We can free old roots now. */
146 spin_lock(&trans->dropped_roots_lock);
147 while (!list_empty(&trans->dropped_roots)) {
148 root = list_first_entry(&trans->dropped_roots,
149 struct btrfs_root, root_list);
150 list_del_init(&root->root_list);
151 spin_unlock(&trans->dropped_roots_lock);
152 btrfs_drop_and_free_fs_root(fs_info, root);
153 spin_lock(&trans->dropped_roots_lock);
154 }
155 spin_unlock(&trans->dropped_roots_lock);
156 up_write(&fs_info->commit_root_sem);
157}
158
159static inline void extwriter_counter_inc(struct btrfs_transaction *trans,
160 unsigned int type)
161{
162 if (type & TRANS_EXTWRITERS)
163 atomic_inc(&trans->num_extwriters);
164}
165
166static inline void extwriter_counter_dec(struct btrfs_transaction *trans,
167 unsigned int type)
168{
169 if (type & TRANS_EXTWRITERS)
170 atomic_dec(&trans->num_extwriters);
171}
172
173static inline void extwriter_counter_init(struct btrfs_transaction *trans,
174 unsigned int type)
175{
176 atomic_set(&trans->num_extwriters, ((type & TRANS_EXTWRITERS) ? 1 : 0));
177}
178
179static inline int extwriter_counter_read(struct btrfs_transaction *trans)
180{
181 return atomic_read(&trans->num_extwriters);
182}
183
184/*
185 * either allocate a new transaction or hop into the existing one
186 */
187static noinline int join_transaction(struct btrfs_fs_info *fs_info,
188 unsigned int type)
189{
190 struct btrfs_transaction *cur_trans;
191
192 spin_lock(&fs_info->trans_lock);
193loop:
194 /* The file system has been taken offline. No new transactions. */
195 if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
196 spin_unlock(&fs_info->trans_lock);
197 return -EROFS;
198 }
199
200 cur_trans = fs_info->running_transaction;
201 if (cur_trans) {
202 if (cur_trans->aborted) {
203 spin_unlock(&fs_info->trans_lock);
204 return cur_trans->aborted;
205 }
206 if (btrfs_blocked_trans_types[cur_trans->state] & type) {
207 spin_unlock(&fs_info->trans_lock);
208 return -EBUSY;
209 }
210 atomic_inc(&cur_trans->use_count);
211 atomic_inc(&cur_trans->num_writers);
212 extwriter_counter_inc(cur_trans, type);
213 spin_unlock(&fs_info->trans_lock);
214 return 0;
215 }
216 spin_unlock(&fs_info->trans_lock);
217
218 /*
219 * If we are ATTACH, we just want to catch the current transaction,
220 * and commit it. If there is no transaction, just return ENOENT.
221 */
222 if (type == TRANS_ATTACH)
223 return -ENOENT;
224
225 /*
226 * JOIN_NOLOCK only happens during the transaction commit, so
227 * it is impossible that ->running_transaction is NULL
228 */
229 BUG_ON(type == TRANS_JOIN_NOLOCK);
230
231 cur_trans = kmem_cache_alloc(btrfs_transaction_cachep, GFP_NOFS);
232 if (!cur_trans)
233 return -ENOMEM;
234
235 spin_lock(&fs_info->trans_lock);
236 if (fs_info->running_transaction) {
237 /*
238 * someone started a transaction after we unlocked. Make sure
239 * to redo the checks above
240 */
241 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
242 goto loop;
243 } else if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
244 spin_unlock(&fs_info->trans_lock);
245 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
246 return -EROFS;
247 }
248
249 cur_trans->fs_info = fs_info;
250 atomic_set(&cur_trans->num_writers, 1);
251 extwriter_counter_init(cur_trans, type);
252 init_waitqueue_head(&cur_trans->writer_wait);
253 init_waitqueue_head(&cur_trans->commit_wait);
254 init_waitqueue_head(&cur_trans->pending_wait);
255 cur_trans->state = TRANS_STATE_RUNNING;
256 /*
257 * One for this trans handle, one so it will live on until we
258 * commit the transaction.
259 */
260 atomic_set(&cur_trans->use_count, 2);
261 atomic_set(&cur_trans->pending_ordered, 0);
262 cur_trans->flags = 0;
263 cur_trans->start_time = get_seconds();
264
265 memset(&cur_trans->delayed_refs, 0, sizeof(cur_trans->delayed_refs));
266
267 cur_trans->delayed_refs.href_root = RB_ROOT;
268 cur_trans->delayed_refs.dirty_extent_root = RB_ROOT;
269 atomic_set(&cur_trans->delayed_refs.num_entries, 0);
270
271 /*
272 * although the tree mod log is per file system and not per transaction,
273 * the log must never go across transaction boundaries.
274 */
275 smp_mb();
276 if (!list_empty(&fs_info->tree_mod_seq_list))
277 WARN(1, KERN_ERR "BTRFS: tree_mod_seq_list not empty when creating a fresh transaction\n");
278 if (!RB_EMPTY_ROOT(&fs_info->tree_mod_log))
279 WARN(1, KERN_ERR "BTRFS: tree_mod_log rb tree not empty when creating a fresh transaction\n");
280 atomic64_set(&fs_info->tree_mod_seq, 0);
281
282 spin_lock_init(&cur_trans->delayed_refs.lock);
283
284 INIT_LIST_HEAD(&cur_trans->pending_snapshots);
285 INIT_LIST_HEAD(&cur_trans->pending_chunks);
286 INIT_LIST_HEAD(&cur_trans->switch_commits);
287 INIT_LIST_HEAD(&cur_trans->dirty_bgs);
288 INIT_LIST_HEAD(&cur_trans->io_bgs);
289 INIT_LIST_HEAD(&cur_trans->dropped_roots);
290 mutex_init(&cur_trans->cache_write_mutex);
291 cur_trans->num_dirty_bgs = 0;
292 spin_lock_init(&cur_trans->dirty_bgs_lock);
293 INIT_LIST_HEAD(&cur_trans->deleted_bgs);
294 spin_lock_init(&cur_trans->dropped_roots_lock);
295 list_add_tail(&cur_trans->list, &fs_info->trans_list);
296 extent_io_tree_init(&cur_trans->dirty_pages,
297 fs_info->btree_inode->i_mapping);
298 fs_info->generation++;
299 cur_trans->transid = fs_info->generation;
300 fs_info->running_transaction = cur_trans;
301 cur_trans->aborted = 0;
302 spin_unlock(&fs_info->trans_lock);
303
304 return 0;
305}
306
307/*
308 * this does all the record keeping required to make sure that a reference
309 * counted root is properly recorded in a given transaction. This is required
310 * to make sure the old root from before we joined the transaction is deleted
311 * when the transaction commits
312 */
313static int record_root_in_trans(struct btrfs_trans_handle *trans,
314 struct btrfs_root *root,
315 int force)
316{
317 struct btrfs_fs_info *fs_info = root->fs_info;
318
319 if ((test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
320 root->last_trans < trans->transid) || force) {
321 WARN_ON(root == fs_info->extent_root);
322 WARN_ON(root->commit_root != root->node);
323
324 /*
325 * see below for IN_TRANS_SETUP usage rules
326 * we have the reloc mutex held now, so there
327 * is only one writer in this function
328 */
329 set_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
330
331 /* make sure readers find IN_TRANS_SETUP before
332 * they find our root->last_trans update
333 */
334 smp_wmb();
335
336 spin_lock(&fs_info->fs_roots_radix_lock);
337 if (root->last_trans == trans->transid && !force) {
338 spin_unlock(&fs_info->fs_roots_radix_lock);
339 return 0;
340 }
341 radix_tree_tag_set(&fs_info->fs_roots_radix,
342 (unsigned long)root->root_key.objectid,
343 BTRFS_ROOT_TRANS_TAG);
344 spin_unlock(&fs_info->fs_roots_radix_lock);
345 root->last_trans = trans->transid;
346
347 /* this is pretty tricky. We don't want to
348 * take the relocation lock in btrfs_record_root_in_trans
349 * unless we're really doing the first setup for this root in
350 * this transaction.
351 *
352 * Normally we'd use root->last_trans as a flag to decide
353 * if we want to take the expensive mutex.
354 *
355 * But, we have to set root->last_trans before we
356 * init the relocation root, otherwise, we trip over warnings
357 * in ctree.c. The solution used here is to flag ourselves
358 * with root IN_TRANS_SETUP. When this is 1, we're still
359 * fixing up the reloc trees and everyone must wait.
360 *
361 * When this is zero, they can trust root->last_trans and fly
362 * through btrfs_record_root_in_trans without having to take the
363 * lock. smp_wmb() makes sure that all the writes above are
364 * done before we pop in the zero below
365 */
366 btrfs_init_reloc_root(trans, root);
367 smp_mb__before_atomic();
368 clear_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
369 }
370 return 0;
371}
372
373
374void btrfs_add_dropped_root(struct btrfs_trans_handle *trans,
375 struct btrfs_root *root)
376{
377 struct btrfs_fs_info *fs_info = root->fs_info;
378 struct btrfs_transaction *cur_trans = trans->transaction;
379
380 /* Add ourselves to the transaction dropped list */
381 spin_lock(&cur_trans->dropped_roots_lock);
382 list_add_tail(&root->root_list, &cur_trans->dropped_roots);
383 spin_unlock(&cur_trans->dropped_roots_lock);
384
385 /* Make sure we don't try to update the root at commit time */
386 spin_lock(&fs_info->fs_roots_radix_lock);
387 radix_tree_tag_clear(&fs_info->fs_roots_radix,
388 (unsigned long)root->root_key.objectid,
389 BTRFS_ROOT_TRANS_TAG);
390 spin_unlock(&fs_info->fs_roots_radix_lock);
391}
392
393int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
394 struct btrfs_root *root)
395{
396 struct btrfs_fs_info *fs_info = root->fs_info;
397
398 if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
399 return 0;
400
401 /*
402 * see record_root_in_trans for comments about IN_TRANS_SETUP usage
403 * and barriers
404 */
405 smp_rmb();
406 if (root->last_trans == trans->transid &&
407 !test_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state))
408 return 0;
409
410 mutex_lock(&fs_info->reloc_mutex);
411 record_root_in_trans(trans, root, 0);
412 mutex_unlock(&fs_info->reloc_mutex);
413
414 return 0;
415}
416
417static inline int is_transaction_blocked(struct btrfs_transaction *trans)
418{
419 return (trans->state >= TRANS_STATE_BLOCKED &&
420 trans->state < TRANS_STATE_UNBLOCKED &&
421 !trans->aborted);
422}
423
424/* wait for commit against the current transaction to become unblocked
425 * when this is done, it is safe to start a new transaction, but the current
426 * transaction might not be fully on disk.
427 */
428static void wait_current_trans(struct btrfs_fs_info *fs_info)
429{
430 struct btrfs_transaction *cur_trans;
431
432 spin_lock(&fs_info->trans_lock);
433 cur_trans = fs_info->running_transaction;
434 if (cur_trans && is_transaction_blocked(cur_trans)) {
435 atomic_inc(&cur_trans->use_count);
436 spin_unlock(&fs_info->trans_lock);
437
438 wait_event(fs_info->transaction_wait,
439 cur_trans->state >= TRANS_STATE_UNBLOCKED ||
440 cur_trans->aborted);
441 btrfs_put_transaction(cur_trans);
442 } else {
443 spin_unlock(&fs_info->trans_lock);
444 }
445}
446
447static int may_wait_transaction(struct btrfs_fs_info *fs_info, int type)
448{
449 if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
450 return 0;
451
452 if (type == TRANS_USERSPACE)
453 return 1;
454
455 if (type == TRANS_START &&
456 !atomic_read(&fs_info->open_ioctl_trans))
457 return 1;
458
459 return 0;
460}
461
462static inline bool need_reserve_reloc_root(struct btrfs_root *root)
463{
464 struct btrfs_fs_info *fs_info = root->fs_info;
465
466 if (!fs_info->reloc_ctl ||
467 !test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
468 root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
469 root->reloc_root)
470 return false;
471
472 return true;
473}
474
475static struct btrfs_trans_handle *
476start_transaction(struct btrfs_root *root, unsigned int num_items,
477 unsigned int type, enum btrfs_reserve_flush_enum flush)
478{
479 struct btrfs_fs_info *fs_info = root->fs_info;
480
481 struct btrfs_trans_handle *h;
482 struct btrfs_transaction *cur_trans;
483 u64 num_bytes = 0;
484 u64 qgroup_reserved = 0;
485 bool reloc_reserved = false;
486 int ret;
487
488 /* Send isn't supposed to start transactions. */
489 ASSERT(current->journal_info != BTRFS_SEND_TRANS_STUB);
490
491 if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
492 return ERR_PTR(-EROFS);
493
494 if (current->journal_info) {
495 WARN_ON(type & TRANS_EXTWRITERS);
496 h = current->journal_info;
497 h->use_count++;
498 WARN_ON(h->use_count > 2);
499 h->orig_rsv = h->block_rsv;
500 h->block_rsv = NULL;
501 goto got_it;
502 }
503
504 /*
505 * Do the reservation before we join the transaction so we can do all
506 * the appropriate flushing if need be.
507 */
508 if (num_items > 0 && root != fs_info->chunk_root) {
509 qgroup_reserved = num_items * fs_info->nodesize;
510 ret = btrfs_qgroup_reserve_meta(root, qgroup_reserved);
511 if (ret)
512 return ERR_PTR(ret);
513
514 num_bytes = btrfs_calc_trans_metadata_size(fs_info, num_items);
515 /*
516 * Do the reservation for the relocation root creation
517 */
518 if (need_reserve_reloc_root(root)) {
519 num_bytes += fs_info->nodesize;
520 reloc_reserved = true;
521 }
522
523 ret = btrfs_block_rsv_add(root, &fs_info->trans_block_rsv,
524 num_bytes, flush);
525 if (ret)
526 goto reserve_fail;
527 }
528again:
529 h = kmem_cache_zalloc(btrfs_trans_handle_cachep, GFP_NOFS);
530 if (!h) {
531 ret = -ENOMEM;
532 goto alloc_fail;
533 }
534
535 /*
536 * If we are JOIN_NOLOCK we're already committing a transaction and
537 * waiting on this guy, so we don't need to do the sb_start_intwrite
538 * because we're already holding a ref. We need this because we could
539 * have raced in and did an fsync() on a file which can kick a commit
540 * and then we deadlock with somebody doing a freeze.
541 *
542 * If we are ATTACH, it means we just want to catch the current
543 * transaction and commit it, so we needn't do sb_start_intwrite().
544 */
545 if (type & __TRANS_FREEZABLE)
546 sb_start_intwrite(fs_info->sb);
547
548 if (may_wait_transaction(fs_info, type))
549 wait_current_trans(fs_info);
550
551 do {
552 ret = join_transaction(fs_info, type);
553 if (ret == -EBUSY) {
554 wait_current_trans(fs_info);
555 if (unlikely(type == TRANS_ATTACH))
556 ret = -ENOENT;
557 }
558 } while (ret == -EBUSY);
559
560 if (ret < 0)
561 goto join_fail;
562
563 cur_trans = fs_info->running_transaction;
564
565 h->transid = cur_trans->transid;
566 h->transaction = cur_trans;
567 h->root = root;
568 h->use_count = 1;
569 h->fs_info = root->fs_info;
570
571 h->type = type;
572 h->can_flush_pending_bgs = true;
573 INIT_LIST_HEAD(&h->qgroup_ref_list);
574 INIT_LIST_HEAD(&h->new_bgs);
575
576 smp_mb();
577 if (cur_trans->state >= TRANS_STATE_BLOCKED &&
578 may_wait_transaction(fs_info, type)) {
579 current->journal_info = h;
580 btrfs_commit_transaction(h);
581 goto again;
582 }
583
584 if (num_bytes) {
585 trace_btrfs_space_reservation(fs_info, "transaction",
586 h->transid, num_bytes, 1);
587 h->block_rsv = &fs_info->trans_block_rsv;
588 h->bytes_reserved = num_bytes;
589 h->reloc_reserved = reloc_reserved;
590 }
591
592got_it:
593 btrfs_record_root_in_trans(h, root);
594
595 if (!current->journal_info && type != TRANS_USERSPACE)
596 current->journal_info = h;
597 return h;
598
599join_fail:
600 if (type & __TRANS_FREEZABLE)
601 sb_end_intwrite(fs_info->sb);
602 kmem_cache_free(btrfs_trans_handle_cachep, h);
603alloc_fail:
604 if (num_bytes)
605 btrfs_block_rsv_release(fs_info, &fs_info->trans_block_rsv,
606 num_bytes);
607reserve_fail:
608 btrfs_qgroup_free_meta(root, qgroup_reserved);
609 return ERR_PTR(ret);
610}
611
612struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
613 unsigned int num_items)
614{
615 return start_transaction(root, num_items, TRANS_START,
616 BTRFS_RESERVE_FLUSH_ALL);
617}
618struct btrfs_trans_handle *btrfs_start_transaction_fallback_global_rsv(
619 struct btrfs_root *root,
620 unsigned int num_items,
621 int min_factor)
622{
623 struct btrfs_fs_info *fs_info = root->fs_info;
624 struct btrfs_trans_handle *trans;
625 u64 num_bytes;
626 int ret;
627
628 trans = btrfs_start_transaction(root, num_items);
629 if (!IS_ERR(trans) || PTR_ERR(trans) != -ENOSPC)
630 return trans;
631
632 trans = btrfs_start_transaction(root, 0);
633 if (IS_ERR(trans))
634 return trans;
635
636 num_bytes = btrfs_calc_trans_metadata_size(fs_info, num_items);
637 ret = btrfs_cond_migrate_bytes(fs_info, &fs_info->trans_block_rsv,
638 num_bytes, min_factor);
639 if (ret) {
640 btrfs_end_transaction(trans);
641 return ERR_PTR(ret);
642 }
643
644 trans->block_rsv = &fs_info->trans_block_rsv;
645 trans->bytes_reserved = num_bytes;
646 trace_btrfs_space_reservation(fs_info, "transaction",
647 trans->transid, num_bytes, 1);
648
649 return trans;
650}
651
652struct btrfs_trans_handle *btrfs_start_transaction_lflush(
653 struct btrfs_root *root,
654 unsigned int num_items)
655{
656 return start_transaction(root, num_items, TRANS_START,
657 BTRFS_RESERVE_FLUSH_LIMIT);
658}
659
660struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
661{
662 return start_transaction(root, 0, TRANS_JOIN,
663 BTRFS_RESERVE_NO_FLUSH);
664}
665
666struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root)
667{
668 return start_transaction(root, 0, TRANS_JOIN_NOLOCK,
669 BTRFS_RESERVE_NO_FLUSH);
670}
671
672struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *root)
673{
674 return start_transaction(root, 0, TRANS_USERSPACE,
675 BTRFS_RESERVE_NO_FLUSH);
676}
677
678/*
679 * btrfs_attach_transaction() - catch the running transaction
680 *
681 * It is used when we want to commit the current the transaction, but
682 * don't want to start a new one.
683 *
684 * Note: If this function return -ENOENT, it just means there is no
685 * running transaction. But it is possible that the inactive transaction
686 * is still in the memory, not fully on disk. If you hope there is no
687 * inactive transaction in the fs when -ENOENT is returned, you should
688 * invoke
689 * btrfs_attach_transaction_barrier()
690 */
691struct btrfs_trans_handle *btrfs_attach_transaction(struct btrfs_root *root)
692{
693 return start_transaction(root, 0, TRANS_ATTACH,
694 BTRFS_RESERVE_NO_FLUSH);
695}
696
697/*
698 * btrfs_attach_transaction_barrier() - catch the running transaction
699 *
700 * It is similar to the above function, the differentia is this one
701 * will wait for all the inactive transactions until they fully
702 * complete.
703 */
704struct btrfs_trans_handle *
705btrfs_attach_transaction_barrier(struct btrfs_root *root)
706{
707 struct btrfs_trans_handle *trans;
708
709 trans = start_transaction(root, 0, TRANS_ATTACH,
710 BTRFS_RESERVE_NO_FLUSH);
711 if (IS_ERR(trans) && PTR_ERR(trans) == -ENOENT)
712 btrfs_wait_for_commit(root->fs_info, 0);
713
714 return trans;
715}
716
717/* wait for a transaction commit to be fully complete */
718static noinline void wait_for_commit(struct btrfs_transaction *commit)
719{
720 wait_event(commit->commit_wait, commit->state == TRANS_STATE_COMPLETED);
721}
722
723int btrfs_wait_for_commit(struct btrfs_fs_info *fs_info, u64 transid)
724{
725 struct btrfs_transaction *cur_trans = NULL, *t;
726 int ret = 0;
727
728 if (transid) {
729 if (transid <= fs_info->last_trans_committed)
730 goto out;
731
732 /* find specified transaction */
733 spin_lock(&fs_info->trans_lock);
734 list_for_each_entry(t, &fs_info->trans_list, list) {
735 if (t->transid == transid) {
736 cur_trans = t;
737 atomic_inc(&cur_trans->use_count);
738 ret = 0;
739 break;
740 }
741 if (t->transid > transid) {
742 ret = 0;
743 break;
744 }
745 }
746 spin_unlock(&fs_info->trans_lock);
747
748 /*
749 * The specified transaction doesn't exist, or we
750 * raced with btrfs_commit_transaction
751 */
752 if (!cur_trans) {
753 if (transid > fs_info->last_trans_committed)
754 ret = -EINVAL;
755 goto out;
756 }
757 } else {
758 /* find newest transaction that is committing | committed */
759 spin_lock(&fs_info->trans_lock);
760 list_for_each_entry_reverse(t, &fs_info->trans_list,
761 list) {
762 if (t->state >= TRANS_STATE_COMMIT_START) {
763 if (t->state == TRANS_STATE_COMPLETED)
764 break;
765 cur_trans = t;
766 atomic_inc(&cur_trans->use_count);
767 break;
768 }
769 }
770 spin_unlock(&fs_info->trans_lock);
771 if (!cur_trans)
772 goto out; /* nothing committing|committed */
773 }
774
775 wait_for_commit(cur_trans);
776 btrfs_put_transaction(cur_trans);
777out:
778 return ret;
779}
780
781void btrfs_throttle(struct btrfs_fs_info *fs_info)
782{
783 if (!atomic_read(&fs_info->open_ioctl_trans))
784 wait_current_trans(fs_info);
785}
786
787static int should_end_transaction(struct btrfs_trans_handle *trans)
788{
789 struct btrfs_fs_info *fs_info = trans->fs_info;
790
791 if (fs_info->global_block_rsv.space_info->full &&
792 btrfs_check_space_for_delayed_refs(trans, fs_info))
793 return 1;
794
795 return !!btrfs_block_rsv_check(&fs_info->global_block_rsv, 5);
796}
797
798int btrfs_should_end_transaction(struct btrfs_trans_handle *trans)
799{
800 struct btrfs_transaction *cur_trans = trans->transaction;
801 struct btrfs_fs_info *fs_info = trans->fs_info;
802 int updates;
803 int err;
804
805 smp_mb();
806 if (cur_trans->state >= TRANS_STATE_BLOCKED ||
807 cur_trans->delayed_refs.flushing)
808 return 1;
809
810 updates = trans->delayed_ref_updates;
811 trans->delayed_ref_updates = 0;
812 if (updates) {
813 err = btrfs_run_delayed_refs(trans, fs_info, updates * 2);
814 if (err) /* Error code will also eval true */
815 return err;
816 }
817
818 return should_end_transaction(trans);
819}
820
821static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
822 int throttle)
823{
824 struct btrfs_fs_info *info = trans->fs_info;
825 struct btrfs_transaction *cur_trans = trans->transaction;
826 u64 transid = trans->transid;
827 unsigned long cur = trans->delayed_ref_updates;
828 int lock = (trans->type != TRANS_JOIN_NOLOCK);
829 int err = 0;
830 int must_run_delayed_refs = 0;
831
832 if (trans->use_count > 1) {
833 trans->use_count--;
834 trans->block_rsv = trans->orig_rsv;
835 return 0;
836 }
837
838 btrfs_trans_release_metadata(trans, info);
839 trans->block_rsv = NULL;
840
841 if (!list_empty(&trans->new_bgs))
842 btrfs_create_pending_block_groups(trans, info);
843
844 trans->delayed_ref_updates = 0;
845 if (!trans->sync) {
846 must_run_delayed_refs =
847 btrfs_should_throttle_delayed_refs(trans, info);
848 cur = max_t(unsigned long, cur, 32);
849
850 /*
851 * don't make the caller wait if they are from a NOLOCK
852 * or ATTACH transaction, it will deadlock with commit
853 */
854 if (must_run_delayed_refs == 1 &&
855 (trans->type & (__TRANS_JOIN_NOLOCK | __TRANS_ATTACH)))
856 must_run_delayed_refs = 2;
857 }
858
859 btrfs_trans_release_metadata(trans, info);
860 trans->block_rsv = NULL;
861
862 if (!list_empty(&trans->new_bgs))
863 btrfs_create_pending_block_groups(trans, info);
864
865 btrfs_trans_release_chunk_metadata(trans);
866
867 if (lock && !atomic_read(&info->open_ioctl_trans) &&
868 should_end_transaction(trans) &&
869 ACCESS_ONCE(cur_trans->state) == TRANS_STATE_RUNNING) {
870 spin_lock(&info->trans_lock);
871 if (cur_trans->state == TRANS_STATE_RUNNING)
872 cur_trans->state = TRANS_STATE_BLOCKED;
873 spin_unlock(&info->trans_lock);
874 }
875
876 if (lock && ACCESS_ONCE(cur_trans->state) == TRANS_STATE_BLOCKED) {
877 if (throttle)
878 return btrfs_commit_transaction(trans);
879 else
880 wake_up_process(info->transaction_kthread);
881 }
882
883 if (trans->type & __TRANS_FREEZABLE)
884 sb_end_intwrite(info->sb);
885
886 WARN_ON(cur_trans != info->running_transaction);
887 WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
888 atomic_dec(&cur_trans->num_writers);
889 extwriter_counter_dec(cur_trans, trans->type);
890
891 /*
892 * Make sure counter is updated before we wake up waiters.
893 */
894 smp_mb();
895 if (waitqueue_active(&cur_trans->writer_wait))
896 wake_up(&cur_trans->writer_wait);
897 btrfs_put_transaction(cur_trans);
898
899 if (current->journal_info == trans)
900 current->journal_info = NULL;
901
902 if (throttle)
903 btrfs_run_delayed_iputs(info);
904
905 if (trans->aborted ||
906 test_bit(BTRFS_FS_STATE_ERROR, &info->fs_state)) {
907 wake_up_process(info->transaction_kthread);
908 err = -EIO;
909 }
910 assert_qgroups_uptodate(trans);
911
912 kmem_cache_free(btrfs_trans_handle_cachep, trans);
913 if (must_run_delayed_refs) {
914 btrfs_async_run_delayed_refs(info, cur, transid,
915 must_run_delayed_refs == 1);
916 }
917 return err;
918}
919
920int btrfs_end_transaction(struct btrfs_trans_handle *trans)
921{
922 return __btrfs_end_transaction(trans, 0);
923}
924
925int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans)
926{
927 return __btrfs_end_transaction(trans, 1);
928}
929
930/*
931 * when btree blocks are allocated, they have some corresponding bits set for
932 * them in one of two extent_io trees. This is used to make sure all of
933 * those extents are sent to disk but does not wait on them
934 */
935int btrfs_write_marked_extents(struct btrfs_fs_info *fs_info,
936 struct extent_io_tree *dirty_pages, int mark)
937{
938 int err = 0;
939 int werr = 0;
940 struct address_space *mapping = fs_info->btree_inode->i_mapping;
941 struct extent_state *cached_state = NULL;
942 u64 start = 0;
943 u64 end;
944
945 while (!find_first_extent_bit(dirty_pages, start, &start, &end,
946 mark, &cached_state)) {
947 bool wait_writeback = false;
948
949 err = convert_extent_bit(dirty_pages, start, end,
950 EXTENT_NEED_WAIT,
951 mark, &cached_state);
952 /*
953 * convert_extent_bit can return -ENOMEM, which is most of the
954 * time a temporary error. So when it happens, ignore the error
955 * and wait for writeback of this range to finish - because we
956 * failed to set the bit EXTENT_NEED_WAIT for the range, a call
957 * to __btrfs_wait_marked_extents() would not know that
958 * writeback for this range started and therefore wouldn't
959 * wait for it to finish - we don't want to commit a
960 * superblock that points to btree nodes/leafs for which
961 * writeback hasn't finished yet (and without errors).
962 * We cleanup any entries left in the io tree when committing
963 * the transaction (through clear_btree_io_tree()).
964 */
965 if (err == -ENOMEM) {
966 err = 0;
967 wait_writeback = true;
968 }
969 if (!err)
970 err = filemap_fdatawrite_range(mapping, start, end);
971 if (err)
972 werr = err;
973 else if (wait_writeback)
974 werr = filemap_fdatawait_range(mapping, start, end);
975 free_extent_state(cached_state);
976 cached_state = NULL;
977 cond_resched();
978 start = end + 1;
979 }
980 return werr;
981}
982
983/*
984 * when btree blocks are allocated, they have some corresponding bits set for
985 * them in one of two extent_io trees. This is used to make sure all of
986 * those extents are on disk for transaction or log commit. We wait
987 * on all the pages and clear them from the dirty pages state tree
988 */
989static int __btrfs_wait_marked_extents(struct btrfs_fs_info *fs_info,
990 struct extent_io_tree *dirty_pages)
991{
992 int err = 0;
993 int werr = 0;
994 struct address_space *mapping = fs_info->btree_inode->i_mapping;
995 struct extent_state *cached_state = NULL;
996 u64 start = 0;
997 u64 end;
998
999 while (!find_first_extent_bit(dirty_pages, start, &start, &end,
1000 EXTENT_NEED_WAIT, &cached_state)) {
1001 /*
1002 * Ignore -ENOMEM errors returned by clear_extent_bit().
1003 * When committing the transaction, we'll remove any entries
1004 * left in the io tree. For a log commit, we don't remove them
1005 * after committing the log because the tree can be accessed
1006 * concurrently - we do it only at transaction commit time when
1007 * it's safe to do it (through clear_btree_io_tree()).
1008 */
1009 err = clear_extent_bit(dirty_pages, start, end,
1010 EXTENT_NEED_WAIT,
1011 0, 0, &cached_state, GFP_NOFS);
1012 if (err == -ENOMEM)
1013 err = 0;
1014 if (!err)
1015 err = filemap_fdatawait_range(mapping, start, end);
1016 if (err)
1017 werr = err;
1018 free_extent_state(cached_state);
1019 cached_state = NULL;
1020 cond_resched();
1021 start = end + 1;
1022 }
1023 if (err)
1024 werr = err;
1025 return werr;
1026}
1027
1028int btrfs_wait_extents(struct btrfs_fs_info *fs_info,
1029 struct extent_io_tree *dirty_pages)
1030{
1031 bool errors = false;
1032 int err;
1033
1034 err = __btrfs_wait_marked_extents(fs_info, dirty_pages);
1035 if (test_and_clear_bit(BTRFS_FS_BTREE_ERR, &fs_info->flags))
1036 errors = true;
1037
1038 if (errors && !err)
1039 err = -EIO;
1040 return err;
1041}
1042
1043int btrfs_wait_tree_log_extents(struct btrfs_root *log_root, int mark)
1044{
1045 struct btrfs_fs_info *fs_info = log_root->fs_info;
1046 struct extent_io_tree *dirty_pages = &log_root->dirty_log_pages;
1047 bool errors = false;
1048 int err;
1049
1050 ASSERT(log_root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
1051
1052 err = __btrfs_wait_marked_extents(fs_info, dirty_pages);
1053 if ((mark & EXTENT_DIRTY) &&
1054 test_and_clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags))
1055 errors = true;
1056
1057 if ((mark & EXTENT_NEW) &&
1058 test_and_clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags))
1059 errors = true;
1060
1061 if (errors && !err)
1062 err = -EIO;
1063 return err;
1064}
1065
1066/*
1067 * when btree blocks are allocated, they have some corresponding bits set for
1068 * them in one of two extent_io trees. This is used to make sure all of
1069 * those extents are on disk for transaction or log commit
1070 */
1071static int btrfs_write_and_wait_marked_extents(struct btrfs_fs_info *fs_info,
1072 struct extent_io_tree *dirty_pages, int mark)
1073{
1074 int ret;
1075 int ret2;
1076 struct blk_plug plug;
1077
1078 blk_start_plug(&plug);
1079 ret = btrfs_write_marked_extents(fs_info, dirty_pages, mark);
1080 blk_finish_plug(&plug);
1081 ret2 = btrfs_wait_extents(fs_info, dirty_pages);
1082
1083 if (ret)
1084 return ret;
1085 if (ret2)
1086 return ret2;
1087 return 0;
1088}
1089
1090static int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
1091 struct btrfs_fs_info *fs_info)
1092{
1093 int ret;
1094
1095 ret = btrfs_write_and_wait_marked_extents(fs_info,
1096 &trans->transaction->dirty_pages,
1097 EXTENT_DIRTY);
1098 clear_btree_io_tree(&trans->transaction->dirty_pages);
1099
1100 return ret;
1101}
1102
1103/*
1104 * this is used to update the root pointer in the tree of tree roots.
1105 *
1106 * But, in the case of the extent allocation tree, updating the root
1107 * pointer may allocate blocks which may change the root of the extent
1108 * allocation tree.
1109 *
1110 * So, this loops and repeats and makes sure the cowonly root didn't
1111 * change while the root pointer was being updated in the metadata.
1112 */
1113static int update_cowonly_root(struct btrfs_trans_handle *trans,
1114 struct btrfs_root *root)
1115{
1116 int ret;
1117 u64 old_root_bytenr;
1118 u64 old_root_used;
1119 struct btrfs_fs_info *fs_info = root->fs_info;
1120 struct btrfs_root *tree_root = fs_info->tree_root;
1121
1122 old_root_used = btrfs_root_used(&root->root_item);
1123
1124 while (1) {
1125 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
1126 if (old_root_bytenr == root->node->start &&
1127 old_root_used == btrfs_root_used(&root->root_item))
1128 break;
1129
1130 btrfs_set_root_node(&root->root_item, root->node);
1131 ret = btrfs_update_root(trans, tree_root,
1132 &root->root_key,
1133 &root->root_item);
1134 if (ret)
1135 return ret;
1136
1137 old_root_used = btrfs_root_used(&root->root_item);
1138 }
1139
1140 return 0;
1141}
1142
1143/*
1144 * update all the cowonly tree roots on disk
1145 *
1146 * The error handling in this function may not be obvious. Any of the
1147 * failures will cause the file system to go offline. We still need
1148 * to clean up the delayed refs.
1149 */
1150static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
1151 struct btrfs_fs_info *fs_info)
1152{
1153 struct list_head *dirty_bgs = &trans->transaction->dirty_bgs;
1154 struct list_head *io_bgs = &trans->transaction->io_bgs;
1155 struct list_head *next;
1156 struct extent_buffer *eb;
1157 int ret;
1158
1159 eb = btrfs_lock_root_node(fs_info->tree_root);
1160 ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL,
1161 0, &eb);
1162 btrfs_tree_unlock(eb);
1163 free_extent_buffer(eb);
1164
1165 if (ret)
1166 return ret;
1167
1168 ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1169 if (ret)
1170 return ret;
1171
1172 ret = btrfs_run_dev_stats(trans, fs_info);
1173 if (ret)
1174 return ret;
1175 ret = btrfs_run_dev_replace(trans, fs_info);
1176 if (ret)
1177 return ret;
1178 ret = btrfs_run_qgroups(trans, fs_info);
1179 if (ret)
1180 return ret;
1181
1182 ret = btrfs_setup_space_cache(trans, fs_info);
1183 if (ret)
1184 return ret;
1185
1186 /* run_qgroups might have added some more refs */
1187 ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1188 if (ret)
1189 return ret;
1190again:
1191 while (!list_empty(&fs_info->dirty_cowonly_roots)) {
1192 struct btrfs_root *root;
1193 next = fs_info->dirty_cowonly_roots.next;
1194 list_del_init(next);
1195 root = list_entry(next, struct btrfs_root, dirty_list);
1196 clear_bit(BTRFS_ROOT_DIRTY, &root->state);
1197
1198 if (root != fs_info->extent_root)
1199 list_add_tail(&root->dirty_list,
1200 &trans->transaction->switch_commits);
1201 ret = update_cowonly_root(trans, root);
1202 if (ret)
1203 return ret;
1204 ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1205 if (ret)
1206 return ret;
1207 }
1208
1209 while (!list_empty(dirty_bgs) || !list_empty(io_bgs)) {
1210 ret = btrfs_write_dirty_block_groups(trans, fs_info);
1211 if (ret)
1212 return ret;
1213 ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1214 if (ret)
1215 return ret;
1216 }
1217
1218 if (!list_empty(&fs_info->dirty_cowonly_roots))
1219 goto again;
1220
1221 list_add_tail(&fs_info->extent_root->dirty_list,
1222 &trans->transaction->switch_commits);
1223 btrfs_after_dev_replace_commit(fs_info);
1224
1225 return 0;
1226}
1227
1228/*
1229 * dead roots are old snapshots that need to be deleted. This allocates
1230 * a dirty root struct and adds it into the list of dead roots that need to
1231 * be deleted
1232 */
1233void btrfs_add_dead_root(struct btrfs_root *root)
1234{
1235 struct btrfs_fs_info *fs_info = root->fs_info;
1236
1237 spin_lock(&fs_info->trans_lock);
1238 if (list_empty(&root->root_list))
1239 list_add_tail(&root->root_list, &fs_info->dead_roots);
1240 spin_unlock(&fs_info->trans_lock);
1241}
1242
1243/*
1244 * update all the cowonly tree roots on disk
1245 */
1246static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
1247 struct btrfs_fs_info *fs_info)
1248{
1249 struct btrfs_root *gang[8];
1250 int i;
1251 int ret;
1252 int err = 0;
1253
1254 spin_lock(&fs_info->fs_roots_radix_lock);
1255 while (1) {
1256 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
1257 (void **)gang, 0,
1258 ARRAY_SIZE(gang),
1259 BTRFS_ROOT_TRANS_TAG);
1260 if (ret == 0)
1261 break;
1262 for (i = 0; i < ret; i++) {
1263 struct btrfs_root *root = gang[i];
1264 radix_tree_tag_clear(&fs_info->fs_roots_radix,
1265 (unsigned long)root->root_key.objectid,
1266 BTRFS_ROOT_TRANS_TAG);
1267 spin_unlock(&fs_info->fs_roots_radix_lock);
1268
1269 btrfs_free_log(trans, root);
1270 btrfs_update_reloc_root(trans, root);
1271 btrfs_orphan_commit_root(trans, root);
1272
1273 btrfs_save_ino_cache(root, trans);
1274
1275 /* see comments in should_cow_block() */
1276 clear_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1277 smp_mb__after_atomic();
1278
1279 if (root->commit_root != root->node) {
1280 list_add_tail(&root->dirty_list,
1281 &trans->transaction->switch_commits);
1282 btrfs_set_root_node(&root->root_item,
1283 root->node);
1284 }
1285
1286 err = btrfs_update_root(trans, fs_info->tree_root,
1287 &root->root_key,
1288 &root->root_item);
1289 spin_lock(&fs_info->fs_roots_radix_lock);
1290 if (err)
1291 break;
1292 btrfs_qgroup_free_meta_all(root);
1293 }
1294 }
1295 spin_unlock(&fs_info->fs_roots_radix_lock);
1296 return err;
1297}
1298
1299/*
1300 * defrag a given btree.
1301 * Every leaf in the btree is read and defragged.
1302 */
1303int btrfs_defrag_root(struct btrfs_root *root)
1304{
1305 struct btrfs_fs_info *info = root->fs_info;
1306 struct btrfs_trans_handle *trans;
1307 int ret;
1308
1309 if (test_and_set_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state))
1310 return 0;
1311
1312 while (1) {
1313 trans = btrfs_start_transaction(root, 0);
1314 if (IS_ERR(trans))
1315 return PTR_ERR(trans);
1316
1317 ret = btrfs_defrag_leaves(trans, root);
1318
1319 btrfs_end_transaction(trans);
1320 btrfs_btree_balance_dirty(info);
1321 cond_resched();
1322
1323 if (btrfs_fs_closing(info) || ret != -EAGAIN)
1324 break;
1325
1326 if (btrfs_defrag_cancelled(info)) {
1327 btrfs_debug(info, "defrag_root cancelled");
1328 ret = -EAGAIN;
1329 break;
1330 }
1331 }
1332 clear_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state);
1333 return ret;
1334}
1335
1336/*
1337 * Do all special snapshot related qgroup dirty hack.
1338 *
1339 * Will do all needed qgroup inherit and dirty hack like switch commit
1340 * roots inside one transaction and write all btree into disk, to make
1341 * qgroup works.
1342 */
1343static int qgroup_account_snapshot(struct btrfs_trans_handle *trans,
1344 struct btrfs_root *src,
1345 struct btrfs_root *parent,
1346 struct btrfs_qgroup_inherit *inherit,
1347 u64 dst_objectid)
1348{
1349 struct btrfs_fs_info *fs_info = src->fs_info;
1350 int ret;
1351
1352 /*
1353 * Save some performance in the case that qgroups are not
1354 * enabled. If this check races with the ioctl, rescan will
1355 * kick in anyway.
1356 */
1357 mutex_lock(&fs_info->qgroup_ioctl_lock);
1358 if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)) {
1359 mutex_unlock(&fs_info->qgroup_ioctl_lock);
1360 return 0;
1361 }
1362 mutex_unlock(&fs_info->qgroup_ioctl_lock);
1363
1364 /*
1365 * We are going to commit transaction, see btrfs_commit_transaction()
1366 * comment for reason locking tree_log_mutex
1367 */
1368 mutex_lock(&fs_info->tree_log_mutex);
1369
1370 ret = commit_fs_roots(trans, fs_info);
1371 if (ret)
1372 goto out;
1373 ret = btrfs_qgroup_prepare_account_extents(trans, fs_info);
1374 if (ret < 0)
1375 goto out;
1376 ret = btrfs_qgroup_account_extents(trans, fs_info);
1377 if (ret < 0)
1378 goto out;
1379
1380 /* Now qgroup are all updated, we can inherit it to new qgroups */
1381 ret = btrfs_qgroup_inherit(trans, fs_info,
1382 src->root_key.objectid, dst_objectid,
1383 inherit);
1384 if (ret < 0)
1385 goto out;
1386
1387 /*
1388 * Now we do a simplified commit transaction, which will:
1389 * 1) commit all subvolume and extent tree
1390 * To ensure all subvolume and extent tree have a valid
1391 * commit_root to accounting later insert_dir_item()
1392 * 2) write all btree blocks onto disk
1393 * This is to make sure later btree modification will be cowed
1394 * Or commit_root can be populated and cause wrong qgroup numbers
1395 * In this simplified commit, we don't really care about other trees
1396 * like chunk and root tree, as they won't affect qgroup.
1397 * And we don't write super to avoid half committed status.
1398 */
1399 ret = commit_cowonly_roots(trans, fs_info);
1400 if (ret)
1401 goto out;
1402 switch_commit_roots(trans->transaction, fs_info);
1403 ret = btrfs_write_and_wait_transaction(trans, fs_info);
1404 if (ret)
1405 btrfs_handle_fs_error(fs_info, ret,
1406 "Error while writing out transaction for qgroup");
1407
1408out:
1409 mutex_unlock(&fs_info->tree_log_mutex);
1410
1411 /*
1412 * Force parent root to be updated, as we recorded it before so its
1413 * last_trans == cur_transid.
1414 * Or it won't be committed again onto disk after later
1415 * insert_dir_item()
1416 */
1417 if (!ret)
1418 record_root_in_trans(trans, parent, 1);
1419 return ret;
1420}
1421
1422/*
1423 * new snapshots need to be created at a very specific time in the
1424 * transaction commit. This does the actual creation.
1425 *
1426 * Note:
1427 * If the error which may affect the commitment of the current transaction
1428 * happens, we should return the error number. If the error which just affect
1429 * the creation of the pending snapshots, just return 0.
1430 */
1431static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
1432 struct btrfs_fs_info *fs_info,
1433 struct btrfs_pending_snapshot *pending)
1434{
1435 struct btrfs_key key;
1436 struct btrfs_root_item *new_root_item;
1437 struct btrfs_root *tree_root = fs_info->tree_root;
1438 struct btrfs_root *root = pending->root;
1439 struct btrfs_root *parent_root;
1440 struct btrfs_block_rsv *rsv;
1441 struct inode *parent_inode;
1442 struct btrfs_path *path;
1443 struct btrfs_dir_item *dir_item;
1444 struct dentry *dentry;
1445 struct extent_buffer *tmp;
1446 struct extent_buffer *old;
1447 struct timespec cur_time;
1448 int ret = 0;
1449 u64 to_reserve = 0;
1450 u64 index = 0;
1451 u64 objectid;
1452 u64 root_flags;
1453 uuid_le new_uuid;
1454
1455 ASSERT(pending->path);
1456 path = pending->path;
1457
1458 ASSERT(pending->root_item);
1459 new_root_item = pending->root_item;
1460
1461 pending->error = btrfs_find_free_objectid(tree_root, &objectid);
1462 if (pending->error)
1463 goto no_free_objectid;
1464
1465 /*
1466 * Make qgroup to skip current new snapshot's qgroupid, as it is
1467 * accounted by later btrfs_qgroup_inherit().
1468 */
1469 btrfs_set_skip_qgroup(trans, objectid);
1470
1471 btrfs_reloc_pre_snapshot(pending, &to_reserve);
1472
1473 if (to_reserve > 0) {
1474 pending->error = btrfs_block_rsv_add(root,
1475 &pending->block_rsv,
1476 to_reserve,
1477 BTRFS_RESERVE_NO_FLUSH);
1478 if (pending->error)
1479 goto clear_skip_qgroup;
1480 }
1481
1482 key.objectid = objectid;
1483 key.offset = (u64)-1;
1484 key.type = BTRFS_ROOT_ITEM_KEY;
1485
1486 rsv = trans->block_rsv;
1487 trans->block_rsv = &pending->block_rsv;
1488 trans->bytes_reserved = trans->block_rsv->reserved;
1489 trace_btrfs_space_reservation(fs_info, "transaction",
1490 trans->transid,
1491 trans->bytes_reserved, 1);
1492 dentry = pending->dentry;
1493 parent_inode = pending->dir;
1494 parent_root = BTRFS_I(parent_inode)->root;
1495 record_root_in_trans(trans, parent_root, 0);
1496
1497 cur_time = current_time(parent_inode);
1498
1499 /*
1500 * insert the directory item
1501 */
1502 ret = btrfs_set_inode_index(parent_inode, &index);
1503 BUG_ON(ret); /* -ENOMEM */
1504
1505 /* check if there is a file/dir which has the same name. */
1506 dir_item = btrfs_lookup_dir_item(NULL, parent_root, path,
1507 btrfs_ino(parent_inode),
1508 dentry->d_name.name,
1509 dentry->d_name.len, 0);
1510 if (dir_item != NULL && !IS_ERR(dir_item)) {
1511 pending->error = -EEXIST;
1512 goto dir_item_existed;
1513 } else if (IS_ERR(dir_item)) {
1514 ret = PTR_ERR(dir_item);
1515 btrfs_abort_transaction(trans, ret);
1516 goto fail;
1517 }
1518 btrfs_release_path(path);
1519
1520 /*
1521 * pull in the delayed directory update
1522 * and the delayed inode item
1523 * otherwise we corrupt the FS during
1524 * snapshot
1525 */
1526 ret = btrfs_run_delayed_items(trans, fs_info);
1527 if (ret) { /* Transaction aborted */
1528 btrfs_abort_transaction(trans, ret);
1529 goto fail;
1530 }
1531
1532 record_root_in_trans(trans, root, 0);
1533 btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
1534 memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
1535 btrfs_check_and_init_root_item(new_root_item);
1536
1537 root_flags = btrfs_root_flags(new_root_item);
1538 if (pending->readonly)
1539 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
1540 else
1541 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
1542 btrfs_set_root_flags(new_root_item, root_flags);
1543
1544 btrfs_set_root_generation_v2(new_root_item,
1545 trans->transid);
1546 uuid_le_gen(&new_uuid);
1547 memcpy(new_root_item->uuid, new_uuid.b, BTRFS_UUID_SIZE);
1548 memcpy(new_root_item->parent_uuid, root->root_item.uuid,
1549 BTRFS_UUID_SIZE);
1550 if (!(root_flags & BTRFS_ROOT_SUBVOL_RDONLY)) {
1551 memset(new_root_item->received_uuid, 0,
1552 sizeof(new_root_item->received_uuid));
1553 memset(&new_root_item->stime, 0, sizeof(new_root_item->stime));
1554 memset(&new_root_item->rtime, 0, sizeof(new_root_item->rtime));
1555 btrfs_set_root_stransid(new_root_item, 0);
1556 btrfs_set_root_rtransid(new_root_item, 0);
1557 }
1558 btrfs_set_stack_timespec_sec(&new_root_item->otime, cur_time.tv_sec);
1559 btrfs_set_stack_timespec_nsec(&new_root_item->otime, cur_time.tv_nsec);
1560 btrfs_set_root_otransid(new_root_item, trans->transid);
1561
1562 old = btrfs_lock_root_node(root);
1563 ret = btrfs_cow_block(trans, root, old, NULL, 0, &old);
1564 if (ret) {
1565 btrfs_tree_unlock(old);
1566 free_extent_buffer(old);
1567 btrfs_abort_transaction(trans, ret);
1568 goto fail;
1569 }
1570
1571 btrfs_set_lock_blocking(old);
1572
1573 ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1574 /* clean up in any case */
1575 btrfs_tree_unlock(old);
1576 free_extent_buffer(old);
1577 if (ret) {
1578 btrfs_abort_transaction(trans, ret);
1579 goto fail;
1580 }
1581 /* see comments in should_cow_block() */
1582 set_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1583 smp_wmb();
1584
1585 btrfs_set_root_node(new_root_item, tmp);
1586 /* record when the snapshot was created in key.offset */
1587 key.offset = trans->transid;
1588 ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1589 btrfs_tree_unlock(tmp);
1590 free_extent_buffer(tmp);
1591 if (ret) {
1592 btrfs_abort_transaction(trans, ret);
1593 goto fail;
1594 }
1595
1596 /*
1597 * insert root back/forward references
1598 */
1599 ret = btrfs_add_root_ref(trans, fs_info, objectid,
1600 parent_root->root_key.objectid,
1601 btrfs_ino(parent_inode), index,
1602 dentry->d_name.name, dentry->d_name.len);
1603 if (ret) {
1604 btrfs_abort_transaction(trans, ret);
1605 goto fail;
1606 }
1607
1608 key.offset = (u64)-1;
1609 pending->snap = btrfs_read_fs_root_no_name(fs_info, &key);
1610 if (IS_ERR(pending->snap)) {
1611 ret = PTR_ERR(pending->snap);
1612 btrfs_abort_transaction(trans, ret);
1613 goto fail;
1614 }
1615
1616 ret = btrfs_reloc_post_snapshot(trans, pending);
1617 if (ret) {
1618 btrfs_abort_transaction(trans, ret);
1619 goto fail;
1620 }
1621
1622 ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1623 if (ret) {
1624 btrfs_abort_transaction(trans, ret);
1625 goto fail;
1626 }
1627
1628 /*
1629 * Do special qgroup accounting for snapshot, as we do some qgroup
1630 * snapshot hack to do fast snapshot.
1631 * To co-operate with that hack, we do hack again.
1632 * Or snapshot will be greatly slowed down by a subtree qgroup rescan
1633 */
1634 ret = qgroup_account_snapshot(trans, root, parent_root,
1635 pending->inherit, objectid);
1636 if (ret < 0)
1637 goto fail;
1638
1639 ret = btrfs_insert_dir_item(trans, parent_root,
1640 dentry->d_name.name, dentry->d_name.len,
1641 parent_inode, &key,
1642 BTRFS_FT_DIR, index);
1643 /* We have check then name at the beginning, so it is impossible. */
1644 BUG_ON(ret == -EEXIST || ret == -EOVERFLOW);
1645 if (ret) {
1646 btrfs_abort_transaction(trans, ret);
1647 goto fail;
1648 }
1649
1650 btrfs_i_size_write(parent_inode, parent_inode->i_size +
1651 dentry->d_name.len * 2);
1652 parent_inode->i_mtime = parent_inode->i_ctime =
1653 current_time(parent_inode);
1654 ret = btrfs_update_inode_fallback(trans, parent_root, parent_inode);
1655 if (ret) {
1656 btrfs_abort_transaction(trans, ret);
1657 goto fail;
1658 }
1659 ret = btrfs_uuid_tree_add(trans, fs_info, new_uuid.b,
1660 BTRFS_UUID_KEY_SUBVOL, objectid);
1661 if (ret) {
1662 btrfs_abort_transaction(trans, ret);
1663 goto fail;
1664 }
1665 if (!btrfs_is_empty_uuid(new_root_item->received_uuid)) {
1666 ret = btrfs_uuid_tree_add(trans, fs_info,
1667 new_root_item->received_uuid,
1668 BTRFS_UUID_KEY_RECEIVED_SUBVOL,
1669 objectid);
1670 if (ret && ret != -EEXIST) {
1671 btrfs_abort_transaction(trans, ret);
1672 goto fail;
1673 }
1674 }
1675
1676 ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1677 if (ret) {
1678 btrfs_abort_transaction(trans, ret);
1679 goto fail;
1680 }
1681
1682fail:
1683 pending->error = ret;
1684dir_item_existed:
1685 trans->block_rsv = rsv;
1686 trans->bytes_reserved = 0;
1687clear_skip_qgroup:
1688 btrfs_clear_skip_qgroup(trans);
1689no_free_objectid:
1690 kfree(new_root_item);
1691 pending->root_item = NULL;
1692 btrfs_free_path(path);
1693 pending->path = NULL;
1694
1695 return ret;
1696}
1697
1698/*
1699 * create all the snapshots we've scheduled for creation
1700 */
1701static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
1702 struct btrfs_fs_info *fs_info)
1703{
1704 struct btrfs_pending_snapshot *pending, *next;
1705 struct list_head *head = &trans->transaction->pending_snapshots;
1706 int ret = 0;
1707
1708 list_for_each_entry_safe(pending, next, head, list) {
1709 list_del(&pending->list);
1710 ret = create_pending_snapshot(trans, fs_info, pending);
1711 if (ret)
1712 break;
1713 }
1714 return ret;
1715}
1716
1717static void update_super_roots(struct btrfs_fs_info *fs_info)
1718{
1719 struct btrfs_root_item *root_item;
1720 struct btrfs_super_block *super;
1721
1722 super = fs_info->super_copy;
1723
1724 root_item = &fs_info->chunk_root->root_item;
1725 super->chunk_root = root_item->bytenr;
1726 super->chunk_root_generation = root_item->generation;
1727 super->chunk_root_level = root_item->level;
1728
1729 root_item = &fs_info->tree_root->root_item;
1730 super->root = root_item->bytenr;
1731 super->generation = root_item->generation;
1732 super->root_level = root_item->level;
1733 if (btrfs_test_opt(fs_info, SPACE_CACHE))
1734 super->cache_generation = root_item->generation;
1735 if (test_bit(BTRFS_FS_UPDATE_UUID_TREE_GEN, &fs_info->flags))
1736 super->uuid_tree_generation = root_item->generation;
1737}
1738
1739int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1740{
1741 struct btrfs_transaction *trans;
1742 int ret = 0;
1743
1744 spin_lock(&info->trans_lock);
1745 trans = info->running_transaction;
1746 if (trans)
1747 ret = (trans->state >= TRANS_STATE_COMMIT_START);
1748 spin_unlock(&info->trans_lock);
1749 return ret;
1750}
1751
1752int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1753{
1754 struct btrfs_transaction *trans;
1755 int ret = 0;
1756
1757 spin_lock(&info->trans_lock);
1758 trans = info->running_transaction;
1759 if (trans)
1760 ret = is_transaction_blocked(trans);
1761 spin_unlock(&info->trans_lock);
1762 return ret;
1763}
1764
1765/*
1766 * wait for the current transaction commit to start and block subsequent
1767 * transaction joins
1768 */
1769static void wait_current_trans_commit_start(struct btrfs_fs_info *fs_info,
1770 struct btrfs_transaction *trans)
1771{
1772 wait_event(fs_info->transaction_blocked_wait,
1773 trans->state >= TRANS_STATE_COMMIT_START || trans->aborted);
1774}
1775
1776/*
1777 * wait for the current transaction to start and then become unblocked.
1778 * caller holds ref.
1779 */
1780static void wait_current_trans_commit_start_and_unblock(
1781 struct btrfs_fs_info *fs_info,
1782 struct btrfs_transaction *trans)
1783{
1784 wait_event(fs_info->transaction_wait,
1785 trans->state >= TRANS_STATE_UNBLOCKED || trans->aborted);
1786}
1787
1788/*
1789 * commit transactions asynchronously. once btrfs_commit_transaction_async
1790 * returns, any subsequent transaction will not be allowed to join.
1791 */
1792struct btrfs_async_commit {
1793 struct btrfs_trans_handle *newtrans;
1794 struct work_struct work;
1795};
1796
1797static void do_async_commit(struct work_struct *work)
1798{
1799 struct btrfs_async_commit *ac =
1800 container_of(work, struct btrfs_async_commit, work);
1801
1802 /*
1803 * We've got freeze protection passed with the transaction.
1804 * Tell lockdep about it.
1805 */
1806 if (ac->newtrans->type & __TRANS_FREEZABLE)
1807 __sb_writers_acquired(ac->newtrans->fs_info->sb, SB_FREEZE_FS);
1808
1809 current->journal_info = ac->newtrans;
1810
1811 btrfs_commit_transaction(ac->newtrans);
1812 kfree(ac);
1813}
1814
1815int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1816 int wait_for_unblock)
1817{
1818 struct btrfs_fs_info *fs_info = trans->fs_info;
1819 struct btrfs_async_commit *ac;
1820 struct btrfs_transaction *cur_trans;
1821
1822 ac = kmalloc(sizeof(*ac), GFP_NOFS);
1823 if (!ac)
1824 return -ENOMEM;
1825
1826 INIT_WORK(&ac->work, do_async_commit);
1827 ac->newtrans = btrfs_join_transaction(trans->root);
1828 if (IS_ERR(ac->newtrans)) {
1829 int err = PTR_ERR(ac->newtrans);
1830 kfree(ac);
1831 return err;
1832 }
1833
1834 /* take transaction reference */
1835 cur_trans = trans->transaction;
1836 atomic_inc(&cur_trans->use_count);
1837
1838 btrfs_end_transaction(trans);
1839
1840 /*
1841 * Tell lockdep we've released the freeze rwsem, since the
1842 * async commit thread will be the one to unlock it.
1843 */
1844 if (ac->newtrans->type & __TRANS_FREEZABLE)
1845 __sb_writers_release(fs_info->sb, SB_FREEZE_FS);
1846
1847 schedule_work(&ac->work);
1848
1849 /* wait for transaction to start and unblock */
1850 if (wait_for_unblock)
1851 wait_current_trans_commit_start_and_unblock(fs_info, cur_trans);
1852 else
1853 wait_current_trans_commit_start(fs_info, cur_trans);
1854
1855 if (current->journal_info == trans)
1856 current->journal_info = NULL;
1857
1858 btrfs_put_transaction(cur_trans);
1859 return 0;
1860}
1861
1862
1863static void cleanup_transaction(struct btrfs_trans_handle *trans,
1864 struct btrfs_root *root, int err)
1865{
1866 struct btrfs_fs_info *fs_info = root->fs_info;
1867 struct btrfs_transaction *cur_trans = trans->transaction;
1868 DEFINE_WAIT(wait);
1869
1870 WARN_ON(trans->use_count > 1);
1871
1872 btrfs_abort_transaction(trans, err);
1873
1874 spin_lock(&fs_info->trans_lock);
1875
1876 /*
1877 * If the transaction is removed from the list, it means this
1878 * transaction has been committed successfully, so it is impossible
1879 * to call the cleanup function.
1880 */
1881 BUG_ON(list_empty(&cur_trans->list));
1882
1883 list_del_init(&cur_trans->list);
1884 if (cur_trans == fs_info->running_transaction) {
1885 cur_trans->state = TRANS_STATE_COMMIT_DOING;
1886 spin_unlock(&fs_info->trans_lock);
1887 wait_event(cur_trans->writer_wait,
1888 atomic_read(&cur_trans->num_writers) == 1);
1889
1890 spin_lock(&fs_info->trans_lock);
1891 }
1892 spin_unlock(&fs_info->trans_lock);
1893
1894 btrfs_cleanup_one_transaction(trans->transaction, fs_info);
1895
1896 spin_lock(&fs_info->trans_lock);
1897 if (cur_trans == fs_info->running_transaction)
1898 fs_info->running_transaction = NULL;
1899 spin_unlock(&fs_info->trans_lock);
1900
1901 if (trans->type & __TRANS_FREEZABLE)
1902 sb_end_intwrite(fs_info->sb);
1903 btrfs_put_transaction(cur_trans);
1904 btrfs_put_transaction(cur_trans);
1905
1906 trace_btrfs_transaction_commit(root);
1907
1908 if (current->journal_info == trans)
1909 current->journal_info = NULL;
1910 btrfs_scrub_cancel(fs_info);
1911
1912 kmem_cache_free(btrfs_trans_handle_cachep, trans);
1913}
1914
1915static inline int btrfs_start_delalloc_flush(struct btrfs_fs_info *fs_info)
1916{
1917 if (btrfs_test_opt(fs_info, FLUSHONCOMMIT))
1918 return btrfs_start_delalloc_roots(fs_info, 1, -1);
1919 return 0;
1920}
1921
1922static inline void btrfs_wait_delalloc_flush(struct btrfs_fs_info *fs_info)
1923{
1924 if (btrfs_test_opt(fs_info, FLUSHONCOMMIT))
1925 btrfs_wait_ordered_roots(fs_info, -1, 0, (u64)-1);
1926}
1927
1928static inline void
1929btrfs_wait_pending_ordered(struct btrfs_transaction *cur_trans)
1930{
1931 wait_event(cur_trans->pending_wait,
1932 atomic_read(&cur_trans->pending_ordered) == 0);
1933}
1934
1935int btrfs_commit_transaction(struct btrfs_trans_handle *trans)
1936{
1937 struct btrfs_fs_info *fs_info = trans->fs_info;
1938 struct btrfs_transaction *cur_trans = trans->transaction;
1939 struct btrfs_transaction *prev_trans = NULL;
1940 int ret;
1941
1942 /* Stop the commit early if ->aborted is set */
1943 if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
1944 ret = cur_trans->aborted;
1945 btrfs_end_transaction(trans);
1946 return ret;
1947 }
1948
1949 /* make a pass through all the delayed refs we have so far
1950 * any runnings procs may add more while we are here
1951 */
1952 ret = btrfs_run_delayed_refs(trans, fs_info, 0);
1953 if (ret) {
1954 btrfs_end_transaction(trans);
1955 return ret;
1956 }
1957
1958 btrfs_trans_release_metadata(trans, fs_info);
1959 trans->block_rsv = NULL;
1960
1961 cur_trans = trans->transaction;
1962
1963 /*
1964 * set the flushing flag so procs in this transaction have to
1965 * start sending their work down.
1966 */
1967 cur_trans->delayed_refs.flushing = 1;
1968 smp_wmb();
1969
1970 if (!list_empty(&trans->new_bgs))
1971 btrfs_create_pending_block_groups(trans, fs_info);
1972
1973 ret = btrfs_run_delayed_refs(trans, fs_info, 0);
1974 if (ret) {
1975 btrfs_end_transaction(trans);
1976 return ret;
1977 }
1978
1979 if (!test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &cur_trans->flags)) {
1980 int run_it = 0;
1981
1982 /* this mutex is also taken before trying to set
1983 * block groups readonly. We need to make sure
1984 * that nobody has set a block group readonly
1985 * after a extents from that block group have been
1986 * allocated for cache files. btrfs_set_block_group_ro
1987 * will wait for the transaction to commit if it
1988 * finds BTRFS_TRANS_DIRTY_BG_RUN set.
1989 *
1990 * The BTRFS_TRANS_DIRTY_BG_RUN flag is also used to make sure
1991 * only one process starts all the block group IO. It wouldn't
1992 * hurt to have more than one go through, but there's no
1993 * real advantage to it either.
1994 */
1995 mutex_lock(&fs_info->ro_block_group_mutex);
1996 if (!test_and_set_bit(BTRFS_TRANS_DIRTY_BG_RUN,
1997 &cur_trans->flags))
1998 run_it = 1;
1999 mutex_unlock(&fs_info->ro_block_group_mutex);
2000
2001 if (run_it)
2002 ret = btrfs_start_dirty_block_groups(trans, fs_info);
2003 }
2004 if (ret) {
2005 btrfs_end_transaction(trans);
2006 return ret;
2007 }
2008
2009 spin_lock(&fs_info->trans_lock);
2010 if (cur_trans->state >= TRANS_STATE_COMMIT_START) {
2011 spin_unlock(&fs_info->trans_lock);
2012 atomic_inc(&cur_trans->use_count);
2013 ret = btrfs_end_transaction(trans);
2014
2015 wait_for_commit(cur_trans);
2016
2017 if (unlikely(cur_trans->aborted))
2018 ret = cur_trans->aborted;
2019
2020 btrfs_put_transaction(cur_trans);
2021
2022 return ret;
2023 }
2024
2025 cur_trans->state = TRANS_STATE_COMMIT_START;
2026 wake_up(&fs_info->transaction_blocked_wait);
2027
2028 if (cur_trans->list.prev != &fs_info->trans_list) {
2029 prev_trans = list_entry(cur_trans->list.prev,
2030 struct btrfs_transaction, list);
2031 if (prev_trans->state != TRANS_STATE_COMPLETED) {
2032 atomic_inc(&prev_trans->use_count);
2033 spin_unlock(&fs_info->trans_lock);
2034
2035 wait_for_commit(prev_trans);
2036 ret = prev_trans->aborted;
2037
2038 btrfs_put_transaction(prev_trans);
2039 if (ret)
2040 goto cleanup_transaction;
2041 } else {
2042 spin_unlock(&fs_info->trans_lock);
2043 }
2044 } else {
2045 spin_unlock(&fs_info->trans_lock);
2046 }
2047
2048 extwriter_counter_dec(cur_trans, trans->type);
2049
2050 ret = btrfs_start_delalloc_flush(fs_info);
2051 if (ret)
2052 goto cleanup_transaction;
2053
2054 ret = btrfs_run_delayed_items(trans, fs_info);
2055 if (ret)
2056 goto cleanup_transaction;
2057
2058 wait_event(cur_trans->writer_wait,
2059 extwriter_counter_read(cur_trans) == 0);
2060
2061 /* some pending stuffs might be added after the previous flush. */
2062 ret = btrfs_run_delayed_items(trans, fs_info);
2063 if (ret)
2064 goto cleanup_transaction;
2065
2066 btrfs_wait_delalloc_flush(fs_info);
2067
2068 btrfs_wait_pending_ordered(cur_trans);
2069
2070 btrfs_scrub_pause(fs_info);
2071 /*
2072 * Ok now we need to make sure to block out any other joins while we
2073 * commit the transaction. We could have started a join before setting
2074 * COMMIT_DOING so make sure to wait for num_writers to == 1 again.
2075 */
2076 spin_lock(&fs_info->trans_lock);
2077 cur_trans->state = TRANS_STATE_COMMIT_DOING;
2078 spin_unlock(&fs_info->trans_lock);
2079 wait_event(cur_trans->writer_wait,
2080 atomic_read(&cur_trans->num_writers) == 1);
2081
2082 /* ->aborted might be set after the previous check, so check it */
2083 if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
2084 ret = cur_trans->aborted;
2085 goto scrub_continue;
2086 }
2087 /*
2088 * the reloc mutex makes sure that we stop
2089 * the balancing code from coming in and moving
2090 * extents around in the middle of the commit
2091 */
2092 mutex_lock(&fs_info->reloc_mutex);
2093
2094 /*
2095 * We needn't worry about the delayed items because we will
2096 * deal with them in create_pending_snapshot(), which is the
2097 * core function of the snapshot creation.
2098 */
2099 ret = create_pending_snapshots(trans, fs_info);
2100 if (ret) {
2101 mutex_unlock(&fs_info->reloc_mutex);
2102 goto scrub_continue;
2103 }
2104
2105 /*
2106 * We insert the dir indexes of the snapshots and update the inode
2107 * of the snapshots' parents after the snapshot creation, so there
2108 * are some delayed items which are not dealt with. Now deal with
2109 * them.
2110 *
2111 * We needn't worry that this operation will corrupt the snapshots,
2112 * because all the tree which are snapshoted will be forced to COW
2113 * the nodes and leaves.
2114 */
2115 ret = btrfs_run_delayed_items(trans, fs_info);
2116 if (ret) {
2117 mutex_unlock(&fs_info->reloc_mutex);
2118 goto scrub_continue;
2119 }
2120
2121 ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
2122 if (ret) {
2123 mutex_unlock(&fs_info->reloc_mutex);
2124 goto scrub_continue;
2125 }
2126
2127 /* Reocrd old roots for later qgroup accounting */
2128 ret = btrfs_qgroup_prepare_account_extents(trans, fs_info);
2129 if (ret) {
2130 mutex_unlock(&fs_info->reloc_mutex);
2131 goto scrub_continue;
2132 }
2133
2134 /*
2135 * make sure none of the code above managed to slip in a
2136 * delayed item
2137 */
2138 btrfs_assert_delayed_root_empty(fs_info);
2139
2140 WARN_ON(cur_trans != trans->transaction);
2141
2142 /* btrfs_commit_tree_roots is responsible for getting the
2143 * various roots consistent with each other. Every pointer
2144 * in the tree of tree roots has to point to the most up to date
2145 * root for every subvolume and other tree. So, we have to keep
2146 * the tree logging code from jumping in and changing any
2147 * of the trees.
2148 *
2149 * At this point in the commit, there can't be any tree-log
2150 * writers, but a little lower down we drop the trans mutex
2151 * and let new people in. By holding the tree_log_mutex
2152 * from now until after the super is written, we avoid races
2153 * with the tree-log code.
2154 */
2155 mutex_lock(&fs_info->tree_log_mutex);
2156
2157 ret = commit_fs_roots(trans, fs_info);
2158 if (ret) {
2159 mutex_unlock(&fs_info->tree_log_mutex);
2160 mutex_unlock(&fs_info->reloc_mutex);
2161 goto scrub_continue;
2162 }
2163
2164 /*
2165 * Since the transaction is done, we can apply the pending changes
2166 * before the next transaction.
2167 */
2168 btrfs_apply_pending_changes(fs_info);
2169
2170 /* commit_fs_roots gets rid of all the tree log roots, it is now
2171 * safe to free the root of tree log roots
2172 */
2173 btrfs_free_log_root_tree(trans, fs_info);
2174
2175 /*
2176 * Since fs roots are all committed, we can get a quite accurate
2177 * new_roots. So let's do quota accounting.
2178 */
2179 ret = btrfs_qgroup_account_extents(trans, fs_info);
2180 if (ret < 0) {
2181 mutex_unlock(&fs_info->tree_log_mutex);
2182 mutex_unlock(&fs_info->reloc_mutex);
2183 goto scrub_continue;
2184 }
2185
2186 ret = commit_cowonly_roots(trans, fs_info);
2187 if (ret) {
2188 mutex_unlock(&fs_info->tree_log_mutex);
2189 mutex_unlock(&fs_info->reloc_mutex);
2190 goto scrub_continue;
2191 }
2192
2193 /*
2194 * The tasks which save the space cache and inode cache may also
2195 * update ->aborted, check it.
2196 */
2197 if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
2198 ret = cur_trans->aborted;
2199 mutex_unlock(&fs_info->tree_log_mutex);
2200 mutex_unlock(&fs_info->reloc_mutex);
2201 goto scrub_continue;
2202 }
2203
2204 btrfs_prepare_extent_commit(trans, fs_info);
2205
2206 cur_trans = fs_info->running_transaction;
2207
2208 btrfs_set_root_node(&fs_info->tree_root->root_item,
2209 fs_info->tree_root->node);
2210 list_add_tail(&fs_info->tree_root->dirty_list,
2211 &cur_trans->switch_commits);
2212
2213 btrfs_set_root_node(&fs_info->chunk_root->root_item,
2214 fs_info->chunk_root->node);
2215 list_add_tail(&fs_info->chunk_root->dirty_list,
2216 &cur_trans->switch_commits);
2217
2218 switch_commit_roots(cur_trans, fs_info);
2219
2220 assert_qgroups_uptodate(trans);
2221 ASSERT(list_empty(&cur_trans->dirty_bgs));
2222 ASSERT(list_empty(&cur_trans->io_bgs));
2223 update_super_roots(fs_info);
2224
2225 btrfs_set_super_log_root(fs_info->super_copy, 0);
2226 btrfs_set_super_log_root_level(fs_info->super_copy, 0);
2227 memcpy(fs_info->super_for_commit, fs_info->super_copy,
2228 sizeof(*fs_info->super_copy));
2229
2230 btrfs_update_commit_device_size(fs_info);
2231 btrfs_update_commit_device_bytes_used(fs_info, cur_trans);
2232
2233 clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags);
2234 clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags);
2235
2236 btrfs_trans_release_chunk_metadata(trans);
2237
2238 spin_lock(&fs_info->trans_lock);
2239 cur_trans->state = TRANS_STATE_UNBLOCKED;
2240 fs_info->running_transaction = NULL;
2241 spin_unlock(&fs_info->trans_lock);
2242 mutex_unlock(&fs_info->reloc_mutex);
2243
2244 wake_up(&fs_info->transaction_wait);
2245
2246 ret = btrfs_write_and_wait_transaction(trans, fs_info);
2247 if (ret) {
2248 btrfs_handle_fs_error(fs_info, ret,
2249 "Error while writing out transaction");
2250 mutex_unlock(&fs_info->tree_log_mutex);
2251 goto scrub_continue;
2252 }
2253
2254 ret = write_ctree_super(trans, fs_info, 0);
2255 if (ret) {
2256 mutex_unlock(&fs_info->tree_log_mutex);
2257 goto scrub_continue;
2258 }
2259
2260 /*
2261 * the super is written, we can safely allow the tree-loggers
2262 * to go about their business
2263 */
2264 mutex_unlock(&fs_info->tree_log_mutex);
2265
2266 btrfs_finish_extent_commit(trans, fs_info);
2267
2268 if (test_bit(BTRFS_TRANS_HAVE_FREE_BGS, &cur_trans->flags))
2269 btrfs_clear_space_info_full(fs_info);
2270
2271 fs_info->last_trans_committed = cur_trans->transid;
2272 /*
2273 * We needn't acquire the lock here because there is no other task
2274 * which can change it.
2275 */
2276 cur_trans->state = TRANS_STATE_COMPLETED;
2277 wake_up(&cur_trans->commit_wait);
2278
2279 spin_lock(&fs_info->trans_lock);
2280 list_del_init(&cur_trans->list);
2281 spin_unlock(&fs_info->trans_lock);
2282
2283 btrfs_put_transaction(cur_trans);
2284 btrfs_put_transaction(cur_trans);
2285
2286 if (trans->type & __TRANS_FREEZABLE)
2287 sb_end_intwrite(fs_info->sb);
2288
2289 trace_btrfs_transaction_commit(trans->root);
2290
2291 btrfs_scrub_continue(fs_info);
2292
2293 if (current->journal_info == trans)
2294 current->journal_info = NULL;
2295
2296 kmem_cache_free(btrfs_trans_handle_cachep, trans);
2297
2298 /*
2299 * If fs has been frozen, we can not handle delayed iputs, otherwise
2300 * it'll result in deadlock about SB_FREEZE_FS.
2301 */
2302 if (current != fs_info->transaction_kthread &&
2303 current != fs_info->cleaner_kthread && !fs_info->fs_frozen)
2304 btrfs_run_delayed_iputs(fs_info);
2305
2306 return ret;
2307
2308scrub_continue:
2309 btrfs_scrub_continue(fs_info);
2310cleanup_transaction:
2311 btrfs_trans_release_metadata(trans, fs_info);
2312 btrfs_trans_release_chunk_metadata(trans);
2313 trans->block_rsv = NULL;
2314 btrfs_warn(fs_info, "Skipping commit of aborted transaction.");
2315 if (current->journal_info == trans)
2316 current->journal_info = NULL;
2317 cleanup_transaction(trans, trans->root, ret);
2318
2319 return ret;
2320}
2321
2322/*
2323 * return < 0 if error
2324 * 0 if there are no more dead_roots at the time of call
2325 * 1 there are more to be processed, call me again
2326 *
2327 * The return value indicates there are certainly more snapshots to delete, but
2328 * if there comes a new one during processing, it may return 0. We don't mind,
2329 * because btrfs_commit_super will poke cleaner thread and it will process it a
2330 * few seconds later.
2331 */
2332int btrfs_clean_one_deleted_snapshot(struct btrfs_root *root)
2333{
2334 int ret;
2335 struct btrfs_fs_info *fs_info = root->fs_info;
2336
2337 spin_lock(&fs_info->trans_lock);
2338 if (list_empty(&fs_info->dead_roots)) {
2339 spin_unlock(&fs_info->trans_lock);
2340 return 0;
2341 }
2342 root = list_first_entry(&fs_info->dead_roots,
2343 struct btrfs_root, root_list);
2344 list_del_init(&root->root_list);
2345 spin_unlock(&fs_info->trans_lock);
2346
2347 btrfs_debug(fs_info, "cleaner removing %llu", root->objectid);
2348
2349 btrfs_kill_all_delayed_nodes(root);
2350
2351 if (btrfs_header_backref_rev(root->node) <
2352 BTRFS_MIXED_BACKREF_REV)
2353 ret = btrfs_drop_snapshot(root, NULL, 0, 0);
2354 else
2355 ret = btrfs_drop_snapshot(root, NULL, 1, 0);
2356
2357 return (ret < 0) ? 0 : 1;
2358}
2359
2360void btrfs_apply_pending_changes(struct btrfs_fs_info *fs_info)
2361{
2362 unsigned long prev;
2363 unsigned long bit;
2364
2365 prev = xchg(&fs_info->pending_changes, 0);
2366 if (!prev)
2367 return;
2368
2369 bit = 1 << BTRFS_PENDING_SET_INODE_MAP_CACHE;
2370 if (prev & bit)
2371 btrfs_set_opt(fs_info->mount_opt, INODE_MAP_CACHE);
2372 prev &= ~bit;
2373
2374 bit = 1 << BTRFS_PENDING_CLEAR_INODE_MAP_CACHE;
2375 if (prev & bit)
2376 btrfs_clear_opt(fs_info->mount_opt, INODE_MAP_CACHE);
2377 prev &= ~bit;
2378
2379 bit = 1 << BTRFS_PENDING_COMMIT;
2380 if (prev & bit)
2381 btrfs_debug(fs_info, "pending commit done");
2382 prev &= ~bit;
2383
2384 if (prev)
2385 btrfs_warn(fs_info,
2386 "unknown pending changes left 0x%lx, ignoring", prev);
2387}