xref: /linux/fs/btrfs/transaction.c (revision 2d7f3d1a5866705be2393150e1ffdf67030ab88d)
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 
38 static struct kmem_cache *btrfs_trans_handle_cachep;
39 
40 /*
41  * Transaction states and transitions
42  *
43  * No running transaction (fs tree blocks are not modified)
44  * |
45  * | To next stage:
46  * |  Call start_transaction() variants. Except btrfs_join_transaction_nostart().
47  * V
48  * Transaction N [[TRANS_STATE_RUNNING]]
49  * |
50  * | New trans handles can be attached to transaction N by calling all
51  * | start_transaction() variants.
52  * |
53  * | To next stage:
54  * |  Call btrfs_commit_transaction() on any trans handle attached to
55  * |  transaction N
56  * V
57  * Transaction N [[TRANS_STATE_COMMIT_PREP]]
58  * |
59  * | If there are simultaneous calls to btrfs_commit_transaction() one will win
60  * | the race and the rest will wait for the winner to commit the transaction.
61  * |
62  * | The winner will wait for previous running transaction to completely finish
63  * | if there is one.
64  * |
65  * Transaction N [[TRANS_STATE_COMMIT_START]]
66  * |
67  * | Then one of the following happens:
68  * | - Wait for all other trans handle holders to release.
69  * |   The btrfs_commit_transaction() caller will do the commit work.
70  * | - Wait for current transaction to be committed by others.
71  * |   Other btrfs_commit_transaction() caller will do the commit work.
72  * |
73  * | At this stage, only btrfs_join_transaction*() variants can attach
74  * | to this running transaction.
75  * | All other variants will wait for current one to finish and attach to
76  * | transaction N+1.
77  * |
78  * | To next stage:
79  * |  Caller is chosen to commit transaction N, and all other trans handle
80  * |  haven been released.
81  * V
82  * Transaction N [[TRANS_STATE_COMMIT_DOING]]
83  * |
84  * | The heavy lifting transaction work is started.
85  * | From running delayed refs (modifying extent tree) to creating pending
86  * | snapshots, running qgroups.
87  * | In short, modify supporting trees to reflect modifications of subvolume
88  * | trees.
89  * |
90  * | At this stage, all start_transaction() calls will wait for this
91  * | transaction to finish and attach to transaction N+1.
92  * |
93  * | To next stage:
94  * |  Until all supporting trees are updated.
95  * V
96  * Transaction N [[TRANS_STATE_UNBLOCKED]]
97  * |						    Transaction N+1
98  * | All needed trees are modified, thus we only    [[TRANS_STATE_RUNNING]]
99  * | need to write them back to disk and update	    |
100  * | super blocks.				    |
101  * |						    |
102  * | At this stage, new transaction is allowed to   |
103  * | start.					    |
104  * | All new start_transaction() calls will be	    |
105  * | attached to transid N+1.			    |
106  * |						    |
107  * | To next stage:				    |
108  * |  Until all tree blocks are super blocks are    |
109  * |  written to block devices			    |
110  * V						    |
111  * Transaction N [[TRANS_STATE_COMPLETED]]	    V
112  *   All tree blocks and super blocks are written.  Transaction N+1
113  *   This transaction is finished and all its	    [[TRANS_STATE_COMMIT_START]]
114  *   data structures will be cleaned up.	    | Life goes on
115  */
116 static const unsigned int btrfs_blocked_trans_types[TRANS_STATE_MAX] = {
117 	[TRANS_STATE_RUNNING]		= 0U,
118 	[TRANS_STATE_COMMIT_PREP]	= 0U,
119 	[TRANS_STATE_COMMIT_START]	= (__TRANS_START | __TRANS_ATTACH),
120 	[TRANS_STATE_COMMIT_DOING]	= (__TRANS_START |
121 					   __TRANS_ATTACH |
122 					   __TRANS_JOIN |
123 					   __TRANS_JOIN_NOSTART),
124 	[TRANS_STATE_UNBLOCKED]		= (__TRANS_START |
125 					   __TRANS_ATTACH |
126 					   __TRANS_JOIN |
127 					   __TRANS_JOIN_NOLOCK |
128 					   __TRANS_JOIN_NOSTART),
129 	[TRANS_STATE_SUPER_COMMITTED]	= (__TRANS_START |
130 					   __TRANS_ATTACH |
131 					   __TRANS_JOIN |
132 					   __TRANS_JOIN_NOLOCK |
133 					   __TRANS_JOIN_NOSTART),
134 	[TRANS_STATE_COMPLETED]		= (__TRANS_START |
135 					   __TRANS_ATTACH |
136 					   __TRANS_JOIN |
137 					   __TRANS_JOIN_NOLOCK |
138 					   __TRANS_JOIN_NOSTART),
139 };
140 
141 void btrfs_put_transaction(struct btrfs_transaction *transaction)
142 {
143 	WARN_ON(refcount_read(&transaction->use_count) == 0);
144 	if (refcount_dec_and_test(&transaction->use_count)) {
145 		BUG_ON(!list_empty(&transaction->list));
146 		WARN_ON(!RB_EMPTY_ROOT(
147 				&transaction->delayed_refs.href_root.rb_root));
148 		WARN_ON(!RB_EMPTY_ROOT(
149 				&transaction->delayed_refs.dirty_extent_root));
150 		if (transaction->delayed_refs.pending_csums)
151 			btrfs_err(transaction->fs_info,
152 				  "pending csums is %llu",
153 				  transaction->delayed_refs.pending_csums);
154 		/*
155 		 * If any block groups are found in ->deleted_bgs then it's
156 		 * because the transaction was aborted and a commit did not
157 		 * happen (things failed before writing the new superblock
158 		 * and calling btrfs_finish_extent_commit()), so we can not
159 		 * discard the physical locations of the block groups.
160 		 */
161 		while (!list_empty(&transaction->deleted_bgs)) {
162 			struct btrfs_block_group *cache;
163 
164 			cache = list_first_entry(&transaction->deleted_bgs,
165 						 struct btrfs_block_group,
166 						 bg_list);
167 			list_del_init(&cache->bg_list);
168 			btrfs_unfreeze_block_group(cache);
169 			btrfs_put_block_group(cache);
170 		}
171 		WARN_ON(!list_empty(&transaction->dev_update_list));
172 		kfree(transaction);
173 	}
174 }
175 
176 static noinline void switch_commit_roots(struct btrfs_trans_handle *trans)
177 {
178 	struct btrfs_transaction *cur_trans = trans->transaction;
179 	struct btrfs_fs_info *fs_info = trans->fs_info;
180 	struct btrfs_root *root, *tmp;
181 
182 	/*
183 	 * At this point no one can be using this transaction to modify any tree
184 	 * and no one can start another transaction to modify any tree either.
185 	 */
186 	ASSERT(cur_trans->state == TRANS_STATE_COMMIT_DOING);
187 
188 	down_write(&fs_info->commit_root_sem);
189 
190 	if (test_bit(BTRFS_FS_RELOC_RUNNING, &fs_info->flags))
191 		fs_info->last_reloc_trans = trans->transid;
192 
193 	list_for_each_entry_safe(root, tmp, &cur_trans->switch_commits,
194 				 dirty_list) {
195 		list_del_init(&root->dirty_list);
196 		free_extent_buffer(root->commit_root);
197 		root->commit_root = btrfs_root_node(root);
198 		extent_io_tree_release(&root->dirty_log_pages);
199 		btrfs_qgroup_clean_swapped_blocks(root);
200 	}
201 
202 	/* We can free old roots now. */
203 	spin_lock(&cur_trans->dropped_roots_lock);
204 	while (!list_empty(&cur_trans->dropped_roots)) {
205 		root = list_first_entry(&cur_trans->dropped_roots,
206 					struct btrfs_root, root_list);
207 		list_del_init(&root->root_list);
208 		spin_unlock(&cur_trans->dropped_roots_lock);
209 		btrfs_free_log(trans, root);
210 		btrfs_drop_and_free_fs_root(fs_info, root);
211 		spin_lock(&cur_trans->dropped_roots_lock);
212 	}
213 	spin_unlock(&cur_trans->dropped_roots_lock);
214 
215 	up_write(&fs_info->commit_root_sem);
216 }
217 
218 static inline void extwriter_counter_inc(struct btrfs_transaction *trans,
219 					 unsigned int type)
220 {
221 	if (type & TRANS_EXTWRITERS)
222 		atomic_inc(&trans->num_extwriters);
223 }
224 
225 static inline void extwriter_counter_dec(struct btrfs_transaction *trans,
226 					 unsigned int type)
227 {
228 	if (type & TRANS_EXTWRITERS)
229 		atomic_dec(&trans->num_extwriters);
230 }
231 
232 static inline void extwriter_counter_init(struct btrfs_transaction *trans,
233 					  unsigned int type)
234 {
235 	atomic_set(&trans->num_extwriters, ((type & TRANS_EXTWRITERS) ? 1 : 0));
236 }
237 
238 static inline int extwriter_counter_read(struct btrfs_transaction *trans)
239 {
240 	return atomic_read(&trans->num_extwriters);
241 }
242 
243 /*
244  * To be called after doing the chunk btree updates right after allocating a new
245  * chunk (after btrfs_chunk_alloc_add_chunk_item() is called), when removing a
246  * chunk after all chunk btree updates and after finishing the second phase of
247  * chunk allocation (btrfs_create_pending_block_groups()) in case some block
248  * group had its chunk item insertion delayed to the second phase.
249  */
250 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
251 {
252 	struct btrfs_fs_info *fs_info = trans->fs_info;
253 
254 	if (!trans->chunk_bytes_reserved)
255 		return;
256 
257 	btrfs_block_rsv_release(fs_info, &fs_info->chunk_block_rsv,
258 				trans->chunk_bytes_reserved, NULL);
259 	trans->chunk_bytes_reserved = 0;
260 }
261 
262 /*
263  * either allocate a new transaction or hop into the existing one
264  */
265 static noinline int join_transaction(struct btrfs_fs_info *fs_info,
266 				     unsigned int type)
267 {
268 	struct btrfs_transaction *cur_trans;
269 
270 	spin_lock(&fs_info->trans_lock);
271 loop:
272 	/* The file system has been taken offline. No new transactions. */
273 	if (BTRFS_FS_ERROR(fs_info)) {
274 		spin_unlock(&fs_info->trans_lock);
275 		return -EROFS;
276 	}
277 
278 	cur_trans = fs_info->running_transaction;
279 	if (cur_trans) {
280 		if (TRANS_ABORTED(cur_trans)) {
281 			spin_unlock(&fs_info->trans_lock);
282 			return cur_trans->aborted;
283 		}
284 		if (btrfs_blocked_trans_types[cur_trans->state] & type) {
285 			spin_unlock(&fs_info->trans_lock);
286 			return -EBUSY;
287 		}
288 		refcount_inc(&cur_trans->use_count);
289 		atomic_inc(&cur_trans->num_writers);
290 		extwriter_counter_inc(cur_trans, type);
291 		spin_unlock(&fs_info->trans_lock);
292 		btrfs_lockdep_acquire(fs_info, btrfs_trans_num_writers);
293 		btrfs_lockdep_acquire(fs_info, btrfs_trans_num_extwriters);
294 		return 0;
295 	}
296 	spin_unlock(&fs_info->trans_lock);
297 
298 	/*
299 	 * If we are ATTACH or TRANS_JOIN_NOSTART, we just want to catch the
300 	 * current transaction, and commit it. If there is no transaction, just
301 	 * return ENOENT.
302 	 */
303 	if (type == TRANS_ATTACH || type == TRANS_JOIN_NOSTART)
304 		return -ENOENT;
305 
306 	/*
307 	 * JOIN_NOLOCK only happens during the transaction commit, so
308 	 * it is impossible that ->running_transaction is NULL
309 	 */
310 	BUG_ON(type == TRANS_JOIN_NOLOCK);
311 
312 	cur_trans = kmalloc(sizeof(*cur_trans), GFP_NOFS);
313 	if (!cur_trans)
314 		return -ENOMEM;
315 
316 	btrfs_lockdep_acquire(fs_info, btrfs_trans_num_writers);
317 	btrfs_lockdep_acquire(fs_info, btrfs_trans_num_extwriters);
318 
319 	spin_lock(&fs_info->trans_lock);
320 	if (fs_info->running_transaction) {
321 		/*
322 		 * someone started a transaction after we unlocked.  Make sure
323 		 * to redo the checks above
324 		 */
325 		btrfs_lockdep_release(fs_info, btrfs_trans_num_extwriters);
326 		btrfs_lockdep_release(fs_info, btrfs_trans_num_writers);
327 		kfree(cur_trans);
328 		goto loop;
329 	} else if (BTRFS_FS_ERROR(fs_info)) {
330 		spin_unlock(&fs_info->trans_lock);
331 		btrfs_lockdep_release(fs_info, btrfs_trans_num_extwriters);
332 		btrfs_lockdep_release(fs_info, btrfs_trans_num_writers);
333 		kfree(cur_trans);
334 		return -EROFS;
335 	}
336 
337 	cur_trans->fs_info = fs_info;
338 	atomic_set(&cur_trans->pending_ordered, 0);
339 	init_waitqueue_head(&cur_trans->pending_wait);
340 	atomic_set(&cur_trans->num_writers, 1);
341 	extwriter_counter_init(cur_trans, type);
342 	init_waitqueue_head(&cur_trans->writer_wait);
343 	init_waitqueue_head(&cur_trans->commit_wait);
344 	cur_trans->state = TRANS_STATE_RUNNING;
345 	/*
346 	 * One for this trans handle, one so it will live on until we
347 	 * commit the transaction.
348 	 */
349 	refcount_set(&cur_trans->use_count, 2);
350 	cur_trans->flags = 0;
351 	cur_trans->start_time = ktime_get_seconds();
352 
353 	memset(&cur_trans->delayed_refs, 0, sizeof(cur_trans->delayed_refs));
354 
355 	cur_trans->delayed_refs.href_root = RB_ROOT_CACHED;
356 	cur_trans->delayed_refs.dirty_extent_root = RB_ROOT;
357 	atomic_set(&cur_trans->delayed_refs.num_entries, 0);
358 
359 	/*
360 	 * although the tree mod log is per file system and not per transaction,
361 	 * the log must never go across transaction boundaries.
362 	 */
363 	smp_mb();
364 	if (!list_empty(&fs_info->tree_mod_seq_list))
365 		WARN(1, KERN_ERR "BTRFS: tree_mod_seq_list not empty when creating a fresh transaction\n");
366 	if (!RB_EMPTY_ROOT(&fs_info->tree_mod_log))
367 		WARN(1, KERN_ERR "BTRFS: tree_mod_log rb tree not empty when creating a fresh transaction\n");
368 	atomic64_set(&fs_info->tree_mod_seq, 0);
369 
370 	spin_lock_init(&cur_trans->delayed_refs.lock);
371 
372 	INIT_LIST_HEAD(&cur_trans->pending_snapshots);
373 	INIT_LIST_HEAD(&cur_trans->dev_update_list);
374 	INIT_LIST_HEAD(&cur_trans->switch_commits);
375 	INIT_LIST_HEAD(&cur_trans->dirty_bgs);
376 	INIT_LIST_HEAD(&cur_trans->io_bgs);
377 	INIT_LIST_HEAD(&cur_trans->dropped_roots);
378 	mutex_init(&cur_trans->cache_write_mutex);
379 	spin_lock_init(&cur_trans->dirty_bgs_lock);
380 	INIT_LIST_HEAD(&cur_trans->deleted_bgs);
381 	spin_lock_init(&cur_trans->dropped_roots_lock);
382 	list_add_tail(&cur_trans->list, &fs_info->trans_list);
383 	extent_io_tree_init(fs_info, &cur_trans->dirty_pages,
384 			IO_TREE_TRANS_DIRTY_PAGES);
385 	extent_io_tree_init(fs_info, &cur_trans->pinned_extents,
386 			IO_TREE_FS_PINNED_EXTENTS);
387 	btrfs_set_fs_generation(fs_info, fs_info->generation + 1);
388 	cur_trans->transid = fs_info->generation;
389 	fs_info->running_transaction = cur_trans;
390 	cur_trans->aborted = 0;
391 	spin_unlock(&fs_info->trans_lock);
392 
393 	return 0;
394 }
395 
396 /*
397  * This does all the record keeping required to make sure that a shareable root
398  * is properly recorded in a given transaction.  This is required to make sure
399  * the old root from before we joined the transaction is deleted when the
400  * transaction commits.
401  */
402 static int record_root_in_trans(struct btrfs_trans_handle *trans,
403 			       struct btrfs_root *root,
404 			       int force)
405 {
406 	struct btrfs_fs_info *fs_info = root->fs_info;
407 	int ret = 0;
408 
409 	if ((test_bit(BTRFS_ROOT_SHAREABLE, &root->state) &&
410 	    root->last_trans < trans->transid) || force) {
411 		WARN_ON(!force && root->commit_root != root->node);
412 
413 		/*
414 		 * see below for IN_TRANS_SETUP usage rules
415 		 * we have the reloc mutex held now, so there
416 		 * is only one writer in this function
417 		 */
418 		set_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
419 
420 		/* make sure readers find IN_TRANS_SETUP before
421 		 * they find our root->last_trans update
422 		 */
423 		smp_wmb();
424 
425 		spin_lock(&fs_info->fs_roots_radix_lock);
426 		if (root->last_trans == trans->transid && !force) {
427 			spin_unlock(&fs_info->fs_roots_radix_lock);
428 			return 0;
429 		}
430 		radix_tree_tag_set(&fs_info->fs_roots_radix,
431 				   (unsigned long)root->root_key.objectid,
432 				   BTRFS_ROOT_TRANS_TAG);
433 		spin_unlock(&fs_info->fs_roots_radix_lock);
434 		root->last_trans = trans->transid;
435 
436 		/* this is pretty tricky.  We don't want to
437 		 * take the relocation lock in btrfs_record_root_in_trans
438 		 * unless we're really doing the first setup for this root in
439 		 * this transaction.
440 		 *
441 		 * Normally we'd use root->last_trans as a flag to decide
442 		 * if we want to take the expensive mutex.
443 		 *
444 		 * But, we have to set root->last_trans before we
445 		 * init the relocation root, otherwise, we trip over warnings
446 		 * in ctree.c.  The solution used here is to flag ourselves
447 		 * with root IN_TRANS_SETUP.  When this is 1, we're still
448 		 * fixing up the reloc trees and everyone must wait.
449 		 *
450 		 * When this is zero, they can trust root->last_trans and fly
451 		 * through btrfs_record_root_in_trans without having to take the
452 		 * lock.  smp_wmb() makes sure that all the writes above are
453 		 * done before we pop in the zero below
454 		 */
455 		ret = btrfs_init_reloc_root(trans, root);
456 		smp_mb__before_atomic();
457 		clear_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
458 	}
459 	return ret;
460 }
461 
462 
463 void btrfs_add_dropped_root(struct btrfs_trans_handle *trans,
464 			    struct btrfs_root *root)
465 {
466 	struct btrfs_fs_info *fs_info = root->fs_info;
467 	struct btrfs_transaction *cur_trans = trans->transaction;
468 
469 	/* Add ourselves to the transaction dropped list */
470 	spin_lock(&cur_trans->dropped_roots_lock);
471 	list_add_tail(&root->root_list, &cur_trans->dropped_roots);
472 	spin_unlock(&cur_trans->dropped_roots_lock);
473 
474 	/* Make sure we don't try to update the root at commit time */
475 	spin_lock(&fs_info->fs_roots_radix_lock);
476 	radix_tree_tag_clear(&fs_info->fs_roots_radix,
477 			     (unsigned long)root->root_key.objectid,
478 			     BTRFS_ROOT_TRANS_TAG);
479 	spin_unlock(&fs_info->fs_roots_radix_lock);
480 }
481 
482 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
483 			       struct btrfs_root *root)
484 {
485 	struct btrfs_fs_info *fs_info = root->fs_info;
486 	int ret;
487 
488 	if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
489 		return 0;
490 
491 	/*
492 	 * see record_root_in_trans for comments about IN_TRANS_SETUP usage
493 	 * and barriers
494 	 */
495 	smp_rmb();
496 	if (root->last_trans == trans->transid &&
497 	    !test_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state))
498 		return 0;
499 
500 	mutex_lock(&fs_info->reloc_mutex);
501 	ret = record_root_in_trans(trans, root, 0);
502 	mutex_unlock(&fs_info->reloc_mutex);
503 
504 	return ret;
505 }
506 
507 static inline int is_transaction_blocked(struct btrfs_transaction *trans)
508 {
509 	return (trans->state >= TRANS_STATE_COMMIT_START &&
510 		trans->state < TRANS_STATE_UNBLOCKED &&
511 		!TRANS_ABORTED(trans));
512 }
513 
514 /* wait for commit against the current transaction to become unblocked
515  * when this is done, it is safe to start a new transaction, but the current
516  * transaction might not be fully on disk.
517  */
518 static void wait_current_trans(struct btrfs_fs_info *fs_info)
519 {
520 	struct btrfs_transaction *cur_trans;
521 
522 	spin_lock(&fs_info->trans_lock);
523 	cur_trans = fs_info->running_transaction;
524 	if (cur_trans && is_transaction_blocked(cur_trans)) {
525 		refcount_inc(&cur_trans->use_count);
526 		spin_unlock(&fs_info->trans_lock);
527 
528 		btrfs_might_wait_for_state(fs_info, BTRFS_LOCKDEP_TRANS_UNBLOCKED);
529 		wait_event(fs_info->transaction_wait,
530 			   cur_trans->state >= TRANS_STATE_UNBLOCKED ||
531 			   TRANS_ABORTED(cur_trans));
532 		btrfs_put_transaction(cur_trans);
533 	} else {
534 		spin_unlock(&fs_info->trans_lock);
535 	}
536 }
537 
538 static int may_wait_transaction(struct btrfs_fs_info *fs_info, int type)
539 {
540 	if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
541 		return 0;
542 
543 	if (type == TRANS_START)
544 		return 1;
545 
546 	return 0;
547 }
548 
549 static inline bool need_reserve_reloc_root(struct btrfs_root *root)
550 {
551 	struct btrfs_fs_info *fs_info = root->fs_info;
552 
553 	if (!fs_info->reloc_ctl ||
554 	    !test_bit(BTRFS_ROOT_SHAREABLE, &root->state) ||
555 	    root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
556 	    root->reloc_root)
557 		return false;
558 
559 	return true;
560 }
561 
562 static int btrfs_reserve_trans_metadata(struct btrfs_fs_info *fs_info,
563 					enum btrfs_reserve_flush_enum flush,
564 					u64 num_bytes,
565 					u64 *delayed_refs_bytes)
566 {
567 	struct btrfs_space_info *si = fs_info->trans_block_rsv.space_info;
568 	u64 bytes = num_bytes + *delayed_refs_bytes;
569 	int ret;
570 
571 	/*
572 	 * We want to reserve all the bytes we may need all at once, so we only
573 	 * do 1 enospc flushing cycle per transaction start.
574 	 */
575 	ret = btrfs_reserve_metadata_bytes(fs_info, si, bytes, flush);
576 
577 	/*
578 	 * If we are an emergency flush, which can steal from the global block
579 	 * reserve, then attempt to not reserve space for the delayed refs, as
580 	 * we will consume space for them from the global block reserve.
581 	 */
582 	if (ret && flush == BTRFS_RESERVE_FLUSH_ALL_STEAL) {
583 		bytes -= *delayed_refs_bytes;
584 		*delayed_refs_bytes = 0;
585 		ret = btrfs_reserve_metadata_bytes(fs_info, si, bytes, flush);
586 	}
587 
588 	return ret;
589 }
590 
591 static struct btrfs_trans_handle *
592 start_transaction(struct btrfs_root *root, unsigned int num_items,
593 		  unsigned int type, enum btrfs_reserve_flush_enum flush,
594 		  bool enforce_qgroups)
595 {
596 	struct btrfs_fs_info *fs_info = root->fs_info;
597 	struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
598 	struct btrfs_block_rsv *trans_rsv = &fs_info->trans_block_rsv;
599 	struct btrfs_trans_handle *h;
600 	struct btrfs_transaction *cur_trans;
601 	u64 num_bytes = 0;
602 	u64 qgroup_reserved = 0;
603 	u64 delayed_refs_bytes = 0;
604 	bool reloc_reserved = false;
605 	bool do_chunk_alloc = false;
606 	int ret;
607 
608 	if (BTRFS_FS_ERROR(fs_info))
609 		return ERR_PTR(-EROFS);
610 
611 	if (current->journal_info) {
612 		WARN_ON(type & TRANS_EXTWRITERS);
613 		h = current->journal_info;
614 		refcount_inc(&h->use_count);
615 		WARN_ON(refcount_read(&h->use_count) > 2);
616 		h->orig_rsv = h->block_rsv;
617 		h->block_rsv = NULL;
618 		goto got_it;
619 	}
620 
621 	/*
622 	 * Do the reservation before we join the transaction so we can do all
623 	 * the appropriate flushing if need be.
624 	 */
625 	if (num_items && root != fs_info->chunk_root) {
626 		qgroup_reserved = num_items * fs_info->nodesize;
627 		/*
628 		 * Use prealloc for now, as there might be a currently running
629 		 * transaction that could free this reserved space prematurely
630 		 * by committing.
631 		 */
632 		ret = btrfs_qgroup_reserve_meta_prealloc(root, qgroup_reserved,
633 							 enforce_qgroups, false);
634 		if (ret)
635 			return ERR_PTR(ret);
636 
637 		num_bytes = btrfs_calc_insert_metadata_size(fs_info, num_items);
638 		/*
639 		 * If we plan to insert/update/delete "num_items" from a btree,
640 		 * we will also generate delayed refs for extent buffers in the
641 		 * respective btree paths, so reserve space for the delayed refs
642 		 * that will be generated by the caller as it modifies btrees.
643 		 * Try to reserve them to avoid excessive use of the global
644 		 * block reserve.
645 		 */
646 		delayed_refs_bytes = btrfs_calc_delayed_ref_bytes(fs_info, num_items);
647 
648 		/*
649 		 * Do the reservation for the relocation root creation
650 		 */
651 		if (need_reserve_reloc_root(root)) {
652 			num_bytes += fs_info->nodesize;
653 			reloc_reserved = true;
654 		}
655 
656 		ret = btrfs_reserve_trans_metadata(fs_info, flush, num_bytes,
657 						   &delayed_refs_bytes);
658 		if (ret)
659 			goto reserve_fail;
660 
661 		btrfs_block_rsv_add_bytes(trans_rsv, num_bytes, true);
662 
663 		if (trans_rsv->space_info->force_alloc)
664 			do_chunk_alloc = true;
665 	} else if (num_items == 0 && flush == BTRFS_RESERVE_FLUSH_ALL &&
666 		   !btrfs_block_rsv_full(delayed_refs_rsv)) {
667 		/*
668 		 * Some people call with btrfs_start_transaction(root, 0)
669 		 * because they can be throttled, but have some other mechanism
670 		 * for reserving space.  We still want these guys to refill the
671 		 * delayed block_rsv so just add 1 items worth of reservation
672 		 * here.
673 		 */
674 		ret = btrfs_delayed_refs_rsv_refill(fs_info, flush);
675 		if (ret)
676 			goto reserve_fail;
677 	}
678 again:
679 	h = kmem_cache_zalloc(btrfs_trans_handle_cachep, GFP_NOFS);
680 	if (!h) {
681 		ret = -ENOMEM;
682 		goto alloc_fail;
683 	}
684 
685 	/*
686 	 * If we are JOIN_NOLOCK we're already committing a transaction and
687 	 * waiting on this guy, so we don't need to do the sb_start_intwrite
688 	 * because we're already holding a ref.  We need this because we could
689 	 * have raced in and did an fsync() on a file which can kick a commit
690 	 * and then we deadlock with somebody doing a freeze.
691 	 *
692 	 * If we are ATTACH, it means we just want to catch the current
693 	 * transaction and commit it, so we needn't do sb_start_intwrite().
694 	 */
695 	if (type & __TRANS_FREEZABLE)
696 		sb_start_intwrite(fs_info->sb);
697 
698 	if (may_wait_transaction(fs_info, type))
699 		wait_current_trans(fs_info);
700 
701 	do {
702 		ret = join_transaction(fs_info, type);
703 		if (ret == -EBUSY) {
704 			wait_current_trans(fs_info);
705 			if (unlikely(type == TRANS_ATTACH ||
706 				     type == TRANS_JOIN_NOSTART))
707 				ret = -ENOENT;
708 		}
709 	} while (ret == -EBUSY);
710 
711 	if (ret < 0)
712 		goto join_fail;
713 
714 	cur_trans = fs_info->running_transaction;
715 
716 	h->transid = cur_trans->transid;
717 	h->transaction = cur_trans;
718 	refcount_set(&h->use_count, 1);
719 	h->fs_info = root->fs_info;
720 
721 	h->type = type;
722 	INIT_LIST_HEAD(&h->new_bgs);
723 	btrfs_init_metadata_block_rsv(fs_info, &h->delayed_rsv, BTRFS_BLOCK_RSV_DELOPS);
724 
725 	smp_mb();
726 	if (cur_trans->state >= TRANS_STATE_COMMIT_START &&
727 	    may_wait_transaction(fs_info, type)) {
728 		current->journal_info = h;
729 		btrfs_commit_transaction(h);
730 		goto again;
731 	}
732 
733 	if (num_bytes) {
734 		trace_btrfs_space_reservation(fs_info, "transaction",
735 					      h->transid, num_bytes, 1);
736 		h->block_rsv = trans_rsv;
737 		h->bytes_reserved = num_bytes;
738 		if (delayed_refs_bytes > 0) {
739 			trace_btrfs_space_reservation(fs_info,
740 						      "local_delayed_refs_rsv",
741 						      h->transid,
742 						      delayed_refs_bytes, 1);
743 			h->delayed_refs_bytes_reserved = delayed_refs_bytes;
744 			btrfs_block_rsv_add_bytes(&h->delayed_rsv, delayed_refs_bytes, true);
745 			delayed_refs_bytes = 0;
746 		}
747 		h->reloc_reserved = reloc_reserved;
748 	}
749 
750 	/*
751 	 * Now that we have found a transaction to be a part of, convert the
752 	 * qgroup reservation from prealloc to pertrans. A different transaction
753 	 * can't race in and free our pertrans out from under us.
754 	 */
755 	if (qgroup_reserved)
756 		btrfs_qgroup_convert_reserved_meta(root, qgroup_reserved);
757 
758 got_it:
759 	if (!current->journal_info)
760 		current->journal_info = h;
761 
762 	/*
763 	 * If the space_info is marked ALLOC_FORCE then we'll get upgraded to
764 	 * ALLOC_FORCE the first run through, and then we won't allocate for
765 	 * anybody else who races in later.  We don't care about the return
766 	 * value here.
767 	 */
768 	if (do_chunk_alloc && num_bytes) {
769 		u64 flags = h->block_rsv->space_info->flags;
770 
771 		btrfs_chunk_alloc(h, btrfs_get_alloc_profile(fs_info, flags),
772 				  CHUNK_ALLOC_NO_FORCE);
773 	}
774 
775 	/*
776 	 * btrfs_record_root_in_trans() needs to alloc new extents, and may
777 	 * call btrfs_join_transaction() while we're also starting a
778 	 * transaction.
779 	 *
780 	 * Thus it need to be called after current->journal_info initialized,
781 	 * or we can deadlock.
782 	 */
783 	ret = btrfs_record_root_in_trans(h, root);
784 	if (ret) {
785 		/*
786 		 * The transaction handle is fully initialized and linked with
787 		 * other structures so it needs to be ended in case of errors,
788 		 * not just freed.
789 		 */
790 		btrfs_end_transaction(h);
791 		return ERR_PTR(ret);
792 	}
793 
794 	return h;
795 
796 join_fail:
797 	if (type & __TRANS_FREEZABLE)
798 		sb_end_intwrite(fs_info->sb);
799 	kmem_cache_free(btrfs_trans_handle_cachep, h);
800 alloc_fail:
801 	if (num_bytes)
802 		btrfs_block_rsv_release(fs_info, trans_rsv, num_bytes, NULL);
803 	if (delayed_refs_bytes)
804 		btrfs_space_info_free_bytes_may_use(fs_info, trans_rsv->space_info,
805 						    delayed_refs_bytes);
806 reserve_fail:
807 	btrfs_qgroup_free_meta_prealloc(root, qgroup_reserved);
808 	return ERR_PTR(ret);
809 }
810 
811 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
812 						   unsigned int num_items)
813 {
814 	return start_transaction(root, num_items, TRANS_START,
815 				 BTRFS_RESERVE_FLUSH_ALL, true);
816 }
817 
818 struct btrfs_trans_handle *btrfs_start_transaction_fallback_global_rsv(
819 					struct btrfs_root *root,
820 					unsigned int num_items)
821 {
822 	return start_transaction(root, num_items, TRANS_START,
823 				 BTRFS_RESERVE_FLUSH_ALL_STEAL, false);
824 }
825 
826 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
827 {
828 	return start_transaction(root, 0, TRANS_JOIN, BTRFS_RESERVE_NO_FLUSH,
829 				 true);
830 }
831 
832 struct btrfs_trans_handle *btrfs_join_transaction_spacecache(struct btrfs_root *root)
833 {
834 	return start_transaction(root, 0, TRANS_JOIN_NOLOCK,
835 				 BTRFS_RESERVE_NO_FLUSH, true);
836 }
837 
838 /*
839  * Similar to regular join but it never starts a transaction when none is
840  * running or when there's a running one at a state >= TRANS_STATE_UNBLOCKED.
841  * This is similar to btrfs_attach_transaction() but it allows the join to
842  * happen if the transaction commit already started but it's not yet in the
843  * "doing" phase (the state is < TRANS_STATE_COMMIT_DOING).
844  */
845 struct btrfs_trans_handle *btrfs_join_transaction_nostart(struct btrfs_root *root)
846 {
847 	return start_transaction(root, 0, TRANS_JOIN_NOSTART,
848 				 BTRFS_RESERVE_NO_FLUSH, true);
849 }
850 
851 /*
852  * Catch the running transaction.
853  *
854  * It is used when we want to commit the current the transaction, but
855  * don't want to start a new one.
856  *
857  * Note: If this function return -ENOENT, it just means there is no
858  * running transaction. But it is possible that the inactive transaction
859  * is still in the memory, not fully on disk. If you hope there is no
860  * inactive transaction in the fs when -ENOENT is returned, you should
861  * invoke
862  *     btrfs_attach_transaction_barrier()
863  */
864 struct btrfs_trans_handle *btrfs_attach_transaction(struct btrfs_root *root)
865 {
866 	return start_transaction(root, 0, TRANS_ATTACH,
867 				 BTRFS_RESERVE_NO_FLUSH, true);
868 }
869 
870 /*
871  * Catch the running transaction.
872  *
873  * It is similar to the above function, the difference is this one
874  * will wait for all the inactive transactions until they fully
875  * complete.
876  */
877 struct btrfs_trans_handle *
878 btrfs_attach_transaction_barrier(struct btrfs_root *root)
879 {
880 	struct btrfs_trans_handle *trans;
881 
882 	trans = start_transaction(root, 0, TRANS_ATTACH,
883 				  BTRFS_RESERVE_NO_FLUSH, true);
884 	if (trans == ERR_PTR(-ENOENT)) {
885 		int ret;
886 
887 		ret = btrfs_wait_for_commit(root->fs_info, 0);
888 		if (ret)
889 			return ERR_PTR(ret);
890 	}
891 
892 	return trans;
893 }
894 
895 /* Wait for a transaction commit to reach at least the given state. */
896 static noinline void wait_for_commit(struct btrfs_transaction *commit,
897 				     const enum btrfs_trans_state min_state)
898 {
899 	struct btrfs_fs_info *fs_info = commit->fs_info;
900 	u64 transid = commit->transid;
901 	bool put = false;
902 
903 	/*
904 	 * At the moment this function is called with min_state either being
905 	 * TRANS_STATE_COMPLETED or TRANS_STATE_SUPER_COMMITTED.
906 	 */
907 	if (min_state == TRANS_STATE_COMPLETED)
908 		btrfs_might_wait_for_state(fs_info, BTRFS_LOCKDEP_TRANS_COMPLETED);
909 	else
910 		btrfs_might_wait_for_state(fs_info, BTRFS_LOCKDEP_TRANS_SUPER_COMMITTED);
911 
912 	while (1) {
913 		wait_event(commit->commit_wait, commit->state >= min_state);
914 		if (put)
915 			btrfs_put_transaction(commit);
916 
917 		if (min_state < TRANS_STATE_COMPLETED)
918 			break;
919 
920 		/*
921 		 * A transaction isn't really completed until all of the
922 		 * previous transactions are completed, but with fsync we can
923 		 * end up with SUPER_COMMITTED transactions before a COMPLETED
924 		 * transaction. Wait for those.
925 		 */
926 
927 		spin_lock(&fs_info->trans_lock);
928 		commit = list_first_entry_or_null(&fs_info->trans_list,
929 						  struct btrfs_transaction,
930 						  list);
931 		if (!commit || commit->transid > transid) {
932 			spin_unlock(&fs_info->trans_lock);
933 			break;
934 		}
935 		refcount_inc(&commit->use_count);
936 		put = true;
937 		spin_unlock(&fs_info->trans_lock);
938 	}
939 }
940 
941 int btrfs_wait_for_commit(struct btrfs_fs_info *fs_info, u64 transid)
942 {
943 	struct btrfs_transaction *cur_trans = NULL, *t;
944 	int ret = 0;
945 
946 	if (transid) {
947 		if (transid <= btrfs_get_last_trans_committed(fs_info))
948 			goto out;
949 
950 		/* find specified transaction */
951 		spin_lock(&fs_info->trans_lock);
952 		list_for_each_entry(t, &fs_info->trans_list, list) {
953 			if (t->transid == transid) {
954 				cur_trans = t;
955 				refcount_inc(&cur_trans->use_count);
956 				ret = 0;
957 				break;
958 			}
959 			if (t->transid > transid) {
960 				ret = 0;
961 				break;
962 			}
963 		}
964 		spin_unlock(&fs_info->trans_lock);
965 
966 		/*
967 		 * The specified transaction doesn't exist, or we
968 		 * raced with btrfs_commit_transaction
969 		 */
970 		if (!cur_trans) {
971 			if (transid > btrfs_get_last_trans_committed(fs_info))
972 				ret = -EINVAL;
973 			goto out;
974 		}
975 	} else {
976 		/* find newest transaction that is committing | committed */
977 		spin_lock(&fs_info->trans_lock);
978 		list_for_each_entry_reverse(t, &fs_info->trans_list,
979 					    list) {
980 			if (t->state >= TRANS_STATE_COMMIT_START) {
981 				if (t->state == TRANS_STATE_COMPLETED)
982 					break;
983 				cur_trans = t;
984 				refcount_inc(&cur_trans->use_count);
985 				break;
986 			}
987 		}
988 		spin_unlock(&fs_info->trans_lock);
989 		if (!cur_trans)
990 			goto out;  /* nothing committing|committed */
991 	}
992 
993 	wait_for_commit(cur_trans, TRANS_STATE_COMPLETED);
994 	ret = cur_trans->aborted;
995 	btrfs_put_transaction(cur_trans);
996 out:
997 	return ret;
998 }
999 
1000 void btrfs_throttle(struct btrfs_fs_info *fs_info)
1001 {
1002 	wait_current_trans(fs_info);
1003 }
1004 
1005 bool btrfs_should_end_transaction(struct btrfs_trans_handle *trans)
1006 {
1007 	struct btrfs_transaction *cur_trans = trans->transaction;
1008 
1009 	if (cur_trans->state >= TRANS_STATE_COMMIT_START ||
1010 	    test_bit(BTRFS_DELAYED_REFS_FLUSHING, &cur_trans->delayed_refs.flags))
1011 		return true;
1012 
1013 	if (btrfs_check_space_for_delayed_refs(trans->fs_info))
1014 		return true;
1015 
1016 	return !!btrfs_block_rsv_check(&trans->fs_info->global_block_rsv, 50);
1017 }
1018 
1019 static void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans)
1020 
1021 {
1022 	struct btrfs_fs_info *fs_info = trans->fs_info;
1023 
1024 	if (!trans->block_rsv) {
1025 		ASSERT(!trans->bytes_reserved);
1026 		ASSERT(!trans->delayed_refs_bytes_reserved);
1027 		return;
1028 	}
1029 
1030 	if (!trans->bytes_reserved) {
1031 		ASSERT(!trans->delayed_refs_bytes_reserved);
1032 		return;
1033 	}
1034 
1035 	ASSERT(trans->block_rsv == &fs_info->trans_block_rsv);
1036 	trace_btrfs_space_reservation(fs_info, "transaction",
1037 				      trans->transid, trans->bytes_reserved, 0);
1038 	btrfs_block_rsv_release(fs_info, trans->block_rsv,
1039 				trans->bytes_reserved, NULL);
1040 	trans->bytes_reserved = 0;
1041 
1042 	if (!trans->delayed_refs_bytes_reserved)
1043 		return;
1044 
1045 	trace_btrfs_space_reservation(fs_info, "local_delayed_refs_rsv",
1046 				      trans->transid,
1047 				      trans->delayed_refs_bytes_reserved, 0);
1048 	btrfs_block_rsv_release(fs_info, &trans->delayed_rsv,
1049 				trans->delayed_refs_bytes_reserved, NULL);
1050 	trans->delayed_refs_bytes_reserved = 0;
1051 }
1052 
1053 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
1054 				   int throttle)
1055 {
1056 	struct btrfs_fs_info *info = trans->fs_info;
1057 	struct btrfs_transaction *cur_trans = trans->transaction;
1058 	int err = 0;
1059 
1060 	if (refcount_read(&trans->use_count) > 1) {
1061 		refcount_dec(&trans->use_count);
1062 		trans->block_rsv = trans->orig_rsv;
1063 		return 0;
1064 	}
1065 
1066 	btrfs_trans_release_metadata(trans);
1067 	trans->block_rsv = NULL;
1068 
1069 	btrfs_create_pending_block_groups(trans);
1070 
1071 	btrfs_trans_release_chunk_metadata(trans);
1072 
1073 	if (trans->type & __TRANS_FREEZABLE)
1074 		sb_end_intwrite(info->sb);
1075 
1076 	WARN_ON(cur_trans != info->running_transaction);
1077 	WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
1078 	atomic_dec(&cur_trans->num_writers);
1079 	extwriter_counter_dec(cur_trans, trans->type);
1080 
1081 	cond_wake_up(&cur_trans->writer_wait);
1082 
1083 	btrfs_lockdep_release(info, btrfs_trans_num_extwriters);
1084 	btrfs_lockdep_release(info, btrfs_trans_num_writers);
1085 
1086 	btrfs_put_transaction(cur_trans);
1087 
1088 	if (current->journal_info == trans)
1089 		current->journal_info = NULL;
1090 
1091 	if (throttle)
1092 		btrfs_run_delayed_iputs(info);
1093 
1094 	if (TRANS_ABORTED(trans) || BTRFS_FS_ERROR(info)) {
1095 		wake_up_process(info->transaction_kthread);
1096 		if (TRANS_ABORTED(trans))
1097 			err = trans->aborted;
1098 		else
1099 			err = -EROFS;
1100 	}
1101 
1102 	kmem_cache_free(btrfs_trans_handle_cachep, trans);
1103 	return err;
1104 }
1105 
1106 int btrfs_end_transaction(struct btrfs_trans_handle *trans)
1107 {
1108 	return __btrfs_end_transaction(trans, 0);
1109 }
1110 
1111 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans)
1112 {
1113 	return __btrfs_end_transaction(trans, 1);
1114 }
1115 
1116 /*
1117  * when btree blocks are allocated, they have some corresponding bits set for
1118  * them in one of two extent_io trees.  This is used to make sure all of
1119  * those extents are sent to disk but does not wait on them
1120  */
1121 int btrfs_write_marked_extents(struct btrfs_fs_info *fs_info,
1122 			       struct extent_io_tree *dirty_pages, int mark)
1123 {
1124 	int err = 0;
1125 	int werr = 0;
1126 	struct address_space *mapping = fs_info->btree_inode->i_mapping;
1127 	struct extent_state *cached_state = NULL;
1128 	u64 start = 0;
1129 	u64 end;
1130 
1131 	while (find_first_extent_bit(dirty_pages, start, &start, &end,
1132 				     mark, &cached_state)) {
1133 		bool wait_writeback = false;
1134 
1135 		err = convert_extent_bit(dirty_pages, start, end,
1136 					 EXTENT_NEED_WAIT,
1137 					 mark, &cached_state);
1138 		/*
1139 		 * convert_extent_bit can return -ENOMEM, which is most of the
1140 		 * time a temporary error. So when it happens, ignore the error
1141 		 * and wait for writeback of this range to finish - because we
1142 		 * failed to set the bit EXTENT_NEED_WAIT for the range, a call
1143 		 * to __btrfs_wait_marked_extents() would not know that
1144 		 * writeback for this range started and therefore wouldn't
1145 		 * wait for it to finish - we don't want to commit a
1146 		 * superblock that points to btree nodes/leafs for which
1147 		 * writeback hasn't finished yet (and without errors).
1148 		 * We cleanup any entries left in the io tree when committing
1149 		 * the transaction (through extent_io_tree_release()).
1150 		 */
1151 		if (err == -ENOMEM) {
1152 			err = 0;
1153 			wait_writeback = true;
1154 		}
1155 		if (!err)
1156 			err = filemap_fdatawrite_range(mapping, start, end);
1157 		if (err)
1158 			werr = err;
1159 		else if (wait_writeback)
1160 			werr = filemap_fdatawait_range(mapping, start, end);
1161 		free_extent_state(cached_state);
1162 		cached_state = NULL;
1163 		cond_resched();
1164 		start = end + 1;
1165 	}
1166 	return werr;
1167 }
1168 
1169 /*
1170  * when btree blocks are allocated, they have some corresponding bits set for
1171  * them in one of two extent_io trees.  This is used to make sure all of
1172  * those extents are on disk for transaction or log commit.  We wait
1173  * on all the pages and clear them from the dirty pages state tree
1174  */
1175 static int __btrfs_wait_marked_extents(struct btrfs_fs_info *fs_info,
1176 				       struct extent_io_tree *dirty_pages)
1177 {
1178 	int err = 0;
1179 	int werr = 0;
1180 	struct address_space *mapping = fs_info->btree_inode->i_mapping;
1181 	struct extent_state *cached_state = NULL;
1182 	u64 start = 0;
1183 	u64 end;
1184 
1185 	while (find_first_extent_bit(dirty_pages, start, &start, &end,
1186 				     EXTENT_NEED_WAIT, &cached_state)) {
1187 		/*
1188 		 * Ignore -ENOMEM errors returned by clear_extent_bit().
1189 		 * When committing the transaction, we'll remove any entries
1190 		 * left in the io tree. For a log commit, we don't remove them
1191 		 * after committing the log because the tree can be accessed
1192 		 * concurrently - we do it only at transaction commit time when
1193 		 * it's safe to do it (through extent_io_tree_release()).
1194 		 */
1195 		err = clear_extent_bit(dirty_pages, start, end,
1196 				       EXTENT_NEED_WAIT, &cached_state);
1197 		if (err == -ENOMEM)
1198 			err = 0;
1199 		if (!err)
1200 			err = filemap_fdatawait_range(mapping, start, end);
1201 		if (err)
1202 			werr = err;
1203 		free_extent_state(cached_state);
1204 		cached_state = NULL;
1205 		cond_resched();
1206 		start = end + 1;
1207 	}
1208 	if (err)
1209 		werr = err;
1210 	return werr;
1211 }
1212 
1213 static int btrfs_wait_extents(struct btrfs_fs_info *fs_info,
1214 		       struct extent_io_tree *dirty_pages)
1215 {
1216 	bool errors = false;
1217 	int err;
1218 
1219 	err = __btrfs_wait_marked_extents(fs_info, dirty_pages);
1220 	if (test_and_clear_bit(BTRFS_FS_BTREE_ERR, &fs_info->flags))
1221 		errors = true;
1222 
1223 	if (errors && !err)
1224 		err = -EIO;
1225 	return err;
1226 }
1227 
1228 int btrfs_wait_tree_log_extents(struct btrfs_root *log_root, int mark)
1229 {
1230 	struct btrfs_fs_info *fs_info = log_root->fs_info;
1231 	struct extent_io_tree *dirty_pages = &log_root->dirty_log_pages;
1232 	bool errors = false;
1233 	int err;
1234 
1235 	ASSERT(log_root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
1236 
1237 	err = __btrfs_wait_marked_extents(fs_info, dirty_pages);
1238 	if ((mark & EXTENT_DIRTY) &&
1239 	    test_and_clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags))
1240 		errors = true;
1241 
1242 	if ((mark & EXTENT_NEW) &&
1243 	    test_and_clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags))
1244 		errors = true;
1245 
1246 	if (errors && !err)
1247 		err = -EIO;
1248 	return err;
1249 }
1250 
1251 /*
1252  * When btree blocks are allocated the corresponding extents are marked dirty.
1253  * This function ensures such extents are persisted on disk for transaction or
1254  * log commit.
1255  *
1256  * @trans: transaction whose dirty pages we'd like to write
1257  */
1258 static int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans)
1259 {
1260 	int ret;
1261 	int ret2;
1262 	struct extent_io_tree *dirty_pages = &trans->transaction->dirty_pages;
1263 	struct btrfs_fs_info *fs_info = trans->fs_info;
1264 	struct blk_plug plug;
1265 
1266 	blk_start_plug(&plug);
1267 	ret = btrfs_write_marked_extents(fs_info, dirty_pages, EXTENT_DIRTY);
1268 	blk_finish_plug(&plug);
1269 	ret2 = btrfs_wait_extents(fs_info, dirty_pages);
1270 
1271 	extent_io_tree_release(&trans->transaction->dirty_pages);
1272 
1273 	if (ret)
1274 		return ret;
1275 	else if (ret2)
1276 		return ret2;
1277 	else
1278 		return 0;
1279 }
1280 
1281 /*
1282  * this is used to update the root pointer in the tree of tree roots.
1283  *
1284  * But, in the case of the extent allocation tree, updating the root
1285  * pointer may allocate blocks which may change the root of the extent
1286  * allocation tree.
1287  *
1288  * So, this loops and repeats and makes sure the cowonly root didn't
1289  * change while the root pointer was being updated in the metadata.
1290  */
1291 static int update_cowonly_root(struct btrfs_trans_handle *trans,
1292 			       struct btrfs_root *root)
1293 {
1294 	int ret;
1295 	u64 old_root_bytenr;
1296 	u64 old_root_used;
1297 	struct btrfs_fs_info *fs_info = root->fs_info;
1298 	struct btrfs_root *tree_root = fs_info->tree_root;
1299 
1300 	old_root_used = btrfs_root_used(&root->root_item);
1301 
1302 	while (1) {
1303 		old_root_bytenr = btrfs_root_bytenr(&root->root_item);
1304 		if (old_root_bytenr == root->node->start &&
1305 		    old_root_used == btrfs_root_used(&root->root_item))
1306 			break;
1307 
1308 		btrfs_set_root_node(&root->root_item, root->node);
1309 		ret = btrfs_update_root(trans, tree_root,
1310 					&root->root_key,
1311 					&root->root_item);
1312 		if (ret)
1313 			return ret;
1314 
1315 		old_root_used = btrfs_root_used(&root->root_item);
1316 	}
1317 
1318 	return 0;
1319 }
1320 
1321 /*
1322  * update all the cowonly tree roots on disk
1323  *
1324  * The error handling in this function may not be obvious. Any of the
1325  * failures will cause the file system to go offline. We still need
1326  * to clean up the delayed refs.
1327  */
1328 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans)
1329 {
1330 	struct btrfs_fs_info *fs_info = trans->fs_info;
1331 	struct list_head *dirty_bgs = &trans->transaction->dirty_bgs;
1332 	struct list_head *io_bgs = &trans->transaction->io_bgs;
1333 	struct list_head *next;
1334 	struct extent_buffer *eb;
1335 	int ret;
1336 
1337 	/*
1338 	 * At this point no one can be using this transaction to modify any tree
1339 	 * and no one can start another transaction to modify any tree either.
1340 	 */
1341 	ASSERT(trans->transaction->state == TRANS_STATE_COMMIT_DOING);
1342 
1343 	eb = btrfs_lock_root_node(fs_info->tree_root);
1344 	ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL,
1345 			      0, &eb, BTRFS_NESTING_COW);
1346 	btrfs_tree_unlock(eb);
1347 	free_extent_buffer(eb);
1348 
1349 	if (ret)
1350 		return ret;
1351 
1352 	ret = btrfs_run_dev_stats(trans);
1353 	if (ret)
1354 		return ret;
1355 	ret = btrfs_run_dev_replace(trans);
1356 	if (ret)
1357 		return ret;
1358 	ret = btrfs_run_qgroups(trans);
1359 	if (ret)
1360 		return ret;
1361 
1362 	ret = btrfs_setup_space_cache(trans);
1363 	if (ret)
1364 		return ret;
1365 
1366 again:
1367 	while (!list_empty(&fs_info->dirty_cowonly_roots)) {
1368 		struct btrfs_root *root;
1369 		next = fs_info->dirty_cowonly_roots.next;
1370 		list_del_init(next);
1371 		root = list_entry(next, struct btrfs_root, dirty_list);
1372 		clear_bit(BTRFS_ROOT_DIRTY, &root->state);
1373 
1374 		list_add_tail(&root->dirty_list,
1375 			      &trans->transaction->switch_commits);
1376 		ret = update_cowonly_root(trans, root);
1377 		if (ret)
1378 			return ret;
1379 	}
1380 
1381 	/* Now flush any delayed refs generated by updating all of the roots */
1382 	ret = btrfs_run_delayed_refs(trans, U64_MAX);
1383 	if (ret)
1384 		return ret;
1385 
1386 	while (!list_empty(dirty_bgs) || !list_empty(io_bgs)) {
1387 		ret = btrfs_write_dirty_block_groups(trans);
1388 		if (ret)
1389 			return ret;
1390 
1391 		/*
1392 		 * We're writing the dirty block groups, which could generate
1393 		 * delayed refs, which could generate more dirty block groups,
1394 		 * so we want to keep this flushing in this loop to make sure
1395 		 * everything gets run.
1396 		 */
1397 		ret = btrfs_run_delayed_refs(trans, U64_MAX);
1398 		if (ret)
1399 			return ret;
1400 	}
1401 
1402 	if (!list_empty(&fs_info->dirty_cowonly_roots))
1403 		goto again;
1404 
1405 	/* Update dev-replace pointer once everything is committed */
1406 	fs_info->dev_replace.committed_cursor_left =
1407 		fs_info->dev_replace.cursor_left_last_write_of_item;
1408 
1409 	return 0;
1410 }
1411 
1412 /*
1413  * If we had a pending drop we need to see if there are any others left in our
1414  * dead roots list, and if not clear our bit and wake any waiters.
1415  */
1416 void btrfs_maybe_wake_unfinished_drop(struct btrfs_fs_info *fs_info)
1417 {
1418 	/*
1419 	 * We put the drop in progress roots at the front of the list, so if the
1420 	 * first entry doesn't have UNFINISHED_DROP set we can wake everybody
1421 	 * up.
1422 	 */
1423 	spin_lock(&fs_info->trans_lock);
1424 	if (!list_empty(&fs_info->dead_roots)) {
1425 		struct btrfs_root *root = list_first_entry(&fs_info->dead_roots,
1426 							   struct btrfs_root,
1427 							   root_list);
1428 		if (test_bit(BTRFS_ROOT_UNFINISHED_DROP, &root->state)) {
1429 			spin_unlock(&fs_info->trans_lock);
1430 			return;
1431 		}
1432 	}
1433 	spin_unlock(&fs_info->trans_lock);
1434 
1435 	btrfs_wake_unfinished_drop(fs_info);
1436 }
1437 
1438 /*
1439  * dead roots are old snapshots that need to be deleted.  This allocates
1440  * a dirty root struct and adds it into the list of dead roots that need to
1441  * be deleted
1442  */
1443 void btrfs_add_dead_root(struct btrfs_root *root)
1444 {
1445 	struct btrfs_fs_info *fs_info = root->fs_info;
1446 
1447 	spin_lock(&fs_info->trans_lock);
1448 	if (list_empty(&root->root_list)) {
1449 		btrfs_grab_root(root);
1450 
1451 		/* We want to process the partially complete drops first. */
1452 		if (test_bit(BTRFS_ROOT_UNFINISHED_DROP, &root->state))
1453 			list_add(&root->root_list, &fs_info->dead_roots);
1454 		else
1455 			list_add_tail(&root->root_list, &fs_info->dead_roots);
1456 	}
1457 	spin_unlock(&fs_info->trans_lock);
1458 }
1459 
1460 /*
1461  * Update each subvolume root and its relocation root, if it exists, in the tree
1462  * of tree roots. Also free log roots if they exist.
1463  */
1464 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans)
1465 {
1466 	struct btrfs_fs_info *fs_info = trans->fs_info;
1467 	struct btrfs_root *gang[8];
1468 	int i;
1469 	int ret;
1470 
1471 	/*
1472 	 * At this point no one can be using this transaction to modify any tree
1473 	 * and no one can start another transaction to modify any tree either.
1474 	 */
1475 	ASSERT(trans->transaction->state == TRANS_STATE_COMMIT_DOING);
1476 
1477 	spin_lock(&fs_info->fs_roots_radix_lock);
1478 	while (1) {
1479 		ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
1480 						 (void **)gang, 0,
1481 						 ARRAY_SIZE(gang),
1482 						 BTRFS_ROOT_TRANS_TAG);
1483 		if (ret == 0)
1484 			break;
1485 		for (i = 0; i < ret; i++) {
1486 			struct btrfs_root *root = gang[i];
1487 			int ret2;
1488 
1489 			/*
1490 			 * At this point we can neither have tasks logging inodes
1491 			 * from a root nor trying to commit a log tree.
1492 			 */
1493 			ASSERT(atomic_read(&root->log_writers) == 0);
1494 			ASSERT(atomic_read(&root->log_commit[0]) == 0);
1495 			ASSERT(atomic_read(&root->log_commit[1]) == 0);
1496 
1497 			radix_tree_tag_clear(&fs_info->fs_roots_radix,
1498 					(unsigned long)root->root_key.objectid,
1499 					BTRFS_ROOT_TRANS_TAG);
1500 			spin_unlock(&fs_info->fs_roots_radix_lock);
1501 
1502 			btrfs_free_log(trans, root);
1503 			ret2 = btrfs_update_reloc_root(trans, root);
1504 			if (ret2)
1505 				return ret2;
1506 
1507 			/* see comments in should_cow_block() */
1508 			clear_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1509 			smp_mb__after_atomic();
1510 
1511 			if (root->commit_root != root->node) {
1512 				list_add_tail(&root->dirty_list,
1513 					&trans->transaction->switch_commits);
1514 				btrfs_set_root_node(&root->root_item,
1515 						    root->node);
1516 			}
1517 
1518 			ret2 = btrfs_update_root(trans, fs_info->tree_root,
1519 						&root->root_key,
1520 						&root->root_item);
1521 			if (ret2)
1522 				return ret2;
1523 			spin_lock(&fs_info->fs_roots_radix_lock);
1524 			btrfs_qgroup_free_meta_all_pertrans(root);
1525 		}
1526 	}
1527 	spin_unlock(&fs_info->fs_roots_radix_lock);
1528 	return 0;
1529 }
1530 
1531 /*
1532  * Do all special snapshot related qgroup dirty hack.
1533  *
1534  * Will do all needed qgroup inherit and dirty hack like switch commit
1535  * roots inside one transaction and write all btree into disk, to make
1536  * qgroup works.
1537  */
1538 static int qgroup_account_snapshot(struct btrfs_trans_handle *trans,
1539 				   struct btrfs_root *src,
1540 				   struct btrfs_root *parent,
1541 				   struct btrfs_qgroup_inherit *inherit,
1542 				   u64 dst_objectid)
1543 {
1544 	struct btrfs_fs_info *fs_info = src->fs_info;
1545 	int ret;
1546 
1547 	/*
1548 	 * Save some performance in the case that qgroups are not enabled. If
1549 	 * this check races with the ioctl, rescan will kick in anyway.
1550 	 */
1551 	if (!btrfs_qgroup_full_accounting(fs_info))
1552 		return 0;
1553 
1554 	/*
1555 	 * Ensure dirty @src will be committed.  Or, after coming
1556 	 * commit_fs_roots() and switch_commit_roots(), any dirty but not
1557 	 * recorded root will never be updated again, causing an outdated root
1558 	 * item.
1559 	 */
1560 	ret = record_root_in_trans(trans, src, 1);
1561 	if (ret)
1562 		return ret;
1563 
1564 	/*
1565 	 * btrfs_qgroup_inherit relies on a consistent view of the usage for the
1566 	 * src root, so we must run the delayed refs here.
1567 	 *
1568 	 * However this isn't particularly fool proof, because there's no
1569 	 * synchronization keeping us from changing the tree after this point
1570 	 * before we do the qgroup_inherit, or even from making changes while
1571 	 * we're doing the qgroup_inherit.  But that's a problem for the future,
1572 	 * for now flush the delayed refs to narrow the race window where the
1573 	 * qgroup counters could end up wrong.
1574 	 */
1575 	ret = btrfs_run_delayed_refs(trans, U64_MAX);
1576 	if (ret) {
1577 		btrfs_abort_transaction(trans, ret);
1578 		return ret;
1579 	}
1580 
1581 	ret = commit_fs_roots(trans);
1582 	if (ret)
1583 		goto out;
1584 	ret = btrfs_qgroup_account_extents(trans);
1585 	if (ret < 0)
1586 		goto out;
1587 
1588 	/* Now qgroup are all updated, we can inherit it to new qgroups */
1589 	ret = btrfs_qgroup_inherit(trans, src->root_key.objectid, dst_objectid,
1590 				   parent->root_key.objectid, inherit);
1591 	if (ret < 0)
1592 		goto out;
1593 
1594 	/*
1595 	 * Now we do a simplified commit transaction, which will:
1596 	 * 1) commit all subvolume and extent tree
1597 	 *    To ensure all subvolume and extent tree have a valid
1598 	 *    commit_root to accounting later insert_dir_item()
1599 	 * 2) write all btree blocks onto disk
1600 	 *    This is to make sure later btree modification will be cowed
1601 	 *    Or commit_root can be populated and cause wrong qgroup numbers
1602 	 * In this simplified commit, we don't really care about other trees
1603 	 * like chunk and root tree, as they won't affect qgroup.
1604 	 * And we don't write super to avoid half committed status.
1605 	 */
1606 	ret = commit_cowonly_roots(trans);
1607 	if (ret)
1608 		goto out;
1609 	switch_commit_roots(trans);
1610 	ret = btrfs_write_and_wait_transaction(trans);
1611 	if (ret)
1612 		btrfs_handle_fs_error(fs_info, ret,
1613 			"Error while writing out transaction for qgroup");
1614 
1615 out:
1616 	/*
1617 	 * Force parent root to be updated, as we recorded it before so its
1618 	 * last_trans == cur_transid.
1619 	 * Or it won't be committed again onto disk after later
1620 	 * insert_dir_item()
1621 	 */
1622 	if (!ret)
1623 		ret = record_root_in_trans(trans, parent, 1);
1624 	return ret;
1625 }
1626 
1627 /*
1628  * new snapshots need to be created at a very specific time in the
1629  * transaction commit.  This does the actual creation.
1630  *
1631  * Note:
1632  * If the error which may affect the commitment of the current transaction
1633  * happens, we should return the error number. If the error which just affect
1634  * the creation of the pending snapshots, just return 0.
1635  */
1636 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
1637 				   struct btrfs_pending_snapshot *pending)
1638 {
1639 
1640 	struct btrfs_fs_info *fs_info = trans->fs_info;
1641 	struct btrfs_key key;
1642 	struct btrfs_root_item *new_root_item;
1643 	struct btrfs_root *tree_root = fs_info->tree_root;
1644 	struct btrfs_root *root = pending->root;
1645 	struct btrfs_root *parent_root;
1646 	struct btrfs_block_rsv *rsv;
1647 	struct inode *parent_inode = pending->dir;
1648 	struct btrfs_path *path;
1649 	struct btrfs_dir_item *dir_item;
1650 	struct extent_buffer *tmp;
1651 	struct extent_buffer *old;
1652 	struct timespec64 cur_time;
1653 	int ret = 0;
1654 	u64 to_reserve = 0;
1655 	u64 index = 0;
1656 	u64 objectid;
1657 	u64 root_flags;
1658 	unsigned int nofs_flags;
1659 	struct fscrypt_name fname;
1660 
1661 	ASSERT(pending->path);
1662 	path = pending->path;
1663 
1664 	ASSERT(pending->root_item);
1665 	new_root_item = pending->root_item;
1666 
1667 	/*
1668 	 * We're inside a transaction and must make sure that any potential
1669 	 * allocations with GFP_KERNEL in fscrypt won't recurse back to
1670 	 * filesystem.
1671 	 */
1672 	nofs_flags = memalloc_nofs_save();
1673 	pending->error = fscrypt_setup_filename(parent_inode,
1674 						&pending->dentry->d_name, 0,
1675 						&fname);
1676 	memalloc_nofs_restore(nofs_flags);
1677 	if (pending->error)
1678 		goto free_pending;
1679 
1680 	pending->error = btrfs_get_free_objectid(tree_root, &objectid);
1681 	if (pending->error)
1682 		goto free_fname;
1683 
1684 	/*
1685 	 * Make qgroup to skip current new snapshot's qgroupid, as it is
1686 	 * accounted by later btrfs_qgroup_inherit().
1687 	 */
1688 	btrfs_set_skip_qgroup(trans, objectid);
1689 
1690 	btrfs_reloc_pre_snapshot(pending, &to_reserve);
1691 
1692 	if (to_reserve > 0) {
1693 		pending->error = btrfs_block_rsv_add(fs_info,
1694 						     &pending->block_rsv,
1695 						     to_reserve,
1696 						     BTRFS_RESERVE_NO_FLUSH);
1697 		if (pending->error)
1698 			goto clear_skip_qgroup;
1699 	}
1700 
1701 	key.objectid = objectid;
1702 	key.offset = (u64)-1;
1703 	key.type = BTRFS_ROOT_ITEM_KEY;
1704 
1705 	rsv = trans->block_rsv;
1706 	trans->block_rsv = &pending->block_rsv;
1707 	trans->bytes_reserved = trans->block_rsv->reserved;
1708 	trace_btrfs_space_reservation(fs_info, "transaction",
1709 				      trans->transid,
1710 				      trans->bytes_reserved, 1);
1711 	parent_root = BTRFS_I(parent_inode)->root;
1712 	ret = record_root_in_trans(trans, parent_root, 0);
1713 	if (ret)
1714 		goto fail;
1715 	cur_time = current_time(parent_inode);
1716 
1717 	/*
1718 	 * insert the directory item
1719 	 */
1720 	ret = btrfs_set_inode_index(BTRFS_I(parent_inode), &index);
1721 	if (ret) {
1722 		btrfs_abort_transaction(trans, ret);
1723 		goto fail;
1724 	}
1725 
1726 	/* check if there is a file/dir which has the same name. */
1727 	dir_item = btrfs_lookup_dir_item(NULL, parent_root, path,
1728 					 btrfs_ino(BTRFS_I(parent_inode)),
1729 					 &fname.disk_name, 0);
1730 	if (dir_item != NULL && !IS_ERR(dir_item)) {
1731 		pending->error = -EEXIST;
1732 		goto dir_item_existed;
1733 	} else if (IS_ERR(dir_item)) {
1734 		ret = PTR_ERR(dir_item);
1735 		btrfs_abort_transaction(trans, ret);
1736 		goto fail;
1737 	}
1738 	btrfs_release_path(path);
1739 
1740 	ret = btrfs_create_qgroup(trans, objectid);
1741 	if (ret && ret != -EEXIST) {
1742 		btrfs_abort_transaction(trans, ret);
1743 		goto fail;
1744 	}
1745 
1746 	/*
1747 	 * pull in the delayed directory update
1748 	 * and the delayed inode item
1749 	 * otherwise we corrupt the FS during
1750 	 * snapshot
1751 	 */
1752 	ret = btrfs_run_delayed_items(trans);
1753 	if (ret) {	/* Transaction aborted */
1754 		btrfs_abort_transaction(trans, ret);
1755 		goto fail;
1756 	}
1757 
1758 	ret = record_root_in_trans(trans, root, 0);
1759 	if (ret) {
1760 		btrfs_abort_transaction(trans, ret);
1761 		goto fail;
1762 	}
1763 	btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
1764 	memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
1765 	btrfs_check_and_init_root_item(new_root_item);
1766 
1767 	root_flags = btrfs_root_flags(new_root_item);
1768 	if (pending->readonly)
1769 		root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
1770 	else
1771 		root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
1772 	btrfs_set_root_flags(new_root_item, root_flags);
1773 
1774 	btrfs_set_root_generation_v2(new_root_item,
1775 			trans->transid);
1776 	generate_random_guid(new_root_item->uuid);
1777 	memcpy(new_root_item->parent_uuid, root->root_item.uuid,
1778 			BTRFS_UUID_SIZE);
1779 	if (!(root_flags & BTRFS_ROOT_SUBVOL_RDONLY)) {
1780 		memset(new_root_item->received_uuid, 0,
1781 		       sizeof(new_root_item->received_uuid));
1782 		memset(&new_root_item->stime, 0, sizeof(new_root_item->stime));
1783 		memset(&new_root_item->rtime, 0, sizeof(new_root_item->rtime));
1784 		btrfs_set_root_stransid(new_root_item, 0);
1785 		btrfs_set_root_rtransid(new_root_item, 0);
1786 	}
1787 	btrfs_set_stack_timespec_sec(&new_root_item->otime, cur_time.tv_sec);
1788 	btrfs_set_stack_timespec_nsec(&new_root_item->otime, cur_time.tv_nsec);
1789 	btrfs_set_root_otransid(new_root_item, trans->transid);
1790 
1791 	old = btrfs_lock_root_node(root);
1792 	ret = btrfs_cow_block(trans, root, old, NULL, 0, &old,
1793 			      BTRFS_NESTING_COW);
1794 	if (ret) {
1795 		btrfs_tree_unlock(old);
1796 		free_extent_buffer(old);
1797 		btrfs_abort_transaction(trans, ret);
1798 		goto fail;
1799 	}
1800 
1801 	ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1802 	/* clean up in any case */
1803 	btrfs_tree_unlock(old);
1804 	free_extent_buffer(old);
1805 	if (ret) {
1806 		btrfs_abort_transaction(trans, ret);
1807 		goto fail;
1808 	}
1809 	/* see comments in should_cow_block() */
1810 	set_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1811 	smp_wmb();
1812 
1813 	btrfs_set_root_node(new_root_item, tmp);
1814 	/* record when the snapshot was created in key.offset */
1815 	key.offset = trans->transid;
1816 	ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1817 	btrfs_tree_unlock(tmp);
1818 	free_extent_buffer(tmp);
1819 	if (ret) {
1820 		btrfs_abort_transaction(trans, ret);
1821 		goto fail;
1822 	}
1823 
1824 	/*
1825 	 * insert root back/forward references
1826 	 */
1827 	ret = btrfs_add_root_ref(trans, objectid,
1828 				 parent_root->root_key.objectid,
1829 				 btrfs_ino(BTRFS_I(parent_inode)), index,
1830 				 &fname.disk_name);
1831 	if (ret) {
1832 		btrfs_abort_transaction(trans, ret);
1833 		goto fail;
1834 	}
1835 
1836 	key.offset = (u64)-1;
1837 	pending->snap = btrfs_get_new_fs_root(fs_info, objectid, pending->anon_dev);
1838 	if (IS_ERR(pending->snap)) {
1839 		ret = PTR_ERR(pending->snap);
1840 		pending->snap = NULL;
1841 		btrfs_abort_transaction(trans, ret);
1842 		goto fail;
1843 	}
1844 
1845 	ret = btrfs_reloc_post_snapshot(trans, pending);
1846 	if (ret) {
1847 		btrfs_abort_transaction(trans, ret);
1848 		goto fail;
1849 	}
1850 
1851 	/*
1852 	 * Do special qgroup accounting for snapshot, as we do some qgroup
1853 	 * snapshot hack to do fast snapshot.
1854 	 * To co-operate with that hack, we do hack again.
1855 	 * Or snapshot will be greatly slowed down by a subtree qgroup rescan
1856 	 */
1857 	if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_FULL)
1858 		ret = qgroup_account_snapshot(trans, root, parent_root,
1859 					      pending->inherit, objectid);
1860 	else if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_SIMPLE)
1861 		ret = btrfs_qgroup_inherit(trans, root->root_key.objectid, objectid,
1862 					   parent_root->root_key.objectid, pending->inherit);
1863 	if (ret < 0)
1864 		goto fail;
1865 
1866 	ret = btrfs_insert_dir_item(trans, &fname.disk_name,
1867 				    BTRFS_I(parent_inode), &key, BTRFS_FT_DIR,
1868 				    index);
1869 	/* We have check then name at the beginning, so it is impossible. */
1870 	BUG_ON(ret == -EEXIST || ret == -EOVERFLOW);
1871 	if (ret) {
1872 		btrfs_abort_transaction(trans, ret);
1873 		goto fail;
1874 	}
1875 
1876 	btrfs_i_size_write(BTRFS_I(parent_inode), parent_inode->i_size +
1877 						  fname.disk_name.len * 2);
1878 	inode_set_mtime_to_ts(parent_inode,
1879 			      inode_set_ctime_current(parent_inode));
1880 	ret = btrfs_update_inode_fallback(trans, BTRFS_I(parent_inode));
1881 	if (ret) {
1882 		btrfs_abort_transaction(trans, ret);
1883 		goto fail;
1884 	}
1885 	ret = btrfs_uuid_tree_add(trans, new_root_item->uuid,
1886 				  BTRFS_UUID_KEY_SUBVOL,
1887 				  objectid);
1888 	if (ret) {
1889 		btrfs_abort_transaction(trans, ret);
1890 		goto fail;
1891 	}
1892 	if (!btrfs_is_empty_uuid(new_root_item->received_uuid)) {
1893 		ret = btrfs_uuid_tree_add(trans, new_root_item->received_uuid,
1894 					  BTRFS_UUID_KEY_RECEIVED_SUBVOL,
1895 					  objectid);
1896 		if (ret && ret != -EEXIST) {
1897 			btrfs_abort_transaction(trans, ret);
1898 			goto fail;
1899 		}
1900 	}
1901 
1902 fail:
1903 	pending->error = ret;
1904 dir_item_existed:
1905 	trans->block_rsv = rsv;
1906 	trans->bytes_reserved = 0;
1907 clear_skip_qgroup:
1908 	btrfs_clear_skip_qgroup(trans);
1909 free_fname:
1910 	fscrypt_free_filename(&fname);
1911 free_pending:
1912 	kfree(new_root_item);
1913 	pending->root_item = NULL;
1914 	btrfs_free_path(path);
1915 	pending->path = NULL;
1916 
1917 	return ret;
1918 }
1919 
1920 /*
1921  * create all the snapshots we've scheduled for creation
1922  */
1923 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans)
1924 {
1925 	struct btrfs_pending_snapshot *pending, *next;
1926 	struct list_head *head = &trans->transaction->pending_snapshots;
1927 	int ret = 0;
1928 
1929 	list_for_each_entry_safe(pending, next, head, list) {
1930 		list_del(&pending->list);
1931 		ret = create_pending_snapshot(trans, pending);
1932 		if (ret)
1933 			break;
1934 	}
1935 	return ret;
1936 }
1937 
1938 static void update_super_roots(struct btrfs_fs_info *fs_info)
1939 {
1940 	struct btrfs_root_item *root_item;
1941 	struct btrfs_super_block *super;
1942 
1943 	super = fs_info->super_copy;
1944 
1945 	root_item = &fs_info->chunk_root->root_item;
1946 	super->chunk_root = root_item->bytenr;
1947 	super->chunk_root_generation = root_item->generation;
1948 	super->chunk_root_level = root_item->level;
1949 
1950 	root_item = &fs_info->tree_root->root_item;
1951 	super->root = root_item->bytenr;
1952 	super->generation = root_item->generation;
1953 	super->root_level = root_item->level;
1954 	if (btrfs_test_opt(fs_info, SPACE_CACHE))
1955 		super->cache_generation = root_item->generation;
1956 	else if (test_bit(BTRFS_FS_CLEANUP_SPACE_CACHE_V1, &fs_info->flags))
1957 		super->cache_generation = 0;
1958 	if (test_bit(BTRFS_FS_UPDATE_UUID_TREE_GEN, &fs_info->flags))
1959 		super->uuid_tree_generation = root_item->generation;
1960 }
1961 
1962 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1963 {
1964 	struct btrfs_transaction *trans;
1965 	int ret = 0;
1966 
1967 	spin_lock(&info->trans_lock);
1968 	trans = info->running_transaction;
1969 	if (trans)
1970 		ret = (trans->state >= TRANS_STATE_COMMIT_START);
1971 	spin_unlock(&info->trans_lock);
1972 	return ret;
1973 }
1974 
1975 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1976 {
1977 	struct btrfs_transaction *trans;
1978 	int ret = 0;
1979 
1980 	spin_lock(&info->trans_lock);
1981 	trans = info->running_transaction;
1982 	if (trans)
1983 		ret = is_transaction_blocked(trans);
1984 	spin_unlock(&info->trans_lock);
1985 	return ret;
1986 }
1987 
1988 void btrfs_commit_transaction_async(struct btrfs_trans_handle *trans)
1989 {
1990 	struct btrfs_fs_info *fs_info = trans->fs_info;
1991 	struct btrfs_transaction *cur_trans;
1992 
1993 	/* Kick the transaction kthread. */
1994 	set_bit(BTRFS_FS_COMMIT_TRANS, &fs_info->flags);
1995 	wake_up_process(fs_info->transaction_kthread);
1996 
1997 	/* take transaction reference */
1998 	cur_trans = trans->transaction;
1999 	refcount_inc(&cur_trans->use_count);
2000 
2001 	btrfs_end_transaction(trans);
2002 
2003 	/*
2004 	 * Wait for the current transaction commit to start and block
2005 	 * subsequent transaction joins
2006 	 */
2007 	btrfs_might_wait_for_state(fs_info, BTRFS_LOCKDEP_TRANS_COMMIT_PREP);
2008 	wait_event(fs_info->transaction_blocked_wait,
2009 		   cur_trans->state >= TRANS_STATE_COMMIT_START ||
2010 		   TRANS_ABORTED(cur_trans));
2011 	btrfs_put_transaction(cur_trans);
2012 }
2013 
2014 static void cleanup_transaction(struct btrfs_trans_handle *trans, int err)
2015 {
2016 	struct btrfs_fs_info *fs_info = trans->fs_info;
2017 	struct btrfs_transaction *cur_trans = trans->transaction;
2018 
2019 	WARN_ON(refcount_read(&trans->use_count) > 1);
2020 
2021 	btrfs_abort_transaction(trans, err);
2022 
2023 	spin_lock(&fs_info->trans_lock);
2024 
2025 	/*
2026 	 * If the transaction is removed from the list, it means this
2027 	 * transaction has been committed successfully, so it is impossible
2028 	 * to call the cleanup function.
2029 	 */
2030 	BUG_ON(list_empty(&cur_trans->list));
2031 
2032 	if (cur_trans == fs_info->running_transaction) {
2033 		cur_trans->state = TRANS_STATE_COMMIT_DOING;
2034 		spin_unlock(&fs_info->trans_lock);
2035 
2036 		/*
2037 		 * The thread has already released the lockdep map as reader
2038 		 * already in btrfs_commit_transaction().
2039 		 */
2040 		btrfs_might_wait_for_event(fs_info, btrfs_trans_num_writers);
2041 		wait_event(cur_trans->writer_wait,
2042 			   atomic_read(&cur_trans->num_writers) == 1);
2043 
2044 		spin_lock(&fs_info->trans_lock);
2045 	}
2046 
2047 	/*
2048 	 * Now that we know no one else is still using the transaction we can
2049 	 * remove the transaction from the list of transactions. This avoids
2050 	 * the transaction kthread from cleaning up the transaction while some
2051 	 * other task is still using it, which could result in a use-after-free
2052 	 * on things like log trees, as it forces the transaction kthread to
2053 	 * wait for this transaction to be cleaned up by us.
2054 	 */
2055 	list_del_init(&cur_trans->list);
2056 
2057 	spin_unlock(&fs_info->trans_lock);
2058 
2059 	btrfs_cleanup_one_transaction(trans->transaction, fs_info);
2060 
2061 	spin_lock(&fs_info->trans_lock);
2062 	if (cur_trans == fs_info->running_transaction)
2063 		fs_info->running_transaction = NULL;
2064 	spin_unlock(&fs_info->trans_lock);
2065 
2066 	if (trans->type & __TRANS_FREEZABLE)
2067 		sb_end_intwrite(fs_info->sb);
2068 	btrfs_put_transaction(cur_trans);
2069 	btrfs_put_transaction(cur_trans);
2070 
2071 	trace_btrfs_transaction_commit(fs_info);
2072 
2073 	if (current->journal_info == trans)
2074 		current->journal_info = NULL;
2075 
2076 	/*
2077 	 * If relocation is running, we can't cancel scrub because that will
2078 	 * result in a deadlock. Before relocating a block group, relocation
2079 	 * pauses scrub, then starts and commits a transaction before unpausing
2080 	 * scrub. If the transaction commit is being done by the relocation
2081 	 * task or triggered by another task and the relocation task is waiting
2082 	 * for the commit, and we end up here due to an error in the commit
2083 	 * path, then calling btrfs_scrub_cancel() will deadlock, as we are
2084 	 * asking for scrub to stop while having it asked to be paused higher
2085 	 * above in relocation code.
2086 	 */
2087 	if (!test_bit(BTRFS_FS_RELOC_RUNNING, &fs_info->flags))
2088 		btrfs_scrub_cancel(fs_info);
2089 
2090 	kmem_cache_free(btrfs_trans_handle_cachep, trans);
2091 }
2092 
2093 /*
2094  * Release reserved delayed ref space of all pending block groups of the
2095  * transaction and remove them from the list
2096  */
2097 static void btrfs_cleanup_pending_block_groups(struct btrfs_trans_handle *trans)
2098 {
2099        struct btrfs_fs_info *fs_info = trans->fs_info;
2100        struct btrfs_block_group *block_group, *tmp;
2101 
2102        list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
2103                btrfs_dec_delayed_refs_rsv_bg_inserts(fs_info);
2104                list_del_init(&block_group->bg_list);
2105        }
2106 }
2107 
2108 static inline int btrfs_start_delalloc_flush(struct btrfs_fs_info *fs_info)
2109 {
2110 	/*
2111 	 * We use try_to_writeback_inodes_sb() here because if we used
2112 	 * btrfs_start_delalloc_roots we would deadlock with fs freeze.
2113 	 * Currently are holding the fs freeze lock, if we do an async flush
2114 	 * we'll do btrfs_join_transaction() and deadlock because we need to
2115 	 * wait for the fs freeze lock.  Using the direct flushing we benefit
2116 	 * from already being in a transaction and our join_transaction doesn't
2117 	 * have to re-take the fs freeze lock.
2118 	 *
2119 	 * Note that try_to_writeback_inodes_sb() will only trigger writeback
2120 	 * if it can read lock sb->s_umount. It will always be able to lock it,
2121 	 * except when the filesystem is being unmounted or being frozen, but in
2122 	 * those cases sync_filesystem() is called, which results in calling
2123 	 * writeback_inodes_sb() while holding a write lock on sb->s_umount.
2124 	 * Note that we don't call writeback_inodes_sb() directly, because it
2125 	 * will emit a warning if sb->s_umount is not locked.
2126 	 */
2127 	if (btrfs_test_opt(fs_info, FLUSHONCOMMIT))
2128 		try_to_writeback_inodes_sb(fs_info->sb, WB_REASON_SYNC);
2129 	return 0;
2130 }
2131 
2132 static inline void btrfs_wait_delalloc_flush(struct btrfs_fs_info *fs_info)
2133 {
2134 	if (btrfs_test_opt(fs_info, FLUSHONCOMMIT))
2135 		btrfs_wait_ordered_roots(fs_info, U64_MAX, 0, (u64)-1);
2136 }
2137 
2138 /*
2139  * Add a pending snapshot associated with the given transaction handle to the
2140  * respective handle. This must be called after the transaction commit started
2141  * and while holding fs_info->trans_lock.
2142  * This serves to guarantee a caller of btrfs_commit_transaction() that it can
2143  * safely free the pending snapshot pointer in case btrfs_commit_transaction()
2144  * returns an error.
2145  */
2146 static void add_pending_snapshot(struct btrfs_trans_handle *trans)
2147 {
2148 	struct btrfs_transaction *cur_trans = trans->transaction;
2149 
2150 	if (!trans->pending_snapshot)
2151 		return;
2152 
2153 	lockdep_assert_held(&trans->fs_info->trans_lock);
2154 	ASSERT(cur_trans->state >= TRANS_STATE_COMMIT_PREP);
2155 
2156 	list_add(&trans->pending_snapshot->list, &cur_trans->pending_snapshots);
2157 }
2158 
2159 static void update_commit_stats(struct btrfs_fs_info *fs_info, ktime_t interval)
2160 {
2161 	fs_info->commit_stats.commit_count++;
2162 	fs_info->commit_stats.last_commit_dur = interval;
2163 	fs_info->commit_stats.max_commit_dur =
2164 			max_t(u64, fs_info->commit_stats.max_commit_dur, interval);
2165 	fs_info->commit_stats.total_commit_dur += interval;
2166 }
2167 
2168 int btrfs_commit_transaction(struct btrfs_trans_handle *trans)
2169 {
2170 	struct btrfs_fs_info *fs_info = trans->fs_info;
2171 	struct btrfs_transaction *cur_trans = trans->transaction;
2172 	struct btrfs_transaction *prev_trans = NULL;
2173 	int ret;
2174 	ktime_t start_time;
2175 	ktime_t interval;
2176 
2177 	ASSERT(refcount_read(&trans->use_count) == 1);
2178 	btrfs_trans_state_lockdep_acquire(fs_info, BTRFS_LOCKDEP_TRANS_COMMIT_PREP);
2179 
2180 	clear_bit(BTRFS_FS_NEED_TRANS_COMMIT, &fs_info->flags);
2181 
2182 	/* Stop the commit early if ->aborted is set */
2183 	if (TRANS_ABORTED(cur_trans)) {
2184 		ret = cur_trans->aborted;
2185 		goto lockdep_trans_commit_start_release;
2186 	}
2187 
2188 	btrfs_trans_release_metadata(trans);
2189 	trans->block_rsv = NULL;
2190 
2191 	/*
2192 	 * We only want one transaction commit doing the flushing so we do not
2193 	 * waste a bunch of time on lock contention on the extent root node.
2194 	 */
2195 	if (!test_and_set_bit(BTRFS_DELAYED_REFS_FLUSHING,
2196 			      &cur_trans->delayed_refs.flags)) {
2197 		/*
2198 		 * Make a pass through all the delayed refs we have so far.
2199 		 * Any running threads may add more while we are here.
2200 		 */
2201 		ret = btrfs_run_delayed_refs(trans, 0);
2202 		if (ret)
2203 			goto lockdep_trans_commit_start_release;
2204 	}
2205 
2206 	btrfs_create_pending_block_groups(trans);
2207 
2208 	if (!test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &cur_trans->flags)) {
2209 		int run_it = 0;
2210 
2211 		/* this mutex is also taken before trying to set
2212 		 * block groups readonly.  We need to make sure
2213 		 * that nobody has set a block group readonly
2214 		 * after a extents from that block group have been
2215 		 * allocated for cache files.  btrfs_set_block_group_ro
2216 		 * will wait for the transaction to commit if it
2217 		 * finds BTRFS_TRANS_DIRTY_BG_RUN set.
2218 		 *
2219 		 * The BTRFS_TRANS_DIRTY_BG_RUN flag is also used to make sure
2220 		 * only one process starts all the block group IO.  It wouldn't
2221 		 * hurt to have more than one go through, but there's no
2222 		 * real advantage to it either.
2223 		 */
2224 		mutex_lock(&fs_info->ro_block_group_mutex);
2225 		if (!test_and_set_bit(BTRFS_TRANS_DIRTY_BG_RUN,
2226 				      &cur_trans->flags))
2227 			run_it = 1;
2228 		mutex_unlock(&fs_info->ro_block_group_mutex);
2229 
2230 		if (run_it) {
2231 			ret = btrfs_start_dirty_block_groups(trans);
2232 			if (ret)
2233 				goto lockdep_trans_commit_start_release;
2234 		}
2235 	}
2236 
2237 	spin_lock(&fs_info->trans_lock);
2238 	if (cur_trans->state >= TRANS_STATE_COMMIT_PREP) {
2239 		enum btrfs_trans_state want_state = TRANS_STATE_COMPLETED;
2240 
2241 		add_pending_snapshot(trans);
2242 
2243 		spin_unlock(&fs_info->trans_lock);
2244 		refcount_inc(&cur_trans->use_count);
2245 
2246 		if (trans->in_fsync)
2247 			want_state = TRANS_STATE_SUPER_COMMITTED;
2248 
2249 		btrfs_trans_state_lockdep_release(fs_info,
2250 						  BTRFS_LOCKDEP_TRANS_COMMIT_PREP);
2251 		ret = btrfs_end_transaction(trans);
2252 		wait_for_commit(cur_trans, want_state);
2253 
2254 		if (TRANS_ABORTED(cur_trans))
2255 			ret = cur_trans->aborted;
2256 
2257 		btrfs_put_transaction(cur_trans);
2258 
2259 		return ret;
2260 	}
2261 
2262 	cur_trans->state = TRANS_STATE_COMMIT_PREP;
2263 	wake_up(&fs_info->transaction_blocked_wait);
2264 	btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_COMMIT_PREP);
2265 
2266 	if (cur_trans->list.prev != &fs_info->trans_list) {
2267 		enum btrfs_trans_state want_state = TRANS_STATE_COMPLETED;
2268 
2269 		if (trans->in_fsync)
2270 			want_state = TRANS_STATE_SUPER_COMMITTED;
2271 
2272 		prev_trans = list_entry(cur_trans->list.prev,
2273 					struct btrfs_transaction, list);
2274 		if (prev_trans->state < want_state) {
2275 			refcount_inc(&prev_trans->use_count);
2276 			spin_unlock(&fs_info->trans_lock);
2277 
2278 			wait_for_commit(prev_trans, want_state);
2279 
2280 			ret = READ_ONCE(prev_trans->aborted);
2281 
2282 			btrfs_put_transaction(prev_trans);
2283 			if (ret)
2284 				goto lockdep_release;
2285 			spin_lock(&fs_info->trans_lock);
2286 		}
2287 	} else {
2288 		/*
2289 		 * The previous transaction was aborted and was already removed
2290 		 * from the list of transactions at fs_info->trans_list. So we
2291 		 * abort to prevent writing a new superblock that reflects a
2292 		 * corrupt state (pointing to trees with unwritten nodes/leafs).
2293 		 */
2294 		if (BTRFS_FS_ERROR(fs_info)) {
2295 			spin_unlock(&fs_info->trans_lock);
2296 			ret = -EROFS;
2297 			goto lockdep_release;
2298 		}
2299 	}
2300 
2301 	cur_trans->state = TRANS_STATE_COMMIT_START;
2302 	wake_up(&fs_info->transaction_blocked_wait);
2303 	spin_unlock(&fs_info->trans_lock);
2304 
2305 	/*
2306 	 * Get the time spent on the work done by the commit thread and not
2307 	 * the time spent waiting on a previous commit
2308 	 */
2309 	start_time = ktime_get_ns();
2310 
2311 	extwriter_counter_dec(cur_trans, trans->type);
2312 
2313 	ret = btrfs_start_delalloc_flush(fs_info);
2314 	if (ret)
2315 		goto lockdep_release;
2316 
2317 	ret = btrfs_run_delayed_items(trans);
2318 	if (ret)
2319 		goto lockdep_release;
2320 
2321 	/*
2322 	 * The thread has started/joined the transaction thus it holds the
2323 	 * lockdep map as a reader. It has to release it before acquiring the
2324 	 * lockdep map as a writer.
2325 	 */
2326 	btrfs_lockdep_release(fs_info, btrfs_trans_num_extwriters);
2327 	btrfs_might_wait_for_event(fs_info, btrfs_trans_num_extwriters);
2328 	wait_event(cur_trans->writer_wait,
2329 		   extwriter_counter_read(cur_trans) == 0);
2330 
2331 	/* some pending stuffs might be added after the previous flush. */
2332 	ret = btrfs_run_delayed_items(trans);
2333 	if (ret) {
2334 		btrfs_lockdep_release(fs_info, btrfs_trans_num_writers);
2335 		goto cleanup_transaction;
2336 	}
2337 
2338 	btrfs_wait_delalloc_flush(fs_info);
2339 
2340 	/*
2341 	 * Wait for all ordered extents started by a fast fsync that joined this
2342 	 * transaction. Otherwise if this transaction commits before the ordered
2343 	 * extents complete we lose logged data after a power failure.
2344 	 */
2345 	btrfs_might_wait_for_event(fs_info, btrfs_trans_pending_ordered);
2346 	wait_event(cur_trans->pending_wait,
2347 		   atomic_read(&cur_trans->pending_ordered) == 0);
2348 
2349 	btrfs_scrub_pause(fs_info);
2350 	/*
2351 	 * Ok now we need to make sure to block out any other joins while we
2352 	 * commit the transaction.  We could have started a join before setting
2353 	 * COMMIT_DOING so make sure to wait for num_writers to == 1 again.
2354 	 */
2355 	spin_lock(&fs_info->trans_lock);
2356 	add_pending_snapshot(trans);
2357 	cur_trans->state = TRANS_STATE_COMMIT_DOING;
2358 	spin_unlock(&fs_info->trans_lock);
2359 
2360 	/*
2361 	 * The thread has started/joined the transaction thus it holds the
2362 	 * lockdep map as a reader. It has to release it before acquiring the
2363 	 * lockdep map as a writer.
2364 	 */
2365 	btrfs_lockdep_release(fs_info, btrfs_trans_num_writers);
2366 	btrfs_might_wait_for_event(fs_info, btrfs_trans_num_writers);
2367 	wait_event(cur_trans->writer_wait,
2368 		   atomic_read(&cur_trans->num_writers) == 1);
2369 
2370 	/*
2371 	 * Make lockdep happy by acquiring the state locks after
2372 	 * btrfs_trans_num_writers is released. If we acquired the state locks
2373 	 * before releasing the btrfs_trans_num_writers lock then lockdep would
2374 	 * complain because we did not follow the reverse order unlocking rule.
2375 	 */
2376 	btrfs_trans_state_lockdep_acquire(fs_info, BTRFS_LOCKDEP_TRANS_COMPLETED);
2377 	btrfs_trans_state_lockdep_acquire(fs_info, BTRFS_LOCKDEP_TRANS_SUPER_COMMITTED);
2378 	btrfs_trans_state_lockdep_acquire(fs_info, BTRFS_LOCKDEP_TRANS_UNBLOCKED);
2379 
2380 	/*
2381 	 * We've started the commit, clear the flag in case we were triggered to
2382 	 * do an async commit but somebody else started before the transaction
2383 	 * kthread could do the work.
2384 	 */
2385 	clear_bit(BTRFS_FS_COMMIT_TRANS, &fs_info->flags);
2386 
2387 	if (TRANS_ABORTED(cur_trans)) {
2388 		ret = cur_trans->aborted;
2389 		btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_UNBLOCKED);
2390 		goto scrub_continue;
2391 	}
2392 	/*
2393 	 * the reloc mutex makes sure that we stop
2394 	 * the balancing code from coming in and moving
2395 	 * extents around in the middle of the commit
2396 	 */
2397 	mutex_lock(&fs_info->reloc_mutex);
2398 
2399 	/*
2400 	 * We needn't worry about the delayed items because we will
2401 	 * deal with them in create_pending_snapshot(), which is the
2402 	 * core function of the snapshot creation.
2403 	 */
2404 	ret = create_pending_snapshots(trans);
2405 	if (ret)
2406 		goto unlock_reloc;
2407 
2408 	/*
2409 	 * We insert the dir indexes of the snapshots and update the inode
2410 	 * of the snapshots' parents after the snapshot creation, so there
2411 	 * are some delayed items which are not dealt with. Now deal with
2412 	 * them.
2413 	 *
2414 	 * We needn't worry that this operation will corrupt the snapshots,
2415 	 * because all the tree which are snapshoted will be forced to COW
2416 	 * the nodes and leaves.
2417 	 */
2418 	ret = btrfs_run_delayed_items(trans);
2419 	if (ret)
2420 		goto unlock_reloc;
2421 
2422 	ret = btrfs_run_delayed_refs(trans, U64_MAX);
2423 	if (ret)
2424 		goto unlock_reloc;
2425 
2426 	/*
2427 	 * make sure none of the code above managed to slip in a
2428 	 * delayed item
2429 	 */
2430 	btrfs_assert_delayed_root_empty(fs_info);
2431 
2432 	WARN_ON(cur_trans != trans->transaction);
2433 
2434 	ret = commit_fs_roots(trans);
2435 	if (ret)
2436 		goto unlock_reloc;
2437 
2438 	/* commit_fs_roots gets rid of all the tree log roots, it is now
2439 	 * safe to free the root of tree log roots
2440 	 */
2441 	btrfs_free_log_root_tree(trans, fs_info);
2442 
2443 	/*
2444 	 * Since fs roots are all committed, we can get a quite accurate
2445 	 * new_roots. So let's do quota accounting.
2446 	 */
2447 	ret = btrfs_qgroup_account_extents(trans);
2448 	if (ret < 0)
2449 		goto unlock_reloc;
2450 
2451 	ret = commit_cowonly_roots(trans);
2452 	if (ret)
2453 		goto unlock_reloc;
2454 
2455 	/*
2456 	 * The tasks which save the space cache and inode cache may also
2457 	 * update ->aborted, check it.
2458 	 */
2459 	if (TRANS_ABORTED(cur_trans)) {
2460 		ret = cur_trans->aborted;
2461 		goto unlock_reloc;
2462 	}
2463 
2464 	cur_trans = fs_info->running_transaction;
2465 
2466 	btrfs_set_root_node(&fs_info->tree_root->root_item,
2467 			    fs_info->tree_root->node);
2468 	list_add_tail(&fs_info->tree_root->dirty_list,
2469 		      &cur_trans->switch_commits);
2470 
2471 	btrfs_set_root_node(&fs_info->chunk_root->root_item,
2472 			    fs_info->chunk_root->node);
2473 	list_add_tail(&fs_info->chunk_root->dirty_list,
2474 		      &cur_trans->switch_commits);
2475 
2476 	if (btrfs_fs_incompat(fs_info, EXTENT_TREE_V2)) {
2477 		btrfs_set_root_node(&fs_info->block_group_root->root_item,
2478 				    fs_info->block_group_root->node);
2479 		list_add_tail(&fs_info->block_group_root->dirty_list,
2480 			      &cur_trans->switch_commits);
2481 	}
2482 
2483 	switch_commit_roots(trans);
2484 
2485 	ASSERT(list_empty(&cur_trans->dirty_bgs));
2486 	ASSERT(list_empty(&cur_trans->io_bgs));
2487 	update_super_roots(fs_info);
2488 
2489 	btrfs_set_super_log_root(fs_info->super_copy, 0);
2490 	btrfs_set_super_log_root_level(fs_info->super_copy, 0);
2491 	memcpy(fs_info->super_for_commit, fs_info->super_copy,
2492 	       sizeof(*fs_info->super_copy));
2493 
2494 	btrfs_commit_device_sizes(cur_trans);
2495 
2496 	clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags);
2497 	clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags);
2498 
2499 	btrfs_trans_release_chunk_metadata(trans);
2500 
2501 	/*
2502 	 * Before changing the transaction state to TRANS_STATE_UNBLOCKED and
2503 	 * setting fs_info->running_transaction to NULL, lock tree_log_mutex to
2504 	 * make sure that before we commit our superblock, no other task can
2505 	 * start a new transaction and commit a log tree before we commit our
2506 	 * superblock. Anyone trying to commit a log tree locks this mutex before
2507 	 * writing its superblock.
2508 	 */
2509 	mutex_lock(&fs_info->tree_log_mutex);
2510 
2511 	spin_lock(&fs_info->trans_lock);
2512 	cur_trans->state = TRANS_STATE_UNBLOCKED;
2513 	fs_info->running_transaction = NULL;
2514 	spin_unlock(&fs_info->trans_lock);
2515 	mutex_unlock(&fs_info->reloc_mutex);
2516 
2517 	wake_up(&fs_info->transaction_wait);
2518 	btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_UNBLOCKED);
2519 
2520 	/* If we have features changed, wake up the cleaner to update sysfs. */
2521 	if (test_bit(BTRFS_FS_FEATURE_CHANGED, &fs_info->flags) &&
2522 	    fs_info->cleaner_kthread)
2523 		wake_up_process(fs_info->cleaner_kthread);
2524 
2525 	ret = btrfs_write_and_wait_transaction(trans);
2526 	if (ret) {
2527 		btrfs_handle_fs_error(fs_info, ret,
2528 				      "Error while writing out transaction");
2529 		mutex_unlock(&fs_info->tree_log_mutex);
2530 		goto scrub_continue;
2531 	}
2532 
2533 	ret = write_all_supers(fs_info, 0);
2534 	/*
2535 	 * the super is written, we can safely allow the tree-loggers
2536 	 * to go about their business
2537 	 */
2538 	mutex_unlock(&fs_info->tree_log_mutex);
2539 	if (ret)
2540 		goto scrub_continue;
2541 
2542 	/*
2543 	 * We needn't acquire the lock here because there is no other task
2544 	 * which can change it.
2545 	 */
2546 	cur_trans->state = TRANS_STATE_SUPER_COMMITTED;
2547 	wake_up(&cur_trans->commit_wait);
2548 	btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_SUPER_COMMITTED);
2549 
2550 	btrfs_finish_extent_commit(trans);
2551 
2552 	if (test_bit(BTRFS_TRANS_HAVE_FREE_BGS, &cur_trans->flags))
2553 		btrfs_clear_space_info_full(fs_info);
2554 
2555 	btrfs_set_last_trans_committed(fs_info, cur_trans->transid);
2556 	/*
2557 	 * We needn't acquire the lock here because there is no other task
2558 	 * which can change it.
2559 	 */
2560 	cur_trans->state = TRANS_STATE_COMPLETED;
2561 	wake_up(&cur_trans->commit_wait);
2562 	btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_COMPLETED);
2563 
2564 	spin_lock(&fs_info->trans_lock);
2565 	list_del_init(&cur_trans->list);
2566 	spin_unlock(&fs_info->trans_lock);
2567 
2568 	btrfs_put_transaction(cur_trans);
2569 	btrfs_put_transaction(cur_trans);
2570 
2571 	if (trans->type & __TRANS_FREEZABLE)
2572 		sb_end_intwrite(fs_info->sb);
2573 
2574 	trace_btrfs_transaction_commit(fs_info);
2575 
2576 	interval = ktime_get_ns() - start_time;
2577 
2578 	btrfs_scrub_continue(fs_info);
2579 
2580 	if (current->journal_info == trans)
2581 		current->journal_info = NULL;
2582 
2583 	kmem_cache_free(btrfs_trans_handle_cachep, trans);
2584 
2585 	update_commit_stats(fs_info, interval);
2586 
2587 	return ret;
2588 
2589 unlock_reloc:
2590 	mutex_unlock(&fs_info->reloc_mutex);
2591 	btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_UNBLOCKED);
2592 scrub_continue:
2593 	btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_SUPER_COMMITTED);
2594 	btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_COMPLETED);
2595 	btrfs_scrub_continue(fs_info);
2596 cleanup_transaction:
2597 	btrfs_trans_release_metadata(trans);
2598 	btrfs_cleanup_pending_block_groups(trans);
2599 	btrfs_trans_release_chunk_metadata(trans);
2600 	trans->block_rsv = NULL;
2601 	btrfs_warn(fs_info, "Skipping commit of aborted transaction.");
2602 	if (current->journal_info == trans)
2603 		current->journal_info = NULL;
2604 	cleanup_transaction(trans, ret);
2605 
2606 	return ret;
2607 
2608 lockdep_release:
2609 	btrfs_lockdep_release(fs_info, btrfs_trans_num_extwriters);
2610 	btrfs_lockdep_release(fs_info, btrfs_trans_num_writers);
2611 	goto cleanup_transaction;
2612 
2613 lockdep_trans_commit_start_release:
2614 	btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_COMMIT_PREP);
2615 	btrfs_end_transaction(trans);
2616 	return ret;
2617 }
2618 
2619 /*
2620  * return < 0 if error
2621  * 0 if there are no more dead_roots at the time of call
2622  * 1 there are more to be processed, call me again
2623  *
2624  * The return value indicates there are certainly more snapshots to delete, but
2625  * if there comes a new one during processing, it may return 0. We don't mind,
2626  * because btrfs_commit_super will poke cleaner thread and it will process it a
2627  * few seconds later.
2628  */
2629 int btrfs_clean_one_deleted_snapshot(struct btrfs_fs_info *fs_info)
2630 {
2631 	struct btrfs_root *root;
2632 	int ret;
2633 
2634 	spin_lock(&fs_info->trans_lock);
2635 	if (list_empty(&fs_info->dead_roots)) {
2636 		spin_unlock(&fs_info->trans_lock);
2637 		return 0;
2638 	}
2639 	root = list_first_entry(&fs_info->dead_roots,
2640 			struct btrfs_root, root_list);
2641 	list_del_init(&root->root_list);
2642 	spin_unlock(&fs_info->trans_lock);
2643 
2644 	btrfs_debug(fs_info, "cleaner removing %llu", root->root_key.objectid);
2645 
2646 	btrfs_kill_all_delayed_nodes(root);
2647 
2648 	if (btrfs_header_backref_rev(root->node) <
2649 			BTRFS_MIXED_BACKREF_REV)
2650 		ret = btrfs_drop_snapshot(root, 0, 0);
2651 	else
2652 		ret = btrfs_drop_snapshot(root, 1, 0);
2653 
2654 	btrfs_put_root(root);
2655 	return (ret < 0) ? 0 : 1;
2656 }
2657 
2658 /*
2659  * We only mark the transaction aborted and then set the file system read-only.
2660  * This will prevent new transactions from starting or trying to join this
2661  * one.
2662  *
2663  * This means that error recovery at the call site is limited to freeing
2664  * any local memory allocations and passing the error code up without
2665  * further cleanup. The transaction should complete as it normally would
2666  * in the call path but will return -EIO.
2667  *
2668  * We'll complete the cleanup in btrfs_end_transaction and
2669  * btrfs_commit_transaction.
2670  */
2671 void __cold __btrfs_abort_transaction(struct btrfs_trans_handle *trans,
2672 				      const char *function,
2673 				      unsigned int line, int error, bool first_hit)
2674 {
2675 	struct btrfs_fs_info *fs_info = trans->fs_info;
2676 
2677 	WRITE_ONCE(trans->aborted, error);
2678 	WRITE_ONCE(trans->transaction->aborted, error);
2679 	if (first_hit && error == -ENOSPC)
2680 		btrfs_dump_space_info_for_trans_abort(fs_info);
2681 	/* Wake up anybody who may be waiting on this transaction */
2682 	wake_up(&fs_info->transaction_wait);
2683 	wake_up(&fs_info->transaction_blocked_wait);
2684 	__btrfs_handle_fs_error(fs_info, function, line, error, NULL);
2685 }
2686 
2687 int __init btrfs_transaction_init(void)
2688 {
2689 	btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
2690 			sizeof(struct btrfs_trans_handle), 0,
2691 			SLAB_TEMPORARY | SLAB_MEM_SPREAD, NULL);
2692 	if (!btrfs_trans_handle_cachep)
2693 		return -ENOMEM;
2694 	return 0;
2695 }
2696 
2697 void __cold btrfs_transaction_exit(void)
2698 {
2699 	kmem_cache_destroy(btrfs_trans_handle_cachep);
2700 }
2701