xref: /linux/fs/btrfs/transaction.c (revision ff19a8dee196d757dbc32a946843260f0b784ca3)
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 #define BTRFS_ROOT_TRANS_TAG 0
41 
42 /*
43  * Transaction states and transitions
44  *
45  * No running transaction (fs tree blocks are not modified)
46  * |
47  * | To next stage:
48  * |  Call start_transaction() variants. Except btrfs_join_transaction_nostart().
49  * V
50  * Transaction N [[TRANS_STATE_RUNNING]]
51  * |
52  * | New trans handles can be attached to transaction N by calling all
53  * | start_transaction() variants.
54  * |
55  * | To next stage:
56  * |  Call btrfs_commit_transaction() on any trans handle attached to
57  * |  transaction N
58  * V
59  * Transaction N [[TRANS_STATE_COMMIT_START]]
60  * |
61  * | Will wait for previous running transaction to completely finish if there
62  * | is one
63  * |
64  * | Then one of the following happes:
65  * | - Wait for all other trans handle holders to release.
66  * |   The btrfs_commit_transaction() caller will do the commit work.
67  * | - Wait for current transaction to be committed by others.
68  * |   Other btrfs_commit_transaction() caller will do the commit work.
69  * |
70  * | At this stage, only btrfs_join_transaction*() variants can attach
71  * | to this running transaction.
72  * | All other variants will wait for current one to finish and attach to
73  * | transaction N+1.
74  * |
75  * | To next stage:
76  * |  Caller is chosen to commit transaction N, and all other trans handle
77  * |  haven been released.
78  * V
79  * Transaction N [[TRANS_STATE_COMMIT_DOING]]
80  * |
81  * | The heavy lifting transaction work is started.
82  * | From running delayed refs (modifying extent tree) to creating pending
83  * | snapshots, running qgroups.
84  * | In short, modify supporting trees to reflect modifications of subvolume
85  * | trees.
86  * |
87  * | At this stage, all start_transaction() calls will wait for this
88  * | transaction to finish and attach to transaction N+1.
89  * |
90  * | To next stage:
91  * |  Until all supporting trees are updated.
92  * V
93  * Transaction N [[TRANS_STATE_UNBLOCKED]]
94  * |						    Transaction N+1
95  * | All needed trees are modified, thus we only    [[TRANS_STATE_RUNNING]]
96  * | need to write them back to disk and update	    |
97  * | super blocks.				    |
98  * |						    |
99  * | At this stage, new transaction is allowed to   |
100  * | start.					    |
101  * | All new start_transaction() calls will be	    |
102  * | attached to transid N+1.			    |
103  * |						    |
104  * | To next stage:				    |
105  * |  Until all tree blocks are super blocks are    |
106  * |  written to block devices			    |
107  * V						    |
108  * Transaction N [[TRANS_STATE_COMPLETED]]	    V
109  *   All tree blocks and super blocks are written.  Transaction N+1
110  *   This transaction is finished and all its	    [[TRANS_STATE_COMMIT_START]]
111  *   data structures will be cleaned up.	    | Life goes on
112  */
113 static const unsigned int btrfs_blocked_trans_types[TRANS_STATE_MAX] = {
114 	[TRANS_STATE_RUNNING]		= 0U,
115 	[TRANS_STATE_COMMIT_START]	= (__TRANS_START | __TRANS_ATTACH),
116 	[TRANS_STATE_COMMIT_DOING]	= (__TRANS_START |
117 					   __TRANS_ATTACH |
118 					   __TRANS_JOIN |
119 					   __TRANS_JOIN_NOSTART),
120 	[TRANS_STATE_UNBLOCKED]		= (__TRANS_START |
121 					   __TRANS_ATTACH |
122 					   __TRANS_JOIN |
123 					   __TRANS_JOIN_NOLOCK |
124 					   __TRANS_JOIN_NOSTART),
125 	[TRANS_STATE_SUPER_COMMITTED]	= (__TRANS_START |
126 					   __TRANS_ATTACH |
127 					   __TRANS_JOIN |
128 					   __TRANS_JOIN_NOLOCK |
129 					   __TRANS_JOIN_NOSTART),
130 	[TRANS_STATE_COMPLETED]		= (__TRANS_START |
131 					   __TRANS_ATTACH |
132 					   __TRANS_JOIN |
133 					   __TRANS_JOIN_NOLOCK |
134 					   __TRANS_JOIN_NOSTART),
135 };
136 
137 void btrfs_put_transaction(struct btrfs_transaction *transaction)
138 {
139 	WARN_ON(refcount_read(&transaction->use_count) == 0);
140 	if (refcount_dec_and_test(&transaction->use_count)) {
141 		BUG_ON(!list_empty(&transaction->list));
142 		WARN_ON(!RB_EMPTY_ROOT(
143 				&transaction->delayed_refs.href_root.rb_root));
144 		WARN_ON(!RB_EMPTY_ROOT(
145 				&transaction->delayed_refs.dirty_extent_root));
146 		if (transaction->delayed_refs.pending_csums)
147 			btrfs_err(transaction->fs_info,
148 				  "pending csums is %llu",
149 				  transaction->delayed_refs.pending_csums);
150 		/*
151 		 * If any block groups are found in ->deleted_bgs then it's
152 		 * because the transaction was aborted and a commit did not
153 		 * happen (things failed before writing the new superblock
154 		 * and calling btrfs_finish_extent_commit()), so we can not
155 		 * discard the physical locations of the block groups.
156 		 */
157 		while (!list_empty(&transaction->deleted_bgs)) {
158 			struct btrfs_block_group *cache;
159 
160 			cache = list_first_entry(&transaction->deleted_bgs,
161 						 struct btrfs_block_group,
162 						 bg_list);
163 			list_del_init(&cache->bg_list);
164 			btrfs_unfreeze_block_group(cache);
165 			btrfs_put_block_group(cache);
166 		}
167 		WARN_ON(!list_empty(&transaction->dev_update_list));
168 		kfree(transaction);
169 	}
170 }
171 
172 static noinline void switch_commit_roots(struct btrfs_trans_handle *trans)
173 {
174 	struct btrfs_transaction *cur_trans = trans->transaction;
175 	struct btrfs_fs_info *fs_info = trans->fs_info;
176 	struct btrfs_root *root, *tmp;
177 
178 	/*
179 	 * At this point no one can be using this transaction to modify any tree
180 	 * and no one can start another transaction to modify any tree either.
181 	 */
182 	ASSERT(cur_trans->state == TRANS_STATE_COMMIT_DOING);
183 
184 	down_write(&fs_info->commit_root_sem);
185 
186 	if (test_bit(BTRFS_FS_RELOC_RUNNING, &fs_info->flags))
187 		fs_info->last_reloc_trans = trans->transid;
188 
189 	list_for_each_entry_safe(root, tmp, &cur_trans->switch_commits,
190 				 dirty_list) {
191 		list_del_init(&root->dirty_list);
192 		free_extent_buffer(root->commit_root);
193 		root->commit_root = btrfs_root_node(root);
194 		extent_io_tree_release(&root->dirty_log_pages);
195 		btrfs_qgroup_clean_swapped_blocks(root);
196 	}
197 
198 	/* We can free old roots now. */
199 	spin_lock(&cur_trans->dropped_roots_lock);
200 	while (!list_empty(&cur_trans->dropped_roots)) {
201 		root = list_first_entry(&cur_trans->dropped_roots,
202 					struct btrfs_root, root_list);
203 		list_del_init(&root->root_list);
204 		spin_unlock(&cur_trans->dropped_roots_lock);
205 		btrfs_free_log(trans, root);
206 		btrfs_drop_and_free_fs_root(fs_info, root);
207 		spin_lock(&cur_trans->dropped_roots_lock);
208 	}
209 	spin_unlock(&cur_trans->dropped_roots_lock);
210 
211 	up_write(&fs_info->commit_root_sem);
212 }
213 
214 static inline void extwriter_counter_inc(struct btrfs_transaction *trans,
215 					 unsigned int type)
216 {
217 	if (type & TRANS_EXTWRITERS)
218 		atomic_inc(&trans->num_extwriters);
219 }
220 
221 static inline void extwriter_counter_dec(struct btrfs_transaction *trans,
222 					 unsigned int type)
223 {
224 	if (type & TRANS_EXTWRITERS)
225 		atomic_dec(&trans->num_extwriters);
226 }
227 
228 static inline void extwriter_counter_init(struct btrfs_transaction *trans,
229 					  unsigned int type)
230 {
231 	atomic_set(&trans->num_extwriters, ((type & TRANS_EXTWRITERS) ? 1 : 0));
232 }
233 
234 static inline int extwriter_counter_read(struct btrfs_transaction *trans)
235 {
236 	return atomic_read(&trans->num_extwriters);
237 }
238 
239 /*
240  * To be called after doing the chunk btree updates right after allocating a new
241  * chunk (after btrfs_chunk_alloc_add_chunk_item() is called), when removing a
242  * chunk after all chunk btree updates and after finishing the second phase of
243  * chunk allocation (btrfs_create_pending_block_groups()) in case some block
244  * group had its chunk item insertion delayed to the second phase.
245  */
246 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
247 {
248 	struct btrfs_fs_info *fs_info = trans->fs_info;
249 
250 	if (!trans->chunk_bytes_reserved)
251 		return;
252 
253 	btrfs_block_rsv_release(fs_info, &fs_info->chunk_block_rsv,
254 				trans->chunk_bytes_reserved, NULL);
255 	trans->chunk_bytes_reserved = 0;
256 }
257 
258 /*
259  * either allocate a new transaction or hop into the existing one
260  */
261 static noinline int join_transaction(struct btrfs_fs_info *fs_info,
262 				     unsigned int type)
263 {
264 	struct btrfs_transaction *cur_trans;
265 
266 	spin_lock(&fs_info->trans_lock);
267 loop:
268 	/* The file system has been taken offline. No new transactions. */
269 	if (BTRFS_FS_ERROR(fs_info)) {
270 		spin_unlock(&fs_info->trans_lock);
271 		return -EROFS;
272 	}
273 
274 	cur_trans = fs_info->running_transaction;
275 	if (cur_trans) {
276 		if (TRANS_ABORTED(cur_trans)) {
277 			spin_unlock(&fs_info->trans_lock);
278 			return cur_trans->aborted;
279 		}
280 		if (btrfs_blocked_trans_types[cur_trans->state] & type) {
281 			spin_unlock(&fs_info->trans_lock);
282 			return -EBUSY;
283 		}
284 		refcount_inc(&cur_trans->use_count);
285 		atomic_inc(&cur_trans->num_writers);
286 		extwriter_counter_inc(cur_trans, type);
287 		spin_unlock(&fs_info->trans_lock);
288 		btrfs_lockdep_acquire(fs_info, btrfs_trans_num_writers);
289 		btrfs_lockdep_acquire(fs_info, btrfs_trans_num_extwriters);
290 		return 0;
291 	}
292 	spin_unlock(&fs_info->trans_lock);
293 
294 	/*
295 	 * If we are ATTACH, we just want to catch the current transaction,
296 	 * and commit it. If there is no transaction, just return ENOENT.
297 	 */
298 	if (type == TRANS_ATTACH)
299 		return -ENOENT;
300 
301 	/*
302 	 * JOIN_NOLOCK only happens during the transaction commit, so
303 	 * it is impossible that ->running_transaction is NULL
304 	 */
305 	BUG_ON(type == TRANS_JOIN_NOLOCK);
306 
307 	cur_trans = kmalloc(sizeof(*cur_trans), GFP_NOFS);
308 	if (!cur_trans)
309 		return -ENOMEM;
310 
311 	btrfs_lockdep_acquire(fs_info, btrfs_trans_num_writers);
312 	btrfs_lockdep_acquire(fs_info, btrfs_trans_num_extwriters);
313 
314 	spin_lock(&fs_info->trans_lock);
315 	if (fs_info->running_transaction) {
316 		/*
317 		 * someone started a transaction after we unlocked.  Make sure
318 		 * to redo the checks above
319 		 */
320 		btrfs_lockdep_release(fs_info, btrfs_trans_num_extwriters);
321 		btrfs_lockdep_release(fs_info, btrfs_trans_num_writers);
322 		kfree(cur_trans);
323 		goto loop;
324 	} else if (BTRFS_FS_ERROR(fs_info)) {
325 		spin_unlock(&fs_info->trans_lock);
326 		btrfs_lockdep_release(fs_info, btrfs_trans_num_extwriters);
327 		btrfs_lockdep_release(fs_info, btrfs_trans_num_writers);
328 		kfree(cur_trans);
329 		return -EROFS;
330 	}
331 
332 	cur_trans->fs_info = fs_info;
333 	atomic_set(&cur_trans->pending_ordered, 0);
334 	init_waitqueue_head(&cur_trans->pending_wait);
335 	atomic_set(&cur_trans->num_writers, 1);
336 	extwriter_counter_init(cur_trans, type);
337 	init_waitqueue_head(&cur_trans->writer_wait);
338 	init_waitqueue_head(&cur_trans->commit_wait);
339 	cur_trans->state = TRANS_STATE_RUNNING;
340 	/*
341 	 * One for this trans handle, one so it will live on until we
342 	 * commit the transaction.
343 	 */
344 	refcount_set(&cur_trans->use_count, 2);
345 	cur_trans->flags = 0;
346 	cur_trans->start_time = ktime_get_seconds();
347 
348 	memset(&cur_trans->delayed_refs, 0, sizeof(cur_trans->delayed_refs));
349 
350 	cur_trans->delayed_refs.href_root = RB_ROOT_CACHED;
351 	cur_trans->delayed_refs.dirty_extent_root = RB_ROOT;
352 	atomic_set(&cur_trans->delayed_refs.num_entries, 0);
353 
354 	/*
355 	 * although the tree mod log is per file system and not per transaction,
356 	 * the log must never go across transaction boundaries.
357 	 */
358 	smp_mb();
359 	if (!list_empty(&fs_info->tree_mod_seq_list))
360 		WARN(1, KERN_ERR "BTRFS: tree_mod_seq_list not empty when creating a fresh transaction\n");
361 	if (!RB_EMPTY_ROOT(&fs_info->tree_mod_log))
362 		WARN(1, KERN_ERR "BTRFS: tree_mod_log rb tree not empty when creating a fresh transaction\n");
363 	atomic64_set(&fs_info->tree_mod_seq, 0);
364 
365 	spin_lock_init(&cur_trans->delayed_refs.lock);
366 
367 	INIT_LIST_HEAD(&cur_trans->pending_snapshots);
368 	INIT_LIST_HEAD(&cur_trans->dev_update_list);
369 	INIT_LIST_HEAD(&cur_trans->switch_commits);
370 	INIT_LIST_HEAD(&cur_trans->dirty_bgs);
371 	INIT_LIST_HEAD(&cur_trans->io_bgs);
372 	INIT_LIST_HEAD(&cur_trans->dropped_roots);
373 	mutex_init(&cur_trans->cache_write_mutex);
374 	spin_lock_init(&cur_trans->dirty_bgs_lock);
375 	INIT_LIST_HEAD(&cur_trans->deleted_bgs);
376 	spin_lock_init(&cur_trans->dropped_roots_lock);
377 	INIT_LIST_HEAD(&cur_trans->releasing_ebs);
378 	spin_lock_init(&cur_trans->releasing_ebs_lock);
379 	list_add_tail(&cur_trans->list, &fs_info->trans_list);
380 	extent_io_tree_init(fs_info, &cur_trans->dirty_pages,
381 			IO_TREE_TRANS_DIRTY_PAGES);
382 	extent_io_tree_init(fs_info, &cur_trans->pinned_extents,
383 			IO_TREE_FS_PINNED_EXTENTS);
384 	fs_info->generation++;
385 	cur_trans->transid = fs_info->generation;
386 	fs_info->running_transaction = cur_trans;
387 	cur_trans->aborted = 0;
388 	spin_unlock(&fs_info->trans_lock);
389 
390 	return 0;
391 }
392 
393 /*
394  * This does all the record keeping required to make sure that a shareable root
395  * is properly recorded in a given transaction.  This is required to make sure
396  * the old root from before we joined the transaction is deleted when the
397  * transaction commits.
398  */
399 static int record_root_in_trans(struct btrfs_trans_handle *trans,
400 			       struct btrfs_root *root,
401 			       int force)
402 {
403 	struct btrfs_fs_info *fs_info = root->fs_info;
404 	int ret = 0;
405 
406 	if ((test_bit(BTRFS_ROOT_SHAREABLE, &root->state) &&
407 	    root->last_trans < trans->transid) || force) {
408 		WARN_ON(!force && root->commit_root != root->node);
409 
410 		/*
411 		 * see below for IN_TRANS_SETUP usage rules
412 		 * we have the reloc mutex held now, so there
413 		 * is only one writer in this function
414 		 */
415 		set_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
416 
417 		/* make sure readers find IN_TRANS_SETUP before
418 		 * they find our root->last_trans update
419 		 */
420 		smp_wmb();
421 
422 		spin_lock(&fs_info->fs_roots_radix_lock);
423 		if (root->last_trans == trans->transid && !force) {
424 			spin_unlock(&fs_info->fs_roots_radix_lock);
425 			return 0;
426 		}
427 		radix_tree_tag_set(&fs_info->fs_roots_radix,
428 				   (unsigned long)root->root_key.objectid,
429 				   BTRFS_ROOT_TRANS_TAG);
430 		spin_unlock(&fs_info->fs_roots_radix_lock);
431 		root->last_trans = trans->transid;
432 
433 		/* this is pretty tricky.  We don't want to
434 		 * take the relocation lock in btrfs_record_root_in_trans
435 		 * unless we're really doing the first setup for this root in
436 		 * this transaction.
437 		 *
438 		 * Normally we'd use root->last_trans as a flag to decide
439 		 * if we want to take the expensive mutex.
440 		 *
441 		 * But, we have to set root->last_trans before we
442 		 * init the relocation root, otherwise, we trip over warnings
443 		 * in ctree.c.  The solution used here is to flag ourselves
444 		 * with root IN_TRANS_SETUP.  When this is 1, we're still
445 		 * fixing up the reloc trees and everyone must wait.
446 		 *
447 		 * When this is zero, they can trust root->last_trans and fly
448 		 * through btrfs_record_root_in_trans without having to take the
449 		 * lock.  smp_wmb() makes sure that all the writes above are
450 		 * done before we pop in the zero below
451 		 */
452 		ret = btrfs_init_reloc_root(trans, root);
453 		smp_mb__before_atomic();
454 		clear_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
455 	}
456 	return ret;
457 }
458 
459 
460 void btrfs_add_dropped_root(struct btrfs_trans_handle *trans,
461 			    struct btrfs_root *root)
462 {
463 	struct btrfs_fs_info *fs_info = root->fs_info;
464 	struct btrfs_transaction *cur_trans = trans->transaction;
465 
466 	/* Add ourselves to the transaction dropped list */
467 	spin_lock(&cur_trans->dropped_roots_lock);
468 	list_add_tail(&root->root_list, &cur_trans->dropped_roots);
469 	spin_unlock(&cur_trans->dropped_roots_lock);
470 
471 	/* Make sure we don't try to update the root at commit time */
472 	spin_lock(&fs_info->fs_roots_radix_lock);
473 	radix_tree_tag_clear(&fs_info->fs_roots_radix,
474 			     (unsigned long)root->root_key.objectid,
475 			     BTRFS_ROOT_TRANS_TAG);
476 	spin_unlock(&fs_info->fs_roots_radix_lock);
477 }
478 
479 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
480 			       struct btrfs_root *root)
481 {
482 	struct btrfs_fs_info *fs_info = root->fs_info;
483 	int ret;
484 
485 	if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
486 		return 0;
487 
488 	/*
489 	 * see record_root_in_trans for comments about IN_TRANS_SETUP usage
490 	 * and barriers
491 	 */
492 	smp_rmb();
493 	if (root->last_trans == trans->transid &&
494 	    !test_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state))
495 		return 0;
496 
497 	mutex_lock(&fs_info->reloc_mutex);
498 	ret = record_root_in_trans(trans, root, 0);
499 	mutex_unlock(&fs_info->reloc_mutex);
500 
501 	return ret;
502 }
503 
504 static inline int is_transaction_blocked(struct btrfs_transaction *trans)
505 {
506 	return (trans->state >= TRANS_STATE_COMMIT_START &&
507 		trans->state < TRANS_STATE_UNBLOCKED &&
508 		!TRANS_ABORTED(trans));
509 }
510 
511 /* wait for commit against the current transaction to become unblocked
512  * when this is done, it is safe to start a new transaction, but the current
513  * transaction might not be fully on disk.
514  */
515 static void wait_current_trans(struct btrfs_fs_info *fs_info)
516 {
517 	struct btrfs_transaction *cur_trans;
518 
519 	spin_lock(&fs_info->trans_lock);
520 	cur_trans = fs_info->running_transaction;
521 	if (cur_trans && is_transaction_blocked(cur_trans)) {
522 		refcount_inc(&cur_trans->use_count);
523 		spin_unlock(&fs_info->trans_lock);
524 
525 		btrfs_might_wait_for_state(fs_info, BTRFS_LOCKDEP_TRANS_UNBLOCKED);
526 		wait_event(fs_info->transaction_wait,
527 			   cur_trans->state >= TRANS_STATE_UNBLOCKED ||
528 			   TRANS_ABORTED(cur_trans));
529 		btrfs_put_transaction(cur_trans);
530 	} else {
531 		spin_unlock(&fs_info->trans_lock);
532 	}
533 }
534 
535 static int may_wait_transaction(struct btrfs_fs_info *fs_info, int type)
536 {
537 	if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
538 		return 0;
539 
540 	if (type == TRANS_START)
541 		return 1;
542 
543 	return 0;
544 }
545 
546 static inline bool need_reserve_reloc_root(struct btrfs_root *root)
547 {
548 	struct btrfs_fs_info *fs_info = root->fs_info;
549 
550 	if (!fs_info->reloc_ctl ||
551 	    !test_bit(BTRFS_ROOT_SHAREABLE, &root->state) ||
552 	    root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
553 	    root->reloc_root)
554 		return false;
555 
556 	return true;
557 }
558 
559 static struct btrfs_trans_handle *
560 start_transaction(struct btrfs_root *root, unsigned int num_items,
561 		  unsigned int type, enum btrfs_reserve_flush_enum flush,
562 		  bool enforce_qgroups)
563 {
564 	struct btrfs_fs_info *fs_info = root->fs_info;
565 	struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
566 	struct btrfs_trans_handle *h;
567 	struct btrfs_transaction *cur_trans;
568 	u64 num_bytes = 0;
569 	u64 qgroup_reserved = 0;
570 	bool reloc_reserved = false;
571 	bool do_chunk_alloc = false;
572 	int ret;
573 
574 	if (BTRFS_FS_ERROR(fs_info))
575 		return ERR_PTR(-EROFS);
576 
577 	if (current->journal_info) {
578 		WARN_ON(type & TRANS_EXTWRITERS);
579 		h = current->journal_info;
580 		refcount_inc(&h->use_count);
581 		WARN_ON(refcount_read(&h->use_count) > 2);
582 		h->orig_rsv = h->block_rsv;
583 		h->block_rsv = NULL;
584 		goto got_it;
585 	}
586 
587 	/*
588 	 * Do the reservation before we join the transaction so we can do all
589 	 * the appropriate flushing if need be.
590 	 */
591 	if (num_items && root != fs_info->chunk_root) {
592 		struct btrfs_block_rsv *rsv = &fs_info->trans_block_rsv;
593 		u64 delayed_refs_bytes = 0;
594 
595 		qgroup_reserved = num_items * fs_info->nodesize;
596 		ret = btrfs_qgroup_reserve_meta_pertrans(root, qgroup_reserved,
597 				enforce_qgroups);
598 		if (ret)
599 			return ERR_PTR(ret);
600 
601 		/*
602 		 * We want to reserve all the bytes we may need all at once, so
603 		 * we only do 1 enospc flushing cycle per transaction start.  We
604 		 * accomplish this by simply assuming we'll do 2 x num_items
605 		 * worth of delayed refs updates in this trans handle, and
606 		 * refill that amount for whatever is missing in the reserve.
607 		 */
608 		num_bytes = btrfs_calc_insert_metadata_size(fs_info, num_items);
609 		if (flush == BTRFS_RESERVE_FLUSH_ALL &&
610 		    btrfs_block_rsv_full(delayed_refs_rsv) == 0) {
611 			delayed_refs_bytes = num_bytes;
612 			num_bytes <<= 1;
613 		}
614 
615 		/*
616 		 * Do the reservation for the relocation root creation
617 		 */
618 		if (need_reserve_reloc_root(root)) {
619 			num_bytes += fs_info->nodesize;
620 			reloc_reserved = true;
621 		}
622 
623 		ret = btrfs_block_rsv_add(fs_info, rsv, num_bytes, flush);
624 		if (ret)
625 			goto reserve_fail;
626 		if (delayed_refs_bytes) {
627 			btrfs_migrate_to_delayed_refs_rsv(fs_info, rsv,
628 							  delayed_refs_bytes);
629 			num_bytes -= delayed_refs_bytes;
630 		}
631 
632 		if (rsv->space_info->force_alloc)
633 			do_chunk_alloc = true;
634 	} else if (num_items == 0 && flush == BTRFS_RESERVE_FLUSH_ALL &&
635 		   !btrfs_block_rsv_full(delayed_refs_rsv)) {
636 		/*
637 		 * Some people call with btrfs_start_transaction(root, 0)
638 		 * because they can be throttled, but have some other mechanism
639 		 * for reserving space.  We still want these guys to refill the
640 		 * delayed block_rsv so just add 1 items worth of reservation
641 		 * here.
642 		 */
643 		ret = btrfs_delayed_refs_rsv_refill(fs_info, flush);
644 		if (ret)
645 			goto reserve_fail;
646 	}
647 again:
648 	h = kmem_cache_zalloc(btrfs_trans_handle_cachep, GFP_NOFS);
649 	if (!h) {
650 		ret = -ENOMEM;
651 		goto alloc_fail;
652 	}
653 
654 	/*
655 	 * If we are JOIN_NOLOCK we're already committing a transaction and
656 	 * waiting on this guy, so we don't need to do the sb_start_intwrite
657 	 * because we're already holding a ref.  We need this because we could
658 	 * have raced in and did an fsync() on a file which can kick a commit
659 	 * and then we deadlock with somebody doing a freeze.
660 	 *
661 	 * If we are ATTACH, it means we just want to catch the current
662 	 * transaction and commit it, so we needn't do sb_start_intwrite().
663 	 */
664 	if (type & __TRANS_FREEZABLE)
665 		sb_start_intwrite(fs_info->sb);
666 
667 	if (may_wait_transaction(fs_info, type))
668 		wait_current_trans(fs_info);
669 
670 	do {
671 		ret = join_transaction(fs_info, type);
672 		if (ret == -EBUSY) {
673 			wait_current_trans(fs_info);
674 			if (unlikely(type == TRANS_ATTACH ||
675 				     type == TRANS_JOIN_NOSTART))
676 				ret = -ENOENT;
677 		}
678 	} while (ret == -EBUSY);
679 
680 	if (ret < 0)
681 		goto join_fail;
682 
683 	cur_trans = fs_info->running_transaction;
684 
685 	h->transid = cur_trans->transid;
686 	h->transaction = cur_trans;
687 	refcount_set(&h->use_count, 1);
688 	h->fs_info = root->fs_info;
689 
690 	h->type = type;
691 	INIT_LIST_HEAD(&h->new_bgs);
692 
693 	smp_mb();
694 	if (cur_trans->state >= TRANS_STATE_COMMIT_START &&
695 	    may_wait_transaction(fs_info, type)) {
696 		current->journal_info = h;
697 		btrfs_commit_transaction(h);
698 		goto again;
699 	}
700 
701 	if (num_bytes) {
702 		trace_btrfs_space_reservation(fs_info, "transaction",
703 					      h->transid, num_bytes, 1);
704 		h->block_rsv = &fs_info->trans_block_rsv;
705 		h->bytes_reserved = num_bytes;
706 		h->reloc_reserved = reloc_reserved;
707 	}
708 
709 got_it:
710 	if (!current->journal_info)
711 		current->journal_info = h;
712 
713 	/*
714 	 * If the space_info is marked ALLOC_FORCE then we'll get upgraded to
715 	 * ALLOC_FORCE the first run through, and then we won't allocate for
716 	 * anybody else who races in later.  We don't care about the return
717 	 * value here.
718 	 */
719 	if (do_chunk_alloc && num_bytes) {
720 		u64 flags = h->block_rsv->space_info->flags;
721 
722 		btrfs_chunk_alloc(h, btrfs_get_alloc_profile(fs_info, flags),
723 				  CHUNK_ALLOC_NO_FORCE);
724 	}
725 
726 	/*
727 	 * btrfs_record_root_in_trans() needs to alloc new extents, and may
728 	 * call btrfs_join_transaction() while we're also starting a
729 	 * transaction.
730 	 *
731 	 * Thus it need to be called after current->journal_info initialized,
732 	 * or we can deadlock.
733 	 */
734 	ret = btrfs_record_root_in_trans(h, root);
735 	if (ret) {
736 		/*
737 		 * The transaction handle is fully initialized and linked with
738 		 * other structures so it needs to be ended in case of errors,
739 		 * not just freed.
740 		 */
741 		btrfs_end_transaction(h);
742 		return ERR_PTR(ret);
743 	}
744 
745 	return h;
746 
747 join_fail:
748 	if (type & __TRANS_FREEZABLE)
749 		sb_end_intwrite(fs_info->sb);
750 	kmem_cache_free(btrfs_trans_handle_cachep, h);
751 alloc_fail:
752 	if (num_bytes)
753 		btrfs_block_rsv_release(fs_info, &fs_info->trans_block_rsv,
754 					num_bytes, NULL);
755 reserve_fail:
756 	btrfs_qgroup_free_meta_pertrans(root, qgroup_reserved);
757 	return ERR_PTR(ret);
758 }
759 
760 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
761 						   unsigned int num_items)
762 {
763 	return start_transaction(root, num_items, TRANS_START,
764 				 BTRFS_RESERVE_FLUSH_ALL, true);
765 }
766 
767 struct btrfs_trans_handle *btrfs_start_transaction_fallback_global_rsv(
768 					struct btrfs_root *root,
769 					unsigned int num_items)
770 {
771 	return start_transaction(root, num_items, TRANS_START,
772 				 BTRFS_RESERVE_FLUSH_ALL_STEAL, false);
773 }
774 
775 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
776 {
777 	return start_transaction(root, 0, TRANS_JOIN, BTRFS_RESERVE_NO_FLUSH,
778 				 true);
779 }
780 
781 struct btrfs_trans_handle *btrfs_join_transaction_spacecache(struct btrfs_root *root)
782 {
783 	return start_transaction(root, 0, TRANS_JOIN_NOLOCK,
784 				 BTRFS_RESERVE_NO_FLUSH, true);
785 }
786 
787 /*
788  * Similar to regular join but it never starts a transaction when none is
789  * running or after waiting for the current one to finish.
790  */
791 struct btrfs_trans_handle *btrfs_join_transaction_nostart(struct btrfs_root *root)
792 {
793 	return start_transaction(root, 0, TRANS_JOIN_NOSTART,
794 				 BTRFS_RESERVE_NO_FLUSH, true);
795 }
796 
797 /*
798  * btrfs_attach_transaction() - catch the running transaction
799  *
800  * It is used when we want to commit the current the transaction, but
801  * don't want to start a new one.
802  *
803  * Note: If this function return -ENOENT, it just means there is no
804  * running transaction. But it is possible that the inactive transaction
805  * is still in the memory, not fully on disk. If you hope there is no
806  * inactive transaction in the fs when -ENOENT is returned, you should
807  * invoke
808  *     btrfs_attach_transaction_barrier()
809  */
810 struct btrfs_trans_handle *btrfs_attach_transaction(struct btrfs_root *root)
811 {
812 	return start_transaction(root, 0, TRANS_ATTACH,
813 				 BTRFS_RESERVE_NO_FLUSH, true);
814 }
815 
816 /*
817  * btrfs_attach_transaction_barrier() - catch the running transaction
818  *
819  * It is similar to the above function, the difference is this one
820  * will wait for all the inactive transactions until they fully
821  * complete.
822  */
823 struct btrfs_trans_handle *
824 btrfs_attach_transaction_barrier(struct btrfs_root *root)
825 {
826 	struct btrfs_trans_handle *trans;
827 
828 	trans = start_transaction(root, 0, TRANS_ATTACH,
829 				  BTRFS_RESERVE_NO_FLUSH, true);
830 	if (trans == ERR_PTR(-ENOENT))
831 		btrfs_wait_for_commit(root->fs_info, 0);
832 
833 	return trans;
834 }
835 
836 /* Wait for a transaction commit to reach at least the given state. */
837 static noinline void wait_for_commit(struct btrfs_transaction *commit,
838 				     const enum btrfs_trans_state min_state)
839 {
840 	struct btrfs_fs_info *fs_info = commit->fs_info;
841 	u64 transid = commit->transid;
842 	bool put = false;
843 
844 	/*
845 	 * At the moment this function is called with min_state either being
846 	 * TRANS_STATE_COMPLETED or TRANS_STATE_SUPER_COMMITTED.
847 	 */
848 	if (min_state == TRANS_STATE_COMPLETED)
849 		btrfs_might_wait_for_state(fs_info, BTRFS_LOCKDEP_TRANS_COMPLETED);
850 	else
851 		btrfs_might_wait_for_state(fs_info, BTRFS_LOCKDEP_TRANS_SUPER_COMMITTED);
852 
853 	while (1) {
854 		wait_event(commit->commit_wait, commit->state >= min_state);
855 		if (put)
856 			btrfs_put_transaction(commit);
857 
858 		if (min_state < TRANS_STATE_COMPLETED)
859 			break;
860 
861 		/*
862 		 * A transaction isn't really completed until all of the
863 		 * previous transactions are completed, but with fsync we can
864 		 * end up with SUPER_COMMITTED transactions before a COMPLETED
865 		 * transaction. Wait for those.
866 		 */
867 
868 		spin_lock(&fs_info->trans_lock);
869 		commit = list_first_entry_or_null(&fs_info->trans_list,
870 						  struct btrfs_transaction,
871 						  list);
872 		if (!commit || commit->transid > transid) {
873 			spin_unlock(&fs_info->trans_lock);
874 			break;
875 		}
876 		refcount_inc(&commit->use_count);
877 		put = true;
878 		spin_unlock(&fs_info->trans_lock);
879 	}
880 }
881 
882 int btrfs_wait_for_commit(struct btrfs_fs_info *fs_info, u64 transid)
883 {
884 	struct btrfs_transaction *cur_trans = NULL, *t;
885 	int ret = 0;
886 
887 	if (transid) {
888 		if (transid <= fs_info->last_trans_committed)
889 			goto out;
890 
891 		/* find specified transaction */
892 		spin_lock(&fs_info->trans_lock);
893 		list_for_each_entry(t, &fs_info->trans_list, list) {
894 			if (t->transid == transid) {
895 				cur_trans = t;
896 				refcount_inc(&cur_trans->use_count);
897 				ret = 0;
898 				break;
899 			}
900 			if (t->transid > transid) {
901 				ret = 0;
902 				break;
903 			}
904 		}
905 		spin_unlock(&fs_info->trans_lock);
906 
907 		/*
908 		 * The specified transaction doesn't exist, or we
909 		 * raced with btrfs_commit_transaction
910 		 */
911 		if (!cur_trans) {
912 			if (transid > fs_info->last_trans_committed)
913 				ret = -EINVAL;
914 			goto out;
915 		}
916 	} else {
917 		/* find newest transaction that is committing | committed */
918 		spin_lock(&fs_info->trans_lock);
919 		list_for_each_entry_reverse(t, &fs_info->trans_list,
920 					    list) {
921 			if (t->state >= TRANS_STATE_COMMIT_START) {
922 				if (t->state == TRANS_STATE_COMPLETED)
923 					break;
924 				cur_trans = t;
925 				refcount_inc(&cur_trans->use_count);
926 				break;
927 			}
928 		}
929 		spin_unlock(&fs_info->trans_lock);
930 		if (!cur_trans)
931 			goto out;  /* nothing committing|committed */
932 	}
933 
934 	wait_for_commit(cur_trans, TRANS_STATE_COMPLETED);
935 	btrfs_put_transaction(cur_trans);
936 out:
937 	return ret;
938 }
939 
940 void btrfs_throttle(struct btrfs_fs_info *fs_info)
941 {
942 	wait_current_trans(fs_info);
943 }
944 
945 static bool should_end_transaction(struct btrfs_trans_handle *trans)
946 {
947 	struct btrfs_fs_info *fs_info = trans->fs_info;
948 
949 	if (btrfs_check_space_for_delayed_refs(fs_info))
950 		return true;
951 
952 	return !!btrfs_block_rsv_check(&fs_info->global_block_rsv, 50);
953 }
954 
955 bool btrfs_should_end_transaction(struct btrfs_trans_handle *trans)
956 {
957 	struct btrfs_transaction *cur_trans = trans->transaction;
958 
959 	if (cur_trans->state >= TRANS_STATE_COMMIT_START ||
960 	    test_bit(BTRFS_DELAYED_REFS_FLUSHING, &cur_trans->delayed_refs.flags))
961 		return true;
962 
963 	return should_end_transaction(trans);
964 }
965 
966 static void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans)
967 
968 {
969 	struct btrfs_fs_info *fs_info = trans->fs_info;
970 
971 	if (!trans->block_rsv) {
972 		ASSERT(!trans->bytes_reserved);
973 		return;
974 	}
975 
976 	if (!trans->bytes_reserved)
977 		return;
978 
979 	ASSERT(trans->block_rsv == &fs_info->trans_block_rsv);
980 	trace_btrfs_space_reservation(fs_info, "transaction",
981 				      trans->transid, trans->bytes_reserved, 0);
982 	btrfs_block_rsv_release(fs_info, trans->block_rsv,
983 				trans->bytes_reserved, NULL);
984 	trans->bytes_reserved = 0;
985 }
986 
987 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
988 				   int throttle)
989 {
990 	struct btrfs_fs_info *info = trans->fs_info;
991 	struct btrfs_transaction *cur_trans = trans->transaction;
992 	int err = 0;
993 
994 	if (refcount_read(&trans->use_count) > 1) {
995 		refcount_dec(&trans->use_count);
996 		trans->block_rsv = trans->orig_rsv;
997 		return 0;
998 	}
999 
1000 	btrfs_trans_release_metadata(trans);
1001 	trans->block_rsv = NULL;
1002 
1003 	btrfs_create_pending_block_groups(trans);
1004 
1005 	btrfs_trans_release_chunk_metadata(trans);
1006 
1007 	if (trans->type & __TRANS_FREEZABLE)
1008 		sb_end_intwrite(info->sb);
1009 
1010 	WARN_ON(cur_trans != info->running_transaction);
1011 	WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
1012 	atomic_dec(&cur_trans->num_writers);
1013 	extwriter_counter_dec(cur_trans, trans->type);
1014 
1015 	cond_wake_up(&cur_trans->writer_wait);
1016 
1017 	btrfs_lockdep_release(info, btrfs_trans_num_extwriters);
1018 	btrfs_lockdep_release(info, btrfs_trans_num_writers);
1019 
1020 	btrfs_put_transaction(cur_trans);
1021 
1022 	if (current->journal_info == trans)
1023 		current->journal_info = NULL;
1024 
1025 	if (throttle)
1026 		btrfs_run_delayed_iputs(info);
1027 
1028 	if (TRANS_ABORTED(trans) || BTRFS_FS_ERROR(info)) {
1029 		wake_up_process(info->transaction_kthread);
1030 		if (TRANS_ABORTED(trans))
1031 			err = trans->aborted;
1032 		else
1033 			err = -EROFS;
1034 	}
1035 
1036 	kmem_cache_free(btrfs_trans_handle_cachep, trans);
1037 	return err;
1038 }
1039 
1040 int btrfs_end_transaction(struct btrfs_trans_handle *trans)
1041 {
1042 	return __btrfs_end_transaction(trans, 0);
1043 }
1044 
1045 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans)
1046 {
1047 	return __btrfs_end_transaction(trans, 1);
1048 }
1049 
1050 /*
1051  * when btree blocks are allocated, they have some corresponding bits set for
1052  * them in one of two extent_io trees.  This is used to make sure all of
1053  * those extents are sent to disk but does not wait on them
1054  */
1055 int btrfs_write_marked_extents(struct btrfs_fs_info *fs_info,
1056 			       struct extent_io_tree *dirty_pages, int mark)
1057 {
1058 	int err = 0;
1059 	int werr = 0;
1060 	struct address_space *mapping = fs_info->btree_inode->i_mapping;
1061 	struct extent_state *cached_state = NULL;
1062 	u64 start = 0;
1063 	u64 end;
1064 
1065 	atomic_inc(&BTRFS_I(fs_info->btree_inode)->sync_writers);
1066 	while (!find_first_extent_bit(dirty_pages, start, &start, &end,
1067 				      mark, &cached_state)) {
1068 		bool wait_writeback = false;
1069 
1070 		err = convert_extent_bit(dirty_pages, start, end,
1071 					 EXTENT_NEED_WAIT,
1072 					 mark, &cached_state);
1073 		/*
1074 		 * convert_extent_bit can return -ENOMEM, which is most of the
1075 		 * time a temporary error. So when it happens, ignore the error
1076 		 * and wait for writeback of this range to finish - because we
1077 		 * failed to set the bit EXTENT_NEED_WAIT for the range, a call
1078 		 * to __btrfs_wait_marked_extents() would not know that
1079 		 * writeback for this range started and therefore wouldn't
1080 		 * wait for it to finish - we don't want to commit a
1081 		 * superblock that points to btree nodes/leafs for which
1082 		 * writeback hasn't finished yet (and without errors).
1083 		 * We cleanup any entries left in the io tree when committing
1084 		 * the transaction (through extent_io_tree_release()).
1085 		 */
1086 		if (err == -ENOMEM) {
1087 			err = 0;
1088 			wait_writeback = true;
1089 		}
1090 		if (!err)
1091 			err = filemap_fdatawrite_range(mapping, start, end);
1092 		if (err)
1093 			werr = err;
1094 		else if (wait_writeback)
1095 			werr = filemap_fdatawait_range(mapping, start, end);
1096 		free_extent_state(cached_state);
1097 		cached_state = NULL;
1098 		cond_resched();
1099 		start = end + 1;
1100 	}
1101 	atomic_dec(&BTRFS_I(fs_info->btree_inode)->sync_writers);
1102 	return werr;
1103 }
1104 
1105 /*
1106  * when btree blocks are allocated, they have some corresponding bits set for
1107  * them in one of two extent_io trees.  This is used to make sure all of
1108  * those extents are on disk for transaction or log commit.  We wait
1109  * on all the pages and clear them from the dirty pages state tree
1110  */
1111 static int __btrfs_wait_marked_extents(struct btrfs_fs_info *fs_info,
1112 				       struct extent_io_tree *dirty_pages)
1113 {
1114 	int err = 0;
1115 	int werr = 0;
1116 	struct address_space *mapping = fs_info->btree_inode->i_mapping;
1117 	struct extent_state *cached_state = NULL;
1118 	u64 start = 0;
1119 	u64 end;
1120 
1121 	while (!find_first_extent_bit(dirty_pages, start, &start, &end,
1122 				      EXTENT_NEED_WAIT, &cached_state)) {
1123 		/*
1124 		 * Ignore -ENOMEM errors returned by clear_extent_bit().
1125 		 * When committing the transaction, we'll remove any entries
1126 		 * left in the io tree. For a log commit, we don't remove them
1127 		 * after committing the log because the tree can be accessed
1128 		 * concurrently - we do it only at transaction commit time when
1129 		 * it's safe to do it (through extent_io_tree_release()).
1130 		 */
1131 		err = clear_extent_bit(dirty_pages, start, end,
1132 				       EXTENT_NEED_WAIT, &cached_state);
1133 		if (err == -ENOMEM)
1134 			err = 0;
1135 		if (!err)
1136 			err = filemap_fdatawait_range(mapping, start, end);
1137 		if (err)
1138 			werr = err;
1139 		free_extent_state(cached_state);
1140 		cached_state = NULL;
1141 		cond_resched();
1142 		start = end + 1;
1143 	}
1144 	if (err)
1145 		werr = err;
1146 	return werr;
1147 }
1148 
1149 static int btrfs_wait_extents(struct btrfs_fs_info *fs_info,
1150 		       struct extent_io_tree *dirty_pages)
1151 {
1152 	bool errors = false;
1153 	int err;
1154 
1155 	err = __btrfs_wait_marked_extents(fs_info, dirty_pages);
1156 	if (test_and_clear_bit(BTRFS_FS_BTREE_ERR, &fs_info->flags))
1157 		errors = true;
1158 
1159 	if (errors && !err)
1160 		err = -EIO;
1161 	return err;
1162 }
1163 
1164 int btrfs_wait_tree_log_extents(struct btrfs_root *log_root, int mark)
1165 {
1166 	struct btrfs_fs_info *fs_info = log_root->fs_info;
1167 	struct extent_io_tree *dirty_pages = &log_root->dirty_log_pages;
1168 	bool errors = false;
1169 	int err;
1170 
1171 	ASSERT(log_root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
1172 
1173 	err = __btrfs_wait_marked_extents(fs_info, dirty_pages);
1174 	if ((mark & EXTENT_DIRTY) &&
1175 	    test_and_clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags))
1176 		errors = true;
1177 
1178 	if ((mark & EXTENT_NEW) &&
1179 	    test_and_clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags))
1180 		errors = true;
1181 
1182 	if (errors && !err)
1183 		err = -EIO;
1184 	return err;
1185 }
1186 
1187 /*
1188  * When btree blocks are allocated the corresponding extents are marked dirty.
1189  * This function ensures such extents are persisted on disk for transaction or
1190  * log commit.
1191  *
1192  * @trans: transaction whose dirty pages we'd like to write
1193  */
1194 static int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans)
1195 {
1196 	int ret;
1197 	int ret2;
1198 	struct extent_io_tree *dirty_pages = &trans->transaction->dirty_pages;
1199 	struct btrfs_fs_info *fs_info = trans->fs_info;
1200 	struct blk_plug plug;
1201 
1202 	blk_start_plug(&plug);
1203 	ret = btrfs_write_marked_extents(fs_info, dirty_pages, EXTENT_DIRTY);
1204 	blk_finish_plug(&plug);
1205 	ret2 = btrfs_wait_extents(fs_info, dirty_pages);
1206 
1207 	extent_io_tree_release(&trans->transaction->dirty_pages);
1208 
1209 	if (ret)
1210 		return ret;
1211 	else if (ret2)
1212 		return ret2;
1213 	else
1214 		return 0;
1215 }
1216 
1217 /*
1218  * this is used to update the root pointer in the tree of tree roots.
1219  *
1220  * But, in the case of the extent allocation tree, updating the root
1221  * pointer may allocate blocks which may change the root of the extent
1222  * allocation tree.
1223  *
1224  * So, this loops and repeats and makes sure the cowonly root didn't
1225  * change while the root pointer was being updated in the metadata.
1226  */
1227 static int update_cowonly_root(struct btrfs_trans_handle *trans,
1228 			       struct btrfs_root *root)
1229 {
1230 	int ret;
1231 	u64 old_root_bytenr;
1232 	u64 old_root_used;
1233 	struct btrfs_fs_info *fs_info = root->fs_info;
1234 	struct btrfs_root *tree_root = fs_info->tree_root;
1235 
1236 	old_root_used = btrfs_root_used(&root->root_item);
1237 
1238 	while (1) {
1239 		old_root_bytenr = btrfs_root_bytenr(&root->root_item);
1240 		if (old_root_bytenr == root->node->start &&
1241 		    old_root_used == btrfs_root_used(&root->root_item))
1242 			break;
1243 
1244 		btrfs_set_root_node(&root->root_item, root->node);
1245 		ret = btrfs_update_root(trans, tree_root,
1246 					&root->root_key,
1247 					&root->root_item);
1248 		if (ret)
1249 			return ret;
1250 
1251 		old_root_used = btrfs_root_used(&root->root_item);
1252 	}
1253 
1254 	return 0;
1255 }
1256 
1257 /*
1258  * update all the cowonly tree roots on disk
1259  *
1260  * The error handling in this function may not be obvious. Any of the
1261  * failures will cause the file system to go offline. We still need
1262  * to clean up the delayed refs.
1263  */
1264 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans)
1265 {
1266 	struct btrfs_fs_info *fs_info = trans->fs_info;
1267 	struct list_head *dirty_bgs = &trans->transaction->dirty_bgs;
1268 	struct list_head *io_bgs = &trans->transaction->io_bgs;
1269 	struct list_head *next;
1270 	struct extent_buffer *eb;
1271 	int ret;
1272 
1273 	/*
1274 	 * At this point no one can be using this transaction to modify any tree
1275 	 * and no one can start another transaction to modify any tree either.
1276 	 */
1277 	ASSERT(trans->transaction->state == TRANS_STATE_COMMIT_DOING);
1278 
1279 	eb = btrfs_lock_root_node(fs_info->tree_root);
1280 	ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL,
1281 			      0, &eb, BTRFS_NESTING_COW);
1282 	btrfs_tree_unlock(eb);
1283 	free_extent_buffer(eb);
1284 
1285 	if (ret)
1286 		return ret;
1287 
1288 	ret = btrfs_run_dev_stats(trans);
1289 	if (ret)
1290 		return ret;
1291 	ret = btrfs_run_dev_replace(trans);
1292 	if (ret)
1293 		return ret;
1294 	ret = btrfs_run_qgroups(trans);
1295 	if (ret)
1296 		return ret;
1297 
1298 	ret = btrfs_setup_space_cache(trans);
1299 	if (ret)
1300 		return ret;
1301 
1302 again:
1303 	while (!list_empty(&fs_info->dirty_cowonly_roots)) {
1304 		struct btrfs_root *root;
1305 		next = fs_info->dirty_cowonly_roots.next;
1306 		list_del_init(next);
1307 		root = list_entry(next, struct btrfs_root, dirty_list);
1308 		clear_bit(BTRFS_ROOT_DIRTY, &root->state);
1309 
1310 		list_add_tail(&root->dirty_list,
1311 			      &trans->transaction->switch_commits);
1312 		ret = update_cowonly_root(trans, root);
1313 		if (ret)
1314 			return ret;
1315 	}
1316 
1317 	/* Now flush any delayed refs generated by updating all of the roots */
1318 	ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
1319 	if (ret)
1320 		return ret;
1321 
1322 	while (!list_empty(dirty_bgs) || !list_empty(io_bgs)) {
1323 		ret = btrfs_write_dirty_block_groups(trans);
1324 		if (ret)
1325 			return ret;
1326 
1327 		/*
1328 		 * We're writing the dirty block groups, which could generate
1329 		 * delayed refs, which could generate more dirty block groups,
1330 		 * so we want to keep this flushing in this loop to make sure
1331 		 * everything gets run.
1332 		 */
1333 		ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
1334 		if (ret)
1335 			return ret;
1336 	}
1337 
1338 	if (!list_empty(&fs_info->dirty_cowonly_roots))
1339 		goto again;
1340 
1341 	/* Update dev-replace pointer once everything is committed */
1342 	fs_info->dev_replace.committed_cursor_left =
1343 		fs_info->dev_replace.cursor_left_last_write_of_item;
1344 
1345 	return 0;
1346 }
1347 
1348 /*
1349  * If we had a pending drop we need to see if there are any others left in our
1350  * dead roots list, and if not clear our bit and wake any waiters.
1351  */
1352 void btrfs_maybe_wake_unfinished_drop(struct btrfs_fs_info *fs_info)
1353 {
1354 	/*
1355 	 * We put the drop in progress roots at the front of the list, so if the
1356 	 * first entry doesn't have UNFINISHED_DROP set we can wake everybody
1357 	 * up.
1358 	 */
1359 	spin_lock(&fs_info->trans_lock);
1360 	if (!list_empty(&fs_info->dead_roots)) {
1361 		struct btrfs_root *root = list_first_entry(&fs_info->dead_roots,
1362 							   struct btrfs_root,
1363 							   root_list);
1364 		if (test_bit(BTRFS_ROOT_UNFINISHED_DROP, &root->state)) {
1365 			spin_unlock(&fs_info->trans_lock);
1366 			return;
1367 		}
1368 	}
1369 	spin_unlock(&fs_info->trans_lock);
1370 
1371 	btrfs_wake_unfinished_drop(fs_info);
1372 }
1373 
1374 /*
1375  * dead roots are old snapshots that need to be deleted.  This allocates
1376  * a dirty root struct and adds it into the list of dead roots that need to
1377  * be deleted
1378  */
1379 void btrfs_add_dead_root(struct btrfs_root *root)
1380 {
1381 	struct btrfs_fs_info *fs_info = root->fs_info;
1382 
1383 	spin_lock(&fs_info->trans_lock);
1384 	if (list_empty(&root->root_list)) {
1385 		btrfs_grab_root(root);
1386 
1387 		/* We want to process the partially complete drops first. */
1388 		if (test_bit(BTRFS_ROOT_UNFINISHED_DROP, &root->state))
1389 			list_add(&root->root_list, &fs_info->dead_roots);
1390 		else
1391 			list_add_tail(&root->root_list, &fs_info->dead_roots);
1392 	}
1393 	spin_unlock(&fs_info->trans_lock);
1394 }
1395 
1396 /*
1397  * Update each subvolume root and its relocation root, if it exists, in the tree
1398  * of tree roots. Also free log roots if they exist.
1399  */
1400 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans)
1401 {
1402 	struct btrfs_fs_info *fs_info = trans->fs_info;
1403 	struct btrfs_root *gang[8];
1404 	int i;
1405 	int ret;
1406 
1407 	/*
1408 	 * At this point no one can be using this transaction to modify any tree
1409 	 * and no one can start another transaction to modify any tree either.
1410 	 */
1411 	ASSERT(trans->transaction->state == TRANS_STATE_COMMIT_DOING);
1412 
1413 	spin_lock(&fs_info->fs_roots_radix_lock);
1414 	while (1) {
1415 		ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
1416 						 (void **)gang, 0,
1417 						 ARRAY_SIZE(gang),
1418 						 BTRFS_ROOT_TRANS_TAG);
1419 		if (ret == 0)
1420 			break;
1421 		for (i = 0; i < ret; i++) {
1422 			struct btrfs_root *root = gang[i];
1423 			int ret2;
1424 
1425 			/*
1426 			 * At this point we can neither have tasks logging inodes
1427 			 * from a root nor trying to commit a log tree.
1428 			 */
1429 			ASSERT(atomic_read(&root->log_writers) == 0);
1430 			ASSERT(atomic_read(&root->log_commit[0]) == 0);
1431 			ASSERT(atomic_read(&root->log_commit[1]) == 0);
1432 
1433 			radix_tree_tag_clear(&fs_info->fs_roots_radix,
1434 					(unsigned long)root->root_key.objectid,
1435 					BTRFS_ROOT_TRANS_TAG);
1436 			spin_unlock(&fs_info->fs_roots_radix_lock);
1437 
1438 			btrfs_free_log(trans, root);
1439 			ret2 = btrfs_update_reloc_root(trans, root);
1440 			if (ret2)
1441 				return ret2;
1442 
1443 			/* see comments in should_cow_block() */
1444 			clear_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1445 			smp_mb__after_atomic();
1446 
1447 			if (root->commit_root != root->node) {
1448 				list_add_tail(&root->dirty_list,
1449 					&trans->transaction->switch_commits);
1450 				btrfs_set_root_node(&root->root_item,
1451 						    root->node);
1452 			}
1453 
1454 			ret2 = btrfs_update_root(trans, fs_info->tree_root,
1455 						&root->root_key,
1456 						&root->root_item);
1457 			if (ret2)
1458 				return ret2;
1459 			spin_lock(&fs_info->fs_roots_radix_lock);
1460 			btrfs_qgroup_free_meta_all_pertrans(root);
1461 		}
1462 	}
1463 	spin_unlock(&fs_info->fs_roots_radix_lock);
1464 	return 0;
1465 }
1466 
1467 /*
1468  * defrag a given btree.
1469  * Every leaf in the btree is read and defragged.
1470  */
1471 int btrfs_defrag_root(struct btrfs_root *root)
1472 {
1473 	struct btrfs_fs_info *info = root->fs_info;
1474 	struct btrfs_trans_handle *trans;
1475 	int ret;
1476 
1477 	if (test_and_set_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state))
1478 		return 0;
1479 
1480 	while (1) {
1481 		trans = btrfs_start_transaction(root, 0);
1482 		if (IS_ERR(trans)) {
1483 			ret = PTR_ERR(trans);
1484 			break;
1485 		}
1486 
1487 		ret = btrfs_defrag_leaves(trans, root);
1488 
1489 		btrfs_end_transaction(trans);
1490 		btrfs_btree_balance_dirty(info);
1491 		cond_resched();
1492 
1493 		if (btrfs_fs_closing(info) || ret != -EAGAIN)
1494 			break;
1495 
1496 		if (btrfs_defrag_cancelled(info)) {
1497 			btrfs_debug(info, "defrag_root cancelled");
1498 			ret = -EAGAIN;
1499 			break;
1500 		}
1501 	}
1502 	clear_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state);
1503 	return ret;
1504 }
1505 
1506 /*
1507  * Do all special snapshot related qgroup dirty hack.
1508  *
1509  * Will do all needed qgroup inherit and dirty hack like switch commit
1510  * roots inside one transaction and write all btree into disk, to make
1511  * qgroup works.
1512  */
1513 static int qgroup_account_snapshot(struct btrfs_trans_handle *trans,
1514 				   struct btrfs_root *src,
1515 				   struct btrfs_root *parent,
1516 				   struct btrfs_qgroup_inherit *inherit,
1517 				   u64 dst_objectid)
1518 {
1519 	struct btrfs_fs_info *fs_info = src->fs_info;
1520 	int ret;
1521 
1522 	/*
1523 	 * Save some performance in the case that qgroups are not
1524 	 * enabled. If this check races with the ioctl, rescan will
1525 	 * kick in anyway.
1526 	 */
1527 	if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
1528 		return 0;
1529 
1530 	/*
1531 	 * Ensure dirty @src will be committed.  Or, after coming
1532 	 * commit_fs_roots() and switch_commit_roots(), any dirty but not
1533 	 * recorded root will never be updated again, causing an outdated root
1534 	 * item.
1535 	 */
1536 	ret = record_root_in_trans(trans, src, 1);
1537 	if (ret)
1538 		return ret;
1539 
1540 	/*
1541 	 * btrfs_qgroup_inherit relies on a consistent view of the usage for the
1542 	 * src root, so we must run the delayed refs here.
1543 	 *
1544 	 * However this isn't particularly fool proof, because there's no
1545 	 * synchronization keeping us from changing the tree after this point
1546 	 * before we do the qgroup_inherit, or even from making changes while
1547 	 * we're doing the qgroup_inherit.  But that's a problem for the future,
1548 	 * for now flush the delayed refs to narrow the race window where the
1549 	 * qgroup counters could end up wrong.
1550 	 */
1551 	ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
1552 	if (ret) {
1553 		btrfs_abort_transaction(trans, ret);
1554 		return ret;
1555 	}
1556 
1557 	ret = commit_fs_roots(trans);
1558 	if (ret)
1559 		goto out;
1560 	ret = btrfs_qgroup_account_extents(trans);
1561 	if (ret < 0)
1562 		goto out;
1563 
1564 	/* Now qgroup are all updated, we can inherit it to new qgroups */
1565 	ret = btrfs_qgroup_inherit(trans, src->root_key.objectid, dst_objectid,
1566 				   inherit);
1567 	if (ret < 0)
1568 		goto out;
1569 
1570 	/*
1571 	 * Now we do a simplified commit transaction, which will:
1572 	 * 1) commit all subvolume and extent tree
1573 	 *    To ensure all subvolume and extent tree have a valid
1574 	 *    commit_root to accounting later insert_dir_item()
1575 	 * 2) write all btree blocks onto disk
1576 	 *    This is to make sure later btree modification will be cowed
1577 	 *    Or commit_root can be populated and cause wrong qgroup numbers
1578 	 * In this simplified commit, we don't really care about other trees
1579 	 * like chunk and root tree, as they won't affect qgroup.
1580 	 * And we don't write super to avoid half committed status.
1581 	 */
1582 	ret = commit_cowonly_roots(trans);
1583 	if (ret)
1584 		goto out;
1585 	switch_commit_roots(trans);
1586 	ret = btrfs_write_and_wait_transaction(trans);
1587 	if (ret)
1588 		btrfs_handle_fs_error(fs_info, ret,
1589 			"Error while writing out transaction for qgroup");
1590 
1591 out:
1592 	/*
1593 	 * Force parent root to be updated, as we recorded it before so its
1594 	 * last_trans == cur_transid.
1595 	 * Or it won't be committed again onto disk after later
1596 	 * insert_dir_item()
1597 	 */
1598 	if (!ret)
1599 		ret = record_root_in_trans(trans, parent, 1);
1600 	return ret;
1601 }
1602 
1603 /*
1604  * new snapshots need to be created at a very specific time in the
1605  * transaction commit.  This does the actual creation.
1606  *
1607  * Note:
1608  * If the error which may affect the commitment of the current transaction
1609  * happens, we should return the error number. If the error which just affect
1610  * the creation of the pending snapshots, just return 0.
1611  */
1612 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
1613 				   struct btrfs_pending_snapshot *pending)
1614 {
1615 
1616 	struct btrfs_fs_info *fs_info = trans->fs_info;
1617 	struct btrfs_key key;
1618 	struct btrfs_root_item *new_root_item;
1619 	struct btrfs_root *tree_root = fs_info->tree_root;
1620 	struct btrfs_root *root = pending->root;
1621 	struct btrfs_root *parent_root;
1622 	struct btrfs_block_rsv *rsv;
1623 	struct inode *parent_inode = pending->dir;
1624 	struct btrfs_path *path;
1625 	struct btrfs_dir_item *dir_item;
1626 	struct extent_buffer *tmp;
1627 	struct extent_buffer *old;
1628 	struct timespec64 cur_time;
1629 	int ret = 0;
1630 	u64 to_reserve = 0;
1631 	u64 index = 0;
1632 	u64 objectid;
1633 	u64 root_flags;
1634 	unsigned int nofs_flags;
1635 	struct fscrypt_name fname;
1636 
1637 	ASSERT(pending->path);
1638 	path = pending->path;
1639 
1640 	ASSERT(pending->root_item);
1641 	new_root_item = pending->root_item;
1642 
1643 	/*
1644 	 * We're inside a transaction and must make sure that any potential
1645 	 * allocations with GFP_KERNEL in fscrypt won't recurse back to
1646 	 * filesystem.
1647 	 */
1648 	nofs_flags = memalloc_nofs_save();
1649 	pending->error = fscrypt_setup_filename(parent_inode,
1650 						&pending->dentry->d_name, 0,
1651 						&fname);
1652 	memalloc_nofs_restore(nofs_flags);
1653 	if (pending->error)
1654 		goto free_pending;
1655 
1656 	pending->error = btrfs_get_free_objectid(tree_root, &objectid);
1657 	if (pending->error)
1658 		goto free_fname;
1659 
1660 	/*
1661 	 * Make qgroup to skip current new snapshot's qgroupid, as it is
1662 	 * accounted by later btrfs_qgroup_inherit().
1663 	 */
1664 	btrfs_set_skip_qgroup(trans, objectid);
1665 
1666 	btrfs_reloc_pre_snapshot(pending, &to_reserve);
1667 
1668 	if (to_reserve > 0) {
1669 		pending->error = btrfs_block_rsv_add(fs_info,
1670 						     &pending->block_rsv,
1671 						     to_reserve,
1672 						     BTRFS_RESERVE_NO_FLUSH);
1673 		if (pending->error)
1674 			goto clear_skip_qgroup;
1675 	}
1676 
1677 	key.objectid = objectid;
1678 	key.offset = (u64)-1;
1679 	key.type = BTRFS_ROOT_ITEM_KEY;
1680 
1681 	rsv = trans->block_rsv;
1682 	trans->block_rsv = &pending->block_rsv;
1683 	trans->bytes_reserved = trans->block_rsv->reserved;
1684 	trace_btrfs_space_reservation(fs_info, "transaction",
1685 				      trans->transid,
1686 				      trans->bytes_reserved, 1);
1687 	parent_root = BTRFS_I(parent_inode)->root;
1688 	ret = record_root_in_trans(trans, parent_root, 0);
1689 	if (ret)
1690 		goto fail;
1691 	cur_time = current_time(parent_inode);
1692 
1693 	/*
1694 	 * insert the directory item
1695 	 */
1696 	ret = btrfs_set_inode_index(BTRFS_I(parent_inode), &index);
1697 	BUG_ON(ret); /* -ENOMEM */
1698 
1699 	/* check if there is a file/dir which has the same name. */
1700 	dir_item = btrfs_lookup_dir_item(NULL, parent_root, path,
1701 					 btrfs_ino(BTRFS_I(parent_inode)),
1702 					 &fname.disk_name, 0);
1703 	if (dir_item != NULL && !IS_ERR(dir_item)) {
1704 		pending->error = -EEXIST;
1705 		goto dir_item_existed;
1706 	} else if (IS_ERR(dir_item)) {
1707 		ret = PTR_ERR(dir_item);
1708 		btrfs_abort_transaction(trans, ret);
1709 		goto fail;
1710 	}
1711 	btrfs_release_path(path);
1712 
1713 	/*
1714 	 * pull in the delayed directory update
1715 	 * and the delayed inode item
1716 	 * otherwise we corrupt the FS during
1717 	 * snapshot
1718 	 */
1719 	ret = btrfs_run_delayed_items(trans);
1720 	if (ret) {	/* Transaction aborted */
1721 		btrfs_abort_transaction(trans, ret);
1722 		goto fail;
1723 	}
1724 
1725 	ret = record_root_in_trans(trans, root, 0);
1726 	if (ret) {
1727 		btrfs_abort_transaction(trans, ret);
1728 		goto fail;
1729 	}
1730 	btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
1731 	memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
1732 	btrfs_check_and_init_root_item(new_root_item);
1733 
1734 	root_flags = btrfs_root_flags(new_root_item);
1735 	if (pending->readonly)
1736 		root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
1737 	else
1738 		root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
1739 	btrfs_set_root_flags(new_root_item, root_flags);
1740 
1741 	btrfs_set_root_generation_v2(new_root_item,
1742 			trans->transid);
1743 	generate_random_guid(new_root_item->uuid);
1744 	memcpy(new_root_item->parent_uuid, root->root_item.uuid,
1745 			BTRFS_UUID_SIZE);
1746 	if (!(root_flags & BTRFS_ROOT_SUBVOL_RDONLY)) {
1747 		memset(new_root_item->received_uuid, 0,
1748 		       sizeof(new_root_item->received_uuid));
1749 		memset(&new_root_item->stime, 0, sizeof(new_root_item->stime));
1750 		memset(&new_root_item->rtime, 0, sizeof(new_root_item->rtime));
1751 		btrfs_set_root_stransid(new_root_item, 0);
1752 		btrfs_set_root_rtransid(new_root_item, 0);
1753 	}
1754 	btrfs_set_stack_timespec_sec(&new_root_item->otime, cur_time.tv_sec);
1755 	btrfs_set_stack_timespec_nsec(&new_root_item->otime, cur_time.tv_nsec);
1756 	btrfs_set_root_otransid(new_root_item, trans->transid);
1757 
1758 	old = btrfs_lock_root_node(root);
1759 	ret = btrfs_cow_block(trans, root, old, NULL, 0, &old,
1760 			      BTRFS_NESTING_COW);
1761 	if (ret) {
1762 		btrfs_tree_unlock(old);
1763 		free_extent_buffer(old);
1764 		btrfs_abort_transaction(trans, ret);
1765 		goto fail;
1766 	}
1767 
1768 	ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1769 	/* clean up in any case */
1770 	btrfs_tree_unlock(old);
1771 	free_extent_buffer(old);
1772 	if (ret) {
1773 		btrfs_abort_transaction(trans, ret);
1774 		goto fail;
1775 	}
1776 	/* see comments in should_cow_block() */
1777 	set_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1778 	smp_wmb();
1779 
1780 	btrfs_set_root_node(new_root_item, tmp);
1781 	/* record when the snapshot was created in key.offset */
1782 	key.offset = trans->transid;
1783 	ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1784 	btrfs_tree_unlock(tmp);
1785 	free_extent_buffer(tmp);
1786 	if (ret) {
1787 		btrfs_abort_transaction(trans, ret);
1788 		goto fail;
1789 	}
1790 
1791 	/*
1792 	 * insert root back/forward references
1793 	 */
1794 	ret = btrfs_add_root_ref(trans, objectid,
1795 				 parent_root->root_key.objectid,
1796 				 btrfs_ino(BTRFS_I(parent_inode)), index,
1797 				 &fname.disk_name);
1798 	if (ret) {
1799 		btrfs_abort_transaction(trans, ret);
1800 		goto fail;
1801 	}
1802 
1803 	key.offset = (u64)-1;
1804 	pending->snap = btrfs_get_new_fs_root(fs_info, objectid, pending->anon_dev);
1805 	if (IS_ERR(pending->snap)) {
1806 		ret = PTR_ERR(pending->snap);
1807 		pending->snap = NULL;
1808 		btrfs_abort_transaction(trans, ret);
1809 		goto fail;
1810 	}
1811 
1812 	ret = btrfs_reloc_post_snapshot(trans, pending);
1813 	if (ret) {
1814 		btrfs_abort_transaction(trans, ret);
1815 		goto fail;
1816 	}
1817 
1818 	/*
1819 	 * Do special qgroup accounting for snapshot, as we do some qgroup
1820 	 * snapshot hack to do fast snapshot.
1821 	 * To co-operate with that hack, we do hack again.
1822 	 * Or snapshot will be greatly slowed down by a subtree qgroup rescan
1823 	 */
1824 	ret = qgroup_account_snapshot(trans, root, parent_root,
1825 				      pending->inherit, objectid);
1826 	if (ret < 0)
1827 		goto fail;
1828 
1829 	ret = btrfs_insert_dir_item(trans, &fname.disk_name,
1830 				    BTRFS_I(parent_inode), &key, BTRFS_FT_DIR,
1831 				    index);
1832 	/* We have check then name at the beginning, so it is impossible. */
1833 	BUG_ON(ret == -EEXIST || ret == -EOVERFLOW);
1834 	if (ret) {
1835 		btrfs_abort_transaction(trans, ret);
1836 		goto fail;
1837 	}
1838 
1839 	btrfs_i_size_write(BTRFS_I(parent_inode), parent_inode->i_size +
1840 						  fname.disk_name.len * 2);
1841 	parent_inode->i_mtime = current_time(parent_inode);
1842 	parent_inode->i_ctime = parent_inode->i_mtime;
1843 	ret = btrfs_update_inode_fallback(trans, parent_root, BTRFS_I(parent_inode));
1844 	if (ret) {
1845 		btrfs_abort_transaction(trans, ret);
1846 		goto fail;
1847 	}
1848 	ret = btrfs_uuid_tree_add(trans, new_root_item->uuid,
1849 				  BTRFS_UUID_KEY_SUBVOL,
1850 				  objectid);
1851 	if (ret) {
1852 		btrfs_abort_transaction(trans, ret);
1853 		goto fail;
1854 	}
1855 	if (!btrfs_is_empty_uuid(new_root_item->received_uuid)) {
1856 		ret = btrfs_uuid_tree_add(trans, new_root_item->received_uuid,
1857 					  BTRFS_UUID_KEY_RECEIVED_SUBVOL,
1858 					  objectid);
1859 		if (ret && ret != -EEXIST) {
1860 			btrfs_abort_transaction(trans, ret);
1861 			goto fail;
1862 		}
1863 	}
1864 
1865 fail:
1866 	pending->error = ret;
1867 dir_item_existed:
1868 	trans->block_rsv = rsv;
1869 	trans->bytes_reserved = 0;
1870 clear_skip_qgroup:
1871 	btrfs_clear_skip_qgroup(trans);
1872 free_fname:
1873 	fscrypt_free_filename(&fname);
1874 free_pending:
1875 	kfree(new_root_item);
1876 	pending->root_item = NULL;
1877 	btrfs_free_path(path);
1878 	pending->path = NULL;
1879 
1880 	return ret;
1881 }
1882 
1883 /*
1884  * create all the snapshots we've scheduled for creation
1885  */
1886 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans)
1887 {
1888 	struct btrfs_pending_snapshot *pending, *next;
1889 	struct list_head *head = &trans->transaction->pending_snapshots;
1890 	int ret = 0;
1891 
1892 	list_for_each_entry_safe(pending, next, head, list) {
1893 		list_del(&pending->list);
1894 		ret = create_pending_snapshot(trans, pending);
1895 		if (ret)
1896 			break;
1897 	}
1898 	return ret;
1899 }
1900 
1901 static void update_super_roots(struct btrfs_fs_info *fs_info)
1902 {
1903 	struct btrfs_root_item *root_item;
1904 	struct btrfs_super_block *super;
1905 
1906 	super = fs_info->super_copy;
1907 
1908 	root_item = &fs_info->chunk_root->root_item;
1909 	super->chunk_root = root_item->bytenr;
1910 	super->chunk_root_generation = root_item->generation;
1911 	super->chunk_root_level = root_item->level;
1912 
1913 	root_item = &fs_info->tree_root->root_item;
1914 	super->root = root_item->bytenr;
1915 	super->generation = root_item->generation;
1916 	super->root_level = root_item->level;
1917 	if (btrfs_test_opt(fs_info, SPACE_CACHE))
1918 		super->cache_generation = root_item->generation;
1919 	else if (test_bit(BTRFS_FS_CLEANUP_SPACE_CACHE_V1, &fs_info->flags))
1920 		super->cache_generation = 0;
1921 	if (test_bit(BTRFS_FS_UPDATE_UUID_TREE_GEN, &fs_info->flags))
1922 		super->uuid_tree_generation = root_item->generation;
1923 }
1924 
1925 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1926 {
1927 	struct btrfs_transaction *trans;
1928 	int ret = 0;
1929 
1930 	spin_lock(&info->trans_lock);
1931 	trans = info->running_transaction;
1932 	if (trans)
1933 		ret = (trans->state >= TRANS_STATE_COMMIT_START);
1934 	spin_unlock(&info->trans_lock);
1935 	return ret;
1936 }
1937 
1938 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1939 {
1940 	struct btrfs_transaction *trans;
1941 	int ret = 0;
1942 
1943 	spin_lock(&info->trans_lock);
1944 	trans = info->running_transaction;
1945 	if (trans)
1946 		ret = is_transaction_blocked(trans);
1947 	spin_unlock(&info->trans_lock);
1948 	return ret;
1949 }
1950 
1951 void btrfs_commit_transaction_async(struct btrfs_trans_handle *trans)
1952 {
1953 	struct btrfs_fs_info *fs_info = trans->fs_info;
1954 	struct btrfs_transaction *cur_trans;
1955 
1956 	/* Kick the transaction kthread. */
1957 	set_bit(BTRFS_FS_COMMIT_TRANS, &fs_info->flags);
1958 	wake_up_process(fs_info->transaction_kthread);
1959 
1960 	/* take transaction reference */
1961 	cur_trans = trans->transaction;
1962 	refcount_inc(&cur_trans->use_count);
1963 
1964 	btrfs_end_transaction(trans);
1965 
1966 	/*
1967 	 * Wait for the current transaction commit to start and block
1968 	 * subsequent transaction joins
1969 	 */
1970 	btrfs_might_wait_for_state(fs_info, BTRFS_LOCKDEP_TRANS_COMMIT_START);
1971 	wait_event(fs_info->transaction_blocked_wait,
1972 		   cur_trans->state >= TRANS_STATE_COMMIT_START ||
1973 		   TRANS_ABORTED(cur_trans));
1974 	btrfs_put_transaction(cur_trans);
1975 }
1976 
1977 static void cleanup_transaction(struct btrfs_trans_handle *trans, int err)
1978 {
1979 	struct btrfs_fs_info *fs_info = trans->fs_info;
1980 	struct btrfs_transaction *cur_trans = trans->transaction;
1981 
1982 	WARN_ON(refcount_read(&trans->use_count) > 1);
1983 
1984 	btrfs_abort_transaction(trans, err);
1985 
1986 	spin_lock(&fs_info->trans_lock);
1987 
1988 	/*
1989 	 * If the transaction is removed from the list, it means this
1990 	 * transaction has been committed successfully, so it is impossible
1991 	 * to call the cleanup function.
1992 	 */
1993 	BUG_ON(list_empty(&cur_trans->list));
1994 
1995 	if (cur_trans == fs_info->running_transaction) {
1996 		cur_trans->state = TRANS_STATE_COMMIT_DOING;
1997 		spin_unlock(&fs_info->trans_lock);
1998 
1999 		/*
2000 		 * The thread has already released the lockdep map as reader
2001 		 * already in btrfs_commit_transaction().
2002 		 */
2003 		btrfs_might_wait_for_event(fs_info, btrfs_trans_num_writers);
2004 		wait_event(cur_trans->writer_wait,
2005 			   atomic_read(&cur_trans->num_writers) == 1);
2006 
2007 		spin_lock(&fs_info->trans_lock);
2008 	}
2009 
2010 	/*
2011 	 * Now that we know no one else is still using the transaction we can
2012 	 * remove the transaction from the list of transactions. This avoids
2013 	 * the transaction kthread from cleaning up the transaction while some
2014 	 * other task is still using it, which could result in a use-after-free
2015 	 * on things like log trees, as it forces the transaction kthread to
2016 	 * wait for this transaction to be cleaned up by us.
2017 	 */
2018 	list_del_init(&cur_trans->list);
2019 
2020 	spin_unlock(&fs_info->trans_lock);
2021 
2022 	btrfs_cleanup_one_transaction(trans->transaction, fs_info);
2023 
2024 	spin_lock(&fs_info->trans_lock);
2025 	if (cur_trans == fs_info->running_transaction)
2026 		fs_info->running_transaction = NULL;
2027 	spin_unlock(&fs_info->trans_lock);
2028 
2029 	if (trans->type & __TRANS_FREEZABLE)
2030 		sb_end_intwrite(fs_info->sb);
2031 	btrfs_put_transaction(cur_trans);
2032 	btrfs_put_transaction(cur_trans);
2033 
2034 	trace_btrfs_transaction_commit(fs_info);
2035 
2036 	if (current->journal_info == trans)
2037 		current->journal_info = NULL;
2038 	btrfs_scrub_cancel(fs_info);
2039 
2040 	kmem_cache_free(btrfs_trans_handle_cachep, trans);
2041 }
2042 
2043 /*
2044  * Release reserved delayed ref space of all pending block groups of the
2045  * transaction and remove them from the list
2046  */
2047 static void btrfs_cleanup_pending_block_groups(struct btrfs_trans_handle *trans)
2048 {
2049        struct btrfs_fs_info *fs_info = trans->fs_info;
2050        struct btrfs_block_group *block_group, *tmp;
2051 
2052        list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
2053                btrfs_delayed_refs_rsv_release(fs_info, 1);
2054                list_del_init(&block_group->bg_list);
2055        }
2056 }
2057 
2058 static inline int btrfs_start_delalloc_flush(struct btrfs_fs_info *fs_info)
2059 {
2060 	/*
2061 	 * We use try_to_writeback_inodes_sb() here because if we used
2062 	 * btrfs_start_delalloc_roots we would deadlock with fs freeze.
2063 	 * Currently are holding the fs freeze lock, if we do an async flush
2064 	 * we'll do btrfs_join_transaction() and deadlock because we need to
2065 	 * wait for the fs freeze lock.  Using the direct flushing we benefit
2066 	 * from already being in a transaction and our join_transaction doesn't
2067 	 * have to re-take the fs freeze lock.
2068 	 *
2069 	 * Note that try_to_writeback_inodes_sb() will only trigger writeback
2070 	 * if it can read lock sb->s_umount. It will always be able to lock it,
2071 	 * except when the filesystem is being unmounted or being frozen, but in
2072 	 * those cases sync_filesystem() is called, which results in calling
2073 	 * writeback_inodes_sb() while holding a write lock on sb->s_umount.
2074 	 * Note that we don't call writeback_inodes_sb() directly, because it
2075 	 * will emit a warning if sb->s_umount is not locked.
2076 	 */
2077 	if (btrfs_test_opt(fs_info, FLUSHONCOMMIT))
2078 		try_to_writeback_inodes_sb(fs_info->sb, WB_REASON_SYNC);
2079 	return 0;
2080 }
2081 
2082 static inline void btrfs_wait_delalloc_flush(struct btrfs_fs_info *fs_info)
2083 {
2084 	if (btrfs_test_opt(fs_info, FLUSHONCOMMIT))
2085 		btrfs_wait_ordered_roots(fs_info, U64_MAX, 0, (u64)-1);
2086 }
2087 
2088 /*
2089  * Add a pending snapshot associated with the given transaction handle to the
2090  * respective handle. This must be called after the transaction commit started
2091  * and while holding fs_info->trans_lock.
2092  * This serves to guarantee a caller of btrfs_commit_transaction() that it can
2093  * safely free the pending snapshot pointer in case btrfs_commit_transaction()
2094  * returns an error.
2095  */
2096 static void add_pending_snapshot(struct btrfs_trans_handle *trans)
2097 {
2098 	struct btrfs_transaction *cur_trans = trans->transaction;
2099 
2100 	if (!trans->pending_snapshot)
2101 		return;
2102 
2103 	lockdep_assert_held(&trans->fs_info->trans_lock);
2104 	ASSERT(cur_trans->state >= TRANS_STATE_COMMIT_START);
2105 
2106 	list_add(&trans->pending_snapshot->list, &cur_trans->pending_snapshots);
2107 }
2108 
2109 static void update_commit_stats(struct btrfs_fs_info *fs_info, ktime_t interval)
2110 {
2111 	fs_info->commit_stats.commit_count++;
2112 	fs_info->commit_stats.last_commit_dur = interval;
2113 	fs_info->commit_stats.max_commit_dur =
2114 			max_t(u64, fs_info->commit_stats.max_commit_dur, interval);
2115 	fs_info->commit_stats.total_commit_dur += interval;
2116 }
2117 
2118 int btrfs_commit_transaction(struct btrfs_trans_handle *trans)
2119 {
2120 	struct btrfs_fs_info *fs_info = trans->fs_info;
2121 	struct btrfs_transaction *cur_trans = trans->transaction;
2122 	struct btrfs_transaction *prev_trans = NULL;
2123 	int ret;
2124 	ktime_t start_time;
2125 	ktime_t interval;
2126 
2127 	ASSERT(refcount_read(&trans->use_count) == 1);
2128 	btrfs_trans_state_lockdep_acquire(fs_info, BTRFS_LOCKDEP_TRANS_COMMIT_START);
2129 
2130 	clear_bit(BTRFS_FS_NEED_TRANS_COMMIT, &fs_info->flags);
2131 
2132 	/* Stop the commit early if ->aborted is set */
2133 	if (TRANS_ABORTED(cur_trans)) {
2134 		ret = cur_trans->aborted;
2135 		goto lockdep_trans_commit_start_release;
2136 	}
2137 
2138 	btrfs_trans_release_metadata(trans);
2139 	trans->block_rsv = NULL;
2140 
2141 	/*
2142 	 * We only want one transaction commit doing the flushing so we do not
2143 	 * waste a bunch of time on lock contention on the extent root node.
2144 	 */
2145 	if (!test_and_set_bit(BTRFS_DELAYED_REFS_FLUSHING,
2146 			      &cur_trans->delayed_refs.flags)) {
2147 		/*
2148 		 * Make a pass through all the delayed refs we have so far.
2149 		 * Any running threads may add more while we are here.
2150 		 */
2151 		ret = btrfs_run_delayed_refs(trans, 0);
2152 		if (ret)
2153 			goto lockdep_trans_commit_start_release;
2154 	}
2155 
2156 	btrfs_create_pending_block_groups(trans);
2157 
2158 	if (!test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &cur_trans->flags)) {
2159 		int run_it = 0;
2160 
2161 		/* this mutex is also taken before trying to set
2162 		 * block groups readonly.  We need to make sure
2163 		 * that nobody has set a block group readonly
2164 		 * after a extents from that block group have been
2165 		 * allocated for cache files.  btrfs_set_block_group_ro
2166 		 * will wait for the transaction to commit if it
2167 		 * finds BTRFS_TRANS_DIRTY_BG_RUN set.
2168 		 *
2169 		 * The BTRFS_TRANS_DIRTY_BG_RUN flag is also used to make sure
2170 		 * only one process starts all the block group IO.  It wouldn't
2171 		 * hurt to have more than one go through, but there's no
2172 		 * real advantage to it either.
2173 		 */
2174 		mutex_lock(&fs_info->ro_block_group_mutex);
2175 		if (!test_and_set_bit(BTRFS_TRANS_DIRTY_BG_RUN,
2176 				      &cur_trans->flags))
2177 			run_it = 1;
2178 		mutex_unlock(&fs_info->ro_block_group_mutex);
2179 
2180 		if (run_it) {
2181 			ret = btrfs_start_dirty_block_groups(trans);
2182 			if (ret)
2183 				goto lockdep_trans_commit_start_release;
2184 		}
2185 	}
2186 
2187 	spin_lock(&fs_info->trans_lock);
2188 	if (cur_trans->state >= TRANS_STATE_COMMIT_START) {
2189 		enum btrfs_trans_state want_state = TRANS_STATE_COMPLETED;
2190 
2191 		add_pending_snapshot(trans);
2192 
2193 		spin_unlock(&fs_info->trans_lock);
2194 		refcount_inc(&cur_trans->use_count);
2195 
2196 		if (trans->in_fsync)
2197 			want_state = TRANS_STATE_SUPER_COMMITTED;
2198 
2199 		btrfs_trans_state_lockdep_release(fs_info,
2200 						  BTRFS_LOCKDEP_TRANS_COMMIT_START);
2201 		ret = btrfs_end_transaction(trans);
2202 		wait_for_commit(cur_trans, want_state);
2203 
2204 		if (TRANS_ABORTED(cur_trans))
2205 			ret = cur_trans->aborted;
2206 
2207 		btrfs_put_transaction(cur_trans);
2208 
2209 		return ret;
2210 	}
2211 
2212 	cur_trans->state = TRANS_STATE_COMMIT_START;
2213 	wake_up(&fs_info->transaction_blocked_wait);
2214 	btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_COMMIT_START);
2215 
2216 	if (cur_trans->list.prev != &fs_info->trans_list) {
2217 		enum btrfs_trans_state want_state = TRANS_STATE_COMPLETED;
2218 
2219 		if (trans->in_fsync)
2220 			want_state = TRANS_STATE_SUPER_COMMITTED;
2221 
2222 		prev_trans = list_entry(cur_trans->list.prev,
2223 					struct btrfs_transaction, list);
2224 		if (prev_trans->state < want_state) {
2225 			refcount_inc(&prev_trans->use_count);
2226 			spin_unlock(&fs_info->trans_lock);
2227 
2228 			wait_for_commit(prev_trans, want_state);
2229 
2230 			ret = READ_ONCE(prev_trans->aborted);
2231 
2232 			btrfs_put_transaction(prev_trans);
2233 			if (ret)
2234 				goto lockdep_release;
2235 		} else {
2236 			spin_unlock(&fs_info->trans_lock);
2237 		}
2238 	} else {
2239 		spin_unlock(&fs_info->trans_lock);
2240 		/*
2241 		 * The previous transaction was aborted and was already removed
2242 		 * from the list of transactions at fs_info->trans_list. So we
2243 		 * abort to prevent writing a new superblock that reflects a
2244 		 * corrupt state (pointing to trees with unwritten nodes/leafs).
2245 		 */
2246 		if (BTRFS_FS_ERROR(fs_info)) {
2247 			ret = -EROFS;
2248 			goto lockdep_release;
2249 		}
2250 	}
2251 
2252 	/*
2253 	 * Get the time spent on the work done by the commit thread and not
2254 	 * the time spent waiting on a previous commit
2255 	 */
2256 	start_time = ktime_get_ns();
2257 
2258 	extwriter_counter_dec(cur_trans, trans->type);
2259 
2260 	ret = btrfs_start_delalloc_flush(fs_info);
2261 	if (ret)
2262 		goto lockdep_release;
2263 
2264 	ret = btrfs_run_delayed_items(trans);
2265 	if (ret)
2266 		goto lockdep_release;
2267 
2268 	/*
2269 	 * The thread has started/joined the transaction thus it holds the
2270 	 * lockdep map as a reader. It has to release it before acquiring the
2271 	 * lockdep map as a writer.
2272 	 */
2273 	btrfs_lockdep_release(fs_info, btrfs_trans_num_extwriters);
2274 	btrfs_might_wait_for_event(fs_info, btrfs_trans_num_extwriters);
2275 	wait_event(cur_trans->writer_wait,
2276 		   extwriter_counter_read(cur_trans) == 0);
2277 
2278 	/* some pending stuffs might be added after the previous flush. */
2279 	ret = btrfs_run_delayed_items(trans);
2280 	if (ret) {
2281 		btrfs_lockdep_release(fs_info, btrfs_trans_num_writers);
2282 		goto cleanup_transaction;
2283 	}
2284 
2285 	btrfs_wait_delalloc_flush(fs_info);
2286 
2287 	/*
2288 	 * Wait for all ordered extents started by a fast fsync that joined this
2289 	 * transaction. Otherwise if this transaction commits before the ordered
2290 	 * extents complete we lose logged data after a power failure.
2291 	 */
2292 	btrfs_might_wait_for_event(fs_info, btrfs_trans_pending_ordered);
2293 	wait_event(cur_trans->pending_wait,
2294 		   atomic_read(&cur_trans->pending_ordered) == 0);
2295 
2296 	btrfs_scrub_pause(fs_info);
2297 	/*
2298 	 * Ok now we need to make sure to block out any other joins while we
2299 	 * commit the transaction.  We could have started a join before setting
2300 	 * COMMIT_DOING so make sure to wait for num_writers to == 1 again.
2301 	 */
2302 	spin_lock(&fs_info->trans_lock);
2303 	add_pending_snapshot(trans);
2304 	cur_trans->state = TRANS_STATE_COMMIT_DOING;
2305 	spin_unlock(&fs_info->trans_lock);
2306 
2307 	/*
2308 	 * The thread has started/joined the transaction thus it holds the
2309 	 * lockdep map as a reader. It has to release it before acquiring the
2310 	 * lockdep map as a writer.
2311 	 */
2312 	btrfs_lockdep_release(fs_info, btrfs_trans_num_writers);
2313 	btrfs_might_wait_for_event(fs_info, btrfs_trans_num_writers);
2314 	wait_event(cur_trans->writer_wait,
2315 		   atomic_read(&cur_trans->num_writers) == 1);
2316 
2317 	/*
2318 	 * Make lockdep happy by acquiring the state locks after
2319 	 * btrfs_trans_num_writers is released. If we acquired the state locks
2320 	 * before releasing the btrfs_trans_num_writers lock then lockdep would
2321 	 * complain because we did not follow the reverse order unlocking rule.
2322 	 */
2323 	btrfs_trans_state_lockdep_acquire(fs_info, BTRFS_LOCKDEP_TRANS_COMPLETED);
2324 	btrfs_trans_state_lockdep_acquire(fs_info, BTRFS_LOCKDEP_TRANS_SUPER_COMMITTED);
2325 	btrfs_trans_state_lockdep_acquire(fs_info, BTRFS_LOCKDEP_TRANS_UNBLOCKED);
2326 
2327 	/*
2328 	 * We've started the commit, clear the flag in case we were triggered to
2329 	 * do an async commit but somebody else started before the transaction
2330 	 * kthread could do the work.
2331 	 */
2332 	clear_bit(BTRFS_FS_COMMIT_TRANS, &fs_info->flags);
2333 
2334 	if (TRANS_ABORTED(cur_trans)) {
2335 		ret = cur_trans->aborted;
2336 		btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_UNBLOCKED);
2337 		goto scrub_continue;
2338 	}
2339 	/*
2340 	 * the reloc mutex makes sure that we stop
2341 	 * the balancing code from coming in and moving
2342 	 * extents around in the middle of the commit
2343 	 */
2344 	mutex_lock(&fs_info->reloc_mutex);
2345 
2346 	/*
2347 	 * We needn't worry about the delayed items because we will
2348 	 * deal with them in create_pending_snapshot(), which is the
2349 	 * core function of the snapshot creation.
2350 	 */
2351 	ret = create_pending_snapshots(trans);
2352 	if (ret)
2353 		goto unlock_reloc;
2354 
2355 	/*
2356 	 * We insert the dir indexes of the snapshots and update the inode
2357 	 * of the snapshots' parents after the snapshot creation, so there
2358 	 * are some delayed items which are not dealt with. Now deal with
2359 	 * them.
2360 	 *
2361 	 * We needn't worry that this operation will corrupt the snapshots,
2362 	 * because all the tree which are snapshoted will be forced to COW
2363 	 * the nodes and leaves.
2364 	 */
2365 	ret = btrfs_run_delayed_items(trans);
2366 	if (ret)
2367 		goto unlock_reloc;
2368 
2369 	ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
2370 	if (ret)
2371 		goto unlock_reloc;
2372 
2373 	/*
2374 	 * make sure none of the code above managed to slip in a
2375 	 * delayed item
2376 	 */
2377 	btrfs_assert_delayed_root_empty(fs_info);
2378 
2379 	WARN_ON(cur_trans != trans->transaction);
2380 
2381 	ret = commit_fs_roots(trans);
2382 	if (ret)
2383 		goto unlock_reloc;
2384 
2385 	/* commit_fs_roots gets rid of all the tree log roots, it is now
2386 	 * safe to free the root of tree log roots
2387 	 */
2388 	btrfs_free_log_root_tree(trans, fs_info);
2389 
2390 	/*
2391 	 * Since fs roots are all committed, we can get a quite accurate
2392 	 * new_roots. So let's do quota accounting.
2393 	 */
2394 	ret = btrfs_qgroup_account_extents(trans);
2395 	if (ret < 0)
2396 		goto unlock_reloc;
2397 
2398 	ret = commit_cowonly_roots(trans);
2399 	if (ret)
2400 		goto unlock_reloc;
2401 
2402 	/*
2403 	 * The tasks which save the space cache and inode cache may also
2404 	 * update ->aborted, check it.
2405 	 */
2406 	if (TRANS_ABORTED(cur_trans)) {
2407 		ret = cur_trans->aborted;
2408 		goto unlock_reloc;
2409 	}
2410 
2411 	cur_trans = fs_info->running_transaction;
2412 
2413 	btrfs_set_root_node(&fs_info->tree_root->root_item,
2414 			    fs_info->tree_root->node);
2415 	list_add_tail(&fs_info->tree_root->dirty_list,
2416 		      &cur_trans->switch_commits);
2417 
2418 	btrfs_set_root_node(&fs_info->chunk_root->root_item,
2419 			    fs_info->chunk_root->node);
2420 	list_add_tail(&fs_info->chunk_root->dirty_list,
2421 		      &cur_trans->switch_commits);
2422 
2423 	if (btrfs_fs_incompat(fs_info, EXTENT_TREE_V2)) {
2424 		btrfs_set_root_node(&fs_info->block_group_root->root_item,
2425 				    fs_info->block_group_root->node);
2426 		list_add_tail(&fs_info->block_group_root->dirty_list,
2427 			      &cur_trans->switch_commits);
2428 	}
2429 
2430 	switch_commit_roots(trans);
2431 
2432 	ASSERT(list_empty(&cur_trans->dirty_bgs));
2433 	ASSERT(list_empty(&cur_trans->io_bgs));
2434 	update_super_roots(fs_info);
2435 
2436 	btrfs_set_super_log_root(fs_info->super_copy, 0);
2437 	btrfs_set_super_log_root_level(fs_info->super_copy, 0);
2438 	memcpy(fs_info->super_for_commit, fs_info->super_copy,
2439 	       sizeof(*fs_info->super_copy));
2440 
2441 	btrfs_commit_device_sizes(cur_trans);
2442 
2443 	clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags);
2444 	clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags);
2445 
2446 	btrfs_trans_release_chunk_metadata(trans);
2447 
2448 	/*
2449 	 * Before changing the transaction state to TRANS_STATE_UNBLOCKED and
2450 	 * setting fs_info->running_transaction to NULL, lock tree_log_mutex to
2451 	 * make sure that before we commit our superblock, no other task can
2452 	 * start a new transaction and commit a log tree before we commit our
2453 	 * superblock. Anyone trying to commit a log tree locks this mutex before
2454 	 * writing its superblock.
2455 	 */
2456 	mutex_lock(&fs_info->tree_log_mutex);
2457 
2458 	spin_lock(&fs_info->trans_lock);
2459 	cur_trans->state = TRANS_STATE_UNBLOCKED;
2460 	fs_info->running_transaction = NULL;
2461 	spin_unlock(&fs_info->trans_lock);
2462 	mutex_unlock(&fs_info->reloc_mutex);
2463 
2464 	wake_up(&fs_info->transaction_wait);
2465 	btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_UNBLOCKED);
2466 
2467 	/* If we have features changed, wake up the cleaner to update sysfs. */
2468 	if (test_bit(BTRFS_FS_FEATURE_CHANGED, &fs_info->flags) &&
2469 	    fs_info->cleaner_kthread)
2470 		wake_up_process(fs_info->cleaner_kthread);
2471 
2472 	ret = btrfs_write_and_wait_transaction(trans);
2473 	if (ret) {
2474 		btrfs_handle_fs_error(fs_info, ret,
2475 				      "Error while writing out transaction");
2476 		mutex_unlock(&fs_info->tree_log_mutex);
2477 		goto scrub_continue;
2478 	}
2479 
2480 	/*
2481 	 * At this point, we should have written all the tree blocks allocated
2482 	 * in this transaction. So it's now safe to free the redirtyied extent
2483 	 * buffers.
2484 	 */
2485 	btrfs_free_redirty_list(cur_trans);
2486 
2487 	ret = write_all_supers(fs_info, 0);
2488 	/*
2489 	 * the super is written, we can safely allow the tree-loggers
2490 	 * to go about their business
2491 	 */
2492 	mutex_unlock(&fs_info->tree_log_mutex);
2493 	if (ret)
2494 		goto scrub_continue;
2495 
2496 	/*
2497 	 * We needn't acquire the lock here because there is no other task
2498 	 * which can change it.
2499 	 */
2500 	cur_trans->state = TRANS_STATE_SUPER_COMMITTED;
2501 	wake_up(&cur_trans->commit_wait);
2502 	btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_SUPER_COMMITTED);
2503 
2504 	btrfs_finish_extent_commit(trans);
2505 
2506 	if (test_bit(BTRFS_TRANS_HAVE_FREE_BGS, &cur_trans->flags))
2507 		btrfs_clear_space_info_full(fs_info);
2508 
2509 	fs_info->last_trans_committed = cur_trans->transid;
2510 	/*
2511 	 * We needn't acquire the lock here because there is no other task
2512 	 * which can change it.
2513 	 */
2514 	cur_trans->state = TRANS_STATE_COMPLETED;
2515 	wake_up(&cur_trans->commit_wait);
2516 	btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_COMPLETED);
2517 
2518 	spin_lock(&fs_info->trans_lock);
2519 	list_del_init(&cur_trans->list);
2520 	spin_unlock(&fs_info->trans_lock);
2521 
2522 	btrfs_put_transaction(cur_trans);
2523 	btrfs_put_transaction(cur_trans);
2524 
2525 	if (trans->type & __TRANS_FREEZABLE)
2526 		sb_end_intwrite(fs_info->sb);
2527 
2528 	trace_btrfs_transaction_commit(fs_info);
2529 
2530 	interval = ktime_get_ns() - start_time;
2531 
2532 	btrfs_scrub_continue(fs_info);
2533 
2534 	if (current->journal_info == trans)
2535 		current->journal_info = NULL;
2536 
2537 	kmem_cache_free(btrfs_trans_handle_cachep, trans);
2538 
2539 	update_commit_stats(fs_info, interval);
2540 
2541 	return ret;
2542 
2543 unlock_reloc:
2544 	mutex_unlock(&fs_info->reloc_mutex);
2545 	btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_UNBLOCKED);
2546 scrub_continue:
2547 	btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_SUPER_COMMITTED);
2548 	btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_COMPLETED);
2549 	btrfs_scrub_continue(fs_info);
2550 cleanup_transaction:
2551 	btrfs_trans_release_metadata(trans);
2552 	btrfs_cleanup_pending_block_groups(trans);
2553 	btrfs_trans_release_chunk_metadata(trans);
2554 	trans->block_rsv = NULL;
2555 	btrfs_warn(fs_info, "Skipping commit of aborted transaction.");
2556 	if (current->journal_info == trans)
2557 		current->journal_info = NULL;
2558 	cleanup_transaction(trans, ret);
2559 
2560 	return ret;
2561 
2562 lockdep_release:
2563 	btrfs_lockdep_release(fs_info, btrfs_trans_num_extwriters);
2564 	btrfs_lockdep_release(fs_info, btrfs_trans_num_writers);
2565 	goto cleanup_transaction;
2566 
2567 lockdep_trans_commit_start_release:
2568 	btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_COMMIT_START);
2569 	btrfs_end_transaction(trans);
2570 	return ret;
2571 }
2572 
2573 /*
2574  * return < 0 if error
2575  * 0 if there are no more dead_roots at the time of call
2576  * 1 there are more to be processed, call me again
2577  *
2578  * The return value indicates there are certainly more snapshots to delete, but
2579  * if there comes a new one during processing, it may return 0. We don't mind,
2580  * because btrfs_commit_super will poke cleaner thread and it will process it a
2581  * few seconds later.
2582  */
2583 int btrfs_clean_one_deleted_snapshot(struct btrfs_fs_info *fs_info)
2584 {
2585 	struct btrfs_root *root;
2586 	int ret;
2587 
2588 	spin_lock(&fs_info->trans_lock);
2589 	if (list_empty(&fs_info->dead_roots)) {
2590 		spin_unlock(&fs_info->trans_lock);
2591 		return 0;
2592 	}
2593 	root = list_first_entry(&fs_info->dead_roots,
2594 			struct btrfs_root, root_list);
2595 	list_del_init(&root->root_list);
2596 	spin_unlock(&fs_info->trans_lock);
2597 
2598 	btrfs_debug(fs_info, "cleaner removing %llu", root->root_key.objectid);
2599 
2600 	btrfs_kill_all_delayed_nodes(root);
2601 
2602 	if (btrfs_header_backref_rev(root->node) <
2603 			BTRFS_MIXED_BACKREF_REV)
2604 		ret = btrfs_drop_snapshot(root, 0, 0);
2605 	else
2606 		ret = btrfs_drop_snapshot(root, 1, 0);
2607 
2608 	btrfs_put_root(root);
2609 	return (ret < 0) ? 0 : 1;
2610 }
2611 
2612 /*
2613  * We only mark the transaction aborted and then set the file system read-only.
2614  * This will prevent new transactions from starting or trying to join this
2615  * one.
2616  *
2617  * This means that error recovery at the call site is limited to freeing
2618  * any local memory allocations and passing the error code up without
2619  * further cleanup. The transaction should complete as it normally would
2620  * in the call path but will return -EIO.
2621  *
2622  * We'll complete the cleanup in btrfs_end_transaction and
2623  * btrfs_commit_transaction.
2624  */
2625 void __cold __btrfs_abort_transaction(struct btrfs_trans_handle *trans,
2626 				      const char *function,
2627 				      unsigned int line, int errno, bool first_hit)
2628 {
2629 	struct btrfs_fs_info *fs_info = trans->fs_info;
2630 
2631 	WRITE_ONCE(trans->aborted, errno);
2632 	WRITE_ONCE(trans->transaction->aborted, errno);
2633 	if (first_hit && errno == -ENOSPC)
2634 		btrfs_dump_space_info_for_trans_abort(fs_info);
2635 	/* Wake up anybody who may be waiting on this transaction */
2636 	wake_up(&fs_info->transaction_wait);
2637 	wake_up(&fs_info->transaction_blocked_wait);
2638 	__btrfs_handle_fs_error(fs_info, function, line, errno, NULL);
2639 }
2640 
2641 int __init btrfs_transaction_init(void)
2642 {
2643 	btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
2644 			sizeof(struct btrfs_trans_handle), 0,
2645 			SLAB_TEMPORARY | SLAB_MEM_SPREAD, NULL);
2646 	if (!btrfs_trans_handle_cachep)
2647 		return -ENOMEM;
2648 	return 0;
2649 }
2650 
2651 void __cold btrfs_transaction_exit(void)
2652 {
2653 	kmem_cache_destroy(btrfs_trans_handle_cachep);
2654 }
2655