xref: /openbmc/linux/fs/btrfs/transaction.c (revision ae3473231e77a3f1909d48cd144cebe5e1d049b3)
1 /*
2  * Copyright (C) 2007 Oracle.  All rights reserved.
3  *
4  * This program is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU General Public
6  * License v2 as published by the Free Software Foundation.
7  *
8  * This program is distributed in the hope that it will be useful,
9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
11  * General Public License for more details.
12  *
13  * You should have received a copy of the GNU General Public
14  * License along with this program; if not, write to the
15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16  * Boston, MA 021110-1307, USA.
17  */
18 
19 #include <linux/fs.h>
20 #include <linux/slab.h>
21 #include <linux/sched.h>
22 #include <linux/writeback.h>
23 #include <linux/pagemap.h>
24 #include <linux/blkdev.h>
25 #include <linux/uuid.h>
26 #include "ctree.h"
27 #include "disk-io.h"
28 #include "transaction.h"
29 #include "locking.h"
30 #include "tree-log.h"
31 #include "inode-map.h"
32 #include "volumes.h"
33 #include "dev-replace.h"
34 #include "qgroup.h"
35 
36 #define BTRFS_ROOT_TRANS_TAG 0
37 
38 static const unsigned int btrfs_blocked_trans_types[TRANS_STATE_MAX] = {
39 	[TRANS_STATE_RUNNING]		= 0U,
40 	[TRANS_STATE_BLOCKED]		= (__TRANS_USERSPACE |
41 					   __TRANS_START),
42 	[TRANS_STATE_COMMIT_START]	= (__TRANS_USERSPACE |
43 					   __TRANS_START |
44 					   __TRANS_ATTACH),
45 	[TRANS_STATE_COMMIT_DOING]	= (__TRANS_USERSPACE |
46 					   __TRANS_START |
47 					   __TRANS_ATTACH |
48 					   __TRANS_JOIN),
49 	[TRANS_STATE_UNBLOCKED]		= (__TRANS_USERSPACE |
50 					   __TRANS_START |
51 					   __TRANS_ATTACH |
52 					   __TRANS_JOIN |
53 					   __TRANS_JOIN_NOLOCK),
54 	[TRANS_STATE_COMPLETED]		= (__TRANS_USERSPACE |
55 					   __TRANS_START |
56 					   __TRANS_ATTACH |
57 					   __TRANS_JOIN |
58 					   __TRANS_JOIN_NOLOCK),
59 };
60 
61 void btrfs_put_transaction(struct btrfs_transaction *transaction)
62 {
63 	WARN_ON(atomic_read(&transaction->use_count) == 0);
64 	if (atomic_dec_and_test(&transaction->use_count)) {
65 		BUG_ON(!list_empty(&transaction->list));
66 		WARN_ON(!RB_EMPTY_ROOT(&transaction->delayed_refs.href_root));
67 		if (transaction->delayed_refs.pending_csums)
68 			btrfs_err(transaction->fs_info,
69 				  "pending csums is %llu",
70 				  transaction->delayed_refs.pending_csums);
71 		while (!list_empty(&transaction->pending_chunks)) {
72 			struct extent_map *em;
73 
74 			em = list_first_entry(&transaction->pending_chunks,
75 					      struct extent_map, list);
76 			list_del_init(&em->list);
77 			free_extent_map(em);
78 		}
79 		/*
80 		 * If any block groups are found in ->deleted_bgs then it's
81 		 * because the transaction was aborted and a commit did not
82 		 * happen (things failed before writing the new superblock
83 		 * and calling btrfs_finish_extent_commit()), so we can not
84 		 * discard the physical locations of the block groups.
85 		 */
86 		while (!list_empty(&transaction->deleted_bgs)) {
87 			struct btrfs_block_group_cache *cache;
88 
89 			cache = list_first_entry(&transaction->deleted_bgs,
90 						 struct btrfs_block_group_cache,
91 						 bg_list);
92 			list_del_init(&cache->bg_list);
93 			btrfs_put_block_group_trimming(cache);
94 			btrfs_put_block_group(cache);
95 		}
96 		kmem_cache_free(btrfs_transaction_cachep, transaction);
97 	}
98 }
99 
100 static void clear_btree_io_tree(struct extent_io_tree *tree)
101 {
102 	spin_lock(&tree->lock);
103 	/*
104 	 * Do a single barrier for the waitqueue_active check here, the state
105 	 * of the waitqueue should not change once clear_btree_io_tree is
106 	 * called.
107 	 */
108 	smp_mb();
109 	while (!RB_EMPTY_ROOT(&tree->state)) {
110 		struct rb_node *node;
111 		struct extent_state *state;
112 
113 		node = rb_first(&tree->state);
114 		state = rb_entry(node, struct extent_state, rb_node);
115 		rb_erase(&state->rb_node, &tree->state);
116 		RB_CLEAR_NODE(&state->rb_node);
117 		/*
118 		 * btree io trees aren't supposed to have tasks waiting for
119 		 * changes in the flags of extent states ever.
120 		 */
121 		ASSERT(!waitqueue_active(&state->wq));
122 		free_extent_state(state);
123 
124 		cond_resched_lock(&tree->lock);
125 	}
126 	spin_unlock(&tree->lock);
127 }
128 
129 static noinline void switch_commit_roots(struct btrfs_transaction *trans,
130 					 struct btrfs_fs_info *fs_info)
131 {
132 	struct btrfs_root *root, *tmp;
133 
134 	down_write(&fs_info->commit_root_sem);
135 	list_for_each_entry_safe(root, tmp, &trans->switch_commits,
136 				 dirty_list) {
137 		list_del_init(&root->dirty_list);
138 		free_extent_buffer(root->commit_root);
139 		root->commit_root = btrfs_root_node(root);
140 		if (is_fstree(root->objectid))
141 			btrfs_unpin_free_ino(root);
142 		clear_btree_io_tree(&root->dirty_log_pages);
143 	}
144 
145 	/* We can free old roots now. */
146 	spin_lock(&trans->dropped_roots_lock);
147 	while (!list_empty(&trans->dropped_roots)) {
148 		root = list_first_entry(&trans->dropped_roots,
149 					struct btrfs_root, root_list);
150 		list_del_init(&root->root_list);
151 		spin_unlock(&trans->dropped_roots_lock);
152 		btrfs_drop_and_free_fs_root(fs_info, root);
153 		spin_lock(&trans->dropped_roots_lock);
154 	}
155 	spin_unlock(&trans->dropped_roots_lock);
156 	up_write(&fs_info->commit_root_sem);
157 }
158 
159 static inline void extwriter_counter_inc(struct btrfs_transaction *trans,
160 					 unsigned int type)
161 {
162 	if (type & TRANS_EXTWRITERS)
163 		atomic_inc(&trans->num_extwriters);
164 }
165 
166 static inline void extwriter_counter_dec(struct btrfs_transaction *trans,
167 					 unsigned int type)
168 {
169 	if (type & TRANS_EXTWRITERS)
170 		atomic_dec(&trans->num_extwriters);
171 }
172 
173 static inline void extwriter_counter_init(struct btrfs_transaction *trans,
174 					  unsigned int type)
175 {
176 	atomic_set(&trans->num_extwriters, ((type & TRANS_EXTWRITERS) ? 1 : 0));
177 }
178 
179 static inline int extwriter_counter_read(struct btrfs_transaction *trans)
180 {
181 	return atomic_read(&trans->num_extwriters);
182 }
183 
184 /*
185  * either allocate a new transaction or hop into the existing one
186  */
187 static noinline int join_transaction(struct btrfs_fs_info *fs_info,
188 				     unsigned int type)
189 {
190 	struct btrfs_transaction *cur_trans;
191 
192 	spin_lock(&fs_info->trans_lock);
193 loop:
194 	/* The file system has been taken offline. No new transactions. */
195 	if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
196 		spin_unlock(&fs_info->trans_lock);
197 		return -EROFS;
198 	}
199 
200 	cur_trans = fs_info->running_transaction;
201 	if (cur_trans) {
202 		if (cur_trans->aborted) {
203 			spin_unlock(&fs_info->trans_lock);
204 			return cur_trans->aborted;
205 		}
206 		if (btrfs_blocked_trans_types[cur_trans->state] & type) {
207 			spin_unlock(&fs_info->trans_lock);
208 			return -EBUSY;
209 		}
210 		atomic_inc(&cur_trans->use_count);
211 		atomic_inc(&cur_trans->num_writers);
212 		extwriter_counter_inc(cur_trans, type);
213 		spin_unlock(&fs_info->trans_lock);
214 		return 0;
215 	}
216 	spin_unlock(&fs_info->trans_lock);
217 
218 	/*
219 	 * If we are ATTACH, we just want to catch the current transaction,
220 	 * and commit it. If there is no transaction, just return ENOENT.
221 	 */
222 	if (type == TRANS_ATTACH)
223 		return -ENOENT;
224 
225 	/*
226 	 * JOIN_NOLOCK only happens during the transaction commit, so
227 	 * it is impossible that ->running_transaction is NULL
228 	 */
229 	BUG_ON(type == TRANS_JOIN_NOLOCK);
230 
231 	cur_trans = kmem_cache_alloc(btrfs_transaction_cachep, GFP_NOFS);
232 	if (!cur_trans)
233 		return -ENOMEM;
234 
235 	spin_lock(&fs_info->trans_lock);
236 	if (fs_info->running_transaction) {
237 		/*
238 		 * someone started a transaction after we unlocked.  Make sure
239 		 * to redo the checks above
240 		 */
241 		kmem_cache_free(btrfs_transaction_cachep, cur_trans);
242 		goto loop;
243 	} else if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
244 		spin_unlock(&fs_info->trans_lock);
245 		kmem_cache_free(btrfs_transaction_cachep, cur_trans);
246 		return -EROFS;
247 	}
248 
249 	cur_trans->fs_info = fs_info;
250 	atomic_set(&cur_trans->num_writers, 1);
251 	extwriter_counter_init(cur_trans, type);
252 	init_waitqueue_head(&cur_trans->writer_wait);
253 	init_waitqueue_head(&cur_trans->commit_wait);
254 	init_waitqueue_head(&cur_trans->pending_wait);
255 	cur_trans->state = TRANS_STATE_RUNNING;
256 	/*
257 	 * One for this trans handle, one so it will live on until we
258 	 * commit the transaction.
259 	 */
260 	atomic_set(&cur_trans->use_count, 2);
261 	atomic_set(&cur_trans->pending_ordered, 0);
262 	cur_trans->flags = 0;
263 	cur_trans->start_time = get_seconds();
264 
265 	memset(&cur_trans->delayed_refs, 0, sizeof(cur_trans->delayed_refs));
266 
267 	cur_trans->delayed_refs.href_root = RB_ROOT;
268 	cur_trans->delayed_refs.dirty_extent_root = RB_ROOT;
269 	atomic_set(&cur_trans->delayed_refs.num_entries, 0);
270 
271 	/*
272 	 * although the tree mod log is per file system and not per transaction,
273 	 * the log must never go across transaction boundaries.
274 	 */
275 	smp_mb();
276 	if (!list_empty(&fs_info->tree_mod_seq_list))
277 		WARN(1, KERN_ERR "BTRFS: tree_mod_seq_list not empty when creating a fresh transaction\n");
278 	if (!RB_EMPTY_ROOT(&fs_info->tree_mod_log))
279 		WARN(1, KERN_ERR "BTRFS: tree_mod_log rb tree not empty when creating a fresh transaction\n");
280 	atomic64_set(&fs_info->tree_mod_seq, 0);
281 
282 	spin_lock_init(&cur_trans->delayed_refs.lock);
283 
284 	INIT_LIST_HEAD(&cur_trans->pending_snapshots);
285 	INIT_LIST_HEAD(&cur_trans->pending_chunks);
286 	INIT_LIST_HEAD(&cur_trans->switch_commits);
287 	INIT_LIST_HEAD(&cur_trans->dirty_bgs);
288 	INIT_LIST_HEAD(&cur_trans->io_bgs);
289 	INIT_LIST_HEAD(&cur_trans->dropped_roots);
290 	mutex_init(&cur_trans->cache_write_mutex);
291 	cur_trans->num_dirty_bgs = 0;
292 	spin_lock_init(&cur_trans->dirty_bgs_lock);
293 	INIT_LIST_HEAD(&cur_trans->deleted_bgs);
294 	spin_lock_init(&cur_trans->dropped_roots_lock);
295 	list_add_tail(&cur_trans->list, &fs_info->trans_list);
296 	extent_io_tree_init(&cur_trans->dirty_pages,
297 			     fs_info->btree_inode->i_mapping);
298 	fs_info->generation++;
299 	cur_trans->transid = fs_info->generation;
300 	fs_info->running_transaction = cur_trans;
301 	cur_trans->aborted = 0;
302 	spin_unlock(&fs_info->trans_lock);
303 
304 	return 0;
305 }
306 
307 /*
308  * this does all the record keeping required to make sure that a reference
309  * counted root is properly recorded in a given transaction.  This is required
310  * to make sure the old root from before we joined the transaction is deleted
311  * when the transaction commits
312  */
313 static int record_root_in_trans(struct btrfs_trans_handle *trans,
314 			       struct btrfs_root *root,
315 			       int force)
316 {
317 	struct btrfs_fs_info *fs_info = root->fs_info;
318 
319 	if ((test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
320 	    root->last_trans < trans->transid) || force) {
321 		WARN_ON(root == fs_info->extent_root);
322 		WARN_ON(root->commit_root != root->node);
323 
324 		/*
325 		 * see below for IN_TRANS_SETUP usage rules
326 		 * we have the reloc mutex held now, so there
327 		 * is only one writer in this function
328 		 */
329 		set_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
330 
331 		/* make sure readers find IN_TRANS_SETUP before
332 		 * they find our root->last_trans update
333 		 */
334 		smp_wmb();
335 
336 		spin_lock(&fs_info->fs_roots_radix_lock);
337 		if (root->last_trans == trans->transid && !force) {
338 			spin_unlock(&fs_info->fs_roots_radix_lock);
339 			return 0;
340 		}
341 		radix_tree_tag_set(&fs_info->fs_roots_radix,
342 				   (unsigned long)root->root_key.objectid,
343 				   BTRFS_ROOT_TRANS_TAG);
344 		spin_unlock(&fs_info->fs_roots_radix_lock);
345 		root->last_trans = trans->transid;
346 
347 		/* this is pretty tricky.  We don't want to
348 		 * take the relocation lock in btrfs_record_root_in_trans
349 		 * unless we're really doing the first setup for this root in
350 		 * this transaction.
351 		 *
352 		 * Normally we'd use root->last_trans as a flag to decide
353 		 * if we want to take the expensive mutex.
354 		 *
355 		 * But, we have to set root->last_trans before we
356 		 * init the relocation root, otherwise, we trip over warnings
357 		 * in ctree.c.  The solution used here is to flag ourselves
358 		 * with root IN_TRANS_SETUP.  When this is 1, we're still
359 		 * fixing up the reloc trees and everyone must wait.
360 		 *
361 		 * When this is zero, they can trust root->last_trans and fly
362 		 * through btrfs_record_root_in_trans without having to take the
363 		 * lock.  smp_wmb() makes sure that all the writes above are
364 		 * done before we pop in the zero below
365 		 */
366 		btrfs_init_reloc_root(trans, root);
367 		smp_mb__before_atomic();
368 		clear_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
369 	}
370 	return 0;
371 }
372 
373 
374 void btrfs_add_dropped_root(struct btrfs_trans_handle *trans,
375 			    struct btrfs_root *root)
376 {
377 	struct btrfs_fs_info *fs_info = root->fs_info;
378 	struct btrfs_transaction *cur_trans = trans->transaction;
379 
380 	/* Add ourselves to the transaction dropped list */
381 	spin_lock(&cur_trans->dropped_roots_lock);
382 	list_add_tail(&root->root_list, &cur_trans->dropped_roots);
383 	spin_unlock(&cur_trans->dropped_roots_lock);
384 
385 	/* Make sure we don't try to update the root at commit time */
386 	spin_lock(&fs_info->fs_roots_radix_lock);
387 	radix_tree_tag_clear(&fs_info->fs_roots_radix,
388 			     (unsigned long)root->root_key.objectid,
389 			     BTRFS_ROOT_TRANS_TAG);
390 	spin_unlock(&fs_info->fs_roots_radix_lock);
391 }
392 
393 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
394 			       struct btrfs_root *root)
395 {
396 	struct btrfs_fs_info *fs_info = root->fs_info;
397 
398 	if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
399 		return 0;
400 
401 	/*
402 	 * see record_root_in_trans for comments about IN_TRANS_SETUP usage
403 	 * and barriers
404 	 */
405 	smp_rmb();
406 	if (root->last_trans == trans->transid &&
407 	    !test_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state))
408 		return 0;
409 
410 	mutex_lock(&fs_info->reloc_mutex);
411 	record_root_in_trans(trans, root, 0);
412 	mutex_unlock(&fs_info->reloc_mutex);
413 
414 	return 0;
415 }
416 
417 static inline int is_transaction_blocked(struct btrfs_transaction *trans)
418 {
419 	return (trans->state >= TRANS_STATE_BLOCKED &&
420 		trans->state < TRANS_STATE_UNBLOCKED &&
421 		!trans->aborted);
422 }
423 
424 /* wait for commit against the current transaction to become unblocked
425  * when this is done, it is safe to start a new transaction, but the current
426  * transaction might not be fully on disk.
427  */
428 static void wait_current_trans(struct btrfs_fs_info *fs_info)
429 {
430 	struct btrfs_transaction *cur_trans;
431 
432 	spin_lock(&fs_info->trans_lock);
433 	cur_trans = fs_info->running_transaction;
434 	if (cur_trans && is_transaction_blocked(cur_trans)) {
435 		atomic_inc(&cur_trans->use_count);
436 		spin_unlock(&fs_info->trans_lock);
437 
438 		wait_event(fs_info->transaction_wait,
439 			   cur_trans->state >= TRANS_STATE_UNBLOCKED ||
440 			   cur_trans->aborted);
441 		btrfs_put_transaction(cur_trans);
442 	} else {
443 		spin_unlock(&fs_info->trans_lock);
444 	}
445 }
446 
447 static int may_wait_transaction(struct btrfs_fs_info *fs_info, int type)
448 {
449 	if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
450 		return 0;
451 
452 	if (type == TRANS_USERSPACE)
453 		return 1;
454 
455 	if (type == TRANS_START &&
456 	    !atomic_read(&fs_info->open_ioctl_trans))
457 		return 1;
458 
459 	return 0;
460 }
461 
462 static inline bool need_reserve_reloc_root(struct btrfs_root *root)
463 {
464 	struct btrfs_fs_info *fs_info = root->fs_info;
465 
466 	if (!fs_info->reloc_ctl ||
467 	    !test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
468 	    root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
469 	    root->reloc_root)
470 		return false;
471 
472 	return true;
473 }
474 
475 static struct btrfs_trans_handle *
476 start_transaction(struct btrfs_root *root, unsigned int num_items,
477 		  unsigned int type, enum btrfs_reserve_flush_enum flush)
478 {
479 	struct btrfs_fs_info *fs_info = root->fs_info;
480 
481 	struct btrfs_trans_handle *h;
482 	struct btrfs_transaction *cur_trans;
483 	u64 num_bytes = 0;
484 	u64 qgroup_reserved = 0;
485 	bool reloc_reserved = false;
486 	int ret;
487 
488 	/* Send isn't supposed to start transactions. */
489 	ASSERT(current->journal_info != BTRFS_SEND_TRANS_STUB);
490 
491 	if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
492 		return ERR_PTR(-EROFS);
493 
494 	if (current->journal_info) {
495 		WARN_ON(type & TRANS_EXTWRITERS);
496 		h = current->journal_info;
497 		h->use_count++;
498 		WARN_ON(h->use_count > 2);
499 		h->orig_rsv = h->block_rsv;
500 		h->block_rsv = NULL;
501 		goto got_it;
502 	}
503 
504 	/*
505 	 * Do the reservation before we join the transaction so we can do all
506 	 * the appropriate flushing if need be.
507 	 */
508 	if (num_items > 0 && root != fs_info->chunk_root) {
509 		qgroup_reserved = num_items * fs_info->nodesize;
510 		ret = btrfs_qgroup_reserve_meta(root, qgroup_reserved);
511 		if (ret)
512 			return ERR_PTR(ret);
513 
514 		num_bytes = btrfs_calc_trans_metadata_size(fs_info, num_items);
515 		/*
516 		 * Do the reservation for the relocation root creation
517 		 */
518 		if (need_reserve_reloc_root(root)) {
519 			num_bytes += fs_info->nodesize;
520 			reloc_reserved = true;
521 		}
522 
523 		ret = btrfs_block_rsv_add(root, &fs_info->trans_block_rsv,
524 					  num_bytes, flush);
525 		if (ret)
526 			goto reserve_fail;
527 	}
528 again:
529 	h = kmem_cache_zalloc(btrfs_trans_handle_cachep, GFP_NOFS);
530 	if (!h) {
531 		ret = -ENOMEM;
532 		goto alloc_fail;
533 	}
534 
535 	/*
536 	 * If we are JOIN_NOLOCK we're already committing a transaction and
537 	 * waiting on this guy, so we don't need to do the sb_start_intwrite
538 	 * because we're already holding a ref.  We need this because we could
539 	 * have raced in and did an fsync() on a file which can kick a commit
540 	 * and then we deadlock with somebody doing a freeze.
541 	 *
542 	 * If we are ATTACH, it means we just want to catch the current
543 	 * transaction and commit it, so we needn't do sb_start_intwrite().
544 	 */
545 	if (type & __TRANS_FREEZABLE)
546 		sb_start_intwrite(fs_info->sb);
547 
548 	if (may_wait_transaction(fs_info, type))
549 		wait_current_trans(fs_info);
550 
551 	do {
552 		ret = join_transaction(fs_info, type);
553 		if (ret == -EBUSY) {
554 			wait_current_trans(fs_info);
555 			if (unlikely(type == TRANS_ATTACH))
556 				ret = -ENOENT;
557 		}
558 	} while (ret == -EBUSY);
559 
560 	if (ret < 0)
561 		goto join_fail;
562 
563 	cur_trans = fs_info->running_transaction;
564 
565 	h->transid = cur_trans->transid;
566 	h->transaction = cur_trans;
567 	h->root = root;
568 	h->use_count = 1;
569 	h->fs_info = root->fs_info;
570 
571 	h->type = type;
572 	h->can_flush_pending_bgs = true;
573 	INIT_LIST_HEAD(&h->qgroup_ref_list);
574 	INIT_LIST_HEAD(&h->new_bgs);
575 
576 	smp_mb();
577 	if (cur_trans->state >= TRANS_STATE_BLOCKED &&
578 	    may_wait_transaction(fs_info, type)) {
579 		current->journal_info = h;
580 		btrfs_commit_transaction(h);
581 		goto again;
582 	}
583 
584 	if (num_bytes) {
585 		trace_btrfs_space_reservation(fs_info, "transaction",
586 					      h->transid, num_bytes, 1);
587 		h->block_rsv = &fs_info->trans_block_rsv;
588 		h->bytes_reserved = num_bytes;
589 		h->reloc_reserved = reloc_reserved;
590 	}
591 
592 got_it:
593 	btrfs_record_root_in_trans(h, root);
594 
595 	if (!current->journal_info && type != TRANS_USERSPACE)
596 		current->journal_info = h;
597 	return h;
598 
599 join_fail:
600 	if (type & __TRANS_FREEZABLE)
601 		sb_end_intwrite(fs_info->sb);
602 	kmem_cache_free(btrfs_trans_handle_cachep, h);
603 alloc_fail:
604 	if (num_bytes)
605 		btrfs_block_rsv_release(fs_info, &fs_info->trans_block_rsv,
606 					num_bytes);
607 reserve_fail:
608 	btrfs_qgroup_free_meta(root, qgroup_reserved);
609 	return ERR_PTR(ret);
610 }
611 
612 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
613 						   unsigned int num_items)
614 {
615 	return start_transaction(root, num_items, TRANS_START,
616 				 BTRFS_RESERVE_FLUSH_ALL);
617 }
618 struct btrfs_trans_handle *btrfs_start_transaction_fallback_global_rsv(
619 					struct btrfs_root *root,
620 					unsigned int num_items,
621 					int min_factor)
622 {
623 	struct btrfs_fs_info *fs_info = root->fs_info;
624 	struct btrfs_trans_handle *trans;
625 	u64 num_bytes;
626 	int ret;
627 
628 	trans = btrfs_start_transaction(root, num_items);
629 	if (!IS_ERR(trans) || PTR_ERR(trans) != -ENOSPC)
630 		return trans;
631 
632 	trans = btrfs_start_transaction(root, 0);
633 	if (IS_ERR(trans))
634 		return trans;
635 
636 	num_bytes = btrfs_calc_trans_metadata_size(fs_info, num_items);
637 	ret = btrfs_cond_migrate_bytes(fs_info, &fs_info->trans_block_rsv,
638 				       num_bytes, min_factor);
639 	if (ret) {
640 		btrfs_end_transaction(trans);
641 		return ERR_PTR(ret);
642 	}
643 
644 	trans->block_rsv = &fs_info->trans_block_rsv;
645 	trans->bytes_reserved = num_bytes;
646 	trace_btrfs_space_reservation(fs_info, "transaction",
647 				      trans->transid, num_bytes, 1);
648 
649 	return trans;
650 }
651 
652 struct btrfs_trans_handle *btrfs_start_transaction_lflush(
653 					struct btrfs_root *root,
654 					unsigned int num_items)
655 {
656 	return start_transaction(root, num_items, TRANS_START,
657 				 BTRFS_RESERVE_FLUSH_LIMIT);
658 }
659 
660 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
661 {
662 	return start_transaction(root, 0, TRANS_JOIN,
663 				 BTRFS_RESERVE_NO_FLUSH);
664 }
665 
666 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root)
667 {
668 	return start_transaction(root, 0, TRANS_JOIN_NOLOCK,
669 				 BTRFS_RESERVE_NO_FLUSH);
670 }
671 
672 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *root)
673 {
674 	return start_transaction(root, 0, TRANS_USERSPACE,
675 				 BTRFS_RESERVE_NO_FLUSH);
676 }
677 
678 /*
679  * btrfs_attach_transaction() - catch the running transaction
680  *
681  * It is used when we want to commit the current the transaction, but
682  * don't want to start a new one.
683  *
684  * Note: If this function return -ENOENT, it just means there is no
685  * running transaction. But it is possible that the inactive transaction
686  * is still in the memory, not fully on disk. If you hope there is no
687  * inactive transaction in the fs when -ENOENT is returned, you should
688  * invoke
689  *     btrfs_attach_transaction_barrier()
690  */
691 struct btrfs_trans_handle *btrfs_attach_transaction(struct btrfs_root *root)
692 {
693 	return start_transaction(root, 0, TRANS_ATTACH,
694 				 BTRFS_RESERVE_NO_FLUSH);
695 }
696 
697 /*
698  * btrfs_attach_transaction_barrier() - catch the running transaction
699  *
700  * It is similar to the above function, the differentia is this one
701  * will wait for all the inactive transactions until they fully
702  * complete.
703  */
704 struct btrfs_trans_handle *
705 btrfs_attach_transaction_barrier(struct btrfs_root *root)
706 {
707 	struct btrfs_trans_handle *trans;
708 
709 	trans = start_transaction(root, 0, TRANS_ATTACH,
710 				  BTRFS_RESERVE_NO_FLUSH);
711 	if (IS_ERR(trans) && PTR_ERR(trans) == -ENOENT)
712 		btrfs_wait_for_commit(root->fs_info, 0);
713 
714 	return trans;
715 }
716 
717 /* wait for a transaction commit to be fully complete */
718 static noinline void wait_for_commit(struct btrfs_transaction *commit)
719 {
720 	wait_event(commit->commit_wait, commit->state == TRANS_STATE_COMPLETED);
721 }
722 
723 int btrfs_wait_for_commit(struct btrfs_fs_info *fs_info, u64 transid)
724 {
725 	struct btrfs_transaction *cur_trans = NULL, *t;
726 	int ret = 0;
727 
728 	if (transid) {
729 		if (transid <= fs_info->last_trans_committed)
730 			goto out;
731 
732 		/* find specified transaction */
733 		spin_lock(&fs_info->trans_lock);
734 		list_for_each_entry(t, &fs_info->trans_list, list) {
735 			if (t->transid == transid) {
736 				cur_trans = t;
737 				atomic_inc(&cur_trans->use_count);
738 				ret = 0;
739 				break;
740 			}
741 			if (t->transid > transid) {
742 				ret = 0;
743 				break;
744 			}
745 		}
746 		spin_unlock(&fs_info->trans_lock);
747 
748 		/*
749 		 * The specified transaction doesn't exist, or we
750 		 * raced with btrfs_commit_transaction
751 		 */
752 		if (!cur_trans) {
753 			if (transid > fs_info->last_trans_committed)
754 				ret = -EINVAL;
755 			goto out;
756 		}
757 	} else {
758 		/* find newest transaction that is committing | committed */
759 		spin_lock(&fs_info->trans_lock);
760 		list_for_each_entry_reverse(t, &fs_info->trans_list,
761 					    list) {
762 			if (t->state >= TRANS_STATE_COMMIT_START) {
763 				if (t->state == TRANS_STATE_COMPLETED)
764 					break;
765 				cur_trans = t;
766 				atomic_inc(&cur_trans->use_count);
767 				break;
768 			}
769 		}
770 		spin_unlock(&fs_info->trans_lock);
771 		if (!cur_trans)
772 			goto out;  /* nothing committing|committed */
773 	}
774 
775 	wait_for_commit(cur_trans);
776 	btrfs_put_transaction(cur_trans);
777 out:
778 	return ret;
779 }
780 
781 void btrfs_throttle(struct btrfs_fs_info *fs_info)
782 {
783 	if (!atomic_read(&fs_info->open_ioctl_trans))
784 		wait_current_trans(fs_info);
785 }
786 
787 static int should_end_transaction(struct btrfs_trans_handle *trans)
788 {
789 	struct btrfs_fs_info *fs_info = trans->fs_info;
790 
791 	if (fs_info->global_block_rsv.space_info->full &&
792 	    btrfs_check_space_for_delayed_refs(trans, fs_info))
793 		return 1;
794 
795 	return !!btrfs_block_rsv_check(&fs_info->global_block_rsv, 5);
796 }
797 
798 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans)
799 {
800 	struct btrfs_transaction *cur_trans = trans->transaction;
801 	struct btrfs_fs_info *fs_info = trans->fs_info;
802 	int updates;
803 	int err;
804 
805 	smp_mb();
806 	if (cur_trans->state >= TRANS_STATE_BLOCKED ||
807 	    cur_trans->delayed_refs.flushing)
808 		return 1;
809 
810 	updates = trans->delayed_ref_updates;
811 	trans->delayed_ref_updates = 0;
812 	if (updates) {
813 		err = btrfs_run_delayed_refs(trans, fs_info, updates * 2);
814 		if (err) /* Error code will also eval true */
815 			return err;
816 	}
817 
818 	return should_end_transaction(trans);
819 }
820 
821 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
822 				   int throttle)
823 {
824 	struct btrfs_fs_info *info = trans->fs_info;
825 	struct btrfs_transaction *cur_trans = trans->transaction;
826 	u64 transid = trans->transid;
827 	unsigned long cur = trans->delayed_ref_updates;
828 	int lock = (trans->type != TRANS_JOIN_NOLOCK);
829 	int err = 0;
830 	int must_run_delayed_refs = 0;
831 
832 	if (trans->use_count > 1) {
833 		trans->use_count--;
834 		trans->block_rsv = trans->orig_rsv;
835 		return 0;
836 	}
837 
838 	btrfs_trans_release_metadata(trans, info);
839 	trans->block_rsv = NULL;
840 
841 	if (!list_empty(&trans->new_bgs))
842 		btrfs_create_pending_block_groups(trans, info);
843 
844 	trans->delayed_ref_updates = 0;
845 	if (!trans->sync) {
846 		must_run_delayed_refs =
847 			btrfs_should_throttle_delayed_refs(trans, info);
848 		cur = max_t(unsigned long, cur, 32);
849 
850 		/*
851 		 * don't make the caller wait if they are from a NOLOCK
852 		 * or ATTACH transaction, it will deadlock with commit
853 		 */
854 		if (must_run_delayed_refs == 1 &&
855 		    (trans->type & (__TRANS_JOIN_NOLOCK | __TRANS_ATTACH)))
856 			must_run_delayed_refs = 2;
857 	}
858 
859 	btrfs_trans_release_metadata(trans, info);
860 	trans->block_rsv = NULL;
861 
862 	if (!list_empty(&trans->new_bgs))
863 		btrfs_create_pending_block_groups(trans, info);
864 
865 	btrfs_trans_release_chunk_metadata(trans);
866 
867 	if (lock && !atomic_read(&info->open_ioctl_trans) &&
868 	    should_end_transaction(trans) &&
869 	    ACCESS_ONCE(cur_trans->state) == TRANS_STATE_RUNNING) {
870 		spin_lock(&info->trans_lock);
871 		if (cur_trans->state == TRANS_STATE_RUNNING)
872 			cur_trans->state = TRANS_STATE_BLOCKED;
873 		spin_unlock(&info->trans_lock);
874 	}
875 
876 	if (lock && ACCESS_ONCE(cur_trans->state) == TRANS_STATE_BLOCKED) {
877 		if (throttle)
878 			return btrfs_commit_transaction(trans);
879 		else
880 			wake_up_process(info->transaction_kthread);
881 	}
882 
883 	if (trans->type & __TRANS_FREEZABLE)
884 		sb_end_intwrite(info->sb);
885 
886 	WARN_ON(cur_trans != info->running_transaction);
887 	WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
888 	atomic_dec(&cur_trans->num_writers);
889 	extwriter_counter_dec(cur_trans, trans->type);
890 
891 	/*
892 	 * Make sure counter is updated before we wake up waiters.
893 	 */
894 	smp_mb();
895 	if (waitqueue_active(&cur_trans->writer_wait))
896 		wake_up(&cur_trans->writer_wait);
897 	btrfs_put_transaction(cur_trans);
898 
899 	if (current->journal_info == trans)
900 		current->journal_info = NULL;
901 
902 	if (throttle)
903 		btrfs_run_delayed_iputs(info);
904 
905 	if (trans->aborted ||
906 	    test_bit(BTRFS_FS_STATE_ERROR, &info->fs_state)) {
907 		wake_up_process(info->transaction_kthread);
908 		err = -EIO;
909 	}
910 	assert_qgroups_uptodate(trans);
911 
912 	kmem_cache_free(btrfs_trans_handle_cachep, trans);
913 	if (must_run_delayed_refs) {
914 		btrfs_async_run_delayed_refs(info, cur, transid,
915 					     must_run_delayed_refs == 1);
916 	}
917 	return err;
918 }
919 
920 int btrfs_end_transaction(struct btrfs_trans_handle *trans)
921 {
922 	return __btrfs_end_transaction(trans, 0);
923 }
924 
925 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans)
926 {
927 	return __btrfs_end_transaction(trans, 1);
928 }
929 
930 /*
931  * when btree blocks are allocated, they have some corresponding bits set for
932  * them in one of two extent_io trees.  This is used to make sure all of
933  * those extents are sent to disk but does not wait on them
934  */
935 int btrfs_write_marked_extents(struct btrfs_fs_info *fs_info,
936 			       struct extent_io_tree *dirty_pages, int mark)
937 {
938 	int err = 0;
939 	int werr = 0;
940 	struct address_space *mapping = fs_info->btree_inode->i_mapping;
941 	struct extent_state *cached_state = NULL;
942 	u64 start = 0;
943 	u64 end;
944 
945 	while (!find_first_extent_bit(dirty_pages, start, &start, &end,
946 				      mark, &cached_state)) {
947 		bool wait_writeback = false;
948 
949 		err = convert_extent_bit(dirty_pages, start, end,
950 					 EXTENT_NEED_WAIT,
951 					 mark, &cached_state);
952 		/*
953 		 * convert_extent_bit can return -ENOMEM, which is most of the
954 		 * time a temporary error. So when it happens, ignore the error
955 		 * and wait for writeback of this range to finish - because we
956 		 * failed to set the bit EXTENT_NEED_WAIT for the range, a call
957 		 * to __btrfs_wait_marked_extents() would not know that
958 		 * writeback for this range started and therefore wouldn't
959 		 * wait for it to finish - we don't want to commit a
960 		 * superblock that points to btree nodes/leafs for which
961 		 * writeback hasn't finished yet (and without errors).
962 		 * We cleanup any entries left in the io tree when committing
963 		 * the transaction (through clear_btree_io_tree()).
964 		 */
965 		if (err == -ENOMEM) {
966 			err = 0;
967 			wait_writeback = true;
968 		}
969 		if (!err)
970 			err = filemap_fdatawrite_range(mapping, start, end);
971 		if (err)
972 			werr = err;
973 		else if (wait_writeback)
974 			werr = filemap_fdatawait_range(mapping, start, end);
975 		free_extent_state(cached_state);
976 		cached_state = NULL;
977 		cond_resched();
978 		start = end + 1;
979 	}
980 	return werr;
981 }
982 
983 /*
984  * when btree blocks are allocated, they have some corresponding bits set for
985  * them in one of two extent_io trees.  This is used to make sure all of
986  * those extents are on disk for transaction or log commit.  We wait
987  * on all the pages and clear them from the dirty pages state tree
988  */
989 static int __btrfs_wait_marked_extents(struct btrfs_fs_info *fs_info,
990 				       struct extent_io_tree *dirty_pages)
991 {
992 	int err = 0;
993 	int werr = 0;
994 	struct address_space *mapping = fs_info->btree_inode->i_mapping;
995 	struct extent_state *cached_state = NULL;
996 	u64 start = 0;
997 	u64 end;
998 
999 	while (!find_first_extent_bit(dirty_pages, start, &start, &end,
1000 				      EXTENT_NEED_WAIT, &cached_state)) {
1001 		/*
1002 		 * Ignore -ENOMEM errors returned by clear_extent_bit().
1003 		 * When committing the transaction, we'll remove any entries
1004 		 * left in the io tree. For a log commit, we don't remove them
1005 		 * after committing the log because the tree can be accessed
1006 		 * concurrently - we do it only at transaction commit time when
1007 		 * it's safe to do it (through clear_btree_io_tree()).
1008 		 */
1009 		err = clear_extent_bit(dirty_pages, start, end,
1010 				       EXTENT_NEED_WAIT,
1011 				       0, 0, &cached_state, GFP_NOFS);
1012 		if (err == -ENOMEM)
1013 			err = 0;
1014 		if (!err)
1015 			err = filemap_fdatawait_range(mapping, start, end);
1016 		if (err)
1017 			werr = err;
1018 		free_extent_state(cached_state);
1019 		cached_state = NULL;
1020 		cond_resched();
1021 		start = end + 1;
1022 	}
1023 	if (err)
1024 		werr = err;
1025 	return werr;
1026 }
1027 
1028 int btrfs_wait_extents(struct btrfs_fs_info *fs_info,
1029 		       struct extent_io_tree *dirty_pages)
1030 {
1031 	bool errors = false;
1032 	int err;
1033 
1034 	err = __btrfs_wait_marked_extents(fs_info, dirty_pages);
1035 	if (test_and_clear_bit(BTRFS_FS_BTREE_ERR, &fs_info->flags))
1036 		errors = true;
1037 
1038 	if (errors && !err)
1039 		err = -EIO;
1040 	return err;
1041 }
1042 
1043 int btrfs_wait_tree_log_extents(struct btrfs_root *log_root, int mark)
1044 {
1045 	struct btrfs_fs_info *fs_info = log_root->fs_info;
1046 	struct extent_io_tree *dirty_pages = &log_root->dirty_log_pages;
1047 	bool errors = false;
1048 	int err;
1049 
1050 	ASSERT(log_root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
1051 
1052 	err = __btrfs_wait_marked_extents(fs_info, dirty_pages);
1053 	if ((mark & EXTENT_DIRTY) &&
1054 	    test_and_clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags))
1055 		errors = true;
1056 
1057 	if ((mark & EXTENT_NEW) &&
1058 	    test_and_clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags))
1059 		errors = true;
1060 
1061 	if (errors && !err)
1062 		err = -EIO;
1063 	return err;
1064 }
1065 
1066 /*
1067  * when btree blocks are allocated, they have some corresponding bits set for
1068  * them in one of two extent_io trees.  This is used to make sure all of
1069  * those extents are on disk for transaction or log commit
1070  */
1071 static int btrfs_write_and_wait_marked_extents(struct btrfs_fs_info *fs_info,
1072 				struct extent_io_tree *dirty_pages, int mark)
1073 {
1074 	int ret;
1075 	int ret2;
1076 	struct blk_plug plug;
1077 
1078 	blk_start_plug(&plug);
1079 	ret = btrfs_write_marked_extents(fs_info, dirty_pages, mark);
1080 	blk_finish_plug(&plug);
1081 	ret2 = btrfs_wait_extents(fs_info, dirty_pages);
1082 
1083 	if (ret)
1084 		return ret;
1085 	if (ret2)
1086 		return ret2;
1087 	return 0;
1088 }
1089 
1090 static int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
1091 					    struct btrfs_fs_info *fs_info)
1092 {
1093 	int ret;
1094 
1095 	ret = btrfs_write_and_wait_marked_extents(fs_info,
1096 					   &trans->transaction->dirty_pages,
1097 					   EXTENT_DIRTY);
1098 	clear_btree_io_tree(&trans->transaction->dirty_pages);
1099 
1100 	return ret;
1101 }
1102 
1103 /*
1104  * this is used to update the root pointer in the tree of tree roots.
1105  *
1106  * But, in the case of the extent allocation tree, updating the root
1107  * pointer may allocate blocks which may change the root of the extent
1108  * allocation tree.
1109  *
1110  * So, this loops and repeats and makes sure the cowonly root didn't
1111  * change while the root pointer was being updated in the metadata.
1112  */
1113 static int update_cowonly_root(struct btrfs_trans_handle *trans,
1114 			       struct btrfs_root *root)
1115 {
1116 	int ret;
1117 	u64 old_root_bytenr;
1118 	u64 old_root_used;
1119 	struct btrfs_fs_info *fs_info = root->fs_info;
1120 	struct btrfs_root *tree_root = fs_info->tree_root;
1121 
1122 	old_root_used = btrfs_root_used(&root->root_item);
1123 
1124 	while (1) {
1125 		old_root_bytenr = btrfs_root_bytenr(&root->root_item);
1126 		if (old_root_bytenr == root->node->start &&
1127 		    old_root_used == btrfs_root_used(&root->root_item))
1128 			break;
1129 
1130 		btrfs_set_root_node(&root->root_item, root->node);
1131 		ret = btrfs_update_root(trans, tree_root,
1132 					&root->root_key,
1133 					&root->root_item);
1134 		if (ret)
1135 			return ret;
1136 
1137 		old_root_used = btrfs_root_used(&root->root_item);
1138 	}
1139 
1140 	return 0;
1141 }
1142 
1143 /*
1144  * update all the cowonly tree roots on disk
1145  *
1146  * The error handling in this function may not be obvious. Any of the
1147  * failures will cause the file system to go offline. We still need
1148  * to clean up the delayed refs.
1149  */
1150 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
1151 					 struct btrfs_fs_info *fs_info)
1152 {
1153 	struct list_head *dirty_bgs = &trans->transaction->dirty_bgs;
1154 	struct list_head *io_bgs = &trans->transaction->io_bgs;
1155 	struct list_head *next;
1156 	struct extent_buffer *eb;
1157 	int ret;
1158 
1159 	eb = btrfs_lock_root_node(fs_info->tree_root);
1160 	ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL,
1161 			      0, &eb);
1162 	btrfs_tree_unlock(eb);
1163 	free_extent_buffer(eb);
1164 
1165 	if (ret)
1166 		return ret;
1167 
1168 	ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1169 	if (ret)
1170 		return ret;
1171 
1172 	ret = btrfs_run_dev_stats(trans, fs_info);
1173 	if (ret)
1174 		return ret;
1175 	ret = btrfs_run_dev_replace(trans, fs_info);
1176 	if (ret)
1177 		return ret;
1178 	ret = btrfs_run_qgroups(trans, fs_info);
1179 	if (ret)
1180 		return ret;
1181 
1182 	ret = btrfs_setup_space_cache(trans, fs_info);
1183 	if (ret)
1184 		return ret;
1185 
1186 	/* run_qgroups might have added some more refs */
1187 	ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1188 	if (ret)
1189 		return ret;
1190 again:
1191 	while (!list_empty(&fs_info->dirty_cowonly_roots)) {
1192 		struct btrfs_root *root;
1193 		next = fs_info->dirty_cowonly_roots.next;
1194 		list_del_init(next);
1195 		root = list_entry(next, struct btrfs_root, dirty_list);
1196 		clear_bit(BTRFS_ROOT_DIRTY, &root->state);
1197 
1198 		if (root != fs_info->extent_root)
1199 			list_add_tail(&root->dirty_list,
1200 				      &trans->transaction->switch_commits);
1201 		ret = update_cowonly_root(trans, root);
1202 		if (ret)
1203 			return ret;
1204 		ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1205 		if (ret)
1206 			return ret;
1207 	}
1208 
1209 	while (!list_empty(dirty_bgs) || !list_empty(io_bgs)) {
1210 		ret = btrfs_write_dirty_block_groups(trans, fs_info);
1211 		if (ret)
1212 			return ret;
1213 		ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1214 		if (ret)
1215 			return ret;
1216 	}
1217 
1218 	if (!list_empty(&fs_info->dirty_cowonly_roots))
1219 		goto again;
1220 
1221 	list_add_tail(&fs_info->extent_root->dirty_list,
1222 		      &trans->transaction->switch_commits);
1223 	btrfs_after_dev_replace_commit(fs_info);
1224 
1225 	return 0;
1226 }
1227 
1228 /*
1229  * dead roots are old snapshots that need to be deleted.  This allocates
1230  * a dirty root struct and adds it into the list of dead roots that need to
1231  * be deleted
1232  */
1233 void btrfs_add_dead_root(struct btrfs_root *root)
1234 {
1235 	struct btrfs_fs_info *fs_info = root->fs_info;
1236 
1237 	spin_lock(&fs_info->trans_lock);
1238 	if (list_empty(&root->root_list))
1239 		list_add_tail(&root->root_list, &fs_info->dead_roots);
1240 	spin_unlock(&fs_info->trans_lock);
1241 }
1242 
1243 /*
1244  * update all the cowonly tree roots on disk
1245  */
1246 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
1247 				    struct btrfs_fs_info *fs_info)
1248 {
1249 	struct btrfs_root *gang[8];
1250 	int i;
1251 	int ret;
1252 	int err = 0;
1253 
1254 	spin_lock(&fs_info->fs_roots_radix_lock);
1255 	while (1) {
1256 		ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
1257 						 (void **)gang, 0,
1258 						 ARRAY_SIZE(gang),
1259 						 BTRFS_ROOT_TRANS_TAG);
1260 		if (ret == 0)
1261 			break;
1262 		for (i = 0; i < ret; i++) {
1263 			struct btrfs_root *root = gang[i];
1264 			radix_tree_tag_clear(&fs_info->fs_roots_radix,
1265 					(unsigned long)root->root_key.objectid,
1266 					BTRFS_ROOT_TRANS_TAG);
1267 			spin_unlock(&fs_info->fs_roots_radix_lock);
1268 
1269 			btrfs_free_log(trans, root);
1270 			btrfs_update_reloc_root(trans, root);
1271 			btrfs_orphan_commit_root(trans, root);
1272 
1273 			btrfs_save_ino_cache(root, trans);
1274 
1275 			/* see comments in should_cow_block() */
1276 			clear_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1277 			smp_mb__after_atomic();
1278 
1279 			if (root->commit_root != root->node) {
1280 				list_add_tail(&root->dirty_list,
1281 					&trans->transaction->switch_commits);
1282 				btrfs_set_root_node(&root->root_item,
1283 						    root->node);
1284 			}
1285 
1286 			err = btrfs_update_root(trans, fs_info->tree_root,
1287 						&root->root_key,
1288 						&root->root_item);
1289 			spin_lock(&fs_info->fs_roots_radix_lock);
1290 			if (err)
1291 				break;
1292 			btrfs_qgroup_free_meta_all(root);
1293 		}
1294 	}
1295 	spin_unlock(&fs_info->fs_roots_radix_lock);
1296 	return err;
1297 }
1298 
1299 /*
1300  * defrag a given btree.
1301  * Every leaf in the btree is read and defragged.
1302  */
1303 int btrfs_defrag_root(struct btrfs_root *root)
1304 {
1305 	struct btrfs_fs_info *info = root->fs_info;
1306 	struct btrfs_trans_handle *trans;
1307 	int ret;
1308 
1309 	if (test_and_set_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state))
1310 		return 0;
1311 
1312 	while (1) {
1313 		trans = btrfs_start_transaction(root, 0);
1314 		if (IS_ERR(trans))
1315 			return PTR_ERR(trans);
1316 
1317 		ret = btrfs_defrag_leaves(trans, root);
1318 
1319 		btrfs_end_transaction(trans);
1320 		btrfs_btree_balance_dirty(info);
1321 		cond_resched();
1322 
1323 		if (btrfs_fs_closing(info) || ret != -EAGAIN)
1324 			break;
1325 
1326 		if (btrfs_defrag_cancelled(info)) {
1327 			btrfs_debug(info, "defrag_root cancelled");
1328 			ret = -EAGAIN;
1329 			break;
1330 		}
1331 	}
1332 	clear_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state);
1333 	return ret;
1334 }
1335 
1336 /*
1337  * Do all special snapshot related qgroup dirty hack.
1338  *
1339  * Will do all needed qgroup inherit and dirty hack like switch commit
1340  * roots inside one transaction and write all btree into disk, to make
1341  * qgroup works.
1342  */
1343 static int qgroup_account_snapshot(struct btrfs_trans_handle *trans,
1344 				   struct btrfs_root *src,
1345 				   struct btrfs_root *parent,
1346 				   struct btrfs_qgroup_inherit *inherit,
1347 				   u64 dst_objectid)
1348 {
1349 	struct btrfs_fs_info *fs_info = src->fs_info;
1350 	int ret;
1351 
1352 	/*
1353 	 * Save some performance in the case that qgroups are not
1354 	 * enabled. If this check races with the ioctl, rescan will
1355 	 * kick in anyway.
1356 	 */
1357 	mutex_lock(&fs_info->qgroup_ioctl_lock);
1358 	if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)) {
1359 		mutex_unlock(&fs_info->qgroup_ioctl_lock);
1360 		return 0;
1361 	}
1362 	mutex_unlock(&fs_info->qgroup_ioctl_lock);
1363 
1364 	/*
1365 	 * We are going to commit transaction, see btrfs_commit_transaction()
1366 	 * comment for reason locking tree_log_mutex
1367 	 */
1368 	mutex_lock(&fs_info->tree_log_mutex);
1369 
1370 	ret = commit_fs_roots(trans, fs_info);
1371 	if (ret)
1372 		goto out;
1373 	ret = btrfs_qgroup_prepare_account_extents(trans, fs_info);
1374 	if (ret < 0)
1375 		goto out;
1376 	ret = btrfs_qgroup_account_extents(trans, fs_info);
1377 	if (ret < 0)
1378 		goto out;
1379 
1380 	/* Now qgroup are all updated, we can inherit it to new qgroups */
1381 	ret = btrfs_qgroup_inherit(trans, fs_info,
1382 				   src->root_key.objectid, dst_objectid,
1383 				   inherit);
1384 	if (ret < 0)
1385 		goto out;
1386 
1387 	/*
1388 	 * Now we do a simplified commit transaction, which will:
1389 	 * 1) commit all subvolume and extent tree
1390 	 *    To ensure all subvolume and extent tree have a valid
1391 	 *    commit_root to accounting later insert_dir_item()
1392 	 * 2) write all btree blocks onto disk
1393 	 *    This is to make sure later btree modification will be cowed
1394 	 *    Or commit_root can be populated and cause wrong qgroup numbers
1395 	 * In this simplified commit, we don't really care about other trees
1396 	 * like chunk and root tree, as they won't affect qgroup.
1397 	 * And we don't write super to avoid half committed status.
1398 	 */
1399 	ret = commit_cowonly_roots(trans, fs_info);
1400 	if (ret)
1401 		goto out;
1402 	switch_commit_roots(trans->transaction, fs_info);
1403 	ret = btrfs_write_and_wait_transaction(trans, fs_info);
1404 	if (ret)
1405 		btrfs_handle_fs_error(fs_info, ret,
1406 			"Error while writing out transaction for qgroup");
1407 
1408 out:
1409 	mutex_unlock(&fs_info->tree_log_mutex);
1410 
1411 	/*
1412 	 * Force parent root to be updated, as we recorded it before so its
1413 	 * last_trans == cur_transid.
1414 	 * Or it won't be committed again onto disk after later
1415 	 * insert_dir_item()
1416 	 */
1417 	if (!ret)
1418 		record_root_in_trans(trans, parent, 1);
1419 	return ret;
1420 }
1421 
1422 /*
1423  * new snapshots need to be created at a very specific time in the
1424  * transaction commit.  This does the actual creation.
1425  *
1426  * Note:
1427  * If the error which may affect the commitment of the current transaction
1428  * happens, we should return the error number. If the error which just affect
1429  * the creation of the pending snapshots, just return 0.
1430  */
1431 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
1432 				   struct btrfs_fs_info *fs_info,
1433 				   struct btrfs_pending_snapshot *pending)
1434 {
1435 	struct btrfs_key key;
1436 	struct btrfs_root_item *new_root_item;
1437 	struct btrfs_root *tree_root = fs_info->tree_root;
1438 	struct btrfs_root *root = pending->root;
1439 	struct btrfs_root *parent_root;
1440 	struct btrfs_block_rsv *rsv;
1441 	struct inode *parent_inode;
1442 	struct btrfs_path *path;
1443 	struct btrfs_dir_item *dir_item;
1444 	struct dentry *dentry;
1445 	struct extent_buffer *tmp;
1446 	struct extent_buffer *old;
1447 	struct timespec cur_time;
1448 	int ret = 0;
1449 	u64 to_reserve = 0;
1450 	u64 index = 0;
1451 	u64 objectid;
1452 	u64 root_flags;
1453 	uuid_le new_uuid;
1454 
1455 	ASSERT(pending->path);
1456 	path = pending->path;
1457 
1458 	ASSERT(pending->root_item);
1459 	new_root_item = pending->root_item;
1460 
1461 	pending->error = btrfs_find_free_objectid(tree_root, &objectid);
1462 	if (pending->error)
1463 		goto no_free_objectid;
1464 
1465 	/*
1466 	 * Make qgroup to skip current new snapshot's qgroupid, as it is
1467 	 * accounted by later btrfs_qgroup_inherit().
1468 	 */
1469 	btrfs_set_skip_qgroup(trans, objectid);
1470 
1471 	btrfs_reloc_pre_snapshot(pending, &to_reserve);
1472 
1473 	if (to_reserve > 0) {
1474 		pending->error = btrfs_block_rsv_add(root,
1475 						     &pending->block_rsv,
1476 						     to_reserve,
1477 						     BTRFS_RESERVE_NO_FLUSH);
1478 		if (pending->error)
1479 			goto clear_skip_qgroup;
1480 	}
1481 
1482 	key.objectid = objectid;
1483 	key.offset = (u64)-1;
1484 	key.type = BTRFS_ROOT_ITEM_KEY;
1485 
1486 	rsv = trans->block_rsv;
1487 	trans->block_rsv = &pending->block_rsv;
1488 	trans->bytes_reserved = trans->block_rsv->reserved;
1489 	trace_btrfs_space_reservation(fs_info, "transaction",
1490 				      trans->transid,
1491 				      trans->bytes_reserved, 1);
1492 	dentry = pending->dentry;
1493 	parent_inode = pending->dir;
1494 	parent_root = BTRFS_I(parent_inode)->root;
1495 	record_root_in_trans(trans, parent_root, 0);
1496 
1497 	cur_time = current_time(parent_inode);
1498 
1499 	/*
1500 	 * insert the directory item
1501 	 */
1502 	ret = btrfs_set_inode_index(parent_inode, &index);
1503 	BUG_ON(ret); /* -ENOMEM */
1504 
1505 	/* check if there is a file/dir which has the same name. */
1506 	dir_item = btrfs_lookup_dir_item(NULL, parent_root, path,
1507 					 btrfs_ino(parent_inode),
1508 					 dentry->d_name.name,
1509 					 dentry->d_name.len, 0);
1510 	if (dir_item != NULL && !IS_ERR(dir_item)) {
1511 		pending->error = -EEXIST;
1512 		goto dir_item_existed;
1513 	} else if (IS_ERR(dir_item)) {
1514 		ret = PTR_ERR(dir_item);
1515 		btrfs_abort_transaction(trans, ret);
1516 		goto fail;
1517 	}
1518 	btrfs_release_path(path);
1519 
1520 	/*
1521 	 * pull in the delayed directory update
1522 	 * and the delayed inode item
1523 	 * otherwise we corrupt the FS during
1524 	 * snapshot
1525 	 */
1526 	ret = btrfs_run_delayed_items(trans, fs_info);
1527 	if (ret) {	/* Transaction aborted */
1528 		btrfs_abort_transaction(trans, ret);
1529 		goto fail;
1530 	}
1531 
1532 	record_root_in_trans(trans, root, 0);
1533 	btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
1534 	memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
1535 	btrfs_check_and_init_root_item(new_root_item);
1536 
1537 	root_flags = btrfs_root_flags(new_root_item);
1538 	if (pending->readonly)
1539 		root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
1540 	else
1541 		root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
1542 	btrfs_set_root_flags(new_root_item, root_flags);
1543 
1544 	btrfs_set_root_generation_v2(new_root_item,
1545 			trans->transid);
1546 	uuid_le_gen(&new_uuid);
1547 	memcpy(new_root_item->uuid, new_uuid.b, BTRFS_UUID_SIZE);
1548 	memcpy(new_root_item->parent_uuid, root->root_item.uuid,
1549 			BTRFS_UUID_SIZE);
1550 	if (!(root_flags & BTRFS_ROOT_SUBVOL_RDONLY)) {
1551 		memset(new_root_item->received_uuid, 0,
1552 		       sizeof(new_root_item->received_uuid));
1553 		memset(&new_root_item->stime, 0, sizeof(new_root_item->stime));
1554 		memset(&new_root_item->rtime, 0, sizeof(new_root_item->rtime));
1555 		btrfs_set_root_stransid(new_root_item, 0);
1556 		btrfs_set_root_rtransid(new_root_item, 0);
1557 	}
1558 	btrfs_set_stack_timespec_sec(&new_root_item->otime, cur_time.tv_sec);
1559 	btrfs_set_stack_timespec_nsec(&new_root_item->otime, cur_time.tv_nsec);
1560 	btrfs_set_root_otransid(new_root_item, trans->transid);
1561 
1562 	old = btrfs_lock_root_node(root);
1563 	ret = btrfs_cow_block(trans, root, old, NULL, 0, &old);
1564 	if (ret) {
1565 		btrfs_tree_unlock(old);
1566 		free_extent_buffer(old);
1567 		btrfs_abort_transaction(trans, ret);
1568 		goto fail;
1569 	}
1570 
1571 	btrfs_set_lock_blocking(old);
1572 
1573 	ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1574 	/* clean up in any case */
1575 	btrfs_tree_unlock(old);
1576 	free_extent_buffer(old);
1577 	if (ret) {
1578 		btrfs_abort_transaction(trans, ret);
1579 		goto fail;
1580 	}
1581 	/* see comments in should_cow_block() */
1582 	set_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1583 	smp_wmb();
1584 
1585 	btrfs_set_root_node(new_root_item, tmp);
1586 	/* record when the snapshot was created in key.offset */
1587 	key.offset = trans->transid;
1588 	ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1589 	btrfs_tree_unlock(tmp);
1590 	free_extent_buffer(tmp);
1591 	if (ret) {
1592 		btrfs_abort_transaction(trans, ret);
1593 		goto fail;
1594 	}
1595 
1596 	/*
1597 	 * insert root back/forward references
1598 	 */
1599 	ret = btrfs_add_root_ref(trans, fs_info, objectid,
1600 				 parent_root->root_key.objectid,
1601 				 btrfs_ino(parent_inode), index,
1602 				 dentry->d_name.name, dentry->d_name.len);
1603 	if (ret) {
1604 		btrfs_abort_transaction(trans, ret);
1605 		goto fail;
1606 	}
1607 
1608 	key.offset = (u64)-1;
1609 	pending->snap = btrfs_read_fs_root_no_name(fs_info, &key);
1610 	if (IS_ERR(pending->snap)) {
1611 		ret = PTR_ERR(pending->snap);
1612 		btrfs_abort_transaction(trans, ret);
1613 		goto fail;
1614 	}
1615 
1616 	ret = btrfs_reloc_post_snapshot(trans, pending);
1617 	if (ret) {
1618 		btrfs_abort_transaction(trans, ret);
1619 		goto fail;
1620 	}
1621 
1622 	ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1623 	if (ret) {
1624 		btrfs_abort_transaction(trans, ret);
1625 		goto fail;
1626 	}
1627 
1628 	/*
1629 	 * Do special qgroup accounting for snapshot, as we do some qgroup
1630 	 * snapshot hack to do fast snapshot.
1631 	 * To co-operate with that hack, we do hack again.
1632 	 * Or snapshot will be greatly slowed down by a subtree qgroup rescan
1633 	 */
1634 	ret = qgroup_account_snapshot(trans, root, parent_root,
1635 				      pending->inherit, objectid);
1636 	if (ret < 0)
1637 		goto fail;
1638 
1639 	ret = btrfs_insert_dir_item(trans, parent_root,
1640 				    dentry->d_name.name, dentry->d_name.len,
1641 				    parent_inode, &key,
1642 				    BTRFS_FT_DIR, index);
1643 	/* We have check then name at the beginning, so it is impossible. */
1644 	BUG_ON(ret == -EEXIST || ret == -EOVERFLOW);
1645 	if (ret) {
1646 		btrfs_abort_transaction(trans, ret);
1647 		goto fail;
1648 	}
1649 
1650 	btrfs_i_size_write(parent_inode, parent_inode->i_size +
1651 					 dentry->d_name.len * 2);
1652 	parent_inode->i_mtime = parent_inode->i_ctime =
1653 		current_time(parent_inode);
1654 	ret = btrfs_update_inode_fallback(trans, parent_root, parent_inode);
1655 	if (ret) {
1656 		btrfs_abort_transaction(trans, ret);
1657 		goto fail;
1658 	}
1659 	ret = btrfs_uuid_tree_add(trans, fs_info, new_uuid.b,
1660 				  BTRFS_UUID_KEY_SUBVOL, objectid);
1661 	if (ret) {
1662 		btrfs_abort_transaction(trans, ret);
1663 		goto fail;
1664 	}
1665 	if (!btrfs_is_empty_uuid(new_root_item->received_uuid)) {
1666 		ret = btrfs_uuid_tree_add(trans, fs_info,
1667 					  new_root_item->received_uuid,
1668 					  BTRFS_UUID_KEY_RECEIVED_SUBVOL,
1669 					  objectid);
1670 		if (ret && ret != -EEXIST) {
1671 			btrfs_abort_transaction(trans, ret);
1672 			goto fail;
1673 		}
1674 	}
1675 
1676 	ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1677 	if (ret) {
1678 		btrfs_abort_transaction(trans, ret);
1679 		goto fail;
1680 	}
1681 
1682 fail:
1683 	pending->error = ret;
1684 dir_item_existed:
1685 	trans->block_rsv = rsv;
1686 	trans->bytes_reserved = 0;
1687 clear_skip_qgroup:
1688 	btrfs_clear_skip_qgroup(trans);
1689 no_free_objectid:
1690 	kfree(new_root_item);
1691 	pending->root_item = NULL;
1692 	btrfs_free_path(path);
1693 	pending->path = NULL;
1694 
1695 	return ret;
1696 }
1697 
1698 /*
1699  * create all the snapshots we've scheduled for creation
1700  */
1701 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
1702 					     struct btrfs_fs_info *fs_info)
1703 {
1704 	struct btrfs_pending_snapshot *pending, *next;
1705 	struct list_head *head = &trans->transaction->pending_snapshots;
1706 	int ret = 0;
1707 
1708 	list_for_each_entry_safe(pending, next, head, list) {
1709 		list_del(&pending->list);
1710 		ret = create_pending_snapshot(trans, fs_info, pending);
1711 		if (ret)
1712 			break;
1713 	}
1714 	return ret;
1715 }
1716 
1717 static void update_super_roots(struct btrfs_fs_info *fs_info)
1718 {
1719 	struct btrfs_root_item *root_item;
1720 	struct btrfs_super_block *super;
1721 
1722 	super = fs_info->super_copy;
1723 
1724 	root_item = &fs_info->chunk_root->root_item;
1725 	super->chunk_root = root_item->bytenr;
1726 	super->chunk_root_generation = root_item->generation;
1727 	super->chunk_root_level = root_item->level;
1728 
1729 	root_item = &fs_info->tree_root->root_item;
1730 	super->root = root_item->bytenr;
1731 	super->generation = root_item->generation;
1732 	super->root_level = root_item->level;
1733 	if (btrfs_test_opt(fs_info, SPACE_CACHE))
1734 		super->cache_generation = root_item->generation;
1735 	if (test_bit(BTRFS_FS_UPDATE_UUID_TREE_GEN, &fs_info->flags))
1736 		super->uuid_tree_generation = root_item->generation;
1737 }
1738 
1739 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1740 {
1741 	struct btrfs_transaction *trans;
1742 	int ret = 0;
1743 
1744 	spin_lock(&info->trans_lock);
1745 	trans = info->running_transaction;
1746 	if (trans)
1747 		ret = (trans->state >= TRANS_STATE_COMMIT_START);
1748 	spin_unlock(&info->trans_lock);
1749 	return ret;
1750 }
1751 
1752 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1753 {
1754 	struct btrfs_transaction *trans;
1755 	int ret = 0;
1756 
1757 	spin_lock(&info->trans_lock);
1758 	trans = info->running_transaction;
1759 	if (trans)
1760 		ret = is_transaction_blocked(trans);
1761 	spin_unlock(&info->trans_lock);
1762 	return ret;
1763 }
1764 
1765 /*
1766  * wait for the current transaction commit to start and block subsequent
1767  * transaction joins
1768  */
1769 static void wait_current_trans_commit_start(struct btrfs_fs_info *fs_info,
1770 					    struct btrfs_transaction *trans)
1771 {
1772 	wait_event(fs_info->transaction_blocked_wait,
1773 		   trans->state >= TRANS_STATE_COMMIT_START || trans->aborted);
1774 }
1775 
1776 /*
1777  * wait for the current transaction to start and then become unblocked.
1778  * caller holds ref.
1779  */
1780 static void wait_current_trans_commit_start_and_unblock(
1781 					struct btrfs_fs_info *fs_info,
1782 					struct btrfs_transaction *trans)
1783 {
1784 	wait_event(fs_info->transaction_wait,
1785 		   trans->state >= TRANS_STATE_UNBLOCKED || trans->aborted);
1786 }
1787 
1788 /*
1789  * commit transactions asynchronously. once btrfs_commit_transaction_async
1790  * returns, any subsequent transaction will not be allowed to join.
1791  */
1792 struct btrfs_async_commit {
1793 	struct btrfs_trans_handle *newtrans;
1794 	struct work_struct work;
1795 };
1796 
1797 static void do_async_commit(struct work_struct *work)
1798 {
1799 	struct btrfs_async_commit *ac =
1800 		container_of(work, struct btrfs_async_commit, work);
1801 
1802 	/*
1803 	 * We've got freeze protection passed with the transaction.
1804 	 * Tell lockdep about it.
1805 	 */
1806 	if (ac->newtrans->type & __TRANS_FREEZABLE)
1807 		__sb_writers_acquired(ac->newtrans->fs_info->sb, SB_FREEZE_FS);
1808 
1809 	current->journal_info = ac->newtrans;
1810 
1811 	btrfs_commit_transaction(ac->newtrans);
1812 	kfree(ac);
1813 }
1814 
1815 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1816 				   int wait_for_unblock)
1817 {
1818 	struct btrfs_fs_info *fs_info = trans->fs_info;
1819 	struct btrfs_async_commit *ac;
1820 	struct btrfs_transaction *cur_trans;
1821 
1822 	ac = kmalloc(sizeof(*ac), GFP_NOFS);
1823 	if (!ac)
1824 		return -ENOMEM;
1825 
1826 	INIT_WORK(&ac->work, do_async_commit);
1827 	ac->newtrans = btrfs_join_transaction(trans->root);
1828 	if (IS_ERR(ac->newtrans)) {
1829 		int err = PTR_ERR(ac->newtrans);
1830 		kfree(ac);
1831 		return err;
1832 	}
1833 
1834 	/* take transaction reference */
1835 	cur_trans = trans->transaction;
1836 	atomic_inc(&cur_trans->use_count);
1837 
1838 	btrfs_end_transaction(trans);
1839 
1840 	/*
1841 	 * Tell lockdep we've released the freeze rwsem, since the
1842 	 * async commit thread will be the one to unlock it.
1843 	 */
1844 	if (ac->newtrans->type & __TRANS_FREEZABLE)
1845 		__sb_writers_release(fs_info->sb, SB_FREEZE_FS);
1846 
1847 	schedule_work(&ac->work);
1848 
1849 	/* wait for transaction to start and unblock */
1850 	if (wait_for_unblock)
1851 		wait_current_trans_commit_start_and_unblock(fs_info, cur_trans);
1852 	else
1853 		wait_current_trans_commit_start(fs_info, cur_trans);
1854 
1855 	if (current->journal_info == trans)
1856 		current->journal_info = NULL;
1857 
1858 	btrfs_put_transaction(cur_trans);
1859 	return 0;
1860 }
1861 
1862 
1863 static void cleanup_transaction(struct btrfs_trans_handle *trans,
1864 				struct btrfs_root *root, int err)
1865 {
1866 	struct btrfs_fs_info *fs_info = root->fs_info;
1867 	struct btrfs_transaction *cur_trans = trans->transaction;
1868 	DEFINE_WAIT(wait);
1869 
1870 	WARN_ON(trans->use_count > 1);
1871 
1872 	btrfs_abort_transaction(trans, err);
1873 
1874 	spin_lock(&fs_info->trans_lock);
1875 
1876 	/*
1877 	 * If the transaction is removed from the list, it means this
1878 	 * transaction has been committed successfully, so it is impossible
1879 	 * to call the cleanup function.
1880 	 */
1881 	BUG_ON(list_empty(&cur_trans->list));
1882 
1883 	list_del_init(&cur_trans->list);
1884 	if (cur_trans == fs_info->running_transaction) {
1885 		cur_trans->state = TRANS_STATE_COMMIT_DOING;
1886 		spin_unlock(&fs_info->trans_lock);
1887 		wait_event(cur_trans->writer_wait,
1888 			   atomic_read(&cur_trans->num_writers) == 1);
1889 
1890 		spin_lock(&fs_info->trans_lock);
1891 	}
1892 	spin_unlock(&fs_info->trans_lock);
1893 
1894 	btrfs_cleanup_one_transaction(trans->transaction, fs_info);
1895 
1896 	spin_lock(&fs_info->trans_lock);
1897 	if (cur_trans == fs_info->running_transaction)
1898 		fs_info->running_transaction = NULL;
1899 	spin_unlock(&fs_info->trans_lock);
1900 
1901 	if (trans->type & __TRANS_FREEZABLE)
1902 		sb_end_intwrite(fs_info->sb);
1903 	btrfs_put_transaction(cur_trans);
1904 	btrfs_put_transaction(cur_trans);
1905 
1906 	trace_btrfs_transaction_commit(root);
1907 
1908 	if (current->journal_info == trans)
1909 		current->journal_info = NULL;
1910 	btrfs_scrub_cancel(fs_info);
1911 
1912 	kmem_cache_free(btrfs_trans_handle_cachep, trans);
1913 }
1914 
1915 static inline int btrfs_start_delalloc_flush(struct btrfs_fs_info *fs_info)
1916 {
1917 	if (btrfs_test_opt(fs_info, FLUSHONCOMMIT))
1918 		return btrfs_start_delalloc_roots(fs_info, 1, -1);
1919 	return 0;
1920 }
1921 
1922 static inline void btrfs_wait_delalloc_flush(struct btrfs_fs_info *fs_info)
1923 {
1924 	if (btrfs_test_opt(fs_info, FLUSHONCOMMIT))
1925 		btrfs_wait_ordered_roots(fs_info, -1, 0, (u64)-1);
1926 }
1927 
1928 static inline void
1929 btrfs_wait_pending_ordered(struct btrfs_transaction *cur_trans)
1930 {
1931 	wait_event(cur_trans->pending_wait,
1932 		   atomic_read(&cur_trans->pending_ordered) == 0);
1933 }
1934 
1935 int btrfs_commit_transaction(struct btrfs_trans_handle *trans)
1936 {
1937 	struct btrfs_fs_info *fs_info = trans->fs_info;
1938 	struct btrfs_transaction *cur_trans = trans->transaction;
1939 	struct btrfs_transaction *prev_trans = NULL;
1940 	int ret;
1941 
1942 	/* Stop the commit early if ->aborted is set */
1943 	if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
1944 		ret = cur_trans->aborted;
1945 		btrfs_end_transaction(trans);
1946 		return ret;
1947 	}
1948 
1949 	/* make a pass through all the delayed refs we have so far
1950 	 * any runnings procs may add more while we are here
1951 	 */
1952 	ret = btrfs_run_delayed_refs(trans, fs_info, 0);
1953 	if (ret) {
1954 		btrfs_end_transaction(trans);
1955 		return ret;
1956 	}
1957 
1958 	btrfs_trans_release_metadata(trans, fs_info);
1959 	trans->block_rsv = NULL;
1960 
1961 	cur_trans = trans->transaction;
1962 
1963 	/*
1964 	 * set the flushing flag so procs in this transaction have to
1965 	 * start sending their work down.
1966 	 */
1967 	cur_trans->delayed_refs.flushing = 1;
1968 	smp_wmb();
1969 
1970 	if (!list_empty(&trans->new_bgs))
1971 		btrfs_create_pending_block_groups(trans, fs_info);
1972 
1973 	ret = btrfs_run_delayed_refs(trans, fs_info, 0);
1974 	if (ret) {
1975 		btrfs_end_transaction(trans);
1976 		return ret;
1977 	}
1978 
1979 	if (!test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &cur_trans->flags)) {
1980 		int run_it = 0;
1981 
1982 		/* this mutex is also taken before trying to set
1983 		 * block groups readonly.  We need to make sure
1984 		 * that nobody has set a block group readonly
1985 		 * after a extents from that block group have been
1986 		 * allocated for cache files.  btrfs_set_block_group_ro
1987 		 * will wait for the transaction to commit if it
1988 		 * finds BTRFS_TRANS_DIRTY_BG_RUN set.
1989 		 *
1990 		 * The BTRFS_TRANS_DIRTY_BG_RUN flag is also used to make sure
1991 		 * only one process starts all the block group IO.  It wouldn't
1992 		 * hurt to have more than one go through, but there's no
1993 		 * real advantage to it either.
1994 		 */
1995 		mutex_lock(&fs_info->ro_block_group_mutex);
1996 		if (!test_and_set_bit(BTRFS_TRANS_DIRTY_BG_RUN,
1997 				      &cur_trans->flags))
1998 			run_it = 1;
1999 		mutex_unlock(&fs_info->ro_block_group_mutex);
2000 
2001 		if (run_it)
2002 			ret = btrfs_start_dirty_block_groups(trans, fs_info);
2003 	}
2004 	if (ret) {
2005 		btrfs_end_transaction(trans);
2006 		return ret;
2007 	}
2008 
2009 	spin_lock(&fs_info->trans_lock);
2010 	if (cur_trans->state >= TRANS_STATE_COMMIT_START) {
2011 		spin_unlock(&fs_info->trans_lock);
2012 		atomic_inc(&cur_trans->use_count);
2013 		ret = btrfs_end_transaction(trans);
2014 
2015 		wait_for_commit(cur_trans);
2016 
2017 		if (unlikely(cur_trans->aborted))
2018 			ret = cur_trans->aborted;
2019 
2020 		btrfs_put_transaction(cur_trans);
2021 
2022 		return ret;
2023 	}
2024 
2025 	cur_trans->state = TRANS_STATE_COMMIT_START;
2026 	wake_up(&fs_info->transaction_blocked_wait);
2027 
2028 	if (cur_trans->list.prev != &fs_info->trans_list) {
2029 		prev_trans = list_entry(cur_trans->list.prev,
2030 					struct btrfs_transaction, list);
2031 		if (prev_trans->state != TRANS_STATE_COMPLETED) {
2032 			atomic_inc(&prev_trans->use_count);
2033 			spin_unlock(&fs_info->trans_lock);
2034 
2035 			wait_for_commit(prev_trans);
2036 			ret = prev_trans->aborted;
2037 
2038 			btrfs_put_transaction(prev_trans);
2039 			if (ret)
2040 				goto cleanup_transaction;
2041 		} else {
2042 			spin_unlock(&fs_info->trans_lock);
2043 		}
2044 	} else {
2045 		spin_unlock(&fs_info->trans_lock);
2046 	}
2047 
2048 	extwriter_counter_dec(cur_trans, trans->type);
2049 
2050 	ret = btrfs_start_delalloc_flush(fs_info);
2051 	if (ret)
2052 		goto cleanup_transaction;
2053 
2054 	ret = btrfs_run_delayed_items(trans, fs_info);
2055 	if (ret)
2056 		goto cleanup_transaction;
2057 
2058 	wait_event(cur_trans->writer_wait,
2059 		   extwriter_counter_read(cur_trans) == 0);
2060 
2061 	/* some pending stuffs might be added after the previous flush. */
2062 	ret = btrfs_run_delayed_items(trans, fs_info);
2063 	if (ret)
2064 		goto cleanup_transaction;
2065 
2066 	btrfs_wait_delalloc_flush(fs_info);
2067 
2068 	btrfs_wait_pending_ordered(cur_trans);
2069 
2070 	btrfs_scrub_pause(fs_info);
2071 	/*
2072 	 * Ok now we need to make sure to block out any other joins while we
2073 	 * commit the transaction.  We could have started a join before setting
2074 	 * COMMIT_DOING so make sure to wait for num_writers to == 1 again.
2075 	 */
2076 	spin_lock(&fs_info->trans_lock);
2077 	cur_trans->state = TRANS_STATE_COMMIT_DOING;
2078 	spin_unlock(&fs_info->trans_lock);
2079 	wait_event(cur_trans->writer_wait,
2080 		   atomic_read(&cur_trans->num_writers) == 1);
2081 
2082 	/* ->aborted might be set after the previous check, so check it */
2083 	if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
2084 		ret = cur_trans->aborted;
2085 		goto scrub_continue;
2086 	}
2087 	/*
2088 	 * the reloc mutex makes sure that we stop
2089 	 * the balancing code from coming in and moving
2090 	 * extents around in the middle of the commit
2091 	 */
2092 	mutex_lock(&fs_info->reloc_mutex);
2093 
2094 	/*
2095 	 * We needn't worry about the delayed items because we will
2096 	 * deal with them in create_pending_snapshot(), which is the
2097 	 * core function of the snapshot creation.
2098 	 */
2099 	ret = create_pending_snapshots(trans, fs_info);
2100 	if (ret) {
2101 		mutex_unlock(&fs_info->reloc_mutex);
2102 		goto scrub_continue;
2103 	}
2104 
2105 	/*
2106 	 * We insert the dir indexes of the snapshots and update the inode
2107 	 * of the snapshots' parents after the snapshot creation, so there
2108 	 * are some delayed items which are not dealt with. Now deal with
2109 	 * them.
2110 	 *
2111 	 * We needn't worry that this operation will corrupt the snapshots,
2112 	 * because all the tree which are snapshoted will be forced to COW
2113 	 * the nodes and leaves.
2114 	 */
2115 	ret = btrfs_run_delayed_items(trans, fs_info);
2116 	if (ret) {
2117 		mutex_unlock(&fs_info->reloc_mutex);
2118 		goto scrub_continue;
2119 	}
2120 
2121 	ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
2122 	if (ret) {
2123 		mutex_unlock(&fs_info->reloc_mutex);
2124 		goto scrub_continue;
2125 	}
2126 
2127 	/* Reocrd old roots for later qgroup accounting */
2128 	ret = btrfs_qgroup_prepare_account_extents(trans, fs_info);
2129 	if (ret) {
2130 		mutex_unlock(&fs_info->reloc_mutex);
2131 		goto scrub_continue;
2132 	}
2133 
2134 	/*
2135 	 * make sure none of the code above managed to slip in a
2136 	 * delayed item
2137 	 */
2138 	btrfs_assert_delayed_root_empty(fs_info);
2139 
2140 	WARN_ON(cur_trans != trans->transaction);
2141 
2142 	/* btrfs_commit_tree_roots is responsible for getting the
2143 	 * various roots consistent with each other.  Every pointer
2144 	 * in the tree of tree roots has to point to the most up to date
2145 	 * root for every subvolume and other tree.  So, we have to keep
2146 	 * the tree logging code from jumping in and changing any
2147 	 * of the trees.
2148 	 *
2149 	 * At this point in the commit, there can't be any tree-log
2150 	 * writers, but a little lower down we drop the trans mutex
2151 	 * and let new people in.  By holding the tree_log_mutex
2152 	 * from now until after the super is written, we avoid races
2153 	 * with the tree-log code.
2154 	 */
2155 	mutex_lock(&fs_info->tree_log_mutex);
2156 
2157 	ret = commit_fs_roots(trans, fs_info);
2158 	if (ret) {
2159 		mutex_unlock(&fs_info->tree_log_mutex);
2160 		mutex_unlock(&fs_info->reloc_mutex);
2161 		goto scrub_continue;
2162 	}
2163 
2164 	/*
2165 	 * Since the transaction is done, we can apply the pending changes
2166 	 * before the next transaction.
2167 	 */
2168 	btrfs_apply_pending_changes(fs_info);
2169 
2170 	/* commit_fs_roots gets rid of all the tree log roots, it is now
2171 	 * safe to free the root of tree log roots
2172 	 */
2173 	btrfs_free_log_root_tree(trans, fs_info);
2174 
2175 	/*
2176 	 * Since fs roots are all committed, we can get a quite accurate
2177 	 * new_roots. So let's do quota accounting.
2178 	 */
2179 	ret = btrfs_qgroup_account_extents(trans, fs_info);
2180 	if (ret < 0) {
2181 		mutex_unlock(&fs_info->tree_log_mutex);
2182 		mutex_unlock(&fs_info->reloc_mutex);
2183 		goto scrub_continue;
2184 	}
2185 
2186 	ret = commit_cowonly_roots(trans, fs_info);
2187 	if (ret) {
2188 		mutex_unlock(&fs_info->tree_log_mutex);
2189 		mutex_unlock(&fs_info->reloc_mutex);
2190 		goto scrub_continue;
2191 	}
2192 
2193 	/*
2194 	 * The tasks which save the space cache and inode cache may also
2195 	 * update ->aborted, check it.
2196 	 */
2197 	if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
2198 		ret = cur_trans->aborted;
2199 		mutex_unlock(&fs_info->tree_log_mutex);
2200 		mutex_unlock(&fs_info->reloc_mutex);
2201 		goto scrub_continue;
2202 	}
2203 
2204 	btrfs_prepare_extent_commit(trans, fs_info);
2205 
2206 	cur_trans = fs_info->running_transaction;
2207 
2208 	btrfs_set_root_node(&fs_info->tree_root->root_item,
2209 			    fs_info->tree_root->node);
2210 	list_add_tail(&fs_info->tree_root->dirty_list,
2211 		      &cur_trans->switch_commits);
2212 
2213 	btrfs_set_root_node(&fs_info->chunk_root->root_item,
2214 			    fs_info->chunk_root->node);
2215 	list_add_tail(&fs_info->chunk_root->dirty_list,
2216 		      &cur_trans->switch_commits);
2217 
2218 	switch_commit_roots(cur_trans, fs_info);
2219 
2220 	assert_qgroups_uptodate(trans);
2221 	ASSERT(list_empty(&cur_trans->dirty_bgs));
2222 	ASSERT(list_empty(&cur_trans->io_bgs));
2223 	update_super_roots(fs_info);
2224 
2225 	btrfs_set_super_log_root(fs_info->super_copy, 0);
2226 	btrfs_set_super_log_root_level(fs_info->super_copy, 0);
2227 	memcpy(fs_info->super_for_commit, fs_info->super_copy,
2228 	       sizeof(*fs_info->super_copy));
2229 
2230 	btrfs_update_commit_device_size(fs_info);
2231 	btrfs_update_commit_device_bytes_used(fs_info, cur_trans);
2232 
2233 	clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags);
2234 	clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags);
2235 
2236 	btrfs_trans_release_chunk_metadata(trans);
2237 
2238 	spin_lock(&fs_info->trans_lock);
2239 	cur_trans->state = TRANS_STATE_UNBLOCKED;
2240 	fs_info->running_transaction = NULL;
2241 	spin_unlock(&fs_info->trans_lock);
2242 	mutex_unlock(&fs_info->reloc_mutex);
2243 
2244 	wake_up(&fs_info->transaction_wait);
2245 
2246 	ret = btrfs_write_and_wait_transaction(trans, fs_info);
2247 	if (ret) {
2248 		btrfs_handle_fs_error(fs_info, ret,
2249 				      "Error while writing out transaction");
2250 		mutex_unlock(&fs_info->tree_log_mutex);
2251 		goto scrub_continue;
2252 	}
2253 
2254 	ret = write_ctree_super(trans, fs_info, 0);
2255 	if (ret) {
2256 		mutex_unlock(&fs_info->tree_log_mutex);
2257 		goto scrub_continue;
2258 	}
2259 
2260 	/*
2261 	 * the super is written, we can safely allow the tree-loggers
2262 	 * to go about their business
2263 	 */
2264 	mutex_unlock(&fs_info->tree_log_mutex);
2265 
2266 	btrfs_finish_extent_commit(trans, fs_info);
2267 
2268 	if (test_bit(BTRFS_TRANS_HAVE_FREE_BGS, &cur_trans->flags))
2269 		btrfs_clear_space_info_full(fs_info);
2270 
2271 	fs_info->last_trans_committed = cur_trans->transid;
2272 	/*
2273 	 * We needn't acquire the lock here because there is no other task
2274 	 * which can change it.
2275 	 */
2276 	cur_trans->state = TRANS_STATE_COMPLETED;
2277 	wake_up(&cur_trans->commit_wait);
2278 
2279 	spin_lock(&fs_info->trans_lock);
2280 	list_del_init(&cur_trans->list);
2281 	spin_unlock(&fs_info->trans_lock);
2282 
2283 	btrfs_put_transaction(cur_trans);
2284 	btrfs_put_transaction(cur_trans);
2285 
2286 	if (trans->type & __TRANS_FREEZABLE)
2287 		sb_end_intwrite(fs_info->sb);
2288 
2289 	trace_btrfs_transaction_commit(trans->root);
2290 
2291 	btrfs_scrub_continue(fs_info);
2292 
2293 	if (current->journal_info == trans)
2294 		current->journal_info = NULL;
2295 
2296 	kmem_cache_free(btrfs_trans_handle_cachep, trans);
2297 
2298 	/*
2299 	 * If fs has been frozen, we can not handle delayed iputs, otherwise
2300 	 * it'll result in deadlock about SB_FREEZE_FS.
2301 	 */
2302 	if (current != fs_info->transaction_kthread &&
2303 	    current != fs_info->cleaner_kthread && !fs_info->fs_frozen)
2304 		btrfs_run_delayed_iputs(fs_info);
2305 
2306 	return ret;
2307 
2308 scrub_continue:
2309 	btrfs_scrub_continue(fs_info);
2310 cleanup_transaction:
2311 	btrfs_trans_release_metadata(trans, fs_info);
2312 	btrfs_trans_release_chunk_metadata(trans);
2313 	trans->block_rsv = NULL;
2314 	btrfs_warn(fs_info, "Skipping commit of aborted transaction.");
2315 	if (current->journal_info == trans)
2316 		current->journal_info = NULL;
2317 	cleanup_transaction(trans, trans->root, ret);
2318 
2319 	return ret;
2320 }
2321 
2322 /*
2323  * return < 0 if error
2324  * 0 if there are no more dead_roots at the time of call
2325  * 1 there are more to be processed, call me again
2326  *
2327  * The return value indicates there are certainly more snapshots to delete, but
2328  * if there comes a new one during processing, it may return 0. We don't mind,
2329  * because btrfs_commit_super will poke cleaner thread and it will process it a
2330  * few seconds later.
2331  */
2332 int btrfs_clean_one_deleted_snapshot(struct btrfs_root *root)
2333 {
2334 	int ret;
2335 	struct btrfs_fs_info *fs_info = root->fs_info;
2336 
2337 	spin_lock(&fs_info->trans_lock);
2338 	if (list_empty(&fs_info->dead_roots)) {
2339 		spin_unlock(&fs_info->trans_lock);
2340 		return 0;
2341 	}
2342 	root = list_first_entry(&fs_info->dead_roots,
2343 			struct btrfs_root, root_list);
2344 	list_del_init(&root->root_list);
2345 	spin_unlock(&fs_info->trans_lock);
2346 
2347 	btrfs_debug(fs_info, "cleaner removing %llu", root->objectid);
2348 
2349 	btrfs_kill_all_delayed_nodes(root);
2350 
2351 	if (btrfs_header_backref_rev(root->node) <
2352 			BTRFS_MIXED_BACKREF_REV)
2353 		ret = btrfs_drop_snapshot(root, NULL, 0, 0);
2354 	else
2355 		ret = btrfs_drop_snapshot(root, NULL, 1, 0);
2356 
2357 	return (ret < 0) ? 0 : 1;
2358 }
2359 
2360 void btrfs_apply_pending_changes(struct btrfs_fs_info *fs_info)
2361 {
2362 	unsigned long prev;
2363 	unsigned long bit;
2364 
2365 	prev = xchg(&fs_info->pending_changes, 0);
2366 	if (!prev)
2367 		return;
2368 
2369 	bit = 1 << BTRFS_PENDING_SET_INODE_MAP_CACHE;
2370 	if (prev & bit)
2371 		btrfs_set_opt(fs_info->mount_opt, INODE_MAP_CACHE);
2372 	prev &= ~bit;
2373 
2374 	bit = 1 << BTRFS_PENDING_CLEAR_INODE_MAP_CACHE;
2375 	if (prev & bit)
2376 		btrfs_clear_opt(fs_info->mount_opt, INODE_MAP_CACHE);
2377 	prev &= ~bit;
2378 
2379 	bit = 1 << BTRFS_PENDING_COMMIT;
2380 	if (prev & bit)
2381 		btrfs_debug(fs_info, "pending commit done");
2382 	prev &= ~bit;
2383 
2384 	if (prev)
2385 		btrfs_warn(fs_info,
2386 			"unknown pending changes left 0x%lx, ignoring", prev);
2387 }
2388