xref: /openbmc/linux/fs/btrfs/transaction.c (revision 42bc47b3)
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/writeback.h>
10 #include <linux/pagemap.h>
11 #include <linux/blkdev.h>
12 #include <linux/uuid.h>
13 #include "ctree.h"
14 #include "disk-io.h"
15 #include "transaction.h"
16 #include "locking.h"
17 #include "tree-log.h"
18 #include "inode-map.h"
19 #include "volumes.h"
20 #include "dev-replace.h"
21 #include "qgroup.h"
22 
23 #define BTRFS_ROOT_TRANS_TAG 0
24 
25 static const unsigned int btrfs_blocked_trans_types[TRANS_STATE_MAX] = {
26 	[TRANS_STATE_RUNNING]		= 0U,
27 	[TRANS_STATE_BLOCKED]		=  __TRANS_START,
28 	[TRANS_STATE_COMMIT_START]	= (__TRANS_START | __TRANS_ATTACH),
29 	[TRANS_STATE_COMMIT_DOING]	= (__TRANS_START |
30 					   __TRANS_ATTACH |
31 					   __TRANS_JOIN),
32 	[TRANS_STATE_UNBLOCKED]		= (__TRANS_START |
33 					   __TRANS_ATTACH |
34 					   __TRANS_JOIN |
35 					   __TRANS_JOIN_NOLOCK),
36 	[TRANS_STATE_COMPLETED]		= (__TRANS_START |
37 					   __TRANS_ATTACH |
38 					   __TRANS_JOIN |
39 					   __TRANS_JOIN_NOLOCK),
40 };
41 
42 void btrfs_put_transaction(struct btrfs_transaction *transaction)
43 {
44 	WARN_ON(refcount_read(&transaction->use_count) == 0);
45 	if (refcount_dec_and_test(&transaction->use_count)) {
46 		BUG_ON(!list_empty(&transaction->list));
47 		WARN_ON(!RB_EMPTY_ROOT(&transaction->delayed_refs.href_root));
48 		if (transaction->delayed_refs.pending_csums)
49 			btrfs_err(transaction->fs_info,
50 				  "pending csums is %llu",
51 				  transaction->delayed_refs.pending_csums);
52 		while (!list_empty(&transaction->pending_chunks)) {
53 			struct extent_map *em;
54 
55 			em = list_first_entry(&transaction->pending_chunks,
56 					      struct extent_map, list);
57 			list_del_init(&em->list);
58 			free_extent_map(em);
59 		}
60 		/*
61 		 * If any block groups are found in ->deleted_bgs then it's
62 		 * because the transaction was aborted and a commit did not
63 		 * happen (things failed before writing the new superblock
64 		 * and calling btrfs_finish_extent_commit()), so we can not
65 		 * discard the physical locations of the block groups.
66 		 */
67 		while (!list_empty(&transaction->deleted_bgs)) {
68 			struct btrfs_block_group_cache *cache;
69 
70 			cache = list_first_entry(&transaction->deleted_bgs,
71 						 struct btrfs_block_group_cache,
72 						 bg_list);
73 			list_del_init(&cache->bg_list);
74 			btrfs_put_block_group_trimming(cache);
75 			btrfs_put_block_group(cache);
76 		}
77 		kfree(transaction);
78 	}
79 }
80 
81 static void clear_btree_io_tree(struct extent_io_tree *tree)
82 {
83 	spin_lock(&tree->lock);
84 	/*
85 	 * Do a single barrier for the waitqueue_active check here, the state
86 	 * of the waitqueue should not change once clear_btree_io_tree is
87 	 * called.
88 	 */
89 	smp_mb();
90 	while (!RB_EMPTY_ROOT(&tree->state)) {
91 		struct rb_node *node;
92 		struct extent_state *state;
93 
94 		node = rb_first(&tree->state);
95 		state = rb_entry(node, struct extent_state, rb_node);
96 		rb_erase(&state->rb_node, &tree->state);
97 		RB_CLEAR_NODE(&state->rb_node);
98 		/*
99 		 * btree io trees aren't supposed to have tasks waiting for
100 		 * changes in the flags of extent states ever.
101 		 */
102 		ASSERT(!waitqueue_active(&state->wq));
103 		free_extent_state(state);
104 
105 		cond_resched_lock(&tree->lock);
106 	}
107 	spin_unlock(&tree->lock);
108 }
109 
110 static noinline void switch_commit_roots(struct btrfs_transaction *trans)
111 {
112 	struct btrfs_fs_info *fs_info = trans->fs_info;
113 	struct btrfs_root *root, *tmp;
114 
115 	down_write(&fs_info->commit_root_sem);
116 	list_for_each_entry_safe(root, tmp, &trans->switch_commits,
117 				 dirty_list) {
118 		list_del_init(&root->dirty_list);
119 		free_extent_buffer(root->commit_root);
120 		root->commit_root = btrfs_root_node(root);
121 		if (is_fstree(root->objectid))
122 			btrfs_unpin_free_ino(root);
123 		clear_btree_io_tree(&root->dirty_log_pages);
124 	}
125 
126 	/* We can free old roots now. */
127 	spin_lock(&trans->dropped_roots_lock);
128 	while (!list_empty(&trans->dropped_roots)) {
129 		root = list_first_entry(&trans->dropped_roots,
130 					struct btrfs_root, root_list);
131 		list_del_init(&root->root_list);
132 		spin_unlock(&trans->dropped_roots_lock);
133 		btrfs_drop_and_free_fs_root(fs_info, root);
134 		spin_lock(&trans->dropped_roots_lock);
135 	}
136 	spin_unlock(&trans->dropped_roots_lock);
137 	up_write(&fs_info->commit_root_sem);
138 }
139 
140 static inline void extwriter_counter_inc(struct btrfs_transaction *trans,
141 					 unsigned int type)
142 {
143 	if (type & TRANS_EXTWRITERS)
144 		atomic_inc(&trans->num_extwriters);
145 }
146 
147 static inline void extwriter_counter_dec(struct btrfs_transaction *trans,
148 					 unsigned int type)
149 {
150 	if (type & TRANS_EXTWRITERS)
151 		atomic_dec(&trans->num_extwriters);
152 }
153 
154 static inline void extwriter_counter_init(struct btrfs_transaction *trans,
155 					  unsigned int type)
156 {
157 	atomic_set(&trans->num_extwriters, ((type & TRANS_EXTWRITERS) ? 1 : 0));
158 }
159 
160 static inline int extwriter_counter_read(struct btrfs_transaction *trans)
161 {
162 	return atomic_read(&trans->num_extwriters);
163 }
164 
165 /*
166  * either allocate a new transaction or hop into the existing one
167  */
168 static noinline int join_transaction(struct btrfs_fs_info *fs_info,
169 				     unsigned int type)
170 {
171 	struct btrfs_transaction *cur_trans;
172 
173 	spin_lock(&fs_info->trans_lock);
174 loop:
175 	/* The file system has been taken offline. No new transactions. */
176 	if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
177 		spin_unlock(&fs_info->trans_lock);
178 		return -EROFS;
179 	}
180 
181 	cur_trans = fs_info->running_transaction;
182 	if (cur_trans) {
183 		if (cur_trans->aborted) {
184 			spin_unlock(&fs_info->trans_lock);
185 			return cur_trans->aborted;
186 		}
187 		if (btrfs_blocked_trans_types[cur_trans->state] & type) {
188 			spin_unlock(&fs_info->trans_lock);
189 			return -EBUSY;
190 		}
191 		refcount_inc(&cur_trans->use_count);
192 		atomic_inc(&cur_trans->num_writers);
193 		extwriter_counter_inc(cur_trans, type);
194 		spin_unlock(&fs_info->trans_lock);
195 		return 0;
196 	}
197 	spin_unlock(&fs_info->trans_lock);
198 
199 	/*
200 	 * If we are ATTACH, we just want to catch the current transaction,
201 	 * and commit it. If there is no transaction, just return ENOENT.
202 	 */
203 	if (type == TRANS_ATTACH)
204 		return -ENOENT;
205 
206 	/*
207 	 * JOIN_NOLOCK only happens during the transaction commit, so
208 	 * it is impossible that ->running_transaction is NULL
209 	 */
210 	BUG_ON(type == TRANS_JOIN_NOLOCK);
211 
212 	cur_trans = kmalloc(sizeof(*cur_trans), GFP_NOFS);
213 	if (!cur_trans)
214 		return -ENOMEM;
215 
216 	spin_lock(&fs_info->trans_lock);
217 	if (fs_info->running_transaction) {
218 		/*
219 		 * someone started a transaction after we unlocked.  Make sure
220 		 * to redo the checks above
221 		 */
222 		kfree(cur_trans);
223 		goto loop;
224 	} else if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
225 		spin_unlock(&fs_info->trans_lock);
226 		kfree(cur_trans);
227 		return -EROFS;
228 	}
229 
230 	cur_trans->fs_info = fs_info;
231 	atomic_set(&cur_trans->num_writers, 1);
232 	extwriter_counter_init(cur_trans, type);
233 	init_waitqueue_head(&cur_trans->writer_wait);
234 	init_waitqueue_head(&cur_trans->commit_wait);
235 	init_waitqueue_head(&cur_trans->pending_wait);
236 	cur_trans->state = TRANS_STATE_RUNNING;
237 	/*
238 	 * One for this trans handle, one so it will live on until we
239 	 * commit the transaction.
240 	 */
241 	refcount_set(&cur_trans->use_count, 2);
242 	atomic_set(&cur_trans->pending_ordered, 0);
243 	cur_trans->flags = 0;
244 	cur_trans->start_time = get_seconds();
245 
246 	memset(&cur_trans->delayed_refs, 0, sizeof(cur_trans->delayed_refs));
247 
248 	cur_trans->delayed_refs.href_root = RB_ROOT;
249 	cur_trans->delayed_refs.dirty_extent_root = RB_ROOT;
250 	atomic_set(&cur_trans->delayed_refs.num_entries, 0);
251 
252 	/*
253 	 * although the tree mod log is per file system and not per transaction,
254 	 * the log must never go across transaction boundaries.
255 	 */
256 	smp_mb();
257 	if (!list_empty(&fs_info->tree_mod_seq_list))
258 		WARN(1, KERN_ERR "BTRFS: tree_mod_seq_list not empty when creating a fresh transaction\n");
259 	if (!RB_EMPTY_ROOT(&fs_info->tree_mod_log))
260 		WARN(1, KERN_ERR "BTRFS: tree_mod_log rb tree not empty when creating a fresh transaction\n");
261 	atomic64_set(&fs_info->tree_mod_seq, 0);
262 
263 	spin_lock_init(&cur_trans->delayed_refs.lock);
264 
265 	INIT_LIST_HEAD(&cur_trans->pending_snapshots);
266 	INIT_LIST_HEAD(&cur_trans->pending_chunks);
267 	INIT_LIST_HEAD(&cur_trans->switch_commits);
268 	INIT_LIST_HEAD(&cur_trans->dirty_bgs);
269 	INIT_LIST_HEAD(&cur_trans->io_bgs);
270 	INIT_LIST_HEAD(&cur_trans->dropped_roots);
271 	mutex_init(&cur_trans->cache_write_mutex);
272 	cur_trans->num_dirty_bgs = 0;
273 	spin_lock_init(&cur_trans->dirty_bgs_lock);
274 	INIT_LIST_HEAD(&cur_trans->deleted_bgs);
275 	spin_lock_init(&cur_trans->dropped_roots_lock);
276 	list_add_tail(&cur_trans->list, &fs_info->trans_list);
277 	extent_io_tree_init(&cur_trans->dirty_pages,
278 			     fs_info->btree_inode);
279 	fs_info->generation++;
280 	cur_trans->transid = fs_info->generation;
281 	fs_info->running_transaction = cur_trans;
282 	cur_trans->aborted = 0;
283 	spin_unlock(&fs_info->trans_lock);
284 
285 	return 0;
286 }
287 
288 /*
289  * this does all the record keeping required to make sure that a reference
290  * counted root is properly recorded in a given transaction.  This is required
291  * to make sure the old root from before we joined the transaction is deleted
292  * when the transaction commits
293  */
294 static int record_root_in_trans(struct btrfs_trans_handle *trans,
295 			       struct btrfs_root *root,
296 			       int force)
297 {
298 	struct btrfs_fs_info *fs_info = root->fs_info;
299 
300 	if ((test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
301 	    root->last_trans < trans->transid) || force) {
302 		WARN_ON(root == fs_info->extent_root);
303 		WARN_ON(!force && root->commit_root != root->node);
304 
305 		/*
306 		 * see below for IN_TRANS_SETUP usage rules
307 		 * we have the reloc mutex held now, so there
308 		 * is only one writer in this function
309 		 */
310 		set_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
311 
312 		/* make sure readers find IN_TRANS_SETUP before
313 		 * they find our root->last_trans update
314 		 */
315 		smp_wmb();
316 
317 		spin_lock(&fs_info->fs_roots_radix_lock);
318 		if (root->last_trans == trans->transid && !force) {
319 			spin_unlock(&fs_info->fs_roots_radix_lock);
320 			return 0;
321 		}
322 		radix_tree_tag_set(&fs_info->fs_roots_radix,
323 				   (unsigned long)root->root_key.objectid,
324 				   BTRFS_ROOT_TRANS_TAG);
325 		spin_unlock(&fs_info->fs_roots_radix_lock);
326 		root->last_trans = trans->transid;
327 
328 		/* this is pretty tricky.  We don't want to
329 		 * take the relocation lock in btrfs_record_root_in_trans
330 		 * unless we're really doing the first setup for this root in
331 		 * this transaction.
332 		 *
333 		 * Normally we'd use root->last_trans as a flag to decide
334 		 * if we want to take the expensive mutex.
335 		 *
336 		 * But, we have to set root->last_trans before we
337 		 * init the relocation root, otherwise, we trip over warnings
338 		 * in ctree.c.  The solution used here is to flag ourselves
339 		 * with root IN_TRANS_SETUP.  When this is 1, we're still
340 		 * fixing up the reloc trees and everyone must wait.
341 		 *
342 		 * When this is zero, they can trust root->last_trans and fly
343 		 * through btrfs_record_root_in_trans without having to take the
344 		 * lock.  smp_wmb() makes sure that all the writes above are
345 		 * done before we pop in the zero below
346 		 */
347 		btrfs_init_reloc_root(trans, root);
348 		smp_mb__before_atomic();
349 		clear_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
350 	}
351 	return 0;
352 }
353 
354 
355 void btrfs_add_dropped_root(struct btrfs_trans_handle *trans,
356 			    struct btrfs_root *root)
357 {
358 	struct btrfs_fs_info *fs_info = root->fs_info;
359 	struct btrfs_transaction *cur_trans = trans->transaction;
360 
361 	/* Add ourselves to the transaction dropped list */
362 	spin_lock(&cur_trans->dropped_roots_lock);
363 	list_add_tail(&root->root_list, &cur_trans->dropped_roots);
364 	spin_unlock(&cur_trans->dropped_roots_lock);
365 
366 	/* Make sure we don't try to update the root at commit time */
367 	spin_lock(&fs_info->fs_roots_radix_lock);
368 	radix_tree_tag_clear(&fs_info->fs_roots_radix,
369 			     (unsigned long)root->root_key.objectid,
370 			     BTRFS_ROOT_TRANS_TAG);
371 	spin_unlock(&fs_info->fs_roots_radix_lock);
372 }
373 
374 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
375 			       struct btrfs_root *root)
376 {
377 	struct btrfs_fs_info *fs_info = root->fs_info;
378 
379 	if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
380 		return 0;
381 
382 	/*
383 	 * see record_root_in_trans for comments about IN_TRANS_SETUP usage
384 	 * and barriers
385 	 */
386 	smp_rmb();
387 	if (root->last_trans == trans->transid &&
388 	    !test_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state))
389 		return 0;
390 
391 	mutex_lock(&fs_info->reloc_mutex);
392 	record_root_in_trans(trans, root, 0);
393 	mutex_unlock(&fs_info->reloc_mutex);
394 
395 	return 0;
396 }
397 
398 static inline int is_transaction_blocked(struct btrfs_transaction *trans)
399 {
400 	return (trans->state >= TRANS_STATE_BLOCKED &&
401 		trans->state < TRANS_STATE_UNBLOCKED &&
402 		!trans->aborted);
403 }
404 
405 /* wait for commit against the current transaction to become unblocked
406  * when this is done, it is safe to start a new transaction, but the current
407  * transaction might not be fully on disk.
408  */
409 static void wait_current_trans(struct btrfs_fs_info *fs_info)
410 {
411 	struct btrfs_transaction *cur_trans;
412 
413 	spin_lock(&fs_info->trans_lock);
414 	cur_trans = fs_info->running_transaction;
415 	if (cur_trans && is_transaction_blocked(cur_trans)) {
416 		refcount_inc(&cur_trans->use_count);
417 		spin_unlock(&fs_info->trans_lock);
418 
419 		wait_event(fs_info->transaction_wait,
420 			   cur_trans->state >= TRANS_STATE_UNBLOCKED ||
421 			   cur_trans->aborted);
422 		btrfs_put_transaction(cur_trans);
423 	} else {
424 		spin_unlock(&fs_info->trans_lock);
425 	}
426 }
427 
428 static int may_wait_transaction(struct btrfs_fs_info *fs_info, int type)
429 {
430 	if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
431 		return 0;
432 
433 	if (type == TRANS_START)
434 		return 1;
435 
436 	return 0;
437 }
438 
439 static inline bool need_reserve_reloc_root(struct btrfs_root *root)
440 {
441 	struct btrfs_fs_info *fs_info = root->fs_info;
442 
443 	if (!fs_info->reloc_ctl ||
444 	    !test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
445 	    root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
446 	    root->reloc_root)
447 		return false;
448 
449 	return true;
450 }
451 
452 static struct btrfs_trans_handle *
453 start_transaction(struct btrfs_root *root, unsigned int num_items,
454 		  unsigned int type, enum btrfs_reserve_flush_enum flush,
455 		  bool enforce_qgroups)
456 {
457 	struct btrfs_fs_info *fs_info = root->fs_info;
458 
459 	struct btrfs_trans_handle *h;
460 	struct btrfs_transaction *cur_trans;
461 	u64 num_bytes = 0;
462 	u64 qgroup_reserved = 0;
463 	bool reloc_reserved = false;
464 	int ret;
465 
466 	/* Send isn't supposed to start transactions. */
467 	ASSERT(current->journal_info != BTRFS_SEND_TRANS_STUB);
468 
469 	if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
470 		return ERR_PTR(-EROFS);
471 
472 	if (current->journal_info) {
473 		WARN_ON(type & TRANS_EXTWRITERS);
474 		h = current->journal_info;
475 		refcount_inc(&h->use_count);
476 		WARN_ON(refcount_read(&h->use_count) > 2);
477 		h->orig_rsv = h->block_rsv;
478 		h->block_rsv = NULL;
479 		goto got_it;
480 	}
481 
482 	/*
483 	 * Do the reservation before we join the transaction so we can do all
484 	 * the appropriate flushing if need be.
485 	 */
486 	if (num_items && root != fs_info->chunk_root) {
487 		qgroup_reserved = num_items * fs_info->nodesize;
488 		ret = btrfs_qgroup_reserve_meta_pertrans(root, qgroup_reserved,
489 				enforce_qgroups);
490 		if (ret)
491 			return ERR_PTR(ret);
492 
493 		num_bytes = btrfs_calc_trans_metadata_size(fs_info, num_items);
494 		/*
495 		 * Do the reservation for the relocation root creation
496 		 */
497 		if (need_reserve_reloc_root(root)) {
498 			num_bytes += fs_info->nodesize;
499 			reloc_reserved = true;
500 		}
501 
502 		ret = btrfs_block_rsv_add(root, &fs_info->trans_block_rsv,
503 					  num_bytes, flush);
504 		if (ret)
505 			goto reserve_fail;
506 	}
507 again:
508 	h = kmem_cache_zalloc(btrfs_trans_handle_cachep, GFP_NOFS);
509 	if (!h) {
510 		ret = -ENOMEM;
511 		goto alloc_fail;
512 	}
513 
514 	/*
515 	 * If we are JOIN_NOLOCK we're already committing a transaction and
516 	 * waiting on this guy, so we don't need to do the sb_start_intwrite
517 	 * because we're already holding a ref.  We need this because we could
518 	 * have raced in and did an fsync() on a file which can kick a commit
519 	 * and then we deadlock with somebody doing a freeze.
520 	 *
521 	 * If we are ATTACH, it means we just want to catch the current
522 	 * transaction and commit it, so we needn't do sb_start_intwrite().
523 	 */
524 	if (type & __TRANS_FREEZABLE)
525 		sb_start_intwrite(fs_info->sb);
526 
527 	if (may_wait_transaction(fs_info, type))
528 		wait_current_trans(fs_info);
529 
530 	do {
531 		ret = join_transaction(fs_info, type);
532 		if (ret == -EBUSY) {
533 			wait_current_trans(fs_info);
534 			if (unlikely(type == TRANS_ATTACH))
535 				ret = -ENOENT;
536 		}
537 	} while (ret == -EBUSY);
538 
539 	if (ret < 0)
540 		goto join_fail;
541 
542 	cur_trans = fs_info->running_transaction;
543 
544 	h->transid = cur_trans->transid;
545 	h->transaction = cur_trans;
546 	h->root = root;
547 	refcount_set(&h->use_count, 1);
548 	h->fs_info = root->fs_info;
549 
550 	h->type = type;
551 	h->can_flush_pending_bgs = true;
552 	INIT_LIST_HEAD(&h->new_bgs);
553 
554 	smp_mb();
555 	if (cur_trans->state >= TRANS_STATE_BLOCKED &&
556 	    may_wait_transaction(fs_info, type)) {
557 		current->journal_info = h;
558 		btrfs_commit_transaction(h);
559 		goto again;
560 	}
561 
562 	if (num_bytes) {
563 		trace_btrfs_space_reservation(fs_info, "transaction",
564 					      h->transid, num_bytes, 1);
565 		h->block_rsv = &fs_info->trans_block_rsv;
566 		h->bytes_reserved = num_bytes;
567 		h->reloc_reserved = reloc_reserved;
568 	}
569 
570 got_it:
571 	btrfs_record_root_in_trans(h, root);
572 
573 	if (!current->journal_info)
574 		current->journal_info = h;
575 	return h;
576 
577 join_fail:
578 	if (type & __TRANS_FREEZABLE)
579 		sb_end_intwrite(fs_info->sb);
580 	kmem_cache_free(btrfs_trans_handle_cachep, h);
581 alloc_fail:
582 	if (num_bytes)
583 		btrfs_block_rsv_release(fs_info, &fs_info->trans_block_rsv,
584 					num_bytes);
585 reserve_fail:
586 	btrfs_qgroup_free_meta_pertrans(root, qgroup_reserved);
587 	return ERR_PTR(ret);
588 }
589 
590 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
591 						   unsigned int num_items)
592 {
593 	return start_transaction(root, num_items, TRANS_START,
594 				 BTRFS_RESERVE_FLUSH_ALL, true);
595 }
596 
597 struct btrfs_trans_handle *btrfs_start_transaction_fallback_global_rsv(
598 					struct btrfs_root *root,
599 					unsigned int num_items,
600 					int min_factor)
601 {
602 	struct btrfs_fs_info *fs_info = root->fs_info;
603 	struct btrfs_trans_handle *trans;
604 	u64 num_bytes;
605 	int ret;
606 
607 	/*
608 	 * We have two callers: unlink and block group removal.  The
609 	 * former should succeed even if we will temporarily exceed
610 	 * quota and the latter operates on the extent root so
611 	 * qgroup enforcement is ignored anyway.
612 	 */
613 	trans = start_transaction(root, num_items, TRANS_START,
614 				  BTRFS_RESERVE_FLUSH_ALL, false);
615 	if (!IS_ERR(trans) || PTR_ERR(trans) != -ENOSPC)
616 		return trans;
617 
618 	trans = btrfs_start_transaction(root, 0);
619 	if (IS_ERR(trans))
620 		return trans;
621 
622 	num_bytes = btrfs_calc_trans_metadata_size(fs_info, num_items);
623 	ret = btrfs_cond_migrate_bytes(fs_info, &fs_info->trans_block_rsv,
624 				       num_bytes, min_factor);
625 	if (ret) {
626 		btrfs_end_transaction(trans);
627 		return ERR_PTR(ret);
628 	}
629 
630 	trans->block_rsv = &fs_info->trans_block_rsv;
631 	trans->bytes_reserved = num_bytes;
632 	trace_btrfs_space_reservation(fs_info, "transaction",
633 				      trans->transid, num_bytes, 1);
634 
635 	return trans;
636 }
637 
638 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
639 {
640 	return start_transaction(root, 0, TRANS_JOIN, BTRFS_RESERVE_NO_FLUSH,
641 				 true);
642 }
643 
644 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root)
645 {
646 	return start_transaction(root, 0, TRANS_JOIN_NOLOCK,
647 				 BTRFS_RESERVE_NO_FLUSH, true);
648 }
649 
650 /*
651  * btrfs_attach_transaction() - catch the running transaction
652  *
653  * It is used when we want to commit the current the transaction, but
654  * don't want to start a new one.
655  *
656  * Note: If this function return -ENOENT, it just means there is no
657  * running transaction. But it is possible that the inactive transaction
658  * is still in the memory, not fully on disk. If you hope there is no
659  * inactive transaction in the fs when -ENOENT is returned, you should
660  * invoke
661  *     btrfs_attach_transaction_barrier()
662  */
663 struct btrfs_trans_handle *btrfs_attach_transaction(struct btrfs_root *root)
664 {
665 	return start_transaction(root, 0, TRANS_ATTACH,
666 				 BTRFS_RESERVE_NO_FLUSH, true);
667 }
668 
669 /*
670  * btrfs_attach_transaction_barrier() - catch the running transaction
671  *
672  * It is similar to the above function, the differentia is this one
673  * will wait for all the inactive transactions until they fully
674  * complete.
675  */
676 struct btrfs_trans_handle *
677 btrfs_attach_transaction_barrier(struct btrfs_root *root)
678 {
679 	struct btrfs_trans_handle *trans;
680 
681 	trans = start_transaction(root, 0, TRANS_ATTACH,
682 				  BTRFS_RESERVE_NO_FLUSH, true);
683 	if (IS_ERR(trans) && PTR_ERR(trans) == -ENOENT)
684 		btrfs_wait_for_commit(root->fs_info, 0);
685 
686 	return trans;
687 }
688 
689 /* wait for a transaction commit to be fully complete */
690 static noinline void wait_for_commit(struct btrfs_transaction *commit)
691 {
692 	wait_event(commit->commit_wait, commit->state == TRANS_STATE_COMPLETED);
693 }
694 
695 int btrfs_wait_for_commit(struct btrfs_fs_info *fs_info, u64 transid)
696 {
697 	struct btrfs_transaction *cur_trans = NULL, *t;
698 	int ret = 0;
699 
700 	if (transid) {
701 		if (transid <= fs_info->last_trans_committed)
702 			goto out;
703 
704 		/* find specified transaction */
705 		spin_lock(&fs_info->trans_lock);
706 		list_for_each_entry(t, &fs_info->trans_list, list) {
707 			if (t->transid == transid) {
708 				cur_trans = t;
709 				refcount_inc(&cur_trans->use_count);
710 				ret = 0;
711 				break;
712 			}
713 			if (t->transid > transid) {
714 				ret = 0;
715 				break;
716 			}
717 		}
718 		spin_unlock(&fs_info->trans_lock);
719 
720 		/*
721 		 * The specified transaction doesn't exist, or we
722 		 * raced with btrfs_commit_transaction
723 		 */
724 		if (!cur_trans) {
725 			if (transid > fs_info->last_trans_committed)
726 				ret = -EINVAL;
727 			goto out;
728 		}
729 	} else {
730 		/* find newest transaction that is committing | committed */
731 		spin_lock(&fs_info->trans_lock);
732 		list_for_each_entry_reverse(t, &fs_info->trans_list,
733 					    list) {
734 			if (t->state >= TRANS_STATE_COMMIT_START) {
735 				if (t->state == TRANS_STATE_COMPLETED)
736 					break;
737 				cur_trans = t;
738 				refcount_inc(&cur_trans->use_count);
739 				break;
740 			}
741 		}
742 		spin_unlock(&fs_info->trans_lock);
743 		if (!cur_trans)
744 			goto out;  /* nothing committing|committed */
745 	}
746 
747 	wait_for_commit(cur_trans);
748 	btrfs_put_transaction(cur_trans);
749 out:
750 	return ret;
751 }
752 
753 void btrfs_throttle(struct btrfs_fs_info *fs_info)
754 {
755 	wait_current_trans(fs_info);
756 }
757 
758 static int should_end_transaction(struct btrfs_trans_handle *trans)
759 {
760 	struct btrfs_fs_info *fs_info = trans->fs_info;
761 
762 	if (btrfs_check_space_for_delayed_refs(trans, fs_info))
763 		return 1;
764 
765 	return !!btrfs_block_rsv_check(&fs_info->global_block_rsv, 5);
766 }
767 
768 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans)
769 {
770 	struct btrfs_transaction *cur_trans = trans->transaction;
771 	int updates;
772 	int err;
773 
774 	smp_mb();
775 	if (cur_trans->state >= TRANS_STATE_BLOCKED ||
776 	    cur_trans->delayed_refs.flushing)
777 		return 1;
778 
779 	updates = trans->delayed_ref_updates;
780 	trans->delayed_ref_updates = 0;
781 	if (updates) {
782 		err = btrfs_run_delayed_refs(trans, updates * 2);
783 		if (err) /* Error code will also eval true */
784 			return err;
785 	}
786 
787 	return should_end_transaction(trans);
788 }
789 
790 static void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans)
791 
792 {
793 	struct btrfs_fs_info *fs_info = trans->fs_info;
794 
795 	if (!trans->block_rsv) {
796 		ASSERT(!trans->bytes_reserved);
797 		return;
798 	}
799 
800 	if (!trans->bytes_reserved)
801 		return;
802 
803 	ASSERT(trans->block_rsv == &fs_info->trans_block_rsv);
804 	trace_btrfs_space_reservation(fs_info, "transaction",
805 				      trans->transid, trans->bytes_reserved, 0);
806 	btrfs_block_rsv_release(fs_info, trans->block_rsv,
807 				trans->bytes_reserved);
808 	trans->bytes_reserved = 0;
809 }
810 
811 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
812 				   int throttle)
813 {
814 	struct btrfs_fs_info *info = trans->fs_info;
815 	struct btrfs_transaction *cur_trans = trans->transaction;
816 	u64 transid = trans->transid;
817 	unsigned long cur = trans->delayed_ref_updates;
818 	int lock = (trans->type != TRANS_JOIN_NOLOCK);
819 	int err = 0;
820 	int must_run_delayed_refs = 0;
821 
822 	if (refcount_read(&trans->use_count) > 1) {
823 		refcount_dec(&trans->use_count);
824 		trans->block_rsv = trans->orig_rsv;
825 		return 0;
826 	}
827 
828 	btrfs_trans_release_metadata(trans);
829 	trans->block_rsv = NULL;
830 
831 	if (!list_empty(&trans->new_bgs))
832 		btrfs_create_pending_block_groups(trans);
833 
834 	trans->delayed_ref_updates = 0;
835 	if (!trans->sync) {
836 		must_run_delayed_refs =
837 			btrfs_should_throttle_delayed_refs(trans, info);
838 		cur = max_t(unsigned long, cur, 32);
839 
840 		/*
841 		 * don't make the caller wait if they are from a NOLOCK
842 		 * or ATTACH transaction, it will deadlock with commit
843 		 */
844 		if (must_run_delayed_refs == 1 &&
845 		    (trans->type & (__TRANS_JOIN_NOLOCK | __TRANS_ATTACH)))
846 			must_run_delayed_refs = 2;
847 	}
848 
849 	btrfs_trans_release_metadata(trans);
850 	trans->block_rsv = NULL;
851 
852 	if (!list_empty(&trans->new_bgs))
853 		btrfs_create_pending_block_groups(trans);
854 
855 	btrfs_trans_release_chunk_metadata(trans);
856 
857 	if (lock && should_end_transaction(trans) &&
858 	    READ_ONCE(cur_trans->state) == TRANS_STATE_RUNNING) {
859 		spin_lock(&info->trans_lock);
860 		if (cur_trans->state == TRANS_STATE_RUNNING)
861 			cur_trans->state = TRANS_STATE_BLOCKED;
862 		spin_unlock(&info->trans_lock);
863 	}
864 
865 	if (lock && READ_ONCE(cur_trans->state) == TRANS_STATE_BLOCKED) {
866 		if (throttle)
867 			return btrfs_commit_transaction(trans);
868 		else
869 			wake_up_process(info->transaction_kthread);
870 	}
871 
872 	if (trans->type & __TRANS_FREEZABLE)
873 		sb_end_intwrite(info->sb);
874 
875 	WARN_ON(cur_trans != info->running_transaction);
876 	WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
877 	atomic_dec(&cur_trans->num_writers);
878 	extwriter_counter_dec(cur_trans, trans->type);
879 
880 	cond_wake_up(&cur_trans->writer_wait);
881 	btrfs_put_transaction(cur_trans);
882 
883 	if (current->journal_info == trans)
884 		current->journal_info = NULL;
885 
886 	if (throttle)
887 		btrfs_run_delayed_iputs(info);
888 
889 	if (trans->aborted ||
890 	    test_bit(BTRFS_FS_STATE_ERROR, &info->fs_state)) {
891 		wake_up_process(info->transaction_kthread);
892 		err = -EIO;
893 	}
894 
895 	kmem_cache_free(btrfs_trans_handle_cachep, trans);
896 	if (must_run_delayed_refs) {
897 		btrfs_async_run_delayed_refs(info, cur, transid,
898 					     must_run_delayed_refs == 1);
899 	}
900 	return err;
901 }
902 
903 int btrfs_end_transaction(struct btrfs_trans_handle *trans)
904 {
905 	return __btrfs_end_transaction(trans, 0);
906 }
907 
908 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans)
909 {
910 	return __btrfs_end_transaction(trans, 1);
911 }
912 
913 /*
914  * when btree blocks are allocated, they have some corresponding bits set for
915  * them in one of two extent_io trees.  This is used to make sure all of
916  * those extents are sent to disk but does not wait on them
917  */
918 int btrfs_write_marked_extents(struct btrfs_fs_info *fs_info,
919 			       struct extent_io_tree *dirty_pages, int mark)
920 {
921 	int err = 0;
922 	int werr = 0;
923 	struct address_space *mapping = fs_info->btree_inode->i_mapping;
924 	struct extent_state *cached_state = NULL;
925 	u64 start = 0;
926 	u64 end;
927 
928 	atomic_inc(&BTRFS_I(fs_info->btree_inode)->sync_writers);
929 	while (!find_first_extent_bit(dirty_pages, start, &start, &end,
930 				      mark, &cached_state)) {
931 		bool wait_writeback = false;
932 
933 		err = convert_extent_bit(dirty_pages, start, end,
934 					 EXTENT_NEED_WAIT,
935 					 mark, &cached_state);
936 		/*
937 		 * convert_extent_bit can return -ENOMEM, which is most of the
938 		 * time a temporary error. So when it happens, ignore the error
939 		 * and wait for writeback of this range to finish - because we
940 		 * failed to set the bit EXTENT_NEED_WAIT for the range, a call
941 		 * to __btrfs_wait_marked_extents() would not know that
942 		 * writeback for this range started and therefore wouldn't
943 		 * wait for it to finish - we don't want to commit a
944 		 * superblock that points to btree nodes/leafs for which
945 		 * writeback hasn't finished yet (and without errors).
946 		 * We cleanup any entries left in the io tree when committing
947 		 * the transaction (through clear_btree_io_tree()).
948 		 */
949 		if (err == -ENOMEM) {
950 			err = 0;
951 			wait_writeback = true;
952 		}
953 		if (!err)
954 			err = filemap_fdatawrite_range(mapping, start, end);
955 		if (err)
956 			werr = err;
957 		else if (wait_writeback)
958 			werr = filemap_fdatawait_range(mapping, start, end);
959 		free_extent_state(cached_state);
960 		cached_state = NULL;
961 		cond_resched();
962 		start = end + 1;
963 	}
964 	atomic_dec(&BTRFS_I(fs_info->btree_inode)->sync_writers);
965 	return werr;
966 }
967 
968 /*
969  * when btree blocks are allocated, they have some corresponding bits set for
970  * them in one of two extent_io trees.  This is used to make sure all of
971  * those extents are on disk for transaction or log commit.  We wait
972  * on all the pages and clear them from the dirty pages state tree
973  */
974 static int __btrfs_wait_marked_extents(struct btrfs_fs_info *fs_info,
975 				       struct extent_io_tree *dirty_pages)
976 {
977 	int err = 0;
978 	int werr = 0;
979 	struct address_space *mapping = fs_info->btree_inode->i_mapping;
980 	struct extent_state *cached_state = NULL;
981 	u64 start = 0;
982 	u64 end;
983 
984 	while (!find_first_extent_bit(dirty_pages, start, &start, &end,
985 				      EXTENT_NEED_WAIT, &cached_state)) {
986 		/*
987 		 * Ignore -ENOMEM errors returned by clear_extent_bit().
988 		 * When committing the transaction, we'll remove any entries
989 		 * left in the io tree. For a log commit, we don't remove them
990 		 * after committing the log because the tree can be accessed
991 		 * concurrently - we do it only at transaction commit time when
992 		 * it's safe to do it (through clear_btree_io_tree()).
993 		 */
994 		err = clear_extent_bit(dirty_pages, start, end,
995 				       EXTENT_NEED_WAIT, 0, 0, &cached_state);
996 		if (err == -ENOMEM)
997 			err = 0;
998 		if (!err)
999 			err = filemap_fdatawait_range(mapping, start, end);
1000 		if (err)
1001 			werr = err;
1002 		free_extent_state(cached_state);
1003 		cached_state = NULL;
1004 		cond_resched();
1005 		start = end + 1;
1006 	}
1007 	if (err)
1008 		werr = err;
1009 	return werr;
1010 }
1011 
1012 int btrfs_wait_extents(struct btrfs_fs_info *fs_info,
1013 		       struct extent_io_tree *dirty_pages)
1014 {
1015 	bool errors = false;
1016 	int err;
1017 
1018 	err = __btrfs_wait_marked_extents(fs_info, dirty_pages);
1019 	if (test_and_clear_bit(BTRFS_FS_BTREE_ERR, &fs_info->flags))
1020 		errors = true;
1021 
1022 	if (errors && !err)
1023 		err = -EIO;
1024 	return err;
1025 }
1026 
1027 int btrfs_wait_tree_log_extents(struct btrfs_root *log_root, int mark)
1028 {
1029 	struct btrfs_fs_info *fs_info = log_root->fs_info;
1030 	struct extent_io_tree *dirty_pages = &log_root->dirty_log_pages;
1031 	bool errors = false;
1032 	int err;
1033 
1034 	ASSERT(log_root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
1035 
1036 	err = __btrfs_wait_marked_extents(fs_info, dirty_pages);
1037 	if ((mark & EXTENT_DIRTY) &&
1038 	    test_and_clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags))
1039 		errors = true;
1040 
1041 	if ((mark & EXTENT_NEW) &&
1042 	    test_and_clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags))
1043 		errors = true;
1044 
1045 	if (errors && !err)
1046 		err = -EIO;
1047 	return err;
1048 }
1049 
1050 /*
1051  * When btree blocks are allocated the corresponding extents are marked dirty.
1052  * This function ensures such extents are persisted on disk for transaction or
1053  * log commit.
1054  *
1055  * @trans: transaction whose dirty pages we'd like to write
1056  */
1057 static int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans)
1058 {
1059 	int ret;
1060 	int ret2;
1061 	struct extent_io_tree *dirty_pages = &trans->transaction->dirty_pages;
1062 	struct btrfs_fs_info *fs_info = trans->fs_info;
1063 	struct blk_plug plug;
1064 
1065 	blk_start_plug(&plug);
1066 	ret = btrfs_write_marked_extents(fs_info, dirty_pages, EXTENT_DIRTY);
1067 	blk_finish_plug(&plug);
1068 	ret2 = btrfs_wait_extents(fs_info, dirty_pages);
1069 
1070 	clear_btree_io_tree(&trans->transaction->dirty_pages);
1071 
1072 	if (ret)
1073 		return ret;
1074 	else if (ret2)
1075 		return ret2;
1076 	else
1077 		return 0;
1078 }
1079 
1080 /*
1081  * this is used to update the root pointer in the tree of tree roots.
1082  *
1083  * But, in the case of the extent allocation tree, updating the root
1084  * pointer may allocate blocks which may change the root of the extent
1085  * allocation tree.
1086  *
1087  * So, this loops and repeats and makes sure the cowonly root didn't
1088  * change while the root pointer was being updated in the metadata.
1089  */
1090 static int update_cowonly_root(struct btrfs_trans_handle *trans,
1091 			       struct btrfs_root *root)
1092 {
1093 	int ret;
1094 	u64 old_root_bytenr;
1095 	u64 old_root_used;
1096 	struct btrfs_fs_info *fs_info = root->fs_info;
1097 	struct btrfs_root *tree_root = fs_info->tree_root;
1098 
1099 	old_root_used = btrfs_root_used(&root->root_item);
1100 
1101 	while (1) {
1102 		old_root_bytenr = btrfs_root_bytenr(&root->root_item);
1103 		if (old_root_bytenr == root->node->start &&
1104 		    old_root_used == btrfs_root_used(&root->root_item))
1105 			break;
1106 
1107 		btrfs_set_root_node(&root->root_item, root->node);
1108 		ret = btrfs_update_root(trans, tree_root,
1109 					&root->root_key,
1110 					&root->root_item);
1111 		if (ret)
1112 			return ret;
1113 
1114 		old_root_used = btrfs_root_used(&root->root_item);
1115 	}
1116 
1117 	return 0;
1118 }
1119 
1120 /*
1121  * update all the cowonly tree roots on disk
1122  *
1123  * The error handling in this function may not be obvious. Any of the
1124  * failures will cause the file system to go offline. We still need
1125  * to clean up the delayed refs.
1126  */
1127 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans)
1128 {
1129 	struct btrfs_fs_info *fs_info = trans->fs_info;
1130 	struct list_head *dirty_bgs = &trans->transaction->dirty_bgs;
1131 	struct list_head *io_bgs = &trans->transaction->io_bgs;
1132 	struct list_head *next;
1133 	struct extent_buffer *eb;
1134 	int ret;
1135 
1136 	eb = btrfs_lock_root_node(fs_info->tree_root);
1137 	ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL,
1138 			      0, &eb);
1139 	btrfs_tree_unlock(eb);
1140 	free_extent_buffer(eb);
1141 
1142 	if (ret)
1143 		return ret;
1144 
1145 	ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
1146 	if (ret)
1147 		return ret;
1148 
1149 	ret = btrfs_run_dev_stats(trans, fs_info);
1150 	if (ret)
1151 		return ret;
1152 	ret = btrfs_run_dev_replace(trans, fs_info);
1153 	if (ret)
1154 		return ret;
1155 	ret = btrfs_run_qgroups(trans, fs_info);
1156 	if (ret)
1157 		return ret;
1158 
1159 	ret = btrfs_setup_space_cache(trans, fs_info);
1160 	if (ret)
1161 		return ret;
1162 
1163 	/* run_qgroups might have added some more refs */
1164 	ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
1165 	if (ret)
1166 		return ret;
1167 again:
1168 	while (!list_empty(&fs_info->dirty_cowonly_roots)) {
1169 		struct btrfs_root *root;
1170 		next = fs_info->dirty_cowonly_roots.next;
1171 		list_del_init(next);
1172 		root = list_entry(next, struct btrfs_root, dirty_list);
1173 		clear_bit(BTRFS_ROOT_DIRTY, &root->state);
1174 
1175 		if (root != fs_info->extent_root)
1176 			list_add_tail(&root->dirty_list,
1177 				      &trans->transaction->switch_commits);
1178 		ret = update_cowonly_root(trans, root);
1179 		if (ret)
1180 			return ret;
1181 		ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
1182 		if (ret)
1183 			return ret;
1184 	}
1185 
1186 	while (!list_empty(dirty_bgs) || !list_empty(io_bgs)) {
1187 		ret = btrfs_write_dirty_block_groups(trans, fs_info);
1188 		if (ret)
1189 			return ret;
1190 		ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
1191 		if (ret)
1192 			return ret;
1193 	}
1194 
1195 	if (!list_empty(&fs_info->dirty_cowonly_roots))
1196 		goto again;
1197 
1198 	list_add_tail(&fs_info->extent_root->dirty_list,
1199 		      &trans->transaction->switch_commits);
1200 	btrfs_after_dev_replace_commit(fs_info);
1201 
1202 	return 0;
1203 }
1204 
1205 /*
1206  * dead roots are old snapshots that need to be deleted.  This allocates
1207  * a dirty root struct and adds it into the list of dead roots that need to
1208  * be deleted
1209  */
1210 void btrfs_add_dead_root(struct btrfs_root *root)
1211 {
1212 	struct btrfs_fs_info *fs_info = root->fs_info;
1213 
1214 	spin_lock(&fs_info->trans_lock);
1215 	if (list_empty(&root->root_list))
1216 		list_add_tail(&root->root_list, &fs_info->dead_roots);
1217 	spin_unlock(&fs_info->trans_lock);
1218 }
1219 
1220 /*
1221  * update all the cowonly tree roots on disk
1222  */
1223 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans)
1224 {
1225 	struct btrfs_fs_info *fs_info = trans->fs_info;
1226 	struct btrfs_root *gang[8];
1227 	int i;
1228 	int ret;
1229 	int err = 0;
1230 
1231 	spin_lock(&fs_info->fs_roots_radix_lock);
1232 	while (1) {
1233 		ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
1234 						 (void **)gang, 0,
1235 						 ARRAY_SIZE(gang),
1236 						 BTRFS_ROOT_TRANS_TAG);
1237 		if (ret == 0)
1238 			break;
1239 		for (i = 0; i < ret; i++) {
1240 			struct btrfs_root *root = gang[i];
1241 			radix_tree_tag_clear(&fs_info->fs_roots_radix,
1242 					(unsigned long)root->root_key.objectid,
1243 					BTRFS_ROOT_TRANS_TAG);
1244 			spin_unlock(&fs_info->fs_roots_radix_lock);
1245 
1246 			btrfs_free_log(trans, root);
1247 			btrfs_update_reloc_root(trans, root);
1248 
1249 			btrfs_save_ino_cache(root, trans);
1250 
1251 			/* see comments in should_cow_block() */
1252 			clear_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1253 			smp_mb__after_atomic();
1254 
1255 			if (root->commit_root != root->node) {
1256 				list_add_tail(&root->dirty_list,
1257 					&trans->transaction->switch_commits);
1258 				btrfs_set_root_node(&root->root_item,
1259 						    root->node);
1260 			}
1261 
1262 			err = btrfs_update_root(trans, fs_info->tree_root,
1263 						&root->root_key,
1264 						&root->root_item);
1265 			spin_lock(&fs_info->fs_roots_radix_lock);
1266 			if (err)
1267 				break;
1268 			btrfs_qgroup_free_meta_all_pertrans(root);
1269 		}
1270 	}
1271 	spin_unlock(&fs_info->fs_roots_radix_lock);
1272 	return err;
1273 }
1274 
1275 /*
1276  * defrag a given btree.
1277  * Every leaf in the btree is read and defragged.
1278  */
1279 int btrfs_defrag_root(struct btrfs_root *root)
1280 {
1281 	struct btrfs_fs_info *info = root->fs_info;
1282 	struct btrfs_trans_handle *trans;
1283 	int ret;
1284 
1285 	if (test_and_set_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state))
1286 		return 0;
1287 
1288 	while (1) {
1289 		trans = btrfs_start_transaction(root, 0);
1290 		if (IS_ERR(trans))
1291 			return PTR_ERR(trans);
1292 
1293 		ret = btrfs_defrag_leaves(trans, root);
1294 
1295 		btrfs_end_transaction(trans);
1296 		btrfs_btree_balance_dirty(info);
1297 		cond_resched();
1298 
1299 		if (btrfs_fs_closing(info) || ret != -EAGAIN)
1300 			break;
1301 
1302 		if (btrfs_defrag_cancelled(info)) {
1303 			btrfs_debug(info, "defrag_root cancelled");
1304 			ret = -EAGAIN;
1305 			break;
1306 		}
1307 	}
1308 	clear_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state);
1309 	return ret;
1310 }
1311 
1312 /*
1313  * Do all special snapshot related qgroup dirty hack.
1314  *
1315  * Will do all needed qgroup inherit and dirty hack like switch commit
1316  * roots inside one transaction and write all btree into disk, to make
1317  * qgroup works.
1318  */
1319 static int qgroup_account_snapshot(struct btrfs_trans_handle *trans,
1320 				   struct btrfs_root *src,
1321 				   struct btrfs_root *parent,
1322 				   struct btrfs_qgroup_inherit *inherit,
1323 				   u64 dst_objectid)
1324 {
1325 	struct btrfs_fs_info *fs_info = src->fs_info;
1326 	int ret;
1327 
1328 	/*
1329 	 * Save some performance in the case that qgroups are not
1330 	 * enabled. If this check races with the ioctl, rescan will
1331 	 * kick in anyway.
1332 	 */
1333 	if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
1334 		return 0;
1335 
1336 	/*
1337 	 * Ensure dirty @src will be commited.  Or, after comming
1338 	 * commit_fs_roots() and switch_commit_roots(), any dirty but not
1339 	 * recorded root will never be updated again, causing an outdated root
1340 	 * item.
1341 	 */
1342 	record_root_in_trans(trans, src, 1);
1343 
1344 	/*
1345 	 * We are going to commit transaction, see btrfs_commit_transaction()
1346 	 * comment for reason locking tree_log_mutex
1347 	 */
1348 	mutex_lock(&fs_info->tree_log_mutex);
1349 
1350 	ret = commit_fs_roots(trans);
1351 	if (ret)
1352 		goto out;
1353 	ret = btrfs_qgroup_account_extents(trans);
1354 	if (ret < 0)
1355 		goto out;
1356 
1357 	/* Now qgroup are all updated, we can inherit it to new qgroups */
1358 	ret = btrfs_qgroup_inherit(trans, fs_info,
1359 				   src->root_key.objectid, dst_objectid,
1360 				   inherit);
1361 	if (ret < 0)
1362 		goto out;
1363 
1364 	/*
1365 	 * Now we do a simplified commit transaction, which will:
1366 	 * 1) commit all subvolume and extent tree
1367 	 *    To ensure all subvolume and extent tree have a valid
1368 	 *    commit_root to accounting later insert_dir_item()
1369 	 * 2) write all btree blocks onto disk
1370 	 *    This is to make sure later btree modification will be cowed
1371 	 *    Or commit_root can be populated and cause wrong qgroup numbers
1372 	 * In this simplified commit, we don't really care about other trees
1373 	 * like chunk and root tree, as they won't affect qgroup.
1374 	 * And we don't write super to avoid half committed status.
1375 	 */
1376 	ret = commit_cowonly_roots(trans);
1377 	if (ret)
1378 		goto out;
1379 	switch_commit_roots(trans->transaction);
1380 	ret = btrfs_write_and_wait_transaction(trans);
1381 	if (ret)
1382 		btrfs_handle_fs_error(fs_info, ret,
1383 			"Error while writing out transaction for qgroup");
1384 
1385 out:
1386 	mutex_unlock(&fs_info->tree_log_mutex);
1387 
1388 	/*
1389 	 * Force parent root to be updated, as we recorded it before so its
1390 	 * last_trans == cur_transid.
1391 	 * Or it won't be committed again onto disk after later
1392 	 * insert_dir_item()
1393 	 */
1394 	if (!ret)
1395 		record_root_in_trans(trans, parent, 1);
1396 	return ret;
1397 }
1398 
1399 /*
1400  * new snapshots need to be created at a very specific time in the
1401  * transaction commit.  This does the actual creation.
1402  *
1403  * Note:
1404  * If the error which may affect the commitment of the current transaction
1405  * happens, we should return the error number. If the error which just affect
1406  * the creation of the pending snapshots, just return 0.
1407  */
1408 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
1409 				   struct btrfs_pending_snapshot *pending)
1410 {
1411 
1412 	struct btrfs_fs_info *fs_info = trans->fs_info;
1413 	struct btrfs_key key;
1414 	struct btrfs_root_item *new_root_item;
1415 	struct btrfs_root *tree_root = fs_info->tree_root;
1416 	struct btrfs_root *root = pending->root;
1417 	struct btrfs_root *parent_root;
1418 	struct btrfs_block_rsv *rsv;
1419 	struct inode *parent_inode;
1420 	struct btrfs_path *path;
1421 	struct btrfs_dir_item *dir_item;
1422 	struct dentry *dentry;
1423 	struct extent_buffer *tmp;
1424 	struct extent_buffer *old;
1425 	struct timespec cur_time;
1426 	int ret = 0;
1427 	u64 to_reserve = 0;
1428 	u64 index = 0;
1429 	u64 objectid;
1430 	u64 root_flags;
1431 	uuid_le new_uuid;
1432 
1433 	ASSERT(pending->path);
1434 	path = pending->path;
1435 
1436 	ASSERT(pending->root_item);
1437 	new_root_item = pending->root_item;
1438 
1439 	pending->error = btrfs_find_free_objectid(tree_root, &objectid);
1440 	if (pending->error)
1441 		goto no_free_objectid;
1442 
1443 	/*
1444 	 * Make qgroup to skip current new snapshot's qgroupid, as it is
1445 	 * accounted by later btrfs_qgroup_inherit().
1446 	 */
1447 	btrfs_set_skip_qgroup(trans, objectid);
1448 
1449 	btrfs_reloc_pre_snapshot(pending, &to_reserve);
1450 
1451 	if (to_reserve > 0) {
1452 		pending->error = btrfs_block_rsv_add(root,
1453 						     &pending->block_rsv,
1454 						     to_reserve,
1455 						     BTRFS_RESERVE_NO_FLUSH);
1456 		if (pending->error)
1457 			goto clear_skip_qgroup;
1458 	}
1459 
1460 	key.objectid = objectid;
1461 	key.offset = (u64)-1;
1462 	key.type = BTRFS_ROOT_ITEM_KEY;
1463 
1464 	rsv = trans->block_rsv;
1465 	trans->block_rsv = &pending->block_rsv;
1466 	trans->bytes_reserved = trans->block_rsv->reserved;
1467 	trace_btrfs_space_reservation(fs_info, "transaction",
1468 				      trans->transid,
1469 				      trans->bytes_reserved, 1);
1470 	dentry = pending->dentry;
1471 	parent_inode = pending->dir;
1472 	parent_root = BTRFS_I(parent_inode)->root;
1473 	record_root_in_trans(trans, parent_root, 0);
1474 
1475 	cur_time = current_time(parent_inode);
1476 
1477 	/*
1478 	 * insert the directory item
1479 	 */
1480 	ret = btrfs_set_inode_index(BTRFS_I(parent_inode), &index);
1481 	BUG_ON(ret); /* -ENOMEM */
1482 
1483 	/* check if there is a file/dir which has the same name. */
1484 	dir_item = btrfs_lookup_dir_item(NULL, parent_root, path,
1485 					 btrfs_ino(BTRFS_I(parent_inode)),
1486 					 dentry->d_name.name,
1487 					 dentry->d_name.len, 0);
1488 	if (dir_item != NULL && !IS_ERR(dir_item)) {
1489 		pending->error = -EEXIST;
1490 		goto dir_item_existed;
1491 	} else if (IS_ERR(dir_item)) {
1492 		ret = PTR_ERR(dir_item);
1493 		btrfs_abort_transaction(trans, ret);
1494 		goto fail;
1495 	}
1496 	btrfs_release_path(path);
1497 
1498 	/*
1499 	 * pull in the delayed directory update
1500 	 * and the delayed inode item
1501 	 * otherwise we corrupt the FS during
1502 	 * snapshot
1503 	 */
1504 	ret = btrfs_run_delayed_items(trans);
1505 	if (ret) {	/* Transaction aborted */
1506 		btrfs_abort_transaction(trans, ret);
1507 		goto fail;
1508 	}
1509 
1510 	record_root_in_trans(trans, root, 0);
1511 	btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
1512 	memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
1513 	btrfs_check_and_init_root_item(new_root_item);
1514 
1515 	root_flags = btrfs_root_flags(new_root_item);
1516 	if (pending->readonly)
1517 		root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
1518 	else
1519 		root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
1520 	btrfs_set_root_flags(new_root_item, root_flags);
1521 
1522 	btrfs_set_root_generation_v2(new_root_item,
1523 			trans->transid);
1524 	uuid_le_gen(&new_uuid);
1525 	memcpy(new_root_item->uuid, new_uuid.b, BTRFS_UUID_SIZE);
1526 	memcpy(new_root_item->parent_uuid, root->root_item.uuid,
1527 			BTRFS_UUID_SIZE);
1528 	if (!(root_flags & BTRFS_ROOT_SUBVOL_RDONLY)) {
1529 		memset(new_root_item->received_uuid, 0,
1530 		       sizeof(new_root_item->received_uuid));
1531 		memset(&new_root_item->stime, 0, sizeof(new_root_item->stime));
1532 		memset(&new_root_item->rtime, 0, sizeof(new_root_item->rtime));
1533 		btrfs_set_root_stransid(new_root_item, 0);
1534 		btrfs_set_root_rtransid(new_root_item, 0);
1535 	}
1536 	btrfs_set_stack_timespec_sec(&new_root_item->otime, cur_time.tv_sec);
1537 	btrfs_set_stack_timespec_nsec(&new_root_item->otime, cur_time.tv_nsec);
1538 	btrfs_set_root_otransid(new_root_item, trans->transid);
1539 
1540 	old = btrfs_lock_root_node(root);
1541 	ret = btrfs_cow_block(trans, root, old, NULL, 0, &old);
1542 	if (ret) {
1543 		btrfs_tree_unlock(old);
1544 		free_extent_buffer(old);
1545 		btrfs_abort_transaction(trans, ret);
1546 		goto fail;
1547 	}
1548 
1549 	btrfs_set_lock_blocking(old);
1550 
1551 	ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1552 	/* clean up in any case */
1553 	btrfs_tree_unlock(old);
1554 	free_extent_buffer(old);
1555 	if (ret) {
1556 		btrfs_abort_transaction(trans, ret);
1557 		goto fail;
1558 	}
1559 	/* see comments in should_cow_block() */
1560 	set_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1561 	smp_wmb();
1562 
1563 	btrfs_set_root_node(new_root_item, tmp);
1564 	/* record when the snapshot was created in key.offset */
1565 	key.offset = trans->transid;
1566 	ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1567 	btrfs_tree_unlock(tmp);
1568 	free_extent_buffer(tmp);
1569 	if (ret) {
1570 		btrfs_abort_transaction(trans, ret);
1571 		goto fail;
1572 	}
1573 
1574 	/*
1575 	 * insert root back/forward references
1576 	 */
1577 	ret = btrfs_add_root_ref(trans, fs_info, objectid,
1578 				 parent_root->root_key.objectid,
1579 				 btrfs_ino(BTRFS_I(parent_inode)), index,
1580 				 dentry->d_name.name, dentry->d_name.len);
1581 	if (ret) {
1582 		btrfs_abort_transaction(trans, ret);
1583 		goto fail;
1584 	}
1585 
1586 	key.offset = (u64)-1;
1587 	pending->snap = btrfs_read_fs_root_no_name(fs_info, &key);
1588 	if (IS_ERR(pending->snap)) {
1589 		ret = PTR_ERR(pending->snap);
1590 		btrfs_abort_transaction(trans, ret);
1591 		goto fail;
1592 	}
1593 
1594 	ret = btrfs_reloc_post_snapshot(trans, pending);
1595 	if (ret) {
1596 		btrfs_abort_transaction(trans, ret);
1597 		goto fail;
1598 	}
1599 
1600 	ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
1601 	if (ret) {
1602 		btrfs_abort_transaction(trans, ret);
1603 		goto fail;
1604 	}
1605 
1606 	/*
1607 	 * Do special qgroup accounting for snapshot, as we do some qgroup
1608 	 * snapshot hack to do fast snapshot.
1609 	 * To co-operate with that hack, we do hack again.
1610 	 * Or snapshot will be greatly slowed down by a subtree qgroup rescan
1611 	 */
1612 	ret = qgroup_account_snapshot(trans, root, parent_root,
1613 				      pending->inherit, objectid);
1614 	if (ret < 0)
1615 		goto fail;
1616 
1617 	ret = btrfs_insert_dir_item(trans, parent_root,
1618 				    dentry->d_name.name, dentry->d_name.len,
1619 				    BTRFS_I(parent_inode), &key,
1620 				    BTRFS_FT_DIR, index);
1621 	/* We have check then name at the beginning, so it is impossible. */
1622 	BUG_ON(ret == -EEXIST || ret == -EOVERFLOW);
1623 	if (ret) {
1624 		btrfs_abort_transaction(trans, ret);
1625 		goto fail;
1626 	}
1627 
1628 	btrfs_i_size_write(BTRFS_I(parent_inode), parent_inode->i_size +
1629 					 dentry->d_name.len * 2);
1630 	parent_inode->i_mtime = parent_inode->i_ctime =
1631 		current_time(parent_inode);
1632 	ret = btrfs_update_inode_fallback(trans, parent_root, parent_inode);
1633 	if (ret) {
1634 		btrfs_abort_transaction(trans, ret);
1635 		goto fail;
1636 	}
1637 	ret = btrfs_uuid_tree_add(trans, new_uuid.b, BTRFS_UUID_KEY_SUBVOL,
1638 				  objectid);
1639 	if (ret) {
1640 		btrfs_abort_transaction(trans, ret);
1641 		goto fail;
1642 	}
1643 	if (!btrfs_is_empty_uuid(new_root_item->received_uuid)) {
1644 		ret = btrfs_uuid_tree_add(trans, new_root_item->received_uuid,
1645 					  BTRFS_UUID_KEY_RECEIVED_SUBVOL,
1646 					  objectid);
1647 		if (ret && ret != -EEXIST) {
1648 			btrfs_abort_transaction(trans, ret);
1649 			goto fail;
1650 		}
1651 	}
1652 
1653 	ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
1654 	if (ret) {
1655 		btrfs_abort_transaction(trans, ret);
1656 		goto fail;
1657 	}
1658 
1659 fail:
1660 	pending->error = ret;
1661 dir_item_existed:
1662 	trans->block_rsv = rsv;
1663 	trans->bytes_reserved = 0;
1664 clear_skip_qgroup:
1665 	btrfs_clear_skip_qgroup(trans);
1666 no_free_objectid:
1667 	kfree(new_root_item);
1668 	pending->root_item = NULL;
1669 	btrfs_free_path(path);
1670 	pending->path = NULL;
1671 
1672 	return ret;
1673 }
1674 
1675 /*
1676  * create all the snapshots we've scheduled for creation
1677  */
1678 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans)
1679 {
1680 	struct btrfs_pending_snapshot *pending, *next;
1681 	struct list_head *head = &trans->transaction->pending_snapshots;
1682 	int ret = 0;
1683 
1684 	list_for_each_entry_safe(pending, next, head, list) {
1685 		list_del(&pending->list);
1686 		ret = create_pending_snapshot(trans, pending);
1687 		if (ret)
1688 			break;
1689 	}
1690 	return ret;
1691 }
1692 
1693 static void update_super_roots(struct btrfs_fs_info *fs_info)
1694 {
1695 	struct btrfs_root_item *root_item;
1696 	struct btrfs_super_block *super;
1697 
1698 	super = fs_info->super_copy;
1699 
1700 	root_item = &fs_info->chunk_root->root_item;
1701 	super->chunk_root = root_item->bytenr;
1702 	super->chunk_root_generation = root_item->generation;
1703 	super->chunk_root_level = root_item->level;
1704 
1705 	root_item = &fs_info->tree_root->root_item;
1706 	super->root = root_item->bytenr;
1707 	super->generation = root_item->generation;
1708 	super->root_level = root_item->level;
1709 	if (btrfs_test_opt(fs_info, SPACE_CACHE))
1710 		super->cache_generation = root_item->generation;
1711 	if (test_bit(BTRFS_FS_UPDATE_UUID_TREE_GEN, &fs_info->flags))
1712 		super->uuid_tree_generation = root_item->generation;
1713 }
1714 
1715 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1716 {
1717 	struct btrfs_transaction *trans;
1718 	int ret = 0;
1719 
1720 	spin_lock(&info->trans_lock);
1721 	trans = info->running_transaction;
1722 	if (trans)
1723 		ret = (trans->state >= TRANS_STATE_COMMIT_START);
1724 	spin_unlock(&info->trans_lock);
1725 	return ret;
1726 }
1727 
1728 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1729 {
1730 	struct btrfs_transaction *trans;
1731 	int ret = 0;
1732 
1733 	spin_lock(&info->trans_lock);
1734 	trans = info->running_transaction;
1735 	if (trans)
1736 		ret = is_transaction_blocked(trans);
1737 	spin_unlock(&info->trans_lock);
1738 	return ret;
1739 }
1740 
1741 /*
1742  * wait for the current transaction commit to start and block subsequent
1743  * transaction joins
1744  */
1745 static void wait_current_trans_commit_start(struct btrfs_fs_info *fs_info,
1746 					    struct btrfs_transaction *trans)
1747 {
1748 	wait_event(fs_info->transaction_blocked_wait,
1749 		   trans->state >= TRANS_STATE_COMMIT_START || trans->aborted);
1750 }
1751 
1752 /*
1753  * wait for the current transaction to start and then become unblocked.
1754  * caller holds ref.
1755  */
1756 static void wait_current_trans_commit_start_and_unblock(
1757 					struct btrfs_fs_info *fs_info,
1758 					struct btrfs_transaction *trans)
1759 {
1760 	wait_event(fs_info->transaction_wait,
1761 		   trans->state >= TRANS_STATE_UNBLOCKED || trans->aborted);
1762 }
1763 
1764 /*
1765  * commit transactions asynchronously. once btrfs_commit_transaction_async
1766  * returns, any subsequent transaction will not be allowed to join.
1767  */
1768 struct btrfs_async_commit {
1769 	struct btrfs_trans_handle *newtrans;
1770 	struct work_struct work;
1771 };
1772 
1773 static void do_async_commit(struct work_struct *work)
1774 {
1775 	struct btrfs_async_commit *ac =
1776 		container_of(work, struct btrfs_async_commit, work);
1777 
1778 	/*
1779 	 * We've got freeze protection passed with the transaction.
1780 	 * Tell lockdep about it.
1781 	 */
1782 	if (ac->newtrans->type & __TRANS_FREEZABLE)
1783 		__sb_writers_acquired(ac->newtrans->fs_info->sb, SB_FREEZE_FS);
1784 
1785 	current->journal_info = ac->newtrans;
1786 
1787 	btrfs_commit_transaction(ac->newtrans);
1788 	kfree(ac);
1789 }
1790 
1791 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1792 				   int wait_for_unblock)
1793 {
1794 	struct btrfs_fs_info *fs_info = trans->fs_info;
1795 	struct btrfs_async_commit *ac;
1796 	struct btrfs_transaction *cur_trans;
1797 
1798 	ac = kmalloc(sizeof(*ac), GFP_NOFS);
1799 	if (!ac)
1800 		return -ENOMEM;
1801 
1802 	INIT_WORK(&ac->work, do_async_commit);
1803 	ac->newtrans = btrfs_join_transaction(trans->root);
1804 	if (IS_ERR(ac->newtrans)) {
1805 		int err = PTR_ERR(ac->newtrans);
1806 		kfree(ac);
1807 		return err;
1808 	}
1809 
1810 	/* take transaction reference */
1811 	cur_trans = trans->transaction;
1812 	refcount_inc(&cur_trans->use_count);
1813 
1814 	btrfs_end_transaction(trans);
1815 
1816 	/*
1817 	 * Tell lockdep we've released the freeze rwsem, since the
1818 	 * async commit thread will be the one to unlock it.
1819 	 */
1820 	if (ac->newtrans->type & __TRANS_FREEZABLE)
1821 		__sb_writers_release(fs_info->sb, SB_FREEZE_FS);
1822 
1823 	schedule_work(&ac->work);
1824 
1825 	/* wait for transaction to start and unblock */
1826 	if (wait_for_unblock)
1827 		wait_current_trans_commit_start_and_unblock(fs_info, cur_trans);
1828 	else
1829 		wait_current_trans_commit_start(fs_info, cur_trans);
1830 
1831 	if (current->journal_info == trans)
1832 		current->journal_info = NULL;
1833 
1834 	btrfs_put_transaction(cur_trans);
1835 	return 0;
1836 }
1837 
1838 
1839 static void cleanup_transaction(struct btrfs_trans_handle *trans, int err)
1840 {
1841 	struct btrfs_fs_info *fs_info = trans->fs_info;
1842 	struct btrfs_transaction *cur_trans = trans->transaction;
1843 	DEFINE_WAIT(wait);
1844 
1845 	WARN_ON(refcount_read(&trans->use_count) > 1);
1846 
1847 	btrfs_abort_transaction(trans, err);
1848 
1849 	spin_lock(&fs_info->trans_lock);
1850 
1851 	/*
1852 	 * If the transaction is removed from the list, it means this
1853 	 * transaction has been committed successfully, so it is impossible
1854 	 * to call the cleanup function.
1855 	 */
1856 	BUG_ON(list_empty(&cur_trans->list));
1857 
1858 	list_del_init(&cur_trans->list);
1859 	if (cur_trans == fs_info->running_transaction) {
1860 		cur_trans->state = TRANS_STATE_COMMIT_DOING;
1861 		spin_unlock(&fs_info->trans_lock);
1862 		wait_event(cur_trans->writer_wait,
1863 			   atomic_read(&cur_trans->num_writers) == 1);
1864 
1865 		spin_lock(&fs_info->trans_lock);
1866 	}
1867 	spin_unlock(&fs_info->trans_lock);
1868 
1869 	btrfs_cleanup_one_transaction(trans->transaction, fs_info);
1870 
1871 	spin_lock(&fs_info->trans_lock);
1872 	if (cur_trans == fs_info->running_transaction)
1873 		fs_info->running_transaction = NULL;
1874 	spin_unlock(&fs_info->trans_lock);
1875 
1876 	if (trans->type & __TRANS_FREEZABLE)
1877 		sb_end_intwrite(fs_info->sb);
1878 	btrfs_put_transaction(cur_trans);
1879 	btrfs_put_transaction(cur_trans);
1880 
1881 	trace_btrfs_transaction_commit(trans->root);
1882 
1883 	if (current->journal_info == trans)
1884 		current->journal_info = NULL;
1885 	btrfs_scrub_cancel(fs_info);
1886 
1887 	kmem_cache_free(btrfs_trans_handle_cachep, trans);
1888 }
1889 
1890 static inline int btrfs_start_delalloc_flush(struct btrfs_fs_info *fs_info)
1891 {
1892 	/*
1893 	 * We use writeback_inodes_sb here because if we used
1894 	 * btrfs_start_delalloc_roots we would deadlock with fs freeze.
1895 	 * Currently are holding the fs freeze lock, if we do an async flush
1896 	 * we'll do btrfs_join_transaction() and deadlock because we need to
1897 	 * wait for the fs freeze lock.  Using the direct flushing we benefit
1898 	 * from already being in a transaction and our join_transaction doesn't
1899 	 * have to re-take the fs freeze lock.
1900 	 */
1901 	if (btrfs_test_opt(fs_info, FLUSHONCOMMIT))
1902 		writeback_inodes_sb(fs_info->sb, WB_REASON_SYNC);
1903 	return 0;
1904 }
1905 
1906 static inline void btrfs_wait_delalloc_flush(struct btrfs_fs_info *fs_info)
1907 {
1908 	if (btrfs_test_opt(fs_info, FLUSHONCOMMIT))
1909 		btrfs_wait_ordered_roots(fs_info, U64_MAX, 0, (u64)-1);
1910 }
1911 
1912 static inline void
1913 btrfs_wait_pending_ordered(struct btrfs_transaction *cur_trans)
1914 {
1915 	wait_event(cur_trans->pending_wait,
1916 		   atomic_read(&cur_trans->pending_ordered) == 0);
1917 }
1918 
1919 int btrfs_commit_transaction(struct btrfs_trans_handle *trans)
1920 {
1921 	struct btrfs_fs_info *fs_info = trans->fs_info;
1922 	struct btrfs_transaction *cur_trans = trans->transaction;
1923 	struct btrfs_transaction *prev_trans = NULL;
1924 	int ret;
1925 
1926 	/* Stop the commit early if ->aborted is set */
1927 	if (unlikely(READ_ONCE(cur_trans->aborted))) {
1928 		ret = cur_trans->aborted;
1929 		btrfs_end_transaction(trans);
1930 		return ret;
1931 	}
1932 
1933 	/* make a pass through all the delayed refs we have so far
1934 	 * any runnings procs may add more while we are here
1935 	 */
1936 	ret = btrfs_run_delayed_refs(trans, 0);
1937 	if (ret) {
1938 		btrfs_end_transaction(trans);
1939 		return ret;
1940 	}
1941 
1942 	btrfs_trans_release_metadata(trans);
1943 	trans->block_rsv = NULL;
1944 
1945 	cur_trans = trans->transaction;
1946 
1947 	/*
1948 	 * set the flushing flag so procs in this transaction have to
1949 	 * start sending their work down.
1950 	 */
1951 	cur_trans->delayed_refs.flushing = 1;
1952 	smp_wmb();
1953 
1954 	if (!list_empty(&trans->new_bgs))
1955 		btrfs_create_pending_block_groups(trans);
1956 
1957 	ret = btrfs_run_delayed_refs(trans, 0);
1958 	if (ret) {
1959 		btrfs_end_transaction(trans);
1960 		return ret;
1961 	}
1962 
1963 	if (!test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &cur_trans->flags)) {
1964 		int run_it = 0;
1965 
1966 		/* this mutex is also taken before trying to set
1967 		 * block groups readonly.  We need to make sure
1968 		 * that nobody has set a block group readonly
1969 		 * after a extents from that block group have been
1970 		 * allocated for cache files.  btrfs_set_block_group_ro
1971 		 * will wait for the transaction to commit if it
1972 		 * finds BTRFS_TRANS_DIRTY_BG_RUN set.
1973 		 *
1974 		 * The BTRFS_TRANS_DIRTY_BG_RUN flag is also used to make sure
1975 		 * only one process starts all the block group IO.  It wouldn't
1976 		 * hurt to have more than one go through, but there's no
1977 		 * real advantage to it either.
1978 		 */
1979 		mutex_lock(&fs_info->ro_block_group_mutex);
1980 		if (!test_and_set_bit(BTRFS_TRANS_DIRTY_BG_RUN,
1981 				      &cur_trans->flags))
1982 			run_it = 1;
1983 		mutex_unlock(&fs_info->ro_block_group_mutex);
1984 
1985 		if (run_it) {
1986 			ret = btrfs_start_dirty_block_groups(trans);
1987 			if (ret) {
1988 				btrfs_end_transaction(trans);
1989 				return ret;
1990 			}
1991 		}
1992 	}
1993 
1994 	spin_lock(&fs_info->trans_lock);
1995 	if (cur_trans->state >= TRANS_STATE_COMMIT_START) {
1996 		spin_unlock(&fs_info->trans_lock);
1997 		refcount_inc(&cur_trans->use_count);
1998 		ret = btrfs_end_transaction(trans);
1999 
2000 		wait_for_commit(cur_trans);
2001 
2002 		if (unlikely(cur_trans->aborted))
2003 			ret = cur_trans->aborted;
2004 
2005 		btrfs_put_transaction(cur_trans);
2006 
2007 		return ret;
2008 	}
2009 
2010 	cur_trans->state = TRANS_STATE_COMMIT_START;
2011 	wake_up(&fs_info->transaction_blocked_wait);
2012 
2013 	if (cur_trans->list.prev != &fs_info->trans_list) {
2014 		prev_trans = list_entry(cur_trans->list.prev,
2015 					struct btrfs_transaction, list);
2016 		if (prev_trans->state != TRANS_STATE_COMPLETED) {
2017 			refcount_inc(&prev_trans->use_count);
2018 			spin_unlock(&fs_info->trans_lock);
2019 
2020 			wait_for_commit(prev_trans);
2021 			ret = prev_trans->aborted;
2022 
2023 			btrfs_put_transaction(prev_trans);
2024 			if (ret)
2025 				goto cleanup_transaction;
2026 		} else {
2027 			spin_unlock(&fs_info->trans_lock);
2028 		}
2029 	} else {
2030 		spin_unlock(&fs_info->trans_lock);
2031 	}
2032 
2033 	extwriter_counter_dec(cur_trans, trans->type);
2034 
2035 	ret = btrfs_start_delalloc_flush(fs_info);
2036 	if (ret)
2037 		goto cleanup_transaction;
2038 
2039 	ret = btrfs_run_delayed_items(trans);
2040 	if (ret)
2041 		goto cleanup_transaction;
2042 
2043 	wait_event(cur_trans->writer_wait,
2044 		   extwriter_counter_read(cur_trans) == 0);
2045 
2046 	/* some pending stuffs might be added after the previous flush. */
2047 	ret = btrfs_run_delayed_items(trans);
2048 	if (ret)
2049 		goto cleanup_transaction;
2050 
2051 	btrfs_wait_delalloc_flush(fs_info);
2052 
2053 	btrfs_wait_pending_ordered(cur_trans);
2054 
2055 	btrfs_scrub_pause(fs_info);
2056 	/*
2057 	 * Ok now we need to make sure to block out any other joins while we
2058 	 * commit the transaction.  We could have started a join before setting
2059 	 * COMMIT_DOING so make sure to wait for num_writers to == 1 again.
2060 	 */
2061 	spin_lock(&fs_info->trans_lock);
2062 	cur_trans->state = TRANS_STATE_COMMIT_DOING;
2063 	spin_unlock(&fs_info->trans_lock);
2064 	wait_event(cur_trans->writer_wait,
2065 		   atomic_read(&cur_trans->num_writers) == 1);
2066 
2067 	/* ->aborted might be set after the previous check, so check it */
2068 	if (unlikely(READ_ONCE(cur_trans->aborted))) {
2069 		ret = cur_trans->aborted;
2070 		goto scrub_continue;
2071 	}
2072 	/*
2073 	 * the reloc mutex makes sure that we stop
2074 	 * the balancing code from coming in and moving
2075 	 * extents around in the middle of the commit
2076 	 */
2077 	mutex_lock(&fs_info->reloc_mutex);
2078 
2079 	/*
2080 	 * We needn't worry about the delayed items because we will
2081 	 * deal with them in create_pending_snapshot(), which is the
2082 	 * core function of the snapshot creation.
2083 	 */
2084 	ret = create_pending_snapshots(trans);
2085 	if (ret) {
2086 		mutex_unlock(&fs_info->reloc_mutex);
2087 		goto scrub_continue;
2088 	}
2089 
2090 	/*
2091 	 * We insert the dir indexes of the snapshots and update the inode
2092 	 * of the snapshots' parents after the snapshot creation, so there
2093 	 * are some delayed items which are not dealt with. Now deal with
2094 	 * them.
2095 	 *
2096 	 * We needn't worry that this operation will corrupt the snapshots,
2097 	 * because all the tree which are snapshoted will be forced to COW
2098 	 * the nodes and leaves.
2099 	 */
2100 	ret = btrfs_run_delayed_items(trans);
2101 	if (ret) {
2102 		mutex_unlock(&fs_info->reloc_mutex);
2103 		goto scrub_continue;
2104 	}
2105 
2106 	ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
2107 	if (ret) {
2108 		mutex_unlock(&fs_info->reloc_mutex);
2109 		goto scrub_continue;
2110 	}
2111 
2112 	/*
2113 	 * make sure none of the code above managed to slip in a
2114 	 * delayed item
2115 	 */
2116 	btrfs_assert_delayed_root_empty(fs_info);
2117 
2118 	WARN_ON(cur_trans != trans->transaction);
2119 
2120 	/* btrfs_commit_tree_roots is responsible for getting the
2121 	 * various roots consistent with each other.  Every pointer
2122 	 * in the tree of tree roots has to point to the most up to date
2123 	 * root for every subvolume and other tree.  So, we have to keep
2124 	 * the tree logging code from jumping in and changing any
2125 	 * of the trees.
2126 	 *
2127 	 * At this point in the commit, there can't be any tree-log
2128 	 * writers, but a little lower down we drop the trans mutex
2129 	 * and let new people in.  By holding the tree_log_mutex
2130 	 * from now until after the super is written, we avoid races
2131 	 * with the tree-log code.
2132 	 */
2133 	mutex_lock(&fs_info->tree_log_mutex);
2134 
2135 	ret = commit_fs_roots(trans);
2136 	if (ret) {
2137 		mutex_unlock(&fs_info->tree_log_mutex);
2138 		mutex_unlock(&fs_info->reloc_mutex);
2139 		goto scrub_continue;
2140 	}
2141 
2142 	/*
2143 	 * Since the transaction is done, we can apply the pending changes
2144 	 * before the next transaction.
2145 	 */
2146 	btrfs_apply_pending_changes(fs_info);
2147 
2148 	/* commit_fs_roots gets rid of all the tree log roots, it is now
2149 	 * safe to free the root of tree log roots
2150 	 */
2151 	btrfs_free_log_root_tree(trans, fs_info);
2152 
2153 	/*
2154 	 * commit_fs_roots() can call btrfs_save_ino_cache(), which generates
2155 	 * new delayed refs. Must handle them or qgroup can be wrong.
2156 	 */
2157 	ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
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 fs roots are all committed, we can get a quite accurate
2166 	 * new_roots. So let's do quota accounting.
2167 	 */
2168 	ret = btrfs_qgroup_account_extents(trans);
2169 	if (ret < 0) {
2170 		mutex_unlock(&fs_info->tree_log_mutex);
2171 		mutex_unlock(&fs_info->reloc_mutex);
2172 		goto scrub_continue;
2173 	}
2174 
2175 	ret = commit_cowonly_roots(trans);
2176 	if (ret) {
2177 		mutex_unlock(&fs_info->tree_log_mutex);
2178 		mutex_unlock(&fs_info->reloc_mutex);
2179 		goto scrub_continue;
2180 	}
2181 
2182 	/*
2183 	 * The tasks which save the space cache and inode cache may also
2184 	 * update ->aborted, check it.
2185 	 */
2186 	if (unlikely(READ_ONCE(cur_trans->aborted))) {
2187 		ret = cur_trans->aborted;
2188 		mutex_unlock(&fs_info->tree_log_mutex);
2189 		mutex_unlock(&fs_info->reloc_mutex);
2190 		goto scrub_continue;
2191 	}
2192 
2193 	btrfs_prepare_extent_commit(fs_info);
2194 
2195 	cur_trans = fs_info->running_transaction;
2196 
2197 	btrfs_set_root_node(&fs_info->tree_root->root_item,
2198 			    fs_info->tree_root->node);
2199 	list_add_tail(&fs_info->tree_root->dirty_list,
2200 		      &cur_trans->switch_commits);
2201 
2202 	btrfs_set_root_node(&fs_info->chunk_root->root_item,
2203 			    fs_info->chunk_root->node);
2204 	list_add_tail(&fs_info->chunk_root->dirty_list,
2205 		      &cur_trans->switch_commits);
2206 
2207 	switch_commit_roots(cur_trans);
2208 
2209 	ASSERT(list_empty(&cur_trans->dirty_bgs));
2210 	ASSERT(list_empty(&cur_trans->io_bgs));
2211 	update_super_roots(fs_info);
2212 
2213 	btrfs_set_super_log_root(fs_info->super_copy, 0);
2214 	btrfs_set_super_log_root_level(fs_info->super_copy, 0);
2215 	memcpy(fs_info->super_for_commit, fs_info->super_copy,
2216 	       sizeof(*fs_info->super_copy));
2217 
2218 	btrfs_update_commit_device_size(fs_info);
2219 	btrfs_update_commit_device_bytes_used(cur_trans);
2220 
2221 	clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags);
2222 	clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags);
2223 
2224 	btrfs_trans_release_chunk_metadata(trans);
2225 
2226 	spin_lock(&fs_info->trans_lock);
2227 	cur_trans->state = TRANS_STATE_UNBLOCKED;
2228 	fs_info->running_transaction = NULL;
2229 	spin_unlock(&fs_info->trans_lock);
2230 	mutex_unlock(&fs_info->reloc_mutex);
2231 
2232 	wake_up(&fs_info->transaction_wait);
2233 
2234 	ret = btrfs_write_and_wait_transaction(trans);
2235 	if (ret) {
2236 		btrfs_handle_fs_error(fs_info, ret,
2237 				      "Error while writing out transaction");
2238 		mutex_unlock(&fs_info->tree_log_mutex);
2239 		goto scrub_continue;
2240 	}
2241 
2242 	ret = write_all_supers(fs_info, 0);
2243 	/*
2244 	 * the super is written, we can safely allow the tree-loggers
2245 	 * to go about their business
2246 	 */
2247 	mutex_unlock(&fs_info->tree_log_mutex);
2248 	if (ret)
2249 		goto scrub_continue;
2250 
2251 	btrfs_finish_extent_commit(trans);
2252 
2253 	if (test_bit(BTRFS_TRANS_HAVE_FREE_BGS, &cur_trans->flags))
2254 		btrfs_clear_space_info_full(fs_info);
2255 
2256 	fs_info->last_trans_committed = cur_trans->transid;
2257 	/*
2258 	 * We needn't acquire the lock here because there is no other task
2259 	 * which can change it.
2260 	 */
2261 	cur_trans->state = TRANS_STATE_COMPLETED;
2262 	wake_up(&cur_trans->commit_wait);
2263 	clear_bit(BTRFS_FS_NEED_ASYNC_COMMIT, &fs_info->flags);
2264 
2265 	spin_lock(&fs_info->trans_lock);
2266 	list_del_init(&cur_trans->list);
2267 	spin_unlock(&fs_info->trans_lock);
2268 
2269 	btrfs_put_transaction(cur_trans);
2270 	btrfs_put_transaction(cur_trans);
2271 
2272 	if (trans->type & __TRANS_FREEZABLE)
2273 		sb_end_intwrite(fs_info->sb);
2274 
2275 	trace_btrfs_transaction_commit(trans->root);
2276 
2277 	btrfs_scrub_continue(fs_info);
2278 
2279 	if (current->journal_info == trans)
2280 		current->journal_info = NULL;
2281 
2282 	kmem_cache_free(btrfs_trans_handle_cachep, trans);
2283 
2284 	/*
2285 	 * If fs has been frozen, we can not handle delayed iputs, otherwise
2286 	 * it'll result in deadlock about SB_FREEZE_FS.
2287 	 */
2288 	if (current != fs_info->transaction_kthread &&
2289 	    current != fs_info->cleaner_kthread &&
2290 	    !test_bit(BTRFS_FS_FROZEN, &fs_info->flags))
2291 		btrfs_run_delayed_iputs(fs_info);
2292 
2293 	return ret;
2294 
2295 scrub_continue:
2296 	btrfs_scrub_continue(fs_info);
2297 cleanup_transaction:
2298 	btrfs_trans_release_metadata(trans);
2299 	btrfs_trans_release_chunk_metadata(trans);
2300 	trans->block_rsv = NULL;
2301 	btrfs_warn(fs_info, "Skipping commit of aborted transaction.");
2302 	if (current->journal_info == trans)
2303 		current->journal_info = NULL;
2304 	cleanup_transaction(trans, ret);
2305 
2306 	return ret;
2307 }
2308 
2309 /*
2310  * return < 0 if error
2311  * 0 if there are no more dead_roots at the time of call
2312  * 1 there are more to be processed, call me again
2313  *
2314  * The return value indicates there are certainly more snapshots to delete, but
2315  * if there comes a new one during processing, it may return 0. We don't mind,
2316  * because btrfs_commit_super will poke cleaner thread and it will process it a
2317  * few seconds later.
2318  */
2319 int btrfs_clean_one_deleted_snapshot(struct btrfs_root *root)
2320 {
2321 	int ret;
2322 	struct btrfs_fs_info *fs_info = root->fs_info;
2323 
2324 	spin_lock(&fs_info->trans_lock);
2325 	if (list_empty(&fs_info->dead_roots)) {
2326 		spin_unlock(&fs_info->trans_lock);
2327 		return 0;
2328 	}
2329 	root = list_first_entry(&fs_info->dead_roots,
2330 			struct btrfs_root, root_list);
2331 	list_del_init(&root->root_list);
2332 	spin_unlock(&fs_info->trans_lock);
2333 
2334 	btrfs_debug(fs_info, "cleaner removing %llu", root->objectid);
2335 
2336 	btrfs_kill_all_delayed_nodes(root);
2337 
2338 	if (btrfs_header_backref_rev(root->node) <
2339 			BTRFS_MIXED_BACKREF_REV)
2340 		ret = btrfs_drop_snapshot(root, NULL, 0, 0);
2341 	else
2342 		ret = btrfs_drop_snapshot(root, NULL, 1, 0);
2343 
2344 	return (ret < 0) ? 0 : 1;
2345 }
2346 
2347 void btrfs_apply_pending_changes(struct btrfs_fs_info *fs_info)
2348 {
2349 	unsigned long prev;
2350 	unsigned long bit;
2351 
2352 	prev = xchg(&fs_info->pending_changes, 0);
2353 	if (!prev)
2354 		return;
2355 
2356 	bit = 1 << BTRFS_PENDING_SET_INODE_MAP_CACHE;
2357 	if (prev & bit)
2358 		btrfs_set_opt(fs_info->mount_opt, INODE_MAP_CACHE);
2359 	prev &= ~bit;
2360 
2361 	bit = 1 << BTRFS_PENDING_CLEAR_INODE_MAP_CACHE;
2362 	if (prev & bit)
2363 		btrfs_clear_opt(fs_info->mount_opt, INODE_MAP_CACHE);
2364 	prev &= ~bit;
2365 
2366 	bit = 1 << BTRFS_PENDING_COMMIT;
2367 	if (prev & bit)
2368 		btrfs_debug(fs_info, "pending commit done");
2369 	prev &= ~bit;
2370 
2371 	if (prev)
2372 		btrfs_warn(fs_info,
2373 			"unknown pending changes left 0x%lx, ignoring", prev);
2374 }
2375