xref: /linux/fs/btrfs/transaction.c (revision e5c86679d5e864947a52fb31e45a425dea3e7fa9)
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 		  bool enforce_qgroups)
479 {
480 	struct btrfs_fs_info *fs_info = root->fs_info;
481 
482 	struct btrfs_trans_handle *h;
483 	struct btrfs_transaction *cur_trans;
484 	u64 num_bytes = 0;
485 	u64 qgroup_reserved = 0;
486 	bool reloc_reserved = false;
487 	int ret;
488 
489 	/* Send isn't supposed to start transactions. */
490 	ASSERT(current->journal_info != BTRFS_SEND_TRANS_STUB);
491 
492 	if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
493 		return ERR_PTR(-EROFS);
494 
495 	if (current->journal_info) {
496 		WARN_ON(type & TRANS_EXTWRITERS);
497 		h = current->journal_info;
498 		h->use_count++;
499 		WARN_ON(h->use_count > 2);
500 		h->orig_rsv = h->block_rsv;
501 		h->block_rsv = NULL;
502 		goto got_it;
503 	}
504 
505 	/*
506 	 * Do the reservation before we join the transaction so we can do all
507 	 * the appropriate flushing if need be.
508 	 */
509 	if (num_items && root != fs_info->chunk_root) {
510 		qgroup_reserved = num_items * fs_info->nodesize;
511 		ret = btrfs_qgroup_reserve_meta(root, qgroup_reserved,
512 						enforce_qgroups);
513 		if (ret)
514 			return ERR_PTR(ret);
515 
516 		num_bytes = btrfs_calc_trans_metadata_size(fs_info, num_items);
517 		/*
518 		 * Do the reservation for the relocation root creation
519 		 */
520 		if (need_reserve_reloc_root(root)) {
521 			num_bytes += fs_info->nodesize;
522 			reloc_reserved = true;
523 		}
524 
525 		ret = btrfs_block_rsv_add(root, &fs_info->trans_block_rsv,
526 					  num_bytes, flush);
527 		if (ret)
528 			goto reserve_fail;
529 	}
530 again:
531 	h = kmem_cache_zalloc(btrfs_trans_handle_cachep, GFP_NOFS);
532 	if (!h) {
533 		ret = -ENOMEM;
534 		goto alloc_fail;
535 	}
536 
537 	/*
538 	 * If we are JOIN_NOLOCK we're already committing a transaction and
539 	 * waiting on this guy, so we don't need to do the sb_start_intwrite
540 	 * because we're already holding a ref.  We need this because we could
541 	 * have raced in and did an fsync() on a file which can kick a commit
542 	 * and then we deadlock with somebody doing a freeze.
543 	 *
544 	 * If we are ATTACH, it means we just want to catch the current
545 	 * transaction and commit it, so we needn't do sb_start_intwrite().
546 	 */
547 	if (type & __TRANS_FREEZABLE)
548 		sb_start_intwrite(fs_info->sb);
549 
550 	if (may_wait_transaction(fs_info, type))
551 		wait_current_trans(fs_info);
552 
553 	do {
554 		ret = join_transaction(fs_info, type);
555 		if (ret == -EBUSY) {
556 			wait_current_trans(fs_info);
557 			if (unlikely(type == TRANS_ATTACH))
558 				ret = -ENOENT;
559 		}
560 	} while (ret == -EBUSY);
561 
562 	if (ret < 0)
563 		goto join_fail;
564 
565 	cur_trans = fs_info->running_transaction;
566 
567 	h->transid = cur_trans->transid;
568 	h->transaction = cur_trans;
569 	h->root = root;
570 	h->use_count = 1;
571 	h->fs_info = root->fs_info;
572 
573 	h->type = type;
574 	h->can_flush_pending_bgs = true;
575 	INIT_LIST_HEAD(&h->qgroup_ref_list);
576 	INIT_LIST_HEAD(&h->new_bgs);
577 
578 	smp_mb();
579 	if (cur_trans->state >= TRANS_STATE_BLOCKED &&
580 	    may_wait_transaction(fs_info, type)) {
581 		current->journal_info = h;
582 		btrfs_commit_transaction(h);
583 		goto again;
584 	}
585 
586 	if (num_bytes) {
587 		trace_btrfs_space_reservation(fs_info, "transaction",
588 					      h->transid, num_bytes, 1);
589 		h->block_rsv = &fs_info->trans_block_rsv;
590 		h->bytes_reserved = num_bytes;
591 		h->reloc_reserved = reloc_reserved;
592 	}
593 
594 got_it:
595 	btrfs_record_root_in_trans(h, root);
596 
597 	if (!current->journal_info && type != TRANS_USERSPACE)
598 		current->journal_info = h;
599 	return h;
600 
601 join_fail:
602 	if (type & __TRANS_FREEZABLE)
603 		sb_end_intwrite(fs_info->sb);
604 	kmem_cache_free(btrfs_trans_handle_cachep, h);
605 alloc_fail:
606 	if (num_bytes)
607 		btrfs_block_rsv_release(fs_info, &fs_info->trans_block_rsv,
608 					num_bytes);
609 reserve_fail:
610 	btrfs_qgroup_free_meta(root, qgroup_reserved);
611 	return ERR_PTR(ret);
612 }
613 
614 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
615 						   unsigned int num_items)
616 {
617 	return start_transaction(root, num_items, TRANS_START,
618 				 BTRFS_RESERVE_FLUSH_ALL, true);
619 }
620 
621 struct btrfs_trans_handle *btrfs_start_transaction_fallback_global_rsv(
622 					struct btrfs_root *root,
623 					unsigned int num_items,
624 					int min_factor)
625 {
626 	struct btrfs_fs_info *fs_info = root->fs_info;
627 	struct btrfs_trans_handle *trans;
628 	u64 num_bytes;
629 	int ret;
630 
631 	/*
632 	 * We have two callers: unlink and block group removal.  The
633 	 * former should succeed even if we will temporarily exceed
634 	 * quota and the latter operates on the extent root so
635 	 * qgroup enforcement is ignored anyway.
636 	 */
637 	trans = start_transaction(root, num_items, TRANS_START,
638 				  BTRFS_RESERVE_FLUSH_ALL, false);
639 	if (!IS_ERR(trans) || PTR_ERR(trans) != -ENOSPC)
640 		return trans;
641 
642 	trans = btrfs_start_transaction(root, 0);
643 	if (IS_ERR(trans))
644 		return trans;
645 
646 	num_bytes = btrfs_calc_trans_metadata_size(fs_info, num_items);
647 	ret = btrfs_cond_migrate_bytes(fs_info, &fs_info->trans_block_rsv,
648 				       num_bytes, min_factor);
649 	if (ret) {
650 		btrfs_end_transaction(trans);
651 		return ERR_PTR(ret);
652 	}
653 
654 	trans->block_rsv = &fs_info->trans_block_rsv;
655 	trans->bytes_reserved = num_bytes;
656 	trace_btrfs_space_reservation(fs_info, "transaction",
657 				      trans->transid, num_bytes, 1);
658 
659 	return trans;
660 }
661 
662 struct btrfs_trans_handle *btrfs_start_transaction_lflush(
663 					struct btrfs_root *root,
664 					unsigned int num_items)
665 {
666 	return start_transaction(root, num_items, TRANS_START,
667 				 BTRFS_RESERVE_FLUSH_LIMIT, true);
668 }
669 
670 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
671 {
672 	return start_transaction(root, 0, TRANS_JOIN, BTRFS_RESERVE_NO_FLUSH,
673 				 true);
674 }
675 
676 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root)
677 {
678 	return start_transaction(root, 0, TRANS_JOIN_NOLOCK,
679 				 BTRFS_RESERVE_NO_FLUSH, true);
680 }
681 
682 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *root)
683 {
684 	return start_transaction(root, 0, TRANS_USERSPACE,
685 				 BTRFS_RESERVE_NO_FLUSH, true);
686 }
687 
688 /*
689  * btrfs_attach_transaction() - catch the running transaction
690  *
691  * It is used when we want to commit the current the transaction, but
692  * don't want to start a new one.
693  *
694  * Note: If this function return -ENOENT, it just means there is no
695  * running transaction. But it is possible that the inactive transaction
696  * is still in the memory, not fully on disk. If you hope there is no
697  * inactive transaction in the fs when -ENOENT is returned, you should
698  * invoke
699  *     btrfs_attach_transaction_barrier()
700  */
701 struct btrfs_trans_handle *btrfs_attach_transaction(struct btrfs_root *root)
702 {
703 	return start_transaction(root, 0, TRANS_ATTACH,
704 				 BTRFS_RESERVE_NO_FLUSH, true);
705 }
706 
707 /*
708  * btrfs_attach_transaction_barrier() - catch the running transaction
709  *
710  * It is similar to the above function, the differentia is this one
711  * will wait for all the inactive transactions until they fully
712  * complete.
713  */
714 struct btrfs_trans_handle *
715 btrfs_attach_transaction_barrier(struct btrfs_root *root)
716 {
717 	struct btrfs_trans_handle *trans;
718 
719 	trans = start_transaction(root, 0, TRANS_ATTACH,
720 				  BTRFS_RESERVE_NO_FLUSH, true);
721 	if (IS_ERR(trans) && PTR_ERR(trans) == -ENOENT)
722 		btrfs_wait_for_commit(root->fs_info, 0);
723 
724 	return trans;
725 }
726 
727 /* wait for a transaction commit to be fully complete */
728 static noinline void wait_for_commit(struct btrfs_transaction *commit)
729 {
730 	wait_event(commit->commit_wait, commit->state == TRANS_STATE_COMPLETED);
731 }
732 
733 int btrfs_wait_for_commit(struct btrfs_fs_info *fs_info, u64 transid)
734 {
735 	struct btrfs_transaction *cur_trans = NULL, *t;
736 	int ret = 0;
737 
738 	if (transid) {
739 		if (transid <= fs_info->last_trans_committed)
740 			goto out;
741 
742 		/* find specified transaction */
743 		spin_lock(&fs_info->trans_lock);
744 		list_for_each_entry(t, &fs_info->trans_list, list) {
745 			if (t->transid == transid) {
746 				cur_trans = t;
747 				atomic_inc(&cur_trans->use_count);
748 				ret = 0;
749 				break;
750 			}
751 			if (t->transid > transid) {
752 				ret = 0;
753 				break;
754 			}
755 		}
756 		spin_unlock(&fs_info->trans_lock);
757 
758 		/*
759 		 * The specified transaction doesn't exist, or we
760 		 * raced with btrfs_commit_transaction
761 		 */
762 		if (!cur_trans) {
763 			if (transid > fs_info->last_trans_committed)
764 				ret = -EINVAL;
765 			goto out;
766 		}
767 	} else {
768 		/* find newest transaction that is committing | committed */
769 		spin_lock(&fs_info->trans_lock);
770 		list_for_each_entry_reverse(t, &fs_info->trans_list,
771 					    list) {
772 			if (t->state >= TRANS_STATE_COMMIT_START) {
773 				if (t->state == TRANS_STATE_COMPLETED)
774 					break;
775 				cur_trans = t;
776 				atomic_inc(&cur_trans->use_count);
777 				break;
778 			}
779 		}
780 		spin_unlock(&fs_info->trans_lock);
781 		if (!cur_trans)
782 			goto out;  /* nothing committing|committed */
783 	}
784 
785 	wait_for_commit(cur_trans);
786 	btrfs_put_transaction(cur_trans);
787 out:
788 	return ret;
789 }
790 
791 void btrfs_throttle(struct btrfs_fs_info *fs_info)
792 {
793 	if (!atomic_read(&fs_info->open_ioctl_trans))
794 		wait_current_trans(fs_info);
795 }
796 
797 static int should_end_transaction(struct btrfs_trans_handle *trans)
798 {
799 	struct btrfs_fs_info *fs_info = trans->fs_info;
800 
801 	if (fs_info->global_block_rsv.space_info->full &&
802 	    btrfs_check_space_for_delayed_refs(trans, fs_info))
803 		return 1;
804 
805 	return !!btrfs_block_rsv_check(&fs_info->global_block_rsv, 5);
806 }
807 
808 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans)
809 {
810 	struct btrfs_transaction *cur_trans = trans->transaction;
811 	struct btrfs_fs_info *fs_info = trans->fs_info;
812 	int updates;
813 	int err;
814 
815 	smp_mb();
816 	if (cur_trans->state >= TRANS_STATE_BLOCKED ||
817 	    cur_trans->delayed_refs.flushing)
818 		return 1;
819 
820 	updates = trans->delayed_ref_updates;
821 	trans->delayed_ref_updates = 0;
822 	if (updates) {
823 		err = btrfs_run_delayed_refs(trans, fs_info, updates * 2);
824 		if (err) /* Error code will also eval true */
825 			return err;
826 	}
827 
828 	return should_end_transaction(trans);
829 }
830 
831 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
832 				   int throttle)
833 {
834 	struct btrfs_fs_info *info = trans->fs_info;
835 	struct btrfs_transaction *cur_trans = trans->transaction;
836 	u64 transid = trans->transid;
837 	unsigned long cur = trans->delayed_ref_updates;
838 	int lock = (trans->type != TRANS_JOIN_NOLOCK);
839 	int err = 0;
840 	int must_run_delayed_refs = 0;
841 
842 	if (trans->use_count > 1) {
843 		trans->use_count--;
844 		trans->block_rsv = trans->orig_rsv;
845 		return 0;
846 	}
847 
848 	btrfs_trans_release_metadata(trans, info);
849 	trans->block_rsv = NULL;
850 
851 	if (!list_empty(&trans->new_bgs))
852 		btrfs_create_pending_block_groups(trans, info);
853 
854 	trans->delayed_ref_updates = 0;
855 	if (!trans->sync) {
856 		must_run_delayed_refs =
857 			btrfs_should_throttle_delayed_refs(trans, info);
858 		cur = max_t(unsigned long, cur, 32);
859 
860 		/*
861 		 * don't make the caller wait if they are from a NOLOCK
862 		 * or ATTACH transaction, it will deadlock with commit
863 		 */
864 		if (must_run_delayed_refs == 1 &&
865 		    (trans->type & (__TRANS_JOIN_NOLOCK | __TRANS_ATTACH)))
866 			must_run_delayed_refs = 2;
867 	}
868 
869 	btrfs_trans_release_metadata(trans, info);
870 	trans->block_rsv = NULL;
871 
872 	if (!list_empty(&trans->new_bgs))
873 		btrfs_create_pending_block_groups(trans, info);
874 
875 	btrfs_trans_release_chunk_metadata(trans);
876 
877 	if (lock && !atomic_read(&info->open_ioctl_trans) &&
878 	    should_end_transaction(trans) &&
879 	    READ_ONCE(cur_trans->state) == TRANS_STATE_RUNNING) {
880 		spin_lock(&info->trans_lock);
881 		if (cur_trans->state == TRANS_STATE_RUNNING)
882 			cur_trans->state = TRANS_STATE_BLOCKED;
883 		spin_unlock(&info->trans_lock);
884 	}
885 
886 	if (lock && READ_ONCE(cur_trans->state) == TRANS_STATE_BLOCKED) {
887 		if (throttle)
888 			return btrfs_commit_transaction(trans);
889 		else
890 			wake_up_process(info->transaction_kthread);
891 	}
892 
893 	if (trans->type & __TRANS_FREEZABLE)
894 		sb_end_intwrite(info->sb);
895 
896 	WARN_ON(cur_trans != info->running_transaction);
897 	WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
898 	atomic_dec(&cur_trans->num_writers);
899 	extwriter_counter_dec(cur_trans, trans->type);
900 
901 	/*
902 	 * Make sure counter is updated before we wake up waiters.
903 	 */
904 	smp_mb();
905 	if (waitqueue_active(&cur_trans->writer_wait))
906 		wake_up(&cur_trans->writer_wait);
907 	btrfs_put_transaction(cur_trans);
908 
909 	if (current->journal_info == trans)
910 		current->journal_info = NULL;
911 
912 	if (throttle)
913 		btrfs_run_delayed_iputs(info);
914 
915 	if (trans->aborted ||
916 	    test_bit(BTRFS_FS_STATE_ERROR, &info->fs_state)) {
917 		wake_up_process(info->transaction_kthread);
918 		err = -EIO;
919 	}
920 	assert_qgroups_uptodate(trans);
921 
922 	kmem_cache_free(btrfs_trans_handle_cachep, trans);
923 	if (must_run_delayed_refs) {
924 		btrfs_async_run_delayed_refs(info, cur, transid,
925 					     must_run_delayed_refs == 1);
926 	}
927 	return err;
928 }
929 
930 int btrfs_end_transaction(struct btrfs_trans_handle *trans)
931 {
932 	return __btrfs_end_transaction(trans, 0);
933 }
934 
935 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans)
936 {
937 	return __btrfs_end_transaction(trans, 1);
938 }
939 
940 /*
941  * when btree blocks are allocated, they have some corresponding bits set for
942  * them in one of two extent_io trees.  This is used to make sure all of
943  * those extents are sent to disk but does not wait on them
944  */
945 int btrfs_write_marked_extents(struct btrfs_fs_info *fs_info,
946 			       struct extent_io_tree *dirty_pages, int mark)
947 {
948 	int err = 0;
949 	int werr = 0;
950 	struct address_space *mapping = fs_info->btree_inode->i_mapping;
951 	struct extent_state *cached_state = NULL;
952 	u64 start = 0;
953 	u64 end;
954 
955 	while (!find_first_extent_bit(dirty_pages, start, &start, &end,
956 				      mark, &cached_state)) {
957 		bool wait_writeback = false;
958 
959 		err = convert_extent_bit(dirty_pages, start, end,
960 					 EXTENT_NEED_WAIT,
961 					 mark, &cached_state);
962 		/*
963 		 * convert_extent_bit can return -ENOMEM, which is most of the
964 		 * time a temporary error. So when it happens, ignore the error
965 		 * and wait for writeback of this range to finish - because we
966 		 * failed to set the bit EXTENT_NEED_WAIT for the range, a call
967 		 * to __btrfs_wait_marked_extents() would not know that
968 		 * writeback for this range started and therefore wouldn't
969 		 * wait for it to finish - we don't want to commit a
970 		 * superblock that points to btree nodes/leafs for which
971 		 * writeback hasn't finished yet (and without errors).
972 		 * We cleanup any entries left in the io tree when committing
973 		 * the transaction (through clear_btree_io_tree()).
974 		 */
975 		if (err == -ENOMEM) {
976 			err = 0;
977 			wait_writeback = true;
978 		}
979 		if (!err)
980 			err = filemap_fdatawrite_range(mapping, start, end);
981 		if (err)
982 			werr = err;
983 		else if (wait_writeback)
984 			werr = filemap_fdatawait_range(mapping, start, end);
985 		free_extent_state(cached_state);
986 		cached_state = NULL;
987 		cond_resched();
988 		start = end + 1;
989 	}
990 	return werr;
991 }
992 
993 /*
994  * when btree blocks are allocated, they have some corresponding bits set for
995  * them in one of two extent_io trees.  This is used to make sure all of
996  * those extents are on disk for transaction or log commit.  We wait
997  * on all the pages and clear them from the dirty pages state tree
998  */
999 static int __btrfs_wait_marked_extents(struct btrfs_fs_info *fs_info,
1000 				       struct extent_io_tree *dirty_pages)
1001 {
1002 	int err = 0;
1003 	int werr = 0;
1004 	struct address_space *mapping = fs_info->btree_inode->i_mapping;
1005 	struct extent_state *cached_state = NULL;
1006 	u64 start = 0;
1007 	u64 end;
1008 
1009 	while (!find_first_extent_bit(dirty_pages, start, &start, &end,
1010 				      EXTENT_NEED_WAIT, &cached_state)) {
1011 		/*
1012 		 * Ignore -ENOMEM errors returned by clear_extent_bit().
1013 		 * When committing the transaction, we'll remove any entries
1014 		 * left in the io tree. For a log commit, we don't remove them
1015 		 * after committing the log because the tree can be accessed
1016 		 * concurrently - we do it only at transaction commit time when
1017 		 * it's safe to do it (through clear_btree_io_tree()).
1018 		 */
1019 		err = clear_extent_bit(dirty_pages, start, end,
1020 				       EXTENT_NEED_WAIT,
1021 				       0, 0, &cached_state, GFP_NOFS);
1022 		if (err == -ENOMEM)
1023 			err = 0;
1024 		if (!err)
1025 			err = filemap_fdatawait_range(mapping, start, end);
1026 		if (err)
1027 			werr = err;
1028 		free_extent_state(cached_state);
1029 		cached_state = NULL;
1030 		cond_resched();
1031 		start = end + 1;
1032 	}
1033 	if (err)
1034 		werr = err;
1035 	return werr;
1036 }
1037 
1038 int btrfs_wait_extents(struct btrfs_fs_info *fs_info,
1039 		       struct extent_io_tree *dirty_pages)
1040 {
1041 	bool errors = false;
1042 	int err;
1043 
1044 	err = __btrfs_wait_marked_extents(fs_info, dirty_pages);
1045 	if (test_and_clear_bit(BTRFS_FS_BTREE_ERR, &fs_info->flags))
1046 		errors = true;
1047 
1048 	if (errors && !err)
1049 		err = -EIO;
1050 	return err;
1051 }
1052 
1053 int btrfs_wait_tree_log_extents(struct btrfs_root *log_root, int mark)
1054 {
1055 	struct btrfs_fs_info *fs_info = log_root->fs_info;
1056 	struct extent_io_tree *dirty_pages = &log_root->dirty_log_pages;
1057 	bool errors = false;
1058 	int err;
1059 
1060 	ASSERT(log_root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
1061 
1062 	err = __btrfs_wait_marked_extents(fs_info, dirty_pages);
1063 	if ((mark & EXTENT_DIRTY) &&
1064 	    test_and_clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags))
1065 		errors = true;
1066 
1067 	if ((mark & EXTENT_NEW) &&
1068 	    test_and_clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags))
1069 		errors = true;
1070 
1071 	if (errors && !err)
1072 		err = -EIO;
1073 	return err;
1074 }
1075 
1076 /*
1077  * when btree blocks are allocated, they have some corresponding bits set for
1078  * them in one of two extent_io trees.  This is used to make sure all of
1079  * those extents are on disk for transaction or log commit
1080  */
1081 static int btrfs_write_and_wait_marked_extents(struct btrfs_fs_info *fs_info,
1082 				struct extent_io_tree *dirty_pages, int mark)
1083 {
1084 	int ret;
1085 	int ret2;
1086 	struct blk_plug plug;
1087 
1088 	blk_start_plug(&plug);
1089 	ret = btrfs_write_marked_extents(fs_info, dirty_pages, mark);
1090 	blk_finish_plug(&plug);
1091 	ret2 = btrfs_wait_extents(fs_info, dirty_pages);
1092 
1093 	if (ret)
1094 		return ret;
1095 	if (ret2)
1096 		return ret2;
1097 	return 0;
1098 }
1099 
1100 static int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
1101 					    struct btrfs_fs_info *fs_info)
1102 {
1103 	int ret;
1104 
1105 	ret = btrfs_write_and_wait_marked_extents(fs_info,
1106 					   &trans->transaction->dirty_pages,
1107 					   EXTENT_DIRTY);
1108 	clear_btree_io_tree(&trans->transaction->dirty_pages);
1109 
1110 	return ret;
1111 }
1112 
1113 /*
1114  * this is used to update the root pointer in the tree of tree roots.
1115  *
1116  * But, in the case of the extent allocation tree, updating the root
1117  * pointer may allocate blocks which may change the root of the extent
1118  * allocation tree.
1119  *
1120  * So, this loops and repeats and makes sure the cowonly root didn't
1121  * change while the root pointer was being updated in the metadata.
1122  */
1123 static int update_cowonly_root(struct btrfs_trans_handle *trans,
1124 			       struct btrfs_root *root)
1125 {
1126 	int ret;
1127 	u64 old_root_bytenr;
1128 	u64 old_root_used;
1129 	struct btrfs_fs_info *fs_info = root->fs_info;
1130 	struct btrfs_root *tree_root = fs_info->tree_root;
1131 
1132 	old_root_used = btrfs_root_used(&root->root_item);
1133 
1134 	while (1) {
1135 		old_root_bytenr = btrfs_root_bytenr(&root->root_item);
1136 		if (old_root_bytenr == root->node->start &&
1137 		    old_root_used == btrfs_root_used(&root->root_item))
1138 			break;
1139 
1140 		btrfs_set_root_node(&root->root_item, root->node);
1141 		ret = btrfs_update_root(trans, tree_root,
1142 					&root->root_key,
1143 					&root->root_item);
1144 		if (ret)
1145 			return ret;
1146 
1147 		old_root_used = btrfs_root_used(&root->root_item);
1148 	}
1149 
1150 	return 0;
1151 }
1152 
1153 /*
1154  * update all the cowonly tree roots on disk
1155  *
1156  * The error handling in this function may not be obvious. Any of the
1157  * failures will cause the file system to go offline. We still need
1158  * to clean up the delayed refs.
1159  */
1160 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
1161 					 struct btrfs_fs_info *fs_info)
1162 {
1163 	struct list_head *dirty_bgs = &trans->transaction->dirty_bgs;
1164 	struct list_head *io_bgs = &trans->transaction->io_bgs;
1165 	struct list_head *next;
1166 	struct extent_buffer *eb;
1167 	int ret;
1168 
1169 	eb = btrfs_lock_root_node(fs_info->tree_root);
1170 	ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL,
1171 			      0, &eb);
1172 	btrfs_tree_unlock(eb);
1173 	free_extent_buffer(eb);
1174 
1175 	if (ret)
1176 		return ret;
1177 
1178 	ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1179 	if (ret)
1180 		return ret;
1181 
1182 	ret = btrfs_run_dev_stats(trans, fs_info);
1183 	if (ret)
1184 		return ret;
1185 	ret = btrfs_run_dev_replace(trans, fs_info);
1186 	if (ret)
1187 		return ret;
1188 	ret = btrfs_run_qgroups(trans, fs_info);
1189 	if (ret)
1190 		return ret;
1191 
1192 	ret = btrfs_setup_space_cache(trans, fs_info);
1193 	if (ret)
1194 		return ret;
1195 
1196 	/* run_qgroups might have added some more refs */
1197 	ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1198 	if (ret)
1199 		return ret;
1200 again:
1201 	while (!list_empty(&fs_info->dirty_cowonly_roots)) {
1202 		struct btrfs_root *root;
1203 		next = fs_info->dirty_cowonly_roots.next;
1204 		list_del_init(next);
1205 		root = list_entry(next, struct btrfs_root, dirty_list);
1206 		clear_bit(BTRFS_ROOT_DIRTY, &root->state);
1207 
1208 		if (root != fs_info->extent_root)
1209 			list_add_tail(&root->dirty_list,
1210 				      &trans->transaction->switch_commits);
1211 		ret = update_cowonly_root(trans, root);
1212 		if (ret)
1213 			return ret;
1214 		ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1215 		if (ret)
1216 			return ret;
1217 	}
1218 
1219 	while (!list_empty(dirty_bgs) || !list_empty(io_bgs)) {
1220 		ret = btrfs_write_dirty_block_groups(trans, fs_info);
1221 		if (ret)
1222 			return ret;
1223 		ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1224 		if (ret)
1225 			return ret;
1226 	}
1227 
1228 	if (!list_empty(&fs_info->dirty_cowonly_roots))
1229 		goto again;
1230 
1231 	list_add_tail(&fs_info->extent_root->dirty_list,
1232 		      &trans->transaction->switch_commits);
1233 	btrfs_after_dev_replace_commit(fs_info);
1234 
1235 	return 0;
1236 }
1237 
1238 /*
1239  * dead roots are old snapshots that need to be deleted.  This allocates
1240  * a dirty root struct and adds it into the list of dead roots that need to
1241  * be deleted
1242  */
1243 void btrfs_add_dead_root(struct btrfs_root *root)
1244 {
1245 	struct btrfs_fs_info *fs_info = root->fs_info;
1246 
1247 	spin_lock(&fs_info->trans_lock);
1248 	if (list_empty(&root->root_list))
1249 		list_add_tail(&root->root_list, &fs_info->dead_roots);
1250 	spin_unlock(&fs_info->trans_lock);
1251 }
1252 
1253 /*
1254  * update all the cowonly tree roots on disk
1255  */
1256 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
1257 				    struct btrfs_fs_info *fs_info)
1258 {
1259 	struct btrfs_root *gang[8];
1260 	int i;
1261 	int ret;
1262 	int err = 0;
1263 
1264 	spin_lock(&fs_info->fs_roots_radix_lock);
1265 	while (1) {
1266 		ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
1267 						 (void **)gang, 0,
1268 						 ARRAY_SIZE(gang),
1269 						 BTRFS_ROOT_TRANS_TAG);
1270 		if (ret == 0)
1271 			break;
1272 		for (i = 0; i < ret; i++) {
1273 			struct btrfs_root *root = gang[i];
1274 			radix_tree_tag_clear(&fs_info->fs_roots_radix,
1275 					(unsigned long)root->root_key.objectid,
1276 					BTRFS_ROOT_TRANS_TAG);
1277 			spin_unlock(&fs_info->fs_roots_radix_lock);
1278 
1279 			btrfs_free_log(trans, root);
1280 			btrfs_update_reloc_root(trans, root);
1281 			btrfs_orphan_commit_root(trans, root);
1282 
1283 			btrfs_save_ino_cache(root, trans);
1284 
1285 			/* see comments in should_cow_block() */
1286 			clear_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1287 			smp_mb__after_atomic();
1288 
1289 			if (root->commit_root != root->node) {
1290 				list_add_tail(&root->dirty_list,
1291 					&trans->transaction->switch_commits);
1292 				btrfs_set_root_node(&root->root_item,
1293 						    root->node);
1294 			}
1295 
1296 			err = btrfs_update_root(trans, fs_info->tree_root,
1297 						&root->root_key,
1298 						&root->root_item);
1299 			spin_lock(&fs_info->fs_roots_radix_lock);
1300 			if (err)
1301 				break;
1302 			btrfs_qgroup_free_meta_all(root);
1303 		}
1304 	}
1305 	spin_unlock(&fs_info->fs_roots_radix_lock);
1306 	return err;
1307 }
1308 
1309 /*
1310  * defrag a given btree.
1311  * Every leaf in the btree is read and defragged.
1312  */
1313 int btrfs_defrag_root(struct btrfs_root *root)
1314 {
1315 	struct btrfs_fs_info *info = root->fs_info;
1316 	struct btrfs_trans_handle *trans;
1317 	int ret;
1318 
1319 	if (test_and_set_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state))
1320 		return 0;
1321 
1322 	while (1) {
1323 		trans = btrfs_start_transaction(root, 0);
1324 		if (IS_ERR(trans))
1325 			return PTR_ERR(trans);
1326 
1327 		ret = btrfs_defrag_leaves(trans, root);
1328 
1329 		btrfs_end_transaction(trans);
1330 		btrfs_btree_balance_dirty(info);
1331 		cond_resched();
1332 
1333 		if (btrfs_fs_closing(info) || ret != -EAGAIN)
1334 			break;
1335 
1336 		if (btrfs_defrag_cancelled(info)) {
1337 			btrfs_debug(info, "defrag_root cancelled");
1338 			ret = -EAGAIN;
1339 			break;
1340 		}
1341 	}
1342 	clear_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state);
1343 	return ret;
1344 }
1345 
1346 /*
1347  * Do all special snapshot related qgroup dirty hack.
1348  *
1349  * Will do all needed qgroup inherit and dirty hack like switch commit
1350  * roots inside one transaction and write all btree into disk, to make
1351  * qgroup works.
1352  */
1353 static int qgroup_account_snapshot(struct btrfs_trans_handle *trans,
1354 				   struct btrfs_root *src,
1355 				   struct btrfs_root *parent,
1356 				   struct btrfs_qgroup_inherit *inherit,
1357 				   u64 dst_objectid)
1358 {
1359 	struct btrfs_fs_info *fs_info = src->fs_info;
1360 	int ret;
1361 
1362 	/*
1363 	 * Save some performance in the case that qgroups are not
1364 	 * enabled. If this check races with the ioctl, rescan will
1365 	 * kick in anyway.
1366 	 */
1367 	if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
1368 		return 0;
1369 
1370 	/*
1371 	 * We are going to commit transaction, see btrfs_commit_transaction()
1372 	 * comment for reason locking tree_log_mutex
1373 	 */
1374 	mutex_lock(&fs_info->tree_log_mutex);
1375 
1376 	ret = commit_fs_roots(trans, fs_info);
1377 	if (ret)
1378 		goto out;
1379 	ret = btrfs_qgroup_prepare_account_extents(trans, fs_info);
1380 	if (ret < 0)
1381 		goto out;
1382 	ret = btrfs_qgroup_account_extents(trans, fs_info);
1383 	if (ret < 0)
1384 		goto out;
1385 
1386 	/* Now qgroup are all updated, we can inherit it to new qgroups */
1387 	ret = btrfs_qgroup_inherit(trans, fs_info,
1388 				   src->root_key.objectid, dst_objectid,
1389 				   inherit);
1390 	if (ret < 0)
1391 		goto out;
1392 
1393 	/*
1394 	 * Now we do a simplified commit transaction, which will:
1395 	 * 1) commit all subvolume and extent tree
1396 	 *    To ensure all subvolume and extent tree have a valid
1397 	 *    commit_root to accounting later insert_dir_item()
1398 	 * 2) write all btree blocks onto disk
1399 	 *    This is to make sure later btree modification will be cowed
1400 	 *    Or commit_root can be populated and cause wrong qgroup numbers
1401 	 * In this simplified commit, we don't really care about other trees
1402 	 * like chunk and root tree, as they won't affect qgroup.
1403 	 * And we don't write super to avoid half committed status.
1404 	 */
1405 	ret = commit_cowonly_roots(trans, fs_info);
1406 	if (ret)
1407 		goto out;
1408 	switch_commit_roots(trans->transaction, fs_info);
1409 	ret = btrfs_write_and_wait_transaction(trans, fs_info);
1410 	if (ret)
1411 		btrfs_handle_fs_error(fs_info, ret,
1412 			"Error while writing out transaction for qgroup");
1413 
1414 out:
1415 	mutex_unlock(&fs_info->tree_log_mutex);
1416 
1417 	/*
1418 	 * Force parent root to be updated, as we recorded it before so its
1419 	 * last_trans == cur_transid.
1420 	 * Or it won't be committed again onto disk after later
1421 	 * insert_dir_item()
1422 	 */
1423 	if (!ret)
1424 		record_root_in_trans(trans, parent, 1);
1425 	return ret;
1426 }
1427 
1428 /*
1429  * new snapshots need to be created at a very specific time in the
1430  * transaction commit.  This does the actual creation.
1431  *
1432  * Note:
1433  * If the error which may affect the commitment of the current transaction
1434  * happens, we should return the error number. If the error which just affect
1435  * the creation of the pending snapshots, just return 0.
1436  */
1437 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
1438 				   struct btrfs_fs_info *fs_info,
1439 				   struct btrfs_pending_snapshot *pending)
1440 {
1441 	struct btrfs_key key;
1442 	struct btrfs_root_item *new_root_item;
1443 	struct btrfs_root *tree_root = fs_info->tree_root;
1444 	struct btrfs_root *root = pending->root;
1445 	struct btrfs_root *parent_root;
1446 	struct btrfs_block_rsv *rsv;
1447 	struct inode *parent_inode;
1448 	struct btrfs_path *path;
1449 	struct btrfs_dir_item *dir_item;
1450 	struct dentry *dentry;
1451 	struct extent_buffer *tmp;
1452 	struct extent_buffer *old;
1453 	struct timespec cur_time;
1454 	int ret = 0;
1455 	u64 to_reserve = 0;
1456 	u64 index = 0;
1457 	u64 objectid;
1458 	u64 root_flags;
1459 	uuid_le new_uuid;
1460 
1461 	ASSERT(pending->path);
1462 	path = pending->path;
1463 
1464 	ASSERT(pending->root_item);
1465 	new_root_item = pending->root_item;
1466 
1467 	pending->error = btrfs_find_free_objectid(tree_root, &objectid);
1468 	if (pending->error)
1469 		goto no_free_objectid;
1470 
1471 	/*
1472 	 * Make qgroup to skip current new snapshot's qgroupid, as it is
1473 	 * accounted by later btrfs_qgroup_inherit().
1474 	 */
1475 	btrfs_set_skip_qgroup(trans, objectid);
1476 
1477 	btrfs_reloc_pre_snapshot(pending, &to_reserve);
1478 
1479 	if (to_reserve > 0) {
1480 		pending->error = btrfs_block_rsv_add(root,
1481 						     &pending->block_rsv,
1482 						     to_reserve,
1483 						     BTRFS_RESERVE_NO_FLUSH);
1484 		if (pending->error)
1485 			goto clear_skip_qgroup;
1486 	}
1487 
1488 	key.objectid = objectid;
1489 	key.offset = (u64)-1;
1490 	key.type = BTRFS_ROOT_ITEM_KEY;
1491 
1492 	rsv = trans->block_rsv;
1493 	trans->block_rsv = &pending->block_rsv;
1494 	trans->bytes_reserved = trans->block_rsv->reserved;
1495 	trace_btrfs_space_reservation(fs_info, "transaction",
1496 				      trans->transid,
1497 				      trans->bytes_reserved, 1);
1498 	dentry = pending->dentry;
1499 	parent_inode = pending->dir;
1500 	parent_root = BTRFS_I(parent_inode)->root;
1501 	record_root_in_trans(trans, parent_root, 0);
1502 
1503 	cur_time = current_time(parent_inode);
1504 
1505 	/*
1506 	 * insert the directory item
1507 	 */
1508 	ret = btrfs_set_inode_index(BTRFS_I(parent_inode), &index);
1509 	BUG_ON(ret); /* -ENOMEM */
1510 
1511 	/* check if there is a file/dir which has the same name. */
1512 	dir_item = btrfs_lookup_dir_item(NULL, parent_root, path,
1513 					 btrfs_ino(BTRFS_I(parent_inode)),
1514 					 dentry->d_name.name,
1515 					 dentry->d_name.len, 0);
1516 	if (dir_item != NULL && !IS_ERR(dir_item)) {
1517 		pending->error = -EEXIST;
1518 		goto dir_item_existed;
1519 	} else if (IS_ERR(dir_item)) {
1520 		ret = PTR_ERR(dir_item);
1521 		btrfs_abort_transaction(trans, ret);
1522 		goto fail;
1523 	}
1524 	btrfs_release_path(path);
1525 
1526 	/*
1527 	 * pull in the delayed directory update
1528 	 * and the delayed inode item
1529 	 * otherwise we corrupt the FS during
1530 	 * snapshot
1531 	 */
1532 	ret = btrfs_run_delayed_items(trans, fs_info);
1533 	if (ret) {	/* Transaction aborted */
1534 		btrfs_abort_transaction(trans, ret);
1535 		goto fail;
1536 	}
1537 
1538 	record_root_in_trans(trans, root, 0);
1539 	btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
1540 	memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
1541 	btrfs_check_and_init_root_item(new_root_item);
1542 
1543 	root_flags = btrfs_root_flags(new_root_item);
1544 	if (pending->readonly)
1545 		root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
1546 	else
1547 		root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
1548 	btrfs_set_root_flags(new_root_item, root_flags);
1549 
1550 	btrfs_set_root_generation_v2(new_root_item,
1551 			trans->transid);
1552 	uuid_le_gen(&new_uuid);
1553 	memcpy(new_root_item->uuid, new_uuid.b, BTRFS_UUID_SIZE);
1554 	memcpy(new_root_item->parent_uuid, root->root_item.uuid,
1555 			BTRFS_UUID_SIZE);
1556 	if (!(root_flags & BTRFS_ROOT_SUBVOL_RDONLY)) {
1557 		memset(new_root_item->received_uuid, 0,
1558 		       sizeof(new_root_item->received_uuid));
1559 		memset(&new_root_item->stime, 0, sizeof(new_root_item->stime));
1560 		memset(&new_root_item->rtime, 0, sizeof(new_root_item->rtime));
1561 		btrfs_set_root_stransid(new_root_item, 0);
1562 		btrfs_set_root_rtransid(new_root_item, 0);
1563 	}
1564 	btrfs_set_stack_timespec_sec(&new_root_item->otime, cur_time.tv_sec);
1565 	btrfs_set_stack_timespec_nsec(&new_root_item->otime, cur_time.tv_nsec);
1566 	btrfs_set_root_otransid(new_root_item, trans->transid);
1567 
1568 	old = btrfs_lock_root_node(root);
1569 	ret = btrfs_cow_block(trans, root, old, NULL, 0, &old);
1570 	if (ret) {
1571 		btrfs_tree_unlock(old);
1572 		free_extent_buffer(old);
1573 		btrfs_abort_transaction(trans, ret);
1574 		goto fail;
1575 	}
1576 
1577 	btrfs_set_lock_blocking(old);
1578 
1579 	ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1580 	/* clean up in any case */
1581 	btrfs_tree_unlock(old);
1582 	free_extent_buffer(old);
1583 	if (ret) {
1584 		btrfs_abort_transaction(trans, ret);
1585 		goto fail;
1586 	}
1587 	/* see comments in should_cow_block() */
1588 	set_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1589 	smp_wmb();
1590 
1591 	btrfs_set_root_node(new_root_item, tmp);
1592 	/* record when the snapshot was created in key.offset */
1593 	key.offset = trans->transid;
1594 	ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1595 	btrfs_tree_unlock(tmp);
1596 	free_extent_buffer(tmp);
1597 	if (ret) {
1598 		btrfs_abort_transaction(trans, ret);
1599 		goto fail;
1600 	}
1601 
1602 	/*
1603 	 * insert root back/forward references
1604 	 */
1605 	ret = btrfs_add_root_ref(trans, fs_info, objectid,
1606 				 parent_root->root_key.objectid,
1607 				 btrfs_ino(BTRFS_I(parent_inode)), index,
1608 				 dentry->d_name.name, dentry->d_name.len);
1609 	if (ret) {
1610 		btrfs_abort_transaction(trans, ret);
1611 		goto fail;
1612 	}
1613 
1614 	key.offset = (u64)-1;
1615 	pending->snap = btrfs_read_fs_root_no_name(fs_info, &key);
1616 	if (IS_ERR(pending->snap)) {
1617 		ret = PTR_ERR(pending->snap);
1618 		btrfs_abort_transaction(trans, ret);
1619 		goto fail;
1620 	}
1621 
1622 	ret = btrfs_reloc_post_snapshot(trans, pending);
1623 	if (ret) {
1624 		btrfs_abort_transaction(trans, ret);
1625 		goto fail;
1626 	}
1627 
1628 	ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1629 	if (ret) {
1630 		btrfs_abort_transaction(trans, ret);
1631 		goto fail;
1632 	}
1633 
1634 	/*
1635 	 * Do special qgroup accounting for snapshot, as we do some qgroup
1636 	 * snapshot hack to do fast snapshot.
1637 	 * To co-operate with that hack, we do hack again.
1638 	 * Or snapshot will be greatly slowed down by a subtree qgroup rescan
1639 	 */
1640 	ret = qgroup_account_snapshot(trans, root, parent_root,
1641 				      pending->inherit, objectid);
1642 	if (ret < 0)
1643 		goto fail;
1644 
1645 	ret = btrfs_insert_dir_item(trans, parent_root,
1646 				    dentry->d_name.name, dentry->d_name.len,
1647 				    BTRFS_I(parent_inode), &key,
1648 				    BTRFS_FT_DIR, index);
1649 	/* We have check then name at the beginning, so it is impossible. */
1650 	BUG_ON(ret == -EEXIST || ret == -EOVERFLOW);
1651 	if (ret) {
1652 		btrfs_abort_transaction(trans, ret);
1653 		goto fail;
1654 	}
1655 
1656 	btrfs_i_size_write(BTRFS_I(parent_inode), parent_inode->i_size +
1657 					 dentry->d_name.len * 2);
1658 	parent_inode->i_mtime = parent_inode->i_ctime =
1659 		current_time(parent_inode);
1660 	ret = btrfs_update_inode_fallback(trans, parent_root, parent_inode);
1661 	if (ret) {
1662 		btrfs_abort_transaction(trans, ret);
1663 		goto fail;
1664 	}
1665 	ret = btrfs_uuid_tree_add(trans, fs_info, new_uuid.b,
1666 				  BTRFS_UUID_KEY_SUBVOL, objectid);
1667 	if (ret) {
1668 		btrfs_abort_transaction(trans, ret);
1669 		goto fail;
1670 	}
1671 	if (!btrfs_is_empty_uuid(new_root_item->received_uuid)) {
1672 		ret = btrfs_uuid_tree_add(trans, fs_info,
1673 					  new_root_item->received_uuid,
1674 					  BTRFS_UUID_KEY_RECEIVED_SUBVOL,
1675 					  objectid);
1676 		if (ret && ret != -EEXIST) {
1677 			btrfs_abort_transaction(trans, ret);
1678 			goto fail;
1679 		}
1680 	}
1681 
1682 	ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1683 	if (ret) {
1684 		btrfs_abort_transaction(trans, ret);
1685 		goto fail;
1686 	}
1687 
1688 fail:
1689 	pending->error = ret;
1690 dir_item_existed:
1691 	trans->block_rsv = rsv;
1692 	trans->bytes_reserved = 0;
1693 clear_skip_qgroup:
1694 	btrfs_clear_skip_qgroup(trans);
1695 no_free_objectid:
1696 	kfree(new_root_item);
1697 	pending->root_item = NULL;
1698 	btrfs_free_path(path);
1699 	pending->path = NULL;
1700 
1701 	return ret;
1702 }
1703 
1704 /*
1705  * create all the snapshots we've scheduled for creation
1706  */
1707 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
1708 					     struct btrfs_fs_info *fs_info)
1709 {
1710 	struct btrfs_pending_snapshot *pending, *next;
1711 	struct list_head *head = &trans->transaction->pending_snapshots;
1712 	int ret = 0;
1713 
1714 	list_for_each_entry_safe(pending, next, head, list) {
1715 		list_del(&pending->list);
1716 		ret = create_pending_snapshot(trans, fs_info, pending);
1717 		if (ret)
1718 			break;
1719 	}
1720 	return ret;
1721 }
1722 
1723 static void update_super_roots(struct btrfs_fs_info *fs_info)
1724 {
1725 	struct btrfs_root_item *root_item;
1726 	struct btrfs_super_block *super;
1727 
1728 	super = fs_info->super_copy;
1729 
1730 	root_item = &fs_info->chunk_root->root_item;
1731 	super->chunk_root = root_item->bytenr;
1732 	super->chunk_root_generation = root_item->generation;
1733 	super->chunk_root_level = root_item->level;
1734 
1735 	root_item = &fs_info->tree_root->root_item;
1736 	super->root = root_item->bytenr;
1737 	super->generation = root_item->generation;
1738 	super->root_level = root_item->level;
1739 	if (btrfs_test_opt(fs_info, SPACE_CACHE))
1740 		super->cache_generation = root_item->generation;
1741 	if (test_bit(BTRFS_FS_UPDATE_UUID_TREE_GEN, &fs_info->flags))
1742 		super->uuid_tree_generation = root_item->generation;
1743 }
1744 
1745 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1746 {
1747 	struct btrfs_transaction *trans;
1748 	int ret = 0;
1749 
1750 	spin_lock(&info->trans_lock);
1751 	trans = info->running_transaction;
1752 	if (trans)
1753 		ret = (trans->state >= TRANS_STATE_COMMIT_START);
1754 	spin_unlock(&info->trans_lock);
1755 	return ret;
1756 }
1757 
1758 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1759 {
1760 	struct btrfs_transaction *trans;
1761 	int ret = 0;
1762 
1763 	spin_lock(&info->trans_lock);
1764 	trans = info->running_transaction;
1765 	if (trans)
1766 		ret = is_transaction_blocked(trans);
1767 	spin_unlock(&info->trans_lock);
1768 	return ret;
1769 }
1770 
1771 /*
1772  * wait for the current transaction commit to start and block subsequent
1773  * transaction joins
1774  */
1775 static void wait_current_trans_commit_start(struct btrfs_fs_info *fs_info,
1776 					    struct btrfs_transaction *trans)
1777 {
1778 	wait_event(fs_info->transaction_blocked_wait,
1779 		   trans->state >= TRANS_STATE_COMMIT_START || trans->aborted);
1780 }
1781 
1782 /*
1783  * wait for the current transaction to start and then become unblocked.
1784  * caller holds ref.
1785  */
1786 static void wait_current_trans_commit_start_and_unblock(
1787 					struct btrfs_fs_info *fs_info,
1788 					struct btrfs_transaction *trans)
1789 {
1790 	wait_event(fs_info->transaction_wait,
1791 		   trans->state >= TRANS_STATE_UNBLOCKED || trans->aborted);
1792 }
1793 
1794 /*
1795  * commit transactions asynchronously. once btrfs_commit_transaction_async
1796  * returns, any subsequent transaction will not be allowed to join.
1797  */
1798 struct btrfs_async_commit {
1799 	struct btrfs_trans_handle *newtrans;
1800 	struct work_struct work;
1801 };
1802 
1803 static void do_async_commit(struct work_struct *work)
1804 {
1805 	struct btrfs_async_commit *ac =
1806 		container_of(work, struct btrfs_async_commit, work);
1807 
1808 	/*
1809 	 * We've got freeze protection passed with the transaction.
1810 	 * Tell lockdep about it.
1811 	 */
1812 	if (ac->newtrans->type & __TRANS_FREEZABLE)
1813 		__sb_writers_acquired(ac->newtrans->fs_info->sb, SB_FREEZE_FS);
1814 
1815 	current->journal_info = ac->newtrans;
1816 
1817 	btrfs_commit_transaction(ac->newtrans);
1818 	kfree(ac);
1819 }
1820 
1821 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1822 				   int wait_for_unblock)
1823 {
1824 	struct btrfs_fs_info *fs_info = trans->fs_info;
1825 	struct btrfs_async_commit *ac;
1826 	struct btrfs_transaction *cur_trans;
1827 
1828 	ac = kmalloc(sizeof(*ac), GFP_NOFS);
1829 	if (!ac)
1830 		return -ENOMEM;
1831 
1832 	INIT_WORK(&ac->work, do_async_commit);
1833 	ac->newtrans = btrfs_join_transaction(trans->root);
1834 	if (IS_ERR(ac->newtrans)) {
1835 		int err = PTR_ERR(ac->newtrans);
1836 		kfree(ac);
1837 		return err;
1838 	}
1839 
1840 	/* take transaction reference */
1841 	cur_trans = trans->transaction;
1842 	atomic_inc(&cur_trans->use_count);
1843 
1844 	btrfs_end_transaction(trans);
1845 
1846 	/*
1847 	 * Tell lockdep we've released the freeze rwsem, since the
1848 	 * async commit thread will be the one to unlock it.
1849 	 */
1850 	if (ac->newtrans->type & __TRANS_FREEZABLE)
1851 		__sb_writers_release(fs_info->sb, SB_FREEZE_FS);
1852 
1853 	schedule_work(&ac->work);
1854 
1855 	/* wait for transaction to start and unblock */
1856 	if (wait_for_unblock)
1857 		wait_current_trans_commit_start_and_unblock(fs_info, cur_trans);
1858 	else
1859 		wait_current_trans_commit_start(fs_info, cur_trans);
1860 
1861 	if (current->journal_info == trans)
1862 		current->journal_info = NULL;
1863 
1864 	btrfs_put_transaction(cur_trans);
1865 	return 0;
1866 }
1867 
1868 
1869 static void cleanup_transaction(struct btrfs_trans_handle *trans,
1870 				struct btrfs_root *root, int err)
1871 {
1872 	struct btrfs_fs_info *fs_info = root->fs_info;
1873 	struct btrfs_transaction *cur_trans = trans->transaction;
1874 	DEFINE_WAIT(wait);
1875 
1876 	WARN_ON(trans->use_count > 1);
1877 
1878 	btrfs_abort_transaction(trans, err);
1879 
1880 	spin_lock(&fs_info->trans_lock);
1881 
1882 	/*
1883 	 * If the transaction is removed from the list, it means this
1884 	 * transaction has been committed successfully, so it is impossible
1885 	 * to call the cleanup function.
1886 	 */
1887 	BUG_ON(list_empty(&cur_trans->list));
1888 
1889 	list_del_init(&cur_trans->list);
1890 	if (cur_trans == fs_info->running_transaction) {
1891 		cur_trans->state = TRANS_STATE_COMMIT_DOING;
1892 		spin_unlock(&fs_info->trans_lock);
1893 		wait_event(cur_trans->writer_wait,
1894 			   atomic_read(&cur_trans->num_writers) == 1);
1895 
1896 		spin_lock(&fs_info->trans_lock);
1897 	}
1898 	spin_unlock(&fs_info->trans_lock);
1899 
1900 	btrfs_cleanup_one_transaction(trans->transaction, fs_info);
1901 
1902 	spin_lock(&fs_info->trans_lock);
1903 	if (cur_trans == fs_info->running_transaction)
1904 		fs_info->running_transaction = NULL;
1905 	spin_unlock(&fs_info->trans_lock);
1906 
1907 	if (trans->type & __TRANS_FREEZABLE)
1908 		sb_end_intwrite(fs_info->sb);
1909 	btrfs_put_transaction(cur_trans);
1910 	btrfs_put_transaction(cur_trans);
1911 
1912 	trace_btrfs_transaction_commit(root);
1913 
1914 	if (current->journal_info == trans)
1915 		current->journal_info = NULL;
1916 	btrfs_scrub_cancel(fs_info);
1917 
1918 	kmem_cache_free(btrfs_trans_handle_cachep, trans);
1919 }
1920 
1921 static inline int btrfs_start_delalloc_flush(struct btrfs_fs_info *fs_info)
1922 {
1923 	if (btrfs_test_opt(fs_info, FLUSHONCOMMIT))
1924 		return btrfs_start_delalloc_roots(fs_info, 1, -1);
1925 	return 0;
1926 }
1927 
1928 static inline void btrfs_wait_delalloc_flush(struct btrfs_fs_info *fs_info)
1929 {
1930 	if (btrfs_test_opt(fs_info, FLUSHONCOMMIT))
1931 		btrfs_wait_ordered_roots(fs_info, -1, 0, (u64)-1);
1932 }
1933 
1934 static inline void
1935 btrfs_wait_pending_ordered(struct btrfs_transaction *cur_trans)
1936 {
1937 	wait_event(cur_trans->pending_wait,
1938 		   atomic_read(&cur_trans->pending_ordered) == 0);
1939 }
1940 
1941 int btrfs_commit_transaction(struct btrfs_trans_handle *trans)
1942 {
1943 	struct btrfs_fs_info *fs_info = trans->fs_info;
1944 	struct btrfs_transaction *cur_trans = trans->transaction;
1945 	struct btrfs_transaction *prev_trans = NULL;
1946 	int ret;
1947 
1948 	/* Stop the commit early if ->aborted is set */
1949 	if (unlikely(READ_ONCE(cur_trans->aborted))) {
1950 		ret = cur_trans->aborted;
1951 		btrfs_end_transaction(trans);
1952 		return ret;
1953 	}
1954 
1955 	/* make a pass through all the delayed refs we have so far
1956 	 * any runnings procs may add more while we are here
1957 	 */
1958 	ret = btrfs_run_delayed_refs(trans, fs_info, 0);
1959 	if (ret) {
1960 		btrfs_end_transaction(trans);
1961 		return ret;
1962 	}
1963 
1964 	btrfs_trans_release_metadata(trans, fs_info);
1965 	trans->block_rsv = NULL;
1966 
1967 	cur_trans = trans->transaction;
1968 
1969 	/*
1970 	 * set the flushing flag so procs in this transaction have to
1971 	 * start sending their work down.
1972 	 */
1973 	cur_trans->delayed_refs.flushing = 1;
1974 	smp_wmb();
1975 
1976 	if (!list_empty(&trans->new_bgs))
1977 		btrfs_create_pending_block_groups(trans, fs_info);
1978 
1979 	ret = btrfs_run_delayed_refs(trans, fs_info, 0);
1980 	if (ret) {
1981 		btrfs_end_transaction(trans);
1982 		return ret;
1983 	}
1984 
1985 	if (!test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &cur_trans->flags)) {
1986 		int run_it = 0;
1987 
1988 		/* this mutex is also taken before trying to set
1989 		 * block groups readonly.  We need to make sure
1990 		 * that nobody has set a block group readonly
1991 		 * after a extents from that block group have been
1992 		 * allocated for cache files.  btrfs_set_block_group_ro
1993 		 * will wait for the transaction to commit if it
1994 		 * finds BTRFS_TRANS_DIRTY_BG_RUN set.
1995 		 *
1996 		 * The BTRFS_TRANS_DIRTY_BG_RUN flag is also used to make sure
1997 		 * only one process starts all the block group IO.  It wouldn't
1998 		 * hurt to have more than one go through, but there's no
1999 		 * real advantage to it either.
2000 		 */
2001 		mutex_lock(&fs_info->ro_block_group_mutex);
2002 		if (!test_and_set_bit(BTRFS_TRANS_DIRTY_BG_RUN,
2003 				      &cur_trans->flags))
2004 			run_it = 1;
2005 		mutex_unlock(&fs_info->ro_block_group_mutex);
2006 
2007 		if (run_it)
2008 			ret = btrfs_start_dirty_block_groups(trans, fs_info);
2009 	}
2010 	if (ret) {
2011 		btrfs_end_transaction(trans);
2012 		return ret;
2013 	}
2014 
2015 	spin_lock(&fs_info->trans_lock);
2016 	if (cur_trans->state >= TRANS_STATE_COMMIT_START) {
2017 		spin_unlock(&fs_info->trans_lock);
2018 		atomic_inc(&cur_trans->use_count);
2019 		ret = btrfs_end_transaction(trans);
2020 
2021 		wait_for_commit(cur_trans);
2022 
2023 		if (unlikely(cur_trans->aborted))
2024 			ret = cur_trans->aborted;
2025 
2026 		btrfs_put_transaction(cur_trans);
2027 
2028 		return ret;
2029 	}
2030 
2031 	cur_trans->state = TRANS_STATE_COMMIT_START;
2032 	wake_up(&fs_info->transaction_blocked_wait);
2033 
2034 	if (cur_trans->list.prev != &fs_info->trans_list) {
2035 		prev_trans = list_entry(cur_trans->list.prev,
2036 					struct btrfs_transaction, list);
2037 		if (prev_trans->state != TRANS_STATE_COMPLETED) {
2038 			atomic_inc(&prev_trans->use_count);
2039 			spin_unlock(&fs_info->trans_lock);
2040 
2041 			wait_for_commit(prev_trans);
2042 			ret = prev_trans->aborted;
2043 
2044 			btrfs_put_transaction(prev_trans);
2045 			if (ret)
2046 				goto cleanup_transaction;
2047 		} else {
2048 			spin_unlock(&fs_info->trans_lock);
2049 		}
2050 	} else {
2051 		spin_unlock(&fs_info->trans_lock);
2052 	}
2053 
2054 	extwriter_counter_dec(cur_trans, trans->type);
2055 
2056 	ret = btrfs_start_delalloc_flush(fs_info);
2057 	if (ret)
2058 		goto cleanup_transaction;
2059 
2060 	ret = btrfs_run_delayed_items(trans, fs_info);
2061 	if (ret)
2062 		goto cleanup_transaction;
2063 
2064 	wait_event(cur_trans->writer_wait,
2065 		   extwriter_counter_read(cur_trans) == 0);
2066 
2067 	/* some pending stuffs might be added after the previous flush. */
2068 	ret = btrfs_run_delayed_items(trans, fs_info);
2069 	if (ret)
2070 		goto cleanup_transaction;
2071 
2072 	btrfs_wait_delalloc_flush(fs_info);
2073 
2074 	btrfs_wait_pending_ordered(cur_trans);
2075 
2076 	btrfs_scrub_pause(fs_info);
2077 	/*
2078 	 * Ok now we need to make sure to block out any other joins while we
2079 	 * commit the transaction.  We could have started a join before setting
2080 	 * COMMIT_DOING so make sure to wait for num_writers to == 1 again.
2081 	 */
2082 	spin_lock(&fs_info->trans_lock);
2083 	cur_trans->state = TRANS_STATE_COMMIT_DOING;
2084 	spin_unlock(&fs_info->trans_lock);
2085 	wait_event(cur_trans->writer_wait,
2086 		   atomic_read(&cur_trans->num_writers) == 1);
2087 
2088 	/* ->aborted might be set after the previous check, so check it */
2089 	if (unlikely(READ_ONCE(cur_trans->aborted))) {
2090 		ret = cur_trans->aborted;
2091 		goto scrub_continue;
2092 	}
2093 	/*
2094 	 * the reloc mutex makes sure that we stop
2095 	 * the balancing code from coming in and moving
2096 	 * extents around in the middle of the commit
2097 	 */
2098 	mutex_lock(&fs_info->reloc_mutex);
2099 
2100 	/*
2101 	 * We needn't worry about the delayed items because we will
2102 	 * deal with them in create_pending_snapshot(), which is the
2103 	 * core function of the snapshot creation.
2104 	 */
2105 	ret = create_pending_snapshots(trans, fs_info);
2106 	if (ret) {
2107 		mutex_unlock(&fs_info->reloc_mutex);
2108 		goto scrub_continue;
2109 	}
2110 
2111 	/*
2112 	 * We insert the dir indexes of the snapshots and update the inode
2113 	 * of the snapshots' parents after the snapshot creation, so there
2114 	 * are some delayed items which are not dealt with. Now deal with
2115 	 * them.
2116 	 *
2117 	 * We needn't worry that this operation will corrupt the snapshots,
2118 	 * because all the tree which are snapshoted will be forced to COW
2119 	 * the nodes and leaves.
2120 	 */
2121 	ret = btrfs_run_delayed_items(trans, fs_info);
2122 	if (ret) {
2123 		mutex_unlock(&fs_info->reloc_mutex);
2124 		goto scrub_continue;
2125 	}
2126 
2127 	ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
2128 	if (ret) {
2129 		mutex_unlock(&fs_info->reloc_mutex);
2130 		goto scrub_continue;
2131 	}
2132 
2133 	/* Reocrd old roots for later qgroup accounting */
2134 	ret = btrfs_qgroup_prepare_account_extents(trans, fs_info);
2135 	if (ret) {
2136 		mutex_unlock(&fs_info->reloc_mutex);
2137 		goto scrub_continue;
2138 	}
2139 
2140 	/*
2141 	 * make sure none of the code above managed to slip in a
2142 	 * delayed item
2143 	 */
2144 	btrfs_assert_delayed_root_empty(fs_info);
2145 
2146 	WARN_ON(cur_trans != trans->transaction);
2147 
2148 	/* btrfs_commit_tree_roots is responsible for getting the
2149 	 * various roots consistent with each other.  Every pointer
2150 	 * in the tree of tree roots has to point to the most up to date
2151 	 * root for every subvolume and other tree.  So, we have to keep
2152 	 * the tree logging code from jumping in and changing any
2153 	 * of the trees.
2154 	 *
2155 	 * At this point in the commit, there can't be any tree-log
2156 	 * writers, but a little lower down we drop the trans mutex
2157 	 * and let new people in.  By holding the tree_log_mutex
2158 	 * from now until after the super is written, we avoid races
2159 	 * with the tree-log code.
2160 	 */
2161 	mutex_lock(&fs_info->tree_log_mutex);
2162 
2163 	ret = commit_fs_roots(trans, fs_info);
2164 	if (ret) {
2165 		mutex_unlock(&fs_info->tree_log_mutex);
2166 		mutex_unlock(&fs_info->reloc_mutex);
2167 		goto scrub_continue;
2168 	}
2169 
2170 	/*
2171 	 * Since the transaction is done, we can apply the pending changes
2172 	 * before the next transaction.
2173 	 */
2174 	btrfs_apply_pending_changes(fs_info);
2175 
2176 	/* commit_fs_roots gets rid of all the tree log roots, it is now
2177 	 * safe to free the root of tree log roots
2178 	 */
2179 	btrfs_free_log_root_tree(trans, fs_info);
2180 
2181 	/*
2182 	 * Since fs roots are all committed, we can get a quite accurate
2183 	 * new_roots. So let's do quota accounting.
2184 	 */
2185 	ret = btrfs_qgroup_account_extents(trans, fs_info);
2186 	if (ret < 0) {
2187 		mutex_unlock(&fs_info->tree_log_mutex);
2188 		mutex_unlock(&fs_info->reloc_mutex);
2189 		goto scrub_continue;
2190 	}
2191 
2192 	ret = commit_cowonly_roots(trans, fs_info);
2193 	if (ret) {
2194 		mutex_unlock(&fs_info->tree_log_mutex);
2195 		mutex_unlock(&fs_info->reloc_mutex);
2196 		goto scrub_continue;
2197 	}
2198 
2199 	/*
2200 	 * The tasks which save the space cache and inode cache may also
2201 	 * update ->aborted, check it.
2202 	 */
2203 	if (unlikely(READ_ONCE(cur_trans->aborted))) {
2204 		ret = cur_trans->aborted;
2205 		mutex_unlock(&fs_info->tree_log_mutex);
2206 		mutex_unlock(&fs_info->reloc_mutex);
2207 		goto scrub_continue;
2208 	}
2209 
2210 	btrfs_prepare_extent_commit(fs_info);
2211 
2212 	cur_trans = fs_info->running_transaction;
2213 
2214 	btrfs_set_root_node(&fs_info->tree_root->root_item,
2215 			    fs_info->tree_root->node);
2216 	list_add_tail(&fs_info->tree_root->dirty_list,
2217 		      &cur_trans->switch_commits);
2218 
2219 	btrfs_set_root_node(&fs_info->chunk_root->root_item,
2220 			    fs_info->chunk_root->node);
2221 	list_add_tail(&fs_info->chunk_root->dirty_list,
2222 		      &cur_trans->switch_commits);
2223 
2224 	switch_commit_roots(cur_trans, fs_info);
2225 
2226 	assert_qgroups_uptodate(trans);
2227 	ASSERT(list_empty(&cur_trans->dirty_bgs));
2228 	ASSERT(list_empty(&cur_trans->io_bgs));
2229 	update_super_roots(fs_info);
2230 
2231 	btrfs_set_super_log_root(fs_info->super_copy, 0);
2232 	btrfs_set_super_log_root_level(fs_info->super_copy, 0);
2233 	memcpy(fs_info->super_for_commit, fs_info->super_copy,
2234 	       sizeof(*fs_info->super_copy));
2235 
2236 	btrfs_update_commit_device_size(fs_info);
2237 	btrfs_update_commit_device_bytes_used(fs_info, cur_trans);
2238 
2239 	clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags);
2240 	clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags);
2241 
2242 	btrfs_trans_release_chunk_metadata(trans);
2243 
2244 	spin_lock(&fs_info->trans_lock);
2245 	cur_trans->state = TRANS_STATE_UNBLOCKED;
2246 	fs_info->running_transaction = NULL;
2247 	spin_unlock(&fs_info->trans_lock);
2248 	mutex_unlock(&fs_info->reloc_mutex);
2249 
2250 	wake_up(&fs_info->transaction_wait);
2251 
2252 	ret = btrfs_write_and_wait_transaction(trans, fs_info);
2253 	if (ret) {
2254 		btrfs_handle_fs_error(fs_info, ret,
2255 				      "Error while writing out transaction");
2256 		mutex_unlock(&fs_info->tree_log_mutex);
2257 		goto scrub_continue;
2258 	}
2259 
2260 	ret = write_all_supers(fs_info, 0);
2261 	if (ret) {
2262 		mutex_unlock(&fs_info->tree_log_mutex);
2263 		goto scrub_continue;
2264 	}
2265 
2266 	/*
2267 	 * the super is written, we can safely allow the tree-loggers
2268 	 * to go about their business
2269 	 */
2270 	mutex_unlock(&fs_info->tree_log_mutex);
2271 
2272 	btrfs_finish_extent_commit(trans, fs_info);
2273 
2274 	if (test_bit(BTRFS_TRANS_HAVE_FREE_BGS, &cur_trans->flags))
2275 		btrfs_clear_space_info_full(fs_info);
2276 
2277 	fs_info->last_trans_committed = cur_trans->transid;
2278 	/*
2279 	 * We needn't acquire the lock here because there is no other task
2280 	 * which can change it.
2281 	 */
2282 	cur_trans->state = TRANS_STATE_COMPLETED;
2283 	wake_up(&cur_trans->commit_wait);
2284 
2285 	spin_lock(&fs_info->trans_lock);
2286 	list_del_init(&cur_trans->list);
2287 	spin_unlock(&fs_info->trans_lock);
2288 
2289 	btrfs_put_transaction(cur_trans);
2290 	btrfs_put_transaction(cur_trans);
2291 
2292 	if (trans->type & __TRANS_FREEZABLE)
2293 		sb_end_intwrite(fs_info->sb);
2294 
2295 	trace_btrfs_transaction_commit(trans->root);
2296 
2297 	btrfs_scrub_continue(fs_info);
2298 
2299 	if (current->journal_info == trans)
2300 		current->journal_info = NULL;
2301 
2302 	kmem_cache_free(btrfs_trans_handle_cachep, trans);
2303 
2304 	/*
2305 	 * If fs has been frozen, we can not handle delayed iputs, otherwise
2306 	 * it'll result in deadlock about SB_FREEZE_FS.
2307 	 */
2308 	if (current != fs_info->transaction_kthread &&
2309 	    current != fs_info->cleaner_kthread && !fs_info->fs_frozen)
2310 		btrfs_run_delayed_iputs(fs_info);
2311 
2312 	return ret;
2313 
2314 scrub_continue:
2315 	btrfs_scrub_continue(fs_info);
2316 cleanup_transaction:
2317 	btrfs_trans_release_metadata(trans, fs_info);
2318 	btrfs_trans_release_chunk_metadata(trans);
2319 	trans->block_rsv = NULL;
2320 	btrfs_warn(fs_info, "Skipping commit of aborted transaction.");
2321 	if (current->journal_info == trans)
2322 		current->journal_info = NULL;
2323 	cleanup_transaction(trans, trans->root, ret);
2324 
2325 	return ret;
2326 }
2327 
2328 /*
2329  * return < 0 if error
2330  * 0 if there are no more dead_roots at the time of call
2331  * 1 there are more to be processed, call me again
2332  *
2333  * The return value indicates there are certainly more snapshots to delete, but
2334  * if there comes a new one during processing, it may return 0. We don't mind,
2335  * because btrfs_commit_super will poke cleaner thread and it will process it a
2336  * few seconds later.
2337  */
2338 int btrfs_clean_one_deleted_snapshot(struct btrfs_root *root)
2339 {
2340 	int ret;
2341 	struct btrfs_fs_info *fs_info = root->fs_info;
2342 
2343 	spin_lock(&fs_info->trans_lock);
2344 	if (list_empty(&fs_info->dead_roots)) {
2345 		spin_unlock(&fs_info->trans_lock);
2346 		return 0;
2347 	}
2348 	root = list_first_entry(&fs_info->dead_roots,
2349 			struct btrfs_root, root_list);
2350 	list_del_init(&root->root_list);
2351 	spin_unlock(&fs_info->trans_lock);
2352 
2353 	btrfs_debug(fs_info, "cleaner removing %llu", root->objectid);
2354 
2355 	btrfs_kill_all_delayed_nodes(root);
2356 
2357 	if (btrfs_header_backref_rev(root->node) <
2358 			BTRFS_MIXED_BACKREF_REV)
2359 		ret = btrfs_drop_snapshot(root, NULL, 0, 0);
2360 	else
2361 		ret = btrfs_drop_snapshot(root, NULL, 1, 0);
2362 
2363 	return (ret < 0) ? 0 : 1;
2364 }
2365 
2366 void btrfs_apply_pending_changes(struct btrfs_fs_info *fs_info)
2367 {
2368 	unsigned long prev;
2369 	unsigned long bit;
2370 
2371 	prev = xchg(&fs_info->pending_changes, 0);
2372 	if (!prev)
2373 		return;
2374 
2375 	bit = 1 << BTRFS_PENDING_SET_INODE_MAP_CACHE;
2376 	if (prev & bit)
2377 		btrfs_set_opt(fs_info->mount_opt, INODE_MAP_CACHE);
2378 	prev &= ~bit;
2379 
2380 	bit = 1 << BTRFS_PENDING_CLEAR_INODE_MAP_CACHE;
2381 	if (prev & bit)
2382 		btrfs_clear_opt(fs_info->mount_opt, INODE_MAP_CACHE);
2383 	prev &= ~bit;
2384 
2385 	bit = 1 << BTRFS_PENDING_COMMIT;
2386 	if (prev & bit)
2387 		btrfs_debug(fs_info, "pending commit done");
2388 	prev &= ~bit;
2389 
2390 	if (prev)
2391 		btrfs_warn(fs_info,
2392 			"unknown pending changes left 0x%lx, ignoring", prev);
2393 }
2394