34efc9773719a1ca53dfff7d0d2270259702480e
[sfrench/cifs-2.6.git] / fs / btrfs / transaction.c
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_START,
41         [TRANS_STATE_COMMIT_START]      = (__TRANS_START | __TRANS_ATTACH),
42         [TRANS_STATE_COMMIT_DOING]      = (__TRANS_START |
43                                            __TRANS_ATTACH |
44                                            __TRANS_JOIN),
45         [TRANS_STATE_UNBLOCKED]         = (__TRANS_START |
46                                            __TRANS_ATTACH |
47                                            __TRANS_JOIN |
48                                            __TRANS_JOIN_NOLOCK),
49         [TRANS_STATE_COMPLETED]         = (__TRANS_START |
50                                            __TRANS_ATTACH |
51                                            __TRANS_JOIN |
52                                            __TRANS_JOIN_NOLOCK),
53 };
54
55 void btrfs_put_transaction(struct btrfs_transaction *transaction)
56 {
57         WARN_ON(refcount_read(&transaction->use_count) == 0);
58         if (refcount_dec_and_test(&transaction->use_count)) {
59                 BUG_ON(!list_empty(&transaction->list));
60                 WARN_ON(!RB_EMPTY_ROOT(&transaction->delayed_refs.href_root));
61                 if (transaction->delayed_refs.pending_csums)
62                         btrfs_err(transaction->fs_info,
63                                   "pending csums is %llu",
64                                   transaction->delayed_refs.pending_csums);
65                 while (!list_empty(&transaction->pending_chunks)) {
66                         struct extent_map *em;
67
68                         em = list_first_entry(&transaction->pending_chunks,
69                                               struct extent_map, list);
70                         list_del_init(&em->list);
71                         free_extent_map(em);
72                 }
73                 /*
74                  * If any block groups are found in ->deleted_bgs then it's
75                  * because the transaction was aborted and a commit did not
76                  * happen (things failed before writing the new superblock
77                  * and calling btrfs_finish_extent_commit()), so we can not
78                  * discard the physical locations of the block groups.
79                  */
80                 while (!list_empty(&transaction->deleted_bgs)) {
81                         struct btrfs_block_group_cache *cache;
82
83                         cache = list_first_entry(&transaction->deleted_bgs,
84                                                  struct btrfs_block_group_cache,
85                                                  bg_list);
86                         list_del_init(&cache->bg_list);
87                         btrfs_put_block_group_trimming(cache);
88                         btrfs_put_block_group(cache);
89                 }
90                 kfree(transaction);
91         }
92 }
93
94 static void clear_btree_io_tree(struct extent_io_tree *tree)
95 {
96         spin_lock(&tree->lock);
97         /*
98          * Do a single barrier for the waitqueue_active check here, the state
99          * of the waitqueue should not change once clear_btree_io_tree is
100          * called.
101          */
102         smp_mb();
103         while (!RB_EMPTY_ROOT(&tree->state)) {
104                 struct rb_node *node;
105                 struct extent_state *state;
106
107                 node = rb_first(&tree->state);
108                 state = rb_entry(node, struct extent_state, rb_node);
109                 rb_erase(&state->rb_node, &tree->state);
110                 RB_CLEAR_NODE(&state->rb_node);
111                 /*
112                  * btree io trees aren't supposed to have tasks waiting for
113                  * changes in the flags of extent states ever.
114                  */
115                 ASSERT(!waitqueue_active(&state->wq));
116                 free_extent_state(state);
117
118                 cond_resched_lock(&tree->lock);
119         }
120         spin_unlock(&tree->lock);
121 }
122
123 static noinline void switch_commit_roots(struct btrfs_transaction *trans)
124 {
125         struct btrfs_fs_info *fs_info = trans->fs_info;
126         struct btrfs_root *root, *tmp;
127
128         down_write(&fs_info->commit_root_sem);
129         list_for_each_entry_safe(root, tmp, &trans->switch_commits,
130                                  dirty_list) {
131                 list_del_init(&root->dirty_list);
132                 free_extent_buffer(root->commit_root);
133                 root->commit_root = btrfs_root_node(root);
134                 if (is_fstree(root->objectid))
135                         btrfs_unpin_free_ino(root);
136                 clear_btree_io_tree(&root->dirty_log_pages);
137         }
138
139         /* We can free old roots now. */
140         spin_lock(&trans->dropped_roots_lock);
141         while (!list_empty(&trans->dropped_roots)) {
142                 root = list_first_entry(&trans->dropped_roots,
143                                         struct btrfs_root, root_list);
144                 list_del_init(&root->root_list);
145                 spin_unlock(&trans->dropped_roots_lock);
146                 btrfs_drop_and_free_fs_root(fs_info, root);
147                 spin_lock(&trans->dropped_roots_lock);
148         }
149         spin_unlock(&trans->dropped_roots_lock);
150         up_write(&fs_info->commit_root_sem);
151 }
152
153 static inline void extwriter_counter_inc(struct btrfs_transaction *trans,
154                                          unsigned int type)
155 {
156         if (type & TRANS_EXTWRITERS)
157                 atomic_inc(&trans->num_extwriters);
158 }
159
160 static inline void extwriter_counter_dec(struct btrfs_transaction *trans,
161                                          unsigned int type)
162 {
163         if (type & TRANS_EXTWRITERS)
164                 atomic_dec(&trans->num_extwriters);
165 }
166
167 static inline void extwriter_counter_init(struct btrfs_transaction *trans,
168                                           unsigned int type)
169 {
170         atomic_set(&trans->num_extwriters, ((type & TRANS_EXTWRITERS) ? 1 : 0));
171 }
172
173 static inline int extwriter_counter_read(struct btrfs_transaction *trans)
174 {
175         return atomic_read(&trans->num_extwriters);
176 }
177
178 /*
179  * either allocate a new transaction or hop into the existing one
180  */
181 static noinline int join_transaction(struct btrfs_fs_info *fs_info,
182                                      unsigned int type)
183 {
184         struct btrfs_transaction *cur_trans;
185
186         spin_lock(&fs_info->trans_lock);
187 loop:
188         /* The file system has been taken offline. No new transactions. */
189         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
190                 spin_unlock(&fs_info->trans_lock);
191                 return -EROFS;
192         }
193
194         cur_trans = fs_info->running_transaction;
195         if (cur_trans) {
196                 if (cur_trans->aborted) {
197                         spin_unlock(&fs_info->trans_lock);
198                         return cur_trans->aborted;
199                 }
200                 if (btrfs_blocked_trans_types[cur_trans->state] & type) {
201                         spin_unlock(&fs_info->trans_lock);
202                         return -EBUSY;
203                 }
204                 refcount_inc(&cur_trans->use_count);
205                 atomic_inc(&cur_trans->num_writers);
206                 extwriter_counter_inc(cur_trans, type);
207                 spin_unlock(&fs_info->trans_lock);
208                 return 0;
209         }
210         spin_unlock(&fs_info->trans_lock);
211
212         /*
213          * If we are ATTACH, we just want to catch the current transaction,
214          * and commit it. If there is no transaction, just return ENOENT.
215          */
216         if (type == TRANS_ATTACH)
217                 return -ENOENT;
218
219         /*
220          * JOIN_NOLOCK only happens during the transaction commit, so
221          * it is impossible that ->running_transaction is NULL
222          */
223         BUG_ON(type == TRANS_JOIN_NOLOCK);
224
225         cur_trans = kmalloc(sizeof(*cur_trans), GFP_NOFS);
226         if (!cur_trans)
227                 return -ENOMEM;
228
229         spin_lock(&fs_info->trans_lock);
230         if (fs_info->running_transaction) {
231                 /*
232                  * someone started a transaction after we unlocked.  Make sure
233                  * to redo the checks above
234                  */
235                 kfree(cur_trans);
236                 goto loop;
237         } else if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
238                 spin_unlock(&fs_info->trans_lock);
239                 kfree(cur_trans);
240                 return -EROFS;
241         }
242
243         cur_trans->fs_info = fs_info;
244         atomic_set(&cur_trans->num_writers, 1);
245         extwriter_counter_init(cur_trans, type);
246         init_waitqueue_head(&cur_trans->writer_wait);
247         init_waitqueue_head(&cur_trans->commit_wait);
248         init_waitqueue_head(&cur_trans->pending_wait);
249         cur_trans->state = TRANS_STATE_RUNNING;
250         /*
251          * One for this trans handle, one so it will live on until we
252          * commit the transaction.
253          */
254         refcount_set(&cur_trans->use_count, 2);
255         atomic_set(&cur_trans->pending_ordered, 0);
256         cur_trans->flags = 0;
257         cur_trans->start_time = get_seconds();
258
259         memset(&cur_trans->delayed_refs, 0, sizeof(cur_trans->delayed_refs));
260
261         cur_trans->delayed_refs.href_root = RB_ROOT;
262         cur_trans->delayed_refs.dirty_extent_root = RB_ROOT;
263         atomic_set(&cur_trans->delayed_refs.num_entries, 0);
264
265         /*
266          * although the tree mod log is per file system and not per transaction,
267          * the log must never go across transaction boundaries.
268          */
269         smp_mb();
270         if (!list_empty(&fs_info->tree_mod_seq_list))
271                 WARN(1, KERN_ERR "BTRFS: tree_mod_seq_list not empty when creating a fresh transaction\n");
272         if (!RB_EMPTY_ROOT(&fs_info->tree_mod_log))
273                 WARN(1, KERN_ERR "BTRFS: tree_mod_log rb tree not empty when creating a fresh transaction\n");
274         atomic64_set(&fs_info->tree_mod_seq, 0);
275
276         spin_lock_init(&cur_trans->delayed_refs.lock);
277
278         INIT_LIST_HEAD(&cur_trans->pending_snapshots);
279         INIT_LIST_HEAD(&cur_trans->pending_chunks);
280         INIT_LIST_HEAD(&cur_trans->switch_commits);
281         INIT_LIST_HEAD(&cur_trans->dirty_bgs);
282         INIT_LIST_HEAD(&cur_trans->io_bgs);
283         INIT_LIST_HEAD(&cur_trans->dropped_roots);
284         mutex_init(&cur_trans->cache_write_mutex);
285         cur_trans->num_dirty_bgs = 0;
286         spin_lock_init(&cur_trans->dirty_bgs_lock);
287         INIT_LIST_HEAD(&cur_trans->deleted_bgs);
288         spin_lock_init(&cur_trans->dropped_roots_lock);
289         list_add_tail(&cur_trans->list, &fs_info->trans_list);
290         extent_io_tree_init(&cur_trans->dirty_pages,
291                              fs_info->btree_inode);
292         fs_info->generation++;
293         cur_trans->transid = fs_info->generation;
294         fs_info->running_transaction = cur_trans;
295         cur_trans->aborted = 0;
296         spin_unlock(&fs_info->trans_lock);
297
298         return 0;
299 }
300
301 /*
302  * this does all the record keeping required to make sure that a reference
303  * counted root is properly recorded in a given transaction.  This is required
304  * to make sure the old root from before we joined the transaction is deleted
305  * when the transaction commits
306  */
307 static int record_root_in_trans(struct btrfs_trans_handle *trans,
308                                struct btrfs_root *root,
309                                int force)
310 {
311         struct btrfs_fs_info *fs_info = root->fs_info;
312
313         if ((test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
314             root->last_trans < trans->transid) || force) {
315                 WARN_ON(root == fs_info->extent_root);
316                 WARN_ON(!force && root->commit_root != root->node);
317
318                 /*
319                  * see below for IN_TRANS_SETUP usage rules
320                  * we have the reloc mutex held now, so there
321                  * is only one writer in this function
322                  */
323                 set_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
324
325                 /* make sure readers find IN_TRANS_SETUP before
326                  * they find our root->last_trans update
327                  */
328                 smp_wmb();
329
330                 spin_lock(&fs_info->fs_roots_radix_lock);
331                 if (root->last_trans == trans->transid && !force) {
332                         spin_unlock(&fs_info->fs_roots_radix_lock);
333                         return 0;
334                 }
335                 radix_tree_tag_set(&fs_info->fs_roots_radix,
336                                    (unsigned long)root->root_key.objectid,
337                                    BTRFS_ROOT_TRANS_TAG);
338                 spin_unlock(&fs_info->fs_roots_radix_lock);
339                 root->last_trans = trans->transid;
340
341                 /* this is pretty tricky.  We don't want to
342                  * take the relocation lock in btrfs_record_root_in_trans
343                  * unless we're really doing the first setup for this root in
344                  * this transaction.
345                  *
346                  * Normally we'd use root->last_trans as a flag to decide
347                  * if we want to take the expensive mutex.
348                  *
349                  * But, we have to set root->last_trans before we
350                  * init the relocation root, otherwise, we trip over warnings
351                  * in ctree.c.  The solution used here is to flag ourselves
352                  * with root IN_TRANS_SETUP.  When this is 1, we're still
353                  * fixing up the reloc trees and everyone must wait.
354                  *
355                  * When this is zero, they can trust root->last_trans and fly
356                  * through btrfs_record_root_in_trans without having to take the
357                  * lock.  smp_wmb() makes sure that all the writes above are
358                  * done before we pop in the zero below
359                  */
360                 btrfs_init_reloc_root(trans, root);
361                 smp_mb__before_atomic();
362                 clear_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
363         }
364         return 0;
365 }
366
367
368 void btrfs_add_dropped_root(struct btrfs_trans_handle *trans,
369                             struct btrfs_root *root)
370 {
371         struct btrfs_fs_info *fs_info = root->fs_info;
372         struct btrfs_transaction *cur_trans = trans->transaction;
373
374         /* Add ourselves to the transaction dropped list */
375         spin_lock(&cur_trans->dropped_roots_lock);
376         list_add_tail(&root->root_list, &cur_trans->dropped_roots);
377         spin_unlock(&cur_trans->dropped_roots_lock);
378
379         /* Make sure we don't try to update the root at commit time */
380         spin_lock(&fs_info->fs_roots_radix_lock);
381         radix_tree_tag_clear(&fs_info->fs_roots_radix,
382                              (unsigned long)root->root_key.objectid,
383                              BTRFS_ROOT_TRANS_TAG);
384         spin_unlock(&fs_info->fs_roots_radix_lock);
385 }
386
387 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
388                                struct btrfs_root *root)
389 {
390         struct btrfs_fs_info *fs_info = root->fs_info;
391
392         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
393                 return 0;
394
395         /*
396          * see record_root_in_trans for comments about IN_TRANS_SETUP usage
397          * and barriers
398          */
399         smp_rmb();
400         if (root->last_trans == trans->transid &&
401             !test_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state))
402                 return 0;
403
404         mutex_lock(&fs_info->reloc_mutex);
405         record_root_in_trans(trans, root, 0);
406         mutex_unlock(&fs_info->reloc_mutex);
407
408         return 0;
409 }
410
411 static inline int is_transaction_blocked(struct btrfs_transaction *trans)
412 {
413         return (trans->state >= TRANS_STATE_BLOCKED &&
414                 trans->state < TRANS_STATE_UNBLOCKED &&
415                 !trans->aborted);
416 }
417
418 /* wait for commit against the current transaction to become unblocked
419  * when this is done, it is safe to start a new transaction, but the current
420  * transaction might not be fully on disk.
421  */
422 static void wait_current_trans(struct btrfs_fs_info *fs_info)
423 {
424         struct btrfs_transaction *cur_trans;
425
426         spin_lock(&fs_info->trans_lock);
427         cur_trans = fs_info->running_transaction;
428         if (cur_trans && is_transaction_blocked(cur_trans)) {
429                 refcount_inc(&cur_trans->use_count);
430                 spin_unlock(&fs_info->trans_lock);
431
432                 wait_event(fs_info->transaction_wait,
433                            cur_trans->state >= TRANS_STATE_UNBLOCKED ||
434                            cur_trans->aborted);
435                 btrfs_put_transaction(cur_trans);
436         } else {
437                 spin_unlock(&fs_info->trans_lock);
438         }
439 }
440
441 static int may_wait_transaction(struct btrfs_fs_info *fs_info, int type)
442 {
443         if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
444                 return 0;
445
446         if (type == TRANS_START &&
447             !atomic_read(&fs_info->open_ioctl_trans))
448                 return 1;
449
450         return 0;
451 }
452
453 static inline bool need_reserve_reloc_root(struct btrfs_root *root)
454 {
455         struct btrfs_fs_info *fs_info = root->fs_info;
456
457         if (!fs_info->reloc_ctl ||
458             !test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
459             root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
460             root->reloc_root)
461                 return false;
462
463         return true;
464 }
465
466 static struct btrfs_trans_handle *
467 start_transaction(struct btrfs_root *root, unsigned int num_items,
468                   unsigned int type, enum btrfs_reserve_flush_enum flush,
469                   bool enforce_qgroups)
470 {
471         struct btrfs_fs_info *fs_info = root->fs_info;
472
473         struct btrfs_trans_handle *h;
474         struct btrfs_transaction *cur_trans;
475         u64 num_bytes = 0;
476         u64 qgroup_reserved = 0;
477         bool reloc_reserved = false;
478         int ret;
479
480         /* Send isn't supposed to start transactions. */
481         ASSERT(current->journal_info != BTRFS_SEND_TRANS_STUB);
482
483         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
484                 return ERR_PTR(-EROFS);
485
486         if (current->journal_info) {
487                 WARN_ON(type & TRANS_EXTWRITERS);
488                 h = current->journal_info;
489                 refcount_inc(&h->use_count);
490                 WARN_ON(refcount_read(&h->use_count) > 2);
491                 h->orig_rsv = h->block_rsv;
492                 h->block_rsv = NULL;
493                 goto got_it;
494         }
495
496         /*
497          * Do the reservation before we join the transaction so we can do all
498          * the appropriate flushing if need be.
499          */
500         if (num_items && root != fs_info->chunk_root) {
501                 qgroup_reserved = num_items * fs_info->nodesize;
502                 ret = btrfs_qgroup_reserve_meta(root, qgroup_reserved,
503                                                 enforce_qgroups);
504                 if (ret)
505                         return ERR_PTR(ret);
506
507                 num_bytes = btrfs_calc_trans_metadata_size(fs_info, num_items);
508                 /*
509                  * Do the reservation for the relocation root creation
510                  */
511                 if (need_reserve_reloc_root(root)) {
512                         num_bytes += fs_info->nodesize;
513                         reloc_reserved = true;
514                 }
515
516                 ret = btrfs_block_rsv_add(root, &fs_info->trans_block_rsv,
517                                           num_bytes, flush);
518                 if (ret)
519                         goto reserve_fail;
520         }
521 again:
522         h = kmem_cache_zalloc(btrfs_trans_handle_cachep, GFP_NOFS);
523         if (!h) {
524                 ret = -ENOMEM;
525                 goto alloc_fail;
526         }
527
528         /*
529          * If we are JOIN_NOLOCK we're already committing a transaction and
530          * waiting on this guy, so we don't need to do the sb_start_intwrite
531          * because we're already holding a ref.  We need this because we could
532          * have raced in and did an fsync() on a file which can kick a commit
533          * and then we deadlock with somebody doing a freeze.
534          *
535          * If we are ATTACH, it means we just want to catch the current
536          * transaction and commit it, so we needn't do sb_start_intwrite(). 
537          */
538         if (type & __TRANS_FREEZABLE)
539                 sb_start_intwrite(fs_info->sb);
540
541         if (may_wait_transaction(fs_info, type))
542                 wait_current_trans(fs_info);
543
544         do {
545                 ret = join_transaction(fs_info, type);
546                 if (ret == -EBUSY) {
547                         wait_current_trans(fs_info);
548                         if (unlikely(type == TRANS_ATTACH))
549                                 ret = -ENOENT;
550                 }
551         } while (ret == -EBUSY);
552
553         if (ret < 0)
554                 goto join_fail;
555
556         cur_trans = fs_info->running_transaction;
557
558         h->transid = cur_trans->transid;
559         h->transaction = cur_trans;
560         h->root = root;
561         refcount_set(&h->use_count, 1);
562         h->fs_info = root->fs_info;
563
564         h->type = type;
565         h->can_flush_pending_bgs = true;
566         INIT_LIST_HEAD(&h->new_bgs);
567
568         smp_mb();
569         if (cur_trans->state >= TRANS_STATE_BLOCKED &&
570             may_wait_transaction(fs_info, type)) {
571                 current->journal_info = h;
572                 btrfs_commit_transaction(h);
573                 goto again;
574         }
575
576         if (num_bytes) {
577                 trace_btrfs_space_reservation(fs_info, "transaction",
578                                               h->transid, num_bytes, 1);
579                 h->block_rsv = &fs_info->trans_block_rsv;
580                 h->bytes_reserved = num_bytes;
581                 h->reloc_reserved = reloc_reserved;
582         }
583
584 got_it:
585         btrfs_record_root_in_trans(h, root);
586
587         if (!current->journal_info)
588                 current->journal_info = h;
589         return h;
590
591 join_fail:
592         if (type & __TRANS_FREEZABLE)
593                 sb_end_intwrite(fs_info->sb);
594         kmem_cache_free(btrfs_trans_handle_cachep, h);
595 alloc_fail:
596         if (num_bytes)
597                 btrfs_block_rsv_release(fs_info, &fs_info->trans_block_rsv,
598                                         num_bytes);
599 reserve_fail:
600         btrfs_qgroup_free_meta(root, qgroup_reserved);
601         return ERR_PTR(ret);
602 }
603
604 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
605                                                    unsigned int num_items)
606 {
607         return start_transaction(root, num_items, TRANS_START,
608                                  BTRFS_RESERVE_FLUSH_ALL, true);
609 }
610
611 struct btrfs_trans_handle *btrfs_start_transaction_fallback_global_rsv(
612                                         struct btrfs_root *root,
613                                         unsigned int num_items,
614                                         int min_factor)
615 {
616         struct btrfs_fs_info *fs_info = root->fs_info;
617         struct btrfs_trans_handle *trans;
618         u64 num_bytes;
619         int ret;
620
621         /*
622          * We have two callers: unlink and block group removal.  The
623          * former should succeed even if we will temporarily exceed
624          * quota and the latter operates on the extent root so
625          * qgroup enforcement is ignored anyway.
626          */
627         trans = start_transaction(root, num_items, TRANS_START,
628                                   BTRFS_RESERVE_FLUSH_ALL, false);
629         if (!IS_ERR(trans) || PTR_ERR(trans) != -ENOSPC)
630                 return trans;
631
632         trans = btrfs_start_transaction(root, 0);
633         if (IS_ERR(trans))
634                 return trans;
635
636         num_bytes = btrfs_calc_trans_metadata_size(fs_info, num_items);
637         ret = btrfs_cond_migrate_bytes(fs_info, &fs_info->trans_block_rsv,
638                                        num_bytes, min_factor);
639         if (ret) {
640                 btrfs_end_transaction(trans);
641                 return ERR_PTR(ret);
642         }
643
644         trans->block_rsv = &fs_info->trans_block_rsv;
645         trans->bytes_reserved = num_bytes;
646         trace_btrfs_space_reservation(fs_info, "transaction",
647                                       trans->transid, num_bytes, 1);
648
649         return trans;
650 }
651
652 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
653 {
654         return start_transaction(root, 0, TRANS_JOIN, BTRFS_RESERVE_NO_FLUSH,
655                                  true);
656 }
657
658 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root)
659 {
660         return start_transaction(root, 0, TRANS_JOIN_NOLOCK,
661                                  BTRFS_RESERVE_NO_FLUSH, true);
662 }
663
664 /*
665  * btrfs_attach_transaction() - catch the running transaction
666  *
667  * It is used when we want to commit the current the transaction, but
668  * don't want to start a new one.
669  *
670  * Note: If this function return -ENOENT, it just means there is no
671  * running transaction. But it is possible that the inactive transaction
672  * is still in the memory, not fully on disk. If you hope there is no
673  * inactive transaction in the fs when -ENOENT is returned, you should
674  * invoke
675  *     btrfs_attach_transaction_barrier()
676  */
677 struct btrfs_trans_handle *btrfs_attach_transaction(struct btrfs_root *root)
678 {
679         return start_transaction(root, 0, TRANS_ATTACH,
680                                  BTRFS_RESERVE_NO_FLUSH, true);
681 }
682
683 /*
684  * btrfs_attach_transaction_barrier() - catch the running transaction
685  *
686  * It is similar to the above function, the differentia is this one
687  * will wait for all the inactive transactions until they fully
688  * complete.
689  */
690 struct btrfs_trans_handle *
691 btrfs_attach_transaction_barrier(struct btrfs_root *root)
692 {
693         struct btrfs_trans_handle *trans;
694
695         trans = start_transaction(root, 0, TRANS_ATTACH,
696                                   BTRFS_RESERVE_NO_FLUSH, true);
697         if (IS_ERR(trans) && PTR_ERR(trans) == -ENOENT)
698                 btrfs_wait_for_commit(root->fs_info, 0);
699
700         return trans;
701 }
702
703 /* wait for a transaction commit to be fully complete */
704 static noinline void wait_for_commit(struct btrfs_transaction *commit)
705 {
706         wait_event(commit->commit_wait, commit->state == TRANS_STATE_COMPLETED);
707 }
708
709 int btrfs_wait_for_commit(struct btrfs_fs_info *fs_info, u64 transid)
710 {
711         struct btrfs_transaction *cur_trans = NULL, *t;
712         int ret = 0;
713
714         if (transid) {
715                 if (transid <= fs_info->last_trans_committed)
716                         goto out;
717
718                 /* find specified transaction */
719                 spin_lock(&fs_info->trans_lock);
720                 list_for_each_entry(t, &fs_info->trans_list, list) {
721                         if (t->transid == transid) {
722                                 cur_trans = t;
723                                 refcount_inc(&cur_trans->use_count);
724                                 ret = 0;
725                                 break;
726                         }
727                         if (t->transid > transid) {
728                                 ret = 0;
729                                 break;
730                         }
731                 }
732                 spin_unlock(&fs_info->trans_lock);
733
734                 /*
735                  * The specified transaction doesn't exist, or we
736                  * raced with btrfs_commit_transaction
737                  */
738                 if (!cur_trans) {
739                         if (transid > fs_info->last_trans_committed)
740                                 ret = -EINVAL;
741                         goto out;
742                 }
743         } else {
744                 /* find newest transaction that is committing | committed */
745                 spin_lock(&fs_info->trans_lock);
746                 list_for_each_entry_reverse(t, &fs_info->trans_list,
747                                             list) {
748                         if (t->state >= TRANS_STATE_COMMIT_START) {
749                                 if (t->state == TRANS_STATE_COMPLETED)
750                                         break;
751                                 cur_trans = t;
752                                 refcount_inc(&cur_trans->use_count);
753                                 break;
754                         }
755                 }
756                 spin_unlock(&fs_info->trans_lock);
757                 if (!cur_trans)
758                         goto out;  /* nothing committing|committed */
759         }
760
761         wait_for_commit(cur_trans);
762         btrfs_put_transaction(cur_trans);
763 out:
764         return ret;
765 }
766
767 void btrfs_throttle(struct btrfs_fs_info *fs_info)
768 {
769         if (!atomic_read(&fs_info->open_ioctl_trans))
770                 wait_current_trans(fs_info);
771 }
772
773 static int should_end_transaction(struct btrfs_trans_handle *trans)
774 {
775         struct btrfs_fs_info *fs_info = trans->fs_info;
776
777         if (btrfs_check_space_for_delayed_refs(trans, fs_info))
778                 return 1;
779
780         return !!btrfs_block_rsv_check(&fs_info->global_block_rsv, 5);
781 }
782
783 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans)
784 {
785         struct btrfs_transaction *cur_trans = trans->transaction;
786         struct btrfs_fs_info *fs_info = trans->fs_info;
787         int updates;
788         int err;
789
790         smp_mb();
791         if (cur_trans->state >= TRANS_STATE_BLOCKED ||
792             cur_trans->delayed_refs.flushing)
793                 return 1;
794
795         updates = trans->delayed_ref_updates;
796         trans->delayed_ref_updates = 0;
797         if (updates) {
798                 err = btrfs_run_delayed_refs(trans, fs_info, updates * 2);
799                 if (err) /* Error code will also eval true */
800                         return err;
801         }
802
803         return should_end_transaction(trans);
804 }
805
806 static void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans)
807
808 {
809         struct btrfs_fs_info *fs_info = trans->fs_info;
810
811         if (!trans->block_rsv) {
812                 ASSERT(!trans->bytes_reserved);
813                 return;
814         }
815
816         if (!trans->bytes_reserved)
817                 return;
818
819         ASSERT(trans->block_rsv == &fs_info->trans_block_rsv);
820         trace_btrfs_space_reservation(fs_info, "transaction",
821                                       trans->transid, trans->bytes_reserved, 0);
822         btrfs_block_rsv_release(fs_info, trans->block_rsv,
823                                 trans->bytes_reserved);
824         trans->bytes_reserved = 0;
825 }
826
827 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
828                                    int throttle)
829 {
830         struct btrfs_fs_info *info = trans->fs_info;
831         struct btrfs_transaction *cur_trans = trans->transaction;
832         u64 transid = trans->transid;
833         unsigned long cur = trans->delayed_ref_updates;
834         int lock = (trans->type != TRANS_JOIN_NOLOCK);
835         int err = 0;
836         int must_run_delayed_refs = 0;
837
838         if (refcount_read(&trans->use_count) > 1) {
839                 refcount_dec(&trans->use_count);
840                 trans->block_rsv = trans->orig_rsv;
841                 return 0;
842         }
843
844         btrfs_trans_release_metadata(trans);
845         trans->block_rsv = NULL;
846
847         if (!list_empty(&trans->new_bgs))
848                 btrfs_create_pending_block_groups(trans);
849
850         trans->delayed_ref_updates = 0;
851         if (!trans->sync) {
852                 must_run_delayed_refs =
853                         btrfs_should_throttle_delayed_refs(trans, info);
854                 cur = max_t(unsigned long, cur, 32);
855
856                 /*
857                  * don't make the caller wait if they are from a NOLOCK
858                  * or ATTACH transaction, it will deadlock with commit
859                  */
860                 if (must_run_delayed_refs == 1 &&
861                     (trans->type & (__TRANS_JOIN_NOLOCK | __TRANS_ATTACH)))
862                         must_run_delayed_refs = 2;
863         }
864
865         btrfs_trans_release_metadata(trans);
866         trans->block_rsv = NULL;
867
868         if (!list_empty(&trans->new_bgs))
869                 btrfs_create_pending_block_groups(trans);
870
871         btrfs_trans_release_chunk_metadata(trans);
872
873         if (lock && !atomic_read(&info->open_ioctl_trans) &&
874             should_end_transaction(trans) &&
875             READ_ONCE(cur_trans->state) == TRANS_STATE_RUNNING) {
876                 spin_lock(&info->trans_lock);
877                 if (cur_trans->state == TRANS_STATE_RUNNING)
878                         cur_trans->state = TRANS_STATE_BLOCKED;
879                 spin_unlock(&info->trans_lock);
880         }
881
882         if (lock && READ_ONCE(cur_trans->state) == TRANS_STATE_BLOCKED) {
883                 if (throttle)
884                         return btrfs_commit_transaction(trans);
885                 else
886                         wake_up_process(info->transaction_kthread);
887         }
888
889         if (trans->type & __TRANS_FREEZABLE)
890                 sb_end_intwrite(info->sb);
891
892         WARN_ON(cur_trans != info->running_transaction);
893         WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
894         atomic_dec(&cur_trans->num_writers);
895         extwriter_counter_dec(cur_trans, trans->type);
896
897         /*
898          * Make sure counter is updated before we wake up waiters.
899          */
900         smp_mb();
901         if (waitqueue_active(&cur_trans->writer_wait))
902                 wake_up(&cur_trans->writer_wait);
903         btrfs_put_transaction(cur_trans);
904
905         if (current->journal_info == trans)
906                 current->journal_info = NULL;
907
908         if (throttle)
909                 btrfs_run_delayed_iputs(info);
910
911         if (trans->aborted ||
912             test_bit(BTRFS_FS_STATE_ERROR, &info->fs_state)) {
913                 wake_up_process(info->transaction_kthread);
914                 err = -EIO;
915         }
916
917         kmem_cache_free(btrfs_trans_handle_cachep, trans);
918         if (must_run_delayed_refs) {
919                 btrfs_async_run_delayed_refs(info, cur, transid,
920                                              must_run_delayed_refs == 1);
921         }
922         return err;
923 }
924
925 int btrfs_end_transaction(struct btrfs_trans_handle *trans)
926 {
927         return __btrfs_end_transaction(trans, 0);
928 }
929
930 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans)
931 {
932         return __btrfs_end_transaction(trans, 1);
933 }
934
935 /*
936  * when btree blocks are allocated, they have some corresponding bits set for
937  * them in one of two extent_io trees.  This is used to make sure all of
938  * those extents are sent to disk but does not wait on them
939  */
940 int btrfs_write_marked_extents(struct btrfs_fs_info *fs_info,
941                                struct extent_io_tree *dirty_pages, int mark)
942 {
943         int err = 0;
944         int werr = 0;
945         struct address_space *mapping = fs_info->btree_inode->i_mapping;
946         struct extent_state *cached_state = NULL;
947         u64 start = 0;
948         u64 end;
949
950         atomic_inc(&BTRFS_I(fs_info->btree_inode)->sync_writers);
951         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
952                                       mark, &cached_state)) {
953                 bool wait_writeback = false;
954
955                 err = convert_extent_bit(dirty_pages, start, end,
956                                          EXTENT_NEED_WAIT,
957                                          mark, &cached_state);
958                 /*
959                  * convert_extent_bit can return -ENOMEM, which is most of the
960                  * time a temporary error. So when it happens, ignore the error
961                  * and wait for writeback of this range to finish - because we
962                  * failed to set the bit EXTENT_NEED_WAIT for the range, a call
963                  * to __btrfs_wait_marked_extents() would not know that
964                  * writeback for this range started and therefore wouldn't
965                  * wait for it to finish - we don't want to commit a
966                  * superblock that points to btree nodes/leafs for which
967                  * writeback hasn't finished yet (and without errors).
968                  * We cleanup any entries left in the io tree when committing
969                  * the transaction (through clear_btree_io_tree()).
970                  */
971                 if (err == -ENOMEM) {
972                         err = 0;
973                         wait_writeback = true;
974                 }
975                 if (!err)
976                         err = filemap_fdatawrite_range(mapping, start, end);
977                 if (err)
978                         werr = err;
979                 else if (wait_writeback)
980                         werr = filemap_fdatawait_range(mapping, start, end);
981                 free_extent_state(cached_state);
982                 cached_state = NULL;
983                 cond_resched();
984                 start = end + 1;
985         }
986         atomic_dec(&BTRFS_I(fs_info->btree_inode)->sync_writers);
987         return werr;
988 }
989
990 /*
991  * when btree blocks are allocated, they have some corresponding bits set for
992  * them in one of two extent_io trees.  This is used to make sure all of
993  * those extents are on disk for transaction or log commit.  We wait
994  * on all the pages and clear them from the dirty pages state tree
995  */
996 static int __btrfs_wait_marked_extents(struct btrfs_fs_info *fs_info,
997                                        struct extent_io_tree *dirty_pages)
998 {
999         int err = 0;
1000         int werr = 0;
1001         struct address_space *mapping = fs_info->btree_inode->i_mapping;
1002         struct extent_state *cached_state = NULL;
1003         u64 start = 0;
1004         u64 end;
1005
1006         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
1007                                       EXTENT_NEED_WAIT, &cached_state)) {
1008                 /*
1009                  * Ignore -ENOMEM errors returned by clear_extent_bit().
1010                  * When committing the transaction, we'll remove any entries
1011                  * left in the io tree. For a log commit, we don't remove them
1012                  * after committing the log because the tree can be accessed
1013                  * concurrently - we do it only at transaction commit time when
1014                  * it's safe to do it (through clear_btree_io_tree()).
1015                  */
1016                 err = clear_extent_bit(dirty_pages, start, end,
1017                                        EXTENT_NEED_WAIT, 0, 0, &cached_state);
1018                 if (err == -ENOMEM)
1019                         err = 0;
1020                 if (!err)
1021                         err = filemap_fdatawait_range(mapping, start, end);
1022                 if (err)
1023                         werr = err;
1024                 free_extent_state(cached_state);
1025                 cached_state = NULL;
1026                 cond_resched();
1027                 start = end + 1;
1028         }
1029         if (err)
1030                 werr = err;
1031         return werr;
1032 }
1033
1034 int btrfs_wait_extents(struct btrfs_fs_info *fs_info,
1035                        struct extent_io_tree *dirty_pages)
1036 {
1037         bool errors = false;
1038         int err;
1039
1040         err = __btrfs_wait_marked_extents(fs_info, dirty_pages);
1041         if (test_and_clear_bit(BTRFS_FS_BTREE_ERR, &fs_info->flags))
1042                 errors = true;
1043
1044         if (errors && !err)
1045                 err = -EIO;
1046         return err;
1047 }
1048
1049 int btrfs_wait_tree_log_extents(struct btrfs_root *log_root, int mark)
1050 {
1051         struct btrfs_fs_info *fs_info = log_root->fs_info;
1052         struct extent_io_tree *dirty_pages = &log_root->dirty_log_pages;
1053         bool errors = false;
1054         int err;
1055
1056         ASSERT(log_root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
1057
1058         err = __btrfs_wait_marked_extents(fs_info, dirty_pages);
1059         if ((mark & EXTENT_DIRTY) &&
1060             test_and_clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags))
1061                 errors = true;
1062
1063         if ((mark & EXTENT_NEW) &&
1064             test_and_clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags))
1065                 errors = true;
1066
1067         if (errors && !err)
1068                 err = -EIO;
1069         return err;
1070 }
1071
1072 /*
1073  * When btree blocks are allocated the corresponding extents are marked dirty.
1074  * This function ensures such extents are persisted on disk for transaction or
1075  * log commit.
1076  *
1077  * @trans: transaction whose dirty pages we'd like to write
1078  */
1079 static int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans)
1080 {
1081         int ret;
1082         int ret2;
1083         struct extent_io_tree *dirty_pages = &trans->transaction->dirty_pages;
1084         struct btrfs_fs_info *fs_info = trans->fs_info;
1085         struct blk_plug plug;
1086
1087         blk_start_plug(&plug);
1088         ret = btrfs_write_marked_extents(fs_info, dirty_pages, EXTENT_DIRTY);
1089         blk_finish_plug(&plug);
1090         ret2 = btrfs_wait_extents(fs_info, dirty_pages);
1091
1092         clear_btree_io_tree(&trans->transaction->dirty_pages);
1093
1094         if (ret)
1095                 return ret;
1096         else if (ret2)
1097                 return ret2;
1098         else
1099                 return 0;
1100 }
1101
1102 /*
1103  * this is used to update the root pointer in the tree of tree roots.
1104  *
1105  * But, in the case of the extent allocation tree, updating the root
1106  * pointer may allocate blocks which may change the root of the extent
1107  * allocation tree.
1108  *
1109  * So, this loops and repeats and makes sure the cowonly root didn't
1110  * change while the root pointer was being updated in the metadata.
1111  */
1112 static int update_cowonly_root(struct btrfs_trans_handle *trans,
1113                                struct btrfs_root *root)
1114 {
1115         int ret;
1116         u64 old_root_bytenr;
1117         u64 old_root_used;
1118         struct btrfs_fs_info *fs_info = root->fs_info;
1119         struct btrfs_root *tree_root = fs_info->tree_root;
1120
1121         old_root_used = btrfs_root_used(&root->root_item);
1122
1123         while (1) {
1124                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
1125                 if (old_root_bytenr == root->node->start &&
1126                     old_root_used == btrfs_root_used(&root->root_item))
1127                         break;
1128
1129                 btrfs_set_root_node(&root->root_item, root->node);
1130                 ret = btrfs_update_root(trans, tree_root,
1131                                         &root->root_key,
1132                                         &root->root_item);
1133                 if (ret)
1134                         return ret;
1135
1136                 old_root_used = btrfs_root_used(&root->root_item);
1137         }
1138
1139         return 0;
1140 }
1141
1142 /*
1143  * update all the cowonly tree roots on disk
1144  *
1145  * The error handling in this function may not be obvious. Any of the
1146  * failures will cause the file system to go offline. We still need
1147  * to clean up the delayed refs.
1148  */
1149 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans)
1150 {
1151         struct btrfs_fs_info *fs_info = trans->fs_info;
1152         struct list_head *dirty_bgs = &trans->transaction->dirty_bgs;
1153         struct list_head *io_bgs = &trans->transaction->io_bgs;
1154         struct list_head *next;
1155         struct extent_buffer *eb;
1156         int ret;
1157
1158         eb = btrfs_lock_root_node(fs_info->tree_root);
1159         ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL,
1160                               0, &eb);
1161         btrfs_tree_unlock(eb);
1162         free_extent_buffer(eb);
1163
1164         if (ret)
1165                 return ret;
1166
1167         ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1168         if (ret)
1169                 return ret;
1170
1171         ret = btrfs_run_dev_stats(trans, fs_info);
1172         if (ret)
1173                 return ret;
1174         ret = btrfs_run_dev_replace(trans, fs_info);
1175         if (ret)
1176                 return ret;
1177         ret = btrfs_run_qgroups(trans, fs_info);
1178         if (ret)
1179                 return ret;
1180
1181         ret = btrfs_setup_space_cache(trans, fs_info);
1182         if (ret)
1183                 return ret;
1184
1185         /* run_qgroups might have added some more refs */
1186         ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1187         if (ret)
1188                 return ret;
1189 again:
1190         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
1191                 struct btrfs_root *root;
1192                 next = fs_info->dirty_cowonly_roots.next;
1193                 list_del_init(next);
1194                 root = list_entry(next, struct btrfs_root, dirty_list);
1195                 clear_bit(BTRFS_ROOT_DIRTY, &root->state);
1196
1197                 if (root != fs_info->extent_root)
1198                         list_add_tail(&root->dirty_list,
1199                                       &trans->transaction->switch_commits);
1200                 ret = update_cowonly_root(trans, root);
1201                 if (ret)
1202                         return ret;
1203                 ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1204                 if (ret)
1205                         return ret;
1206         }
1207
1208         while (!list_empty(dirty_bgs) || !list_empty(io_bgs)) {
1209                 ret = btrfs_write_dirty_block_groups(trans, fs_info);
1210                 if (ret)
1211                         return ret;
1212                 ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1213                 if (ret)
1214                         return ret;
1215         }
1216
1217         if (!list_empty(&fs_info->dirty_cowonly_roots))
1218                 goto again;
1219
1220         list_add_tail(&fs_info->extent_root->dirty_list,
1221                       &trans->transaction->switch_commits);
1222         btrfs_after_dev_replace_commit(fs_info);
1223
1224         return 0;
1225 }
1226
1227 /*
1228  * dead roots are old snapshots that need to be deleted.  This allocates
1229  * a dirty root struct and adds it into the list of dead roots that need to
1230  * be deleted
1231  */
1232 void btrfs_add_dead_root(struct btrfs_root *root)
1233 {
1234         struct btrfs_fs_info *fs_info = root->fs_info;
1235
1236         spin_lock(&fs_info->trans_lock);
1237         if (list_empty(&root->root_list))
1238                 list_add_tail(&root->root_list, &fs_info->dead_roots);
1239         spin_unlock(&fs_info->trans_lock);
1240 }
1241
1242 /*
1243  * update all the cowonly tree roots on disk
1244  */
1245 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans)
1246 {
1247         struct btrfs_fs_info *fs_info = trans->fs_info;
1248         struct btrfs_root *gang[8];
1249         int i;
1250         int ret;
1251         int err = 0;
1252
1253         spin_lock(&fs_info->fs_roots_radix_lock);
1254         while (1) {
1255                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
1256                                                  (void **)gang, 0,
1257                                                  ARRAY_SIZE(gang),
1258                                                  BTRFS_ROOT_TRANS_TAG);
1259                 if (ret == 0)
1260                         break;
1261                 for (i = 0; i < ret; i++) {
1262                         struct btrfs_root *root = gang[i];
1263                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
1264                                         (unsigned long)root->root_key.objectid,
1265                                         BTRFS_ROOT_TRANS_TAG);
1266                         spin_unlock(&fs_info->fs_roots_radix_lock);
1267
1268                         btrfs_free_log(trans, root);
1269                         btrfs_update_reloc_root(trans, root);
1270                         btrfs_orphan_commit_root(trans, root);
1271
1272                         btrfs_save_ino_cache(root, trans);
1273
1274                         /* see comments in should_cow_block() */
1275                         clear_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1276                         smp_mb__after_atomic();
1277
1278                         if (root->commit_root != root->node) {
1279                                 list_add_tail(&root->dirty_list,
1280                                         &trans->transaction->switch_commits);
1281                                 btrfs_set_root_node(&root->root_item,
1282                                                     root->node);
1283                         }
1284
1285                         err = btrfs_update_root(trans, fs_info->tree_root,
1286                                                 &root->root_key,
1287                                                 &root->root_item);
1288                         spin_lock(&fs_info->fs_roots_radix_lock);
1289                         if (err)
1290                                 break;
1291                         btrfs_qgroup_free_meta_all(root);
1292                 }
1293         }
1294         spin_unlock(&fs_info->fs_roots_radix_lock);
1295         return err;
1296 }
1297
1298 /*
1299  * defrag a given btree.
1300  * Every leaf in the btree is read and defragged.
1301  */
1302 int btrfs_defrag_root(struct btrfs_root *root)
1303 {
1304         struct btrfs_fs_info *info = root->fs_info;
1305         struct btrfs_trans_handle *trans;
1306         int ret;
1307
1308         if (test_and_set_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state))
1309                 return 0;
1310
1311         while (1) {
1312                 trans = btrfs_start_transaction(root, 0);
1313                 if (IS_ERR(trans))
1314                         return PTR_ERR(trans);
1315
1316                 ret = btrfs_defrag_leaves(trans, root);
1317
1318                 btrfs_end_transaction(trans);
1319                 btrfs_btree_balance_dirty(info);
1320                 cond_resched();
1321
1322                 if (btrfs_fs_closing(info) || ret != -EAGAIN)
1323                         break;
1324
1325                 if (btrfs_defrag_cancelled(info)) {
1326                         btrfs_debug(info, "defrag_root cancelled");
1327                         ret = -EAGAIN;
1328                         break;
1329                 }
1330         }
1331         clear_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state);
1332         return ret;
1333 }
1334
1335 /*
1336  * Do all special snapshot related qgroup dirty hack.
1337  *
1338  * Will do all needed qgroup inherit and dirty hack like switch commit
1339  * roots inside one transaction and write all btree into disk, to make
1340  * qgroup works.
1341  */
1342 static int qgroup_account_snapshot(struct btrfs_trans_handle *trans,
1343                                    struct btrfs_root *src,
1344                                    struct btrfs_root *parent,
1345                                    struct btrfs_qgroup_inherit *inherit,
1346                                    u64 dst_objectid)
1347 {
1348         struct btrfs_fs_info *fs_info = src->fs_info;
1349         int ret;
1350
1351         /*
1352          * Save some performance in the case that qgroups are not
1353          * enabled. If this check races with the ioctl, rescan will
1354          * kick in anyway.
1355          */
1356         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
1357                 return 0;
1358
1359         /*
1360          * Ensure dirty @src will be commited.  Or, after comming
1361          * commit_fs_roots() and switch_commit_roots(), any dirty but not
1362          * recorded root will never be updated again, causing an outdated root
1363          * item.
1364          */
1365         record_root_in_trans(trans, src, 1);
1366
1367         /*
1368          * We are going to commit transaction, see btrfs_commit_transaction()
1369          * comment for reason locking tree_log_mutex
1370          */
1371         mutex_lock(&fs_info->tree_log_mutex);
1372
1373         ret = commit_fs_roots(trans);
1374         if (ret)
1375                 goto out;
1376         ret = btrfs_qgroup_account_extents(trans, fs_info);
1377         if (ret < 0)
1378                 goto out;
1379
1380         /* Now qgroup are all updated, we can inherit it to new qgroups */
1381         ret = btrfs_qgroup_inherit(trans, fs_info,
1382                                    src->root_key.objectid, dst_objectid,
1383                                    inherit);
1384         if (ret < 0)
1385                 goto out;
1386
1387         /*
1388          * Now we do a simplified commit transaction, which will:
1389          * 1) commit all subvolume and extent tree
1390          *    To ensure all subvolume and extent tree have a valid
1391          *    commit_root to accounting later insert_dir_item()
1392          * 2) write all btree blocks onto disk
1393          *    This is to make sure later btree modification will be cowed
1394          *    Or commit_root can be populated and cause wrong qgroup numbers
1395          * In this simplified commit, we don't really care about other trees
1396          * like chunk and root tree, as they won't affect qgroup.
1397          * And we don't write super to avoid half committed status.
1398          */
1399         ret = commit_cowonly_roots(trans);
1400         if (ret)
1401                 goto out;
1402         switch_commit_roots(trans->transaction);
1403         ret = btrfs_write_and_wait_transaction(trans);
1404         if (ret)
1405                 btrfs_handle_fs_error(fs_info, ret,
1406                         "Error while writing out transaction for qgroup");
1407
1408 out:
1409         mutex_unlock(&fs_info->tree_log_mutex);
1410
1411         /*
1412          * Force parent root to be updated, as we recorded it before so its
1413          * last_trans == cur_transid.
1414          * Or it won't be committed again onto disk after later
1415          * insert_dir_item()
1416          */
1417         if (!ret)
1418                 record_root_in_trans(trans, parent, 1);
1419         return ret;
1420 }
1421
1422 /*
1423  * new snapshots need to be created at a very specific time in the
1424  * transaction commit.  This does the actual creation.
1425  *
1426  * Note:
1427  * If the error which may affect the commitment of the current transaction
1428  * happens, we should return the error number. If the error which just affect
1429  * the creation of the pending snapshots, just return 0.
1430  */
1431 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
1432                                    struct btrfs_pending_snapshot *pending)
1433 {
1434
1435         struct btrfs_fs_info *fs_info = trans->fs_info;
1436         struct btrfs_key key;
1437         struct btrfs_root_item *new_root_item;
1438         struct btrfs_root *tree_root = fs_info->tree_root;
1439         struct btrfs_root *root = pending->root;
1440         struct btrfs_root *parent_root;
1441         struct btrfs_block_rsv *rsv;
1442         struct inode *parent_inode;
1443         struct btrfs_path *path;
1444         struct btrfs_dir_item *dir_item;
1445         struct dentry *dentry;
1446         struct extent_buffer *tmp;
1447         struct extent_buffer *old;
1448         struct timespec cur_time;
1449         int ret = 0;
1450         u64 to_reserve = 0;
1451         u64 index = 0;
1452         u64 objectid;
1453         u64 root_flags;
1454         uuid_le new_uuid;
1455
1456         ASSERT(pending->path);
1457         path = pending->path;
1458
1459         ASSERT(pending->root_item);
1460         new_root_item = pending->root_item;
1461
1462         pending->error = btrfs_find_free_objectid(tree_root, &objectid);
1463         if (pending->error)
1464                 goto no_free_objectid;
1465
1466         /*
1467          * Make qgroup to skip current new snapshot's qgroupid, as it is
1468          * accounted by later btrfs_qgroup_inherit().
1469          */
1470         btrfs_set_skip_qgroup(trans, objectid);
1471
1472         btrfs_reloc_pre_snapshot(pending, &to_reserve);
1473
1474         if (to_reserve > 0) {
1475                 pending->error = btrfs_block_rsv_add(root,
1476                                                      &pending->block_rsv,
1477                                                      to_reserve,
1478                                                      BTRFS_RESERVE_NO_FLUSH);
1479                 if (pending->error)
1480                         goto clear_skip_qgroup;
1481         }
1482
1483         key.objectid = objectid;
1484         key.offset = (u64)-1;
1485         key.type = BTRFS_ROOT_ITEM_KEY;
1486
1487         rsv = trans->block_rsv;
1488         trans->block_rsv = &pending->block_rsv;
1489         trans->bytes_reserved = trans->block_rsv->reserved;
1490         trace_btrfs_space_reservation(fs_info, "transaction",
1491                                       trans->transid,
1492                                       trans->bytes_reserved, 1);
1493         dentry = pending->dentry;
1494         parent_inode = pending->dir;
1495         parent_root = BTRFS_I(parent_inode)->root;
1496         record_root_in_trans(trans, parent_root, 0);
1497
1498         cur_time = current_time(parent_inode);
1499
1500         /*
1501          * insert the directory item
1502          */
1503         ret = btrfs_set_inode_index(BTRFS_I(parent_inode), &index);
1504         BUG_ON(ret); /* -ENOMEM */
1505
1506         /* check if there is a file/dir which has the same name. */
1507         dir_item = btrfs_lookup_dir_item(NULL, parent_root, path,
1508                                          btrfs_ino(BTRFS_I(parent_inode)),
1509                                          dentry->d_name.name,
1510                                          dentry->d_name.len, 0);
1511         if (dir_item != NULL && !IS_ERR(dir_item)) {
1512                 pending->error = -EEXIST;
1513                 goto dir_item_existed;
1514         } else if (IS_ERR(dir_item)) {
1515                 ret = PTR_ERR(dir_item);
1516                 btrfs_abort_transaction(trans, ret);
1517                 goto fail;
1518         }
1519         btrfs_release_path(path);
1520
1521         /*
1522          * pull in the delayed directory update
1523          * and the delayed inode item
1524          * otherwise we corrupt the FS during
1525          * snapshot
1526          */
1527         ret = btrfs_run_delayed_items(trans);
1528         if (ret) {      /* Transaction aborted */
1529                 btrfs_abort_transaction(trans, ret);
1530                 goto fail;
1531         }
1532
1533         record_root_in_trans(trans, root, 0);
1534         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
1535         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
1536         btrfs_check_and_init_root_item(new_root_item);
1537
1538         root_flags = btrfs_root_flags(new_root_item);
1539         if (pending->readonly)
1540                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
1541         else
1542                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
1543         btrfs_set_root_flags(new_root_item, root_flags);
1544
1545         btrfs_set_root_generation_v2(new_root_item,
1546                         trans->transid);
1547         uuid_le_gen(&new_uuid);
1548         memcpy(new_root_item->uuid, new_uuid.b, BTRFS_UUID_SIZE);
1549         memcpy(new_root_item->parent_uuid, root->root_item.uuid,
1550                         BTRFS_UUID_SIZE);
1551         if (!(root_flags & BTRFS_ROOT_SUBVOL_RDONLY)) {
1552                 memset(new_root_item->received_uuid, 0,
1553                        sizeof(new_root_item->received_uuid));
1554                 memset(&new_root_item->stime, 0, sizeof(new_root_item->stime));
1555                 memset(&new_root_item->rtime, 0, sizeof(new_root_item->rtime));
1556                 btrfs_set_root_stransid(new_root_item, 0);
1557                 btrfs_set_root_rtransid(new_root_item, 0);
1558         }
1559         btrfs_set_stack_timespec_sec(&new_root_item->otime, cur_time.tv_sec);
1560         btrfs_set_stack_timespec_nsec(&new_root_item->otime, cur_time.tv_nsec);
1561         btrfs_set_root_otransid(new_root_item, trans->transid);
1562
1563         old = btrfs_lock_root_node(root);
1564         ret = btrfs_cow_block(trans, root, old, NULL, 0, &old);
1565         if (ret) {
1566                 btrfs_tree_unlock(old);
1567                 free_extent_buffer(old);
1568                 btrfs_abort_transaction(trans, ret);
1569                 goto fail;
1570         }
1571
1572         btrfs_set_lock_blocking(old);
1573
1574         ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1575         /* clean up in any case */
1576         btrfs_tree_unlock(old);
1577         free_extent_buffer(old);
1578         if (ret) {
1579                 btrfs_abort_transaction(trans, ret);
1580                 goto fail;
1581         }
1582         /* see comments in should_cow_block() */
1583         set_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1584         smp_wmb();
1585
1586         btrfs_set_root_node(new_root_item, tmp);
1587         /* record when the snapshot was created in key.offset */
1588         key.offset = trans->transid;
1589         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1590         btrfs_tree_unlock(tmp);
1591         free_extent_buffer(tmp);
1592         if (ret) {
1593                 btrfs_abort_transaction(trans, ret);
1594                 goto fail;
1595         }
1596
1597         /*
1598          * insert root back/forward references
1599          */
1600         ret = btrfs_add_root_ref(trans, fs_info, objectid,
1601                                  parent_root->root_key.objectid,
1602                                  btrfs_ino(BTRFS_I(parent_inode)), index,
1603                                  dentry->d_name.name, dentry->d_name.len);
1604         if (ret) {
1605                 btrfs_abort_transaction(trans, ret);
1606                 goto fail;
1607         }
1608
1609         key.offset = (u64)-1;
1610         pending->snap = btrfs_read_fs_root_no_name(fs_info, &key);
1611         if (IS_ERR(pending->snap)) {
1612                 ret = PTR_ERR(pending->snap);
1613                 btrfs_abort_transaction(trans, ret);
1614                 goto fail;
1615         }
1616
1617         ret = btrfs_reloc_post_snapshot(trans, pending);
1618         if (ret) {
1619                 btrfs_abort_transaction(trans, ret);
1620                 goto fail;
1621         }
1622
1623         ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1624         if (ret) {
1625                 btrfs_abort_transaction(trans, ret);
1626                 goto fail;
1627         }
1628
1629         /*
1630          * Do special qgroup accounting for snapshot, as we do some qgroup
1631          * snapshot hack to do fast snapshot.
1632          * To co-operate with that hack, we do hack again.
1633          * Or snapshot will be greatly slowed down by a subtree qgroup rescan
1634          */
1635         ret = qgroup_account_snapshot(trans, root, parent_root,
1636                                       pending->inherit, objectid);
1637         if (ret < 0)
1638                 goto fail;
1639
1640         ret = btrfs_insert_dir_item(trans, parent_root,
1641                                     dentry->d_name.name, dentry->d_name.len,
1642                                     BTRFS_I(parent_inode), &key,
1643                                     BTRFS_FT_DIR, index);
1644         /* We have check then name at the beginning, so it is impossible. */
1645         BUG_ON(ret == -EEXIST || ret == -EOVERFLOW);
1646         if (ret) {
1647                 btrfs_abort_transaction(trans, ret);
1648                 goto fail;
1649         }
1650
1651         btrfs_i_size_write(BTRFS_I(parent_inode), parent_inode->i_size +
1652                                          dentry->d_name.len * 2);
1653         parent_inode->i_mtime = parent_inode->i_ctime =
1654                 current_time(parent_inode);
1655         ret = btrfs_update_inode_fallback(trans, parent_root, parent_inode);
1656         if (ret) {
1657                 btrfs_abort_transaction(trans, ret);
1658                 goto fail;
1659         }
1660         ret = btrfs_uuid_tree_add(trans, fs_info, new_uuid.b,
1661                                   BTRFS_UUID_KEY_SUBVOL, objectid);
1662         if (ret) {
1663                 btrfs_abort_transaction(trans, ret);
1664                 goto fail;
1665         }
1666         if (!btrfs_is_empty_uuid(new_root_item->received_uuid)) {
1667                 ret = btrfs_uuid_tree_add(trans, fs_info,
1668                                           new_root_item->received_uuid,
1669                                           BTRFS_UUID_KEY_RECEIVED_SUBVOL,
1670                                           objectid);
1671                 if (ret && ret != -EEXIST) {
1672                         btrfs_abort_transaction(trans, ret);
1673                         goto fail;
1674                 }
1675         }
1676
1677         ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1678         if (ret) {
1679                 btrfs_abort_transaction(trans, ret);
1680                 goto fail;
1681         }
1682
1683 fail:
1684         pending->error = ret;
1685 dir_item_existed:
1686         trans->block_rsv = rsv;
1687         trans->bytes_reserved = 0;
1688 clear_skip_qgroup:
1689         btrfs_clear_skip_qgroup(trans);
1690 no_free_objectid:
1691         kfree(new_root_item);
1692         pending->root_item = NULL;
1693         btrfs_free_path(path);
1694         pending->path = NULL;
1695
1696         return ret;
1697 }
1698
1699 /*
1700  * create all the snapshots we've scheduled for creation
1701  */
1702 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans)
1703 {
1704         struct btrfs_pending_snapshot *pending, *next;
1705         struct list_head *head = &trans->transaction->pending_snapshots;
1706         int ret = 0;
1707
1708         list_for_each_entry_safe(pending, next, head, list) {
1709                 list_del(&pending->list);
1710                 ret = create_pending_snapshot(trans, pending);
1711                 if (ret)
1712                         break;
1713         }
1714         return ret;
1715 }
1716
1717 static void update_super_roots(struct btrfs_fs_info *fs_info)
1718 {
1719         struct btrfs_root_item *root_item;
1720         struct btrfs_super_block *super;
1721
1722         super = fs_info->super_copy;
1723
1724         root_item = &fs_info->chunk_root->root_item;
1725         super->chunk_root = root_item->bytenr;
1726         super->chunk_root_generation = root_item->generation;
1727         super->chunk_root_level = root_item->level;
1728
1729         root_item = &fs_info->tree_root->root_item;
1730         super->root = root_item->bytenr;
1731         super->generation = root_item->generation;
1732         super->root_level = root_item->level;
1733         if (btrfs_test_opt(fs_info, SPACE_CACHE))
1734                 super->cache_generation = root_item->generation;
1735         if (test_bit(BTRFS_FS_UPDATE_UUID_TREE_GEN, &fs_info->flags))
1736                 super->uuid_tree_generation = root_item->generation;
1737 }
1738
1739 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1740 {
1741         struct btrfs_transaction *trans;
1742         int ret = 0;
1743
1744         spin_lock(&info->trans_lock);
1745         trans = info->running_transaction;
1746         if (trans)
1747                 ret = (trans->state >= TRANS_STATE_COMMIT_START);
1748         spin_unlock(&info->trans_lock);
1749         return ret;
1750 }
1751
1752 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1753 {
1754         struct btrfs_transaction *trans;
1755         int ret = 0;
1756
1757         spin_lock(&info->trans_lock);
1758         trans = info->running_transaction;
1759         if (trans)
1760                 ret = is_transaction_blocked(trans);
1761         spin_unlock(&info->trans_lock);
1762         return ret;
1763 }
1764
1765 /*
1766  * wait for the current transaction commit to start and block subsequent
1767  * transaction joins
1768  */
1769 static void wait_current_trans_commit_start(struct btrfs_fs_info *fs_info,
1770                                             struct btrfs_transaction *trans)
1771 {
1772         wait_event(fs_info->transaction_blocked_wait,
1773                    trans->state >= TRANS_STATE_COMMIT_START || trans->aborted);
1774 }
1775
1776 /*
1777  * wait for the current transaction to start and then become unblocked.
1778  * caller holds ref.
1779  */
1780 static void wait_current_trans_commit_start_and_unblock(
1781                                         struct btrfs_fs_info *fs_info,
1782                                         struct btrfs_transaction *trans)
1783 {
1784         wait_event(fs_info->transaction_wait,
1785                    trans->state >= TRANS_STATE_UNBLOCKED || trans->aborted);
1786 }
1787
1788 /*
1789  * commit transactions asynchronously. once btrfs_commit_transaction_async
1790  * returns, any subsequent transaction will not be allowed to join.
1791  */
1792 struct btrfs_async_commit {
1793         struct btrfs_trans_handle *newtrans;
1794         struct work_struct work;
1795 };
1796
1797 static void do_async_commit(struct work_struct *work)
1798 {
1799         struct btrfs_async_commit *ac =
1800                 container_of(work, struct btrfs_async_commit, work);
1801
1802         /*
1803          * We've got freeze protection passed with the transaction.
1804          * Tell lockdep about it.
1805          */
1806         if (ac->newtrans->type & __TRANS_FREEZABLE)
1807                 __sb_writers_acquired(ac->newtrans->fs_info->sb, SB_FREEZE_FS);
1808
1809         current->journal_info = ac->newtrans;
1810
1811         btrfs_commit_transaction(ac->newtrans);
1812         kfree(ac);
1813 }
1814
1815 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1816                                    int wait_for_unblock)
1817 {
1818         struct btrfs_fs_info *fs_info = trans->fs_info;
1819         struct btrfs_async_commit *ac;
1820         struct btrfs_transaction *cur_trans;
1821
1822         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1823         if (!ac)
1824                 return -ENOMEM;
1825
1826         INIT_WORK(&ac->work, do_async_commit);
1827         ac->newtrans = btrfs_join_transaction(trans->root);
1828         if (IS_ERR(ac->newtrans)) {
1829                 int err = PTR_ERR(ac->newtrans);
1830                 kfree(ac);
1831                 return err;
1832         }
1833
1834         /* take transaction reference */
1835         cur_trans = trans->transaction;
1836         refcount_inc(&cur_trans->use_count);
1837
1838         btrfs_end_transaction(trans);
1839
1840         /*
1841          * Tell lockdep we've released the freeze rwsem, since the
1842          * async commit thread will be the one to unlock it.
1843          */
1844         if (ac->newtrans->type & __TRANS_FREEZABLE)
1845                 __sb_writers_release(fs_info->sb, SB_FREEZE_FS);
1846
1847         schedule_work(&ac->work);
1848
1849         /* wait for transaction to start and unblock */
1850         if (wait_for_unblock)
1851                 wait_current_trans_commit_start_and_unblock(fs_info, cur_trans);
1852         else
1853                 wait_current_trans_commit_start(fs_info, cur_trans);
1854
1855         if (current->journal_info == trans)
1856                 current->journal_info = NULL;
1857
1858         btrfs_put_transaction(cur_trans);
1859         return 0;
1860 }
1861
1862
1863 static void cleanup_transaction(struct btrfs_trans_handle *trans, int err)
1864 {
1865         struct btrfs_fs_info *fs_info = trans->fs_info;
1866         struct btrfs_transaction *cur_trans = trans->transaction;
1867         DEFINE_WAIT(wait);
1868
1869         WARN_ON(refcount_read(&trans->use_count) > 1);
1870
1871         btrfs_abort_transaction(trans, err);
1872
1873         spin_lock(&fs_info->trans_lock);
1874
1875         /*
1876          * If the transaction is removed from the list, it means this
1877          * transaction has been committed successfully, so it is impossible
1878          * to call the cleanup function.
1879          */
1880         BUG_ON(list_empty(&cur_trans->list));
1881
1882         list_del_init(&cur_trans->list);
1883         if (cur_trans == fs_info->running_transaction) {
1884                 cur_trans->state = TRANS_STATE_COMMIT_DOING;
1885                 spin_unlock(&fs_info->trans_lock);
1886                 wait_event(cur_trans->writer_wait,
1887                            atomic_read(&cur_trans->num_writers) == 1);
1888
1889                 spin_lock(&fs_info->trans_lock);
1890         }
1891         spin_unlock(&fs_info->trans_lock);
1892
1893         btrfs_cleanup_one_transaction(trans->transaction, fs_info);
1894
1895         spin_lock(&fs_info->trans_lock);
1896         if (cur_trans == fs_info->running_transaction)
1897                 fs_info->running_transaction = NULL;
1898         spin_unlock(&fs_info->trans_lock);
1899
1900         if (trans->type & __TRANS_FREEZABLE)
1901                 sb_end_intwrite(fs_info->sb);
1902         btrfs_put_transaction(cur_trans);
1903         btrfs_put_transaction(cur_trans);
1904
1905         trace_btrfs_transaction_commit(trans->root);
1906
1907         if (current->journal_info == trans)
1908                 current->journal_info = NULL;
1909         btrfs_scrub_cancel(fs_info);
1910
1911         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1912 }
1913
1914 static inline int btrfs_start_delalloc_flush(struct btrfs_fs_info *fs_info)
1915 {
1916         /*
1917          * We use writeback_inodes_sb here because if we used
1918          * btrfs_start_delalloc_roots we would deadlock with fs freeze.
1919          * Currently are holding the fs freeze lock, if we do an async flush
1920          * we'll do btrfs_join_transaction() and deadlock because we need to
1921          * wait for the fs freeze lock.  Using the direct flushing we benefit
1922          * from already being in a transaction and our join_transaction doesn't
1923          * have to re-take the fs freeze lock.
1924          */
1925         if (btrfs_test_opt(fs_info, FLUSHONCOMMIT))
1926                 writeback_inodes_sb(fs_info->sb, WB_REASON_SYNC);
1927         return 0;
1928 }
1929
1930 static inline void btrfs_wait_delalloc_flush(struct btrfs_fs_info *fs_info)
1931 {
1932         if (btrfs_test_opt(fs_info, FLUSHONCOMMIT))
1933                 btrfs_wait_ordered_roots(fs_info, U64_MAX, 0, (u64)-1);
1934 }
1935
1936 static inline void
1937 btrfs_wait_pending_ordered(struct btrfs_transaction *cur_trans)
1938 {
1939         wait_event(cur_trans->pending_wait,
1940                    atomic_read(&cur_trans->pending_ordered) == 0);
1941 }
1942
1943 int btrfs_commit_transaction(struct btrfs_trans_handle *trans)
1944 {
1945         struct btrfs_fs_info *fs_info = trans->fs_info;
1946         struct btrfs_transaction *cur_trans = trans->transaction;
1947         struct btrfs_transaction *prev_trans = NULL;
1948         int ret;
1949
1950         /* Stop the commit early if ->aborted is set */
1951         if (unlikely(READ_ONCE(cur_trans->aborted))) {
1952                 ret = cur_trans->aborted;
1953                 btrfs_end_transaction(trans);
1954                 return ret;
1955         }
1956
1957         /* make a pass through all the delayed refs we have so far
1958          * any runnings procs may add more while we are here
1959          */
1960         ret = btrfs_run_delayed_refs(trans, fs_info, 0);
1961         if (ret) {
1962                 btrfs_end_transaction(trans);
1963                 return ret;
1964         }
1965
1966         btrfs_trans_release_metadata(trans);
1967         trans->block_rsv = NULL;
1968
1969         cur_trans = trans->transaction;
1970
1971         /*
1972          * set the flushing flag so procs in this transaction have to
1973          * start sending their work down.
1974          */
1975         cur_trans->delayed_refs.flushing = 1;
1976         smp_wmb();
1977
1978         if (!list_empty(&trans->new_bgs))
1979                 btrfs_create_pending_block_groups(trans);
1980
1981         ret = btrfs_run_delayed_refs(trans, fs_info, 0);
1982         if (ret) {
1983                 btrfs_end_transaction(trans);
1984                 return ret;
1985         }
1986
1987         if (!test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &cur_trans->flags)) {
1988                 int run_it = 0;
1989
1990                 /* this mutex is also taken before trying to set
1991                  * block groups readonly.  We need to make sure
1992                  * that nobody has set a block group readonly
1993                  * after a extents from that block group have been
1994                  * allocated for cache files.  btrfs_set_block_group_ro
1995                  * will wait for the transaction to commit if it
1996                  * finds BTRFS_TRANS_DIRTY_BG_RUN set.
1997                  *
1998                  * The BTRFS_TRANS_DIRTY_BG_RUN flag is also used to make sure
1999                  * only one process starts all the block group IO.  It wouldn't
2000                  * hurt to have more than one go through, but there's no
2001                  * real advantage to it either.
2002                  */
2003                 mutex_lock(&fs_info->ro_block_group_mutex);
2004                 if (!test_and_set_bit(BTRFS_TRANS_DIRTY_BG_RUN,
2005                                       &cur_trans->flags))
2006                         run_it = 1;
2007                 mutex_unlock(&fs_info->ro_block_group_mutex);
2008
2009                 if (run_it) {
2010                         ret = btrfs_start_dirty_block_groups(trans);
2011                         if (ret) {
2012                                 btrfs_end_transaction(trans);
2013                                 return ret;
2014                         }
2015                 }
2016         }
2017
2018         spin_lock(&fs_info->trans_lock);
2019         if (cur_trans->state >= TRANS_STATE_COMMIT_START) {
2020                 spin_unlock(&fs_info->trans_lock);
2021                 refcount_inc(&cur_trans->use_count);
2022                 ret = btrfs_end_transaction(trans);
2023
2024                 wait_for_commit(cur_trans);
2025
2026                 if (unlikely(cur_trans->aborted))
2027                         ret = cur_trans->aborted;
2028
2029                 btrfs_put_transaction(cur_trans);
2030
2031                 return ret;
2032         }
2033
2034         cur_trans->state = TRANS_STATE_COMMIT_START;
2035         wake_up(&fs_info->transaction_blocked_wait);
2036
2037         if (cur_trans->list.prev != &fs_info->trans_list) {
2038                 prev_trans = list_entry(cur_trans->list.prev,
2039                                         struct btrfs_transaction, list);
2040                 if (prev_trans->state != TRANS_STATE_COMPLETED) {
2041                         refcount_inc(&prev_trans->use_count);
2042                         spin_unlock(&fs_info->trans_lock);
2043
2044                         wait_for_commit(prev_trans);
2045                         ret = prev_trans->aborted;
2046
2047                         btrfs_put_transaction(prev_trans);
2048                         if (ret)
2049                                 goto cleanup_transaction;
2050                 } else {
2051                         spin_unlock(&fs_info->trans_lock);
2052                 }
2053         } else {
2054                 spin_unlock(&fs_info->trans_lock);
2055         }
2056
2057         extwriter_counter_dec(cur_trans, trans->type);
2058
2059         ret = btrfs_start_delalloc_flush(fs_info);
2060         if (ret)
2061                 goto cleanup_transaction;
2062
2063         ret = btrfs_run_delayed_items(trans);
2064         if (ret)
2065                 goto cleanup_transaction;
2066
2067         wait_event(cur_trans->writer_wait,
2068                    extwriter_counter_read(cur_trans) == 0);
2069
2070         /* some pending stuffs might be added after the previous flush. */
2071         ret = btrfs_run_delayed_items(trans);
2072         if (ret)
2073                 goto cleanup_transaction;
2074
2075         btrfs_wait_delalloc_flush(fs_info);
2076
2077         btrfs_wait_pending_ordered(cur_trans);
2078
2079         btrfs_scrub_pause(fs_info);
2080         /*
2081          * Ok now we need to make sure to block out any other joins while we
2082          * commit the transaction.  We could have started a join before setting
2083          * COMMIT_DOING so make sure to wait for num_writers to == 1 again.
2084          */
2085         spin_lock(&fs_info->trans_lock);
2086         cur_trans->state = TRANS_STATE_COMMIT_DOING;
2087         spin_unlock(&fs_info->trans_lock);
2088         wait_event(cur_trans->writer_wait,
2089                    atomic_read(&cur_trans->num_writers) == 1);
2090
2091         /* ->aborted might be set after the previous check, so check it */
2092         if (unlikely(READ_ONCE(cur_trans->aborted))) {
2093                 ret = cur_trans->aborted;
2094                 goto scrub_continue;
2095         }
2096         /*
2097          * the reloc mutex makes sure that we stop
2098          * the balancing code from coming in and moving
2099          * extents around in the middle of the commit
2100          */
2101         mutex_lock(&fs_info->reloc_mutex);
2102
2103         /*
2104          * We needn't worry about the delayed items because we will
2105          * deal with them in create_pending_snapshot(), which is the
2106          * core function of the snapshot creation.
2107          */
2108         ret = create_pending_snapshots(trans);
2109         if (ret) {
2110                 mutex_unlock(&fs_info->reloc_mutex);
2111                 goto scrub_continue;
2112         }
2113
2114         /*
2115          * We insert the dir indexes of the snapshots and update the inode
2116          * of the snapshots' parents after the snapshot creation, so there
2117          * are some delayed items which are not dealt with. Now deal with
2118          * them.
2119          *
2120          * We needn't worry that this operation will corrupt the snapshots,
2121          * because all the tree which are snapshoted will be forced to COW
2122          * the nodes and leaves.
2123          */
2124         ret = btrfs_run_delayed_items(trans);
2125         if (ret) {
2126                 mutex_unlock(&fs_info->reloc_mutex);
2127                 goto scrub_continue;
2128         }
2129
2130         ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
2131         if (ret) {
2132                 mutex_unlock(&fs_info->reloc_mutex);
2133                 goto scrub_continue;
2134         }
2135
2136         /*
2137          * make sure none of the code above managed to slip in a
2138          * delayed item
2139          */
2140         btrfs_assert_delayed_root_empty(fs_info);
2141
2142         WARN_ON(cur_trans != trans->transaction);
2143
2144         /* btrfs_commit_tree_roots is responsible for getting the
2145          * various roots consistent with each other.  Every pointer
2146          * in the tree of tree roots has to point to the most up to date
2147          * root for every subvolume and other tree.  So, we have to keep
2148          * the tree logging code from jumping in and changing any
2149          * of the trees.
2150          *
2151          * At this point in the commit, there can't be any tree-log
2152          * writers, but a little lower down we drop the trans mutex
2153          * and let new people in.  By holding the tree_log_mutex
2154          * from now until after the super is written, we avoid races
2155          * with the tree-log code.
2156          */
2157         mutex_lock(&fs_info->tree_log_mutex);
2158
2159         ret = commit_fs_roots(trans);
2160         if (ret) {
2161                 mutex_unlock(&fs_info->tree_log_mutex);
2162                 mutex_unlock(&fs_info->reloc_mutex);
2163                 goto scrub_continue;
2164         }
2165
2166         /*
2167          * Since the transaction is done, we can apply the pending changes
2168          * before the next transaction.
2169          */
2170         btrfs_apply_pending_changes(fs_info);
2171
2172         /* commit_fs_roots gets rid of all the tree log roots, it is now
2173          * safe to free the root of tree log roots
2174          */
2175         btrfs_free_log_root_tree(trans, fs_info);
2176
2177         /*
2178          * commit_fs_roots() can call btrfs_save_ino_cache(), which generates
2179          * new delayed refs. Must handle them or qgroup can be wrong.
2180          */
2181         ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
2182         if (ret) {
2183                 mutex_unlock(&fs_info->tree_log_mutex);
2184                 mutex_unlock(&fs_info->reloc_mutex);
2185                 goto scrub_continue;
2186         }
2187
2188         /*
2189          * Since fs roots are all committed, we can get a quite accurate
2190          * new_roots. So let's do quota accounting.
2191          */
2192         ret = btrfs_qgroup_account_extents(trans, fs_info);
2193         if (ret < 0) {
2194                 mutex_unlock(&fs_info->tree_log_mutex);
2195                 mutex_unlock(&fs_info->reloc_mutex);
2196                 goto scrub_continue;
2197         }
2198
2199         ret = commit_cowonly_roots(trans);
2200         if (ret) {
2201                 mutex_unlock(&fs_info->tree_log_mutex);
2202                 mutex_unlock(&fs_info->reloc_mutex);
2203                 goto scrub_continue;
2204         }
2205
2206         /*
2207          * The tasks which save the space cache and inode cache may also
2208          * update ->aborted, check it.
2209          */
2210         if (unlikely(READ_ONCE(cur_trans->aborted))) {
2211                 ret = cur_trans->aborted;
2212                 mutex_unlock(&fs_info->tree_log_mutex);
2213                 mutex_unlock(&fs_info->reloc_mutex);
2214                 goto scrub_continue;
2215         }
2216
2217         btrfs_prepare_extent_commit(fs_info);
2218
2219         cur_trans = fs_info->running_transaction;
2220
2221         btrfs_set_root_node(&fs_info->tree_root->root_item,
2222                             fs_info->tree_root->node);
2223         list_add_tail(&fs_info->tree_root->dirty_list,
2224                       &cur_trans->switch_commits);
2225
2226         btrfs_set_root_node(&fs_info->chunk_root->root_item,
2227                             fs_info->chunk_root->node);
2228         list_add_tail(&fs_info->chunk_root->dirty_list,
2229                       &cur_trans->switch_commits);
2230
2231         switch_commit_roots(cur_trans);
2232
2233         ASSERT(list_empty(&cur_trans->dirty_bgs));
2234         ASSERT(list_empty(&cur_trans->io_bgs));
2235         update_super_roots(fs_info);
2236
2237         btrfs_set_super_log_root(fs_info->super_copy, 0);
2238         btrfs_set_super_log_root_level(fs_info->super_copy, 0);
2239         memcpy(fs_info->super_for_commit, fs_info->super_copy,
2240                sizeof(*fs_info->super_copy));
2241
2242         btrfs_update_commit_device_size(fs_info);
2243         btrfs_update_commit_device_bytes_used(cur_trans);
2244
2245         clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags);
2246         clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags);
2247
2248         btrfs_trans_release_chunk_metadata(trans);
2249
2250         spin_lock(&fs_info->trans_lock);
2251         cur_trans->state = TRANS_STATE_UNBLOCKED;
2252         fs_info->running_transaction = NULL;
2253         spin_unlock(&fs_info->trans_lock);
2254         mutex_unlock(&fs_info->reloc_mutex);
2255
2256         wake_up(&fs_info->transaction_wait);
2257
2258         ret = btrfs_write_and_wait_transaction(trans);
2259         if (ret) {
2260                 btrfs_handle_fs_error(fs_info, ret,
2261                                       "Error while writing out transaction");
2262                 mutex_unlock(&fs_info->tree_log_mutex);
2263                 goto scrub_continue;
2264         }
2265
2266         ret = write_all_supers(fs_info, 0);
2267         /*
2268          * the super is written, we can safely allow the tree-loggers
2269          * to go about their business
2270          */
2271         mutex_unlock(&fs_info->tree_log_mutex);
2272         if (ret)
2273                 goto scrub_continue;
2274
2275         btrfs_finish_extent_commit(trans, fs_info);
2276
2277         if (test_bit(BTRFS_TRANS_HAVE_FREE_BGS, &cur_trans->flags))
2278                 btrfs_clear_space_info_full(fs_info);
2279
2280         fs_info->last_trans_committed = cur_trans->transid;
2281         /*
2282          * We needn't acquire the lock here because there is no other task
2283          * which can change it.
2284          */
2285         cur_trans->state = TRANS_STATE_COMPLETED;
2286         wake_up(&cur_trans->commit_wait);
2287
2288         spin_lock(&fs_info->trans_lock);
2289         list_del_init(&cur_trans->list);
2290         spin_unlock(&fs_info->trans_lock);
2291
2292         btrfs_put_transaction(cur_trans);
2293         btrfs_put_transaction(cur_trans);
2294
2295         if (trans->type & __TRANS_FREEZABLE)
2296                 sb_end_intwrite(fs_info->sb);
2297
2298         trace_btrfs_transaction_commit(trans->root);
2299
2300         btrfs_scrub_continue(fs_info);
2301
2302         if (current->journal_info == trans)
2303                 current->journal_info = NULL;
2304
2305         kmem_cache_free(btrfs_trans_handle_cachep, trans);
2306
2307         /*
2308          * If fs has been frozen, we can not handle delayed iputs, otherwise
2309          * it'll result in deadlock about SB_FREEZE_FS.
2310          */
2311         if (current != fs_info->transaction_kthread &&
2312             current != fs_info->cleaner_kthread &&
2313             !test_bit(BTRFS_FS_FROZEN, &fs_info->flags))
2314                 btrfs_run_delayed_iputs(fs_info);
2315
2316         return ret;
2317
2318 scrub_continue:
2319         btrfs_scrub_continue(fs_info);
2320 cleanup_transaction:
2321         btrfs_trans_release_metadata(trans);
2322         btrfs_trans_release_chunk_metadata(trans);
2323         trans->block_rsv = NULL;
2324         btrfs_warn(fs_info, "Skipping commit of aborted transaction.");
2325         if (current->journal_info == trans)
2326                 current->journal_info = NULL;
2327         cleanup_transaction(trans, ret);
2328
2329         return ret;
2330 }
2331
2332 /*
2333  * return < 0 if error
2334  * 0 if there are no more dead_roots at the time of call
2335  * 1 there are more to be processed, call me again
2336  *
2337  * The return value indicates there are certainly more snapshots to delete, but
2338  * if there comes a new one during processing, it may return 0. We don't mind,
2339  * because btrfs_commit_super will poke cleaner thread and it will process it a
2340  * few seconds later.
2341  */
2342 int btrfs_clean_one_deleted_snapshot(struct btrfs_root *root)
2343 {
2344         int ret;
2345         struct btrfs_fs_info *fs_info = root->fs_info;
2346
2347         spin_lock(&fs_info->trans_lock);
2348         if (list_empty(&fs_info->dead_roots)) {
2349                 spin_unlock(&fs_info->trans_lock);
2350                 return 0;
2351         }
2352         root = list_first_entry(&fs_info->dead_roots,
2353                         struct btrfs_root, root_list);
2354         list_del_init(&root->root_list);
2355         spin_unlock(&fs_info->trans_lock);
2356
2357         btrfs_debug(fs_info, "cleaner removing %llu", root->objectid);
2358
2359         btrfs_kill_all_delayed_nodes(root);
2360
2361         if (btrfs_header_backref_rev(root->node) <
2362                         BTRFS_MIXED_BACKREF_REV)
2363                 ret = btrfs_drop_snapshot(root, NULL, 0, 0);
2364         else
2365                 ret = btrfs_drop_snapshot(root, NULL, 1, 0);
2366
2367         return (ret < 0) ? 0 : 1;
2368 }
2369
2370 void btrfs_apply_pending_changes(struct btrfs_fs_info *fs_info)
2371 {
2372         unsigned long prev;
2373         unsigned long bit;
2374
2375         prev = xchg(&fs_info->pending_changes, 0);
2376         if (!prev)
2377                 return;
2378
2379         bit = 1 << BTRFS_PENDING_SET_INODE_MAP_CACHE;
2380         if (prev & bit)
2381                 btrfs_set_opt(fs_info->mount_opt, INODE_MAP_CACHE);
2382         prev &= ~bit;
2383
2384         bit = 1 << BTRFS_PENDING_CLEAR_INODE_MAP_CACHE;
2385         if (prev & bit)
2386                 btrfs_clear_opt(fs_info->mount_opt, INODE_MAP_CACHE);
2387         prev &= ~bit;
2388
2389         bit = 1 << BTRFS_PENDING_COMMIT;
2390         if (prev & bit)
2391                 btrfs_debug(fs_info, "pending commit done");
2392         prev &= ~bit;
2393
2394         if (prev)
2395                 btrfs_warn(fs_info,
2396                         "unknown pending changes left 0x%lx, ignoring", prev);
2397 }