7c815885ac01f9cbdfab5cef85db065d33ced655
[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                 return 1;
448
449         return 0;
450 }
451
452 static inline bool need_reserve_reloc_root(struct btrfs_root *root)
453 {
454         struct btrfs_fs_info *fs_info = root->fs_info;
455
456         if (!fs_info->reloc_ctl ||
457             !test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
458             root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
459             root->reloc_root)
460                 return false;
461
462         return true;
463 }
464
465 static struct btrfs_trans_handle *
466 start_transaction(struct btrfs_root *root, unsigned int num_items,
467                   unsigned int type, enum btrfs_reserve_flush_enum flush,
468                   bool enforce_qgroups)
469 {
470         struct btrfs_fs_info *fs_info = root->fs_info;
471
472         struct btrfs_trans_handle *h;
473         struct btrfs_transaction *cur_trans;
474         u64 num_bytes = 0;
475         u64 qgroup_reserved = 0;
476         bool reloc_reserved = false;
477         int ret;
478
479         /* Send isn't supposed to start transactions. */
480         ASSERT(current->journal_info != BTRFS_SEND_TRANS_STUB);
481
482         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
483                 return ERR_PTR(-EROFS);
484
485         if (current->journal_info) {
486                 WARN_ON(type & TRANS_EXTWRITERS);
487                 h = current->journal_info;
488                 refcount_inc(&h->use_count);
489                 WARN_ON(refcount_read(&h->use_count) > 2);
490                 h->orig_rsv = h->block_rsv;
491                 h->block_rsv = NULL;
492                 goto got_it;
493         }
494
495         /*
496          * Do the reservation before we join the transaction so we can do all
497          * the appropriate flushing if need be.
498          */
499         if (num_items && root != fs_info->chunk_root) {
500                 qgroup_reserved = num_items * fs_info->nodesize;
501                 ret = btrfs_qgroup_reserve_meta(root, qgroup_reserved,
502                                                 enforce_qgroups);
503                 if (ret)
504                         return ERR_PTR(ret);
505
506                 num_bytes = btrfs_calc_trans_metadata_size(fs_info, num_items);
507                 /*
508                  * Do the reservation for the relocation root creation
509                  */
510                 if (need_reserve_reloc_root(root)) {
511                         num_bytes += fs_info->nodesize;
512                         reloc_reserved = true;
513                 }
514
515                 ret = btrfs_block_rsv_add(root, &fs_info->trans_block_rsv,
516                                           num_bytes, flush);
517                 if (ret)
518                         goto reserve_fail;
519         }
520 again:
521         h = kmem_cache_zalloc(btrfs_trans_handle_cachep, GFP_NOFS);
522         if (!h) {
523                 ret = -ENOMEM;
524                 goto alloc_fail;
525         }
526
527         /*
528          * If we are JOIN_NOLOCK we're already committing a transaction and
529          * waiting on this guy, so we don't need to do the sb_start_intwrite
530          * because we're already holding a ref.  We need this because we could
531          * have raced in and did an fsync() on a file which can kick a commit
532          * and then we deadlock with somebody doing a freeze.
533          *
534          * If we are ATTACH, it means we just want to catch the current
535          * transaction and commit it, so we needn't do sb_start_intwrite(). 
536          */
537         if (type & __TRANS_FREEZABLE)
538                 sb_start_intwrite(fs_info->sb);
539
540         if (may_wait_transaction(fs_info, type))
541                 wait_current_trans(fs_info);
542
543         do {
544                 ret = join_transaction(fs_info, type);
545                 if (ret == -EBUSY) {
546                         wait_current_trans(fs_info);
547                         if (unlikely(type == TRANS_ATTACH))
548                                 ret = -ENOENT;
549                 }
550         } while (ret == -EBUSY);
551
552         if (ret < 0)
553                 goto join_fail;
554
555         cur_trans = fs_info->running_transaction;
556
557         h->transid = cur_trans->transid;
558         h->transaction = cur_trans;
559         h->root = root;
560         refcount_set(&h->use_count, 1);
561         h->fs_info = root->fs_info;
562
563         h->type = type;
564         h->can_flush_pending_bgs = true;
565         INIT_LIST_HEAD(&h->new_bgs);
566
567         smp_mb();
568         if (cur_trans->state >= TRANS_STATE_BLOCKED &&
569             may_wait_transaction(fs_info, type)) {
570                 current->journal_info = h;
571                 btrfs_commit_transaction(h);
572                 goto again;
573         }
574
575         if (num_bytes) {
576                 trace_btrfs_space_reservation(fs_info, "transaction",
577                                               h->transid, num_bytes, 1);
578                 h->block_rsv = &fs_info->trans_block_rsv;
579                 h->bytes_reserved = num_bytes;
580                 h->reloc_reserved = reloc_reserved;
581         }
582
583 got_it:
584         btrfs_record_root_in_trans(h, root);
585
586         if (!current->journal_info)
587                 current->journal_info = h;
588         return h;
589
590 join_fail:
591         if (type & __TRANS_FREEZABLE)
592                 sb_end_intwrite(fs_info->sb);
593         kmem_cache_free(btrfs_trans_handle_cachep, h);
594 alloc_fail:
595         if (num_bytes)
596                 btrfs_block_rsv_release(fs_info, &fs_info->trans_block_rsv,
597                                         num_bytes);
598 reserve_fail:
599         btrfs_qgroup_free_meta(root, qgroup_reserved);
600         return ERR_PTR(ret);
601 }
602
603 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
604                                                    unsigned int num_items)
605 {
606         return start_transaction(root, num_items, TRANS_START,
607                                  BTRFS_RESERVE_FLUSH_ALL, true);
608 }
609
610 struct btrfs_trans_handle *btrfs_start_transaction_fallback_global_rsv(
611                                         struct btrfs_root *root,
612                                         unsigned int num_items,
613                                         int min_factor)
614 {
615         struct btrfs_fs_info *fs_info = root->fs_info;
616         struct btrfs_trans_handle *trans;
617         u64 num_bytes;
618         int ret;
619
620         /*
621          * We have two callers: unlink and block group removal.  The
622          * former should succeed even if we will temporarily exceed
623          * quota and the latter operates on the extent root so
624          * qgroup enforcement is ignored anyway.
625          */
626         trans = start_transaction(root, num_items, TRANS_START,
627                                   BTRFS_RESERVE_FLUSH_ALL, false);
628         if (!IS_ERR(trans) || PTR_ERR(trans) != -ENOSPC)
629                 return trans;
630
631         trans = btrfs_start_transaction(root, 0);
632         if (IS_ERR(trans))
633                 return trans;
634
635         num_bytes = btrfs_calc_trans_metadata_size(fs_info, num_items);
636         ret = btrfs_cond_migrate_bytes(fs_info, &fs_info->trans_block_rsv,
637                                        num_bytes, min_factor);
638         if (ret) {
639                 btrfs_end_transaction(trans);
640                 return ERR_PTR(ret);
641         }
642
643         trans->block_rsv = &fs_info->trans_block_rsv;
644         trans->bytes_reserved = num_bytes;
645         trace_btrfs_space_reservation(fs_info, "transaction",
646                                       trans->transid, num_bytes, 1);
647
648         return trans;
649 }
650
651 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
652 {
653         return start_transaction(root, 0, TRANS_JOIN, BTRFS_RESERVE_NO_FLUSH,
654                                  true);
655 }
656
657 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root)
658 {
659         return start_transaction(root, 0, TRANS_JOIN_NOLOCK,
660                                  BTRFS_RESERVE_NO_FLUSH, true);
661 }
662
663 /*
664  * btrfs_attach_transaction() - catch the running transaction
665  *
666  * It is used when we want to commit the current the transaction, but
667  * don't want to start a new one.
668  *
669  * Note: If this function return -ENOENT, it just means there is no
670  * running transaction. But it is possible that the inactive transaction
671  * is still in the memory, not fully on disk. If you hope there is no
672  * inactive transaction in the fs when -ENOENT is returned, you should
673  * invoke
674  *     btrfs_attach_transaction_barrier()
675  */
676 struct btrfs_trans_handle *btrfs_attach_transaction(struct btrfs_root *root)
677 {
678         return start_transaction(root, 0, TRANS_ATTACH,
679                                  BTRFS_RESERVE_NO_FLUSH, true);
680 }
681
682 /*
683  * btrfs_attach_transaction_barrier() - catch the running transaction
684  *
685  * It is similar to the above function, the differentia is this one
686  * will wait for all the inactive transactions until they fully
687  * complete.
688  */
689 struct btrfs_trans_handle *
690 btrfs_attach_transaction_barrier(struct btrfs_root *root)
691 {
692         struct btrfs_trans_handle *trans;
693
694         trans = start_transaction(root, 0, TRANS_ATTACH,
695                                   BTRFS_RESERVE_NO_FLUSH, true);
696         if (IS_ERR(trans) && PTR_ERR(trans) == -ENOENT)
697                 btrfs_wait_for_commit(root->fs_info, 0);
698
699         return trans;
700 }
701
702 /* wait for a transaction commit to be fully complete */
703 static noinline void wait_for_commit(struct btrfs_transaction *commit)
704 {
705         wait_event(commit->commit_wait, commit->state == TRANS_STATE_COMPLETED);
706 }
707
708 int btrfs_wait_for_commit(struct btrfs_fs_info *fs_info, u64 transid)
709 {
710         struct btrfs_transaction *cur_trans = NULL, *t;
711         int ret = 0;
712
713         if (transid) {
714                 if (transid <= fs_info->last_trans_committed)
715                         goto out;
716
717                 /* find specified transaction */
718                 spin_lock(&fs_info->trans_lock);
719                 list_for_each_entry(t, &fs_info->trans_list, list) {
720                         if (t->transid == transid) {
721                                 cur_trans = t;
722                                 refcount_inc(&cur_trans->use_count);
723                                 ret = 0;
724                                 break;
725                         }
726                         if (t->transid > transid) {
727                                 ret = 0;
728                                 break;
729                         }
730                 }
731                 spin_unlock(&fs_info->trans_lock);
732
733                 /*
734                  * The specified transaction doesn't exist, or we
735                  * raced with btrfs_commit_transaction
736                  */
737                 if (!cur_trans) {
738                         if (transid > fs_info->last_trans_committed)
739                                 ret = -EINVAL;
740                         goto out;
741                 }
742         } else {
743                 /* find newest transaction that is committing | committed */
744                 spin_lock(&fs_info->trans_lock);
745                 list_for_each_entry_reverse(t, &fs_info->trans_list,
746                                             list) {
747                         if (t->state >= TRANS_STATE_COMMIT_START) {
748                                 if (t->state == TRANS_STATE_COMPLETED)
749                                         break;
750                                 cur_trans = t;
751                                 refcount_inc(&cur_trans->use_count);
752                                 break;
753                         }
754                 }
755                 spin_unlock(&fs_info->trans_lock);
756                 if (!cur_trans)
757                         goto out;  /* nothing committing|committed */
758         }
759
760         wait_for_commit(cur_trans);
761         btrfs_put_transaction(cur_trans);
762 out:
763         return ret;
764 }
765
766 void btrfs_throttle(struct btrfs_fs_info *fs_info)
767 {
768         wait_current_trans(fs_info);
769 }
770
771 static int should_end_transaction(struct btrfs_trans_handle *trans)
772 {
773         struct btrfs_fs_info *fs_info = trans->fs_info;
774
775         if (btrfs_check_space_for_delayed_refs(trans, fs_info))
776                 return 1;
777
778         return !!btrfs_block_rsv_check(&fs_info->global_block_rsv, 5);
779 }
780
781 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans)
782 {
783         struct btrfs_transaction *cur_trans = trans->transaction;
784         int updates;
785         int err;
786
787         smp_mb();
788         if (cur_trans->state >= TRANS_STATE_BLOCKED ||
789             cur_trans->delayed_refs.flushing)
790                 return 1;
791
792         updates = trans->delayed_ref_updates;
793         trans->delayed_ref_updates = 0;
794         if (updates) {
795                 err = btrfs_run_delayed_refs(trans, updates * 2);
796                 if (err) /* Error code will also eval true */
797                         return err;
798         }
799
800         return should_end_transaction(trans);
801 }
802
803 static void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans)
804
805 {
806         struct btrfs_fs_info *fs_info = trans->fs_info;
807
808         if (!trans->block_rsv) {
809                 ASSERT(!trans->bytes_reserved);
810                 return;
811         }
812
813         if (!trans->bytes_reserved)
814                 return;
815
816         ASSERT(trans->block_rsv == &fs_info->trans_block_rsv);
817         trace_btrfs_space_reservation(fs_info, "transaction",
818                                       trans->transid, trans->bytes_reserved, 0);
819         btrfs_block_rsv_release(fs_info, trans->block_rsv,
820                                 trans->bytes_reserved);
821         trans->bytes_reserved = 0;
822 }
823
824 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
825                                    int throttle)
826 {
827         struct btrfs_fs_info *info = trans->fs_info;
828         struct btrfs_transaction *cur_trans = trans->transaction;
829         u64 transid = trans->transid;
830         unsigned long cur = trans->delayed_ref_updates;
831         int lock = (trans->type != TRANS_JOIN_NOLOCK);
832         int err = 0;
833         int must_run_delayed_refs = 0;
834
835         if (refcount_read(&trans->use_count) > 1) {
836                 refcount_dec(&trans->use_count);
837                 trans->block_rsv = trans->orig_rsv;
838                 return 0;
839         }
840
841         btrfs_trans_release_metadata(trans);
842         trans->block_rsv = NULL;
843
844         if (!list_empty(&trans->new_bgs))
845                 btrfs_create_pending_block_groups(trans);
846
847         trans->delayed_ref_updates = 0;
848         if (!trans->sync) {
849                 must_run_delayed_refs =
850                         btrfs_should_throttle_delayed_refs(trans, info);
851                 cur = max_t(unsigned long, cur, 32);
852
853                 /*
854                  * don't make the caller wait if they are from a NOLOCK
855                  * or ATTACH transaction, it will deadlock with commit
856                  */
857                 if (must_run_delayed_refs == 1 &&
858                     (trans->type & (__TRANS_JOIN_NOLOCK | __TRANS_ATTACH)))
859                         must_run_delayed_refs = 2;
860         }
861
862         btrfs_trans_release_metadata(trans);
863         trans->block_rsv = NULL;
864
865         if (!list_empty(&trans->new_bgs))
866                 btrfs_create_pending_block_groups(trans);
867
868         btrfs_trans_release_chunk_metadata(trans);
869
870         if (lock && should_end_transaction(trans) &&
871             READ_ONCE(cur_trans->state) == TRANS_STATE_RUNNING) {
872                 spin_lock(&info->trans_lock);
873                 if (cur_trans->state == TRANS_STATE_RUNNING)
874                         cur_trans->state = TRANS_STATE_BLOCKED;
875                 spin_unlock(&info->trans_lock);
876         }
877
878         if (lock && READ_ONCE(cur_trans->state) == TRANS_STATE_BLOCKED) {
879                 if (throttle)
880                         return btrfs_commit_transaction(trans);
881                 else
882                         wake_up_process(info->transaction_kthread);
883         }
884
885         if (trans->type & __TRANS_FREEZABLE)
886                 sb_end_intwrite(info->sb);
887
888         WARN_ON(cur_trans != info->running_transaction);
889         WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
890         atomic_dec(&cur_trans->num_writers);
891         extwriter_counter_dec(cur_trans, trans->type);
892
893         /*
894          * Make sure counter is updated before we wake up waiters.
895          */
896         smp_mb();
897         if (waitqueue_active(&cur_trans->writer_wait))
898                 wake_up(&cur_trans->writer_wait);
899         btrfs_put_transaction(cur_trans);
900
901         if (current->journal_info == trans)
902                 current->journal_info = NULL;
903
904         if (throttle)
905                 btrfs_run_delayed_iputs(info);
906
907         if (trans->aborted ||
908             test_bit(BTRFS_FS_STATE_ERROR, &info->fs_state)) {
909                 wake_up_process(info->transaction_kthread);
910                 err = -EIO;
911         }
912
913         kmem_cache_free(btrfs_trans_handle_cachep, trans);
914         if (must_run_delayed_refs) {
915                 btrfs_async_run_delayed_refs(info, cur, transid,
916                                              must_run_delayed_refs == 1);
917         }
918         return err;
919 }
920
921 int btrfs_end_transaction(struct btrfs_trans_handle *trans)
922 {
923         return __btrfs_end_transaction(trans, 0);
924 }
925
926 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans)
927 {
928         return __btrfs_end_transaction(trans, 1);
929 }
930
931 /*
932  * when btree blocks are allocated, they have some corresponding bits set for
933  * them in one of two extent_io trees.  This is used to make sure all of
934  * those extents are sent to disk but does not wait on them
935  */
936 int btrfs_write_marked_extents(struct btrfs_fs_info *fs_info,
937                                struct extent_io_tree *dirty_pages, int mark)
938 {
939         int err = 0;
940         int werr = 0;
941         struct address_space *mapping = fs_info->btree_inode->i_mapping;
942         struct extent_state *cached_state = NULL;
943         u64 start = 0;
944         u64 end;
945
946         atomic_inc(&BTRFS_I(fs_info->btree_inode)->sync_writers);
947         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
948                                       mark, &cached_state)) {
949                 bool wait_writeback = false;
950
951                 err = convert_extent_bit(dirty_pages, start, end,
952                                          EXTENT_NEED_WAIT,
953                                          mark, &cached_state);
954                 /*
955                  * convert_extent_bit can return -ENOMEM, which is most of the
956                  * time a temporary error. So when it happens, ignore the error
957                  * and wait for writeback of this range to finish - because we
958                  * failed to set the bit EXTENT_NEED_WAIT for the range, a call
959                  * to __btrfs_wait_marked_extents() would not know that
960                  * writeback for this range started and therefore wouldn't
961                  * wait for it to finish - we don't want to commit a
962                  * superblock that points to btree nodes/leafs for which
963                  * writeback hasn't finished yet (and without errors).
964                  * We cleanup any entries left in the io tree when committing
965                  * the transaction (through clear_btree_io_tree()).
966                  */
967                 if (err == -ENOMEM) {
968                         err = 0;
969                         wait_writeback = true;
970                 }
971                 if (!err)
972                         err = filemap_fdatawrite_range(mapping, start, end);
973                 if (err)
974                         werr = err;
975                 else if (wait_writeback)
976                         werr = filemap_fdatawait_range(mapping, start, end);
977                 free_extent_state(cached_state);
978                 cached_state = NULL;
979                 cond_resched();
980                 start = end + 1;
981         }
982         atomic_dec(&BTRFS_I(fs_info->btree_inode)->sync_writers);
983         return werr;
984 }
985
986 /*
987  * when btree blocks are allocated, they have some corresponding bits set for
988  * them in one of two extent_io trees.  This is used to make sure all of
989  * those extents are on disk for transaction or log commit.  We wait
990  * on all the pages and clear them from the dirty pages state tree
991  */
992 static int __btrfs_wait_marked_extents(struct btrfs_fs_info *fs_info,
993                                        struct extent_io_tree *dirty_pages)
994 {
995         int err = 0;
996         int werr = 0;
997         struct address_space *mapping = fs_info->btree_inode->i_mapping;
998         struct extent_state *cached_state = NULL;
999         u64 start = 0;
1000         u64 end;
1001
1002         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
1003                                       EXTENT_NEED_WAIT, &cached_state)) {
1004                 /*
1005                  * Ignore -ENOMEM errors returned by clear_extent_bit().
1006                  * When committing the transaction, we'll remove any entries
1007                  * left in the io tree. For a log commit, we don't remove them
1008                  * after committing the log because the tree can be accessed
1009                  * concurrently - we do it only at transaction commit time when
1010                  * it's safe to do it (through clear_btree_io_tree()).
1011                  */
1012                 err = clear_extent_bit(dirty_pages, start, end,
1013                                        EXTENT_NEED_WAIT, 0, 0, &cached_state);
1014                 if (err == -ENOMEM)
1015                         err = 0;
1016                 if (!err)
1017                         err = filemap_fdatawait_range(mapping, start, end);
1018                 if (err)
1019                         werr = err;
1020                 free_extent_state(cached_state);
1021                 cached_state = NULL;
1022                 cond_resched();
1023                 start = end + 1;
1024         }
1025         if (err)
1026                 werr = err;
1027         return werr;
1028 }
1029
1030 int btrfs_wait_extents(struct btrfs_fs_info *fs_info,
1031                        struct extent_io_tree *dirty_pages)
1032 {
1033         bool errors = false;
1034         int err;
1035
1036         err = __btrfs_wait_marked_extents(fs_info, dirty_pages);
1037         if (test_and_clear_bit(BTRFS_FS_BTREE_ERR, &fs_info->flags))
1038                 errors = true;
1039
1040         if (errors && !err)
1041                 err = -EIO;
1042         return err;
1043 }
1044
1045 int btrfs_wait_tree_log_extents(struct btrfs_root *log_root, int mark)
1046 {
1047         struct btrfs_fs_info *fs_info = log_root->fs_info;
1048         struct extent_io_tree *dirty_pages = &log_root->dirty_log_pages;
1049         bool errors = false;
1050         int err;
1051
1052         ASSERT(log_root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
1053
1054         err = __btrfs_wait_marked_extents(fs_info, dirty_pages);
1055         if ((mark & EXTENT_DIRTY) &&
1056             test_and_clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags))
1057                 errors = true;
1058
1059         if ((mark & EXTENT_NEW) &&
1060             test_and_clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags))
1061                 errors = true;
1062
1063         if (errors && !err)
1064                 err = -EIO;
1065         return err;
1066 }
1067
1068 /*
1069  * When btree blocks are allocated the corresponding extents are marked dirty.
1070  * This function ensures such extents are persisted on disk for transaction or
1071  * log commit.
1072  *
1073  * @trans: transaction whose dirty pages we'd like to write
1074  */
1075 static int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans)
1076 {
1077         int ret;
1078         int ret2;
1079         struct extent_io_tree *dirty_pages = &trans->transaction->dirty_pages;
1080         struct btrfs_fs_info *fs_info = trans->fs_info;
1081         struct blk_plug plug;
1082
1083         blk_start_plug(&plug);
1084         ret = btrfs_write_marked_extents(fs_info, dirty_pages, EXTENT_DIRTY);
1085         blk_finish_plug(&plug);
1086         ret2 = btrfs_wait_extents(fs_info, dirty_pages);
1087
1088         clear_btree_io_tree(&trans->transaction->dirty_pages);
1089
1090         if (ret)
1091                 return ret;
1092         else if (ret2)
1093                 return ret2;
1094         else
1095                 return 0;
1096 }
1097
1098 /*
1099  * this is used to update the root pointer in the tree of tree roots.
1100  *
1101  * But, in the case of the extent allocation tree, updating the root
1102  * pointer may allocate blocks which may change the root of the extent
1103  * allocation tree.
1104  *
1105  * So, this loops and repeats and makes sure the cowonly root didn't
1106  * change while the root pointer was being updated in the metadata.
1107  */
1108 static int update_cowonly_root(struct btrfs_trans_handle *trans,
1109                                struct btrfs_root *root)
1110 {
1111         int ret;
1112         u64 old_root_bytenr;
1113         u64 old_root_used;
1114         struct btrfs_fs_info *fs_info = root->fs_info;
1115         struct btrfs_root *tree_root = fs_info->tree_root;
1116
1117         old_root_used = btrfs_root_used(&root->root_item);
1118
1119         while (1) {
1120                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
1121                 if (old_root_bytenr == root->node->start &&
1122                     old_root_used == btrfs_root_used(&root->root_item))
1123                         break;
1124
1125                 btrfs_set_root_node(&root->root_item, root->node);
1126                 ret = btrfs_update_root(trans, tree_root,
1127                                         &root->root_key,
1128                                         &root->root_item);
1129                 if (ret)
1130                         return ret;
1131
1132                 old_root_used = btrfs_root_used(&root->root_item);
1133         }
1134
1135         return 0;
1136 }
1137
1138 /*
1139  * update all the cowonly tree roots on disk
1140  *
1141  * The error handling in this function may not be obvious. Any of the
1142  * failures will cause the file system to go offline. We still need
1143  * to clean up the delayed refs.
1144  */
1145 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans)
1146 {
1147         struct btrfs_fs_info *fs_info = trans->fs_info;
1148         struct list_head *dirty_bgs = &trans->transaction->dirty_bgs;
1149         struct list_head *io_bgs = &trans->transaction->io_bgs;
1150         struct list_head *next;
1151         struct extent_buffer *eb;
1152         int ret;
1153
1154         eb = btrfs_lock_root_node(fs_info->tree_root);
1155         ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL,
1156                               0, &eb);
1157         btrfs_tree_unlock(eb);
1158         free_extent_buffer(eb);
1159
1160         if (ret)
1161                 return ret;
1162
1163         ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
1164         if (ret)
1165                 return ret;
1166
1167         ret = btrfs_run_dev_stats(trans, fs_info);
1168         if (ret)
1169                 return ret;
1170         ret = btrfs_run_dev_replace(trans, fs_info);
1171         if (ret)
1172                 return ret;
1173         ret = btrfs_run_qgroups(trans, fs_info);
1174         if (ret)
1175                 return ret;
1176
1177         ret = btrfs_setup_space_cache(trans, fs_info);
1178         if (ret)
1179                 return ret;
1180
1181         /* run_qgroups might have added some more refs */
1182         ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
1183         if (ret)
1184                 return ret;
1185 again:
1186         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
1187                 struct btrfs_root *root;
1188                 next = fs_info->dirty_cowonly_roots.next;
1189                 list_del_init(next);
1190                 root = list_entry(next, struct btrfs_root, dirty_list);
1191                 clear_bit(BTRFS_ROOT_DIRTY, &root->state);
1192
1193                 if (root != fs_info->extent_root)
1194                         list_add_tail(&root->dirty_list,
1195                                       &trans->transaction->switch_commits);
1196                 ret = update_cowonly_root(trans, root);
1197                 if (ret)
1198                         return ret;
1199                 ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
1200                 if (ret)
1201                         return ret;
1202         }
1203
1204         while (!list_empty(dirty_bgs) || !list_empty(io_bgs)) {
1205                 ret = btrfs_write_dirty_block_groups(trans, fs_info);
1206                 if (ret)
1207                         return ret;
1208                 ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
1209                 if (ret)
1210                         return ret;
1211         }
1212
1213         if (!list_empty(&fs_info->dirty_cowonly_roots))
1214                 goto again;
1215
1216         list_add_tail(&fs_info->extent_root->dirty_list,
1217                       &trans->transaction->switch_commits);
1218         btrfs_after_dev_replace_commit(fs_info);
1219
1220         return 0;
1221 }
1222
1223 /*
1224  * dead roots are old snapshots that need to be deleted.  This allocates
1225  * a dirty root struct and adds it into the list of dead roots that need to
1226  * be deleted
1227  */
1228 void btrfs_add_dead_root(struct btrfs_root *root)
1229 {
1230         struct btrfs_fs_info *fs_info = root->fs_info;
1231
1232         spin_lock(&fs_info->trans_lock);
1233         if (list_empty(&root->root_list))
1234                 list_add_tail(&root->root_list, &fs_info->dead_roots);
1235         spin_unlock(&fs_info->trans_lock);
1236 }
1237
1238 /*
1239  * update all the cowonly tree roots on disk
1240  */
1241 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans)
1242 {
1243         struct btrfs_fs_info *fs_info = trans->fs_info;
1244         struct btrfs_root *gang[8];
1245         int i;
1246         int ret;
1247         int err = 0;
1248
1249         spin_lock(&fs_info->fs_roots_radix_lock);
1250         while (1) {
1251                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
1252                                                  (void **)gang, 0,
1253                                                  ARRAY_SIZE(gang),
1254                                                  BTRFS_ROOT_TRANS_TAG);
1255                 if (ret == 0)
1256                         break;
1257                 for (i = 0; i < ret; i++) {
1258                         struct btrfs_root *root = gang[i];
1259                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
1260                                         (unsigned long)root->root_key.objectid,
1261                                         BTRFS_ROOT_TRANS_TAG);
1262                         spin_unlock(&fs_info->fs_roots_radix_lock);
1263
1264                         btrfs_free_log(trans, root);
1265                         btrfs_update_reloc_root(trans, root);
1266                         btrfs_orphan_commit_root(trans, root);
1267
1268                         btrfs_save_ino_cache(root, trans);
1269
1270                         /* see comments in should_cow_block() */
1271                         clear_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1272                         smp_mb__after_atomic();
1273
1274                         if (root->commit_root != root->node) {
1275                                 list_add_tail(&root->dirty_list,
1276                                         &trans->transaction->switch_commits);
1277                                 btrfs_set_root_node(&root->root_item,
1278                                                     root->node);
1279                         }
1280
1281                         err = btrfs_update_root(trans, fs_info->tree_root,
1282                                                 &root->root_key,
1283                                                 &root->root_item);
1284                         spin_lock(&fs_info->fs_roots_radix_lock);
1285                         if (err)
1286                                 break;
1287                         btrfs_qgroup_free_meta_all(root);
1288                 }
1289         }
1290         spin_unlock(&fs_info->fs_roots_radix_lock);
1291         return err;
1292 }
1293
1294 /*
1295  * defrag a given btree.
1296  * Every leaf in the btree is read and defragged.
1297  */
1298 int btrfs_defrag_root(struct btrfs_root *root)
1299 {
1300         struct btrfs_fs_info *info = root->fs_info;
1301         struct btrfs_trans_handle *trans;
1302         int ret;
1303
1304         if (test_and_set_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state))
1305                 return 0;
1306
1307         while (1) {
1308                 trans = btrfs_start_transaction(root, 0);
1309                 if (IS_ERR(trans))
1310                         return PTR_ERR(trans);
1311
1312                 ret = btrfs_defrag_leaves(trans, root);
1313
1314                 btrfs_end_transaction(trans);
1315                 btrfs_btree_balance_dirty(info);
1316                 cond_resched();
1317
1318                 if (btrfs_fs_closing(info) || ret != -EAGAIN)
1319                         break;
1320
1321                 if (btrfs_defrag_cancelled(info)) {
1322                         btrfs_debug(info, "defrag_root cancelled");
1323                         ret = -EAGAIN;
1324                         break;
1325                 }
1326         }
1327         clear_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state);
1328         return ret;
1329 }
1330
1331 /*
1332  * Do all special snapshot related qgroup dirty hack.
1333  *
1334  * Will do all needed qgroup inherit and dirty hack like switch commit
1335  * roots inside one transaction and write all btree into disk, to make
1336  * qgroup works.
1337  */
1338 static int qgroup_account_snapshot(struct btrfs_trans_handle *trans,
1339                                    struct btrfs_root *src,
1340                                    struct btrfs_root *parent,
1341                                    struct btrfs_qgroup_inherit *inherit,
1342                                    u64 dst_objectid)
1343 {
1344         struct btrfs_fs_info *fs_info = src->fs_info;
1345         int ret;
1346
1347         /*
1348          * Save some performance in the case that qgroups are not
1349          * enabled. If this check races with the ioctl, rescan will
1350          * kick in anyway.
1351          */
1352         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
1353                 return 0;
1354
1355         /*
1356          * Ensure dirty @src will be commited.  Or, after comming
1357          * commit_fs_roots() and switch_commit_roots(), any dirty but not
1358          * recorded root will never be updated again, causing an outdated root
1359          * item.
1360          */
1361         record_root_in_trans(trans, src, 1);
1362
1363         /*
1364          * We are going to commit transaction, see btrfs_commit_transaction()
1365          * comment for reason locking tree_log_mutex
1366          */
1367         mutex_lock(&fs_info->tree_log_mutex);
1368
1369         ret = commit_fs_roots(trans);
1370         if (ret)
1371                 goto out;
1372         ret = btrfs_qgroup_account_extents(trans);
1373         if (ret < 0)
1374                 goto out;
1375
1376         /* Now qgroup are all updated, we can inherit it to new qgroups */
1377         ret = btrfs_qgroup_inherit(trans, fs_info,
1378                                    src->root_key.objectid, dst_objectid,
1379                                    inherit);
1380         if (ret < 0)
1381                 goto out;
1382
1383         /*
1384          * Now we do a simplified commit transaction, which will:
1385          * 1) commit all subvolume and extent tree
1386          *    To ensure all subvolume and extent tree have a valid
1387          *    commit_root to accounting later insert_dir_item()
1388          * 2) write all btree blocks onto disk
1389          *    This is to make sure later btree modification will be cowed
1390          *    Or commit_root can be populated and cause wrong qgroup numbers
1391          * In this simplified commit, we don't really care about other trees
1392          * like chunk and root tree, as they won't affect qgroup.
1393          * And we don't write super to avoid half committed status.
1394          */
1395         ret = commit_cowonly_roots(trans);
1396         if (ret)
1397                 goto out;
1398         switch_commit_roots(trans->transaction);
1399         ret = btrfs_write_and_wait_transaction(trans);
1400         if (ret)
1401                 btrfs_handle_fs_error(fs_info, ret,
1402                         "Error while writing out transaction for qgroup");
1403
1404 out:
1405         mutex_unlock(&fs_info->tree_log_mutex);
1406
1407         /*
1408          * Force parent root to be updated, as we recorded it before so its
1409          * last_trans == cur_transid.
1410          * Or it won't be committed again onto disk after later
1411          * insert_dir_item()
1412          */
1413         if (!ret)
1414                 record_root_in_trans(trans, parent, 1);
1415         return ret;
1416 }
1417
1418 /*
1419  * new snapshots need to be created at a very specific time in the
1420  * transaction commit.  This does the actual creation.
1421  *
1422  * Note:
1423  * If the error which may affect the commitment of the current transaction
1424  * happens, we should return the error number. If the error which just affect
1425  * the creation of the pending snapshots, just return 0.
1426  */
1427 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
1428                                    struct btrfs_pending_snapshot *pending)
1429 {
1430
1431         struct btrfs_fs_info *fs_info = trans->fs_info;
1432         struct btrfs_key key;
1433         struct btrfs_root_item *new_root_item;
1434         struct btrfs_root *tree_root = fs_info->tree_root;
1435         struct btrfs_root *root = pending->root;
1436         struct btrfs_root *parent_root;
1437         struct btrfs_block_rsv *rsv;
1438         struct inode *parent_inode;
1439         struct btrfs_path *path;
1440         struct btrfs_dir_item *dir_item;
1441         struct dentry *dentry;
1442         struct extent_buffer *tmp;
1443         struct extent_buffer *old;
1444         struct timespec cur_time;
1445         int ret = 0;
1446         u64 to_reserve = 0;
1447         u64 index = 0;
1448         u64 objectid;
1449         u64 root_flags;
1450         uuid_le new_uuid;
1451
1452         ASSERT(pending->path);
1453         path = pending->path;
1454
1455         ASSERT(pending->root_item);
1456         new_root_item = pending->root_item;
1457
1458         pending->error = btrfs_find_free_objectid(tree_root, &objectid);
1459         if (pending->error)
1460                 goto no_free_objectid;
1461
1462         /*
1463          * Make qgroup to skip current new snapshot's qgroupid, as it is
1464          * accounted by later btrfs_qgroup_inherit().
1465          */
1466         btrfs_set_skip_qgroup(trans, objectid);
1467
1468         btrfs_reloc_pre_snapshot(pending, &to_reserve);
1469
1470         if (to_reserve > 0) {
1471                 pending->error = btrfs_block_rsv_add(root,
1472                                                      &pending->block_rsv,
1473                                                      to_reserve,
1474                                                      BTRFS_RESERVE_NO_FLUSH);
1475                 if (pending->error)
1476                         goto clear_skip_qgroup;
1477         }
1478
1479         key.objectid = objectid;
1480         key.offset = (u64)-1;
1481         key.type = BTRFS_ROOT_ITEM_KEY;
1482
1483         rsv = trans->block_rsv;
1484         trans->block_rsv = &pending->block_rsv;
1485         trans->bytes_reserved = trans->block_rsv->reserved;
1486         trace_btrfs_space_reservation(fs_info, "transaction",
1487                                       trans->transid,
1488                                       trans->bytes_reserved, 1);
1489         dentry = pending->dentry;
1490         parent_inode = pending->dir;
1491         parent_root = BTRFS_I(parent_inode)->root;
1492         record_root_in_trans(trans, parent_root, 0);
1493
1494         cur_time = current_time(parent_inode);
1495
1496         /*
1497          * insert the directory item
1498          */
1499         ret = btrfs_set_inode_index(BTRFS_I(parent_inode), &index);
1500         BUG_ON(ret); /* -ENOMEM */
1501
1502         /* check if there is a file/dir which has the same name. */
1503         dir_item = btrfs_lookup_dir_item(NULL, parent_root, path,
1504                                          btrfs_ino(BTRFS_I(parent_inode)),
1505                                          dentry->d_name.name,
1506                                          dentry->d_name.len, 0);
1507         if (dir_item != NULL && !IS_ERR(dir_item)) {
1508                 pending->error = -EEXIST;
1509                 goto dir_item_existed;
1510         } else if (IS_ERR(dir_item)) {
1511                 ret = PTR_ERR(dir_item);
1512                 btrfs_abort_transaction(trans, ret);
1513                 goto fail;
1514         }
1515         btrfs_release_path(path);
1516
1517         /*
1518          * pull in the delayed directory update
1519          * and the delayed inode item
1520          * otherwise we corrupt the FS during
1521          * snapshot
1522          */
1523         ret = btrfs_run_delayed_items(trans);
1524         if (ret) {      /* Transaction aborted */
1525                 btrfs_abort_transaction(trans, ret);
1526                 goto fail;
1527         }
1528
1529         record_root_in_trans(trans, root, 0);
1530         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
1531         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
1532         btrfs_check_and_init_root_item(new_root_item);
1533
1534         root_flags = btrfs_root_flags(new_root_item);
1535         if (pending->readonly)
1536                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
1537         else
1538                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
1539         btrfs_set_root_flags(new_root_item, root_flags);
1540
1541         btrfs_set_root_generation_v2(new_root_item,
1542                         trans->transid);
1543         uuid_le_gen(&new_uuid);
1544         memcpy(new_root_item->uuid, new_uuid.b, BTRFS_UUID_SIZE);
1545         memcpy(new_root_item->parent_uuid, root->root_item.uuid,
1546                         BTRFS_UUID_SIZE);
1547         if (!(root_flags & BTRFS_ROOT_SUBVOL_RDONLY)) {
1548                 memset(new_root_item->received_uuid, 0,
1549                        sizeof(new_root_item->received_uuid));
1550                 memset(&new_root_item->stime, 0, sizeof(new_root_item->stime));
1551                 memset(&new_root_item->rtime, 0, sizeof(new_root_item->rtime));
1552                 btrfs_set_root_stransid(new_root_item, 0);
1553                 btrfs_set_root_rtransid(new_root_item, 0);
1554         }
1555         btrfs_set_stack_timespec_sec(&new_root_item->otime, cur_time.tv_sec);
1556         btrfs_set_stack_timespec_nsec(&new_root_item->otime, cur_time.tv_nsec);
1557         btrfs_set_root_otransid(new_root_item, trans->transid);
1558
1559         old = btrfs_lock_root_node(root);
1560         ret = btrfs_cow_block(trans, root, old, NULL, 0, &old);
1561         if (ret) {
1562                 btrfs_tree_unlock(old);
1563                 free_extent_buffer(old);
1564                 btrfs_abort_transaction(trans, ret);
1565                 goto fail;
1566         }
1567
1568         btrfs_set_lock_blocking(old);
1569
1570         ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1571         /* clean up in any case */
1572         btrfs_tree_unlock(old);
1573         free_extent_buffer(old);
1574         if (ret) {
1575                 btrfs_abort_transaction(trans, ret);
1576                 goto fail;
1577         }
1578         /* see comments in should_cow_block() */
1579         set_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1580         smp_wmb();
1581
1582         btrfs_set_root_node(new_root_item, tmp);
1583         /* record when the snapshot was created in key.offset */
1584         key.offset = trans->transid;
1585         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1586         btrfs_tree_unlock(tmp);
1587         free_extent_buffer(tmp);
1588         if (ret) {
1589                 btrfs_abort_transaction(trans, ret);
1590                 goto fail;
1591         }
1592
1593         /*
1594          * insert root back/forward references
1595          */
1596         ret = btrfs_add_root_ref(trans, fs_info, objectid,
1597                                  parent_root->root_key.objectid,
1598                                  btrfs_ino(BTRFS_I(parent_inode)), index,
1599                                  dentry->d_name.name, dentry->d_name.len);
1600         if (ret) {
1601                 btrfs_abort_transaction(trans, ret);
1602                 goto fail;
1603         }
1604
1605         key.offset = (u64)-1;
1606         pending->snap = btrfs_read_fs_root_no_name(fs_info, &key);
1607         if (IS_ERR(pending->snap)) {
1608                 ret = PTR_ERR(pending->snap);
1609                 btrfs_abort_transaction(trans, ret);
1610                 goto fail;
1611         }
1612
1613         ret = btrfs_reloc_post_snapshot(trans, pending);
1614         if (ret) {
1615                 btrfs_abort_transaction(trans, ret);
1616                 goto fail;
1617         }
1618
1619         ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
1620         if (ret) {
1621                 btrfs_abort_transaction(trans, ret);
1622                 goto fail;
1623         }
1624
1625         /*
1626          * Do special qgroup accounting for snapshot, as we do some qgroup
1627          * snapshot hack to do fast snapshot.
1628          * To co-operate with that hack, we do hack again.
1629          * Or snapshot will be greatly slowed down by a subtree qgroup rescan
1630          */
1631         ret = qgroup_account_snapshot(trans, root, parent_root,
1632                                       pending->inherit, objectid);
1633         if (ret < 0)
1634                 goto fail;
1635
1636         ret = btrfs_insert_dir_item(trans, parent_root,
1637                                     dentry->d_name.name, dentry->d_name.len,
1638                                     BTRFS_I(parent_inode), &key,
1639                                     BTRFS_FT_DIR, index);
1640         /* We have check then name at the beginning, so it is impossible. */
1641         BUG_ON(ret == -EEXIST || ret == -EOVERFLOW);
1642         if (ret) {
1643                 btrfs_abort_transaction(trans, ret);
1644                 goto fail;
1645         }
1646
1647         btrfs_i_size_write(BTRFS_I(parent_inode), parent_inode->i_size +
1648                                          dentry->d_name.len * 2);
1649         parent_inode->i_mtime = parent_inode->i_ctime =
1650                 current_time(parent_inode);
1651         ret = btrfs_update_inode_fallback(trans, parent_root, parent_inode);
1652         if (ret) {
1653                 btrfs_abort_transaction(trans, ret);
1654                 goto fail;
1655         }
1656         ret = btrfs_uuid_tree_add(trans, fs_info, new_uuid.b,
1657                                   BTRFS_UUID_KEY_SUBVOL, objectid);
1658         if (ret) {
1659                 btrfs_abort_transaction(trans, ret);
1660                 goto fail;
1661         }
1662         if (!btrfs_is_empty_uuid(new_root_item->received_uuid)) {
1663                 ret = btrfs_uuid_tree_add(trans, fs_info,
1664                                           new_root_item->received_uuid,
1665                                           BTRFS_UUID_KEY_RECEIVED_SUBVOL,
1666                                           objectid);
1667                 if (ret && ret != -EEXIST) {
1668                         btrfs_abort_transaction(trans, ret);
1669                         goto fail;
1670                 }
1671         }
1672
1673         ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
1674         if (ret) {
1675                 btrfs_abort_transaction(trans, ret);
1676                 goto fail;
1677         }
1678
1679 fail:
1680         pending->error = ret;
1681 dir_item_existed:
1682         trans->block_rsv = rsv;
1683         trans->bytes_reserved = 0;
1684 clear_skip_qgroup:
1685         btrfs_clear_skip_qgroup(trans);
1686 no_free_objectid:
1687         kfree(new_root_item);
1688         pending->root_item = NULL;
1689         btrfs_free_path(path);
1690         pending->path = NULL;
1691
1692         return ret;
1693 }
1694
1695 /*
1696  * create all the snapshots we've scheduled for creation
1697  */
1698 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans)
1699 {
1700         struct btrfs_pending_snapshot *pending, *next;
1701         struct list_head *head = &trans->transaction->pending_snapshots;
1702         int ret = 0;
1703
1704         list_for_each_entry_safe(pending, next, head, list) {
1705                 list_del(&pending->list);
1706                 ret = create_pending_snapshot(trans, pending);
1707                 if (ret)
1708                         break;
1709         }
1710         return ret;
1711 }
1712
1713 static void update_super_roots(struct btrfs_fs_info *fs_info)
1714 {
1715         struct btrfs_root_item *root_item;
1716         struct btrfs_super_block *super;
1717
1718         super = fs_info->super_copy;
1719
1720         root_item = &fs_info->chunk_root->root_item;
1721         super->chunk_root = root_item->bytenr;
1722         super->chunk_root_generation = root_item->generation;
1723         super->chunk_root_level = root_item->level;
1724
1725         root_item = &fs_info->tree_root->root_item;
1726         super->root = root_item->bytenr;
1727         super->generation = root_item->generation;
1728         super->root_level = root_item->level;
1729         if (btrfs_test_opt(fs_info, SPACE_CACHE))
1730                 super->cache_generation = root_item->generation;
1731         if (test_bit(BTRFS_FS_UPDATE_UUID_TREE_GEN, &fs_info->flags))
1732                 super->uuid_tree_generation = root_item->generation;
1733 }
1734
1735 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1736 {
1737         struct btrfs_transaction *trans;
1738         int ret = 0;
1739
1740         spin_lock(&info->trans_lock);
1741         trans = info->running_transaction;
1742         if (trans)
1743                 ret = (trans->state >= TRANS_STATE_COMMIT_START);
1744         spin_unlock(&info->trans_lock);
1745         return ret;
1746 }
1747
1748 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1749 {
1750         struct btrfs_transaction *trans;
1751         int ret = 0;
1752
1753         spin_lock(&info->trans_lock);
1754         trans = info->running_transaction;
1755         if (trans)
1756                 ret = is_transaction_blocked(trans);
1757         spin_unlock(&info->trans_lock);
1758         return ret;
1759 }
1760
1761 /*
1762  * wait for the current transaction commit to start and block subsequent
1763  * transaction joins
1764  */
1765 static void wait_current_trans_commit_start(struct btrfs_fs_info *fs_info,
1766                                             struct btrfs_transaction *trans)
1767 {
1768         wait_event(fs_info->transaction_blocked_wait,
1769                    trans->state >= TRANS_STATE_COMMIT_START || trans->aborted);
1770 }
1771
1772 /*
1773  * wait for the current transaction to start and then become unblocked.
1774  * caller holds ref.
1775  */
1776 static void wait_current_trans_commit_start_and_unblock(
1777                                         struct btrfs_fs_info *fs_info,
1778                                         struct btrfs_transaction *trans)
1779 {
1780         wait_event(fs_info->transaction_wait,
1781                    trans->state >= TRANS_STATE_UNBLOCKED || trans->aborted);
1782 }
1783
1784 /*
1785  * commit transactions asynchronously. once btrfs_commit_transaction_async
1786  * returns, any subsequent transaction will not be allowed to join.
1787  */
1788 struct btrfs_async_commit {
1789         struct btrfs_trans_handle *newtrans;
1790         struct work_struct work;
1791 };
1792
1793 static void do_async_commit(struct work_struct *work)
1794 {
1795         struct btrfs_async_commit *ac =
1796                 container_of(work, struct btrfs_async_commit, work);
1797
1798         /*
1799          * We've got freeze protection passed with the transaction.
1800          * Tell lockdep about it.
1801          */
1802         if (ac->newtrans->type & __TRANS_FREEZABLE)
1803                 __sb_writers_acquired(ac->newtrans->fs_info->sb, SB_FREEZE_FS);
1804
1805         current->journal_info = ac->newtrans;
1806
1807         btrfs_commit_transaction(ac->newtrans);
1808         kfree(ac);
1809 }
1810
1811 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1812                                    int wait_for_unblock)
1813 {
1814         struct btrfs_fs_info *fs_info = trans->fs_info;
1815         struct btrfs_async_commit *ac;
1816         struct btrfs_transaction *cur_trans;
1817
1818         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1819         if (!ac)
1820                 return -ENOMEM;
1821
1822         INIT_WORK(&ac->work, do_async_commit);
1823         ac->newtrans = btrfs_join_transaction(trans->root);
1824         if (IS_ERR(ac->newtrans)) {
1825                 int err = PTR_ERR(ac->newtrans);
1826                 kfree(ac);
1827                 return err;
1828         }
1829
1830         /* take transaction reference */
1831         cur_trans = trans->transaction;
1832         refcount_inc(&cur_trans->use_count);
1833
1834         btrfs_end_transaction(trans);
1835
1836         /*
1837          * Tell lockdep we've released the freeze rwsem, since the
1838          * async commit thread will be the one to unlock it.
1839          */
1840         if (ac->newtrans->type & __TRANS_FREEZABLE)
1841                 __sb_writers_release(fs_info->sb, SB_FREEZE_FS);
1842
1843         schedule_work(&ac->work);
1844
1845         /* wait for transaction to start and unblock */
1846         if (wait_for_unblock)
1847                 wait_current_trans_commit_start_and_unblock(fs_info, cur_trans);
1848         else
1849                 wait_current_trans_commit_start(fs_info, cur_trans);
1850
1851         if (current->journal_info == trans)
1852                 current->journal_info = NULL;
1853
1854         btrfs_put_transaction(cur_trans);
1855         return 0;
1856 }
1857
1858
1859 static void cleanup_transaction(struct btrfs_trans_handle *trans, int err)
1860 {
1861         struct btrfs_fs_info *fs_info = trans->fs_info;
1862         struct btrfs_transaction *cur_trans = trans->transaction;
1863         DEFINE_WAIT(wait);
1864
1865         WARN_ON(refcount_read(&trans->use_count) > 1);
1866
1867         btrfs_abort_transaction(trans, err);
1868
1869         spin_lock(&fs_info->trans_lock);
1870
1871         /*
1872          * If the transaction is removed from the list, it means this
1873          * transaction has been committed successfully, so it is impossible
1874          * to call the cleanup function.
1875          */
1876         BUG_ON(list_empty(&cur_trans->list));
1877
1878         list_del_init(&cur_trans->list);
1879         if (cur_trans == fs_info->running_transaction) {
1880                 cur_trans->state = TRANS_STATE_COMMIT_DOING;
1881                 spin_unlock(&fs_info->trans_lock);
1882                 wait_event(cur_trans->writer_wait,
1883                            atomic_read(&cur_trans->num_writers) == 1);
1884
1885                 spin_lock(&fs_info->trans_lock);
1886         }
1887         spin_unlock(&fs_info->trans_lock);
1888
1889         btrfs_cleanup_one_transaction(trans->transaction, fs_info);
1890
1891         spin_lock(&fs_info->trans_lock);
1892         if (cur_trans == fs_info->running_transaction)
1893                 fs_info->running_transaction = NULL;
1894         spin_unlock(&fs_info->trans_lock);
1895
1896         if (trans->type & __TRANS_FREEZABLE)
1897                 sb_end_intwrite(fs_info->sb);
1898         btrfs_put_transaction(cur_trans);
1899         btrfs_put_transaction(cur_trans);
1900
1901         trace_btrfs_transaction_commit(trans->root);
1902
1903         if (current->journal_info == trans)
1904                 current->journal_info = NULL;
1905         btrfs_scrub_cancel(fs_info);
1906
1907         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1908 }
1909
1910 static inline int btrfs_start_delalloc_flush(struct btrfs_fs_info *fs_info)
1911 {
1912         /*
1913          * We use writeback_inodes_sb here because if we used
1914          * btrfs_start_delalloc_roots we would deadlock with fs freeze.
1915          * Currently are holding the fs freeze lock, if we do an async flush
1916          * we'll do btrfs_join_transaction() and deadlock because we need to
1917          * wait for the fs freeze lock.  Using the direct flushing we benefit
1918          * from already being in a transaction and our join_transaction doesn't
1919          * have to re-take the fs freeze lock.
1920          */
1921         if (btrfs_test_opt(fs_info, FLUSHONCOMMIT))
1922                 writeback_inodes_sb(fs_info->sb, WB_REASON_SYNC);
1923         return 0;
1924 }
1925
1926 static inline void btrfs_wait_delalloc_flush(struct btrfs_fs_info *fs_info)
1927 {
1928         if (btrfs_test_opt(fs_info, FLUSHONCOMMIT))
1929                 btrfs_wait_ordered_roots(fs_info, U64_MAX, 0, (u64)-1);
1930 }
1931
1932 static inline void
1933 btrfs_wait_pending_ordered(struct btrfs_transaction *cur_trans)
1934 {
1935         wait_event(cur_trans->pending_wait,
1936                    atomic_read(&cur_trans->pending_ordered) == 0);
1937 }
1938
1939 int btrfs_commit_transaction(struct btrfs_trans_handle *trans)
1940 {
1941         struct btrfs_fs_info *fs_info = trans->fs_info;
1942         struct btrfs_transaction *cur_trans = trans->transaction;
1943         struct btrfs_transaction *prev_trans = NULL;
1944         int ret;
1945
1946         /* Stop the commit early if ->aborted is set */
1947         if (unlikely(READ_ONCE(cur_trans->aborted))) {
1948                 ret = cur_trans->aborted;
1949                 btrfs_end_transaction(trans);
1950                 return ret;
1951         }
1952
1953         /* make a pass through all the delayed refs we have so far
1954          * any runnings procs may add more while we are here
1955          */
1956         ret = btrfs_run_delayed_refs(trans, 0);
1957         if (ret) {
1958                 btrfs_end_transaction(trans);
1959                 return ret;
1960         }
1961
1962         btrfs_trans_release_metadata(trans);
1963         trans->block_rsv = NULL;
1964
1965         cur_trans = trans->transaction;
1966
1967         /*
1968          * set the flushing flag so procs in this transaction have to
1969          * start sending their work down.
1970          */
1971         cur_trans->delayed_refs.flushing = 1;
1972         smp_wmb();
1973
1974         if (!list_empty(&trans->new_bgs))
1975                 btrfs_create_pending_block_groups(trans);
1976
1977         ret = btrfs_run_delayed_refs(trans, 0);
1978         if (ret) {
1979                 btrfs_end_transaction(trans);
1980                 return ret;
1981         }
1982
1983         if (!test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &cur_trans->flags)) {
1984                 int run_it = 0;
1985
1986                 /* this mutex is also taken before trying to set
1987                  * block groups readonly.  We need to make sure
1988                  * that nobody has set a block group readonly
1989                  * after a extents from that block group have been
1990                  * allocated for cache files.  btrfs_set_block_group_ro
1991                  * will wait for the transaction to commit if it
1992                  * finds BTRFS_TRANS_DIRTY_BG_RUN set.
1993                  *
1994                  * The BTRFS_TRANS_DIRTY_BG_RUN flag is also used to make sure
1995                  * only one process starts all the block group IO.  It wouldn't
1996                  * hurt to have more than one go through, but there's no
1997                  * real advantage to it either.
1998                  */
1999                 mutex_lock(&fs_info->ro_block_group_mutex);
2000                 if (!test_and_set_bit(BTRFS_TRANS_DIRTY_BG_RUN,
2001                                       &cur_trans->flags))
2002                         run_it = 1;
2003                 mutex_unlock(&fs_info->ro_block_group_mutex);
2004
2005                 if (run_it) {
2006                         ret = btrfs_start_dirty_block_groups(trans);
2007                         if (ret) {
2008                                 btrfs_end_transaction(trans);
2009                                 return ret;
2010                         }
2011                 }
2012         }
2013
2014         spin_lock(&fs_info->trans_lock);
2015         if (cur_trans->state >= TRANS_STATE_COMMIT_START) {
2016                 spin_unlock(&fs_info->trans_lock);
2017                 refcount_inc(&cur_trans->use_count);
2018                 ret = btrfs_end_transaction(trans);
2019
2020                 wait_for_commit(cur_trans);
2021
2022                 if (unlikely(cur_trans->aborted))
2023                         ret = cur_trans->aborted;
2024
2025                 btrfs_put_transaction(cur_trans);
2026
2027                 return ret;
2028         }
2029
2030         cur_trans->state = TRANS_STATE_COMMIT_START;
2031         wake_up(&fs_info->transaction_blocked_wait);
2032
2033         if (cur_trans->list.prev != &fs_info->trans_list) {
2034                 prev_trans = list_entry(cur_trans->list.prev,
2035                                         struct btrfs_transaction, list);
2036                 if (prev_trans->state != TRANS_STATE_COMPLETED) {
2037                         refcount_inc(&prev_trans->use_count);
2038                         spin_unlock(&fs_info->trans_lock);
2039
2040                         wait_for_commit(prev_trans);
2041                         ret = prev_trans->aborted;
2042
2043                         btrfs_put_transaction(prev_trans);
2044                         if (ret)
2045                                 goto cleanup_transaction;
2046                 } else {
2047                         spin_unlock(&fs_info->trans_lock);
2048                 }
2049         } else {
2050                 spin_unlock(&fs_info->trans_lock);
2051         }
2052
2053         extwriter_counter_dec(cur_trans, trans->type);
2054
2055         ret = btrfs_start_delalloc_flush(fs_info);
2056         if (ret)
2057                 goto cleanup_transaction;
2058
2059         ret = btrfs_run_delayed_items(trans);
2060         if (ret)
2061                 goto cleanup_transaction;
2062
2063         wait_event(cur_trans->writer_wait,
2064                    extwriter_counter_read(cur_trans) == 0);
2065
2066         /* some pending stuffs might be added after the previous flush. */
2067         ret = btrfs_run_delayed_items(trans);
2068         if (ret)
2069                 goto cleanup_transaction;
2070
2071         btrfs_wait_delalloc_flush(fs_info);
2072
2073         btrfs_wait_pending_ordered(cur_trans);
2074
2075         btrfs_scrub_pause(fs_info);
2076         /*
2077          * Ok now we need to make sure to block out any other joins while we
2078          * commit the transaction.  We could have started a join before setting
2079          * COMMIT_DOING so make sure to wait for num_writers to == 1 again.
2080          */
2081         spin_lock(&fs_info->trans_lock);
2082         cur_trans->state = TRANS_STATE_COMMIT_DOING;
2083         spin_unlock(&fs_info->trans_lock);
2084         wait_event(cur_trans->writer_wait,
2085                    atomic_read(&cur_trans->num_writers) == 1);
2086
2087         /* ->aborted might be set after the previous check, so check it */
2088         if (unlikely(READ_ONCE(cur_trans->aborted))) {
2089                 ret = cur_trans->aborted;
2090                 goto scrub_continue;
2091         }
2092         /*
2093          * the reloc mutex makes sure that we stop
2094          * the balancing code from coming in and moving
2095          * extents around in the middle of the commit
2096          */
2097         mutex_lock(&fs_info->reloc_mutex);
2098
2099         /*
2100          * We needn't worry about the delayed items because we will
2101          * deal with them in create_pending_snapshot(), which is the
2102          * core function of the snapshot creation.
2103          */
2104         ret = create_pending_snapshots(trans);
2105         if (ret) {
2106                 mutex_unlock(&fs_info->reloc_mutex);
2107                 goto scrub_continue;
2108         }
2109
2110         /*
2111          * We insert the dir indexes of the snapshots and update the inode
2112          * of the snapshots' parents after the snapshot creation, so there
2113          * are some delayed items which are not dealt with. Now deal with
2114          * them.
2115          *
2116          * We needn't worry that this operation will corrupt the snapshots,
2117          * because all the tree which are snapshoted will be forced to COW
2118          * the nodes and leaves.
2119          */
2120         ret = btrfs_run_delayed_items(trans);
2121         if (ret) {
2122                 mutex_unlock(&fs_info->reloc_mutex);
2123                 goto scrub_continue;
2124         }
2125
2126         ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
2127         if (ret) {
2128                 mutex_unlock(&fs_info->reloc_mutex);
2129                 goto scrub_continue;
2130         }
2131
2132         /*
2133          * make sure none of the code above managed to slip in a
2134          * delayed item
2135          */
2136         btrfs_assert_delayed_root_empty(fs_info);
2137
2138         WARN_ON(cur_trans != trans->transaction);
2139
2140         /* btrfs_commit_tree_roots is responsible for getting the
2141          * various roots consistent with each other.  Every pointer
2142          * in the tree of tree roots has to point to the most up to date
2143          * root for every subvolume and other tree.  So, we have to keep
2144          * the tree logging code from jumping in and changing any
2145          * of the trees.
2146          *
2147          * At this point in the commit, there can't be any tree-log
2148          * writers, but a little lower down we drop the trans mutex
2149          * and let new people in.  By holding the tree_log_mutex
2150          * from now until after the super is written, we avoid races
2151          * with the tree-log code.
2152          */
2153         mutex_lock(&fs_info->tree_log_mutex);
2154
2155         ret = commit_fs_roots(trans);
2156         if (ret) {
2157                 mutex_unlock(&fs_info->tree_log_mutex);
2158                 mutex_unlock(&fs_info->reloc_mutex);
2159                 goto scrub_continue;
2160         }
2161
2162         /*
2163          * Since the transaction is done, we can apply the pending changes
2164          * before the next transaction.
2165          */
2166         btrfs_apply_pending_changes(fs_info);
2167
2168         /* commit_fs_roots gets rid of all the tree log roots, it is now
2169          * safe to free the root of tree log roots
2170          */
2171         btrfs_free_log_root_tree(trans, fs_info);
2172
2173         /*
2174          * commit_fs_roots() can call btrfs_save_ino_cache(), which generates
2175          * new delayed refs. Must handle them or qgroup can be wrong.
2176          */
2177         ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
2178         if (ret) {
2179                 mutex_unlock(&fs_info->tree_log_mutex);
2180                 mutex_unlock(&fs_info->reloc_mutex);
2181                 goto scrub_continue;
2182         }
2183
2184         /*
2185          * Since fs roots are all committed, we can get a quite accurate
2186          * new_roots. So let's do quota accounting.
2187          */
2188         ret = btrfs_qgroup_account_extents(trans);
2189         if (ret < 0) {
2190                 mutex_unlock(&fs_info->tree_log_mutex);
2191                 mutex_unlock(&fs_info->reloc_mutex);
2192                 goto scrub_continue;
2193         }
2194
2195         ret = commit_cowonly_roots(trans);
2196         if (ret) {
2197                 mutex_unlock(&fs_info->tree_log_mutex);
2198                 mutex_unlock(&fs_info->reloc_mutex);
2199                 goto scrub_continue;
2200         }
2201
2202         /*
2203          * The tasks which save the space cache and inode cache may also
2204          * update ->aborted, check it.
2205          */
2206         if (unlikely(READ_ONCE(cur_trans->aborted))) {
2207                 ret = cur_trans->aborted;
2208                 mutex_unlock(&fs_info->tree_log_mutex);
2209                 mutex_unlock(&fs_info->reloc_mutex);
2210                 goto scrub_continue;
2211         }
2212
2213         btrfs_prepare_extent_commit(fs_info);
2214
2215         cur_trans = fs_info->running_transaction;
2216
2217         btrfs_set_root_node(&fs_info->tree_root->root_item,
2218                             fs_info->tree_root->node);
2219         list_add_tail(&fs_info->tree_root->dirty_list,
2220                       &cur_trans->switch_commits);
2221
2222         btrfs_set_root_node(&fs_info->chunk_root->root_item,
2223                             fs_info->chunk_root->node);
2224         list_add_tail(&fs_info->chunk_root->dirty_list,
2225                       &cur_trans->switch_commits);
2226
2227         switch_commit_roots(cur_trans);
2228
2229         ASSERT(list_empty(&cur_trans->dirty_bgs));
2230         ASSERT(list_empty(&cur_trans->io_bgs));
2231         update_super_roots(fs_info);
2232
2233         btrfs_set_super_log_root(fs_info->super_copy, 0);
2234         btrfs_set_super_log_root_level(fs_info->super_copy, 0);
2235         memcpy(fs_info->super_for_commit, fs_info->super_copy,
2236                sizeof(*fs_info->super_copy));
2237
2238         btrfs_update_commit_device_size(fs_info);
2239         btrfs_update_commit_device_bytes_used(cur_trans);
2240
2241         clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags);
2242         clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags);
2243
2244         btrfs_trans_release_chunk_metadata(trans);
2245
2246         spin_lock(&fs_info->trans_lock);
2247         cur_trans->state = TRANS_STATE_UNBLOCKED;
2248         fs_info->running_transaction = NULL;
2249         spin_unlock(&fs_info->trans_lock);
2250         mutex_unlock(&fs_info->reloc_mutex);
2251
2252         wake_up(&fs_info->transaction_wait);
2253
2254         ret = btrfs_write_and_wait_transaction(trans);
2255         if (ret) {
2256                 btrfs_handle_fs_error(fs_info, ret,
2257                                       "Error while writing out transaction");
2258                 mutex_unlock(&fs_info->tree_log_mutex);
2259                 goto scrub_continue;
2260         }
2261
2262         ret = write_all_supers(fs_info, 0);
2263         /*
2264          * the super is written, we can safely allow the tree-loggers
2265          * to go about their business
2266          */
2267         mutex_unlock(&fs_info->tree_log_mutex);
2268         if (ret)
2269                 goto scrub_continue;
2270
2271         btrfs_finish_extent_commit(trans, fs_info);
2272
2273         if (test_bit(BTRFS_TRANS_HAVE_FREE_BGS, &cur_trans->flags))
2274                 btrfs_clear_space_info_full(fs_info);
2275
2276         fs_info->last_trans_committed = cur_trans->transid;
2277         /*
2278          * We needn't acquire the lock here because there is no other task
2279          * which can change it.
2280          */
2281         cur_trans->state = TRANS_STATE_COMPLETED;
2282         wake_up(&cur_trans->commit_wait);
2283
2284         spin_lock(&fs_info->trans_lock);
2285         list_del_init(&cur_trans->list);
2286         spin_unlock(&fs_info->trans_lock);
2287
2288         btrfs_put_transaction(cur_trans);
2289         btrfs_put_transaction(cur_trans);
2290
2291         if (trans->type & __TRANS_FREEZABLE)
2292                 sb_end_intwrite(fs_info->sb);
2293
2294         trace_btrfs_transaction_commit(trans->root);
2295
2296         btrfs_scrub_continue(fs_info);
2297
2298         if (current->journal_info == trans)
2299                 current->journal_info = NULL;
2300
2301         kmem_cache_free(btrfs_trans_handle_cachep, trans);
2302
2303         /*
2304          * If fs has been frozen, we can not handle delayed iputs, otherwise
2305          * it'll result in deadlock about SB_FREEZE_FS.
2306          */
2307         if (current != fs_info->transaction_kthread &&
2308             current != fs_info->cleaner_kthread &&
2309             !test_bit(BTRFS_FS_FROZEN, &fs_info->flags))
2310                 btrfs_run_delayed_iputs(fs_info);
2311
2312         return ret;
2313
2314 scrub_continue:
2315         btrfs_scrub_continue(fs_info);
2316 cleanup_transaction:
2317         btrfs_trans_release_metadata(trans);
2318         btrfs_trans_release_chunk_metadata(trans);
2319         trans->block_rsv = NULL;
2320         btrfs_warn(fs_info, "Skipping commit of aborted transaction.");
2321         if (current->journal_info == trans)
2322                 current->journal_info = NULL;
2323         cleanup_transaction(trans, ret);
2324
2325         return ret;
2326 }
2327
2328 /*
2329  * return < 0 if error
2330  * 0 if there are no more dead_roots at the time of call
2331  * 1 there are more to be processed, call me again
2332  *
2333  * The return value indicates there are certainly more snapshots to delete, but
2334  * if there comes a new one during processing, it may return 0. We don't mind,
2335  * because btrfs_commit_super will poke cleaner thread and it will process it a
2336  * few seconds later.
2337  */
2338 int btrfs_clean_one_deleted_snapshot(struct btrfs_root *root)
2339 {
2340         int ret;
2341         struct btrfs_fs_info *fs_info = root->fs_info;
2342
2343         spin_lock(&fs_info->trans_lock);
2344         if (list_empty(&fs_info->dead_roots)) {
2345                 spin_unlock(&fs_info->trans_lock);
2346                 return 0;
2347         }
2348         root = list_first_entry(&fs_info->dead_roots,
2349                         struct btrfs_root, root_list);
2350         list_del_init(&root->root_list);
2351         spin_unlock(&fs_info->trans_lock);
2352
2353         btrfs_debug(fs_info, "cleaner removing %llu", root->objectid);
2354
2355         btrfs_kill_all_delayed_nodes(root);
2356
2357         if (btrfs_header_backref_rev(root->node) <
2358                         BTRFS_MIXED_BACKREF_REV)
2359                 ret = btrfs_drop_snapshot(root, NULL, 0, 0);
2360         else
2361                 ret = btrfs_drop_snapshot(root, NULL, 1, 0);
2362
2363         return (ret < 0) ? 0 : 1;
2364 }
2365
2366 void btrfs_apply_pending_changes(struct btrfs_fs_info *fs_info)
2367 {
2368         unsigned long prev;
2369         unsigned long bit;
2370
2371         prev = xchg(&fs_info->pending_changes, 0);
2372         if (!prev)
2373                 return;
2374
2375         bit = 1 << BTRFS_PENDING_SET_INODE_MAP_CACHE;
2376         if (prev & bit)
2377                 btrfs_set_opt(fs_info->mount_opt, INODE_MAP_CACHE);
2378         prev &= ~bit;
2379
2380         bit = 1 << BTRFS_PENDING_CLEAR_INODE_MAP_CACHE;
2381         if (prev & bit)
2382                 btrfs_clear_opt(fs_info->mount_opt, INODE_MAP_CACHE);
2383         prev &= ~bit;
2384
2385         bit = 1 << BTRFS_PENDING_COMMIT;
2386         if (prev & bit)
2387                 btrfs_debug(fs_info, "pending commit done");
2388         prev &= ~bit;
2389
2390         if (prev)
2391                 btrfs_warn(fs_info,
2392                         "unknown pending changes left 0x%lx, ignoring", prev);
2393 }