btrfs: Remove unused flush var in shrink_delalloc
[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         struct btrfs_fs_info *fs_info = trans->fs_info;
785         int updates;
786         int err;
787
788         smp_mb();
789         if (cur_trans->state >= TRANS_STATE_BLOCKED ||
790             cur_trans->delayed_refs.flushing)
791                 return 1;
792
793         updates = trans->delayed_ref_updates;
794         trans->delayed_ref_updates = 0;
795         if (updates) {
796                 err = btrfs_run_delayed_refs(trans, fs_info, updates * 2);
797                 if (err) /* Error code will also eval true */
798                         return err;
799         }
800
801         return should_end_transaction(trans);
802 }
803
804 static void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans)
805
806 {
807         struct btrfs_fs_info *fs_info = trans->fs_info;
808
809         if (!trans->block_rsv) {
810                 ASSERT(!trans->bytes_reserved);
811                 return;
812         }
813
814         if (!trans->bytes_reserved)
815                 return;
816
817         ASSERT(trans->block_rsv == &fs_info->trans_block_rsv);
818         trace_btrfs_space_reservation(fs_info, "transaction",
819                                       trans->transid, trans->bytes_reserved, 0);
820         btrfs_block_rsv_release(fs_info, trans->block_rsv,
821                                 trans->bytes_reserved);
822         trans->bytes_reserved = 0;
823 }
824
825 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
826                                    int throttle)
827 {
828         struct btrfs_fs_info *info = trans->fs_info;
829         struct btrfs_transaction *cur_trans = trans->transaction;
830         u64 transid = trans->transid;
831         unsigned long cur = trans->delayed_ref_updates;
832         int lock = (trans->type != TRANS_JOIN_NOLOCK);
833         int err = 0;
834         int must_run_delayed_refs = 0;
835
836         if (refcount_read(&trans->use_count) > 1) {
837                 refcount_dec(&trans->use_count);
838                 trans->block_rsv = trans->orig_rsv;
839                 return 0;
840         }
841
842         btrfs_trans_release_metadata(trans);
843         trans->block_rsv = NULL;
844
845         if (!list_empty(&trans->new_bgs))
846                 btrfs_create_pending_block_groups(trans);
847
848         trans->delayed_ref_updates = 0;
849         if (!trans->sync) {
850                 must_run_delayed_refs =
851                         btrfs_should_throttle_delayed_refs(trans, info);
852                 cur = max_t(unsigned long, cur, 32);
853
854                 /*
855                  * don't make the caller wait if they are from a NOLOCK
856                  * or ATTACH transaction, it will deadlock with commit
857                  */
858                 if (must_run_delayed_refs == 1 &&
859                     (trans->type & (__TRANS_JOIN_NOLOCK | __TRANS_ATTACH)))
860                         must_run_delayed_refs = 2;
861         }
862
863         btrfs_trans_release_metadata(trans);
864         trans->block_rsv = NULL;
865
866         if (!list_empty(&trans->new_bgs))
867                 btrfs_create_pending_block_groups(trans);
868
869         btrfs_trans_release_chunk_metadata(trans);
870
871         if (lock && should_end_transaction(trans) &&
872             READ_ONCE(cur_trans->state) == TRANS_STATE_RUNNING) {
873                 spin_lock(&info->trans_lock);
874                 if (cur_trans->state == TRANS_STATE_RUNNING)
875                         cur_trans->state = TRANS_STATE_BLOCKED;
876                 spin_unlock(&info->trans_lock);
877         }
878
879         if (lock && READ_ONCE(cur_trans->state) == TRANS_STATE_BLOCKED) {
880                 if (throttle)
881                         return btrfs_commit_transaction(trans);
882                 else
883                         wake_up_process(info->transaction_kthread);
884         }
885
886         if (trans->type & __TRANS_FREEZABLE)
887                 sb_end_intwrite(info->sb);
888
889         WARN_ON(cur_trans != info->running_transaction);
890         WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
891         atomic_dec(&cur_trans->num_writers);
892         extwriter_counter_dec(cur_trans, trans->type);
893
894         /*
895          * Make sure counter is updated before we wake up waiters.
896          */
897         smp_mb();
898         if (waitqueue_active(&cur_trans->writer_wait))
899                 wake_up(&cur_trans->writer_wait);
900         btrfs_put_transaction(cur_trans);
901
902         if (current->journal_info == trans)
903                 current->journal_info = NULL;
904
905         if (throttle)
906                 btrfs_run_delayed_iputs(info);
907
908         if (trans->aborted ||
909             test_bit(BTRFS_FS_STATE_ERROR, &info->fs_state)) {
910                 wake_up_process(info->transaction_kthread);
911                 err = -EIO;
912         }
913
914         kmem_cache_free(btrfs_trans_handle_cachep, trans);
915         if (must_run_delayed_refs) {
916                 btrfs_async_run_delayed_refs(info, cur, transid,
917                                              must_run_delayed_refs == 1);
918         }
919         return err;
920 }
921
922 int btrfs_end_transaction(struct btrfs_trans_handle *trans)
923 {
924         return __btrfs_end_transaction(trans, 0);
925 }
926
927 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans)
928 {
929         return __btrfs_end_transaction(trans, 1);
930 }
931
932 /*
933  * when btree blocks are allocated, they have some corresponding bits set for
934  * them in one of two extent_io trees.  This is used to make sure all of
935  * those extents are sent to disk but does not wait on them
936  */
937 int btrfs_write_marked_extents(struct btrfs_fs_info *fs_info,
938                                struct extent_io_tree *dirty_pages, int mark)
939 {
940         int err = 0;
941         int werr = 0;
942         struct address_space *mapping = fs_info->btree_inode->i_mapping;
943         struct extent_state *cached_state = NULL;
944         u64 start = 0;
945         u64 end;
946
947         atomic_inc(&BTRFS_I(fs_info->btree_inode)->sync_writers);
948         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
949                                       mark, &cached_state)) {
950                 bool wait_writeback = false;
951
952                 err = convert_extent_bit(dirty_pages, start, end,
953                                          EXTENT_NEED_WAIT,
954                                          mark, &cached_state);
955                 /*
956                  * convert_extent_bit can return -ENOMEM, which is most of the
957                  * time a temporary error. So when it happens, ignore the error
958                  * and wait for writeback of this range to finish - because we
959                  * failed to set the bit EXTENT_NEED_WAIT for the range, a call
960                  * to __btrfs_wait_marked_extents() would not know that
961                  * writeback for this range started and therefore wouldn't
962                  * wait for it to finish - we don't want to commit a
963                  * superblock that points to btree nodes/leafs for which
964                  * writeback hasn't finished yet (and without errors).
965                  * We cleanup any entries left in the io tree when committing
966                  * the transaction (through clear_btree_io_tree()).
967                  */
968                 if (err == -ENOMEM) {
969                         err = 0;
970                         wait_writeback = true;
971                 }
972                 if (!err)
973                         err = filemap_fdatawrite_range(mapping, start, end);
974                 if (err)
975                         werr = err;
976                 else if (wait_writeback)
977                         werr = filemap_fdatawait_range(mapping, start, end);
978                 free_extent_state(cached_state);
979                 cached_state = NULL;
980                 cond_resched();
981                 start = end + 1;
982         }
983         atomic_dec(&BTRFS_I(fs_info->btree_inode)->sync_writers);
984         return werr;
985 }
986
987 /*
988  * when btree blocks are allocated, they have some corresponding bits set for
989  * them in one of two extent_io trees.  This is used to make sure all of
990  * those extents are on disk for transaction or log commit.  We wait
991  * on all the pages and clear them from the dirty pages state tree
992  */
993 static int __btrfs_wait_marked_extents(struct btrfs_fs_info *fs_info,
994                                        struct extent_io_tree *dirty_pages)
995 {
996         int err = 0;
997         int werr = 0;
998         struct address_space *mapping = fs_info->btree_inode->i_mapping;
999         struct extent_state *cached_state = NULL;
1000         u64 start = 0;
1001         u64 end;
1002
1003         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
1004                                       EXTENT_NEED_WAIT, &cached_state)) {
1005                 /*
1006                  * Ignore -ENOMEM errors returned by clear_extent_bit().
1007                  * When committing the transaction, we'll remove any entries
1008                  * left in the io tree. For a log commit, we don't remove them
1009                  * after committing the log because the tree can be accessed
1010                  * concurrently - we do it only at transaction commit time when
1011                  * it's safe to do it (through clear_btree_io_tree()).
1012                  */
1013                 err = clear_extent_bit(dirty_pages, start, end,
1014                                        EXTENT_NEED_WAIT, 0, 0, &cached_state);
1015                 if (err == -ENOMEM)
1016                         err = 0;
1017                 if (!err)
1018                         err = filemap_fdatawait_range(mapping, start, end);
1019                 if (err)
1020                         werr = err;
1021                 free_extent_state(cached_state);
1022                 cached_state = NULL;
1023                 cond_resched();
1024                 start = end + 1;
1025         }
1026         if (err)
1027                 werr = err;
1028         return werr;
1029 }
1030
1031 int btrfs_wait_extents(struct btrfs_fs_info *fs_info,
1032                        struct extent_io_tree *dirty_pages)
1033 {
1034         bool errors = false;
1035         int err;
1036
1037         err = __btrfs_wait_marked_extents(fs_info, dirty_pages);
1038         if (test_and_clear_bit(BTRFS_FS_BTREE_ERR, &fs_info->flags))
1039                 errors = true;
1040
1041         if (errors && !err)
1042                 err = -EIO;
1043         return err;
1044 }
1045
1046 int btrfs_wait_tree_log_extents(struct btrfs_root *log_root, int mark)
1047 {
1048         struct btrfs_fs_info *fs_info = log_root->fs_info;
1049         struct extent_io_tree *dirty_pages = &log_root->dirty_log_pages;
1050         bool errors = false;
1051         int err;
1052
1053         ASSERT(log_root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
1054
1055         err = __btrfs_wait_marked_extents(fs_info, dirty_pages);
1056         if ((mark & EXTENT_DIRTY) &&
1057             test_and_clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags))
1058                 errors = true;
1059
1060         if ((mark & EXTENT_NEW) &&
1061             test_and_clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags))
1062                 errors = true;
1063
1064         if (errors && !err)
1065                 err = -EIO;
1066         return err;
1067 }
1068
1069 /*
1070  * When btree blocks are allocated the corresponding extents are marked dirty.
1071  * This function ensures such extents are persisted on disk for transaction or
1072  * log commit.
1073  *
1074  * @trans: transaction whose dirty pages we'd like to write
1075  */
1076 static int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans)
1077 {
1078         int ret;
1079         int ret2;
1080         struct extent_io_tree *dirty_pages = &trans->transaction->dirty_pages;
1081         struct btrfs_fs_info *fs_info = trans->fs_info;
1082         struct blk_plug plug;
1083
1084         blk_start_plug(&plug);
1085         ret = btrfs_write_marked_extents(fs_info, dirty_pages, EXTENT_DIRTY);
1086         blk_finish_plug(&plug);
1087         ret2 = btrfs_wait_extents(fs_info, dirty_pages);
1088
1089         clear_btree_io_tree(&trans->transaction->dirty_pages);
1090
1091         if (ret)
1092                 return ret;
1093         else if (ret2)
1094                 return ret2;
1095         else
1096                 return 0;
1097 }
1098
1099 /*
1100  * this is used to update the root pointer in the tree of tree roots.
1101  *
1102  * But, in the case of the extent allocation tree, updating the root
1103  * pointer may allocate blocks which may change the root of the extent
1104  * allocation tree.
1105  *
1106  * So, this loops and repeats and makes sure the cowonly root didn't
1107  * change while the root pointer was being updated in the metadata.
1108  */
1109 static int update_cowonly_root(struct btrfs_trans_handle *trans,
1110                                struct btrfs_root *root)
1111 {
1112         int ret;
1113         u64 old_root_bytenr;
1114         u64 old_root_used;
1115         struct btrfs_fs_info *fs_info = root->fs_info;
1116         struct btrfs_root *tree_root = fs_info->tree_root;
1117
1118         old_root_used = btrfs_root_used(&root->root_item);
1119
1120         while (1) {
1121                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
1122                 if (old_root_bytenr == root->node->start &&
1123                     old_root_used == btrfs_root_used(&root->root_item))
1124                         break;
1125
1126                 btrfs_set_root_node(&root->root_item, root->node);
1127                 ret = btrfs_update_root(trans, tree_root,
1128                                         &root->root_key,
1129                                         &root->root_item);
1130                 if (ret)
1131                         return ret;
1132
1133                 old_root_used = btrfs_root_used(&root->root_item);
1134         }
1135
1136         return 0;
1137 }
1138
1139 /*
1140  * update all the cowonly tree roots on disk
1141  *
1142  * The error handling in this function may not be obvious. Any of the
1143  * failures will cause the file system to go offline. We still need
1144  * to clean up the delayed refs.
1145  */
1146 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans)
1147 {
1148         struct btrfs_fs_info *fs_info = trans->fs_info;
1149         struct list_head *dirty_bgs = &trans->transaction->dirty_bgs;
1150         struct list_head *io_bgs = &trans->transaction->io_bgs;
1151         struct list_head *next;
1152         struct extent_buffer *eb;
1153         int ret;
1154
1155         eb = btrfs_lock_root_node(fs_info->tree_root);
1156         ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL,
1157                               0, &eb);
1158         btrfs_tree_unlock(eb);
1159         free_extent_buffer(eb);
1160
1161         if (ret)
1162                 return ret;
1163
1164         ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1165         if (ret)
1166                 return ret;
1167
1168         ret = btrfs_run_dev_stats(trans, fs_info);
1169         if (ret)
1170                 return ret;
1171         ret = btrfs_run_dev_replace(trans, fs_info);
1172         if (ret)
1173                 return ret;
1174         ret = btrfs_run_qgroups(trans, fs_info);
1175         if (ret)
1176                 return ret;
1177
1178         ret = btrfs_setup_space_cache(trans, fs_info);
1179         if (ret)
1180                 return ret;
1181
1182         /* run_qgroups might have added some more refs */
1183         ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1184         if (ret)
1185                 return ret;
1186 again:
1187         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
1188                 struct btrfs_root *root;
1189                 next = fs_info->dirty_cowonly_roots.next;
1190                 list_del_init(next);
1191                 root = list_entry(next, struct btrfs_root, dirty_list);
1192                 clear_bit(BTRFS_ROOT_DIRTY, &root->state);
1193
1194                 if (root != fs_info->extent_root)
1195                         list_add_tail(&root->dirty_list,
1196                                       &trans->transaction->switch_commits);
1197                 ret = update_cowonly_root(trans, root);
1198                 if (ret)
1199                         return ret;
1200                 ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1201                 if (ret)
1202                         return ret;
1203         }
1204
1205         while (!list_empty(dirty_bgs) || !list_empty(io_bgs)) {
1206                 ret = btrfs_write_dirty_block_groups(trans, fs_info);
1207                 if (ret)
1208                         return ret;
1209                 ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1210                 if (ret)
1211                         return ret;
1212         }
1213
1214         if (!list_empty(&fs_info->dirty_cowonly_roots))
1215                 goto again;
1216
1217         list_add_tail(&fs_info->extent_root->dirty_list,
1218                       &trans->transaction->switch_commits);
1219         btrfs_after_dev_replace_commit(fs_info);
1220
1221         return 0;
1222 }
1223
1224 /*
1225  * dead roots are old snapshots that need to be deleted.  This allocates
1226  * a dirty root struct and adds it into the list of dead roots that need to
1227  * be deleted
1228  */
1229 void btrfs_add_dead_root(struct btrfs_root *root)
1230 {
1231         struct btrfs_fs_info *fs_info = root->fs_info;
1232
1233         spin_lock(&fs_info->trans_lock);
1234         if (list_empty(&root->root_list))
1235                 list_add_tail(&root->root_list, &fs_info->dead_roots);
1236         spin_unlock(&fs_info->trans_lock);
1237 }
1238
1239 /*
1240  * update all the cowonly tree roots on disk
1241  */
1242 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans)
1243 {
1244         struct btrfs_fs_info *fs_info = trans->fs_info;
1245         struct btrfs_root *gang[8];
1246         int i;
1247         int ret;
1248         int err = 0;
1249
1250         spin_lock(&fs_info->fs_roots_radix_lock);
1251         while (1) {
1252                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
1253                                                  (void **)gang, 0,
1254                                                  ARRAY_SIZE(gang),
1255                                                  BTRFS_ROOT_TRANS_TAG);
1256                 if (ret == 0)
1257                         break;
1258                 for (i = 0; i < ret; i++) {
1259                         struct btrfs_root *root = gang[i];
1260                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
1261                                         (unsigned long)root->root_key.objectid,
1262                                         BTRFS_ROOT_TRANS_TAG);
1263                         spin_unlock(&fs_info->fs_roots_radix_lock);
1264
1265                         btrfs_free_log(trans, root);
1266                         btrfs_update_reloc_root(trans, root);
1267                         btrfs_orphan_commit_root(trans, root);
1268
1269                         btrfs_save_ino_cache(root, trans);
1270
1271                         /* see comments in should_cow_block() */
1272                         clear_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1273                         smp_mb__after_atomic();
1274
1275                         if (root->commit_root != root->node) {
1276                                 list_add_tail(&root->dirty_list,
1277                                         &trans->transaction->switch_commits);
1278                                 btrfs_set_root_node(&root->root_item,
1279                                                     root->node);
1280                         }
1281
1282                         err = btrfs_update_root(trans, fs_info->tree_root,
1283                                                 &root->root_key,
1284                                                 &root->root_item);
1285                         spin_lock(&fs_info->fs_roots_radix_lock);
1286                         if (err)
1287                                 break;
1288                         btrfs_qgroup_free_meta_all(root);
1289                 }
1290         }
1291         spin_unlock(&fs_info->fs_roots_radix_lock);
1292         return err;
1293 }
1294
1295 /*
1296  * defrag a given btree.
1297  * Every leaf in the btree is read and defragged.
1298  */
1299 int btrfs_defrag_root(struct btrfs_root *root)
1300 {
1301         struct btrfs_fs_info *info = root->fs_info;
1302         struct btrfs_trans_handle *trans;
1303         int ret;
1304
1305         if (test_and_set_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state))
1306                 return 0;
1307
1308         while (1) {
1309                 trans = btrfs_start_transaction(root, 0);
1310                 if (IS_ERR(trans))
1311                         return PTR_ERR(trans);
1312
1313                 ret = btrfs_defrag_leaves(trans, root);
1314
1315                 btrfs_end_transaction(trans);
1316                 btrfs_btree_balance_dirty(info);
1317                 cond_resched();
1318
1319                 if (btrfs_fs_closing(info) || ret != -EAGAIN)
1320                         break;
1321
1322                 if (btrfs_defrag_cancelled(info)) {
1323                         btrfs_debug(info, "defrag_root cancelled");
1324                         ret = -EAGAIN;
1325                         break;
1326                 }
1327         }
1328         clear_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state);
1329         return ret;
1330 }
1331
1332 /*
1333  * Do all special snapshot related qgroup dirty hack.
1334  *
1335  * Will do all needed qgroup inherit and dirty hack like switch commit
1336  * roots inside one transaction and write all btree into disk, to make
1337  * qgroup works.
1338  */
1339 static int qgroup_account_snapshot(struct btrfs_trans_handle *trans,
1340                                    struct btrfs_root *src,
1341                                    struct btrfs_root *parent,
1342                                    struct btrfs_qgroup_inherit *inherit,
1343                                    u64 dst_objectid)
1344 {
1345         struct btrfs_fs_info *fs_info = src->fs_info;
1346         int ret;
1347
1348         /*
1349          * Save some performance in the case that qgroups are not
1350          * enabled. If this check races with the ioctl, rescan will
1351          * kick in anyway.
1352          */
1353         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
1354                 return 0;
1355
1356         /*
1357          * Ensure dirty @src will be commited.  Or, after comming
1358          * commit_fs_roots() and switch_commit_roots(), any dirty but not
1359          * recorded root will never be updated again, causing an outdated root
1360          * item.
1361          */
1362         record_root_in_trans(trans, src, 1);
1363
1364         /*
1365          * We are going to commit transaction, see btrfs_commit_transaction()
1366          * comment for reason locking tree_log_mutex
1367          */
1368         mutex_lock(&fs_info->tree_log_mutex);
1369
1370         ret = commit_fs_roots(trans);
1371         if (ret)
1372                 goto out;
1373         ret = btrfs_qgroup_account_extents(trans, fs_info);
1374         if (ret < 0)
1375                 goto out;
1376
1377         /* Now qgroup are all updated, we can inherit it to new qgroups */
1378         ret = btrfs_qgroup_inherit(trans, fs_info,
1379                                    src->root_key.objectid, dst_objectid,
1380                                    inherit);
1381         if (ret < 0)
1382                 goto out;
1383
1384         /*
1385          * Now we do a simplified commit transaction, which will:
1386          * 1) commit all subvolume and extent tree
1387          *    To ensure all subvolume and extent tree have a valid
1388          *    commit_root to accounting later insert_dir_item()
1389          * 2) write all btree blocks onto disk
1390          *    This is to make sure later btree modification will be cowed
1391          *    Or commit_root can be populated and cause wrong qgroup numbers
1392          * In this simplified commit, we don't really care about other trees
1393          * like chunk and root tree, as they won't affect qgroup.
1394          * And we don't write super to avoid half committed status.
1395          */
1396         ret = commit_cowonly_roots(trans);
1397         if (ret)
1398                 goto out;
1399         switch_commit_roots(trans->transaction);
1400         ret = btrfs_write_and_wait_transaction(trans);
1401         if (ret)
1402                 btrfs_handle_fs_error(fs_info, ret,
1403                         "Error while writing out transaction for qgroup");
1404
1405 out:
1406         mutex_unlock(&fs_info->tree_log_mutex);
1407
1408         /*
1409          * Force parent root to be updated, as we recorded it before so its
1410          * last_trans == cur_transid.
1411          * Or it won't be committed again onto disk after later
1412          * insert_dir_item()
1413          */
1414         if (!ret)
1415                 record_root_in_trans(trans, parent, 1);
1416         return ret;
1417 }
1418
1419 /*
1420  * new snapshots need to be created at a very specific time in the
1421  * transaction commit.  This does the actual creation.
1422  *
1423  * Note:
1424  * If the error which may affect the commitment of the current transaction
1425  * happens, we should return the error number. If the error which just affect
1426  * the creation of the pending snapshots, just return 0.
1427  */
1428 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
1429                                    struct btrfs_pending_snapshot *pending)
1430 {
1431
1432         struct btrfs_fs_info *fs_info = trans->fs_info;
1433         struct btrfs_key key;
1434         struct btrfs_root_item *new_root_item;
1435         struct btrfs_root *tree_root = fs_info->tree_root;
1436         struct btrfs_root *root = pending->root;
1437         struct btrfs_root *parent_root;
1438         struct btrfs_block_rsv *rsv;
1439         struct inode *parent_inode;
1440         struct btrfs_path *path;
1441         struct btrfs_dir_item *dir_item;
1442         struct dentry *dentry;
1443         struct extent_buffer *tmp;
1444         struct extent_buffer *old;
1445         struct timespec cur_time;
1446         int ret = 0;
1447         u64 to_reserve = 0;
1448         u64 index = 0;
1449         u64 objectid;
1450         u64 root_flags;
1451         uuid_le new_uuid;
1452
1453         ASSERT(pending->path);
1454         path = pending->path;
1455
1456         ASSERT(pending->root_item);
1457         new_root_item = pending->root_item;
1458
1459         pending->error = btrfs_find_free_objectid(tree_root, &objectid);
1460         if (pending->error)
1461                 goto no_free_objectid;
1462
1463         /*
1464          * Make qgroup to skip current new snapshot's qgroupid, as it is
1465          * accounted by later btrfs_qgroup_inherit().
1466          */
1467         btrfs_set_skip_qgroup(trans, objectid);
1468
1469         btrfs_reloc_pre_snapshot(pending, &to_reserve);
1470
1471         if (to_reserve > 0) {
1472                 pending->error = btrfs_block_rsv_add(root,
1473                                                      &pending->block_rsv,
1474                                                      to_reserve,
1475                                                      BTRFS_RESERVE_NO_FLUSH);
1476                 if (pending->error)
1477                         goto clear_skip_qgroup;
1478         }
1479
1480         key.objectid = objectid;
1481         key.offset = (u64)-1;
1482         key.type = BTRFS_ROOT_ITEM_KEY;
1483
1484         rsv = trans->block_rsv;
1485         trans->block_rsv = &pending->block_rsv;
1486         trans->bytes_reserved = trans->block_rsv->reserved;
1487         trace_btrfs_space_reservation(fs_info, "transaction",
1488                                       trans->transid,
1489                                       trans->bytes_reserved, 1);
1490         dentry = pending->dentry;
1491         parent_inode = pending->dir;
1492         parent_root = BTRFS_I(parent_inode)->root;
1493         record_root_in_trans(trans, parent_root, 0);
1494
1495         cur_time = current_time(parent_inode);
1496
1497         /*
1498          * insert the directory item
1499          */
1500         ret = btrfs_set_inode_index(BTRFS_I(parent_inode), &index);
1501         BUG_ON(ret); /* -ENOMEM */
1502
1503         /* check if there is a file/dir which has the same name. */
1504         dir_item = btrfs_lookup_dir_item(NULL, parent_root, path,
1505                                          btrfs_ino(BTRFS_I(parent_inode)),
1506                                          dentry->d_name.name,
1507                                          dentry->d_name.len, 0);
1508         if (dir_item != NULL && !IS_ERR(dir_item)) {
1509                 pending->error = -EEXIST;
1510                 goto dir_item_existed;
1511         } else if (IS_ERR(dir_item)) {
1512                 ret = PTR_ERR(dir_item);
1513                 btrfs_abort_transaction(trans, ret);
1514                 goto fail;
1515         }
1516         btrfs_release_path(path);
1517
1518         /*
1519          * pull in the delayed directory update
1520          * and the delayed inode item
1521          * otherwise we corrupt the FS during
1522          * snapshot
1523          */
1524         ret = btrfs_run_delayed_items(trans);
1525         if (ret) {      /* Transaction aborted */
1526                 btrfs_abort_transaction(trans, ret);
1527                 goto fail;
1528         }
1529
1530         record_root_in_trans(trans, root, 0);
1531         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
1532         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
1533         btrfs_check_and_init_root_item(new_root_item);
1534
1535         root_flags = btrfs_root_flags(new_root_item);
1536         if (pending->readonly)
1537                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
1538         else
1539                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
1540         btrfs_set_root_flags(new_root_item, root_flags);
1541
1542         btrfs_set_root_generation_v2(new_root_item,
1543                         trans->transid);
1544         uuid_le_gen(&new_uuid);
1545         memcpy(new_root_item->uuid, new_uuid.b, BTRFS_UUID_SIZE);
1546         memcpy(new_root_item->parent_uuid, root->root_item.uuid,
1547                         BTRFS_UUID_SIZE);
1548         if (!(root_flags & BTRFS_ROOT_SUBVOL_RDONLY)) {
1549                 memset(new_root_item->received_uuid, 0,
1550                        sizeof(new_root_item->received_uuid));
1551                 memset(&new_root_item->stime, 0, sizeof(new_root_item->stime));
1552                 memset(&new_root_item->rtime, 0, sizeof(new_root_item->rtime));
1553                 btrfs_set_root_stransid(new_root_item, 0);
1554                 btrfs_set_root_rtransid(new_root_item, 0);
1555         }
1556         btrfs_set_stack_timespec_sec(&new_root_item->otime, cur_time.tv_sec);
1557         btrfs_set_stack_timespec_nsec(&new_root_item->otime, cur_time.tv_nsec);
1558         btrfs_set_root_otransid(new_root_item, trans->transid);
1559
1560         old = btrfs_lock_root_node(root);
1561         ret = btrfs_cow_block(trans, root, old, NULL, 0, &old);
1562         if (ret) {
1563                 btrfs_tree_unlock(old);
1564                 free_extent_buffer(old);
1565                 btrfs_abort_transaction(trans, ret);
1566                 goto fail;
1567         }
1568
1569         btrfs_set_lock_blocking(old);
1570
1571         ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1572         /* clean up in any case */
1573         btrfs_tree_unlock(old);
1574         free_extent_buffer(old);
1575         if (ret) {
1576                 btrfs_abort_transaction(trans, ret);
1577                 goto fail;
1578         }
1579         /* see comments in should_cow_block() */
1580         set_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1581         smp_wmb();
1582
1583         btrfs_set_root_node(new_root_item, tmp);
1584         /* record when the snapshot was created in key.offset */
1585         key.offset = trans->transid;
1586         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1587         btrfs_tree_unlock(tmp);
1588         free_extent_buffer(tmp);
1589         if (ret) {
1590                 btrfs_abort_transaction(trans, ret);
1591                 goto fail;
1592         }
1593
1594         /*
1595          * insert root back/forward references
1596          */
1597         ret = btrfs_add_root_ref(trans, fs_info, objectid,
1598                                  parent_root->root_key.objectid,
1599                                  btrfs_ino(BTRFS_I(parent_inode)), index,
1600                                  dentry->d_name.name, dentry->d_name.len);
1601         if (ret) {
1602                 btrfs_abort_transaction(trans, ret);
1603                 goto fail;
1604         }
1605
1606         key.offset = (u64)-1;
1607         pending->snap = btrfs_read_fs_root_no_name(fs_info, &key);
1608         if (IS_ERR(pending->snap)) {
1609                 ret = PTR_ERR(pending->snap);
1610                 btrfs_abort_transaction(trans, ret);
1611                 goto fail;
1612         }
1613
1614         ret = btrfs_reloc_post_snapshot(trans, pending);
1615         if (ret) {
1616                 btrfs_abort_transaction(trans, ret);
1617                 goto fail;
1618         }
1619
1620         ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1621         if (ret) {
1622                 btrfs_abort_transaction(trans, ret);
1623                 goto fail;
1624         }
1625
1626         /*
1627          * Do special qgroup accounting for snapshot, as we do some qgroup
1628          * snapshot hack to do fast snapshot.
1629          * To co-operate with that hack, we do hack again.
1630          * Or snapshot will be greatly slowed down by a subtree qgroup rescan
1631          */
1632         ret = qgroup_account_snapshot(trans, root, parent_root,
1633                                       pending->inherit, objectid);
1634         if (ret < 0)
1635                 goto fail;
1636
1637         ret = btrfs_insert_dir_item(trans, parent_root,
1638                                     dentry->d_name.name, dentry->d_name.len,
1639                                     BTRFS_I(parent_inode), &key,
1640                                     BTRFS_FT_DIR, index);
1641         /* We have check then name at the beginning, so it is impossible. */
1642         BUG_ON(ret == -EEXIST || ret == -EOVERFLOW);
1643         if (ret) {
1644                 btrfs_abort_transaction(trans, ret);
1645                 goto fail;
1646         }
1647
1648         btrfs_i_size_write(BTRFS_I(parent_inode), parent_inode->i_size +
1649                                          dentry->d_name.len * 2);
1650         parent_inode->i_mtime = parent_inode->i_ctime =
1651                 current_time(parent_inode);
1652         ret = btrfs_update_inode_fallback(trans, parent_root, parent_inode);
1653         if (ret) {
1654                 btrfs_abort_transaction(trans, ret);
1655                 goto fail;
1656         }
1657         ret = btrfs_uuid_tree_add(trans, fs_info, new_uuid.b,
1658                                   BTRFS_UUID_KEY_SUBVOL, objectid);
1659         if (ret) {
1660                 btrfs_abort_transaction(trans, ret);
1661                 goto fail;
1662         }
1663         if (!btrfs_is_empty_uuid(new_root_item->received_uuid)) {
1664                 ret = btrfs_uuid_tree_add(trans, fs_info,
1665                                           new_root_item->received_uuid,
1666                                           BTRFS_UUID_KEY_RECEIVED_SUBVOL,
1667                                           objectid);
1668                 if (ret && ret != -EEXIST) {
1669                         btrfs_abort_transaction(trans, ret);
1670                         goto fail;
1671                 }
1672         }
1673
1674         ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1675         if (ret) {
1676                 btrfs_abort_transaction(trans, ret);
1677                 goto fail;
1678         }
1679
1680 fail:
1681         pending->error = ret;
1682 dir_item_existed:
1683         trans->block_rsv = rsv;
1684         trans->bytes_reserved = 0;
1685 clear_skip_qgroup:
1686         btrfs_clear_skip_qgroup(trans);
1687 no_free_objectid:
1688         kfree(new_root_item);
1689         pending->root_item = NULL;
1690         btrfs_free_path(path);
1691         pending->path = NULL;
1692
1693         return ret;
1694 }
1695
1696 /*
1697  * create all the snapshots we've scheduled for creation
1698  */
1699 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans)
1700 {
1701         struct btrfs_pending_snapshot *pending, *next;
1702         struct list_head *head = &trans->transaction->pending_snapshots;
1703         int ret = 0;
1704
1705         list_for_each_entry_safe(pending, next, head, list) {
1706                 list_del(&pending->list);
1707                 ret = create_pending_snapshot(trans, pending);
1708                 if (ret)
1709                         break;
1710         }
1711         return ret;
1712 }
1713
1714 static void update_super_roots(struct btrfs_fs_info *fs_info)
1715 {
1716         struct btrfs_root_item *root_item;
1717         struct btrfs_super_block *super;
1718
1719         super = fs_info->super_copy;
1720
1721         root_item = &fs_info->chunk_root->root_item;
1722         super->chunk_root = root_item->bytenr;
1723         super->chunk_root_generation = root_item->generation;
1724         super->chunk_root_level = root_item->level;
1725
1726         root_item = &fs_info->tree_root->root_item;
1727         super->root = root_item->bytenr;
1728         super->generation = root_item->generation;
1729         super->root_level = root_item->level;
1730         if (btrfs_test_opt(fs_info, SPACE_CACHE))
1731                 super->cache_generation = root_item->generation;
1732         if (test_bit(BTRFS_FS_UPDATE_UUID_TREE_GEN, &fs_info->flags))
1733                 super->uuid_tree_generation = root_item->generation;
1734 }
1735
1736 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1737 {
1738         struct btrfs_transaction *trans;
1739         int ret = 0;
1740
1741         spin_lock(&info->trans_lock);
1742         trans = info->running_transaction;
1743         if (trans)
1744                 ret = (trans->state >= TRANS_STATE_COMMIT_START);
1745         spin_unlock(&info->trans_lock);
1746         return ret;
1747 }
1748
1749 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1750 {
1751         struct btrfs_transaction *trans;
1752         int ret = 0;
1753
1754         spin_lock(&info->trans_lock);
1755         trans = info->running_transaction;
1756         if (trans)
1757                 ret = is_transaction_blocked(trans);
1758         spin_unlock(&info->trans_lock);
1759         return ret;
1760 }
1761
1762 /*
1763  * wait for the current transaction commit to start and block subsequent
1764  * transaction joins
1765  */
1766 static void wait_current_trans_commit_start(struct btrfs_fs_info *fs_info,
1767                                             struct btrfs_transaction *trans)
1768 {
1769         wait_event(fs_info->transaction_blocked_wait,
1770                    trans->state >= TRANS_STATE_COMMIT_START || trans->aborted);
1771 }
1772
1773 /*
1774  * wait for the current transaction to start and then become unblocked.
1775  * caller holds ref.
1776  */
1777 static void wait_current_trans_commit_start_and_unblock(
1778                                         struct btrfs_fs_info *fs_info,
1779                                         struct btrfs_transaction *trans)
1780 {
1781         wait_event(fs_info->transaction_wait,
1782                    trans->state >= TRANS_STATE_UNBLOCKED || trans->aborted);
1783 }
1784
1785 /*
1786  * commit transactions asynchronously. once btrfs_commit_transaction_async
1787  * returns, any subsequent transaction will not be allowed to join.
1788  */
1789 struct btrfs_async_commit {
1790         struct btrfs_trans_handle *newtrans;
1791         struct work_struct work;
1792 };
1793
1794 static void do_async_commit(struct work_struct *work)
1795 {
1796         struct btrfs_async_commit *ac =
1797                 container_of(work, struct btrfs_async_commit, work);
1798
1799         /*
1800          * We've got freeze protection passed with the transaction.
1801          * Tell lockdep about it.
1802          */
1803         if (ac->newtrans->type & __TRANS_FREEZABLE)
1804                 __sb_writers_acquired(ac->newtrans->fs_info->sb, SB_FREEZE_FS);
1805
1806         current->journal_info = ac->newtrans;
1807
1808         btrfs_commit_transaction(ac->newtrans);
1809         kfree(ac);
1810 }
1811
1812 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1813                                    int wait_for_unblock)
1814 {
1815         struct btrfs_fs_info *fs_info = trans->fs_info;
1816         struct btrfs_async_commit *ac;
1817         struct btrfs_transaction *cur_trans;
1818
1819         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1820         if (!ac)
1821                 return -ENOMEM;
1822
1823         INIT_WORK(&ac->work, do_async_commit);
1824         ac->newtrans = btrfs_join_transaction(trans->root);
1825         if (IS_ERR(ac->newtrans)) {
1826                 int err = PTR_ERR(ac->newtrans);
1827                 kfree(ac);
1828                 return err;
1829         }
1830
1831         /* take transaction reference */
1832         cur_trans = trans->transaction;
1833         refcount_inc(&cur_trans->use_count);
1834
1835         btrfs_end_transaction(trans);
1836
1837         /*
1838          * Tell lockdep we've released the freeze rwsem, since the
1839          * async commit thread will be the one to unlock it.
1840          */
1841         if (ac->newtrans->type & __TRANS_FREEZABLE)
1842                 __sb_writers_release(fs_info->sb, SB_FREEZE_FS);
1843
1844         schedule_work(&ac->work);
1845
1846         /* wait for transaction to start and unblock */
1847         if (wait_for_unblock)
1848                 wait_current_trans_commit_start_and_unblock(fs_info, cur_trans);
1849         else
1850                 wait_current_trans_commit_start(fs_info, cur_trans);
1851
1852         if (current->journal_info == trans)
1853                 current->journal_info = NULL;
1854
1855         btrfs_put_transaction(cur_trans);
1856         return 0;
1857 }
1858
1859
1860 static void cleanup_transaction(struct btrfs_trans_handle *trans, int err)
1861 {
1862         struct btrfs_fs_info *fs_info = trans->fs_info;
1863         struct btrfs_transaction *cur_trans = trans->transaction;
1864         DEFINE_WAIT(wait);
1865
1866         WARN_ON(refcount_read(&trans->use_count) > 1);
1867
1868         btrfs_abort_transaction(trans, err);
1869
1870         spin_lock(&fs_info->trans_lock);
1871
1872         /*
1873          * If the transaction is removed from the list, it means this
1874          * transaction has been committed successfully, so it is impossible
1875          * to call the cleanup function.
1876          */
1877         BUG_ON(list_empty(&cur_trans->list));
1878
1879         list_del_init(&cur_trans->list);
1880         if (cur_trans == fs_info->running_transaction) {
1881                 cur_trans->state = TRANS_STATE_COMMIT_DOING;
1882                 spin_unlock(&fs_info->trans_lock);
1883                 wait_event(cur_trans->writer_wait,
1884                            atomic_read(&cur_trans->num_writers) == 1);
1885
1886                 spin_lock(&fs_info->trans_lock);
1887         }
1888         spin_unlock(&fs_info->trans_lock);
1889
1890         btrfs_cleanup_one_transaction(trans->transaction, fs_info);
1891
1892         spin_lock(&fs_info->trans_lock);
1893         if (cur_trans == fs_info->running_transaction)
1894                 fs_info->running_transaction = NULL;
1895         spin_unlock(&fs_info->trans_lock);
1896
1897         if (trans->type & __TRANS_FREEZABLE)
1898                 sb_end_intwrite(fs_info->sb);
1899         btrfs_put_transaction(cur_trans);
1900         btrfs_put_transaction(cur_trans);
1901
1902         trace_btrfs_transaction_commit(trans->root);
1903
1904         if (current->journal_info == trans)
1905                 current->journal_info = NULL;
1906         btrfs_scrub_cancel(fs_info);
1907
1908         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1909 }
1910
1911 static inline int btrfs_start_delalloc_flush(struct btrfs_fs_info *fs_info)
1912 {
1913         /*
1914          * We use writeback_inodes_sb here because if we used
1915          * btrfs_start_delalloc_roots we would deadlock with fs freeze.
1916          * Currently are holding the fs freeze lock, if we do an async flush
1917          * we'll do btrfs_join_transaction() and deadlock because we need to
1918          * wait for the fs freeze lock.  Using the direct flushing we benefit
1919          * from already being in a transaction and our join_transaction doesn't
1920          * have to re-take the fs freeze lock.
1921          */
1922         if (btrfs_test_opt(fs_info, FLUSHONCOMMIT))
1923                 writeback_inodes_sb(fs_info->sb, WB_REASON_SYNC);
1924         return 0;
1925 }
1926
1927 static inline void btrfs_wait_delalloc_flush(struct btrfs_fs_info *fs_info)
1928 {
1929         if (btrfs_test_opt(fs_info, FLUSHONCOMMIT))
1930                 btrfs_wait_ordered_roots(fs_info, U64_MAX, 0, (u64)-1);
1931 }
1932
1933 static inline void
1934 btrfs_wait_pending_ordered(struct btrfs_transaction *cur_trans)
1935 {
1936         wait_event(cur_trans->pending_wait,
1937                    atomic_read(&cur_trans->pending_ordered) == 0);
1938 }
1939
1940 int btrfs_commit_transaction(struct btrfs_trans_handle *trans)
1941 {
1942         struct btrfs_fs_info *fs_info = trans->fs_info;
1943         struct btrfs_transaction *cur_trans = trans->transaction;
1944         struct btrfs_transaction *prev_trans = NULL;
1945         int ret;
1946
1947         /* Stop the commit early if ->aborted is set */
1948         if (unlikely(READ_ONCE(cur_trans->aborted))) {
1949                 ret = cur_trans->aborted;
1950                 btrfs_end_transaction(trans);
1951                 return ret;
1952         }
1953
1954         /* make a pass through all the delayed refs we have so far
1955          * any runnings procs may add more while we are here
1956          */
1957         ret = btrfs_run_delayed_refs(trans, fs_info, 0);
1958         if (ret) {
1959                 btrfs_end_transaction(trans);
1960                 return ret;
1961         }
1962
1963         btrfs_trans_release_metadata(trans);
1964         trans->block_rsv = NULL;
1965
1966         cur_trans = trans->transaction;
1967
1968         /*
1969          * set the flushing flag so procs in this transaction have to
1970          * start sending their work down.
1971          */
1972         cur_trans->delayed_refs.flushing = 1;
1973         smp_wmb();
1974
1975         if (!list_empty(&trans->new_bgs))
1976                 btrfs_create_pending_block_groups(trans);
1977
1978         ret = btrfs_run_delayed_refs(trans, fs_info, 0);
1979         if (ret) {
1980                 btrfs_end_transaction(trans);
1981                 return ret;
1982         }
1983
1984         if (!test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &cur_trans->flags)) {
1985                 int run_it = 0;
1986
1987                 /* this mutex is also taken before trying to set
1988                  * block groups readonly.  We need to make sure
1989                  * that nobody has set a block group readonly
1990                  * after a extents from that block group have been
1991                  * allocated for cache files.  btrfs_set_block_group_ro
1992                  * will wait for the transaction to commit if it
1993                  * finds BTRFS_TRANS_DIRTY_BG_RUN set.
1994                  *
1995                  * The BTRFS_TRANS_DIRTY_BG_RUN flag is also used to make sure
1996                  * only one process starts all the block group IO.  It wouldn't
1997                  * hurt to have more than one go through, but there's no
1998                  * real advantage to it either.
1999                  */
2000                 mutex_lock(&fs_info->ro_block_group_mutex);
2001                 if (!test_and_set_bit(BTRFS_TRANS_DIRTY_BG_RUN,
2002                                       &cur_trans->flags))
2003                         run_it = 1;
2004                 mutex_unlock(&fs_info->ro_block_group_mutex);
2005
2006                 if (run_it) {
2007                         ret = btrfs_start_dirty_block_groups(trans);
2008                         if (ret) {
2009                                 btrfs_end_transaction(trans);
2010                                 return ret;
2011                         }
2012                 }
2013         }
2014
2015         spin_lock(&fs_info->trans_lock);
2016         if (cur_trans->state >= TRANS_STATE_COMMIT_START) {
2017                 spin_unlock(&fs_info->trans_lock);
2018                 refcount_inc(&cur_trans->use_count);
2019                 ret = btrfs_end_transaction(trans);
2020
2021                 wait_for_commit(cur_trans);
2022
2023                 if (unlikely(cur_trans->aborted))
2024                         ret = cur_trans->aborted;
2025
2026                 btrfs_put_transaction(cur_trans);
2027
2028                 return ret;
2029         }
2030
2031         cur_trans->state = TRANS_STATE_COMMIT_START;
2032         wake_up(&fs_info->transaction_blocked_wait);
2033
2034         if (cur_trans->list.prev != &fs_info->trans_list) {
2035                 prev_trans = list_entry(cur_trans->list.prev,
2036                                         struct btrfs_transaction, list);
2037                 if (prev_trans->state != TRANS_STATE_COMPLETED) {
2038                         refcount_inc(&prev_trans->use_count);
2039                         spin_unlock(&fs_info->trans_lock);
2040
2041                         wait_for_commit(prev_trans);
2042                         ret = prev_trans->aborted;
2043
2044                         btrfs_put_transaction(prev_trans);
2045                         if (ret)
2046                                 goto cleanup_transaction;
2047                 } else {
2048                         spin_unlock(&fs_info->trans_lock);
2049                 }
2050         } else {
2051                 spin_unlock(&fs_info->trans_lock);
2052         }
2053
2054         extwriter_counter_dec(cur_trans, trans->type);
2055
2056         ret = btrfs_start_delalloc_flush(fs_info);
2057         if (ret)
2058                 goto cleanup_transaction;
2059
2060         ret = btrfs_run_delayed_items(trans);
2061         if (ret)
2062                 goto cleanup_transaction;
2063
2064         wait_event(cur_trans->writer_wait,
2065                    extwriter_counter_read(cur_trans) == 0);
2066
2067         /* some pending stuffs might be added after the previous flush. */
2068         ret = btrfs_run_delayed_items(trans);
2069         if (ret)
2070                 goto cleanup_transaction;
2071
2072         btrfs_wait_delalloc_flush(fs_info);
2073
2074         btrfs_wait_pending_ordered(cur_trans);
2075
2076         btrfs_scrub_pause(fs_info);
2077         /*
2078          * Ok now we need to make sure to block out any other joins while we
2079          * commit the transaction.  We could have started a join before setting
2080          * COMMIT_DOING so make sure to wait for num_writers to == 1 again.
2081          */
2082         spin_lock(&fs_info->trans_lock);
2083         cur_trans->state = TRANS_STATE_COMMIT_DOING;
2084         spin_unlock(&fs_info->trans_lock);
2085         wait_event(cur_trans->writer_wait,
2086                    atomic_read(&cur_trans->num_writers) == 1);
2087
2088         /* ->aborted might be set after the previous check, so check it */
2089         if (unlikely(READ_ONCE(cur_trans->aborted))) {
2090                 ret = cur_trans->aborted;
2091                 goto scrub_continue;
2092         }
2093         /*
2094          * the reloc mutex makes sure that we stop
2095          * the balancing code from coming in and moving
2096          * extents around in the middle of the commit
2097          */
2098         mutex_lock(&fs_info->reloc_mutex);
2099
2100         /*
2101          * We needn't worry about the delayed items because we will
2102          * deal with them in create_pending_snapshot(), which is the
2103          * core function of the snapshot creation.
2104          */
2105         ret = create_pending_snapshots(trans);
2106         if (ret) {
2107                 mutex_unlock(&fs_info->reloc_mutex);
2108                 goto scrub_continue;
2109         }
2110
2111         /*
2112          * We insert the dir indexes of the snapshots and update the inode
2113          * of the snapshots' parents after the snapshot creation, so there
2114          * are some delayed items which are not dealt with. Now deal with
2115          * them.
2116          *
2117          * We needn't worry that this operation will corrupt the snapshots,
2118          * because all the tree which are snapshoted will be forced to COW
2119          * the nodes and leaves.
2120          */
2121         ret = btrfs_run_delayed_items(trans);
2122         if (ret) {
2123                 mutex_unlock(&fs_info->reloc_mutex);
2124                 goto scrub_continue;
2125         }
2126
2127         ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
2128         if (ret) {
2129                 mutex_unlock(&fs_info->reloc_mutex);
2130                 goto scrub_continue;
2131         }
2132
2133         /*
2134          * make sure none of the code above managed to slip in a
2135          * delayed item
2136          */
2137         btrfs_assert_delayed_root_empty(fs_info);
2138
2139         WARN_ON(cur_trans != trans->transaction);
2140
2141         /* btrfs_commit_tree_roots is responsible for getting the
2142          * various roots consistent with each other.  Every pointer
2143          * in the tree of tree roots has to point to the most up to date
2144          * root for every subvolume and other tree.  So, we have to keep
2145          * the tree logging code from jumping in and changing any
2146          * of the trees.
2147          *
2148          * At this point in the commit, there can't be any tree-log
2149          * writers, but a little lower down we drop the trans mutex
2150          * and let new people in.  By holding the tree_log_mutex
2151          * from now until after the super is written, we avoid races
2152          * with the tree-log code.
2153          */
2154         mutex_lock(&fs_info->tree_log_mutex);
2155
2156         ret = commit_fs_roots(trans);
2157         if (ret) {
2158                 mutex_unlock(&fs_info->tree_log_mutex);
2159                 mutex_unlock(&fs_info->reloc_mutex);
2160                 goto scrub_continue;
2161         }
2162
2163         /*
2164          * Since the transaction is done, we can apply the pending changes
2165          * before the next transaction.
2166          */
2167         btrfs_apply_pending_changes(fs_info);
2168
2169         /* commit_fs_roots gets rid of all the tree log roots, it is now
2170          * safe to free the root of tree log roots
2171          */
2172         btrfs_free_log_root_tree(trans, fs_info);
2173
2174         /*
2175          * commit_fs_roots() can call btrfs_save_ino_cache(), which generates
2176          * new delayed refs. Must handle them or qgroup can be wrong.
2177          */
2178         ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
2179         if (ret) {
2180                 mutex_unlock(&fs_info->tree_log_mutex);
2181                 mutex_unlock(&fs_info->reloc_mutex);
2182                 goto scrub_continue;
2183         }
2184
2185         /*
2186          * Since fs roots are all committed, we can get a quite accurate
2187          * new_roots. So let's do quota accounting.
2188          */
2189         ret = btrfs_qgroup_account_extents(trans, fs_info);
2190         if (ret < 0) {
2191                 mutex_unlock(&fs_info->tree_log_mutex);
2192                 mutex_unlock(&fs_info->reloc_mutex);
2193                 goto scrub_continue;
2194         }
2195
2196         ret = commit_cowonly_roots(trans);
2197         if (ret) {
2198                 mutex_unlock(&fs_info->tree_log_mutex);
2199                 mutex_unlock(&fs_info->reloc_mutex);
2200                 goto scrub_continue;
2201         }
2202
2203         /*
2204          * The tasks which save the space cache and inode cache may also
2205          * update ->aborted, check it.
2206          */
2207         if (unlikely(READ_ONCE(cur_trans->aborted))) {
2208                 ret = cur_trans->aborted;
2209                 mutex_unlock(&fs_info->tree_log_mutex);
2210                 mutex_unlock(&fs_info->reloc_mutex);
2211                 goto scrub_continue;
2212         }
2213
2214         btrfs_prepare_extent_commit(fs_info);
2215
2216         cur_trans = fs_info->running_transaction;
2217
2218         btrfs_set_root_node(&fs_info->tree_root->root_item,
2219                             fs_info->tree_root->node);
2220         list_add_tail(&fs_info->tree_root->dirty_list,
2221                       &cur_trans->switch_commits);
2222
2223         btrfs_set_root_node(&fs_info->chunk_root->root_item,
2224                             fs_info->chunk_root->node);
2225         list_add_tail(&fs_info->chunk_root->dirty_list,
2226                       &cur_trans->switch_commits);
2227
2228         switch_commit_roots(cur_trans);
2229
2230         ASSERT(list_empty(&cur_trans->dirty_bgs));
2231         ASSERT(list_empty(&cur_trans->io_bgs));
2232         update_super_roots(fs_info);
2233
2234         btrfs_set_super_log_root(fs_info->super_copy, 0);
2235         btrfs_set_super_log_root_level(fs_info->super_copy, 0);
2236         memcpy(fs_info->super_for_commit, fs_info->super_copy,
2237                sizeof(*fs_info->super_copy));
2238
2239         btrfs_update_commit_device_size(fs_info);
2240         btrfs_update_commit_device_bytes_used(cur_trans);
2241
2242         clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags);
2243         clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags);
2244
2245         btrfs_trans_release_chunk_metadata(trans);
2246
2247         spin_lock(&fs_info->trans_lock);
2248         cur_trans->state = TRANS_STATE_UNBLOCKED;
2249         fs_info->running_transaction = NULL;
2250         spin_unlock(&fs_info->trans_lock);
2251         mutex_unlock(&fs_info->reloc_mutex);
2252
2253         wake_up(&fs_info->transaction_wait);
2254
2255         ret = btrfs_write_and_wait_transaction(trans);
2256         if (ret) {
2257                 btrfs_handle_fs_error(fs_info, ret,
2258                                       "Error while writing out transaction");
2259                 mutex_unlock(&fs_info->tree_log_mutex);
2260                 goto scrub_continue;
2261         }
2262
2263         ret = write_all_supers(fs_info, 0);
2264         /*
2265          * the super is written, we can safely allow the tree-loggers
2266          * to go about their business
2267          */
2268         mutex_unlock(&fs_info->tree_log_mutex);
2269         if (ret)
2270                 goto scrub_continue;
2271
2272         btrfs_finish_extent_commit(trans, fs_info);
2273
2274         if (test_bit(BTRFS_TRANS_HAVE_FREE_BGS, &cur_trans->flags))
2275                 btrfs_clear_space_info_full(fs_info);
2276
2277         fs_info->last_trans_committed = cur_trans->transid;
2278         /*
2279          * We needn't acquire the lock here because there is no other task
2280          * which can change it.
2281          */
2282         cur_trans->state = TRANS_STATE_COMPLETED;
2283         wake_up(&cur_trans->commit_wait);
2284
2285         spin_lock(&fs_info->trans_lock);
2286         list_del_init(&cur_trans->list);
2287         spin_unlock(&fs_info->trans_lock);
2288
2289         btrfs_put_transaction(cur_trans);
2290         btrfs_put_transaction(cur_trans);
2291
2292         if (trans->type & __TRANS_FREEZABLE)
2293                 sb_end_intwrite(fs_info->sb);
2294
2295         trace_btrfs_transaction_commit(trans->root);
2296
2297         btrfs_scrub_continue(fs_info);
2298
2299         if (current->journal_info == trans)
2300                 current->journal_info = NULL;
2301
2302         kmem_cache_free(btrfs_trans_handle_cachep, trans);
2303
2304         /*
2305          * If fs has been frozen, we can not handle delayed iputs, otherwise
2306          * it'll result in deadlock about SB_FREEZE_FS.
2307          */
2308         if (current != fs_info->transaction_kthread &&
2309             current != fs_info->cleaner_kthread &&
2310             !test_bit(BTRFS_FS_FROZEN, &fs_info->flags))
2311                 btrfs_run_delayed_iputs(fs_info);
2312
2313         return ret;
2314
2315 scrub_continue:
2316         btrfs_scrub_continue(fs_info);
2317 cleanup_transaction:
2318         btrfs_trans_release_metadata(trans);
2319         btrfs_trans_release_chunk_metadata(trans);
2320         trans->block_rsv = NULL;
2321         btrfs_warn(fs_info, "Skipping commit of aborted transaction.");
2322         if (current->journal_info == trans)
2323                 current->journal_info = NULL;
2324         cleanup_transaction(trans, ret);
2325
2326         return ret;
2327 }
2328
2329 /*
2330  * return < 0 if error
2331  * 0 if there are no more dead_roots at the time of call
2332  * 1 there are more to be processed, call me again
2333  *
2334  * The return value indicates there are certainly more snapshots to delete, but
2335  * if there comes a new one during processing, it may return 0. We don't mind,
2336  * because btrfs_commit_super will poke cleaner thread and it will process it a
2337  * few seconds later.
2338  */
2339 int btrfs_clean_one_deleted_snapshot(struct btrfs_root *root)
2340 {
2341         int ret;
2342         struct btrfs_fs_info *fs_info = root->fs_info;
2343
2344         spin_lock(&fs_info->trans_lock);
2345         if (list_empty(&fs_info->dead_roots)) {
2346                 spin_unlock(&fs_info->trans_lock);
2347                 return 0;
2348         }
2349         root = list_first_entry(&fs_info->dead_roots,
2350                         struct btrfs_root, root_list);
2351         list_del_init(&root->root_list);
2352         spin_unlock(&fs_info->trans_lock);
2353
2354         btrfs_debug(fs_info, "cleaner removing %llu", root->objectid);
2355
2356         btrfs_kill_all_delayed_nodes(root);
2357
2358         if (btrfs_header_backref_rev(root->node) <
2359                         BTRFS_MIXED_BACKREF_REV)
2360                 ret = btrfs_drop_snapshot(root, NULL, 0, 0);
2361         else
2362                 ret = btrfs_drop_snapshot(root, NULL, 1, 0);
2363
2364         return (ret < 0) ? 0 : 1;
2365 }
2366
2367 void btrfs_apply_pending_changes(struct btrfs_fs_info *fs_info)
2368 {
2369         unsigned long prev;
2370         unsigned long bit;
2371
2372         prev = xchg(&fs_info->pending_changes, 0);
2373         if (!prev)
2374                 return;
2375
2376         bit = 1 << BTRFS_PENDING_SET_INODE_MAP_CACHE;
2377         if (prev & bit)
2378                 btrfs_set_opt(fs_info->mount_opt, INODE_MAP_CACHE);
2379         prev &= ~bit;
2380
2381         bit = 1 << BTRFS_PENDING_CLEAR_INODE_MAP_CACHE;
2382         if (prev & bit)
2383                 btrfs_clear_opt(fs_info->mount_opt, INODE_MAP_CACHE);
2384         prev &= ~bit;
2385
2386         bit = 1 << BTRFS_PENDING_COMMIT;
2387         if (prev & bit)
2388                 btrfs_debug(fs_info, "pending commit done");
2389         prev &= ~bit;
2390
2391         if (prev)
2392                 btrfs_warn(fs_info,
2393                         "unknown pending changes left 0x%lx, ignoring", prev);
2394 }