beca2563578722615afeb3a7e287fd4c0b9b969a
[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_USERSPACE |
41                                            __TRANS_START),
42         [TRANS_STATE_COMMIT_START]      = (__TRANS_USERSPACE |
43                                            __TRANS_START |
44                                            __TRANS_ATTACH),
45         [TRANS_STATE_COMMIT_DOING]      = (__TRANS_USERSPACE |
46                                            __TRANS_START |
47                                            __TRANS_ATTACH |
48                                            __TRANS_JOIN),
49         [TRANS_STATE_UNBLOCKED]         = (__TRANS_USERSPACE |
50                                            __TRANS_START |
51                                            __TRANS_ATTACH |
52                                            __TRANS_JOIN |
53                                            __TRANS_JOIN_NOLOCK),
54         [TRANS_STATE_COMPLETED]         = (__TRANS_USERSPACE |
55                                            __TRANS_START |
56                                            __TRANS_ATTACH |
57                                            __TRANS_JOIN |
58                                            __TRANS_JOIN_NOLOCK),
59 };
60
61 void btrfs_put_transaction(struct btrfs_transaction *transaction)
62 {
63         WARN_ON(refcount_read(&transaction->use_count) == 0);
64         if (refcount_dec_and_test(&transaction->use_count)) {
65                 BUG_ON(!list_empty(&transaction->list));
66                 WARN_ON(!RB_EMPTY_ROOT(&transaction->delayed_refs.href_root));
67                 if (transaction->delayed_refs.pending_csums)
68                         btrfs_err(transaction->fs_info,
69                                   "pending csums is %llu",
70                                   transaction->delayed_refs.pending_csums);
71                 while (!list_empty(&transaction->pending_chunks)) {
72                         struct extent_map *em;
73
74                         em = list_first_entry(&transaction->pending_chunks,
75                                               struct extent_map, list);
76                         list_del_init(&em->list);
77                         free_extent_map(em);
78                 }
79                 /*
80                  * If any block groups are found in ->deleted_bgs then it's
81                  * because the transaction was aborted and a commit did not
82                  * happen (things failed before writing the new superblock
83                  * and calling btrfs_finish_extent_commit()), so we can not
84                  * discard the physical locations of the block groups.
85                  */
86                 while (!list_empty(&transaction->deleted_bgs)) {
87                         struct btrfs_block_group_cache *cache;
88
89                         cache = list_first_entry(&transaction->deleted_bgs,
90                                                  struct btrfs_block_group_cache,
91                                                  bg_list);
92                         list_del_init(&cache->bg_list);
93                         btrfs_put_block_group_trimming(cache);
94                         btrfs_put_block_group(cache);
95                 }
96                 kfree(transaction);
97         }
98 }
99
100 static void clear_btree_io_tree(struct extent_io_tree *tree)
101 {
102         spin_lock(&tree->lock);
103         /*
104          * Do a single barrier for the waitqueue_active check here, the state
105          * of the waitqueue should not change once clear_btree_io_tree is
106          * called.
107          */
108         smp_mb();
109         while (!RB_EMPTY_ROOT(&tree->state)) {
110                 struct rb_node *node;
111                 struct extent_state *state;
112
113                 node = rb_first(&tree->state);
114                 state = rb_entry(node, struct extent_state, rb_node);
115                 rb_erase(&state->rb_node, &tree->state);
116                 RB_CLEAR_NODE(&state->rb_node);
117                 /*
118                  * btree io trees aren't supposed to have tasks waiting for
119                  * changes in the flags of extent states ever.
120                  */
121                 ASSERT(!waitqueue_active(&state->wq));
122                 free_extent_state(state);
123
124                 cond_resched_lock(&tree->lock);
125         }
126         spin_unlock(&tree->lock);
127 }
128
129 static noinline void switch_commit_roots(struct btrfs_transaction *trans,
130                                          struct btrfs_fs_info *fs_info)
131 {
132         struct btrfs_root *root, *tmp;
133
134         down_write(&fs_info->commit_root_sem);
135         list_for_each_entry_safe(root, tmp, &trans->switch_commits,
136                                  dirty_list) {
137                 list_del_init(&root->dirty_list);
138                 free_extent_buffer(root->commit_root);
139                 root->commit_root = btrfs_root_node(root);
140                 if (is_fstree(root->objectid))
141                         btrfs_unpin_free_ino(root);
142                 clear_btree_io_tree(&root->dirty_log_pages);
143         }
144
145         /* We can free old roots now. */
146         spin_lock(&trans->dropped_roots_lock);
147         while (!list_empty(&trans->dropped_roots)) {
148                 root = list_first_entry(&trans->dropped_roots,
149                                         struct btrfs_root, root_list);
150                 list_del_init(&root->root_list);
151                 spin_unlock(&trans->dropped_roots_lock);
152                 btrfs_drop_and_free_fs_root(fs_info, root);
153                 spin_lock(&trans->dropped_roots_lock);
154         }
155         spin_unlock(&trans->dropped_roots_lock);
156         up_write(&fs_info->commit_root_sem);
157 }
158
159 static inline void extwriter_counter_inc(struct btrfs_transaction *trans,
160                                          unsigned int type)
161 {
162         if (type & TRANS_EXTWRITERS)
163                 atomic_inc(&trans->num_extwriters);
164 }
165
166 static inline void extwriter_counter_dec(struct btrfs_transaction *trans,
167                                          unsigned int type)
168 {
169         if (type & TRANS_EXTWRITERS)
170                 atomic_dec(&trans->num_extwriters);
171 }
172
173 static inline void extwriter_counter_init(struct btrfs_transaction *trans,
174                                           unsigned int type)
175 {
176         atomic_set(&trans->num_extwriters, ((type & TRANS_EXTWRITERS) ? 1 : 0));
177 }
178
179 static inline int extwriter_counter_read(struct btrfs_transaction *trans)
180 {
181         return atomic_read(&trans->num_extwriters);
182 }
183
184 /*
185  * either allocate a new transaction or hop into the existing one
186  */
187 static noinline int join_transaction(struct btrfs_fs_info *fs_info,
188                                      unsigned int type)
189 {
190         struct btrfs_transaction *cur_trans;
191
192         spin_lock(&fs_info->trans_lock);
193 loop:
194         /* The file system has been taken offline. No new transactions. */
195         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
196                 spin_unlock(&fs_info->trans_lock);
197                 return -EROFS;
198         }
199
200         cur_trans = fs_info->running_transaction;
201         if (cur_trans) {
202                 if (cur_trans->aborted) {
203                         spin_unlock(&fs_info->trans_lock);
204                         return cur_trans->aborted;
205                 }
206                 if (btrfs_blocked_trans_types[cur_trans->state] & type) {
207                         spin_unlock(&fs_info->trans_lock);
208                         return -EBUSY;
209                 }
210                 refcount_inc(&cur_trans->use_count);
211                 atomic_inc(&cur_trans->num_writers);
212                 extwriter_counter_inc(cur_trans, type);
213                 spin_unlock(&fs_info->trans_lock);
214                 return 0;
215         }
216         spin_unlock(&fs_info->trans_lock);
217
218         /*
219          * If we are ATTACH, we just want to catch the current transaction,
220          * and commit it. If there is no transaction, just return ENOENT.
221          */
222         if (type == TRANS_ATTACH)
223                 return -ENOENT;
224
225         /*
226          * JOIN_NOLOCK only happens during the transaction commit, so
227          * it is impossible that ->running_transaction is NULL
228          */
229         BUG_ON(type == TRANS_JOIN_NOLOCK);
230
231         cur_trans = kmalloc(sizeof(*cur_trans), GFP_NOFS);
232         if (!cur_trans)
233                 return -ENOMEM;
234
235         spin_lock(&fs_info->trans_lock);
236         if (fs_info->running_transaction) {
237                 /*
238                  * someone started a transaction after we unlocked.  Make sure
239                  * to redo the checks above
240                  */
241                 kfree(cur_trans);
242                 goto loop;
243         } else if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
244                 spin_unlock(&fs_info->trans_lock);
245                 kfree(cur_trans);
246                 return -EROFS;
247         }
248
249         cur_trans->fs_info = fs_info;
250         atomic_set(&cur_trans->num_writers, 1);
251         extwriter_counter_init(cur_trans, type);
252         init_waitqueue_head(&cur_trans->writer_wait);
253         init_waitqueue_head(&cur_trans->commit_wait);
254         init_waitqueue_head(&cur_trans->pending_wait);
255         cur_trans->state = TRANS_STATE_RUNNING;
256         /*
257          * One for this trans handle, one so it will live on until we
258          * commit the transaction.
259          */
260         refcount_set(&cur_trans->use_count, 2);
261         atomic_set(&cur_trans->pending_ordered, 0);
262         cur_trans->flags = 0;
263         cur_trans->start_time = get_seconds();
264
265         memset(&cur_trans->delayed_refs, 0, sizeof(cur_trans->delayed_refs));
266
267         cur_trans->delayed_refs.href_root = RB_ROOT;
268         cur_trans->delayed_refs.dirty_extent_root = RB_ROOT;
269         atomic_set(&cur_trans->delayed_refs.num_entries, 0);
270
271         /*
272          * although the tree mod log is per file system and not per transaction,
273          * the log must never go across transaction boundaries.
274          */
275         smp_mb();
276         if (!list_empty(&fs_info->tree_mod_seq_list))
277                 WARN(1, KERN_ERR "BTRFS: tree_mod_seq_list not empty when creating a fresh transaction\n");
278         if (!RB_EMPTY_ROOT(&fs_info->tree_mod_log))
279                 WARN(1, KERN_ERR "BTRFS: tree_mod_log rb tree not empty when creating a fresh transaction\n");
280         atomic64_set(&fs_info->tree_mod_seq, 0);
281
282         spin_lock_init(&cur_trans->delayed_refs.lock);
283
284         INIT_LIST_HEAD(&cur_trans->pending_snapshots);
285         INIT_LIST_HEAD(&cur_trans->pending_chunks);
286         INIT_LIST_HEAD(&cur_trans->switch_commits);
287         INIT_LIST_HEAD(&cur_trans->dirty_bgs);
288         INIT_LIST_HEAD(&cur_trans->io_bgs);
289         INIT_LIST_HEAD(&cur_trans->dropped_roots);
290         mutex_init(&cur_trans->cache_write_mutex);
291         cur_trans->num_dirty_bgs = 0;
292         spin_lock_init(&cur_trans->dirty_bgs_lock);
293         INIT_LIST_HEAD(&cur_trans->deleted_bgs);
294         spin_lock_init(&cur_trans->dropped_roots_lock);
295         list_add_tail(&cur_trans->list, &fs_info->trans_list);
296         extent_io_tree_init(&cur_trans->dirty_pages,
297                              fs_info->btree_inode);
298         fs_info->generation++;
299         cur_trans->transid = fs_info->generation;
300         fs_info->running_transaction = cur_trans;
301         cur_trans->aborted = 0;
302         spin_unlock(&fs_info->trans_lock);
303
304         return 0;
305 }
306
307 /*
308  * this does all the record keeping required to make sure that a reference
309  * counted root is properly recorded in a given transaction.  This is required
310  * to make sure the old root from before we joined the transaction is deleted
311  * when the transaction commits
312  */
313 static int record_root_in_trans(struct btrfs_trans_handle *trans,
314                                struct btrfs_root *root,
315                                int force)
316 {
317         struct btrfs_fs_info *fs_info = root->fs_info;
318
319         if ((test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
320             root->last_trans < trans->transid) || force) {
321                 WARN_ON(root == fs_info->extent_root);
322                 WARN_ON(root->commit_root != root->node);
323
324                 /*
325                  * see below for IN_TRANS_SETUP usage rules
326                  * we have the reloc mutex held now, so there
327                  * is only one writer in this function
328                  */
329                 set_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
330
331                 /* make sure readers find IN_TRANS_SETUP before
332                  * they find our root->last_trans update
333                  */
334                 smp_wmb();
335
336                 spin_lock(&fs_info->fs_roots_radix_lock);
337                 if (root->last_trans == trans->transid && !force) {
338                         spin_unlock(&fs_info->fs_roots_radix_lock);
339                         return 0;
340                 }
341                 radix_tree_tag_set(&fs_info->fs_roots_radix,
342                                    (unsigned long)root->root_key.objectid,
343                                    BTRFS_ROOT_TRANS_TAG);
344                 spin_unlock(&fs_info->fs_roots_radix_lock);
345                 root->last_trans = trans->transid;
346
347                 /* this is pretty tricky.  We don't want to
348                  * take the relocation lock in btrfs_record_root_in_trans
349                  * unless we're really doing the first setup for this root in
350                  * this transaction.
351                  *
352                  * Normally we'd use root->last_trans as a flag to decide
353                  * if we want to take the expensive mutex.
354                  *
355                  * But, we have to set root->last_trans before we
356                  * init the relocation root, otherwise, we trip over warnings
357                  * in ctree.c.  The solution used here is to flag ourselves
358                  * with root IN_TRANS_SETUP.  When this is 1, we're still
359                  * fixing up the reloc trees and everyone must wait.
360                  *
361                  * When this is zero, they can trust root->last_trans and fly
362                  * through btrfs_record_root_in_trans without having to take the
363                  * lock.  smp_wmb() makes sure that all the writes above are
364                  * done before we pop in the zero below
365                  */
366                 btrfs_init_reloc_root(trans, root);
367                 smp_mb__before_atomic();
368                 clear_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
369         }
370         return 0;
371 }
372
373
374 void btrfs_add_dropped_root(struct btrfs_trans_handle *trans,
375                             struct btrfs_root *root)
376 {
377         struct btrfs_fs_info *fs_info = root->fs_info;
378         struct btrfs_transaction *cur_trans = trans->transaction;
379
380         /* Add ourselves to the transaction dropped list */
381         spin_lock(&cur_trans->dropped_roots_lock);
382         list_add_tail(&root->root_list, &cur_trans->dropped_roots);
383         spin_unlock(&cur_trans->dropped_roots_lock);
384
385         /* Make sure we don't try to update the root at commit time */
386         spin_lock(&fs_info->fs_roots_radix_lock);
387         radix_tree_tag_clear(&fs_info->fs_roots_radix,
388                              (unsigned long)root->root_key.objectid,
389                              BTRFS_ROOT_TRANS_TAG);
390         spin_unlock(&fs_info->fs_roots_radix_lock);
391 }
392
393 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
394                                struct btrfs_root *root)
395 {
396         struct btrfs_fs_info *fs_info = root->fs_info;
397
398         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
399                 return 0;
400
401         /*
402          * see record_root_in_trans for comments about IN_TRANS_SETUP usage
403          * and barriers
404          */
405         smp_rmb();
406         if (root->last_trans == trans->transid &&
407             !test_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state))
408                 return 0;
409
410         mutex_lock(&fs_info->reloc_mutex);
411         record_root_in_trans(trans, root, 0);
412         mutex_unlock(&fs_info->reloc_mutex);
413
414         return 0;
415 }
416
417 static inline int is_transaction_blocked(struct btrfs_transaction *trans)
418 {
419         return (trans->state >= TRANS_STATE_BLOCKED &&
420                 trans->state < TRANS_STATE_UNBLOCKED &&
421                 !trans->aborted);
422 }
423
424 /* wait for commit against the current transaction to become unblocked
425  * when this is done, it is safe to start a new transaction, but the current
426  * transaction might not be fully on disk.
427  */
428 static void wait_current_trans(struct btrfs_fs_info *fs_info)
429 {
430         struct btrfs_transaction *cur_trans;
431
432         spin_lock(&fs_info->trans_lock);
433         cur_trans = fs_info->running_transaction;
434         if (cur_trans && is_transaction_blocked(cur_trans)) {
435                 refcount_inc(&cur_trans->use_count);
436                 spin_unlock(&fs_info->trans_lock);
437
438                 wait_event(fs_info->transaction_wait,
439                            cur_trans->state >= TRANS_STATE_UNBLOCKED ||
440                            cur_trans->aborted);
441                 btrfs_put_transaction(cur_trans);
442         } else {
443                 spin_unlock(&fs_info->trans_lock);
444         }
445 }
446
447 static int may_wait_transaction(struct btrfs_fs_info *fs_info, int type)
448 {
449         if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
450                 return 0;
451
452         if (type == TRANS_USERSPACE)
453                 return 1;
454
455         if (type == TRANS_START &&
456             !atomic_read(&fs_info->open_ioctl_trans))
457                 return 1;
458
459         return 0;
460 }
461
462 static inline bool need_reserve_reloc_root(struct btrfs_root *root)
463 {
464         struct btrfs_fs_info *fs_info = root->fs_info;
465
466         if (!fs_info->reloc_ctl ||
467             !test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
468             root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
469             root->reloc_root)
470                 return false;
471
472         return true;
473 }
474
475 static struct btrfs_trans_handle *
476 start_transaction(struct btrfs_root *root, unsigned int num_items,
477                   unsigned int type, enum btrfs_reserve_flush_enum flush,
478                   bool enforce_qgroups)
479 {
480         struct btrfs_fs_info *fs_info = root->fs_info;
481
482         struct btrfs_trans_handle *h;
483         struct btrfs_transaction *cur_trans;
484         u64 num_bytes = 0;
485         u64 qgroup_reserved = 0;
486         bool reloc_reserved = false;
487         int ret;
488
489         /* Send isn't supposed to start transactions. */
490         ASSERT(current->journal_info != BTRFS_SEND_TRANS_STUB);
491
492         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
493                 return ERR_PTR(-EROFS);
494
495         if (current->journal_info) {
496                 WARN_ON(type & TRANS_EXTWRITERS);
497                 h = current->journal_info;
498                 refcount_inc(&h->use_count);
499                 WARN_ON(refcount_read(&h->use_count) > 2);
500                 h->orig_rsv = h->block_rsv;
501                 h->block_rsv = NULL;
502                 goto got_it;
503         }
504
505         /*
506          * Do the reservation before we join the transaction so we can do all
507          * the appropriate flushing if need be.
508          */
509         if (num_items && root != fs_info->chunk_root) {
510                 qgroup_reserved = num_items * fs_info->nodesize;
511                 ret = btrfs_qgroup_reserve_meta(root, qgroup_reserved,
512                                                 enforce_qgroups);
513                 if (ret)
514                         return ERR_PTR(ret);
515
516                 num_bytes = btrfs_calc_trans_metadata_size(fs_info, num_items);
517                 /*
518                  * Do the reservation for the relocation root creation
519                  */
520                 if (need_reserve_reloc_root(root)) {
521                         num_bytes += fs_info->nodesize;
522                         reloc_reserved = true;
523                 }
524
525                 ret = btrfs_block_rsv_add(root, &fs_info->trans_block_rsv,
526                                           num_bytes, flush);
527                 if (ret)
528                         goto reserve_fail;
529         }
530 again:
531         h = kmem_cache_zalloc(btrfs_trans_handle_cachep, GFP_NOFS);
532         if (!h) {
533                 ret = -ENOMEM;
534                 goto alloc_fail;
535         }
536
537         /*
538          * If we are JOIN_NOLOCK we're already committing a transaction and
539          * waiting on this guy, so we don't need to do the sb_start_intwrite
540          * because we're already holding a ref.  We need this because we could
541          * have raced in and did an fsync() on a file which can kick a commit
542          * and then we deadlock with somebody doing a freeze.
543          *
544          * If we are ATTACH, it means we just want to catch the current
545          * transaction and commit it, so we needn't do sb_start_intwrite(). 
546          */
547         if (type & __TRANS_FREEZABLE)
548                 sb_start_intwrite(fs_info->sb);
549
550         if (may_wait_transaction(fs_info, type))
551                 wait_current_trans(fs_info);
552
553         do {
554                 ret = join_transaction(fs_info, type);
555                 if (ret == -EBUSY) {
556                         wait_current_trans(fs_info);
557                         if (unlikely(type == TRANS_ATTACH))
558                                 ret = -ENOENT;
559                 }
560         } while (ret == -EBUSY);
561
562         if (ret < 0)
563                 goto join_fail;
564
565         cur_trans = fs_info->running_transaction;
566
567         h->transid = cur_trans->transid;
568         h->transaction = cur_trans;
569         h->root = root;
570         refcount_set(&h->use_count, 1);
571         h->fs_info = root->fs_info;
572
573         h->type = type;
574         h->can_flush_pending_bgs = true;
575         INIT_LIST_HEAD(&h->new_bgs);
576
577         smp_mb();
578         if (cur_trans->state >= TRANS_STATE_BLOCKED &&
579             may_wait_transaction(fs_info, type)) {
580                 current->journal_info = h;
581                 btrfs_commit_transaction(h);
582                 goto again;
583         }
584
585         if (num_bytes) {
586                 trace_btrfs_space_reservation(fs_info, "transaction",
587                                               h->transid, num_bytes, 1);
588                 h->block_rsv = &fs_info->trans_block_rsv;
589                 h->bytes_reserved = num_bytes;
590                 h->reloc_reserved = reloc_reserved;
591         }
592
593 got_it:
594         btrfs_record_root_in_trans(h, root);
595
596         if (!current->journal_info && type != TRANS_USERSPACE)
597                 current->journal_info = h;
598         return h;
599
600 join_fail:
601         if (type & __TRANS_FREEZABLE)
602                 sb_end_intwrite(fs_info->sb);
603         kmem_cache_free(btrfs_trans_handle_cachep, h);
604 alloc_fail:
605         if (num_bytes)
606                 btrfs_block_rsv_release(fs_info, &fs_info->trans_block_rsv,
607                                         num_bytes);
608 reserve_fail:
609         btrfs_qgroup_free_meta(root, qgroup_reserved);
610         return ERR_PTR(ret);
611 }
612
613 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
614                                                    unsigned int num_items)
615 {
616         return start_transaction(root, num_items, TRANS_START,
617                                  BTRFS_RESERVE_FLUSH_ALL, true);
618 }
619
620 struct btrfs_trans_handle *btrfs_start_transaction_fallback_global_rsv(
621                                         struct btrfs_root *root,
622                                         unsigned int num_items,
623                                         int min_factor)
624 {
625         struct btrfs_fs_info *fs_info = root->fs_info;
626         struct btrfs_trans_handle *trans;
627         u64 num_bytes;
628         int ret;
629
630         /*
631          * We have two callers: unlink and block group removal.  The
632          * former should succeed even if we will temporarily exceed
633          * quota and the latter operates on the extent root so
634          * qgroup enforcement is ignored anyway.
635          */
636         trans = start_transaction(root, num_items, TRANS_START,
637                                   BTRFS_RESERVE_FLUSH_ALL, false);
638         if (!IS_ERR(trans) || PTR_ERR(trans) != -ENOSPC)
639                 return trans;
640
641         trans = btrfs_start_transaction(root, 0);
642         if (IS_ERR(trans))
643                 return trans;
644
645         num_bytes = btrfs_calc_trans_metadata_size(fs_info, num_items);
646         ret = btrfs_cond_migrate_bytes(fs_info, &fs_info->trans_block_rsv,
647                                        num_bytes, min_factor);
648         if (ret) {
649                 btrfs_end_transaction(trans);
650                 return ERR_PTR(ret);
651         }
652
653         trans->block_rsv = &fs_info->trans_block_rsv;
654         trans->bytes_reserved = num_bytes;
655         trace_btrfs_space_reservation(fs_info, "transaction",
656                                       trans->transid, num_bytes, 1);
657
658         return trans;
659 }
660
661 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
662 {
663         return start_transaction(root, 0, TRANS_JOIN, BTRFS_RESERVE_NO_FLUSH,
664                                  true);
665 }
666
667 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root)
668 {
669         return start_transaction(root, 0, TRANS_JOIN_NOLOCK,
670                                  BTRFS_RESERVE_NO_FLUSH, true);
671 }
672
673 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *root)
674 {
675         return start_transaction(root, 0, TRANS_USERSPACE,
676                                  BTRFS_RESERVE_NO_FLUSH, true);
677 }
678
679 /*
680  * btrfs_attach_transaction() - catch the running transaction
681  *
682  * It is used when we want to commit the current the transaction, but
683  * don't want to start a new one.
684  *
685  * Note: If this function return -ENOENT, it just means there is no
686  * running transaction. But it is possible that the inactive transaction
687  * is still in the memory, not fully on disk. If you hope there is no
688  * inactive transaction in the fs when -ENOENT is returned, you should
689  * invoke
690  *     btrfs_attach_transaction_barrier()
691  */
692 struct btrfs_trans_handle *btrfs_attach_transaction(struct btrfs_root *root)
693 {
694         return start_transaction(root, 0, TRANS_ATTACH,
695                                  BTRFS_RESERVE_NO_FLUSH, true);
696 }
697
698 /*
699  * btrfs_attach_transaction_barrier() - catch the running transaction
700  *
701  * It is similar to the above function, the differentia is this one
702  * will wait for all the inactive transactions until they fully
703  * complete.
704  */
705 struct btrfs_trans_handle *
706 btrfs_attach_transaction_barrier(struct btrfs_root *root)
707 {
708         struct btrfs_trans_handle *trans;
709
710         trans = start_transaction(root, 0, TRANS_ATTACH,
711                                   BTRFS_RESERVE_NO_FLUSH, true);
712         if (IS_ERR(trans) && PTR_ERR(trans) == -ENOENT)
713                 btrfs_wait_for_commit(root->fs_info, 0);
714
715         return trans;
716 }
717
718 /* wait for a transaction commit to be fully complete */
719 static noinline void wait_for_commit(struct btrfs_transaction *commit)
720 {
721         wait_event(commit->commit_wait, commit->state == TRANS_STATE_COMPLETED);
722 }
723
724 int btrfs_wait_for_commit(struct btrfs_fs_info *fs_info, u64 transid)
725 {
726         struct btrfs_transaction *cur_trans = NULL, *t;
727         int ret = 0;
728
729         if (transid) {
730                 if (transid <= fs_info->last_trans_committed)
731                         goto out;
732
733                 /* find specified transaction */
734                 spin_lock(&fs_info->trans_lock);
735                 list_for_each_entry(t, &fs_info->trans_list, list) {
736                         if (t->transid == transid) {
737                                 cur_trans = t;
738                                 refcount_inc(&cur_trans->use_count);
739                                 ret = 0;
740                                 break;
741                         }
742                         if (t->transid > transid) {
743                                 ret = 0;
744                                 break;
745                         }
746                 }
747                 spin_unlock(&fs_info->trans_lock);
748
749                 /*
750                  * The specified transaction doesn't exist, or we
751                  * raced with btrfs_commit_transaction
752                  */
753                 if (!cur_trans) {
754                         if (transid > fs_info->last_trans_committed)
755                                 ret = -EINVAL;
756                         goto out;
757                 }
758         } else {
759                 /* find newest transaction that is committing | committed */
760                 spin_lock(&fs_info->trans_lock);
761                 list_for_each_entry_reverse(t, &fs_info->trans_list,
762                                             list) {
763                         if (t->state >= TRANS_STATE_COMMIT_START) {
764                                 if (t->state == TRANS_STATE_COMPLETED)
765                                         break;
766                                 cur_trans = t;
767                                 refcount_inc(&cur_trans->use_count);
768                                 break;
769                         }
770                 }
771                 spin_unlock(&fs_info->trans_lock);
772                 if (!cur_trans)
773                         goto out;  /* nothing committing|committed */
774         }
775
776         wait_for_commit(cur_trans);
777         btrfs_put_transaction(cur_trans);
778 out:
779         return ret;
780 }
781
782 void btrfs_throttle(struct btrfs_fs_info *fs_info)
783 {
784         if (!atomic_read(&fs_info->open_ioctl_trans))
785                 wait_current_trans(fs_info);
786 }
787
788 static int should_end_transaction(struct btrfs_trans_handle *trans)
789 {
790         struct btrfs_fs_info *fs_info = trans->fs_info;
791
792         if (btrfs_check_space_for_delayed_refs(trans, fs_info))
793                 return 1;
794
795         return !!btrfs_block_rsv_check(&fs_info->global_block_rsv, 5);
796 }
797
798 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans)
799 {
800         struct btrfs_transaction *cur_trans = trans->transaction;
801         struct btrfs_fs_info *fs_info = trans->fs_info;
802         int updates;
803         int err;
804
805         smp_mb();
806         if (cur_trans->state >= TRANS_STATE_BLOCKED ||
807             cur_trans->delayed_refs.flushing)
808                 return 1;
809
810         updates = trans->delayed_ref_updates;
811         trans->delayed_ref_updates = 0;
812         if (updates) {
813                 err = btrfs_run_delayed_refs(trans, fs_info, updates * 2);
814                 if (err) /* Error code will also eval true */
815                         return err;
816         }
817
818         return should_end_transaction(trans);
819 }
820
821 static void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
822                                   struct btrfs_fs_info *fs_info)
823 {
824         if (!trans->block_rsv) {
825                 ASSERT(!trans->bytes_reserved);
826                 return;
827         }
828
829         if (!trans->bytes_reserved)
830                 return;
831
832         ASSERT(trans->block_rsv == &fs_info->trans_block_rsv);
833         trace_btrfs_space_reservation(fs_info, "transaction",
834                                       trans->transid, trans->bytes_reserved, 0);
835         btrfs_block_rsv_release(fs_info, trans->block_rsv,
836                                 trans->bytes_reserved);
837         trans->bytes_reserved = 0;
838 }
839
840 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
841                                    int throttle)
842 {
843         struct btrfs_fs_info *info = trans->fs_info;
844         struct btrfs_transaction *cur_trans = trans->transaction;
845         u64 transid = trans->transid;
846         unsigned long cur = trans->delayed_ref_updates;
847         int lock = (trans->type != TRANS_JOIN_NOLOCK);
848         int err = 0;
849         int must_run_delayed_refs = 0;
850
851         if (refcount_read(&trans->use_count) > 1) {
852                 refcount_dec(&trans->use_count);
853                 trans->block_rsv = trans->orig_rsv;
854                 return 0;
855         }
856
857         btrfs_trans_release_metadata(trans, info);
858         trans->block_rsv = NULL;
859
860         if (!list_empty(&trans->new_bgs))
861                 btrfs_create_pending_block_groups(trans, info);
862
863         trans->delayed_ref_updates = 0;
864         if (!trans->sync) {
865                 must_run_delayed_refs =
866                         btrfs_should_throttle_delayed_refs(trans, info);
867                 cur = max_t(unsigned long, cur, 32);
868
869                 /*
870                  * don't make the caller wait if they are from a NOLOCK
871                  * or ATTACH transaction, it will deadlock with commit
872                  */
873                 if (must_run_delayed_refs == 1 &&
874                     (trans->type & (__TRANS_JOIN_NOLOCK | __TRANS_ATTACH)))
875                         must_run_delayed_refs = 2;
876         }
877
878         btrfs_trans_release_metadata(trans, info);
879         trans->block_rsv = NULL;
880
881         if (!list_empty(&trans->new_bgs))
882                 btrfs_create_pending_block_groups(trans, info);
883
884         btrfs_trans_release_chunk_metadata(trans);
885
886         if (lock && !atomic_read(&info->open_ioctl_trans) &&
887             should_end_transaction(trans) &&
888             READ_ONCE(cur_trans->state) == TRANS_STATE_RUNNING) {
889                 spin_lock(&info->trans_lock);
890                 if (cur_trans->state == TRANS_STATE_RUNNING)
891                         cur_trans->state = TRANS_STATE_BLOCKED;
892                 spin_unlock(&info->trans_lock);
893         }
894
895         if (lock && READ_ONCE(cur_trans->state) == TRANS_STATE_BLOCKED) {
896                 if (throttle)
897                         return btrfs_commit_transaction(trans);
898                 else
899                         wake_up_process(info->transaction_kthread);
900         }
901
902         if (trans->type & __TRANS_FREEZABLE)
903                 sb_end_intwrite(info->sb);
904
905         WARN_ON(cur_trans != info->running_transaction);
906         WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
907         atomic_dec(&cur_trans->num_writers);
908         extwriter_counter_dec(cur_trans, trans->type);
909
910         /*
911          * Make sure counter is updated before we wake up waiters.
912          */
913         smp_mb();
914         if (waitqueue_active(&cur_trans->writer_wait))
915                 wake_up(&cur_trans->writer_wait);
916         btrfs_put_transaction(cur_trans);
917
918         if (current->journal_info == trans)
919                 current->journal_info = NULL;
920
921         if (throttle)
922                 btrfs_run_delayed_iputs(info);
923
924         if (trans->aborted ||
925             test_bit(BTRFS_FS_STATE_ERROR, &info->fs_state)) {
926                 wake_up_process(info->transaction_kthread);
927                 err = -EIO;
928         }
929
930         kmem_cache_free(btrfs_trans_handle_cachep, trans);
931         if (must_run_delayed_refs) {
932                 btrfs_async_run_delayed_refs(info, cur, transid,
933                                              must_run_delayed_refs == 1);
934         }
935         return err;
936 }
937
938 int btrfs_end_transaction(struct btrfs_trans_handle *trans)
939 {
940         return __btrfs_end_transaction(trans, 0);
941 }
942
943 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans)
944 {
945         return __btrfs_end_transaction(trans, 1);
946 }
947
948 /*
949  * when btree blocks are allocated, they have some corresponding bits set for
950  * them in one of two extent_io trees.  This is used to make sure all of
951  * those extents are sent to disk but does not wait on them
952  */
953 int btrfs_write_marked_extents(struct btrfs_fs_info *fs_info,
954                                struct extent_io_tree *dirty_pages, int mark)
955 {
956         int err = 0;
957         int werr = 0;
958         struct address_space *mapping = fs_info->btree_inode->i_mapping;
959         struct extent_state *cached_state = NULL;
960         u64 start = 0;
961         u64 end;
962
963         atomic_inc(&BTRFS_I(fs_info->btree_inode)->sync_writers);
964         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
965                                       mark, &cached_state)) {
966                 bool wait_writeback = false;
967
968                 err = convert_extent_bit(dirty_pages, start, end,
969                                          EXTENT_NEED_WAIT,
970                                          mark, &cached_state);
971                 /*
972                  * convert_extent_bit can return -ENOMEM, which is most of the
973                  * time a temporary error. So when it happens, ignore the error
974                  * and wait for writeback of this range to finish - because we
975                  * failed to set the bit EXTENT_NEED_WAIT for the range, a call
976                  * to __btrfs_wait_marked_extents() would not know that
977                  * writeback for this range started and therefore wouldn't
978                  * wait for it to finish - we don't want to commit a
979                  * superblock that points to btree nodes/leafs for which
980                  * writeback hasn't finished yet (and without errors).
981                  * We cleanup any entries left in the io tree when committing
982                  * the transaction (through clear_btree_io_tree()).
983                  */
984                 if (err == -ENOMEM) {
985                         err = 0;
986                         wait_writeback = true;
987                 }
988                 if (!err)
989                         err = filemap_fdatawrite_range(mapping, start, end);
990                 if (err)
991                         werr = err;
992                 else if (wait_writeback)
993                         werr = filemap_fdatawait_range(mapping, start, end);
994                 free_extent_state(cached_state);
995                 cached_state = NULL;
996                 cond_resched();
997                 start = end + 1;
998         }
999         atomic_dec(&BTRFS_I(fs_info->btree_inode)->sync_writers);
1000         return werr;
1001 }
1002
1003 /*
1004  * when btree blocks are allocated, they have some corresponding bits set for
1005  * them in one of two extent_io trees.  This is used to make sure all of
1006  * those extents are on disk for transaction or log commit.  We wait
1007  * on all the pages and clear them from the dirty pages state tree
1008  */
1009 static int __btrfs_wait_marked_extents(struct btrfs_fs_info *fs_info,
1010                                        struct extent_io_tree *dirty_pages)
1011 {
1012         int err = 0;
1013         int werr = 0;
1014         struct address_space *mapping = fs_info->btree_inode->i_mapping;
1015         struct extent_state *cached_state = NULL;
1016         u64 start = 0;
1017         u64 end;
1018
1019         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
1020                                       EXTENT_NEED_WAIT, &cached_state)) {
1021                 /*
1022                  * Ignore -ENOMEM errors returned by clear_extent_bit().
1023                  * When committing the transaction, we'll remove any entries
1024                  * left in the io tree. For a log commit, we don't remove them
1025                  * after committing the log because the tree can be accessed
1026                  * concurrently - we do it only at transaction commit time when
1027                  * it's safe to do it (through clear_btree_io_tree()).
1028                  */
1029                 err = clear_extent_bit(dirty_pages, start, end,
1030                                        EXTENT_NEED_WAIT, 0, 0, &cached_state);
1031                 if (err == -ENOMEM)
1032                         err = 0;
1033                 if (!err)
1034                         err = filemap_fdatawait_range(mapping, start, end);
1035                 if (err)
1036                         werr = err;
1037                 free_extent_state(cached_state);
1038                 cached_state = NULL;
1039                 cond_resched();
1040                 start = end + 1;
1041         }
1042         if (err)
1043                 werr = err;
1044         return werr;
1045 }
1046
1047 int btrfs_wait_extents(struct btrfs_fs_info *fs_info,
1048                        struct extent_io_tree *dirty_pages)
1049 {
1050         bool errors = false;
1051         int err;
1052
1053         err = __btrfs_wait_marked_extents(fs_info, dirty_pages);
1054         if (test_and_clear_bit(BTRFS_FS_BTREE_ERR, &fs_info->flags))
1055                 errors = true;
1056
1057         if (errors && !err)
1058                 err = -EIO;
1059         return err;
1060 }
1061
1062 int btrfs_wait_tree_log_extents(struct btrfs_root *log_root, int mark)
1063 {
1064         struct btrfs_fs_info *fs_info = log_root->fs_info;
1065         struct extent_io_tree *dirty_pages = &log_root->dirty_log_pages;
1066         bool errors = false;
1067         int err;
1068
1069         ASSERT(log_root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
1070
1071         err = __btrfs_wait_marked_extents(fs_info, dirty_pages);
1072         if ((mark & EXTENT_DIRTY) &&
1073             test_and_clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags))
1074                 errors = true;
1075
1076         if ((mark & EXTENT_NEW) &&
1077             test_and_clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags))
1078                 errors = true;
1079
1080         if (errors && !err)
1081                 err = -EIO;
1082         return err;
1083 }
1084
1085 /*
1086  * when btree blocks are allocated, they have some corresponding bits set for
1087  * them in one of two extent_io trees.  This is used to make sure all of
1088  * those extents are on disk for transaction or log commit
1089  */
1090 static int btrfs_write_and_wait_marked_extents(struct btrfs_fs_info *fs_info,
1091                                 struct extent_io_tree *dirty_pages, int mark)
1092 {
1093         int ret;
1094         int ret2;
1095         struct blk_plug plug;
1096
1097         blk_start_plug(&plug);
1098         ret = btrfs_write_marked_extents(fs_info, dirty_pages, mark);
1099         blk_finish_plug(&plug);
1100         ret2 = btrfs_wait_extents(fs_info, dirty_pages);
1101
1102         if (ret)
1103                 return ret;
1104         if (ret2)
1105                 return ret2;
1106         return 0;
1107 }
1108
1109 static int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
1110                                             struct btrfs_fs_info *fs_info)
1111 {
1112         int ret;
1113
1114         ret = btrfs_write_and_wait_marked_extents(fs_info,
1115                                            &trans->transaction->dirty_pages,
1116                                            EXTENT_DIRTY);
1117         clear_btree_io_tree(&trans->transaction->dirty_pages);
1118
1119         return ret;
1120 }
1121
1122 /*
1123  * this is used to update the root pointer in the tree of tree roots.
1124  *
1125  * But, in the case of the extent allocation tree, updating the root
1126  * pointer may allocate blocks which may change the root of the extent
1127  * allocation tree.
1128  *
1129  * So, this loops and repeats and makes sure the cowonly root didn't
1130  * change while the root pointer was being updated in the metadata.
1131  */
1132 static int update_cowonly_root(struct btrfs_trans_handle *trans,
1133                                struct btrfs_root *root)
1134 {
1135         int ret;
1136         u64 old_root_bytenr;
1137         u64 old_root_used;
1138         struct btrfs_fs_info *fs_info = root->fs_info;
1139         struct btrfs_root *tree_root = fs_info->tree_root;
1140
1141         old_root_used = btrfs_root_used(&root->root_item);
1142
1143         while (1) {
1144                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
1145                 if (old_root_bytenr == root->node->start &&
1146                     old_root_used == btrfs_root_used(&root->root_item))
1147                         break;
1148
1149                 btrfs_set_root_node(&root->root_item, root->node);
1150                 ret = btrfs_update_root(trans, tree_root,
1151                                         &root->root_key,
1152                                         &root->root_item);
1153                 if (ret)
1154                         return ret;
1155
1156                 old_root_used = btrfs_root_used(&root->root_item);
1157         }
1158
1159         return 0;
1160 }
1161
1162 /*
1163  * update all the cowonly tree roots on disk
1164  *
1165  * The error handling in this function may not be obvious. Any of the
1166  * failures will cause the file system to go offline. We still need
1167  * to clean up the delayed refs.
1168  */
1169 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
1170                                          struct btrfs_fs_info *fs_info)
1171 {
1172         struct list_head *dirty_bgs = &trans->transaction->dirty_bgs;
1173         struct list_head *io_bgs = &trans->transaction->io_bgs;
1174         struct list_head *next;
1175         struct extent_buffer *eb;
1176         int ret;
1177
1178         eb = btrfs_lock_root_node(fs_info->tree_root);
1179         ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL,
1180                               0, &eb);
1181         btrfs_tree_unlock(eb);
1182         free_extent_buffer(eb);
1183
1184         if (ret)
1185                 return ret;
1186
1187         ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1188         if (ret)
1189                 return ret;
1190
1191         ret = btrfs_run_dev_stats(trans, fs_info);
1192         if (ret)
1193                 return ret;
1194         ret = btrfs_run_dev_replace(trans, fs_info);
1195         if (ret)
1196                 return ret;
1197         ret = btrfs_run_qgroups(trans, fs_info);
1198         if (ret)
1199                 return ret;
1200
1201         ret = btrfs_setup_space_cache(trans, fs_info);
1202         if (ret)
1203                 return ret;
1204
1205         /* run_qgroups might have added some more refs */
1206         ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1207         if (ret)
1208                 return ret;
1209 again:
1210         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
1211                 struct btrfs_root *root;
1212                 next = fs_info->dirty_cowonly_roots.next;
1213                 list_del_init(next);
1214                 root = list_entry(next, struct btrfs_root, dirty_list);
1215                 clear_bit(BTRFS_ROOT_DIRTY, &root->state);
1216
1217                 if (root != fs_info->extent_root)
1218                         list_add_tail(&root->dirty_list,
1219                                       &trans->transaction->switch_commits);
1220                 ret = update_cowonly_root(trans, root);
1221                 if (ret)
1222                         return ret;
1223                 ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1224                 if (ret)
1225                         return ret;
1226         }
1227
1228         while (!list_empty(dirty_bgs) || !list_empty(io_bgs)) {
1229                 ret = btrfs_write_dirty_block_groups(trans, fs_info);
1230                 if (ret)
1231                         return ret;
1232                 ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1233                 if (ret)
1234                         return ret;
1235         }
1236
1237         if (!list_empty(&fs_info->dirty_cowonly_roots))
1238                 goto again;
1239
1240         list_add_tail(&fs_info->extent_root->dirty_list,
1241                       &trans->transaction->switch_commits);
1242         btrfs_after_dev_replace_commit(fs_info);
1243
1244         return 0;
1245 }
1246
1247 /*
1248  * dead roots are old snapshots that need to be deleted.  This allocates
1249  * a dirty root struct and adds it into the list of dead roots that need to
1250  * be deleted
1251  */
1252 void btrfs_add_dead_root(struct btrfs_root *root)
1253 {
1254         struct btrfs_fs_info *fs_info = root->fs_info;
1255
1256         spin_lock(&fs_info->trans_lock);
1257         if (list_empty(&root->root_list))
1258                 list_add_tail(&root->root_list, &fs_info->dead_roots);
1259         spin_unlock(&fs_info->trans_lock);
1260 }
1261
1262 /*
1263  * update all the cowonly tree roots on disk
1264  */
1265 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
1266                                     struct btrfs_fs_info *fs_info)
1267 {
1268         struct btrfs_root *gang[8];
1269         int i;
1270         int ret;
1271         int err = 0;
1272
1273         spin_lock(&fs_info->fs_roots_radix_lock);
1274         while (1) {
1275                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
1276                                                  (void **)gang, 0,
1277                                                  ARRAY_SIZE(gang),
1278                                                  BTRFS_ROOT_TRANS_TAG);
1279                 if (ret == 0)
1280                         break;
1281                 for (i = 0; i < ret; i++) {
1282                         struct btrfs_root *root = gang[i];
1283                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
1284                                         (unsigned long)root->root_key.objectid,
1285                                         BTRFS_ROOT_TRANS_TAG);
1286                         spin_unlock(&fs_info->fs_roots_radix_lock);
1287
1288                         btrfs_free_log(trans, root);
1289                         btrfs_update_reloc_root(trans, root);
1290                         btrfs_orphan_commit_root(trans, root);
1291
1292                         btrfs_save_ino_cache(root, trans);
1293
1294                         /* see comments in should_cow_block() */
1295                         clear_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1296                         smp_mb__after_atomic();
1297
1298                         if (root->commit_root != root->node) {
1299                                 list_add_tail(&root->dirty_list,
1300                                         &trans->transaction->switch_commits);
1301                                 btrfs_set_root_node(&root->root_item,
1302                                                     root->node);
1303                         }
1304
1305                         err = btrfs_update_root(trans, fs_info->tree_root,
1306                                                 &root->root_key,
1307                                                 &root->root_item);
1308                         spin_lock(&fs_info->fs_roots_radix_lock);
1309                         if (err)
1310                                 break;
1311                         btrfs_qgroup_free_meta_all(root);
1312                 }
1313         }
1314         spin_unlock(&fs_info->fs_roots_radix_lock);
1315         return err;
1316 }
1317
1318 /*
1319  * defrag a given btree.
1320  * Every leaf in the btree is read and defragged.
1321  */
1322 int btrfs_defrag_root(struct btrfs_root *root)
1323 {
1324         struct btrfs_fs_info *info = root->fs_info;
1325         struct btrfs_trans_handle *trans;
1326         int ret;
1327
1328         if (test_and_set_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state))
1329                 return 0;
1330
1331         while (1) {
1332                 trans = btrfs_start_transaction(root, 0);
1333                 if (IS_ERR(trans))
1334                         return PTR_ERR(trans);
1335
1336                 ret = btrfs_defrag_leaves(trans, root);
1337
1338                 btrfs_end_transaction(trans);
1339                 btrfs_btree_balance_dirty(info);
1340                 cond_resched();
1341
1342                 if (btrfs_fs_closing(info) || ret != -EAGAIN)
1343                         break;
1344
1345                 if (btrfs_defrag_cancelled(info)) {
1346                         btrfs_debug(info, "defrag_root cancelled");
1347                         ret = -EAGAIN;
1348                         break;
1349                 }
1350         }
1351         clear_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state);
1352         return ret;
1353 }
1354
1355 /*
1356  * Do all special snapshot related qgroup dirty hack.
1357  *
1358  * Will do all needed qgroup inherit and dirty hack like switch commit
1359  * roots inside one transaction and write all btree into disk, to make
1360  * qgroup works.
1361  */
1362 static int qgroup_account_snapshot(struct btrfs_trans_handle *trans,
1363                                    struct btrfs_root *src,
1364                                    struct btrfs_root *parent,
1365                                    struct btrfs_qgroup_inherit *inherit,
1366                                    u64 dst_objectid)
1367 {
1368         struct btrfs_fs_info *fs_info = src->fs_info;
1369         int ret;
1370
1371         /*
1372          * Save some performance in the case that qgroups are not
1373          * enabled. If this check races with the ioctl, rescan will
1374          * kick in anyway.
1375          */
1376         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
1377                 return 0;
1378
1379         /*
1380          * We are going to commit transaction, see btrfs_commit_transaction()
1381          * comment for reason locking tree_log_mutex
1382          */
1383         mutex_lock(&fs_info->tree_log_mutex);
1384
1385         ret = commit_fs_roots(trans, fs_info);
1386         if (ret)
1387                 goto out;
1388         ret = btrfs_qgroup_account_extents(trans, fs_info);
1389         if (ret < 0)
1390                 goto out;
1391
1392         /* Now qgroup are all updated, we can inherit it to new qgroups */
1393         ret = btrfs_qgroup_inherit(trans, fs_info,
1394                                    src->root_key.objectid, dst_objectid,
1395                                    inherit);
1396         if (ret < 0)
1397                 goto out;
1398
1399         /*
1400          * Now we do a simplified commit transaction, which will:
1401          * 1) commit all subvolume and extent tree
1402          *    To ensure all subvolume and extent tree have a valid
1403          *    commit_root to accounting later insert_dir_item()
1404          * 2) write all btree blocks onto disk
1405          *    This is to make sure later btree modification will be cowed
1406          *    Or commit_root can be populated and cause wrong qgroup numbers
1407          * In this simplified commit, we don't really care about other trees
1408          * like chunk and root tree, as they won't affect qgroup.
1409          * And we don't write super to avoid half committed status.
1410          */
1411         ret = commit_cowonly_roots(trans, fs_info);
1412         if (ret)
1413                 goto out;
1414         switch_commit_roots(trans->transaction, fs_info);
1415         ret = btrfs_write_and_wait_transaction(trans, fs_info);
1416         if (ret)
1417                 btrfs_handle_fs_error(fs_info, ret,
1418                         "Error while writing out transaction for qgroup");
1419
1420 out:
1421         mutex_unlock(&fs_info->tree_log_mutex);
1422
1423         /*
1424          * Force parent root to be updated, as we recorded it before so its
1425          * last_trans == cur_transid.
1426          * Or it won't be committed again onto disk after later
1427          * insert_dir_item()
1428          */
1429         if (!ret)
1430                 record_root_in_trans(trans, parent, 1);
1431         return ret;
1432 }
1433
1434 /*
1435  * new snapshots need to be created at a very specific time in the
1436  * transaction commit.  This does the actual creation.
1437  *
1438  * Note:
1439  * If the error which may affect the commitment of the current transaction
1440  * happens, we should return the error number. If the error which just affect
1441  * the creation of the pending snapshots, just return 0.
1442  */
1443 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
1444                                    struct btrfs_fs_info *fs_info,
1445                                    struct btrfs_pending_snapshot *pending)
1446 {
1447         struct btrfs_key key;
1448         struct btrfs_root_item *new_root_item;
1449         struct btrfs_root *tree_root = fs_info->tree_root;
1450         struct btrfs_root *root = pending->root;
1451         struct btrfs_root *parent_root;
1452         struct btrfs_block_rsv *rsv;
1453         struct inode *parent_inode;
1454         struct btrfs_path *path;
1455         struct btrfs_dir_item *dir_item;
1456         struct dentry *dentry;
1457         struct extent_buffer *tmp;
1458         struct extent_buffer *old;
1459         struct timespec cur_time;
1460         int ret = 0;
1461         u64 to_reserve = 0;
1462         u64 index = 0;
1463         u64 objectid;
1464         u64 root_flags;
1465         uuid_le new_uuid;
1466
1467         ASSERT(pending->path);
1468         path = pending->path;
1469
1470         ASSERT(pending->root_item);
1471         new_root_item = pending->root_item;
1472
1473         pending->error = btrfs_find_free_objectid(tree_root, &objectid);
1474         if (pending->error)
1475                 goto no_free_objectid;
1476
1477         /*
1478          * Make qgroup to skip current new snapshot's qgroupid, as it is
1479          * accounted by later btrfs_qgroup_inherit().
1480          */
1481         btrfs_set_skip_qgroup(trans, objectid);
1482
1483         btrfs_reloc_pre_snapshot(pending, &to_reserve);
1484
1485         if (to_reserve > 0) {
1486                 pending->error = btrfs_block_rsv_add(root,
1487                                                      &pending->block_rsv,
1488                                                      to_reserve,
1489                                                      BTRFS_RESERVE_NO_FLUSH);
1490                 if (pending->error)
1491                         goto clear_skip_qgroup;
1492         }
1493
1494         key.objectid = objectid;
1495         key.offset = (u64)-1;
1496         key.type = BTRFS_ROOT_ITEM_KEY;
1497
1498         rsv = trans->block_rsv;
1499         trans->block_rsv = &pending->block_rsv;
1500         trans->bytes_reserved = trans->block_rsv->reserved;
1501         trace_btrfs_space_reservation(fs_info, "transaction",
1502                                       trans->transid,
1503                                       trans->bytes_reserved, 1);
1504         dentry = pending->dentry;
1505         parent_inode = pending->dir;
1506         parent_root = BTRFS_I(parent_inode)->root;
1507         record_root_in_trans(trans, parent_root, 0);
1508
1509         cur_time = current_time(parent_inode);
1510
1511         /*
1512          * insert the directory item
1513          */
1514         ret = btrfs_set_inode_index(BTRFS_I(parent_inode), &index);
1515         BUG_ON(ret); /* -ENOMEM */
1516
1517         /* check if there is a file/dir which has the same name. */
1518         dir_item = btrfs_lookup_dir_item(NULL, parent_root, path,
1519                                          btrfs_ino(BTRFS_I(parent_inode)),
1520                                          dentry->d_name.name,
1521                                          dentry->d_name.len, 0);
1522         if (dir_item != NULL && !IS_ERR(dir_item)) {
1523                 pending->error = -EEXIST;
1524                 goto dir_item_existed;
1525         } else if (IS_ERR(dir_item)) {
1526                 ret = PTR_ERR(dir_item);
1527                 btrfs_abort_transaction(trans, ret);
1528                 goto fail;
1529         }
1530         btrfs_release_path(path);
1531
1532         /*
1533          * pull in the delayed directory update
1534          * and the delayed inode item
1535          * otherwise we corrupt the FS during
1536          * snapshot
1537          */
1538         ret = btrfs_run_delayed_items(trans, fs_info);
1539         if (ret) {      /* Transaction aborted */
1540                 btrfs_abort_transaction(trans, ret);
1541                 goto fail;
1542         }
1543
1544         record_root_in_trans(trans, root, 0);
1545         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
1546         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
1547         btrfs_check_and_init_root_item(new_root_item);
1548
1549         root_flags = btrfs_root_flags(new_root_item);
1550         if (pending->readonly)
1551                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
1552         else
1553                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
1554         btrfs_set_root_flags(new_root_item, root_flags);
1555
1556         btrfs_set_root_generation_v2(new_root_item,
1557                         trans->transid);
1558         uuid_le_gen(&new_uuid);
1559         memcpy(new_root_item->uuid, new_uuid.b, BTRFS_UUID_SIZE);
1560         memcpy(new_root_item->parent_uuid, root->root_item.uuid,
1561                         BTRFS_UUID_SIZE);
1562         if (!(root_flags & BTRFS_ROOT_SUBVOL_RDONLY)) {
1563                 memset(new_root_item->received_uuid, 0,
1564                        sizeof(new_root_item->received_uuid));
1565                 memset(&new_root_item->stime, 0, sizeof(new_root_item->stime));
1566                 memset(&new_root_item->rtime, 0, sizeof(new_root_item->rtime));
1567                 btrfs_set_root_stransid(new_root_item, 0);
1568                 btrfs_set_root_rtransid(new_root_item, 0);
1569         }
1570         btrfs_set_stack_timespec_sec(&new_root_item->otime, cur_time.tv_sec);
1571         btrfs_set_stack_timespec_nsec(&new_root_item->otime, cur_time.tv_nsec);
1572         btrfs_set_root_otransid(new_root_item, trans->transid);
1573
1574         old = btrfs_lock_root_node(root);
1575         ret = btrfs_cow_block(trans, root, old, NULL, 0, &old);
1576         if (ret) {
1577                 btrfs_tree_unlock(old);
1578                 free_extent_buffer(old);
1579                 btrfs_abort_transaction(trans, ret);
1580                 goto fail;
1581         }
1582
1583         btrfs_set_lock_blocking(old);
1584
1585         ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1586         /* clean up in any case */
1587         btrfs_tree_unlock(old);
1588         free_extent_buffer(old);
1589         if (ret) {
1590                 btrfs_abort_transaction(trans, ret);
1591                 goto fail;
1592         }
1593         /* see comments in should_cow_block() */
1594         set_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1595         smp_wmb();
1596
1597         btrfs_set_root_node(new_root_item, tmp);
1598         /* record when the snapshot was created in key.offset */
1599         key.offset = trans->transid;
1600         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1601         btrfs_tree_unlock(tmp);
1602         free_extent_buffer(tmp);
1603         if (ret) {
1604                 btrfs_abort_transaction(trans, ret);
1605                 goto fail;
1606         }
1607
1608         /*
1609          * insert root back/forward references
1610          */
1611         ret = btrfs_add_root_ref(trans, fs_info, objectid,
1612                                  parent_root->root_key.objectid,
1613                                  btrfs_ino(BTRFS_I(parent_inode)), index,
1614                                  dentry->d_name.name, dentry->d_name.len);
1615         if (ret) {
1616                 btrfs_abort_transaction(trans, ret);
1617                 goto fail;
1618         }
1619
1620         key.offset = (u64)-1;
1621         pending->snap = btrfs_read_fs_root_no_name(fs_info, &key);
1622         if (IS_ERR(pending->snap)) {
1623                 ret = PTR_ERR(pending->snap);
1624                 btrfs_abort_transaction(trans, ret);
1625                 goto fail;
1626         }
1627
1628         ret = btrfs_reloc_post_snapshot(trans, pending);
1629         if (ret) {
1630                 btrfs_abort_transaction(trans, ret);
1631                 goto fail;
1632         }
1633
1634         ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1635         if (ret) {
1636                 btrfs_abort_transaction(trans, ret);
1637                 goto fail;
1638         }
1639
1640         /*
1641          * Do special qgroup accounting for snapshot, as we do some qgroup
1642          * snapshot hack to do fast snapshot.
1643          * To co-operate with that hack, we do hack again.
1644          * Or snapshot will be greatly slowed down by a subtree qgroup rescan
1645          */
1646         ret = qgroup_account_snapshot(trans, root, parent_root,
1647                                       pending->inherit, objectid);
1648         if (ret < 0)
1649                 goto fail;
1650
1651         ret = btrfs_insert_dir_item(trans, parent_root,
1652                                     dentry->d_name.name, dentry->d_name.len,
1653                                     BTRFS_I(parent_inode), &key,
1654                                     BTRFS_FT_DIR, index);
1655         /* We have check then name at the beginning, so it is impossible. */
1656         BUG_ON(ret == -EEXIST || ret == -EOVERFLOW);
1657         if (ret) {
1658                 btrfs_abort_transaction(trans, ret);
1659                 goto fail;
1660         }
1661
1662         btrfs_i_size_write(BTRFS_I(parent_inode), parent_inode->i_size +
1663                                          dentry->d_name.len * 2);
1664         parent_inode->i_mtime = parent_inode->i_ctime =
1665                 current_time(parent_inode);
1666         ret = btrfs_update_inode_fallback(trans, parent_root, parent_inode);
1667         if (ret) {
1668                 btrfs_abort_transaction(trans, ret);
1669                 goto fail;
1670         }
1671         ret = btrfs_uuid_tree_add(trans, fs_info, new_uuid.b,
1672                                   BTRFS_UUID_KEY_SUBVOL, objectid);
1673         if (ret) {
1674                 btrfs_abort_transaction(trans, ret);
1675                 goto fail;
1676         }
1677         if (!btrfs_is_empty_uuid(new_root_item->received_uuid)) {
1678                 ret = btrfs_uuid_tree_add(trans, fs_info,
1679                                           new_root_item->received_uuid,
1680                                           BTRFS_UUID_KEY_RECEIVED_SUBVOL,
1681                                           objectid);
1682                 if (ret && ret != -EEXIST) {
1683                         btrfs_abort_transaction(trans, ret);
1684                         goto fail;
1685                 }
1686         }
1687
1688         ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1689         if (ret) {
1690                 btrfs_abort_transaction(trans, ret);
1691                 goto fail;
1692         }
1693
1694 fail:
1695         pending->error = ret;
1696 dir_item_existed:
1697         trans->block_rsv = rsv;
1698         trans->bytes_reserved = 0;
1699 clear_skip_qgroup:
1700         btrfs_clear_skip_qgroup(trans);
1701 no_free_objectid:
1702         kfree(new_root_item);
1703         pending->root_item = NULL;
1704         btrfs_free_path(path);
1705         pending->path = NULL;
1706
1707         return ret;
1708 }
1709
1710 /*
1711  * create all the snapshots we've scheduled for creation
1712  */
1713 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
1714                                              struct btrfs_fs_info *fs_info)
1715 {
1716         struct btrfs_pending_snapshot *pending, *next;
1717         struct list_head *head = &trans->transaction->pending_snapshots;
1718         int ret = 0;
1719
1720         list_for_each_entry_safe(pending, next, head, list) {
1721                 list_del(&pending->list);
1722                 ret = create_pending_snapshot(trans, fs_info, pending);
1723                 if (ret)
1724                         break;
1725         }
1726         return ret;
1727 }
1728
1729 static void update_super_roots(struct btrfs_fs_info *fs_info)
1730 {
1731         struct btrfs_root_item *root_item;
1732         struct btrfs_super_block *super;
1733
1734         super = fs_info->super_copy;
1735
1736         root_item = &fs_info->chunk_root->root_item;
1737         super->chunk_root = root_item->bytenr;
1738         super->chunk_root_generation = root_item->generation;
1739         super->chunk_root_level = root_item->level;
1740
1741         root_item = &fs_info->tree_root->root_item;
1742         super->root = root_item->bytenr;
1743         super->generation = root_item->generation;
1744         super->root_level = root_item->level;
1745         if (btrfs_test_opt(fs_info, SPACE_CACHE))
1746                 super->cache_generation = root_item->generation;
1747         if (test_bit(BTRFS_FS_UPDATE_UUID_TREE_GEN, &fs_info->flags))
1748                 super->uuid_tree_generation = root_item->generation;
1749 }
1750
1751 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1752 {
1753         struct btrfs_transaction *trans;
1754         int ret = 0;
1755
1756         spin_lock(&info->trans_lock);
1757         trans = info->running_transaction;
1758         if (trans)
1759                 ret = (trans->state >= TRANS_STATE_COMMIT_START);
1760         spin_unlock(&info->trans_lock);
1761         return ret;
1762 }
1763
1764 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1765 {
1766         struct btrfs_transaction *trans;
1767         int ret = 0;
1768
1769         spin_lock(&info->trans_lock);
1770         trans = info->running_transaction;
1771         if (trans)
1772                 ret = is_transaction_blocked(trans);
1773         spin_unlock(&info->trans_lock);
1774         return ret;
1775 }
1776
1777 /*
1778  * wait for the current transaction commit to start and block subsequent
1779  * transaction joins
1780  */
1781 static void wait_current_trans_commit_start(struct btrfs_fs_info *fs_info,
1782                                             struct btrfs_transaction *trans)
1783 {
1784         wait_event(fs_info->transaction_blocked_wait,
1785                    trans->state >= TRANS_STATE_COMMIT_START || trans->aborted);
1786 }
1787
1788 /*
1789  * wait for the current transaction to start and then become unblocked.
1790  * caller holds ref.
1791  */
1792 static void wait_current_trans_commit_start_and_unblock(
1793                                         struct btrfs_fs_info *fs_info,
1794                                         struct btrfs_transaction *trans)
1795 {
1796         wait_event(fs_info->transaction_wait,
1797                    trans->state >= TRANS_STATE_UNBLOCKED || trans->aborted);
1798 }
1799
1800 /*
1801  * commit transactions asynchronously. once btrfs_commit_transaction_async
1802  * returns, any subsequent transaction will not be allowed to join.
1803  */
1804 struct btrfs_async_commit {
1805         struct btrfs_trans_handle *newtrans;
1806         struct work_struct work;
1807 };
1808
1809 static void do_async_commit(struct work_struct *work)
1810 {
1811         struct btrfs_async_commit *ac =
1812                 container_of(work, struct btrfs_async_commit, work);
1813
1814         /*
1815          * We've got freeze protection passed with the transaction.
1816          * Tell lockdep about it.
1817          */
1818         if (ac->newtrans->type & __TRANS_FREEZABLE)
1819                 __sb_writers_acquired(ac->newtrans->fs_info->sb, SB_FREEZE_FS);
1820
1821         current->journal_info = ac->newtrans;
1822
1823         btrfs_commit_transaction(ac->newtrans);
1824         kfree(ac);
1825 }
1826
1827 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1828                                    int wait_for_unblock)
1829 {
1830         struct btrfs_fs_info *fs_info = trans->fs_info;
1831         struct btrfs_async_commit *ac;
1832         struct btrfs_transaction *cur_trans;
1833
1834         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1835         if (!ac)
1836                 return -ENOMEM;
1837
1838         INIT_WORK(&ac->work, do_async_commit);
1839         ac->newtrans = btrfs_join_transaction(trans->root);
1840         if (IS_ERR(ac->newtrans)) {
1841                 int err = PTR_ERR(ac->newtrans);
1842                 kfree(ac);
1843                 return err;
1844         }
1845
1846         /* take transaction reference */
1847         cur_trans = trans->transaction;
1848         refcount_inc(&cur_trans->use_count);
1849
1850         btrfs_end_transaction(trans);
1851
1852         /*
1853          * Tell lockdep we've released the freeze rwsem, since the
1854          * async commit thread will be the one to unlock it.
1855          */
1856         if (ac->newtrans->type & __TRANS_FREEZABLE)
1857                 __sb_writers_release(fs_info->sb, SB_FREEZE_FS);
1858
1859         schedule_work(&ac->work);
1860
1861         /* wait for transaction to start and unblock */
1862         if (wait_for_unblock)
1863                 wait_current_trans_commit_start_and_unblock(fs_info, cur_trans);
1864         else
1865                 wait_current_trans_commit_start(fs_info, cur_trans);
1866
1867         if (current->journal_info == trans)
1868                 current->journal_info = NULL;
1869
1870         btrfs_put_transaction(cur_trans);
1871         return 0;
1872 }
1873
1874
1875 static void cleanup_transaction(struct btrfs_trans_handle *trans,
1876                                 struct btrfs_root *root, int err)
1877 {
1878         struct btrfs_fs_info *fs_info = root->fs_info;
1879         struct btrfs_transaction *cur_trans = trans->transaction;
1880         DEFINE_WAIT(wait);
1881
1882         WARN_ON(refcount_read(&trans->use_count) > 1);
1883
1884         btrfs_abort_transaction(trans, err);
1885
1886         spin_lock(&fs_info->trans_lock);
1887
1888         /*
1889          * If the transaction is removed from the list, it means this
1890          * transaction has been committed successfully, so it is impossible
1891          * to call the cleanup function.
1892          */
1893         BUG_ON(list_empty(&cur_trans->list));
1894
1895         list_del_init(&cur_trans->list);
1896         if (cur_trans == fs_info->running_transaction) {
1897                 cur_trans->state = TRANS_STATE_COMMIT_DOING;
1898                 spin_unlock(&fs_info->trans_lock);
1899                 wait_event(cur_trans->writer_wait,
1900                            atomic_read(&cur_trans->num_writers) == 1);
1901
1902                 spin_lock(&fs_info->trans_lock);
1903         }
1904         spin_unlock(&fs_info->trans_lock);
1905
1906         btrfs_cleanup_one_transaction(trans->transaction, fs_info);
1907
1908         spin_lock(&fs_info->trans_lock);
1909         if (cur_trans == fs_info->running_transaction)
1910                 fs_info->running_transaction = NULL;
1911         spin_unlock(&fs_info->trans_lock);
1912
1913         if (trans->type & __TRANS_FREEZABLE)
1914                 sb_end_intwrite(fs_info->sb);
1915         btrfs_put_transaction(cur_trans);
1916         btrfs_put_transaction(cur_trans);
1917
1918         trace_btrfs_transaction_commit(root);
1919
1920         if (current->journal_info == trans)
1921                 current->journal_info = NULL;
1922         btrfs_scrub_cancel(fs_info);
1923
1924         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1925 }
1926
1927 static inline int btrfs_start_delalloc_flush(struct btrfs_fs_info *fs_info)
1928 {
1929         /*
1930          * We use writeback_inodes_sb here because if we used
1931          * btrfs_start_delalloc_roots we would deadlock with fs freeze.
1932          * Currently are holding the fs freeze lock, if we do an async flush
1933          * we'll do btrfs_join_transaction() and deadlock because we need to
1934          * wait for the fs freeze lock.  Using the direct flushing we benefit
1935          * from already being in a transaction and our join_transaction doesn't
1936          * have to re-take the fs freeze lock.
1937          */
1938         if (btrfs_test_opt(fs_info, FLUSHONCOMMIT))
1939                 writeback_inodes_sb(fs_info->sb, WB_REASON_SYNC);
1940         return 0;
1941 }
1942
1943 static inline void btrfs_wait_delalloc_flush(struct btrfs_fs_info *fs_info)
1944 {
1945         if (btrfs_test_opt(fs_info, FLUSHONCOMMIT))
1946                 btrfs_wait_ordered_roots(fs_info, U64_MAX, 0, (u64)-1);
1947 }
1948
1949 static inline void
1950 btrfs_wait_pending_ordered(struct btrfs_transaction *cur_trans)
1951 {
1952         wait_event(cur_trans->pending_wait,
1953                    atomic_read(&cur_trans->pending_ordered) == 0);
1954 }
1955
1956 int btrfs_commit_transaction(struct btrfs_trans_handle *trans)
1957 {
1958         struct btrfs_fs_info *fs_info = trans->fs_info;
1959         struct btrfs_transaction *cur_trans = trans->transaction;
1960         struct btrfs_transaction *prev_trans = NULL;
1961         int ret;
1962
1963         /* Stop the commit early if ->aborted is set */
1964         if (unlikely(READ_ONCE(cur_trans->aborted))) {
1965                 ret = cur_trans->aborted;
1966                 btrfs_end_transaction(trans);
1967                 return ret;
1968         }
1969
1970         /* make a pass through all the delayed refs we have so far
1971          * any runnings procs may add more while we are here
1972          */
1973         ret = btrfs_run_delayed_refs(trans, fs_info, 0);
1974         if (ret) {
1975                 btrfs_end_transaction(trans);
1976                 return ret;
1977         }
1978
1979         btrfs_trans_release_metadata(trans, fs_info);
1980         trans->block_rsv = NULL;
1981
1982         cur_trans = trans->transaction;
1983
1984         /*
1985          * set the flushing flag so procs in this transaction have to
1986          * start sending their work down.
1987          */
1988         cur_trans->delayed_refs.flushing = 1;
1989         smp_wmb();
1990
1991         if (!list_empty(&trans->new_bgs))
1992                 btrfs_create_pending_block_groups(trans, fs_info);
1993
1994         ret = btrfs_run_delayed_refs(trans, fs_info, 0);
1995         if (ret) {
1996                 btrfs_end_transaction(trans);
1997                 return ret;
1998         }
1999
2000         if (!test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &cur_trans->flags)) {
2001                 int run_it = 0;
2002
2003                 /* this mutex is also taken before trying to set
2004                  * block groups readonly.  We need to make sure
2005                  * that nobody has set a block group readonly
2006                  * after a extents from that block group have been
2007                  * allocated for cache files.  btrfs_set_block_group_ro
2008                  * will wait for the transaction to commit if it
2009                  * finds BTRFS_TRANS_DIRTY_BG_RUN set.
2010                  *
2011                  * The BTRFS_TRANS_DIRTY_BG_RUN flag is also used to make sure
2012                  * only one process starts all the block group IO.  It wouldn't
2013                  * hurt to have more than one go through, but there's no
2014                  * real advantage to it either.
2015                  */
2016                 mutex_lock(&fs_info->ro_block_group_mutex);
2017                 if (!test_and_set_bit(BTRFS_TRANS_DIRTY_BG_RUN,
2018                                       &cur_trans->flags))
2019                         run_it = 1;
2020                 mutex_unlock(&fs_info->ro_block_group_mutex);
2021
2022                 if (run_it)
2023                         ret = btrfs_start_dirty_block_groups(trans, fs_info);
2024         }
2025         if (ret) {
2026                 btrfs_end_transaction(trans);
2027                 return ret;
2028         }
2029
2030         spin_lock(&fs_info->trans_lock);
2031         if (cur_trans->state >= TRANS_STATE_COMMIT_START) {
2032                 spin_unlock(&fs_info->trans_lock);
2033                 refcount_inc(&cur_trans->use_count);
2034                 ret = btrfs_end_transaction(trans);
2035
2036                 wait_for_commit(cur_trans);
2037
2038                 if (unlikely(cur_trans->aborted))
2039                         ret = cur_trans->aborted;
2040
2041                 btrfs_put_transaction(cur_trans);
2042
2043                 return ret;
2044         }
2045
2046         cur_trans->state = TRANS_STATE_COMMIT_START;
2047         wake_up(&fs_info->transaction_blocked_wait);
2048
2049         if (cur_trans->list.prev != &fs_info->trans_list) {
2050                 prev_trans = list_entry(cur_trans->list.prev,
2051                                         struct btrfs_transaction, list);
2052                 if (prev_trans->state != TRANS_STATE_COMPLETED) {
2053                         refcount_inc(&prev_trans->use_count);
2054                         spin_unlock(&fs_info->trans_lock);
2055
2056                         wait_for_commit(prev_trans);
2057                         ret = prev_trans->aborted;
2058
2059                         btrfs_put_transaction(prev_trans);
2060                         if (ret)
2061                                 goto cleanup_transaction;
2062                 } else {
2063                         spin_unlock(&fs_info->trans_lock);
2064                 }
2065         } else {
2066                 spin_unlock(&fs_info->trans_lock);
2067         }
2068
2069         extwriter_counter_dec(cur_trans, trans->type);
2070
2071         ret = btrfs_start_delalloc_flush(fs_info);
2072         if (ret)
2073                 goto cleanup_transaction;
2074
2075         ret = btrfs_run_delayed_items(trans, fs_info);
2076         if (ret)
2077                 goto cleanup_transaction;
2078
2079         wait_event(cur_trans->writer_wait,
2080                    extwriter_counter_read(cur_trans) == 0);
2081
2082         /* some pending stuffs might be added after the previous flush. */
2083         ret = btrfs_run_delayed_items(trans, fs_info);
2084         if (ret)
2085                 goto cleanup_transaction;
2086
2087         btrfs_wait_delalloc_flush(fs_info);
2088
2089         btrfs_wait_pending_ordered(cur_trans);
2090
2091         btrfs_scrub_pause(fs_info);
2092         /*
2093          * Ok now we need to make sure to block out any other joins while we
2094          * commit the transaction.  We could have started a join before setting
2095          * COMMIT_DOING so make sure to wait for num_writers to == 1 again.
2096          */
2097         spin_lock(&fs_info->trans_lock);
2098         cur_trans->state = TRANS_STATE_COMMIT_DOING;
2099         spin_unlock(&fs_info->trans_lock);
2100         wait_event(cur_trans->writer_wait,
2101                    atomic_read(&cur_trans->num_writers) == 1);
2102
2103         /* ->aborted might be set after the previous check, so check it */
2104         if (unlikely(READ_ONCE(cur_trans->aborted))) {
2105                 ret = cur_trans->aborted;
2106                 goto scrub_continue;
2107         }
2108         /*
2109          * the reloc mutex makes sure that we stop
2110          * the balancing code from coming in and moving
2111          * extents around in the middle of the commit
2112          */
2113         mutex_lock(&fs_info->reloc_mutex);
2114
2115         /*
2116          * We needn't worry about the delayed items because we will
2117          * deal with them in create_pending_snapshot(), which is the
2118          * core function of the snapshot creation.
2119          */
2120         ret = create_pending_snapshots(trans, fs_info);
2121         if (ret) {
2122                 mutex_unlock(&fs_info->reloc_mutex);
2123                 goto scrub_continue;
2124         }
2125
2126         /*
2127          * We insert the dir indexes of the snapshots and update the inode
2128          * of the snapshots' parents after the snapshot creation, so there
2129          * are some delayed items which are not dealt with. Now deal with
2130          * them.
2131          *
2132          * We needn't worry that this operation will corrupt the snapshots,
2133          * because all the tree which are snapshoted will be forced to COW
2134          * the nodes and leaves.
2135          */
2136         ret = btrfs_run_delayed_items(trans, fs_info);
2137         if (ret) {
2138                 mutex_unlock(&fs_info->reloc_mutex);
2139                 goto scrub_continue;
2140         }
2141
2142         ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
2143         if (ret) {
2144                 mutex_unlock(&fs_info->reloc_mutex);
2145                 goto scrub_continue;
2146         }
2147
2148         /*
2149          * make sure none of the code above managed to slip in a
2150          * delayed item
2151          */
2152         btrfs_assert_delayed_root_empty(fs_info);
2153
2154         WARN_ON(cur_trans != trans->transaction);
2155
2156         /* btrfs_commit_tree_roots is responsible for getting the
2157          * various roots consistent with each other.  Every pointer
2158          * in the tree of tree roots has to point to the most up to date
2159          * root for every subvolume and other tree.  So, we have to keep
2160          * the tree logging code from jumping in and changing any
2161          * of the trees.
2162          *
2163          * At this point in the commit, there can't be any tree-log
2164          * writers, but a little lower down we drop the trans mutex
2165          * and let new people in.  By holding the tree_log_mutex
2166          * from now until after the super is written, we avoid races
2167          * with the tree-log code.
2168          */
2169         mutex_lock(&fs_info->tree_log_mutex);
2170
2171         ret = commit_fs_roots(trans, fs_info);
2172         if (ret) {
2173                 mutex_unlock(&fs_info->tree_log_mutex);
2174                 mutex_unlock(&fs_info->reloc_mutex);
2175                 goto scrub_continue;
2176         }
2177
2178         /*
2179          * Since the transaction is done, we can apply the pending changes
2180          * before the next transaction.
2181          */
2182         btrfs_apply_pending_changes(fs_info);
2183
2184         /* commit_fs_roots gets rid of all the tree log roots, it is now
2185          * safe to free the root of tree log roots
2186          */
2187         btrfs_free_log_root_tree(trans, fs_info);
2188
2189         /*
2190          * commit_fs_roots() can call btrfs_save_ino_cache(), which generates
2191          * new delayed refs. Must handle them or qgroup can be wrong.
2192          */
2193         ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
2194         if (ret) {
2195                 mutex_unlock(&fs_info->tree_log_mutex);
2196                 mutex_unlock(&fs_info->reloc_mutex);
2197                 goto scrub_continue;
2198         }
2199
2200         /*
2201          * Since fs roots are all committed, we can get a quite accurate
2202          * new_roots. So let's do quota accounting.
2203          */
2204         ret = btrfs_qgroup_account_extents(trans, fs_info);
2205         if (ret < 0) {
2206                 mutex_unlock(&fs_info->tree_log_mutex);
2207                 mutex_unlock(&fs_info->reloc_mutex);
2208                 goto scrub_continue;
2209         }
2210
2211         ret = commit_cowonly_roots(trans, fs_info);
2212         if (ret) {
2213                 mutex_unlock(&fs_info->tree_log_mutex);
2214                 mutex_unlock(&fs_info->reloc_mutex);
2215                 goto scrub_continue;
2216         }
2217
2218         /*
2219          * The tasks which save the space cache and inode cache may also
2220          * update ->aborted, check it.
2221          */
2222         if (unlikely(READ_ONCE(cur_trans->aborted))) {
2223                 ret = cur_trans->aborted;
2224                 mutex_unlock(&fs_info->tree_log_mutex);
2225                 mutex_unlock(&fs_info->reloc_mutex);
2226                 goto scrub_continue;
2227         }
2228
2229         btrfs_prepare_extent_commit(fs_info);
2230
2231         cur_trans = fs_info->running_transaction;
2232
2233         btrfs_set_root_node(&fs_info->tree_root->root_item,
2234                             fs_info->tree_root->node);
2235         list_add_tail(&fs_info->tree_root->dirty_list,
2236                       &cur_trans->switch_commits);
2237
2238         btrfs_set_root_node(&fs_info->chunk_root->root_item,
2239                             fs_info->chunk_root->node);
2240         list_add_tail(&fs_info->chunk_root->dirty_list,
2241                       &cur_trans->switch_commits);
2242
2243         switch_commit_roots(cur_trans, fs_info);
2244
2245         ASSERT(list_empty(&cur_trans->dirty_bgs));
2246         ASSERT(list_empty(&cur_trans->io_bgs));
2247         update_super_roots(fs_info);
2248
2249         btrfs_set_super_log_root(fs_info->super_copy, 0);
2250         btrfs_set_super_log_root_level(fs_info->super_copy, 0);
2251         memcpy(fs_info->super_for_commit, fs_info->super_copy,
2252                sizeof(*fs_info->super_copy));
2253
2254         btrfs_update_commit_device_size(fs_info);
2255         btrfs_update_commit_device_bytes_used(fs_info, cur_trans);
2256
2257         clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags);
2258         clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags);
2259
2260         btrfs_trans_release_chunk_metadata(trans);
2261
2262         spin_lock(&fs_info->trans_lock);
2263         cur_trans->state = TRANS_STATE_UNBLOCKED;
2264         fs_info->running_transaction = NULL;
2265         spin_unlock(&fs_info->trans_lock);
2266         mutex_unlock(&fs_info->reloc_mutex);
2267
2268         wake_up(&fs_info->transaction_wait);
2269
2270         ret = btrfs_write_and_wait_transaction(trans, fs_info);
2271         if (ret) {
2272                 btrfs_handle_fs_error(fs_info, ret,
2273                                       "Error while writing out transaction");
2274                 mutex_unlock(&fs_info->tree_log_mutex);
2275                 goto scrub_continue;
2276         }
2277
2278         ret = write_all_supers(fs_info, 0);
2279         /*
2280          * the super is written, we can safely allow the tree-loggers
2281          * to go about their business
2282          */
2283         mutex_unlock(&fs_info->tree_log_mutex);
2284         if (ret)
2285                 goto scrub_continue;
2286
2287         btrfs_finish_extent_commit(trans, fs_info);
2288
2289         if (test_bit(BTRFS_TRANS_HAVE_FREE_BGS, &cur_trans->flags))
2290                 btrfs_clear_space_info_full(fs_info);
2291
2292         fs_info->last_trans_committed = cur_trans->transid;
2293         /*
2294          * We needn't acquire the lock here because there is no other task
2295          * which can change it.
2296          */
2297         cur_trans->state = TRANS_STATE_COMPLETED;
2298         wake_up(&cur_trans->commit_wait);
2299
2300         spin_lock(&fs_info->trans_lock);
2301         list_del_init(&cur_trans->list);
2302         spin_unlock(&fs_info->trans_lock);
2303
2304         btrfs_put_transaction(cur_trans);
2305         btrfs_put_transaction(cur_trans);
2306
2307         if (trans->type & __TRANS_FREEZABLE)
2308                 sb_end_intwrite(fs_info->sb);
2309
2310         trace_btrfs_transaction_commit(trans->root);
2311
2312         btrfs_scrub_continue(fs_info);
2313
2314         if (current->journal_info == trans)
2315                 current->journal_info = NULL;
2316
2317         kmem_cache_free(btrfs_trans_handle_cachep, trans);
2318
2319         /*
2320          * If fs has been frozen, we can not handle delayed iputs, otherwise
2321          * it'll result in deadlock about SB_FREEZE_FS.
2322          */
2323         if (current != fs_info->transaction_kthread &&
2324             current != fs_info->cleaner_kthread &&
2325             !test_bit(BTRFS_FS_FROZEN, &fs_info->flags))
2326                 btrfs_run_delayed_iputs(fs_info);
2327
2328         return ret;
2329
2330 scrub_continue:
2331         btrfs_scrub_continue(fs_info);
2332 cleanup_transaction:
2333         btrfs_trans_release_metadata(trans, fs_info);
2334         btrfs_trans_release_chunk_metadata(trans);
2335         trans->block_rsv = NULL;
2336         btrfs_warn(fs_info, "Skipping commit of aborted transaction.");
2337         if (current->journal_info == trans)
2338                 current->journal_info = NULL;
2339         cleanup_transaction(trans, trans->root, ret);
2340
2341         return ret;
2342 }
2343
2344 /*
2345  * return < 0 if error
2346  * 0 if there are no more dead_roots at the time of call
2347  * 1 there are more to be processed, call me again
2348  *
2349  * The return value indicates there are certainly more snapshots to delete, but
2350  * if there comes a new one during processing, it may return 0. We don't mind,
2351  * because btrfs_commit_super will poke cleaner thread and it will process it a
2352  * few seconds later.
2353  */
2354 int btrfs_clean_one_deleted_snapshot(struct btrfs_root *root)
2355 {
2356         int ret;
2357         struct btrfs_fs_info *fs_info = root->fs_info;
2358
2359         spin_lock(&fs_info->trans_lock);
2360         if (list_empty(&fs_info->dead_roots)) {
2361                 spin_unlock(&fs_info->trans_lock);
2362                 return 0;
2363         }
2364         root = list_first_entry(&fs_info->dead_roots,
2365                         struct btrfs_root, root_list);
2366         list_del_init(&root->root_list);
2367         spin_unlock(&fs_info->trans_lock);
2368
2369         btrfs_debug(fs_info, "cleaner removing %llu", root->objectid);
2370
2371         btrfs_kill_all_delayed_nodes(root);
2372
2373         if (btrfs_header_backref_rev(root->node) <
2374                         BTRFS_MIXED_BACKREF_REV)
2375                 ret = btrfs_drop_snapshot(root, NULL, 0, 0);
2376         else
2377                 ret = btrfs_drop_snapshot(root, NULL, 1, 0);
2378
2379         return (ret < 0) ? 0 : 1;
2380 }
2381
2382 void btrfs_apply_pending_changes(struct btrfs_fs_info *fs_info)
2383 {
2384         unsigned long prev;
2385         unsigned long bit;
2386
2387         prev = xchg(&fs_info->pending_changes, 0);
2388         if (!prev)
2389                 return;
2390
2391         bit = 1 << BTRFS_PENDING_SET_INODE_MAP_CACHE;
2392         if (prev & bit)
2393                 btrfs_set_opt(fs_info->mount_opt, INODE_MAP_CACHE);
2394         prev &= ~bit;
2395
2396         bit = 1 << BTRFS_PENDING_CLEAR_INODE_MAP_CACHE;
2397         if (prev & bit)
2398                 btrfs_clear_opt(fs_info->mount_opt, INODE_MAP_CACHE);
2399         prev &= ~bit;
2400
2401         bit = 1 << BTRFS_PENDING_COMMIT;
2402         if (prev & bit)
2403                 btrfs_debug(fs_info, "pending commit done");
2404         prev &= ~bit;
2405
2406         if (prev)
2407                 btrfs_warn(fs_info,
2408                         "unknown pending changes left 0x%lx, ignoring", prev);
2409 }