Merge tag 'pci-v4.4-fixes-1' of git://git.kernel.org/pub/scm/linux/kernel/git/helgaas/pci
[firefly-linux-kernel-4.4.55.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(atomic_read(&transaction->use_count) == 0);
64         if (atomic_dec_and_test(&transaction->use_count)) {
65                 BUG_ON(!list_empty(&transaction->list));
66                 WARN_ON(!RB_EMPTY_ROOT(&transaction->delayed_refs.href_root));
67                 if (transaction->delayed_refs.pending_csums)
68                         printk(KERN_ERR "pending csums is %llu\n",
69                                transaction->delayed_refs.pending_csums);
70                 while (!list_empty(&transaction->pending_chunks)) {
71                         struct extent_map *em;
72
73                         em = list_first_entry(&transaction->pending_chunks,
74                                               struct extent_map, list);
75                         list_del_init(&em->list);
76                         free_extent_map(em);
77                 }
78                 kmem_cache_free(btrfs_transaction_cachep, transaction);
79         }
80 }
81
82 static void clear_btree_io_tree(struct extent_io_tree *tree)
83 {
84         spin_lock(&tree->lock);
85         /*
86          * Do a single barrier for the waitqueue_active check here, the state
87          * of the waitqueue should not change once clear_btree_io_tree is
88          * called.
89          */
90         smp_mb();
91         while (!RB_EMPTY_ROOT(&tree->state)) {
92                 struct rb_node *node;
93                 struct extent_state *state;
94
95                 node = rb_first(&tree->state);
96                 state = rb_entry(node, struct extent_state, rb_node);
97                 rb_erase(&state->rb_node, &tree->state);
98                 RB_CLEAR_NODE(&state->rb_node);
99                 /*
100                  * btree io trees aren't supposed to have tasks waiting for
101                  * changes in the flags of extent states ever.
102                  */
103                 ASSERT(!waitqueue_active(&state->wq));
104                 free_extent_state(state);
105
106                 cond_resched_lock(&tree->lock);
107         }
108         spin_unlock(&tree->lock);
109 }
110
111 static noinline void switch_commit_roots(struct btrfs_transaction *trans,
112                                          struct btrfs_fs_info *fs_info)
113 {
114         struct btrfs_root *root, *tmp;
115
116         down_write(&fs_info->commit_root_sem);
117         list_for_each_entry_safe(root, tmp, &trans->switch_commits,
118                                  dirty_list) {
119                 list_del_init(&root->dirty_list);
120                 free_extent_buffer(root->commit_root);
121                 root->commit_root = btrfs_root_node(root);
122                 if (is_fstree(root->objectid))
123                         btrfs_unpin_free_ino(root);
124                 clear_btree_io_tree(&root->dirty_log_pages);
125         }
126
127         /* We can free old roots now. */
128         spin_lock(&trans->dropped_roots_lock);
129         while (!list_empty(&trans->dropped_roots)) {
130                 root = list_first_entry(&trans->dropped_roots,
131                                         struct btrfs_root, root_list);
132                 list_del_init(&root->root_list);
133                 spin_unlock(&trans->dropped_roots_lock);
134                 btrfs_drop_and_free_fs_root(fs_info, root);
135                 spin_lock(&trans->dropped_roots_lock);
136         }
137         spin_unlock(&trans->dropped_roots_lock);
138         up_write(&fs_info->commit_root_sem);
139 }
140
141 static inline void extwriter_counter_inc(struct btrfs_transaction *trans,
142                                          unsigned int type)
143 {
144         if (type & TRANS_EXTWRITERS)
145                 atomic_inc(&trans->num_extwriters);
146 }
147
148 static inline void extwriter_counter_dec(struct btrfs_transaction *trans,
149                                          unsigned int type)
150 {
151         if (type & TRANS_EXTWRITERS)
152                 atomic_dec(&trans->num_extwriters);
153 }
154
155 static inline void extwriter_counter_init(struct btrfs_transaction *trans,
156                                           unsigned int type)
157 {
158         atomic_set(&trans->num_extwriters, ((type & TRANS_EXTWRITERS) ? 1 : 0));
159 }
160
161 static inline int extwriter_counter_read(struct btrfs_transaction *trans)
162 {
163         return atomic_read(&trans->num_extwriters);
164 }
165
166 /*
167  * either allocate a new transaction or hop into the existing one
168  */
169 static noinline int join_transaction(struct btrfs_root *root, unsigned int type)
170 {
171         struct btrfs_transaction *cur_trans;
172         struct btrfs_fs_info *fs_info = root->fs_info;
173
174         spin_lock(&fs_info->trans_lock);
175 loop:
176         /* The file system has been taken offline. No new transactions. */
177         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
178                 spin_unlock(&fs_info->trans_lock);
179                 return -EROFS;
180         }
181
182         cur_trans = fs_info->running_transaction;
183         if (cur_trans) {
184                 if (cur_trans->aborted) {
185                         spin_unlock(&fs_info->trans_lock);
186                         return cur_trans->aborted;
187                 }
188                 if (btrfs_blocked_trans_types[cur_trans->state] & type) {
189                         spin_unlock(&fs_info->trans_lock);
190                         return -EBUSY;
191                 }
192                 atomic_inc(&cur_trans->use_count);
193                 atomic_inc(&cur_trans->num_writers);
194                 extwriter_counter_inc(cur_trans, type);
195                 spin_unlock(&fs_info->trans_lock);
196                 return 0;
197         }
198         spin_unlock(&fs_info->trans_lock);
199
200         /*
201          * If we are ATTACH, we just want to catch the current transaction,
202          * and commit it. If there is no transaction, just return ENOENT.
203          */
204         if (type == TRANS_ATTACH)
205                 return -ENOENT;
206
207         /*
208          * JOIN_NOLOCK only happens during the transaction commit, so
209          * it is impossible that ->running_transaction is NULL
210          */
211         BUG_ON(type == TRANS_JOIN_NOLOCK);
212
213         cur_trans = kmem_cache_alloc(btrfs_transaction_cachep, GFP_NOFS);
214         if (!cur_trans)
215                 return -ENOMEM;
216
217         spin_lock(&fs_info->trans_lock);
218         if (fs_info->running_transaction) {
219                 /*
220                  * someone started a transaction after we unlocked.  Make sure
221                  * to redo the checks above
222                  */
223                 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
224                 goto loop;
225         } else if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
226                 spin_unlock(&fs_info->trans_lock);
227                 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
228                 return -EROFS;
229         }
230
231         atomic_set(&cur_trans->num_writers, 1);
232         extwriter_counter_init(cur_trans, type);
233         init_waitqueue_head(&cur_trans->writer_wait);
234         init_waitqueue_head(&cur_trans->commit_wait);
235         init_waitqueue_head(&cur_trans->pending_wait);
236         cur_trans->state = TRANS_STATE_RUNNING;
237         /*
238          * One for this trans handle, one so it will live on until we
239          * commit the transaction.
240          */
241         atomic_set(&cur_trans->use_count, 2);
242         atomic_set(&cur_trans->pending_ordered, 0);
243         cur_trans->flags = 0;
244         cur_trans->start_time = get_seconds();
245
246         memset(&cur_trans->delayed_refs, 0, sizeof(cur_trans->delayed_refs));
247
248         cur_trans->delayed_refs.href_root = RB_ROOT;
249         cur_trans->delayed_refs.dirty_extent_root = RB_ROOT;
250         atomic_set(&cur_trans->delayed_refs.num_entries, 0);
251
252         /*
253          * although the tree mod log is per file system and not per transaction,
254          * the log must never go across transaction boundaries.
255          */
256         smp_mb();
257         if (!list_empty(&fs_info->tree_mod_seq_list))
258                 WARN(1, KERN_ERR "BTRFS: tree_mod_seq_list not empty when "
259                         "creating a fresh transaction\n");
260         if (!RB_EMPTY_ROOT(&fs_info->tree_mod_log))
261                 WARN(1, KERN_ERR "BTRFS: tree_mod_log rb tree not empty when "
262                         "creating a fresh transaction\n");
263         atomic64_set(&fs_info->tree_mod_seq, 0);
264
265         spin_lock_init(&cur_trans->delayed_refs.lock);
266
267         INIT_LIST_HEAD(&cur_trans->pending_snapshots);
268         INIT_LIST_HEAD(&cur_trans->pending_chunks);
269         INIT_LIST_HEAD(&cur_trans->switch_commits);
270         INIT_LIST_HEAD(&cur_trans->dirty_bgs);
271         INIT_LIST_HEAD(&cur_trans->io_bgs);
272         INIT_LIST_HEAD(&cur_trans->dropped_roots);
273         mutex_init(&cur_trans->cache_write_mutex);
274         cur_trans->num_dirty_bgs = 0;
275         spin_lock_init(&cur_trans->dirty_bgs_lock);
276         INIT_LIST_HEAD(&cur_trans->deleted_bgs);
277         spin_lock_init(&cur_trans->deleted_bgs_lock);
278         spin_lock_init(&cur_trans->dropped_roots_lock);
279         list_add_tail(&cur_trans->list, &fs_info->trans_list);
280         extent_io_tree_init(&cur_trans->dirty_pages,
281                              fs_info->btree_inode->i_mapping);
282         fs_info->generation++;
283         cur_trans->transid = fs_info->generation;
284         fs_info->running_transaction = cur_trans;
285         cur_trans->aborted = 0;
286         spin_unlock(&fs_info->trans_lock);
287
288         return 0;
289 }
290
291 /*
292  * this does all the record keeping required to make sure that a reference
293  * counted root is properly recorded in a given transaction.  This is required
294  * to make sure the old root from before we joined the transaction is deleted
295  * when the transaction commits
296  */
297 static int record_root_in_trans(struct btrfs_trans_handle *trans,
298                                struct btrfs_root *root)
299 {
300         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
301             root->last_trans < trans->transid) {
302                 WARN_ON(root == root->fs_info->extent_root);
303                 WARN_ON(root->commit_root != root->node);
304
305                 /*
306                  * see below for IN_TRANS_SETUP usage rules
307                  * we have the reloc mutex held now, so there
308                  * is only one writer in this function
309                  */
310                 set_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
311
312                 /* make sure readers find IN_TRANS_SETUP before
313                  * they find our root->last_trans update
314                  */
315                 smp_wmb();
316
317                 spin_lock(&root->fs_info->fs_roots_radix_lock);
318                 if (root->last_trans == trans->transid) {
319                         spin_unlock(&root->fs_info->fs_roots_radix_lock);
320                         return 0;
321                 }
322                 radix_tree_tag_set(&root->fs_info->fs_roots_radix,
323                            (unsigned long)root->root_key.objectid,
324                            BTRFS_ROOT_TRANS_TAG);
325                 spin_unlock(&root->fs_info->fs_roots_radix_lock);
326                 root->last_trans = trans->transid;
327
328                 /* this is pretty tricky.  We don't want to
329                  * take the relocation lock in btrfs_record_root_in_trans
330                  * unless we're really doing the first setup for this root in
331                  * this transaction.
332                  *
333                  * Normally we'd use root->last_trans as a flag to decide
334                  * if we want to take the expensive mutex.
335                  *
336                  * But, we have to set root->last_trans before we
337                  * init the relocation root, otherwise, we trip over warnings
338                  * in ctree.c.  The solution used here is to flag ourselves
339                  * with root IN_TRANS_SETUP.  When this is 1, we're still
340                  * fixing up the reloc trees and everyone must wait.
341                  *
342                  * When this is zero, they can trust root->last_trans and fly
343                  * through btrfs_record_root_in_trans without having to take the
344                  * lock.  smp_wmb() makes sure that all the writes above are
345                  * done before we pop in the zero below
346                  */
347                 btrfs_init_reloc_root(trans, root);
348                 smp_mb__before_atomic();
349                 clear_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
350         }
351         return 0;
352 }
353
354
355 void btrfs_add_dropped_root(struct btrfs_trans_handle *trans,
356                             struct btrfs_root *root)
357 {
358         struct btrfs_transaction *cur_trans = trans->transaction;
359
360         /* Add ourselves to the transaction dropped list */
361         spin_lock(&cur_trans->dropped_roots_lock);
362         list_add_tail(&root->root_list, &cur_trans->dropped_roots);
363         spin_unlock(&cur_trans->dropped_roots_lock);
364
365         /* Make sure we don't try to update the root at commit time */
366         spin_lock(&root->fs_info->fs_roots_radix_lock);
367         radix_tree_tag_clear(&root->fs_info->fs_roots_radix,
368                              (unsigned long)root->root_key.objectid,
369                              BTRFS_ROOT_TRANS_TAG);
370         spin_unlock(&root->fs_info->fs_roots_radix_lock);
371 }
372
373 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
374                                struct btrfs_root *root)
375 {
376         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
377                 return 0;
378
379         /*
380          * see record_root_in_trans for comments about IN_TRANS_SETUP usage
381          * and barriers
382          */
383         smp_rmb();
384         if (root->last_trans == trans->transid &&
385             !test_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state))
386                 return 0;
387
388         mutex_lock(&root->fs_info->reloc_mutex);
389         record_root_in_trans(trans, root);
390         mutex_unlock(&root->fs_info->reloc_mutex);
391
392         return 0;
393 }
394
395 static inline int is_transaction_blocked(struct btrfs_transaction *trans)
396 {
397         return (trans->state >= TRANS_STATE_BLOCKED &&
398                 trans->state < TRANS_STATE_UNBLOCKED &&
399                 !trans->aborted);
400 }
401
402 /* wait for commit against the current transaction to become unblocked
403  * when this is done, it is safe to start a new transaction, but the current
404  * transaction might not be fully on disk.
405  */
406 static void wait_current_trans(struct btrfs_root *root)
407 {
408         struct btrfs_transaction *cur_trans;
409
410         spin_lock(&root->fs_info->trans_lock);
411         cur_trans = root->fs_info->running_transaction;
412         if (cur_trans && is_transaction_blocked(cur_trans)) {
413                 atomic_inc(&cur_trans->use_count);
414                 spin_unlock(&root->fs_info->trans_lock);
415
416                 wait_event(root->fs_info->transaction_wait,
417                            cur_trans->state >= TRANS_STATE_UNBLOCKED ||
418                            cur_trans->aborted);
419                 btrfs_put_transaction(cur_trans);
420         } else {
421                 spin_unlock(&root->fs_info->trans_lock);
422         }
423 }
424
425 static int may_wait_transaction(struct btrfs_root *root, int type)
426 {
427         if (root->fs_info->log_root_recovering)
428                 return 0;
429
430         if (type == TRANS_USERSPACE)
431                 return 1;
432
433         if (type == TRANS_START &&
434             !atomic_read(&root->fs_info->open_ioctl_trans))
435                 return 1;
436
437         return 0;
438 }
439
440 static inline bool need_reserve_reloc_root(struct btrfs_root *root)
441 {
442         if (!root->fs_info->reloc_ctl ||
443             !test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
444             root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
445             root->reloc_root)
446                 return false;
447
448         return true;
449 }
450
451 static struct btrfs_trans_handle *
452 start_transaction(struct btrfs_root *root, unsigned int num_items,
453                   unsigned int type, enum btrfs_reserve_flush_enum flush)
454 {
455         struct btrfs_trans_handle *h;
456         struct btrfs_transaction *cur_trans;
457         u64 num_bytes = 0;
458         u64 qgroup_reserved = 0;
459         bool reloc_reserved = false;
460         int ret;
461
462         /* Send isn't supposed to start transactions. */
463         ASSERT(current->journal_info != BTRFS_SEND_TRANS_STUB);
464
465         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
466                 return ERR_PTR(-EROFS);
467
468         if (current->journal_info) {
469                 WARN_ON(type & TRANS_EXTWRITERS);
470                 h = current->journal_info;
471                 h->use_count++;
472                 WARN_ON(h->use_count > 2);
473                 h->orig_rsv = h->block_rsv;
474                 h->block_rsv = NULL;
475                 goto got_it;
476         }
477
478         /*
479          * Do the reservation before we join the transaction so we can do all
480          * the appropriate flushing if need be.
481          */
482         if (num_items > 0 && root != root->fs_info->chunk_root) {
483                 qgroup_reserved = num_items * root->nodesize;
484                 ret = btrfs_qgroup_reserve_meta(root, qgroup_reserved);
485                 if (ret)
486                         return ERR_PTR(ret);
487
488                 num_bytes = btrfs_calc_trans_metadata_size(root, num_items);
489                 /*
490                  * Do the reservation for the relocation root creation
491                  */
492                 if (need_reserve_reloc_root(root)) {
493                         num_bytes += root->nodesize;
494                         reloc_reserved = true;
495                 }
496
497                 ret = btrfs_block_rsv_add(root,
498                                           &root->fs_info->trans_block_rsv,
499                                           num_bytes, flush);
500                 if (ret)
501                         goto reserve_fail;
502         }
503 again:
504         h = kmem_cache_zalloc(btrfs_trans_handle_cachep, GFP_NOFS);
505         if (!h) {
506                 ret = -ENOMEM;
507                 goto alloc_fail;
508         }
509
510         /*
511          * If we are JOIN_NOLOCK we're already committing a transaction and
512          * waiting on this guy, so we don't need to do the sb_start_intwrite
513          * because we're already holding a ref.  We need this because we could
514          * have raced in and did an fsync() on a file which can kick a commit
515          * and then we deadlock with somebody doing a freeze.
516          *
517          * If we are ATTACH, it means we just want to catch the current
518          * transaction and commit it, so we needn't do sb_start_intwrite(). 
519          */
520         if (type & __TRANS_FREEZABLE)
521                 sb_start_intwrite(root->fs_info->sb);
522
523         if (may_wait_transaction(root, type))
524                 wait_current_trans(root);
525
526         do {
527                 ret = join_transaction(root, type);
528                 if (ret == -EBUSY) {
529                         wait_current_trans(root);
530                         if (unlikely(type == TRANS_ATTACH))
531                                 ret = -ENOENT;
532                 }
533         } while (ret == -EBUSY);
534
535         if (ret < 0) {
536                 /* We must get the transaction if we are JOIN_NOLOCK. */
537                 BUG_ON(type == TRANS_JOIN_NOLOCK);
538                 goto join_fail;
539         }
540
541         cur_trans = root->fs_info->running_transaction;
542
543         h->transid = cur_trans->transid;
544         h->transaction = cur_trans;
545         h->root = root;
546         h->use_count = 1;
547
548         h->type = type;
549         h->can_flush_pending_bgs = true;
550         INIT_LIST_HEAD(&h->qgroup_ref_list);
551         INIT_LIST_HEAD(&h->new_bgs);
552
553         smp_mb();
554         if (cur_trans->state >= TRANS_STATE_BLOCKED &&
555             may_wait_transaction(root, type)) {
556                 current->journal_info = h;
557                 btrfs_commit_transaction(h, root);
558                 goto again;
559         }
560
561         if (num_bytes) {
562                 trace_btrfs_space_reservation(root->fs_info, "transaction",
563                                               h->transid, num_bytes, 1);
564                 h->block_rsv = &root->fs_info->trans_block_rsv;
565                 h->bytes_reserved = num_bytes;
566                 h->reloc_reserved = reloc_reserved;
567         }
568
569 got_it:
570         btrfs_record_root_in_trans(h, root);
571
572         if (!current->journal_info && type != TRANS_USERSPACE)
573                 current->journal_info = h;
574         return h;
575
576 join_fail:
577         if (type & __TRANS_FREEZABLE)
578                 sb_end_intwrite(root->fs_info->sb);
579         kmem_cache_free(btrfs_trans_handle_cachep, h);
580 alloc_fail:
581         if (num_bytes)
582                 btrfs_block_rsv_release(root, &root->fs_info->trans_block_rsv,
583                                         num_bytes);
584 reserve_fail:
585         btrfs_qgroup_free_meta(root, qgroup_reserved);
586         return ERR_PTR(ret);
587 }
588
589 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
590                                                    unsigned int num_items)
591 {
592         return start_transaction(root, num_items, TRANS_START,
593                                  BTRFS_RESERVE_FLUSH_ALL);
594 }
595 struct btrfs_trans_handle *btrfs_start_transaction_fallback_global_rsv(
596                                         struct btrfs_root *root,
597                                         unsigned int num_items,
598                                         int min_factor)
599 {
600         struct btrfs_trans_handle *trans;
601         u64 num_bytes;
602         int ret;
603
604         trans = btrfs_start_transaction(root, num_items);
605         if (!IS_ERR(trans) || PTR_ERR(trans) != -ENOSPC)
606                 return trans;
607
608         trans = btrfs_start_transaction(root, 0);
609         if (IS_ERR(trans))
610                 return trans;
611
612         num_bytes = btrfs_calc_trans_metadata_size(root, num_items);
613         ret = btrfs_cond_migrate_bytes(root->fs_info,
614                                        &root->fs_info->trans_block_rsv,
615                                        num_bytes,
616                                        min_factor);
617         if (ret) {
618                 btrfs_end_transaction(trans, root);
619                 return ERR_PTR(ret);
620         }
621
622         trans->block_rsv = &root->fs_info->trans_block_rsv;
623         trans->bytes_reserved = num_bytes;
624
625         return trans;
626 }
627
628 struct btrfs_trans_handle *btrfs_start_transaction_lflush(
629                                         struct btrfs_root *root,
630                                         unsigned int num_items)
631 {
632         return start_transaction(root, num_items, TRANS_START,
633                                  BTRFS_RESERVE_FLUSH_LIMIT);
634 }
635
636 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
637 {
638         return start_transaction(root, 0, TRANS_JOIN, 0);
639 }
640
641 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root)
642 {
643         return start_transaction(root, 0, TRANS_JOIN_NOLOCK, 0);
644 }
645
646 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *root)
647 {
648         return start_transaction(root, 0, TRANS_USERSPACE, 0);
649 }
650
651 /*
652  * btrfs_attach_transaction() - catch the running transaction
653  *
654  * It is used when we want to commit the current the transaction, but
655  * don't want to start a new one.
656  *
657  * Note: If this function return -ENOENT, it just means there is no
658  * running transaction. But it is possible that the inactive transaction
659  * is still in the memory, not fully on disk. If you hope there is no
660  * inactive transaction in the fs when -ENOENT is returned, you should
661  * invoke
662  *     btrfs_attach_transaction_barrier()
663  */
664 struct btrfs_trans_handle *btrfs_attach_transaction(struct btrfs_root *root)
665 {
666         return start_transaction(root, 0, TRANS_ATTACH, 0);
667 }
668
669 /*
670  * btrfs_attach_transaction_barrier() - catch the running transaction
671  *
672  * It is similar to the above function, the differentia is this one
673  * will wait for all the inactive transactions until they fully
674  * complete.
675  */
676 struct btrfs_trans_handle *
677 btrfs_attach_transaction_barrier(struct btrfs_root *root)
678 {
679         struct btrfs_trans_handle *trans;
680
681         trans = start_transaction(root, 0, TRANS_ATTACH, 0);
682         if (IS_ERR(trans) && PTR_ERR(trans) == -ENOENT)
683                 btrfs_wait_for_commit(root, 0);
684
685         return trans;
686 }
687
688 /* wait for a transaction commit to be fully complete */
689 static noinline void wait_for_commit(struct btrfs_root *root,
690                                     struct btrfs_transaction *commit)
691 {
692         wait_event(commit->commit_wait, commit->state == TRANS_STATE_COMPLETED);
693 }
694
695 int btrfs_wait_for_commit(struct btrfs_root *root, u64 transid)
696 {
697         struct btrfs_transaction *cur_trans = NULL, *t;
698         int ret = 0;
699
700         if (transid) {
701                 if (transid <= root->fs_info->last_trans_committed)
702                         goto out;
703
704                 /* find specified transaction */
705                 spin_lock(&root->fs_info->trans_lock);
706                 list_for_each_entry(t, &root->fs_info->trans_list, list) {
707                         if (t->transid == transid) {
708                                 cur_trans = t;
709                                 atomic_inc(&cur_trans->use_count);
710                                 ret = 0;
711                                 break;
712                         }
713                         if (t->transid > transid) {
714                                 ret = 0;
715                                 break;
716                         }
717                 }
718                 spin_unlock(&root->fs_info->trans_lock);
719
720                 /*
721                  * The specified transaction doesn't exist, or we
722                  * raced with btrfs_commit_transaction
723                  */
724                 if (!cur_trans) {
725                         if (transid > root->fs_info->last_trans_committed)
726                                 ret = -EINVAL;
727                         goto out;
728                 }
729         } else {
730                 /* find newest transaction that is committing | committed */
731                 spin_lock(&root->fs_info->trans_lock);
732                 list_for_each_entry_reverse(t, &root->fs_info->trans_list,
733                                             list) {
734                         if (t->state >= TRANS_STATE_COMMIT_START) {
735                                 if (t->state == TRANS_STATE_COMPLETED)
736                                         break;
737                                 cur_trans = t;
738                                 atomic_inc(&cur_trans->use_count);
739                                 break;
740                         }
741                 }
742                 spin_unlock(&root->fs_info->trans_lock);
743                 if (!cur_trans)
744                         goto out;  /* nothing committing|committed */
745         }
746
747         wait_for_commit(root, cur_trans);
748         btrfs_put_transaction(cur_trans);
749 out:
750         return ret;
751 }
752
753 void btrfs_throttle(struct btrfs_root *root)
754 {
755         if (!atomic_read(&root->fs_info->open_ioctl_trans))
756                 wait_current_trans(root);
757 }
758
759 static int should_end_transaction(struct btrfs_trans_handle *trans,
760                                   struct btrfs_root *root)
761 {
762         if (root->fs_info->global_block_rsv.space_info->full &&
763             btrfs_check_space_for_delayed_refs(trans, root))
764                 return 1;
765
766         return !!btrfs_block_rsv_check(root, &root->fs_info->global_block_rsv, 5);
767 }
768
769 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans,
770                                  struct btrfs_root *root)
771 {
772         struct btrfs_transaction *cur_trans = trans->transaction;
773         int updates;
774         int err;
775
776         smp_mb();
777         if (cur_trans->state >= TRANS_STATE_BLOCKED ||
778             cur_trans->delayed_refs.flushing)
779                 return 1;
780
781         updates = trans->delayed_ref_updates;
782         trans->delayed_ref_updates = 0;
783         if (updates) {
784                 err = btrfs_run_delayed_refs(trans, root, updates * 2);
785                 if (err) /* Error code will also eval true */
786                         return err;
787         }
788
789         return should_end_transaction(trans, root);
790 }
791
792 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
793                           struct btrfs_root *root, int throttle)
794 {
795         struct btrfs_transaction *cur_trans = trans->transaction;
796         struct btrfs_fs_info *info = root->fs_info;
797         unsigned long cur = trans->delayed_ref_updates;
798         int lock = (trans->type != TRANS_JOIN_NOLOCK);
799         int err = 0;
800         int must_run_delayed_refs = 0;
801
802         if (trans->use_count > 1) {
803                 trans->use_count--;
804                 trans->block_rsv = trans->orig_rsv;
805                 return 0;
806         }
807
808         btrfs_trans_release_metadata(trans, root);
809         trans->block_rsv = NULL;
810
811         if (!list_empty(&trans->new_bgs))
812                 btrfs_create_pending_block_groups(trans, root);
813
814         trans->delayed_ref_updates = 0;
815         if (!trans->sync) {
816                 must_run_delayed_refs =
817                         btrfs_should_throttle_delayed_refs(trans, root);
818                 cur = max_t(unsigned long, cur, 32);
819
820                 /*
821                  * don't make the caller wait if they are from a NOLOCK
822                  * or ATTACH transaction, it will deadlock with commit
823                  */
824                 if (must_run_delayed_refs == 1 &&
825                     (trans->type & (__TRANS_JOIN_NOLOCK | __TRANS_ATTACH)))
826                         must_run_delayed_refs = 2;
827         }
828
829         btrfs_trans_release_metadata(trans, root);
830         trans->block_rsv = NULL;
831
832         if (!list_empty(&trans->new_bgs))
833                 btrfs_create_pending_block_groups(trans, root);
834
835         btrfs_trans_release_chunk_metadata(trans);
836
837         if (lock && !atomic_read(&root->fs_info->open_ioctl_trans) &&
838             should_end_transaction(trans, root) &&
839             ACCESS_ONCE(cur_trans->state) == TRANS_STATE_RUNNING) {
840                 spin_lock(&info->trans_lock);
841                 if (cur_trans->state == TRANS_STATE_RUNNING)
842                         cur_trans->state = TRANS_STATE_BLOCKED;
843                 spin_unlock(&info->trans_lock);
844         }
845
846         if (lock && ACCESS_ONCE(cur_trans->state) == TRANS_STATE_BLOCKED) {
847                 if (throttle)
848                         return btrfs_commit_transaction(trans, root);
849                 else
850                         wake_up_process(info->transaction_kthread);
851         }
852
853         if (trans->type & __TRANS_FREEZABLE)
854                 sb_end_intwrite(root->fs_info->sb);
855
856         WARN_ON(cur_trans != info->running_transaction);
857         WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
858         atomic_dec(&cur_trans->num_writers);
859         extwriter_counter_dec(cur_trans, trans->type);
860
861         /*
862          * Make sure counter is updated before we wake up waiters.
863          */
864         smp_mb();
865         if (waitqueue_active(&cur_trans->writer_wait))
866                 wake_up(&cur_trans->writer_wait);
867         btrfs_put_transaction(cur_trans);
868
869         if (current->journal_info == trans)
870                 current->journal_info = NULL;
871
872         if (throttle)
873                 btrfs_run_delayed_iputs(root);
874
875         if (trans->aborted ||
876             test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state)) {
877                 wake_up_process(info->transaction_kthread);
878                 err = -EIO;
879         }
880         assert_qgroups_uptodate(trans);
881
882         kmem_cache_free(btrfs_trans_handle_cachep, trans);
883         if (must_run_delayed_refs) {
884                 btrfs_async_run_delayed_refs(root, cur,
885                                              must_run_delayed_refs == 1);
886         }
887         return err;
888 }
889
890 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
891                           struct btrfs_root *root)
892 {
893         return __btrfs_end_transaction(trans, root, 0);
894 }
895
896 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
897                                    struct btrfs_root *root)
898 {
899         return __btrfs_end_transaction(trans, root, 1);
900 }
901
902 /*
903  * when btree blocks are allocated, they have some corresponding bits set for
904  * them in one of two extent_io trees.  This is used to make sure all of
905  * those extents are sent to disk but does not wait on them
906  */
907 int btrfs_write_marked_extents(struct btrfs_root *root,
908                                struct extent_io_tree *dirty_pages, int mark)
909 {
910         int err = 0;
911         int werr = 0;
912         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
913         struct extent_state *cached_state = NULL;
914         u64 start = 0;
915         u64 end;
916
917         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
918                                       mark, &cached_state)) {
919                 bool wait_writeback = false;
920
921                 err = convert_extent_bit(dirty_pages, start, end,
922                                          EXTENT_NEED_WAIT,
923                                          mark, &cached_state, GFP_NOFS);
924                 /*
925                  * convert_extent_bit can return -ENOMEM, which is most of the
926                  * time a temporary error. So when it happens, ignore the error
927                  * and wait for writeback of this range to finish - because we
928                  * failed to set the bit EXTENT_NEED_WAIT for the range, a call
929                  * to btrfs_wait_marked_extents() would not know that writeback
930                  * for this range started and therefore wouldn't wait for it to
931                  * finish - we don't want to commit a superblock that points to
932                  * btree nodes/leafs for which writeback hasn't finished yet
933                  * (and without errors).
934                  * We cleanup any entries left in the io tree when committing
935                  * the transaction (through clear_btree_io_tree()).
936                  */
937                 if (err == -ENOMEM) {
938                         err = 0;
939                         wait_writeback = true;
940                 }
941                 if (!err)
942                         err = filemap_fdatawrite_range(mapping, start, end);
943                 if (err)
944                         werr = err;
945                 else if (wait_writeback)
946                         werr = filemap_fdatawait_range(mapping, start, end);
947                 free_extent_state(cached_state);
948                 cached_state = NULL;
949                 cond_resched();
950                 start = end + 1;
951         }
952         return werr;
953 }
954
955 /*
956  * when btree blocks are allocated, they have some corresponding bits set for
957  * them in one of two extent_io trees.  This is used to make sure all of
958  * those extents are on disk for transaction or log commit.  We wait
959  * on all the pages and clear them from the dirty pages state tree
960  */
961 int btrfs_wait_marked_extents(struct btrfs_root *root,
962                               struct extent_io_tree *dirty_pages, int mark)
963 {
964         int err = 0;
965         int werr = 0;
966         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
967         struct extent_state *cached_state = NULL;
968         u64 start = 0;
969         u64 end;
970         struct btrfs_inode *btree_ino = BTRFS_I(root->fs_info->btree_inode);
971         bool errors = false;
972
973         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
974                                       EXTENT_NEED_WAIT, &cached_state)) {
975                 /*
976                  * Ignore -ENOMEM errors returned by clear_extent_bit().
977                  * When committing the transaction, we'll remove any entries
978                  * left in the io tree. For a log commit, we don't remove them
979                  * after committing the log because the tree can be accessed
980                  * concurrently - we do it only at transaction commit time when
981                  * it's safe to do it (through clear_btree_io_tree()).
982                  */
983                 err = clear_extent_bit(dirty_pages, start, end,
984                                        EXTENT_NEED_WAIT,
985                                        0, 0, &cached_state, GFP_NOFS);
986                 if (err == -ENOMEM)
987                         err = 0;
988                 if (!err)
989                         err = filemap_fdatawait_range(mapping, start, end);
990                 if (err)
991                         werr = err;
992                 free_extent_state(cached_state);
993                 cached_state = NULL;
994                 cond_resched();
995                 start = end + 1;
996         }
997         if (err)
998                 werr = err;
999
1000         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
1001                 if ((mark & EXTENT_DIRTY) &&
1002                     test_and_clear_bit(BTRFS_INODE_BTREE_LOG1_ERR,
1003                                        &btree_ino->runtime_flags))
1004                         errors = true;
1005
1006                 if ((mark & EXTENT_NEW) &&
1007                     test_and_clear_bit(BTRFS_INODE_BTREE_LOG2_ERR,
1008                                        &btree_ino->runtime_flags))
1009                         errors = true;
1010         } else {
1011                 if (test_and_clear_bit(BTRFS_INODE_BTREE_ERR,
1012                                        &btree_ino->runtime_flags))
1013                         errors = true;
1014         }
1015
1016         if (errors && !werr)
1017                 werr = -EIO;
1018
1019         return werr;
1020 }
1021
1022 /*
1023  * when btree blocks are allocated, they have some corresponding bits set for
1024  * them in one of two extent_io trees.  This is used to make sure all of
1025  * those extents are on disk for transaction or log commit
1026  */
1027 static int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
1028                                 struct extent_io_tree *dirty_pages, int mark)
1029 {
1030         int ret;
1031         int ret2;
1032         struct blk_plug plug;
1033
1034         blk_start_plug(&plug);
1035         ret = btrfs_write_marked_extents(root, dirty_pages, mark);
1036         blk_finish_plug(&plug);
1037         ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
1038
1039         if (ret)
1040                 return ret;
1041         if (ret2)
1042                 return ret2;
1043         return 0;
1044 }
1045
1046 static int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
1047                                      struct btrfs_root *root)
1048 {
1049         int ret;
1050
1051         ret = btrfs_write_and_wait_marked_extents(root,
1052                                            &trans->transaction->dirty_pages,
1053                                            EXTENT_DIRTY);
1054         clear_btree_io_tree(&trans->transaction->dirty_pages);
1055
1056         return ret;
1057 }
1058
1059 /*
1060  * this is used to update the root pointer in the tree of tree roots.
1061  *
1062  * But, in the case of the extent allocation tree, updating the root
1063  * pointer may allocate blocks which may change the root of the extent
1064  * allocation tree.
1065  *
1066  * So, this loops and repeats and makes sure the cowonly root didn't
1067  * change while the root pointer was being updated in the metadata.
1068  */
1069 static int update_cowonly_root(struct btrfs_trans_handle *trans,
1070                                struct btrfs_root *root)
1071 {
1072         int ret;
1073         u64 old_root_bytenr;
1074         u64 old_root_used;
1075         struct btrfs_root *tree_root = root->fs_info->tree_root;
1076
1077         old_root_used = btrfs_root_used(&root->root_item);
1078
1079         while (1) {
1080                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
1081                 if (old_root_bytenr == root->node->start &&
1082                     old_root_used == btrfs_root_used(&root->root_item))
1083                         break;
1084
1085                 btrfs_set_root_node(&root->root_item, root->node);
1086                 ret = btrfs_update_root(trans, tree_root,
1087                                         &root->root_key,
1088                                         &root->root_item);
1089                 if (ret)
1090                         return ret;
1091
1092                 old_root_used = btrfs_root_used(&root->root_item);
1093         }
1094
1095         return 0;
1096 }
1097
1098 /*
1099  * update all the cowonly tree roots on disk
1100  *
1101  * The error handling in this function may not be obvious. Any of the
1102  * failures will cause the file system to go offline. We still need
1103  * to clean up the delayed refs.
1104  */
1105 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
1106                                          struct btrfs_root *root)
1107 {
1108         struct btrfs_fs_info *fs_info = root->fs_info;
1109         struct list_head *dirty_bgs = &trans->transaction->dirty_bgs;
1110         struct list_head *io_bgs = &trans->transaction->io_bgs;
1111         struct list_head *next;
1112         struct extent_buffer *eb;
1113         int ret;
1114
1115         eb = btrfs_lock_root_node(fs_info->tree_root);
1116         ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL,
1117                               0, &eb);
1118         btrfs_tree_unlock(eb);
1119         free_extent_buffer(eb);
1120
1121         if (ret)
1122                 return ret;
1123
1124         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1125         if (ret)
1126                 return ret;
1127
1128         ret = btrfs_run_dev_stats(trans, root->fs_info);
1129         if (ret)
1130                 return ret;
1131         ret = btrfs_run_dev_replace(trans, root->fs_info);
1132         if (ret)
1133                 return ret;
1134         ret = btrfs_run_qgroups(trans, root->fs_info);
1135         if (ret)
1136                 return ret;
1137
1138         ret = btrfs_setup_space_cache(trans, root);
1139         if (ret)
1140                 return ret;
1141
1142         /* run_qgroups might have added some more refs */
1143         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1144         if (ret)
1145                 return ret;
1146 again:
1147         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
1148                 next = fs_info->dirty_cowonly_roots.next;
1149                 list_del_init(next);
1150                 root = list_entry(next, struct btrfs_root, dirty_list);
1151                 clear_bit(BTRFS_ROOT_DIRTY, &root->state);
1152
1153                 if (root != fs_info->extent_root)
1154                         list_add_tail(&root->dirty_list,
1155                                       &trans->transaction->switch_commits);
1156                 ret = update_cowonly_root(trans, root);
1157                 if (ret)
1158                         return ret;
1159                 ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1160                 if (ret)
1161                         return ret;
1162         }
1163
1164         while (!list_empty(dirty_bgs) || !list_empty(io_bgs)) {
1165                 ret = btrfs_write_dirty_block_groups(trans, root);
1166                 if (ret)
1167                         return ret;
1168                 ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1169                 if (ret)
1170                         return ret;
1171         }
1172
1173         if (!list_empty(&fs_info->dirty_cowonly_roots))
1174                 goto again;
1175
1176         list_add_tail(&fs_info->extent_root->dirty_list,
1177                       &trans->transaction->switch_commits);
1178         btrfs_after_dev_replace_commit(fs_info);
1179
1180         return 0;
1181 }
1182
1183 /*
1184  * dead roots are old snapshots that need to be deleted.  This allocates
1185  * a dirty root struct and adds it into the list of dead roots that need to
1186  * be deleted
1187  */
1188 void btrfs_add_dead_root(struct btrfs_root *root)
1189 {
1190         spin_lock(&root->fs_info->trans_lock);
1191         if (list_empty(&root->root_list))
1192                 list_add_tail(&root->root_list, &root->fs_info->dead_roots);
1193         spin_unlock(&root->fs_info->trans_lock);
1194 }
1195
1196 /*
1197  * update all the cowonly tree roots on disk
1198  */
1199 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
1200                                     struct btrfs_root *root)
1201 {
1202         struct btrfs_root *gang[8];
1203         struct btrfs_fs_info *fs_info = root->fs_info;
1204         int i;
1205         int ret;
1206         int err = 0;
1207
1208         spin_lock(&fs_info->fs_roots_radix_lock);
1209         while (1) {
1210                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
1211                                                  (void **)gang, 0,
1212                                                  ARRAY_SIZE(gang),
1213                                                  BTRFS_ROOT_TRANS_TAG);
1214                 if (ret == 0)
1215                         break;
1216                 for (i = 0; i < ret; i++) {
1217                         root = gang[i];
1218                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
1219                                         (unsigned long)root->root_key.objectid,
1220                                         BTRFS_ROOT_TRANS_TAG);
1221                         spin_unlock(&fs_info->fs_roots_radix_lock);
1222
1223                         btrfs_free_log(trans, root);
1224                         btrfs_update_reloc_root(trans, root);
1225                         btrfs_orphan_commit_root(trans, root);
1226
1227                         btrfs_save_ino_cache(root, trans);
1228
1229                         /* see comments in should_cow_block() */
1230                         clear_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1231                         smp_mb__after_atomic();
1232
1233                         if (root->commit_root != root->node) {
1234                                 list_add_tail(&root->dirty_list,
1235                                         &trans->transaction->switch_commits);
1236                                 btrfs_set_root_node(&root->root_item,
1237                                                     root->node);
1238                         }
1239
1240                         err = btrfs_update_root(trans, fs_info->tree_root,
1241                                                 &root->root_key,
1242                                                 &root->root_item);
1243                         spin_lock(&fs_info->fs_roots_radix_lock);
1244                         if (err)
1245                                 break;
1246                         btrfs_qgroup_free_meta_all(root);
1247                 }
1248         }
1249         spin_unlock(&fs_info->fs_roots_radix_lock);
1250         return err;
1251 }
1252
1253 /*
1254  * defrag a given btree.
1255  * Every leaf in the btree is read and defragged.
1256  */
1257 int btrfs_defrag_root(struct btrfs_root *root)
1258 {
1259         struct btrfs_fs_info *info = root->fs_info;
1260         struct btrfs_trans_handle *trans;
1261         int ret;
1262
1263         if (test_and_set_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state))
1264                 return 0;
1265
1266         while (1) {
1267                 trans = btrfs_start_transaction(root, 0);
1268                 if (IS_ERR(trans))
1269                         return PTR_ERR(trans);
1270
1271                 ret = btrfs_defrag_leaves(trans, root);
1272
1273                 btrfs_end_transaction(trans, root);
1274                 btrfs_btree_balance_dirty(info->tree_root);
1275                 cond_resched();
1276
1277                 if (btrfs_fs_closing(root->fs_info) || ret != -EAGAIN)
1278                         break;
1279
1280                 if (btrfs_defrag_cancelled(root->fs_info)) {
1281                         pr_debug("BTRFS: defrag_root cancelled\n");
1282                         ret = -EAGAIN;
1283                         break;
1284                 }
1285         }
1286         clear_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state);
1287         return ret;
1288 }
1289
1290 /*
1291  * new snapshots need to be created at a very specific time in the
1292  * transaction commit.  This does the actual creation.
1293  *
1294  * Note:
1295  * If the error which may affect the commitment of the current transaction
1296  * happens, we should return the error number. If the error which just affect
1297  * the creation of the pending snapshots, just return 0.
1298  */
1299 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
1300                                    struct btrfs_fs_info *fs_info,
1301                                    struct btrfs_pending_snapshot *pending)
1302 {
1303         struct btrfs_key key;
1304         struct btrfs_root_item *new_root_item;
1305         struct btrfs_root *tree_root = fs_info->tree_root;
1306         struct btrfs_root *root = pending->root;
1307         struct btrfs_root *parent_root;
1308         struct btrfs_block_rsv *rsv;
1309         struct inode *parent_inode;
1310         struct btrfs_path *path;
1311         struct btrfs_dir_item *dir_item;
1312         struct dentry *dentry;
1313         struct extent_buffer *tmp;
1314         struct extent_buffer *old;
1315         struct timespec cur_time = CURRENT_TIME;
1316         int ret = 0;
1317         u64 to_reserve = 0;
1318         u64 index = 0;
1319         u64 objectid;
1320         u64 root_flags;
1321         uuid_le new_uuid;
1322
1323         path = btrfs_alloc_path();
1324         if (!path) {
1325                 pending->error = -ENOMEM;
1326                 return 0;
1327         }
1328
1329         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
1330         if (!new_root_item) {
1331                 pending->error = -ENOMEM;
1332                 goto root_item_alloc_fail;
1333         }
1334
1335         pending->error = btrfs_find_free_objectid(tree_root, &objectid);
1336         if (pending->error)
1337                 goto no_free_objectid;
1338
1339         /*
1340          * Make qgroup to skip current new snapshot's qgroupid, as it is
1341          * accounted by later btrfs_qgroup_inherit().
1342          */
1343         btrfs_set_skip_qgroup(trans, objectid);
1344
1345         btrfs_reloc_pre_snapshot(pending, &to_reserve);
1346
1347         if (to_reserve > 0) {
1348                 pending->error = btrfs_block_rsv_add(root,
1349                                                      &pending->block_rsv,
1350                                                      to_reserve,
1351                                                      BTRFS_RESERVE_NO_FLUSH);
1352                 if (pending->error)
1353                         goto clear_skip_qgroup;
1354         }
1355
1356         key.objectid = objectid;
1357         key.offset = (u64)-1;
1358         key.type = BTRFS_ROOT_ITEM_KEY;
1359
1360         rsv = trans->block_rsv;
1361         trans->block_rsv = &pending->block_rsv;
1362         trans->bytes_reserved = trans->block_rsv->reserved;
1363
1364         dentry = pending->dentry;
1365         parent_inode = pending->dir;
1366         parent_root = BTRFS_I(parent_inode)->root;
1367         record_root_in_trans(trans, parent_root);
1368
1369         /*
1370          * insert the directory item
1371          */
1372         ret = btrfs_set_inode_index(parent_inode, &index);
1373         BUG_ON(ret); /* -ENOMEM */
1374
1375         /* check if there is a file/dir which has the same name. */
1376         dir_item = btrfs_lookup_dir_item(NULL, parent_root, path,
1377                                          btrfs_ino(parent_inode),
1378                                          dentry->d_name.name,
1379                                          dentry->d_name.len, 0);
1380         if (dir_item != NULL && !IS_ERR(dir_item)) {
1381                 pending->error = -EEXIST;
1382                 goto dir_item_existed;
1383         } else if (IS_ERR(dir_item)) {
1384                 ret = PTR_ERR(dir_item);
1385                 btrfs_abort_transaction(trans, root, ret);
1386                 goto fail;
1387         }
1388         btrfs_release_path(path);
1389
1390         /*
1391          * pull in the delayed directory update
1392          * and the delayed inode item
1393          * otherwise we corrupt the FS during
1394          * snapshot
1395          */
1396         ret = btrfs_run_delayed_items(trans, root);
1397         if (ret) {      /* Transaction aborted */
1398                 btrfs_abort_transaction(trans, root, ret);
1399                 goto fail;
1400         }
1401
1402         record_root_in_trans(trans, root);
1403         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
1404         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
1405         btrfs_check_and_init_root_item(new_root_item);
1406
1407         root_flags = btrfs_root_flags(new_root_item);
1408         if (pending->readonly)
1409                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
1410         else
1411                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
1412         btrfs_set_root_flags(new_root_item, root_flags);
1413
1414         btrfs_set_root_generation_v2(new_root_item,
1415                         trans->transid);
1416         uuid_le_gen(&new_uuid);
1417         memcpy(new_root_item->uuid, new_uuid.b, BTRFS_UUID_SIZE);
1418         memcpy(new_root_item->parent_uuid, root->root_item.uuid,
1419                         BTRFS_UUID_SIZE);
1420         if (!(root_flags & BTRFS_ROOT_SUBVOL_RDONLY)) {
1421                 memset(new_root_item->received_uuid, 0,
1422                        sizeof(new_root_item->received_uuid));
1423                 memset(&new_root_item->stime, 0, sizeof(new_root_item->stime));
1424                 memset(&new_root_item->rtime, 0, sizeof(new_root_item->rtime));
1425                 btrfs_set_root_stransid(new_root_item, 0);
1426                 btrfs_set_root_rtransid(new_root_item, 0);
1427         }
1428         btrfs_set_stack_timespec_sec(&new_root_item->otime, cur_time.tv_sec);
1429         btrfs_set_stack_timespec_nsec(&new_root_item->otime, cur_time.tv_nsec);
1430         btrfs_set_root_otransid(new_root_item, trans->transid);
1431
1432         old = btrfs_lock_root_node(root);
1433         ret = btrfs_cow_block(trans, root, old, NULL, 0, &old);
1434         if (ret) {
1435                 btrfs_tree_unlock(old);
1436                 free_extent_buffer(old);
1437                 btrfs_abort_transaction(trans, root, ret);
1438                 goto fail;
1439         }
1440
1441         btrfs_set_lock_blocking(old);
1442
1443         ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1444         /* clean up in any case */
1445         btrfs_tree_unlock(old);
1446         free_extent_buffer(old);
1447         if (ret) {
1448                 btrfs_abort_transaction(trans, root, ret);
1449                 goto fail;
1450         }
1451         /* see comments in should_cow_block() */
1452         set_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1453         smp_wmb();
1454
1455         btrfs_set_root_node(new_root_item, tmp);
1456         /* record when the snapshot was created in key.offset */
1457         key.offset = trans->transid;
1458         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1459         btrfs_tree_unlock(tmp);
1460         free_extent_buffer(tmp);
1461         if (ret) {
1462                 btrfs_abort_transaction(trans, root, ret);
1463                 goto fail;
1464         }
1465
1466         /*
1467          * insert root back/forward references
1468          */
1469         ret = btrfs_add_root_ref(trans, tree_root, objectid,
1470                                  parent_root->root_key.objectid,
1471                                  btrfs_ino(parent_inode), index,
1472                                  dentry->d_name.name, dentry->d_name.len);
1473         if (ret) {
1474                 btrfs_abort_transaction(trans, root, ret);
1475                 goto fail;
1476         }
1477
1478         key.offset = (u64)-1;
1479         pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
1480         if (IS_ERR(pending->snap)) {
1481                 ret = PTR_ERR(pending->snap);
1482                 btrfs_abort_transaction(trans, root, ret);
1483                 goto fail;
1484         }
1485
1486         ret = btrfs_reloc_post_snapshot(trans, pending);
1487         if (ret) {
1488                 btrfs_abort_transaction(trans, root, ret);
1489                 goto fail;
1490         }
1491
1492         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1493         if (ret) {
1494                 btrfs_abort_transaction(trans, root, ret);
1495                 goto fail;
1496         }
1497
1498         ret = btrfs_insert_dir_item(trans, parent_root,
1499                                     dentry->d_name.name, dentry->d_name.len,
1500                                     parent_inode, &key,
1501                                     BTRFS_FT_DIR, index);
1502         /* We have check then name at the beginning, so it is impossible. */
1503         BUG_ON(ret == -EEXIST || ret == -EOVERFLOW);
1504         if (ret) {
1505                 btrfs_abort_transaction(trans, root, ret);
1506                 goto fail;
1507         }
1508
1509         btrfs_i_size_write(parent_inode, parent_inode->i_size +
1510                                          dentry->d_name.len * 2);
1511         parent_inode->i_mtime = parent_inode->i_ctime = CURRENT_TIME;
1512         ret = btrfs_update_inode_fallback(trans, parent_root, parent_inode);
1513         if (ret) {
1514                 btrfs_abort_transaction(trans, root, ret);
1515                 goto fail;
1516         }
1517         ret = btrfs_uuid_tree_add(trans, fs_info->uuid_root, new_uuid.b,
1518                                   BTRFS_UUID_KEY_SUBVOL, objectid);
1519         if (ret) {
1520                 btrfs_abort_transaction(trans, root, ret);
1521                 goto fail;
1522         }
1523         if (!btrfs_is_empty_uuid(new_root_item->received_uuid)) {
1524                 ret = btrfs_uuid_tree_add(trans, fs_info->uuid_root,
1525                                           new_root_item->received_uuid,
1526                                           BTRFS_UUID_KEY_RECEIVED_SUBVOL,
1527                                           objectid);
1528                 if (ret && ret != -EEXIST) {
1529                         btrfs_abort_transaction(trans, root, ret);
1530                         goto fail;
1531                 }
1532         }
1533
1534         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1535         if (ret) {
1536                 btrfs_abort_transaction(trans, root, ret);
1537                 goto fail;
1538         }
1539
1540         /*
1541          * account qgroup counters before qgroup_inherit()
1542          */
1543         ret = btrfs_qgroup_prepare_account_extents(trans, fs_info);
1544         if (ret)
1545                 goto fail;
1546         ret = btrfs_qgroup_account_extents(trans, fs_info);
1547         if (ret)
1548                 goto fail;
1549         ret = btrfs_qgroup_inherit(trans, fs_info,
1550                                    root->root_key.objectid,
1551                                    objectid, pending->inherit);
1552         if (ret) {
1553                 btrfs_abort_transaction(trans, root, ret);
1554                 goto fail;
1555         }
1556
1557 fail:
1558         pending->error = ret;
1559 dir_item_existed:
1560         trans->block_rsv = rsv;
1561         trans->bytes_reserved = 0;
1562 clear_skip_qgroup:
1563         btrfs_clear_skip_qgroup(trans);
1564 no_free_objectid:
1565         kfree(new_root_item);
1566 root_item_alloc_fail:
1567         btrfs_free_path(path);
1568         return ret;
1569 }
1570
1571 /*
1572  * create all the snapshots we've scheduled for creation
1573  */
1574 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
1575                                              struct btrfs_fs_info *fs_info)
1576 {
1577         struct btrfs_pending_snapshot *pending, *next;
1578         struct list_head *head = &trans->transaction->pending_snapshots;
1579         int ret = 0;
1580
1581         list_for_each_entry_safe(pending, next, head, list) {
1582                 list_del(&pending->list);
1583                 ret = create_pending_snapshot(trans, fs_info, pending);
1584                 if (ret)
1585                         break;
1586         }
1587         return ret;
1588 }
1589
1590 static void update_super_roots(struct btrfs_root *root)
1591 {
1592         struct btrfs_root_item *root_item;
1593         struct btrfs_super_block *super;
1594
1595         super = root->fs_info->super_copy;
1596
1597         root_item = &root->fs_info->chunk_root->root_item;
1598         super->chunk_root = root_item->bytenr;
1599         super->chunk_root_generation = root_item->generation;
1600         super->chunk_root_level = root_item->level;
1601
1602         root_item = &root->fs_info->tree_root->root_item;
1603         super->root = root_item->bytenr;
1604         super->generation = root_item->generation;
1605         super->root_level = root_item->level;
1606         if (btrfs_test_opt(root, SPACE_CACHE))
1607                 super->cache_generation = root_item->generation;
1608         if (root->fs_info->update_uuid_tree_gen)
1609                 super->uuid_tree_generation = root_item->generation;
1610 }
1611
1612 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1613 {
1614         struct btrfs_transaction *trans;
1615         int ret = 0;
1616
1617         spin_lock(&info->trans_lock);
1618         trans = info->running_transaction;
1619         if (trans)
1620                 ret = (trans->state >= TRANS_STATE_COMMIT_START);
1621         spin_unlock(&info->trans_lock);
1622         return ret;
1623 }
1624
1625 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1626 {
1627         struct btrfs_transaction *trans;
1628         int ret = 0;
1629
1630         spin_lock(&info->trans_lock);
1631         trans = info->running_transaction;
1632         if (trans)
1633                 ret = is_transaction_blocked(trans);
1634         spin_unlock(&info->trans_lock);
1635         return ret;
1636 }
1637
1638 /*
1639  * wait for the current transaction commit to start and block subsequent
1640  * transaction joins
1641  */
1642 static void wait_current_trans_commit_start(struct btrfs_root *root,
1643                                             struct btrfs_transaction *trans)
1644 {
1645         wait_event(root->fs_info->transaction_blocked_wait,
1646                    trans->state >= TRANS_STATE_COMMIT_START ||
1647                    trans->aborted);
1648 }
1649
1650 /*
1651  * wait for the current transaction to start and then become unblocked.
1652  * caller holds ref.
1653  */
1654 static void wait_current_trans_commit_start_and_unblock(struct btrfs_root *root,
1655                                          struct btrfs_transaction *trans)
1656 {
1657         wait_event(root->fs_info->transaction_wait,
1658                    trans->state >= TRANS_STATE_UNBLOCKED ||
1659                    trans->aborted);
1660 }
1661
1662 /*
1663  * commit transactions asynchronously. once btrfs_commit_transaction_async
1664  * returns, any subsequent transaction will not be allowed to join.
1665  */
1666 struct btrfs_async_commit {
1667         struct btrfs_trans_handle *newtrans;
1668         struct btrfs_root *root;
1669         struct work_struct work;
1670 };
1671
1672 static void do_async_commit(struct work_struct *work)
1673 {
1674         struct btrfs_async_commit *ac =
1675                 container_of(work, struct btrfs_async_commit, work);
1676
1677         /*
1678          * We've got freeze protection passed with the transaction.
1679          * Tell lockdep about it.
1680          */
1681         if (ac->newtrans->type & __TRANS_FREEZABLE)
1682                 __sb_writers_acquired(ac->root->fs_info->sb, SB_FREEZE_FS);
1683
1684         current->journal_info = ac->newtrans;
1685
1686         btrfs_commit_transaction(ac->newtrans, ac->root);
1687         kfree(ac);
1688 }
1689
1690 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1691                                    struct btrfs_root *root,
1692                                    int wait_for_unblock)
1693 {
1694         struct btrfs_async_commit *ac;
1695         struct btrfs_transaction *cur_trans;
1696
1697         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1698         if (!ac)
1699                 return -ENOMEM;
1700
1701         INIT_WORK(&ac->work, do_async_commit);
1702         ac->root = root;
1703         ac->newtrans = btrfs_join_transaction(root);
1704         if (IS_ERR(ac->newtrans)) {
1705                 int err = PTR_ERR(ac->newtrans);
1706                 kfree(ac);
1707                 return err;
1708         }
1709
1710         /* take transaction reference */
1711         cur_trans = trans->transaction;
1712         atomic_inc(&cur_trans->use_count);
1713
1714         btrfs_end_transaction(trans, root);
1715
1716         /*
1717          * Tell lockdep we've released the freeze rwsem, since the
1718          * async commit thread will be the one to unlock it.
1719          */
1720         if (ac->newtrans->type & __TRANS_FREEZABLE)
1721                 __sb_writers_release(root->fs_info->sb, SB_FREEZE_FS);
1722
1723         schedule_work(&ac->work);
1724
1725         /* wait for transaction to start and unblock */
1726         if (wait_for_unblock)
1727                 wait_current_trans_commit_start_and_unblock(root, cur_trans);
1728         else
1729                 wait_current_trans_commit_start(root, cur_trans);
1730
1731         if (current->journal_info == trans)
1732                 current->journal_info = NULL;
1733
1734         btrfs_put_transaction(cur_trans);
1735         return 0;
1736 }
1737
1738
1739 static void cleanup_transaction(struct btrfs_trans_handle *trans,
1740                                 struct btrfs_root *root, int err)
1741 {
1742         struct btrfs_transaction *cur_trans = trans->transaction;
1743         DEFINE_WAIT(wait);
1744
1745         WARN_ON(trans->use_count > 1);
1746
1747         btrfs_abort_transaction(trans, root, err);
1748
1749         spin_lock(&root->fs_info->trans_lock);
1750
1751         /*
1752          * If the transaction is removed from the list, it means this
1753          * transaction has been committed successfully, so it is impossible
1754          * to call the cleanup function.
1755          */
1756         BUG_ON(list_empty(&cur_trans->list));
1757
1758         list_del_init(&cur_trans->list);
1759         if (cur_trans == root->fs_info->running_transaction) {
1760                 cur_trans->state = TRANS_STATE_COMMIT_DOING;
1761                 spin_unlock(&root->fs_info->trans_lock);
1762                 wait_event(cur_trans->writer_wait,
1763                            atomic_read(&cur_trans->num_writers) == 1);
1764
1765                 spin_lock(&root->fs_info->trans_lock);
1766         }
1767         spin_unlock(&root->fs_info->trans_lock);
1768
1769         btrfs_cleanup_one_transaction(trans->transaction, root);
1770
1771         spin_lock(&root->fs_info->trans_lock);
1772         if (cur_trans == root->fs_info->running_transaction)
1773                 root->fs_info->running_transaction = NULL;
1774         spin_unlock(&root->fs_info->trans_lock);
1775
1776         if (trans->type & __TRANS_FREEZABLE)
1777                 sb_end_intwrite(root->fs_info->sb);
1778         btrfs_put_transaction(cur_trans);
1779         btrfs_put_transaction(cur_trans);
1780
1781         trace_btrfs_transaction_commit(root);
1782
1783         if (current->journal_info == trans)
1784                 current->journal_info = NULL;
1785         btrfs_scrub_cancel(root->fs_info);
1786
1787         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1788 }
1789
1790 static inline int btrfs_start_delalloc_flush(struct btrfs_fs_info *fs_info)
1791 {
1792         if (btrfs_test_opt(fs_info->tree_root, FLUSHONCOMMIT))
1793                 return btrfs_start_delalloc_roots(fs_info, 1, -1);
1794         return 0;
1795 }
1796
1797 static inline void btrfs_wait_delalloc_flush(struct btrfs_fs_info *fs_info)
1798 {
1799         if (btrfs_test_opt(fs_info->tree_root, FLUSHONCOMMIT))
1800                 btrfs_wait_ordered_roots(fs_info, -1);
1801 }
1802
1803 static inline void
1804 btrfs_wait_pending_ordered(struct btrfs_transaction *cur_trans)
1805 {
1806         wait_event(cur_trans->pending_wait,
1807                    atomic_read(&cur_trans->pending_ordered) == 0);
1808 }
1809
1810 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
1811                              struct btrfs_root *root)
1812 {
1813         struct btrfs_transaction *cur_trans = trans->transaction;
1814         struct btrfs_transaction *prev_trans = NULL;
1815         struct btrfs_inode *btree_ino = BTRFS_I(root->fs_info->btree_inode);
1816         int ret;
1817
1818         /* Stop the commit early if ->aborted is set */
1819         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
1820                 ret = cur_trans->aborted;
1821                 btrfs_end_transaction(trans, root);
1822                 return ret;
1823         }
1824
1825         /* make a pass through all the delayed refs we have so far
1826          * any runnings procs may add more while we are here
1827          */
1828         ret = btrfs_run_delayed_refs(trans, root, 0);
1829         if (ret) {
1830                 btrfs_end_transaction(trans, root);
1831                 return ret;
1832         }
1833
1834         btrfs_trans_release_metadata(trans, root);
1835         trans->block_rsv = NULL;
1836
1837         cur_trans = trans->transaction;
1838
1839         /*
1840          * set the flushing flag so procs in this transaction have to
1841          * start sending their work down.
1842          */
1843         cur_trans->delayed_refs.flushing = 1;
1844         smp_wmb();
1845
1846         if (!list_empty(&trans->new_bgs))
1847                 btrfs_create_pending_block_groups(trans, root);
1848
1849         ret = btrfs_run_delayed_refs(trans, root, 0);
1850         if (ret) {
1851                 btrfs_end_transaction(trans, root);
1852                 return ret;
1853         }
1854
1855         if (!test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &cur_trans->flags)) {
1856                 int run_it = 0;
1857
1858                 /* this mutex is also taken before trying to set
1859                  * block groups readonly.  We need to make sure
1860                  * that nobody has set a block group readonly
1861                  * after a extents from that block group have been
1862                  * allocated for cache files.  btrfs_set_block_group_ro
1863                  * will wait for the transaction to commit if it
1864                  * finds BTRFS_TRANS_DIRTY_BG_RUN set.
1865                  *
1866                  * The BTRFS_TRANS_DIRTY_BG_RUN flag is also used to make sure
1867                  * only one process starts all the block group IO.  It wouldn't
1868                  * hurt to have more than one go through, but there's no
1869                  * real advantage to it either.
1870                  */
1871                 mutex_lock(&root->fs_info->ro_block_group_mutex);
1872                 if (!test_and_set_bit(BTRFS_TRANS_DIRTY_BG_RUN,
1873                                       &cur_trans->flags))
1874                         run_it = 1;
1875                 mutex_unlock(&root->fs_info->ro_block_group_mutex);
1876
1877                 if (run_it)
1878                         ret = btrfs_start_dirty_block_groups(trans, root);
1879         }
1880         if (ret) {
1881                 btrfs_end_transaction(trans, root);
1882                 return ret;
1883         }
1884
1885         spin_lock(&root->fs_info->trans_lock);
1886         if (cur_trans->state >= TRANS_STATE_COMMIT_START) {
1887                 spin_unlock(&root->fs_info->trans_lock);
1888                 atomic_inc(&cur_trans->use_count);
1889                 ret = btrfs_end_transaction(trans, root);
1890
1891                 wait_for_commit(root, cur_trans);
1892
1893                 if (unlikely(cur_trans->aborted))
1894                         ret = cur_trans->aborted;
1895
1896                 btrfs_put_transaction(cur_trans);
1897
1898                 return ret;
1899         }
1900
1901         cur_trans->state = TRANS_STATE_COMMIT_START;
1902         wake_up(&root->fs_info->transaction_blocked_wait);
1903
1904         if (cur_trans->list.prev != &root->fs_info->trans_list) {
1905                 prev_trans = list_entry(cur_trans->list.prev,
1906                                         struct btrfs_transaction, list);
1907                 if (prev_trans->state != TRANS_STATE_COMPLETED) {
1908                         atomic_inc(&prev_trans->use_count);
1909                         spin_unlock(&root->fs_info->trans_lock);
1910
1911                         wait_for_commit(root, prev_trans);
1912                         ret = prev_trans->aborted;
1913
1914                         btrfs_put_transaction(prev_trans);
1915                         if (ret)
1916                                 goto cleanup_transaction;
1917                 } else {
1918                         spin_unlock(&root->fs_info->trans_lock);
1919                 }
1920         } else {
1921                 spin_unlock(&root->fs_info->trans_lock);
1922         }
1923
1924         extwriter_counter_dec(cur_trans, trans->type);
1925
1926         ret = btrfs_start_delalloc_flush(root->fs_info);
1927         if (ret)
1928                 goto cleanup_transaction;
1929
1930         ret = btrfs_run_delayed_items(trans, root);
1931         if (ret)
1932                 goto cleanup_transaction;
1933
1934         wait_event(cur_trans->writer_wait,
1935                    extwriter_counter_read(cur_trans) == 0);
1936
1937         /* some pending stuffs might be added after the previous flush. */
1938         ret = btrfs_run_delayed_items(trans, root);
1939         if (ret)
1940                 goto cleanup_transaction;
1941
1942         btrfs_wait_delalloc_flush(root->fs_info);
1943
1944         btrfs_wait_pending_ordered(cur_trans);
1945
1946         btrfs_scrub_pause(root);
1947         /*
1948          * Ok now we need to make sure to block out any other joins while we
1949          * commit the transaction.  We could have started a join before setting
1950          * COMMIT_DOING so make sure to wait for num_writers to == 1 again.
1951          */
1952         spin_lock(&root->fs_info->trans_lock);
1953         cur_trans->state = TRANS_STATE_COMMIT_DOING;
1954         spin_unlock(&root->fs_info->trans_lock);
1955         wait_event(cur_trans->writer_wait,
1956                    atomic_read(&cur_trans->num_writers) == 1);
1957
1958         /* ->aborted might be set after the previous check, so check it */
1959         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
1960                 ret = cur_trans->aborted;
1961                 goto scrub_continue;
1962         }
1963         /*
1964          * the reloc mutex makes sure that we stop
1965          * the balancing code from coming in and moving
1966          * extents around in the middle of the commit
1967          */
1968         mutex_lock(&root->fs_info->reloc_mutex);
1969
1970         /*
1971          * We needn't worry about the delayed items because we will
1972          * deal with them in create_pending_snapshot(), which is the
1973          * core function of the snapshot creation.
1974          */
1975         ret = create_pending_snapshots(trans, root->fs_info);
1976         if (ret) {
1977                 mutex_unlock(&root->fs_info->reloc_mutex);
1978                 goto scrub_continue;
1979         }
1980
1981         /*
1982          * We insert the dir indexes of the snapshots and update the inode
1983          * of the snapshots' parents after the snapshot creation, so there
1984          * are some delayed items which are not dealt with. Now deal with
1985          * them.
1986          *
1987          * We needn't worry that this operation will corrupt the snapshots,
1988          * because all the tree which are snapshoted will be forced to COW
1989          * the nodes and leaves.
1990          */
1991         ret = btrfs_run_delayed_items(trans, root);
1992         if (ret) {
1993                 mutex_unlock(&root->fs_info->reloc_mutex);
1994                 goto scrub_continue;
1995         }
1996
1997         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1998         if (ret) {
1999                 mutex_unlock(&root->fs_info->reloc_mutex);
2000                 goto scrub_continue;
2001         }
2002
2003         /* Reocrd old roots for later qgroup accounting */
2004         ret = btrfs_qgroup_prepare_account_extents(trans, root->fs_info);
2005         if (ret) {
2006                 mutex_unlock(&root->fs_info->reloc_mutex);
2007                 goto scrub_continue;
2008         }
2009
2010         /*
2011          * make sure none of the code above managed to slip in a
2012          * delayed item
2013          */
2014         btrfs_assert_delayed_root_empty(root);
2015
2016         WARN_ON(cur_trans != trans->transaction);
2017
2018         /* btrfs_commit_tree_roots is responsible for getting the
2019          * various roots consistent with each other.  Every pointer
2020          * in the tree of tree roots has to point to the most up to date
2021          * root for every subvolume and other tree.  So, we have to keep
2022          * the tree logging code from jumping in and changing any
2023          * of the trees.
2024          *
2025          * At this point in the commit, there can't be any tree-log
2026          * writers, but a little lower down we drop the trans mutex
2027          * and let new people in.  By holding the tree_log_mutex
2028          * from now until after the super is written, we avoid races
2029          * with the tree-log code.
2030          */
2031         mutex_lock(&root->fs_info->tree_log_mutex);
2032
2033         ret = commit_fs_roots(trans, root);
2034         if (ret) {
2035                 mutex_unlock(&root->fs_info->tree_log_mutex);
2036                 mutex_unlock(&root->fs_info->reloc_mutex);
2037                 goto scrub_continue;
2038         }
2039
2040         /*
2041          * Since the transaction is done, we can apply the pending changes
2042          * before the next transaction.
2043          */
2044         btrfs_apply_pending_changes(root->fs_info);
2045
2046         /* commit_fs_roots gets rid of all the tree log roots, it is now
2047          * safe to free the root of tree log roots
2048          */
2049         btrfs_free_log_root_tree(trans, root->fs_info);
2050
2051         /*
2052          * Since fs roots are all committed, we can get a quite accurate
2053          * new_roots. So let's do quota accounting.
2054          */
2055         ret = btrfs_qgroup_account_extents(trans, root->fs_info);
2056         if (ret < 0) {
2057                 mutex_unlock(&root->fs_info->tree_log_mutex);
2058                 mutex_unlock(&root->fs_info->reloc_mutex);
2059                 goto scrub_continue;
2060         }
2061
2062         ret = commit_cowonly_roots(trans, root);
2063         if (ret) {
2064                 mutex_unlock(&root->fs_info->tree_log_mutex);
2065                 mutex_unlock(&root->fs_info->reloc_mutex);
2066                 goto scrub_continue;
2067         }
2068
2069         /*
2070          * The tasks which save the space cache and inode cache may also
2071          * update ->aborted, check it.
2072          */
2073         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
2074                 ret = cur_trans->aborted;
2075                 mutex_unlock(&root->fs_info->tree_log_mutex);
2076                 mutex_unlock(&root->fs_info->reloc_mutex);
2077                 goto scrub_continue;
2078         }
2079
2080         btrfs_prepare_extent_commit(trans, root);
2081
2082         cur_trans = root->fs_info->running_transaction;
2083
2084         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
2085                             root->fs_info->tree_root->node);
2086         list_add_tail(&root->fs_info->tree_root->dirty_list,
2087                       &cur_trans->switch_commits);
2088
2089         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
2090                             root->fs_info->chunk_root->node);
2091         list_add_tail(&root->fs_info->chunk_root->dirty_list,
2092                       &cur_trans->switch_commits);
2093
2094         switch_commit_roots(cur_trans, root->fs_info);
2095
2096         assert_qgroups_uptodate(trans);
2097         ASSERT(list_empty(&cur_trans->dirty_bgs));
2098         ASSERT(list_empty(&cur_trans->io_bgs));
2099         update_super_roots(root);
2100
2101         btrfs_set_super_log_root(root->fs_info->super_copy, 0);
2102         btrfs_set_super_log_root_level(root->fs_info->super_copy, 0);
2103         memcpy(root->fs_info->super_for_commit, root->fs_info->super_copy,
2104                sizeof(*root->fs_info->super_copy));
2105
2106         btrfs_update_commit_device_size(root->fs_info);
2107         btrfs_update_commit_device_bytes_used(root, cur_trans);
2108
2109         clear_bit(BTRFS_INODE_BTREE_LOG1_ERR, &btree_ino->runtime_flags);
2110         clear_bit(BTRFS_INODE_BTREE_LOG2_ERR, &btree_ino->runtime_flags);
2111
2112         btrfs_trans_release_chunk_metadata(trans);
2113
2114         spin_lock(&root->fs_info->trans_lock);
2115         cur_trans->state = TRANS_STATE_UNBLOCKED;
2116         root->fs_info->running_transaction = NULL;
2117         spin_unlock(&root->fs_info->trans_lock);
2118         mutex_unlock(&root->fs_info->reloc_mutex);
2119
2120         wake_up(&root->fs_info->transaction_wait);
2121
2122         ret = btrfs_write_and_wait_transaction(trans, root);
2123         if (ret) {
2124                 btrfs_std_error(root->fs_info, ret,
2125                             "Error while writing out transaction");
2126                 mutex_unlock(&root->fs_info->tree_log_mutex);
2127                 goto scrub_continue;
2128         }
2129
2130         ret = write_ctree_super(trans, root, 0);
2131         if (ret) {
2132                 mutex_unlock(&root->fs_info->tree_log_mutex);
2133                 goto scrub_continue;
2134         }
2135
2136         /*
2137          * the super is written, we can safely allow the tree-loggers
2138          * to go about their business
2139          */
2140         mutex_unlock(&root->fs_info->tree_log_mutex);
2141
2142         btrfs_finish_extent_commit(trans, root);
2143
2144         if (test_bit(BTRFS_TRANS_HAVE_FREE_BGS, &cur_trans->flags))
2145                 btrfs_clear_space_info_full(root->fs_info);
2146
2147         root->fs_info->last_trans_committed = cur_trans->transid;
2148         /*
2149          * We needn't acquire the lock here because there is no other task
2150          * which can change it.
2151          */
2152         cur_trans->state = TRANS_STATE_COMPLETED;
2153         wake_up(&cur_trans->commit_wait);
2154
2155         spin_lock(&root->fs_info->trans_lock);
2156         list_del_init(&cur_trans->list);
2157         spin_unlock(&root->fs_info->trans_lock);
2158
2159         btrfs_put_transaction(cur_trans);
2160         btrfs_put_transaction(cur_trans);
2161
2162         if (trans->type & __TRANS_FREEZABLE)
2163                 sb_end_intwrite(root->fs_info->sb);
2164
2165         trace_btrfs_transaction_commit(root);
2166
2167         btrfs_scrub_continue(root);
2168
2169         if (current->journal_info == trans)
2170                 current->journal_info = NULL;
2171
2172         kmem_cache_free(btrfs_trans_handle_cachep, trans);
2173
2174         if (current != root->fs_info->transaction_kthread &&
2175             current != root->fs_info->cleaner_kthread)
2176                 btrfs_run_delayed_iputs(root);
2177
2178         return ret;
2179
2180 scrub_continue:
2181         btrfs_scrub_continue(root);
2182 cleanup_transaction:
2183         btrfs_trans_release_metadata(trans, root);
2184         btrfs_trans_release_chunk_metadata(trans);
2185         trans->block_rsv = NULL;
2186         btrfs_warn(root->fs_info, "Skipping commit of aborted transaction.");
2187         if (current->journal_info == trans)
2188                 current->journal_info = NULL;
2189         cleanup_transaction(trans, root, ret);
2190
2191         return ret;
2192 }
2193
2194 /*
2195  * return < 0 if error
2196  * 0 if there are no more dead_roots at the time of call
2197  * 1 there are more to be processed, call me again
2198  *
2199  * The return value indicates there are certainly more snapshots to delete, but
2200  * if there comes a new one during processing, it may return 0. We don't mind,
2201  * because btrfs_commit_super will poke cleaner thread and it will process it a
2202  * few seconds later.
2203  */
2204 int btrfs_clean_one_deleted_snapshot(struct btrfs_root *root)
2205 {
2206         int ret;
2207         struct btrfs_fs_info *fs_info = root->fs_info;
2208
2209         spin_lock(&fs_info->trans_lock);
2210         if (list_empty(&fs_info->dead_roots)) {
2211                 spin_unlock(&fs_info->trans_lock);
2212                 return 0;
2213         }
2214         root = list_first_entry(&fs_info->dead_roots,
2215                         struct btrfs_root, root_list);
2216         list_del_init(&root->root_list);
2217         spin_unlock(&fs_info->trans_lock);
2218
2219         pr_debug("BTRFS: cleaner removing %llu\n", root->objectid);
2220
2221         btrfs_kill_all_delayed_nodes(root);
2222
2223         if (btrfs_header_backref_rev(root->node) <
2224                         BTRFS_MIXED_BACKREF_REV)
2225                 ret = btrfs_drop_snapshot(root, NULL, 0, 0);
2226         else
2227                 ret = btrfs_drop_snapshot(root, NULL, 1, 0);
2228
2229         return (ret < 0) ? 0 : 1;
2230 }
2231
2232 void btrfs_apply_pending_changes(struct btrfs_fs_info *fs_info)
2233 {
2234         unsigned long prev;
2235         unsigned long bit;
2236
2237         prev = xchg(&fs_info->pending_changes, 0);
2238         if (!prev)
2239                 return;
2240
2241         bit = 1 << BTRFS_PENDING_SET_INODE_MAP_CACHE;
2242         if (prev & bit)
2243                 btrfs_set_opt(fs_info->mount_opt, INODE_MAP_CACHE);
2244         prev &= ~bit;
2245
2246         bit = 1 << BTRFS_PENDING_CLEAR_INODE_MAP_CACHE;
2247         if (prev & bit)
2248                 btrfs_clear_opt(fs_info->mount_opt, INODE_MAP_CACHE);
2249         prev &= ~bit;
2250
2251         bit = 1 << BTRFS_PENDING_COMMIT;
2252         if (prev & bit)
2253                 btrfs_debug(fs_info, "pending commit done");
2254         prev &= ~bit;
2255
2256         if (prev)
2257                 btrfs_warn(fs_info,
2258                         "unknown pending changes left 0x%lx, ignoring", prev);
2259 }