Btrfs: fix a bug of per-file nocow
[linux-3.10.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
35 #define BTRFS_ROOT_TRANS_TAG 0
36
37 void put_transaction(struct btrfs_transaction *transaction)
38 {
39         WARN_ON(atomic_read(&transaction->use_count) == 0);
40         if (atomic_dec_and_test(&transaction->use_count)) {
41                 BUG_ON(!list_empty(&transaction->list));
42                 WARN_ON(transaction->delayed_refs.root.rb_node);
43                 memset(transaction, 0, sizeof(*transaction));
44                 kmem_cache_free(btrfs_transaction_cachep, transaction);
45         }
46 }
47
48 static noinline void switch_commit_root(struct btrfs_root *root)
49 {
50         free_extent_buffer(root->commit_root);
51         root->commit_root = btrfs_root_node(root);
52 }
53
54 /*
55  * either allocate a new transaction or hop into the existing one
56  */
57 static noinline int join_transaction(struct btrfs_root *root, int type)
58 {
59         struct btrfs_transaction *cur_trans;
60         struct btrfs_fs_info *fs_info = root->fs_info;
61
62         spin_lock(&fs_info->trans_lock);
63 loop:
64         /* The file system has been taken offline. No new transactions. */
65         if (fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR) {
66                 spin_unlock(&fs_info->trans_lock);
67                 return -EROFS;
68         }
69
70         if (fs_info->trans_no_join) {
71                 /* 
72                  * If we are JOIN_NOLOCK we're already committing a current
73                  * transaction, we just need a handle to deal with something
74                  * when committing the transaction, such as inode cache and
75                  * space cache. It is a special case.
76                  */
77                 if (type != TRANS_JOIN_NOLOCK) {
78                         spin_unlock(&fs_info->trans_lock);
79                         return -EBUSY;
80                 }
81         }
82
83         cur_trans = fs_info->running_transaction;
84         if (cur_trans) {
85                 if (cur_trans->aborted) {
86                         spin_unlock(&fs_info->trans_lock);
87                         return cur_trans->aborted;
88                 }
89                 atomic_inc(&cur_trans->use_count);
90                 atomic_inc(&cur_trans->num_writers);
91                 cur_trans->num_joined++;
92                 spin_unlock(&fs_info->trans_lock);
93                 return 0;
94         }
95         spin_unlock(&fs_info->trans_lock);
96
97         /*
98          * If we are ATTACH, we just want to catch the current transaction,
99          * and commit it. If there is no transaction, just return ENOENT.
100          */
101         if (type == TRANS_ATTACH)
102                 return -ENOENT;
103
104         cur_trans = kmem_cache_alloc(btrfs_transaction_cachep, GFP_NOFS);
105         if (!cur_trans)
106                 return -ENOMEM;
107
108         spin_lock(&fs_info->trans_lock);
109         if (fs_info->running_transaction) {
110                 /*
111                  * someone started a transaction after we unlocked.  Make sure
112                  * to redo the trans_no_join checks above
113                  */
114                 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
115                 cur_trans = fs_info->running_transaction;
116                 goto loop;
117         } else if (fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR) {
118                 spin_unlock(&fs_info->trans_lock);
119                 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
120                 return -EROFS;
121         }
122
123         atomic_set(&cur_trans->num_writers, 1);
124         cur_trans->num_joined = 0;
125         init_waitqueue_head(&cur_trans->writer_wait);
126         init_waitqueue_head(&cur_trans->commit_wait);
127         cur_trans->in_commit = 0;
128         cur_trans->blocked = 0;
129         /*
130          * One for this trans handle, one so it will live on until we
131          * commit the transaction.
132          */
133         atomic_set(&cur_trans->use_count, 2);
134         cur_trans->commit_done = 0;
135         cur_trans->start_time = get_seconds();
136
137         cur_trans->delayed_refs.root = RB_ROOT;
138         cur_trans->delayed_refs.num_entries = 0;
139         cur_trans->delayed_refs.num_heads_ready = 0;
140         cur_trans->delayed_refs.num_heads = 0;
141         cur_trans->delayed_refs.flushing = 0;
142         cur_trans->delayed_refs.run_delayed_start = 0;
143
144         /*
145          * although the tree mod log is per file system and not per transaction,
146          * the log must never go across transaction boundaries.
147          */
148         smp_mb();
149         if (!list_empty(&fs_info->tree_mod_seq_list))
150                 WARN(1, KERN_ERR "btrfs: tree_mod_seq_list not empty when "
151                         "creating a fresh transaction\n");
152         if (!RB_EMPTY_ROOT(&fs_info->tree_mod_log))
153                 WARN(1, KERN_ERR "btrfs: tree_mod_log rb tree not empty when "
154                         "creating a fresh transaction\n");
155         atomic_set(&fs_info->tree_mod_seq, 0);
156
157         spin_lock_init(&cur_trans->commit_lock);
158         spin_lock_init(&cur_trans->delayed_refs.lock);
159
160         INIT_LIST_HEAD(&cur_trans->pending_snapshots);
161         list_add_tail(&cur_trans->list, &fs_info->trans_list);
162         extent_io_tree_init(&cur_trans->dirty_pages,
163                              fs_info->btree_inode->i_mapping);
164         fs_info->generation++;
165         cur_trans->transid = fs_info->generation;
166         fs_info->running_transaction = cur_trans;
167         cur_trans->aborted = 0;
168         spin_unlock(&fs_info->trans_lock);
169
170         return 0;
171 }
172
173 /*
174  * this does all the record keeping required to make sure that a reference
175  * counted root is properly recorded in a given transaction.  This is required
176  * to make sure the old root from before we joined the transaction is deleted
177  * when the transaction commits
178  */
179 static int record_root_in_trans(struct btrfs_trans_handle *trans,
180                                struct btrfs_root *root)
181 {
182         if (root->ref_cows && root->last_trans < trans->transid) {
183                 WARN_ON(root == root->fs_info->extent_root);
184                 WARN_ON(root->commit_root != root->node);
185
186                 /*
187                  * see below for in_trans_setup usage rules
188                  * we have the reloc mutex held now, so there
189                  * is only one writer in this function
190                  */
191                 root->in_trans_setup = 1;
192
193                 /* make sure readers find in_trans_setup before
194                  * they find our root->last_trans update
195                  */
196                 smp_wmb();
197
198                 spin_lock(&root->fs_info->fs_roots_radix_lock);
199                 if (root->last_trans == trans->transid) {
200                         spin_unlock(&root->fs_info->fs_roots_radix_lock);
201                         return 0;
202                 }
203                 radix_tree_tag_set(&root->fs_info->fs_roots_radix,
204                            (unsigned long)root->root_key.objectid,
205                            BTRFS_ROOT_TRANS_TAG);
206                 spin_unlock(&root->fs_info->fs_roots_radix_lock);
207                 root->last_trans = trans->transid;
208
209                 /* this is pretty tricky.  We don't want to
210                  * take the relocation lock in btrfs_record_root_in_trans
211                  * unless we're really doing the first setup for this root in
212                  * this transaction.
213                  *
214                  * Normally we'd use root->last_trans as a flag to decide
215                  * if we want to take the expensive mutex.
216                  *
217                  * But, we have to set root->last_trans before we
218                  * init the relocation root, otherwise, we trip over warnings
219                  * in ctree.c.  The solution used here is to flag ourselves
220                  * with root->in_trans_setup.  When this is 1, we're still
221                  * fixing up the reloc trees and everyone must wait.
222                  *
223                  * When this is zero, they can trust root->last_trans and fly
224                  * through btrfs_record_root_in_trans without having to take the
225                  * lock.  smp_wmb() makes sure that all the writes above are
226                  * done before we pop in the zero below
227                  */
228                 btrfs_init_reloc_root(trans, root);
229                 smp_wmb();
230                 root->in_trans_setup = 0;
231         }
232         return 0;
233 }
234
235
236 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
237                                struct btrfs_root *root)
238 {
239         if (!root->ref_cows)
240                 return 0;
241
242         /*
243          * see record_root_in_trans for comments about in_trans_setup usage
244          * and barriers
245          */
246         smp_rmb();
247         if (root->last_trans == trans->transid &&
248             !root->in_trans_setup)
249                 return 0;
250
251         mutex_lock(&root->fs_info->reloc_mutex);
252         record_root_in_trans(trans, root);
253         mutex_unlock(&root->fs_info->reloc_mutex);
254
255         return 0;
256 }
257
258 /* wait for commit against the current transaction to become unblocked
259  * when this is done, it is safe to start a new transaction, but the current
260  * transaction might not be fully on disk.
261  */
262 static void wait_current_trans(struct btrfs_root *root)
263 {
264         struct btrfs_transaction *cur_trans;
265
266         spin_lock(&root->fs_info->trans_lock);
267         cur_trans = root->fs_info->running_transaction;
268         if (cur_trans && cur_trans->blocked) {
269                 atomic_inc(&cur_trans->use_count);
270                 spin_unlock(&root->fs_info->trans_lock);
271
272                 wait_event(root->fs_info->transaction_wait,
273                            !cur_trans->blocked);
274                 put_transaction(cur_trans);
275         } else {
276                 spin_unlock(&root->fs_info->trans_lock);
277         }
278 }
279
280 static int may_wait_transaction(struct btrfs_root *root, int type)
281 {
282         if (root->fs_info->log_root_recovering)
283                 return 0;
284
285         if (type == TRANS_USERSPACE)
286                 return 1;
287
288         if (type == TRANS_START &&
289             !atomic_read(&root->fs_info->open_ioctl_trans))
290                 return 1;
291
292         return 0;
293 }
294
295 static struct btrfs_trans_handle *
296 start_transaction(struct btrfs_root *root, u64 num_items, int type,
297                   enum btrfs_reserve_flush_enum flush)
298 {
299         struct btrfs_trans_handle *h;
300         struct btrfs_transaction *cur_trans;
301         u64 num_bytes = 0;
302         int ret;
303         u64 qgroup_reserved = 0;
304
305         if (root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR)
306                 return ERR_PTR(-EROFS);
307
308         if (current->journal_info) {
309                 WARN_ON(type != TRANS_JOIN && type != TRANS_JOIN_NOLOCK);
310                 h = current->journal_info;
311                 h->use_count++;
312                 WARN_ON(h->use_count > 2);
313                 h->orig_rsv = h->block_rsv;
314                 h->block_rsv = NULL;
315                 goto got_it;
316         }
317
318         /*
319          * Do the reservation before we join the transaction so we can do all
320          * the appropriate flushing if need be.
321          */
322         if (num_items > 0 && root != root->fs_info->chunk_root) {
323                 if (root->fs_info->quota_enabled &&
324                     is_fstree(root->root_key.objectid)) {
325                         qgroup_reserved = num_items * root->leafsize;
326                         ret = btrfs_qgroup_reserve(root, qgroup_reserved);
327                         if (ret)
328                                 return ERR_PTR(ret);
329                 }
330
331                 num_bytes = btrfs_calc_trans_metadata_size(root, num_items);
332                 ret = btrfs_block_rsv_add(root,
333                                           &root->fs_info->trans_block_rsv,
334                                           num_bytes, flush);
335                 if (ret)
336                         return ERR_PTR(ret);
337         }
338 again:
339         h = kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
340         if (!h)
341                 return ERR_PTR(-ENOMEM);
342
343         /*
344          * If we are JOIN_NOLOCK we're already committing a transaction and
345          * waiting on this guy, so we don't need to do the sb_start_intwrite
346          * because we're already holding a ref.  We need this because we could
347          * have raced in and did an fsync() on a file which can kick a commit
348          * and then we deadlock with somebody doing a freeze.
349          *
350          * If we are ATTACH, it means we just want to catch the current
351          * transaction and commit it, so we needn't do sb_start_intwrite(). 
352          */
353         if (type < TRANS_JOIN_NOLOCK)
354                 sb_start_intwrite(root->fs_info->sb);
355
356         if (may_wait_transaction(root, type))
357                 wait_current_trans(root);
358
359         do {
360                 ret = join_transaction(root, type);
361                 if (ret == -EBUSY)
362                         wait_current_trans(root);
363         } while (ret == -EBUSY);
364
365         if (ret < 0) {
366                 /* We must get the transaction if we are JOIN_NOLOCK. */
367                 BUG_ON(type == TRANS_JOIN_NOLOCK);
368
369                 if (type < TRANS_JOIN_NOLOCK)
370                         sb_end_intwrite(root->fs_info->sb);
371                 kmem_cache_free(btrfs_trans_handle_cachep, h);
372                 return ERR_PTR(ret);
373         }
374
375         cur_trans = root->fs_info->running_transaction;
376
377         h->transid = cur_trans->transid;
378         h->transaction = cur_trans;
379         h->blocks_used = 0;
380         h->bytes_reserved = 0;
381         h->root = root;
382         h->delayed_ref_updates = 0;
383         h->use_count = 1;
384         h->adding_csums = 0;
385         h->block_rsv = NULL;
386         h->orig_rsv = NULL;
387         h->aborted = 0;
388         h->qgroup_reserved = qgroup_reserved;
389         h->delayed_ref_elem.seq = 0;
390         h->type = type;
391         INIT_LIST_HEAD(&h->qgroup_ref_list);
392         INIT_LIST_HEAD(&h->new_bgs);
393
394         smp_mb();
395         if (cur_trans->blocked && may_wait_transaction(root, type)) {
396                 btrfs_commit_transaction(h, root);
397                 goto again;
398         }
399
400         if (num_bytes) {
401                 trace_btrfs_space_reservation(root->fs_info, "transaction",
402                                               h->transid, num_bytes, 1);
403                 h->block_rsv = &root->fs_info->trans_block_rsv;
404                 h->bytes_reserved = num_bytes;
405         }
406
407 got_it:
408         btrfs_record_root_in_trans(h, root);
409
410         if (!current->journal_info && type != TRANS_USERSPACE)
411                 current->journal_info = h;
412         return h;
413 }
414
415 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
416                                                    int num_items)
417 {
418         return start_transaction(root, num_items, TRANS_START,
419                                  BTRFS_RESERVE_FLUSH_ALL);
420 }
421
422 struct btrfs_trans_handle *btrfs_start_transaction_lflush(
423                                         struct btrfs_root *root, int num_items)
424 {
425         return start_transaction(root, num_items, TRANS_START,
426                                  BTRFS_RESERVE_FLUSH_LIMIT);
427 }
428
429 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
430 {
431         return start_transaction(root, 0, TRANS_JOIN, 0);
432 }
433
434 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root)
435 {
436         return start_transaction(root, 0, TRANS_JOIN_NOLOCK, 0);
437 }
438
439 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *root)
440 {
441         return start_transaction(root, 0, TRANS_USERSPACE, 0);
442 }
443
444 struct btrfs_trans_handle *btrfs_attach_transaction(struct btrfs_root *root)
445 {
446         return start_transaction(root, 0, TRANS_ATTACH, 0);
447 }
448
449 /* wait for a transaction commit to be fully complete */
450 static noinline void wait_for_commit(struct btrfs_root *root,
451                                     struct btrfs_transaction *commit)
452 {
453         wait_event(commit->commit_wait, commit->commit_done);
454 }
455
456 int btrfs_wait_for_commit(struct btrfs_root *root, u64 transid)
457 {
458         struct btrfs_transaction *cur_trans = NULL, *t;
459         int ret = 0;
460
461         if (transid) {
462                 if (transid <= root->fs_info->last_trans_committed)
463                         goto out;
464
465                 ret = -EINVAL;
466                 /* find specified transaction */
467                 spin_lock(&root->fs_info->trans_lock);
468                 list_for_each_entry(t, &root->fs_info->trans_list, list) {
469                         if (t->transid == transid) {
470                                 cur_trans = t;
471                                 atomic_inc(&cur_trans->use_count);
472                                 ret = 0;
473                                 break;
474                         }
475                         if (t->transid > transid) {
476                                 ret = 0;
477                                 break;
478                         }
479                 }
480                 spin_unlock(&root->fs_info->trans_lock);
481                 /* The specified transaction doesn't exist */
482                 if (!cur_trans)
483                         goto out;
484         } else {
485                 /* find newest transaction that is committing | committed */
486                 spin_lock(&root->fs_info->trans_lock);
487                 list_for_each_entry_reverse(t, &root->fs_info->trans_list,
488                                             list) {
489                         if (t->in_commit) {
490                                 if (t->commit_done)
491                                         break;
492                                 cur_trans = t;
493                                 atomic_inc(&cur_trans->use_count);
494                                 break;
495                         }
496                 }
497                 spin_unlock(&root->fs_info->trans_lock);
498                 if (!cur_trans)
499                         goto out;  /* nothing committing|committed */
500         }
501
502         wait_for_commit(root, cur_trans);
503         put_transaction(cur_trans);
504 out:
505         return ret;
506 }
507
508 void btrfs_throttle(struct btrfs_root *root)
509 {
510         if (!atomic_read(&root->fs_info->open_ioctl_trans))
511                 wait_current_trans(root);
512 }
513
514 static int should_end_transaction(struct btrfs_trans_handle *trans,
515                                   struct btrfs_root *root)
516 {
517         int ret;
518
519         ret = btrfs_block_rsv_check(root, &root->fs_info->global_block_rsv, 5);
520         return ret ? 1 : 0;
521 }
522
523 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans,
524                                  struct btrfs_root *root)
525 {
526         struct btrfs_transaction *cur_trans = trans->transaction;
527         int updates;
528         int err;
529
530         smp_mb();
531         if (cur_trans->blocked || cur_trans->delayed_refs.flushing)
532                 return 1;
533
534         updates = trans->delayed_ref_updates;
535         trans->delayed_ref_updates = 0;
536         if (updates) {
537                 err = btrfs_run_delayed_refs(trans, root, updates);
538                 if (err) /* Error code will also eval true */
539                         return err;
540         }
541
542         return should_end_transaction(trans, root);
543 }
544
545 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
546                           struct btrfs_root *root, int throttle)
547 {
548         struct btrfs_transaction *cur_trans = trans->transaction;
549         struct btrfs_fs_info *info = root->fs_info;
550         int count = 0;
551         int lock = (trans->type != TRANS_JOIN_NOLOCK);
552         int err = 0;
553
554         if (--trans->use_count) {
555                 trans->block_rsv = trans->orig_rsv;
556                 return 0;
557         }
558
559         /*
560          * do the qgroup accounting as early as possible
561          */
562         err = btrfs_delayed_refs_qgroup_accounting(trans, info);
563
564         btrfs_trans_release_metadata(trans, root);
565         trans->block_rsv = NULL;
566         /*
567          * the same root has to be passed to start_transaction and
568          * end_transaction. Subvolume quota depends on this.
569          */
570         WARN_ON(trans->root != root);
571
572         if (trans->qgroup_reserved) {
573                 btrfs_qgroup_free(root, trans->qgroup_reserved);
574                 trans->qgroup_reserved = 0;
575         }
576
577         if (!list_empty(&trans->new_bgs))
578                 btrfs_create_pending_block_groups(trans, root);
579
580         while (count < 2) {
581                 unsigned long cur = trans->delayed_ref_updates;
582                 trans->delayed_ref_updates = 0;
583                 if (cur &&
584                     trans->transaction->delayed_refs.num_heads_ready > 64) {
585                         trans->delayed_ref_updates = 0;
586                         btrfs_run_delayed_refs(trans, root, cur);
587                 } else {
588                         break;
589                 }
590                 count++;
591         }
592         btrfs_trans_release_metadata(trans, root);
593         trans->block_rsv = NULL;
594
595         if (!list_empty(&trans->new_bgs))
596                 btrfs_create_pending_block_groups(trans, root);
597
598         if (lock && !atomic_read(&root->fs_info->open_ioctl_trans) &&
599             should_end_transaction(trans, root)) {
600                 trans->transaction->blocked = 1;
601                 smp_wmb();
602         }
603
604         if (lock && cur_trans->blocked && !cur_trans->in_commit) {
605                 if (throttle) {
606                         /*
607                          * We may race with somebody else here so end up having
608                          * to call end_transaction on ourselves again, so inc
609                          * our use_count.
610                          */
611                         trans->use_count++;
612                         return btrfs_commit_transaction(trans, root);
613                 } else {
614                         wake_up_process(info->transaction_kthread);
615                 }
616         }
617
618         if (trans->type < TRANS_JOIN_NOLOCK)
619                 sb_end_intwrite(root->fs_info->sb);
620
621         WARN_ON(cur_trans != info->running_transaction);
622         WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
623         atomic_dec(&cur_trans->num_writers);
624
625         smp_mb();
626         if (waitqueue_active(&cur_trans->writer_wait))
627                 wake_up(&cur_trans->writer_wait);
628         put_transaction(cur_trans);
629
630         if (current->journal_info == trans)
631                 current->journal_info = NULL;
632
633         if (throttle)
634                 btrfs_run_delayed_iputs(root);
635
636         if (trans->aborted ||
637             root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR) {
638                 err = -EIO;
639         }
640         assert_qgroups_uptodate(trans);
641
642         memset(trans, 0, sizeof(*trans));
643         kmem_cache_free(btrfs_trans_handle_cachep, trans);
644         return err;
645 }
646
647 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
648                           struct btrfs_root *root)
649 {
650         int ret;
651
652         ret = __btrfs_end_transaction(trans, root, 0);
653         if (ret)
654                 return ret;
655         return 0;
656 }
657
658 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
659                                    struct btrfs_root *root)
660 {
661         int ret;
662
663         ret = __btrfs_end_transaction(trans, root, 1);
664         if (ret)
665                 return ret;
666         return 0;
667 }
668
669 int btrfs_end_transaction_dmeta(struct btrfs_trans_handle *trans,
670                                 struct btrfs_root *root)
671 {
672         return __btrfs_end_transaction(trans, root, 1);
673 }
674
675 /*
676  * when btree blocks are allocated, they have some corresponding bits set for
677  * them in one of two extent_io trees.  This is used to make sure all of
678  * those extents are sent to disk but does not wait on them
679  */
680 int btrfs_write_marked_extents(struct btrfs_root *root,
681                                struct extent_io_tree *dirty_pages, int mark)
682 {
683         int err = 0;
684         int werr = 0;
685         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
686         struct extent_state *cached_state = NULL;
687         u64 start = 0;
688         u64 end;
689
690         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
691                                       mark, &cached_state)) {
692                 convert_extent_bit(dirty_pages, start, end, EXTENT_NEED_WAIT,
693                                    mark, &cached_state, GFP_NOFS);
694                 cached_state = NULL;
695                 err = filemap_fdatawrite_range(mapping, start, end);
696                 if (err)
697                         werr = err;
698                 cond_resched();
699                 start = end + 1;
700         }
701         if (err)
702                 werr = err;
703         return werr;
704 }
705
706 /*
707  * when btree blocks are allocated, they have some corresponding bits set for
708  * them in one of two extent_io trees.  This is used to make sure all of
709  * those extents are on disk for transaction or log commit.  We wait
710  * on all the pages and clear them from the dirty pages state tree
711  */
712 int btrfs_wait_marked_extents(struct btrfs_root *root,
713                               struct extent_io_tree *dirty_pages, int mark)
714 {
715         int err = 0;
716         int werr = 0;
717         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
718         struct extent_state *cached_state = NULL;
719         u64 start = 0;
720         u64 end;
721
722         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
723                                       EXTENT_NEED_WAIT, &cached_state)) {
724                 clear_extent_bit(dirty_pages, start, end, EXTENT_NEED_WAIT,
725                                  0, 0, &cached_state, GFP_NOFS);
726                 err = filemap_fdatawait_range(mapping, start, end);
727                 if (err)
728                         werr = err;
729                 cond_resched();
730                 start = end + 1;
731         }
732         if (err)
733                 werr = err;
734         return werr;
735 }
736
737 /*
738  * when btree blocks are allocated, they have some corresponding bits set for
739  * them in one of two extent_io trees.  This is used to make sure all of
740  * those extents are on disk for transaction or log commit
741  */
742 int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
743                                 struct extent_io_tree *dirty_pages, int mark)
744 {
745         int ret;
746         int ret2;
747
748         ret = btrfs_write_marked_extents(root, dirty_pages, mark);
749         ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
750
751         if (ret)
752                 return ret;
753         if (ret2)
754                 return ret2;
755         return 0;
756 }
757
758 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
759                                      struct btrfs_root *root)
760 {
761         if (!trans || !trans->transaction) {
762                 struct inode *btree_inode;
763                 btree_inode = root->fs_info->btree_inode;
764                 return filemap_write_and_wait(btree_inode->i_mapping);
765         }
766         return btrfs_write_and_wait_marked_extents(root,
767                                            &trans->transaction->dirty_pages,
768                                            EXTENT_DIRTY);
769 }
770
771 /*
772  * this is used to update the root pointer in the tree of tree roots.
773  *
774  * But, in the case of the extent allocation tree, updating the root
775  * pointer may allocate blocks which may change the root of the extent
776  * allocation tree.
777  *
778  * So, this loops and repeats and makes sure the cowonly root didn't
779  * change while the root pointer was being updated in the metadata.
780  */
781 static int update_cowonly_root(struct btrfs_trans_handle *trans,
782                                struct btrfs_root *root)
783 {
784         int ret;
785         u64 old_root_bytenr;
786         u64 old_root_used;
787         struct btrfs_root *tree_root = root->fs_info->tree_root;
788
789         old_root_used = btrfs_root_used(&root->root_item);
790         btrfs_write_dirty_block_groups(trans, root);
791
792         while (1) {
793                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
794                 if (old_root_bytenr == root->node->start &&
795                     old_root_used == btrfs_root_used(&root->root_item))
796                         break;
797
798                 btrfs_set_root_node(&root->root_item, root->node);
799                 ret = btrfs_update_root(trans, tree_root,
800                                         &root->root_key,
801                                         &root->root_item);
802                 if (ret)
803                         return ret;
804
805                 old_root_used = btrfs_root_used(&root->root_item);
806                 ret = btrfs_write_dirty_block_groups(trans, root);
807                 if (ret)
808                         return ret;
809         }
810
811         if (root != root->fs_info->extent_root)
812                 switch_commit_root(root);
813
814         return 0;
815 }
816
817 /*
818  * update all the cowonly tree roots on disk
819  *
820  * The error handling in this function may not be obvious. Any of the
821  * failures will cause the file system to go offline. We still need
822  * to clean up the delayed refs.
823  */
824 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
825                                          struct btrfs_root *root)
826 {
827         struct btrfs_fs_info *fs_info = root->fs_info;
828         struct list_head *next;
829         struct extent_buffer *eb;
830         int ret;
831
832         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
833         if (ret)
834                 return ret;
835
836         eb = btrfs_lock_root_node(fs_info->tree_root);
837         ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL,
838                               0, &eb);
839         btrfs_tree_unlock(eb);
840         free_extent_buffer(eb);
841
842         if (ret)
843                 return ret;
844
845         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
846         if (ret)
847                 return ret;
848
849         ret = btrfs_run_dev_stats(trans, root->fs_info);
850         WARN_ON(ret);
851         ret = btrfs_run_dev_replace(trans, root->fs_info);
852         WARN_ON(ret);
853
854         ret = btrfs_run_qgroups(trans, root->fs_info);
855         BUG_ON(ret);
856
857         /* run_qgroups might have added some more refs */
858         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
859         BUG_ON(ret);
860
861         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
862                 next = fs_info->dirty_cowonly_roots.next;
863                 list_del_init(next);
864                 root = list_entry(next, struct btrfs_root, dirty_list);
865
866                 ret = update_cowonly_root(trans, root);
867                 if (ret)
868                         return ret;
869         }
870
871         down_write(&fs_info->extent_commit_sem);
872         switch_commit_root(fs_info->extent_root);
873         up_write(&fs_info->extent_commit_sem);
874
875         btrfs_after_dev_replace_commit(fs_info);
876
877         return 0;
878 }
879
880 /*
881  * dead roots are old snapshots that need to be deleted.  This allocates
882  * a dirty root struct and adds it into the list of dead roots that need to
883  * be deleted
884  */
885 int btrfs_add_dead_root(struct btrfs_root *root)
886 {
887         spin_lock(&root->fs_info->trans_lock);
888         list_add(&root->root_list, &root->fs_info->dead_roots);
889         spin_unlock(&root->fs_info->trans_lock);
890         return 0;
891 }
892
893 /*
894  * update all the cowonly tree roots on disk
895  */
896 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
897                                     struct btrfs_root *root)
898 {
899         struct btrfs_root *gang[8];
900         struct btrfs_fs_info *fs_info = root->fs_info;
901         int i;
902         int ret;
903         int err = 0;
904
905         spin_lock(&fs_info->fs_roots_radix_lock);
906         while (1) {
907                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
908                                                  (void **)gang, 0,
909                                                  ARRAY_SIZE(gang),
910                                                  BTRFS_ROOT_TRANS_TAG);
911                 if (ret == 0)
912                         break;
913                 for (i = 0; i < ret; i++) {
914                         root = gang[i];
915                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
916                                         (unsigned long)root->root_key.objectid,
917                                         BTRFS_ROOT_TRANS_TAG);
918                         spin_unlock(&fs_info->fs_roots_radix_lock);
919
920                         btrfs_free_log(trans, root);
921                         btrfs_update_reloc_root(trans, root);
922                         btrfs_orphan_commit_root(trans, root);
923
924                         btrfs_save_ino_cache(root, trans);
925
926                         /* see comments in should_cow_block() */
927                         root->force_cow = 0;
928                         smp_wmb();
929
930                         if (root->commit_root != root->node) {
931                                 mutex_lock(&root->fs_commit_mutex);
932                                 switch_commit_root(root);
933                                 btrfs_unpin_free_ino(root);
934                                 mutex_unlock(&root->fs_commit_mutex);
935
936                                 btrfs_set_root_node(&root->root_item,
937                                                     root->node);
938                         }
939
940                         err = btrfs_update_root(trans, fs_info->tree_root,
941                                                 &root->root_key,
942                                                 &root->root_item);
943                         spin_lock(&fs_info->fs_roots_radix_lock);
944                         if (err)
945                                 break;
946                 }
947         }
948         spin_unlock(&fs_info->fs_roots_radix_lock);
949         return err;
950 }
951
952 /*
953  * defrag a given btree.  If cacheonly == 1, this won't read from the disk,
954  * otherwise every leaf in the btree is read and defragged.
955  */
956 int btrfs_defrag_root(struct btrfs_root *root, int cacheonly)
957 {
958         struct btrfs_fs_info *info = root->fs_info;
959         struct btrfs_trans_handle *trans;
960         int ret;
961
962         if (xchg(&root->defrag_running, 1))
963                 return 0;
964
965         while (1) {
966                 trans = btrfs_start_transaction(root, 0);
967                 if (IS_ERR(trans))
968                         return PTR_ERR(trans);
969
970                 ret = btrfs_defrag_leaves(trans, root, cacheonly);
971
972                 btrfs_end_transaction(trans, root);
973                 btrfs_btree_balance_dirty(info->tree_root);
974                 cond_resched();
975
976                 if (btrfs_fs_closing(root->fs_info) || ret != -EAGAIN)
977                         break;
978         }
979         root->defrag_running = 0;
980         return ret;
981 }
982
983 /*
984  * new snapshots need to be created at a very specific time in the
985  * transaction commit.  This does the actual creation
986  */
987 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
988                                    struct btrfs_fs_info *fs_info,
989                                    struct btrfs_pending_snapshot *pending)
990 {
991         struct btrfs_key key;
992         struct btrfs_root_item *new_root_item;
993         struct btrfs_root *tree_root = fs_info->tree_root;
994         struct btrfs_root *root = pending->root;
995         struct btrfs_root *parent_root;
996         struct btrfs_block_rsv *rsv;
997         struct inode *parent_inode;
998         struct btrfs_path *path;
999         struct btrfs_dir_item *dir_item;
1000         struct dentry *parent;
1001         struct dentry *dentry;
1002         struct extent_buffer *tmp;
1003         struct extent_buffer *old;
1004         struct timespec cur_time = CURRENT_TIME;
1005         int ret;
1006         u64 to_reserve = 0;
1007         u64 index = 0;
1008         u64 objectid;
1009         u64 root_flags;
1010         uuid_le new_uuid;
1011
1012         path = btrfs_alloc_path();
1013         if (!path) {
1014                 ret = pending->error = -ENOMEM;
1015                 goto path_alloc_fail;
1016         }
1017
1018         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
1019         if (!new_root_item) {
1020                 ret = pending->error = -ENOMEM;
1021                 goto root_item_alloc_fail;
1022         }
1023
1024         ret = btrfs_find_free_objectid(tree_root, &objectid);
1025         if (ret) {
1026                 pending->error = ret;
1027                 goto no_free_objectid;
1028         }
1029
1030         btrfs_reloc_pre_snapshot(trans, pending, &to_reserve);
1031
1032         if (to_reserve > 0) {
1033                 ret = btrfs_block_rsv_add(root, &pending->block_rsv,
1034                                           to_reserve,
1035                                           BTRFS_RESERVE_NO_FLUSH);
1036                 if (ret) {
1037                         pending->error = ret;
1038                         goto no_free_objectid;
1039                 }
1040         }
1041
1042         ret = btrfs_qgroup_inherit(trans, fs_info, root->root_key.objectid,
1043                                    objectid, pending->inherit);
1044         if (ret) {
1045                 pending->error = ret;
1046                 goto no_free_objectid;
1047         }
1048
1049         key.objectid = objectid;
1050         key.offset = (u64)-1;
1051         key.type = BTRFS_ROOT_ITEM_KEY;
1052
1053         rsv = trans->block_rsv;
1054         trans->block_rsv = &pending->block_rsv;
1055
1056         dentry = pending->dentry;
1057         parent = dget_parent(dentry);
1058         parent_inode = parent->d_inode;
1059         parent_root = BTRFS_I(parent_inode)->root;
1060         record_root_in_trans(trans, parent_root);
1061
1062         /*
1063          * insert the directory item
1064          */
1065         ret = btrfs_set_inode_index(parent_inode, &index);
1066         BUG_ON(ret); /* -ENOMEM */
1067
1068         /* check if there is a file/dir which has the same name. */
1069         dir_item = btrfs_lookup_dir_item(NULL, parent_root, path,
1070                                          btrfs_ino(parent_inode),
1071                                          dentry->d_name.name,
1072                                          dentry->d_name.len, 0);
1073         if (dir_item != NULL && !IS_ERR(dir_item)) {
1074                 pending->error = -EEXIST;
1075                 goto fail;
1076         } else if (IS_ERR(dir_item)) {
1077                 ret = PTR_ERR(dir_item);
1078                 btrfs_abort_transaction(trans, root, ret);
1079                 goto fail;
1080         }
1081         btrfs_release_path(path);
1082
1083         /*
1084          * pull in the delayed directory update
1085          * and the delayed inode item
1086          * otherwise we corrupt the FS during
1087          * snapshot
1088          */
1089         ret = btrfs_run_delayed_items(trans, root);
1090         if (ret) {      /* Transaction aborted */
1091                 btrfs_abort_transaction(trans, root, ret);
1092                 goto fail;
1093         }
1094
1095         record_root_in_trans(trans, root);
1096         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
1097         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
1098         btrfs_check_and_init_root_item(new_root_item);
1099
1100         root_flags = btrfs_root_flags(new_root_item);
1101         if (pending->readonly)
1102                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
1103         else
1104                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
1105         btrfs_set_root_flags(new_root_item, root_flags);
1106
1107         btrfs_set_root_generation_v2(new_root_item,
1108                         trans->transid);
1109         uuid_le_gen(&new_uuid);
1110         memcpy(new_root_item->uuid, new_uuid.b, BTRFS_UUID_SIZE);
1111         memcpy(new_root_item->parent_uuid, root->root_item.uuid,
1112                         BTRFS_UUID_SIZE);
1113         new_root_item->otime.sec = cpu_to_le64(cur_time.tv_sec);
1114         new_root_item->otime.nsec = cpu_to_le32(cur_time.tv_nsec);
1115         btrfs_set_root_otransid(new_root_item, trans->transid);
1116         memset(&new_root_item->stime, 0, sizeof(new_root_item->stime));
1117         memset(&new_root_item->rtime, 0, sizeof(new_root_item->rtime));
1118         btrfs_set_root_stransid(new_root_item, 0);
1119         btrfs_set_root_rtransid(new_root_item, 0);
1120
1121         old = btrfs_lock_root_node(root);
1122         ret = btrfs_cow_block(trans, root, old, NULL, 0, &old);
1123         if (ret) {
1124                 btrfs_tree_unlock(old);
1125                 free_extent_buffer(old);
1126                 btrfs_abort_transaction(trans, root, ret);
1127                 goto fail;
1128         }
1129
1130         btrfs_set_lock_blocking(old);
1131
1132         ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1133         /* clean up in any case */
1134         btrfs_tree_unlock(old);
1135         free_extent_buffer(old);
1136         if (ret) {
1137                 btrfs_abort_transaction(trans, root, ret);
1138                 goto fail;
1139         }
1140
1141         /* see comments in should_cow_block() */
1142         root->force_cow = 1;
1143         smp_wmb();
1144
1145         btrfs_set_root_node(new_root_item, tmp);
1146         /* record when the snapshot was created in key.offset */
1147         key.offset = trans->transid;
1148         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1149         btrfs_tree_unlock(tmp);
1150         free_extent_buffer(tmp);
1151         if (ret) {
1152                 btrfs_abort_transaction(trans, root, ret);
1153                 goto fail;
1154         }
1155
1156         /*
1157          * insert root back/forward references
1158          */
1159         ret = btrfs_add_root_ref(trans, tree_root, objectid,
1160                                  parent_root->root_key.objectid,
1161                                  btrfs_ino(parent_inode), index,
1162                                  dentry->d_name.name, dentry->d_name.len);
1163         if (ret) {
1164                 btrfs_abort_transaction(trans, root, ret);
1165                 goto fail;
1166         }
1167
1168         key.offset = (u64)-1;
1169         pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
1170         if (IS_ERR(pending->snap)) {
1171                 ret = PTR_ERR(pending->snap);
1172                 btrfs_abort_transaction(trans, root, ret);
1173                 goto fail;
1174         }
1175
1176         ret = btrfs_reloc_post_snapshot(trans, pending);
1177         if (ret) {
1178                 btrfs_abort_transaction(trans, root, ret);
1179                 goto fail;
1180         }
1181
1182         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1183         if (ret) {
1184                 btrfs_abort_transaction(trans, root, ret);
1185                 goto fail;
1186         }
1187
1188         ret = btrfs_insert_dir_item(trans, parent_root,
1189                                     dentry->d_name.name, dentry->d_name.len,
1190                                     parent_inode, &key,
1191                                     BTRFS_FT_DIR, index);
1192         /* We have check then name at the beginning, so it is impossible. */
1193         BUG_ON(ret == -EEXIST || ret == -EOVERFLOW);
1194         if (ret) {
1195                 btrfs_abort_transaction(trans, root, ret);
1196                 goto fail;
1197         }
1198
1199         btrfs_i_size_write(parent_inode, parent_inode->i_size +
1200                                          dentry->d_name.len * 2);
1201         parent_inode->i_mtime = parent_inode->i_ctime = CURRENT_TIME;
1202         ret = btrfs_update_inode_fallback(trans, parent_root, parent_inode);
1203         if (ret)
1204                 btrfs_abort_transaction(trans, root, ret);
1205 fail:
1206         dput(parent);
1207         trans->block_rsv = rsv;
1208 no_free_objectid:
1209         kfree(new_root_item);
1210 root_item_alloc_fail:
1211         btrfs_free_path(path);
1212 path_alloc_fail:
1213         btrfs_block_rsv_release(root, &pending->block_rsv, (u64)-1);
1214         return ret;
1215 }
1216
1217 /*
1218  * create all the snapshots we've scheduled for creation
1219  */
1220 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
1221                                              struct btrfs_fs_info *fs_info)
1222 {
1223         struct btrfs_pending_snapshot *pending;
1224         struct list_head *head = &trans->transaction->pending_snapshots;
1225
1226         list_for_each_entry(pending, head, list)
1227                 create_pending_snapshot(trans, fs_info, pending);
1228         return 0;
1229 }
1230
1231 static void update_super_roots(struct btrfs_root *root)
1232 {
1233         struct btrfs_root_item *root_item;
1234         struct btrfs_super_block *super;
1235
1236         super = root->fs_info->super_copy;
1237
1238         root_item = &root->fs_info->chunk_root->root_item;
1239         super->chunk_root = root_item->bytenr;
1240         super->chunk_root_generation = root_item->generation;
1241         super->chunk_root_level = root_item->level;
1242
1243         root_item = &root->fs_info->tree_root->root_item;
1244         super->root = root_item->bytenr;
1245         super->generation = root_item->generation;
1246         super->root_level = root_item->level;
1247         if (btrfs_test_opt(root, SPACE_CACHE))
1248                 super->cache_generation = root_item->generation;
1249 }
1250
1251 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1252 {
1253         int ret = 0;
1254         spin_lock(&info->trans_lock);
1255         if (info->running_transaction)
1256                 ret = info->running_transaction->in_commit;
1257         spin_unlock(&info->trans_lock);
1258         return ret;
1259 }
1260
1261 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1262 {
1263         int ret = 0;
1264         spin_lock(&info->trans_lock);
1265         if (info->running_transaction)
1266                 ret = info->running_transaction->blocked;
1267         spin_unlock(&info->trans_lock);
1268         return ret;
1269 }
1270
1271 /*
1272  * wait for the current transaction commit to start and block subsequent
1273  * transaction joins
1274  */
1275 static void wait_current_trans_commit_start(struct btrfs_root *root,
1276                                             struct btrfs_transaction *trans)
1277 {
1278         wait_event(root->fs_info->transaction_blocked_wait, trans->in_commit);
1279 }
1280
1281 /*
1282  * wait for the current transaction to start and then become unblocked.
1283  * caller holds ref.
1284  */
1285 static void wait_current_trans_commit_start_and_unblock(struct btrfs_root *root,
1286                                          struct btrfs_transaction *trans)
1287 {
1288         wait_event(root->fs_info->transaction_wait,
1289                    trans->commit_done || (trans->in_commit && !trans->blocked));
1290 }
1291
1292 /*
1293  * commit transactions asynchronously. once btrfs_commit_transaction_async
1294  * returns, any subsequent transaction will not be allowed to join.
1295  */
1296 struct btrfs_async_commit {
1297         struct btrfs_trans_handle *newtrans;
1298         struct btrfs_root *root;
1299         struct delayed_work work;
1300 };
1301
1302 static void do_async_commit(struct work_struct *work)
1303 {
1304         struct btrfs_async_commit *ac =
1305                 container_of(work, struct btrfs_async_commit, work.work);
1306
1307         /*
1308          * We've got freeze protection passed with the transaction.
1309          * Tell lockdep about it.
1310          */
1311         if (ac->newtrans->type < TRANS_JOIN_NOLOCK)
1312                 rwsem_acquire_read(
1313                      &ac->root->fs_info->sb->s_writers.lock_map[SB_FREEZE_FS-1],
1314                      0, 1, _THIS_IP_);
1315
1316         current->journal_info = ac->newtrans;
1317
1318         btrfs_commit_transaction(ac->newtrans, ac->root);
1319         kfree(ac);
1320 }
1321
1322 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1323                                    struct btrfs_root *root,
1324                                    int wait_for_unblock)
1325 {
1326         struct btrfs_async_commit *ac;
1327         struct btrfs_transaction *cur_trans;
1328
1329         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1330         if (!ac)
1331                 return -ENOMEM;
1332
1333         INIT_DELAYED_WORK(&ac->work, do_async_commit);
1334         ac->root = root;
1335         ac->newtrans = btrfs_join_transaction(root);
1336         if (IS_ERR(ac->newtrans)) {
1337                 int err = PTR_ERR(ac->newtrans);
1338                 kfree(ac);
1339                 return err;
1340         }
1341
1342         /* take transaction reference */
1343         cur_trans = trans->transaction;
1344         atomic_inc(&cur_trans->use_count);
1345
1346         btrfs_end_transaction(trans, root);
1347
1348         /*
1349          * Tell lockdep we've released the freeze rwsem, since the
1350          * async commit thread will be the one to unlock it.
1351          */
1352         if (trans->type < TRANS_JOIN_NOLOCK)
1353                 rwsem_release(
1354                         &root->fs_info->sb->s_writers.lock_map[SB_FREEZE_FS-1],
1355                         1, _THIS_IP_);
1356
1357         schedule_delayed_work(&ac->work, 0);
1358
1359         /* wait for transaction to start and unblock */
1360         if (wait_for_unblock)
1361                 wait_current_trans_commit_start_and_unblock(root, cur_trans);
1362         else
1363                 wait_current_trans_commit_start(root, cur_trans);
1364
1365         if (current->journal_info == trans)
1366                 current->journal_info = NULL;
1367
1368         put_transaction(cur_trans);
1369         return 0;
1370 }
1371
1372
1373 static void cleanup_transaction(struct btrfs_trans_handle *trans,
1374                                 struct btrfs_root *root, int err)
1375 {
1376         struct btrfs_transaction *cur_trans = trans->transaction;
1377
1378         WARN_ON(trans->use_count > 1);
1379
1380         btrfs_abort_transaction(trans, root, err);
1381
1382         spin_lock(&root->fs_info->trans_lock);
1383         list_del_init(&cur_trans->list);
1384         if (cur_trans == root->fs_info->running_transaction) {
1385                 root->fs_info->running_transaction = NULL;
1386                 root->fs_info->trans_no_join = 0;
1387         }
1388         spin_unlock(&root->fs_info->trans_lock);
1389
1390         btrfs_cleanup_one_transaction(trans->transaction, root);
1391
1392         put_transaction(cur_trans);
1393         put_transaction(cur_trans);
1394
1395         trace_btrfs_transaction_commit(root);
1396
1397         btrfs_scrub_continue(root);
1398
1399         if (current->journal_info == trans)
1400                 current->journal_info = NULL;
1401
1402         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1403 }
1404
1405 static int btrfs_flush_all_pending_stuffs(struct btrfs_trans_handle *trans,
1406                                           struct btrfs_root *root)
1407 {
1408         int flush_on_commit = btrfs_test_opt(root, FLUSHONCOMMIT);
1409         int snap_pending = 0;
1410         int ret;
1411
1412         if (!flush_on_commit) {
1413                 spin_lock(&root->fs_info->trans_lock);
1414                 if (!list_empty(&trans->transaction->pending_snapshots))
1415                         snap_pending = 1;
1416                 spin_unlock(&root->fs_info->trans_lock);
1417         }
1418
1419         if (flush_on_commit || snap_pending) {
1420                 btrfs_start_delalloc_inodes(root, 1);
1421                 btrfs_wait_ordered_extents(root, 1);
1422         }
1423
1424         ret = btrfs_run_delayed_items(trans, root);
1425         if (ret)
1426                 return ret;
1427
1428         /*
1429          * running the delayed items may have added new refs. account
1430          * them now so that they hinder processing of more delayed refs
1431          * as little as possible.
1432          */
1433         btrfs_delayed_refs_qgroup_accounting(trans, root->fs_info);
1434
1435         /*
1436          * rename don't use btrfs_join_transaction, so, once we
1437          * set the transaction to blocked above, we aren't going
1438          * to get any new ordered operations.  We can safely run
1439          * it here and no for sure that nothing new will be added
1440          * to the list
1441          */
1442         btrfs_run_ordered_operations(root, 1);
1443
1444         return 0;
1445 }
1446
1447 /*
1448  * btrfs_transaction state sequence:
1449  *    in_commit = 0, blocked = 0  (initial)
1450  *    in_commit = 1, blocked = 1
1451  *    blocked = 0
1452  *    commit_done = 1
1453  */
1454 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
1455                              struct btrfs_root *root)
1456 {
1457         unsigned long joined = 0;
1458         struct btrfs_transaction *cur_trans = trans->transaction;
1459         struct btrfs_transaction *prev_trans = NULL;
1460         DEFINE_WAIT(wait);
1461         int ret;
1462         int should_grow = 0;
1463         unsigned long now = get_seconds();
1464
1465         ret = btrfs_run_ordered_operations(root, 0);
1466         if (ret) {
1467                 btrfs_abort_transaction(trans, root, ret);
1468                 goto cleanup_transaction;
1469         }
1470
1471         if (cur_trans->aborted) {
1472                 ret = cur_trans->aborted;
1473                 goto cleanup_transaction;
1474         }
1475
1476         /* make a pass through all the delayed refs we have so far
1477          * any runnings procs may add more while we are here
1478          */
1479         ret = btrfs_run_delayed_refs(trans, root, 0);
1480         if (ret)
1481                 goto cleanup_transaction;
1482
1483         btrfs_trans_release_metadata(trans, root);
1484         trans->block_rsv = NULL;
1485
1486         cur_trans = trans->transaction;
1487
1488         /*
1489          * set the flushing flag so procs in this transaction have to
1490          * start sending their work down.
1491          */
1492         cur_trans->delayed_refs.flushing = 1;
1493
1494         if (!list_empty(&trans->new_bgs))
1495                 btrfs_create_pending_block_groups(trans, root);
1496
1497         ret = btrfs_run_delayed_refs(trans, root, 0);
1498         if (ret)
1499                 goto cleanup_transaction;
1500
1501         spin_lock(&cur_trans->commit_lock);
1502         if (cur_trans->in_commit) {
1503                 spin_unlock(&cur_trans->commit_lock);
1504                 atomic_inc(&cur_trans->use_count);
1505                 ret = btrfs_end_transaction(trans, root);
1506
1507                 wait_for_commit(root, cur_trans);
1508
1509                 put_transaction(cur_trans);
1510
1511                 return ret;
1512         }
1513
1514         trans->transaction->in_commit = 1;
1515         trans->transaction->blocked = 1;
1516         spin_unlock(&cur_trans->commit_lock);
1517         wake_up(&root->fs_info->transaction_blocked_wait);
1518
1519         spin_lock(&root->fs_info->trans_lock);
1520         if (cur_trans->list.prev != &root->fs_info->trans_list) {
1521                 prev_trans = list_entry(cur_trans->list.prev,
1522                                         struct btrfs_transaction, list);
1523                 if (!prev_trans->commit_done) {
1524                         atomic_inc(&prev_trans->use_count);
1525                         spin_unlock(&root->fs_info->trans_lock);
1526
1527                         wait_for_commit(root, prev_trans);
1528
1529                         put_transaction(prev_trans);
1530                 } else {
1531                         spin_unlock(&root->fs_info->trans_lock);
1532                 }
1533         } else {
1534                 spin_unlock(&root->fs_info->trans_lock);
1535         }
1536
1537         if (!btrfs_test_opt(root, SSD) &&
1538             (now < cur_trans->start_time || now - cur_trans->start_time < 1))
1539                 should_grow = 1;
1540
1541         do {
1542                 joined = cur_trans->num_joined;
1543
1544                 WARN_ON(cur_trans != trans->transaction);
1545
1546                 ret = btrfs_flush_all_pending_stuffs(trans, root);
1547                 if (ret)
1548                         goto cleanup_transaction;
1549
1550                 prepare_to_wait(&cur_trans->writer_wait, &wait,
1551                                 TASK_UNINTERRUPTIBLE);
1552
1553                 if (atomic_read(&cur_trans->num_writers) > 1)
1554                         schedule_timeout(MAX_SCHEDULE_TIMEOUT);
1555                 else if (should_grow)
1556                         schedule_timeout(1);
1557
1558                 finish_wait(&cur_trans->writer_wait, &wait);
1559         } while (atomic_read(&cur_trans->num_writers) > 1 ||
1560                  (should_grow && cur_trans->num_joined != joined));
1561
1562         ret = btrfs_flush_all_pending_stuffs(trans, root);
1563         if (ret)
1564                 goto cleanup_transaction;
1565
1566         /*
1567          * Ok now we need to make sure to block out any other joins while we
1568          * commit the transaction.  We could have started a join before setting
1569          * no_join so make sure to wait for num_writers to == 1 again.
1570          */
1571         spin_lock(&root->fs_info->trans_lock);
1572         root->fs_info->trans_no_join = 1;
1573         spin_unlock(&root->fs_info->trans_lock);
1574         wait_event(cur_trans->writer_wait,
1575                    atomic_read(&cur_trans->num_writers) == 1);
1576
1577         /*
1578          * the reloc mutex makes sure that we stop
1579          * the balancing code from coming in and moving
1580          * extents around in the middle of the commit
1581          */
1582         mutex_lock(&root->fs_info->reloc_mutex);
1583
1584         /*
1585          * We needn't worry about the delayed items because we will
1586          * deal with them in create_pending_snapshot(), which is the
1587          * core function of the snapshot creation.
1588          */
1589         ret = create_pending_snapshots(trans, root->fs_info);
1590         if (ret) {
1591                 mutex_unlock(&root->fs_info->reloc_mutex);
1592                 goto cleanup_transaction;
1593         }
1594
1595         /*
1596          * We insert the dir indexes of the snapshots and update the inode
1597          * of the snapshots' parents after the snapshot creation, so there
1598          * are some delayed items which are not dealt with. Now deal with
1599          * them.
1600          *
1601          * We needn't worry that this operation will corrupt the snapshots,
1602          * because all the tree which are snapshoted will be forced to COW
1603          * the nodes and leaves.
1604          */
1605         ret = btrfs_run_delayed_items(trans, root);
1606         if (ret) {
1607                 mutex_unlock(&root->fs_info->reloc_mutex);
1608                 goto cleanup_transaction;
1609         }
1610
1611         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1612         if (ret) {
1613                 mutex_unlock(&root->fs_info->reloc_mutex);
1614                 goto cleanup_transaction;
1615         }
1616
1617         /*
1618          * make sure none of the code above managed to slip in a
1619          * delayed item
1620          */
1621         btrfs_assert_delayed_root_empty(root);
1622
1623         WARN_ON(cur_trans != trans->transaction);
1624
1625         btrfs_scrub_pause(root);
1626         /* btrfs_commit_tree_roots is responsible for getting the
1627          * various roots consistent with each other.  Every pointer
1628          * in the tree of tree roots has to point to the most up to date
1629          * root for every subvolume and other tree.  So, we have to keep
1630          * the tree logging code from jumping in and changing any
1631          * of the trees.
1632          *
1633          * At this point in the commit, there can't be any tree-log
1634          * writers, but a little lower down we drop the trans mutex
1635          * and let new people in.  By holding the tree_log_mutex
1636          * from now until after the super is written, we avoid races
1637          * with the tree-log code.
1638          */
1639         mutex_lock(&root->fs_info->tree_log_mutex);
1640
1641         ret = commit_fs_roots(trans, root);
1642         if (ret) {
1643                 mutex_unlock(&root->fs_info->tree_log_mutex);
1644                 mutex_unlock(&root->fs_info->reloc_mutex);
1645                 goto cleanup_transaction;
1646         }
1647
1648         /* commit_fs_roots gets rid of all the tree log roots, it is now
1649          * safe to free the root of tree log roots
1650          */
1651         btrfs_free_log_root_tree(trans, root->fs_info);
1652
1653         ret = commit_cowonly_roots(trans, root);
1654         if (ret) {
1655                 mutex_unlock(&root->fs_info->tree_log_mutex);
1656                 mutex_unlock(&root->fs_info->reloc_mutex);
1657                 goto cleanup_transaction;
1658         }
1659
1660         btrfs_prepare_extent_commit(trans, root);
1661
1662         cur_trans = root->fs_info->running_transaction;
1663
1664         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
1665                             root->fs_info->tree_root->node);
1666         switch_commit_root(root->fs_info->tree_root);
1667
1668         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
1669                             root->fs_info->chunk_root->node);
1670         switch_commit_root(root->fs_info->chunk_root);
1671
1672         assert_qgroups_uptodate(trans);
1673         update_super_roots(root);
1674
1675         if (!root->fs_info->log_root_recovering) {
1676                 btrfs_set_super_log_root(root->fs_info->super_copy, 0);
1677                 btrfs_set_super_log_root_level(root->fs_info->super_copy, 0);
1678         }
1679
1680         memcpy(root->fs_info->super_for_commit, root->fs_info->super_copy,
1681                sizeof(*root->fs_info->super_copy));
1682
1683         trans->transaction->blocked = 0;
1684         spin_lock(&root->fs_info->trans_lock);
1685         root->fs_info->running_transaction = NULL;
1686         root->fs_info->trans_no_join = 0;
1687         spin_unlock(&root->fs_info->trans_lock);
1688         mutex_unlock(&root->fs_info->reloc_mutex);
1689
1690         wake_up(&root->fs_info->transaction_wait);
1691
1692         ret = btrfs_write_and_wait_transaction(trans, root);
1693         if (ret) {
1694                 btrfs_error(root->fs_info, ret,
1695                             "Error while writing out transaction.");
1696                 mutex_unlock(&root->fs_info->tree_log_mutex);
1697                 goto cleanup_transaction;
1698         }
1699
1700         ret = write_ctree_super(trans, root, 0);
1701         if (ret) {
1702                 mutex_unlock(&root->fs_info->tree_log_mutex);
1703                 goto cleanup_transaction;
1704         }
1705
1706         /*
1707          * the super is written, we can safely allow the tree-loggers
1708          * to go about their business
1709          */
1710         mutex_unlock(&root->fs_info->tree_log_mutex);
1711
1712         btrfs_finish_extent_commit(trans, root);
1713
1714         cur_trans->commit_done = 1;
1715
1716         root->fs_info->last_trans_committed = cur_trans->transid;
1717
1718         wake_up(&cur_trans->commit_wait);
1719
1720         spin_lock(&root->fs_info->trans_lock);
1721         list_del_init(&cur_trans->list);
1722         spin_unlock(&root->fs_info->trans_lock);
1723
1724         put_transaction(cur_trans);
1725         put_transaction(cur_trans);
1726
1727         if (trans->type < TRANS_JOIN_NOLOCK)
1728                 sb_end_intwrite(root->fs_info->sb);
1729
1730         trace_btrfs_transaction_commit(root);
1731
1732         btrfs_scrub_continue(root);
1733
1734         if (current->journal_info == trans)
1735                 current->journal_info = NULL;
1736
1737         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1738
1739         if (current != root->fs_info->transaction_kthread)
1740                 btrfs_run_delayed_iputs(root);
1741
1742         return ret;
1743
1744 cleanup_transaction:
1745         btrfs_trans_release_metadata(trans, root);
1746         trans->block_rsv = NULL;
1747         btrfs_printk(root->fs_info, "Skipping commit of aborted transaction.\n");
1748 //      WARN_ON(1);
1749         if (current->journal_info == trans)
1750                 current->journal_info = NULL;
1751         cleanup_transaction(trans, root, ret);
1752
1753         return ret;
1754 }
1755
1756 /*
1757  * interface function to delete all the snapshots we have scheduled for deletion
1758  */
1759 int btrfs_clean_old_snapshots(struct btrfs_root *root)
1760 {
1761         LIST_HEAD(list);
1762         struct btrfs_fs_info *fs_info = root->fs_info;
1763
1764         spin_lock(&fs_info->trans_lock);
1765         list_splice_init(&fs_info->dead_roots, &list);
1766         spin_unlock(&fs_info->trans_lock);
1767
1768         while (!list_empty(&list)) {
1769                 int ret;
1770
1771                 root = list_entry(list.next, struct btrfs_root, root_list);
1772                 list_del(&root->root_list);
1773
1774                 btrfs_kill_all_delayed_nodes(root);
1775
1776                 if (btrfs_header_backref_rev(root->node) <
1777                     BTRFS_MIXED_BACKREF_REV)
1778                         ret = btrfs_drop_snapshot(root, NULL, 0, 0);
1779                 else
1780                         ret =btrfs_drop_snapshot(root, NULL, 1, 0);
1781                 BUG_ON(ret < 0);
1782         }
1783         return 0;
1784 }