Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/mason/linux...
[~shefty/rdma-dev.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                 kmem_cache_free(btrfs_transaction_cachep, transaction);
44         }
45 }
46
47 static noinline void switch_commit_root(struct btrfs_root *root)
48 {
49         free_extent_buffer(root->commit_root);
50         root->commit_root = btrfs_root_node(root);
51 }
52
53 static inline int can_join_transaction(struct btrfs_transaction *trans,
54                                        int type)
55 {
56         return !(trans->in_commit &&
57                  type != TRANS_JOIN &&
58                  type != TRANS_JOIN_NOLOCK);
59 }
60
61 /*
62  * either allocate a new transaction or hop into the existing one
63  */
64 static noinline int join_transaction(struct btrfs_root *root, int type)
65 {
66         struct btrfs_transaction *cur_trans;
67         struct btrfs_fs_info *fs_info = root->fs_info;
68
69         spin_lock(&fs_info->trans_lock);
70 loop:
71         /* The file system has been taken offline. No new transactions. */
72         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
73                 spin_unlock(&fs_info->trans_lock);
74                 return -EROFS;
75         }
76
77         if (fs_info->trans_no_join) {
78                 /* 
79                  * If we are JOIN_NOLOCK we're already committing a current
80                  * transaction, we just need a handle to deal with something
81                  * when committing the transaction, such as inode cache and
82                  * space cache. It is a special case.
83                  */
84                 if (type != TRANS_JOIN_NOLOCK) {
85                         spin_unlock(&fs_info->trans_lock);
86                         return -EBUSY;
87                 }
88         }
89
90         cur_trans = fs_info->running_transaction;
91         if (cur_trans) {
92                 if (cur_trans->aborted) {
93                         spin_unlock(&fs_info->trans_lock);
94                         return cur_trans->aborted;
95                 }
96                 if (!can_join_transaction(cur_trans, type)) {
97                         spin_unlock(&fs_info->trans_lock);
98                         return -EBUSY;
99                 }
100                 atomic_inc(&cur_trans->use_count);
101                 atomic_inc(&cur_trans->num_writers);
102                 cur_trans->num_joined++;
103                 spin_unlock(&fs_info->trans_lock);
104                 return 0;
105         }
106         spin_unlock(&fs_info->trans_lock);
107
108         /*
109          * If we are ATTACH, we just want to catch the current transaction,
110          * and commit it. If there is no transaction, just return ENOENT.
111          */
112         if (type == TRANS_ATTACH)
113                 return -ENOENT;
114
115         cur_trans = kmem_cache_alloc(btrfs_transaction_cachep, GFP_NOFS);
116         if (!cur_trans)
117                 return -ENOMEM;
118
119         spin_lock(&fs_info->trans_lock);
120         if (fs_info->running_transaction) {
121                 /*
122                  * someone started a transaction after we unlocked.  Make sure
123                  * to redo the trans_no_join checks above
124                  */
125                 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
126                 goto loop;
127         } else if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
128                 spin_unlock(&fs_info->trans_lock);
129                 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
130                 return -EROFS;
131         }
132
133         atomic_set(&cur_trans->num_writers, 1);
134         cur_trans->num_joined = 0;
135         init_waitqueue_head(&cur_trans->writer_wait);
136         init_waitqueue_head(&cur_trans->commit_wait);
137         cur_trans->in_commit = 0;
138         cur_trans->blocked = 0;
139         /*
140          * One for this trans handle, one so it will live on until we
141          * commit the transaction.
142          */
143         atomic_set(&cur_trans->use_count, 2);
144         cur_trans->commit_done = 0;
145         cur_trans->start_time = get_seconds();
146
147         cur_trans->delayed_refs.root = RB_ROOT;
148         cur_trans->delayed_refs.num_entries = 0;
149         cur_trans->delayed_refs.num_heads_ready = 0;
150         cur_trans->delayed_refs.num_heads = 0;
151         cur_trans->delayed_refs.flushing = 0;
152         cur_trans->delayed_refs.run_delayed_start = 0;
153
154         /*
155          * although the tree mod log is per file system and not per transaction,
156          * the log must never go across transaction boundaries.
157          */
158         smp_mb();
159         if (!list_empty(&fs_info->tree_mod_seq_list))
160                 WARN(1, KERN_ERR "btrfs: tree_mod_seq_list not empty when "
161                         "creating a fresh transaction\n");
162         if (!RB_EMPTY_ROOT(&fs_info->tree_mod_log))
163                 WARN(1, KERN_ERR "btrfs: tree_mod_log rb tree not empty when "
164                         "creating a fresh transaction\n");
165         atomic_set(&fs_info->tree_mod_seq, 0);
166
167         spin_lock_init(&cur_trans->commit_lock);
168         spin_lock_init(&cur_trans->delayed_refs.lock);
169         atomic_set(&cur_trans->delayed_refs.procs_running_refs, 0);
170         atomic_set(&cur_trans->delayed_refs.ref_seq, 0);
171         init_waitqueue_head(&cur_trans->delayed_refs.wait);
172
173         INIT_LIST_HEAD(&cur_trans->pending_snapshots);
174         INIT_LIST_HEAD(&cur_trans->ordered_operations);
175         list_add_tail(&cur_trans->list, &fs_info->trans_list);
176         extent_io_tree_init(&cur_trans->dirty_pages,
177                              fs_info->btree_inode->i_mapping);
178         fs_info->generation++;
179         cur_trans->transid = fs_info->generation;
180         fs_info->running_transaction = cur_trans;
181         cur_trans->aborted = 0;
182         spin_unlock(&fs_info->trans_lock);
183
184         return 0;
185 }
186
187 /*
188  * this does all the record keeping required to make sure that a reference
189  * counted root is properly recorded in a given transaction.  This is required
190  * to make sure the old root from before we joined the transaction is deleted
191  * when the transaction commits
192  */
193 static int record_root_in_trans(struct btrfs_trans_handle *trans,
194                                struct btrfs_root *root)
195 {
196         if (root->ref_cows && root->last_trans < trans->transid) {
197                 WARN_ON(root == root->fs_info->extent_root);
198                 WARN_ON(root->commit_root != root->node);
199
200                 /*
201                  * see below for in_trans_setup usage rules
202                  * we have the reloc mutex held now, so there
203                  * is only one writer in this function
204                  */
205                 root->in_trans_setup = 1;
206
207                 /* make sure readers find in_trans_setup before
208                  * they find our root->last_trans update
209                  */
210                 smp_wmb();
211
212                 spin_lock(&root->fs_info->fs_roots_radix_lock);
213                 if (root->last_trans == trans->transid) {
214                         spin_unlock(&root->fs_info->fs_roots_radix_lock);
215                         return 0;
216                 }
217                 radix_tree_tag_set(&root->fs_info->fs_roots_radix,
218                            (unsigned long)root->root_key.objectid,
219                            BTRFS_ROOT_TRANS_TAG);
220                 spin_unlock(&root->fs_info->fs_roots_radix_lock);
221                 root->last_trans = trans->transid;
222
223                 /* this is pretty tricky.  We don't want to
224                  * take the relocation lock in btrfs_record_root_in_trans
225                  * unless we're really doing the first setup for this root in
226                  * this transaction.
227                  *
228                  * Normally we'd use root->last_trans as a flag to decide
229                  * if we want to take the expensive mutex.
230                  *
231                  * But, we have to set root->last_trans before we
232                  * init the relocation root, otherwise, we trip over warnings
233                  * in ctree.c.  The solution used here is to flag ourselves
234                  * with root->in_trans_setup.  When this is 1, we're still
235                  * fixing up the reloc trees and everyone must wait.
236                  *
237                  * When this is zero, they can trust root->last_trans and fly
238                  * through btrfs_record_root_in_trans without having to take the
239                  * lock.  smp_wmb() makes sure that all the writes above are
240                  * done before we pop in the zero below
241                  */
242                 btrfs_init_reloc_root(trans, root);
243                 smp_wmb();
244                 root->in_trans_setup = 0;
245         }
246         return 0;
247 }
248
249
250 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
251                                struct btrfs_root *root)
252 {
253         if (!root->ref_cows)
254                 return 0;
255
256         /*
257          * see record_root_in_trans for comments about in_trans_setup usage
258          * and barriers
259          */
260         smp_rmb();
261         if (root->last_trans == trans->transid &&
262             !root->in_trans_setup)
263                 return 0;
264
265         mutex_lock(&root->fs_info->reloc_mutex);
266         record_root_in_trans(trans, root);
267         mutex_unlock(&root->fs_info->reloc_mutex);
268
269         return 0;
270 }
271
272 /* wait for commit against the current transaction to become unblocked
273  * when this is done, it is safe to start a new transaction, but the current
274  * transaction might not be fully on disk.
275  */
276 static void wait_current_trans(struct btrfs_root *root)
277 {
278         struct btrfs_transaction *cur_trans;
279
280         spin_lock(&root->fs_info->trans_lock);
281         cur_trans = root->fs_info->running_transaction;
282         if (cur_trans && cur_trans->blocked) {
283                 atomic_inc(&cur_trans->use_count);
284                 spin_unlock(&root->fs_info->trans_lock);
285
286                 wait_event(root->fs_info->transaction_wait,
287                            !cur_trans->blocked);
288                 put_transaction(cur_trans);
289         } else {
290                 spin_unlock(&root->fs_info->trans_lock);
291         }
292 }
293
294 static int may_wait_transaction(struct btrfs_root *root, int type)
295 {
296         if (root->fs_info->log_root_recovering)
297                 return 0;
298
299         if (type == TRANS_USERSPACE)
300                 return 1;
301
302         if (type == TRANS_START &&
303             !atomic_read(&root->fs_info->open_ioctl_trans))
304                 return 1;
305
306         return 0;
307 }
308
309 static struct btrfs_trans_handle *
310 start_transaction(struct btrfs_root *root, u64 num_items, int type,
311                   enum btrfs_reserve_flush_enum flush)
312 {
313         struct btrfs_trans_handle *h;
314         struct btrfs_transaction *cur_trans;
315         u64 num_bytes = 0;
316         int ret;
317         u64 qgroup_reserved = 0;
318
319         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
320                 return ERR_PTR(-EROFS);
321
322         if (current->journal_info) {
323                 WARN_ON(type != TRANS_JOIN && type != TRANS_JOIN_NOLOCK);
324                 h = current->journal_info;
325                 h->use_count++;
326                 WARN_ON(h->use_count > 2);
327                 h->orig_rsv = h->block_rsv;
328                 h->block_rsv = NULL;
329                 goto got_it;
330         }
331
332         /*
333          * Do the reservation before we join the transaction so we can do all
334          * the appropriate flushing if need be.
335          */
336         if (num_items > 0 && root != root->fs_info->chunk_root) {
337                 if (root->fs_info->quota_enabled &&
338                     is_fstree(root->root_key.objectid)) {
339                         qgroup_reserved = num_items * root->leafsize;
340                         ret = btrfs_qgroup_reserve(root, qgroup_reserved);
341                         if (ret)
342                                 return ERR_PTR(ret);
343                 }
344
345                 num_bytes = btrfs_calc_trans_metadata_size(root, num_items);
346                 ret = btrfs_block_rsv_add(root,
347                                           &root->fs_info->trans_block_rsv,
348                                           num_bytes, flush);
349                 if (ret)
350                         goto reserve_fail;
351         }
352 again:
353         h = kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
354         if (!h) {
355                 ret = -ENOMEM;
356                 goto alloc_fail;
357         }
358
359         /*
360          * If we are JOIN_NOLOCK we're already committing a transaction and
361          * waiting on this guy, so we don't need to do the sb_start_intwrite
362          * because we're already holding a ref.  We need this because we could
363          * have raced in and did an fsync() on a file which can kick a commit
364          * and then we deadlock with somebody doing a freeze.
365          *
366          * If we are ATTACH, it means we just want to catch the current
367          * transaction and commit it, so we needn't do sb_start_intwrite(). 
368          */
369         if (type < TRANS_JOIN_NOLOCK)
370                 sb_start_intwrite(root->fs_info->sb);
371
372         if (may_wait_transaction(root, type))
373                 wait_current_trans(root);
374
375         do {
376                 ret = join_transaction(root, type);
377                 if (ret == -EBUSY) {
378                         wait_current_trans(root);
379                         if (unlikely(type == TRANS_ATTACH))
380                                 ret = -ENOENT;
381                 }
382         } while (ret == -EBUSY);
383
384         if (ret < 0) {
385                 /* We must get the transaction if we are JOIN_NOLOCK. */
386                 BUG_ON(type == TRANS_JOIN_NOLOCK);
387                 goto join_fail;
388         }
389
390         cur_trans = root->fs_info->running_transaction;
391
392         h->transid = cur_trans->transid;
393         h->transaction = cur_trans;
394         h->blocks_used = 0;
395         h->bytes_reserved = 0;
396         h->root = root;
397         h->delayed_ref_updates = 0;
398         h->use_count = 1;
399         h->adding_csums = 0;
400         h->block_rsv = NULL;
401         h->orig_rsv = NULL;
402         h->aborted = 0;
403         h->qgroup_reserved = 0;
404         h->delayed_ref_elem.seq = 0;
405         h->type = type;
406         h->allocating_chunk = false;
407         INIT_LIST_HEAD(&h->qgroup_ref_list);
408         INIT_LIST_HEAD(&h->new_bgs);
409
410         smp_mb();
411         if (cur_trans->blocked && may_wait_transaction(root, type)) {
412                 btrfs_commit_transaction(h, root);
413                 goto again;
414         }
415
416         if (num_bytes) {
417                 trace_btrfs_space_reservation(root->fs_info, "transaction",
418                                               h->transid, num_bytes, 1);
419                 h->block_rsv = &root->fs_info->trans_block_rsv;
420                 h->bytes_reserved = num_bytes;
421         }
422         h->qgroup_reserved = qgroup_reserved;
423
424 got_it:
425         btrfs_record_root_in_trans(h, root);
426
427         if (!current->journal_info && type != TRANS_USERSPACE)
428                 current->journal_info = h;
429         return h;
430
431 join_fail:
432         if (type < TRANS_JOIN_NOLOCK)
433                 sb_end_intwrite(root->fs_info->sb);
434         kmem_cache_free(btrfs_trans_handle_cachep, h);
435 alloc_fail:
436         if (num_bytes)
437                 btrfs_block_rsv_release(root, &root->fs_info->trans_block_rsv,
438                                         num_bytes);
439 reserve_fail:
440         if (qgroup_reserved)
441                 btrfs_qgroup_free(root, qgroup_reserved);
442         return ERR_PTR(ret);
443 }
444
445 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
446                                                    int num_items)
447 {
448         return start_transaction(root, num_items, TRANS_START,
449                                  BTRFS_RESERVE_FLUSH_ALL);
450 }
451
452 struct btrfs_trans_handle *btrfs_start_transaction_lflush(
453                                         struct btrfs_root *root, int num_items)
454 {
455         return start_transaction(root, num_items, TRANS_START,
456                                  BTRFS_RESERVE_FLUSH_LIMIT);
457 }
458
459 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
460 {
461         return start_transaction(root, 0, TRANS_JOIN, 0);
462 }
463
464 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root)
465 {
466         return start_transaction(root, 0, TRANS_JOIN_NOLOCK, 0);
467 }
468
469 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *root)
470 {
471         return start_transaction(root, 0, TRANS_USERSPACE, 0);
472 }
473
474 /*
475  * btrfs_attach_transaction() - catch the running transaction
476  *
477  * It is used when we want to commit the current the transaction, but
478  * don't want to start a new one.
479  *
480  * Note: If this function return -ENOENT, it just means there is no
481  * running transaction. But it is possible that the inactive transaction
482  * is still in the memory, not fully on disk. If you hope there is no
483  * inactive transaction in the fs when -ENOENT is returned, you should
484  * invoke
485  *     btrfs_attach_transaction_barrier()
486  */
487 struct btrfs_trans_handle *btrfs_attach_transaction(struct btrfs_root *root)
488 {
489         return start_transaction(root, 0, TRANS_ATTACH, 0);
490 }
491
492 /*
493  * btrfs_attach_transaction() - catch the running transaction
494  *
495  * It is similar to the above function, the differentia is this one
496  * will wait for all the inactive transactions until they fully
497  * complete.
498  */
499 struct btrfs_trans_handle *
500 btrfs_attach_transaction_barrier(struct btrfs_root *root)
501 {
502         struct btrfs_trans_handle *trans;
503
504         trans = start_transaction(root, 0, TRANS_ATTACH, 0);
505         if (IS_ERR(trans) && PTR_ERR(trans) == -ENOENT)
506                 btrfs_wait_for_commit(root, 0);
507
508         return trans;
509 }
510
511 /* wait for a transaction commit to be fully complete */
512 static noinline void wait_for_commit(struct btrfs_root *root,
513                                     struct btrfs_transaction *commit)
514 {
515         wait_event(commit->commit_wait, commit->commit_done);
516 }
517
518 int btrfs_wait_for_commit(struct btrfs_root *root, u64 transid)
519 {
520         struct btrfs_transaction *cur_trans = NULL, *t;
521         int ret = 0;
522
523         if (transid) {
524                 if (transid <= root->fs_info->last_trans_committed)
525                         goto out;
526
527                 ret = -EINVAL;
528                 /* find specified transaction */
529                 spin_lock(&root->fs_info->trans_lock);
530                 list_for_each_entry(t, &root->fs_info->trans_list, list) {
531                         if (t->transid == transid) {
532                                 cur_trans = t;
533                                 atomic_inc(&cur_trans->use_count);
534                                 ret = 0;
535                                 break;
536                         }
537                         if (t->transid > transid) {
538                                 ret = 0;
539                                 break;
540                         }
541                 }
542                 spin_unlock(&root->fs_info->trans_lock);
543                 /* The specified transaction doesn't exist */
544                 if (!cur_trans)
545                         goto out;
546         } else {
547                 /* find newest transaction that is committing | committed */
548                 spin_lock(&root->fs_info->trans_lock);
549                 list_for_each_entry_reverse(t, &root->fs_info->trans_list,
550                                             list) {
551                         if (t->in_commit) {
552                                 if (t->commit_done)
553                                         break;
554                                 cur_trans = t;
555                                 atomic_inc(&cur_trans->use_count);
556                                 break;
557                         }
558                 }
559                 spin_unlock(&root->fs_info->trans_lock);
560                 if (!cur_trans)
561                         goto out;  /* nothing committing|committed */
562         }
563
564         wait_for_commit(root, cur_trans);
565         put_transaction(cur_trans);
566 out:
567         return ret;
568 }
569
570 void btrfs_throttle(struct btrfs_root *root)
571 {
572         if (!atomic_read(&root->fs_info->open_ioctl_trans))
573                 wait_current_trans(root);
574 }
575
576 static int should_end_transaction(struct btrfs_trans_handle *trans,
577                                   struct btrfs_root *root)
578 {
579         int ret;
580
581         ret = btrfs_block_rsv_check(root, &root->fs_info->global_block_rsv, 5);
582         return ret ? 1 : 0;
583 }
584
585 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans,
586                                  struct btrfs_root *root)
587 {
588         struct btrfs_transaction *cur_trans = trans->transaction;
589         int updates;
590         int err;
591
592         smp_mb();
593         if (cur_trans->blocked || cur_trans->delayed_refs.flushing)
594                 return 1;
595
596         updates = trans->delayed_ref_updates;
597         trans->delayed_ref_updates = 0;
598         if (updates) {
599                 err = btrfs_run_delayed_refs(trans, root, updates);
600                 if (err) /* Error code will also eval true */
601                         return err;
602         }
603
604         return should_end_transaction(trans, root);
605 }
606
607 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
608                           struct btrfs_root *root, int throttle)
609 {
610         struct btrfs_transaction *cur_trans = trans->transaction;
611         struct btrfs_fs_info *info = root->fs_info;
612         int count = 0;
613         int lock = (trans->type != TRANS_JOIN_NOLOCK);
614         int err = 0;
615
616         if (--trans->use_count) {
617                 trans->block_rsv = trans->orig_rsv;
618                 return 0;
619         }
620
621         /*
622          * do the qgroup accounting as early as possible
623          */
624         err = btrfs_delayed_refs_qgroup_accounting(trans, info);
625
626         btrfs_trans_release_metadata(trans, root);
627         trans->block_rsv = NULL;
628
629         if (trans->qgroup_reserved) {
630                 /*
631                  * the same root has to be passed here between start_transaction
632                  * and end_transaction. Subvolume quota depends on this.
633                  */
634                 btrfs_qgroup_free(trans->root, trans->qgroup_reserved);
635                 trans->qgroup_reserved = 0;
636         }
637
638         if (!list_empty(&trans->new_bgs))
639                 btrfs_create_pending_block_groups(trans, root);
640
641         while (count < 1) {
642                 unsigned long cur = trans->delayed_ref_updates;
643                 trans->delayed_ref_updates = 0;
644                 if (cur &&
645                     trans->transaction->delayed_refs.num_heads_ready > 64) {
646                         trans->delayed_ref_updates = 0;
647                         btrfs_run_delayed_refs(trans, root, cur);
648                 } else {
649                         break;
650                 }
651                 count++;
652         }
653
654         btrfs_trans_release_metadata(trans, root);
655         trans->block_rsv = NULL;
656
657         if (!list_empty(&trans->new_bgs))
658                 btrfs_create_pending_block_groups(trans, root);
659
660         if (lock && !atomic_read(&root->fs_info->open_ioctl_trans) &&
661             should_end_transaction(trans, root)) {
662                 trans->transaction->blocked = 1;
663                 smp_wmb();
664         }
665
666         if (lock && cur_trans->blocked && !cur_trans->in_commit) {
667                 if (throttle) {
668                         /*
669                          * We may race with somebody else here so end up having
670                          * to call end_transaction on ourselves again, so inc
671                          * our use_count.
672                          */
673                         trans->use_count++;
674                         return btrfs_commit_transaction(trans, root);
675                 } else {
676                         wake_up_process(info->transaction_kthread);
677                 }
678         }
679
680         if (trans->type < TRANS_JOIN_NOLOCK)
681                 sb_end_intwrite(root->fs_info->sb);
682
683         WARN_ON(cur_trans != info->running_transaction);
684         WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
685         atomic_dec(&cur_trans->num_writers);
686
687         smp_mb();
688         if (waitqueue_active(&cur_trans->writer_wait))
689                 wake_up(&cur_trans->writer_wait);
690         put_transaction(cur_trans);
691
692         if (current->journal_info == trans)
693                 current->journal_info = NULL;
694
695         if (throttle)
696                 btrfs_run_delayed_iputs(root);
697
698         if (trans->aborted ||
699             test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
700                 err = -EIO;
701         assert_qgroups_uptodate(trans);
702
703         kmem_cache_free(btrfs_trans_handle_cachep, trans);
704         return err;
705 }
706
707 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
708                           struct btrfs_root *root)
709 {
710         int ret;
711
712         ret = __btrfs_end_transaction(trans, root, 0);
713         if (ret)
714                 return ret;
715         return 0;
716 }
717
718 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
719                                    struct btrfs_root *root)
720 {
721         int ret;
722
723         ret = __btrfs_end_transaction(trans, root, 1);
724         if (ret)
725                 return ret;
726         return 0;
727 }
728
729 int btrfs_end_transaction_dmeta(struct btrfs_trans_handle *trans,
730                                 struct btrfs_root *root)
731 {
732         return __btrfs_end_transaction(trans, root, 1);
733 }
734
735 /*
736  * when btree blocks are allocated, they have some corresponding bits set for
737  * them in one of two extent_io trees.  This is used to make sure all of
738  * those extents are sent to disk but does not wait on them
739  */
740 int btrfs_write_marked_extents(struct btrfs_root *root,
741                                struct extent_io_tree *dirty_pages, int mark)
742 {
743         int err = 0;
744         int werr = 0;
745         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
746         struct extent_state *cached_state = NULL;
747         u64 start = 0;
748         u64 end;
749         struct blk_plug plug;
750
751         blk_start_plug(&plug);
752         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
753                                       mark, &cached_state)) {
754                 convert_extent_bit(dirty_pages, start, end, EXTENT_NEED_WAIT,
755                                    mark, &cached_state, GFP_NOFS);
756                 cached_state = NULL;
757                 err = filemap_fdatawrite_range(mapping, start, end);
758                 if (err)
759                         werr = err;
760                 cond_resched();
761                 start = end + 1;
762         }
763         if (err)
764                 werr = err;
765         blk_finish_plug(&plug);
766         return werr;
767 }
768
769 /*
770  * when btree blocks are allocated, they have some corresponding bits set for
771  * them in one of two extent_io trees.  This is used to make sure all of
772  * those extents are on disk for transaction or log commit.  We wait
773  * on all the pages and clear them from the dirty pages state tree
774  */
775 int btrfs_wait_marked_extents(struct btrfs_root *root,
776                               struct extent_io_tree *dirty_pages, int mark)
777 {
778         int err = 0;
779         int werr = 0;
780         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
781         struct extent_state *cached_state = NULL;
782         u64 start = 0;
783         u64 end;
784
785         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
786                                       EXTENT_NEED_WAIT, &cached_state)) {
787                 clear_extent_bit(dirty_pages, start, end, EXTENT_NEED_WAIT,
788                                  0, 0, &cached_state, GFP_NOFS);
789                 err = filemap_fdatawait_range(mapping, start, end);
790                 if (err)
791                         werr = err;
792                 cond_resched();
793                 start = end + 1;
794         }
795         if (err)
796                 werr = err;
797         return werr;
798 }
799
800 /*
801  * when btree blocks are allocated, they have some corresponding bits set for
802  * them in one of two extent_io trees.  This is used to make sure all of
803  * those extents are on disk for transaction or log commit
804  */
805 int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
806                                 struct extent_io_tree *dirty_pages, int mark)
807 {
808         int ret;
809         int ret2;
810
811         ret = btrfs_write_marked_extents(root, dirty_pages, mark);
812         ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
813
814         if (ret)
815                 return ret;
816         if (ret2)
817                 return ret2;
818         return 0;
819 }
820
821 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
822                                      struct btrfs_root *root)
823 {
824         if (!trans || !trans->transaction) {
825                 struct inode *btree_inode;
826                 btree_inode = root->fs_info->btree_inode;
827                 return filemap_write_and_wait(btree_inode->i_mapping);
828         }
829         return btrfs_write_and_wait_marked_extents(root,
830                                            &trans->transaction->dirty_pages,
831                                            EXTENT_DIRTY);
832 }
833
834 /*
835  * this is used to update the root pointer in the tree of tree roots.
836  *
837  * But, in the case of the extent allocation tree, updating the root
838  * pointer may allocate blocks which may change the root of the extent
839  * allocation tree.
840  *
841  * So, this loops and repeats and makes sure the cowonly root didn't
842  * change while the root pointer was being updated in the metadata.
843  */
844 static int update_cowonly_root(struct btrfs_trans_handle *trans,
845                                struct btrfs_root *root)
846 {
847         int ret;
848         u64 old_root_bytenr;
849         u64 old_root_used;
850         struct btrfs_root *tree_root = root->fs_info->tree_root;
851
852         old_root_used = btrfs_root_used(&root->root_item);
853         btrfs_write_dirty_block_groups(trans, root);
854
855         while (1) {
856                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
857                 if (old_root_bytenr == root->node->start &&
858                     old_root_used == btrfs_root_used(&root->root_item))
859                         break;
860
861                 btrfs_set_root_node(&root->root_item, root->node);
862                 ret = btrfs_update_root(trans, tree_root,
863                                         &root->root_key,
864                                         &root->root_item);
865                 if (ret)
866                         return ret;
867
868                 old_root_used = btrfs_root_used(&root->root_item);
869                 ret = btrfs_write_dirty_block_groups(trans, root);
870                 if (ret)
871                         return ret;
872         }
873
874         if (root != root->fs_info->extent_root)
875                 switch_commit_root(root);
876
877         return 0;
878 }
879
880 /*
881  * update all the cowonly tree roots on disk
882  *
883  * The error handling in this function may not be obvious. Any of the
884  * failures will cause the file system to go offline. We still need
885  * to clean up the delayed refs.
886  */
887 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
888                                          struct btrfs_root *root)
889 {
890         struct btrfs_fs_info *fs_info = root->fs_info;
891         struct list_head *next;
892         struct extent_buffer *eb;
893         int ret;
894
895         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
896         if (ret)
897                 return ret;
898
899         eb = btrfs_lock_root_node(fs_info->tree_root);
900         ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL,
901                               0, &eb);
902         btrfs_tree_unlock(eb);
903         free_extent_buffer(eb);
904
905         if (ret)
906                 return ret;
907
908         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
909         if (ret)
910                 return ret;
911
912         ret = btrfs_run_dev_stats(trans, root->fs_info);
913         WARN_ON(ret);
914         ret = btrfs_run_dev_replace(trans, root->fs_info);
915         WARN_ON(ret);
916
917         ret = btrfs_run_qgroups(trans, root->fs_info);
918         BUG_ON(ret);
919
920         /* run_qgroups might have added some more refs */
921         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
922         BUG_ON(ret);
923
924         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
925                 next = fs_info->dirty_cowonly_roots.next;
926                 list_del_init(next);
927                 root = list_entry(next, struct btrfs_root, dirty_list);
928
929                 ret = update_cowonly_root(trans, root);
930                 if (ret)
931                         return ret;
932         }
933
934         down_write(&fs_info->extent_commit_sem);
935         switch_commit_root(fs_info->extent_root);
936         up_write(&fs_info->extent_commit_sem);
937
938         btrfs_after_dev_replace_commit(fs_info);
939
940         return 0;
941 }
942
943 /*
944  * dead roots are old snapshots that need to be deleted.  This allocates
945  * a dirty root struct and adds it into the list of dead roots that need to
946  * be deleted
947  */
948 int btrfs_add_dead_root(struct btrfs_root *root)
949 {
950         spin_lock(&root->fs_info->trans_lock);
951         list_add(&root->root_list, &root->fs_info->dead_roots);
952         spin_unlock(&root->fs_info->trans_lock);
953         return 0;
954 }
955
956 /*
957  * update all the cowonly tree roots on disk
958  */
959 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
960                                     struct btrfs_root *root)
961 {
962         struct btrfs_root *gang[8];
963         struct btrfs_fs_info *fs_info = root->fs_info;
964         int i;
965         int ret;
966         int err = 0;
967
968         spin_lock(&fs_info->fs_roots_radix_lock);
969         while (1) {
970                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
971                                                  (void **)gang, 0,
972                                                  ARRAY_SIZE(gang),
973                                                  BTRFS_ROOT_TRANS_TAG);
974                 if (ret == 0)
975                         break;
976                 for (i = 0; i < ret; i++) {
977                         root = gang[i];
978                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
979                                         (unsigned long)root->root_key.objectid,
980                                         BTRFS_ROOT_TRANS_TAG);
981                         spin_unlock(&fs_info->fs_roots_radix_lock);
982
983                         btrfs_free_log(trans, root);
984                         btrfs_update_reloc_root(trans, root);
985                         btrfs_orphan_commit_root(trans, root);
986
987                         btrfs_save_ino_cache(root, trans);
988
989                         /* see comments in should_cow_block() */
990                         root->force_cow = 0;
991                         smp_wmb();
992
993                         if (root->commit_root != root->node) {
994                                 mutex_lock(&root->fs_commit_mutex);
995                                 switch_commit_root(root);
996                                 btrfs_unpin_free_ino(root);
997                                 mutex_unlock(&root->fs_commit_mutex);
998
999                                 btrfs_set_root_node(&root->root_item,
1000                                                     root->node);
1001                         }
1002
1003                         err = btrfs_update_root(trans, fs_info->tree_root,
1004                                                 &root->root_key,
1005                                                 &root->root_item);
1006                         spin_lock(&fs_info->fs_roots_radix_lock);
1007                         if (err)
1008                                 break;
1009                 }
1010         }
1011         spin_unlock(&fs_info->fs_roots_radix_lock);
1012         return err;
1013 }
1014
1015 /*
1016  * defrag a given btree.
1017  * Every leaf in the btree is read and defragged.
1018  */
1019 int btrfs_defrag_root(struct btrfs_root *root)
1020 {
1021         struct btrfs_fs_info *info = root->fs_info;
1022         struct btrfs_trans_handle *trans;
1023         int ret;
1024
1025         if (xchg(&root->defrag_running, 1))
1026                 return 0;
1027
1028         while (1) {
1029                 trans = btrfs_start_transaction(root, 0);
1030                 if (IS_ERR(trans))
1031                         return PTR_ERR(trans);
1032
1033                 ret = btrfs_defrag_leaves(trans, root);
1034
1035                 btrfs_end_transaction(trans, root);
1036                 btrfs_btree_balance_dirty(info->tree_root);
1037                 cond_resched();
1038
1039                 if (btrfs_fs_closing(root->fs_info) || ret != -EAGAIN)
1040                         break;
1041
1042                 if (btrfs_defrag_cancelled(root->fs_info)) {
1043                         printk(KERN_DEBUG "btrfs: defrag_root cancelled\n");
1044                         ret = -EAGAIN;
1045                         break;
1046                 }
1047         }
1048         root->defrag_running = 0;
1049         return ret;
1050 }
1051
1052 /*
1053  * new snapshots need to be created at a very specific time in the
1054  * transaction commit.  This does the actual creation.
1055  *
1056  * Note:
1057  * If the error which may affect the commitment of the current transaction
1058  * happens, we should return the error number. If the error which just affect
1059  * the creation of the pending snapshots, just return 0.
1060  */
1061 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
1062                                    struct btrfs_fs_info *fs_info,
1063                                    struct btrfs_pending_snapshot *pending)
1064 {
1065         struct btrfs_key key;
1066         struct btrfs_root_item *new_root_item;
1067         struct btrfs_root *tree_root = fs_info->tree_root;
1068         struct btrfs_root *root = pending->root;
1069         struct btrfs_root *parent_root;
1070         struct btrfs_block_rsv *rsv;
1071         struct inode *parent_inode;
1072         struct btrfs_path *path;
1073         struct btrfs_dir_item *dir_item;
1074         struct dentry *dentry;
1075         struct extent_buffer *tmp;
1076         struct extent_buffer *old;
1077         struct timespec cur_time = CURRENT_TIME;
1078         int ret = 0;
1079         u64 to_reserve = 0;
1080         u64 index = 0;
1081         u64 objectid;
1082         u64 root_flags;
1083         uuid_le new_uuid;
1084
1085         path = btrfs_alloc_path();
1086         if (!path) {
1087                 pending->error = -ENOMEM;
1088                 return 0;
1089         }
1090
1091         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
1092         if (!new_root_item) {
1093                 pending->error = -ENOMEM;
1094                 goto root_item_alloc_fail;
1095         }
1096
1097         pending->error = btrfs_find_free_objectid(tree_root, &objectid);
1098         if (pending->error)
1099                 goto no_free_objectid;
1100
1101         btrfs_reloc_pre_snapshot(trans, pending, &to_reserve);
1102
1103         if (to_reserve > 0) {
1104                 pending->error = btrfs_block_rsv_add(root,
1105                                                      &pending->block_rsv,
1106                                                      to_reserve,
1107                                                      BTRFS_RESERVE_NO_FLUSH);
1108                 if (pending->error)
1109                         goto no_free_objectid;
1110         }
1111
1112         pending->error = btrfs_qgroup_inherit(trans, fs_info,
1113                                               root->root_key.objectid,
1114                                               objectid, pending->inherit);
1115         if (pending->error)
1116                 goto no_free_objectid;
1117
1118         key.objectid = objectid;
1119         key.offset = (u64)-1;
1120         key.type = BTRFS_ROOT_ITEM_KEY;
1121
1122         rsv = trans->block_rsv;
1123         trans->block_rsv = &pending->block_rsv;
1124         trans->bytes_reserved = trans->block_rsv->reserved;
1125
1126         dentry = pending->dentry;
1127         parent_inode = pending->dir;
1128         parent_root = BTRFS_I(parent_inode)->root;
1129         record_root_in_trans(trans, parent_root);
1130
1131         /*
1132          * insert the directory item
1133          */
1134         ret = btrfs_set_inode_index(parent_inode, &index);
1135         BUG_ON(ret); /* -ENOMEM */
1136
1137         /* check if there is a file/dir which has the same name. */
1138         dir_item = btrfs_lookup_dir_item(NULL, parent_root, path,
1139                                          btrfs_ino(parent_inode),
1140                                          dentry->d_name.name,
1141                                          dentry->d_name.len, 0);
1142         if (dir_item != NULL && !IS_ERR(dir_item)) {
1143                 pending->error = -EEXIST;
1144                 goto dir_item_existed;
1145         } else if (IS_ERR(dir_item)) {
1146                 ret = PTR_ERR(dir_item);
1147                 btrfs_abort_transaction(trans, root, ret);
1148                 goto fail;
1149         }
1150         btrfs_release_path(path);
1151
1152         /*
1153          * pull in the delayed directory update
1154          * and the delayed inode item
1155          * otherwise we corrupt the FS during
1156          * snapshot
1157          */
1158         ret = btrfs_run_delayed_items(trans, root);
1159         if (ret) {      /* Transaction aborted */
1160                 btrfs_abort_transaction(trans, root, ret);
1161                 goto fail;
1162         }
1163
1164         record_root_in_trans(trans, root);
1165         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
1166         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
1167         btrfs_check_and_init_root_item(new_root_item);
1168
1169         root_flags = btrfs_root_flags(new_root_item);
1170         if (pending->readonly)
1171                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
1172         else
1173                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
1174         btrfs_set_root_flags(new_root_item, root_flags);
1175
1176         btrfs_set_root_generation_v2(new_root_item,
1177                         trans->transid);
1178         uuid_le_gen(&new_uuid);
1179         memcpy(new_root_item->uuid, new_uuid.b, BTRFS_UUID_SIZE);
1180         memcpy(new_root_item->parent_uuid, root->root_item.uuid,
1181                         BTRFS_UUID_SIZE);
1182         new_root_item->otime.sec = cpu_to_le64(cur_time.tv_sec);
1183         new_root_item->otime.nsec = cpu_to_le32(cur_time.tv_nsec);
1184         btrfs_set_root_otransid(new_root_item, trans->transid);
1185         memset(&new_root_item->stime, 0, sizeof(new_root_item->stime));
1186         memset(&new_root_item->rtime, 0, sizeof(new_root_item->rtime));
1187         btrfs_set_root_stransid(new_root_item, 0);
1188         btrfs_set_root_rtransid(new_root_item, 0);
1189
1190         old = btrfs_lock_root_node(root);
1191         ret = btrfs_cow_block(trans, root, old, NULL, 0, &old);
1192         if (ret) {
1193                 btrfs_tree_unlock(old);
1194                 free_extent_buffer(old);
1195                 btrfs_abort_transaction(trans, root, ret);
1196                 goto fail;
1197         }
1198
1199         btrfs_set_lock_blocking(old);
1200
1201         ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1202         /* clean up in any case */
1203         btrfs_tree_unlock(old);
1204         free_extent_buffer(old);
1205         if (ret) {
1206                 btrfs_abort_transaction(trans, root, ret);
1207                 goto fail;
1208         }
1209
1210         /* see comments in should_cow_block() */
1211         root->force_cow = 1;
1212         smp_wmb();
1213
1214         btrfs_set_root_node(new_root_item, tmp);
1215         /* record when the snapshot was created in key.offset */
1216         key.offset = trans->transid;
1217         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1218         btrfs_tree_unlock(tmp);
1219         free_extent_buffer(tmp);
1220         if (ret) {
1221                 btrfs_abort_transaction(trans, root, ret);
1222                 goto fail;
1223         }
1224
1225         /*
1226          * insert root back/forward references
1227          */
1228         ret = btrfs_add_root_ref(trans, tree_root, objectid,
1229                                  parent_root->root_key.objectid,
1230                                  btrfs_ino(parent_inode), index,
1231                                  dentry->d_name.name, dentry->d_name.len);
1232         if (ret) {
1233                 btrfs_abort_transaction(trans, root, ret);
1234                 goto fail;
1235         }
1236
1237         key.offset = (u64)-1;
1238         pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
1239         if (IS_ERR(pending->snap)) {
1240                 ret = PTR_ERR(pending->snap);
1241                 btrfs_abort_transaction(trans, root, ret);
1242                 goto fail;
1243         }
1244
1245         ret = btrfs_reloc_post_snapshot(trans, pending);
1246         if (ret) {
1247                 btrfs_abort_transaction(trans, root, ret);
1248                 goto fail;
1249         }
1250
1251         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1252         if (ret) {
1253                 btrfs_abort_transaction(trans, root, ret);
1254                 goto fail;
1255         }
1256
1257         ret = btrfs_insert_dir_item(trans, parent_root,
1258                                     dentry->d_name.name, dentry->d_name.len,
1259                                     parent_inode, &key,
1260                                     BTRFS_FT_DIR, index);
1261         /* We have check then name at the beginning, so it is impossible. */
1262         BUG_ON(ret == -EEXIST || ret == -EOVERFLOW);
1263         if (ret) {
1264                 btrfs_abort_transaction(trans, root, ret);
1265                 goto fail;
1266         }
1267
1268         btrfs_i_size_write(parent_inode, parent_inode->i_size +
1269                                          dentry->d_name.len * 2);
1270         parent_inode->i_mtime = parent_inode->i_ctime = CURRENT_TIME;
1271         ret = btrfs_update_inode_fallback(trans, parent_root, parent_inode);
1272         if (ret)
1273                 btrfs_abort_transaction(trans, root, ret);
1274 fail:
1275         pending->error = ret;
1276 dir_item_existed:
1277         trans->block_rsv = rsv;
1278         trans->bytes_reserved = 0;
1279 no_free_objectid:
1280         kfree(new_root_item);
1281 root_item_alloc_fail:
1282         btrfs_free_path(path);
1283         return ret;
1284 }
1285
1286 /*
1287  * create all the snapshots we've scheduled for creation
1288  */
1289 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
1290                                              struct btrfs_fs_info *fs_info)
1291 {
1292         struct btrfs_pending_snapshot *pending, *next;
1293         struct list_head *head = &trans->transaction->pending_snapshots;
1294         int ret = 0;
1295
1296         list_for_each_entry_safe(pending, next, head, list) {
1297                 list_del(&pending->list);
1298                 ret = create_pending_snapshot(trans, fs_info, pending);
1299                 if (ret)
1300                         break;
1301         }
1302         return ret;
1303 }
1304
1305 static void update_super_roots(struct btrfs_root *root)
1306 {
1307         struct btrfs_root_item *root_item;
1308         struct btrfs_super_block *super;
1309
1310         super = root->fs_info->super_copy;
1311
1312         root_item = &root->fs_info->chunk_root->root_item;
1313         super->chunk_root = root_item->bytenr;
1314         super->chunk_root_generation = root_item->generation;
1315         super->chunk_root_level = root_item->level;
1316
1317         root_item = &root->fs_info->tree_root->root_item;
1318         super->root = root_item->bytenr;
1319         super->generation = root_item->generation;
1320         super->root_level = root_item->level;
1321         if (btrfs_test_opt(root, SPACE_CACHE))
1322                 super->cache_generation = root_item->generation;
1323 }
1324
1325 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1326 {
1327         int ret = 0;
1328         spin_lock(&info->trans_lock);
1329         if (info->running_transaction)
1330                 ret = info->running_transaction->in_commit;
1331         spin_unlock(&info->trans_lock);
1332         return ret;
1333 }
1334
1335 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1336 {
1337         int ret = 0;
1338         spin_lock(&info->trans_lock);
1339         if (info->running_transaction)
1340                 ret = info->running_transaction->blocked;
1341         spin_unlock(&info->trans_lock);
1342         return ret;
1343 }
1344
1345 /*
1346  * wait for the current transaction commit to start and block subsequent
1347  * transaction joins
1348  */
1349 static void wait_current_trans_commit_start(struct btrfs_root *root,
1350                                             struct btrfs_transaction *trans)
1351 {
1352         wait_event(root->fs_info->transaction_blocked_wait, trans->in_commit);
1353 }
1354
1355 /*
1356  * wait for the current transaction to start and then become unblocked.
1357  * caller holds ref.
1358  */
1359 static void wait_current_trans_commit_start_and_unblock(struct btrfs_root *root,
1360                                          struct btrfs_transaction *trans)
1361 {
1362         wait_event(root->fs_info->transaction_wait,
1363                    trans->commit_done || (trans->in_commit && !trans->blocked));
1364 }
1365
1366 /*
1367  * commit transactions asynchronously. once btrfs_commit_transaction_async
1368  * returns, any subsequent transaction will not be allowed to join.
1369  */
1370 struct btrfs_async_commit {
1371         struct btrfs_trans_handle *newtrans;
1372         struct btrfs_root *root;
1373         struct work_struct work;
1374 };
1375
1376 static void do_async_commit(struct work_struct *work)
1377 {
1378         struct btrfs_async_commit *ac =
1379                 container_of(work, struct btrfs_async_commit, work);
1380
1381         /*
1382          * We've got freeze protection passed with the transaction.
1383          * Tell lockdep about it.
1384          */
1385         if (ac->newtrans->type < TRANS_JOIN_NOLOCK)
1386                 rwsem_acquire_read(
1387                      &ac->root->fs_info->sb->s_writers.lock_map[SB_FREEZE_FS-1],
1388                      0, 1, _THIS_IP_);
1389
1390         current->journal_info = ac->newtrans;
1391
1392         btrfs_commit_transaction(ac->newtrans, ac->root);
1393         kfree(ac);
1394 }
1395
1396 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1397                                    struct btrfs_root *root,
1398                                    int wait_for_unblock)
1399 {
1400         struct btrfs_async_commit *ac;
1401         struct btrfs_transaction *cur_trans;
1402
1403         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1404         if (!ac)
1405                 return -ENOMEM;
1406
1407         INIT_WORK(&ac->work, do_async_commit);
1408         ac->root = root;
1409         ac->newtrans = btrfs_join_transaction(root);
1410         if (IS_ERR(ac->newtrans)) {
1411                 int err = PTR_ERR(ac->newtrans);
1412                 kfree(ac);
1413                 return err;
1414         }
1415
1416         /* take transaction reference */
1417         cur_trans = trans->transaction;
1418         atomic_inc(&cur_trans->use_count);
1419
1420         btrfs_end_transaction(trans, root);
1421
1422         /*
1423          * Tell lockdep we've released the freeze rwsem, since the
1424          * async commit thread will be the one to unlock it.
1425          */
1426         if (trans->type < TRANS_JOIN_NOLOCK)
1427                 rwsem_release(
1428                         &root->fs_info->sb->s_writers.lock_map[SB_FREEZE_FS-1],
1429                         1, _THIS_IP_);
1430
1431         schedule_work(&ac->work);
1432
1433         /* wait for transaction to start and unblock */
1434         if (wait_for_unblock)
1435                 wait_current_trans_commit_start_and_unblock(root, cur_trans);
1436         else
1437                 wait_current_trans_commit_start(root, cur_trans);
1438
1439         if (current->journal_info == trans)
1440                 current->journal_info = NULL;
1441
1442         put_transaction(cur_trans);
1443         return 0;
1444 }
1445
1446
1447 static void cleanup_transaction(struct btrfs_trans_handle *trans,
1448                                 struct btrfs_root *root, int err)
1449 {
1450         struct btrfs_transaction *cur_trans = trans->transaction;
1451         DEFINE_WAIT(wait);
1452
1453         WARN_ON(trans->use_count > 1);
1454
1455         btrfs_abort_transaction(trans, root, err);
1456
1457         spin_lock(&root->fs_info->trans_lock);
1458
1459         if (list_empty(&cur_trans->list)) {
1460                 spin_unlock(&root->fs_info->trans_lock);
1461                 btrfs_end_transaction(trans, root);
1462                 return;
1463         }
1464
1465         list_del_init(&cur_trans->list);
1466         if (cur_trans == root->fs_info->running_transaction) {
1467                 root->fs_info->trans_no_join = 1;
1468                 spin_unlock(&root->fs_info->trans_lock);
1469                 wait_event(cur_trans->writer_wait,
1470                            atomic_read(&cur_trans->num_writers) == 1);
1471
1472                 spin_lock(&root->fs_info->trans_lock);
1473                 root->fs_info->running_transaction = NULL;
1474         }
1475         spin_unlock(&root->fs_info->trans_lock);
1476
1477         btrfs_cleanup_one_transaction(trans->transaction, root);
1478
1479         put_transaction(cur_trans);
1480         put_transaction(cur_trans);
1481
1482         trace_btrfs_transaction_commit(root);
1483
1484         btrfs_scrub_continue(root);
1485
1486         if (current->journal_info == trans)
1487                 current->journal_info = NULL;
1488
1489         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1490 }
1491
1492 static int btrfs_flush_all_pending_stuffs(struct btrfs_trans_handle *trans,
1493                                           struct btrfs_root *root)
1494 {
1495         int flush_on_commit = btrfs_test_opt(root, FLUSHONCOMMIT);
1496         int snap_pending = 0;
1497         int ret;
1498
1499         if (!flush_on_commit) {
1500                 spin_lock(&root->fs_info->trans_lock);
1501                 if (!list_empty(&trans->transaction->pending_snapshots))
1502                         snap_pending = 1;
1503                 spin_unlock(&root->fs_info->trans_lock);
1504         }
1505
1506         if (flush_on_commit || snap_pending) {
1507                 ret = btrfs_start_delalloc_inodes(root, 1);
1508                 if (ret)
1509                         return ret;
1510                 btrfs_wait_ordered_extents(root, 1);
1511         }
1512
1513         ret = btrfs_run_delayed_items(trans, root);
1514         if (ret)
1515                 return ret;
1516
1517         /*
1518          * running the delayed items may have added new refs. account
1519          * them now so that they hinder processing of more delayed refs
1520          * as little as possible.
1521          */
1522         btrfs_delayed_refs_qgroup_accounting(trans, root->fs_info);
1523
1524         /*
1525          * rename don't use btrfs_join_transaction, so, once we
1526          * set the transaction to blocked above, we aren't going
1527          * to get any new ordered operations.  We can safely run
1528          * it here and no for sure that nothing new will be added
1529          * to the list
1530          */
1531         ret = btrfs_run_ordered_operations(trans, root, 1);
1532
1533         return ret;
1534 }
1535
1536 /*
1537  * btrfs_transaction state sequence:
1538  *    in_commit = 0, blocked = 0  (initial)
1539  *    in_commit = 1, blocked = 1
1540  *    blocked = 0
1541  *    commit_done = 1
1542  */
1543 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
1544                              struct btrfs_root *root)
1545 {
1546         unsigned long joined = 0;
1547         struct btrfs_transaction *cur_trans = trans->transaction;
1548         struct btrfs_transaction *prev_trans = NULL;
1549         DEFINE_WAIT(wait);
1550         int ret;
1551         int should_grow = 0;
1552         unsigned long now = get_seconds();
1553
1554         ret = btrfs_run_ordered_operations(trans, root, 0);
1555         if (ret) {
1556                 btrfs_abort_transaction(trans, root, ret);
1557                 btrfs_end_transaction(trans, root);
1558                 return ret;
1559         }
1560
1561         /* Stop the commit early if ->aborted is set */
1562         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
1563                 ret = cur_trans->aborted;
1564                 btrfs_end_transaction(trans, root);
1565                 return ret;
1566         }
1567
1568         /* make a pass through all the delayed refs we have so far
1569          * any runnings procs may add more while we are here
1570          */
1571         ret = btrfs_run_delayed_refs(trans, root, 0);
1572         if (ret) {
1573                 btrfs_end_transaction(trans, root);
1574                 return ret;
1575         }
1576
1577         btrfs_trans_release_metadata(trans, root);
1578         trans->block_rsv = NULL;
1579         if (trans->qgroup_reserved) {
1580                 btrfs_qgroup_free(root, trans->qgroup_reserved);
1581                 trans->qgroup_reserved = 0;
1582         }
1583
1584         cur_trans = trans->transaction;
1585
1586         /*
1587          * set the flushing flag so procs in this transaction have to
1588          * start sending their work down.
1589          */
1590         cur_trans->delayed_refs.flushing = 1;
1591
1592         if (!list_empty(&trans->new_bgs))
1593                 btrfs_create_pending_block_groups(trans, root);
1594
1595         ret = btrfs_run_delayed_refs(trans, root, 0);
1596         if (ret) {
1597                 btrfs_end_transaction(trans, root);
1598                 return ret;
1599         }
1600
1601         spin_lock(&cur_trans->commit_lock);
1602         if (cur_trans->in_commit) {
1603                 spin_unlock(&cur_trans->commit_lock);
1604                 atomic_inc(&cur_trans->use_count);
1605                 ret = btrfs_end_transaction(trans, root);
1606
1607                 wait_for_commit(root, cur_trans);
1608
1609                 put_transaction(cur_trans);
1610
1611                 return ret;
1612         }
1613
1614         trans->transaction->in_commit = 1;
1615         trans->transaction->blocked = 1;
1616         spin_unlock(&cur_trans->commit_lock);
1617         wake_up(&root->fs_info->transaction_blocked_wait);
1618
1619         spin_lock(&root->fs_info->trans_lock);
1620         if (cur_trans->list.prev != &root->fs_info->trans_list) {
1621                 prev_trans = list_entry(cur_trans->list.prev,
1622                                         struct btrfs_transaction, list);
1623                 if (!prev_trans->commit_done) {
1624                         atomic_inc(&prev_trans->use_count);
1625                         spin_unlock(&root->fs_info->trans_lock);
1626
1627                         wait_for_commit(root, prev_trans);
1628
1629                         put_transaction(prev_trans);
1630                 } else {
1631                         spin_unlock(&root->fs_info->trans_lock);
1632                 }
1633         } else {
1634                 spin_unlock(&root->fs_info->trans_lock);
1635         }
1636
1637         if (!btrfs_test_opt(root, SSD) &&
1638             (now < cur_trans->start_time || now - cur_trans->start_time < 1))
1639                 should_grow = 1;
1640
1641         do {
1642                 joined = cur_trans->num_joined;
1643
1644                 WARN_ON(cur_trans != trans->transaction);
1645
1646                 ret = btrfs_flush_all_pending_stuffs(trans, root);
1647                 if (ret)
1648                         goto cleanup_transaction;
1649
1650                 prepare_to_wait(&cur_trans->writer_wait, &wait,
1651                                 TASK_UNINTERRUPTIBLE);
1652
1653                 if (atomic_read(&cur_trans->num_writers) > 1)
1654                         schedule_timeout(MAX_SCHEDULE_TIMEOUT);
1655                 else if (should_grow)
1656                         schedule_timeout(1);
1657
1658                 finish_wait(&cur_trans->writer_wait, &wait);
1659         } while (atomic_read(&cur_trans->num_writers) > 1 ||
1660                  (should_grow && cur_trans->num_joined != joined));
1661
1662         ret = btrfs_flush_all_pending_stuffs(trans, root);
1663         if (ret)
1664                 goto cleanup_transaction;
1665
1666         /*
1667          * Ok now we need to make sure to block out any other joins while we
1668          * commit the transaction.  We could have started a join before setting
1669          * no_join so make sure to wait for num_writers to == 1 again.
1670          */
1671         spin_lock(&root->fs_info->trans_lock);
1672         root->fs_info->trans_no_join = 1;
1673         spin_unlock(&root->fs_info->trans_lock);
1674         wait_event(cur_trans->writer_wait,
1675                    atomic_read(&cur_trans->num_writers) == 1);
1676
1677         /* ->aborted might be set after the previous check, so check it */
1678         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
1679                 ret = cur_trans->aborted;
1680                 goto cleanup_transaction;
1681         }
1682         /*
1683          * the reloc mutex makes sure that we stop
1684          * the balancing code from coming in and moving
1685          * extents around in the middle of the commit
1686          */
1687         mutex_lock(&root->fs_info->reloc_mutex);
1688
1689         /*
1690          * We needn't worry about the delayed items because we will
1691          * deal with them in create_pending_snapshot(), which is the
1692          * core function of the snapshot creation.
1693          */
1694         ret = create_pending_snapshots(trans, root->fs_info);
1695         if (ret) {
1696                 mutex_unlock(&root->fs_info->reloc_mutex);
1697                 goto cleanup_transaction;
1698         }
1699
1700         /*
1701          * We insert the dir indexes of the snapshots and update the inode
1702          * of the snapshots' parents after the snapshot creation, so there
1703          * are some delayed items which are not dealt with. Now deal with
1704          * them.
1705          *
1706          * We needn't worry that this operation will corrupt the snapshots,
1707          * because all the tree which are snapshoted will be forced to COW
1708          * the nodes and leaves.
1709          */
1710         ret = btrfs_run_delayed_items(trans, root);
1711         if (ret) {
1712                 mutex_unlock(&root->fs_info->reloc_mutex);
1713                 goto cleanup_transaction;
1714         }
1715
1716         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1717         if (ret) {
1718                 mutex_unlock(&root->fs_info->reloc_mutex);
1719                 goto cleanup_transaction;
1720         }
1721
1722         /*
1723          * make sure none of the code above managed to slip in a
1724          * delayed item
1725          */
1726         btrfs_assert_delayed_root_empty(root);
1727
1728         WARN_ON(cur_trans != trans->transaction);
1729
1730         btrfs_scrub_pause(root);
1731         /* btrfs_commit_tree_roots is responsible for getting the
1732          * various roots consistent with each other.  Every pointer
1733          * in the tree of tree roots has to point to the most up to date
1734          * root for every subvolume and other tree.  So, we have to keep
1735          * the tree logging code from jumping in and changing any
1736          * of the trees.
1737          *
1738          * At this point in the commit, there can't be any tree-log
1739          * writers, but a little lower down we drop the trans mutex
1740          * and let new people in.  By holding the tree_log_mutex
1741          * from now until after the super is written, we avoid races
1742          * with the tree-log code.
1743          */
1744         mutex_lock(&root->fs_info->tree_log_mutex);
1745
1746         ret = commit_fs_roots(trans, root);
1747         if (ret) {
1748                 mutex_unlock(&root->fs_info->tree_log_mutex);
1749                 mutex_unlock(&root->fs_info->reloc_mutex);
1750                 goto cleanup_transaction;
1751         }
1752
1753         /* commit_fs_roots gets rid of all the tree log roots, it is now
1754          * safe to free the root of tree log roots
1755          */
1756         btrfs_free_log_root_tree(trans, root->fs_info);
1757
1758         ret = commit_cowonly_roots(trans, root);
1759         if (ret) {
1760                 mutex_unlock(&root->fs_info->tree_log_mutex);
1761                 mutex_unlock(&root->fs_info->reloc_mutex);
1762                 goto cleanup_transaction;
1763         }
1764
1765         /*
1766          * The tasks which save the space cache and inode cache may also
1767          * update ->aborted, check it.
1768          */
1769         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
1770                 ret = cur_trans->aborted;
1771                 mutex_unlock(&root->fs_info->tree_log_mutex);
1772                 mutex_unlock(&root->fs_info->reloc_mutex);
1773                 goto cleanup_transaction;
1774         }
1775
1776         btrfs_prepare_extent_commit(trans, root);
1777
1778         cur_trans = root->fs_info->running_transaction;
1779
1780         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
1781                             root->fs_info->tree_root->node);
1782         switch_commit_root(root->fs_info->tree_root);
1783
1784         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
1785                             root->fs_info->chunk_root->node);
1786         switch_commit_root(root->fs_info->chunk_root);
1787
1788         assert_qgroups_uptodate(trans);
1789         update_super_roots(root);
1790
1791         if (!root->fs_info->log_root_recovering) {
1792                 btrfs_set_super_log_root(root->fs_info->super_copy, 0);
1793                 btrfs_set_super_log_root_level(root->fs_info->super_copy, 0);
1794         }
1795
1796         memcpy(root->fs_info->super_for_commit, root->fs_info->super_copy,
1797                sizeof(*root->fs_info->super_copy));
1798
1799         trans->transaction->blocked = 0;
1800         spin_lock(&root->fs_info->trans_lock);
1801         root->fs_info->running_transaction = NULL;
1802         root->fs_info->trans_no_join = 0;
1803         spin_unlock(&root->fs_info->trans_lock);
1804         mutex_unlock(&root->fs_info->reloc_mutex);
1805
1806         wake_up(&root->fs_info->transaction_wait);
1807
1808         ret = btrfs_write_and_wait_transaction(trans, root);
1809         if (ret) {
1810                 btrfs_error(root->fs_info, ret,
1811                             "Error while writing out transaction.");
1812                 mutex_unlock(&root->fs_info->tree_log_mutex);
1813                 goto cleanup_transaction;
1814         }
1815
1816         ret = write_ctree_super(trans, root, 0);
1817         if (ret) {
1818                 mutex_unlock(&root->fs_info->tree_log_mutex);
1819                 goto cleanup_transaction;
1820         }
1821
1822         /*
1823          * the super is written, we can safely allow the tree-loggers
1824          * to go about their business
1825          */
1826         mutex_unlock(&root->fs_info->tree_log_mutex);
1827
1828         btrfs_finish_extent_commit(trans, root);
1829
1830         cur_trans->commit_done = 1;
1831
1832         root->fs_info->last_trans_committed = cur_trans->transid;
1833
1834         wake_up(&cur_trans->commit_wait);
1835
1836         spin_lock(&root->fs_info->trans_lock);
1837         list_del_init(&cur_trans->list);
1838         spin_unlock(&root->fs_info->trans_lock);
1839
1840         put_transaction(cur_trans);
1841         put_transaction(cur_trans);
1842
1843         if (trans->type < TRANS_JOIN_NOLOCK)
1844                 sb_end_intwrite(root->fs_info->sb);
1845
1846         trace_btrfs_transaction_commit(root);
1847
1848         btrfs_scrub_continue(root);
1849
1850         if (current->journal_info == trans)
1851                 current->journal_info = NULL;
1852
1853         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1854
1855         if (current != root->fs_info->transaction_kthread)
1856                 btrfs_run_delayed_iputs(root);
1857
1858         return ret;
1859
1860 cleanup_transaction:
1861         btrfs_trans_release_metadata(trans, root);
1862         trans->block_rsv = NULL;
1863         if (trans->qgroup_reserved) {
1864                 btrfs_qgroup_free(root, trans->qgroup_reserved);
1865                 trans->qgroup_reserved = 0;
1866         }
1867         btrfs_printk(root->fs_info, "Skipping commit of aborted transaction.\n");
1868 //      WARN_ON(1);
1869         if (current->journal_info == trans)
1870                 current->journal_info = NULL;
1871         cleanup_transaction(trans, root, ret);
1872
1873         return ret;
1874 }
1875
1876 /*
1877  * interface function to delete all the snapshots we have scheduled for deletion
1878  */
1879 int btrfs_clean_old_snapshots(struct btrfs_root *root)
1880 {
1881         LIST_HEAD(list);
1882         struct btrfs_fs_info *fs_info = root->fs_info;
1883
1884         spin_lock(&fs_info->trans_lock);
1885         list_splice_init(&fs_info->dead_roots, &list);
1886         spin_unlock(&fs_info->trans_lock);
1887
1888         while (!list_empty(&list)) {
1889                 int ret;
1890
1891                 root = list_entry(list.next, struct btrfs_root, root_list);
1892                 list_del(&root->root_list);
1893
1894                 btrfs_kill_all_delayed_nodes(root);
1895
1896                 if (btrfs_header_backref_rev(root->node) <
1897                     BTRFS_MIXED_BACKREF_REV)
1898                         ret = btrfs_drop_snapshot(root, NULL, 0, 0);
1899                 else
1900                         ret =btrfs_drop_snapshot(root, NULL, 1, 0);
1901                 BUG_ON(ret < 0);
1902         }
1903         return 0;
1904 }