]> bbs.cooldavid.org Git - net-next-2.6.git/blob - fs/btrfs/transaction.c
Btrfs: async transaction commit
[net-next-2.6.git] / fs / btrfs / transaction.c
1 /*
2  * Copyright (C) 2007 Oracle.  All rights reserved.
3  *
4  * This program is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU General Public
6  * License v2 as published by the Free Software Foundation.
7  *
8  * This program is distributed in the hope that it will be useful,
9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
11  * General Public License for more details.
12  *
13  * You should have received a copy of the GNU General Public
14  * License along with this program; if not, write to the
15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16  * Boston, MA 021110-1307, USA.
17  */
18
19 #include <linux/fs.h>
20 #include <linux/slab.h>
21 #include <linux/sched.h>
22 #include <linux/writeback.h>
23 #include <linux/pagemap.h>
24 #include <linux/blkdev.h>
25 #include "ctree.h"
26 #include "disk-io.h"
27 #include "transaction.h"
28 #include "locking.h"
29 #include "tree-log.h"
30
31 #define BTRFS_ROOT_TRANS_TAG 0
32
33 static noinline void put_transaction(struct btrfs_transaction *transaction)
34 {
35         WARN_ON(transaction->use_count == 0);
36         transaction->use_count--;
37         if (transaction->use_count == 0) {
38                 list_del_init(&transaction->list);
39                 memset(transaction, 0, sizeof(*transaction));
40                 kmem_cache_free(btrfs_transaction_cachep, transaction);
41         }
42 }
43
44 static noinline void switch_commit_root(struct btrfs_root *root)
45 {
46         free_extent_buffer(root->commit_root);
47         root->commit_root = btrfs_root_node(root);
48 }
49
50 /*
51  * either allocate a new transaction or hop into the existing one
52  */
53 static noinline int join_transaction(struct btrfs_root *root)
54 {
55         struct btrfs_transaction *cur_trans;
56         cur_trans = root->fs_info->running_transaction;
57         if (!cur_trans) {
58                 cur_trans = kmem_cache_alloc(btrfs_transaction_cachep,
59                                              GFP_NOFS);
60                 BUG_ON(!cur_trans);
61                 root->fs_info->generation++;
62                 cur_trans->num_writers = 1;
63                 cur_trans->num_joined = 0;
64                 cur_trans->transid = root->fs_info->generation;
65                 init_waitqueue_head(&cur_trans->writer_wait);
66                 init_waitqueue_head(&cur_trans->commit_wait);
67                 cur_trans->in_commit = 0;
68                 cur_trans->blocked = 0;
69                 cur_trans->use_count = 1;
70                 cur_trans->commit_done = 0;
71                 cur_trans->start_time = get_seconds();
72
73                 cur_trans->delayed_refs.root = RB_ROOT;
74                 cur_trans->delayed_refs.num_entries = 0;
75                 cur_trans->delayed_refs.num_heads_ready = 0;
76                 cur_trans->delayed_refs.num_heads = 0;
77                 cur_trans->delayed_refs.flushing = 0;
78                 cur_trans->delayed_refs.run_delayed_start = 0;
79                 spin_lock_init(&cur_trans->delayed_refs.lock);
80
81                 INIT_LIST_HEAD(&cur_trans->pending_snapshots);
82                 list_add_tail(&cur_trans->list, &root->fs_info->trans_list);
83                 extent_io_tree_init(&cur_trans->dirty_pages,
84                                      root->fs_info->btree_inode->i_mapping,
85                                      GFP_NOFS);
86                 spin_lock(&root->fs_info->new_trans_lock);
87                 root->fs_info->running_transaction = cur_trans;
88                 spin_unlock(&root->fs_info->new_trans_lock);
89         } else {
90                 cur_trans->num_writers++;
91                 cur_trans->num_joined++;
92         }
93
94         return 0;
95 }
96
97 /*
98  * this does all the record keeping required to make sure that a reference
99  * counted root is properly recorded in a given transaction.  This is required
100  * to make sure the old root from before we joined the transaction is deleted
101  * when the transaction commits
102  */
103 static noinline int record_root_in_trans(struct btrfs_trans_handle *trans,
104                                          struct btrfs_root *root)
105 {
106         if (root->ref_cows && root->last_trans < trans->transid) {
107                 WARN_ON(root == root->fs_info->extent_root);
108                 WARN_ON(root->commit_root != root->node);
109
110                 radix_tree_tag_set(&root->fs_info->fs_roots_radix,
111                            (unsigned long)root->root_key.objectid,
112                            BTRFS_ROOT_TRANS_TAG);
113                 root->last_trans = trans->transid;
114                 btrfs_init_reloc_root(trans, root);
115         }
116         return 0;
117 }
118
119 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
120                                struct btrfs_root *root)
121 {
122         if (!root->ref_cows)
123                 return 0;
124
125         mutex_lock(&root->fs_info->trans_mutex);
126         if (root->last_trans == trans->transid) {
127                 mutex_unlock(&root->fs_info->trans_mutex);
128                 return 0;
129         }
130
131         record_root_in_trans(trans, root);
132         mutex_unlock(&root->fs_info->trans_mutex);
133         return 0;
134 }
135
136 /* wait for commit against the current transaction to become unblocked
137  * when this is done, it is safe to start a new transaction, but the current
138  * transaction might not be fully on disk.
139  */
140 static void wait_current_trans(struct btrfs_root *root)
141 {
142         struct btrfs_transaction *cur_trans;
143
144         cur_trans = root->fs_info->running_transaction;
145         if (cur_trans && cur_trans->blocked) {
146                 DEFINE_WAIT(wait);
147                 cur_trans->use_count++;
148                 while (1) {
149                         prepare_to_wait(&root->fs_info->transaction_wait, &wait,
150                                         TASK_UNINTERRUPTIBLE);
151                         if (!cur_trans->blocked)
152                                 break;
153                         mutex_unlock(&root->fs_info->trans_mutex);
154                         schedule();
155                         mutex_lock(&root->fs_info->trans_mutex);
156                 }
157                 finish_wait(&root->fs_info->transaction_wait, &wait);
158                 put_transaction(cur_trans);
159         }
160 }
161
162 enum btrfs_trans_type {
163         TRANS_START,
164         TRANS_JOIN,
165         TRANS_USERSPACE,
166         TRANS_JOIN_NOLOCK,
167 };
168
169 static int may_wait_transaction(struct btrfs_root *root, int type)
170 {
171         if (!root->fs_info->log_root_recovering &&
172             ((type == TRANS_START && !root->fs_info->open_ioctl_trans) ||
173              type == TRANS_USERSPACE))
174                 return 1;
175         return 0;
176 }
177
178 static struct btrfs_trans_handle *start_transaction(struct btrfs_root *root,
179                                                     u64 num_items, int type)
180 {
181         struct btrfs_trans_handle *h;
182         struct btrfs_transaction *cur_trans;
183         int ret;
184 again:
185         h = kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
186         if (!h)
187                 return ERR_PTR(-ENOMEM);
188
189         if (type != TRANS_JOIN_NOLOCK)
190                 mutex_lock(&root->fs_info->trans_mutex);
191         if (may_wait_transaction(root, type))
192                 wait_current_trans(root);
193
194         ret = join_transaction(root);
195         BUG_ON(ret);
196
197         cur_trans = root->fs_info->running_transaction;
198         cur_trans->use_count++;
199         if (type != TRANS_JOIN_NOLOCK)
200                 mutex_unlock(&root->fs_info->trans_mutex);
201
202         h->transid = cur_trans->transid;
203         h->transaction = cur_trans;
204         h->blocks_used = 0;
205         h->block_group = 0;
206         h->bytes_reserved = 0;
207         h->delayed_ref_updates = 0;
208         h->block_rsv = NULL;
209
210         smp_mb();
211         if (cur_trans->blocked && may_wait_transaction(root, type)) {
212                 btrfs_commit_transaction(h, root);
213                 goto again;
214         }
215
216         if (num_items > 0) {
217                 ret = btrfs_trans_reserve_metadata(h, root, num_items);
218                 if (ret == -EAGAIN) {
219                         btrfs_commit_transaction(h, root);
220                         goto again;
221                 }
222                 if (ret < 0) {
223                         btrfs_end_transaction(h, root);
224                         return ERR_PTR(ret);
225                 }
226         }
227
228         if (type != TRANS_JOIN_NOLOCK)
229                 mutex_lock(&root->fs_info->trans_mutex);
230         record_root_in_trans(h, root);
231         if (type != TRANS_JOIN_NOLOCK)
232                 mutex_unlock(&root->fs_info->trans_mutex);
233
234         if (!current->journal_info && type != TRANS_USERSPACE)
235                 current->journal_info = h;
236         return h;
237 }
238
239 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
240                                                    int num_items)
241 {
242         return start_transaction(root, num_items, TRANS_START);
243 }
244 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root,
245                                                    int num_blocks)
246 {
247         return start_transaction(root, 0, TRANS_JOIN);
248 }
249
250 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root,
251                                                           int num_blocks)
252 {
253         return start_transaction(root, 0, TRANS_JOIN_NOLOCK);
254 }
255
256 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *r,
257                                                          int num_blocks)
258 {
259         return start_transaction(r, 0, TRANS_USERSPACE);
260 }
261
262 /* wait for a transaction commit to be fully complete */
263 static noinline int wait_for_commit(struct btrfs_root *root,
264                                     struct btrfs_transaction *commit)
265 {
266         DEFINE_WAIT(wait);
267         mutex_lock(&root->fs_info->trans_mutex);
268         while (!commit->commit_done) {
269                 prepare_to_wait(&commit->commit_wait, &wait,
270                                 TASK_UNINTERRUPTIBLE);
271                 if (commit->commit_done)
272                         break;
273                 mutex_unlock(&root->fs_info->trans_mutex);
274                 schedule();
275                 mutex_lock(&root->fs_info->trans_mutex);
276         }
277         mutex_unlock(&root->fs_info->trans_mutex);
278         finish_wait(&commit->commit_wait, &wait);
279         return 0;
280 }
281
282 #if 0
283 /*
284  * rate limit against the drop_snapshot code.  This helps to slow down new
285  * operations if the drop_snapshot code isn't able to keep up.
286  */
287 static void throttle_on_drops(struct btrfs_root *root)
288 {
289         struct btrfs_fs_info *info = root->fs_info;
290         int harder_count = 0;
291
292 harder:
293         if (atomic_read(&info->throttles)) {
294                 DEFINE_WAIT(wait);
295                 int thr;
296                 thr = atomic_read(&info->throttle_gen);
297
298                 do {
299                         prepare_to_wait(&info->transaction_throttle,
300                                         &wait, TASK_UNINTERRUPTIBLE);
301                         if (!atomic_read(&info->throttles)) {
302                                 finish_wait(&info->transaction_throttle, &wait);
303                                 break;
304                         }
305                         schedule();
306                         finish_wait(&info->transaction_throttle, &wait);
307                 } while (thr == atomic_read(&info->throttle_gen));
308                 harder_count++;
309
310                 if (root->fs_info->total_ref_cache_size > 1 * 1024 * 1024 &&
311                     harder_count < 2)
312                         goto harder;
313
314                 if (root->fs_info->total_ref_cache_size > 5 * 1024 * 1024 &&
315                     harder_count < 10)
316                         goto harder;
317
318                 if (root->fs_info->total_ref_cache_size > 10 * 1024 * 1024 &&
319                     harder_count < 20)
320                         goto harder;
321         }
322 }
323 #endif
324
325 void btrfs_throttle(struct btrfs_root *root)
326 {
327         mutex_lock(&root->fs_info->trans_mutex);
328         if (!root->fs_info->open_ioctl_trans)
329                 wait_current_trans(root);
330         mutex_unlock(&root->fs_info->trans_mutex);
331 }
332
333 static int should_end_transaction(struct btrfs_trans_handle *trans,
334                                   struct btrfs_root *root)
335 {
336         int ret;
337         ret = btrfs_block_rsv_check(trans, root,
338                                     &root->fs_info->global_block_rsv, 0, 5);
339         return ret ? 1 : 0;
340 }
341
342 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans,
343                                  struct btrfs_root *root)
344 {
345         struct btrfs_transaction *cur_trans = trans->transaction;
346         int updates;
347
348         if (cur_trans->blocked || cur_trans->delayed_refs.flushing)
349                 return 1;
350
351         updates = trans->delayed_ref_updates;
352         trans->delayed_ref_updates = 0;
353         if (updates)
354                 btrfs_run_delayed_refs(trans, root, updates);
355
356         return should_end_transaction(trans, root);
357 }
358
359 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
360                           struct btrfs_root *root, int throttle, int lock)
361 {
362         struct btrfs_transaction *cur_trans = trans->transaction;
363         struct btrfs_fs_info *info = root->fs_info;
364         int count = 0;
365
366         while (count < 4) {
367                 unsigned long cur = trans->delayed_ref_updates;
368                 trans->delayed_ref_updates = 0;
369                 if (cur &&
370                     trans->transaction->delayed_refs.num_heads_ready > 64) {
371                         trans->delayed_ref_updates = 0;
372
373                         /*
374                          * do a full flush if the transaction is trying
375                          * to close
376                          */
377                         if (trans->transaction->delayed_refs.flushing)
378                                 cur = 0;
379                         btrfs_run_delayed_refs(trans, root, cur);
380                 } else {
381                         break;
382                 }
383                 count++;
384         }
385
386         btrfs_trans_release_metadata(trans, root);
387
388         if (lock && !root->fs_info->open_ioctl_trans &&
389             should_end_transaction(trans, root))
390                 trans->transaction->blocked = 1;
391
392         if (lock && cur_trans->blocked && !cur_trans->in_commit) {
393                 if (throttle)
394                         return btrfs_commit_transaction(trans, root);
395                 else
396                         wake_up_process(info->transaction_kthread);
397         }
398
399         if (lock)
400                 mutex_lock(&info->trans_mutex);
401         WARN_ON(cur_trans != info->running_transaction);
402         WARN_ON(cur_trans->num_writers < 1);
403         cur_trans->num_writers--;
404
405         smp_mb();
406         if (waitqueue_active(&cur_trans->writer_wait))
407                 wake_up(&cur_trans->writer_wait);
408         put_transaction(cur_trans);
409         if (lock)
410                 mutex_unlock(&info->trans_mutex);
411
412         if (current->journal_info == trans)
413                 current->journal_info = NULL;
414         memset(trans, 0, sizeof(*trans));
415         kmem_cache_free(btrfs_trans_handle_cachep, trans);
416
417         if (throttle)
418                 btrfs_run_delayed_iputs(root);
419
420         return 0;
421 }
422
423 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
424                           struct btrfs_root *root)
425 {
426         return __btrfs_end_transaction(trans, root, 0, 1);
427 }
428
429 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
430                                    struct btrfs_root *root)
431 {
432         return __btrfs_end_transaction(trans, root, 1, 1);
433 }
434
435 int btrfs_end_transaction_nolock(struct btrfs_trans_handle *trans,
436                                  struct btrfs_root *root)
437 {
438         return __btrfs_end_transaction(trans, root, 0, 0);
439 }
440
441 /*
442  * when btree blocks are allocated, they have some corresponding bits set for
443  * them in one of two extent_io trees.  This is used to make sure all of
444  * those extents are sent to disk but does not wait on them
445  */
446 int btrfs_write_marked_extents(struct btrfs_root *root,
447                                struct extent_io_tree *dirty_pages, int mark)
448 {
449         int ret;
450         int err = 0;
451         int werr = 0;
452         struct page *page;
453         struct inode *btree_inode = root->fs_info->btree_inode;
454         u64 start = 0;
455         u64 end;
456         unsigned long index;
457
458         while (1) {
459                 ret = find_first_extent_bit(dirty_pages, start, &start, &end,
460                                             mark);
461                 if (ret)
462                         break;
463                 while (start <= end) {
464                         cond_resched();
465
466                         index = start >> PAGE_CACHE_SHIFT;
467                         start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
468                         page = find_get_page(btree_inode->i_mapping, index);
469                         if (!page)
470                                 continue;
471
472                         btree_lock_page_hook(page);
473                         if (!page->mapping) {
474                                 unlock_page(page);
475                                 page_cache_release(page);
476                                 continue;
477                         }
478
479                         if (PageWriteback(page)) {
480                                 if (PageDirty(page))
481                                         wait_on_page_writeback(page);
482                                 else {
483                                         unlock_page(page);
484                                         page_cache_release(page);
485                                         continue;
486                                 }
487                         }
488                         err = write_one_page(page, 0);
489                         if (err)
490                                 werr = err;
491                         page_cache_release(page);
492                 }
493         }
494         if (err)
495                 werr = err;
496         return werr;
497 }
498
499 /*
500  * when btree blocks are allocated, they have some corresponding bits set for
501  * them in one of two extent_io trees.  This is used to make sure all of
502  * those extents are on disk for transaction or log commit.  We wait
503  * on all the pages and clear them from the dirty pages state tree
504  */
505 int btrfs_wait_marked_extents(struct btrfs_root *root,
506                               struct extent_io_tree *dirty_pages, int mark)
507 {
508         int ret;
509         int err = 0;
510         int werr = 0;
511         struct page *page;
512         struct inode *btree_inode = root->fs_info->btree_inode;
513         u64 start = 0;
514         u64 end;
515         unsigned long index;
516
517         while (1) {
518                 ret = find_first_extent_bit(dirty_pages, start, &start, &end,
519                                             mark);
520                 if (ret)
521                         break;
522
523                 clear_extent_bits(dirty_pages, start, end, mark, GFP_NOFS);
524                 while (start <= end) {
525                         index = start >> PAGE_CACHE_SHIFT;
526                         start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
527                         page = find_get_page(btree_inode->i_mapping, index);
528                         if (!page)
529                                 continue;
530                         if (PageDirty(page)) {
531                                 btree_lock_page_hook(page);
532                                 wait_on_page_writeback(page);
533                                 err = write_one_page(page, 0);
534                                 if (err)
535                                         werr = err;
536                         }
537                         wait_on_page_writeback(page);
538                         page_cache_release(page);
539                         cond_resched();
540                 }
541         }
542         if (err)
543                 werr = err;
544         return werr;
545 }
546
547 /*
548  * when btree blocks are allocated, they have some corresponding bits set for
549  * them in one of two extent_io trees.  This is used to make sure all of
550  * those extents are on disk for transaction or log commit
551  */
552 int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
553                                 struct extent_io_tree *dirty_pages, int mark)
554 {
555         int ret;
556         int ret2;
557
558         ret = btrfs_write_marked_extents(root, dirty_pages, mark);
559         ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
560         return ret || ret2;
561 }
562
563 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
564                                      struct btrfs_root *root)
565 {
566         if (!trans || !trans->transaction) {
567                 struct inode *btree_inode;
568                 btree_inode = root->fs_info->btree_inode;
569                 return filemap_write_and_wait(btree_inode->i_mapping);
570         }
571         return btrfs_write_and_wait_marked_extents(root,
572                                            &trans->transaction->dirty_pages,
573                                            EXTENT_DIRTY);
574 }
575
576 /*
577  * this is used to update the root pointer in the tree of tree roots.
578  *
579  * But, in the case of the extent allocation tree, updating the root
580  * pointer may allocate blocks which may change the root of the extent
581  * allocation tree.
582  *
583  * So, this loops and repeats and makes sure the cowonly root didn't
584  * change while the root pointer was being updated in the metadata.
585  */
586 static int update_cowonly_root(struct btrfs_trans_handle *trans,
587                                struct btrfs_root *root)
588 {
589         int ret;
590         u64 old_root_bytenr;
591         u64 old_root_used;
592         struct btrfs_root *tree_root = root->fs_info->tree_root;
593
594         old_root_used = btrfs_root_used(&root->root_item);
595         btrfs_write_dirty_block_groups(trans, root);
596
597         while (1) {
598                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
599                 if (old_root_bytenr == root->node->start &&
600                     old_root_used == btrfs_root_used(&root->root_item))
601                         break;
602
603                 btrfs_set_root_node(&root->root_item, root->node);
604                 ret = btrfs_update_root(trans, tree_root,
605                                         &root->root_key,
606                                         &root->root_item);
607                 BUG_ON(ret);
608
609                 old_root_used = btrfs_root_used(&root->root_item);
610                 ret = btrfs_write_dirty_block_groups(trans, root);
611                 BUG_ON(ret);
612         }
613
614         if (root != root->fs_info->extent_root)
615                 switch_commit_root(root);
616
617         return 0;
618 }
619
620 /*
621  * update all the cowonly tree roots on disk
622  */
623 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
624                                          struct btrfs_root *root)
625 {
626         struct btrfs_fs_info *fs_info = root->fs_info;
627         struct list_head *next;
628         struct extent_buffer *eb;
629         int ret;
630
631         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
632         BUG_ON(ret);
633
634         eb = btrfs_lock_root_node(fs_info->tree_root);
635         btrfs_cow_block(trans, fs_info->tree_root, eb, NULL, 0, &eb);
636         btrfs_tree_unlock(eb);
637         free_extent_buffer(eb);
638
639         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
640         BUG_ON(ret);
641
642         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
643                 next = fs_info->dirty_cowonly_roots.next;
644                 list_del_init(next);
645                 root = list_entry(next, struct btrfs_root, dirty_list);
646
647                 update_cowonly_root(trans, root);
648         }
649
650         down_write(&fs_info->extent_commit_sem);
651         switch_commit_root(fs_info->extent_root);
652         up_write(&fs_info->extent_commit_sem);
653
654         return 0;
655 }
656
657 /*
658  * dead roots are old snapshots that need to be deleted.  This allocates
659  * a dirty root struct and adds it into the list of dead roots that need to
660  * be deleted
661  */
662 int btrfs_add_dead_root(struct btrfs_root *root)
663 {
664         mutex_lock(&root->fs_info->trans_mutex);
665         list_add(&root->root_list, &root->fs_info->dead_roots);
666         mutex_unlock(&root->fs_info->trans_mutex);
667         return 0;
668 }
669
670 /*
671  * update all the cowonly tree roots on disk
672  */
673 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
674                                     struct btrfs_root *root)
675 {
676         struct btrfs_root *gang[8];
677         struct btrfs_fs_info *fs_info = root->fs_info;
678         int i;
679         int ret;
680         int err = 0;
681
682         while (1) {
683                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
684                                                  (void **)gang, 0,
685                                                  ARRAY_SIZE(gang),
686                                                  BTRFS_ROOT_TRANS_TAG);
687                 if (ret == 0)
688                         break;
689                 for (i = 0; i < ret; i++) {
690                         root = gang[i];
691                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
692                                         (unsigned long)root->root_key.objectid,
693                                         BTRFS_ROOT_TRANS_TAG);
694
695                         btrfs_free_log(trans, root);
696                         btrfs_update_reloc_root(trans, root);
697                         btrfs_orphan_commit_root(trans, root);
698
699                         if (root->commit_root != root->node) {
700                                 switch_commit_root(root);
701                                 btrfs_set_root_node(&root->root_item,
702                                                     root->node);
703                         }
704
705                         err = btrfs_update_root(trans, fs_info->tree_root,
706                                                 &root->root_key,
707                                                 &root->root_item);
708                         if (err)
709                                 break;
710                 }
711         }
712         return err;
713 }
714
715 /*
716  * defrag a given btree.  If cacheonly == 1, this won't read from the disk,
717  * otherwise every leaf in the btree is read and defragged.
718  */
719 int btrfs_defrag_root(struct btrfs_root *root, int cacheonly)
720 {
721         struct btrfs_fs_info *info = root->fs_info;
722         struct btrfs_trans_handle *trans;
723         int ret;
724         unsigned long nr;
725
726         if (xchg(&root->defrag_running, 1))
727                 return 0;
728
729         while (1) {
730                 trans = btrfs_start_transaction(root, 0);
731                 if (IS_ERR(trans))
732                         return PTR_ERR(trans);
733
734                 ret = btrfs_defrag_leaves(trans, root, cacheonly);
735
736                 nr = trans->blocks_used;
737                 btrfs_end_transaction(trans, root);
738                 btrfs_btree_balance_dirty(info->tree_root, nr);
739                 cond_resched();
740
741                 if (root->fs_info->closing || ret != -EAGAIN)
742                         break;
743         }
744         root->defrag_running = 0;
745         return ret;
746 }
747
748 #if 0
749 /*
750  * when dropping snapshots, we generate a ton of delayed refs, and it makes
751  * sense not to join the transaction while it is trying to flush the current
752  * queue of delayed refs out.
753  *
754  * This is used by the drop snapshot code only
755  */
756 static noinline int wait_transaction_pre_flush(struct btrfs_fs_info *info)
757 {
758         DEFINE_WAIT(wait);
759
760         mutex_lock(&info->trans_mutex);
761         while (info->running_transaction &&
762                info->running_transaction->delayed_refs.flushing) {
763                 prepare_to_wait(&info->transaction_wait, &wait,
764                                 TASK_UNINTERRUPTIBLE);
765                 mutex_unlock(&info->trans_mutex);
766
767                 schedule();
768
769                 mutex_lock(&info->trans_mutex);
770                 finish_wait(&info->transaction_wait, &wait);
771         }
772         mutex_unlock(&info->trans_mutex);
773         return 0;
774 }
775
776 /*
777  * Given a list of roots that need to be deleted, call btrfs_drop_snapshot on
778  * all of them
779  */
780 int btrfs_drop_dead_root(struct btrfs_root *root)
781 {
782         struct btrfs_trans_handle *trans;
783         struct btrfs_root *tree_root = root->fs_info->tree_root;
784         unsigned long nr;
785         int ret;
786
787         while (1) {
788                 /*
789                  * we don't want to jump in and create a bunch of
790                  * delayed refs if the transaction is starting to close
791                  */
792                 wait_transaction_pre_flush(tree_root->fs_info);
793                 trans = btrfs_start_transaction(tree_root, 1);
794
795                 /*
796                  * we've joined a transaction, make sure it isn't
797                  * closing right now
798                  */
799                 if (trans->transaction->delayed_refs.flushing) {
800                         btrfs_end_transaction(trans, tree_root);
801                         continue;
802                 }
803
804                 ret = btrfs_drop_snapshot(trans, root);
805                 if (ret != -EAGAIN)
806                         break;
807
808                 ret = btrfs_update_root(trans, tree_root,
809                                         &root->root_key,
810                                         &root->root_item);
811                 if (ret)
812                         break;
813
814                 nr = trans->blocks_used;
815                 ret = btrfs_end_transaction(trans, tree_root);
816                 BUG_ON(ret);
817
818                 btrfs_btree_balance_dirty(tree_root, nr);
819                 cond_resched();
820         }
821         BUG_ON(ret);
822
823         ret = btrfs_del_root(trans, tree_root, &root->root_key);
824         BUG_ON(ret);
825
826         nr = trans->blocks_used;
827         ret = btrfs_end_transaction(trans, tree_root);
828         BUG_ON(ret);
829
830         free_extent_buffer(root->node);
831         free_extent_buffer(root->commit_root);
832         kfree(root);
833
834         btrfs_btree_balance_dirty(tree_root, nr);
835         return ret;
836 }
837 #endif
838
839 /*
840  * new snapshots need to be created at a very specific time in the
841  * transaction commit.  This does the actual creation
842  */
843 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
844                                    struct btrfs_fs_info *fs_info,
845                                    struct btrfs_pending_snapshot *pending)
846 {
847         struct btrfs_key key;
848         struct btrfs_root_item *new_root_item;
849         struct btrfs_root *tree_root = fs_info->tree_root;
850         struct btrfs_root *root = pending->root;
851         struct btrfs_root *parent_root;
852         struct inode *parent_inode;
853         struct dentry *dentry;
854         struct extent_buffer *tmp;
855         struct extent_buffer *old;
856         int ret;
857         u64 to_reserve = 0;
858         u64 index = 0;
859         u64 objectid;
860
861         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
862         if (!new_root_item) {
863                 pending->error = -ENOMEM;
864                 goto fail;
865         }
866
867         ret = btrfs_find_free_objectid(trans, tree_root, 0, &objectid);
868         if (ret) {
869                 pending->error = ret;
870                 goto fail;
871         }
872
873         btrfs_reloc_pre_snapshot(trans, pending, &to_reserve);
874         btrfs_orphan_pre_snapshot(trans, pending, &to_reserve);
875
876         if (to_reserve > 0) {
877                 ret = btrfs_block_rsv_add(trans, root, &pending->block_rsv,
878                                           to_reserve);
879                 if (ret) {
880                         pending->error = ret;
881                         goto fail;
882                 }
883         }
884
885         key.objectid = objectid;
886         key.offset = (u64)-1;
887         key.type = BTRFS_ROOT_ITEM_KEY;
888
889         trans->block_rsv = &pending->block_rsv;
890
891         dentry = pending->dentry;
892         parent_inode = dentry->d_parent->d_inode;
893         parent_root = BTRFS_I(parent_inode)->root;
894         record_root_in_trans(trans, parent_root);
895
896         /*
897          * insert the directory item
898          */
899         ret = btrfs_set_inode_index(parent_inode, &index);
900         BUG_ON(ret);
901         ret = btrfs_insert_dir_item(trans, parent_root,
902                                 dentry->d_name.name, dentry->d_name.len,
903                                 parent_inode->i_ino, &key,
904                                 BTRFS_FT_DIR, index);
905         BUG_ON(ret);
906
907         btrfs_i_size_write(parent_inode, parent_inode->i_size +
908                                          dentry->d_name.len * 2);
909         ret = btrfs_update_inode(trans, parent_root, parent_inode);
910         BUG_ON(ret);
911
912         record_root_in_trans(trans, root);
913         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
914         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
915
916         old = btrfs_lock_root_node(root);
917         btrfs_cow_block(trans, root, old, NULL, 0, &old);
918         btrfs_set_lock_blocking(old);
919
920         btrfs_copy_root(trans, root, old, &tmp, objectid);
921         btrfs_tree_unlock(old);
922         free_extent_buffer(old);
923
924         btrfs_set_root_node(new_root_item, tmp);
925         /* record when the snapshot was created in key.offset */
926         key.offset = trans->transid;
927         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
928         btrfs_tree_unlock(tmp);
929         free_extent_buffer(tmp);
930         BUG_ON(ret);
931
932         /*
933          * insert root back/forward references
934          */
935         ret = btrfs_add_root_ref(trans, tree_root, objectid,
936                                  parent_root->root_key.objectid,
937                                  parent_inode->i_ino, index,
938                                  dentry->d_name.name, dentry->d_name.len);
939         BUG_ON(ret);
940
941         key.offset = (u64)-1;
942         pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
943         BUG_ON(IS_ERR(pending->snap));
944
945         btrfs_reloc_post_snapshot(trans, pending);
946         btrfs_orphan_post_snapshot(trans, pending);
947 fail:
948         kfree(new_root_item);
949         btrfs_block_rsv_release(root, &pending->block_rsv, (u64)-1);
950         return 0;
951 }
952
953 /*
954  * create all the snapshots we've scheduled for creation
955  */
956 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
957                                              struct btrfs_fs_info *fs_info)
958 {
959         struct btrfs_pending_snapshot *pending;
960         struct list_head *head = &trans->transaction->pending_snapshots;
961         int ret;
962
963         list_for_each_entry(pending, head, list) {
964                 ret = create_pending_snapshot(trans, fs_info, pending);
965                 BUG_ON(ret);
966         }
967         return 0;
968 }
969
970 static void update_super_roots(struct btrfs_root *root)
971 {
972         struct btrfs_root_item *root_item;
973         struct btrfs_super_block *super;
974
975         super = &root->fs_info->super_copy;
976
977         root_item = &root->fs_info->chunk_root->root_item;
978         super->chunk_root = root_item->bytenr;
979         super->chunk_root_generation = root_item->generation;
980         super->chunk_root_level = root_item->level;
981
982         root_item = &root->fs_info->tree_root->root_item;
983         super->root = root_item->bytenr;
984         super->generation = root_item->generation;
985         super->root_level = root_item->level;
986         if (super->cache_generation != 0 || btrfs_test_opt(root, SPACE_CACHE))
987                 super->cache_generation = root_item->generation;
988 }
989
990 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
991 {
992         int ret = 0;
993         spin_lock(&info->new_trans_lock);
994         if (info->running_transaction)
995                 ret = info->running_transaction->in_commit;
996         spin_unlock(&info->new_trans_lock);
997         return ret;
998 }
999
1000 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1001 {
1002         int ret = 0;
1003         spin_lock(&info->new_trans_lock);
1004         if (info->running_transaction)
1005                 ret = info->running_transaction->blocked;
1006         spin_unlock(&info->new_trans_lock);
1007         return ret;
1008 }
1009
1010 /*
1011  * wait for the current transaction commit to start and block subsequent
1012  * transaction joins
1013  */
1014 static void wait_current_trans_commit_start(struct btrfs_root *root,
1015                                             struct btrfs_transaction *trans)
1016 {
1017         DEFINE_WAIT(wait);
1018
1019         if (trans->in_commit)
1020                 return;
1021
1022         while (1) {
1023                 prepare_to_wait(&root->fs_info->transaction_blocked_wait, &wait,
1024                                 TASK_UNINTERRUPTIBLE);
1025                 if (trans->in_commit) {
1026                         finish_wait(&root->fs_info->transaction_blocked_wait,
1027                                     &wait);
1028                         break;
1029                 }
1030                 mutex_unlock(&root->fs_info->trans_mutex);
1031                 schedule();
1032                 mutex_lock(&root->fs_info->trans_mutex);
1033                 finish_wait(&root->fs_info->transaction_blocked_wait, &wait);
1034         }
1035 }
1036
1037 /*
1038  * wait for the current transaction to start and then become unblocked.
1039  * caller holds ref.
1040  */
1041 static void wait_current_trans_commit_start_and_unblock(struct btrfs_root *root,
1042                                          struct btrfs_transaction *trans)
1043 {
1044         DEFINE_WAIT(wait);
1045
1046         if (trans->commit_done || (trans->in_commit && !trans->blocked))
1047                 return;
1048
1049         while (1) {
1050                 prepare_to_wait(&root->fs_info->transaction_wait, &wait,
1051                                 TASK_UNINTERRUPTIBLE);
1052                 if (trans->commit_done ||
1053                     (trans->in_commit && !trans->blocked)) {
1054                         finish_wait(&root->fs_info->transaction_wait,
1055                                     &wait);
1056                         break;
1057                 }
1058                 mutex_unlock(&root->fs_info->trans_mutex);
1059                 schedule();
1060                 mutex_lock(&root->fs_info->trans_mutex);
1061                 finish_wait(&root->fs_info->transaction_wait,
1062                             &wait);
1063         }
1064 }
1065
1066 /*
1067  * commit transactions asynchronously. once btrfs_commit_transaction_async
1068  * returns, any subsequent transaction will not be allowed to join.
1069  */
1070 struct btrfs_async_commit {
1071         struct btrfs_trans_handle *newtrans;
1072         struct btrfs_root *root;
1073         struct delayed_work work;
1074 };
1075
1076 static void do_async_commit(struct work_struct *work)
1077 {
1078         struct btrfs_async_commit *ac =
1079                 container_of(work, struct btrfs_async_commit, work.work);
1080
1081         btrfs_commit_transaction(ac->newtrans, ac->root);
1082         kfree(ac);
1083 }
1084
1085 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1086                                    struct btrfs_root *root,
1087                                    int wait_for_unblock)
1088 {
1089         struct btrfs_async_commit *ac;
1090         struct btrfs_transaction *cur_trans;
1091
1092         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1093         BUG_ON(!ac);
1094
1095         INIT_DELAYED_WORK(&ac->work, do_async_commit);
1096         ac->root = root;
1097         ac->newtrans = btrfs_join_transaction(root, 0);
1098
1099         /* take transaction reference */
1100         mutex_lock(&root->fs_info->trans_mutex);
1101         cur_trans = trans->transaction;
1102         cur_trans->use_count++;
1103         mutex_unlock(&root->fs_info->trans_mutex);
1104
1105         btrfs_end_transaction(trans, root);
1106         schedule_delayed_work(&ac->work, 0);
1107
1108         /* wait for transaction to start and unblock */
1109         mutex_lock(&root->fs_info->trans_mutex);
1110         if (wait_for_unblock)
1111                 wait_current_trans_commit_start_and_unblock(root, cur_trans);
1112         else
1113                 wait_current_trans_commit_start(root, cur_trans);
1114         put_transaction(cur_trans);
1115         mutex_unlock(&root->fs_info->trans_mutex);
1116
1117         return 0;
1118 }
1119
1120 /*
1121  * btrfs_transaction state sequence:
1122  *    in_commit = 0, blocked = 0  (initial)
1123  *    in_commit = 1, blocked = 1
1124  *    blocked = 0
1125  *    commit_done = 1
1126  */
1127 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
1128                              struct btrfs_root *root)
1129 {
1130         unsigned long joined = 0;
1131         struct btrfs_transaction *cur_trans;
1132         struct btrfs_transaction *prev_trans = NULL;
1133         DEFINE_WAIT(wait);
1134         int ret;
1135         int should_grow = 0;
1136         unsigned long now = get_seconds();
1137         int flush_on_commit = btrfs_test_opt(root, FLUSHONCOMMIT);
1138
1139         btrfs_run_ordered_operations(root, 0);
1140
1141         /* make a pass through all the delayed refs we have so far
1142          * any runnings procs may add more while we are here
1143          */
1144         ret = btrfs_run_delayed_refs(trans, root, 0);
1145         BUG_ON(ret);
1146
1147         btrfs_trans_release_metadata(trans, root);
1148
1149         cur_trans = trans->transaction;
1150         /*
1151          * set the flushing flag so procs in this transaction have to
1152          * start sending their work down.
1153          */
1154         cur_trans->delayed_refs.flushing = 1;
1155
1156         ret = btrfs_run_delayed_refs(trans, root, 0);
1157         BUG_ON(ret);
1158
1159         mutex_lock(&root->fs_info->trans_mutex);
1160         if (cur_trans->in_commit) {
1161                 cur_trans->use_count++;
1162                 mutex_unlock(&root->fs_info->trans_mutex);
1163                 btrfs_end_transaction(trans, root);
1164
1165                 ret = wait_for_commit(root, cur_trans);
1166                 BUG_ON(ret);
1167
1168                 mutex_lock(&root->fs_info->trans_mutex);
1169                 put_transaction(cur_trans);
1170                 mutex_unlock(&root->fs_info->trans_mutex);
1171
1172                 return 0;
1173         }
1174
1175         trans->transaction->in_commit = 1;
1176         trans->transaction->blocked = 1;
1177         wake_up(&root->fs_info->transaction_blocked_wait);
1178
1179         if (cur_trans->list.prev != &root->fs_info->trans_list) {
1180                 prev_trans = list_entry(cur_trans->list.prev,
1181                                         struct btrfs_transaction, list);
1182                 if (!prev_trans->commit_done) {
1183                         prev_trans->use_count++;
1184                         mutex_unlock(&root->fs_info->trans_mutex);
1185
1186                         wait_for_commit(root, prev_trans);
1187
1188                         mutex_lock(&root->fs_info->trans_mutex);
1189                         put_transaction(prev_trans);
1190                 }
1191         }
1192
1193         if (now < cur_trans->start_time || now - cur_trans->start_time < 1)
1194                 should_grow = 1;
1195
1196         do {
1197                 int snap_pending = 0;
1198                 joined = cur_trans->num_joined;
1199                 if (!list_empty(&trans->transaction->pending_snapshots))
1200                         snap_pending = 1;
1201
1202                 WARN_ON(cur_trans != trans->transaction);
1203                 mutex_unlock(&root->fs_info->trans_mutex);
1204
1205                 if (flush_on_commit || snap_pending) {
1206                         btrfs_start_delalloc_inodes(root, 1);
1207                         ret = btrfs_wait_ordered_extents(root, 0, 1);
1208                         BUG_ON(ret);
1209                 }
1210
1211                 /*
1212                  * rename don't use btrfs_join_transaction, so, once we
1213                  * set the transaction to blocked above, we aren't going
1214                  * to get any new ordered operations.  We can safely run
1215                  * it here and no for sure that nothing new will be added
1216                  * to the list
1217                  */
1218                 btrfs_run_ordered_operations(root, 1);
1219
1220                 prepare_to_wait(&cur_trans->writer_wait, &wait,
1221                                 TASK_UNINTERRUPTIBLE);
1222
1223                 smp_mb();
1224                 if (cur_trans->num_writers > 1)
1225                         schedule_timeout(MAX_SCHEDULE_TIMEOUT);
1226                 else if (should_grow)
1227                         schedule_timeout(1);
1228
1229                 mutex_lock(&root->fs_info->trans_mutex);
1230                 finish_wait(&cur_trans->writer_wait, &wait);
1231         } while (cur_trans->num_writers > 1 ||
1232                  (should_grow && cur_trans->num_joined != joined));
1233
1234         ret = create_pending_snapshots(trans, root->fs_info);
1235         BUG_ON(ret);
1236
1237         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1238         BUG_ON(ret);
1239
1240         WARN_ON(cur_trans != trans->transaction);
1241
1242         /* btrfs_commit_tree_roots is responsible for getting the
1243          * various roots consistent with each other.  Every pointer
1244          * in the tree of tree roots has to point to the most up to date
1245          * root for every subvolume and other tree.  So, we have to keep
1246          * the tree logging code from jumping in and changing any
1247          * of the trees.
1248          *
1249          * At this point in the commit, there can't be any tree-log
1250          * writers, but a little lower down we drop the trans mutex
1251          * and let new people in.  By holding the tree_log_mutex
1252          * from now until after the super is written, we avoid races
1253          * with the tree-log code.
1254          */
1255         mutex_lock(&root->fs_info->tree_log_mutex);
1256
1257         ret = commit_fs_roots(trans, root);
1258         BUG_ON(ret);
1259
1260         /* commit_fs_roots gets rid of all the tree log roots, it is now
1261          * safe to free the root of tree log roots
1262          */
1263         btrfs_free_log_root_tree(trans, root->fs_info);
1264
1265         ret = commit_cowonly_roots(trans, root);
1266         BUG_ON(ret);
1267
1268         btrfs_prepare_extent_commit(trans, root);
1269
1270         cur_trans = root->fs_info->running_transaction;
1271         spin_lock(&root->fs_info->new_trans_lock);
1272         root->fs_info->running_transaction = NULL;
1273         spin_unlock(&root->fs_info->new_trans_lock);
1274
1275         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
1276                             root->fs_info->tree_root->node);
1277         switch_commit_root(root->fs_info->tree_root);
1278
1279         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
1280                             root->fs_info->chunk_root->node);
1281         switch_commit_root(root->fs_info->chunk_root);
1282
1283         update_super_roots(root);
1284
1285         if (!root->fs_info->log_root_recovering) {
1286                 btrfs_set_super_log_root(&root->fs_info->super_copy, 0);
1287                 btrfs_set_super_log_root_level(&root->fs_info->super_copy, 0);
1288         }
1289
1290         memcpy(&root->fs_info->super_for_commit, &root->fs_info->super_copy,
1291                sizeof(root->fs_info->super_copy));
1292
1293         trans->transaction->blocked = 0;
1294
1295         wake_up(&root->fs_info->transaction_wait);
1296
1297         mutex_unlock(&root->fs_info->trans_mutex);
1298         ret = btrfs_write_and_wait_transaction(trans, root);
1299         BUG_ON(ret);
1300         write_ctree_super(trans, root, 0);
1301
1302         /*
1303          * the super is written, we can safely allow the tree-loggers
1304          * to go about their business
1305          */
1306         mutex_unlock(&root->fs_info->tree_log_mutex);
1307
1308         btrfs_finish_extent_commit(trans, root);
1309
1310         mutex_lock(&root->fs_info->trans_mutex);
1311
1312         cur_trans->commit_done = 1;
1313
1314         root->fs_info->last_trans_committed = cur_trans->transid;
1315
1316         wake_up(&cur_trans->commit_wait);
1317
1318         put_transaction(cur_trans);
1319         put_transaction(cur_trans);
1320
1321         mutex_unlock(&root->fs_info->trans_mutex);
1322
1323         if (current->journal_info == trans)
1324                 current->journal_info = NULL;
1325
1326         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1327
1328         if (current != root->fs_info->transaction_kthread)
1329                 btrfs_run_delayed_iputs(root);
1330
1331         return ret;
1332 }
1333
1334 /*
1335  * interface function to delete all the snapshots we have scheduled for deletion
1336  */
1337 int btrfs_clean_old_snapshots(struct btrfs_root *root)
1338 {
1339         LIST_HEAD(list);
1340         struct btrfs_fs_info *fs_info = root->fs_info;
1341
1342         mutex_lock(&fs_info->trans_mutex);
1343         list_splice_init(&fs_info->dead_roots, &list);
1344         mutex_unlock(&fs_info->trans_mutex);
1345
1346         while (!list_empty(&list)) {
1347                 root = list_entry(list.next, struct btrfs_root, root_list);
1348                 list_del(&root->root_list);
1349
1350                 if (btrfs_header_backref_rev(root->node) <
1351                     BTRFS_MIXED_BACKREF_REV)
1352                         btrfs_drop_snapshot(root, NULL, 0);
1353                 else
1354                         btrfs_drop_snapshot(root, NULL, 1);
1355         }
1356         return 0;
1357 }