Btrfs: be more polite in the async caching threads
[linux-2.6] / 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/sched.h>
21 #include <linux/writeback.h>
22 #include <linux/pagemap.h>
23 #include <linux/blkdev.h>
24 #include "ctree.h"
25 #include "disk-io.h"
26 #include "transaction.h"
27 #include "locking.h"
28 #include "tree-log.h"
29
30 #define BTRFS_ROOT_TRANS_TAG 0
31
32 static noinline void put_transaction(struct btrfs_transaction *transaction)
33 {
34         WARN_ON(transaction->use_count == 0);
35         transaction->use_count--;
36         if (transaction->use_count == 0) {
37                 list_del_init(&transaction->list);
38                 memset(transaction, 0, sizeof(*transaction));
39                 kmem_cache_free(btrfs_transaction_cachep, transaction);
40         }
41 }
42
43 static noinline void switch_commit_root(struct btrfs_root *root)
44 {
45         free_extent_buffer(root->commit_root);
46         root->commit_root = btrfs_root_node(root);
47 }
48
49 /*
50  * either allocate a new transaction or hop into the existing one
51  */
52 static noinline int join_transaction(struct btrfs_root *root)
53 {
54         struct btrfs_transaction *cur_trans;
55         cur_trans = root->fs_info->running_transaction;
56         if (!cur_trans) {
57                 cur_trans = kmem_cache_alloc(btrfs_transaction_cachep,
58                                              GFP_NOFS);
59                 BUG_ON(!cur_trans);
60                 root->fs_info->generation++;
61                 cur_trans->num_writers = 1;
62                 cur_trans->num_joined = 0;
63                 cur_trans->transid = root->fs_info->generation;
64                 init_waitqueue_head(&cur_trans->writer_wait);
65                 init_waitqueue_head(&cur_trans->commit_wait);
66                 cur_trans->in_commit = 0;
67                 cur_trans->blocked = 0;
68                 cur_trans->use_count = 1;
69                 cur_trans->commit_done = 0;
70                 cur_trans->start_time = get_seconds();
71
72                 cur_trans->delayed_refs.root.rb_node = NULL;
73                 cur_trans->delayed_refs.num_entries = 0;
74                 cur_trans->delayed_refs.num_heads_ready = 0;
75                 cur_trans->delayed_refs.num_heads = 0;
76                 cur_trans->delayed_refs.flushing = 0;
77                 cur_trans->delayed_refs.run_delayed_start = 0;
78                 spin_lock_init(&cur_trans->delayed_refs.lock);
79
80                 INIT_LIST_HEAD(&cur_trans->pending_snapshots);
81                 list_add_tail(&cur_trans->list, &root->fs_info->trans_list);
82                 extent_io_tree_init(&cur_trans->dirty_pages,
83                                      root->fs_info->btree_inode->i_mapping,
84                                      GFP_NOFS);
85                 spin_lock(&root->fs_info->new_trans_lock);
86                 root->fs_info->running_transaction = cur_trans;
87                 spin_unlock(&root->fs_info->new_trans_lock);
88         } else {
89                 cur_trans->num_writers++;
90                 cur_trans->num_joined++;
91         }
92
93         return 0;
94 }
95
96 /*
97  * this does all the record keeping required to make sure that a reference
98  * counted root is properly recorded in a given transaction.  This is required
99  * to make sure the old root from before we joined the transaction is deleted
100  * when the transaction commits
101  */
102 static noinline int record_root_in_trans(struct btrfs_trans_handle *trans,
103                                          struct btrfs_root *root)
104 {
105         if (root->ref_cows && root->last_trans < trans->transid) {
106                 WARN_ON(root == root->fs_info->extent_root);
107                 WARN_ON(root->root_item.refs == 0);
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                                 mutex_unlock(&root->fs_info->trans_mutex);
153                                 schedule();
154                                 mutex_lock(&root->fs_info->trans_mutex);
155                                 finish_wait(&root->fs_info->transaction_wait,
156                                             &wait);
157                         } else {
158                                 finish_wait(&root->fs_info->transaction_wait,
159                                             &wait);
160                                 break;
161                         }
162                 }
163                 put_transaction(cur_trans);
164         }
165 }
166
167 static struct btrfs_trans_handle *start_transaction(struct btrfs_root *root,
168                                              int num_blocks, int wait)
169 {
170         struct btrfs_trans_handle *h =
171                 kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
172         int ret;
173
174         mutex_lock(&root->fs_info->trans_mutex);
175         if (!root->fs_info->log_root_recovering &&
176             ((wait == 1 && !root->fs_info->open_ioctl_trans) || wait == 2))
177                 wait_current_trans(root);
178         ret = join_transaction(root);
179         BUG_ON(ret);
180
181         h->transid = root->fs_info->running_transaction->transid;
182         h->transaction = root->fs_info->running_transaction;
183         h->blocks_reserved = num_blocks;
184         h->blocks_used = 0;
185         h->block_group = 0;
186         h->alloc_exclude_nr = 0;
187         h->alloc_exclude_start = 0;
188         h->delayed_ref_updates = 0;
189
190         root->fs_info->running_transaction->use_count++;
191         record_root_in_trans(h, root);
192         mutex_unlock(&root->fs_info->trans_mutex);
193         return h;
194 }
195
196 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
197                                                    int num_blocks)
198 {
199         return start_transaction(root, num_blocks, 1);
200 }
201 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root,
202                                                    int num_blocks)
203 {
204         return start_transaction(root, num_blocks, 0);
205 }
206
207 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *r,
208                                                          int num_blocks)
209 {
210         return start_transaction(r, num_blocks, 2);
211 }
212
213 /* wait for a transaction commit to be fully complete */
214 static noinline int wait_for_commit(struct btrfs_root *root,
215                                     struct btrfs_transaction *commit)
216 {
217         DEFINE_WAIT(wait);
218         mutex_lock(&root->fs_info->trans_mutex);
219         while (!commit->commit_done) {
220                 prepare_to_wait(&commit->commit_wait, &wait,
221                                 TASK_UNINTERRUPTIBLE);
222                 if (commit->commit_done)
223                         break;
224                 mutex_unlock(&root->fs_info->trans_mutex);
225                 schedule();
226                 mutex_lock(&root->fs_info->trans_mutex);
227         }
228         mutex_unlock(&root->fs_info->trans_mutex);
229         finish_wait(&commit->commit_wait, &wait);
230         return 0;
231 }
232
233 #if 0
234 /*
235  * rate limit against the drop_snapshot code.  This helps to slow down new
236  * operations if the drop_snapshot code isn't able to keep up.
237  */
238 static void throttle_on_drops(struct btrfs_root *root)
239 {
240         struct btrfs_fs_info *info = root->fs_info;
241         int harder_count = 0;
242
243 harder:
244         if (atomic_read(&info->throttles)) {
245                 DEFINE_WAIT(wait);
246                 int thr;
247                 thr = atomic_read(&info->throttle_gen);
248
249                 do {
250                         prepare_to_wait(&info->transaction_throttle,
251                                         &wait, TASK_UNINTERRUPTIBLE);
252                         if (!atomic_read(&info->throttles)) {
253                                 finish_wait(&info->transaction_throttle, &wait);
254                                 break;
255                         }
256                         schedule();
257                         finish_wait(&info->transaction_throttle, &wait);
258                 } while (thr == atomic_read(&info->throttle_gen));
259                 harder_count++;
260
261                 if (root->fs_info->total_ref_cache_size > 1 * 1024 * 1024 &&
262                     harder_count < 2)
263                         goto harder;
264
265                 if (root->fs_info->total_ref_cache_size > 5 * 1024 * 1024 &&
266                     harder_count < 10)
267                         goto harder;
268
269                 if (root->fs_info->total_ref_cache_size > 10 * 1024 * 1024 &&
270                     harder_count < 20)
271                         goto harder;
272         }
273 }
274 #endif
275
276 void btrfs_throttle(struct btrfs_root *root)
277 {
278         mutex_lock(&root->fs_info->trans_mutex);
279         if (!root->fs_info->open_ioctl_trans)
280                 wait_current_trans(root);
281         mutex_unlock(&root->fs_info->trans_mutex);
282 }
283
284 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
285                           struct btrfs_root *root, int throttle)
286 {
287         struct btrfs_transaction *cur_trans;
288         struct btrfs_fs_info *info = root->fs_info;
289         int count = 0;
290
291         while (count < 4) {
292                 unsigned long cur = trans->delayed_ref_updates;
293                 trans->delayed_ref_updates = 0;
294                 if (cur &&
295                     trans->transaction->delayed_refs.num_heads_ready > 64) {
296                         trans->delayed_ref_updates = 0;
297
298                         /*
299                          * do a full flush if the transaction is trying
300                          * to close
301                          */
302                         if (trans->transaction->delayed_refs.flushing)
303                                 cur = 0;
304                         btrfs_run_delayed_refs(trans, root, cur);
305                 } else {
306                         break;
307                 }
308                 count++;
309         }
310
311         mutex_lock(&info->trans_mutex);
312         cur_trans = info->running_transaction;
313         WARN_ON(cur_trans != trans->transaction);
314         WARN_ON(cur_trans->num_writers < 1);
315         cur_trans->num_writers--;
316
317         if (waitqueue_active(&cur_trans->writer_wait))
318                 wake_up(&cur_trans->writer_wait);
319         put_transaction(cur_trans);
320         mutex_unlock(&info->trans_mutex);
321         memset(trans, 0, sizeof(*trans));
322         kmem_cache_free(btrfs_trans_handle_cachep, trans);
323
324         return 0;
325 }
326
327 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
328                           struct btrfs_root *root)
329 {
330         return __btrfs_end_transaction(trans, root, 0);
331 }
332
333 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
334                                    struct btrfs_root *root)
335 {
336         return __btrfs_end_transaction(trans, root, 1);
337 }
338
339 /*
340  * when btree blocks are allocated, they have some corresponding bits set for
341  * them in one of two extent_io trees.  This is used to make sure all of
342  * those extents are on disk for transaction or log commit
343  */
344 int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
345                                         struct extent_io_tree *dirty_pages)
346 {
347         int ret;
348         int err = 0;
349         int werr = 0;
350         struct page *page;
351         struct inode *btree_inode = root->fs_info->btree_inode;
352         u64 start = 0;
353         u64 end;
354         unsigned long index;
355
356         while (1) {
357                 ret = find_first_extent_bit(dirty_pages, start, &start, &end,
358                                             EXTENT_DIRTY);
359                 if (ret)
360                         break;
361                 while (start <= end) {
362                         cond_resched();
363
364                         index = start >> PAGE_CACHE_SHIFT;
365                         start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
366                         page = find_get_page(btree_inode->i_mapping, index);
367                         if (!page)
368                                 continue;
369
370                         btree_lock_page_hook(page);
371                         if (!page->mapping) {
372                                 unlock_page(page);
373                                 page_cache_release(page);
374                                 continue;
375                         }
376
377                         if (PageWriteback(page)) {
378                                 if (PageDirty(page))
379                                         wait_on_page_writeback(page);
380                                 else {
381                                         unlock_page(page);
382                                         page_cache_release(page);
383                                         continue;
384                                 }
385                         }
386                         err = write_one_page(page, 0);
387                         if (err)
388                                 werr = err;
389                         page_cache_release(page);
390                 }
391         }
392         while (1) {
393                 ret = find_first_extent_bit(dirty_pages, 0, &start, &end,
394                                             EXTENT_DIRTY);
395                 if (ret)
396                         break;
397
398                 clear_extent_dirty(dirty_pages, start, end, GFP_NOFS);
399                 while (start <= end) {
400                         index = start >> PAGE_CACHE_SHIFT;
401                         start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
402                         page = find_get_page(btree_inode->i_mapping, index);
403                         if (!page)
404                                 continue;
405                         if (PageDirty(page)) {
406                                 btree_lock_page_hook(page);
407                                 wait_on_page_writeback(page);
408                                 err = write_one_page(page, 0);
409                                 if (err)
410                                         werr = err;
411                         }
412                         wait_on_page_writeback(page);
413                         page_cache_release(page);
414                         cond_resched();
415                 }
416         }
417         if (err)
418                 werr = err;
419         return werr;
420 }
421
422 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
423                                      struct btrfs_root *root)
424 {
425         if (!trans || !trans->transaction) {
426                 struct inode *btree_inode;
427                 btree_inode = root->fs_info->btree_inode;
428                 return filemap_write_and_wait(btree_inode->i_mapping);
429         }
430         return btrfs_write_and_wait_marked_extents(root,
431                                            &trans->transaction->dirty_pages);
432 }
433
434 /*
435  * this is used to update the root pointer in the tree of tree roots.
436  *
437  * But, in the case of the extent allocation tree, updating the root
438  * pointer may allocate blocks which may change the root of the extent
439  * allocation tree.
440  *
441  * So, this loops and repeats and makes sure the cowonly root didn't
442  * change while the root pointer was being updated in the metadata.
443  */
444 static int update_cowonly_root(struct btrfs_trans_handle *trans,
445                                struct btrfs_root *root)
446 {
447         int ret;
448         u64 old_root_bytenr;
449         struct btrfs_root *tree_root = root->fs_info->tree_root;
450
451         btrfs_write_dirty_block_groups(trans, root);
452
453         while (1) {
454                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
455                 if (old_root_bytenr == root->node->start)
456                         break;
457
458                 btrfs_set_root_node(&root->root_item, root->node);
459                 ret = btrfs_update_root(trans, tree_root,
460                                         &root->root_key,
461                                         &root->root_item);
462                 BUG_ON(ret);
463
464                 ret = btrfs_write_dirty_block_groups(trans, root);
465                 BUG_ON(ret);
466         }
467
468         if (root != root->fs_info->extent_root)
469                 switch_commit_root(root);
470
471         return 0;
472 }
473
474 /*
475  * update all the cowonly tree roots on disk
476  */
477 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
478                                          struct btrfs_root *root)
479 {
480         struct btrfs_fs_info *fs_info = root->fs_info;
481         struct list_head *next;
482         struct extent_buffer *eb;
483         int ret;
484
485         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
486         BUG_ON(ret);
487
488         eb = btrfs_lock_root_node(fs_info->tree_root);
489         btrfs_cow_block(trans, fs_info->tree_root, eb, NULL, 0, &eb);
490         btrfs_tree_unlock(eb);
491         free_extent_buffer(eb);
492
493         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
494         BUG_ON(ret);
495
496         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
497                 next = fs_info->dirty_cowonly_roots.next;
498                 list_del_init(next);
499                 root = list_entry(next, struct btrfs_root, dirty_list);
500
501                 update_cowonly_root(trans, root);
502         }
503
504         down_write(&fs_info->extent_commit_sem);
505         switch_commit_root(fs_info->extent_root);
506         up_write(&fs_info->extent_commit_sem);
507
508         return 0;
509 }
510
511 /*
512  * dead roots are old snapshots that need to be deleted.  This allocates
513  * a dirty root struct and adds it into the list of dead roots that need to
514  * be deleted
515  */
516 int btrfs_add_dead_root(struct btrfs_root *root)
517 {
518         mutex_lock(&root->fs_info->trans_mutex);
519         list_add(&root->root_list, &root->fs_info->dead_roots);
520         mutex_unlock(&root->fs_info->trans_mutex);
521         return 0;
522 }
523
524 /*
525  * update all the cowonly tree roots on disk
526  */
527 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
528                                     struct btrfs_root *root)
529 {
530         struct btrfs_root *gang[8];
531         struct btrfs_fs_info *fs_info = root->fs_info;
532         int i;
533         int ret;
534         int err = 0;
535
536         while (1) {
537                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
538                                                  (void **)gang, 0,
539                                                  ARRAY_SIZE(gang),
540                                                  BTRFS_ROOT_TRANS_TAG);
541                 if (ret == 0)
542                         break;
543                 for (i = 0; i < ret; i++) {
544                         root = gang[i];
545                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
546                                         (unsigned long)root->root_key.objectid,
547                                         BTRFS_ROOT_TRANS_TAG);
548
549                         btrfs_free_log(trans, root);
550                         btrfs_update_reloc_root(trans, root);
551
552                         if (root->commit_root != root->node) {
553                                 switch_commit_root(root);
554                                 btrfs_set_root_node(&root->root_item,
555                                                     root->node);
556                         }
557
558                         err = btrfs_update_root(trans, fs_info->tree_root,
559                                                 &root->root_key,
560                                                 &root->root_item);
561                         if (err)
562                                 break;
563                 }
564         }
565         return err;
566 }
567
568 /*
569  * defrag a given btree.  If cacheonly == 1, this won't read from the disk,
570  * otherwise every leaf in the btree is read and defragged.
571  */
572 int btrfs_defrag_root(struct btrfs_root *root, int cacheonly)
573 {
574         struct btrfs_fs_info *info = root->fs_info;
575         int ret;
576         struct btrfs_trans_handle *trans;
577         unsigned long nr;
578
579         smp_mb();
580         if (root->defrag_running)
581                 return 0;
582         trans = btrfs_start_transaction(root, 1);
583         while (1) {
584                 root->defrag_running = 1;
585                 ret = btrfs_defrag_leaves(trans, root, cacheonly);
586                 nr = trans->blocks_used;
587                 btrfs_end_transaction(trans, root);
588                 btrfs_btree_balance_dirty(info->tree_root, nr);
589                 cond_resched();
590
591                 trans = btrfs_start_transaction(root, 1);
592                 if (root->fs_info->closing || ret != -EAGAIN)
593                         break;
594         }
595         root->defrag_running = 0;
596         smp_mb();
597         btrfs_end_transaction(trans, root);
598         return 0;
599 }
600
601 #if 0
602 /*
603  * when dropping snapshots, we generate a ton of delayed refs, and it makes
604  * sense not to join the transaction while it is trying to flush the current
605  * queue of delayed refs out.
606  *
607  * This is used by the drop snapshot code only
608  */
609 static noinline int wait_transaction_pre_flush(struct btrfs_fs_info *info)
610 {
611         DEFINE_WAIT(wait);
612
613         mutex_lock(&info->trans_mutex);
614         while (info->running_transaction &&
615                info->running_transaction->delayed_refs.flushing) {
616                 prepare_to_wait(&info->transaction_wait, &wait,
617                                 TASK_UNINTERRUPTIBLE);
618                 mutex_unlock(&info->trans_mutex);
619
620                 schedule();
621
622                 mutex_lock(&info->trans_mutex);
623                 finish_wait(&info->transaction_wait, &wait);
624         }
625         mutex_unlock(&info->trans_mutex);
626         return 0;
627 }
628
629 /*
630  * Given a list of roots that need to be deleted, call btrfs_drop_snapshot on
631  * all of them
632  */
633 int btrfs_drop_dead_root(struct btrfs_root *root)
634 {
635         struct btrfs_trans_handle *trans;
636         struct btrfs_root *tree_root = root->fs_info->tree_root;
637         unsigned long nr;
638         int ret;
639
640         while (1) {
641                 /*
642                  * we don't want to jump in and create a bunch of
643                  * delayed refs if the transaction is starting to close
644                  */
645                 wait_transaction_pre_flush(tree_root->fs_info);
646                 trans = btrfs_start_transaction(tree_root, 1);
647
648                 /*
649                  * we've joined a transaction, make sure it isn't
650                  * closing right now
651                  */
652                 if (trans->transaction->delayed_refs.flushing) {
653                         btrfs_end_transaction(trans, tree_root);
654                         continue;
655                 }
656
657                 ret = btrfs_drop_snapshot(trans, root);
658                 if (ret != -EAGAIN)
659                         break;
660
661                 ret = btrfs_update_root(trans, tree_root,
662                                         &root->root_key,
663                                         &root->root_item);
664                 if (ret)
665                         break;
666
667                 nr = trans->blocks_used;
668                 ret = btrfs_end_transaction(trans, tree_root);
669                 BUG_ON(ret);
670
671                 btrfs_btree_balance_dirty(tree_root, nr);
672                 cond_resched();
673         }
674         BUG_ON(ret);
675
676         ret = btrfs_del_root(trans, tree_root, &root->root_key);
677         BUG_ON(ret);
678
679         nr = trans->blocks_used;
680         ret = btrfs_end_transaction(trans, tree_root);
681         BUG_ON(ret);
682
683         free_extent_buffer(root->node);
684         free_extent_buffer(root->commit_root);
685         kfree(root);
686
687         btrfs_btree_balance_dirty(tree_root, nr);
688         return ret;
689 }
690 #endif
691
692 /*
693  * new snapshots need to be created at a very specific time in the
694  * transaction commit.  This does the actual creation
695  */
696 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
697                                    struct btrfs_fs_info *fs_info,
698                                    struct btrfs_pending_snapshot *pending)
699 {
700         struct btrfs_key key;
701         struct btrfs_root_item *new_root_item;
702         struct btrfs_root *tree_root = fs_info->tree_root;
703         struct btrfs_root *root = pending->root;
704         struct extent_buffer *tmp;
705         struct extent_buffer *old;
706         int ret;
707         u64 objectid;
708
709         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
710         if (!new_root_item) {
711                 ret = -ENOMEM;
712                 goto fail;
713         }
714         ret = btrfs_find_free_objectid(trans, tree_root, 0, &objectid);
715         if (ret)
716                 goto fail;
717
718         record_root_in_trans(trans, root);
719         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
720         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
721
722         key.objectid = objectid;
723         key.offset = 0;
724         btrfs_set_key_type(&key, BTRFS_ROOT_ITEM_KEY);
725
726         old = btrfs_lock_root_node(root);
727         btrfs_cow_block(trans, root, old, NULL, 0, &old);
728         btrfs_set_lock_blocking(old);
729
730         btrfs_copy_root(trans, root, old, &tmp, objectid);
731         btrfs_tree_unlock(old);
732         free_extent_buffer(old);
733
734         btrfs_set_root_node(new_root_item, tmp);
735         ret = btrfs_insert_root(trans, root->fs_info->tree_root, &key,
736                                 new_root_item);
737         btrfs_tree_unlock(tmp);
738         free_extent_buffer(tmp);
739         if (ret)
740                 goto fail;
741
742         key.offset = (u64)-1;
743         memcpy(&pending->root_key, &key, sizeof(key));
744 fail:
745         kfree(new_root_item);
746         return ret;
747 }
748
749 static noinline int finish_pending_snapshot(struct btrfs_fs_info *fs_info,
750                                    struct btrfs_pending_snapshot *pending)
751 {
752         int ret;
753         int namelen;
754         u64 index = 0;
755         struct btrfs_trans_handle *trans;
756         struct inode *parent_inode;
757         struct inode *inode;
758         struct btrfs_root *parent_root;
759
760         parent_inode = pending->dentry->d_parent->d_inode;
761         parent_root = BTRFS_I(parent_inode)->root;
762         trans = btrfs_join_transaction(parent_root, 1);
763
764         /*
765          * insert the directory item
766          */
767         namelen = strlen(pending->name);
768         ret = btrfs_set_inode_index(parent_inode, &index);
769         ret = btrfs_insert_dir_item(trans, parent_root,
770                             pending->name, namelen,
771                             parent_inode->i_ino,
772                             &pending->root_key, BTRFS_FT_DIR, index);
773
774         if (ret)
775                 goto fail;
776
777         btrfs_i_size_write(parent_inode, parent_inode->i_size + namelen * 2);
778         ret = btrfs_update_inode(trans, parent_root, parent_inode);
779         BUG_ON(ret);
780
781         /* add the backref first */
782         ret = btrfs_add_root_ref(trans, parent_root->fs_info->tree_root,
783                                  pending->root_key.objectid,
784                                  BTRFS_ROOT_BACKREF_KEY,
785                                  parent_root->root_key.objectid,
786                                  parent_inode->i_ino, index, pending->name,
787                                  namelen);
788
789         BUG_ON(ret);
790
791         /* now add the forward ref */
792         ret = btrfs_add_root_ref(trans, parent_root->fs_info->tree_root,
793                                  parent_root->root_key.objectid,
794                                  BTRFS_ROOT_REF_KEY,
795                                  pending->root_key.objectid,
796                                  parent_inode->i_ino, index, pending->name,
797                                  namelen);
798
799         inode = btrfs_lookup_dentry(parent_inode, pending->dentry);
800         d_instantiate(pending->dentry, inode);
801 fail:
802         btrfs_end_transaction(trans, fs_info->fs_root);
803         return ret;
804 }
805
806 /*
807  * create all the snapshots we've scheduled for creation
808  */
809 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
810                                              struct btrfs_fs_info *fs_info)
811 {
812         struct btrfs_pending_snapshot *pending;
813         struct list_head *head = &trans->transaction->pending_snapshots;
814         int ret;
815
816         list_for_each_entry(pending, head, list) {
817                 ret = create_pending_snapshot(trans, fs_info, pending);
818                 BUG_ON(ret);
819         }
820         return 0;
821 }
822
823 static noinline int finish_pending_snapshots(struct btrfs_trans_handle *trans,
824                                              struct btrfs_fs_info *fs_info)
825 {
826         struct btrfs_pending_snapshot *pending;
827         struct list_head *head = &trans->transaction->pending_snapshots;
828         int ret;
829
830         while (!list_empty(head)) {
831                 pending = list_entry(head->next,
832                                      struct btrfs_pending_snapshot, list);
833                 ret = finish_pending_snapshot(fs_info, pending);
834                 BUG_ON(ret);
835                 list_del(&pending->list);
836                 kfree(pending->name);
837                 kfree(pending);
838         }
839         return 0;
840 }
841
842 static void update_super_roots(struct btrfs_root *root)
843 {
844         struct btrfs_root_item *root_item;
845         struct btrfs_super_block *super;
846
847         super = &root->fs_info->super_copy;
848
849         root_item = &root->fs_info->chunk_root->root_item;
850         super->chunk_root = root_item->bytenr;
851         super->chunk_root_generation = root_item->generation;
852         super->chunk_root_level = root_item->level;
853
854         root_item = &root->fs_info->tree_root->root_item;
855         super->root = root_item->bytenr;
856         super->generation = root_item->generation;
857         super->root_level = root_item->level;
858 }
859
860 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
861 {
862         int ret = 0;
863         spin_lock(&info->new_trans_lock);
864         if (info->running_transaction)
865                 ret = info->running_transaction->in_commit;
866         spin_unlock(&info->new_trans_lock);
867         return ret;
868 }
869
870 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
871                              struct btrfs_root *root)
872 {
873         unsigned long joined = 0;
874         unsigned long timeout = 1;
875         struct btrfs_transaction *cur_trans;
876         struct btrfs_transaction *prev_trans = NULL;
877         struct extent_io_tree *pinned_copy;
878         DEFINE_WAIT(wait);
879         int ret;
880         int should_grow = 0;
881         unsigned long now = get_seconds();
882         int flush_on_commit = btrfs_test_opt(root, FLUSHONCOMMIT);
883
884         btrfs_run_ordered_operations(root, 0);
885
886         /* make a pass through all the delayed refs we have so far
887          * any runnings procs may add more while we are here
888          */
889         ret = btrfs_run_delayed_refs(trans, root, 0);
890         BUG_ON(ret);
891
892         cur_trans = trans->transaction;
893         /*
894          * set the flushing flag so procs in this transaction have to
895          * start sending their work down.
896          */
897         cur_trans->delayed_refs.flushing = 1;
898
899         ret = btrfs_run_delayed_refs(trans, root, 0);
900         BUG_ON(ret);
901
902         mutex_lock(&root->fs_info->trans_mutex);
903         if (cur_trans->in_commit) {
904                 cur_trans->use_count++;
905                 mutex_unlock(&root->fs_info->trans_mutex);
906                 btrfs_end_transaction(trans, root);
907
908                 ret = wait_for_commit(root, cur_trans);
909                 BUG_ON(ret);
910
911                 mutex_lock(&root->fs_info->trans_mutex);
912                 put_transaction(cur_trans);
913                 mutex_unlock(&root->fs_info->trans_mutex);
914
915                 return 0;
916         }
917
918         pinned_copy = kmalloc(sizeof(*pinned_copy), GFP_NOFS);
919         if (!pinned_copy)
920                 return -ENOMEM;
921
922         extent_io_tree_init(pinned_copy,
923                              root->fs_info->btree_inode->i_mapping, GFP_NOFS);
924
925         trans->transaction->in_commit = 1;
926         trans->transaction->blocked = 1;
927         if (cur_trans->list.prev != &root->fs_info->trans_list) {
928                 prev_trans = list_entry(cur_trans->list.prev,
929                                         struct btrfs_transaction, list);
930                 if (!prev_trans->commit_done) {
931                         prev_trans->use_count++;
932                         mutex_unlock(&root->fs_info->trans_mutex);
933
934                         wait_for_commit(root, prev_trans);
935
936                         mutex_lock(&root->fs_info->trans_mutex);
937                         put_transaction(prev_trans);
938                 }
939         }
940
941         if (now < cur_trans->start_time || now - cur_trans->start_time < 1)
942                 should_grow = 1;
943
944         do {
945                 int snap_pending = 0;
946                 joined = cur_trans->num_joined;
947                 if (!list_empty(&trans->transaction->pending_snapshots))
948                         snap_pending = 1;
949
950                 WARN_ON(cur_trans != trans->transaction);
951                 prepare_to_wait(&cur_trans->writer_wait, &wait,
952                                 TASK_UNINTERRUPTIBLE);
953
954                 if (cur_trans->num_writers > 1)
955                         timeout = MAX_SCHEDULE_TIMEOUT;
956                 else if (should_grow)
957                         timeout = 1;
958
959                 mutex_unlock(&root->fs_info->trans_mutex);
960
961                 if (flush_on_commit) {
962                         btrfs_start_delalloc_inodes(root);
963                         ret = btrfs_wait_ordered_extents(root, 0);
964                         BUG_ON(ret);
965                 } else if (snap_pending) {
966                         ret = btrfs_wait_ordered_extents(root, 1);
967                         BUG_ON(ret);
968                 }
969
970                 /*
971                  * rename don't use btrfs_join_transaction, so, once we
972                  * set the transaction to blocked above, we aren't going
973                  * to get any new ordered operations.  We can safely run
974                  * it here and no for sure that nothing new will be added
975                  * to the list
976                  */
977                 btrfs_run_ordered_operations(root, 1);
978
979                 smp_mb();
980                 if (cur_trans->num_writers > 1 || should_grow)
981                         schedule_timeout(timeout);
982
983                 mutex_lock(&root->fs_info->trans_mutex);
984                 finish_wait(&cur_trans->writer_wait, &wait);
985         } while (cur_trans->num_writers > 1 ||
986                  (should_grow && cur_trans->num_joined != joined));
987
988         ret = create_pending_snapshots(trans, root->fs_info);
989         BUG_ON(ret);
990
991         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
992         BUG_ON(ret);
993
994         WARN_ON(cur_trans != trans->transaction);
995
996         /* btrfs_commit_tree_roots is responsible for getting the
997          * various roots consistent with each other.  Every pointer
998          * in the tree of tree roots has to point to the most up to date
999          * root for every subvolume and other tree.  So, we have to keep
1000          * the tree logging code from jumping in and changing any
1001          * of the trees.
1002          *
1003          * At this point in the commit, there can't be any tree-log
1004          * writers, but a little lower down we drop the trans mutex
1005          * and let new people in.  By holding the tree_log_mutex
1006          * from now until after the super is written, we avoid races
1007          * with the tree-log code.
1008          */
1009         mutex_lock(&root->fs_info->tree_log_mutex);
1010
1011         ret = commit_fs_roots(trans, root);
1012         BUG_ON(ret);
1013
1014         /* commit_fs_roots gets rid of all the tree log roots, it is now
1015          * safe to free the root of tree log roots
1016          */
1017         btrfs_free_log_root_tree(trans, root->fs_info);
1018
1019         ret = commit_cowonly_roots(trans, root);
1020         BUG_ON(ret);
1021
1022         cur_trans = root->fs_info->running_transaction;
1023         spin_lock(&root->fs_info->new_trans_lock);
1024         root->fs_info->running_transaction = NULL;
1025         spin_unlock(&root->fs_info->new_trans_lock);
1026
1027         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
1028                             root->fs_info->tree_root->node);
1029         switch_commit_root(root->fs_info->tree_root);
1030
1031         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
1032                             root->fs_info->chunk_root->node);
1033         switch_commit_root(root->fs_info->chunk_root);
1034
1035         update_super_roots(root);
1036
1037         if (!root->fs_info->log_root_recovering) {
1038                 btrfs_set_super_log_root(&root->fs_info->super_copy, 0);
1039                 btrfs_set_super_log_root_level(&root->fs_info->super_copy, 0);
1040         }
1041
1042         memcpy(&root->fs_info->super_for_commit, &root->fs_info->super_copy,
1043                sizeof(root->fs_info->super_copy));
1044
1045         btrfs_copy_pinned(root, pinned_copy);
1046
1047         trans->transaction->blocked = 0;
1048
1049         wake_up(&root->fs_info->transaction_wait);
1050
1051         mutex_unlock(&root->fs_info->trans_mutex);
1052         ret = btrfs_write_and_wait_transaction(trans, root);
1053         BUG_ON(ret);
1054         write_ctree_super(trans, root, 0);
1055
1056         /*
1057          * the super is written, we can safely allow the tree-loggers
1058          * to go about their business
1059          */
1060         mutex_unlock(&root->fs_info->tree_log_mutex);
1061
1062         btrfs_finish_extent_commit(trans, root, pinned_copy);
1063         kfree(pinned_copy);
1064
1065         /* do the directory inserts of any pending snapshot creations */
1066         finish_pending_snapshots(trans, root->fs_info);
1067
1068         mutex_lock(&root->fs_info->trans_mutex);
1069
1070         cur_trans->commit_done = 1;
1071
1072         root->fs_info->last_trans_committed = cur_trans->transid;
1073
1074         wake_up(&cur_trans->commit_wait);
1075
1076         put_transaction(cur_trans);
1077         put_transaction(cur_trans);
1078
1079         mutex_unlock(&root->fs_info->trans_mutex);
1080
1081         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1082         return ret;
1083 }
1084
1085 /*
1086  * interface function to delete all the snapshots we have scheduled for deletion
1087  */
1088 int btrfs_clean_old_snapshots(struct btrfs_root *root)
1089 {
1090         LIST_HEAD(list);
1091         struct btrfs_fs_info *fs_info = root->fs_info;
1092
1093         mutex_lock(&fs_info->trans_mutex);
1094         list_splice_init(&fs_info->dead_roots, &list);
1095         mutex_unlock(&fs_info->trans_mutex);
1096
1097         while (!list_empty(&list)) {
1098                 root = list_entry(list.next, struct btrfs_root, root_list);
1099                 list_del_init(&root->root_list);
1100                 btrfs_drop_snapshot(root, 0);
1101         }
1102         return 0;
1103 }