Btrfs: delay commits during fsync to allow more writers
[safe/jmp/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 "ctree.h"
22 #include "disk-io.h"
23 #include "transaction.h"
24
25 static int total_trans = 0;
26 extern struct kmem_cache *btrfs_trans_handle_cachep;
27 extern struct kmem_cache *btrfs_transaction_cachep;
28
29 static struct workqueue_struct *trans_wq;
30
31 #define BTRFS_ROOT_TRANS_TAG 0
32 #define BTRFS_ROOT_DEFRAG_TAG 1
33
34 static void put_transaction(struct btrfs_transaction *transaction)
35 {
36         WARN_ON(transaction->use_count == 0);
37         transaction->use_count--;
38         if (transaction->use_count == 0) {
39                 WARN_ON(total_trans == 0);
40                 total_trans--;
41                 list_del_init(&transaction->list);
42                 memset(transaction, 0, sizeof(*transaction));
43                 kmem_cache_free(btrfs_transaction_cachep, transaction);
44         }
45 }
46
47 static int join_transaction(struct btrfs_root *root)
48 {
49         struct btrfs_transaction *cur_trans;
50         cur_trans = root->fs_info->running_transaction;
51         if (!cur_trans) {
52                 cur_trans = kmem_cache_alloc(btrfs_transaction_cachep,
53                                              GFP_NOFS);
54                 total_trans++;
55                 BUG_ON(!cur_trans);
56                 root->fs_info->generation++;
57                 root->fs_info->running_transaction = cur_trans;
58                 cur_trans->num_writers = 1;
59                 cur_trans->num_joined = 0;
60                 cur_trans->transid = root->fs_info->generation;
61                 init_waitqueue_head(&cur_trans->writer_wait);
62                 init_waitqueue_head(&cur_trans->commit_wait);
63                 cur_trans->in_commit = 0;
64                 cur_trans->use_count = 1;
65                 cur_trans->commit_done = 0;
66                 cur_trans->start_time = get_seconds();
67                 list_add_tail(&cur_trans->list, &root->fs_info->trans_list);
68                 init_bit_radix(&cur_trans->dirty_pages);
69         } else {
70                 cur_trans->num_writers++;
71                 cur_trans->num_joined++;
72         }
73
74         return 0;
75 }
76
77 static int record_root_in_trans(struct btrfs_root *root)
78 {
79         u64 running_trans_id = root->fs_info->running_transaction->transid;
80         if (root->ref_cows && root->last_trans < running_trans_id) {
81                 WARN_ON(root == root->fs_info->extent_root);
82                 if (root->root_item.refs != 0) {
83                         radix_tree_tag_set(&root->fs_info->fs_roots_radix,
84                                    (unsigned long)root->root_key.objectid,
85                                    BTRFS_ROOT_TRANS_TAG);
86                         radix_tree_tag_set(&root->fs_info->fs_roots_radix,
87                                    (unsigned long)root->root_key.objectid,
88                                    BTRFS_ROOT_DEFRAG_TAG);
89                         root->commit_root = root->node;
90                         get_bh(root->node);
91                 } else {
92                         WARN_ON(1);
93                 }
94                 root->last_trans = running_trans_id;
95         }
96         return 0;
97 }
98
99 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
100                                                    int num_blocks)
101 {
102         struct btrfs_trans_handle *h =
103                 kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
104         int ret;
105
106         mutex_lock(&root->fs_info->trans_mutex);
107         ret = join_transaction(root);
108         BUG_ON(ret);
109
110         record_root_in_trans(root);
111         h->transid = root->fs_info->running_transaction->transid;
112         h->transaction = root->fs_info->running_transaction;
113         h->blocks_reserved = num_blocks;
114         h->blocks_used = 0;
115         h->block_group = NULL;
116         h->alloc_exclude_nr = 0;
117         h->alloc_exclude_start = 0;
118         root->fs_info->running_transaction->use_count++;
119         mutex_unlock(&root->fs_info->trans_mutex);
120         return h;
121 }
122
123 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
124                           struct btrfs_root *root)
125 {
126         struct btrfs_transaction *cur_trans;
127
128         mutex_lock(&root->fs_info->trans_mutex);
129         cur_trans = root->fs_info->running_transaction;
130         WARN_ON(cur_trans != trans->transaction);
131         WARN_ON(cur_trans->num_writers < 1);
132         cur_trans->num_writers--;
133         if (waitqueue_active(&cur_trans->writer_wait))
134                 wake_up(&cur_trans->writer_wait);
135         put_transaction(cur_trans);
136         mutex_unlock(&root->fs_info->trans_mutex);
137         memset(trans, 0, sizeof(*trans));
138         kmem_cache_free(btrfs_trans_handle_cachep, trans);
139         return 0;
140 }
141
142
143 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
144                                      struct btrfs_root *root)
145 {
146         unsigned long gang[16];
147         int ret;
148         int i;
149         int err;
150         int werr = 0;
151         struct page *page;
152         struct radix_tree_root *dirty_pages;
153         struct inode *btree_inode = root->fs_info->btree_inode;
154
155         if (!trans || !trans->transaction) {
156                 return filemap_write_and_wait(btree_inode->i_mapping);
157         }
158         dirty_pages = &trans->transaction->dirty_pages;
159         while(1) {
160                 ret = find_first_radix_bit(dirty_pages, gang,
161                                            0, ARRAY_SIZE(gang));
162                 if (!ret)
163                         break;
164                 for (i = 0; i < ret; i++) {
165                         /* FIXME EIO */
166                         clear_radix_bit(dirty_pages, gang[i]);
167                         page = find_lock_page(btree_inode->i_mapping,
168                                               gang[i]);
169                         if (!page)
170                                 continue;
171                         if (PageWriteback(page)) {
172                                 if (PageDirty(page))
173                                         wait_on_page_writeback(page);
174                                 else {
175                                         unlock_page(page);
176                                         page_cache_release(page);
177                                         continue;
178                                 }
179                         }
180                         err = write_one_page(page, 0);
181                         if (err)
182                                 werr = err;
183                         page_cache_release(page);
184                 }
185         }
186         err = filemap_fdatawait(btree_inode->i_mapping);
187         if (err)
188                 werr = err;
189         return werr;
190 }
191
192 int btrfs_commit_tree_roots(struct btrfs_trans_handle *trans,
193                             struct btrfs_root *root)
194 {
195         int ret;
196         u64 old_extent_block;
197         struct btrfs_fs_info *fs_info = root->fs_info;
198         struct btrfs_root *tree_root = fs_info->tree_root;
199         struct btrfs_root *extent_root = fs_info->extent_root;
200
201         btrfs_write_dirty_block_groups(trans, extent_root);
202         while(1) {
203                 old_extent_block = btrfs_root_blocknr(&extent_root->root_item);
204                 if (old_extent_block == bh_blocknr(extent_root->node))
205                         break;
206                 btrfs_set_root_blocknr(&extent_root->root_item,
207                                        bh_blocknr(extent_root->node));
208                 ret = btrfs_update_root(trans, tree_root,
209                                         &extent_root->root_key,
210                                         &extent_root->root_item);
211                 BUG_ON(ret);
212                 btrfs_write_dirty_block_groups(trans, extent_root);
213         }
214         return 0;
215 }
216
217 static int wait_for_commit(struct btrfs_root *root,
218                            struct btrfs_transaction *commit)
219 {
220         DEFINE_WAIT(wait);
221         mutex_lock(&root->fs_info->trans_mutex);
222         while(!commit->commit_done) {
223                 prepare_to_wait(&commit->commit_wait, &wait,
224                                 TASK_UNINTERRUPTIBLE);
225                 if (commit->commit_done)
226                         break;
227                 mutex_unlock(&root->fs_info->trans_mutex);
228                 schedule();
229                 mutex_lock(&root->fs_info->trans_mutex);
230         }
231         mutex_unlock(&root->fs_info->trans_mutex);
232         finish_wait(&commit->commit_wait, &wait);
233         return 0;
234 }
235
236 struct dirty_root {
237         struct list_head list;
238         struct btrfs_root *root;
239 };
240
241 int btrfs_add_dead_root(struct btrfs_root *root, struct list_head *dead_list)
242 {
243         struct dirty_root *dirty;
244
245         dirty = kmalloc(sizeof(*dirty), GFP_NOFS);
246         if (!dirty)
247                 return -ENOMEM;
248         dirty->root = root;
249         list_add(&dirty->list, dead_list);
250         return 0;
251 }
252
253 static int add_dirty_roots(struct btrfs_trans_handle *trans,
254                            struct radix_tree_root *radix,
255                            struct list_head *list)
256 {
257         struct dirty_root *dirty;
258         struct btrfs_root *gang[8];
259         struct btrfs_root *root;
260         int i;
261         int ret;
262         int err = 0;
263         u32 refs;
264
265         while(1) {
266                 ret = radix_tree_gang_lookup_tag(radix, (void **)gang, 0,
267                                                  ARRAY_SIZE(gang),
268                                                  BTRFS_ROOT_TRANS_TAG);
269                 if (ret == 0)
270                         break;
271                 for (i = 0; i < ret; i++) {
272                         root = gang[i];
273                         radix_tree_tag_clear(radix,
274                                      (unsigned long)root->root_key.objectid,
275                                      BTRFS_ROOT_TRANS_TAG);
276                         if (root->commit_root == root->node) {
277                                 WARN_ON(bh_blocknr(root->node) !=
278                                         btrfs_root_blocknr(&root->root_item));
279                                 brelse(root->commit_root);
280                                 root->commit_root = NULL;
281                                 continue;
282                         }
283                         dirty = kmalloc(sizeof(*dirty), GFP_NOFS);
284                         BUG_ON(!dirty);
285                         dirty->root = kmalloc(sizeof(*dirty->root), GFP_NOFS);
286                         BUG_ON(!dirty->root);
287
288                         memset(&root->root_item.drop_progress, 0,
289                                sizeof(struct btrfs_disk_key));
290                         root->root_item.drop_level = 0;
291
292                         memcpy(dirty->root, root, sizeof(*root));
293                         dirty->root->node = root->commit_root;
294                         root->commit_root = NULL;
295
296                         root->root_key.offset = root->fs_info->generation;
297                         btrfs_set_root_blocknr(&root->root_item,
298                                                bh_blocknr(root->node));
299                         err = btrfs_insert_root(trans, root->fs_info->tree_root,
300                                                 &root->root_key,
301                                                 &root->root_item);
302                         if (err)
303                                 break;
304
305                         refs = btrfs_root_refs(&dirty->root->root_item);
306                         btrfs_set_root_refs(&dirty->root->root_item, refs - 1);
307                         err = btrfs_update_root(trans, root->fs_info->tree_root,
308                                                 &dirty->root->root_key,
309                                                 &dirty->root->root_item);
310
311                         BUG_ON(err);
312                         if (refs == 1) {
313                                 list_add(&dirty->list, list);
314                         } else {
315                                 WARN_ON(1);
316                                 kfree(dirty->root);
317                                 kfree(dirty);
318                         }
319                 }
320         }
321         return err;
322 }
323
324 int btrfs_defrag_root(struct btrfs_root *root, int cacheonly)
325 {
326         struct btrfs_fs_info *info = root->fs_info;
327         int ret;
328         struct btrfs_trans_handle *trans;
329
330         if (root->defrag_running)
331                 return 0;
332
333         trans = btrfs_start_transaction(root, 1);
334         while (1) {
335                 root->defrag_running = 1;
336                 ret = btrfs_defrag_leaves(trans, root, cacheonly);
337                 btrfs_end_transaction(trans, root);
338                 mutex_unlock(&info->fs_mutex);
339
340                 btrfs_btree_balance_dirty(root);
341                 cond_resched();
342
343                 mutex_lock(&info->fs_mutex);
344                 trans = btrfs_start_transaction(root, 1);
345                 if (ret != -EAGAIN)
346                         break;
347         }
348         root->defrag_running = 0;
349         radix_tree_tag_clear(&info->fs_roots_radix,
350                      (unsigned long)root->root_key.objectid,
351                      BTRFS_ROOT_DEFRAG_TAG);
352         btrfs_end_transaction(trans, root);
353         return 0;
354 }
355
356 int btrfs_defrag_dirty_roots(struct btrfs_fs_info *info)
357 {
358         struct btrfs_root *gang[1];
359         struct btrfs_root *root;
360         int i;
361         int ret;
362         int err = 0;
363         u64 last = 0;
364
365         while(1) {
366                 ret = radix_tree_gang_lookup_tag(&info->fs_roots_radix,
367                                                  (void **)gang, last,
368                                                  ARRAY_SIZE(gang),
369                                                  BTRFS_ROOT_DEFRAG_TAG);
370                 if (ret == 0)
371                         break;
372                 for (i = 0; i < ret; i++) {
373                         root = gang[i];
374                         last = root->root_key.objectid + 1;
375                         btrfs_defrag_root(root, 1);
376                 }
377         }
378         btrfs_defrag_root(info->extent_root, 1);
379         return err;
380 }
381
382 static int drop_dirty_roots(struct btrfs_root *tree_root,
383                             struct list_head *list)
384 {
385         struct dirty_root *dirty;
386         struct btrfs_trans_handle *trans;
387         int ret = 0;
388         int err;
389
390         while(!list_empty(list)) {
391                 mutex_lock(&tree_root->fs_info->fs_mutex);
392                 dirty = list_entry(list->next, struct dirty_root, list);
393                 list_del_init(&dirty->list);
394
395                 while(1) {
396                         trans = btrfs_start_transaction(tree_root, 1);
397                         ret = btrfs_drop_snapshot(trans, dirty->root);
398                         if (ret != -EAGAIN) {
399                                 break;
400                         }
401                         err = btrfs_update_root(trans,
402                                         tree_root,
403                                         &dirty->root->root_key,
404                                         &dirty->root->root_item);
405                         if (err)
406                                 ret = err;
407                         ret = btrfs_end_transaction(trans, tree_root);
408                         BUG_ON(ret);
409                         mutex_unlock(&tree_root->fs_info->fs_mutex);
410
411                         btrfs_btree_balance_dirty(tree_root);
412                         schedule();
413
414                         mutex_lock(&tree_root->fs_info->fs_mutex);
415                 }
416                 BUG_ON(ret);
417                 ret = btrfs_del_root(trans, tree_root, &dirty->root->root_key);
418                 if (ret)
419                         break;
420                 ret = btrfs_end_transaction(trans, tree_root);
421                 BUG_ON(ret);
422
423                 kfree(dirty->root);
424                 kfree(dirty);
425                 mutex_unlock(&tree_root->fs_info->fs_mutex);
426                 btrfs_btree_balance_dirty(tree_root);
427                 schedule();
428         }
429         return ret;
430 }
431
432 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
433                              struct btrfs_root *root)
434 {
435         unsigned long joined = 0;
436         unsigned long timeout = 1;
437         struct btrfs_transaction *cur_trans;
438         struct btrfs_transaction *prev_trans = NULL;
439         struct list_head dirty_fs_roots;
440         struct radix_tree_root pinned_copy;
441         DEFINE_WAIT(wait);
442         int ret;
443
444         init_bit_radix(&pinned_copy);
445         INIT_LIST_HEAD(&dirty_fs_roots);
446
447         mutex_lock(&root->fs_info->trans_mutex);
448         if (trans->transaction->in_commit) {
449                 cur_trans = trans->transaction;
450                 trans->transaction->use_count++;
451                 mutex_unlock(&root->fs_info->trans_mutex);
452                 btrfs_end_transaction(trans, root);
453
454                 mutex_unlock(&root->fs_info->fs_mutex);
455                 ret = wait_for_commit(root, cur_trans);
456                 BUG_ON(ret);
457
458                 mutex_lock(&root->fs_info->trans_mutex);
459                 put_transaction(cur_trans);
460                 mutex_unlock(&root->fs_info->trans_mutex);
461
462                 mutex_lock(&root->fs_info->fs_mutex);
463                 return 0;
464         }
465         trans->transaction->in_commit = 1;
466         cur_trans = trans->transaction;
467         if (cur_trans->list.prev != &root->fs_info->trans_list) {
468                 prev_trans = list_entry(cur_trans->list.prev,
469                                         struct btrfs_transaction, list);
470                 if (!prev_trans->commit_done) {
471                         prev_trans->use_count++;
472                         mutex_unlock(&root->fs_info->fs_mutex);
473                         mutex_unlock(&root->fs_info->trans_mutex);
474
475                         wait_for_commit(root, prev_trans);
476
477                         mutex_lock(&root->fs_info->fs_mutex);
478                         mutex_lock(&root->fs_info->trans_mutex);
479                         put_transaction(prev_trans);
480                 }
481         }
482
483         do {
484                 joined = cur_trans->num_joined;
485                 WARN_ON(cur_trans != trans->transaction);
486                 prepare_to_wait(&cur_trans->writer_wait, &wait,
487                                 TASK_UNINTERRUPTIBLE);
488
489                 if (cur_trans->num_writers > 1)
490                         timeout = MAX_SCHEDULE_TIMEOUT;
491                 else
492                         timeout = 1;
493
494                 mutex_unlock(&root->fs_info->fs_mutex);
495                 mutex_unlock(&root->fs_info->trans_mutex);
496
497                 schedule_timeout(timeout);
498
499                 mutex_lock(&root->fs_info->fs_mutex);
500                 mutex_lock(&root->fs_info->trans_mutex);
501                 finish_wait(&cur_trans->writer_wait, &wait);
502         } while (cur_trans->num_writers > 1 ||
503                  (cur_trans->num_joined != joined));
504
505         WARN_ON(cur_trans != trans->transaction);
506         ret = add_dirty_roots(trans, &root->fs_info->fs_roots_radix,
507                               &dirty_fs_roots);
508         BUG_ON(ret);
509
510         ret = btrfs_commit_tree_roots(trans, root);
511         BUG_ON(ret);
512
513         cur_trans = root->fs_info->running_transaction;
514         root->fs_info->running_transaction = NULL;
515         btrfs_set_super_generation(&root->fs_info->super_copy,
516                                    cur_trans->transid);
517         btrfs_set_super_root(&root->fs_info->super_copy,
518                              bh_blocknr(root->fs_info->tree_root->node));
519         memcpy(root->fs_info->disk_super, &root->fs_info->super_copy,
520                sizeof(root->fs_info->super_copy));
521
522         btrfs_copy_pinned(root, &pinned_copy);
523
524         mutex_unlock(&root->fs_info->trans_mutex);
525         mutex_unlock(&root->fs_info->fs_mutex);
526         ret = btrfs_write_and_wait_transaction(trans, root);
527         BUG_ON(ret);
528         write_ctree_super(trans, root);
529         mutex_lock(&root->fs_info->fs_mutex);
530         btrfs_finish_extent_commit(trans, root, &pinned_copy);
531         mutex_lock(&root->fs_info->trans_mutex);
532         cur_trans->commit_done = 1;
533         root->fs_info->last_trans_committed = cur_trans->transid;
534         wake_up(&cur_trans->commit_wait);
535         put_transaction(cur_trans);
536         put_transaction(cur_trans);
537         if (root->fs_info->closing)
538                 list_splice_init(&root->fs_info->dead_roots, &dirty_fs_roots);
539         else
540                 list_splice_init(&dirty_fs_roots, &root->fs_info->dead_roots);
541         mutex_unlock(&root->fs_info->trans_mutex);
542         kmem_cache_free(btrfs_trans_handle_cachep, trans);
543
544         if (root->fs_info->closing) {
545                 mutex_unlock(&root->fs_info->fs_mutex);
546                 drop_dirty_roots(root->fs_info->tree_root, &dirty_fs_roots);
547                 mutex_lock(&root->fs_info->fs_mutex);
548         }
549         return ret;
550 }
551
552 int btrfs_clean_old_snapshots(struct btrfs_root *root)
553 {
554         struct list_head dirty_roots;
555         INIT_LIST_HEAD(&dirty_roots);
556
557         mutex_lock(&root->fs_info->trans_mutex);
558         list_splice_init(&root->fs_info->dead_roots, &dirty_roots);
559         mutex_unlock(&root->fs_info->trans_mutex);
560
561         if (!list_empty(&dirty_roots)) {
562                 drop_dirty_roots(root, &dirty_roots);
563         }
564         return 0;
565 }
566 void btrfs_transaction_cleaner(struct work_struct *work)
567 {
568         struct btrfs_fs_info *fs_info = container_of(work,
569                                                      struct btrfs_fs_info,
570                                                      trans_work.work);
571
572         struct btrfs_root *root = fs_info->tree_root;
573         struct btrfs_transaction *cur;
574         struct btrfs_trans_handle *trans;
575         unsigned long now;
576         unsigned long delay = HZ * 30;
577         int ret;
578
579         mutex_lock(&root->fs_info->fs_mutex);
580         mutex_lock(&root->fs_info->trans_mutex);
581         cur = root->fs_info->running_transaction;
582         if (!cur) {
583                 mutex_unlock(&root->fs_info->trans_mutex);
584                 goto out;
585         }
586         now = get_seconds();
587         if (now < cur->start_time || now - cur->start_time < 30) {
588                 mutex_unlock(&root->fs_info->trans_mutex);
589                 delay = HZ * 5;
590                 goto out;
591         }
592         mutex_unlock(&root->fs_info->trans_mutex);
593         btrfs_defrag_dirty_roots(root->fs_info);
594         trans = btrfs_start_transaction(root, 1);
595         ret = btrfs_commit_transaction(trans, root);
596 out:
597         mutex_unlock(&root->fs_info->fs_mutex);
598         btrfs_clean_old_snapshots(root);
599         btrfs_transaction_queue_work(root, delay);
600 }
601
602 void btrfs_transaction_queue_work(struct btrfs_root *root, int delay)
603 {
604         queue_delayed_work(trans_wq, &root->fs_info->trans_work, delay);
605 }
606
607 void btrfs_transaction_flush_work(struct btrfs_root *root)
608 {
609         cancel_rearming_delayed_workqueue(trans_wq, &root->fs_info->trans_work);
610         flush_workqueue(trans_wq);
611 }
612
613 void __init btrfs_init_transaction_sys(void)
614 {
615         trans_wq = create_workqueue("btrfs");
616 }
617
618 void __exit btrfs_exit_transaction_sys(void)
619 {
620         destroy_workqueue(trans_wq);
621 }
622