Loading...
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#include "inode-map.h"
31
32#define BTRFS_ROOT_TRANS_TAG 0
33
34static noinline void put_transaction(struct btrfs_transaction *transaction)
35{
36 WARN_ON(atomic_read(&transaction->use_count) == 0);
37 if (atomic_dec_and_test(&transaction->use_count)) {
38 BUG_ON(!list_empty(&transaction->list));
39 memset(transaction, 0, sizeof(*transaction));
40 kmem_cache_free(btrfs_transaction_cachep, transaction);
41 }
42}
43
44static 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 */
53static noinline int join_transaction(struct btrfs_root *root, int nofail)
54{
55 struct btrfs_transaction *cur_trans;
56
57 spin_lock(&root->fs_info->trans_lock);
58 if (root->fs_info->trans_no_join) {
59 if (!nofail) {
60 spin_unlock(&root->fs_info->trans_lock);
61 return -EBUSY;
62 }
63 }
64
65 cur_trans = root->fs_info->running_transaction;
66 if (cur_trans) {
67 atomic_inc(&cur_trans->use_count);
68 atomic_inc(&cur_trans->num_writers);
69 cur_trans->num_joined++;
70 spin_unlock(&root->fs_info->trans_lock);
71 return 0;
72 }
73 spin_unlock(&root->fs_info->trans_lock);
74
75 cur_trans = kmem_cache_alloc(btrfs_transaction_cachep, GFP_NOFS);
76 if (!cur_trans)
77 return -ENOMEM;
78 spin_lock(&root->fs_info->trans_lock);
79 if (root->fs_info->running_transaction) {
80 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
81 cur_trans = root->fs_info->running_transaction;
82 atomic_inc(&cur_trans->use_count);
83 atomic_inc(&cur_trans->num_writers);
84 cur_trans->num_joined++;
85 spin_unlock(&root->fs_info->trans_lock);
86 return 0;
87 }
88 atomic_set(&cur_trans->num_writers, 1);
89 cur_trans->num_joined = 0;
90 init_waitqueue_head(&cur_trans->writer_wait);
91 init_waitqueue_head(&cur_trans->commit_wait);
92 cur_trans->in_commit = 0;
93 cur_trans->blocked = 0;
94 /*
95 * One for this trans handle, one so it will live on until we
96 * commit the transaction.
97 */
98 atomic_set(&cur_trans->use_count, 2);
99 cur_trans->commit_done = 0;
100 cur_trans->start_time = get_seconds();
101
102 cur_trans->delayed_refs.root = RB_ROOT;
103 cur_trans->delayed_refs.num_entries = 0;
104 cur_trans->delayed_refs.num_heads_ready = 0;
105 cur_trans->delayed_refs.num_heads = 0;
106 cur_trans->delayed_refs.flushing = 0;
107 cur_trans->delayed_refs.run_delayed_start = 0;
108 spin_lock_init(&cur_trans->commit_lock);
109 spin_lock_init(&cur_trans->delayed_refs.lock);
110
111 INIT_LIST_HEAD(&cur_trans->pending_snapshots);
112 list_add_tail(&cur_trans->list, &root->fs_info->trans_list);
113 extent_io_tree_init(&cur_trans->dirty_pages,
114 root->fs_info->btree_inode->i_mapping);
115 root->fs_info->generation++;
116 cur_trans->transid = root->fs_info->generation;
117 root->fs_info->running_transaction = cur_trans;
118 spin_unlock(&root->fs_info->trans_lock);
119
120 return 0;
121}
122
123/*
124 * this does all the record keeping required to make sure that a reference
125 * counted root is properly recorded in a given transaction. This is required
126 * to make sure the old root from before we joined the transaction is deleted
127 * when the transaction commits
128 */
129static int record_root_in_trans(struct btrfs_trans_handle *trans,
130 struct btrfs_root *root)
131{
132 if (root->ref_cows && root->last_trans < trans->transid) {
133 WARN_ON(root == root->fs_info->extent_root);
134 WARN_ON(root->commit_root != root->node);
135
136 /*
137 * see below for in_trans_setup usage rules
138 * we have the reloc mutex held now, so there
139 * is only one writer in this function
140 */
141 root->in_trans_setup = 1;
142
143 /* make sure readers find in_trans_setup before
144 * they find our root->last_trans update
145 */
146 smp_wmb();
147
148 spin_lock(&root->fs_info->fs_roots_radix_lock);
149 if (root->last_trans == trans->transid) {
150 spin_unlock(&root->fs_info->fs_roots_radix_lock);
151 return 0;
152 }
153 radix_tree_tag_set(&root->fs_info->fs_roots_radix,
154 (unsigned long)root->root_key.objectid,
155 BTRFS_ROOT_TRANS_TAG);
156 spin_unlock(&root->fs_info->fs_roots_radix_lock);
157 root->last_trans = trans->transid;
158
159 /* this is pretty tricky. We don't want to
160 * take the relocation lock in btrfs_record_root_in_trans
161 * unless we're really doing the first setup for this root in
162 * this transaction.
163 *
164 * Normally we'd use root->last_trans as a flag to decide
165 * if we want to take the expensive mutex.
166 *
167 * But, we have to set root->last_trans before we
168 * init the relocation root, otherwise, we trip over warnings
169 * in ctree.c. The solution used here is to flag ourselves
170 * with root->in_trans_setup. When this is 1, we're still
171 * fixing up the reloc trees and everyone must wait.
172 *
173 * When this is zero, they can trust root->last_trans and fly
174 * through btrfs_record_root_in_trans without having to take the
175 * lock. smp_wmb() makes sure that all the writes above are
176 * done before we pop in the zero below
177 */
178 btrfs_init_reloc_root(trans, root);
179 smp_wmb();
180 root->in_trans_setup = 0;
181 }
182 return 0;
183}
184
185
186int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
187 struct btrfs_root *root)
188{
189 if (!root->ref_cows)
190 return 0;
191
192 /*
193 * see record_root_in_trans for comments about in_trans_setup usage
194 * and barriers
195 */
196 smp_rmb();
197 if (root->last_trans == trans->transid &&
198 !root->in_trans_setup)
199 return 0;
200
201 mutex_lock(&root->fs_info->reloc_mutex);
202 record_root_in_trans(trans, root);
203 mutex_unlock(&root->fs_info->reloc_mutex);
204
205 return 0;
206}
207
208/* wait for commit against the current transaction to become unblocked
209 * when this is done, it is safe to start a new transaction, but the current
210 * transaction might not be fully on disk.
211 */
212static void wait_current_trans(struct btrfs_root *root)
213{
214 struct btrfs_transaction *cur_trans;
215
216 spin_lock(&root->fs_info->trans_lock);
217 cur_trans = root->fs_info->running_transaction;
218 if (cur_trans && cur_trans->blocked) {
219 atomic_inc(&cur_trans->use_count);
220 spin_unlock(&root->fs_info->trans_lock);
221
222 wait_event(root->fs_info->transaction_wait,
223 !cur_trans->blocked);
224 put_transaction(cur_trans);
225 } else {
226 spin_unlock(&root->fs_info->trans_lock);
227 }
228}
229
230enum btrfs_trans_type {
231 TRANS_START,
232 TRANS_JOIN,
233 TRANS_USERSPACE,
234 TRANS_JOIN_NOLOCK,
235};
236
237static int may_wait_transaction(struct btrfs_root *root, int type)
238{
239 if (root->fs_info->log_root_recovering)
240 return 0;
241
242 if (type == TRANS_USERSPACE)
243 return 1;
244
245 if (type == TRANS_START &&
246 !atomic_read(&root->fs_info->open_ioctl_trans))
247 return 1;
248
249 return 0;
250}
251
252static struct btrfs_trans_handle *start_transaction(struct btrfs_root *root,
253 u64 num_items, int type)
254{
255 struct btrfs_trans_handle *h;
256 struct btrfs_transaction *cur_trans;
257 u64 num_bytes = 0;
258 int ret;
259
260 if (root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR)
261 return ERR_PTR(-EROFS);
262
263 if (current->journal_info) {
264 WARN_ON(type != TRANS_JOIN && type != TRANS_JOIN_NOLOCK);
265 h = current->journal_info;
266 h->use_count++;
267 h->orig_rsv = h->block_rsv;
268 h->block_rsv = NULL;
269 goto got_it;
270 }
271
272 /*
273 * Do the reservation before we join the transaction so we can do all
274 * the appropriate flushing if need be.
275 */
276 if (num_items > 0 && root != root->fs_info->chunk_root) {
277 num_bytes = btrfs_calc_trans_metadata_size(root, num_items);
278 ret = btrfs_block_rsv_add(NULL, root,
279 &root->fs_info->trans_block_rsv,
280 num_bytes);
281 if (ret)
282 return ERR_PTR(ret);
283 }
284again:
285 h = kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
286 if (!h)
287 return ERR_PTR(-ENOMEM);
288
289 if (may_wait_transaction(root, type))
290 wait_current_trans(root);
291
292 do {
293 ret = join_transaction(root, type == TRANS_JOIN_NOLOCK);
294 if (ret == -EBUSY)
295 wait_current_trans(root);
296 } while (ret == -EBUSY);
297
298 if (ret < 0) {
299 kmem_cache_free(btrfs_trans_handle_cachep, h);
300 return ERR_PTR(ret);
301 }
302
303 cur_trans = root->fs_info->running_transaction;
304
305 h->transid = cur_trans->transid;
306 h->transaction = cur_trans;
307 h->blocks_used = 0;
308 h->bytes_reserved = 0;
309 h->delayed_ref_updates = 0;
310 h->use_count = 1;
311 h->block_rsv = NULL;
312 h->orig_rsv = NULL;
313
314 smp_mb();
315 if (cur_trans->blocked && may_wait_transaction(root, type)) {
316 btrfs_commit_transaction(h, root);
317 goto again;
318 }
319
320 if (num_bytes) {
321 h->block_rsv = &root->fs_info->trans_block_rsv;
322 h->bytes_reserved = num_bytes;
323 }
324
325got_it:
326 btrfs_record_root_in_trans(h, root);
327
328 if (!current->journal_info && type != TRANS_USERSPACE)
329 current->journal_info = h;
330 return h;
331}
332
333struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
334 int num_items)
335{
336 return start_transaction(root, num_items, TRANS_START);
337}
338struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
339{
340 return start_transaction(root, 0, TRANS_JOIN);
341}
342
343struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root)
344{
345 return start_transaction(root, 0, TRANS_JOIN_NOLOCK);
346}
347
348struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *root)
349{
350 return start_transaction(root, 0, TRANS_USERSPACE);
351}
352
353/* wait for a transaction commit to be fully complete */
354static noinline void wait_for_commit(struct btrfs_root *root,
355 struct btrfs_transaction *commit)
356{
357 wait_event(commit->commit_wait, commit->commit_done);
358}
359
360int btrfs_wait_for_commit(struct btrfs_root *root, u64 transid)
361{
362 struct btrfs_transaction *cur_trans = NULL, *t;
363 int ret;
364
365 ret = 0;
366 if (transid) {
367 if (transid <= root->fs_info->last_trans_committed)
368 goto out;
369
370 /* find specified transaction */
371 spin_lock(&root->fs_info->trans_lock);
372 list_for_each_entry(t, &root->fs_info->trans_list, list) {
373 if (t->transid == transid) {
374 cur_trans = t;
375 atomic_inc(&cur_trans->use_count);
376 break;
377 }
378 if (t->transid > transid)
379 break;
380 }
381 spin_unlock(&root->fs_info->trans_lock);
382 ret = -EINVAL;
383 if (!cur_trans)
384 goto out; /* bad transid */
385 } else {
386 /* find newest transaction that is committing | committed */
387 spin_lock(&root->fs_info->trans_lock);
388 list_for_each_entry_reverse(t, &root->fs_info->trans_list,
389 list) {
390 if (t->in_commit) {
391 if (t->commit_done)
392 break;
393 cur_trans = t;
394 atomic_inc(&cur_trans->use_count);
395 break;
396 }
397 }
398 spin_unlock(&root->fs_info->trans_lock);
399 if (!cur_trans)
400 goto out; /* nothing committing|committed */
401 }
402
403 wait_for_commit(root, cur_trans);
404
405 put_transaction(cur_trans);
406 ret = 0;
407out:
408 return ret;
409}
410
411void btrfs_throttle(struct btrfs_root *root)
412{
413 if (!atomic_read(&root->fs_info->open_ioctl_trans))
414 wait_current_trans(root);
415}
416
417static int should_end_transaction(struct btrfs_trans_handle *trans,
418 struct btrfs_root *root)
419{
420 int ret;
421 ret = btrfs_block_rsv_check(trans, root,
422 &root->fs_info->global_block_rsv, 0, 5);
423 return ret ? 1 : 0;
424}
425
426int btrfs_should_end_transaction(struct btrfs_trans_handle *trans,
427 struct btrfs_root *root)
428{
429 struct btrfs_transaction *cur_trans = trans->transaction;
430 int updates;
431
432 smp_mb();
433 if (cur_trans->blocked || cur_trans->delayed_refs.flushing)
434 return 1;
435
436 updates = trans->delayed_ref_updates;
437 trans->delayed_ref_updates = 0;
438 if (updates)
439 btrfs_run_delayed_refs(trans, root, updates);
440
441 return should_end_transaction(trans, root);
442}
443
444static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
445 struct btrfs_root *root, int throttle, int lock)
446{
447 struct btrfs_transaction *cur_trans = trans->transaction;
448 struct btrfs_fs_info *info = root->fs_info;
449 int count = 0;
450
451 if (--trans->use_count) {
452 trans->block_rsv = trans->orig_rsv;
453 return 0;
454 }
455
456 while (count < 4) {
457 unsigned long cur = trans->delayed_ref_updates;
458 trans->delayed_ref_updates = 0;
459 if (cur &&
460 trans->transaction->delayed_refs.num_heads_ready > 64) {
461 trans->delayed_ref_updates = 0;
462
463 /*
464 * do a full flush if the transaction is trying
465 * to close
466 */
467 if (trans->transaction->delayed_refs.flushing)
468 cur = 0;
469 btrfs_run_delayed_refs(trans, root, cur);
470 } else {
471 break;
472 }
473 count++;
474 }
475
476 btrfs_trans_release_metadata(trans, root);
477
478 if (lock && !atomic_read(&root->fs_info->open_ioctl_trans) &&
479 should_end_transaction(trans, root)) {
480 trans->transaction->blocked = 1;
481 smp_wmb();
482 }
483
484 if (lock && cur_trans->blocked && !cur_trans->in_commit) {
485 if (throttle) {
486 /*
487 * We may race with somebody else here so end up having
488 * to call end_transaction on ourselves again, so inc
489 * our use_count.
490 */
491 trans->use_count++;
492 return btrfs_commit_transaction(trans, root);
493 } else {
494 wake_up_process(info->transaction_kthread);
495 }
496 }
497
498 WARN_ON(cur_trans != info->running_transaction);
499 WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
500 atomic_dec(&cur_trans->num_writers);
501
502 smp_mb();
503 if (waitqueue_active(&cur_trans->writer_wait))
504 wake_up(&cur_trans->writer_wait);
505 put_transaction(cur_trans);
506
507 if (current->journal_info == trans)
508 current->journal_info = NULL;
509 memset(trans, 0, sizeof(*trans));
510 kmem_cache_free(btrfs_trans_handle_cachep, trans);
511
512 if (throttle)
513 btrfs_run_delayed_iputs(root);
514
515 return 0;
516}
517
518int btrfs_end_transaction(struct btrfs_trans_handle *trans,
519 struct btrfs_root *root)
520{
521 int ret;
522
523 ret = __btrfs_end_transaction(trans, root, 0, 1);
524 if (ret)
525 return ret;
526 return 0;
527}
528
529int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
530 struct btrfs_root *root)
531{
532 int ret;
533
534 ret = __btrfs_end_transaction(trans, root, 1, 1);
535 if (ret)
536 return ret;
537 return 0;
538}
539
540int btrfs_end_transaction_nolock(struct btrfs_trans_handle *trans,
541 struct btrfs_root *root)
542{
543 int ret;
544
545 ret = __btrfs_end_transaction(trans, root, 0, 0);
546 if (ret)
547 return ret;
548 return 0;
549}
550
551int btrfs_end_transaction_dmeta(struct btrfs_trans_handle *trans,
552 struct btrfs_root *root)
553{
554 return __btrfs_end_transaction(trans, root, 1, 1);
555}
556
557/*
558 * when btree blocks are allocated, they have some corresponding bits set for
559 * them in one of two extent_io trees. This is used to make sure all of
560 * those extents are sent to disk but does not wait on them
561 */
562int btrfs_write_marked_extents(struct btrfs_root *root,
563 struct extent_io_tree *dirty_pages, int mark)
564{
565 int ret;
566 int err = 0;
567 int werr = 0;
568 struct page *page;
569 struct inode *btree_inode = root->fs_info->btree_inode;
570 u64 start = 0;
571 u64 end;
572 unsigned long index;
573
574 while (1) {
575 ret = find_first_extent_bit(dirty_pages, start, &start, &end,
576 mark);
577 if (ret)
578 break;
579 while (start <= end) {
580 cond_resched();
581
582 index = start >> PAGE_CACHE_SHIFT;
583 start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
584 page = find_get_page(btree_inode->i_mapping, index);
585 if (!page)
586 continue;
587
588 btree_lock_page_hook(page);
589 if (!page->mapping) {
590 unlock_page(page);
591 page_cache_release(page);
592 continue;
593 }
594
595 if (PageWriteback(page)) {
596 if (PageDirty(page))
597 wait_on_page_writeback(page);
598 else {
599 unlock_page(page);
600 page_cache_release(page);
601 continue;
602 }
603 }
604 err = write_one_page(page, 0);
605 if (err)
606 werr = err;
607 page_cache_release(page);
608 }
609 }
610 if (err)
611 werr = err;
612 return werr;
613}
614
615/*
616 * when btree blocks are allocated, they have some corresponding bits set for
617 * them in one of two extent_io trees. This is used to make sure all of
618 * those extents are on disk for transaction or log commit. We wait
619 * on all the pages and clear them from the dirty pages state tree
620 */
621int btrfs_wait_marked_extents(struct btrfs_root *root,
622 struct extent_io_tree *dirty_pages, int mark)
623{
624 int ret;
625 int err = 0;
626 int werr = 0;
627 struct page *page;
628 struct inode *btree_inode = root->fs_info->btree_inode;
629 u64 start = 0;
630 u64 end;
631 unsigned long index;
632
633 while (1) {
634 ret = find_first_extent_bit(dirty_pages, start, &start, &end,
635 mark);
636 if (ret)
637 break;
638
639 clear_extent_bits(dirty_pages, start, end, mark, GFP_NOFS);
640 while (start <= end) {
641 index = start >> PAGE_CACHE_SHIFT;
642 start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
643 page = find_get_page(btree_inode->i_mapping, index);
644 if (!page)
645 continue;
646 if (PageDirty(page)) {
647 btree_lock_page_hook(page);
648 wait_on_page_writeback(page);
649 err = write_one_page(page, 0);
650 if (err)
651 werr = err;
652 }
653 wait_on_page_writeback(page);
654 page_cache_release(page);
655 cond_resched();
656 }
657 }
658 if (err)
659 werr = err;
660 return werr;
661}
662
663/*
664 * when btree blocks are allocated, they have some corresponding bits set for
665 * them in one of two extent_io trees. This is used to make sure all of
666 * those extents are on disk for transaction or log commit
667 */
668int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
669 struct extent_io_tree *dirty_pages, int mark)
670{
671 int ret;
672 int ret2;
673
674 ret = btrfs_write_marked_extents(root, dirty_pages, mark);
675 ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
676 return ret || ret2;
677}
678
679int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
680 struct btrfs_root *root)
681{
682 if (!trans || !trans->transaction) {
683 struct inode *btree_inode;
684 btree_inode = root->fs_info->btree_inode;
685 return filemap_write_and_wait(btree_inode->i_mapping);
686 }
687 return btrfs_write_and_wait_marked_extents(root,
688 &trans->transaction->dirty_pages,
689 EXTENT_DIRTY);
690}
691
692/*
693 * this is used to update the root pointer in the tree of tree roots.
694 *
695 * But, in the case of the extent allocation tree, updating the root
696 * pointer may allocate blocks which may change the root of the extent
697 * allocation tree.
698 *
699 * So, this loops and repeats and makes sure the cowonly root didn't
700 * change while the root pointer was being updated in the metadata.
701 */
702static int update_cowonly_root(struct btrfs_trans_handle *trans,
703 struct btrfs_root *root)
704{
705 int ret;
706 u64 old_root_bytenr;
707 u64 old_root_used;
708 struct btrfs_root *tree_root = root->fs_info->tree_root;
709
710 old_root_used = btrfs_root_used(&root->root_item);
711 btrfs_write_dirty_block_groups(trans, root);
712
713 while (1) {
714 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
715 if (old_root_bytenr == root->node->start &&
716 old_root_used == btrfs_root_used(&root->root_item))
717 break;
718
719 btrfs_set_root_node(&root->root_item, root->node);
720 ret = btrfs_update_root(trans, tree_root,
721 &root->root_key,
722 &root->root_item);
723 BUG_ON(ret);
724
725 old_root_used = btrfs_root_used(&root->root_item);
726 ret = btrfs_write_dirty_block_groups(trans, root);
727 BUG_ON(ret);
728 }
729
730 if (root != root->fs_info->extent_root)
731 switch_commit_root(root);
732
733 return 0;
734}
735
736/*
737 * update all the cowonly tree roots on disk
738 */
739static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
740 struct btrfs_root *root)
741{
742 struct btrfs_fs_info *fs_info = root->fs_info;
743 struct list_head *next;
744 struct extent_buffer *eb;
745 int ret;
746
747 ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
748 BUG_ON(ret);
749
750 eb = btrfs_lock_root_node(fs_info->tree_root);
751 btrfs_cow_block(trans, fs_info->tree_root, eb, NULL, 0, &eb);
752 btrfs_tree_unlock(eb);
753 free_extent_buffer(eb);
754
755 ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
756 BUG_ON(ret);
757
758 while (!list_empty(&fs_info->dirty_cowonly_roots)) {
759 next = fs_info->dirty_cowonly_roots.next;
760 list_del_init(next);
761 root = list_entry(next, struct btrfs_root, dirty_list);
762
763 update_cowonly_root(trans, root);
764 }
765
766 down_write(&fs_info->extent_commit_sem);
767 switch_commit_root(fs_info->extent_root);
768 up_write(&fs_info->extent_commit_sem);
769
770 return 0;
771}
772
773/*
774 * dead roots are old snapshots that need to be deleted. This allocates
775 * a dirty root struct and adds it into the list of dead roots that need to
776 * be deleted
777 */
778int btrfs_add_dead_root(struct btrfs_root *root)
779{
780 spin_lock(&root->fs_info->trans_lock);
781 list_add(&root->root_list, &root->fs_info->dead_roots);
782 spin_unlock(&root->fs_info->trans_lock);
783 return 0;
784}
785
786/*
787 * update all the cowonly tree roots on disk
788 */
789static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
790 struct btrfs_root *root)
791{
792 struct btrfs_root *gang[8];
793 struct btrfs_fs_info *fs_info = root->fs_info;
794 int i;
795 int ret;
796 int err = 0;
797
798 spin_lock(&fs_info->fs_roots_radix_lock);
799 while (1) {
800 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
801 (void **)gang, 0,
802 ARRAY_SIZE(gang),
803 BTRFS_ROOT_TRANS_TAG);
804 if (ret == 0)
805 break;
806 for (i = 0; i < ret; i++) {
807 root = gang[i];
808 radix_tree_tag_clear(&fs_info->fs_roots_radix,
809 (unsigned long)root->root_key.objectid,
810 BTRFS_ROOT_TRANS_TAG);
811 spin_unlock(&fs_info->fs_roots_radix_lock);
812
813 btrfs_free_log(trans, root);
814 btrfs_update_reloc_root(trans, root);
815 btrfs_orphan_commit_root(trans, root);
816
817 btrfs_save_ino_cache(root, trans);
818
819 if (root->commit_root != root->node) {
820 mutex_lock(&root->fs_commit_mutex);
821 switch_commit_root(root);
822 btrfs_unpin_free_ino(root);
823 mutex_unlock(&root->fs_commit_mutex);
824
825 btrfs_set_root_node(&root->root_item,
826 root->node);
827 }
828
829 err = btrfs_update_root(trans, fs_info->tree_root,
830 &root->root_key,
831 &root->root_item);
832 spin_lock(&fs_info->fs_roots_radix_lock);
833 if (err)
834 break;
835 }
836 }
837 spin_unlock(&fs_info->fs_roots_radix_lock);
838 return err;
839}
840
841/*
842 * defrag a given btree. If cacheonly == 1, this won't read from the disk,
843 * otherwise every leaf in the btree is read and defragged.
844 */
845int btrfs_defrag_root(struct btrfs_root *root, int cacheonly)
846{
847 struct btrfs_fs_info *info = root->fs_info;
848 struct btrfs_trans_handle *trans;
849 int ret;
850 unsigned long nr;
851
852 if (xchg(&root->defrag_running, 1))
853 return 0;
854
855 while (1) {
856 trans = btrfs_start_transaction(root, 0);
857 if (IS_ERR(trans))
858 return PTR_ERR(trans);
859
860 ret = btrfs_defrag_leaves(trans, root, cacheonly);
861
862 nr = trans->blocks_used;
863 btrfs_end_transaction(trans, root);
864 btrfs_btree_balance_dirty(info->tree_root, nr);
865 cond_resched();
866
867 if (btrfs_fs_closing(root->fs_info) || ret != -EAGAIN)
868 break;
869 }
870 root->defrag_running = 0;
871 return ret;
872}
873
874/*
875 * new snapshots need to be created at a very specific time in the
876 * transaction commit. This does the actual creation
877 */
878static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
879 struct btrfs_fs_info *fs_info,
880 struct btrfs_pending_snapshot *pending)
881{
882 struct btrfs_key key;
883 struct btrfs_root_item *new_root_item;
884 struct btrfs_root *tree_root = fs_info->tree_root;
885 struct btrfs_root *root = pending->root;
886 struct btrfs_root *parent_root;
887 struct btrfs_block_rsv *rsv;
888 struct inode *parent_inode;
889 struct dentry *parent;
890 struct dentry *dentry;
891 struct extent_buffer *tmp;
892 struct extent_buffer *old;
893 int ret;
894 u64 to_reserve = 0;
895 u64 index = 0;
896 u64 objectid;
897 u64 root_flags;
898
899 rsv = trans->block_rsv;
900
901 new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
902 if (!new_root_item) {
903 pending->error = -ENOMEM;
904 goto fail;
905 }
906
907 ret = btrfs_find_free_objectid(tree_root, &objectid);
908 if (ret) {
909 pending->error = ret;
910 goto fail;
911 }
912
913 btrfs_reloc_pre_snapshot(trans, pending, &to_reserve);
914 btrfs_orphan_pre_snapshot(trans, pending, &to_reserve);
915
916 if (to_reserve > 0) {
917 ret = btrfs_block_rsv_add(trans, root, &pending->block_rsv,
918 to_reserve);
919 if (ret) {
920 pending->error = ret;
921 goto fail;
922 }
923 }
924
925 key.objectid = objectid;
926 key.offset = (u64)-1;
927 key.type = BTRFS_ROOT_ITEM_KEY;
928
929 trans->block_rsv = &pending->block_rsv;
930
931 dentry = pending->dentry;
932 parent = dget_parent(dentry);
933 parent_inode = parent->d_inode;
934 parent_root = BTRFS_I(parent_inode)->root;
935 record_root_in_trans(trans, parent_root);
936
937 /*
938 * insert the directory item
939 */
940 ret = btrfs_set_inode_index(parent_inode, &index);
941 BUG_ON(ret);
942 ret = btrfs_insert_dir_item(trans, parent_root,
943 dentry->d_name.name, dentry->d_name.len,
944 parent_inode, &key,
945 BTRFS_FT_DIR, index);
946 BUG_ON(ret);
947
948 btrfs_i_size_write(parent_inode, parent_inode->i_size +
949 dentry->d_name.len * 2);
950 ret = btrfs_update_inode(trans, parent_root, parent_inode);
951 BUG_ON(ret);
952
953 /*
954 * pull in the delayed directory update
955 * and the delayed inode item
956 * otherwise we corrupt the FS during
957 * snapshot
958 */
959 ret = btrfs_run_delayed_items(trans, root);
960 BUG_ON(ret);
961
962 record_root_in_trans(trans, root);
963 btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
964 memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
965 btrfs_check_and_init_root_item(new_root_item);
966
967 root_flags = btrfs_root_flags(new_root_item);
968 if (pending->readonly)
969 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
970 else
971 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
972 btrfs_set_root_flags(new_root_item, root_flags);
973
974 old = btrfs_lock_root_node(root);
975 btrfs_cow_block(trans, root, old, NULL, 0, &old);
976 btrfs_set_lock_blocking(old);
977
978 btrfs_copy_root(trans, root, old, &tmp, objectid);
979 btrfs_tree_unlock(old);
980 free_extent_buffer(old);
981
982 btrfs_set_root_node(new_root_item, tmp);
983 /* record when the snapshot was created in key.offset */
984 key.offset = trans->transid;
985 ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
986 btrfs_tree_unlock(tmp);
987 free_extent_buffer(tmp);
988 BUG_ON(ret);
989
990 /*
991 * insert root back/forward references
992 */
993 ret = btrfs_add_root_ref(trans, tree_root, objectid,
994 parent_root->root_key.objectid,
995 btrfs_ino(parent_inode), index,
996 dentry->d_name.name, dentry->d_name.len);
997 BUG_ON(ret);
998 dput(parent);
999
1000 key.offset = (u64)-1;
1001 pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
1002 BUG_ON(IS_ERR(pending->snap));
1003
1004 btrfs_reloc_post_snapshot(trans, pending);
1005 btrfs_orphan_post_snapshot(trans, pending);
1006fail:
1007 kfree(new_root_item);
1008 trans->block_rsv = rsv;
1009 btrfs_block_rsv_release(root, &pending->block_rsv, (u64)-1);
1010 return 0;
1011}
1012
1013/*
1014 * create all the snapshots we've scheduled for creation
1015 */
1016static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
1017 struct btrfs_fs_info *fs_info)
1018{
1019 struct btrfs_pending_snapshot *pending;
1020 struct list_head *head = &trans->transaction->pending_snapshots;
1021 int ret;
1022
1023 list_for_each_entry(pending, head, list) {
1024 ret = create_pending_snapshot(trans, fs_info, pending);
1025 BUG_ON(ret);
1026 }
1027 return 0;
1028}
1029
1030static void update_super_roots(struct btrfs_root *root)
1031{
1032 struct btrfs_root_item *root_item;
1033 struct btrfs_super_block *super;
1034
1035 super = &root->fs_info->super_copy;
1036
1037 root_item = &root->fs_info->chunk_root->root_item;
1038 super->chunk_root = root_item->bytenr;
1039 super->chunk_root_generation = root_item->generation;
1040 super->chunk_root_level = root_item->level;
1041
1042 root_item = &root->fs_info->tree_root->root_item;
1043 super->root = root_item->bytenr;
1044 super->generation = root_item->generation;
1045 super->root_level = root_item->level;
1046 if (super->cache_generation != 0 || btrfs_test_opt(root, SPACE_CACHE))
1047 super->cache_generation = root_item->generation;
1048}
1049
1050int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1051{
1052 int ret = 0;
1053 spin_lock(&info->trans_lock);
1054 if (info->running_transaction)
1055 ret = info->running_transaction->in_commit;
1056 spin_unlock(&info->trans_lock);
1057 return ret;
1058}
1059
1060int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1061{
1062 int ret = 0;
1063 spin_lock(&info->trans_lock);
1064 if (info->running_transaction)
1065 ret = info->running_transaction->blocked;
1066 spin_unlock(&info->trans_lock);
1067 return ret;
1068}
1069
1070/*
1071 * wait for the current transaction commit to start and block subsequent
1072 * transaction joins
1073 */
1074static void wait_current_trans_commit_start(struct btrfs_root *root,
1075 struct btrfs_transaction *trans)
1076{
1077 wait_event(root->fs_info->transaction_blocked_wait, trans->in_commit);
1078}
1079
1080/*
1081 * wait for the current transaction to start and then become unblocked.
1082 * caller holds ref.
1083 */
1084static void wait_current_trans_commit_start_and_unblock(struct btrfs_root *root,
1085 struct btrfs_transaction *trans)
1086{
1087 wait_event(root->fs_info->transaction_wait,
1088 trans->commit_done || (trans->in_commit && !trans->blocked));
1089}
1090
1091/*
1092 * commit transactions asynchronously. once btrfs_commit_transaction_async
1093 * returns, any subsequent transaction will not be allowed to join.
1094 */
1095struct btrfs_async_commit {
1096 struct btrfs_trans_handle *newtrans;
1097 struct btrfs_root *root;
1098 struct delayed_work work;
1099};
1100
1101static void do_async_commit(struct work_struct *work)
1102{
1103 struct btrfs_async_commit *ac =
1104 container_of(work, struct btrfs_async_commit, work.work);
1105
1106 btrfs_commit_transaction(ac->newtrans, ac->root);
1107 kfree(ac);
1108}
1109
1110int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1111 struct btrfs_root *root,
1112 int wait_for_unblock)
1113{
1114 struct btrfs_async_commit *ac;
1115 struct btrfs_transaction *cur_trans;
1116
1117 ac = kmalloc(sizeof(*ac), GFP_NOFS);
1118 if (!ac)
1119 return -ENOMEM;
1120
1121 INIT_DELAYED_WORK(&ac->work, do_async_commit);
1122 ac->root = root;
1123 ac->newtrans = btrfs_join_transaction(root);
1124 if (IS_ERR(ac->newtrans)) {
1125 int err = PTR_ERR(ac->newtrans);
1126 kfree(ac);
1127 return err;
1128 }
1129
1130 /* take transaction reference */
1131 cur_trans = trans->transaction;
1132 atomic_inc(&cur_trans->use_count);
1133
1134 btrfs_end_transaction(trans, root);
1135 schedule_delayed_work(&ac->work, 0);
1136
1137 /* wait for transaction to start and unblock */
1138 if (wait_for_unblock)
1139 wait_current_trans_commit_start_and_unblock(root, cur_trans);
1140 else
1141 wait_current_trans_commit_start(root, cur_trans);
1142
1143 if (current->journal_info == trans)
1144 current->journal_info = NULL;
1145
1146 put_transaction(cur_trans);
1147 return 0;
1148}
1149
1150/*
1151 * btrfs_transaction state sequence:
1152 * in_commit = 0, blocked = 0 (initial)
1153 * in_commit = 1, blocked = 1
1154 * blocked = 0
1155 * commit_done = 1
1156 */
1157int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
1158 struct btrfs_root *root)
1159{
1160 unsigned long joined = 0;
1161 struct btrfs_transaction *cur_trans;
1162 struct btrfs_transaction *prev_trans = NULL;
1163 DEFINE_WAIT(wait);
1164 int ret;
1165 int should_grow = 0;
1166 unsigned long now = get_seconds();
1167 int flush_on_commit = btrfs_test_opt(root, FLUSHONCOMMIT);
1168
1169 btrfs_run_ordered_operations(root, 0);
1170
1171 /* make a pass through all the delayed refs we have so far
1172 * any runnings procs may add more while we are here
1173 */
1174 ret = btrfs_run_delayed_refs(trans, root, 0);
1175 BUG_ON(ret);
1176
1177 btrfs_trans_release_metadata(trans, root);
1178
1179 cur_trans = trans->transaction;
1180 /*
1181 * set the flushing flag so procs in this transaction have to
1182 * start sending their work down.
1183 */
1184 cur_trans->delayed_refs.flushing = 1;
1185
1186 ret = btrfs_run_delayed_refs(trans, root, 0);
1187 BUG_ON(ret);
1188
1189 spin_lock(&cur_trans->commit_lock);
1190 if (cur_trans->in_commit) {
1191 spin_unlock(&cur_trans->commit_lock);
1192 atomic_inc(&cur_trans->use_count);
1193 btrfs_end_transaction(trans, root);
1194
1195 wait_for_commit(root, cur_trans);
1196
1197 put_transaction(cur_trans);
1198
1199 return 0;
1200 }
1201
1202 trans->transaction->in_commit = 1;
1203 trans->transaction->blocked = 1;
1204 spin_unlock(&cur_trans->commit_lock);
1205 wake_up(&root->fs_info->transaction_blocked_wait);
1206
1207 spin_lock(&root->fs_info->trans_lock);
1208 if (cur_trans->list.prev != &root->fs_info->trans_list) {
1209 prev_trans = list_entry(cur_trans->list.prev,
1210 struct btrfs_transaction, list);
1211 if (!prev_trans->commit_done) {
1212 atomic_inc(&prev_trans->use_count);
1213 spin_unlock(&root->fs_info->trans_lock);
1214
1215 wait_for_commit(root, prev_trans);
1216
1217 put_transaction(prev_trans);
1218 } else {
1219 spin_unlock(&root->fs_info->trans_lock);
1220 }
1221 } else {
1222 spin_unlock(&root->fs_info->trans_lock);
1223 }
1224
1225 if (now < cur_trans->start_time || now - cur_trans->start_time < 1)
1226 should_grow = 1;
1227
1228 do {
1229 int snap_pending = 0;
1230
1231 joined = cur_trans->num_joined;
1232 if (!list_empty(&trans->transaction->pending_snapshots))
1233 snap_pending = 1;
1234
1235 WARN_ON(cur_trans != trans->transaction);
1236
1237 if (flush_on_commit || snap_pending) {
1238 btrfs_start_delalloc_inodes(root, 1);
1239 ret = btrfs_wait_ordered_extents(root, 0, 1);
1240 BUG_ON(ret);
1241 }
1242
1243 ret = btrfs_run_delayed_items(trans, root);
1244 BUG_ON(ret);
1245
1246 /*
1247 * rename don't use btrfs_join_transaction, so, once we
1248 * set the transaction to blocked above, we aren't going
1249 * to get any new ordered operations. We can safely run
1250 * it here and no for sure that nothing new will be added
1251 * to the list
1252 */
1253 btrfs_run_ordered_operations(root, 1);
1254
1255 prepare_to_wait(&cur_trans->writer_wait, &wait,
1256 TASK_UNINTERRUPTIBLE);
1257
1258 if (atomic_read(&cur_trans->num_writers) > 1)
1259 schedule_timeout(MAX_SCHEDULE_TIMEOUT);
1260 else if (should_grow)
1261 schedule_timeout(1);
1262
1263 finish_wait(&cur_trans->writer_wait, &wait);
1264 } while (atomic_read(&cur_trans->num_writers) > 1 ||
1265 (should_grow && cur_trans->num_joined != joined));
1266
1267 /*
1268 * Ok now we need to make sure to block out any other joins while we
1269 * commit the transaction. We could have started a join before setting
1270 * no_join so make sure to wait for num_writers to == 1 again.
1271 */
1272 spin_lock(&root->fs_info->trans_lock);
1273 root->fs_info->trans_no_join = 1;
1274 spin_unlock(&root->fs_info->trans_lock);
1275 wait_event(cur_trans->writer_wait,
1276 atomic_read(&cur_trans->num_writers) == 1);
1277
1278 /*
1279 * the reloc mutex makes sure that we stop
1280 * the balancing code from coming in and moving
1281 * extents around in the middle of the commit
1282 */
1283 mutex_lock(&root->fs_info->reloc_mutex);
1284
1285 ret = btrfs_run_delayed_items(trans, root);
1286 BUG_ON(ret);
1287
1288 ret = create_pending_snapshots(trans, root->fs_info);
1289 BUG_ON(ret);
1290
1291 ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1292 BUG_ON(ret);
1293
1294 /*
1295 * make sure none of the code above managed to slip in a
1296 * delayed item
1297 */
1298 btrfs_assert_delayed_root_empty(root);
1299
1300 WARN_ON(cur_trans != trans->transaction);
1301
1302 btrfs_scrub_pause(root);
1303 /* btrfs_commit_tree_roots is responsible for getting the
1304 * various roots consistent with each other. Every pointer
1305 * in the tree of tree roots has to point to the most up to date
1306 * root for every subvolume and other tree. So, we have to keep
1307 * the tree logging code from jumping in and changing any
1308 * of the trees.
1309 *
1310 * At this point in the commit, there can't be any tree-log
1311 * writers, but a little lower down we drop the trans mutex
1312 * and let new people in. By holding the tree_log_mutex
1313 * from now until after the super is written, we avoid races
1314 * with the tree-log code.
1315 */
1316 mutex_lock(&root->fs_info->tree_log_mutex);
1317
1318 ret = commit_fs_roots(trans, root);
1319 BUG_ON(ret);
1320
1321 /* commit_fs_roots gets rid of all the tree log roots, it is now
1322 * safe to free the root of tree log roots
1323 */
1324 btrfs_free_log_root_tree(trans, root->fs_info);
1325
1326 ret = commit_cowonly_roots(trans, root);
1327 BUG_ON(ret);
1328
1329 btrfs_prepare_extent_commit(trans, root);
1330
1331 cur_trans = root->fs_info->running_transaction;
1332
1333 btrfs_set_root_node(&root->fs_info->tree_root->root_item,
1334 root->fs_info->tree_root->node);
1335 switch_commit_root(root->fs_info->tree_root);
1336
1337 btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
1338 root->fs_info->chunk_root->node);
1339 switch_commit_root(root->fs_info->chunk_root);
1340
1341 update_super_roots(root);
1342
1343 if (!root->fs_info->log_root_recovering) {
1344 btrfs_set_super_log_root(&root->fs_info->super_copy, 0);
1345 btrfs_set_super_log_root_level(&root->fs_info->super_copy, 0);
1346 }
1347
1348 memcpy(&root->fs_info->super_for_commit, &root->fs_info->super_copy,
1349 sizeof(root->fs_info->super_copy));
1350
1351 trans->transaction->blocked = 0;
1352 spin_lock(&root->fs_info->trans_lock);
1353 root->fs_info->running_transaction = NULL;
1354 root->fs_info->trans_no_join = 0;
1355 spin_unlock(&root->fs_info->trans_lock);
1356 mutex_unlock(&root->fs_info->reloc_mutex);
1357
1358 wake_up(&root->fs_info->transaction_wait);
1359
1360 ret = btrfs_write_and_wait_transaction(trans, root);
1361 BUG_ON(ret);
1362 write_ctree_super(trans, root, 0);
1363
1364 /*
1365 * the super is written, we can safely allow the tree-loggers
1366 * to go about their business
1367 */
1368 mutex_unlock(&root->fs_info->tree_log_mutex);
1369
1370 btrfs_finish_extent_commit(trans, root);
1371
1372 cur_trans->commit_done = 1;
1373
1374 root->fs_info->last_trans_committed = cur_trans->transid;
1375
1376 wake_up(&cur_trans->commit_wait);
1377
1378 spin_lock(&root->fs_info->trans_lock);
1379 list_del_init(&cur_trans->list);
1380 spin_unlock(&root->fs_info->trans_lock);
1381
1382 put_transaction(cur_trans);
1383 put_transaction(cur_trans);
1384
1385 trace_btrfs_transaction_commit(root);
1386
1387 btrfs_scrub_continue(root);
1388
1389 if (current->journal_info == trans)
1390 current->journal_info = NULL;
1391
1392 kmem_cache_free(btrfs_trans_handle_cachep, trans);
1393
1394 if (current != root->fs_info->transaction_kthread)
1395 btrfs_run_delayed_iputs(root);
1396
1397 return ret;
1398}
1399
1400/*
1401 * interface function to delete all the snapshots we have scheduled for deletion
1402 */
1403int btrfs_clean_old_snapshots(struct btrfs_root *root)
1404{
1405 LIST_HEAD(list);
1406 struct btrfs_fs_info *fs_info = root->fs_info;
1407
1408 spin_lock(&fs_info->trans_lock);
1409 list_splice_init(&fs_info->dead_roots, &list);
1410 spin_unlock(&fs_info->trans_lock);
1411
1412 while (!list_empty(&list)) {
1413 root = list_entry(list.next, struct btrfs_root, root_list);
1414 list_del(&root->root_list);
1415
1416 btrfs_kill_all_delayed_nodes(root);
1417
1418 if (btrfs_header_backref_rev(root->node) <
1419 BTRFS_MIXED_BACKREF_REV)
1420 btrfs_drop_snapshot(root, NULL, 0);
1421 else
1422 btrfs_drop_snapshot(root, NULL, 1);
1423 }
1424 return 0;
1425}
1/*
2 * Copyright (C) 2007 Oracle. All rights reserved.
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public
6 * License v2 as published by the Free Software Foundation.
7 *
8 * This program is distributed in the hope that it will be useful,
9 * but WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
11 * General Public License for more details.
12 *
13 * You should have received a copy of the GNU General Public
14 * License along with this program; if not, write to the
15 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16 * Boston, MA 021110-1307, USA.
17 */
18
19#include <linux/fs.h>
20#include <linux/slab.h>
21#include <linux/sched.h>
22#include <linux/writeback.h>
23#include <linux/pagemap.h>
24#include <linux/blkdev.h>
25#include <linux/uuid.h>
26#include "ctree.h"
27#include "disk-io.h"
28#include "transaction.h"
29#include "locking.h"
30#include "tree-log.h"
31#include "inode-map.h"
32#include "volumes.h"
33#include "dev-replace.h"
34#include "qgroup.h"
35
36#define BTRFS_ROOT_TRANS_TAG 0
37
38static const unsigned int btrfs_blocked_trans_types[TRANS_STATE_MAX] = {
39 [TRANS_STATE_RUNNING] = 0U,
40 [TRANS_STATE_BLOCKED] = (__TRANS_USERSPACE |
41 __TRANS_START),
42 [TRANS_STATE_COMMIT_START] = (__TRANS_USERSPACE |
43 __TRANS_START |
44 __TRANS_ATTACH),
45 [TRANS_STATE_COMMIT_DOING] = (__TRANS_USERSPACE |
46 __TRANS_START |
47 __TRANS_ATTACH |
48 __TRANS_JOIN),
49 [TRANS_STATE_UNBLOCKED] = (__TRANS_USERSPACE |
50 __TRANS_START |
51 __TRANS_ATTACH |
52 __TRANS_JOIN |
53 __TRANS_JOIN_NOLOCK),
54 [TRANS_STATE_COMPLETED] = (__TRANS_USERSPACE |
55 __TRANS_START |
56 __TRANS_ATTACH |
57 __TRANS_JOIN |
58 __TRANS_JOIN_NOLOCK),
59};
60
61void btrfs_put_transaction(struct btrfs_transaction *transaction)
62{
63 WARN_ON(atomic_read(&transaction->use_count) == 0);
64 if (atomic_dec_and_test(&transaction->use_count)) {
65 BUG_ON(!list_empty(&transaction->list));
66 WARN_ON(!RB_EMPTY_ROOT(&transaction->delayed_refs.href_root));
67 if (transaction->delayed_refs.pending_csums)
68 printk(KERN_ERR "pending csums is %llu\n",
69 transaction->delayed_refs.pending_csums);
70 while (!list_empty(&transaction->pending_chunks)) {
71 struct extent_map *em;
72
73 em = list_first_entry(&transaction->pending_chunks,
74 struct extent_map, list);
75 list_del_init(&em->list);
76 free_extent_map(em);
77 }
78 /*
79 * If any block groups are found in ->deleted_bgs then it's
80 * because the transaction was aborted and a commit did not
81 * happen (things failed before writing the new superblock
82 * and calling btrfs_finish_extent_commit()), so we can not
83 * discard the physical locations of the block groups.
84 */
85 while (!list_empty(&transaction->deleted_bgs)) {
86 struct btrfs_block_group_cache *cache;
87
88 cache = list_first_entry(&transaction->deleted_bgs,
89 struct btrfs_block_group_cache,
90 bg_list);
91 list_del_init(&cache->bg_list);
92 btrfs_put_block_group_trimming(cache);
93 btrfs_put_block_group(cache);
94 }
95 kmem_cache_free(btrfs_transaction_cachep, transaction);
96 }
97}
98
99static void clear_btree_io_tree(struct extent_io_tree *tree)
100{
101 spin_lock(&tree->lock);
102 /*
103 * Do a single barrier for the waitqueue_active check here, the state
104 * of the waitqueue should not change once clear_btree_io_tree is
105 * called.
106 */
107 smp_mb();
108 while (!RB_EMPTY_ROOT(&tree->state)) {
109 struct rb_node *node;
110 struct extent_state *state;
111
112 node = rb_first(&tree->state);
113 state = rb_entry(node, struct extent_state, rb_node);
114 rb_erase(&state->rb_node, &tree->state);
115 RB_CLEAR_NODE(&state->rb_node);
116 /*
117 * btree io trees aren't supposed to have tasks waiting for
118 * changes in the flags of extent states ever.
119 */
120 ASSERT(!waitqueue_active(&state->wq));
121 free_extent_state(state);
122
123 cond_resched_lock(&tree->lock);
124 }
125 spin_unlock(&tree->lock);
126}
127
128static noinline void switch_commit_roots(struct btrfs_transaction *trans,
129 struct btrfs_fs_info *fs_info)
130{
131 struct btrfs_root *root, *tmp;
132
133 down_write(&fs_info->commit_root_sem);
134 list_for_each_entry_safe(root, tmp, &trans->switch_commits,
135 dirty_list) {
136 list_del_init(&root->dirty_list);
137 free_extent_buffer(root->commit_root);
138 root->commit_root = btrfs_root_node(root);
139 if (is_fstree(root->objectid))
140 btrfs_unpin_free_ino(root);
141 clear_btree_io_tree(&root->dirty_log_pages);
142 }
143
144 /* We can free old roots now. */
145 spin_lock(&trans->dropped_roots_lock);
146 while (!list_empty(&trans->dropped_roots)) {
147 root = list_first_entry(&trans->dropped_roots,
148 struct btrfs_root, root_list);
149 list_del_init(&root->root_list);
150 spin_unlock(&trans->dropped_roots_lock);
151 btrfs_drop_and_free_fs_root(fs_info, root);
152 spin_lock(&trans->dropped_roots_lock);
153 }
154 spin_unlock(&trans->dropped_roots_lock);
155 up_write(&fs_info->commit_root_sem);
156}
157
158static inline void extwriter_counter_inc(struct btrfs_transaction *trans,
159 unsigned int type)
160{
161 if (type & TRANS_EXTWRITERS)
162 atomic_inc(&trans->num_extwriters);
163}
164
165static inline void extwriter_counter_dec(struct btrfs_transaction *trans,
166 unsigned int type)
167{
168 if (type & TRANS_EXTWRITERS)
169 atomic_dec(&trans->num_extwriters);
170}
171
172static inline void extwriter_counter_init(struct btrfs_transaction *trans,
173 unsigned int type)
174{
175 atomic_set(&trans->num_extwriters, ((type & TRANS_EXTWRITERS) ? 1 : 0));
176}
177
178static inline int extwriter_counter_read(struct btrfs_transaction *trans)
179{
180 return atomic_read(&trans->num_extwriters);
181}
182
183/*
184 * either allocate a new transaction or hop into the existing one
185 */
186static noinline int join_transaction(struct btrfs_root *root, unsigned int type)
187{
188 struct btrfs_transaction *cur_trans;
189 struct btrfs_fs_info *fs_info = root->fs_info;
190
191 spin_lock(&fs_info->trans_lock);
192loop:
193 /* The file system has been taken offline. No new transactions. */
194 if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
195 spin_unlock(&fs_info->trans_lock);
196 return -EROFS;
197 }
198
199 cur_trans = fs_info->running_transaction;
200 if (cur_trans) {
201 if (cur_trans->aborted) {
202 spin_unlock(&fs_info->trans_lock);
203 return cur_trans->aborted;
204 }
205 if (btrfs_blocked_trans_types[cur_trans->state] & type) {
206 spin_unlock(&fs_info->trans_lock);
207 return -EBUSY;
208 }
209 atomic_inc(&cur_trans->use_count);
210 atomic_inc(&cur_trans->num_writers);
211 extwriter_counter_inc(cur_trans, type);
212 spin_unlock(&fs_info->trans_lock);
213 return 0;
214 }
215 spin_unlock(&fs_info->trans_lock);
216
217 /*
218 * If we are ATTACH, we just want to catch the current transaction,
219 * and commit it. If there is no transaction, just return ENOENT.
220 */
221 if (type == TRANS_ATTACH)
222 return -ENOENT;
223
224 /*
225 * JOIN_NOLOCK only happens during the transaction commit, so
226 * it is impossible that ->running_transaction is NULL
227 */
228 BUG_ON(type == TRANS_JOIN_NOLOCK);
229
230 cur_trans = kmem_cache_alloc(btrfs_transaction_cachep, GFP_NOFS);
231 if (!cur_trans)
232 return -ENOMEM;
233
234 spin_lock(&fs_info->trans_lock);
235 if (fs_info->running_transaction) {
236 /*
237 * someone started a transaction after we unlocked. Make sure
238 * to redo the checks above
239 */
240 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
241 goto loop;
242 } else if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
243 spin_unlock(&fs_info->trans_lock);
244 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
245 return -EROFS;
246 }
247
248 atomic_set(&cur_trans->num_writers, 1);
249 extwriter_counter_init(cur_trans, type);
250 init_waitqueue_head(&cur_trans->writer_wait);
251 init_waitqueue_head(&cur_trans->commit_wait);
252 init_waitqueue_head(&cur_trans->pending_wait);
253 cur_trans->state = TRANS_STATE_RUNNING;
254 /*
255 * One for this trans handle, one so it will live on until we
256 * commit the transaction.
257 */
258 atomic_set(&cur_trans->use_count, 2);
259 atomic_set(&cur_trans->pending_ordered, 0);
260 cur_trans->flags = 0;
261 cur_trans->start_time = get_seconds();
262
263 memset(&cur_trans->delayed_refs, 0, sizeof(cur_trans->delayed_refs));
264
265 cur_trans->delayed_refs.href_root = RB_ROOT;
266 cur_trans->delayed_refs.dirty_extent_root = RB_ROOT;
267 atomic_set(&cur_trans->delayed_refs.num_entries, 0);
268
269 /*
270 * although the tree mod log is per file system and not per transaction,
271 * the log must never go across transaction boundaries.
272 */
273 smp_mb();
274 if (!list_empty(&fs_info->tree_mod_seq_list))
275 WARN(1, KERN_ERR "BTRFS: tree_mod_seq_list not empty when "
276 "creating a fresh transaction\n");
277 if (!RB_EMPTY_ROOT(&fs_info->tree_mod_log))
278 WARN(1, KERN_ERR "BTRFS: tree_mod_log rb tree not empty when "
279 "creating a fresh transaction\n");
280 atomic64_set(&fs_info->tree_mod_seq, 0);
281
282 spin_lock_init(&cur_trans->delayed_refs.lock);
283
284 INIT_LIST_HEAD(&cur_trans->pending_snapshots);
285 INIT_LIST_HEAD(&cur_trans->pending_chunks);
286 INIT_LIST_HEAD(&cur_trans->switch_commits);
287 INIT_LIST_HEAD(&cur_trans->dirty_bgs);
288 INIT_LIST_HEAD(&cur_trans->io_bgs);
289 INIT_LIST_HEAD(&cur_trans->dropped_roots);
290 mutex_init(&cur_trans->cache_write_mutex);
291 cur_trans->num_dirty_bgs = 0;
292 spin_lock_init(&cur_trans->dirty_bgs_lock);
293 INIT_LIST_HEAD(&cur_trans->deleted_bgs);
294 spin_lock_init(&cur_trans->dropped_roots_lock);
295 list_add_tail(&cur_trans->list, &fs_info->trans_list);
296 extent_io_tree_init(&cur_trans->dirty_pages,
297 fs_info->btree_inode->i_mapping);
298 fs_info->generation++;
299 cur_trans->transid = fs_info->generation;
300 fs_info->running_transaction = cur_trans;
301 cur_trans->aborted = 0;
302 spin_unlock(&fs_info->trans_lock);
303
304 return 0;
305}
306
307/*
308 * this does all the record keeping required to make sure that a reference
309 * counted root is properly recorded in a given transaction. This is required
310 * to make sure the old root from before we joined the transaction is deleted
311 * when the transaction commits
312 */
313static int record_root_in_trans(struct btrfs_trans_handle *trans,
314 struct btrfs_root *root)
315{
316 if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
317 root->last_trans < trans->transid) {
318 WARN_ON(root == root->fs_info->extent_root);
319 WARN_ON(root->commit_root != root->node);
320
321 /*
322 * see below for IN_TRANS_SETUP usage rules
323 * we have the reloc mutex held now, so there
324 * is only one writer in this function
325 */
326 set_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
327
328 /* make sure readers find IN_TRANS_SETUP before
329 * they find our root->last_trans update
330 */
331 smp_wmb();
332
333 spin_lock(&root->fs_info->fs_roots_radix_lock);
334 if (root->last_trans == trans->transid) {
335 spin_unlock(&root->fs_info->fs_roots_radix_lock);
336 return 0;
337 }
338 radix_tree_tag_set(&root->fs_info->fs_roots_radix,
339 (unsigned long)root->root_key.objectid,
340 BTRFS_ROOT_TRANS_TAG);
341 spin_unlock(&root->fs_info->fs_roots_radix_lock);
342 root->last_trans = trans->transid;
343
344 /* this is pretty tricky. We don't want to
345 * take the relocation lock in btrfs_record_root_in_trans
346 * unless we're really doing the first setup for this root in
347 * this transaction.
348 *
349 * Normally we'd use root->last_trans as a flag to decide
350 * if we want to take the expensive mutex.
351 *
352 * But, we have to set root->last_trans before we
353 * init the relocation root, otherwise, we trip over warnings
354 * in ctree.c. The solution used here is to flag ourselves
355 * with root IN_TRANS_SETUP. When this is 1, we're still
356 * fixing up the reloc trees and everyone must wait.
357 *
358 * When this is zero, they can trust root->last_trans and fly
359 * through btrfs_record_root_in_trans without having to take the
360 * lock. smp_wmb() makes sure that all the writes above are
361 * done before we pop in the zero below
362 */
363 btrfs_init_reloc_root(trans, root);
364 smp_mb__before_atomic();
365 clear_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
366 }
367 return 0;
368}
369
370
371void btrfs_add_dropped_root(struct btrfs_trans_handle *trans,
372 struct btrfs_root *root)
373{
374 struct btrfs_transaction *cur_trans = trans->transaction;
375
376 /* Add ourselves to the transaction dropped list */
377 spin_lock(&cur_trans->dropped_roots_lock);
378 list_add_tail(&root->root_list, &cur_trans->dropped_roots);
379 spin_unlock(&cur_trans->dropped_roots_lock);
380
381 /* Make sure we don't try to update the root at commit time */
382 spin_lock(&root->fs_info->fs_roots_radix_lock);
383 radix_tree_tag_clear(&root->fs_info->fs_roots_radix,
384 (unsigned long)root->root_key.objectid,
385 BTRFS_ROOT_TRANS_TAG);
386 spin_unlock(&root->fs_info->fs_roots_radix_lock);
387}
388
389int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
390 struct btrfs_root *root)
391{
392 if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
393 return 0;
394
395 /*
396 * see record_root_in_trans for comments about IN_TRANS_SETUP usage
397 * and barriers
398 */
399 smp_rmb();
400 if (root->last_trans == trans->transid &&
401 !test_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state))
402 return 0;
403
404 mutex_lock(&root->fs_info->reloc_mutex);
405 record_root_in_trans(trans, root);
406 mutex_unlock(&root->fs_info->reloc_mutex);
407
408 return 0;
409}
410
411static inline int is_transaction_blocked(struct btrfs_transaction *trans)
412{
413 return (trans->state >= TRANS_STATE_BLOCKED &&
414 trans->state < TRANS_STATE_UNBLOCKED &&
415 !trans->aborted);
416}
417
418/* wait for commit against the current transaction to become unblocked
419 * when this is done, it is safe to start a new transaction, but the current
420 * transaction might not be fully on disk.
421 */
422static void wait_current_trans(struct btrfs_root *root)
423{
424 struct btrfs_transaction *cur_trans;
425
426 spin_lock(&root->fs_info->trans_lock);
427 cur_trans = root->fs_info->running_transaction;
428 if (cur_trans && is_transaction_blocked(cur_trans)) {
429 atomic_inc(&cur_trans->use_count);
430 spin_unlock(&root->fs_info->trans_lock);
431
432 wait_event(root->fs_info->transaction_wait,
433 cur_trans->state >= TRANS_STATE_UNBLOCKED ||
434 cur_trans->aborted);
435 btrfs_put_transaction(cur_trans);
436 } else {
437 spin_unlock(&root->fs_info->trans_lock);
438 }
439}
440
441static int may_wait_transaction(struct btrfs_root *root, int type)
442{
443 if (root->fs_info->log_root_recovering)
444 return 0;
445
446 if (type == TRANS_USERSPACE)
447 return 1;
448
449 if (type == TRANS_START &&
450 !atomic_read(&root->fs_info->open_ioctl_trans))
451 return 1;
452
453 return 0;
454}
455
456static inline bool need_reserve_reloc_root(struct btrfs_root *root)
457{
458 if (!root->fs_info->reloc_ctl ||
459 !test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
460 root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
461 root->reloc_root)
462 return false;
463
464 return true;
465}
466
467static struct btrfs_trans_handle *
468start_transaction(struct btrfs_root *root, unsigned int num_items,
469 unsigned int type, enum btrfs_reserve_flush_enum flush)
470{
471 struct btrfs_trans_handle *h;
472 struct btrfs_transaction *cur_trans;
473 u64 num_bytes = 0;
474 u64 qgroup_reserved = 0;
475 bool reloc_reserved = false;
476 int ret;
477
478 /* Send isn't supposed to start transactions. */
479 ASSERT(current->journal_info != BTRFS_SEND_TRANS_STUB);
480
481 if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
482 return ERR_PTR(-EROFS);
483
484 if (current->journal_info) {
485 WARN_ON(type & TRANS_EXTWRITERS);
486 h = current->journal_info;
487 h->use_count++;
488 WARN_ON(h->use_count > 2);
489 h->orig_rsv = h->block_rsv;
490 h->block_rsv = NULL;
491 goto got_it;
492 }
493
494 /*
495 * Do the reservation before we join the transaction so we can do all
496 * the appropriate flushing if need be.
497 */
498 if (num_items > 0 && root != root->fs_info->chunk_root) {
499 qgroup_reserved = num_items * root->nodesize;
500 ret = btrfs_qgroup_reserve_meta(root, qgroup_reserved);
501 if (ret)
502 return ERR_PTR(ret);
503
504 num_bytes = btrfs_calc_trans_metadata_size(root, num_items);
505 /*
506 * Do the reservation for the relocation root creation
507 */
508 if (need_reserve_reloc_root(root)) {
509 num_bytes += root->nodesize;
510 reloc_reserved = true;
511 }
512
513 ret = btrfs_block_rsv_add(root,
514 &root->fs_info->trans_block_rsv,
515 num_bytes, flush);
516 if (ret)
517 goto reserve_fail;
518 }
519again:
520 h = kmem_cache_zalloc(btrfs_trans_handle_cachep, GFP_NOFS);
521 if (!h) {
522 ret = -ENOMEM;
523 goto alloc_fail;
524 }
525
526 /*
527 * If we are JOIN_NOLOCK we're already committing a transaction and
528 * waiting on this guy, so we don't need to do the sb_start_intwrite
529 * because we're already holding a ref. We need this because we could
530 * have raced in and did an fsync() on a file which can kick a commit
531 * and then we deadlock with somebody doing a freeze.
532 *
533 * If we are ATTACH, it means we just want to catch the current
534 * transaction and commit it, so we needn't do sb_start_intwrite().
535 */
536 if (type & __TRANS_FREEZABLE)
537 sb_start_intwrite(root->fs_info->sb);
538
539 if (may_wait_transaction(root, type))
540 wait_current_trans(root);
541
542 do {
543 ret = join_transaction(root, type);
544 if (ret == -EBUSY) {
545 wait_current_trans(root);
546 if (unlikely(type == TRANS_ATTACH))
547 ret = -ENOENT;
548 }
549 } while (ret == -EBUSY);
550
551 if (ret < 0) {
552 /* We must get the transaction if we are JOIN_NOLOCK. */
553 BUG_ON(type == TRANS_JOIN_NOLOCK);
554 goto join_fail;
555 }
556
557 cur_trans = root->fs_info->running_transaction;
558
559 h->transid = cur_trans->transid;
560 h->transaction = cur_trans;
561 h->root = root;
562 h->use_count = 1;
563
564 h->type = type;
565 h->can_flush_pending_bgs = true;
566 INIT_LIST_HEAD(&h->qgroup_ref_list);
567 INIT_LIST_HEAD(&h->new_bgs);
568
569 smp_mb();
570 if (cur_trans->state >= TRANS_STATE_BLOCKED &&
571 may_wait_transaction(root, type)) {
572 current->journal_info = h;
573 btrfs_commit_transaction(h, root);
574 goto again;
575 }
576
577 if (num_bytes) {
578 trace_btrfs_space_reservation(root->fs_info, "transaction",
579 h->transid, num_bytes, 1);
580 h->block_rsv = &root->fs_info->trans_block_rsv;
581 h->bytes_reserved = num_bytes;
582 h->reloc_reserved = reloc_reserved;
583 }
584
585got_it:
586 btrfs_record_root_in_trans(h, root);
587
588 if (!current->journal_info && type != TRANS_USERSPACE)
589 current->journal_info = h;
590 return h;
591
592join_fail:
593 if (type & __TRANS_FREEZABLE)
594 sb_end_intwrite(root->fs_info->sb);
595 kmem_cache_free(btrfs_trans_handle_cachep, h);
596alloc_fail:
597 if (num_bytes)
598 btrfs_block_rsv_release(root, &root->fs_info->trans_block_rsv,
599 num_bytes);
600reserve_fail:
601 btrfs_qgroup_free_meta(root, qgroup_reserved);
602 return ERR_PTR(ret);
603}
604
605struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
606 unsigned int num_items)
607{
608 return start_transaction(root, num_items, TRANS_START,
609 BTRFS_RESERVE_FLUSH_ALL);
610}
611struct btrfs_trans_handle *btrfs_start_transaction_fallback_global_rsv(
612 struct btrfs_root *root,
613 unsigned int num_items,
614 int min_factor)
615{
616 struct btrfs_trans_handle *trans;
617 u64 num_bytes;
618 int ret;
619
620 trans = btrfs_start_transaction(root, num_items);
621 if (!IS_ERR(trans) || PTR_ERR(trans) != -ENOSPC)
622 return trans;
623
624 trans = btrfs_start_transaction(root, 0);
625 if (IS_ERR(trans))
626 return trans;
627
628 num_bytes = btrfs_calc_trans_metadata_size(root, num_items);
629 ret = btrfs_cond_migrate_bytes(root->fs_info,
630 &root->fs_info->trans_block_rsv,
631 num_bytes,
632 min_factor);
633 if (ret) {
634 btrfs_end_transaction(trans, root);
635 return ERR_PTR(ret);
636 }
637
638 trans->block_rsv = &root->fs_info->trans_block_rsv;
639 trans->bytes_reserved = num_bytes;
640 trace_btrfs_space_reservation(root->fs_info, "transaction",
641 trans->transid, num_bytes, 1);
642
643 return trans;
644}
645
646struct btrfs_trans_handle *btrfs_start_transaction_lflush(
647 struct btrfs_root *root,
648 unsigned int num_items)
649{
650 return start_transaction(root, num_items, TRANS_START,
651 BTRFS_RESERVE_FLUSH_LIMIT);
652}
653
654struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
655{
656 return start_transaction(root, 0, TRANS_JOIN,
657 BTRFS_RESERVE_NO_FLUSH);
658}
659
660struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root)
661{
662 return start_transaction(root, 0, TRANS_JOIN_NOLOCK,
663 BTRFS_RESERVE_NO_FLUSH);
664}
665
666struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *root)
667{
668 return start_transaction(root, 0, TRANS_USERSPACE,
669 BTRFS_RESERVE_NO_FLUSH);
670}
671
672/*
673 * btrfs_attach_transaction() - catch the running transaction
674 *
675 * It is used when we want to commit the current the transaction, but
676 * don't want to start a new one.
677 *
678 * Note: If this function return -ENOENT, it just means there is no
679 * running transaction. But it is possible that the inactive transaction
680 * is still in the memory, not fully on disk. If you hope there is no
681 * inactive transaction in the fs when -ENOENT is returned, you should
682 * invoke
683 * btrfs_attach_transaction_barrier()
684 */
685struct btrfs_trans_handle *btrfs_attach_transaction(struct btrfs_root *root)
686{
687 return start_transaction(root, 0, TRANS_ATTACH,
688 BTRFS_RESERVE_NO_FLUSH);
689}
690
691/*
692 * btrfs_attach_transaction_barrier() - catch the running transaction
693 *
694 * It is similar to the above function, the differentia is this one
695 * will wait for all the inactive transactions until they fully
696 * complete.
697 */
698struct btrfs_trans_handle *
699btrfs_attach_transaction_barrier(struct btrfs_root *root)
700{
701 struct btrfs_trans_handle *trans;
702
703 trans = start_transaction(root, 0, TRANS_ATTACH,
704 BTRFS_RESERVE_NO_FLUSH);
705 if (IS_ERR(trans) && PTR_ERR(trans) == -ENOENT)
706 btrfs_wait_for_commit(root, 0);
707
708 return trans;
709}
710
711/* wait for a transaction commit to be fully complete */
712static noinline void wait_for_commit(struct btrfs_root *root,
713 struct btrfs_transaction *commit)
714{
715 wait_event(commit->commit_wait, commit->state == TRANS_STATE_COMPLETED);
716}
717
718int btrfs_wait_for_commit(struct btrfs_root *root, u64 transid)
719{
720 struct btrfs_transaction *cur_trans = NULL, *t;
721 int ret = 0;
722
723 if (transid) {
724 if (transid <= root->fs_info->last_trans_committed)
725 goto out;
726
727 /* find specified transaction */
728 spin_lock(&root->fs_info->trans_lock);
729 list_for_each_entry(t, &root->fs_info->trans_list, list) {
730 if (t->transid == transid) {
731 cur_trans = t;
732 atomic_inc(&cur_trans->use_count);
733 ret = 0;
734 break;
735 }
736 if (t->transid > transid) {
737 ret = 0;
738 break;
739 }
740 }
741 spin_unlock(&root->fs_info->trans_lock);
742
743 /*
744 * The specified transaction doesn't exist, or we
745 * raced with btrfs_commit_transaction
746 */
747 if (!cur_trans) {
748 if (transid > root->fs_info->last_trans_committed)
749 ret = -EINVAL;
750 goto out;
751 }
752 } else {
753 /* find newest transaction that is committing | committed */
754 spin_lock(&root->fs_info->trans_lock);
755 list_for_each_entry_reverse(t, &root->fs_info->trans_list,
756 list) {
757 if (t->state >= TRANS_STATE_COMMIT_START) {
758 if (t->state == TRANS_STATE_COMPLETED)
759 break;
760 cur_trans = t;
761 atomic_inc(&cur_trans->use_count);
762 break;
763 }
764 }
765 spin_unlock(&root->fs_info->trans_lock);
766 if (!cur_trans)
767 goto out; /* nothing committing|committed */
768 }
769
770 wait_for_commit(root, cur_trans);
771 btrfs_put_transaction(cur_trans);
772out:
773 return ret;
774}
775
776void btrfs_throttle(struct btrfs_root *root)
777{
778 if (!atomic_read(&root->fs_info->open_ioctl_trans))
779 wait_current_trans(root);
780}
781
782static int should_end_transaction(struct btrfs_trans_handle *trans,
783 struct btrfs_root *root)
784{
785 if (root->fs_info->global_block_rsv.space_info->full &&
786 btrfs_check_space_for_delayed_refs(trans, root))
787 return 1;
788
789 return !!btrfs_block_rsv_check(root, &root->fs_info->global_block_rsv, 5);
790}
791
792int btrfs_should_end_transaction(struct btrfs_trans_handle *trans,
793 struct btrfs_root *root)
794{
795 struct btrfs_transaction *cur_trans = trans->transaction;
796 int updates;
797 int err;
798
799 smp_mb();
800 if (cur_trans->state >= TRANS_STATE_BLOCKED ||
801 cur_trans->delayed_refs.flushing)
802 return 1;
803
804 updates = trans->delayed_ref_updates;
805 trans->delayed_ref_updates = 0;
806 if (updates) {
807 err = btrfs_run_delayed_refs(trans, root, updates * 2);
808 if (err) /* Error code will also eval true */
809 return err;
810 }
811
812 return should_end_transaction(trans, root);
813}
814
815static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
816 struct btrfs_root *root, int throttle)
817{
818 struct btrfs_transaction *cur_trans = trans->transaction;
819 struct btrfs_fs_info *info = root->fs_info;
820 unsigned long cur = trans->delayed_ref_updates;
821 int lock = (trans->type != TRANS_JOIN_NOLOCK);
822 int err = 0;
823 int must_run_delayed_refs = 0;
824
825 if (trans->use_count > 1) {
826 trans->use_count--;
827 trans->block_rsv = trans->orig_rsv;
828 return 0;
829 }
830
831 btrfs_trans_release_metadata(trans, root);
832 trans->block_rsv = NULL;
833
834 if (!list_empty(&trans->new_bgs))
835 btrfs_create_pending_block_groups(trans, root);
836
837 trans->delayed_ref_updates = 0;
838 if (!trans->sync) {
839 must_run_delayed_refs =
840 btrfs_should_throttle_delayed_refs(trans, root);
841 cur = max_t(unsigned long, cur, 32);
842
843 /*
844 * don't make the caller wait if they are from a NOLOCK
845 * or ATTACH transaction, it will deadlock with commit
846 */
847 if (must_run_delayed_refs == 1 &&
848 (trans->type & (__TRANS_JOIN_NOLOCK | __TRANS_ATTACH)))
849 must_run_delayed_refs = 2;
850 }
851
852 btrfs_trans_release_metadata(trans, root);
853 trans->block_rsv = NULL;
854
855 if (!list_empty(&trans->new_bgs))
856 btrfs_create_pending_block_groups(trans, root);
857
858 btrfs_trans_release_chunk_metadata(trans);
859
860 if (lock && !atomic_read(&root->fs_info->open_ioctl_trans) &&
861 should_end_transaction(trans, root) &&
862 ACCESS_ONCE(cur_trans->state) == TRANS_STATE_RUNNING) {
863 spin_lock(&info->trans_lock);
864 if (cur_trans->state == TRANS_STATE_RUNNING)
865 cur_trans->state = TRANS_STATE_BLOCKED;
866 spin_unlock(&info->trans_lock);
867 }
868
869 if (lock && ACCESS_ONCE(cur_trans->state) == TRANS_STATE_BLOCKED) {
870 if (throttle)
871 return btrfs_commit_transaction(trans, root);
872 else
873 wake_up_process(info->transaction_kthread);
874 }
875
876 if (trans->type & __TRANS_FREEZABLE)
877 sb_end_intwrite(root->fs_info->sb);
878
879 WARN_ON(cur_trans != info->running_transaction);
880 WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
881 atomic_dec(&cur_trans->num_writers);
882 extwriter_counter_dec(cur_trans, trans->type);
883
884 /*
885 * Make sure counter is updated before we wake up waiters.
886 */
887 smp_mb();
888 if (waitqueue_active(&cur_trans->writer_wait))
889 wake_up(&cur_trans->writer_wait);
890 btrfs_put_transaction(cur_trans);
891
892 if (current->journal_info == trans)
893 current->journal_info = NULL;
894
895 if (throttle)
896 btrfs_run_delayed_iputs(root);
897
898 if (trans->aborted ||
899 test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state)) {
900 wake_up_process(info->transaction_kthread);
901 err = -EIO;
902 }
903 assert_qgroups_uptodate(trans);
904
905 kmem_cache_free(btrfs_trans_handle_cachep, trans);
906 if (must_run_delayed_refs) {
907 btrfs_async_run_delayed_refs(root, cur,
908 must_run_delayed_refs == 1);
909 }
910 return err;
911}
912
913int btrfs_end_transaction(struct btrfs_trans_handle *trans,
914 struct btrfs_root *root)
915{
916 return __btrfs_end_transaction(trans, root, 0);
917}
918
919int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
920 struct btrfs_root *root)
921{
922 return __btrfs_end_transaction(trans, root, 1);
923}
924
925/*
926 * when btree blocks are allocated, they have some corresponding bits set for
927 * them in one of two extent_io trees. This is used to make sure all of
928 * those extents are sent to disk but does not wait on them
929 */
930int btrfs_write_marked_extents(struct btrfs_root *root,
931 struct extent_io_tree *dirty_pages, int mark)
932{
933 int err = 0;
934 int werr = 0;
935 struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
936 struct extent_state *cached_state = NULL;
937 u64 start = 0;
938 u64 end;
939
940 while (!find_first_extent_bit(dirty_pages, start, &start, &end,
941 mark, &cached_state)) {
942 bool wait_writeback = false;
943
944 err = convert_extent_bit(dirty_pages, start, end,
945 EXTENT_NEED_WAIT,
946 mark, &cached_state, GFP_NOFS);
947 /*
948 * convert_extent_bit can return -ENOMEM, which is most of the
949 * time a temporary error. So when it happens, ignore the error
950 * and wait for writeback of this range to finish - because we
951 * failed to set the bit EXTENT_NEED_WAIT for the range, a call
952 * to btrfs_wait_marked_extents() would not know that writeback
953 * for this range started and therefore wouldn't wait for it to
954 * finish - we don't want to commit a superblock that points to
955 * btree nodes/leafs for which writeback hasn't finished yet
956 * (and without errors).
957 * We cleanup any entries left in the io tree when committing
958 * the transaction (through clear_btree_io_tree()).
959 */
960 if (err == -ENOMEM) {
961 err = 0;
962 wait_writeback = true;
963 }
964 if (!err)
965 err = filemap_fdatawrite_range(mapping, start, end);
966 if (err)
967 werr = err;
968 else if (wait_writeback)
969 werr = filemap_fdatawait_range(mapping, start, end);
970 free_extent_state(cached_state);
971 cached_state = NULL;
972 cond_resched();
973 start = end + 1;
974 }
975 return werr;
976}
977
978/*
979 * when btree blocks are allocated, they have some corresponding bits set for
980 * them in one of two extent_io trees. This is used to make sure all of
981 * those extents are on disk for transaction or log commit. We wait
982 * on all the pages and clear them from the dirty pages state tree
983 */
984int btrfs_wait_marked_extents(struct btrfs_root *root,
985 struct extent_io_tree *dirty_pages, int mark)
986{
987 int err = 0;
988 int werr = 0;
989 struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
990 struct extent_state *cached_state = NULL;
991 u64 start = 0;
992 u64 end;
993 struct btrfs_inode *btree_ino = BTRFS_I(root->fs_info->btree_inode);
994 bool errors = false;
995
996 while (!find_first_extent_bit(dirty_pages, start, &start, &end,
997 EXTENT_NEED_WAIT, &cached_state)) {
998 /*
999 * Ignore -ENOMEM errors returned by clear_extent_bit().
1000 * When committing the transaction, we'll remove any entries
1001 * left in the io tree. For a log commit, we don't remove them
1002 * after committing the log because the tree can be accessed
1003 * concurrently - we do it only at transaction commit time when
1004 * it's safe to do it (through clear_btree_io_tree()).
1005 */
1006 err = clear_extent_bit(dirty_pages, start, end,
1007 EXTENT_NEED_WAIT,
1008 0, 0, &cached_state, GFP_NOFS);
1009 if (err == -ENOMEM)
1010 err = 0;
1011 if (!err)
1012 err = filemap_fdatawait_range(mapping, start, end);
1013 if (err)
1014 werr = err;
1015 free_extent_state(cached_state);
1016 cached_state = NULL;
1017 cond_resched();
1018 start = end + 1;
1019 }
1020 if (err)
1021 werr = err;
1022
1023 if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
1024 if ((mark & EXTENT_DIRTY) &&
1025 test_and_clear_bit(BTRFS_INODE_BTREE_LOG1_ERR,
1026 &btree_ino->runtime_flags))
1027 errors = true;
1028
1029 if ((mark & EXTENT_NEW) &&
1030 test_and_clear_bit(BTRFS_INODE_BTREE_LOG2_ERR,
1031 &btree_ino->runtime_flags))
1032 errors = true;
1033 } else {
1034 if (test_and_clear_bit(BTRFS_INODE_BTREE_ERR,
1035 &btree_ino->runtime_flags))
1036 errors = true;
1037 }
1038
1039 if (errors && !werr)
1040 werr = -EIO;
1041
1042 return werr;
1043}
1044
1045/*
1046 * when btree blocks are allocated, they have some corresponding bits set for
1047 * them in one of two extent_io trees. This is used to make sure all of
1048 * those extents are on disk for transaction or log commit
1049 */
1050static int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
1051 struct extent_io_tree *dirty_pages, int mark)
1052{
1053 int ret;
1054 int ret2;
1055 struct blk_plug plug;
1056
1057 blk_start_plug(&plug);
1058 ret = btrfs_write_marked_extents(root, dirty_pages, mark);
1059 blk_finish_plug(&plug);
1060 ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
1061
1062 if (ret)
1063 return ret;
1064 if (ret2)
1065 return ret2;
1066 return 0;
1067}
1068
1069static int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
1070 struct btrfs_root *root)
1071{
1072 int ret;
1073
1074 ret = btrfs_write_and_wait_marked_extents(root,
1075 &trans->transaction->dirty_pages,
1076 EXTENT_DIRTY);
1077 clear_btree_io_tree(&trans->transaction->dirty_pages);
1078
1079 return ret;
1080}
1081
1082/*
1083 * this is used to update the root pointer in the tree of tree roots.
1084 *
1085 * But, in the case of the extent allocation tree, updating the root
1086 * pointer may allocate blocks which may change the root of the extent
1087 * allocation tree.
1088 *
1089 * So, this loops and repeats and makes sure the cowonly root didn't
1090 * change while the root pointer was being updated in the metadata.
1091 */
1092static int update_cowonly_root(struct btrfs_trans_handle *trans,
1093 struct btrfs_root *root)
1094{
1095 int ret;
1096 u64 old_root_bytenr;
1097 u64 old_root_used;
1098 struct btrfs_root *tree_root = root->fs_info->tree_root;
1099
1100 old_root_used = btrfs_root_used(&root->root_item);
1101
1102 while (1) {
1103 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
1104 if (old_root_bytenr == root->node->start &&
1105 old_root_used == btrfs_root_used(&root->root_item))
1106 break;
1107
1108 btrfs_set_root_node(&root->root_item, root->node);
1109 ret = btrfs_update_root(trans, tree_root,
1110 &root->root_key,
1111 &root->root_item);
1112 if (ret)
1113 return ret;
1114
1115 old_root_used = btrfs_root_used(&root->root_item);
1116 }
1117
1118 return 0;
1119}
1120
1121/*
1122 * update all the cowonly tree roots on disk
1123 *
1124 * The error handling in this function may not be obvious. Any of the
1125 * failures will cause the file system to go offline. We still need
1126 * to clean up the delayed refs.
1127 */
1128static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
1129 struct btrfs_root *root)
1130{
1131 struct btrfs_fs_info *fs_info = root->fs_info;
1132 struct list_head *dirty_bgs = &trans->transaction->dirty_bgs;
1133 struct list_head *io_bgs = &trans->transaction->io_bgs;
1134 struct list_head *next;
1135 struct extent_buffer *eb;
1136 int ret;
1137
1138 eb = btrfs_lock_root_node(fs_info->tree_root);
1139 ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL,
1140 0, &eb);
1141 btrfs_tree_unlock(eb);
1142 free_extent_buffer(eb);
1143
1144 if (ret)
1145 return ret;
1146
1147 ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1148 if (ret)
1149 return ret;
1150
1151 ret = btrfs_run_dev_stats(trans, root->fs_info);
1152 if (ret)
1153 return ret;
1154 ret = btrfs_run_dev_replace(trans, root->fs_info);
1155 if (ret)
1156 return ret;
1157 ret = btrfs_run_qgroups(trans, root->fs_info);
1158 if (ret)
1159 return ret;
1160
1161 ret = btrfs_setup_space_cache(trans, root);
1162 if (ret)
1163 return ret;
1164
1165 /* run_qgroups might have added some more refs */
1166 ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1167 if (ret)
1168 return ret;
1169again:
1170 while (!list_empty(&fs_info->dirty_cowonly_roots)) {
1171 next = fs_info->dirty_cowonly_roots.next;
1172 list_del_init(next);
1173 root = list_entry(next, struct btrfs_root, dirty_list);
1174 clear_bit(BTRFS_ROOT_DIRTY, &root->state);
1175
1176 if (root != fs_info->extent_root)
1177 list_add_tail(&root->dirty_list,
1178 &trans->transaction->switch_commits);
1179 ret = update_cowonly_root(trans, root);
1180 if (ret)
1181 return ret;
1182 ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1183 if (ret)
1184 return ret;
1185 }
1186
1187 while (!list_empty(dirty_bgs) || !list_empty(io_bgs)) {
1188 ret = btrfs_write_dirty_block_groups(trans, root);
1189 if (ret)
1190 return ret;
1191 ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1192 if (ret)
1193 return ret;
1194 }
1195
1196 if (!list_empty(&fs_info->dirty_cowonly_roots))
1197 goto again;
1198
1199 list_add_tail(&fs_info->extent_root->dirty_list,
1200 &trans->transaction->switch_commits);
1201 btrfs_after_dev_replace_commit(fs_info);
1202
1203 return 0;
1204}
1205
1206/*
1207 * dead roots are old snapshots that need to be deleted. This allocates
1208 * a dirty root struct and adds it into the list of dead roots that need to
1209 * be deleted
1210 */
1211void btrfs_add_dead_root(struct btrfs_root *root)
1212{
1213 spin_lock(&root->fs_info->trans_lock);
1214 if (list_empty(&root->root_list))
1215 list_add_tail(&root->root_list, &root->fs_info->dead_roots);
1216 spin_unlock(&root->fs_info->trans_lock);
1217}
1218
1219/*
1220 * update all the cowonly tree roots on disk
1221 */
1222static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
1223 struct btrfs_root *root)
1224{
1225 struct btrfs_root *gang[8];
1226 struct btrfs_fs_info *fs_info = root->fs_info;
1227 int i;
1228 int ret;
1229 int err = 0;
1230
1231 spin_lock(&fs_info->fs_roots_radix_lock);
1232 while (1) {
1233 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
1234 (void **)gang, 0,
1235 ARRAY_SIZE(gang),
1236 BTRFS_ROOT_TRANS_TAG);
1237 if (ret == 0)
1238 break;
1239 for (i = 0; i < ret; i++) {
1240 root = gang[i];
1241 radix_tree_tag_clear(&fs_info->fs_roots_radix,
1242 (unsigned long)root->root_key.objectid,
1243 BTRFS_ROOT_TRANS_TAG);
1244 spin_unlock(&fs_info->fs_roots_radix_lock);
1245
1246 btrfs_free_log(trans, root);
1247 btrfs_update_reloc_root(trans, root);
1248 btrfs_orphan_commit_root(trans, root);
1249
1250 btrfs_save_ino_cache(root, trans);
1251
1252 /* see comments in should_cow_block() */
1253 clear_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1254 smp_mb__after_atomic();
1255
1256 if (root->commit_root != root->node) {
1257 list_add_tail(&root->dirty_list,
1258 &trans->transaction->switch_commits);
1259 btrfs_set_root_node(&root->root_item,
1260 root->node);
1261 }
1262
1263 err = btrfs_update_root(trans, fs_info->tree_root,
1264 &root->root_key,
1265 &root->root_item);
1266 spin_lock(&fs_info->fs_roots_radix_lock);
1267 if (err)
1268 break;
1269 btrfs_qgroup_free_meta_all(root);
1270 }
1271 }
1272 spin_unlock(&fs_info->fs_roots_radix_lock);
1273 return err;
1274}
1275
1276/*
1277 * defrag a given btree.
1278 * Every leaf in the btree is read and defragged.
1279 */
1280int btrfs_defrag_root(struct btrfs_root *root)
1281{
1282 struct btrfs_fs_info *info = root->fs_info;
1283 struct btrfs_trans_handle *trans;
1284 int ret;
1285
1286 if (test_and_set_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state))
1287 return 0;
1288
1289 while (1) {
1290 trans = btrfs_start_transaction(root, 0);
1291 if (IS_ERR(trans))
1292 return PTR_ERR(trans);
1293
1294 ret = btrfs_defrag_leaves(trans, root);
1295
1296 btrfs_end_transaction(trans, root);
1297 btrfs_btree_balance_dirty(info->tree_root);
1298 cond_resched();
1299
1300 if (btrfs_fs_closing(root->fs_info) || ret != -EAGAIN)
1301 break;
1302
1303 if (btrfs_defrag_cancelled(root->fs_info)) {
1304 pr_debug("BTRFS: defrag_root cancelled\n");
1305 ret = -EAGAIN;
1306 break;
1307 }
1308 }
1309 clear_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state);
1310 return ret;
1311}
1312
1313/*
1314 * new snapshots need to be created at a very specific time in the
1315 * transaction commit. This does the actual creation.
1316 *
1317 * Note:
1318 * If the error which may affect the commitment of the current transaction
1319 * happens, we should return the error number. If the error which just affect
1320 * the creation of the pending snapshots, just return 0.
1321 */
1322static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
1323 struct btrfs_fs_info *fs_info,
1324 struct btrfs_pending_snapshot *pending)
1325{
1326 struct btrfs_key key;
1327 struct btrfs_root_item *new_root_item;
1328 struct btrfs_root *tree_root = fs_info->tree_root;
1329 struct btrfs_root *root = pending->root;
1330 struct btrfs_root *parent_root;
1331 struct btrfs_block_rsv *rsv;
1332 struct inode *parent_inode;
1333 struct btrfs_path *path;
1334 struct btrfs_dir_item *dir_item;
1335 struct dentry *dentry;
1336 struct extent_buffer *tmp;
1337 struct extent_buffer *old;
1338 struct timespec cur_time;
1339 int ret = 0;
1340 u64 to_reserve = 0;
1341 u64 index = 0;
1342 u64 objectid;
1343 u64 root_flags;
1344 uuid_le new_uuid;
1345
1346 ASSERT(pending->path);
1347 path = pending->path;
1348
1349 ASSERT(pending->root_item);
1350 new_root_item = pending->root_item;
1351
1352 pending->error = btrfs_find_free_objectid(tree_root, &objectid);
1353 if (pending->error)
1354 goto no_free_objectid;
1355
1356 /*
1357 * Make qgroup to skip current new snapshot's qgroupid, as it is
1358 * accounted by later btrfs_qgroup_inherit().
1359 */
1360 btrfs_set_skip_qgroup(trans, objectid);
1361
1362 btrfs_reloc_pre_snapshot(pending, &to_reserve);
1363
1364 if (to_reserve > 0) {
1365 pending->error = btrfs_block_rsv_add(root,
1366 &pending->block_rsv,
1367 to_reserve,
1368 BTRFS_RESERVE_NO_FLUSH);
1369 if (pending->error)
1370 goto clear_skip_qgroup;
1371 }
1372
1373 key.objectid = objectid;
1374 key.offset = (u64)-1;
1375 key.type = BTRFS_ROOT_ITEM_KEY;
1376
1377 rsv = trans->block_rsv;
1378 trans->block_rsv = &pending->block_rsv;
1379 trans->bytes_reserved = trans->block_rsv->reserved;
1380 trace_btrfs_space_reservation(root->fs_info, "transaction",
1381 trans->transid,
1382 trans->bytes_reserved, 1);
1383 dentry = pending->dentry;
1384 parent_inode = pending->dir;
1385 parent_root = BTRFS_I(parent_inode)->root;
1386 record_root_in_trans(trans, parent_root);
1387
1388 cur_time = current_fs_time(parent_inode->i_sb);
1389
1390 /*
1391 * insert the directory item
1392 */
1393 ret = btrfs_set_inode_index(parent_inode, &index);
1394 BUG_ON(ret); /* -ENOMEM */
1395
1396 /* check if there is a file/dir which has the same name. */
1397 dir_item = btrfs_lookup_dir_item(NULL, parent_root, path,
1398 btrfs_ino(parent_inode),
1399 dentry->d_name.name,
1400 dentry->d_name.len, 0);
1401 if (dir_item != NULL && !IS_ERR(dir_item)) {
1402 pending->error = -EEXIST;
1403 goto dir_item_existed;
1404 } else if (IS_ERR(dir_item)) {
1405 ret = PTR_ERR(dir_item);
1406 btrfs_abort_transaction(trans, root, ret);
1407 goto fail;
1408 }
1409 btrfs_release_path(path);
1410
1411 /*
1412 * pull in the delayed directory update
1413 * and the delayed inode item
1414 * otherwise we corrupt the FS during
1415 * snapshot
1416 */
1417 ret = btrfs_run_delayed_items(trans, root);
1418 if (ret) { /* Transaction aborted */
1419 btrfs_abort_transaction(trans, root, ret);
1420 goto fail;
1421 }
1422
1423 record_root_in_trans(trans, root);
1424 btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
1425 memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
1426 btrfs_check_and_init_root_item(new_root_item);
1427
1428 root_flags = btrfs_root_flags(new_root_item);
1429 if (pending->readonly)
1430 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
1431 else
1432 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
1433 btrfs_set_root_flags(new_root_item, root_flags);
1434
1435 btrfs_set_root_generation_v2(new_root_item,
1436 trans->transid);
1437 uuid_le_gen(&new_uuid);
1438 memcpy(new_root_item->uuid, new_uuid.b, BTRFS_UUID_SIZE);
1439 memcpy(new_root_item->parent_uuid, root->root_item.uuid,
1440 BTRFS_UUID_SIZE);
1441 if (!(root_flags & BTRFS_ROOT_SUBVOL_RDONLY)) {
1442 memset(new_root_item->received_uuid, 0,
1443 sizeof(new_root_item->received_uuid));
1444 memset(&new_root_item->stime, 0, sizeof(new_root_item->stime));
1445 memset(&new_root_item->rtime, 0, sizeof(new_root_item->rtime));
1446 btrfs_set_root_stransid(new_root_item, 0);
1447 btrfs_set_root_rtransid(new_root_item, 0);
1448 }
1449 btrfs_set_stack_timespec_sec(&new_root_item->otime, cur_time.tv_sec);
1450 btrfs_set_stack_timespec_nsec(&new_root_item->otime, cur_time.tv_nsec);
1451 btrfs_set_root_otransid(new_root_item, trans->transid);
1452
1453 old = btrfs_lock_root_node(root);
1454 ret = btrfs_cow_block(trans, root, old, NULL, 0, &old);
1455 if (ret) {
1456 btrfs_tree_unlock(old);
1457 free_extent_buffer(old);
1458 btrfs_abort_transaction(trans, root, ret);
1459 goto fail;
1460 }
1461
1462 btrfs_set_lock_blocking(old);
1463
1464 ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1465 /* clean up in any case */
1466 btrfs_tree_unlock(old);
1467 free_extent_buffer(old);
1468 if (ret) {
1469 btrfs_abort_transaction(trans, root, ret);
1470 goto fail;
1471 }
1472 /* see comments in should_cow_block() */
1473 set_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1474 smp_wmb();
1475
1476 btrfs_set_root_node(new_root_item, tmp);
1477 /* record when the snapshot was created in key.offset */
1478 key.offset = trans->transid;
1479 ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1480 btrfs_tree_unlock(tmp);
1481 free_extent_buffer(tmp);
1482 if (ret) {
1483 btrfs_abort_transaction(trans, root, ret);
1484 goto fail;
1485 }
1486
1487 /*
1488 * insert root back/forward references
1489 */
1490 ret = btrfs_add_root_ref(trans, tree_root, objectid,
1491 parent_root->root_key.objectid,
1492 btrfs_ino(parent_inode), index,
1493 dentry->d_name.name, dentry->d_name.len);
1494 if (ret) {
1495 btrfs_abort_transaction(trans, root, ret);
1496 goto fail;
1497 }
1498
1499 key.offset = (u64)-1;
1500 pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
1501 if (IS_ERR(pending->snap)) {
1502 ret = PTR_ERR(pending->snap);
1503 btrfs_abort_transaction(trans, root, ret);
1504 goto fail;
1505 }
1506
1507 ret = btrfs_reloc_post_snapshot(trans, pending);
1508 if (ret) {
1509 btrfs_abort_transaction(trans, root, ret);
1510 goto fail;
1511 }
1512
1513 ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1514 if (ret) {
1515 btrfs_abort_transaction(trans, root, ret);
1516 goto fail;
1517 }
1518
1519 ret = btrfs_insert_dir_item(trans, parent_root,
1520 dentry->d_name.name, dentry->d_name.len,
1521 parent_inode, &key,
1522 BTRFS_FT_DIR, index);
1523 /* We have check then name at the beginning, so it is impossible. */
1524 BUG_ON(ret == -EEXIST || ret == -EOVERFLOW);
1525 if (ret) {
1526 btrfs_abort_transaction(trans, root, ret);
1527 goto fail;
1528 }
1529
1530 btrfs_i_size_write(parent_inode, parent_inode->i_size +
1531 dentry->d_name.len * 2);
1532 parent_inode->i_mtime = parent_inode->i_ctime =
1533 current_fs_time(parent_inode->i_sb);
1534 ret = btrfs_update_inode_fallback(trans, parent_root, parent_inode);
1535 if (ret) {
1536 btrfs_abort_transaction(trans, root, ret);
1537 goto fail;
1538 }
1539 ret = btrfs_uuid_tree_add(trans, fs_info->uuid_root, new_uuid.b,
1540 BTRFS_UUID_KEY_SUBVOL, objectid);
1541 if (ret) {
1542 btrfs_abort_transaction(trans, root, ret);
1543 goto fail;
1544 }
1545 if (!btrfs_is_empty_uuid(new_root_item->received_uuid)) {
1546 ret = btrfs_uuid_tree_add(trans, fs_info->uuid_root,
1547 new_root_item->received_uuid,
1548 BTRFS_UUID_KEY_RECEIVED_SUBVOL,
1549 objectid);
1550 if (ret && ret != -EEXIST) {
1551 btrfs_abort_transaction(trans, root, ret);
1552 goto fail;
1553 }
1554 }
1555
1556 ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1557 if (ret) {
1558 btrfs_abort_transaction(trans, root, ret);
1559 goto fail;
1560 }
1561
1562 /*
1563 * account qgroup counters before qgroup_inherit()
1564 */
1565 ret = btrfs_qgroup_prepare_account_extents(trans, fs_info);
1566 if (ret)
1567 goto fail;
1568 ret = btrfs_qgroup_account_extents(trans, fs_info);
1569 if (ret)
1570 goto fail;
1571 ret = btrfs_qgroup_inherit(trans, fs_info,
1572 root->root_key.objectid,
1573 objectid, pending->inherit);
1574 if (ret) {
1575 btrfs_abort_transaction(trans, root, ret);
1576 goto fail;
1577 }
1578
1579fail:
1580 pending->error = ret;
1581dir_item_existed:
1582 trans->block_rsv = rsv;
1583 trans->bytes_reserved = 0;
1584clear_skip_qgroup:
1585 btrfs_clear_skip_qgroup(trans);
1586no_free_objectid:
1587 kfree(new_root_item);
1588 pending->root_item = NULL;
1589 btrfs_free_path(path);
1590 pending->path = NULL;
1591
1592 return ret;
1593}
1594
1595/*
1596 * create all the snapshots we've scheduled for creation
1597 */
1598static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
1599 struct btrfs_fs_info *fs_info)
1600{
1601 struct btrfs_pending_snapshot *pending, *next;
1602 struct list_head *head = &trans->transaction->pending_snapshots;
1603 int ret = 0;
1604
1605 list_for_each_entry_safe(pending, next, head, list) {
1606 list_del(&pending->list);
1607 ret = create_pending_snapshot(trans, fs_info, pending);
1608 if (ret)
1609 break;
1610 }
1611 return ret;
1612}
1613
1614static void update_super_roots(struct btrfs_root *root)
1615{
1616 struct btrfs_root_item *root_item;
1617 struct btrfs_super_block *super;
1618
1619 super = root->fs_info->super_copy;
1620
1621 root_item = &root->fs_info->chunk_root->root_item;
1622 super->chunk_root = root_item->bytenr;
1623 super->chunk_root_generation = root_item->generation;
1624 super->chunk_root_level = root_item->level;
1625
1626 root_item = &root->fs_info->tree_root->root_item;
1627 super->root = root_item->bytenr;
1628 super->generation = root_item->generation;
1629 super->root_level = root_item->level;
1630 if (btrfs_test_opt(root, SPACE_CACHE))
1631 super->cache_generation = root_item->generation;
1632 if (root->fs_info->update_uuid_tree_gen)
1633 super->uuid_tree_generation = root_item->generation;
1634}
1635
1636int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1637{
1638 struct btrfs_transaction *trans;
1639 int ret = 0;
1640
1641 spin_lock(&info->trans_lock);
1642 trans = info->running_transaction;
1643 if (trans)
1644 ret = (trans->state >= TRANS_STATE_COMMIT_START);
1645 spin_unlock(&info->trans_lock);
1646 return ret;
1647}
1648
1649int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1650{
1651 struct btrfs_transaction *trans;
1652 int ret = 0;
1653
1654 spin_lock(&info->trans_lock);
1655 trans = info->running_transaction;
1656 if (trans)
1657 ret = is_transaction_blocked(trans);
1658 spin_unlock(&info->trans_lock);
1659 return ret;
1660}
1661
1662/*
1663 * wait for the current transaction commit to start and block subsequent
1664 * transaction joins
1665 */
1666static void wait_current_trans_commit_start(struct btrfs_root *root,
1667 struct btrfs_transaction *trans)
1668{
1669 wait_event(root->fs_info->transaction_blocked_wait,
1670 trans->state >= TRANS_STATE_COMMIT_START ||
1671 trans->aborted);
1672}
1673
1674/*
1675 * wait for the current transaction to start and then become unblocked.
1676 * caller holds ref.
1677 */
1678static void wait_current_trans_commit_start_and_unblock(struct btrfs_root *root,
1679 struct btrfs_transaction *trans)
1680{
1681 wait_event(root->fs_info->transaction_wait,
1682 trans->state >= TRANS_STATE_UNBLOCKED ||
1683 trans->aborted);
1684}
1685
1686/*
1687 * commit transactions asynchronously. once btrfs_commit_transaction_async
1688 * returns, any subsequent transaction will not be allowed to join.
1689 */
1690struct btrfs_async_commit {
1691 struct btrfs_trans_handle *newtrans;
1692 struct btrfs_root *root;
1693 struct work_struct work;
1694};
1695
1696static void do_async_commit(struct work_struct *work)
1697{
1698 struct btrfs_async_commit *ac =
1699 container_of(work, struct btrfs_async_commit, work);
1700
1701 /*
1702 * We've got freeze protection passed with the transaction.
1703 * Tell lockdep about it.
1704 */
1705 if (ac->newtrans->type & __TRANS_FREEZABLE)
1706 __sb_writers_acquired(ac->root->fs_info->sb, SB_FREEZE_FS);
1707
1708 current->journal_info = ac->newtrans;
1709
1710 btrfs_commit_transaction(ac->newtrans, ac->root);
1711 kfree(ac);
1712}
1713
1714int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1715 struct btrfs_root *root,
1716 int wait_for_unblock)
1717{
1718 struct btrfs_async_commit *ac;
1719 struct btrfs_transaction *cur_trans;
1720
1721 ac = kmalloc(sizeof(*ac), GFP_NOFS);
1722 if (!ac)
1723 return -ENOMEM;
1724
1725 INIT_WORK(&ac->work, do_async_commit);
1726 ac->root = root;
1727 ac->newtrans = btrfs_join_transaction(root);
1728 if (IS_ERR(ac->newtrans)) {
1729 int err = PTR_ERR(ac->newtrans);
1730 kfree(ac);
1731 return err;
1732 }
1733
1734 /* take transaction reference */
1735 cur_trans = trans->transaction;
1736 atomic_inc(&cur_trans->use_count);
1737
1738 btrfs_end_transaction(trans, root);
1739
1740 /*
1741 * Tell lockdep we've released the freeze rwsem, since the
1742 * async commit thread will be the one to unlock it.
1743 */
1744 if (ac->newtrans->type & __TRANS_FREEZABLE)
1745 __sb_writers_release(root->fs_info->sb, SB_FREEZE_FS);
1746
1747 schedule_work(&ac->work);
1748
1749 /* wait for transaction to start and unblock */
1750 if (wait_for_unblock)
1751 wait_current_trans_commit_start_and_unblock(root, cur_trans);
1752 else
1753 wait_current_trans_commit_start(root, cur_trans);
1754
1755 if (current->journal_info == trans)
1756 current->journal_info = NULL;
1757
1758 btrfs_put_transaction(cur_trans);
1759 return 0;
1760}
1761
1762
1763static void cleanup_transaction(struct btrfs_trans_handle *trans,
1764 struct btrfs_root *root, int err)
1765{
1766 struct btrfs_transaction *cur_trans = trans->transaction;
1767 DEFINE_WAIT(wait);
1768
1769 WARN_ON(trans->use_count > 1);
1770
1771 btrfs_abort_transaction(trans, root, err);
1772
1773 spin_lock(&root->fs_info->trans_lock);
1774
1775 /*
1776 * If the transaction is removed from the list, it means this
1777 * transaction has been committed successfully, so it is impossible
1778 * to call the cleanup function.
1779 */
1780 BUG_ON(list_empty(&cur_trans->list));
1781
1782 list_del_init(&cur_trans->list);
1783 if (cur_trans == root->fs_info->running_transaction) {
1784 cur_trans->state = TRANS_STATE_COMMIT_DOING;
1785 spin_unlock(&root->fs_info->trans_lock);
1786 wait_event(cur_trans->writer_wait,
1787 atomic_read(&cur_trans->num_writers) == 1);
1788
1789 spin_lock(&root->fs_info->trans_lock);
1790 }
1791 spin_unlock(&root->fs_info->trans_lock);
1792
1793 btrfs_cleanup_one_transaction(trans->transaction, root);
1794
1795 spin_lock(&root->fs_info->trans_lock);
1796 if (cur_trans == root->fs_info->running_transaction)
1797 root->fs_info->running_transaction = NULL;
1798 spin_unlock(&root->fs_info->trans_lock);
1799
1800 if (trans->type & __TRANS_FREEZABLE)
1801 sb_end_intwrite(root->fs_info->sb);
1802 btrfs_put_transaction(cur_trans);
1803 btrfs_put_transaction(cur_trans);
1804
1805 trace_btrfs_transaction_commit(root);
1806
1807 if (current->journal_info == trans)
1808 current->journal_info = NULL;
1809 btrfs_scrub_cancel(root->fs_info);
1810
1811 kmem_cache_free(btrfs_trans_handle_cachep, trans);
1812}
1813
1814static inline int btrfs_start_delalloc_flush(struct btrfs_fs_info *fs_info)
1815{
1816 if (btrfs_test_opt(fs_info->tree_root, FLUSHONCOMMIT))
1817 return btrfs_start_delalloc_roots(fs_info, 1, -1);
1818 return 0;
1819}
1820
1821static inline void btrfs_wait_delalloc_flush(struct btrfs_fs_info *fs_info)
1822{
1823 if (btrfs_test_opt(fs_info->tree_root, FLUSHONCOMMIT))
1824 btrfs_wait_ordered_roots(fs_info, -1);
1825}
1826
1827static inline void
1828btrfs_wait_pending_ordered(struct btrfs_transaction *cur_trans)
1829{
1830 wait_event(cur_trans->pending_wait,
1831 atomic_read(&cur_trans->pending_ordered) == 0);
1832}
1833
1834int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
1835 struct btrfs_root *root)
1836{
1837 struct btrfs_transaction *cur_trans = trans->transaction;
1838 struct btrfs_transaction *prev_trans = NULL;
1839 struct btrfs_inode *btree_ino = BTRFS_I(root->fs_info->btree_inode);
1840 int ret;
1841
1842 /* Stop the commit early if ->aborted is set */
1843 if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
1844 ret = cur_trans->aborted;
1845 btrfs_end_transaction(trans, root);
1846 return ret;
1847 }
1848
1849 /* make a pass through all the delayed refs we have so far
1850 * any runnings procs may add more while we are here
1851 */
1852 ret = btrfs_run_delayed_refs(trans, root, 0);
1853 if (ret) {
1854 btrfs_end_transaction(trans, root);
1855 return ret;
1856 }
1857
1858 btrfs_trans_release_metadata(trans, root);
1859 trans->block_rsv = NULL;
1860
1861 cur_trans = trans->transaction;
1862
1863 /*
1864 * set the flushing flag so procs in this transaction have to
1865 * start sending their work down.
1866 */
1867 cur_trans->delayed_refs.flushing = 1;
1868 smp_wmb();
1869
1870 if (!list_empty(&trans->new_bgs))
1871 btrfs_create_pending_block_groups(trans, root);
1872
1873 ret = btrfs_run_delayed_refs(trans, root, 0);
1874 if (ret) {
1875 btrfs_end_transaction(trans, root);
1876 return ret;
1877 }
1878
1879 if (!test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &cur_trans->flags)) {
1880 int run_it = 0;
1881
1882 /* this mutex is also taken before trying to set
1883 * block groups readonly. We need to make sure
1884 * that nobody has set a block group readonly
1885 * after a extents from that block group have been
1886 * allocated for cache files. btrfs_set_block_group_ro
1887 * will wait for the transaction to commit if it
1888 * finds BTRFS_TRANS_DIRTY_BG_RUN set.
1889 *
1890 * The BTRFS_TRANS_DIRTY_BG_RUN flag is also used to make sure
1891 * only one process starts all the block group IO. It wouldn't
1892 * hurt to have more than one go through, but there's no
1893 * real advantage to it either.
1894 */
1895 mutex_lock(&root->fs_info->ro_block_group_mutex);
1896 if (!test_and_set_bit(BTRFS_TRANS_DIRTY_BG_RUN,
1897 &cur_trans->flags))
1898 run_it = 1;
1899 mutex_unlock(&root->fs_info->ro_block_group_mutex);
1900
1901 if (run_it)
1902 ret = btrfs_start_dirty_block_groups(trans, root);
1903 }
1904 if (ret) {
1905 btrfs_end_transaction(trans, root);
1906 return ret;
1907 }
1908
1909 spin_lock(&root->fs_info->trans_lock);
1910 if (cur_trans->state >= TRANS_STATE_COMMIT_START) {
1911 spin_unlock(&root->fs_info->trans_lock);
1912 atomic_inc(&cur_trans->use_count);
1913 ret = btrfs_end_transaction(trans, root);
1914
1915 wait_for_commit(root, cur_trans);
1916
1917 if (unlikely(cur_trans->aborted))
1918 ret = cur_trans->aborted;
1919
1920 btrfs_put_transaction(cur_trans);
1921
1922 return ret;
1923 }
1924
1925 cur_trans->state = TRANS_STATE_COMMIT_START;
1926 wake_up(&root->fs_info->transaction_blocked_wait);
1927
1928 if (cur_trans->list.prev != &root->fs_info->trans_list) {
1929 prev_trans = list_entry(cur_trans->list.prev,
1930 struct btrfs_transaction, list);
1931 if (prev_trans->state != TRANS_STATE_COMPLETED) {
1932 atomic_inc(&prev_trans->use_count);
1933 spin_unlock(&root->fs_info->trans_lock);
1934
1935 wait_for_commit(root, prev_trans);
1936 ret = prev_trans->aborted;
1937
1938 btrfs_put_transaction(prev_trans);
1939 if (ret)
1940 goto cleanup_transaction;
1941 } else {
1942 spin_unlock(&root->fs_info->trans_lock);
1943 }
1944 } else {
1945 spin_unlock(&root->fs_info->trans_lock);
1946 }
1947
1948 extwriter_counter_dec(cur_trans, trans->type);
1949
1950 ret = btrfs_start_delalloc_flush(root->fs_info);
1951 if (ret)
1952 goto cleanup_transaction;
1953
1954 ret = btrfs_run_delayed_items(trans, root);
1955 if (ret)
1956 goto cleanup_transaction;
1957
1958 wait_event(cur_trans->writer_wait,
1959 extwriter_counter_read(cur_trans) == 0);
1960
1961 /* some pending stuffs might be added after the previous flush. */
1962 ret = btrfs_run_delayed_items(trans, root);
1963 if (ret)
1964 goto cleanup_transaction;
1965
1966 btrfs_wait_delalloc_flush(root->fs_info);
1967
1968 btrfs_wait_pending_ordered(cur_trans);
1969
1970 btrfs_scrub_pause(root);
1971 /*
1972 * Ok now we need to make sure to block out any other joins while we
1973 * commit the transaction. We could have started a join before setting
1974 * COMMIT_DOING so make sure to wait for num_writers to == 1 again.
1975 */
1976 spin_lock(&root->fs_info->trans_lock);
1977 cur_trans->state = TRANS_STATE_COMMIT_DOING;
1978 spin_unlock(&root->fs_info->trans_lock);
1979 wait_event(cur_trans->writer_wait,
1980 atomic_read(&cur_trans->num_writers) == 1);
1981
1982 /* ->aborted might be set after the previous check, so check it */
1983 if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
1984 ret = cur_trans->aborted;
1985 goto scrub_continue;
1986 }
1987 /*
1988 * the reloc mutex makes sure that we stop
1989 * the balancing code from coming in and moving
1990 * extents around in the middle of the commit
1991 */
1992 mutex_lock(&root->fs_info->reloc_mutex);
1993
1994 /*
1995 * We needn't worry about the delayed items because we will
1996 * deal with them in create_pending_snapshot(), which is the
1997 * core function of the snapshot creation.
1998 */
1999 ret = create_pending_snapshots(trans, root->fs_info);
2000 if (ret) {
2001 mutex_unlock(&root->fs_info->reloc_mutex);
2002 goto scrub_continue;
2003 }
2004
2005 /*
2006 * We insert the dir indexes of the snapshots and update the inode
2007 * of the snapshots' parents after the snapshot creation, so there
2008 * are some delayed items which are not dealt with. Now deal with
2009 * them.
2010 *
2011 * We needn't worry that this operation will corrupt the snapshots,
2012 * because all the tree which are snapshoted will be forced to COW
2013 * the nodes and leaves.
2014 */
2015 ret = btrfs_run_delayed_items(trans, root);
2016 if (ret) {
2017 mutex_unlock(&root->fs_info->reloc_mutex);
2018 goto scrub_continue;
2019 }
2020
2021 ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
2022 if (ret) {
2023 mutex_unlock(&root->fs_info->reloc_mutex);
2024 goto scrub_continue;
2025 }
2026
2027 /* Reocrd old roots for later qgroup accounting */
2028 ret = btrfs_qgroup_prepare_account_extents(trans, root->fs_info);
2029 if (ret) {
2030 mutex_unlock(&root->fs_info->reloc_mutex);
2031 goto scrub_continue;
2032 }
2033
2034 /*
2035 * make sure none of the code above managed to slip in a
2036 * delayed item
2037 */
2038 btrfs_assert_delayed_root_empty(root);
2039
2040 WARN_ON(cur_trans != trans->transaction);
2041
2042 /* btrfs_commit_tree_roots is responsible for getting the
2043 * various roots consistent with each other. Every pointer
2044 * in the tree of tree roots has to point to the most up to date
2045 * root for every subvolume and other tree. So, we have to keep
2046 * the tree logging code from jumping in and changing any
2047 * of the trees.
2048 *
2049 * At this point in the commit, there can't be any tree-log
2050 * writers, but a little lower down we drop the trans mutex
2051 * and let new people in. By holding the tree_log_mutex
2052 * from now until after the super is written, we avoid races
2053 * with the tree-log code.
2054 */
2055 mutex_lock(&root->fs_info->tree_log_mutex);
2056
2057 ret = commit_fs_roots(trans, root);
2058 if (ret) {
2059 mutex_unlock(&root->fs_info->tree_log_mutex);
2060 mutex_unlock(&root->fs_info->reloc_mutex);
2061 goto scrub_continue;
2062 }
2063
2064 /*
2065 * Since the transaction is done, we can apply the pending changes
2066 * before the next transaction.
2067 */
2068 btrfs_apply_pending_changes(root->fs_info);
2069
2070 /* commit_fs_roots gets rid of all the tree log roots, it is now
2071 * safe to free the root of tree log roots
2072 */
2073 btrfs_free_log_root_tree(trans, root->fs_info);
2074
2075 /*
2076 * Since fs roots are all committed, we can get a quite accurate
2077 * new_roots. So let's do quota accounting.
2078 */
2079 ret = btrfs_qgroup_account_extents(trans, root->fs_info);
2080 if (ret < 0) {
2081 mutex_unlock(&root->fs_info->tree_log_mutex);
2082 mutex_unlock(&root->fs_info->reloc_mutex);
2083 goto scrub_continue;
2084 }
2085
2086 ret = commit_cowonly_roots(trans, root);
2087 if (ret) {
2088 mutex_unlock(&root->fs_info->tree_log_mutex);
2089 mutex_unlock(&root->fs_info->reloc_mutex);
2090 goto scrub_continue;
2091 }
2092
2093 /*
2094 * The tasks which save the space cache and inode cache may also
2095 * update ->aborted, check it.
2096 */
2097 if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
2098 ret = cur_trans->aborted;
2099 mutex_unlock(&root->fs_info->tree_log_mutex);
2100 mutex_unlock(&root->fs_info->reloc_mutex);
2101 goto scrub_continue;
2102 }
2103
2104 btrfs_prepare_extent_commit(trans, root);
2105
2106 cur_trans = root->fs_info->running_transaction;
2107
2108 btrfs_set_root_node(&root->fs_info->tree_root->root_item,
2109 root->fs_info->tree_root->node);
2110 list_add_tail(&root->fs_info->tree_root->dirty_list,
2111 &cur_trans->switch_commits);
2112
2113 btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
2114 root->fs_info->chunk_root->node);
2115 list_add_tail(&root->fs_info->chunk_root->dirty_list,
2116 &cur_trans->switch_commits);
2117
2118 switch_commit_roots(cur_trans, root->fs_info);
2119
2120 assert_qgroups_uptodate(trans);
2121 ASSERT(list_empty(&cur_trans->dirty_bgs));
2122 ASSERT(list_empty(&cur_trans->io_bgs));
2123 update_super_roots(root);
2124
2125 btrfs_set_super_log_root(root->fs_info->super_copy, 0);
2126 btrfs_set_super_log_root_level(root->fs_info->super_copy, 0);
2127 memcpy(root->fs_info->super_for_commit, root->fs_info->super_copy,
2128 sizeof(*root->fs_info->super_copy));
2129
2130 btrfs_update_commit_device_size(root->fs_info);
2131 btrfs_update_commit_device_bytes_used(root, cur_trans);
2132
2133 clear_bit(BTRFS_INODE_BTREE_LOG1_ERR, &btree_ino->runtime_flags);
2134 clear_bit(BTRFS_INODE_BTREE_LOG2_ERR, &btree_ino->runtime_flags);
2135
2136 btrfs_trans_release_chunk_metadata(trans);
2137
2138 spin_lock(&root->fs_info->trans_lock);
2139 cur_trans->state = TRANS_STATE_UNBLOCKED;
2140 root->fs_info->running_transaction = NULL;
2141 spin_unlock(&root->fs_info->trans_lock);
2142 mutex_unlock(&root->fs_info->reloc_mutex);
2143
2144 wake_up(&root->fs_info->transaction_wait);
2145
2146 ret = btrfs_write_and_wait_transaction(trans, root);
2147 if (ret) {
2148 btrfs_std_error(root->fs_info, ret,
2149 "Error while writing out transaction");
2150 mutex_unlock(&root->fs_info->tree_log_mutex);
2151 goto scrub_continue;
2152 }
2153
2154 ret = write_ctree_super(trans, root, 0);
2155 if (ret) {
2156 mutex_unlock(&root->fs_info->tree_log_mutex);
2157 goto scrub_continue;
2158 }
2159
2160 /*
2161 * the super is written, we can safely allow the tree-loggers
2162 * to go about their business
2163 */
2164 mutex_unlock(&root->fs_info->tree_log_mutex);
2165
2166 btrfs_finish_extent_commit(trans, root);
2167
2168 if (test_bit(BTRFS_TRANS_HAVE_FREE_BGS, &cur_trans->flags))
2169 btrfs_clear_space_info_full(root->fs_info);
2170
2171 root->fs_info->last_trans_committed = cur_trans->transid;
2172 /*
2173 * We needn't acquire the lock here because there is no other task
2174 * which can change it.
2175 */
2176 cur_trans->state = TRANS_STATE_COMPLETED;
2177 wake_up(&cur_trans->commit_wait);
2178
2179 spin_lock(&root->fs_info->trans_lock);
2180 list_del_init(&cur_trans->list);
2181 spin_unlock(&root->fs_info->trans_lock);
2182
2183 btrfs_put_transaction(cur_trans);
2184 btrfs_put_transaction(cur_trans);
2185
2186 if (trans->type & __TRANS_FREEZABLE)
2187 sb_end_intwrite(root->fs_info->sb);
2188
2189 trace_btrfs_transaction_commit(root);
2190
2191 btrfs_scrub_continue(root);
2192
2193 if (current->journal_info == trans)
2194 current->journal_info = NULL;
2195
2196 kmem_cache_free(btrfs_trans_handle_cachep, trans);
2197
2198 if (current != root->fs_info->transaction_kthread &&
2199 current != root->fs_info->cleaner_kthread)
2200 btrfs_run_delayed_iputs(root);
2201
2202 return ret;
2203
2204scrub_continue:
2205 btrfs_scrub_continue(root);
2206cleanup_transaction:
2207 btrfs_trans_release_metadata(trans, root);
2208 btrfs_trans_release_chunk_metadata(trans);
2209 trans->block_rsv = NULL;
2210 btrfs_warn(root->fs_info, "Skipping commit of aborted transaction.");
2211 if (current->journal_info == trans)
2212 current->journal_info = NULL;
2213 cleanup_transaction(trans, root, ret);
2214
2215 return ret;
2216}
2217
2218/*
2219 * return < 0 if error
2220 * 0 if there are no more dead_roots at the time of call
2221 * 1 there are more to be processed, call me again
2222 *
2223 * The return value indicates there are certainly more snapshots to delete, but
2224 * if there comes a new one during processing, it may return 0. We don't mind,
2225 * because btrfs_commit_super will poke cleaner thread and it will process it a
2226 * few seconds later.
2227 */
2228int btrfs_clean_one_deleted_snapshot(struct btrfs_root *root)
2229{
2230 int ret;
2231 struct btrfs_fs_info *fs_info = root->fs_info;
2232
2233 spin_lock(&fs_info->trans_lock);
2234 if (list_empty(&fs_info->dead_roots)) {
2235 spin_unlock(&fs_info->trans_lock);
2236 return 0;
2237 }
2238 root = list_first_entry(&fs_info->dead_roots,
2239 struct btrfs_root, root_list);
2240 list_del_init(&root->root_list);
2241 spin_unlock(&fs_info->trans_lock);
2242
2243 pr_debug("BTRFS: cleaner removing %llu\n", root->objectid);
2244
2245 btrfs_kill_all_delayed_nodes(root);
2246
2247 if (btrfs_header_backref_rev(root->node) <
2248 BTRFS_MIXED_BACKREF_REV)
2249 ret = btrfs_drop_snapshot(root, NULL, 0, 0);
2250 else
2251 ret = btrfs_drop_snapshot(root, NULL, 1, 0);
2252
2253 return (ret < 0) ? 0 : 1;
2254}
2255
2256void btrfs_apply_pending_changes(struct btrfs_fs_info *fs_info)
2257{
2258 unsigned long prev;
2259 unsigned long bit;
2260
2261 prev = xchg(&fs_info->pending_changes, 0);
2262 if (!prev)
2263 return;
2264
2265 bit = 1 << BTRFS_PENDING_SET_INODE_MAP_CACHE;
2266 if (prev & bit)
2267 btrfs_set_opt(fs_info->mount_opt, INODE_MAP_CACHE);
2268 prev &= ~bit;
2269
2270 bit = 1 << BTRFS_PENDING_CLEAR_INODE_MAP_CACHE;
2271 if (prev & bit)
2272 btrfs_clear_opt(fs_info->mount_opt, INODE_MAP_CACHE);
2273 prev &= ~bit;
2274
2275 bit = 1 << BTRFS_PENDING_COMMIT;
2276 if (prev & bit)
2277 btrfs_debug(fs_info, "pending commit done");
2278 prev &= ~bit;
2279
2280 if (prev)
2281 btrfs_warn(fs_info,
2282 "unknown pending changes left 0x%lx, ignoring", prev);
2283}