Merge tag 'rtc-5.14' of git://git.kernel.org/pub/scm/linux/kernel/git/abelloni/linux
[linux-2.6-microblaze.git] / fs / btrfs / transaction.c
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2007 Oracle.  All rights reserved.
4  */
5
6 #include <linux/fs.h>
7 #include <linux/slab.h>
8 #include <linux/sched.h>
9 #include <linux/writeback.h>
10 #include <linux/pagemap.h>
11 #include <linux/blkdev.h>
12 #include <linux/uuid.h>
13 #include "misc.h"
14 #include "ctree.h"
15 #include "disk-io.h"
16 #include "transaction.h"
17 #include "locking.h"
18 #include "tree-log.h"
19 #include "volumes.h"
20 #include "dev-replace.h"
21 #include "qgroup.h"
22 #include "block-group.h"
23 #include "space-info.h"
24 #include "zoned.h"
25
26 #define BTRFS_ROOT_TRANS_TAG 0
27
28 /*
29  * Transaction states and transitions
30  *
31  * No running transaction (fs tree blocks are not modified)
32  * |
33  * | To next stage:
34  * |  Call start_transaction() variants. Except btrfs_join_transaction_nostart().
35  * V
36  * Transaction N [[TRANS_STATE_RUNNING]]
37  * |
38  * | New trans handles can be attached to transaction N by calling all
39  * | start_transaction() variants.
40  * |
41  * | To next stage:
42  * |  Call btrfs_commit_transaction() on any trans handle attached to
43  * |  transaction N
44  * V
45  * Transaction N [[TRANS_STATE_COMMIT_START]]
46  * |
47  * | Will wait for previous running transaction to completely finish if there
48  * | is one
49  * |
50  * | Then one of the following happes:
51  * | - Wait for all other trans handle holders to release.
52  * |   The btrfs_commit_transaction() caller will do the commit work.
53  * | - Wait for current transaction to be committed by others.
54  * |   Other btrfs_commit_transaction() caller will do the commit work.
55  * |
56  * | At this stage, only btrfs_join_transaction*() variants can attach
57  * | to this running transaction.
58  * | All other variants will wait for current one to finish and attach to
59  * | transaction N+1.
60  * |
61  * | To next stage:
62  * |  Caller is chosen to commit transaction N, and all other trans handle
63  * |  haven been released.
64  * V
65  * Transaction N [[TRANS_STATE_COMMIT_DOING]]
66  * |
67  * | The heavy lifting transaction work is started.
68  * | From running delayed refs (modifying extent tree) to creating pending
69  * | snapshots, running qgroups.
70  * | In short, modify supporting trees to reflect modifications of subvolume
71  * | trees.
72  * |
73  * | At this stage, all start_transaction() calls will wait for this
74  * | transaction to finish and attach to transaction N+1.
75  * |
76  * | To next stage:
77  * |  Until all supporting trees are updated.
78  * V
79  * Transaction N [[TRANS_STATE_UNBLOCKED]]
80  * |                                                Transaction N+1
81  * | All needed trees are modified, thus we only    [[TRANS_STATE_RUNNING]]
82  * | need to write them back to disk and update     |
83  * | super blocks.                                  |
84  * |                                                |
85  * | At this stage, new transaction is allowed to   |
86  * | start.                                         |
87  * | All new start_transaction() calls will be      |
88  * | attached to transid N+1.                       |
89  * |                                                |
90  * | To next stage:                                 |
91  * |  Until all tree blocks are super blocks are    |
92  * |  written to block devices                      |
93  * V                                                |
94  * Transaction N [[TRANS_STATE_COMPLETED]]          V
95  *   All tree blocks and super blocks are written.  Transaction N+1
96  *   This transaction is finished and all its       [[TRANS_STATE_COMMIT_START]]
97  *   data structures will be cleaned up.            | Life goes on
98  */
99 static const unsigned int btrfs_blocked_trans_types[TRANS_STATE_MAX] = {
100         [TRANS_STATE_RUNNING]           = 0U,
101         [TRANS_STATE_COMMIT_START]      = (__TRANS_START | __TRANS_ATTACH),
102         [TRANS_STATE_COMMIT_DOING]      = (__TRANS_START |
103                                            __TRANS_ATTACH |
104                                            __TRANS_JOIN |
105                                            __TRANS_JOIN_NOSTART),
106         [TRANS_STATE_UNBLOCKED]         = (__TRANS_START |
107                                            __TRANS_ATTACH |
108                                            __TRANS_JOIN |
109                                            __TRANS_JOIN_NOLOCK |
110                                            __TRANS_JOIN_NOSTART),
111         [TRANS_STATE_SUPER_COMMITTED]   = (__TRANS_START |
112                                            __TRANS_ATTACH |
113                                            __TRANS_JOIN |
114                                            __TRANS_JOIN_NOLOCK |
115                                            __TRANS_JOIN_NOSTART),
116         [TRANS_STATE_COMPLETED]         = (__TRANS_START |
117                                            __TRANS_ATTACH |
118                                            __TRANS_JOIN |
119                                            __TRANS_JOIN_NOLOCK |
120                                            __TRANS_JOIN_NOSTART),
121 };
122
123 void btrfs_put_transaction(struct btrfs_transaction *transaction)
124 {
125         WARN_ON(refcount_read(&transaction->use_count) == 0);
126         if (refcount_dec_and_test(&transaction->use_count)) {
127                 BUG_ON(!list_empty(&transaction->list));
128                 WARN_ON(!RB_EMPTY_ROOT(
129                                 &transaction->delayed_refs.href_root.rb_root));
130                 WARN_ON(!RB_EMPTY_ROOT(
131                                 &transaction->delayed_refs.dirty_extent_root));
132                 if (transaction->delayed_refs.pending_csums)
133                         btrfs_err(transaction->fs_info,
134                                   "pending csums is %llu",
135                                   transaction->delayed_refs.pending_csums);
136                 /*
137                  * If any block groups are found in ->deleted_bgs then it's
138                  * because the transaction was aborted and a commit did not
139                  * happen (things failed before writing the new superblock
140                  * and calling btrfs_finish_extent_commit()), so we can not
141                  * discard the physical locations of the block groups.
142                  */
143                 while (!list_empty(&transaction->deleted_bgs)) {
144                         struct btrfs_block_group *cache;
145
146                         cache = list_first_entry(&transaction->deleted_bgs,
147                                                  struct btrfs_block_group,
148                                                  bg_list);
149                         list_del_init(&cache->bg_list);
150                         btrfs_unfreeze_block_group(cache);
151                         btrfs_put_block_group(cache);
152                 }
153                 WARN_ON(!list_empty(&transaction->dev_update_list));
154                 kfree(transaction);
155         }
156 }
157
158 static noinline void switch_commit_roots(struct btrfs_trans_handle *trans)
159 {
160         struct btrfs_transaction *cur_trans = trans->transaction;
161         struct btrfs_fs_info *fs_info = trans->fs_info;
162         struct btrfs_root *root, *tmp;
163         struct btrfs_caching_control *caching_ctl, *next;
164
165         down_write(&fs_info->commit_root_sem);
166         list_for_each_entry_safe(root, tmp, &cur_trans->switch_commits,
167                                  dirty_list) {
168                 list_del_init(&root->dirty_list);
169                 free_extent_buffer(root->commit_root);
170                 root->commit_root = btrfs_root_node(root);
171                 extent_io_tree_release(&root->dirty_log_pages);
172                 btrfs_qgroup_clean_swapped_blocks(root);
173         }
174
175         /* We can free old roots now. */
176         spin_lock(&cur_trans->dropped_roots_lock);
177         while (!list_empty(&cur_trans->dropped_roots)) {
178                 root = list_first_entry(&cur_trans->dropped_roots,
179                                         struct btrfs_root, root_list);
180                 list_del_init(&root->root_list);
181                 spin_unlock(&cur_trans->dropped_roots_lock);
182                 btrfs_free_log(trans, root);
183                 btrfs_drop_and_free_fs_root(fs_info, root);
184                 spin_lock(&cur_trans->dropped_roots_lock);
185         }
186         spin_unlock(&cur_trans->dropped_roots_lock);
187
188         /*
189          * We have to update the last_byte_to_unpin under the commit_root_sem,
190          * at the same time we swap out the commit roots.
191          *
192          * This is because we must have a real view of the last spot the caching
193          * kthreads were while caching.  Consider the following views of the
194          * extent tree for a block group
195          *
196          * commit root
197          * +----+----+----+----+----+----+----+
198          * |\\\\|    |\\\\|\\\\|    |\\\\|\\\\|
199          * +----+----+----+----+----+----+----+
200          * 0    1    2    3    4    5    6    7
201          *
202          * new commit root
203          * +----+----+----+----+----+----+----+
204          * |    |    |    |\\\\|    |    |\\\\|
205          * +----+----+----+----+----+----+----+
206          * 0    1    2    3    4    5    6    7
207          *
208          * If the cache_ctl->progress was at 3, then we are only allowed to
209          * unpin [0,1) and [2,3], because the caching thread has already
210          * processed those extents.  We are not allowed to unpin [5,6), because
211          * the caching thread will re-start it's search from 3, and thus find
212          * the hole from [4,6) to add to the free space cache.
213          */
214         spin_lock(&fs_info->block_group_cache_lock);
215         list_for_each_entry_safe(caching_ctl, next,
216                                  &fs_info->caching_block_groups, list) {
217                 struct btrfs_block_group *cache = caching_ctl->block_group;
218
219                 if (btrfs_block_group_done(cache)) {
220                         cache->last_byte_to_unpin = (u64)-1;
221                         list_del_init(&caching_ctl->list);
222                         btrfs_put_caching_control(caching_ctl);
223                 } else {
224                         cache->last_byte_to_unpin = caching_ctl->progress;
225                 }
226         }
227         spin_unlock(&fs_info->block_group_cache_lock);
228         up_write(&fs_info->commit_root_sem);
229 }
230
231 static inline void extwriter_counter_inc(struct btrfs_transaction *trans,
232                                          unsigned int type)
233 {
234         if (type & TRANS_EXTWRITERS)
235                 atomic_inc(&trans->num_extwriters);
236 }
237
238 static inline void extwriter_counter_dec(struct btrfs_transaction *trans,
239                                          unsigned int type)
240 {
241         if (type & TRANS_EXTWRITERS)
242                 atomic_dec(&trans->num_extwriters);
243 }
244
245 static inline void extwriter_counter_init(struct btrfs_transaction *trans,
246                                           unsigned int type)
247 {
248         atomic_set(&trans->num_extwriters, ((type & TRANS_EXTWRITERS) ? 1 : 0));
249 }
250
251 static inline int extwriter_counter_read(struct btrfs_transaction *trans)
252 {
253         return atomic_read(&trans->num_extwriters);
254 }
255
256 /*
257  * To be called after all the new block groups attached to the transaction
258  * handle have been created (btrfs_create_pending_block_groups()).
259  */
260 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
261 {
262         struct btrfs_fs_info *fs_info = trans->fs_info;
263         struct btrfs_transaction *cur_trans = trans->transaction;
264
265         if (!trans->chunk_bytes_reserved)
266                 return;
267
268         WARN_ON_ONCE(!list_empty(&trans->new_bgs));
269
270         btrfs_block_rsv_release(fs_info, &fs_info->chunk_block_rsv,
271                                 trans->chunk_bytes_reserved, NULL);
272         atomic64_sub(trans->chunk_bytes_reserved, &cur_trans->chunk_bytes_reserved);
273         cond_wake_up(&cur_trans->chunk_reserve_wait);
274         trans->chunk_bytes_reserved = 0;
275 }
276
277 /*
278  * either allocate a new transaction or hop into the existing one
279  */
280 static noinline int join_transaction(struct btrfs_fs_info *fs_info,
281                                      unsigned int type)
282 {
283         struct btrfs_transaction *cur_trans;
284
285         spin_lock(&fs_info->trans_lock);
286 loop:
287         /* The file system has been taken offline. No new transactions. */
288         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
289                 spin_unlock(&fs_info->trans_lock);
290                 return -EROFS;
291         }
292
293         cur_trans = fs_info->running_transaction;
294         if (cur_trans) {
295                 if (TRANS_ABORTED(cur_trans)) {
296                         spin_unlock(&fs_info->trans_lock);
297                         return cur_trans->aborted;
298                 }
299                 if (btrfs_blocked_trans_types[cur_trans->state] & type) {
300                         spin_unlock(&fs_info->trans_lock);
301                         return -EBUSY;
302                 }
303                 refcount_inc(&cur_trans->use_count);
304                 atomic_inc(&cur_trans->num_writers);
305                 extwriter_counter_inc(cur_trans, type);
306                 spin_unlock(&fs_info->trans_lock);
307                 return 0;
308         }
309         spin_unlock(&fs_info->trans_lock);
310
311         /*
312          * If we are ATTACH, we just want to catch the current transaction,
313          * and commit it. If there is no transaction, just return ENOENT.
314          */
315         if (type == TRANS_ATTACH)
316                 return -ENOENT;
317
318         /*
319          * JOIN_NOLOCK only happens during the transaction commit, so
320          * it is impossible that ->running_transaction is NULL
321          */
322         BUG_ON(type == TRANS_JOIN_NOLOCK);
323
324         cur_trans = kmalloc(sizeof(*cur_trans), GFP_NOFS);
325         if (!cur_trans)
326                 return -ENOMEM;
327
328         spin_lock(&fs_info->trans_lock);
329         if (fs_info->running_transaction) {
330                 /*
331                  * someone started a transaction after we unlocked.  Make sure
332                  * to redo the checks above
333                  */
334                 kfree(cur_trans);
335                 goto loop;
336         } else if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
337                 spin_unlock(&fs_info->trans_lock);
338                 kfree(cur_trans);
339                 return -EROFS;
340         }
341
342         cur_trans->fs_info = fs_info;
343         atomic_set(&cur_trans->pending_ordered, 0);
344         init_waitqueue_head(&cur_trans->pending_wait);
345         atomic_set(&cur_trans->num_writers, 1);
346         extwriter_counter_init(cur_trans, type);
347         init_waitqueue_head(&cur_trans->writer_wait);
348         init_waitqueue_head(&cur_trans->commit_wait);
349         cur_trans->state = TRANS_STATE_RUNNING;
350         /*
351          * One for this trans handle, one so it will live on until we
352          * commit the transaction.
353          */
354         refcount_set(&cur_trans->use_count, 2);
355         cur_trans->flags = 0;
356         cur_trans->start_time = ktime_get_seconds();
357
358         memset(&cur_trans->delayed_refs, 0, sizeof(cur_trans->delayed_refs));
359
360         cur_trans->delayed_refs.href_root = RB_ROOT_CACHED;
361         cur_trans->delayed_refs.dirty_extent_root = RB_ROOT;
362         atomic_set(&cur_trans->delayed_refs.num_entries, 0);
363
364         /*
365          * although the tree mod log is per file system and not per transaction,
366          * the log must never go across transaction boundaries.
367          */
368         smp_mb();
369         if (!list_empty(&fs_info->tree_mod_seq_list))
370                 WARN(1, KERN_ERR "BTRFS: tree_mod_seq_list not empty when creating a fresh transaction\n");
371         if (!RB_EMPTY_ROOT(&fs_info->tree_mod_log))
372                 WARN(1, KERN_ERR "BTRFS: tree_mod_log rb tree not empty when creating a fresh transaction\n");
373         atomic64_set(&fs_info->tree_mod_seq, 0);
374
375         spin_lock_init(&cur_trans->delayed_refs.lock);
376
377         INIT_LIST_HEAD(&cur_trans->pending_snapshots);
378         INIT_LIST_HEAD(&cur_trans->dev_update_list);
379         INIT_LIST_HEAD(&cur_trans->switch_commits);
380         INIT_LIST_HEAD(&cur_trans->dirty_bgs);
381         INIT_LIST_HEAD(&cur_trans->io_bgs);
382         INIT_LIST_HEAD(&cur_trans->dropped_roots);
383         mutex_init(&cur_trans->cache_write_mutex);
384         spin_lock_init(&cur_trans->dirty_bgs_lock);
385         INIT_LIST_HEAD(&cur_trans->deleted_bgs);
386         spin_lock_init(&cur_trans->dropped_roots_lock);
387         INIT_LIST_HEAD(&cur_trans->releasing_ebs);
388         spin_lock_init(&cur_trans->releasing_ebs_lock);
389         atomic64_set(&cur_trans->chunk_bytes_reserved, 0);
390         init_waitqueue_head(&cur_trans->chunk_reserve_wait);
391         list_add_tail(&cur_trans->list, &fs_info->trans_list);
392         extent_io_tree_init(fs_info, &cur_trans->dirty_pages,
393                         IO_TREE_TRANS_DIRTY_PAGES, fs_info->btree_inode);
394         extent_io_tree_init(fs_info, &cur_trans->pinned_extents,
395                         IO_TREE_FS_PINNED_EXTENTS, NULL);
396         fs_info->generation++;
397         cur_trans->transid = fs_info->generation;
398         fs_info->running_transaction = cur_trans;
399         cur_trans->aborted = 0;
400         spin_unlock(&fs_info->trans_lock);
401
402         return 0;
403 }
404
405 /*
406  * This does all the record keeping required to make sure that a shareable root
407  * is properly recorded in a given transaction.  This is required to make sure
408  * the old root from before we joined the transaction is deleted when the
409  * transaction commits.
410  */
411 static int record_root_in_trans(struct btrfs_trans_handle *trans,
412                                struct btrfs_root *root,
413                                int force)
414 {
415         struct btrfs_fs_info *fs_info = root->fs_info;
416         int ret = 0;
417
418         if ((test_bit(BTRFS_ROOT_SHAREABLE, &root->state) &&
419             root->last_trans < trans->transid) || force) {
420                 WARN_ON(root == fs_info->extent_root);
421                 WARN_ON(!force && root->commit_root != root->node);
422
423                 /*
424                  * see below for IN_TRANS_SETUP usage rules
425                  * we have the reloc mutex held now, so there
426                  * is only one writer in this function
427                  */
428                 set_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
429
430                 /* make sure readers find IN_TRANS_SETUP before
431                  * they find our root->last_trans update
432                  */
433                 smp_wmb();
434
435                 spin_lock(&fs_info->fs_roots_radix_lock);
436                 if (root->last_trans == trans->transid && !force) {
437                         spin_unlock(&fs_info->fs_roots_radix_lock);
438                         return 0;
439                 }
440                 radix_tree_tag_set(&fs_info->fs_roots_radix,
441                                    (unsigned long)root->root_key.objectid,
442                                    BTRFS_ROOT_TRANS_TAG);
443                 spin_unlock(&fs_info->fs_roots_radix_lock);
444                 root->last_trans = trans->transid;
445
446                 /* this is pretty tricky.  We don't want to
447                  * take the relocation lock in btrfs_record_root_in_trans
448                  * unless we're really doing the first setup for this root in
449                  * this transaction.
450                  *
451                  * Normally we'd use root->last_trans as a flag to decide
452                  * if we want to take the expensive mutex.
453                  *
454                  * But, we have to set root->last_trans before we
455                  * init the relocation root, otherwise, we trip over warnings
456                  * in ctree.c.  The solution used here is to flag ourselves
457                  * with root IN_TRANS_SETUP.  When this is 1, we're still
458                  * fixing up the reloc trees and everyone must wait.
459                  *
460                  * When this is zero, they can trust root->last_trans and fly
461                  * through btrfs_record_root_in_trans without having to take the
462                  * lock.  smp_wmb() makes sure that all the writes above are
463                  * done before we pop in the zero below
464                  */
465                 ret = btrfs_init_reloc_root(trans, root);
466                 smp_mb__before_atomic();
467                 clear_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
468         }
469         return ret;
470 }
471
472
473 void btrfs_add_dropped_root(struct btrfs_trans_handle *trans,
474                             struct btrfs_root *root)
475 {
476         struct btrfs_fs_info *fs_info = root->fs_info;
477         struct btrfs_transaction *cur_trans = trans->transaction;
478
479         /* Add ourselves to the transaction dropped list */
480         spin_lock(&cur_trans->dropped_roots_lock);
481         list_add_tail(&root->root_list, &cur_trans->dropped_roots);
482         spin_unlock(&cur_trans->dropped_roots_lock);
483
484         /* Make sure we don't try to update the root at commit time */
485         spin_lock(&fs_info->fs_roots_radix_lock);
486         radix_tree_tag_clear(&fs_info->fs_roots_radix,
487                              (unsigned long)root->root_key.objectid,
488                              BTRFS_ROOT_TRANS_TAG);
489         spin_unlock(&fs_info->fs_roots_radix_lock);
490 }
491
492 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
493                                struct btrfs_root *root)
494 {
495         struct btrfs_fs_info *fs_info = root->fs_info;
496         int ret;
497
498         if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
499                 return 0;
500
501         /*
502          * see record_root_in_trans for comments about IN_TRANS_SETUP usage
503          * and barriers
504          */
505         smp_rmb();
506         if (root->last_trans == trans->transid &&
507             !test_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state))
508                 return 0;
509
510         mutex_lock(&fs_info->reloc_mutex);
511         ret = record_root_in_trans(trans, root, 0);
512         mutex_unlock(&fs_info->reloc_mutex);
513
514         return ret;
515 }
516
517 static inline int is_transaction_blocked(struct btrfs_transaction *trans)
518 {
519         return (trans->state >= TRANS_STATE_COMMIT_START &&
520                 trans->state < TRANS_STATE_UNBLOCKED &&
521                 !TRANS_ABORTED(trans));
522 }
523
524 /* wait for commit against the current transaction to become unblocked
525  * when this is done, it is safe to start a new transaction, but the current
526  * transaction might not be fully on disk.
527  */
528 static void wait_current_trans(struct btrfs_fs_info *fs_info)
529 {
530         struct btrfs_transaction *cur_trans;
531
532         spin_lock(&fs_info->trans_lock);
533         cur_trans = fs_info->running_transaction;
534         if (cur_trans && is_transaction_blocked(cur_trans)) {
535                 refcount_inc(&cur_trans->use_count);
536                 spin_unlock(&fs_info->trans_lock);
537
538                 wait_event(fs_info->transaction_wait,
539                            cur_trans->state >= TRANS_STATE_UNBLOCKED ||
540                            TRANS_ABORTED(cur_trans));
541                 btrfs_put_transaction(cur_trans);
542         } else {
543                 spin_unlock(&fs_info->trans_lock);
544         }
545 }
546
547 static int may_wait_transaction(struct btrfs_fs_info *fs_info, int type)
548 {
549         if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
550                 return 0;
551
552         if (type == TRANS_START)
553                 return 1;
554
555         return 0;
556 }
557
558 static inline bool need_reserve_reloc_root(struct btrfs_root *root)
559 {
560         struct btrfs_fs_info *fs_info = root->fs_info;
561
562         if (!fs_info->reloc_ctl ||
563             !test_bit(BTRFS_ROOT_SHAREABLE, &root->state) ||
564             root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
565             root->reloc_root)
566                 return false;
567
568         return true;
569 }
570
571 static struct btrfs_trans_handle *
572 start_transaction(struct btrfs_root *root, unsigned int num_items,
573                   unsigned int type, enum btrfs_reserve_flush_enum flush,
574                   bool enforce_qgroups)
575 {
576         struct btrfs_fs_info *fs_info = root->fs_info;
577         struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
578         struct btrfs_trans_handle *h;
579         struct btrfs_transaction *cur_trans;
580         u64 num_bytes = 0;
581         u64 qgroup_reserved = 0;
582         bool reloc_reserved = false;
583         bool do_chunk_alloc = false;
584         int ret;
585
586         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
587                 return ERR_PTR(-EROFS);
588
589         if (current->journal_info) {
590                 WARN_ON(type & TRANS_EXTWRITERS);
591                 h = current->journal_info;
592                 refcount_inc(&h->use_count);
593                 WARN_ON(refcount_read(&h->use_count) > 2);
594                 h->orig_rsv = h->block_rsv;
595                 h->block_rsv = NULL;
596                 goto got_it;
597         }
598
599         /*
600          * Do the reservation before we join the transaction so we can do all
601          * the appropriate flushing if need be.
602          */
603         if (num_items && root != fs_info->chunk_root) {
604                 struct btrfs_block_rsv *rsv = &fs_info->trans_block_rsv;
605                 u64 delayed_refs_bytes = 0;
606
607                 qgroup_reserved = num_items * fs_info->nodesize;
608                 ret = btrfs_qgroup_reserve_meta_pertrans(root, qgroup_reserved,
609                                 enforce_qgroups);
610                 if (ret)
611                         return ERR_PTR(ret);
612
613                 /*
614                  * We want to reserve all the bytes we may need all at once, so
615                  * we only do 1 enospc flushing cycle per transaction start.  We
616                  * accomplish this by simply assuming we'll do 2 x num_items
617                  * worth of delayed refs updates in this trans handle, and
618                  * refill that amount for whatever is missing in the reserve.
619                  */
620                 num_bytes = btrfs_calc_insert_metadata_size(fs_info, num_items);
621                 if (flush == BTRFS_RESERVE_FLUSH_ALL &&
622                     delayed_refs_rsv->full == 0) {
623                         delayed_refs_bytes = num_bytes;
624                         num_bytes <<= 1;
625                 }
626
627                 /*
628                  * Do the reservation for the relocation root creation
629                  */
630                 if (need_reserve_reloc_root(root)) {
631                         num_bytes += fs_info->nodesize;
632                         reloc_reserved = true;
633                 }
634
635                 ret = btrfs_block_rsv_add(root, rsv, num_bytes, flush);
636                 if (ret)
637                         goto reserve_fail;
638                 if (delayed_refs_bytes) {
639                         btrfs_migrate_to_delayed_refs_rsv(fs_info, rsv,
640                                                           delayed_refs_bytes);
641                         num_bytes -= delayed_refs_bytes;
642                 }
643
644                 if (rsv->space_info->force_alloc)
645                         do_chunk_alloc = true;
646         } else if (num_items == 0 && flush == BTRFS_RESERVE_FLUSH_ALL &&
647                    !delayed_refs_rsv->full) {
648                 /*
649                  * Some people call with btrfs_start_transaction(root, 0)
650                  * because they can be throttled, but have some other mechanism
651                  * for reserving space.  We still want these guys to refill the
652                  * delayed block_rsv so just add 1 items worth of reservation
653                  * here.
654                  */
655                 ret = btrfs_delayed_refs_rsv_refill(fs_info, flush);
656                 if (ret)
657                         goto reserve_fail;
658         }
659 again:
660         h = kmem_cache_zalloc(btrfs_trans_handle_cachep, GFP_NOFS);
661         if (!h) {
662                 ret = -ENOMEM;
663                 goto alloc_fail;
664         }
665
666         /*
667          * If we are JOIN_NOLOCK we're already committing a transaction and
668          * waiting on this guy, so we don't need to do the sb_start_intwrite
669          * because we're already holding a ref.  We need this because we could
670          * have raced in and did an fsync() on a file which can kick a commit
671          * and then we deadlock with somebody doing a freeze.
672          *
673          * If we are ATTACH, it means we just want to catch the current
674          * transaction and commit it, so we needn't do sb_start_intwrite(). 
675          */
676         if (type & __TRANS_FREEZABLE)
677                 sb_start_intwrite(fs_info->sb);
678
679         if (may_wait_transaction(fs_info, type))
680                 wait_current_trans(fs_info);
681
682         do {
683                 ret = join_transaction(fs_info, type);
684                 if (ret == -EBUSY) {
685                         wait_current_trans(fs_info);
686                         if (unlikely(type == TRANS_ATTACH ||
687                                      type == TRANS_JOIN_NOSTART))
688                                 ret = -ENOENT;
689                 }
690         } while (ret == -EBUSY);
691
692         if (ret < 0)
693                 goto join_fail;
694
695         cur_trans = fs_info->running_transaction;
696
697         h->transid = cur_trans->transid;
698         h->transaction = cur_trans;
699         h->root = root;
700         refcount_set(&h->use_count, 1);
701         h->fs_info = root->fs_info;
702
703         h->type = type;
704         h->can_flush_pending_bgs = true;
705         INIT_LIST_HEAD(&h->new_bgs);
706
707         smp_mb();
708         if (cur_trans->state >= TRANS_STATE_COMMIT_START &&
709             may_wait_transaction(fs_info, type)) {
710                 current->journal_info = h;
711                 btrfs_commit_transaction(h);
712                 goto again;
713         }
714
715         if (num_bytes) {
716                 trace_btrfs_space_reservation(fs_info, "transaction",
717                                               h->transid, num_bytes, 1);
718                 h->block_rsv = &fs_info->trans_block_rsv;
719                 h->bytes_reserved = num_bytes;
720                 h->reloc_reserved = reloc_reserved;
721         }
722
723 got_it:
724         if (!current->journal_info)
725                 current->journal_info = h;
726
727         /*
728          * If the space_info is marked ALLOC_FORCE then we'll get upgraded to
729          * ALLOC_FORCE the first run through, and then we won't allocate for
730          * anybody else who races in later.  We don't care about the return
731          * value here.
732          */
733         if (do_chunk_alloc && num_bytes) {
734                 u64 flags = h->block_rsv->space_info->flags;
735
736                 btrfs_chunk_alloc(h, btrfs_get_alloc_profile(fs_info, flags),
737                                   CHUNK_ALLOC_NO_FORCE);
738         }
739
740         /*
741          * btrfs_record_root_in_trans() needs to alloc new extents, and may
742          * call btrfs_join_transaction() while we're also starting a
743          * transaction.
744          *
745          * Thus it need to be called after current->journal_info initialized,
746          * or we can deadlock.
747          */
748         ret = btrfs_record_root_in_trans(h, root);
749         if (ret) {
750                 /*
751                  * The transaction handle is fully initialized and linked with
752                  * other structures so it needs to be ended in case of errors,
753                  * not just freed.
754                  */
755                 btrfs_end_transaction(h);
756                 return ERR_PTR(ret);
757         }
758
759         return h;
760
761 join_fail:
762         if (type & __TRANS_FREEZABLE)
763                 sb_end_intwrite(fs_info->sb);
764         kmem_cache_free(btrfs_trans_handle_cachep, h);
765 alloc_fail:
766         if (num_bytes)
767                 btrfs_block_rsv_release(fs_info, &fs_info->trans_block_rsv,
768                                         num_bytes, NULL);
769 reserve_fail:
770         btrfs_qgroup_free_meta_pertrans(root, qgroup_reserved);
771         return ERR_PTR(ret);
772 }
773
774 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
775                                                    unsigned int num_items)
776 {
777         return start_transaction(root, num_items, TRANS_START,
778                                  BTRFS_RESERVE_FLUSH_ALL, true);
779 }
780
781 struct btrfs_trans_handle *btrfs_start_transaction_fallback_global_rsv(
782                                         struct btrfs_root *root,
783                                         unsigned int num_items)
784 {
785         return start_transaction(root, num_items, TRANS_START,
786                                  BTRFS_RESERVE_FLUSH_ALL_STEAL, false);
787 }
788
789 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
790 {
791         return start_transaction(root, 0, TRANS_JOIN, BTRFS_RESERVE_NO_FLUSH,
792                                  true);
793 }
794
795 struct btrfs_trans_handle *btrfs_join_transaction_spacecache(struct btrfs_root *root)
796 {
797         return start_transaction(root, 0, TRANS_JOIN_NOLOCK,
798                                  BTRFS_RESERVE_NO_FLUSH, true);
799 }
800
801 /*
802  * Similar to regular join but it never starts a transaction when none is
803  * running or after waiting for the current one to finish.
804  */
805 struct btrfs_trans_handle *btrfs_join_transaction_nostart(struct btrfs_root *root)
806 {
807         return start_transaction(root, 0, TRANS_JOIN_NOSTART,
808                                  BTRFS_RESERVE_NO_FLUSH, true);
809 }
810
811 /*
812  * btrfs_attach_transaction() - catch the running transaction
813  *
814  * It is used when we want to commit the current the transaction, but
815  * don't want to start a new one.
816  *
817  * Note: If this function return -ENOENT, it just means there is no
818  * running transaction. But it is possible that the inactive transaction
819  * is still in the memory, not fully on disk. If you hope there is no
820  * inactive transaction in the fs when -ENOENT is returned, you should
821  * invoke
822  *     btrfs_attach_transaction_barrier()
823  */
824 struct btrfs_trans_handle *btrfs_attach_transaction(struct btrfs_root *root)
825 {
826         return start_transaction(root, 0, TRANS_ATTACH,
827                                  BTRFS_RESERVE_NO_FLUSH, true);
828 }
829
830 /*
831  * btrfs_attach_transaction_barrier() - catch the running transaction
832  *
833  * It is similar to the above function, the difference is this one
834  * will wait for all the inactive transactions until they fully
835  * complete.
836  */
837 struct btrfs_trans_handle *
838 btrfs_attach_transaction_barrier(struct btrfs_root *root)
839 {
840         struct btrfs_trans_handle *trans;
841
842         trans = start_transaction(root, 0, TRANS_ATTACH,
843                                   BTRFS_RESERVE_NO_FLUSH, true);
844         if (trans == ERR_PTR(-ENOENT))
845                 btrfs_wait_for_commit(root->fs_info, 0);
846
847         return trans;
848 }
849
850 /* Wait for a transaction commit to reach at least the given state. */
851 static noinline void wait_for_commit(struct btrfs_transaction *commit,
852                                      const enum btrfs_trans_state min_state)
853 {
854         wait_event(commit->commit_wait, commit->state >= min_state);
855 }
856
857 int btrfs_wait_for_commit(struct btrfs_fs_info *fs_info, u64 transid)
858 {
859         struct btrfs_transaction *cur_trans = NULL, *t;
860         int ret = 0;
861
862         if (transid) {
863                 if (transid <= fs_info->last_trans_committed)
864                         goto out;
865
866                 /* find specified transaction */
867                 spin_lock(&fs_info->trans_lock);
868                 list_for_each_entry(t, &fs_info->trans_list, list) {
869                         if (t->transid == transid) {
870                                 cur_trans = t;
871                                 refcount_inc(&cur_trans->use_count);
872                                 ret = 0;
873                                 break;
874                         }
875                         if (t->transid > transid) {
876                                 ret = 0;
877                                 break;
878                         }
879                 }
880                 spin_unlock(&fs_info->trans_lock);
881
882                 /*
883                  * The specified transaction doesn't exist, or we
884                  * raced with btrfs_commit_transaction
885                  */
886                 if (!cur_trans) {
887                         if (transid > fs_info->last_trans_committed)
888                                 ret = -EINVAL;
889                         goto out;
890                 }
891         } else {
892                 /* find newest transaction that is committing | committed */
893                 spin_lock(&fs_info->trans_lock);
894                 list_for_each_entry_reverse(t, &fs_info->trans_list,
895                                             list) {
896                         if (t->state >= TRANS_STATE_COMMIT_START) {
897                                 if (t->state == TRANS_STATE_COMPLETED)
898                                         break;
899                                 cur_trans = t;
900                                 refcount_inc(&cur_trans->use_count);
901                                 break;
902                         }
903                 }
904                 spin_unlock(&fs_info->trans_lock);
905                 if (!cur_trans)
906                         goto out;  /* nothing committing|committed */
907         }
908
909         wait_for_commit(cur_trans, TRANS_STATE_COMPLETED);
910         btrfs_put_transaction(cur_trans);
911 out:
912         return ret;
913 }
914
915 void btrfs_throttle(struct btrfs_fs_info *fs_info)
916 {
917         wait_current_trans(fs_info);
918 }
919
920 static bool should_end_transaction(struct btrfs_trans_handle *trans)
921 {
922         struct btrfs_fs_info *fs_info = trans->fs_info;
923
924         if (btrfs_check_space_for_delayed_refs(fs_info))
925                 return true;
926
927         return !!btrfs_block_rsv_check(&fs_info->global_block_rsv, 5);
928 }
929
930 bool btrfs_should_end_transaction(struct btrfs_trans_handle *trans)
931 {
932         struct btrfs_transaction *cur_trans = trans->transaction;
933
934         if (cur_trans->state >= TRANS_STATE_COMMIT_START ||
935             test_bit(BTRFS_DELAYED_REFS_FLUSHING, &cur_trans->delayed_refs.flags))
936                 return true;
937
938         return should_end_transaction(trans);
939 }
940
941 static void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans)
942
943 {
944         struct btrfs_fs_info *fs_info = trans->fs_info;
945
946         if (!trans->block_rsv) {
947                 ASSERT(!trans->bytes_reserved);
948                 return;
949         }
950
951         if (!trans->bytes_reserved)
952                 return;
953
954         ASSERT(trans->block_rsv == &fs_info->trans_block_rsv);
955         trace_btrfs_space_reservation(fs_info, "transaction",
956                                       trans->transid, trans->bytes_reserved, 0);
957         btrfs_block_rsv_release(fs_info, trans->block_rsv,
958                                 trans->bytes_reserved, NULL);
959         trans->bytes_reserved = 0;
960 }
961
962 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
963                                    int throttle)
964 {
965         struct btrfs_fs_info *info = trans->fs_info;
966         struct btrfs_transaction *cur_trans = trans->transaction;
967         int err = 0;
968
969         if (refcount_read(&trans->use_count) > 1) {
970                 refcount_dec(&trans->use_count);
971                 trans->block_rsv = trans->orig_rsv;
972                 return 0;
973         }
974
975         btrfs_trans_release_metadata(trans);
976         trans->block_rsv = NULL;
977
978         btrfs_create_pending_block_groups(trans);
979
980         btrfs_trans_release_chunk_metadata(trans);
981
982         if (trans->type & __TRANS_FREEZABLE)
983                 sb_end_intwrite(info->sb);
984
985         WARN_ON(cur_trans != info->running_transaction);
986         WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
987         atomic_dec(&cur_trans->num_writers);
988         extwriter_counter_dec(cur_trans, trans->type);
989
990         cond_wake_up(&cur_trans->writer_wait);
991         btrfs_put_transaction(cur_trans);
992
993         if (current->journal_info == trans)
994                 current->journal_info = NULL;
995
996         if (throttle)
997                 btrfs_run_delayed_iputs(info);
998
999         if (TRANS_ABORTED(trans) ||
1000             test_bit(BTRFS_FS_STATE_ERROR, &info->fs_state)) {
1001                 wake_up_process(info->transaction_kthread);
1002                 if (TRANS_ABORTED(trans))
1003                         err = trans->aborted;
1004                 else
1005                         err = -EROFS;
1006         }
1007
1008         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1009         return err;
1010 }
1011
1012 int btrfs_end_transaction(struct btrfs_trans_handle *trans)
1013 {
1014         return __btrfs_end_transaction(trans, 0);
1015 }
1016
1017 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans)
1018 {
1019         return __btrfs_end_transaction(trans, 1);
1020 }
1021
1022 /*
1023  * when btree blocks are allocated, they have some corresponding bits set for
1024  * them in one of two extent_io trees.  This is used to make sure all of
1025  * those extents are sent to disk but does not wait on them
1026  */
1027 int btrfs_write_marked_extents(struct btrfs_fs_info *fs_info,
1028                                struct extent_io_tree *dirty_pages, int mark)
1029 {
1030         int err = 0;
1031         int werr = 0;
1032         struct address_space *mapping = fs_info->btree_inode->i_mapping;
1033         struct extent_state *cached_state = NULL;
1034         u64 start = 0;
1035         u64 end;
1036
1037         atomic_inc(&BTRFS_I(fs_info->btree_inode)->sync_writers);
1038         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
1039                                       mark, &cached_state)) {
1040                 bool wait_writeback = false;
1041
1042                 err = convert_extent_bit(dirty_pages, start, end,
1043                                          EXTENT_NEED_WAIT,
1044                                          mark, &cached_state);
1045                 /*
1046                  * convert_extent_bit can return -ENOMEM, which is most of the
1047                  * time a temporary error. So when it happens, ignore the error
1048                  * and wait for writeback of this range to finish - because we
1049                  * failed to set the bit EXTENT_NEED_WAIT for the range, a call
1050                  * to __btrfs_wait_marked_extents() would not know that
1051                  * writeback for this range started and therefore wouldn't
1052                  * wait for it to finish - we don't want to commit a
1053                  * superblock that points to btree nodes/leafs for which
1054                  * writeback hasn't finished yet (and without errors).
1055                  * We cleanup any entries left in the io tree when committing
1056                  * the transaction (through extent_io_tree_release()).
1057                  */
1058                 if (err == -ENOMEM) {
1059                         err = 0;
1060                         wait_writeback = true;
1061                 }
1062                 if (!err)
1063                         err = filemap_fdatawrite_range(mapping, start, end);
1064                 if (err)
1065                         werr = err;
1066                 else if (wait_writeback)
1067                         werr = filemap_fdatawait_range(mapping, start, end);
1068                 free_extent_state(cached_state);
1069                 cached_state = NULL;
1070                 cond_resched();
1071                 start = end + 1;
1072         }
1073         atomic_dec(&BTRFS_I(fs_info->btree_inode)->sync_writers);
1074         return werr;
1075 }
1076
1077 /*
1078  * when btree blocks are allocated, they have some corresponding bits set for
1079  * them in one of two extent_io trees.  This is used to make sure all of
1080  * those extents are on disk for transaction or log commit.  We wait
1081  * on all the pages and clear them from the dirty pages state tree
1082  */
1083 static int __btrfs_wait_marked_extents(struct btrfs_fs_info *fs_info,
1084                                        struct extent_io_tree *dirty_pages)
1085 {
1086         int err = 0;
1087         int werr = 0;
1088         struct address_space *mapping = fs_info->btree_inode->i_mapping;
1089         struct extent_state *cached_state = NULL;
1090         u64 start = 0;
1091         u64 end;
1092
1093         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
1094                                       EXTENT_NEED_WAIT, &cached_state)) {
1095                 /*
1096                  * Ignore -ENOMEM errors returned by clear_extent_bit().
1097                  * When committing the transaction, we'll remove any entries
1098                  * left in the io tree. For a log commit, we don't remove them
1099                  * after committing the log because the tree can be accessed
1100                  * concurrently - we do it only at transaction commit time when
1101                  * it's safe to do it (through extent_io_tree_release()).
1102                  */
1103                 err = clear_extent_bit(dirty_pages, start, end,
1104                                        EXTENT_NEED_WAIT, 0, 0, &cached_state);
1105                 if (err == -ENOMEM)
1106                         err = 0;
1107                 if (!err)
1108                         err = filemap_fdatawait_range(mapping, start, end);
1109                 if (err)
1110                         werr = err;
1111                 free_extent_state(cached_state);
1112                 cached_state = NULL;
1113                 cond_resched();
1114                 start = end + 1;
1115         }
1116         if (err)
1117                 werr = err;
1118         return werr;
1119 }
1120
1121 static int btrfs_wait_extents(struct btrfs_fs_info *fs_info,
1122                        struct extent_io_tree *dirty_pages)
1123 {
1124         bool errors = false;
1125         int err;
1126
1127         err = __btrfs_wait_marked_extents(fs_info, dirty_pages);
1128         if (test_and_clear_bit(BTRFS_FS_BTREE_ERR, &fs_info->flags))
1129                 errors = true;
1130
1131         if (errors && !err)
1132                 err = -EIO;
1133         return err;
1134 }
1135
1136 int btrfs_wait_tree_log_extents(struct btrfs_root *log_root, int mark)
1137 {
1138         struct btrfs_fs_info *fs_info = log_root->fs_info;
1139         struct extent_io_tree *dirty_pages = &log_root->dirty_log_pages;
1140         bool errors = false;
1141         int err;
1142
1143         ASSERT(log_root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
1144
1145         err = __btrfs_wait_marked_extents(fs_info, dirty_pages);
1146         if ((mark & EXTENT_DIRTY) &&
1147             test_and_clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags))
1148                 errors = true;
1149
1150         if ((mark & EXTENT_NEW) &&
1151             test_and_clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags))
1152                 errors = true;
1153
1154         if (errors && !err)
1155                 err = -EIO;
1156         return err;
1157 }
1158
1159 /*
1160  * When btree blocks are allocated the corresponding extents are marked dirty.
1161  * This function ensures such extents are persisted on disk for transaction or
1162  * log commit.
1163  *
1164  * @trans: transaction whose dirty pages we'd like to write
1165  */
1166 static int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans)
1167 {
1168         int ret;
1169         int ret2;
1170         struct extent_io_tree *dirty_pages = &trans->transaction->dirty_pages;
1171         struct btrfs_fs_info *fs_info = trans->fs_info;
1172         struct blk_plug plug;
1173
1174         blk_start_plug(&plug);
1175         ret = btrfs_write_marked_extents(fs_info, dirty_pages, EXTENT_DIRTY);
1176         blk_finish_plug(&plug);
1177         ret2 = btrfs_wait_extents(fs_info, dirty_pages);
1178
1179         extent_io_tree_release(&trans->transaction->dirty_pages);
1180
1181         if (ret)
1182                 return ret;
1183         else if (ret2)
1184                 return ret2;
1185         else
1186                 return 0;
1187 }
1188
1189 /*
1190  * this is used to update the root pointer in the tree of tree roots.
1191  *
1192  * But, in the case of the extent allocation tree, updating the root
1193  * pointer may allocate blocks which may change the root of the extent
1194  * allocation tree.
1195  *
1196  * So, this loops and repeats and makes sure the cowonly root didn't
1197  * change while the root pointer was being updated in the metadata.
1198  */
1199 static int update_cowonly_root(struct btrfs_trans_handle *trans,
1200                                struct btrfs_root *root)
1201 {
1202         int ret;
1203         u64 old_root_bytenr;
1204         u64 old_root_used;
1205         struct btrfs_fs_info *fs_info = root->fs_info;
1206         struct btrfs_root *tree_root = fs_info->tree_root;
1207
1208         old_root_used = btrfs_root_used(&root->root_item);
1209
1210         while (1) {
1211                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
1212                 if (old_root_bytenr == root->node->start &&
1213                     old_root_used == btrfs_root_used(&root->root_item))
1214                         break;
1215
1216                 btrfs_set_root_node(&root->root_item, root->node);
1217                 ret = btrfs_update_root(trans, tree_root,
1218                                         &root->root_key,
1219                                         &root->root_item);
1220                 if (ret)
1221                         return ret;
1222
1223                 old_root_used = btrfs_root_used(&root->root_item);
1224         }
1225
1226         return 0;
1227 }
1228
1229 /*
1230  * update all the cowonly tree roots on disk
1231  *
1232  * The error handling in this function may not be obvious. Any of the
1233  * failures will cause the file system to go offline. We still need
1234  * to clean up the delayed refs.
1235  */
1236 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans)
1237 {
1238         struct btrfs_fs_info *fs_info = trans->fs_info;
1239         struct list_head *dirty_bgs = &trans->transaction->dirty_bgs;
1240         struct list_head *io_bgs = &trans->transaction->io_bgs;
1241         struct list_head *next;
1242         struct extent_buffer *eb;
1243         int ret;
1244
1245         eb = btrfs_lock_root_node(fs_info->tree_root);
1246         ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL,
1247                               0, &eb, BTRFS_NESTING_COW);
1248         btrfs_tree_unlock(eb);
1249         free_extent_buffer(eb);
1250
1251         if (ret)
1252                 return ret;
1253
1254         ret = btrfs_run_dev_stats(trans);
1255         if (ret)
1256                 return ret;
1257         ret = btrfs_run_dev_replace(trans);
1258         if (ret)
1259                 return ret;
1260         ret = btrfs_run_qgroups(trans);
1261         if (ret)
1262                 return ret;
1263
1264         ret = btrfs_setup_space_cache(trans);
1265         if (ret)
1266                 return ret;
1267
1268 again:
1269         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
1270                 struct btrfs_root *root;
1271                 next = fs_info->dirty_cowonly_roots.next;
1272                 list_del_init(next);
1273                 root = list_entry(next, struct btrfs_root, dirty_list);
1274                 clear_bit(BTRFS_ROOT_DIRTY, &root->state);
1275
1276                 if (root != fs_info->extent_root)
1277                         list_add_tail(&root->dirty_list,
1278                                       &trans->transaction->switch_commits);
1279                 ret = update_cowonly_root(trans, root);
1280                 if (ret)
1281                         return ret;
1282         }
1283
1284         /* Now flush any delayed refs generated by updating all of the roots */
1285         ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
1286         if (ret)
1287                 return ret;
1288
1289         while (!list_empty(dirty_bgs) || !list_empty(io_bgs)) {
1290                 ret = btrfs_write_dirty_block_groups(trans);
1291                 if (ret)
1292                         return ret;
1293
1294                 /*
1295                  * We're writing the dirty block groups, which could generate
1296                  * delayed refs, which could generate more dirty block groups,
1297                  * so we want to keep this flushing in this loop to make sure
1298                  * everything gets run.
1299                  */
1300                 ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
1301                 if (ret)
1302                         return ret;
1303         }
1304
1305         if (!list_empty(&fs_info->dirty_cowonly_roots))
1306                 goto again;
1307
1308         list_add_tail(&fs_info->extent_root->dirty_list,
1309                       &trans->transaction->switch_commits);
1310
1311         /* Update dev-replace pointer once everything is committed */
1312         fs_info->dev_replace.committed_cursor_left =
1313                 fs_info->dev_replace.cursor_left_last_write_of_item;
1314
1315         return 0;
1316 }
1317
1318 /*
1319  * dead roots are old snapshots that need to be deleted.  This allocates
1320  * a dirty root struct and adds it into the list of dead roots that need to
1321  * be deleted
1322  */
1323 void btrfs_add_dead_root(struct btrfs_root *root)
1324 {
1325         struct btrfs_fs_info *fs_info = root->fs_info;
1326
1327         spin_lock(&fs_info->trans_lock);
1328         if (list_empty(&root->root_list)) {
1329                 btrfs_grab_root(root);
1330                 list_add_tail(&root->root_list, &fs_info->dead_roots);
1331         }
1332         spin_unlock(&fs_info->trans_lock);
1333 }
1334
1335 /*
1336  * update all the cowonly tree roots on disk
1337  */
1338 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans)
1339 {
1340         struct btrfs_fs_info *fs_info = trans->fs_info;
1341         struct btrfs_root *gang[8];
1342         int i;
1343         int ret;
1344
1345         spin_lock(&fs_info->fs_roots_radix_lock);
1346         while (1) {
1347                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
1348                                                  (void **)gang, 0,
1349                                                  ARRAY_SIZE(gang),
1350                                                  BTRFS_ROOT_TRANS_TAG);
1351                 if (ret == 0)
1352                         break;
1353                 for (i = 0; i < ret; i++) {
1354                         struct btrfs_root *root = gang[i];
1355                         int ret2;
1356
1357                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
1358                                         (unsigned long)root->root_key.objectid,
1359                                         BTRFS_ROOT_TRANS_TAG);
1360                         spin_unlock(&fs_info->fs_roots_radix_lock);
1361
1362                         btrfs_free_log(trans, root);
1363                         ret2 = btrfs_update_reloc_root(trans, root);
1364                         if (ret2)
1365                                 return ret2;
1366
1367                         /* see comments in should_cow_block() */
1368                         clear_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1369                         smp_mb__after_atomic();
1370
1371                         if (root->commit_root != root->node) {
1372                                 list_add_tail(&root->dirty_list,
1373                                         &trans->transaction->switch_commits);
1374                                 btrfs_set_root_node(&root->root_item,
1375                                                     root->node);
1376                         }
1377
1378                         ret2 = btrfs_update_root(trans, fs_info->tree_root,
1379                                                 &root->root_key,
1380                                                 &root->root_item);
1381                         if (ret2)
1382                                 return ret2;
1383                         spin_lock(&fs_info->fs_roots_radix_lock);
1384                         btrfs_qgroup_free_meta_all_pertrans(root);
1385                 }
1386         }
1387         spin_unlock(&fs_info->fs_roots_radix_lock);
1388         return 0;
1389 }
1390
1391 /*
1392  * defrag a given btree.
1393  * Every leaf in the btree is read and defragged.
1394  */
1395 int btrfs_defrag_root(struct btrfs_root *root)
1396 {
1397         struct btrfs_fs_info *info = root->fs_info;
1398         struct btrfs_trans_handle *trans;
1399         int ret;
1400
1401         if (test_and_set_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state))
1402                 return 0;
1403
1404         while (1) {
1405                 trans = btrfs_start_transaction(root, 0);
1406                 if (IS_ERR(trans)) {
1407                         ret = PTR_ERR(trans);
1408                         break;
1409                 }
1410
1411                 ret = btrfs_defrag_leaves(trans, root);
1412
1413                 btrfs_end_transaction(trans);
1414                 btrfs_btree_balance_dirty(info);
1415                 cond_resched();
1416
1417                 if (btrfs_fs_closing(info) || ret != -EAGAIN)
1418                         break;
1419
1420                 if (btrfs_defrag_cancelled(info)) {
1421                         btrfs_debug(info, "defrag_root cancelled");
1422                         ret = -EAGAIN;
1423                         break;
1424                 }
1425         }
1426         clear_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state);
1427         return ret;
1428 }
1429
1430 /*
1431  * Do all special snapshot related qgroup dirty hack.
1432  *
1433  * Will do all needed qgroup inherit and dirty hack like switch commit
1434  * roots inside one transaction and write all btree into disk, to make
1435  * qgroup works.
1436  */
1437 static int qgroup_account_snapshot(struct btrfs_trans_handle *trans,
1438                                    struct btrfs_root *src,
1439                                    struct btrfs_root *parent,
1440                                    struct btrfs_qgroup_inherit *inherit,
1441                                    u64 dst_objectid)
1442 {
1443         struct btrfs_fs_info *fs_info = src->fs_info;
1444         int ret;
1445
1446         /*
1447          * Save some performance in the case that qgroups are not
1448          * enabled. If this check races with the ioctl, rescan will
1449          * kick in anyway.
1450          */
1451         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
1452                 return 0;
1453
1454         /*
1455          * Ensure dirty @src will be committed.  Or, after coming
1456          * commit_fs_roots() and switch_commit_roots(), any dirty but not
1457          * recorded root will never be updated again, causing an outdated root
1458          * item.
1459          */
1460         ret = record_root_in_trans(trans, src, 1);
1461         if (ret)
1462                 return ret;
1463
1464         /*
1465          * btrfs_qgroup_inherit relies on a consistent view of the usage for the
1466          * src root, so we must run the delayed refs here.
1467          *
1468          * However this isn't particularly fool proof, because there's no
1469          * synchronization keeping us from changing the tree after this point
1470          * before we do the qgroup_inherit, or even from making changes while
1471          * we're doing the qgroup_inherit.  But that's a problem for the future,
1472          * for now flush the delayed refs to narrow the race window where the
1473          * qgroup counters could end up wrong.
1474          */
1475         ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
1476         if (ret) {
1477                 btrfs_abort_transaction(trans, ret);
1478                 return ret;
1479         }
1480
1481         /*
1482          * We are going to commit transaction, see btrfs_commit_transaction()
1483          * comment for reason locking tree_log_mutex
1484          */
1485         mutex_lock(&fs_info->tree_log_mutex);
1486
1487         ret = commit_fs_roots(trans);
1488         if (ret)
1489                 goto out;
1490         ret = btrfs_qgroup_account_extents(trans);
1491         if (ret < 0)
1492                 goto out;
1493
1494         /* Now qgroup are all updated, we can inherit it to new qgroups */
1495         ret = btrfs_qgroup_inherit(trans, src->root_key.objectid, dst_objectid,
1496                                    inherit);
1497         if (ret < 0)
1498                 goto out;
1499
1500         /*
1501          * Now we do a simplified commit transaction, which will:
1502          * 1) commit all subvolume and extent tree
1503          *    To ensure all subvolume and extent tree have a valid
1504          *    commit_root to accounting later insert_dir_item()
1505          * 2) write all btree blocks onto disk
1506          *    This is to make sure later btree modification will be cowed
1507          *    Or commit_root can be populated and cause wrong qgroup numbers
1508          * In this simplified commit, we don't really care about other trees
1509          * like chunk and root tree, as they won't affect qgroup.
1510          * And we don't write super to avoid half committed status.
1511          */
1512         ret = commit_cowonly_roots(trans);
1513         if (ret)
1514                 goto out;
1515         switch_commit_roots(trans);
1516         ret = btrfs_write_and_wait_transaction(trans);
1517         if (ret)
1518                 btrfs_handle_fs_error(fs_info, ret,
1519                         "Error while writing out transaction for qgroup");
1520
1521 out:
1522         mutex_unlock(&fs_info->tree_log_mutex);
1523
1524         /*
1525          * Force parent root to be updated, as we recorded it before so its
1526          * last_trans == cur_transid.
1527          * Or it won't be committed again onto disk after later
1528          * insert_dir_item()
1529          */
1530         if (!ret)
1531                 ret = record_root_in_trans(trans, parent, 1);
1532         return ret;
1533 }
1534
1535 /*
1536  * new snapshots need to be created at a very specific time in the
1537  * transaction commit.  This does the actual creation.
1538  *
1539  * Note:
1540  * If the error which may affect the commitment of the current transaction
1541  * happens, we should return the error number. If the error which just affect
1542  * the creation of the pending snapshots, just return 0.
1543  */
1544 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
1545                                    struct btrfs_pending_snapshot *pending)
1546 {
1547
1548         struct btrfs_fs_info *fs_info = trans->fs_info;
1549         struct btrfs_key key;
1550         struct btrfs_root_item *new_root_item;
1551         struct btrfs_root *tree_root = fs_info->tree_root;
1552         struct btrfs_root *root = pending->root;
1553         struct btrfs_root *parent_root;
1554         struct btrfs_block_rsv *rsv;
1555         struct inode *parent_inode;
1556         struct btrfs_path *path;
1557         struct btrfs_dir_item *dir_item;
1558         struct dentry *dentry;
1559         struct extent_buffer *tmp;
1560         struct extent_buffer *old;
1561         struct timespec64 cur_time;
1562         int ret = 0;
1563         u64 to_reserve = 0;
1564         u64 index = 0;
1565         u64 objectid;
1566         u64 root_flags;
1567
1568         ASSERT(pending->path);
1569         path = pending->path;
1570
1571         ASSERT(pending->root_item);
1572         new_root_item = pending->root_item;
1573
1574         pending->error = btrfs_get_free_objectid(tree_root, &objectid);
1575         if (pending->error)
1576                 goto no_free_objectid;
1577
1578         /*
1579          * Make qgroup to skip current new snapshot's qgroupid, as it is
1580          * accounted by later btrfs_qgroup_inherit().
1581          */
1582         btrfs_set_skip_qgroup(trans, objectid);
1583
1584         btrfs_reloc_pre_snapshot(pending, &to_reserve);
1585
1586         if (to_reserve > 0) {
1587                 pending->error = btrfs_block_rsv_add(root,
1588                                                      &pending->block_rsv,
1589                                                      to_reserve,
1590                                                      BTRFS_RESERVE_NO_FLUSH);
1591                 if (pending->error)
1592                         goto clear_skip_qgroup;
1593         }
1594
1595         key.objectid = objectid;
1596         key.offset = (u64)-1;
1597         key.type = BTRFS_ROOT_ITEM_KEY;
1598
1599         rsv = trans->block_rsv;
1600         trans->block_rsv = &pending->block_rsv;
1601         trans->bytes_reserved = trans->block_rsv->reserved;
1602         trace_btrfs_space_reservation(fs_info, "transaction",
1603                                       trans->transid,
1604                                       trans->bytes_reserved, 1);
1605         dentry = pending->dentry;
1606         parent_inode = pending->dir;
1607         parent_root = BTRFS_I(parent_inode)->root;
1608         ret = record_root_in_trans(trans, parent_root, 0);
1609         if (ret)
1610                 goto fail;
1611         cur_time = current_time(parent_inode);
1612
1613         /*
1614          * insert the directory item
1615          */
1616         ret = btrfs_set_inode_index(BTRFS_I(parent_inode), &index);
1617         BUG_ON(ret); /* -ENOMEM */
1618
1619         /* check if there is a file/dir which has the same name. */
1620         dir_item = btrfs_lookup_dir_item(NULL, parent_root, path,
1621                                          btrfs_ino(BTRFS_I(parent_inode)),
1622                                          dentry->d_name.name,
1623                                          dentry->d_name.len, 0);
1624         if (dir_item != NULL && !IS_ERR(dir_item)) {
1625                 pending->error = -EEXIST;
1626                 goto dir_item_existed;
1627         } else if (IS_ERR(dir_item)) {
1628                 ret = PTR_ERR(dir_item);
1629                 btrfs_abort_transaction(trans, ret);
1630                 goto fail;
1631         }
1632         btrfs_release_path(path);
1633
1634         /*
1635          * pull in the delayed directory update
1636          * and the delayed inode item
1637          * otherwise we corrupt the FS during
1638          * snapshot
1639          */
1640         ret = btrfs_run_delayed_items(trans);
1641         if (ret) {      /* Transaction aborted */
1642                 btrfs_abort_transaction(trans, ret);
1643                 goto fail;
1644         }
1645
1646         ret = record_root_in_trans(trans, root, 0);
1647         if (ret) {
1648                 btrfs_abort_transaction(trans, ret);
1649                 goto fail;
1650         }
1651         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
1652         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
1653         btrfs_check_and_init_root_item(new_root_item);
1654
1655         root_flags = btrfs_root_flags(new_root_item);
1656         if (pending->readonly)
1657                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
1658         else
1659                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
1660         btrfs_set_root_flags(new_root_item, root_flags);
1661
1662         btrfs_set_root_generation_v2(new_root_item,
1663                         trans->transid);
1664         generate_random_guid(new_root_item->uuid);
1665         memcpy(new_root_item->parent_uuid, root->root_item.uuid,
1666                         BTRFS_UUID_SIZE);
1667         if (!(root_flags & BTRFS_ROOT_SUBVOL_RDONLY)) {
1668                 memset(new_root_item->received_uuid, 0,
1669                        sizeof(new_root_item->received_uuid));
1670                 memset(&new_root_item->stime, 0, sizeof(new_root_item->stime));
1671                 memset(&new_root_item->rtime, 0, sizeof(new_root_item->rtime));
1672                 btrfs_set_root_stransid(new_root_item, 0);
1673                 btrfs_set_root_rtransid(new_root_item, 0);
1674         }
1675         btrfs_set_stack_timespec_sec(&new_root_item->otime, cur_time.tv_sec);
1676         btrfs_set_stack_timespec_nsec(&new_root_item->otime, cur_time.tv_nsec);
1677         btrfs_set_root_otransid(new_root_item, trans->transid);
1678
1679         old = btrfs_lock_root_node(root);
1680         ret = btrfs_cow_block(trans, root, old, NULL, 0, &old,
1681                               BTRFS_NESTING_COW);
1682         if (ret) {
1683                 btrfs_tree_unlock(old);
1684                 free_extent_buffer(old);
1685                 btrfs_abort_transaction(trans, ret);
1686                 goto fail;
1687         }
1688
1689         ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1690         /* clean up in any case */
1691         btrfs_tree_unlock(old);
1692         free_extent_buffer(old);
1693         if (ret) {
1694                 btrfs_abort_transaction(trans, ret);
1695                 goto fail;
1696         }
1697         /* see comments in should_cow_block() */
1698         set_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1699         smp_wmb();
1700
1701         btrfs_set_root_node(new_root_item, tmp);
1702         /* record when the snapshot was created in key.offset */
1703         key.offset = trans->transid;
1704         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1705         btrfs_tree_unlock(tmp);
1706         free_extent_buffer(tmp);
1707         if (ret) {
1708                 btrfs_abort_transaction(trans, ret);
1709                 goto fail;
1710         }
1711
1712         /*
1713          * insert root back/forward references
1714          */
1715         ret = btrfs_add_root_ref(trans, objectid,
1716                                  parent_root->root_key.objectid,
1717                                  btrfs_ino(BTRFS_I(parent_inode)), index,
1718                                  dentry->d_name.name, dentry->d_name.len);
1719         if (ret) {
1720                 btrfs_abort_transaction(trans, ret);
1721                 goto fail;
1722         }
1723
1724         key.offset = (u64)-1;
1725         pending->snap = btrfs_get_new_fs_root(fs_info, objectid, pending->anon_dev);
1726         if (IS_ERR(pending->snap)) {
1727                 ret = PTR_ERR(pending->snap);
1728                 pending->snap = NULL;
1729                 btrfs_abort_transaction(trans, ret);
1730                 goto fail;
1731         }
1732
1733         ret = btrfs_reloc_post_snapshot(trans, pending);
1734         if (ret) {
1735                 btrfs_abort_transaction(trans, ret);
1736                 goto fail;
1737         }
1738
1739         /*
1740          * Do special qgroup accounting for snapshot, as we do some qgroup
1741          * snapshot hack to do fast snapshot.
1742          * To co-operate with that hack, we do hack again.
1743          * Or snapshot will be greatly slowed down by a subtree qgroup rescan
1744          */
1745         ret = qgroup_account_snapshot(trans, root, parent_root,
1746                                       pending->inherit, objectid);
1747         if (ret < 0)
1748                 goto fail;
1749
1750         ret = btrfs_insert_dir_item(trans, dentry->d_name.name,
1751                                     dentry->d_name.len, BTRFS_I(parent_inode),
1752                                     &key, BTRFS_FT_DIR, index);
1753         /* We have check then name at the beginning, so it is impossible. */
1754         BUG_ON(ret == -EEXIST || ret == -EOVERFLOW);
1755         if (ret) {
1756                 btrfs_abort_transaction(trans, ret);
1757                 goto fail;
1758         }
1759
1760         btrfs_i_size_write(BTRFS_I(parent_inode), parent_inode->i_size +
1761                                          dentry->d_name.len * 2);
1762         parent_inode->i_mtime = parent_inode->i_ctime =
1763                 current_time(parent_inode);
1764         ret = btrfs_update_inode_fallback(trans, parent_root, BTRFS_I(parent_inode));
1765         if (ret) {
1766                 btrfs_abort_transaction(trans, ret);
1767                 goto fail;
1768         }
1769         ret = btrfs_uuid_tree_add(trans, new_root_item->uuid,
1770                                   BTRFS_UUID_KEY_SUBVOL,
1771                                   objectid);
1772         if (ret) {
1773                 btrfs_abort_transaction(trans, ret);
1774                 goto fail;
1775         }
1776         if (!btrfs_is_empty_uuid(new_root_item->received_uuid)) {
1777                 ret = btrfs_uuid_tree_add(trans, new_root_item->received_uuid,
1778                                           BTRFS_UUID_KEY_RECEIVED_SUBVOL,
1779                                           objectid);
1780                 if (ret && ret != -EEXIST) {
1781                         btrfs_abort_transaction(trans, ret);
1782                         goto fail;
1783                 }
1784         }
1785
1786 fail:
1787         pending->error = ret;
1788 dir_item_existed:
1789         trans->block_rsv = rsv;
1790         trans->bytes_reserved = 0;
1791 clear_skip_qgroup:
1792         btrfs_clear_skip_qgroup(trans);
1793 no_free_objectid:
1794         kfree(new_root_item);
1795         pending->root_item = NULL;
1796         btrfs_free_path(path);
1797         pending->path = NULL;
1798
1799         return ret;
1800 }
1801
1802 /*
1803  * create all the snapshots we've scheduled for creation
1804  */
1805 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans)
1806 {
1807         struct btrfs_pending_snapshot *pending, *next;
1808         struct list_head *head = &trans->transaction->pending_snapshots;
1809         int ret = 0;
1810
1811         list_for_each_entry_safe(pending, next, head, list) {
1812                 list_del(&pending->list);
1813                 ret = create_pending_snapshot(trans, pending);
1814                 if (ret)
1815                         break;
1816         }
1817         return ret;
1818 }
1819
1820 static void update_super_roots(struct btrfs_fs_info *fs_info)
1821 {
1822         struct btrfs_root_item *root_item;
1823         struct btrfs_super_block *super;
1824
1825         super = fs_info->super_copy;
1826
1827         root_item = &fs_info->chunk_root->root_item;
1828         super->chunk_root = root_item->bytenr;
1829         super->chunk_root_generation = root_item->generation;
1830         super->chunk_root_level = root_item->level;
1831
1832         root_item = &fs_info->tree_root->root_item;
1833         super->root = root_item->bytenr;
1834         super->generation = root_item->generation;
1835         super->root_level = root_item->level;
1836         if (btrfs_test_opt(fs_info, SPACE_CACHE))
1837                 super->cache_generation = root_item->generation;
1838         else if (test_bit(BTRFS_FS_CLEANUP_SPACE_CACHE_V1, &fs_info->flags))
1839                 super->cache_generation = 0;
1840         if (test_bit(BTRFS_FS_UPDATE_UUID_TREE_GEN, &fs_info->flags))
1841                 super->uuid_tree_generation = root_item->generation;
1842 }
1843
1844 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1845 {
1846         struct btrfs_transaction *trans;
1847         int ret = 0;
1848
1849         spin_lock(&info->trans_lock);
1850         trans = info->running_transaction;
1851         if (trans)
1852                 ret = (trans->state >= TRANS_STATE_COMMIT_START);
1853         spin_unlock(&info->trans_lock);
1854         return ret;
1855 }
1856
1857 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1858 {
1859         struct btrfs_transaction *trans;
1860         int ret = 0;
1861
1862         spin_lock(&info->trans_lock);
1863         trans = info->running_transaction;
1864         if (trans)
1865                 ret = is_transaction_blocked(trans);
1866         spin_unlock(&info->trans_lock);
1867         return ret;
1868 }
1869
1870 /*
1871  * commit transactions asynchronously. once btrfs_commit_transaction_async
1872  * returns, any subsequent transaction will not be allowed to join.
1873  */
1874 struct btrfs_async_commit {
1875         struct btrfs_trans_handle *newtrans;
1876         struct work_struct work;
1877 };
1878
1879 static void do_async_commit(struct work_struct *work)
1880 {
1881         struct btrfs_async_commit *ac =
1882                 container_of(work, struct btrfs_async_commit, work);
1883
1884         /*
1885          * We've got freeze protection passed with the transaction.
1886          * Tell lockdep about it.
1887          */
1888         if (ac->newtrans->type & __TRANS_FREEZABLE)
1889                 __sb_writers_acquired(ac->newtrans->fs_info->sb, SB_FREEZE_FS);
1890
1891         current->journal_info = ac->newtrans;
1892
1893         btrfs_commit_transaction(ac->newtrans);
1894         kfree(ac);
1895 }
1896
1897 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans)
1898 {
1899         struct btrfs_fs_info *fs_info = trans->fs_info;
1900         struct btrfs_async_commit *ac;
1901         struct btrfs_transaction *cur_trans;
1902
1903         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1904         if (!ac)
1905                 return -ENOMEM;
1906
1907         INIT_WORK(&ac->work, do_async_commit);
1908         ac->newtrans = btrfs_join_transaction(trans->root);
1909         if (IS_ERR(ac->newtrans)) {
1910                 int err = PTR_ERR(ac->newtrans);
1911                 kfree(ac);
1912                 return err;
1913         }
1914
1915         /* take transaction reference */
1916         cur_trans = trans->transaction;
1917         refcount_inc(&cur_trans->use_count);
1918
1919         btrfs_end_transaction(trans);
1920
1921         /*
1922          * Tell lockdep we've released the freeze rwsem, since the
1923          * async commit thread will be the one to unlock it.
1924          */
1925         if (ac->newtrans->type & __TRANS_FREEZABLE)
1926                 __sb_writers_release(fs_info->sb, SB_FREEZE_FS);
1927
1928         schedule_work(&ac->work);
1929         /*
1930          * Wait for the current transaction commit to start and block
1931          * subsequent transaction joins
1932          */
1933         wait_event(fs_info->transaction_blocked_wait,
1934                    cur_trans->state >= TRANS_STATE_COMMIT_START ||
1935                    TRANS_ABORTED(cur_trans));
1936         if (current->journal_info == trans)
1937                 current->journal_info = NULL;
1938
1939         btrfs_put_transaction(cur_trans);
1940         return 0;
1941 }
1942
1943
1944 static void cleanup_transaction(struct btrfs_trans_handle *trans, int err)
1945 {
1946         struct btrfs_fs_info *fs_info = trans->fs_info;
1947         struct btrfs_transaction *cur_trans = trans->transaction;
1948
1949         WARN_ON(refcount_read(&trans->use_count) > 1);
1950
1951         btrfs_abort_transaction(trans, err);
1952
1953         spin_lock(&fs_info->trans_lock);
1954
1955         /*
1956          * If the transaction is removed from the list, it means this
1957          * transaction has been committed successfully, so it is impossible
1958          * to call the cleanup function.
1959          */
1960         BUG_ON(list_empty(&cur_trans->list));
1961
1962         if (cur_trans == fs_info->running_transaction) {
1963                 cur_trans->state = TRANS_STATE_COMMIT_DOING;
1964                 spin_unlock(&fs_info->trans_lock);
1965                 wait_event(cur_trans->writer_wait,
1966                            atomic_read(&cur_trans->num_writers) == 1);
1967
1968                 spin_lock(&fs_info->trans_lock);
1969         }
1970
1971         /*
1972          * Now that we know no one else is still using the transaction we can
1973          * remove the transaction from the list of transactions. This avoids
1974          * the transaction kthread from cleaning up the transaction while some
1975          * other task is still using it, which could result in a use-after-free
1976          * on things like log trees, as it forces the transaction kthread to
1977          * wait for this transaction to be cleaned up by us.
1978          */
1979         list_del_init(&cur_trans->list);
1980
1981         spin_unlock(&fs_info->trans_lock);
1982
1983         btrfs_cleanup_one_transaction(trans->transaction, fs_info);
1984
1985         spin_lock(&fs_info->trans_lock);
1986         if (cur_trans == fs_info->running_transaction)
1987                 fs_info->running_transaction = NULL;
1988         spin_unlock(&fs_info->trans_lock);
1989
1990         if (trans->type & __TRANS_FREEZABLE)
1991                 sb_end_intwrite(fs_info->sb);
1992         btrfs_put_transaction(cur_trans);
1993         btrfs_put_transaction(cur_trans);
1994
1995         trace_btrfs_transaction_commit(trans->root);
1996
1997         if (current->journal_info == trans)
1998                 current->journal_info = NULL;
1999         btrfs_scrub_cancel(fs_info);
2000
2001         kmem_cache_free(btrfs_trans_handle_cachep, trans);
2002 }
2003
2004 /*
2005  * Release reserved delayed ref space of all pending block groups of the
2006  * transaction and remove them from the list
2007  */
2008 static void btrfs_cleanup_pending_block_groups(struct btrfs_trans_handle *trans)
2009 {
2010        struct btrfs_fs_info *fs_info = trans->fs_info;
2011        struct btrfs_block_group *block_group, *tmp;
2012
2013        list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
2014                btrfs_delayed_refs_rsv_release(fs_info, 1);
2015                list_del_init(&block_group->bg_list);
2016        }
2017 }
2018
2019 static inline int btrfs_start_delalloc_flush(struct btrfs_fs_info *fs_info)
2020 {
2021         /*
2022          * We use writeback_inodes_sb here because if we used
2023          * btrfs_start_delalloc_roots we would deadlock with fs freeze.
2024          * Currently are holding the fs freeze lock, if we do an async flush
2025          * we'll do btrfs_join_transaction() and deadlock because we need to
2026          * wait for the fs freeze lock.  Using the direct flushing we benefit
2027          * from already being in a transaction and our join_transaction doesn't
2028          * have to re-take the fs freeze lock.
2029          */
2030         if (btrfs_test_opt(fs_info, FLUSHONCOMMIT))
2031                 writeback_inodes_sb(fs_info->sb, WB_REASON_SYNC);
2032         return 0;
2033 }
2034
2035 static inline void btrfs_wait_delalloc_flush(struct btrfs_fs_info *fs_info)
2036 {
2037         if (btrfs_test_opt(fs_info, FLUSHONCOMMIT))
2038                 btrfs_wait_ordered_roots(fs_info, U64_MAX, 0, (u64)-1);
2039 }
2040
2041 int btrfs_commit_transaction(struct btrfs_trans_handle *trans)
2042 {
2043         struct btrfs_fs_info *fs_info = trans->fs_info;
2044         struct btrfs_transaction *cur_trans = trans->transaction;
2045         struct btrfs_transaction *prev_trans = NULL;
2046         int ret;
2047
2048         ASSERT(refcount_read(&trans->use_count) == 1);
2049
2050         /* Stop the commit early if ->aborted is set */
2051         if (TRANS_ABORTED(cur_trans)) {
2052                 ret = cur_trans->aborted;
2053                 btrfs_end_transaction(trans);
2054                 return ret;
2055         }
2056
2057         btrfs_trans_release_metadata(trans);
2058         trans->block_rsv = NULL;
2059
2060         /*
2061          * We only want one transaction commit doing the flushing so we do not
2062          * waste a bunch of time on lock contention on the extent root node.
2063          */
2064         if (!test_and_set_bit(BTRFS_DELAYED_REFS_FLUSHING,
2065                               &cur_trans->delayed_refs.flags)) {
2066                 /*
2067                  * Make a pass through all the delayed refs we have so far.
2068                  * Any running threads may add more while we are here.
2069                  */
2070                 ret = btrfs_run_delayed_refs(trans, 0);
2071                 if (ret) {
2072                         btrfs_end_transaction(trans);
2073                         return ret;
2074                 }
2075         }
2076
2077         btrfs_create_pending_block_groups(trans);
2078
2079         if (!test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &cur_trans->flags)) {
2080                 int run_it = 0;
2081
2082                 /* this mutex is also taken before trying to set
2083                  * block groups readonly.  We need to make sure
2084                  * that nobody has set a block group readonly
2085                  * after a extents from that block group have been
2086                  * allocated for cache files.  btrfs_set_block_group_ro
2087                  * will wait for the transaction to commit if it
2088                  * finds BTRFS_TRANS_DIRTY_BG_RUN set.
2089                  *
2090                  * The BTRFS_TRANS_DIRTY_BG_RUN flag is also used to make sure
2091                  * only one process starts all the block group IO.  It wouldn't
2092                  * hurt to have more than one go through, but there's no
2093                  * real advantage to it either.
2094                  */
2095                 mutex_lock(&fs_info->ro_block_group_mutex);
2096                 if (!test_and_set_bit(BTRFS_TRANS_DIRTY_BG_RUN,
2097                                       &cur_trans->flags))
2098                         run_it = 1;
2099                 mutex_unlock(&fs_info->ro_block_group_mutex);
2100
2101                 if (run_it) {
2102                         ret = btrfs_start_dirty_block_groups(trans);
2103                         if (ret) {
2104                                 btrfs_end_transaction(trans);
2105                                 return ret;
2106                         }
2107                 }
2108         }
2109
2110         spin_lock(&fs_info->trans_lock);
2111         if (cur_trans->state >= TRANS_STATE_COMMIT_START) {
2112                 enum btrfs_trans_state want_state = TRANS_STATE_COMPLETED;
2113
2114                 spin_unlock(&fs_info->trans_lock);
2115                 refcount_inc(&cur_trans->use_count);
2116
2117                 if (trans->in_fsync)
2118                         want_state = TRANS_STATE_SUPER_COMMITTED;
2119                 ret = btrfs_end_transaction(trans);
2120                 wait_for_commit(cur_trans, want_state);
2121
2122                 if (TRANS_ABORTED(cur_trans))
2123                         ret = cur_trans->aborted;
2124
2125                 btrfs_put_transaction(cur_trans);
2126
2127                 return ret;
2128         }
2129
2130         cur_trans->state = TRANS_STATE_COMMIT_START;
2131         wake_up(&fs_info->transaction_blocked_wait);
2132
2133         if (cur_trans->list.prev != &fs_info->trans_list) {
2134                 enum btrfs_trans_state want_state = TRANS_STATE_COMPLETED;
2135
2136                 if (trans->in_fsync)
2137                         want_state = TRANS_STATE_SUPER_COMMITTED;
2138
2139                 prev_trans = list_entry(cur_trans->list.prev,
2140                                         struct btrfs_transaction, list);
2141                 if (prev_trans->state < want_state) {
2142                         refcount_inc(&prev_trans->use_count);
2143                         spin_unlock(&fs_info->trans_lock);
2144
2145                         wait_for_commit(prev_trans, want_state);
2146
2147                         ret = READ_ONCE(prev_trans->aborted);
2148
2149                         btrfs_put_transaction(prev_trans);
2150                         if (ret)
2151                                 goto cleanup_transaction;
2152                 } else {
2153                         spin_unlock(&fs_info->trans_lock);
2154                 }
2155         } else {
2156                 spin_unlock(&fs_info->trans_lock);
2157                 /*
2158                  * The previous transaction was aborted and was already removed
2159                  * from the list of transactions at fs_info->trans_list. So we
2160                  * abort to prevent writing a new superblock that reflects a
2161                  * corrupt state (pointing to trees with unwritten nodes/leafs).
2162                  */
2163                 if (test_bit(BTRFS_FS_STATE_TRANS_ABORTED, &fs_info->fs_state)) {
2164                         ret = -EROFS;
2165                         goto cleanup_transaction;
2166                 }
2167         }
2168
2169         extwriter_counter_dec(cur_trans, trans->type);
2170
2171         ret = btrfs_start_delalloc_flush(fs_info);
2172         if (ret)
2173                 goto cleanup_transaction;
2174
2175         ret = btrfs_run_delayed_items(trans);
2176         if (ret)
2177                 goto cleanup_transaction;
2178
2179         wait_event(cur_trans->writer_wait,
2180                    extwriter_counter_read(cur_trans) == 0);
2181
2182         /* some pending stuffs might be added after the previous flush. */
2183         ret = btrfs_run_delayed_items(trans);
2184         if (ret)
2185                 goto cleanup_transaction;
2186
2187         btrfs_wait_delalloc_flush(fs_info);
2188
2189         /*
2190          * Wait for all ordered extents started by a fast fsync that joined this
2191          * transaction. Otherwise if this transaction commits before the ordered
2192          * extents complete we lose logged data after a power failure.
2193          */
2194         wait_event(cur_trans->pending_wait,
2195                    atomic_read(&cur_trans->pending_ordered) == 0);
2196
2197         btrfs_scrub_pause(fs_info);
2198         /*
2199          * Ok now we need to make sure to block out any other joins while we
2200          * commit the transaction.  We could have started a join before setting
2201          * COMMIT_DOING so make sure to wait for num_writers to == 1 again.
2202          */
2203         spin_lock(&fs_info->trans_lock);
2204         cur_trans->state = TRANS_STATE_COMMIT_DOING;
2205         spin_unlock(&fs_info->trans_lock);
2206         wait_event(cur_trans->writer_wait,
2207                    atomic_read(&cur_trans->num_writers) == 1);
2208
2209         if (TRANS_ABORTED(cur_trans)) {
2210                 ret = cur_trans->aborted;
2211                 goto scrub_continue;
2212         }
2213         /*
2214          * the reloc mutex makes sure that we stop
2215          * the balancing code from coming in and moving
2216          * extents around in the middle of the commit
2217          */
2218         mutex_lock(&fs_info->reloc_mutex);
2219
2220         /*
2221          * We needn't worry about the delayed items because we will
2222          * deal with them in create_pending_snapshot(), which is the
2223          * core function of the snapshot creation.
2224          */
2225         ret = create_pending_snapshots(trans);
2226         if (ret)
2227                 goto unlock_reloc;
2228
2229         /*
2230          * We insert the dir indexes of the snapshots and update the inode
2231          * of the snapshots' parents after the snapshot creation, so there
2232          * are some delayed items which are not dealt with. Now deal with
2233          * them.
2234          *
2235          * We needn't worry that this operation will corrupt the snapshots,
2236          * because all the tree which are snapshoted will be forced to COW
2237          * the nodes and leaves.
2238          */
2239         ret = btrfs_run_delayed_items(trans);
2240         if (ret)
2241                 goto unlock_reloc;
2242
2243         ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
2244         if (ret)
2245                 goto unlock_reloc;
2246
2247         /*
2248          * make sure none of the code above managed to slip in a
2249          * delayed item
2250          */
2251         btrfs_assert_delayed_root_empty(fs_info);
2252
2253         WARN_ON(cur_trans != trans->transaction);
2254
2255         /* btrfs_commit_tree_roots is responsible for getting the
2256          * various roots consistent with each other.  Every pointer
2257          * in the tree of tree roots has to point to the most up to date
2258          * root for every subvolume and other tree.  So, we have to keep
2259          * the tree logging code from jumping in and changing any
2260          * of the trees.
2261          *
2262          * At this point in the commit, there can't be any tree-log
2263          * writers, but a little lower down we drop the trans mutex
2264          * and let new people in.  By holding the tree_log_mutex
2265          * from now until after the super is written, we avoid races
2266          * with the tree-log code.
2267          */
2268         mutex_lock(&fs_info->tree_log_mutex);
2269
2270         ret = commit_fs_roots(trans);
2271         if (ret)
2272                 goto unlock_tree_log;
2273
2274         /*
2275          * Since the transaction is done, we can apply the pending changes
2276          * before the next transaction.
2277          */
2278         btrfs_apply_pending_changes(fs_info);
2279
2280         /* commit_fs_roots gets rid of all the tree log roots, it is now
2281          * safe to free the root of tree log roots
2282          */
2283         btrfs_free_log_root_tree(trans, fs_info);
2284
2285         /*
2286          * Since fs roots are all committed, we can get a quite accurate
2287          * new_roots. So let's do quota accounting.
2288          */
2289         ret = btrfs_qgroup_account_extents(trans);
2290         if (ret < 0)
2291                 goto unlock_tree_log;
2292
2293         ret = commit_cowonly_roots(trans);
2294         if (ret)
2295                 goto unlock_tree_log;
2296
2297         /*
2298          * The tasks which save the space cache and inode cache may also
2299          * update ->aborted, check it.
2300          */
2301         if (TRANS_ABORTED(cur_trans)) {
2302                 ret = cur_trans->aborted;
2303                 goto unlock_tree_log;
2304         }
2305
2306         cur_trans = fs_info->running_transaction;
2307
2308         btrfs_set_root_node(&fs_info->tree_root->root_item,
2309                             fs_info->tree_root->node);
2310         list_add_tail(&fs_info->tree_root->dirty_list,
2311                       &cur_trans->switch_commits);
2312
2313         btrfs_set_root_node(&fs_info->chunk_root->root_item,
2314                             fs_info->chunk_root->node);
2315         list_add_tail(&fs_info->chunk_root->dirty_list,
2316                       &cur_trans->switch_commits);
2317
2318         switch_commit_roots(trans);
2319
2320         ASSERT(list_empty(&cur_trans->dirty_bgs));
2321         ASSERT(list_empty(&cur_trans->io_bgs));
2322         update_super_roots(fs_info);
2323
2324         btrfs_set_super_log_root(fs_info->super_copy, 0);
2325         btrfs_set_super_log_root_level(fs_info->super_copy, 0);
2326         memcpy(fs_info->super_for_commit, fs_info->super_copy,
2327                sizeof(*fs_info->super_copy));
2328
2329         btrfs_commit_device_sizes(cur_trans);
2330
2331         clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags);
2332         clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags);
2333
2334         btrfs_trans_release_chunk_metadata(trans);
2335
2336         spin_lock(&fs_info->trans_lock);
2337         cur_trans->state = TRANS_STATE_UNBLOCKED;
2338         fs_info->running_transaction = NULL;
2339         spin_unlock(&fs_info->trans_lock);
2340         mutex_unlock(&fs_info->reloc_mutex);
2341
2342         wake_up(&fs_info->transaction_wait);
2343
2344         ret = btrfs_write_and_wait_transaction(trans);
2345         if (ret) {
2346                 btrfs_handle_fs_error(fs_info, ret,
2347                                       "Error while writing out transaction");
2348                 /*
2349                  * reloc_mutex has been unlocked, tree_log_mutex is still held
2350                  * but we can't jump to unlock_tree_log causing double unlock
2351                  */
2352                 mutex_unlock(&fs_info->tree_log_mutex);
2353                 goto scrub_continue;
2354         }
2355
2356         /*
2357          * At this point, we should have written all the tree blocks allocated
2358          * in this transaction. So it's now safe to free the redirtyied extent
2359          * buffers.
2360          */
2361         btrfs_free_redirty_list(cur_trans);
2362
2363         ret = write_all_supers(fs_info, 0);
2364         /*
2365          * the super is written, we can safely allow the tree-loggers
2366          * to go about their business
2367          */
2368         mutex_unlock(&fs_info->tree_log_mutex);
2369         if (ret)
2370                 goto scrub_continue;
2371
2372         /*
2373          * We needn't acquire the lock here because there is no other task
2374          * which can change it.
2375          */
2376         cur_trans->state = TRANS_STATE_SUPER_COMMITTED;
2377         wake_up(&cur_trans->commit_wait);
2378
2379         btrfs_finish_extent_commit(trans);
2380
2381         if (test_bit(BTRFS_TRANS_HAVE_FREE_BGS, &cur_trans->flags))
2382                 btrfs_clear_space_info_full(fs_info);
2383
2384         fs_info->last_trans_committed = cur_trans->transid;
2385         /*
2386          * We needn't acquire the lock here because there is no other task
2387          * which can change it.
2388          */
2389         cur_trans->state = TRANS_STATE_COMPLETED;
2390         wake_up(&cur_trans->commit_wait);
2391
2392         spin_lock(&fs_info->trans_lock);
2393         list_del_init(&cur_trans->list);
2394         spin_unlock(&fs_info->trans_lock);
2395
2396         btrfs_put_transaction(cur_trans);
2397         btrfs_put_transaction(cur_trans);
2398
2399         if (trans->type & __TRANS_FREEZABLE)
2400                 sb_end_intwrite(fs_info->sb);
2401
2402         trace_btrfs_transaction_commit(trans->root);
2403
2404         btrfs_scrub_continue(fs_info);
2405
2406         if (current->journal_info == trans)
2407                 current->journal_info = NULL;
2408
2409         kmem_cache_free(btrfs_trans_handle_cachep, trans);
2410
2411         return ret;
2412
2413 unlock_tree_log:
2414         mutex_unlock(&fs_info->tree_log_mutex);
2415 unlock_reloc:
2416         mutex_unlock(&fs_info->reloc_mutex);
2417 scrub_continue:
2418         btrfs_scrub_continue(fs_info);
2419 cleanup_transaction:
2420         btrfs_trans_release_metadata(trans);
2421         btrfs_cleanup_pending_block_groups(trans);
2422         btrfs_trans_release_chunk_metadata(trans);
2423         trans->block_rsv = NULL;
2424         btrfs_warn(fs_info, "Skipping commit of aborted transaction.");
2425         if (current->journal_info == trans)
2426                 current->journal_info = NULL;
2427         cleanup_transaction(trans, ret);
2428
2429         return ret;
2430 }
2431
2432 /*
2433  * return < 0 if error
2434  * 0 if there are no more dead_roots at the time of call
2435  * 1 there are more to be processed, call me again
2436  *
2437  * The return value indicates there are certainly more snapshots to delete, but
2438  * if there comes a new one during processing, it may return 0. We don't mind,
2439  * because btrfs_commit_super will poke cleaner thread and it will process it a
2440  * few seconds later.
2441  */
2442 int btrfs_clean_one_deleted_snapshot(struct btrfs_root *root)
2443 {
2444         int ret;
2445         struct btrfs_fs_info *fs_info = root->fs_info;
2446
2447         spin_lock(&fs_info->trans_lock);
2448         if (list_empty(&fs_info->dead_roots)) {
2449                 spin_unlock(&fs_info->trans_lock);
2450                 return 0;
2451         }
2452         root = list_first_entry(&fs_info->dead_roots,
2453                         struct btrfs_root, root_list);
2454         list_del_init(&root->root_list);
2455         spin_unlock(&fs_info->trans_lock);
2456
2457         btrfs_debug(fs_info, "cleaner removing %llu", root->root_key.objectid);
2458
2459         btrfs_kill_all_delayed_nodes(root);
2460
2461         if (btrfs_header_backref_rev(root->node) <
2462                         BTRFS_MIXED_BACKREF_REV)
2463                 ret = btrfs_drop_snapshot(root, 0, 0);
2464         else
2465                 ret = btrfs_drop_snapshot(root, 1, 0);
2466
2467         btrfs_put_root(root);
2468         return (ret < 0) ? 0 : 1;
2469 }
2470
2471 void btrfs_apply_pending_changes(struct btrfs_fs_info *fs_info)
2472 {
2473         unsigned long prev;
2474         unsigned long bit;
2475
2476         prev = xchg(&fs_info->pending_changes, 0);
2477         if (!prev)
2478                 return;
2479
2480         bit = 1 << BTRFS_PENDING_COMMIT;
2481         if (prev & bit)
2482                 btrfs_debug(fs_info, "pending commit done");
2483         prev &= ~bit;
2484
2485         if (prev)
2486                 btrfs_warn(fs_info,
2487                         "unknown pending changes left 0x%lx, ignoring", prev);
2488 }