1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 2007 Oracle. All rights reserved.
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>
15 #include "transaction.h"
18 #include "inode-map.h"
20 #include "dev-replace.h"
23 #define BTRFS_ROOT_TRANS_TAG 0
25 static const unsigned int btrfs_blocked_trans_types[TRANS_STATE_MAX] = {
26 [TRANS_STATE_RUNNING] = 0U,
27 [TRANS_STATE_BLOCKED] = __TRANS_START,
28 [TRANS_STATE_COMMIT_START] = (__TRANS_START | __TRANS_ATTACH),
29 [TRANS_STATE_COMMIT_DOING] = (__TRANS_START |
32 [TRANS_STATE_UNBLOCKED] = (__TRANS_START |
36 [TRANS_STATE_COMPLETED] = (__TRANS_START |
42 void btrfs_put_transaction(struct btrfs_transaction *transaction)
44 WARN_ON(refcount_read(&transaction->use_count) == 0);
45 if (refcount_dec_and_test(&transaction->use_count)) {
46 BUG_ON(!list_empty(&transaction->list));
47 WARN_ON(!RB_EMPTY_ROOT(
48 &transaction->delayed_refs.href_root.rb_root));
49 if (transaction->delayed_refs.pending_csums)
50 btrfs_err(transaction->fs_info,
51 "pending csums is %llu",
52 transaction->delayed_refs.pending_csums);
53 while (!list_empty(&transaction->pending_chunks)) {
54 struct extent_map *em;
56 em = list_first_entry(&transaction->pending_chunks,
57 struct extent_map, list);
58 list_del_init(&em->list);
62 * If any block groups are found in ->deleted_bgs then it's
63 * because the transaction was aborted and a commit did not
64 * happen (things failed before writing the new superblock
65 * and calling btrfs_finish_extent_commit()), so we can not
66 * discard the physical locations of the block groups.
68 while (!list_empty(&transaction->deleted_bgs)) {
69 struct btrfs_block_group_cache *cache;
71 cache = list_first_entry(&transaction->deleted_bgs,
72 struct btrfs_block_group_cache,
74 list_del_init(&cache->bg_list);
75 btrfs_put_block_group_trimming(cache);
76 btrfs_put_block_group(cache);
82 static void clear_btree_io_tree(struct extent_io_tree *tree)
84 spin_lock(&tree->lock);
86 * Do a single barrier for the waitqueue_active check here, the state
87 * of the waitqueue should not change once clear_btree_io_tree is
91 while (!RB_EMPTY_ROOT(&tree->state)) {
93 struct extent_state *state;
95 node = rb_first(&tree->state);
96 state = rb_entry(node, struct extent_state, rb_node);
97 rb_erase(&state->rb_node, &tree->state);
98 RB_CLEAR_NODE(&state->rb_node);
100 * btree io trees aren't supposed to have tasks waiting for
101 * changes in the flags of extent states ever.
103 ASSERT(!waitqueue_active(&state->wq));
104 free_extent_state(state);
106 cond_resched_lock(&tree->lock);
108 spin_unlock(&tree->lock);
111 static noinline void switch_commit_roots(struct btrfs_transaction *trans)
113 struct btrfs_fs_info *fs_info = trans->fs_info;
114 struct btrfs_root *root, *tmp;
116 down_write(&fs_info->commit_root_sem);
117 list_for_each_entry_safe(root, tmp, &trans->switch_commits,
119 list_del_init(&root->dirty_list);
120 free_extent_buffer(root->commit_root);
121 root->commit_root = btrfs_root_node(root);
122 if (is_fstree(root->root_key.objectid))
123 btrfs_unpin_free_ino(root);
124 clear_btree_io_tree(&root->dirty_log_pages);
127 /* We can free old roots now. */
128 spin_lock(&trans->dropped_roots_lock);
129 while (!list_empty(&trans->dropped_roots)) {
130 root = list_first_entry(&trans->dropped_roots,
131 struct btrfs_root, root_list);
132 list_del_init(&root->root_list);
133 spin_unlock(&trans->dropped_roots_lock);
134 btrfs_drop_and_free_fs_root(fs_info, root);
135 spin_lock(&trans->dropped_roots_lock);
137 spin_unlock(&trans->dropped_roots_lock);
138 up_write(&fs_info->commit_root_sem);
141 static inline void extwriter_counter_inc(struct btrfs_transaction *trans,
144 if (type & TRANS_EXTWRITERS)
145 atomic_inc(&trans->num_extwriters);
148 static inline void extwriter_counter_dec(struct btrfs_transaction *trans,
151 if (type & TRANS_EXTWRITERS)
152 atomic_dec(&trans->num_extwriters);
155 static inline void extwriter_counter_init(struct btrfs_transaction *trans,
158 atomic_set(&trans->num_extwriters, ((type & TRANS_EXTWRITERS) ? 1 : 0));
161 static inline int extwriter_counter_read(struct btrfs_transaction *trans)
163 return atomic_read(&trans->num_extwriters);
167 * either allocate a new transaction or hop into the existing one
169 static noinline int join_transaction(struct btrfs_fs_info *fs_info,
172 struct btrfs_transaction *cur_trans;
174 spin_lock(&fs_info->trans_lock);
176 /* The file system has been taken offline. No new transactions. */
177 if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
178 spin_unlock(&fs_info->trans_lock);
182 cur_trans = fs_info->running_transaction;
184 if (cur_trans->aborted) {
185 spin_unlock(&fs_info->trans_lock);
186 return cur_trans->aborted;
188 if (btrfs_blocked_trans_types[cur_trans->state] & type) {
189 spin_unlock(&fs_info->trans_lock);
192 refcount_inc(&cur_trans->use_count);
193 atomic_inc(&cur_trans->num_writers);
194 extwriter_counter_inc(cur_trans, type);
195 spin_unlock(&fs_info->trans_lock);
198 spin_unlock(&fs_info->trans_lock);
201 * If we are ATTACH, we just want to catch the current transaction,
202 * and commit it. If there is no transaction, just return ENOENT.
204 if (type == TRANS_ATTACH)
208 * JOIN_NOLOCK only happens during the transaction commit, so
209 * it is impossible that ->running_transaction is NULL
211 BUG_ON(type == TRANS_JOIN_NOLOCK);
213 cur_trans = kmalloc(sizeof(*cur_trans), GFP_NOFS);
217 spin_lock(&fs_info->trans_lock);
218 if (fs_info->running_transaction) {
220 * someone started a transaction after we unlocked. Make sure
221 * to redo the checks above
225 } else if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
226 spin_unlock(&fs_info->trans_lock);
231 cur_trans->fs_info = fs_info;
232 atomic_set(&cur_trans->num_writers, 1);
233 extwriter_counter_init(cur_trans, type);
234 init_waitqueue_head(&cur_trans->writer_wait);
235 init_waitqueue_head(&cur_trans->commit_wait);
236 init_waitqueue_head(&cur_trans->pending_wait);
237 cur_trans->state = TRANS_STATE_RUNNING;
239 * One for this trans handle, one so it will live on until we
240 * commit the transaction.
242 refcount_set(&cur_trans->use_count, 2);
243 atomic_set(&cur_trans->pending_ordered, 0);
244 cur_trans->flags = 0;
245 cur_trans->start_time = ktime_get_seconds();
247 memset(&cur_trans->delayed_refs, 0, sizeof(cur_trans->delayed_refs));
249 cur_trans->delayed_refs.href_root = RB_ROOT_CACHED;
250 cur_trans->delayed_refs.dirty_extent_root = RB_ROOT;
251 atomic_set(&cur_trans->delayed_refs.num_entries, 0);
254 * although the tree mod log is per file system and not per transaction,
255 * the log must never go across transaction boundaries.
258 if (!list_empty(&fs_info->tree_mod_seq_list))
259 WARN(1, KERN_ERR "BTRFS: tree_mod_seq_list not empty when creating a fresh transaction\n");
260 if (!RB_EMPTY_ROOT(&fs_info->tree_mod_log))
261 WARN(1, KERN_ERR "BTRFS: tree_mod_log rb tree not empty when creating a fresh transaction\n");
262 atomic64_set(&fs_info->tree_mod_seq, 0);
264 spin_lock_init(&cur_trans->delayed_refs.lock);
266 INIT_LIST_HEAD(&cur_trans->pending_snapshots);
267 INIT_LIST_HEAD(&cur_trans->pending_chunks);
268 INIT_LIST_HEAD(&cur_trans->switch_commits);
269 INIT_LIST_HEAD(&cur_trans->dirty_bgs);
270 INIT_LIST_HEAD(&cur_trans->io_bgs);
271 INIT_LIST_HEAD(&cur_trans->dropped_roots);
272 mutex_init(&cur_trans->cache_write_mutex);
273 cur_trans->num_dirty_bgs = 0;
274 spin_lock_init(&cur_trans->dirty_bgs_lock);
275 INIT_LIST_HEAD(&cur_trans->deleted_bgs);
276 spin_lock_init(&cur_trans->dropped_roots_lock);
277 list_add_tail(&cur_trans->list, &fs_info->trans_list);
278 extent_io_tree_init(&cur_trans->dirty_pages,
279 fs_info->btree_inode);
280 fs_info->generation++;
281 cur_trans->transid = fs_info->generation;
282 fs_info->running_transaction = cur_trans;
283 cur_trans->aborted = 0;
284 spin_unlock(&fs_info->trans_lock);
290 * this does all the record keeping required to make sure that a reference
291 * counted root is properly recorded in a given transaction. This is required
292 * to make sure the old root from before we joined the transaction is deleted
293 * when the transaction commits
295 static int record_root_in_trans(struct btrfs_trans_handle *trans,
296 struct btrfs_root *root,
299 struct btrfs_fs_info *fs_info = root->fs_info;
301 if ((test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
302 root->last_trans < trans->transid) || force) {
303 WARN_ON(root == fs_info->extent_root);
304 WARN_ON(!force && root->commit_root != root->node);
307 * see below for IN_TRANS_SETUP usage rules
308 * we have the reloc mutex held now, so there
309 * is only one writer in this function
311 set_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
313 /* make sure readers find IN_TRANS_SETUP before
314 * they find our root->last_trans update
318 spin_lock(&fs_info->fs_roots_radix_lock);
319 if (root->last_trans == trans->transid && !force) {
320 spin_unlock(&fs_info->fs_roots_radix_lock);
323 radix_tree_tag_set(&fs_info->fs_roots_radix,
324 (unsigned long)root->root_key.objectid,
325 BTRFS_ROOT_TRANS_TAG);
326 spin_unlock(&fs_info->fs_roots_radix_lock);
327 root->last_trans = trans->transid;
329 /* this is pretty tricky. We don't want to
330 * take the relocation lock in btrfs_record_root_in_trans
331 * unless we're really doing the first setup for this root in
334 * Normally we'd use root->last_trans as a flag to decide
335 * if we want to take the expensive mutex.
337 * But, we have to set root->last_trans before we
338 * init the relocation root, otherwise, we trip over warnings
339 * in ctree.c. The solution used here is to flag ourselves
340 * with root IN_TRANS_SETUP. When this is 1, we're still
341 * fixing up the reloc trees and everyone must wait.
343 * When this is zero, they can trust root->last_trans and fly
344 * through btrfs_record_root_in_trans without having to take the
345 * lock. smp_wmb() makes sure that all the writes above are
346 * done before we pop in the zero below
348 btrfs_init_reloc_root(trans, root);
349 smp_mb__before_atomic();
350 clear_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
356 void btrfs_add_dropped_root(struct btrfs_trans_handle *trans,
357 struct btrfs_root *root)
359 struct btrfs_fs_info *fs_info = root->fs_info;
360 struct btrfs_transaction *cur_trans = trans->transaction;
362 /* Add ourselves to the transaction dropped list */
363 spin_lock(&cur_trans->dropped_roots_lock);
364 list_add_tail(&root->root_list, &cur_trans->dropped_roots);
365 spin_unlock(&cur_trans->dropped_roots_lock);
367 /* Make sure we don't try to update the root at commit time */
368 spin_lock(&fs_info->fs_roots_radix_lock);
369 radix_tree_tag_clear(&fs_info->fs_roots_radix,
370 (unsigned long)root->root_key.objectid,
371 BTRFS_ROOT_TRANS_TAG);
372 spin_unlock(&fs_info->fs_roots_radix_lock);
375 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
376 struct btrfs_root *root)
378 struct btrfs_fs_info *fs_info = root->fs_info;
380 if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
384 * see record_root_in_trans for comments about IN_TRANS_SETUP usage
388 if (root->last_trans == trans->transid &&
389 !test_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state))
392 mutex_lock(&fs_info->reloc_mutex);
393 record_root_in_trans(trans, root, 0);
394 mutex_unlock(&fs_info->reloc_mutex);
399 static inline int is_transaction_blocked(struct btrfs_transaction *trans)
401 return (trans->state >= TRANS_STATE_BLOCKED &&
402 trans->state < TRANS_STATE_UNBLOCKED &&
406 /* wait for commit against the current transaction to become unblocked
407 * when this is done, it is safe to start a new transaction, but the current
408 * transaction might not be fully on disk.
410 static void wait_current_trans(struct btrfs_fs_info *fs_info)
412 struct btrfs_transaction *cur_trans;
414 spin_lock(&fs_info->trans_lock);
415 cur_trans = fs_info->running_transaction;
416 if (cur_trans && is_transaction_blocked(cur_trans)) {
417 refcount_inc(&cur_trans->use_count);
418 spin_unlock(&fs_info->trans_lock);
420 wait_event(fs_info->transaction_wait,
421 cur_trans->state >= TRANS_STATE_UNBLOCKED ||
423 btrfs_put_transaction(cur_trans);
425 spin_unlock(&fs_info->trans_lock);
429 static int may_wait_transaction(struct btrfs_fs_info *fs_info, int type)
431 if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
434 if (type == TRANS_START)
440 static inline bool need_reserve_reloc_root(struct btrfs_root *root)
442 struct btrfs_fs_info *fs_info = root->fs_info;
444 if (!fs_info->reloc_ctl ||
445 !test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
446 root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
453 static struct btrfs_trans_handle *
454 start_transaction(struct btrfs_root *root, unsigned int num_items,
455 unsigned int type, enum btrfs_reserve_flush_enum flush,
456 bool enforce_qgroups)
458 struct btrfs_fs_info *fs_info = root->fs_info;
460 struct btrfs_trans_handle *h;
461 struct btrfs_transaction *cur_trans;
463 u64 qgroup_reserved = 0;
464 bool reloc_reserved = false;
467 /* Send isn't supposed to start transactions. */
468 ASSERT(current->journal_info != BTRFS_SEND_TRANS_STUB);
470 if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
471 return ERR_PTR(-EROFS);
473 if (current->journal_info) {
474 WARN_ON(type & TRANS_EXTWRITERS);
475 h = current->journal_info;
476 refcount_inc(&h->use_count);
477 WARN_ON(refcount_read(&h->use_count) > 2);
478 h->orig_rsv = h->block_rsv;
484 * Do the reservation before we join the transaction so we can do all
485 * the appropriate flushing if need be.
487 if (num_items && root != fs_info->chunk_root) {
488 qgroup_reserved = num_items * fs_info->nodesize;
489 ret = btrfs_qgroup_reserve_meta_pertrans(root, qgroup_reserved,
494 num_bytes = btrfs_calc_trans_metadata_size(fs_info, num_items);
496 * Do the reservation for the relocation root creation
498 if (need_reserve_reloc_root(root)) {
499 num_bytes += fs_info->nodesize;
500 reloc_reserved = true;
503 ret = btrfs_block_rsv_add(root, &fs_info->trans_block_rsv,
509 h = kmem_cache_zalloc(btrfs_trans_handle_cachep, GFP_NOFS);
516 * If we are JOIN_NOLOCK we're already committing a transaction and
517 * waiting on this guy, so we don't need to do the sb_start_intwrite
518 * because we're already holding a ref. We need this because we could
519 * have raced in and did an fsync() on a file which can kick a commit
520 * and then we deadlock with somebody doing a freeze.
522 * If we are ATTACH, it means we just want to catch the current
523 * transaction and commit it, so we needn't do sb_start_intwrite().
525 if (type & __TRANS_FREEZABLE)
526 sb_start_intwrite(fs_info->sb);
528 if (may_wait_transaction(fs_info, type))
529 wait_current_trans(fs_info);
532 ret = join_transaction(fs_info, type);
534 wait_current_trans(fs_info);
535 if (unlikely(type == TRANS_ATTACH))
538 } while (ret == -EBUSY);
543 cur_trans = fs_info->running_transaction;
545 h->transid = cur_trans->transid;
546 h->transaction = cur_trans;
548 refcount_set(&h->use_count, 1);
549 h->fs_info = root->fs_info;
552 h->can_flush_pending_bgs = true;
553 INIT_LIST_HEAD(&h->new_bgs);
556 if (cur_trans->state >= TRANS_STATE_BLOCKED &&
557 may_wait_transaction(fs_info, type)) {
558 current->journal_info = h;
559 btrfs_commit_transaction(h);
564 trace_btrfs_space_reservation(fs_info, "transaction",
565 h->transid, num_bytes, 1);
566 h->block_rsv = &fs_info->trans_block_rsv;
567 h->bytes_reserved = num_bytes;
568 h->reloc_reserved = reloc_reserved;
572 btrfs_record_root_in_trans(h, root);
574 if (!current->journal_info)
575 current->journal_info = h;
579 if (type & __TRANS_FREEZABLE)
580 sb_end_intwrite(fs_info->sb);
581 kmem_cache_free(btrfs_trans_handle_cachep, h);
584 btrfs_block_rsv_release(fs_info, &fs_info->trans_block_rsv,
587 btrfs_qgroup_free_meta_pertrans(root, qgroup_reserved);
591 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
592 unsigned int num_items)
594 return start_transaction(root, num_items, TRANS_START,
595 BTRFS_RESERVE_FLUSH_ALL, true);
598 struct btrfs_trans_handle *btrfs_start_transaction_fallback_global_rsv(
599 struct btrfs_root *root,
600 unsigned int num_items,
603 struct btrfs_fs_info *fs_info = root->fs_info;
604 struct btrfs_trans_handle *trans;
609 * We have two callers: unlink and block group removal. The
610 * former should succeed even if we will temporarily exceed
611 * quota and the latter operates on the extent root so
612 * qgroup enforcement is ignored anyway.
614 trans = start_transaction(root, num_items, TRANS_START,
615 BTRFS_RESERVE_FLUSH_ALL, false);
616 if (!IS_ERR(trans) || PTR_ERR(trans) != -ENOSPC)
619 trans = btrfs_start_transaction(root, 0);
623 num_bytes = btrfs_calc_trans_metadata_size(fs_info, num_items);
624 ret = btrfs_cond_migrate_bytes(fs_info, &fs_info->trans_block_rsv,
625 num_bytes, min_factor);
627 btrfs_end_transaction(trans);
631 trans->block_rsv = &fs_info->trans_block_rsv;
632 trans->bytes_reserved = num_bytes;
633 trace_btrfs_space_reservation(fs_info, "transaction",
634 trans->transid, num_bytes, 1);
639 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
641 return start_transaction(root, 0, TRANS_JOIN, BTRFS_RESERVE_NO_FLUSH,
645 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root)
647 return start_transaction(root, 0, TRANS_JOIN_NOLOCK,
648 BTRFS_RESERVE_NO_FLUSH, true);
652 * btrfs_attach_transaction() - catch the running transaction
654 * It is used when we want to commit the current the transaction, but
655 * don't want to start a new one.
657 * Note: If this function return -ENOENT, it just means there is no
658 * running transaction. But it is possible that the inactive transaction
659 * is still in the memory, not fully on disk. If you hope there is no
660 * inactive transaction in the fs when -ENOENT is returned, you should
662 * btrfs_attach_transaction_barrier()
664 struct btrfs_trans_handle *btrfs_attach_transaction(struct btrfs_root *root)
666 return start_transaction(root, 0, TRANS_ATTACH,
667 BTRFS_RESERVE_NO_FLUSH, true);
671 * btrfs_attach_transaction_barrier() - catch the running transaction
673 * It is similar to the above function, the differentia is this one
674 * will wait for all the inactive transactions until they fully
677 struct btrfs_trans_handle *
678 btrfs_attach_transaction_barrier(struct btrfs_root *root)
680 struct btrfs_trans_handle *trans;
682 trans = start_transaction(root, 0, TRANS_ATTACH,
683 BTRFS_RESERVE_NO_FLUSH, true);
684 if (trans == ERR_PTR(-ENOENT))
685 btrfs_wait_for_commit(root->fs_info, 0);
690 /* wait for a transaction commit to be fully complete */
691 static noinline void wait_for_commit(struct btrfs_transaction *commit)
693 wait_event(commit->commit_wait, commit->state == TRANS_STATE_COMPLETED);
696 int btrfs_wait_for_commit(struct btrfs_fs_info *fs_info, u64 transid)
698 struct btrfs_transaction *cur_trans = NULL, *t;
702 if (transid <= fs_info->last_trans_committed)
705 /* find specified transaction */
706 spin_lock(&fs_info->trans_lock);
707 list_for_each_entry(t, &fs_info->trans_list, list) {
708 if (t->transid == transid) {
710 refcount_inc(&cur_trans->use_count);
714 if (t->transid > transid) {
719 spin_unlock(&fs_info->trans_lock);
722 * The specified transaction doesn't exist, or we
723 * raced with btrfs_commit_transaction
726 if (transid > fs_info->last_trans_committed)
731 /* find newest transaction that is committing | committed */
732 spin_lock(&fs_info->trans_lock);
733 list_for_each_entry_reverse(t, &fs_info->trans_list,
735 if (t->state >= TRANS_STATE_COMMIT_START) {
736 if (t->state == TRANS_STATE_COMPLETED)
739 refcount_inc(&cur_trans->use_count);
743 spin_unlock(&fs_info->trans_lock);
745 goto out; /* nothing committing|committed */
748 wait_for_commit(cur_trans);
749 btrfs_put_transaction(cur_trans);
754 void btrfs_throttle(struct btrfs_fs_info *fs_info)
756 wait_current_trans(fs_info);
759 static int should_end_transaction(struct btrfs_trans_handle *trans)
761 struct btrfs_fs_info *fs_info = trans->fs_info;
763 if (btrfs_check_space_for_delayed_refs(trans, fs_info))
766 return !!btrfs_block_rsv_check(&fs_info->global_block_rsv, 5);
769 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans)
771 struct btrfs_transaction *cur_trans = trans->transaction;
776 if (cur_trans->state >= TRANS_STATE_BLOCKED ||
777 cur_trans->delayed_refs.flushing)
780 updates = trans->delayed_ref_updates;
781 trans->delayed_ref_updates = 0;
783 err = btrfs_run_delayed_refs(trans, updates * 2);
784 if (err) /* Error code will also eval true */
788 return should_end_transaction(trans);
791 static void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans)
794 struct btrfs_fs_info *fs_info = trans->fs_info;
796 if (!trans->block_rsv) {
797 ASSERT(!trans->bytes_reserved);
801 if (!trans->bytes_reserved)
804 ASSERT(trans->block_rsv == &fs_info->trans_block_rsv);
805 trace_btrfs_space_reservation(fs_info, "transaction",
806 trans->transid, trans->bytes_reserved, 0);
807 btrfs_block_rsv_release(fs_info, trans->block_rsv,
808 trans->bytes_reserved);
809 trans->bytes_reserved = 0;
812 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
815 struct btrfs_fs_info *info = trans->fs_info;
816 struct btrfs_transaction *cur_trans = trans->transaction;
817 u64 transid = trans->transid;
818 unsigned long cur = trans->delayed_ref_updates;
819 int lock = (trans->type != TRANS_JOIN_NOLOCK);
821 int must_run_delayed_refs = 0;
823 if (refcount_read(&trans->use_count) > 1) {
824 refcount_dec(&trans->use_count);
825 trans->block_rsv = trans->orig_rsv;
829 btrfs_trans_release_metadata(trans);
830 trans->block_rsv = NULL;
832 if (!list_empty(&trans->new_bgs))
833 btrfs_create_pending_block_groups(trans);
835 trans->delayed_ref_updates = 0;
837 must_run_delayed_refs =
838 btrfs_should_throttle_delayed_refs(trans, info);
839 cur = max_t(unsigned long, cur, 32);
842 * don't make the caller wait if they are from a NOLOCK
843 * or ATTACH transaction, it will deadlock with commit
845 if (must_run_delayed_refs == 1 &&
846 (trans->type & (__TRANS_JOIN_NOLOCK | __TRANS_ATTACH)))
847 must_run_delayed_refs = 2;
850 btrfs_trans_release_metadata(trans);
851 trans->block_rsv = NULL;
853 if (!list_empty(&trans->new_bgs))
854 btrfs_create_pending_block_groups(trans);
856 btrfs_trans_release_chunk_metadata(trans);
858 if (lock && should_end_transaction(trans) &&
859 READ_ONCE(cur_trans->state) == TRANS_STATE_RUNNING) {
860 spin_lock(&info->trans_lock);
861 if (cur_trans->state == TRANS_STATE_RUNNING)
862 cur_trans->state = TRANS_STATE_BLOCKED;
863 spin_unlock(&info->trans_lock);
866 if (lock && READ_ONCE(cur_trans->state) == TRANS_STATE_BLOCKED) {
868 return btrfs_commit_transaction(trans);
870 wake_up_process(info->transaction_kthread);
873 if (trans->type & __TRANS_FREEZABLE)
874 sb_end_intwrite(info->sb);
876 WARN_ON(cur_trans != info->running_transaction);
877 WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
878 atomic_dec(&cur_trans->num_writers);
879 extwriter_counter_dec(cur_trans, trans->type);
881 cond_wake_up(&cur_trans->writer_wait);
882 btrfs_put_transaction(cur_trans);
884 if (current->journal_info == trans)
885 current->journal_info = NULL;
888 btrfs_run_delayed_iputs(info);
890 if (trans->aborted ||
891 test_bit(BTRFS_FS_STATE_ERROR, &info->fs_state)) {
892 wake_up_process(info->transaction_kthread);
896 kmem_cache_free(btrfs_trans_handle_cachep, trans);
897 if (must_run_delayed_refs) {
898 btrfs_async_run_delayed_refs(info, cur, transid,
899 must_run_delayed_refs == 1);
904 int btrfs_end_transaction(struct btrfs_trans_handle *trans)
906 return __btrfs_end_transaction(trans, 0);
909 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans)
911 return __btrfs_end_transaction(trans, 1);
915 * when btree blocks are allocated, they have some corresponding bits set for
916 * them in one of two extent_io trees. This is used to make sure all of
917 * those extents are sent to disk but does not wait on them
919 int btrfs_write_marked_extents(struct btrfs_fs_info *fs_info,
920 struct extent_io_tree *dirty_pages, int mark)
924 struct address_space *mapping = fs_info->btree_inode->i_mapping;
925 struct extent_state *cached_state = NULL;
929 atomic_inc(&BTRFS_I(fs_info->btree_inode)->sync_writers);
930 while (!find_first_extent_bit(dirty_pages, start, &start, &end,
931 mark, &cached_state)) {
932 bool wait_writeback = false;
934 err = convert_extent_bit(dirty_pages, start, end,
936 mark, &cached_state);
938 * convert_extent_bit can return -ENOMEM, which is most of the
939 * time a temporary error. So when it happens, ignore the error
940 * and wait for writeback of this range to finish - because we
941 * failed to set the bit EXTENT_NEED_WAIT for the range, a call
942 * to __btrfs_wait_marked_extents() would not know that
943 * writeback for this range started and therefore wouldn't
944 * wait for it to finish - we don't want to commit a
945 * superblock that points to btree nodes/leafs for which
946 * writeback hasn't finished yet (and without errors).
947 * We cleanup any entries left in the io tree when committing
948 * the transaction (through clear_btree_io_tree()).
950 if (err == -ENOMEM) {
952 wait_writeback = true;
955 err = filemap_fdatawrite_range(mapping, start, end);
958 else if (wait_writeback)
959 werr = filemap_fdatawait_range(mapping, start, end);
960 free_extent_state(cached_state);
965 atomic_dec(&BTRFS_I(fs_info->btree_inode)->sync_writers);
970 * when btree blocks are allocated, they have some corresponding bits set for
971 * them in one of two extent_io trees. This is used to make sure all of
972 * those extents are on disk for transaction or log commit. We wait
973 * on all the pages and clear them from the dirty pages state tree
975 static int __btrfs_wait_marked_extents(struct btrfs_fs_info *fs_info,
976 struct extent_io_tree *dirty_pages)
980 struct address_space *mapping = fs_info->btree_inode->i_mapping;
981 struct extent_state *cached_state = NULL;
985 while (!find_first_extent_bit(dirty_pages, start, &start, &end,
986 EXTENT_NEED_WAIT, &cached_state)) {
988 * Ignore -ENOMEM errors returned by clear_extent_bit().
989 * When committing the transaction, we'll remove any entries
990 * left in the io tree. For a log commit, we don't remove them
991 * after committing the log because the tree can be accessed
992 * concurrently - we do it only at transaction commit time when
993 * it's safe to do it (through clear_btree_io_tree()).
995 err = clear_extent_bit(dirty_pages, start, end,
996 EXTENT_NEED_WAIT, 0, 0, &cached_state);
1000 err = filemap_fdatawait_range(mapping, start, end);
1003 free_extent_state(cached_state);
1004 cached_state = NULL;
1013 int btrfs_wait_extents(struct btrfs_fs_info *fs_info,
1014 struct extent_io_tree *dirty_pages)
1016 bool errors = false;
1019 err = __btrfs_wait_marked_extents(fs_info, dirty_pages);
1020 if (test_and_clear_bit(BTRFS_FS_BTREE_ERR, &fs_info->flags))
1028 int btrfs_wait_tree_log_extents(struct btrfs_root *log_root, int mark)
1030 struct btrfs_fs_info *fs_info = log_root->fs_info;
1031 struct extent_io_tree *dirty_pages = &log_root->dirty_log_pages;
1032 bool errors = false;
1035 ASSERT(log_root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
1037 err = __btrfs_wait_marked_extents(fs_info, dirty_pages);
1038 if ((mark & EXTENT_DIRTY) &&
1039 test_and_clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags))
1042 if ((mark & EXTENT_NEW) &&
1043 test_and_clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags))
1052 * When btree blocks are allocated the corresponding extents are marked dirty.
1053 * This function ensures such extents are persisted on disk for transaction or
1056 * @trans: transaction whose dirty pages we'd like to write
1058 static int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans)
1062 struct extent_io_tree *dirty_pages = &trans->transaction->dirty_pages;
1063 struct btrfs_fs_info *fs_info = trans->fs_info;
1064 struct blk_plug plug;
1066 blk_start_plug(&plug);
1067 ret = btrfs_write_marked_extents(fs_info, dirty_pages, EXTENT_DIRTY);
1068 blk_finish_plug(&plug);
1069 ret2 = btrfs_wait_extents(fs_info, dirty_pages);
1071 clear_btree_io_tree(&trans->transaction->dirty_pages);
1082 * this is used to update the root pointer in the tree of tree roots.
1084 * But, in the case of the extent allocation tree, updating the root
1085 * pointer may allocate blocks which may change the root of the extent
1088 * So, this loops and repeats and makes sure the cowonly root didn't
1089 * change while the root pointer was being updated in the metadata.
1091 static int update_cowonly_root(struct btrfs_trans_handle *trans,
1092 struct btrfs_root *root)
1095 u64 old_root_bytenr;
1097 struct btrfs_fs_info *fs_info = root->fs_info;
1098 struct btrfs_root *tree_root = fs_info->tree_root;
1100 old_root_used = btrfs_root_used(&root->root_item);
1103 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
1104 if (old_root_bytenr == root->node->start &&
1105 old_root_used == btrfs_root_used(&root->root_item))
1108 btrfs_set_root_node(&root->root_item, root->node);
1109 ret = btrfs_update_root(trans, tree_root,
1115 old_root_used = btrfs_root_used(&root->root_item);
1122 * update all the cowonly tree roots on disk
1124 * The error handling in this function may not be obvious. Any of the
1125 * failures will cause the file system to go offline. We still need
1126 * to clean up the delayed refs.
1128 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans)
1130 struct btrfs_fs_info *fs_info = trans->fs_info;
1131 struct list_head *dirty_bgs = &trans->transaction->dirty_bgs;
1132 struct list_head *io_bgs = &trans->transaction->io_bgs;
1133 struct list_head *next;
1134 struct extent_buffer *eb;
1137 eb = btrfs_lock_root_node(fs_info->tree_root);
1138 ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL,
1140 btrfs_tree_unlock(eb);
1141 free_extent_buffer(eb);
1146 ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
1150 ret = btrfs_run_dev_stats(trans, fs_info);
1153 ret = btrfs_run_dev_replace(trans, fs_info);
1156 ret = btrfs_run_qgroups(trans);
1160 ret = btrfs_setup_space_cache(trans, fs_info);
1164 /* run_qgroups might have added some more refs */
1165 ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
1169 while (!list_empty(&fs_info->dirty_cowonly_roots)) {
1170 struct btrfs_root *root;
1171 next = fs_info->dirty_cowonly_roots.next;
1172 list_del_init(next);
1173 root = list_entry(next, struct btrfs_root, dirty_list);
1174 clear_bit(BTRFS_ROOT_DIRTY, &root->state);
1176 if (root != fs_info->extent_root)
1177 list_add_tail(&root->dirty_list,
1178 &trans->transaction->switch_commits);
1179 ret = update_cowonly_root(trans, root);
1182 ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
1187 while (!list_empty(dirty_bgs) || !list_empty(io_bgs)) {
1188 ret = btrfs_write_dirty_block_groups(trans, fs_info);
1191 ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
1196 if (!list_empty(&fs_info->dirty_cowonly_roots))
1199 list_add_tail(&fs_info->extent_root->dirty_list,
1200 &trans->transaction->switch_commits);
1201 btrfs_after_dev_replace_commit(fs_info);
1207 * dead roots are old snapshots that need to be deleted. This allocates
1208 * a dirty root struct and adds it into the list of dead roots that need to
1211 void btrfs_add_dead_root(struct btrfs_root *root)
1213 struct btrfs_fs_info *fs_info = root->fs_info;
1215 spin_lock(&fs_info->trans_lock);
1216 if (list_empty(&root->root_list))
1217 list_add_tail(&root->root_list, &fs_info->dead_roots);
1218 spin_unlock(&fs_info->trans_lock);
1222 * update all the cowonly tree roots on disk
1224 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans)
1226 struct btrfs_fs_info *fs_info = trans->fs_info;
1227 struct btrfs_root *gang[8];
1232 spin_lock(&fs_info->fs_roots_radix_lock);
1234 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
1237 BTRFS_ROOT_TRANS_TAG);
1240 for (i = 0; i < ret; i++) {
1241 struct btrfs_root *root = gang[i];
1242 radix_tree_tag_clear(&fs_info->fs_roots_radix,
1243 (unsigned long)root->root_key.objectid,
1244 BTRFS_ROOT_TRANS_TAG);
1245 spin_unlock(&fs_info->fs_roots_radix_lock);
1247 btrfs_free_log(trans, root);
1248 btrfs_update_reloc_root(trans, root);
1250 btrfs_save_ino_cache(root, trans);
1252 /* see comments in should_cow_block() */
1253 clear_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1254 smp_mb__after_atomic();
1256 if (root->commit_root != root->node) {
1257 list_add_tail(&root->dirty_list,
1258 &trans->transaction->switch_commits);
1259 btrfs_set_root_node(&root->root_item,
1263 err = btrfs_update_root(trans, fs_info->tree_root,
1266 spin_lock(&fs_info->fs_roots_radix_lock);
1269 btrfs_qgroup_free_meta_all_pertrans(root);
1272 spin_unlock(&fs_info->fs_roots_radix_lock);
1277 * defrag a given btree.
1278 * Every leaf in the btree is read and defragged.
1280 int btrfs_defrag_root(struct btrfs_root *root)
1282 struct btrfs_fs_info *info = root->fs_info;
1283 struct btrfs_trans_handle *trans;
1286 if (test_and_set_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state))
1290 trans = btrfs_start_transaction(root, 0);
1292 return PTR_ERR(trans);
1294 ret = btrfs_defrag_leaves(trans, root);
1296 btrfs_end_transaction(trans);
1297 btrfs_btree_balance_dirty(info);
1300 if (btrfs_fs_closing(info) || ret != -EAGAIN)
1303 if (btrfs_defrag_cancelled(info)) {
1304 btrfs_debug(info, "defrag_root cancelled");
1309 clear_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state);
1314 * Do all special snapshot related qgroup dirty hack.
1316 * Will do all needed qgroup inherit and dirty hack like switch commit
1317 * roots inside one transaction and write all btree into disk, to make
1320 static int qgroup_account_snapshot(struct btrfs_trans_handle *trans,
1321 struct btrfs_root *src,
1322 struct btrfs_root *parent,
1323 struct btrfs_qgroup_inherit *inherit,
1326 struct btrfs_fs_info *fs_info = src->fs_info;
1330 * Save some performance in the case that qgroups are not
1331 * enabled. If this check races with the ioctl, rescan will
1334 if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
1338 * Ensure dirty @src will be commited. Or, after comming
1339 * commit_fs_roots() and switch_commit_roots(), any dirty but not
1340 * recorded root will never be updated again, causing an outdated root
1343 record_root_in_trans(trans, src, 1);
1346 * We are going to commit transaction, see btrfs_commit_transaction()
1347 * comment for reason locking tree_log_mutex
1349 mutex_lock(&fs_info->tree_log_mutex);
1351 ret = commit_fs_roots(trans);
1354 ret = btrfs_qgroup_account_extents(trans);
1358 /* Now qgroup are all updated, we can inherit it to new qgroups */
1359 ret = btrfs_qgroup_inherit(trans, src->root_key.objectid, dst_objectid,
1365 * Now we do a simplified commit transaction, which will:
1366 * 1) commit all subvolume and extent tree
1367 * To ensure all subvolume and extent tree have a valid
1368 * commit_root to accounting later insert_dir_item()
1369 * 2) write all btree blocks onto disk
1370 * This is to make sure later btree modification will be cowed
1371 * Or commit_root can be populated and cause wrong qgroup numbers
1372 * In this simplified commit, we don't really care about other trees
1373 * like chunk and root tree, as they won't affect qgroup.
1374 * And we don't write super to avoid half committed status.
1376 ret = commit_cowonly_roots(trans);
1379 switch_commit_roots(trans->transaction);
1380 ret = btrfs_write_and_wait_transaction(trans);
1382 btrfs_handle_fs_error(fs_info, ret,
1383 "Error while writing out transaction for qgroup");
1386 mutex_unlock(&fs_info->tree_log_mutex);
1389 * Force parent root to be updated, as we recorded it before so its
1390 * last_trans == cur_transid.
1391 * Or it won't be committed again onto disk after later
1395 record_root_in_trans(trans, parent, 1);
1400 * new snapshots need to be created at a very specific time in the
1401 * transaction commit. This does the actual creation.
1404 * If the error which may affect the commitment of the current transaction
1405 * happens, we should return the error number. If the error which just affect
1406 * the creation of the pending snapshots, just return 0.
1408 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
1409 struct btrfs_pending_snapshot *pending)
1412 struct btrfs_fs_info *fs_info = trans->fs_info;
1413 struct btrfs_key key;
1414 struct btrfs_root_item *new_root_item;
1415 struct btrfs_root *tree_root = fs_info->tree_root;
1416 struct btrfs_root *root = pending->root;
1417 struct btrfs_root *parent_root;
1418 struct btrfs_block_rsv *rsv;
1419 struct inode *parent_inode;
1420 struct btrfs_path *path;
1421 struct btrfs_dir_item *dir_item;
1422 struct dentry *dentry;
1423 struct extent_buffer *tmp;
1424 struct extent_buffer *old;
1425 struct timespec64 cur_time;
1433 ASSERT(pending->path);
1434 path = pending->path;
1436 ASSERT(pending->root_item);
1437 new_root_item = pending->root_item;
1439 pending->error = btrfs_find_free_objectid(tree_root, &objectid);
1441 goto no_free_objectid;
1444 * Make qgroup to skip current new snapshot's qgroupid, as it is
1445 * accounted by later btrfs_qgroup_inherit().
1447 btrfs_set_skip_qgroup(trans, objectid);
1449 btrfs_reloc_pre_snapshot(pending, &to_reserve);
1451 if (to_reserve > 0) {
1452 pending->error = btrfs_block_rsv_add(root,
1453 &pending->block_rsv,
1455 BTRFS_RESERVE_NO_FLUSH);
1457 goto clear_skip_qgroup;
1460 key.objectid = objectid;
1461 key.offset = (u64)-1;
1462 key.type = BTRFS_ROOT_ITEM_KEY;
1464 rsv = trans->block_rsv;
1465 trans->block_rsv = &pending->block_rsv;
1466 trans->bytes_reserved = trans->block_rsv->reserved;
1467 trace_btrfs_space_reservation(fs_info, "transaction",
1469 trans->bytes_reserved, 1);
1470 dentry = pending->dentry;
1471 parent_inode = pending->dir;
1472 parent_root = BTRFS_I(parent_inode)->root;
1473 record_root_in_trans(trans, parent_root, 0);
1475 cur_time = current_time(parent_inode);
1478 * insert the directory item
1480 ret = btrfs_set_inode_index(BTRFS_I(parent_inode), &index);
1481 BUG_ON(ret); /* -ENOMEM */
1483 /* check if there is a file/dir which has the same name. */
1484 dir_item = btrfs_lookup_dir_item(NULL, parent_root, path,
1485 btrfs_ino(BTRFS_I(parent_inode)),
1486 dentry->d_name.name,
1487 dentry->d_name.len, 0);
1488 if (dir_item != NULL && !IS_ERR(dir_item)) {
1489 pending->error = -EEXIST;
1490 goto dir_item_existed;
1491 } else if (IS_ERR(dir_item)) {
1492 ret = PTR_ERR(dir_item);
1493 btrfs_abort_transaction(trans, ret);
1496 btrfs_release_path(path);
1499 * pull in the delayed directory update
1500 * and the delayed inode item
1501 * otherwise we corrupt the FS during
1504 ret = btrfs_run_delayed_items(trans);
1505 if (ret) { /* Transaction aborted */
1506 btrfs_abort_transaction(trans, ret);
1510 record_root_in_trans(trans, root, 0);
1511 btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
1512 memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
1513 btrfs_check_and_init_root_item(new_root_item);
1515 root_flags = btrfs_root_flags(new_root_item);
1516 if (pending->readonly)
1517 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
1519 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
1520 btrfs_set_root_flags(new_root_item, root_flags);
1522 btrfs_set_root_generation_v2(new_root_item,
1524 uuid_le_gen(&new_uuid);
1525 memcpy(new_root_item->uuid, new_uuid.b, BTRFS_UUID_SIZE);
1526 memcpy(new_root_item->parent_uuid, root->root_item.uuid,
1528 if (!(root_flags & BTRFS_ROOT_SUBVOL_RDONLY)) {
1529 memset(new_root_item->received_uuid, 0,
1530 sizeof(new_root_item->received_uuid));
1531 memset(&new_root_item->stime, 0, sizeof(new_root_item->stime));
1532 memset(&new_root_item->rtime, 0, sizeof(new_root_item->rtime));
1533 btrfs_set_root_stransid(new_root_item, 0);
1534 btrfs_set_root_rtransid(new_root_item, 0);
1536 btrfs_set_stack_timespec_sec(&new_root_item->otime, cur_time.tv_sec);
1537 btrfs_set_stack_timespec_nsec(&new_root_item->otime, cur_time.tv_nsec);
1538 btrfs_set_root_otransid(new_root_item, trans->transid);
1540 old = btrfs_lock_root_node(root);
1541 ret = btrfs_cow_block(trans, root, old, NULL, 0, &old);
1543 btrfs_tree_unlock(old);
1544 free_extent_buffer(old);
1545 btrfs_abort_transaction(trans, ret);
1549 btrfs_set_lock_blocking(old);
1551 ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1552 /* clean up in any case */
1553 btrfs_tree_unlock(old);
1554 free_extent_buffer(old);
1556 btrfs_abort_transaction(trans, ret);
1559 /* see comments in should_cow_block() */
1560 set_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1563 btrfs_set_root_node(new_root_item, tmp);
1564 /* record when the snapshot was created in key.offset */
1565 key.offset = trans->transid;
1566 ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1567 btrfs_tree_unlock(tmp);
1568 free_extent_buffer(tmp);
1570 btrfs_abort_transaction(trans, ret);
1575 * insert root back/forward references
1577 ret = btrfs_add_root_ref(trans, objectid,
1578 parent_root->root_key.objectid,
1579 btrfs_ino(BTRFS_I(parent_inode)), index,
1580 dentry->d_name.name, dentry->d_name.len);
1582 btrfs_abort_transaction(trans, ret);
1586 key.offset = (u64)-1;
1587 pending->snap = btrfs_read_fs_root_no_name(fs_info, &key);
1588 if (IS_ERR(pending->snap)) {
1589 ret = PTR_ERR(pending->snap);
1590 btrfs_abort_transaction(trans, ret);
1594 ret = btrfs_reloc_post_snapshot(trans, pending);
1596 btrfs_abort_transaction(trans, ret);
1600 ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
1602 btrfs_abort_transaction(trans, ret);
1607 * Do special qgroup accounting for snapshot, as we do some qgroup
1608 * snapshot hack to do fast snapshot.
1609 * To co-operate with that hack, we do hack again.
1610 * Or snapshot will be greatly slowed down by a subtree qgroup rescan
1612 ret = qgroup_account_snapshot(trans, root, parent_root,
1613 pending->inherit, objectid);
1617 ret = btrfs_insert_dir_item(trans, dentry->d_name.name,
1618 dentry->d_name.len, BTRFS_I(parent_inode),
1619 &key, BTRFS_FT_DIR, index);
1620 /* We have check then name at the beginning, so it is impossible. */
1621 BUG_ON(ret == -EEXIST || ret == -EOVERFLOW);
1623 btrfs_abort_transaction(trans, ret);
1627 btrfs_i_size_write(BTRFS_I(parent_inode), parent_inode->i_size +
1628 dentry->d_name.len * 2);
1629 parent_inode->i_mtime = parent_inode->i_ctime =
1630 current_time(parent_inode);
1631 ret = btrfs_update_inode_fallback(trans, parent_root, parent_inode);
1633 btrfs_abort_transaction(trans, ret);
1636 ret = btrfs_uuid_tree_add(trans, new_uuid.b, BTRFS_UUID_KEY_SUBVOL,
1639 btrfs_abort_transaction(trans, ret);
1642 if (!btrfs_is_empty_uuid(new_root_item->received_uuid)) {
1643 ret = btrfs_uuid_tree_add(trans, new_root_item->received_uuid,
1644 BTRFS_UUID_KEY_RECEIVED_SUBVOL,
1646 if (ret && ret != -EEXIST) {
1647 btrfs_abort_transaction(trans, ret);
1652 ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
1654 btrfs_abort_transaction(trans, ret);
1659 pending->error = ret;
1661 trans->block_rsv = rsv;
1662 trans->bytes_reserved = 0;
1664 btrfs_clear_skip_qgroup(trans);
1666 kfree(new_root_item);
1667 pending->root_item = NULL;
1668 btrfs_free_path(path);
1669 pending->path = NULL;
1675 * create all the snapshots we've scheduled for creation
1677 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans)
1679 struct btrfs_pending_snapshot *pending, *next;
1680 struct list_head *head = &trans->transaction->pending_snapshots;
1683 list_for_each_entry_safe(pending, next, head, list) {
1684 list_del(&pending->list);
1685 ret = create_pending_snapshot(trans, pending);
1692 static void update_super_roots(struct btrfs_fs_info *fs_info)
1694 struct btrfs_root_item *root_item;
1695 struct btrfs_super_block *super;
1697 super = fs_info->super_copy;
1699 root_item = &fs_info->chunk_root->root_item;
1700 super->chunk_root = root_item->bytenr;
1701 super->chunk_root_generation = root_item->generation;
1702 super->chunk_root_level = root_item->level;
1704 root_item = &fs_info->tree_root->root_item;
1705 super->root = root_item->bytenr;
1706 super->generation = root_item->generation;
1707 super->root_level = root_item->level;
1708 if (btrfs_test_opt(fs_info, SPACE_CACHE))
1709 super->cache_generation = root_item->generation;
1710 if (test_bit(BTRFS_FS_UPDATE_UUID_TREE_GEN, &fs_info->flags))
1711 super->uuid_tree_generation = root_item->generation;
1714 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1716 struct btrfs_transaction *trans;
1719 spin_lock(&info->trans_lock);
1720 trans = info->running_transaction;
1722 ret = (trans->state >= TRANS_STATE_COMMIT_START);
1723 spin_unlock(&info->trans_lock);
1727 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1729 struct btrfs_transaction *trans;
1732 spin_lock(&info->trans_lock);
1733 trans = info->running_transaction;
1735 ret = is_transaction_blocked(trans);
1736 spin_unlock(&info->trans_lock);
1741 * wait for the current transaction commit to start and block subsequent
1744 static void wait_current_trans_commit_start(struct btrfs_fs_info *fs_info,
1745 struct btrfs_transaction *trans)
1747 wait_event(fs_info->transaction_blocked_wait,
1748 trans->state >= TRANS_STATE_COMMIT_START || trans->aborted);
1752 * wait for the current transaction to start and then become unblocked.
1755 static void wait_current_trans_commit_start_and_unblock(
1756 struct btrfs_fs_info *fs_info,
1757 struct btrfs_transaction *trans)
1759 wait_event(fs_info->transaction_wait,
1760 trans->state >= TRANS_STATE_UNBLOCKED || trans->aborted);
1764 * commit transactions asynchronously. once btrfs_commit_transaction_async
1765 * returns, any subsequent transaction will not be allowed to join.
1767 struct btrfs_async_commit {
1768 struct btrfs_trans_handle *newtrans;
1769 struct work_struct work;
1772 static void do_async_commit(struct work_struct *work)
1774 struct btrfs_async_commit *ac =
1775 container_of(work, struct btrfs_async_commit, work);
1778 * We've got freeze protection passed with the transaction.
1779 * Tell lockdep about it.
1781 if (ac->newtrans->type & __TRANS_FREEZABLE)
1782 __sb_writers_acquired(ac->newtrans->fs_info->sb, SB_FREEZE_FS);
1784 current->journal_info = ac->newtrans;
1786 btrfs_commit_transaction(ac->newtrans);
1790 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1791 int wait_for_unblock)
1793 struct btrfs_fs_info *fs_info = trans->fs_info;
1794 struct btrfs_async_commit *ac;
1795 struct btrfs_transaction *cur_trans;
1797 ac = kmalloc(sizeof(*ac), GFP_NOFS);
1801 INIT_WORK(&ac->work, do_async_commit);
1802 ac->newtrans = btrfs_join_transaction(trans->root);
1803 if (IS_ERR(ac->newtrans)) {
1804 int err = PTR_ERR(ac->newtrans);
1809 /* take transaction reference */
1810 cur_trans = trans->transaction;
1811 refcount_inc(&cur_trans->use_count);
1813 btrfs_end_transaction(trans);
1816 * Tell lockdep we've released the freeze rwsem, since the
1817 * async commit thread will be the one to unlock it.
1819 if (ac->newtrans->type & __TRANS_FREEZABLE)
1820 __sb_writers_release(fs_info->sb, SB_FREEZE_FS);
1822 schedule_work(&ac->work);
1824 /* wait for transaction to start and unblock */
1825 if (wait_for_unblock)
1826 wait_current_trans_commit_start_and_unblock(fs_info, cur_trans);
1828 wait_current_trans_commit_start(fs_info, cur_trans);
1830 if (current->journal_info == trans)
1831 current->journal_info = NULL;
1833 btrfs_put_transaction(cur_trans);
1838 static void cleanup_transaction(struct btrfs_trans_handle *trans, int err)
1840 struct btrfs_fs_info *fs_info = trans->fs_info;
1841 struct btrfs_transaction *cur_trans = trans->transaction;
1844 WARN_ON(refcount_read(&trans->use_count) > 1);
1846 btrfs_abort_transaction(trans, err);
1848 spin_lock(&fs_info->trans_lock);
1851 * If the transaction is removed from the list, it means this
1852 * transaction has been committed successfully, so it is impossible
1853 * to call the cleanup function.
1855 BUG_ON(list_empty(&cur_trans->list));
1857 list_del_init(&cur_trans->list);
1858 if (cur_trans == fs_info->running_transaction) {
1859 cur_trans->state = TRANS_STATE_COMMIT_DOING;
1860 spin_unlock(&fs_info->trans_lock);
1861 wait_event(cur_trans->writer_wait,
1862 atomic_read(&cur_trans->num_writers) == 1);
1864 spin_lock(&fs_info->trans_lock);
1866 spin_unlock(&fs_info->trans_lock);
1868 btrfs_cleanup_one_transaction(trans->transaction, fs_info);
1870 spin_lock(&fs_info->trans_lock);
1871 if (cur_trans == fs_info->running_transaction)
1872 fs_info->running_transaction = NULL;
1873 spin_unlock(&fs_info->trans_lock);
1875 if (trans->type & __TRANS_FREEZABLE)
1876 sb_end_intwrite(fs_info->sb);
1877 btrfs_put_transaction(cur_trans);
1878 btrfs_put_transaction(cur_trans);
1880 trace_btrfs_transaction_commit(trans->root);
1882 if (current->journal_info == trans)
1883 current->journal_info = NULL;
1884 btrfs_scrub_cancel(fs_info);
1886 kmem_cache_free(btrfs_trans_handle_cachep, trans);
1889 static inline int btrfs_start_delalloc_flush(struct btrfs_fs_info *fs_info)
1892 * We use writeback_inodes_sb here because if we used
1893 * btrfs_start_delalloc_roots we would deadlock with fs freeze.
1894 * Currently are holding the fs freeze lock, if we do an async flush
1895 * we'll do btrfs_join_transaction() and deadlock because we need to
1896 * wait for the fs freeze lock. Using the direct flushing we benefit
1897 * from already being in a transaction and our join_transaction doesn't
1898 * have to re-take the fs freeze lock.
1900 if (btrfs_test_opt(fs_info, FLUSHONCOMMIT))
1901 writeback_inodes_sb(fs_info->sb, WB_REASON_SYNC);
1905 static inline void btrfs_wait_delalloc_flush(struct btrfs_fs_info *fs_info)
1907 if (btrfs_test_opt(fs_info, FLUSHONCOMMIT))
1908 btrfs_wait_ordered_roots(fs_info, U64_MAX, 0, (u64)-1);
1912 btrfs_wait_pending_ordered(struct btrfs_transaction *cur_trans)
1914 wait_event(cur_trans->pending_wait,
1915 atomic_read(&cur_trans->pending_ordered) == 0);
1918 int btrfs_commit_transaction(struct btrfs_trans_handle *trans)
1920 struct btrfs_fs_info *fs_info = trans->fs_info;
1921 struct btrfs_transaction *cur_trans = trans->transaction;
1922 struct btrfs_transaction *prev_trans = NULL;
1925 /* Stop the commit early if ->aborted is set */
1926 if (unlikely(READ_ONCE(cur_trans->aborted))) {
1927 ret = cur_trans->aborted;
1928 btrfs_end_transaction(trans);
1932 /* make a pass through all the delayed refs we have so far
1933 * any runnings procs may add more while we are here
1935 ret = btrfs_run_delayed_refs(trans, 0);
1937 btrfs_end_transaction(trans);
1941 btrfs_trans_release_metadata(trans);
1942 trans->block_rsv = NULL;
1944 cur_trans = trans->transaction;
1947 * set the flushing flag so procs in this transaction have to
1948 * start sending their work down.
1950 cur_trans->delayed_refs.flushing = 1;
1953 if (!list_empty(&trans->new_bgs))
1954 btrfs_create_pending_block_groups(trans);
1956 ret = btrfs_run_delayed_refs(trans, 0);
1958 btrfs_end_transaction(trans);
1962 if (!test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &cur_trans->flags)) {
1965 /* this mutex is also taken before trying to set
1966 * block groups readonly. We need to make sure
1967 * that nobody has set a block group readonly
1968 * after a extents from that block group have been
1969 * allocated for cache files. btrfs_set_block_group_ro
1970 * will wait for the transaction to commit if it
1971 * finds BTRFS_TRANS_DIRTY_BG_RUN set.
1973 * The BTRFS_TRANS_DIRTY_BG_RUN flag is also used to make sure
1974 * only one process starts all the block group IO. It wouldn't
1975 * hurt to have more than one go through, but there's no
1976 * real advantage to it either.
1978 mutex_lock(&fs_info->ro_block_group_mutex);
1979 if (!test_and_set_bit(BTRFS_TRANS_DIRTY_BG_RUN,
1982 mutex_unlock(&fs_info->ro_block_group_mutex);
1985 ret = btrfs_start_dirty_block_groups(trans);
1987 btrfs_end_transaction(trans);
1993 spin_lock(&fs_info->trans_lock);
1994 if (cur_trans->state >= TRANS_STATE_COMMIT_START) {
1995 spin_unlock(&fs_info->trans_lock);
1996 refcount_inc(&cur_trans->use_count);
1997 ret = btrfs_end_transaction(trans);
1999 wait_for_commit(cur_trans);
2001 if (unlikely(cur_trans->aborted))
2002 ret = cur_trans->aborted;
2004 btrfs_put_transaction(cur_trans);
2009 cur_trans->state = TRANS_STATE_COMMIT_START;
2010 wake_up(&fs_info->transaction_blocked_wait);
2012 if (cur_trans->list.prev != &fs_info->trans_list) {
2013 prev_trans = list_entry(cur_trans->list.prev,
2014 struct btrfs_transaction, list);
2015 if (prev_trans->state != TRANS_STATE_COMPLETED) {
2016 refcount_inc(&prev_trans->use_count);
2017 spin_unlock(&fs_info->trans_lock);
2019 wait_for_commit(prev_trans);
2020 ret = prev_trans->aborted;
2022 btrfs_put_transaction(prev_trans);
2024 goto cleanup_transaction;
2026 spin_unlock(&fs_info->trans_lock);
2029 spin_unlock(&fs_info->trans_lock);
2032 extwriter_counter_dec(cur_trans, trans->type);
2034 ret = btrfs_start_delalloc_flush(fs_info);
2036 goto cleanup_transaction;
2038 ret = btrfs_run_delayed_items(trans);
2040 goto cleanup_transaction;
2042 wait_event(cur_trans->writer_wait,
2043 extwriter_counter_read(cur_trans) == 0);
2045 /* some pending stuffs might be added after the previous flush. */
2046 ret = btrfs_run_delayed_items(trans);
2048 goto cleanup_transaction;
2050 btrfs_wait_delalloc_flush(fs_info);
2052 btrfs_wait_pending_ordered(cur_trans);
2054 btrfs_scrub_pause(fs_info);
2056 * Ok now we need to make sure to block out any other joins while we
2057 * commit the transaction. We could have started a join before setting
2058 * COMMIT_DOING so make sure to wait for num_writers to == 1 again.
2060 spin_lock(&fs_info->trans_lock);
2061 cur_trans->state = TRANS_STATE_COMMIT_DOING;
2062 spin_unlock(&fs_info->trans_lock);
2063 wait_event(cur_trans->writer_wait,
2064 atomic_read(&cur_trans->num_writers) == 1);
2066 /* ->aborted might be set after the previous check, so check it */
2067 if (unlikely(READ_ONCE(cur_trans->aborted))) {
2068 ret = cur_trans->aborted;
2069 goto scrub_continue;
2072 * the reloc mutex makes sure that we stop
2073 * the balancing code from coming in and moving
2074 * extents around in the middle of the commit
2076 mutex_lock(&fs_info->reloc_mutex);
2079 * We needn't worry about the delayed items because we will
2080 * deal with them in create_pending_snapshot(), which is the
2081 * core function of the snapshot creation.
2083 ret = create_pending_snapshots(trans);
2085 mutex_unlock(&fs_info->reloc_mutex);
2086 goto scrub_continue;
2090 * We insert the dir indexes of the snapshots and update the inode
2091 * of the snapshots' parents after the snapshot creation, so there
2092 * are some delayed items which are not dealt with. Now deal with
2095 * We needn't worry that this operation will corrupt the snapshots,
2096 * because all the tree which are snapshoted will be forced to COW
2097 * the nodes and leaves.
2099 ret = btrfs_run_delayed_items(trans);
2101 mutex_unlock(&fs_info->reloc_mutex);
2102 goto scrub_continue;
2105 ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
2107 mutex_unlock(&fs_info->reloc_mutex);
2108 goto scrub_continue;
2112 * make sure none of the code above managed to slip in a
2115 btrfs_assert_delayed_root_empty(fs_info);
2117 WARN_ON(cur_trans != trans->transaction);
2119 /* btrfs_commit_tree_roots is responsible for getting the
2120 * various roots consistent with each other. Every pointer
2121 * in the tree of tree roots has to point to the most up to date
2122 * root for every subvolume and other tree. So, we have to keep
2123 * the tree logging code from jumping in and changing any
2126 * At this point in the commit, there can't be any tree-log
2127 * writers, but a little lower down we drop the trans mutex
2128 * and let new people in. By holding the tree_log_mutex
2129 * from now until after the super is written, we avoid races
2130 * with the tree-log code.
2132 mutex_lock(&fs_info->tree_log_mutex);
2134 ret = commit_fs_roots(trans);
2136 mutex_unlock(&fs_info->tree_log_mutex);
2137 mutex_unlock(&fs_info->reloc_mutex);
2138 goto scrub_continue;
2142 * Since the transaction is done, we can apply the pending changes
2143 * before the next transaction.
2145 btrfs_apply_pending_changes(fs_info);
2147 /* commit_fs_roots gets rid of all the tree log roots, it is now
2148 * safe to free the root of tree log roots
2150 btrfs_free_log_root_tree(trans, fs_info);
2153 * commit_fs_roots() can call btrfs_save_ino_cache(), which generates
2154 * new delayed refs. Must handle them or qgroup can be wrong.
2156 ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
2158 mutex_unlock(&fs_info->tree_log_mutex);
2159 mutex_unlock(&fs_info->reloc_mutex);
2160 goto scrub_continue;
2164 * Since fs roots are all committed, we can get a quite accurate
2165 * new_roots. So let's do quota accounting.
2167 ret = btrfs_qgroup_account_extents(trans);
2169 mutex_unlock(&fs_info->tree_log_mutex);
2170 mutex_unlock(&fs_info->reloc_mutex);
2171 goto scrub_continue;
2174 ret = commit_cowonly_roots(trans);
2176 mutex_unlock(&fs_info->tree_log_mutex);
2177 mutex_unlock(&fs_info->reloc_mutex);
2178 goto scrub_continue;
2182 * The tasks which save the space cache and inode cache may also
2183 * update ->aborted, check it.
2185 if (unlikely(READ_ONCE(cur_trans->aborted))) {
2186 ret = cur_trans->aborted;
2187 mutex_unlock(&fs_info->tree_log_mutex);
2188 mutex_unlock(&fs_info->reloc_mutex);
2189 goto scrub_continue;
2192 btrfs_prepare_extent_commit(fs_info);
2194 cur_trans = fs_info->running_transaction;
2196 btrfs_set_root_node(&fs_info->tree_root->root_item,
2197 fs_info->tree_root->node);
2198 list_add_tail(&fs_info->tree_root->dirty_list,
2199 &cur_trans->switch_commits);
2201 btrfs_set_root_node(&fs_info->chunk_root->root_item,
2202 fs_info->chunk_root->node);
2203 list_add_tail(&fs_info->chunk_root->dirty_list,
2204 &cur_trans->switch_commits);
2206 switch_commit_roots(cur_trans);
2208 ASSERT(list_empty(&cur_trans->dirty_bgs));
2209 ASSERT(list_empty(&cur_trans->io_bgs));
2210 update_super_roots(fs_info);
2212 btrfs_set_super_log_root(fs_info->super_copy, 0);
2213 btrfs_set_super_log_root_level(fs_info->super_copy, 0);
2214 memcpy(fs_info->super_for_commit, fs_info->super_copy,
2215 sizeof(*fs_info->super_copy));
2217 btrfs_update_commit_device_size(fs_info);
2218 btrfs_update_commit_device_bytes_used(cur_trans);
2220 clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags);
2221 clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags);
2223 btrfs_trans_release_chunk_metadata(trans);
2225 spin_lock(&fs_info->trans_lock);
2226 cur_trans->state = TRANS_STATE_UNBLOCKED;
2227 fs_info->running_transaction = NULL;
2228 spin_unlock(&fs_info->trans_lock);
2229 mutex_unlock(&fs_info->reloc_mutex);
2231 wake_up(&fs_info->transaction_wait);
2233 ret = btrfs_write_and_wait_transaction(trans);
2235 btrfs_handle_fs_error(fs_info, ret,
2236 "Error while writing out transaction");
2237 mutex_unlock(&fs_info->tree_log_mutex);
2238 goto scrub_continue;
2241 ret = write_all_supers(fs_info, 0);
2243 * the super is written, we can safely allow the tree-loggers
2244 * to go about their business
2246 mutex_unlock(&fs_info->tree_log_mutex);
2248 goto scrub_continue;
2250 btrfs_finish_extent_commit(trans);
2252 if (test_bit(BTRFS_TRANS_HAVE_FREE_BGS, &cur_trans->flags))
2253 btrfs_clear_space_info_full(fs_info);
2255 fs_info->last_trans_committed = cur_trans->transid;
2257 * We needn't acquire the lock here because there is no other task
2258 * which can change it.
2260 cur_trans->state = TRANS_STATE_COMPLETED;
2261 wake_up(&cur_trans->commit_wait);
2262 clear_bit(BTRFS_FS_NEED_ASYNC_COMMIT, &fs_info->flags);
2264 spin_lock(&fs_info->trans_lock);
2265 list_del_init(&cur_trans->list);
2266 spin_unlock(&fs_info->trans_lock);
2268 btrfs_put_transaction(cur_trans);
2269 btrfs_put_transaction(cur_trans);
2271 if (trans->type & __TRANS_FREEZABLE)
2272 sb_end_intwrite(fs_info->sb);
2274 trace_btrfs_transaction_commit(trans->root);
2276 btrfs_scrub_continue(fs_info);
2278 if (current->journal_info == trans)
2279 current->journal_info = NULL;
2281 kmem_cache_free(btrfs_trans_handle_cachep, trans);
2284 * If fs has been frozen, we can not handle delayed iputs, otherwise
2285 * it'll result in deadlock about SB_FREEZE_FS.
2287 if (current != fs_info->transaction_kthread &&
2288 current != fs_info->cleaner_kthread &&
2289 !test_bit(BTRFS_FS_FROZEN, &fs_info->flags))
2290 btrfs_run_delayed_iputs(fs_info);
2295 btrfs_scrub_continue(fs_info);
2296 cleanup_transaction:
2297 btrfs_trans_release_metadata(trans);
2298 btrfs_trans_release_chunk_metadata(trans);
2299 trans->block_rsv = NULL;
2300 btrfs_warn(fs_info, "Skipping commit of aborted transaction.");
2301 if (current->journal_info == trans)
2302 current->journal_info = NULL;
2303 cleanup_transaction(trans, ret);
2309 * return < 0 if error
2310 * 0 if there are no more dead_roots at the time of call
2311 * 1 there are more to be processed, call me again
2313 * The return value indicates there are certainly more snapshots to delete, but
2314 * if there comes a new one during processing, it may return 0. We don't mind,
2315 * because btrfs_commit_super will poke cleaner thread and it will process it a
2316 * few seconds later.
2318 int btrfs_clean_one_deleted_snapshot(struct btrfs_root *root)
2321 struct btrfs_fs_info *fs_info = root->fs_info;
2323 spin_lock(&fs_info->trans_lock);
2324 if (list_empty(&fs_info->dead_roots)) {
2325 spin_unlock(&fs_info->trans_lock);
2328 root = list_first_entry(&fs_info->dead_roots,
2329 struct btrfs_root, root_list);
2330 list_del_init(&root->root_list);
2331 spin_unlock(&fs_info->trans_lock);
2333 btrfs_debug(fs_info, "cleaner removing %llu", root->root_key.objectid);
2335 btrfs_kill_all_delayed_nodes(root);
2337 if (btrfs_header_backref_rev(root->node) <
2338 BTRFS_MIXED_BACKREF_REV)
2339 ret = btrfs_drop_snapshot(root, NULL, 0, 0);
2341 ret = btrfs_drop_snapshot(root, NULL, 1, 0);
2343 return (ret < 0) ? 0 : 1;
2346 void btrfs_apply_pending_changes(struct btrfs_fs_info *fs_info)
2351 prev = xchg(&fs_info->pending_changes, 0);
2355 bit = 1 << BTRFS_PENDING_SET_INODE_MAP_CACHE;
2357 btrfs_set_opt(fs_info->mount_opt, INODE_MAP_CACHE);
2360 bit = 1 << BTRFS_PENDING_CLEAR_INODE_MAP_CACHE;
2362 btrfs_clear_opt(fs_info->mount_opt, INODE_MAP_CACHE);
2365 bit = 1 << BTRFS_PENDING_COMMIT;
2367 btrfs_debug(fs_info, "pending commit done");
2372 "unknown pending changes left 0x%lx, ignoring", prev);