1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 2011 Fujitsu. All rights reserved.
4 * Written by Miao Xie <miaox@cn.fujitsu.com>
7 #include <linux/slab.h>
8 #include <linux/iversion.h>
10 #include "delayed-inode.h"
12 #include "transaction.h"
16 #include "inode-item.h"
18 #define BTRFS_DELAYED_WRITEBACK 512
19 #define BTRFS_DELAYED_BACKGROUND 128
20 #define BTRFS_DELAYED_BATCH 16
22 static struct kmem_cache *delayed_node_cache;
24 int __init btrfs_delayed_inode_init(void)
26 delayed_node_cache = kmem_cache_create("btrfs_delayed_node",
27 sizeof(struct btrfs_delayed_node),
31 if (!delayed_node_cache)
36 void __cold btrfs_delayed_inode_exit(void)
38 kmem_cache_destroy(delayed_node_cache);
41 static inline void btrfs_init_delayed_node(
42 struct btrfs_delayed_node *delayed_node,
43 struct btrfs_root *root, u64 inode_id)
45 delayed_node->root = root;
46 delayed_node->inode_id = inode_id;
47 refcount_set(&delayed_node->refs, 0);
48 delayed_node->ins_root = RB_ROOT_CACHED;
49 delayed_node->del_root = RB_ROOT_CACHED;
50 mutex_init(&delayed_node->mutex);
51 INIT_LIST_HEAD(&delayed_node->n_list);
52 INIT_LIST_HEAD(&delayed_node->p_list);
55 static struct btrfs_delayed_node *btrfs_get_delayed_node(
56 struct btrfs_inode *btrfs_inode)
58 struct btrfs_root *root = btrfs_inode->root;
59 u64 ino = btrfs_ino(btrfs_inode);
60 struct btrfs_delayed_node *node;
62 node = READ_ONCE(btrfs_inode->delayed_node);
64 refcount_inc(&node->refs);
68 spin_lock(&root->inode_lock);
69 node = radix_tree_lookup(&root->delayed_nodes_tree, ino);
72 if (btrfs_inode->delayed_node) {
73 refcount_inc(&node->refs); /* can be accessed */
74 BUG_ON(btrfs_inode->delayed_node != node);
75 spin_unlock(&root->inode_lock);
80 * It's possible that we're racing into the middle of removing
81 * this node from the radix tree. In this case, the refcount
82 * was zero and it should never go back to one. Just return
83 * NULL like it was never in the radix at all; our release
84 * function is in the process of removing it.
86 * Some implementations of refcount_inc refuse to bump the
87 * refcount once it has hit zero. If we don't do this dance
88 * here, refcount_inc() may decide to just WARN_ONCE() instead
89 * of actually bumping the refcount.
91 * If this node is properly in the radix, we want to bump the
92 * refcount twice, once for the inode and once for this get
95 if (refcount_inc_not_zero(&node->refs)) {
96 refcount_inc(&node->refs);
97 btrfs_inode->delayed_node = node;
102 spin_unlock(&root->inode_lock);
105 spin_unlock(&root->inode_lock);
110 /* Will return either the node or PTR_ERR(-ENOMEM) */
111 static struct btrfs_delayed_node *btrfs_get_or_create_delayed_node(
112 struct btrfs_inode *btrfs_inode)
114 struct btrfs_delayed_node *node;
115 struct btrfs_root *root = btrfs_inode->root;
116 u64 ino = btrfs_ino(btrfs_inode);
120 node = btrfs_get_delayed_node(btrfs_inode);
124 node = kmem_cache_zalloc(delayed_node_cache, GFP_NOFS);
126 return ERR_PTR(-ENOMEM);
127 btrfs_init_delayed_node(node, root, ino);
129 /* cached in the btrfs inode and can be accessed */
130 refcount_set(&node->refs, 2);
132 ret = radix_tree_preload(GFP_NOFS);
134 kmem_cache_free(delayed_node_cache, node);
138 spin_lock(&root->inode_lock);
139 ret = radix_tree_insert(&root->delayed_nodes_tree, ino, node);
140 if (ret == -EEXIST) {
141 spin_unlock(&root->inode_lock);
142 kmem_cache_free(delayed_node_cache, node);
143 radix_tree_preload_end();
146 btrfs_inode->delayed_node = node;
147 spin_unlock(&root->inode_lock);
148 radix_tree_preload_end();
154 * Call it when holding delayed_node->mutex
156 * If mod = 1, add this node into the prepared list.
158 static void btrfs_queue_delayed_node(struct btrfs_delayed_root *root,
159 struct btrfs_delayed_node *node,
162 spin_lock(&root->lock);
163 if (test_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags)) {
164 if (!list_empty(&node->p_list))
165 list_move_tail(&node->p_list, &root->prepare_list);
167 list_add_tail(&node->p_list, &root->prepare_list);
169 list_add_tail(&node->n_list, &root->node_list);
170 list_add_tail(&node->p_list, &root->prepare_list);
171 refcount_inc(&node->refs); /* inserted into list */
173 set_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags);
175 spin_unlock(&root->lock);
178 /* Call it when holding delayed_node->mutex */
179 static void btrfs_dequeue_delayed_node(struct btrfs_delayed_root *root,
180 struct btrfs_delayed_node *node)
182 spin_lock(&root->lock);
183 if (test_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags)) {
185 refcount_dec(&node->refs); /* not in the list */
186 list_del_init(&node->n_list);
187 if (!list_empty(&node->p_list))
188 list_del_init(&node->p_list);
189 clear_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags);
191 spin_unlock(&root->lock);
194 static struct btrfs_delayed_node *btrfs_first_delayed_node(
195 struct btrfs_delayed_root *delayed_root)
198 struct btrfs_delayed_node *node = NULL;
200 spin_lock(&delayed_root->lock);
201 if (list_empty(&delayed_root->node_list))
204 p = delayed_root->node_list.next;
205 node = list_entry(p, struct btrfs_delayed_node, n_list);
206 refcount_inc(&node->refs);
208 spin_unlock(&delayed_root->lock);
213 static struct btrfs_delayed_node *btrfs_next_delayed_node(
214 struct btrfs_delayed_node *node)
216 struct btrfs_delayed_root *delayed_root;
218 struct btrfs_delayed_node *next = NULL;
220 delayed_root = node->root->fs_info->delayed_root;
221 spin_lock(&delayed_root->lock);
222 if (!test_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags)) {
223 /* not in the list */
224 if (list_empty(&delayed_root->node_list))
226 p = delayed_root->node_list.next;
227 } else if (list_is_last(&node->n_list, &delayed_root->node_list))
230 p = node->n_list.next;
232 next = list_entry(p, struct btrfs_delayed_node, n_list);
233 refcount_inc(&next->refs);
235 spin_unlock(&delayed_root->lock);
240 static void __btrfs_release_delayed_node(
241 struct btrfs_delayed_node *delayed_node,
244 struct btrfs_delayed_root *delayed_root;
249 delayed_root = delayed_node->root->fs_info->delayed_root;
251 mutex_lock(&delayed_node->mutex);
252 if (delayed_node->count)
253 btrfs_queue_delayed_node(delayed_root, delayed_node, mod);
255 btrfs_dequeue_delayed_node(delayed_root, delayed_node);
256 mutex_unlock(&delayed_node->mutex);
258 if (refcount_dec_and_test(&delayed_node->refs)) {
259 struct btrfs_root *root = delayed_node->root;
261 spin_lock(&root->inode_lock);
263 * Once our refcount goes to zero, nobody is allowed to bump it
264 * back up. We can delete it now.
266 ASSERT(refcount_read(&delayed_node->refs) == 0);
267 radix_tree_delete(&root->delayed_nodes_tree,
268 delayed_node->inode_id);
269 spin_unlock(&root->inode_lock);
270 kmem_cache_free(delayed_node_cache, delayed_node);
274 static inline void btrfs_release_delayed_node(struct btrfs_delayed_node *node)
276 __btrfs_release_delayed_node(node, 0);
279 static struct btrfs_delayed_node *btrfs_first_prepared_delayed_node(
280 struct btrfs_delayed_root *delayed_root)
283 struct btrfs_delayed_node *node = NULL;
285 spin_lock(&delayed_root->lock);
286 if (list_empty(&delayed_root->prepare_list))
289 p = delayed_root->prepare_list.next;
291 node = list_entry(p, struct btrfs_delayed_node, p_list);
292 refcount_inc(&node->refs);
294 spin_unlock(&delayed_root->lock);
299 static inline void btrfs_release_prepared_delayed_node(
300 struct btrfs_delayed_node *node)
302 __btrfs_release_delayed_node(node, 1);
305 static struct btrfs_delayed_item *btrfs_alloc_delayed_item(u16 data_len,
306 struct btrfs_delayed_node *node,
307 enum btrfs_delayed_item_type type)
309 struct btrfs_delayed_item *item;
311 item = kmalloc(sizeof(*item) + data_len, GFP_NOFS);
313 item->data_len = data_len;
315 item->bytes_reserved = 0;
316 item->delayed_node = node;
317 RB_CLEAR_NODE(&item->rb_node);
318 INIT_LIST_HEAD(&item->log_list);
319 item->logged = false;
320 refcount_set(&item->refs, 1);
326 * __btrfs_lookup_delayed_item - look up the delayed item by key
327 * @delayed_node: pointer to the delayed node
328 * @index: the dir index value to lookup (offset of a dir index key)
330 * Note: if we don't find the right item, we will return the prev item and
333 static struct btrfs_delayed_item *__btrfs_lookup_delayed_item(
334 struct rb_root *root,
337 struct rb_node *node = root->rb_node;
338 struct btrfs_delayed_item *delayed_item = NULL;
341 delayed_item = rb_entry(node, struct btrfs_delayed_item,
343 if (delayed_item->index < index)
344 node = node->rb_right;
345 else if (delayed_item->index > index)
346 node = node->rb_left;
354 static int __btrfs_add_delayed_item(struct btrfs_delayed_node *delayed_node,
355 struct btrfs_delayed_item *ins)
357 struct rb_node **p, *node;
358 struct rb_node *parent_node = NULL;
359 struct rb_root_cached *root;
360 struct btrfs_delayed_item *item;
361 bool leftmost = true;
363 if (ins->type == BTRFS_DELAYED_INSERTION_ITEM)
364 root = &delayed_node->ins_root;
366 root = &delayed_node->del_root;
368 p = &root->rb_root.rb_node;
369 node = &ins->rb_node;
373 item = rb_entry(parent_node, struct btrfs_delayed_item,
376 if (item->index < ins->index) {
379 } else if (item->index > ins->index) {
386 rb_link_node(node, parent_node, p);
387 rb_insert_color_cached(node, root, leftmost);
389 if (ins->type == BTRFS_DELAYED_INSERTION_ITEM &&
390 ins->index >= delayed_node->index_cnt)
391 delayed_node->index_cnt = ins->index + 1;
393 delayed_node->count++;
394 atomic_inc(&delayed_node->root->fs_info->delayed_root->items);
398 static void finish_one_item(struct btrfs_delayed_root *delayed_root)
400 int seq = atomic_inc_return(&delayed_root->items_seq);
402 /* atomic_dec_return implies a barrier */
403 if ((atomic_dec_return(&delayed_root->items) <
404 BTRFS_DELAYED_BACKGROUND || seq % BTRFS_DELAYED_BATCH == 0))
405 cond_wake_up_nomb(&delayed_root->wait);
408 static void __btrfs_remove_delayed_item(struct btrfs_delayed_item *delayed_item)
410 struct btrfs_delayed_node *delayed_node = delayed_item->delayed_node;
411 struct rb_root_cached *root;
412 struct btrfs_delayed_root *delayed_root;
414 /* Not inserted, ignore it. */
415 if (RB_EMPTY_NODE(&delayed_item->rb_node))
418 /* If it's in a rbtree, then we need to have delayed node locked. */
419 lockdep_assert_held(&delayed_node->mutex);
421 delayed_root = delayed_node->root->fs_info->delayed_root;
423 BUG_ON(!delayed_root);
425 if (delayed_item->type == BTRFS_DELAYED_INSERTION_ITEM)
426 root = &delayed_node->ins_root;
428 root = &delayed_node->del_root;
430 rb_erase_cached(&delayed_item->rb_node, root);
431 RB_CLEAR_NODE(&delayed_item->rb_node);
432 delayed_node->count--;
434 finish_one_item(delayed_root);
437 static void btrfs_release_delayed_item(struct btrfs_delayed_item *item)
440 __btrfs_remove_delayed_item(item);
441 if (refcount_dec_and_test(&item->refs))
446 static struct btrfs_delayed_item *__btrfs_first_delayed_insertion_item(
447 struct btrfs_delayed_node *delayed_node)
450 struct btrfs_delayed_item *item = NULL;
452 p = rb_first_cached(&delayed_node->ins_root);
454 item = rb_entry(p, struct btrfs_delayed_item, rb_node);
459 static struct btrfs_delayed_item *__btrfs_first_delayed_deletion_item(
460 struct btrfs_delayed_node *delayed_node)
463 struct btrfs_delayed_item *item = NULL;
465 p = rb_first_cached(&delayed_node->del_root);
467 item = rb_entry(p, struct btrfs_delayed_item, rb_node);
472 static struct btrfs_delayed_item *__btrfs_next_delayed_item(
473 struct btrfs_delayed_item *item)
476 struct btrfs_delayed_item *next = NULL;
478 p = rb_next(&item->rb_node);
480 next = rb_entry(p, struct btrfs_delayed_item, rb_node);
485 static int btrfs_delayed_item_reserve_metadata(struct btrfs_trans_handle *trans,
486 struct btrfs_delayed_item *item)
488 struct btrfs_block_rsv *src_rsv;
489 struct btrfs_block_rsv *dst_rsv;
490 struct btrfs_fs_info *fs_info = trans->fs_info;
494 if (!trans->bytes_reserved)
497 src_rsv = trans->block_rsv;
498 dst_rsv = &fs_info->delayed_block_rsv;
500 num_bytes = btrfs_calc_insert_metadata_size(fs_info, 1);
503 * Here we migrate space rsv from transaction rsv, since have already
504 * reserved space when starting a transaction. So no need to reserve
507 ret = btrfs_block_rsv_migrate(src_rsv, dst_rsv, num_bytes, true);
509 trace_btrfs_space_reservation(fs_info, "delayed_item",
510 item->delayed_node->inode_id,
513 * For insertions we track reserved metadata space by accounting
514 * for the number of leaves that will be used, based on the delayed
515 * node's index_items_size field.
517 if (item->type == BTRFS_DELAYED_DELETION_ITEM)
518 item->bytes_reserved = num_bytes;
524 static void btrfs_delayed_item_release_metadata(struct btrfs_root *root,
525 struct btrfs_delayed_item *item)
527 struct btrfs_block_rsv *rsv;
528 struct btrfs_fs_info *fs_info = root->fs_info;
530 if (!item->bytes_reserved)
533 rsv = &fs_info->delayed_block_rsv;
535 * Check btrfs_delayed_item_reserve_metadata() to see why we don't need
536 * to release/reserve qgroup space.
538 trace_btrfs_space_reservation(fs_info, "delayed_item",
539 item->delayed_node->inode_id,
540 item->bytes_reserved, 0);
541 btrfs_block_rsv_release(fs_info, rsv, item->bytes_reserved, NULL);
544 static void btrfs_delayed_item_release_leaves(struct btrfs_delayed_node *node,
545 unsigned int num_leaves)
547 struct btrfs_fs_info *fs_info = node->root->fs_info;
548 const u64 bytes = btrfs_calc_insert_metadata_size(fs_info, num_leaves);
550 /* There are no space reservations during log replay, bail out. */
551 if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
554 trace_btrfs_space_reservation(fs_info, "delayed_item", node->inode_id,
556 btrfs_block_rsv_release(fs_info, &fs_info->delayed_block_rsv, bytes, NULL);
559 static int btrfs_delayed_inode_reserve_metadata(
560 struct btrfs_trans_handle *trans,
561 struct btrfs_root *root,
562 struct btrfs_delayed_node *node)
564 struct btrfs_fs_info *fs_info = root->fs_info;
565 struct btrfs_block_rsv *src_rsv;
566 struct btrfs_block_rsv *dst_rsv;
570 src_rsv = trans->block_rsv;
571 dst_rsv = &fs_info->delayed_block_rsv;
573 num_bytes = btrfs_calc_metadata_size(fs_info, 1);
576 * btrfs_dirty_inode will update the inode under btrfs_join_transaction
577 * which doesn't reserve space for speed. This is a problem since we
578 * still need to reserve space for this update, so try to reserve the
581 * Now if src_rsv == delalloc_block_rsv we'll let it just steal since
582 * we always reserve enough to update the inode item.
584 if (!src_rsv || (!trans->bytes_reserved &&
585 src_rsv->type != BTRFS_BLOCK_RSV_DELALLOC)) {
586 ret = btrfs_qgroup_reserve_meta(root, num_bytes,
587 BTRFS_QGROUP_RSV_META_PREALLOC, true);
590 ret = btrfs_block_rsv_add(fs_info, dst_rsv, num_bytes,
591 BTRFS_RESERVE_NO_FLUSH);
592 /* NO_FLUSH could only fail with -ENOSPC */
593 ASSERT(ret == 0 || ret == -ENOSPC);
595 btrfs_qgroup_free_meta_prealloc(root, num_bytes);
597 ret = btrfs_block_rsv_migrate(src_rsv, dst_rsv, num_bytes, true);
601 trace_btrfs_space_reservation(fs_info, "delayed_inode",
602 node->inode_id, num_bytes, 1);
603 node->bytes_reserved = num_bytes;
609 static void btrfs_delayed_inode_release_metadata(struct btrfs_fs_info *fs_info,
610 struct btrfs_delayed_node *node,
613 struct btrfs_block_rsv *rsv;
615 if (!node->bytes_reserved)
618 rsv = &fs_info->delayed_block_rsv;
619 trace_btrfs_space_reservation(fs_info, "delayed_inode",
620 node->inode_id, node->bytes_reserved, 0);
621 btrfs_block_rsv_release(fs_info, rsv, node->bytes_reserved, NULL);
623 btrfs_qgroup_free_meta_prealloc(node->root,
624 node->bytes_reserved);
626 btrfs_qgroup_convert_reserved_meta(node->root,
627 node->bytes_reserved);
628 node->bytes_reserved = 0;
632 * Insert a single delayed item or a batch of delayed items, as many as possible
633 * that fit in a leaf. The delayed items (dir index keys) are sorted by their key
634 * in the rbtree, and if there's a gap between two consecutive dir index items,
635 * then it means at some point we had delayed dir indexes to add but they got
636 * removed (by btrfs_delete_delayed_dir_index()) before we attempted to flush them
637 * into the subvolume tree. Dir index keys also have their offsets coming from a
638 * monotonically increasing counter, so we can't get new keys with an offset that
639 * fits within a gap between delayed dir index items.
641 static int btrfs_insert_delayed_item(struct btrfs_trans_handle *trans,
642 struct btrfs_root *root,
643 struct btrfs_path *path,
644 struct btrfs_delayed_item *first_item)
646 struct btrfs_fs_info *fs_info = root->fs_info;
647 struct btrfs_delayed_node *node = first_item->delayed_node;
648 LIST_HEAD(item_list);
649 struct btrfs_delayed_item *curr;
650 struct btrfs_delayed_item *next;
651 const int max_size = BTRFS_LEAF_DATA_SIZE(fs_info);
652 struct btrfs_item_batch batch;
653 struct btrfs_key first_key;
654 const u32 first_data_size = first_item->data_len;
656 char *ins_data = NULL;
658 bool continuous_keys_only = false;
660 lockdep_assert_held(&node->mutex);
663 * During normal operation the delayed index offset is continuously
664 * increasing, so we can batch insert all items as there will not be any
665 * overlapping keys in the tree.
667 * The exception to this is log replay, where we may have interleaved
668 * offsets in the tree, so our batch needs to be continuous keys only in
669 * order to ensure we do not end up with out of order items in our leaf.
671 if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
672 continuous_keys_only = true;
675 * For delayed items to insert, we track reserved metadata bytes based
676 * on the number of leaves that we will use.
677 * See btrfs_insert_delayed_dir_index() and
678 * btrfs_delayed_item_reserve_metadata()).
680 ASSERT(first_item->bytes_reserved == 0);
682 list_add_tail(&first_item->tree_list, &item_list);
683 batch.total_data_size = first_data_size;
685 total_size = first_data_size + sizeof(struct btrfs_item);
691 next = __btrfs_next_delayed_item(curr);
696 * We cannot allow gaps in the key space if we're doing log
699 if (continuous_keys_only && (next->index != curr->index + 1))
702 ASSERT(next->bytes_reserved == 0);
704 next_size = next->data_len + sizeof(struct btrfs_item);
705 if (total_size + next_size > max_size)
708 list_add_tail(&next->tree_list, &item_list);
710 total_size += next_size;
711 batch.total_data_size += next->data_len;
716 first_key.objectid = node->inode_id;
717 first_key.type = BTRFS_DIR_INDEX_KEY;
718 first_key.offset = first_item->index;
719 batch.keys = &first_key;
720 batch.data_sizes = &first_data_size;
722 struct btrfs_key *ins_keys;
726 ins_data = kmalloc(batch.nr * sizeof(u32) +
727 batch.nr * sizeof(struct btrfs_key), GFP_NOFS);
732 ins_sizes = (u32 *)ins_data;
733 ins_keys = (struct btrfs_key *)(ins_data + batch.nr * sizeof(u32));
734 batch.keys = ins_keys;
735 batch.data_sizes = ins_sizes;
736 list_for_each_entry(curr, &item_list, tree_list) {
737 ins_keys[i].objectid = node->inode_id;
738 ins_keys[i].type = BTRFS_DIR_INDEX_KEY;
739 ins_keys[i].offset = curr->index;
740 ins_sizes[i] = curr->data_len;
745 ret = btrfs_insert_empty_items(trans, root, path, &batch);
749 list_for_each_entry(curr, &item_list, tree_list) {
752 data_ptr = btrfs_item_ptr(path->nodes[0], path->slots[0], char);
753 write_extent_buffer(path->nodes[0], &curr->data,
754 (unsigned long)data_ptr, curr->data_len);
759 * Now release our path before releasing the delayed items and their
760 * metadata reservations, so that we don't block other tasks for more
763 btrfs_release_path(path);
765 ASSERT(node->index_item_leaves > 0);
768 * For normal operations we will batch an entire leaf's worth of delayed
769 * items, so if there are more items to process we can decrement
770 * index_item_leaves by 1 as we inserted 1 leaf's worth of items.
772 * However for log replay we may not have inserted an entire leaf's
773 * worth of items, we may have not had continuous items, so decrementing
774 * here would mess up the index_item_leaves accounting. For this case
775 * only clean up the accounting when there are no items left.
777 if (next && !continuous_keys_only) {
779 * We inserted one batch of items into a leaf a there are more
780 * items to flush in a future batch, now release one unit of
781 * metadata space from the delayed block reserve, corresponding
782 * the leaf we just flushed to.
784 btrfs_delayed_item_release_leaves(node, 1);
785 node->index_item_leaves--;
788 * There are no more items to insert. We can have a number of
789 * reserved leaves > 1 here - this happens when many dir index
790 * items are added and then removed before they are flushed (file
791 * names with a very short life, never span a transaction). So
792 * release all remaining leaves.
794 btrfs_delayed_item_release_leaves(node, node->index_item_leaves);
795 node->index_item_leaves = 0;
798 list_for_each_entry_safe(curr, next, &item_list, tree_list) {
799 list_del(&curr->tree_list);
800 btrfs_release_delayed_item(curr);
807 static int btrfs_insert_delayed_items(struct btrfs_trans_handle *trans,
808 struct btrfs_path *path,
809 struct btrfs_root *root,
810 struct btrfs_delayed_node *node)
815 struct btrfs_delayed_item *curr;
817 mutex_lock(&node->mutex);
818 curr = __btrfs_first_delayed_insertion_item(node);
820 mutex_unlock(&node->mutex);
823 ret = btrfs_insert_delayed_item(trans, root, path, curr);
824 mutex_unlock(&node->mutex);
830 static int btrfs_batch_delete_items(struct btrfs_trans_handle *trans,
831 struct btrfs_root *root,
832 struct btrfs_path *path,
833 struct btrfs_delayed_item *item)
835 const u64 ino = item->delayed_node->inode_id;
836 struct btrfs_fs_info *fs_info = root->fs_info;
837 struct btrfs_delayed_item *curr, *next;
838 struct extent_buffer *leaf = path->nodes[0];
839 LIST_HEAD(batch_list);
840 int nitems, slot, last_slot;
842 u64 total_reserved_size = item->bytes_reserved;
844 ASSERT(leaf != NULL);
846 slot = path->slots[0];
847 last_slot = btrfs_header_nritems(leaf) - 1;
849 * Our caller always gives us a path pointing to an existing item, so
850 * this can not happen.
852 ASSERT(slot <= last_slot);
853 if (WARN_ON(slot > last_slot))
858 list_add_tail(&curr->tree_list, &batch_list);
861 * Keep checking if the next delayed item matches the next item in the
862 * leaf - if so, we can add it to the batch of items to delete from the
865 while (slot < last_slot) {
866 struct btrfs_key key;
868 next = __btrfs_next_delayed_item(curr);
873 btrfs_item_key_to_cpu(leaf, &key, slot);
874 if (key.objectid != ino ||
875 key.type != BTRFS_DIR_INDEX_KEY ||
876 key.offset != next->index)
880 list_add_tail(&curr->tree_list, &batch_list);
881 total_reserved_size += curr->bytes_reserved;
884 ret = btrfs_del_items(trans, root, path, path->slots[0], nitems);
888 /* In case of BTRFS_FS_LOG_RECOVERING items won't have reserved space */
889 if (total_reserved_size > 0) {
891 * Check btrfs_delayed_item_reserve_metadata() to see why we
892 * don't need to release/reserve qgroup space.
894 trace_btrfs_space_reservation(fs_info, "delayed_item", ino,
895 total_reserved_size, 0);
896 btrfs_block_rsv_release(fs_info, &fs_info->delayed_block_rsv,
897 total_reserved_size, NULL);
900 list_for_each_entry_safe(curr, next, &batch_list, tree_list) {
901 list_del(&curr->tree_list);
902 btrfs_release_delayed_item(curr);
908 static int btrfs_delete_delayed_items(struct btrfs_trans_handle *trans,
909 struct btrfs_path *path,
910 struct btrfs_root *root,
911 struct btrfs_delayed_node *node)
913 struct btrfs_key key;
916 key.objectid = node->inode_id;
917 key.type = BTRFS_DIR_INDEX_KEY;
920 struct btrfs_delayed_item *item;
922 mutex_lock(&node->mutex);
923 item = __btrfs_first_delayed_deletion_item(node);
925 mutex_unlock(&node->mutex);
929 key.offset = item->index;
930 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
933 * There's no matching item in the leaf. This means we
934 * have already deleted this item in a past run of the
935 * delayed items. We ignore errors when running delayed
936 * items from an async context, through a work queue job
937 * running btrfs_async_run_delayed_root(), and don't
938 * release delayed items that failed to complete. This
939 * is because we will retry later, and at transaction
940 * commit time we always run delayed items and will
941 * then deal with errors if they fail to run again.
943 * So just release delayed items for which we can't find
944 * an item in the tree, and move to the next item.
946 btrfs_release_path(path);
947 btrfs_release_delayed_item(item);
949 } else if (ret == 0) {
950 ret = btrfs_batch_delete_items(trans, root, path, item);
951 btrfs_release_path(path);
955 * We unlock and relock on each iteration, this is to prevent
956 * blocking other tasks for too long while we are being run from
957 * the async context (work queue job). Those tasks are typically
958 * running system calls like creat/mkdir/rename/unlink/etc which
959 * need to add delayed items to this delayed node.
961 mutex_unlock(&node->mutex);
967 static void btrfs_release_delayed_inode(struct btrfs_delayed_node *delayed_node)
969 struct btrfs_delayed_root *delayed_root;
972 test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
973 BUG_ON(!delayed_node->root);
974 clear_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags);
975 delayed_node->count--;
977 delayed_root = delayed_node->root->fs_info->delayed_root;
978 finish_one_item(delayed_root);
982 static void btrfs_release_delayed_iref(struct btrfs_delayed_node *delayed_node)
985 if (test_and_clear_bit(BTRFS_DELAYED_NODE_DEL_IREF, &delayed_node->flags)) {
986 struct btrfs_delayed_root *delayed_root;
988 ASSERT(delayed_node->root);
989 delayed_node->count--;
991 delayed_root = delayed_node->root->fs_info->delayed_root;
992 finish_one_item(delayed_root);
996 static int __btrfs_update_delayed_inode(struct btrfs_trans_handle *trans,
997 struct btrfs_root *root,
998 struct btrfs_path *path,
999 struct btrfs_delayed_node *node)
1001 struct btrfs_fs_info *fs_info = root->fs_info;
1002 struct btrfs_key key;
1003 struct btrfs_inode_item *inode_item;
1004 struct extent_buffer *leaf;
1008 key.objectid = node->inode_id;
1009 key.type = BTRFS_INODE_ITEM_KEY;
1012 if (test_bit(BTRFS_DELAYED_NODE_DEL_IREF, &node->flags))
1017 ret = btrfs_lookup_inode(trans, root, path, &key, mod);
1023 leaf = path->nodes[0];
1024 inode_item = btrfs_item_ptr(leaf, path->slots[0],
1025 struct btrfs_inode_item);
1026 write_extent_buffer(leaf, &node->inode_item, (unsigned long)inode_item,
1027 sizeof(struct btrfs_inode_item));
1028 btrfs_mark_buffer_dirty(leaf);
1030 if (!test_bit(BTRFS_DELAYED_NODE_DEL_IREF, &node->flags))
1034 if (path->slots[0] >= btrfs_header_nritems(leaf))
1037 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1038 if (key.objectid != node->inode_id)
1041 if (key.type != BTRFS_INODE_REF_KEY &&
1042 key.type != BTRFS_INODE_EXTREF_KEY)
1046 * Delayed iref deletion is for the inode who has only one link,
1047 * so there is only one iref. The case that several irefs are
1048 * in the same item doesn't exist.
1050 btrfs_del_item(trans, root, path);
1052 btrfs_release_delayed_iref(node);
1053 btrfs_release_path(path);
1055 btrfs_delayed_inode_release_metadata(fs_info, node, (ret < 0));
1056 btrfs_release_delayed_inode(node);
1059 * If we fail to update the delayed inode we need to abort the
1060 * transaction, because we could leave the inode with the improper
1063 if (ret && ret != -ENOENT)
1064 btrfs_abort_transaction(trans, ret);
1069 btrfs_release_path(path);
1071 key.type = BTRFS_INODE_EXTREF_KEY;
1074 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1080 leaf = path->nodes[0];
1085 static inline int btrfs_update_delayed_inode(struct btrfs_trans_handle *trans,
1086 struct btrfs_root *root,
1087 struct btrfs_path *path,
1088 struct btrfs_delayed_node *node)
1092 mutex_lock(&node->mutex);
1093 if (!test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &node->flags)) {
1094 mutex_unlock(&node->mutex);
1098 ret = __btrfs_update_delayed_inode(trans, root, path, node);
1099 mutex_unlock(&node->mutex);
1104 __btrfs_commit_inode_delayed_items(struct btrfs_trans_handle *trans,
1105 struct btrfs_path *path,
1106 struct btrfs_delayed_node *node)
1110 ret = btrfs_insert_delayed_items(trans, path, node->root, node);
1114 ret = btrfs_delete_delayed_items(trans, path, node->root, node);
1118 ret = btrfs_record_root_in_trans(trans, node->root);
1121 ret = btrfs_update_delayed_inode(trans, node->root, path, node);
1126 * Called when committing the transaction.
1127 * Returns 0 on success.
1128 * Returns < 0 on error and returns with an aborted transaction with any
1129 * outstanding delayed items cleaned up.
1131 static int __btrfs_run_delayed_items(struct btrfs_trans_handle *trans, int nr)
1133 struct btrfs_fs_info *fs_info = trans->fs_info;
1134 struct btrfs_delayed_root *delayed_root;
1135 struct btrfs_delayed_node *curr_node, *prev_node;
1136 struct btrfs_path *path;
1137 struct btrfs_block_rsv *block_rsv;
1139 bool count = (nr > 0);
1141 if (TRANS_ABORTED(trans))
1144 path = btrfs_alloc_path();
1148 block_rsv = trans->block_rsv;
1149 trans->block_rsv = &fs_info->delayed_block_rsv;
1151 delayed_root = fs_info->delayed_root;
1153 curr_node = btrfs_first_delayed_node(delayed_root);
1154 while (curr_node && (!count || nr--)) {
1155 ret = __btrfs_commit_inode_delayed_items(trans, path,
1158 btrfs_abort_transaction(trans, ret);
1162 prev_node = curr_node;
1163 curr_node = btrfs_next_delayed_node(curr_node);
1165 * See the comment below about releasing path before releasing
1166 * node. If the commit of delayed items was successful the path
1167 * should always be released, but in case of an error, it may
1168 * point to locked extent buffers (a leaf at the very least).
1170 ASSERT(path->nodes[0] == NULL);
1171 btrfs_release_delayed_node(prev_node);
1175 * Release the path to avoid a potential deadlock and lockdep splat when
1176 * releasing the delayed node, as that requires taking the delayed node's
1177 * mutex. If another task starts running delayed items before we take
1178 * the mutex, it will first lock the mutex and then it may try to lock
1179 * the same btree path (leaf).
1181 btrfs_free_path(path);
1184 btrfs_release_delayed_node(curr_node);
1185 trans->block_rsv = block_rsv;
1190 int btrfs_run_delayed_items(struct btrfs_trans_handle *trans)
1192 return __btrfs_run_delayed_items(trans, -1);
1195 int btrfs_run_delayed_items_nr(struct btrfs_trans_handle *trans, int nr)
1197 return __btrfs_run_delayed_items(trans, nr);
1200 int btrfs_commit_inode_delayed_items(struct btrfs_trans_handle *trans,
1201 struct btrfs_inode *inode)
1203 struct btrfs_delayed_node *delayed_node = btrfs_get_delayed_node(inode);
1204 struct btrfs_path *path;
1205 struct btrfs_block_rsv *block_rsv;
1211 mutex_lock(&delayed_node->mutex);
1212 if (!delayed_node->count) {
1213 mutex_unlock(&delayed_node->mutex);
1214 btrfs_release_delayed_node(delayed_node);
1217 mutex_unlock(&delayed_node->mutex);
1219 path = btrfs_alloc_path();
1221 btrfs_release_delayed_node(delayed_node);
1225 block_rsv = trans->block_rsv;
1226 trans->block_rsv = &delayed_node->root->fs_info->delayed_block_rsv;
1228 ret = __btrfs_commit_inode_delayed_items(trans, path, delayed_node);
1230 btrfs_release_delayed_node(delayed_node);
1231 btrfs_free_path(path);
1232 trans->block_rsv = block_rsv;
1237 int btrfs_commit_inode_delayed_inode(struct btrfs_inode *inode)
1239 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1240 struct btrfs_trans_handle *trans;
1241 struct btrfs_delayed_node *delayed_node = btrfs_get_delayed_node(inode);
1242 struct btrfs_path *path;
1243 struct btrfs_block_rsv *block_rsv;
1249 mutex_lock(&delayed_node->mutex);
1250 if (!test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
1251 mutex_unlock(&delayed_node->mutex);
1252 btrfs_release_delayed_node(delayed_node);
1255 mutex_unlock(&delayed_node->mutex);
1257 trans = btrfs_join_transaction(delayed_node->root);
1258 if (IS_ERR(trans)) {
1259 ret = PTR_ERR(trans);
1263 path = btrfs_alloc_path();
1269 block_rsv = trans->block_rsv;
1270 trans->block_rsv = &fs_info->delayed_block_rsv;
1272 mutex_lock(&delayed_node->mutex);
1273 if (test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags))
1274 ret = __btrfs_update_delayed_inode(trans, delayed_node->root,
1275 path, delayed_node);
1278 mutex_unlock(&delayed_node->mutex);
1280 btrfs_free_path(path);
1281 trans->block_rsv = block_rsv;
1283 btrfs_end_transaction(trans);
1284 btrfs_btree_balance_dirty(fs_info);
1286 btrfs_release_delayed_node(delayed_node);
1291 void btrfs_remove_delayed_node(struct btrfs_inode *inode)
1293 struct btrfs_delayed_node *delayed_node;
1295 delayed_node = READ_ONCE(inode->delayed_node);
1299 inode->delayed_node = NULL;
1300 btrfs_release_delayed_node(delayed_node);
1303 struct btrfs_async_delayed_work {
1304 struct btrfs_delayed_root *delayed_root;
1306 struct btrfs_work work;
1309 static void btrfs_async_run_delayed_root(struct btrfs_work *work)
1311 struct btrfs_async_delayed_work *async_work;
1312 struct btrfs_delayed_root *delayed_root;
1313 struct btrfs_trans_handle *trans;
1314 struct btrfs_path *path;
1315 struct btrfs_delayed_node *delayed_node = NULL;
1316 struct btrfs_root *root;
1317 struct btrfs_block_rsv *block_rsv;
1320 async_work = container_of(work, struct btrfs_async_delayed_work, work);
1321 delayed_root = async_work->delayed_root;
1323 path = btrfs_alloc_path();
1328 if (atomic_read(&delayed_root->items) <
1329 BTRFS_DELAYED_BACKGROUND / 2)
1332 delayed_node = btrfs_first_prepared_delayed_node(delayed_root);
1336 root = delayed_node->root;
1338 trans = btrfs_join_transaction(root);
1339 if (IS_ERR(trans)) {
1340 btrfs_release_path(path);
1341 btrfs_release_prepared_delayed_node(delayed_node);
1346 block_rsv = trans->block_rsv;
1347 trans->block_rsv = &root->fs_info->delayed_block_rsv;
1349 __btrfs_commit_inode_delayed_items(trans, path, delayed_node);
1351 trans->block_rsv = block_rsv;
1352 btrfs_end_transaction(trans);
1353 btrfs_btree_balance_dirty_nodelay(root->fs_info);
1355 btrfs_release_path(path);
1356 btrfs_release_prepared_delayed_node(delayed_node);
1359 } while ((async_work->nr == 0 && total_done < BTRFS_DELAYED_WRITEBACK)
1360 || total_done < async_work->nr);
1362 btrfs_free_path(path);
1364 wake_up(&delayed_root->wait);
1369 static int btrfs_wq_run_delayed_node(struct btrfs_delayed_root *delayed_root,
1370 struct btrfs_fs_info *fs_info, int nr)
1372 struct btrfs_async_delayed_work *async_work;
1374 async_work = kmalloc(sizeof(*async_work), GFP_NOFS);
1378 async_work->delayed_root = delayed_root;
1379 btrfs_init_work(&async_work->work, btrfs_async_run_delayed_root, NULL,
1381 async_work->nr = nr;
1383 btrfs_queue_work(fs_info->delayed_workers, &async_work->work);
1387 void btrfs_assert_delayed_root_empty(struct btrfs_fs_info *fs_info)
1389 WARN_ON(btrfs_first_delayed_node(fs_info->delayed_root));
1392 static int could_end_wait(struct btrfs_delayed_root *delayed_root, int seq)
1394 int val = atomic_read(&delayed_root->items_seq);
1396 if (val < seq || val >= seq + BTRFS_DELAYED_BATCH)
1399 if (atomic_read(&delayed_root->items) < BTRFS_DELAYED_BACKGROUND)
1405 void btrfs_balance_delayed_items(struct btrfs_fs_info *fs_info)
1407 struct btrfs_delayed_root *delayed_root = fs_info->delayed_root;
1409 if ((atomic_read(&delayed_root->items) < BTRFS_DELAYED_BACKGROUND) ||
1410 btrfs_workqueue_normal_congested(fs_info->delayed_workers))
1413 if (atomic_read(&delayed_root->items) >= BTRFS_DELAYED_WRITEBACK) {
1417 seq = atomic_read(&delayed_root->items_seq);
1419 ret = btrfs_wq_run_delayed_node(delayed_root, fs_info, 0);
1423 wait_event_interruptible(delayed_root->wait,
1424 could_end_wait(delayed_root, seq));
1428 btrfs_wq_run_delayed_node(delayed_root, fs_info, BTRFS_DELAYED_BATCH);
1431 static void btrfs_release_dir_index_item_space(struct btrfs_trans_handle *trans)
1433 struct btrfs_fs_info *fs_info = trans->fs_info;
1434 const u64 bytes = btrfs_calc_insert_metadata_size(fs_info, 1);
1436 if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
1440 * Adding the new dir index item does not require touching another
1441 * leaf, so we can release 1 unit of metadata that was previously
1442 * reserved when starting the transaction. This applies only to
1443 * the case where we had a transaction start and excludes the
1444 * transaction join case (when replaying log trees).
1446 trace_btrfs_space_reservation(fs_info, "transaction",
1447 trans->transid, bytes, 0);
1448 btrfs_block_rsv_release(fs_info, trans->block_rsv, bytes, NULL);
1449 ASSERT(trans->bytes_reserved >= bytes);
1450 trans->bytes_reserved -= bytes;
1453 /* Will return 0, -ENOMEM or -EEXIST (index number collision, unexpected). */
1454 int btrfs_insert_delayed_dir_index(struct btrfs_trans_handle *trans,
1455 const char *name, int name_len,
1456 struct btrfs_inode *dir,
1457 struct btrfs_disk_key *disk_key, u8 type,
1460 struct btrfs_fs_info *fs_info = trans->fs_info;
1461 const unsigned int leaf_data_size = BTRFS_LEAF_DATA_SIZE(fs_info);
1462 struct btrfs_delayed_node *delayed_node;
1463 struct btrfs_delayed_item *delayed_item;
1464 struct btrfs_dir_item *dir_item;
1465 bool reserve_leaf_space;
1469 delayed_node = btrfs_get_or_create_delayed_node(dir);
1470 if (IS_ERR(delayed_node))
1471 return PTR_ERR(delayed_node);
1473 delayed_item = btrfs_alloc_delayed_item(sizeof(*dir_item) + name_len,
1475 BTRFS_DELAYED_INSERTION_ITEM);
1476 if (!delayed_item) {
1481 delayed_item->index = index;
1483 dir_item = (struct btrfs_dir_item *)delayed_item->data;
1484 dir_item->location = *disk_key;
1485 btrfs_set_stack_dir_transid(dir_item, trans->transid);
1486 btrfs_set_stack_dir_data_len(dir_item, 0);
1487 btrfs_set_stack_dir_name_len(dir_item, name_len);
1488 btrfs_set_stack_dir_type(dir_item, type);
1489 memcpy((char *)(dir_item + 1), name, name_len);
1491 data_len = delayed_item->data_len + sizeof(struct btrfs_item);
1493 mutex_lock(&delayed_node->mutex);
1496 * First attempt to insert the delayed item. This is to make the error
1497 * handling path simpler in case we fail (-EEXIST). There's no risk of
1498 * any other task coming in and running the delayed item before we do
1499 * the metadata space reservation below, because we are holding the
1500 * delayed node's mutex and that mutex must also be locked before the
1501 * node's delayed items can be run.
1503 ret = __btrfs_add_delayed_item(delayed_node, delayed_item);
1504 if (unlikely(ret)) {
1505 btrfs_err(trans->fs_info,
1506 "error adding delayed dir index item, name: %.*s, index: %llu, root: %llu, dir: %llu, dir->index_cnt: %llu, delayed_node->index_cnt: %llu, error: %d",
1507 name_len, name, index, btrfs_root_id(delayed_node->root),
1508 delayed_node->inode_id, dir->index_cnt,
1509 delayed_node->index_cnt, ret);
1510 btrfs_release_delayed_item(delayed_item);
1511 btrfs_release_dir_index_item_space(trans);
1512 mutex_unlock(&delayed_node->mutex);
1516 if (delayed_node->index_item_leaves == 0 ||
1517 delayed_node->curr_index_batch_size + data_len > leaf_data_size) {
1518 delayed_node->curr_index_batch_size = data_len;
1519 reserve_leaf_space = true;
1521 delayed_node->curr_index_batch_size += data_len;
1522 reserve_leaf_space = false;
1525 if (reserve_leaf_space) {
1526 ret = btrfs_delayed_item_reserve_metadata(trans, delayed_item);
1528 * Space was reserved for a dir index item insertion when we
1529 * started the transaction, so getting a failure here should be
1533 btrfs_release_delayed_item(delayed_item);
1534 mutex_unlock(&delayed_node->mutex);
1538 delayed_node->index_item_leaves++;
1540 btrfs_release_dir_index_item_space(trans);
1542 mutex_unlock(&delayed_node->mutex);
1545 btrfs_release_delayed_node(delayed_node);
1549 static int btrfs_delete_delayed_insertion_item(struct btrfs_fs_info *fs_info,
1550 struct btrfs_delayed_node *node,
1553 struct btrfs_delayed_item *item;
1555 mutex_lock(&node->mutex);
1556 item = __btrfs_lookup_delayed_item(&node->ins_root.rb_root, index);
1558 mutex_unlock(&node->mutex);
1563 * For delayed items to insert, we track reserved metadata bytes based
1564 * on the number of leaves that we will use.
1565 * See btrfs_insert_delayed_dir_index() and
1566 * btrfs_delayed_item_reserve_metadata()).
1568 ASSERT(item->bytes_reserved == 0);
1569 ASSERT(node->index_item_leaves > 0);
1572 * If there's only one leaf reserved, we can decrement this item from the
1573 * current batch, otherwise we can not because we don't know which leaf
1574 * it belongs to. With the current limit on delayed items, we rarely
1575 * accumulate enough dir index items to fill more than one leaf (even
1576 * when using a leaf size of 4K).
1578 if (node->index_item_leaves == 1) {
1579 const u32 data_len = item->data_len + sizeof(struct btrfs_item);
1581 ASSERT(node->curr_index_batch_size >= data_len);
1582 node->curr_index_batch_size -= data_len;
1585 btrfs_release_delayed_item(item);
1587 /* If we now have no more dir index items, we can release all leaves. */
1588 if (RB_EMPTY_ROOT(&node->ins_root.rb_root)) {
1589 btrfs_delayed_item_release_leaves(node, node->index_item_leaves);
1590 node->index_item_leaves = 0;
1593 mutex_unlock(&node->mutex);
1597 int btrfs_delete_delayed_dir_index(struct btrfs_trans_handle *trans,
1598 struct btrfs_inode *dir, u64 index)
1600 struct btrfs_delayed_node *node;
1601 struct btrfs_delayed_item *item;
1604 node = btrfs_get_or_create_delayed_node(dir);
1606 return PTR_ERR(node);
1608 ret = btrfs_delete_delayed_insertion_item(trans->fs_info, node, index);
1612 item = btrfs_alloc_delayed_item(0, node, BTRFS_DELAYED_DELETION_ITEM);
1618 item->index = index;
1620 ret = btrfs_delayed_item_reserve_metadata(trans, item);
1622 * we have reserved enough space when we start a new transaction,
1623 * so reserving metadata failure is impossible.
1626 btrfs_err(trans->fs_info,
1627 "metadata reservation failed for delayed dir item deltiona, should have been reserved");
1628 btrfs_release_delayed_item(item);
1632 mutex_lock(&node->mutex);
1633 ret = __btrfs_add_delayed_item(node, item);
1634 if (unlikely(ret)) {
1635 btrfs_err(trans->fs_info,
1636 "err add delayed dir index item(index: %llu) into the deletion tree of the delayed node(root id: %llu, inode id: %llu, errno: %d)",
1637 index, node->root->root_key.objectid,
1638 node->inode_id, ret);
1639 btrfs_delayed_item_release_metadata(dir->root, item);
1640 btrfs_release_delayed_item(item);
1642 mutex_unlock(&node->mutex);
1644 btrfs_release_delayed_node(node);
1648 int btrfs_inode_delayed_dir_index_count(struct btrfs_inode *inode)
1650 struct btrfs_delayed_node *delayed_node = btrfs_get_delayed_node(inode);
1656 * Since we have held i_mutex of this directory, it is impossible that
1657 * a new directory index is added into the delayed node and index_cnt
1658 * is updated now. So we needn't lock the delayed node.
1660 if (!delayed_node->index_cnt) {
1661 btrfs_release_delayed_node(delayed_node);
1665 inode->index_cnt = delayed_node->index_cnt;
1666 btrfs_release_delayed_node(delayed_node);
1670 bool btrfs_readdir_get_delayed_items(struct inode *inode,
1672 struct list_head *ins_list,
1673 struct list_head *del_list)
1675 struct btrfs_delayed_node *delayed_node;
1676 struct btrfs_delayed_item *item;
1678 delayed_node = btrfs_get_delayed_node(BTRFS_I(inode));
1683 * We can only do one readdir with delayed items at a time because of
1684 * item->readdir_list.
1686 btrfs_inode_unlock(inode, BTRFS_ILOCK_SHARED);
1687 btrfs_inode_lock(inode, 0);
1689 mutex_lock(&delayed_node->mutex);
1690 item = __btrfs_first_delayed_insertion_item(delayed_node);
1691 while (item && item->index <= last_index) {
1692 refcount_inc(&item->refs);
1693 list_add_tail(&item->readdir_list, ins_list);
1694 item = __btrfs_next_delayed_item(item);
1697 item = __btrfs_first_delayed_deletion_item(delayed_node);
1698 while (item && item->index <= last_index) {
1699 refcount_inc(&item->refs);
1700 list_add_tail(&item->readdir_list, del_list);
1701 item = __btrfs_next_delayed_item(item);
1703 mutex_unlock(&delayed_node->mutex);
1705 * This delayed node is still cached in the btrfs inode, so refs
1706 * must be > 1 now, and we needn't check it is going to be freed
1709 * Besides that, this function is used to read dir, we do not
1710 * insert/delete delayed items in this period. So we also needn't
1711 * requeue or dequeue this delayed node.
1713 refcount_dec(&delayed_node->refs);
1718 void btrfs_readdir_put_delayed_items(struct inode *inode,
1719 struct list_head *ins_list,
1720 struct list_head *del_list)
1722 struct btrfs_delayed_item *curr, *next;
1724 list_for_each_entry_safe(curr, next, ins_list, readdir_list) {
1725 list_del(&curr->readdir_list);
1726 if (refcount_dec_and_test(&curr->refs))
1730 list_for_each_entry_safe(curr, next, del_list, readdir_list) {
1731 list_del(&curr->readdir_list);
1732 if (refcount_dec_and_test(&curr->refs))
1737 * The VFS is going to do up_read(), so we need to downgrade back to a
1740 downgrade_write(&inode->i_rwsem);
1743 int btrfs_should_delete_dir_index(struct list_head *del_list,
1746 struct btrfs_delayed_item *curr;
1749 list_for_each_entry(curr, del_list, readdir_list) {
1750 if (curr->index > index)
1752 if (curr->index == index) {
1761 * btrfs_readdir_delayed_dir_index - read dir info stored in the delayed tree
1764 int btrfs_readdir_delayed_dir_index(struct dir_context *ctx,
1765 struct list_head *ins_list)
1767 struct btrfs_dir_item *di;
1768 struct btrfs_delayed_item *curr, *next;
1769 struct btrfs_key location;
1773 unsigned char d_type;
1775 if (list_empty(ins_list))
1779 * Changing the data of the delayed item is impossible. So
1780 * we needn't lock them. And we have held i_mutex of the
1781 * directory, nobody can delete any directory indexes now.
1783 list_for_each_entry_safe(curr, next, ins_list, readdir_list) {
1784 list_del(&curr->readdir_list);
1786 if (curr->index < ctx->pos) {
1787 if (refcount_dec_and_test(&curr->refs))
1792 ctx->pos = curr->index;
1794 di = (struct btrfs_dir_item *)curr->data;
1795 name = (char *)(di + 1);
1796 name_len = btrfs_stack_dir_name_len(di);
1798 d_type = fs_ftype_to_dtype(di->type);
1799 btrfs_disk_key_to_cpu(&location, &di->location);
1801 over = !dir_emit(ctx, name, name_len,
1802 location.objectid, d_type);
1804 if (refcount_dec_and_test(&curr->refs))
1814 static void fill_stack_inode_item(struct btrfs_trans_handle *trans,
1815 struct btrfs_inode_item *inode_item,
1816 struct inode *inode)
1820 btrfs_set_stack_inode_uid(inode_item, i_uid_read(inode));
1821 btrfs_set_stack_inode_gid(inode_item, i_gid_read(inode));
1822 btrfs_set_stack_inode_size(inode_item, BTRFS_I(inode)->disk_i_size);
1823 btrfs_set_stack_inode_mode(inode_item, inode->i_mode);
1824 btrfs_set_stack_inode_nlink(inode_item, inode->i_nlink);
1825 btrfs_set_stack_inode_nbytes(inode_item, inode_get_bytes(inode));
1826 btrfs_set_stack_inode_generation(inode_item,
1827 BTRFS_I(inode)->generation);
1828 btrfs_set_stack_inode_sequence(inode_item,
1829 inode_peek_iversion(inode));
1830 btrfs_set_stack_inode_transid(inode_item, trans->transid);
1831 btrfs_set_stack_inode_rdev(inode_item, inode->i_rdev);
1832 flags = btrfs_inode_combine_flags(BTRFS_I(inode)->flags,
1833 BTRFS_I(inode)->ro_flags);
1834 btrfs_set_stack_inode_flags(inode_item, flags);
1835 btrfs_set_stack_inode_block_group(inode_item, 0);
1837 btrfs_set_stack_timespec_sec(&inode_item->atime,
1838 inode->i_atime.tv_sec);
1839 btrfs_set_stack_timespec_nsec(&inode_item->atime,
1840 inode->i_atime.tv_nsec);
1842 btrfs_set_stack_timespec_sec(&inode_item->mtime,
1843 inode->i_mtime.tv_sec);
1844 btrfs_set_stack_timespec_nsec(&inode_item->mtime,
1845 inode->i_mtime.tv_nsec);
1847 btrfs_set_stack_timespec_sec(&inode_item->ctime,
1848 inode->i_ctime.tv_sec);
1849 btrfs_set_stack_timespec_nsec(&inode_item->ctime,
1850 inode->i_ctime.tv_nsec);
1852 btrfs_set_stack_timespec_sec(&inode_item->otime,
1853 BTRFS_I(inode)->i_otime.tv_sec);
1854 btrfs_set_stack_timespec_nsec(&inode_item->otime,
1855 BTRFS_I(inode)->i_otime.tv_nsec);
1858 int btrfs_fill_inode(struct inode *inode, u32 *rdev)
1860 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
1861 struct btrfs_delayed_node *delayed_node;
1862 struct btrfs_inode_item *inode_item;
1864 delayed_node = btrfs_get_delayed_node(BTRFS_I(inode));
1868 mutex_lock(&delayed_node->mutex);
1869 if (!test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
1870 mutex_unlock(&delayed_node->mutex);
1871 btrfs_release_delayed_node(delayed_node);
1875 inode_item = &delayed_node->inode_item;
1877 i_uid_write(inode, btrfs_stack_inode_uid(inode_item));
1878 i_gid_write(inode, btrfs_stack_inode_gid(inode_item));
1879 btrfs_i_size_write(BTRFS_I(inode), btrfs_stack_inode_size(inode_item));
1880 btrfs_inode_set_file_extent_range(BTRFS_I(inode), 0,
1881 round_up(i_size_read(inode), fs_info->sectorsize));
1882 inode->i_mode = btrfs_stack_inode_mode(inode_item);
1883 set_nlink(inode, btrfs_stack_inode_nlink(inode_item));
1884 inode_set_bytes(inode, btrfs_stack_inode_nbytes(inode_item));
1885 BTRFS_I(inode)->generation = btrfs_stack_inode_generation(inode_item);
1886 BTRFS_I(inode)->last_trans = btrfs_stack_inode_transid(inode_item);
1888 inode_set_iversion_queried(inode,
1889 btrfs_stack_inode_sequence(inode_item));
1891 *rdev = btrfs_stack_inode_rdev(inode_item);
1892 btrfs_inode_split_flags(btrfs_stack_inode_flags(inode_item),
1893 &BTRFS_I(inode)->flags, &BTRFS_I(inode)->ro_flags);
1895 inode->i_atime.tv_sec = btrfs_stack_timespec_sec(&inode_item->atime);
1896 inode->i_atime.tv_nsec = btrfs_stack_timespec_nsec(&inode_item->atime);
1898 inode->i_mtime.tv_sec = btrfs_stack_timespec_sec(&inode_item->mtime);
1899 inode->i_mtime.tv_nsec = btrfs_stack_timespec_nsec(&inode_item->mtime);
1901 inode->i_ctime.tv_sec = btrfs_stack_timespec_sec(&inode_item->ctime);
1902 inode->i_ctime.tv_nsec = btrfs_stack_timespec_nsec(&inode_item->ctime);
1904 BTRFS_I(inode)->i_otime.tv_sec =
1905 btrfs_stack_timespec_sec(&inode_item->otime);
1906 BTRFS_I(inode)->i_otime.tv_nsec =
1907 btrfs_stack_timespec_nsec(&inode_item->otime);
1909 inode->i_generation = BTRFS_I(inode)->generation;
1910 BTRFS_I(inode)->index_cnt = (u64)-1;
1912 mutex_unlock(&delayed_node->mutex);
1913 btrfs_release_delayed_node(delayed_node);
1917 int btrfs_delayed_update_inode(struct btrfs_trans_handle *trans,
1918 struct btrfs_root *root,
1919 struct btrfs_inode *inode)
1921 struct btrfs_delayed_node *delayed_node;
1924 delayed_node = btrfs_get_or_create_delayed_node(inode);
1925 if (IS_ERR(delayed_node))
1926 return PTR_ERR(delayed_node);
1928 mutex_lock(&delayed_node->mutex);
1929 if (test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
1930 fill_stack_inode_item(trans, &delayed_node->inode_item,
1935 ret = btrfs_delayed_inode_reserve_metadata(trans, root, delayed_node);
1939 fill_stack_inode_item(trans, &delayed_node->inode_item, &inode->vfs_inode);
1940 set_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags);
1941 delayed_node->count++;
1942 atomic_inc(&root->fs_info->delayed_root->items);
1944 mutex_unlock(&delayed_node->mutex);
1945 btrfs_release_delayed_node(delayed_node);
1949 int btrfs_delayed_delete_inode_ref(struct btrfs_inode *inode)
1951 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1952 struct btrfs_delayed_node *delayed_node;
1955 * we don't do delayed inode updates during log recovery because it
1956 * leads to enospc problems. This means we also can't do
1957 * delayed inode refs
1959 if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
1962 delayed_node = btrfs_get_or_create_delayed_node(inode);
1963 if (IS_ERR(delayed_node))
1964 return PTR_ERR(delayed_node);
1967 * We don't reserve space for inode ref deletion is because:
1968 * - We ONLY do async inode ref deletion for the inode who has only
1969 * one link(i_nlink == 1), it means there is only one inode ref.
1970 * And in most case, the inode ref and the inode item are in the
1971 * same leaf, and we will deal with them at the same time.
1972 * Since we are sure we will reserve the space for the inode item,
1973 * it is unnecessary to reserve space for inode ref deletion.
1974 * - If the inode ref and the inode item are not in the same leaf,
1975 * We also needn't worry about enospc problem, because we reserve
1976 * much more space for the inode update than it needs.
1977 * - At the worst, we can steal some space from the global reservation.
1980 mutex_lock(&delayed_node->mutex);
1981 if (test_bit(BTRFS_DELAYED_NODE_DEL_IREF, &delayed_node->flags))
1984 set_bit(BTRFS_DELAYED_NODE_DEL_IREF, &delayed_node->flags);
1985 delayed_node->count++;
1986 atomic_inc(&fs_info->delayed_root->items);
1988 mutex_unlock(&delayed_node->mutex);
1989 btrfs_release_delayed_node(delayed_node);
1993 static void __btrfs_kill_delayed_node(struct btrfs_delayed_node *delayed_node)
1995 struct btrfs_root *root = delayed_node->root;
1996 struct btrfs_fs_info *fs_info = root->fs_info;
1997 struct btrfs_delayed_item *curr_item, *prev_item;
1999 mutex_lock(&delayed_node->mutex);
2000 curr_item = __btrfs_first_delayed_insertion_item(delayed_node);
2002 prev_item = curr_item;
2003 curr_item = __btrfs_next_delayed_item(prev_item);
2004 btrfs_release_delayed_item(prev_item);
2007 if (delayed_node->index_item_leaves > 0) {
2008 btrfs_delayed_item_release_leaves(delayed_node,
2009 delayed_node->index_item_leaves);
2010 delayed_node->index_item_leaves = 0;
2013 curr_item = __btrfs_first_delayed_deletion_item(delayed_node);
2015 btrfs_delayed_item_release_metadata(root, curr_item);
2016 prev_item = curr_item;
2017 curr_item = __btrfs_next_delayed_item(prev_item);
2018 btrfs_release_delayed_item(prev_item);
2021 btrfs_release_delayed_iref(delayed_node);
2023 if (test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
2024 btrfs_delayed_inode_release_metadata(fs_info, delayed_node, false);
2025 btrfs_release_delayed_inode(delayed_node);
2027 mutex_unlock(&delayed_node->mutex);
2030 void btrfs_kill_delayed_inode_items(struct btrfs_inode *inode)
2032 struct btrfs_delayed_node *delayed_node;
2034 delayed_node = btrfs_get_delayed_node(inode);
2038 __btrfs_kill_delayed_node(delayed_node);
2039 btrfs_release_delayed_node(delayed_node);
2042 void btrfs_kill_all_delayed_nodes(struct btrfs_root *root)
2045 struct btrfs_delayed_node *delayed_nodes[8];
2049 spin_lock(&root->inode_lock);
2050 n = radix_tree_gang_lookup(&root->delayed_nodes_tree,
2051 (void **)delayed_nodes, inode_id,
2052 ARRAY_SIZE(delayed_nodes));
2054 spin_unlock(&root->inode_lock);
2058 inode_id = delayed_nodes[n - 1]->inode_id + 1;
2059 for (i = 0; i < n; i++) {
2061 * Don't increase refs in case the node is dead and
2062 * about to be removed from the tree in the loop below
2064 if (!refcount_inc_not_zero(&delayed_nodes[i]->refs))
2065 delayed_nodes[i] = NULL;
2067 spin_unlock(&root->inode_lock);
2069 for (i = 0; i < n; i++) {
2070 if (!delayed_nodes[i])
2072 __btrfs_kill_delayed_node(delayed_nodes[i]);
2073 btrfs_release_delayed_node(delayed_nodes[i]);
2078 void btrfs_destroy_delayed_inodes(struct btrfs_fs_info *fs_info)
2080 struct btrfs_delayed_node *curr_node, *prev_node;
2082 curr_node = btrfs_first_delayed_node(fs_info->delayed_root);
2084 __btrfs_kill_delayed_node(curr_node);
2086 prev_node = curr_node;
2087 curr_node = btrfs_next_delayed_node(curr_node);
2088 btrfs_release_delayed_node(prev_node);
2092 void btrfs_log_get_delayed_items(struct btrfs_inode *inode,
2093 struct list_head *ins_list,
2094 struct list_head *del_list)
2096 struct btrfs_delayed_node *node;
2097 struct btrfs_delayed_item *item;
2099 node = btrfs_get_delayed_node(inode);
2103 mutex_lock(&node->mutex);
2104 item = __btrfs_first_delayed_insertion_item(node);
2107 * It's possible that the item is already in a log list. This
2108 * can happen in case two tasks are trying to log the same
2109 * directory. For example if we have tasks A and task B:
2111 * Task A collected the delayed items into a log list while
2112 * under the inode's log_mutex (at btrfs_log_inode()), but it
2113 * only releases the items after logging the inodes they point
2114 * to (if they are new inodes), which happens after unlocking
2117 * Task B enters btrfs_log_inode() and acquires the log_mutex
2118 * of the same directory inode, before task B releases the
2119 * delayed items. This can happen for example when logging some
2120 * inode we need to trigger logging of its parent directory, so
2121 * logging two files that have the same parent directory can
2124 * If this happens, just ignore delayed items already in a log
2125 * list. All the tasks logging the directory are under a log
2126 * transaction and whichever finishes first can not sync the log
2127 * before the other completes and leaves the log transaction.
2129 if (!item->logged && list_empty(&item->log_list)) {
2130 refcount_inc(&item->refs);
2131 list_add_tail(&item->log_list, ins_list);
2133 item = __btrfs_next_delayed_item(item);
2136 item = __btrfs_first_delayed_deletion_item(node);
2138 /* It may be non-empty, for the same reason mentioned above. */
2139 if (!item->logged && list_empty(&item->log_list)) {
2140 refcount_inc(&item->refs);
2141 list_add_tail(&item->log_list, del_list);
2143 item = __btrfs_next_delayed_item(item);
2145 mutex_unlock(&node->mutex);
2148 * We are called during inode logging, which means the inode is in use
2149 * and can not be evicted before we finish logging the inode. So we never
2150 * have the last reference on the delayed inode.
2151 * Also, we don't use btrfs_release_delayed_node() because that would
2152 * requeue the delayed inode (change its order in the list of prepared
2153 * nodes) and we don't want to do such change because we don't create or
2154 * delete delayed items.
2156 ASSERT(refcount_read(&node->refs) > 1);
2157 refcount_dec(&node->refs);
2160 void btrfs_log_put_delayed_items(struct btrfs_inode *inode,
2161 struct list_head *ins_list,
2162 struct list_head *del_list)
2164 struct btrfs_delayed_node *node;
2165 struct btrfs_delayed_item *item;
2166 struct btrfs_delayed_item *next;
2168 node = btrfs_get_delayed_node(inode);
2172 mutex_lock(&node->mutex);
2174 list_for_each_entry_safe(item, next, ins_list, log_list) {
2175 item->logged = true;
2176 list_del_init(&item->log_list);
2177 if (refcount_dec_and_test(&item->refs))
2181 list_for_each_entry_safe(item, next, del_list, log_list) {
2182 item->logged = true;
2183 list_del_init(&item->log_list);
2184 if (refcount_dec_and_test(&item->refs))
2188 mutex_unlock(&node->mutex);
2191 * We are called during inode logging, which means the inode is in use
2192 * and can not be evicted before we finish logging the inode. So we never
2193 * have the last reference on the delayed inode.
2194 * Also, we don't use btrfs_release_delayed_node() because that would
2195 * requeue the delayed inode (change its order in the list of prepared
2196 * nodes) and we don't want to do such change because we don't create or
2197 * delete delayed items.
2199 ASSERT(refcount_read(&node->refs) > 1);
2200 refcount_dec(&node->refs);