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>
13 #include "delayed-inode.h"
15 #include "transaction.h"
18 #include "inode-item.h"
19 #include "space-info.h"
20 #include "accessors.h"
21 #include "file-item.h"
23 #define BTRFS_DELAYED_WRITEBACK 512
24 #define BTRFS_DELAYED_BACKGROUND 128
25 #define BTRFS_DELAYED_BATCH 16
27 static struct kmem_cache *delayed_node_cache;
29 int __init btrfs_delayed_inode_init(void)
31 delayed_node_cache = kmem_cache_create("btrfs_delayed_node",
32 sizeof(struct btrfs_delayed_node),
36 if (!delayed_node_cache)
41 void __cold btrfs_delayed_inode_exit(void)
43 kmem_cache_destroy(delayed_node_cache);
46 static inline void btrfs_init_delayed_node(
47 struct btrfs_delayed_node *delayed_node,
48 struct btrfs_root *root, u64 inode_id)
50 delayed_node->root = root;
51 delayed_node->inode_id = inode_id;
52 refcount_set(&delayed_node->refs, 0);
53 delayed_node->ins_root = RB_ROOT_CACHED;
54 delayed_node->del_root = RB_ROOT_CACHED;
55 mutex_init(&delayed_node->mutex);
56 INIT_LIST_HEAD(&delayed_node->n_list);
57 INIT_LIST_HEAD(&delayed_node->p_list);
60 static struct btrfs_delayed_node *btrfs_get_delayed_node(
61 struct btrfs_inode *btrfs_inode)
63 struct btrfs_root *root = btrfs_inode->root;
64 u64 ino = btrfs_ino(btrfs_inode);
65 struct btrfs_delayed_node *node;
67 node = READ_ONCE(btrfs_inode->delayed_node);
69 refcount_inc(&node->refs);
73 spin_lock(&root->inode_lock);
74 node = radix_tree_lookup(&root->delayed_nodes_tree, ino);
77 if (btrfs_inode->delayed_node) {
78 refcount_inc(&node->refs); /* can be accessed */
79 BUG_ON(btrfs_inode->delayed_node != node);
80 spin_unlock(&root->inode_lock);
85 * It's possible that we're racing into the middle of removing
86 * this node from the radix tree. In this case, the refcount
87 * was zero and it should never go back to one. Just return
88 * NULL like it was never in the radix at all; our release
89 * function is in the process of removing it.
91 * Some implementations of refcount_inc refuse to bump the
92 * refcount once it has hit zero. If we don't do this dance
93 * here, refcount_inc() may decide to just WARN_ONCE() instead
94 * of actually bumping the refcount.
96 * If this node is properly in the radix, we want to bump the
97 * refcount twice, once for the inode and once for this get
100 if (refcount_inc_not_zero(&node->refs)) {
101 refcount_inc(&node->refs);
102 btrfs_inode->delayed_node = node;
107 spin_unlock(&root->inode_lock);
110 spin_unlock(&root->inode_lock);
115 /* Will return either the node or PTR_ERR(-ENOMEM) */
116 static struct btrfs_delayed_node *btrfs_get_or_create_delayed_node(
117 struct btrfs_inode *btrfs_inode)
119 struct btrfs_delayed_node *node;
120 struct btrfs_root *root = btrfs_inode->root;
121 u64 ino = btrfs_ino(btrfs_inode);
125 node = btrfs_get_delayed_node(btrfs_inode);
129 node = kmem_cache_zalloc(delayed_node_cache, GFP_NOFS);
131 return ERR_PTR(-ENOMEM);
132 btrfs_init_delayed_node(node, root, ino);
134 /* cached in the btrfs inode and can be accessed */
135 refcount_set(&node->refs, 2);
137 ret = radix_tree_preload(GFP_NOFS);
139 kmem_cache_free(delayed_node_cache, node);
143 spin_lock(&root->inode_lock);
144 ret = radix_tree_insert(&root->delayed_nodes_tree, ino, node);
145 if (ret == -EEXIST) {
146 spin_unlock(&root->inode_lock);
147 kmem_cache_free(delayed_node_cache, node);
148 radix_tree_preload_end();
151 btrfs_inode->delayed_node = node;
152 spin_unlock(&root->inode_lock);
153 radix_tree_preload_end();
159 * Call it when holding delayed_node->mutex
161 * If mod = 1, add this node into the prepared list.
163 static void btrfs_queue_delayed_node(struct btrfs_delayed_root *root,
164 struct btrfs_delayed_node *node,
167 spin_lock(&root->lock);
168 if (test_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags)) {
169 if (!list_empty(&node->p_list))
170 list_move_tail(&node->p_list, &root->prepare_list);
172 list_add_tail(&node->p_list, &root->prepare_list);
174 list_add_tail(&node->n_list, &root->node_list);
175 list_add_tail(&node->p_list, &root->prepare_list);
176 refcount_inc(&node->refs); /* inserted into list */
178 set_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags);
180 spin_unlock(&root->lock);
183 /* Call it when holding delayed_node->mutex */
184 static void btrfs_dequeue_delayed_node(struct btrfs_delayed_root *root,
185 struct btrfs_delayed_node *node)
187 spin_lock(&root->lock);
188 if (test_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags)) {
190 refcount_dec(&node->refs); /* not in the list */
191 list_del_init(&node->n_list);
192 if (!list_empty(&node->p_list))
193 list_del_init(&node->p_list);
194 clear_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags);
196 spin_unlock(&root->lock);
199 static struct btrfs_delayed_node *btrfs_first_delayed_node(
200 struct btrfs_delayed_root *delayed_root)
203 struct btrfs_delayed_node *node = NULL;
205 spin_lock(&delayed_root->lock);
206 if (list_empty(&delayed_root->node_list))
209 p = delayed_root->node_list.next;
210 node = list_entry(p, struct btrfs_delayed_node, n_list);
211 refcount_inc(&node->refs);
213 spin_unlock(&delayed_root->lock);
218 static struct btrfs_delayed_node *btrfs_next_delayed_node(
219 struct btrfs_delayed_node *node)
221 struct btrfs_delayed_root *delayed_root;
223 struct btrfs_delayed_node *next = NULL;
225 delayed_root = node->root->fs_info->delayed_root;
226 spin_lock(&delayed_root->lock);
227 if (!test_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags)) {
228 /* not in the list */
229 if (list_empty(&delayed_root->node_list))
231 p = delayed_root->node_list.next;
232 } else if (list_is_last(&node->n_list, &delayed_root->node_list))
235 p = node->n_list.next;
237 next = list_entry(p, struct btrfs_delayed_node, n_list);
238 refcount_inc(&next->refs);
240 spin_unlock(&delayed_root->lock);
245 static void __btrfs_release_delayed_node(
246 struct btrfs_delayed_node *delayed_node,
249 struct btrfs_delayed_root *delayed_root;
254 delayed_root = delayed_node->root->fs_info->delayed_root;
256 mutex_lock(&delayed_node->mutex);
257 if (delayed_node->count)
258 btrfs_queue_delayed_node(delayed_root, delayed_node, mod);
260 btrfs_dequeue_delayed_node(delayed_root, delayed_node);
261 mutex_unlock(&delayed_node->mutex);
263 if (refcount_dec_and_test(&delayed_node->refs)) {
264 struct btrfs_root *root = delayed_node->root;
266 spin_lock(&root->inode_lock);
268 * Once our refcount goes to zero, nobody is allowed to bump it
269 * back up. We can delete it now.
271 ASSERT(refcount_read(&delayed_node->refs) == 0);
272 radix_tree_delete(&root->delayed_nodes_tree,
273 delayed_node->inode_id);
274 spin_unlock(&root->inode_lock);
275 kmem_cache_free(delayed_node_cache, delayed_node);
279 static inline void btrfs_release_delayed_node(struct btrfs_delayed_node *node)
281 __btrfs_release_delayed_node(node, 0);
284 static struct btrfs_delayed_node *btrfs_first_prepared_delayed_node(
285 struct btrfs_delayed_root *delayed_root)
288 struct btrfs_delayed_node *node = NULL;
290 spin_lock(&delayed_root->lock);
291 if (list_empty(&delayed_root->prepare_list))
294 p = delayed_root->prepare_list.next;
296 node = list_entry(p, struct btrfs_delayed_node, p_list);
297 refcount_inc(&node->refs);
299 spin_unlock(&delayed_root->lock);
304 static inline void btrfs_release_prepared_delayed_node(
305 struct btrfs_delayed_node *node)
307 __btrfs_release_delayed_node(node, 1);
310 static struct btrfs_delayed_item *btrfs_alloc_delayed_item(u16 data_len,
311 struct btrfs_delayed_node *node,
312 enum btrfs_delayed_item_type type)
314 struct btrfs_delayed_item *item;
316 item = kmalloc(struct_size(item, data, data_len), GFP_NOFS);
318 item->data_len = data_len;
320 item->bytes_reserved = 0;
321 item->delayed_node = node;
322 RB_CLEAR_NODE(&item->rb_node);
323 INIT_LIST_HEAD(&item->log_list);
324 item->logged = false;
325 refcount_set(&item->refs, 1);
331 * Look up the delayed item by key.
333 * @delayed_node: pointer to the delayed node
334 * @index: the dir index value to lookup (offset of a dir index key)
336 * Note: if we don't find the right item, we will return the prev item and
339 static struct btrfs_delayed_item *__btrfs_lookup_delayed_item(
340 struct rb_root *root,
343 struct rb_node *node = root->rb_node;
344 struct btrfs_delayed_item *delayed_item = NULL;
347 delayed_item = rb_entry(node, struct btrfs_delayed_item,
349 if (delayed_item->index < index)
350 node = node->rb_right;
351 else if (delayed_item->index > index)
352 node = node->rb_left;
360 static int __btrfs_add_delayed_item(struct btrfs_delayed_node *delayed_node,
361 struct btrfs_delayed_item *ins)
363 struct rb_node **p, *node;
364 struct rb_node *parent_node = NULL;
365 struct rb_root_cached *root;
366 struct btrfs_delayed_item *item;
367 bool leftmost = true;
369 if (ins->type == BTRFS_DELAYED_INSERTION_ITEM)
370 root = &delayed_node->ins_root;
372 root = &delayed_node->del_root;
374 p = &root->rb_root.rb_node;
375 node = &ins->rb_node;
379 item = rb_entry(parent_node, struct btrfs_delayed_item,
382 if (item->index < ins->index) {
385 } else if (item->index > ins->index) {
392 rb_link_node(node, parent_node, p);
393 rb_insert_color_cached(node, root, leftmost);
395 if (ins->type == BTRFS_DELAYED_INSERTION_ITEM &&
396 ins->index >= delayed_node->index_cnt)
397 delayed_node->index_cnt = ins->index + 1;
399 delayed_node->count++;
400 atomic_inc(&delayed_node->root->fs_info->delayed_root->items);
404 static void finish_one_item(struct btrfs_delayed_root *delayed_root)
406 int seq = atomic_inc_return(&delayed_root->items_seq);
408 /* atomic_dec_return implies a barrier */
409 if ((atomic_dec_return(&delayed_root->items) <
410 BTRFS_DELAYED_BACKGROUND || seq % BTRFS_DELAYED_BATCH == 0))
411 cond_wake_up_nomb(&delayed_root->wait);
414 static void __btrfs_remove_delayed_item(struct btrfs_delayed_item *delayed_item)
416 struct btrfs_delayed_node *delayed_node = delayed_item->delayed_node;
417 struct rb_root_cached *root;
418 struct btrfs_delayed_root *delayed_root;
420 /* Not inserted, ignore it. */
421 if (RB_EMPTY_NODE(&delayed_item->rb_node))
424 /* If it's in a rbtree, then we need to have delayed node locked. */
425 lockdep_assert_held(&delayed_node->mutex);
427 delayed_root = delayed_node->root->fs_info->delayed_root;
429 BUG_ON(!delayed_root);
431 if (delayed_item->type == BTRFS_DELAYED_INSERTION_ITEM)
432 root = &delayed_node->ins_root;
434 root = &delayed_node->del_root;
436 rb_erase_cached(&delayed_item->rb_node, root);
437 RB_CLEAR_NODE(&delayed_item->rb_node);
438 delayed_node->count--;
440 finish_one_item(delayed_root);
443 static void btrfs_release_delayed_item(struct btrfs_delayed_item *item)
446 __btrfs_remove_delayed_item(item);
447 if (refcount_dec_and_test(&item->refs))
452 static struct btrfs_delayed_item *__btrfs_first_delayed_insertion_item(
453 struct btrfs_delayed_node *delayed_node)
456 struct btrfs_delayed_item *item = NULL;
458 p = rb_first_cached(&delayed_node->ins_root);
460 item = rb_entry(p, struct btrfs_delayed_item, rb_node);
465 static struct btrfs_delayed_item *__btrfs_first_delayed_deletion_item(
466 struct btrfs_delayed_node *delayed_node)
469 struct btrfs_delayed_item *item = NULL;
471 p = rb_first_cached(&delayed_node->del_root);
473 item = rb_entry(p, struct btrfs_delayed_item, rb_node);
478 static struct btrfs_delayed_item *__btrfs_next_delayed_item(
479 struct btrfs_delayed_item *item)
482 struct btrfs_delayed_item *next = NULL;
484 p = rb_next(&item->rb_node);
486 next = rb_entry(p, struct btrfs_delayed_item, rb_node);
491 static int btrfs_delayed_item_reserve_metadata(struct btrfs_trans_handle *trans,
492 struct btrfs_delayed_item *item)
494 struct btrfs_block_rsv *src_rsv;
495 struct btrfs_block_rsv *dst_rsv;
496 struct btrfs_fs_info *fs_info = trans->fs_info;
500 if (!trans->bytes_reserved)
503 src_rsv = trans->block_rsv;
504 dst_rsv = &fs_info->delayed_block_rsv;
506 num_bytes = btrfs_calc_insert_metadata_size(fs_info, 1);
509 * Here we migrate space rsv from transaction rsv, since have already
510 * reserved space when starting a transaction. So no need to reserve
513 ret = btrfs_block_rsv_migrate(src_rsv, dst_rsv, num_bytes, true);
515 trace_btrfs_space_reservation(fs_info, "delayed_item",
516 item->delayed_node->inode_id,
519 * For insertions we track reserved metadata space by accounting
520 * for the number of leaves that will be used, based on the delayed
521 * node's curr_index_batch_size and index_item_leaves fields.
523 if (item->type == BTRFS_DELAYED_DELETION_ITEM)
524 item->bytes_reserved = num_bytes;
530 static void btrfs_delayed_item_release_metadata(struct btrfs_root *root,
531 struct btrfs_delayed_item *item)
533 struct btrfs_block_rsv *rsv;
534 struct btrfs_fs_info *fs_info = root->fs_info;
536 if (!item->bytes_reserved)
539 rsv = &fs_info->delayed_block_rsv;
541 * Check btrfs_delayed_item_reserve_metadata() to see why we don't need
542 * to release/reserve qgroup space.
544 trace_btrfs_space_reservation(fs_info, "delayed_item",
545 item->delayed_node->inode_id,
546 item->bytes_reserved, 0);
547 btrfs_block_rsv_release(fs_info, rsv, item->bytes_reserved, NULL);
550 static void btrfs_delayed_item_release_leaves(struct btrfs_delayed_node *node,
551 unsigned int num_leaves)
553 struct btrfs_fs_info *fs_info = node->root->fs_info;
554 const u64 bytes = btrfs_calc_insert_metadata_size(fs_info, num_leaves);
556 /* There are no space reservations during log replay, bail out. */
557 if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
560 trace_btrfs_space_reservation(fs_info, "delayed_item", node->inode_id,
562 btrfs_block_rsv_release(fs_info, &fs_info->delayed_block_rsv, bytes, NULL);
565 static int btrfs_delayed_inode_reserve_metadata(
566 struct btrfs_trans_handle *trans,
567 struct btrfs_root *root,
568 struct btrfs_delayed_node *node)
570 struct btrfs_fs_info *fs_info = root->fs_info;
571 struct btrfs_block_rsv *src_rsv;
572 struct btrfs_block_rsv *dst_rsv;
576 src_rsv = trans->block_rsv;
577 dst_rsv = &fs_info->delayed_block_rsv;
579 num_bytes = btrfs_calc_metadata_size(fs_info, 1);
582 * btrfs_dirty_inode will update the inode under btrfs_join_transaction
583 * which doesn't reserve space for speed. This is a problem since we
584 * still need to reserve space for this update, so try to reserve the
587 * Now if src_rsv == delalloc_block_rsv we'll let it just steal since
588 * we always reserve enough to update the inode item.
590 if (!src_rsv || (!trans->bytes_reserved &&
591 src_rsv->type != BTRFS_BLOCK_RSV_DELALLOC)) {
592 ret = btrfs_qgroup_reserve_meta(root, num_bytes,
593 BTRFS_QGROUP_RSV_META_PREALLOC, true);
596 ret = btrfs_block_rsv_add(fs_info, dst_rsv, num_bytes,
597 BTRFS_RESERVE_NO_FLUSH);
598 /* NO_FLUSH could only fail with -ENOSPC */
599 ASSERT(ret == 0 || ret == -ENOSPC);
601 btrfs_qgroup_free_meta_prealloc(root, num_bytes);
603 ret = btrfs_block_rsv_migrate(src_rsv, dst_rsv, num_bytes, true);
607 trace_btrfs_space_reservation(fs_info, "delayed_inode",
608 node->inode_id, num_bytes, 1);
609 node->bytes_reserved = num_bytes;
615 static void btrfs_delayed_inode_release_metadata(struct btrfs_fs_info *fs_info,
616 struct btrfs_delayed_node *node,
619 struct btrfs_block_rsv *rsv;
621 if (!node->bytes_reserved)
624 rsv = &fs_info->delayed_block_rsv;
625 trace_btrfs_space_reservation(fs_info, "delayed_inode",
626 node->inode_id, node->bytes_reserved, 0);
627 btrfs_block_rsv_release(fs_info, rsv, node->bytes_reserved, NULL);
629 btrfs_qgroup_free_meta_prealloc(node->root,
630 node->bytes_reserved);
632 btrfs_qgroup_convert_reserved_meta(node->root,
633 node->bytes_reserved);
634 node->bytes_reserved = 0;
638 * Insert a single delayed item or a batch of delayed items, as many as possible
639 * that fit in a leaf. The delayed items (dir index keys) are sorted by their key
640 * in the rbtree, and if there's a gap between two consecutive dir index items,
641 * then it means at some point we had delayed dir indexes to add but they got
642 * removed (by btrfs_delete_delayed_dir_index()) before we attempted to flush them
643 * into the subvolume tree. Dir index keys also have their offsets coming from a
644 * monotonically increasing counter, so we can't get new keys with an offset that
645 * fits within a gap between delayed dir index items.
647 static int btrfs_insert_delayed_item(struct btrfs_trans_handle *trans,
648 struct btrfs_root *root,
649 struct btrfs_path *path,
650 struct btrfs_delayed_item *first_item)
652 struct btrfs_fs_info *fs_info = root->fs_info;
653 struct btrfs_delayed_node *node = first_item->delayed_node;
654 LIST_HEAD(item_list);
655 struct btrfs_delayed_item *curr;
656 struct btrfs_delayed_item *next;
657 const int max_size = BTRFS_LEAF_DATA_SIZE(fs_info);
658 struct btrfs_item_batch batch;
659 struct btrfs_key first_key;
660 const u32 first_data_size = first_item->data_len;
662 char *ins_data = NULL;
664 bool continuous_keys_only = false;
666 lockdep_assert_held(&node->mutex);
669 * During normal operation the delayed index offset is continuously
670 * increasing, so we can batch insert all items as there will not be any
671 * overlapping keys in the tree.
673 * The exception to this is log replay, where we may have interleaved
674 * offsets in the tree, so our batch needs to be continuous keys only in
675 * order to ensure we do not end up with out of order items in our leaf.
677 if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
678 continuous_keys_only = true;
681 * For delayed items to insert, we track reserved metadata bytes based
682 * on the number of leaves that we will use.
683 * See btrfs_insert_delayed_dir_index() and
684 * btrfs_delayed_item_reserve_metadata()).
686 ASSERT(first_item->bytes_reserved == 0);
688 list_add_tail(&first_item->tree_list, &item_list);
689 batch.total_data_size = first_data_size;
691 total_size = first_data_size + sizeof(struct btrfs_item);
697 next = __btrfs_next_delayed_item(curr);
702 * We cannot allow gaps in the key space if we're doing log
705 if (continuous_keys_only && (next->index != curr->index + 1))
708 ASSERT(next->bytes_reserved == 0);
710 next_size = next->data_len + sizeof(struct btrfs_item);
711 if (total_size + next_size > max_size)
714 list_add_tail(&next->tree_list, &item_list);
716 total_size += next_size;
717 batch.total_data_size += next->data_len;
722 first_key.objectid = node->inode_id;
723 first_key.type = BTRFS_DIR_INDEX_KEY;
724 first_key.offset = first_item->index;
725 batch.keys = &first_key;
726 batch.data_sizes = &first_data_size;
728 struct btrfs_key *ins_keys;
732 ins_data = kmalloc(batch.nr * sizeof(u32) +
733 batch.nr * sizeof(struct btrfs_key), GFP_NOFS);
738 ins_sizes = (u32 *)ins_data;
739 ins_keys = (struct btrfs_key *)(ins_data + batch.nr * sizeof(u32));
740 batch.keys = ins_keys;
741 batch.data_sizes = ins_sizes;
742 list_for_each_entry(curr, &item_list, tree_list) {
743 ins_keys[i].objectid = node->inode_id;
744 ins_keys[i].type = BTRFS_DIR_INDEX_KEY;
745 ins_keys[i].offset = curr->index;
746 ins_sizes[i] = curr->data_len;
751 ret = btrfs_insert_empty_items(trans, root, path, &batch);
755 list_for_each_entry(curr, &item_list, tree_list) {
758 data_ptr = btrfs_item_ptr(path->nodes[0], path->slots[0], char);
759 write_extent_buffer(path->nodes[0], &curr->data,
760 (unsigned long)data_ptr, curr->data_len);
765 * Now release our path before releasing the delayed items and their
766 * metadata reservations, so that we don't block other tasks for more
769 btrfs_release_path(path);
771 ASSERT(node->index_item_leaves > 0);
774 * For normal operations we will batch an entire leaf's worth of delayed
775 * items, so if there are more items to process we can decrement
776 * index_item_leaves by 1 as we inserted 1 leaf's worth of items.
778 * However for log replay we may not have inserted an entire leaf's
779 * worth of items, we may have not had continuous items, so decrementing
780 * here would mess up the index_item_leaves accounting. For this case
781 * only clean up the accounting when there are no items left.
783 if (next && !continuous_keys_only) {
785 * We inserted one batch of items into a leaf a there are more
786 * items to flush in a future batch, now release one unit of
787 * metadata space from the delayed block reserve, corresponding
788 * the leaf we just flushed to.
790 btrfs_delayed_item_release_leaves(node, 1);
791 node->index_item_leaves--;
794 * There are no more items to insert. We can have a number of
795 * reserved leaves > 1 here - this happens when many dir index
796 * items are added and then removed before they are flushed (file
797 * names with a very short life, never span a transaction). So
798 * release all remaining leaves.
800 btrfs_delayed_item_release_leaves(node, node->index_item_leaves);
801 node->index_item_leaves = 0;
804 list_for_each_entry_safe(curr, next, &item_list, tree_list) {
805 list_del(&curr->tree_list);
806 btrfs_release_delayed_item(curr);
813 static int btrfs_insert_delayed_items(struct btrfs_trans_handle *trans,
814 struct btrfs_path *path,
815 struct btrfs_root *root,
816 struct btrfs_delayed_node *node)
821 struct btrfs_delayed_item *curr;
823 mutex_lock(&node->mutex);
824 curr = __btrfs_first_delayed_insertion_item(node);
826 mutex_unlock(&node->mutex);
829 ret = btrfs_insert_delayed_item(trans, root, path, curr);
830 mutex_unlock(&node->mutex);
836 static int btrfs_batch_delete_items(struct btrfs_trans_handle *trans,
837 struct btrfs_root *root,
838 struct btrfs_path *path,
839 struct btrfs_delayed_item *item)
841 const u64 ino = item->delayed_node->inode_id;
842 struct btrfs_fs_info *fs_info = root->fs_info;
843 struct btrfs_delayed_item *curr, *next;
844 struct extent_buffer *leaf = path->nodes[0];
845 LIST_HEAD(batch_list);
846 int nitems, slot, last_slot;
848 u64 total_reserved_size = item->bytes_reserved;
850 ASSERT(leaf != NULL);
852 slot = path->slots[0];
853 last_slot = btrfs_header_nritems(leaf) - 1;
855 * Our caller always gives us a path pointing to an existing item, so
856 * this can not happen.
858 ASSERT(slot <= last_slot);
859 if (WARN_ON(slot > last_slot))
864 list_add_tail(&curr->tree_list, &batch_list);
867 * Keep checking if the next delayed item matches the next item in the
868 * leaf - if so, we can add it to the batch of items to delete from the
871 while (slot < last_slot) {
872 struct btrfs_key key;
874 next = __btrfs_next_delayed_item(curr);
879 btrfs_item_key_to_cpu(leaf, &key, slot);
880 if (key.objectid != ino ||
881 key.type != BTRFS_DIR_INDEX_KEY ||
882 key.offset != next->index)
886 list_add_tail(&curr->tree_list, &batch_list);
887 total_reserved_size += curr->bytes_reserved;
890 ret = btrfs_del_items(trans, root, path, path->slots[0], nitems);
894 /* In case of BTRFS_FS_LOG_RECOVERING items won't have reserved space */
895 if (total_reserved_size > 0) {
897 * Check btrfs_delayed_item_reserve_metadata() to see why we
898 * don't need to release/reserve qgroup space.
900 trace_btrfs_space_reservation(fs_info, "delayed_item", ino,
901 total_reserved_size, 0);
902 btrfs_block_rsv_release(fs_info, &fs_info->delayed_block_rsv,
903 total_reserved_size, NULL);
906 list_for_each_entry_safe(curr, next, &batch_list, tree_list) {
907 list_del(&curr->tree_list);
908 btrfs_release_delayed_item(curr);
914 static int btrfs_delete_delayed_items(struct btrfs_trans_handle *trans,
915 struct btrfs_path *path,
916 struct btrfs_root *root,
917 struct btrfs_delayed_node *node)
919 struct btrfs_key key;
922 key.objectid = node->inode_id;
923 key.type = BTRFS_DIR_INDEX_KEY;
926 struct btrfs_delayed_item *item;
928 mutex_lock(&node->mutex);
929 item = __btrfs_first_delayed_deletion_item(node);
931 mutex_unlock(&node->mutex);
935 key.offset = item->index;
936 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
939 * There's no matching item in the leaf. This means we
940 * have already deleted this item in a past run of the
941 * delayed items. We ignore errors when running delayed
942 * items from an async context, through a work queue job
943 * running btrfs_async_run_delayed_root(), and don't
944 * release delayed items that failed to complete. This
945 * is because we will retry later, and at transaction
946 * commit time we always run delayed items and will
947 * then deal with errors if they fail to run again.
949 * So just release delayed items for which we can't find
950 * an item in the tree, and move to the next item.
952 btrfs_release_path(path);
953 btrfs_release_delayed_item(item);
955 } else if (ret == 0) {
956 ret = btrfs_batch_delete_items(trans, root, path, item);
957 btrfs_release_path(path);
961 * We unlock and relock on each iteration, this is to prevent
962 * blocking other tasks for too long while we are being run from
963 * the async context (work queue job). Those tasks are typically
964 * running system calls like creat/mkdir/rename/unlink/etc which
965 * need to add delayed items to this delayed node.
967 mutex_unlock(&node->mutex);
973 static void btrfs_release_delayed_inode(struct btrfs_delayed_node *delayed_node)
975 struct btrfs_delayed_root *delayed_root;
978 test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
979 BUG_ON(!delayed_node->root);
980 clear_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags);
981 delayed_node->count--;
983 delayed_root = delayed_node->root->fs_info->delayed_root;
984 finish_one_item(delayed_root);
988 static void btrfs_release_delayed_iref(struct btrfs_delayed_node *delayed_node)
991 if (test_and_clear_bit(BTRFS_DELAYED_NODE_DEL_IREF, &delayed_node->flags)) {
992 struct btrfs_delayed_root *delayed_root;
994 ASSERT(delayed_node->root);
995 delayed_node->count--;
997 delayed_root = delayed_node->root->fs_info->delayed_root;
998 finish_one_item(delayed_root);
1002 static int __btrfs_update_delayed_inode(struct btrfs_trans_handle *trans,
1003 struct btrfs_root *root,
1004 struct btrfs_path *path,
1005 struct btrfs_delayed_node *node)
1007 struct btrfs_fs_info *fs_info = root->fs_info;
1008 struct btrfs_key key;
1009 struct btrfs_inode_item *inode_item;
1010 struct extent_buffer *leaf;
1014 key.objectid = node->inode_id;
1015 key.type = BTRFS_INODE_ITEM_KEY;
1018 if (test_bit(BTRFS_DELAYED_NODE_DEL_IREF, &node->flags))
1023 ret = btrfs_lookup_inode(trans, root, path, &key, mod);
1029 leaf = path->nodes[0];
1030 inode_item = btrfs_item_ptr(leaf, path->slots[0],
1031 struct btrfs_inode_item);
1032 write_extent_buffer(leaf, &node->inode_item, (unsigned long)inode_item,
1033 sizeof(struct btrfs_inode_item));
1034 btrfs_mark_buffer_dirty(trans, leaf);
1036 if (!test_bit(BTRFS_DELAYED_NODE_DEL_IREF, &node->flags))
1040 if (path->slots[0] >= btrfs_header_nritems(leaf))
1043 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1044 if (key.objectid != node->inode_id)
1047 if (key.type != BTRFS_INODE_REF_KEY &&
1048 key.type != BTRFS_INODE_EXTREF_KEY)
1052 * Delayed iref deletion is for the inode who has only one link,
1053 * so there is only one iref. The case that several irefs are
1054 * in the same item doesn't exist.
1056 ret = btrfs_del_item(trans, root, path);
1058 btrfs_release_delayed_iref(node);
1059 btrfs_release_path(path);
1061 btrfs_delayed_inode_release_metadata(fs_info, node, (ret < 0));
1062 btrfs_release_delayed_inode(node);
1065 * If we fail to update the delayed inode we need to abort the
1066 * transaction, because we could leave the inode with the improper
1069 if (ret && ret != -ENOENT)
1070 btrfs_abort_transaction(trans, ret);
1075 btrfs_release_path(path);
1077 key.type = BTRFS_INODE_EXTREF_KEY;
1080 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1086 leaf = path->nodes[0];
1091 static inline int btrfs_update_delayed_inode(struct btrfs_trans_handle *trans,
1092 struct btrfs_root *root,
1093 struct btrfs_path *path,
1094 struct btrfs_delayed_node *node)
1098 mutex_lock(&node->mutex);
1099 if (!test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &node->flags)) {
1100 mutex_unlock(&node->mutex);
1104 ret = __btrfs_update_delayed_inode(trans, root, path, node);
1105 mutex_unlock(&node->mutex);
1110 __btrfs_commit_inode_delayed_items(struct btrfs_trans_handle *trans,
1111 struct btrfs_path *path,
1112 struct btrfs_delayed_node *node)
1116 ret = btrfs_insert_delayed_items(trans, path, node->root, node);
1120 ret = btrfs_delete_delayed_items(trans, path, node->root, node);
1124 ret = btrfs_update_delayed_inode(trans, node->root, path, node);
1129 * Called when committing the transaction.
1130 * Returns 0 on success.
1131 * Returns < 0 on error and returns with an aborted transaction with any
1132 * outstanding delayed items cleaned up.
1134 static int __btrfs_run_delayed_items(struct btrfs_trans_handle *trans, int nr)
1136 struct btrfs_fs_info *fs_info = trans->fs_info;
1137 struct btrfs_delayed_root *delayed_root;
1138 struct btrfs_delayed_node *curr_node, *prev_node;
1139 struct btrfs_path *path;
1140 struct btrfs_block_rsv *block_rsv;
1142 bool count = (nr > 0);
1144 if (TRANS_ABORTED(trans))
1147 path = btrfs_alloc_path();
1151 block_rsv = trans->block_rsv;
1152 trans->block_rsv = &fs_info->delayed_block_rsv;
1154 delayed_root = fs_info->delayed_root;
1156 curr_node = btrfs_first_delayed_node(delayed_root);
1157 while (curr_node && (!count || nr--)) {
1158 ret = __btrfs_commit_inode_delayed_items(trans, path,
1161 btrfs_abort_transaction(trans, ret);
1165 prev_node = curr_node;
1166 curr_node = btrfs_next_delayed_node(curr_node);
1168 * See the comment below about releasing path before releasing
1169 * node. If the commit of delayed items was successful the path
1170 * should always be released, but in case of an error, it may
1171 * point to locked extent buffers (a leaf at the very least).
1173 ASSERT(path->nodes[0] == NULL);
1174 btrfs_release_delayed_node(prev_node);
1178 * Release the path to avoid a potential deadlock and lockdep splat when
1179 * releasing the delayed node, as that requires taking the delayed node's
1180 * mutex. If another task starts running delayed items before we take
1181 * the mutex, it will first lock the mutex and then it may try to lock
1182 * the same btree path (leaf).
1184 btrfs_free_path(path);
1187 btrfs_release_delayed_node(curr_node);
1188 trans->block_rsv = block_rsv;
1193 int btrfs_run_delayed_items(struct btrfs_trans_handle *trans)
1195 return __btrfs_run_delayed_items(trans, -1);
1198 int btrfs_run_delayed_items_nr(struct btrfs_trans_handle *trans, int nr)
1200 return __btrfs_run_delayed_items(trans, nr);
1203 int btrfs_commit_inode_delayed_items(struct btrfs_trans_handle *trans,
1204 struct btrfs_inode *inode)
1206 struct btrfs_delayed_node *delayed_node = btrfs_get_delayed_node(inode);
1207 struct btrfs_path *path;
1208 struct btrfs_block_rsv *block_rsv;
1214 mutex_lock(&delayed_node->mutex);
1215 if (!delayed_node->count) {
1216 mutex_unlock(&delayed_node->mutex);
1217 btrfs_release_delayed_node(delayed_node);
1220 mutex_unlock(&delayed_node->mutex);
1222 path = btrfs_alloc_path();
1224 btrfs_release_delayed_node(delayed_node);
1228 block_rsv = trans->block_rsv;
1229 trans->block_rsv = &delayed_node->root->fs_info->delayed_block_rsv;
1231 ret = __btrfs_commit_inode_delayed_items(trans, path, delayed_node);
1233 btrfs_release_delayed_node(delayed_node);
1234 btrfs_free_path(path);
1235 trans->block_rsv = block_rsv;
1240 int btrfs_commit_inode_delayed_inode(struct btrfs_inode *inode)
1242 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1243 struct btrfs_trans_handle *trans;
1244 struct btrfs_delayed_node *delayed_node = btrfs_get_delayed_node(inode);
1245 struct btrfs_path *path;
1246 struct btrfs_block_rsv *block_rsv;
1252 mutex_lock(&delayed_node->mutex);
1253 if (!test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
1254 mutex_unlock(&delayed_node->mutex);
1255 btrfs_release_delayed_node(delayed_node);
1258 mutex_unlock(&delayed_node->mutex);
1260 trans = btrfs_join_transaction(delayed_node->root);
1261 if (IS_ERR(trans)) {
1262 ret = PTR_ERR(trans);
1266 path = btrfs_alloc_path();
1272 block_rsv = trans->block_rsv;
1273 trans->block_rsv = &fs_info->delayed_block_rsv;
1275 mutex_lock(&delayed_node->mutex);
1276 if (test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags))
1277 ret = __btrfs_update_delayed_inode(trans, delayed_node->root,
1278 path, delayed_node);
1281 mutex_unlock(&delayed_node->mutex);
1283 btrfs_free_path(path);
1284 trans->block_rsv = block_rsv;
1286 btrfs_end_transaction(trans);
1287 btrfs_btree_balance_dirty(fs_info);
1289 btrfs_release_delayed_node(delayed_node);
1294 void btrfs_remove_delayed_node(struct btrfs_inode *inode)
1296 struct btrfs_delayed_node *delayed_node;
1298 delayed_node = READ_ONCE(inode->delayed_node);
1302 inode->delayed_node = NULL;
1303 btrfs_release_delayed_node(delayed_node);
1306 struct btrfs_async_delayed_work {
1307 struct btrfs_delayed_root *delayed_root;
1309 struct btrfs_work work;
1312 static void btrfs_async_run_delayed_root(struct btrfs_work *work)
1314 struct btrfs_async_delayed_work *async_work;
1315 struct btrfs_delayed_root *delayed_root;
1316 struct btrfs_trans_handle *trans;
1317 struct btrfs_path *path;
1318 struct btrfs_delayed_node *delayed_node = NULL;
1319 struct btrfs_root *root;
1320 struct btrfs_block_rsv *block_rsv;
1323 async_work = container_of(work, struct btrfs_async_delayed_work, work);
1324 delayed_root = async_work->delayed_root;
1326 path = btrfs_alloc_path();
1331 if (atomic_read(&delayed_root->items) <
1332 BTRFS_DELAYED_BACKGROUND / 2)
1335 delayed_node = btrfs_first_prepared_delayed_node(delayed_root);
1339 root = delayed_node->root;
1341 trans = btrfs_join_transaction(root);
1342 if (IS_ERR(trans)) {
1343 btrfs_release_path(path);
1344 btrfs_release_prepared_delayed_node(delayed_node);
1349 block_rsv = trans->block_rsv;
1350 trans->block_rsv = &root->fs_info->delayed_block_rsv;
1352 __btrfs_commit_inode_delayed_items(trans, path, delayed_node);
1354 trans->block_rsv = block_rsv;
1355 btrfs_end_transaction(trans);
1356 btrfs_btree_balance_dirty_nodelay(root->fs_info);
1358 btrfs_release_path(path);
1359 btrfs_release_prepared_delayed_node(delayed_node);
1362 } while ((async_work->nr == 0 && total_done < BTRFS_DELAYED_WRITEBACK)
1363 || total_done < async_work->nr);
1365 btrfs_free_path(path);
1367 wake_up(&delayed_root->wait);
1372 static int btrfs_wq_run_delayed_node(struct btrfs_delayed_root *delayed_root,
1373 struct btrfs_fs_info *fs_info, int nr)
1375 struct btrfs_async_delayed_work *async_work;
1377 async_work = kmalloc(sizeof(*async_work), GFP_NOFS);
1381 async_work->delayed_root = delayed_root;
1382 btrfs_init_work(&async_work->work, btrfs_async_run_delayed_root, NULL);
1383 async_work->nr = nr;
1385 btrfs_queue_work(fs_info->delayed_workers, &async_work->work);
1389 void btrfs_assert_delayed_root_empty(struct btrfs_fs_info *fs_info)
1391 WARN_ON(btrfs_first_delayed_node(fs_info->delayed_root));
1394 static int could_end_wait(struct btrfs_delayed_root *delayed_root, int seq)
1396 int val = atomic_read(&delayed_root->items_seq);
1398 if (val < seq || val >= seq + BTRFS_DELAYED_BATCH)
1401 if (atomic_read(&delayed_root->items) < BTRFS_DELAYED_BACKGROUND)
1407 void btrfs_balance_delayed_items(struct btrfs_fs_info *fs_info)
1409 struct btrfs_delayed_root *delayed_root = fs_info->delayed_root;
1411 if ((atomic_read(&delayed_root->items) < BTRFS_DELAYED_BACKGROUND) ||
1412 btrfs_workqueue_normal_congested(fs_info->delayed_workers))
1415 if (atomic_read(&delayed_root->items) >= BTRFS_DELAYED_WRITEBACK) {
1419 seq = atomic_read(&delayed_root->items_seq);
1421 ret = btrfs_wq_run_delayed_node(delayed_root, fs_info, 0);
1425 wait_event_interruptible(delayed_root->wait,
1426 could_end_wait(delayed_root, seq));
1430 btrfs_wq_run_delayed_node(delayed_root, fs_info, BTRFS_DELAYED_BATCH);
1433 static void btrfs_release_dir_index_item_space(struct btrfs_trans_handle *trans)
1435 struct btrfs_fs_info *fs_info = trans->fs_info;
1436 const u64 bytes = btrfs_calc_insert_metadata_size(fs_info, 1);
1438 if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
1442 * Adding the new dir index item does not require touching another
1443 * leaf, so we can release 1 unit of metadata that was previously
1444 * reserved when starting the transaction. This applies only to
1445 * the case where we had a transaction start and excludes the
1446 * transaction join case (when replaying log trees).
1448 trace_btrfs_space_reservation(fs_info, "transaction",
1449 trans->transid, bytes, 0);
1450 btrfs_block_rsv_release(fs_info, trans->block_rsv, bytes, NULL);
1451 ASSERT(trans->bytes_reserved >= bytes);
1452 trans->bytes_reserved -= bytes;
1455 /* Will return 0, -ENOMEM or -EEXIST (index number collision, unexpected). */
1456 int btrfs_insert_delayed_dir_index(struct btrfs_trans_handle *trans,
1457 const char *name, int name_len,
1458 struct btrfs_inode *dir,
1459 struct btrfs_disk_key *disk_key, u8 flags,
1462 struct btrfs_fs_info *fs_info = trans->fs_info;
1463 const unsigned int leaf_data_size = BTRFS_LEAF_DATA_SIZE(fs_info);
1464 struct btrfs_delayed_node *delayed_node;
1465 struct btrfs_delayed_item *delayed_item;
1466 struct btrfs_dir_item *dir_item;
1467 bool reserve_leaf_space;
1471 delayed_node = btrfs_get_or_create_delayed_node(dir);
1472 if (IS_ERR(delayed_node))
1473 return PTR_ERR(delayed_node);
1475 delayed_item = btrfs_alloc_delayed_item(sizeof(*dir_item) + name_len,
1477 BTRFS_DELAYED_INSERTION_ITEM);
1478 if (!delayed_item) {
1483 delayed_item->index = index;
1485 dir_item = (struct btrfs_dir_item *)delayed_item->data;
1486 dir_item->location = *disk_key;
1487 btrfs_set_stack_dir_transid(dir_item, trans->transid);
1488 btrfs_set_stack_dir_data_len(dir_item, 0);
1489 btrfs_set_stack_dir_name_len(dir_item, name_len);
1490 btrfs_set_stack_dir_flags(dir_item, flags);
1491 memcpy((char *)(dir_item + 1), name, name_len);
1493 data_len = delayed_item->data_len + sizeof(struct btrfs_item);
1495 mutex_lock(&delayed_node->mutex);
1498 * First attempt to insert the delayed item. This is to make the error
1499 * handling path simpler in case we fail (-EEXIST). There's no risk of
1500 * any other task coming in and running the delayed item before we do
1501 * the metadata space reservation below, because we are holding the
1502 * delayed node's mutex and that mutex must also be locked before the
1503 * node's delayed items can be run.
1505 ret = __btrfs_add_delayed_item(delayed_node, delayed_item);
1506 if (unlikely(ret)) {
1507 btrfs_err(trans->fs_info,
1508 "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",
1509 name_len, name, index, btrfs_root_id(delayed_node->root),
1510 delayed_node->inode_id, dir->index_cnt,
1511 delayed_node->index_cnt, ret);
1512 btrfs_release_delayed_item(delayed_item);
1513 btrfs_release_dir_index_item_space(trans);
1514 mutex_unlock(&delayed_node->mutex);
1518 if (delayed_node->index_item_leaves == 0 ||
1519 delayed_node->curr_index_batch_size + data_len > leaf_data_size) {
1520 delayed_node->curr_index_batch_size = data_len;
1521 reserve_leaf_space = true;
1523 delayed_node->curr_index_batch_size += data_len;
1524 reserve_leaf_space = false;
1527 if (reserve_leaf_space) {
1528 ret = btrfs_delayed_item_reserve_metadata(trans, delayed_item);
1530 * Space was reserved for a dir index item insertion when we
1531 * started the transaction, so getting a failure here should be
1535 btrfs_release_delayed_item(delayed_item);
1536 mutex_unlock(&delayed_node->mutex);
1540 delayed_node->index_item_leaves++;
1542 btrfs_release_dir_index_item_space(trans);
1544 mutex_unlock(&delayed_node->mutex);
1547 btrfs_release_delayed_node(delayed_node);
1551 static int btrfs_delete_delayed_insertion_item(struct btrfs_fs_info *fs_info,
1552 struct btrfs_delayed_node *node,
1555 struct btrfs_delayed_item *item;
1557 mutex_lock(&node->mutex);
1558 item = __btrfs_lookup_delayed_item(&node->ins_root.rb_root, index);
1560 mutex_unlock(&node->mutex);
1565 * For delayed items to insert, we track reserved metadata bytes based
1566 * on the number of leaves that we will use.
1567 * See btrfs_insert_delayed_dir_index() and
1568 * btrfs_delayed_item_reserve_metadata()).
1570 ASSERT(item->bytes_reserved == 0);
1571 ASSERT(node->index_item_leaves > 0);
1574 * If there's only one leaf reserved, we can decrement this item from the
1575 * current batch, otherwise we can not because we don't know which leaf
1576 * it belongs to. With the current limit on delayed items, we rarely
1577 * accumulate enough dir index items to fill more than one leaf (even
1578 * when using a leaf size of 4K).
1580 if (node->index_item_leaves == 1) {
1581 const u32 data_len = item->data_len + sizeof(struct btrfs_item);
1583 ASSERT(node->curr_index_batch_size >= data_len);
1584 node->curr_index_batch_size -= data_len;
1587 btrfs_release_delayed_item(item);
1589 /* If we now have no more dir index items, we can release all leaves. */
1590 if (RB_EMPTY_ROOT(&node->ins_root.rb_root)) {
1591 btrfs_delayed_item_release_leaves(node, node->index_item_leaves);
1592 node->index_item_leaves = 0;
1595 mutex_unlock(&node->mutex);
1599 int btrfs_delete_delayed_dir_index(struct btrfs_trans_handle *trans,
1600 struct btrfs_inode *dir, u64 index)
1602 struct btrfs_delayed_node *node;
1603 struct btrfs_delayed_item *item;
1606 node = btrfs_get_or_create_delayed_node(dir);
1608 return PTR_ERR(node);
1610 ret = btrfs_delete_delayed_insertion_item(trans->fs_info, node, index);
1614 item = btrfs_alloc_delayed_item(0, node, BTRFS_DELAYED_DELETION_ITEM);
1620 item->index = index;
1622 ret = btrfs_delayed_item_reserve_metadata(trans, item);
1624 * we have reserved enough space when we start a new transaction,
1625 * so reserving metadata failure is impossible.
1628 btrfs_err(trans->fs_info,
1629 "metadata reservation failed for delayed dir item deltiona, should have been reserved");
1630 btrfs_release_delayed_item(item);
1634 mutex_lock(&node->mutex);
1635 ret = __btrfs_add_delayed_item(node, item);
1636 if (unlikely(ret)) {
1637 btrfs_err(trans->fs_info,
1638 "err add delayed dir index item(index: %llu) into the deletion tree of the delayed node(root id: %llu, inode id: %llu, errno: %d)",
1639 index, node->root->root_key.objectid,
1640 node->inode_id, ret);
1641 btrfs_delayed_item_release_metadata(dir->root, item);
1642 btrfs_release_delayed_item(item);
1644 mutex_unlock(&node->mutex);
1646 btrfs_release_delayed_node(node);
1650 int btrfs_inode_delayed_dir_index_count(struct btrfs_inode *inode)
1652 struct btrfs_delayed_node *delayed_node = btrfs_get_delayed_node(inode);
1658 * Since we have held i_mutex of this directory, it is impossible that
1659 * a new directory index is added into the delayed node and index_cnt
1660 * is updated now. So we needn't lock the delayed node.
1662 if (!delayed_node->index_cnt) {
1663 btrfs_release_delayed_node(delayed_node);
1667 inode->index_cnt = delayed_node->index_cnt;
1668 btrfs_release_delayed_node(delayed_node);
1672 bool btrfs_readdir_get_delayed_items(struct inode *inode,
1674 struct list_head *ins_list,
1675 struct list_head *del_list)
1677 struct btrfs_delayed_node *delayed_node;
1678 struct btrfs_delayed_item *item;
1680 delayed_node = btrfs_get_delayed_node(BTRFS_I(inode));
1685 * We can only do one readdir with delayed items at a time because of
1686 * item->readdir_list.
1688 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_SHARED);
1689 btrfs_inode_lock(BTRFS_I(inode), 0);
1691 mutex_lock(&delayed_node->mutex);
1692 item = __btrfs_first_delayed_insertion_item(delayed_node);
1693 while (item && item->index <= last_index) {
1694 refcount_inc(&item->refs);
1695 list_add_tail(&item->readdir_list, ins_list);
1696 item = __btrfs_next_delayed_item(item);
1699 item = __btrfs_first_delayed_deletion_item(delayed_node);
1700 while (item && item->index <= last_index) {
1701 refcount_inc(&item->refs);
1702 list_add_tail(&item->readdir_list, del_list);
1703 item = __btrfs_next_delayed_item(item);
1705 mutex_unlock(&delayed_node->mutex);
1707 * This delayed node is still cached in the btrfs inode, so refs
1708 * must be > 1 now, and we needn't check it is going to be freed
1711 * Besides that, this function is used to read dir, we do not
1712 * insert/delete delayed items in this period. So we also needn't
1713 * requeue or dequeue this delayed node.
1715 refcount_dec(&delayed_node->refs);
1720 void btrfs_readdir_put_delayed_items(struct inode *inode,
1721 struct list_head *ins_list,
1722 struct list_head *del_list)
1724 struct btrfs_delayed_item *curr, *next;
1726 list_for_each_entry_safe(curr, next, ins_list, readdir_list) {
1727 list_del(&curr->readdir_list);
1728 if (refcount_dec_and_test(&curr->refs))
1732 list_for_each_entry_safe(curr, next, del_list, readdir_list) {
1733 list_del(&curr->readdir_list);
1734 if (refcount_dec_and_test(&curr->refs))
1739 * The VFS is going to do up_read(), so we need to downgrade back to a
1742 downgrade_write(&inode->i_rwsem);
1745 int btrfs_should_delete_dir_index(struct list_head *del_list,
1748 struct btrfs_delayed_item *curr;
1751 list_for_each_entry(curr, del_list, readdir_list) {
1752 if (curr->index > index)
1754 if (curr->index == index) {
1763 * Read dir info stored in the delayed tree.
1765 int btrfs_readdir_delayed_dir_index(struct dir_context *ctx,
1766 struct list_head *ins_list)
1768 struct btrfs_dir_item *di;
1769 struct btrfs_delayed_item *curr, *next;
1770 struct btrfs_key location;
1774 unsigned char d_type;
1777 * Changing the data of the delayed item is impossible. So
1778 * we needn't lock them. And we have held i_mutex of the
1779 * directory, nobody can delete any directory indexes now.
1781 list_for_each_entry_safe(curr, next, ins_list, readdir_list) {
1782 list_del(&curr->readdir_list);
1784 if (curr->index < ctx->pos) {
1785 if (refcount_dec_and_test(&curr->refs))
1790 ctx->pos = curr->index;
1792 di = (struct btrfs_dir_item *)curr->data;
1793 name = (char *)(di + 1);
1794 name_len = btrfs_stack_dir_name_len(di);
1796 d_type = fs_ftype_to_dtype(btrfs_dir_flags_to_ftype(di->type));
1797 btrfs_disk_key_to_cpu(&location, &di->location);
1799 over = !dir_emit(ctx, name, name_len,
1800 location.objectid, d_type);
1802 if (refcount_dec_and_test(&curr->refs))
1812 static void fill_stack_inode_item(struct btrfs_trans_handle *trans,
1813 struct btrfs_inode_item *inode_item,
1814 struct inode *inode)
1818 btrfs_set_stack_inode_uid(inode_item, i_uid_read(inode));
1819 btrfs_set_stack_inode_gid(inode_item, i_gid_read(inode));
1820 btrfs_set_stack_inode_size(inode_item, BTRFS_I(inode)->disk_i_size);
1821 btrfs_set_stack_inode_mode(inode_item, inode->i_mode);
1822 btrfs_set_stack_inode_nlink(inode_item, inode->i_nlink);
1823 btrfs_set_stack_inode_nbytes(inode_item, inode_get_bytes(inode));
1824 btrfs_set_stack_inode_generation(inode_item,
1825 BTRFS_I(inode)->generation);
1826 btrfs_set_stack_inode_sequence(inode_item,
1827 inode_peek_iversion(inode));
1828 btrfs_set_stack_inode_transid(inode_item, trans->transid);
1829 btrfs_set_stack_inode_rdev(inode_item, inode->i_rdev);
1830 flags = btrfs_inode_combine_flags(BTRFS_I(inode)->flags,
1831 BTRFS_I(inode)->ro_flags);
1832 btrfs_set_stack_inode_flags(inode_item, flags);
1833 btrfs_set_stack_inode_block_group(inode_item, 0);
1835 btrfs_set_stack_timespec_sec(&inode_item->atime,
1836 inode->i_atime.tv_sec);
1837 btrfs_set_stack_timespec_nsec(&inode_item->atime,
1838 inode->i_atime.tv_nsec);
1840 btrfs_set_stack_timespec_sec(&inode_item->mtime,
1841 inode->i_mtime.tv_sec);
1842 btrfs_set_stack_timespec_nsec(&inode_item->mtime,
1843 inode->i_mtime.tv_nsec);
1845 btrfs_set_stack_timespec_sec(&inode_item->ctime,
1846 inode_get_ctime(inode).tv_sec);
1847 btrfs_set_stack_timespec_nsec(&inode_item->ctime,
1848 inode_get_ctime(inode).tv_nsec);
1850 btrfs_set_stack_timespec_sec(&inode_item->otime,
1851 BTRFS_I(inode)->i_otime.tv_sec);
1852 btrfs_set_stack_timespec_nsec(&inode_item->otime,
1853 BTRFS_I(inode)->i_otime.tv_nsec);
1856 int btrfs_fill_inode(struct inode *inode, u32 *rdev)
1858 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
1859 struct btrfs_delayed_node *delayed_node;
1860 struct btrfs_inode_item *inode_item;
1862 delayed_node = btrfs_get_delayed_node(BTRFS_I(inode));
1866 mutex_lock(&delayed_node->mutex);
1867 if (!test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
1868 mutex_unlock(&delayed_node->mutex);
1869 btrfs_release_delayed_node(delayed_node);
1873 inode_item = &delayed_node->inode_item;
1875 i_uid_write(inode, btrfs_stack_inode_uid(inode_item));
1876 i_gid_write(inode, btrfs_stack_inode_gid(inode_item));
1877 btrfs_i_size_write(BTRFS_I(inode), btrfs_stack_inode_size(inode_item));
1878 btrfs_inode_set_file_extent_range(BTRFS_I(inode), 0,
1879 round_up(i_size_read(inode), fs_info->sectorsize));
1880 inode->i_mode = btrfs_stack_inode_mode(inode_item);
1881 set_nlink(inode, btrfs_stack_inode_nlink(inode_item));
1882 inode_set_bytes(inode, btrfs_stack_inode_nbytes(inode_item));
1883 BTRFS_I(inode)->generation = btrfs_stack_inode_generation(inode_item);
1884 BTRFS_I(inode)->last_trans = btrfs_stack_inode_transid(inode_item);
1886 inode_set_iversion_queried(inode,
1887 btrfs_stack_inode_sequence(inode_item));
1889 *rdev = btrfs_stack_inode_rdev(inode_item);
1890 btrfs_inode_split_flags(btrfs_stack_inode_flags(inode_item),
1891 &BTRFS_I(inode)->flags, &BTRFS_I(inode)->ro_flags);
1893 inode->i_atime.tv_sec = btrfs_stack_timespec_sec(&inode_item->atime);
1894 inode->i_atime.tv_nsec = btrfs_stack_timespec_nsec(&inode_item->atime);
1896 inode->i_mtime.tv_sec = btrfs_stack_timespec_sec(&inode_item->mtime);
1897 inode->i_mtime.tv_nsec = btrfs_stack_timespec_nsec(&inode_item->mtime);
1899 inode_set_ctime(inode, btrfs_stack_timespec_sec(&inode_item->ctime),
1900 btrfs_stack_timespec_nsec(&inode_item->ctime));
1902 BTRFS_I(inode)->i_otime.tv_sec =
1903 btrfs_stack_timespec_sec(&inode_item->otime);
1904 BTRFS_I(inode)->i_otime.tv_nsec =
1905 btrfs_stack_timespec_nsec(&inode_item->otime);
1907 inode->i_generation = BTRFS_I(inode)->generation;
1908 BTRFS_I(inode)->index_cnt = (u64)-1;
1910 mutex_unlock(&delayed_node->mutex);
1911 btrfs_release_delayed_node(delayed_node);
1915 int btrfs_delayed_update_inode(struct btrfs_trans_handle *trans,
1916 struct btrfs_inode *inode)
1918 struct btrfs_root *root = inode->root;
1919 struct btrfs_delayed_node *delayed_node;
1922 delayed_node = btrfs_get_or_create_delayed_node(inode);
1923 if (IS_ERR(delayed_node))
1924 return PTR_ERR(delayed_node);
1926 mutex_lock(&delayed_node->mutex);
1927 if (test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
1928 fill_stack_inode_item(trans, &delayed_node->inode_item,
1933 ret = btrfs_delayed_inode_reserve_metadata(trans, root, delayed_node);
1937 fill_stack_inode_item(trans, &delayed_node->inode_item, &inode->vfs_inode);
1938 set_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags);
1939 delayed_node->count++;
1940 atomic_inc(&root->fs_info->delayed_root->items);
1942 mutex_unlock(&delayed_node->mutex);
1943 btrfs_release_delayed_node(delayed_node);
1947 int btrfs_delayed_delete_inode_ref(struct btrfs_inode *inode)
1949 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1950 struct btrfs_delayed_node *delayed_node;
1953 * we don't do delayed inode updates during log recovery because it
1954 * leads to enospc problems. This means we also can't do
1955 * delayed inode refs
1957 if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
1960 delayed_node = btrfs_get_or_create_delayed_node(inode);
1961 if (IS_ERR(delayed_node))
1962 return PTR_ERR(delayed_node);
1965 * We don't reserve space for inode ref deletion is because:
1966 * - We ONLY do async inode ref deletion for the inode who has only
1967 * one link(i_nlink == 1), it means there is only one inode ref.
1968 * And in most case, the inode ref and the inode item are in the
1969 * same leaf, and we will deal with them at the same time.
1970 * Since we are sure we will reserve the space for the inode item,
1971 * it is unnecessary to reserve space for inode ref deletion.
1972 * - If the inode ref and the inode item are not in the same leaf,
1973 * We also needn't worry about enospc problem, because we reserve
1974 * much more space for the inode update than it needs.
1975 * - At the worst, we can steal some space from the global reservation.
1978 mutex_lock(&delayed_node->mutex);
1979 if (test_bit(BTRFS_DELAYED_NODE_DEL_IREF, &delayed_node->flags))
1982 set_bit(BTRFS_DELAYED_NODE_DEL_IREF, &delayed_node->flags);
1983 delayed_node->count++;
1984 atomic_inc(&fs_info->delayed_root->items);
1986 mutex_unlock(&delayed_node->mutex);
1987 btrfs_release_delayed_node(delayed_node);
1991 static void __btrfs_kill_delayed_node(struct btrfs_delayed_node *delayed_node)
1993 struct btrfs_root *root = delayed_node->root;
1994 struct btrfs_fs_info *fs_info = root->fs_info;
1995 struct btrfs_delayed_item *curr_item, *prev_item;
1997 mutex_lock(&delayed_node->mutex);
1998 curr_item = __btrfs_first_delayed_insertion_item(delayed_node);
2000 prev_item = curr_item;
2001 curr_item = __btrfs_next_delayed_item(prev_item);
2002 btrfs_release_delayed_item(prev_item);
2005 if (delayed_node->index_item_leaves > 0) {
2006 btrfs_delayed_item_release_leaves(delayed_node,
2007 delayed_node->index_item_leaves);
2008 delayed_node->index_item_leaves = 0;
2011 curr_item = __btrfs_first_delayed_deletion_item(delayed_node);
2013 btrfs_delayed_item_release_metadata(root, curr_item);
2014 prev_item = curr_item;
2015 curr_item = __btrfs_next_delayed_item(prev_item);
2016 btrfs_release_delayed_item(prev_item);
2019 btrfs_release_delayed_iref(delayed_node);
2021 if (test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
2022 btrfs_delayed_inode_release_metadata(fs_info, delayed_node, false);
2023 btrfs_release_delayed_inode(delayed_node);
2025 mutex_unlock(&delayed_node->mutex);
2028 void btrfs_kill_delayed_inode_items(struct btrfs_inode *inode)
2030 struct btrfs_delayed_node *delayed_node;
2032 delayed_node = btrfs_get_delayed_node(inode);
2036 __btrfs_kill_delayed_node(delayed_node);
2037 btrfs_release_delayed_node(delayed_node);
2040 void btrfs_kill_all_delayed_nodes(struct btrfs_root *root)
2043 struct btrfs_delayed_node *delayed_nodes[8];
2047 spin_lock(&root->inode_lock);
2048 n = radix_tree_gang_lookup(&root->delayed_nodes_tree,
2049 (void **)delayed_nodes, inode_id,
2050 ARRAY_SIZE(delayed_nodes));
2052 spin_unlock(&root->inode_lock);
2056 inode_id = delayed_nodes[n - 1]->inode_id + 1;
2057 for (i = 0; i < n; i++) {
2059 * Don't increase refs in case the node is dead and
2060 * about to be removed from the tree in the loop below
2062 if (!refcount_inc_not_zero(&delayed_nodes[i]->refs))
2063 delayed_nodes[i] = NULL;
2065 spin_unlock(&root->inode_lock);
2067 for (i = 0; i < n; i++) {
2068 if (!delayed_nodes[i])
2070 __btrfs_kill_delayed_node(delayed_nodes[i]);
2071 btrfs_release_delayed_node(delayed_nodes[i]);
2076 void btrfs_destroy_delayed_inodes(struct btrfs_fs_info *fs_info)
2078 struct btrfs_delayed_node *curr_node, *prev_node;
2080 curr_node = btrfs_first_delayed_node(fs_info->delayed_root);
2082 __btrfs_kill_delayed_node(curr_node);
2084 prev_node = curr_node;
2085 curr_node = btrfs_next_delayed_node(curr_node);
2086 btrfs_release_delayed_node(prev_node);
2090 void btrfs_log_get_delayed_items(struct btrfs_inode *inode,
2091 struct list_head *ins_list,
2092 struct list_head *del_list)
2094 struct btrfs_delayed_node *node;
2095 struct btrfs_delayed_item *item;
2097 node = btrfs_get_delayed_node(inode);
2101 mutex_lock(&node->mutex);
2102 item = __btrfs_first_delayed_insertion_item(node);
2105 * It's possible that the item is already in a log list. This
2106 * can happen in case two tasks are trying to log the same
2107 * directory. For example if we have tasks A and task B:
2109 * Task A collected the delayed items into a log list while
2110 * under the inode's log_mutex (at btrfs_log_inode()), but it
2111 * only releases the items after logging the inodes they point
2112 * to (if they are new inodes), which happens after unlocking
2115 * Task B enters btrfs_log_inode() and acquires the log_mutex
2116 * of the same directory inode, before task B releases the
2117 * delayed items. This can happen for example when logging some
2118 * inode we need to trigger logging of its parent directory, so
2119 * logging two files that have the same parent directory can
2122 * If this happens, just ignore delayed items already in a log
2123 * list. All the tasks logging the directory are under a log
2124 * transaction and whichever finishes first can not sync the log
2125 * before the other completes and leaves the log transaction.
2127 if (!item->logged && list_empty(&item->log_list)) {
2128 refcount_inc(&item->refs);
2129 list_add_tail(&item->log_list, ins_list);
2131 item = __btrfs_next_delayed_item(item);
2134 item = __btrfs_first_delayed_deletion_item(node);
2136 /* It may be non-empty, for the same reason mentioned above. */
2137 if (!item->logged && list_empty(&item->log_list)) {
2138 refcount_inc(&item->refs);
2139 list_add_tail(&item->log_list, del_list);
2141 item = __btrfs_next_delayed_item(item);
2143 mutex_unlock(&node->mutex);
2146 * We are called during inode logging, which means the inode is in use
2147 * and can not be evicted before we finish logging the inode. So we never
2148 * have the last reference on the delayed inode.
2149 * Also, we don't use btrfs_release_delayed_node() because that would
2150 * requeue the delayed inode (change its order in the list of prepared
2151 * nodes) and we don't want to do such change because we don't create or
2152 * delete delayed items.
2154 ASSERT(refcount_read(&node->refs) > 1);
2155 refcount_dec(&node->refs);
2158 void btrfs_log_put_delayed_items(struct btrfs_inode *inode,
2159 struct list_head *ins_list,
2160 struct list_head *del_list)
2162 struct btrfs_delayed_node *node;
2163 struct btrfs_delayed_item *item;
2164 struct btrfs_delayed_item *next;
2166 node = btrfs_get_delayed_node(inode);
2170 mutex_lock(&node->mutex);
2172 list_for_each_entry_safe(item, next, ins_list, log_list) {
2173 item->logged = true;
2174 list_del_init(&item->log_list);
2175 if (refcount_dec_and_test(&item->refs))
2179 list_for_each_entry_safe(item, next, del_list, log_list) {
2180 item->logged = true;
2181 list_del_init(&item->log_list);
2182 if (refcount_dec_and_test(&item->refs))
2186 mutex_unlock(&node->mutex);
2189 * We are called during inode logging, which means the inode is in use
2190 * and can not be evicted before we finish logging the inode. So we never
2191 * have the last reference on the delayed inode.
2192 * Also, we don't use btrfs_release_delayed_node() because that would
2193 * requeue the delayed inode (change its order in the list of prepared
2194 * nodes) and we don't want to do such change because we don't create or
2195 * delete delayed items.
2197 ASSERT(refcount_read(&node->refs) > 1);
2198 refcount_dec(&node->refs);