Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2011 Fujitsu. All rights reserved.
4 * Written by Miao Xie <miaox@cn.fujitsu.com>
5 */
6
7#include <linux/slab.h>
8#include <linux/iversion.h>
9#include "misc.h"
10#include "delayed-inode.h"
11#include "disk-io.h"
12#include "transaction.h"
13#include "ctree.h"
14#include "qgroup.h"
15
16#define BTRFS_DELAYED_WRITEBACK 512
17#define BTRFS_DELAYED_BACKGROUND 128
18#define BTRFS_DELAYED_BATCH 16
19
20static struct kmem_cache *delayed_node_cache;
21
22int __init btrfs_delayed_inode_init(void)
23{
24 delayed_node_cache = kmem_cache_create("btrfs_delayed_node",
25 sizeof(struct btrfs_delayed_node),
26 0,
27 SLAB_MEM_SPREAD,
28 NULL);
29 if (!delayed_node_cache)
30 return -ENOMEM;
31 return 0;
32}
33
34void __cold btrfs_delayed_inode_exit(void)
35{
36 kmem_cache_destroy(delayed_node_cache);
37}
38
39static inline void btrfs_init_delayed_node(
40 struct btrfs_delayed_node *delayed_node,
41 struct btrfs_root *root, u64 inode_id)
42{
43 delayed_node->root = root;
44 delayed_node->inode_id = inode_id;
45 refcount_set(&delayed_node->refs, 0);
46 delayed_node->ins_root = RB_ROOT_CACHED;
47 delayed_node->del_root = RB_ROOT_CACHED;
48 mutex_init(&delayed_node->mutex);
49 INIT_LIST_HEAD(&delayed_node->n_list);
50 INIT_LIST_HEAD(&delayed_node->p_list);
51}
52
53static inline int btrfs_is_continuous_delayed_item(
54 struct btrfs_delayed_item *item1,
55 struct btrfs_delayed_item *item2)
56{
57 if (item1->key.type == BTRFS_DIR_INDEX_KEY &&
58 item1->key.objectid == item2->key.objectid &&
59 item1->key.type == item2->key.type &&
60 item1->key.offset + 1 == item2->key.offset)
61 return 1;
62 return 0;
63}
64
65static struct btrfs_delayed_node *btrfs_get_delayed_node(
66 struct btrfs_inode *btrfs_inode)
67{
68 struct btrfs_root *root = btrfs_inode->root;
69 u64 ino = btrfs_ino(btrfs_inode);
70 struct btrfs_delayed_node *node;
71
72 node = READ_ONCE(btrfs_inode->delayed_node);
73 if (node) {
74 refcount_inc(&node->refs);
75 return node;
76 }
77
78 spin_lock(&root->inode_lock);
79 node = radix_tree_lookup(&root->delayed_nodes_tree, ino);
80
81 if (node) {
82 if (btrfs_inode->delayed_node) {
83 refcount_inc(&node->refs); /* can be accessed */
84 BUG_ON(btrfs_inode->delayed_node != node);
85 spin_unlock(&root->inode_lock);
86 return node;
87 }
88
89 /*
90 * It's possible that we're racing into the middle of removing
91 * this node from the radix tree. In this case, the refcount
92 * was zero and it should never go back to one. Just return
93 * NULL like it was never in the radix at all; our release
94 * function is in the process of removing it.
95 *
96 * Some implementations of refcount_inc refuse to bump the
97 * refcount once it has hit zero. If we don't do this dance
98 * here, refcount_inc() may decide to just WARN_ONCE() instead
99 * of actually bumping the refcount.
100 *
101 * If this node is properly in the radix, we want to bump the
102 * refcount twice, once for the inode and once for this get
103 * operation.
104 */
105 if (refcount_inc_not_zero(&node->refs)) {
106 refcount_inc(&node->refs);
107 btrfs_inode->delayed_node = node;
108 } else {
109 node = NULL;
110 }
111
112 spin_unlock(&root->inode_lock);
113 return node;
114 }
115 spin_unlock(&root->inode_lock);
116
117 return NULL;
118}
119
120/* Will return either the node or PTR_ERR(-ENOMEM) */
121static struct btrfs_delayed_node *btrfs_get_or_create_delayed_node(
122 struct btrfs_inode *btrfs_inode)
123{
124 struct btrfs_delayed_node *node;
125 struct btrfs_root *root = btrfs_inode->root;
126 u64 ino = btrfs_ino(btrfs_inode);
127 int ret;
128
129again:
130 node = btrfs_get_delayed_node(btrfs_inode);
131 if (node)
132 return node;
133
134 node = kmem_cache_zalloc(delayed_node_cache, GFP_NOFS);
135 if (!node)
136 return ERR_PTR(-ENOMEM);
137 btrfs_init_delayed_node(node, root, ino);
138
139 /* cached in the btrfs inode and can be accessed */
140 refcount_set(&node->refs, 2);
141
142 ret = radix_tree_preload(GFP_NOFS);
143 if (ret) {
144 kmem_cache_free(delayed_node_cache, node);
145 return ERR_PTR(ret);
146 }
147
148 spin_lock(&root->inode_lock);
149 ret = radix_tree_insert(&root->delayed_nodes_tree, ino, node);
150 if (ret == -EEXIST) {
151 spin_unlock(&root->inode_lock);
152 kmem_cache_free(delayed_node_cache, node);
153 radix_tree_preload_end();
154 goto again;
155 }
156 btrfs_inode->delayed_node = node;
157 spin_unlock(&root->inode_lock);
158 radix_tree_preload_end();
159
160 return node;
161}
162
163/*
164 * Call it when holding delayed_node->mutex
165 *
166 * If mod = 1, add this node into the prepared list.
167 */
168static void btrfs_queue_delayed_node(struct btrfs_delayed_root *root,
169 struct btrfs_delayed_node *node,
170 int mod)
171{
172 spin_lock(&root->lock);
173 if (test_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags)) {
174 if (!list_empty(&node->p_list))
175 list_move_tail(&node->p_list, &root->prepare_list);
176 else if (mod)
177 list_add_tail(&node->p_list, &root->prepare_list);
178 } else {
179 list_add_tail(&node->n_list, &root->node_list);
180 list_add_tail(&node->p_list, &root->prepare_list);
181 refcount_inc(&node->refs); /* inserted into list */
182 root->nodes++;
183 set_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags);
184 }
185 spin_unlock(&root->lock);
186}
187
188/* Call it when holding delayed_node->mutex */
189static void btrfs_dequeue_delayed_node(struct btrfs_delayed_root *root,
190 struct btrfs_delayed_node *node)
191{
192 spin_lock(&root->lock);
193 if (test_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags)) {
194 root->nodes--;
195 refcount_dec(&node->refs); /* not in the list */
196 list_del_init(&node->n_list);
197 if (!list_empty(&node->p_list))
198 list_del_init(&node->p_list);
199 clear_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags);
200 }
201 spin_unlock(&root->lock);
202}
203
204static struct btrfs_delayed_node *btrfs_first_delayed_node(
205 struct btrfs_delayed_root *delayed_root)
206{
207 struct list_head *p;
208 struct btrfs_delayed_node *node = NULL;
209
210 spin_lock(&delayed_root->lock);
211 if (list_empty(&delayed_root->node_list))
212 goto out;
213
214 p = delayed_root->node_list.next;
215 node = list_entry(p, struct btrfs_delayed_node, n_list);
216 refcount_inc(&node->refs);
217out:
218 spin_unlock(&delayed_root->lock);
219
220 return node;
221}
222
223static struct btrfs_delayed_node *btrfs_next_delayed_node(
224 struct btrfs_delayed_node *node)
225{
226 struct btrfs_delayed_root *delayed_root;
227 struct list_head *p;
228 struct btrfs_delayed_node *next = NULL;
229
230 delayed_root = node->root->fs_info->delayed_root;
231 spin_lock(&delayed_root->lock);
232 if (!test_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags)) {
233 /* not in the list */
234 if (list_empty(&delayed_root->node_list))
235 goto out;
236 p = delayed_root->node_list.next;
237 } else if (list_is_last(&node->n_list, &delayed_root->node_list))
238 goto out;
239 else
240 p = node->n_list.next;
241
242 next = list_entry(p, struct btrfs_delayed_node, n_list);
243 refcount_inc(&next->refs);
244out:
245 spin_unlock(&delayed_root->lock);
246
247 return next;
248}
249
250static void __btrfs_release_delayed_node(
251 struct btrfs_delayed_node *delayed_node,
252 int mod)
253{
254 struct btrfs_delayed_root *delayed_root;
255
256 if (!delayed_node)
257 return;
258
259 delayed_root = delayed_node->root->fs_info->delayed_root;
260
261 mutex_lock(&delayed_node->mutex);
262 if (delayed_node->count)
263 btrfs_queue_delayed_node(delayed_root, delayed_node, mod);
264 else
265 btrfs_dequeue_delayed_node(delayed_root, delayed_node);
266 mutex_unlock(&delayed_node->mutex);
267
268 if (refcount_dec_and_test(&delayed_node->refs)) {
269 struct btrfs_root *root = delayed_node->root;
270
271 spin_lock(&root->inode_lock);
272 /*
273 * Once our refcount goes to zero, nobody is allowed to bump it
274 * back up. We can delete it now.
275 */
276 ASSERT(refcount_read(&delayed_node->refs) == 0);
277 radix_tree_delete(&root->delayed_nodes_tree,
278 delayed_node->inode_id);
279 spin_unlock(&root->inode_lock);
280 kmem_cache_free(delayed_node_cache, delayed_node);
281 }
282}
283
284static inline void btrfs_release_delayed_node(struct btrfs_delayed_node *node)
285{
286 __btrfs_release_delayed_node(node, 0);
287}
288
289static struct btrfs_delayed_node *btrfs_first_prepared_delayed_node(
290 struct btrfs_delayed_root *delayed_root)
291{
292 struct list_head *p;
293 struct btrfs_delayed_node *node = NULL;
294
295 spin_lock(&delayed_root->lock);
296 if (list_empty(&delayed_root->prepare_list))
297 goto out;
298
299 p = delayed_root->prepare_list.next;
300 list_del_init(p);
301 node = list_entry(p, struct btrfs_delayed_node, p_list);
302 refcount_inc(&node->refs);
303out:
304 spin_unlock(&delayed_root->lock);
305
306 return node;
307}
308
309static inline void btrfs_release_prepared_delayed_node(
310 struct btrfs_delayed_node *node)
311{
312 __btrfs_release_delayed_node(node, 1);
313}
314
315static struct btrfs_delayed_item *btrfs_alloc_delayed_item(u32 data_len)
316{
317 struct btrfs_delayed_item *item;
318 item = kmalloc(sizeof(*item) + data_len, GFP_NOFS);
319 if (item) {
320 item->data_len = data_len;
321 item->ins_or_del = 0;
322 item->bytes_reserved = 0;
323 item->delayed_node = NULL;
324 refcount_set(&item->refs, 1);
325 }
326 return item;
327}
328
329/*
330 * __btrfs_lookup_delayed_item - look up the delayed item by key
331 * @delayed_node: pointer to the delayed node
332 * @key: the key to look up
333 * @prev: used to store the prev item if the right item isn't found
334 * @next: used to store the next item if the right item isn't found
335 *
336 * Note: if we don't find the right item, we will return the prev item and
337 * the next item.
338 */
339static struct btrfs_delayed_item *__btrfs_lookup_delayed_item(
340 struct rb_root *root,
341 struct btrfs_key *key,
342 struct btrfs_delayed_item **prev,
343 struct btrfs_delayed_item **next)
344{
345 struct rb_node *node, *prev_node = NULL;
346 struct btrfs_delayed_item *delayed_item = NULL;
347 int ret = 0;
348
349 node = root->rb_node;
350
351 while (node) {
352 delayed_item = rb_entry(node, struct btrfs_delayed_item,
353 rb_node);
354 prev_node = node;
355 ret = btrfs_comp_cpu_keys(&delayed_item->key, key);
356 if (ret < 0)
357 node = node->rb_right;
358 else if (ret > 0)
359 node = node->rb_left;
360 else
361 return delayed_item;
362 }
363
364 if (prev) {
365 if (!prev_node)
366 *prev = NULL;
367 else if (ret < 0)
368 *prev = delayed_item;
369 else if ((node = rb_prev(prev_node)) != NULL) {
370 *prev = rb_entry(node, struct btrfs_delayed_item,
371 rb_node);
372 } else
373 *prev = NULL;
374 }
375
376 if (next) {
377 if (!prev_node)
378 *next = NULL;
379 else if (ret > 0)
380 *next = delayed_item;
381 else if ((node = rb_next(prev_node)) != NULL) {
382 *next = rb_entry(node, struct btrfs_delayed_item,
383 rb_node);
384 } else
385 *next = NULL;
386 }
387 return NULL;
388}
389
390static struct btrfs_delayed_item *__btrfs_lookup_delayed_insertion_item(
391 struct btrfs_delayed_node *delayed_node,
392 struct btrfs_key *key)
393{
394 return __btrfs_lookup_delayed_item(&delayed_node->ins_root.rb_root, key,
395 NULL, NULL);
396}
397
398static int __btrfs_add_delayed_item(struct btrfs_delayed_node *delayed_node,
399 struct btrfs_delayed_item *ins,
400 int action)
401{
402 struct rb_node **p, *node;
403 struct rb_node *parent_node = NULL;
404 struct rb_root_cached *root;
405 struct btrfs_delayed_item *item;
406 int cmp;
407 bool leftmost = true;
408
409 if (action == BTRFS_DELAYED_INSERTION_ITEM)
410 root = &delayed_node->ins_root;
411 else if (action == BTRFS_DELAYED_DELETION_ITEM)
412 root = &delayed_node->del_root;
413 else
414 BUG();
415 p = &root->rb_root.rb_node;
416 node = &ins->rb_node;
417
418 while (*p) {
419 parent_node = *p;
420 item = rb_entry(parent_node, struct btrfs_delayed_item,
421 rb_node);
422
423 cmp = btrfs_comp_cpu_keys(&item->key, &ins->key);
424 if (cmp < 0) {
425 p = &(*p)->rb_right;
426 leftmost = false;
427 } else if (cmp > 0) {
428 p = &(*p)->rb_left;
429 } else {
430 return -EEXIST;
431 }
432 }
433
434 rb_link_node(node, parent_node, p);
435 rb_insert_color_cached(node, root, leftmost);
436 ins->delayed_node = delayed_node;
437 ins->ins_or_del = action;
438
439 if (ins->key.type == BTRFS_DIR_INDEX_KEY &&
440 action == BTRFS_DELAYED_INSERTION_ITEM &&
441 ins->key.offset >= delayed_node->index_cnt)
442 delayed_node->index_cnt = ins->key.offset + 1;
443
444 delayed_node->count++;
445 atomic_inc(&delayed_node->root->fs_info->delayed_root->items);
446 return 0;
447}
448
449static int __btrfs_add_delayed_insertion_item(struct btrfs_delayed_node *node,
450 struct btrfs_delayed_item *item)
451{
452 return __btrfs_add_delayed_item(node, item,
453 BTRFS_DELAYED_INSERTION_ITEM);
454}
455
456static int __btrfs_add_delayed_deletion_item(struct btrfs_delayed_node *node,
457 struct btrfs_delayed_item *item)
458{
459 return __btrfs_add_delayed_item(node, item,
460 BTRFS_DELAYED_DELETION_ITEM);
461}
462
463static void finish_one_item(struct btrfs_delayed_root *delayed_root)
464{
465 int seq = atomic_inc_return(&delayed_root->items_seq);
466
467 /* atomic_dec_return implies a barrier */
468 if ((atomic_dec_return(&delayed_root->items) <
469 BTRFS_DELAYED_BACKGROUND || seq % BTRFS_DELAYED_BATCH == 0))
470 cond_wake_up_nomb(&delayed_root->wait);
471}
472
473static void __btrfs_remove_delayed_item(struct btrfs_delayed_item *delayed_item)
474{
475 struct rb_root_cached *root;
476 struct btrfs_delayed_root *delayed_root;
477
478 /* Not associated with any delayed_node */
479 if (!delayed_item->delayed_node)
480 return;
481 delayed_root = delayed_item->delayed_node->root->fs_info->delayed_root;
482
483 BUG_ON(!delayed_root);
484 BUG_ON(delayed_item->ins_or_del != BTRFS_DELAYED_DELETION_ITEM &&
485 delayed_item->ins_or_del != BTRFS_DELAYED_INSERTION_ITEM);
486
487 if (delayed_item->ins_or_del == BTRFS_DELAYED_INSERTION_ITEM)
488 root = &delayed_item->delayed_node->ins_root;
489 else
490 root = &delayed_item->delayed_node->del_root;
491
492 rb_erase_cached(&delayed_item->rb_node, root);
493 delayed_item->delayed_node->count--;
494
495 finish_one_item(delayed_root);
496}
497
498static void btrfs_release_delayed_item(struct btrfs_delayed_item *item)
499{
500 if (item) {
501 __btrfs_remove_delayed_item(item);
502 if (refcount_dec_and_test(&item->refs))
503 kfree(item);
504 }
505}
506
507static struct btrfs_delayed_item *__btrfs_first_delayed_insertion_item(
508 struct btrfs_delayed_node *delayed_node)
509{
510 struct rb_node *p;
511 struct btrfs_delayed_item *item = NULL;
512
513 p = rb_first_cached(&delayed_node->ins_root);
514 if (p)
515 item = rb_entry(p, struct btrfs_delayed_item, rb_node);
516
517 return item;
518}
519
520static struct btrfs_delayed_item *__btrfs_first_delayed_deletion_item(
521 struct btrfs_delayed_node *delayed_node)
522{
523 struct rb_node *p;
524 struct btrfs_delayed_item *item = NULL;
525
526 p = rb_first_cached(&delayed_node->del_root);
527 if (p)
528 item = rb_entry(p, struct btrfs_delayed_item, rb_node);
529
530 return item;
531}
532
533static struct btrfs_delayed_item *__btrfs_next_delayed_item(
534 struct btrfs_delayed_item *item)
535{
536 struct rb_node *p;
537 struct btrfs_delayed_item *next = NULL;
538
539 p = rb_next(&item->rb_node);
540 if (p)
541 next = rb_entry(p, struct btrfs_delayed_item, rb_node);
542
543 return next;
544}
545
546static int btrfs_delayed_item_reserve_metadata(struct btrfs_trans_handle *trans,
547 struct btrfs_root *root,
548 struct btrfs_delayed_item *item)
549{
550 struct btrfs_block_rsv *src_rsv;
551 struct btrfs_block_rsv *dst_rsv;
552 struct btrfs_fs_info *fs_info = root->fs_info;
553 u64 num_bytes;
554 int ret;
555
556 if (!trans->bytes_reserved)
557 return 0;
558
559 src_rsv = trans->block_rsv;
560 dst_rsv = &fs_info->delayed_block_rsv;
561
562 num_bytes = btrfs_calc_insert_metadata_size(fs_info, 1);
563
564 /*
565 * Here we migrate space rsv from transaction rsv, since have already
566 * reserved space when starting a transaction. So no need to reserve
567 * qgroup space here.
568 */
569 ret = btrfs_block_rsv_migrate(src_rsv, dst_rsv, num_bytes, true);
570 if (!ret) {
571 trace_btrfs_space_reservation(fs_info, "delayed_item",
572 item->key.objectid,
573 num_bytes, 1);
574 item->bytes_reserved = num_bytes;
575 }
576
577 return ret;
578}
579
580static void btrfs_delayed_item_release_metadata(struct btrfs_root *root,
581 struct btrfs_delayed_item *item)
582{
583 struct btrfs_block_rsv *rsv;
584 struct btrfs_fs_info *fs_info = root->fs_info;
585
586 if (!item->bytes_reserved)
587 return;
588
589 rsv = &fs_info->delayed_block_rsv;
590 /*
591 * Check btrfs_delayed_item_reserve_metadata() to see why we don't need
592 * to release/reserve qgroup space.
593 */
594 trace_btrfs_space_reservation(fs_info, "delayed_item",
595 item->key.objectid, item->bytes_reserved,
596 0);
597 btrfs_block_rsv_release(fs_info, rsv,
598 item->bytes_reserved);
599}
600
601static int btrfs_delayed_inode_reserve_metadata(
602 struct btrfs_trans_handle *trans,
603 struct btrfs_root *root,
604 struct btrfs_inode *inode,
605 struct btrfs_delayed_node *node)
606{
607 struct btrfs_fs_info *fs_info = root->fs_info;
608 struct btrfs_block_rsv *src_rsv;
609 struct btrfs_block_rsv *dst_rsv;
610 u64 num_bytes;
611 int ret;
612
613 src_rsv = trans->block_rsv;
614 dst_rsv = &fs_info->delayed_block_rsv;
615
616 num_bytes = btrfs_calc_metadata_size(fs_info, 1);
617
618 /*
619 * btrfs_dirty_inode will update the inode under btrfs_join_transaction
620 * which doesn't reserve space for speed. This is a problem since we
621 * still need to reserve space for this update, so try to reserve the
622 * space.
623 *
624 * Now if src_rsv == delalloc_block_rsv we'll let it just steal since
625 * we always reserve enough to update the inode item.
626 */
627 if (!src_rsv || (!trans->bytes_reserved &&
628 src_rsv->type != BTRFS_BLOCK_RSV_DELALLOC)) {
629 ret = btrfs_qgroup_reserve_meta_prealloc(root,
630 fs_info->nodesize, true);
631 if (ret < 0)
632 return ret;
633 ret = btrfs_block_rsv_add(root, dst_rsv, num_bytes,
634 BTRFS_RESERVE_NO_FLUSH);
635 /*
636 * Since we're under a transaction reserve_metadata_bytes could
637 * try to commit the transaction which will make it return
638 * EAGAIN to make us stop the transaction we have, so return
639 * ENOSPC instead so that btrfs_dirty_inode knows what to do.
640 */
641 if (ret == -EAGAIN) {
642 ret = -ENOSPC;
643 btrfs_qgroup_free_meta_prealloc(root, num_bytes);
644 }
645 if (!ret) {
646 node->bytes_reserved = num_bytes;
647 trace_btrfs_space_reservation(fs_info,
648 "delayed_inode",
649 btrfs_ino(inode),
650 num_bytes, 1);
651 } else {
652 btrfs_qgroup_free_meta_prealloc(root, fs_info->nodesize);
653 }
654 return ret;
655 }
656
657 ret = btrfs_block_rsv_migrate(src_rsv, dst_rsv, num_bytes, true);
658 if (!ret) {
659 trace_btrfs_space_reservation(fs_info, "delayed_inode",
660 btrfs_ino(inode), num_bytes, 1);
661 node->bytes_reserved = num_bytes;
662 }
663
664 return ret;
665}
666
667static void btrfs_delayed_inode_release_metadata(struct btrfs_fs_info *fs_info,
668 struct btrfs_delayed_node *node,
669 bool qgroup_free)
670{
671 struct btrfs_block_rsv *rsv;
672
673 if (!node->bytes_reserved)
674 return;
675
676 rsv = &fs_info->delayed_block_rsv;
677 trace_btrfs_space_reservation(fs_info, "delayed_inode",
678 node->inode_id, node->bytes_reserved, 0);
679 btrfs_block_rsv_release(fs_info, rsv,
680 node->bytes_reserved);
681 if (qgroup_free)
682 btrfs_qgroup_free_meta_prealloc(node->root,
683 node->bytes_reserved);
684 else
685 btrfs_qgroup_convert_reserved_meta(node->root,
686 node->bytes_reserved);
687 node->bytes_reserved = 0;
688}
689
690/*
691 * This helper will insert some continuous items into the same leaf according
692 * to the free space of the leaf.
693 */
694static int btrfs_batch_insert_items(struct btrfs_root *root,
695 struct btrfs_path *path,
696 struct btrfs_delayed_item *item)
697{
698 struct btrfs_delayed_item *curr, *next;
699 int free_space;
700 int total_data_size = 0, total_size = 0;
701 struct extent_buffer *leaf;
702 char *data_ptr;
703 struct btrfs_key *keys;
704 u32 *data_size;
705 struct list_head head;
706 int slot;
707 int nitems;
708 int i;
709 int ret = 0;
710
711 BUG_ON(!path->nodes[0]);
712
713 leaf = path->nodes[0];
714 free_space = btrfs_leaf_free_space(leaf);
715 INIT_LIST_HEAD(&head);
716
717 next = item;
718 nitems = 0;
719
720 /*
721 * count the number of the continuous items that we can insert in batch
722 */
723 while (total_size + next->data_len + sizeof(struct btrfs_item) <=
724 free_space) {
725 total_data_size += next->data_len;
726 total_size += next->data_len + sizeof(struct btrfs_item);
727 list_add_tail(&next->tree_list, &head);
728 nitems++;
729
730 curr = next;
731 next = __btrfs_next_delayed_item(curr);
732 if (!next)
733 break;
734
735 if (!btrfs_is_continuous_delayed_item(curr, next))
736 break;
737 }
738
739 if (!nitems) {
740 ret = 0;
741 goto out;
742 }
743
744 /*
745 * we need allocate some memory space, but it might cause the task
746 * to sleep, so we set all locked nodes in the path to blocking locks
747 * first.
748 */
749 btrfs_set_path_blocking(path);
750
751 keys = kmalloc_array(nitems, sizeof(struct btrfs_key), GFP_NOFS);
752 if (!keys) {
753 ret = -ENOMEM;
754 goto out;
755 }
756
757 data_size = kmalloc_array(nitems, sizeof(u32), GFP_NOFS);
758 if (!data_size) {
759 ret = -ENOMEM;
760 goto error;
761 }
762
763 /* get keys of all the delayed items */
764 i = 0;
765 list_for_each_entry(next, &head, tree_list) {
766 keys[i] = next->key;
767 data_size[i] = next->data_len;
768 i++;
769 }
770
771 /* insert the keys of the items */
772 setup_items_for_insert(root, path, keys, data_size,
773 total_data_size, total_size, nitems);
774
775 /* insert the dir index items */
776 slot = path->slots[0];
777 list_for_each_entry_safe(curr, next, &head, tree_list) {
778 data_ptr = btrfs_item_ptr(leaf, slot, char);
779 write_extent_buffer(leaf, &curr->data,
780 (unsigned long)data_ptr,
781 curr->data_len);
782 slot++;
783
784 btrfs_delayed_item_release_metadata(root, curr);
785
786 list_del(&curr->tree_list);
787 btrfs_release_delayed_item(curr);
788 }
789
790error:
791 kfree(data_size);
792 kfree(keys);
793out:
794 return ret;
795}
796
797/*
798 * This helper can just do simple insertion that needn't extend item for new
799 * data, such as directory name index insertion, inode insertion.
800 */
801static int btrfs_insert_delayed_item(struct btrfs_trans_handle *trans,
802 struct btrfs_root *root,
803 struct btrfs_path *path,
804 struct btrfs_delayed_item *delayed_item)
805{
806 struct extent_buffer *leaf;
807 char *ptr;
808 int ret;
809
810 ret = btrfs_insert_empty_item(trans, root, path, &delayed_item->key,
811 delayed_item->data_len);
812 if (ret < 0 && ret != -EEXIST)
813 return ret;
814
815 leaf = path->nodes[0];
816
817 ptr = btrfs_item_ptr(leaf, path->slots[0], char);
818
819 write_extent_buffer(leaf, delayed_item->data, (unsigned long)ptr,
820 delayed_item->data_len);
821 btrfs_mark_buffer_dirty(leaf);
822
823 btrfs_delayed_item_release_metadata(root, delayed_item);
824 return 0;
825}
826
827/*
828 * we insert an item first, then if there are some continuous items, we try
829 * to insert those items into the same leaf.
830 */
831static int btrfs_insert_delayed_items(struct btrfs_trans_handle *trans,
832 struct btrfs_path *path,
833 struct btrfs_root *root,
834 struct btrfs_delayed_node *node)
835{
836 struct btrfs_delayed_item *curr, *prev;
837 int ret = 0;
838
839do_again:
840 mutex_lock(&node->mutex);
841 curr = __btrfs_first_delayed_insertion_item(node);
842 if (!curr)
843 goto insert_end;
844
845 ret = btrfs_insert_delayed_item(trans, root, path, curr);
846 if (ret < 0) {
847 btrfs_release_path(path);
848 goto insert_end;
849 }
850
851 prev = curr;
852 curr = __btrfs_next_delayed_item(prev);
853 if (curr && btrfs_is_continuous_delayed_item(prev, curr)) {
854 /* insert the continuous items into the same leaf */
855 path->slots[0]++;
856 btrfs_batch_insert_items(root, path, curr);
857 }
858 btrfs_release_delayed_item(prev);
859 btrfs_mark_buffer_dirty(path->nodes[0]);
860
861 btrfs_release_path(path);
862 mutex_unlock(&node->mutex);
863 goto do_again;
864
865insert_end:
866 mutex_unlock(&node->mutex);
867 return ret;
868}
869
870static int btrfs_batch_delete_items(struct btrfs_trans_handle *trans,
871 struct btrfs_root *root,
872 struct btrfs_path *path,
873 struct btrfs_delayed_item *item)
874{
875 struct btrfs_delayed_item *curr, *next;
876 struct extent_buffer *leaf;
877 struct btrfs_key key;
878 struct list_head head;
879 int nitems, i, last_item;
880 int ret = 0;
881
882 BUG_ON(!path->nodes[0]);
883
884 leaf = path->nodes[0];
885
886 i = path->slots[0];
887 last_item = btrfs_header_nritems(leaf) - 1;
888 if (i > last_item)
889 return -ENOENT; /* FIXME: Is errno suitable? */
890
891 next = item;
892 INIT_LIST_HEAD(&head);
893 btrfs_item_key_to_cpu(leaf, &key, i);
894 nitems = 0;
895 /*
896 * count the number of the dir index items that we can delete in batch
897 */
898 while (btrfs_comp_cpu_keys(&next->key, &key) == 0) {
899 list_add_tail(&next->tree_list, &head);
900 nitems++;
901
902 curr = next;
903 next = __btrfs_next_delayed_item(curr);
904 if (!next)
905 break;
906
907 if (!btrfs_is_continuous_delayed_item(curr, next))
908 break;
909
910 i++;
911 if (i > last_item)
912 break;
913 btrfs_item_key_to_cpu(leaf, &key, i);
914 }
915
916 if (!nitems)
917 return 0;
918
919 ret = btrfs_del_items(trans, root, path, path->slots[0], nitems);
920 if (ret)
921 goto out;
922
923 list_for_each_entry_safe(curr, next, &head, tree_list) {
924 btrfs_delayed_item_release_metadata(root, curr);
925 list_del(&curr->tree_list);
926 btrfs_release_delayed_item(curr);
927 }
928
929out:
930 return ret;
931}
932
933static int btrfs_delete_delayed_items(struct btrfs_trans_handle *trans,
934 struct btrfs_path *path,
935 struct btrfs_root *root,
936 struct btrfs_delayed_node *node)
937{
938 struct btrfs_delayed_item *curr, *prev;
939 int ret = 0;
940
941do_again:
942 mutex_lock(&node->mutex);
943 curr = __btrfs_first_delayed_deletion_item(node);
944 if (!curr)
945 goto delete_fail;
946
947 ret = btrfs_search_slot(trans, root, &curr->key, path, -1, 1);
948 if (ret < 0)
949 goto delete_fail;
950 else if (ret > 0) {
951 /*
952 * can't find the item which the node points to, so this node
953 * is invalid, just drop it.
954 */
955 prev = curr;
956 curr = __btrfs_next_delayed_item(prev);
957 btrfs_release_delayed_item(prev);
958 ret = 0;
959 btrfs_release_path(path);
960 if (curr) {
961 mutex_unlock(&node->mutex);
962 goto do_again;
963 } else
964 goto delete_fail;
965 }
966
967 btrfs_batch_delete_items(trans, root, path, curr);
968 btrfs_release_path(path);
969 mutex_unlock(&node->mutex);
970 goto do_again;
971
972delete_fail:
973 btrfs_release_path(path);
974 mutex_unlock(&node->mutex);
975 return ret;
976}
977
978static void btrfs_release_delayed_inode(struct btrfs_delayed_node *delayed_node)
979{
980 struct btrfs_delayed_root *delayed_root;
981
982 if (delayed_node &&
983 test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
984 BUG_ON(!delayed_node->root);
985 clear_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags);
986 delayed_node->count--;
987
988 delayed_root = delayed_node->root->fs_info->delayed_root;
989 finish_one_item(delayed_root);
990 }
991}
992
993static void btrfs_release_delayed_iref(struct btrfs_delayed_node *delayed_node)
994{
995 struct btrfs_delayed_root *delayed_root;
996
997 ASSERT(delayed_node->root);
998 clear_bit(BTRFS_DELAYED_NODE_DEL_IREF, &delayed_node->flags);
999 delayed_node->count--;
1000
1001 delayed_root = delayed_node->root->fs_info->delayed_root;
1002 finish_one_item(delayed_root);
1003}
1004
1005static int __btrfs_update_delayed_inode(struct btrfs_trans_handle *trans,
1006 struct btrfs_root *root,
1007 struct btrfs_path *path,
1008 struct btrfs_delayed_node *node)
1009{
1010 struct btrfs_fs_info *fs_info = root->fs_info;
1011 struct btrfs_key key;
1012 struct btrfs_inode_item *inode_item;
1013 struct extent_buffer *leaf;
1014 int mod;
1015 int ret;
1016
1017 key.objectid = node->inode_id;
1018 key.type = BTRFS_INODE_ITEM_KEY;
1019 key.offset = 0;
1020
1021 if (test_bit(BTRFS_DELAYED_NODE_DEL_IREF, &node->flags))
1022 mod = -1;
1023 else
1024 mod = 1;
1025
1026 ret = btrfs_lookup_inode(trans, root, path, &key, mod);
1027 if (ret > 0) {
1028 btrfs_release_path(path);
1029 return -ENOENT;
1030 } else if (ret < 0) {
1031 return ret;
1032 }
1033
1034 leaf = path->nodes[0];
1035 inode_item = btrfs_item_ptr(leaf, path->slots[0],
1036 struct btrfs_inode_item);
1037 write_extent_buffer(leaf, &node->inode_item, (unsigned long)inode_item,
1038 sizeof(struct btrfs_inode_item));
1039 btrfs_mark_buffer_dirty(leaf);
1040
1041 if (!test_bit(BTRFS_DELAYED_NODE_DEL_IREF, &node->flags))
1042 goto no_iref;
1043
1044 path->slots[0]++;
1045 if (path->slots[0] >= btrfs_header_nritems(leaf))
1046 goto search;
1047again:
1048 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1049 if (key.objectid != node->inode_id)
1050 goto out;
1051
1052 if (key.type != BTRFS_INODE_REF_KEY &&
1053 key.type != BTRFS_INODE_EXTREF_KEY)
1054 goto out;
1055
1056 /*
1057 * Delayed iref deletion is for the inode who has only one link,
1058 * so there is only one iref. The case that several irefs are
1059 * in the same item doesn't exist.
1060 */
1061 btrfs_del_item(trans, root, path);
1062out:
1063 btrfs_release_delayed_iref(node);
1064no_iref:
1065 btrfs_release_path(path);
1066err_out:
1067 btrfs_delayed_inode_release_metadata(fs_info, node, (ret < 0));
1068 btrfs_release_delayed_inode(node);
1069
1070 return ret;
1071
1072search:
1073 btrfs_release_path(path);
1074
1075 key.type = BTRFS_INODE_EXTREF_KEY;
1076 key.offset = -1;
1077 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1078 if (ret < 0)
1079 goto err_out;
1080 ASSERT(ret);
1081
1082 ret = 0;
1083 leaf = path->nodes[0];
1084 path->slots[0]--;
1085 goto again;
1086}
1087
1088static inline int btrfs_update_delayed_inode(struct btrfs_trans_handle *trans,
1089 struct btrfs_root *root,
1090 struct btrfs_path *path,
1091 struct btrfs_delayed_node *node)
1092{
1093 int ret;
1094
1095 mutex_lock(&node->mutex);
1096 if (!test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &node->flags)) {
1097 mutex_unlock(&node->mutex);
1098 return 0;
1099 }
1100
1101 ret = __btrfs_update_delayed_inode(trans, root, path, node);
1102 mutex_unlock(&node->mutex);
1103 return ret;
1104}
1105
1106static inline int
1107__btrfs_commit_inode_delayed_items(struct btrfs_trans_handle *trans,
1108 struct btrfs_path *path,
1109 struct btrfs_delayed_node *node)
1110{
1111 int ret;
1112
1113 ret = btrfs_insert_delayed_items(trans, path, node->root, node);
1114 if (ret)
1115 return ret;
1116
1117 ret = btrfs_delete_delayed_items(trans, path, node->root, node);
1118 if (ret)
1119 return ret;
1120
1121 ret = btrfs_update_delayed_inode(trans, node->root, path, node);
1122 return ret;
1123}
1124
1125/*
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.
1130 */
1131static int __btrfs_run_delayed_items(struct btrfs_trans_handle *trans, int nr)
1132{
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;
1138 int ret = 0;
1139 bool count = (nr > 0);
1140
1141 if (trans->aborted)
1142 return -EIO;
1143
1144 path = btrfs_alloc_path();
1145 if (!path)
1146 return -ENOMEM;
1147 path->leave_spinning = 1;
1148
1149 block_rsv = trans->block_rsv;
1150 trans->block_rsv = &fs_info->delayed_block_rsv;
1151
1152 delayed_root = fs_info->delayed_root;
1153
1154 curr_node = btrfs_first_delayed_node(delayed_root);
1155 while (curr_node && (!count || (count && nr--))) {
1156 ret = __btrfs_commit_inode_delayed_items(trans, path,
1157 curr_node);
1158 if (ret) {
1159 btrfs_release_delayed_node(curr_node);
1160 curr_node = NULL;
1161 btrfs_abort_transaction(trans, ret);
1162 break;
1163 }
1164
1165 prev_node = curr_node;
1166 curr_node = btrfs_next_delayed_node(curr_node);
1167 btrfs_release_delayed_node(prev_node);
1168 }
1169
1170 if (curr_node)
1171 btrfs_release_delayed_node(curr_node);
1172 btrfs_free_path(path);
1173 trans->block_rsv = block_rsv;
1174
1175 return ret;
1176}
1177
1178int btrfs_run_delayed_items(struct btrfs_trans_handle *trans)
1179{
1180 return __btrfs_run_delayed_items(trans, -1);
1181}
1182
1183int btrfs_run_delayed_items_nr(struct btrfs_trans_handle *trans, int nr)
1184{
1185 return __btrfs_run_delayed_items(trans, nr);
1186}
1187
1188int btrfs_commit_inode_delayed_items(struct btrfs_trans_handle *trans,
1189 struct btrfs_inode *inode)
1190{
1191 struct btrfs_delayed_node *delayed_node = btrfs_get_delayed_node(inode);
1192 struct btrfs_path *path;
1193 struct btrfs_block_rsv *block_rsv;
1194 int ret;
1195
1196 if (!delayed_node)
1197 return 0;
1198
1199 mutex_lock(&delayed_node->mutex);
1200 if (!delayed_node->count) {
1201 mutex_unlock(&delayed_node->mutex);
1202 btrfs_release_delayed_node(delayed_node);
1203 return 0;
1204 }
1205 mutex_unlock(&delayed_node->mutex);
1206
1207 path = btrfs_alloc_path();
1208 if (!path) {
1209 btrfs_release_delayed_node(delayed_node);
1210 return -ENOMEM;
1211 }
1212 path->leave_spinning = 1;
1213
1214 block_rsv = trans->block_rsv;
1215 trans->block_rsv = &delayed_node->root->fs_info->delayed_block_rsv;
1216
1217 ret = __btrfs_commit_inode_delayed_items(trans, path, delayed_node);
1218
1219 btrfs_release_delayed_node(delayed_node);
1220 btrfs_free_path(path);
1221 trans->block_rsv = block_rsv;
1222
1223 return ret;
1224}
1225
1226int btrfs_commit_inode_delayed_inode(struct btrfs_inode *inode)
1227{
1228 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1229 struct btrfs_trans_handle *trans;
1230 struct btrfs_delayed_node *delayed_node = btrfs_get_delayed_node(inode);
1231 struct btrfs_path *path;
1232 struct btrfs_block_rsv *block_rsv;
1233 int ret;
1234
1235 if (!delayed_node)
1236 return 0;
1237
1238 mutex_lock(&delayed_node->mutex);
1239 if (!test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
1240 mutex_unlock(&delayed_node->mutex);
1241 btrfs_release_delayed_node(delayed_node);
1242 return 0;
1243 }
1244 mutex_unlock(&delayed_node->mutex);
1245
1246 trans = btrfs_join_transaction(delayed_node->root);
1247 if (IS_ERR(trans)) {
1248 ret = PTR_ERR(trans);
1249 goto out;
1250 }
1251
1252 path = btrfs_alloc_path();
1253 if (!path) {
1254 ret = -ENOMEM;
1255 goto trans_out;
1256 }
1257 path->leave_spinning = 1;
1258
1259 block_rsv = trans->block_rsv;
1260 trans->block_rsv = &fs_info->delayed_block_rsv;
1261
1262 mutex_lock(&delayed_node->mutex);
1263 if (test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags))
1264 ret = __btrfs_update_delayed_inode(trans, delayed_node->root,
1265 path, delayed_node);
1266 else
1267 ret = 0;
1268 mutex_unlock(&delayed_node->mutex);
1269
1270 btrfs_free_path(path);
1271 trans->block_rsv = block_rsv;
1272trans_out:
1273 btrfs_end_transaction(trans);
1274 btrfs_btree_balance_dirty(fs_info);
1275out:
1276 btrfs_release_delayed_node(delayed_node);
1277
1278 return ret;
1279}
1280
1281void btrfs_remove_delayed_node(struct btrfs_inode *inode)
1282{
1283 struct btrfs_delayed_node *delayed_node;
1284
1285 delayed_node = READ_ONCE(inode->delayed_node);
1286 if (!delayed_node)
1287 return;
1288
1289 inode->delayed_node = NULL;
1290 btrfs_release_delayed_node(delayed_node);
1291}
1292
1293struct btrfs_async_delayed_work {
1294 struct btrfs_delayed_root *delayed_root;
1295 int nr;
1296 struct btrfs_work work;
1297};
1298
1299static void btrfs_async_run_delayed_root(struct btrfs_work *work)
1300{
1301 struct btrfs_async_delayed_work *async_work;
1302 struct btrfs_delayed_root *delayed_root;
1303 struct btrfs_trans_handle *trans;
1304 struct btrfs_path *path;
1305 struct btrfs_delayed_node *delayed_node = NULL;
1306 struct btrfs_root *root;
1307 struct btrfs_block_rsv *block_rsv;
1308 int total_done = 0;
1309
1310 async_work = container_of(work, struct btrfs_async_delayed_work, work);
1311 delayed_root = async_work->delayed_root;
1312
1313 path = btrfs_alloc_path();
1314 if (!path)
1315 goto out;
1316
1317 do {
1318 if (atomic_read(&delayed_root->items) <
1319 BTRFS_DELAYED_BACKGROUND / 2)
1320 break;
1321
1322 delayed_node = btrfs_first_prepared_delayed_node(delayed_root);
1323 if (!delayed_node)
1324 break;
1325
1326 path->leave_spinning = 1;
1327 root = delayed_node->root;
1328
1329 trans = btrfs_join_transaction(root);
1330 if (IS_ERR(trans)) {
1331 btrfs_release_path(path);
1332 btrfs_release_prepared_delayed_node(delayed_node);
1333 total_done++;
1334 continue;
1335 }
1336
1337 block_rsv = trans->block_rsv;
1338 trans->block_rsv = &root->fs_info->delayed_block_rsv;
1339
1340 __btrfs_commit_inode_delayed_items(trans, path, delayed_node);
1341
1342 trans->block_rsv = block_rsv;
1343 btrfs_end_transaction(trans);
1344 btrfs_btree_balance_dirty_nodelay(root->fs_info);
1345
1346 btrfs_release_path(path);
1347 btrfs_release_prepared_delayed_node(delayed_node);
1348 total_done++;
1349
1350 } while ((async_work->nr == 0 && total_done < BTRFS_DELAYED_WRITEBACK)
1351 || total_done < async_work->nr);
1352
1353 btrfs_free_path(path);
1354out:
1355 wake_up(&delayed_root->wait);
1356 kfree(async_work);
1357}
1358
1359
1360static int btrfs_wq_run_delayed_node(struct btrfs_delayed_root *delayed_root,
1361 struct btrfs_fs_info *fs_info, int nr)
1362{
1363 struct btrfs_async_delayed_work *async_work;
1364
1365 async_work = kmalloc(sizeof(*async_work), GFP_NOFS);
1366 if (!async_work)
1367 return -ENOMEM;
1368
1369 async_work->delayed_root = delayed_root;
1370 btrfs_init_work(&async_work->work, btrfs_delayed_meta_helper,
1371 btrfs_async_run_delayed_root, NULL, NULL);
1372 async_work->nr = nr;
1373
1374 btrfs_queue_work(fs_info->delayed_workers, &async_work->work);
1375 return 0;
1376}
1377
1378void btrfs_assert_delayed_root_empty(struct btrfs_fs_info *fs_info)
1379{
1380 WARN_ON(btrfs_first_delayed_node(fs_info->delayed_root));
1381}
1382
1383static int could_end_wait(struct btrfs_delayed_root *delayed_root, int seq)
1384{
1385 int val = atomic_read(&delayed_root->items_seq);
1386
1387 if (val < seq || val >= seq + BTRFS_DELAYED_BATCH)
1388 return 1;
1389
1390 if (atomic_read(&delayed_root->items) < BTRFS_DELAYED_BACKGROUND)
1391 return 1;
1392
1393 return 0;
1394}
1395
1396void btrfs_balance_delayed_items(struct btrfs_fs_info *fs_info)
1397{
1398 struct btrfs_delayed_root *delayed_root = fs_info->delayed_root;
1399
1400 if ((atomic_read(&delayed_root->items) < BTRFS_DELAYED_BACKGROUND) ||
1401 btrfs_workqueue_normal_congested(fs_info->delayed_workers))
1402 return;
1403
1404 if (atomic_read(&delayed_root->items) >= BTRFS_DELAYED_WRITEBACK) {
1405 int seq;
1406 int ret;
1407
1408 seq = atomic_read(&delayed_root->items_seq);
1409
1410 ret = btrfs_wq_run_delayed_node(delayed_root, fs_info, 0);
1411 if (ret)
1412 return;
1413
1414 wait_event_interruptible(delayed_root->wait,
1415 could_end_wait(delayed_root, seq));
1416 return;
1417 }
1418
1419 btrfs_wq_run_delayed_node(delayed_root, fs_info, BTRFS_DELAYED_BATCH);
1420}
1421
1422/* Will return 0 or -ENOMEM */
1423int btrfs_insert_delayed_dir_index(struct btrfs_trans_handle *trans,
1424 const char *name, int name_len,
1425 struct btrfs_inode *dir,
1426 struct btrfs_disk_key *disk_key, u8 type,
1427 u64 index)
1428{
1429 struct btrfs_delayed_node *delayed_node;
1430 struct btrfs_delayed_item *delayed_item;
1431 struct btrfs_dir_item *dir_item;
1432 int ret;
1433
1434 delayed_node = btrfs_get_or_create_delayed_node(dir);
1435 if (IS_ERR(delayed_node))
1436 return PTR_ERR(delayed_node);
1437
1438 delayed_item = btrfs_alloc_delayed_item(sizeof(*dir_item) + name_len);
1439 if (!delayed_item) {
1440 ret = -ENOMEM;
1441 goto release_node;
1442 }
1443
1444 delayed_item->key.objectid = btrfs_ino(dir);
1445 delayed_item->key.type = BTRFS_DIR_INDEX_KEY;
1446 delayed_item->key.offset = index;
1447
1448 dir_item = (struct btrfs_dir_item *)delayed_item->data;
1449 dir_item->location = *disk_key;
1450 btrfs_set_stack_dir_transid(dir_item, trans->transid);
1451 btrfs_set_stack_dir_data_len(dir_item, 0);
1452 btrfs_set_stack_dir_name_len(dir_item, name_len);
1453 btrfs_set_stack_dir_type(dir_item, type);
1454 memcpy((char *)(dir_item + 1), name, name_len);
1455
1456 ret = btrfs_delayed_item_reserve_metadata(trans, dir->root, delayed_item);
1457 /*
1458 * we have reserved enough space when we start a new transaction,
1459 * so reserving metadata failure is impossible
1460 */
1461 BUG_ON(ret);
1462
1463 mutex_lock(&delayed_node->mutex);
1464 ret = __btrfs_add_delayed_insertion_item(delayed_node, delayed_item);
1465 if (unlikely(ret)) {
1466 btrfs_err(trans->fs_info,
1467 "err add delayed dir index item(name: %.*s) into the insertion tree of the delayed node(root id: %llu, inode id: %llu, errno: %d)",
1468 name_len, name, delayed_node->root->root_key.objectid,
1469 delayed_node->inode_id, ret);
1470 BUG();
1471 }
1472 mutex_unlock(&delayed_node->mutex);
1473
1474release_node:
1475 btrfs_release_delayed_node(delayed_node);
1476 return ret;
1477}
1478
1479static int btrfs_delete_delayed_insertion_item(struct btrfs_fs_info *fs_info,
1480 struct btrfs_delayed_node *node,
1481 struct btrfs_key *key)
1482{
1483 struct btrfs_delayed_item *item;
1484
1485 mutex_lock(&node->mutex);
1486 item = __btrfs_lookup_delayed_insertion_item(node, key);
1487 if (!item) {
1488 mutex_unlock(&node->mutex);
1489 return 1;
1490 }
1491
1492 btrfs_delayed_item_release_metadata(node->root, item);
1493 btrfs_release_delayed_item(item);
1494 mutex_unlock(&node->mutex);
1495 return 0;
1496}
1497
1498int btrfs_delete_delayed_dir_index(struct btrfs_trans_handle *trans,
1499 struct btrfs_inode *dir, u64 index)
1500{
1501 struct btrfs_delayed_node *node;
1502 struct btrfs_delayed_item *item;
1503 struct btrfs_key item_key;
1504 int ret;
1505
1506 node = btrfs_get_or_create_delayed_node(dir);
1507 if (IS_ERR(node))
1508 return PTR_ERR(node);
1509
1510 item_key.objectid = btrfs_ino(dir);
1511 item_key.type = BTRFS_DIR_INDEX_KEY;
1512 item_key.offset = index;
1513
1514 ret = btrfs_delete_delayed_insertion_item(trans->fs_info, node,
1515 &item_key);
1516 if (!ret)
1517 goto end;
1518
1519 item = btrfs_alloc_delayed_item(0);
1520 if (!item) {
1521 ret = -ENOMEM;
1522 goto end;
1523 }
1524
1525 item->key = item_key;
1526
1527 ret = btrfs_delayed_item_reserve_metadata(trans, dir->root, item);
1528 /*
1529 * we have reserved enough space when we start a new transaction,
1530 * so reserving metadata failure is impossible.
1531 */
1532 if (ret < 0) {
1533 btrfs_err(trans->fs_info,
1534"metadata reservation failed for delayed dir item deltiona, should have been reserved");
1535 btrfs_release_delayed_item(item);
1536 goto end;
1537 }
1538
1539 mutex_lock(&node->mutex);
1540 ret = __btrfs_add_delayed_deletion_item(node, item);
1541 if (unlikely(ret)) {
1542 btrfs_err(trans->fs_info,
1543 "err add delayed dir index item(index: %llu) into the deletion tree of the delayed node(root id: %llu, inode id: %llu, errno: %d)",
1544 index, node->root->root_key.objectid,
1545 node->inode_id, ret);
1546 btrfs_delayed_item_release_metadata(dir->root, item);
1547 btrfs_release_delayed_item(item);
1548 }
1549 mutex_unlock(&node->mutex);
1550end:
1551 btrfs_release_delayed_node(node);
1552 return ret;
1553}
1554
1555int btrfs_inode_delayed_dir_index_count(struct btrfs_inode *inode)
1556{
1557 struct btrfs_delayed_node *delayed_node = btrfs_get_delayed_node(inode);
1558
1559 if (!delayed_node)
1560 return -ENOENT;
1561
1562 /*
1563 * Since we have held i_mutex of this directory, it is impossible that
1564 * a new directory index is added into the delayed node and index_cnt
1565 * is updated now. So we needn't lock the delayed node.
1566 */
1567 if (!delayed_node->index_cnt) {
1568 btrfs_release_delayed_node(delayed_node);
1569 return -EINVAL;
1570 }
1571
1572 inode->index_cnt = delayed_node->index_cnt;
1573 btrfs_release_delayed_node(delayed_node);
1574 return 0;
1575}
1576
1577bool btrfs_readdir_get_delayed_items(struct inode *inode,
1578 struct list_head *ins_list,
1579 struct list_head *del_list)
1580{
1581 struct btrfs_delayed_node *delayed_node;
1582 struct btrfs_delayed_item *item;
1583
1584 delayed_node = btrfs_get_delayed_node(BTRFS_I(inode));
1585 if (!delayed_node)
1586 return false;
1587
1588 /*
1589 * We can only do one readdir with delayed items at a time because of
1590 * item->readdir_list.
1591 */
1592 inode_unlock_shared(inode);
1593 inode_lock(inode);
1594
1595 mutex_lock(&delayed_node->mutex);
1596 item = __btrfs_first_delayed_insertion_item(delayed_node);
1597 while (item) {
1598 refcount_inc(&item->refs);
1599 list_add_tail(&item->readdir_list, ins_list);
1600 item = __btrfs_next_delayed_item(item);
1601 }
1602
1603 item = __btrfs_first_delayed_deletion_item(delayed_node);
1604 while (item) {
1605 refcount_inc(&item->refs);
1606 list_add_tail(&item->readdir_list, del_list);
1607 item = __btrfs_next_delayed_item(item);
1608 }
1609 mutex_unlock(&delayed_node->mutex);
1610 /*
1611 * This delayed node is still cached in the btrfs inode, so refs
1612 * must be > 1 now, and we needn't check it is going to be freed
1613 * or not.
1614 *
1615 * Besides that, this function is used to read dir, we do not
1616 * insert/delete delayed items in this period. So we also needn't
1617 * requeue or dequeue this delayed node.
1618 */
1619 refcount_dec(&delayed_node->refs);
1620
1621 return true;
1622}
1623
1624void btrfs_readdir_put_delayed_items(struct inode *inode,
1625 struct list_head *ins_list,
1626 struct list_head *del_list)
1627{
1628 struct btrfs_delayed_item *curr, *next;
1629
1630 list_for_each_entry_safe(curr, next, ins_list, readdir_list) {
1631 list_del(&curr->readdir_list);
1632 if (refcount_dec_and_test(&curr->refs))
1633 kfree(curr);
1634 }
1635
1636 list_for_each_entry_safe(curr, next, del_list, readdir_list) {
1637 list_del(&curr->readdir_list);
1638 if (refcount_dec_and_test(&curr->refs))
1639 kfree(curr);
1640 }
1641
1642 /*
1643 * The VFS is going to do up_read(), so we need to downgrade back to a
1644 * read lock.
1645 */
1646 downgrade_write(&inode->i_rwsem);
1647}
1648
1649int btrfs_should_delete_dir_index(struct list_head *del_list,
1650 u64 index)
1651{
1652 struct btrfs_delayed_item *curr;
1653 int ret = 0;
1654
1655 list_for_each_entry(curr, del_list, readdir_list) {
1656 if (curr->key.offset > index)
1657 break;
1658 if (curr->key.offset == index) {
1659 ret = 1;
1660 break;
1661 }
1662 }
1663 return ret;
1664}
1665
1666/*
1667 * btrfs_readdir_delayed_dir_index - read dir info stored in the delayed tree
1668 *
1669 */
1670int btrfs_readdir_delayed_dir_index(struct dir_context *ctx,
1671 struct list_head *ins_list)
1672{
1673 struct btrfs_dir_item *di;
1674 struct btrfs_delayed_item *curr, *next;
1675 struct btrfs_key location;
1676 char *name;
1677 int name_len;
1678 int over = 0;
1679 unsigned char d_type;
1680
1681 if (list_empty(ins_list))
1682 return 0;
1683
1684 /*
1685 * Changing the data of the delayed item is impossible. So
1686 * we needn't lock them. And we have held i_mutex of the
1687 * directory, nobody can delete any directory indexes now.
1688 */
1689 list_for_each_entry_safe(curr, next, ins_list, readdir_list) {
1690 list_del(&curr->readdir_list);
1691
1692 if (curr->key.offset < ctx->pos) {
1693 if (refcount_dec_and_test(&curr->refs))
1694 kfree(curr);
1695 continue;
1696 }
1697
1698 ctx->pos = curr->key.offset;
1699
1700 di = (struct btrfs_dir_item *)curr->data;
1701 name = (char *)(di + 1);
1702 name_len = btrfs_stack_dir_name_len(di);
1703
1704 d_type = fs_ftype_to_dtype(di->type);
1705 btrfs_disk_key_to_cpu(&location, &di->location);
1706
1707 over = !dir_emit(ctx, name, name_len,
1708 location.objectid, d_type);
1709
1710 if (refcount_dec_and_test(&curr->refs))
1711 kfree(curr);
1712
1713 if (over)
1714 return 1;
1715 ctx->pos++;
1716 }
1717 return 0;
1718}
1719
1720static void fill_stack_inode_item(struct btrfs_trans_handle *trans,
1721 struct btrfs_inode_item *inode_item,
1722 struct inode *inode)
1723{
1724 btrfs_set_stack_inode_uid(inode_item, i_uid_read(inode));
1725 btrfs_set_stack_inode_gid(inode_item, i_gid_read(inode));
1726 btrfs_set_stack_inode_size(inode_item, BTRFS_I(inode)->disk_i_size);
1727 btrfs_set_stack_inode_mode(inode_item, inode->i_mode);
1728 btrfs_set_stack_inode_nlink(inode_item, inode->i_nlink);
1729 btrfs_set_stack_inode_nbytes(inode_item, inode_get_bytes(inode));
1730 btrfs_set_stack_inode_generation(inode_item,
1731 BTRFS_I(inode)->generation);
1732 btrfs_set_stack_inode_sequence(inode_item,
1733 inode_peek_iversion(inode));
1734 btrfs_set_stack_inode_transid(inode_item, trans->transid);
1735 btrfs_set_stack_inode_rdev(inode_item, inode->i_rdev);
1736 btrfs_set_stack_inode_flags(inode_item, BTRFS_I(inode)->flags);
1737 btrfs_set_stack_inode_block_group(inode_item, 0);
1738
1739 btrfs_set_stack_timespec_sec(&inode_item->atime,
1740 inode->i_atime.tv_sec);
1741 btrfs_set_stack_timespec_nsec(&inode_item->atime,
1742 inode->i_atime.tv_nsec);
1743
1744 btrfs_set_stack_timespec_sec(&inode_item->mtime,
1745 inode->i_mtime.tv_sec);
1746 btrfs_set_stack_timespec_nsec(&inode_item->mtime,
1747 inode->i_mtime.tv_nsec);
1748
1749 btrfs_set_stack_timespec_sec(&inode_item->ctime,
1750 inode->i_ctime.tv_sec);
1751 btrfs_set_stack_timespec_nsec(&inode_item->ctime,
1752 inode->i_ctime.tv_nsec);
1753
1754 btrfs_set_stack_timespec_sec(&inode_item->otime,
1755 BTRFS_I(inode)->i_otime.tv_sec);
1756 btrfs_set_stack_timespec_nsec(&inode_item->otime,
1757 BTRFS_I(inode)->i_otime.tv_nsec);
1758}
1759
1760int btrfs_fill_inode(struct inode *inode, u32 *rdev)
1761{
1762 struct btrfs_delayed_node *delayed_node;
1763 struct btrfs_inode_item *inode_item;
1764
1765 delayed_node = btrfs_get_delayed_node(BTRFS_I(inode));
1766 if (!delayed_node)
1767 return -ENOENT;
1768
1769 mutex_lock(&delayed_node->mutex);
1770 if (!test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
1771 mutex_unlock(&delayed_node->mutex);
1772 btrfs_release_delayed_node(delayed_node);
1773 return -ENOENT;
1774 }
1775
1776 inode_item = &delayed_node->inode_item;
1777
1778 i_uid_write(inode, btrfs_stack_inode_uid(inode_item));
1779 i_gid_write(inode, btrfs_stack_inode_gid(inode_item));
1780 btrfs_i_size_write(BTRFS_I(inode), btrfs_stack_inode_size(inode_item));
1781 inode->i_mode = btrfs_stack_inode_mode(inode_item);
1782 set_nlink(inode, btrfs_stack_inode_nlink(inode_item));
1783 inode_set_bytes(inode, btrfs_stack_inode_nbytes(inode_item));
1784 BTRFS_I(inode)->generation = btrfs_stack_inode_generation(inode_item);
1785 BTRFS_I(inode)->last_trans = btrfs_stack_inode_transid(inode_item);
1786
1787 inode_set_iversion_queried(inode,
1788 btrfs_stack_inode_sequence(inode_item));
1789 inode->i_rdev = 0;
1790 *rdev = btrfs_stack_inode_rdev(inode_item);
1791 BTRFS_I(inode)->flags = btrfs_stack_inode_flags(inode_item);
1792
1793 inode->i_atime.tv_sec = btrfs_stack_timespec_sec(&inode_item->atime);
1794 inode->i_atime.tv_nsec = btrfs_stack_timespec_nsec(&inode_item->atime);
1795
1796 inode->i_mtime.tv_sec = btrfs_stack_timespec_sec(&inode_item->mtime);
1797 inode->i_mtime.tv_nsec = btrfs_stack_timespec_nsec(&inode_item->mtime);
1798
1799 inode->i_ctime.tv_sec = btrfs_stack_timespec_sec(&inode_item->ctime);
1800 inode->i_ctime.tv_nsec = btrfs_stack_timespec_nsec(&inode_item->ctime);
1801
1802 BTRFS_I(inode)->i_otime.tv_sec =
1803 btrfs_stack_timespec_sec(&inode_item->otime);
1804 BTRFS_I(inode)->i_otime.tv_nsec =
1805 btrfs_stack_timespec_nsec(&inode_item->otime);
1806
1807 inode->i_generation = BTRFS_I(inode)->generation;
1808 BTRFS_I(inode)->index_cnt = (u64)-1;
1809
1810 mutex_unlock(&delayed_node->mutex);
1811 btrfs_release_delayed_node(delayed_node);
1812 return 0;
1813}
1814
1815int btrfs_delayed_update_inode(struct btrfs_trans_handle *trans,
1816 struct btrfs_root *root, struct inode *inode)
1817{
1818 struct btrfs_delayed_node *delayed_node;
1819 int ret = 0;
1820
1821 delayed_node = btrfs_get_or_create_delayed_node(BTRFS_I(inode));
1822 if (IS_ERR(delayed_node))
1823 return PTR_ERR(delayed_node);
1824
1825 mutex_lock(&delayed_node->mutex);
1826 if (test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
1827 fill_stack_inode_item(trans, &delayed_node->inode_item, inode);
1828 goto release_node;
1829 }
1830
1831 ret = btrfs_delayed_inode_reserve_metadata(trans, root, BTRFS_I(inode),
1832 delayed_node);
1833 if (ret)
1834 goto release_node;
1835
1836 fill_stack_inode_item(trans, &delayed_node->inode_item, inode);
1837 set_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags);
1838 delayed_node->count++;
1839 atomic_inc(&root->fs_info->delayed_root->items);
1840release_node:
1841 mutex_unlock(&delayed_node->mutex);
1842 btrfs_release_delayed_node(delayed_node);
1843 return ret;
1844}
1845
1846int btrfs_delayed_delete_inode_ref(struct btrfs_inode *inode)
1847{
1848 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1849 struct btrfs_delayed_node *delayed_node;
1850
1851 /*
1852 * we don't do delayed inode updates during log recovery because it
1853 * leads to enospc problems. This means we also can't do
1854 * delayed inode refs
1855 */
1856 if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
1857 return -EAGAIN;
1858
1859 delayed_node = btrfs_get_or_create_delayed_node(inode);
1860 if (IS_ERR(delayed_node))
1861 return PTR_ERR(delayed_node);
1862
1863 /*
1864 * We don't reserve space for inode ref deletion is because:
1865 * - We ONLY do async inode ref deletion for the inode who has only
1866 * one link(i_nlink == 1), it means there is only one inode ref.
1867 * And in most case, the inode ref and the inode item are in the
1868 * same leaf, and we will deal with them at the same time.
1869 * Since we are sure we will reserve the space for the inode item,
1870 * it is unnecessary to reserve space for inode ref deletion.
1871 * - If the inode ref and the inode item are not in the same leaf,
1872 * We also needn't worry about enospc problem, because we reserve
1873 * much more space for the inode update than it needs.
1874 * - At the worst, we can steal some space from the global reservation.
1875 * It is very rare.
1876 */
1877 mutex_lock(&delayed_node->mutex);
1878 if (test_bit(BTRFS_DELAYED_NODE_DEL_IREF, &delayed_node->flags))
1879 goto release_node;
1880
1881 set_bit(BTRFS_DELAYED_NODE_DEL_IREF, &delayed_node->flags);
1882 delayed_node->count++;
1883 atomic_inc(&fs_info->delayed_root->items);
1884release_node:
1885 mutex_unlock(&delayed_node->mutex);
1886 btrfs_release_delayed_node(delayed_node);
1887 return 0;
1888}
1889
1890static void __btrfs_kill_delayed_node(struct btrfs_delayed_node *delayed_node)
1891{
1892 struct btrfs_root *root = delayed_node->root;
1893 struct btrfs_fs_info *fs_info = root->fs_info;
1894 struct btrfs_delayed_item *curr_item, *prev_item;
1895
1896 mutex_lock(&delayed_node->mutex);
1897 curr_item = __btrfs_first_delayed_insertion_item(delayed_node);
1898 while (curr_item) {
1899 btrfs_delayed_item_release_metadata(root, curr_item);
1900 prev_item = curr_item;
1901 curr_item = __btrfs_next_delayed_item(prev_item);
1902 btrfs_release_delayed_item(prev_item);
1903 }
1904
1905 curr_item = __btrfs_first_delayed_deletion_item(delayed_node);
1906 while (curr_item) {
1907 btrfs_delayed_item_release_metadata(root, curr_item);
1908 prev_item = curr_item;
1909 curr_item = __btrfs_next_delayed_item(prev_item);
1910 btrfs_release_delayed_item(prev_item);
1911 }
1912
1913 if (test_bit(BTRFS_DELAYED_NODE_DEL_IREF, &delayed_node->flags))
1914 btrfs_release_delayed_iref(delayed_node);
1915
1916 if (test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
1917 btrfs_delayed_inode_release_metadata(fs_info, delayed_node, false);
1918 btrfs_release_delayed_inode(delayed_node);
1919 }
1920 mutex_unlock(&delayed_node->mutex);
1921}
1922
1923void btrfs_kill_delayed_inode_items(struct btrfs_inode *inode)
1924{
1925 struct btrfs_delayed_node *delayed_node;
1926
1927 delayed_node = btrfs_get_delayed_node(inode);
1928 if (!delayed_node)
1929 return;
1930
1931 __btrfs_kill_delayed_node(delayed_node);
1932 btrfs_release_delayed_node(delayed_node);
1933}
1934
1935void btrfs_kill_all_delayed_nodes(struct btrfs_root *root)
1936{
1937 u64 inode_id = 0;
1938 struct btrfs_delayed_node *delayed_nodes[8];
1939 int i, n;
1940
1941 while (1) {
1942 spin_lock(&root->inode_lock);
1943 n = radix_tree_gang_lookup(&root->delayed_nodes_tree,
1944 (void **)delayed_nodes, inode_id,
1945 ARRAY_SIZE(delayed_nodes));
1946 if (!n) {
1947 spin_unlock(&root->inode_lock);
1948 break;
1949 }
1950
1951 inode_id = delayed_nodes[n - 1]->inode_id + 1;
1952
1953 for (i = 0; i < n; i++)
1954 refcount_inc(&delayed_nodes[i]->refs);
1955 spin_unlock(&root->inode_lock);
1956
1957 for (i = 0; i < n; i++) {
1958 __btrfs_kill_delayed_node(delayed_nodes[i]);
1959 btrfs_release_delayed_node(delayed_nodes[i]);
1960 }
1961 }
1962}
1963
1964void btrfs_destroy_delayed_inodes(struct btrfs_fs_info *fs_info)
1965{
1966 struct btrfs_delayed_node *curr_node, *prev_node;
1967
1968 curr_node = btrfs_first_delayed_node(fs_info->delayed_root);
1969 while (curr_node) {
1970 __btrfs_kill_delayed_node(curr_node);
1971
1972 prev_node = curr_node;
1973 curr_node = btrfs_next_delayed_node(curr_node);
1974 btrfs_release_delayed_node(prev_node);
1975 }
1976}
1977
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2011 Fujitsu. All rights reserved.
4 * Written by Miao Xie <miaox@cn.fujitsu.com>
5 */
6
7#include <linux/slab.h>
8#include <linux/iversion.h>
9#include "ctree.h"
10#include "fs.h"
11#include "messages.h"
12#include "misc.h"
13#include "delayed-inode.h"
14#include "disk-io.h"
15#include "transaction.h"
16#include "qgroup.h"
17#include "locking.h"
18#include "inode-item.h"
19#include "space-info.h"
20#include "accessors.h"
21#include "file-item.h"
22
23#define BTRFS_DELAYED_WRITEBACK 512
24#define BTRFS_DELAYED_BACKGROUND 128
25#define BTRFS_DELAYED_BATCH 16
26
27static struct kmem_cache *delayed_node_cache;
28
29int __init btrfs_delayed_inode_init(void)
30{
31 delayed_node_cache = KMEM_CACHE(btrfs_delayed_node, 0);
32 if (!delayed_node_cache)
33 return -ENOMEM;
34 return 0;
35}
36
37void __cold btrfs_delayed_inode_exit(void)
38{
39 kmem_cache_destroy(delayed_node_cache);
40}
41
42void btrfs_init_delayed_root(struct btrfs_delayed_root *delayed_root)
43{
44 atomic_set(&delayed_root->items, 0);
45 atomic_set(&delayed_root->items_seq, 0);
46 delayed_root->nodes = 0;
47 spin_lock_init(&delayed_root->lock);
48 init_waitqueue_head(&delayed_root->wait);
49 INIT_LIST_HEAD(&delayed_root->node_list);
50 INIT_LIST_HEAD(&delayed_root->prepare_list);
51}
52
53static inline void btrfs_init_delayed_node(
54 struct btrfs_delayed_node *delayed_node,
55 struct btrfs_root *root, u64 inode_id)
56{
57 delayed_node->root = root;
58 delayed_node->inode_id = inode_id;
59 refcount_set(&delayed_node->refs, 0);
60 delayed_node->ins_root = RB_ROOT_CACHED;
61 delayed_node->del_root = RB_ROOT_CACHED;
62 mutex_init(&delayed_node->mutex);
63 INIT_LIST_HEAD(&delayed_node->n_list);
64 INIT_LIST_HEAD(&delayed_node->p_list);
65}
66
67static struct btrfs_delayed_node *btrfs_get_delayed_node(
68 struct btrfs_inode *btrfs_inode)
69{
70 struct btrfs_root *root = btrfs_inode->root;
71 u64 ino = btrfs_ino(btrfs_inode);
72 struct btrfs_delayed_node *node;
73
74 node = READ_ONCE(btrfs_inode->delayed_node);
75 if (node) {
76 refcount_inc(&node->refs);
77 return node;
78 }
79
80 xa_lock(&root->delayed_nodes);
81 node = xa_load(&root->delayed_nodes, ino);
82
83 if (node) {
84 if (btrfs_inode->delayed_node) {
85 refcount_inc(&node->refs); /* can be accessed */
86 BUG_ON(btrfs_inode->delayed_node != node);
87 xa_unlock(&root->delayed_nodes);
88 return node;
89 }
90
91 /*
92 * It's possible that we're racing into the middle of removing
93 * this node from the xarray. In this case, the refcount
94 * was zero and it should never go back to one. Just return
95 * NULL like it was never in the xarray at all; our release
96 * function is in the process of removing it.
97 *
98 * Some implementations of refcount_inc refuse to bump the
99 * refcount once it has hit zero. If we don't do this dance
100 * here, refcount_inc() may decide to just WARN_ONCE() instead
101 * of actually bumping the refcount.
102 *
103 * If this node is properly in the xarray, we want to bump the
104 * refcount twice, once for the inode and once for this get
105 * operation.
106 */
107 if (refcount_inc_not_zero(&node->refs)) {
108 refcount_inc(&node->refs);
109 btrfs_inode->delayed_node = node;
110 } else {
111 node = NULL;
112 }
113
114 xa_unlock(&root->delayed_nodes);
115 return node;
116 }
117 xa_unlock(&root->delayed_nodes);
118
119 return NULL;
120}
121
122/* Will return either the node or PTR_ERR(-ENOMEM) */
123static struct btrfs_delayed_node *btrfs_get_or_create_delayed_node(
124 struct btrfs_inode *btrfs_inode)
125{
126 struct btrfs_delayed_node *node;
127 struct btrfs_root *root = btrfs_inode->root;
128 u64 ino = btrfs_ino(btrfs_inode);
129 int ret;
130 void *ptr;
131
132again:
133 node = btrfs_get_delayed_node(btrfs_inode);
134 if (node)
135 return node;
136
137 node = kmem_cache_zalloc(delayed_node_cache, GFP_NOFS);
138 if (!node)
139 return ERR_PTR(-ENOMEM);
140 btrfs_init_delayed_node(node, root, ino);
141
142 /* Cached in the inode and can be accessed. */
143 refcount_set(&node->refs, 2);
144
145 /* Allocate and reserve the slot, from now it can return a NULL from xa_load(). */
146 ret = xa_reserve(&root->delayed_nodes, ino, GFP_NOFS);
147 if (ret == -ENOMEM) {
148 kmem_cache_free(delayed_node_cache, node);
149 return ERR_PTR(-ENOMEM);
150 }
151 xa_lock(&root->delayed_nodes);
152 ptr = xa_load(&root->delayed_nodes, ino);
153 if (ptr) {
154 /* Somebody inserted it, go back and read it. */
155 xa_unlock(&root->delayed_nodes);
156 kmem_cache_free(delayed_node_cache, node);
157 node = NULL;
158 goto again;
159 }
160 ptr = __xa_store(&root->delayed_nodes, ino, node, GFP_ATOMIC);
161 ASSERT(xa_err(ptr) != -EINVAL);
162 ASSERT(xa_err(ptr) != -ENOMEM);
163 ASSERT(ptr == NULL);
164 btrfs_inode->delayed_node = node;
165 xa_unlock(&root->delayed_nodes);
166
167 return node;
168}
169
170/*
171 * Call it when holding delayed_node->mutex
172 *
173 * If mod = 1, add this node into the prepared list.
174 */
175static void btrfs_queue_delayed_node(struct btrfs_delayed_root *root,
176 struct btrfs_delayed_node *node,
177 int mod)
178{
179 spin_lock(&root->lock);
180 if (test_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags)) {
181 if (!list_empty(&node->p_list))
182 list_move_tail(&node->p_list, &root->prepare_list);
183 else if (mod)
184 list_add_tail(&node->p_list, &root->prepare_list);
185 } else {
186 list_add_tail(&node->n_list, &root->node_list);
187 list_add_tail(&node->p_list, &root->prepare_list);
188 refcount_inc(&node->refs); /* inserted into list */
189 root->nodes++;
190 set_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags);
191 }
192 spin_unlock(&root->lock);
193}
194
195/* Call it when holding delayed_node->mutex */
196static void btrfs_dequeue_delayed_node(struct btrfs_delayed_root *root,
197 struct btrfs_delayed_node *node)
198{
199 spin_lock(&root->lock);
200 if (test_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags)) {
201 root->nodes--;
202 refcount_dec(&node->refs); /* not in the list */
203 list_del_init(&node->n_list);
204 if (!list_empty(&node->p_list))
205 list_del_init(&node->p_list);
206 clear_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags);
207 }
208 spin_unlock(&root->lock);
209}
210
211static struct btrfs_delayed_node *btrfs_first_delayed_node(
212 struct btrfs_delayed_root *delayed_root)
213{
214 struct list_head *p;
215 struct btrfs_delayed_node *node = NULL;
216
217 spin_lock(&delayed_root->lock);
218 if (list_empty(&delayed_root->node_list))
219 goto out;
220
221 p = delayed_root->node_list.next;
222 node = list_entry(p, struct btrfs_delayed_node, n_list);
223 refcount_inc(&node->refs);
224out:
225 spin_unlock(&delayed_root->lock);
226
227 return node;
228}
229
230static struct btrfs_delayed_node *btrfs_next_delayed_node(
231 struct btrfs_delayed_node *node)
232{
233 struct btrfs_delayed_root *delayed_root;
234 struct list_head *p;
235 struct btrfs_delayed_node *next = NULL;
236
237 delayed_root = node->root->fs_info->delayed_root;
238 spin_lock(&delayed_root->lock);
239 if (!test_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags)) {
240 /* not in the list */
241 if (list_empty(&delayed_root->node_list))
242 goto out;
243 p = delayed_root->node_list.next;
244 } else if (list_is_last(&node->n_list, &delayed_root->node_list))
245 goto out;
246 else
247 p = node->n_list.next;
248
249 next = list_entry(p, struct btrfs_delayed_node, n_list);
250 refcount_inc(&next->refs);
251out:
252 spin_unlock(&delayed_root->lock);
253
254 return next;
255}
256
257static void __btrfs_release_delayed_node(
258 struct btrfs_delayed_node *delayed_node,
259 int mod)
260{
261 struct btrfs_delayed_root *delayed_root;
262
263 if (!delayed_node)
264 return;
265
266 delayed_root = delayed_node->root->fs_info->delayed_root;
267
268 mutex_lock(&delayed_node->mutex);
269 if (delayed_node->count)
270 btrfs_queue_delayed_node(delayed_root, delayed_node, mod);
271 else
272 btrfs_dequeue_delayed_node(delayed_root, delayed_node);
273 mutex_unlock(&delayed_node->mutex);
274
275 if (refcount_dec_and_test(&delayed_node->refs)) {
276 struct btrfs_root *root = delayed_node->root;
277
278 xa_erase(&root->delayed_nodes, delayed_node->inode_id);
279 /*
280 * Once our refcount goes to zero, nobody is allowed to bump it
281 * back up. We can delete it now.
282 */
283 ASSERT(refcount_read(&delayed_node->refs) == 0);
284 kmem_cache_free(delayed_node_cache, delayed_node);
285 }
286}
287
288static inline void btrfs_release_delayed_node(struct btrfs_delayed_node *node)
289{
290 __btrfs_release_delayed_node(node, 0);
291}
292
293static struct btrfs_delayed_node *btrfs_first_prepared_delayed_node(
294 struct btrfs_delayed_root *delayed_root)
295{
296 struct list_head *p;
297 struct btrfs_delayed_node *node = NULL;
298
299 spin_lock(&delayed_root->lock);
300 if (list_empty(&delayed_root->prepare_list))
301 goto out;
302
303 p = delayed_root->prepare_list.next;
304 list_del_init(p);
305 node = list_entry(p, struct btrfs_delayed_node, p_list);
306 refcount_inc(&node->refs);
307out:
308 spin_unlock(&delayed_root->lock);
309
310 return node;
311}
312
313static inline void btrfs_release_prepared_delayed_node(
314 struct btrfs_delayed_node *node)
315{
316 __btrfs_release_delayed_node(node, 1);
317}
318
319static struct btrfs_delayed_item *btrfs_alloc_delayed_item(u16 data_len,
320 struct btrfs_delayed_node *node,
321 enum btrfs_delayed_item_type type)
322{
323 struct btrfs_delayed_item *item;
324
325 item = kmalloc(struct_size(item, data, data_len), GFP_NOFS);
326 if (item) {
327 item->data_len = data_len;
328 item->type = type;
329 item->bytes_reserved = 0;
330 item->delayed_node = node;
331 RB_CLEAR_NODE(&item->rb_node);
332 INIT_LIST_HEAD(&item->log_list);
333 item->logged = false;
334 refcount_set(&item->refs, 1);
335 }
336 return item;
337}
338
339/*
340 * Look up the delayed item by key.
341 *
342 * @delayed_node: pointer to the delayed node
343 * @index: the dir index value to lookup (offset of a dir index key)
344 *
345 * Note: if we don't find the right item, we will return the prev item and
346 * the next item.
347 */
348static struct btrfs_delayed_item *__btrfs_lookup_delayed_item(
349 struct rb_root *root,
350 u64 index)
351{
352 struct rb_node *node = root->rb_node;
353 struct btrfs_delayed_item *delayed_item = NULL;
354
355 while (node) {
356 delayed_item = rb_entry(node, struct btrfs_delayed_item,
357 rb_node);
358 if (delayed_item->index < index)
359 node = node->rb_right;
360 else if (delayed_item->index > index)
361 node = node->rb_left;
362 else
363 return delayed_item;
364 }
365
366 return NULL;
367}
368
369static int __btrfs_add_delayed_item(struct btrfs_delayed_node *delayed_node,
370 struct btrfs_delayed_item *ins)
371{
372 struct rb_node **p, *node;
373 struct rb_node *parent_node = NULL;
374 struct rb_root_cached *root;
375 struct btrfs_delayed_item *item;
376 bool leftmost = true;
377
378 if (ins->type == BTRFS_DELAYED_INSERTION_ITEM)
379 root = &delayed_node->ins_root;
380 else
381 root = &delayed_node->del_root;
382
383 p = &root->rb_root.rb_node;
384 node = &ins->rb_node;
385
386 while (*p) {
387 parent_node = *p;
388 item = rb_entry(parent_node, struct btrfs_delayed_item,
389 rb_node);
390
391 if (item->index < ins->index) {
392 p = &(*p)->rb_right;
393 leftmost = false;
394 } else if (item->index > ins->index) {
395 p = &(*p)->rb_left;
396 } else {
397 return -EEXIST;
398 }
399 }
400
401 rb_link_node(node, parent_node, p);
402 rb_insert_color_cached(node, root, leftmost);
403
404 if (ins->type == BTRFS_DELAYED_INSERTION_ITEM &&
405 ins->index >= delayed_node->index_cnt)
406 delayed_node->index_cnt = ins->index + 1;
407
408 delayed_node->count++;
409 atomic_inc(&delayed_node->root->fs_info->delayed_root->items);
410 return 0;
411}
412
413static void finish_one_item(struct btrfs_delayed_root *delayed_root)
414{
415 int seq = atomic_inc_return(&delayed_root->items_seq);
416
417 /* atomic_dec_return implies a barrier */
418 if ((atomic_dec_return(&delayed_root->items) <
419 BTRFS_DELAYED_BACKGROUND || seq % BTRFS_DELAYED_BATCH == 0))
420 cond_wake_up_nomb(&delayed_root->wait);
421}
422
423static void __btrfs_remove_delayed_item(struct btrfs_delayed_item *delayed_item)
424{
425 struct btrfs_delayed_node *delayed_node = delayed_item->delayed_node;
426 struct rb_root_cached *root;
427 struct btrfs_delayed_root *delayed_root;
428
429 /* Not inserted, ignore it. */
430 if (RB_EMPTY_NODE(&delayed_item->rb_node))
431 return;
432
433 /* If it's in a rbtree, then we need to have delayed node locked. */
434 lockdep_assert_held(&delayed_node->mutex);
435
436 delayed_root = delayed_node->root->fs_info->delayed_root;
437
438 if (delayed_item->type == BTRFS_DELAYED_INSERTION_ITEM)
439 root = &delayed_node->ins_root;
440 else
441 root = &delayed_node->del_root;
442
443 rb_erase_cached(&delayed_item->rb_node, root);
444 RB_CLEAR_NODE(&delayed_item->rb_node);
445 delayed_node->count--;
446
447 finish_one_item(delayed_root);
448}
449
450static void btrfs_release_delayed_item(struct btrfs_delayed_item *item)
451{
452 if (item) {
453 __btrfs_remove_delayed_item(item);
454 if (refcount_dec_and_test(&item->refs))
455 kfree(item);
456 }
457}
458
459static struct btrfs_delayed_item *__btrfs_first_delayed_insertion_item(
460 struct btrfs_delayed_node *delayed_node)
461{
462 struct rb_node *p;
463 struct btrfs_delayed_item *item = NULL;
464
465 p = rb_first_cached(&delayed_node->ins_root);
466 if (p)
467 item = rb_entry(p, struct btrfs_delayed_item, rb_node);
468
469 return item;
470}
471
472static struct btrfs_delayed_item *__btrfs_first_delayed_deletion_item(
473 struct btrfs_delayed_node *delayed_node)
474{
475 struct rb_node *p;
476 struct btrfs_delayed_item *item = NULL;
477
478 p = rb_first_cached(&delayed_node->del_root);
479 if (p)
480 item = rb_entry(p, struct btrfs_delayed_item, rb_node);
481
482 return item;
483}
484
485static struct btrfs_delayed_item *__btrfs_next_delayed_item(
486 struct btrfs_delayed_item *item)
487{
488 struct rb_node *p;
489 struct btrfs_delayed_item *next = NULL;
490
491 p = rb_next(&item->rb_node);
492 if (p)
493 next = rb_entry(p, struct btrfs_delayed_item, rb_node);
494
495 return next;
496}
497
498static int btrfs_delayed_item_reserve_metadata(struct btrfs_trans_handle *trans,
499 struct btrfs_delayed_item *item)
500{
501 struct btrfs_block_rsv *src_rsv;
502 struct btrfs_block_rsv *dst_rsv;
503 struct btrfs_fs_info *fs_info = trans->fs_info;
504 u64 num_bytes;
505 int ret;
506
507 if (!trans->bytes_reserved)
508 return 0;
509
510 src_rsv = trans->block_rsv;
511 dst_rsv = &fs_info->delayed_block_rsv;
512
513 num_bytes = btrfs_calc_insert_metadata_size(fs_info, 1);
514
515 /*
516 * Here we migrate space rsv from transaction rsv, since have already
517 * reserved space when starting a transaction. So no need to reserve
518 * qgroup space here.
519 */
520 ret = btrfs_block_rsv_migrate(src_rsv, dst_rsv, num_bytes, true);
521 if (!ret) {
522 trace_btrfs_space_reservation(fs_info, "delayed_item",
523 item->delayed_node->inode_id,
524 num_bytes, 1);
525 /*
526 * For insertions we track reserved metadata space by accounting
527 * for the number of leaves that will be used, based on the delayed
528 * node's curr_index_batch_size and index_item_leaves fields.
529 */
530 if (item->type == BTRFS_DELAYED_DELETION_ITEM)
531 item->bytes_reserved = num_bytes;
532 }
533
534 return ret;
535}
536
537static void btrfs_delayed_item_release_metadata(struct btrfs_root *root,
538 struct btrfs_delayed_item *item)
539{
540 struct btrfs_block_rsv *rsv;
541 struct btrfs_fs_info *fs_info = root->fs_info;
542
543 if (!item->bytes_reserved)
544 return;
545
546 rsv = &fs_info->delayed_block_rsv;
547 /*
548 * Check btrfs_delayed_item_reserve_metadata() to see why we don't need
549 * to release/reserve qgroup space.
550 */
551 trace_btrfs_space_reservation(fs_info, "delayed_item",
552 item->delayed_node->inode_id,
553 item->bytes_reserved, 0);
554 btrfs_block_rsv_release(fs_info, rsv, item->bytes_reserved, NULL);
555}
556
557static void btrfs_delayed_item_release_leaves(struct btrfs_delayed_node *node,
558 unsigned int num_leaves)
559{
560 struct btrfs_fs_info *fs_info = node->root->fs_info;
561 const u64 bytes = btrfs_calc_insert_metadata_size(fs_info, num_leaves);
562
563 /* There are no space reservations during log replay, bail out. */
564 if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
565 return;
566
567 trace_btrfs_space_reservation(fs_info, "delayed_item", node->inode_id,
568 bytes, 0);
569 btrfs_block_rsv_release(fs_info, &fs_info->delayed_block_rsv, bytes, NULL);
570}
571
572static int btrfs_delayed_inode_reserve_metadata(
573 struct btrfs_trans_handle *trans,
574 struct btrfs_root *root,
575 struct btrfs_delayed_node *node)
576{
577 struct btrfs_fs_info *fs_info = root->fs_info;
578 struct btrfs_block_rsv *src_rsv;
579 struct btrfs_block_rsv *dst_rsv;
580 u64 num_bytes;
581 int ret;
582
583 src_rsv = trans->block_rsv;
584 dst_rsv = &fs_info->delayed_block_rsv;
585
586 num_bytes = btrfs_calc_metadata_size(fs_info, 1);
587
588 /*
589 * btrfs_dirty_inode will update the inode under btrfs_join_transaction
590 * which doesn't reserve space for speed. This is a problem since we
591 * still need to reserve space for this update, so try to reserve the
592 * space.
593 *
594 * Now if src_rsv == delalloc_block_rsv we'll let it just steal since
595 * we always reserve enough to update the inode item.
596 */
597 if (!src_rsv || (!trans->bytes_reserved &&
598 src_rsv->type != BTRFS_BLOCK_RSV_DELALLOC)) {
599 ret = btrfs_qgroup_reserve_meta(root, num_bytes,
600 BTRFS_QGROUP_RSV_META_PREALLOC, true);
601 if (ret < 0)
602 return ret;
603 ret = btrfs_block_rsv_add(fs_info, dst_rsv, num_bytes,
604 BTRFS_RESERVE_NO_FLUSH);
605 /* NO_FLUSH could only fail with -ENOSPC */
606 ASSERT(ret == 0 || ret == -ENOSPC);
607 if (ret)
608 btrfs_qgroup_free_meta_prealloc(root, num_bytes);
609 } else {
610 ret = btrfs_block_rsv_migrate(src_rsv, dst_rsv, num_bytes, true);
611 }
612
613 if (!ret) {
614 trace_btrfs_space_reservation(fs_info, "delayed_inode",
615 node->inode_id, num_bytes, 1);
616 node->bytes_reserved = num_bytes;
617 }
618
619 return ret;
620}
621
622static void btrfs_delayed_inode_release_metadata(struct btrfs_fs_info *fs_info,
623 struct btrfs_delayed_node *node,
624 bool qgroup_free)
625{
626 struct btrfs_block_rsv *rsv;
627
628 if (!node->bytes_reserved)
629 return;
630
631 rsv = &fs_info->delayed_block_rsv;
632 trace_btrfs_space_reservation(fs_info, "delayed_inode",
633 node->inode_id, node->bytes_reserved, 0);
634 btrfs_block_rsv_release(fs_info, rsv, node->bytes_reserved, NULL);
635 if (qgroup_free)
636 btrfs_qgroup_free_meta_prealloc(node->root,
637 node->bytes_reserved);
638 else
639 btrfs_qgroup_convert_reserved_meta(node->root,
640 node->bytes_reserved);
641 node->bytes_reserved = 0;
642}
643
644/*
645 * Insert a single delayed item or a batch of delayed items, as many as possible
646 * that fit in a leaf. The delayed items (dir index keys) are sorted by their key
647 * in the rbtree, and if there's a gap between two consecutive dir index items,
648 * then it means at some point we had delayed dir indexes to add but they got
649 * removed (by btrfs_delete_delayed_dir_index()) before we attempted to flush them
650 * into the subvolume tree. Dir index keys also have their offsets coming from a
651 * monotonically increasing counter, so we can't get new keys with an offset that
652 * fits within a gap between delayed dir index items.
653 */
654static int btrfs_insert_delayed_item(struct btrfs_trans_handle *trans,
655 struct btrfs_root *root,
656 struct btrfs_path *path,
657 struct btrfs_delayed_item *first_item)
658{
659 struct btrfs_fs_info *fs_info = root->fs_info;
660 struct btrfs_delayed_node *node = first_item->delayed_node;
661 LIST_HEAD(item_list);
662 struct btrfs_delayed_item *curr;
663 struct btrfs_delayed_item *next;
664 const int max_size = BTRFS_LEAF_DATA_SIZE(fs_info);
665 struct btrfs_item_batch batch;
666 struct btrfs_key first_key;
667 const u32 first_data_size = first_item->data_len;
668 int total_size;
669 char *ins_data = NULL;
670 int ret;
671 bool continuous_keys_only = false;
672
673 lockdep_assert_held(&node->mutex);
674
675 /*
676 * During normal operation the delayed index offset is continuously
677 * increasing, so we can batch insert all items as there will not be any
678 * overlapping keys in the tree.
679 *
680 * The exception to this is log replay, where we may have interleaved
681 * offsets in the tree, so our batch needs to be continuous keys only in
682 * order to ensure we do not end up with out of order items in our leaf.
683 */
684 if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
685 continuous_keys_only = true;
686
687 /*
688 * For delayed items to insert, we track reserved metadata bytes based
689 * on the number of leaves that we will use.
690 * See btrfs_insert_delayed_dir_index() and
691 * btrfs_delayed_item_reserve_metadata()).
692 */
693 ASSERT(first_item->bytes_reserved == 0);
694
695 list_add_tail(&first_item->tree_list, &item_list);
696 batch.total_data_size = first_data_size;
697 batch.nr = 1;
698 total_size = first_data_size + sizeof(struct btrfs_item);
699 curr = first_item;
700
701 while (true) {
702 int next_size;
703
704 next = __btrfs_next_delayed_item(curr);
705 if (!next)
706 break;
707
708 /*
709 * We cannot allow gaps in the key space if we're doing log
710 * replay.
711 */
712 if (continuous_keys_only && (next->index != curr->index + 1))
713 break;
714
715 ASSERT(next->bytes_reserved == 0);
716
717 next_size = next->data_len + sizeof(struct btrfs_item);
718 if (total_size + next_size > max_size)
719 break;
720
721 list_add_tail(&next->tree_list, &item_list);
722 batch.nr++;
723 total_size += next_size;
724 batch.total_data_size += next->data_len;
725 curr = next;
726 }
727
728 if (batch.nr == 1) {
729 first_key.objectid = node->inode_id;
730 first_key.type = BTRFS_DIR_INDEX_KEY;
731 first_key.offset = first_item->index;
732 batch.keys = &first_key;
733 batch.data_sizes = &first_data_size;
734 } else {
735 struct btrfs_key *ins_keys;
736 u32 *ins_sizes;
737 int i = 0;
738
739 ins_data = kmalloc(batch.nr * sizeof(u32) +
740 batch.nr * sizeof(struct btrfs_key), GFP_NOFS);
741 if (!ins_data) {
742 ret = -ENOMEM;
743 goto out;
744 }
745 ins_sizes = (u32 *)ins_data;
746 ins_keys = (struct btrfs_key *)(ins_data + batch.nr * sizeof(u32));
747 batch.keys = ins_keys;
748 batch.data_sizes = ins_sizes;
749 list_for_each_entry(curr, &item_list, tree_list) {
750 ins_keys[i].objectid = node->inode_id;
751 ins_keys[i].type = BTRFS_DIR_INDEX_KEY;
752 ins_keys[i].offset = curr->index;
753 ins_sizes[i] = curr->data_len;
754 i++;
755 }
756 }
757
758 ret = btrfs_insert_empty_items(trans, root, path, &batch);
759 if (ret)
760 goto out;
761
762 list_for_each_entry(curr, &item_list, tree_list) {
763 char *data_ptr;
764
765 data_ptr = btrfs_item_ptr(path->nodes[0], path->slots[0], char);
766 write_extent_buffer(path->nodes[0], &curr->data,
767 (unsigned long)data_ptr, curr->data_len);
768 path->slots[0]++;
769 }
770
771 /*
772 * Now release our path before releasing the delayed items and their
773 * metadata reservations, so that we don't block other tasks for more
774 * time than needed.
775 */
776 btrfs_release_path(path);
777
778 ASSERT(node->index_item_leaves > 0);
779
780 /*
781 * For normal operations we will batch an entire leaf's worth of delayed
782 * items, so if there are more items to process we can decrement
783 * index_item_leaves by 1 as we inserted 1 leaf's worth of items.
784 *
785 * However for log replay we may not have inserted an entire leaf's
786 * worth of items, we may have not had continuous items, so decrementing
787 * here would mess up the index_item_leaves accounting. For this case
788 * only clean up the accounting when there are no items left.
789 */
790 if (next && !continuous_keys_only) {
791 /*
792 * We inserted one batch of items into a leaf a there are more
793 * items to flush in a future batch, now release one unit of
794 * metadata space from the delayed block reserve, corresponding
795 * the leaf we just flushed to.
796 */
797 btrfs_delayed_item_release_leaves(node, 1);
798 node->index_item_leaves--;
799 } else if (!next) {
800 /*
801 * There are no more items to insert. We can have a number of
802 * reserved leaves > 1 here - this happens when many dir index
803 * items are added and then removed before they are flushed (file
804 * names with a very short life, never span a transaction). So
805 * release all remaining leaves.
806 */
807 btrfs_delayed_item_release_leaves(node, node->index_item_leaves);
808 node->index_item_leaves = 0;
809 }
810
811 list_for_each_entry_safe(curr, next, &item_list, tree_list) {
812 list_del(&curr->tree_list);
813 btrfs_release_delayed_item(curr);
814 }
815out:
816 kfree(ins_data);
817 return ret;
818}
819
820static int btrfs_insert_delayed_items(struct btrfs_trans_handle *trans,
821 struct btrfs_path *path,
822 struct btrfs_root *root,
823 struct btrfs_delayed_node *node)
824{
825 int ret = 0;
826
827 while (ret == 0) {
828 struct btrfs_delayed_item *curr;
829
830 mutex_lock(&node->mutex);
831 curr = __btrfs_first_delayed_insertion_item(node);
832 if (!curr) {
833 mutex_unlock(&node->mutex);
834 break;
835 }
836 ret = btrfs_insert_delayed_item(trans, root, path, curr);
837 mutex_unlock(&node->mutex);
838 }
839
840 return ret;
841}
842
843static int btrfs_batch_delete_items(struct btrfs_trans_handle *trans,
844 struct btrfs_root *root,
845 struct btrfs_path *path,
846 struct btrfs_delayed_item *item)
847{
848 const u64 ino = item->delayed_node->inode_id;
849 struct btrfs_fs_info *fs_info = root->fs_info;
850 struct btrfs_delayed_item *curr, *next;
851 struct extent_buffer *leaf = path->nodes[0];
852 LIST_HEAD(batch_list);
853 int nitems, slot, last_slot;
854 int ret;
855 u64 total_reserved_size = item->bytes_reserved;
856
857 ASSERT(leaf != NULL);
858
859 slot = path->slots[0];
860 last_slot = btrfs_header_nritems(leaf) - 1;
861 /*
862 * Our caller always gives us a path pointing to an existing item, so
863 * this can not happen.
864 */
865 ASSERT(slot <= last_slot);
866 if (WARN_ON(slot > last_slot))
867 return -ENOENT;
868
869 nitems = 1;
870 curr = item;
871 list_add_tail(&curr->tree_list, &batch_list);
872
873 /*
874 * Keep checking if the next delayed item matches the next item in the
875 * leaf - if so, we can add it to the batch of items to delete from the
876 * leaf.
877 */
878 while (slot < last_slot) {
879 struct btrfs_key key;
880
881 next = __btrfs_next_delayed_item(curr);
882 if (!next)
883 break;
884
885 slot++;
886 btrfs_item_key_to_cpu(leaf, &key, slot);
887 if (key.objectid != ino ||
888 key.type != BTRFS_DIR_INDEX_KEY ||
889 key.offset != next->index)
890 break;
891 nitems++;
892 curr = next;
893 list_add_tail(&curr->tree_list, &batch_list);
894 total_reserved_size += curr->bytes_reserved;
895 }
896
897 ret = btrfs_del_items(trans, root, path, path->slots[0], nitems);
898 if (ret)
899 return ret;
900
901 /* In case of BTRFS_FS_LOG_RECOVERING items won't have reserved space */
902 if (total_reserved_size > 0) {
903 /*
904 * Check btrfs_delayed_item_reserve_metadata() to see why we
905 * don't need to release/reserve qgroup space.
906 */
907 trace_btrfs_space_reservation(fs_info, "delayed_item", ino,
908 total_reserved_size, 0);
909 btrfs_block_rsv_release(fs_info, &fs_info->delayed_block_rsv,
910 total_reserved_size, NULL);
911 }
912
913 list_for_each_entry_safe(curr, next, &batch_list, tree_list) {
914 list_del(&curr->tree_list);
915 btrfs_release_delayed_item(curr);
916 }
917
918 return 0;
919}
920
921static int btrfs_delete_delayed_items(struct btrfs_trans_handle *trans,
922 struct btrfs_path *path,
923 struct btrfs_root *root,
924 struct btrfs_delayed_node *node)
925{
926 struct btrfs_key key;
927 int ret = 0;
928
929 key.objectid = node->inode_id;
930 key.type = BTRFS_DIR_INDEX_KEY;
931
932 while (ret == 0) {
933 struct btrfs_delayed_item *item;
934
935 mutex_lock(&node->mutex);
936 item = __btrfs_first_delayed_deletion_item(node);
937 if (!item) {
938 mutex_unlock(&node->mutex);
939 break;
940 }
941
942 key.offset = item->index;
943 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
944 if (ret > 0) {
945 /*
946 * There's no matching item in the leaf. This means we
947 * have already deleted this item in a past run of the
948 * delayed items. We ignore errors when running delayed
949 * items from an async context, through a work queue job
950 * running btrfs_async_run_delayed_root(), and don't
951 * release delayed items that failed to complete. This
952 * is because we will retry later, and at transaction
953 * commit time we always run delayed items and will
954 * then deal with errors if they fail to run again.
955 *
956 * So just release delayed items for which we can't find
957 * an item in the tree, and move to the next item.
958 */
959 btrfs_release_path(path);
960 btrfs_release_delayed_item(item);
961 ret = 0;
962 } else if (ret == 0) {
963 ret = btrfs_batch_delete_items(trans, root, path, item);
964 btrfs_release_path(path);
965 }
966
967 /*
968 * We unlock and relock on each iteration, this is to prevent
969 * blocking other tasks for too long while we are being run from
970 * the async context (work queue job). Those tasks are typically
971 * running system calls like creat/mkdir/rename/unlink/etc which
972 * need to add delayed items to this delayed node.
973 */
974 mutex_unlock(&node->mutex);
975 }
976
977 return ret;
978}
979
980static void btrfs_release_delayed_inode(struct btrfs_delayed_node *delayed_node)
981{
982 struct btrfs_delayed_root *delayed_root;
983
984 if (delayed_node &&
985 test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
986 ASSERT(delayed_node->root);
987 clear_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags);
988 delayed_node->count--;
989
990 delayed_root = delayed_node->root->fs_info->delayed_root;
991 finish_one_item(delayed_root);
992 }
993}
994
995static void btrfs_release_delayed_iref(struct btrfs_delayed_node *delayed_node)
996{
997
998 if (test_and_clear_bit(BTRFS_DELAYED_NODE_DEL_IREF, &delayed_node->flags)) {
999 struct btrfs_delayed_root *delayed_root;
1000
1001 ASSERT(delayed_node->root);
1002 delayed_node->count--;
1003
1004 delayed_root = delayed_node->root->fs_info->delayed_root;
1005 finish_one_item(delayed_root);
1006 }
1007}
1008
1009static int __btrfs_update_delayed_inode(struct btrfs_trans_handle *trans,
1010 struct btrfs_root *root,
1011 struct btrfs_path *path,
1012 struct btrfs_delayed_node *node)
1013{
1014 struct btrfs_fs_info *fs_info = root->fs_info;
1015 struct btrfs_key key;
1016 struct btrfs_inode_item *inode_item;
1017 struct extent_buffer *leaf;
1018 int mod;
1019 int ret;
1020
1021 key.objectid = node->inode_id;
1022 key.type = BTRFS_INODE_ITEM_KEY;
1023 key.offset = 0;
1024
1025 if (test_bit(BTRFS_DELAYED_NODE_DEL_IREF, &node->flags))
1026 mod = -1;
1027 else
1028 mod = 1;
1029
1030 ret = btrfs_lookup_inode(trans, root, path, &key, mod);
1031 if (ret > 0)
1032 ret = -ENOENT;
1033 if (ret < 0)
1034 goto out;
1035
1036 leaf = path->nodes[0];
1037 inode_item = btrfs_item_ptr(leaf, path->slots[0],
1038 struct btrfs_inode_item);
1039 write_extent_buffer(leaf, &node->inode_item, (unsigned long)inode_item,
1040 sizeof(struct btrfs_inode_item));
1041 btrfs_mark_buffer_dirty(trans, leaf);
1042
1043 if (!test_bit(BTRFS_DELAYED_NODE_DEL_IREF, &node->flags))
1044 goto out;
1045
1046 /*
1047 * Now we're going to delete the INODE_REF/EXTREF, which should be the
1048 * only one ref left. Check if the next item is an INODE_REF/EXTREF.
1049 *
1050 * But if we're the last item already, release and search for the last
1051 * INODE_REF/EXTREF.
1052 */
1053 if (path->slots[0] + 1 >= btrfs_header_nritems(leaf)) {
1054 key.objectid = node->inode_id;
1055 key.type = BTRFS_INODE_EXTREF_KEY;
1056 key.offset = (u64)-1;
1057
1058 btrfs_release_path(path);
1059 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1060 if (ret < 0)
1061 goto err_out;
1062 ASSERT(ret > 0);
1063 ASSERT(path->slots[0] > 0);
1064 ret = 0;
1065 path->slots[0]--;
1066 leaf = path->nodes[0];
1067 } else {
1068 path->slots[0]++;
1069 }
1070 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1071 if (key.objectid != node->inode_id)
1072 goto out;
1073 if (key.type != BTRFS_INODE_REF_KEY &&
1074 key.type != BTRFS_INODE_EXTREF_KEY)
1075 goto out;
1076
1077 /*
1078 * Delayed iref deletion is for the inode who has only one link,
1079 * so there is only one iref. The case that several irefs are
1080 * in the same item doesn't exist.
1081 */
1082 ret = btrfs_del_item(trans, root, path);
1083out:
1084 btrfs_release_delayed_iref(node);
1085 btrfs_release_path(path);
1086err_out:
1087 btrfs_delayed_inode_release_metadata(fs_info, node, (ret < 0));
1088 btrfs_release_delayed_inode(node);
1089
1090 /*
1091 * If we fail to update the delayed inode we need to abort the
1092 * transaction, because we could leave the inode with the improper
1093 * counts behind.
1094 */
1095 if (ret && ret != -ENOENT)
1096 btrfs_abort_transaction(trans, ret);
1097
1098 return ret;
1099}
1100
1101static inline int btrfs_update_delayed_inode(struct btrfs_trans_handle *trans,
1102 struct btrfs_root *root,
1103 struct btrfs_path *path,
1104 struct btrfs_delayed_node *node)
1105{
1106 int ret;
1107
1108 mutex_lock(&node->mutex);
1109 if (!test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &node->flags)) {
1110 mutex_unlock(&node->mutex);
1111 return 0;
1112 }
1113
1114 ret = __btrfs_update_delayed_inode(trans, root, path, node);
1115 mutex_unlock(&node->mutex);
1116 return ret;
1117}
1118
1119static inline int
1120__btrfs_commit_inode_delayed_items(struct btrfs_trans_handle *trans,
1121 struct btrfs_path *path,
1122 struct btrfs_delayed_node *node)
1123{
1124 int ret;
1125
1126 ret = btrfs_insert_delayed_items(trans, path, node->root, node);
1127 if (ret)
1128 return ret;
1129
1130 ret = btrfs_delete_delayed_items(trans, path, node->root, node);
1131 if (ret)
1132 return ret;
1133
1134 ret = btrfs_record_root_in_trans(trans, node->root);
1135 if (ret)
1136 return ret;
1137 ret = btrfs_update_delayed_inode(trans, node->root, path, node);
1138 return ret;
1139}
1140
1141/*
1142 * Called when committing the transaction.
1143 * Returns 0 on success.
1144 * Returns < 0 on error and returns with an aborted transaction with any
1145 * outstanding delayed items cleaned up.
1146 */
1147static int __btrfs_run_delayed_items(struct btrfs_trans_handle *trans, int nr)
1148{
1149 struct btrfs_fs_info *fs_info = trans->fs_info;
1150 struct btrfs_delayed_root *delayed_root;
1151 struct btrfs_delayed_node *curr_node, *prev_node;
1152 struct btrfs_path *path;
1153 struct btrfs_block_rsv *block_rsv;
1154 int ret = 0;
1155 bool count = (nr > 0);
1156
1157 if (TRANS_ABORTED(trans))
1158 return -EIO;
1159
1160 path = btrfs_alloc_path();
1161 if (!path)
1162 return -ENOMEM;
1163
1164 block_rsv = trans->block_rsv;
1165 trans->block_rsv = &fs_info->delayed_block_rsv;
1166
1167 delayed_root = fs_info->delayed_root;
1168
1169 curr_node = btrfs_first_delayed_node(delayed_root);
1170 while (curr_node && (!count || nr--)) {
1171 ret = __btrfs_commit_inode_delayed_items(trans, path,
1172 curr_node);
1173 if (ret) {
1174 btrfs_abort_transaction(trans, ret);
1175 break;
1176 }
1177
1178 prev_node = curr_node;
1179 curr_node = btrfs_next_delayed_node(curr_node);
1180 /*
1181 * See the comment below about releasing path before releasing
1182 * node. If the commit of delayed items was successful the path
1183 * should always be released, but in case of an error, it may
1184 * point to locked extent buffers (a leaf at the very least).
1185 */
1186 ASSERT(path->nodes[0] == NULL);
1187 btrfs_release_delayed_node(prev_node);
1188 }
1189
1190 /*
1191 * Release the path to avoid a potential deadlock and lockdep splat when
1192 * releasing the delayed node, as that requires taking the delayed node's
1193 * mutex. If another task starts running delayed items before we take
1194 * the mutex, it will first lock the mutex and then it may try to lock
1195 * the same btree path (leaf).
1196 */
1197 btrfs_free_path(path);
1198
1199 if (curr_node)
1200 btrfs_release_delayed_node(curr_node);
1201 trans->block_rsv = block_rsv;
1202
1203 return ret;
1204}
1205
1206int btrfs_run_delayed_items(struct btrfs_trans_handle *trans)
1207{
1208 return __btrfs_run_delayed_items(trans, -1);
1209}
1210
1211int btrfs_run_delayed_items_nr(struct btrfs_trans_handle *trans, int nr)
1212{
1213 return __btrfs_run_delayed_items(trans, nr);
1214}
1215
1216int btrfs_commit_inode_delayed_items(struct btrfs_trans_handle *trans,
1217 struct btrfs_inode *inode)
1218{
1219 struct btrfs_delayed_node *delayed_node = btrfs_get_delayed_node(inode);
1220 struct btrfs_path *path;
1221 struct btrfs_block_rsv *block_rsv;
1222 int ret;
1223
1224 if (!delayed_node)
1225 return 0;
1226
1227 mutex_lock(&delayed_node->mutex);
1228 if (!delayed_node->count) {
1229 mutex_unlock(&delayed_node->mutex);
1230 btrfs_release_delayed_node(delayed_node);
1231 return 0;
1232 }
1233 mutex_unlock(&delayed_node->mutex);
1234
1235 path = btrfs_alloc_path();
1236 if (!path) {
1237 btrfs_release_delayed_node(delayed_node);
1238 return -ENOMEM;
1239 }
1240
1241 block_rsv = trans->block_rsv;
1242 trans->block_rsv = &delayed_node->root->fs_info->delayed_block_rsv;
1243
1244 ret = __btrfs_commit_inode_delayed_items(trans, path, delayed_node);
1245
1246 btrfs_release_delayed_node(delayed_node);
1247 btrfs_free_path(path);
1248 trans->block_rsv = block_rsv;
1249
1250 return ret;
1251}
1252
1253int btrfs_commit_inode_delayed_inode(struct btrfs_inode *inode)
1254{
1255 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1256 struct btrfs_trans_handle *trans;
1257 struct btrfs_delayed_node *delayed_node = btrfs_get_delayed_node(inode);
1258 struct btrfs_path *path;
1259 struct btrfs_block_rsv *block_rsv;
1260 int ret;
1261
1262 if (!delayed_node)
1263 return 0;
1264
1265 mutex_lock(&delayed_node->mutex);
1266 if (!test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
1267 mutex_unlock(&delayed_node->mutex);
1268 btrfs_release_delayed_node(delayed_node);
1269 return 0;
1270 }
1271 mutex_unlock(&delayed_node->mutex);
1272
1273 trans = btrfs_join_transaction(delayed_node->root);
1274 if (IS_ERR(trans)) {
1275 ret = PTR_ERR(trans);
1276 goto out;
1277 }
1278
1279 path = btrfs_alloc_path();
1280 if (!path) {
1281 ret = -ENOMEM;
1282 goto trans_out;
1283 }
1284
1285 block_rsv = trans->block_rsv;
1286 trans->block_rsv = &fs_info->delayed_block_rsv;
1287
1288 mutex_lock(&delayed_node->mutex);
1289 if (test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags))
1290 ret = __btrfs_update_delayed_inode(trans, delayed_node->root,
1291 path, delayed_node);
1292 else
1293 ret = 0;
1294 mutex_unlock(&delayed_node->mutex);
1295
1296 btrfs_free_path(path);
1297 trans->block_rsv = block_rsv;
1298trans_out:
1299 btrfs_end_transaction(trans);
1300 btrfs_btree_balance_dirty(fs_info);
1301out:
1302 btrfs_release_delayed_node(delayed_node);
1303
1304 return ret;
1305}
1306
1307void btrfs_remove_delayed_node(struct btrfs_inode *inode)
1308{
1309 struct btrfs_delayed_node *delayed_node;
1310
1311 delayed_node = READ_ONCE(inode->delayed_node);
1312 if (!delayed_node)
1313 return;
1314
1315 inode->delayed_node = NULL;
1316 btrfs_release_delayed_node(delayed_node);
1317}
1318
1319struct btrfs_async_delayed_work {
1320 struct btrfs_delayed_root *delayed_root;
1321 int nr;
1322 struct btrfs_work work;
1323};
1324
1325static void btrfs_async_run_delayed_root(struct btrfs_work *work)
1326{
1327 struct btrfs_async_delayed_work *async_work;
1328 struct btrfs_delayed_root *delayed_root;
1329 struct btrfs_trans_handle *trans;
1330 struct btrfs_path *path;
1331 struct btrfs_delayed_node *delayed_node = NULL;
1332 struct btrfs_root *root;
1333 struct btrfs_block_rsv *block_rsv;
1334 int total_done = 0;
1335
1336 async_work = container_of(work, struct btrfs_async_delayed_work, work);
1337 delayed_root = async_work->delayed_root;
1338
1339 path = btrfs_alloc_path();
1340 if (!path)
1341 goto out;
1342
1343 do {
1344 if (atomic_read(&delayed_root->items) <
1345 BTRFS_DELAYED_BACKGROUND / 2)
1346 break;
1347
1348 delayed_node = btrfs_first_prepared_delayed_node(delayed_root);
1349 if (!delayed_node)
1350 break;
1351
1352 root = delayed_node->root;
1353
1354 trans = btrfs_join_transaction(root);
1355 if (IS_ERR(trans)) {
1356 btrfs_release_path(path);
1357 btrfs_release_prepared_delayed_node(delayed_node);
1358 total_done++;
1359 continue;
1360 }
1361
1362 block_rsv = trans->block_rsv;
1363 trans->block_rsv = &root->fs_info->delayed_block_rsv;
1364
1365 __btrfs_commit_inode_delayed_items(trans, path, delayed_node);
1366
1367 trans->block_rsv = block_rsv;
1368 btrfs_end_transaction(trans);
1369 btrfs_btree_balance_dirty_nodelay(root->fs_info);
1370
1371 btrfs_release_path(path);
1372 btrfs_release_prepared_delayed_node(delayed_node);
1373 total_done++;
1374
1375 } while ((async_work->nr == 0 && total_done < BTRFS_DELAYED_WRITEBACK)
1376 || total_done < async_work->nr);
1377
1378 btrfs_free_path(path);
1379out:
1380 wake_up(&delayed_root->wait);
1381 kfree(async_work);
1382}
1383
1384
1385static int btrfs_wq_run_delayed_node(struct btrfs_delayed_root *delayed_root,
1386 struct btrfs_fs_info *fs_info, int nr)
1387{
1388 struct btrfs_async_delayed_work *async_work;
1389
1390 async_work = kmalloc(sizeof(*async_work), GFP_NOFS);
1391 if (!async_work)
1392 return -ENOMEM;
1393
1394 async_work->delayed_root = delayed_root;
1395 btrfs_init_work(&async_work->work, btrfs_async_run_delayed_root, NULL);
1396 async_work->nr = nr;
1397
1398 btrfs_queue_work(fs_info->delayed_workers, &async_work->work);
1399 return 0;
1400}
1401
1402void btrfs_assert_delayed_root_empty(struct btrfs_fs_info *fs_info)
1403{
1404 WARN_ON(btrfs_first_delayed_node(fs_info->delayed_root));
1405}
1406
1407static int could_end_wait(struct btrfs_delayed_root *delayed_root, int seq)
1408{
1409 int val = atomic_read(&delayed_root->items_seq);
1410
1411 if (val < seq || val >= seq + BTRFS_DELAYED_BATCH)
1412 return 1;
1413
1414 if (atomic_read(&delayed_root->items) < BTRFS_DELAYED_BACKGROUND)
1415 return 1;
1416
1417 return 0;
1418}
1419
1420void btrfs_balance_delayed_items(struct btrfs_fs_info *fs_info)
1421{
1422 struct btrfs_delayed_root *delayed_root = fs_info->delayed_root;
1423
1424 if ((atomic_read(&delayed_root->items) < BTRFS_DELAYED_BACKGROUND) ||
1425 btrfs_workqueue_normal_congested(fs_info->delayed_workers))
1426 return;
1427
1428 if (atomic_read(&delayed_root->items) >= BTRFS_DELAYED_WRITEBACK) {
1429 int seq;
1430 int ret;
1431
1432 seq = atomic_read(&delayed_root->items_seq);
1433
1434 ret = btrfs_wq_run_delayed_node(delayed_root, fs_info, 0);
1435 if (ret)
1436 return;
1437
1438 wait_event_interruptible(delayed_root->wait,
1439 could_end_wait(delayed_root, seq));
1440 return;
1441 }
1442
1443 btrfs_wq_run_delayed_node(delayed_root, fs_info, BTRFS_DELAYED_BATCH);
1444}
1445
1446static void btrfs_release_dir_index_item_space(struct btrfs_trans_handle *trans)
1447{
1448 struct btrfs_fs_info *fs_info = trans->fs_info;
1449 const u64 bytes = btrfs_calc_insert_metadata_size(fs_info, 1);
1450
1451 if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
1452 return;
1453
1454 /*
1455 * Adding the new dir index item does not require touching another
1456 * leaf, so we can release 1 unit of metadata that was previously
1457 * reserved when starting the transaction. This applies only to
1458 * the case where we had a transaction start and excludes the
1459 * transaction join case (when replaying log trees).
1460 */
1461 trace_btrfs_space_reservation(fs_info, "transaction",
1462 trans->transid, bytes, 0);
1463 btrfs_block_rsv_release(fs_info, trans->block_rsv, bytes, NULL);
1464 ASSERT(trans->bytes_reserved >= bytes);
1465 trans->bytes_reserved -= bytes;
1466}
1467
1468/* Will return 0, -ENOMEM or -EEXIST (index number collision, unexpected). */
1469int btrfs_insert_delayed_dir_index(struct btrfs_trans_handle *trans,
1470 const char *name, int name_len,
1471 struct btrfs_inode *dir,
1472 const struct btrfs_disk_key *disk_key, u8 flags,
1473 u64 index)
1474{
1475 struct btrfs_fs_info *fs_info = trans->fs_info;
1476 const unsigned int leaf_data_size = BTRFS_LEAF_DATA_SIZE(fs_info);
1477 struct btrfs_delayed_node *delayed_node;
1478 struct btrfs_delayed_item *delayed_item;
1479 struct btrfs_dir_item *dir_item;
1480 bool reserve_leaf_space;
1481 u32 data_len;
1482 int ret;
1483
1484 delayed_node = btrfs_get_or_create_delayed_node(dir);
1485 if (IS_ERR(delayed_node))
1486 return PTR_ERR(delayed_node);
1487
1488 delayed_item = btrfs_alloc_delayed_item(sizeof(*dir_item) + name_len,
1489 delayed_node,
1490 BTRFS_DELAYED_INSERTION_ITEM);
1491 if (!delayed_item) {
1492 ret = -ENOMEM;
1493 goto release_node;
1494 }
1495
1496 delayed_item->index = index;
1497
1498 dir_item = (struct btrfs_dir_item *)delayed_item->data;
1499 dir_item->location = *disk_key;
1500 btrfs_set_stack_dir_transid(dir_item, trans->transid);
1501 btrfs_set_stack_dir_data_len(dir_item, 0);
1502 btrfs_set_stack_dir_name_len(dir_item, name_len);
1503 btrfs_set_stack_dir_flags(dir_item, flags);
1504 memcpy((char *)(dir_item + 1), name, name_len);
1505
1506 data_len = delayed_item->data_len + sizeof(struct btrfs_item);
1507
1508 mutex_lock(&delayed_node->mutex);
1509
1510 /*
1511 * First attempt to insert the delayed item. This is to make the error
1512 * handling path simpler in case we fail (-EEXIST). There's no risk of
1513 * any other task coming in and running the delayed item before we do
1514 * the metadata space reservation below, because we are holding the
1515 * delayed node's mutex and that mutex must also be locked before the
1516 * node's delayed items can be run.
1517 */
1518 ret = __btrfs_add_delayed_item(delayed_node, delayed_item);
1519 if (unlikely(ret)) {
1520 btrfs_err(trans->fs_info,
1521"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",
1522 name_len, name, index, btrfs_root_id(delayed_node->root),
1523 delayed_node->inode_id, dir->index_cnt,
1524 delayed_node->index_cnt, ret);
1525 btrfs_release_delayed_item(delayed_item);
1526 btrfs_release_dir_index_item_space(trans);
1527 mutex_unlock(&delayed_node->mutex);
1528 goto release_node;
1529 }
1530
1531 if (delayed_node->index_item_leaves == 0 ||
1532 delayed_node->curr_index_batch_size + data_len > leaf_data_size) {
1533 delayed_node->curr_index_batch_size = data_len;
1534 reserve_leaf_space = true;
1535 } else {
1536 delayed_node->curr_index_batch_size += data_len;
1537 reserve_leaf_space = false;
1538 }
1539
1540 if (reserve_leaf_space) {
1541 ret = btrfs_delayed_item_reserve_metadata(trans, delayed_item);
1542 /*
1543 * Space was reserved for a dir index item insertion when we
1544 * started the transaction, so getting a failure here should be
1545 * impossible.
1546 */
1547 if (WARN_ON(ret)) {
1548 btrfs_release_delayed_item(delayed_item);
1549 mutex_unlock(&delayed_node->mutex);
1550 goto release_node;
1551 }
1552
1553 delayed_node->index_item_leaves++;
1554 } else {
1555 btrfs_release_dir_index_item_space(trans);
1556 }
1557 mutex_unlock(&delayed_node->mutex);
1558
1559release_node:
1560 btrfs_release_delayed_node(delayed_node);
1561 return ret;
1562}
1563
1564static int btrfs_delete_delayed_insertion_item(struct btrfs_fs_info *fs_info,
1565 struct btrfs_delayed_node *node,
1566 u64 index)
1567{
1568 struct btrfs_delayed_item *item;
1569
1570 mutex_lock(&node->mutex);
1571 item = __btrfs_lookup_delayed_item(&node->ins_root.rb_root, index);
1572 if (!item) {
1573 mutex_unlock(&node->mutex);
1574 return 1;
1575 }
1576
1577 /*
1578 * For delayed items to insert, we track reserved metadata bytes based
1579 * on the number of leaves that we will use.
1580 * See btrfs_insert_delayed_dir_index() and
1581 * btrfs_delayed_item_reserve_metadata()).
1582 */
1583 ASSERT(item->bytes_reserved == 0);
1584 ASSERT(node->index_item_leaves > 0);
1585
1586 /*
1587 * If there's only one leaf reserved, we can decrement this item from the
1588 * current batch, otherwise we can not because we don't know which leaf
1589 * it belongs to. With the current limit on delayed items, we rarely
1590 * accumulate enough dir index items to fill more than one leaf (even
1591 * when using a leaf size of 4K).
1592 */
1593 if (node->index_item_leaves == 1) {
1594 const u32 data_len = item->data_len + sizeof(struct btrfs_item);
1595
1596 ASSERT(node->curr_index_batch_size >= data_len);
1597 node->curr_index_batch_size -= data_len;
1598 }
1599
1600 btrfs_release_delayed_item(item);
1601
1602 /* If we now have no more dir index items, we can release all leaves. */
1603 if (RB_EMPTY_ROOT(&node->ins_root.rb_root)) {
1604 btrfs_delayed_item_release_leaves(node, node->index_item_leaves);
1605 node->index_item_leaves = 0;
1606 }
1607
1608 mutex_unlock(&node->mutex);
1609 return 0;
1610}
1611
1612int btrfs_delete_delayed_dir_index(struct btrfs_trans_handle *trans,
1613 struct btrfs_inode *dir, u64 index)
1614{
1615 struct btrfs_delayed_node *node;
1616 struct btrfs_delayed_item *item;
1617 int ret;
1618
1619 node = btrfs_get_or_create_delayed_node(dir);
1620 if (IS_ERR(node))
1621 return PTR_ERR(node);
1622
1623 ret = btrfs_delete_delayed_insertion_item(trans->fs_info, node, index);
1624 if (!ret)
1625 goto end;
1626
1627 item = btrfs_alloc_delayed_item(0, node, BTRFS_DELAYED_DELETION_ITEM);
1628 if (!item) {
1629 ret = -ENOMEM;
1630 goto end;
1631 }
1632
1633 item->index = index;
1634
1635 ret = btrfs_delayed_item_reserve_metadata(trans, item);
1636 /*
1637 * we have reserved enough space when we start a new transaction,
1638 * so reserving metadata failure is impossible.
1639 */
1640 if (ret < 0) {
1641 btrfs_err(trans->fs_info,
1642"metadata reservation failed for delayed dir item deltiona, should have been reserved");
1643 btrfs_release_delayed_item(item);
1644 goto end;
1645 }
1646
1647 mutex_lock(&node->mutex);
1648 ret = __btrfs_add_delayed_item(node, item);
1649 if (unlikely(ret)) {
1650 btrfs_err(trans->fs_info,
1651 "err add delayed dir index item(index: %llu) into the deletion tree of the delayed node(root id: %llu, inode id: %llu, errno: %d)",
1652 index, btrfs_root_id(node->root),
1653 node->inode_id, ret);
1654 btrfs_delayed_item_release_metadata(dir->root, item);
1655 btrfs_release_delayed_item(item);
1656 }
1657 mutex_unlock(&node->mutex);
1658end:
1659 btrfs_release_delayed_node(node);
1660 return ret;
1661}
1662
1663int btrfs_inode_delayed_dir_index_count(struct btrfs_inode *inode)
1664{
1665 struct btrfs_delayed_node *delayed_node = btrfs_get_delayed_node(inode);
1666
1667 if (!delayed_node)
1668 return -ENOENT;
1669
1670 /*
1671 * Since we have held i_mutex of this directory, it is impossible that
1672 * a new directory index is added into the delayed node and index_cnt
1673 * is updated now. So we needn't lock the delayed node.
1674 */
1675 if (!delayed_node->index_cnt) {
1676 btrfs_release_delayed_node(delayed_node);
1677 return -EINVAL;
1678 }
1679
1680 inode->index_cnt = delayed_node->index_cnt;
1681 btrfs_release_delayed_node(delayed_node);
1682 return 0;
1683}
1684
1685bool btrfs_readdir_get_delayed_items(struct btrfs_inode *inode,
1686 u64 last_index,
1687 struct list_head *ins_list,
1688 struct list_head *del_list)
1689{
1690 struct btrfs_delayed_node *delayed_node;
1691 struct btrfs_delayed_item *item;
1692
1693 delayed_node = btrfs_get_delayed_node(inode);
1694 if (!delayed_node)
1695 return false;
1696
1697 /*
1698 * We can only do one readdir with delayed items at a time because of
1699 * item->readdir_list.
1700 */
1701 btrfs_inode_unlock(inode, BTRFS_ILOCK_SHARED);
1702 btrfs_inode_lock(inode, 0);
1703
1704 mutex_lock(&delayed_node->mutex);
1705 item = __btrfs_first_delayed_insertion_item(delayed_node);
1706 while (item && item->index <= last_index) {
1707 refcount_inc(&item->refs);
1708 list_add_tail(&item->readdir_list, ins_list);
1709 item = __btrfs_next_delayed_item(item);
1710 }
1711
1712 item = __btrfs_first_delayed_deletion_item(delayed_node);
1713 while (item && item->index <= last_index) {
1714 refcount_inc(&item->refs);
1715 list_add_tail(&item->readdir_list, del_list);
1716 item = __btrfs_next_delayed_item(item);
1717 }
1718 mutex_unlock(&delayed_node->mutex);
1719 /*
1720 * This delayed node is still cached in the btrfs inode, so refs
1721 * must be > 1 now, and we needn't check it is going to be freed
1722 * or not.
1723 *
1724 * Besides that, this function is used to read dir, we do not
1725 * insert/delete delayed items in this period. So we also needn't
1726 * requeue or dequeue this delayed node.
1727 */
1728 refcount_dec(&delayed_node->refs);
1729
1730 return true;
1731}
1732
1733void btrfs_readdir_put_delayed_items(struct btrfs_inode *inode,
1734 struct list_head *ins_list,
1735 struct list_head *del_list)
1736{
1737 struct btrfs_delayed_item *curr, *next;
1738
1739 list_for_each_entry_safe(curr, next, ins_list, readdir_list) {
1740 list_del(&curr->readdir_list);
1741 if (refcount_dec_and_test(&curr->refs))
1742 kfree(curr);
1743 }
1744
1745 list_for_each_entry_safe(curr, next, del_list, readdir_list) {
1746 list_del(&curr->readdir_list);
1747 if (refcount_dec_and_test(&curr->refs))
1748 kfree(curr);
1749 }
1750
1751 /*
1752 * The VFS is going to do up_read(), so we need to downgrade back to a
1753 * read lock.
1754 */
1755 downgrade_write(&inode->vfs_inode.i_rwsem);
1756}
1757
1758int btrfs_should_delete_dir_index(const struct list_head *del_list,
1759 u64 index)
1760{
1761 struct btrfs_delayed_item *curr;
1762 int ret = 0;
1763
1764 list_for_each_entry(curr, del_list, readdir_list) {
1765 if (curr->index > index)
1766 break;
1767 if (curr->index == index) {
1768 ret = 1;
1769 break;
1770 }
1771 }
1772 return ret;
1773}
1774
1775/*
1776 * Read dir info stored in the delayed tree.
1777 */
1778int btrfs_readdir_delayed_dir_index(struct dir_context *ctx,
1779 const struct list_head *ins_list)
1780{
1781 struct btrfs_dir_item *di;
1782 struct btrfs_delayed_item *curr, *next;
1783 struct btrfs_key location;
1784 char *name;
1785 int name_len;
1786 int over = 0;
1787 unsigned char d_type;
1788
1789 /*
1790 * Changing the data of the delayed item is impossible. So
1791 * we needn't lock them. And we have held i_mutex of the
1792 * directory, nobody can delete any directory indexes now.
1793 */
1794 list_for_each_entry_safe(curr, next, ins_list, readdir_list) {
1795 list_del(&curr->readdir_list);
1796
1797 if (curr->index < ctx->pos) {
1798 if (refcount_dec_and_test(&curr->refs))
1799 kfree(curr);
1800 continue;
1801 }
1802
1803 ctx->pos = curr->index;
1804
1805 di = (struct btrfs_dir_item *)curr->data;
1806 name = (char *)(di + 1);
1807 name_len = btrfs_stack_dir_name_len(di);
1808
1809 d_type = fs_ftype_to_dtype(btrfs_dir_flags_to_ftype(di->type));
1810 btrfs_disk_key_to_cpu(&location, &di->location);
1811
1812 over = !dir_emit(ctx, name, name_len,
1813 location.objectid, d_type);
1814
1815 if (refcount_dec_and_test(&curr->refs))
1816 kfree(curr);
1817
1818 if (over)
1819 return 1;
1820 ctx->pos++;
1821 }
1822 return 0;
1823}
1824
1825static void fill_stack_inode_item(struct btrfs_trans_handle *trans,
1826 struct btrfs_inode_item *inode_item,
1827 struct inode *inode)
1828{
1829 u64 flags;
1830
1831 btrfs_set_stack_inode_uid(inode_item, i_uid_read(inode));
1832 btrfs_set_stack_inode_gid(inode_item, i_gid_read(inode));
1833 btrfs_set_stack_inode_size(inode_item, BTRFS_I(inode)->disk_i_size);
1834 btrfs_set_stack_inode_mode(inode_item, inode->i_mode);
1835 btrfs_set_stack_inode_nlink(inode_item, inode->i_nlink);
1836 btrfs_set_stack_inode_nbytes(inode_item, inode_get_bytes(inode));
1837 btrfs_set_stack_inode_generation(inode_item,
1838 BTRFS_I(inode)->generation);
1839 btrfs_set_stack_inode_sequence(inode_item,
1840 inode_peek_iversion(inode));
1841 btrfs_set_stack_inode_transid(inode_item, trans->transid);
1842 btrfs_set_stack_inode_rdev(inode_item, inode->i_rdev);
1843 flags = btrfs_inode_combine_flags(BTRFS_I(inode)->flags,
1844 BTRFS_I(inode)->ro_flags);
1845 btrfs_set_stack_inode_flags(inode_item, flags);
1846 btrfs_set_stack_inode_block_group(inode_item, 0);
1847
1848 btrfs_set_stack_timespec_sec(&inode_item->atime,
1849 inode_get_atime_sec(inode));
1850 btrfs_set_stack_timespec_nsec(&inode_item->atime,
1851 inode_get_atime_nsec(inode));
1852
1853 btrfs_set_stack_timespec_sec(&inode_item->mtime,
1854 inode_get_mtime_sec(inode));
1855 btrfs_set_stack_timespec_nsec(&inode_item->mtime,
1856 inode_get_mtime_nsec(inode));
1857
1858 btrfs_set_stack_timespec_sec(&inode_item->ctime,
1859 inode_get_ctime_sec(inode));
1860 btrfs_set_stack_timespec_nsec(&inode_item->ctime,
1861 inode_get_ctime_nsec(inode));
1862
1863 btrfs_set_stack_timespec_sec(&inode_item->otime, BTRFS_I(inode)->i_otime_sec);
1864 btrfs_set_stack_timespec_nsec(&inode_item->otime, BTRFS_I(inode)->i_otime_nsec);
1865}
1866
1867int btrfs_fill_inode(struct inode *inode, u32 *rdev)
1868{
1869 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
1870 struct btrfs_delayed_node *delayed_node;
1871 struct btrfs_inode_item *inode_item;
1872
1873 delayed_node = btrfs_get_delayed_node(BTRFS_I(inode));
1874 if (!delayed_node)
1875 return -ENOENT;
1876
1877 mutex_lock(&delayed_node->mutex);
1878 if (!test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
1879 mutex_unlock(&delayed_node->mutex);
1880 btrfs_release_delayed_node(delayed_node);
1881 return -ENOENT;
1882 }
1883
1884 inode_item = &delayed_node->inode_item;
1885
1886 i_uid_write(inode, btrfs_stack_inode_uid(inode_item));
1887 i_gid_write(inode, btrfs_stack_inode_gid(inode_item));
1888 btrfs_i_size_write(BTRFS_I(inode), btrfs_stack_inode_size(inode_item));
1889 btrfs_inode_set_file_extent_range(BTRFS_I(inode), 0,
1890 round_up(i_size_read(inode), fs_info->sectorsize));
1891 inode->i_mode = btrfs_stack_inode_mode(inode_item);
1892 set_nlink(inode, btrfs_stack_inode_nlink(inode_item));
1893 inode_set_bytes(inode, btrfs_stack_inode_nbytes(inode_item));
1894 BTRFS_I(inode)->generation = btrfs_stack_inode_generation(inode_item);
1895 BTRFS_I(inode)->last_trans = btrfs_stack_inode_transid(inode_item);
1896
1897 inode_set_iversion_queried(inode,
1898 btrfs_stack_inode_sequence(inode_item));
1899 inode->i_rdev = 0;
1900 *rdev = btrfs_stack_inode_rdev(inode_item);
1901 btrfs_inode_split_flags(btrfs_stack_inode_flags(inode_item),
1902 &BTRFS_I(inode)->flags, &BTRFS_I(inode)->ro_flags);
1903
1904 inode_set_atime(inode, btrfs_stack_timespec_sec(&inode_item->atime),
1905 btrfs_stack_timespec_nsec(&inode_item->atime));
1906
1907 inode_set_mtime(inode, btrfs_stack_timespec_sec(&inode_item->mtime),
1908 btrfs_stack_timespec_nsec(&inode_item->mtime));
1909
1910 inode_set_ctime(inode, btrfs_stack_timespec_sec(&inode_item->ctime),
1911 btrfs_stack_timespec_nsec(&inode_item->ctime));
1912
1913 BTRFS_I(inode)->i_otime_sec = btrfs_stack_timespec_sec(&inode_item->otime);
1914 BTRFS_I(inode)->i_otime_nsec = btrfs_stack_timespec_nsec(&inode_item->otime);
1915
1916 inode->i_generation = BTRFS_I(inode)->generation;
1917 if (S_ISDIR(inode->i_mode))
1918 BTRFS_I(inode)->index_cnt = (u64)-1;
1919
1920 mutex_unlock(&delayed_node->mutex);
1921 btrfs_release_delayed_node(delayed_node);
1922 return 0;
1923}
1924
1925int btrfs_delayed_update_inode(struct btrfs_trans_handle *trans,
1926 struct btrfs_inode *inode)
1927{
1928 struct btrfs_root *root = inode->root;
1929 struct btrfs_delayed_node *delayed_node;
1930 int ret = 0;
1931
1932 delayed_node = btrfs_get_or_create_delayed_node(inode);
1933 if (IS_ERR(delayed_node))
1934 return PTR_ERR(delayed_node);
1935
1936 mutex_lock(&delayed_node->mutex);
1937 if (test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
1938 fill_stack_inode_item(trans, &delayed_node->inode_item,
1939 &inode->vfs_inode);
1940 goto release_node;
1941 }
1942
1943 ret = btrfs_delayed_inode_reserve_metadata(trans, root, delayed_node);
1944 if (ret)
1945 goto release_node;
1946
1947 fill_stack_inode_item(trans, &delayed_node->inode_item, &inode->vfs_inode);
1948 set_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags);
1949 delayed_node->count++;
1950 atomic_inc(&root->fs_info->delayed_root->items);
1951release_node:
1952 mutex_unlock(&delayed_node->mutex);
1953 btrfs_release_delayed_node(delayed_node);
1954 return ret;
1955}
1956
1957int btrfs_delayed_delete_inode_ref(struct btrfs_inode *inode)
1958{
1959 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1960 struct btrfs_delayed_node *delayed_node;
1961
1962 /*
1963 * we don't do delayed inode updates during log recovery because it
1964 * leads to enospc problems. This means we also can't do
1965 * delayed inode refs
1966 */
1967 if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
1968 return -EAGAIN;
1969
1970 delayed_node = btrfs_get_or_create_delayed_node(inode);
1971 if (IS_ERR(delayed_node))
1972 return PTR_ERR(delayed_node);
1973
1974 /*
1975 * We don't reserve space for inode ref deletion is because:
1976 * - We ONLY do async inode ref deletion for the inode who has only
1977 * one link(i_nlink == 1), it means there is only one inode ref.
1978 * And in most case, the inode ref and the inode item are in the
1979 * same leaf, and we will deal with them at the same time.
1980 * Since we are sure we will reserve the space for the inode item,
1981 * it is unnecessary to reserve space for inode ref deletion.
1982 * - If the inode ref and the inode item are not in the same leaf,
1983 * We also needn't worry about enospc problem, because we reserve
1984 * much more space for the inode update than it needs.
1985 * - At the worst, we can steal some space from the global reservation.
1986 * It is very rare.
1987 */
1988 mutex_lock(&delayed_node->mutex);
1989 if (test_bit(BTRFS_DELAYED_NODE_DEL_IREF, &delayed_node->flags))
1990 goto release_node;
1991
1992 set_bit(BTRFS_DELAYED_NODE_DEL_IREF, &delayed_node->flags);
1993 delayed_node->count++;
1994 atomic_inc(&fs_info->delayed_root->items);
1995release_node:
1996 mutex_unlock(&delayed_node->mutex);
1997 btrfs_release_delayed_node(delayed_node);
1998 return 0;
1999}
2000
2001static void __btrfs_kill_delayed_node(struct btrfs_delayed_node *delayed_node)
2002{
2003 struct btrfs_root *root = delayed_node->root;
2004 struct btrfs_fs_info *fs_info = root->fs_info;
2005 struct btrfs_delayed_item *curr_item, *prev_item;
2006
2007 mutex_lock(&delayed_node->mutex);
2008 curr_item = __btrfs_first_delayed_insertion_item(delayed_node);
2009 while (curr_item) {
2010 prev_item = curr_item;
2011 curr_item = __btrfs_next_delayed_item(prev_item);
2012 btrfs_release_delayed_item(prev_item);
2013 }
2014
2015 if (delayed_node->index_item_leaves > 0) {
2016 btrfs_delayed_item_release_leaves(delayed_node,
2017 delayed_node->index_item_leaves);
2018 delayed_node->index_item_leaves = 0;
2019 }
2020
2021 curr_item = __btrfs_first_delayed_deletion_item(delayed_node);
2022 while (curr_item) {
2023 btrfs_delayed_item_release_metadata(root, curr_item);
2024 prev_item = curr_item;
2025 curr_item = __btrfs_next_delayed_item(prev_item);
2026 btrfs_release_delayed_item(prev_item);
2027 }
2028
2029 btrfs_release_delayed_iref(delayed_node);
2030
2031 if (test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
2032 btrfs_delayed_inode_release_metadata(fs_info, delayed_node, false);
2033 btrfs_release_delayed_inode(delayed_node);
2034 }
2035 mutex_unlock(&delayed_node->mutex);
2036}
2037
2038void btrfs_kill_delayed_inode_items(struct btrfs_inode *inode)
2039{
2040 struct btrfs_delayed_node *delayed_node;
2041
2042 delayed_node = btrfs_get_delayed_node(inode);
2043 if (!delayed_node)
2044 return;
2045
2046 __btrfs_kill_delayed_node(delayed_node);
2047 btrfs_release_delayed_node(delayed_node);
2048}
2049
2050void btrfs_kill_all_delayed_nodes(struct btrfs_root *root)
2051{
2052 unsigned long index = 0;
2053 struct btrfs_delayed_node *delayed_nodes[8];
2054
2055 while (1) {
2056 struct btrfs_delayed_node *node;
2057 int count;
2058
2059 xa_lock(&root->delayed_nodes);
2060 if (xa_empty(&root->delayed_nodes)) {
2061 xa_unlock(&root->delayed_nodes);
2062 return;
2063 }
2064
2065 count = 0;
2066 xa_for_each_start(&root->delayed_nodes, index, node, index) {
2067 /*
2068 * Don't increase refs in case the node is dead and
2069 * about to be removed from the tree in the loop below
2070 */
2071 if (refcount_inc_not_zero(&node->refs)) {
2072 delayed_nodes[count] = node;
2073 count++;
2074 }
2075 if (count >= ARRAY_SIZE(delayed_nodes))
2076 break;
2077 }
2078 xa_unlock(&root->delayed_nodes);
2079 index++;
2080
2081 for (int i = 0; i < count; i++) {
2082 __btrfs_kill_delayed_node(delayed_nodes[i]);
2083 btrfs_release_delayed_node(delayed_nodes[i]);
2084 }
2085 }
2086}
2087
2088void btrfs_destroy_delayed_inodes(struct btrfs_fs_info *fs_info)
2089{
2090 struct btrfs_delayed_node *curr_node, *prev_node;
2091
2092 curr_node = btrfs_first_delayed_node(fs_info->delayed_root);
2093 while (curr_node) {
2094 __btrfs_kill_delayed_node(curr_node);
2095
2096 prev_node = curr_node;
2097 curr_node = btrfs_next_delayed_node(curr_node);
2098 btrfs_release_delayed_node(prev_node);
2099 }
2100}
2101
2102void btrfs_log_get_delayed_items(struct btrfs_inode *inode,
2103 struct list_head *ins_list,
2104 struct list_head *del_list)
2105{
2106 struct btrfs_delayed_node *node;
2107 struct btrfs_delayed_item *item;
2108
2109 node = btrfs_get_delayed_node(inode);
2110 if (!node)
2111 return;
2112
2113 mutex_lock(&node->mutex);
2114 item = __btrfs_first_delayed_insertion_item(node);
2115 while (item) {
2116 /*
2117 * It's possible that the item is already in a log list. This
2118 * can happen in case two tasks are trying to log the same
2119 * directory. For example if we have tasks A and task B:
2120 *
2121 * Task A collected the delayed items into a log list while
2122 * under the inode's log_mutex (at btrfs_log_inode()), but it
2123 * only releases the items after logging the inodes they point
2124 * to (if they are new inodes), which happens after unlocking
2125 * the log mutex;
2126 *
2127 * Task B enters btrfs_log_inode() and acquires the log_mutex
2128 * of the same directory inode, before task B releases the
2129 * delayed items. This can happen for example when logging some
2130 * inode we need to trigger logging of its parent directory, so
2131 * logging two files that have the same parent directory can
2132 * lead to this.
2133 *
2134 * If this happens, just ignore delayed items already in a log
2135 * list. All the tasks logging the directory are under a log
2136 * transaction and whichever finishes first can not sync the log
2137 * before the other completes and leaves the log transaction.
2138 */
2139 if (!item->logged && list_empty(&item->log_list)) {
2140 refcount_inc(&item->refs);
2141 list_add_tail(&item->log_list, ins_list);
2142 }
2143 item = __btrfs_next_delayed_item(item);
2144 }
2145
2146 item = __btrfs_first_delayed_deletion_item(node);
2147 while (item) {
2148 /* It may be non-empty, for the same reason mentioned above. */
2149 if (!item->logged && list_empty(&item->log_list)) {
2150 refcount_inc(&item->refs);
2151 list_add_tail(&item->log_list, del_list);
2152 }
2153 item = __btrfs_next_delayed_item(item);
2154 }
2155 mutex_unlock(&node->mutex);
2156
2157 /*
2158 * We are called during inode logging, which means the inode is in use
2159 * and can not be evicted before we finish logging the inode. So we never
2160 * have the last reference on the delayed inode.
2161 * Also, we don't use btrfs_release_delayed_node() because that would
2162 * requeue the delayed inode (change its order in the list of prepared
2163 * nodes) and we don't want to do such change because we don't create or
2164 * delete delayed items.
2165 */
2166 ASSERT(refcount_read(&node->refs) > 1);
2167 refcount_dec(&node->refs);
2168}
2169
2170void btrfs_log_put_delayed_items(struct btrfs_inode *inode,
2171 struct list_head *ins_list,
2172 struct list_head *del_list)
2173{
2174 struct btrfs_delayed_node *node;
2175 struct btrfs_delayed_item *item;
2176 struct btrfs_delayed_item *next;
2177
2178 node = btrfs_get_delayed_node(inode);
2179 if (!node)
2180 return;
2181
2182 mutex_lock(&node->mutex);
2183
2184 list_for_each_entry_safe(item, next, ins_list, log_list) {
2185 item->logged = true;
2186 list_del_init(&item->log_list);
2187 if (refcount_dec_and_test(&item->refs))
2188 kfree(item);
2189 }
2190
2191 list_for_each_entry_safe(item, next, del_list, log_list) {
2192 item->logged = true;
2193 list_del_init(&item->log_list);
2194 if (refcount_dec_and_test(&item->refs))
2195 kfree(item);
2196 }
2197
2198 mutex_unlock(&node->mutex);
2199
2200 /*
2201 * We are called during inode logging, which means the inode is in use
2202 * and can not be evicted before we finish logging the inode. So we never
2203 * have the last reference on the delayed inode.
2204 * Also, we don't use btrfs_release_delayed_node() because that would
2205 * requeue the delayed inode (change its order in the list of prepared
2206 * nodes) and we don't want to do such change because we don't create or
2207 * delete delayed items.
2208 */
2209 ASSERT(refcount_read(&node->refs) > 1);
2210 refcount_dec(&node->refs);
2211}