Loading...
1/*
2 * Copyright (C) 2011 Fujitsu. All rights reserved.
3 * Written by Miao Xie <miaox@cn.fujitsu.com>
4 *
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public
7 * License v2 as published by the Free Software Foundation.
8 *
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
12 * General Public License for more details.
13 *
14 * You should have received a copy of the GNU General Public
15 * License along with this program; if not, write to the
16 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
17 * Boston, MA 021110-1307, USA.
18 */
19
20#include <linux/slab.h>
21#include "delayed-inode.h"
22#include "disk-io.h"
23#include "transaction.h"
24#include "ctree.h"
25
26#define BTRFS_DELAYED_WRITEBACK 512
27#define BTRFS_DELAYED_BACKGROUND 128
28#define BTRFS_DELAYED_BATCH 16
29
30static struct kmem_cache *delayed_node_cache;
31
32int __init btrfs_delayed_inode_init(void)
33{
34 delayed_node_cache = kmem_cache_create("btrfs_delayed_node",
35 sizeof(struct btrfs_delayed_node),
36 0,
37 SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD,
38 NULL);
39 if (!delayed_node_cache)
40 return -ENOMEM;
41 return 0;
42}
43
44void btrfs_delayed_inode_exit(void)
45{
46 if (delayed_node_cache)
47 kmem_cache_destroy(delayed_node_cache);
48}
49
50static inline void btrfs_init_delayed_node(
51 struct btrfs_delayed_node *delayed_node,
52 struct btrfs_root *root, u64 inode_id)
53{
54 delayed_node->root = root;
55 delayed_node->inode_id = inode_id;
56 atomic_set(&delayed_node->refs, 0);
57 delayed_node->count = 0;
58 delayed_node->flags = 0;
59 delayed_node->ins_root = RB_ROOT;
60 delayed_node->del_root = RB_ROOT;
61 mutex_init(&delayed_node->mutex);
62 delayed_node->index_cnt = 0;
63 INIT_LIST_HEAD(&delayed_node->n_list);
64 INIT_LIST_HEAD(&delayed_node->p_list);
65 delayed_node->bytes_reserved = 0;
66 memset(&delayed_node->inode_item, 0, sizeof(delayed_node->inode_item));
67}
68
69static inline int btrfs_is_continuous_delayed_item(
70 struct btrfs_delayed_item *item1,
71 struct btrfs_delayed_item *item2)
72{
73 if (item1->key.type == BTRFS_DIR_INDEX_KEY &&
74 item1->key.objectid == item2->key.objectid &&
75 item1->key.type == item2->key.type &&
76 item1->key.offset + 1 == item2->key.offset)
77 return 1;
78 return 0;
79}
80
81static inline struct btrfs_delayed_root *btrfs_get_delayed_root(
82 struct btrfs_root *root)
83{
84 return root->fs_info->delayed_root;
85}
86
87static struct btrfs_delayed_node *btrfs_get_delayed_node(struct inode *inode)
88{
89 struct btrfs_inode *btrfs_inode = BTRFS_I(inode);
90 struct btrfs_root *root = btrfs_inode->root;
91 u64 ino = btrfs_ino(inode);
92 struct btrfs_delayed_node *node;
93
94 node = ACCESS_ONCE(btrfs_inode->delayed_node);
95 if (node) {
96 atomic_inc(&node->refs);
97 return node;
98 }
99
100 spin_lock(&root->inode_lock);
101 node = radix_tree_lookup(&root->delayed_nodes_tree, ino);
102 if (node) {
103 if (btrfs_inode->delayed_node) {
104 atomic_inc(&node->refs); /* can be accessed */
105 BUG_ON(btrfs_inode->delayed_node != node);
106 spin_unlock(&root->inode_lock);
107 return node;
108 }
109 btrfs_inode->delayed_node = node;
110 /* can be accessed and cached in the inode */
111 atomic_add(2, &node->refs);
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 inode *inode)
123{
124 struct btrfs_delayed_node *node;
125 struct btrfs_inode *btrfs_inode = BTRFS_I(inode);
126 struct btrfs_root *root = btrfs_inode->root;
127 u64 ino = btrfs_ino(inode);
128 int ret;
129
130again:
131 node = btrfs_get_delayed_node(inode);
132 if (node)
133 return node;
134
135 node = kmem_cache_alloc(delayed_node_cache, GFP_NOFS);
136 if (!node)
137 return ERR_PTR(-ENOMEM);
138 btrfs_init_delayed_node(node, root, ino);
139
140 /* cached in the btrfs inode and can be accessed */
141 atomic_add(2, &node->refs);
142
143 ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
144 if (ret) {
145 kmem_cache_free(delayed_node_cache, node);
146 return ERR_PTR(ret);
147 }
148
149 spin_lock(&root->inode_lock);
150 ret = radix_tree_insert(&root->delayed_nodes_tree, ino, node);
151 if (ret == -EEXIST) {
152 kmem_cache_free(delayed_node_cache, node);
153 spin_unlock(&root->inode_lock);
154 radix_tree_preload_end();
155 goto again;
156 }
157 btrfs_inode->delayed_node = node;
158 spin_unlock(&root->inode_lock);
159 radix_tree_preload_end();
160
161 return node;
162}
163
164/*
165 * Call it when holding delayed_node->mutex
166 *
167 * If mod = 1, add this node into the prepared list.
168 */
169static void btrfs_queue_delayed_node(struct btrfs_delayed_root *root,
170 struct btrfs_delayed_node *node,
171 int mod)
172{
173 spin_lock(&root->lock);
174 if (test_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags)) {
175 if (!list_empty(&node->p_list))
176 list_move_tail(&node->p_list, &root->prepare_list);
177 else if (mod)
178 list_add_tail(&node->p_list, &root->prepare_list);
179 } else {
180 list_add_tail(&node->n_list, &root->node_list);
181 list_add_tail(&node->p_list, &root->prepare_list);
182 atomic_inc(&node->refs); /* inserted into list */
183 root->nodes++;
184 set_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags);
185 }
186 spin_unlock(&root->lock);
187}
188
189/* Call it when holding delayed_node->mutex */
190static void btrfs_dequeue_delayed_node(struct btrfs_delayed_root *root,
191 struct btrfs_delayed_node *node)
192{
193 spin_lock(&root->lock);
194 if (test_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags)) {
195 root->nodes--;
196 atomic_dec(&node->refs); /* not in the list */
197 list_del_init(&node->n_list);
198 if (!list_empty(&node->p_list))
199 list_del_init(&node->p_list);
200 clear_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags);
201 }
202 spin_unlock(&root->lock);
203}
204
205static struct btrfs_delayed_node *btrfs_first_delayed_node(
206 struct btrfs_delayed_root *delayed_root)
207{
208 struct list_head *p;
209 struct btrfs_delayed_node *node = NULL;
210
211 spin_lock(&delayed_root->lock);
212 if (list_empty(&delayed_root->node_list))
213 goto out;
214
215 p = delayed_root->node_list.next;
216 node = list_entry(p, struct btrfs_delayed_node, n_list);
217 atomic_inc(&node->refs);
218out:
219 spin_unlock(&delayed_root->lock);
220
221 return node;
222}
223
224static struct btrfs_delayed_node *btrfs_next_delayed_node(
225 struct btrfs_delayed_node *node)
226{
227 struct btrfs_delayed_root *delayed_root;
228 struct list_head *p;
229 struct btrfs_delayed_node *next = NULL;
230
231 delayed_root = node->root->fs_info->delayed_root;
232 spin_lock(&delayed_root->lock);
233 if (!test_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags)) {
234 /* not in the list */
235 if (list_empty(&delayed_root->node_list))
236 goto out;
237 p = delayed_root->node_list.next;
238 } else if (list_is_last(&node->n_list, &delayed_root->node_list))
239 goto out;
240 else
241 p = node->n_list.next;
242
243 next = list_entry(p, struct btrfs_delayed_node, n_list);
244 atomic_inc(&next->refs);
245out:
246 spin_unlock(&delayed_root->lock);
247
248 return next;
249}
250
251static void __btrfs_release_delayed_node(
252 struct btrfs_delayed_node *delayed_node,
253 int mod)
254{
255 struct btrfs_delayed_root *delayed_root;
256
257 if (!delayed_node)
258 return;
259
260 delayed_root = delayed_node->root->fs_info->delayed_root;
261
262 mutex_lock(&delayed_node->mutex);
263 if (delayed_node->count)
264 btrfs_queue_delayed_node(delayed_root, delayed_node, mod);
265 else
266 btrfs_dequeue_delayed_node(delayed_root, delayed_node);
267 mutex_unlock(&delayed_node->mutex);
268
269 if (atomic_dec_and_test(&delayed_node->refs)) {
270 struct btrfs_root *root = delayed_node->root;
271 spin_lock(&root->inode_lock);
272 if (atomic_read(&delayed_node->refs) == 0) {
273 radix_tree_delete(&root->delayed_nodes_tree,
274 delayed_node->inode_id);
275 kmem_cache_free(delayed_node_cache, delayed_node);
276 }
277 spin_unlock(&root->inode_lock);
278 }
279}
280
281static inline void btrfs_release_delayed_node(struct btrfs_delayed_node *node)
282{
283 __btrfs_release_delayed_node(node, 0);
284}
285
286static struct btrfs_delayed_node *btrfs_first_prepared_delayed_node(
287 struct btrfs_delayed_root *delayed_root)
288{
289 struct list_head *p;
290 struct btrfs_delayed_node *node = NULL;
291
292 spin_lock(&delayed_root->lock);
293 if (list_empty(&delayed_root->prepare_list))
294 goto out;
295
296 p = delayed_root->prepare_list.next;
297 list_del_init(p);
298 node = list_entry(p, struct btrfs_delayed_node, p_list);
299 atomic_inc(&node->refs);
300out:
301 spin_unlock(&delayed_root->lock);
302
303 return node;
304}
305
306static inline void btrfs_release_prepared_delayed_node(
307 struct btrfs_delayed_node *node)
308{
309 __btrfs_release_delayed_node(node, 1);
310}
311
312static struct btrfs_delayed_item *btrfs_alloc_delayed_item(u32 data_len)
313{
314 struct btrfs_delayed_item *item;
315 item = kmalloc(sizeof(*item) + data_len, GFP_NOFS);
316 if (item) {
317 item->data_len = data_len;
318 item->ins_or_del = 0;
319 item->bytes_reserved = 0;
320 item->delayed_node = NULL;
321 atomic_set(&item->refs, 1);
322 }
323 return item;
324}
325
326/*
327 * __btrfs_lookup_delayed_item - look up the delayed item by key
328 * @delayed_node: pointer to the delayed node
329 * @key: the key to look up
330 * @prev: used to store the prev item if the right item isn't found
331 * @next: used to store the next item if the right item isn't found
332 *
333 * Note: if we don't find the right item, we will return the prev item and
334 * the next item.
335 */
336static struct btrfs_delayed_item *__btrfs_lookup_delayed_item(
337 struct rb_root *root,
338 struct btrfs_key *key,
339 struct btrfs_delayed_item **prev,
340 struct btrfs_delayed_item **next)
341{
342 struct rb_node *node, *prev_node = NULL;
343 struct btrfs_delayed_item *delayed_item = NULL;
344 int ret = 0;
345
346 node = root->rb_node;
347
348 while (node) {
349 delayed_item = rb_entry(node, struct btrfs_delayed_item,
350 rb_node);
351 prev_node = node;
352 ret = btrfs_comp_cpu_keys(&delayed_item->key, key);
353 if (ret < 0)
354 node = node->rb_right;
355 else if (ret > 0)
356 node = node->rb_left;
357 else
358 return delayed_item;
359 }
360
361 if (prev) {
362 if (!prev_node)
363 *prev = NULL;
364 else if (ret < 0)
365 *prev = delayed_item;
366 else if ((node = rb_prev(prev_node)) != NULL) {
367 *prev = rb_entry(node, struct btrfs_delayed_item,
368 rb_node);
369 } else
370 *prev = NULL;
371 }
372
373 if (next) {
374 if (!prev_node)
375 *next = NULL;
376 else if (ret > 0)
377 *next = delayed_item;
378 else if ((node = rb_next(prev_node)) != NULL) {
379 *next = rb_entry(node, struct btrfs_delayed_item,
380 rb_node);
381 } else
382 *next = NULL;
383 }
384 return NULL;
385}
386
387static struct btrfs_delayed_item *__btrfs_lookup_delayed_insertion_item(
388 struct btrfs_delayed_node *delayed_node,
389 struct btrfs_key *key)
390{
391 struct btrfs_delayed_item *item;
392
393 item = __btrfs_lookup_delayed_item(&delayed_node->ins_root, key,
394 NULL, NULL);
395 return item;
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 *root;
405 struct btrfs_delayed_item *item;
406 int cmp;
407
408 if (action == BTRFS_DELAYED_INSERTION_ITEM)
409 root = &delayed_node->ins_root;
410 else if (action == BTRFS_DELAYED_DELETION_ITEM)
411 root = &delayed_node->del_root;
412 else
413 BUG();
414 p = &root->rb_node;
415 node = &ins->rb_node;
416
417 while (*p) {
418 parent_node = *p;
419 item = rb_entry(parent_node, struct btrfs_delayed_item,
420 rb_node);
421
422 cmp = btrfs_comp_cpu_keys(&item->key, &ins->key);
423 if (cmp < 0)
424 p = &(*p)->rb_right;
425 else if (cmp > 0)
426 p = &(*p)->rb_left;
427 else
428 return -EEXIST;
429 }
430
431 rb_link_node(node, parent_node, p);
432 rb_insert_color(node, root);
433 ins->delayed_node = delayed_node;
434 ins->ins_or_del = action;
435
436 if (ins->key.type == BTRFS_DIR_INDEX_KEY &&
437 action == BTRFS_DELAYED_INSERTION_ITEM &&
438 ins->key.offset >= delayed_node->index_cnt)
439 delayed_node->index_cnt = ins->key.offset + 1;
440
441 delayed_node->count++;
442 atomic_inc(&delayed_node->root->fs_info->delayed_root->items);
443 return 0;
444}
445
446static int __btrfs_add_delayed_insertion_item(struct btrfs_delayed_node *node,
447 struct btrfs_delayed_item *item)
448{
449 return __btrfs_add_delayed_item(node, item,
450 BTRFS_DELAYED_INSERTION_ITEM);
451}
452
453static int __btrfs_add_delayed_deletion_item(struct btrfs_delayed_node *node,
454 struct btrfs_delayed_item *item)
455{
456 return __btrfs_add_delayed_item(node, item,
457 BTRFS_DELAYED_DELETION_ITEM);
458}
459
460static void finish_one_item(struct btrfs_delayed_root *delayed_root)
461{
462 int seq = atomic_inc_return(&delayed_root->items_seq);
463 if ((atomic_dec_return(&delayed_root->items) <
464 BTRFS_DELAYED_BACKGROUND || seq % BTRFS_DELAYED_BATCH == 0) &&
465 waitqueue_active(&delayed_root->wait))
466 wake_up(&delayed_root->wait);
467}
468
469static void __btrfs_remove_delayed_item(struct btrfs_delayed_item *delayed_item)
470{
471 struct rb_root *root;
472 struct btrfs_delayed_root *delayed_root;
473
474 delayed_root = delayed_item->delayed_node->root->fs_info->delayed_root;
475
476 BUG_ON(!delayed_root);
477 BUG_ON(delayed_item->ins_or_del != BTRFS_DELAYED_DELETION_ITEM &&
478 delayed_item->ins_or_del != BTRFS_DELAYED_INSERTION_ITEM);
479
480 if (delayed_item->ins_or_del == BTRFS_DELAYED_INSERTION_ITEM)
481 root = &delayed_item->delayed_node->ins_root;
482 else
483 root = &delayed_item->delayed_node->del_root;
484
485 rb_erase(&delayed_item->rb_node, root);
486 delayed_item->delayed_node->count--;
487
488 finish_one_item(delayed_root);
489}
490
491static void btrfs_release_delayed_item(struct btrfs_delayed_item *item)
492{
493 if (item) {
494 __btrfs_remove_delayed_item(item);
495 if (atomic_dec_and_test(&item->refs))
496 kfree(item);
497 }
498}
499
500static struct btrfs_delayed_item *__btrfs_first_delayed_insertion_item(
501 struct btrfs_delayed_node *delayed_node)
502{
503 struct rb_node *p;
504 struct btrfs_delayed_item *item = NULL;
505
506 p = rb_first(&delayed_node->ins_root);
507 if (p)
508 item = rb_entry(p, struct btrfs_delayed_item, rb_node);
509
510 return item;
511}
512
513static struct btrfs_delayed_item *__btrfs_first_delayed_deletion_item(
514 struct btrfs_delayed_node *delayed_node)
515{
516 struct rb_node *p;
517 struct btrfs_delayed_item *item = NULL;
518
519 p = rb_first(&delayed_node->del_root);
520 if (p)
521 item = rb_entry(p, struct btrfs_delayed_item, rb_node);
522
523 return item;
524}
525
526static struct btrfs_delayed_item *__btrfs_next_delayed_item(
527 struct btrfs_delayed_item *item)
528{
529 struct rb_node *p;
530 struct btrfs_delayed_item *next = NULL;
531
532 p = rb_next(&item->rb_node);
533 if (p)
534 next = rb_entry(p, struct btrfs_delayed_item, rb_node);
535
536 return next;
537}
538
539static int btrfs_delayed_item_reserve_metadata(struct btrfs_trans_handle *trans,
540 struct btrfs_root *root,
541 struct btrfs_delayed_item *item)
542{
543 struct btrfs_block_rsv *src_rsv;
544 struct btrfs_block_rsv *dst_rsv;
545 u64 num_bytes;
546 int ret;
547
548 if (!trans->bytes_reserved)
549 return 0;
550
551 src_rsv = trans->block_rsv;
552 dst_rsv = &root->fs_info->delayed_block_rsv;
553
554 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
555 ret = btrfs_block_rsv_migrate(src_rsv, dst_rsv, num_bytes);
556 if (!ret) {
557 trace_btrfs_space_reservation(root->fs_info, "delayed_item",
558 item->key.objectid,
559 num_bytes, 1);
560 item->bytes_reserved = num_bytes;
561 }
562
563 return ret;
564}
565
566static void btrfs_delayed_item_release_metadata(struct btrfs_root *root,
567 struct btrfs_delayed_item *item)
568{
569 struct btrfs_block_rsv *rsv;
570
571 if (!item->bytes_reserved)
572 return;
573
574 rsv = &root->fs_info->delayed_block_rsv;
575 trace_btrfs_space_reservation(root->fs_info, "delayed_item",
576 item->key.objectid, item->bytes_reserved,
577 0);
578 btrfs_block_rsv_release(root, rsv,
579 item->bytes_reserved);
580}
581
582static int btrfs_delayed_inode_reserve_metadata(
583 struct btrfs_trans_handle *trans,
584 struct btrfs_root *root,
585 struct inode *inode,
586 struct btrfs_delayed_node *node)
587{
588 struct btrfs_block_rsv *src_rsv;
589 struct btrfs_block_rsv *dst_rsv;
590 u64 num_bytes;
591 int ret;
592 bool release = false;
593
594 src_rsv = trans->block_rsv;
595 dst_rsv = &root->fs_info->delayed_block_rsv;
596
597 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
598
599 /*
600 * btrfs_dirty_inode will update the inode under btrfs_join_transaction
601 * which doesn't reserve space for speed. This is a problem since we
602 * still need to reserve space for this update, so try to reserve the
603 * space.
604 *
605 * Now if src_rsv == delalloc_block_rsv we'll let it just steal since
606 * we're accounted for.
607 */
608 if (!src_rsv || (!trans->bytes_reserved &&
609 src_rsv->type != BTRFS_BLOCK_RSV_DELALLOC)) {
610 ret = btrfs_block_rsv_add(root, dst_rsv, num_bytes,
611 BTRFS_RESERVE_NO_FLUSH);
612 /*
613 * Since we're under a transaction reserve_metadata_bytes could
614 * try to commit the transaction which will make it return
615 * EAGAIN to make us stop the transaction we have, so return
616 * ENOSPC instead so that btrfs_dirty_inode knows what to do.
617 */
618 if (ret == -EAGAIN)
619 ret = -ENOSPC;
620 if (!ret) {
621 node->bytes_reserved = num_bytes;
622 trace_btrfs_space_reservation(root->fs_info,
623 "delayed_inode",
624 btrfs_ino(inode),
625 num_bytes, 1);
626 }
627 return ret;
628 } else if (src_rsv->type == BTRFS_BLOCK_RSV_DELALLOC) {
629 spin_lock(&BTRFS_I(inode)->lock);
630 if (test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
631 &BTRFS_I(inode)->runtime_flags)) {
632 spin_unlock(&BTRFS_I(inode)->lock);
633 release = true;
634 goto migrate;
635 }
636 spin_unlock(&BTRFS_I(inode)->lock);
637
638 /* Ok we didn't have space pre-reserved. This shouldn't happen
639 * too often but it can happen if we do delalloc to an existing
640 * inode which gets dirtied because of the time update, and then
641 * isn't touched again until after the transaction commits and
642 * then we try to write out the data. First try to be nice and
643 * reserve something strictly for us. If not be a pain and try
644 * to steal from the delalloc block rsv.
645 */
646 ret = btrfs_block_rsv_add(root, dst_rsv, num_bytes,
647 BTRFS_RESERVE_NO_FLUSH);
648 if (!ret)
649 goto out;
650
651 ret = btrfs_block_rsv_migrate(src_rsv, dst_rsv, num_bytes);
652 if (!WARN_ON(ret))
653 goto out;
654
655 /*
656 * Ok this is a problem, let's just steal from the global rsv
657 * since this really shouldn't happen that often.
658 */
659 ret = btrfs_block_rsv_migrate(&root->fs_info->global_block_rsv,
660 dst_rsv, num_bytes);
661 goto out;
662 }
663
664migrate:
665 ret = btrfs_block_rsv_migrate(src_rsv, dst_rsv, num_bytes);
666
667out:
668 /*
669 * Migrate only takes a reservation, it doesn't touch the size of the
670 * block_rsv. This is to simplify people who don't normally have things
671 * migrated from their block rsv. If they go to release their
672 * reservation, that will decrease the size as well, so if migrate
673 * reduced size we'd end up with a negative size. But for the
674 * delalloc_meta_reserved stuff we will only know to drop 1 reservation,
675 * but we could in fact do this reserve/migrate dance several times
676 * between the time we did the original reservation and we'd clean it
677 * up. So to take care of this, release the space for the meta
678 * reservation here. I think it may be time for a documentation page on
679 * how block rsvs. work.
680 */
681 if (!ret) {
682 trace_btrfs_space_reservation(root->fs_info, "delayed_inode",
683 btrfs_ino(inode), num_bytes, 1);
684 node->bytes_reserved = num_bytes;
685 }
686
687 if (release) {
688 trace_btrfs_space_reservation(root->fs_info, "delalloc",
689 btrfs_ino(inode), num_bytes, 0);
690 btrfs_block_rsv_release(root, src_rsv, num_bytes);
691 }
692
693 return ret;
694}
695
696static void btrfs_delayed_inode_release_metadata(struct btrfs_root *root,
697 struct btrfs_delayed_node *node)
698{
699 struct btrfs_block_rsv *rsv;
700
701 if (!node->bytes_reserved)
702 return;
703
704 rsv = &root->fs_info->delayed_block_rsv;
705 trace_btrfs_space_reservation(root->fs_info, "delayed_inode",
706 node->inode_id, node->bytes_reserved, 0);
707 btrfs_block_rsv_release(root, rsv,
708 node->bytes_reserved);
709 node->bytes_reserved = 0;
710}
711
712/*
713 * This helper will insert some continuous items into the same leaf according
714 * to the free space of the leaf.
715 */
716static int btrfs_batch_insert_items(struct btrfs_root *root,
717 struct btrfs_path *path,
718 struct btrfs_delayed_item *item)
719{
720 struct btrfs_delayed_item *curr, *next;
721 int free_space;
722 int total_data_size = 0, total_size = 0;
723 struct extent_buffer *leaf;
724 char *data_ptr;
725 struct btrfs_key *keys;
726 u32 *data_size;
727 struct list_head head;
728 int slot;
729 int nitems;
730 int i;
731 int ret = 0;
732
733 BUG_ON(!path->nodes[0]);
734
735 leaf = path->nodes[0];
736 free_space = btrfs_leaf_free_space(root, leaf);
737 INIT_LIST_HEAD(&head);
738
739 next = item;
740 nitems = 0;
741
742 /*
743 * count the number of the continuous items that we can insert in batch
744 */
745 while (total_size + next->data_len + sizeof(struct btrfs_item) <=
746 free_space) {
747 total_data_size += next->data_len;
748 total_size += next->data_len + sizeof(struct btrfs_item);
749 list_add_tail(&next->tree_list, &head);
750 nitems++;
751
752 curr = next;
753 next = __btrfs_next_delayed_item(curr);
754 if (!next)
755 break;
756
757 if (!btrfs_is_continuous_delayed_item(curr, next))
758 break;
759 }
760
761 if (!nitems) {
762 ret = 0;
763 goto out;
764 }
765
766 /*
767 * we need allocate some memory space, but it might cause the task
768 * to sleep, so we set all locked nodes in the path to blocking locks
769 * first.
770 */
771 btrfs_set_path_blocking(path);
772
773 keys = kmalloc_array(nitems, sizeof(struct btrfs_key), GFP_NOFS);
774 if (!keys) {
775 ret = -ENOMEM;
776 goto out;
777 }
778
779 data_size = kmalloc_array(nitems, sizeof(u32), GFP_NOFS);
780 if (!data_size) {
781 ret = -ENOMEM;
782 goto error;
783 }
784
785 /* get keys of all the delayed items */
786 i = 0;
787 list_for_each_entry(next, &head, tree_list) {
788 keys[i] = next->key;
789 data_size[i] = next->data_len;
790 i++;
791 }
792
793 /* reset all the locked nodes in the patch to spinning locks. */
794 btrfs_clear_path_blocking(path, NULL, 0);
795
796 /* insert the keys of the items */
797 setup_items_for_insert(root, path, keys, data_size,
798 total_data_size, total_size, nitems);
799
800 /* insert the dir index items */
801 slot = path->slots[0];
802 list_for_each_entry_safe(curr, next, &head, tree_list) {
803 data_ptr = btrfs_item_ptr(leaf, slot, char);
804 write_extent_buffer(leaf, &curr->data,
805 (unsigned long)data_ptr,
806 curr->data_len);
807 slot++;
808
809 btrfs_delayed_item_release_metadata(root, curr);
810
811 list_del(&curr->tree_list);
812 btrfs_release_delayed_item(curr);
813 }
814
815error:
816 kfree(data_size);
817 kfree(keys);
818out:
819 return ret;
820}
821
822/*
823 * This helper can just do simple insertion that needn't extend item for new
824 * data, such as directory name index insertion, inode insertion.
825 */
826static int btrfs_insert_delayed_item(struct btrfs_trans_handle *trans,
827 struct btrfs_root *root,
828 struct btrfs_path *path,
829 struct btrfs_delayed_item *delayed_item)
830{
831 struct extent_buffer *leaf;
832 char *ptr;
833 int ret;
834
835 ret = btrfs_insert_empty_item(trans, root, path, &delayed_item->key,
836 delayed_item->data_len);
837 if (ret < 0 && ret != -EEXIST)
838 return ret;
839
840 leaf = path->nodes[0];
841
842 ptr = btrfs_item_ptr(leaf, path->slots[0], char);
843
844 write_extent_buffer(leaf, delayed_item->data, (unsigned long)ptr,
845 delayed_item->data_len);
846 btrfs_mark_buffer_dirty(leaf);
847
848 btrfs_delayed_item_release_metadata(root, delayed_item);
849 return 0;
850}
851
852/*
853 * we insert an item first, then if there are some continuous items, we try
854 * to insert those items into the same leaf.
855 */
856static int btrfs_insert_delayed_items(struct btrfs_trans_handle *trans,
857 struct btrfs_path *path,
858 struct btrfs_root *root,
859 struct btrfs_delayed_node *node)
860{
861 struct btrfs_delayed_item *curr, *prev;
862 int ret = 0;
863
864do_again:
865 mutex_lock(&node->mutex);
866 curr = __btrfs_first_delayed_insertion_item(node);
867 if (!curr)
868 goto insert_end;
869
870 ret = btrfs_insert_delayed_item(trans, root, path, curr);
871 if (ret < 0) {
872 btrfs_release_path(path);
873 goto insert_end;
874 }
875
876 prev = curr;
877 curr = __btrfs_next_delayed_item(prev);
878 if (curr && btrfs_is_continuous_delayed_item(prev, curr)) {
879 /* insert the continuous items into the same leaf */
880 path->slots[0]++;
881 btrfs_batch_insert_items(root, path, curr);
882 }
883 btrfs_release_delayed_item(prev);
884 btrfs_mark_buffer_dirty(path->nodes[0]);
885
886 btrfs_release_path(path);
887 mutex_unlock(&node->mutex);
888 goto do_again;
889
890insert_end:
891 mutex_unlock(&node->mutex);
892 return ret;
893}
894
895static int btrfs_batch_delete_items(struct btrfs_trans_handle *trans,
896 struct btrfs_root *root,
897 struct btrfs_path *path,
898 struct btrfs_delayed_item *item)
899{
900 struct btrfs_delayed_item *curr, *next;
901 struct extent_buffer *leaf;
902 struct btrfs_key key;
903 struct list_head head;
904 int nitems, i, last_item;
905 int ret = 0;
906
907 BUG_ON(!path->nodes[0]);
908
909 leaf = path->nodes[0];
910
911 i = path->slots[0];
912 last_item = btrfs_header_nritems(leaf) - 1;
913 if (i > last_item)
914 return -ENOENT; /* FIXME: Is errno suitable? */
915
916 next = item;
917 INIT_LIST_HEAD(&head);
918 btrfs_item_key_to_cpu(leaf, &key, i);
919 nitems = 0;
920 /*
921 * count the number of the dir index items that we can delete in batch
922 */
923 while (btrfs_comp_cpu_keys(&next->key, &key) == 0) {
924 list_add_tail(&next->tree_list, &head);
925 nitems++;
926
927 curr = next;
928 next = __btrfs_next_delayed_item(curr);
929 if (!next)
930 break;
931
932 if (!btrfs_is_continuous_delayed_item(curr, next))
933 break;
934
935 i++;
936 if (i > last_item)
937 break;
938 btrfs_item_key_to_cpu(leaf, &key, i);
939 }
940
941 if (!nitems)
942 return 0;
943
944 ret = btrfs_del_items(trans, root, path, path->slots[0], nitems);
945 if (ret)
946 goto out;
947
948 list_for_each_entry_safe(curr, next, &head, tree_list) {
949 btrfs_delayed_item_release_metadata(root, curr);
950 list_del(&curr->tree_list);
951 btrfs_release_delayed_item(curr);
952 }
953
954out:
955 return ret;
956}
957
958static int btrfs_delete_delayed_items(struct btrfs_trans_handle *trans,
959 struct btrfs_path *path,
960 struct btrfs_root *root,
961 struct btrfs_delayed_node *node)
962{
963 struct btrfs_delayed_item *curr, *prev;
964 int ret = 0;
965
966do_again:
967 mutex_lock(&node->mutex);
968 curr = __btrfs_first_delayed_deletion_item(node);
969 if (!curr)
970 goto delete_fail;
971
972 ret = btrfs_search_slot(trans, root, &curr->key, path, -1, 1);
973 if (ret < 0)
974 goto delete_fail;
975 else if (ret > 0) {
976 /*
977 * can't find the item which the node points to, so this node
978 * is invalid, just drop it.
979 */
980 prev = curr;
981 curr = __btrfs_next_delayed_item(prev);
982 btrfs_release_delayed_item(prev);
983 ret = 0;
984 btrfs_release_path(path);
985 if (curr) {
986 mutex_unlock(&node->mutex);
987 goto do_again;
988 } else
989 goto delete_fail;
990 }
991
992 btrfs_batch_delete_items(trans, root, path, curr);
993 btrfs_release_path(path);
994 mutex_unlock(&node->mutex);
995 goto do_again;
996
997delete_fail:
998 btrfs_release_path(path);
999 mutex_unlock(&node->mutex);
1000 return ret;
1001}
1002
1003static void btrfs_release_delayed_inode(struct btrfs_delayed_node *delayed_node)
1004{
1005 struct btrfs_delayed_root *delayed_root;
1006
1007 if (delayed_node &&
1008 test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
1009 BUG_ON(!delayed_node->root);
1010 clear_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags);
1011 delayed_node->count--;
1012
1013 delayed_root = delayed_node->root->fs_info->delayed_root;
1014 finish_one_item(delayed_root);
1015 }
1016}
1017
1018static void btrfs_release_delayed_iref(struct btrfs_delayed_node *delayed_node)
1019{
1020 struct btrfs_delayed_root *delayed_root;
1021
1022 ASSERT(delayed_node->root);
1023 clear_bit(BTRFS_DELAYED_NODE_DEL_IREF, &delayed_node->flags);
1024 delayed_node->count--;
1025
1026 delayed_root = delayed_node->root->fs_info->delayed_root;
1027 finish_one_item(delayed_root);
1028}
1029
1030static int __btrfs_update_delayed_inode(struct btrfs_trans_handle *trans,
1031 struct btrfs_root *root,
1032 struct btrfs_path *path,
1033 struct btrfs_delayed_node *node)
1034{
1035 struct btrfs_key key;
1036 struct btrfs_inode_item *inode_item;
1037 struct extent_buffer *leaf;
1038 int mod;
1039 int ret;
1040
1041 key.objectid = node->inode_id;
1042 btrfs_set_key_type(&key, BTRFS_INODE_ITEM_KEY);
1043 key.offset = 0;
1044
1045 if (test_bit(BTRFS_DELAYED_NODE_DEL_IREF, &node->flags))
1046 mod = -1;
1047 else
1048 mod = 1;
1049
1050 ret = btrfs_lookup_inode(trans, root, path, &key, mod);
1051 if (ret > 0) {
1052 btrfs_release_path(path);
1053 return -ENOENT;
1054 } else if (ret < 0) {
1055 return ret;
1056 }
1057
1058 leaf = path->nodes[0];
1059 inode_item = btrfs_item_ptr(leaf, path->slots[0],
1060 struct btrfs_inode_item);
1061 write_extent_buffer(leaf, &node->inode_item, (unsigned long)inode_item,
1062 sizeof(struct btrfs_inode_item));
1063 btrfs_mark_buffer_dirty(leaf);
1064
1065 if (!test_bit(BTRFS_DELAYED_NODE_DEL_IREF, &node->flags))
1066 goto no_iref;
1067
1068 path->slots[0]++;
1069 if (path->slots[0] >= btrfs_header_nritems(leaf))
1070 goto search;
1071again:
1072 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1073 if (key.objectid != node->inode_id)
1074 goto out;
1075
1076 if (key.type != BTRFS_INODE_REF_KEY &&
1077 key.type != BTRFS_INODE_EXTREF_KEY)
1078 goto out;
1079
1080 /*
1081 * Delayed iref deletion is for the inode who has only one link,
1082 * so there is only one iref. The case that several irefs are
1083 * in the same item doesn't exist.
1084 */
1085 btrfs_del_item(trans, root, path);
1086out:
1087 btrfs_release_delayed_iref(node);
1088no_iref:
1089 btrfs_release_path(path);
1090err_out:
1091 btrfs_delayed_inode_release_metadata(root, node);
1092 btrfs_release_delayed_inode(node);
1093
1094 return ret;
1095
1096search:
1097 btrfs_release_path(path);
1098
1099 btrfs_set_key_type(&key, BTRFS_INODE_EXTREF_KEY);
1100 key.offset = -1;
1101 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1102 if (ret < 0)
1103 goto err_out;
1104 ASSERT(ret);
1105
1106 ret = 0;
1107 leaf = path->nodes[0];
1108 path->slots[0]--;
1109 goto again;
1110}
1111
1112static inline int btrfs_update_delayed_inode(struct btrfs_trans_handle *trans,
1113 struct btrfs_root *root,
1114 struct btrfs_path *path,
1115 struct btrfs_delayed_node *node)
1116{
1117 int ret;
1118
1119 mutex_lock(&node->mutex);
1120 if (!test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &node->flags)) {
1121 mutex_unlock(&node->mutex);
1122 return 0;
1123 }
1124
1125 ret = __btrfs_update_delayed_inode(trans, root, path, node);
1126 mutex_unlock(&node->mutex);
1127 return ret;
1128}
1129
1130static inline int
1131__btrfs_commit_inode_delayed_items(struct btrfs_trans_handle *trans,
1132 struct btrfs_path *path,
1133 struct btrfs_delayed_node *node)
1134{
1135 int ret;
1136
1137 ret = btrfs_insert_delayed_items(trans, path, node->root, node);
1138 if (ret)
1139 return ret;
1140
1141 ret = btrfs_delete_delayed_items(trans, path, node->root, node);
1142 if (ret)
1143 return ret;
1144
1145 ret = btrfs_update_delayed_inode(trans, node->root, path, node);
1146 return ret;
1147}
1148
1149/*
1150 * Called when committing the transaction.
1151 * Returns 0 on success.
1152 * Returns < 0 on error and returns with an aborted transaction with any
1153 * outstanding delayed items cleaned up.
1154 */
1155static int __btrfs_run_delayed_items(struct btrfs_trans_handle *trans,
1156 struct btrfs_root *root, int nr)
1157{
1158 struct btrfs_delayed_root *delayed_root;
1159 struct btrfs_delayed_node *curr_node, *prev_node;
1160 struct btrfs_path *path;
1161 struct btrfs_block_rsv *block_rsv;
1162 int ret = 0;
1163 bool count = (nr > 0);
1164
1165 if (trans->aborted)
1166 return -EIO;
1167
1168 path = btrfs_alloc_path();
1169 if (!path)
1170 return -ENOMEM;
1171 path->leave_spinning = 1;
1172
1173 block_rsv = trans->block_rsv;
1174 trans->block_rsv = &root->fs_info->delayed_block_rsv;
1175
1176 delayed_root = btrfs_get_delayed_root(root);
1177
1178 curr_node = btrfs_first_delayed_node(delayed_root);
1179 while (curr_node && (!count || (count && nr--))) {
1180 ret = __btrfs_commit_inode_delayed_items(trans, path,
1181 curr_node);
1182 if (ret) {
1183 btrfs_release_delayed_node(curr_node);
1184 curr_node = NULL;
1185 btrfs_abort_transaction(trans, root, ret);
1186 break;
1187 }
1188
1189 prev_node = curr_node;
1190 curr_node = btrfs_next_delayed_node(curr_node);
1191 btrfs_release_delayed_node(prev_node);
1192 }
1193
1194 if (curr_node)
1195 btrfs_release_delayed_node(curr_node);
1196 btrfs_free_path(path);
1197 trans->block_rsv = block_rsv;
1198
1199 return ret;
1200}
1201
1202int btrfs_run_delayed_items(struct btrfs_trans_handle *trans,
1203 struct btrfs_root *root)
1204{
1205 return __btrfs_run_delayed_items(trans, root, -1);
1206}
1207
1208int btrfs_run_delayed_items_nr(struct btrfs_trans_handle *trans,
1209 struct btrfs_root *root, int nr)
1210{
1211 return __btrfs_run_delayed_items(trans, root, nr);
1212}
1213
1214int btrfs_commit_inode_delayed_items(struct btrfs_trans_handle *trans,
1215 struct inode *inode)
1216{
1217 struct btrfs_delayed_node *delayed_node = btrfs_get_delayed_node(inode);
1218 struct btrfs_path *path;
1219 struct btrfs_block_rsv *block_rsv;
1220 int ret;
1221
1222 if (!delayed_node)
1223 return 0;
1224
1225 mutex_lock(&delayed_node->mutex);
1226 if (!delayed_node->count) {
1227 mutex_unlock(&delayed_node->mutex);
1228 btrfs_release_delayed_node(delayed_node);
1229 return 0;
1230 }
1231 mutex_unlock(&delayed_node->mutex);
1232
1233 path = btrfs_alloc_path();
1234 if (!path) {
1235 btrfs_release_delayed_node(delayed_node);
1236 return -ENOMEM;
1237 }
1238 path->leave_spinning = 1;
1239
1240 block_rsv = trans->block_rsv;
1241 trans->block_rsv = &delayed_node->root->fs_info->delayed_block_rsv;
1242
1243 ret = __btrfs_commit_inode_delayed_items(trans, path, delayed_node);
1244
1245 btrfs_release_delayed_node(delayed_node);
1246 btrfs_free_path(path);
1247 trans->block_rsv = block_rsv;
1248
1249 return ret;
1250}
1251
1252int btrfs_commit_inode_delayed_inode(struct inode *inode)
1253{
1254 struct btrfs_trans_handle *trans;
1255 struct btrfs_delayed_node *delayed_node = btrfs_get_delayed_node(inode);
1256 struct btrfs_path *path;
1257 struct btrfs_block_rsv *block_rsv;
1258 int ret;
1259
1260 if (!delayed_node)
1261 return 0;
1262
1263 mutex_lock(&delayed_node->mutex);
1264 if (!test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
1265 mutex_unlock(&delayed_node->mutex);
1266 btrfs_release_delayed_node(delayed_node);
1267 return 0;
1268 }
1269 mutex_unlock(&delayed_node->mutex);
1270
1271 trans = btrfs_join_transaction(delayed_node->root);
1272 if (IS_ERR(trans)) {
1273 ret = PTR_ERR(trans);
1274 goto out;
1275 }
1276
1277 path = btrfs_alloc_path();
1278 if (!path) {
1279 ret = -ENOMEM;
1280 goto trans_out;
1281 }
1282 path->leave_spinning = 1;
1283
1284 block_rsv = trans->block_rsv;
1285 trans->block_rsv = &delayed_node->root->fs_info->delayed_block_rsv;
1286
1287 mutex_lock(&delayed_node->mutex);
1288 if (test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags))
1289 ret = __btrfs_update_delayed_inode(trans, delayed_node->root,
1290 path, delayed_node);
1291 else
1292 ret = 0;
1293 mutex_unlock(&delayed_node->mutex);
1294
1295 btrfs_free_path(path);
1296 trans->block_rsv = block_rsv;
1297trans_out:
1298 btrfs_end_transaction(trans, delayed_node->root);
1299 btrfs_btree_balance_dirty(delayed_node->root);
1300out:
1301 btrfs_release_delayed_node(delayed_node);
1302
1303 return ret;
1304}
1305
1306void btrfs_remove_delayed_node(struct inode *inode)
1307{
1308 struct btrfs_delayed_node *delayed_node;
1309
1310 delayed_node = ACCESS_ONCE(BTRFS_I(inode)->delayed_node);
1311 if (!delayed_node)
1312 return;
1313
1314 BTRFS_I(inode)->delayed_node = NULL;
1315 btrfs_release_delayed_node(delayed_node);
1316}
1317
1318struct btrfs_async_delayed_work {
1319 struct btrfs_delayed_root *delayed_root;
1320 int nr;
1321 struct btrfs_work work;
1322};
1323
1324static void btrfs_async_run_delayed_root(struct btrfs_work *work)
1325{
1326 struct btrfs_async_delayed_work *async_work;
1327 struct btrfs_delayed_root *delayed_root;
1328 struct btrfs_trans_handle *trans;
1329 struct btrfs_path *path;
1330 struct btrfs_delayed_node *delayed_node = NULL;
1331 struct btrfs_root *root;
1332 struct btrfs_block_rsv *block_rsv;
1333 int total_done = 0;
1334
1335 async_work = container_of(work, struct btrfs_async_delayed_work, work);
1336 delayed_root = async_work->delayed_root;
1337
1338 path = btrfs_alloc_path();
1339 if (!path)
1340 goto out;
1341
1342again:
1343 if (atomic_read(&delayed_root->items) < BTRFS_DELAYED_BACKGROUND / 2)
1344 goto free_path;
1345
1346 delayed_node = btrfs_first_prepared_delayed_node(delayed_root);
1347 if (!delayed_node)
1348 goto free_path;
1349
1350 path->leave_spinning = 1;
1351 root = delayed_node->root;
1352
1353 trans = btrfs_join_transaction(root);
1354 if (IS_ERR(trans))
1355 goto release_path;
1356
1357 block_rsv = trans->block_rsv;
1358 trans->block_rsv = &root->fs_info->delayed_block_rsv;
1359
1360 __btrfs_commit_inode_delayed_items(trans, path, delayed_node);
1361
1362 trans->block_rsv = block_rsv;
1363 btrfs_end_transaction(trans, root);
1364 btrfs_btree_balance_dirty_nodelay(root);
1365
1366release_path:
1367 btrfs_release_path(path);
1368 total_done++;
1369
1370 btrfs_release_prepared_delayed_node(delayed_node);
1371 if (async_work->nr == 0 || total_done < async_work->nr)
1372 goto again;
1373
1374free_path:
1375 btrfs_free_path(path);
1376out:
1377 wake_up(&delayed_root->wait);
1378 kfree(async_work);
1379}
1380
1381
1382static int btrfs_wq_run_delayed_node(struct btrfs_delayed_root *delayed_root,
1383 struct btrfs_root *root, int nr)
1384{
1385 struct btrfs_async_delayed_work *async_work;
1386
1387 if (atomic_read(&delayed_root->items) < BTRFS_DELAYED_BACKGROUND)
1388 return 0;
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,
1396 NULL, NULL);
1397 async_work->nr = nr;
1398
1399 btrfs_queue_work(root->fs_info->delayed_workers, &async_work->work);
1400 return 0;
1401}
1402
1403void btrfs_assert_delayed_root_empty(struct btrfs_root *root)
1404{
1405 struct btrfs_delayed_root *delayed_root;
1406 delayed_root = btrfs_get_delayed_root(root);
1407 WARN_ON(btrfs_first_delayed_node(delayed_root));
1408}
1409
1410static int could_end_wait(struct btrfs_delayed_root *delayed_root, int seq)
1411{
1412 int val = atomic_read(&delayed_root->items_seq);
1413
1414 if (val < seq || val >= seq + BTRFS_DELAYED_BATCH)
1415 return 1;
1416
1417 if (atomic_read(&delayed_root->items) < BTRFS_DELAYED_BACKGROUND)
1418 return 1;
1419
1420 return 0;
1421}
1422
1423void btrfs_balance_delayed_items(struct btrfs_root *root)
1424{
1425 struct btrfs_delayed_root *delayed_root;
1426
1427 delayed_root = btrfs_get_delayed_root(root);
1428
1429 if (atomic_read(&delayed_root->items) < BTRFS_DELAYED_BACKGROUND)
1430 return;
1431
1432 if (atomic_read(&delayed_root->items) >= BTRFS_DELAYED_WRITEBACK) {
1433 int seq;
1434 int ret;
1435
1436 seq = atomic_read(&delayed_root->items_seq);
1437
1438 ret = btrfs_wq_run_delayed_node(delayed_root, root, 0);
1439 if (ret)
1440 return;
1441
1442 wait_event_interruptible(delayed_root->wait,
1443 could_end_wait(delayed_root, seq));
1444 return;
1445 }
1446
1447 btrfs_wq_run_delayed_node(delayed_root, root, BTRFS_DELAYED_BATCH);
1448}
1449
1450/* Will return 0 or -ENOMEM */
1451int btrfs_insert_delayed_dir_index(struct btrfs_trans_handle *trans,
1452 struct btrfs_root *root, const char *name,
1453 int name_len, struct inode *dir,
1454 struct btrfs_disk_key *disk_key, u8 type,
1455 u64 index)
1456{
1457 struct btrfs_delayed_node *delayed_node;
1458 struct btrfs_delayed_item *delayed_item;
1459 struct btrfs_dir_item *dir_item;
1460 int ret;
1461
1462 delayed_node = btrfs_get_or_create_delayed_node(dir);
1463 if (IS_ERR(delayed_node))
1464 return PTR_ERR(delayed_node);
1465
1466 delayed_item = btrfs_alloc_delayed_item(sizeof(*dir_item) + name_len);
1467 if (!delayed_item) {
1468 ret = -ENOMEM;
1469 goto release_node;
1470 }
1471
1472 delayed_item->key.objectid = btrfs_ino(dir);
1473 btrfs_set_key_type(&delayed_item->key, BTRFS_DIR_INDEX_KEY);
1474 delayed_item->key.offset = index;
1475
1476 dir_item = (struct btrfs_dir_item *)delayed_item->data;
1477 dir_item->location = *disk_key;
1478 btrfs_set_stack_dir_transid(dir_item, trans->transid);
1479 btrfs_set_stack_dir_data_len(dir_item, 0);
1480 btrfs_set_stack_dir_name_len(dir_item, name_len);
1481 btrfs_set_stack_dir_type(dir_item, type);
1482 memcpy((char *)(dir_item + 1), name, name_len);
1483
1484 ret = btrfs_delayed_item_reserve_metadata(trans, root, delayed_item);
1485 /*
1486 * we have reserved enough space when we start a new transaction,
1487 * so reserving metadata failure is impossible
1488 */
1489 BUG_ON(ret);
1490
1491
1492 mutex_lock(&delayed_node->mutex);
1493 ret = __btrfs_add_delayed_insertion_item(delayed_node, delayed_item);
1494 if (unlikely(ret)) {
1495 btrfs_err(root->fs_info, "err add delayed dir index item(name: %.*s) "
1496 "into the insertion tree of the delayed node"
1497 "(root id: %llu, inode id: %llu, errno: %d)",
1498 name_len, name, delayed_node->root->objectid,
1499 delayed_node->inode_id, ret);
1500 BUG();
1501 }
1502 mutex_unlock(&delayed_node->mutex);
1503
1504release_node:
1505 btrfs_release_delayed_node(delayed_node);
1506 return ret;
1507}
1508
1509static int btrfs_delete_delayed_insertion_item(struct btrfs_root *root,
1510 struct btrfs_delayed_node *node,
1511 struct btrfs_key *key)
1512{
1513 struct btrfs_delayed_item *item;
1514
1515 mutex_lock(&node->mutex);
1516 item = __btrfs_lookup_delayed_insertion_item(node, key);
1517 if (!item) {
1518 mutex_unlock(&node->mutex);
1519 return 1;
1520 }
1521
1522 btrfs_delayed_item_release_metadata(root, item);
1523 btrfs_release_delayed_item(item);
1524 mutex_unlock(&node->mutex);
1525 return 0;
1526}
1527
1528int btrfs_delete_delayed_dir_index(struct btrfs_trans_handle *trans,
1529 struct btrfs_root *root, struct inode *dir,
1530 u64 index)
1531{
1532 struct btrfs_delayed_node *node;
1533 struct btrfs_delayed_item *item;
1534 struct btrfs_key item_key;
1535 int ret;
1536
1537 node = btrfs_get_or_create_delayed_node(dir);
1538 if (IS_ERR(node))
1539 return PTR_ERR(node);
1540
1541 item_key.objectid = btrfs_ino(dir);
1542 btrfs_set_key_type(&item_key, BTRFS_DIR_INDEX_KEY);
1543 item_key.offset = index;
1544
1545 ret = btrfs_delete_delayed_insertion_item(root, node, &item_key);
1546 if (!ret)
1547 goto end;
1548
1549 item = btrfs_alloc_delayed_item(0);
1550 if (!item) {
1551 ret = -ENOMEM;
1552 goto end;
1553 }
1554
1555 item->key = item_key;
1556
1557 ret = btrfs_delayed_item_reserve_metadata(trans, root, item);
1558 /*
1559 * we have reserved enough space when we start a new transaction,
1560 * so reserving metadata failure is impossible.
1561 */
1562 BUG_ON(ret);
1563
1564 mutex_lock(&node->mutex);
1565 ret = __btrfs_add_delayed_deletion_item(node, item);
1566 if (unlikely(ret)) {
1567 btrfs_err(root->fs_info, "err add delayed dir index item(index: %llu) "
1568 "into the deletion tree of the delayed node"
1569 "(root id: %llu, inode id: %llu, errno: %d)",
1570 index, node->root->objectid, node->inode_id,
1571 ret);
1572 BUG();
1573 }
1574 mutex_unlock(&node->mutex);
1575end:
1576 btrfs_release_delayed_node(node);
1577 return ret;
1578}
1579
1580int btrfs_inode_delayed_dir_index_count(struct inode *inode)
1581{
1582 struct btrfs_delayed_node *delayed_node = btrfs_get_delayed_node(inode);
1583
1584 if (!delayed_node)
1585 return -ENOENT;
1586
1587 /*
1588 * Since we have held i_mutex of this directory, it is impossible that
1589 * a new directory index is added into the delayed node and index_cnt
1590 * is updated now. So we needn't lock the delayed node.
1591 */
1592 if (!delayed_node->index_cnt) {
1593 btrfs_release_delayed_node(delayed_node);
1594 return -EINVAL;
1595 }
1596
1597 BTRFS_I(inode)->index_cnt = delayed_node->index_cnt;
1598 btrfs_release_delayed_node(delayed_node);
1599 return 0;
1600}
1601
1602void btrfs_get_delayed_items(struct inode *inode, struct list_head *ins_list,
1603 struct list_head *del_list)
1604{
1605 struct btrfs_delayed_node *delayed_node;
1606 struct btrfs_delayed_item *item;
1607
1608 delayed_node = btrfs_get_delayed_node(inode);
1609 if (!delayed_node)
1610 return;
1611
1612 mutex_lock(&delayed_node->mutex);
1613 item = __btrfs_first_delayed_insertion_item(delayed_node);
1614 while (item) {
1615 atomic_inc(&item->refs);
1616 list_add_tail(&item->readdir_list, ins_list);
1617 item = __btrfs_next_delayed_item(item);
1618 }
1619
1620 item = __btrfs_first_delayed_deletion_item(delayed_node);
1621 while (item) {
1622 atomic_inc(&item->refs);
1623 list_add_tail(&item->readdir_list, del_list);
1624 item = __btrfs_next_delayed_item(item);
1625 }
1626 mutex_unlock(&delayed_node->mutex);
1627 /*
1628 * This delayed node is still cached in the btrfs inode, so refs
1629 * must be > 1 now, and we needn't check it is going to be freed
1630 * or not.
1631 *
1632 * Besides that, this function is used to read dir, we do not
1633 * insert/delete delayed items in this period. So we also needn't
1634 * requeue or dequeue this delayed node.
1635 */
1636 atomic_dec(&delayed_node->refs);
1637}
1638
1639void btrfs_put_delayed_items(struct list_head *ins_list,
1640 struct list_head *del_list)
1641{
1642 struct btrfs_delayed_item *curr, *next;
1643
1644 list_for_each_entry_safe(curr, next, ins_list, readdir_list) {
1645 list_del(&curr->readdir_list);
1646 if (atomic_dec_and_test(&curr->refs))
1647 kfree(curr);
1648 }
1649
1650 list_for_each_entry_safe(curr, next, del_list, readdir_list) {
1651 list_del(&curr->readdir_list);
1652 if (atomic_dec_and_test(&curr->refs))
1653 kfree(curr);
1654 }
1655}
1656
1657int btrfs_should_delete_dir_index(struct list_head *del_list,
1658 u64 index)
1659{
1660 struct btrfs_delayed_item *curr, *next;
1661 int ret;
1662
1663 if (list_empty(del_list))
1664 return 0;
1665
1666 list_for_each_entry_safe(curr, next, del_list, readdir_list) {
1667 if (curr->key.offset > index)
1668 break;
1669
1670 list_del(&curr->readdir_list);
1671 ret = (curr->key.offset == index);
1672
1673 if (atomic_dec_and_test(&curr->refs))
1674 kfree(curr);
1675
1676 if (ret)
1677 return 1;
1678 else
1679 continue;
1680 }
1681 return 0;
1682}
1683
1684/*
1685 * btrfs_readdir_delayed_dir_index - read dir info stored in the delayed tree
1686 *
1687 */
1688int btrfs_readdir_delayed_dir_index(struct dir_context *ctx,
1689 struct list_head *ins_list)
1690{
1691 struct btrfs_dir_item *di;
1692 struct btrfs_delayed_item *curr, *next;
1693 struct btrfs_key location;
1694 char *name;
1695 int name_len;
1696 int over = 0;
1697 unsigned char d_type;
1698
1699 if (list_empty(ins_list))
1700 return 0;
1701
1702 /*
1703 * Changing the data of the delayed item is impossible. So
1704 * we needn't lock them. And we have held i_mutex of the
1705 * directory, nobody can delete any directory indexes now.
1706 */
1707 list_for_each_entry_safe(curr, next, ins_list, readdir_list) {
1708 list_del(&curr->readdir_list);
1709
1710 if (curr->key.offset < ctx->pos) {
1711 if (atomic_dec_and_test(&curr->refs))
1712 kfree(curr);
1713 continue;
1714 }
1715
1716 ctx->pos = curr->key.offset;
1717
1718 di = (struct btrfs_dir_item *)curr->data;
1719 name = (char *)(di + 1);
1720 name_len = btrfs_stack_dir_name_len(di);
1721
1722 d_type = btrfs_filetype_table[di->type];
1723 btrfs_disk_key_to_cpu(&location, &di->location);
1724
1725 over = !dir_emit(ctx, name, name_len,
1726 location.objectid, d_type);
1727
1728 if (atomic_dec_and_test(&curr->refs))
1729 kfree(curr);
1730
1731 if (over)
1732 return 1;
1733 }
1734 return 0;
1735}
1736
1737static void fill_stack_inode_item(struct btrfs_trans_handle *trans,
1738 struct btrfs_inode_item *inode_item,
1739 struct inode *inode)
1740{
1741 btrfs_set_stack_inode_uid(inode_item, i_uid_read(inode));
1742 btrfs_set_stack_inode_gid(inode_item, i_gid_read(inode));
1743 btrfs_set_stack_inode_size(inode_item, BTRFS_I(inode)->disk_i_size);
1744 btrfs_set_stack_inode_mode(inode_item, inode->i_mode);
1745 btrfs_set_stack_inode_nlink(inode_item, inode->i_nlink);
1746 btrfs_set_stack_inode_nbytes(inode_item, inode_get_bytes(inode));
1747 btrfs_set_stack_inode_generation(inode_item,
1748 BTRFS_I(inode)->generation);
1749 btrfs_set_stack_inode_sequence(inode_item, inode->i_version);
1750 btrfs_set_stack_inode_transid(inode_item, trans->transid);
1751 btrfs_set_stack_inode_rdev(inode_item, inode->i_rdev);
1752 btrfs_set_stack_inode_flags(inode_item, BTRFS_I(inode)->flags);
1753 btrfs_set_stack_inode_block_group(inode_item, 0);
1754
1755 btrfs_set_stack_timespec_sec(btrfs_inode_atime(inode_item),
1756 inode->i_atime.tv_sec);
1757 btrfs_set_stack_timespec_nsec(btrfs_inode_atime(inode_item),
1758 inode->i_atime.tv_nsec);
1759
1760 btrfs_set_stack_timespec_sec(btrfs_inode_mtime(inode_item),
1761 inode->i_mtime.tv_sec);
1762 btrfs_set_stack_timespec_nsec(btrfs_inode_mtime(inode_item),
1763 inode->i_mtime.tv_nsec);
1764
1765 btrfs_set_stack_timespec_sec(btrfs_inode_ctime(inode_item),
1766 inode->i_ctime.tv_sec);
1767 btrfs_set_stack_timespec_nsec(btrfs_inode_ctime(inode_item),
1768 inode->i_ctime.tv_nsec);
1769}
1770
1771int btrfs_fill_inode(struct inode *inode, u32 *rdev)
1772{
1773 struct btrfs_delayed_node *delayed_node;
1774 struct btrfs_inode_item *inode_item;
1775 struct btrfs_timespec *tspec;
1776
1777 delayed_node = btrfs_get_delayed_node(inode);
1778 if (!delayed_node)
1779 return -ENOENT;
1780
1781 mutex_lock(&delayed_node->mutex);
1782 if (!test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
1783 mutex_unlock(&delayed_node->mutex);
1784 btrfs_release_delayed_node(delayed_node);
1785 return -ENOENT;
1786 }
1787
1788 inode_item = &delayed_node->inode_item;
1789
1790 i_uid_write(inode, btrfs_stack_inode_uid(inode_item));
1791 i_gid_write(inode, btrfs_stack_inode_gid(inode_item));
1792 btrfs_i_size_write(inode, btrfs_stack_inode_size(inode_item));
1793 inode->i_mode = btrfs_stack_inode_mode(inode_item);
1794 set_nlink(inode, btrfs_stack_inode_nlink(inode_item));
1795 inode_set_bytes(inode, btrfs_stack_inode_nbytes(inode_item));
1796 BTRFS_I(inode)->generation = btrfs_stack_inode_generation(inode_item);
1797 inode->i_version = btrfs_stack_inode_sequence(inode_item);
1798 inode->i_rdev = 0;
1799 *rdev = btrfs_stack_inode_rdev(inode_item);
1800 BTRFS_I(inode)->flags = btrfs_stack_inode_flags(inode_item);
1801
1802 tspec = btrfs_inode_atime(inode_item);
1803 inode->i_atime.tv_sec = btrfs_stack_timespec_sec(tspec);
1804 inode->i_atime.tv_nsec = btrfs_stack_timespec_nsec(tspec);
1805
1806 tspec = btrfs_inode_mtime(inode_item);
1807 inode->i_mtime.tv_sec = btrfs_stack_timespec_sec(tspec);
1808 inode->i_mtime.tv_nsec = btrfs_stack_timespec_nsec(tspec);
1809
1810 tspec = btrfs_inode_ctime(inode_item);
1811 inode->i_ctime.tv_sec = btrfs_stack_timespec_sec(tspec);
1812 inode->i_ctime.tv_nsec = btrfs_stack_timespec_nsec(tspec);
1813
1814 inode->i_generation = BTRFS_I(inode)->generation;
1815 BTRFS_I(inode)->index_cnt = (u64)-1;
1816
1817 mutex_unlock(&delayed_node->mutex);
1818 btrfs_release_delayed_node(delayed_node);
1819 return 0;
1820}
1821
1822int btrfs_delayed_update_inode(struct btrfs_trans_handle *trans,
1823 struct btrfs_root *root, struct inode *inode)
1824{
1825 struct btrfs_delayed_node *delayed_node;
1826 int ret = 0;
1827
1828 delayed_node = btrfs_get_or_create_delayed_node(inode);
1829 if (IS_ERR(delayed_node))
1830 return PTR_ERR(delayed_node);
1831
1832 mutex_lock(&delayed_node->mutex);
1833 if (test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
1834 fill_stack_inode_item(trans, &delayed_node->inode_item, inode);
1835 goto release_node;
1836 }
1837
1838 ret = btrfs_delayed_inode_reserve_metadata(trans, root, inode,
1839 delayed_node);
1840 if (ret)
1841 goto release_node;
1842
1843 fill_stack_inode_item(trans, &delayed_node->inode_item, inode);
1844 set_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags);
1845 delayed_node->count++;
1846 atomic_inc(&root->fs_info->delayed_root->items);
1847release_node:
1848 mutex_unlock(&delayed_node->mutex);
1849 btrfs_release_delayed_node(delayed_node);
1850 return ret;
1851}
1852
1853int btrfs_delayed_delete_inode_ref(struct inode *inode)
1854{
1855 struct btrfs_delayed_node *delayed_node;
1856
1857 delayed_node = btrfs_get_or_create_delayed_node(inode);
1858 if (IS_ERR(delayed_node))
1859 return PTR_ERR(delayed_node);
1860
1861 /*
1862 * We don't reserve space for inode ref deletion is because:
1863 * - We ONLY do async inode ref deletion for the inode who has only
1864 * one link(i_nlink == 1), it means there is only one inode ref.
1865 * And in most case, the inode ref and the inode item are in the
1866 * same leaf, and we will deal with them at the same time.
1867 * Since we are sure we will reserve the space for the inode item,
1868 * it is unnecessary to reserve space for inode ref deletion.
1869 * - If the inode ref and the inode item are not in the same leaf,
1870 * We also needn't worry about enospc problem, because we reserve
1871 * much more space for the inode update than it needs.
1872 * - At the worst, we can steal some space from the global reservation.
1873 * It is very rare.
1874 */
1875 mutex_lock(&delayed_node->mutex);
1876 if (test_bit(BTRFS_DELAYED_NODE_DEL_IREF, &delayed_node->flags))
1877 goto release_node;
1878
1879 set_bit(BTRFS_DELAYED_NODE_DEL_IREF, &delayed_node->flags);
1880 delayed_node->count++;
1881 atomic_inc(&BTRFS_I(inode)->root->fs_info->delayed_root->items);
1882release_node:
1883 mutex_unlock(&delayed_node->mutex);
1884 btrfs_release_delayed_node(delayed_node);
1885 return 0;
1886}
1887
1888static void __btrfs_kill_delayed_node(struct btrfs_delayed_node *delayed_node)
1889{
1890 struct btrfs_root *root = delayed_node->root;
1891 struct btrfs_delayed_item *curr_item, *prev_item;
1892
1893 mutex_lock(&delayed_node->mutex);
1894 curr_item = __btrfs_first_delayed_insertion_item(delayed_node);
1895 while (curr_item) {
1896 btrfs_delayed_item_release_metadata(root, curr_item);
1897 prev_item = curr_item;
1898 curr_item = __btrfs_next_delayed_item(prev_item);
1899 btrfs_release_delayed_item(prev_item);
1900 }
1901
1902 curr_item = __btrfs_first_delayed_deletion_item(delayed_node);
1903 while (curr_item) {
1904 btrfs_delayed_item_release_metadata(root, curr_item);
1905 prev_item = curr_item;
1906 curr_item = __btrfs_next_delayed_item(prev_item);
1907 btrfs_release_delayed_item(prev_item);
1908 }
1909
1910 if (test_bit(BTRFS_DELAYED_NODE_DEL_IREF, &delayed_node->flags))
1911 btrfs_release_delayed_iref(delayed_node);
1912
1913 if (test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
1914 btrfs_delayed_inode_release_metadata(root, delayed_node);
1915 btrfs_release_delayed_inode(delayed_node);
1916 }
1917 mutex_unlock(&delayed_node->mutex);
1918}
1919
1920void btrfs_kill_delayed_inode_items(struct inode *inode)
1921{
1922 struct btrfs_delayed_node *delayed_node;
1923
1924 delayed_node = btrfs_get_delayed_node(inode);
1925 if (!delayed_node)
1926 return;
1927
1928 __btrfs_kill_delayed_node(delayed_node);
1929 btrfs_release_delayed_node(delayed_node);
1930}
1931
1932void btrfs_kill_all_delayed_nodes(struct btrfs_root *root)
1933{
1934 u64 inode_id = 0;
1935 struct btrfs_delayed_node *delayed_nodes[8];
1936 int i, n;
1937
1938 while (1) {
1939 spin_lock(&root->inode_lock);
1940 n = radix_tree_gang_lookup(&root->delayed_nodes_tree,
1941 (void **)delayed_nodes, inode_id,
1942 ARRAY_SIZE(delayed_nodes));
1943 if (!n) {
1944 spin_unlock(&root->inode_lock);
1945 break;
1946 }
1947
1948 inode_id = delayed_nodes[n - 1]->inode_id + 1;
1949
1950 for (i = 0; i < n; i++)
1951 atomic_inc(&delayed_nodes[i]->refs);
1952 spin_unlock(&root->inode_lock);
1953
1954 for (i = 0; i < n; i++) {
1955 __btrfs_kill_delayed_node(delayed_nodes[i]);
1956 btrfs_release_delayed_node(delayed_nodes[i]);
1957 }
1958 }
1959}
1960
1961void btrfs_destroy_delayed_inodes(struct btrfs_root *root)
1962{
1963 struct btrfs_delayed_root *delayed_root;
1964 struct btrfs_delayed_node *curr_node, *prev_node;
1965
1966 delayed_root = btrfs_get_delayed_root(root);
1967
1968 curr_node = btrfs_first_delayed_node(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}