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 kmem_cache_destroy(delayed_node_cache);
47}
48
49static inline void btrfs_init_delayed_node(
50 struct btrfs_delayed_node *delayed_node,
51 struct btrfs_root *root, u64 inode_id)
52{
53 delayed_node->root = root;
54 delayed_node->inode_id = inode_id;
55 atomic_set(&delayed_node->refs, 0);
56 delayed_node->ins_root = RB_ROOT;
57 delayed_node->del_root = RB_ROOT;
58 mutex_init(&delayed_node->mutex);
59 INIT_LIST_HEAD(&delayed_node->n_list);
60 INIT_LIST_HEAD(&delayed_node->p_list);
61}
62
63static inline int btrfs_is_continuous_delayed_item(
64 struct btrfs_delayed_item *item1,
65 struct btrfs_delayed_item *item2)
66{
67 if (item1->key.type == BTRFS_DIR_INDEX_KEY &&
68 item1->key.objectid == item2->key.objectid &&
69 item1->key.type == item2->key.type &&
70 item1->key.offset + 1 == item2->key.offset)
71 return 1;
72 return 0;
73}
74
75static inline struct btrfs_delayed_root *btrfs_get_delayed_root(
76 struct btrfs_root *root)
77{
78 return root->fs_info->delayed_root;
79}
80
81static struct btrfs_delayed_node *btrfs_get_delayed_node(struct inode *inode)
82{
83 struct btrfs_inode *btrfs_inode = BTRFS_I(inode);
84 struct btrfs_root *root = btrfs_inode->root;
85 u64 ino = btrfs_ino(inode);
86 struct btrfs_delayed_node *node;
87
88 node = ACCESS_ONCE(btrfs_inode->delayed_node);
89 if (node) {
90 atomic_inc(&node->refs);
91 return node;
92 }
93
94 spin_lock(&root->inode_lock);
95 node = radix_tree_lookup(&root->delayed_nodes_tree, ino);
96 if (node) {
97 if (btrfs_inode->delayed_node) {
98 atomic_inc(&node->refs); /* can be accessed */
99 BUG_ON(btrfs_inode->delayed_node != node);
100 spin_unlock(&root->inode_lock);
101 return node;
102 }
103 btrfs_inode->delayed_node = node;
104 /* can be accessed and cached in the inode */
105 atomic_add(2, &node->refs);
106 spin_unlock(&root->inode_lock);
107 return node;
108 }
109 spin_unlock(&root->inode_lock);
110
111 return NULL;
112}
113
114/* Will return either the node or PTR_ERR(-ENOMEM) */
115static struct btrfs_delayed_node *btrfs_get_or_create_delayed_node(
116 struct inode *inode)
117{
118 struct btrfs_delayed_node *node;
119 struct btrfs_inode *btrfs_inode = BTRFS_I(inode);
120 struct btrfs_root *root = btrfs_inode->root;
121 u64 ino = btrfs_ino(inode);
122 int ret;
123
124again:
125 node = btrfs_get_delayed_node(inode);
126 if (node)
127 return node;
128
129 node = kmem_cache_zalloc(delayed_node_cache, GFP_NOFS);
130 if (!node)
131 return ERR_PTR(-ENOMEM);
132 btrfs_init_delayed_node(node, root, ino);
133
134 /* cached in the btrfs inode and can be accessed */
135 atomic_add(2, &node->refs);
136
137 ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
138 if (ret) {
139 kmem_cache_free(delayed_node_cache, node);
140 return ERR_PTR(ret);
141 }
142
143 spin_lock(&root->inode_lock);
144 ret = radix_tree_insert(&root->delayed_nodes_tree, ino, node);
145 if (ret == -EEXIST) {
146 spin_unlock(&root->inode_lock);
147 kmem_cache_free(delayed_node_cache, node);
148 radix_tree_preload_end();
149 goto again;
150 }
151 btrfs_inode->delayed_node = node;
152 spin_unlock(&root->inode_lock);
153 radix_tree_preload_end();
154
155 return node;
156}
157
158/*
159 * Call it when holding delayed_node->mutex
160 *
161 * If mod = 1, add this node into the prepared list.
162 */
163static void btrfs_queue_delayed_node(struct btrfs_delayed_root *root,
164 struct btrfs_delayed_node *node,
165 int mod)
166{
167 spin_lock(&root->lock);
168 if (test_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags)) {
169 if (!list_empty(&node->p_list))
170 list_move_tail(&node->p_list, &root->prepare_list);
171 else if (mod)
172 list_add_tail(&node->p_list, &root->prepare_list);
173 } else {
174 list_add_tail(&node->n_list, &root->node_list);
175 list_add_tail(&node->p_list, &root->prepare_list);
176 atomic_inc(&node->refs); /* inserted into list */
177 root->nodes++;
178 set_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags);
179 }
180 spin_unlock(&root->lock);
181}
182
183/* Call it when holding delayed_node->mutex */
184static void btrfs_dequeue_delayed_node(struct btrfs_delayed_root *root,
185 struct btrfs_delayed_node *node)
186{
187 spin_lock(&root->lock);
188 if (test_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags)) {
189 root->nodes--;
190 atomic_dec(&node->refs); /* not in the list */
191 list_del_init(&node->n_list);
192 if (!list_empty(&node->p_list))
193 list_del_init(&node->p_list);
194 clear_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags);
195 }
196 spin_unlock(&root->lock);
197}
198
199static struct btrfs_delayed_node *btrfs_first_delayed_node(
200 struct btrfs_delayed_root *delayed_root)
201{
202 struct list_head *p;
203 struct btrfs_delayed_node *node = NULL;
204
205 spin_lock(&delayed_root->lock);
206 if (list_empty(&delayed_root->node_list))
207 goto out;
208
209 p = delayed_root->node_list.next;
210 node = list_entry(p, struct btrfs_delayed_node, n_list);
211 atomic_inc(&node->refs);
212out:
213 spin_unlock(&delayed_root->lock);
214
215 return node;
216}
217
218static struct btrfs_delayed_node *btrfs_next_delayed_node(
219 struct btrfs_delayed_node *node)
220{
221 struct btrfs_delayed_root *delayed_root;
222 struct list_head *p;
223 struct btrfs_delayed_node *next = NULL;
224
225 delayed_root = node->root->fs_info->delayed_root;
226 spin_lock(&delayed_root->lock);
227 if (!test_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags)) {
228 /* not in the list */
229 if (list_empty(&delayed_root->node_list))
230 goto out;
231 p = delayed_root->node_list.next;
232 } else if (list_is_last(&node->n_list, &delayed_root->node_list))
233 goto out;
234 else
235 p = node->n_list.next;
236
237 next = list_entry(p, struct btrfs_delayed_node, n_list);
238 atomic_inc(&next->refs);
239out:
240 spin_unlock(&delayed_root->lock);
241
242 return next;
243}
244
245static void __btrfs_release_delayed_node(
246 struct btrfs_delayed_node *delayed_node,
247 int mod)
248{
249 struct btrfs_delayed_root *delayed_root;
250
251 if (!delayed_node)
252 return;
253
254 delayed_root = delayed_node->root->fs_info->delayed_root;
255
256 mutex_lock(&delayed_node->mutex);
257 if (delayed_node->count)
258 btrfs_queue_delayed_node(delayed_root, delayed_node, mod);
259 else
260 btrfs_dequeue_delayed_node(delayed_root, delayed_node);
261 mutex_unlock(&delayed_node->mutex);
262
263 if (atomic_dec_and_test(&delayed_node->refs)) {
264 bool free = false;
265 struct btrfs_root *root = delayed_node->root;
266 spin_lock(&root->inode_lock);
267 if (atomic_read(&delayed_node->refs) == 0) {
268 radix_tree_delete(&root->delayed_nodes_tree,
269 delayed_node->inode_id);
270 free = true;
271 }
272 spin_unlock(&root->inode_lock);
273 if (free)
274 kmem_cache_free(delayed_node_cache, delayed_node);
275 }
276}
277
278static inline void btrfs_release_delayed_node(struct btrfs_delayed_node *node)
279{
280 __btrfs_release_delayed_node(node, 0);
281}
282
283static struct btrfs_delayed_node *btrfs_first_prepared_delayed_node(
284 struct btrfs_delayed_root *delayed_root)
285{
286 struct list_head *p;
287 struct btrfs_delayed_node *node = NULL;
288
289 spin_lock(&delayed_root->lock);
290 if (list_empty(&delayed_root->prepare_list))
291 goto out;
292
293 p = delayed_root->prepare_list.next;
294 list_del_init(p);
295 node = list_entry(p, struct btrfs_delayed_node, p_list);
296 atomic_inc(&node->refs);
297out:
298 spin_unlock(&delayed_root->lock);
299
300 return node;
301}
302
303static inline void btrfs_release_prepared_delayed_node(
304 struct btrfs_delayed_node *node)
305{
306 __btrfs_release_delayed_node(node, 1);
307}
308
309static struct btrfs_delayed_item *btrfs_alloc_delayed_item(u32 data_len)
310{
311 struct btrfs_delayed_item *item;
312 item = kmalloc(sizeof(*item) + data_len, GFP_NOFS);
313 if (item) {
314 item->data_len = data_len;
315 item->ins_or_del = 0;
316 item->bytes_reserved = 0;
317 item->delayed_node = NULL;
318 atomic_set(&item->refs, 1);
319 }
320 return item;
321}
322
323/*
324 * __btrfs_lookup_delayed_item - look up the delayed item by key
325 * @delayed_node: pointer to the delayed node
326 * @key: the key to look up
327 * @prev: used to store the prev item if the right item isn't found
328 * @next: used to store the next item if the right item isn't found
329 *
330 * Note: if we don't find the right item, we will return the prev item and
331 * the next item.
332 */
333static struct btrfs_delayed_item *__btrfs_lookup_delayed_item(
334 struct rb_root *root,
335 struct btrfs_key *key,
336 struct btrfs_delayed_item **prev,
337 struct btrfs_delayed_item **next)
338{
339 struct rb_node *node, *prev_node = NULL;
340 struct btrfs_delayed_item *delayed_item = NULL;
341 int ret = 0;
342
343 node = root->rb_node;
344
345 while (node) {
346 delayed_item = rb_entry(node, struct btrfs_delayed_item,
347 rb_node);
348 prev_node = node;
349 ret = btrfs_comp_cpu_keys(&delayed_item->key, key);
350 if (ret < 0)
351 node = node->rb_right;
352 else if (ret > 0)
353 node = node->rb_left;
354 else
355 return delayed_item;
356 }
357
358 if (prev) {
359 if (!prev_node)
360 *prev = NULL;
361 else if (ret < 0)
362 *prev = delayed_item;
363 else if ((node = rb_prev(prev_node)) != NULL) {
364 *prev = rb_entry(node, struct btrfs_delayed_item,
365 rb_node);
366 } else
367 *prev = NULL;
368 }
369
370 if (next) {
371 if (!prev_node)
372 *next = NULL;
373 else if (ret > 0)
374 *next = delayed_item;
375 else if ((node = rb_next(prev_node)) != NULL) {
376 *next = rb_entry(node, struct btrfs_delayed_item,
377 rb_node);
378 } else
379 *next = NULL;
380 }
381 return NULL;
382}
383
384static struct btrfs_delayed_item *__btrfs_lookup_delayed_insertion_item(
385 struct btrfs_delayed_node *delayed_node,
386 struct btrfs_key *key)
387{
388 struct btrfs_delayed_item *item;
389
390 item = __btrfs_lookup_delayed_item(&delayed_node->ins_root, key,
391 NULL, NULL);
392 return item;
393}
394
395static int __btrfs_add_delayed_item(struct btrfs_delayed_node *delayed_node,
396 struct btrfs_delayed_item *ins,
397 int action)
398{
399 struct rb_node **p, *node;
400 struct rb_node *parent_node = NULL;
401 struct rb_root *root;
402 struct btrfs_delayed_item *item;
403 int cmp;
404
405 if (action == BTRFS_DELAYED_INSERTION_ITEM)
406 root = &delayed_node->ins_root;
407 else if (action == BTRFS_DELAYED_DELETION_ITEM)
408 root = &delayed_node->del_root;
409 else
410 BUG();
411 p = &root->rb_node;
412 node = &ins->rb_node;
413
414 while (*p) {
415 parent_node = *p;
416 item = rb_entry(parent_node, struct btrfs_delayed_item,
417 rb_node);
418
419 cmp = btrfs_comp_cpu_keys(&item->key, &ins->key);
420 if (cmp < 0)
421 p = &(*p)->rb_right;
422 else if (cmp > 0)
423 p = &(*p)->rb_left;
424 else
425 return -EEXIST;
426 }
427
428 rb_link_node(node, parent_node, p);
429 rb_insert_color(node, root);
430 ins->delayed_node = delayed_node;
431 ins->ins_or_del = action;
432
433 if (ins->key.type == BTRFS_DIR_INDEX_KEY &&
434 action == BTRFS_DELAYED_INSERTION_ITEM &&
435 ins->key.offset >= delayed_node->index_cnt)
436 delayed_node->index_cnt = ins->key.offset + 1;
437
438 delayed_node->count++;
439 atomic_inc(&delayed_node->root->fs_info->delayed_root->items);
440 return 0;
441}
442
443static int __btrfs_add_delayed_insertion_item(struct btrfs_delayed_node *node,
444 struct btrfs_delayed_item *item)
445{
446 return __btrfs_add_delayed_item(node, item,
447 BTRFS_DELAYED_INSERTION_ITEM);
448}
449
450static int __btrfs_add_delayed_deletion_item(struct btrfs_delayed_node *node,
451 struct btrfs_delayed_item *item)
452{
453 return __btrfs_add_delayed_item(node, item,
454 BTRFS_DELAYED_DELETION_ITEM);
455}
456
457static void finish_one_item(struct btrfs_delayed_root *delayed_root)
458{
459 int seq = atomic_inc_return(&delayed_root->items_seq);
460
461 /*
462 * atomic_dec_return implies a barrier for waitqueue_active
463 */
464 if ((atomic_dec_return(&delayed_root->items) <
465 BTRFS_DELAYED_BACKGROUND || seq % BTRFS_DELAYED_BATCH == 0) &&
466 waitqueue_active(&delayed_root->wait))
467 wake_up(&delayed_root->wait);
468}
469
470static void __btrfs_remove_delayed_item(struct btrfs_delayed_item *delayed_item)
471{
472 struct rb_root *root;
473 struct btrfs_delayed_root *delayed_root;
474
475 delayed_root = delayed_item->delayed_node->root->fs_info->delayed_root;
476
477 BUG_ON(!delayed_root);
478 BUG_ON(delayed_item->ins_or_del != BTRFS_DELAYED_DELETION_ITEM &&
479 delayed_item->ins_or_del != BTRFS_DELAYED_INSERTION_ITEM);
480
481 if (delayed_item->ins_or_del == BTRFS_DELAYED_INSERTION_ITEM)
482 root = &delayed_item->delayed_node->ins_root;
483 else
484 root = &delayed_item->delayed_node->del_root;
485
486 rb_erase(&delayed_item->rb_node, root);
487 delayed_item->delayed_node->count--;
488
489 finish_one_item(delayed_root);
490}
491
492static void btrfs_release_delayed_item(struct btrfs_delayed_item *item)
493{
494 if (item) {
495 __btrfs_remove_delayed_item(item);
496 if (atomic_dec_and_test(&item->refs))
497 kfree(item);
498 }
499}
500
501static struct btrfs_delayed_item *__btrfs_first_delayed_insertion_item(
502 struct btrfs_delayed_node *delayed_node)
503{
504 struct rb_node *p;
505 struct btrfs_delayed_item *item = NULL;
506
507 p = rb_first(&delayed_node->ins_root);
508 if (p)
509 item = rb_entry(p, struct btrfs_delayed_item, rb_node);
510
511 return item;
512}
513
514static struct btrfs_delayed_item *__btrfs_first_delayed_deletion_item(
515 struct btrfs_delayed_node *delayed_node)
516{
517 struct rb_node *p;
518 struct btrfs_delayed_item *item = NULL;
519
520 p = rb_first(&delayed_node->del_root);
521 if (p)
522 item = rb_entry(p, struct btrfs_delayed_item, rb_node);
523
524 return item;
525}
526
527static struct btrfs_delayed_item *__btrfs_next_delayed_item(
528 struct btrfs_delayed_item *item)
529{
530 struct rb_node *p;
531 struct btrfs_delayed_item *next = NULL;
532
533 p = rb_next(&item->rb_node);
534 if (p)
535 next = rb_entry(p, struct btrfs_delayed_item, rb_node);
536
537 return next;
538}
539
540static int btrfs_delayed_item_reserve_metadata(struct btrfs_trans_handle *trans,
541 struct btrfs_root *root,
542 struct btrfs_delayed_item *item)
543{
544 struct btrfs_block_rsv *src_rsv;
545 struct btrfs_block_rsv *dst_rsv;
546 u64 num_bytes;
547 int ret;
548
549 if (!trans->bytes_reserved)
550 return 0;
551
552 src_rsv = trans->block_rsv;
553 dst_rsv = &root->fs_info->delayed_block_rsv;
554
555 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
556 ret = btrfs_block_rsv_migrate(src_rsv, dst_rsv, num_bytes);
557 if (!ret) {
558 trace_btrfs_space_reservation(root->fs_info, "delayed_item",
559 item->key.objectid,
560 num_bytes, 1);
561 item->bytes_reserved = num_bytes;
562 }
563
564 return ret;
565}
566
567static void btrfs_delayed_item_release_metadata(struct btrfs_root *root,
568 struct btrfs_delayed_item *item)
569{
570 struct btrfs_block_rsv *rsv;
571
572 if (!item->bytes_reserved)
573 return;
574
575 rsv = &root->fs_info->delayed_block_rsv;
576 trace_btrfs_space_reservation(root->fs_info, "delayed_item",
577 item->key.objectid, item->bytes_reserved,
578 0);
579 btrfs_block_rsv_release(root, rsv,
580 item->bytes_reserved);
581}
582
583static int btrfs_delayed_inode_reserve_metadata(
584 struct btrfs_trans_handle *trans,
585 struct btrfs_root *root,
586 struct inode *inode,
587 struct btrfs_delayed_node *node)
588{
589 struct btrfs_block_rsv *src_rsv;
590 struct btrfs_block_rsv *dst_rsv;
591 u64 num_bytes;
592 int ret;
593 bool release = false;
594
595 src_rsv = trans->block_rsv;
596 dst_rsv = &root->fs_info->delayed_block_rsv;
597
598 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
599
600 /*
601 * btrfs_dirty_inode will update the inode under btrfs_join_transaction
602 * which doesn't reserve space for speed. This is a problem since we
603 * still need to reserve space for this update, so try to reserve the
604 * space.
605 *
606 * Now if src_rsv == delalloc_block_rsv we'll let it just steal since
607 * we're accounted for.
608 */
609 if (!src_rsv || (!trans->bytes_reserved &&
610 src_rsv->type != BTRFS_BLOCK_RSV_DELALLOC)) {
611 ret = btrfs_block_rsv_add(root, dst_rsv, num_bytes,
612 BTRFS_RESERVE_NO_FLUSH);
613 /*
614 * Since we're under a transaction reserve_metadata_bytes could
615 * try to commit the transaction which will make it return
616 * EAGAIN to make us stop the transaction we have, so return
617 * ENOSPC instead so that btrfs_dirty_inode knows what to do.
618 */
619 if (ret == -EAGAIN)
620 ret = -ENOSPC;
621 if (!ret) {
622 node->bytes_reserved = num_bytes;
623 trace_btrfs_space_reservation(root->fs_info,
624 "delayed_inode",
625 btrfs_ino(inode),
626 num_bytes, 1);
627 }
628 return ret;
629 } else if (src_rsv->type == BTRFS_BLOCK_RSV_DELALLOC) {
630 spin_lock(&BTRFS_I(inode)->lock);
631 if (test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
632 &BTRFS_I(inode)->runtime_flags)) {
633 spin_unlock(&BTRFS_I(inode)->lock);
634 release = true;
635 goto migrate;
636 }
637 spin_unlock(&BTRFS_I(inode)->lock);
638
639 /* Ok we didn't have space pre-reserved. This shouldn't happen
640 * too often but it can happen if we do delalloc to an existing
641 * inode which gets dirtied because of the time update, and then
642 * isn't touched again until after the transaction commits and
643 * then we try to write out the data. First try to be nice and
644 * reserve something strictly for us. If not be a pain and try
645 * to steal from the delalloc block rsv.
646 */
647 ret = btrfs_block_rsv_add(root, dst_rsv, num_bytes,
648 BTRFS_RESERVE_NO_FLUSH);
649 if (!ret)
650 goto out;
651
652 ret = btrfs_block_rsv_migrate(src_rsv, dst_rsv, num_bytes);
653 if (!ret)
654 goto out;
655
656 if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
657 btrfs_debug(root->fs_info,
658 "block rsv migrate returned %d", ret);
659 WARN_ON(1);
660 }
661 /*
662 * Ok this is a problem, let's just steal from the global rsv
663 * since this really shouldn't happen that often.
664 */
665 ret = btrfs_block_rsv_migrate(&root->fs_info->global_block_rsv,
666 dst_rsv, num_bytes);
667 goto out;
668 }
669
670migrate:
671 ret = btrfs_block_rsv_migrate(src_rsv, dst_rsv, num_bytes);
672
673out:
674 /*
675 * Migrate only takes a reservation, it doesn't touch the size of the
676 * block_rsv. This is to simplify people who don't normally have things
677 * migrated from their block rsv. If they go to release their
678 * reservation, that will decrease the size as well, so if migrate
679 * reduced size we'd end up with a negative size. But for the
680 * delalloc_meta_reserved stuff we will only know to drop 1 reservation,
681 * but we could in fact do this reserve/migrate dance several times
682 * between the time we did the original reservation and we'd clean it
683 * up. So to take care of this, release the space for the meta
684 * reservation here. I think it may be time for a documentation page on
685 * how block rsvs. work.
686 */
687 if (!ret) {
688 trace_btrfs_space_reservation(root->fs_info, "delayed_inode",
689 btrfs_ino(inode), num_bytes, 1);
690 node->bytes_reserved = num_bytes;
691 }
692
693 if (release) {
694 trace_btrfs_space_reservation(root->fs_info, "delalloc",
695 btrfs_ino(inode), num_bytes, 0);
696 btrfs_block_rsv_release(root, src_rsv, num_bytes);
697 }
698
699 return ret;
700}
701
702static void btrfs_delayed_inode_release_metadata(struct btrfs_root *root,
703 struct btrfs_delayed_node *node)
704{
705 struct btrfs_block_rsv *rsv;
706
707 if (!node->bytes_reserved)
708 return;
709
710 rsv = &root->fs_info->delayed_block_rsv;
711 trace_btrfs_space_reservation(root->fs_info, "delayed_inode",
712 node->inode_id, node->bytes_reserved, 0);
713 btrfs_block_rsv_release(root, rsv,
714 node->bytes_reserved);
715 node->bytes_reserved = 0;
716}
717
718/*
719 * This helper will insert some continuous items into the same leaf according
720 * to the free space of the leaf.
721 */
722static int btrfs_batch_insert_items(struct btrfs_root *root,
723 struct btrfs_path *path,
724 struct btrfs_delayed_item *item)
725{
726 struct btrfs_delayed_item *curr, *next;
727 int free_space;
728 int total_data_size = 0, total_size = 0;
729 struct extent_buffer *leaf;
730 char *data_ptr;
731 struct btrfs_key *keys;
732 u32 *data_size;
733 struct list_head head;
734 int slot;
735 int nitems;
736 int i;
737 int ret = 0;
738
739 BUG_ON(!path->nodes[0]);
740
741 leaf = path->nodes[0];
742 free_space = btrfs_leaf_free_space(root, leaf);
743 INIT_LIST_HEAD(&head);
744
745 next = item;
746 nitems = 0;
747
748 /*
749 * count the number of the continuous items that we can insert in batch
750 */
751 while (total_size + next->data_len + sizeof(struct btrfs_item) <=
752 free_space) {
753 total_data_size += next->data_len;
754 total_size += next->data_len + sizeof(struct btrfs_item);
755 list_add_tail(&next->tree_list, &head);
756 nitems++;
757
758 curr = next;
759 next = __btrfs_next_delayed_item(curr);
760 if (!next)
761 break;
762
763 if (!btrfs_is_continuous_delayed_item(curr, next))
764 break;
765 }
766
767 if (!nitems) {
768 ret = 0;
769 goto out;
770 }
771
772 /*
773 * we need allocate some memory space, but it might cause the task
774 * to sleep, so we set all locked nodes in the path to blocking locks
775 * first.
776 */
777 btrfs_set_path_blocking(path);
778
779 keys = kmalloc_array(nitems, sizeof(struct btrfs_key), GFP_NOFS);
780 if (!keys) {
781 ret = -ENOMEM;
782 goto out;
783 }
784
785 data_size = kmalloc_array(nitems, sizeof(u32), GFP_NOFS);
786 if (!data_size) {
787 ret = -ENOMEM;
788 goto error;
789 }
790
791 /* get keys of all the delayed items */
792 i = 0;
793 list_for_each_entry(next, &head, tree_list) {
794 keys[i] = next->key;
795 data_size[i] = next->data_len;
796 i++;
797 }
798
799 /* reset all the locked nodes in the patch to spinning locks. */
800 btrfs_clear_path_blocking(path, NULL, 0);
801
802 /* insert the keys of the items */
803 setup_items_for_insert(root, path, keys, data_size,
804 total_data_size, total_size, nitems);
805
806 /* insert the dir index items */
807 slot = path->slots[0];
808 list_for_each_entry_safe(curr, next, &head, tree_list) {
809 data_ptr = btrfs_item_ptr(leaf, slot, char);
810 write_extent_buffer(leaf, &curr->data,
811 (unsigned long)data_ptr,
812 curr->data_len);
813 slot++;
814
815 btrfs_delayed_item_release_metadata(root, curr);
816
817 list_del(&curr->tree_list);
818 btrfs_release_delayed_item(curr);
819 }
820
821error:
822 kfree(data_size);
823 kfree(keys);
824out:
825 return ret;
826}
827
828/*
829 * This helper can just do simple insertion that needn't extend item for new
830 * data, such as directory name index insertion, inode insertion.
831 */
832static int btrfs_insert_delayed_item(struct btrfs_trans_handle *trans,
833 struct btrfs_root *root,
834 struct btrfs_path *path,
835 struct btrfs_delayed_item *delayed_item)
836{
837 struct extent_buffer *leaf;
838 char *ptr;
839 int ret;
840
841 ret = btrfs_insert_empty_item(trans, root, path, &delayed_item->key,
842 delayed_item->data_len);
843 if (ret < 0 && ret != -EEXIST)
844 return ret;
845
846 leaf = path->nodes[0];
847
848 ptr = btrfs_item_ptr(leaf, path->slots[0], char);
849
850 write_extent_buffer(leaf, delayed_item->data, (unsigned long)ptr,
851 delayed_item->data_len);
852 btrfs_mark_buffer_dirty(leaf);
853
854 btrfs_delayed_item_release_metadata(root, delayed_item);
855 return 0;
856}
857
858/*
859 * we insert an item first, then if there are some continuous items, we try
860 * to insert those items into the same leaf.
861 */
862static int btrfs_insert_delayed_items(struct btrfs_trans_handle *trans,
863 struct btrfs_path *path,
864 struct btrfs_root *root,
865 struct btrfs_delayed_node *node)
866{
867 struct btrfs_delayed_item *curr, *prev;
868 int ret = 0;
869
870do_again:
871 mutex_lock(&node->mutex);
872 curr = __btrfs_first_delayed_insertion_item(node);
873 if (!curr)
874 goto insert_end;
875
876 ret = btrfs_insert_delayed_item(trans, root, path, curr);
877 if (ret < 0) {
878 btrfs_release_path(path);
879 goto insert_end;
880 }
881
882 prev = curr;
883 curr = __btrfs_next_delayed_item(prev);
884 if (curr && btrfs_is_continuous_delayed_item(prev, curr)) {
885 /* insert the continuous items into the same leaf */
886 path->slots[0]++;
887 btrfs_batch_insert_items(root, path, curr);
888 }
889 btrfs_release_delayed_item(prev);
890 btrfs_mark_buffer_dirty(path->nodes[0]);
891
892 btrfs_release_path(path);
893 mutex_unlock(&node->mutex);
894 goto do_again;
895
896insert_end:
897 mutex_unlock(&node->mutex);
898 return ret;
899}
900
901static int btrfs_batch_delete_items(struct btrfs_trans_handle *trans,
902 struct btrfs_root *root,
903 struct btrfs_path *path,
904 struct btrfs_delayed_item *item)
905{
906 struct btrfs_delayed_item *curr, *next;
907 struct extent_buffer *leaf;
908 struct btrfs_key key;
909 struct list_head head;
910 int nitems, i, last_item;
911 int ret = 0;
912
913 BUG_ON(!path->nodes[0]);
914
915 leaf = path->nodes[0];
916
917 i = path->slots[0];
918 last_item = btrfs_header_nritems(leaf) - 1;
919 if (i > last_item)
920 return -ENOENT; /* FIXME: Is errno suitable? */
921
922 next = item;
923 INIT_LIST_HEAD(&head);
924 btrfs_item_key_to_cpu(leaf, &key, i);
925 nitems = 0;
926 /*
927 * count the number of the dir index items that we can delete in batch
928 */
929 while (btrfs_comp_cpu_keys(&next->key, &key) == 0) {
930 list_add_tail(&next->tree_list, &head);
931 nitems++;
932
933 curr = next;
934 next = __btrfs_next_delayed_item(curr);
935 if (!next)
936 break;
937
938 if (!btrfs_is_continuous_delayed_item(curr, next))
939 break;
940
941 i++;
942 if (i > last_item)
943 break;
944 btrfs_item_key_to_cpu(leaf, &key, i);
945 }
946
947 if (!nitems)
948 return 0;
949
950 ret = btrfs_del_items(trans, root, path, path->slots[0], nitems);
951 if (ret)
952 goto out;
953
954 list_for_each_entry_safe(curr, next, &head, tree_list) {
955 btrfs_delayed_item_release_metadata(root, curr);
956 list_del(&curr->tree_list);
957 btrfs_release_delayed_item(curr);
958 }
959
960out:
961 return ret;
962}
963
964static int btrfs_delete_delayed_items(struct btrfs_trans_handle *trans,
965 struct btrfs_path *path,
966 struct btrfs_root *root,
967 struct btrfs_delayed_node *node)
968{
969 struct btrfs_delayed_item *curr, *prev;
970 int ret = 0;
971
972do_again:
973 mutex_lock(&node->mutex);
974 curr = __btrfs_first_delayed_deletion_item(node);
975 if (!curr)
976 goto delete_fail;
977
978 ret = btrfs_search_slot(trans, root, &curr->key, path, -1, 1);
979 if (ret < 0)
980 goto delete_fail;
981 else if (ret > 0) {
982 /*
983 * can't find the item which the node points to, so this node
984 * is invalid, just drop it.
985 */
986 prev = curr;
987 curr = __btrfs_next_delayed_item(prev);
988 btrfs_release_delayed_item(prev);
989 ret = 0;
990 btrfs_release_path(path);
991 if (curr) {
992 mutex_unlock(&node->mutex);
993 goto do_again;
994 } else
995 goto delete_fail;
996 }
997
998 btrfs_batch_delete_items(trans, root, path, curr);
999 btrfs_release_path(path);
1000 mutex_unlock(&node->mutex);
1001 goto do_again;
1002
1003delete_fail:
1004 btrfs_release_path(path);
1005 mutex_unlock(&node->mutex);
1006 return ret;
1007}
1008
1009static void btrfs_release_delayed_inode(struct btrfs_delayed_node *delayed_node)
1010{
1011 struct btrfs_delayed_root *delayed_root;
1012
1013 if (delayed_node &&
1014 test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
1015 BUG_ON(!delayed_node->root);
1016 clear_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags);
1017 delayed_node->count--;
1018
1019 delayed_root = delayed_node->root->fs_info->delayed_root;
1020 finish_one_item(delayed_root);
1021 }
1022}
1023
1024static void btrfs_release_delayed_iref(struct btrfs_delayed_node *delayed_node)
1025{
1026 struct btrfs_delayed_root *delayed_root;
1027
1028 ASSERT(delayed_node->root);
1029 clear_bit(BTRFS_DELAYED_NODE_DEL_IREF, &delayed_node->flags);
1030 delayed_node->count--;
1031
1032 delayed_root = delayed_node->root->fs_info->delayed_root;
1033 finish_one_item(delayed_root);
1034}
1035
1036static int __btrfs_update_delayed_inode(struct btrfs_trans_handle *trans,
1037 struct btrfs_root *root,
1038 struct btrfs_path *path,
1039 struct btrfs_delayed_node *node)
1040{
1041 struct btrfs_key key;
1042 struct btrfs_inode_item *inode_item;
1043 struct extent_buffer *leaf;
1044 int mod;
1045 int ret;
1046
1047 key.objectid = node->inode_id;
1048 key.type = BTRFS_INODE_ITEM_KEY;
1049 key.offset = 0;
1050
1051 if (test_bit(BTRFS_DELAYED_NODE_DEL_IREF, &node->flags))
1052 mod = -1;
1053 else
1054 mod = 1;
1055
1056 ret = btrfs_lookup_inode(trans, root, path, &key, mod);
1057 if (ret > 0) {
1058 btrfs_release_path(path);
1059 return -ENOENT;
1060 } else if (ret < 0) {
1061 return ret;
1062 }
1063
1064 leaf = path->nodes[0];
1065 inode_item = btrfs_item_ptr(leaf, path->slots[0],
1066 struct btrfs_inode_item);
1067 write_extent_buffer(leaf, &node->inode_item, (unsigned long)inode_item,
1068 sizeof(struct btrfs_inode_item));
1069 btrfs_mark_buffer_dirty(leaf);
1070
1071 if (!test_bit(BTRFS_DELAYED_NODE_DEL_IREF, &node->flags))
1072 goto no_iref;
1073
1074 path->slots[0]++;
1075 if (path->slots[0] >= btrfs_header_nritems(leaf))
1076 goto search;
1077again:
1078 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1079 if (key.objectid != node->inode_id)
1080 goto out;
1081
1082 if (key.type != BTRFS_INODE_REF_KEY &&
1083 key.type != BTRFS_INODE_EXTREF_KEY)
1084 goto out;
1085
1086 /*
1087 * Delayed iref deletion is for the inode who has only one link,
1088 * so there is only one iref. The case that several irefs are
1089 * in the same item doesn't exist.
1090 */
1091 btrfs_del_item(trans, root, path);
1092out:
1093 btrfs_release_delayed_iref(node);
1094no_iref:
1095 btrfs_release_path(path);
1096err_out:
1097 btrfs_delayed_inode_release_metadata(root, node);
1098 btrfs_release_delayed_inode(node);
1099
1100 return ret;
1101
1102search:
1103 btrfs_release_path(path);
1104
1105 key.type = BTRFS_INODE_EXTREF_KEY;
1106 key.offset = -1;
1107 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1108 if (ret < 0)
1109 goto err_out;
1110 ASSERT(ret);
1111
1112 ret = 0;
1113 leaf = path->nodes[0];
1114 path->slots[0]--;
1115 goto again;
1116}
1117
1118static inline int btrfs_update_delayed_inode(struct btrfs_trans_handle *trans,
1119 struct btrfs_root *root,
1120 struct btrfs_path *path,
1121 struct btrfs_delayed_node *node)
1122{
1123 int ret;
1124
1125 mutex_lock(&node->mutex);
1126 if (!test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &node->flags)) {
1127 mutex_unlock(&node->mutex);
1128 return 0;
1129 }
1130
1131 ret = __btrfs_update_delayed_inode(trans, root, path, node);
1132 mutex_unlock(&node->mutex);
1133 return ret;
1134}
1135
1136static inline int
1137__btrfs_commit_inode_delayed_items(struct btrfs_trans_handle *trans,
1138 struct btrfs_path *path,
1139 struct btrfs_delayed_node *node)
1140{
1141 int ret;
1142
1143 ret = btrfs_insert_delayed_items(trans, path, node->root, node);
1144 if (ret)
1145 return ret;
1146
1147 ret = btrfs_delete_delayed_items(trans, path, node->root, node);
1148 if (ret)
1149 return ret;
1150
1151 ret = btrfs_update_delayed_inode(trans, node->root, path, node);
1152 return ret;
1153}
1154
1155/*
1156 * Called when committing the transaction.
1157 * Returns 0 on success.
1158 * Returns < 0 on error and returns with an aborted transaction with any
1159 * outstanding delayed items cleaned up.
1160 */
1161static int __btrfs_run_delayed_items(struct btrfs_trans_handle *trans,
1162 struct btrfs_root *root, int nr)
1163{
1164 struct btrfs_delayed_root *delayed_root;
1165 struct btrfs_delayed_node *curr_node, *prev_node;
1166 struct btrfs_path *path;
1167 struct btrfs_block_rsv *block_rsv;
1168 int ret = 0;
1169 bool count = (nr > 0);
1170
1171 if (trans->aborted)
1172 return -EIO;
1173
1174 path = btrfs_alloc_path();
1175 if (!path)
1176 return -ENOMEM;
1177 path->leave_spinning = 1;
1178
1179 block_rsv = trans->block_rsv;
1180 trans->block_rsv = &root->fs_info->delayed_block_rsv;
1181
1182 delayed_root = btrfs_get_delayed_root(root);
1183
1184 curr_node = btrfs_first_delayed_node(delayed_root);
1185 while (curr_node && (!count || (count && nr--))) {
1186 ret = __btrfs_commit_inode_delayed_items(trans, path,
1187 curr_node);
1188 if (ret) {
1189 btrfs_release_delayed_node(curr_node);
1190 curr_node = NULL;
1191 btrfs_abort_transaction(trans, root, ret);
1192 break;
1193 }
1194
1195 prev_node = curr_node;
1196 curr_node = btrfs_next_delayed_node(curr_node);
1197 btrfs_release_delayed_node(prev_node);
1198 }
1199
1200 if (curr_node)
1201 btrfs_release_delayed_node(curr_node);
1202 btrfs_free_path(path);
1203 trans->block_rsv = block_rsv;
1204
1205 return ret;
1206}
1207
1208int btrfs_run_delayed_items(struct btrfs_trans_handle *trans,
1209 struct btrfs_root *root)
1210{
1211 return __btrfs_run_delayed_items(trans, root, -1);
1212}
1213
1214int btrfs_run_delayed_items_nr(struct btrfs_trans_handle *trans,
1215 struct btrfs_root *root, int nr)
1216{
1217 return __btrfs_run_delayed_items(trans, root, nr);
1218}
1219
1220int btrfs_commit_inode_delayed_items(struct btrfs_trans_handle *trans,
1221 struct inode *inode)
1222{
1223 struct btrfs_delayed_node *delayed_node = btrfs_get_delayed_node(inode);
1224 struct btrfs_path *path;
1225 struct btrfs_block_rsv *block_rsv;
1226 int ret;
1227
1228 if (!delayed_node)
1229 return 0;
1230
1231 mutex_lock(&delayed_node->mutex);
1232 if (!delayed_node->count) {
1233 mutex_unlock(&delayed_node->mutex);
1234 btrfs_release_delayed_node(delayed_node);
1235 return 0;
1236 }
1237 mutex_unlock(&delayed_node->mutex);
1238
1239 path = btrfs_alloc_path();
1240 if (!path) {
1241 btrfs_release_delayed_node(delayed_node);
1242 return -ENOMEM;
1243 }
1244 path->leave_spinning = 1;
1245
1246 block_rsv = trans->block_rsv;
1247 trans->block_rsv = &delayed_node->root->fs_info->delayed_block_rsv;
1248
1249 ret = __btrfs_commit_inode_delayed_items(trans, path, delayed_node);
1250
1251 btrfs_release_delayed_node(delayed_node);
1252 btrfs_free_path(path);
1253 trans->block_rsv = block_rsv;
1254
1255 return ret;
1256}
1257
1258int btrfs_commit_inode_delayed_inode(struct inode *inode)
1259{
1260 struct btrfs_trans_handle *trans;
1261 struct btrfs_delayed_node *delayed_node = btrfs_get_delayed_node(inode);
1262 struct btrfs_path *path;
1263 struct btrfs_block_rsv *block_rsv;
1264 int ret;
1265
1266 if (!delayed_node)
1267 return 0;
1268
1269 mutex_lock(&delayed_node->mutex);
1270 if (!test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
1271 mutex_unlock(&delayed_node->mutex);
1272 btrfs_release_delayed_node(delayed_node);
1273 return 0;
1274 }
1275 mutex_unlock(&delayed_node->mutex);
1276
1277 trans = btrfs_join_transaction(delayed_node->root);
1278 if (IS_ERR(trans)) {
1279 ret = PTR_ERR(trans);
1280 goto out;
1281 }
1282
1283 path = btrfs_alloc_path();
1284 if (!path) {
1285 ret = -ENOMEM;
1286 goto trans_out;
1287 }
1288 path->leave_spinning = 1;
1289
1290 block_rsv = trans->block_rsv;
1291 trans->block_rsv = &delayed_node->root->fs_info->delayed_block_rsv;
1292
1293 mutex_lock(&delayed_node->mutex);
1294 if (test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags))
1295 ret = __btrfs_update_delayed_inode(trans, delayed_node->root,
1296 path, delayed_node);
1297 else
1298 ret = 0;
1299 mutex_unlock(&delayed_node->mutex);
1300
1301 btrfs_free_path(path);
1302 trans->block_rsv = block_rsv;
1303trans_out:
1304 btrfs_end_transaction(trans, delayed_node->root);
1305 btrfs_btree_balance_dirty(delayed_node->root);
1306out:
1307 btrfs_release_delayed_node(delayed_node);
1308
1309 return ret;
1310}
1311
1312void btrfs_remove_delayed_node(struct inode *inode)
1313{
1314 struct btrfs_delayed_node *delayed_node;
1315
1316 delayed_node = ACCESS_ONCE(BTRFS_I(inode)->delayed_node);
1317 if (!delayed_node)
1318 return;
1319
1320 BTRFS_I(inode)->delayed_node = NULL;
1321 btrfs_release_delayed_node(delayed_node);
1322}
1323
1324struct btrfs_async_delayed_work {
1325 struct btrfs_delayed_root *delayed_root;
1326 int nr;
1327 struct btrfs_work work;
1328};
1329
1330static void btrfs_async_run_delayed_root(struct btrfs_work *work)
1331{
1332 struct btrfs_async_delayed_work *async_work;
1333 struct btrfs_delayed_root *delayed_root;
1334 struct btrfs_trans_handle *trans;
1335 struct btrfs_path *path;
1336 struct btrfs_delayed_node *delayed_node = NULL;
1337 struct btrfs_root *root;
1338 struct btrfs_block_rsv *block_rsv;
1339 int total_done = 0;
1340
1341 async_work = container_of(work, struct btrfs_async_delayed_work, work);
1342 delayed_root = async_work->delayed_root;
1343
1344 path = btrfs_alloc_path();
1345 if (!path)
1346 goto out;
1347
1348again:
1349 if (atomic_read(&delayed_root->items) < BTRFS_DELAYED_BACKGROUND / 2)
1350 goto free_path;
1351
1352 delayed_node = btrfs_first_prepared_delayed_node(delayed_root);
1353 if (!delayed_node)
1354 goto free_path;
1355
1356 path->leave_spinning = 1;
1357 root = delayed_node->root;
1358
1359 trans = btrfs_join_transaction(root);
1360 if (IS_ERR(trans))
1361 goto release_path;
1362
1363 block_rsv = trans->block_rsv;
1364 trans->block_rsv = &root->fs_info->delayed_block_rsv;
1365
1366 __btrfs_commit_inode_delayed_items(trans, path, delayed_node);
1367
1368 trans->block_rsv = block_rsv;
1369 btrfs_end_transaction(trans, root);
1370 btrfs_btree_balance_dirty_nodelay(root);
1371
1372release_path:
1373 btrfs_release_path(path);
1374 total_done++;
1375
1376 btrfs_release_prepared_delayed_node(delayed_node);
1377 if (async_work->nr == 0 || total_done < async_work->nr)
1378 goto again;
1379
1380free_path:
1381 btrfs_free_path(path);
1382out:
1383 wake_up(&delayed_root->wait);
1384 kfree(async_work);
1385}
1386
1387
1388static int btrfs_wq_run_delayed_node(struct btrfs_delayed_root *delayed_root,
1389 struct btrfs_fs_info *fs_info, int nr)
1390{
1391 struct btrfs_async_delayed_work *async_work;
1392
1393 if (atomic_read(&delayed_root->items) < BTRFS_DELAYED_BACKGROUND)
1394 return 0;
1395
1396 async_work = kmalloc(sizeof(*async_work), GFP_NOFS);
1397 if (!async_work)
1398 return -ENOMEM;
1399
1400 async_work->delayed_root = delayed_root;
1401 btrfs_init_work(&async_work->work, btrfs_delayed_meta_helper,
1402 btrfs_async_run_delayed_root, NULL, NULL);
1403 async_work->nr = nr;
1404
1405 btrfs_queue_work(fs_info->delayed_workers, &async_work->work);
1406 return 0;
1407}
1408
1409void btrfs_assert_delayed_root_empty(struct btrfs_root *root)
1410{
1411 struct btrfs_delayed_root *delayed_root;
1412 delayed_root = btrfs_get_delayed_root(root);
1413 WARN_ON(btrfs_first_delayed_node(delayed_root));
1414}
1415
1416static int could_end_wait(struct btrfs_delayed_root *delayed_root, int seq)
1417{
1418 int val = atomic_read(&delayed_root->items_seq);
1419
1420 if (val < seq || val >= seq + BTRFS_DELAYED_BATCH)
1421 return 1;
1422
1423 if (atomic_read(&delayed_root->items) < BTRFS_DELAYED_BACKGROUND)
1424 return 1;
1425
1426 return 0;
1427}
1428
1429void btrfs_balance_delayed_items(struct btrfs_root *root)
1430{
1431 struct btrfs_delayed_root *delayed_root;
1432 struct btrfs_fs_info *fs_info = root->fs_info;
1433
1434 delayed_root = btrfs_get_delayed_root(root);
1435
1436 if (atomic_read(&delayed_root->items) < BTRFS_DELAYED_BACKGROUND)
1437 return;
1438
1439 if (atomic_read(&delayed_root->items) >= BTRFS_DELAYED_WRITEBACK) {
1440 int seq;
1441 int ret;
1442
1443 seq = atomic_read(&delayed_root->items_seq);
1444
1445 ret = btrfs_wq_run_delayed_node(delayed_root, fs_info, 0);
1446 if (ret)
1447 return;
1448
1449 wait_event_interruptible(delayed_root->wait,
1450 could_end_wait(delayed_root, seq));
1451 return;
1452 }
1453
1454 btrfs_wq_run_delayed_node(delayed_root, fs_info, BTRFS_DELAYED_BATCH);
1455}
1456
1457/* Will return 0 or -ENOMEM */
1458int btrfs_insert_delayed_dir_index(struct btrfs_trans_handle *trans,
1459 struct btrfs_root *root, const char *name,
1460 int name_len, struct inode *dir,
1461 struct btrfs_disk_key *disk_key, u8 type,
1462 u64 index)
1463{
1464 struct btrfs_delayed_node *delayed_node;
1465 struct btrfs_delayed_item *delayed_item;
1466 struct btrfs_dir_item *dir_item;
1467 int ret;
1468
1469 delayed_node = btrfs_get_or_create_delayed_node(dir);
1470 if (IS_ERR(delayed_node))
1471 return PTR_ERR(delayed_node);
1472
1473 delayed_item = btrfs_alloc_delayed_item(sizeof(*dir_item) + name_len);
1474 if (!delayed_item) {
1475 ret = -ENOMEM;
1476 goto release_node;
1477 }
1478
1479 delayed_item->key.objectid = btrfs_ino(dir);
1480 delayed_item->key.type = BTRFS_DIR_INDEX_KEY;
1481 delayed_item->key.offset = index;
1482
1483 dir_item = (struct btrfs_dir_item *)delayed_item->data;
1484 dir_item->location = *disk_key;
1485 btrfs_set_stack_dir_transid(dir_item, trans->transid);
1486 btrfs_set_stack_dir_data_len(dir_item, 0);
1487 btrfs_set_stack_dir_name_len(dir_item, name_len);
1488 btrfs_set_stack_dir_type(dir_item, type);
1489 memcpy((char *)(dir_item + 1), name, name_len);
1490
1491 ret = btrfs_delayed_item_reserve_metadata(trans, root, delayed_item);
1492 /*
1493 * we have reserved enough space when we start a new transaction,
1494 * so reserving metadata failure is impossible
1495 */
1496 BUG_ON(ret);
1497
1498
1499 mutex_lock(&delayed_node->mutex);
1500 ret = __btrfs_add_delayed_insertion_item(delayed_node, delayed_item);
1501 if (unlikely(ret)) {
1502 btrfs_err(root->fs_info, "err add delayed dir index item(name: %.*s) "
1503 "into the insertion tree of the delayed node"
1504 "(root id: %llu, inode id: %llu, errno: %d)",
1505 name_len, name, delayed_node->root->objectid,
1506 delayed_node->inode_id, ret);
1507 BUG();
1508 }
1509 mutex_unlock(&delayed_node->mutex);
1510
1511release_node:
1512 btrfs_release_delayed_node(delayed_node);
1513 return ret;
1514}
1515
1516static int btrfs_delete_delayed_insertion_item(struct btrfs_root *root,
1517 struct btrfs_delayed_node *node,
1518 struct btrfs_key *key)
1519{
1520 struct btrfs_delayed_item *item;
1521
1522 mutex_lock(&node->mutex);
1523 item = __btrfs_lookup_delayed_insertion_item(node, key);
1524 if (!item) {
1525 mutex_unlock(&node->mutex);
1526 return 1;
1527 }
1528
1529 btrfs_delayed_item_release_metadata(root, item);
1530 btrfs_release_delayed_item(item);
1531 mutex_unlock(&node->mutex);
1532 return 0;
1533}
1534
1535int btrfs_delete_delayed_dir_index(struct btrfs_trans_handle *trans,
1536 struct btrfs_root *root, struct inode *dir,
1537 u64 index)
1538{
1539 struct btrfs_delayed_node *node;
1540 struct btrfs_delayed_item *item;
1541 struct btrfs_key item_key;
1542 int ret;
1543
1544 node = btrfs_get_or_create_delayed_node(dir);
1545 if (IS_ERR(node))
1546 return PTR_ERR(node);
1547
1548 item_key.objectid = btrfs_ino(dir);
1549 item_key.type = BTRFS_DIR_INDEX_KEY;
1550 item_key.offset = index;
1551
1552 ret = btrfs_delete_delayed_insertion_item(root, node, &item_key);
1553 if (!ret)
1554 goto end;
1555
1556 item = btrfs_alloc_delayed_item(0);
1557 if (!item) {
1558 ret = -ENOMEM;
1559 goto end;
1560 }
1561
1562 item->key = item_key;
1563
1564 ret = btrfs_delayed_item_reserve_metadata(trans, root, item);
1565 /*
1566 * we have reserved enough space when we start a new transaction,
1567 * so reserving metadata failure is impossible.
1568 */
1569 BUG_ON(ret);
1570
1571 mutex_lock(&node->mutex);
1572 ret = __btrfs_add_delayed_deletion_item(node, item);
1573 if (unlikely(ret)) {
1574 btrfs_err(root->fs_info, "err add delayed dir index item(index: %llu) "
1575 "into the deletion tree of the delayed node"
1576 "(root id: %llu, inode id: %llu, errno: %d)",
1577 index, node->root->objectid, node->inode_id,
1578 ret);
1579 BUG();
1580 }
1581 mutex_unlock(&node->mutex);
1582end:
1583 btrfs_release_delayed_node(node);
1584 return ret;
1585}
1586
1587int btrfs_inode_delayed_dir_index_count(struct inode *inode)
1588{
1589 struct btrfs_delayed_node *delayed_node = btrfs_get_delayed_node(inode);
1590
1591 if (!delayed_node)
1592 return -ENOENT;
1593
1594 /*
1595 * Since we have held i_mutex of this directory, it is impossible that
1596 * a new directory index is added into the delayed node and index_cnt
1597 * is updated now. So we needn't lock the delayed node.
1598 */
1599 if (!delayed_node->index_cnt) {
1600 btrfs_release_delayed_node(delayed_node);
1601 return -EINVAL;
1602 }
1603
1604 BTRFS_I(inode)->index_cnt = delayed_node->index_cnt;
1605 btrfs_release_delayed_node(delayed_node);
1606 return 0;
1607}
1608
1609void btrfs_get_delayed_items(struct inode *inode, struct list_head *ins_list,
1610 struct list_head *del_list)
1611{
1612 struct btrfs_delayed_node *delayed_node;
1613 struct btrfs_delayed_item *item;
1614
1615 delayed_node = btrfs_get_delayed_node(inode);
1616 if (!delayed_node)
1617 return;
1618
1619 mutex_lock(&delayed_node->mutex);
1620 item = __btrfs_first_delayed_insertion_item(delayed_node);
1621 while (item) {
1622 atomic_inc(&item->refs);
1623 list_add_tail(&item->readdir_list, ins_list);
1624 item = __btrfs_next_delayed_item(item);
1625 }
1626
1627 item = __btrfs_first_delayed_deletion_item(delayed_node);
1628 while (item) {
1629 atomic_inc(&item->refs);
1630 list_add_tail(&item->readdir_list, del_list);
1631 item = __btrfs_next_delayed_item(item);
1632 }
1633 mutex_unlock(&delayed_node->mutex);
1634 /*
1635 * This delayed node is still cached in the btrfs inode, so refs
1636 * must be > 1 now, and we needn't check it is going to be freed
1637 * or not.
1638 *
1639 * Besides that, this function is used to read dir, we do not
1640 * insert/delete delayed items in this period. So we also needn't
1641 * requeue or dequeue this delayed node.
1642 */
1643 atomic_dec(&delayed_node->refs);
1644}
1645
1646void btrfs_put_delayed_items(struct list_head *ins_list,
1647 struct list_head *del_list)
1648{
1649 struct btrfs_delayed_item *curr, *next;
1650
1651 list_for_each_entry_safe(curr, next, ins_list, readdir_list) {
1652 list_del(&curr->readdir_list);
1653 if (atomic_dec_and_test(&curr->refs))
1654 kfree(curr);
1655 }
1656
1657 list_for_each_entry_safe(curr, next, del_list, readdir_list) {
1658 list_del(&curr->readdir_list);
1659 if (atomic_dec_and_test(&curr->refs))
1660 kfree(curr);
1661 }
1662}
1663
1664int btrfs_should_delete_dir_index(struct list_head *del_list,
1665 u64 index)
1666{
1667 struct btrfs_delayed_item *curr, *next;
1668 int ret;
1669
1670 if (list_empty(del_list))
1671 return 0;
1672
1673 list_for_each_entry_safe(curr, next, del_list, readdir_list) {
1674 if (curr->key.offset > index)
1675 break;
1676
1677 list_del(&curr->readdir_list);
1678 ret = (curr->key.offset == index);
1679
1680 if (atomic_dec_and_test(&curr->refs))
1681 kfree(curr);
1682
1683 if (ret)
1684 return 1;
1685 else
1686 continue;
1687 }
1688 return 0;
1689}
1690
1691/*
1692 * btrfs_readdir_delayed_dir_index - read dir info stored in the delayed tree
1693 *
1694 */
1695int btrfs_readdir_delayed_dir_index(struct dir_context *ctx,
1696 struct list_head *ins_list, bool *emitted)
1697{
1698 struct btrfs_dir_item *di;
1699 struct btrfs_delayed_item *curr, *next;
1700 struct btrfs_key location;
1701 char *name;
1702 int name_len;
1703 int over = 0;
1704 unsigned char d_type;
1705
1706 if (list_empty(ins_list))
1707 return 0;
1708
1709 /*
1710 * Changing the data of the delayed item is impossible. So
1711 * we needn't lock them. And we have held i_mutex of the
1712 * directory, nobody can delete any directory indexes now.
1713 */
1714 list_for_each_entry_safe(curr, next, ins_list, readdir_list) {
1715 list_del(&curr->readdir_list);
1716
1717 if (curr->key.offset < ctx->pos) {
1718 if (atomic_dec_and_test(&curr->refs))
1719 kfree(curr);
1720 continue;
1721 }
1722
1723 ctx->pos = curr->key.offset;
1724
1725 di = (struct btrfs_dir_item *)curr->data;
1726 name = (char *)(di + 1);
1727 name_len = btrfs_stack_dir_name_len(di);
1728
1729 d_type = btrfs_filetype_table[di->type];
1730 btrfs_disk_key_to_cpu(&location, &di->location);
1731
1732 over = !dir_emit(ctx, name, name_len,
1733 location.objectid, d_type);
1734
1735 if (atomic_dec_and_test(&curr->refs))
1736 kfree(curr);
1737
1738 if (over)
1739 return 1;
1740 *emitted = true;
1741 }
1742 return 0;
1743}
1744
1745static void fill_stack_inode_item(struct btrfs_trans_handle *trans,
1746 struct btrfs_inode_item *inode_item,
1747 struct inode *inode)
1748{
1749 btrfs_set_stack_inode_uid(inode_item, i_uid_read(inode));
1750 btrfs_set_stack_inode_gid(inode_item, i_gid_read(inode));
1751 btrfs_set_stack_inode_size(inode_item, BTRFS_I(inode)->disk_i_size);
1752 btrfs_set_stack_inode_mode(inode_item, inode->i_mode);
1753 btrfs_set_stack_inode_nlink(inode_item, inode->i_nlink);
1754 btrfs_set_stack_inode_nbytes(inode_item, inode_get_bytes(inode));
1755 btrfs_set_stack_inode_generation(inode_item,
1756 BTRFS_I(inode)->generation);
1757 btrfs_set_stack_inode_sequence(inode_item, inode->i_version);
1758 btrfs_set_stack_inode_transid(inode_item, trans->transid);
1759 btrfs_set_stack_inode_rdev(inode_item, inode->i_rdev);
1760 btrfs_set_stack_inode_flags(inode_item, BTRFS_I(inode)->flags);
1761 btrfs_set_stack_inode_block_group(inode_item, 0);
1762
1763 btrfs_set_stack_timespec_sec(&inode_item->atime,
1764 inode->i_atime.tv_sec);
1765 btrfs_set_stack_timespec_nsec(&inode_item->atime,
1766 inode->i_atime.tv_nsec);
1767
1768 btrfs_set_stack_timespec_sec(&inode_item->mtime,
1769 inode->i_mtime.tv_sec);
1770 btrfs_set_stack_timespec_nsec(&inode_item->mtime,
1771 inode->i_mtime.tv_nsec);
1772
1773 btrfs_set_stack_timespec_sec(&inode_item->ctime,
1774 inode->i_ctime.tv_sec);
1775 btrfs_set_stack_timespec_nsec(&inode_item->ctime,
1776 inode->i_ctime.tv_nsec);
1777
1778 btrfs_set_stack_timespec_sec(&inode_item->otime,
1779 BTRFS_I(inode)->i_otime.tv_sec);
1780 btrfs_set_stack_timespec_nsec(&inode_item->otime,
1781 BTRFS_I(inode)->i_otime.tv_nsec);
1782}
1783
1784int btrfs_fill_inode(struct inode *inode, u32 *rdev)
1785{
1786 struct btrfs_delayed_node *delayed_node;
1787 struct btrfs_inode_item *inode_item;
1788
1789 delayed_node = btrfs_get_delayed_node(inode);
1790 if (!delayed_node)
1791 return -ENOENT;
1792
1793 mutex_lock(&delayed_node->mutex);
1794 if (!test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
1795 mutex_unlock(&delayed_node->mutex);
1796 btrfs_release_delayed_node(delayed_node);
1797 return -ENOENT;
1798 }
1799
1800 inode_item = &delayed_node->inode_item;
1801
1802 i_uid_write(inode, btrfs_stack_inode_uid(inode_item));
1803 i_gid_write(inode, btrfs_stack_inode_gid(inode_item));
1804 btrfs_i_size_write(inode, btrfs_stack_inode_size(inode_item));
1805 inode->i_mode = btrfs_stack_inode_mode(inode_item);
1806 set_nlink(inode, btrfs_stack_inode_nlink(inode_item));
1807 inode_set_bytes(inode, btrfs_stack_inode_nbytes(inode_item));
1808 BTRFS_I(inode)->generation = btrfs_stack_inode_generation(inode_item);
1809 BTRFS_I(inode)->last_trans = btrfs_stack_inode_transid(inode_item);
1810
1811 inode->i_version = btrfs_stack_inode_sequence(inode_item);
1812 inode->i_rdev = 0;
1813 *rdev = btrfs_stack_inode_rdev(inode_item);
1814 BTRFS_I(inode)->flags = btrfs_stack_inode_flags(inode_item);
1815
1816 inode->i_atime.tv_sec = btrfs_stack_timespec_sec(&inode_item->atime);
1817 inode->i_atime.tv_nsec = btrfs_stack_timespec_nsec(&inode_item->atime);
1818
1819 inode->i_mtime.tv_sec = btrfs_stack_timespec_sec(&inode_item->mtime);
1820 inode->i_mtime.tv_nsec = btrfs_stack_timespec_nsec(&inode_item->mtime);
1821
1822 inode->i_ctime.tv_sec = btrfs_stack_timespec_sec(&inode_item->ctime);
1823 inode->i_ctime.tv_nsec = btrfs_stack_timespec_nsec(&inode_item->ctime);
1824
1825 BTRFS_I(inode)->i_otime.tv_sec =
1826 btrfs_stack_timespec_sec(&inode_item->otime);
1827 BTRFS_I(inode)->i_otime.tv_nsec =
1828 btrfs_stack_timespec_nsec(&inode_item->otime);
1829
1830 inode->i_generation = BTRFS_I(inode)->generation;
1831 BTRFS_I(inode)->index_cnt = (u64)-1;
1832
1833 mutex_unlock(&delayed_node->mutex);
1834 btrfs_release_delayed_node(delayed_node);
1835 return 0;
1836}
1837
1838int btrfs_delayed_update_inode(struct btrfs_trans_handle *trans,
1839 struct btrfs_root *root, struct inode *inode)
1840{
1841 struct btrfs_delayed_node *delayed_node;
1842 int ret = 0;
1843
1844 delayed_node = btrfs_get_or_create_delayed_node(inode);
1845 if (IS_ERR(delayed_node))
1846 return PTR_ERR(delayed_node);
1847
1848 mutex_lock(&delayed_node->mutex);
1849 if (test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
1850 fill_stack_inode_item(trans, &delayed_node->inode_item, inode);
1851 goto release_node;
1852 }
1853
1854 ret = btrfs_delayed_inode_reserve_metadata(trans, root, inode,
1855 delayed_node);
1856 if (ret)
1857 goto release_node;
1858
1859 fill_stack_inode_item(trans, &delayed_node->inode_item, inode);
1860 set_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags);
1861 delayed_node->count++;
1862 atomic_inc(&root->fs_info->delayed_root->items);
1863release_node:
1864 mutex_unlock(&delayed_node->mutex);
1865 btrfs_release_delayed_node(delayed_node);
1866 return ret;
1867}
1868
1869int btrfs_delayed_delete_inode_ref(struct inode *inode)
1870{
1871 struct btrfs_delayed_node *delayed_node;
1872
1873 /*
1874 * we don't do delayed inode updates during log recovery because it
1875 * leads to enospc problems. This means we also can't do
1876 * delayed inode refs
1877 */
1878 if (BTRFS_I(inode)->root->fs_info->log_root_recovering)
1879 return -EAGAIN;
1880
1881 delayed_node = btrfs_get_or_create_delayed_node(inode);
1882 if (IS_ERR(delayed_node))
1883 return PTR_ERR(delayed_node);
1884
1885 /*
1886 * We don't reserve space for inode ref deletion is because:
1887 * - We ONLY do async inode ref deletion for the inode who has only
1888 * one link(i_nlink == 1), it means there is only one inode ref.
1889 * And in most case, the inode ref and the inode item are in the
1890 * same leaf, and we will deal with them at the same time.
1891 * Since we are sure we will reserve the space for the inode item,
1892 * it is unnecessary to reserve space for inode ref deletion.
1893 * - If the inode ref and the inode item are not in the same leaf,
1894 * We also needn't worry about enospc problem, because we reserve
1895 * much more space for the inode update than it needs.
1896 * - At the worst, we can steal some space from the global reservation.
1897 * It is very rare.
1898 */
1899 mutex_lock(&delayed_node->mutex);
1900 if (test_bit(BTRFS_DELAYED_NODE_DEL_IREF, &delayed_node->flags))
1901 goto release_node;
1902
1903 set_bit(BTRFS_DELAYED_NODE_DEL_IREF, &delayed_node->flags);
1904 delayed_node->count++;
1905 atomic_inc(&BTRFS_I(inode)->root->fs_info->delayed_root->items);
1906release_node:
1907 mutex_unlock(&delayed_node->mutex);
1908 btrfs_release_delayed_node(delayed_node);
1909 return 0;
1910}
1911
1912static void __btrfs_kill_delayed_node(struct btrfs_delayed_node *delayed_node)
1913{
1914 struct btrfs_root *root = delayed_node->root;
1915 struct btrfs_delayed_item *curr_item, *prev_item;
1916
1917 mutex_lock(&delayed_node->mutex);
1918 curr_item = __btrfs_first_delayed_insertion_item(delayed_node);
1919 while (curr_item) {
1920 btrfs_delayed_item_release_metadata(root, curr_item);
1921 prev_item = curr_item;
1922 curr_item = __btrfs_next_delayed_item(prev_item);
1923 btrfs_release_delayed_item(prev_item);
1924 }
1925
1926 curr_item = __btrfs_first_delayed_deletion_item(delayed_node);
1927 while (curr_item) {
1928 btrfs_delayed_item_release_metadata(root, curr_item);
1929 prev_item = curr_item;
1930 curr_item = __btrfs_next_delayed_item(prev_item);
1931 btrfs_release_delayed_item(prev_item);
1932 }
1933
1934 if (test_bit(BTRFS_DELAYED_NODE_DEL_IREF, &delayed_node->flags))
1935 btrfs_release_delayed_iref(delayed_node);
1936
1937 if (test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
1938 btrfs_delayed_inode_release_metadata(root, delayed_node);
1939 btrfs_release_delayed_inode(delayed_node);
1940 }
1941 mutex_unlock(&delayed_node->mutex);
1942}
1943
1944void btrfs_kill_delayed_inode_items(struct inode *inode)
1945{
1946 struct btrfs_delayed_node *delayed_node;
1947
1948 delayed_node = btrfs_get_delayed_node(inode);
1949 if (!delayed_node)
1950 return;
1951
1952 __btrfs_kill_delayed_node(delayed_node);
1953 btrfs_release_delayed_node(delayed_node);
1954}
1955
1956void btrfs_kill_all_delayed_nodes(struct btrfs_root *root)
1957{
1958 u64 inode_id = 0;
1959 struct btrfs_delayed_node *delayed_nodes[8];
1960 int i, n;
1961
1962 while (1) {
1963 spin_lock(&root->inode_lock);
1964 n = radix_tree_gang_lookup(&root->delayed_nodes_tree,
1965 (void **)delayed_nodes, inode_id,
1966 ARRAY_SIZE(delayed_nodes));
1967 if (!n) {
1968 spin_unlock(&root->inode_lock);
1969 break;
1970 }
1971
1972 inode_id = delayed_nodes[n - 1]->inode_id + 1;
1973
1974 for (i = 0; i < n; i++)
1975 atomic_inc(&delayed_nodes[i]->refs);
1976 spin_unlock(&root->inode_lock);
1977
1978 for (i = 0; i < n; i++) {
1979 __btrfs_kill_delayed_node(delayed_nodes[i]);
1980 btrfs_release_delayed_node(delayed_nodes[i]);
1981 }
1982 }
1983}
1984
1985void btrfs_destroy_delayed_inodes(struct btrfs_root *root)
1986{
1987 struct btrfs_delayed_root *delayed_root;
1988 struct btrfs_delayed_node *curr_node, *prev_node;
1989
1990 delayed_root = btrfs_get_delayed_root(root);
1991
1992 curr_node = btrfs_first_delayed_node(delayed_root);
1993 while (curr_node) {
1994 __btrfs_kill_delayed_node(curr_node);
1995
1996 prev_node = curr_node;
1997 curr_node = btrfs_next_delayed_node(curr_node);
1998 btrfs_release_delayed_node(prev_node);
1999 }
2000}
2001
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 spin_lock(&root->inode_lock);
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 spin_unlock(&root->inode_lock);
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 spin_unlock(&root->inode_lock);
115 return node;
116 }
117 spin_unlock(&root->inode_lock);
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 spin_lock(&root->inode_lock);
152 ptr = xa_load(&root->delayed_nodes, ino);
153 if (ptr) {
154 /* Somebody inserted it, go back and read it. */
155 spin_unlock(&root->inode_lock);
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 spin_unlock(&root->inode_lock);
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 spin_lock(&root->inode_lock);
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 xa_erase(&root->delayed_nodes, delayed_node->inode_id);
285 spin_unlock(&root->inode_lock);
286 kmem_cache_free(delayed_node_cache, delayed_node);
287 }
288}
289
290static inline void btrfs_release_delayed_node(struct btrfs_delayed_node *node)
291{
292 __btrfs_release_delayed_node(node, 0);
293}
294
295static struct btrfs_delayed_node *btrfs_first_prepared_delayed_node(
296 struct btrfs_delayed_root *delayed_root)
297{
298 struct list_head *p;
299 struct btrfs_delayed_node *node = NULL;
300
301 spin_lock(&delayed_root->lock);
302 if (list_empty(&delayed_root->prepare_list))
303 goto out;
304
305 p = delayed_root->prepare_list.next;
306 list_del_init(p);
307 node = list_entry(p, struct btrfs_delayed_node, p_list);
308 refcount_inc(&node->refs);
309out:
310 spin_unlock(&delayed_root->lock);
311
312 return node;
313}
314
315static inline void btrfs_release_prepared_delayed_node(
316 struct btrfs_delayed_node *node)
317{
318 __btrfs_release_delayed_node(node, 1);
319}
320
321static struct btrfs_delayed_item *btrfs_alloc_delayed_item(u16 data_len,
322 struct btrfs_delayed_node *node,
323 enum btrfs_delayed_item_type type)
324{
325 struct btrfs_delayed_item *item;
326
327 item = kmalloc(struct_size(item, data, data_len), GFP_NOFS);
328 if (item) {
329 item->data_len = data_len;
330 item->type = type;
331 item->bytes_reserved = 0;
332 item->delayed_node = node;
333 RB_CLEAR_NODE(&item->rb_node);
334 INIT_LIST_HEAD(&item->log_list);
335 item->logged = false;
336 refcount_set(&item->refs, 1);
337 }
338 return item;
339}
340
341/*
342 * Look up the delayed item by key.
343 *
344 * @delayed_node: pointer to the delayed node
345 * @index: the dir index value to lookup (offset of a dir index key)
346 *
347 * Note: if we don't find the right item, we will return the prev item and
348 * the next item.
349 */
350static struct btrfs_delayed_item *__btrfs_lookup_delayed_item(
351 struct rb_root *root,
352 u64 index)
353{
354 struct rb_node *node = root->rb_node;
355 struct btrfs_delayed_item *delayed_item = NULL;
356
357 while (node) {
358 delayed_item = rb_entry(node, struct btrfs_delayed_item,
359 rb_node);
360 if (delayed_item->index < index)
361 node = node->rb_right;
362 else if (delayed_item->index > index)
363 node = node->rb_left;
364 else
365 return delayed_item;
366 }
367
368 return NULL;
369}
370
371static int __btrfs_add_delayed_item(struct btrfs_delayed_node *delayed_node,
372 struct btrfs_delayed_item *ins)
373{
374 struct rb_node **p, *node;
375 struct rb_node *parent_node = NULL;
376 struct rb_root_cached *root;
377 struct btrfs_delayed_item *item;
378 bool leftmost = true;
379
380 if (ins->type == BTRFS_DELAYED_INSERTION_ITEM)
381 root = &delayed_node->ins_root;
382 else
383 root = &delayed_node->del_root;
384
385 p = &root->rb_root.rb_node;
386 node = &ins->rb_node;
387
388 while (*p) {
389 parent_node = *p;
390 item = rb_entry(parent_node, struct btrfs_delayed_item,
391 rb_node);
392
393 if (item->index < ins->index) {
394 p = &(*p)->rb_right;
395 leftmost = false;
396 } else if (item->index > ins->index) {
397 p = &(*p)->rb_left;
398 } else {
399 return -EEXIST;
400 }
401 }
402
403 rb_link_node(node, parent_node, p);
404 rb_insert_color_cached(node, root, leftmost);
405
406 if (ins->type == BTRFS_DELAYED_INSERTION_ITEM &&
407 ins->index >= delayed_node->index_cnt)
408 delayed_node->index_cnt = ins->index + 1;
409
410 delayed_node->count++;
411 atomic_inc(&delayed_node->root->fs_info->delayed_root->items);
412 return 0;
413}
414
415static void finish_one_item(struct btrfs_delayed_root *delayed_root)
416{
417 int seq = atomic_inc_return(&delayed_root->items_seq);
418
419 /* atomic_dec_return implies a barrier */
420 if ((atomic_dec_return(&delayed_root->items) <
421 BTRFS_DELAYED_BACKGROUND || seq % BTRFS_DELAYED_BATCH == 0))
422 cond_wake_up_nomb(&delayed_root->wait);
423}
424
425static void __btrfs_remove_delayed_item(struct btrfs_delayed_item *delayed_item)
426{
427 struct btrfs_delayed_node *delayed_node = delayed_item->delayed_node;
428 struct rb_root_cached *root;
429 struct btrfs_delayed_root *delayed_root;
430
431 /* Not inserted, ignore it. */
432 if (RB_EMPTY_NODE(&delayed_item->rb_node))
433 return;
434
435 /* If it's in a rbtree, then we need to have delayed node locked. */
436 lockdep_assert_held(&delayed_node->mutex);
437
438 delayed_root = delayed_node->root->fs_info->delayed_root;
439
440 if (delayed_item->type == BTRFS_DELAYED_INSERTION_ITEM)
441 root = &delayed_node->ins_root;
442 else
443 root = &delayed_node->del_root;
444
445 rb_erase_cached(&delayed_item->rb_node, root);
446 RB_CLEAR_NODE(&delayed_item->rb_node);
447 delayed_node->count--;
448
449 finish_one_item(delayed_root);
450}
451
452static void btrfs_release_delayed_item(struct btrfs_delayed_item *item)
453{
454 if (item) {
455 __btrfs_remove_delayed_item(item);
456 if (refcount_dec_and_test(&item->refs))
457 kfree(item);
458 }
459}
460
461static struct btrfs_delayed_item *__btrfs_first_delayed_insertion_item(
462 struct btrfs_delayed_node *delayed_node)
463{
464 struct rb_node *p;
465 struct btrfs_delayed_item *item = NULL;
466
467 p = rb_first_cached(&delayed_node->ins_root);
468 if (p)
469 item = rb_entry(p, struct btrfs_delayed_item, rb_node);
470
471 return item;
472}
473
474static struct btrfs_delayed_item *__btrfs_first_delayed_deletion_item(
475 struct btrfs_delayed_node *delayed_node)
476{
477 struct rb_node *p;
478 struct btrfs_delayed_item *item = NULL;
479
480 p = rb_first_cached(&delayed_node->del_root);
481 if (p)
482 item = rb_entry(p, struct btrfs_delayed_item, rb_node);
483
484 return item;
485}
486
487static struct btrfs_delayed_item *__btrfs_next_delayed_item(
488 struct btrfs_delayed_item *item)
489{
490 struct rb_node *p;
491 struct btrfs_delayed_item *next = NULL;
492
493 p = rb_next(&item->rb_node);
494 if (p)
495 next = rb_entry(p, struct btrfs_delayed_item, rb_node);
496
497 return next;
498}
499
500static int btrfs_delayed_item_reserve_metadata(struct btrfs_trans_handle *trans,
501 struct btrfs_delayed_item *item)
502{
503 struct btrfs_block_rsv *src_rsv;
504 struct btrfs_block_rsv *dst_rsv;
505 struct btrfs_fs_info *fs_info = trans->fs_info;
506 u64 num_bytes;
507 int ret;
508
509 if (!trans->bytes_reserved)
510 return 0;
511
512 src_rsv = trans->block_rsv;
513 dst_rsv = &fs_info->delayed_block_rsv;
514
515 num_bytes = btrfs_calc_insert_metadata_size(fs_info, 1);
516
517 /*
518 * Here we migrate space rsv from transaction rsv, since have already
519 * reserved space when starting a transaction. So no need to reserve
520 * qgroup space here.
521 */
522 ret = btrfs_block_rsv_migrate(src_rsv, dst_rsv, num_bytes, true);
523 if (!ret) {
524 trace_btrfs_space_reservation(fs_info, "delayed_item",
525 item->delayed_node->inode_id,
526 num_bytes, 1);
527 /*
528 * For insertions we track reserved metadata space by accounting
529 * for the number of leaves that will be used, based on the delayed
530 * node's curr_index_batch_size and index_item_leaves fields.
531 */
532 if (item->type == BTRFS_DELAYED_DELETION_ITEM)
533 item->bytes_reserved = num_bytes;
534 }
535
536 return ret;
537}
538
539static void btrfs_delayed_item_release_metadata(struct btrfs_root *root,
540 struct btrfs_delayed_item *item)
541{
542 struct btrfs_block_rsv *rsv;
543 struct btrfs_fs_info *fs_info = root->fs_info;
544
545 if (!item->bytes_reserved)
546 return;
547
548 rsv = &fs_info->delayed_block_rsv;
549 /*
550 * Check btrfs_delayed_item_reserve_metadata() to see why we don't need
551 * to release/reserve qgroup space.
552 */
553 trace_btrfs_space_reservation(fs_info, "delayed_item",
554 item->delayed_node->inode_id,
555 item->bytes_reserved, 0);
556 btrfs_block_rsv_release(fs_info, rsv, item->bytes_reserved, NULL);
557}
558
559static void btrfs_delayed_item_release_leaves(struct btrfs_delayed_node *node,
560 unsigned int num_leaves)
561{
562 struct btrfs_fs_info *fs_info = node->root->fs_info;
563 const u64 bytes = btrfs_calc_insert_metadata_size(fs_info, num_leaves);
564
565 /* There are no space reservations during log replay, bail out. */
566 if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
567 return;
568
569 trace_btrfs_space_reservation(fs_info, "delayed_item", node->inode_id,
570 bytes, 0);
571 btrfs_block_rsv_release(fs_info, &fs_info->delayed_block_rsv, bytes, NULL);
572}
573
574static int btrfs_delayed_inode_reserve_metadata(
575 struct btrfs_trans_handle *trans,
576 struct btrfs_root *root,
577 struct btrfs_delayed_node *node)
578{
579 struct btrfs_fs_info *fs_info = root->fs_info;
580 struct btrfs_block_rsv *src_rsv;
581 struct btrfs_block_rsv *dst_rsv;
582 u64 num_bytes;
583 int ret;
584
585 src_rsv = trans->block_rsv;
586 dst_rsv = &fs_info->delayed_block_rsv;
587
588 num_bytes = btrfs_calc_metadata_size(fs_info, 1);
589
590 /*
591 * btrfs_dirty_inode will update the inode under btrfs_join_transaction
592 * which doesn't reserve space for speed. This is a problem since we
593 * still need to reserve space for this update, so try to reserve the
594 * space.
595 *
596 * Now if src_rsv == delalloc_block_rsv we'll let it just steal since
597 * we always reserve enough to update the inode item.
598 */
599 if (!src_rsv || (!trans->bytes_reserved &&
600 src_rsv->type != BTRFS_BLOCK_RSV_DELALLOC)) {
601 ret = btrfs_qgroup_reserve_meta(root, num_bytes,
602 BTRFS_QGROUP_RSV_META_PREALLOC, true);
603 if (ret < 0)
604 return ret;
605 ret = btrfs_block_rsv_add(fs_info, dst_rsv, num_bytes,
606 BTRFS_RESERVE_NO_FLUSH);
607 /* NO_FLUSH could only fail with -ENOSPC */
608 ASSERT(ret == 0 || ret == -ENOSPC);
609 if (ret)
610 btrfs_qgroup_free_meta_prealloc(root, num_bytes);
611 } else {
612 ret = btrfs_block_rsv_migrate(src_rsv, dst_rsv, num_bytes, true);
613 }
614
615 if (!ret) {
616 trace_btrfs_space_reservation(fs_info, "delayed_inode",
617 node->inode_id, num_bytes, 1);
618 node->bytes_reserved = num_bytes;
619 }
620
621 return ret;
622}
623
624static void btrfs_delayed_inode_release_metadata(struct btrfs_fs_info *fs_info,
625 struct btrfs_delayed_node *node,
626 bool qgroup_free)
627{
628 struct btrfs_block_rsv *rsv;
629
630 if (!node->bytes_reserved)
631 return;
632
633 rsv = &fs_info->delayed_block_rsv;
634 trace_btrfs_space_reservation(fs_info, "delayed_inode",
635 node->inode_id, node->bytes_reserved, 0);
636 btrfs_block_rsv_release(fs_info, rsv, node->bytes_reserved, NULL);
637 if (qgroup_free)
638 btrfs_qgroup_free_meta_prealloc(node->root,
639 node->bytes_reserved);
640 else
641 btrfs_qgroup_convert_reserved_meta(node->root,
642 node->bytes_reserved);
643 node->bytes_reserved = 0;
644}
645
646/*
647 * Insert a single delayed item or a batch of delayed items, as many as possible
648 * that fit in a leaf. The delayed items (dir index keys) are sorted by their key
649 * in the rbtree, and if there's a gap between two consecutive dir index items,
650 * then it means at some point we had delayed dir indexes to add but they got
651 * removed (by btrfs_delete_delayed_dir_index()) before we attempted to flush them
652 * into the subvolume tree. Dir index keys also have their offsets coming from a
653 * monotonically increasing counter, so we can't get new keys with an offset that
654 * fits within a gap between delayed dir index items.
655 */
656static int btrfs_insert_delayed_item(struct btrfs_trans_handle *trans,
657 struct btrfs_root *root,
658 struct btrfs_path *path,
659 struct btrfs_delayed_item *first_item)
660{
661 struct btrfs_fs_info *fs_info = root->fs_info;
662 struct btrfs_delayed_node *node = first_item->delayed_node;
663 LIST_HEAD(item_list);
664 struct btrfs_delayed_item *curr;
665 struct btrfs_delayed_item *next;
666 const int max_size = BTRFS_LEAF_DATA_SIZE(fs_info);
667 struct btrfs_item_batch batch;
668 struct btrfs_key first_key;
669 const u32 first_data_size = first_item->data_len;
670 int total_size;
671 char *ins_data = NULL;
672 int ret;
673 bool continuous_keys_only = false;
674
675 lockdep_assert_held(&node->mutex);
676
677 /*
678 * During normal operation the delayed index offset is continuously
679 * increasing, so we can batch insert all items as there will not be any
680 * overlapping keys in the tree.
681 *
682 * The exception to this is log replay, where we may have interleaved
683 * offsets in the tree, so our batch needs to be continuous keys only in
684 * order to ensure we do not end up with out of order items in our leaf.
685 */
686 if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
687 continuous_keys_only = true;
688
689 /*
690 * For delayed items to insert, we track reserved metadata bytes based
691 * on the number of leaves that we will use.
692 * See btrfs_insert_delayed_dir_index() and
693 * btrfs_delayed_item_reserve_metadata()).
694 */
695 ASSERT(first_item->bytes_reserved == 0);
696
697 list_add_tail(&first_item->tree_list, &item_list);
698 batch.total_data_size = first_data_size;
699 batch.nr = 1;
700 total_size = first_data_size + sizeof(struct btrfs_item);
701 curr = first_item;
702
703 while (true) {
704 int next_size;
705
706 next = __btrfs_next_delayed_item(curr);
707 if (!next)
708 break;
709
710 /*
711 * We cannot allow gaps in the key space if we're doing log
712 * replay.
713 */
714 if (continuous_keys_only && (next->index != curr->index + 1))
715 break;
716
717 ASSERT(next->bytes_reserved == 0);
718
719 next_size = next->data_len + sizeof(struct btrfs_item);
720 if (total_size + next_size > max_size)
721 break;
722
723 list_add_tail(&next->tree_list, &item_list);
724 batch.nr++;
725 total_size += next_size;
726 batch.total_data_size += next->data_len;
727 curr = next;
728 }
729
730 if (batch.nr == 1) {
731 first_key.objectid = node->inode_id;
732 first_key.type = BTRFS_DIR_INDEX_KEY;
733 first_key.offset = first_item->index;
734 batch.keys = &first_key;
735 batch.data_sizes = &first_data_size;
736 } else {
737 struct btrfs_key *ins_keys;
738 u32 *ins_sizes;
739 int i = 0;
740
741 ins_data = kmalloc(batch.nr * sizeof(u32) +
742 batch.nr * sizeof(struct btrfs_key), GFP_NOFS);
743 if (!ins_data) {
744 ret = -ENOMEM;
745 goto out;
746 }
747 ins_sizes = (u32 *)ins_data;
748 ins_keys = (struct btrfs_key *)(ins_data + batch.nr * sizeof(u32));
749 batch.keys = ins_keys;
750 batch.data_sizes = ins_sizes;
751 list_for_each_entry(curr, &item_list, tree_list) {
752 ins_keys[i].objectid = node->inode_id;
753 ins_keys[i].type = BTRFS_DIR_INDEX_KEY;
754 ins_keys[i].offset = curr->index;
755 ins_sizes[i] = curr->data_len;
756 i++;
757 }
758 }
759
760 ret = btrfs_insert_empty_items(trans, root, path, &batch);
761 if (ret)
762 goto out;
763
764 list_for_each_entry(curr, &item_list, tree_list) {
765 char *data_ptr;
766
767 data_ptr = btrfs_item_ptr(path->nodes[0], path->slots[0], char);
768 write_extent_buffer(path->nodes[0], &curr->data,
769 (unsigned long)data_ptr, curr->data_len);
770 path->slots[0]++;
771 }
772
773 /*
774 * Now release our path before releasing the delayed items and their
775 * metadata reservations, so that we don't block other tasks for more
776 * time than needed.
777 */
778 btrfs_release_path(path);
779
780 ASSERT(node->index_item_leaves > 0);
781
782 /*
783 * For normal operations we will batch an entire leaf's worth of delayed
784 * items, so if there are more items to process we can decrement
785 * index_item_leaves by 1 as we inserted 1 leaf's worth of items.
786 *
787 * However for log replay we may not have inserted an entire leaf's
788 * worth of items, we may have not had continuous items, so decrementing
789 * here would mess up the index_item_leaves accounting. For this case
790 * only clean up the accounting when there are no items left.
791 */
792 if (next && !continuous_keys_only) {
793 /*
794 * We inserted one batch of items into a leaf a there are more
795 * items to flush in a future batch, now release one unit of
796 * metadata space from the delayed block reserve, corresponding
797 * the leaf we just flushed to.
798 */
799 btrfs_delayed_item_release_leaves(node, 1);
800 node->index_item_leaves--;
801 } else if (!next) {
802 /*
803 * There are no more items to insert. We can have a number of
804 * reserved leaves > 1 here - this happens when many dir index
805 * items are added and then removed before they are flushed (file
806 * names with a very short life, never span a transaction). So
807 * release all remaining leaves.
808 */
809 btrfs_delayed_item_release_leaves(node, node->index_item_leaves);
810 node->index_item_leaves = 0;
811 }
812
813 list_for_each_entry_safe(curr, next, &item_list, tree_list) {
814 list_del(&curr->tree_list);
815 btrfs_release_delayed_item(curr);
816 }
817out:
818 kfree(ins_data);
819 return ret;
820}
821
822static int btrfs_insert_delayed_items(struct btrfs_trans_handle *trans,
823 struct btrfs_path *path,
824 struct btrfs_root *root,
825 struct btrfs_delayed_node *node)
826{
827 int ret = 0;
828
829 while (ret == 0) {
830 struct btrfs_delayed_item *curr;
831
832 mutex_lock(&node->mutex);
833 curr = __btrfs_first_delayed_insertion_item(node);
834 if (!curr) {
835 mutex_unlock(&node->mutex);
836 break;
837 }
838 ret = btrfs_insert_delayed_item(trans, root, path, curr);
839 mutex_unlock(&node->mutex);
840 }
841
842 return ret;
843}
844
845static int btrfs_batch_delete_items(struct btrfs_trans_handle *trans,
846 struct btrfs_root *root,
847 struct btrfs_path *path,
848 struct btrfs_delayed_item *item)
849{
850 const u64 ino = item->delayed_node->inode_id;
851 struct btrfs_fs_info *fs_info = root->fs_info;
852 struct btrfs_delayed_item *curr, *next;
853 struct extent_buffer *leaf = path->nodes[0];
854 LIST_HEAD(batch_list);
855 int nitems, slot, last_slot;
856 int ret;
857 u64 total_reserved_size = item->bytes_reserved;
858
859 ASSERT(leaf != NULL);
860
861 slot = path->slots[0];
862 last_slot = btrfs_header_nritems(leaf) - 1;
863 /*
864 * Our caller always gives us a path pointing to an existing item, so
865 * this can not happen.
866 */
867 ASSERT(slot <= last_slot);
868 if (WARN_ON(slot > last_slot))
869 return -ENOENT;
870
871 nitems = 1;
872 curr = item;
873 list_add_tail(&curr->tree_list, &batch_list);
874
875 /*
876 * Keep checking if the next delayed item matches the next item in the
877 * leaf - if so, we can add it to the batch of items to delete from the
878 * leaf.
879 */
880 while (slot < last_slot) {
881 struct btrfs_key key;
882
883 next = __btrfs_next_delayed_item(curr);
884 if (!next)
885 break;
886
887 slot++;
888 btrfs_item_key_to_cpu(leaf, &key, slot);
889 if (key.objectid != ino ||
890 key.type != BTRFS_DIR_INDEX_KEY ||
891 key.offset != next->index)
892 break;
893 nitems++;
894 curr = next;
895 list_add_tail(&curr->tree_list, &batch_list);
896 total_reserved_size += curr->bytes_reserved;
897 }
898
899 ret = btrfs_del_items(trans, root, path, path->slots[0], nitems);
900 if (ret)
901 return ret;
902
903 /* In case of BTRFS_FS_LOG_RECOVERING items won't have reserved space */
904 if (total_reserved_size > 0) {
905 /*
906 * Check btrfs_delayed_item_reserve_metadata() to see why we
907 * don't need to release/reserve qgroup space.
908 */
909 trace_btrfs_space_reservation(fs_info, "delayed_item", ino,
910 total_reserved_size, 0);
911 btrfs_block_rsv_release(fs_info, &fs_info->delayed_block_rsv,
912 total_reserved_size, NULL);
913 }
914
915 list_for_each_entry_safe(curr, next, &batch_list, tree_list) {
916 list_del(&curr->tree_list);
917 btrfs_release_delayed_item(curr);
918 }
919
920 return 0;
921}
922
923static int btrfs_delete_delayed_items(struct btrfs_trans_handle *trans,
924 struct btrfs_path *path,
925 struct btrfs_root *root,
926 struct btrfs_delayed_node *node)
927{
928 struct btrfs_key key;
929 int ret = 0;
930
931 key.objectid = node->inode_id;
932 key.type = BTRFS_DIR_INDEX_KEY;
933
934 while (ret == 0) {
935 struct btrfs_delayed_item *item;
936
937 mutex_lock(&node->mutex);
938 item = __btrfs_first_delayed_deletion_item(node);
939 if (!item) {
940 mutex_unlock(&node->mutex);
941 break;
942 }
943
944 key.offset = item->index;
945 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
946 if (ret > 0) {
947 /*
948 * There's no matching item in the leaf. This means we
949 * have already deleted this item in a past run of the
950 * delayed items. We ignore errors when running delayed
951 * items from an async context, through a work queue job
952 * running btrfs_async_run_delayed_root(), and don't
953 * release delayed items that failed to complete. This
954 * is because we will retry later, and at transaction
955 * commit time we always run delayed items and will
956 * then deal with errors if they fail to run again.
957 *
958 * So just release delayed items for which we can't find
959 * an item in the tree, and move to the next item.
960 */
961 btrfs_release_path(path);
962 btrfs_release_delayed_item(item);
963 ret = 0;
964 } else if (ret == 0) {
965 ret = btrfs_batch_delete_items(trans, root, path, item);
966 btrfs_release_path(path);
967 }
968
969 /*
970 * We unlock and relock on each iteration, this is to prevent
971 * blocking other tasks for too long while we are being run from
972 * the async context (work queue job). Those tasks are typically
973 * running system calls like creat/mkdir/rename/unlink/etc which
974 * need to add delayed items to this delayed node.
975 */
976 mutex_unlock(&node->mutex);
977 }
978
979 return ret;
980}
981
982static void btrfs_release_delayed_inode(struct btrfs_delayed_node *delayed_node)
983{
984 struct btrfs_delayed_root *delayed_root;
985
986 if (delayed_node &&
987 test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
988 ASSERT(delayed_node->root);
989 clear_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags);
990 delayed_node->count--;
991
992 delayed_root = delayed_node->root->fs_info->delayed_root;
993 finish_one_item(delayed_root);
994 }
995}
996
997static void btrfs_release_delayed_iref(struct btrfs_delayed_node *delayed_node)
998{
999
1000 if (test_and_clear_bit(BTRFS_DELAYED_NODE_DEL_IREF, &delayed_node->flags)) {
1001 struct btrfs_delayed_root *delayed_root;
1002
1003 ASSERT(delayed_node->root);
1004 delayed_node->count--;
1005
1006 delayed_root = delayed_node->root->fs_info->delayed_root;
1007 finish_one_item(delayed_root);
1008 }
1009}
1010
1011static int __btrfs_update_delayed_inode(struct btrfs_trans_handle *trans,
1012 struct btrfs_root *root,
1013 struct btrfs_path *path,
1014 struct btrfs_delayed_node *node)
1015{
1016 struct btrfs_fs_info *fs_info = root->fs_info;
1017 struct btrfs_key key;
1018 struct btrfs_inode_item *inode_item;
1019 struct extent_buffer *leaf;
1020 int mod;
1021 int ret;
1022
1023 key.objectid = node->inode_id;
1024 key.type = BTRFS_INODE_ITEM_KEY;
1025 key.offset = 0;
1026
1027 if (test_bit(BTRFS_DELAYED_NODE_DEL_IREF, &node->flags))
1028 mod = -1;
1029 else
1030 mod = 1;
1031
1032 ret = btrfs_lookup_inode(trans, root, path, &key, mod);
1033 if (ret > 0)
1034 ret = -ENOENT;
1035 if (ret < 0)
1036 goto out;
1037
1038 leaf = path->nodes[0];
1039 inode_item = btrfs_item_ptr(leaf, path->slots[0],
1040 struct btrfs_inode_item);
1041 write_extent_buffer(leaf, &node->inode_item, (unsigned long)inode_item,
1042 sizeof(struct btrfs_inode_item));
1043 btrfs_mark_buffer_dirty(trans, leaf);
1044
1045 if (!test_bit(BTRFS_DELAYED_NODE_DEL_IREF, &node->flags))
1046 goto out;
1047
1048 /*
1049 * Now we're going to delete the INODE_REF/EXTREF, which should be the
1050 * only one ref left. Check if the next item is an INODE_REF/EXTREF.
1051 *
1052 * But if we're the last item already, release and search for the last
1053 * INODE_REF/EXTREF.
1054 */
1055 if (path->slots[0] + 1 >= btrfs_header_nritems(leaf)) {
1056 key.objectid = node->inode_id;
1057 key.type = BTRFS_INODE_EXTREF_KEY;
1058 key.offset = (u64)-1;
1059
1060 btrfs_release_path(path);
1061 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1062 if (ret < 0)
1063 goto err_out;
1064 ASSERT(ret > 0);
1065 ASSERT(path->slots[0] > 0);
1066 ret = 0;
1067 path->slots[0]--;
1068 leaf = path->nodes[0];
1069 } else {
1070 path->slots[0]++;
1071 }
1072 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1073 if (key.objectid != node->inode_id)
1074 goto out;
1075 if (key.type != BTRFS_INODE_REF_KEY &&
1076 key.type != BTRFS_INODE_EXTREF_KEY)
1077 goto out;
1078
1079 /*
1080 * Delayed iref deletion is for the inode who has only one link,
1081 * so there is only one iref. The case that several irefs are
1082 * in the same item doesn't exist.
1083 */
1084 ret = btrfs_del_item(trans, root, path);
1085out:
1086 btrfs_release_delayed_iref(node);
1087 btrfs_release_path(path);
1088err_out:
1089 btrfs_delayed_inode_release_metadata(fs_info, node, (ret < 0));
1090 btrfs_release_delayed_inode(node);
1091
1092 /*
1093 * If we fail to update the delayed inode we need to abort the
1094 * transaction, because we could leave the inode with the improper
1095 * counts behind.
1096 */
1097 if (ret && ret != -ENOENT)
1098 btrfs_abort_transaction(trans, ret);
1099
1100 return ret;
1101}
1102
1103static inline int btrfs_update_delayed_inode(struct btrfs_trans_handle *trans,
1104 struct btrfs_root *root,
1105 struct btrfs_path *path,
1106 struct btrfs_delayed_node *node)
1107{
1108 int ret;
1109
1110 mutex_lock(&node->mutex);
1111 if (!test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &node->flags)) {
1112 mutex_unlock(&node->mutex);
1113 return 0;
1114 }
1115
1116 ret = __btrfs_update_delayed_inode(trans, root, path, node);
1117 mutex_unlock(&node->mutex);
1118 return ret;
1119}
1120
1121static inline int
1122__btrfs_commit_inode_delayed_items(struct btrfs_trans_handle *trans,
1123 struct btrfs_path *path,
1124 struct btrfs_delayed_node *node)
1125{
1126 int ret;
1127
1128 ret = btrfs_insert_delayed_items(trans, path, node->root, node);
1129 if (ret)
1130 return ret;
1131
1132 ret = btrfs_delete_delayed_items(trans, path, node->root, node);
1133 if (ret)
1134 return ret;
1135
1136 ret = btrfs_record_root_in_trans(trans, node->root);
1137 if (ret)
1138 return ret;
1139 ret = btrfs_update_delayed_inode(trans, node->root, path, node);
1140 return ret;
1141}
1142
1143/*
1144 * Called when committing the transaction.
1145 * Returns 0 on success.
1146 * Returns < 0 on error and returns with an aborted transaction with any
1147 * outstanding delayed items cleaned up.
1148 */
1149static int __btrfs_run_delayed_items(struct btrfs_trans_handle *trans, int nr)
1150{
1151 struct btrfs_fs_info *fs_info = trans->fs_info;
1152 struct btrfs_delayed_root *delayed_root;
1153 struct btrfs_delayed_node *curr_node, *prev_node;
1154 struct btrfs_path *path;
1155 struct btrfs_block_rsv *block_rsv;
1156 int ret = 0;
1157 bool count = (nr > 0);
1158
1159 if (TRANS_ABORTED(trans))
1160 return -EIO;
1161
1162 path = btrfs_alloc_path();
1163 if (!path)
1164 return -ENOMEM;
1165
1166 block_rsv = trans->block_rsv;
1167 trans->block_rsv = &fs_info->delayed_block_rsv;
1168
1169 delayed_root = fs_info->delayed_root;
1170
1171 curr_node = btrfs_first_delayed_node(delayed_root);
1172 while (curr_node && (!count || nr--)) {
1173 ret = __btrfs_commit_inode_delayed_items(trans, path,
1174 curr_node);
1175 if (ret) {
1176 btrfs_abort_transaction(trans, ret);
1177 break;
1178 }
1179
1180 prev_node = curr_node;
1181 curr_node = btrfs_next_delayed_node(curr_node);
1182 /*
1183 * See the comment below about releasing path before releasing
1184 * node. If the commit of delayed items was successful the path
1185 * should always be released, but in case of an error, it may
1186 * point to locked extent buffers (a leaf at the very least).
1187 */
1188 ASSERT(path->nodes[0] == NULL);
1189 btrfs_release_delayed_node(prev_node);
1190 }
1191
1192 /*
1193 * Release the path to avoid a potential deadlock and lockdep splat when
1194 * releasing the delayed node, as that requires taking the delayed node's
1195 * mutex. If another task starts running delayed items before we take
1196 * the mutex, it will first lock the mutex and then it may try to lock
1197 * the same btree path (leaf).
1198 */
1199 btrfs_free_path(path);
1200
1201 if (curr_node)
1202 btrfs_release_delayed_node(curr_node);
1203 trans->block_rsv = block_rsv;
1204
1205 return ret;
1206}
1207
1208int btrfs_run_delayed_items(struct btrfs_trans_handle *trans)
1209{
1210 return __btrfs_run_delayed_items(trans, -1);
1211}
1212
1213int btrfs_run_delayed_items_nr(struct btrfs_trans_handle *trans, int nr)
1214{
1215 return __btrfs_run_delayed_items(trans, nr);
1216}
1217
1218int btrfs_commit_inode_delayed_items(struct btrfs_trans_handle *trans,
1219 struct btrfs_inode *inode)
1220{
1221 struct btrfs_delayed_node *delayed_node = btrfs_get_delayed_node(inode);
1222 struct btrfs_path *path;
1223 struct btrfs_block_rsv *block_rsv;
1224 int ret;
1225
1226 if (!delayed_node)
1227 return 0;
1228
1229 mutex_lock(&delayed_node->mutex);
1230 if (!delayed_node->count) {
1231 mutex_unlock(&delayed_node->mutex);
1232 btrfs_release_delayed_node(delayed_node);
1233 return 0;
1234 }
1235 mutex_unlock(&delayed_node->mutex);
1236
1237 path = btrfs_alloc_path();
1238 if (!path) {
1239 btrfs_release_delayed_node(delayed_node);
1240 return -ENOMEM;
1241 }
1242
1243 block_rsv = trans->block_rsv;
1244 trans->block_rsv = &delayed_node->root->fs_info->delayed_block_rsv;
1245
1246 ret = __btrfs_commit_inode_delayed_items(trans, path, delayed_node);
1247
1248 btrfs_release_delayed_node(delayed_node);
1249 btrfs_free_path(path);
1250 trans->block_rsv = block_rsv;
1251
1252 return ret;
1253}
1254
1255int btrfs_commit_inode_delayed_inode(struct btrfs_inode *inode)
1256{
1257 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1258 struct btrfs_trans_handle *trans;
1259 struct btrfs_delayed_node *delayed_node = btrfs_get_delayed_node(inode);
1260 struct btrfs_path *path;
1261 struct btrfs_block_rsv *block_rsv;
1262 int ret;
1263
1264 if (!delayed_node)
1265 return 0;
1266
1267 mutex_lock(&delayed_node->mutex);
1268 if (!test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
1269 mutex_unlock(&delayed_node->mutex);
1270 btrfs_release_delayed_node(delayed_node);
1271 return 0;
1272 }
1273 mutex_unlock(&delayed_node->mutex);
1274
1275 trans = btrfs_join_transaction(delayed_node->root);
1276 if (IS_ERR(trans)) {
1277 ret = PTR_ERR(trans);
1278 goto out;
1279 }
1280
1281 path = btrfs_alloc_path();
1282 if (!path) {
1283 ret = -ENOMEM;
1284 goto trans_out;
1285 }
1286
1287 block_rsv = trans->block_rsv;
1288 trans->block_rsv = &fs_info->delayed_block_rsv;
1289
1290 mutex_lock(&delayed_node->mutex);
1291 if (test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags))
1292 ret = __btrfs_update_delayed_inode(trans, delayed_node->root,
1293 path, delayed_node);
1294 else
1295 ret = 0;
1296 mutex_unlock(&delayed_node->mutex);
1297
1298 btrfs_free_path(path);
1299 trans->block_rsv = block_rsv;
1300trans_out:
1301 btrfs_end_transaction(trans);
1302 btrfs_btree_balance_dirty(fs_info);
1303out:
1304 btrfs_release_delayed_node(delayed_node);
1305
1306 return ret;
1307}
1308
1309void btrfs_remove_delayed_node(struct btrfs_inode *inode)
1310{
1311 struct btrfs_delayed_node *delayed_node;
1312
1313 delayed_node = READ_ONCE(inode->delayed_node);
1314 if (!delayed_node)
1315 return;
1316
1317 inode->delayed_node = NULL;
1318 btrfs_release_delayed_node(delayed_node);
1319}
1320
1321struct btrfs_async_delayed_work {
1322 struct btrfs_delayed_root *delayed_root;
1323 int nr;
1324 struct btrfs_work work;
1325};
1326
1327static void btrfs_async_run_delayed_root(struct btrfs_work *work)
1328{
1329 struct btrfs_async_delayed_work *async_work;
1330 struct btrfs_delayed_root *delayed_root;
1331 struct btrfs_trans_handle *trans;
1332 struct btrfs_path *path;
1333 struct btrfs_delayed_node *delayed_node = NULL;
1334 struct btrfs_root *root;
1335 struct btrfs_block_rsv *block_rsv;
1336 int total_done = 0;
1337
1338 async_work = container_of(work, struct btrfs_async_delayed_work, work);
1339 delayed_root = async_work->delayed_root;
1340
1341 path = btrfs_alloc_path();
1342 if (!path)
1343 goto out;
1344
1345 do {
1346 if (atomic_read(&delayed_root->items) <
1347 BTRFS_DELAYED_BACKGROUND / 2)
1348 break;
1349
1350 delayed_node = btrfs_first_prepared_delayed_node(delayed_root);
1351 if (!delayed_node)
1352 break;
1353
1354 root = delayed_node->root;
1355
1356 trans = btrfs_join_transaction(root);
1357 if (IS_ERR(trans)) {
1358 btrfs_release_path(path);
1359 btrfs_release_prepared_delayed_node(delayed_node);
1360 total_done++;
1361 continue;
1362 }
1363
1364 block_rsv = trans->block_rsv;
1365 trans->block_rsv = &root->fs_info->delayed_block_rsv;
1366
1367 __btrfs_commit_inode_delayed_items(trans, path, delayed_node);
1368
1369 trans->block_rsv = block_rsv;
1370 btrfs_end_transaction(trans);
1371 btrfs_btree_balance_dirty_nodelay(root->fs_info);
1372
1373 btrfs_release_path(path);
1374 btrfs_release_prepared_delayed_node(delayed_node);
1375 total_done++;
1376
1377 } while ((async_work->nr == 0 && total_done < BTRFS_DELAYED_WRITEBACK)
1378 || total_done < async_work->nr);
1379
1380 btrfs_free_path(path);
1381out:
1382 wake_up(&delayed_root->wait);
1383 kfree(async_work);
1384}
1385
1386
1387static int btrfs_wq_run_delayed_node(struct btrfs_delayed_root *delayed_root,
1388 struct btrfs_fs_info *fs_info, int nr)
1389{
1390 struct btrfs_async_delayed_work *async_work;
1391
1392 async_work = kmalloc(sizeof(*async_work), GFP_NOFS);
1393 if (!async_work)
1394 return -ENOMEM;
1395
1396 async_work->delayed_root = delayed_root;
1397 btrfs_init_work(&async_work->work, btrfs_async_run_delayed_root, NULL);
1398 async_work->nr = nr;
1399
1400 btrfs_queue_work(fs_info->delayed_workers, &async_work->work);
1401 return 0;
1402}
1403
1404void btrfs_assert_delayed_root_empty(struct btrfs_fs_info *fs_info)
1405{
1406 WARN_ON(btrfs_first_delayed_node(fs_info->delayed_root));
1407}
1408
1409static int could_end_wait(struct btrfs_delayed_root *delayed_root, int seq)
1410{
1411 int val = atomic_read(&delayed_root->items_seq);
1412
1413 if (val < seq || val >= seq + BTRFS_DELAYED_BATCH)
1414 return 1;
1415
1416 if (atomic_read(&delayed_root->items) < BTRFS_DELAYED_BACKGROUND)
1417 return 1;
1418
1419 return 0;
1420}
1421
1422void btrfs_balance_delayed_items(struct btrfs_fs_info *fs_info)
1423{
1424 struct btrfs_delayed_root *delayed_root = fs_info->delayed_root;
1425
1426 if ((atomic_read(&delayed_root->items) < BTRFS_DELAYED_BACKGROUND) ||
1427 btrfs_workqueue_normal_congested(fs_info->delayed_workers))
1428 return;
1429
1430 if (atomic_read(&delayed_root->items) >= BTRFS_DELAYED_WRITEBACK) {
1431 int seq;
1432 int ret;
1433
1434 seq = atomic_read(&delayed_root->items_seq);
1435
1436 ret = btrfs_wq_run_delayed_node(delayed_root, fs_info, 0);
1437 if (ret)
1438 return;
1439
1440 wait_event_interruptible(delayed_root->wait,
1441 could_end_wait(delayed_root, seq));
1442 return;
1443 }
1444
1445 btrfs_wq_run_delayed_node(delayed_root, fs_info, BTRFS_DELAYED_BATCH);
1446}
1447
1448static void btrfs_release_dir_index_item_space(struct btrfs_trans_handle *trans)
1449{
1450 struct btrfs_fs_info *fs_info = trans->fs_info;
1451 const u64 bytes = btrfs_calc_insert_metadata_size(fs_info, 1);
1452
1453 if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
1454 return;
1455
1456 /*
1457 * Adding the new dir index item does not require touching another
1458 * leaf, so we can release 1 unit of metadata that was previously
1459 * reserved when starting the transaction. This applies only to
1460 * the case where we had a transaction start and excludes the
1461 * transaction join case (when replaying log trees).
1462 */
1463 trace_btrfs_space_reservation(fs_info, "transaction",
1464 trans->transid, bytes, 0);
1465 btrfs_block_rsv_release(fs_info, trans->block_rsv, bytes, NULL);
1466 ASSERT(trans->bytes_reserved >= bytes);
1467 trans->bytes_reserved -= bytes;
1468}
1469
1470/* Will return 0, -ENOMEM or -EEXIST (index number collision, unexpected). */
1471int btrfs_insert_delayed_dir_index(struct btrfs_trans_handle *trans,
1472 const char *name, int name_len,
1473 struct btrfs_inode *dir,
1474 struct btrfs_disk_key *disk_key, u8 flags,
1475 u64 index)
1476{
1477 struct btrfs_fs_info *fs_info = trans->fs_info;
1478 const unsigned int leaf_data_size = BTRFS_LEAF_DATA_SIZE(fs_info);
1479 struct btrfs_delayed_node *delayed_node;
1480 struct btrfs_delayed_item *delayed_item;
1481 struct btrfs_dir_item *dir_item;
1482 bool reserve_leaf_space;
1483 u32 data_len;
1484 int ret;
1485
1486 delayed_node = btrfs_get_or_create_delayed_node(dir);
1487 if (IS_ERR(delayed_node))
1488 return PTR_ERR(delayed_node);
1489
1490 delayed_item = btrfs_alloc_delayed_item(sizeof(*dir_item) + name_len,
1491 delayed_node,
1492 BTRFS_DELAYED_INSERTION_ITEM);
1493 if (!delayed_item) {
1494 ret = -ENOMEM;
1495 goto release_node;
1496 }
1497
1498 delayed_item->index = index;
1499
1500 dir_item = (struct btrfs_dir_item *)delayed_item->data;
1501 dir_item->location = *disk_key;
1502 btrfs_set_stack_dir_transid(dir_item, trans->transid);
1503 btrfs_set_stack_dir_data_len(dir_item, 0);
1504 btrfs_set_stack_dir_name_len(dir_item, name_len);
1505 btrfs_set_stack_dir_flags(dir_item, flags);
1506 memcpy((char *)(dir_item + 1), name, name_len);
1507
1508 data_len = delayed_item->data_len + sizeof(struct btrfs_item);
1509
1510 mutex_lock(&delayed_node->mutex);
1511
1512 /*
1513 * First attempt to insert the delayed item. This is to make the error
1514 * handling path simpler in case we fail (-EEXIST). There's no risk of
1515 * any other task coming in and running the delayed item before we do
1516 * the metadata space reservation below, because we are holding the
1517 * delayed node's mutex and that mutex must also be locked before the
1518 * node's delayed items can be run.
1519 */
1520 ret = __btrfs_add_delayed_item(delayed_node, delayed_item);
1521 if (unlikely(ret)) {
1522 btrfs_err(trans->fs_info,
1523"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",
1524 name_len, name, index, btrfs_root_id(delayed_node->root),
1525 delayed_node->inode_id, dir->index_cnt,
1526 delayed_node->index_cnt, ret);
1527 btrfs_release_delayed_item(delayed_item);
1528 btrfs_release_dir_index_item_space(trans);
1529 mutex_unlock(&delayed_node->mutex);
1530 goto release_node;
1531 }
1532
1533 if (delayed_node->index_item_leaves == 0 ||
1534 delayed_node->curr_index_batch_size + data_len > leaf_data_size) {
1535 delayed_node->curr_index_batch_size = data_len;
1536 reserve_leaf_space = true;
1537 } else {
1538 delayed_node->curr_index_batch_size += data_len;
1539 reserve_leaf_space = false;
1540 }
1541
1542 if (reserve_leaf_space) {
1543 ret = btrfs_delayed_item_reserve_metadata(trans, delayed_item);
1544 /*
1545 * Space was reserved for a dir index item insertion when we
1546 * started the transaction, so getting a failure here should be
1547 * impossible.
1548 */
1549 if (WARN_ON(ret)) {
1550 btrfs_release_delayed_item(delayed_item);
1551 mutex_unlock(&delayed_node->mutex);
1552 goto release_node;
1553 }
1554
1555 delayed_node->index_item_leaves++;
1556 } else {
1557 btrfs_release_dir_index_item_space(trans);
1558 }
1559 mutex_unlock(&delayed_node->mutex);
1560
1561release_node:
1562 btrfs_release_delayed_node(delayed_node);
1563 return ret;
1564}
1565
1566static int btrfs_delete_delayed_insertion_item(struct btrfs_fs_info *fs_info,
1567 struct btrfs_delayed_node *node,
1568 u64 index)
1569{
1570 struct btrfs_delayed_item *item;
1571
1572 mutex_lock(&node->mutex);
1573 item = __btrfs_lookup_delayed_item(&node->ins_root.rb_root, index);
1574 if (!item) {
1575 mutex_unlock(&node->mutex);
1576 return 1;
1577 }
1578
1579 /*
1580 * For delayed items to insert, we track reserved metadata bytes based
1581 * on the number of leaves that we will use.
1582 * See btrfs_insert_delayed_dir_index() and
1583 * btrfs_delayed_item_reserve_metadata()).
1584 */
1585 ASSERT(item->bytes_reserved == 0);
1586 ASSERT(node->index_item_leaves > 0);
1587
1588 /*
1589 * If there's only one leaf reserved, we can decrement this item from the
1590 * current batch, otherwise we can not because we don't know which leaf
1591 * it belongs to. With the current limit on delayed items, we rarely
1592 * accumulate enough dir index items to fill more than one leaf (even
1593 * when using a leaf size of 4K).
1594 */
1595 if (node->index_item_leaves == 1) {
1596 const u32 data_len = item->data_len + sizeof(struct btrfs_item);
1597
1598 ASSERT(node->curr_index_batch_size >= data_len);
1599 node->curr_index_batch_size -= data_len;
1600 }
1601
1602 btrfs_release_delayed_item(item);
1603
1604 /* If we now have no more dir index items, we can release all leaves. */
1605 if (RB_EMPTY_ROOT(&node->ins_root.rb_root)) {
1606 btrfs_delayed_item_release_leaves(node, node->index_item_leaves);
1607 node->index_item_leaves = 0;
1608 }
1609
1610 mutex_unlock(&node->mutex);
1611 return 0;
1612}
1613
1614int btrfs_delete_delayed_dir_index(struct btrfs_trans_handle *trans,
1615 struct btrfs_inode *dir, u64 index)
1616{
1617 struct btrfs_delayed_node *node;
1618 struct btrfs_delayed_item *item;
1619 int ret;
1620
1621 node = btrfs_get_or_create_delayed_node(dir);
1622 if (IS_ERR(node))
1623 return PTR_ERR(node);
1624
1625 ret = btrfs_delete_delayed_insertion_item(trans->fs_info, node, index);
1626 if (!ret)
1627 goto end;
1628
1629 item = btrfs_alloc_delayed_item(0, node, BTRFS_DELAYED_DELETION_ITEM);
1630 if (!item) {
1631 ret = -ENOMEM;
1632 goto end;
1633 }
1634
1635 item->index = index;
1636
1637 ret = btrfs_delayed_item_reserve_metadata(trans, item);
1638 /*
1639 * we have reserved enough space when we start a new transaction,
1640 * so reserving metadata failure is impossible.
1641 */
1642 if (ret < 0) {
1643 btrfs_err(trans->fs_info,
1644"metadata reservation failed for delayed dir item deltiona, should have been reserved");
1645 btrfs_release_delayed_item(item);
1646 goto end;
1647 }
1648
1649 mutex_lock(&node->mutex);
1650 ret = __btrfs_add_delayed_item(node, item);
1651 if (unlikely(ret)) {
1652 btrfs_err(trans->fs_info,
1653 "err add delayed dir index item(index: %llu) into the deletion tree of the delayed node(root id: %llu, inode id: %llu, errno: %d)",
1654 index, node->root->root_key.objectid,
1655 node->inode_id, ret);
1656 btrfs_delayed_item_release_metadata(dir->root, item);
1657 btrfs_release_delayed_item(item);
1658 }
1659 mutex_unlock(&node->mutex);
1660end:
1661 btrfs_release_delayed_node(node);
1662 return ret;
1663}
1664
1665int btrfs_inode_delayed_dir_index_count(struct btrfs_inode *inode)
1666{
1667 struct btrfs_delayed_node *delayed_node = btrfs_get_delayed_node(inode);
1668
1669 if (!delayed_node)
1670 return -ENOENT;
1671
1672 /*
1673 * Since we have held i_mutex of this directory, it is impossible that
1674 * a new directory index is added into the delayed node and index_cnt
1675 * is updated now. So we needn't lock the delayed node.
1676 */
1677 if (!delayed_node->index_cnt) {
1678 btrfs_release_delayed_node(delayed_node);
1679 return -EINVAL;
1680 }
1681
1682 inode->index_cnt = delayed_node->index_cnt;
1683 btrfs_release_delayed_node(delayed_node);
1684 return 0;
1685}
1686
1687bool btrfs_readdir_get_delayed_items(struct inode *inode,
1688 u64 last_index,
1689 struct list_head *ins_list,
1690 struct list_head *del_list)
1691{
1692 struct btrfs_delayed_node *delayed_node;
1693 struct btrfs_delayed_item *item;
1694
1695 delayed_node = btrfs_get_delayed_node(BTRFS_I(inode));
1696 if (!delayed_node)
1697 return false;
1698
1699 /*
1700 * We can only do one readdir with delayed items at a time because of
1701 * item->readdir_list.
1702 */
1703 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_SHARED);
1704 btrfs_inode_lock(BTRFS_I(inode), 0);
1705
1706 mutex_lock(&delayed_node->mutex);
1707 item = __btrfs_first_delayed_insertion_item(delayed_node);
1708 while (item && item->index <= last_index) {
1709 refcount_inc(&item->refs);
1710 list_add_tail(&item->readdir_list, ins_list);
1711 item = __btrfs_next_delayed_item(item);
1712 }
1713
1714 item = __btrfs_first_delayed_deletion_item(delayed_node);
1715 while (item && item->index <= last_index) {
1716 refcount_inc(&item->refs);
1717 list_add_tail(&item->readdir_list, del_list);
1718 item = __btrfs_next_delayed_item(item);
1719 }
1720 mutex_unlock(&delayed_node->mutex);
1721 /*
1722 * This delayed node is still cached in the btrfs inode, so refs
1723 * must be > 1 now, and we needn't check it is going to be freed
1724 * or not.
1725 *
1726 * Besides that, this function is used to read dir, we do not
1727 * insert/delete delayed items in this period. So we also needn't
1728 * requeue or dequeue this delayed node.
1729 */
1730 refcount_dec(&delayed_node->refs);
1731
1732 return true;
1733}
1734
1735void btrfs_readdir_put_delayed_items(struct inode *inode,
1736 struct list_head *ins_list,
1737 struct list_head *del_list)
1738{
1739 struct btrfs_delayed_item *curr, *next;
1740
1741 list_for_each_entry_safe(curr, next, ins_list, readdir_list) {
1742 list_del(&curr->readdir_list);
1743 if (refcount_dec_and_test(&curr->refs))
1744 kfree(curr);
1745 }
1746
1747 list_for_each_entry_safe(curr, next, del_list, readdir_list) {
1748 list_del(&curr->readdir_list);
1749 if (refcount_dec_and_test(&curr->refs))
1750 kfree(curr);
1751 }
1752
1753 /*
1754 * The VFS is going to do up_read(), so we need to downgrade back to a
1755 * read lock.
1756 */
1757 downgrade_write(&inode->i_rwsem);
1758}
1759
1760int btrfs_should_delete_dir_index(struct list_head *del_list,
1761 u64 index)
1762{
1763 struct btrfs_delayed_item *curr;
1764 int ret = 0;
1765
1766 list_for_each_entry(curr, del_list, readdir_list) {
1767 if (curr->index > index)
1768 break;
1769 if (curr->index == index) {
1770 ret = 1;
1771 break;
1772 }
1773 }
1774 return ret;
1775}
1776
1777/*
1778 * Read dir info stored in the delayed tree.
1779 */
1780int btrfs_readdir_delayed_dir_index(struct dir_context *ctx,
1781 struct list_head *ins_list)
1782{
1783 struct btrfs_dir_item *di;
1784 struct btrfs_delayed_item *curr, *next;
1785 struct btrfs_key location;
1786 char *name;
1787 int name_len;
1788 int over = 0;
1789 unsigned char d_type;
1790
1791 /*
1792 * Changing the data of the delayed item is impossible. So
1793 * we needn't lock them. And we have held i_mutex of the
1794 * directory, nobody can delete any directory indexes now.
1795 */
1796 list_for_each_entry_safe(curr, next, ins_list, readdir_list) {
1797 list_del(&curr->readdir_list);
1798
1799 if (curr->index < ctx->pos) {
1800 if (refcount_dec_and_test(&curr->refs))
1801 kfree(curr);
1802 continue;
1803 }
1804
1805 ctx->pos = curr->index;
1806
1807 di = (struct btrfs_dir_item *)curr->data;
1808 name = (char *)(di + 1);
1809 name_len = btrfs_stack_dir_name_len(di);
1810
1811 d_type = fs_ftype_to_dtype(btrfs_dir_flags_to_ftype(di->type));
1812 btrfs_disk_key_to_cpu(&location, &di->location);
1813
1814 over = !dir_emit(ctx, name, name_len,
1815 location.objectid, d_type);
1816
1817 if (refcount_dec_and_test(&curr->refs))
1818 kfree(curr);
1819
1820 if (over)
1821 return 1;
1822 ctx->pos++;
1823 }
1824 return 0;
1825}
1826
1827static void fill_stack_inode_item(struct btrfs_trans_handle *trans,
1828 struct btrfs_inode_item *inode_item,
1829 struct inode *inode)
1830{
1831 u64 flags;
1832
1833 btrfs_set_stack_inode_uid(inode_item, i_uid_read(inode));
1834 btrfs_set_stack_inode_gid(inode_item, i_gid_read(inode));
1835 btrfs_set_stack_inode_size(inode_item, BTRFS_I(inode)->disk_i_size);
1836 btrfs_set_stack_inode_mode(inode_item, inode->i_mode);
1837 btrfs_set_stack_inode_nlink(inode_item, inode->i_nlink);
1838 btrfs_set_stack_inode_nbytes(inode_item, inode_get_bytes(inode));
1839 btrfs_set_stack_inode_generation(inode_item,
1840 BTRFS_I(inode)->generation);
1841 btrfs_set_stack_inode_sequence(inode_item,
1842 inode_peek_iversion(inode));
1843 btrfs_set_stack_inode_transid(inode_item, trans->transid);
1844 btrfs_set_stack_inode_rdev(inode_item, inode->i_rdev);
1845 flags = btrfs_inode_combine_flags(BTRFS_I(inode)->flags,
1846 BTRFS_I(inode)->ro_flags);
1847 btrfs_set_stack_inode_flags(inode_item, flags);
1848 btrfs_set_stack_inode_block_group(inode_item, 0);
1849
1850 btrfs_set_stack_timespec_sec(&inode_item->atime,
1851 inode_get_atime_sec(inode));
1852 btrfs_set_stack_timespec_nsec(&inode_item->atime,
1853 inode_get_atime_nsec(inode));
1854
1855 btrfs_set_stack_timespec_sec(&inode_item->mtime,
1856 inode_get_mtime_sec(inode));
1857 btrfs_set_stack_timespec_nsec(&inode_item->mtime,
1858 inode_get_mtime_nsec(inode));
1859
1860 btrfs_set_stack_timespec_sec(&inode_item->ctime,
1861 inode_get_ctime_sec(inode));
1862 btrfs_set_stack_timespec_nsec(&inode_item->ctime,
1863 inode_get_ctime_nsec(inode));
1864
1865 btrfs_set_stack_timespec_sec(&inode_item->otime, BTRFS_I(inode)->i_otime_sec);
1866 btrfs_set_stack_timespec_nsec(&inode_item->otime, BTRFS_I(inode)->i_otime_nsec);
1867}
1868
1869int btrfs_fill_inode(struct inode *inode, u32 *rdev)
1870{
1871 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
1872 struct btrfs_delayed_node *delayed_node;
1873 struct btrfs_inode_item *inode_item;
1874
1875 delayed_node = btrfs_get_delayed_node(BTRFS_I(inode));
1876 if (!delayed_node)
1877 return -ENOENT;
1878
1879 mutex_lock(&delayed_node->mutex);
1880 if (!test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
1881 mutex_unlock(&delayed_node->mutex);
1882 btrfs_release_delayed_node(delayed_node);
1883 return -ENOENT;
1884 }
1885
1886 inode_item = &delayed_node->inode_item;
1887
1888 i_uid_write(inode, btrfs_stack_inode_uid(inode_item));
1889 i_gid_write(inode, btrfs_stack_inode_gid(inode_item));
1890 btrfs_i_size_write(BTRFS_I(inode), btrfs_stack_inode_size(inode_item));
1891 btrfs_inode_set_file_extent_range(BTRFS_I(inode), 0,
1892 round_up(i_size_read(inode), fs_info->sectorsize));
1893 inode->i_mode = btrfs_stack_inode_mode(inode_item);
1894 set_nlink(inode, btrfs_stack_inode_nlink(inode_item));
1895 inode_set_bytes(inode, btrfs_stack_inode_nbytes(inode_item));
1896 BTRFS_I(inode)->generation = btrfs_stack_inode_generation(inode_item);
1897 BTRFS_I(inode)->last_trans = btrfs_stack_inode_transid(inode_item);
1898
1899 inode_set_iversion_queried(inode,
1900 btrfs_stack_inode_sequence(inode_item));
1901 inode->i_rdev = 0;
1902 *rdev = btrfs_stack_inode_rdev(inode_item);
1903 btrfs_inode_split_flags(btrfs_stack_inode_flags(inode_item),
1904 &BTRFS_I(inode)->flags, &BTRFS_I(inode)->ro_flags);
1905
1906 inode_set_atime(inode, btrfs_stack_timespec_sec(&inode_item->atime),
1907 btrfs_stack_timespec_nsec(&inode_item->atime));
1908
1909 inode_set_mtime(inode, btrfs_stack_timespec_sec(&inode_item->mtime),
1910 btrfs_stack_timespec_nsec(&inode_item->mtime));
1911
1912 inode_set_ctime(inode, btrfs_stack_timespec_sec(&inode_item->ctime),
1913 btrfs_stack_timespec_nsec(&inode_item->ctime));
1914
1915 BTRFS_I(inode)->i_otime_sec = btrfs_stack_timespec_sec(&inode_item->otime);
1916 BTRFS_I(inode)->i_otime_nsec = btrfs_stack_timespec_nsec(&inode_item->otime);
1917
1918 inode->i_generation = BTRFS_I(inode)->generation;
1919 BTRFS_I(inode)->index_cnt = (u64)-1;
1920
1921 mutex_unlock(&delayed_node->mutex);
1922 btrfs_release_delayed_node(delayed_node);
1923 return 0;
1924}
1925
1926int btrfs_delayed_update_inode(struct btrfs_trans_handle *trans,
1927 struct btrfs_inode *inode)
1928{
1929 struct btrfs_root *root = inode->root;
1930 struct btrfs_delayed_node *delayed_node;
1931 int ret = 0;
1932
1933 delayed_node = btrfs_get_or_create_delayed_node(inode);
1934 if (IS_ERR(delayed_node))
1935 return PTR_ERR(delayed_node);
1936
1937 mutex_lock(&delayed_node->mutex);
1938 if (test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
1939 fill_stack_inode_item(trans, &delayed_node->inode_item,
1940 &inode->vfs_inode);
1941 goto release_node;
1942 }
1943
1944 ret = btrfs_delayed_inode_reserve_metadata(trans, root, delayed_node);
1945 if (ret)
1946 goto release_node;
1947
1948 fill_stack_inode_item(trans, &delayed_node->inode_item, &inode->vfs_inode);
1949 set_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags);
1950 delayed_node->count++;
1951 atomic_inc(&root->fs_info->delayed_root->items);
1952release_node:
1953 mutex_unlock(&delayed_node->mutex);
1954 btrfs_release_delayed_node(delayed_node);
1955 return ret;
1956}
1957
1958int btrfs_delayed_delete_inode_ref(struct btrfs_inode *inode)
1959{
1960 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1961 struct btrfs_delayed_node *delayed_node;
1962
1963 /*
1964 * we don't do delayed inode updates during log recovery because it
1965 * leads to enospc problems. This means we also can't do
1966 * delayed inode refs
1967 */
1968 if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
1969 return -EAGAIN;
1970
1971 delayed_node = btrfs_get_or_create_delayed_node(inode);
1972 if (IS_ERR(delayed_node))
1973 return PTR_ERR(delayed_node);
1974
1975 /*
1976 * We don't reserve space for inode ref deletion is because:
1977 * - We ONLY do async inode ref deletion for the inode who has only
1978 * one link(i_nlink == 1), it means there is only one inode ref.
1979 * And in most case, the inode ref and the inode item are in the
1980 * same leaf, and we will deal with them at the same time.
1981 * Since we are sure we will reserve the space for the inode item,
1982 * it is unnecessary to reserve space for inode ref deletion.
1983 * - If the inode ref and the inode item are not in the same leaf,
1984 * We also needn't worry about enospc problem, because we reserve
1985 * much more space for the inode update than it needs.
1986 * - At the worst, we can steal some space from the global reservation.
1987 * It is very rare.
1988 */
1989 mutex_lock(&delayed_node->mutex);
1990 if (test_bit(BTRFS_DELAYED_NODE_DEL_IREF, &delayed_node->flags))
1991 goto release_node;
1992
1993 set_bit(BTRFS_DELAYED_NODE_DEL_IREF, &delayed_node->flags);
1994 delayed_node->count++;
1995 atomic_inc(&fs_info->delayed_root->items);
1996release_node:
1997 mutex_unlock(&delayed_node->mutex);
1998 btrfs_release_delayed_node(delayed_node);
1999 return 0;
2000}
2001
2002static void __btrfs_kill_delayed_node(struct btrfs_delayed_node *delayed_node)
2003{
2004 struct btrfs_root *root = delayed_node->root;
2005 struct btrfs_fs_info *fs_info = root->fs_info;
2006 struct btrfs_delayed_item *curr_item, *prev_item;
2007
2008 mutex_lock(&delayed_node->mutex);
2009 curr_item = __btrfs_first_delayed_insertion_item(delayed_node);
2010 while (curr_item) {
2011 prev_item = curr_item;
2012 curr_item = __btrfs_next_delayed_item(prev_item);
2013 btrfs_release_delayed_item(prev_item);
2014 }
2015
2016 if (delayed_node->index_item_leaves > 0) {
2017 btrfs_delayed_item_release_leaves(delayed_node,
2018 delayed_node->index_item_leaves);
2019 delayed_node->index_item_leaves = 0;
2020 }
2021
2022 curr_item = __btrfs_first_delayed_deletion_item(delayed_node);
2023 while (curr_item) {
2024 btrfs_delayed_item_release_metadata(root, curr_item);
2025 prev_item = curr_item;
2026 curr_item = __btrfs_next_delayed_item(prev_item);
2027 btrfs_release_delayed_item(prev_item);
2028 }
2029
2030 btrfs_release_delayed_iref(delayed_node);
2031
2032 if (test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
2033 btrfs_delayed_inode_release_metadata(fs_info, delayed_node, false);
2034 btrfs_release_delayed_inode(delayed_node);
2035 }
2036 mutex_unlock(&delayed_node->mutex);
2037}
2038
2039void btrfs_kill_delayed_inode_items(struct btrfs_inode *inode)
2040{
2041 struct btrfs_delayed_node *delayed_node;
2042
2043 delayed_node = btrfs_get_delayed_node(inode);
2044 if (!delayed_node)
2045 return;
2046
2047 __btrfs_kill_delayed_node(delayed_node);
2048 btrfs_release_delayed_node(delayed_node);
2049}
2050
2051void btrfs_kill_all_delayed_nodes(struct btrfs_root *root)
2052{
2053 unsigned long index = 0;
2054 struct btrfs_delayed_node *delayed_nodes[8];
2055
2056 while (1) {
2057 struct btrfs_delayed_node *node;
2058 int count;
2059
2060 spin_lock(&root->inode_lock);
2061 if (xa_empty(&root->delayed_nodes)) {
2062 spin_unlock(&root->inode_lock);
2063 return;
2064 }
2065
2066 count = 0;
2067 xa_for_each_start(&root->delayed_nodes, index, node, index) {
2068 /*
2069 * Don't increase refs in case the node is dead and
2070 * about to be removed from the tree in the loop below
2071 */
2072 if (refcount_inc_not_zero(&node->refs)) {
2073 delayed_nodes[count] = node;
2074 count++;
2075 }
2076 if (count >= ARRAY_SIZE(delayed_nodes))
2077 break;
2078 }
2079 spin_unlock(&root->inode_lock);
2080 index++;
2081
2082 for (int i = 0; i < count; i++) {
2083 __btrfs_kill_delayed_node(delayed_nodes[i]);
2084 btrfs_release_delayed_node(delayed_nodes[i]);
2085 }
2086 }
2087}
2088
2089void btrfs_destroy_delayed_inodes(struct btrfs_fs_info *fs_info)
2090{
2091 struct btrfs_delayed_node *curr_node, *prev_node;
2092
2093 curr_node = btrfs_first_delayed_node(fs_info->delayed_root);
2094 while (curr_node) {
2095 __btrfs_kill_delayed_node(curr_node);
2096
2097 prev_node = curr_node;
2098 curr_node = btrfs_next_delayed_node(curr_node);
2099 btrfs_release_delayed_node(prev_node);
2100 }
2101}
2102
2103void btrfs_log_get_delayed_items(struct btrfs_inode *inode,
2104 struct list_head *ins_list,
2105 struct list_head *del_list)
2106{
2107 struct btrfs_delayed_node *node;
2108 struct btrfs_delayed_item *item;
2109
2110 node = btrfs_get_delayed_node(inode);
2111 if (!node)
2112 return;
2113
2114 mutex_lock(&node->mutex);
2115 item = __btrfs_first_delayed_insertion_item(node);
2116 while (item) {
2117 /*
2118 * It's possible that the item is already in a log list. This
2119 * can happen in case two tasks are trying to log the same
2120 * directory. For example if we have tasks A and task B:
2121 *
2122 * Task A collected the delayed items into a log list while
2123 * under the inode's log_mutex (at btrfs_log_inode()), but it
2124 * only releases the items after logging the inodes they point
2125 * to (if they are new inodes), which happens after unlocking
2126 * the log mutex;
2127 *
2128 * Task B enters btrfs_log_inode() and acquires the log_mutex
2129 * of the same directory inode, before task B releases the
2130 * delayed items. This can happen for example when logging some
2131 * inode we need to trigger logging of its parent directory, so
2132 * logging two files that have the same parent directory can
2133 * lead to this.
2134 *
2135 * If this happens, just ignore delayed items already in a log
2136 * list. All the tasks logging the directory are under a log
2137 * transaction and whichever finishes first can not sync the log
2138 * before the other completes and leaves the log transaction.
2139 */
2140 if (!item->logged && list_empty(&item->log_list)) {
2141 refcount_inc(&item->refs);
2142 list_add_tail(&item->log_list, ins_list);
2143 }
2144 item = __btrfs_next_delayed_item(item);
2145 }
2146
2147 item = __btrfs_first_delayed_deletion_item(node);
2148 while (item) {
2149 /* It may be non-empty, for the same reason mentioned above. */
2150 if (!item->logged && list_empty(&item->log_list)) {
2151 refcount_inc(&item->refs);
2152 list_add_tail(&item->log_list, del_list);
2153 }
2154 item = __btrfs_next_delayed_item(item);
2155 }
2156 mutex_unlock(&node->mutex);
2157
2158 /*
2159 * We are called during inode logging, which means the inode is in use
2160 * and can not be evicted before we finish logging the inode. So we never
2161 * have the last reference on the delayed inode.
2162 * Also, we don't use btrfs_release_delayed_node() because that would
2163 * requeue the delayed inode (change its order in the list of prepared
2164 * nodes) and we don't want to do such change because we don't create or
2165 * delete delayed items.
2166 */
2167 ASSERT(refcount_read(&node->refs) > 1);
2168 refcount_dec(&node->refs);
2169}
2170
2171void btrfs_log_put_delayed_items(struct btrfs_inode *inode,
2172 struct list_head *ins_list,
2173 struct list_head *del_list)
2174{
2175 struct btrfs_delayed_node *node;
2176 struct btrfs_delayed_item *item;
2177 struct btrfs_delayed_item *next;
2178
2179 node = btrfs_get_delayed_node(inode);
2180 if (!node)
2181 return;
2182
2183 mutex_lock(&node->mutex);
2184
2185 list_for_each_entry_safe(item, next, ins_list, log_list) {
2186 item->logged = true;
2187 list_del_init(&item->log_list);
2188 if (refcount_dec_and_test(&item->refs))
2189 kfree(item);
2190 }
2191
2192 list_for_each_entry_safe(item, next, del_list, log_list) {
2193 item->logged = true;
2194 list_del_init(&item->log_list);
2195 if (refcount_dec_and_test(&item->refs))
2196 kfree(item);
2197 }
2198
2199 mutex_unlock(&node->mutex);
2200
2201 /*
2202 * We are called during inode logging, which means the inode is in use
2203 * and can not be evicted before we finish logging the inode. So we never
2204 * have the last reference on the delayed inode.
2205 * Also, we don't use btrfs_release_delayed_node() because that would
2206 * requeue the delayed inode (change its order in the list of prepared
2207 * nodes) and we don't want to do such change because we don't create or
2208 * delete delayed items.
2209 */
2210 ASSERT(refcount_read(&node->refs) > 1);
2211 refcount_dec(&node->refs);
2212}