Loading...
1/*
2 * Copyright (C) 2009 Oracle. All rights reserved.
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public
6 * License v2 as published by the Free Software Foundation.
7 *
8 * This program is distributed in the hope that it will be useful,
9 * but WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
11 * General Public License for more details.
12 *
13 * You should have received a copy of the GNU General Public
14 * License along with this program; if not, write to the
15 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16 * Boston, MA 021110-1307, USA.
17 */
18
19#include <linux/sched.h>
20#include <linux/slab.h>
21#include <linux/sort.h>
22#include "ctree.h"
23#include "delayed-ref.h"
24#include "transaction.h"
25
26/*
27 * delayed back reference update tracking. For subvolume trees
28 * we queue up extent allocations and backref maintenance for
29 * delayed processing. This avoids deep call chains where we
30 * add extents in the middle of btrfs_search_slot, and it allows
31 * us to buffer up frequently modified backrefs in an rb tree instead
32 * of hammering updates on the extent allocation tree.
33 */
34
35/*
36 * compare two delayed tree backrefs with same bytenr and type
37 */
38static int comp_tree_refs(struct btrfs_delayed_tree_ref *ref2,
39 struct btrfs_delayed_tree_ref *ref1)
40{
41 if (ref1->node.type == BTRFS_TREE_BLOCK_REF_KEY) {
42 if (ref1->root < ref2->root)
43 return -1;
44 if (ref1->root > ref2->root)
45 return 1;
46 } else {
47 if (ref1->parent < ref2->parent)
48 return -1;
49 if (ref1->parent > ref2->parent)
50 return 1;
51 }
52 return 0;
53}
54
55/*
56 * compare two delayed data backrefs with same bytenr and type
57 */
58static int comp_data_refs(struct btrfs_delayed_data_ref *ref2,
59 struct btrfs_delayed_data_ref *ref1)
60{
61 if (ref1->node.type == BTRFS_EXTENT_DATA_REF_KEY) {
62 if (ref1->root < ref2->root)
63 return -1;
64 if (ref1->root > ref2->root)
65 return 1;
66 if (ref1->objectid < ref2->objectid)
67 return -1;
68 if (ref1->objectid > ref2->objectid)
69 return 1;
70 if (ref1->offset < ref2->offset)
71 return -1;
72 if (ref1->offset > ref2->offset)
73 return 1;
74 } else {
75 if (ref1->parent < ref2->parent)
76 return -1;
77 if (ref1->parent > ref2->parent)
78 return 1;
79 }
80 return 0;
81}
82
83/*
84 * entries in the rb tree are ordered by the byte number of the extent,
85 * type of the delayed backrefs and content of delayed backrefs.
86 */
87static int comp_entry(struct btrfs_delayed_ref_node *ref2,
88 struct btrfs_delayed_ref_node *ref1)
89{
90 if (ref1->bytenr < ref2->bytenr)
91 return -1;
92 if (ref1->bytenr > ref2->bytenr)
93 return 1;
94 if (ref1->is_head && ref2->is_head)
95 return 0;
96 if (ref2->is_head)
97 return -1;
98 if (ref1->is_head)
99 return 1;
100 if (ref1->type < ref2->type)
101 return -1;
102 if (ref1->type > ref2->type)
103 return 1;
104 if (ref1->type == BTRFS_TREE_BLOCK_REF_KEY ||
105 ref1->type == BTRFS_SHARED_BLOCK_REF_KEY) {
106 return comp_tree_refs(btrfs_delayed_node_to_tree_ref(ref2),
107 btrfs_delayed_node_to_tree_ref(ref1));
108 } else if (ref1->type == BTRFS_EXTENT_DATA_REF_KEY ||
109 ref1->type == BTRFS_SHARED_DATA_REF_KEY) {
110 return comp_data_refs(btrfs_delayed_node_to_data_ref(ref2),
111 btrfs_delayed_node_to_data_ref(ref1));
112 }
113 BUG();
114 return 0;
115}
116
117/*
118 * insert a new ref into the rbtree. This returns any existing refs
119 * for the same (bytenr,parent) tuple, or NULL if the new node was properly
120 * inserted.
121 */
122static struct btrfs_delayed_ref_node *tree_insert(struct rb_root *root,
123 struct rb_node *node)
124{
125 struct rb_node **p = &root->rb_node;
126 struct rb_node *parent_node = NULL;
127 struct btrfs_delayed_ref_node *entry;
128 struct btrfs_delayed_ref_node *ins;
129 int cmp;
130
131 ins = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
132 while (*p) {
133 parent_node = *p;
134 entry = rb_entry(parent_node, struct btrfs_delayed_ref_node,
135 rb_node);
136
137 cmp = comp_entry(entry, ins);
138 if (cmp < 0)
139 p = &(*p)->rb_left;
140 else if (cmp > 0)
141 p = &(*p)->rb_right;
142 else
143 return entry;
144 }
145
146 rb_link_node(node, parent_node, p);
147 rb_insert_color(node, root);
148 return NULL;
149}
150
151/*
152 * find an head entry based on bytenr. This returns the delayed ref
153 * head if it was able to find one, or NULL if nothing was in that spot
154 */
155static struct btrfs_delayed_ref_node *find_ref_head(struct rb_root *root,
156 u64 bytenr,
157 struct btrfs_delayed_ref_node **last)
158{
159 struct rb_node *n = root->rb_node;
160 struct btrfs_delayed_ref_node *entry;
161 int cmp;
162
163 while (n) {
164 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
165 WARN_ON(!entry->in_tree);
166 if (last)
167 *last = entry;
168
169 if (bytenr < entry->bytenr)
170 cmp = -1;
171 else if (bytenr > entry->bytenr)
172 cmp = 1;
173 else if (!btrfs_delayed_ref_is_head(entry))
174 cmp = 1;
175 else
176 cmp = 0;
177
178 if (cmp < 0)
179 n = n->rb_left;
180 else if (cmp > 0)
181 n = n->rb_right;
182 else
183 return entry;
184 }
185 return NULL;
186}
187
188int btrfs_delayed_ref_lock(struct btrfs_trans_handle *trans,
189 struct btrfs_delayed_ref_head *head)
190{
191 struct btrfs_delayed_ref_root *delayed_refs;
192
193 delayed_refs = &trans->transaction->delayed_refs;
194 assert_spin_locked(&delayed_refs->lock);
195 if (mutex_trylock(&head->mutex))
196 return 0;
197
198 atomic_inc(&head->node.refs);
199 spin_unlock(&delayed_refs->lock);
200
201 mutex_lock(&head->mutex);
202 spin_lock(&delayed_refs->lock);
203 if (!head->node.in_tree) {
204 mutex_unlock(&head->mutex);
205 btrfs_put_delayed_ref(&head->node);
206 return -EAGAIN;
207 }
208 btrfs_put_delayed_ref(&head->node);
209 return 0;
210}
211
212int btrfs_find_ref_cluster(struct btrfs_trans_handle *trans,
213 struct list_head *cluster, u64 start)
214{
215 int count = 0;
216 struct btrfs_delayed_ref_root *delayed_refs;
217 struct rb_node *node;
218 struct btrfs_delayed_ref_node *ref;
219 struct btrfs_delayed_ref_head *head;
220
221 delayed_refs = &trans->transaction->delayed_refs;
222 if (start == 0) {
223 node = rb_first(&delayed_refs->root);
224 } else {
225 ref = NULL;
226 find_ref_head(&delayed_refs->root, start, &ref);
227 if (ref) {
228 struct btrfs_delayed_ref_node *tmp;
229
230 node = rb_prev(&ref->rb_node);
231 while (node) {
232 tmp = rb_entry(node,
233 struct btrfs_delayed_ref_node,
234 rb_node);
235 if (tmp->bytenr < start)
236 break;
237 ref = tmp;
238 node = rb_prev(&ref->rb_node);
239 }
240 node = &ref->rb_node;
241 } else
242 node = rb_first(&delayed_refs->root);
243 }
244again:
245 while (node && count < 32) {
246 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
247 if (btrfs_delayed_ref_is_head(ref)) {
248 head = btrfs_delayed_node_to_head(ref);
249 if (list_empty(&head->cluster)) {
250 list_add_tail(&head->cluster, cluster);
251 delayed_refs->run_delayed_start =
252 head->node.bytenr;
253 count++;
254
255 WARN_ON(delayed_refs->num_heads_ready == 0);
256 delayed_refs->num_heads_ready--;
257 } else if (count) {
258 /* the goal of the clustering is to find extents
259 * that are likely to end up in the same extent
260 * leaf on disk. So, we don't want them spread
261 * all over the tree. Stop now if we've hit
262 * a head that was already in use
263 */
264 break;
265 }
266 }
267 node = rb_next(node);
268 }
269 if (count) {
270 return 0;
271 } else if (start) {
272 /*
273 * we've gone to the end of the rbtree without finding any
274 * clusters. start from the beginning and try again
275 */
276 start = 0;
277 node = rb_first(&delayed_refs->root);
278 goto again;
279 }
280 return 1;
281}
282
283/*
284 * helper function to update an extent delayed ref in the
285 * rbtree. existing and update must both have the same
286 * bytenr and parent
287 *
288 * This may free existing if the update cancels out whatever
289 * operation it was doing.
290 */
291static noinline void
292update_existing_ref(struct btrfs_trans_handle *trans,
293 struct btrfs_delayed_ref_root *delayed_refs,
294 struct btrfs_delayed_ref_node *existing,
295 struct btrfs_delayed_ref_node *update)
296{
297 if (update->action != existing->action) {
298 /*
299 * this is effectively undoing either an add or a
300 * drop. We decrement the ref_mod, and if it goes
301 * down to zero we just delete the entry without
302 * every changing the extent allocation tree.
303 */
304 existing->ref_mod--;
305 if (existing->ref_mod == 0) {
306 rb_erase(&existing->rb_node,
307 &delayed_refs->root);
308 existing->in_tree = 0;
309 btrfs_put_delayed_ref(existing);
310 delayed_refs->num_entries--;
311 if (trans->delayed_ref_updates)
312 trans->delayed_ref_updates--;
313 } else {
314 WARN_ON(existing->type == BTRFS_TREE_BLOCK_REF_KEY ||
315 existing->type == BTRFS_SHARED_BLOCK_REF_KEY);
316 }
317 } else {
318 WARN_ON(existing->type == BTRFS_TREE_BLOCK_REF_KEY ||
319 existing->type == BTRFS_SHARED_BLOCK_REF_KEY);
320 /*
321 * the action on the existing ref matches
322 * the action on the ref we're trying to add.
323 * Bump the ref_mod by one so the backref that
324 * is eventually added/removed has the correct
325 * reference count
326 */
327 existing->ref_mod += update->ref_mod;
328 }
329}
330
331/*
332 * helper function to update the accounting in the head ref
333 * existing and update must have the same bytenr
334 */
335static noinline void
336update_existing_head_ref(struct btrfs_delayed_ref_node *existing,
337 struct btrfs_delayed_ref_node *update)
338{
339 struct btrfs_delayed_ref_head *existing_ref;
340 struct btrfs_delayed_ref_head *ref;
341
342 existing_ref = btrfs_delayed_node_to_head(existing);
343 ref = btrfs_delayed_node_to_head(update);
344 BUG_ON(existing_ref->is_data != ref->is_data);
345
346 if (ref->must_insert_reserved) {
347 /* if the extent was freed and then
348 * reallocated before the delayed ref
349 * entries were processed, we can end up
350 * with an existing head ref without
351 * the must_insert_reserved flag set.
352 * Set it again here
353 */
354 existing_ref->must_insert_reserved = ref->must_insert_reserved;
355
356 /*
357 * update the num_bytes so we make sure the accounting
358 * is done correctly
359 */
360 existing->num_bytes = update->num_bytes;
361
362 }
363
364 if (ref->extent_op) {
365 if (!existing_ref->extent_op) {
366 existing_ref->extent_op = ref->extent_op;
367 } else {
368 if (ref->extent_op->update_key) {
369 memcpy(&existing_ref->extent_op->key,
370 &ref->extent_op->key,
371 sizeof(ref->extent_op->key));
372 existing_ref->extent_op->update_key = 1;
373 }
374 if (ref->extent_op->update_flags) {
375 existing_ref->extent_op->flags_to_set |=
376 ref->extent_op->flags_to_set;
377 existing_ref->extent_op->update_flags = 1;
378 }
379 kfree(ref->extent_op);
380 }
381 }
382 /*
383 * update the reference mod on the head to reflect this new operation
384 */
385 existing->ref_mod += update->ref_mod;
386}
387
388/*
389 * helper function to actually insert a head node into the rbtree.
390 * this does all the dirty work in terms of maintaining the correct
391 * overall modification count.
392 */
393static noinline int add_delayed_ref_head(struct btrfs_trans_handle *trans,
394 struct btrfs_delayed_ref_node *ref,
395 u64 bytenr, u64 num_bytes,
396 int action, int is_data)
397{
398 struct btrfs_delayed_ref_node *existing;
399 struct btrfs_delayed_ref_head *head_ref = NULL;
400 struct btrfs_delayed_ref_root *delayed_refs;
401 int count_mod = 1;
402 int must_insert_reserved = 0;
403
404 /*
405 * the head node stores the sum of all the mods, so dropping a ref
406 * should drop the sum in the head node by one.
407 */
408 if (action == BTRFS_UPDATE_DELAYED_HEAD)
409 count_mod = 0;
410 else if (action == BTRFS_DROP_DELAYED_REF)
411 count_mod = -1;
412
413 /*
414 * BTRFS_ADD_DELAYED_EXTENT means that we need to update
415 * the reserved accounting when the extent is finally added, or
416 * if a later modification deletes the delayed ref without ever
417 * inserting the extent into the extent allocation tree.
418 * ref->must_insert_reserved is the flag used to record
419 * that accounting mods are required.
420 *
421 * Once we record must_insert_reserved, switch the action to
422 * BTRFS_ADD_DELAYED_REF because other special casing is not required.
423 */
424 if (action == BTRFS_ADD_DELAYED_EXTENT)
425 must_insert_reserved = 1;
426 else
427 must_insert_reserved = 0;
428
429 delayed_refs = &trans->transaction->delayed_refs;
430
431 /* first set the basic ref node struct up */
432 atomic_set(&ref->refs, 1);
433 ref->bytenr = bytenr;
434 ref->num_bytes = num_bytes;
435 ref->ref_mod = count_mod;
436 ref->type = 0;
437 ref->action = 0;
438 ref->is_head = 1;
439 ref->in_tree = 1;
440
441 head_ref = btrfs_delayed_node_to_head(ref);
442 head_ref->must_insert_reserved = must_insert_reserved;
443 head_ref->is_data = is_data;
444
445 INIT_LIST_HEAD(&head_ref->cluster);
446 mutex_init(&head_ref->mutex);
447
448 trace_btrfs_delayed_ref_head(ref, head_ref, action);
449
450 existing = tree_insert(&delayed_refs->root, &ref->rb_node);
451
452 if (existing) {
453 update_existing_head_ref(existing, ref);
454 /*
455 * we've updated the existing ref, free the newly
456 * allocated ref
457 */
458 kfree(ref);
459 } else {
460 delayed_refs->num_heads++;
461 delayed_refs->num_heads_ready++;
462 delayed_refs->num_entries++;
463 trans->delayed_ref_updates++;
464 }
465 return 0;
466}
467
468/*
469 * helper to insert a delayed tree ref into the rbtree.
470 */
471static noinline int add_delayed_tree_ref(struct btrfs_trans_handle *trans,
472 struct btrfs_delayed_ref_node *ref,
473 u64 bytenr, u64 num_bytes, u64 parent,
474 u64 ref_root, int level, int action)
475{
476 struct btrfs_delayed_ref_node *existing;
477 struct btrfs_delayed_tree_ref *full_ref;
478 struct btrfs_delayed_ref_root *delayed_refs;
479
480 if (action == BTRFS_ADD_DELAYED_EXTENT)
481 action = BTRFS_ADD_DELAYED_REF;
482
483 delayed_refs = &trans->transaction->delayed_refs;
484
485 /* first set the basic ref node struct up */
486 atomic_set(&ref->refs, 1);
487 ref->bytenr = bytenr;
488 ref->num_bytes = num_bytes;
489 ref->ref_mod = 1;
490 ref->action = action;
491 ref->is_head = 0;
492 ref->in_tree = 1;
493
494 full_ref = btrfs_delayed_node_to_tree_ref(ref);
495 if (parent) {
496 full_ref->parent = parent;
497 ref->type = BTRFS_SHARED_BLOCK_REF_KEY;
498 } else {
499 full_ref->root = ref_root;
500 ref->type = BTRFS_TREE_BLOCK_REF_KEY;
501 }
502 full_ref->level = level;
503
504 trace_btrfs_delayed_tree_ref(ref, full_ref, action);
505
506 existing = tree_insert(&delayed_refs->root, &ref->rb_node);
507
508 if (existing) {
509 update_existing_ref(trans, delayed_refs, existing, ref);
510 /*
511 * we've updated the existing ref, free the newly
512 * allocated ref
513 */
514 kfree(ref);
515 } else {
516 delayed_refs->num_entries++;
517 trans->delayed_ref_updates++;
518 }
519 return 0;
520}
521
522/*
523 * helper to insert a delayed data ref into the rbtree.
524 */
525static noinline int add_delayed_data_ref(struct btrfs_trans_handle *trans,
526 struct btrfs_delayed_ref_node *ref,
527 u64 bytenr, u64 num_bytes, u64 parent,
528 u64 ref_root, u64 owner, u64 offset,
529 int action)
530{
531 struct btrfs_delayed_ref_node *existing;
532 struct btrfs_delayed_data_ref *full_ref;
533 struct btrfs_delayed_ref_root *delayed_refs;
534
535 if (action == BTRFS_ADD_DELAYED_EXTENT)
536 action = BTRFS_ADD_DELAYED_REF;
537
538 delayed_refs = &trans->transaction->delayed_refs;
539
540 /* first set the basic ref node struct up */
541 atomic_set(&ref->refs, 1);
542 ref->bytenr = bytenr;
543 ref->num_bytes = num_bytes;
544 ref->ref_mod = 1;
545 ref->action = action;
546 ref->is_head = 0;
547 ref->in_tree = 1;
548
549 full_ref = btrfs_delayed_node_to_data_ref(ref);
550 if (parent) {
551 full_ref->parent = parent;
552 ref->type = BTRFS_SHARED_DATA_REF_KEY;
553 } else {
554 full_ref->root = ref_root;
555 ref->type = BTRFS_EXTENT_DATA_REF_KEY;
556 }
557 full_ref->objectid = owner;
558 full_ref->offset = offset;
559
560 trace_btrfs_delayed_data_ref(ref, full_ref, action);
561
562 existing = tree_insert(&delayed_refs->root, &ref->rb_node);
563
564 if (existing) {
565 update_existing_ref(trans, delayed_refs, existing, ref);
566 /*
567 * we've updated the existing ref, free the newly
568 * allocated ref
569 */
570 kfree(ref);
571 } else {
572 delayed_refs->num_entries++;
573 trans->delayed_ref_updates++;
574 }
575 return 0;
576}
577
578/*
579 * add a delayed tree ref. This does all of the accounting required
580 * to make sure the delayed ref is eventually processed before this
581 * transaction commits.
582 */
583int btrfs_add_delayed_tree_ref(struct btrfs_trans_handle *trans,
584 u64 bytenr, u64 num_bytes, u64 parent,
585 u64 ref_root, int level, int action,
586 struct btrfs_delayed_extent_op *extent_op)
587{
588 struct btrfs_delayed_tree_ref *ref;
589 struct btrfs_delayed_ref_head *head_ref;
590 struct btrfs_delayed_ref_root *delayed_refs;
591 int ret;
592
593 BUG_ON(extent_op && extent_op->is_data);
594 ref = kmalloc(sizeof(*ref), GFP_NOFS);
595 if (!ref)
596 return -ENOMEM;
597
598 head_ref = kmalloc(sizeof(*head_ref), GFP_NOFS);
599 if (!head_ref) {
600 kfree(ref);
601 return -ENOMEM;
602 }
603
604 head_ref->extent_op = extent_op;
605
606 delayed_refs = &trans->transaction->delayed_refs;
607 spin_lock(&delayed_refs->lock);
608
609 /*
610 * insert both the head node and the new ref without dropping
611 * the spin lock
612 */
613 ret = add_delayed_ref_head(trans, &head_ref->node, bytenr, num_bytes,
614 action, 0);
615 BUG_ON(ret);
616
617 ret = add_delayed_tree_ref(trans, &ref->node, bytenr, num_bytes,
618 parent, ref_root, level, action);
619 BUG_ON(ret);
620 spin_unlock(&delayed_refs->lock);
621 return 0;
622}
623
624/*
625 * add a delayed data ref. it's similar to btrfs_add_delayed_tree_ref.
626 */
627int btrfs_add_delayed_data_ref(struct btrfs_trans_handle *trans,
628 u64 bytenr, u64 num_bytes,
629 u64 parent, u64 ref_root,
630 u64 owner, u64 offset, int action,
631 struct btrfs_delayed_extent_op *extent_op)
632{
633 struct btrfs_delayed_data_ref *ref;
634 struct btrfs_delayed_ref_head *head_ref;
635 struct btrfs_delayed_ref_root *delayed_refs;
636 int ret;
637
638 BUG_ON(extent_op && !extent_op->is_data);
639 ref = kmalloc(sizeof(*ref), GFP_NOFS);
640 if (!ref)
641 return -ENOMEM;
642
643 head_ref = kmalloc(sizeof(*head_ref), GFP_NOFS);
644 if (!head_ref) {
645 kfree(ref);
646 return -ENOMEM;
647 }
648
649 head_ref->extent_op = extent_op;
650
651 delayed_refs = &trans->transaction->delayed_refs;
652 spin_lock(&delayed_refs->lock);
653
654 /*
655 * insert both the head node and the new ref without dropping
656 * the spin lock
657 */
658 ret = add_delayed_ref_head(trans, &head_ref->node, bytenr, num_bytes,
659 action, 1);
660 BUG_ON(ret);
661
662 ret = add_delayed_data_ref(trans, &ref->node, bytenr, num_bytes,
663 parent, ref_root, owner, offset, action);
664 BUG_ON(ret);
665 spin_unlock(&delayed_refs->lock);
666 return 0;
667}
668
669int btrfs_add_delayed_extent_op(struct btrfs_trans_handle *trans,
670 u64 bytenr, u64 num_bytes,
671 struct btrfs_delayed_extent_op *extent_op)
672{
673 struct btrfs_delayed_ref_head *head_ref;
674 struct btrfs_delayed_ref_root *delayed_refs;
675 int ret;
676
677 head_ref = kmalloc(sizeof(*head_ref), GFP_NOFS);
678 if (!head_ref)
679 return -ENOMEM;
680
681 head_ref->extent_op = extent_op;
682
683 delayed_refs = &trans->transaction->delayed_refs;
684 spin_lock(&delayed_refs->lock);
685
686 ret = add_delayed_ref_head(trans, &head_ref->node, bytenr,
687 num_bytes, BTRFS_UPDATE_DELAYED_HEAD,
688 extent_op->is_data);
689 BUG_ON(ret);
690
691 spin_unlock(&delayed_refs->lock);
692 return 0;
693}
694
695/*
696 * this does a simple search for the head node for a given extent.
697 * It must be called with the delayed ref spinlock held, and it returns
698 * the head node if any where found, or NULL if not.
699 */
700struct btrfs_delayed_ref_head *
701btrfs_find_delayed_ref_head(struct btrfs_trans_handle *trans, u64 bytenr)
702{
703 struct btrfs_delayed_ref_node *ref;
704 struct btrfs_delayed_ref_root *delayed_refs;
705
706 delayed_refs = &trans->transaction->delayed_refs;
707 ref = find_ref_head(&delayed_refs->root, bytenr, NULL);
708 if (ref)
709 return btrfs_delayed_node_to_head(ref);
710 return NULL;
711}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2009 Oracle. All rights reserved.
4 */
5
6#include <linux/sched.h>
7#include <linux/slab.h>
8#include <linux/sort.h>
9#include "ctree.h"
10#include "delayed-ref.h"
11#include "transaction.h"
12#include "qgroup.h"
13#include "space-info.h"
14#include "tree-mod-log.h"
15
16struct kmem_cache *btrfs_delayed_ref_head_cachep;
17struct kmem_cache *btrfs_delayed_tree_ref_cachep;
18struct kmem_cache *btrfs_delayed_data_ref_cachep;
19struct kmem_cache *btrfs_delayed_extent_op_cachep;
20/*
21 * delayed back reference update tracking. For subvolume trees
22 * we queue up extent allocations and backref maintenance for
23 * delayed processing. This avoids deep call chains where we
24 * add extents in the middle of btrfs_search_slot, and it allows
25 * us to buffer up frequently modified backrefs in an rb tree instead
26 * of hammering updates on the extent allocation tree.
27 */
28
29bool btrfs_check_space_for_delayed_refs(struct btrfs_fs_info *fs_info)
30{
31 struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
32 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
33 bool ret = false;
34 u64 reserved;
35
36 spin_lock(&global_rsv->lock);
37 reserved = global_rsv->reserved;
38 spin_unlock(&global_rsv->lock);
39
40 /*
41 * Since the global reserve is just kind of magic we don't really want
42 * to rely on it to save our bacon, so if our size is more than the
43 * delayed_refs_rsv and the global rsv then it's time to think about
44 * bailing.
45 */
46 spin_lock(&delayed_refs_rsv->lock);
47 reserved += delayed_refs_rsv->reserved;
48 if (delayed_refs_rsv->size >= reserved)
49 ret = true;
50 spin_unlock(&delayed_refs_rsv->lock);
51 return ret;
52}
53
54int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans)
55{
56 u64 num_entries =
57 atomic_read(&trans->transaction->delayed_refs.num_entries);
58 u64 avg_runtime;
59 u64 val;
60
61 smp_mb();
62 avg_runtime = trans->fs_info->avg_delayed_ref_runtime;
63 val = num_entries * avg_runtime;
64 if (val >= NSEC_PER_SEC)
65 return 1;
66 if (val >= NSEC_PER_SEC / 2)
67 return 2;
68
69 return btrfs_check_space_for_delayed_refs(trans->fs_info);
70}
71
72/**
73 * Release a ref head's reservation
74 *
75 * @fs_info: the filesystem
76 * @nr: number of items to drop
77 *
78 * This drops the delayed ref head's count from the delayed refs rsv and frees
79 * any excess reservation we had.
80 */
81void btrfs_delayed_refs_rsv_release(struct btrfs_fs_info *fs_info, int nr)
82{
83 struct btrfs_block_rsv *block_rsv = &fs_info->delayed_refs_rsv;
84 u64 num_bytes = btrfs_calc_insert_metadata_size(fs_info, nr);
85 u64 released = 0;
86
87 released = btrfs_block_rsv_release(fs_info, block_rsv, num_bytes, NULL);
88 if (released)
89 trace_btrfs_space_reservation(fs_info, "delayed_refs_rsv",
90 0, released, 0);
91}
92
93/*
94 * btrfs_update_delayed_refs_rsv - adjust the size of the delayed refs rsv
95 * @trans - the trans that may have generated delayed refs
96 *
97 * This is to be called anytime we may have adjusted trans->delayed_ref_updates,
98 * it'll calculate the additional size and add it to the delayed_refs_rsv.
99 */
100void btrfs_update_delayed_refs_rsv(struct btrfs_trans_handle *trans)
101{
102 struct btrfs_fs_info *fs_info = trans->fs_info;
103 struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_refs_rsv;
104 u64 num_bytes;
105
106 if (!trans->delayed_ref_updates)
107 return;
108
109 num_bytes = btrfs_calc_insert_metadata_size(fs_info,
110 trans->delayed_ref_updates);
111 spin_lock(&delayed_rsv->lock);
112 delayed_rsv->size += num_bytes;
113 delayed_rsv->full = 0;
114 spin_unlock(&delayed_rsv->lock);
115 trans->delayed_ref_updates = 0;
116}
117
118/**
119 * Transfer bytes to our delayed refs rsv
120 *
121 * @fs_info: the filesystem
122 * @src: source block rsv to transfer from
123 * @num_bytes: number of bytes to transfer
124 *
125 * This transfers up to the num_bytes amount from the src rsv to the
126 * delayed_refs_rsv. Any extra bytes are returned to the space info.
127 */
128void btrfs_migrate_to_delayed_refs_rsv(struct btrfs_fs_info *fs_info,
129 struct btrfs_block_rsv *src,
130 u64 num_bytes)
131{
132 struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
133 u64 to_free = 0;
134
135 spin_lock(&src->lock);
136 src->reserved -= num_bytes;
137 src->size -= num_bytes;
138 spin_unlock(&src->lock);
139
140 spin_lock(&delayed_refs_rsv->lock);
141 if (delayed_refs_rsv->size > delayed_refs_rsv->reserved) {
142 u64 delta = delayed_refs_rsv->size -
143 delayed_refs_rsv->reserved;
144 if (num_bytes > delta) {
145 to_free = num_bytes - delta;
146 num_bytes = delta;
147 }
148 } else {
149 to_free = num_bytes;
150 num_bytes = 0;
151 }
152
153 if (num_bytes)
154 delayed_refs_rsv->reserved += num_bytes;
155 if (delayed_refs_rsv->reserved >= delayed_refs_rsv->size)
156 delayed_refs_rsv->full = 1;
157 spin_unlock(&delayed_refs_rsv->lock);
158
159 if (num_bytes)
160 trace_btrfs_space_reservation(fs_info, "delayed_refs_rsv",
161 0, num_bytes, 1);
162 if (to_free)
163 btrfs_space_info_free_bytes_may_use(fs_info,
164 delayed_refs_rsv->space_info, to_free);
165}
166
167/**
168 * Refill based on our delayed refs usage
169 *
170 * @fs_info: the filesystem
171 * @flush: control how we can flush for this reservation.
172 *
173 * This will refill the delayed block_rsv up to 1 items size worth of space and
174 * will return -ENOSPC if we can't make the reservation.
175 */
176int btrfs_delayed_refs_rsv_refill(struct btrfs_fs_info *fs_info,
177 enum btrfs_reserve_flush_enum flush)
178{
179 struct btrfs_block_rsv *block_rsv = &fs_info->delayed_refs_rsv;
180 u64 limit = btrfs_calc_insert_metadata_size(fs_info, 1);
181 u64 num_bytes = 0;
182 int ret = -ENOSPC;
183
184 spin_lock(&block_rsv->lock);
185 if (block_rsv->reserved < block_rsv->size) {
186 num_bytes = block_rsv->size - block_rsv->reserved;
187 num_bytes = min(num_bytes, limit);
188 }
189 spin_unlock(&block_rsv->lock);
190
191 if (!num_bytes)
192 return 0;
193
194 ret = btrfs_reserve_metadata_bytes(fs_info->extent_root, block_rsv,
195 num_bytes, flush);
196 if (ret)
197 return ret;
198 btrfs_block_rsv_add_bytes(block_rsv, num_bytes, 0);
199 trace_btrfs_space_reservation(fs_info, "delayed_refs_rsv",
200 0, num_bytes, 1);
201 return 0;
202}
203
204/*
205 * compare two delayed tree backrefs with same bytenr and type
206 */
207static int comp_tree_refs(struct btrfs_delayed_tree_ref *ref1,
208 struct btrfs_delayed_tree_ref *ref2)
209{
210 if (ref1->node.type == BTRFS_TREE_BLOCK_REF_KEY) {
211 if (ref1->root < ref2->root)
212 return -1;
213 if (ref1->root > ref2->root)
214 return 1;
215 } else {
216 if (ref1->parent < ref2->parent)
217 return -1;
218 if (ref1->parent > ref2->parent)
219 return 1;
220 }
221 return 0;
222}
223
224/*
225 * compare two delayed data backrefs with same bytenr and type
226 */
227static int comp_data_refs(struct btrfs_delayed_data_ref *ref1,
228 struct btrfs_delayed_data_ref *ref2)
229{
230 if (ref1->node.type == BTRFS_EXTENT_DATA_REF_KEY) {
231 if (ref1->root < ref2->root)
232 return -1;
233 if (ref1->root > ref2->root)
234 return 1;
235 if (ref1->objectid < ref2->objectid)
236 return -1;
237 if (ref1->objectid > ref2->objectid)
238 return 1;
239 if (ref1->offset < ref2->offset)
240 return -1;
241 if (ref1->offset > ref2->offset)
242 return 1;
243 } else {
244 if (ref1->parent < ref2->parent)
245 return -1;
246 if (ref1->parent > ref2->parent)
247 return 1;
248 }
249 return 0;
250}
251
252static int comp_refs(struct btrfs_delayed_ref_node *ref1,
253 struct btrfs_delayed_ref_node *ref2,
254 bool check_seq)
255{
256 int ret = 0;
257
258 if (ref1->type < ref2->type)
259 return -1;
260 if (ref1->type > ref2->type)
261 return 1;
262 if (ref1->type == BTRFS_TREE_BLOCK_REF_KEY ||
263 ref1->type == BTRFS_SHARED_BLOCK_REF_KEY)
264 ret = comp_tree_refs(btrfs_delayed_node_to_tree_ref(ref1),
265 btrfs_delayed_node_to_tree_ref(ref2));
266 else
267 ret = comp_data_refs(btrfs_delayed_node_to_data_ref(ref1),
268 btrfs_delayed_node_to_data_ref(ref2));
269 if (ret)
270 return ret;
271 if (check_seq) {
272 if (ref1->seq < ref2->seq)
273 return -1;
274 if (ref1->seq > ref2->seq)
275 return 1;
276 }
277 return 0;
278}
279
280/* insert a new ref to head ref rbtree */
281static struct btrfs_delayed_ref_head *htree_insert(struct rb_root_cached *root,
282 struct rb_node *node)
283{
284 struct rb_node **p = &root->rb_root.rb_node;
285 struct rb_node *parent_node = NULL;
286 struct btrfs_delayed_ref_head *entry;
287 struct btrfs_delayed_ref_head *ins;
288 u64 bytenr;
289 bool leftmost = true;
290
291 ins = rb_entry(node, struct btrfs_delayed_ref_head, href_node);
292 bytenr = ins->bytenr;
293 while (*p) {
294 parent_node = *p;
295 entry = rb_entry(parent_node, struct btrfs_delayed_ref_head,
296 href_node);
297
298 if (bytenr < entry->bytenr) {
299 p = &(*p)->rb_left;
300 } else if (bytenr > entry->bytenr) {
301 p = &(*p)->rb_right;
302 leftmost = false;
303 } else {
304 return entry;
305 }
306 }
307
308 rb_link_node(node, parent_node, p);
309 rb_insert_color_cached(node, root, leftmost);
310 return NULL;
311}
312
313static struct btrfs_delayed_ref_node* tree_insert(struct rb_root_cached *root,
314 struct btrfs_delayed_ref_node *ins)
315{
316 struct rb_node **p = &root->rb_root.rb_node;
317 struct rb_node *node = &ins->ref_node;
318 struct rb_node *parent_node = NULL;
319 struct btrfs_delayed_ref_node *entry;
320 bool leftmost = true;
321
322 while (*p) {
323 int comp;
324
325 parent_node = *p;
326 entry = rb_entry(parent_node, struct btrfs_delayed_ref_node,
327 ref_node);
328 comp = comp_refs(ins, entry, true);
329 if (comp < 0) {
330 p = &(*p)->rb_left;
331 } else if (comp > 0) {
332 p = &(*p)->rb_right;
333 leftmost = false;
334 } else {
335 return entry;
336 }
337 }
338
339 rb_link_node(node, parent_node, p);
340 rb_insert_color_cached(node, root, leftmost);
341 return NULL;
342}
343
344static struct btrfs_delayed_ref_head *find_first_ref_head(
345 struct btrfs_delayed_ref_root *dr)
346{
347 struct rb_node *n;
348 struct btrfs_delayed_ref_head *entry;
349
350 n = rb_first_cached(&dr->href_root);
351 if (!n)
352 return NULL;
353
354 entry = rb_entry(n, struct btrfs_delayed_ref_head, href_node);
355
356 return entry;
357}
358
359/*
360 * Find a head entry based on bytenr. This returns the delayed ref head if it
361 * was able to find one, or NULL if nothing was in that spot. If return_bigger
362 * is given, the next bigger entry is returned if no exact match is found.
363 */
364static struct btrfs_delayed_ref_head *find_ref_head(
365 struct btrfs_delayed_ref_root *dr, u64 bytenr,
366 bool return_bigger)
367{
368 struct rb_root *root = &dr->href_root.rb_root;
369 struct rb_node *n;
370 struct btrfs_delayed_ref_head *entry;
371
372 n = root->rb_node;
373 entry = NULL;
374 while (n) {
375 entry = rb_entry(n, struct btrfs_delayed_ref_head, href_node);
376
377 if (bytenr < entry->bytenr)
378 n = n->rb_left;
379 else if (bytenr > entry->bytenr)
380 n = n->rb_right;
381 else
382 return entry;
383 }
384 if (entry && return_bigger) {
385 if (bytenr > entry->bytenr) {
386 n = rb_next(&entry->href_node);
387 if (!n)
388 return NULL;
389 entry = rb_entry(n, struct btrfs_delayed_ref_head,
390 href_node);
391 }
392 return entry;
393 }
394 return NULL;
395}
396
397int btrfs_delayed_ref_lock(struct btrfs_delayed_ref_root *delayed_refs,
398 struct btrfs_delayed_ref_head *head)
399{
400 lockdep_assert_held(&delayed_refs->lock);
401 if (mutex_trylock(&head->mutex))
402 return 0;
403
404 refcount_inc(&head->refs);
405 spin_unlock(&delayed_refs->lock);
406
407 mutex_lock(&head->mutex);
408 spin_lock(&delayed_refs->lock);
409 if (RB_EMPTY_NODE(&head->href_node)) {
410 mutex_unlock(&head->mutex);
411 btrfs_put_delayed_ref_head(head);
412 return -EAGAIN;
413 }
414 btrfs_put_delayed_ref_head(head);
415 return 0;
416}
417
418static inline void drop_delayed_ref(struct btrfs_trans_handle *trans,
419 struct btrfs_delayed_ref_root *delayed_refs,
420 struct btrfs_delayed_ref_head *head,
421 struct btrfs_delayed_ref_node *ref)
422{
423 lockdep_assert_held(&head->lock);
424 rb_erase_cached(&ref->ref_node, &head->ref_tree);
425 RB_CLEAR_NODE(&ref->ref_node);
426 if (!list_empty(&ref->add_list))
427 list_del(&ref->add_list);
428 ref->in_tree = 0;
429 btrfs_put_delayed_ref(ref);
430 atomic_dec(&delayed_refs->num_entries);
431}
432
433static bool merge_ref(struct btrfs_trans_handle *trans,
434 struct btrfs_delayed_ref_root *delayed_refs,
435 struct btrfs_delayed_ref_head *head,
436 struct btrfs_delayed_ref_node *ref,
437 u64 seq)
438{
439 struct btrfs_delayed_ref_node *next;
440 struct rb_node *node = rb_next(&ref->ref_node);
441 bool done = false;
442
443 while (!done && node) {
444 int mod;
445
446 next = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
447 node = rb_next(node);
448 if (seq && next->seq >= seq)
449 break;
450 if (comp_refs(ref, next, false))
451 break;
452
453 if (ref->action == next->action) {
454 mod = next->ref_mod;
455 } else {
456 if (ref->ref_mod < next->ref_mod) {
457 swap(ref, next);
458 done = true;
459 }
460 mod = -next->ref_mod;
461 }
462
463 drop_delayed_ref(trans, delayed_refs, head, next);
464 ref->ref_mod += mod;
465 if (ref->ref_mod == 0) {
466 drop_delayed_ref(trans, delayed_refs, head, ref);
467 done = true;
468 } else {
469 /*
470 * Can't have multiples of the same ref on a tree block.
471 */
472 WARN_ON(ref->type == BTRFS_TREE_BLOCK_REF_KEY ||
473 ref->type == BTRFS_SHARED_BLOCK_REF_KEY);
474 }
475 }
476
477 return done;
478}
479
480void btrfs_merge_delayed_refs(struct btrfs_trans_handle *trans,
481 struct btrfs_delayed_ref_root *delayed_refs,
482 struct btrfs_delayed_ref_head *head)
483{
484 struct btrfs_fs_info *fs_info = trans->fs_info;
485 struct btrfs_delayed_ref_node *ref;
486 struct rb_node *node;
487 u64 seq = 0;
488
489 lockdep_assert_held(&head->lock);
490
491 if (RB_EMPTY_ROOT(&head->ref_tree.rb_root))
492 return;
493
494 /* We don't have too many refs to merge for data. */
495 if (head->is_data)
496 return;
497
498 seq = btrfs_tree_mod_log_lowest_seq(fs_info);
499again:
500 for (node = rb_first_cached(&head->ref_tree); node;
501 node = rb_next(node)) {
502 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
503 if (seq && ref->seq >= seq)
504 continue;
505 if (merge_ref(trans, delayed_refs, head, ref, seq))
506 goto again;
507 }
508}
509
510int btrfs_check_delayed_seq(struct btrfs_fs_info *fs_info, u64 seq)
511{
512 int ret = 0;
513 u64 min_seq = btrfs_tree_mod_log_lowest_seq(fs_info);
514
515 if (min_seq != 0 && seq >= min_seq) {
516 btrfs_debug(fs_info,
517 "holding back delayed_ref %llu, lowest is %llu",
518 seq, min_seq);
519 ret = 1;
520 }
521
522 return ret;
523}
524
525struct btrfs_delayed_ref_head *btrfs_select_ref_head(
526 struct btrfs_delayed_ref_root *delayed_refs)
527{
528 struct btrfs_delayed_ref_head *head;
529
530again:
531 head = find_ref_head(delayed_refs, delayed_refs->run_delayed_start,
532 true);
533 if (!head && delayed_refs->run_delayed_start != 0) {
534 delayed_refs->run_delayed_start = 0;
535 head = find_first_ref_head(delayed_refs);
536 }
537 if (!head)
538 return NULL;
539
540 while (head->processing) {
541 struct rb_node *node;
542
543 node = rb_next(&head->href_node);
544 if (!node) {
545 if (delayed_refs->run_delayed_start == 0)
546 return NULL;
547 delayed_refs->run_delayed_start = 0;
548 goto again;
549 }
550 head = rb_entry(node, struct btrfs_delayed_ref_head,
551 href_node);
552 }
553
554 head->processing = 1;
555 WARN_ON(delayed_refs->num_heads_ready == 0);
556 delayed_refs->num_heads_ready--;
557 delayed_refs->run_delayed_start = head->bytenr +
558 head->num_bytes;
559 return head;
560}
561
562void btrfs_delete_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
563 struct btrfs_delayed_ref_head *head)
564{
565 lockdep_assert_held(&delayed_refs->lock);
566 lockdep_assert_held(&head->lock);
567
568 rb_erase_cached(&head->href_node, &delayed_refs->href_root);
569 RB_CLEAR_NODE(&head->href_node);
570 atomic_dec(&delayed_refs->num_entries);
571 delayed_refs->num_heads--;
572 if (head->processing == 0)
573 delayed_refs->num_heads_ready--;
574}
575
576/*
577 * Helper to insert the ref_node to the tail or merge with tail.
578 *
579 * Return 0 for insert.
580 * Return >0 for merge.
581 */
582static int insert_delayed_ref(struct btrfs_trans_handle *trans,
583 struct btrfs_delayed_ref_root *root,
584 struct btrfs_delayed_ref_head *href,
585 struct btrfs_delayed_ref_node *ref)
586{
587 struct btrfs_delayed_ref_node *exist;
588 int mod;
589 int ret = 0;
590
591 spin_lock(&href->lock);
592 exist = tree_insert(&href->ref_tree, ref);
593 if (!exist)
594 goto inserted;
595
596 /* Now we are sure we can merge */
597 ret = 1;
598 if (exist->action == ref->action) {
599 mod = ref->ref_mod;
600 } else {
601 /* Need to change action */
602 if (exist->ref_mod < ref->ref_mod) {
603 exist->action = ref->action;
604 mod = -exist->ref_mod;
605 exist->ref_mod = ref->ref_mod;
606 if (ref->action == BTRFS_ADD_DELAYED_REF)
607 list_add_tail(&exist->add_list,
608 &href->ref_add_list);
609 else if (ref->action == BTRFS_DROP_DELAYED_REF) {
610 ASSERT(!list_empty(&exist->add_list));
611 list_del(&exist->add_list);
612 } else {
613 ASSERT(0);
614 }
615 } else
616 mod = -ref->ref_mod;
617 }
618 exist->ref_mod += mod;
619
620 /* remove existing tail if its ref_mod is zero */
621 if (exist->ref_mod == 0)
622 drop_delayed_ref(trans, root, href, exist);
623 spin_unlock(&href->lock);
624 return ret;
625inserted:
626 if (ref->action == BTRFS_ADD_DELAYED_REF)
627 list_add_tail(&ref->add_list, &href->ref_add_list);
628 atomic_inc(&root->num_entries);
629 spin_unlock(&href->lock);
630 return ret;
631}
632
633/*
634 * helper function to update the accounting in the head ref
635 * existing and update must have the same bytenr
636 */
637static noinline void update_existing_head_ref(struct btrfs_trans_handle *trans,
638 struct btrfs_delayed_ref_head *existing,
639 struct btrfs_delayed_ref_head *update)
640{
641 struct btrfs_delayed_ref_root *delayed_refs =
642 &trans->transaction->delayed_refs;
643 struct btrfs_fs_info *fs_info = trans->fs_info;
644 int old_ref_mod;
645
646 BUG_ON(existing->is_data != update->is_data);
647
648 spin_lock(&existing->lock);
649 if (update->must_insert_reserved) {
650 /* if the extent was freed and then
651 * reallocated before the delayed ref
652 * entries were processed, we can end up
653 * with an existing head ref without
654 * the must_insert_reserved flag set.
655 * Set it again here
656 */
657 existing->must_insert_reserved = update->must_insert_reserved;
658
659 /*
660 * update the num_bytes so we make sure the accounting
661 * is done correctly
662 */
663 existing->num_bytes = update->num_bytes;
664
665 }
666
667 if (update->extent_op) {
668 if (!existing->extent_op) {
669 existing->extent_op = update->extent_op;
670 } else {
671 if (update->extent_op->update_key) {
672 memcpy(&existing->extent_op->key,
673 &update->extent_op->key,
674 sizeof(update->extent_op->key));
675 existing->extent_op->update_key = true;
676 }
677 if (update->extent_op->update_flags) {
678 existing->extent_op->flags_to_set |=
679 update->extent_op->flags_to_set;
680 existing->extent_op->update_flags = true;
681 }
682 btrfs_free_delayed_extent_op(update->extent_op);
683 }
684 }
685 /*
686 * update the reference mod on the head to reflect this new operation,
687 * only need the lock for this case cause we could be processing it
688 * currently, for refs we just added we know we're a-ok.
689 */
690 old_ref_mod = existing->total_ref_mod;
691 existing->ref_mod += update->ref_mod;
692 existing->total_ref_mod += update->ref_mod;
693
694 /*
695 * If we are going to from a positive ref mod to a negative or vice
696 * versa we need to make sure to adjust pending_csums accordingly.
697 */
698 if (existing->is_data) {
699 u64 csum_leaves =
700 btrfs_csum_bytes_to_leaves(fs_info,
701 existing->num_bytes);
702
703 if (existing->total_ref_mod >= 0 && old_ref_mod < 0) {
704 delayed_refs->pending_csums -= existing->num_bytes;
705 btrfs_delayed_refs_rsv_release(fs_info, csum_leaves);
706 }
707 if (existing->total_ref_mod < 0 && old_ref_mod >= 0) {
708 delayed_refs->pending_csums += existing->num_bytes;
709 trans->delayed_ref_updates += csum_leaves;
710 }
711 }
712
713 spin_unlock(&existing->lock);
714}
715
716static void init_delayed_ref_head(struct btrfs_delayed_ref_head *head_ref,
717 struct btrfs_qgroup_extent_record *qrecord,
718 u64 bytenr, u64 num_bytes, u64 ref_root,
719 u64 reserved, int action, bool is_data,
720 bool is_system)
721{
722 int count_mod = 1;
723 int must_insert_reserved = 0;
724
725 /* If reserved is provided, it must be a data extent. */
726 BUG_ON(!is_data && reserved);
727
728 /*
729 * The head node stores the sum of all the mods, so dropping a ref
730 * should drop the sum in the head node by one.
731 */
732 if (action == BTRFS_UPDATE_DELAYED_HEAD)
733 count_mod = 0;
734 else if (action == BTRFS_DROP_DELAYED_REF)
735 count_mod = -1;
736
737 /*
738 * BTRFS_ADD_DELAYED_EXTENT means that we need to update the reserved
739 * accounting when the extent is finally added, or if a later
740 * modification deletes the delayed ref without ever inserting the
741 * extent into the extent allocation tree. ref->must_insert_reserved
742 * is the flag used to record that accounting mods are required.
743 *
744 * Once we record must_insert_reserved, switch the action to
745 * BTRFS_ADD_DELAYED_REF because other special casing is not required.
746 */
747 if (action == BTRFS_ADD_DELAYED_EXTENT)
748 must_insert_reserved = 1;
749 else
750 must_insert_reserved = 0;
751
752 refcount_set(&head_ref->refs, 1);
753 head_ref->bytenr = bytenr;
754 head_ref->num_bytes = num_bytes;
755 head_ref->ref_mod = count_mod;
756 head_ref->must_insert_reserved = must_insert_reserved;
757 head_ref->is_data = is_data;
758 head_ref->is_system = is_system;
759 head_ref->ref_tree = RB_ROOT_CACHED;
760 INIT_LIST_HEAD(&head_ref->ref_add_list);
761 RB_CLEAR_NODE(&head_ref->href_node);
762 head_ref->processing = 0;
763 head_ref->total_ref_mod = count_mod;
764 spin_lock_init(&head_ref->lock);
765 mutex_init(&head_ref->mutex);
766
767 if (qrecord) {
768 if (ref_root && reserved) {
769 qrecord->data_rsv = reserved;
770 qrecord->data_rsv_refroot = ref_root;
771 }
772 qrecord->bytenr = bytenr;
773 qrecord->num_bytes = num_bytes;
774 qrecord->old_roots = NULL;
775 }
776}
777
778/*
779 * helper function to actually insert a head node into the rbtree.
780 * this does all the dirty work in terms of maintaining the correct
781 * overall modification count.
782 */
783static noinline struct btrfs_delayed_ref_head *
784add_delayed_ref_head(struct btrfs_trans_handle *trans,
785 struct btrfs_delayed_ref_head *head_ref,
786 struct btrfs_qgroup_extent_record *qrecord,
787 int action, int *qrecord_inserted_ret)
788{
789 struct btrfs_delayed_ref_head *existing;
790 struct btrfs_delayed_ref_root *delayed_refs;
791 int qrecord_inserted = 0;
792
793 delayed_refs = &trans->transaction->delayed_refs;
794
795 /* Record qgroup extent info if provided */
796 if (qrecord) {
797 if (btrfs_qgroup_trace_extent_nolock(trans->fs_info,
798 delayed_refs, qrecord))
799 kfree(qrecord);
800 else
801 qrecord_inserted = 1;
802 }
803
804 trace_add_delayed_ref_head(trans->fs_info, head_ref, action);
805
806 existing = htree_insert(&delayed_refs->href_root,
807 &head_ref->href_node);
808 if (existing) {
809 update_existing_head_ref(trans, existing, head_ref);
810 /*
811 * we've updated the existing ref, free the newly
812 * allocated ref
813 */
814 kmem_cache_free(btrfs_delayed_ref_head_cachep, head_ref);
815 head_ref = existing;
816 } else {
817 if (head_ref->is_data && head_ref->ref_mod < 0) {
818 delayed_refs->pending_csums += head_ref->num_bytes;
819 trans->delayed_ref_updates +=
820 btrfs_csum_bytes_to_leaves(trans->fs_info,
821 head_ref->num_bytes);
822 }
823 delayed_refs->num_heads++;
824 delayed_refs->num_heads_ready++;
825 atomic_inc(&delayed_refs->num_entries);
826 trans->delayed_ref_updates++;
827 }
828 if (qrecord_inserted_ret)
829 *qrecord_inserted_ret = qrecord_inserted;
830
831 return head_ref;
832}
833
834/*
835 * init_delayed_ref_common - Initialize the structure which represents a
836 * modification to a an extent.
837 *
838 * @fs_info: Internal to the mounted filesystem mount structure.
839 *
840 * @ref: The structure which is going to be initialized.
841 *
842 * @bytenr: The logical address of the extent for which a modification is
843 * going to be recorded.
844 *
845 * @num_bytes: Size of the extent whose modification is being recorded.
846 *
847 * @ref_root: The id of the root where this modification has originated, this
848 * can be either one of the well-known metadata trees or the
849 * subvolume id which references this extent.
850 *
851 * @action: Can be one of BTRFS_ADD_DELAYED_REF/BTRFS_DROP_DELAYED_REF or
852 * BTRFS_ADD_DELAYED_EXTENT
853 *
854 * @ref_type: Holds the type of the extent which is being recorded, can be
855 * one of BTRFS_SHARED_BLOCK_REF_KEY/BTRFS_TREE_BLOCK_REF_KEY
856 * when recording a metadata extent or BTRFS_SHARED_DATA_REF_KEY/
857 * BTRFS_EXTENT_DATA_REF_KEY when recording data extent
858 */
859static void init_delayed_ref_common(struct btrfs_fs_info *fs_info,
860 struct btrfs_delayed_ref_node *ref,
861 u64 bytenr, u64 num_bytes, u64 ref_root,
862 int action, u8 ref_type)
863{
864 u64 seq = 0;
865
866 if (action == BTRFS_ADD_DELAYED_EXTENT)
867 action = BTRFS_ADD_DELAYED_REF;
868
869 if (is_fstree(ref_root))
870 seq = atomic64_read(&fs_info->tree_mod_seq);
871
872 refcount_set(&ref->refs, 1);
873 ref->bytenr = bytenr;
874 ref->num_bytes = num_bytes;
875 ref->ref_mod = 1;
876 ref->action = action;
877 ref->is_head = 0;
878 ref->in_tree = 1;
879 ref->seq = seq;
880 ref->type = ref_type;
881 RB_CLEAR_NODE(&ref->ref_node);
882 INIT_LIST_HEAD(&ref->add_list);
883}
884
885/*
886 * add a delayed tree ref. This does all of the accounting required
887 * to make sure the delayed ref is eventually processed before this
888 * transaction commits.
889 */
890int btrfs_add_delayed_tree_ref(struct btrfs_trans_handle *trans,
891 struct btrfs_ref *generic_ref,
892 struct btrfs_delayed_extent_op *extent_op)
893{
894 struct btrfs_fs_info *fs_info = trans->fs_info;
895 struct btrfs_delayed_tree_ref *ref;
896 struct btrfs_delayed_ref_head *head_ref;
897 struct btrfs_delayed_ref_root *delayed_refs;
898 struct btrfs_qgroup_extent_record *record = NULL;
899 int qrecord_inserted;
900 bool is_system;
901 int action = generic_ref->action;
902 int level = generic_ref->tree_ref.level;
903 int ret;
904 u64 bytenr = generic_ref->bytenr;
905 u64 num_bytes = generic_ref->len;
906 u64 parent = generic_ref->parent;
907 u8 ref_type;
908
909 is_system = (generic_ref->real_root == BTRFS_CHUNK_TREE_OBJECTID);
910
911 ASSERT(generic_ref->type == BTRFS_REF_METADATA && generic_ref->action);
912 BUG_ON(extent_op && extent_op->is_data);
913 ref = kmem_cache_alloc(btrfs_delayed_tree_ref_cachep, GFP_NOFS);
914 if (!ref)
915 return -ENOMEM;
916
917 head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
918 if (!head_ref) {
919 kmem_cache_free(btrfs_delayed_tree_ref_cachep, ref);
920 return -ENOMEM;
921 }
922
923 if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) &&
924 is_fstree(generic_ref->real_root) &&
925 is_fstree(generic_ref->tree_ref.root) &&
926 !generic_ref->skip_qgroup) {
927 record = kzalloc(sizeof(*record), GFP_NOFS);
928 if (!record) {
929 kmem_cache_free(btrfs_delayed_tree_ref_cachep, ref);
930 kmem_cache_free(btrfs_delayed_ref_head_cachep, head_ref);
931 return -ENOMEM;
932 }
933 }
934
935 if (parent)
936 ref_type = BTRFS_SHARED_BLOCK_REF_KEY;
937 else
938 ref_type = BTRFS_TREE_BLOCK_REF_KEY;
939
940 init_delayed_ref_common(fs_info, &ref->node, bytenr, num_bytes,
941 generic_ref->tree_ref.root, action, ref_type);
942 ref->root = generic_ref->tree_ref.root;
943 ref->parent = parent;
944 ref->level = level;
945
946 init_delayed_ref_head(head_ref, record, bytenr, num_bytes,
947 generic_ref->tree_ref.root, 0, action, false,
948 is_system);
949 head_ref->extent_op = extent_op;
950
951 delayed_refs = &trans->transaction->delayed_refs;
952 spin_lock(&delayed_refs->lock);
953
954 /*
955 * insert both the head node and the new ref without dropping
956 * the spin lock
957 */
958 head_ref = add_delayed_ref_head(trans, head_ref, record,
959 action, &qrecord_inserted);
960
961 ret = insert_delayed_ref(trans, delayed_refs, head_ref, &ref->node);
962 spin_unlock(&delayed_refs->lock);
963
964 /*
965 * Need to update the delayed_refs_rsv with any changes we may have
966 * made.
967 */
968 btrfs_update_delayed_refs_rsv(trans);
969
970 trace_add_delayed_tree_ref(fs_info, &ref->node, ref,
971 action == BTRFS_ADD_DELAYED_EXTENT ?
972 BTRFS_ADD_DELAYED_REF : action);
973 if (ret > 0)
974 kmem_cache_free(btrfs_delayed_tree_ref_cachep, ref);
975
976 if (qrecord_inserted)
977 btrfs_qgroup_trace_extent_post(trans, record);
978
979 return 0;
980}
981
982/*
983 * add a delayed data ref. it's similar to btrfs_add_delayed_tree_ref.
984 */
985int btrfs_add_delayed_data_ref(struct btrfs_trans_handle *trans,
986 struct btrfs_ref *generic_ref,
987 u64 reserved)
988{
989 struct btrfs_fs_info *fs_info = trans->fs_info;
990 struct btrfs_delayed_data_ref *ref;
991 struct btrfs_delayed_ref_head *head_ref;
992 struct btrfs_delayed_ref_root *delayed_refs;
993 struct btrfs_qgroup_extent_record *record = NULL;
994 int qrecord_inserted;
995 int action = generic_ref->action;
996 int ret;
997 u64 bytenr = generic_ref->bytenr;
998 u64 num_bytes = generic_ref->len;
999 u64 parent = generic_ref->parent;
1000 u64 ref_root = generic_ref->data_ref.ref_root;
1001 u64 owner = generic_ref->data_ref.ino;
1002 u64 offset = generic_ref->data_ref.offset;
1003 u8 ref_type;
1004
1005 ASSERT(generic_ref->type == BTRFS_REF_DATA && action);
1006 ref = kmem_cache_alloc(btrfs_delayed_data_ref_cachep, GFP_NOFS);
1007 if (!ref)
1008 return -ENOMEM;
1009
1010 if (parent)
1011 ref_type = BTRFS_SHARED_DATA_REF_KEY;
1012 else
1013 ref_type = BTRFS_EXTENT_DATA_REF_KEY;
1014 init_delayed_ref_common(fs_info, &ref->node, bytenr, num_bytes,
1015 ref_root, action, ref_type);
1016 ref->root = ref_root;
1017 ref->parent = parent;
1018 ref->objectid = owner;
1019 ref->offset = offset;
1020
1021
1022 head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
1023 if (!head_ref) {
1024 kmem_cache_free(btrfs_delayed_data_ref_cachep, ref);
1025 return -ENOMEM;
1026 }
1027
1028 if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) &&
1029 is_fstree(ref_root) &&
1030 is_fstree(generic_ref->real_root) &&
1031 !generic_ref->skip_qgroup) {
1032 record = kzalloc(sizeof(*record), GFP_NOFS);
1033 if (!record) {
1034 kmem_cache_free(btrfs_delayed_data_ref_cachep, ref);
1035 kmem_cache_free(btrfs_delayed_ref_head_cachep,
1036 head_ref);
1037 return -ENOMEM;
1038 }
1039 }
1040
1041 init_delayed_ref_head(head_ref, record, bytenr, num_bytes, ref_root,
1042 reserved, action, true, false);
1043 head_ref->extent_op = NULL;
1044
1045 delayed_refs = &trans->transaction->delayed_refs;
1046 spin_lock(&delayed_refs->lock);
1047
1048 /*
1049 * insert both the head node and the new ref without dropping
1050 * the spin lock
1051 */
1052 head_ref = add_delayed_ref_head(trans, head_ref, record,
1053 action, &qrecord_inserted);
1054
1055 ret = insert_delayed_ref(trans, delayed_refs, head_ref, &ref->node);
1056 spin_unlock(&delayed_refs->lock);
1057
1058 /*
1059 * Need to update the delayed_refs_rsv with any changes we may have
1060 * made.
1061 */
1062 btrfs_update_delayed_refs_rsv(trans);
1063
1064 trace_add_delayed_data_ref(trans->fs_info, &ref->node, ref,
1065 action == BTRFS_ADD_DELAYED_EXTENT ?
1066 BTRFS_ADD_DELAYED_REF : action);
1067 if (ret > 0)
1068 kmem_cache_free(btrfs_delayed_data_ref_cachep, ref);
1069
1070
1071 if (qrecord_inserted)
1072 return btrfs_qgroup_trace_extent_post(trans, record);
1073 return 0;
1074}
1075
1076int btrfs_add_delayed_extent_op(struct btrfs_trans_handle *trans,
1077 u64 bytenr, u64 num_bytes,
1078 struct btrfs_delayed_extent_op *extent_op)
1079{
1080 struct btrfs_delayed_ref_head *head_ref;
1081 struct btrfs_delayed_ref_root *delayed_refs;
1082
1083 head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
1084 if (!head_ref)
1085 return -ENOMEM;
1086
1087 init_delayed_ref_head(head_ref, NULL, bytenr, num_bytes, 0, 0,
1088 BTRFS_UPDATE_DELAYED_HEAD, extent_op->is_data,
1089 false);
1090 head_ref->extent_op = extent_op;
1091
1092 delayed_refs = &trans->transaction->delayed_refs;
1093 spin_lock(&delayed_refs->lock);
1094
1095 add_delayed_ref_head(trans, head_ref, NULL, BTRFS_UPDATE_DELAYED_HEAD,
1096 NULL);
1097
1098 spin_unlock(&delayed_refs->lock);
1099
1100 /*
1101 * Need to update the delayed_refs_rsv with any changes we may have
1102 * made.
1103 */
1104 btrfs_update_delayed_refs_rsv(trans);
1105 return 0;
1106}
1107
1108/*
1109 * This does a simple search for the head node for a given extent. Returns the
1110 * head node if found, or NULL if not.
1111 */
1112struct btrfs_delayed_ref_head *
1113btrfs_find_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs, u64 bytenr)
1114{
1115 lockdep_assert_held(&delayed_refs->lock);
1116
1117 return find_ref_head(delayed_refs, bytenr, false);
1118}
1119
1120void __cold btrfs_delayed_ref_exit(void)
1121{
1122 kmem_cache_destroy(btrfs_delayed_ref_head_cachep);
1123 kmem_cache_destroy(btrfs_delayed_tree_ref_cachep);
1124 kmem_cache_destroy(btrfs_delayed_data_ref_cachep);
1125 kmem_cache_destroy(btrfs_delayed_extent_op_cachep);
1126}
1127
1128int __init btrfs_delayed_ref_init(void)
1129{
1130 btrfs_delayed_ref_head_cachep = kmem_cache_create(
1131 "btrfs_delayed_ref_head",
1132 sizeof(struct btrfs_delayed_ref_head), 0,
1133 SLAB_MEM_SPREAD, NULL);
1134 if (!btrfs_delayed_ref_head_cachep)
1135 goto fail;
1136
1137 btrfs_delayed_tree_ref_cachep = kmem_cache_create(
1138 "btrfs_delayed_tree_ref",
1139 sizeof(struct btrfs_delayed_tree_ref), 0,
1140 SLAB_MEM_SPREAD, NULL);
1141 if (!btrfs_delayed_tree_ref_cachep)
1142 goto fail;
1143
1144 btrfs_delayed_data_ref_cachep = kmem_cache_create(
1145 "btrfs_delayed_data_ref",
1146 sizeof(struct btrfs_delayed_data_ref), 0,
1147 SLAB_MEM_SPREAD, NULL);
1148 if (!btrfs_delayed_data_ref_cachep)
1149 goto fail;
1150
1151 btrfs_delayed_extent_op_cachep = kmem_cache_create(
1152 "btrfs_delayed_extent_op",
1153 sizeof(struct btrfs_delayed_extent_op), 0,
1154 SLAB_MEM_SPREAD, NULL);
1155 if (!btrfs_delayed_extent_op_cachep)
1156 goto fail;
1157
1158 return 0;
1159fail:
1160 btrfs_delayed_ref_exit();
1161 return -ENOMEM;
1162}