Loading...
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 "messages.h"
10#include "ctree.h"
11#include "delayed-ref.h"
12#include "extent-tree.h"
13#include "transaction.h"
14#include "qgroup.h"
15#include "space-info.h"
16#include "tree-mod-log.h"
17#include "fs.h"
18
19struct kmem_cache *btrfs_delayed_ref_head_cachep;
20struct kmem_cache *btrfs_delayed_ref_node_cachep;
21struct kmem_cache *btrfs_delayed_extent_op_cachep;
22/*
23 * delayed back reference update tracking. For subvolume trees
24 * we queue up extent allocations and backref maintenance for
25 * delayed processing. This avoids deep call chains where we
26 * add extents in the middle of btrfs_search_slot, and it allows
27 * us to buffer up frequently modified backrefs in an rb tree instead
28 * of hammering updates on the extent allocation tree.
29 */
30
31bool btrfs_check_space_for_delayed_refs(struct btrfs_fs_info *fs_info)
32{
33 struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
34 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
35 bool ret = false;
36 u64 reserved;
37
38 spin_lock(&global_rsv->lock);
39 reserved = global_rsv->reserved;
40 spin_unlock(&global_rsv->lock);
41
42 /*
43 * Since the global reserve is just kind of magic we don't really want
44 * to rely on it to save our bacon, so if our size is more than the
45 * delayed_refs_rsv and the global rsv then it's time to think about
46 * bailing.
47 */
48 spin_lock(&delayed_refs_rsv->lock);
49 reserved += delayed_refs_rsv->reserved;
50 if (delayed_refs_rsv->size >= reserved)
51 ret = true;
52 spin_unlock(&delayed_refs_rsv->lock);
53 return ret;
54}
55
56/*
57 * Release a ref head's reservation.
58 *
59 * @fs_info: the filesystem
60 * @nr_refs: number of delayed refs to drop
61 * @nr_csums: number of csum items to drop
62 *
63 * Drops the delayed ref head's count from the delayed refs rsv and free any
64 * excess reservation we had.
65 */
66void btrfs_delayed_refs_rsv_release(struct btrfs_fs_info *fs_info, int nr_refs, int nr_csums)
67{
68 struct btrfs_block_rsv *block_rsv = &fs_info->delayed_refs_rsv;
69 u64 num_bytes;
70 u64 released;
71
72 num_bytes = btrfs_calc_delayed_ref_bytes(fs_info, nr_refs);
73 num_bytes += btrfs_calc_delayed_ref_csum_bytes(fs_info, nr_csums);
74
75 released = btrfs_block_rsv_release(fs_info, block_rsv, num_bytes, NULL);
76 if (released)
77 trace_btrfs_space_reservation(fs_info, "delayed_refs_rsv",
78 0, released, 0);
79}
80
81/*
82 * Adjust the size of the delayed refs rsv.
83 *
84 * This is to be called anytime we may have adjusted trans->delayed_ref_updates
85 * or trans->delayed_ref_csum_deletions, it'll calculate the additional size and
86 * add it to the delayed_refs_rsv.
87 */
88void btrfs_update_delayed_refs_rsv(struct btrfs_trans_handle *trans)
89{
90 struct btrfs_fs_info *fs_info = trans->fs_info;
91 struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_refs_rsv;
92 struct btrfs_block_rsv *local_rsv = &trans->delayed_rsv;
93 u64 num_bytes;
94 u64 reserved_bytes;
95
96 num_bytes = btrfs_calc_delayed_ref_bytes(fs_info, trans->delayed_ref_updates);
97 num_bytes += btrfs_calc_delayed_ref_csum_bytes(fs_info,
98 trans->delayed_ref_csum_deletions);
99
100 if (num_bytes == 0)
101 return;
102
103 /*
104 * Try to take num_bytes from the transaction's local delayed reserve.
105 * If not possible, try to take as much as it's available. If the local
106 * reserve doesn't have enough reserved space, the delayed refs reserve
107 * will be refilled next time btrfs_delayed_refs_rsv_refill() is called
108 * by someone or if a transaction commit is triggered before that, the
109 * global block reserve will be used. We want to minimize using the
110 * global block reserve for cases we can account for in advance, to
111 * avoid exhausting it and reach -ENOSPC during a transaction commit.
112 */
113 spin_lock(&local_rsv->lock);
114 reserved_bytes = min(num_bytes, local_rsv->reserved);
115 local_rsv->reserved -= reserved_bytes;
116 local_rsv->full = (local_rsv->reserved >= local_rsv->size);
117 spin_unlock(&local_rsv->lock);
118
119 spin_lock(&delayed_rsv->lock);
120 delayed_rsv->size += num_bytes;
121 delayed_rsv->reserved += reserved_bytes;
122 delayed_rsv->full = (delayed_rsv->reserved >= delayed_rsv->size);
123 spin_unlock(&delayed_rsv->lock);
124 trans->delayed_ref_updates = 0;
125 trans->delayed_ref_csum_deletions = 0;
126}
127
128/*
129 * Adjust the size of the delayed refs block reserve for 1 block group item
130 * insertion, used after allocating a block group.
131 */
132void btrfs_inc_delayed_refs_rsv_bg_inserts(struct btrfs_fs_info *fs_info)
133{
134 struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_refs_rsv;
135
136 spin_lock(&delayed_rsv->lock);
137 /*
138 * Inserting a block group item does not require changing the free space
139 * tree, only the extent tree or the block group tree, so this is all we
140 * need.
141 */
142 delayed_rsv->size += btrfs_calc_insert_metadata_size(fs_info, 1);
143 delayed_rsv->full = false;
144 spin_unlock(&delayed_rsv->lock);
145}
146
147/*
148 * Adjust the size of the delayed refs block reserve to release space for 1
149 * block group item insertion.
150 */
151void btrfs_dec_delayed_refs_rsv_bg_inserts(struct btrfs_fs_info *fs_info)
152{
153 struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_refs_rsv;
154 const u64 num_bytes = btrfs_calc_insert_metadata_size(fs_info, 1);
155 u64 released;
156
157 released = btrfs_block_rsv_release(fs_info, delayed_rsv, num_bytes, NULL);
158 if (released > 0)
159 trace_btrfs_space_reservation(fs_info, "delayed_refs_rsv",
160 0, released, 0);
161}
162
163/*
164 * Adjust the size of the delayed refs block reserve for 1 block group item
165 * update.
166 */
167void btrfs_inc_delayed_refs_rsv_bg_updates(struct btrfs_fs_info *fs_info)
168{
169 struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_refs_rsv;
170
171 spin_lock(&delayed_rsv->lock);
172 /*
173 * Updating a block group item does not result in new nodes/leaves and
174 * does not require changing the free space tree, only the extent tree
175 * or the block group tree, so this is all we need.
176 */
177 delayed_rsv->size += btrfs_calc_metadata_size(fs_info, 1);
178 delayed_rsv->full = false;
179 spin_unlock(&delayed_rsv->lock);
180}
181
182/*
183 * Adjust the size of the delayed refs block reserve to release space for 1
184 * block group item update.
185 */
186void btrfs_dec_delayed_refs_rsv_bg_updates(struct btrfs_fs_info *fs_info)
187{
188 struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_refs_rsv;
189 const u64 num_bytes = btrfs_calc_metadata_size(fs_info, 1);
190 u64 released;
191
192 released = btrfs_block_rsv_release(fs_info, delayed_rsv, num_bytes, NULL);
193 if (released > 0)
194 trace_btrfs_space_reservation(fs_info, "delayed_refs_rsv",
195 0, released, 0);
196}
197
198/*
199 * Refill based on our delayed refs usage.
200 *
201 * @fs_info: the filesystem
202 * @flush: control how we can flush for this reservation.
203 *
204 * This will refill the delayed block_rsv up to 1 items size worth of space and
205 * will return -ENOSPC if we can't make the reservation.
206 */
207int btrfs_delayed_refs_rsv_refill(struct btrfs_fs_info *fs_info,
208 enum btrfs_reserve_flush_enum flush)
209{
210 struct btrfs_block_rsv *block_rsv = &fs_info->delayed_refs_rsv;
211 struct btrfs_space_info *space_info = block_rsv->space_info;
212 u64 limit = btrfs_calc_delayed_ref_bytes(fs_info, 1);
213 u64 num_bytes = 0;
214 u64 refilled_bytes;
215 u64 to_free;
216 int ret = -ENOSPC;
217
218 spin_lock(&block_rsv->lock);
219 if (block_rsv->reserved < block_rsv->size) {
220 num_bytes = block_rsv->size - block_rsv->reserved;
221 num_bytes = min(num_bytes, limit);
222 }
223 spin_unlock(&block_rsv->lock);
224
225 if (!num_bytes)
226 return 0;
227
228 ret = btrfs_reserve_metadata_bytes(fs_info, space_info, num_bytes, flush);
229 if (ret)
230 return ret;
231
232 /*
233 * We may have raced with someone else, so check again if we the block
234 * reserve is still not full and release any excess space.
235 */
236 spin_lock(&block_rsv->lock);
237 if (block_rsv->reserved < block_rsv->size) {
238 u64 needed = block_rsv->size - block_rsv->reserved;
239
240 if (num_bytes >= needed) {
241 block_rsv->reserved += needed;
242 block_rsv->full = true;
243 to_free = num_bytes - needed;
244 refilled_bytes = needed;
245 } else {
246 block_rsv->reserved += num_bytes;
247 to_free = 0;
248 refilled_bytes = num_bytes;
249 }
250 } else {
251 to_free = num_bytes;
252 refilled_bytes = 0;
253 }
254 spin_unlock(&block_rsv->lock);
255
256 if (to_free > 0)
257 btrfs_space_info_free_bytes_may_use(fs_info, space_info, to_free);
258
259 if (refilled_bytes > 0)
260 trace_btrfs_space_reservation(fs_info, "delayed_refs_rsv", 0,
261 refilled_bytes, 1);
262 return 0;
263}
264
265/*
266 * compare two delayed data backrefs with same bytenr and type
267 */
268static int comp_data_refs(struct btrfs_delayed_ref_node *ref1,
269 struct btrfs_delayed_ref_node *ref2)
270{
271 if (ref1->data_ref.objectid < ref2->data_ref.objectid)
272 return -1;
273 if (ref1->data_ref.objectid > ref2->data_ref.objectid)
274 return 1;
275 if (ref1->data_ref.offset < ref2->data_ref.offset)
276 return -1;
277 if (ref1->data_ref.offset > ref2->data_ref.offset)
278 return 1;
279 return 0;
280}
281
282static int comp_refs(struct btrfs_delayed_ref_node *ref1,
283 struct btrfs_delayed_ref_node *ref2,
284 bool check_seq)
285{
286 int ret = 0;
287
288 if (ref1->type < ref2->type)
289 return -1;
290 if (ref1->type > ref2->type)
291 return 1;
292 if (ref1->type == BTRFS_SHARED_BLOCK_REF_KEY ||
293 ref1->type == BTRFS_SHARED_DATA_REF_KEY) {
294 if (ref1->parent < ref2->parent)
295 return -1;
296 if (ref1->parent > ref2->parent)
297 return 1;
298 } else {
299 if (ref1->ref_root < ref2->ref_root)
300 return -1;
301 if (ref1->ref_root > ref2->ref_root)
302 return 1;
303 if (ref1->type == BTRFS_EXTENT_DATA_REF_KEY)
304 ret = comp_data_refs(ref1, ref2);
305 }
306 if (ret)
307 return ret;
308 if (check_seq) {
309 if (ref1->seq < ref2->seq)
310 return -1;
311 if (ref1->seq > ref2->seq)
312 return 1;
313 }
314 return 0;
315}
316
317static struct btrfs_delayed_ref_node* tree_insert(struct rb_root_cached *root,
318 struct btrfs_delayed_ref_node *ins)
319{
320 struct rb_node **p = &root->rb_root.rb_node;
321 struct rb_node *node = &ins->ref_node;
322 struct rb_node *parent_node = NULL;
323 struct btrfs_delayed_ref_node *entry;
324 bool leftmost = true;
325
326 while (*p) {
327 int comp;
328
329 parent_node = *p;
330 entry = rb_entry(parent_node, struct btrfs_delayed_ref_node,
331 ref_node);
332 comp = comp_refs(ins, entry, true);
333 if (comp < 0) {
334 p = &(*p)->rb_left;
335 } else if (comp > 0) {
336 p = &(*p)->rb_right;
337 leftmost = false;
338 } else {
339 return entry;
340 }
341 }
342
343 rb_link_node(node, parent_node, p);
344 rb_insert_color_cached(node, root, leftmost);
345 return NULL;
346}
347
348static struct btrfs_delayed_ref_head *find_first_ref_head(
349 struct btrfs_delayed_ref_root *dr)
350{
351 unsigned long from = 0;
352
353 lockdep_assert_held(&dr->lock);
354
355 return xa_find(&dr->head_refs, &from, ULONG_MAX, XA_PRESENT);
356}
357
358static bool btrfs_delayed_ref_lock(struct btrfs_delayed_ref_root *delayed_refs,
359 struct btrfs_delayed_ref_head *head)
360{
361 lockdep_assert_held(&delayed_refs->lock);
362 if (mutex_trylock(&head->mutex))
363 return true;
364
365 refcount_inc(&head->refs);
366 spin_unlock(&delayed_refs->lock);
367
368 mutex_lock(&head->mutex);
369 spin_lock(&delayed_refs->lock);
370 if (!head->tracked) {
371 mutex_unlock(&head->mutex);
372 btrfs_put_delayed_ref_head(head);
373 return false;
374 }
375 btrfs_put_delayed_ref_head(head);
376 return true;
377}
378
379static inline void drop_delayed_ref(struct btrfs_fs_info *fs_info,
380 struct btrfs_delayed_ref_root *delayed_refs,
381 struct btrfs_delayed_ref_head *head,
382 struct btrfs_delayed_ref_node *ref)
383{
384 lockdep_assert_held(&head->lock);
385 rb_erase_cached(&ref->ref_node, &head->ref_tree);
386 RB_CLEAR_NODE(&ref->ref_node);
387 if (!list_empty(&ref->add_list))
388 list_del(&ref->add_list);
389 btrfs_put_delayed_ref(ref);
390 btrfs_delayed_refs_rsv_release(fs_info, 1, 0);
391}
392
393static bool merge_ref(struct btrfs_fs_info *fs_info,
394 struct btrfs_delayed_ref_root *delayed_refs,
395 struct btrfs_delayed_ref_head *head,
396 struct btrfs_delayed_ref_node *ref,
397 u64 seq)
398{
399 struct btrfs_delayed_ref_node *next;
400 struct rb_node *node = rb_next(&ref->ref_node);
401 bool done = false;
402
403 while (!done && node) {
404 int mod;
405
406 next = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
407 node = rb_next(node);
408 if (seq && next->seq >= seq)
409 break;
410 if (comp_refs(ref, next, false))
411 break;
412
413 if (ref->action == next->action) {
414 mod = next->ref_mod;
415 } else {
416 if (ref->ref_mod < next->ref_mod) {
417 swap(ref, next);
418 done = true;
419 }
420 mod = -next->ref_mod;
421 }
422
423 drop_delayed_ref(fs_info, delayed_refs, head, next);
424 ref->ref_mod += mod;
425 if (ref->ref_mod == 0) {
426 drop_delayed_ref(fs_info, delayed_refs, head, ref);
427 done = true;
428 } else {
429 /*
430 * Can't have multiples of the same ref on a tree block.
431 */
432 WARN_ON(ref->type == BTRFS_TREE_BLOCK_REF_KEY ||
433 ref->type == BTRFS_SHARED_BLOCK_REF_KEY);
434 }
435 }
436
437 return done;
438}
439
440void btrfs_merge_delayed_refs(struct btrfs_fs_info *fs_info,
441 struct btrfs_delayed_ref_root *delayed_refs,
442 struct btrfs_delayed_ref_head *head)
443{
444 struct btrfs_delayed_ref_node *ref;
445 struct rb_node *node;
446 u64 seq = 0;
447
448 lockdep_assert_held(&head->lock);
449
450 if (RB_EMPTY_ROOT(&head->ref_tree.rb_root))
451 return;
452
453 /* We don't have too many refs to merge for data. */
454 if (head->is_data)
455 return;
456
457 seq = btrfs_tree_mod_log_lowest_seq(fs_info);
458again:
459 for (node = rb_first_cached(&head->ref_tree); node;
460 node = rb_next(node)) {
461 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
462 if (seq && ref->seq >= seq)
463 continue;
464 if (merge_ref(fs_info, delayed_refs, head, ref, seq))
465 goto again;
466 }
467}
468
469int btrfs_check_delayed_seq(struct btrfs_fs_info *fs_info, u64 seq)
470{
471 int ret = 0;
472 u64 min_seq = btrfs_tree_mod_log_lowest_seq(fs_info);
473
474 if (min_seq != 0 && seq >= min_seq) {
475 btrfs_debug(fs_info,
476 "holding back delayed_ref %llu, lowest is %llu",
477 seq, min_seq);
478 ret = 1;
479 }
480
481 return ret;
482}
483
484struct btrfs_delayed_ref_head *btrfs_select_ref_head(
485 const struct btrfs_fs_info *fs_info,
486 struct btrfs_delayed_ref_root *delayed_refs)
487{
488 struct btrfs_delayed_ref_head *head;
489 unsigned long start_index;
490 unsigned long found_index;
491 bool found_head = false;
492 bool locked;
493
494 spin_lock(&delayed_refs->lock);
495again:
496 start_index = (delayed_refs->run_delayed_start >> fs_info->sectorsize_bits);
497 xa_for_each_start(&delayed_refs->head_refs, found_index, head, start_index) {
498 if (!head->processing) {
499 found_head = true;
500 break;
501 }
502 }
503 if (!found_head) {
504 if (delayed_refs->run_delayed_start == 0) {
505 spin_unlock(&delayed_refs->lock);
506 return NULL;
507 }
508 delayed_refs->run_delayed_start = 0;
509 goto again;
510 }
511
512 head->processing = true;
513 WARN_ON(delayed_refs->num_heads_ready == 0);
514 delayed_refs->num_heads_ready--;
515 delayed_refs->run_delayed_start = head->bytenr +
516 head->num_bytes;
517
518 locked = btrfs_delayed_ref_lock(delayed_refs, head);
519 spin_unlock(&delayed_refs->lock);
520
521 /*
522 * We may have dropped the spin lock to get the head mutex lock, and
523 * that might have given someone else time to free the head. If that's
524 * true, it has been removed from our list and we can move on.
525 */
526 if (!locked)
527 return ERR_PTR(-EAGAIN);
528
529 return head;
530}
531
532void btrfs_unselect_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
533 struct btrfs_delayed_ref_head *head)
534{
535 spin_lock(&delayed_refs->lock);
536 head->processing = false;
537 delayed_refs->num_heads_ready++;
538 spin_unlock(&delayed_refs->lock);
539 btrfs_delayed_ref_unlock(head);
540}
541
542void btrfs_delete_ref_head(const struct btrfs_fs_info *fs_info,
543 struct btrfs_delayed_ref_root *delayed_refs,
544 struct btrfs_delayed_ref_head *head)
545{
546 const unsigned long index = (head->bytenr >> fs_info->sectorsize_bits);
547
548 lockdep_assert_held(&delayed_refs->lock);
549 lockdep_assert_held(&head->lock);
550
551 xa_erase(&delayed_refs->head_refs, index);
552 head->tracked = false;
553 delayed_refs->num_heads--;
554 if (!head->processing)
555 delayed_refs->num_heads_ready--;
556}
557
558/*
559 * Helper to insert the ref_node to the tail or merge with tail.
560 *
561 * Return false if the ref was inserted.
562 * Return true if the ref was merged into an existing one (and therefore can be
563 * freed by the caller).
564 */
565static bool insert_delayed_ref(struct btrfs_trans_handle *trans,
566 struct btrfs_delayed_ref_head *href,
567 struct btrfs_delayed_ref_node *ref)
568{
569 struct btrfs_delayed_ref_root *root = &trans->transaction->delayed_refs;
570 struct btrfs_delayed_ref_node *exist;
571 int mod;
572
573 spin_lock(&href->lock);
574 exist = tree_insert(&href->ref_tree, ref);
575 if (!exist) {
576 if (ref->action == BTRFS_ADD_DELAYED_REF)
577 list_add_tail(&ref->add_list, &href->ref_add_list);
578 spin_unlock(&href->lock);
579 trans->delayed_ref_updates++;
580 return false;
581 }
582
583 /* Now we are sure we can merge */
584 if (exist->action == ref->action) {
585 mod = ref->ref_mod;
586 } else {
587 /* Need to change action */
588 if (exist->ref_mod < ref->ref_mod) {
589 exist->action = ref->action;
590 mod = -exist->ref_mod;
591 exist->ref_mod = ref->ref_mod;
592 if (ref->action == BTRFS_ADD_DELAYED_REF)
593 list_add_tail(&exist->add_list,
594 &href->ref_add_list);
595 else if (ref->action == BTRFS_DROP_DELAYED_REF) {
596 ASSERT(!list_empty(&exist->add_list));
597 list_del_init(&exist->add_list);
598 } else {
599 ASSERT(0);
600 }
601 } else
602 mod = -ref->ref_mod;
603 }
604 exist->ref_mod += mod;
605
606 /* remove existing tail if its ref_mod is zero */
607 if (exist->ref_mod == 0)
608 drop_delayed_ref(trans->fs_info, root, href, exist);
609 spin_unlock(&href->lock);
610 return true;
611}
612
613/*
614 * helper function to update the accounting in the head ref
615 * existing and update must have the same bytenr
616 */
617static noinline void update_existing_head_ref(struct btrfs_trans_handle *trans,
618 struct btrfs_delayed_ref_head *existing,
619 struct btrfs_delayed_ref_head *update)
620{
621 struct btrfs_delayed_ref_root *delayed_refs =
622 &trans->transaction->delayed_refs;
623 struct btrfs_fs_info *fs_info = trans->fs_info;
624 int old_ref_mod;
625
626 BUG_ON(existing->is_data != update->is_data);
627
628 spin_lock(&existing->lock);
629
630 /*
631 * When freeing an extent, we may not know the owning root when we
632 * first create the head_ref. However, some deref before the last deref
633 * will know it, so we just need to update the head_ref accordingly.
634 */
635 if (!existing->owning_root)
636 existing->owning_root = update->owning_root;
637
638 if (update->must_insert_reserved) {
639 /* if the extent was freed and then
640 * reallocated before the delayed ref
641 * entries were processed, we can end up
642 * with an existing head ref without
643 * the must_insert_reserved flag set.
644 * Set it again here
645 */
646 existing->must_insert_reserved = update->must_insert_reserved;
647 existing->owning_root = update->owning_root;
648
649 /*
650 * update the num_bytes so we make sure the accounting
651 * is done correctly
652 */
653 existing->num_bytes = update->num_bytes;
654
655 }
656
657 if (update->extent_op) {
658 if (!existing->extent_op) {
659 existing->extent_op = update->extent_op;
660 } else {
661 if (update->extent_op->update_key) {
662 memcpy(&existing->extent_op->key,
663 &update->extent_op->key,
664 sizeof(update->extent_op->key));
665 existing->extent_op->update_key = true;
666 }
667 if (update->extent_op->update_flags) {
668 existing->extent_op->flags_to_set |=
669 update->extent_op->flags_to_set;
670 existing->extent_op->update_flags = true;
671 }
672 btrfs_free_delayed_extent_op(update->extent_op);
673 }
674 }
675 /*
676 * update the reference mod on the head to reflect this new operation,
677 * only need the lock for this case cause we could be processing it
678 * currently, for refs we just added we know we're a-ok.
679 */
680 old_ref_mod = existing->total_ref_mod;
681 existing->ref_mod += update->ref_mod;
682 existing->total_ref_mod += update->ref_mod;
683
684 /*
685 * If we are going to from a positive ref mod to a negative or vice
686 * versa we need to make sure to adjust pending_csums accordingly.
687 * We reserve bytes for csum deletion when adding or updating a ref head
688 * see add_delayed_ref_head() for more details.
689 */
690 if (existing->is_data) {
691 u64 csum_leaves =
692 btrfs_csum_bytes_to_leaves(fs_info,
693 existing->num_bytes);
694
695 if (existing->total_ref_mod >= 0 && old_ref_mod < 0) {
696 delayed_refs->pending_csums -= existing->num_bytes;
697 btrfs_delayed_refs_rsv_release(fs_info, 0, csum_leaves);
698 }
699 if (existing->total_ref_mod < 0 && old_ref_mod >= 0) {
700 delayed_refs->pending_csums += existing->num_bytes;
701 trans->delayed_ref_csum_deletions += csum_leaves;
702 }
703 }
704
705 spin_unlock(&existing->lock);
706}
707
708static void init_delayed_ref_head(struct btrfs_delayed_ref_head *head_ref,
709 struct btrfs_ref *generic_ref,
710 struct btrfs_qgroup_extent_record *qrecord,
711 u64 reserved)
712{
713 int count_mod = 1;
714 bool must_insert_reserved = false;
715
716 /* If reserved is provided, it must be a data extent. */
717 BUG_ON(generic_ref->type != BTRFS_REF_DATA && reserved);
718
719 switch (generic_ref->action) {
720 case BTRFS_ADD_DELAYED_REF:
721 /* count_mod is already set to 1. */
722 break;
723 case BTRFS_UPDATE_DELAYED_HEAD:
724 count_mod = 0;
725 break;
726 case BTRFS_DROP_DELAYED_REF:
727 /*
728 * The head node stores the sum of all the mods, so dropping a ref
729 * should drop the sum in the head node by one.
730 */
731 count_mod = -1;
732 break;
733 case BTRFS_ADD_DELAYED_EXTENT:
734 /*
735 * BTRFS_ADD_DELAYED_EXTENT means that we need to update the
736 * reserved accounting when the extent is finally added, or if a
737 * later modification deletes the delayed ref without ever
738 * inserting the extent into the extent allocation tree.
739 * ref->must_insert_reserved is the flag used to record that
740 * accounting mods are required.
741 *
742 * Once we record must_insert_reserved, switch the action to
743 * BTRFS_ADD_DELAYED_REF because other special casing is not
744 * required.
745 */
746 must_insert_reserved = true;
747 break;
748 }
749
750 refcount_set(&head_ref->refs, 1);
751 head_ref->bytenr = generic_ref->bytenr;
752 head_ref->num_bytes = generic_ref->num_bytes;
753 head_ref->ref_mod = count_mod;
754 head_ref->reserved_bytes = reserved;
755 head_ref->must_insert_reserved = must_insert_reserved;
756 head_ref->owning_root = generic_ref->owning_root;
757 head_ref->is_data = (generic_ref->type == BTRFS_REF_DATA);
758 head_ref->is_system = (generic_ref->ref_root == BTRFS_CHUNK_TREE_OBJECTID);
759 head_ref->ref_tree = RB_ROOT_CACHED;
760 INIT_LIST_HEAD(&head_ref->ref_add_list);
761 head_ref->tracked = false;
762 head_ref->processing = false;
763 head_ref->total_ref_mod = count_mod;
764 spin_lock_init(&head_ref->lock);
765 mutex_init(&head_ref->mutex);
766
767 /* If not metadata set an impossible level to help debugging. */
768 if (generic_ref->type == BTRFS_REF_METADATA)
769 head_ref->level = generic_ref->tree_ref.level;
770 else
771 head_ref->level = U8_MAX;
772
773 if (qrecord) {
774 if (generic_ref->ref_root && reserved) {
775 qrecord->data_rsv = reserved;
776 qrecord->data_rsv_refroot = generic_ref->ref_root;
777 }
778 qrecord->num_bytes = generic_ref->num_bytes;
779 qrecord->old_roots = NULL;
780 }
781}
782
783/*
784 * helper function to actually insert a head node into the rbtree.
785 * this does all the dirty work in terms of maintaining the correct
786 * overall modification count.
787 *
788 * Returns an error pointer in case of an error.
789 */
790static noinline struct btrfs_delayed_ref_head *
791add_delayed_ref_head(struct btrfs_trans_handle *trans,
792 struct btrfs_delayed_ref_head *head_ref,
793 struct btrfs_qgroup_extent_record *qrecord,
794 int action, bool *qrecord_inserted_ret)
795{
796 struct btrfs_fs_info *fs_info = trans->fs_info;
797 struct btrfs_delayed_ref_head *existing;
798 struct btrfs_delayed_ref_root *delayed_refs;
799 const unsigned long index = (head_ref->bytenr >> fs_info->sectorsize_bits);
800 bool qrecord_inserted = false;
801
802 delayed_refs = &trans->transaction->delayed_refs;
803 lockdep_assert_held(&delayed_refs->lock);
804
805#if BITS_PER_LONG == 32
806 if (head_ref->bytenr >= MAX_LFS_FILESIZE) {
807 if (qrecord)
808 xa_release(&delayed_refs->dirty_extents, index);
809 btrfs_err_rl(fs_info,
810"delayed ref head %llu is beyond 32bit page cache and xarray index limit",
811 head_ref->bytenr);
812 btrfs_err_32bit_limit(fs_info);
813 return ERR_PTR(-EOVERFLOW);
814 }
815#endif
816
817 /* Record qgroup extent info if provided */
818 if (qrecord) {
819 int ret;
820
821 ret = btrfs_qgroup_trace_extent_nolock(fs_info, delayed_refs, qrecord,
822 head_ref->bytenr);
823 if (ret) {
824 /* Clean up if insertion fails or item exists. */
825 xa_release(&delayed_refs->dirty_extents, index);
826 /* Caller responsible for freeing qrecord on error. */
827 if (ret < 0)
828 return ERR_PTR(ret);
829 kfree(qrecord);
830 } else {
831 qrecord_inserted = true;
832 }
833 }
834
835 trace_add_delayed_ref_head(fs_info, head_ref, action);
836
837 existing = xa_load(&delayed_refs->head_refs, index);
838 if (existing) {
839 update_existing_head_ref(trans, existing, head_ref);
840 /*
841 * we've updated the existing ref, free the newly
842 * allocated ref
843 */
844 kmem_cache_free(btrfs_delayed_ref_head_cachep, head_ref);
845 head_ref = existing;
846 } else {
847 existing = xa_store(&delayed_refs->head_refs, index, head_ref, GFP_ATOMIC);
848 if (xa_is_err(existing)) {
849 /* Memory was preallocated by the caller. */
850 ASSERT(xa_err(existing) != -ENOMEM);
851 return ERR_PTR(xa_err(existing));
852 } else if (WARN_ON(existing)) {
853 /*
854 * Shouldn't happen we just did a lookup before under
855 * delayed_refs->lock.
856 */
857 return ERR_PTR(-EEXIST);
858 }
859 head_ref->tracked = true;
860 /*
861 * We reserve the amount of bytes needed to delete csums when
862 * adding the ref head and not when adding individual drop refs
863 * since the csum items are deleted only after running the last
864 * delayed drop ref (the data extent's ref count drops to 0).
865 */
866 if (head_ref->is_data && head_ref->ref_mod < 0) {
867 delayed_refs->pending_csums += head_ref->num_bytes;
868 trans->delayed_ref_csum_deletions +=
869 btrfs_csum_bytes_to_leaves(fs_info, head_ref->num_bytes);
870 }
871 delayed_refs->num_heads++;
872 delayed_refs->num_heads_ready++;
873 }
874 if (qrecord_inserted_ret)
875 *qrecord_inserted_ret = qrecord_inserted;
876
877 return head_ref;
878}
879
880/*
881 * Initialize the structure which represents a modification to a an extent.
882 *
883 * @fs_info: Internal to the mounted filesystem mount structure.
884 *
885 * @ref: The structure which is going to be initialized.
886 *
887 * @bytenr: The logical address of the extent for which a modification is
888 * going to be recorded.
889 *
890 * @num_bytes: Size of the extent whose modification is being recorded.
891 *
892 * @ref_root: The id of the root where this modification has originated, this
893 * can be either one of the well-known metadata trees or the
894 * subvolume id which references this extent.
895 *
896 * @action: Can be one of BTRFS_ADD_DELAYED_REF/BTRFS_DROP_DELAYED_REF or
897 * BTRFS_ADD_DELAYED_EXTENT
898 *
899 * @ref_type: Holds the type of the extent which is being recorded, can be
900 * one of BTRFS_SHARED_BLOCK_REF_KEY/BTRFS_TREE_BLOCK_REF_KEY
901 * when recording a metadata extent or BTRFS_SHARED_DATA_REF_KEY/
902 * BTRFS_EXTENT_DATA_REF_KEY when recording data extent
903 */
904static void init_delayed_ref_common(struct btrfs_fs_info *fs_info,
905 struct btrfs_delayed_ref_node *ref,
906 struct btrfs_ref *generic_ref)
907{
908 int action = generic_ref->action;
909 u64 seq = 0;
910
911 if (action == BTRFS_ADD_DELAYED_EXTENT)
912 action = BTRFS_ADD_DELAYED_REF;
913
914 if (is_fstree(generic_ref->ref_root))
915 seq = atomic64_read(&fs_info->tree_mod_seq);
916
917 refcount_set(&ref->refs, 1);
918 ref->bytenr = generic_ref->bytenr;
919 ref->num_bytes = generic_ref->num_bytes;
920 ref->ref_mod = 1;
921 ref->action = action;
922 ref->seq = seq;
923 ref->type = btrfs_ref_type(generic_ref);
924 ref->ref_root = generic_ref->ref_root;
925 ref->parent = generic_ref->parent;
926 RB_CLEAR_NODE(&ref->ref_node);
927 INIT_LIST_HEAD(&ref->add_list);
928
929 if (generic_ref->type == BTRFS_REF_DATA)
930 ref->data_ref = generic_ref->data_ref;
931 else
932 ref->tree_ref = generic_ref->tree_ref;
933}
934
935void btrfs_init_tree_ref(struct btrfs_ref *generic_ref, int level, u64 mod_root,
936 bool skip_qgroup)
937{
938#ifdef CONFIG_BTRFS_FS_REF_VERIFY
939 /* If @real_root not set, use @root as fallback */
940 generic_ref->real_root = mod_root ?: generic_ref->ref_root;
941#endif
942 generic_ref->tree_ref.level = level;
943 generic_ref->type = BTRFS_REF_METADATA;
944 if (skip_qgroup || !(is_fstree(generic_ref->ref_root) &&
945 (!mod_root || is_fstree(mod_root))))
946 generic_ref->skip_qgroup = true;
947 else
948 generic_ref->skip_qgroup = false;
949
950}
951
952void btrfs_init_data_ref(struct btrfs_ref *generic_ref, u64 ino, u64 offset,
953 u64 mod_root, bool skip_qgroup)
954{
955#ifdef CONFIG_BTRFS_FS_REF_VERIFY
956 /* If @real_root not set, use @root as fallback */
957 generic_ref->real_root = mod_root ?: generic_ref->ref_root;
958#endif
959 generic_ref->data_ref.objectid = ino;
960 generic_ref->data_ref.offset = offset;
961 generic_ref->type = BTRFS_REF_DATA;
962 if (skip_qgroup || !(is_fstree(generic_ref->ref_root) &&
963 (!mod_root || is_fstree(mod_root))))
964 generic_ref->skip_qgroup = true;
965 else
966 generic_ref->skip_qgroup = false;
967}
968
969static int add_delayed_ref(struct btrfs_trans_handle *trans,
970 struct btrfs_ref *generic_ref,
971 struct btrfs_delayed_extent_op *extent_op,
972 u64 reserved)
973{
974 struct btrfs_fs_info *fs_info = trans->fs_info;
975 struct btrfs_delayed_ref_node *node;
976 struct btrfs_delayed_ref_head *head_ref;
977 struct btrfs_delayed_ref_head *new_head_ref;
978 struct btrfs_delayed_ref_root *delayed_refs;
979 struct btrfs_qgroup_extent_record *record = NULL;
980 const unsigned long index = (generic_ref->bytenr >> fs_info->sectorsize_bits);
981 bool qrecord_reserved = false;
982 bool qrecord_inserted;
983 int action = generic_ref->action;
984 bool merged;
985 int ret;
986
987 node = kmem_cache_alloc(btrfs_delayed_ref_node_cachep, GFP_NOFS);
988 if (!node)
989 return -ENOMEM;
990
991 head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
992 if (!head_ref) {
993 ret = -ENOMEM;
994 goto free_node;
995 }
996
997 delayed_refs = &trans->transaction->delayed_refs;
998
999 if (btrfs_qgroup_full_accounting(fs_info) && !generic_ref->skip_qgroup) {
1000 record = kzalloc(sizeof(*record), GFP_NOFS);
1001 if (!record) {
1002 ret = -ENOMEM;
1003 goto free_head_ref;
1004 }
1005 if (xa_reserve(&delayed_refs->dirty_extents, index, GFP_NOFS)) {
1006 ret = -ENOMEM;
1007 goto free_record;
1008 }
1009 qrecord_reserved = true;
1010 }
1011
1012 ret = xa_reserve(&delayed_refs->head_refs, index, GFP_NOFS);
1013 if (ret) {
1014 if (qrecord_reserved)
1015 xa_release(&delayed_refs->dirty_extents, index);
1016 goto free_record;
1017 }
1018
1019 init_delayed_ref_common(fs_info, node, generic_ref);
1020 init_delayed_ref_head(head_ref, generic_ref, record, reserved);
1021 head_ref->extent_op = extent_op;
1022
1023 spin_lock(&delayed_refs->lock);
1024
1025 /*
1026 * insert both the head node and the new ref without dropping
1027 * the spin lock
1028 */
1029 new_head_ref = add_delayed_ref_head(trans, head_ref, record,
1030 action, &qrecord_inserted);
1031 if (IS_ERR(new_head_ref)) {
1032 xa_release(&delayed_refs->head_refs, index);
1033 spin_unlock(&delayed_refs->lock);
1034 ret = PTR_ERR(new_head_ref);
1035 goto free_record;
1036 }
1037 head_ref = new_head_ref;
1038
1039 merged = insert_delayed_ref(trans, head_ref, node);
1040 spin_unlock(&delayed_refs->lock);
1041
1042 /*
1043 * Need to update the delayed_refs_rsv with any changes we may have
1044 * made.
1045 */
1046 btrfs_update_delayed_refs_rsv(trans);
1047
1048 if (generic_ref->type == BTRFS_REF_DATA)
1049 trace_add_delayed_data_ref(trans->fs_info, node);
1050 else
1051 trace_add_delayed_tree_ref(trans->fs_info, node);
1052 if (merged)
1053 kmem_cache_free(btrfs_delayed_ref_node_cachep, node);
1054
1055 if (qrecord_inserted)
1056 return btrfs_qgroup_trace_extent_post(trans, record, generic_ref->bytenr);
1057 return 0;
1058
1059free_record:
1060 kfree(record);
1061free_head_ref:
1062 kmem_cache_free(btrfs_delayed_ref_head_cachep, head_ref);
1063free_node:
1064 kmem_cache_free(btrfs_delayed_ref_node_cachep, node);
1065 return ret;
1066}
1067
1068/*
1069 * Add a delayed tree ref. This does all of the accounting required to make sure
1070 * the delayed ref is eventually processed before this transaction commits.
1071 */
1072int btrfs_add_delayed_tree_ref(struct btrfs_trans_handle *trans,
1073 struct btrfs_ref *generic_ref,
1074 struct btrfs_delayed_extent_op *extent_op)
1075{
1076 ASSERT(generic_ref->type == BTRFS_REF_METADATA && generic_ref->action);
1077 return add_delayed_ref(trans, generic_ref, extent_op, 0);
1078}
1079
1080/*
1081 * add a delayed data ref. it's similar to btrfs_add_delayed_tree_ref.
1082 */
1083int btrfs_add_delayed_data_ref(struct btrfs_trans_handle *trans,
1084 struct btrfs_ref *generic_ref,
1085 u64 reserved)
1086{
1087 ASSERT(generic_ref->type == BTRFS_REF_DATA && generic_ref->action);
1088 return add_delayed_ref(trans, generic_ref, NULL, reserved);
1089}
1090
1091int btrfs_add_delayed_extent_op(struct btrfs_trans_handle *trans,
1092 u64 bytenr, u64 num_bytes, u8 level,
1093 struct btrfs_delayed_extent_op *extent_op)
1094{
1095 const unsigned long index = (bytenr >> trans->fs_info->sectorsize_bits);
1096 struct btrfs_delayed_ref_head *head_ref;
1097 struct btrfs_delayed_ref_head *head_ref_ret;
1098 struct btrfs_delayed_ref_root *delayed_refs;
1099 struct btrfs_ref generic_ref = {
1100 .type = BTRFS_REF_METADATA,
1101 .action = BTRFS_UPDATE_DELAYED_HEAD,
1102 .bytenr = bytenr,
1103 .num_bytes = num_bytes,
1104 .tree_ref.level = level,
1105 };
1106 int ret;
1107
1108 head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
1109 if (!head_ref)
1110 return -ENOMEM;
1111
1112 init_delayed_ref_head(head_ref, &generic_ref, NULL, 0);
1113 head_ref->extent_op = extent_op;
1114
1115 delayed_refs = &trans->transaction->delayed_refs;
1116
1117 ret = xa_reserve(&delayed_refs->head_refs, index, GFP_NOFS);
1118 if (ret) {
1119 kmem_cache_free(btrfs_delayed_ref_head_cachep, head_ref);
1120 return ret;
1121 }
1122
1123 spin_lock(&delayed_refs->lock);
1124 head_ref_ret = add_delayed_ref_head(trans, head_ref, NULL,
1125 BTRFS_UPDATE_DELAYED_HEAD, NULL);
1126 if (IS_ERR(head_ref_ret)) {
1127 xa_release(&delayed_refs->head_refs, index);
1128 spin_unlock(&delayed_refs->lock);
1129 kmem_cache_free(btrfs_delayed_ref_head_cachep, head_ref);
1130 return PTR_ERR(head_ref_ret);
1131 }
1132 spin_unlock(&delayed_refs->lock);
1133
1134 /*
1135 * Need to update the delayed_refs_rsv with any changes we may have
1136 * made.
1137 */
1138 btrfs_update_delayed_refs_rsv(trans);
1139 return 0;
1140}
1141
1142void btrfs_put_delayed_ref(struct btrfs_delayed_ref_node *ref)
1143{
1144 if (refcount_dec_and_test(&ref->refs)) {
1145 WARN_ON(!RB_EMPTY_NODE(&ref->ref_node));
1146 kmem_cache_free(btrfs_delayed_ref_node_cachep, ref);
1147 }
1148}
1149
1150/*
1151 * This does a simple search for the head node for a given extent. Returns the
1152 * head node if found, or NULL if not.
1153 */
1154struct btrfs_delayed_ref_head *
1155btrfs_find_delayed_ref_head(const struct btrfs_fs_info *fs_info,
1156 struct btrfs_delayed_ref_root *delayed_refs,
1157 u64 bytenr)
1158{
1159 const unsigned long index = (bytenr >> fs_info->sectorsize_bits);
1160
1161 lockdep_assert_held(&delayed_refs->lock);
1162
1163 return xa_load(&delayed_refs->head_refs, index);
1164}
1165
1166static int find_comp(struct btrfs_delayed_ref_node *entry, u64 root, u64 parent)
1167{
1168 int type = parent ? BTRFS_SHARED_BLOCK_REF_KEY : BTRFS_TREE_BLOCK_REF_KEY;
1169
1170 if (type < entry->type)
1171 return -1;
1172 if (type > entry->type)
1173 return 1;
1174
1175 if (type == BTRFS_TREE_BLOCK_REF_KEY) {
1176 if (root < entry->ref_root)
1177 return -1;
1178 if (root > entry->ref_root)
1179 return 1;
1180 } else {
1181 if (parent < entry->parent)
1182 return -1;
1183 if (parent > entry->parent)
1184 return 1;
1185 }
1186 return 0;
1187}
1188
1189/*
1190 * Check to see if a given root/parent reference is attached to the head. This
1191 * only checks for BTRFS_ADD_DELAYED_REF references that match, as that
1192 * indicates the reference exists for the given root or parent. This is for
1193 * tree blocks only.
1194 *
1195 * @head: the head of the bytenr we're searching.
1196 * @root: the root objectid of the reference if it is a normal reference.
1197 * @parent: the parent if this is a shared backref.
1198 */
1199bool btrfs_find_delayed_tree_ref(struct btrfs_delayed_ref_head *head,
1200 u64 root, u64 parent)
1201{
1202 struct rb_node *node;
1203 bool found = false;
1204
1205 lockdep_assert_held(&head->mutex);
1206
1207 spin_lock(&head->lock);
1208 node = head->ref_tree.rb_root.rb_node;
1209 while (node) {
1210 struct btrfs_delayed_ref_node *entry;
1211 int ret;
1212
1213 entry = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
1214 ret = find_comp(entry, root, parent);
1215 if (ret < 0) {
1216 node = node->rb_left;
1217 } else if (ret > 0) {
1218 node = node->rb_right;
1219 } else {
1220 /*
1221 * We only want to count ADD actions, as drops mean the
1222 * ref doesn't exist.
1223 */
1224 if (entry->action == BTRFS_ADD_DELAYED_REF)
1225 found = true;
1226 break;
1227 }
1228 }
1229 spin_unlock(&head->lock);
1230 return found;
1231}
1232
1233void btrfs_destroy_delayed_refs(struct btrfs_transaction *trans)
1234{
1235 struct btrfs_delayed_ref_root *delayed_refs = &trans->delayed_refs;
1236 struct btrfs_fs_info *fs_info = trans->fs_info;
1237
1238 spin_lock(&delayed_refs->lock);
1239 while (true) {
1240 struct btrfs_delayed_ref_head *head;
1241 struct rb_node *n;
1242 bool pin_bytes = false;
1243
1244 head = find_first_ref_head(delayed_refs);
1245 if (!head)
1246 break;
1247
1248 if (!btrfs_delayed_ref_lock(delayed_refs, head))
1249 continue;
1250
1251 spin_lock(&head->lock);
1252 while ((n = rb_first_cached(&head->ref_tree)) != NULL) {
1253 struct btrfs_delayed_ref_node *ref;
1254
1255 ref = rb_entry(n, struct btrfs_delayed_ref_node, ref_node);
1256 drop_delayed_ref(fs_info, delayed_refs, head, ref);
1257 }
1258 if (head->must_insert_reserved)
1259 pin_bytes = true;
1260 btrfs_free_delayed_extent_op(head->extent_op);
1261 btrfs_delete_ref_head(fs_info, delayed_refs, head);
1262 spin_unlock(&head->lock);
1263 spin_unlock(&delayed_refs->lock);
1264 mutex_unlock(&head->mutex);
1265
1266 if (pin_bytes) {
1267 struct btrfs_block_group *bg;
1268
1269 bg = btrfs_lookup_block_group(fs_info, head->bytenr);
1270 if (WARN_ON_ONCE(bg == NULL)) {
1271 /*
1272 * Unexpected and there's nothing we can do here
1273 * because we are in a transaction abort path,
1274 * so any errors can only be ignored or reported
1275 * while attempting to cleanup all resources.
1276 */
1277 btrfs_err(fs_info,
1278"block group for delayed ref at %llu was not found while destroying ref head",
1279 head->bytenr);
1280 } else {
1281 spin_lock(&bg->space_info->lock);
1282 spin_lock(&bg->lock);
1283 bg->pinned += head->num_bytes;
1284 btrfs_space_info_update_bytes_pinned(fs_info,
1285 bg->space_info,
1286 head->num_bytes);
1287 bg->reserved -= head->num_bytes;
1288 bg->space_info->bytes_reserved -= head->num_bytes;
1289 spin_unlock(&bg->lock);
1290 spin_unlock(&bg->space_info->lock);
1291
1292 btrfs_put_block_group(bg);
1293 }
1294
1295 btrfs_error_unpin_extent_range(fs_info, head->bytenr,
1296 head->bytenr + head->num_bytes - 1);
1297 }
1298 btrfs_cleanup_ref_head_accounting(fs_info, delayed_refs, head);
1299 btrfs_put_delayed_ref_head(head);
1300 cond_resched();
1301 spin_lock(&delayed_refs->lock);
1302 }
1303 btrfs_qgroup_destroy_extent_records(trans);
1304
1305 spin_unlock(&delayed_refs->lock);
1306}
1307
1308void __cold btrfs_delayed_ref_exit(void)
1309{
1310 kmem_cache_destroy(btrfs_delayed_ref_head_cachep);
1311 kmem_cache_destroy(btrfs_delayed_ref_node_cachep);
1312 kmem_cache_destroy(btrfs_delayed_extent_op_cachep);
1313}
1314
1315int __init btrfs_delayed_ref_init(void)
1316{
1317 btrfs_delayed_ref_head_cachep = KMEM_CACHE(btrfs_delayed_ref_head, 0);
1318 if (!btrfs_delayed_ref_head_cachep)
1319 goto fail;
1320
1321 btrfs_delayed_ref_node_cachep = KMEM_CACHE(btrfs_delayed_ref_node, 0);
1322 if (!btrfs_delayed_ref_node_cachep)
1323 goto fail;
1324
1325 btrfs_delayed_extent_op_cachep = KMEM_CACHE(btrfs_delayed_extent_op, 0);
1326 if (!btrfs_delayed_extent_op_cachep)
1327 goto fail;
1328
1329 return 0;
1330fail:
1331 btrfs_delayed_ref_exit();
1332 return -ENOMEM;
1333}
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#include "qgroup.h"
26
27struct kmem_cache *btrfs_delayed_ref_head_cachep;
28struct kmem_cache *btrfs_delayed_tree_ref_cachep;
29struct kmem_cache *btrfs_delayed_data_ref_cachep;
30struct kmem_cache *btrfs_delayed_extent_op_cachep;
31/*
32 * delayed back reference update tracking. For subvolume trees
33 * we queue up extent allocations and backref maintenance for
34 * delayed processing. This avoids deep call chains where we
35 * add extents in the middle of btrfs_search_slot, and it allows
36 * us to buffer up frequently modified backrefs in an rb tree instead
37 * of hammering updates on the extent allocation tree.
38 */
39
40/*
41 * compare two delayed tree backrefs with same bytenr and type
42 */
43static int comp_tree_refs(struct btrfs_delayed_tree_ref *ref2,
44 struct btrfs_delayed_tree_ref *ref1, int type)
45{
46 if (type == BTRFS_TREE_BLOCK_REF_KEY) {
47 if (ref1->root < ref2->root)
48 return -1;
49 if (ref1->root > ref2->root)
50 return 1;
51 } else {
52 if (ref1->parent < ref2->parent)
53 return -1;
54 if (ref1->parent > ref2->parent)
55 return 1;
56 }
57 return 0;
58}
59
60/*
61 * compare two delayed data backrefs with same bytenr and type
62 */
63static int comp_data_refs(struct btrfs_delayed_data_ref *ref2,
64 struct btrfs_delayed_data_ref *ref1)
65{
66 if (ref1->node.type == BTRFS_EXTENT_DATA_REF_KEY) {
67 if (ref1->root < ref2->root)
68 return -1;
69 if (ref1->root > ref2->root)
70 return 1;
71 if (ref1->objectid < ref2->objectid)
72 return -1;
73 if (ref1->objectid > ref2->objectid)
74 return 1;
75 if (ref1->offset < ref2->offset)
76 return -1;
77 if (ref1->offset > ref2->offset)
78 return 1;
79 } else {
80 if (ref1->parent < ref2->parent)
81 return -1;
82 if (ref1->parent > ref2->parent)
83 return 1;
84 }
85 return 0;
86}
87
88/* insert a new ref to head ref rbtree */
89static struct btrfs_delayed_ref_head *htree_insert(struct rb_root *root,
90 struct rb_node *node)
91{
92 struct rb_node **p = &root->rb_node;
93 struct rb_node *parent_node = NULL;
94 struct btrfs_delayed_ref_head *entry;
95 struct btrfs_delayed_ref_head *ins;
96 u64 bytenr;
97
98 ins = rb_entry(node, struct btrfs_delayed_ref_head, href_node);
99 bytenr = ins->node.bytenr;
100 while (*p) {
101 parent_node = *p;
102 entry = rb_entry(parent_node, struct btrfs_delayed_ref_head,
103 href_node);
104
105 if (bytenr < entry->node.bytenr)
106 p = &(*p)->rb_left;
107 else if (bytenr > entry->node.bytenr)
108 p = &(*p)->rb_right;
109 else
110 return entry;
111 }
112
113 rb_link_node(node, parent_node, p);
114 rb_insert_color(node, root);
115 return NULL;
116}
117
118/*
119 * find an head entry based on bytenr. This returns the delayed ref
120 * head if it was able to find one, or NULL if nothing was in that spot.
121 * If return_bigger is given, the next bigger entry is returned if no exact
122 * match is found.
123 */
124static struct btrfs_delayed_ref_head *
125find_ref_head(struct rb_root *root, u64 bytenr,
126 int return_bigger)
127{
128 struct rb_node *n;
129 struct btrfs_delayed_ref_head *entry;
130
131 n = root->rb_node;
132 entry = NULL;
133 while (n) {
134 entry = rb_entry(n, struct btrfs_delayed_ref_head, href_node);
135
136 if (bytenr < entry->node.bytenr)
137 n = n->rb_left;
138 else if (bytenr > entry->node.bytenr)
139 n = n->rb_right;
140 else
141 return entry;
142 }
143 if (entry && return_bigger) {
144 if (bytenr > entry->node.bytenr) {
145 n = rb_next(&entry->href_node);
146 if (!n)
147 n = rb_first(root);
148 entry = rb_entry(n, struct btrfs_delayed_ref_head,
149 href_node);
150 return entry;
151 }
152 return entry;
153 }
154 return NULL;
155}
156
157int btrfs_delayed_ref_lock(struct btrfs_trans_handle *trans,
158 struct btrfs_delayed_ref_head *head)
159{
160 struct btrfs_delayed_ref_root *delayed_refs;
161
162 delayed_refs = &trans->transaction->delayed_refs;
163 assert_spin_locked(&delayed_refs->lock);
164 if (mutex_trylock(&head->mutex))
165 return 0;
166
167 atomic_inc(&head->node.refs);
168 spin_unlock(&delayed_refs->lock);
169
170 mutex_lock(&head->mutex);
171 spin_lock(&delayed_refs->lock);
172 if (!head->node.in_tree) {
173 mutex_unlock(&head->mutex);
174 btrfs_put_delayed_ref(&head->node);
175 return -EAGAIN;
176 }
177 btrfs_put_delayed_ref(&head->node);
178 return 0;
179}
180
181static inline void drop_delayed_ref(struct btrfs_trans_handle *trans,
182 struct btrfs_delayed_ref_root *delayed_refs,
183 struct btrfs_delayed_ref_head *head,
184 struct btrfs_delayed_ref_node *ref)
185{
186 if (btrfs_delayed_ref_is_head(ref)) {
187 head = btrfs_delayed_node_to_head(ref);
188 rb_erase(&head->href_node, &delayed_refs->href_root);
189 } else {
190 assert_spin_locked(&head->lock);
191 list_del(&ref->list);
192 }
193 ref->in_tree = 0;
194 btrfs_put_delayed_ref(ref);
195 atomic_dec(&delayed_refs->num_entries);
196 if (trans->delayed_ref_updates)
197 trans->delayed_ref_updates--;
198}
199
200static bool merge_ref(struct btrfs_trans_handle *trans,
201 struct btrfs_delayed_ref_root *delayed_refs,
202 struct btrfs_delayed_ref_head *head,
203 struct btrfs_delayed_ref_node *ref,
204 u64 seq)
205{
206 struct btrfs_delayed_ref_node *next;
207 bool done = false;
208
209 next = list_first_entry(&head->ref_list, struct btrfs_delayed_ref_node,
210 list);
211 while (!done && &next->list != &head->ref_list) {
212 int mod;
213 struct btrfs_delayed_ref_node *next2;
214
215 next2 = list_next_entry(next, list);
216
217 if (next == ref)
218 goto next;
219
220 if (seq && next->seq >= seq)
221 goto next;
222
223 if (next->type != ref->type)
224 goto next;
225
226 if ((ref->type == BTRFS_TREE_BLOCK_REF_KEY ||
227 ref->type == BTRFS_SHARED_BLOCK_REF_KEY) &&
228 comp_tree_refs(btrfs_delayed_node_to_tree_ref(ref),
229 btrfs_delayed_node_to_tree_ref(next),
230 ref->type))
231 goto next;
232 if ((ref->type == BTRFS_EXTENT_DATA_REF_KEY ||
233 ref->type == BTRFS_SHARED_DATA_REF_KEY) &&
234 comp_data_refs(btrfs_delayed_node_to_data_ref(ref),
235 btrfs_delayed_node_to_data_ref(next)))
236 goto next;
237
238 if (ref->action == next->action) {
239 mod = next->ref_mod;
240 } else {
241 if (ref->ref_mod < next->ref_mod) {
242 swap(ref, next);
243 done = true;
244 }
245 mod = -next->ref_mod;
246 }
247
248 drop_delayed_ref(trans, delayed_refs, head, next);
249 ref->ref_mod += mod;
250 if (ref->ref_mod == 0) {
251 drop_delayed_ref(trans, delayed_refs, head, ref);
252 done = true;
253 } else {
254 /*
255 * Can't have multiples of the same ref on a tree block.
256 */
257 WARN_ON(ref->type == BTRFS_TREE_BLOCK_REF_KEY ||
258 ref->type == BTRFS_SHARED_BLOCK_REF_KEY);
259 }
260next:
261 next = next2;
262 }
263
264 return done;
265}
266
267void btrfs_merge_delayed_refs(struct btrfs_trans_handle *trans,
268 struct btrfs_fs_info *fs_info,
269 struct btrfs_delayed_ref_root *delayed_refs,
270 struct btrfs_delayed_ref_head *head)
271{
272 struct btrfs_delayed_ref_node *ref;
273 u64 seq = 0;
274
275 assert_spin_locked(&head->lock);
276
277 if (list_empty(&head->ref_list))
278 return;
279
280 /* We don't have too many refs to merge for data. */
281 if (head->is_data)
282 return;
283
284 spin_lock(&fs_info->tree_mod_seq_lock);
285 if (!list_empty(&fs_info->tree_mod_seq_list)) {
286 struct seq_list *elem;
287
288 elem = list_first_entry(&fs_info->tree_mod_seq_list,
289 struct seq_list, list);
290 seq = elem->seq;
291 }
292 spin_unlock(&fs_info->tree_mod_seq_lock);
293
294 ref = list_first_entry(&head->ref_list, struct btrfs_delayed_ref_node,
295 list);
296 while (&ref->list != &head->ref_list) {
297 if (seq && ref->seq >= seq)
298 goto next;
299
300 if (merge_ref(trans, delayed_refs, head, ref, seq)) {
301 if (list_empty(&head->ref_list))
302 break;
303 ref = list_first_entry(&head->ref_list,
304 struct btrfs_delayed_ref_node,
305 list);
306 continue;
307 }
308next:
309 ref = list_next_entry(ref, list);
310 }
311}
312
313int btrfs_check_delayed_seq(struct btrfs_fs_info *fs_info,
314 struct btrfs_delayed_ref_root *delayed_refs,
315 u64 seq)
316{
317 struct seq_list *elem;
318 int ret = 0;
319
320 spin_lock(&fs_info->tree_mod_seq_lock);
321 if (!list_empty(&fs_info->tree_mod_seq_list)) {
322 elem = list_first_entry(&fs_info->tree_mod_seq_list,
323 struct seq_list, list);
324 if (seq >= elem->seq) {
325 pr_debug("holding back delayed_ref %#x.%x, lowest is %#x.%x (%p)\n",
326 (u32)(seq >> 32), (u32)seq,
327 (u32)(elem->seq >> 32), (u32)elem->seq,
328 delayed_refs);
329 ret = 1;
330 }
331 }
332
333 spin_unlock(&fs_info->tree_mod_seq_lock);
334 return ret;
335}
336
337struct btrfs_delayed_ref_head *
338btrfs_select_ref_head(struct btrfs_trans_handle *trans)
339{
340 struct btrfs_delayed_ref_root *delayed_refs;
341 struct btrfs_delayed_ref_head *head;
342 u64 start;
343 bool loop = false;
344
345 delayed_refs = &trans->transaction->delayed_refs;
346
347again:
348 start = delayed_refs->run_delayed_start;
349 head = find_ref_head(&delayed_refs->href_root, start, 1);
350 if (!head && !loop) {
351 delayed_refs->run_delayed_start = 0;
352 start = 0;
353 loop = true;
354 head = find_ref_head(&delayed_refs->href_root, start, 1);
355 if (!head)
356 return NULL;
357 } else if (!head && loop) {
358 return NULL;
359 }
360
361 while (head->processing) {
362 struct rb_node *node;
363
364 node = rb_next(&head->href_node);
365 if (!node) {
366 if (loop)
367 return NULL;
368 delayed_refs->run_delayed_start = 0;
369 start = 0;
370 loop = true;
371 goto again;
372 }
373 head = rb_entry(node, struct btrfs_delayed_ref_head,
374 href_node);
375 }
376
377 head->processing = 1;
378 WARN_ON(delayed_refs->num_heads_ready == 0);
379 delayed_refs->num_heads_ready--;
380 delayed_refs->run_delayed_start = head->node.bytenr +
381 head->node.num_bytes;
382 return head;
383}
384
385/*
386 * Helper to insert the ref_node to the tail or merge with tail.
387 *
388 * Return 0 for insert.
389 * Return >0 for merge.
390 */
391static int
392add_delayed_ref_tail_merge(struct btrfs_trans_handle *trans,
393 struct btrfs_delayed_ref_root *root,
394 struct btrfs_delayed_ref_head *href,
395 struct btrfs_delayed_ref_node *ref)
396{
397 struct btrfs_delayed_ref_node *exist;
398 int mod;
399 int ret = 0;
400
401 spin_lock(&href->lock);
402 /* Check whether we can merge the tail node with ref */
403 if (list_empty(&href->ref_list))
404 goto add_tail;
405 exist = list_entry(href->ref_list.prev, struct btrfs_delayed_ref_node,
406 list);
407 /* No need to compare bytenr nor is_head */
408 if (exist->type != ref->type || exist->seq != ref->seq)
409 goto add_tail;
410
411 if ((exist->type == BTRFS_TREE_BLOCK_REF_KEY ||
412 exist->type == BTRFS_SHARED_BLOCK_REF_KEY) &&
413 comp_tree_refs(btrfs_delayed_node_to_tree_ref(exist),
414 btrfs_delayed_node_to_tree_ref(ref),
415 ref->type))
416 goto add_tail;
417 if ((exist->type == BTRFS_EXTENT_DATA_REF_KEY ||
418 exist->type == BTRFS_SHARED_DATA_REF_KEY) &&
419 comp_data_refs(btrfs_delayed_node_to_data_ref(exist),
420 btrfs_delayed_node_to_data_ref(ref)))
421 goto add_tail;
422
423 /* Now we are sure we can merge */
424 ret = 1;
425 if (exist->action == ref->action) {
426 mod = ref->ref_mod;
427 } else {
428 /* Need to change action */
429 if (exist->ref_mod < ref->ref_mod) {
430 exist->action = ref->action;
431 mod = -exist->ref_mod;
432 exist->ref_mod = ref->ref_mod;
433 } else
434 mod = -ref->ref_mod;
435 }
436 exist->ref_mod += mod;
437
438 /* remove existing tail if its ref_mod is zero */
439 if (exist->ref_mod == 0)
440 drop_delayed_ref(trans, root, href, exist);
441 spin_unlock(&href->lock);
442 return ret;
443
444add_tail:
445 list_add_tail(&ref->list, &href->ref_list);
446 atomic_inc(&root->num_entries);
447 trans->delayed_ref_updates++;
448 spin_unlock(&href->lock);
449 return ret;
450}
451
452/*
453 * helper function to update the accounting in the head ref
454 * existing and update must have the same bytenr
455 */
456static noinline void
457update_existing_head_ref(struct btrfs_delayed_ref_root *delayed_refs,
458 struct btrfs_delayed_ref_node *existing,
459 struct btrfs_delayed_ref_node *update)
460{
461 struct btrfs_delayed_ref_head *existing_ref;
462 struct btrfs_delayed_ref_head *ref;
463 int old_ref_mod;
464
465 existing_ref = btrfs_delayed_node_to_head(existing);
466 ref = btrfs_delayed_node_to_head(update);
467 BUG_ON(existing_ref->is_data != ref->is_data);
468
469 spin_lock(&existing_ref->lock);
470 if (ref->must_insert_reserved) {
471 /* if the extent was freed and then
472 * reallocated before the delayed ref
473 * entries were processed, we can end up
474 * with an existing head ref without
475 * the must_insert_reserved flag set.
476 * Set it again here
477 */
478 existing_ref->must_insert_reserved = ref->must_insert_reserved;
479
480 /*
481 * update the num_bytes so we make sure the accounting
482 * is done correctly
483 */
484 existing->num_bytes = update->num_bytes;
485
486 }
487
488 if (ref->extent_op) {
489 if (!existing_ref->extent_op) {
490 existing_ref->extent_op = ref->extent_op;
491 } else {
492 if (ref->extent_op->update_key) {
493 memcpy(&existing_ref->extent_op->key,
494 &ref->extent_op->key,
495 sizeof(ref->extent_op->key));
496 existing_ref->extent_op->update_key = true;
497 }
498 if (ref->extent_op->update_flags) {
499 existing_ref->extent_op->flags_to_set |=
500 ref->extent_op->flags_to_set;
501 existing_ref->extent_op->update_flags = true;
502 }
503 btrfs_free_delayed_extent_op(ref->extent_op);
504 }
505 }
506 /*
507 * update the reference mod on the head to reflect this new operation,
508 * only need the lock for this case cause we could be processing it
509 * currently, for refs we just added we know we're a-ok.
510 */
511 old_ref_mod = existing_ref->total_ref_mod;
512 existing->ref_mod += update->ref_mod;
513 existing_ref->total_ref_mod += update->ref_mod;
514
515 /*
516 * If we are going to from a positive ref mod to a negative or vice
517 * versa we need to make sure to adjust pending_csums accordingly.
518 */
519 if (existing_ref->is_data) {
520 if (existing_ref->total_ref_mod >= 0 && old_ref_mod < 0)
521 delayed_refs->pending_csums -= existing->num_bytes;
522 if (existing_ref->total_ref_mod < 0 && old_ref_mod >= 0)
523 delayed_refs->pending_csums += existing->num_bytes;
524 }
525 spin_unlock(&existing_ref->lock);
526}
527
528/*
529 * helper function to actually insert a head node into the rbtree.
530 * this does all the dirty work in terms of maintaining the correct
531 * overall modification count.
532 */
533static noinline struct btrfs_delayed_ref_head *
534add_delayed_ref_head(struct btrfs_fs_info *fs_info,
535 struct btrfs_trans_handle *trans,
536 struct btrfs_delayed_ref_node *ref,
537 struct btrfs_qgroup_extent_record *qrecord,
538 u64 bytenr, u64 num_bytes, u64 ref_root, u64 reserved,
539 int action, int is_data)
540{
541 struct btrfs_delayed_ref_head *existing;
542 struct btrfs_delayed_ref_head *head_ref = NULL;
543 struct btrfs_delayed_ref_root *delayed_refs;
544 struct btrfs_qgroup_extent_record *qexisting;
545 int count_mod = 1;
546 int must_insert_reserved = 0;
547
548 /* If reserved is provided, it must be a data extent. */
549 BUG_ON(!is_data && reserved);
550
551 /*
552 * the head node stores the sum of all the mods, so dropping a ref
553 * should drop the sum in the head node by one.
554 */
555 if (action == BTRFS_UPDATE_DELAYED_HEAD)
556 count_mod = 0;
557 else if (action == BTRFS_DROP_DELAYED_REF)
558 count_mod = -1;
559
560 /*
561 * BTRFS_ADD_DELAYED_EXTENT means that we need to update
562 * the reserved accounting when the extent is finally added, or
563 * if a later modification deletes the delayed ref without ever
564 * inserting the extent into the extent allocation tree.
565 * ref->must_insert_reserved is the flag used to record
566 * that accounting mods are required.
567 *
568 * Once we record must_insert_reserved, switch the action to
569 * BTRFS_ADD_DELAYED_REF because other special casing is not required.
570 */
571 if (action == BTRFS_ADD_DELAYED_EXTENT)
572 must_insert_reserved = 1;
573 else
574 must_insert_reserved = 0;
575
576 delayed_refs = &trans->transaction->delayed_refs;
577
578 /* first set the basic ref node struct up */
579 atomic_set(&ref->refs, 1);
580 ref->bytenr = bytenr;
581 ref->num_bytes = num_bytes;
582 ref->ref_mod = count_mod;
583 ref->type = 0;
584 ref->action = 0;
585 ref->is_head = 1;
586 ref->in_tree = 1;
587 ref->seq = 0;
588
589 head_ref = btrfs_delayed_node_to_head(ref);
590 head_ref->must_insert_reserved = must_insert_reserved;
591 head_ref->is_data = is_data;
592 INIT_LIST_HEAD(&head_ref->ref_list);
593 head_ref->processing = 0;
594 head_ref->total_ref_mod = count_mod;
595 head_ref->qgroup_reserved = 0;
596 head_ref->qgroup_ref_root = 0;
597
598 /* Record qgroup extent info if provided */
599 if (qrecord) {
600 if (ref_root && reserved) {
601 head_ref->qgroup_ref_root = ref_root;
602 head_ref->qgroup_reserved = reserved;
603 }
604
605 qrecord->bytenr = bytenr;
606 qrecord->num_bytes = num_bytes;
607 qrecord->old_roots = NULL;
608
609 qexisting = btrfs_qgroup_insert_dirty_extent(delayed_refs,
610 qrecord);
611 if (qexisting)
612 kfree(qrecord);
613 }
614
615 spin_lock_init(&head_ref->lock);
616 mutex_init(&head_ref->mutex);
617
618 trace_add_delayed_ref_head(ref, head_ref, action);
619
620 existing = htree_insert(&delayed_refs->href_root,
621 &head_ref->href_node);
622 if (existing) {
623 WARN_ON(ref_root && reserved && existing->qgroup_ref_root
624 && existing->qgroup_reserved);
625 update_existing_head_ref(delayed_refs, &existing->node, ref);
626 /*
627 * we've updated the existing ref, free the newly
628 * allocated ref
629 */
630 kmem_cache_free(btrfs_delayed_ref_head_cachep, head_ref);
631 head_ref = existing;
632 } else {
633 if (is_data && count_mod < 0)
634 delayed_refs->pending_csums += num_bytes;
635 delayed_refs->num_heads++;
636 delayed_refs->num_heads_ready++;
637 atomic_inc(&delayed_refs->num_entries);
638 trans->delayed_ref_updates++;
639 }
640 return head_ref;
641}
642
643/*
644 * helper to insert a delayed tree ref into the rbtree.
645 */
646static noinline void
647add_delayed_tree_ref(struct btrfs_fs_info *fs_info,
648 struct btrfs_trans_handle *trans,
649 struct btrfs_delayed_ref_head *head_ref,
650 struct btrfs_delayed_ref_node *ref, u64 bytenr,
651 u64 num_bytes, u64 parent, u64 ref_root, int level,
652 int action)
653{
654 struct btrfs_delayed_tree_ref *full_ref;
655 struct btrfs_delayed_ref_root *delayed_refs;
656 u64 seq = 0;
657 int ret;
658
659 if (action == BTRFS_ADD_DELAYED_EXTENT)
660 action = BTRFS_ADD_DELAYED_REF;
661
662 if (is_fstree(ref_root))
663 seq = atomic64_read(&fs_info->tree_mod_seq);
664 delayed_refs = &trans->transaction->delayed_refs;
665
666 /* first set the basic ref node struct up */
667 atomic_set(&ref->refs, 1);
668 ref->bytenr = bytenr;
669 ref->num_bytes = num_bytes;
670 ref->ref_mod = 1;
671 ref->action = action;
672 ref->is_head = 0;
673 ref->in_tree = 1;
674 ref->seq = seq;
675
676 full_ref = btrfs_delayed_node_to_tree_ref(ref);
677 full_ref->parent = parent;
678 full_ref->root = ref_root;
679 if (parent)
680 ref->type = BTRFS_SHARED_BLOCK_REF_KEY;
681 else
682 ref->type = BTRFS_TREE_BLOCK_REF_KEY;
683 full_ref->level = level;
684
685 trace_add_delayed_tree_ref(ref, full_ref, action);
686
687 ret = add_delayed_ref_tail_merge(trans, delayed_refs, head_ref, ref);
688
689 /*
690 * XXX: memory should be freed at the same level allocated.
691 * But bad practice is anywhere... Follow it now. Need cleanup.
692 */
693 if (ret > 0)
694 kmem_cache_free(btrfs_delayed_tree_ref_cachep, full_ref);
695}
696
697/*
698 * helper to insert a delayed data ref into the rbtree.
699 */
700static noinline void
701add_delayed_data_ref(struct btrfs_fs_info *fs_info,
702 struct btrfs_trans_handle *trans,
703 struct btrfs_delayed_ref_head *head_ref,
704 struct btrfs_delayed_ref_node *ref, u64 bytenr,
705 u64 num_bytes, u64 parent, u64 ref_root, u64 owner,
706 u64 offset, int action)
707{
708 struct btrfs_delayed_data_ref *full_ref;
709 struct btrfs_delayed_ref_root *delayed_refs;
710 u64 seq = 0;
711 int ret;
712
713 if (action == BTRFS_ADD_DELAYED_EXTENT)
714 action = BTRFS_ADD_DELAYED_REF;
715
716 delayed_refs = &trans->transaction->delayed_refs;
717
718 if (is_fstree(ref_root))
719 seq = atomic64_read(&fs_info->tree_mod_seq);
720
721 /* first set the basic ref node struct up */
722 atomic_set(&ref->refs, 1);
723 ref->bytenr = bytenr;
724 ref->num_bytes = num_bytes;
725 ref->ref_mod = 1;
726 ref->action = action;
727 ref->is_head = 0;
728 ref->in_tree = 1;
729 ref->seq = seq;
730
731 full_ref = btrfs_delayed_node_to_data_ref(ref);
732 full_ref->parent = parent;
733 full_ref->root = ref_root;
734 if (parent)
735 ref->type = BTRFS_SHARED_DATA_REF_KEY;
736 else
737 ref->type = BTRFS_EXTENT_DATA_REF_KEY;
738
739 full_ref->objectid = owner;
740 full_ref->offset = offset;
741
742 trace_add_delayed_data_ref(ref, full_ref, action);
743
744 ret = add_delayed_ref_tail_merge(trans, delayed_refs, head_ref, ref);
745
746 if (ret > 0)
747 kmem_cache_free(btrfs_delayed_data_ref_cachep, full_ref);
748}
749
750/*
751 * add a delayed tree ref. This does all of the accounting required
752 * to make sure the delayed ref is eventually processed before this
753 * transaction commits.
754 */
755int btrfs_add_delayed_tree_ref(struct btrfs_fs_info *fs_info,
756 struct btrfs_trans_handle *trans,
757 u64 bytenr, u64 num_bytes, u64 parent,
758 u64 ref_root, int level, int action,
759 struct btrfs_delayed_extent_op *extent_op)
760{
761 struct btrfs_delayed_tree_ref *ref;
762 struct btrfs_delayed_ref_head *head_ref;
763 struct btrfs_delayed_ref_root *delayed_refs;
764 struct btrfs_qgroup_extent_record *record = NULL;
765
766 BUG_ON(extent_op && extent_op->is_data);
767 ref = kmem_cache_alloc(btrfs_delayed_tree_ref_cachep, GFP_NOFS);
768 if (!ref)
769 return -ENOMEM;
770
771 head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
772 if (!head_ref)
773 goto free_ref;
774
775 if (fs_info->quota_enabled && is_fstree(ref_root)) {
776 record = kmalloc(sizeof(*record), GFP_NOFS);
777 if (!record)
778 goto free_head_ref;
779 }
780
781 head_ref->extent_op = extent_op;
782
783 delayed_refs = &trans->transaction->delayed_refs;
784 spin_lock(&delayed_refs->lock);
785
786 /*
787 * insert both the head node and the new ref without dropping
788 * the spin lock
789 */
790 head_ref = add_delayed_ref_head(fs_info, trans, &head_ref->node, record,
791 bytenr, num_bytes, 0, 0, action, 0);
792
793 add_delayed_tree_ref(fs_info, trans, head_ref, &ref->node, bytenr,
794 num_bytes, parent, ref_root, level, action);
795 spin_unlock(&delayed_refs->lock);
796
797 return 0;
798
799free_head_ref:
800 kmem_cache_free(btrfs_delayed_ref_head_cachep, head_ref);
801free_ref:
802 kmem_cache_free(btrfs_delayed_tree_ref_cachep, ref);
803
804 return -ENOMEM;
805}
806
807/*
808 * add a delayed data ref. it's similar to btrfs_add_delayed_tree_ref.
809 */
810int btrfs_add_delayed_data_ref(struct btrfs_fs_info *fs_info,
811 struct btrfs_trans_handle *trans,
812 u64 bytenr, u64 num_bytes,
813 u64 parent, u64 ref_root,
814 u64 owner, u64 offset, u64 reserved, int action,
815 struct btrfs_delayed_extent_op *extent_op)
816{
817 struct btrfs_delayed_data_ref *ref;
818 struct btrfs_delayed_ref_head *head_ref;
819 struct btrfs_delayed_ref_root *delayed_refs;
820 struct btrfs_qgroup_extent_record *record = NULL;
821
822 BUG_ON(extent_op && !extent_op->is_data);
823 ref = kmem_cache_alloc(btrfs_delayed_data_ref_cachep, GFP_NOFS);
824 if (!ref)
825 return -ENOMEM;
826
827 head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
828 if (!head_ref) {
829 kmem_cache_free(btrfs_delayed_data_ref_cachep, ref);
830 return -ENOMEM;
831 }
832
833 if (fs_info->quota_enabled && is_fstree(ref_root)) {
834 record = kmalloc(sizeof(*record), GFP_NOFS);
835 if (!record) {
836 kmem_cache_free(btrfs_delayed_data_ref_cachep, ref);
837 kmem_cache_free(btrfs_delayed_ref_head_cachep,
838 head_ref);
839 return -ENOMEM;
840 }
841 }
842
843 head_ref->extent_op = extent_op;
844
845 delayed_refs = &trans->transaction->delayed_refs;
846 spin_lock(&delayed_refs->lock);
847
848 /*
849 * insert both the head node and the new ref without dropping
850 * the spin lock
851 */
852 head_ref = add_delayed_ref_head(fs_info, trans, &head_ref->node, record,
853 bytenr, num_bytes, ref_root, reserved,
854 action, 1);
855
856 add_delayed_data_ref(fs_info, trans, head_ref, &ref->node, bytenr,
857 num_bytes, parent, ref_root, owner, offset,
858 action);
859 spin_unlock(&delayed_refs->lock);
860
861 return 0;
862}
863
864int btrfs_add_delayed_qgroup_reserve(struct btrfs_fs_info *fs_info,
865 struct btrfs_trans_handle *trans,
866 u64 ref_root, u64 bytenr, u64 num_bytes)
867{
868 struct btrfs_delayed_ref_root *delayed_refs;
869 struct btrfs_delayed_ref_head *ref_head;
870 int ret = 0;
871
872 if (!fs_info->quota_enabled || !is_fstree(ref_root))
873 return 0;
874
875 delayed_refs = &trans->transaction->delayed_refs;
876
877 spin_lock(&delayed_refs->lock);
878 ref_head = find_ref_head(&delayed_refs->href_root, bytenr, 0);
879 if (!ref_head) {
880 ret = -ENOENT;
881 goto out;
882 }
883 WARN_ON(ref_head->qgroup_reserved || ref_head->qgroup_ref_root);
884 ref_head->qgroup_ref_root = ref_root;
885 ref_head->qgroup_reserved = num_bytes;
886out:
887 spin_unlock(&delayed_refs->lock);
888 return ret;
889}
890
891int btrfs_add_delayed_extent_op(struct btrfs_fs_info *fs_info,
892 struct btrfs_trans_handle *trans,
893 u64 bytenr, u64 num_bytes,
894 struct btrfs_delayed_extent_op *extent_op)
895{
896 struct btrfs_delayed_ref_head *head_ref;
897 struct btrfs_delayed_ref_root *delayed_refs;
898
899 head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
900 if (!head_ref)
901 return -ENOMEM;
902
903 head_ref->extent_op = extent_op;
904
905 delayed_refs = &trans->transaction->delayed_refs;
906 spin_lock(&delayed_refs->lock);
907
908 add_delayed_ref_head(fs_info, trans, &head_ref->node, NULL, bytenr,
909 num_bytes, 0, 0, BTRFS_UPDATE_DELAYED_HEAD,
910 extent_op->is_data);
911
912 spin_unlock(&delayed_refs->lock);
913 return 0;
914}
915
916/*
917 * this does a simple search for the head node for a given extent.
918 * It must be called with the delayed ref spinlock held, and it returns
919 * the head node if any where found, or NULL if not.
920 */
921struct btrfs_delayed_ref_head *
922btrfs_find_delayed_ref_head(struct btrfs_trans_handle *trans, u64 bytenr)
923{
924 struct btrfs_delayed_ref_root *delayed_refs;
925
926 delayed_refs = &trans->transaction->delayed_refs;
927 return find_ref_head(&delayed_refs->href_root, bytenr, 0);
928}
929
930void btrfs_delayed_ref_exit(void)
931{
932 kmem_cache_destroy(btrfs_delayed_ref_head_cachep);
933 kmem_cache_destroy(btrfs_delayed_tree_ref_cachep);
934 kmem_cache_destroy(btrfs_delayed_data_ref_cachep);
935 kmem_cache_destroy(btrfs_delayed_extent_op_cachep);
936}
937
938int btrfs_delayed_ref_init(void)
939{
940 btrfs_delayed_ref_head_cachep = kmem_cache_create(
941 "btrfs_delayed_ref_head",
942 sizeof(struct btrfs_delayed_ref_head), 0,
943 SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
944 if (!btrfs_delayed_ref_head_cachep)
945 goto fail;
946
947 btrfs_delayed_tree_ref_cachep = kmem_cache_create(
948 "btrfs_delayed_tree_ref",
949 sizeof(struct btrfs_delayed_tree_ref), 0,
950 SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
951 if (!btrfs_delayed_tree_ref_cachep)
952 goto fail;
953
954 btrfs_delayed_data_ref_cachep = kmem_cache_create(
955 "btrfs_delayed_data_ref",
956 sizeof(struct btrfs_delayed_data_ref), 0,
957 SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
958 if (!btrfs_delayed_data_ref_cachep)
959 goto fail;
960
961 btrfs_delayed_extent_op_cachep = kmem_cache_create(
962 "btrfs_delayed_extent_op",
963 sizeof(struct btrfs_delayed_extent_op), 0,
964 SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
965 if (!btrfs_delayed_extent_op_cachep)
966 goto fail;
967
968 return 0;
969fail:
970 btrfs_delayed_ref_exit();
971 return -ENOMEM;
972}