xref: /linux/fs/btrfs/delayed-ref.c (revision ca55b2fef3a9373fcfc30f82fd26bc7fccbda732)
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 
27 struct kmem_cache *btrfs_delayed_ref_head_cachep;
28 struct kmem_cache *btrfs_delayed_tree_ref_cachep;
29 struct kmem_cache *btrfs_delayed_data_ref_cachep;
30 struct 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  */
43 static 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  */
63 static 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 */
89 static 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  */
124 static struct btrfs_delayed_ref_head *
125 find_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 
157 int 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 
181 static 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 
200 int btrfs_check_delayed_seq(struct btrfs_fs_info *fs_info,
201 			    struct btrfs_delayed_ref_root *delayed_refs,
202 			    u64 seq)
203 {
204 	struct seq_list *elem;
205 	int ret = 0;
206 
207 	spin_lock(&fs_info->tree_mod_seq_lock);
208 	if (!list_empty(&fs_info->tree_mod_seq_list)) {
209 		elem = list_first_entry(&fs_info->tree_mod_seq_list,
210 					struct seq_list, list);
211 		if (seq >= elem->seq) {
212 			pr_debug("holding back delayed_ref %#x.%x, lowest is %#x.%x (%p)\n",
213 				 (u32)(seq >> 32), (u32)seq,
214 				 (u32)(elem->seq >> 32), (u32)elem->seq,
215 				 delayed_refs);
216 			ret = 1;
217 		}
218 	}
219 
220 	spin_unlock(&fs_info->tree_mod_seq_lock);
221 	return ret;
222 }
223 
224 struct btrfs_delayed_ref_head *
225 btrfs_select_ref_head(struct btrfs_trans_handle *trans)
226 {
227 	struct btrfs_delayed_ref_root *delayed_refs;
228 	struct btrfs_delayed_ref_head *head;
229 	u64 start;
230 	bool loop = false;
231 
232 	delayed_refs = &trans->transaction->delayed_refs;
233 
234 again:
235 	start = delayed_refs->run_delayed_start;
236 	head = find_ref_head(&delayed_refs->href_root, start, 1);
237 	if (!head && !loop) {
238 		delayed_refs->run_delayed_start = 0;
239 		start = 0;
240 		loop = true;
241 		head = find_ref_head(&delayed_refs->href_root, start, 1);
242 		if (!head)
243 			return NULL;
244 	} else if (!head && loop) {
245 		return NULL;
246 	}
247 
248 	while (head->processing) {
249 		struct rb_node *node;
250 
251 		node = rb_next(&head->href_node);
252 		if (!node) {
253 			if (loop)
254 				return NULL;
255 			delayed_refs->run_delayed_start = 0;
256 			start = 0;
257 			loop = true;
258 			goto again;
259 		}
260 		head = rb_entry(node, struct btrfs_delayed_ref_head,
261 				href_node);
262 	}
263 
264 	head->processing = 1;
265 	WARN_ON(delayed_refs->num_heads_ready == 0);
266 	delayed_refs->num_heads_ready--;
267 	delayed_refs->run_delayed_start = head->node.bytenr +
268 		head->node.num_bytes;
269 	return head;
270 }
271 
272 /*
273  * Helper to insert the ref_node to the tail or merge with tail.
274  *
275  * Return 0 for insert.
276  * Return >0 for merge.
277  */
278 static int
279 add_delayed_ref_tail_merge(struct btrfs_trans_handle *trans,
280 			   struct btrfs_delayed_ref_root *root,
281 			   struct btrfs_delayed_ref_head *href,
282 			   struct btrfs_delayed_ref_node *ref)
283 {
284 	struct btrfs_delayed_ref_node *exist;
285 	int mod;
286 	int ret = 0;
287 
288 	spin_lock(&href->lock);
289 	/* Check whether we can merge the tail node with ref */
290 	if (list_empty(&href->ref_list))
291 		goto add_tail;
292 	exist = list_entry(href->ref_list.prev, struct btrfs_delayed_ref_node,
293 			   list);
294 	/* No need to compare bytenr nor is_head */
295 	if (exist->type != ref->type || exist->no_quota != ref->no_quota ||
296 	    exist->seq != ref->seq)
297 		goto add_tail;
298 
299 	if ((exist->type == BTRFS_TREE_BLOCK_REF_KEY ||
300 	     exist->type == BTRFS_SHARED_BLOCK_REF_KEY) &&
301 	    comp_tree_refs(btrfs_delayed_node_to_tree_ref(exist),
302 			   btrfs_delayed_node_to_tree_ref(ref),
303 			   ref->type))
304 		goto add_tail;
305 	if ((exist->type == BTRFS_EXTENT_DATA_REF_KEY ||
306 	     exist->type == BTRFS_SHARED_DATA_REF_KEY) &&
307 	    comp_data_refs(btrfs_delayed_node_to_data_ref(exist),
308 			   btrfs_delayed_node_to_data_ref(ref)))
309 		goto add_tail;
310 
311 	/* Now we are sure we can merge */
312 	ret = 1;
313 	if (exist->action == ref->action) {
314 		mod = ref->ref_mod;
315 	} else {
316 		/* Need to change action */
317 		if (exist->ref_mod < ref->ref_mod) {
318 			exist->action = ref->action;
319 			mod = -exist->ref_mod;
320 			exist->ref_mod = ref->ref_mod;
321 		} else
322 			mod = -ref->ref_mod;
323 	}
324 	exist->ref_mod += mod;
325 
326 	/* remove existing tail if its ref_mod is zero */
327 	if (exist->ref_mod == 0)
328 		drop_delayed_ref(trans, root, href, exist);
329 	spin_unlock(&href->lock);
330 	return ret;
331 
332 add_tail:
333 	list_add_tail(&ref->list, &href->ref_list);
334 	atomic_inc(&root->num_entries);
335 	trans->delayed_ref_updates++;
336 	spin_unlock(&href->lock);
337 	return ret;
338 }
339 
340 /*
341  * helper function to update the accounting in the head ref
342  * existing and update must have the same bytenr
343  */
344 static noinline void
345 update_existing_head_ref(struct btrfs_delayed_ref_root *delayed_refs,
346 			 struct btrfs_delayed_ref_node *existing,
347 			 struct btrfs_delayed_ref_node *update)
348 {
349 	struct btrfs_delayed_ref_head *existing_ref;
350 	struct btrfs_delayed_ref_head *ref;
351 	int old_ref_mod;
352 
353 	existing_ref = btrfs_delayed_node_to_head(existing);
354 	ref = btrfs_delayed_node_to_head(update);
355 	BUG_ON(existing_ref->is_data != ref->is_data);
356 
357 	spin_lock(&existing_ref->lock);
358 	if (ref->must_insert_reserved) {
359 		/* if the extent was freed and then
360 		 * reallocated before the delayed ref
361 		 * entries were processed, we can end up
362 		 * with an existing head ref without
363 		 * the must_insert_reserved flag set.
364 		 * Set it again here
365 		 */
366 		existing_ref->must_insert_reserved = ref->must_insert_reserved;
367 
368 		/*
369 		 * update the num_bytes so we make sure the accounting
370 		 * is done correctly
371 		 */
372 		existing->num_bytes = update->num_bytes;
373 
374 	}
375 
376 	if (ref->extent_op) {
377 		if (!existing_ref->extent_op) {
378 			existing_ref->extent_op = ref->extent_op;
379 		} else {
380 			if (ref->extent_op->update_key) {
381 				memcpy(&existing_ref->extent_op->key,
382 				       &ref->extent_op->key,
383 				       sizeof(ref->extent_op->key));
384 				existing_ref->extent_op->update_key = 1;
385 			}
386 			if (ref->extent_op->update_flags) {
387 				existing_ref->extent_op->flags_to_set |=
388 					ref->extent_op->flags_to_set;
389 				existing_ref->extent_op->update_flags = 1;
390 			}
391 			btrfs_free_delayed_extent_op(ref->extent_op);
392 		}
393 	}
394 	/*
395 	 * update the reference mod on the head to reflect this new operation,
396 	 * only need the lock for this case cause we could be processing it
397 	 * currently, for refs we just added we know we're a-ok.
398 	 */
399 	old_ref_mod = existing_ref->total_ref_mod;
400 	existing->ref_mod += update->ref_mod;
401 	existing_ref->total_ref_mod += update->ref_mod;
402 
403 	/*
404 	 * If we are going to from a positive ref mod to a negative or vice
405 	 * versa we need to make sure to adjust pending_csums accordingly.
406 	 */
407 	if (existing_ref->is_data) {
408 		if (existing_ref->total_ref_mod >= 0 && old_ref_mod < 0)
409 			delayed_refs->pending_csums -= existing->num_bytes;
410 		if (existing_ref->total_ref_mod < 0 && old_ref_mod >= 0)
411 			delayed_refs->pending_csums += existing->num_bytes;
412 	}
413 	spin_unlock(&existing_ref->lock);
414 }
415 
416 /*
417  * helper function to actually insert a head node into the rbtree.
418  * this does all the dirty work in terms of maintaining the correct
419  * overall modification count.
420  */
421 static noinline struct btrfs_delayed_ref_head *
422 add_delayed_ref_head(struct btrfs_fs_info *fs_info,
423 		     struct btrfs_trans_handle *trans,
424 		     struct btrfs_delayed_ref_node *ref,
425 		     struct btrfs_qgroup_extent_record *qrecord,
426 		     u64 bytenr, u64 num_bytes, int action, int is_data)
427 {
428 	struct btrfs_delayed_ref_head *existing;
429 	struct btrfs_delayed_ref_head *head_ref = NULL;
430 	struct btrfs_delayed_ref_root *delayed_refs;
431 	struct btrfs_qgroup_extent_record *qexisting;
432 	int count_mod = 1;
433 	int must_insert_reserved = 0;
434 
435 	/*
436 	 * the head node stores the sum of all the mods, so dropping a ref
437 	 * should drop the sum in the head node by one.
438 	 */
439 	if (action == BTRFS_UPDATE_DELAYED_HEAD)
440 		count_mod = 0;
441 	else if (action == BTRFS_DROP_DELAYED_REF)
442 		count_mod = -1;
443 
444 	/*
445 	 * BTRFS_ADD_DELAYED_EXTENT means that we need to update
446 	 * the reserved accounting when the extent is finally added, or
447 	 * if a later modification deletes the delayed ref without ever
448 	 * inserting the extent into the extent allocation tree.
449 	 * ref->must_insert_reserved is the flag used to record
450 	 * that accounting mods are required.
451 	 *
452 	 * Once we record must_insert_reserved, switch the action to
453 	 * BTRFS_ADD_DELAYED_REF because other special casing is not required.
454 	 */
455 	if (action == BTRFS_ADD_DELAYED_EXTENT)
456 		must_insert_reserved = 1;
457 	else
458 		must_insert_reserved = 0;
459 
460 	delayed_refs = &trans->transaction->delayed_refs;
461 
462 	/* first set the basic ref node struct up */
463 	atomic_set(&ref->refs, 1);
464 	ref->bytenr = bytenr;
465 	ref->num_bytes = num_bytes;
466 	ref->ref_mod = count_mod;
467 	ref->type  = 0;
468 	ref->action  = 0;
469 	ref->is_head = 1;
470 	ref->in_tree = 1;
471 	ref->seq = 0;
472 
473 	head_ref = btrfs_delayed_node_to_head(ref);
474 	head_ref->must_insert_reserved = must_insert_reserved;
475 	head_ref->is_data = is_data;
476 	INIT_LIST_HEAD(&head_ref->ref_list);
477 	head_ref->processing = 0;
478 	head_ref->total_ref_mod = count_mod;
479 
480 	/* Record qgroup extent info if provided */
481 	if (qrecord) {
482 		qrecord->bytenr = bytenr;
483 		qrecord->num_bytes = num_bytes;
484 		qrecord->old_roots = NULL;
485 
486 		qexisting = btrfs_qgroup_insert_dirty_extent(delayed_refs,
487 							     qrecord);
488 		if (qexisting)
489 			kfree(qrecord);
490 	}
491 
492 	spin_lock_init(&head_ref->lock);
493 	mutex_init(&head_ref->mutex);
494 
495 	trace_add_delayed_ref_head(ref, head_ref, action);
496 
497 	existing = htree_insert(&delayed_refs->href_root,
498 				&head_ref->href_node);
499 	if (existing) {
500 		update_existing_head_ref(delayed_refs, &existing->node, ref);
501 		/*
502 		 * we've updated the existing ref, free the newly
503 		 * allocated ref
504 		 */
505 		kmem_cache_free(btrfs_delayed_ref_head_cachep, head_ref);
506 		head_ref = existing;
507 	} else {
508 		if (is_data && count_mod < 0)
509 			delayed_refs->pending_csums += num_bytes;
510 		delayed_refs->num_heads++;
511 		delayed_refs->num_heads_ready++;
512 		atomic_inc(&delayed_refs->num_entries);
513 		trans->delayed_ref_updates++;
514 	}
515 	return head_ref;
516 }
517 
518 /*
519  * helper to insert a delayed tree ref into the rbtree.
520  */
521 static noinline void
522 add_delayed_tree_ref(struct btrfs_fs_info *fs_info,
523 		     struct btrfs_trans_handle *trans,
524 		     struct btrfs_delayed_ref_head *head_ref,
525 		     struct btrfs_delayed_ref_node *ref, u64 bytenr,
526 		     u64 num_bytes, u64 parent, u64 ref_root, int level,
527 		     int action, int no_quota)
528 {
529 	struct btrfs_delayed_tree_ref *full_ref;
530 	struct btrfs_delayed_ref_root *delayed_refs;
531 	u64 seq = 0;
532 	int ret;
533 
534 	if (action == BTRFS_ADD_DELAYED_EXTENT)
535 		action = BTRFS_ADD_DELAYED_REF;
536 
537 	if (is_fstree(ref_root))
538 		seq = atomic64_read(&fs_info->tree_mod_seq);
539 	delayed_refs = &trans->transaction->delayed_refs;
540 
541 	/* first set the basic ref node struct up */
542 	atomic_set(&ref->refs, 1);
543 	ref->bytenr = bytenr;
544 	ref->num_bytes = num_bytes;
545 	ref->ref_mod = 1;
546 	ref->action = action;
547 	ref->is_head = 0;
548 	ref->in_tree = 1;
549 	ref->no_quota = no_quota;
550 	ref->seq = seq;
551 
552 	full_ref = btrfs_delayed_node_to_tree_ref(ref);
553 	full_ref->parent = parent;
554 	full_ref->root = ref_root;
555 	if (parent)
556 		ref->type = BTRFS_SHARED_BLOCK_REF_KEY;
557 	else
558 		ref->type = BTRFS_TREE_BLOCK_REF_KEY;
559 	full_ref->level = level;
560 
561 	trace_add_delayed_tree_ref(ref, full_ref, action);
562 
563 	ret = add_delayed_ref_tail_merge(trans, delayed_refs, head_ref, ref);
564 
565 	/*
566 	 * XXX: memory should be freed at the same level allocated.
567 	 * But bad practice is anywhere... Follow it now. Need cleanup.
568 	 */
569 	if (ret > 0)
570 		kmem_cache_free(btrfs_delayed_tree_ref_cachep, full_ref);
571 }
572 
573 /*
574  * helper to insert a delayed data ref into the rbtree.
575  */
576 static noinline void
577 add_delayed_data_ref(struct btrfs_fs_info *fs_info,
578 		     struct btrfs_trans_handle *trans,
579 		     struct btrfs_delayed_ref_head *head_ref,
580 		     struct btrfs_delayed_ref_node *ref, u64 bytenr,
581 		     u64 num_bytes, u64 parent, u64 ref_root, u64 owner,
582 		     u64 offset, int action, int no_quota)
583 {
584 	struct btrfs_delayed_data_ref *full_ref;
585 	struct btrfs_delayed_ref_root *delayed_refs;
586 	u64 seq = 0;
587 	int ret;
588 
589 	if (action == BTRFS_ADD_DELAYED_EXTENT)
590 		action = BTRFS_ADD_DELAYED_REF;
591 
592 	delayed_refs = &trans->transaction->delayed_refs;
593 
594 	if (is_fstree(ref_root))
595 		seq = atomic64_read(&fs_info->tree_mod_seq);
596 
597 	/* first set the basic ref node struct up */
598 	atomic_set(&ref->refs, 1);
599 	ref->bytenr = bytenr;
600 	ref->num_bytes = num_bytes;
601 	ref->ref_mod = 1;
602 	ref->action = action;
603 	ref->is_head = 0;
604 	ref->in_tree = 1;
605 	ref->no_quota = no_quota;
606 	ref->seq = seq;
607 
608 	full_ref = btrfs_delayed_node_to_data_ref(ref);
609 	full_ref->parent = parent;
610 	full_ref->root = ref_root;
611 	if (parent)
612 		ref->type = BTRFS_SHARED_DATA_REF_KEY;
613 	else
614 		ref->type = BTRFS_EXTENT_DATA_REF_KEY;
615 
616 	full_ref->objectid = owner;
617 	full_ref->offset = offset;
618 
619 	trace_add_delayed_data_ref(ref, full_ref, action);
620 
621 	ret = add_delayed_ref_tail_merge(trans, delayed_refs, head_ref, ref);
622 
623 	if (ret > 0)
624 		kmem_cache_free(btrfs_delayed_data_ref_cachep, full_ref);
625 }
626 
627 /*
628  * add a delayed tree ref.  This does all of the accounting required
629  * to make sure the delayed ref is eventually processed before this
630  * transaction commits.
631  */
632 int btrfs_add_delayed_tree_ref(struct btrfs_fs_info *fs_info,
633 			       struct btrfs_trans_handle *trans,
634 			       u64 bytenr, u64 num_bytes, u64 parent,
635 			       u64 ref_root,  int level, int action,
636 			       struct btrfs_delayed_extent_op *extent_op,
637 			       int no_quota)
638 {
639 	struct btrfs_delayed_tree_ref *ref;
640 	struct btrfs_delayed_ref_head *head_ref;
641 	struct btrfs_delayed_ref_root *delayed_refs;
642 	struct btrfs_qgroup_extent_record *record = NULL;
643 
644 	if (!is_fstree(ref_root) || !fs_info->quota_enabled)
645 		no_quota = 0;
646 
647 	BUG_ON(extent_op && extent_op->is_data);
648 	ref = kmem_cache_alloc(btrfs_delayed_tree_ref_cachep, GFP_NOFS);
649 	if (!ref)
650 		return -ENOMEM;
651 
652 	head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
653 	if (!head_ref)
654 		goto free_ref;
655 
656 	if (fs_info->quota_enabled && is_fstree(ref_root)) {
657 		record = kmalloc(sizeof(*record), GFP_NOFS);
658 		if (!record)
659 			goto free_head_ref;
660 	}
661 
662 	head_ref->extent_op = extent_op;
663 
664 	delayed_refs = &trans->transaction->delayed_refs;
665 	spin_lock(&delayed_refs->lock);
666 
667 	/*
668 	 * insert both the head node and the new ref without dropping
669 	 * the spin lock
670 	 */
671 	head_ref = add_delayed_ref_head(fs_info, trans, &head_ref->node, record,
672 					bytenr, num_bytes, action, 0);
673 
674 	add_delayed_tree_ref(fs_info, trans, head_ref, &ref->node, bytenr,
675 				   num_bytes, parent, ref_root, level, action,
676 				   no_quota);
677 	spin_unlock(&delayed_refs->lock);
678 
679 	return 0;
680 
681 free_head_ref:
682 	kmem_cache_free(btrfs_delayed_ref_head_cachep, head_ref);
683 free_ref:
684 	kmem_cache_free(btrfs_delayed_tree_ref_cachep, ref);
685 
686 	return -ENOMEM;
687 }
688 
689 /*
690  * add a delayed data ref. it's similar to btrfs_add_delayed_tree_ref.
691  */
692 int btrfs_add_delayed_data_ref(struct btrfs_fs_info *fs_info,
693 			       struct btrfs_trans_handle *trans,
694 			       u64 bytenr, u64 num_bytes,
695 			       u64 parent, u64 ref_root,
696 			       u64 owner, u64 offset, int action,
697 			       struct btrfs_delayed_extent_op *extent_op,
698 			       int no_quota)
699 {
700 	struct btrfs_delayed_data_ref *ref;
701 	struct btrfs_delayed_ref_head *head_ref;
702 	struct btrfs_delayed_ref_root *delayed_refs;
703 	struct btrfs_qgroup_extent_record *record = NULL;
704 
705 	if (!is_fstree(ref_root) || !fs_info->quota_enabled)
706 		no_quota = 0;
707 
708 	BUG_ON(extent_op && !extent_op->is_data);
709 	ref = kmem_cache_alloc(btrfs_delayed_data_ref_cachep, GFP_NOFS);
710 	if (!ref)
711 		return -ENOMEM;
712 
713 	head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
714 	if (!head_ref) {
715 		kmem_cache_free(btrfs_delayed_data_ref_cachep, ref);
716 		return -ENOMEM;
717 	}
718 
719 	if (fs_info->quota_enabled && is_fstree(ref_root)) {
720 		record = kmalloc(sizeof(*record), GFP_NOFS);
721 		if (!record) {
722 			kmem_cache_free(btrfs_delayed_data_ref_cachep, ref);
723 			kmem_cache_free(btrfs_delayed_ref_head_cachep,
724 					head_ref);
725 			return -ENOMEM;
726 		}
727 	}
728 
729 	head_ref->extent_op = extent_op;
730 
731 	delayed_refs = &trans->transaction->delayed_refs;
732 	spin_lock(&delayed_refs->lock);
733 
734 	/*
735 	 * insert both the head node and the new ref without dropping
736 	 * the spin lock
737 	 */
738 	head_ref = add_delayed_ref_head(fs_info, trans, &head_ref->node, record,
739 					bytenr, num_bytes, action, 1);
740 
741 	add_delayed_data_ref(fs_info, trans, head_ref, &ref->node, bytenr,
742 				   num_bytes, parent, ref_root, owner, offset,
743 				   action, no_quota);
744 	spin_unlock(&delayed_refs->lock);
745 
746 	return 0;
747 }
748 
749 int btrfs_add_delayed_extent_op(struct btrfs_fs_info *fs_info,
750 				struct btrfs_trans_handle *trans,
751 				u64 bytenr, u64 num_bytes,
752 				struct btrfs_delayed_extent_op *extent_op)
753 {
754 	struct btrfs_delayed_ref_head *head_ref;
755 	struct btrfs_delayed_ref_root *delayed_refs;
756 
757 	head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
758 	if (!head_ref)
759 		return -ENOMEM;
760 
761 	head_ref->extent_op = extent_op;
762 
763 	delayed_refs = &trans->transaction->delayed_refs;
764 	spin_lock(&delayed_refs->lock);
765 
766 	add_delayed_ref_head(fs_info, trans, &head_ref->node, NULL, bytenr,
767 			     num_bytes, BTRFS_UPDATE_DELAYED_HEAD,
768 			     extent_op->is_data);
769 
770 	spin_unlock(&delayed_refs->lock);
771 	return 0;
772 }
773 
774 /*
775  * this does a simple search for the head node for a given extent.
776  * It must be called with the delayed ref spinlock held, and it returns
777  * the head node if any where found, or NULL if not.
778  */
779 struct btrfs_delayed_ref_head *
780 btrfs_find_delayed_ref_head(struct btrfs_trans_handle *trans, u64 bytenr)
781 {
782 	struct btrfs_delayed_ref_root *delayed_refs;
783 
784 	delayed_refs = &trans->transaction->delayed_refs;
785 	return find_ref_head(&delayed_refs->href_root, bytenr, 0);
786 }
787 
788 void btrfs_delayed_ref_exit(void)
789 {
790 	if (btrfs_delayed_ref_head_cachep)
791 		kmem_cache_destroy(btrfs_delayed_ref_head_cachep);
792 	if (btrfs_delayed_tree_ref_cachep)
793 		kmem_cache_destroy(btrfs_delayed_tree_ref_cachep);
794 	if (btrfs_delayed_data_ref_cachep)
795 		kmem_cache_destroy(btrfs_delayed_data_ref_cachep);
796 	if (btrfs_delayed_extent_op_cachep)
797 		kmem_cache_destroy(btrfs_delayed_extent_op_cachep);
798 }
799 
800 int btrfs_delayed_ref_init(void)
801 {
802 	btrfs_delayed_ref_head_cachep = kmem_cache_create(
803 				"btrfs_delayed_ref_head",
804 				sizeof(struct btrfs_delayed_ref_head), 0,
805 				SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
806 	if (!btrfs_delayed_ref_head_cachep)
807 		goto fail;
808 
809 	btrfs_delayed_tree_ref_cachep = kmem_cache_create(
810 				"btrfs_delayed_tree_ref",
811 				sizeof(struct btrfs_delayed_tree_ref), 0,
812 				SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
813 	if (!btrfs_delayed_tree_ref_cachep)
814 		goto fail;
815 
816 	btrfs_delayed_data_ref_cachep = kmem_cache_create(
817 				"btrfs_delayed_data_ref",
818 				sizeof(struct btrfs_delayed_data_ref), 0,
819 				SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
820 	if (!btrfs_delayed_data_ref_cachep)
821 		goto fail;
822 
823 	btrfs_delayed_extent_op_cachep = kmem_cache_create(
824 				"btrfs_delayed_extent_op",
825 				sizeof(struct btrfs_delayed_extent_op), 0,
826 				SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
827 	if (!btrfs_delayed_extent_op_cachep)
828 		goto fail;
829 
830 	return 0;
831 fail:
832 	btrfs_delayed_ref_exit();
833 	return -ENOMEM;
834 }
835