xref: /openbmc/linux/fs/btrfs/delayed-ref.c (revision 31b90347)
1 /*
2  * Copyright (C) 2009 Oracle.  All rights reserved.
3  *
4  * This program is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU General Public
6  * License v2 as published by the Free Software Foundation.
7  *
8  * This program is distributed in the hope that it will be useful,
9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
11  * General Public License for more details.
12  *
13  * You should have received a copy of the GNU General Public
14  * License along with this program; if not, write to the
15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16  * Boston, MA 021110-1307, USA.
17  */
18 
19 #include <linux/sched.h>
20 #include <linux/slab.h>
21 #include <linux/sort.h>
22 #include "ctree.h"
23 #include "delayed-ref.h"
24 #include "transaction.h"
25 
26 struct kmem_cache *btrfs_delayed_ref_head_cachep;
27 struct kmem_cache *btrfs_delayed_tree_ref_cachep;
28 struct kmem_cache *btrfs_delayed_data_ref_cachep;
29 struct kmem_cache *btrfs_delayed_extent_op_cachep;
30 /*
31  * delayed back reference update tracking.  For subvolume trees
32  * we queue up extent allocations and backref maintenance for
33  * delayed processing.   This avoids deep call chains where we
34  * add extents in the middle of btrfs_search_slot, and it allows
35  * us to buffer up frequently modified backrefs in an rb tree instead
36  * of hammering updates on the extent allocation tree.
37  */
38 
39 /*
40  * compare two delayed tree backrefs with same bytenr and type
41  */
42 static int comp_tree_refs(struct btrfs_delayed_tree_ref *ref2,
43 			  struct btrfs_delayed_tree_ref *ref1, int type)
44 {
45 	if (type == BTRFS_TREE_BLOCK_REF_KEY) {
46 		if (ref1->root < ref2->root)
47 			return -1;
48 		if (ref1->root > ref2->root)
49 			return 1;
50 	} else {
51 		if (ref1->parent < ref2->parent)
52 			return -1;
53 		if (ref1->parent > ref2->parent)
54 			return 1;
55 	}
56 	return 0;
57 }
58 
59 /*
60  * compare two delayed data backrefs with same bytenr and type
61  */
62 static int comp_data_refs(struct btrfs_delayed_data_ref *ref2,
63 			  struct btrfs_delayed_data_ref *ref1)
64 {
65 	if (ref1->node.type == BTRFS_EXTENT_DATA_REF_KEY) {
66 		if (ref1->root < ref2->root)
67 			return -1;
68 		if (ref1->root > ref2->root)
69 			return 1;
70 		if (ref1->objectid < ref2->objectid)
71 			return -1;
72 		if (ref1->objectid > ref2->objectid)
73 			return 1;
74 		if (ref1->offset < ref2->offset)
75 			return -1;
76 		if (ref1->offset > ref2->offset)
77 			return 1;
78 	} else {
79 		if (ref1->parent < ref2->parent)
80 			return -1;
81 		if (ref1->parent > ref2->parent)
82 			return 1;
83 	}
84 	return 0;
85 }
86 
87 /*
88  * entries in the rb tree are ordered by the byte number of the extent,
89  * type of the delayed backrefs and content of delayed backrefs.
90  */
91 static int comp_entry(struct btrfs_delayed_ref_node *ref2,
92 		      struct btrfs_delayed_ref_node *ref1,
93 		      bool compare_seq)
94 {
95 	if (ref1->bytenr < ref2->bytenr)
96 		return -1;
97 	if (ref1->bytenr > ref2->bytenr)
98 		return 1;
99 	if (ref1->is_head && ref2->is_head)
100 		return 0;
101 	if (ref2->is_head)
102 		return -1;
103 	if (ref1->is_head)
104 		return 1;
105 	if (ref1->type < ref2->type)
106 		return -1;
107 	if (ref1->type > ref2->type)
108 		return 1;
109 	/* merging of sequenced refs is not allowed */
110 	if (compare_seq) {
111 		if (ref1->seq < ref2->seq)
112 			return -1;
113 		if (ref1->seq > ref2->seq)
114 			return 1;
115 	}
116 	if (ref1->type == BTRFS_TREE_BLOCK_REF_KEY ||
117 	    ref1->type == BTRFS_SHARED_BLOCK_REF_KEY) {
118 		return comp_tree_refs(btrfs_delayed_node_to_tree_ref(ref2),
119 				      btrfs_delayed_node_to_tree_ref(ref1),
120 				      ref1->type);
121 	} else if (ref1->type == BTRFS_EXTENT_DATA_REF_KEY ||
122 		   ref1->type == BTRFS_SHARED_DATA_REF_KEY) {
123 		return comp_data_refs(btrfs_delayed_node_to_data_ref(ref2),
124 				      btrfs_delayed_node_to_data_ref(ref1));
125 	}
126 	BUG();
127 	return 0;
128 }
129 
130 /*
131  * insert a new ref into the rbtree.  This returns any existing refs
132  * for the same (bytenr,parent) tuple, or NULL if the new node was properly
133  * inserted.
134  */
135 static struct btrfs_delayed_ref_node *tree_insert(struct rb_root *root,
136 						  struct rb_node *node)
137 {
138 	struct rb_node **p = &root->rb_node;
139 	struct rb_node *parent_node = NULL;
140 	struct btrfs_delayed_ref_node *entry;
141 	struct btrfs_delayed_ref_node *ins;
142 	int cmp;
143 
144 	ins = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
145 	while (*p) {
146 		parent_node = *p;
147 		entry = rb_entry(parent_node, struct btrfs_delayed_ref_node,
148 				 rb_node);
149 
150 		cmp = comp_entry(entry, ins, 1);
151 		if (cmp < 0)
152 			p = &(*p)->rb_left;
153 		else if (cmp > 0)
154 			p = &(*p)->rb_right;
155 		else
156 			return entry;
157 	}
158 
159 	rb_link_node(node, parent_node, p);
160 	rb_insert_color(node, root);
161 	return NULL;
162 }
163 
164 /*
165  * find an head entry based on bytenr. This returns the delayed ref
166  * head if it was able to find one, or NULL if nothing was in that spot.
167  * If return_bigger is given, the next bigger entry is returned if no exact
168  * match is found.
169  */
170 static struct btrfs_delayed_ref_node *find_ref_head(struct rb_root *root,
171 				  u64 bytenr,
172 				  struct btrfs_delayed_ref_node **last,
173 				  int return_bigger)
174 {
175 	struct rb_node *n;
176 	struct btrfs_delayed_ref_node *entry;
177 	int cmp = 0;
178 
179 again:
180 	n = root->rb_node;
181 	entry = NULL;
182 	while (n) {
183 		entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
184 		WARN_ON(!entry->in_tree);
185 		if (last)
186 			*last = entry;
187 
188 		if (bytenr < entry->bytenr)
189 			cmp = -1;
190 		else if (bytenr > entry->bytenr)
191 			cmp = 1;
192 		else if (!btrfs_delayed_ref_is_head(entry))
193 			cmp = 1;
194 		else
195 			cmp = 0;
196 
197 		if (cmp < 0)
198 			n = n->rb_left;
199 		else if (cmp > 0)
200 			n = n->rb_right;
201 		else
202 			return entry;
203 	}
204 	if (entry && return_bigger) {
205 		if (cmp > 0) {
206 			n = rb_next(&entry->rb_node);
207 			if (!n)
208 				n = rb_first(root);
209 			entry = rb_entry(n, struct btrfs_delayed_ref_node,
210 					 rb_node);
211 			bytenr = entry->bytenr;
212 			return_bigger = 0;
213 			goto again;
214 		}
215 		return entry;
216 	}
217 	return NULL;
218 }
219 
220 int btrfs_delayed_ref_lock(struct btrfs_trans_handle *trans,
221 			   struct btrfs_delayed_ref_head *head)
222 {
223 	struct btrfs_delayed_ref_root *delayed_refs;
224 
225 	delayed_refs = &trans->transaction->delayed_refs;
226 	assert_spin_locked(&delayed_refs->lock);
227 	if (mutex_trylock(&head->mutex))
228 		return 0;
229 
230 	atomic_inc(&head->node.refs);
231 	spin_unlock(&delayed_refs->lock);
232 
233 	mutex_lock(&head->mutex);
234 	spin_lock(&delayed_refs->lock);
235 	if (!head->node.in_tree) {
236 		mutex_unlock(&head->mutex);
237 		btrfs_put_delayed_ref(&head->node);
238 		return -EAGAIN;
239 	}
240 	btrfs_put_delayed_ref(&head->node);
241 	return 0;
242 }
243 
244 static inline void drop_delayed_ref(struct btrfs_trans_handle *trans,
245 				    struct btrfs_delayed_ref_root *delayed_refs,
246 				    struct btrfs_delayed_ref_node *ref)
247 {
248 	rb_erase(&ref->rb_node, &delayed_refs->root);
249 	ref->in_tree = 0;
250 	btrfs_put_delayed_ref(ref);
251 	delayed_refs->num_entries--;
252 	if (trans->delayed_ref_updates)
253 		trans->delayed_ref_updates--;
254 }
255 
256 static int merge_ref(struct btrfs_trans_handle *trans,
257 		     struct btrfs_delayed_ref_root *delayed_refs,
258 		     struct btrfs_delayed_ref_node *ref, u64 seq)
259 {
260 	struct rb_node *node;
261 	int merged = 0;
262 	int mod = 0;
263 	int done = 0;
264 
265 	node = rb_prev(&ref->rb_node);
266 	while (node) {
267 		struct btrfs_delayed_ref_node *next;
268 
269 		next = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
270 		node = rb_prev(node);
271 		if (next->bytenr != ref->bytenr)
272 			break;
273 		if (seq && next->seq >= seq)
274 			break;
275 		if (comp_entry(ref, next, 0))
276 			continue;
277 
278 		if (ref->action == next->action) {
279 			mod = next->ref_mod;
280 		} else {
281 			if (ref->ref_mod < next->ref_mod) {
282 				struct btrfs_delayed_ref_node *tmp;
283 
284 				tmp = ref;
285 				ref = next;
286 				next = tmp;
287 				done = 1;
288 			}
289 			mod = -next->ref_mod;
290 		}
291 
292 		merged++;
293 		drop_delayed_ref(trans, delayed_refs, next);
294 		ref->ref_mod += mod;
295 		if (ref->ref_mod == 0) {
296 			drop_delayed_ref(trans, delayed_refs, ref);
297 			break;
298 		} else {
299 			/*
300 			 * You can't have multiples of the same ref on a tree
301 			 * block.
302 			 */
303 			WARN_ON(ref->type == BTRFS_TREE_BLOCK_REF_KEY ||
304 				ref->type == BTRFS_SHARED_BLOCK_REF_KEY);
305 		}
306 
307 		if (done)
308 			break;
309 		node = rb_prev(&ref->rb_node);
310 	}
311 
312 	return merged;
313 }
314 
315 void btrfs_merge_delayed_refs(struct btrfs_trans_handle *trans,
316 			      struct btrfs_fs_info *fs_info,
317 			      struct btrfs_delayed_ref_root *delayed_refs,
318 			      struct btrfs_delayed_ref_head *head)
319 {
320 	struct rb_node *node;
321 	u64 seq = 0;
322 
323 	spin_lock(&fs_info->tree_mod_seq_lock);
324 	if (!list_empty(&fs_info->tree_mod_seq_list)) {
325 		struct seq_list *elem;
326 
327 		elem = list_first_entry(&fs_info->tree_mod_seq_list,
328 					struct seq_list, list);
329 		seq = elem->seq;
330 	}
331 	spin_unlock(&fs_info->tree_mod_seq_lock);
332 
333 	node = rb_prev(&head->node.rb_node);
334 	while (node) {
335 		struct btrfs_delayed_ref_node *ref;
336 
337 		ref = rb_entry(node, struct btrfs_delayed_ref_node,
338 			       rb_node);
339 		if (ref->bytenr != head->node.bytenr)
340 			break;
341 
342 		/* We can't merge refs that are outside of our seq count */
343 		if (seq && ref->seq >= seq)
344 			break;
345 		if (merge_ref(trans, delayed_refs, ref, seq))
346 			node = rb_prev(&head->node.rb_node);
347 		else
348 			node = rb_prev(node);
349 	}
350 }
351 
352 int btrfs_check_delayed_seq(struct btrfs_fs_info *fs_info,
353 			    struct btrfs_delayed_ref_root *delayed_refs,
354 			    u64 seq)
355 {
356 	struct seq_list *elem;
357 	int ret = 0;
358 
359 	spin_lock(&fs_info->tree_mod_seq_lock);
360 	if (!list_empty(&fs_info->tree_mod_seq_list)) {
361 		elem = list_first_entry(&fs_info->tree_mod_seq_list,
362 					struct seq_list, list);
363 		if (seq >= elem->seq) {
364 			pr_debug("holding back delayed_ref %#x.%x, lowest is %#x.%x (%p)\n",
365 				 (u32)(seq >> 32), (u32)seq,
366 				 (u32)(elem->seq >> 32), (u32)elem->seq,
367 				 delayed_refs);
368 			ret = 1;
369 		}
370 	}
371 
372 	spin_unlock(&fs_info->tree_mod_seq_lock);
373 	return ret;
374 }
375 
376 int btrfs_find_ref_cluster(struct btrfs_trans_handle *trans,
377 			   struct list_head *cluster, u64 start)
378 {
379 	int count = 0;
380 	struct btrfs_delayed_ref_root *delayed_refs;
381 	struct rb_node *node;
382 	struct btrfs_delayed_ref_node *ref;
383 	struct btrfs_delayed_ref_head *head;
384 
385 	delayed_refs = &trans->transaction->delayed_refs;
386 	if (start == 0) {
387 		node = rb_first(&delayed_refs->root);
388 	} else {
389 		ref = NULL;
390 		find_ref_head(&delayed_refs->root, start + 1, &ref, 1);
391 		if (ref) {
392 			node = &ref->rb_node;
393 		} else
394 			node = rb_first(&delayed_refs->root);
395 	}
396 again:
397 	while (node && count < 32) {
398 		ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
399 		if (btrfs_delayed_ref_is_head(ref)) {
400 			head = btrfs_delayed_node_to_head(ref);
401 			if (list_empty(&head->cluster)) {
402 				list_add_tail(&head->cluster, cluster);
403 				delayed_refs->run_delayed_start =
404 					head->node.bytenr;
405 				count++;
406 
407 				WARN_ON(delayed_refs->num_heads_ready == 0);
408 				delayed_refs->num_heads_ready--;
409 			} else if (count) {
410 				/* the goal of the clustering is to find extents
411 				 * that are likely to end up in the same extent
412 				 * leaf on disk.  So, we don't want them spread
413 				 * all over the tree.  Stop now if we've hit
414 				 * a head that was already in use
415 				 */
416 				break;
417 			}
418 		}
419 		node = rb_next(node);
420 	}
421 	if (count) {
422 		return 0;
423 	} else if (start) {
424 		/*
425 		 * we've gone to the end of the rbtree without finding any
426 		 * clusters.  start from the beginning and try again
427 		 */
428 		start = 0;
429 		node = rb_first(&delayed_refs->root);
430 		goto again;
431 	}
432 	return 1;
433 }
434 
435 void btrfs_release_ref_cluster(struct list_head *cluster)
436 {
437 	struct list_head *pos, *q;
438 
439 	list_for_each_safe(pos, q, cluster)
440 		list_del_init(pos);
441 }
442 
443 /*
444  * helper function to update an extent delayed ref in the
445  * rbtree.  existing and update must both have the same
446  * bytenr and parent
447  *
448  * This may free existing if the update cancels out whatever
449  * operation it was doing.
450  */
451 static noinline void
452 update_existing_ref(struct btrfs_trans_handle *trans,
453 		    struct btrfs_delayed_ref_root *delayed_refs,
454 		    struct btrfs_delayed_ref_node *existing,
455 		    struct btrfs_delayed_ref_node *update)
456 {
457 	if (update->action != existing->action) {
458 		/*
459 		 * this is effectively undoing either an add or a
460 		 * drop.  We decrement the ref_mod, and if it goes
461 		 * down to zero we just delete the entry without
462 		 * every changing the extent allocation tree.
463 		 */
464 		existing->ref_mod--;
465 		if (existing->ref_mod == 0)
466 			drop_delayed_ref(trans, delayed_refs, existing);
467 		else
468 			WARN_ON(existing->type == BTRFS_TREE_BLOCK_REF_KEY ||
469 				existing->type == BTRFS_SHARED_BLOCK_REF_KEY);
470 	} else {
471 		WARN_ON(existing->type == BTRFS_TREE_BLOCK_REF_KEY ||
472 			existing->type == BTRFS_SHARED_BLOCK_REF_KEY);
473 		/*
474 		 * the action on the existing ref matches
475 		 * the action on the ref we're trying to add.
476 		 * Bump the ref_mod by one so the backref that
477 		 * is eventually added/removed has the correct
478 		 * reference count
479 		 */
480 		existing->ref_mod += update->ref_mod;
481 	}
482 }
483 
484 /*
485  * helper function to update the accounting in the head ref
486  * existing and update must have the same bytenr
487  */
488 static noinline void
489 update_existing_head_ref(struct btrfs_delayed_ref_node *existing,
490 			 struct btrfs_delayed_ref_node *update)
491 {
492 	struct btrfs_delayed_ref_head *existing_ref;
493 	struct btrfs_delayed_ref_head *ref;
494 
495 	existing_ref = btrfs_delayed_node_to_head(existing);
496 	ref = btrfs_delayed_node_to_head(update);
497 	BUG_ON(existing_ref->is_data != ref->is_data);
498 
499 	if (ref->must_insert_reserved) {
500 		/* if the extent was freed and then
501 		 * reallocated before the delayed ref
502 		 * entries were processed, we can end up
503 		 * with an existing head ref without
504 		 * the must_insert_reserved flag set.
505 		 * Set it again here
506 		 */
507 		existing_ref->must_insert_reserved = ref->must_insert_reserved;
508 
509 		/*
510 		 * update the num_bytes so we make sure the accounting
511 		 * is done correctly
512 		 */
513 		existing->num_bytes = update->num_bytes;
514 
515 	}
516 
517 	if (ref->extent_op) {
518 		if (!existing_ref->extent_op) {
519 			existing_ref->extent_op = ref->extent_op;
520 		} else {
521 			if (ref->extent_op->update_key) {
522 				memcpy(&existing_ref->extent_op->key,
523 				       &ref->extent_op->key,
524 				       sizeof(ref->extent_op->key));
525 				existing_ref->extent_op->update_key = 1;
526 			}
527 			if (ref->extent_op->update_flags) {
528 				existing_ref->extent_op->flags_to_set |=
529 					ref->extent_op->flags_to_set;
530 				existing_ref->extent_op->update_flags = 1;
531 			}
532 			btrfs_free_delayed_extent_op(ref->extent_op);
533 		}
534 	}
535 	/*
536 	 * update the reference mod on the head to reflect this new operation
537 	 */
538 	existing->ref_mod += update->ref_mod;
539 }
540 
541 /*
542  * helper function to actually insert a head node into the rbtree.
543  * this does all the dirty work in terms of maintaining the correct
544  * overall modification count.
545  */
546 static noinline void add_delayed_ref_head(struct btrfs_fs_info *fs_info,
547 					struct btrfs_trans_handle *trans,
548 					struct btrfs_delayed_ref_node *ref,
549 					u64 bytenr, u64 num_bytes,
550 					int action, int is_data)
551 {
552 	struct btrfs_delayed_ref_node *existing;
553 	struct btrfs_delayed_ref_head *head_ref = NULL;
554 	struct btrfs_delayed_ref_root *delayed_refs;
555 	int count_mod = 1;
556 	int must_insert_reserved = 0;
557 
558 	/*
559 	 * the head node stores the sum of all the mods, so dropping a ref
560 	 * should drop the sum in the head node by one.
561 	 */
562 	if (action == BTRFS_UPDATE_DELAYED_HEAD)
563 		count_mod = 0;
564 	else if (action == BTRFS_DROP_DELAYED_REF)
565 		count_mod = -1;
566 
567 	/*
568 	 * BTRFS_ADD_DELAYED_EXTENT means that we need to update
569 	 * the reserved accounting when the extent is finally added, or
570 	 * if a later modification deletes the delayed ref without ever
571 	 * inserting the extent into the extent allocation tree.
572 	 * ref->must_insert_reserved is the flag used to record
573 	 * that accounting mods are required.
574 	 *
575 	 * Once we record must_insert_reserved, switch the action to
576 	 * BTRFS_ADD_DELAYED_REF because other special casing is not required.
577 	 */
578 	if (action == BTRFS_ADD_DELAYED_EXTENT)
579 		must_insert_reserved = 1;
580 	else
581 		must_insert_reserved = 0;
582 
583 	delayed_refs = &trans->transaction->delayed_refs;
584 
585 	/* first set the basic ref node struct up */
586 	atomic_set(&ref->refs, 1);
587 	ref->bytenr = bytenr;
588 	ref->num_bytes = num_bytes;
589 	ref->ref_mod = count_mod;
590 	ref->type  = 0;
591 	ref->action  = 0;
592 	ref->is_head = 1;
593 	ref->in_tree = 1;
594 	ref->seq = 0;
595 
596 	head_ref = btrfs_delayed_node_to_head(ref);
597 	head_ref->must_insert_reserved = must_insert_reserved;
598 	head_ref->is_data = is_data;
599 
600 	INIT_LIST_HEAD(&head_ref->cluster);
601 	mutex_init(&head_ref->mutex);
602 
603 	trace_add_delayed_ref_head(ref, head_ref, action);
604 
605 	existing = tree_insert(&delayed_refs->root, &ref->rb_node);
606 
607 	if (existing) {
608 		update_existing_head_ref(existing, ref);
609 		/*
610 		 * we've updated the existing ref, free the newly
611 		 * allocated ref
612 		 */
613 		kmem_cache_free(btrfs_delayed_ref_head_cachep, head_ref);
614 	} else {
615 		delayed_refs->num_heads++;
616 		delayed_refs->num_heads_ready++;
617 		delayed_refs->num_entries++;
618 		trans->delayed_ref_updates++;
619 	}
620 }
621 
622 /*
623  * helper to insert a delayed tree ref into the rbtree.
624  */
625 static noinline void add_delayed_tree_ref(struct btrfs_fs_info *fs_info,
626 					 struct btrfs_trans_handle *trans,
627 					 struct btrfs_delayed_ref_node *ref,
628 					 u64 bytenr, u64 num_bytes, u64 parent,
629 					 u64 ref_root, int level, int action,
630 					 int for_cow)
631 {
632 	struct btrfs_delayed_ref_node *existing;
633 	struct btrfs_delayed_tree_ref *full_ref;
634 	struct btrfs_delayed_ref_root *delayed_refs;
635 	u64 seq = 0;
636 
637 	if (action == BTRFS_ADD_DELAYED_EXTENT)
638 		action = BTRFS_ADD_DELAYED_REF;
639 
640 	delayed_refs = &trans->transaction->delayed_refs;
641 
642 	/* first set the basic ref node struct up */
643 	atomic_set(&ref->refs, 1);
644 	ref->bytenr = bytenr;
645 	ref->num_bytes = num_bytes;
646 	ref->ref_mod = 1;
647 	ref->action = action;
648 	ref->is_head = 0;
649 	ref->in_tree = 1;
650 
651 	if (need_ref_seq(for_cow, ref_root))
652 		seq = btrfs_get_tree_mod_seq(fs_info, &trans->delayed_ref_elem);
653 	ref->seq = seq;
654 
655 	full_ref = btrfs_delayed_node_to_tree_ref(ref);
656 	full_ref->parent = parent;
657 	full_ref->root = ref_root;
658 	if (parent)
659 		ref->type = BTRFS_SHARED_BLOCK_REF_KEY;
660 	else
661 		ref->type = BTRFS_TREE_BLOCK_REF_KEY;
662 	full_ref->level = level;
663 
664 	trace_add_delayed_tree_ref(ref, full_ref, action);
665 
666 	existing = tree_insert(&delayed_refs->root, &ref->rb_node);
667 
668 	if (existing) {
669 		update_existing_ref(trans, delayed_refs, existing, ref);
670 		/*
671 		 * we've updated the existing ref, free the newly
672 		 * allocated ref
673 		 */
674 		kmem_cache_free(btrfs_delayed_tree_ref_cachep, full_ref);
675 	} else {
676 		delayed_refs->num_entries++;
677 		trans->delayed_ref_updates++;
678 	}
679 }
680 
681 /*
682  * helper to insert a delayed data ref into the rbtree.
683  */
684 static noinline void add_delayed_data_ref(struct btrfs_fs_info *fs_info,
685 					 struct btrfs_trans_handle *trans,
686 					 struct btrfs_delayed_ref_node *ref,
687 					 u64 bytenr, u64 num_bytes, u64 parent,
688 					 u64 ref_root, u64 owner, u64 offset,
689 					 int action, int for_cow)
690 {
691 	struct btrfs_delayed_ref_node *existing;
692 	struct btrfs_delayed_data_ref *full_ref;
693 	struct btrfs_delayed_ref_root *delayed_refs;
694 	u64 seq = 0;
695 
696 	if (action == BTRFS_ADD_DELAYED_EXTENT)
697 		action = BTRFS_ADD_DELAYED_REF;
698 
699 	delayed_refs = &trans->transaction->delayed_refs;
700 
701 	/* first set the basic ref node struct up */
702 	atomic_set(&ref->refs, 1);
703 	ref->bytenr = bytenr;
704 	ref->num_bytes = num_bytes;
705 	ref->ref_mod = 1;
706 	ref->action = action;
707 	ref->is_head = 0;
708 	ref->in_tree = 1;
709 
710 	if (need_ref_seq(for_cow, ref_root))
711 		seq = btrfs_get_tree_mod_seq(fs_info, &trans->delayed_ref_elem);
712 	ref->seq = seq;
713 
714 	full_ref = btrfs_delayed_node_to_data_ref(ref);
715 	full_ref->parent = parent;
716 	full_ref->root = ref_root;
717 	if (parent)
718 		ref->type = BTRFS_SHARED_DATA_REF_KEY;
719 	else
720 		ref->type = BTRFS_EXTENT_DATA_REF_KEY;
721 
722 	full_ref->objectid = owner;
723 	full_ref->offset = offset;
724 
725 	trace_add_delayed_data_ref(ref, full_ref, action);
726 
727 	existing = tree_insert(&delayed_refs->root, &ref->rb_node);
728 
729 	if (existing) {
730 		update_existing_ref(trans, delayed_refs, existing, ref);
731 		/*
732 		 * we've updated the existing ref, free the newly
733 		 * allocated ref
734 		 */
735 		kmem_cache_free(btrfs_delayed_data_ref_cachep, full_ref);
736 	} else {
737 		delayed_refs->num_entries++;
738 		trans->delayed_ref_updates++;
739 	}
740 }
741 
742 /*
743  * add a delayed tree ref.  This does all of the accounting required
744  * to make sure the delayed ref is eventually processed before this
745  * transaction commits.
746  */
747 int btrfs_add_delayed_tree_ref(struct btrfs_fs_info *fs_info,
748 			       struct btrfs_trans_handle *trans,
749 			       u64 bytenr, u64 num_bytes, u64 parent,
750 			       u64 ref_root,  int level, int action,
751 			       struct btrfs_delayed_extent_op *extent_op,
752 			       int for_cow)
753 {
754 	struct btrfs_delayed_tree_ref *ref;
755 	struct btrfs_delayed_ref_head *head_ref;
756 	struct btrfs_delayed_ref_root *delayed_refs;
757 
758 	BUG_ON(extent_op && extent_op->is_data);
759 	ref = kmem_cache_alloc(btrfs_delayed_tree_ref_cachep, GFP_NOFS);
760 	if (!ref)
761 		return -ENOMEM;
762 
763 	head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
764 	if (!head_ref) {
765 		kmem_cache_free(btrfs_delayed_tree_ref_cachep, ref);
766 		return -ENOMEM;
767 	}
768 
769 	head_ref->extent_op = extent_op;
770 
771 	delayed_refs = &trans->transaction->delayed_refs;
772 	spin_lock(&delayed_refs->lock);
773 
774 	/*
775 	 * insert both the head node and the new ref without dropping
776 	 * the spin lock
777 	 */
778 	add_delayed_ref_head(fs_info, trans, &head_ref->node, bytenr,
779 				   num_bytes, action, 0);
780 
781 	add_delayed_tree_ref(fs_info, trans, &ref->node, bytenr,
782 				   num_bytes, parent, ref_root, level, action,
783 				   for_cow);
784 	spin_unlock(&delayed_refs->lock);
785 	if (need_ref_seq(for_cow, ref_root))
786 		btrfs_qgroup_record_ref(trans, &ref->node, extent_op);
787 
788 	return 0;
789 }
790 
791 /*
792  * add a delayed data ref. it's similar to btrfs_add_delayed_tree_ref.
793  */
794 int btrfs_add_delayed_data_ref(struct btrfs_fs_info *fs_info,
795 			       struct btrfs_trans_handle *trans,
796 			       u64 bytenr, u64 num_bytes,
797 			       u64 parent, u64 ref_root,
798 			       u64 owner, u64 offset, int action,
799 			       struct btrfs_delayed_extent_op *extent_op,
800 			       int for_cow)
801 {
802 	struct btrfs_delayed_data_ref *ref;
803 	struct btrfs_delayed_ref_head *head_ref;
804 	struct btrfs_delayed_ref_root *delayed_refs;
805 
806 	BUG_ON(extent_op && !extent_op->is_data);
807 	ref = kmem_cache_alloc(btrfs_delayed_data_ref_cachep, GFP_NOFS);
808 	if (!ref)
809 		return -ENOMEM;
810 
811 	head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
812 	if (!head_ref) {
813 		kmem_cache_free(btrfs_delayed_data_ref_cachep, ref);
814 		return -ENOMEM;
815 	}
816 
817 	head_ref->extent_op = extent_op;
818 
819 	delayed_refs = &trans->transaction->delayed_refs;
820 	spin_lock(&delayed_refs->lock);
821 
822 	/*
823 	 * insert both the head node and the new ref without dropping
824 	 * the spin lock
825 	 */
826 	add_delayed_ref_head(fs_info, trans, &head_ref->node, bytenr,
827 				   num_bytes, action, 1);
828 
829 	add_delayed_data_ref(fs_info, trans, &ref->node, bytenr,
830 				   num_bytes, parent, ref_root, owner, offset,
831 				   action, for_cow);
832 	spin_unlock(&delayed_refs->lock);
833 	if (need_ref_seq(for_cow, ref_root))
834 		btrfs_qgroup_record_ref(trans, &ref->node, extent_op);
835 
836 	return 0;
837 }
838 
839 int btrfs_add_delayed_extent_op(struct btrfs_fs_info *fs_info,
840 				struct btrfs_trans_handle *trans,
841 				u64 bytenr, u64 num_bytes,
842 				struct btrfs_delayed_extent_op *extent_op)
843 {
844 	struct btrfs_delayed_ref_head *head_ref;
845 	struct btrfs_delayed_ref_root *delayed_refs;
846 
847 	head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
848 	if (!head_ref)
849 		return -ENOMEM;
850 
851 	head_ref->extent_op = extent_op;
852 
853 	delayed_refs = &trans->transaction->delayed_refs;
854 	spin_lock(&delayed_refs->lock);
855 
856 	add_delayed_ref_head(fs_info, trans, &head_ref->node, bytenr,
857 				   num_bytes, BTRFS_UPDATE_DELAYED_HEAD,
858 				   extent_op->is_data);
859 
860 	spin_unlock(&delayed_refs->lock);
861 	return 0;
862 }
863 
864 /*
865  * this does a simple search for the head node for a given extent.
866  * It must be called with the delayed ref spinlock held, and it returns
867  * the head node if any where found, or NULL if not.
868  */
869 struct btrfs_delayed_ref_head *
870 btrfs_find_delayed_ref_head(struct btrfs_trans_handle *trans, u64 bytenr)
871 {
872 	struct btrfs_delayed_ref_node *ref;
873 	struct btrfs_delayed_ref_root *delayed_refs;
874 
875 	delayed_refs = &trans->transaction->delayed_refs;
876 	ref = find_ref_head(&delayed_refs->root, bytenr, NULL, 0);
877 	if (ref)
878 		return btrfs_delayed_node_to_head(ref);
879 	return NULL;
880 }
881 
882 void btrfs_delayed_ref_exit(void)
883 {
884 	if (btrfs_delayed_ref_head_cachep)
885 		kmem_cache_destroy(btrfs_delayed_ref_head_cachep);
886 	if (btrfs_delayed_tree_ref_cachep)
887 		kmem_cache_destroy(btrfs_delayed_tree_ref_cachep);
888 	if (btrfs_delayed_data_ref_cachep)
889 		kmem_cache_destroy(btrfs_delayed_data_ref_cachep);
890 	if (btrfs_delayed_extent_op_cachep)
891 		kmem_cache_destroy(btrfs_delayed_extent_op_cachep);
892 }
893 
894 int btrfs_delayed_ref_init(void)
895 {
896 	btrfs_delayed_ref_head_cachep = kmem_cache_create(
897 				"btrfs_delayed_ref_head",
898 				sizeof(struct btrfs_delayed_ref_head), 0,
899 				SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
900 	if (!btrfs_delayed_ref_head_cachep)
901 		goto fail;
902 
903 	btrfs_delayed_tree_ref_cachep = kmem_cache_create(
904 				"btrfs_delayed_tree_ref",
905 				sizeof(struct btrfs_delayed_tree_ref), 0,
906 				SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
907 	if (!btrfs_delayed_tree_ref_cachep)
908 		goto fail;
909 
910 	btrfs_delayed_data_ref_cachep = kmem_cache_create(
911 				"btrfs_delayed_data_ref",
912 				sizeof(struct btrfs_delayed_data_ref), 0,
913 				SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
914 	if (!btrfs_delayed_data_ref_cachep)
915 		goto fail;
916 
917 	btrfs_delayed_extent_op_cachep = kmem_cache_create(
918 				"btrfs_delayed_extent_op",
919 				sizeof(struct btrfs_delayed_extent_op), 0,
920 				SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
921 	if (!btrfs_delayed_extent_op_cachep)
922 		goto fail;
923 
924 	return 0;
925 fail:
926 	btrfs_delayed_ref_exit();
927 	return -ENOMEM;
928 }
929