xref: /openbmc/linux/fs/btrfs/delayed-ref.c (revision bc5aa3a0)
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 static 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 		}
260 next:
261 		next = next2;
262 	}
263 
264 	return done;
265 }
266 
267 void 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 		}
308 next:
309 		ref = list_next_entry(ref, list);
310 	}
311 }
312 
313 int 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 
337 struct btrfs_delayed_ref_head *
338 btrfs_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 
347 again:
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  */
391 static int
392 add_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 
444 add_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  */
456 static noinline void
457 update_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  */
533 static noinline struct btrfs_delayed_ref_head *
534 add_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 	int count_mod = 1;
545 	int must_insert_reserved = 0;
546 
547 	/* If reserved is provided, it must be a data extent. */
548 	BUG_ON(!is_data && reserved);
549 
550 	/*
551 	 * the head node stores the sum of all the mods, so dropping a ref
552 	 * should drop the sum in the head node by one.
553 	 */
554 	if (action == BTRFS_UPDATE_DELAYED_HEAD)
555 		count_mod = 0;
556 	else if (action == BTRFS_DROP_DELAYED_REF)
557 		count_mod = -1;
558 
559 	/*
560 	 * BTRFS_ADD_DELAYED_EXTENT means that we need to update
561 	 * the reserved accounting when the extent is finally added, or
562 	 * if a later modification deletes the delayed ref without ever
563 	 * inserting the extent into the extent allocation tree.
564 	 * ref->must_insert_reserved is the flag used to record
565 	 * that accounting mods are required.
566 	 *
567 	 * Once we record must_insert_reserved, switch the action to
568 	 * BTRFS_ADD_DELAYED_REF because other special casing is not required.
569 	 */
570 	if (action == BTRFS_ADD_DELAYED_EXTENT)
571 		must_insert_reserved = 1;
572 	else
573 		must_insert_reserved = 0;
574 
575 	delayed_refs = &trans->transaction->delayed_refs;
576 
577 	/* first set the basic ref node struct up */
578 	atomic_set(&ref->refs, 1);
579 	ref->bytenr = bytenr;
580 	ref->num_bytes = num_bytes;
581 	ref->ref_mod = count_mod;
582 	ref->type  = 0;
583 	ref->action  = 0;
584 	ref->is_head = 1;
585 	ref->in_tree = 1;
586 	ref->seq = 0;
587 
588 	head_ref = btrfs_delayed_node_to_head(ref);
589 	head_ref->must_insert_reserved = must_insert_reserved;
590 	head_ref->is_data = is_data;
591 	INIT_LIST_HEAD(&head_ref->ref_list);
592 	head_ref->processing = 0;
593 	head_ref->total_ref_mod = count_mod;
594 	head_ref->qgroup_reserved = 0;
595 	head_ref->qgroup_ref_root = 0;
596 
597 	/* Record qgroup extent info if provided */
598 	if (qrecord) {
599 		if (ref_root && reserved) {
600 			head_ref->qgroup_ref_root = ref_root;
601 			head_ref->qgroup_reserved = reserved;
602 		}
603 
604 		qrecord->bytenr = bytenr;
605 		qrecord->num_bytes = num_bytes;
606 		qrecord->old_roots = NULL;
607 
608 		if(btrfs_qgroup_insert_dirty_extent_nolock(fs_info,
609 					delayed_refs, qrecord))
610 			kfree(qrecord);
611 	}
612 
613 	spin_lock_init(&head_ref->lock);
614 	mutex_init(&head_ref->mutex);
615 
616 	trace_add_delayed_ref_head(fs_info, ref, head_ref, action);
617 
618 	existing = htree_insert(&delayed_refs->href_root,
619 				&head_ref->href_node);
620 	if (existing) {
621 		WARN_ON(ref_root && reserved && existing->qgroup_ref_root
622 			&& existing->qgroup_reserved);
623 		update_existing_head_ref(delayed_refs, &existing->node, ref);
624 		/*
625 		 * we've updated the existing ref, free the newly
626 		 * allocated ref
627 		 */
628 		kmem_cache_free(btrfs_delayed_ref_head_cachep, head_ref);
629 		head_ref = existing;
630 	} else {
631 		if (is_data && count_mod < 0)
632 			delayed_refs->pending_csums += num_bytes;
633 		delayed_refs->num_heads++;
634 		delayed_refs->num_heads_ready++;
635 		atomic_inc(&delayed_refs->num_entries);
636 		trans->delayed_ref_updates++;
637 	}
638 	return head_ref;
639 }
640 
641 /*
642  * helper to insert a delayed tree ref into the rbtree.
643  */
644 static noinline void
645 add_delayed_tree_ref(struct btrfs_fs_info *fs_info,
646 		     struct btrfs_trans_handle *trans,
647 		     struct btrfs_delayed_ref_head *head_ref,
648 		     struct btrfs_delayed_ref_node *ref, u64 bytenr,
649 		     u64 num_bytes, u64 parent, u64 ref_root, int level,
650 		     int action)
651 {
652 	struct btrfs_delayed_tree_ref *full_ref;
653 	struct btrfs_delayed_ref_root *delayed_refs;
654 	u64 seq = 0;
655 	int ret;
656 
657 	if (action == BTRFS_ADD_DELAYED_EXTENT)
658 		action = BTRFS_ADD_DELAYED_REF;
659 
660 	if (is_fstree(ref_root))
661 		seq = atomic64_read(&fs_info->tree_mod_seq);
662 	delayed_refs = &trans->transaction->delayed_refs;
663 
664 	/* first set the basic ref node struct up */
665 	atomic_set(&ref->refs, 1);
666 	ref->bytenr = bytenr;
667 	ref->num_bytes = num_bytes;
668 	ref->ref_mod = 1;
669 	ref->action = action;
670 	ref->is_head = 0;
671 	ref->in_tree = 1;
672 	ref->seq = seq;
673 
674 	full_ref = btrfs_delayed_node_to_tree_ref(ref);
675 	full_ref->parent = parent;
676 	full_ref->root = ref_root;
677 	if (parent)
678 		ref->type = BTRFS_SHARED_BLOCK_REF_KEY;
679 	else
680 		ref->type = BTRFS_TREE_BLOCK_REF_KEY;
681 	full_ref->level = level;
682 
683 	trace_add_delayed_tree_ref(fs_info, ref, full_ref, action);
684 
685 	ret = add_delayed_ref_tail_merge(trans, delayed_refs, head_ref, ref);
686 
687 	/*
688 	 * XXX: memory should be freed at the same level allocated.
689 	 * But bad practice is anywhere... Follow it now. Need cleanup.
690 	 */
691 	if (ret > 0)
692 		kmem_cache_free(btrfs_delayed_tree_ref_cachep, full_ref);
693 }
694 
695 /*
696  * helper to insert a delayed data ref into the rbtree.
697  */
698 static noinline void
699 add_delayed_data_ref(struct btrfs_fs_info *fs_info,
700 		     struct btrfs_trans_handle *trans,
701 		     struct btrfs_delayed_ref_head *head_ref,
702 		     struct btrfs_delayed_ref_node *ref, u64 bytenr,
703 		     u64 num_bytes, u64 parent, u64 ref_root, u64 owner,
704 		     u64 offset, int action)
705 {
706 	struct btrfs_delayed_data_ref *full_ref;
707 	struct btrfs_delayed_ref_root *delayed_refs;
708 	u64 seq = 0;
709 	int ret;
710 
711 	if (action == BTRFS_ADD_DELAYED_EXTENT)
712 		action = BTRFS_ADD_DELAYED_REF;
713 
714 	delayed_refs = &trans->transaction->delayed_refs;
715 
716 	if (is_fstree(ref_root))
717 		seq = atomic64_read(&fs_info->tree_mod_seq);
718 
719 	/* first set the basic ref node struct up */
720 	atomic_set(&ref->refs, 1);
721 	ref->bytenr = bytenr;
722 	ref->num_bytes = num_bytes;
723 	ref->ref_mod = 1;
724 	ref->action = action;
725 	ref->is_head = 0;
726 	ref->in_tree = 1;
727 	ref->seq = seq;
728 
729 	full_ref = btrfs_delayed_node_to_data_ref(ref);
730 	full_ref->parent = parent;
731 	full_ref->root = ref_root;
732 	if (parent)
733 		ref->type = BTRFS_SHARED_DATA_REF_KEY;
734 	else
735 		ref->type = BTRFS_EXTENT_DATA_REF_KEY;
736 
737 	full_ref->objectid = owner;
738 	full_ref->offset = offset;
739 
740 	trace_add_delayed_data_ref(fs_info, ref, full_ref, action);
741 
742 	ret = add_delayed_ref_tail_merge(trans, delayed_refs, head_ref, ref);
743 
744 	if (ret > 0)
745 		kmem_cache_free(btrfs_delayed_data_ref_cachep, full_ref);
746 }
747 
748 /*
749  * add a delayed tree ref.  This does all of the accounting required
750  * to make sure the delayed ref is eventually processed before this
751  * transaction commits.
752  */
753 int btrfs_add_delayed_tree_ref(struct btrfs_fs_info *fs_info,
754 			       struct btrfs_trans_handle *trans,
755 			       u64 bytenr, u64 num_bytes, u64 parent,
756 			       u64 ref_root,  int level, int action,
757 			       struct btrfs_delayed_extent_op *extent_op)
758 {
759 	struct btrfs_delayed_tree_ref *ref;
760 	struct btrfs_delayed_ref_head *head_ref;
761 	struct btrfs_delayed_ref_root *delayed_refs;
762 	struct btrfs_qgroup_extent_record *record = NULL;
763 
764 	BUG_ON(extent_op && extent_op->is_data);
765 	ref = kmem_cache_alloc(btrfs_delayed_tree_ref_cachep, GFP_NOFS);
766 	if (!ref)
767 		return -ENOMEM;
768 
769 	head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
770 	if (!head_ref)
771 		goto free_ref;
772 
773 	if (fs_info->quota_enabled && is_fstree(ref_root)) {
774 		record = kmalloc(sizeof(*record), GFP_NOFS);
775 		if (!record)
776 			goto free_head_ref;
777 	}
778 
779 	head_ref->extent_op = extent_op;
780 
781 	delayed_refs = &trans->transaction->delayed_refs;
782 	spin_lock(&delayed_refs->lock);
783 
784 	/*
785 	 * insert both the head node and the new ref without dropping
786 	 * the spin lock
787 	 */
788 	head_ref = add_delayed_ref_head(fs_info, trans, &head_ref->node, record,
789 					bytenr, num_bytes, 0, 0, action, 0);
790 
791 	add_delayed_tree_ref(fs_info, trans, head_ref, &ref->node, bytenr,
792 			     num_bytes, parent, ref_root, level, action);
793 	spin_unlock(&delayed_refs->lock);
794 
795 	return 0;
796 
797 free_head_ref:
798 	kmem_cache_free(btrfs_delayed_ref_head_cachep, head_ref);
799 free_ref:
800 	kmem_cache_free(btrfs_delayed_tree_ref_cachep, ref);
801 
802 	return -ENOMEM;
803 }
804 
805 /*
806  * add a delayed data ref. it's similar to btrfs_add_delayed_tree_ref.
807  */
808 int btrfs_add_delayed_data_ref(struct btrfs_fs_info *fs_info,
809 			       struct btrfs_trans_handle *trans,
810 			       u64 bytenr, u64 num_bytes,
811 			       u64 parent, u64 ref_root,
812 			       u64 owner, u64 offset, u64 reserved, int action,
813 			       struct btrfs_delayed_extent_op *extent_op)
814 {
815 	struct btrfs_delayed_data_ref *ref;
816 	struct btrfs_delayed_ref_head *head_ref;
817 	struct btrfs_delayed_ref_root *delayed_refs;
818 	struct btrfs_qgroup_extent_record *record = NULL;
819 
820 	BUG_ON(extent_op && !extent_op->is_data);
821 	ref = kmem_cache_alloc(btrfs_delayed_data_ref_cachep, GFP_NOFS);
822 	if (!ref)
823 		return -ENOMEM;
824 
825 	head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
826 	if (!head_ref) {
827 		kmem_cache_free(btrfs_delayed_data_ref_cachep, ref);
828 		return -ENOMEM;
829 	}
830 
831 	if (fs_info->quota_enabled && is_fstree(ref_root)) {
832 		record = kmalloc(sizeof(*record), GFP_NOFS);
833 		if (!record) {
834 			kmem_cache_free(btrfs_delayed_data_ref_cachep, ref);
835 			kmem_cache_free(btrfs_delayed_ref_head_cachep,
836 					head_ref);
837 			return -ENOMEM;
838 		}
839 	}
840 
841 	head_ref->extent_op = extent_op;
842 
843 	delayed_refs = &trans->transaction->delayed_refs;
844 	spin_lock(&delayed_refs->lock);
845 
846 	/*
847 	 * insert both the head node and the new ref without dropping
848 	 * the spin lock
849 	 */
850 	head_ref = add_delayed_ref_head(fs_info, trans, &head_ref->node, record,
851 					bytenr, num_bytes, ref_root, reserved,
852 					action, 1);
853 
854 	add_delayed_data_ref(fs_info, trans, head_ref, &ref->node, bytenr,
855 				   num_bytes, parent, ref_root, owner, offset,
856 				   action);
857 	spin_unlock(&delayed_refs->lock);
858 
859 	return 0;
860 }
861 
862 int btrfs_add_delayed_extent_op(struct btrfs_fs_info *fs_info,
863 				struct btrfs_trans_handle *trans,
864 				u64 bytenr, u64 num_bytes,
865 				struct btrfs_delayed_extent_op *extent_op)
866 {
867 	struct btrfs_delayed_ref_head *head_ref;
868 	struct btrfs_delayed_ref_root *delayed_refs;
869 
870 	head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
871 	if (!head_ref)
872 		return -ENOMEM;
873 
874 	head_ref->extent_op = extent_op;
875 
876 	delayed_refs = &trans->transaction->delayed_refs;
877 	spin_lock(&delayed_refs->lock);
878 
879 	add_delayed_ref_head(fs_info, trans, &head_ref->node, NULL, bytenr,
880 			     num_bytes, 0, 0, BTRFS_UPDATE_DELAYED_HEAD,
881 			     extent_op->is_data);
882 
883 	spin_unlock(&delayed_refs->lock);
884 	return 0;
885 }
886 
887 /*
888  * this does a simple search for the head node for a given extent.
889  * It must be called with the delayed ref spinlock held, and it returns
890  * the head node if any where found, or NULL if not.
891  */
892 struct btrfs_delayed_ref_head *
893 btrfs_find_delayed_ref_head(struct btrfs_trans_handle *trans, u64 bytenr)
894 {
895 	struct btrfs_delayed_ref_root *delayed_refs;
896 
897 	delayed_refs = &trans->transaction->delayed_refs;
898 	return find_ref_head(&delayed_refs->href_root, bytenr, 0);
899 }
900 
901 void btrfs_delayed_ref_exit(void)
902 {
903 	kmem_cache_destroy(btrfs_delayed_ref_head_cachep);
904 	kmem_cache_destroy(btrfs_delayed_tree_ref_cachep);
905 	kmem_cache_destroy(btrfs_delayed_data_ref_cachep);
906 	kmem_cache_destroy(btrfs_delayed_extent_op_cachep);
907 }
908 
909 int btrfs_delayed_ref_init(void)
910 {
911 	btrfs_delayed_ref_head_cachep = kmem_cache_create(
912 				"btrfs_delayed_ref_head",
913 				sizeof(struct btrfs_delayed_ref_head), 0,
914 				SLAB_MEM_SPREAD, NULL);
915 	if (!btrfs_delayed_ref_head_cachep)
916 		goto fail;
917 
918 	btrfs_delayed_tree_ref_cachep = kmem_cache_create(
919 				"btrfs_delayed_tree_ref",
920 				sizeof(struct btrfs_delayed_tree_ref), 0,
921 				SLAB_MEM_SPREAD, NULL);
922 	if (!btrfs_delayed_tree_ref_cachep)
923 		goto fail;
924 
925 	btrfs_delayed_data_ref_cachep = kmem_cache_create(
926 				"btrfs_delayed_data_ref",
927 				sizeof(struct btrfs_delayed_data_ref), 0,
928 				SLAB_MEM_SPREAD, NULL);
929 	if (!btrfs_delayed_data_ref_cachep)
930 		goto fail;
931 
932 	btrfs_delayed_extent_op_cachep = kmem_cache_create(
933 				"btrfs_delayed_extent_op",
934 				sizeof(struct btrfs_delayed_extent_op), 0,
935 				SLAB_MEM_SPREAD, NULL);
936 	if (!btrfs_delayed_extent_op_cachep)
937 		goto fail;
938 
939 	return 0;
940 fail:
941 	btrfs_delayed_ref_exit();
942 	return -ENOMEM;
943 }
944