xref: /openbmc/linux/fs/btrfs/delayed-ref.c (revision 703e7713)
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 "transaction.h"
13 #include "qgroup.h"
14 #include "space-info.h"
15 #include "tree-mod-log.h"
16 #include "fs.h"
17 
18 struct kmem_cache *btrfs_delayed_ref_head_cachep;
19 struct kmem_cache *btrfs_delayed_tree_ref_cachep;
20 struct kmem_cache *btrfs_delayed_data_ref_cachep;
21 struct 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 
31 bool 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:       number of items to drop
61  *
62  * Drops the delayed ref head's count from the delayed refs rsv and free any
63  * excess reservation we had.
64  */
65 void btrfs_delayed_refs_rsv_release(struct btrfs_fs_info *fs_info, int nr)
66 {
67 	struct btrfs_block_rsv *block_rsv = &fs_info->delayed_refs_rsv;
68 	const u64 num_bytes = btrfs_calc_delayed_ref_bytes(fs_info, nr);
69 	u64 released = 0;
70 
71 	released = btrfs_block_rsv_release(fs_info, block_rsv, num_bytes, NULL);
72 	if (released)
73 		trace_btrfs_space_reservation(fs_info, "delayed_refs_rsv",
74 					      0, released, 0);
75 }
76 
77 /*
78  * Adjust the size of the delayed refs rsv.
79  *
80  * This is to be called anytime we may have adjusted trans->delayed_ref_updates,
81  * it'll calculate the additional size and add it to the delayed_refs_rsv.
82  */
83 void btrfs_update_delayed_refs_rsv(struct btrfs_trans_handle *trans)
84 {
85 	struct btrfs_fs_info *fs_info = trans->fs_info;
86 	struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_refs_rsv;
87 	u64 num_bytes;
88 
89 	if (!trans->delayed_ref_updates)
90 		return;
91 
92 	num_bytes = btrfs_calc_delayed_ref_bytes(fs_info,
93 						 trans->delayed_ref_updates);
94 
95 	spin_lock(&delayed_rsv->lock);
96 	delayed_rsv->size += num_bytes;
97 	delayed_rsv->full = false;
98 	spin_unlock(&delayed_rsv->lock);
99 	trans->delayed_ref_updates = 0;
100 }
101 
102 /*
103  * Transfer bytes to our delayed refs rsv.
104  *
105  * @fs_info:   the filesystem
106  * @src:       source block rsv to transfer from
107  * @num_bytes: number of bytes to transfer
108  *
109  * This transfers up to the num_bytes amount from the src rsv to the
110  * delayed_refs_rsv.  Any extra bytes are returned to the space info.
111  */
112 void btrfs_migrate_to_delayed_refs_rsv(struct btrfs_fs_info *fs_info,
113 				       struct btrfs_block_rsv *src,
114 				       u64 num_bytes)
115 {
116 	struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
117 	u64 to_free = 0;
118 
119 	spin_lock(&src->lock);
120 	src->reserved -= num_bytes;
121 	src->size -= num_bytes;
122 	spin_unlock(&src->lock);
123 
124 	spin_lock(&delayed_refs_rsv->lock);
125 	if (delayed_refs_rsv->size > delayed_refs_rsv->reserved) {
126 		u64 delta = delayed_refs_rsv->size -
127 			delayed_refs_rsv->reserved;
128 		if (num_bytes > delta) {
129 			to_free = num_bytes - delta;
130 			num_bytes = delta;
131 		}
132 	} else {
133 		to_free = num_bytes;
134 		num_bytes = 0;
135 	}
136 
137 	if (num_bytes)
138 		delayed_refs_rsv->reserved += num_bytes;
139 	if (delayed_refs_rsv->reserved >= delayed_refs_rsv->size)
140 		delayed_refs_rsv->full = true;
141 	spin_unlock(&delayed_refs_rsv->lock);
142 
143 	if (num_bytes)
144 		trace_btrfs_space_reservation(fs_info, "delayed_refs_rsv",
145 					      0, num_bytes, 1);
146 	if (to_free)
147 		btrfs_space_info_free_bytes_may_use(fs_info,
148 				delayed_refs_rsv->space_info, to_free);
149 }
150 
151 /*
152  * Refill based on our delayed refs usage.
153  *
154  * @fs_info: the filesystem
155  * @flush:   control how we can flush for this reservation.
156  *
157  * This will refill the delayed block_rsv up to 1 items size worth of space and
158  * will return -ENOSPC if we can't make the reservation.
159  */
160 int btrfs_delayed_refs_rsv_refill(struct btrfs_fs_info *fs_info,
161 				  enum btrfs_reserve_flush_enum flush)
162 {
163 	struct btrfs_block_rsv *block_rsv = &fs_info->delayed_refs_rsv;
164 	u64 limit = btrfs_calc_delayed_ref_bytes(fs_info, 1);
165 	u64 num_bytes = 0;
166 	int ret = -ENOSPC;
167 
168 	spin_lock(&block_rsv->lock);
169 	if (block_rsv->reserved < block_rsv->size) {
170 		num_bytes = block_rsv->size - block_rsv->reserved;
171 		num_bytes = min(num_bytes, limit);
172 	}
173 	spin_unlock(&block_rsv->lock);
174 
175 	if (!num_bytes)
176 		return 0;
177 
178 	ret = btrfs_reserve_metadata_bytes(fs_info, block_rsv, num_bytes, flush);
179 	if (ret)
180 		return ret;
181 	btrfs_block_rsv_add_bytes(block_rsv, num_bytes, false);
182 	trace_btrfs_space_reservation(fs_info, "delayed_refs_rsv",
183 				      0, num_bytes, 1);
184 	return 0;
185 }
186 
187 /*
188  * compare two delayed tree backrefs with same bytenr and type
189  */
190 static int comp_tree_refs(struct btrfs_delayed_tree_ref *ref1,
191 			  struct btrfs_delayed_tree_ref *ref2)
192 {
193 	if (ref1->node.type == BTRFS_TREE_BLOCK_REF_KEY) {
194 		if (ref1->root < ref2->root)
195 			return -1;
196 		if (ref1->root > ref2->root)
197 			return 1;
198 	} else {
199 		if (ref1->parent < ref2->parent)
200 			return -1;
201 		if (ref1->parent > ref2->parent)
202 			return 1;
203 	}
204 	return 0;
205 }
206 
207 /*
208  * compare two delayed data backrefs with same bytenr and type
209  */
210 static int comp_data_refs(struct btrfs_delayed_data_ref *ref1,
211 			  struct btrfs_delayed_data_ref *ref2)
212 {
213 	if (ref1->node.type == BTRFS_EXTENT_DATA_REF_KEY) {
214 		if (ref1->root < ref2->root)
215 			return -1;
216 		if (ref1->root > ref2->root)
217 			return 1;
218 		if (ref1->objectid < ref2->objectid)
219 			return -1;
220 		if (ref1->objectid > ref2->objectid)
221 			return 1;
222 		if (ref1->offset < ref2->offset)
223 			return -1;
224 		if (ref1->offset > ref2->offset)
225 			return 1;
226 	} else {
227 		if (ref1->parent < ref2->parent)
228 			return -1;
229 		if (ref1->parent > ref2->parent)
230 			return 1;
231 	}
232 	return 0;
233 }
234 
235 static int comp_refs(struct btrfs_delayed_ref_node *ref1,
236 		     struct btrfs_delayed_ref_node *ref2,
237 		     bool check_seq)
238 {
239 	int ret = 0;
240 
241 	if (ref1->type < ref2->type)
242 		return -1;
243 	if (ref1->type > ref2->type)
244 		return 1;
245 	if (ref1->type == BTRFS_TREE_BLOCK_REF_KEY ||
246 	    ref1->type == BTRFS_SHARED_BLOCK_REF_KEY)
247 		ret = comp_tree_refs(btrfs_delayed_node_to_tree_ref(ref1),
248 				     btrfs_delayed_node_to_tree_ref(ref2));
249 	else
250 		ret = comp_data_refs(btrfs_delayed_node_to_data_ref(ref1),
251 				     btrfs_delayed_node_to_data_ref(ref2));
252 	if (ret)
253 		return ret;
254 	if (check_seq) {
255 		if (ref1->seq < ref2->seq)
256 			return -1;
257 		if (ref1->seq > ref2->seq)
258 			return 1;
259 	}
260 	return 0;
261 }
262 
263 /* insert a new ref to head ref rbtree */
264 static struct btrfs_delayed_ref_head *htree_insert(struct rb_root_cached *root,
265 						   struct rb_node *node)
266 {
267 	struct rb_node **p = &root->rb_root.rb_node;
268 	struct rb_node *parent_node = NULL;
269 	struct btrfs_delayed_ref_head *entry;
270 	struct btrfs_delayed_ref_head *ins;
271 	u64 bytenr;
272 	bool leftmost = true;
273 
274 	ins = rb_entry(node, struct btrfs_delayed_ref_head, href_node);
275 	bytenr = ins->bytenr;
276 	while (*p) {
277 		parent_node = *p;
278 		entry = rb_entry(parent_node, struct btrfs_delayed_ref_head,
279 				 href_node);
280 
281 		if (bytenr < entry->bytenr) {
282 			p = &(*p)->rb_left;
283 		} else if (bytenr > entry->bytenr) {
284 			p = &(*p)->rb_right;
285 			leftmost = false;
286 		} else {
287 			return entry;
288 		}
289 	}
290 
291 	rb_link_node(node, parent_node, p);
292 	rb_insert_color_cached(node, root, leftmost);
293 	return NULL;
294 }
295 
296 static struct btrfs_delayed_ref_node* tree_insert(struct rb_root_cached *root,
297 		struct btrfs_delayed_ref_node *ins)
298 {
299 	struct rb_node **p = &root->rb_root.rb_node;
300 	struct rb_node *node = &ins->ref_node;
301 	struct rb_node *parent_node = NULL;
302 	struct btrfs_delayed_ref_node *entry;
303 	bool leftmost = true;
304 
305 	while (*p) {
306 		int comp;
307 
308 		parent_node = *p;
309 		entry = rb_entry(parent_node, struct btrfs_delayed_ref_node,
310 				 ref_node);
311 		comp = comp_refs(ins, entry, true);
312 		if (comp < 0) {
313 			p = &(*p)->rb_left;
314 		} else if (comp > 0) {
315 			p = &(*p)->rb_right;
316 			leftmost = false;
317 		} else {
318 			return entry;
319 		}
320 	}
321 
322 	rb_link_node(node, parent_node, p);
323 	rb_insert_color_cached(node, root, leftmost);
324 	return NULL;
325 }
326 
327 static struct btrfs_delayed_ref_head *find_first_ref_head(
328 		struct btrfs_delayed_ref_root *dr)
329 {
330 	struct rb_node *n;
331 	struct btrfs_delayed_ref_head *entry;
332 
333 	n = rb_first_cached(&dr->href_root);
334 	if (!n)
335 		return NULL;
336 
337 	entry = rb_entry(n, struct btrfs_delayed_ref_head, href_node);
338 
339 	return entry;
340 }
341 
342 /*
343  * Find a head entry based on bytenr. This returns the delayed ref head if it
344  * was able to find one, or NULL if nothing was in that spot.  If return_bigger
345  * is given, the next bigger entry is returned if no exact match is found.
346  */
347 static struct btrfs_delayed_ref_head *find_ref_head(
348 		struct btrfs_delayed_ref_root *dr, u64 bytenr,
349 		bool return_bigger)
350 {
351 	struct rb_root *root = &dr->href_root.rb_root;
352 	struct rb_node *n;
353 	struct btrfs_delayed_ref_head *entry;
354 
355 	n = root->rb_node;
356 	entry = NULL;
357 	while (n) {
358 		entry = rb_entry(n, struct btrfs_delayed_ref_head, href_node);
359 
360 		if (bytenr < entry->bytenr)
361 			n = n->rb_left;
362 		else if (bytenr > entry->bytenr)
363 			n = n->rb_right;
364 		else
365 			return entry;
366 	}
367 	if (entry && return_bigger) {
368 		if (bytenr > entry->bytenr) {
369 			n = rb_next(&entry->href_node);
370 			if (!n)
371 				return NULL;
372 			entry = rb_entry(n, struct btrfs_delayed_ref_head,
373 					 href_node);
374 		}
375 		return entry;
376 	}
377 	return NULL;
378 }
379 
380 int btrfs_delayed_ref_lock(struct btrfs_delayed_ref_root *delayed_refs,
381 			   struct btrfs_delayed_ref_head *head)
382 {
383 	lockdep_assert_held(&delayed_refs->lock);
384 	if (mutex_trylock(&head->mutex))
385 		return 0;
386 
387 	refcount_inc(&head->refs);
388 	spin_unlock(&delayed_refs->lock);
389 
390 	mutex_lock(&head->mutex);
391 	spin_lock(&delayed_refs->lock);
392 	if (RB_EMPTY_NODE(&head->href_node)) {
393 		mutex_unlock(&head->mutex);
394 		btrfs_put_delayed_ref_head(head);
395 		return -EAGAIN;
396 	}
397 	btrfs_put_delayed_ref_head(head);
398 	return 0;
399 }
400 
401 static inline void drop_delayed_ref(struct btrfs_delayed_ref_root *delayed_refs,
402 				    struct btrfs_delayed_ref_head *head,
403 				    struct btrfs_delayed_ref_node *ref)
404 {
405 	lockdep_assert_held(&head->lock);
406 	rb_erase_cached(&ref->ref_node, &head->ref_tree);
407 	RB_CLEAR_NODE(&ref->ref_node);
408 	if (!list_empty(&ref->add_list))
409 		list_del(&ref->add_list);
410 	btrfs_put_delayed_ref(ref);
411 	atomic_dec(&delayed_refs->num_entries);
412 }
413 
414 static bool merge_ref(struct btrfs_delayed_ref_root *delayed_refs,
415 		      struct btrfs_delayed_ref_head *head,
416 		      struct btrfs_delayed_ref_node *ref,
417 		      u64 seq)
418 {
419 	struct btrfs_delayed_ref_node *next;
420 	struct rb_node *node = rb_next(&ref->ref_node);
421 	bool done = false;
422 
423 	while (!done && node) {
424 		int mod;
425 
426 		next = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
427 		node = rb_next(node);
428 		if (seq && next->seq >= seq)
429 			break;
430 		if (comp_refs(ref, next, false))
431 			break;
432 
433 		if (ref->action == next->action) {
434 			mod = next->ref_mod;
435 		} else {
436 			if (ref->ref_mod < next->ref_mod) {
437 				swap(ref, next);
438 				done = true;
439 			}
440 			mod = -next->ref_mod;
441 		}
442 
443 		drop_delayed_ref(delayed_refs, head, next);
444 		ref->ref_mod += mod;
445 		if (ref->ref_mod == 0) {
446 			drop_delayed_ref(delayed_refs, head, ref);
447 			done = true;
448 		} else {
449 			/*
450 			 * Can't have multiples of the same ref on a tree block.
451 			 */
452 			WARN_ON(ref->type == BTRFS_TREE_BLOCK_REF_KEY ||
453 				ref->type == BTRFS_SHARED_BLOCK_REF_KEY);
454 		}
455 	}
456 
457 	return done;
458 }
459 
460 void btrfs_merge_delayed_refs(struct btrfs_fs_info *fs_info,
461 			      struct btrfs_delayed_ref_root *delayed_refs,
462 			      struct btrfs_delayed_ref_head *head)
463 {
464 	struct btrfs_delayed_ref_node *ref;
465 	struct rb_node *node;
466 	u64 seq = 0;
467 
468 	lockdep_assert_held(&head->lock);
469 
470 	if (RB_EMPTY_ROOT(&head->ref_tree.rb_root))
471 		return;
472 
473 	/* We don't have too many refs to merge for data. */
474 	if (head->is_data)
475 		return;
476 
477 	seq = btrfs_tree_mod_log_lowest_seq(fs_info);
478 again:
479 	for (node = rb_first_cached(&head->ref_tree); node;
480 	     node = rb_next(node)) {
481 		ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
482 		if (seq && ref->seq >= seq)
483 			continue;
484 		if (merge_ref(delayed_refs, head, ref, seq))
485 			goto again;
486 	}
487 }
488 
489 int btrfs_check_delayed_seq(struct btrfs_fs_info *fs_info, u64 seq)
490 {
491 	int ret = 0;
492 	u64 min_seq = btrfs_tree_mod_log_lowest_seq(fs_info);
493 
494 	if (min_seq != 0 && seq >= min_seq) {
495 		btrfs_debug(fs_info,
496 			    "holding back delayed_ref %llu, lowest is %llu",
497 			    seq, min_seq);
498 		ret = 1;
499 	}
500 
501 	return ret;
502 }
503 
504 struct btrfs_delayed_ref_head *btrfs_select_ref_head(
505 		struct btrfs_delayed_ref_root *delayed_refs)
506 {
507 	struct btrfs_delayed_ref_head *head;
508 
509 	lockdep_assert_held(&delayed_refs->lock);
510 again:
511 	head = find_ref_head(delayed_refs, delayed_refs->run_delayed_start,
512 			     true);
513 	if (!head && delayed_refs->run_delayed_start != 0) {
514 		delayed_refs->run_delayed_start = 0;
515 		head = find_first_ref_head(delayed_refs);
516 	}
517 	if (!head)
518 		return NULL;
519 
520 	while (head->processing) {
521 		struct rb_node *node;
522 
523 		node = rb_next(&head->href_node);
524 		if (!node) {
525 			if (delayed_refs->run_delayed_start == 0)
526 				return NULL;
527 			delayed_refs->run_delayed_start = 0;
528 			goto again;
529 		}
530 		head = rb_entry(node, struct btrfs_delayed_ref_head,
531 				href_node);
532 	}
533 
534 	head->processing = true;
535 	WARN_ON(delayed_refs->num_heads_ready == 0);
536 	delayed_refs->num_heads_ready--;
537 	delayed_refs->run_delayed_start = head->bytenr +
538 		head->num_bytes;
539 	return head;
540 }
541 
542 void btrfs_delete_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
543 			   struct btrfs_delayed_ref_head *head)
544 {
545 	lockdep_assert_held(&delayed_refs->lock);
546 	lockdep_assert_held(&head->lock);
547 
548 	rb_erase_cached(&head->href_node, &delayed_refs->href_root);
549 	RB_CLEAR_NODE(&head->href_node);
550 	atomic_dec(&delayed_refs->num_entries);
551 	delayed_refs->num_heads--;
552 	if (!head->processing)
553 		delayed_refs->num_heads_ready--;
554 }
555 
556 /*
557  * Helper to insert the ref_node to the tail or merge with tail.
558  *
559  * Return false if the ref was inserted.
560  * Return true if the ref was merged into an existing one (and therefore can be
561  * freed by the caller).
562  */
563 static bool insert_delayed_ref(struct btrfs_delayed_ref_root *root,
564 			       struct btrfs_delayed_ref_head *href,
565 			       struct btrfs_delayed_ref_node *ref)
566 {
567 	struct btrfs_delayed_ref_node *exist;
568 	int mod;
569 
570 	spin_lock(&href->lock);
571 	exist = tree_insert(&href->ref_tree, ref);
572 	if (!exist) {
573 		if (ref->action == BTRFS_ADD_DELAYED_REF)
574 			list_add_tail(&ref->add_list, &href->ref_add_list);
575 		atomic_inc(&root->num_entries);
576 		spin_unlock(&href->lock);
577 		return false;
578 	}
579 
580 	/* Now we are sure we can merge */
581 	if (exist->action == ref->action) {
582 		mod = ref->ref_mod;
583 	} else {
584 		/* Need to change action */
585 		if (exist->ref_mod < ref->ref_mod) {
586 			exist->action = ref->action;
587 			mod = -exist->ref_mod;
588 			exist->ref_mod = ref->ref_mod;
589 			if (ref->action == BTRFS_ADD_DELAYED_REF)
590 				list_add_tail(&exist->add_list,
591 					      &href->ref_add_list);
592 			else if (ref->action == BTRFS_DROP_DELAYED_REF) {
593 				ASSERT(!list_empty(&exist->add_list));
594 				list_del(&exist->add_list);
595 			} else {
596 				ASSERT(0);
597 			}
598 		} else
599 			mod = -ref->ref_mod;
600 	}
601 	exist->ref_mod += mod;
602 
603 	/* remove existing tail if its ref_mod is zero */
604 	if (exist->ref_mod == 0)
605 		drop_delayed_ref(root, href, exist);
606 	spin_unlock(&href->lock);
607 	return true;
608 }
609 
610 /*
611  * helper function to update the accounting in the head ref
612  * existing and update must have the same bytenr
613  */
614 static noinline void update_existing_head_ref(struct btrfs_trans_handle *trans,
615 			 struct btrfs_delayed_ref_head *existing,
616 			 struct btrfs_delayed_ref_head *update)
617 {
618 	struct btrfs_delayed_ref_root *delayed_refs =
619 		&trans->transaction->delayed_refs;
620 	struct btrfs_fs_info *fs_info = trans->fs_info;
621 	int old_ref_mod;
622 
623 	BUG_ON(existing->is_data != update->is_data);
624 
625 	spin_lock(&existing->lock);
626 	if (update->must_insert_reserved) {
627 		/* if the extent was freed and then
628 		 * reallocated before the delayed ref
629 		 * entries were processed, we can end up
630 		 * with an existing head ref without
631 		 * the must_insert_reserved flag set.
632 		 * Set it again here
633 		 */
634 		existing->must_insert_reserved = update->must_insert_reserved;
635 
636 		/*
637 		 * update the num_bytes so we make sure the accounting
638 		 * is done correctly
639 		 */
640 		existing->num_bytes = update->num_bytes;
641 
642 	}
643 
644 	if (update->extent_op) {
645 		if (!existing->extent_op) {
646 			existing->extent_op = update->extent_op;
647 		} else {
648 			if (update->extent_op->update_key) {
649 				memcpy(&existing->extent_op->key,
650 				       &update->extent_op->key,
651 				       sizeof(update->extent_op->key));
652 				existing->extent_op->update_key = true;
653 			}
654 			if (update->extent_op->update_flags) {
655 				existing->extent_op->flags_to_set |=
656 					update->extent_op->flags_to_set;
657 				existing->extent_op->update_flags = true;
658 			}
659 			btrfs_free_delayed_extent_op(update->extent_op);
660 		}
661 	}
662 	/*
663 	 * update the reference mod on the head to reflect this new operation,
664 	 * only need the lock for this case cause we could be processing it
665 	 * currently, for refs we just added we know we're a-ok.
666 	 */
667 	old_ref_mod = existing->total_ref_mod;
668 	existing->ref_mod += update->ref_mod;
669 	existing->total_ref_mod += update->ref_mod;
670 
671 	/*
672 	 * If we are going to from a positive ref mod to a negative or vice
673 	 * versa we need to make sure to adjust pending_csums accordingly.
674 	 */
675 	if (existing->is_data) {
676 		u64 csum_leaves =
677 			btrfs_csum_bytes_to_leaves(fs_info,
678 						   existing->num_bytes);
679 
680 		if (existing->total_ref_mod >= 0 && old_ref_mod < 0) {
681 			delayed_refs->pending_csums -= existing->num_bytes;
682 			btrfs_delayed_refs_rsv_release(fs_info, csum_leaves);
683 		}
684 		if (existing->total_ref_mod < 0 && old_ref_mod >= 0) {
685 			delayed_refs->pending_csums += existing->num_bytes;
686 			trans->delayed_ref_updates += csum_leaves;
687 		}
688 	}
689 
690 	spin_unlock(&existing->lock);
691 }
692 
693 static void init_delayed_ref_head(struct btrfs_delayed_ref_head *head_ref,
694 				  struct btrfs_qgroup_extent_record *qrecord,
695 				  u64 bytenr, u64 num_bytes, u64 ref_root,
696 				  u64 reserved, int action, bool is_data,
697 				  bool is_system)
698 {
699 	int count_mod = 1;
700 	bool must_insert_reserved = false;
701 
702 	/* If reserved is provided, it must be a data extent. */
703 	BUG_ON(!is_data && reserved);
704 
705 	switch (action) {
706 	case BTRFS_UPDATE_DELAYED_HEAD:
707 		count_mod = 0;
708 		break;
709 	case BTRFS_DROP_DELAYED_REF:
710 		/*
711 		 * The head node stores the sum of all the mods, so dropping a ref
712 		 * should drop the sum in the head node by one.
713 		 */
714 		count_mod = -1;
715 		break;
716 	case BTRFS_ADD_DELAYED_EXTENT:
717 		/*
718 		 * BTRFS_ADD_DELAYED_EXTENT means that we need to update the
719 		 * reserved accounting when the extent is finally added, or if a
720 		 * later modification deletes the delayed ref without ever
721 		 * inserting the extent into the extent allocation tree.
722 		 * ref->must_insert_reserved is the flag used to record that
723 		 * accounting mods are required.
724 		 *
725 		 * Once we record must_insert_reserved, switch the action to
726 		 * BTRFS_ADD_DELAYED_REF because other special casing is not
727 		 * required.
728 		 */
729 		must_insert_reserved = true;
730 		break;
731 	}
732 
733 	refcount_set(&head_ref->refs, 1);
734 	head_ref->bytenr = bytenr;
735 	head_ref->num_bytes = num_bytes;
736 	head_ref->ref_mod = count_mod;
737 	head_ref->must_insert_reserved = must_insert_reserved;
738 	head_ref->is_data = is_data;
739 	head_ref->is_system = is_system;
740 	head_ref->ref_tree = RB_ROOT_CACHED;
741 	INIT_LIST_HEAD(&head_ref->ref_add_list);
742 	RB_CLEAR_NODE(&head_ref->href_node);
743 	head_ref->processing = false;
744 	head_ref->total_ref_mod = count_mod;
745 	spin_lock_init(&head_ref->lock);
746 	mutex_init(&head_ref->mutex);
747 
748 	if (qrecord) {
749 		if (ref_root && reserved) {
750 			qrecord->data_rsv = reserved;
751 			qrecord->data_rsv_refroot = ref_root;
752 		}
753 		qrecord->bytenr = bytenr;
754 		qrecord->num_bytes = num_bytes;
755 		qrecord->old_roots = NULL;
756 	}
757 }
758 
759 /*
760  * helper function to actually insert a head node into the rbtree.
761  * this does all the dirty work in terms of maintaining the correct
762  * overall modification count.
763  */
764 static noinline struct btrfs_delayed_ref_head *
765 add_delayed_ref_head(struct btrfs_trans_handle *trans,
766 		     struct btrfs_delayed_ref_head *head_ref,
767 		     struct btrfs_qgroup_extent_record *qrecord,
768 		     int action, bool *qrecord_inserted_ret)
769 {
770 	struct btrfs_delayed_ref_head *existing;
771 	struct btrfs_delayed_ref_root *delayed_refs;
772 	bool qrecord_inserted = false;
773 
774 	delayed_refs = &trans->transaction->delayed_refs;
775 
776 	/* Record qgroup extent info if provided */
777 	if (qrecord) {
778 		if (btrfs_qgroup_trace_extent_nolock(trans->fs_info,
779 					delayed_refs, qrecord))
780 			kfree(qrecord);
781 		else
782 			qrecord_inserted = true;
783 	}
784 
785 	trace_add_delayed_ref_head(trans->fs_info, head_ref, action);
786 
787 	existing = htree_insert(&delayed_refs->href_root,
788 				&head_ref->href_node);
789 	if (existing) {
790 		update_existing_head_ref(trans, existing, head_ref);
791 		/*
792 		 * we've updated the existing ref, free the newly
793 		 * allocated ref
794 		 */
795 		kmem_cache_free(btrfs_delayed_ref_head_cachep, head_ref);
796 		head_ref = existing;
797 	} else {
798 		if (head_ref->is_data && head_ref->ref_mod < 0) {
799 			delayed_refs->pending_csums += head_ref->num_bytes;
800 			trans->delayed_ref_updates +=
801 				btrfs_csum_bytes_to_leaves(trans->fs_info,
802 							   head_ref->num_bytes);
803 		}
804 		delayed_refs->num_heads++;
805 		delayed_refs->num_heads_ready++;
806 		atomic_inc(&delayed_refs->num_entries);
807 		trans->delayed_ref_updates++;
808 	}
809 	if (qrecord_inserted_ret)
810 		*qrecord_inserted_ret = qrecord_inserted;
811 
812 	return head_ref;
813 }
814 
815 /*
816  * init_delayed_ref_common - Initialize the structure which represents a
817  *			     modification to a an extent.
818  *
819  * @fs_info:    Internal to the mounted filesystem mount structure.
820  *
821  * @ref:	The structure which is going to be initialized.
822  *
823  * @bytenr:	The logical address of the extent for which a modification is
824  *		going to be recorded.
825  *
826  * @num_bytes:  Size of the extent whose modification is being recorded.
827  *
828  * @ref_root:	The id of the root where this modification has originated, this
829  *		can be either one of the well-known metadata trees or the
830  *		subvolume id which references this extent.
831  *
832  * @action:	Can be one of BTRFS_ADD_DELAYED_REF/BTRFS_DROP_DELAYED_REF or
833  *		BTRFS_ADD_DELAYED_EXTENT
834  *
835  * @ref_type:	Holds the type of the extent which is being recorded, can be
836  *		one of BTRFS_SHARED_BLOCK_REF_KEY/BTRFS_TREE_BLOCK_REF_KEY
837  *		when recording a metadata extent or BTRFS_SHARED_DATA_REF_KEY/
838  *		BTRFS_EXTENT_DATA_REF_KEY when recording data extent
839  */
840 static void init_delayed_ref_common(struct btrfs_fs_info *fs_info,
841 				    struct btrfs_delayed_ref_node *ref,
842 				    u64 bytenr, u64 num_bytes, u64 ref_root,
843 				    int action, u8 ref_type)
844 {
845 	u64 seq = 0;
846 
847 	if (action == BTRFS_ADD_DELAYED_EXTENT)
848 		action = BTRFS_ADD_DELAYED_REF;
849 
850 	if (is_fstree(ref_root))
851 		seq = atomic64_read(&fs_info->tree_mod_seq);
852 
853 	refcount_set(&ref->refs, 1);
854 	ref->bytenr = bytenr;
855 	ref->num_bytes = num_bytes;
856 	ref->ref_mod = 1;
857 	ref->action = action;
858 	ref->seq = seq;
859 	ref->type = ref_type;
860 	RB_CLEAR_NODE(&ref->ref_node);
861 	INIT_LIST_HEAD(&ref->add_list);
862 }
863 
864 /*
865  * add a delayed tree ref.  This does all of the accounting required
866  * to make sure the delayed ref is eventually processed before this
867  * transaction commits.
868  */
869 int btrfs_add_delayed_tree_ref(struct btrfs_trans_handle *trans,
870 			       struct btrfs_ref *generic_ref,
871 			       struct btrfs_delayed_extent_op *extent_op)
872 {
873 	struct btrfs_fs_info *fs_info = trans->fs_info;
874 	struct btrfs_delayed_tree_ref *ref;
875 	struct btrfs_delayed_ref_head *head_ref;
876 	struct btrfs_delayed_ref_root *delayed_refs;
877 	struct btrfs_qgroup_extent_record *record = NULL;
878 	bool qrecord_inserted;
879 	bool is_system;
880 	bool merged;
881 	int action = generic_ref->action;
882 	int level = generic_ref->tree_ref.level;
883 	u64 bytenr = generic_ref->bytenr;
884 	u64 num_bytes = generic_ref->len;
885 	u64 parent = generic_ref->parent;
886 	u8 ref_type;
887 
888 	is_system = (generic_ref->tree_ref.owning_root == BTRFS_CHUNK_TREE_OBJECTID);
889 
890 	ASSERT(generic_ref->type == BTRFS_REF_METADATA && generic_ref->action);
891 	ref = kmem_cache_alloc(btrfs_delayed_tree_ref_cachep, GFP_NOFS);
892 	if (!ref)
893 		return -ENOMEM;
894 
895 	head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
896 	if (!head_ref) {
897 		kmem_cache_free(btrfs_delayed_tree_ref_cachep, ref);
898 		return -ENOMEM;
899 	}
900 
901 	if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) &&
902 	    !generic_ref->skip_qgroup) {
903 		record = kzalloc(sizeof(*record), GFP_NOFS);
904 		if (!record) {
905 			kmem_cache_free(btrfs_delayed_tree_ref_cachep, ref);
906 			kmem_cache_free(btrfs_delayed_ref_head_cachep, head_ref);
907 			return -ENOMEM;
908 		}
909 	}
910 
911 	if (parent)
912 		ref_type = BTRFS_SHARED_BLOCK_REF_KEY;
913 	else
914 		ref_type = BTRFS_TREE_BLOCK_REF_KEY;
915 
916 	init_delayed_ref_common(fs_info, &ref->node, bytenr, num_bytes,
917 				generic_ref->tree_ref.owning_root, action,
918 				ref_type);
919 	ref->root = generic_ref->tree_ref.owning_root;
920 	ref->parent = parent;
921 	ref->level = level;
922 
923 	init_delayed_ref_head(head_ref, record, bytenr, num_bytes,
924 			      generic_ref->tree_ref.owning_root, 0, action,
925 			      false, is_system);
926 	head_ref->extent_op = extent_op;
927 
928 	delayed_refs = &trans->transaction->delayed_refs;
929 	spin_lock(&delayed_refs->lock);
930 
931 	/*
932 	 * insert both the head node and the new ref without dropping
933 	 * the spin lock
934 	 */
935 	head_ref = add_delayed_ref_head(trans, head_ref, record,
936 					action, &qrecord_inserted);
937 
938 	merged = insert_delayed_ref(delayed_refs, head_ref, &ref->node);
939 	spin_unlock(&delayed_refs->lock);
940 
941 	/*
942 	 * Need to update the delayed_refs_rsv with any changes we may have
943 	 * made.
944 	 */
945 	btrfs_update_delayed_refs_rsv(trans);
946 
947 	trace_add_delayed_tree_ref(fs_info, &ref->node, ref,
948 				   action == BTRFS_ADD_DELAYED_EXTENT ?
949 				   BTRFS_ADD_DELAYED_REF : action);
950 	if (merged)
951 		kmem_cache_free(btrfs_delayed_tree_ref_cachep, ref);
952 
953 	if (qrecord_inserted)
954 		btrfs_qgroup_trace_extent_post(trans, record);
955 
956 	return 0;
957 }
958 
959 /*
960  * add a delayed data ref. it's similar to btrfs_add_delayed_tree_ref.
961  */
962 int btrfs_add_delayed_data_ref(struct btrfs_trans_handle *trans,
963 			       struct btrfs_ref *generic_ref,
964 			       u64 reserved)
965 {
966 	struct btrfs_fs_info *fs_info = trans->fs_info;
967 	struct btrfs_delayed_data_ref *ref;
968 	struct btrfs_delayed_ref_head *head_ref;
969 	struct btrfs_delayed_ref_root *delayed_refs;
970 	struct btrfs_qgroup_extent_record *record = NULL;
971 	bool qrecord_inserted;
972 	int action = generic_ref->action;
973 	bool merged;
974 	u64 bytenr = generic_ref->bytenr;
975 	u64 num_bytes = generic_ref->len;
976 	u64 parent = generic_ref->parent;
977 	u64 ref_root = generic_ref->data_ref.owning_root;
978 	u64 owner = generic_ref->data_ref.ino;
979 	u64 offset = generic_ref->data_ref.offset;
980 	u8 ref_type;
981 
982 	ASSERT(generic_ref->type == BTRFS_REF_DATA && action);
983 	ref = kmem_cache_alloc(btrfs_delayed_data_ref_cachep, GFP_NOFS);
984 	if (!ref)
985 		return -ENOMEM;
986 
987 	if (parent)
988 	        ref_type = BTRFS_SHARED_DATA_REF_KEY;
989 	else
990 	        ref_type = BTRFS_EXTENT_DATA_REF_KEY;
991 	init_delayed_ref_common(fs_info, &ref->node, bytenr, num_bytes,
992 				ref_root, action, ref_type);
993 	ref->root = ref_root;
994 	ref->parent = parent;
995 	ref->objectid = owner;
996 	ref->offset = offset;
997 
998 
999 	head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
1000 	if (!head_ref) {
1001 		kmem_cache_free(btrfs_delayed_data_ref_cachep, ref);
1002 		return -ENOMEM;
1003 	}
1004 
1005 	if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) &&
1006 	    !generic_ref->skip_qgroup) {
1007 		record = kzalloc(sizeof(*record), GFP_NOFS);
1008 		if (!record) {
1009 			kmem_cache_free(btrfs_delayed_data_ref_cachep, ref);
1010 			kmem_cache_free(btrfs_delayed_ref_head_cachep,
1011 					head_ref);
1012 			return -ENOMEM;
1013 		}
1014 	}
1015 
1016 	init_delayed_ref_head(head_ref, record, bytenr, num_bytes, ref_root,
1017 			      reserved, action, true, false);
1018 	head_ref->extent_op = NULL;
1019 
1020 	delayed_refs = &trans->transaction->delayed_refs;
1021 	spin_lock(&delayed_refs->lock);
1022 
1023 	/*
1024 	 * insert both the head node and the new ref without dropping
1025 	 * the spin lock
1026 	 */
1027 	head_ref = add_delayed_ref_head(trans, head_ref, record,
1028 					action, &qrecord_inserted);
1029 
1030 	merged = insert_delayed_ref(delayed_refs, head_ref, &ref->node);
1031 	spin_unlock(&delayed_refs->lock);
1032 
1033 	/*
1034 	 * Need to update the delayed_refs_rsv with any changes we may have
1035 	 * made.
1036 	 */
1037 	btrfs_update_delayed_refs_rsv(trans);
1038 
1039 	trace_add_delayed_data_ref(trans->fs_info, &ref->node, ref,
1040 				   action == BTRFS_ADD_DELAYED_EXTENT ?
1041 				   BTRFS_ADD_DELAYED_REF : action);
1042 	if (merged)
1043 		kmem_cache_free(btrfs_delayed_data_ref_cachep, ref);
1044 
1045 
1046 	if (qrecord_inserted)
1047 		return btrfs_qgroup_trace_extent_post(trans, record);
1048 	return 0;
1049 }
1050 
1051 int btrfs_add_delayed_extent_op(struct btrfs_trans_handle *trans,
1052 				u64 bytenr, u64 num_bytes,
1053 				struct btrfs_delayed_extent_op *extent_op)
1054 {
1055 	struct btrfs_delayed_ref_head *head_ref;
1056 	struct btrfs_delayed_ref_root *delayed_refs;
1057 
1058 	head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
1059 	if (!head_ref)
1060 		return -ENOMEM;
1061 
1062 	init_delayed_ref_head(head_ref, NULL, bytenr, num_bytes, 0, 0,
1063 			      BTRFS_UPDATE_DELAYED_HEAD, false, false);
1064 	head_ref->extent_op = extent_op;
1065 
1066 	delayed_refs = &trans->transaction->delayed_refs;
1067 	spin_lock(&delayed_refs->lock);
1068 
1069 	add_delayed_ref_head(trans, head_ref, NULL, BTRFS_UPDATE_DELAYED_HEAD,
1070 			     NULL);
1071 
1072 	spin_unlock(&delayed_refs->lock);
1073 
1074 	/*
1075 	 * Need to update the delayed_refs_rsv with any changes we may have
1076 	 * made.
1077 	 */
1078 	btrfs_update_delayed_refs_rsv(trans);
1079 	return 0;
1080 }
1081 
1082 /*
1083  * This does a simple search for the head node for a given extent.  Returns the
1084  * head node if found, or NULL if not.
1085  */
1086 struct btrfs_delayed_ref_head *
1087 btrfs_find_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs, u64 bytenr)
1088 {
1089 	lockdep_assert_held(&delayed_refs->lock);
1090 
1091 	return find_ref_head(delayed_refs, bytenr, false);
1092 }
1093 
1094 void __cold btrfs_delayed_ref_exit(void)
1095 {
1096 	kmem_cache_destroy(btrfs_delayed_ref_head_cachep);
1097 	kmem_cache_destroy(btrfs_delayed_tree_ref_cachep);
1098 	kmem_cache_destroy(btrfs_delayed_data_ref_cachep);
1099 	kmem_cache_destroy(btrfs_delayed_extent_op_cachep);
1100 }
1101 
1102 int __init btrfs_delayed_ref_init(void)
1103 {
1104 	btrfs_delayed_ref_head_cachep = kmem_cache_create(
1105 				"btrfs_delayed_ref_head",
1106 				sizeof(struct btrfs_delayed_ref_head), 0,
1107 				SLAB_MEM_SPREAD, NULL);
1108 	if (!btrfs_delayed_ref_head_cachep)
1109 		goto fail;
1110 
1111 	btrfs_delayed_tree_ref_cachep = kmem_cache_create(
1112 				"btrfs_delayed_tree_ref",
1113 				sizeof(struct btrfs_delayed_tree_ref), 0,
1114 				SLAB_MEM_SPREAD, NULL);
1115 	if (!btrfs_delayed_tree_ref_cachep)
1116 		goto fail;
1117 
1118 	btrfs_delayed_data_ref_cachep = kmem_cache_create(
1119 				"btrfs_delayed_data_ref",
1120 				sizeof(struct btrfs_delayed_data_ref), 0,
1121 				SLAB_MEM_SPREAD, NULL);
1122 	if (!btrfs_delayed_data_ref_cachep)
1123 		goto fail;
1124 
1125 	btrfs_delayed_extent_op_cachep = kmem_cache_create(
1126 				"btrfs_delayed_extent_op",
1127 				sizeof(struct btrfs_delayed_extent_op), 0,
1128 				SLAB_MEM_SPREAD, NULL);
1129 	if (!btrfs_delayed_extent_op_cachep)
1130 		goto fail;
1131 
1132 	return 0;
1133 fail:
1134 	btrfs_delayed_ref_exit();
1135 	return -ENOMEM;
1136 }
1137