xref: /openbmc/linux/fs/btrfs/delayed-ref.c (revision f79e4d5f)
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 "ctree.h"
10 #include "delayed-ref.h"
11 #include "transaction.h"
12 #include "qgroup.h"
13 
14 struct kmem_cache *btrfs_delayed_ref_head_cachep;
15 struct kmem_cache *btrfs_delayed_tree_ref_cachep;
16 struct kmem_cache *btrfs_delayed_data_ref_cachep;
17 struct kmem_cache *btrfs_delayed_extent_op_cachep;
18 /*
19  * delayed back reference update tracking.  For subvolume trees
20  * we queue up extent allocations and backref maintenance for
21  * delayed processing.   This avoids deep call chains where we
22  * add extents in the middle of btrfs_search_slot, and it allows
23  * us to buffer up frequently modified backrefs in an rb tree instead
24  * of hammering updates on the extent allocation tree.
25  */
26 
27 /*
28  * compare two delayed tree backrefs with same bytenr and type
29  */
30 static int comp_tree_refs(struct btrfs_delayed_tree_ref *ref1,
31 			  struct btrfs_delayed_tree_ref *ref2)
32 {
33 	if (ref1->node.type == BTRFS_TREE_BLOCK_REF_KEY) {
34 		if (ref1->root < ref2->root)
35 			return -1;
36 		if (ref1->root > ref2->root)
37 			return 1;
38 	} else {
39 		if (ref1->parent < ref2->parent)
40 			return -1;
41 		if (ref1->parent > ref2->parent)
42 			return 1;
43 	}
44 	return 0;
45 }
46 
47 /*
48  * compare two delayed data backrefs with same bytenr and type
49  */
50 static int comp_data_refs(struct btrfs_delayed_data_ref *ref1,
51 			  struct btrfs_delayed_data_ref *ref2)
52 {
53 	if (ref1->node.type == BTRFS_EXTENT_DATA_REF_KEY) {
54 		if (ref1->root < ref2->root)
55 			return -1;
56 		if (ref1->root > ref2->root)
57 			return 1;
58 		if (ref1->objectid < ref2->objectid)
59 			return -1;
60 		if (ref1->objectid > ref2->objectid)
61 			return 1;
62 		if (ref1->offset < ref2->offset)
63 			return -1;
64 		if (ref1->offset > ref2->offset)
65 			return 1;
66 	} else {
67 		if (ref1->parent < ref2->parent)
68 			return -1;
69 		if (ref1->parent > ref2->parent)
70 			return 1;
71 	}
72 	return 0;
73 }
74 
75 static int comp_refs(struct btrfs_delayed_ref_node *ref1,
76 		     struct btrfs_delayed_ref_node *ref2,
77 		     bool check_seq)
78 {
79 	int ret = 0;
80 
81 	if (ref1->type < ref2->type)
82 		return -1;
83 	if (ref1->type > ref2->type)
84 		return 1;
85 	if (ref1->type == BTRFS_TREE_BLOCK_REF_KEY ||
86 	    ref1->type == BTRFS_SHARED_BLOCK_REF_KEY)
87 		ret = comp_tree_refs(btrfs_delayed_node_to_tree_ref(ref1),
88 				     btrfs_delayed_node_to_tree_ref(ref2));
89 	else
90 		ret = comp_data_refs(btrfs_delayed_node_to_data_ref(ref1),
91 				     btrfs_delayed_node_to_data_ref(ref2));
92 	if (ret)
93 		return ret;
94 	if (check_seq) {
95 		if (ref1->seq < ref2->seq)
96 			return -1;
97 		if (ref1->seq > ref2->seq)
98 			return 1;
99 	}
100 	return 0;
101 }
102 
103 /* insert a new ref to head ref rbtree */
104 static struct btrfs_delayed_ref_head *htree_insert(struct rb_root *root,
105 						   struct rb_node *node)
106 {
107 	struct rb_node **p = &root->rb_node;
108 	struct rb_node *parent_node = NULL;
109 	struct btrfs_delayed_ref_head *entry;
110 	struct btrfs_delayed_ref_head *ins;
111 	u64 bytenr;
112 
113 	ins = rb_entry(node, struct btrfs_delayed_ref_head, href_node);
114 	bytenr = ins->bytenr;
115 	while (*p) {
116 		parent_node = *p;
117 		entry = rb_entry(parent_node, struct btrfs_delayed_ref_head,
118 				 href_node);
119 
120 		if (bytenr < entry->bytenr)
121 			p = &(*p)->rb_left;
122 		else if (bytenr > entry->bytenr)
123 			p = &(*p)->rb_right;
124 		else
125 			return entry;
126 	}
127 
128 	rb_link_node(node, parent_node, p);
129 	rb_insert_color(node, root);
130 	return NULL;
131 }
132 
133 static struct btrfs_delayed_ref_node* tree_insert(struct rb_root *root,
134 		struct btrfs_delayed_ref_node *ins)
135 {
136 	struct rb_node **p = &root->rb_node;
137 	struct rb_node *node = &ins->ref_node;
138 	struct rb_node *parent_node = NULL;
139 	struct btrfs_delayed_ref_node *entry;
140 
141 	while (*p) {
142 		int comp;
143 
144 		parent_node = *p;
145 		entry = rb_entry(parent_node, struct btrfs_delayed_ref_node,
146 				 ref_node);
147 		comp = comp_refs(ins, entry, true);
148 		if (comp < 0)
149 			p = &(*p)->rb_left;
150 		else if (comp > 0)
151 			p = &(*p)->rb_right;
152 		else
153 			return entry;
154 	}
155 
156 	rb_link_node(node, parent_node, p);
157 	rb_insert_color(node, root);
158 	return NULL;
159 }
160 
161 /*
162  * find an head entry based on bytenr. This returns the delayed ref
163  * head if it was able to find one, or NULL if nothing was in that spot.
164  * If return_bigger is given, the next bigger entry is returned if no exact
165  * match is found.
166  */
167 static struct btrfs_delayed_ref_head *
168 find_ref_head(struct rb_root *root, u64 bytenr,
169 	      int return_bigger)
170 {
171 	struct rb_node *n;
172 	struct btrfs_delayed_ref_head *entry;
173 
174 	n = root->rb_node;
175 	entry = NULL;
176 	while (n) {
177 		entry = rb_entry(n, struct btrfs_delayed_ref_head, href_node);
178 
179 		if (bytenr < entry->bytenr)
180 			n = n->rb_left;
181 		else if (bytenr > entry->bytenr)
182 			n = n->rb_right;
183 		else
184 			return entry;
185 	}
186 	if (entry && return_bigger) {
187 		if (bytenr > entry->bytenr) {
188 			n = rb_next(&entry->href_node);
189 			if (!n)
190 				n = rb_first(root);
191 			entry = rb_entry(n, struct btrfs_delayed_ref_head,
192 					 href_node);
193 			return entry;
194 		}
195 		return entry;
196 	}
197 	return NULL;
198 }
199 
200 int btrfs_delayed_ref_lock(struct btrfs_trans_handle *trans,
201 			   struct btrfs_delayed_ref_head *head)
202 {
203 	struct btrfs_delayed_ref_root *delayed_refs;
204 
205 	delayed_refs = &trans->transaction->delayed_refs;
206 	lockdep_assert_held(&delayed_refs->lock);
207 	if (mutex_trylock(&head->mutex))
208 		return 0;
209 
210 	refcount_inc(&head->refs);
211 	spin_unlock(&delayed_refs->lock);
212 
213 	mutex_lock(&head->mutex);
214 	spin_lock(&delayed_refs->lock);
215 	if (RB_EMPTY_NODE(&head->href_node)) {
216 		mutex_unlock(&head->mutex);
217 		btrfs_put_delayed_ref_head(head);
218 		return -EAGAIN;
219 	}
220 	btrfs_put_delayed_ref_head(head);
221 	return 0;
222 }
223 
224 static inline void drop_delayed_ref(struct btrfs_trans_handle *trans,
225 				    struct btrfs_delayed_ref_root *delayed_refs,
226 				    struct btrfs_delayed_ref_head *head,
227 				    struct btrfs_delayed_ref_node *ref)
228 {
229 	lockdep_assert_held(&head->lock);
230 	rb_erase(&ref->ref_node, &head->ref_tree);
231 	RB_CLEAR_NODE(&ref->ref_node);
232 	if (!list_empty(&ref->add_list))
233 		list_del(&ref->add_list);
234 	ref->in_tree = 0;
235 	btrfs_put_delayed_ref(ref);
236 	atomic_dec(&delayed_refs->num_entries);
237 	if (trans->delayed_ref_updates)
238 		trans->delayed_ref_updates--;
239 }
240 
241 static bool merge_ref(struct btrfs_trans_handle *trans,
242 		      struct btrfs_delayed_ref_root *delayed_refs,
243 		      struct btrfs_delayed_ref_head *head,
244 		      struct btrfs_delayed_ref_node *ref,
245 		      u64 seq)
246 {
247 	struct btrfs_delayed_ref_node *next;
248 	struct rb_node *node = rb_next(&ref->ref_node);
249 	bool done = false;
250 
251 	while (!done && node) {
252 		int mod;
253 
254 		next = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
255 		node = rb_next(node);
256 		if (seq && next->seq >= seq)
257 			break;
258 		if (comp_refs(ref, next, false))
259 			break;
260 
261 		if (ref->action == next->action) {
262 			mod = next->ref_mod;
263 		} else {
264 			if (ref->ref_mod < next->ref_mod) {
265 				swap(ref, next);
266 				done = true;
267 			}
268 			mod = -next->ref_mod;
269 		}
270 
271 		drop_delayed_ref(trans, delayed_refs, head, next);
272 		ref->ref_mod += mod;
273 		if (ref->ref_mod == 0) {
274 			drop_delayed_ref(trans, delayed_refs, head, ref);
275 			done = true;
276 		} else {
277 			/*
278 			 * Can't have multiples of the same ref on a tree block.
279 			 */
280 			WARN_ON(ref->type == BTRFS_TREE_BLOCK_REF_KEY ||
281 				ref->type == BTRFS_SHARED_BLOCK_REF_KEY);
282 		}
283 	}
284 
285 	return done;
286 }
287 
288 void btrfs_merge_delayed_refs(struct btrfs_trans_handle *trans,
289 			      struct btrfs_delayed_ref_root *delayed_refs,
290 			      struct btrfs_delayed_ref_head *head)
291 {
292 	struct btrfs_fs_info *fs_info = trans->fs_info;
293 	struct btrfs_delayed_ref_node *ref;
294 	struct rb_node *node;
295 	u64 seq = 0;
296 
297 	lockdep_assert_held(&head->lock);
298 
299 	if (RB_EMPTY_ROOT(&head->ref_tree))
300 		return;
301 
302 	/* We don't have too many refs to merge for data. */
303 	if (head->is_data)
304 		return;
305 
306 	spin_lock(&fs_info->tree_mod_seq_lock);
307 	if (!list_empty(&fs_info->tree_mod_seq_list)) {
308 		struct seq_list *elem;
309 
310 		elem = list_first_entry(&fs_info->tree_mod_seq_list,
311 					struct seq_list, list);
312 		seq = elem->seq;
313 	}
314 	spin_unlock(&fs_info->tree_mod_seq_lock);
315 
316 again:
317 	for (node = rb_first(&head->ref_tree); node; node = rb_next(node)) {
318 		ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
319 		if (seq && ref->seq >= seq)
320 			continue;
321 		if (merge_ref(trans, delayed_refs, head, ref, seq))
322 			goto again;
323 	}
324 }
325 
326 int btrfs_check_delayed_seq(struct btrfs_fs_info *fs_info, u64 seq)
327 {
328 	struct seq_list *elem;
329 	int ret = 0;
330 
331 	spin_lock(&fs_info->tree_mod_seq_lock);
332 	if (!list_empty(&fs_info->tree_mod_seq_list)) {
333 		elem = list_first_entry(&fs_info->tree_mod_seq_list,
334 					struct seq_list, list);
335 		if (seq >= elem->seq) {
336 			btrfs_debug(fs_info,
337 				"holding back delayed_ref %#x.%x, lowest is %#x.%x",
338 				(u32)(seq >> 32), (u32)seq,
339 				(u32)(elem->seq >> 32), (u32)elem->seq);
340 			ret = 1;
341 		}
342 	}
343 
344 	spin_unlock(&fs_info->tree_mod_seq_lock);
345 	return ret;
346 }
347 
348 struct btrfs_delayed_ref_head *
349 btrfs_select_ref_head(struct btrfs_trans_handle *trans)
350 {
351 	struct btrfs_delayed_ref_root *delayed_refs;
352 	struct btrfs_delayed_ref_head *head;
353 	u64 start;
354 	bool loop = false;
355 
356 	delayed_refs = &trans->transaction->delayed_refs;
357 
358 again:
359 	start = delayed_refs->run_delayed_start;
360 	head = find_ref_head(&delayed_refs->href_root, start, 1);
361 	if (!head && !loop) {
362 		delayed_refs->run_delayed_start = 0;
363 		start = 0;
364 		loop = true;
365 		head = find_ref_head(&delayed_refs->href_root, start, 1);
366 		if (!head)
367 			return NULL;
368 	} else if (!head && loop) {
369 		return NULL;
370 	}
371 
372 	while (head->processing) {
373 		struct rb_node *node;
374 
375 		node = rb_next(&head->href_node);
376 		if (!node) {
377 			if (loop)
378 				return NULL;
379 			delayed_refs->run_delayed_start = 0;
380 			start = 0;
381 			loop = true;
382 			goto again;
383 		}
384 		head = rb_entry(node, struct btrfs_delayed_ref_head,
385 				href_node);
386 	}
387 
388 	head->processing = 1;
389 	WARN_ON(delayed_refs->num_heads_ready == 0);
390 	delayed_refs->num_heads_ready--;
391 	delayed_refs->run_delayed_start = head->bytenr +
392 		head->num_bytes;
393 	return head;
394 }
395 
396 /*
397  * Helper to insert the ref_node to the tail or merge with tail.
398  *
399  * Return 0 for insert.
400  * Return >0 for merge.
401  */
402 static int insert_delayed_ref(struct btrfs_trans_handle *trans,
403 			      struct btrfs_delayed_ref_root *root,
404 			      struct btrfs_delayed_ref_head *href,
405 			      struct btrfs_delayed_ref_node *ref)
406 {
407 	struct btrfs_delayed_ref_node *exist;
408 	int mod;
409 	int ret = 0;
410 
411 	spin_lock(&href->lock);
412 	exist = tree_insert(&href->ref_tree, ref);
413 	if (!exist)
414 		goto inserted;
415 
416 	/* Now we are sure we can merge */
417 	ret = 1;
418 	if (exist->action == ref->action) {
419 		mod = ref->ref_mod;
420 	} else {
421 		/* Need to change action */
422 		if (exist->ref_mod < ref->ref_mod) {
423 			exist->action = ref->action;
424 			mod = -exist->ref_mod;
425 			exist->ref_mod = ref->ref_mod;
426 			if (ref->action == BTRFS_ADD_DELAYED_REF)
427 				list_add_tail(&exist->add_list,
428 					      &href->ref_add_list);
429 			else if (ref->action == BTRFS_DROP_DELAYED_REF) {
430 				ASSERT(!list_empty(&exist->add_list));
431 				list_del(&exist->add_list);
432 			} else {
433 				ASSERT(0);
434 			}
435 		} else
436 			mod = -ref->ref_mod;
437 	}
438 	exist->ref_mod += mod;
439 
440 	/* remove existing tail if its ref_mod is zero */
441 	if (exist->ref_mod == 0)
442 		drop_delayed_ref(trans, root, href, exist);
443 	spin_unlock(&href->lock);
444 	return ret;
445 inserted:
446 	if (ref->action == BTRFS_ADD_DELAYED_REF)
447 		list_add_tail(&ref->add_list, &href->ref_add_list);
448 	atomic_inc(&root->num_entries);
449 	trans->delayed_ref_updates++;
450 	spin_unlock(&href->lock);
451 	return ret;
452 }
453 
454 /*
455  * helper function to update the accounting in the head ref
456  * existing and update must have the same bytenr
457  */
458 static noinline void
459 update_existing_head_ref(struct btrfs_delayed_ref_root *delayed_refs,
460 			 struct btrfs_delayed_ref_head *existing,
461 			 struct btrfs_delayed_ref_head *update,
462 			 int *old_ref_mod_ret)
463 {
464 	int old_ref_mod;
465 
466 	BUG_ON(existing->is_data != update->is_data);
467 
468 	spin_lock(&existing->lock);
469 	if (update->must_insert_reserved) {
470 		/* if the extent was freed and then
471 		 * reallocated before the delayed ref
472 		 * entries were processed, we can end up
473 		 * with an existing head ref without
474 		 * the must_insert_reserved flag set.
475 		 * Set it again here
476 		 */
477 		existing->must_insert_reserved = update->must_insert_reserved;
478 
479 		/*
480 		 * update the num_bytes so we make sure the accounting
481 		 * is done correctly
482 		 */
483 		existing->num_bytes = update->num_bytes;
484 
485 	}
486 
487 	if (update->extent_op) {
488 		if (!existing->extent_op) {
489 			existing->extent_op = update->extent_op;
490 		} else {
491 			if (update->extent_op->update_key) {
492 				memcpy(&existing->extent_op->key,
493 				       &update->extent_op->key,
494 				       sizeof(update->extent_op->key));
495 				existing->extent_op->update_key = true;
496 			}
497 			if (update->extent_op->update_flags) {
498 				existing->extent_op->flags_to_set |=
499 					update->extent_op->flags_to_set;
500 				existing->extent_op->update_flags = true;
501 			}
502 			btrfs_free_delayed_extent_op(update->extent_op);
503 		}
504 	}
505 	/*
506 	 * update the reference mod on the head to reflect this new operation,
507 	 * only need the lock for this case cause we could be processing it
508 	 * currently, for refs we just added we know we're a-ok.
509 	 */
510 	old_ref_mod = existing->total_ref_mod;
511 	if (old_ref_mod_ret)
512 		*old_ref_mod_ret = old_ref_mod;
513 	existing->ref_mod += update->ref_mod;
514 	existing->total_ref_mod += update->ref_mod;
515 
516 	/*
517 	 * If we are going to from a positive ref mod to a negative or vice
518 	 * versa we need to make sure to adjust pending_csums accordingly.
519 	 */
520 	if (existing->is_data) {
521 		if (existing->total_ref_mod >= 0 && old_ref_mod < 0)
522 			delayed_refs->pending_csums -= existing->num_bytes;
523 		if (existing->total_ref_mod < 0 && old_ref_mod >= 0)
524 			delayed_refs->pending_csums += existing->num_bytes;
525 	}
526 	spin_unlock(&existing->lock);
527 }
528 
529 static void init_delayed_ref_head(struct btrfs_delayed_ref_head *head_ref,
530 				  struct btrfs_qgroup_extent_record *qrecord,
531 				  u64 bytenr, u64 num_bytes, u64 ref_root,
532 				  u64 reserved, int action, bool is_data,
533 				  bool is_system)
534 {
535 	int count_mod = 1;
536 	int must_insert_reserved = 0;
537 
538 	/* If reserved is provided, it must be a data extent. */
539 	BUG_ON(!is_data && reserved);
540 
541 	/*
542 	 * The head node stores the sum of all the mods, so dropping a ref
543 	 * should drop the sum in the head node by one.
544 	 */
545 	if (action == BTRFS_UPDATE_DELAYED_HEAD)
546 		count_mod = 0;
547 	else if (action == BTRFS_DROP_DELAYED_REF)
548 		count_mod = -1;
549 
550 	/*
551 	 * BTRFS_ADD_DELAYED_EXTENT means that we need to update the reserved
552 	 * accounting when the extent is finally added, or if a later
553 	 * modification deletes the delayed ref without ever inserting the
554 	 * extent into the extent allocation tree.  ref->must_insert_reserved
555 	 * is the flag used to record that accounting mods are required.
556 	 *
557 	 * Once we record must_insert_reserved, switch the action to
558 	 * BTRFS_ADD_DELAYED_REF because other special casing is not required.
559 	 */
560 	if (action == BTRFS_ADD_DELAYED_EXTENT)
561 		must_insert_reserved = 1;
562 	else
563 		must_insert_reserved = 0;
564 
565 	refcount_set(&head_ref->refs, 1);
566 	head_ref->bytenr = bytenr;
567 	head_ref->num_bytes = num_bytes;
568 	head_ref->ref_mod = count_mod;
569 	head_ref->must_insert_reserved = must_insert_reserved;
570 	head_ref->is_data = is_data;
571 	head_ref->is_system = is_system;
572 	head_ref->ref_tree = RB_ROOT;
573 	INIT_LIST_HEAD(&head_ref->ref_add_list);
574 	RB_CLEAR_NODE(&head_ref->href_node);
575 	head_ref->processing = 0;
576 	head_ref->total_ref_mod = count_mod;
577 	head_ref->qgroup_reserved = 0;
578 	head_ref->qgroup_ref_root = 0;
579 	spin_lock_init(&head_ref->lock);
580 	mutex_init(&head_ref->mutex);
581 
582 	if (qrecord) {
583 		if (ref_root && reserved) {
584 			head_ref->qgroup_ref_root = ref_root;
585 			head_ref->qgroup_reserved = reserved;
586 		}
587 
588 		qrecord->bytenr = bytenr;
589 		qrecord->num_bytes = num_bytes;
590 		qrecord->old_roots = NULL;
591 	}
592 }
593 
594 /*
595  * helper function to actually insert a head node into the rbtree.
596  * this does all the dirty work in terms of maintaining the correct
597  * overall modification count.
598  */
599 static noinline struct btrfs_delayed_ref_head *
600 add_delayed_ref_head(struct btrfs_trans_handle *trans,
601 		     struct btrfs_delayed_ref_head *head_ref,
602 		     struct btrfs_qgroup_extent_record *qrecord,
603 		     int action, int *qrecord_inserted_ret,
604 		     int *old_ref_mod, int *new_ref_mod)
605 {
606 	struct btrfs_delayed_ref_head *existing;
607 	struct btrfs_delayed_ref_root *delayed_refs;
608 	int qrecord_inserted = 0;
609 
610 	delayed_refs = &trans->transaction->delayed_refs;
611 
612 	/* Record qgroup extent info if provided */
613 	if (qrecord) {
614 		if (btrfs_qgroup_trace_extent_nolock(trans->fs_info,
615 					delayed_refs, qrecord))
616 			kfree(qrecord);
617 		else
618 			qrecord_inserted = 1;
619 	}
620 
621 	trace_add_delayed_ref_head(trans->fs_info, head_ref, action);
622 
623 	existing = htree_insert(&delayed_refs->href_root,
624 				&head_ref->href_node);
625 	if (existing) {
626 		WARN_ON(qrecord && head_ref->qgroup_ref_root
627 			&& head_ref->qgroup_reserved
628 			&& existing->qgroup_ref_root
629 			&& existing->qgroup_reserved);
630 		update_existing_head_ref(delayed_refs, existing, head_ref,
631 					 old_ref_mod);
632 		/*
633 		 * we've updated the existing ref, free the newly
634 		 * allocated ref
635 		 */
636 		kmem_cache_free(btrfs_delayed_ref_head_cachep, head_ref);
637 		head_ref = existing;
638 	} else {
639 		if (old_ref_mod)
640 			*old_ref_mod = 0;
641 		if (head_ref->is_data && head_ref->ref_mod < 0)
642 			delayed_refs->pending_csums += head_ref->num_bytes;
643 		delayed_refs->num_heads++;
644 		delayed_refs->num_heads_ready++;
645 		atomic_inc(&delayed_refs->num_entries);
646 		trans->delayed_ref_updates++;
647 	}
648 	if (qrecord_inserted_ret)
649 		*qrecord_inserted_ret = qrecord_inserted;
650 	if (new_ref_mod)
651 		*new_ref_mod = head_ref->total_ref_mod;
652 
653 	return head_ref;
654 }
655 
656 /*
657  * init_delayed_ref_common - Initialize the structure which represents a
658  *			     modification to a an extent.
659  *
660  * @fs_info:    Internal to the mounted filesystem mount structure.
661  *
662  * @ref:	The structure which is going to be initialized.
663  *
664  * @bytenr:	The logical address of the extent for which a modification is
665  *		going to be recorded.
666  *
667  * @num_bytes:  Size of the extent whose modification is being recorded.
668  *
669  * @ref_root:	The id of the root where this modification has originated, this
670  *		can be either one of the well-known metadata trees or the
671  *		subvolume id which references this extent.
672  *
673  * @action:	Can be one of BTRFS_ADD_DELAYED_REF/BTRFS_DROP_DELAYED_REF or
674  *		BTRFS_ADD_DELAYED_EXTENT
675  *
676  * @ref_type:	Holds the type of the extent which is being recorded, can be
677  *		one of BTRFS_SHARED_BLOCK_REF_KEY/BTRFS_TREE_BLOCK_REF_KEY
678  *		when recording a metadata extent or BTRFS_SHARED_DATA_REF_KEY/
679  *		BTRFS_EXTENT_DATA_REF_KEY when recording data extent
680  */
681 static void init_delayed_ref_common(struct btrfs_fs_info *fs_info,
682 				    struct btrfs_delayed_ref_node *ref,
683 				    u64 bytenr, u64 num_bytes, u64 ref_root,
684 				    int action, u8 ref_type)
685 {
686 	u64 seq = 0;
687 
688 	if (action == BTRFS_ADD_DELAYED_EXTENT)
689 		action = BTRFS_ADD_DELAYED_REF;
690 
691 	if (is_fstree(ref_root))
692 		seq = atomic64_read(&fs_info->tree_mod_seq);
693 
694 	refcount_set(&ref->refs, 1);
695 	ref->bytenr = bytenr;
696 	ref->num_bytes = num_bytes;
697 	ref->ref_mod = 1;
698 	ref->action = action;
699 	ref->is_head = 0;
700 	ref->in_tree = 1;
701 	ref->seq = seq;
702 	ref->type = ref_type;
703 	RB_CLEAR_NODE(&ref->ref_node);
704 	INIT_LIST_HEAD(&ref->add_list);
705 }
706 
707 /*
708  * add a delayed tree ref.  This does all of the accounting required
709  * to make sure the delayed ref is eventually processed before this
710  * transaction commits.
711  */
712 int btrfs_add_delayed_tree_ref(struct btrfs_fs_info *fs_info,
713 			       struct btrfs_trans_handle *trans,
714 			       u64 bytenr, u64 num_bytes, u64 parent,
715 			       u64 ref_root,  int level, int action,
716 			       struct btrfs_delayed_extent_op *extent_op,
717 			       int *old_ref_mod, int *new_ref_mod)
718 {
719 	struct btrfs_delayed_tree_ref *ref;
720 	struct btrfs_delayed_ref_head *head_ref;
721 	struct btrfs_delayed_ref_root *delayed_refs;
722 	struct btrfs_qgroup_extent_record *record = NULL;
723 	int qrecord_inserted;
724 	bool is_system = (ref_root == BTRFS_CHUNK_TREE_OBJECTID);
725 	int ret;
726 	u8 ref_type;
727 
728 	BUG_ON(extent_op && extent_op->is_data);
729 	ref = kmem_cache_alloc(btrfs_delayed_tree_ref_cachep, GFP_NOFS);
730 	if (!ref)
731 		return -ENOMEM;
732 
733 	if (parent)
734 		ref_type = BTRFS_SHARED_BLOCK_REF_KEY;
735 	else
736 		ref_type = BTRFS_TREE_BLOCK_REF_KEY;
737 	init_delayed_ref_common(fs_info, &ref->node, bytenr, num_bytes,
738 				ref_root, action, ref_type);
739 	ref->root = ref_root;
740 	ref->parent = parent;
741 	ref->level = level;
742 
743 	head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
744 	if (!head_ref)
745 		goto free_ref;
746 
747 	if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) &&
748 	    is_fstree(ref_root)) {
749 		record = kmalloc(sizeof(*record), GFP_NOFS);
750 		if (!record)
751 			goto free_head_ref;
752 	}
753 
754 	init_delayed_ref_head(head_ref, record, bytenr, num_bytes,
755 			      ref_root, 0, action, false, is_system);
756 	head_ref->extent_op = extent_op;
757 
758 	delayed_refs = &trans->transaction->delayed_refs;
759 	spin_lock(&delayed_refs->lock);
760 
761 	/*
762 	 * insert both the head node and the new ref without dropping
763 	 * the spin lock
764 	 */
765 	head_ref = add_delayed_ref_head(trans, head_ref, record,
766 					action, &qrecord_inserted,
767 					old_ref_mod, new_ref_mod);
768 
769 	ret = insert_delayed_ref(trans, delayed_refs, head_ref, &ref->node);
770 	spin_unlock(&delayed_refs->lock);
771 
772 	trace_add_delayed_tree_ref(fs_info, &ref->node, ref,
773 				   action == BTRFS_ADD_DELAYED_EXTENT ?
774 				   BTRFS_ADD_DELAYED_REF : action);
775 	if (ret > 0)
776 		kmem_cache_free(btrfs_delayed_tree_ref_cachep, ref);
777 
778 	if (qrecord_inserted)
779 		btrfs_qgroup_trace_extent_post(fs_info, record);
780 
781 	return 0;
782 
783 free_head_ref:
784 	kmem_cache_free(btrfs_delayed_ref_head_cachep, head_ref);
785 free_ref:
786 	kmem_cache_free(btrfs_delayed_tree_ref_cachep, ref);
787 
788 	return -ENOMEM;
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, u64 reserved, int action,
799 			       int *old_ref_mod, int *new_ref_mod)
800 {
801 	struct btrfs_delayed_data_ref *ref;
802 	struct btrfs_delayed_ref_head *head_ref;
803 	struct btrfs_delayed_ref_root *delayed_refs;
804 	struct btrfs_qgroup_extent_record *record = NULL;
805 	int qrecord_inserted;
806 	int ret;
807 	u8 ref_type;
808 
809 	ref = kmem_cache_alloc(btrfs_delayed_data_ref_cachep, GFP_NOFS);
810 	if (!ref)
811 		return -ENOMEM;
812 
813 	if (parent)
814 	        ref_type = BTRFS_SHARED_DATA_REF_KEY;
815 	else
816 	        ref_type = BTRFS_EXTENT_DATA_REF_KEY;
817 	init_delayed_ref_common(fs_info, &ref->node, bytenr, num_bytes,
818 				ref_root, action, ref_type);
819 	ref->root = ref_root;
820 	ref->parent = parent;
821 	ref->objectid = owner;
822 	ref->offset = offset;
823 
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 (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) &&
832 	    is_fstree(ref_root)) {
833 		record = kmalloc(sizeof(*record), GFP_NOFS);
834 		if (!record) {
835 			kmem_cache_free(btrfs_delayed_data_ref_cachep, ref);
836 			kmem_cache_free(btrfs_delayed_ref_head_cachep,
837 					head_ref);
838 			return -ENOMEM;
839 		}
840 	}
841 
842 	init_delayed_ref_head(head_ref, record, bytenr, num_bytes, ref_root,
843 			      reserved, action, true, false);
844 	head_ref->extent_op = NULL;
845 
846 	delayed_refs = &trans->transaction->delayed_refs;
847 	spin_lock(&delayed_refs->lock);
848 
849 	/*
850 	 * insert both the head node and the new ref without dropping
851 	 * the spin lock
852 	 */
853 	head_ref = add_delayed_ref_head(trans, head_ref, record,
854 					action, &qrecord_inserted,
855 					old_ref_mod, new_ref_mod);
856 
857 	ret = insert_delayed_ref(trans, delayed_refs, head_ref, &ref->node);
858 	spin_unlock(&delayed_refs->lock);
859 
860 	trace_add_delayed_data_ref(trans->fs_info, &ref->node, ref,
861 				   action == BTRFS_ADD_DELAYED_EXTENT ?
862 				   BTRFS_ADD_DELAYED_REF : action);
863 	if (ret > 0)
864 		kmem_cache_free(btrfs_delayed_data_ref_cachep, ref);
865 
866 
867 	if (qrecord_inserted)
868 		return btrfs_qgroup_trace_extent_post(fs_info, record);
869 	return 0;
870 }
871 
872 int btrfs_add_delayed_extent_op(struct btrfs_fs_info *fs_info,
873 				struct btrfs_trans_handle *trans,
874 				u64 bytenr, u64 num_bytes,
875 				struct btrfs_delayed_extent_op *extent_op)
876 {
877 	struct btrfs_delayed_ref_head *head_ref;
878 	struct btrfs_delayed_ref_root *delayed_refs;
879 
880 	head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
881 	if (!head_ref)
882 		return -ENOMEM;
883 
884 	init_delayed_ref_head(head_ref, NULL, bytenr, num_bytes, 0, 0,
885 			      BTRFS_UPDATE_DELAYED_HEAD, extent_op->is_data,
886 			      false);
887 	head_ref->extent_op = extent_op;
888 
889 	delayed_refs = &trans->transaction->delayed_refs;
890 	spin_lock(&delayed_refs->lock);
891 
892 	add_delayed_ref_head(trans, head_ref, NULL, BTRFS_UPDATE_DELAYED_HEAD,
893 			     NULL, NULL, NULL);
894 
895 	spin_unlock(&delayed_refs->lock);
896 	return 0;
897 }
898 
899 /*
900  * this does a simple search for the head node for a given extent.
901  * It must be called with the delayed ref spinlock held, and it returns
902  * the head node if any where found, or NULL if not.
903  */
904 struct btrfs_delayed_ref_head *
905 btrfs_find_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs, u64 bytenr)
906 {
907 	return find_ref_head(&delayed_refs->href_root, bytenr, 0);
908 }
909 
910 void __cold btrfs_delayed_ref_exit(void)
911 {
912 	kmem_cache_destroy(btrfs_delayed_ref_head_cachep);
913 	kmem_cache_destroy(btrfs_delayed_tree_ref_cachep);
914 	kmem_cache_destroy(btrfs_delayed_data_ref_cachep);
915 	kmem_cache_destroy(btrfs_delayed_extent_op_cachep);
916 }
917 
918 int __init btrfs_delayed_ref_init(void)
919 {
920 	btrfs_delayed_ref_head_cachep = kmem_cache_create(
921 				"btrfs_delayed_ref_head",
922 				sizeof(struct btrfs_delayed_ref_head), 0,
923 				SLAB_MEM_SPREAD, NULL);
924 	if (!btrfs_delayed_ref_head_cachep)
925 		goto fail;
926 
927 	btrfs_delayed_tree_ref_cachep = kmem_cache_create(
928 				"btrfs_delayed_tree_ref",
929 				sizeof(struct btrfs_delayed_tree_ref), 0,
930 				SLAB_MEM_SPREAD, NULL);
931 	if (!btrfs_delayed_tree_ref_cachep)
932 		goto fail;
933 
934 	btrfs_delayed_data_ref_cachep = kmem_cache_create(
935 				"btrfs_delayed_data_ref",
936 				sizeof(struct btrfs_delayed_data_ref), 0,
937 				SLAB_MEM_SPREAD, NULL);
938 	if (!btrfs_delayed_data_ref_cachep)
939 		goto fail;
940 
941 	btrfs_delayed_extent_op_cachep = kmem_cache_create(
942 				"btrfs_delayed_extent_op",
943 				sizeof(struct btrfs_delayed_extent_op), 0,
944 				SLAB_MEM_SPREAD, NULL);
945 	if (!btrfs_delayed_extent_op_cachep)
946 		goto fail;
947 
948 	return 0;
949 fail:
950 	btrfs_delayed_ref_exit();
951 	return -ENOMEM;
952 }
953