xref: /openbmc/linux/fs/btrfs/qgroup.c (revision ac73d4bf)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2011 STRATO.  All rights reserved.
4  */
5 
6 #include <linux/sched.h>
7 #include <linux/pagemap.h>
8 #include <linux/writeback.h>
9 #include <linux/blkdev.h>
10 #include <linux/rbtree.h>
11 #include <linux/slab.h>
12 #include <linux/workqueue.h>
13 #include <linux/btrfs.h>
14 #include <linux/sched/mm.h>
15 
16 #include "ctree.h"
17 #include "transaction.h"
18 #include "disk-io.h"
19 #include "locking.h"
20 #include "ulist.h"
21 #include "backref.h"
22 #include "extent_io.h"
23 #include "qgroup.h"
24 #include "block-group.h"
25 #include "sysfs.h"
26 #include "tree-mod-log.h"
27 
28 /*
29  * Helpers to access qgroup reservation
30  *
31  * Callers should ensure the lock context and type are valid
32  */
33 
34 static u64 qgroup_rsv_total(const struct btrfs_qgroup *qgroup)
35 {
36 	u64 ret = 0;
37 	int i;
38 
39 	for (i = 0; i < BTRFS_QGROUP_RSV_LAST; i++)
40 		ret += qgroup->rsv.values[i];
41 
42 	return ret;
43 }
44 
45 #ifdef CONFIG_BTRFS_DEBUG
46 static const char *qgroup_rsv_type_str(enum btrfs_qgroup_rsv_type type)
47 {
48 	if (type == BTRFS_QGROUP_RSV_DATA)
49 		return "data";
50 	if (type == BTRFS_QGROUP_RSV_META_PERTRANS)
51 		return "meta_pertrans";
52 	if (type == BTRFS_QGROUP_RSV_META_PREALLOC)
53 		return "meta_prealloc";
54 	return NULL;
55 }
56 #endif
57 
58 static void qgroup_rsv_add(struct btrfs_fs_info *fs_info,
59 			   struct btrfs_qgroup *qgroup, u64 num_bytes,
60 			   enum btrfs_qgroup_rsv_type type)
61 {
62 	trace_qgroup_update_reserve(fs_info, qgroup, num_bytes, type);
63 	qgroup->rsv.values[type] += num_bytes;
64 }
65 
66 static void qgroup_rsv_release(struct btrfs_fs_info *fs_info,
67 			       struct btrfs_qgroup *qgroup, u64 num_bytes,
68 			       enum btrfs_qgroup_rsv_type type)
69 {
70 	trace_qgroup_update_reserve(fs_info, qgroup, -(s64)num_bytes, type);
71 	if (qgroup->rsv.values[type] >= num_bytes) {
72 		qgroup->rsv.values[type] -= num_bytes;
73 		return;
74 	}
75 #ifdef CONFIG_BTRFS_DEBUG
76 	WARN_RATELIMIT(1,
77 		"qgroup %llu %s reserved space underflow, have %llu to free %llu",
78 		qgroup->qgroupid, qgroup_rsv_type_str(type),
79 		qgroup->rsv.values[type], num_bytes);
80 #endif
81 	qgroup->rsv.values[type] = 0;
82 }
83 
84 static void qgroup_rsv_add_by_qgroup(struct btrfs_fs_info *fs_info,
85 				     struct btrfs_qgroup *dest,
86 				     struct btrfs_qgroup *src)
87 {
88 	int i;
89 
90 	for (i = 0; i < BTRFS_QGROUP_RSV_LAST; i++)
91 		qgroup_rsv_add(fs_info, dest, src->rsv.values[i], i);
92 }
93 
94 static void qgroup_rsv_release_by_qgroup(struct btrfs_fs_info *fs_info,
95 					 struct btrfs_qgroup *dest,
96 					  struct btrfs_qgroup *src)
97 {
98 	int i;
99 
100 	for (i = 0; i < BTRFS_QGROUP_RSV_LAST; i++)
101 		qgroup_rsv_release(fs_info, dest, src->rsv.values[i], i);
102 }
103 
104 static void btrfs_qgroup_update_old_refcnt(struct btrfs_qgroup *qg, u64 seq,
105 					   int mod)
106 {
107 	if (qg->old_refcnt < seq)
108 		qg->old_refcnt = seq;
109 	qg->old_refcnt += mod;
110 }
111 
112 static void btrfs_qgroup_update_new_refcnt(struct btrfs_qgroup *qg, u64 seq,
113 					   int mod)
114 {
115 	if (qg->new_refcnt < seq)
116 		qg->new_refcnt = seq;
117 	qg->new_refcnt += mod;
118 }
119 
120 static inline u64 btrfs_qgroup_get_old_refcnt(struct btrfs_qgroup *qg, u64 seq)
121 {
122 	if (qg->old_refcnt < seq)
123 		return 0;
124 	return qg->old_refcnt - seq;
125 }
126 
127 static inline u64 btrfs_qgroup_get_new_refcnt(struct btrfs_qgroup *qg, u64 seq)
128 {
129 	if (qg->new_refcnt < seq)
130 		return 0;
131 	return qg->new_refcnt - seq;
132 }
133 
134 /*
135  * glue structure to represent the relations between qgroups.
136  */
137 struct btrfs_qgroup_list {
138 	struct list_head next_group;
139 	struct list_head next_member;
140 	struct btrfs_qgroup *group;
141 	struct btrfs_qgroup *member;
142 };
143 
144 static inline u64 qgroup_to_aux(struct btrfs_qgroup *qg)
145 {
146 	return (u64)(uintptr_t)qg;
147 }
148 
149 static inline struct btrfs_qgroup* unode_aux_to_qgroup(struct ulist_node *n)
150 {
151 	return (struct btrfs_qgroup *)(uintptr_t)n->aux;
152 }
153 
154 static int
155 qgroup_rescan_init(struct btrfs_fs_info *fs_info, u64 progress_objectid,
156 		   int init_flags);
157 static void qgroup_rescan_zero_tracking(struct btrfs_fs_info *fs_info);
158 
159 /* must be called with qgroup_ioctl_lock held */
160 static struct btrfs_qgroup *find_qgroup_rb(struct btrfs_fs_info *fs_info,
161 					   u64 qgroupid)
162 {
163 	struct rb_node *n = fs_info->qgroup_tree.rb_node;
164 	struct btrfs_qgroup *qgroup;
165 
166 	while (n) {
167 		qgroup = rb_entry(n, struct btrfs_qgroup, node);
168 		if (qgroup->qgroupid < qgroupid)
169 			n = n->rb_left;
170 		else if (qgroup->qgroupid > qgroupid)
171 			n = n->rb_right;
172 		else
173 			return qgroup;
174 	}
175 	return NULL;
176 }
177 
178 /* must be called with qgroup_lock held */
179 static struct btrfs_qgroup *add_qgroup_rb(struct btrfs_fs_info *fs_info,
180 					  u64 qgroupid)
181 {
182 	struct rb_node **p = &fs_info->qgroup_tree.rb_node;
183 	struct rb_node *parent = NULL;
184 	struct btrfs_qgroup *qgroup;
185 
186 	while (*p) {
187 		parent = *p;
188 		qgroup = rb_entry(parent, struct btrfs_qgroup, node);
189 
190 		if (qgroup->qgroupid < qgroupid)
191 			p = &(*p)->rb_left;
192 		else if (qgroup->qgroupid > qgroupid)
193 			p = &(*p)->rb_right;
194 		else
195 			return qgroup;
196 	}
197 
198 	qgroup = kzalloc(sizeof(*qgroup), GFP_ATOMIC);
199 	if (!qgroup)
200 		return ERR_PTR(-ENOMEM);
201 
202 	qgroup->qgroupid = qgroupid;
203 	INIT_LIST_HEAD(&qgroup->groups);
204 	INIT_LIST_HEAD(&qgroup->members);
205 	INIT_LIST_HEAD(&qgroup->dirty);
206 
207 	rb_link_node(&qgroup->node, parent, p);
208 	rb_insert_color(&qgroup->node, &fs_info->qgroup_tree);
209 
210 	return qgroup;
211 }
212 
213 static void __del_qgroup_rb(struct btrfs_fs_info *fs_info,
214 			    struct btrfs_qgroup *qgroup)
215 {
216 	struct btrfs_qgroup_list *list;
217 
218 	list_del(&qgroup->dirty);
219 	while (!list_empty(&qgroup->groups)) {
220 		list = list_first_entry(&qgroup->groups,
221 					struct btrfs_qgroup_list, next_group);
222 		list_del(&list->next_group);
223 		list_del(&list->next_member);
224 		kfree(list);
225 	}
226 
227 	while (!list_empty(&qgroup->members)) {
228 		list = list_first_entry(&qgroup->members,
229 					struct btrfs_qgroup_list, next_member);
230 		list_del(&list->next_group);
231 		list_del(&list->next_member);
232 		kfree(list);
233 	}
234 }
235 
236 /* must be called with qgroup_lock held */
237 static int del_qgroup_rb(struct btrfs_fs_info *fs_info, u64 qgroupid)
238 {
239 	struct btrfs_qgroup *qgroup = find_qgroup_rb(fs_info, qgroupid);
240 
241 	if (!qgroup)
242 		return -ENOENT;
243 
244 	rb_erase(&qgroup->node, &fs_info->qgroup_tree);
245 	__del_qgroup_rb(fs_info, qgroup);
246 	return 0;
247 }
248 
249 /*
250  * Add relation specified by two qgroups.
251  *
252  * Must be called with qgroup_lock held.
253  *
254  * Return: 0        on success
255  *         -ENOENT  if one of the qgroups is NULL
256  *         <0       other errors
257  */
258 static int __add_relation_rb(struct btrfs_qgroup *member, struct btrfs_qgroup *parent)
259 {
260 	struct btrfs_qgroup_list *list;
261 
262 	if (!member || !parent)
263 		return -ENOENT;
264 
265 	list = kzalloc(sizeof(*list), GFP_ATOMIC);
266 	if (!list)
267 		return -ENOMEM;
268 
269 	list->group = parent;
270 	list->member = member;
271 	list_add_tail(&list->next_group, &member->groups);
272 	list_add_tail(&list->next_member, &parent->members);
273 
274 	return 0;
275 }
276 
277 /*
278  * Add relation specified by two qgroup ids.
279  *
280  * Must be called with qgroup_lock held.
281  *
282  * Return: 0        on success
283  *         -ENOENT  if one of the ids does not exist
284  *         <0       other errors
285  */
286 static int add_relation_rb(struct btrfs_fs_info *fs_info, u64 memberid, u64 parentid)
287 {
288 	struct btrfs_qgroup *member;
289 	struct btrfs_qgroup *parent;
290 
291 	member = find_qgroup_rb(fs_info, memberid);
292 	parent = find_qgroup_rb(fs_info, parentid);
293 
294 	return __add_relation_rb(member, parent);
295 }
296 
297 /* Must be called with qgroup_lock held */
298 static int del_relation_rb(struct btrfs_fs_info *fs_info,
299 			   u64 memberid, u64 parentid)
300 {
301 	struct btrfs_qgroup *member;
302 	struct btrfs_qgroup *parent;
303 	struct btrfs_qgroup_list *list;
304 
305 	member = find_qgroup_rb(fs_info, memberid);
306 	parent = find_qgroup_rb(fs_info, parentid);
307 	if (!member || !parent)
308 		return -ENOENT;
309 
310 	list_for_each_entry(list, &member->groups, next_group) {
311 		if (list->group == parent) {
312 			list_del(&list->next_group);
313 			list_del(&list->next_member);
314 			kfree(list);
315 			return 0;
316 		}
317 	}
318 	return -ENOENT;
319 }
320 
321 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
322 int btrfs_verify_qgroup_counts(struct btrfs_fs_info *fs_info, u64 qgroupid,
323 			       u64 rfer, u64 excl)
324 {
325 	struct btrfs_qgroup *qgroup;
326 
327 	qgroup = find_qgroup_rb(fs_info, qgroupid);
328 	if (!qgroup)
329 		return -EINVAL;
330 	if (qgroup->rfer != rfer || qgroup->excl != excl)
331 		return -EINVAL;
332 	return 0;
333 }
334 #endif
335 
336 static void qgroup_mark_inconsistent(struct btrfs_fs_info *fs_info)
337 {
338 	fs_info->qgroup_flags |= (BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT |
339 				  BTRFS_QGROUP_RUNTIME_FLAG_CANCEL_RESCAN |
340 				  BTRFS_QGROUP_RUNTIME_FLAG_NO_ACCOUNTING);
341 }
342 
343 /*
344  * The full config is read in one go, only called from open_ctree()
345  * It doesn't use any locking, as at this point we're still single-threaded
346  */
347 int btrfs_read_qgroup_config(struct btrfs_fs_info *fs_info)
348 {
349 	struct btrfs_key key;
350 	struct btrfs_key found_key;
351 	struct btrfs_root *quota_root = fs_info->quota_root;
352 	struct btrfs_path *path = NULL;
353 	struct extent_buffer *l;
354 	int slot;
355 	int ret = 0;
356 	u64 flags = 0;
357 	u64 rescan_progress = 0;
358 
359 	if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
360 		return 0;
361 
362 	fs_info->qgroup_ulist = ulist_alloc(GFP_KERNEL);
363 	if (!fs_info->qgroup_ulist) {
364 		ret = -ENOMEM;
365 		goto out;
366 	}
367 
368 	path = btrfs_alloc_path();
369 	if (!path) {
370 		ret = -ENOMEM;
371 		goto out;
372 	}
373 
374 	ret = btrfs_sysfs_add_qgroups(fs_info);
375 	if (ret < 0)
376 		goto out;
377 	/* default this to quota off, in case no status key is found */
378 	fs_info->qgroup_flags = 0;
379 
380 	/*
381 	 * pass 1: read status, all qgroup infos and limits
382 	 */
383 	key.objectid = 0;
384 	key.type = 0;
385 	key.offset = 0;
386 	ret = btrfs_search_slot_for_read(quota_root, &key, path, 1, 1);
387 	if (ret)
388 		goto out;
389 
390 	while (1) {
391 		struct btrfs_qgroup *qgroup;
392 
393 		slot = path->slots[0];
394 		l = path->nodes[0];
395 		btrfs_item_key_to_cpu(l, &found_key, slot);
396 
397 		if (found_key.type == BTRFS_QGROUP_STATUS_KEY) {
398 			struct btrfs_qgroup_status_item *ptr;
399 
400 			ptr = btrfs_item_ptr(l, slot,
401 					     struct btrfs_qgroup_status_item);
402 
403 			if (btrfs_qgroup_status_version(l, ptr) !=
404 			    BTRFS_QGROUP_STATUS_VERSION) {
405 				btrfs_err(fs_info,
406 				 "old qgroup version, quota disabled");
407 				goto out;
408 			}
409 			if (btrfs_qgroup_status_generation(l, ptr) !=
410 			    fs_info->generation) {
411 				qgroup_mark_inconsistent(fs_info);
412 				btrfs_err(fs_info,
413 					"qgroup generation mismatch, marked as inconsistent");
414 			}
415 			fs_info->qgroup_flags = btrfs_qgroup_status_flags(l,
416 									  ptr);
417 			rescan_progress = btrfs_qgroup_status_rescan(l, ptr);
418 			goto next1;
419 		}
420 
421 		if (found_key.type != BTRFS_QGROUP_INFO_KEY &&
422 		    found_key.type != BTRFS_QGROUP_LIMIT_KEY)
423 			goto next1;
424 
425 		qgroup = find_qgroup_rb(fs_info, found_key.offset);
426 		if ((qgroup && found_key.type == BTRFS_QGROUP_INFO_KEY) ||
427 		    (!qgroup && found_key.type == BTRFS_QGROUP_LIMIT_KEY)) {
428 			btrfs_err(fs_info, "inconsistent qgroup config");
429 			qgroup_mark_inconsistent(fs_info);
430 		}
431 		if (!qgroup) {
432 			qgroup = add_qgroup_rb(fs_info, found_key.offset);
433 			if (IS_ERR(qgroup)) {
434 				ret = PTR_ERR(qgroup);
435 				goto out;
436 			}
437 		}
438 		ret = btrfs_sysfs_add_one_qgroup(fs_info, qgroup);
439 		if (ret < 0)
440 			goto out;
441 
442 		switch (found_key.type) {
443 		case BTRFS_QGROUP_INFO_KEY: {
444 			struct btrfs_qgroup_info_item *ptr;
445 
446 			ptr = btrfs_item_ptr(l, slot,
447 					     struct btrfs_qgroup_info_item);
448 			qgroup->rfer = btrfs_qgroup_info_rfer(l, ptr);
449 			qgroup->rfer_cmpr = btrfs_qgroup_info_rfer_cmpr(l, ptr);
450 			qgroup->excl = btrfs_qgroup_info_excl(l, ptr);
451 			qgroup->excl_cmpr = btrfs_qgroup_info_excl_cmpr(l, ptr);
452 			/* generation currently unused */
453 			break;
454 		}
455 		case BTRFS_QGROUP_LIMIT_KEY: {
456 			struct btrfs_qgroup_limit_item *ptr;
457 
458 			ptr = btrfs_item_ptr(l, slot,
459 					     struct btrfs_qgroup_limit_item);
460 			qgroup->lim_flags = btrfs_qgroup_limit_flags(l, ptr);
461 			qgroup->max_rfer = btrfs_qgroup_limit_max_rfer(l, ptr);
462 			qgroup->max_excl = btrfs_qgroup_limit_max_excl(l, ptr);
463 			qgroup->rsv_rfer = btrfs_qgroup_limit_rsv_rfer(l, ptr);
464 			qgroup->rsv_excl = btrfs_qgroup_limit_rsv_excl(l, ptr);
465 			break;
466 		}
467 		}
468 next1:
469 		ret = btrfs_next_item(quota_root, path);
470 		if (ret < 0)
471 			goto out;
472 		if (ret)
473 			break;
474 	}
475 	btrfs_release_path(path);
476 
477 	/*
478 	 * pass 2: read all qgroup relations
479 	 */
480 	key.objectid = 0;
481 	key.type = BTRFS_QGROUP_RELATION_KEY;
482 	key.offset = 0;
483 	ret = btrfs_search_slot_for_read(quota_root, &key, path, 1, 0);
484 	if (ret)
485 		goto out;
486 	while (1) {
487 		slot = path->slots[0];
488 		l = path->nodes[0];
489 		btrfs_item_key_to_cpu(l, &found_key, slot);
490 
491 		if (found_key.type != BTRFS_QGROUP_RELATION_KEY)
492 			goto next2;
493 
494 		if (found_key.objectid > found_key.offset) {
495 			/* parent <- member, not needed to build config */
496 			/* FIXME should we omit the key completely? */
497 			goto next2;
498 		}
499 
500 		ret = add_relation_rb(fs_info, found_key.objectid,
501 				      found_key.offset);
502 		if (ret == -ENOENT) {
503 			btrfs_warn(fs_info,
504 				"orphan qgroup relation 0x%llx->0x%llx",
505 				found_key.objectid, found_key.offset);
506 			ret = 0;	/* ignore the error */
507 		}
508 		if (ret)
509 			goto out;
510 next2:
511 		ret = btrfs_next_item(quota_root, path);
512 		if (ret < 0)
513 			goto out;
514 		if (ret)
515 			break;
516 	}
517 out:
518 	btrfs_free_path(path);
519 	fs_info->qgroup_flags |= flags;
520 	if (!(fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_ON))
521 		clear_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags);
522 	else if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN &&
523 		 ret >= 0)
524 		ret = qgroup_rescan_init(fs_info, rescan_progress, 0);
525 
526 	if (ret < 0) {
527 		ulist_free(fs_info->qgroup_ulist);
528 		fs_info->qgroup_ulist = NULL;
529 		fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
530 		btrfs_sysfs_del_qgroups(fs_info);
531 	}
532 
533 	return ret < 0 ? ret : 0;
534 }
535 
536 /*
537  * Called in close_ctree() when quota is still enabled.  This verifies we don't
538  * leak some reserved space.
539  *
540  * Return false if no reserved space is left.
541  * Return true if some reserved space is leaked.
542  */
543 bool btrfs_check_quota_leak(struct btrfs_fs_info *fs_info)
544 {
545 	struct rb_node *node;
546 	bool ret = false;
547 
548 	if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
549 		return ret;
550 	/*
551 	 * Since we're unmounting, there is no race and no need to grab qgroup
552 	 * lock.  And here we don't go post-order to provide a more user
553 	 * friendly sorted result.
554 	 */
555 	for (node = rb_first(&fs_info->qgroup_tree); node; node = rb_next(node)) {
556 		struct btrfs_qgroup *qgroup;
557 		int i;
558 
559 		qgroup = rb_entry(node, struct btrfs_qgroup, node);
560 		for (i = 0; i < BTRFS_QGROUP_RSV_LAST; i++) {
561 			if (qgroup->rsv.values[i]) {
562 				ret = true;
563 				btrfs_warn(fs_info,
564 		"qgroup %hu/%llu has unreleased space, type %d rsv %llu",
565 				   btrfs_qgroup_level(qgroup->qgroupid),
566 				   btrfs_qgroup_subvolid(qgroup->qgroupid),
567 				   i, qgroup->rsv.values[i]);
568 			}
569 		}
570 	}
571 	return ret;
572 }
573 
574 /*
575  * This is called from close_ctree() or open_ctree() or btrfs_quota_disable(),
576  * first two are in single-threaded paths.And for the third one, we have set
577  * quota_root to be null with qgroup_lock held before, so it is safe to clean
578  * up the in-memory structures without qgroup_lock held.
579  */
580 void btrfs_free_qgroup_config(struct btrfs_fs_info *fs_info)
581 {
582 	struct rb_node *n;
583 	struct btrfs_qgroup *qgroup;
584 
585 	while ((n = rb_first(&fs_info->qgroup_tree))) {
586 		qgroup = rb_entry(n, struct btrfs_qgroup, node);
587 		rb_erase(n, &fs_info->qgroup_tree);
588 		__del_qgroup_rb(fs_info, qgroup);
589 		btrfs_sysfs_del_one_qgroup(fs_info, qgroup);
590 		kfree(qgroup);
591 	}
592 	/*
593 	 * We call btrfs_free_qgroup_config() when unmounting
594 	 * filesystem and disabling quota, so we set qgroup_ulist
595 	 * to be null here to avoid double free.
596 	 */
597 	ulist_free(fs_info->qgroup_ulist);
598 	fs_info->qgroup_ulist = NULL;
599 	btrfs_sysfs_del_qgroups(fs_info);
600 }
601 
602 static int add_qgroup_relation_item(struct btrfs_trans_handle *trans, u64 src,
603 				    u64 dst)
604 {
605 	int ret;
606 	struct btrfs_root *quota_root = trans->fs_info->quota_root;
607 	struct btrfs_path *path;
608 	struct btrfs_key key;
609 
610 	path = btrfs_alloc_path();
611 	if (!path)
612 		return -ENOMEM;
613 
614 	key.objectid = src;
615 	key.type = BTRFS_QGROUP_RELATION_KEY;
616 	key.offset = dst;
617 
618 	ret = btrfs_insert_empty_item(trans, quota_root, path, &key, 0);
619 
620 	btrfs_mark_buffer_dirty(path->nodes[0]);
621 
622 	btrfs_free_path(path);
623 	return ret;
624 }
625 
626 static int del_qgroup_relation_item(struct btrfs_trans_handle *trans, u64 src,
627 				    u64 dst)
628 {
629 	int ret;
630 	struct btrfs_root *quota_root = trans->fs_info->quota_root;
631 	struct btrfs_path *path;
632 	struct btrfs_key key;
633 
634 	path = btrfs_alloc_path();
635 	if (!path)
636 		return -ENOMEM;
637 
638 	key.objectid = src;
639 	key.type = BTRFS_QGROUP_RELATION_KEY;
640 	key.offset = dst;
641 
642 	ret = btrfs_search_slot(trans, quota_root, &key, path, -1, 1);
643 	if (ret < 0)
644 		goto out;
645 
646 	if (ret > 0) {
647 		ret = -ENOENT;
648 		goto out;
649 	}
650 
651 	ret = btrfs_del_item(trans, quota_root, path);
652 out:
653 	btrfs_free_path(path);
654 	return ret;
655 }
656 
657 static int add_qgroup_item(struct btrfs_trans_handle *trans,
658 			   struct btrfs_root *quota_root, u64 qgroupid)
659 {
660 	int ret;
661 	struct btrfs_path *path;
662 	struct btrfs_qgroup_info_item *qgroup_info;
663 	struct btrfs_qgroup_limit_item *qgroup_limit;
664 	struct extent_buffer *leaf;
665 	struct btrfs_key key;
666 
667 	if (btrfs_is_testing(quota_root->fs_info))
668 		return 0;
669 
670 	path = btrfs_alloc_path();
671 	if (!path)
672 		return -ENOMEM;
673 
674 	key.objectid = 0;
675 	key.type = BTRFS_QGROUP_INFO_KEY;
676 	key.offset = qgroupid;
677 
678 	/*
679 	 * Avoid a transaction abort by catching -EEXIST here. In that
680 	 * case, we proceed by re-initializing the existing structure
681 	 * on disk.
682 	 */
683 
684 	ret = btrfs_insert_empty_item(trans, quota_root, path, &key,
685 				      sizeof(*qgroup_info));
686 	if (ret && ret != -EEXIST)
687 		goto out;
688 
689 	leaf = path->nodes[0];
690 	qgroup_info = btrfs_item_ptr(leaf, path->slots[0],
691 				 struct btrfs_qgroup_info_item);
692 	btrfs_set_qgroup_info_generation(leaf, qgroup_info, trans->transid);
693 	btrfs_set_qgroup_info_rfer(leaf, qgroup_info, 0);
694 	btrfs_set_qgroup_info_rfer_cmpr(leaf, qgroup_info, 0);
695 	btrfs_set_qgroup_info_excl(leaf, qgroup_info, 0);
696 	btrfs_set_qgroup_info_excl_cmpr(leaf, qgroup_info, 0);
697 
698 	btrfs_mark_buffer_dirty(leaf);
699 
700 	btrfs_release_path(path);
701 
702 	key.type = BTRFS_QGROUP_LIMIT_KEY;
703 	ret = btrfs_insert_empty_item(trans, quota_root, path, &key,
704 				      sizeof(*qgroup_limit));
705 	if (ret && ret != -EEXIST)
706 		goto out;
707 
708 	leaf = path->nodes[0];
709 	qgroup_limit = btrfs_item_ptr(leaf, path->slots[0],
710 				  struct btrfs_qgroup_limit_item);
711 	btrfs_set_qgroup_limit_flags(leaf, qgroup_limit, 0);
712 	btrfs_set_qgroup_limit_max_rfer(leaf, qgroup_limit, 0);
713 	btrfs_set_qgroup_limit_max_excl(leaf, qgroup_limit, 0);
714 	btrfs_set_qgroup_limit_rsv_rfer(leaf, qgroup_limit, 0);
715 	btrfs_set_qgroup_limit_rsv_excl(leaf, qgroup_limit, 0);
716 
717 	btrfs_mark_buffer_dirty(leaf);
718 
719 	ret = 0;
720 out:
721 	btrfs_free_path(path);
722 	return ret;
723 }
724 
725 static int del_qgroup_item(struct btrfs_trans_handle *trans, u64 qgroupid)
726 {
727 	int ret;
728 	struct btrfs_root *quota_root = trans->fs_info->quota_root;
729 	struct btrfs_path *path;
730 	struct btrfs_key key;
731 
732 	path = btrfs_alloc_path();
733 	if (!path)
734 		return -ENOMEM;
735 
736 	key.objectid = 0;
737 	key.type = BTRFS_QGROUP_INFO_KEY;
738 	key.offset = qgroupid;
739 	ret = btrfs_search_slot(trans, quota_root, &key, path, -1, 1);
740 	if (ret < 0)
741 		goto out;
742 
743 	if (ret > 0) {
744 		ret = -ENOENT;
745 		goto out;
746 	}
747 
748 	ret = btrfs_del_item(trans, quota_root, path);
749 	if (ret)
750 		goto out;
751 
752 	btrfs_release_path(path);
753 
754 	key.type = BTRFS_QGROUP_LIMIT_KEY;
755 	ret = btrfs_search_slot(trans, quota_root, &key, path, -1, 1);
756 	if (ret < 0)
757 		goto out;
758 
759 	if (ret > 0) {
760 		ret = -ENOENT;
761 		goto out;
762 	}
763 
764 	ret = btrfs_del_item(trans, quota_root, path);
765 
766 out:
767 	btrfs_free_path(path);
768 	return ret;
769 }
770 
771 static int update_qgroup_limit_item(struct btrfs_trans_handle *trans,
772 				    struct btrfs_qgroup *qgroup)
773 {
774 	struct btrfs_root *quota_root = trans->fs_info->quota_root;
775 	struct btrfs_path *path;
776 	struct btrfs_key key;
777 	struct extent_buffer *l;
778 	struct btrfs_qgroup_limit_item *qgroup_limit;
779 	int ret;
780 	int slot;
781 
782 	key.objectid = 0;
783 	key.type = BTRFS_QGROUP_LIMIT_KEY;
784 	key.offset = qgroup->qgroupid;
785 
786 	path = btrfs_alloc_path();
787 	if (!path)
788 		return -ENOMEM;
789 
790 	ret = btrfs_search_slot(trans, quota_root, &key, path, 0, 1);
791 	if (ret > 0)
792 		ret = -ENOENT;
793 
794 	if (ret)
795 		goto out;
796 
797 	l = path->nodes[0];
798 	slot = path->slots[0];
799 	qgroup_limit = btrfs_item_ptr(l, slot, struct btrfs_qgroup_limit_item);
800 	btrfs_set_qgroup_limit_flags(l, qgroup_limit, qgroup->lim_flags);
801 	btrfs_set_qgroup_limit_max_rfer(l, qgroup_limit, qgroup->max_rfer);
802 	btrfs_set_qgroup_limit_max_excl(l, qgroup_limit, qgroup->max_excl);
803 	btrfs_set_qgroup_limit_rsv_rfer(l, qgroup_limit, qgroup->rsv_rfer);
804 	btrfs_set_qgroup_limit_rsv_excl(l, qgroup_limit, qgroup->rsv_excl);
805 
806 	btrfs_mark_buffer_dirty(l);
807 
808 out:
809 	btrfs_free_path(path);
810 	return ret;
811 }
812 
813 static int update_qgroup_info_item(struct btrfs_trans_handle *trans,
814 				   struct btrfs_qgroup *qgroup)
815 {
816 	struct btrfs_fs_info *fs_info = trans->fs_info;
817 	struct btrfs_root *quota_root = fs_info->quota_root;
818 	struct btrfs_path *path;
819 	struct btrfs_key key;
820 	struct extent_buffer *l;
821 	struct btrfs_qgroup_info_item *qgroup_info;
822 	int ret;
823 	int slot;
824 
825 	if (btrfs_is_testing(fs_info))
826 		return 0;
827 
828 	key.objectid = 0;
829 	key.type = BTRFS_QGROUP_INFO_KEY;
830 	key.offset = qgroup->qgroupid;
831 
832 	path = btrfs_alloc_path();
833 	if (!path)
834 		return -ENOMEM;
835 
836 	ret = btrfs_search_slot(trans, quota_root, &key, path, 0, 1);
837 	if (ret > 0)
838 		ret = -ENOENT;
839 
840 	if (ret)
841 		goto out;
842 
843 	l = path->nodes[0];
844 	slot = path->slots[0];
845 	qgroup_info = btrfs_item_ptr(l, slot, struct btrfs_qgroup_info_item);
846 	btrfs_set_qgroup_info_generation(l, qgroup_info, trans->transid);
847 	btrfs_set_qgroup_info_rfer(l, qgroup_info, qgroup->rfer);
848 	btrfs_set_qgroup_info_rfer_cmpr(l, qgroup_info, qgroup->rfer_cmpr);
849 	btrfs_set_qgroup_info_excl(l, qgroup_info, qgroup->excl);
850 	btrfs_set_qgroup_info_excl_cmpr(l, qgroup_info, qgroup->excl_cmpr);
851 
852 	btrfs_mark_buffer_dirty(l);
853 
854 out:
855 	btrfs_free_path(path);
856 	return ret;
857 }
858 
859 static int update_qgroup_status_item(struct btrfs_trans_handle *trans)
860 {
861 	struct btrfs_fs_info *fs_info = trans->fs_info;
862 	struct btrfs_root *quota_root = fs_info->quota_root;
863 	struct btrfs_path *path;
864 	struct btrfs_key key;
865 	struct extent_buffer *l;
866 	struct btrfs_qgroup_status_item *ptr;
867 	int ret;
868 	int slot;
869 
870 	key.objectid = 0;
871 	key.type = BTRFS_QGROUP_STATUS_KEY;
872 	key.offset = 0;
873 
874 	path = btrfs_alloc_path();
875 	if (!path)
876 		return -ENOMEM;
877 
878 	ret = btrfs_search_slot(trans, quota_root, &key, path, 0, 1);
879 	if (ret > 0)
880 		ret = -ENOENT;
881 
882 	if (ret)
883 		goto out;
884 
885 	l = path->nodes[0];
886 	slot = path->slots[0];
887 	ptr = btrfs_item_ptr(l, slot, struct btrfs_qgroup_status_item);
888 	btrfs_set_qgroup_status_flags(l, ptr, fs_info->qgroup_flags &
889 				      BTRFS_QGROUP_STATUS_FLAGS_MASK);
890 	btrfs_set_qgroup_status_generation(l, ptr, trans->transid);
891 	btrfs_set_qgroup_status_rescan(l, ptr,
892 				fs_info->qgroup_rescan_progress.objectid);
893 
894 	btrfs_mark_buffer_dirty(l);
895 
896 out:
897 	btrfs_free_path(path);
898 	return ret;
899 }
900 
901 /*
902  * called with qgroup_lock held
903  */
904 static int btrfs_clean_quota_tree(struct btrfs_trans_handle *trans,
905 				  struct btrfs_root *root)
906 {
907 	struct btrfs_path *path;
908 	struct btrfs_key key;
909 	struct extent_buffer *leaf = NULL;
910 	int ret;
911 	int nr = 0;
912 
913 	path = btrfs_alloc_path();
914 	if (!path)
915 		return -ENOMEM;
916 
917 	key.objectid = 0;
918 	key.offset = 0;
919 	key.type = 0;
920 
921 	while (1) {
922 		ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
923 		if (ret < 0)
924 			goto out;
925 		leaf = path->nodes[0];
926 		nr = btrfs_header_nritems(leaf);
927 		if (!nr)
928 			break;
929 		/*
930 		 * delete the leaf one by one
931 		 * since the whole tree is going
932 		 * to be deleted.
933 		 */
934 		path->slots[0] = 0;
935 		ret = btrfs_del_items(trans, root, path, 0, nr);
936 		if (ret)
937 			goto out;
938 
939 		btrfs_release_path(path);
940 	}
941 	ret = 0;
942 out:
943 	btrfs_free_path(path);
944 	return ret;
945 }
946 
947 int btrfs_quota_enable(struct btrfs_fs_info *fs_info)
948 {
949 	struct btrfs_root *quota_root;
950 	struct btrfs_root *tree_root = fs_info->tree_root;
951 	struct btrfs_path *path = NULL;
952 	struct btrfs_qgroup_status_item *ptr;
953 	struct extent_buffer *leaf;
954 	struct btrfs_key key;
955 	struct btrfs_key found_key;
956 	struct btrfs_qgroup *qgroup = NULL;
957 	struct btrfs_trans_handle *trans = NULL;
958 	struct ulist *ulist = NULL;
959 	int ret = 0;
960 	int slot;
961 
962 	/*
963 	 * We need to have subvol_sem write locked, to prevent races between
964 	 * concurrent tasks trying to enable quotas, because we will unlock
965 	 * and relock qgroup_ioctl_lock before setting fs_info->quota_root
966 	 * and before setting BTRFS_FS_QUOTA_ENABLED.
967 	 */
968 	lockdep_assert_held_write(&fs_info->subvol_sem);
969 
970 	if (btrfs_fs_incompat(fs_info, EXTENT_TREE_V2)) {
971 		btrfs_err(fs_info,
972 			  "qgroups are currently unsupported in extent tree v2");
973 		return -EINVAL;
974 	}
975 
976 	mutex_lock(&fs_info->qgroup_ioctl_lock);
977 	if (fs_info->quota_root)
978 		goto out;
979 
980 	ulist = ulist_alloc(GFP_KERNEL);
981 	if (!ulist) {
982 		ret = -ENOMEM;
983 		goto out;
984 	}
985 
986 	ret = btrfs_sysfs_add_qgroups(fs_info);
987 	if (ret < 0)
988 		goto out;
989 
990 	/*
991 	 * Unlock qgroup_ioctl_lock before starting the transaction. This is to
992 	 * avoid lock acquisition inversion problems (reported by lockdep) between
993 	 * qgroup_ioctl_lock and the vfs freeze semaphores, acquired when we
994 	 * start a transaction.
995 	 * After we started the transaction lock qgroup_ioctl_lock again and
996 	 * check if someone else created the quota root in the meanwhile. If so,
997 	 * just return success and release the transaction handle.
998 	 *
999 	 * Also we don't need to worry about someone else calling
1000 	 * btrfs_sysfs_add_qgroups() after we unlock and getting an error because
1001 	 * that function returns 0 (success) when the sysfs entries already exist.
1002 	 */
1003 	mutex_unlock(&fs_info->qgroup_ioctl_lock);
1004 
1005 	/*
1006 	 * 1 for quota root item
1007 	 * 1 for BTRFS_QGROUP_STATUS item
1008 	 *
1009 	 * Yet we also need 2*n items for a QGROUP_INFO/QGROUP_LIMIT items
1010 	 * per subvolume. However those are not currently reserved since it
1011 	 * would be a lot of overkill.
1012 	 */
1013 	trans = btrfs_start_transaction(tree_root, 2);
1014 
1015 	mutex_lock(&fs_info->qgroup_ioctl_lock);
1016 	if (IS_ERR(trans)) {
1017 		ret = PTR_ERR(trans);
1018 		trans = NULL;
1019 		goto out;
1020 	}
1021 
1022 	if (fs_info->quota_root)
1023 		goto out;
1024 
1025 	fs_info->qgroup_ulist = ulist;
1026 	ulist = NULL;
1027 
1028 	/*
1029 	 * initially create the quota tree
1030 	 */
1031 	quota_root = btrfs_create_tree(trans, BTRFS_QUOTA_TREE_OBJECTID);
1032 	if (IS_ERR(quota_root)) {
1033 		ret =  PTR_ERR(quota_root);
1034 		btrfs_abort_transaction(trans, ret);
1035 		goto out;
1036 	}
1037 
1038 	path = btrfs_alloc_path();
1039 	if (!path) {
1040 		ret = -ENOMEM;
1041 		btrfs_abort_transaction(trans, ret);
1042 		goto out_free_root;
1043 	}
1044 
1045 	key.objectid = 0;
1046 	key.type = BTRFS_QGROUP_STATUS_KEY;
1047 	key.offset = 0;
1048 
1049 	ret = btrfs_insert_empty_item(trans, quota_root, path, &key,
1050 				      sizeof(*ptr));
1051 	if (ret) {
1052 		btrfs_abort_transaction(trans, ret);
1053 		goto out_free_path;
1054 	}
1055 
1056 	leaf = path->nodes[0];
1057 	ptr = btrfs_item_ptr(leaf, path->slots[0],
1058 				 struct btrfs_qgroup_status_item);
1059 	btrfs_set_qgroup_status_generation(leaf, ptr, trans->transid);
1060 	btrfs_set_qgroup_status_version(leaf, ptr, BTRFS_QGROUP_STATUS_VERSION);
1061 	fs_info->qgroup_flags = BTRFS_QGROUP_STATUS_FLAG_ON |
1062 				BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
1063 	btrfs_set_qgroup_status_flags(leaf, ptr, fs_info->qgroup_flags &
1064 				      BTRFS_QGROUP_STATUS_FLAGS_MASK);
1065 	btrfs_set_qgroup_status_rescan(leaf, ptr, 0);
1066 
1067 	btrfs_mark_buffer_dirty(leaf);
1068 
1069 	key.objectid = 0;
1070 	key.type = BTRFS_ROOT_REF_KEY;
1071 	key.offset = 0;
1072 
1073 	btrfs_release_path(path);
1074 	ret = btrfs_search_slot_for_read(tree_root, &key, path, 1, 0);
1075 	if (ret > 0)
1076 		goto out_add_root;
1077 	if (ret < 0) {
1078 		btrfs_abort_transaction(trans, ret);
1079 		goto out_free_path;
1080 	}
1081 
1082 	while (1) {
1083 		slot = path->slots[0];
1084 		leaf = path->nodes[0];
1085 		btrfs_item_key_to_cpu(leaf, &found_key, slot);
1086 
1087 		if (found_key.type == BTRFS_ROOT_REF_KEY) {
1088 
1089 			/* Release locks on tree_root before we access quota_root */
1090 			btrfs_release_path(path);
1091 
1092 			ret = add_qgroup_item(trans, quota_root,
1093 					      found_key.offset);
1094 			if (ret) {
1095 				btrfs_abort_transaction(trans, ret);
1096 				goto out_free_path;
1097 			}
1098 
1099 			qgroup = add_qgroup_rb(fs_info, found_key.offset);
1100 			if (IS_ERR(qgroup)) {
1101 				ret = PTR_ERR(qgroup);
1102 				btrfs_abort_transaction(trans, ret);
1103 				goto out_free_path;
1104 			}
1105 			ret = btrfs_sysfs_add_one_qgroup(fs_info, qgroup);
1106 			if (ret < 0) {
1107 				btrfs_abort_transaction(trans, ret);
1108 				goto out_free_path;
1109 			}
1110 			ret = btrfs_search_slot_for_read(tree_root, &found_key,
1111 							 path, 1, 0);
1112 			if (ret < 0) {
1113 				btrfs_abort_transaction(trans, ret);
1114 				goto out_free_path;
1115 			}
1116 			if (ret > 0) {
1117 				/*
1118 				 * Shouldn't happen, but in case it does we
1119 				 * don't need to do the btrfs_next_item, just
1120 				 * continue.
1121 				 */
1122 				continue;
1123 			}
1124 		}
1125 		ret = btrfs_next_item(tree_root, path);
1126 		if (ret < 0) {
1127 			btrfs_abort_transaction(trans, ret);
1128 			goto out_free_path;
1129 		}
1130 		if (ret)
1131 			break;
1132 	}
1133 
1134 out_add_root:
1135 	btrfs_release_path(path);
1136 	ret = add_qgroup_item(trans, quota_root, BTRFS_FS_TREE_OBJECTID);
1137 	if (ret) {
1138 		btrfs_abort_transaction(trans, ret);
1139 		goto out_free_path;
1140 	}
1141 
1142 	qgroup = add_qgroup_rb(fs_info, BTRFS_FS_TREE_OBJECTID);
1143 	if (IS_ERR(qgroup)) {
1144 		ret = PTR_ERR(qgroup);
1145 		btrfs_abort_transaction(trans, ret);
1146 		goto out_free_path;
1147 	}
1148 	ret = btrfs_sysfs_add_one_qgroup(fs_info, qgroup);
1149 	if (ret < 0) {
1150 		btrfs_abort_transaction(trans, ret);
1151 		goto out_free_path;
1152 	}
1153 
1154 	mutex_unlock(&fs_info->qgroup_ioctl_lock);
1155 	/*
1156 	 * Commit the transaction while not holding qgroup_ioctl_lock, to avoid
1157 	 * a deadlock with tasks concurrently doing other qgroup operations, such
1158 	 * adding/removing qgroups or adding/deleting qgroup relations for example,
1159 	 * because all qgroup operations first start or join a transaction and then
1160 	 * lock the qgroup_ioctl_lock mutex.
1161 	 * We are safe from a concurrent task trying to enable quotas, by calling
1162 	 * this function, since we are serialized by fs_info->subvol_sem.
1163 	 */
1164 	ret = btrfs_commit_transaction(trans);
1165 	trans = NULL;
1166 	mutex_lock(&fs_info->qgroup_ioctl_lock);
1167 	if (ret)
1168 		goto out_free_path;
1169 
1170 	/*
1171 	 * Set quota enabled flag after committing the transaction, to avoid
1172 	 * deadlocks on fs_info->qgroup_ioctl_lock with concurrent snapshot
1173 	 * creation.
1174 	 */
1175 	spin_lock(&fs_info->qgroup_lock);
1176 	fs_info->quota_root = quota_root;
1177 	set_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags);
1178 	spin_unlock(&fs_info->qgroup_lock);
1179 
1180 	ret = qgroup_rescan_init(fs_info, 0, 1);
1181 	if (!ret) {
1182 	        qgroup_rescan_zero_tracking(fs_info);
1183 		fs_info->qgroup_rescan_running = true;
1184 	        btrfs_queue_work(fs_info->qgroup_rescan_workers,
1185 	                         &fs_info->qgroup_rescan_work);
1186 	} else {
1187 		/*
1188 		 * We have set both BTRFS_FS_QUOTA_ENABLED and
1189 		 * BTRFS_QGROUP_STATUS_FLAG_ON, so we can only fail with
1190 		 * -EINPROGRESS. That can happen because someone started the
1191 		 * rescan worker by calling quota rescan ioctl before we
1192 		 * attempted to initialize the rescan worker. Failure due to
1193 		 * quotas disabled in the meanwhile is not possible, because
1194 		 * we are holding a write lock on fs_info->subvol_sem, which
1195 		 * is also acquired when disabling quotas.
1196 		 * Ignore such error, and any other error would need to undo
1197 		 * everything we did in the transaction we just committed.
1198 		 */
1199 		ASSERT(ret == -EINPROGRESS);
1200 		ret = 0;
1201 	}
1202 
1203 out_free_path:
1204 	btrfs_free_path(path);
1205 out_free_root:
1206 	if (ret)
1207 		btrfs_put_root(quota_root);
1208 out:
1209 	if (ret) {
1210 		ulist_free(fs_info->qgroup_ulist);
1211 		fs_info->qgroup_ulist = NULL;
1212 		btrfs_sysfs_del_qgroups(fs_info);
1213 	}
1214 	mutex_unlock(&fs_info->qgroup_ioctl_lock);
1215 	if (ret && trans)
1216 		btrfs_end_transaction(trans);
1217 	else if (trans)
1218 		ret = btrfs_end_transaction(trans);
1219 	ulist_free(ulist);
1220 	return ret;
1221 }
1222 
1223 int btrfs_quota_disable(struct btrfs_fs_info *fs_info)
1224 {
1225 	struct btrfs_root *quota_root;
1226 	struct btrfs_trans_handle *trans = NULL;
1227 	int ret = 0;
1228 
1229 	/*
1230 	 * We need to have subvol_sem write locked, to prevent races between
1231 	 * concurrent tasks trying to disable quotas, because we will unlock
1232 	 * and relock qgroup_ioctl_lock across BTRFS_FS_QUOTA_ENABLED changes.
1233 	 */
1234 	lockdep_assert_held_write(&fs_info->subvol_sem);
1235 
1236 	mutex_lock(&fs_info->qgroup_ioctl_lock);
1237 	if (!fs_info->quota_root)
1238 		goto out;
1239 
1240 	/*
1241 	 * Unlock the qgroup_ioctl_lock mutex before waiting for the rescan worker to
1242 	 * complete. Otherwise we can deadlock because btrfs_remove_qgroup() needs
1243 	 * to lock that mutex while holding a transaction handle and the rescan
1244 	 * worker needs to commit a transaction.
1245 	 */
1246 	mutex_unlock(&fs_info->qgroup_ioctl_lock);
1247 
1248 	/*
1249 	 * Request qgroup rescan worker to complete and wait for it. This wait
1250 	 * must be done before transaction start for quota disable since it may
1251 	 * deadlock with transaction by the qgroup rescan worker.
1252 	 */
1253 	clear_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags);
1254 	btrfs_qgroup_wait_for_completion(fs_info, false);
1255 
1256 	/*
1257 	 * 1 For the root item
1258 	 *
1259 	 * We should also reserve enough items for the quota tree deletion in
1260 	 * btrfs_clean_quota_tree but this is not done.
1261 	 *
1262 	 * Also, we must always start a transaction without holding the mutex
1263 	 * qgroup_ioctl_lock, see btrfs_quota_enable().
1264 	 */
1265 	trans = btrfs_start_transaction(fs_info->tree_root, 1);
1266 
1267 	mutex_lock(&fs_info->qgroup_ioctl_lock);
1268 	if (IS_ERR(trans)) {
1269 		ret = PTR_ERR(trans);
1270 		trans = NULL;
1271 		set_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags);
1272 		goto out;
1273 	}
1274 
1275 	if (!fs_info->quota_root)
1276 		goto out;
1277 
1278 	spin_lock(&fs_info->qgroup_lock);
1279 	quota_root = fs_info->quota_root;
1280 	fs_info->quota_root = NULL;
1281 	fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_ON;
1282 	fs_info->qgroup_drop_subtree_thres = BTRFS_MAX_LEVEL;
1283 	spin_unlock(&fs_info->qgroup_lock);
1284 
1285 	btrfs_free_qgroup_config(fs_info);
1286 
1287 	ret = btrfs_clean_quota_tree(trans, quota_root);
1288 	if (ret) {
1289 		btrfs_abort_transaction(trans, ret);
1290 		goto out;
1291 	}
1292 
1293 	ret = btrfs_del_root(trans, &quota_root->root_key);
1294 	if (ret) {
1295 		btrfs_abort_transaction(trans, ret);
1296 		goto out;
1297 	}
1298 
1299 	list_del(&quota_root->dirty_list);
1300 
1301 	btrfs_tree_lock(quota_root->node);
1302 	btrfs_clean_tree_block(quota_root->node);
1303 	btrfs_tree_unlock(quota_root->node);
1304 	btrfs_free_tree_block(trans, btrfs_root_id(quota_root),
1305 			      quota_root->node, 0, 1);
1306 
1307 	btrfs_put_root(quota_root);
1308 
1309 out:
1310 	mutex_unlock(&fs_info->qgroup_ioctl_lock);
1311 	if (ret && trans)
1312 		btrfs_end_transaction(trans);
1313 	else if (trans)
1314 		ret = btrfs_end_transaction(trans);
1315 
1316 	return ret;
1317 }
1318 
1319 static void qgroup_dirty(struct btrfs_fs_info *fs_info,
1320 			 struct btrfs_qgroup *qgroup)
1321 {
1322 	if (list_empty(&qgroup->dirty))
1323 		list_add(&qgroup->dirty, &fs_info->dirty_qgroups);
1324 }
1325 
1326 /*
1327  * The easy accounting, we're updating qgroup relationship whose child qgroup
1328  * only has exclusive extents.
1329  *
1330  * In this case, all exclusive extents will also be exclusive for parent, so
1331  * excl/rfer just get added/removed.
1332  *
1333  * So is qgroup reservation space, which should also be added/removed to
1334  * parent.
1335  * Or when child tries to release reservation space, parent will underflow its
1336  * reservation (for relationship adding case).
1337  *
1338  * Caller should hold fs_info->qgroup_lock.
1339  */
1340 static int __qgroup_excl_accounting(struct btrfs_fs_info *fs_info,
1341 				    struct ulist *tmp, u64 ref_root,
1342 				    struct btrfs_qgroup *src, int sign)
1343 {
1344 	struct btrfs_qgroup *qgroup;
1345 	struct btrfs_qgroup_list *glist;
1346 	struct ulist_node *unode;
1347 	struct ulist_iterator uiter;
1348 	u64 num_bytes = src->excl;
1349 	int ret = 0;
1350 
1351 	qgroup = find_qgroup_rb(fs_info, ref_root);
1352 	if (!qgroup)
1353 		goto out;
1354 
1355 	qgroup->rfer += sign * num_bytes;
1356 	qgroup->rfer_cmpr += sign * num_bytes;
1357 
1358 	WARN_ON(sign < 0 && qgroup->excl < num_bytes);
1359 	qgroup->excl += sign * num_bytes;
1360 	qgroup->excl_cmpr += sign * num_bytes;
1361 
1362 	if (sign > 0)
1363 		qgroup_rsv_add_by_qgroup(fs_info, qgroup, src);
1364 	else
1365 		qgroup_rsv_release_by_qgroup(fs_info, qgroup, src);
1366 
1367 	qgroup_dirty(fs_info, qgroup);
1368 
1369 	/* Get all of the parent groups that contain this qgroup */
1370 	list_for_each_entry(glist, &qgroup->groups, next_group) {
1371 		ret = ulist_add(tmp, glist->group->qgroupid,
1372 				qgroup_to_aux(glist->group), GFP_ATOMIC);
1373 		if (ret < 0)
1374 			goto out;
1375 	}
1376 
1377 	/* Iterate all of the parents and adjust their reference counts */
1378 	ULIST_ITER_INIT(&uiter);
1379 	while ((unode = ulist_next(tmp, &uiter))) {
1380 		qgroup = unode_aux_to_qgroup(unode);
1381 		qgroup->rfer += sign * num_bytes;
1382 		qgroup->rfer_cmpr += sign * num_bytes;
1383 		WARN_ON(sign < 0 && qgroup->excl < num_bytes);
1384 		qgroup->excl += sign * num_bytes;
1385 		if (sign > 0)
1386 			qgroup_rsv_add_by_qgroup(fs_info, qgroup, src);
1387 		else
1388 			qgroup_rsv_release_by_qgroup(fs_info, qgroup, src);
1389 		qgroup->excl_cmpr += sign * num_bytes;
1390 		qgroup_dirty(fs_info, qgroup);
1391 
1392 		/* Add any parents of the parents */
1393 		list_for_each_entry(glist, &qgroup->groups, next_group) {
1394 			ret = ulist_add(tmp, glist->group->qgroupid,
1395 					qgroup_to_aux(glist->group), GFP_ATOMIC);
1396 			if (ret < 0)
1397 				goto out;
1398 		}
1399 	}
1400 	ret = 0;
1401 out:
1402 	return ret;
1403 }
1404 
1405 
1406 /*
1407  * Quick path for updating qgroup with only excl refs.
1408  *
1409  * In that case, just update all parent will be enough.
1410  * Or we needs to do a full rescan.
1411  * Caller should also hold fs_info->qgroup_lock.
1412  *
1413  * Return 0 for quick update, return >0 for need to full rescan
1414  * and mark INCONSISTENT flag.
1415  * Return < 0 for other error.
1416  */
1417 static int quick_update_accounting(struct btrfs_fs_info *fs_info,
1418 				   struct ulist *tmp, u64 src, u64 dst,
1419 				   int sign)
1420 {
1421 	struct btrfs_qgroup *qgroup;
1422 	int ret = 1;
1423 	int err = 0;
1424 
1425 	qgroup = find_qgroup_rb(fs_info, src);
1426 	if (!qgroup)
1427 		goto out;
1428 	if (qgroup->excl == qgroup->rfer) {
1429 		ret = 0;
1430 		err = __qgroup_excl_accounting(fs_info, tmp, dst,
1431 					       qgroup, sign);
1432 		if (err < 0) {
1433 			ret = err;
1434 			goto out;
1435 		}
1436 	}
1437 out:
1438 	if (ret)
1439 		fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
1440 	return ret;
1441 }
1442 
1443 int btrfs_add_qgroup_relation(struct btrfs_trans_handle *trans, u64 src,
1444 			      u64 dst)
1445 {
1446 	struct btrfs_fs_info *fs_info = trans->fs_info;
1447 	struct btrfs_qgroup *parent;
1448 	struct btrfs_qgroup *member;
1449 	struct btrfs_qgroup_list *list;
1450 	struct ulist *tmp;
1451 	unsigned int nofs_flag;
1452 	int ret = 0;
1453 
1454 	/* Check the level of src and dst first */
1455 	if (btrfs_qgroup_level(src) >= btrfs_qgroup_level(dst))
1456 		return -EINVAL;
1457 
1458 	/* We hold a transaction handle open, must do a NOFS allocation. */
1459 	nofs_flag = memalloc_nofs_save();
1460 	tmp = ulist_alloc(GFP_KERNEL);
1461 	memalloc_nofs_restore(nofs_flag);
1462 	if (!tmp)
1463 		return -ENOMEM;
1464 
1465 	mutex_lock(&fs_info->qgroup_ioctl_lock);
1466 	if (!fs_info->quota_root) {
1467 		ret = -ENOTCONN;
1468 		goto out;
1469 	}
1470 	member = find_qgroup_rb(fs_info, src);
1471 	parent = find_qgroup_rb(fs_info, dst);
1472 	if (!member || !parent) {
1473 		ret = -EINVAL;
1474 		goto out;
1475 	}
1476 
1477 	/* check if such qgroup relation exist firstly */
1478 	list_for_each_entry(list, &member->groups, next_group) {
1479 		if (list->group == parent) {
1480 			ret = -EEXIST;
1481 			goto out;
1482 		}
1483 	}
1484 
1485 	ret = add_qgroup_relation_item(trans, src, dst);
1486 	if (ret)
1487 		goto out;
1488 
1489 	ret = add_qgroup_relation_item(trans, dst, src);
1490 	if (ret) {
1491 		del_qgroup_relation_item(trans, src, dst);
1492 		goto out;
1493 	}
1494 
1495 	spin_lock(&fs_info->qgroup_lock);
1496 	ret = __add_relation_rb(member, parent);
1497 	if (ret < 0) {
1498 		spin_unlock(&fs_info->qgroup_lock);
1499 		goto out;
1500 	}
1501 	ret = quick_update_accounting(fs_info, tmp, src, dst, 1);
1502 	spin_unlock(&fs_info->qgroup_lock);
1503 out:
1504 	mutex_unlock(&fs_info->qgroup_ioctl_lock);
1505 	ulist_free(tmp);
1506 	return ret;
1507 }
1508 
1509 static int __del_qgroup_relation(struct btrfs_trans_handle *trans, u64 src,
1510 				 u64 dst)
1511 {
1512 	struct btrfs_fs_info *fs_info = trans->fs_info;
1513 	struct btrfs_qgroup *parent;
1514 	struct btrfs_qgroup *member;
1515 	struct btrfs_qgroup_list *list;
1516 	struct ulist *tmp;
1517 	bool found = false;
1518 	unsigned int nofs_flag;
1519 	int ret = 0;
1520 	int ret2;
1521 
1522 	/* We hold a transaction handle open, must do a NOFS allocation. */
1523 	nofs_flag = memalloc_nofs_save();
1524 	tmp = ulist_alloc(GFP_KERNEL);
1525 	memalloc_nofs_restore(nofs_flag);
1526 	if (!tmp)
1527 		return -ENOMEM;
1528 
1529 	if (!fs_info->quota_root) {
1530 		ret = -ENOTCONN;
1531 		goto out;
1532 	}
1533 
1534 	member = find_qgroup_rb(fs_info, src);
1535 	parent = find_qgroup_rb(fs_info, dst);
1536 	/*
1537 	 * The parent/member pair doesn't exist, then try to delete the dead
1538 	 * relation items only.
1539 	 */
1540 	if (!member || !parent)
1541 		goto delete_item;
1542 
1543 	/* check if such qgroup relation exist firstly */
1544 	list_for_each_entry(list, &member->groups, next_group) {
1545 		if (list->group == parent) {
1546 			found = true;
1547 			break;
1548 		}
1549 	}
1550 
1551 delete_item:
1552 	ret = del_qgroup_relation_item(trans, src, dst);
1553 	if (ret < 0 && ret != -ENOENT)
1554 		goto out;
1555 	ret2 = del_qgroup_relation_item(trans, dst, src);
1556 	if (ret2 < 0 && ret2 != -ENOENT)
1557 		goto out;
1558 
1559 	/* At least one deletion succeeded, return 0 */
1560 	if (!ret || !ret2)
1561 		ret = 0;
1562 
1563 	if (found) {
1564 		spin_lock(&fs_info->qgroup_lock);
1565 		del_relation_rb(fs_info, src, dst);
1566 		ret = quick_update_accounting(fs_info, tmp, src, dst, -1);
1567 		spin_unlock(&fs_info->qgroup_lock);
1568 	}
1569 out:
1570 	ulist_free(tmp);
1571 	return ret;
1572 }
1573 
1574 int btrfs_del_qgroup_relation(struct btrfs_trans_handle *trans, u64 src,
1575 			      u64 dst)
1576 {
1577 	struct btrfs_fs_info *fs_info = trans->fs_info;
1578 	int ret = 0;
1579 
1580 	mutex_lock(&fs_info->qgroup_ioctl_lock);
1581 	ret = __del_qgroup_relation(trans, src, dst);
1582 	mutex_unlock(&fs_info->qgroup_ioctl_lock);
1583 
1584 	return ret;
1585 }
1586 
1587 int btrfs_create_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid)
1588 {
1589 	struct btrfs_fs_info *fs_info = trans->fs_info;
1590 	struct btrfs_root *quota_root;
1591 	struct btrfs_qgroup *qgroup;
1592 	int ret = 0;
1593 
1594 	mutex_lock(&fs_info->qgroup_ioctl_lock);
1595 	if (!fs_info->quota_root) {
1596 		ret = -ENOTCONN;
1597 		goto out;
1598 	}
1599 	quota_root = fs_info->quota_root;
1600 	qgroup = find_qgroup_rb(fs_info, qgroupid);
1601 	if (qgroup) {
1602 		ret = -EEXIST;
1603 		goto out;
1604 	}
1605 
1606 	ret = add_qgroup_item(trans, quota_root, qgroupid);
1607 	if (ret)
1608 		goto out;
1609 
1610 	spin_lock(&fs_info->qgroup_lock);
1611 	qgroup = add_qgroup_rb(fs_info, qgroupid);
1612 	spin_unlock(&fs_info->qgroup_lock);
1613 
1614 	if (IS_ERR(qgroup)) {
1615 		ret = PTR_ERR(qgroup);
1616 		goto out;
1617 	}
1618 	ret = btrfs_sysfs_add_one_qgroup(fs_info, qgroup);
1619 out:
1620 	mutex_unlock(&fs_info->qgroup_ioctl_lock);
1621 	return ret;
1622 }
1623 
1624 int btrfs_remove_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid)
1625 {
1626 	struct btrfs_fs_info *fs_info = trans->fs_info;
1627 	struct btrfs_qgroup *qgroup;
1628 	struct btrfs_qgroup_list *list;
1629 	int ret = 0;
1630 
1631 	mutex_lock(&fs_info->qgroup_ioctl_lock);
1632 	if (!fs_info->quota_root) {
1633 		ret = -ENOTCONN;
1634 		goto out;
1635 	}
1636 
1637 	qgroup = find_qgroup_rb(fs_info, qgroupid);
1638 	if (!qgroup) {
1639 		ret = -ENOENT;
1640 		goto out;
1641 	}
1642 
1643 	/* Check if there are no children of this qgroup */
1644 	if (!list_empty(&qgroup->members)) {
1645 		ret = -EBUSY;
1646 		goto out;
1647 	}
1648 
1649 	ret = del_qgroup_item(trans, qgroupid);
1650 	if (ret && ret != -ENOENT)
1651 		goto out;
1652 
1653 	while (!list_empty(&qgroup->groups)) {
1654 		list = list_first_entry(&qgroup->groups,
1655 					struct btrfs_qgroup_list, next_group);
1656 		ret = __del_qgroup_relation(trans, qgroupid,
1657 					    list->group->qgroupid);
1658 		if (ret)
1659 			goto out;
1660 	}
1661 
1662 	spin_lock(&fs_info->qgroup_lock);
1663 	del_qgroup_rb(fs_info, qgroupid);
1664 	spin_unlock(&fs_info->qgroup_lock);
1665 
1666 	/*
1667 	 * Remove the qgroup from sysfs now without holding the qgroup_lock
1668 	 * spinlock, since the sysfs_remove_group() function needs to take
1669 	 * the mutex kernfs_mutex through kernfs_remove_by_name_ns().
1670 	 */
1671 	btrfs_sysfs_del_one_qgroup(fs_info, qgroup);
1672 	kfree(qgroup);
1673 out:
1674 	mutex_unlock(&fs_info->qgroup_ioctl_lock);
1675 	return ret;
1676 }
1677 
1678 int btrfs_limit_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid,
1679 		       struct btrfs_qgroup_limit *limit)
1680 {
1681 	struct btrfs_fs_info *fs_info = trans->fs_info;
1682 	struct btrfs_qgroup *qgroup;
1683 	int ret = 0;
1684 	/* Sometimes we would want to clear the limit on this qgroup.
1685 	 * To meet this requirement, we treat the -1 as a special value
1686 	 * which tell kernel to clear the limit on this qgroup.
1687 	 */
1688 	const u64 CLEAR_VALUE = -1;
1689 
1690 	mutex_lock(&fs_info->qgroup_ioctl_lock);
1691 	if (!fs_info->quota_root) {
1692 		ret = -ENOTCONN;
1693 		goto out;
1694 	}
1695 
1696 	qgroup = find_qgroup_rb(fs_info, qgroupid);
1697 	if (!qgroup) {
1698 		ret = -ENOENT;
1699 		goto out;
1700 	}
1701 
1702 	spin_lock(&fs_info->qgroup_lock);
1703 	if (limit->flags & BTRFS_QGROUP_LIMIT_MAX_RFER) {
1704 		if (limit->max_rfer == CLEAR_VALUE) {
1705 			qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_MAX_RFER;
1706 			limit->flags &= ~BTRFS_QGROUP_LIMIT_MAX_RFER;
1707 			qgroup->max_rfer = 0;
1708 		} else {
1709 			qgroup->max_rfer = limit->max_rfer;
1710 		}
1711 	}
1712 	if (limit->flags & BTRFS_QGROUP_LIMIT_MAX_EXCL) {
1713 		if (limit->max_excl == CLEAR_VALUE) {
1714 			qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_MAX_EXCL;
1715 			limit->flags &= ~BTRFS_QGROUP_LIMIT_MAX_EXCL;
1716 			qgroup->max_excl = 0;
1717 		} else {
1718 			qgroup->max_excl = limit->max_excl;
1719 		}
1720 	}
1721 	if (limit->flags & BTRFS_QGROUP_LIMIT_RSV_RFER) {
1722 		if (limit->rsv_rfer == CLEAR_VALUE) {
1723 			qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_RSV_RFER;
1724 			limit->flags &= ~BTRFS_QGROUP_LIMIT_RSV_RFER;
1725 			qgroup->rsv_rfer = 0;
1726 		} else {
1727 			qgroup->rsv_rfer = limit->rsv_rfer;
1728 		}
1729 	}
1730 	if (limit->flags & BTRFS_QGROUP_LIMIT_RSV_EXCL) {
1731 		if (limit->rsv_excl == CLEAR_VALUE) {
1732 			qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_RSV_EXCL;
1733 			limit->flags &= ~BTRFS_QGROUP_LIMIT_RSV_EXCL;
1734 			qgroup->rsv_excl = 0;
1735 		} else {
1736 			qgroup->rsv_excl = limit->rsv_excl;
1737 		}
1738 	}
1739 	qgroup->lim_flags |= limit->flags;
1740 
1741 	spin_unlock(&fs_info->qgroup_lock);
1742 
1743 	ret = update_qgroup_limit_item(trans, qgroup);
1744 	if (ret) {
1745 		qgroup_mark_inconsistent(fs_info);
1746 		btrfs_info(fs_info, "unable to update quota limit for %llu",
1747 		       qgroupid);
1748 	}
1749 
1750 out:
1751 	mutex_unlock(&fs_info->qgroup_ioctl_lock);
1752 	return ret;
1753 }
1754 
1755 int btrfs_qgroup_trace_extent_nolock(struct btrfs_fs_info *fs_info,
1756 				struct btrfs_delayed_ref_root *delayed_refs,
1757 				struct btrfs_qgroup_extent_record *record)
1758 {
1759 	struct rb_node **p = &delayed_refs->dirty_extent_root.rb_node;
1760 	struct rb_node *parent_node = NULL;
1761 	struct btrfs_qgroup_extent_record *entry;
1762 	u64 bytenr = record->bytenr;
1763 
1764 	lockdep_assert_held(&delayed_refs->lock);
1765 	trace_btrfs_qgroup_trace_extent(fs_info, record);
1766 
1767 	while (*p) {
1768 		parent_node = *p;
1769 		entry = rb_entry(parent_node, struct btrfs_qgroup_extent_record,
1770 				 node);
1771 		if (bytenr < entry->bytenr) {
1772 			p = &(*p)->rb_left;
1773 		} else if (bytenr > entry->bytenr) {
1774 			p = &(*p)->rb_right;
1775 		} else {
1776 			if (record->data_rsv && !entry->data_rsv) {
1777 				entry->data_rsv = record->data_rsv;
1778 				entry->data_rsv_refroot =
1779 					record->data_rsv_refroot;
1780 			}
1781 			return 1;
1782 		}
1783 	}
1784 
1785 	rb_link_node(&record->node, parent_node, p);
1786 	rb_insert_color(&record->node, &delayed_refs->dirty_extent_root);
1787 	return 0;
1788 }
1789 
1790 int btrfs_qgroup_trace_extent_post(struct btrfs_trans_handle *trans,
1791 				   struct btrfs_qgroup_extent_record *qrecord)
1792 {
1793 	struct ulist *old_root;
1794 	u64 bytenr = qrecord->bytenr;
1795 	int ret;
1796 
1797 	/*
1798 	 * We are always called in a context where we are already holding a
1799 	 * transaction handle. Often we are called when adding a data delayed
1800 	 * reference from btrfs_truncate_inode_items() (truncating or unlinking),
1801 	 * in which case we will be holding a write lock on extent buffer from a
1802 	 * subvolume tree. In this case we can't allow btrfs_find_all_roots() to
1803 	 * acquire fs_info->commit_root_sem, because that is a higher level lock
1804 	 * that must be acquired before locking any extent buffers.
1805 	 *
1806 	 * So we want btrfs_find_all_roots() to not acquire the commit_root_sem
1807 	 * but we can't pass it a non-NULL transaction handle, because otherwise
1808 	 * it would not use commit roots and would lock extent buffers, causing
1809 	 * a deadlock if it ends up trying to read lock the same extent buffer
1810 	 * that was previously write locked at btrfs_truncate_inode_items().
1811 	 *
1812 	 * So pass a NULL transaction handle to btrfs_find_all_roots() and
1813 	 * explicitly tell it to not acquire the commit_root_sem - if we are
1814 	 * holding a transaction handle we don't need its protection.
1815 	 */
1816 	ASSERT(trans != NULL);
1817 
1818 	if (trans->fs_info->qgroup_flags & BTRFS_QGROUP_RUNTIME_FLAG_NO_ACCOUNTING)
1819 		return 0;
1820 
1821 	ret = btrfs_find_all_roots(NULL, trans->fs_info, bytenr, 0, &old_root,
1822 				   true);
1823 	if (ret < 0) {
1824 		qgroup_mark_inconsistent(trans->fs_info);
1825 		btrfs_warn(trans->fs_info,
1826 "error accounting new delayed refs extent (err code: %d), quota inconsistent",
1827 			ret);
1828 		return 0;
1829 	}
1830 
1831 	/*
1832 	 * Here we don't need to get the lock of
1833 	 * trans->transaction->delayed_refs, since inserted qrecord won't
1834 	 * be deleted, only qrecord->node may be modified (new qrecord insert)
1835 	 *
1836 	 * So modifying qrecord->old_roots is safe here
1837 	 */
1838 	qrecord->old_roots = old_root;
1839 	return 0;
1840 }
1841 
1842 int btrfs_qgroup_trace_extent(struct btrfs_trans_handle *trans, u64 bytenr,
1843 			      u64 num_bytes, gfp_t gfp_flag)
1844 {
1845 	struct btrfs_fs_info *fs_info = trans->fs_info;
1846 	struct btrfs_qgroup_extent_record *record;
1847 	struct btrfs_delayed_ref_root *delayed_refs;
1848 	int ret;
1849 
1850 	if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)
1851 	    || bytenr == 0 || num_bytes == 0)
1852 		return 0;
1853 	record = kzalloc(sizeof(*record), gfp_flag);
1854 	if (!record)
1855 		return -ENOMEM;
1856 
1857 	delayed_refs = &trans->transaction->delayed_refs;
1858 	record->bytenr = bytenr;
1859 	record->num_bytes = num_bytes;
1860 	record->old_roots = NULL;
1861 
1862 	spin_lock(&delayed_refs->lock);
1863 	ret = btrfs_qgroup_trace_extent_nolock(fs_info, delayed_refs, record);
1864 	spin_unlock(&delayed_refs->lock);
1865 	if (ret > 0) {
1866 		kfree(record);
1867 		return 0;
1868 	}
1869 	return btrfs_qgroup_trace_extent_post(trans, record);
1870 }
1871 
1872 int btrfs_qgroup_trace_leaf_items(struct btrfs_trans_handle *trans,
1873 				  struct extent_buffer *eb)
1874 {
1875 	struct btrfs_fs_info *fs_info = trans->fs_info;
1876 	int nr = btrfs_header_nritems(eb);
1877 	int i, extent_type, ret;
1878 	struct btrfs_key key;
1879 	struct btrfs_file_extent_item *fi;
1880 	u64 bytenr, num_bytes;
1881 
1882 	/* We can be called directly from walk_up_proc() */
1883 	if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
1884 		return 0;
1885 
1886 	for (i = 0; i < nr; i++) {
1887 		btrfs_item_key_to_cpu(eb, &key, i);
1888 
1889 		if (key.type != BTRFS_EXTENT_DATA_KEY)
1890 			continue;
1891 
1892 		fi = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
1893 		/* filter out non qgroup-accountable extents  */
1894 		extent_type = btrfs_file_extent_type(eb, fi);
1895 
1896 		if (extent_type == BTRFS_FILE_EXTENT_INLINE)
1897 			continue;
1898 
1899 		bytenr = btrfs_file_extent_disk_bytenr(eb, fi);
1900 		if (!bytenr)
1901 			continue;
1902 
1903 		num_bytes = btrfs_file_extent_disk_num_bytes(eb, fi);
1904 
1905 		ret = btrfs_qgroup_trace_extent(trans, bytenr, num_bytes,
1906 						GFP_NOFS);
1907 		if (ret)
1908 			return ret;
1909 	}
1910 	cond_resched();
1911 	return 0;
1912 }
1913 
1914 /*
1915  * Walk up the tree from the bottom, freeing leaves and any interior
1916  * nodes which have had all slots visited. If a node (leaf or
1917  * interior) is freed, the node above it will have it's slot
1918  * incremented. The root node will never be freed.
1919  *
1920  * At the end of this function, we should have a path which has all
1921  * slots incremented to the next position for a search. If we need to
1922  * read a new node it will be NULL and the node above it will have the
1923  * correct slot selected for a later read.
1924  *
1925  * If we increment the root nodes slot counter past the number of
1926  * elements, 1 is returned to signal completion of the search.
1927  */
1928 static int adjust_slots_upwards(struct btrfs_path *path, int root_level)
1929 {
1930 	int level = 0;
1931 	int nr, slot;
1932 	struct extent_buffer *eb;
1933 
1934 	if (root_level == 0)
1935 		return 1;
1936 
1937 	while (level <= root_level) {
1938 		eb = path->nodes[level];
1939 		nr = btrfs_header_nritems(eb);
1940 		path->slots[level]++;
1941 		slot = path->slots[level];
1942 		if (slot >= nr || level == 0) {
1943 			/*
1944 			 * Don't free the root -  we will detect this
1945 			 * condition after our loop and return a
1946 			 * positive value for caller to stop walking the tree.
1947 			 */
1948 			if (level != root_level) {
1949 				btrfs_tree_unlock_rw(eb, path->locks[level]);
1950 				path->locks[level] = 0;
1951 
1952 				free_extent_buffer(eb);
1953 				path->nodes[level] = NULL;
1954 				path->slots[level] = 0;
1955 			}
1956 		} else {
1957 			/*
1958 			 * We have a valid slot to walk back down
1959 			 * from. Stop here so caller can process these
1960 			 * new nodes.
1961 			 */
1962 			break;
1963 		}
1964 
1965 		level++;
1966 	}
1967 
1968 	eb = path->nodes[root_level];
1969 	if (path->slots[root_level] >= btrfs_header_nritems(eb))
1970 		return 1;
1971 
1972 	return 0;
1973 }
1974 
1975 /*
1976  * Helper function to trace a subtree tree block swap.
1977  *
1978  * The swap will happen in highest tree block, but there may be a lot of
1979  * tree blocks involved.
1980  *
1981  * For example:
1982  *  OO = Old tree blocks
1983  *  NN = New tree blocks allocated during balance
1984  *
1985  *           File tree (257)                  Reloc tree for 257
1986  * L2              OO                                NN
1987  *               /    \                            /    \
1988  * L1          OO      OO (a)                    OO      NN (a)
1989  *            / \     / \                       / \     / \
1990  * L0       OO   OO OO   OO                   OO   OO NN   NN
1991  *                  (b)  (c)                          (b)  (c)
1992  *
1993  * When calling qgroup_trace_extent_swap(), we will pass:
1994  * @src_eb = OO(a)
1995  * @dst_path = [ nodes[1] = NN(a), nodes[0] = NN(c) ]
1996  * @dst_level = 0
1997  * @root_level = 1
1998  *
1999  * In that case, qgroup_trace_extent_swap() will search from OO(a) to
2000  * reach OO(c), then mark both OO(c) and NN(c) as qgroup dirty.
2001  *
2002  * The main work of qgroup_trace_extent_swap() can be split into 3 parts:
2003  *
2004  * 1) Tree search from @src_eb
2005  *    It should acts as a simplified btrfs_search_slot().
2006  *    The key for search can be extracted from @dst_path->nodes[dst_level]
2007  *    (first key).
2008  *
2009  * 2) Mark the final tree blocks in @src_path and @dst_path qgroup dirty
2010  *    NOTE: In above case, OO(a) and NN(a) won't be marked qgroup dirty.
2011  *    They should be marked during previous (@dst_level = 1) iteration.
2012  *
2013  * 3) Mark file extents in leaves dirty
2014  *    We don't have good way to pick out new file extents only.
2015  *    So we still follow the old method by scanning all file extents in
2016  *    the leave.
2017  *
2018  * This function can free us from keeping two paths, thus later we only need
2019  * to care about how to iterate all new tree blocks in reloc tree.
2020  */
2021 static int qgroup_trace_extent_swap(struct btrfs_trans_handle* trans,
2022 				    struct extent_buffer *src_eb,
2023 				    struct btrfs_path *dst_path,
2024 				    int dst_level, int root_level,
2025 				    bool trace_leaf)
2026 {
2027 	struct btrfs_key key;
2028 	struct btrfs_path *src_path;
2029 	struct btrfs_fs_info *fs_info = trans->fs_info;
2030 	u32 nodesize = fs_info->nodesize;
2031 	int cur_level = root_level;
2032 	int ret;
2033 
2034 	BUG_ON(dst_level > root_level);
2035 	/* Level mismatch */
2036 	if (btrfs_header_level(src_eb) != root_level)
2037 		return -EINVAL;
2038 
2039 	src_path = btrfs_alloc_path();
2040 	if (!src_path) {
2041 		ret = -ENOMEM;
2042 		goto out;
2043 	}
2044 
2045 	if (dst_level)
2046 		btrfs_node_key_to_cpu(dst_path->nodes[dst_level], &key, 0);
2047 	else
2048 		btrfs_item_key_to_cpu(dst_path->nodes[dst_level], &key, 0);
2049 
2050 	/* For src_path */
2051 	atomic_inc(&src_eb->refs);
2052 	src_path->nodes[root_level] = src_eb;
2053 	src_path->slots[root_level] = dst_path->slots[root_level];
2054 	src_path->locks[root_level] = 0;
2055 
2056 	/* A simplified version of btrfs_search_slot() */
2057 	while (cur_level >= dst_level) {
2058 		struct btrfs_key src_key;
2059 		struct btrfs_key dst_key;
2060 
2061 		if (src_path->nodes[cur_level] == NULL) {
2062 			struct extent_buffer *eb;
2063 			int parent_slot;
2064 
2065 			eb = src_path->nodes[cur_level + 1];
2066 			parent_slot = src_path->slots[cur_level + 1];
2067 
2068 			eb = btrfs_read_node_slot(eb, parent_slot);
2069 			if (IS_ERR(eb)) {
2070 				ret = PTR_ERR(eb);
2071 				goto out;
2072 			}
2073 
2074 			src_path->nodes[cur_level] = eb;
2075 
2076 			btrfs_tree_read_lock(eb);
2077 			src_path->locks[cur_level] = BTRFS_READ_LOCK;
2078 		}
2079 
2080 		src_path->slots[cur_level] = dst_path->slots[cur_level];
2081 		if (cur_level) {
2082 			btrfs_node_key_to_cpu(dst_path->nodes[cur_level],
2083 					&dst_key, dst_path->slots[cur_level]);
2084 			btrfs_node_key_to_cpu(src_path->nodes[cur_level],
2085 					&src_key, src_path->slots[cur_level]);
2086 		} else {
2087 			btrfs_item_key_to_cpu(dst_path->nodes[cur_level],
2088 					&dst_key, dst_path->slots[cur_level]);
2089 			btrfs_item_key_to_cpu(src_path->nodes[cur_level],
2090 					&src_key, src_path->slots[cur_level]);
2091 		}
2092 		/* Content mismatch, something went wrong */
2093 		if (btrfs_comp_cpu_keys(&dst_key, &src_key)) {
2094 			ret = -ENOENT;
2095 			goto out;
2096 		}
2097 		cur_level--;
2098 	}
2099 
2100 	/*
2101 	 * Now both @dst_path and @src_path have been populated, record the tree
2102 	 * blocks for qgroup accounting.
2103 	 */
2104 	ret = btrfs_qgroup_trace_extent(trans, src_path->nodes[dst_level]->start,
2105 			nodesize, GFP_NOFS);
2106 	if (ret < 0)
2107 		goto out;
2108 	ret = btrfs_qgroup_trace_extent(trans,
2109 			dst_path->nodes[dst_level]->start,
2110 			nodesize, GFP_NOFS);
2111 	if (ret < 0)
2112 		goto out;
2113 
2114 	/* Record leaf file extents */
2115 	if (dst_level == 0 && trace_leaf) {
2116 		ret = btrfs_qgroup_trace_leaf_items(trans, src_path->nodes[0]);
2117 		if (ret < 0)
2118 			goto out;
2119 		ret = btrfs_qgroup_trace_leaf_items(trans, dst_path->nodes[0]);
2120 	}
2121 out:
2122 	btrfs_free_path(src_path);
2123 	return ret;
2124 }
2125 
2126 /*
2127  * Helper function to do recursive generation-aware depth-first search, to
2128  * locate all new tree blocks in a subtree of reloc tree.
2129  *
2130  * E.g. (OO = Old tree blocks, NN = New tree blocks, whose gen == last_snapshot)
2131  *         reloc tree
2132  * L2         NN (a)
2133  *          /    \
2134  * L1    OO        NN (b)
2135  *      /  \      /  \
2136  * L0  OO  OO    OO  NN
2137  *               (c) (d)
2138  * If we pass:
2139  * @dst_path = [ nodes[1] = NN(b), nodes[0] = NULL ],
2140  * @cur_level = 1
2141  * @root_level = 1
2142  *
2143  * We will iterate through tree blocks NN(b), NN(d) and info qgroup to trace
2144  * above tree blocks along with their counter parts in file tree.
2145  * While during search, old tree blocks OO(c) will be skipped as tree block swap
2146  * won't affect OO(c).
2147  */
2148 static int qgroup_trace_new_subtree_blocks(struct btrfs_trans_handle* trans,
2149 					   struct extent_buffer *src_eb,
2150 					   struct btrfs_path *dst_path,
2151 					   int cur_level, int root_level,
2152 					   u64 last_snapshot, bool trace_leaf)
2153 {
2154 	struct btrfs_fs_info *fs_info = trans->fs_info;
2155 	struct extent_buffer *eb;
2156 	bool need_cleanup = false;
2157 	int ret = 0;
2158 	int i;
2159 
2160 	/* Level sanity check */
2161 	if (cur_level < 0 || cur_level >= BTRFS_MAX_LEVEL - 1 ||
2162 	    root_level < 0 || root_level >= BTRFS_MAX_LEVEL - 1 ||
2163 	    root_level < cur_level) {
2164 		btrfs_err_rl(fs_info,
2165 			"%s: bad levels, cur_level=%d root_level=%d",
2166 			__func__, cur_level, root_level);
2167 		return -EUCLEAN;
2168 	}
2169 
2170 	/* Read the tree block if needed */
2171 	if (dst_path->nodes[cur_level] == NULL) {
2172 		int parent_slot;
2173 		u64 child_gen;
2174 
2175 		/*
2176 		 * dst_path->nodes[root_level] must be initialized before
2177 		 * calling this function.
2178 		 */
2179 		if (cur_level == root_level) {
2180 			btrfs_err_rl(fs_info,
2181 	"%s: dst_path->nodes[%d] not initialized, root_level=%d cur_level=%d",
2182 				__func__, root_level, root_level, cur_level);
2183 			return -EUCLEAN;
2184 		}
2185 
2186 		/*
2187 		 * We need to get child blockptr/gen from parent before we can
2188 		 * read it.
2189 		  */
2190 		eb = dst_path->nodes[cur_level + 1];
2191 		parent_slot = dst_path->slots[cur_level + 1];
2192 		child_gen = btrfs_node_ptr_generation(eb, parent_slot);
2193 
2194 		/* This node is old, no need to trace */
2195 		if (child_gen < last_snapshot)
2196 			goto out;
2197 
2198 		eb = btrfs_read_node_slot(eb, parent_slot);
2199 		if (IS_ERR(eb)) {
2200 			ret = PTR_ERR(eb);
2201 			goto out;
2202 		}
2203 
2204 		dst_path->nodes[cur_level] = eb;
2205 		dst_path->slots[cur_level] = 0;
2206 
2207 		btrfs_tree_read_lock(eb);
2208 		dst_path->locks[cur_level] = BTRFS_READ_LOCK;
2209 		need_cleanup = true;
2210 	}
2211 
2212 	/* Now record this tree block and its counter part for qgroups */
2213 	ret = qgroup_trace_extent_swap(trans, src_eb, dst_path, cur_level,
2214 				       root_level, trace_leaf);
2215 	if (ret < 0)
2216 		goto cleanup;
2217 
2218 	eb = dst_path->nodes[cur_level];
2219 
2220 	if (cur_level > 0) {
2221 		/* Iterate all child tree blocks */
2222 		for (i = 0; i < btrfs_header_nritems(eb); i++) {
2223 			/* Skip old tree blocks as they won't be swapped */
2224 			if (btrfs_node_ptr_generation(eb, i) < last_snapshot)
2225 				continue;
2226 			dst_path->slots[cur_level] = i;
2227 
2228 			/* Recursive call (at most 7 times) */
2229 			ret = qgroup_trace_new_subtree_blocks(trans, src_eb,
2230 					dst_path, cur_level - 1, root_level,
2231 					last_snapshot, trace_leaf);
2232 			if (ret < 0)
2233 				goto cleanup;
2234 		}
2235 	}
2236 
2237 cleanup:
2238 	if (need_cleanup) {
2239 		/* Clean up */
2240 		btrfs_tree_unlock_rw(dst_path->nodes[cur_level],
2241 				     dst_path->locks[cur_level]);
2242 		free_extent_buffer(dst_path->nodes[cur_level]);
2243 		dst_path->nodes[cur_level] = NULL;
2244 		dst_path->slots[cur_level] = 0;
2245 		dst_path->locks[cur_level] = 0;
2246 	}
2247 out:
2248 	return ret;
2249 }
2250 
2251 static int qgroup_trace_subtree_swap(struct btrfs_trans_handle *trans,
2252 				struct extent_buffer *src_eb,
2253 				struct extent_buffer *dst_eb,
2254 				u64 last_snapshot, bool trace_leaf)
2255 {
2256 	struct btrfs_fs_info *fs_info = trans->fs_info;
2257 	struct btrfs_path *dst_path = NULL;
2258 	int level;
2259 	int ret;
2260 
2261 	if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
2262 		return 0;
2263 
2264 	/* Wrong parameter order */
2265 	if (btrfs_header_generation(src_eb) > btrfs_header_generation(dst_eb)) {
2266 		btrfs_err_rl(fs_info,
2267 		"%s: bad parameter order, src_gen=%llu dst_gen=%llu", __func__,
2268 			     btrfs_header_generation(src_eb),
2269 			     btrfs_header_generation(dst_eb));
2270 		return -EUCLEAN;
2271 	}
2272 
2273 	if (!extent_buffer_uptodate(src_eb) || !extent_buffer_uptodate(dst_eb)) {
2274 		ret = -EIO;
2275 		goto out;
2276 	}
2277 
2278 	level = btrfs_header_level(dst_eb);
2279 	dst_path = btrfs_alloc_path();
2280 	if (!dst_path) {
2281 		ret = -ENOMEM;
2282 		goto out;
2283 	}
2284 	/* For dst_path */
2285 	atomic_inc(&dst_eb->refs);
2286 	dst_path->nodes[level] = dst_eb;
2287 	dst_path->slots[level] = 0;
2288 	dst_path->locks[level] = 0;
2289 
2290 	/* Do the generation aware breadth-first search */
2291 	ret = qgroup_trace_new_subtree_blocks(trans, src_eb, dst_path, level,
2292 					      level, last_snapshot, trace_leaf);
2293 	if (ret < 0)
2294 		goto out;
2295 	ret = 0;
2296 
2297 out:
2298 	btrfs_free_path(dst_path);
2299 	if (ret < 0)
2300 		qgroup_mark_inconsistent(fs_info);
2301 	return ret;
2302 }
2303 
2304 int btrfs_qgroup_trace_subtree(struct btrfs_trans_handle *trans,
2305 			       struct extent_buffer *root_eb,
2306 			       u64 root_gen, int root_level)
2307 {
2308 	struct btrfs_fs_info *fs_info = trans->fs_info;
2309 	int ret = 0;
2310 	int level;
2311 	u8 drop_subptree_thres;
2312 	struct extent_buffer *eb = root_eb;
2313 	struct btrfs_path *path = NULL;
2314 
2315 	BUG_ON(root_level < 0 || root_level >= BTRFS_MAX_LEVEL);
2316 	BUG_ON(root_eb == NULL);
2317 
2318 	if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
2319 		return 0;
2320 
2321 	spin_lock(&fs_info->qgroup_lock);
2322 	drop_subptree_thres = fs_info->qgroup_drop_subtree_thres;
2323 	spin_unlock(&fs_info->qgroup_lock);
2324 
2325 	/*
2326 	 * This function only gets called for snapshot drop, if we hit a high
2327 	 * node here, it means we are going to change ownership for quite a lot
2328 	 * of extents, which will greatly slow down btrfs_commit_transaction().
2329 	 *
2330 	 * So here if we find a high tree here, we just skip the accounting and
2331 	 * mark qgroup inconsistent.
2332 	 */
2333 	if (root_level >= drop_subptree_thres) {
2334 		qgroup_mark_inconsistent(fs_info);
2335 		return 0;
2336 	}
2337 
2338 	if (!extent_buffer_uptodate(root_eb)) {
2339 		ret = btrfs_read_extent_buffer(root_eb, root_gen, root_level, NULL);
2340 		if (ret)
2341 			goto out;
2342 	}
2343 
2344 	if (root_level == 0) {
2345 		ret = btrfs_qgroup_trace_leaf_items(trans, root_eb);
2346 		goto out;
2347 	}
2348 
2349 	path = btrfs_alloc_path();
2350 	if (!path)
2351 		return -ENOMEM;
2352 
2353 	/*
2354 	 * Walk down the tree.  Missing extent blocks are filled in as
2355 	 * we go. Metadata is accounted every time we read a new
2356 	 * extent block.
2357 	 *
2358 	 * When we reach a leaf, we account for file extent items in it,
2359 	 * walk back up the tree (adjusting slot pointers as we go)
2360 	 * and restart the search process.
2361 	 */
2362 	atomic_inc(&root_eb->refs);	/* For path */
2363 	path->nodes[root_level] = root_eb;
2364 	path->slots[root_level] = 0;
2365 	path->locks[root_level] = 0; /* so release_path doesn't try to unlock */
2366 walk_down:
2367 	level = root_level;
2368 	while (level >= 0) {
2369 		if (path->nodes[level] == NULL) {
2370 			int parent_slot;
2371 			u64 child_bytenr;
2372 
2373 			/*
2374 			 * We need to get child blockptr from parent before we
2375 			 * can read it.
2376 			  */
2377 			eb = path->nodes[level + 1];
2378 			parent_slot = path->slots[level + 1];
2379 			child_bytenr = btrfs_node_blockptr(eb, parent_slot);
2380 
2381 			eb = btrfs_read_node_slot(eb, parent_slot);
2382 			if (IS_ERR(eb)) {
2383 				ret = PTR_ERR(eb);
2384 				goto out;
2385 			}
2386 
2387 			path->nodes[level] = eb;
2388 			path->slots[level] = 0;
2389 
2390 			btrfs_tree_read_lock(eb);
2391 			path->locks[level] = BTRFS_READ_LOCK;
2392 
2393 			ret = btrfs_qgroup_trace_extent(trans, child_bytenr,
2394 							fs_info->nodesize,
2395 							GFP_NOFS);
2396 			if (ret)
2397 				goto out;
2398 		}
2399 
2400 		if (level == 0) {
2401 			ret = btrfs_qgroup_trace_leaf_items(trans,
2402 							    path->nodes[level]);
2403 			if (ret)
2404 				goto out;
2405 
2406 			/* Nonzero return here means we completed our search */
2407 			ret = adjust_slots_upwards(path, root_level);
2408 			if (ret)
2409 				break;
2410 
2411 			/* Restart search with new slots */
2412 			goto walk_down;
2413 		}
2414 
2415 		level--;
2416 	}
2417 
2418 	ret = 0;
2419 out:
2420 	btrfs_free_path(path);
2421 
2422 	return ret;
2423 }
2424 
2425 #define UPDATE_NEW	0
2426 #define UPDATE_OLD	1
2427 /*
2428  * Walk all of the roots that points to the bytenr and adjust their refcnts.
2429  */
2430 static int qgroup_update_refcnt(struct btrfs_fs_info *fs_info,
2431 				struct ulist *roots, struct ulist *tmp,
2432 				struct ulist *qgroups, u64 seq, int update_old)
2433 {
2434 	struct ulist_node *unode;
2435 	struct ulist_iterator uiter;
2436 	struct ulist_node *tmp_unode;
2437 	struct ulist_iterator tmp_uiter;
2438 	struct btrfs_qgroup *qg;
2439 	int ret = 0;
2440 
2441 	if (!roots)
2442 		return 0;
2443 	ULIST_ITER_INIT(&uiter);
2444 	while ((unode = ulist_next(roots, &uiter))) {
2445 		qg = find_qgroup_rb(fs_info, unode->val);
2446 		if (!qg)
2447 			continue;
2448 
2449 		ulist_reinit(tmp);
2450 		ret = ulist_add(qgroups, qg->qgroupid, qgroup_to_aux(qg),
2451 				GFP_ATOMIC);
2452 		if (ret < 0)
2453 			return ret;
2454 		ret = ulist_add(tmp, qg->qgroupid, qgroup_to_aux(qg), GFP_ATOMIC);
2455 		if (ret < 0)
2456 			return ret;
2457 		ULIST_ITER_INIT(&tmp_uiter);
2458 		while ((tmp_unode = ulist_next(tmp, &tmp_uiter))) {
2459 			struct btrfs_qgroup_list *glist;
2460 
2461 			qg = unode_aux_to_qgroup(tmp_unode);
2462 			if (update_old)
2463 				btrfs_qgroup_update_old_refcnt(qg, seq, 1);
2464 			else
2465 				btrfs_qgroup_update_new_refcnt(qg, seq, 1);
2466 			list_for_each_entry(glist, &qg->groups, next_group) {
2467 				ret = ulist_add(qgroups, glist->group->qgroupid,
2468 						qgroup_to_aux(glist->group),
2469 						GFP_ATOMIC);
2470 				if (ret < 0)
2471 					return ret;
2472 				ret = ulist_add(tmp, glist->group->qgroupid,
2473 						qgroup_to_aux(glist->group),
2474 						GFP_ATOMIC);
2475 				if (ret < 0)
2476 					return ret;
2477 			}
2478 		}
2479 	}
2480 	return 0;
2481 }
2482 
2483 /*
2484  * Update qgroup rfer/excl counters.
2485  * Rfer update is easy, codes can explain themselves.
2486  *
2487  * Excl update is tricky, the update is split into 2 parts.
2488  * Part 1: Possible exclusive <-> sharing detect:
2489  *	|	A	|	!A	|
2490  *  -------------------------------------
2491  *  B	|	*	|	-	|
2492  *  -------------------------------------
2493  *  !B	|	+	|	**	|
2494  *  -------------------------------------
2495  *
2496  * Conditions:
2497  * A:	cur_old_roots < nr_old_roots	(not exclusive before)
2498  * !A:	cur_old_roots == nr_old_roots	(possible exclusive before)
2499  * B:	cur_new_roots < nr_new_roots	(not exclusive now)
2500  * !B:	cur_new_roots == nr_new_roots	(possible exclusive now)
2501  *
2502  * Results:
2503  * +: Possible sharing -> exclusive	-: Possible exclusive -> sharing
2504  * *: Definitely not changed.		**: Possible unchanged.
2505  *
2506  * For !A and !B condition, the exception is cur_old/new_roots == 0 case.
2507  *
2508  * To make the logic clear, we first use condition A and B to split
2509  * combination into 4 results.
2510  *
2511  * Then, for result "+" and "-", check old/new_roots == 0 case, as in them
2512  * only on variant maybe 0.
2513  *
2514  * Lastly, check result **, since there are 2 variants maybe 0, split them
2515  * again(2x2).
2516  * But this time we don't need to consider other things, the codes and logic
2517  * is easy to understand now.
2518  */
2519 static int qgroup_update_counters(struct btrfs_fs_info *fs_info,
2520 				  struct ulist *qgroups,
2521 				  u64 nr_old_roots,
2522 				  u64 nr_new_roots,
2523 				  u64 num_bytes, u64 seq)
2524 {
2525 	struct ulist_node *unode;
2526 	struct ulist_iterator uiter;
2527 	struct btrfs_qgroup *qg;
2528 	u64 cur_new_count, cur_old_count;
2529 
2530 	ULIST_ITER_INIT(&uiter);
2531 	while ((unode = ulist_next(qgroups, &uiter))) {
2532 		bool dirty = false;
2533 
2534 		qg = unode_aux_to_qgroup(unode);
2535 		cur_old_count = btrfs_qgroup_get_old_refcnt(qg, seq);
2536 		cur_new_count = btrfs_qgroup_get_new_refcnt(qg, seq);
2537 
2538 		trace_qgroup_update_counters(fs_info, qg, cur_old_count,
2539 					     cur_new_count);
2540 
2541 		/* Rfer update part */
2542 		if (cur_old_count == 0 && cur_new_count > 0) {
2543 			qg->rfer += num_bytes;
2544 			qg->rfer_cmpr += num_bytes;
2545 			dirty = true;
2546 		}
2547 		if (cur_old_count > 0 && cur_new_count == 0) {
2548 			qg->rfer -= num_bytes;
2549 			qg->rfer_cmpr -= num_bytes;
2550 			dirty = true;
2551 		}
2552 
2553 		/* Excl update part */
2554 		/* Exclusive/none -> shared case */
2555 		if (cur_old_count == nr_old_roots &&
2556 		    cur_new_count < nr_new_roots) {
2557 			/* Exclusive -> shared */
2558 			if (cur_old_count != 0) {
2559 				qg->excl -= num_bytes;
2560 				qg->excl_cmpr -= num_bytes;
2561 				dirty = true;
2562 			}
2563 		}
2564 
2565 		/* Shared -> exclusive/none case */
2566 		if (cur_old_count < nr_old_roots &&
2567 		    cur_new_count == nr_new_roots) {
2568 			/* Shared->exclusive */
2569 			if (cur_new_count != 0) {
2570 				qg->excl += num_bytes;
2571 				qg->excl_cmpr += num_bytes;
2572 				dirty = true;
2573 			}
2574 		}
2575 
2576 		/* Exclusive/none -> exclusive/none case */
2577 		if (cur_old_count == nr_old_roots &&
2578 		    cur_new_count == nr_new_roots) {
2579 			if (cur_old_count == 0) {
2580 				/* None -> exclusive/none */
2581 
2582 				if (cur_new_count != 0) {
2583 					/* None -> exclusive */
2584 					qg->excl += num_bytes;
2585 					qg->excl_cmpr += num_bytes;
2586 					dirty = true;
2587 				}
2588 				/* None -> none, nothing changed */
2589 			} else {
2590 				/* Exclusive -> exclusive/none */
2591 
2592 				if (cur_new_count == 0) {
2593 					/* Exclusive -> none */
2594 					qg->excl -= num_bytes;
2595 					qg->excl_cmpr -= num_bytes;
2596 					dirty = true;
2597 				}
2598 				/* Exclusive -> exclusive, nothing changed */
2599 			}
2600 		}
2601 
2602 		if (dirty)
2603 			qgroup_dirty(fs_info, qg);
2604 	}
2605 	return 0;
2606 }
2607 
2608 /*
2609  * Check if the @roots potentially is a list of fs tree roots
2610  *
2611  * Return 0 for definitely not a fs/subvol tree roots ulist
2612  * Return 1 for possible fs/subvol tree roots in the list (considering an empty
2613  *          one as well)
2614  */
2615 static int maybe_fs_roots(struct ulist *roots)
2616 {
2617 	struct ulist_node *unode;
2618 	struct ulist_iterator uiter;
2619 
2620 	/* Empty one, still possible for fs roots */
2621 	if (!roots || roots->nnodes == 0)
2622 		return 1;
2623 
2624 	ULIST_ITER_INIT(&uiter);
2625 	unode = ulist_next(roots, &uiter);
2626 	if (!unode)
2627 		return 1;
2628 
2629 	/*
2630 	 * If it contains fs tree roots, then it must belong to fs/subvol
2631 	 * trees.
2632 	 * If it contains a non-fs tree, it won't be shared with fs/subvol trees.
2633 	 */
2634 	return is_fstree(unode->val);
2635 }
2636 
2637 int btrfs_qgroup_account_extent(struct btrfs_trans_handle *trans, u64 bytenr,
2638 				u64 num_bytes, struct ulist *old_roots,
2639 				struct ulist *new_roots)
2640 {
2641 	struct btrfs_fs_info *fs_info = trans->fs_info;
2642 	struct ulist *qgroups = NULL;
2643 	struct ulist *tmp = NULL;
2644 	u64 seq;
2645 	u64 nr_new_roots = 0;
2646 	u64 nr_old_roots = 0;
2647 	int ret = 0;
2648 
2649 	/*
2650 	 * If quotas get disabled meanwhile, the resources need to be freed and
2651 	 * we can't just exit here.
2652 	 */
2653 	if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) ||
2654 	    fs_info->qgroup_flags & BTRFS_QGROUP_RUNTIME_FLAG_NO_ACCOUNTING)
2655 		goto out_free;
2656 
2657 	if (new_roots) {
2658 		if (!maybe_fs_roots(new_roots))
2659 			goto out_free;
2660 		nr_new_roots = new_roots->nnodes;
2661 	}
2662 	if (old_roots) {
2663 		if (!maybe_fs_roots(old_roots))
2664 			goto out_free;
2665 		nr_old_roots = old_roots->nnodes;
2666 	}
2667 
2668 	/* Quick exit, either not fs tree roots, or won't affect any qgroup */
2669 	if (nr_old_roots == 0 && nr_new_roots == 0)
2670 		goto out_free;
2671 
2672 	BUG_ON(!fs_info->quota_root);
2673 
2674 	trace_btrfs_qgroup_account_extent(fs_info, trans->transid, bytenr,
2675 					num_bytes, nr_old_roots, nr_new_roots);
2676 
2677 	qgroups = ulist_alloc(GFP_NOFS);
2678 	if (!qgroups) {
2679 		ret = -ENOMEM;
2680 		goto out_free;
2681 	}
2682 	tmp = ulist_alloc(GFP_NOFS);
2683 	if (!tmp) {
2684 		ret = -ENOMEM;
2685 		goto out_free;
2686 	}
2687 
2688 	mutex_lock(&fs_info->qgroup_rescan_lock);
2689 	if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN) {
2690 		if (fs_info->qgroup_rescan_progress.objectid <= bytenr) {
2691 			mutex_unlock(&fs_info->qgroup_rescan_lock);
2692 			ret = 0;
2693 			goto out_free;
2694 		}
2695 	}
2696 	mutex_unlock(&fs_info->qgroup_rescan_lock);
2697 
2698 	spin_lock(&fs_info->qgroup_lock);
2699 	seq = fs_info->qgroup_seq;
2700 
2701 	/* Update old refcnts using old_roots */
2702 	ret = qgroup_update_refcnt(fs_info, old_roots, tmp, qgroups, seq,
2703 				   UPDATE_OLD);
2704 	if (ret < 0)
2705 		goto out;
2706 
2707 	/* Update new refcnts using new_roots */
2708 	ret = qgroup_update_refcnt(fs_info, new_roots, tmp, qgroups, seq,
2709 				   UPDATE_NEW);
2710 	if (ret < 0)
2711 		goto out;
2712 
2713 	qgroup_update_counters(fs_info, qgroups, nr_old_roots, nr_new_roots,
2714 			       num_bytes, seq);
2715 
2716 	/*
2717 	 * Bump qgroup_seq to avoid seq overlap
2718 	 */
2719 	fs_info->qgroup_seq += max(nr_old_roots, nr_new_roots) + 1;
2720 out:
2721 	spin_unlock(&fs_info->qgroup_lock);
2722 out_free:
2723 	ulist_free(tmp);
2724 	ulist_free(qgroups);
2725 	ulist_free(old_roots);
2726 	ulist_free(new_roots);
2727 	return ret;
2728 }
2729 
2730 int btrfs_qgroup_account_extents(struct btrfs_trans_handle *trans)
2731 {
2732 	struct btrfs_fs_info *fs_info = trans->fs_info;
2733 	struct btrfs_qgroup_extent_record *record;
2734 	struct btrfs_delayed_ref_root *delayed_refs;
2735 	struct ulist *new_roots = NULL;
2736 	struct rb_node *node;
2737 	u64 num_dirty_extents = 0;
2738 	u64 qgroup_to_skip;
2739 	int ret = 0;
2740 
2741 	delayed_refs = &trans->transaction->delayed_refs;
2742 	qgroup_to_skip = delayed_refs->qgroup_to_skip;
2743 	while ((node = rb_first(&delayed_refs->dirty_extent_root))) {
2744 		record = rb_entry(node, struct btrfs_qgroup_extent_record,
2745 				  node);
2746 
2747 		num_dirty_extents++;
2748 		trace_btrfs_qgroup_account_extents(fs_info, record);
2749 
2750 		if (!ret && !(fs_info->qgroup_flags &
2751 			      BTRFS_QGROUP_RUNTIME_FLAG_NO_ACCOUNTING)) {
2752 			/*
2753 			 * Old roots should be searched when inserting qgroup
2754 			 * extent record
2755 			 */
2756 			if (WARN_ON(!record->old_roots)) {
2757 				/* Search commit root to find old_roots */
2758 				ret = btrfs_find_all_roots(NULL, fs_info,
2759 						record->bytenr, 0,
2760 						&record->old_roots, false);
2761 				if (ret < 0)
2762 					goto cleanup;
2763 			}
2764 
2765 			/* Free the reserved data space */
2766 			btrfs_qgroup_free_refroot(fs_info,
2767 					record->data_rsv_refroot,
2768 					record->data_rsv,
2769 					BTRFS_QGROUP_RSV_DATA);
2770 			/*
2771 			 * Use BTRFS_SEQ_LAST as time_seq to do special search,
2772 			 * which doesn't lock tree or delayed_refs and search
2773 			 * current root. It's safe inside commit_transaction().
2774 			 */
2775 			ret = btrfs_find_all_roots(trans, fs_info,
2776 			   record->bytenr, BTRFS_SEQ_LAST, &new_roots, false);
2777 			if (ret < 0)
2778 				goto cleanup;
2779 			if (qgroup_to_skip) {
2780 				ulist_del(new_roots, qgroup_to_skip, 0);
2781 				ulist_del(record->old_roots, qgroup_to_skip,
2782 					  0);
2783 			}
2784 			ret = btrfs_qgroup_account_extent(trans, record->bytenr,
2785 							  record->num_bytes,
2786 							  record->old_roots,
2787 							  new_roots);
2788 			record->old_roots = NULL;
2789 			new_roots = NULL;
2790 		}
2791 cleanup:
2792 		ulist_free(record->old_roots);
2793 		ulist_free(new_roots);
2794 		new_roots = NULL;
2795 		rb_erase(node, &delayed_refs->dirty_extent_root);
2796 		kfree(record);
2797 
2798 	}
2799 	trace_qgroup_num_dirty_extents(fs_info, trans->transid,
2800 				       num_dirty_extents);
2801 	return ret;
2802 }
2803 
2804 /*
2805  * called from commit_transaction. Writes all changed qgroups to disk.
2806  */
2807 int btrfs_run_qgroups(struct btrfs_trans_handle *trans)
2808 {
2809 	struct btrfs_fs_info *fs_info = trans->fs_info;
2810 	int ret = 0;
2811 
2812 	if (!fs_info->quota_root)
2813 		return ret;
2814 
2815 	spin_lock(&fs_info->qgroup_lock);
2816 	while (!list_empty(&fs_info->dirty_qgroups)) {
2817 		struct btrfs_qgroup *qgroup;
2818 		qgroup = list_first_entry(&fs_info->dirty_qgroups,
2819 					  struct btrfs_qgroup, dirty);
2820 		list_del_init(&qgroup->dirty);
2821 		spin_unlock(&fs_info->qgroup_lock);
2822 		ret = update_qgroup_info_item(trans, qgroup);
2823 		if (ret)
2824 			qgroup_mark_inconsistent(fs_info);
2825 		ret = update_qgroup_limit_item(trans, qgroup);
2826 		if (ret)
2827 			qgroup_mark_inconsistent(fs_info);
2828 		spin_lock(&fs_info->qgroup_lock);
2829 	}
2830 	if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
2831 		fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_ON;
2832 	else
2833 		fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_ON;
2834 	spin_unlock(&fs_info->qgroup_lock);
2835 
2836 	ret = update_qgroup_status_item(trans);
2837 	if (ret)
2838 		qgroup_mark_inconsistent(fs_info);
2839 
2840 	return ret;
2841 }
2842 
2843 /*
2844  * Copy the accounting information between qgroups. This is necessary
2845  * when a snapshot or a subvolume is created. Throwing an error will
2846  * cause a transaction abort so we take extra care here to only error
2847  * when a readonly fs is a reasonable outcome.
2848  */
2849 int btrfs_qgroup_inherit(struct btrfs_trans_handle *trans, u64 srcid,
2850 			 u64 objectid, struct btrfs_qgroup_inherit *inherit)
2851 {
2852 	int ret = 0;
2853 	int i;
2854 	u64 *i_qgroups;
2855 	bool committing = false;
2856 	struct btrfs_fs_info *fs_info = trans->fs_info;
2857 	struct btrfs_root *quota_root;
2858 	struct btrfs_qgroup *srcgroup;
2859 	struct btrfs_qgroup *dstgroup;
2860 	bool need_rescan = false;
2861 	u32 level_size = 0;
2862 	u64 nums;
2863 
2864 	/*
2865 	 * There are only two callers of this function.
2866 	 *
2867 	 * One in create_subvol() in the ioctl context, which needs to hold
2868 	 * the qgroup_ioctl_lock.
2869 	 *
2870 	 * The other one in create_pending_snapshot() where no other qgroup
2871 	 * code can modify the fs as they all need to either start a new trans
2872 	 * or hold a trans handler, thus we don't need to hold
2873 	 * qgroup_ioctl_lock.
2874 	 * This would avoid long and complex lock chain and make lockdep happy.
2875 	 */
2876 	spin_lock(&fs_info->trans_lock);
2877 	if (trans->transaction->state == TRANS_STATE_COMMIT_DOING)
2878 		committing = true;
2879 	spin_unlock(&fs_info->trans_lock);
2880 
2881 	if (!committing)
2882 		mutex_lock(&fs_info->qgroup_ioctl_lock);
2883 	if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
2884 		goto out;
2885 
2886 	quota_root = fs_info->quota_root;
2887 	if (!quota_root) {
2888 		ret = -EINVAL;
2889 		goto out;
2890 	}
2891 
2892 	if (inherit) {
2893 		i_qgroups = (u64 *)(inherit + 1);
2894 		nums = inherit->num_qgroups + 2 * inherit->num_ref_copies +
2895 		       2 * inherit->num_excl_copies;
2896 		for (i = 0; i < nums; ++i) {
2897 			srcgroup = find_qgroup_rb(fs_info, *i_qgroups);
2898 
2899 			/*
2900 			 * Zero out invalid groups so we can ignore
2901 			 * them later.
2902 			 */
2903 			if (!srcgroup ||
2904 			    ((srcgroup->qgroupid >> 48) <= (objectid >> 48)))
2905 				*i_qgroups = 0ULL;
2906 
2907 			++i_qgroups;
2908 		}
2909 	}
2910 
2911 	/*
2912 	 * create a tracking group for the subvol itself
2913 	 */
2914 	ret = add_qgroup_item(trans, quota_root, objectid);
2915 	if (ret)
2916 		goto out;
2917 
2918 	/*
2919 	 * add qgroup to all inherited groups
2920 	 */
2921 	if (inherit) {
2922 		i_qgroups = (u64 *)(inherit + 1);
2923 		for (i = 0; i < inherit->num_qgroups; ++i, ++i_qgroups) {
2924 			if (*i_qgroups == 0)
2925 				continue;
2926 			ret = add_qgroup_relation_item(trans, objectid,
2927 						       *i_qgroups);
2928 			if (ret && ret != -EEXIST)
2929 				goto out;
2930 			ret = add_qgroup_relation_item(trans, *i_qgroups,
2931 						       objectid);
2932 			if (ret && ret != -EEXIST)
2933 				goto out;
2934 		}
2935 		ret = 0;
2936 	}
2937 
2938 
2939 	spin_lock(&fs_info->qgroup_lock);
2940 
2941 	dstgroup = add_qgroup_rb(fs_info, objectid);
2942 	if (IS_ERR(dstgroup)) {
2943 		ret = PTR_ERR(dstgroup);
2944 		goto unlock;
2945 	}
2946 
2947 	if (inherit && inherit->flags & BTRFS_QGROUP_INHERIT_SET_LIMITS) {
2948 		dstgroup->lim_flags = inherit->lim.flags;
2949 		dstgroup->max_rfer = inherit->lim.max_rfer;
2950 		dstgroup->max_excl = inherit->lim.max_excl;
2951 		dstgroup->rsv_rfer = inherit->lim.rsv_rfer;
2952 		dstgroup->rsv_excl = inherit->lim.rsv_excl;
2953 
2954 		ret = update_qgroup_limit_item(trans, dstgroup);
2955 		if (ret) {
2956 			qgroup_mark_inconsistent(fs_info);
2957 			btrfs_info(fs_info,
2958 				   "unable to update quota limit for %llu",
2959 				   dstgroup->qgroupid);
2960 			goto unlock;
2961 		}
2962 	}
2963 
2964 	if (srcid) {
2965 		srcgroup = find_qgroup_rb(fs_info, srcid);
2966 		if (!srcgroup)
2967 			goto unlock;
2968 
2969 		/*
2970 		 * We call inherit after we clone the root in order to make sure
2971 		 * our counts don't go crazy, so at this point the only
2972 		 * difference between the two roots should be the root node.
2973 		 */
2974 		level_size = fs_info->nodesize;
2975 		dstgroup->rfer = srcgroup->rfer;
2976 		dstgroup->rfer_cmpr = srcgroup->rfer_cmpr;
2977 		dstgroup->excl = level_size;
2978 		dstgroup->excl_cmpr = level_size;
2979 		srcgroup->excl = level_size;
2980 		srcgroup->excl_cmpr = level_size;
2981 
2982 		/* inherit the limit info */
2983 		dstgroup->lim_flags = srcgroup->lim_flags;
2984 		dstgroup->max_rfer = srcgroup->max_rfer;
2985 		dstgroup->max_excl = srcgroup->max_excl;
2986 		dstgroup->rsv_rfer = srcgroup->rsv_rfer;
2987 		dstgroup->rsv_excl = srcgroup->rsv_excl;
2988 
2989 		qgroup_dirty(fs_info, dstgroup);
2990 		qgroup_dirty(fs_info, srcgroup);
2991 	}
2992 
2993 	if (!inherit)
2994 		goto unlock;
2995 
2996 	i_qgroups = (u64 *)(inherit + 1);
2997 	for (i = 0; i < inherit->num_qgroups; ++i) {
2998 		if (*i_qgroups) {
2999 			ret = add_relation_rb(fs_info, objectid, *i_qgroups);
3000 			if (ret)
3001 				goto unlock;
3002 		}
3003 		++i_qgroups;
3004 
3005 		/*
3006 		 * If we're doing a snapshot, and adding the snapshot to a new
3007 		 * qgroup, the numbers are guaranteed to be incorrect.
3008 		 */
3009 		if (srcid)
3010 			need_rescan = true;
3011 	}
3012 
3013 	for (i = 0; i <  inherit->num_ref_copies; ++i, i_qgroups += 2) {
3014 		struct btrfs_qgroup *src;
3015 		struct btrfs_qgroup *dst;
3016 
3017 		if (!i_qgroups[0] || !i_qgroups[1])
3018 			continue;
3019 
3020 		src = find_qgroup_rb(fs_info, i_qgroups[0]);
3021 		dst = find_qgroup_rb(fs_info, i_qgroups[1]);
3022 
3023 		if (!src || !dst) {
3024 			ret = -EINVAL;
3025 			goto unlock;
3026 		}
3027 
3028 		dst->rfer = src->rfer - level_size;
3029 		dst->rfer_cmpr = src->rfer_cmpr - level_size;
3030 
3031 		/* Manually tweaking numbers certainly needs a rescan */
3032 		need_rescan = true;
3033 	}
3034 	for (i = 0; i <  inherit->num_excl_copies; ++i, i_qgroups += 2) {
3035 		struct btrfs_qgroup *src;
3036 		struct btrfs_qgroup *dst;
3037 
3038 		if (!i_qgroups[0] || !i_qgroups[1])
3039 			continue;
3040 
3041 		src = find_qgroup_rb(fs_info, i_qgroups[0]);
3042 		dst = find_qgroup_rb(fs_info, i_qgroups[1]);
3043 
3044 		if (!src || !dst) {
3045 			ret = -EINVAL;
3046 			goto unlock;
3047 		}
3048 
3049 		dst->excl = src->excl + level_size;
3050 		dst->excl_cmpr = src->excl_cmpr + level_size;
3051 		need_rescan = true;
3052 	}
3053 
3054 unlock:
3055 	spin_unlock(&fs_info->qgroup_lock);
3056 	if (!ret)
3057 		ret = btrfs_sysfs_add_one_qgroup(fs_info, dstgroup);
3058 out:
3059 	if (!committing)
3060 		mutex_unlock(&fs_info->qgroup_ioctl_lock);
3061 	if (need_rescan)
3062 		qgroup_mark_inconsistent(fs_info);
3063 	return ret;
3064 }
3065 
3066 static bool qgroup_check_limits(const struct btrfs_qgroup *qg, u64 num_bytes)
3067 {
3068 	if ((qg->lim_flags & BTRFS_QGROUP_LIMIT_MAX_RFER) &&
3069 	    qgroup_rsv_total(qg) + (s64)qg->rfer + num_bytes > qg->max_rfer)
3070 		return false;
3071 
3072 	if ((qg->lim_flags & BTRFS_QGROUP_LIMIT_MAX_EXCL) &&
3073 	    qgroup_rsv_total(qg) + (s64)qg->excl + num_bytes > qg->max_excl)
3074 		return false;
3075 
3076 	return true;
3077 }
3078 
3079 static int qgroup_reserve(struct btrfs_root *root, u64 num_bytes, bool enforce,
3080 			  enum btrfs_qgroup_rsv_type type)
3081 {
3082 	struct btrfs_qgroup *qgroup;
3083 	struct btrfs_fs_info *fs_info = root->fs_info;
3084 	u64 ref_root = root->root_key.objectid;
3085 	int ret = 0;
3086 	struct ulist_node *unode;
3087 	struct ulist_iterator uiter;
3088 
3089 	if (!is_fstree(ref_root))
3090 		return 0;
3091 
3092 	if (num_bytes == 0)
3093 		return 0;
3094 
3095 	if (test_bit(BTRFS_FS_QUOTA_OVERRIDE, &fs_info->flags) &&
3096 	    capable(CAP_SYS_RESOURCE))
3097 		enforce = false;
3098 
3099 	spin_lock(&fs_info->qgroup_lock);
3100 	if (!fs_info->quota_root)
3101 		goto out;
3102 
3103 	qgroup = find_qgroup_rb(fs_info, ref_root);
3104 	if (!qgroup)
3105 		goto out;
3106 
3107 	/*
3108 	 * in a first step, we check all affected qgroups if any limits would
3109 	 * be exceeded
3110 	 */
3111 	ulist_reinit(fs_info->qgroup_ulist);
3112 	ret = ulist_add(fs_info->qgroup_ulist, qgroup->qgroupid,
3113 			qgroup_to_aux(qgroup), GFP_ATOMIC);
3114 	if (ret < 0)
3115 		goto out;
3116 	ULIST_ITER_INIT(&uiter);
3117 	while ((unode = ulist_next(fs_info->qgroup_ulist, &uiter))) {
3118 		struct btrfs_qgroup *qg;
3119 		struct btrfs_qgroup_list *glist;
3120 
3121 		qg = unode_aux_to_qgroup(unode);
3122 
3123 		if (enforce && !qgroup_check_limits(qg, num_bytes)) {
3124 			ret = -EDQUOT;
3125 			goto out;
3126 		}
3127 
3128 		list_for_each_entry(glist, &qg->groups, next_group) {
3129 			ret = ulist_add(fs_info->qgroup_ulist,
3130 					glist->group->qgroupid,
3131 					qgroup_to_aux(glist->group), GFP_ATOMIC);
3132 			if (ret < 0)
3133 				goto out;
3134 		}
3135 	}
3136 	ret = 0;
3137 	/*
3138 	 * no limits exceeded, now record the reservation into all qgroups
3139 	 */
3140 	ULIST_ITER_INIT(&uiter);
3141 	while ((unode = ulist_next(fs_info->qgroup_ulist, &uiter))) {
3142 		struct btrfs_qgroup *qg;
3143 
3144 		qg = unode_aux_to_qgroup(unode);
3145 
3146 		qgroup_rsv_add(fs_info, qg, num_bytes, type);
3147 	}
3148 
3149 out:
3150 	spin_unlock(&fs_info->qgroup_lock);
3151 	return ret;
3152 }
3153 
3154 /*
3155  * Free @num_bytes of reserved space with @type for qgroup.  (Normally level 0
3156  * qgroup).
3157  *
3158  * Will handle all higher level qgroup too.
3159  *
3160  * NOTE: If @num_bytes is (u64)-1, this means to free all bytes of this qgroup.
3161  * This special case is only used for META_PERTRANS type.
3162  */
3163 void btrfs_qgroup_free_refroot(struct btrfs_fs_info *fs_info,
3164 			       u64 ref_root, u64 num_bytes,
3165 			       enum btrfs_qgroup_rsv_type type)
3166 {
3167 	struct btrfs_qgroup *qgroup;
3168 	struct ulist_node *unode;
3169 	struct ulist_iterator uiter;
3170 	int ret = 0;
3171 
3172 	if (!is_fstree(ref_root))
3173 		return;
3174 
3175 	if (num_bytes == 0)
3176 		return;
3177 
3178 	if (num_bytes == (u64)-1 && type != BTRFS_QGROUP_RSV_META_PERTRANS) {
3179 		WARN(1, "%s: Invalid type to free", __func__);
3180 		return;
3181 	}
3182 	spin_lock(&fs_info->qgroup_lock);
3183 
3184 	if (!fs_info->quota_root)
3185 		goto out;
3186 
3187 	qgroup = find_qgroup_rb(fs_info, ref_root);
3188 	if (!qgroup)
3189 		goto out;
3190 
3191 	if (num_bytes == (u64)-1)
3192 		/*
3193 		 * We're freeing all pertrans rsv, get reserved value from
3194 		 * level 0 qgroup as real num_bytes to free.
3195 		 */
3196 		num_bytes = qgroup->rsv.values[type];
3197 
3198 	ulist_reinit(fs_info->qgroup_ulist);
3199 	ret = ulist_add(fs_info->qgroup_ulist, qgroup->qgroupid,
3200 			qgroup_to_aux(qgroup), GFP_ATOMIC);
3201 	if (ret < 0)
3202 		goto out;
3203 	ULIST_ITER_INIT(&uiter);
3204 	while ((unode = ulist_next(fs_info->qgroup_ulist, &uiter))) {
3205 		struct btrfs_qgroup *qg;
3206 		struct btrfs_qgroup_list *glist;
3207 
3208 		qg = unode_aux_to_qgroup(unode);
3209 
3210 		qgroup_rsv_release(fs_info, qg, num_bytes, type);
3211 
3212 		list_for_each_entry(glist, &qg->groups, next_group) {
3213 			ret = ulist_add(fs_info->qgroup_ulist,
3214 					glist->group->qgroupid,
3215 					qgroup_to_aux(glist->group), GFP_ATOMIC);
3216 			if (ret < 0)
3217 				goto out;
3218 		}
3219 	}
3220 
3221 out:
3222 	spin_unlock(&fs_info->qgroup_lock);
3223 }
3224 
3225 /*
3226  * Check if the leaf is the last leaf. Which means all node pointers
3227  * are at their last position.
3228  */
3229 static bool is_last_leaf(struct btrfs_path *path)
3230 {
3231 	int i;
3232 
3233 	for (i = 1; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
3234 		if (path->slots[i] != btrfs_header_nritems(path->nodes[i]) - 1)
3235 			return false;
3236 	}
3237 	return true;
3238 }
3239 
3240 /*
3241  * returns < 0 on error, 0 when more leafs are to be scanned.
3242  * returns 1 when done.
3243  */
3244 static int qgroup_rescan_leaf(struct btrfs_trans_handle *trans,
3245 			      struct btrfs_path *path)
3246 {
3247 	struct btrfs_fs_info *fs_info = trans->fs_info;
3248 	struct btrfs_root *extent_root;
3249 	struct btrfs_key found;
3250 	struct extent_buffer *scratch_leaf = NULL;
3251 	struct ulist *roots = NULL;
3252 	u64 num_bytes;
3253 	bool done;
3254 	int slot;
3255 	int ret;
3256 
3257 	mutex_lock(&fs_info->qgroup_rescan_lock);
3258 	extent_root = btrfs_extent_root(fs_info,
3259 				fs_info->qgroup_rescan_progress.objectid);
3260 	ret = btrfs_search_slot_for_read(extent_root,
3261 					 &fs_info->qgroup_rescan_progress,
3262 					 path, 1, 0);
3263 
3264 	btrfs_debug(fs_info,
3265 		"current progress key (%llu %u %llu), search_slot ret %d",
3266 		fs_info->qgroup_rescan_progress.objectid,
3267 		fs_info->qgroup_rescan_progress.type,
3268 		fs_info->qgroup_rescan_progress.offset, ret);
3269 
3270 	if (ret) {
3271 		/*
3272 		 * The rescan is about to end, we will not be scanning any
3273 		 * further blocks. We cannot unset the RESCAN flag here, because
3274 		 * we want to commit the transaction if everything went well.
3275 		 * To make the live accounting work in this phase, we set our
3276 		 * scan progress pointer such that every real extent objectid
3277 		 * will be smaller.
3278 		 */
3279 		fs_info->qgroup_rescan_progress.objectid = (u64)-1;
3280 		btrfs_release_path(path);
3281 		mutex_unlock(&fs_info->qgroup_rescan_lock);
3282 		return ret;
3283 	}
3284 	done = is_last_leaf(path);
3285 
3286 	btrfs_item_key_to_cpu(path->nodes[0], &found,
3287 			      btrfs_header_nritems(path->nodes[0]) - 1);
3288 	fs_info->qgroup_rescan_progress.objectid = found.objectid + 1;
3289 
3290 	scratch_leaf = btrfs_clone_extent_buffer(path->nodes[0]);
3291 	if (!scratch_leaf) {
3292 		ret = -ENOMEM;
3293 		mutex_unlock(&fs_info->qgroup_rescan_lock);
3294 		goto out;
3295 	}
3296 	slot = path->slots[0];
3297 	btrfs_release_path(path);
3298 	mutex_unlock(&fs_info->qgroup_rescan_lock);
3299 
3300 	for (; slot < btrfs_header_nritems(scratch_leaf); ++slot) {
3301 		btrfs_item_key_to_cpu(scratch_leaf, &found, slot);
3302 		if (found.type != BTRFS_EXTENT_ITEM_KEY &&
3303 		    found.type != BTRFS_METADATA_ITEM_KEY)
3304 			continue;
3305 		if (found.type == BTRFS_METADATA_ITEM_KEY)
3306 			num_bytes = fs_info->nodesize;
3307 		else
3308 			num_bytes = found.offset;
3309 
3310 		ret = btrfs_find_all_roots(NULL, fs_info, found.objectid, 0,
3311 					   &roots, false);
3312 		if (ret < 0)
3313 			goto out;
3314 		/* For rescan, just pass old_roots as NULL */
3315 		ret = btrfs_qgroup_account_extent(trans, found.objectid,
3316 						  num_bytes, NULL, roots);
3317 		if (ret < 0)
3318 			goto out;
3319 	}
3320 out:
3321 	if (scratch_leaf)
3322 		free_extent_buffer(scratch_leaf);
3323 
3324 	if (done && !ret) {
3325 		ret = 1;
3326 		fs_info->qgroup_rescan_progress.objectid = (u64)-1;
3327 	}
3328 	return ret;
3329 }
3330 
3331 static bool rescan_should_stop(struct btrfs_fs_info *fs_info)
3332 {
3333 	return btrfs_fs_closing(fs_info) ||
3334 		test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state) ||
3335 		!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) ||
3336 			  fs_info->qgroup_flags & BTRFS_QGROUP_RUNTIME_FLAG_CANCEL_RESCAN;
3337 }
3338 
3339 static void btrfs_qgroup_rescan_worker(struct btrfs_work *work)
3340 {
3341 	struct btrfs_fs_info *fs_info = container_of(work, struct btrfs_fs_info,
3342 						     qgroup_rescan_work);
3343 	struct btrfs_path *path;
3344 	struct btrfs_trans_handle *trans = NULL;
3345 	int err = -ENOMEM;
3346 	int ret = 0;
3347 	bool stopped = false;
3348 
3349 	path = btrfs_alloc_path();
3350 	if (!path)
3351 		goto out;
3352 	/*
3353 	 * Rescan should only search for commit root, and any later difference
3354 	 * should be recorded by qgroup
3355 	 */
3356 	path->search_commit_root = 1;
3357 	path->skip_locking = 1;
3358 
3359 	err = 0;
3360 	while (!err && !(stopped = rescan_should_stop(fs_info))) {
3361 		trans = btrfs_start_transaction(fs_info->fs_root, 0);
3362 		if (IS_ERR(trans)) {
3363 			err = PTR_ERR(trans);
3364 			break;
3365 		}
3366 
3367 		err = qgroup_rescan_leaf(trans, path);
3368 
3369 		if (err > 0)
3370 			btrfs_commit_transaction(trans);
3371 		else
3372 			btrfs_end_transaction(trans);
3373 	}
3374 
3375 out:
3376 	btrfs_free_path(path);
3377 
3378 	mutex_lock(&fs_info->qgroup_rescan_lock);
3379 	if (err > 0 &&
3380 	    fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT) {
3381 		fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
3382 	} else if (err < 0 || stopped) {
3383 		fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
3384 	}
3385 	mutex_unlock(&fs_info->qgroup_rescan_lock);
3386 
3387 	/*
3388 	 * only update status, since the previous part has already updated the
3389 	 * qgroup info.
3390 	 */
3391 	trans = btrfs_start_transaction(fs_info->quota_root, 1);
3392 	if (IS_ERR(trans)) {
3393 		err = PTR_ERR(trans);
3394 		trans = NULL;
3395 		btrfs_err(fs_info,
3396 			  "fail to start transaction for status update: %d",
3397 			  err);
3398 	}
3399 
3400 	mutex_lock(&fs_info->qgroup_rescan_lock);
3401 	if (!stopped ||
3402 	    fs_info->qgroup_flags & BTRFS_QGROUP_RUNTIME_FLAG_CANCEL_RESCAN)
3403 		fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
3404 	if (trans) {
3405 		ret = update_qgroup_status_item(trans);
3406 		if (ret < 0) {
3407 			err = ret;
3408 			btrfs_err(fs_info, "fail to update qgroup status: %d",
3409 				  err);
3410 		}
3411 	}
3412 	fs_info->qgroup_rescan_running = false;
3413 	fs_info->qgroup_flags &= ~BTRFS_QGROUP_RUNTIME_FLAG_CANCEL_RESCAN;
3414 	complete_all(&fs_info->qgroup_rescan_completion);
3415 	mutex_unlock(&fs_info->qgroup_rescan_lock);
3416 
3417 	if (!trans)
3418 		return;
3419 
3420 	btrfs_end_transaction(trans);
3421 
3422 	if (stopped) {
3423 		btrfs_info(fs_info, "qgroup scan paused");
3424 	} else if (fs_info->qgroup_flags & BTRFS_QGROUP_RUNTIME_FLAG_CANCEL_RESCAN) {
3425 		btrfs_info(fs_info, "qgroup scan cancelled");
3426 	} else if (err >= 0) {
3427 		btrfs_info(fs_info, "qgroup scan completed%s",
3428 			err > 0 ? " (inconsistency flag cleared)" : "");
3429 	} else {
3430 		btrfs_err(fs_info, "qgroup scan failed with %d", err);
3431 	}
3432 }
3433 
3434 /*
3435  * Checks that (a) no rescan is running and (b) quota is enabled. Allocates all
3436  * memory required for the rescan context.
3437  */
3438 static int
3439 qgroup_rescan_init(struct btrfs_fs_info *fs_info, u64 progress_objectid,
3440 		   int init_flags)
3441 {
3442 	int ret = 0;
3443 
3444 	if (!init_flags) {
3445 		/* we're resuming qgroup rescan at mount time */
3446 		if (!(fs_info->qgroup_flags &
3447 		      BTRFS_QGROUP_STATUS_FLAG_RESCAN)) {
3448 			btrfs_warn(fs_info,
3449 			"qgroup rescan init failed, qgroup rescan is not queued");
3450 			ret = -EINVAL;
3451 		} else if (!(fs_info->qgroup_flags &
3452 			     BTRFS_QGROUP_STATUS_FLAG_ON)) {
3453 			btrfs_warn(fs_info,
3454 			"qgroup rescan init failed, qgroup is not enabled");
3455 			ret = -EINVAL;
3456 		}
3457 
3458 		if (ret)
3459 			return ret;
3460 	}
3461 
3462 	mutex_lock(&fs_info->qgroup_rescan_lock);
3463 
3464 	if (init_flags) {
3465 		if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN) {
3466 			btrfs_warn(fs_info,
3467 				   "qgroup rescan is already in progress");
3468 			ret = -EINPROGRESS;
3469 		} else if (!(fs_info->qgroup_flags &
3470 			     BTRFS_QGROUP_STATUS_FLAG_ON)) {
3471 			btrfs_warn(fs_info,
3472 			"qgroup rescan init failed, qgroup is not enabled");
3473 			ret = -EINVAL;
3474 		} else if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)) {
3475 			/* Quota disable is in progress */
3476 			ret = -EBUSY;
3477 		}
3478 
3479 		if (ret) {
3480 			mutex_unlock(&fs_info->qgroup_rescan_lock);
3481 			return ret;
3482 		}
3483 		fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_RESCAN;
3484 	}
3485 
3486 	memset(&fs_info->qgroup_rescan_progress, 0,
3487 		sizeof(fs_info->qgroup_rescan_progress));
3488 	fs_info->qgroup_flags &= ~(BTRFS_QGROUP_RUNTIME_FLAG_CANCEL_RESCAN |
3489 				   BTRFS_QGROUP_RUNTIME_FLAG_NO_ACCOUNTING);
3490 	fs_info->qgroup_rescan_progress.objectid = progress_objectid;
3491 	init_completion(&fs_info->qgroup_rescan_completion);
3492 	mutex_unlock(&fs_info->qgroup_rescan_lock);
3493 
3494 	btrfs_init_work(&fs_info->qgroup_rescan_work,
3495 			btrfs_qgroup_rescan_worker, NULL, NULL);
3496 	return 0;
3497 }
3498 
3499 static void
3500 qgroup_rescan_zero_tracking(struct btrfs_fs_info *fs_info)
3501 {
3502 	struct rb_node *n;
3503 	struct btrfs_qgroup *qgroup;
3504 
3505 	spin_lock(&fs_info->qgroup_lock);
3506 	/* clear all current qgroup tracking information */
3507 	for (n = rb_first(&fs_info->qgroup_tree); n; n = rb_next(n)) {
3508 		qgroup = rb_entry(n, struct btrfs_qgroup, node);
3509 		qgroup->rfer = 0;
3510 		qgroup->rfer_cmpr = 0;
3511 		qgroup->excl = 0;
3512 		qgroup->excl_cmpr = 0;
3513 		qgroup_dirty(fs_info, qgroup);
3514 	}
3515 	spin_unlock(&fs_info->qgroup_lock);
3516 }
3517 
3518 int
3519 btrfs_qgroup_rescan(struct btrfs_fs_info *fs_info)
3520 {
3521 	int ret = 0;
3522 	struct btrfs_trans_handle *trans;
3523 
3524 	ret = qgroup_rescan_init(fs_info, 0, 1);
3525 	if (ret)
3526 		return ret;
3527 
3528 	/*
3529 	 * We have set the rescan_progress to 0, which means no more
3530 	 * delayed refs will be accounted by btrfs_qgroup_account_ref.
3531 	 * However, btrfs_qgroup_account_ref may be right after its call
3532 	 * to btrfs_find_all_roots, in which case it would still do the
3533 	 * accounting.
3534 	 * To solve this, we're committing the transaction, which will
3535 	 * ensure we run all delayed refs and only after that, we are
3536 	 * going to clear all tracking information for a clean start.
3537 	 */
3538 
3539 	trans = btrfs_join_transaction(fs_info->fs_root);
3540 	if (IS_ERR(trans)) {
3541 		fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
3542 		return PTR_ERR(trans);
3543 	}
3544 	ret = btrfs_commit_transaction(trans);
3545 	if (ret) {
3546 		fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
3547 		return ret;
3548 	}
3549 
3550 	qgroup_rescan_zero_tracking(fs_info);
3551 
3552 	mutex_lock(&fs_info->qgroup_rescan_lock);
3553 	fs_info->qgroup_rescan_running = true;
3554 	btrfs_queue_work(fs_info->qgroup_rescan_workers,
3555 			 &fs_info->qgroup_rescan_work);
3556 	mutex_unlock(&fs_info->qgroup_rescan_lock);
3557 
3558 	return 0;
3559 }
3560 
3561 int btrfs_qgroup_wait_for_completion(struct btrfs_fs_info *fs_info,
3562 				     bool interruptible)
3563 {
3564 	int running;
3565 	int ret = 0;
3566 
3567 	mutex_lock(&fs_info->qgroup_rescan_lock);
3568 	running = fs_info->qgroup_rescan_running;
3569 	mutex_unlock(&fs_info->qgroup_rescan_lock);
3570 
3571 	if (!running)
3572 		return 0;
3573 
3574 	if (interruptible)
3575 		ret = wait_for_completion_interruptible(
3576 					&fs_info->qgroup_rescan_completion);
3577 	else
3578 		wait_for_completion(&fs_info->qgroup_rescan_completion);
3579 
3580 	return ret;
3581 }
3582 
3583 /*
3584  * this is only called from open_ctree where we're still single threaded, thus
3585  * locking is omitted here.
3586  */
3587 void
3588 btrfs_qgroup_rescan_resume(struct btrfs_fs_info *fs_info)
3589 {
3590 	if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN) {
3591 		mutex_lock(&fs_info->qgroup_rescan_lock);
3592 		fs_info->qgroup_rescan_running = true;
3593 		btrfs_queue_work(fs_info->qgroup_rescan_workers,
3594 				 &fs_info->qgroup_rescan_work);
3595 		mutex_unlock(&fs_info->qgroup_rescan_lock);
3596 	}
3597 }
3598 
3599 #define rbtree_iterate_from_safe(node, next, start)				\
3600        for (node = start; node && ({ next = rb_next(node); 1;}); node = next)
3601 
3602 static int qgroup_unreserve_range(struct btrfs_inode *inode,
3603 				  struct extent_changeset *reserved, u64 start,
3604 				  u64 len)
3605 {
3606 	struct rb_node *node;
3607 	struct rb_node *next;
3608 	struct ulist_node *entry;
3609 	int ret = 0;
3610 
3611 	node = reserved->range_changed.root.rb_node;
3612 	if (!node)
3613 		return 0;
3614 	while (node) {
3615 		entry = rb_entry(node, struct ulist_node, rb_node);
3616 		if (entry->val < start)
3617 			node = node->rb_right;
3618 		else
3619 			node = node->rb_left;
3620 	}
3621 
3622 	if (entry->val > start && rb_prev(&entry->rb_node))
3623 		entry = rb_entry(rb_prev(&entry->rb_node), struct ulist_node,
3624 				 rb_node);
3625 
3626 	rbtree_iterate_from_safe(node, next, &entry->rb_node) {
3627 		u64 entry_start;
3628 		u64 entry_end;
3629 		u64 entry_len;
3630 		int clear_ret;
3631 
3632 		entry = rb_entry(node, struct ulist_node, rb_node);
3633 		entry_start = entry->val;
3634 		entry_end = entry->aux;
3635 		entry_len = entry_end - entry_start + 1;
3636 
3637 		if (entry_start >= start + len)
3638 			break;
3639 		if (entry_start + entry_len <= start)
3640 			continue;
3641 		/*
3642 		 * Now the entry is in [start, start + len), revert the
3643 		 * EXTENT_QGROUP_RESERVED bit.
3644 		 */
3645 		clear_ret = clear_extent_bits(&inode->io_tree, entry_start,
3646 					      entry_end, EXTENT_QGROUP_RESERVED);
3647 		if (!ret && clear_ret < 0)
3648 			ret = clear_ret;
3649 
3650 		ulist_del(&reserved->range_changed, entry->val, entry->aux);
3651 		if (likely(reserved->bytes_changed >= entry_len)) {
3652 			reserved->bytes_changed -= entry_len;
3653 		} else {
3654 			WARN_ON(1);
3655 			reserved->bytes_changed = 0;
3656 		}
3657 	}
3658 
3659 	return ret;
3660 }
3661 
3662 /*
3663  * Try to free some space for qgroup.
3664  *
3665  * For qgroup, there are only 3 ways to free qgroup space:
3666  * - Flush nodatacow write
3667  *   Any nodatacow write will free its reserved data space at run_delalloc_range().
3668  *   In theory, we should only flush nodatacow inodes, but it's not yet
3669  *   possible, so we need to flush the whole root.
3670  *
3671  * - Wait for ordered extents
3672  *   When ordered extents are finished, their reserved metadata is finally
3673  *   converted to per_trans status, which can be freed by later commit
3674  *   transaction.
3675  *
3676  * - Commit transaction
3677  *   This would free the meta_per_trans space.
3678  *   In theory this shouldn't provide much space, but any more qgroup space
3679  *   is needed.
3680  */
3681 static int try_flush_qgroup(struct btrfs_root *root)
3682 {
3683 	struct btrfs_trans_handle *trans;
3684 	int ret;
3685 
3686 	/* Can't hold an open transaction or we run the risk of deadlocking. */
3687 	ASSERT(current->journal_info == NULL);
3688 	if (WARN_ON(current->journal_info))
3689 		return 0;
3690 
3691 	/*
3692 	 * We don't want to run flush again and again, so if there is a running
3693 	 * one, we won't try to start a new flush, but exit directly.
3694 	 */
3695 	if (test_and_set_bit(BTRFS_ROOT_QGROUP_FLUSHING, &root->state)) {
3696 		wait_event(root->qgroup_flush_wait,
3697 			!test_bit(BTRFS_ROOT_QGROUP_FLUSHING, &root->state));
3698 		return 0;
3699 	}
3700 
3701 	ret = btrfs_start_delalloc_snapshot(root, true);
3702 	if (ret < 0)
3703 		goto out;
3704 	btrfs_wait_ordered_extents(root, U64_MAX, 0, (u64)-1);
3705 
3706 	trans = btrfs_join_transaction(root);
3707 	if (IS_ERR(trans)) {
3708 		ret = PTR_ERR(trans);
3709 		goto out;
3710 	}
3711 
3712 	ret = btrfs_commit_transaction(trans);
3713 out:
3714 	clear_bit(BTRFS_ROOT_QGROUP_FLUSHING, &root->state);
3715 	wake_up(&root->qgroup_flush_wait);
3716 	return ret;
3717 }
3718 
3719 static int qgroup_reserve_data(struct btrfs_inode *inode,
3720 			struct extent_changeset **reserved_ret, u64 start,
3721 			u64 len)
3722 {
3723 	struct btrfs_root *root = inode->root;
3724 	struct extent_changeset *reserved;
3725 	bool new_reserved = false;
3726 	u64 orig_reserved;
3727 	u64 to_reserve;
3728 	int ret;
3729 
3730 	if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &root->fs_info->flags) ||
3731 	    !is_fstree(root->root_key.objectid) || len == 0)
3732 		return 0;
3733 
3734 	/* @reserved parameter is mandatory for qgroup */
3735 	if (WARN_ON(!reserved_ret))
3736 		return -EINVAL;
3737 	if (!*reserved_ret) {
3738 		new_reserved = true;
3739 		*reserved_ret = extent_changeset_alloc();
3740 		if (!*reserved_ret)
3741 			return -ENOMEM;
3742 	}
3743 	reserved = *reserved_ret;
3744 	/* Record already reserved space */
3745 	orig_reserved = reserved->bytes_changed;
3746 	ret = set_record_extent_bits(&inode->io_tree, start,
3747 			start + len -1, EXTENT_QGROUP_RESERVED, reserved);
3748 
3749 	/* Newly reserved space */
3750 	to_reserve = reserved->bytes_changed - orig_reserved;
3751 	trace_btrfs_qgroup_reserve_data(&inode->vfs_inode, start, len,
3752 					to_reserve, QGROUP_RESERVE);
3753 	if (ret < 0)
3754 		goto out;
3755 	ret = qgroup_reserve(root, to_reserve, true, BTRFS_QGROUP_RSV_DATA);
3756 	if (ret < 0)
3757 		goto cleanup;
3758 
3759 	return ret;
3760 
3761 cleanup:
3762 	qgroup_unreserve_range(inode, reserved, start, len);
3763 out:
3764 	if (new_reserved) {
3765 		extent_changeset_free(reserved);
3766 		*reserved_ret = NULL;
3767 	}
3768 	return ret;
3769 }
3770 
3771 /*
3772  * Reserve qgroup space for range [start, start + len).
3773  *
3774  * This function will either reserve space from related qgroups or do nothing
3775  * if the range is already reserved.
3776  *
3777  * Return 0 for successful reservation
3778  * Return <0 for error (including -EQUOT)
3779  *
3780  * NOTE: This function may sleep for memory allocation, dirty page flushing and
3781  *	 commit transaction. So caller should not hold any dirty page locked.
3782  */
3783 int btrfs_qgroup_reserve_data(struct btrfs_inode *inode,
3784 			struct extent_changeset **reserved_ret, u64 start,
3785 			u64 len)
3786 {
3787 	int ret;
3788 
3789 	ret = qgroup_reserve_data(inode, reserved_ret, start, len);
3790 	if (ret <= 0 && ret != -EDQUOT)
3791 		return ret;
3792 
3793 	ret = try_flush_qgroup(inode->root);
3794 	if (ret < 0)
3795 		return ret;
3796 	return qgroup_reserve_data(inode, reserved_ret, start, len);
3797 }
3798 
3799 /* Free ranges specified by @reserved, normally in error path */
3800 static int qgroup_free_reserved_data(struct btrfs_inode *inode,
3801 			struct extent_changeset *reserved, u64 start, u64 len)
3802 {
3803 	struct btrfs_root *root = inode->root;
3804 	struct ulist_node *unode;
3805 	struct ulist_iterator uiter;
3806 	struct extent_changeset changeset;
3807 	int freed = 0;
3808 	int ret;
3809 
3810 	extent_changeset_init(&changeset);
3811 	len = round_up(start + len, root->fs_info->sectorsize);
3812 	start = round_down(start, root->fs_info->sectorsize);
3813 
3814 	ULIST_ITER_INIT(&uiter);
3815 	while ((unode = ulist_next(&reserved->range_changed, &uiter))) {
3816 		u64 range_start = unode->val;
3817 		/* unode->aux is the inclusive end */
3818 		u64 range_len = unode->aux - range_start + 1;
3819 		u64 free_start;
3820 		u64 free_len;
3821 
3822 		extent_changeset_release(&changeset);
3823 
3824 		/* Only free range in range [start, start + len) */
3825 		if (range_start >= start + len ||
3826 		    range_start + range_len <= start)
3827 			continue;
3828 		free_start = max(range_start, start);
3829 		free_len = min(start + len, range_start + range_len) -
3830 			   free_start;
3831 		/*
3832 		 * TODO: To also modify reserved->ranges_reserved to reflect
3833 		 * the modification.
3834 		 *
3835 		 * However as long as we free qgroup reserved according to
3836 		 * EXTENT_QGROUP_RESERVED, we won't double free.
3837 		 * So not need to rush.
3838 		 */
3839 		ret = clear_record_extent_bits(&inode->io_tree, free_start,
3840 				free_start + free_len - 1,
3841 				EXTENT_QGROUP_RESERVED, &changeset);
3842 		if (ret < 0)
3843 			goto out;
3844 		freed += changeset.bytes_changed;
3845 	}
3846 	btrfs_qgroup_free_refroot(root->fs_info, root->root_key.objectid, freed,
3847 				  BTRFS_QGROUP_RSV_DATA);
3848 	ret = freed;
3849 out:
3850 	extent_changeset_release(&changeset);
3851 	return ret;
3852 }
3853 
3854 static int __btrfs_qgroup_release_data(struct btrfs_inode *inode,
3855 			struct extent_changeset *reserved, u64 start, u64 len,
3856 			int free)
3857 {
3858 	struct extent_changeset changeset;
3859 	int trace_op = QGROUP_RELEASE;
3860 	int ret;
3861 
3862 	if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &inode->root->fs_info->flags))
3863 		return 0;
3864 
3865 	/* In release case, we shouldn't have @reserved */
3866 	WARN_ON(!free && reserved);
3867 	if (free && reserved)
3868 		return qgroup_free_reserved_data(inode, reserved, start, len);
3869 	extent_changeset_init(&changeset);
3870 	ret = clear_record_extent_bits(&inode->io_tree, start, start + len -1,
3871 				       EXTENT_QGROUP_RESERVED, &changeset);
3872 	if (ret < 0)
3873 		goto out;
3874 
3875 	if (free)
3876 		trace_op = QGROUP_FREE;
3877 	trace_btrfs_qgroup_release_data(&inode->vfs_inode, start, len,
3878 					changeset.bytes_changed, trace_op);
3879 	if (free)
3880 		btrfs_qgroup_free_refroot(inode->root->fs_info,
3881 				inode->root->root_key.objectid,
3882 				changeset.bytes_changed, BTRFS_QGROUP_RSV_DATA);
3883 	ret = changeset.bytes_changed;
3884 out:
3885 	extent_changeset_release(&changeset);
3886 	return ret;
3887 }
3888 
3889 /*
3890  * Free a reserved space range from io_tree and related qgroups
3891  *
3892  * Should be called when a range of pages get invalidated before reaching disk.
3893  * Or for error cleanup case.
3894  * if @reserved is given, only reserved range in [@start, @start + @len) will
3895  * be freed.
3896  *
3897  * For data written to disk, use btrfs_qgroup_release_data().
3898  *
3899  * NOTE: This function may sleep for memory allocation.
3900  */
3901 int btrfs_qgroup_free_data(struct btrfs_inode *inode,
3902 			struct extent_changeset *reserved, u64 start, u64 len)
3903 {
3904 	return __btrfs_qgroup_release_data(inode, reserved, start, len, 1);
3905 }
3906 
3907 /*
3908  * Release a reserved space range from io_tree only.
3909  *
3910  * Should be called when a range of pages get written to disk and corresponding
3911  * FILE_EXTENT is inserted into corresponding root.
3912  *
3913  * Since new qgroup accounting framework will only update qgroup numbers at
3914  * commit_transaction() time, its reserved space shouldn't be freed from
3915  * related qgroups.
3916  *
3917  * But we should release the range from io_tree, to allow further write to be
3918  * COWed.
3919  *
3920  * NOTE: This function may sleep for memory allocation.
3921  */
3922 int btrfs_qgroup_release_data(struct btrfs_inode *inode, u64 start, u64 len)
3923 {
3924 	return __btrfs_qgroup_release_data(inode, NULL, start, len, 0);
3925 }
3926 
3927 static void add_root_meta_rsv(struct btrfs_root *root, int num_bytes,
3928 			      enum btrfs_qgroup_rsv_type type)
3929 {
3930 	if (type != BTRFS_QGROUP_RSV_META_PREALLOC &&
3931 	    type != BTRFS_QGROUP_RSV_META_PERTRANS)
3932 		return;
3933 	if (num_bytes == 0)
3934 		return;
3935 
3936 	spin_lock(&root->qgroup_meta_rsv_lock);
3937 	if (type == BTRFS_QGROUP_RSV_META_PREALLOC)
3938 		root->qgroup_meta_rsv_prealloc += num_bytes;
3939 	else
3940 		root->qgroup_meta_rsv_pertrans += num_bytes;
3941 	spin_unlock(&root->qgroup_meta_rsv_lock);
3942 }
3943 
3944 static int sub_root_meta_rsv(struct btrfs_root *root, int num_bytes,
3945 			     enum btrfs_qgroup_rsv_type type)
3946 {
3947 	if (type != BTRFS_QGROUP_RSV_META_PREALLOC &&
3948 	    type != BTRFS_QGROUP_RSV_META_PERTRANS)
3949 		return 0;
3950 	if (num_bytes == 0)
3951 		return 0;
3952 
3953 	spin_lock(&root->qgroup_meta_rsv_lock);
3954 	if (type == BTRFS_QGROUP_RSV_META_PREALLOC) {
3955 		num_bytes = min_t(u64, root->qgroup_meta_rsv_prealloc,
3956 				  num_bytes);
3957 		root->qgroup_meta_rsv_prealloc -= num_bytes;
3958 	} else {
3959 		num_bytes = min_t(u64, root->qgroup_meta_rsv_pertrans,
3960 				  num_bytes);
3961 		root->qgroup_meta_rsv_pertrans -= num_bytes;
3962 	}
3963 	spin_unlock(&root->qgroup_meta_rsv_lock);
3964 	return num_bytes;
3965 }
3966 
3967 int btrfs_qgroup_reserve_meta(struct btrfs_root *root, int num_bytes,
3968 			      enum btrfs_qgroup_rsv_type type, bool enforce)
3969 {
3970 	struct btrfs_fs_info *fs_info = root->fs_info;
3971 	int ret;
3972 
3973 	if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) ||
3974 	    !is_fstree(root->root_key.objectid) || num_bytes == 0)
3975 		return 0;
3976 
3977 	BUG_ON(num_bytes != round_down(num_bytes, fs_info->nodesize));
3978 	trace_qgroup_meta_reserve(root, (s64)num_bytes, type);
3979 	ret = qgroup_reserve(root, num_bytes, enforce, type);
3980 	if (ret < 0)
3981 		return ret;
3982 	/*
3983 	 * Record what we have reserved into root.
3984 	 *
3985 	 * To avoid quota disabled->enabled underflow.
3986 	 * In that case, we may try to free space we haven't reserved
3987 	 * (since quota was disabled), so record what we reserved into root.
3988 	 * And ensure later release won't underflow this number.
3989 	 */
3990 	add_root_meta_rsv(root, num_bytes, type);
3991 	return ret;
3992 }
3993 
3994 int __btrfs_qgroup_reserve_meta(struct btrfs_root *root, int num_bytes,
3995 				enum btrfs_qgroup_rsv_type type, bool enforce,
3996 				bool noflush)
3997 {
3998 	int ret;
3999 
4000 	ret = btrfs_qgroup_reserve_meta(root, num_bytes, type, enforce);
4001 	if ((ret <= 0 && ret != -EDQUOT) || noflush)
4002 		return ret;
4003 
4004 	ret = try_flush_qgroup(root);
4005 	if (ret < 0)
4006 		return ret;
4007 	return btrfs_qgroup_reserve_meta(root, num_bytes, type, enforce);
4008 }
4009 
4010 void btrfs_qgroup_free_meta_all_pertrans(struct btrfs_root *root)
4011 {
4012 	struct btrfs_fs_info *fs_info = root->fs_info;
4013 
4014 	if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) ||
4015 	    !is_fstree(root->root_key.objectid))
4016 		return;
4017 
4018 	/* TODO: Update trace point to handle such free */
4019 	trace_qgroup_meta_free_all_pertrans(root);
4020 	/* Special value -1 means to free all reserved space */
4021 	btrfs_qgroup_free_refroot(fs_info, root->root_key.objectid, (u64)-1,
4022 				  BTRFS_QGROUP_RSV_META_PERTRANS);
4023 }
4024 
4025 void __btrfs_qgroup_free_meta(struct btrfs_root *root, int num_bytes,
4026 			      enum btrfs_qgroup_rsv_type type)
4027 {
4028 	struct btrfs_fs_info *fs_info = root->fs_info;
4029 
4030 	if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) ||
4031 	    !is_fstree(root->root_key.objectid))
4032 		return;
4033 
4034 	/*
4035 	 * reservation for META_PREALLOC can happen before quota is enabled,
4036 	 * which can lead to underflow.
4037 	 * Here ensure we will only free what we really have reserved.
4038 	 */
4039 	num_bytes = sub_root_meta_rsv(root, num_bytes, type);
4040 	BUG_ON(num_bytes != round_down(num_bytes, fs_info->nodesize));
4041 	trace_qgroup_meta_reserve(root, -(s64)num_bytes, type);
4042 	btrfs_qgroup_free_refroot(fs_info, root->root_key.objectid,
4043 				  num_bytes, type);
4044 }
4045 
4046 static void qgroup_convert_meta(struct btrfs_fs_info *fs_info, u64 ref_root,
4047 				int num_bytes)
4048 {
4049 	struct btrfs_qgroup *qgroup;
4050 	struct ulist_node *unode;
4051 	struct ulist_iterator uiter;
4052 	int ret = 0;
4053 
4054 	if (num_bytes == 0)
4055 		return;
4056 	if (!fs_info->quota_root)
4057 		return;
4058 
4059 	spin_lock(&fs_info->qgroup_lock);
4060 	qgroup = find_qgroup_rb(fs_info, ref_root);
4061 	if (!qgroup)
4062 		goto out;
4063 	ulist_reinit(fs_info->qgroup_ulist);
4064 	ret = ulist_add(fs_info->qgroup_ulist, qgroup->qgroupid,
4065 		       qgroup_to_aux(qgroup), GFP_ATOMIC);
4066 	if (ret < 0)
4067 		goto out;
4068 	ULIST_ITER_INIT(&uiter);
4069 	while ((unode = ulist_next(fs_info->qgroup_ulist, &uiter))) {
4070 		struct btrfs_qgroup *qg;
4071 		struct btrfs_qgroup_list *glist;
4072 
4073 		qg = unode_aux_to_qgroup(unode);
4074 
4075 		qgroup_rsv_release(fs_info, qg, num_bytes,
4076 				BTRFS_QGROUP_RSV_META_PREALLOC);
4077 		qgroup_rsv_add(fs_info, qg, num_bytes,
4078 				BTRFS_QGROUP_RSV_META_PERTRANS);
4079 		list_for_each_entry(glist, &qg->groups, next_group) {
4080 			ret = ulist_add(fs_info->qgroup_ulist,
4081 					glist->group->qgroupid,
4082 					qgroup_to_aux(glist->group), GFP_ATOMIC);
4083 			if (ret < 0)
4084 				goto out;
4085 		}
4086 	}
4087 out:
4088 	spin_unlock(&fs_info->qgroup_lock);
4089 }
4090 
4091 void btrfs_qgroup_convert_reserved_meta(struct btrfs_root *root, int num_bytes)
4092 {
4093 	struct btrfs_fs_info *fs_info = root->fs_info;
4094 
4095 	if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) ||
4096 	    !is_fstree(root->root_key.objectid))
4097 		return;
4098 	/* Same as btrfs_qgroup_free_meta_prealloc() */
4099 	num_bytes = sub_root_meta_rsv(root, num_bytes,
4100 				      BTRFS_QGROUP_RSV_META_PREALLOC);
4101 	trace_qgroup_meta_convert(root, num_bytes);
4102 	qgroup_convert_meta(fs_info, root->root_key.objectid, num_bytes);
4103 }
4104 
4105 /*
4106  * Check qgroup reserved space leaking, normally at destroy inode
4107  * time
4108  */
4109 void btrfs_qgroup_check_reserved_leak(struct btrfs_inode *inode)
4110 {
4111 	struct extent_changeset changeset;
4112 	struct ulist_node *unode;
4113 	struct ulist_iterator iter;
4114 	int ret;
4115 
4116 	extent_changeset_init(&changeset);
4117 	ret = clear_record_extent_bits(&inode->io_tree, 0, (u64)-1,
4118 			EXTENT_QGROUP_RESERVED, &changeset);
4119 
4120 	WARN_ON(ret < 0);
4121 	if (WARN_ON(changeset.bytes_changed)) {
4122 		ULIST_ITER_INIT(&iter);
4123 		while ((unode = ulist_next(&changeset.range_changed, &iter))) {
4124 			btrfs_warn(inode->root->fs_info,
4125 		"leaking qgroup reserved space, ino: %llu, start: %llu, end: %llu",
4126 				btrfs_ino(inode), unode->val, unode->aux);
4127 		}
4128 		btrfs_qgroup_free_refroot(inode->root->fs_info,
4129 				inode->root->root_key.objectid,
4130 				changeset.bytes_changed, BTRFS_QGROUP_RSV_DATA);
4131 
4132 	}
4133 	extent_changeset_release(&changeset);
4134 }
4135 
4136 void btrfs_qgroup_init_swapped_blocks(
4137 	struct btrfs_qgroup_swapped_blocks *swapped_blocks)
4138 {
4139 	int i;
4140 
4141 	spin_lock_init(&swapped_blocks->lock);
4142 	for (i = 0; i < BTRFS_MAX_LEVEL; i++)
4143 		swapped_blocks->blocks[i] = RB_ROOT;
4144 	swapped_blocks->swapped = false;
4145 }
4146 
4147 /*
4148  * Delete all swapped blocks record of @root.
4149  * Every record here means we skipped a full subtree scan for qgroup.
4150  *
4151  * Gets called when committing one transaction.
4152  */
4153 void btrfs_qgroup_clean_swapped_blocks(struct btrfs_root *root)
4154 {
4155 	struct btrfs_qgroup_swapped_blocks *swapped_blocks;
4156 	int i;
4157 
4158 	swapped_blocks = &root->swapped_blocks;
4159 
4160 	spin_lock(&swapped_blocks->lock);
4161 	if (!swapped_blocks->swapped)
4162 		goto out;
4163 	for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
4164 		struct rb_root *cur_root = &swapped_blocks->blocks[i];
4165 		struct btrfs_qgroup_swapped_block *entry;
4166 		struct btrfs_qgroup_swapped_block *next;
4167 
4168 		rbtree_postorder_for_each_entry_safe(entry, next, cur_root,
4169 						     node)
4170 			kfree(entry);
4171 		swapped_blocks->blocks[i] = RB_ROOT;
4172 	}
4173 	swapped_blocks->swapped = false;
4174 out:
4175 	spin_unlock(&swapped_blocks->lock);
4176 }
4177 
4178 /*
4179  * Add subtree roots record into @subvol_root.
4180  *
4181  * @subvol_root:	tree root of the subvolume tree get swapped
4182  * @bg:			block group under balance
4183  * @subvol_parent/slot:	pointer to the subtree root in subvolume tree
4184  * @reloc_parent/slot:	pointer to the subtree root in reloc tree
4185  *			BOTH POINTERS ARE BEFORE TREE SWAP
4186  * @last_snapshot:	last snapshot generation of the subvolume tree
4187  */
4188 int btrfs_qgroup_add_swapped_blocks(struct btrfs_trans_handle *trans,
4189 		struct btrfs_root *subvol_root,
4190 		struct btrfs_block_group *bg,
4191 		struct extent_buffer *subvol_parent, int subvol_slot,
4192 		struct extent_buffer *reloc_parent, int reloc_slot,
4193 		u64 last_snapshot)
4194 {
4195 	struct btrfs_fs_info *fs_info = subvol_root->fs_info;
4196 	struct btrfs_qgroup_swapped_blocks *blocks = &subvol_root->swapped_blocks;
4197 	struct btrfs_qgroup_swapped_block *block;
4198 	struct rb_node **cur;
4199 	struct rb_node *parent = NULL;
4200 	int level = btrfs_header_level(subvol_parent) - 1;
4201 	int ret = 0;
4202 
4203 	if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
4204 		return 0;
4205 
4206 	if (btrfs_node_ptr_generation(subvol_parent, subvol_slot) >
4207 	    btrfs_node_ptr_generation(reloc_parent, reloc_slot)) {
4208 		btrfs_err_rl(fs_info,
4209 		"%s: bad parameter order, subvol_gen=%llu reloc_gen=%llu",
4210 			__func__,
4211 			btrfs_node_ptr_generation(subvol_parent, subvol_slot),
4212 			btrfs_node_ptr_generation(reloc_parent, reloc_slot));
4213 		return -EUCLEAN;
4214 	}
4215 
4216 	block = kmalloc(sizeof(*block), GFP_NOFS);
4217 	if (!block) {
4218 		ret = -ENOMEM;
4219 		goto out;
4220 	}
4221 
4222 	/*
4223 	 * @reloc_parent/slot is still before swap, while @block is going to
4224 	 * record the bytenr after swap, so we do the swap here.
4225 	 */
4226 	block->subvol_bytenr = btrfs_node_blockptr(reloc_parent, reloc_slot);
4227 	block->subvol_generation = btrfs_node_ptr_generation(reloc_parent,
4228 							     reloc_slot);
4229 	block->reloc_bytenr = btrfs_node_blockptr(subvol_parent, subvol_slot);
4230 	block->reloc_generation = btrfs_node_ptr_generation(subvol_parent,
4231 							    subvol_slot);
4232 	block->last_snapshot = last_snapshot;
4233 	block->level = level;
4234 
4235 	/*
4236 	 * If we have bg == NULL, we're called from btrfs_recover_relocation(),
4237 	 * no one else can modify tree blocks thus we qgroup will not change
4238 	 * no matter the value of trace_leaf.
4239 	 */
4240 	if (bg && bg->flags & BTRFS_BLOCK_GROUP_DATA)
4241 		block->trace_leaf = true;
4242 	else
4243 		block->trace_leaf = false;
4244 	btrfs_node_key_to_cpu(reloc_parent, &block->first_key, reloc_slot);
4245 
4246 	/* Insert @block into @blocks */
4247 	spin_lock(&blocks->lock);
4248 	cur = &blocks->blocks[level].rb_node;
4249 	while (*cur) {
4250 		struct btrfs_qgroup_swapped_block *entry;
4251 
4252 		parent = *cur;
4253 		entry = rb_entry(parent, struct btrfs_qgroup_swapped_block,
4254 				 node);
4255 
4256 		if (entry->subvol_bytenr < block->subvol_bytenr) {
4257 			cur = &(*cur)->rb_left;
4258 		} else if (entry->subvol_bytenr > block->subvol_bytenr) {
4259 			cur = &(*cur)->rb_right;
4260 		} else {
4261 			if (entry->subvol_generation !=
4262 					block->subvol_generation ||
4263 			    entry->reloc_bytenr != block->reloc_bytenr ||
4264 			    entry->reloc_generation !=
4265 					block->reloc_generation) {
4266 				/*
4267 				 * Duplicated but mismatch entry found.
4268 				 * Shouldn't happen.
4269 				 *
4270 				 * Marking qgroup inconsistent should be enough
4271 				 * for end users.
4272 				 */
4273 				WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
4274 				ret = -EEXIST;
4275 			}
4276 			kfree(block);
4277 			goto out_unlock;
4278 		}
4279 	}
4280 	rb_link_node(&block->node, parent, cur);
4281 	rb_insert_color(&block->node, &blocks->blocks[level]);
4282 	blocks->swapped = true;
4283 out_unlock:
4284 	spin_unlock(&blocks->lock);
4285 out:
4286 	if (ret < 0)
4287 		qgroup_mark_inconsistent(fs_info);
4288 	return ret;
4289 }
4290 
4291 /*
4292  * Check if the tree block is a subtree root, and if so do the needed
4293  * delayed subtree trace for qgroup.
4294  *
4295  * This is called during btrfs_cow_block().
4296  */
4297 int btrfs_qgroup_trace_subtree_after_cow(struct btrfs_trans_handle *trans,
4298 					 struct btrfs_root *root,
4299 					 struct extent_buffer *subvol_eb)
4300 {
4301 	struct btrfs_fs_info *fs_info = root->fs_info;
4302 	struct btrfs_qgroup_swapped_blocks *blocks = &root->swapped_blocks;
4303 	struct btrfs_qgroup_swapped_block *block;
4304 	struct extent_buffer *reloc_eb = NULL;
4305 	struct rb_node *node;
4306 	bool found = false;
4307 	bool swapped = false;
4308 	int level = btrfs_header_level(subvol_eb);
4309 	int ret = 0;
4310 	int i;
4311 
4312 	if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
4313 		return 0;
4314 	if (!is_fstree(root->root_key.objectid) || !root->reloc_root)
4315 		return 0;
4316 
4317 	spin_lock(&blocks->lock);
4318 	if (!blocks->swapped) {
4319 		spin_unlock(&blocks->lock);
4320 		return 0;
4321 	}
4322 	node = blocks->blocks[level].rb_node;
4323 
4324 	while (node) {
4325 		block = rb_entry(node, struct btrfs_qgroup_swapped_block, node);
4326 		if (block->subvol_bytenr < subvol_eb->start) {
4327 			node = node->rb_left;
4328 		} else if (block->subvol_bytenr > subvol_eb->start) {
4329 			node = node->rb_right;
4330 		} else {
4331 			found = true;
4332 			break;
4333 		}
4334 	}
4335 	if (!found) {
4336 		spin_unlock(&blocks->lock);
4337 		goto out;
4338 	}
4339 	/* Found one, remove it from @blocks first and update blocks->swapped */
4340 	rb_erase(&block->node, &blocks->blocks[level]);
4341 	for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
4342 		if (RB_EMPTY_ROOT(&blocks->blocks[i])) {
4343 			swapped = true;
4344 			break;
4345 		}
4346 	}
4347 	blocks->swapped = swapped;
4348 	spin_unlock(&blocks->lock);
4349 
4350 	/* Read out reloc subtree root */
4351 	reloc_eb = read_tree_block(fs_info, block->reloc_bytenr, 0,
4352 				   block->reloc_generation, block->level,
4353 				   &block->first_key);
4354 	if (IS_ERR(reloc_eb)) {
4355 		ret = PTR_ERR(reloc_eb);
4356 		reloc_eb = NULL;
4357 		goto free_out;
4358 	}
4359 	if (!extent_buffer_uptodate(reloc_eb)) {
4360 		ret = -EIO;
4361 		goto free_out;
4362 	}
4363 
4364 	ret = qgroup_trace_subtree_swap(trans, reloc_eb, subvol_eb,
4365 			block->last_snapshot, block->trace_leaf);
4366 free_out:
4367 	kfree(block);
4368 	free_extent_buffer(reloc_eb);
4369 out:
4370 	if (ret < 0) {
4371 		btrfs_err_rl(fs_info,
4372 			     "failed to account subtree at bytenr %llu: %d",
4373 			     subvol_eb->start, ret);
4374 		qgroup_mark_inconsistent(fs_info);
4375 	}
4376 	return ret;
4377 }
4378 
4379 void btrfs_qgroup_destroy_extent_records(struct btrfs_transaction *trans)
4380 {
4381 	struct btrfs_qgroup_extent_record *entry;
4382 	struct btrfs_qgroup_extent_record *next;
4383 	struct rb_root *root;
4384 
4385 	root = &trans->delayed_refs.dirty_extent_root;
4386 	rbtree_postorder_for_each_entry_safe(entry, next, root, node) {
4387 		ulist_free(entry->old_roots);
4388 		kfree(entry);
4389 	}
4390 }
4391