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 #include "fs.h"
28 #include "accessors.h"
29 #include "extent-tree.h"
30 #include "root-tree.h"
31 #include "tree-checker.h"
32
33 enum btrfs_qgroup_mode btrfs_qgroup_mode(const struct btrfs_fs_info *fs_info)
34 {
35 if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
36 return BTRFS_QGROUP_MODE_DISABLED;
37 if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_SIMPLE_MODE)
38 return BTRFS_QGROUP_MODE_SIMPLE;
39 return BTRFS_QGROUP_MODE_FULL;
40 }
41
42 bool btrfs_qgroup_enabled(const struct btrfs_fs_info *fs_info)
43 {
44 return btrfs_qgroup_mode(fs_info) != BTRFS_QGROUP_MODE_DISABLED;
45 }
46
47 bool btrfs_qgroup_full_accounting(const struct btrfs_fs_info *fs_info)
48 {
49 return btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_FULL;
50 }
51
52 /*
53 * Helpers to access qgroup reservation
54 *
55 * Callers should ensure the lock context and type are valid
56 */
57
58 static u64 qgroup_rsv_total(const struct btrfs_qgroup *qgroup)
59 {
60 u64 ret = 0;
61 int i;
62
63 for (i = 0; i < BTRFS_QGROUP_RSV_LAST; i++)
64 ret += qgroup->rsv.values[i];
65
66 return ret;
67 }
68
69 #ifdef CONFIG_BTRFS_DEBUG
70 static const char *qgroup_rsv_type_str(enum btrfs_qgroup_rsv_type type)
71 {
72 if (type == BTRFS_QGROUP_RSV_DATA)
73 return "data";
74 if (type == BTRFS_QGROUP_RSV_META_PERTRANS)
75 return "meta_pertrans";
76 if (type == BTRFS_QGROUP_RSV_META_PREALLOC)
77 return "meta_prealloc";
78 return NULL;
79 }
80 #endif
81
82 static void qgroup_rsv_add(struct btrfs_fs_info *fs_info,
83 struct btrfs_qgroup *qgroup, u64 num_bytes,
84 enum btrfs_qgroup_rsv_type type)
85 {
86 trace_qgroup_update_reserve(fs_info, qgroup, num_bytes, type);
87 qgroup->rsv.values[type] += num_bytes;
88 }
89
90 static void qgroup_rsv_release(struct btrfs_fs_info *fs_info,
91 struct btrfs_qgroup *qgroup, u64 num_bytes,
92 enum btrfs_qgroup_rsv_type type)
93 {
94 trace_qgroup_update_reserve(fs_info, qgroup, -(s64)num_bytes, type);
95 if (qgroup->rsv.values[type] >= num_bytes) {
96 qgroup->rsv.values[type] -= num_bytes;
97 return;
98 }
99 #ifdef CONFIG_BTRFS_DEBUG
100 WARN_RATELIMIT(1,
101 "qgroup %llu %s reserved space underflow, have %llu to free %llu",
102 qgroup->qgroupid, qgroup_rsv_type_str(type),
103 qgroup->rsv.values[type], num_bytes);
104 #endif
105 qgroup->rsv.values[type] = 0;
106 }
107
108 static void qgroup_rsv_add_by_qgroup(struct btrfs_fs_info *fs_info,
109 struct btrfs_qgroup *dest,
110 const struct btrfs_qgroup *src)
111 {
112 int i;
113
114 for (i = 0; i < BTRFS_QGROUP_RSV_LAST; i++)
115 qgroup_rsv_add(fs_info, dest, src->rsv.values[i], i);
116 }
117
118 static void qgroup_rsv_release_by_qgroup(struct btrfs_fs_info *fs_info,
119 struct btrfs_qgroup *dest,
120 const struct btrfs_qgroup *src)
121 {
122 int i;
123
124 for (i = 0; i < BTRFS_QGROUP_RSV_LAST; i++)
125 qgroup_rsv_release(fs_info, dest, src->rsv.values[i], i);
126 }
127
128 static void btrfs_qgroup_update_old_refcnt(struct btrfs_qgroup *qg, u64 seq,
129 int mod)
130 {
131 if (qg->old_refcnt < seq)
132 qg->old_refcnt = seq;
133 qg->old_refcnt += mod;
134 }
135
136 static void btrfs_qgroup_update_new_refcnt(struct btrfs_qgroup *qg, u64 seq,
137 int mod)
138 {
139 if (qg->new_refcnt < seq)
140 qg->new_refcnt = seq;
141 qg->new_refcnt += mod;
142 }
143
144 static inline u64 btrfs_qgroup_get_old_refcnt(const struct btrfs_qgroup *qg, u64 seq)
145 {
146 if (qg->old_refcnt < seq)
147 return 0;
148 return qg->old_refcnt - seq;
149 }
150
151 static inline u64 btrfs_qgroup_get_new_refcnt(const struct btrfs_qgroup *qg, u64 seq)
152 {
153 if (qg->new_refcnt < seq)
154 return 0;
155 return qg->new_refcnt - seq;
156 }
157
158 static int
159 qgroup_rescan_init(struct btrfs_fs_info *fs_info, u64 progress_objectid,
160 int init_flags);
161 static void qgroup_rescan_zero_tracking(struct btrfs_fs_info *fs_info);
162
163 /* must be called with qgroup_ioctl_lock held */
164 static struct btrfs_qgroup *find_qgroup_rb(const struct btrfs_fs_info *fs_info,
165 u64 qgroupid)
166 {
167 struct rb_node *n = fs_info->qgroup_tree.rb_node;
168 struct btrfs_qgroup *qgroup;
169
170 while (n) {
171 qgroup = rb_entry(n, struct btrfs_qgroup, node);
172 if (qgroup->qgroupid < qgroupid)
173 n = n->rb_left;
174 else if (qgroup->qgroupid > qgroupid)
175 n = n->rb_right;
176 else
177 return qgroup;
178 }
179 return NULL;
180 }
181
182 /*
183 * Add qgroup to the filesystem's qgroup tree.
184 *
185 * Must be called with qgroup_lock held and @prealloc preallocated.
186 *
187 * The control on the lifespan of @prealloc would be transferred to this
188 * function, thus caller should no longer touch @prealloc.
189 */
190 static struct btrfs_qgroup *add_qgroup_rb(struct btrfs_fs_info *fs_info,
191 struct btrfs_qgroup *prealloc,
192 u64 qgroupid)
193 {
194 struct rb_node **p = &fs_info->qgroup_tree.rb_node;
195 struct rb_node *parent = NULL;
196 struct btrfs_qgroup *qgroup;
197
198 /* Caller must have pre-allocated @prealloc. */
199 ASSERT(prealloc);
200
201 while (*p) {
202 parent = *p;
203 qgroup = rb_entry(parent, struct btrfs_qgroup, node);
204
205 if (qgroup->qgroupid < qgroupid) {
206 p = &(*p)->rb_left;
207 } else if (qgroup->qgroupid > qgroupid) {
208 p = &(*p)->rb_right;
209 } else {
210 kfree(prealloc);
211 return qgroup;
212 }
213 }
214
215 qgroup = prealloc;
216 qgroup->qgroupid = qgroupid;
217 INIT_LIST_HEAD(&qgroup->groups);
218 INIT_LIST_HEAD(&qgroup->members);
219 INIT_LIST_HEAD(&qgroup->dirty);
220 INIT_LIST_HEAD(&qgroup->iterator);
221 INIT_LIST_HEAD(&qgroup->nested_iterator);
222
223 rb_link_node(&qgroup->node, parent, p);
224 rb_insert_color(&qgroup->node, &fs_info->qgroup_tree);
225
226 return qgroup;
227 }
228
229 static void __del_qgroup_rb(struct btrfs_fs_info *fs_info,
230 struct btrfs_qgroup *qgroup)
231 {
232 struct btrfs_qgroup_list *list;
233
234 list_del(&qgroup->dirty);
235 while (!list_empty(&qgroup->groups)) {
236 list = list_first_entry(&qgroup->groups,
237 struct btrfs_qgroup_list, next_group);
238 list_del(&list->next_group);
239 list_del(&list->next_member);
240 kfree(list);
241 }
242
243 while (!list_empty(&qgroup->members)) {
244 list = list_first_entry(&qgroup->members,
245 struct btrfs_qgroup_list, next_member);
246 list_del(&list->next_group);
247 list_del(&list->next_member);
248 kfree(list);
249 }
250 }
251
252 /* must be called with qgroup_lock held */
253 static int del_qgroup_rb(struct btrfs_fs_info *fs_info, u64 qgroupid)
254 {
255 struct btrfs_qgroup *qgroup = find_qgroup_rb(fs_info, qgroupid);
256
257 if (!qgroup)
258 return -ENOENT;
259
260 rb_erase(&qgroup->node, &fs_info->qgroup_tree);
261 __del_qgroup_rb(fs_info, qgroup);
262 return 0;
263 }
264
265 /*
266 * Add relation specified by two qgroups.
267 *
268 * Must be called with qgroup_lock held, the ownership of @prealloc is
269 * transferred to this function and caller should not touch it anymore.
270 *
271 * Return: 0 on success
272 * -ENOENT if one of the qgroups is NULL
273 * <0 other errors
274 */
275 static int __add_relation_rb(struct btrfs_qgroup_list *prealloc,
276 struct btrfs_qgroup *member,
277 struct btrfs_qgroup *parent)
278 {
279 if (!member || !parent) {
280 kfree(prealloc);
281 return -ENOENT;
282 }
283
284 prealloc->group = parent;
285 prealloc->member = member;
286 list_add_tail(&prealloc->next_group, &member->groups);
287 list_add_tail(&prealloc->next_member, &parent->members);
288
289 return 0;
290 }
291
292 /*
293 * Add relation specified by two qgroup ids.
294 *
295 * Must be called with qgroup_lock held.
296 *
297 * Return: 0 on success
298 * -ENOENT if one of the ids does not exist
299 * <0 other errors
300 */
301 static int add_relation_rb(struct btrfs_fs_info *fs_info,
302 struct btrfs_qgroup_list *prealloc,
303 u64 memberid, u64 parentid)
304 {
305 struct btrfs_qgroup *member;
306 struct btrfs_qgroup *parent;
307
308 member = find_qgroup_rb(fs_info, memberid);
309 parent = find_qgroup_rb(fs_info, parentid);
310
311 return __add_relation_rb(prealloc, member, parent);
312 }
313
314 /* Must be called with qgroup_lock held */
315 static int del_relation_rb(struct btrfs_fs_info *fs_info,
316 u64 memberid, u64 parentid)
317 {
318 struct btrfs_qgroup *member;
319 struct btrfs_qgroup *parent;
320 struct btrfs_qgroup_list *list;
321
322 member = find_qgroup_rb(fs_info, memberid);
323 parent = find_qgroup_rb(fs_info, parentid);
324 if (!member || !parent)
325 return -ENOENT;
326
327 list_for_each_entry(list, &member->groups, next_group) {
328 if (list->group == parent) {
329 list_del(&list->next_group);
330 list_del(&list->next_member);
331 kfree(list);
332 return 0;
333 }
334 }
335 return -ENOENT;
336 }
337
338 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
339 int btrfs_verify_qgroup_counts(const struct btrfs_fs_info *fs_info, u64 qgroupid,
340 u64 rfer, u64 excl)
341 {
342 struct btrfs_qgroup *qgroup;
343
344 qgroup = find_qgroup_rb(fs_info, qgroupid);
345 if (!qgroup)
346 return -EINVAL;
347 if (qgroup->rfer != rfer || qgroup->excl != excl)
348 return -EINVAL;
349 return 0;
350 }
351 #endif
352
353 static void qgroup_mark_inconsistent(struct btrfs_fs_info *fs_info)
354 {
355 if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_SIMPLE)
356 return;
357 fs_info->qgroup_flags |= (BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT |
358 BTRFS_QGROUP_RUNTIME_FLAG_CANCEL_RESCAN |
359 BTRFS_QGROUP_RUNTIME_FLAG_NO_ACCOUNTING);
360 }
361
362 static void qgroup_read_enable_gen(struct btrfs_fs_info *fs_info,
363 struct extent_buffer *leaf, int slot,
364 struct btrfs_qgroup_status_item *ptr)
365 {
366 ASSERT(btrfs_fs_incompat(fs_info, SIMPLE_QUOTA));
367 ASSERT(btrfs_item_size(leaf, slot) >= sizeof(*ptr));
368 fs_info->qgroup_enable_gen = btrfs_qgroup_status_enable_gen(leaf, ptr);
369 }
370
371 /*
372 * The full config is read in one go, only called from open_ctree()
373 * It doesn't use any locking, as at this point we're still single-threaded
374 */
375 int btrfs_read_qgroup_config(struct btrfs_fs_info *fs_info)
376 {
377 struct btrfs_key key;
378 struct btrfs_key found_key;
379 struct btrfs_root *quota_root = fs_info->quota_root;
380 struct btrfs_path *path = NULL;
381 struct extent_buffer *l;
382 int slot;
383 int ret = 0;
384 u64 flags = 0;
385 u64 rescan_progress = 0;
386
387 if (!fs_info->quota_root)
388 return 0;
389
390 fs_info->qgroup_ulist = ulist_alloc(GFP_KERNEL);
391 if (!fs_info->qgroup_ulist) {
392 ret = -ENOMEM;
393 goto out;
394 }
395
396 path = btrfs_alloc_path();
397 if (!path) {
398 ret = -ENOMEM;
399 goto out;
400 }
401
402 ret = btrfs_sysfs_add_qgroups(fs_info);
403 if (ret < 0)
404 goto out;
405 /* default this to quota off, in case no status key is found */
406 fs_info->qgroup_flags = 0;
407
408 /*
409 * pass 1: read status, all qgroup infos and limits
410 */
411 key.objectid = 0;
412 key.type = 0;
413 key.offset = 0;
414 ret = btrfs_search_slot_for_read(quota_root, &key, path, 1, 1);
415 if (ret)
416 goto out;
417
418 while (1) {
419 struct btrfs_qgroup *qgroup;
420
421 slot = path->slots[0];
422 l = path->nodes[0];
423 btrfs_item_key_to_cpu(l, &found_key, slot);
424
425 if (found_key.type == BTRFS_QGROUP_STATUS_KEY) {
426 struct btrfs_qgroup_status_item *ptr;
427
428 ptr = btrfs_item_ptr(l, slot,
429 struct btrfs_qgroup_status_item);
430
431 if (btrfs_qgroup_status_version(l, ptr) !=
432 BTRFS_QGROUP_STATUS_VERSION) {
433 btrfs_err(fs_info,
434 "old qgroup version, quota disabled");
435 goto out;
436 }
437 fs_info->qgroup_flags = btrfs_qgroup_status_flags(l, ptr);
438 if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_SIMPLE_MODE) {
439 qgroup_read_enable_gen(fs_info, l, slot, ptr);
440 } else if (btrfs_qgroup_status_generation(l, ptr) != fs_info->generation) {
441 qgroup_mark_inconsistent(fs_info);
442 btrfs_err(fs_info,
443 "qgroup generation mismatch, marked as inconsistent");
444 }
445 rescan_progress = btrfs_qgroup_status_rescan(l, ptr);
446 goto next1;
447 }
448
449 if (found_key.type != BTRFS_QGROUP_INFO_KEY &&
450 found_key.type != BTRFS_QGROUP_LIMIT_KEY)
451 goto next1;
452
453 qgroup = find_qgroup_rb(fs_info, found_key.offset);
454 if ((qgroup && found_key.type == BTRFS_QGROUP_INFO_KEY) ||
455 (!qgroup && found_key.type == BTRFS_QGROUP_LIMIT_KEY)) {
456 btrfs_err(fs_info, "inconsistent qgroup config");
457 qgroup_mark_inconsistent(fs_info);
458 }
459 if (!qgroup) {
460 struct btrfs_qgroup *prealloc;
461 struct btrfs_root *tree_root = fs_info->tree_root;
462
463 prealloc = kzalloc(sizeof(*prealloc), GFP_KERNEL);
464 if (!prealloc) {
465 ret = -ENOMEM;
466 goto out;
467 }
468 qgroup = add_qgroup_rb(fs_info, prealloc, found_key.offset);
469 /*
470 * If a qgroup exists for a subvolume ID, it is possible
471 * that subvolume has been deleted, in which case
472 * re-using that ID would lead to incorrect accounting.
473 *
474 * Ensure that we skip any such subvol ids.
475 *
476 * We don't need to lock because this is only called
477 * during mount before we start doing things like creating
478 * subvolumes.
479 */
480 if (is_fstree(qgroup->qgroupid) &&
481 qgroup->qgroupid > tree_root->free_objectid)
482 /*
483 * Don't need to check against BTRFS_LAST_FREE_OBJECTID,
484 * as it will get checked on the next call to
485 * btrfs_get_free_objectid.
486 */
487 tree_root->free_objectid = qgroup->qgroupid + 1;
488 }
489 ret = btrfs_sysfs_add_one_qgroup(fs_info, qgroup);
490 if (ret < 0)
491 goto out;
492
493 switch (found_key.type) {
494 case BTRFS_QGROUP_INFO_KEY: {
495 struct btrfs_qgroup_info_item *ptr;
496
497 ptr = btrfs_item_ptr(l, slot,
498 struct btrfs_qgroup_info_item);
499 qgroup->rfer = btrfs_qgroup_info_rfer(l, ptr);
500 qgroup->rfer_cmpr = btrfs_qgroup_info_rfer_cmpr(l, ptr);
501 qgroup->excl = btrfs_qgroup_info_excl(l, ptr);
502 qgroup->excl_cmpr = btrfs_qgroup_info_excl_cmpr(l, ptr);
503 /* generation currently unused */
504 break;
505 }
506 case BTRFS_QGROUP_LIMIT_KEY: {
507 struct btrfs_qgroup_limit_item *ptr;
508
509 ptr = btrfs_item_ptr(l, slot,
510 struct btrfs_qgroup_limit_item);
511 qgroup->lim_flags = btrfs_qgroup_limit_flags(l, ptr);
512 qgroup->max_rfer = btrfs_qgroup_limit_max_rfer(l, ptr);
513 qgroup->max_excl = btrfs_qgroup_limit_max_excl(l, ptr);
514 qgroup->rsv_rfer = btrfs_qgroup_limit_rsv_rfer(l, ptr);
515 qgroup->rsv_excl = btrfs_qgroup_limit_rsv_excl(l, ptr);
516 break;
517 }
518 }
519 next1:
520 ret = btrfs_next_item(quota_root, path);
521 if (ret < 0)
522 goto out;
523 if (ret)
524 break;
525 }
526 btrfs_release_path(path);
527
528 /*
529 * pass 2: read all qgroup relations
530 */
531 key.objectid = 0;
532 key.type = BTRFS_QGROUP_RELATION_KEY;
533 key.offset = 0;
534 ret = btrfs_search_slot_for_read(quota_root, &key, path, 1, 0);
535 if (ret)
536 goto out;
537 while (1) {
538 struct btrfs_qgroup_list *list = NULL;
539
540 slot = path->slots[0];
541 l = path->nodes[0];
542 btrfs_item_key_to_cpu(l, &found_key, slot);
543
544 if (found_key.type != BTRFS_QGROUP_RELATION_KEY)
545 goto next2;
546
547 if (found_key.objectid > found_key.offset) {
548 /* parent <- member, not needed to build config */
549 /* FIXME should we omit the key completely? */
550 goto next2;
551 }
552
553 list = kzalloc(sizeof(*list), GFP_KERNEL);
554 if (!list) {
555 ret = -ENOMEM;
556 goto out;
557 }
558 ret = add_relation_rb(fs_info, list, found_key.objectid,
559 found_key.offset);
560 list = NULL;
561 if (ret == -ENOENT) {
562 btrfs_warn(fs_info,
563 "orphan qgroup relation 0x%llx->0x%llx",
564 found_key.objectid, found_key.offset);
565 ret = 0; /* ignore the error */
566 }
567 if (ret)
568 goto out;
569 next2:
570 ret = btrfs_next_item(quota_root, path);
571 if (ret < 0)
572 goto out;
573 if (ret)
574 break;
575 }
576 out:
577 btrfs_free_path(path);
578 fs_info->qgroup_flags |= flags;
579 if (ret >= 0) {
580 if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_ON)
581 set_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags);
582 if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN)
583 ret = qgroup_rescan_init(fs_info, rescan_progress, 0);
584 } else {
585 ulist_free(fs_info->qgroup_ulist);
586 fs_info->qgroup_ulist = NULL;
587 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
588 btrfs_sysfs_del_qgroups(fs_info);
589 }
590
591 return ret < 0 ? ret : 0;
592 }
593
594 /*
595 * Called in close_ctree() when quota is still enabled. This verifies we don't
596 * leak some reserved space.
597 *
598 * Return false if no reserved space is left.
599 * Return true if some reserved space is leaked.
600 */
601 bool btrfs_check_quota_leak(const struct btrfs_fs_info *fs_info)
602 {
603 struct rb_node *node;
604 bool ret = false;
605
606 if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_DISABLED)
607 return ret;
608 /*
609 * Since we're unmounting, there is no race and no need to grab qgroup
610 * lock. And here we don't go post-order to provide a more user
611 * friendly sorted result.
612 */
613 for (node = rb_first(&fs_info->qgroup_tree); node; node = rb_next(node)) {
614 struct btrfs_qgroup *qgroup;
615 int i;
616
617 qgroup = rb_entry(node, struct btrfs_qgroup, node);
618 for (i = 0; i < BTRFS_QGROUP_RSV_LAST; i++) {
619 if (qgroup->rsv.values[i]) {
620 ret = true;
621 btrfs_warn(fs_info,
622 "qgroup %hu/%llu has unreleased space, type %d rsv %llu",
623 btrfs_qgroup_level(qgroup->qgroupid),
624 btrfs_qgroup_subvolid(qgroup->qgroupid),
625 i, qgroup->rsv.values[i]);
626 }
627 }
628 }
629 return ret;
630 }
631
632 /*
633 * This is called from close_ctree() or open_ctree() or btrfs_quota_disable(),
634 * first two are in single-threaded paths.And for the third one, we have set
635 * quota_root to be null with qgroup_lock held before, so it is safe to clean
636 * up the in-memory structures without qgroup_lock held.
637 */
638 void btrfs_free_qgroup_config(struct btrfs_fs_info *fs_info)
639 {
640 struct rb_node *n;
641 struct btrfs_qgroup *qgroup;
642
643 while ((n = rb_first(&fs_info->qgroup_tree))) {
644 qgroup = rb_entry(n, struct btrfs_qgroup, node);
645 rb_erase(n, &fs_info->qgroup_tree);
646 __del_qgroup_rb(fs_info, qgroup);
647 btrfs_sysfs_del_one_qgroup(fs_info, qgroup);
648 kfree(qgroup);
649 }
650 /*
651 * We call btrfs_free_qgroup_config() when unmounting
652 * filesystem and disabling quota, so we set qgroup_ulist
653 * to be null here to avoid double free.
654 */
655 ulist_free(fs_info->qgroup_ulist);
656 fs_info->qgroup_ulist = NULL;
657 btrfs_sysfs_del_qgroups(fs_info);
658 }
659
660 static int add_qgroup_relation_item(struct btrfs_trans_handle *trans, u64 src,
661 u64 dst)
662 {
663 int ret;
664 struct btrfs_root *quota_root = trans->fs_info->quota_root;
665 struct btrfs_path *path;
666 struct btrfs_key key;
667
668 path = btrfs_alloc_path();
669 if (!path)
670 return -ENOMEM;
671
672 key.objectid = src;
673 key.type = BTRFS_QGROUP_RELATION_KEY;
674 key.offset = dst;
675
676 ret = btrfs_insert_empty_item(trans, quota_root, path, &key, 0);
677
678 btrfs_mark_buffer_dirty(trans, path->nodes[0]);
679
680 btrfs_free_path(path);
681 return ret;
682 }
683
684 static int del_qgroup_relation_item(struct btrfs_trans_handle *trans, u64 src,
685 u64 dst)
686 {
687 int ret;
688 struct btrfs_root *quota_root = trans->fs_info->quota_root;
689 struct btrfs_path *path;
690 struct btrfs_key key;
691
692 path = btrfs_alloc_path();
693 if (!path)
694 return -ENOMEM;
695
696 key.objectid = src;
697 key.type = BTRFS_QGROUP_RELATION_KEY;
698 key.offset = dst;
699
700 ret = btrfs_search_slot(trans, quota_root, &key, path, -1, 1);
701 if (ret < 0)
702 goto out;
703
704 if (ret > 0) {
705 ret = -ENOENT;
706 goto out;
707 }
708
709 ret = btrfs_del_item(trans, quota_root, path);
710 out:
711 btrfs_free_path(path);
712 return ret;
713 }
714
715 static int add_qgroup_item(struct btrfs_trans_handle *trans,
716 struct btrfs_root *quota_root, u64 qgroupid)
717 {
718 int ret;
719 struct btrfs_path *path;
720 struct btrfs_qgroup_info_item *qgroup_info;
721 struct btrfs_qgroup_limit_item *qgroup_limit;
722 struct extent_buffer *leaf;
723 struct btrfs_key key;
724
725 if (btrfs_is_testing(quota_root->fs_info))
726 return 0;
727
728 path = btrfs_alloc_path();
729 if (!path)
730 return -ENOMEM;
731
732 key.objectid = 0;
733 key.type = BTRFS_QGROUP_INFO_KEY;
734 key.offset = qgroupid;
735
736 /*
737 * Avoid a transaction abort by catching -EEXIST here. In that
738 * case, we proceed by re-initializing the existing structure
739 * on disk.
740 */
741
742 ret = btrfs_insert_empty_item(trans, quota_root, path, &key,
743 sizeof(*qgroup_info));
744 if (ret && ret != -EEXIST)
745 goto out;
746
747 leaf = path->nodes[0];
748 qgroup_info = btrfs_item_ptr(leaf, path->slots[0],
749 struct btrfs_qgroup_info_item);
750 btrfs_set_qgroup_info_generation(leaf, qgroup_info, trans->transid);
751 btrfs_set_qgroup_info_rfer(leaf, qgroup_info, 0);
752 btrfs_set_qgroup_info_rfer_cmpr(leaf, qgroup_info, 0);
753 btrfs_set_qgroup_info_excl(leaf, qgroup_info, 0);
754 btrfs_set_qgroup_info_excl_cmpr(leaf, qgroup_info, 0);
755
756 btrfs_mark_buffer_dirty(trans, leaf);
757
758 btrfs_release_path(path);
759
760 key.type = BTRFS_QGROUP_LIMIT_KEY;
761 ret = btrfs_insert_empty_item(trans, quota_root, path, &key,
762 sizeof(*qgroup_limit));
763 if (ret && ret != -EEXIST)
764 goto out;
765
766 leaf = path->nodes[0];
767 qgroup_limit = btrfs_item_ptr(leaf, path->slots[0],
768 struct btrfs_qgroup_limit_item);
769 btrfs_set_qgroup_limit_flags(leaf, qgroup_limit, 0);
770 btrfs_set_qgroup_limit_max_rfer(leaf, qgroup_limit, 0);
771 btrfs_set_qgroup_limit_max_excl(leaf, qgroup_limit, 0);
772 btrfs_set_qgroup_limit_rsv_rfer(leaf, qgroup_limit, 0);
773 btrfs_set_qgroup_limit_rsv_excl(leaf, qgroup_limit, 0);
774
775 btrfs_mark_buffer_dirty(trans, leaf);
776
777 ret = 0;
778 out:
779 btrfs_free_path(path);
780 return ret;
781 }
782
783 static int del_qgroup_item(struct btrfs_trans_handle *trans, u64 qgroupid)
784 {
785 int ret;
786 struct btrfs_root *quota_root = trans->fs_info->quota_root;
787 struct btrfs_path *path;
788 struct btrfs_key key;
789
790 path = btrfs_alloc_path();
791 if (!path)
792 return -ENOMEM;
793
794 key.objectid = 0;
795 key.type = BTRFS_QGROUP_INFO_KEY;
796 key.offset = qgroupid;
797 ret = btrfs_search_slot(trans, quota_root, &key, path, -1, 1);
798 if (ret < 0)
799 goto out;
800
801 if (ret > 0) {
802 ret = -ENOENT;
803 goto out;
804 }
805
806 ret = btrfs_del_item(trans, quota_root, path);
807 if (ret)
808 goto out;
809
810 btrfs_release_path(path);
811
812 key.type = BTRFS_QGROUP_LIMIT_KEY;
813 ret = btrfs_search_slot(trans, quota_root, &key, path, -1, 1);
814 if (ret < 0)
815 goto out;
816
817 if (ret > 0) {
818 ret = -ENOENT;
819 goto out;
820 }
821
822 ret = btrfs_del_item(trans, quota_root, path);
823
824 out:
825 btrfs_free_path(path);
826 return ret;
827 }
828
829 static int update_qgroup_limit_item(struct btrfs_trans_handle *trans,
830 struct btrfs_qgroup *qgroup)
831 {
832 struct btrfs_root *quota_root = trans->fs_info->quota_root;
833 struct btrfs_path *path;
834 struct btrfs_key key;
835 struct extent_buffer *l;
836 struct btrfs_qgroup_limit_item *qgroup_limit;
837 int ret;
838 int slot;
839
840 key.objectid = 0;
841 key.type = BTRFS_QGROUP_LIMIT_KEY;
842 key.offset = qgroup->qgroupid;
843
844 path = btrfs_alloc_path();
845 if (!path)
846 return -ENOMEM;
847
848 ret = btrfs_search_slot(trans, quota_root, &key, path, 0, 1);
849 if (ret > 0)
850 ret = -ENOENT;
851
852 if (ret)
853 goto out;
854
855 l = path->nodes[0];
856 slot = path->slots[0];
857 qgroup_limit = btrfs_item_ptr(l, slot, struct btrfs_qgroup_limit_item);
858 btrfs_set_qgroup_limit_flags(l, qgroup_limit, qgroup->lim_flags);
859 btrfs_set_qgroup_limit_max_rfer(l, qgroup_limit, qgroup->max_rfer);
860 btrfs_set_qgroup_limit_max_excl(l, qgroup_limit, qgroup->max_excl);
861 btrfs_set_qgroup_limit_rsv_rfer(l, qgroup_limit, qgroup->rsv_rfer);
862 btrfs_set_qgroup_limit_rsv_excl(l, qgroup_limit, qgroup->rsv_excl);
863
864 btrfs_mark_buffer_dirty(trans, l);
865
866 out:
867 btrfs_free_path(path);
868 return ret;
869 }
870
871 static int update_qgroup_info_item(struct btrfs_trans_handle *trans,
872 struct btrfs_qgroup *qgroup)
873 {
874 struct btrfs_fs_info *fs_info = trans->fs_info;
875 struct btrfs_root *quota_root = fs_info->quota_root;
876 struct btrfs_path *path;
877 struct btrfs_key key;
878 struct extent_buffer *l;
879 struct btrfs_qgroup_info_item *qgroup_info;
880 int ret;
881 int slot;
882
883 if (btrfs_is_testing(fs_info))
884 return 0;
885
886 key.objectid = 0;
887 key.type = BTRFS_QGROUP_INFO_KEY;
888 key.offset = qgroup->qgroupid;
889
890 path = btrfs_alloc_path();
891 if (!path)
892 return -ENOMEM;
893
894 ret = btrfs_search_slot(trans, quota_root, &key, path, 0, 1);
895 if (ret > 0)
896 ret = -ENOENT;
897
898 if (ret)
899 goto out;
900
901 l = path->nodes[0];
902 slot = path->slots[0];
903 qgroup_info = btrfs_item_ptr(l, slot, struct btrfs_qgroup_info_item);
904 btrfs_set_qgroup_info_generation(l, qgroup_info, trans->transid);
905 btrfs_set_qgroup_info_rfer(l, qgroup_info, qgroup->rfer);
906 btrfs_set_qgroup_info_rfer_cmpr(l, qgroup_info, qgroup->rfer_cmpr);
907 btrfs_set_qgroup_info_excl(l, qgroup_info, qgroup->excl);
908 btrfs_set_qgroup_info_excl_cmpr(l, qgroup_info, qgroup->excl_cmpr);
909
910 btrfs_mark_buffer_dirty(trans, l);
911
912 out:
913 btrfs_free_path(path);
914 return ret;
915 }
916
917 static int update_qgroup_status_item(struct btrfs_trans_handle *trans)
918 {
919 struct btrfs_fs_info *fs_info = trans->fs_info;
920 struct btrfs_root *quota_root = fs_info->quota_root;
921 struct btrfs_path *path;
922 struct btrfs_key key;
923 struct extent_buffer *l;
924 struct btrfs_qgroup_status_item *ptr;
925 int ret;
926 int slot;
927
928 key.objectid = 0;
929 key.type = BTRFS_QGROUP_STATUS_KEY;
930 key.offset = 0;
931
932 path = btrfs_alloc_path();
933 if (!path)
934 return -ENOMEM;
935
936 ret = btrfs_search_slot(trans, quota_root, &key, path, 0, 1);
937 if (ret > 0)
938 ret = -ENOENT;
939
940 if (ret)
941 goto out;
942
943 l = path->nodes[0];
944 slot = path->slots[0];
945 ptr = btrfs_item_ptr(l, slot, struct btrfs_qgroup_status_item);
946 btrfs_set_qgroup_status_flags(l, ptr, fs_info->qgroup_flags &
947 BTRFS_QGROUP_STATUS_FLAGS_MASK);
948 btrfs_set_qgroup_status_generation(l, ptr, trans->transid);
949 btrfs_set_qgroup_status_rescan(l, ptr,
950 fs_info->qgroup_rescan_progress.objectid);
951
952 btrfs_mark_buffer_dirty(trans, l);
953
954 out:
955 btrfs_free_path(path);
956 return ret;
957 }
958
959 /*
960 * called with qgroup_lock held
961 */
962 static int btrfs_clean_quota_tree(struct btrfs_trans_handle *trans,
963 struct btrfs_root *root)
964 {
965 struct btrfs_path *path;
966 struct btrfs_key key;
967 struct extent_buffer *leaf = NULL;
968 int ret;
969 int nr = 0;
970
971 path = btrfs_alloc_path();
972 if (!path)
973 return -ENOMEM;
974
975 key.objectid = 0;
976 key.offset = 0;
977 key.type = 0;
978
979 while (1) {
980 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
981 if (ret < 0)
982 goto out;
983 leaf = path->nodes[0];
984 nr = btrfs_header_nritems(leaf);
985 if (!nr)
986 break;
987 /*
988 * delete the leaf one by one
989 * since the whole tree is going
990 * to be deleted.
991 */
992 path->slots[0] = 0;
993 ret = btrfs_del_items(trans, root, path, 0, nr);
994 if (ret)
995 goto out;
996
997 btrfs_release_path(path);
998 }
999 ret = 0;
1000 out:
1001 btrfs_free_path(path);
1002 return ret;
1003 }
1004
1005 int btrfs_quota_enable(struct btrfs_fs_info *fs_info,
1006 struct btrfs_ioctl_quota_ctl_args *quota_ctl_args)
1007 {
1008 struct btrfs_root *quota_root;
1009 struct btrfs_root *tree_root = fs_info->tree_root;
1010 struct btrfs_path *path = NULL;
1011 struct btrfs_qgroup_status_item *ptr;
1012 struct extent_buffer *leaf;
1013 struct btrfs_key key;
1014 struct btrfs_key found_key;
1015 struct btrfs_qgroup *qgroup = NULL;
1016 struct btrfs_qgroup *prealloc = NULL;
1017 struct btrfs_trans_handle *trans = NULL;
1018 struct ulist *ulist = NULL;
1019 const bool simple = (quota_ctl_args->cmd == BTRFS_QUOTA_CTL_ENABLE_SIMPLE_QUOTA);
1020 int ret = 0;
1021 int slot;
1022
1023 /*
1024 * We need to have subvol_sem write locked, to prevent races between
1025 * concurrent tasks trying to enable quotas, because we will unlock
1026 * and relock qgroup_ioctl_lock before setting fs_info->quota_root
1027 * and before setting BTRFS_FS_QUOTA_ENABLED.
1028 */
1029 lockdep_assert_held_write(&fs_info->subvol_sem);
1030
1031 if (btrfs_fs_incompat(fs_info, EXTENT_TREE_V2)) {
1032 btrfs_err(fs_info,
1033 "qgroups are currently unsupported in extent tree v2");
1034 return -EINVAL;
1035 }
1036
1037 mutex_lock(&fs_info->qgroup_ioctl_lock);
1038 if (fs_info->quota_root)
1039 goto out;
1040
1041 ulist = ulist_alloc(GFP_KERNEL);
1042 if (!ulist) {
1043 ret = -ENOMEM;
1044 goto out;
1045 }
1046
1047 ret = btrfs_sysfs_add_qgroups(fs_info);
1048 if (ret < 0)
1049 goto out;
1050
1051 /*
1052 * Unlock qgroup_ioctl_lock before starting the transaction. This is to
1053 * avoid lock acquisition inversion problems (reported by lockdep) between
1054 * qgroup_ioctl_lock and the vfs freeze semaphores, acquired when we
1055 * start a transaction.
1056 * After we started the transaction lock qgroup_ioctl_lock again and
1057 * check if someone else created the quota root in the meanwhile. If so,
1058 * just return success and release the transaction handle.
1059 *
1060 * Also we don't need to worry about someone else calling
1061 * btrfs_sysfs_add_qgroups() after we unlock and getting an error because
1062 * that function returns 0 (success) when the sysfs entries already exist.
1063 */
1064 mutex_unlock(&fs_info->qgroup_ioctl_lock);
1065
1066 /*
1067 * 1 for quota root item
1068 * 1 for BTRFS_QGROUP_STATUS item
1069 *
1070 * Yet we also need 2*n items for a QGROUP_INFO/QGROUP_LIMIT items
1071 * per subvolume. However those are not currently reserved since it
1072 * would be a lot of overkill.
1073 */
1074 trans = btrfs_start_transaction(tree_root, 2);
1075
1076 mutex_lock(&fs_info->qgroup_ioctl_lock);
1077 if (IS_ERR(trans)) {
1078 ret = PTR_ERR(trans);
1079 trans = NULL;
1080 goto out;
1081 }
1082
1083 if (fs_info->quota_root)
1084 goto out;
1085
1086 fs_info->qgroup_ulist = ulist;
1087 ulist = NULL;
1088
1089 /*
1090 * initially create the quota tree
1091 */
1092 quota_root = btrfs_create_tree(trans, BTRFS_QUOTA_TREE_OBJECTID);
1093 if (IS_ERR(quota_root)) {
1094 ret = PTR_ERR(quota_root);
1095 btrfs_abort_transaction(trans, ret);
1096 goto out;
1097 }
1098
1099 path = btrfs_alloc_path();
1100 if (!path) {
1101 ret = -ENOMEM;
1102 btrfs_abort_transaction(trans, ret);
1103 goto out_free_root;
1104 }
1105
1106 key.objectid = 0;
1107 key.type = BTRFS_QGROUP_STATUS_KEY;
1108 key.offset = 0;
1109
1110 ret = btrfs_insert_empty_item(trans, quota_root, path, &key,
1111 sizeof(*ptr));
1112 if (ret) {
1113 btrfs_abort_transaction(trans, ret);
1114 goto out_free_path;
1115 }
1116
1117 leaf = path->nodes[0];
1118 ptr = btrfs_item_ptr(leaf, path->slots[0],
1119 struct btrfs_qgroup_status_item);
1120 btrfs_set_qgroup_status_generation(leaf, ptr, trans->transid);
1121 btrfs_set_qgroup_status_version(leaf, ptr, BTRFS_QGROUP_STATUS_VERSION);
1122 fs_info->qgroup_flags = BTRFS_QGROUP_STATUS_FLAG_ON;
1123 if (simple) {
1124 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_SIMPLE_MODE;
1125 btrfs_set_qgroup_status_enable_gen(leaf, ptr, trans->transid);
1126 } else {
1127 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
1128 }
1129 btrfs_set_qgroup_status_flags(leaf, ptr, fs_info->qgroup_flags &
1130 BTRFS_QGROUP_STATUS_FLAGS_MASK);
1131 btrfs_set_qgroup_status_rescan(leaf, ptr, 0);
1132
1133 btrfs_mark_buffer_dirty(trans, leaf);
1134
1135 key.objectid = 0;
1136 key.type = BTRFS_ROOT_REF_KEY;
1137 key.offset = 0;
1138
1139 btrfs_release_path(path);
1140 ret = btrfs_search_slot_for_read(tree_root, &key, path, 1, 0);
1141 if (ret > 0)
1142 goto out_add_root;
1143 if (ret < 0) {
1144 btrfs_abort_transaction(trans, ret);
1145 goto out_free_path;
1146 }
1147
1148 while (1) {
1149 slot = path->slots[0];
1150 leaf = path->nodes[0];
1151 btrfs_item_key_to_cpu(leaf, &found_key, slot);
1152
1153 if (found_key.type == BTRFS_ROOT_REF_KEY) {
1154
1155 /* Release locks on tree_root before we access quota_root */
1156 btrfs_release_path(path);
1157
1158 /* We should not have a stray @prealloc pointer. */
1159 ASSERT(prealloc == NULL);
1160 prealloc = kzalloc(sizeof(*prealloc), GFP_NOFS);
1161 if (!prealloc) {
1162 ret = -ENOMEM;
1163 btrfs_abort_transaction(trans, ret);
1164 goto out_free_path;
1165 }
1166
1167 ret = add_qgroup_item(trans, quota_root,
1168 found_key.offset);
1169 if (ret) {
1170 btrfs_abort_transaction(trans, ret);
1171 goto out_free_path;
1172 }
1173
1174 qgroup = add_qgroup_rb(fs_info, prealloc, found_key.offset);
1175 prealloc = NULL;
1176 if (IS_ERR(qgroup)) {
1177 ret = PTR_ERR(qgroup);
1178 btrfs_abort_transaction(trans, ret);
1179 goto out_free_path;
1180 }
1181 ret = btrfs_sysfs_add_one_qgroup(fs_info, qgroup);
1182 if (ret < 0) {
1183 btrfs_abort_transaction(trans, ret);
1184 goto out_free_path;
1185 }
1186 ret = btrfs_search_slot_for_read(tree_root, &found_key,
1187 path, 1, 0);
1188 if (ret < 0) {
1189 btrfs_abort_transaction(trans, ret);
1190 goto out_free_path;
1191 }
1192 if (ret > 0) {
1193 /*
1194 * Shouldn't happen, but in case it does we
1195 * don't need to do the btrfs_next_item, just
1196 * continue.
1197 */
1198 continue;
1199 }
1200 }
1201 ret = btrfs_next_item(tree_root, path);
1202 if (ret < 0) {
1203 btrfs_abort_transaction(trans, ret);
1204 goto out_free_path;
1205 }
1206 if (ret)
1207 break;
1208 }
1209
1210 out_add_root:
1211 btrfs_release_path(path);
1212 ret = add_qgroup_item(trans, quota_root, BTRFS_FS_TREE_OBJECTID);
1213 if (ret) {
1214 btrfs_abort_transaction(trans, ret);
1215 goto out_free_path;
1216 }
1217
1218 ASSERT(prealloc == NULL);
1219 prealloc = kzalloc(sizeof(*prealloc), GFP_NOFS);
1220 if (!prealloc) {
1221 ret = -ENOMEM;
1222 goto out_free_path;
1223 }
1224 qgroup = add_qgroup_rb(fs_info, prealloc, BTRFS_FS_TREE_OBJECTID);
1225 prealloc = NULL;
1226 ret = btrfs_sysfs_add_one_qgroup(fs_info, qgroup);
1227 if (ret < 0) {
1228 btrfs_abort_transaction(trans, ret);
1229 goto out_free_path;
1230 }
1231
1232 fs_info->qgroup_enable_gen = trans->transid;
1233
1234 mutex_unlock(&fs_info->qgroup_ioctl_lock);
1235 /*
1236 * Commit the transaction while not holding qgroup_ioctl_lock, to avoid
1237 * a deadlock with tasks concurrently doing other qgroup operations, such
1238 * adding/removing qgroups or adding/deleting qgroup relations for example,
1239 * because all qgroup operations first start or join a transaction and then
1240 * lock the qgroup_ioctl_lock mutex.
1241 * We are safe from a concurrent task trying to enable quotas, by calling
1242 * this function, since we are serialized by fs_info->subvol_sem.
1243 */
1244 ret = btrfs_commit_transaction(trans);
1245 trans = NULL;
1246 mutex_lock(&fs_info->qgroup_ioctl_lock);
1247 if (ret)
1248 goto out_free_path;
1249
1250 /*
1251 * Set quota enabled flag after committing the transaction, to avoid
1252 * deadlocks on fs_info->qgroup_ioctl_lock with concurrent snapshot
1253 * creation.
1254 */
1255 spin_lock(&fs_info->qgroup_lock);
1256 fs_info->quota_root = quota_root;
1257 set_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags);
1258 if (simple)
1259 btrfs_set_fs_incompat(fs_info, SIMPLE_QUOTA);
1260 spin_unlock(&fs_info->qgroup_lock);
1261
1262 /* Skip rescan for simple qgroups. */
1263 if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_SIMPLE)
1264 goto out_free_path;
1265
1266 ret = qgroup_rescan_init(fs_info, 0, 1);
1267 if (!ret) {
1268 qgroup_rescan_zero_tracking(fs_info);
1269 fs_info->qgroup_rescan_running = true;
1270 btrfs_queue_work(fs_info->qgroup_rescan_workers,
1271 &fs_info->qgroup_rescan_work);
1272 } else {
1273 /*
1274 * We have set both BTRFS_FS_QUOTA_ENABLED and
1275 * BTRFS_QGROUP_STATUS_FLAG_ON, so we can only fail with
1276 * -EINPROGRESS. That can happen because someone started the
1277 * rescan worker by calling quota rescan ioctl before we
1278 * attempted to initialize the rescan worker. Failure due to
1279 * quotas disabled in the meanwhile is not possible, because
1280 * we are holding a write lock on fs_info->subvol_sem, which
1281 * is also acquired when disabling quotas.
1282 * Ignore such error, and any other error would need to undo
1283 * everything we did in the transaction we just committed.
1284 */
1285 ASSERT(ret == -EINPROGRESS);
1286 ret = 0;
1287 }
1288
1289 out_free_path:
1290 btrfs_free_path(path);
1291 out_free_root:
1292 if (ret)
1293 btrfs_put_root(quota_root);
1294 out:
1295 if (ret) {
1296 ulist_free(fs_info->qgroup_ulist);
1297 fs_info->qgroup_ulist = NULL;
1298 btrfs_sysfs_del_qgroups(fs_info);
1299 }
1300 mutex_unlock(&fs_info->qgroup_ioctl_lock);
1301 if (ret && trans)
1302 btrfs_end_transaction(trans);
1303 else if (trans)
1304 ret = btrfs_end_transaction(trans);
1305 ulist_free(ulist);
1306 kfree(prealloc);
1307 return ret;
1308 }
1309
1310 /*
1311 * It is possible to have outstanding ordered extents which reserved bytes
1312 * before we disabled. We need to fully flush delalloc, ordered extents, and a
1313 * commit to ensure that we don't leak such reservations, only to have them
1314 * come back if we re-enable.
1315 *
1316 * - enable simple quotas
1317 * - reserve space
1318 * - release it, store rsv_bytes in OE
1319 * - disable quotas
1320 * - enable simple quotas (qgroup rsv are all 0)
1321 * - OE finishes
1322 * - run delayed refs
1323 * - free rsv_bytes, resulting in miscounting or even underflow
1324 */
1325 static int flush_reservations(struct btrfs_fs_info *fs_info)
1326 {
1327 int ret;
1328
1329 ret = btrfs_start_delalloc_roots(fs_info, LONG_MAX, false);
1330 if (ret)
1331 return ret;
1332 btrfs_wait_ordered_roots(fs_info, U64_MAX, NULL);
1333
1334 return btrfs_commit_current_transaction(fs_info->tree_root);
1335 }
1336
1337 int btrfs_quota_disable(struct btrfs_fs_info *fs_info)
1338 {
1339 struct btrfs_root *quota_root = NULL;
1340 struct btrfs_trans_handle *trans = NULL;
1341 int ret = 0;
1342
1343 /*
1344 * We need to have subvol_sem write locked to prevent races with
1345 * snapshot creation.
1346 */
1347 lockdep_assert_held_write(&fs_info->subvol_sem);
1348
1349 /*
1350 * Relocation will mess with backrefs, so make sure we have the
1351 * cleaner_mutex held to protect us from relocate.
1352 */
1353 lockdep_assert_held(&fs_info->cleaner_mutex);
1354
1355 mutex_lock(&fs_info->qgroup_ioctl_lock);
1356 if (!fs_info->quota_root)
1357 goto out;
1358
1359 /*
1360 * Unlock the qgroup_ioctl_lock mutex before waiting for the rescan worker to
1361 * complete. Otherwise we can deadlock because btrfs_remove_qgroup() needs
1362 * to lock that mutex while holding a transaction handle and the rescan
1363 * worker needs to commit a transaction.
1364 */
1365 mutex_unlock(&fs_info->qgroup_ioctl_lock);
1366
1367 /*
1368 * Request qgroup rescan worker to complete and wait for it. This wait
1369 * must be done before transaction start for quota disable since it may
1370 * deadlock with transaction by the qgroup rescan worker.
1371 */
1372 clear_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags);
1373 btrfs_qgroup_wait_for_completion(fs_info, false);
1374
1375 /*
1376 * We have nothing held here and no trans handle, just return the error
1377 * if there is one.
1378 */
1379 ret = flush_reservations(fs_info);
1380 if (ret)
1381 return ret;
1382
1383 /*
1384 * 1 For the root item
1385 *
1386 * We should also reserve enough items for the quota tree deletion in
1387 * btrfs_clean_quota_tree but this is not done.
1388 *
1389 * Also, we must always start a transaction without holding the mutex
1390 * qgroup_ioctl_lock, see btrfs_quota_enable().
1391 */
1392 trans = btrfs_start_transaction(fs_info->tree_root, 1);
1393
1394 mutex_lock(&fs_info->qgroup_ioctl_lock);
1395 if (IS_ERR(trans)) {
1396 ret = PTR_ERR(trans);
1397 trans = NULL;
1398 set_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags);
1399 goto out;
1400 }
1401
1402 if (!fs_info->quota_root)
1403 goto out;
1404
1405 spin_lock(&fs_info->qgroup_lock);
1406 quota_root = fs_info->quota_root;
1407 fs_info->quota_root = NULL;
1408 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_ON;
1409 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_SIMPLE_MODE;
1410 fs_info->qgroup_drop_subtree_thres = BTRFS_MAX_LEVEL;
1411 spin_unlock(&fs_info->qgroup_lock);
1412
1413 btrfs_free_qgroup_config(fs_info);
1414
1415 ret = btrfs_clean_quota_tree(trans, quota_root);
1416 if (ret) {
1417 btrfs_abort_transaction(trans, ret);
1418 goto out;
1419 }
1420
1421 ret = btrfs_del_root(trans, "a_root->root_key);
1422 if (ret) {
1423 btrfs_abort_transaction(trans, ret);
1424 goto out;
1425 }
1426
1427 spin_lock(&fs_info->trans_lock);
1428 list_del("a_root->dirty_list);
1429 spin_unlock(&fs_info->trans_lock);
1430
1431 btrfs_tree_lock(quota_root->node);
1432 btrfs_clear_buffer_dirty(trans, quota_root->node);
1433 btrfs_tree_unlock(quota_root->node);
1434 ret = btrfs_free_tree_block(trans, btrfs_root_id(quota_root),
1435 quota_root->node, 0, 1);
1436
1437 if (ret < 0)
1438 btrfs_abort_transaction(trans, ret);
1439
1440 out:
1441 btrfs_put_root(quota_root);
1442 mutex_unlock(&fs_info->qgroup_ioctl_lock);
1443 if (ret && trans)
1444 btrfs_end_transaction(trans);
1445 else if (trans)
1446 ret = btrfs_commit_transaction(trans);
1447 return ret;
1448 }
1449
1450 static void qgroup_dirty(struct btrfs_fs_info *fs_info,
1451 struct btrfs_qgroup *qgroup)
1452 {
1453 if (list_empty(&qgroup->dirty))
1454 list_add(&qgroup->dirty, &fs_info->dirty_qgroups);
1455 }
1456
1457 static void qgroup_iterator_add(struct list_head *head, struct btrfs_qgroup *qgroup)
1458 {
1459 if (!list_empty(&qgroup->iterator))
1460 return;
1461
1462 list_add_tail(&qgroup->iterator, head);
1463 }
1464
1465 static void qgroup_iterator_clean(struct list_head *head)
1466 {
1467 while (!list_empty(head)) {
1468 struct btrfs_qgroup *qgroup;
1469
1470 qgroup = list_first_entry(head, struct btrfs_qgroup, iterator);
1471 list_del_init(&qgroup->iterator);
1472 }
1473 }
1474
1475 /*
1476 * The easy accounting, we're updating qgroup relationship whose child qgroup
1477 * only has exclusive extents.
1478 *
1479 * In this case, all exclusive extents will also be exclusive for parent, so
1480 * excl/rfer just get added/removed.
1481 *
1482 * So is qgroup reservation space, which should also be added/removed to
1483 * parent.
1484 * Or when child tries to release reservation space, parent will underflow its
1485 * reservation (for relationship adding case).
1486 *
1487 * Caller should hold fs_info->qgroup_lock.
1488 */
1489 static int __qgroup_excl_accounting(struct btrfs_fs_info *fs_info, u64 ref_root,
1490 struct btrfs_qgroup *src, int sign)
1491 {
1492 struct btrfs_qgroup *qgroup;
1493 struct btrfs_qgroup *cur;
1494 LIST_HEAD(qgroup_list);
1495 u64 num_bytes = src->excl;
1496 int ret = 0;
1497
1498 qgroup = find_qgroup_rb(fs_info, ref_root);
1499 if (!qgroup)
1500 goto out;
1501
1502 qgroup_iterator_add(&qgroup_list, qgroup);
1503 list_for_each_entry(cur, &qgroup_list, iterator) {
1504 struct btrfs_qgroup_list *glist;
1505
1506 qgroup->rfer += sign * num_bytes;
1507 qgroup->rfer_cmpr += sign * num_bytes;
1508
1509 WARN_ON(sign < 0 && qgroup->excl < num_bytes);
1510 qgroup->excl += sign * num_bytes;
1511 qgroup->excl_cmpr += sign * num_bytes;
1512
1513 if (sign > 0)
1514 qgroup_rsv_add_by_qgroup(fs_info, qgroup, src);
1515 else
1516 qgroup_rsv_release_by_qgroup(fs_info, qgroup, src);
1517 qgroup_dirty(fs_info, qgroup);
1518
1519 /* Append parent qgroups to @qgroup_list. */
1520 list_for_each_entry(glist, &qgroup->groups, next_group)
1521 qgroup_iterator_add(&qgroup_list, glist->group);
1522 }
1523 ret = 0;
1524 out:
1525 qgroup_iterator_clean(&qgroup_list);
1526 return ret;
1527 }
1528
1529
1530 /*
1531 * Quick path for updating qgroup with only excl refs.
1532 *
1533 * In that case, just update all parent will be enough.
1534 * Or we needs to do a full rescan.
1535 * Caller should also hold fs_info->qgroup_lock.
1536 *
1537 * Return 0 for quick update, return >0 for need to full rescan
1538 * and mark INCONSISTENT flag.
1539 * Return < 0 for other error.
1540 */
1541 static int quick_update_accounting(struct btrfs_fs_info *fs_info,
1542 u64 src, u64 dst, int sign)
1543 {
1544 struct btrfs_qgroup *qgroup;
1545 int ret = 1;
1546
1547 qgroup = find_qgroup_rb(fs_info, src);
1548 if (!qgroup)
1549 goto out;
1550 if (qgroup->excl == qgroup->rfer) {
1551 ret = __qgroup_excl_accounting(fs_info, dst, qgroup, sign);
1552 if (ret < 0)
1553 goto out;
1554 ret = 0;
1555 }
1556 out:
1557 if (ret)
1558 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
1559 return ret;
1560 }
1561
1562 /*
1563 * Add relation between @src and @dst qgroup. The @prealloc is allocated by the
1564 * callers and transferred here (either used or freed on error).
1565 */
1566 int btrfs_add_qgroup_relation(struct btrfs_trans_handle *trans, u64 src, u64 dst,
1567 struct btrfs_qgroup_list *prealloc)
1568 {
1569 struct btrfs_fs_info *fs_info = trans->fs_info;
1570 struct btrfs_qgroup *parent;
1571 struct btrfs_qgroup *member;
1572 struct btrfs_qgroup_list *list;
1573 int ret = 0;
1574
1575 ASSERT(prealloc);
1576
1577 /* Check the level of src and dst first */
1578 if (btrfs_qgroup_level(src) >= btrfs_qgroup_level(dst))
1579 return -EINVAL;
1580
1581 mutex_lock(&fs_info->qgroup_ioctl_lock);
1582 if (!fs_info->quota_root) {
1583 ret = -ENOTCONN;
1584 goto out;
1585 }
1586 member = find_qgroup_rb(fs_info, src);
1587 parent = find_qgroup_rb(fs_info, dst);
1588 if (!member || !parent) {
1589 ret = -EINVAL;
1590 goto out;
1591 }
1592
1593 /* check if such qgroup relation exist firstly */
1594 list_for_each_entry(list, &member->groups, next_group) {
1595 if (list->group == parent) {
1596 ret = -EEXIST;
1597 goto out;
1598 }
1599 }
1600
1601 ret = add_qgroup_relation_item(trans, src, dst);
1602 if (ret)
1603 goto out;
1604
1605 ret = add_qgroup_relation_item(trans, dst, src);
1606 if (ret) {
1607 del_qgroup_relation_item(trans, src, dst);
1608 goto out;
1609 }
1610
1611 spin_lock(&fs_info->qgroup_lock);
1612 ret = __add_relation_rb(prealloc, member, parent);
1613 prealloc = NULL;
1614 if (ret < 0) {
1615 spin_unlock(&fs_info->qgroup_lock);
1616 goto out;
1617 }
1618 ret = quick_update_accounting(fs_info, src, dst, 1);
1619 spin_unlock(&fs_info->qgroup_lock);
1620 out:
1621 kfree(prealloc);
1622 mutex_unlock(&fs_info->qgroup_ioctl_lock);
1623 return ret;
1624 }
1625
1626 static int __del_qgroup_relation(struct btrfs_trans_handle *trans, u64 src,
1627 u64 dst)
1628 {
1629 struct btrfs_fs_info *fs_info = trans->fs_info;
1630 struct btrfs_qgroup *parent;
1631 struct btrfs_qgroup *member;
1632 struct btrfs_qgroup_list *list;
1633 bool found = false;
1634 int ret = 0;
1635 int ret2;
1636
1637 if (!fs_info->quota_root) {
1638 ret = -ENOTCONN;
1639 goto out;
1640 }
1641
1642 member = find_qgroup_rb(fs_info, src);
1643 parent = find_qgroup_rb(fs_info, dst);
1644 /*
1645 * The parent/member pair doesn't exist, then try to delete the dead
1646 * relation items only.
1647 */
1648 if (!member || !parent)
1649 goto delete_item;
1650
1651 /* check if such qgroup relation exist firstly */
1652 list_for_each_entry(list, &member->groups, next_group) {
1653 if (list->group == parent) {
1654 found = true;
1655 break;
1656 }
1657 }
1658
1659 delete_item:
1660 ret = del_qgroup_relation_item(trans, src, dst);
1661 if (ret < 0 && ret != -ENOENT)
1662 goto out;
1663 ret2 = del_qgroup_relation_item(trans, dst, src);
1664 if (ret2 < 0 && ret2 != -ENOENT)
1665 goto out;
1666
1667 /* At least one deletion succeeded, return 0 */
1668 if (!ret || !ret2)
1669 ret = 0;
1670
1671 if (found) {
1672 spin_lock(&fs_info->qgroup_lock);
1673 del_relation_rb(fs_info, src, dst);
1674 ret = quick_update_accounting(fs_info, src, dst, -1);
1675 spin_unlock(&fs_info->qgroup_lock);
1676 }
1677 out:
1678 return ret;
1679 }
1680
1681 int btrfs_del_qgroup_relation(struct btrfs_trans_handle *trans, u64 src,
1682 u64 dst)
1683 {
1684 struct btrfs_fs_info *fs_info = trans->fs_info;
1685 int ret = 0;
1686
1687 mutex_lock(&fs_info->qgroup_ioctl_lock);
1688 ret = __del_qgroup_relation(trans, src, dst);
1689 mutex_unlock(&fs_info->qgroup_ioctl_lock);
1690
1691 return ret;
1692 }
1693
1694 int btrfs_create_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid)
1695 {
1696 struct btrfs_fs_info *fs_info = trans->fs_info;
1697 struct btrfs_root *quota_root;
1698 struct btrfs_qgroup *qgroup;
1699 struct btrfs_qgroup *prealloc = NULL;
1700 int ret = 0;
1701
1702 if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_DISABLED)
1703 return 0;
1704
1705 mutex_lock(&fs_info->qgroup_ioctl_lock);
1706 if (!fs_info->quota_root) {
1707 ret = -ENOTCONN;
1708 goto out;
1709 }
1710 quota_root = fs_info->quota_root;
1711 qgroup = find_qgroup_rb(fs_info, qgroupid);
1712 if (qgroup) {
1713 ret = -EEXIST;
1714 goto out;
1715 }
1716
1717 prealloc = kzalloc(sizeof(*prealloc), GFP_NOFS);
1718 if (!prealloc) {
1719 ret = -ENOMEM;
1720 goto out;
1721 }
1722
1723 ret = add_qgroup_item(trans, quota_root, qgroupid);
1724 if (ret)
1725 goto out;
1726
1727 spin_lock(&fs_info->qgroup_lock);
1728 qgroup = add_qgroup_rb(fs_info, prealloc, qgroupid);
1729 spin_unlock(&fs_info->qgroup_lock);
1730 prealloc = NULL;
1731
1732 ret = btrfs_sysfs_add_one_qgroup(fs_info, qgroup);
1733 out:
1734 mutex_unlock(&fs_info->qgroup_ioctl_lock);
1735 kfree(prealloc);
1736 return ret;
1737 }
1738
1739 /*
1740 * Return 0 if we can not delete the qgroup (not empty or has children etc).
1741 * Return >0 if we can delete the qgroup.
1742 * Return <0 for other errors during tree search.
1743 */
1744 static int can_delete_qgroup(struct btrfs_fs_info *fs_info, struct btrfs_qgroup *qgroup)
1745 {
1746 struct btrfs_key key;
1747 struct btrfs_path *path;
1748 int ret;
1749
1750 /*
1751 * Squota would never be inconsistent, but there can still be case
1752 * where a dropped subvolume still has qgroup numbers, and squota
1753 * relies on such qgroup for future accounting.
1754 *
1755 * So for squota, do not allow dropping any non-zero qgroup.
1756 */
1757 if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_SIMPLE &&
1758 (qgroup->rfer || qgroup->excl || qgroup->excl_cmpr || qgroup->rfer_cmpr))
1759 return 0;
1760
1761 /* For higher level qgroup, we can only delete it if it has no child. */
1762 if (btrfs_qgroup_level(qgroup->qgroupid)) {
1763 if (!list_empty(&qgroup->members))
1764 return 0;
1765 return 1;
1766 }
1767
1768 /*
1769 * For level-0 qgroups, we can only delete it if it has no subvolume
1770 * for it.
1771 * This means even a subvolume is unlinked but not yet fully dropped,
1772 * we can not delete the qgroup.
1773 */
1774 key.objectid = qgroup->qgroupid;
1775 key.type = BTRFS_ROOT_ITEM_KEY;
1776 key.offset = -1ULL;
1777 path = btrfs_alloc_path();
1778 if (!path)
1779 return -ENOMEM;
1780
1781 ret = btrfs_find_root(fs_info->tree_root, &key, path, NULL, NULL);
1782 btrfs_free_path(path);
1783 /*
1784 * The @ret from btrfs_find_root() exactly matches our definition for
1785 * the return value, thus can be returned directly.
1786 */
1787 return ret;
1788 }
1789
1790 int btrfs_remove_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid)
1791 {
1792 struct btrfs_fs_info *fs_info = trans->fs_info;
1793 struct btrfs_qgroup *qgroup;
1794 struct btrfs_qgroup_list *list;
1795 int ret = 0;
1796
1797 mutex_lock(&fs_info->qgroup_ioctl_lock);
1798 if (!fs_info->quota_root) {
1799 ret = -ENOTCONN;
1800 goto out;
1801 }
1802
1803 qgroup = find_qgroup_rb(fs_info, qgroupid);
1804 if (!qgroup) {
1805 ret = -ENOENT;
1806 goto out;
1807 }
1808
1809 ret = can_delete_qgroup(fs_info, qgroup);
1810 if (ret < 0)
1811 goto out;
1812 if (ret == 0) {
1813 ret = -EBUSY;
1814 goto out;
1815 }
1816
1817 /* Check if there are no children of this qgroup */
1818 if (!list_empty(&qgroup->members)) {
1819 ret = -EBUSY;
1820 goto out;
1821 }
1822
1823 ret = del_qgroup_item(trans, qgroupid);
1824 if (ret && ret != -ENOENT)
1825 goto out;
1826
1827 while (!list_empty(&qgroup->groups)) {
1828 list = list_first_entry(&qgroup->groups,
1829 struct btrfs_qgroup_list, next_group);
1830 ret = __del_qgroup_relation(trans, qgroupid,
1831 list->group->qgroupid);
1832 if (ret)
1833 goto out;
1834 }
1835
1836 spin_lock(&fs_info->qgroup_lock);
1837 /*
1838 * Warn on reserved space. The subvolume should has no child nor
1839 * corresponding subvolume.
1840 * Thus its reserved space should all be zero, no matter if qgroup
1841 * is consistent or the mode.
1842 */
1843 WARN_ON(qgroup->rsv.values[BTRFS_QGROUP_RSV_DATA] ||
1844 qgroup->rsv.values[BTRFS_QGROUP_RSV_META_PREALLOC] ||
1845 qgroup->rsv.values[BTRFS_QGROUP_RSV_META_PERTRANS]);
1846 /*
1847 * The same for rfer/excl numbers, but that's only if our qgroup is
1848 * consistent and if it's in regular qgroup mode.
1849 * For simple mode it's not as accurate thus we can hit non-zero values
1850 * very frequently.
1851 */
1852 if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_FULL &&
1853 !(fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT)) {
1854 if (WARN_ON(qgroup->rfer || qgroup->excl ||
1855 qgroup->rfer_cmpr || qgroup->excl_cmpr)) {
1856 btrfs_warn_rl(fs_info,
1857 "to be deleted qgroup %u/%llu has non-zero numbers, rfer %llu rfer_cmpr %llu excl %llu excl_cmpr %llu",
1858 btrfs_qgroup_level(qgroup->qgroupid),
1859 btrfs_qgroup_subvolid(qgroup->qgroupid),
1860 qgroup->rfer, qgroup->rfer_cmpr,
1861 qgroup->excl, qgroup->excl_cmpr);
1862 qgroup_mark_inconsistent(fs_info);
1863 }
1864 }
1865 del_qgroup_rb(fs_info, qgroupid);
1866 spin_unlock(&fs_info->qgroup_lock);
1867
1868 /*
1869 * Remove the qgroup from sysfs now without holding the qgroup_lock
1870 * spinlock, since the sysfs_remove_group() function needs to take
1871 * the mutex kernfs_mutex through kernfs_remove_by_name_ns().
1872 */
1873 btrfs_sysfs_del_one_qgroup(fs_info, qgroup);
1874 kfree(qgroup);
1875 out:
1876 mutex_unlock(&fs_info->qgroup_ioctl_lock);
1877 return ret;
1878 }
1879
1880 int btrfs_qgroup_cleanup_dropped_subvolume(struct btrfs_fs_info *fs_info, u64 subvolid)
1881 {
1882 struct btrfs_trans_handle *trans;
1883 int ret;
1884
1885 if (!is_fstree(subvolid) || !btrfs_qgroup_enabled(fs_info) || !fs_info->quota_root)
1886 return 0;
1887
1888 /*
1889 * Commit current transaction to make sure all the rfer/excl numbers
1890 * get updated.
1891 */
1892 trans = btrfs_start_transaction(fs_info->quota_root, 0);
1893 if (IS_ERR(trans))
1894 return PTR_ERR(trans);
1895
1896 ret = btrfs_commit_transaction(trans);
1897 if (ret < 0)
1898 return ret;
1899
1900 /* Start new trans to delete the qgroup info and limit items. */
1901 trans = btrfs_start_transaction(fs_info->quota_root, 2);
1902 if (IS_ERR(trans))
1903 return PTR_ERR(trans);
1904 ret = btrfs_remove_qgroup(trans, subvolid);
1905 btrfs_end_transaction(trans);
1906 /*
1907 * It's squota and the subvolume still has numbers needed for future
1908 * accounting, in this case we can not delete it. Just skip it.
1909 */
1910 if (ret == -EBUSY)
1911 ret = 0;
1912 return ret;
1913 }
1914
1915 int btrfs_limit_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid,
1916 struct btrfs_qgroup_limit *limit)
1917 {
1918 struct btrfs_fs_info *fs_info = trans->fs_info;
1919 struct btrfs_qgroup *qgroup;
1920 int ret = 0;
1921 /* Sometimes we would want to clear the limit on this qgroup.
1922 * To meet this requirement, we treat the -1 as a special value
1923 * which tell kernel to clear the limit on this qgroup.
1924 */
1925 const u64 CLEAR_VALUE = -1;
1926
1927 mutex_lock(&fs_info->qgroup_ioctl_lock);
1928 if (!fs_info->quota_root) {
1929 ret = -ENOTCONN;
1930 goto out;
1931 }
1932
1933 qgroup = find_qgroup_rb(fs_info, qgroupid);
1934 if (!qgroup) {
1935 ret = -ENOENT;
1936 goto out;
1937 }
1938
1939 spin_lock(&fs_info->qgroup_lock);
1940 if (limit->flags & BTRFS_QGROUP_LIMIT_MAX_RFER) {
1941 if (limit->max_rfer == CLEAR_VALUE) {
1942 qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_MAX_RFER;
1943 limit->flags &= ~BTRFS_QGROUP_LIMIT_MAX_RFER;
1944 qgroup->max_rfer = 0;
1945 } else {
1946 qgroup->max_rfer = limit->max_rfer;
1947 }
1948 }
1949 if (limit->flags & BTRFS_QGROUP_LIMIT_MAX_EXCL) {
1950 if (limit->max_excl == CLEAR_VALUE) {
1951 qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_MAX_EXCL;
1952 limit->flags &= ~BTRFS_QGROUP_LIMIT_MAX_EXCL;
1953 qgroup->max_excl = 0;
1954 } else {
1955 qgroup->max_excl = limit->max_excl;
1956 }
1957 }
1958 if (limit->flags & BTRFS_QGROUP_LIMIT_RSV_RFER) {
1959 if (limit->rsv_rfer == CLEAR_VALUE) {
1960 qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_RSV_RFER;
1961 limit->flags &= ~BTRFS_QGROUP_LIMIT_RSV_RFER;
1962 qgroup->rsv_rfer = 0;
1963 } else {
1964 qgroup->rsv_rfer = limit->rsv_rfer;
1965 }
1966 }
1967 if (limit->flags & BTRFS_QGROUP_LIMIT_RSV_EXCL) {
1968 if (limit->rsv_excl == CLEAR_VALUE) {
1969 qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_RSV_EXCL;
1970 limit->flags &= ~BTRFS_QGROUP_LIMIT_RSV_EXCL;
1971 qgroup->rsv_excl = 0;
1972 } else {
1973 qgroup->rsv_excl = limit->rsv_excl;
1974 }
1975 }
1976 qgroup->lim_flags |= limit->flags;
1977
1978 spin_unlock(&fs_info->qgroup_lock);
1979
1980 ret = update_qgroup_limit_item(trans, qgroup);
1981 if (ret) {
1982 qgroup_mark_inconsistent(fs_info);
1983 btrfs_info(fs_info, "unable to update quota limit for %llu",
1984 qgroupid);
1985 }
1986
1987 out:
1988 mutex_unlock(&fs_info->qgroup_ioctl_lock);
1989 return ret;
1990 }
1991
1992 /*
1993 * Inform qgroup to trace one dirty extent, its info is recorded in @record.
1994 * So qgroup can account it at transaction committing time.
1995 *
1996 * No lock version, caller must acquire delayed ref lock and allocated memory,
1997 * then call btrfs_qgroup_trace_extent_post() after exiting lock context.
1998 *
1999 * Return 0 for success insert
2000 * Return >0 for existing record, caller can free @record safely.
2001 * Error is not possible
2002 */
2003 int btrfs_qgroup_trace_extent_nolock(struct btrfs_fs_info *fs_info,
2004 struct btrfs_delayed_ref_root *delayed_refs,
2005 struct btrfs_qgroup_extent_record *record)
2006 {
2007 struct rb_node **p = &delayed_refs->dirty_extent_root.rb_node;
2008 struct rb_node *parent_node = NULL;
2009 struct btrfs_qgroup_extent_record *entry;
2010 u64 bytenr = record->bytenr;
2011
2012 if (!btrfs_qgroup_full_accounting(fs_info))
2013 return 1;
2014
2015 lockdep_assert_held(&delayed_refs->lock);
2016 trace_btrfs_qgroup_trace_extent(fs_info, record);
2017
2018 while (*p) {
2019 parent_node = *p;
2020 entry = rb_entry(parent_node, struct btrfs_qgroup_extent_record,
2021 node);
2022 if (bytenr < entry->bytenr) {
2023 p = &(*p)->rb_left;
2024 } else if (bytenr > entry->bytenr) {
2025 p = &(*p)->rb_right;
2026 } else {
2027 if (record->data_rsv && !entry->data_rsv) {
2028 entry->data_rsv = record->data_rsv;
2029 entry->data_rsv_refroot =
2030 record->data_rsv_refroot;
2031 }
2032 return 1;
2033 }
2034 }
2035
2036 rb_link_node(&record->node, parent_node, p);
2037 rb_insert_color(&record->node, &delayed_refs->dirty_extent_root);
2038 return 0;
2039 }
2040
2041 /*
2042 * Post handler after qgroup_trace_extent_nolock().
2043 *
2044 * NOTE: Current qgroup does the expensive backref walk at transaction
2045 * committing time with TRANS_STATE_COMMIT_DOING, this blocks incoming
2046 * new transaction.
2047 * This is designed to allow btrfs_find_all_roots() to get correct new_roots
2048 * result.
2049 *
2050 * However for old_roots there is no need to do backref walk at that time,
2051 * since we search commit roots to walk backref and result will always be
2052 * correct.
2053 *
2054 * Due to the nature of no lock version, we can't do backref there.
2055 * So we must call btrfs_qgroup_trace_extent_post() after exiting
2056 * spinlock context.
2057 *
2058 * TODO: If we can fix and prove btrfs_find_all_roots() can get correct result
2059 * using current root, then we can move all expensive backref walk out of
2060 * transaction committing, but not now as qgroup accounting will be wrong again.
2061 */
2062 int btrfs_qgroup_trace_extent_post(struct btrfs_trans_handle *trans,
2063 struct btrfs_qgroup_extent_record *qrecord)
2064 {
2065 struct btrfs_backref_walk_ctx ctx = { 0 };
2066 int ret;
2067
2068 if (!btrfs_qgroup_full_accounting(trans->fs_info))
2069 return 0;
2070 /*
2071 * We are always called in a context where we are already holding a
2072 * transaction handle. Often we are called when adding a data delayed
2073 * reference from btrfs_truncate_inode_items() (truncating or unlinking),
2074 * in which case we will be holding a write lock on extent buffer from a
2075 * subvolume tree. In this case we can't allow btrfs_find_all_roots() to
2076 * acquire fs_info->commit_root_sem, because that is a higher level lock
2077 * that must be acquired before locking any extent buffers.
2078 *
2079 * So we want btrfs_find_all_roots() to not acquire the commit_root_sem
2080 * but we can't pass it a non-NULL transaction handle, because otherwise
2081 * it would not use commit roots and would lock extent buffers, causing
2082 * a deadlock if it ends up trying to read lock the same extent buffer
2083 * that was previously write locked at btrfs_truncate_inode_items().
2084 *
2085 * So pass a NULL transaction handle to btrfs_find_all_roots() and
2086 * explicitly tell it to not acquire the commit_root_sem - if we are
2087 * holding a transaction handle we don't need its protection.
2088 */
2089 ASSERT(trans != NULL);
2090
2091 if (trans->fs_info->qgroup_flags & BTRFS_QGROUP_RUNTIME_FLAG_NO_ACCOUNTING)
2092 return 0;
2093
2094 ctx.bytenr = qrecord->bytenr;
2095 ctx.fs_info = trans->fs_info;
2096
2097 ret = btrfs_find_all_roots(&ctx, true);
2098 if (ret < 0) {
2099 qgroup_mark_inconsistent(trans->fs_info);
2100 btrfs_warn(trans->fs_info,
2101 "error accounting new delayed refs extent (err code: %d), quota inconsistent",
2102 ret);
2103 return 0;
2104 }
2105
2106 /*
2107 * Here we don't need to get the lock of
2108 * trans->transaction->delayed_refs, since inserted qrecord won't
2109 * be deleted, only qrecord->node may be modified (new qrecord insert)
2110 *
2111 * So modifying qrecord->old_roots is safe here
2112 */
2113 qrecord->old_roots = ctx.roots;
2114 return 0;
2115 }
2116
2117 /*
2118 * Inform qgroup to trace one dirty extent, specified by @bytenr and
2119 * @num_bytes.
2120 * So qgroup can account it at commit trans time.
2121 *
2122 * Better encapsulated version, with memory allocation and backref walk for
2123 * commit roots.
2124 * So this can sleep.
2125 *
2126 * Return 0 if the operation is done.
2127 * Return <0 for error, like memory allocation failure or invalid parameter
2128 * (NULL trans)
2129 */
2130 int btrfs_qgroup_trace_extent(struct btrfs_trans_handle *trans, u64 bytenr,
2131 u64 num_bytes)
2132 {
2133 struct btrfs_fs_info *fs_info = trans->fs_info;
2134 struct btrfs_qgroup_extent_record *record;
2135 struct btrfs_delayed_ref_root *delayed_refs;
2136 int ret;
2137
2138 if (!btrfs_qgroup_full_accounting(fs_info) || bytenr == 0 || num_bytes == 0)
2139 return 0;
2140 record = kzalloc(sizeof(*record), GFP_NOFS);
2141 if (!record)
2142 return -ENOMEM;
2143
2144 delayed_refs = &trans->transaction->delayed_refs;
2145 record->bytenr = bytenr;
2146 record->num_bytes = num_bytes;
2147 record->old_roots = NULL;
2148
2149 spin_lock(&delayed_refs->lock);
2150 ret = btrfs_qgroup_trace_extent_nolock(fs_info, delayed_refs, record);
2151 spin_unlock(&delayed_refs->lock);
2152 if (ret > 0) {
2153 kfree(record);
2154 return 0;
2155 }
2156 return btrfs_qgroup_trace_extent_post(trans, record);
2157 }
2158
2159 /*
2160 * Inform qgroup to trace all leaf items of data
2161 *
2162 * Return 0 for success
2163 * Return <0 for error(ENOMEM)
2164 */
2165 int btrfs_qgroup_trace_leaf_items(struct btrfs_trans_handle *trans,
2166 struct extent_buffer *eb)
2167 {
2168 struct btrfs_fs_info *fs_info = trans->fs_info;
2169 int nr = btrfs_header_nritems(eb);
2170 int i, extent_type, ret;
2171 struct btrfs_key key;
2172 struct btrfs_file_extent_item *fi;
2173 u64 bytenr, num_bytes;
2174
2175 /* We can be called directly from walk_up_proc() */
2176 if (!btrfs_qgroup_full_accounting(fs_info))
2177 return 0;
2178
2179 for (i = 0; i < nr; i++) {
2180 btrfs_item_key_to_cpu(eb, &key, i);
2181
2182 if (key.type != BTRFS_EXTENT_DATA_KEY)
2183 continue;
2184
2185 fi = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
2186 /* filter out non qgroup-accountable extents */
2187 extent_type = btrfs_file_extent_type(eb, fi);
2188
2189 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
2190 continue;
2191
2192 bytenr = btrfs_file_extent_disk_bytenr(eb, fi);
2193 if (!bytenr)
2194 continue;
2195
2196 num_bytes = btrfs_file_extent_disk_num_bytes(eb, fi);
2197
2198 ret = btrfs_qgroup_trace_extent(trans, bytenr, num_bytes);
2199 if (ret)
2200 return ret;
2201 }
2202 cond_resched();
2203 return 0;
2204 }
2205
2206 /*
2207 * Walk up the tree from the bottom, freeing leaves and any interior
2208 * nodes which have had all slots visited. If a node (leaf or
2209 * interior) is freed, the node above it will have it's slot
2210 * incremented. The root node will never be freed.
2211 *
2212 * At the end of this function, we should have a path which has all
2213 * slots incremented to the next position for a search. If we need to
2214 * read a new node it will be NULL and the node above it will have the
2215 * correct slot selected for a later read.
2216 *
2217 * If we increment the root nodes slot counter past the number of
2218 * elements, 1 is returned to signal completion of the search.
2219 */
2220 static int adjust_slots_upwards(struct btrfs_path *path, int root_level)
2221 {
2222 int level = 0;
2223 int nr, slot;
2224 struct extent_buffer *eb;
2225
2226 if (root_level == 0)
2227 return 1;
2228
2229 while (level <= root_level) {
2230 eb = path->nodes[level];
2231 nr = btrfs_header_nritems(eb);
2232 path->slots[level]++;
2233 slot = path->slots[level];
2234 if (slot >= nr || level == 0) {
2235 /*
2236 * Don't free the root - we will detect this
2237 * condition after our loop and return a
2238 * positive value for caller to stop walking the tree.
2239 */
2240 if (level != root_level) {
2241 btrfs_tree_unlock_rw(eb, path->locks[level]);
2242 path->locks[level] = 0;
2243
2244 free_extent_buffer(eb);
2245 path->nodes[level] = NULL;
2246 path->slots[level] = 0;
2247 }
2248 } else {
2249 /*
2250 * We have a valid slot to walk back down
2251 * from. Stop here so caller can process these
2252 * new nodes.
2253 */
2254 break;
2255 }
2256
2257 level++;
2258 }
2259
2260 eb = path->nodes[root_level];
2261 if (path->slots[root_level] >= btrfs_header_nritems(eb))
2262 return 1;
2263
2264 return 0;
2265 }
2266
2267 /*
2268 * Helper function to trace a subtree tree block swap.
2269 *
2270 * The swap will happen in highest tree block, but there may be a lot of
2271 * tree blocks involved.
2272 *
2273 * For example:
2274 * OO = Old tree blocks
2275 * NN = New tree blocks allocated during balance
2276 *
2277 * File tree (257) Reloc tree for 257
2278 * L2 OO NN
2279 * / \ / \
2280 * L1 OO OO (a) OO NN (a)
2281 * / \ / \ / \ / \
2282 * L0 OO OO OO OO OO OO NN NN
2283 * (b) (c) (b) (c)
2284 *
2285 * When calling qgroup_trace_extent_swap(), we will pass:
2286 * @src_eb = OO(a)
2287 * @dst_path = [ nodes[1] = NN(a), nodes[0] = NN(c) ]
2288 * @dst_level = 0
2289 * @root_level = 1
2290 *
2291 * In that case, qgroup_trace_extent_swap() will search from OO(a) to
2292 * reach OO(c), then mark both OO(c) and NN(c) as qgroup dirty.
2293 *
2294 * The main work of qgroup_trace_extent_swap() can be split into 3 parts:
2295 *
2296 * 1) Tree search from @src_eb
2297 * It should acts as a simplified btrfs_search_slot().
2298 * The key for search can be extracted from @dst_path->nodes[dst_level]
2299 * (first key).
2300 *
2301 * 2) Mark the final tree blocks in @src_path and @dst_path qgroup dirty
2302 * NOTE: In above case, OO(a) and NN(a) won't be marked qgroup dirty.
2303 * They should be marked during previous (@dst_level = 1) iteration.
2304 *
2305 * 3) Mark file extents in leaves dirty
2306 * We don't have good way to pick out new file extents only.
2307 * So we still follow the old method by scanning all file extents in
2308 * the leave.
2309 *
2310 * This function can free us from keeping two paths, thus later we only need
2311 * to care about how to iterate all new tree blocks in reloc tree.
2312 */
2313 static int qgroup_trace_extent_swap(struct btrfs_trans_handle* trans,
2314 struct extent_buffer *src_eb,
2315 struct btrfs_path *dst_path,
2316 int dst_level, int root_level,
2317 bool trace_leaf)
2318 {
2319 struct btrfs_key key;
2320 struct btrfs_path *src_path;
2321 struct btrfs_fs_info *fs_info = trans->fs_info;
2322 u32 nodesize = fs_info->nodesize;
2323 int cur_level = root_level;
2324 int ret;
2325
2326 BUG_ON(dst_level > root_level);
2327 /* Level mismatch */
2328 if (btrfs_header_level(src_eb) != root_level)
2329 return -EINVAL;
2330
2331 src_path = btrfs_alloc_path();
2332 if (!src_path) {
2333 ret = -ENOMEM;
2334 goto out;
2335 }
2336
2337 if (dst_level)
2338 btrfs_node_key_to_cpu(dst_path->nodes[dst_level], &key, 0);
2339 else
2340 btrfs_item_key_to_cpu(dst_path->nodes[dst_level], &key, 0);
2341
2342 /* For src_path */
2343 atomic_inc(&src_eb->refs);
2344 src_path->nodes[root_level] = src_eb;
2345 src_path->slots[root_level] = dst_path->slots[root_level];
2346 src_path->locks[root_level] = 0;
2347
2348 /* A simplified version of btrfs_search_slot() */
2349 while (cur_level >= dst_level) {
2350 struct btrfs_key src_key;
2351 struct btrfs_key dst_key;
2352
2353 if (src_path->nodes[cur_level] == NULL) {
2354 struct extent_buffer *eb;
2355 int parent_slot;
2356
2357 eb = src_path->nodes[cur_level + 1];
2358 parent_slot = src_path->slots[cur_level + 1];
2359
2360 eb = btrfs_read_node_slot(eb, parent_slot);
2361 if (IS_ERR(eb)) {
2362 ret = PTR_ERR(eb);
2363 goto out;
2364 }
2365
2366 src_path->nodes[cur_level] = eb;
2367
2368 btrfs_tree_read_lock(eb);
2369 src_path->locks[cur_level] = BTRFS_READ_LOCK;
2370 }
2371
2372 src_path->slots[cur_level] = dst_path->slots[cur_level];
2373 if (cur_level) {
2374 btrfs_node_key_to_cpu(dst_path->nodes[cur_level],
2375 &dst_key, dst_path->slots[cur_level]);
2376 btrfs_node_key_to_cpu(src_path->nodes[cur_level],
2377 &src_key, src_path->slots[cur_level]);
2378 } else {
2379 btrfs_item_key_to_cpu(dst_path->nodes[cur_level],
2380 &dst_key, dst_path->slots[cur_level]);
2381 btrfs_item_key_to_cpu(src_path->nodes[cur_level],
2382 &src_key, src_path->slots[cur_level]);
2383 }
2384 /* Content mismatch, something went wrong */
2385 if (btrfs_comp_cpu_keys(&dst_key, &src_key)) {
2386 ret = -ENOENT;
2387 goto out;
2388 }
2389 cur_level--;
2390 }
2391
2392 /*
2393 * Now both @dst_path and @src_path have been populated, record the tree
2394 * blocks for qgroup accounting.
2395 */
2396 ret = btrfs_qgroup_trace_extent(trans, src_path->nodes[dst_level]->start,
2397 nodesize);
2398 if (ret < 0)
2399 goto out;
2400 ret = btrfs_qgroup_trace_extent(trans, dst_path->nodes[dst_level]->start,
2401 nodesize);
2402 if (ret < 0)
2403 goto out;
2404
2405 /* Record leaf file extents */
2406 if (dst_level == 0 && trace_leaf) {
2407 ret = btrfs_qgroup_trace_leaf_items(trans, src_path->nodes[0]);
2408 if (ret < 0)
2409 goto out;
2410 ret = btrfs_qgroup_trace_leaf_items(trans, dst_path->nodes[0]);
2411 }
2412 out:
2413 btrfs_free_path(src_path);
2414 return ret;
2415 }
2416
2417 /*
2418 * Helper function to do recursive generation-aware depth-first search, to
2419 * locate all new tree blocks in a subtree of reloc tree.
2420 *
2421 * E.g. (OO = Old tree blocks, NN = New tree blocks, whose gen == last_snapshot)
2422 * reloc tree
2423 * L2 NN (a)
2424 * / \
2425 * L1 OO NN (b)
2426 * / \ / \
2427 * L0 OO OO OO NN
2428 * (c) (d)
2429 * If we pass:
2430 * @dst_path = [ nodes[1] = NN(b), nodes[0] = NULL ],
2431 * @cur_level = 1
2432 * @root_level = 1
2433 *
2434 * We will iterate through tree blocks NN(b), NN(d) and info qgroup to trace
2435 * above tree blocks along with their counter parts in file tree.
2436 * While during search, old tree blocks OO(c) will be skipped as tree block swap
2437 * won't affect OO(c).
2438 */
2439 static int qgroup_trace_new_subtree_blocks(struct btrfs_trans_handle* trans,
2440 struct extent_buffer *src_eb,
2441 struct btrfs_path *dst_path,
2442 int cur_level, int root_level,
2443 u64 last_snapshot, bool trace_leaf)
2444 {
2445 struct btrfs_fs_info *fs_info = trans->fs_info;
2446 struct extent_buffer *eb;
2447 bool need_cleanup = false;
2448 int ret = 0;
2449 int i;
2450
2451 /* Level sanity check */
2452 if (cur_level < 0 || cur_level >= BTRFS_MAX_LEVEL - 1 ||
2453 root_level < 0 || root_level >= BTRFS_MAX_LEVEL - 1 ||
2454 root_level < cur_level) {
2455 btrfs_err_rl(fs_info,
2456 "%s: bad levels, cur_level=%d root_level=%d",
2457 __func__, cur_level, root_level);
2458 return -EUCLEAN;
2459 }
2460
2461 /* Read the tree block if needed */
2462 if (dst_path->nodes[cur_level] == NULL) {
2463 int parent_slot;
2464 u64 child_gen;
2465
2466 /*
2467 * dst_path->nodes[root_level] must be initialized before
2468 * calling this function.
2469 */
2470 if (cur_level == root_level) {
2471 btrfs_err_rl(fs_info,
2472 "%s: dst_path->nodes[%d] not initialized, root_level=%d cur_level=%d",
2473 __func__, root_level, root_level, cur_level);
2474 return -EUCLEAN;
2475 }
2476
2477 /*
2478 * We need to get child blockptr/gen from parent before we can
2479 * read it.
2480 */
2481 eb = dst_path->nodes[cur_level + 1];
2482 parent_slot = dst_path->slots[cur_level + 1];
2483 child_gen = btrfs_node_ptr_generation(eb, parent_slot);
2484
2485 /* This node is old, no need to trace */
2486 if (child_gen < last_snapshot)
2487 goto out;
2488
2489 eb = btrfs_read_node_slot(eb, parent_slot);
2490 if (IS_ERR(eb)) {
2491 ret = PTR_ERR(eb);
2492 goto out;
2493 }
2494
2495 dst_path->nodes[cur_level] = eb;
2496 dst_path->slots[cur_level] = 0;
2497
2498 btrfs_tree_read_lock(eb);
2499 dst_path->locks[cur_level] = BTRFS_READ_LOCK;
2500 need_cleanup = true;
2501 }
2502
2503 /* Now record this tree block and its counter part for qgroups */
2504 ret = qgroup_trace_extent_swap(trans, src_eb, dst_path, cur_level,
2505 root_level, trace_leaf);
2506 if (ret < 0)
2507 goto cleanup;
2508
2509 eb = dst_path->nodes[cur_level];
2510
2511 if (cur_level > 0) {
2512 /* Iterate all child tree blocks */
2513 for (i = 0; i < btrfs_header_nritems(eb); i++) {
2514 /* Skip old tree blocks as they won't be swapped */
2515 if (btrfs_node_ptr_generation(eb, i) < last_snapshot)
2516 continue;
2517 dst_path->slots[cur_level] = i;
2518
2519 /* Recursive call (at most 7 times) */
2520 ret = qgroup_trace_new_subtree_blocks(trans, src_eb,
2521 dst_path, cur_level - 1, root_level,
2522 last_snapshot, trace_leaf);
2523 if (ret < 0)
2524 goto cleanup;
2525 }
2526 }
2527
2528 cleanup:
2529 if (need_cleanup) {
2530 /* Clean up */
2531 btrfs_tree_unlock_rw(dst_path->nodes[cur_level],
2532 dst_path->locks[cur_level]);
2533 free_extent_buffer(dst_path->nodes[cur_level]);
2534 dst_path->nodes[cur_level] = NULL;
2535 dst_path->slots[cur_level] = 0;
2536 dst_path->locks[cur_level] = 0;
2537 }
2538 out:
2539 return ret;
2540 }
2541
2542 static int qgroup_trace_subtree_swap(struct btrfs_trans_handle *trans,
2543 struct extent_buffer *src_eb,
2544 struct extent_buffer *dst_eb,
2545 u64 last_snapshot, bool trace_leaf)
2546 {
2547 struct btrfs_fs_info *fs_info = trans->fs_info;
2548 struct btrfs_path *dst_path = NULL;
2549 int level;
2550 int ret;
2551
2552 if (!btrfs_qgroup_full_accounting(fs_info))
2553 return 0;
2554
2555 /* Wrong parameter order */
2556 if (btrfs_header_generation(src_eb) > btrfs_header_generation(dst_eb)) {
2557 btrfs_err_rl(fs_info,
2558 "%s: bad parameter order, src_gen=%llu dst_gen=%llu", __func__,
2559 btrfs_header_generation(src_eb),
2560 btrfs_header_generation(dst_eb));
2561 return -EUCLEAN;
2562 }
2563
2564 if (!extent_buffer_uptodate(src_eb) || !extent_buffer_uptodate(dst_eb)) {
2565 ret = -EIO;
2566 goto out;
2567 }
2568
2569 level = btrfs_header_level(dst_eb);
2570 dst_path = btrfs_alloc_path();
2571 if (!dst_path) {
2572 ret = -ENOMEM;
2573 goto out;
2574 }
2575 /* For dst_path */
2576 atomic_inc(&dst_eb->refs);
2577 dst_path->nodes[level] = dst_eb;
2578 dst_path->slots[level] = 0;
2579 dst_path->locks[level] = 0;
2580
2581 /* Do the generation aware breadth-first search */
2582 ret = qgroup_trace_new_subtree_blocks(trans, src_eb, dst_path, level,
2583 level, last_snapshot, trace_leaf);
2584 if (ret < 0)
2585 goto out;
2586 ret = 0;
2587
2588 out:
2589 btrfs_free_path(dst_path);
2590 if (ret < 0)
2591 qgroup_mark_inconsistent(fs_info);
2592 return ret;
2593 }
2594
2595 /*
2596 * Inform qgroup to trace a whole subtree, including all its child tree
2597 * blocks and data.
2598 * The root tree block is specified by @root_eb.
2599 *
2600 * Normally used by relocation(tree block swap) and subvolume deletion.
2601 *
2602 * Return 0 for success
2603 * Return <0 for error(ENOMEM or tree search error)
2604 */
2605 int btrfs_qgroup_trace_subtree(struct btrfs_trans_handle *trans,
2606 struct extent_buffer *root_eb,
2607 u64 root_gen, int root_level)
2608 {
2609 struct btrfs_fs_info *fs_info = trans->fs_info;
2610 int ret = 0;
2611 int level;
2612 u8 drop_subptree_thres;
2613 struct extent_buffer *eb = root_eb;
2614 struct btrfs_path *path = NULL;
2615
2616 ASSERT(0 <= root_level && root_level < BTRFS_MAX_LEVEL);
2617 ASSERT(root_eb != NULL);
2618
2619 if (!btrfs_qgroup_full_accounting(fs_info))
2620 return 0;
2621
2622 spin_lock(&fs_info->qgroup_lock);
2623 drop_subptree_thres = fs_info->qgroup_drop_subtree_thres;
2624 spin_unlock(&fs_info->qgroup_lock);
2625
2626 /*
2627 * This function only gets called for snapshot drop, if we hit a high
2628 * node here, it means we are going to change ownership for quite a lot
2629 * of extents, which will greatly slow down btrfs_commit_transaction().
2630 *
2631 * So here if we find a high tree here, we just skip the accounting and
2632 * mark qgroup inconsistent.
2633 */
2634 if (root_level >= drop_subptree_thres) {
2635 qgroup_mark_inconsistent(fs_info);
2636 return 0;
2637 }
2638
2639 if (!extent_buffer_uptodate(root_eb)) {
2640 struct btrfs_tree_parent_check check = {
2641 .has_first_key = false,
2642 .transid = root_gen,
2643 .level = root_level
2644 };
2645
2646 ret = btrfs_read_extent_buffer(root_eb, &check);
2647 if (ret)
2648 goto out;
2649 }
2650
2651 if (root_level == 0) {
2652 ret = btrfs_qgroup_trace_leaf_items(trans, root_eb);
2653 goto out;
2654 }
2655
2656 path = btrfs_alloc_path();
2657 if (!path)
2658 return -ENOMEM;
2659
2660 /*
2661 * Walk down the tree. Missing extent blocks are filled in as
2662 * we go. Metadata is accounted every time we read a new
2663 * extent block.
2664 *
2665 * When we reach a leaf, we account for file extent items in it,
2666 * walk back up the tree (adjusting slot pointers as we go)
2667 * and restart the search process.
2668 */
2669 atomic_inc(&root_eb->refs); /* For path */
2670 path->nodes[root_level] = root_eb;
2671 path->slots[root_level] = 0;
2672 path->locks[root_level] = 0; /* so release_path doesn't try to unlock */
2673 walk_down:
2674 level = root_level;
2675 while (level >= 0) {
2676 if (path->nodes[level] == NULL) {
2677 int parent_slot;
2678 u64 child_bytenr;
2679
2680 /*
2681 * We need to get child blockptr from parent before we
2682 * can read it.
2683 */
2684 eb = path->nodes[level + 1];
2685 parent_slot = path->slots[level + 1];
2686 child_bytenr = btrfs_node_blockptr(eb, parent_slot);
2687
2688 eb = btrfs_read_node_slot(eb, parent_slot);
2689 if (IS_ERR(eb)) {
2690 ret = PTR_ERR(eb);
2691 goto out;
2692 }
2693
2694 path->nodes[level] = eb;
2695 path->slots[level] = 0;
2696
2697 btrfs_tree_read_lock(eb);
2698 path->locks[level] = BTRFS_READ_LOCK;
2699
2700 ret = btrfs_qgroup_trace_extent(trans, child_bytenr,
2701 fs_info->nodesize);
2702 if (ret)
2703 goto out;
2704 }
2705
2706 if (level == 0) {
2707 ret = btrfs_qgroup_trace_leaf_items(trans,
2708 path->nodes[level]);
2709 if (ret)
2710 goto out;
2711
2712 /* Nonzero return here means we completed our search */
2713 ret = adjust_slots_upwards(path, root_level);
2714 if (ret)
2715 break;
2716
2717 /* Restart search with new slots */
2718 goto walk_down;
2719 }
2720
2721 level--;
2722 }
2723
2724 ret = 0;
2725 out:
2726 btrfs_free_path(path);
2727
2728 return ret;
2729 }
2730
2731 static void qgroup_iterator_nested_add(struct list_head *head, struct btrfs_qgroup *qgroup)
2732 {
2733 if (!list_empty(&qgroup->nested_iterator))
2734 return;
2735
2736 list_add_tail(&qgroup->nested_iterator, head);
2737 }
2738
2739 static void qgroup_iterator_nested_clean(struct list_head *head)
2740 {
2741 while (!list_empty(head)) {
2742 struct btrfs_qgroup *qgroup;
2743
2744 qgroup = list_first_entry(head, struct btrfs_qgroup, nested_iterator);
2745 list_del_init(&qgroup->nested_iterator);
2746 }
2747 }
2748
2749 #define UPDATE_NEW 0
2750 #define UPDATE_OLD 1
2751 /*
2752 * Walk all of the roots that points to the bytenr and adjust their refcnts.
2753 */
2754 static void qgroup_update_refcnt(struct btrfs_fs_info *fs_info,
2755 struct ulist *roots, struct list_head *qgroups,
2756 u64 seq, int update_old)
2757 {
2758 struct ulist_node *unode;
2759 struct ulist_iterator uiter;
2760 struct btrfs_qgroup *qg;
2761
2762 if (!roots)
2763 return;
2764 ULIST_ITER_INIT(&uiter);
2765 while ((unode = ulist_next(roots, &uiter))) {
2766 LIST_HEAD(tmp);
2767
2768 qg = find_qgroup_rb(fs_info, unode->val);
2769 if (!qg)
2770 continue;
2771
2772 qgroup_iterator_nested_add(qgroups, qg);
2773 qgroup_iterator_add(&tmp, qg);
2774 list_for_each_entry(qg, &tmp, iterator) {
2775 struct btrfs_qgroup_list *glist;
2776
2777 if (update_old)
2778 btrfs_qgroup_update_old_refcnt(qg, seq, 1);
2779 else
2780 btrfs_qgroup_update_new_refcnt(qg, seq, 1);
2781
2782 list_for_each_entry(glist, &qg->groups, next_group) {
2783 qgroup_iterator_nested_add(qgroups, glist->group);
2784 qgroup_iterator_add(&tmp, glist->group);
2785 }
2786 }
2787 qgroup_iterator_clean(&tmp);
2788 }
2789 }
2790
2791 /*
2792 * Update qgroup rfer/excl counters.
2793 * Rfer update is easy, codes can explain themselves.
2794 *
2795 * Excl update is tricky, the update is split into 2 parts.
2796 * Part 1: Possible exclusive <-> sharing detect:
2797 * | A | !A |
2798 * -------------------------------------
2799 * B | * | - |
2800 * -------------------------------------
2801 * !B | + | ** |
2802 * -------------------------------------
2803 *
2804 * Conditions:
2805 * A: cur_old_roots < nr_old_roots (not exclusive before)
2806 * !A: cur_old_roots == nr_old_roots (possible exclusive before)
2807 * B: cur_new_roots < nr_new_roots (not exclusive now)
2808 * !B: cur_new_roots == nr_new_roots (possible exclusive now)
2809 *
2810 * Results:
2811 * +: Possible sharing -> exclusive -: Possible exclusive -> sharing
2812 * *: Definitely not changed. **: Possible unchanged.
2813 *
2814 * For !A and !B condition, the exception is cur_old/new_roots == 0 case.
2815 *
2816 * To make the logic clear, we first use condition A and B to split
2817 * combination into 4 results.
2818 *
2819 * Then, for result "+" and "-", check old/new_roots == 0 case, as in them
2820 * only on variant maybe 0.
2821 *
2822 * Lastly, check result **, since there are 2 variants maybe 0, split them
2823 * again(2x2).
2824 * But this time we don't need to consider other things, the codes and logic
2825 * is easy to understand now.
2826 */
2827 static void qgroup_update_counters(struct btrfs_fs_info *fs_info,
2828 struct list_head *qgroups, u64 nr_old_roots,
2829 u64 nr_new_roots, u64 num_bytes, u64 seq)
2830 {
2831 struct btrfs_qgroup *qg;
2832
2833 list_for_each_entry(qg, qgroups, nested_iterator) {
2834 u64 cur_new_count, cur_old_count;
2835 bool dirty = false;
2836
2837 cur_old_count = btrfs_qgroup_get_old_refcnt(qg, seq);
2838 cur_new_count = btrfs_qgroup_get_new_refcnt(qg, seq);
2839
2840 trace_qgroup_update_counters(fs_info, qg, cur_old_count,
2841 cur_new_count);
2842
2843 /* Rfer update part */
2844 if (cur_old_count == 0 && cur_new_count > 0) {
2845 qg->rfer += num_bytes;
2846 qg->rfer_cmpr += num_bytes;
2847 dirty = true;
2848 }
2849 if (cur_old_count > 0 && cur_new_count == 0) {
2850 qg->rfer -= num_bytes;
2851 qg->rfer_cmpr -= num_bytes;
2852 dirty = true;
2853 }
2854
2855 /* Excl update part */
2856 /* Exclusive/none -> shared case */
2857 if (cur_old_count == nr_old_roots &&
2858 cur_new_count < nr_new_roots) {
2859 /* Exclusive -> shared */
2860 if (cur_old_count != 0) {
2861 qg->excl -= num_bytes;
2862 qg->excl_cmpr -= num_bytes;
2863 dirty = true;
2864 }
2865 }
2866
2867 /* Shared -> exclusive/none case */
2868 if (cur_old_count < nr_old_roots &&
2869 cur_new_count == nr_new_roots) {
2870 /* Shared->exclusive */
2871 if (cur_new_count != 0) {
2872 qg->excl += num_bytes;
2873 qg->excl_cmpr += num_bytes;
2874 dirty = true;
2875 }
2876 }
2877
2878 /* Exclusive/none -> exclusive/none case */
2879 if (cur_old_count == nr_old_roots &&
2880 cur_new_count == nr_new_roots) {
2881 if (cur_old_count == 0) {
2882 /* None -> exclusive/none */
2883
2884 if (cur_new_count != 0) {
2885 /* None -> exclusive */
2886 qg->excl += num_bytes;
2887 qg->excl_cmpr += num_bytes;
2888 dirty = true;
2889 }
2890 /* None -> none, nothing changed */
2891 } else {
2892 /* Exclusive -> exclusive/none */
2893
2894 if (cur_new_count == 0) {
2895 /* Exclusive -> none */
2896 qg->excl -= num_bytes;
2897 qg->excl_cmpr -= num_bytes;
2898 dirty = true;
2899 }
2900 /* Exclusive -> exclusive, nothing changed */
2901 }
2902 }
2903
2904 if (dirty)
2905 qgroup_dirty(fs_info, qg);
2906 }
2907 }
2908
2909 /*
2910 * Check if the @roots potentially is a list of fs tree roots
2911 *
2912 * Return 0 for definitely not a fs/subvol tree roots ulist
2913 * Return 1 for possible fs/subvol tree roots in the list (considering an empty
2914 * one as well)
2915 */
2916 static int maybe_fs_roots(struct ulist *roots)
2917 {
2918 struct ulist_node *unode;
2919 struct ulist_iterator uiter;
2920
2921 /* Empty one, still possible for fs roots */
2922 if (!roots || roots->nnodes == 0)
2923 return 1;
2924
2925 ULIST_ITER_INIT(&uiter);
2926 unode = ulist_next(roots, &uiter);
2927 if (!unode)
2928 return 1;
2929
2930 /*
2931 * If it contains fs tree roots, then it must belong to fs/subvol
2932 * trees.
2933 * If it contains a non-fs tree, it won't be shared with fs/subvol trees.
2934 */
2935 return is_fstree(unode->val);
2936 }
2937
2938 int btrfs_qgroup_account_extent(struct btrfs_trans_handle *trans, u64 bytenr,
2939 u64 num_bytes, struct ulist *old_roots,
2940 struct ulist *new_roots)
2941 {
2942 struct btrfs_fs_info *fs_info = trans->fs_info;
2943 LIST_HEAD(qgroups);
2944 u64 seq;
2945 u64 nr_new_roots = 0;
2946 u64 nr_old_roots = 0;
2947 int ret = 0;
2948
2949 /*
2950 * If quotas get disabled meanwhile, the resources need to be freed and
2951 * we can't just exit here.
2952 */
2953 if (!btrfs_qgroup_full_accounting(fs_info) ||
2954 fs_info->qgroup_flags & BTRFS_QGROUP_RUNTIME_FLAG_NO_ACCOUNTING)
2955 goto out_free;
2956
2957 if (new_roots) {
2958 if (!maybe_fs_roots(new_roots))
2959 goto out_free;
2960 nr_new_roots = new_roots->nnodes;
2961 }
2962 if (old_roots) {
2963 if (!maybe_fs_roots(old_roots))
2964 goto out_free;
2965 nr_old_roots = old_roots->nnodes;
2966 }
2967
2968 /* Quick exit, either not fs tree roots, or won't affect any qgroup */
2969 if (nr_old_roots == 0 && nr_new_roots == 0)
2970 goto out_free;
2971
2972 trace_btrfs_qgroup_account_extent(fs_info, trans->transid, bytenr,
2973 num_bytes, nr_old_roots, nr_new_roots);
2974
2975 mutex_lock(&fs_info->qgroup_rescan_lock);
2976 if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN) {
2977 if (fs_info->qgroup_rescan_progress.objectid <= bytenr) {
2978 mutex_unlock(&fs_info->qgroup_rescan_lock);
2979 ret = 0;
2980 goto out_free;
2981 }
2982 }
2983 mutex_unlock(&fs_info->qgroup_rescan_lock);
2984
2985 spin_lock(&fs_info->qgroup_lock);
2986 seq = fs_info->qgroup_seq;
2987
2988 /* Update old refcnts using old_roots */
2989 qgroup_update_refcnt(fs_info, old_roots, &qgroups, seq, UPDATE_OLD);
2990
2991 /* Update new refcnts using new_roots */
2992 qgroup_update_refcnt(fs_info, new_roots, &qgroups, seq, UPDATE_NEW);
2993
2994 qgroup_update_counters(fs_info, &qgroups, nr_old_roots, nr_new_roots,
2995 num_bytes, seq);
2996
2997 /*
2998 * We're done using the iterator, release all its qgroups while holding
2999 * fs_info->qgroup_lock so that we don't race with btrfs_remove_qgroup()
3000 * and trigger use-after-free accesses to qgroups.
3001 */
3002 qgroup_iterator_nested_clean(&qgroups);
3003
3004 /*
3005 * Bump qgroup_seq to avoid seq overlap
3006 */
3007 fs_info->qgroup_seq += max(nr_old_roots, nr_new_roots) + 1;
3008 spin_unlock(&fs_info->qgroup_lock);
3009 out_free:
3010 ulist_free(old_roots);
3011 ulist_free(new_roots);
3012 return ret;
3013 }
3014
3015 int btrfs_qgroup_account_extents(struct btrfs_trans_handle *trans)
3016 {
3017 struct btrfs_fs_info *fs_info = trans->fs_info;
3018 struct btrfs_qgroup_extent_record *record;
3019 struct btrfs_delayed_ref_root *delayed_refs;
3020 struct ulist *new_roots = NULL;
3021 struct rb_node *node;
3022 u64 num_dirty_extents = 0;
3023 u64 qgroup_to_skip;
3024 int ret = 0;
3025
3026 if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_SIMPLE)
3027 return 0;
3028
3029 delayed_refs = &trans->transaction->delayed_refs;
3030 qgroup_to_skip = delayed_refs->qgroup_to_skip;
3031 while ((node = rb_first(&delayed_refs->dirty_extent_root))) {
3032 record = rb_entry(node, struct btrfs_qgroup_extent_record,
3033 node);
3034
3035 num_dirty_extents++;
3036 trace_btrfs_qgroup_account_extents(fs_info, record);
3037
3038 if (!ret && !(fs_info->qgroup_flags &
3039 BTRFS_QGROUP_RUNTIME_FLAG_NO_ACCOUNTING)) {
3040 struct btrfs_backref_walk_ctx ctx = { 0 };
3041
3042 ctx.bytenr = record->bytenr;
3043 ctx.fs_info = fs_info;
3044
3045 /*
3046 * Old roots should be searched when inserting qgroup
3047 * extent record.
3048 *
3049 * But for INCONSISTENT (NO_ACCOUNTING) -> rescan case,
3050 * we may have some record inserted during
3051 * NO_ACCOUNTING (thus no old_roots populated), but
3052 * later we start rescan, which clears NO_ACCOUNTING,
3053 * leaving some inserted records without old_roots
3054 * populated.
3055 *
3056 * Those cases are rare and should not cause too much
3057 * time spent during commit_transaction().
3058 */
3059 if (!record->old_roots) {
3060 /* Search commit root to find old_roots */
3061 ret = btrfs_find_all_roots(&ctx, false);
3062 if (ret < 0)
3063 goto cleanup;
3064 record->old_roots = ctx.roots;
3065 ctx.roots = NULL;
3066 }
3067
3068 /*
3069 * Use BTRFS_SEQ_LAST as time_seq to do special search,
3070 * which doesn't lock tree or delayed_refs and search
3071 * current root. It's safe inside commit_transaction().
3072 */
3073 ctx.trans = trans;
3074 ctx.time_seq = BTRFS_SEQ_LAST;
3075 ret = btrfs_find_all_roots(&ctx, false);
3076 if (ret < 0)
3077 goto cleanup;
3078 new_roots = ctx.roots;
3079 if (qgroup_to_skip) {
3080 ulist_del(new_roots, qgroup_to_skip, 0);
3081 ulist_del(record->old_roots, qgroup_to_skip,
3082 0);
3083 }
3084 ret = btrfs_qgroup_account_extent(trans, record->bytenr,
3085 record->num_bytes,
3086 record->old_roots,
3087 new_roots);
3088 record->old_roots = NULL;
3089 new_roots = NULL;
3090 }
3091 /* Free the reserved data space */
3092 btrfs_qgroup_free_refroot(fs_info,
3093 record->data_rsv_refroot,
3094 record->data_rsv,
3095 BTRFS_QGROUP_RSV_DATA);
3096 cleanup:
3097 ulist_free(record->old_roots);
3098 ulist_free(new_roots);
3099 new_roots = NULL;
3100 rb_erase(node, &delayed_refs->dirty_extent_root);
3101 kfree(record);
3102
3103 }
3104 trace_qgroup_num_dirty_extents(fs_info, trans->transid,
3105 num_dirty_extents);
3106 return ret;
3107 }
3108
3109 /*
3110 * Writes all changed qgroups to disk.
3111 * Called by the transaction commit path and the qgroup assign ioctl.
3112 */
3113 int btrfs_run_qgroups(struct btrfs_trans_handle *trans)
3114 {
3115 struct btrfs_fs_info *fs_info = trans->fs_info;
3116 int ret = 0;
3117
3118 /*
3119 * In case we are called from the qgroup assign ioctl, assert that we
3120 * are holding the qgroup_ioctl_lock, otherwise we can race with a quota
3121 * disable operation (ioctl) and access a freed quota root.
3122 */
3123 if (trans->transaction->state != TRANS_STATE_COMMIT_DOING)
3124 lockdep_assert_held(&fs_info->qgroup_ioctl_lock);
3125
3126 if (!fs_info->quota_root)
3127 return ret;
3128
3129 spin_lock(&fs_info->qgroup_lock);
3130 while (!list_empty(&fs_info->dirty_qgroups)) {
3131 struct btrfs_qgroup *qgroup;
3132 qgroup = list_first_entry(&fs_info->dirty_qgroups,
3133 struct btrfs_qgroup, dirty);
3134 list_del_init(&qgroup->dirty);
3135 spin_unlock(&fs_info->qgroup_lock);
3136 ret = update_qgroup_info_item(trans, qgroup);
3137 if (ret)
3138 qgroup_mark_inconsistent(fs_info);
3139 ret = update_qgroup_limit_item(trans, qgroup);
3140 if (ret)
3141 qgroup_mark_inconsistent(fs_info);
3142 spin_lock(&fs_info->qgroup_lock);
3143 }
3144 if (btrfs_qgroup_enabled(fs_info))
3145 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_ON;
3146 else
3147 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_ON;
3148 spin_unlock(&fs_info->qgroup_lock);
3149
3150 ret = update_qgroup_status_item(trans);
3151 if (ret)
3152 qgroup_mark_inconsistent(fs_info);
3153
3154 return ret;
3155 }
3156
3157 int btrfs_qgroup_check_inherit(struct btrfs_fs_info *fs_info,
3158 struct btrfs_qgroup_inherit *inherit,
3159 size_t size)
3160 {
3161 if (inherit->flags & ~BTRFS_QGROUP_INHERIT_FLAGS_SUPP)
3162 return -EOPNOTSUPP;
3163 if (size < sizeof(*inherit) || size > PAGE_SIZE)
3164 return -EINVAL;
3165
3166 /*
3167 * In the past we allowed btrfs_qgroup_inherit to specify to copy
3168 * rfer/excl numbers directly from other qgroups. This behavior has
3169 * been disabled in userspace for a very long time, but here we should
3170 * also disable it in kernel, as this behavior is known to mark qgroup
3171 * inconsistent, and a rescan would wipe out the changes anyway.
3172 *
3173 * Reject any btrfs_qgroup_inherit with num_ref_copies or num_excl_copies.
3174 */
3175 if (inherit->num_ref_copies > 0 || inherit->num_excl_copies > 0)
3176 return -EINVAL;
3177
3178 if (size != struct_size(inherit, qgroups, inherit->num_qgroups))
3179 return -EINVAL;
3180
3181 /*
3182 * Skip the inherit source qgroups check if qgroup is not enabled.
3183 * Qgroup can still be later enabled causing problems, but in that case
3184 * btrfs_qgroup_inherit() would just ignore those invalid ones.
3185 */
3186 if (!btrfs_qgroup_enabled(fs_info))
3187 return 0;
3188
3189 /*
3190 * Now check all the remaining qgroups, they should all:
3191 *
3192 * - Exist
3193 * - Be higher level qgroups.
3194 */
3195 for (int i = 0; i < inherit->num_qgroups; i++) {
3196 struct btrfs_qgroup *qgroup;
3197 u64 qgroupid = inherit->qgroups[i];
3198
3199 if (btrfs_qgroup_level(qgroupid) == 0)
3200 return -EINVAL;
3201
3202 spin_lock(&fs_info->qgroup_lock);
3203 qgroup = find_qgroup_rb(fs_info, qgroupid);
3204 if (!qgroup) {
3205 spin_unlock(&fs_info->qgroup_lock);
3206 return -ENOENT;
3207 }
3208 spin_unlock(&fs_info->qgroup_lock);
3209 }
3210 return 0;
3211 }
3212
3213 static int qgroup_auto_inherit(struct btrfs_fs_info *fs_info,
3214 u64 inode_rootid,
3215 struct btrfs_qgroup_inherit **inherit)
3216 {
3217 int i = 0;
3218 u64 num_qgroups = 0;
3219 struct btrfs_qgroup *inode_qg;
3220 struct btrfs_qgroup_list *qg_list;
3221 struct btrfs_qgroup_inherit *res;
3222 size_t struct_sz;
3223 u64 *qgids;
3224
3225 if (*inherit)
3226 return -EEXIST;
3227
3228 inode_qg = find_qgroup_rb(fs_info, inode_rootid);
3229 if (!inode_qg)
3230 return -ENOENT;
3231
3232 num_qgroups = list_count_nodes(&inode_qg->groups);
3233
3234 if (!num_qgroups)
3235 return 0;
3236
3237 struct_sz = struct_size(res, qgroups, num_qgroups);
3238 if (struct_sz == SIZE_MAX)
3239 return -ERANGE;
3240
3241 res = kzalloc(struct_sz, GFP_NOFS);
3242 if (!res)
3243 return -ENOMEM;
3244 res->num_qgroups = num_qgroups;
3245 qgids = res->qgroups;
3246
3247 list_for_each_entry(qg_list, &inode_qg->groups, next_group)
3248 qgids[i++] = qg_list->group->qgroupid;
3249
3250 *inherit = res;
3251 return 0;
3252 }
3253
3254 /*
3255 * Check if we can skip rescan when inheriting qgroups. If @src has a single
3256 * @parent, and that @parent is owning all its bytes exclusively, we can skip
3257 * the full rescan, by just adding nodesize to the @parent's excl/rfer.
3258 *
3259 * Return <0 for fatal errors (like srcid/parentid has no qgroup).
3260 * Return 0 if a quick inherit is done.
3261 * Return >0 if a quick inherit is not possible, and a full rescan is needed.
3262 */
3263 static int qgroup_snapshot_quick_inherit(struct btrfs_fs_info *fs_info,
3264 u64 srcid, u64 parentid)
3265 {
3266 struct btrfs_qgroup *src;
3267 struct btrfs_qgroup *parent;
3268 struct btrfs_qgroup_list *list;
3269 int nr_parents = 0;
3270
3271 src = find_qgroup_rb(fs_info, srcid);
3272 if (!src)
3273 return -ENOENT;
3274 parent = find_qgroup_rb(fs_info, parentid);
3275 if (!parent)
3276 return -ENOENT;
3277
3278 /*
3279 * Source has no parent qgroup, but our new qgroup would have one.
3280 * Qgroup numbers would become inconsistent.
3281 */
3282 if (list_empty(&src->groups))
3283 return 1;
3284
3285 list_for_each_entry(list, &src->groups, next_group) {
3286 /* The parent is not the same, quick update is not possible. */
3287 if (list->group->qgroupid != parentid)
3288 return 1;
3289 nr_parents++;
3290 /*
3291 * More than one parent qgroup, we can't be sure about accounting
3292 * consistency.
3293 */
3294 if (nr_parents > 1)
3295 return 1;
3296 }
3297
3298 /*
3299 * The parent is not exclusively owning all its bytes. We're not sure
3300 * if the source has any bytes not fully owned by the parent.
3301 */
3302 if (parent->excl != parent->rfer)
3303 return 1;
3304
3305 parent->excl += fs_info->nodesize;
3306 parent->rfer += fs_info->nodesize;
3307 return 0;
3308 }
3309
3310 /*
3311 * Copy the accounting information between qgroups. This is necessary
3312 * when a snapshot or a subvolume is created. Throwing an error will
3313 * cause a transaction abort so we take extra care here to only error
3314 * when a readonly fs is a reasonable outcome.
3315 */
3316 int btrfs_qgroup_inherit(struct btrfs_trans_handle *trans, u64 srcid,
3317 u64 objectid, u64 inode_rootid,
3318 struct btrfs_qgroup_inherit *inherit)
3319 {
3320 int ret = 0;
3321 u64 *i_qgroups;
3322 bool committing = false;
3323 struct btrfs_fs_info *fs_info = trans->fs_info;
3324 struct btrfs_root *quota_root;
3325 struct btrfs_qgroup *srcgroup;
3326 struct btrfs_qgroup *dstgroup;
3327 struct btrfs_qgroup *prealloc;
3328 struct btrfs_qgroup_list **qlist_prealloc = NULL;
3329 bool free_inherit = false;
3330 bool need_rescan = false;
3331 u32 level_size = 0;
3332 u64 nums;
3333
3334 prealloc = kzalloc(sizeof(*prealloc), GFP_NOFS);
3335 if (!prealloc)
3336 return -ENOMEM;
3337
3338 /*
3339 * There are only two callers of this function.
3340 *
3341 * One in create_subvol() in the ioctl context, which needs to hold
3342 * the qgroup_ioctl_lock.
3343 *
3344 * The other one in create_pending_snapshot() where no other qgroup
3345 * code can modify the fs as they all need to either start a new trans
3346 * or hold a trans handler, thus we don't need to hold
3347 * qgroup_ioctl_lock.
3348 * This would avoid long and complex lock chain and make lockdep happy.
3349 */
3350 spin_lock(&fs_info->trans_lock);
3351 if (trans->transaction->state == TRANS_STATE_COMMIT_DOING)
3352 committing = true;
3353 spin_unlock(&fs_info->trans_lock);
3354
3355 if (!committing)
3356 mutex_lock(&fs_info->qgroup_ioctl_lock);
3357 if (!btrfs_qgroup_enabled(fs_info))
3358 goto out;
3359
3360 quota_root = fs_info->quota_root;
3361 if (!quota_root) {
3362 ret = -EINVAL;
3363 goto out;
3364 }
3365
3366 if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_SIMPLE && !inherit) {
3367 ret = qgroup_auto_inherit(fs_info, inode_rootid, &inherit);
3368 if (ret)
3369 goto out;
3370 free_inherit = true;
3371 }
3372
3373 if (inherit) {
3374 i_qgroups = (u64 *)(inherit + 1);
3375 nums = inherit->num_qgroups + 2 * inherit->num_ref_copies +
3376 2 * inherit->num_excl_copies;
3377 for (int i = 0; i < nums; i++) {
3378 srcgroup = find_qgroup_rb(fs_info, *i_qgroups);
3379
3380 /*
3381 * Zero out invalid groups so we can ignore
3382 * them later.
3383 */
3384 if (!srcgroup ||
3385 ((srcgroup->qgroupid >> 48) <= (objectid >> 48)))
3386 *i_qgroups = 0ULL;
3387
3388 ++i_qgroups;
3389 }
3390 }
3391
3392 /*
3393 * create a tracking group for the subvol itself
3394 */
3395 ret = add_qgroup_item(trans, quota_root, objectid);
3396 if (ret)
3397 goto out;
3398
3399 /*
3400 * add qgroup to all inherited groups
3401 */
3402 if (inherit) {
3403 i_qgroups = (u64 *)(inherit + 1);
3404 for (int i = 0; i < inherit->num_qgroups; i++, i_qgroups++) {
3405 if (*i_qgroups == 0)
3406 continue;
3407 ret = add_qgroup_relation_item(trans, objectid,
3408 *i_qgroups);
3409 if (ret && ret != -EEXIST)
3410 goto out;
3411 ret = add_qgroup_relation_item(trans, *i_qgroups,
3412 objectid);
3413 if (ret && ret != -EEXIST)
3414 goto out;
3415 }
3416 ret = 0;
3417
3418 qlist_prealloc = kcalloc(inherit->num_qgroups,
3419 sizeof(struct btrfs_qgroup_list *),
3420 GFP_NOFS);
3421 if (!qlist_prealloc) {
3422 ret = -ENOMEM;
3423 goto out;
3424 }
3425 for (int i = 0; i < inherit->num_qgroups; i++) {
3426 qlist_prealloc[i] = kzalloc(sizeof(struct btrfs_qgroup_list),
3427 GFP_NOFS);
3428 if (!qlist_prealloc[i]) {
3429 ret = -ENOMEM;
3430 goto out;
3431 }
3432 }
3433 }
3434
3435 spin_lock(&fs_info->qgroup_lock);
3436
3437 dstgroup = add_qgroup_rb(fs_info, prealloc, objectid);
3438 prealloc = NULL;
3439
3440 if (inherit && inherit->flags & BTRFS_QGROUP_INHERIT_SET_LIMITS) {
3441 dstgroup->lim_flags = inherit->lim.flags;
3442 dstgroup->max_rfer = inherit->lim.max_rfer;
3443 dstgroup->max_excl = inherit->lim.max_excl;
3444 dstgroup->rsv_rfer = inherit->lim.rsv_rfer;
3445 dstgroup->rsv_excl = inherit->lim.rsv_excl;
3446
3447 qgroup_dirty(fs_info, dstgroup);
3448 }
3449
3450 if (srcid && btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_FULL) {
3451 srcgroup = find_qgroup_rb(fs_info, srcid);
3452 if (!srcgroup)
3453 goto unlock;
3454
3455 /*
3456 * We call inherit after we clone the root in order to make sure
3457 * our counts don't go crazy, so at this point the only
3458 * difference between the two roots should be the root node.
3459 */
3460 level_size = fs_info->nodesize;
3461 dstgroup->rfer = srcgroup->rfer;
3462 dstgroup->rfer_cmpr = srcgroup->rfer_cmpr;
3463 dstgroup->excl = level_size;
3464 dstgroup->excl_cmpr = level_size;
3465 srcgroup->excl = level_size;
3466 srcgroup->excl_cmpr = level_size;
3467
3468 /* inherit the limit info */
3469 dstgroup->lim_flags = srcgroup->lim_flags;
3470 dstgroup->max_rfer = srcgroup->max_rfer;
3471 dstgroup->max_excl = srcgroup->max_excl;
3472 dstgroup->rsv_rfer = srcgroup->rsv_rfer;
3473 dstgroup->rsv_excl = srcgroup->rsv_excl;
3474
3475 qgroup_dirty(fs_info, dstgroup);
3476 qgroup_dirty(fs_info, srcgroup);
3477
3478 /*
3479 * If the source qgroup has parent but the new one doesn't,
3480 * we need a full rescan.
3481 */
3482 if (!inherit && !list_empty(&srcgroup->groups))
3483 need_rescan = true;
3484 }
3485
3486 if (!inherit)
3487 goto unlock;
3488
3489 i_qgroups = (u64 *)(inherit + 1);
3490 for (int i = 0; i < inherit->num_qgroups; i++) {
3491 if (*i_qgroups) {
3492 ret = add_relation_rb(fs_info, qlist_prealloc[i], objectid,
3493 *i_qgroups);
3494 qlist_prealloc[i] = NULL;
3495 if (ret)
3496 goto unlock;
3497 }
3498 if (srcid) {
3499 /* Check if we can do a quick inherit. */
3500 ret = qgroup_snapshot_quick_inherit(fs_info, srcid, *i_qgroups);
3501 if (ret < 0)
3502 goto unlock;
3503 if (ret > 0)
3504 need_rescan = true;
3505 ret = 0;
3506 }
3507 ++i_qgroups;
3508 }
3509
3510 for (int i = 0; i < inherit->num_ref_copies; i++, i_qgroups += 2) {
3511 struct btrfs_qgroup *src;
3512 struct btrfs_qgroup *dst;
3513
3514 if (!i_qgroups[0] || !i_qgroups[1])
3515 continue;
3516
3517 src = find_qgroup_rb(fs_info, i_qgroups[0]);
3518 dst = find_qgroup_rb(fs_info, i_qgroups[1]);
3519
3520 if (!src || !dst) {
3521 ret = -EINVAL;
3522 goto unlock;
3523 }
3524
3525 dst->rfer = src->rfer - level_size;
3526 dst->rfer_cmpr = src->rfer_cmpr - level_size;
3527
3528 /* Manually tweaking numbers certainly needs a rescan */
3529 need_rescan = true;
3530 }
3531 for (int i = 0; i < inherit->num_excl_copies; i++, i_qgroups += 2) {
3532 struct btrfs_qgroup *src;
3533 struct btrfs_qgroup *dst;
3534
3535 if (!i_qgroups[0] || !i_qgroups[1])
3536 continue;
3537
3538 src = find_qgroup_rb(fs_info, i_qgroups[0]);
3539 dst = find_qgroup_rb(fs_info, i_qgroups[1]);
3540
3541 if (!src || !dst) {
3542 ret = -EINVAL;
3543 goto unlock;
3544 }
3545
3546 dst->excl = src->excl + level_size;
3547 dst->excl_cmpr = src->excl_cmpr + level_size;
3548 need_rescan = true;
3549 }
3550
3551 unlock:
3552 spin_unlock(&fs_info->qgroup_lock);
3553 if (!ret)
3554 ret = btrfs_sysfs_add_one_qgroup(fs_info, dstgroup);
3555 out:
3556 if (!committing)
3557 mutex_unlock(&fs_info->qgroup_ioctl_lock);
3558 if (need_rescan)
3559 qgroup_mark_inconsistent(fs_info);
3560 if (qlist_prealloc) {
3561 for (int i = 0; i < inherit->num_qgroups; i++)
3562 kfree(qlist_prealloc[i]);
3563 kfree(qlist_prealloc);
3564 }
3565 if (free_inherit)
3566 kfree(inherit);
3567 kfree(prealloc);
3568 return ret;
3569 }
3570
3571 static bool qgroup_check_limits(const struct btrfs_qgroup *qg, u64 num_bytes)
3572 {
3573 if ((qg->lim_flags & BTRFS_QGROUP_LIMIT_MAX_RFER) &&
3574 qgroup_rsv_total(qg) + (s64)qg->rfer + num_bytes > qg->max_rfer)
3575 return false;
3576
3577 if ((qg->lim_flags & BTRFS_QGROUP_LIMIT_MAX_EXCL) &&
3578 qgroup_rsv_total(qg) + (s64)qg->excl + num_bytes > qg->max_excl)
3579 return false;
3580
3581 return true;
3582 }
3583
3584 static int qgroup_reserve(struct btrfs_root *root, u64 num_bytes, bool enforce,
3585 enum btrfs_qgroup_rsv_type type)
3586 {
3587 struct btrfs_qgroup *qgroup;
3588 struct btrfs_fs_info *fs_info = root->fs_info;
3589 u64 ref_root = btrfs_root_id(root);
3590 int ret = 0;
3591 LIST_HEAD(qgroup_list);
3592
3593 if (!is_fstree(ref_root))
3594 return 0;
3595
3596 if (num_bytes == 0)
3597 return 0;
3598
3599 if (test_bit(BTRFS_FS_QUOTA_OVERRIDE, &fs_info->flags) &&
3600 capable(CAP_SYS_RESOURCE))
3601 enforce = false;
3602
3603 spin_lock(&fs_info->qgroup_lock);
3604 if (!fs_info->quota_root)
3605 goto out;
3606
3607 qgroup = find_qgroup_rb(fs_info, ref_root);
3608 if (!qgroup)
3609 goto out;
3610
3611 qgroup_iterator_add(&qgroup_list, qgroup);
3612 list_for_each_entry(qgroup, &qgroup_list, iterator) {
3613 struct btrfs_qgroup_list *glist;
3614
3615 if (enforce && !qgroup_check_limits(qgroup, num_bytes)) {
3616 ret = -EDQUOT;
3617 goto out;
3618 }
3619
3620 list_for_each_entry(glist, &qgroup->groups, next_group)
3621 qgroup_iterator_add(&qgroup_list, glist->group);
3622 }
3623
3624 ret = 0;
3625 /*
3626 * no limits exceeded, now record the reservation into all qgroups
3627 */
3628 list_for_each_entry(qgroup, &qgroup_list, iterator)
3629 qgroup_rsv_add(fs_info, qgroup, num_bytes, type);
3630
3631 out:
3632 qgroup_iterator_clean(&qgroup_list);
3633 spin_unlock(&fs_info->qgroup_lock);
3634 return ret;
3635 }
3636
3637 /*
3638 * Free @num_bytes of reserved space with @type for qgroup. (Normally level 0
3639 * qgroup).
3640 *
3641 * Will handle all higher level qgroup too.
3642 *
3643 * NOTE: If @num_bytes is (u64)-1, this means to free all bytes of this qgroup.
3644 * This special case is only used for META_PERTRANS type.
3645 */
3646 void btrfs_qgroup_free_refroot(struct btrfs_fs_info *fs_info,
3647 u64 ref_root, u64 num_bytes,
3648 enum btrfs_qgroup_rsv_type type)
3649 {
3650 struct btrfs_qgroup *qgroup;
3651 LIST_HEAD(qgroup_list);
3652
3653 if (!is_fstree(ref_root))
3654 return;
3655
3656 if (num_bytes == 0)
3657 return;
3658
3659 if (num_bytes == (u64)-1 && type != BTRFS_QGROUP_RSV_META_PERTRANS) {
3660 WARN(1, "%s: Invalid type to free", __func__);
3661 return;
3662 }
3663 spin_lock(&fs_info->qgroup_lock);
3664
3665 if (!fs_info->quota_root)
3666 goto out;
3667
3668 qgroup = find_qgroup_rb(fs_info, ref_root);
3669 if (!qgroup)
3670 goto out;
3671
3672 if (num_bytes == (u64)-1)
3673 /*
3674 * We're freeing all pertrans rsv, get reserved value from
3675 * level 0 qgroup as real num_bytes to free.
3676 */
3677 num_bytes = qgroup->rsv.values[type];
3678
3679 qgroup_iterator_add(&qgroup_list, qgroup);
3680 list_for_each_entry(qgroup, &qgroup_list, iterator) {
3681 struct btrfs_qgroup_list *glist;
3682
3683 qgroup_rsv_release(fs_info, qgroup, num_bytes, type);
3684 list_for_each_entry(glist, &qgroup->groups, next_group) {
3685 qgroup_iterator_add(&qgroup_list, glist->group);
3686 }
3687 }
3688 out:
3689 qgroup_iterator_clean(&qgroup_list);
3690 spin_unlock(&fs_info->qgroup_lock);
3691 }
3692
3693 /*
3694 * Check if the leaf is the last leaf. Which means all node pointers
3695 * are at their last position.
3696 */
3697 static bool is_last_leaf(struct btrfs_path *path)
3698 {
3699 int i;
3700
3701 for (i = 1; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
3702 if (path->slots[i] != btrfs_header_nritems(path->nodes[i]) - 1)
3703 return false;
3704 }
3705 return true;
3706 }
3707
3708 /*
3709 * returns < 0 on error, 0 when more leafs are to be scanned.
3710 * returns 1 when done.
3711 */
3712 static int qgroup_rescan_leaf(struct btrfs_trans_handle *trans,
3713 struct btrfs_path *path)
3714 {
3715 struct btrfs_fs_info *fs_info = trans->fs_info;
3716 struct btrfs_root *extent_root;
3717 struct btrfs_key found;
3718 struct extent_buffer *scratch_leaf = NULL;
3719 u64 num_bytes;
3720 bool done;
3721 int slot;
3722 int ret;
3723
3724 if (!btrfs_qgroup_full_accounting(fs_info))
3725 return 1;
3726
3727 mutex_lock(&fs_info->qgroup_rescan_lock);
3728 extent_root = btrfs_extent_root(fs_info,
3729 fs_info->qgroup_rescan_progress.objectid);
3730 ret = btrfs_search_slot_for_read(extent_root,
3731 &fs_info->qgroup_rescan_progress,
3732 path, 1, 0);
3733
3734 btrfs_debug(fs_info,
3735 "current progress key (%llu %u %llu), search_slot ret %d",
3736 fs_info->qgroup_rescan_progress.objectid,
3737 fs_info->qgroup_rescan_progress.type,
3738 fs_info->qgroup_rescan_progress.offset, ret);
3739
3740 if (ret) {
3741 /*
3742 * The rescan is about to end, we will not be scanning any
3743 * further blocks. We cannot unset the RESCAN flag here, because
3744 * we want to commit the transaction if everything went well.
3745 * To make the live accounting work in this phase, we set our
3746 * scan progress pointer such that every real extent objectid
3747 * will be smaller.
3748 */
3749 fs_info->qgroup_rescan_progress.objectid = (u64)-1;
3750 btrfs_release_path(path);
3751 mutex_unlock(&fs_info->qgroup_rescan_lock);
3752 return ret;
3753 }
3754 done = is_last_leaf(path);
3755
3756 btrfs_item_key_to_cpu(path->nodes[0], &found,
3757 btrfs_header_nritems(path->nodes[0]) - 1);
3758 fs_info->qgroup_rescan_progress.objectid = found.objectid + 1;
3759
3760 scratch_leaf = btrfs_clone_extent_buffer(path->nodes[0]);
3761 if (!scratch_leaf) {
3762 ret = -ENOMEM;
3763 mutex_unlock(&fs_info->qgroup_rescan_lock);
3764 goto out;
3765 }
3766 slot = path->slots[0];
3767 btrfs_release_path(path);
3768 mutex_unlock(&fs_info->qgroup_rescan_lock);
3769
3770 for (; slot < btrfs_header_nritems(scratch_leaf); ++slot) {
3771 struct btrfs_backref_walk_ctx ctx = { 0 };
3772
3773 btrfs_item_key_to_cpu(scratch_leaf, &found, slot);
3774 if (found.type != BTRFS_EXTENT_ITEM_KEY &&
3775 found.type != BTRFS_METADATA_ITEM_KEY)
3776 continue;
3777 if (found.type == BTRFS_METADATA_ITEM_KEY)
3778 num_bytes = fs_info->nodesize;
3779 else
3780 num_bytes = found.offset;
3781
3782 ctx.bytenr = found.objectid;
3783 ctx.fs_info = fs_info;
3784
3785 ret = btrfs_find_all_roots(&ctx, false);
3786 if (ret < 0)
3787 goto out;
3788 /* For rescan, just pass old_roots as NULL */
3789 ret = btrfs_qgroup_account_extent(trans, found.objectid,
3790 num_bytes, NULL, ctx.roots);
3791 if (ret < 0)
3792 goto out;
3793 }
3794 out:
3795 if (scratch_leaf)
3796 free_extent_buffer(scratch_leaf);
3797
3798 if (done && !ret) {
3799 ret = 1;
3800 fs_info->qgroup_rescan_progress.objectid = (u64)-1;
3801 }
3802 return ret;
3803 }
3804
3805 static bool rescan_should_stop(struct btrfs_fs_info *fs_info)
3806 {
3807 if (btrfs_fs_closing(fs_info))
3808 return true;
3809 if (test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state))
3810 return true;
3811 if (!btrfs_qgroup_enabled(fs_info))
3812 return true;
3813 if (fs_info->qgroup_flags & BTRFS_QGROUP_RUNTIME_FLAG_CANCEL_RESCAN)
3814 return true;
3815 return false;
3816 }
3817
3818 static void btrfs_qgroup_rescan_worker(struct btrfs_work *work)
3819 {
3820 struct btrfs_fs_info *fs_info = container_of(work, struct btrfs_fs_info,
3821 qgroup_rescan_work);
3822 struct btrfs_path *path;
3823 struct btrfs_trans_handle *trans = NULL;
3824 int ret = 0;
3825 bool stopped = false;
3826 bool did_leaf_rescans = false;
3827
3828 if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_SIMPLE)
3829 return;
3830
3831 path = btrfs_alloc_path();
3832 if (!path) {
3833 ret = -ENOMEM;
3834 goto out;
3835 }
3836 /*
3837 * Rescan should only search for commit root, and any later difference
3838 * should be recorded by qgroup
3839 */
3840 path->search_commit_root = 1;
3841 path->skip_locking = 1;
3842
3843 while (!ret && !(stopped = rescan_should_stop(fs_info))) {
3844 trans = btrfs_start_transaction(fs_info->fs_root, 0);
3845 if (IS_ERR(trans)) {
3846 ret = PTR_ERR(trans);
3847 break;
3848 }
3849
3850 ret = qgroup_rescan_leaf(trans, path);
3851 did_leaf_rescans = true;
3852
3853 if (ret > 0)
3854 btrfs_commit_transaction(trans);
3855 else
3856 btrfs_end_transaction(trans);
3857 }
3858
3859 out:
3860 btrfs_free_path(path);
3861
3862 mutex_lock(&fs_info->qgroup_rescan_lock);
3863 if (ret > 0 &&
3864 fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT) {
3865 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
3866 } else if (ret < 0 || stopped) {
3867 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
3868 }
3869 mutex_unlock(&fs_info->qgroup_rescan_lock);
3870
3871 /*
3872 * Only update status, since the previous part has already updated the
3873 * qgroup info, and only if we did any actual work. This also prevents
3874 * race with a concurrent quota disable, which has already set
3875 * fs_info->quota_root to NULL and cleared BTRFS_FS_QUOTA_ENABLED at
3876 * btrfs_quota_disable().
3877 */
3878 if (did_leaf_rescans) {
3879 trans = btrfs_start_transaction(fs_info->quota_root, 1);
3880 if (IS_ERR(trans)) {
3881 ret = PTR_ERR(trans);
3882 trans = NULL;
3883 btrfs_err(fs_info,
3884 "fail to start transaction for status update: %d",
3885 ret);
3886 }
3887 } else {
3888 trans = NULL;
3889 }
3890
3891 mutex_lock(&fs_info->qgroup_rescan_lock);
3892 if (!stopped ||
3893 fs_info->qgroup_flags & BTRFS_QGROUP_RUNTIME_FLAG_CANCEL_RESCAN)
3894 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
3895 if (trans) {
3896 int ret2 = update_qgroup_status_item(trans);
3897
3898 if (ret2 < 0) {
3899 ret = ret2;
3900 btrfs_err(fs_info, "fail to update qgroup status: %d", ret);
3901 }
3902 }
3903 fs_info->qgroup_rescan_running = false;
3904 fs_info->qgroup_flags &= ~BTRFS_QGROUP_RUNTIME_FLAG_CANCEL_RESCAN;
3905 complete_all(&fs_info->qgroup_rescan_completion);
3906 mutex_unlock(&fs_info->qgroup_rescan_lock);
3907
3908 if (!trans)
3909 return;
3910
3911 btrfs_end_transaction(trans);
3912
3913 if (stopped) {
3914 btrfs_info(fs_info, "qgroup scan paused");
3915 } else if (fs_info->qgroup_flags & BTRFS_QGROUP_RUNTIME_FLAG_CANCEL_RESCAN) {
3916 btrfs_info(fs_info, "qgroup scan cancelled");
3917 } else if (ret >= 0) {
3918 btrfs_info(fs_info, "qgroup scan completed%s",
3919 ret > 0 ? " (inconsistency flag cleared)" : "");
3920 } else {
3921 btrfs_err(fs_info, "qgroup scan failed with %d", ret);
3922 }
3923 }
3924
3925 /*
3926 * Checks that (a) no rescan is running and (b) quota is enabled. Allocates all
3927 * memory required for the rescan context.
3928 */
3929 static int
3930 qgroup_rescan_init(struct btrfs_fs_info *fs_info, u64 progress_objectid,
3931 int init_flags)
3932 {
3933 int ret = 0;
3934
3935 if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_SIMPLE) {
3936 btrfs_warn(fs_info, "qgroup rescan init failed, running in simple mode");
3937 return -EINVAL;
3938 }
3939
3940 if (!init_flags) {
3941 /* we're resuming qgroup rescan at mount time */
3942 if (!(fs_info->qgroup_flags &
3943 BTRFS_QGROUP_STATUS_FLAG_RESCAN)) {
3944 btrfs_debug(fs_info,
3945 "qgroup rescan init failed, qgroup rescan is not queued");
3946 ret = -EINVAL;
3947 } else if (!(fs_info->qgroup_flags &
3948 BTRFS_QGROUP_STATUS_FLAG_ON)) {
3949 btrfs_debug(fs_info,
3950 "qgroup rescan init failed, qgroup is not enabled");
3951 ret = -ENOTCONN;
3952 }
3953
3954 if (ret)
3955 return ret;
3956 }
3957
3958 mutex_lock(&fs_info->qgroup_rescan_lock);
3959
3960 if (init_flags) {
3961 if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN) {
3962 ret = -EINPROGRESS;
3963 } else if (!(fs_info->qgroup_flags &
3964 BTRFS_QGROUP_STATUS_FLAG_ON)) {
3965 btrfs_debug(fs_info,
3966 "qgroup rescan init failed, qgroup is not enabled");
3967 ret = -ENOTCONN;
3968 } else if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_DISABLED) {
3969 /* Quota disable is in progress */
3970 ret = -EBUSY;
3971 }
3972
3973 if (ret) {
3974 mutex_unlock(&fs_info->qgroup_rescan_lock);
3975 return ret;
3976 }
3977 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_RESCAN;
3978 }
3979
3980 memset(&fs_info->qgroup_rescan_progress, 0,
3981 sizeof(fs_info->qgroup_rescan_progress));
3982 fs_info->qgroup_flags &= ~(BTRFS_QGROUP_RUNTIME_FLAG_CANCEL_RESCAN |
3983 BTRFS_QGROUP_RUNTIME_FLAG_NO_ACCOUNTING);
3984 fs_info->qgroup_rescan_progress.objectid = progress_objectid;
3985 init_completion(&fs_info->qgroup_rescan_completion);
3986 mutex_unlock(&fs_info->qgroup_rescan_lock);
3987
3988 btrfs_init_work(&fs_info->qgroup_rescan_work,
3989 btrfs_qgroup_rescan_worker, NULL);
3990 return 0;
3991 }
3992
3993 static void
3994 qgroup_rescan_zero_tracking(struct btrfs_fs_info *fs_info)
3995 {
3996 struct rb_node *n;
3997 struct btrfs_qgroup *qgroup;
3998
3999 spin_lock(&fs_info->qgroup_lock);
4000 /* clear all current qgroup tracking information */
4001 for (n = rb_first(&fs_info->qgroup_tree); n; n = rb_next(n)) {
4002 qgroup = rb_entry(n, struct btrfs_qgroup, node);
4003 qgroup->rfer = 0;
4004 qgroup->rfer_cmpr = 0;
4005 qgroup->excl = 0;
4006 qgroup->excl_cmpr = 0;
4007 qgroup_dirty(fs_info, qgroup);
4008 }
4009 spin_unlock(&fs_info->qgroup_lock);
4010 }
4011
4012 int
4013 btrfs_qgroup_rescan(struct btrfs_fs_info *fs_info)
4014 {
4015 int ret = 0;
4016
4017 ret = qgroup_rescan_init(fs_info, 0, 1);
4018 if (ret)
4019 return ret;
4020
4021 /*
4022 * We have set the rescan_progress to 0, which means no more
4023 * delayed refs will be accounted by btrfs_qgroup_account_ref.
4024 * However, btrfs_qgroup_account_ref may be right after its call
4025 * to btrfs_find_all_roots, in which case it would still do the
4026 * accounting.
4027 * To solve this, we're committing the transaction, which will
4028 * ensure we run all delayed refs and only after that, we are
4029 * going to clear all tracking information for a clean start.
4030 */
4031
4032 ret = btrfs_commit_current_transaction(fs_info->fs_root);
4033 if (ret) {
4034 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
4035 return ret;
4036 }
4037
4038 qgroup_rescan_zero_tracking(fs_info);
4039
4040 mutex_lock(&fs_info->qgroup_rescan_lock);
4041 fs_info->qgroup_rescan_running = true;
4042 btrfs_queue_work(fs_info->qgroup_rescan_workers,
4043 &fs_info->qgroup_rescan_work);
4044 mutex_unlock(&fs_info->qgroup_rescan_lock);
4045
4046 return 0;
4047 }
4048
4049 int btrfs_qgroup_wait_for_completion(struct btrfs_fs_info *fs_info,
4050 bool interruptible)
4051 {
4052 int running;
4053 int ret = 0;
4054
4055 mutex_lock(&fs_info->qgroup_rescan_lock);
4056 running = fs_info->qgroup_rescan_running;
4057 mutex_unlock(&fs_info->qgroup_rescan_lock);
4058
4059 if (!running)
4060 return 0;
4061
4062 if (interruptible)
4063 ret = wait_for_completion_interruptible(
4064 &fs_info->qgroup_rescan_completion);
4065 else
4066 wait_for_completion(&fs_info->qgroup_rescan_completion);
4067
4068 return ret;
4069 }
4070
4071 /*
4072 * this is only called from open_ctree where we're still single threaded, thus
4073 * locking is omitted here.
4074 */
4075 void
4076 btrfs_qgroup_rescan_resume(struct btrfs_fs_info *fs_info)
4077 {
4078 if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN) {
4079 mutex_lock(&fs_info->qgroup_rescan_lock);
4080 fs_info->qgroup_rescan_running = true;
4081 btrfs_queue_work(fs_info->qgroup_rescan_workers,
4082 &fs_info->qgroup_rescan_work);
4083 mutex_unlock(&fs_info->qgroup_rescan_lock);
4084 }
4085 }
4086
4087 #define rbtree_iterate_from_safe(node, next, start) \
4088 for (node = start; node && ({ next = rb_next(node); 1;}); node = next)
4089
4090 static int qgroup_unreserve_range(struct btrfs_inode *inode,
4091 struct extent_changeset *reserved, u64 start,
4092 u64 len)
4093 {
4094 struct rb_node *node;
4095 struct rb_node *next;
4096 struct ulist_node *entry;
4097 int ret = 0;
4098
4099 node = reserved->range_changed.root.rb_node;
4100 if (!node)
4101 return 0;
4102 while (node) {
4103 entry = rb_entry(node, struct ulist_node, rb_node);
4104 if (entry->val < start)
4105 node = node->rb_right;
4106 else
4107 node = node->rb_left;
4108 }
4109
4110 if (entry->val > start && rb_prev(&entry->rb_node))
4111 entry = rb_entry(rb_prev(&entry->rb_node), struct ulist_node,
4112 rb_node);
4113
4114 rbtree_iterate_from_safe(node, next, &entry->rb_node) {
4115 u64 entry_start;
4116 u64 entry_end;
4117 u64 entry_len;
4118 int clear_ret;
4119
4120 entry = rb_entry(node, struct ulist_node, rb_node);
4121 entry_start = entry->val;
4122 entry_end = entry->aux;
4123 entry_len = entry_end - entry_start + 1;
4124
4125 if (entry_start >= start + len)
4126 break;
4127 if (entry_start + entry_len <= start)
4128 continue;
4129 /*
4130 * Now the entry is in [start, start + len), revert the
4131 * EXTENT_QGROUP_RESERVED bit.
4132 */
4133 clear_ret = clear_extent_bits(&inode->io_tree, entry_start,
4134 entry_end, EXTENT_QGROUP_RESERVED);
4135 if (!ret && clear_ret < 0)
4136 ret = clear_ret;
4137
4138 ulist_del(&reserved->range_changed, entry->val, entry->aux);
4139 if (likely(reserved->bytes_changed >= entry_len)) {
4140 reserved->bytes_changed -= entry_len;
4141 } else {
4142 WARN_ON(1);
4143 reserved->bytes_changed = 0;
4144 }
4145 }
4146
4147 return ret;
4148 }
4149
4150 /*
4151 * Try to free some space for qgroup.
4152 *
4153 * For qgroup, there are only 3 ways to free qgroup space:
4154 * - Flush nodatacow write
4155 * Any nodatacow write will free its reserved data space at run_delalloc_range().
4156 * In theory, we should only flush nodatacow inodes, but it's not yet
4157 * possible, so we need to flush the whole root.
4158 *
4159 * - Wait for ordered extents
4160 * When ordered extents are finished, their reserved metadata is finally
4161 * converted to per_trans status, which can be freed by later commit
4162 * transaction.
4163 *
4164 * - Commit transaction
4165 * This would free the meta_per_trans space.
4166 * In theory this shouldn't provide much space, but any more qgroup space
4167 * is needed.
4168 */
4169 static int try_flush_qgroup(struct btrfs_root *root)
4170 {
4171 int ret;
4172
4173 /* Can't hold an open transaction or we run the risk of deadlocking. */
4174 ASSERT(current->journal_info == NULL);
4175 if (WARN_ON(current->journal_info))
4176 return 0;
4177
4178 /*
4179 * We don't want to run flush again and again, so if there is a running
4180 * one, we won't try to start a new flush, but exit directly.
4181 */
4182 if (test_and_set_bit(BTRFS_ROOT_QGROUP_FLUSHING, &root->state)) {
4183 wait_event(root->qgroup_flush_wait,
4184 !test_bit(BTRFS_ROOT_QGROUP_FLUSHING, &root->state));
4185 return 0;
4186 }
4187
4188 btrfs_run_delayed_iputs(root->fs_info);
4189 btrfs_wait_on_delayed_iputs(root->fs_info);
4190 ret = btrfs_start_delalloc_snapshot(root, true);
4191 if (ret < 0)
4192 goto out;
4193 btrfs_wait_ordered_extents(root, U64_MAX, NULL);
4194
4195 ret = btrfs_commit_current_transaction(root);
4196 out:
4197 clear_bit(BTRFS_ROOT_QGROUP_FLUSHING, &root->state);
4198 wake_up(&root->qgroup_flush_wait);
4199 return ret;
4200 }
4201
4202 static int qgroup_reserve_data(struct btrfs_inode *inode,
4203 struct extent_changeset **reserved_ret, u64 start,
4204 u64 len)
4205 {
4206 struct btrfs_root *root = inode->root;
4207 struct extent_changeset *reserved;
4208 bool new_reserved = false;
4209 u64 orig_reserved;
4210 u64 to_reserve;
4211 int ret;
4212
4213 if (btrfs_qgroup_mode(root->fs_info) == BTRFS_QGROUP_MODE_DISABLED ||
4214 !is_fstree(btrfs_root_id(root)) || len == 0)
4215 return 0;
4216
4217 /* @reserved parameter is mandatory for qgroup */
4218 if (WARN_ON(!reserved_ret))
4219 return -EINVAL;
4220 if (!*reserved_ret) {
4221 new_reserved = true;
4222 *reserved_ret = extent_changeset_alloc();
4223 if (!*reserved_ret)
4224 return -ENOMEM;
4225 }
4226 reserved = *reserved_ret;
4227 /* Record already reserved space */
4228 orig_reserved = reserved->bytes_changed;
4229 ret = set_record_extent_bits(&inode->io_tree, start,
4230 start + len -1, EXTENT_QGROUP_RESERVED, reserved);
4231
4232 /* Newly reserved space */
4233 to_reserve = reserved->bytes_changed - orig_reserved;
4234 trace_btrfs_qgroup_reserve_data(&inode->vfs_inode, start, len,
4235 to_reserve, QGROUP_RESERVE);
4236 if (ret < 0)
4237 goto out;
4238 ret = qgroup_reserve(root, to_reserve, true, BTRFS_QGROUP_RSV_DATA);
4239 if (ret < 0)
4240 goto cleanup;
4241
4242 return ret;
4243
4244 cleanup:
4245 qgroup_unreserve_range(inode, reserved, start, len);
4246 out:
4247 if (new_reserved) {
4248 extent_changeset_free(reserved);
4249 *reserved_ret = NULL;
4250 }
4251 return ret;
4252 }
4253
4254 /*
4255 * Reserve qgroup space for range [start, start + len).
4256 *
4257 * This function will either reserve space from related qgroups or do nothing
4258 * if the range is already reserved.
4259 *
4260 * Return 0 for successful reservation
4261 * Return <0 for error (including -EQUOT)
4262 *
4263 * NOTE: This function may sleep for memory allocation, dirty page flushing and
4264 * commit transaction. So caller should not hold any dirty page locked.
4265 */
4266 int btrfs_qgroup_reserve_data(struct btrfs_inode *inode,
4267 struct extent_changeset **reserved_ret, u64 start,
4268 u64 len)
4269 {
4270 int ret;
4271
4272 ret = qgroup_reserve_data(inode, reserved_ret, start, len);
4273 if (ret <= 0 && ret != -EDQUOT)
4274 return ret;
4275
4276 ret = try_flush_qgroup(inode->root);
4277 if (ret < 0)
4278 return ret;
4279 return qgroup_reserve_data(inode, reserved_ret, start, len);
4280 }
4281
4282 /* Free ranges specified by @reserved, normally in error path */
4283 static int qgroup_free_reserved_data(struct btrfs_inode *inode,
4284 struct extent_changeset *reserved,
4285 u64 start, u64 len, u64 *freed_ret)
4286 {
4287 struct btrfs_root *root = inode->root;
4288 struct ulist_node *unode;
4289 struct ulist_iterator uiter;
4290 struct extent_changeset changeset;
4291 u64 freed = 0;
4292 int ret;
4293
4294 extent_changeset_init(&changeset);
4295 len = round_up(start + len, root->fs_info->sectorsize);
4296 start = round_down(start, root->fs_info->sectorsize);
4297
4298 ULIST_ITER_INIT(&uiter);
4299 while ((unode = ulist_next(&reserved->range_changed, &uiter))) {
4300 u64 range_start = unode->val;
4301 /* unode->aux is the inclusive end */
4302 u64 range_len = unode->aux - range_start + 1;
4303 u64 free_start;
4304 u64 free_len;
4305
4306 extent_changeset_release(&changeset);
4307
4308 /* Only free range in range [start, start + len) */
4309 if (range_start >= start + len ||
4310 range_start + range_len <= start)
4311 continue;
4312 free_start = max(range_start, start);
4313 free_len = min(start + len, range_start + range_len) -
4314 free_start;
4315 /*
4316 * TODO: To also modify reserved->ranges_reserved to reflect
4317 * the modification.
4318 *
4319 * However as long as we free qgroup reserved according to
4320 * EXTENT_QGROUP_RESERVED, we won't double free.
4321 * So not need to rush.
4322 */
4323 ret = clear_record_extent_bits(&inode->io_tree, free_start,
4324 free_start + free_len - 1,
4325 EXTENT_QGROUP_RESERVED, &changeset);
4326 if (ret < 0)
4327 goto out;
4328 freed += changeset.bytes_changed;
4329 }
4330 btrfs_qgroup_free_refroot(root->fs_info, btrfs_root_id(root), freed,
4331 BTRFS_QGROUP_RSV_DATA);
4332 if (freed_ret)
4333 *freed_ret = freed;
4334 ret = 0;
4335 out:
4336 extent_changeset_release(&changeset);
4337 return ret;
4338 }
4339
4340 static int __btrfs_qgroup_release_data(struct btrfs_inode *inode,
4341 struct extent_changeset *reserved, u64 start, u64 len,
4342 u64 *released, int free)
4343 {
4344 struct extent_changeset changeset;
4345 int trace_op = QGROUP_RELEASE;
4346 int ret;
4347
4348 if (btrfs_qgroup_mode(inode->root->fs_info) == BTRFS_QGROUP_MODE_DISABLED) {
4349 return clear_record_extent_bits(&inode->io_tree, start,
4350 start + len - 1,
4351 EXTENT_QGROUP_RESERVED, NULL);
4352 }
4353
4354 /* In release case, we shouldn't have @reserved */
4355 WARN_ON(!free && reserved);
4356 if (free && reserved)
4357 return qgroup_free_reserved_data(inode, reserved, start, len, released);
4358 extent_changeset_init(&changeset);
4359 ret = clear_record_extent_bits(&inode->io_tree, start, start + len -1,
4360 EXTENT_QGROUP_RESERVED, &changeset);
4361 if (ret < 0)
4362 goto out;
4363
4364 if (free)
4365 trace_op = QGROUP_FREE;
4366 trace_btrfs_qgroup_release_data(&inode->vfs_inode, start, len,
4367 changeset.bytes_changed, trace_op);
4368 if (free)
4369 btrfs_qgroup_free_refroot(inode->root->fs_info,
4370 btrfs_root_id(inode->root),
4371 changeset.bytes_changed, BTRFS_QGROUP_RSV_DATA);
4372 if (released)
4373 *released = changeset.bytes_changed;
4374 out:
4375 extent_changeset_release(&changeset);
4376 return ret;
4377 }
4378
4379 /*
4380 * Free a reserved space range from io_tree and related qgroups
4381 *
4382 * Should be called when a range of pages get invalidated before reaching disk.
4383 * Or for error cleanup case.
4384 * if @reserved is given, only reserved range in [@start, @start + @len) will
4385 * be freed.
4386 *
4387 * For data written to disk, use btrfs_qgroup_release_data().
4388 *
4389 * NOTE: This function may sleep for memory allocation.
4390 */
4391 int btrfs_qgroup_free_data(struct btrfs_inode *inode,
4392 struct extent_changeset *reserved,
4393 u64 start, u64 len, u64 *freed)
4394 {
4395 return __btrfs_qgroup_release_data(inode, reserved, start, len, freed, 1);
4396 }
4397
4398 /*
4399 * Release a reserved space range from io_tree only.
4400 *
4401 * Should be called when a range of pages get written to disk and corresponding
4402 * FILE_EXTENT is inserted into corresponding root.
4403 *
4404 * Since new qgroup accounting framework will only update qgroup numbers at
4405 * commit_transaction() time, its reserved space shouldn't be freed from
4406 * related qgroups.
4407 *
4408 * But we should release the range from io_tree, to allow further write to be
4409 * COWed.
4410 *
4411 * NOTE: This function may sleep for memory allocation.
4412 */
4413 int btrfs_qgroup_release_data(struct btrfs_inode *inode, u64 start, u64 len, u64 *released)
4414 {
4415 return __btrfs_qgroup_release_data(inode, NULL, start, len, released, 0);
4416 }
4417
4418 static void add_root_meta_rsv(struct btrfs_root *root, int num_bytes,
4419 enum btrfs_qgroup_rsv_type type)
4420 {
4421 if (type != BTRFS_QGROUP_RSV_META_PREALLOC &&
4422 type != BTRFS_QGROUP_RSV_META_PERTRANS)
4423 return;
4424 if (num_bytes == 0)
4425 return;
4426
4427 spin_lock(&root->qgroup_meta_rsv_lock);
4428 if (type == BTRFS_QGROUP_RSV_META_PREALLOC)
4429 root->qgroup_meta_rsv_prealloc += num_bytes;
4430 else
4431 root->qgroup_meta_rsv_pertrans += num_bytes;
4432 spin_unlock(&root->qgroup_meta_rsv_lock);
4433 }
4434
4435 static int sub_root_meta_rsv(struct btrfs_root *root, int num_bytes,
4436 enum btrfs_qgroup_rsv_type type)
4437 {
4438 if (type != BTRFS_QGROUP_RSV_META_PREALLOC &&
4439 type != BTRFS_QGROUP_RSV_META_PERTRANS)
4440 return 0;
4441 if (num_bytes == 0)
4442 return 0;
4443
4444 spin_lock(&root->qgroup_meta_rsv_lock);
4445 if (type == BTRFS_QGROUP_RSV_META_PREALLOC) {
4446 num_bytes = min_t(u64, root->qgroup_meta_rsv_prealloc,
4447 num_bytes);
4448 root->qgroup_meta_rsv_prealloc -= num_bytes;
4449 } else {
4450 num_bytes = min_t(u64, root->qgroup_meta_rsv_pertrans,
4451 num_bytes);
4452 root->qgroup_meta_rsv_pertrans -= num_bytes;
4453 }
4454 spin_unlock(&root->qgroup_meta_rsv_lock);
4455 return num_bytes;
4456 }
4457
4458 int btrfs_qgroup_reserve_meta(struct btrfs_root *root, int num_bytes,
4459 enum btrfs_qgroup_rsv_type type, bool enforce)
4460 {
4461 struct btrfs_fs_info *fs_info = root->fs_info;
4462 int ret;
4463
4464 if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_DISABLED ||
4465 !is_fstree(btrfs_root_id(root)) || num_bytes == 0)
4466 return 0;
4467
4468 BUG_ON(num_bytes != round_down(num_bytes, fs_info->nodesize));
4469 trace_qgroup_meta_reserve(root, (s64)num_bytes, type);
4470 ret = qgroup_reserve(root, num_bytes, enforce, type);
4471 if (ret < 0)
4472 return ret;
4473 /*
4474 * Record what we have reserved into root.
4475 *
4476 * To avoid quota disabled->enabled underflow.
4477 * In that case, we may try to free space we haven't reserved
4478 * (since quota was disabled), so record what we reserved into root.
4479 * And ensure later release won't underflow this number.
4480 */
4481 add_root_meta_rsv(root, num_bytes, type);
4482 return ret;
4483 }
4484
4485 int __btrfs_qgroup_reserve_meta(struct btrfs_root *root, int num_bytes,
4486 enum btrfs_qgroup_rsv_type type, bool enforce,
4487 bool noflush)
4488 {
4489 int ret;
4490
4491 ret = btrfs_qgroup_reserve_meta(root, num_bytes, type, enforce);
4492 if ((ret <= 0 && ret != -EDQUOT) || noflush)
4493 return ret;
4494
4495 ret = try_flush_qgroup(root);
4496 if (ret < 0)
4497 return ret;
4498 return btrfs_qgroup_reserve_meta(root, num_bytes, type, enforce);
4499 }
4500
4501 /*
4502 * Per-transaction meta reservation should be all freed at transaction commit
4503 * time
4504 */
4505 void btrfs_qgroup_free_meta_all_pertrans(struct btrfs_root *root)
4506 {
4507 struct btrfs_fs_info *fs_info = root->fs_info;
4508
4509 if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_DISABLED ||
4510 !is_fstree(btrfs_root_id(root)))
4511 return;
4512
4513 /* TODO: Update trace point to handle such free */
4514 trace_qgroup_meta_free_all_pertrans(root);
4515 /* Special value -1 means to free all reserved space */
4516 btrfs_qgroup_free_refroot(fs_info, btrfs_root_id(root), (u64)-1,
4517 BTRFS_QGROUP_RSV_META_PERTRANS);
4518 }
4519
4520 void __btrfs_qgroup_free_meta(struct btrfs_root *root, int num_bytes,
4521 enum btrfs_qgroup_rsv_type type)
4522 {
4523 struct btrfs_fs_info *fs_info = root->fs_info;
4524
4525 if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_DISABLED ||
4526 !is_fstree(btrfs_root_id(root)))
4527 return;
4528
4529 /*
4530 * reservation for META_PREALLOC can happen before quota is enabled,
4531 * which can lead to underflow.
4532 * Here ensure we will only free what we really have reserved.
4533 */
4534 num_bytes = sub_root_meta_rsv(root, num_bytes, type);
4535 BUG_ON(num_bytes != round_down(num_bytes, fs_info->nodesize));
4536 trace_qgroup_meta_reserve(root, -(s64)num_bytes, type);
4537 btrfs_qgroup_free_refroot(fs_info, btrfs_root_id(root), num_bytes, type);
4538 }
4539
4540 static void qgroup_convert_meta(struct btrfs_fs_info *fs_info, u64 ref_root,
4541 int num_bytes)
4542 {
4543 struct btrfs_qgroup *qgroup;
4544 LIST_HEAD(qgroup_list);
4545
4546 if (num_bytes == 0)
4547 return;
4548 if (!fs_info->quota_root)
4549 return;
4550
4551 spin_lock(&fs_info->qgroup_lock);
4552 qgroup = find_qgroup_rb(fs_info, ref_root);
4553 if (!qgroup)
4554 goto out;
4555
4556 qgroup_iterator_add(&qgroup_list, qgroup);
4557 list_for_each_entry(qgroup, &qgroup_list, iterator) {
4558 struct btrfs_qgroup_list *glist;
4559
4560 qgroup_rsv_release(fs_info, qgroup, num_bytes,
4561 BTRFS_QGROUP_RSV_META_PREALLOC);
4562 if (!sb_rdonly(fs_info->sb))
4563 qgroup_rsv_add(fs_info, qgroup, num_bytes,
4564 BTRFS_QGROUP_RSV_META_PERTRANS);
4565
4566 list_for_each_entry(glist, &qgroup->groups, next_group)
4567 qgroup_iterator_add(&qgroup_list, glist->group);
4568 }
4569 out:
4570 qgroup_iterator_clean(&qgroup_list);
4571 spin_unlock(&fs_info->qgroup_lock);
4572 }
4573
4574 /*
4575 * Convert @num_bytes of META_PREALLOCATED reservation to META_PERTRANS.
4576 *
4577 * This is called when preallocated meta reservation needs to be used.
4578 * Normally after btrfs_join_transaction() call.
4579 */
4580 void btrfs_qgroup_convert_reserved_meta(struct btrfs_root *root, int num_bytes)
4581 {
4582 struct btrfs_fs_info *fs_info = root->fs_info;
4583
4584 if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_DISABLED ||
4585 !is_fstree(btrfs_root_id(root)))
4586 return;
4587 /* Same as btrfs_qgroup_free_meta_prealloc() */
4588 num_bytes = sub_root_meta_rsv(root, num_bytes,
4589 BTRFS_QGROUP_RSV_META_PREALLOC);
4590 trace_qgroup_meta_convert(root, num_bytes);
4591 qgroup_convert_meta(fs_info, btrfs_root_id(root), num_bytes);
4592 if (!sb_rdonly(fs_info->sb))
4593 add_root_meta_rsv(root, num_bytes, BTRFS_QGROUP_RSV_META_PERTRANS);
4594 }
4595
4596 /*
4597 * Check qgroup reserved space leaking, normally at destroy inode
4598 * time
4599 */
4600 void btrfs_qgroup_check_reserved_leak(struct btrfs_inode *inode)
4601 {
4602 struct extent_changeset changeset;
4603 struct ulist_node *unode;
4604 struct ulist_iterator iter;
4605 int ret;
4606
4607 extent_changeset_init(&changeset);
4608 ret = clear_record_extent_bits(&inode->io_tree, 0, (u64)-1,
4609 EXTENT_QGROUP_RESERVED, &changeset);
4610
4611 WARN_ON(ret < 0);
4612 if (WARN_ON(changeset.bytes_changed)) {
4613 ULIST_ITER_INIT(&iter);
4614 while ((unode = ulist_next(&changeset.range_changed, &iter))) {
4615 btrfs_warn(inode->root->fs_info,
4616 "leaking qgroup reserved space, ino: %llu, start: %llu, end: %llu",
4617 btrfs_ino(inode), unode->val, unode->aux);
4618 }
4619 btrfs_qgroup_free_refroot(inode->root->fs_info,
4620 btrfs_root_id(inode->root),
4621 changeset.bytes_changed, BTRFS_QGROUP_RSV_DATA);
4622
4623 }
4624 extent_changeset_release(&changeset);
4625 }
4626
4627 void btrfs_qgroup_init_swapped_blocks(
4628 struct btrfs_qgroup_swapped_blocks *swapped_blocks)
4629 {
4630 int i;
4631
4632 spin_lock_init(&swapped_blocks->lock);
4633 for (i = 0; i < BTRFS_MAX_LEVEL; i++)
4634 swapped_blocks->blocks[i] = RB_ROOT;
4635 swapped_blocks->swapped = false;
4636 }
4637
4638 /*
4639 * Delete all swapped blocks record of @root.
4640 * Every record here means we skipped a full subtree scan for qgroup.
4641 *
4642 * Gets called when committing one transaction.
4643 */
4644 void btrfs_qgroup_clean_swapped_blocks(struct btrfs_root *root)
4645 {
4646 struct btrfs_qgroup_swapped_blocks *swapped_blocks;
4647 int i;
4648
4649 swapped_blocks = &root->swapped_blocks;
4650
4651 spin_lock(&swapped_blocks->lock);
4652 if (!swapped_blocks->swapped)
4653 goto out;
4654 for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
4655 struct rb_root *cur_root = &swapped_blocks->blocks[i];
4656 struct btrfs_qgroup_swapped_block *entry;
4657 struct btrfs_qgroup_swapped_block *next;
4658
4659 rbtree_postorder_for_each_entry_safe(entry, next, cur_root,
4660 node)
4661 kfree(entry);
4662 swapped_blocks->blocks[i] = RB_ROOT;
4663 }
4664 swapped_blocks->swapped = false;
4665 out:
4666 spin_unlock(&swapped_blocks->lock);
4667 }
4668
4669 /*
4670 * Add subtree roots record into @subvol_root.
4671 *
4672 * @subvol_root: tree root of the subvolume tree get swapped
4673 * @bg: block group under balance
4674 * @subvol_parent/slot: pointer to the subtree root in subvolume tree
4675 * @reloc_parent/slot: pointer to the subtree root in reloc tree
4676 * BOTH POINTERS ARE BEFORE TREE SWAP
4677 * @last_snapshot: last snapshot generation of the subvolume tree
4678 */
4679 int btrfs_qgroup_add_swapped_blocks(struct btrfs_trans_handle *trans,
4680 struct btrfs_root *subvol_root,
4681 struct btrfs_block_group *bg,
4682 struct extent_buffer *subvol_parent, int subvol_slot,
4683 struct extent_buffer *reloc_parent, int reloc_slot,
4684 u64 last_snapshot)
4685 {
4686 struct btrfs_fs_info *fs_info = subvol_root->fs_info;
4687 struct btrfs_qgroup_swapped_blocks *blocks = &subvol_root->swapped_blocks;
4688 struct btrfs_qgroup_swapped_block *block;
4689 struct rb_node **cur;
4690 struct rb_node *parent = NULL;
4691 int level = btrfs_header_level(subvol_parent) - 1;
4692 int ret = 0;
4693
4694 if (!btrfs_qgroup_full_accounting(fs_info))
4695 return 0;
4696
4697 if (btrfs_node_ptr_generation(subvol_parent, subvol_slot) >
4698 btrfs_node_ptr_generation(reloc_parent, reloc_slot)) {
4699 btrfs_err_rl(fs_info,
4700 "%s: bad parameter order, subvol_gen=%llu reloc_gen=%llu",
4701 __func__,
4702 btrfs_node_ptr_generation(subvol_parent, subvol_slot),
4703 btrfs_node_ptr_generation(reloc_parent, reloc_slot));
4704 return -EUCLEAN;
4705 }
4706
4707 block = kmalloc(sizeof(*block), GFP_NOFS);
4708 if (!block) {
4709 ret = -ENOMEM;
4710 goto out;
4711 }
4712
4713 /*
4714 * @reloc_parent/slot is still before swap, while @block is going to
4715 * record the bytenr after swap, so we do the swap here.
4716 */
4717 block->subvol_bytenr = btrfs_node_blockptr(reloc_parent, reloc_slot);
4718 block->subvol_generation = btrfs_node_ptr_generation(reloc_parent,
4719 reloc_slot);
4720 block->reloc_bytenr = btrfs_node_blockptr(subvol_parent, subvol_slot);
4721 block->reloc_generation = btrfs_node_ptr_generation(subvol_parent,
4722 subvol_slot);
4723 block->last_snapshot = last_snapshot;
4724 block->level = level;
4725
4726 /*
4727 * If we have bg == NULL, we're called from btrfs_recover_relocation(),
4728 * no one else can modify tree blocks thus we qgroup will not change
4729 * no matter the value of trace_leaf.
4730 */
4731 if (bg && bg->flags & BTRFS_BLOCK_GROUP_DATA)
4732 block->trace_leaf = true;
4733 else
4734 block->trace_leaf = false;
4735 btrfs_node_key_to_cpu(reloc_parent, &block->first_key, reloc_slot);
4736
4737 /* Insert @block into @blocks */
4738 spin_lock(&blocks->lock);
4739 cur = &blocks->blocks[level].rb_node;
4740 while (*cur) {
4741 struct btrfs_qgroup_swapped_block *entry;
4742
4743 parent = *cur;
4744 entry = rb_entry(parent, struct btrfs_qgroup_swapped_block,
4745 node);
4746
4747 if (entry->subvol_bytenr < block->subvol_bytenr) {
4748 cur = &(*cur)->rb_left;
4749 } else if (entry->subvol_bytenr > block->subvol_bytenr) {
4750 cur = &(*cur)->rb_right;
4751 } else {
4752 if (entry->subvol_generation !=
4753 block->subvol_generation ||
4754 entry->reloc_bytenr != block->reloc_bytenr ||
4755 entry->reloc_generation !=
4756 block->reloc_generation) {
4757 /*
4758 * Duplicated but mismatch entry found.
4759 * Shouldn't happen.
4760 *
4761 * Marking qgroup inconsistent should be enough
4762 * for end users.
4763 */
4764 WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
4765 ret = -EEXIST;
4766 }
4767 kfree(block);
4768 goto out_unlock;
4769 }
4770 }
4771 rb_link_node(&block->node, parent, cur);
4772 rb_insert_color(&block->node, &blocks->blocks[level]);
4773 blocks->swapped = true;
4774 out_unlock:
4775 spin_unlock(&blocks->lock);
4776 out:
4777 if (ret < 0)
4778 qgroup_mark_inconsistent(fs_info);
4779 return ret;
4780 }
4781
4782 /*
4783 * Check if the tree block is a subtree root, and if so do the needed
4784 * delayed subtree trace for qgroup.
4785 *
4786 * This is called during btrfs_cow_block().
4787 */
4788 int btrfs_qgroup_trace_subtree_after_cow(struct btrfs_trans_handle *trans,
4789 struct btrfs_root *root,
4790 struct extent_buffer *subvol_eb)
4791 {
4792 struct btrfs_fs_info *fs_info = root->fs_info;
4793 struct btrfs_tree_parent_check check = { 0 };
4794 struct btrfs_qgroup_swapped_blocks *blocks = &root->swapped_blocks;
4795 struct btrfs_qgroup_swapped_block *block;
4796 struct extent_buffer *reloc_eb = NULL;
4797 struct rb_node *node;
4798 bool found = false;
4799 bool swapped = false;
4800 int level = btrfs_header_level(subvol_eb);
4801 int ret = 0;
4802 int i;
4803
4804 if (!btrfs_qgroup_full_accounting(fs_info))
4805 return 0;
4806 if (!is_fstree(btrfs_root_id(root)) || !root->reloc_root)
4807 return 0;
4808
4809 spin_lock(&blocks->lock);
4810 if (!blocks->swapped) {
4811 spin_unlock(&blocks->lock);
4812 return 0;
4813 }
4814 node = blocks->blocks[level].rb_node;
4815
4816 while (node) {
4817 block = rb_entry(node, struct btrfs_qgroup_swapped_block, node);
4818 if (block->subvol_bytenr < subvol_eb->start) {
4819 node = node->rb_left;
4820 } else if (block->subvol_bytenr > subvol_eb->start) {
4821 node = node->rb_right;
4822 } else {
4823 found = true;
4824 break;
4825 }
4826 }
4827 if (!found) {
4828 spin_unlock(&blocks->lock);
4829 goto out;
4830 }
4831 /* Found one, remove it from @blocks first and update blocks->swapped */
4832 rb_erase(&block->node, &blocks->blocks[level]);
4833 for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
4834 if (RB_EMPTY_ROOT(&blocks->blocks[i])) {
4835 swapped = true;
4836 break;
4837 }
4838 }
4839 blocks->swapped = swapped;
4840 spin_unlock(&blocks->lock);
4841
4842 check.level = block->level;
4843 check.transid = block->reloc_generation;
4844 check.has_first_key = true;
4845 memcpy(&check.first_key, &block->first_key, sizeof(check.first_key));
4846
4847 /* Read out reloc subtree root */
4848 reloc_eb = read_tree_block(fs_info, block->reloc_bytenr, &check);
4849 if (IS_ERR(reloc_eb)) {
4850 ret = PTR_ERR(reloc_eb);
4851 reloc_eb = NULL;
4852 goto free_out;
4853 }
4854 if (!extent_buffer_uptodate(reloc_eb)) {
4855 ret = -EIO;
4856 goto free_out;
4857 }
4858
4859 ret = qgroup_trace_subtree_swap(trans, reloc_eb, subvol_eb,
4860 block->last_snapshot, block->trace_leaf);
4861 free_out:
4862 kfree(block);
4863 free_extent_buffer(reloc_eb);
4864 out:
4865 if (ret < 0) {
4866 btrfs_err_rl(fs_info,
4867 "failed to account subtree at bytenr %llu: %d",
4868 subvol_eb->start, ret);
4869 qgroup_mark_inconsistent(fs_info);
4870 }
4871 return ret;
4872 }
4873
4874 void btrfs_qgroup_destroy_extent_records(struct btrfs_transaction *trans)
4875 {
4876 struct btrfs_qgroup_extent_record *entry;
4877 struct btrfs_qgroup_extent_record *next;
4878 struct rb_root *root;
4879
4880 root = &trans->delayed_refs.dirty_extent_root;
4881 rbtree_postorder_for_each_entry_safe(entry, next, root, node) {
4882 ulist_free(entry->old_roots);
4883 kfree(entry);
4884 }
4885 *root = RB_ROOT;
4886 }
4887
4888 void btrfs_free_squota_rsv(struct btrfs_fs_info *fs_info, u64 root, u64 rsv_bytes)
4889 {
4890 if (btrfs_qgroup_mode(fs_info) != BTRFS_QGROUP_MODE_SIMPLE)
4891 return;
4892
4893 if (!is_fstree(root))
4894 return;
4895
4896 btrfs_qgroup_free_refroot(fs_info, root, rsv_bytes, BTRFS_QGROUP_RSV_DATA);
4897 }
4898
4899 int btrfs_record_squota_delta(struct btrfs_fs_info *fs_info,
4900 const struct btrfs_squota_delta *delta)
4901 {
4902 int ret;
4903 struct btrfs_qgroup *qgroup;
4904 struct btrfs_qgroup *qg;
4905 LIST_HEAD(qgroup_list);
4906 u64 root = delta->root;
4907 u64 num_bytes = delta->num_bytes;
4908 const int sign = (delta->is_inc ? 1 : -1);
4909
4910 if (btrfs_qgroup_mode(fs_info) != BTRFS_QGROUP_MODE_SIMPLE)
4911 return 0;
4912
4913 if (!is_fstree(root))
4914 return 0;
4915
4916 /* If the extent predates enabling quotas, don't count it. */
4917 if (delta->generation < fs_info->qgroup_enable_gen)
4918 return 0;
4919
4920 spin_lock(&fs_info->qgroup_lock);
4921 qgroup = find_qgroup_rb(fs_info, root);
4922 if (!qgroup) {
4923 ret = -ENOENT;
4924 goto out;
4925 }
4926
4927 ret = 0;
4928 qgroup_iterator_add(&qgroup_list, qgroup);
4929 list_for_each_entry(qg, &qgroup_list, iterator) {
4930 struct btrfs_qgroup_list *glist;
4931
4932 qg->excl += num_bytes * sign;
4933 qg->rfer += num_bytes * sign;
4934 qgroup_dirty(fs_info, qg);
4935
4936 list_for_each_entry(glist, &qg->groups, next_group)
4937 qgroup_iterator_add(&qgroup_list, glist->group);
4938 }
4939 qgroup_iterator_clean(&qgroup_list);
4940
4941 out:
4942 spin_unlock(&fs_info->qgroup_lock);
4943 return ret;
4944 }
4945
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.