Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2014 Facebook. All rights reserved.
4 */
5
6#include <linux/sched.h>
7#include <linux/stacktrace.h>
8#include "messages.h"
9#include "ctree.h"
10#include "disk-io.h"
11#include "locking.h"
12#include "delayed-ref.h"
13#include "ref-verify.h"
14#include "fs.h"
15#include "accessors.h"
16
17/*
18 * Used to keep track the roots and number of refs each root has for a given
19 * bytenr. This just tracks the number of direct references, no shared
20 * references.
21 */
22struct root_entry {
23 u64 root_objectid;
24 u64 num_refs;
25 struct rb_node node;
26};
27
28/*
29 * These are meant to represent what should exist in the extent tree, these can
30 * be used to verify the extent tree is consistent as these should all match
31 * what the extent tree says.
32 */
33struct ref_entry {
34 u64 root_objectid;
35 u64 parent;
36 u64 owner;
37 u64 offset;
38 u64 num_refs;
39 struct rb_node node;
40};
41
42#define MAX_TRACE 16
43
44/*
45 * Whenever we add/remove a reference we record the action. The action maps
46 * back to the delayed ref action. We hold the ref we are changing in the
47 * action so we can account for the history properly, and we record the root we
48 * were called with since it could be different from ref_root. We also store
49 * stack traces because that's how I roll.
50 */
51struct ref_action {
52 int action;
53 u64 root;
54 struct ref_entry ref;
55 struct list_head list;
56 unsigned long trace[MAX_TRACE];
57 unsigned int trace_len;
58};
59
60/*
61 * One of these for every block we reference, it holds the roots and references
62 * to it as well as all of the ref actions that have occurred to it. We never
63 * free it until we unmount the file system in order to make sure re-allocations
64 * are happening properly.
65 */
66struct block_entry {
67 u64 bytenr;
68 u64 len;
69 u64 num_refs;
70 int metadata;
71 int from_disk;
72 struct rb_root roots;
73 struct rb_root refs;
74 struct rb_node node;
75 struct list_head actions;
76};
77
78static struct block_entry *insert_block_entry(struct rb_root *root,
79 struct block_entry *be)
80{
81 struct rb_node **p = &root->rb_node;
82 struct rb_node *parent_node = NULL;
83 struct block_entry *entry;
84
85 while (*p) {
86 parent_node = *p;
87 entry = rb_entry(parent_node, struct block_entry, node);
88 if (entry->bytenr > be->bytenr)
89 p = &(*p)->rb_left;
90 else if (entry->bytenr < be->bytenr)
91 p = &(*p)->rb_right;
92 else
93 return entry;
94 }
95
96 rb_link_node(&be->node, parent_node, p);
97 rb_insert_color(&be->node, root);
98 return NULL;
99}
100
101static struct block_entry *lookup_block_entry(struct rb_root *root, u64 bytenr)
102{
103 struct rb_node *n;
104 struct block_entry *entry = NULL;
105
106 n = root->rb_node;
107 while (n) {
108 entry = rb_entry(n, struct block_entry, node);
109 if (entry->bytenr < bytenr)
110 n = n->rb_right;
111 else if (entry->bytenr > bytenr)
112 n = n->rb_left;
113 else
114 return entry;
115 }
116 return NULL;
117}
118
119static struct root_entry *insert_root_entry(struct rb_root *root,
120 struct root_entry *re)
121{
122 struct rb_node **p = &root->rb_node;
123 struct rb_node *parent_node = NULL;
124 struct root_entry *entry;
125
126 while (*p) {
127 parent_node = *p;
128 entry = rb_entry(parent_node, struct root_entry, node);
129 if (entry->root_objectid > re->root_objectid)
130 p = &(*p)->rb_left;
131 else if (entry->root_objectid < re->root_objectid)
132 p = &(*p)->rb_right;
133 else
134 return entry;
135 }
136
137 rb_link_node(&re->node, parent_node, p);
138 rb_insert_color(&re->node, root);
139 return NULL;
140
141}
142
143static int comp_refs(struct ref_entry *ref1, struct ref_entry *ref2)
144{
145 if (ref1->root_objectid < ref2->root_objectid)
146 return -1;
147 if (ref1->root_objectid > ref2->root_objectid)
148 return 1;
149 if (ref1->parent < ref2->parent)
150 return -1;
151 if (ref1->parent > ref2->parent)
152 return 1;
153 if (ref1->owner < ref2->owner)
154 return -1;
155 if (ref1->owner > ref2->owner)
156 return 1;
157 if (ref1->offset < ref2->offset)
158 return -1;
159 if (ref1->offset > ref2->offset)
160 return 1;
161 return 0;
162}
163
164static struct ref_entry *insert_ref_entry(struct rb_root *root,
165 struct ref_entry *ref)
166{
167 struct rb_node **p = &root->rb_node;
168 struct rb_node *parent_node = NULL;
169 struct ref_entry *entry;
170 int cmp;
171
172 while (*p) {
173 parent_node = *p;
174 entry = rb_entry(parent_node, struct ref_entry, node);
175 cmp = comp_refs(entry, ref);
176 if (cmp > 0)
177 p = &(*p)->rb_left;
178 else if (cmp < 0)
179 p = &(*p)->rb_right;
180 else
181 return entry;
182 }
183
184 rb_link_node(&ref->node, parent_node, p);
185 rb_insert_color(&ref->node, root);
186 return NULL;
187
188}
189
190static struct root_entry *lookup_root_entry(struct rb_root *root, u64 objectid)
191{
192 struct rb_node *n;
193 struct root_entry *entry = NULL;
194
195 n = root->rb_node;
196 while (n) {
197 entry = rb_entry(n, struct root_entry, node);
198 if (entry->root_objectid < objectid)
199 n = n->rb_right;
200 else if (entry->root_objectid > objectid)
201 n = n->rb_left;
202 else
203 return entry;
204 }
205 return NULL;
206}
207
208#ifdef CONFIG_STACKTRACE
209static void __save_stack_trace(struct ref_action *ra)
210{
211 ra->trace_len = stack_trace_save(ra->trace, MAX_TRACE, 2);
212}
213
214static void __print_stack_trace(struct btrfs_fs_info *fs_info,
215 struct ref_action *ra)
216{
217 if (ra->trace_len == 0) {
218 btrfs_err(fs_info, " ref-verify: no stacktrace");
219 return;
220 }
221 stack_trace_print(ra->trace, ra->trace_len, 2);
222}
223#else
224static inline void __save_stack_trace(struct ref_action *ra)
225{
226}
227
228static inline void __print_stack_trace(struct btrfs_fs_info *fs_info,
229 struct ref_action *ra)
230{
231 btrfs_err(fs_info, " ref-verify: no stacktrace support");
232}
233#endif
234
235static void free_block_entry(struct block_entry *be)
236{
237 struct root_entry *re;
238 struct ref_entry *ref;
239 struct ref_action *ra;
240 struct rb_node *n;
241
242 while ((n = rb_first(&be->roots))) {
243 re = rb_entry(n, struct root_entry, node);
244 rb_erase(&re->node, &be->roots);
245 kfree(re);
246 }
247
248 while((n = rb_first(&be->refs))) {
249 ref = rb_entry(n, struct ref_entry, node);
250 rb_erase(&ref->node, &be->refs);
251 kfree(ref);
252 }
253
254 while (!list_empty(&be->actions)) {
255 ra = list_first_entry(&be->actions, struct ref_action,
256 list);
257 list_del(&ra->list);
258 kfree(ra);
259 }
260 kfree(be);
261}
262
263static struct block_entry *add_block_entry(struct btrfs_fs_info *fs_info,
264 u64 bytenr, u64 len,
265 u64 root_objectid)
266{
267 struct block_entry *be = NULL, *exist;
268 struct root_entry *re = NULL;
269
270 re = kzalloc(sizeof(struct root_entry), GFP_NOFS);
271 be = kzalloc(sizeof(struct block_entry), GFP_NOFS);
272 if (!be || !re) {
273 kfree(re);
274 kfree(be);
275 return ERR_PTR(-ENOMEM);
276 }
277 be->bytenr = bytenr;
278 be->len = len;
279
280 re->root_objectid = root_objectid;
281 re->num_refs = 0;
282
283 spin_lock(&fs_info->ref_verify_lock);
284 exist = insert_block_entry(&fs_info->block_tree, be);
285 if (exist) {
286 if (root_objectid) {
287 struct root_entry *exist_re;
288
289 exist_re = insert_root_entry(&exist->roots, re);
290 if (exist_re)
291 kfree(re);
292 } else {
293 kfree(re);
294 }
295 kfree(be);
296 return exist;
297 }
298
299 be->num_refs = 0;
300 be->metadata = 0;
301 be->from_disk = 0;
302 be->roots = RB_ROOT;
303 be->refs = RB_ROOT;
304 INIT_LIST_HEAD(&be->actions);
305 if (root_objectid)
306 insert_root_entry(&be->roots, re);
307 else
308 kfree(re);
309 return be;
310}
311
312static int add_tree_block(struct btrfs_fs_info *fs_info, u64 ref_root,
313 u64 parent, u64 bytenr, int level)
314{
315 struct block_entry *be;
316 struct root_entry *re;
317 struct ref_entry *ref = NULL, *exist;
318
319 ref = kmalloc(sizeof(struct ref_entry), GFP_NOFS);
320 if (!ref)
321 return -ENOMEM;
322
323 if (parent)
324 ref->root_objectid = 0;
325 else
326 ref->root_objectid = ref_root;
327 ref->parent = parent;
328 ref->owner = level;
329 ref->offset = 0;
330 ref->num_refs = 1;
331
332 be = add_block_entry(fs_info, bytenr, fs_info->nodesize, ref_root);
333 if (IS_ERR(be)) {
334 kfree(ref);
335 return PTR_ERR(be);
336 }
337 be->num_refs++;
338 be->from_disk = 1;
339 be->metadata = 1;
340
341 if (!parent) {
342 ASSERT(ref_root);
343 re = lookup_root_entry(&be->roots, ref_root);
344 ASSERT(re);
345 re->num_refs++;
346 }
347 exist = insert_ref_entry(&be->refs, ref);
348 if (exist) {
349 exist->num_refs++;
350 kfree(ref);
351 }
352 spin_unlock(&fs_info->ref_verify_lock);
353
354 return 0;
355}
356
357static int add_shared_data_ref(struct btrfs_fs_info *fs_info,
358 u64 parent, u32 num_refs, u64 bytenr,
359 u64 num_bytes)
360{
361 struct block_entry *be;
362 struct ref_entry *ref;
363
364 ref = kzalloc(sizeof(struct ref_entry), GFP_NOFS);
365 if (!ref)
366 return -ENOMEM;
367 be = add_block_entry(fs_info, bytenr, num_bytes, 0);
368 if (IS_ERR(be)) {
369 kfree(ref);
370 return PTR_ERR(be);
371 }
372 be->num_refs += num_refs;
373
374 ref->parent = parent;
375 ref->num_refs = num_refs;
376 if (insert_ref_entry(&be->refs, ref)) {
377 spin_unlock(&fs_info->ref_verify_lock);
378 btrfs_err(fs_info, "existing shared ref when reading from disk?");
379 kfree(ref);
380 return -EINVAL;
381 }
382 spin_unlock(&fs_info->ref_verify_lock);
383 return 0;
384}
385
386static int add_extent_data_ref(struct btrfs_fs_info *fs_info,
387 struct extent_buffer *leaf,
388 struct btrfs_extent_data_ref *dref,
389 u64 bytenr, u64 num_bytes)
390{
391 struct block_entry *be;
392 struct ref_entry *ref;
393 struct root_entry *re;
394 u64 ref_root = btrfs_extent_data_ref_root(leaf, dref);
395 u64 owner = btrfs_extent_data_ref_objectid(leaf, dref);
396 u64 offset = btrfs_extent_data_ref_offset(leaf, dref);
397 u32 num_refs = btrfs_extent_data_ref_count(leaf, dref);
398
399 ref = kzalloc(sizeof(struct ref_entry), GFP_NOFS);
400 if (!ref)
401 return -ENOMEM;
402 be = add_block_entry(fs_info, bytenr, num_bytes, ref_root);
403 if (IS_ERR(be)) {
404 kfree(ref);
405 return PTR_ERR(be);
406 }
407 be->num_refs += num_refs;
408
409 ref->parent = 0;
410 ref->owner = owner;
411 ref->root_objectid = ref_root;
412 ref->offset = offset;
413 ref->num_refs = num_refs;
414 if (insert_ref_entry(&be->refs, ref)) {
415 spin_unlock(&fs_info->ref_verify_lock);
416 btrfs_err(fs_info, "existing ref when reading from disk?");
417 kfree(ref);
418 return -EINVAL;
419 }
420
421 re = lookup_root_entry(&be->roots, ref_root);
422 if (!re) {
423 spin_unlock(&fs_info->ref_verify_lock);
424 btrfs_err(fs_info, "missing root in new block entry?");
425 return -EINVAL;
426 }
427 re->num_refs += num_refs;
428 spin_unlock(&fs_info->ref_verify_lock);
429 return 0;
430}
431
432static int process_extent_item(struct btrfs_fs_info *fs_info,
433 struct btrfs_path *path, struct btrfs_key *key,
434 int slot, int *tree_block_level)
435{
436 struct btrfs_extent_item *ei;
437 struct btrfs_extent_inline_ref *iref;
438 struct btrfs_extent_data_ref *dref;
439 struct btrfs_shared_data_ref *sref;
440 struct extent_buffer *leaf = path->nodes[0];
441 u32 item_size = btrfs_item_size(leaf, slot);
442 unsigned long end, ptr;
443 u64 offset, flags, count;
444 int type;
445 int ret = 0;
446
447 ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
448 flags = btrfs_extent_flags(leaf, ei);
449
450 if ((key->type == BTRFS_EXTENT_ITEM_KEY) &&
451 flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
452 struct btrfs_tree_block_info *info;
453
454 info = (struct btrfs_tree_block_info *)(ei + 1);
455 *tree_block_level = btrfs_tree_block_level(leaf, info);
456 iref = (struct btrfs_extent_inline_ref *)(info + 1);
457 } else {
458 if (key->type == BTRFS_METADATA_ITEM_KEY)
459 *tree_block_level = key->offset;
460 iref = (struct btrfs_extent_inline_ref *)(ei + 1);
461 }
462
463 ptr = (unsigned long)iref;
464 end = (unsigned long)ei + item_size;
465 while (ptr < end) {
466 iref = (struct btrfs_extent_inline_ref *)ptr;
467 type = btrfs_extent_inline_ref_type(leaf, iref);
468 offset = btrfs_extent_inline_ref_offset(leaf, iref);
469 switch (type) {
470 case BTRFS_TREE_BLOCK_REF_KEY:
471 ret = add_tree_block(fs_info, offset, 0, key->objectid,
472 *tree_block_level);
473 break;
474 case BTRFS_SHARED_BLOCK_REF_KEY:
475 ret = add_tree_block(fs_info, 0, offset, key->objectid,
476 *tree_block_level);
477 break;
478 case BTRFS_EXTENT_DATA_REF_KEY:
479 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
480 ret = add_extent_data_ref(fs_info, leaf, dref,
481 key->objectid, key->offset);
482 break;
483 case BTRFS_SHARED_DATA_REF_KEY:
484 sref = (struct btrfs_shared_data_ref *)(iref + 1);
485 count = btrfs_shared_data_ref_count(leaf, sref);
486 ret = add_shared_data_ref(fs_info, offset, count,
487 key->objectid, key->offset);
488 break;
489 case BTRFS_EXTENT_OWNER_REF_KEY:
490 if (!btrfs_fs_incompat(fs_info, SIMPLE_QUOTA)) {
491 btrfs_err(fs_info,
492 "found extent owner ref without simple quotas enabled");
493 ret = -EINVAL;
494 }
495 break;
496 default:
497 btrfs_err(fs_info, "invalid key type in iref");
498 ret = -EINVAL;
499 break;
500 }
501 if (ret)
502 break;
503 ptr += btrfs_extent_inline_ref_size(type);
504 }
505 return ret;
506}
507
508static int process_leaf(struct btrfs_root *root,
509 struct btrfs_path *path, u64 *bytenr, u64 *num_bytes,
510 int *tree_block_level)
511{
512 struct btrfs_fs_info *fs_info = root->fs_info;
513 struct extent_buffer *leaf = path->nodes[0];
514 struct btrfs_extent_data_ref *dref;
515 struct btrfs_shared_data_ref *sref;
516 u32 count;
517 int i = 0, ret = 0;
518 struct btrfs_key key;
519 int nritems = btrfs_header_nritems(leaf);
520
521 for (i = 0; i < nritems; i++) {
522 btrfs_item_key_to_cpu(leaf, &key, i);
523 switch (key.type) {
524 case BTRFS_EXTENT_ITEM_KEY:
525 *num_bytes = key.offset;
526 fallthrough;
527 case BTRFS_METADATA_ITEM_KEY:
528 *bytenr = key.objectid;
529 ret = process_extent_item(fs_info, path, &key, i,
530 tree_block_level);
531 break;
532 case BTRFS_TREE_BLOCK_REF_KEY:
533 ret = add_tree_block(fs_info, key.offset, 0,
534 key.objectid, *tree_block_level);
535 break;
536 case BTRFS_SHARED_BLOCK_REF_KEY:
537 ret = add_tree_block(fs_info, 0, key.offset,
538 key.objectid, *tree_block_level);
539 break;
540 case BTRFS_EXTENT_DATA_REF_KEY:
541 dref = btrfs_item_ptr(leaf, i,
542 struct btrfs_extent_data_ref);
543 ret = add_extent_data_ref(fs_info, leaf, dref, *bytenr,
544 *num_bytes);
545 break;
546 case BTRFS_SHARED_DATA_REF_KEY:
547 sref = btrfs_item_ptr(leaf, i,
548 struct btrfs_shared_data_ref);
549 count = btrfs_shared_data_ref_count(leaf, sref);
550 ret = add_shared_data_ref(fs_info, key.offset, count,
551 *bytenr, *num_bytes);
552 break;
553 default:
554 break;
555 }
556 if (ret)
557 break;
558 }
559 return ret;
560}
561
562/* Walk down to the leaf from the given level */
563static int walk_down_tree(struct btrfs_root *root, struct btrfs_path *path,
564 int level, u64 *bytenr, u64 *num_bytes,
565 int *tree_block_level)
566{
567 struct extent_buffer *eb;
568 int ret = 0;
569
570 while (level >= 0) {
571 if (level) {
572 eb = btrfs_read_node_slot(path->nodes[level],
573 path->slots[level]);
574 if (IS_ERR(eb))
575 return PTR_ERR(eb);
576 btrfs_tree_read_lock(eb);
577 path->nodes[level-1] = eb;
578 path->slots[level-1] = 0;
579 path->locks[level-1] = BTRFS_READ_LOCK;
580 } else {
581 ret = process_leaf(root, path, bytenr, num_bytes,
582 tree_block_level);
583 if (ret)
584 break;
585 }
586 level--;
587 }
588 return ret;
589}
590
591/* Walk up to the next node that needs to be processed */
592static int walk_up_tree(struct btrfs_path *path, int *level)
593{
594 int l;
595
596 for (l = 0; l < BTRFS_MAX_LEVEL; l++) {
597 if (!path->nodes[l])
598 continue;
599 if (l) {
600 path->slots[l]++;
601 if (path->slots[l] <
602 btrfs_header_nritems(path->nodes[l])) {
603 *level = l;
604 return 0;
605 }
606 }
607 btrfs_tree_unlock_rw(path->nodes[l], path->locks[l]);
608 free_extent_buffer(path->nodes[l]);
609 path->nodes[l] = NULL;
610 path->slots[l] = 0;
611 path->locks[l] = 0;
612 }
613
614 return 1;
615}
616
617static void dump_ref_action(struct btrfs_fs_info *fs_info,
618 struct ref_action *ra)
619{
620 btrfs_err(fs_info,
621" Ref action %d, root %llu, ref_root %llu, parent %llu, owner %llu, offset %llu, num_refs %llu",
622 ra->action, ra->root, ra->ref.root_objectid, ra->ref.parent,
623 ra->ref.owner, ra->ref.offset, ra->ref.num_refs);
624 __print_stack_trace(fs_info, ra);
625}
626
627/*
628 * Dumps all the information from the block entry to printk, it's going to be
629 * awesome.
630 */
631static void dump_block_entry(struct btrfs_fs_info *fs_info,
632 struct block_entry *be)
633{
634 struct ref_entry *ref;
635 struct root_entry *re;
636 struct ref_action *ra;
637 struct rb_node *n;
638
639 btrfs_err(fs_info,
640"dumping block entry [%llu %llu], num_refs %llu, metadata %d, from disk %d",
641 be->bytenr, be->len, be->num_refs, be->metadata,
642 be->from_disk);
643
644 for (n = rb_first(&be->refs); n; n = rb_next(n)) {
645 ref = rb_entry(n, struct ref_entry, node);
646 btrfs_err(fs_info,
647" ref root %llu, parent %llu, owner %llu, offset %llu, num_refs %llu",
648 ref->root_objectid, ref->parent, ref->owner,
649 ref->offset, ref->num_refs);
650 }
651
652 for (n = rb_first(&be->roots); n; n = rb_next(n)) {
653 re = rb_entry(n, struct root_entry, node);
654 btrfs_err(fs_info, " root entry %llu, num_refs %llu",
655 re->root_objectid, re->num_refs);
656 }
657
658 list_for_each_entry(ra, &be->actions, list)
659 dump_ref_action(fs_info, ra);
660}
661
662/*
663 * Called when we modify a ref for a bytenr.
664 *
665 * This will add an action item to the given bytenr and do sanity checks to make
666 * sure we haven't messed something up. If we are making a new allocation and
667 * this block entry has history we will delete all previous actions as long as
668 * our sanity checks pass as they are no longer needed.
669 */
670int btrfs_ref_tree_mod(struct btrfs_fs_info *fs_info,
671 struct btrfs_ref *generic_ref)
672{
673 struct ref_entry *ref = NULL, *exist;
674 struct ref_action *ra = NULL;
675 struct block_entry *be = NULL;
676 struct root_entry *re = NULL;
677 int action = generic_ref->action;
678 int ret = 0;
679 bool metadata;
680 u64 bytenr = generic_ref->bytenr;
681 u64 num_bytes = generic_ref->num_bytes;
682 u64 parent = generic_ref->parent;
683 u64 ref_root = 0;
684 u64 owner = 0;
685 u64 offset = 0;
686
687 if (!btrfs_test_opt(fs_info, REF_VERIFY))
688 return 0;
689
690 if (generic_ref->type == BTRFS_REF_METADATA) {
691 if (!parent)
692 ref_root = generic_ref->ref_root;
693 owner = generic_ref->tree_ref.level;
694 } else if (!parent) {
695 ref_root = generic_ref->ref_root;
696 owner = generic_ref->data_ref.objectid;
697 offset = generic_ref->data_ref.offset;
698 }
699 metadata = owner < BTRFS_FIRST_FREE_OBJECTID;
700
701 ref = kzalloc(sizeof(struct ref_entry), GFP_NOFS);
702 ra = kmalloc(sizeof(struct ref_action), GFP_NOFS);
703 if (!ra || !ref) {
704 kfree(ref);
705 kfree(ra);
706 ret = -ENOMEM;
707 goto out;
708 }
709
710 ref->parent = parent;
711 ref->owner = owner;
712 ref->root_objectid = ref_root;
713 ref->offset = offset;
714 ref->num_refs = (action == BTRFS_DROP_DELAYED_REF) ? -1 : 1;
715
716 memcpy(&ra->ref, ref, sizeof(struct ref_entry));
717 /*
718 * Save the extra info from the delayed ref in the ref action to make it
719 * easier to figure out what is happening. The real ref's we add to the
720 * ref tree need to reflect what we save on disk so it matches any
721 * on-disk refs we pre-loaded.
722 */
723 ra->ref.owner = owner;
724 ra->ref.offset = offset;
725 ra->ref.root_objectid = ref_root;
726 __save_stack_trace(ra);
727
728 INIT_LIST_HEAD(&ra->list);
729 ra->action = action;
730 ra->root = generic_ref->real_root;
731
732 /*
733 * This is an allocation, preallocate the block_entry in case we haven't
734 * used it before.
735 */
736 ret = -EINVAL;
737 if (action == BTRFS_ADD_DELAYED_EXTENT) {
738 /*
739 * For subvol_create we'll just pass in whatever the parent root
740 * is and the new root objectid, so let's not treat the passed
741 * in root as if it really has a ref for this bytenr.
742 */
743 be = add_block_entry(fs_info, bytenr, num_bytes, ref_root);
744 if (IS_ERR(be)) {
745 kfree(ref);
746 kfree(ra);
747 ret = PTR_ERR(be);
748 goto out;
749 }
750 be->num_refs++;
751 if (metadata)
752 be->metadata = 1;
753
754 if (be->num_refs != 1) {
755 btrfs_err(fs_info,
756 "re-allocated a block that still has references to it!");
757 dump_block_entry(fs_info, be);
758 dump_ref_action(fs_info, ra);
759 kfree(ref);
760 kfree(ra);
761 goto out_unlock;
762 }
763
764 while (!list_empty(&be->actions)) {
765 struct ref_action *tmp;
766
767 tmp = list_first_entry(&be->actions, struct ref_action,
768 list);
769 list_del(&tmp->list);
770 kfree(tmp);
771 }
772 } else {
773 struct root_entry *tmp;
774
775 if (!parent) {
776 re = kmalloc(sizeof(struct root_entry), GFP_NOFS);
777 if (!re) {
778 kfree(ref);
779 kfree(ra);
780 ret = -ENOMEM;
781 goto out;
782 }
783 /*
784 * This is the root that is modifying us, so it's the
785 * one we want to lookup below when we modify the
786 * re->num_refs.
787 */
788 ref_root = generic_ref->real_root;
789 re->root_objectid = generic_ref->real_root;
790 re->num_refs = 0;
791 }
792
793 spin_lock(&fs_info->ref_verify_lock);
794 be = lookup_block_entry(&fs_info->block_tree, bytenr);
795 if (!be) {
796 btrfs_err(fs_info,
797"trying to do action %d to bytenr %llu num_bytes %llu but there is no existing entry!",
798 action, bytenr, num_bytes);
799 dump_ref_action(fs_info, ra);
800 kfree(ref);
801 kfree(ra);
802 kfree(re);
803 goto out_unlock;
804 } else if (be->num_refs == 0) {
805 btrfs_err(fs_info,
806 "trying to do action %d for a bytenr that has 0 total references",
807 action);
808 dump_block_entry(fs_info, be);
809 dump_ref_action(fs_info, ra);
810 kfree(ref);
811 kfree(ra);
812 kfree(re);
813 goto out_unlock;
814 }
815
816 if (!parent) {
817 tmp = insert_root_entry(&be->roots, re);
818 if (tmp) {
819 kfree(re);
820 re = tmp;
821 }
822 }
823 }
824
825 exist = insert_ref_entry(&be->refs, ref);
826 if (exist) {
827 if (action == BTRFS_DROP_DELAYED_REF) {
828 if (exist->num_refs == 0) {
829 btrfs_err(fs_info,
830"dropping a ref for a existing root that doesn't have a ref on the block");
831 dump_block_entry(fs_info, be);
832 dump_ref_action(fs_info, ra);
833 kfree(ref);
834 kfree(ra);
835 goto out_unlock;
836 }
837 exist->num_refs--;
838 if (exist->num_refs == 0) {
839 rb_erase(&exist->node, &be->refs);
840 kfree(exist);
841 }
842 } else if (!be->metadata) {
843 exist->num_refs++;
844 } else {
845 btrfs_err(fs_info,
846"attempting to add another ref for an existing ref on a tree block");
847 dump_block_entry(fs_info, be);
848 dump_ref_action(fs_info, ra);
849 kfree(ref);
850 kfree(ra);
851 goto out_unlock;
852 }
853 kfree(ref);
854 } else {
855 if (action == BTRFS_DROP_DELAYED_REF) {
856 btrfs_err(fs_info,
857"dropping a ref for a root that doesn't have a ref on the block");
858 dump_block_entry(fs_info, be);
859 dump_ref_action(fs_info, ra);
860 rb_erase(&ref->node, &be->refs);
861 kfree(ref);
862 kfree(ra);
863 goto out_unlock;
864 }
865 }
866
867 if (!parent && !re) {
868 re = lookup_root_entry(&be->roots, ref_root);
869 if (!re) {
870 /*
871 * This shouldn't happen because we will add our re
872 * above when we lookup the be with !parent, but just in
873 * case catch this case so we don't panic because I
874 * didn't think of some other corner case.
875 */
876 btrfs_err(fs_info, "failed to find root %llu for %llu",
877 generic_ref->real_root, be->bytenr);
878 dump_block_entry(fs_info, be);
879 dump_ref_action(fs_info, ra);
880 kfree(ra);
881 goto out_unlock;
882 }
883 }
884 if (action == BTRFS_DROP_DELAYED_REF) {
885 if (re)
886 re->num_refs--;
887 be->num_refs--;
888 } else if (action == BTRFS_ADD_DELAYED_REF) {
889 be->num_refs++;
890 if (re)
891 re->num_refs++;
892 }
893 list_add_tail(&ra->list, &be->actions);
894 ret = 0;
895out_unlock:
896 spin_unlock(&fs_info->ref_verify_lock);
897out:
898 if (ret) {
899 btrfs_free_ref_cache(fs_info);
900 btrfs_clear_opt(fs_info->mount_opt, REF_VERIFY);
901 }
902 return ret;
903}
904
905/* Free up the ref cache */
906void btrfs_free_ref_cache(struct btrfs_fs_info *fs_info)
907{
908 struct block_entry *be;
909 struct rb_node *n;
910
911 if (!btrfs_test_opt(fs_info, REF_VERIFY))
912 return;
913
914 spin_lock(&fs_info->ref_verify_lock);
915 while ((n = rb_first(&fs_info->block_tree))) {
916 be = rb_entry(n, struct block_entry, node);
917 rb_erase(&be->node, &fs_info->block_tree);
918 free_block_entry(be);
919 cond_resched_lock(&fs_info->ref_verify_lock);
920 }
921 spin_unlock(&fs_info->ref_verify_lock);
922}
923
924void btrfs_free_ref_tree_range(struct btrfs_fs_info *fs_info, u64 start,
925 u64 len)
926{
927 struct block_entry *be = NULL, *entry;
928 struct rb_node *n;
929
930 if (!btrfs_test_opt(fs_info, REF_VERIFY))
931 return;
932
933 spin_lock(&fs_info->ref_verify_lock);
934 n = fs_info->block_tree.rb_node;
935 while (n) {
936 entry = rb_entry(n, struct block_entry, node);
937 if (entry->bytenr < start) {
938 n = n->rb_right;
939 } else if (entry->bytenr > start) {
940 n = n->rb_left;
941 } else {
942 be = entry;
943 break;
944 }
945 /* We want to get as close to start as possible */
946 if (be == NULL ||
947 (entry->bytenr < start && be->bytenr > start) ||
948 (entry->bytenr < start && entry->bytenr > be->bytenr))
949 be = entry;
950 }
951
952 /*
953 * Could have an empty block group, maybe have something to check for
954 * this case to verify we were actually empty?
955 */
956 if (!be) {
957 spin_unlock(&fs_info->ref_verify_lock);
958 return;
959 }
960
961 n = &be->node;
962 while (n) {
963 be = rb_entry(n, struct block_entry, node);
964 n = rb_next(n);
965 if (be->bytenr < start && be->bytenr + be->len > start) {
966 btrfs_err(fs_info,
967 "block entry overlaps a block group [%llu,%llu]!",
968 start, len);
969 dump_block_entry(fs_info, be);
970 continue;
971 }
972 if (be->bytenr < start)
973 continue;
974 if (be->bytenr >= start + len)
975 break;
976 if (be->bytenr + be->len > start + len) {
977 btrfs_err(fs_info,
978 "block entry overlaps a block group [%llu,%llu]!",
979 start, len);
980 dump_block_entry(fs_info, be);
981 }
982 rb_erase(&be->node, &fs_info->block_tree);
983 free_block_entry(be);
984 }
985 spin_unlock(&fs_info->ref_verify_lock);
986}
987
988/* Walk down all roots and build the ref tree, meant to be called at mount */
989int btrfs_build_ref_tree(struct btrfs_fs_info *fs_info)
990{
991 struct btrfs_root *extent_root;
992 struct btrfs_path *path;
993 struct extent_buffer *eb;
994 int tree_block_level = 0;
995 u64 bytenr = 0, num_bytes = 0;
996 int ret, level;
997
998 if (!btrfs_test_opt(fs_info, REF_VERIFY))
999 return 0;
1000
1001 path = btrfs_alloc_path();
1002 if (!path)
1003 return -ENOMEM;
1004
1005 extent_root = btrfs_extent_root(fs_info, 0);
1006 eb = btrfs_read_lock_root_node(extent_root);
1007 level = btrfs_header_level(eb);
1008 path->nodes[level] = eb;
1009 path->slots[level] = 0;
1010 path->locks[level] = BTRFS_READ_LOCK;
1011
1012 while (1) {
1013 /*
1014 * We have to keep track of the bytenr/num_bytes we last hit
1015 * because we could have run out of space for an inline ref, and
1016 * would have had to added a ref key item which may appear on a
1017 * different leaf from the original extent item.
1018 */
1019 ret = walk_down_tree(extent_root, path, level,
1020 &bytenr, &num_bytes, &tree_block_level);
1021 if (ret)
1022 break;
1023 ret = walk_up_tree(path, &level);
1024 if (ret < 0)
1025 break;
1026 if (ret > 0) {
1027 ret = 0;
1028 break;
1029 }
1030 }
1031 if (ret) {
1032 btrfs_free_ref_cache(fs_info);
1033 btrfs_clear_opt(fs_info->mount_opt, REF_VERIFY);
1034 }
1035 btrfs_free_path(path);
1036 return ret;
1037}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2014 Facebook. All rights reserved.
4 */
5
6#include <linux/sched.h>
7#include <linux/stacktrace.h>
8#include "ctree.h"
9#include "disk-io.h"
10#include "locking.h"
11#include "delayed-ref.h"
12#include "ref-verify.h"
13
14/*
15 * Used to keep track the roots and number of refs each root has for a given
16 * bytenr. This just tracks the number of direct references, no shared
17 * references.
18 */
19struct root_entry {
20 u64 root_objectid;
21 u64 num_refs;
22 struct rb_node node;
23};
24
25/*
26 * These are meant to represent what should exist in the extent tree, these can
27 * be used to verify the extent tree is consistent as these should all match
28 * what the extent tree says.
29 */
30struct ref_entry {
31 u64 root_objectid;
32 u64 parent;
33 u64 owner;
34 u64 offset;
35 u64 num_refs;
36 struct rb_node node;
37};
38
39#define MAX_TRACE 16
40
41/*
42 * Whenever we add/remove a reference we record the action. The action maps
43 * back to the delayed ref action. We hold the ref we are changing in the
44 * action so we can account for the history properly, and we record the root we
45 * were called with since it could be different from ref_root. We also store
46 * stack traces because that's how I roll.
47 */
48struct ref_action {
49 int action;
50 u64 root;
51 struct ref_entry ref;
52 struct list_head list;
53 unsigned long trace[MAX_TRACE];
54 unsigned int trace_len;
55};
56
57/*
58 * One of these for every block we reference, it holds the roots and references
59 * to it as well as all of the ref actions that have occurred to it. We never
60 * free it until we unmount the file system in order to make sure re-allocations
61 * are happening properly.
62 */
63struct block_entry {
64 u64 bytenr;
65 u64 len;
66 u64 num_refs;
67 int metadata;
68 int from_disk;
69 struct rb_root roots;
70 struct rb_root refs;
71 struct rb_node node;
72 struct list_head actions;
73};
74
75static struct block_entry *insert_block_entry(struct rb_root *root,
76 struct block_entry *be)
77{
78 struct rb_node **p = &root->rb_node;
79 struct rb_node *parent_node = NULL;
80 struct block_entry *entry;
81
82 while (*p) {
83 parent_node = *p;
84 entry = rb_entry(parent_node, struct block_entry, node);
85 if (entry->bytenr > be->bytenr)
86 p = &(*p)->rb_left;
87 else if (entry->bytenr < be->bytenr)
88 p = &(*p)->rb_right;
89 else
90 return entry;
91 }
92
93 rb_link_node(&be->node, parent_node, p);
94 rb_insert_color(&be->node, root);
95 return NULL;
96}
97
98static struct block_entry *lookup_block_entry(struct rb_root *root, u64 bytenr)
99{
100 struct rb_node *n;
101 struct block_entry *entry = NULL;
102
103 n = root->rb_node;
104 while (n) {
105 entry = rb_entry(n, struct block_entry, node);
106 if (entry->bytenr < bytenr)
107 n = n->rb_right;
108 else if (entry->bytenr > bytenr)
109 n = n->rb_left;
110 else
111 return entry;
112 }
113 return NULL;
114}
115
116static struct root_entry *insert_root_entry(struct rb_root *root,
117 struct root_entry *re)
118{
119 struct rb_node **p = &root->rb_node;
120 struct rb_node *parent_node = NULL;
121 struct root_entry *entry;
122
123 while (*p) {
124 parent_node = *p;
125 entry = rb_entry(parent_node, struct root_entry, node);
126 if (entry->root_objectid > re->root_objectid)
127 p = &(*p)->rb_left;
128 else if (entry->root_objectid < re->root_objectid)
129 p = &(*p)->rb_right;
130 else
131 return entry;
132 }
133
134 rb_link_node(&re->node, parent_node, p);
135 rb_insert_color(&re->node, root);
136 return NULL;
137
138}
139
140static int comp_refs(struct ref_entry *ref1, struct ref_entry *ref2)
141{
142 if (ref1->root_objectid < ref2->root_objectid)
143 return -1;
144 if (ref1->root_objectid > ref2->root_objectid)
145 return 1;
146 if (ref1->parent < ref2->parent)
147 return -1;
148 if (ref1->parent > ref2->parent)
149 return 1;
150 if (ref1->owner < ref2->owner)
151 return -1;
152 if (ref1->owner > ref2->owner)
153 return 1;
154 if (ref1->offset < ref2->offset)
155 return -1;
156 if (ref1->offset > ref2->offset)
157 return 1;
158 return 0;
159}
160
161static struct ref_entry *insert_ref_entry(struct rb_root *root,
162 struct ref_entry *ref)
163{
164 struct rb_node **p = &root->rb_node;
165 struct rb_node *parent_node = NULL;
166 struct ref_entry *entry;
167 int cmp;
168
169 while (*p) {
170 parent_node = *p;
171 entry = rb_entry(parent_node, struct ref_entry, node);
172 cmp = comp_refs(entry, ref);
173 if (cmp > 0)
174 p = &(*p)->rb_left;
175 else if (cmp < 0)
176 p = &(*p)->rb_right;
177 else
178 return entry;
179 }
180
181 rb_link_node(&ref->node, parent_node, p);
182 rb_insert_color(&ref->node, root);
183 return NULL;
184
185}
186
187static struct root_entry *lookup_root_entry(struct rb_root *root, u64 objectid)
188{
189 struct rb_node *n;
190 struct root_entry *entry = NULL;
191
192 n = root->rb_node;
193 while (n) {
194 entry = rb_entry(n, struct root_entry, node);
195 if (entry->root_objectid < objectid)
196 n = n->rb_right;
197 else if (entry->root_objectid > objectid)
198 n = n->rb_left;
199 else
200 return entry;
201 }
202 return NULL;
203}
204
205#ifdef CONFIG_STACKTRACE
206static void __save_stack_trace(struct ref_action *ra)
207{
208 ra->trace_len = stack_trace_save(ra->trace, MAX_TRACE, 2);
209}
210
211static void __print_stack_trace(struct btrfs_fs_info *fs_info,
212 struct ref_action *ra)
213{
214 if (ra->trace_len == 0) {
215 btrfs_err(fs_info, " ref-verify: no stacktrace");
216 return;
217 }
218 stack_trace_print(ra->trace, ra->trace_len, 2);
219}
220#else
221static void inline __save_stack_trace(struct ref_action *ra)
222{
223}
224
225static void inline __print_stack_trace(struct btrfs_fs_info *fs_info,
226 struct ref_action *ra)
227{
228 btrfs_err(fs_info, " ref-verify: no stacktrace support");
229}
230#endif
231
232static void free_block_entry(struct block_entry *be)
233{
234 struct root_entry *re;
235 struct ref_entry *ref;
236 struct ref_action *ra;
237 struct rb_node *n;
238
239 while ((n = rb_first(&be->roots))) {
240 re = rb_entry(n, struct root_entry, node);
241 rb_erase(&re->node, &be->roots);
242 kfree(re);
243 }
244
245 while((n = rb_first(&be->refs))) {
246 ref = rb_entry(n, struct ref_entry, node);
247 rb_erase(&ref->node, &be->refs);
248 kfree(ref);
249 }
250
251 while (!list_empty(&be->actions)) {
252 ra = list_first_entry(&be->actions, struct ref_action,
253 list);
254 list_del(&ra->list);
255 kfree(ra);
256 }
257 kfree(be);
258}
259
260static struct block_entry *add_block_entry(struct btrfs_fs_info *fs_info,
261 u64 bytenr, u64 len,
262 u64 root_objectid)
263{
264 struct block_entry *be = NULL, *exist;
265 struct root_entry *re = NULL;
266
267 re = kzalloc(sizeof(struct root_entry), GFP_KERNEL);
268 be = kzalloc(sizeof(struct block_entry), GFP_KERNEL);
269 if (!be || !re) {
270 kfree(re);
271 kfree(be);
272 return ERR_PTR(-ENOMEM);
273 }
274 be->bytenr = bytenr;
275 be->len = len;
276
277 re->root_objectid = root_objectid;
278 re->num_refs = 0;
279
280 spin_lock(&fs_info->ref_verify_lock);
281 exist = insert_block_entry(&fs_info->block_tree, be);
282 if (exist) {
283 if (root_objectid) {
284 struct root_entry *exist_re;
285
286 exist_re = insert_root_entry(&exist->roots, re);
287 if (exist_re)
288 kfree(re);
289 } else {
290 kfree(re);
291 }
292 kfree(be);
293 return exist;
294 }
295
296 be->num_refs = 0;
297 be->metadata = 0;
298 be->from_disk = 0;
299 be->roots = RB_ROOT;
300 be->refs = RB_ROOT;
301 INIT_LIST_HEAD(&be->actions);
302 if (root_objectid)
303 insert_root_entry(&be->roots, re);
304 else
305 kfree(re);
306 return be;
307}
308
309static int add_tree_block(struct btrfs_fs_info *fs_info, u64 ref_root,
310 u64 parent, u64 bytenr, int level)
311{
312 struct block_entry *be;
313 struct root_entry *re;
314 struct ref_entry *ref = NULL, *exist;
315
316 ref = kmalloc(sizeof(struct ref_entry), GFP_KERNEL);
317 if (!ref)
318 return -ENOMEM;
319
320 if (parent)
321 ref->root_objectid = 0;
322 else
323 ref->root_objectid = ref_root;
324 ref->parent = parent;
325 ref->owner = level;
326 ref->offset = 0;
327 ref->num_refs = 1;
328
329 be = add_block_entry(fs_info, bytenr, fs_info->nodesize, ref_root);
330 if (IS_ERR(be)) {
331 kfree(ref);
332 return PTR_ERR(be);
333 }
334 be->num_refs++;
335 be->from_disk = 1;
336 be->metadata = 1;
337
338 if (!parent) {
339 ASSERT(ref_root);
340 re = lookup_root_entry(&be->roots, ref_root);
341 ASSERT(re);
342 re->num_refs++;
343 }
344 exist = insert_ref_entry(&be->refs, ref);
345 if (exist) {
346 exist->num_refs++;
347 kfree(ref);
348 }
349 spin_unlock(&fs_info->ref_verify_lock);
350
351 return 0;
352}
353
354static int add_shared_data_ref(struct btrfs_fs_info *fs_info,
355 u64 parent, u32 num_refs, u64 bytenr,
356 u64 num_bytes)
357{
358 struct block_entry *be;
359 struct ref_entry *ref;
360
361 ref = kzalloc(sizeof(struct ref_entry), GFP_KERNEL);
362 if (!ref)
363 return -ENOMEM;
364 be = add_block_entry(fs_info, bytenr, num_bytes, 0);
365 if (IS_ERR(be)) {
366 kfree(ref);
367 return PTR_ERR(be);
368 }
369 be->num_refs += num_refs;
370
371 ref->parent = parent;
372 ref->num_refs = num_refs;
373 if (insert_ref_entry(&be->refs, ref)) {
374 spin_unlock(&fs_info->ref_verify_lock);
375 btrfs_err(fs_info, "existing shared ref when reading from disk?");
376 kfree(ref);
377 return -EINVAL;
378 }
379 spin_unlock(&fs_info->ref_verify_lock);
380 return 0;
381}
382
383static int add_extent_data_ref(struct btrfs_fs_info *fs_info,
384 struct extent_buffer *leaf,
385 struct btrfs_extent_data_ref *dref,
386 u64 bytenr, u64 num_bytes)
387{
388 struct block_entry *be;
389 struct ref_entry *ref;
390 struct root_entry *re;
391 u64 ref_root = btrfs_extent_data_ref_root(leaf, dref);
392 u64 owner = btrfs_extent_data_ref_objectid(leaf, dref);
393 u64 offset = btrfs_extent_data_ref_offset(leaf, dref);
394 u32 num_refs = btrfs_extent_data_ref_count(leaf, dref);
395
396 ref = kzalloc(sizeof(struct ref_entry), GFP_KERNEL);
397 if (!ref)
398 return -ENOMEM;
399 be = add_block_entry(fs_info, bytenr, num_bytes, ref_root);
400 if (IS_ERR(be)) {
401 kfree(ref);
402 return PTR_ERR(be);
403 }
404 be->num_refs += num_refs;
405
406 ref->parent = 0;
407 ref->owner = owner;
408 ref->root_objectid = ref_root;
409 ref->offset = offset;
410 ref->num_refs = num_refs;
411 if (insert_ref_entry(&be->refs, ref)) {
412 spin_unlock(&fs_info->ref_verify_lock);
413 btrfs_err(fs_info, "existing ref when reading from disk?");
414 kfree(ref);
415 return -EINVAL;
416 }
417
418 re = lookup_root_entry(&be->roots, ref_root);
419 if (!re) {
420 spin_unlock(&fs_info->ref_verify_lock);
421 btrfs_err(fs_info, "missing root in new block entry?");
422 return -EINVAL;
423 }
424 re->num_refs += num_refs;
425 spin_unlock(&fs_info->ref_verify_lock);
426 return 0;
427}
428
429static int process_extent_item(struct btrfs_fs_info *fs_info,
430 struct btrfs_path *path, struct btrfs_key *key,
431 int slot, int *tree_block_level)
432{
433 struct btrfs_extent_item *ei;
434 struct btrfs_extent_inline_ref *iref;
435 struct btrfs_extent_data_ref *dref;
436 struct btrfs_shared_data_ref *sref;
437 struct extent_buffer *leaf = path->nodes[0];
438 u32 item_size = btrfs_item_size_nr(leaf, slot);
439 unsigned long end, ptr;
440 u64 offset, flags, count;
441 int type, ret;
442
443 ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
444 flags = btrfs_extent_flags(leaf, ei);
445
446 if ((key->type == BTRFS_EXTENT_ITEM_KEY) &&
447 flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
448 struct btrfs_tree_block_info *info;
449
450 info = (struct btrfs_tree_block_info *)(ei + 1);
451 *tree_block_level = btrfs_tree_block_level(leaf, info);
452 iref = (struct btrfs_extent_inline_ref *)(info + 1);
453 } else {
454 if (key->type == BTRFS_METADATA_ITEM_KEY)
455 *tree_block_level = key->offset;
456 iref = (struct btrfs_extent_inline_ref *)(ei + 1);
457 }
458
459 ptr = (unsigned long)iref;
460 end = (unsigned long)ei + item_size;
461 while (ptr < end) {
462 iref = (struct btrfs_extent_inline_ref *)ptr;
463 type = btrfs_extent_inline_ref_type(leaf, iref);
464 offset = btrfs_extent_inline_ref_offset(leaf, iref);
465 switch (type) {
466 case BTRFS_TREE_BLOCK_REF_KEY:
467 ret = add_tree_block(fs_info, offset, 0, key->objectid,
468 *tree_block_level);
469 break;
470 case BTRFS_SHARED_BLOCK_REF_KEY:
471 ret = add_tree_block(fs_info, 0, offset, key->objectid,
472 *tree_block_level);
473 break;
474 case BTRFS_EXTENT_DATA_REF_KEY:
475 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
476 ret = add_extent_data_ref(fs_info, leaf, dref,
477 key->objectid, key->offset);
478 break;
479 case BTRFS_SHARED_DATA_REF_KEY:
480 sref = (struct btrfs_shared_data_ref *)(iref + 1);
481 count = btrfs_shared_data_ref_count(leaf, sref);
482 ret = add_shared_data_ref(fs_info, offset, count,
483 key->objectid, key->offset);
484 break;
485 default:
486 btrfs_err(fs_info, "invalid key type in iref");
487 ret = -EINVAL;
488 break;
489 }
490 if (ret)
491 break;
492 ptr += btrfs_extent_inline_ref_size(type);
493 }
494 return ret;
495}
496
497static int process_leaf(struct btrfs_root *root,
498 struct btrfs_path *path, u64 *bytenr, u64 *num_bytes)
499{
500 struct btrfs_fs_info *fs_info = root->fs_info;
501 struct extent_buffer *leaf = path->nodes[0];
502 struct btrfs_extent_data_ref *dref;
503 struct btrfs_shared_data_ref *sref;
504 u32 count;
505 int i = 0, tree_block_level = 0, ret = 0;
506 struct btrfs_key key;
507 int nritems = btrfs_header_nritems(leaf);
508
509 for (i = 0; i < nritems; i++) {
510 btrfs_item_key_to_cpu(leaf, &key, i);
511 switch (key.type) {
512 case BTRFS_EXTENT_ITEM_KEY:
513 *num_bytes = key.offset;
514 fallthrough;
515 case BTRFS_METADATA_ITEM_KEY:
516 *bytenr = key.objectid;
517 ret = process_extent_item(fs_info, path, &key, i,
518 &tree_block_level);
519 break;
520 case BTRFS_TREE_BLOCK_REF_KEY:
521 ret = add_tree_block(fs_info, key.offset, 0,
522 key.objectid, tree_block_level);
523 break;
524 case BTRFS_SHARED_BLOCK_REF_KEY:
525 ret = add_tree_block(fs_info, 0, key.offset,
526 key.objectid, tree_block_level);
527 break;
528 case BTRFS_EXTENT_DATA_REF_KEY:
529 dref = btrfs_item_ptr(leaf, i,
530 struct btrfs_extent_data_ref);
531 ret = add_extent_data_ref(fs_info, leaf, dref, *bytenr,
532 *num_bytes);
533 break;
534 case BTRFS_SHARED_DATA_REF_KEY:
535 sref = btrfs_item_ptr(leaf, i,
536 struct btrfs_shared_data_ref);
537 count = btrfs_shared_data_ref_count(leaf, sref);
538 ret = add_shared_data_ref(fs_info, key.offset, count,
539 *bytenr, *num_bytes);
540 break;
541 default:
542 break;
543 }
544 if (ret)
545 break;
546 }
547 return ret;
548}
549
550/* Walk down to the leaf from the given level */
551static int walk_down_tree(struct btrfs_root *root, struct btrfs_path *path,
552 int level, u64 *bytenr, u64 *num_bytes)
553{
554 struct btrfs_fs_info *fs_info = root->fs_info;
555 struct extent_buffer *eb;
556 u64 block_bytenr, gen;
557 int ret = 0;
558
559 while (level >= 0) {
560 if (level) {
561 struct btrfs_key first_key;
562
563 block_bytenr = btrfs_node_blockptr(path->nodes[level],
564 path->slots[level]);
565 gen = btrfs_node_ptr_generation(path->nodes[level],
566 path->slots[level]);
567 btrfs_node_key_to_cpu(path->nodes[level], &first_key,
568 path->slots[level]);
569 eb = read_tree_block(fs_info, block_bytenr, gen,
570 level - 1, &first_key);
571 if (IS_ERR(eb))
572 return PTR_ERR(eb);
573 if (!extent_buffer_uptodate(eb)) {
574 free_extent_buffer(eb);
575 return -EIO;
576 }
577 btrfs_tree_read_lock(eb);
578 btrfs_set_lock_blocking_read(eb);
579 path->nodes[level-1] = eb;
580 path->slots[level-1] = 0;
581 path->locks[level-1] = BTRFS_READ_LOCK_BLOCKING;
582 } else {
583 ret = process_leaf(root, path, bytenr, num_bytes);
584 if (ret)
585 break;
586 }
587 level--;
588 }
589 return ret;
590}
591
592/* Walk up to the next node that needs to be processed */
593static int walk_up_tree(struct btrfs_path *path, int *level)
594{
595 int l;
596
597 for (l = 0; l < BTRFS_MAX_LEVEL; l++) {
598 if (!path->nodes[l])
599 continue;
600 if (l) {
601 path->slots[l]++;
602 if (path->slots[l] <
603 btrfs_header_nritems(path->nodes[l])) {
604 *level = l;
605 return 0;
606 }
607 }
608 btrfs_tree_unlock_rw(path->nodes[l], path->locks[l]);
609 free_extent_buffer(path->nodes[l]);
610 path->nodes[l] = NULL;
611 path->slots[l] = 0;
612 path->locks[l] = 0;
613 }
614
615 return 1;
616}
617
618static void dump_ref_action(struct btrfs_fs_info *fs_info,
619 struct ref_action *ra)
620{
621 btrfs_err(fs_info,
622" Ref action %d, root %llu, ref_root %llu, parent %llu, owner %llu, offset %llu, num_refs %llu",
623 ra->action, ra->root, ra->ref.root_objectid, ra->ref.parent,
624 ra->ref.owner, ra->ref.offset, ra->ref.num_refs);
625 __print_stack_trace(fs_info, ra);
626}
627
628/*
629 * Dumps all the information from the block entry to printk, it's going to be
630 * awesome.
631 */
632static void dump_block_entry(struct btrfs_fs_info *fs_info,
633 struct block_entry *be)
634{
635 struct ref_entry *ref;
636 struct root_entry *re;
637 struct ref_action *ra;
638 struct rb_node *n;
639
640 btrfs_err(fs_info,
641"dumping block entry [%llu %llu], num_refs %llu, metadata %d, from disk %d",
642 be->bytenr, be->len, be->num_refs, be->metadata,
643 be->from_disk);
644
645 for (n = rb_first(&be->refs); n; n = rb_next(n)) {
646 ref = rb_entry(n, struct ref_entry, node);
647 btrfs_err(fs_info,
648" ref root %llu, parent %llu, owner %llu, offset %llu, num_refs %llu",
649 ref->root_objectid, ref->parent, ref->owner,
650 ref->offset, ref->num_refs);
651 }
652
653 for (n = rb_first(&be->roots); n; n = rb_next(n)) {
654 re = rb_entry(n, struct root_entry, node);
655 btrfs_err(fs_info, " root entry %llu, num_refs %llu",
656 re->root_objectid, re->num_refs);
657 }
658
659 list_for_each_entry(ra, &be->actions, list)
660 dump_ref_action(fs_info, ra);
661}
662
663/*
664 * btrfs_ref_tree_mod: called when we modify a ref for a bytenr
665 *
666 * This will add an action item to the given bytenr and do sanity checks to make
667 * sure we haven't messed something up. If we are making a new allocation and
668 * this block entry has history we will delete all previous actions as long as
669 * our sanity checks pass as they are no longer needed.
670 */
671int btrfs_ref_tree_mod(struct btrfs_fs_info *fs_info,
672 struct btrfs_ref *generic_ref)
673{
674 struct ref_entry *ref = NULL, *exist;
675 struct ref_action *ra = NULL;
676 struct block_entry *be = NULL;
677 struct root_entry *re = NULL;
678 int action = generic_ref->action;
679 int ret = 0;
680 bool metadata;
681 u64 bytenr = generic_ref->bytenr;
682 u64 num_bytes = generic_ref->len;
683 u64 parent = generic_ref->parent;
684 u64 ref_root;
685 u64 owner;
686 u64 offset;
687
688 if (!btrfs_test_opt(fs_info, REF_VERIFY))
689 return 0;
690
691 if (generic_ref->type == BTRFS_REF_METADATA) {
692 ref_root = generic_ref->tree_ref.root;
693 owner = generic_ref->tree_ref.level;
694 offset = 0;
695 } else {
696 ref_root = generic_ref->data_ref.ref_root;
697 owner = generic_ref->data_ref.ino;
698 offset = generic_ref->data_ref.offset;
699 }
700 metadata = owner < BTRFS_FIRST_FREE_OBJECTID;
701
702 ref = kzalloc(sizeof(struct ref_entry), GFP_NOFS);
703 ra = kmalloc(sizeof(struct ref_action), GFP_NOFS);
704 if (!ra || !ref) {
705 kfree(ref);
706 kfree(ra);
707 ret = -ENOMEM;
708 goto out;
709 }
710
711 if (parent) {
712 ref->parent = parent;
713 } else {
714 ref->root_objectid = ref_root;
715 ref->owner = owner;
716 ref->offset = offset;
717 }
718 ref->num_refs = (action == BTRFS_DROP_DELAYED_REF) ? -1 : 1;
719
720 memcpy(&ra->ref, ref, sizeof(struct ref_entry));
721 /*
722 * Save the extra info from the delayed ref in the ref action to make it
723 * easier to figure out what is happening. The real ref's we add to the
724 * ref tree need to reflect what we save on disk so it matches any
725 * on-disk refs we pre-loaded.
726 */
727 ra->ref.owner = owner;
728 ra->ref.offset = offset;
729 ra->ref.root_objectid = ref_root;
730 __save_stack_trace(ra);
731
732 INIT_LIST_HEAD(&ra->list);
733 ra->action = action;
734 ra->root = generic_ref->real_root;
735
736 /*
737 * This is an allocation, preallocate the block_entry in case we haven't
738 * used it before.
739 */
740 ret = -EINVAL;
741 if (action == BTRFS_ADD_DELAYED_EXTENT) {
742 /*
743 * For subvol_create we'll just pass in whatever the parent root
744 * is and the new root objectid, so let's not treat the passed
745 * in root as if it really has a ref for this bytenr.
746 */
747 be = add_block_entry(fs_info, bytenr, num_bytes, ref_root);
748 if (IS_ERR(be)) {
749 kfree(ref);
750 kfree(ra);
751 ret = PTR_ERR(be);
752 goto out;
753 }
754 be->num_refs++;
755 if (metadata)
756 be->metadata = 1;
757
758 if (be->num_refs != 1) {
759 btrfs_err(fs_info,
760 "re-allocated a block that still has references to it!");
761 dump_block_entry(fs_info, be);
762 dump_ref_action(fs_info, ra);
763 kfree(ref);
764 kfree(ra);
765 goto out_unlock;
766 }
767
768 while (!list_empty(&be->actions)) {
769 struct ref_action *tmp;
770
771 tmp = list_first_entry(&be->actions, struct ref_action,
772 list);
773 list_del(&tmp->list);
774 kfree(tmp);
775 }
776 } else {
777 struct root_entry *tmp;
778
779 if (!parent) {
780 re = kmalloc(sizeof(struct root_entry), GFP_NOFS);
781 if (!re) {
782 kfree(ref);
783 kfree(ra);
784 ret = -ENOMEM;
785 goto out;
786 }
787 /*
788 * This is the root that is modifying us, so it's the
789 * one we want to lookup below when we modify the
790 * re->num_refs.
791 */
792 ref_root = generic_ref->real_root;
793 re->root_objectid = generic_ref->real_root;
794 re->num_refs = 0;
795 }
796
797 spin_lock(&fs_info->ref_verify_lock);
798 be = lookup_block_entry(&fs_info->block_tree, bytenr);
799 if (!be) {
800 btrfs_err(fs_info,
801"trying to do action %d to bytenr %llu num_bytes %llu but there is no existing entry!",
802 action, (unsigned long long)bytenr,
803 (unsigned long long)num_bytes);
804 dump_ref_action(fs_info, ra);
805 kfree(ref);
806 kfree(ra);
807 goto out_unlock;
808 } else if (be->num_refs == 0) {
809 btrfs_err(fs_info,
810 "trying to do action %d for a bytenr that has 0 total references",
811 action);
812 dump_block_entry(fs_info, be);
813 dump_ref_action(fs_info, ra);
814 kfree(ref);
815 kfree(ra);
816 goto out_unlock;
817 }
818
819 if (!parent) {
820 tmp = insert_root_entry(&be->roots, re);
821 if (tmp) {
822 kfree(re);
823 re = tmp;
824 }
825 }
826 }
827
828 exist = insert_ref_entry(&be->refs, ref);
829 if (exist) {
830 if (action == BTRFS_DROP_DELAYED_REF) {
831 if (exist->num_refs == 0) {
832 btrfs_err(fs_info,
833"dropping a ref for a existing root that doesn't have a ref on the block");
834 dump_block_entry(fs_info, be);
835 dump_ref_action(fs_info, ra);
836 kfree(ref);
837 kfree(ra);
838 goto out_unlock;
839 }
840 exist->num_refs--;
841 if (exist->num_refs == 0) {
842 rb_erase(&exist->node, &be->refs);
843 kfree(exist);
844 }
845 } else if (!be->metadata) {
846 exist->num_refs++;
847 } else {
848 btrfs_err(fs_info,
849"attempting to add another ref for an existing ref on a tree block");
850 dump_block_entry(fs_info, be);
851 dump_ref_action(fs_info, ra);
852 kfree(ref);
853 kfree(ra);
854 goto out_unlock;
855 }
856 kfree(ref);
857 } else {
858 if (action == BTRFS_DROP_DELAYED_REF) {
859 btrfs_err(fs_info,
860"dropping a ref for a root that doesn't have a ref on the block");
861 dump_block_entry(fs_info, be);
862 dump_ref_action(fs_info, ra);
863 kfree(ra);
864 goto out_unlock;
865 }
866 }
867
868 if (!parent && !re) {
869 re = lookup_root_entry(&be->roots, ref_root);
870 if (!re) {
871 /*
872 * This shouldn't happen because we will add our re
873 * above when we lookup the be with !parent, but just in
874 * case catch this case so we don't panic because I
875 * didn't think of some other corner case.
876 */
877 btrfs_err(fs_info, "failed to find root %llu for %llu",
878 generic_ref->real_root, be->bytenr);
879 dump_block_entry(fs_info, be);
880 dump_ref_action(fs_info, ra);
881 kfree(ra);
882 goto out_unlock;
883 }
884 }
885 if (action == BTRFS_DROP_DELAYED_REF) {
886 if (re)
887 re->num_refs--;
888 be->num_refs--;
889 } else if (action == BTRFS_ADD_DELAYED_REF) {
890 be->num_refs++;
891 if (re)
892 re->num_refs++;
893 }
894 list_add_tail(&ra->list, &be->actions);
895 ret = 0;
896out_unlock:
897 spin_unlock(&fs_info->ref_verify_lock);
898out:
899 if (ret)
900 btrfs_clear_opt(fs_info->mount_opt, REF_VERIFY);
901 return ret;
902}
903
904/* Free up the ref cache */
905void btrfs_free_ref_cache(struct btrfs_fs_info *fs_info)
906{
907 struct block_entry *be;
908 struct rb_node *n;
909
910 if (!btrfs_test_opt(fs_info, REF_VERIFY))
911 return;
912
913 spin_lock(&fs_info->ref_verify_lock);
914 while ((n = rb_first(&fs_info->block_tree))) {
915 be = rb_entry(n, struct block_entry, node);
916 rb_erase(&be->node, &fs_info->block_tree);
917 free_block_entry(be);
918 cond_resched_lock(&fs_info->ref_verify_lock);
919 }
920 spin_unlock(&fs_info->ref_verify_lock);
921}
922
923void btrfs_free_ref_tree_range(struct btrfs_fs_info *fs_info, u64 start,
924 u64 len)
925{
926 struct block_entry *be = NULL, *entry;
927 struct rb_node *n;
928
929 if (!btrfs_test_opt(fs_info, REF_VERIFY))
930 return;
931
932 spin_lock(&fs_info->ref_verify_lock);
933 n = fs_info->block_tree.rb_node;
934 while (n) {
935 entry = rb_entry(n, struct block_entry, node);
936 if (entry->bytenr < start) {
937 n = n->rb_right;
938 } else if (entry->bytenr > start) {
939 n = n->rb_left;
940 } else {
941 be = entry;
942 break;
943 }
944 /* We want to get as close to start as possible */
945 if (be == NULL ||
946 (entry->bytenr < start && be->bytenr > start) ||
947 (entry->bytenr < start && entry->bytenr > be->bytenr))
948 be = entry;
949 }
950
951 /*
952 * Could have an empty block group, maybe have something to check for
953 * this case to verify we were actually empty?
954 */
955 if (!be) {
956 spin_unlock(&fs_info->ref_verify_lock);
957 return;
958 }
959
960 n = &be->node;
961 while (n) {
962 be = rb_entry(n, struct block_entry, node);
963 n = rb_next(n);
964 if (be->bytenr < start && be->bytenr + be->len > start) {
965 btrfs_err(fs_info,
966 "block entry overlaps a block group [%llu,%llu]!",
967 start, len);
968 dump_block_entry(fs_info, be);
969 continue;
970 }
971 if (be->bytenr < start)
972 continue;
973 if (be->bytenr >= start + len)
974 break;
975 if (be->bytenr + be->len > start + len) {
976 btrfs_err(fs_info,
977 "block entry overlaps a block group [%llu,%llu]!",
978 start, len);
979 dump_block_entry(fs_info, be);
980 }
981 rb_erase(&be->node, &fs_info->block_tree);
982 free_block_entry(be);
983 }
984 spin_unlock(&fs_info->ref_verify_lock);
985}
986
987/* Walk down all roots and build the ref tree, meant to be called at mount */
988int btrfs_build_ref_tree(struct btrfs_fs_info *fs_info)
989{
990 struct btrfs_path *path;
991 struct extent_buffer *eb;
992 u64 bytenr = 0, num_bytes = 0;
993 int ret, level;
994
995 if (!btrfs_test_opt(fs_info, REF_VERIFY))
996 return 0;
997
998 path = btrfs_alloc_path();
999 if (!path)
1000 return -ENOMEM;
1001
1002 eb = btrfs_read_lock_root_node(fs_info->extent_root);
1003 btrfs_set_lock_blocking_read(eb);
1004 level = btrfs_header_level(eb);
1005 path->nodes[level] = eb;
1006 path->slots[level] = 0;
1007 path->locks[level] = BTRFS_READ_LOCK_BLOCKING;
1008
1009 while (1) {
1010 /*
1011 * We have to keep track of the bytenr/num_bytes we last hit
1012 * because we could have run out of space for an inline ref, and
1013 * would have had to added a ref key item which may appear on a
1014 * different leaf from the original extent item.
1015 */
1016 ret = walk_down_tree(fs_info->extent_root, path, level,
1017 &bytenr, &num_bytes);
1018 if (ret)
1019 break;
1020 ret = walk_up_tree(path, &level);
1021 if (ret < 0)
1022 break;
1023 if (ret > 0) {
1024 ret = 0;
1025 break;
1026 }
1027 }
1028 if (ret) {
1029 btrfs_clear_opt(fs_info->mount_opt, REF_VERIFY);
1030 btrfs_free_ref_cache(fs_info);
1031 }
1032 btrfs_free_path(path);
1033 return ret;
1034}