]> git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/backpointers.c
Update bcachefs sources to dfaf9a6ee2 lib/printbuf: Clean up headers
[bcachefs-tools-debian] / libbcachefs / backpointers.c
1 // SPDX-License-Identifier: GPL-2.0
2 #include "bcachefs.h"
3 #include "alloc_background.h"
4 #include "backpointers.h"
5 #include "btree_cache.h"
6 #include "btree_update.h"
7 #include "error.h"
8
9 #define MAX_EXTENT_COMPRESS_RATIO_SHIFT         10
10
11 /*
12  * Convert from pos in backpointer btree to pos of corresponding bucket in alloc
13  * btree:
14  */
15 static inline struct bpos bp_pos_to_bucket(const struct bch_fs *c,
16                                            struct bpos bp_pos)
17 {
18         struct bch_dev *ca = bch_dev_bkey_exists(c, bp_pos.inode);
19         u64 bucket_sector = bp_pos.offset >> MAX_EXTENT_COMPRESS_RATIO_SHIFT;
20
21         return POS(bp_pos.inode, sector_to_bucket(ca, bucket_sector));
22 }
23
24 /*
25  * Convert from pos in alloc btree + bucket offset to pos in backpointer btree:
26  */
27 static inline struct bpos bucket_pos_to_bp(const struct bch_fs *c,
28                                            struct bpos bucket,
29                                            u64 bucket_offset)
30 {
31         struct bch_dev *ca = bch_dev_bkey_exists(c, bucket.inode);
32
33         return POS(bucket.inode,
34                    (bucket_to_sector(ca, bucket.offset) <<
35                     MAX_EXTENT_COMPRESS_RATIO_SHIFT) + bucket_offset);
36 }
37
38 void bch2_extent_ptr_to_bp(struct bch_fs *c,
39                            enum btree_id btree_id, unsigned level,
40                            struct bkey_s_c k, struct extent_ptr_decoded p,
41                            struct bpos *bucket_pos, struct bch_backpointer *bp)
42 {
43         enum bch_data_type data_type = level ? BCH_DATA_btree : BCH_DATA_user;
44         s64 sectors = level ? btree_sectors(c) : k.k->size;
45         u32 bucket_offset;
46
47         *bucket_pos = PTR_BUCKET_POS_OFFSET(c, &p.ptr, &bucket_offset);
48         *bp = (struct bch_backpointer) {
49                 .btree_id       = btree_id,
50                 .level          = level,
51                 .data_type      = data_type,
52                 .bucket_offset  = ((u64) bucket_offset << MAX_EXTENT_COMPRESS_RATIO_SHIFT) +
53                         p.crc.offset,
54                 .bucket_len     = ptr_disk_sectors(sectors, p),
55                 .pos            = k.k->p,
56         };
57 }
58
59 static bool extent_matches_bp(struct bch_fs *c,
60                               enum btree_id btree_id, unsigned level,
61                               struct bkey_s_c k,
62                               struct bpos bucket,
63                               struct bch_backpointer bp)
64 {
65         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
66         const union bch_extent_entry *entry;
67         struct extent_ptr_decoded p;
68
69         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
70                 struct bpos bucket2;
71                 struct bch_backpointer bp2;
72
73                 if (p.ptr.cached)
74                         continue;
75
76                 bch2_extent_ptr_to_bp(c, btree_id, level, k, p,
77                                       &bucket2, &bp2);
78                 if (!bpos_cmp(bucket, bucket2) &&
79                     !memcmp(&bp, &bp2, sizeof(bp)))
80                         return true;
81         }
82
83         return false;
84 }
85
86 int bch2_backpointer_invalid(const struct bch_fs *c, struct bkey_s_c k,
87                              int rw, struct printbuf *err)
88 {
89         struct bkey_s_c_backpointer bp = bkey_s_c_to_backpointer(k);
90         struct bpos bucket = bp_pos_to_bucket(c, bp.k->p);
91
92         if (bkey_val_bytes(bp.k) < sizeof(*bp.v)) {
93                 prt_str(err, "incorrect value size");
94                 return -EINVAL;
95         }
96
97         if (bpos_cmp(bp.k->p, bucket_pos_to_bp(c, bucket, bp.v->bucket_offset))) {
98                 prt_str(err, "backpointer at wrong pos");
99                 return -EINVAL;
100         }
101
102         return 0;
103 }
104
105 void bch2_backpointer_to_text(struct printbuf *out, const struct bch_backpointer *bp)
106 {
107         prt_printf(out, "btree=%s l=%u offset=%llu:%u len=%u pos=",
108                bch2_btree_ids[bp->btree_id],
109                bp->level,
110                (u64) (bp->bucket_offset >> MAX_EXTENT_COMPRESS_RATIO_SHIFT),
111                (u32) bp->bucket_offset & ~(~0U << MAX_EXTENT_COMPRESS_RATIO_SHIFT),
112                bp->bucket_len);
113         bch2_bpos_to_text(out, bp->pos);
114 }
115
116 void bch2_backpointer_k_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k)
117 {
118         bch2_backpointer_to_text(out, bkey_s_c_to_backpointer(k).v);
119 }
120
121 void bch2_backpointer_swab(struct bkey_s k)
122 {
123         struct bkey_s_backpointer bp = bkey_s_to_backpointer(k);
124
125         bp.v->bucket_offset     = swab32(bp.v->bucket_offset);
126         bp.v->bucket_len        = swab32(bp.v->bucket_len);
127         bch2_bpos_swab(&bp.v->pos);
128 }
129
130 #define BACKPOINTER_OFFSET_MAX  ((1ULL << 40) - 1)
131
132 static inline int backpointer_cmp(struct bch_backpointer l, struct bch_backpointer r)
133 {
134         return cmp_int(l.bucket_offset, r.bucket_offset);
135 }
136
137 static int bch2_backpointer_del_by_offset(struct btree_trans *trans,
138                                           struct bpos bucket,
139                                           u64 bp_offset,
140                                           struct bch_backpointer bp)
141 {
142         struct bch_fs *c = trans->c;
143         struct btree_iter iter;
144         struct bkey_s_c k;
145         int ret;
146
147         if (bp_offset < BACKPOINTER_OFFSET_MAX) {
148                 struct bch_backpointer *bps;
149                 struct bkey_i_alloc_v4 *a;
150                 unsigned i, nr;
151
152                 bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
153                                      bucket,
154                                      BTREE_ITER_INTENT|
155                                      BTREE_ITER_SLOTS|
156                                      BTREE_ITER_WITH_UPDATES);
157                 k = bch2_btree_iter_peek_slot(&iter);
158                 ret = bkey_err(k);
159                 if (ret)
160                         goto err;
161
162                 if (k.k->type != KEY_TYPE_alloc_v4) {
163                         ret = -ENOENT;
164                         goto err;
165                 }
166
167                 a = bch2_alloc_to_v4_mut(trans, k);
168                 ret = PTR_ERR_OR_ZERO(a);
169                 if (ret)
170                         goto err;
171                 bps = alloc_v4_backpointers(&a->v);
172                 nr = BCH_ALLOC_V4_NR_BACKPOINTERS(&a->v);
173
174                 for (i = 0; i < nr; i++) {
175                         if (bps[i].bucket_offset == bp_offset)
176                                 goto found;
177                         if (bps[i].bucket_offset > bp_offset)
178                                 break;
179                 }
180
181                 ret = -ENOENT;
182                 goto err;
183 found:
184                 if (memcmp(&bps[i], &bp, sizeof(bp))) {
185                         ret = -ENOENT;
186                         goto err;
187                 }
188                 array_remove_item(bps, nr, i);
189                 SET_BCH_ALLOC_V4_NR_BACKPOINTERS(&a->v, nr);
190                 set_alloc_v4_u64s(a);
191                 ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
192         } else {
193                 bp_offset -= BACKPOINTER_OFFSET_MAX;
194
195                 bch2_trans_iter_init(trans, &iter, BTREE_ID_backpointers,
196                                      bucket_pos_to_bp(c, bucket, bp_offset),
197                                      BTREE_ITER_INTENT|
198                                      BTREE_ITER_SLOTS|
199                                      BTREE_ITER_WITH_UPDATES);
200                 k = bch2_btree_iter_peek_slot(&iter);
201                 ret = bkey_err(k);
202                 if (ret)
203                         goto err;
204
205                 if (k.k->type != KEY_TYPE_backpointer ||
206                     memcmp(bkey_s_c_to_backpointer(k).v, &bp, sizeof(bp))) {
207                         ret = -ENOENT;
208                         goto err;
209                 }
210
211                 ret = bch2_btree_delete_at(trans, &iter, 0);
212         }
213 err:
214         bch2_trans_iter_exit(trans, &iter);
215         return ret;
216 }
217
218 int bch2_bucket_backpointer_del(struct btree_trans *trans,
219                                 struct bkey_i_alloc_v4 *a,
220                                 struct bch_backpointer bp,
221                                 struct bkey_s_c orig_k)
222 {
223         struct bch_fs *c = trans->c;
224         struct bch_backpointer *bps = alloc_v4_backpointers(&a->v);
225         unsigned i, nr = BCH_ALLOC_V4_NR_BACKPOINTERS(&a->v);
226         struct btree_iter bp_iter;
227         struct bkey_s_c k;
228         int ret;
229
230         for (i = 0; i < nr; i++) {
231                 int cmp = backpointer_cmp(bps[i], bp) ?:
232                         memcmp(&bps[i], &bp, sizeof(bp));
233                 if (!cmp)
234                         goto found;
235                 if (cmp >= 0)
236                         break;
237         }
238
239         goto btree;
240 found:
241         array_remove_item(bps, nr, i);
242         SET_BCH_ALLOC_V4_NR_BACKPOINTERS(&a->v, nr);
243         set_alloc_v4_u64s(a);
244         return 0;
245 btree:
246         bch2_trans_iter_init(trans, &bp_iter, BTREE_ID_backpointers,
247                              bucket_pos_to_bp(c, a->k.p, bp.bucket_offset),
248                              BTREE_ITER_INTENT|
249                              BTREE_ITER_SLOTS|
250                              BTREE_ITER_WITH_UPDATES);
251         k = bch2_btree_iter_peek_slot(&bp_iter);
252         ret = bkey_err(k);
253         if (ret)
254                 goto err;
255
256         if (k.k->type != KEY_TYPE_backpointer ||
257             memcmp(bkey_s_c_to_backpointer(k).v, &bp, sizeof(bp))) {
258                 struct printbuf buf = PRINTBUF;
259
260                 prt_printf(&buf, "backpointer not found when deleting");
261                 prt_newline(&buf);
262                 printbuf_indent_add(&buf, 2);
263
264                 prt_printf(&buf, "searching for ");
265                 bch2_backpointer_to_text(&buf, &bp);
266                 prt_newline(&buf);
267
268                 prt_printf(&buf, "got ");
269                 bch2_bkey_val_to_text(&buf, c, k);
270                 prt_newline(&buf);
271
272                 prt_str(&buf, "alloc ");
273                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&a->k_i));
274                 prt_newline(&buf);
275
276                 prt_printf(&buf, "for ");
277                 bch2_bkey_val_to_text(&buf, c, orig_k);
278
279                 if (!test_bit(BCH_FS_CHECK_BACKPOINTERS_DONE, &c->flags)) {
280                         bch_err(c, "%s", buf.buf);
281                 } else {
282                         ret = -EIO;
283                         bch2_trans_inconsistent(trans, "%s", buf.buf);
284                 }
285                 printbuf_exit(&buf);
286                 goto err;
287         }
288
289         ret = bch2_btree_delete_at(trans, &bp_iter, 0);
290 err:
291         bch2_trans_iter_exit(trans, &bp_iter);
292         return ret;
293 }
294
295 int bch2_bucket_backpointer_add(struct btree_trans *trans,
296                                 struct bkey_i_alloc_v4 *a,
297                                 struct bch_backpointer bp,
298                                 struct bkey_s_c orig_k)
299 {
300         struct bch_fs *c = trans->c;
301         struct bch_dev *ca;
302         struct bch_backpointer *bps = alloc_v4_backpointers(&a->v);
303         unsigned i, nr = BCH_ALLOC_V4_NR_BACKPOINTERS(&a->v);
304         struct bkey_i_backpointer *bp_k;
305         struct btree_iter bp_iter;
306         struct bkey_s_c k;
307         int ret;
308
309         /* Check for duplicates: */
310         for (i = 0; i < nr; i++) {
311                 int cmp = backpointer_cmp(bps[i], bp);
312                 if (cmp >= 0)
313                         break;
314         }
315
316         if ((i &&
317              (bps[i - 1].bucket_offset +
318               bps[i - 1].bucket_len > bp.bucket_offset)) ||
319             (i < nr &&
320              (bp.bucket_offset + bp.bucket_len > bps[i].bucket_offset))) {
321                 struct printbuf buf = PRINTBUF;
322
323                 prt_printf(&buf, "overlapping backpointer found when inserting ");
324                 bch2_backpointer_to_text(&buf, &bp);
325                 prt_newline(&buf);
326                 printbuf_indent_add(&buf, 2);
327
328                 prt_printf(&buf, "into ");
329                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&a->k_i));
330                 prt_newline(&buf);
331
332                 prt_printf(&buf, "for ");
333                 bch2_bkey_val_to_text(&buf, c, orig_k);
334
335                 if (!test_bit(BCH_FS_CHECK_BACKPOINTERS_DONE, &c->flags))
336                         bch_err(c, "%s", buf.buf);
337                 else {
338                         bch2_trans_inconsistent(trans, "%s", buf.buf);
339                         printbuf_exit(&buf);
340                         return -EIO;
341                 }
342         }
343
344         if (nr < BCH_ALLOC_V4_NR_BACKPOINTERS_MAX) {
345                 array_insert_item(bps, nr, i, bp);
346                 SET_BCH_ALLOC_V4_NR_BACKPOINTERS(&a->v, nr);
347                 set_alloc_v4_u64s(a);
348                 return 0;
349         }
350
351         /* Overflow: use backpointer btree */
352         bp_k = bch2_trans_kmalloc(trans, sizeof(*bp_k));
353         ret = PTR_ERR_OR_ZERO(bp_k);
354         if (ret)
355                 return ret;
356
357         ca = bch_dev_bkey_exists(c, a->k.p.inode);
358
359         bkey_backpointer_init(&bp_k->k_i);
360         bp_k->k.p = bucket_pos_to_bp(c, a->k.p, bp.bucket_offset);
361         bp_k->v = bp;
362
363         bch2_trans_iter_init(trans, &bp_iter, BTREE_ID_backpointers, bp_k->k.p,
364                              BTREE_ITER_INTENT|
365                              BTREE_ITER_SLOTS|
366                              BTREE_ITER_WITH_UPDATES);
367         k = bch2_btree_iter_peek_slot(&bp_iter);
368         ret = bkey_err(k);
369         if (ret)
370                 goto err;
371
372         if (k.k->type) {
373                 struct printbuf buf = PRINTBUF;
374
375                 prt_printf(&buf, "existing btree backpointer key found when inserting ");
376                 bch2_backpointer_to_text(&buf, &bp);
377                 prt_newline(&buf);
378                 printbuf_indent_add(&buf, 2);
379
380                 prt_printf(&buf, "found ");
381                 bch2_bkey_val_to_text(&buf, c, k);
382                 prt_newline(&buf);
383
384                 prt_printf(&buf, "for ");
385                 bch2_bkey_val_to_text(&buf, c, orig_k);
386
387                 if (!test_bit(BCH_FS_CHECK_BACKPOINTERS_DONE, &c->flags))
388                         bch_err(c, "%s", buf.buf);
389                 else {
390                         bch2_trans_inconsistent(trans, "%s", buf.buf);
391                         printbuf_exit(&buf);
392                         ret = -EIO;
393                         goto err;
394                 }
395         }
396
397         ret = bch2_trans_update(trans, &bp_iter, &bp_k->k_i, 0);
398 err:
399         bch2_trans_iter_exit(trans, &bp_iter);
400         return ret;
401 }
402
403 /*
404  * Find the next backpointer >= *bp_offset:
405  */
406 int bch2_get_next_backpointer(struct btree_trans *trans,
407                               struct bpos bucket, int gen,
408                               u64 *bp_offset,
409                               struct bch_backpointer *dst)
410 {
411         struct bch_fs *c = trans->c;
412         struct bpos bp_pos =
413                 bucket_pos_to_bp(c, bucket,
414                                 max(*bp_offset, BACKPOINTER_OFFSET_MAX) - BACKPOINTER_OFFSET_MAX);
415         struct bpos bp_end_pos =
416                 bucket_pos_to_bp(c, bpos_nosnap_successor(bucket), 0);
417         struct btree_iter alloc_iter, bp_iter = { NULL };
418         struct bkey_s_c k;
419         struct bkey_s_c_alloc_v4 a;
420         size_t i;
421         int ret;
422
423         bch2_trans_iter_init(trans, &alloc_iter, BTREE_ID_alloc,
424                              bucket, BTREE_ITER_CACHED);
425         k = bch2_btree_iter_peek_slot(&alloc_iter);
426         ret = bkey_err(k);
427         if (ret)
428                 goto out;
429
430         if (k.k->type != KEY_TYPE_alloc_v4)
431                 goto done;
432
433         a = bkey_s_c_to_alloc_v4(k);
434         if (gen >= 0 && a.v->gen != gen)
435                 goto done;
436
437         for (i = 0; i < BCH_ALLOC_V4_NR_BACKPOINTERS(a.v); i++) {
438                 if (alloc_v4_backpointers_c(a.v)[i].bucket_offset < *bp_offset)
439                         continue;
440
441                 *dst = alloc_v4_backpointers_c(a.v)[i];
442                 *bp_offset = dst->bucket_offset;
443                 goto out;
444         }
445
446         for_each_btree_key_norestart(trans, bp_iter, BTREE_ID_backpointers,
447                                      bp_pos, 0, k, ret) {
448                 if (bpos_cmp(k.k->p, bp_end_pos) >= 0)
449                         break;
450
451                 if (k.k->type != KEY_TYPE_backpointer)
452                         continue;
453
454                 *dst = *bkey_s_c_to_backpointer(k).v;
455                 *bp_offset = dst->bucket_offset + BACKPOINTER_OFFSET_MAX;
456                 goto out;
457         }
458 done:
459         *bp_offset = U64_MAX;
460 out:
461         bch2_trans_iter_exit(trans, &bp_iter);
462         bch2_trans_iter_exit(trans, &alloc_iter);
463         return ret;
464 }
465
466 static void backpointer_not_found(struct btree_trans *trans,
467                                   struct bpos bucket,
468                                   u64 bp_offset,
469                                   struct bch_backpointer bp,
470                                   struct bkey_s_c k,
471                                   const char *thing_it_points_to)
472 {
473         struct bch_fs *c = trans->c;
474         struct printbuf buf = PRINTBUF;
475
476         prt_printf(&buf, "backpointer doesn't match %s it points to:\n  ",
477                    thing_it_points_to);
478         prt_printf(&buf, "bucket: ");
479         bch2_bpos_to_text(&buf, bucket);
480         prt_printf(&buf, "\n  ");
481
482         if (bp_offset >= BACKPOINTER_OFFSET_MAX) {
483                 struct bpos bp_pos =
484                         bucket_pos_to_bp(c, bucket,
485                                         bp_offset - BACKPOINTER_OFFSET_MAX);
486                 prt_printf(&buf, "backpointer pos: ");
487                 bch2_bpos_to_text(&buf, bp_pos);
488                 prt_printf(&buf, "\n  ");
489         }
490
491         bch2_backpointer_to_text(&buf, &bp);
492         prt_printf(&buf, "\n  ");
493         bch2_bkey_val_to_text(&buf, c, k);
494         if (!test_bit(BCH_FS_CHECK_BACKPOINTERS_DONE, &c->flags))
495                 bch_err_ratelimited(c, "%s", buf.buf);
496         else
497                 bch2_trans_inconsistent(trans, "%s", buf.buf);
498
499         printbuf_exit(&buf);
500 }
501
502 struct bkey_s_c bch2_backpointer_get_key(struct btree_trans *trans,
503                                          struct btree_iter *iter,
504                                          struct bpos bucket,
505                                          u64 bp_offset,
506                                          struct bch_backpointer bp)
507 {
508         struct bch_fs *c = trans->c;
509         struct bkey_s_c k;
510
511         bch2_trans_node_iter_init(trans, iter,
512                                   bp.btree_id,
513                                   bp.pos,
514                                   0,
515                                   min(bp.level, c->btree_roots[bp.btree_id].level),
516                                   0);
517         k = bch2_btree_iter_peek_slot(iter);
518         if (bkey_err(k)) {
519                 bch2_trans_iter_exit(trans, iter);
520                 return k;
521         }
522
523         if (bp.level == c->btree_roots[bp.btree_id].level + 1)
524                 k = bkey_i_to_s_c(&c->btree_roots[bp.btree_id].key);
525
526         if (extent_matches_bp(c, bp.btree_id, bp.level, k, bucket, bp))
527                 return k;
528
529         bch2_trans_iter_exit(trans, iter);
530
531         if (bp.level) {
532                 /*
533                  * If a backpointer for a btree node wasn't found, it may be
534                  * because it was overwritten by a new btree node that hasn't
535                  * been written out yet - backpointer_get_node() checks for
536                  * this:
537                  */
538                 bch2_backpointer_get_node(trans, iter, bucket, bp_offset, bp);
539                 bch2_trans_iter_exit(trans, iter);
540                 return bkey_s_c_null;
541         }
542
543         backpointer_not_found(trans, bucket, bp_offset, bp, k, "extent");
544         return bkey_s_c_null;
545 }
546
547 struct btree *bch2_backpointer_get_node(struct btree_trans *trans,
548                                         struct btree_iter *iter,
549                                         struct bpos bucket,
550                                         u64 bp_offset,
551                                         struct bch_backpointer bp)
552 {
553         struct bch_fs *c = trans->c;
554         struct btree *b;
555
556         BUG_ON(!bp.level);
557
558         bch2_trans_node_iter_init(trans, iter,
559                                   bp.btree_id,
560                                   bp.pos,
561                                   0,
562                                   bp.level - 1,
563                                   0);
564         b = bch2_btree_iter_peek_node(iter);
565         if (IS_ERR(b))
566                 goto err;
567
568         if (extent_matches_bp(c, bp.btree_id, bp.level,
569                               bkey_i_to_s_c(&b->key),
570                               bucket, bp))
571                 return b;
572
573         if (btree_node_will_make_reachable(b)) {
574                 b = ERR_PTR(-BCH_ERR_backpointer_to_overwritten_btree_node);
575         } else {
576                 backpointer_not_found(trans, bucket, bp_offset, bp,
577                                       bkey_i_to_s_c(&b->key), "btree node");
578                 b = NULL;
579         }
580 err:
581         bch2_trans_iter_exit(trans, iter);
582         return b;
583 }
584
585 static int bch2_check_btree_backpointer(struct btree_trans *trans, struct btree_iter *bp_iter,
586                                         struct bkey_s_c k)
587 {
588         struct bch_fs *c = trans->c;
589         struct btree_iter alloc_iter = { NULL };
590         struct bch_dev *ca;
591         struct bkey_s_c alloc_k;
592         struct printbuf buf = PRINTBUF;
593         int ret = 0;
594
595         if (fsck_err_on(!bch2_dev_exists2(c, k.k->p.inode), c,
596                         "backpointer for mising device:\n%s",
597                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
598                 ret = bch2_btree_delete_at(trans, bp_iter, 0);
599                 goto out;
600         }
601
602         ca = bch_dev_bkey_exists(c, k.k->p.inode);
603
604         bch2_trans_iter_init(trans, &alloc_iter, BTREE_ID_alloc,
605                              bp_pos_to_bucket(c, k.k->p), 0);
606
607         alloc_k = bch2_btree_iter_peek_slot(&alloc_iter);
608         ret = bkey_err(alloc_k);
609         if (ret)
610                 goto out;
611
612         if (fsck_err_on(alloc_k.k->type != KEY_TYPE_alloc_v4, c,
613                         "backpointer for nonexistent alloc key: %llu:%llu:0\n%s",
614                         alloc_iter.pos.inode, alloc_iter.pos.offset,
615                         (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
616                 ret = bch2_btree_delete_at(trans, bp_iter, 0);
617                 goto out;
618         }
619 out:
620 fsck_err:
621         bch2_trans_iter_exit(trans, &alloc_iter);
622         printbuf_exit(&buf);
623         return ret;
624 }
625
626 /* verify that every backpointer has a corresponding alloc key */
627 int bch2_check_btree_backpointers(struct bch_fs *c)
628 {
629         struct btree_iter iter;
630         struct bkey_s_c k;
631
632         return bch2_trans_run(c,
633                 for_each_btree_key_commit(&trans, iter,
634                         BTREE_ID_backpointers, POS_MIN, 0, k,
635                         NULL, NULL, BTREE_INSERT_LAZY_RW|BTREE_INSERT_NOFAIL,
636                   bch2_check_btree_backpointer(&trans, &iter, k)));
637 }
638
639 static int check_bp_exists(struct btree_trans *trans,
640                            struct bpos bucket_pos,
641                            struct bch_backpointer bp,
642                            struct bkey_s_c orig_k)
643 {
644         struct bch_fs *c = trans->c;
645         struct btree_iter alloc_iter, bp_iter = { NULL };
646         struct printbuf buf = PRINTBUF;
647         struct bkey_s_c alloc_k, bp_k;
648         int ret;
649
650         bch2_trans_iter_init(trans, &alloc_iter, BTREE_ID_alloc, bucket_pos, 0);
651         alloc_k = bch2_btree_iter_peek_slot(&alloc_iter);
652         ret = bkey_err(alloc_k);
653         if (ret)
654                 goto err;
655
656         if (alloc_k.k->type == KEY_TYPE_alloc_v4) {
657                 struct bkey_s_c_alloc_v4 a = bkey_s_c_to_alloc_v4(alloc_k);
658                 const struct bch_backpointer *bps = alloc_v4_backpointers_c(a.v);
659                 unsigned i, nr = BCH_ALLOC_V4_NR_BACKPOINTERS(a.v);
660
661                 for (i = 0; i < nr; i++) {
662                         int cmp = backpointer_cmp(bps[i], bp) ?:
663                                 memcmp(&bps[i], &bp, sizeof(bp));
664                         if (!cmp)
665                                 goto out;
666                         if (cmp >= 0)
667                                 break;
668                 }
669         } else {
670                 goto missing;
671         }
672
673         bch2_trans_iter_init(trans, &bp_iter, BTREE_ID_backpointers,
674                              bucket_pos_to_bp(c, bucket_pos, bp.bucket_offset),
675                              0);
676         bp_k = bch2_btree_iter_peek_slot(&bp_iter);
677         ret = bkey_err(bp_k);
678         if (ret)
679                 goto err;
680
681         if (bp_k.k->type != KEY_TYPE_backpointer ||
682             memcmp(bkey_s_c_to_backpointer(bp_k).v, &bp, sizeof(bp)))
683                 goto missing;
684 out:
685 err:
686 fsck_err:
687         bch2_trans_iter_exit(trans, &bp_iter);
688         bch2_trans_iter_exit(trans, &alloc_iter);
689         printbuf_exit(&buf);
690         return ret;
691 missing:
692         prt_printf(&buf, "missing backpointer for btree=%s l=%u ",
693                bch2_btree_ids[bp.btree_id], bp.level);
694         bch2_bkey_val_to_text(&buf, c, orig_k);
695         prt_printf(&buf, "\nin alloc key ");
696         bch2_bkey_val_to_text(&buf, c, alloc_k);
697
698         if (c->sb.version < bcachefs_metadata_version_backpointers ||
699             c->opts.reconstruct_alloc ||
700             fsck_err(c, "%s", buf.buf)) {
701                 struct bkey_i_alloc_v4 *a = bch2_alloc_to_v4_mut(trans, alloc_k);
702
703                 ret   = PTR_ERR_OR_ZERO(a) ?:
704                         bch2_bucket_backpointer_add(trans, a, bp, orig_k) ?:
705                         bch2_trans_update(trans, &alloc_iter, &a->k_i, 0);
706         }
707
708         goto out;
709 }
710
711 static int check_extent_to_backpointers(struct btree_trans *trans,
712                                         struct btree_iter *iter)
713 {
714         struct bch_fs *c = trans->c;
715         struct bkey_ptrs_c ptrs;
716         const union bch_extent_entry *entry;
717         struct extent_ptr_decoded p;
718         struct bkey_s_c k;
719         int ret;
720
721         k = bch2_btree_iter_peek_all_levels(iter);
722         ret = bkey_err(k);
723         if (ret)
724                 return ret;
725         if (!k.k)
726                 return 0;
727
728         ptrs = bch2_bkey_ptrs_c(k);
729         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
730                 struct bpos bucket_pos;
731                 struct bch_backpointer bp;
732
733                 if (p.ptr.cached)
734                         continue;
735
736                 bch2_extent_ptr_to_bp(c, iter->btree_id, iter->path->level,
737                                       k, p, &bucket_pos, &bp);
738
739                 ret = check_bp_exists(trans, bucket_pos, bp, k);
740                 if (ret)
741                         return ret;
742         }
743
744         return 0;
745 }
746
747 static int check_btree_root_to_backpointers(struct btree_trans *trans,
748                                             enum btree_id btree_id)
749 {
750         struct bch_fs *c = trans->c;
751         struct btree_iter iter;
752         struct btree *b;
753         struct bkey_s_c k;
754         struct bkey_ptrs_c ptrs;
755         struct extent_ptr_decoded p;
756         const union bch_extent_entry *entry;
757         int ret;
758
759         bch2_trans_node_iter_init(trans, &iter, btree_id, POS_MIN, 0,
760                                   c->btree_roots[btree_id].level, 0);
761         b = bch2_btree_iter_peek_node(&iter);
762         ret = PTR_ERR_OR_ZERO(b);
763         if (ret)
764                 goto err;
765
766         BUG_ON(b != btree_node_root(c, b));
767
768         k = bkey_i_to_s_c(&b->key);
769         ptrs = bch2_bkey_ptrs_c(k);
770         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
771                 struct bpos bucket_pos;
772                 struct bch_backpointer bp;
773
774                 if (p.ptr.cached)
775                         continue;
776
777                 bch2_extent_ptr_to_bp(c, iter.btree_id, iter.path->level + 1,
778                                       k, p, &bucket_pos, &bp);
779
780                 ret = check_bp_exists(trans, bucket_pos, bp, k);
781                 if (ret)
782                         goto err;
783         }
784 err:
785         bch2_trans_iter_exit(trans, &iter);
786         return ret;
787 }
788
789 int bch2_check_extents_to_backpointers(struct bch_fs *c)
790 {
791         struct btree_trans trans;
792         struct btree_iter iter;
793         enum btree_id btree_id;
794         int ret = 0;
795
796         bch2_trans_init(&trans, c, 0, 0);
797         for (btree_id = 0; btree_id < BTREE_ID_NR; btree_id++) {
798                 bch2_trans_node_iter_init(&trans, &iter, btree_id, POS_MIN, 0,
799                                           0,
800                                           BTREE_ITER_ALL_LEVELS|
801                                           BTREE_ITER_PREFETCH);
802
803                 do {
804                         ret = commit_do(&trans, NULL, NULL,
805                                               BTREE_INSERT_LAZY_RW|
806                                               BTREE_INSERT_NOFAIL,
807                                               check_extent_to_backpointers(&trans, &iter));
808                         if (ret)
809                                 break;
810                 } while (!bch2_btree_iter_advance(&iter));
811
812                 bch2_trans_iter_exit(&trans, &iter);
813
814                 if (ret)
815                         break;
816
817                 ret = commit_do(&trans, NULL, NULL,
818                                       BTREE_INSERT_LAZY_RW|
819                                       BTREE_INSERT_NOFAIL,
820                                       check_btree_root_to_backpointers(&trans, btree_id));
821                 if (ret)
822                         break;
823         }
824         bch2_trans_exit(&trans);
825         return ret;
826 }
827
828 static int check_one_backpointer(struct btree_trans *trans,
829                                  struct bpos bucket,
830                                  u64 *bp_offset)
831 {
832         struct btree_iter iter;
833         struct bch_backpointer bp;
834         struct bkey_s_c k;
835         struct printbuf buf = PRINTBUF;
836         int ret;
837
838         ret = bch2_get_next_backpointer(trans, bucket, -1,
839                                         bp_offset, &bp);
840         if (ret || *bp_offset == U64_MAX)
841                 return ret;
842
843         k = bch2_backpointer_get_key(trans, &iter, bucket, *bp_offset, bp);
844         ret = bkey_err(k);
845         if (ret == -BCH_ERR_backpointer_to_overwritten_btree_node)
846                 return 0;
847         if (ret)
848                 return ret;
849
850         if (fsck_err_on(!k.k, trans->c,
851                         "%s backpointer points to missing extent\n%s",
852                         *bp_offset < BACKPOINTER_OFFSET_MAX ? "alloc" : "btree",
853                         (bch2_backpointer_to_text(&buf, &bp), buf.buf))) {
854                 ret = bch2_backpointer_del_by_offset(trans, bucket, *bp_offset, bp);
855                 if (ret == -ENOENT)
856                         bch_err(trans->c, "backpointer at %llu not found", *bp_offset);
857         }
858
859         bch2_trans_iter_exit(trans, &iter);
860 fsck_err:
861         printbuf_exit(&buf);
862         return ret;
863 }
864
865 int bch2_check_backpointers_to_extents(struct bch_fs *c)
866 {
867         struct btree_trans trans;
868         struct btree_iter iter;
869         struct bkey_s_c k;
870         int ret = 0;
871
872         bch2_trans_init(&trans, c, 0, 0);
873         for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
874                            BTREE_ITER_PREFETCH, k, ret) {
875                 u64 bp_offset = 0;
876
877                 while (!(ret = commit_do(&trans, NULL, NULL,
878                                                BTREE_INSERT_LAZY_RW|
879                                                BTREE_INSERT_NOFAIL,
880                                 check_one_backpointer(&trans, iter.pos, &bp_offset))) &&
881                        bp_offset < U64_MAX)
882                         bp_offset++;
883
884                 if (ret)
885                         break;
886         }
887         bch2_trans_iter_exit(&trans, &iter);
888         bch2_trans_exit(&trans);
889         return ret < 0 ? ret : 0;
890 }