]> git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/backpointers.c
Update bcachefs sources to 83edfdeb29 bcachefs: Run check_extents_to_backpointers...
[bcachefs-tools-debian] / libbcachefs / backpointers.c
1 // SPDX-License-Identifier: GPL-2.0
2 #include "bcachefs.h"
3 #include "bbpos.h"
4 #include "alloc_background.h"
5 #include "backpointers.h"
6 #include "btree_cache.h"
7 #include "btree_update.h"
8 #include "error.h"
9
10 #include <linux/mm.h>
11
12 #define MAX_EXTENT_COMPRESS_RATIO_SHIFT         10
13
14 /*
15  * Convert from pos in backpointer btree to pos of corresponding bucket in alloc
16  * btree:
17  */
18 static inline struct bpos bp_pos_to_bucket(const struct bch_fs *c,
19                                            struct bpos bp_pos)
20 {
21         struct bch_dev *ca = bch_dev_bkey_exists(c, bp_pos.inode);
22         u64 bucket_sector = bp_pos.offset >> MAX_EXTENT_COMPRESS_RATIO_SHIFT;
23
24         return POS(bp_pos.inode, sector_to_bucket(ca, bucket_sector));
25 }
26
27 /*
28  * Convert from pos in alloc btree + bucket offset to pos in backpointer btree:
29  */
30 static inline struct bpos bucket_pos_to_bp(const struct bch_fs *c,
31                                            struct bpos bucket,
32                                            u64 bucket_offset)
33 {
34         struct bch_dev *ca = bch_dev_bkey_exists(c, bucket.inode);
35         struct bpos ret;
36
37         ret = POS(bucket.inode,
38                   (bucket_to_sector(ca, bucket.offset) <<
39                    MAX_EXTENT_COMPRESS_RATIO_SHIFT) + bucket_offset);
40
41         BUG_ON(bkey_cmp(bucket, bp_pos_to_bucket(c, ret)));
42
43         return ret;
44 }
45
46 void bch2_extent_ptr_to_bp(struct bch_fs *c,
47                            enum btree_id btree_id, unsigned level,
48                            struct bkey_s_c k, struct extent_ptr_decoded p,
49                            struct bpos *bucket_pos, struct bch_backpointer *bp)
50 {
51         enum bch_data_type data_type = level ? BCH_DATA_btree : BCH_DATA_user;
52         s64 sectors = level ? btree_sectors(c) : k.k->size;
53         u32 bucket_offset;
54
55         *bucket_pos = PTR_BUCKET_POS_OFFSET(c, &p.ptr, &bucket_offset);
56         *bp = (struct bch_backpointer) {
57                 .btree_id       = btree_id,
58                 .level          = level,
59                 .data_type      = data_type,
60                 .bucket_offset  = ((u64) bucket_offset << MAX_EXTENT_COMPRESS_RATIO_SHIFT) +
61                         p.crc.offset,
62                 .bucket_len     = ptr_disk_sectors(sectors, p),
63                 .pos            = k.k->p,
64         };
65 }
66
67 static bool extent_matches_bp(struct bch_fs *c,
68                               enum btree_id btree_id, unsigned level,
69                               struct bkey_s_c k,
70                               struct bpos bucket,
71                               struct bch_backpointer bp)
72 {
73         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
74         const union bch_extent_entry *entry;
75         struct extent_ptr_decoded p;
76
77         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
78                 struct bpos bucket2;
79                 struct bch_backpointer bp2;
80
81                 if (p.ptr.cached)
82                         continue;
83
84                 bch2_extent_ptr_to_bp(c, btree_id, level, k, p,
85                                       &bucket2, &bp2);
86                 if (!bpos_cmp(bucket, bucket2) &&
87                     !memcmp(&bp, &bp2, sizeof(bp)))
88                         return true;
89         }
90
91         return false;
92 }
93
94 int bch2_backpointer_invalid(const struct bch_fs *c, struct bkey_s_c k,
95                              int rw, struct printbuf *err)
96 {
97         struct bkey_s_c_backpointer bp = bkey_s_c_to_backpointer(k);
98         struct bpos bucket = bp_pos_to_bucket(c, bp.k->p);
99
100         if (bkey_val_bytes(bp.k) < sizeof(*bp.v)) {
101                 prt_str(err, "incorrect value size");
102                 return -EINVAL;
103         }
104
105         if (bpos_cmp(bp.k->p, bucket_pos_to_bp(c, bucket, bp.v->bucket_offset))) {
106                 prt_str(err, "backpointer at wrong pos");
107                 return -EINVAL;
108         }
109
110         return 0;
111 }
112
113 void bch2_backpointer_to_text(struct printbuf *out, const struct bch_backpointer *bp)
114 {
115         prt_printf(out, "btree=%s l=%u offset=%llu:%u len=%u pos=",
116                bch2_btree_ids[bp->btree_id],
117                bp->level,
118                (u64) (bp->bucket_offset >> MAX_EXTENT_COMPRESS_RATIO_SHIFT),
119                (u32) bp->bucket_offset & ~(~0U << MAX_EXTENT_COMPRESS_RATIO_SHIFT),
120                bp->bucket_len);
121         bch2_bpos_to_text(out, bp->pos);
122 }
123
124 void bch2_backpointer_k_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k)
125 {
126         bch2_backpointer_to_text(out, bkey_s_c_to_backpointer(k).v);
127 }
128
129 void bch2_backpointer_swab(struct bkey_s k)
130 {
131         struct bkey_s_backpointer bp = bkey_s_to_backpointer(k);
132
133         bp.v->bucket_offset     = swab32(bp.v->bucket_offset);
134         bp.v->bucket_len        = swab32(bp.v->bucket_len);
135         bch2_bpos_swab(&bp.v->pos);
136 }
137
138 #define BACKPOINTER_OFFSET_MAX  ((1ULL << 40) - 1)
139
140 static inline int backpointer_cmp(struct bch_backpointer l, struct bch_backpointer r)
141 {
142         return cmp_int(l.bucket_offset, r.bucket_offset);
143 }
144
145 static int bch2_backpointer_del_by_offset(struct btree_trans *trans,
146                                           struct bpos bucket,
147                                           u64 bp_offset,
148                                           struct bch_backpointer bp)
149 {
150         struct bch_fs *c = trans->c;
151         struct btree_iter iter;
152         struct bkey_s_c k;
153         int ret;
154
155         if (bp_offset < BACKPOINTER_OFFSET_MAX) {
156                 struct bch_backpointer *bps;
157                 struct bkey_i_alloc_v4 *a;
158                 unsigned i, nr;
159
160                 bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
161                                      bucket,
162                                      BTREE_ITER_INTENT|
163                                      BTREE_ITER_SLOTS|
164                                      BTREE_ITER_WITH_UPDATES);
165                 k = bch2_btree_iter_peek_slot(&iter);
166                 ret = bkey_err(k);
167                 if (ret)
168                         goto err;
169
170                 if (k.k->type != KEY_TYPE_alloc_v4) {
171                         ret = -ENOENT;
172                         goto err;
173                 }
174
175                 a = bch2_alloc_to_v4_mut(trans, k);
176                 ret = PTR_ERR_OR_ZERO(a);
177                 if (ret)
178                         goto err;
179                 bps = alloc_v4_backpointers(&a->v);
180                 nr = BCH_ALLOC_V4_NR_BACKPOINTERS(&a->v);
181
182                 for (i = 0; i < nr; i++) {
183                         if (bps[i].bucket_offset == bp_offset)
184                                 goto found;
185                         if (bps[i].bucket_offset > bp_offset)
186                                 break;
187                 }
188
189                 ret = -ENOENT;
190                 goto err;
191 found:
192                 if (memcmp(&bps[i], &bp, sizeof(bp))) {
193                         ret = -ENOENT;
194                         goto err;
195                 }
196                 array_remove_item(bps, nr, i);
197                 SET_BCH_ALLOC_V4_NR_BACKPOINTERS(&a->v, nr);
198                 set_alloc_v4_u64s(a);
199                 ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
200         } else {
201                 bp_offset -= BACKPOINTER_OFFSET_MAX;
202
203                 bch2_trans_iter_init(trans, &iter, BTREE_ID_backpointers,
204                                      bucket_pos_to_bp(c, bucket, bp_offset),
205                                      BTREE_ITER_INTENT|
206                                      BTREE_ITER_SLOTS|
207                                      BTREE_ITER_WITH_UPDATES);
208                 k = bch2_btree_iter_peek_slot(&iter);
209                 ret = bkey_err(k);
210                 if (ret)
211                         goto err;
212
213                 if (k.k->type != KEY_TYPE_backpointer ||
214                     memcmp(bkey_s_c_to_backpointer(k).v, &bp, sizeof(bp))) {
215                         ret = -ENOENT;
216                         goto err;
217                 }
218
219                 ret = bch2_btree_delete_at(trans, &iter, 0);
220         }
221 err:
222         bch2_trans_iter_exit(trans, &iter);
223         return ret;
224 }
225
226 int bch2_bucket_backpointer_del(struct btree_trans *trans,
227                                 struct bkey_i_alloc_v4 *a,
228                                 struct bch_backpointer bp,
229                                 struct bkey_s_c orig_k)
230 {
231         struct bch_fs *c = trans->c;
232         struct bch_backpointer *bps = alloc_v4_backpointers(&a->v);
233         unsigned i, nr = BCH_ALLOC_V4_NR_BACKPOINTERS(&a->v);
234         struct btree_iter bp_iter;
235         struct bkey_s_c k;
236         int ret;
237
238         for (i = 0; i < nr; i++) {
239                 int cmp = backpointer_cmp(bps[i], bp) ?:
240                         memcmp(&bps[i], &bp, sizeof(bp));
241                 if (!cmp)
242                         goto found;
243                 if (cmp >= 0)
244                         break;
245         }
246
247         goto btree;
248 found:
249         array_remove_item(bps, nr, i);
250         SET_BCH_ALLOC_V4_NR_BACKPOINTERS(&a->v, nr);
251         set_alloc_v4_u64s(a);
252         return 0;
253 btree:
254         bch2_trans_iter_init(trans, &bp_iter, BTREE_ID_backpointers,
255                              bucket_pos_to_bp(c, a->k.p, bp.bucket_offset),
256                              BTREE_ITER_INTENT|
257                              BTREE_ITER_SLOTS|
258                              BTREE_ITER_WITH_UPDATES);
259         k = bch2_btree_iter_peek_slot(&bp_iter);
260         ret = bkey_err(k);
261         if (ret)
262                 goto err;
263
264         if (k.k->type != KEY_TYPE_backpointer ||
265             memcmp(bkey_s_c_to_backpointer(k).v, &bp, sizeof(bp))) {
266                 struct printbuf buf = PRINTBUF;
267
268                 prt_printf(&buf, "backpointer not found when deleting");
269                 prt_newline(&buf);
270                 printbuf_indent_add(&buf, 2);
271
272                 prt_printf(&buf, "searching for ");
273                 bch2_backpointer_to_text(&buf, &bp);
274                 prt_newline(&buf);
275
276                 prt_printf(&buf, "got ");
277                 bch2_bkey_val_to_text(&buf, c, k);
278                 prt_newline(&buf);
279
280                 prt_str(&buf, "alloc ");
281                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&a->k_i));
282                 prt_newline(&buf);
283
284                 prt_printf(&buf, "for ");
285                 bch2_bkey_val_to_text(&buf, c, orig_k);
286
287                 if (!test_bit(BCH_FS_CHECK_BACKPOINTERS_DONE, &c->flags)) {
288                         bch_err(c, "%s", buf.buf);
289                 } else {
290                         ret = -EIO;
291                         bch2_trans_inconsistent(trans, "%s", buf.buf);
292                 }
293                 printbuf_exit(&buf);
294                 goto err;
295         }
296
297         ret = bch2_btree_delete_at(trans, &bp_iter, 0);
298 err:
299         bch2_trans_iter_exit(trans, &bp_iter);
300         return ret;
301 }
302
303 int bch2_bucket_backpointer_add(struct btree_trans *trans,
304                                 struct bkey_i_alloc_v4 *a,
305                                 struct bch_backpointer bp,
306                                 struct bkey_s_c orig_k)
307 {
308         struct bch_fs *c = trans->c;
309         struct bch_dev *ca;
310         struct bch_backpointer *bps = alloc_v4_backpointers(&a->v);
311         unsigned i, nr = BCH_ALLOC_V4_NR_BACKPOINTERS(&a->v);
312         struct bkey_i_backpointer *bp_k;
313         struct btree_iter bp_iter;
314         struct bkey_s_c k;
315         int ret;
316
317         /* Check for duplicates: */
318         for (i = 0; i < nr; i++) {
319                 int cmp = backpointer_cmp(bps[i], bp);
320                 if (cmp >= 0)
321                         break;
322         }
323
324         if ((i &&
325              (bps[i - 1].bucket_offset +
326               bps[i - 1].bucket_len > bp.bucket_offset)) ||
327             (i < nr &&
328              (bp.bucket_offset + bp.bucket_len > bps[i].bucket_offset))) {
329                 struct printbuf buf = PRINTBUF;
330
331                 prt_printf(&buf, "overlapping backpointer found when inserting ");
332                 bch2_backpointer_to_text(&buf, &bp);
333                 prt_newline(&buf);
334                 printbuf_indent_add(&buf, 2);
335
336                 prt_printf(&buf, "into ");
337                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&a->k_i));
338                 prt_newline(&buf);
339
340                 prt_printf(&buf, "for ");
341                 bch2_bkey_val_to_text(&buf, c, orig_k);
342
343                 if (!test_bit(BCH_FS_CHECK_BACKPOINTERS_DONE, &c->flags))
344                         bch_err(c, "%s", buf.buf);
345                 else {
346                         bch2_trans_inconsistent(trans, "%s", buf.buf);
347                         printbuf_exit(&buf);
348                         return -EIO;
349                 }
350         }
351
352         if (nr < BCH_ALLOC_V4_NR_BACKPOINTERS_MAX) {
353                 array_insert_item(bps, nr, i, bp);
354                 SET_BCH_ALLOC_V4_NR_BACKPOINTERS(&a->v, nr);
355                 set_alloc_v4_u64s(a);
356                 return 0;
357         }
358
359         /* Overflow: use backpointer btree */
360         bp_k = bch2_trans_kmalloc(trans, sizeof(*bp_k));
361         ret = PTR_ERR_OR_ZERO(bp_k);
362         if (ret)
363                 return ret;
364
365         ca = bch_dev_bkey_exists(c, a->k.p.inode);
366
367         bkey_backpointer_init(&bp_k->k_i);
368         bp_k->k.p = bucket_pos_to_bp(c, a->k.p, bp.bucket_offset);
369         bp_k->v = bp;
370
371         bch2_trans_iter_init(trans, &bp_iter, BTREE_ID_backpointers, bp_k->k.p,
372                              BTREE_ITER_INTENT|
373                              BTREE_ITER_SLOTS|
374                              BTREE_ITER_WITH_UPDATES);
375         k = bch2_btree_iter_peek_slot(&bp_iter);
376         ret = bkey_err(k);
377         if (ret)
378                 goto err;
379
380         if (k.k->type) {
381                 struct printbuf buf = PRINTBUF;
382
383                 prt_printf(&buf, "existing btree backpointer key found when inserting ");
384                 bch2_backpointer_to_text(&buf, &bp);
385                 prt_newline(&buf);
386                 printbuf_indent_add(&buf, 2);
387
388                 prt_printf(&buf, "found ");
389                 bch2_bkey_val_to_text(&buf, c, k);
390                 prt_newline(&buf);
391
392                 prt_printf(&buf, "for ");
393                 bch2_bkey_val_to_text(&buf, c, orig_k);
394
395                 if (!test_bit(BCH_FS_CHECK_BACKPOINTERS_DONE, &c->flags))
396                         bch_err(c, "%s", buf.buf);
397                 else {
398                         bch2_trans_inconsistent(trans, "%s", buf.buf);
399                         printbuf_exit(&buf);
400                         ret = -EIO;
401                         goto err;
402                 }
403         }
404
405         ret = bch2_trans_update(trans, &bp_iter, &bp_k->k_i, 0);
406 err:
407         bch2_trans_iter_exit(trans, &bp_iter);
408         return ret;
409 }
410
411 /*
412  * Find the next backpointer >= *bp_offset:
413  */
414 int bch2_get_next_backpointer(struct btree_trans *trans,
415                               struct bpos bucket, int gen,
416                               u64 *bp_offset,
417                               struct bch_backpointer *dst)
418 {
419         struct bch_fs *c = trans->c;
420         struct bpos bp_pos, bp_end_pos;
421         struct btree_iter alloc_iter, bp_iter = { NULL };
422         struct bkey_s_c k;
423         struct bkey_s_c_alloc_v4 a;
424         size_t i;
425         int ret;
426
427         if (*bp_offset == U64_MAX)
428                 return 0;
429
430         bp_pos = bucket_pos_to_bp(c, bucket,
431                                   max(*bp_offset, BACKPOINTER_OFFSET_MAX) - BACKPOINTER_OFFSET_MAX);
432         bp_end_pos = bucket_pos_to_bp(c, bpos_nosnap_successor(bucket), 0);
433
434         bch2_trans_iter_init(trans, &alloc_iter, BTREE_ID_alloc,
435                              bucket, BTREE_ITER_CACHED);
436         k = bch2_btree_iter_peek_slot(&alloc_iter);
437         ret = bkey_err(k);
438         if (ret)
439                 goto out;
440
441         if (k.k->type != KEY_TYPE_alloc_v4)
442                 goto done;
443
444         a = bkey_s_c_to_alloc_v4(k);
445         if (gen >= 0 && a.v->gen != gen)
446                 goto done;
447
448         for (i = 0; i < BCH_ALLOC_V4_NR_BACKPOINTERS(a.v); i++) {
449                 if (alloc_v4_backpointers_c(a.v)[i].bucket_offset < *bp_offset)
450                         continue;
451
452                 *dst = alloc_v4_backpointers_c(a.v)[i];
453                 *bp_offset = dst->bucket_offset;
454                 goto out;
455         }
456
457         for_each_btree_key_norestart(trans, bp_iter, BTREE_ID_backpointers,
458                                      bp_pos, 0, k, ret) {
459                 if (bpos_cmp(k.k->p, bp_end_pos) >= 0)
460                         break;
461
462                 if (k.k->type != KEY_TYPE_backpointer)
463                         continue;
464
465                 *dst = *bkey_s_c_to_backpointer(k).v;
466                 *bp_offset = dst->bucket_offset + BACKPOINTER_OFFSET_MAX;
467                 goto out;
468         }
469 done:
470         *bp_offset = U64_MAX;
471 out:
472         bch2_trans_iter_exit(trans, &bp_iter);
473         bch2_trans_iter_exit(trans, &alloc_iter);
474         return ret;
475 }
476
477 static void backpointer_not_found(struct btree_trans *trans,
478                                   struct bpos bucket,
479                                   u64 bp_offset,
480                                   struct bch_backpointer bp,
481                                   struct bkey_s_c k,
482                                   const char *thing_it_points_to)
483 {
484         struct bch_fs *c = trans->c;
485         struct printbuf buf = PRINTBUF;
486
487         prt_printf(&buf, "backpointer doesn't match %s it points to:\n  ",
488                    thing_it_points_to);
489         prt_printf(&buf, "bucket: ");
490         bch2_bpos_to_text(&buf, bucket);
491         prt_printf(&buf, "\n  ");
492
493         if (bp_offset >= BACKPOINTER_OFFSET_MAX) {
494                 struct bpos bp_pos =
495                         bucket_pos_to_bp(c, bucket,
496                                         bp_offset - BACKPOINTER_OFFSET_MAX);
497                 prt_printf(&buf, "backpointer pos: ");
498                 bch2_bpos_to_text(&buf, bp_pos);
499                 prt_printf(&buf, "\n  ");
500         }
501
502         bch2_backpointer_to_text(&buf, &bp);
503         prt_printf(&buf, "\n  ");
504         bch2_bkey_val_to_text(&buf, c, k);
505         if (!test_bit(BCH_FS_CHECK_BACKPOINTERS_DONE, &c->flags))
506                 bch_err_ratelimited(c, "%s", buf.buf);
507         else
508                 bch2_trans_inconsistent(trans, "%s", buf.buf);
509
510         printbuf_exit(&buf);
511 }
512
513 struct bkey_s_c bch2_backpointer_get_key(struct btree_trans *trans,
514                                          struct btree_iter *iter,
515                                          struct bpos bucket,
516                                          u64 bp_offset,
517                                          struct bch_backpointer bp)
518 {
519         struct bch_fs *c = trans->c;
520         struct bkey_s_c k;
521
522         bch2_trans_node_iter_init(trans, iter,
523                                   bp.btree_id,
524                                   bp.pos,
525                                   0,
526                                   min(bp.level, c->btree_roots[bp.btree_id].level),
527                                   0);
528         k = bch2_btree_iter_peek_slot(iter);
529         if (bkey_err(k)) {
530                 bch2_trans_iter_exit(trans, iter);
531                 return k;
532         }
533
534         if (bp.level == c->btree_roots[bp.btree_id].level + 1)
535                 k = bkey_i_to_s_c(&c->btree_roots[bp.btree_id].key);
536
537         if (extent_matches_bp(c, bp.btree_id, bp.level, k, bucket, bp))
538                 return k;
539
540         bch2_trans_iter_exit(trans, iter);
541
542         if (bp.level) {
543                 struct btree *b;
544
545                 /*
546                  * If a backpointer for a btree node wasn't found, it may be
547                  * because it was overwritten by a new btree node that hasn't
548                  * been written out yet - backpointer_get_node() checks for
549                  * this:
550                  */
551                 b = bch2_backpointer_get_node(trans, iter, bucket, bp_offset, bp);
552                 if (!IS_ERR_OR_NULL(b))
553                         return bkey_i_to_s_c(&b->key);
554
555                 bch2_trans_iter_exit(trans, iter);
556
557                 if (IS_ERR(b))
558                         return bkey_s_c_err(PTR_ERR(b));
559                 return bkey_s_c_null;
560         }
561
562         backpointer_not_found(trans, bucket, bp_offset, bp, k, "extent");
563         return bkey_s_c_null;
564 }
565
566 struct btree *bch2_backpointer_get_node(struct btree_trans *trans,
567                                         struct btree_iter *iter,
568                                         struct bpos bucket,
569                                         u64 bp_offset,
570                                         struct bch_backpointer bp)
571 {
572         struct bch_fs *c = trans->c;
573         struct btree *b;
574
575         BUG_ON(!bp.level);
576
577         bch2_trans_node_iter_init(trans, iter,
578                                   bp.btree_id,
579                                   bp.pos,
580                                   0,
581                                   bp.level - 1,
582                                   0);
583         b = bch2_btree_iter_peek_node(iter);
584         if (IS_ERR(b))
585                 goto err;
586
587         if (extent_matches_bp(c, bp.btree_id, bp.level,
588                               bkey_i_to_s_c(&b->key),
589                               bucket, bp))
590                 return b;
591
592         if (btree_node_will_make_reachable(b)) {
593                 b = ERR_PTR(-BCH_ERR_backpointer_to_overwritten_btree_node);
594         } else {
595                 backpointer_not_found(trans, bucket, bp_offset, bp,
596                                       bkey_i_to_s_c(&b->key), "btree node");
597                 b = NULL;
598         }
599 err:
600         bch2_trans_iter_exit(trans, iter);
601         return b;
602 }
603
604 static int bch2_check_btree_backpointer(struct btree_trans *trans, struct btree_iter *bp_iter,
605                                         struct bkey_s_c k)
606 {
607         struct bch_fs *c = trans->c;
608         struct btree_iter alloc_iter = { NULL };
609         struct bch_dev *ca;
610         struct bkey_s_c alloc_k;
611         struct printbuf buf = PRINTBUF;
612         int ret = 0;
613
614         if (fsck_err_on(!bch2_dev_exists2(c, k.k->p.inode), c,
615                         "backpointer for mising device:\n%s",
616                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
617                 ret = bch2_btree_delete_at(trans, bp_iter, 0);
618                 goto out;
619         }
620
621         ca = bch_dev_bkey_exists(c, k.k->p.inode);
622
623         bch2_trans_iter_init(trans, &alloc_iter, BTREE_ID_alloc,
624                              bp_pos_to_bucket(c, k.k->p), 0);
625
626         alloc_k = bch2_btree_iter_peek_slot(&alloc_iter);
627         ret = bkey_err(alloc_k);
628         if (ret)
629                 goto out;
630
631         if (fsck_err_on(alloc_k.k->type != KEY_TYPE_alloc_v4, c,
632                         "backpointer for nonexistent alloc key: %llu:%llu:0\n%s",
633                         alloc_iter.pos.inode, alloc_iter.pos.offset,
634                         (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
635                 ret = bch2_btree_delete_at(trans, bp_iter, 0);
636                 goto out;
637         }
638 out:
639 fsck_err:
640         bch2_trans_iter_exit(trans, &alloc_iter);
641         printbuf_exit(&buf);
642         return ret;
643 }
644
645 /* verify that every backpointer has a corresponding alloc key */
646 int bch2_check_btree_backpointers(struct bch_fs *c)
647 {
648         struct btree_iter iter;
649         struct bkey_s_c k;
650
651         return bch2_trans_run(c,
652                 for_each_btree_key_commit(&trans, iter,
653                         BTREE_ID_backpointers, POS_MIN, 0, k,
654                         NULL, NULL, BTREE_INSERT_LAZY_RW|BTREE_INSERT_NOFAIL,
655                   bch2_check_btree_backpointer(&trans, &iter, k)));
656 }
657
658 static int check_bp_exists(struct btree_trans *trans,
659                            struct bpos bucket_pos,
660                            struct bch_backpointer bp,
661                            struct bkey_s_c orig_k,
662                            struct bpos bucket_start,
663                            struct bpos bucket_end)
664 {
665         struct bch_fs *c = trans->c;
666         struct btree_iter alloc_iter, bp_iter = { NULL };
667         struct printbuf buf = PRINTBUF;
668         struct bkey_s_c alloc_k, bp_k;
669         int ret;
670
671         if (bpos_cmp(bucket_pos, bucket_start) < 0 ||
672             bpos_cmp(bucket_pos, bucket_end) > 0)
673                 return 0;
674
675         bch2_trans_iter_init(trans, &alloc_iter, BTREE_ID_alloc, bucket_pos, 0);
676         alloc_k = bch2_btree_iter_peek_slot(&alloc_iter);
677         ret = bkey_err(alloc_k);
678         if (ret)
679                 goto err;
680
681         if (alloc_k.k->type == KEY_TYPE_alloc_v4) {
682                 struct bkey_s_c_alloc_v4 a = bkey_s_c_to_alloc_v4(alloc_k);
683                 const struct bch_backpointer *bps = alloc_v4_backpointers_c(a.v);
684                 unsigned i, nr = BCH_ALLOC_V4_NR_BACKPOINTERS(a.v);
685
686                 for (i = 0; i < nr; i++) {
687                         int cmp = backpointer_cmp(bps[i], bp) ?:
688                                 memcmp(&bps[i], &bp, sizeof(bp));
689                         if (!cmp)
690                                 goto out;
691                         if (cmp >= 0)
692                                 break;
693                 }
694         } else {
695                 goto missing;
696         }
697
698         bch2_trans_iter_init(trans, &bp_iter, BTREE_ID_backpointers,
699                              bucket_pos_to_bp(c, bucket_pos, bp.bucket_offset),
700                              0);
701         bp_k = bch2_btree_iter_peek_slot(&bp_iter);
702         ret = bkey_err(bp_k);
703         if (ret)
704                 goto err;
705
706         if (bp_k.k->type != KEY_TYPE_backpointer ||
707             memcmp(bkey_s_c_to_backpointer(bp_k).v, &bp, sizeof(bp)))
708                 goto missing;
709 out:
710 err:
711 fsck_err:
712         bch2_trans_iter_exit(trans, &bp_iter);
713         bch2_trans_iter_exit(trans, &alloc_iter);
714         printbuf_exit(&buf);
715         return ret;
716 missing:
717         prt_printf(&buf, "missing backpointer for btree=%s l=%u ",
718                bch2_btree_ids[bp.btree_id], bp.level);
719         bch2_bkey_val_to_text(&buf, c, orig_k);
720         prt_printf(&buf, "\nin alloc key ");
721         bch2_bkey_val_to_text(&buf, c, alloc_k);
722
723         if (c->sb.version < bcachefs_metadata_version_backpointers ||
724             c->opts.reconstruct_alloc ||
725             fsck_err(c, "%s", buf.buf)) {
726                 struct bkey_i_alloc_v4 *a = bch2_alloc_to_v4_mut(trans, alloc_k);
727
728                 ret   = PTR_ERR_OR_ZERO(a) ?:
729                         bch2_bucket_backpointer_add(trans, a, bp, orig_k) ?:
730                         bch2_trans_update(trans, &alloc_iter, &a->k_i, 0);
731         }
732
733         goto out;
734 }
735
736 static int check_extent_to_backpointers(struct btree_trans *trans,
737                                         struct btree_iter *iter,
738                                         struct bpos bucket_start,
739                                         struct bpos bucket_end)
740 {
741         struct bch_fs *c = trans->c;
742         struct bkey_ptrs_c ptrs;
743         const union bch_extent_entry *entry;
744         struct extent_ptr_decoded p;
745         struct bkey_s_c k;
746         int ret;
747
748         k = bch2_btree_iter_peek_all_levels(iter);
749         ret = bkey_err(k);
750         if (ret)
751                 return ret;
752         if (!k.k)
753                 return 0;
754
755         ptrs = bch2_bkey_ptrs_c(k);
756         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
757                 struct bpos bucket_pos;
758                 struct bch_backpointer bp;
759
760                 if (p.ptr.cached)
761                         continue;
762
763                 bch2_extent_ptr_to_bp(c, iter->btree_id, iter->path->level,
764                                       k, p, &bucket_pos, &bp);
765
766                 ret = check_bp_exists(trans, bucket_pos, bp, k, bucket_start, bucket_end);
767                 if (ret)
768                         return ret;
769         }
770
771         return 0;
772 }
773
774 static int check_btree_root_to_backpointers(struct btree_trans *trans,
775                                             enum btree_id btree_id,
776                                             struct bpos bucket_start,
777                                             struct bpos bucket_end)
778 {
779         struct bch_fs *c = trans->c;
780         struct btree_iter iter;
781         struct btree *b;
782         struct bkey_s_c k;
783         struct bkey_ptrs_c ptrs;
784         struct extent_ptr_decoded p;
785         const union bch_extent_entry *entry;
786         int ret;
787
788         bch2_trans_node_iter_init(trans, &iter, btree_id, POS_MIN, 0,
789                                   c->btree_roots[btree_id].level, 0);
790         b = bch2_btree_iter_peek_node(&iter);
791         ret = PTR_ERR_OR_ZERO(b);
792         if (ret)
793                 goto err;
794
795         BUG_ON(b != btree_node_root(c, b));
796
797         k = bkey_i_to_s_c(&b->key);
798         ptrs = bch2_bkey_ptrs_c(k);
799         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
800                 struct bpos bucket_pos;
801                 struct bch_backpointer bp;
802
803                 if (p.ptr.cached)
804                         continue;
805
806                 bch2_extent_ptr_to_bp(c, iter.btree_id, iter.path->level + 1,
807                                       k, p, &bucket_pos, &bp);
808
809                 ret = check_bp_exists(trans, bucket_pos, bp, k, bucket_start, bucket_end);
810                 if (ret)
811                         goto err;
812         }
813 err:
814         bch2_trans_iter_exit(trans, &iter);
815         return ret;
816 }
817
818 static inline struct bbpos bp_to_bbpos(struct bch_backpointer bp)
819 {
820         return (struct bbpos) {
821                 .btree  = bp.btree_id,
822                 .pos    = bp.pos,
823         };
824 }
825
826 static size_t btree_nodes_fit_in_ram(struct bch_fs *c)
827 {
828         struct sysinfo i;
829         u64 mem_bytes;
830
831         si_meminfo(&i);
832         mem_bytes = i.totalram * i.mem_unit;
833         return (mem_bytes >> 1) / btree_bytes(c);
834 }
835
836 int bch2_get_btree_in_memory_pos(struct btree_trans *trans,
837                                  unsigned btree_leaf_mask,
838                                  unsigned btree_interior_mask,
839                                  struct bbpos start, struct bbpos *end)
840 {
841         struct btree_iter iter;
842         struct bkey_s_c k;
843         size_t btree_nodes = btree_nodes_fit_in_ram(trans->c);
844         enum btree_id btree;
845         int ret = 0;
846
847         for (btree = start.btree; btree < BTREE_ID_NR && !ret; btree++) {
848                 unsigned depth = ((1U << btree) & btree_leaf_mask) ? 1 : 2;
849
850                 if (!((1U << btree) & btree_leaf_mask) &&
851                     !((1U << btree) & btree_interior_mask))
852                         continue;
853
854                 bch2_trans_node_iter_init(trans, &iter, btree,
855                                           btree == start.btree ? start.pos : POS_MIN,
856                                           0, depth, 0);
857                 /*
858                  * for_each_btree_key_contineu() doesn't check the return value
859                  * from bch2_btree_iter_advance(), which is needed when
860                  * iterating over interior nodes where we'll see keys at
861                  * SPOS_MAX:
862                  */
863                 do {
864                         k = __bch2_btree_iter_peek_and_restart(trans, &iter, 0);
865                         ret = bkey_err(k);
866                         if (!k.k || ret)
867                                 break;
868
869                         --btree_nodes;
870                         if (!btree_nodes) {
871                                 *end = BBPOS(btree, k.k->p);
872                                 bch2_trans_iter_exit(trans, &iter);
873                                 return 0;
874                         }
875                 } while (bch2_btree_iter_advance(&iter));
876                 bch2_trans_iter_exit(trans, &iter);
877         }
878
879         *end = BBPOS_MAX;
880         return ret;
881 }
882
883 static int bch2_check_extents_to_backpointers_pass(struct btree_trans *trans,
884                                                    struct bpos bucket_start,
885                                                    struct bpos bucket_end)
886 {
887         struct btree_iter iter;
888         enum btree_id btree_id;
889         int ret = 0;
890
891         for (btree_id = 0; btree_id < BTREE_ID_NR; btree_id++) {
892                 unsigned depth = btree_type_has_ptrs(btree_id) ? 0 : 1;
893
894                 bch2_trans_node_iter_init(trans, &iter, btree_id, POS_MIN, 0,
895                                           depth,
896                                           BTREE_ITER_ALL_LEVELS|
897                                           BTREE_ITER_PREFETCH);
898
899                 do {
900                         ret = commit_do(trans, NULL, NULL,
901                                         BTREE_INSERT_LAZY_RW|
902                                         BTREE_INSERT_NOFAIL,
903                                         check_extent_to_backpointers(trans, &iter,
904                                                                 bucket_start, bucket_end));
905                         if (ret)
906                                 break;
907                 } while (!bch2_btree_iter_advance(&iter));
908
909                 bch2_trans_iter_exit(trans, &iter);
910
911                 if (ret)
912                         break;
913
914                 ret = commit_do(trans, NULL, NULL,
915                                 BTREE_INSERT_LAZY_RW|
916                                 BTREE_INSERT_NOFAIL,
917                                 check_btree_root_to_backpointers(trans, btree_id,
918                                                         bucket_start, bucket_end));
919                 if (ret)
920                         break;
921         }
922         return ret;
923 }
924
925 int bch2_get_alloc_in_memory_pos(struct btree_trans *trans,
926                                  struct bpos start, struct bpos *end)
927 {
928         struct btree_iter alloc_iter;
929         struct btree_iter bp_iter;
930         struct bkey_s_c alloc_k, bp_k;
931         size_t btree_nodes = btree_nodes_fit_in_ram(trans->c);
932         bool alloc_end = false, bp_end = false;
933         int ret = 0;
934
935         bch2_trans_node_iter_init(trans, &alloc_iter, BTREE_ID_alloc,
936                                   start, 0, 1, 0);
937         bch2_trans_node_iter_init(trans, &bp_iter, BTREE_ID_backpointers,
938                                   bucket_pos_to_bp(trans->c, start, 0), 0, 1, 0);
939         while (1) {
940                 alloc_k = !alloc_end
941                         ? __bch2_btree_iter_peek_and_restart(trans, &alloc_iter, 0)
942                         : bkey_s_c_null;
943                 bp_k = !bp_end
944                         ? __bch2_btree_iter_peek_and_restart(trans, &bp_iter, 0)
945                         : bkey_s_c_null;
946
947                 ret = bkey_err(alloc_k) ?: bkey_err(bp_k);
948                 if ((!alloc_k.k && !bp_k.k) || ret) {
949                         *end = SPOS_MAX;
950                         break;
951                 }
952
953                 --btree_nodes;
954                 if (!btree_nodes) {
955                         *end = alloc_k.k->p;
956                         break;
957                 }
958
959                 if (bpos_cmp(alloc_iter.pos, SPOS_MAX) &&
960                     bpos_cmp(bucket_pos_to_bp(trans->c, alloc_iter.pos, 0), bp_iter.pos) < 0) {
961                         if (!bch2_btree_iter_advance(&alloc_iter))
962                                 alloc_end = true;
963                 } else {
964                         if (!bch2_btree_iter_advance(&bp_iter))
965                                 bp_end = true;
966                 }
967         }
968         bch2_trans_iter_exit(trans, &bp_iter);
969         bch2_trans_iter_exit(trans, &alloc_iter);
970         return ret;
971 }
972
973 int bch2_check_extents_to_backpointers(struct bch_fs *c)
974 {
975         struct btree_trans trans;
976         struct bpos start = POS_MIN, end;
977         int ret;
978
979         bch2_trans_init(&trans, c, 0, 0);
980         while (1) {
981                 ret = bch2_get_alloc_in_memory_pos(&trans, start, &end);
982                 if (ret)
983                         break;
984
985                 if (!bpos_cmp(start, POS_MIN) && bpos_cmp(end, SPOS_MAX))
986                         bch_verbose(c, "check_extents_to_backpointers(): alloc info does not fit in ram,"
987                                     "running in multiple passes with %zu nodes per pass",
988                                     btree_nodes_fit_in_ram(c));
989
990                 if (bpos_cmp(start, POS_MIN) || bpos_cmp(end, SPOS_MAX)) {
991                         struct printbuf buf = PRINTBUF;
992
993                         prt_str(&buf, "check_extents_to_backpointers(): ");
994                         bch2_bpos_to_text(&buf, start);
995                         prt_str(&buf, "-");
996                         bch2_bpos_to_text(&buf, end);
997
998                         bch_verbose(c, "%s", buf.buf);
999                         printbuf_exit(&buf);
1000                 }
1001
1002                 ret = bch2_check_extents_to_backpointers_pass(&trans, start, end);
1003                 if (ret || !bpos_cmp(end, SPOS_MAX))
1004                         break;
1005
1006                 start = bpos_successor(end);
1007         }
1008         bch2_trans_exit(&trans);
1009
1010         return ret;
1011 }
1012
1013 static int check_one_backpointer(struct btree_trans *trans,
1014                                  struct bpos bucket,
1015                                  u64 *bp_offset,
1016                                  struct bbpos start,
1017                                  struct bbpos end)
1018 {
1019         struct btree_iter iter;
1020         struct bch_backpointer bp;
1021         struct bbpos pos;
1022         struct bkey_s_c k;
1023         struct printbuf buf = PRINTBUF;
1024         int ret;
1025
1026         ret = bch2_get_next_backpointer(trans, bucket, -1, bp_offset, &bp);
1027         if (ret || *bp_offset == U64_MAX)
1028                 return ret;
1029
1030         pos = bp_to_bbpos(bp);
1031         if (bbpos_cmp(pos, start) < 0 ||
1032             bbpos_cmp(pos, end) > 0)
1033                 return 0;
1034
1035         k = bch2_backpointer_get_key(trans, &iter, bucket, *bp_offset, bp);
1036         ret = bkey_err(k);
1037         if (ret == -BCH_ERR_backpointer_to_overwritten_btree_node)
1038                 return 0;
1039         if (ret)
1040                 return ret;
1041
1042         if (fsck_err_on(!k.k, trans->c,
1043                         "%s backpointer points to missing extent\n%s",
1044                         *bp_offset < BACKPOINTER_OFFSET_MAX ? "alloc" : "btree",
1045                         (bch2_backpointer_to_text(&buf, &bp), buf.buf))) {
1046                 ret = bch2_backpointer_del_by_offset(trans, bucket, *bp_offset, bp);
1047                 if (ret == -ENOENT)
1048                         bch_err(trans->c, "backpointer at %llu not found", *bp_offset);
1049         }
1050
1051         bch2_trans_iter_exit(trans, &iter);
1052 fsck_err:
1053         printbuf_exit(&buf);
1054         return ret;
1055 }
1056
1057 static int bch2_check_backpointers_to_extents_pass(struct btree_trans *trans,
1058                                                    struct bbpos start,
1059                                                    struct bbpos end)
1060 {
1061         struct btree_iter iter;
1062         struct bkey_s_c k;
1063         int ret = 0;
1064
1065         for_each_btree_key(trans, iter, BTREE_ID_alloc, POS_MIN,
1066                            BTREE_ITER_PREFETCH, k, ret) {
1067                 u64 bp_offset = 0;
1068
1069                 while (!(ret = commit_do(trans, NULL, NULL,
1070                                          BTREE_INSERT_LAZY_RW|
1071                                          BTREE_INSERT_NOFAIL,
1072                                 check_one_backpointer(trans, iter.pos, &bp_offset, start, end))) &&
1073                        bp_offset < U64_MAX)
1074                         bp_offset++;
1075
1076                 if (ret)
1077                         break;
1078         }
1079         bch2_trans_iter_exit(trans, &iter);
1080         return ret < 0 ? ret : 0;
1081 }
1082
1083 int bch2_check_backpointers_to_extents(struct bch_fs *c)
1084 {
1085         struct btree_trans trans;
1086         struct bbpos start = (struct bbpos) { .btree = 0, .pos = POS_MIN, }, end;
1087         int ret;
1088
1089         bch2_trans_init(&trans, c, 0, 0);
1090         while (1) {
1091                 ret = bch2_get_btree_in_memory_pos(&trans,
1092                                                    (1U << BTREE_ID_extents)|
1093                                                    (1U << BTREE_ID_reflink),
1094                                                    ~0,
1095                                                    start, &end);
1096                 if (ret)
1097                         break;
1098
1099                 if (!bbpos_cmp(start, BBPOS_MIN) &&
1100                     bbpos_cmp(end, BBPOS_MAX))
1101                         bch_verbose(c, "check_backpointers_to_extents(): extents do not fit in ram,"
1102                                     "running in multiple passes with %zu nodes per pass",
1103                                     btree_nodes_fit_in_ram(c));
1104
1105                 if (bbpos_cmp(start, BBPOS_MIN) ||
1106                     bbpos_cmp(end, BBPOS_MAX)) {
1107                         struct printbuf buf = PRINTBUF;
1108
1109                         prt_str(&buf, "check_backpointers_to_extents(): ");
1110                         bch2_bbpos_to_text(&buf, start);
1111                         prt_str(&buf, "-");
1112                         bch2_bbpos_to_text(&buf, end);
1113
1114                         bch_verbose(c, "%s", buf.buf);
1115                         printbuf_exit(&buf);
1116                 }
1117
1118                 ret = bch2_check_backpointers_to_extents_pass(&trans, start, end);
1119                 if (ret || !bbpos_cmp(end, BBPOS_MAX))
1120                         break;
1121
1122                 start = bbpos_successor(end);
1123         }
1124         bch2_trans_exit(&trans);
1125
1126         return ret;
1127 }