]> git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/btree_iter.c
Update bcachefs sources to e67d924a90 bcachefs: Introduce bch2_journal_keys_peek_...
[bcachefs-tools-debian] / libbcachefs / btree_iter.c
1 // SPDX-License-Identifier: GPL-2.0
2
3 #include "bcachefs.h"
4 #include "bkey_methods.h"
5 #include "bkey_buf.h"
6 #include "btree_cache.h"
7 #include "btree_iter.h"
8 #include "btree_key_cache.h"
9 #include "btree_locking.h"
10 #include "btree_update.h"
11 #include "debug.h"
12 #include "error.h"
13 #include "extents.h"
14 #include "journal.h"
15 #include "recovery.h"
16 #include "replicas.h"
17 #include "subvolume.h"
18
19 #include <linux/prefetch.h>
20 #include <trace/events/bcachefs.h>
21
22 static void btree_trans_verify_sorted(struct btree_trans *);
23 inline void bch2_btree_path_check_sort(struct btree_trans *, struct btree_path *, int);
24
25 static inline void btree_path_list_remove(struct btree_trans *, struct btree_path *);
26 static inline void btree_path_list_add(struct btree_trans *, struct btree_path *,
27                                        struct btree_path *);
28
29 static inline unsigned long btree_iter_ip_allocated(struct btree_iter *iter)
30 {
31 #ifdef CONFIG_BCACHEFS_DEBUG
32         return iter->ip_allocated;
33 #else
34         return 0;
35 #endif
36 }
37
38 static struct btree_path *btree_path_alloc(struct btree_trans *, struct btree_path *);
39
40 /*
41  * Unlocks before scheduling
42  * Note: does not revalidate iterator
43  */
44 static inline int bch2_trans_cond_resched(struct btree_trans *trans)
45 {
46         if (need_resched() || race_fault()) {
47                 bch2_trans_unlock(trans);
48                 schedule();
49                 return bch2_trans_relock(trans) ? 0 : -EINTR;
50         } else {
51                 return 0;
52         }
53 }
54
55 static inline int __btree_path_cmp(const struct btree_path *l,
56                                    enum btree_id        r_btree_id,
57                                    bool                 r_cached,
58                                    struct bpos          r_pos,
59                                    unsigned             r_level)
60 {
61         /*
62          * Must match lock ordering as defined by __bch2_btree_node_lock:
63          */
64         return   cmp_int(l->btree_id,   r_btree_id) ?:
65                  cmp_int((int) l->cached,       (int) r_cached) ?:
66                  bpos_cmp(l->pos,       r_pos) ?:
67                 -cmp_int(l->level,      r_level);
68 }
69
70 static inline int btree_path_cmp(const struct btree_path *l,
71                                  const struct btree_path *r)
72 {
73         return __btree_path_cmp(l, r->btree_id, r->cached, r->pos, r->level);
74 }
75
76 static inline struct bpos bkey_successor(struct btree_iter *iter, struct bpos p)
77 {
78         /* Are we iterating over keys in all snapshots? */
79         if (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) {
80                 p = bpos_successor(p);
81         } else {
82                 p = bpos_nosnap_successor(p);
83                 p.snapshot = iter->snapshot;
84         }
85
86         return p;
87 }
88
89 static inline struct bpos bkey_predecessor(struct btree_iter *iter, struct bpos p)
90 {
91         /* Are we iterating over keys in all snapshots? */
92         if (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) {
93                 p = bpos_predecessor(p);
94         } else {
95                 p = bpos_nosnap_predecessor(p);
96                 p.snapshot = iter->snapshot;
97         }
98
99         return p;
100 }
101
102 static inline bool is_btree_node(struct btree_path *path, unsigned l)
103 {
104         return l < BTREE_MAX_DEPTH &&
105                 (unsigned long) path->l[l].b >= 128;
106 }
107
108 static inline struct bpos btree_iter_search_key(struct btree_iter *iter)
109 {
110         struct bpos pos = iter->pos;
111
112         if ((iter->flags & BTREE_ITER_IS_EXTENTS) &&
113             bkey_cmp(pos, POS_MAX))
114                 pos = bkey_successor(iter, pos);
115         return pos;
116 }
117
118 static inline bool btree_path_pos_before_node(struct btree_path *path,
119                                               struct btree *b)
120 {
121         return bpos_cmp(path->pos, b->data->min_key) < 0;
122 }
123
124 static inline bool btree_path_pos_after_node(struct btree_path *path,
125                                              struct btree *b)
126 {
127         return bpos_cmp(b->key.k.p, path->pos) < 0;
128 }
129
130 static inline bool btree_path_pos_in_node(struct btree_path *path,
131                                           struct btree *b)
132 {
133         return path->btree_id == b->c.btree_id &&
134                 !btree_path_pos_before_node(path, b) &&
135                 !btree_path_pos_after_node(path, b);
136 }
137
138 /* Btree node locking: */
139
140 void bch2_btree_node_unlock_write(struct btree_trans *trans,
141                         struct btree_path *path, struct btree *b)
142 {
143         bch2_btree_node_unlock_write_inlined(trans, path, b);
144 }
145
146 void __bch2_btree_node_lock_write(struct btree_trans *trans, struct btree *b)
147 {
148         struct btree_path *linked;
149         unsigned readers = 0;
150
151         trans_for_each_path(trans, linked)
152                 if (linked->l[b->c.level].b == b &&
153                     btree_node_read_locked(linked, b->c.level))
154                         readers++;
155
156         /*
157          * Must drop our read locks before calling six_lock_write() -
158          * six_unlock() won't do wakeups until the reader count
159          * goes to 0, and it's safe because we have the node intent
160          * locked:
161          */
162         if (!b->c.lock.readers)
163                 atomic64_sub(__SIX_VAL(read_lock, readers),
164                              &b->c.lock.state.counter);
165         else
166                 this_cpu_sub(*b->c.lock.readers, readers);
167
168         six_lock_write(&b->c.lock, NULL, NULL);
169
170         if (!b->c.lock.readers)
171                 atomic64_add(__SIX_VAL(read_lock, readers),
172                              &b->c.lock.state.counter);
173         else
174                 this_cpu_add(*b->c.lock.readers, readers);
175 }
176
177 bool __bch2_btree_node_relock(struct btree_trans *trans,
178                               struct btree_path *path, unsigned level)
179 {
180         struct btree *b = btree_path_node(path, level);
181         int want = __btree_lock_want(path, level);
182
183         if (!is_btree_node(path, level))
184                 goto fail;
185
186         if (race_fault())
187                 goto fail;
188
189         if (six_relock_type(&b->c.lock, want, path->l[level].lock_seq) ||
190             (btree_node_lock_seq_matches(path, b, level) &&
191              btree_node_lock_increment(trans, b, level, want))) {
192                 mark_btree_node_locked(trans, path, level, want);
193                 return true;
194         }
195 fail:
196         trace_btree_node_relock_fail(trans->fn, _RET_IP_,
197                                      path->btree_id,
198                                      &path->pos,
199                                      (unsigned long) b,
200                                      path->l[level].lock_seq,
201                                      is_btree_node(path, level) ? b->c.lock.state.seq : 0);
202         return false;
203 }
204
205 bool bch2_btree_node_upgrade(struct btree_trans *trans,
206                              struct btree_path *path, unsigned level)
207 {
208         struct btree *b = path->l[level].b;
209
210         if (!is_btree_node(path, level))
211                 return false;
212
213         switch (btree_lock_want(path, level)) {
214         case BTREE_NODE_UNLOCKED:
215                 BUG_ON(btree_node_locked(path, level));
216                 return true;
217         case BTREE_NODE_READ_LOCKED:
218                 BUG_ON(btree_node_intent_locked(path, level));
219                 return bch2_btree_node_relock(trans, path, level);
220         case BTREE_NODE_INTENT_LOCKED:
221                 break;
222         }
223
224         if (btree_node_intent_locked(path, level))
225                 return true;
226
227         if (race_fault())
228                 return false;
229
230         if (btree_node_locked(path, level)
231             ? six_lock_tryupgrade(&b->c.lock)
232             : six_relock_type(&b->c.lock, SIX_LOCK_intent, path->l[level].lock_seq))
233                 goto success;
234
235         if (btree_node_lock_seq_matches(path, b, level) &&
236             btree_node_lock_increment(trans, b, level, BTREE_NODE_INTENT_LOCKED)) {
237                 btree_node_unlock(path, level);
238                 goto success;
239         }
240
241         return false;
242 success:
243         mark_btree_node_intent_locked(trans, path, level);
244         return true;
245 }
246
247 static inline bool btree_path_get_locks(struct btree_trans *trans,
248                                         struct btree_path *path,
249                                         bool upgrade)
250 {
251         unsigned l = path->level;
252         int fail_idx = -1;
253
254         do {
255                 if (!btree_path_node(path, l))
256                         break;
257
258                 if (!(upgrade
259                       ? bch2_btree_node_upgrade(trans, path, l)
260                       : bch2_btree_node_relock(trans, path, l)))
261                         fail_idx = l;
262
263                 l++;
264         } while (l < path->locks_want);
265
266         /*
267          * When we fail to get a lock, we have to ensure that any child nodes
268          * can't be relocked so bch2_btree_path_traverse has to walk back up to
269          * the node that we failed to relock:
270          */
271         if (fail_idx >= 0) {
272                 __bch2_btree_path_unlock(path);
273                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
274
275                 do {
276                         path->l[fail_idx].b = BTREE_ITER_NO_NODE_GET_LOCKS;
277                         --fail_idx;
278                 } while (fail_idx >= 0);
279         }
280
281         if (path->uptodate == BTREE_ITER_NEED_RELOCK)
282                 path->uptodate = BTREE_ITER_UPTODATE;
283
284         bch2_trans_verify_locks(trans);
285
286         return path->uptodate < BTREE_ITER_NEED_RELOCK;
287 }
288
289 static struct bpos btree_node_pos(struct btree_bkey_cached_common *_b,
290                                   bool cached)
291 {
292         return !cached
293                 ? container_of(_b, struct btree, c)->key.k.p
294                 : container_of(_b, struct bkey_cached, c)->key.pos;
295 }
296
297 /* Slowpath: */
298 bool __bch2_btree_node_lock(struct btree_trans *trans,
299                             struct btree_path *path,
300                             struct btree *b,
301                             struct bpos pos, unsigned level,
302                             enum six_lock_type type,
303                             six_lock_should_sleep_fn should_sleep_fn, void *p,
304                             unsigned long ip)
305 {
306         struct btree_path *linked;
307         unsigned reason;
308
309         /* Check if it's safe to block: */
310         trans_for_each_path(trans, linked) {
311                 if (!linked->nodes_locked)
312                         continue;
313
314                 /*
315                  * Can't block taking an intent lock if we have _any_ nodes read
316                  * locked:
317                  *
318                  * - Our read lock blocks another thread with an intent lock on
319                  *   the same node from getting a write lock, and thus from
320                  *   dropping its intent lock
321                  *
322                  * - And the other thread may have multiple nodes intent locked:
323                  *   both the node we want to intent lock, and the node we
324                  *   already have read locked - deadlock:
325                  */
326                 if (type == SIX_LOCK_intent &&
327                     linked->nodes_locked != linked->nodes_intent_locked) {
328                         reason = 1;
329                         goto deadlock;
330                 }
331
332                 if (linked->btree_id != path->btree_id) {
333                         if (linked->btree_id < path->btree_id)
334                                 continue;
335
336                         reason = 3;
337                         goto deadlock;
338                 }
339
340                 /*
341                  * Within the same btree, non-cached paths come before cached
342                  * paths:
343                  */
344                 if (linked->cached != path->cached) {
345                         if (!linked->cached)
346                                 continue;
347
348                         reason = 4;
349                         goto deadlock;
350                 }
351
352                 /*
353                  * Interior nodes must be locked before their descendants: if
354                  * another path has possible descendants locked of the node
355                  * we're about to lock, it must have the ancestors locked too:
356                  */
357                 if (level > __fls(linked->nodes_locked)) {
358                         reason = 5;
359                         goto deadlock;
360                 }
361
362                 /* Must lock btree nodes in key order: */
363                 if (btree_node_locked(linked, level) &&
364                     bpos_cmp(pos, btree_node_pos((void *) linked->l[level].b,
365                                                  linked->cached)) <= 0) {
366                         BUG_ON(trans->in_traverse_all);
367                         reason = 7;
368                         goto deadlock;
369                 }
370         }
371
372         return btree_node_lock_type(trans, path, b, pos, level,
373                                     type, should_sleep_fn, p);
374 deadlock:
375         trace_trans_restart_would_deadlock(trans->fn, ip,
376                         trans->in_traverse_all, reason,
377                         linked->btree_id,
378                         linked->cached,
379                         &linked->pos,
380                         path->btree_id,
381                         path->cached,
382                         &pos);
383         btree_trans_restart(trans);
384         return false;
385 }
386
387 /* Btree iterator locking: */
388
389 #ifdef CONFIG_BCACHEFS_DEBUG
390
391 static void bch2_btree_path_verify_locks(struct btree_path *path)
392 {
393         unsigned l;
394
395         if (!path->nodes_locked) {
396                 BUG_ON(path->uptodate == BTREE_ITER_UPTODATE &&
397                        btree_path_node(path, path->level));
398                 return;
399         }
400
401         for (l = 0; btree_path_node(path, l); l++)
402                 BUG_ON(btree_lock_want(path, l) !=
403                        btree_node_locked_type(path, l));
404 }
405
406 void bch2_trans_verify_locks(struct btree_trans *trans)
407 {
408         struct btree_path *path;
409
410         trans_for_each_path(trans, path)
411                 bch2_btree_path_verify_locks(path);
412 }
413 #else
414 static inline void bch2_btree_path_verify_locks(struct btree_path *path) {}
415 #endif
416
417 /* Btree path locking: */
418
419 /*
420  * Only for btree_cache.c - only relocks intent locks
421  */
422 bool bch2_btree_path_relock_intent(struct btree_trans *trans,
423                                    struct btree_path *path)
424 {
425         unsigned l;
426
427         for (l = path->level;
428              l < path->locks_want && btree_path_node(path, l);
429              l++) {
430                 if (!bch2_btree_node_relock(trans, path, l)) {
431                         __bch2_btree_path_unlock(path);
432                         btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
433                         trace_trans_restart_relock_path_intent(trans->fn, _RET_IP_,
434                                                    path->btree_id, &path->pos);
435                         btree_trans_restart(trans);
436                         return false;
437                 }
438         }
439
440         return true;
441 }
442
443 __flatten
444 static bool bch2_btree_path_relock(struct btree_trans *trans,
445                         struct btree_path *path, unsigned long trace_ip)
446 {
447         bool ret = btree_path_get_locks(trans, path, false);
448
449         if (!ret) {
450                 trace_trans_restart_relock_path(trans->fn, trace_ip,
451                                                 path->btree_id, &path->pos);
452                 btree_trans_restart(trans);
453         }
454         return ret;
455 }
456
457 bool __bch2_btree_path_upgrade(struct btree_trans *trans,
458                                struct btree_path *path,
459                                unsigned new_locks_want)
460 {
461         struct btree_path *linked;
462
463         EBUG_ON(path->locks_want >= new_locks_want);
464
465         path->locks_want = new_locks_want;
466
467         if (btree_path_get_locks(trans, path, true))
468                 return true;
469
470         /*
471          * XXX: this is ugly - we'd prefer to not be mucking with other
472          * iterators in the btree_trans here.
473          *
474          * On failure to upgrade the iterator, setting iter->locks_want and
475          * calling get_locks() is sufficient to make bch2_btree_path_traverse()
476          * get the locks we want on transaction restart.
477          *
478          * But if this iterator was a clone, on transaction restart what we did
479          * to this iterator isn't going to be preserved.
480          *
481          * Possibly we could add an iterator field for the parent iterator when
482          * an iterator is a copy - for now, we'll just upgrade any other
483          * iterators with the same btree id.
484          *
485          * The code below used to be needed to ensure ancestor nodes get locked
486          * before interior nodes - now that's handled by
487          * bch2_btree_path_traverse_all().
488          */
489         if (!path->cached && !trans->in_traverse_all)
490                 trans_for_each_path(trans, linked)
491                         if (linked != path &&
492                             linked->cached == path->cached &&
493                             linked->btree_id == path->btree_id &&
494                             linked->locks_want < new_locks_want) {
495                                 linked->locks_want = new_locks_want;
496                                 btree_path_get_locks(trans, linked, true);
497                         }
498
499         return false;
500 }
501
502 void __bch2_btree_path_downgrade(struct btree_path *path,
503                                  unsigned new_locks_want)
504 {
505         unsigned l;
506
507         EBUG_ON(path->locks_want < new_locks_want);
508
509         path->locks_want = new_locks_want;
510
511         while (path->nodes_locked &&
512                (l = __fls(path->nodes_locked)) >= path->locks_want) {
513                 if (l > path->level) {
514                         btree_node_unlock(path, l);
515                 } else {
516                         if (btree_node_intent_locked(path, l)) {
517                                 six_lock_downgrade(&path->l[l].b->c.lock);
518                                 path->nodes_intent_locked ^= 1 << l;
519                         }
520                         break;
521                 }
522         }
523
524         bch2_btree_path_verify_locks(path);
525 }
526
527 void bch2_trans_downgrade(struct btree_trans *trans)
528 {
529         struct btree_path *path;
530
531         trans_for_each_path(trans, path)
532                 bch2_btree_path_downgrade(path);
533 }
534
535 /* Btree transaction locking: */
536
537 bool bch2_trans_relock(struct btree_trans *trans)
538 {
539         struct btree_path *path;
540
541         if (unlikely(trans->restarted))
542                 return false;
543
544         trans_for_each_path(trans, path)
545                 if (path->should_be_locked &&
546                     !bch2_btree_path_relock(trans, path, _RET_IP_)) {
547                         trace_trans_restart_relock(trans->fn, _RET_IP_,
548                                         path->btree_id, &path->pos);
549                         BUG_ON(!trans->restarted);
550                         return false;
551                 }
552         return true;
553 }
554
555 void bch2_trans_unlock(struct btree_trans *trans)
556 {
557         struct btree_path *path;
558
559         trans_for_each_path(trans, path)
560                 __bch2_btree_path_unlock(path);
561
562         /*
563          * bch2_gc_btree_init_recurse() doesn't use btree iterators for walking
564          * btree nodes, it implements its own walking:
565          */
566         BUG_ON(!trans->is_initial_gc &&
567                lock_class_is_held(&bch2_btree_node_lock_key));
568 }
569
570 /* Btree iterator: */
571
572 #ifdef CONFIG_BCACHEFS_DEBUG
573
574 static void bch2_btree_path_verify_cached(struct btree_trans *trans,
575                                           struct btree_path *path)
576 {
577         struct bkey_cached *ck;
578         bool locked = btree_node_locked(path, 0);
579
580         if (!bch2_btree_node_relock(trans, path, 0))
581                 return;
582
583         ck = (void *) path->l[0].b;
584         BUG_ON(ck->key.btree_id != path->btree_id ||
585                bkey_cmp(ck->key.pos, path->pos));
586
587         if (!locked)
588                 btree_node_unlock(path, 0);
589 }
590
591 static void bch2_btree_path_verify_level(struct btree_trans *trans,
592                                 struct btree_path *path, unsigned level)
593 {
594         struct btree_path_level *l;
595         struct btree_node_iter tmp;
596         bool locked;
597         struct bkey_packed *p, *k;
598         struct printbuf buf1 = PRINTBUF;
599         struct printbuf buf2 = PRINTBUF;
600         struct printbuf buf3 = PRINTBUF;
601         const char *msg;
602
603         if (!bch2_debug_check_iterators)
604                 return;
605
606         l       = &path->l[level];
607         tmp     = l->iter;
608         locked  = btree_node_locked(path, level);
609
610         if (path->cached) {
611                 if (!level)
612                         bch2_btree_path_verify_cached(trans, path);
613                 return;
614         }
615
616         if (!btree_path_node(path, level))
617                 return;
618
619         if (!bch2_btree_node_relock(trans, path, level))
620                 return;
621
622         BUG_ON(!btree_path_pos_in_node(path, l->b));
623
624         bch2_btree_node_iter_verify(&l->iter, l->b);
625
626         /*
627          * For interior nodes, the iterator will have skipped past deleted keys:
628          */
629         p = level
630                 ? bch2_btree_node_iter_prev(&tmp, l->b)
631                 : bch2_btree_node_iter_prev_all(&tmp, l->b);
632         k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
633
634         if (p && bkey_iter_pos_cmp(l->b, p, &path->pos) >= 0) {
635                 msg = "before";
636                 goto err;
637         }
638
639         if (k && bkey_iter_pos_cmp(l->b, k, &path->pos) < 0) {
640                 msg = "after";
641                 goto err;
642         }
643
644         if (!locked)
645                 btree_node_unlock(path, level);
646         return;
647 err:
648         bch2_bpos_to_text(&buf1, path->pos);
649
650         if (p) {
651                 struct bkey uk = bkey_unpack_key(l->b, p);
652                 bch2_bkey_to_text(&buf2, &uk);
653         } else {
654                 pr_buf(&buf2, "(none)");
655         }
656
657         if (k) {
658                 struct bkey uk = bkey_unpack_key(l->b, k);
659                 bch2_bkey_to_text(&buf3, &uk);
660         } else {
661                 pr_buf(&buf3, "(none)");
662         }
663
664         panic("path should be %s key at level %u:\n"
665               "path pos %s\n"
666               "prev key %s\n"
667               "cur  key %s\n",
668               msg, level, buf1.buf, buf2.buf, buf3.buf);
669 }
670
671 static void bch2_btree_path_verify(struct btree_trans *trans,
672                                    struct btree_path *path)
673 {
674         struct bch_fs *c = trans->c;
675         unsigned i;
676
677         EBUG_ON(path->btree_id >= BTREE_ID_NR);
678
679         for (i = 0; i < (!path->cached ? BTREE_MAX_DEPTH : 1); i++) {
680                 if (!path->l[i].b) {
681                         BUG_ON(!path->cached &&
682                                c->btree_roots[path->btree_id].b->c.level > i);
683                         break;
684                 }
685
686                 bch2_btree_path_verify_level(trans, path, i);
687         }
688
689         bch2_btree_path_verify_locks(path);
690 }
691
692 void bch2_trans_verify_paths(struct btree_trans *trans)
693 {
694         struct btree_path *path;
695
696         trans_for_each_path(trans, path)
697                 bch2_btree_path_verify(trans, path);
698 }
699
700 static void bch2_btree_iter_verify(struct btree_iter *iter)
701 {
702         struct btree_trans *trans = iter->trans;
703
704         BUG_ON(iter->btree_id >= BTREE_ID_NR);
705
706         BUG_ON(!!(iter->flags & BTREE_ITER_CACHED) != iter->path->cached);
707
708         BUG_ON((iter->flags & BTREE_ITER_IS_EXTENTS) &&
709                (iter->flags & BTREE_ITER_ALL_SNAPSHOTS));
710
711         BUG_ON(!(iter->flags & __BTREE_ITER_ALL_SNAPSHOTS) &&
712                (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) &&
713                !btree_type_has_snapshots(iter->btree_id));
714
715         if (iter->update_path)
716                 bch2_btree_path_verify(trans, iter->update_path);
717         bch2_btree_path_verify(trans, iter->path);
718 }
719
720 static void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter)
721 {
722         BUG_ON((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
723                !iter->pos.snapshot);
724
725         BUG_ON(!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS) &&
726                iter->pos.snapshot != iter->snapshot);
727
728         BUG_ON(bkey_cmp(iter->pos, bkey_start_pos(&iter->k)) < 0 ||
729                bkey_cmp(iter->pos, iter->k.p) > 0);
730 }
731
732 static int bch2_btree_iter_verify_ret(struct btree_iter *iter, struct bkey_s_c k)
733 {
734         struct btree_trans *trans = iter->trans;
735         struct btree_iter copy;
736         struct bkey_s_c prev;
737         int ret = 0;
738
739         if (!bch2_debug_check_iterators)
740                 return 0;
741
742         if (!(iter->flags & BTREE_ITER_FILTER_SNAPSHOTS))
743                 return 0;
744
745         if (bkey_err(k) || !k.k)
746                 return 0;
747
748         BUG_ON(!bch2_snapshot_is_ancestor(trans->c,
749                                           iter->snapshot,
750                                           k.k->p.snapshot));
751
752         bch2_trans_iter_init(trans, &copy, iter->btree_id, iter->pos,
753                              BTREE_ITER_NOPRESERVE|
754                              BTREE_ITER_ALL_SNAPSHOTS);
755         prev = bch2_btree_iter_prev(&copy);
756         if (!prev.k)
757                 goto out;
758
759         ret = bkey_err(prev);
760         if (ret)
761                 goto out;
762
763         if (!bkey_cmp(prev.k->p, k.k->p) &&
764             bch2_snapshot_is_ancestor(trans->c, iter->snapshot,
765                                       prev.k->p.snapshot) > 0) {
766                 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
767
768                 bch2_bkey_to_text(&buf1, k.k);
769                 bch2_bkey_to_text(&buf2, prev.k);
770
771                 panic("iter snap %u\n"
772                       "k    %s\n"
773                       "prev %s\n",
774                       iter->snapshot,
775                       buf1.buf, buf2.buf);
776         }
777 out:
778         bch2_trans_iter_exit(trans, &copy);
779         return ret;
780 }
781
782 void bch2_assert_pos_locked(struct btree_trans *trans, enum btree_id id,
783                             struct bpos pos, bool key_cache)
784 {
785         struct btree_path *path;
786         unsigned idx;
787         struct printbuf buf = PRINTBUF;
788
789         trans_for_each_path_inorder(trans, path, idx) {
790                 int cmp = cmp_int(path->btree_id, id) ?:
791                         cmp_int(path->cached, key_cache);
792
793                 if (cmp > 0)
794                         break;
795                 if (cmp < 0)
796                         continue;
797
798                 if (!(path->nodes_locked & 1) ||
799                     !path->should_be_locked)
800                         continue;
801
802                 if (!key_cache) {
803                         if (bkey_cmp(pos, path->l[0].b->data->min_key) >= 0 &&
804                             bkey_cmp(pos, path->l[0].b->key.k.p) <= 0)
805                                 return;
806                 } else {
807                         if (!bkey_cmp(pos, path->pos))
808                                 return;
809                 }
810         }
811
812         bch2_dump_trans_paths_updates(trans);
813         bch2_bpos_to_text(&buf, pos);
814
815         panic("not locked: %s %s%s\n",
816               bch2_btree_ids[id], buf.buf,
817               key_cache ? " cached" : "");
818 }
819
820 #else
821
822 static inline void bch2_btree_path_verify_level(struct btree_trans *trans,
823                                                 struct btree_path *path, unsigned l) {}
824 static inline void bch2_btree_path_verify(struct btree_trans *trans,
825                                           struct btree_path *path) {}
826 static inline void bch2_btree_iter_verify(struct btree_iter *iter) {}
827 static inline void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter) {}
828 static inline int bch2_btree_iter_verify_ret(struct btree_iter *iter, struct bkey_s_c k) { return 0; }
829
830 #endif
831
832 /* Btree path: fixups after btree updates */
833
834 static void btree_node_iter_set_set_pos(struct btree_node_iter *iter,
835                                         struct btree *b,
836                                         struct bset_tree *t,
837                                         struct bkey_packed *k)
838 {
839         struct btree_node_iter_set *set;
840
841         btree_node_iter_for_each(iter, set)
842                 if (set->end == t->end_offset) {
843                         set->k = __btree_node_key_to_offset(b, k);
844                         bch2_btree_node_iter_sort(iter, b);
845                         return;
846                 }
847
848         bch2_btree_node_iter_push(iter, b, k, btree_bkey_last(b, t));
849 }
850
851 static void __bch2_btree_path_fix_key_modified(struct btree_path *path,
852                                                struct btree *b,
853                                                struct bkey_packed *where)
854 {
855         struct btree_path_level *l = &path->l[b->c.level];
856
857         if (where != bch2_btree_node_iter_peek_all(&l->iter, l->b))
858                 return;
859
860         if (bkey_iter_pos_cmp(l->b, where, &path->pos) < 0)
861                 bch2_btree_node_iter_advance(&l->iter, l->b);
862 }
863
864 void bch2_btree_path_fix_key_modified(struct btree_trans *trans,
865                                       struct btree *b,
866                                       struct bkey_packed *where)
867 {
868         struct btree_path *path;
869
870         trans_for_each_path_with_node(trans, b, path) {
871                 __bch2_btree_path_fix_key_modified(path, b, where);
872                 bch2_btree_path_verify_level(trans, path, b->c.level);
873         }
874 }
875
876 static void __bch2_btree_node_iter_fix(struct btree_path *path,
877                                        struct btree *b,
878                                        struct btree_node_iter *node_iter,
879                                        struct bset_tree *t,
880                                        struct bkey_packed *where,
881                                        unsigned clobber_u64s,
882                                        unsigned new_u64s)
883 {
884         const struct bkey_packed *end = btree_bkey_last(b, t);
885         struct btree_node_iter_set *set;
886         unsigned offset = __btree_node_key_to_offset(b, where);
887         int shift = new_u64s - clobber_u64s;
888         unsigned old_end = t->end_offset - shift;
889         unsigned orig_iter_pos = node_iter->data[0].k;
890         bool iter_current_key_modified =
891                 orig_iter_pos >= offset &&
892                 orig_iter_pos <= offset + clobber_u64s;
893
894         btree_node_iter_for_each(node_iter, set)
895                 if (set->end == old_end)
896                         goto found;
897
898         /* didn't find the bset in the iterator - might have to readd it: */
899         if (new_u64s &&
900             bkey_iter_pos_cmp(b, where, &path->pos) >= 0) {
901                 bch2_btree_node_iter_push(node_iter, b, where, end);
902                 goto fixup_done;
903         } else {
904                 /* Iterator is after key that changed */
905                 return;
906         }
907 found:
908         set->end = t->end_offset;
909
910         /* Iterator hasn't gotten to the key that changed yet: */
911         if (set->k < offset)
912                 return;
913
914         if (new_u64s &&
915             bkey_iter_pos_cmp(b, where, &path->pos) >= 0) {
916                 set->k = offset;
917         } else if (set->k < offset + clobber_u64s) {
918                 set->k = offset + new_u64s;
919                 if (set->k == set->end)
920                         bch2_btree_node_iter_set_drop(node_iter, set);
921         } else {
922                 /* Iterator is after key that changed */
923                 set->k = (int) set->k + shift;
924                 return;
925         }
926
927         bch2_btree_node_iter_sort(node_iter, b);
928 fixup_done:
929         if (node_iter->data[0].k != orig_iter_pos)
930                 iter_current_key_modified = true;
931
932         /*
933          * When a new key is added, and the node iterator now points to that
934          * key, the iterator might have skipped past deleted keys that should
935          * come after the key the iterator now points to. We have to rewind to
936          * before those deleted keys - otherwise
937          * bch2_btree_node_iter_prev_all() breaks:
938          */
939         if (!bch2_btree_node_iter_end(node_iter) &&
940             iter_current_key_modified &&
941             b->c.level) {
942                 struct bset_tree *t;
943                 struct bkey_packed *k, *k2, *p;
944
945                 k = bch2_btree_node_iter_peek_all(node_iter, b);
946
947                 for_each_bset(b, t) {
948                         bool set_pos = false;
949
950                         if (node_iter->data[0].end == t->end_offset)
951                                 continue;
952
953                         k2 = bch2_btree_node_iter_bset_pos(node_iter, b, t);
954
955                         while ((p = bch2_bkey_prev_all(b, t, k2)) &&
956                                bkey_iter_cmp(b, k, p) < 0) {
957                                 k2 = p;
958                                 set_pos = true;
959                         }
960
961                         if (set_pos)
962                                 btree_node_iter_set_set_pos(node_iter,
963                                                             b, t, k2);
964                 }
965         }
966 }
967
968 void bch2_btree_node_iter_fix(struct btree_trans *trans,
969                               struct btree_path *path,
970                               struct btree *b,
971                               struct btree_node_iter *node_iter,
972                               struct bkey_packed *where,
973                               unsigned clobber_u64s,
974                               unsigned new_u64s)
975 {
976         struct bset_tree *t = bch2_bkey_to_bset(b, where);
977         struct btree_path *linked;
978
979         if (node_iter != &path->l[b->c.level].iter) {
980                 __bch2_btree_node_iter_fix(path, b, node_iter, t,
981                                            where, clobber_u64s, new_u64s);
982
983                 if (bch2_debug_check_iterators)
984                         bch2_btree_node_iter_verify(node_iter, b);
985         }
986
987         trans_for_each_path_with_node(trans, b, linked) {
988                 __bch2_btree_node_iter_fix(linked, b,
989                                            &linked->l[b->c.level].iter, t,
990                                            where, clobber_u64s, new_u64s);
991                 bch2_btree_path_verify_level(trans, linked, b->c.level);
992         }
993 }
994
995 /* Btree path level: pointer to a particular btree node and node iter */
996
997 static inline struct bkey_s_c __btree_iter_unpack(struct bch_fs *c,
998                                                   struct btree_path_level *l,
999                                                   struct bkey *u,
1000                                                   struct bkey_packed *k)
1001 {
1002         if (unlikely(!k)) {
1003                 /*
1004                  * signal to bch2_btree_iter_peek_slot() that we're currently at
1005                  * a hole
1006                  */
1007                 u->type = KEY_TYPE_deleted;
1008                 return bkey_s_c_null;
1009         }
1010
1011         return bkey_disassemble(l->b, k, u);
1012 }
1013
1014 static inline struct bkey_s_c btree_path_level_peek_all(struct bch_fs *c,
1015                                                         struct btree_path_level *l,
1016                                                         struct bkey *u)
1017 {
1018         return __btree_iter_unpack(c, l, u,
1019                         bch2_btree_node_iter_peek_all(&l->iter, l->b));
1020 }
1021
1022 static inline struct bkey_s_c btree_path_level_peek(struct bch_fs *c,
1023                                                     struct btree_path *path,
1024                                                     struct btree_path_level *l,
1025                                                     struct bkey *u)
1026 {
1027         struct bkey_s_c k = __btree_iter_unpack(c, l, u,
1028                         bch2_btree_node_iter_peek(&l->iter, l->b));
1029
1030         path->pos = k.k ? k.k->p : l->b->key.k.p;
1031         return k;
1032 }
1033
1034 static inline struct bkey_s_c btree_path_level_prev(struct bch_fs *c,
1035                                                     struct btree_path *path,
1036                                                     struct btree_path_level *l,
1037                                                     struct bkey *u)
1038 {
1039         struct bkey_s_c k = __btree_iter_unpack(c, l, u,
1040                         bch2_btree_node_iter_prev(&l->iter, l->b));
1041
1042         path->pos = k.k ? k.k->p : l->b->data->min_key;
1043         return k;
1044 }
1045
1046 static inline bool btree_path_advance_to_pos(struct btree_path *path,
1047                                              struct btree_path_level *l,
1048                                              int max_advance)
1049 {
1050         struct bkey_packed *k;
1051         int nr_advanced = 0;
1052
1053         while ((k = bch2_btree_node_iter_peek_all(&l->iter, l->b)) &&
1054                bkey_iter_pos_cmp(l->b, k, &path->pos) < 0) {
1055                 if (max_advance > 0 && nr_advanced >= max_advance)
1056                         return false;
1057
1058                 bch2_btree_node_iter_advance(&l->iter, l->b);
1059                 nr_advanced++;
1060         }
1061
1062         return true;
1063 }
1064
1065 /*
1066  * Verify that iterator for parent node points to child node:
1067  */
1068 static void btree_path_verify_new_node(struct btree_trans *trans,
1069                                        struct btree_path *path, struct btree *b)
1070 {
1071         struct bch_fs *c = trans->c;
1072         struct btree_path_level *l;
1073         unsigned plevel;
1074         bool parent_locked;
1075         struct bkey_packed *k;
1076
1077         if (!IS_ENABLED(CONFIG_BCACHEFS_DEBUG))
1078                 return;
1079
1080         if (!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))
1081                 return;
1082
1083         plevel = b->c.level + 1;
1084         if (!btree_path_node(path, plevel))
1085                 return;
1086
1087         parent_locked = btree_node_locked(path, plevel);
1088
1089         if (!bch2_btree_node_relock(trans, path, plevel))
1090                 return;
1091
1092         l = &path->l[plevel];
1093         k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
1094         if (!k ||
1095             bkey_deleted(k) ||
1096             bkey_cmp_left_packed(l->b, k, &b->key.k.p)) {
1097                 struct printbuf buf1 = PRINTBUF;
1098                 struct printbuf buf2 = PRINTBUF;
1099                 struct printbuf buf3 = PRINTBUF;
1100                 struct printbuf buf4 = PRINTBUF;
1101                 struct bkey uk = bkey_unpack_key(b, k);
1102
1103                 bch2_dump_btree_node(c, l->b);
1104                 bch2_bpos_to_text(&buf1, path->pos);
1105                 bch2_bkey_to_text(&buf2, &uk);
1106                 bch2_bpos_to_text(&buf3, b->data->min_key);
1107                 bch2_bpos_to_text(&buf3, b->data->max_key);
1108                 panic("parent iter doesn't point to new node:\n"
1109                       "iter pos %s %s\n"
1110                       "iter key %s\n"
1111                       "new node %s-%s\n",
1112                       bch2_btree_ids[path->btree_id],
1113                       buf1.buf, buf2.buf, buf3.buf, buf4.buf);
1114         }
1115
1116         if (!parent_locked)
1117                 btree_node_unlock(path, plevel);
1118 }
1119
1120 static inline void __btree_path_level_init(struct btree_path *path,
1121                                            unsigned level)
1122 {
1123         struct btree_path_level *l = &path->l[level];
1124
1125         bch2_btree_node_iter_init(&l->iter, l->b, &path->pos);
1126
1127         /*
1128          * Iterators to interior nodes should always be pointed at the first non
1129          * whiteout:
1130          */
1131         if (level)
1132                 bch2_btree_node_iter_peek(&l->iter, l->b);
1133 }
1134
1135 static inline void btree_path_level_init(struct btree_trans *trans,
1136                                          struct btree_path *path,
1137                                          struct btree *b)
1138 {
1139         BUG_ON(path->cached);
1140
1141         btree_path_verify_new_node(trans, path, b);
1142
1143         EBUG_ON(!btree_path_pos_in_node(path, b));
1144         EBUG_ON(b->c.lock.state.seq & 1);
1145
1146         path->l[b->c.level].lock_seq = b->c.lock.state.seq;
1147         path->l[b->c.level].b = b;
1148         __btree_path_level_init(path, b->c.level);
1149 }
1150
1151 /* Btree path: fixups after btree node updates: */
1152
1153 /*
1154  * A btree node is being replaced - update the iterator to point to the new
1155  * node:
1156  */
1157 void bch2_trans_node_add(struct btree_trans *trans, struct btree *b)
1158 {
1159         struct btree_path *path;
1160
1161         trans_for_each_path(trans, path)
1162                 if (!path->cached &&
1163                     btree_path_pos_in_node(path, b)) {
1164                         enum btree_node_locked_type t =
1165                                 btree_lock_want(path, b->c.level);
1166
1167                         if (path->nodes_locked &&
1168                             t != BTREE_NODE_UNLOCKED) {
1169                                 btree_node_unlock(path, b->c.level);
1170                                 six_lock_increment(&b->c.lock, t);
1171                                 mark_btree_node_locked(trans, path, b->c.level, t);
1172                         }
1173
1174                         btree_path_level_init(trans, path, b);
1175                 }
1176 }
1177
1178 /*
1179  * A btree node has been modified in such a way as to invalidate iterators - fix
1180  * them:
1181  */
1182 void bch2_trans_node_reinit_iter(struct btree_trans *trans, struct btree *b)
1183 {
1184         struct btree_path *path;
1185
1186         trans_for_each_path_with_node(trans, b, path)
1187                 __btree_path_level_init(path, b->c.level);
1188 }
1189
1190 /* Btree path: traverse, set_pos: */
1191
1192 static int lock_root_check_fn(struct six_lock *lock, void *p)
1193 {
1194         struct btree *b = container_of(lock, struct btree, c.lock);
1195         struct btree **rootp = p;
1196
1197         return b == *rootp ? 0 : -1;
1198 }
1199
1200 static inline int btree_path_lock_root(struct btree_trans *trans,
1201                                        struct btree_path *path,
1202                                        unsigned depth_want,
1203                                        unsigned long trace_ip)
1204 {
1205         struct bch_fs *c = trans->c;
1206         struct btree *b, **rootp = &c->btree_roots[path->btree_id].b;
1207         enum six_lock_type lock_type;
1208         unsigned i;
1209
1210         EBUG_ON(path->nodes_locked);
1211
1212         while (1) {
1213                 b = READ_ONCE(*rootp);
1214                 path->level = READ_ONCE(b->c.level);
1215
1216                 if (unlikely(path->level < depth_want)) {
1217                         /*
1218                          * the root is at a lower depth than the depth we want:
1219                          * got to the end of the btree, or we're walking nodes
1220                          * greater than some depth and there are no nodes >=
1221                          * that depth
1222                          */
1223                         path->level = depth_want;
1224                         for (i = path->level; i < BTREE_MAX_DEPTH; i++)
1225                                 path->l[i].b = NULL;
1226                         return 1;
1227                 }
1228
1229                 lock_type = __btree_lock_want(path, path->level);
1230                 if (unlikely(!btree_node_lock(trans, path, b, SPOS_MAX,
1231                                               path->level, lock_type,
1232                                               lock_root_check_fn, rootp,
1233                                               trace_ip))) {
1234                         if (trans->restarted)
1235                                 return -EINTR;
1236                         continue;
1237                 }
1238
1239                 if (likely(b == READ_ONCE(*rootp) &&
1240                            b->c.level == path->level &&
1241                            !race_fault())) {
1242                         for (i = 0; i < path->level; i++)
1243                                 path->l[i].b = BTREE_ITER_NO_NODE_LOCK_ROOT;
1244                         path->l[path->level].b = b;
1245                         for (i = path->level + 1; i < BTREE_MAX_DEPTH; i++)
1246                                 path->l[i].b = NULL;
1247
1248                         mark_btree_node_locked(trans, path, path->level, lock_type);
1249                         btree_path_level_init(trans, path, b);
1250                         return 0;
1251                 }
1252
1253                 six_unlock_type(&b->c.lock, lock_type);
1254         }
1255 }
1256
1257 noinline
1258 static int btree_path_prefetch(struct btree_trans *trans, struct btree_path *path)
1259 {
1260         struct bch_fs *c = trans->c;
1261         struct btree_path_level *l = path_l(path);
1262         struct btree_node_iter node_iter = l->iter;
1263         struct bkey_packed *k;
1264         struct bkey_buf tmp;
1265         unsigned nr = test_bit(BCH_FS_STARTED, &c->flags)
1266                 ? (path->level > 1 ? 0 :  2)
1267                 : (path->level > 1 ? 1 : 16);
1268         bool was_locked = btree_node_locked(path, path->level);
1269         int ret = 0;
1270
1271         bch2_bkey_buf_init(&tmp);
1272
1273         while (nr && !ret) {
1274                 if (!bch2_btree_node_relock(trans, path, path->level))
1275                         break;
1276
1277                 bch2_btree_node_iter_advance(&node_iter, l->b);
1278                 k = bch2_btree_node_iter_peek(&node_iter, l->b);
1279                 if (!k)
1280                         break;
1281
1282                 bch2_bkey_buf_unpack(&tmp, c, l->b, k);
1283                 ret = bch2_btree_node_prefetch(c, trans, path, tmp.k, path->btree_id,
1284                                                path->level - 1);
1285         }
1286
1287         if (!was_locked)
1288                 btree_node_unlock(path, path->level);
1289
1290         bch2_bkey_buf_exit(&tmp, c);
1291         return ret;
1292 }
1293
1294 static int btree_path_prefetch_j(struct btree_trans *trans, struct btree_path *path,
1295                                  struct btree_and_journal_iter *jiter)
1296 {
1297         struct bch_fs *c = trans->c;
1298         struct bkey_s_c k;
1299         struct bkey_buf tmp;
1300         unsigned nr = test_bit(BCH_FS_STARTED, &c->flags)
1301                 ? (path->level > 1 ? 0 :  2)
1302                 : (path->level > 1 ? 1 : 16);
1303         bool was_locked = btree_node_locked(path, path->level);
1304         int ret = 0;
1305
1306         bch2_bkey_buf_init(&tmp);
1307
1308         while (nr && !ret) {
1309                 if (!bch2_btree_node_relock(trans, path, path->level))
1310                         break;
1311
1312                 bch2_btree_and_journal_iter_advance(jiter);
1313                 k = bch2_btree_and_journal_iter_peek(jiter);
1314                 if (!k.k)
1315                         break;
1316
1317                 bch2_bkey_buf_reassemble(&tmp, c, k);
1318                 ret = bch2_btree_node_prefetch(c, trans, path, tmp.k, path->btree_id,
1319                                                path->level - 1);
1320         }
1321
1322         if (!was_locked)
1323                 btree_node_unlock(path, path->level);
1324
1325         bch2_bkey_buf_exit(&tmp, c);
1326         return ret;
1327 }
1328
1329 static noinline void btree_node_mem_ptr_set(struct btree_trans *trans,
1330                                             struct btree_path *path,
1331                                             unsigned plevel, struct btree *b)
1332 {
1333         struct btree_path_level *l = &path->l[plevel];
1334         bool locked = btree_node_locked(path, plevel);
1335         struct bkey_packed *k;
1336         struct bch_btree_ptr_v2 *bp;
1337
1338         if (!bch2_btree_node_relock(trans, path, plevel))
1339                 return;
1340
1341         k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
1342         BUG_ON(k->type != KEY_TYPE_btree_ptr_v2);
1343
1344         bp = (void *) bkeyp_val(&l->b->format, k);
1345         bp->mem_ptr = (unsigned long)b;
1346
1347         if (!locked)
1348                 btree_node_unlock(path, plevel);
1349 }
1350
1351 static noinline int btree_node_iter_and_journal_peek(struct btree_trans *trans,
1352                                                      struct btree_path *path,
1353                                                      unsigned flags,
1354                                                      struct bkey_buf *out)
1355 {
1356         struct bch_fs *c = trans->c;
1357         struct btree_path_level *l = path_l(path);
1358         struct btree_and_journal_iter jiter;
1359         struct bkey_s_c k;
1360         int ret = 0;
1361
1362         __bch2_btree_and_journal_iter_init_node_iter(&jiter, c, l->b, l->iter, path->pos);
1363
1364         k = bch2_btree_and_journal_iter_peek(&jiter);
1365
1366         bch2_bkey_buf_reassemble(out, c, k);
1367
1368         if (flags & BTREE_ITER_PREFETCH)
1369                 ret = btree_path_prefetch_j(trans, path, &jiter);
1370
1371         bch2_btree_and_journal_iter_exit(&jiter);
1372         return ret;
1373 }
1374
1375 static __always_inline int btree_path_down(struct btree_trans *trans,
1376                                            struct btree_path *path,
1377                                            unsigned flags,
1378                                            unsigned long trace_ip)
1379 {
1380         struct bch_fs *c = trans->c;
1381         struct btree_path_level *l = path_l(path);
1382         struct btree *b;
1383         unsigned level = path->level - 1;
1384         enum six_lock_type lock_type = __btree_lock_want(path, level);
1385         bool replay_done = test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags);
1386         struct bkey_buf tmp;
1387         int ret;
1388
1389         EBUG_ON(!btree_node_locked(path, path->level));
1390
1391         bch2_bkey_buf_init(&tmp);
1392
1393         if (unlikely(!replay_done)) {
1394                 ret = btree_node_iter_and_journal_peek(trans, path, flags, &tmp);
1395                 if (ret)
1396                         goto err;
1397         } else {
1398                 bch2_bkey_buf_unpack(&tmp, c, l->b,
1399                                  bch2_btree_node_iter_peek(&l->iter, l->b));
1400
1401                 if (flags & BTREE_ITER_PREFETCH) {
1402                         ret = btree_path_prefetch(trans, path);
1403                         if (ret)
1404                                 goto err;
1405                 }
1406         }
1407
1408         b = bch2_btree_node_get(trans, path, tmp.k, level, lock_type, trace_ip);
1409         ret = PTR_ERR_OR_ZERO(b);
1410         if (unlikely(ret))
1411                 goto err;
1412
1413         mark_btree_node_locked(trans, path, level, lock_type);
1414         btree_path_level_init(trans, path, b);
1415
1416         if (likely(replay_done && tmp.k->k.type == KEY_TYPE_btree_ptr_v2) &&
1417             unlikely(b != btree_node_mem_ptr(tmp.k)))
1418                 btree_node_mem_ptr_set(trans, path, level + 1, b);
1419
1420         if (btree_node_read_locked(path, level + 1))
1421                 btree_node_unlock(path, level + 1);
1422         path->level = level;
1423
1424         bch2_btree_path_verify_locks(path);
1425 err:
1426         bch2_bkey_buf_exit(&tmp, c);
1427         return ret;
1428 }
1429
1430 static int btree_path_traverse_one(struct btree_trans *, struct btree_path *,
1431                                    unsigned, unsigned long);
1432
1433 static int bch2_btree_path_traverse_all(struct btree_trans *trans)
1434 {
1435         struct bch_fs *c = trans->c;
1436         struct btree_path *path;
1437         unsigned long trace_ip = _RET_IP_;
1438         int i, ret = 0;
1439
1440         if (trans->in_traverse_all)
1441                 return -EINTR;
1442
1443         trans->in_traverse_all = true;
1444 retry_all:
1445         trans->restarted = false;
1446         trans->traverse_all_idx = U8_MAX;
1447
1448         trans_for_each_path(trans, path)
1449                 path->should_be_locked = false;
1450
1451         btree_trans_verify_sorted(trans);
1452
1453         for (i = trans->nr_sorted - 2; i >= 0; --i) {
1454                 struct btree_path *path1 = trans->paths + trans->sorted[i];
1455                 struct btree_path *path2 = trans->paths + trans->sorted[i + 1];
1456
1457                 if (path1->btree_id == path2->btree_id &&
1458                     path1->locks_want < path2->locks_want)
1459                         __bch2_btree_path_upgrade(trans, path1, path2->locks_want);
1460                 else if (!path1->locks_want && path2->locks_want)
1461                         __bch2_btree_path_upgrade(trans, path1, 1);
1462         }
1463
1464         bch2_trans_unlock(trans);
1465         cond_resched();
1466
1467         if (unlikely(trans->memory_allocation_failure)) {
1468                 struct closure cl;
1469
1470                 closure_init_stack(&cl);
1471
1472                 do {
1473                         ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1474                         closure_sync(&cl);
1475                 } while (ret);
1476         }
1477
1478         /* Now, redo traversals in correct order: */
1479         trans->traverse_all_idx = 0;
1480         while (trans->traverse_all_idx < trans->nr_sorted) {
1481                 path = trans->paths + trans->sorted[trans->traverse_all_idx];
1482
1483                 /*
1484                  * Traversing a path can cause another path to be added at about
1485                  * the same position:
1486                  */
1487                 if (path->uptodate) {
1488                         ret = btree_path_traverse_one(trans, path, 0, _THIS_IP_);
1489                         if (ret == -EINTR || ret == -ENOMEM)
1490                                 goto retry_all;
1491                         if (ret)
1492                                 goto err;
1493                         BUG_ON(path->uptodate);
1494                 } else {
1495                         trans->traverse_all_idx++;
1496                 }
1497         }
1498
1499         /*
1500          * BTREE_ITER_NEED_RELOCK is ok here - if we called bch2_trans_unlock()
1501          * and relock(), relock() won't relock since path->should_be_locked
1502          * isn't set yet, which is all fine
1503          */
1504         trans_for_each_path(trans, path)
1505                 BUG_ON(path->uptodate >= BTREE_ITER_NEED_TRAVERSE);
1506 err:
1507         bch2_btree_cache_cannibalize_unlock(c);
1508
1509         trans->in_traverse_all = false;
1510
1511         trace_trans_traverse_all(trans->fn, trace_ip);
1512         return ret;
1513 }
1514
1515 static inline bool btree_path_good_node(struct btree_trans *trans,
1516                                         struct btree_path *path,
1517                                         unsigned l, int check_pos)
1518 {
1519         if (!is_btree_node(path, l) ||
1520             !bch2_btree_node_relock(trans, path, l))
1521                 return false;
1522
1523         if (check_pos < 0 && btree_path_pos_before_node(path, path->l[l].b))
1524                 return false;
1525         if (check_pos > 0 && btree_path_pos_after_node(path, path->l[l].b))
1526                 return false;
1527         return true;
1528 }
1529
1530 static inline unsigned btree_path_up_until_good_node(struct btree_trans *trans,
1531                                                      struct btree_path *path,
1532                                                      int check_pos)
1533 {
1534         unsigned i, l = path->level;
1535
1536         while (btree_path_node(path, l) &&
1537                !btree_path_good_node(trans, path, l, check_pos)) {
1538                 btree_node_unlock(path, l);
1539                 path->l[l].b = BTREE_ITER_NO_NODE_UP;
1540                 l++;
1541         }
1542
1543         /* If we need intent locks, take them too: */
1544         for (i = l + 1;
1545              i < path->locks_want && btree_path_node(path, i);
1546              i++)
1547                 if (!bch2_btree_node_relock(trans, path, i))
1548                         while (l <= i) {
1549                                 btree_node_unlock(path, l);
1550                                 path->l[l].b = BTREE_ITER_NO_NODE_UP;
1551                                 l++;
1552                         }
1553
1554         return l;
1555 }
1556
1557 /*
1558  * This is the main state machine for walking down the btree - walks down to a
1559  * specified depth
1560  *
1561  * Returns 0 on success, -EIO on error (error reading in a btree node).
1562  *
1563  * On error, caller (peek_node()/peek_key()) must return NULL; the error is
1564  * stashed in the iterator and returned from bch2_trans_exit().
1565  */
1566 static int btree_path_traverse_one(struct btree_trans *trans,
1567                                    struct btree_path *path,
1568                                    unsigned flags,
1569                                    unsigned long trace_ip)
1570 {
1571         unsigned depth_want = path->level;
1572         int ret = 0;
1573
1574         if (unlikely(trans->restarted)) {
1575                 ret = -EINTR;
1576                 goto out;
1577         }
1578
1579         /*
1580          * Ensure we obey path->should_be_locked: if it's set, we can't unlock
1581          * and re-traverse the path without a transaction restart:
1582          */
1583         if (path->should_be_locked) {
1584                 ret = bch2_btree_path_relock(trans, path, trace_ip) ? 0 : -EINTR;
1585                 goto out;
1586         }
1587
1588         if (path->cached) {
1589                 ret = bch2_btree_path_traverse_cached(trans, path, flags);
1590                 goto out;
1591         }
1592
1593         if (unlikely(path->level >= BTREE_MAX_DEPTH))
1594                 goto out;
1595
1596         path->level = btree_path_up_until_good_node(trans, path, 0);
1597
1598         /*
1599          * Note: path->nodes[path->level] may be temporarily NULL here - that
1600          * would indicate to other code that we got to the end of the btree,
1601          * here it indicates that relocking the root failed - it's critical that
1602          * btree_path_lock_root() comes next and that it can't fail
1603          */
1604         while (path->level > depth_want) {
1605                 ret = btree_path_node(path, path->level)
1606                         ? btree_path_down(trans, path, flags, trace_ip)
1607                         : btree_path_lock_root(trans, path, depth_want, trace_ip);
1608                 if (unlikely(ret)) {
1609                         if (ret == 1) {
1610                                 /*
1611                                  * No nodes at this level - got to the end of
1612                                  * the btree:
1613                                  */
1614                                 ret = 0;
1615                                 goto out;
1616                         }
1617
1618                         __bch2_btree_path_unlock(path);
1619                         path->level = depth_want;
1620
1621                         if (ret == -EIO)
1622                                 path->l[path->level].b =
1623                                         BTREE_ITER_NO_NODE_ERROR;
1624                         else
1625                                 path->l[path->level].b =
1626                                         BTREE_ITER_NO_NODE_DOWN;
1627                         goto out;
1628                 }
1629         }
1630
1631         path->uptodate = BTREE_ITER_UPTODATE;
1632 out:
1633         BUG_ON((ret == -EINTR) != !!trans->restarted);
1634         bch2_btree_path_verify(trans, path);
1635         return ret;
1636 }
1637
1638 int __must_check bch2_btree_path_traverse(struct btree_trans *trans,
1639                                           struct btree_path *path, unsigned flags)
1640 {
1641         if (path->uptodate < BTREE_ITER_NEED_RELOCK)
1642                 return 0;
1643
1644         return  bch2_trans_cond_resched(trans) ?:
1645                 btree_path_traverse_one(trans, path, flags, _RET_IP_);
1646 }
1647
1648 static void btree_path_copy(struct btree_trans *trans, struct btree_path *dst,
1649                             struct btree_path *src)
1650 {
1651         unsigned i;
1652
1653         memcpy(&dst->pos, &src->pos,
1654                sizeof(struct btree_path) - offsetof(struct btree_path, pos));
1655
1656         for (i = 0; i < BTREE_MAX_DEPTH; i++)
1657                 if (btree_node_locked(dst, i))
1658                         six_lock_increment(&dst->l[i].b->c.lock,
1659                                            __btree_lock_want(dst, i));
1660
1661         bch2_btree_path_check_sort(trans, dst, 0);
1662 }
1663
1664 static struct btree_path *btree_path_clone(struct btree_trans *trans, struct btree_path *src,
1665                                            bool intent)
1666 {
1667         struct btree_path *new = btree_path_alloc(trans, src);
1668
1669         btree_path_copy(trans, new, src);
1670         __btree_path_get(new, intent);
1671         return new;
1672 }
1673
1674 inline struct btree_path * __must_check
1675 bch2_btree_path_make_mut(struct btree_trans *trans,
1676                          struct btree_path *path, bool intent,
1677                          unsigned long ip)
1678 {
1679         if (path->ref > 1 || path->preserve) {
1680                 __btree_path_put(path, intent);
1681                 path = btree_path_clone(trans, path, intent);
1682                 path->preserve = false;
1683 #ifdef CONFIG_BCACHEFS_DEBUG
1684                 path->ip_allocated = ip;
1685 #endif
1686                 btree_trans_verify_sorted(trans);
1687         }
1688
1689         path->should_be_locked = false;
1690         return path;
1691 }
1692
1693 struct btree_path * __must_check
1694 bch2_btree_path_set_pos(struct btree_trans *trans,
1695                    struct btree_path *path, struct bpos new_pos,
1696                    bool intent, unsigned long ip)
1697 {
1698         int cmp = bpos_cmp(new_pos, path->pos);
1699         unsigned l = path->level;
1700
1701         EBUG_ON(trans->restarted);
1702         EBUG_ON(!path->ref);
1703
1704         if (!cmp)
1705                 return path;
1706
1707         path = bch2_btree_path_make_mut(trans, path, intent, ip);
1708
1709         path->pos = new_pos;
1710
1711         bch2_btree_path_check_sort(trans, path, cmp);
1712
1713         if (unlikely(path->cached)) {
1714                 btree_node_unlock(path, 0);
1715                 path->l[0].b = BTREE_ITER_NO_NODE_CACHED;
1716                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1717                 goto out;
1718         }
1719
1720         l = btree_path_up_until_good_node(trans, path, cmp);
1721
1722         if (btree_path_node(path, l)) {
1723                 BUG_ON(!btree_node_locked(path, l));
1724                 /*
1725                  * We might have to skip over many keys, or just a few: try
1726                  * advancing the node iterator, and if we have to skip over too
1727                  * many keys just reinit it (or if we're rewinding, since that
1728                  * is expensive).
1729                  */
1730                 if (cmp < 0 ||
1731                     !btree_path_advance_to_pos(path, &path->l[l], 8))
1732                         __btree_path_level_init(path, l);
1733         }
1734
1735         if (l != path->level) {
1736                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1737                 __bch2_btree_path_unlock(path);
1738         }
1739 out:
1740         bch2_btree_path_verify(trans, path);
1741         return path;
1742 }
1743
1744 /* Btree path: main interface: */
1745
1746 static struct btree_path *have_path_at_pos(struct btree_trans *trans, struct btree_path *path)
1747 {
1748         struct btree_path *next;
1749
1750         next = prev_btree_path(trans, path);
1751         if (next && !btree_path_cmp(next, path))
1752                 return next;
1753
1754         next = next_btree_path(trans, path);
1755         if (next && !btree_path_cmp(next, path))
1756                 return next;
1757
1758         return NULL;
1759 }
1760
1761 static struct btree_path *have_node_at_pos(struct btree_trans *trans, struct btree_path *path)
1762 {
1763         struct btree_path *next;
1764
1765         next = prev_btree_path(trans, path);
1766         if (next && next->level == path->level && path_l(next)->b == path_l(path)->b)
1767                 return next;
1768
1769         next = next_btree_path(trans, path);
1770         if (next && next->level == path->level && path_l(next)->b == path_l(path)->b)
1771                 return next;
1772
1773         return NULL;
1774 }
1775
1776 static inline void __bch2_path_free(struct btree_trans *trans, struct btree_path *path)
1777 {
1778         __bch2_btree_path_unlock(path);
1779         btree_path_list_remove(trans, path);
1780         trans->paths_allocated &= ~(1ULL << path->idx);
1781 }
1782
1783 void bch2_path_put(struct btree_trans *trans, struct btree_path *path, bool intent)
1784 {
1785         struct btree_path *dup;
1786
1787         EBUG_ON(trans->paths + path->idx != path);
1788         EBUG_ON(!path->ref);
1789
1790         if (!__btree_path_put(path, intent))
1791                 return;
1792
1793         /*
1794          * Perhaps instead we should check for duplicate paths in traverse_all:
1795          */
1796         if (path->preserve &&
1797             (dup = have_path_at_pos(trans, path))) {
1798                 dup->preserve = true;
1799                 path->preserve = false;
1800                 goto free;
1801         }
1802
1803         if (!path->preserve &&
1804             (dup = have_node_at_pos(trans, path)))
1805                 goto free;
1806         return;
1807 free:
1808         if (path->should_be_locked &&
1809             !btree_node_locked(dup, path->level))
1810                 return;
1811
1812         dup->should_be_locked |= path->should_be_locked;
1813         __bch2_path_free(trans, path);
1814 }
1815
1816 void bch2_trans_updates_to_text(struct printbuf *buf, struct btree_trans *trans)
1817 {
1818         struct btree_insert_entry *i;
1819
1820         pr_buf(buf, "transaction updates for %s journal seq %llu",
1821                trans->fn, trans->journal_res.seq);
1822         pr_newline(buf);
1823         pr_indent_push(buf, 2);
1824
1825         trans_for_each_update(trans, i) {
1826                 struct bkey_s_c old = { &i->old_k, i->old_v };
1827
1828                 pr_buf(buf, "update: btree %s %pS",
1829                        bch2_btree_ids[i->btree_id],
1830                        (void *) i->ip_allocated);
1831                 pr_newline(buf);
1832
1833                 pr_buf(buf, "  old ");
1834                 bch2_bkey_val_to_text(buf, trans->c, old);
1835                 pr_newline(buf);
1836
1837                 pr_buf(buf, "  new ");
1838                 bch2_bkey_val_to_text(buf, trans->c, bkey_i_to_s_c(i->k));
1839                 pr_newline(buf);
1840         }
1841
1842         pr_indent_pop(buf, 2);
1843 }
1844
1845 noinline __cold
1846 void bch2_dump_trans_updates(struct btree_trans *trans)
1847 {
1848         struct printbuf buf = PRINTBUF;
1849
1850         bch2_trans_updates_to_text(&buf, trans);
1851         bch_err(trans->c, "%s", buf.buf);
1852         printbuf_exit(&buf);
1853 }
1854
1855 noinline __cold
1856 void bch2_dump_trans_paths_updates(struct btree_trans *trans)
1857 {
1858         struct btree_path *path;
1859         struct printbuf buf = PRINTBUF;
1860         unsigned idx;
1861
1862         trans_for_each_path_inorder(trans, path, idx) {
1863                 printbuf_reset(&buf);
1864
1865                 bch2_bpos_to_text(&buf, path->pos);
1866
1867                 printk(KERN_ERR "path: idx %u ref %u:%u%s%s btree=%s l=%u pos %s locks %u %pS\n",
1868                        path->idx, path->ref, path->intent_ref,
1869                        path->should_be_locked ? " S" : "",
1870                        path->preserve ? " P" : "",
1871                        bch2_btree_ids[path->btree_id],
1872                        path->level,
1873                        buf.buf,
1874                        path->nodes_locked,
1875 #ifdef CONFIG_BCACHEFS_DEBUG
1876                        (void *) path->ip_allocated
1877 #else
1878                        NULL
1879 #endif
1880                        );
1881         }
1882
1883         printbuf_exit(&buf);
1884
1885         bch2_dump_trans_updates(trans);
1886 }
1887
1888 static struct btree_path *btree_path_alloc(struct btree_trans *trans,
1889                                            struct btree_path *pos)
1890 {
1891         struct btree_path *path;
1892         unsigned idx;
1893
1894         if (unlikely(trans->paths_allocated ==
1895                      ~((~0ULL << 1) << (BTREE_ITER_MAX - 1)))) {
1896                 bch2_dump_trans_paths_updates(trans);
1897                 panic("trans path oveflow\n");
1898         }
1899
1900         idx = __ffs64(~trans->paths_allocated);
1901         trans->paths_allocated |= 1ULL << idx;
1902
1903         path = &trans->paths[idx];
1904
1905         path->idx               = idx;
1906         path->ref               = 0;
1907         path->intent_ref        = 0;
1908         path->nodes_locked      = 0;
1909         path->nodes_intent_locked = 0;
1910
1911         btree_path_list_add(trans, pos, path);
1912         return path;
1913 }
1914
1915 struct btree_path *bch2_path_get(struct btree_trans *trans,
1916                                  enum btree_id btree_id, struct bpos pos,
1917                                  unsigned locks_want, unsigned level,
1918                                  unsigned flags, unsigned long ip)
1919 {
1920         struct btree_path *path, *path_pos = NULL;
1921         bool cached = flags & BTREE_ITER_CACHED;
1922         bool intent = flags & BTREE_ITER_INTENT;
1923         int i;
1924
1925         BUG_ON(trans->restarted);
1926         btree_trans_verify_sorted(trans);
1927         bch2_trans_verify_locks(trans);
1928
1929         trans_for_each_path_inorder(trans, path, i) {
1930                 if (__btree_path_cmp(path,
1931                                      btree_id,
1932                                      cached,
1933                                      pos,
1934                                      level) > 0)
1935                         break;
1936
1937                 path_pos = path;
1938         }
1939
1940         if (path_pos &&
1941             path_pos->cached    == cached &&
1942             path_pos->btree_id  == btree_id &&
1943             path_pos->level     == level) {
1944                 __btree_path_get(path_pos, intent);
1945                 path = bch2_btree_path_set_pos(trans, path_pos, pos, intent, ip);
1946         } else {
1947                 path = btree_path_alloc(trans, path_pos);
1948                 path_pos = NULL;
1949
1950                 __btree_path_get(path, intent);
1951                 path->pos                       = pos;
1952                 path->btree_id                  = btree_id;
1953                 path->cached                    = cached;
1954                 path->uptodate                  = BTREE_ITER_NEED_TRAVERSE;
1955                 path->should_be_locked          = false;
1956                 path->level                     = level;
1957                 path->locks_want                = locks_want;
1958                 path->nodes_locked              = 0;
1959                 path->nodes_intent_locked       = 0;
1960                 for (i = 0; i < ARRAY_SIZE(path->l); i++)
1961                         path->l[i].b            = BTREE_ITER_NO_NODE_INIT;
1962 #ifdef CONFIG_BCACHEFS_DEBUG
1963                 path->ip_allocated              = ip;
1964 #endif
1965                 btree_trans_verify_sorted(trans);
1966         }
1967
1968         if (!(flags & BTREE_ITER_NOPRESERVE))
1969                 path->preserve = true;
1970
1971         if (path->intent_ref)
1972                 locks_want = max(locks_want, level + 1);
1973
1974         /*
1975          * If the path has locks_want greater than requested, we don't downgrade
1976          * it here - on transaction restart because btree node split needs to
1977          * upgrade locks, we might be putting/getting the iterator again.
1978          * Downgrading iterators only happens via bch2_trans_downgrade(), after
1979          * a successful transaction commit.
1980          */
1981
1982         locks_want = min(locks_want, BTREE_MAX_DEPTH);
1983         if (locks_want > path->locks_want) {
1984                 path->locks_want = locks_want;
1985                 btree_path_get_locks(trans, path, true);
1986         }
1987
1988         return path;
1989 }
1990
1991 inline struct bkey_s_c bch2_btree_path_peek_slot(struct btree_path *path, struct bkey *u)
1992 {
1993
1994         struct bkey_s_c k;
1995
1996         if (!path->cached) {
1997                 struct btree_path_level *l = path_l(path);
1998                 struct bkey_packed *_k;
1999
2000                 EBUG_ON(path->uptodate != BTREE_ITER_UPTODATE);
2001
2002                 _k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
2003                 k = _k ? bkey_disassemble(l->b, _k, u) : bkey_s_c_null;
2004
2005                 EBUG_ON(k.k && bkey_deleted(k.k) && bpos_cmp(k.k->p, path->pos) == 0);
2006
2007                 if (!k.k || bpos_cmp(path->pos, k.k->p))
2008                         goto hole;
2009         } else {
2010                 struct bkey_cached *ck = (void *) path->l[0].b;
2011
2012                 EBUG_ON(ck &&
2013                         (path->btree_id != ck->key.btree_id ||
2014                          bkey_cmp(path->pos, ck->key.pos)));
2015
2016                 /* BTREE_ITER_CACHED_NOFILL|BTREE_ITER_CACHED_NOCREATE? */
2017                 if (unlikely(!ck || !ck->valid))
2018                         return bkey_s_c_null;
2019
2020                 EBUG_ON(path->uptodate != BTREE_ITER_UPTODATE);
2021
2022                 *u = ck->k->k;
2023                 k = bkey_i_to_s_c(ck->k);
2024         }
2025
2026         return k;
2027 hole:
2028         bkey_init(u);
2029         u->p = path->pos;
2030         return (struct bkey_s_c) { u, NULL };
2031 }
2032
2033 /* Btree iterators: */
2034
2035 int __must_check
2036 __bch2_btree_iter_traverse(struct btree_iter *iter)
2037 {
2038         return bch2_btree_path_traverse(iter->trans, iter->path, iter->flags);
2039 }
2040
2041 int __must_check
2042 bch2_btree_iter_traverse(struct btree_iter *iter)
2043 {
2044         int ret;
2045
2046         iter->path = bch2_btree_path_set_pos(iter->trans, iter->path,
2047                                         btree_iter_search_key(iter),
2048                                         iter->flags & BTREE_ITER_INTENT,
2049                                         btree_iter_ip_allocated(iter));
2050
2051         ret = bch2_btree_path_traverse(iter->trans, iter->path, iter->flags);
2052         if (ret)
2053                 return ret;
2054
2055         iter->path->should_be_locked = true;
2056         return 0;
2057 }
2058
2059 /* Iterate across nodes (leaf and interior nodes) */
2060
2061 struct btree *bch2_btree_iter_peek_node(struct btree_iter *iter)
2062 {
2063         struct btree_trans *trans = iter->trans;
2064         struct btree *b = NULL;
2065         int ret;
2066
2067         EBUG_ON(iter->path->cached);
2068         bch2_btree_iter_verify(iter);
2069
2070         ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2071         if (ret)
2072                 goto err;
2073
2074         b = btree_path_node(iter->path, iter->path->level);
2075         if (!b)
2076                 goto out;
2077
2078         BUG_ON(bpos_cmp(b->key.k.p, iter->pos) < 0);
2079
2080         bkey_init(&iter->k);
2081         iter->k.p = iter->pos = b->key.k.p;
2082
2083         iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p,
2084                                         iter->flags & BTREE_ITER_INTENT,
2085                                         btree_iter_ip_allocated(iter));
2086         iter->path->should_be_locked = true;
2087         BUG_ON(iter->path->uptodate);
2088 out:
2089         bch2_btree_iter_verify_entry_exit(iter);
2090         bch2_btree_iter_verify(iter);
2091
2092         return b;
2093 err:
2094         b = ERR_PTR(ret);
2095         goto out;
2096 }
2097
2098 struct btree *bch2_btree_iter_next_node(struct btree_iter *iter)
2099 {
2100         struct btree_trans *trans = iter->trans;
2101         struct btree_path *path = iter->path;
2102         struct btree *b = NULL;
2103         unsigned l;
2104         int ret;
2105
2106         BUG_ON(trans->restarted);
2107         EBUG_ON(iter->path->cached);
2108         bch2_btree_iter_verify(iter);
2109
2110         /* already at end? */
2111         if (!btree_path_node(path, path->level))
2112                 return NULL;
2113
2114         /* got to end? */
2115         if (!btree_path_node(path, path->level + 1)) {
2116                 btree_node_unlock(path, path->level);
2117                 path->l[path->level].b = BTREE_ITER_NO_NODE_UP;
2118                 path->level++;
2119                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
2120                 return NULL;
2121         }
2122
2123         if (!bch2_btree_node_relock(trans, path, path->level + 1)) {
2124                 __bch2_btree_path_unlock(path);
2125                 path->l[path->level].b = BTREE_ITER_NO_NODE_GET_LOCKS;
2126                 path->l[path->level + 1].b = BTREE_ITER_NO_NODE_GET_LOCKS;
2127                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
2128                 trace_trans_restart_relock_next_node(trans->fn, _THIS_IP_,
2129                                            path->btree_id, &path->pos);
2130                 btree_trans_restart(trans);
2131                 ret = -EINTR;
2132                 goto err;
2133         }
2134
2135         b = btree_path_node(path, path->level + 1);
2136
2137         if (!bpos_cmp(iter->pos, b->key.k.p)) {
2138                 btree_node_unlock(path, path->level);
2139                 path->l[path->level].b = BTREE_ITER_NO_NODE_UP;
2140                 path->level++;
2141         } else {
2142                 /*
2143                  * Haven't gotten to the end of the parent node: go back down to
2144                  * the next child node
2145                  */
2146                 path = iter->path =
2147                         bch2_btree_path_set_pos(trans, path, bpos_successor(iter->pos),
2148                                            iter->flags & BTREE_ITER_INTENT,
2149                                            btree_iter_ip_allocated(iter));
2150
2151                 path->level = iter->min_depth;
2152
2153                 for (l = path->level + 1; l < BTREE_MAX_DEPTH; l++)
2154                         if (btree_lock_want(path, l) == BTREE_NODE_UNLOCKED)
2155                                 btree_node_unlock(path, l);
2156
2157                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
2158                 bch2_btree_iter_verify(iter);
2159
2160                 ret = bch2_btree_path_traverse(trans, path, iter->flags);
2161                 if (ret)
2162                         goto err;
2163
2164                 b = path->l[path->level].b;
2165         }
2166
2167         bkey_init(&iter->k);
2168         iter->k.p = iter->pos = b->key.k.p;
2169
2170         iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p,
2171                                         iter->flags & BTREE_ITER_INTENT,
2172                                         btree_iter_ip_allocated(iter));
2173         iter->path->should_be_locked = true;
2174         BUG_ON(iter->path->uptodate);
2175 out:
2176         bch2_btree_iter_verify_entry_exit(iter);
2177         bch2_btree_iter_verify(iter);
2178
2179         return b;
2180 err:
2181         b = ERR_PTR(ret);
2182         goto out;
2183 }
2184
2185 /* Iterate across keys (in leaf nodes only) */
2186
2187 inline bool bch2_btree_iter_advance(struct btree_iter *iter)
2188 {
2189         struct bpos pos = iter->k.p;
2190         bool ret = (iter->flags & BTREE_ITER_ALL_SNAPSHOTS
2191                     ? bpos_cmp(pos, SPOS_MAX)
2192                     : bkey_cmp(pos, SPOS_MAX)) != 0;
2193
2194         if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
2195                 pos = bkey_successor(iter, pos);
2196         bch2_btree_iter_set_pos(iter, pos);
2197         return ret;
2198 }
2199
2200 inline bool bch2_btree_iter_rewind(struct btree_iter *iter)
2201 {
2202         struct bpos pos = bkey_start_pos(&iter->k);
2203         bool ret = (iter->flags & BTREE_ITER_ALL_SNAPSHOTS
2204                     ? bpos_cmp(pos, POS_MIN)
2205                     : bkey_cmp(pos, POS_MIN)) != 0;
2206
2207         if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
2208                 pos = bkey_predecessor(iter, pos);
2209         bch2_btree_iter_set_pos(iter, pos);
2210         return ret;
2211 }
2212
2213 static inline struct bkey_i *btree_trans_peek_updates(struct btree_trans *trans,
2214                                                       enum btree_id btree_id,
2215                                                       struct bpos pos)
2216 {
2217         struct btree_insert_entry *i;
2218
2219         trans_for_each_update(trans, i)
2220                 if ((cmp_int(btree_id,  i->btree_id) ?:
2221                      bpos_cmp(pos,      i->k->k.p)) <= 0) {
2222                         if (btree_id == i->btree_id)
2223                                 return i->k;
2224                         break;
2225                 }
2226
2227         return NULL;
2228 }
2229
2230 static noinline
2231 struct bkey_s_c btree_trans_peek_journal(struct btree_trans *trans,
2232                                          struct btree_iter *iter,
2233                                          struct bkey_s_c k)
2234 {
2235         struct bkey_i *next_journal =
2236                 bch2_journal_keys_peek_upto(trans->c, iter->btree_id, 0,
2237                                 iter->path->pos,
2238                                 k.k ? k.k->p : iter->path->l[0].b->key.k.p);
2239
2240         if (next_journal) {
2241                 iter->k = next_journal->k;
2242                 k = bkey_i_to_s_c(next_journal);
2243         }
2244
2245         return k;
2246 }
2247
2248 /*
2249  * Checks btree key cache for key at iter->pos and returns it if present, or
2250  * bkey_s_c_null:
2251  */
2252 static noinline
2253 struct bkey_s_c btree_trans_peek_key_cache(struct btree_iter *iter, struct bpos pos)
2254 {
2255         struct btree_trans *trans = iter->trans;
2256         struct bch_fs *c = trans->c;
2257         struct bkey u;
2258         int ret;
2259
2260         if (!bch2_btree_key_cache_find(c, iter->btree_id, pos))
2261                 return bkey_s_c_null;
2262
2263         if (!iter->key_cache_path)
2264                 iter->key_cache_path = bch2_path_get(trans, iter->btree_id, pos,
2265                                                      iter->flags & BTREE_ITER_INTENT, 0,
2266                                                      iter->flags|BTREE_ITER_CACHED,
2267                                                      _THIS_IP_);
2268
2269         iter->key_cache_path = bch2_btree_path_set_pos(trans, iter->key_cache_path, pos,
2270                                         iter->flags & BTREE_ITER_INTENT,
2271                                         btree_iter_ip_allocated(iter));
2272
2273         ret = bch2_btree_path_traverse(trans, iter->key_cache_path, iter->flags|BTREE_ITER_CACHED);
2274         if (unlikely(ret))
2275                 return bkey_s_c_err(ret);
2276
2277         iter->key_cache_path->should_be_locked = true;
2278
2279         return bch2_btree_path_peek_slot(iter->key_cache_path, &u);
2280 }
2281
2282 static struct bkey_s_c __bch2_btree_iter_peek(struct btree_iter *iter, struct bpos search_key)
2283 {
2284         struct btree_trans *trans = iter->trans;
2285         struct bkey_i *next_update;
2286         struct bkey_s_c k, k2;
2287         int ret;
2288
2289         EBUG_ON(iter->path->cached || iter->path->level);
2290         bch2_btree_iter_verify(iter);
2291
2292         while (1) {
2293                 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2294                                         iter->flags & BTREE_ITER_INTENT,
2295                                         btree_iter_ip_allocated(iter));
2296
2297                 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2298                 if (unlikely(ret)) {
2299                         /* ensure that iter->k is consistent with iter->pos: */
2300                         bch2_btree_iter_set_pos(iter, iter->pos);
2301                         k = bkey_s_c_err(ret);
2302                         goto out;
2303                 }
2304
2305                 iter->path->should_be_locked = true;
2306
2307                 k = btree_path_level_peek_all(trans->c, &iter->path->l[0], &iter->k);
2308
2309                 if (unlikely(iter->flags & BTREE_ITER_WITH_KEY_CACHE) &&
2310                     k.k &&
2311                     (k2 = btree_trans_peek_key_cache(iter, k.k->p)).k) {
2312                         ret = bkey_err(k2);
2313                         if (ret) {
2314                                 k = k2;
2315                                 bch2_btree_iter_set_pos(iter, iter->pos);
2316                                 goto out;
2317                         }
2318
2319                         k = k2;
2320                         iter->k = *k.k;
2321                 }
2322
2323                 if (unlikely(iter->flags & BTREE_ITER_WITH_JOURNAL))
2324                         k = btree_trans_peek_journal(trans, iter, k);
2325
2326                 next_update = iter->flags & BTREE_ITER_WITH_UPDATES
2327                         ? btree_trans_peek_updates(trans, iter->btree_id, search_key)
2328                         : NULL;
2329                 if (next_update &&
2330                     bpos_cmp(next_update->k.p,
2331                              k.k ? k.k->p : iter->path->l[0].b->key.k.p) <= 0) {
2332                         iter->k = next_update->k;
2333                         k = bkey_i_to_s_c(next_update);
2334                 }
2335
2336                 if (k.k && bkey_deleted(k.k)) {
2337                         /*
2338                          * If we've got a whiteout, and it's after the search
2339                          * key, advance the search key to the whiteout instead
2340                          * of just after the whiteout - it might be a btree
2341                          * whiteout, with a real key at the same position, since
2342                          * in the btree deleted keys sort before non deleted.
2343                          */
2344                         search_key = bpos_cmp(search_key, k.k->p)
2345                                 ? k.k->p
2346                                 : bpos_successor(k.k->p);
2347                         continue;
2348                 }
2349
2350                 if (likely(k.k)) {
2351                         break;
2352                 } else if (likely(bpos_cmp(iter->path->l[0].b->key.k.p, SPOS_MAX))) {
2353                         /* Advance to next leaf node: */
2354                         search_key = bpos_successor(iter->path->l[0].b->key.k.p);
2355                 } else {
2356                         /* End of btree: */
2357                         bch2_btree_iter_set_pos(iter, SPOS_MAX);
2358                         k = bkey_s_c_null;
2359                         goto out;
2360                 }
2361         }
2362 out:
2363         bch2_btree_iter_verify(iter);
2364
2365         return k;
2366 }
2367
2368 /**
2369  * bch2_btree_iter_peek: returns first key greater than or equal to iterator's
2370  * current position
2371  */
2372 struct bkey_s_c bch2_btree_iter_peek_upto(struct btree_iter *iter, struct bpos end)
2373 {
2374         struct btree_trans *trans = iter->trans;
2375         struct bpos search_key = btree_iter_search_key(iter);
2376         struct bkey_s_c k;
2377         struct bpos iter_pos;
2378         int ret;
2379
2380         if (iter->update_path) {
2381                 bch2_path_put(trans, iter->update_path,
2382                               iter->flags & BTREE_ITER_INTENT);
2383                 iter->update_path = NULL;
2384         }
2385
2386         bch2_btree_iter_verify_entry_exit(iter);
2387
2388         while (1) {
2389                 k = __bch2_btree_iter_peek(iter, search_key);
2390                 if (!k.k || bkey_err(k))
2391                         goto out;
2392
2393                 /*
2394                  * iter->pos should be mononotically increasing, and always be
2395                  * equal to the key we just returned - except extents can
2396                  * straddle iter->pos:
2397                  */
2398                 if (!(iter->flags & BTREE_ITER_IS_EXTENTS))
2399                         iter_pos = k.k->p;
2400                 else if (bkey_cmp(bkey_start_pos(k.k), iter->pos) > 0)
2401                         iter_pos = bkey_start_pos(k.k);
2402                 else
2403                         iter_pos = iter->pos;
2404
2405                 if (bkey_cmp(iter_pos, end) > 0) {
2406                         bch2_btree_iter_set_pos(iter, end);
2407                         k = bkey_s_c_null;
2408                         goto out;
2409                 }
2410
2411                 if (iter->update_path &&
2412                     bkey_cmp(iter->update_path->pos, k.k->p)) {
2413                         bch2_path_put(trans, iter->update_path,
2414                                       iter->flags & BTREE_ITER_INTENT);
2415                         iter->update_path = NULL;
2416                 }
2417
2418                 if ((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
2419                     (iter->flags & BTREE_ITER_INTENT) &&
2420                     !(iter->flags & BTREE_ITER_IS_EXTENTS) &&
2421                     !iter->update_path) {
2422                         struct bpos pos = k.k->p;
2423
2424                         if (pos.snapshot < iter->snapshot) {
2425                                 search_key = bpos_successor(k.k->p);
2426                                 continue;
2427                         }
2428
2429                         pos.snapshot = iter->snapshot;
2430
2431                         /*
2432                          * advance, same as on exit for iter->path, but only up
2433                          * to snapshot
2434                          */
2435                         __btree_path_get(iter->path, iter->flags & BTREE_ITER_INTENT);
2436                         iter->update_path = iter->path;
2437
2438                         iter->update_path = bch2_btree_path_set_pos(trans,
2439                                                 iter->update_path, pos,
2440                                                 iter->flags & BTREE_ITER_INTENT,
2441                                                 _THIS_IP_);
2442                 }
2443
2444                 /*
2445                  * We can never have a key in a leaf node at POS_MAX, so
2446                  * we don't have to check these successor() calls:
2447                  */
2448                 if ((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
2449                     !bch2_snapshot_is_ancestor(trans->c,
2450                                                iter->snapshot,
2451                                                k.k->p.snapshot)) {
2452                         search_key = bpos_successor(k.k->p);
2453                         continue;
2454                 }
2455
2456                 if (bkey_whiteout(k.k) &&
2457                     !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS)) {
2458                         search_key = bkey_successor(iter, k.k->p);
2459                         continue;
2460                 }
2461
2462                 break;
2463         }
2464
2465         iter->pos = iter_pos;
2466
2467         iter->path = bch2_btree_path_set_pos(trans, iter->path, k.k->p,
2468                                 iter->flags & BTREE_ITER_INTENT,
2469                                 btree_iter_ip_allocated(iter));
2470         BUG_ON(!iter->path->nodes_locked);
2471 out:
2472         if (iter->update_path) {
2473                 if (iter->update_path->uptodate &&
2474                     !bch2_btree_path_relock(trans, iter->update_path, _THIS_IP_)) {
2475                         k = bkey_s_c_err(-EINTR);
2476                 } else {
2477                         BUG_ON(!(iter->update_path->nodes_locked & 1));
2478                         iter->update_path->should_be_locked = true;
2479                 }
2480         }
2481         iter->path->should_be_locked = true;
2482
2483         if (!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS))
2484                 iter->pos.snapshot = iter->snapshot;
2485
2486         ret = bch2_btree_iter_verify_ret(iter, k);
2487         if (unlikely(ret)) {
2488                 bch2_btree_iter_set_pos(iter, iter->pos);
2489                 k = bkey_s_c_err(ret);
2490         }
2491
2492         bch2_btree_iter_verify_entry_exit(iter);
2493
2494         return k;
2495 }
2496
2497 /**
2498  * bch2_btree_iter_next: returns first key greater than iterator's current
2499  * position
2500  */
2501 struct bkey_s_c bch2_btree_iter_next(struct btree_iter *iter)
2502 {
2503         if (!bch2_btree_iter_advance(iter))
2504                 return bkey_s_c_null;
2505
2506         return bch2_btree_iter_peek(iter);
2507 }
2508
2509 /**
2510  * bch2_btree_iter_peek_prev: returns first key less than or equal to
2511  * iterator's current position
2512  */
2513 struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *iter)
2514 {
2515         struct btree_trans *trans = iter->trans;
2516         struct bpos search_key = iter->pos;
2517         struct btree_path *saved_path = NULL;
2518         struct bkey_s_c k;
2519         struct bkey saved_k;
2520         const struct bch_val *saved_v;
2521         int ret;
2522
2523         EBUG_ON(iter->path->cached || iter->path->level);
2524         EBUG_ON(iter->flags & BTREE_ITER_WITH_UPDATES);
2525
2526         if (iter->flags & BTREE_ITER_WITH_JOURNAL)
2527                 return bkey_s_c_err(-EIO);
2528
2529         bch2_btree_iter_verify(iter);
2530         bch2_btree_iter_verify_entry_exit(iter);
2531
2532         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2533                 search_key.snapshot = U32_MAX;
2534
2535         while (1) {
2536                 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2537                                                 iter->flags & BTREE_ITER_INTENT,
2538                                                 btree_iter_ip_allocated(iter));
2539
2540                 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2541                 if (unlikely(ret)) {
2542                         /* ensure that iter->k is consistent with iter->pos: */
2543                         bch2_btree_iter_set_pos(iter, iter->pos);
2544                         k = bkey_s_c_err(ret);
2545                         goto out;
2546                 }
2547
2548                 k = btree_path_level_peek(trans->c, iter->path,
2549                                           &iter->path->l[0], &iter->k);
2550                 if (!k.k ||
2551                     ((iter->flags & BTREE_ITER_IS_EXTENTS)
2552                      ? bpos_cmp(bkey_start_pos(k.k), search_key) >= 0
2553                      : bpos_cmp(k.k->p, search_key) > 0))
2554                         k = btree_path_level_prev(trans->c, iter->path,
2555                                                   &iter->path->l[0], &iter->k);
2556
2557                 bch2_btree_path_check_sort(trans, iter->path, 0);
2558
2559                 if (likely(k.k)) {
2560                         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) {
2561                                 if (k.k->p.snapshot == iter->snapshot)
2562                                         goto got_key;
2563
2564                                 /*
2565                                  * If we have a saved candidate, and we're no
2566                                  * longer at the same _key_ (not pos), return
2567                                  * that candidate
2568                                  */
2569                                 if (saved_path && bkey_cmp(k.k->p, saved_k.p)) {
2570                                         bch2_path_put(trans, iter->path,
2571                                                       iter->flags & BTREE_ITER_INTENT);
2572                                         iter->path = saved_path;
2573                                         saved_path = NULL;
2574                                         iter->k = saved_k;
2575                                         k.v     = saved_v;
2576                                         goto got_key;
2577                                 }
2578
2579                                 if (bch2_snapshot_is_ancestor(iter->trans->c,
2580                                                               iter->snapshot,
2581                                                               k.k->p.snapshot)) {
2582                                         if (saved_path)
2583                                                 bch2_path_put(trans, saved_path,
2584                                                       iter->flags & BTREE_ITER_INTENT);
2585                                         saved_path = btree_path_clone(trans, iter->path,
2586                                                                 iter->flags & BTREE_ITER_INTENT);
2587                                         saved_k = *k.k;
2588                                         saved_v = k.v;
2589                                 }
2590
2591                                 search_key = bpos_predecessor(k.k->p);
2592                                 continue;
2593                         }
2594 got_key:
2595                         if (bkey_whiteout(k.k) &&
2596                             !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS)) {
2597                                 search_key = bkey_predecessor(iter, k.k->p);
2598                                 if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2599                                         search_key.snapshot = U32_MAX;
2600                                 continue;
2601                         }
2602
2603                         break;
2604                 } else if (likely(bpos_cmp(iter->path->l[0].b->data->min_key, POS_MIN))) {
2605                         /* Advance to previous leaf node: */
2606                         search_key = bpos_predecessor(iter->path->l[0].b->data->min_key);
2607                 } else {
2608                         /* Start of btree: */
2609                         bch2_btree_iter_set_pos(iter, POS_MIN);
2610                         k = bkey_s_c_null;
2611                         goto out;
2612                 }
2613         }
2614
2615         EBUG_ON(bkey_cmp(bkey_start_pos(k.k), iter->pos) > 0);
2616
2617         /* Extents can straddle iter->pos: */
2618         if (bkey_cmp(k.k->p, iter->pos) < 0)
2619                 iter->pos = k.k->p;
2620
2621         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2622                 iter->pos.snapshot = iter->snapshot;
2623 out:
2624         if (saved_path)
2625                 bch2_path_put(trans, saved_path, iter->flags & BTREE_ITER_INTENT);
2626         iter->path->should_be_locked = true;
2627
2628         bch2_btree_iter_verify_entry_exit(iter);
2629         bch2_btree_iter_verify(iter);
2630
2631         return k;
2632 }
2633
2634 /**
2635  * bch2_btree_iter_prev: returns first key less than iterator's current
2636  * position
2637  */
2638 struct bkey_s_c bch2_btree_iter_prev(struct btree_iter *iter)
2639 {
2640         if (!bch2_btree_iter_rewind(iter))
2641                 return bkey_s_c_null;
2642
2643         return bch2_btree_iter_peek_prev(iter);
2644 }
2645
2646 struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *iter)
2647 {
2648         struct btree_trans *trans = iter->trans;
2649         struct bpos search_key;
2650         struct bkey_s_c k;
2651         int ret;
2652
2653         EBUG_ON(iter->path->level);
2654         bch2_btree_iter_verify(iter);
2655         bch2_btree_iter_verify_entry_exit(iter);
2656
2657         /* extents can't span inode numbers: */
2658         if ((iter->flags & BTREE_ITER_IS_EXTENTS) &&
2659             unlikely(iter->pos.offset == KEY_OFFSET_MAX)) {
2660                 if (iter->pos.inode == KEY_INODE_MAX)
2661                         return bkey_s_c_null;
2662
2663                 bch2_btree_iter_set_pos(iter, bpos_nosnap_successor(iter->pos));
2664         }
2665
2666         search_key = btree_iter_search_key(iter);
2667         iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2668                                         iter->flags & BTREE_ITER_INTENT,
2669                                         btree_iter_ip_allocated(iter));
2670
2671         ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2672         if (unlikely(ret))
2673                 return bkey_s_c_err(ret);
2674
2675         if ((iter->flags & BTREE_ITER_CACHED) ||
2676             !(iter->flags & (BTREE_ITER_IS_EXTENTS|BTREE_ITER_FILTER_SNAPSHOTS))) {
2677                 struct bkey_i *next_update;
2678
2679                 if ((iter->flags & BTREE_ITER_WITH_UPDATES) &&
2680                     (next_update = btree_trans_peek_updates(trans,
2681                                                 iter->btree_id, search_key)) &&
2682                     !bpos_cmp(next_update->k.p, iter->pos)) {
2683                         iter->k = next_update->k;
2684                         k = bkey_i_to_s_c(next_update);
2685                         goto out;
2686                 }
2687
2688                 if (unlikely(iter->flags & BTREE_ITER_WITH_JOURNAL) &&
2689                     (next_update = bch2_journal_keys_peek_slot(trans->c,
2690                                         iter->btree_id, 0, iter->pos))) {
2691                         iter->k = next_update->k;
2692                         k = bkey_i_to_s_c(next_update);
2693                         goto out;
2694                 }
2695
2696                 if (unlikely(iter->flags & BTREE_ITER_WITH_KEY_CACHE) &&
2697                     (k = btree_trans_peek_key_cache(iter, iter->pos)).k) {
2698                         if (!bkey_err(k))
2699                                 iter->k = *k.k;
2700                         goto out;
2701                 }
2702
2703                 k = bch2_btree_path_peek_slot(iter->path, &iter->k);
2704         } else {
2705                 struct bpos next;
2706
2707                 if (iter->flags & BTREE_ITER_INTENT) {
2708                         struct btree_iter iter2;
2709                         struct bpos end = iter->pos;
2710
2711                         if (iter->flags & BTREE_ITER_IS_EXTENTS)
2712                                 end.offset = U64_MAX;
2713
2714                         bch2_trans_copy_iter(&iter2, iter);
2715                         k = bch2_btree_iter_peek_upto(&iter2, end);
2716
2717                         if (k.k && !bkey_err(k)) {
2718                                 iter->k = iter2.k;
2719                                 k.k = &iter->k;
2720                         }
2721                         bch2_trans_iter_exit(trans, &iter2);
2722                 } else {
2723                         struct bpos pos = iter->pos;
2724
2725                         k = bch2_btree_iter_peek(iter);
2726                         iter->pos = pos;
2727                 }
2728
2729                 if (unlikely(bkey_err(k)))
2730                         return k;
2731
2732                 next = k.k ? bkey_start_pos(k.k) : POS_MAX;
2733
2734                 if (bkey_cmp(iter->pos, next) < 0) {
2735                         bkey_init(&iter->k);
2736                         iter->k.p = iter->pos;
2737
2738                         if (iter->flags & BTREE_ITER_IS_EXTENTS) {
2739                                 bch2_key_resize(&iter->k,
2740                                                 min_t(u64, KEY_SIZE_MAX,
2741                                                       (next.inode == iter->pos.inode
2742                                                        ? next.offset
2743                                                        : KEY_OFFSET_MAX) -
2744                                                       iter->pos.offset));
2745                                 EBUG_ON(!iter->k.size);
2746                         }
2747
2748                         k = (struct bkey_s_c) { &iter->k, NULL };
2749                 }
2750         }
2751 out:
2752         iter->path->should_be_locked = true;
2753
2754         bch2_btree_iter_verify_entry_exit(iter);
2755         bch2_btree_iter_verify(iter);
2756         ret = bch2_btree_iter_verify_ret(iter, k);
2757         if (unlikely(ret))
2758                 return bkey_s_c_err(ret);
2759
2760         return k;
2761 }
2762
2763 struct bkey_s_c bch2_btree_iter_next_slot(struct btree_iter *iter)
2764 {
2765         if (!bch2_btree_iter_advance(iter))
2766                 return bkey_s_c_null;
2767
2768         return bch2_btree_iter_peek_slot(iter);
2769 }
2770
2771 struct bkey_s_c bch2_btree_iter_prev_slot(struct btree_iter *iter)
2772 {
2773         if (!bch2_btree_iter_rewind(iter))
2774                 return bkey_s_c_null;
2775
2776         return bch2_btree_iter_peek_slot(iter);
2777 }
2778
2779 /* new transactional stuff: */
2780
2781 static inline void btree_path_verify_sorted_ref(struct btree_trans *trans,
2782                                                 struct btree_path *path)
2783 {
2784         EBUG_ON(path->sorted_idx >= trans->nr_sorted);
2785         EBUG_ON(trans->sorted[path->sorted_idx] != path->idx);
2786         EBUG_ON(!(trans->paths_allocated & (1ULL << path->idx)));
2787 }
2788
2789 static inline void btree_trans_verify_sorted_refs(struct btree_trans *trans)
2790 {
2791 #ifdef CONFIG_BCACHEFS_DEBUG
2792         unsigned i;
2793
2794         for (i = 0; i < trans->nr_sorted; i++)
2795                 btree_path_verify_sorted_ref(trans, trans->paths + trans->sorted[i]);
2796 #endif
2797 }
2798
2799 static void btree_trans_verify_sorted(struct btree_trans *trans)
2800 {
2801 #ifdef CONFIG_BCACHEFS_DEBUG
2802         struct btree_path *path, *prev = NULL;
2803         unsigned i;
2804
2805         trans_for_each_path_inorder(trans, path, i) {
2806                 if (prev && btree_path_cmp(prev, path) > 0) {
2807                         bch2_dump_trans_paths_updates(trans);
2808                         panic("trans paths out of order!\n");
2809                 }
2810                 prev = path;
2811         }
2812 #endif
2813 }
2814
2815 static inline void btree_path_swap(struct btree_trans *trans,
2816                                    struct btree_path *l, struct btree_path *r)
2817 {
2818         swap(l->sorted_idx, r->sorted_idx);
2819         swap(trans->sorted[l->sorted_idx],
2820              trans->sorted[r->sorted_idx]);
2821
2822         btree_path_verify_sorted_ref(trans, l);
2823         btree_path_verify_sorted_ref(trans, r);
2824 }
2825
2826 inline void bch2_btree_path_check_sort(struct btree_trans *trans, struct btree_path *path,
2827                                        int cmp)
2828 {
2829         struct btree_path *n;
2830
2831         if (cmp <= 0) {
2832                 n = prev_btree_path(trans, path);
2833                 if (n && btree_path_cmp(n, path) > 0) {
2834                         do {
2835                                 btree_path_swap(trans, n, path);
2836                                 n = prev_btree_path(trans, path);
2837                         } while (n && btree_path_cmp(n, path) > 0);
2838
2839                         goto out;
2840                 }
2841         }
2842
2843         if (cmp >= 0) {
2844                 n = next_btree_path(trans, path);
2845                 if (n && btree_path_cmp(path, n) > 0) {
2846                         do {
2847                                 btree_path_swap(trans, path, n);
2848                                 n = next_btree_path(trans, path);
2849                         } while (n && btree_path_cmp(path, n) > 0);
2850                 }
2851         }
2852 out:
2853         btree_trans_verify_sorted(trans);
2854 }
2855
2856 static inline void btree_path_list_remove(struct btree_trans *trans,
2857                                           struct btree_path *path)
2858 {
2859         unsigned i;
2860
2861         EBUG_ON(path->sorted_idx >= trans->nr_sorted);
2862
2863         array_remove_item(trans->sorted, trans->nr_sorted, path->sorted_idx);
2864
2865         for (i = path->sorted_idx; i < trans->nr_sorted; i++)
2866                 trans->paths[trans->sorted[i]].sorted_idx = i;
2867
2868         path->sorted_idx = U8_MAX;
2869
2870         btree_trans_verify_sorted_refs(trans);
2871 }
2872
2873 static inline void btree_path_list_add(struct btree_trans *trans,
2874                                        struct btree_path *pos,
2875                                        struct btree_path *path)
2876 {
2877         unsigned i;
2878
2879         btree_trans_verify_sorted_refs(trans);
2880
2881         path->sorted_idx = pos ? pos->sorted_idx + 1 : 0;
2882
2883         if (trans->in_traverse_all &&
2884             trans->traverse_all_idx != U8_MAX &&
2885             trans->traverse_all_idx >= path->sorted_idx)
2886                 trans->traverse_all_idx++;
2887
2888         array_insert_item(trans->sorted, trans->nr_sorted, path->sorted_idx, path->idx);
2889
2890         for (i = path->sorted_idx; i < trans->nr_sorted; i++)
2891                 trans->paths[trans->sorted[i]].sorted_idx = i;
2892
2893         btree_trans_verify_sorted_refs(trans);
2894 }
2895
2896 void bch2_trans_iter_exit(struct btree_trans *trans, struct btree_iter *iter)
2897 {
2898         if (iter->path)
2899                 bch2_path_put(trans, iter->path,
2900                               iter->flags & BTREE_ITER_INTENT);
2901         if (iter->update_path)
2902                 bch2_path_put(trans, iter->update_path,
2903                               iter->flags & BTREE_ITER_INTENT);
2904         if (iter->key_cache_path)
2905                 bch2_path_put(trans, iter->key_cache_path,
2906                               iter->flags & BTREE_ITER_INTENT);
2907         iter->path = NULL;
2908         iter->update_path = NULL;
2909         iter->key_cache_path = NULL;
2910 }
2911
2912 static void __bch2_trans_iter_init(struct btree_trans *trans,
2913                                    struct btree_iter *iter,
2914                                    unsigned btree_id, struct bpos pos,
2915                                    unsigned locks_want,
2916                                    unsigned depth,
2917                                    unsigned flags,
2918                                    unsigned long ip)
2919 {
2920         EBUG_ON(trans->restarted);
2921
2922         if (!(flags & (BTREE_ITER_ALL_SNAPSHOTS|BTREE_ITER_NOT_EXTENTS)) &&
2923             btree_node_type_is_extents(btree_id))
2924                 flags |= BTREE_ITER_IS_EXTENTS;
2925
2926         if (!(flags & __BTREE_ITER_ALL_SNAPSHOTS) &&
2927             !btree_type_has_snapshots(btree_id))
2928                 flags &= ~BTREE_ITER_ALL_SNAPSHOTS;
2929
2930         if (!(flags & BTREE_ITER_ALL_SNAPSHOTS) &&
2931             btree_type_has_snapshots(btree_id))
2932                 flags |= BTREE_ITER_FILTER_SNAPSHOTS;
2933
2934         if (!test_bit(JOURNAL_REPLAY_DONE, &trans->c->journal.flags))
2935                 flags |= BTREE_ITER_WITH_JOURNAL;
2936
2937         if (!btree_id_cached(trans->c, btree_id)) {
2938                 flags &= ~BTREE_ITER_CACHED;
2939                 flags &= ~BTREE_ITER_WITH_KEY_CACHE;
2940         } else if (!(flags & BTREE_ITER_CACHED))
2941                 flags |= BTREE_ITER_WITH_KEY_CACHE;
2942
2943         iter->trans     = trans;
2944         iter->path      = NULL;
2945         iter->update_path = NULL;
2946         iter->key_cache_path = NULL;
2947         iter->btree_id  = btree_id;
2948         iter->min_depth = depth;
2949         iter->flags     = flags;
2950         iter->snapshot  = pos.snapshot;
2951         iter->pos       = pos;
2952         iter->k.type    = KEY_TYPE_deleted;
2953         iter->k.p       = pos;
2954         iter->k.size    = 0;
2955 #ifdef CONFIG_BCACHEFS_DEBUG
2956         iter->ip_allocated = ip;
2957 #endif
2958
2959         iter->path = bch2_path_get(trans, btree_id, iter->pos,
2960                                    locks_want, depth, flags, ip);
2961 }
2962
2963 void bch2_trans_iter_init(struct btree_trans *trans,
2964                           struct btree_iter *iter,
2965                           unsigned btree_id, struct bpos pos,
2966                           unsigned flags)
2967 {
2968         __bch2_trans_iter_init(trans, iter, btree_id, pos,
2969                                0, 0, flags, _RET_IP_);
2970 }
2971
2972 void bch2_trans_node_iter_init(struct btree_trans *trans,
2973                                struct btree_iter *iter,
2974                                enum btree_id btree_id,
2975                                struct bpos pos,
2976                                unsigned locks_want,
2977                                unsigned depth,
2978                                unsigned flags)
2979 {
2980         __bch2_trans_iter_init(trans, iter, btree_id, pos, locks_want, depth,
2981                                BTREE_ITER_NOT_EXTENTS|
2982                                __BTREE_ITER_ALL_SNAPSHOTS|
2983                                BTREE_ITER_ALL_SNAPSHOTS|
2984                                flags, _RET_IP_);
2985         BUG_ON(iter->path->locks_want    < min(locks_want, BTREE_MAX_DEPTH));
2986         BUG_ON(iter->path->level        != depth);
2987         BUG_ON(iter->min_depth          != depth);
2988 }
2989
2990 void bch2_trans_copy_iter(struct btree_iter *dst, struct btree_iter *src)
2991 {
2992         *dst = *src;
2993         if (src->path)
2994                 __btree_path_get(src->path, src->flags & BTREE_ITER_INTENT);
2995         if (src->update_path)
2996                 __btree_path_get(src->update_path, src->flags & BTREE_ITER_INTENT);
2997         dst->key_cache_path = NULL;
2998 }
2999
3000 void *bch2_trans_kmalloc(struct btree_trans *trans, size_t size)
3001 {
3002         size_t new_top = trans->mem_top + size;
3003         void *p;
3004
3005         if (new_top > trans->mem_bytes) {
3006                 size_t old_bytes = trans->mem_bytes;
3007                 size_t new_bytes = roundup_pow_of_two(new_top);
3008                 void *new_mem;
3009
3010                 WARN_ON_ONCE(new_bytes > BTREE_TRANS_MEM_MAX);
3011
3012                 new_mem = krealloc(trans->mem, new_bytes, GFP_NOFS);
3013                 if (!new_mem && new_bytes <= BTREE_TRANS_MEM_MAX) {
3014                         new_mem = mempool_alloc(&trans->c->btree_trans_mem_pool, GFP_KERNEL);
3015                         new_bytes = BTREE_TRANS_MEM_MAX;
3016                         kfree(trans->mem);
3017                 }
3018
3019                 if (!new_mem)
3020                         return ERR_PTR(-ENOMEM);
3021
3022                 trans->mem = new_mem;
3023                 trans->mem_bytes = new_bytes;
3024
3025                 if (old_bytes) {
3026                         trace_trans_restart_mem_realloced(trans->fn, _RET_IP_, new_bytes);
3027                         btree_trans_restart(trans);
3028                         return ERR_PTR(-EINTR);
3029                 }
3030         }
3031
3032         p = trans->mem + trans->mem_top;
3033         trans->mem_top += size;
3034         memset(p, 0, size);
3035         return p;
3036 }
3037
3038 /**
3039  * bch2_trans_begin() - reset a transaction after a interrupted attempt
3040  * @trans: transaction to reset
3041  *
3042  * While iterating over nodes or updating nodes a attempt to lock a btree
3043  * node may return EINTR when the trylock fails. When this occurs
3044  * bch2_trans_begin() should be called and the transaction retried.
3045  */
3046 void bch2_trans_begin(struct btree_trans *trans)
3047 {
3048         struct btree_insert_entry *i;
3049         struct btree_path *path;
3050
3051         trans_for_each_update(trans, i)
3052                 __btree_path_put(i->path, true);
3053
3054         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
3055         trans->extra_journal_res        = 0;
3056         trans->nr_updates               = 0;
3057         trans->mem_top                  = 0;
3058
3059         trans->hooks                    = NULL;
3060         trans->extra_journal_entries.nr = 0;
3061
3062         if (trans->fs_usage_deltas) {
3063                 trans->fs_usage_deltas->used = 0;
3064                 memset(&trans->fs_usage_deltas->memset_start, 0,
3065                        (void *) &trans->fs_usage_deltas->memset_end -
3066                        (void *) &trans->fs_usage_deltas->memset_start);
3067         }
3068
3069         trans_for_each_path(trans, path) {
3070                 path->should_be_locked = false;
3071
3072                 /*
3073                  * If the transaction wasn't restarted, we're presuming to be
3074                  * doing something new: dont keep iterators excpt the ones that
3075                  * are in use - except for the subvolumes btree:
3076                  */
3077                 if (!trans->restarted && path->btree_id != BTREE_ID_subvolumes)
3078                         path->preserve = false;
3079
3080                 /*
3081                  * XXX: we probably shouldn't be doing this if the transaction
3082                  * was restarted, but currently we still overflow transaction
3083                  * iterators if we do that
3084                  */
3085                 if (!path->ref && !path->preserve)
3086                         __bch2_path_free(trans, path);
3087                 else
3088                         path->preserve = false;
3089         }
3090
3091         bch2_trans_cond_resched(trans);
3092
3093         if (trans->restarted)
3094                 bch2_btree_path_traverse_all(trans);
3095
3096         trans->restarted = false;
3097 }
3098
3099 static void bch2_trans_alloc_paths(struct btree_trans *trans, struct bch_fs *c)
3100 {
3101         size_t paths_bytes      = sizeof(struct btree_path) * BTREE_ITER_MAX;
3102         size_t updates_bytes    = sizeof(struct btree_insert_entry) * BTREE_ITER_MAX;
3103         void *p = NULL;
3104
3105         BUG_ON(trans->used_mempool);
3106
3107 #ifdef __KERNEL__
3108         p = this_cpu_xchg(c->btree_paths_bufs->path , NULL);
3109 #endif
3110         if (!p)
3111                 p = mempool_alloc(&trans->c->btree_paths_pool, GFP_NOFS);
3112
3113         trans->paths            = p; p += paths_bytes;
3114         trans->updates          = p; p += updates_bytes;
3115 }
3116
3117 void __bch2_trans_init(struct btree_trans *trans, struct bch_fs *c,
3118                        unsigned expected_nr_iters,
3119                        size_t expected_mem_bytes,
3120                        const char *fn)
3121         __acquires(&c->btree_trans_barrier)
3122 {
3123         BUG_ON(lock_class_is_held(&bch2_btree_node_lock_key));
3124
3125         memset(trans, 0, sizeof(*trans));
3126         trans->c                = c;
3127         trans->fn               = fn;
3128
3129         bch2_trans_alloc_paths(trans, c);
3130
3131         if (expected_mem_bytes) {
3132                 trans->mem_bytes = roundup_pow_of_two(expected_mem_bytes);
3133                 trans->mem = kmalloc(trans->mem_bytes, GFP_KERNEL|__GFP_NOFAIL);
3134
3135                 if (!unlikely(trans->mem)) {
3136                         trans->mem = mempool_alloc(&c->btree_trans_mem_pool, GFP_KERNEL);
3137                         trans->mem_bytes = BTREE_TRANS_MEM_MAX;
3138                 }
3139         }
3140
3141         trans->srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
3142
3143         trans->pid = current->pid;
3144         mutex_lock(&c->btree_trans_lock);
3145         list_add(&trans->list, &c->btree_trans_list);
3146         mutex_unlock(&c->btree_trans_lock);
3147 }
3148
3149 static void check_btree_paths_leaked(struct btree_trans *trans)
3150 {
3151 #ifdef CONFIG_BCACHEFS_DEBUG
3152         struct bch_fs *c = trans->c;
3153         struct btree_path *path;
3154
3155         trans_for_each_path(trans, path)
3156                 if (path->ref)
3157                         goto leaked;
3158         return;
3159 leaked:
3160         bch_err(c, "btree paths leaked from %s!", trans->fn);
3161         trans_for_each_path(trans, path)
3162                 if (path->ref)
3163                         printk(KERN_ERR "  btree %s %pS\n",
3164                                bch2_btree_ids[path->btree_id],
3165                                (void *) path->ip_allocated);
3166         /* Be noisy about this: */
3167         bch2_fatal_error(c);
3168 #endif
3169 }
3170
3171 void bch2_trans_exit(struct btree_trans *trans)
3172         __releases(&c->btree_trans_barrier)
3173 {
3174         struct btree_insert_entry *i;
3175         struct bch_fs *c = trans->c;
3176
3177         bch2_trans_unlock(trans);
3178
3179         trans_for_each_update(trans, i)
3180                 __btree_path_put(i->path, true);
3181         trans->nr_updates               = 0;
3182
3183         check_btree_paths_leaked(trans);
3184
3185         mutex_lock(&c->btree_trans_lock);
3186         list_del(&trans->list);
3187         mutex_unlock(&c->btree_trans_lock);
3188
3189         srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx);
3190
3191         bch2_journal_preres_put(&c->journal, &trans->journal_preres);
3192
3193         kfree(trans->extra_journal_entries.data);
3194
3195         if (trans->fs_usage_deltas) {
3196                 if (trans->fs_usage_deltas->size + sizeof(trans->fs_usage_deltas) ==
3197                     REPLICAS_DELTA_LIST_MAX)
3198                         mempool_free(trans->fs_usage_deltas,
3199                                      &c->replicas_delta_pool);
3200                 else
3201                         kfree(trans->fs_usage_deltas);
3202         }
3203
3204         if (trans->mem_bytes == BTREE_TRANS_MEM_MAX)
3205                 mempool_free(trans->mem, &c->btree_trans_mem_pool);
3206         else
3207                 kfree(trans->mem);
3208
3209 #ifdef __KERNEL__
3210         /*
3211          * Userspace doesn't have a real percpu implementation:
3212          */
3213         trans->paths = this_cpu_xchg(c->btree_paths_bufs->path, trans->paths);
3214 #endif
3215
3216         if (trans->paths)
3217                 mempool_free(trans->paths, &c->btree_paths_pool);
3218
3219         trans->mem      = (void *) 0x1;
3220         trans->paths    = (void *) 0x1;
3221 }
3222
3223 static void __maybe_unused
3224 bch2_btree_path_node_to_text(struct printbuf *out,
3225                              struct btree_bkey_cached_common *_b,
3226                              bool cached)
3227 {
3228         pr_buf(out, "    l=%u %s:",
3229                _b->level, bch2_btree_ids[_b->btree_id]);
3230         bch2_bpos_to_text(out, btree_node_pos(_b, cached));
3231 }
3232
3233 static bool trans_has_locks(struct btree_trans *trans)
3234 {
3235         struct btree_path *path;
3236
3237         trans_for_each_path(trans, path)
3238                 if (path->nodes_locked)
3239                         return true;
3240         return false;
3241 }
3242
3243 void bch2_btree_trans_to_text(struct printbuf *out, struct bch_fs *c)
3244 {
3245         struct btree_trans *trans;
3246         struct btree_path *path;
3247         struct btree *b;
3248         static char lock_types[] = { 'r', 'i', 'w' };
3249         unsigned l;
3250
3251         mutex_lock(&c->btree_trans_lock);
3252         list_for_each_entry(trans, &c->btree_trans_list, list) {
3253                 if (!trans_has_locks(trans))
3254                         continue;
3255
3256                 pr_buf(out, "%i %s\n", trans->pid, trans->fn);
3257
3258                 trans_for_each_path(trans, path) {
3259                         if (!path->nodes_locked)
3260                                 continue;
3261
3262                         pr_buf(out, "  path %u %c l=%u %s:",
3263                                path->idx,
3264                                path->cached ? 'c' : 'b',
3265                                path->level,
3266                                bch2_btree_ids[path->btree_id]);
3267                         bch2_bpos_to_text(out, path->pos);
3268                         pr_buf(out, "\n");
3269
3270                         for (l = 0; l < BTREE_MAX_DEPTH; l++) {
3271                                 if (btree_node_locked(path, l)) {
3272                                         pr_buf(out, "    %s l=%u ",
3273                                                btree_node_intent_locked(path, l) ? "i" : "r", l);
3274                                         bch2_btree_path_node_to_text(out,
3275                                                         (void *) path->l[l].b,
3276                                                         path->cached);
3277                                         pr_buf(out, "\n");
3278                                 }
3279                         }
3280                 }
3281
3282                 b = READ_ONCE(trans->locking);
3283                 if (b) {
3284                         path = &trans->paths[trans->locking_path_idx];
3285                         pr_buf(out, "  locking path %u %c l=%u %c %s:",
3286                                trans->locking_path_idx,
3287                                path->cached ? 'c' : 'b',
3288                                trans->locking_level,
3289                                lock_types[trans->locking_lock_type],
3290                                bch2_btree_ids[trans->locking_btree_id]);
3291                         bch2_bpos_to_text(out, trans->locking_pos);
3292
3293                         pr_buf(out, " node ");
3294                         bch2_btree_path_node_to_text(out,
3295                                         (void *) b, path->cached);
3296                         pr_buf(out, "\n");
3297                 }
3298         }
3299         mutex_unlock(&c->btree_trans_lock);
3300 }
3301
3302 void bch2_fs_btree_iter_exit(struct bch_fs *c)
3303 {
3304         if (c->btree_trans_barrier_initialized)
3305                 cleanup_srcu_struct(&c->btree_trans_barrier);
3306         mempool_exit(&c->btree_trans_mem_pool);
3307         mempool_exit(&c->btree_paths_pool);
3308 }
3309
3310 int bch2_fs_btree_iter_init(struct bch_fs *c)
3311 {
3312         unsigned nr = BTREE_ITER_MAX;
3313         int ret;
3314
3315         INIT_LIST_HEAD(&c->btree_trans_list);
3316         mutex_init(&c->btree_trans_lock);
3317
3318         ret   = mempool_init_kmalloc_pool(&c->btree_paths_pool, 1,
3319                         sizeof(struct btree_path) * nr +
3320                         sizeof(struct btree_insert_entry) * nr) ?:
3321                 mempool_init_kmalloc_pool(&c->btree_trans_mem_pool, 1,
3322                                           BTREE_TRANS_MEM_MAX) ?:
3323                 init_srcu_struct(&c->btree_trans_barrier);
3324         if (!ret)
3325                 c->btree_trans_barrier_initialized = true;
3326         return ret;
3327 }