1 // SPDX-License-Identifier: GPL-2.0
4 #include "bkey_methods.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"
17 #include "subvolume.h"
19 #include <linux/prandom.h>
20 #include <linux/prefetch.h>
21 #include <trace/events/bcachefs.h>
23 static inline void btree_path_list_remove(struct btree_trans *, struct btree_path *);
24 static inline void btree_path_list_add(struct btree_trans *, struct btree_path *,
27 static inline unsigned long btree_iter_ip_allocated(struct btree_iter *iter)
29 #ifdef TRACK_PATH_ALLOCATED
30 return iter->ip_allocated;
36 static struct btree_path *btree_path_alloc(struct btree_trans *, struct btree_path *);
39 * Unlocks before scheduling
40 * Note: does not revalidate iterator
42 static inline int bch2_trans_cond_resched(struct btree_trans *trans)
44 if (need_resched() || race_fault()) {
45 bch2_trans_unlock(trans);
47 return bch2_trans_relock(trans);
53 static inline int __btree_path_cmp(const struct btree_path *l,
54 enum btree_id r_btree_id,
60 * Must match lock ordering as defined by __bch2_btree_node_lock:
62 return cmp_int(l->btree_id, r_btree_id) ?:
63 cmp_int((int) l->cached, (int) r_cached) ?:
64 bpos_cmp(l->pos, r_pos) ?:
65 -cmp_int(l->level, r_level);
68 static inline int btree_path_cmp(const struct btree_path *l,
69 const struct btree_path *r)
71 return __btree_path_cmp(l, r->btree_id, r->cached, r->pos, r->level);
74 static inline struct bpos bkey_successor(struct btree_iter *iter, struct bpos p)
76 /* Are we iterating over keys in all snapshots? */
77 if (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) {
78 p = bpos_successor(p);
80 p = bpos_nosnap_successor(p);
81 p.snapshot = iter->snapshot;
87 static inline struct bpos bkey_predecessor(struct btree_iter *iter, struct bpos p)
89 /* Are we iterating over keys in all snapshots? */
90 if (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) {
91 p = bpos_predecessor(p);
93 p = bpos_nosnap_predecessor(p);
94 p.snapshot = iter->snapshot;
100 static inline struct bpos btree_iter_search_key(struct btree_iter *iter)
102 struct bpos pos = iter->pos;
104 if ((iter->flags & BTREE_ITER_IS_EXTENTS) &&
105 !bkey_eq(pos, POS_MAX))
106 pos = bkey_successor(iter, pos);
110 static inline bool btree_path_pos_before_node(struct btree_path *path,
113 return bpos_lt(path->pos, b->data->min_key);
116 static inline bool btree_path_pos_after_node(struct btree_path *path,
119 return bpos_gt(path->pos, b->key.k.p);
122 static inline bool btree_path_pos_in_node(struct btree_path *path,
125 return path->btree_id == b->c.btree_id &&
126 !btree_path_pos_before_node(path, b) &&
127 !btree_path_pos_after_node(path, b);
130 /* Btree iterator: */
132 #ifdef CONFIG_BCACHEFS_DEBUG
134 static void bch2_btree_path_verify_cached(struct btree_trans *trans,
135 struct btree_path *path)
137 struct bkey_cached *ck;
138 bool locked = btree_node_locked(path, 0);
140 if (!bch2_btree_node_relock(trans, path, 0))
143 ck = (void *) path->l[0].b;
144 BUG_ON(ck->key.btree_id != path->btree_id ||
145 !bkey_eq(ck->key.pos, path->pos));
148 btree_node_unlock(trans, path, 0);
151 static void bch2_btree_path_verify_level(struct btree_trans *trans,
152 struct btree_path *path, unsigned level)
154 struct btree_path_level *l;
155 struct btree_node_iter tmp;
157 struct bkey_packed *p, *k;
158 struct printbuf buf1 = PRINTBUF;
159 struct printbuf buf2 = PRINTBUF;
160 struct printbuf buf3 = PRINTBUF;
163 if (!bch2_debug_check_iterators)
168 locked = btree_node_locked(path, level);
172 bch2_btree_path_verify_cached(trans, path);
176 if (!btree_path_node(path, level))
179 if (!bch2_btree_node_relock_notrace(trans, path, level))
182 BUG_ON(!btree_path_pos_in_node(path, l->b));
184 bch2_btree_node_iter_verify(&l->iter, l->b);
187 * For interior nodes, the iterator will have skipped past deleted keys:
190 ? bch2_btree_node_iter_prev(&tmp, l->b)
191 : bch2_btree_node_iter_prev_all(&tmp, l->b);
192 k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
194 if (p && bkey_iter_pos_cmp(l->b, p, &path->pos) >= 0) {
199 if (k && bkey_iter_pos_cmp(l->b, k, &path->pos) < 0) {
205 btree_node_unlock(trans, path, level);
208 bch2_bpos_to_text(&buf1, path->pos);
211 struct bkey uk = bkey_unpack_key(l->b, p);
213 bch2_bkey_to_text(&buf2, &uk);
215 prt_printf(&buf2, "(none)");
219 struct bkey uk = bkey_unpack_key(l->b, k);
221 bch2_bkey_to_text(&buf3, &uk);
223 prt_printf(&buf3, "(none)");
226 panic("path should be %s key at level %u:\n"
230 msg, level, buf1.buf, buf2.buf, buf3.buf);
233 static void bch2_btree_path_verify(struct btree_trans *trans,
234 struct btree_path *path)
236 struct bch_fs *c = trans->c;
239 EBUG_ON(path->btree_id >= BTREE_ID_NR);
241 for (i = 0; i < (!path->cached ? BTREE_MAX_DEPTH : 1); i++) {
243 BUG_ON(!path->cached &&
244 c->btree_roots[path->btree_id].b->c.level > i);
248 bch2_btree_path_verify_level(trans, path, i);
251 bch2_btree_path_verify_locks(path);
254 void bch2_trans_verify_paths(struct btree_trans *trans)
256 struct btree_path *path;
258 trans_for_each_path(trans, path)
259 bch2_btree_path_verify(trans, path);
262 static void bch2_btree_iter_verify(struct btree_iter *iter)
264 struct btree_trans *trans = iter->trans;
266 BUG_ON(iter->btree_id >= BTREE_ID_NR);
268 BUG_ON(!!(iter->flags & BTREE_ITER_CACHED) != iter->path->cached);
270 BUG_ON((iter->flags & BTREE_ITER_IS_EXTENTS) &&
271 (iter->flags & BTREE_ITER_ALL_SNAPSHOTS));
273 BUG_ON(!(iter->flags & __BTREE_ITER_ALL_SNAPSHOTS) &&
274 (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) &&
275 !btree_type_has_snapshots(iter->btree_id));
277 if (iter->update_path)
278 bch2_btree_path_verify(trans, iter->update_path);
279 bch2_btree_path_verify(trans, iter->path);
282 static void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter)
284 BUG_ON((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
285 !iter->pos.snapshot);
287 BUG_ON(!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS) &&
288 iter->pos.snapshot != iter->snapshot);
290 BUG_ON(bkey_lt(iter->pos, bkey_start_pos(&iter->k)) ||
291 bkey_gt(iter->pos, iter->k.p));
294 static int bch2_btree_iter_verify_ret(struct btree_iter *iter, struct bkey_s_c k)
296 struct btree_trans *trans = iter->trans;
297 struct btree_iter copy;
298 struct bkey_s_c prev;
301 if (!bch2_debug_check_iterators)
304 if (!(iter->flags & BTREE_ITER_FILTER_SNAPSHOTS))
307 if (bkey_err(k) || !k.k)
310 BUG_ON(!bch2_snapshot_is_ancestor(trans->c,
314 bch2_trans_iter_init(trans, ©, iter->btree_id, iter->pos,
315 BTREE_ITER_NOPRESERVE|
316 BTREE_ITER_ALL_SNAPSHOTS);
317 prev = bch2_btree_iter_prev(©);
321 ret = bkey_err(prev);
325 if (bkey_eq(prev.k->p, k.k->p) &&
326 bch2_snapshot_is_ancestor(trans->c, iter->snapshot,
327 prev.k->p.snapshot) > 0) {
328 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
330 bch2_bkey_to_text(&buf1, k.k);
331 bch2_bkey_to_text(&buf2, prev.k);
333 panic("iter snap %u\n"
340 bch2_trans_iter_exit(trans, ©);
344 void bch2_assert_pos_locked(struct btree_trans *trans, enum btree_id id,
345 struct bpos pos, bool key_cache)
347 struct btree_path *path;
349 struct printbuf buf = PRINTBUF;
351 btree_trans_sort_paths(trans);
353 trans_for_each_path_inorder(trans, path, idx) {
354 int cmp = cmp_int(path->btree_id, id) ?:
355 cmp_int(path->cached, key_cache);
362 if (!btree_node_locked(path, 0) ||
363 !path->should_be_locked)
367 if (bkey_ge(pos, path->l[0].b->data->min_key) &&
368 bkey_le(pos, path->l[0].b->key.k.p))
371 if (bkey_eq(pos, path->pos))
376 bch2_dump_trans_paths_updates(trans);
377 bch2_bpos_to_text(&buf, pos);
379 panic("not locked: %s %s%s\n",
380 bch2_btree_ids[id], buf.buf,
381 key_cache ? " cached" : "");
386 static inline void bch2_btree_path_verify_level(struct btree_trans *trans,
387 struct btree_path *path, unsigned l) {}
388 static inline void bch2_btree_path_verify(struct btree_trans *trans,
389 struct btree_path *path) {}
390 static inline void bch2_btree_iter_verify(struct btree_iter *iter) {}
391 static inline void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter) {}
392 static inline int bch2_btree_iter_verify_ret(struct btree_iter *iter, struct bkey_s_c k) { return 0; }
396 /* Btree path: fixups after btree updates */
398 static void btree_node_iter_set_set_pos(struct btree_node_iter *iter,
401 struct bkey_packed *k)
403 struct btree_node_iter_set *set;
405 btree_node_iter_for_each(iter, set)
406 if (set->end == t->end_offset) {
407 set->k = __btree_node_key_to_offset(b, k);
408 bch2_btree_node_iter_sort(iter, b);
412 bch2_btree_node_iter_push(iter, b, k, btree_bkey_last(b, t));
415 static void __bch2_btree_path_fix_key_modified(struct btree_path *path,
417 struct bkey_packed *where)
419 struct btree_path_level *l = &path->l[b->c.level];
421 if (where != bch2_btree_node_iter_peek_all(&l->iter, l->b))
424 if (bkey_iter_pos_cmp(l->b, where, &path->pos) < 0)
425 bch2_btree_node_iter_advance(&l->iter, l->b);
428 void bch2_btree_path_fix_key_modified(struct btree_trans *trans,
430 struct bkey_packed *where)
432 struct btree_path *path;
434 trans_for_each_path_with_node(trans, b, path) {
435 __bch2_btree_path_fix_key_modified(path, b, where);
436 bch2_btree_path_verify_level(trans, path, b->c.level);
440 static void __bch2_btree_node_iter_fix(struct btree_path *path,
442 struct btree_node_iter *node_iter,
444 struct bkey_packed *where,
445 unsigned clobber_u64s,
448 const struct bkey_packed *end = btree_bkey_last(b, t);
449 struct btree_node_iter_set *set;
450 unsigned offset = __btree_node_key_to_offset(b, where);
451 int shift = new_u64s - clobber_u64s;
452 unsigned old_end = t->end_offset - shift;
453 unsigned orig_iter_pos = node_iter->data[0].k;
454 bool iter_current_key_modified =
455 orig_iter_pos >= offset &&
456 orig_iter_pos <= offset + clobber_u64s;
458 btree_node_iter_for_each(node_iter, set)
459 if (set->end == old_end)
462 /* didn't find the bset in the iterator - might have to readd it: */
464 bkey_iter_pos_cmp(b, where, &path->pos) >= 0) {
465 bch2_btree_node_iter_push(node_iter, b, where, end);
468 /* Iterator is after key that changed */
472 set->end = t->end_offset;
474 /* Iterator hasn't gotten to the key that changed yet: */
479 bkey_iter_pos_cmp(b, where, &path->pos) >= 0) {
481 } else if (set->k < offset + clobber_u64s) {
482 set->k = offset + new_u64s;
483 if (set->k == set->end)
484 bch2_btree_node_iter_set_drop(node_iter, set);
486 /* Iterator is after key that changed */
487 set->k = (int) set->k + shift;
491 bch2_btree_node_iter_sort(node_iter, b);
493 if (node_iter->data[0].k != orig_iter_pos)
494 iter_current_key_modified = true;
497 * When a new key is added, and the node iterator now points to that
498 * key, the iterator might have skipped past deleted keys that should
499 * come after the key the iterator now points to. We have to rewind to
500 * before those deleted keys - otherwise
501 * bch2_btree_node_iter_prev_all() breaks:
503 if (!bch2_btree_node_iter_end(node_iter) &&
504 iter_current_key_modified &&
507 struct bkey_packed *k, *k2, *p;
509 k = bch2_btree_node_iter_peek_all(node_iter, b);
511 for_each_bset(b, t) {
512 bool set_pos = false;
514 if (node_iter->data[0].end == t->end_offset)
517 k2 = bch2_btree_node_iter_bset_pos(node_iter, b, t);
519 while ((p = bch2_bkey_prev_all(b, t, k2)) &&
520 bkey_iter_cmp(b, k, p) < 0) {
526 btree_node_iter_set_set_pos(node_iter,
532 void bch2_btree_node_iter_fix(struct btree_trans *trans,
533 struct btree_path *path,
535 struct btree_node_iter *node_iter,
536 struct bkey_packed *where,
537 unsigned clobber_u64s,
540 struct bset_tree *t = bch2_bkey_to_bset_inlined(b, where);
541 struct btree_path *linked;
543 if (node_iter != &path->l[b->c.level].iter) {
544 __bch2_btree_node_iter_fix(path, b, node_iter, t,
545 where, clobber_u64s, new_u64s);
547 if (bch2_debug_check_iterators)
548 bch2_btree_node_iter_verify(node_iter, b);
551 trans_for_each_path_with_node(trans, b, linked) {
552 __bch2_btree_node_iter_fix(linked, b,
553 &linked->l[b->c.level].iter, t,
554 where, clobber_u64s, new_u64s);
555 bch2_btree_path_verify_level(trans, linked, b->c.level);
559 /* Btree path level: pointer to a particular btree node and node iter */
561 static inline struct bkey_s_c __btree_iter_unpack(struct bch_fs *c,
562 struct btree_path_level *l,
564 struct bkey_packed *k)
568 * signal to bch2_btree_iter_peek_slot() that we're currently at
571 u->type = KEY_TYPE_deleted;
572 return bkey_s_c_null;
575 return bkey_disassemble(l->b, k, u);
578 static inline struct bkey_s_c btree_path_level_peek_all(struct bch_fs *c,
579 struct btree_path_level *l,
582 return __btree_iter_unpack(c, l, u,
583 bch2_btree_node_iter_peek_all(&l->iter, l->b));
586 static inline struct bkey_s_c btree_path_level_peek(struct btree_trans *trans,
587 struct btree_path *path,
588 struct btree_path_level *l,
591 struct bkey_s_c k = __btree_iter_unpack(trans->c, l, u,
592 bch2_btree_node_iter_peek(&l->iter, l->b));
594 path->pos = k.k ? k.k->p : l->b->key.k.p;
595 trans->paths_sorted = false;
596 bch2_btree_path_verify_level(trans, path, l - path->l);
600 static inline struct bkey_s_c btree_path_level_prev(struct btree_trans *trans,
601 struct btree_path *path,
602 struct btree_path_level *l,
605 struct bkey_s_c k = __btree_iter_unpack(trans->c, l, u,
606 bch2_btree_node_iter_prev(&l->iter, l->b));
608 path->pos = k.k ? k.k->p : l->b->data->min_key;
609 trans->paths_sorted = false;
610 bch2_btree_path_verify_level(trans, path, l - path->l);
614 static inline bool btree_path_advance_to_pos(struct btree_path *path,
615 struct btree_path_level *l,
618 struct bkey_packed *k;
621 while ((k = bch2_btree_node_iter_peek_all(&l->iter, l->b)) &&
622 bkey_iter_pos_cmp(l->b, k, &path->pos) < 0) {
623 if (max_advance > 0 && nr_advanced >= max_advance)
626 bch2_btree_node_iter_advance(&l->iter, l->b);
633 static inline void __btree_path_level_init(struct btree_path *path,
636 struct btree_path_level *l = &path->l[level];
638 bch2_btree_node_iter_init(&l->iter, l->b, &path->pos);
641 * Iterators to interior nodes should always be pointed at the first non
645 bch2_btree_node_iter_peek(&l->iter, l->b);
648 void bch2_btree_path_level_init(struct btree_trans *trans,
649 struct btree_path *path,
652 BUG_ON(path->cached);
654 EBUG_ON(!btree_path_pos_in_node(path, b));
655 EBUG_ON(b->c.lock.state.seq & 1);
657 path->l[b->c.level].lock_seq = b->c.lock.state.seq;
658 path->l[b->c.level].b = b;
659 __btree_path_level_init(path, b->c.level);
662 /* Btree path: fixups after btree node updates: */
664 static void bch2_trans_revalidate_updates_in_node(struct btree_trans *trans, struct btree *b)
666 struct bch_fs *c = trans->c;
667 struct btree_insert_entry *i;
669 trans_for_each_update(trans, i)
671 i->level == b->c.level &&
672 i->btree_id == b->c.btree_id &&
673 bpos_cmp(i->k->k.p, b->data->min_key) >= 0 &&
674 bpos_cmp(i->k->k.p, b->data->max_key) <= 0) {
675 i->old_v = bch2_btree_path_peek_slot(i->path, &i->old_k).v;
677 if (unlikely(trans->journal_replay_not_finished)) {
679 bch2_journal_keys_peek_slot(c, i->btree_id, i->level,
691 * A btree node is being replaced - update the iterator to point to the new
694 void bch2_trans_node_add(struct btree_trans *trans, struct btree *b)
696 struct btree_path *path;
698 trans_for_each_path(trans, path)
699 if (path->uptodate == BTREE_ITER_UPTODATE &&
701 btree_path_pos_in_node(path, b)) {
702 enum btree_node_locked_type t =
703 btree_lock_want(path, b->c.level);
705 if (t != BTREE_NODE_UNLOCKED) {
706 btree_node_unlock(trans, path, b->c.level);
707 six_lock_increment(&b->c.lock, t);
708 mark_btree_node_locked(trans, path, b->c.level, t);
711 bch2_btree_path_level_init(trans, path, b);
714 bch2_trans_revalidate_updates_in_node(trans, b);
718 * A btree node has been modified in such a way as to invalidate iterators - fix
721 void bch2_trans_node_reinit_iter(struct btree_trans *trans, struct btree *b)
723 struct btree_path *path;
725 trans_for_each_path_with_node(trans, b, path)
726 __btree_path_level_init(path, b->c.level);
728 bch2_trans_revalidate_updates_in_node(trans, b);
731 /* Btree path: traverse, set_pos: */
733 static inline int btree_path_lock_root(struct btree_trans *trans,
734 struct btree_path *path,
736 unsigned long trace_ip)
738 struct bch_fs *c = trans->c;
739 struct btree *b, **rootp = &c->btree_roots[path->btree_id].b;
740 enum six_lock_type lock_type;
744 EBUG_ON(path->nodes_locked);
747 b = READ_ONCE(*rootp);
748 path->level = READ_ONCE(b->c.level);
750 if (unlikely(path->level < depth_want)) {
752 * the root is at a lower depth than the depth we want:
753 * got to the end of the btree, or we're walking nodes
754 * greater than some depth and there are no nodes >=
757 path->level = depth_want;
758 for (i = path->level; i < BTREE_MAX_DEPTH; i++)
763 lock_type = __btree_lock_want(path, path->level);
764 ret = btree_node_lock(trans, path, &b->c,
765 path->level, lock_type, trace_ip);
767 if (bch2_err_matches(ret, BCH_ERR_lock_fail_root_changed))
769 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
774 if (likely(b == READ_ONCE(*rootp) &&
775 b->c.level == path->level &&
777 for (i = 0; i < path->level; i++)
778 path->l[i].b = ERR_PTR(-BCH_ERR_no_btree_node_lock_root);
779 path->l[path->level].b = b;
780 for (i = path->level + 1; i < BTREE_MAX_DEPTH; i++)
783 mark_btree_node_locked(trans, path, path->level, lock_type);
784 bch2_btree_path_level_init(trans, path, b);
788 six_unlock_type(&b->c.lock, lock_type);
793 static int btree_path_prefetch(struct btree_trans *trans, struct btree_path *path)
795 struct bch_fs *c = trans->c;
796 struct btree_path_level *l = path_l(path);
797 struct btree_node_iter node_iter = l->iter;
798 struct bkey_packed *k;
800 unsigned nr = test_bit(BCH_FS_STARTED, &c->flags)
801 ? (path->level > 1 ? 0 : 2)
802 : (path->level > 1 ? 1 : 16);
803 bool was_locked = btree_node_locked(path, path->level);
806 bch2_bkey_buf_init(&tmp);
808 while (nr-- && !ret) {
809 if (!bch2_btree_node_relock(trans, path, path->level))
812 bch2_btree_node_iter_advance(&node_iter, l->b);
813 k = bch2_btree_node_iter_peek(&node_iter, l->b);
817 bch2_bkey_buf_unpack(&tmp, c, l->b, k);
818 ret = bch2_btree_node_prefetch(c, trans, path, tmp.k, path->btree_id,
823 btree_node_unlock(trans, path, path->level);
825 bch2_bkey_buf_exit(&tmp, c);
829 static int btree_path_prefetch_j(struct btree_trans *trans, struct btree_path *path,
830 struct btree_and_journal_iter *jiter)
832 struct bch_fs *c = trans->c;
835 unsigned nr = test_bit(BCH_FS_STARTED, &c->flags)
836 ? (path->level > 1 ? 0 : 2)
837 : (path->level > 1 ? 1 : 16);
838 bool was_locked = btree_node_locked(path, path->level);
841 bch2_bkey_buf_init(&tmp);
843 while (nr-- && !ret) {
844 if (!bch2_btree_node_relock(trans, path, path->level))
847 bch2_btree_and_journal_iter_advance(jiter);
848 k = bch2_btree_and_journal_iter_peek(jiter);
852 bch2_bkey_buf_reassemble(&tmp, c, k);
853 ret = bch2_btree_node_prefetch(c, trans, path, tmp.k, path->btree_id,
858 btree_node_unlock(trans, path, path->level);
860 bch2_bkey_buf_exit(&tmp, c);
864 static noinline void btree_node_mem_ptr_set(struct btree_trans *trans,
865 struct btree_path *path,
866 unsigned plevel, struct btree *b)
868 struct btree_path_level *l = &path->l[plevel];
869 bool locked = btree_node_locked(path, plevel);
870 struct bkey_packed *k;
871 struct bch_btree_ptr_v2 *bp;
873 if (!bch2_btree_node_relock(trans, path, plevel))
876 k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
877 BUG_ON(k->type != KEY_TYPE_btree_ptr_v2);
879 bp = (void *) bkeyp_val(&l->b->format, k);
880 bp->mem_ptr = (unsigned long)b;
883 btree_node_unlock(trans, path, plevel);
886 static noinline int btree_node_iter_and_journal_peek(struct btree_trans *trans,
887 struct btree_path *path,
889 struct bkey_buf *out)
891 struct bch_fs *c = trans->c;
892 struct btree_path_level *l = path_l(path);
893 struct btree_and_journal_iter jiter;
897 __bch2_btree_and_journal_iter_init_node_iter(&jiter, c, l->b, l->iter, path->pos);
899 k = bch2_btree_and_journal_iter_peek(&jiter);
901 bch2_bkey_buf_reassemble(out, c, k);
903 if (flags & BTREE_ITER_PREFETCH)
904 ret = btree_path_prefetch_j(trans, path, &jiter);
906 bch2_btree_and_journal_iter_exit(&jiter);
910 static __always_inline int btree_path_down(struct btree_trans *trans,
911 struct btree_path *path,
913 unsigned long trace_ip)
915 struct bch_fs *c = trans->c;
916 struct btree_path_level *l = path_l(path);
918 unsigned level = path->level - 1;
919 enum six_lock_type lock_type = __btree_lock_want(path, level);
923 EBUG_ON(!btree_node_locked(path, path->level));
925 bch2_bkey_buf_init(&tmp);
927 if (unlikely(trans->journal_replay_not_finished)) {
928 ret = btree_node_iter_and_journal_peek(trans, path, flags, &tmp);
932 bch2_bkey_buf_unpack(&tmp, c, l->b,
933 bch2_btree_node_iter_peek(&l->iter, l->b));
935 if (flags & BTREE_ITER_PREFETCH) {
936 ret = btree_path_prefetch(trans, path);
942 b = bch2_btree_node_get(trans, path, tmp.k, level, lock_type, trace_ip);
943 ret = PTR_ERR_OR_ZERO(b);
947 if (likely(!trans->journal_replay_not_finished &&
948 tmp.k->k.type == KEY_TYPE_btree_ptr_v2) &&
949 unlikely(b != btree_node_mem_ptr(tmp.k)))
950 btree_node_mem_ptr_set(trans, path, level + 1, b);
952 if (btree_node_read_locked(path, level + 1))
953 btree_node_unlock(trans, path, level + 1);
955 mark_btree_node_locked(trans, path, level, lock_type);
957 bch2_btree_path_level_init(trans, path, b);
959 bch2_btree_path_verify_locks(path);
961 bch2_bkey_buf_exit(&tmp, c);
966 static int bch2_btree_path_traverse_all(struct btree_trans *trans)
968 struct bch_fs *c = trans->c;
969 struct btree_path *path;
970 unsigned long trace_ip = _RET_IP_;
973 if (trans->in_traverse_all)
974 return -BCH_ERR_transaction_restart_in_traverse_all;
976 trans->in_traverse_all = true;
978 trans->restarted = 0;
980 trans_for_each_path(trans, path)
981 path->should_be_locked = false;
983 btree_trans_sort_paths(trans);
985 bch2_trans_unlock(trans);
988 if (unlikely(trans->memory_allocation_failure)) {
991 closure_init_stack(&cl);
994 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
999 /* Now, redo traversals in correct order: */
1001 while (i < trans->nr_sorted) {
1002 path = trans->paths + trans->sorted[i];
1005 * Traversing a path can cause another path to be added at about
1006 * the same position:
1008 if (path->uptodate) {
1009 __btree_path_get(path, false);
1010 ret = bch2_btree_path_traverse_one(trans, path, 0, _THIS_IP_);
1011 __btree_path_put(path, false);
1013 if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
1024 * We used to assert that all paths had been traversed here
1025 * (path->uptodate < BTREE_ITER_NEED_TRAVERSE); however, since
1026 * path->Should_be_locked is not set yet, we we might have unlocked and
1027 * then failed to relock a path - that's fine.
1030 bch2_btree_cache_cannibalize_unlock(c);
1032 trans->in_traverse_all = false;
1034 trace_and_count(c, trans_traverse_all, trans, trace_ip);
1038 static inline bool btree_path_check_pos_in_node(struct btree_path *path,
1039 unsigned l, int check_pos)
1041 if (check_pos < 0 && btree_path_pos_before_node(path, path->l[l].b))
1043 if (check_pos > 0 && btree_path_pos_after_node(path, path->l[l].b))
1048 static inline bool btree_path_good_node(struct btree_trans *trans,
1049 struct btree_path *path,
1050 unsigned l, int check_pos)
1052 return is_btree_node(path, l) &&
1053 bch2_btree_node_relock(trans, path, l) &&
1054 btree_path_check_pos_in_node(path, l, check_pos);
1057 static void btree_path_set_level_down(struct btree_trans *trans,
1058 struct btree_path *path,
1063 path->level = new_level;
1065 for (l = path->level + 1; l < BTREE_MAX_DEPTH; l++)
1066 if (btree_lock_want(path, l) == BTREE_NODE_UNLOCKED)
1067 btree_node_unlock(trans, path, l);
1069 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1070 bch2_btree_path_verify(trans, path);
1073 static noinline unsigned __btree_path_up_until_good_node(struct btree_trans *trans,
1074 struct btree_path *path,
1077 unsigned i, l = path->level;
1079 while (btree_path_node(path, l) &&
1080 !btree_path_good_node(trans, path, l, check_pos))
1081 __btree_path_set_level_up(trans, path, l++);
1083 /* If we need intent locks, take them too: */
1085 i < path->locks_want && btree_path_node(path, i);
1087 if (!bch2_btree_node_relock(trans, path, i)) {
1089 __btree_path_set_level_up(trans, path, l++);
1096 static inline unsigned btree_path_up_until_good_node(struct btree_trans *trans,
1097 struct btree_path *path,
1100 return likely(btree_node_locked(path, path->level) &&
1101 btree_path_check_pos_in_node(path, path->level, check_pos))
1103 : __btree_path_up_until_good_node(trans, path, check_pos);
1107 * This is the main state machine for walking down the btree - walks down to a
1110 * Returns 0 on success, -EIO on error (error reading in a btree node).
1112 * On error, caller (peek_node()/peek_key()) must return NULL; the error is
1113 * stashed in the iterator and returned from bch2_trans_exit().
1115 int bch2_btree_path_traverse_one(struct btree_trans *trans,
1116 struct btree_path *path,
1118 unsigned long trace_ip)
1120 unsigned depth_want = path->level;
1121 int ret = -((int) trans->restarted);
1127 * Ensure we obey path->should_be_locked: if it's set, we can't unlock
1128 * and re-traverse the path without a transaction restart:
1130 if (path->should_be_locked) {
1131 ret = bch2_btree_path_relock(trans, path, trace_ip);
1136 ret = bch2_btree_path_traverse_cached(trans, path, flags);
1140 if (unlikely(path->level >= BTREE_MAX_DEPTH))
1143 path->level = btree_path_up_until_good_node(trans, path, 0);
1145 EBUG_ON(btree_path_node(path, path->level) &&
1146 !btree_node_locked(path, path->level));
1149 * Note: path->nodes[path->level] may be temporarily NULL here - that
1150 * would indicate to other code that we got to the end of the btree,
1151 * here it indicates that relocking the root failed - it's critical that
1152 * btree_path_lock_root() comes next and that it can't fail
1154 while (path->level > depth_want) {
1155 ret = btree_path_node(path, path->level)
1156 ? btree_path_down(trans, path, flags, trace_ip)
1157 : btree_path_lock_root(trans, path, depth_want, trace_ip);
1158 if (unlikely(ret)) {
1161 * No nodes at this level - got to the end of
1168 __bch2_btree_path_unlock(trans, path);
1169 path->level = depth_want;
1170 path->l[path->level].b = ERR_PTR(ret);
1175 path->uptodate = BTREE_ITER_UPTODATE;
1177 if (bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted) {
1178 struct printbuf buf = PRINTBUF;
1180 prt_printf(&buf, "ret %s (%i) trans->restarted %s (%i)\n",
1181 bch2_err_str(ret), ret,
1182 bch2_err_str(trans->restarted), trans->restarted);
1183 #ifdef CONFIG_BCACHEFS_DEBUG
1184 bch2_prt_backtrace(&buf, &trans->last_restarted);
1186 panic("%s", buf.buf);
1188 bch2_btree_path_verify(trans, path);
1192 static inline void btree_path_copy(struct btree_trans *trans, struct btree_path *dst,
1193 struct btree_path *src)
1195 unsigned i, offset = offsetof(struct btree_path, pos);
1197 memcpy((void *) dst + offset,
1198 (void *) src + offset,
1199 sizeof(struct btree_path) - offset);
1201 for (i = 0; i < BTREE_MAX_DEPTH; i++) {
1202 unsigned t = btree_node_locked_type(dst, i);
1204 if (t != BTREE_NODE_UNLOCKED)
1205 six_lock_increment(&dst->l[i].b->c.lock, t);
1209 static struct btree_path *btree_path_clone(struct btree_trans *trans, struct btree_path *src,
1212 struct btree_path *new = btree_path_alloc(trans, src);
1214 btree_path_copy(trans, new, src);
1215 __btree_path_get(new, intent);
1220 struct btree_path *__bch2_btree_path_make_mut(struct btree_trans *trans,
1221 struct btree_path *path, bool intent,
1224 __btree_path_put(path, intent);
1225 path = btree_path_clone(trans, path, intent);
1226 path->preserve = false;
1230 struct btree_path * __must_check
1231 __bch2_btree_path_set_pos(struct btree_trans *trans,
1232 struct btree_path *path, struct bpos new_pos,
1233 bool intent, unsigned long ip, int cmp)
1235 unsigned level = path->level;
1237 bch2_trans_verify_not_in_restart(trans);
1238 EBUG_ON(!path->ref);
1240 path = bch2_btree_path_make_mut(trans, path, intent, ip);
1242 path->pos = new_pos;
1243 trans->paths_sorted = false;
1245 if (unlikely(path->cached)) {
1246 btree_node_unlock(trans, path, 0);
1247 path->l[0].b = ERR_PTR(-BCH_ERR_no_btree_node_up);
1248 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1252 level = btree_path_up_until_good_node(trans, path, cmp);
1254 if (btree_path_node(path, level)) {
1255 struct btree_path_level *l = &path->l[level];
1257 BUG_ON(!btree_node_locked(path, level));
1259 * We might have to skip over many keys, or just a few: try
1260 * advancing the node iterator, and if we have to skip over too
1261 * many keys just reinit it (or if we're rewinding, since that
1265 !btree_path_advance_to_pos(path, l, 8))
1266 bch2_btree_node_iter_init(&l->iter, l->b, &path->pos);
1269 * Iterators to interior nodes should always be pointed at the first non
1272 if (unlikely(level))
1273 bch2_btree_node_iter_peek(&l->iter, l->b);
1276 if (unlikely(level != path->level)) {
1277 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1278 __bch2_btree_path_unlock(trans, path);
1281 bch2_btree_path_verify(trans, path);
1285 /* Btree path: main interface: */
1287 static struct btree_path *have_path_at_pos(struct btree_trans *trans, struct btree_path *path)
1289 struct btree_path *sib;
1291 sib = prev_btree_path(trans, path);
1292 if (sib && !btree_path_cmp(sib, path))
1295 sib = next_btree_path(trans, path);
1296 if (sib && !btree_path_cmp(sib, path))
1302 static struct btree_path *have_node_at_pos(struct btree_trans *trans, struct btree_path *path)
1304 struct btree_path *sib;
1306 sib = prev_btree_path(trans, path);
1307 if (sib && sib->level == path->level && path_l(sib)->b == path_l(path)->b)
1310 sib = next_btree_path(trans, path);
1311 if (sib && sib->level == path->level && path_l(sib)->b == path_l(path)->b)
1317 static inline void __bch2_path_free(struct btree_trans *trans, struct btree_path *path)
1319 __bch2_btree_path_unlock(trans, path);
1320 btree_path_list_remove(trans, path);
1321 trans->paths_allocated &= ~(1ULL << path->idx);
1324 void bch2_path_put(struct btree_trans *trans, struct btree_path *path, bool intent)
1326 struct btree_path *dup;
1328 EBUG_ON(trans->paths + path->idx != path);
1329 EBUG_ON(!path->ref);
1331 if (!__btree_path_put(path, intent))
1334 dup = path->preserve
1335 ? have_path_at_pos(trans, path)
1336 : have_node_at_pos(trans, path);
1338 if (!dup && !(!path->preserve && !is_btree_node(path, path->level)))
1341 if (path->should_be_locked &&
1342 !trans->restarted &&
1343 (!dup || !bch2_btree_path_relock_norestart(trans, dup, _THIS_IP_)))
1347 dup->preserve |= path->preserve;
1348 dup->should_be_locked |= path->should_be_locked;
1351 __bch2_path_free(trans, path);
1354 static void bch2_path_put_nokeep(struct btree_trans *trans, struct btree_path *path,
1357 EBUG_ON(trans->paths + path->idx != path);
1358 EBUG_ON(!path->ref);
1360 if (!__btree_path_put(path, intent))
1363 __bch2_path_free(trans, path);
1366 void bch2_trans_restart_error(struct btree_trans *trans, u32 restart_count)
1368 panic("trans->restart_count %u, should be %u, last restarted by %pS\n",
1369 trans->restart_count, restart_count,
1370 (void *) trans->last_begin_ip);
1373 void bch2_trans_in_restart_error(struct btree_trans *trans)
1375 panic("in transaction restart: %s, last restarted by %pS\n",
1376 bch2_err_str(trans->restarted),
1377 (void *) trans->last_begin_ip);
1381 void bch2_trans_updates_to_text(struct printbuf *buf, struct btree_trans *trans)
1383 struct btree_insert_entry *i;
1384 struct btree_write_buffered_key *wb;
1386 prt_printf(buf, "transaction updates for %s journal seq %llu",
1387 trans->fn, trans->journal_res.seq);
1389 printbuf_indent_add(buf, 2);
1391 trans_for_each_update(trans, i) {
1392 struct bkey_s_c old = { &i->old_k, i->old_v };
1394 prt_printf(buf, "update: btree=%s cached=%u %pS",
1395 bch2_btree_ids[i->btree_id],
1397 (void *) i->ip_allocated);
1400 prt_printf(buf, " old ");
1401 bch2_bkey_val_to_text(buf, trans->c, old);
1404 prt_printf(buf, " new ");
1405 bch2_bkey_val_to_text(buf, trans->c, bkey_i_to_s_c(i->k));
1409 trans_for_each_wb_update(trans, wb) {
1410 prt_printf(buf, "update: btree=%s wb=1 %pS",
1411 bch2_btree_ids[wb->btree],
1412 (void *) i->ip_allocated);
1415 prt_printf(buf, " new ");
1416 bch2_bkey_val_to_text(buf, trans->c, bkey_i_to_s_c(&wb->k));
1420 printbuf_indent_sub(buf, 2);
1424 void bch2_dump_trans_updates(struct btree_trans *trans)
1426 struct printbuf buf = PRINTBUF;
1428 bch2_trans_updates_to_text(&buf, trans);
1429 bch2_print_string_as_lines(KERN_ERR, buf.buf);
1430 printbuf_exit(&buf);
1434 void bch2_btree_path_to_text(struct printbuf *out, struct btree_path *path)
1436 prt_printf(out, "path: idx %2u ref %u:%u %c %c btree=%s l=%u pos ",
1437 path->idx, path->ref, path->intent_ref,
1438 path->preserve ? 'P' : ' ',
1439 path->should_be_locked ? 'S' : ' ',
1440 bch2_btree_ids[path->btree_id],
1442 bch2_bpos_to_text(out, path->pos);
1444 prt_printf(out, " locks %u", path->nodes_locked);
1445 #ifdef TRACK_PATH_ALLOCATED
1446 prt_printf(out, " %pS", (void *) path->ip_allocated);
1452 void __bch2_trans_paths_to_text(struct printbuf *out, struct btree_trans *trans,
1455 struct btree_path *path;
1459 btree_trans_sort_paths(trans);
1461 trans_for_each_path_inorder(trans, path, idx)
1462 bch2_btree_path_to_text(out, path);
1466 void bch2_trans_paths_to_text(struct printbuf *out, struct btree_trans *trans)
1468 __bch2_trans_paths_to_text(out, trans, false);
1472 void __bch2_dump_trans_paths_updates(struct btree_trans *trans, bool nosort)
1474 struct printbuf buf = PRINTBUF;
1476 __bch2_trans_paths_to_text(&buf, trans, nosort);
1477 bch2_trans_updates_to_text(&buf, trans);
1479 bch2_print_string_as_lines(KERN_ERR, buf.buf);
1480 printbuf_exit(&buf);
1484 void bch2_dump_trans_paths_updates(struct btree_trans *trans)
1486 __bch2_dump_trans_paths_updates(trans, false);
1490 static void bch2_trans_update_max_paths(struct btree_trans *trans)
1492 struct btree_transaction_stats *s = btree_trans_stats(trans);
1493 struct printbuf buf = PRINTBUF;
1498 bch2_trans_paths_to_text(&buf, trans);
1500 if (!buf.allocation_failure) {
1501 mutex_lock(&s->lock);
1502 if (s->nr_max_paths < hweight64(trans->paths_allocated)) {
1503 s->nr_max_paths = trans->nr_max_paths =
1504 hweight64(trans->paths_allocated);
1505 swap(s->max_paths_text, buf.buf);
1507 mutex_unlock(&s->lock);
1510 printbuf_exit(&buf);
1512 trans->nr_max_paths = hweight64(trans->paths_allocated);
1515 static noinline void btree_path_overflow(struct btree_trans *trans)
1517 bch2_dump_trans_paths_updates(trans);
1518 panic("trans path oveflow\n");
1521 static inline struct btree_path *btree_path_alloc(struct btree_trans *trans,
1522 struct btree_path *pos)
1524 struct btree_path *path;
1527 if (unlikely(trans->paths_allocated ==
1528 ~((~0ULL << 1) << (BTREE_ITER_MAX - 1))))
1529 btree_path_overflow(trans);
1531 idx = __ffs64(~trans->paths_allocated);
1534 * Do this before marking the new path as allocated, since it won't be
1537 if (unlikely(idx > trans->nr_max_paths))
1538 bch2_trans_update_max_paths(trans);
1540 trans->paths_allocated |= 1ULL << idx;
1542 path = &trans->paths[idx];
1545 path->intent_ref = 0;
1546 path->nodes_locked = 0;
1548 btree_path_list_add(trans, pos, path);
1549 trans->paths_sorted = false;
1553 struct btree_path *bch2_path_get(struct btree_trans *trans,
1554 enum btree_id btree_id, struct bpos pos,
1555 unsigned locks_want, unsigned level,
1556 unsigned flags, unsigned long ip)
1558 struct btree_path *path, *path_pos = NULL;
1559 bool cached = flags & BTREE_ITER_CACHED;
1560 bool intent = flags & BTREE_ITER_INTENT;
1563 bch2_trans_verify_not_in_restart(trans);
1564 bch2_trans_verify_locks(trans);
1566 btree_trans_sort_paths(trans);
1568 trans_for_each_path_inorder(trans, path, i) {
1569 if (__btree_path_cmp(path,
1580 path_pos->cached == cached &&
1581 path_pos->btree_id == btree_id &&
1582 path_pos->level == level) {
1583 __btree_path_get(path_pos, intent);
1584 path = bch2_btree_path_set_pos(trans, path_pos, pos, intent, ip);
1586 path = btree_path_alloc(trans, path_pos);
1589 __btree_path_get(path, intent);
1591 path->btree_id = btree_id;
1592 path->cached = cached;
1593 path->uptodate = BTREE_ITER_NEED_TRAVERSE;
1594 path->should_be_locked = false;
1595 path->level = level;
1596 path->locks_want = locks_want;
1597 path->nodes_locked = 0;
1598 for (i = 0; i < ARRAY_SIZE(path->l); i++)
1599 path->l[i].b = ERR_PTR(-BCH_ERR_no_btree_node_init);
1600 #ifdef TRACK_PATH_ALLOCATED
1601 path->ip_allocated = ip;
1603 trans->paths_sorted = false;
1606 if (!(flags & BTREE_ITER_NOPRESERVE))
1607 path->preserve = true;
1609 if (path->intent_ref)
1610 locks_want = max(locks_want, level + 1);
1613 * If the path has locks_want greater than requested, we don't downgrade
1614 * it here - on transaction restart because btree node split needs to
1615 * upgrade locks, we might be putting/getting the iterator again.
1616 * Downgrading iterators only happens via bch2_trans_downgrade(), after
1617 * a successful transaction commit.
1620 locks_want = min(locks_want, BTREE_MAX_DEPTH);
1621 if (locks_want > path->locks_want)
1622 bch2_btree_path_upgrade_noupgrade_sibs(trans, path, locks_want);
1627 struct bkey_s_c bch2_btree_path_peek_slot(struct btree_path *path, struct bkey *u)
1630 struct btree_path_level *l = path_l(path);
1631 struct bkey_packed *_k;
1634 if (unlikely(!l->b))
1635 return bkey_s_c_null;
1637 EBUG_ON(path->uptodate != BTREE_ITER_UPTODATE);
1638 EBUG_ON(!btree_node_locked(path, path->level));
1640 if (!path->cached) {
1641 _k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
1642 k = _k ? bkey_disassemble(l->b, _k, u) : bkey_s_c_null;
1644 EBUG_ON(k.k && bkey_deleted(k.k) && bpos_eq(k.k->p, path->pos));
1646 if (!k.k || !bpos_eq(path->pos, k.k->p))
1649 struct bkey_cached *ck = (void *) path->l[0].b;
1652 (path->btree_id != ck->key.btree_id ||
1653 !bkey_eq(path->pos, ck->key.pos)));
1654 if (!ck || !ck->valid)
1655 return bkey_s_c_null;
1658 k = bkey_i_to_s_c(ck->k);
1665 return (struct bkey_s_c) { u, NULL };
1668 /* Btree iterators: */
1671 __bch2_btree_iter_traverse(struct btree_iter *iter)
1673 return bch2_btree_path_traverse(iter->trans, iter->path, iter->flags);
1677 bch2_btree_iter_traverse(struct btree_iter *iter)
1681 iter->path = bch2_btree_path_set_pos(iter->trans, iter->path,
1682 btree_iter_search_key(iter),
1683 iter->flags & BTREE_ITER_INTENT,
1684 btree_iter_ip_allocated(iter));
1686 ret = bch2_btree_path_traverse(iter->trans, iter->path, iter->flags);
1690 btree_path_set_should_be_locked(iter->path);
1694 /* Iterate across nodes (leaf and interior nodes) */
1696 struct btree *bch2_btree_iter_peek_node(struct btree_iter *iter)
1698 struct btree_trans *trans = iter->trans;
1699 struct btree *b = NULL;
1702 EBUG_ON(iter->path->cached);
1703 bch2_btree_iter_verify(iter);
1705 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
1709 b = btree_path_node(iter->path, iter->path->level);
1713 BUG_ON(bpos_lt(b->key.k.p, iter->pos));
1715 bkey_init(&iter->k);
1716 iter->k.p = iter->pos = b->key.k.p;
1718 iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p,
1719 iter->flags & BTREE_ITER_INTENT,
1720 btree_iter_ip_allocated(iter));
1721 btree_path_set_should_be_locked(iter->path);
1723 bch2_btree_iter_verify_entry_exit(iter);
1724 bch2_btree_iter_verify(iter);
1732 struct btree *bch2_btree_iter_next_node(struct btree_iter *iter)
1734 struct btree_trans *trans = iter->trans;
1735 struct btree_path *path = iter->path;
1736 struct btree *b = NULL;
1739 bch2_trans_verify_not_in_restart(trans);
1740 EBUG_ON(iter->path->cached);
1741 bch2_btree_iter_verify(iter);
1743 /* already at end? */
1744 if (!btree_path_node(path, path->level))
1748 if (!btree_path_node(path, path->level + 1)) {
1749 btree_path_set_level_up(trans, path);
1753 if (!bch2_btree_node_relock(trans, path, path->level + 1)) {
1754 __bch2_btree_path_unlock(trans, path);
1755 path->l[path->level].b = ERR_PTR(-BCH_ERR_no_btree_node_relock);
1756 path->l[path->level + 1].b = ERR_PTR(-BCH_ERR_no_btree_node_relock);
1757 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1758 trace_and_count(trans->c, trans_restart_relock_next_node, trans, _THIS_IP_, path);
1759 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_relock);
1763 b = btree_path_node(path, path->level + 1);
1765 if (bpos_eq(iter->pos, b->key.k.p)) {
1766 __btree_path_set_level_up(trans, path, path->level++);
1769 * Haven't gotten to the end of the parent node: go back down to
1770 * the next child node
1773 bch2_btree_path_set_pos(trans, path, bpos_successor(iter->pos),
1774 iter->flags & BTREE_ITER_INTENT,
1775 btree_iter_ip_allocated(iter));
1777 btree_path_set_level_down(trans, path, iter->min_depth);
1779 ret = bch2_btree_path_traverse(trans, path, iter->flags);
1783 b = path->l[path->level].b;
1786 bkey_init(&iter->k);
1787 iter->k.p = iter->pos = b->key.k.p;
1789 iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p,
1790 iter->flags & BTREE_ITER_INTENT,
1791 btree_iter_ip_allocated(iter));
1792 btree_path_set_should_be_locked(iter->path);
1793 BUG_ON(iter->path->uptodate);
1795 bch2_btree_iter_verify_entry_exit(iter);
1796 bch2_btree_iter_verify(iter);
1804 /* Iterate across keys (in leaf nodes only) */
1806 inline bool bch2_btree_iter_advance(struct btree_iter *iter)
1808 if (likely(!(iter->flags & BTREE_ITER_ALL_LEVELS))) {
1809 struct bpos pos = iter->k.p;
1810 bool ret = !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS
1811 ? bpos_eq(pos, SPOS_MAX)
1812 : bkey_eq(pos, SPOS_MAX));
1814 if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
1815 pos = bkey_successor(iter, pos);
1816 bch2_btree_iter_set_pos(iter, pos);
1819 if (!btree_path_node(iter->path, iter->path->level))
1822 iter->advanced = true;
1827 inline bool bch2_btree_iter_rewind(struct btree_iter *iter)
1829 struct bpos pos = bkey_start_pos(&iter->k);
1830 bool ret = !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS
1831 ? bpos_eq(pos, POS_MIN)
1832 : bkey_eq(pos, POS_MIN));
1834 if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
1835 pos = bkey_predecessor(iter, pos);
1836 bch2_btree_iter_set_pos(iter, pos);
1841 struct bkey_i *__bch2_btree_trans_peek_updates(struct btree_iter *iter)
1843 struct btree_insert_entry *i;
1844 struct bkey_i *ret = NULL;
1846 trans_for_each_update(iter->trans, i) {
1847 if (i->btree_id < iter->btree_id)
1849 if (i->btree_id > iter->btree_id)
1851 if (bpos_lt(i->k->k.p, iter->path->pos))
1853 if (i->key_cache_already_flushed)
1855 if (!ret || bpos_lt(i->k->k.p, ret->k.p))
1862 static inline struct bkey_i *btree_trans_peek_updates(struct btree_iter *iter)
1864 return iter->flags & BTREE_ITER_WITH_UPDATES
1865 ? __bch2_btree_trans_peek_updates(iter)
1869 struct bkey_i *bch2_btree_journal_peek(struct btree_trans *trans,
1870 struct btree_iter *iter,
1871 struct bpos end_pos)
1875 if (bpos_lt(iter->path->pos, iter->journal_pos))
1876 iter->journal_idx = 0;
1878 k = bch2_journal_keys_peek_upto(trans->c, iter->btree_id,
1882 &iter->journal_idx);
1884 iter->journal_pos = k ? k->k.p : end_pos;
1889 struct bkey_s_c btree_trans_peek_slot_journal(struct btree_trans *trans,
1890 struct btree_iter *iter)
1892 struct bkey_i *k = bch2_btree_journal_peek(trans, iter, iter->path->pos);
1896 return bkey_i_to_s_c(k);
1898 return bkey_s_c_null;
1903 struct bkey_s_c btree_trans_peek_journal(struct btree_trans *trans,
1904 struct btree_iter *iter,
1907 struct bkey_i *next_journal =
1908 bch2_btree_journal_peek(trans, iter,
1909 k.k ? k.k->p : path_l(iter->path)->b->key.k.p);
1912 iter->k = next_journal->k;
1913 k = bkey_i_to_s_c(next_journal);
1920 * Checks btree key cache for key at iter->pos and returns it if present, or
1924 struct bkey_s_c btree_trans_peek_key_cache(struct btree_iter *iter, struct bpos pos)
1926 struct btree_trans *trans = iter->trans;
1927 struct bch_fs *c = trans->c;
1932 if ((iter->flags & BTREE_ITER_KEY_CACHE_FILL) &&
1933 bpos_eq(iter->pos, pos))
1934 return bkey_s_c_null;
1936 if (!bch2_btree_key_cache_find(c, iter->btree_id, pos))
1937 return bkey_s_c_null;
1939 if (!iter->key_cache_path)
1940 iter->key_cache_path = bch2_path_get(trans, iter->btree_id, pos,
1941 iter->flags & BTREE_ITER_INTENT, 0,
1942 iter->flags|BTREE_ITER_CACHED|
1943 BTREE_ITER_CACHED_NOFILL,
1946 iter->key_cache_path = bch2_btree_path_set_pos(trans, iter->key_cache_path, pos,
1947 iter->flags & BTREE_ITER_INTENT,
1948 btree_iter_ip_allocated(iter));
1950 ret = bch2_btree_path_traverse(trans, iter->key_cache_path,
1951 iter->flags|BTREE_ITER_CACHED) ?:
1952 bch2_btree_path_relock(trans, iter->path, _THIS_IP_);
1954 return bkey_s_c_err(ret);
1956 btree_path_set_should_be_locked(iter->key_cache_path);
1958 k = bch2_btree_path_peek_slot(iter->key_cache_path, &u);
1959 if (k.k && !bkey_err(k)) {
1966 static struct bkey_s_c __bch2_btree_iter_peek(struct btree_iter *iter, struct bpos search_key)
1968 struct btree_trans *trans = iter->trans;
1969 struct bkey_i *next_update;
1970 struct bkey_s_c k, k2;
1973 EBUG_ON(iter->path->cached);
1974 bch2_btree_iter_verify(iter);
1977 struct btree_path_level *l;
1979 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
1980 iter->flags & BTREE_ITER_INTENT,
1981 btree_iter_ip_allocated(iter));
1983 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
1984 if (unlikely(ret)) {
1985 /* ensure that iter->k is consistent with iter->pos: */
1986 bch2_btree_iter_set_pos(iter, iter->pos);
1987 k = bkey_s_c_err(ret);
1991 l = path_l(iter->path);
1993 if (unlikely(!l->b)) {
1994 /* No btree nodes at requested level: */
1995 bch2_btree_iter_set_pos(iter, SPOS_MAX);
2000 btree_path_set_should_be_locked(iter->path);
2002 k = btree_path_level_peek_all(trans->c, l, &iter->k);
2004 if (unlikely(iter->flags & BTREE_ITER_WITH_KEY_CACHE) &&
2006 (k2 = btree_trans_peek_key_cache(iter, k.k->p)).k) {
2010 bch2_btree_iter_set_pos(iter, iter->pos);
2015 if (unlikely(iter->flags & BTREE_ITER_WITH_JOURNAL))
2016 k = btree_trans_peek_journal(trans, iter, k);
2018 next_update = btree_trans_peek_updates(iter);
2021 bpos_le(next_update->k.p,
2022 k.k ? k.k->p : l->b->key.k.p)) {
2023 iter->k = next_update->k;
2024 k = bkey_i_to_s_c(next_update);
2027 if (k.k && bkey_deleted(k.k)) {
2029 * If we've got a whiteout, and it's after the search
2030 * key, advance the search key to the whiteout instead
2031 * of just after the whiteout - it might be a btree
2032 * whiteout, with a real key at the same position, since
2033 * in the btree deleted keys sort before non deleted.
2035 search_key = !bpos_eq(search_key, k.k->p)
2037 : bpos_successor(k.k->p);
2043 } else if (likely(!bpos_eq(l->b->key.k.p, SPOS_MAX))) {
2044 /* Advance to next leaf node: */
2045 search_key = bpos_successor(l->b->key.k.p);
2048 bch2_btree_iter_set_pos(iter, SPOS_MAX);
2054 bch2_btree_iter_verify(iter);
2060 * bch2_btree_iter_peek: returns first key greater than or equal to iterator's
2063 struct bkey_s_c bch2_btree_iter_peek_upto(struct btree_iter *iter, struct bpos end)
2065 struct btree_trans *trans = iter->trans;
2066 struct bpos search_key = btree_iter_search_key(iter);
2068 struct bpos iter_pos;
2071 EBUG_ON(iter->flags & BTREE_ITER_ALL_LEVELS);
2072 EBUG_ON((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) && bkey_eq(end, POS_MAX));
2074 if (iter->update_path) {
2075 bch2_path_put_nokeep(trans, iter->update_path,
2076 iter->flags & BTREE_ITER_INTENT);
2077 iter->update_path = NULL;
2080 bch2_btree_iter_verify_entry_exit(iter);
2083 k = __bch2_btree_iter_peek(iter, search_key);
2086 if (unlikely(bkey_err(k)))
2090 * iter->pos should be mononotically increasing, and always be
2091 * equal to the key we just returned - except extents can
2092 * straddle iter->pos:
2094 if (!(iter->flags & BTREE_ITER_IS_EXTENTS))
2097 iter_pos = bkey_max(iter->pos, bkey_start_pos(k.k));
2099 if (unlikely(!(iter->flags & BTREE_ITER_IS_EXTENTS)
2100 ? bkey_gt(iter_pos, end)
2101 : bkey_ge(iter_pos, end)))
2104 if (iter->update_path &&
2105 !bkey_eq(iter->update_path->pos, k.k->p)) {
2106 bch2_path_put_nokeep(trans, iter->update_path,
2107 iter->flags & BTREE_ITER_INTENT);
2108 iter->update_path = NULL;
2111 if ((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
2112 (iter->flags & BTREE_ITER_INTENT) &&
2113 !(iter->flags & BTREE_ITER_IS_EXTENTS) &&
2114 !iter->update_path) {
2115 struct bpos pos = k.k->p;
2117 if (pos.snapshot < iter->snapshot) {
2118 search_key = bpos_successor(k.k->p);
2122 pos.snapshot = iter->snapshot;
2125 * advance, same as on exit for iter->path, but only up
2128 __btree_path_get(iter->path, iter->flags & BTREE_ITER_INTENT);
2129 iter->update_path = iter->path;
2131 iter->update_path = bch2_btree_path_set_pos(trans,
2132 iter->update_path, pos,
2133 iter->flags & BTREE_ITER_INTENT,
2135 ret = bch2_btree_path_traverse(trans, iter->update_path, iter->flags);
2136 if (unlikely(ret)) {
2137 k = bkey_s_c_err(ret);
2143 * We can never have a key in a leaf node at POS_MAX, so
2144 * we don't have to check these successor() calls:
2146 if ((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
2147 !bch2_snapshot_is_ancestor(trans->c,
2150 search_key = bpos_successor(k.k->p);
2154 if (bkey_whiteout(k.k) &&
2155 !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS)) {
2156 search_key = bkey_successor(iter, k.k->p);
2163 iter->pos = iter_pos;
2165 iter->path = bch2_btree_path_set_pos(trans, iter->path, k.k->p,
2166 iter->flags & BTREE_ITER_INTENT,
2167 btree_iter_ip_allocated(iter));
2169 btree_path_set_should_be_locked(iter->path);
2171 if (iter->update_path) {
2172 ret = bch2_btree_path_relock(trans, iter->update_path, _THIS_IP_);
2174 k = bkey_s_c_err(ret);
2176 btree_path_set_should_be_locked(iter->update_path);
2179 if (!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS))
2180 iter->pos.snapshot = iter->snapshot;
2182 ret = bch2_btree_iter_verify_ret(iter, k);
2183 if (unlikely(ret)) {
2184 bch2_btree_iter_set_pos(iter, iter->pos);
2185 k = bkey_s_c_err(ret);
2188 bch2_btree_iter_verify_entry_exit(iter);
2192 bch2_btree_iter_set_pos(iter, end);
2198 * bch2_btree_iter_peek_all_levels: returns the first key greater than or equal
2199 * to iterator's current position, returning keys from every level of the btree.
2200 * For keys at different levels of the btree that compare equal, the key from
2201 * the lower level (leaf) is returned first.
2203 struct bkey_s_c bch2_btree_iter_peek_all_levels(struct btree_iter *iter)
2205 struct btree_trans *trans = iter->trans;
2209 EBUG_ON(iter->path->cached);
2210 bch2_btree_iter_verify(iter);
2211 BUG_ON(iter->path->level < iter->min_depth);
2212 BUG_ON(!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS));
2213 EBUG_ON(!(iter->flags & BTREE_ITER_ALL_LEVELS));
2216 iter->path = bch2_btree_path_set_pos(trans, iter->path, iter->pos,
2217 iter->flags & BTREE_ITER_INTENT,
2218 btree_iter_ip_allocated(iter));
2220 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2221 if (unlikely(ret)) {
2222 /* ensure that iter->k is consistent with iter->pos: */
2223 bch2_btree_iter_set_pos(iter, iter->pos);
2224 k = bkey_s_c_err(ret);
2228 /* Already at end? */
2229 if (!btree_path_node(iter->path, iter->path->level)) {
2234 k = btree_path_level_peek_all(trans->c,
2235 &iter->path->l[iter->path->level], &iter->k);
2237 /* Check if we should go up to the parent node: */
2240 bpos_eq(path_l(iter->path)->b->key.k.p, iter->pos))) {
2241 iter->pos = path_l(iter->path)->b->key.k.p;
2242 btree_path_set_level_up(trans, iter->path);
2243 iter->advanced = false;
2248 * Check if we should go back down to a leaf:
2249 * If we're not in a leaf node, we only return the current key
2250 * if it exactly matches iter->pos - otherwise we first have to
2251 * go back to the leaf:
2253 if (iter->path->level != iter->min_depth &&
2256 !bpos_eq(iter->pos, k.k->p))) {
2257 btree_path_set_level_down(trans, iter->path, iter->min_depth);
2258 iter->pos = bpos_successor(iter->pos);
2259 iter->advanced = false;
2263 /* Check if we should go to the next key: */
2264 if (iter->path->level == iter->min_depth &&
2267 bpos_eq(iter->pos, k.k->p)) {
2268 iter->pos = bpos_successor(iter->pos);
2269 iter->advanced = false;
2273 if (iter->advanced &&
2274 iter->path->level == iter->min_depth &&
2275 !bpos_eq(k.k->p, iter->pos))
2276 iter->advanced = false;
2278 BUG_ON(iter->advanced);
2284 btree_path_set_should_be_locked(iter->path);
2286 bch2_btree_iter_verify(iter);
2292 * bch2_btree_iter_next: returns first key greater than iterator's current
2295 struct bkey_s_c bch2_btree_iter_next(struct btree_iter *iter)
2297 if (!bch2_btree_iter_advance(iter))
2298 return bkey_s_c_null;
2300 return bch2_btree_iter_peek(iter);
2304 * bch2_btree_iter_peek_prev: returns first key less than or equal to
2305 * iterator's current position
2307 struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *iter)
2309 struct btree_trans *trans = iter->trans;
2310 struct bpos search_key = iter->pos;
2311 struct btree_path *saved_path = NULL;
2313 struct bkey saved_k;
2314 const struct bch_val *saved_v;
2317 EBUG_ON(iter->path->cached || iter->path->level);
2318 EBUG_ON(iter->flags & BTREE_ITER_WITH_UPDATES);
2320 if (iter->flags & BTREE_ITER_WITH_JOURNAL)
2321 return bkey_s_c_err(-EIO);
2323 bch2_btree_iter_verify(iter);
2324 bch2_btree_iter_verify_entry_exit(iter);
2326 if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2327 search_key.snapshot = U32_MAX;
2330 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2331 iter->flags & BTREE_ITER_INTENT,
2332 btree_iter_ip_allocated(iter));
2334 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2335 if (unlikely(ret)) {
2336 /* ensure that iter->k is consistent with iter->pos: */
2337 bch2_btree_iter_set_pos(iter, iter->pos);
2338 k = bkey_s_c_err(ret);
2342 k = btree_path_level_peek(trans, iter->path,
2343 &iter->path->l[0], &iter->k);
2345 ((iter->flags & BTREE_ITER_IS_EXTENTS)
2346 ? bpos_ge(bkey_start_pos(k.k), search_key)
2347 : bpos_gt(k.k->p, search_key)))
2348 k = btree_path_level_prev(trans, iter->path,
2349 &iter->path->l[0], &iter->k);
2352 if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) {
2353 if (k.k->p.snapshot == iter->snapshot)
2357 * If we have a saved candidate, and we're no
2358 * longer at the same _key_ (not pos), return
2361 if (saved_path && !bkey_eq(k.k->p, saved_k.p)) {
2362 bch2_path_put_nokeep(trans, iter->path,
2363 iter->flags & BTREE_ITER_INTENT);
2364 iter->path = saved_path;
2371 if (bch2_snapshot_is_ancestor(iter->trans->c,
2375 bch2_path_put_nokeep(trans, saved_path,
2376 iter->flags & BTREE_ITER_INTENT);
2377 saved_path = btree_path_clone(trans, iter->path,
2378 iter->flags & BTREE_ITER_INTENT);
2383 search_key = bpos_predecessor(k.k->p);
2387 if (bkey_whiteout(k.k) &&
2388 !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS)) {
2389 search_key = bkey_predecessor(iter, k.k->p);
2390 if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2391 search_key.snapshot = U32_MAX;
2396 } else if (likely(!bpos_eq(iter->path->l[0].b->data->min_key, POS_MIN))) {
2397 /* Advance to previous leaf node: */
2398 search_key = bpos_predecessor(iter->path->l[0].b->data->min_key);
2400 /* Start of btree: */
2401 bch2_btree_iter_set_pos(iter, POS_MIN);
2407 EBUG_ON(bkey_gt(bkey_start_pos(k.k), iter->pos));
2409 /* Extents can straddle iter->pos: */
2410 if (bkey_lt(k.k->p, iter->pos))
2413 if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2414 iter->pos.snapshot = iter->snapshot;
2416 btree_path_set_should_be_locked(iter->path);
2419 bch2_path_put_nokeep(trans, saved_path, iter->flags & BTREE_ITER_INTENT);
2421 bch2_btree_iter_verify_entry_exit(iter);
2422 bch2_btree_iter_verify(iter);
2428 * bch2_btree_iter_prev: returns first key less than iterator's current
2431 struct bkey_s_c bch2_btree_iter_prev(struct btree_iter *iter)
2433 if (!bch2_btree_iter_rewind(iter))
2434 return bkey_s_c_null;
2436 return bch2_btree_iter_peek_prev(iter);
2439 struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *iter)
2441 struct btree_trans *trans = iter->trans;
2442 struct bpos search_key;
2446 bch2_btree_iter_verify(iter);
2447 bch2_btree_iter_verify_entry_exit(iter);
2448 EBUG_ON(iter->flags & BTREE_ITER_ALL_LEVELS);
2449 EBUG_ON(iter->path->level && (iter->flags & BTREE_ITER_WITH_KEY_CACHE));
2451 /* extents can't span inode numbers: */
2452 if ((iter->flags & BTREE_ITER_IS_EXTENTS) &&
2453 unlikely(iter->pos.offset == KEY_OFFSET_MAX)) {
2454 if (iter->pos.inode == KEY_INODE_MAX)
2455 return bkey_s_c_null;
2457 bch2_btree_iter_set_pos(iter, bpos_nosnap_successor(iter->pos));
2460 search_key = btree_iter_search_key(iter);
2461 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2462 iter->flags & BTREE_ITER_INTENT,
2463 btree_iter_ip_allocated(iter));
2465 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2466 if (unlikely(ret)) {
2467 k = bkey_s_c_err(ret);
2471 if ((iter->flags & BTREE_ITER_CACHED) ||
2472 !(iter->flags & (BTREE_ITER_IS_EXTENTS|BTREE_ITER_FILTER_SNAPSHOTS))) {
2473 struct bkey_i *next_update;
2475 if ((next_update = btree_trans_peek_updates(iter)) &&
2476 bpos_eq(next_update->k.p, iter->pos)) {
2477 iter->k = next_update->k;
2478 k = bkey_i_to_s_c(next_update);
2482 if (unlikely(iter->flags & BTREE_ITER_WITH_JOURNAL) &&
2483 (k = btree_trans_peek_slot_journal(trans, iter)).k)
2486 if (unlikely(iter->flags & BTREE_ITER_WITH_KEY_CACHE) &&
2487 (k = btree_trans_peek_key_cache(iter, iter->pos)).k) {
2490 /* We're not returning a key from iter->path: */
2494 k = bch2_btree_path_peek_slot(iter->path, &iter->k);
2499 struct bpos end = iter->pos;
2501 if (iter->flags & BTREE_ITER_IS_EXTENTS)
2502 end.offset = U64_MAX;
2504 EBUG_ON(iter->path->level);
2506 if (iter->flags & BTREE_ITER_INTENT) {
2507 struct btree_iter iter2;
2509 bch2_trans_copy_iter(&iter2, iter);
2510 k = bch2_btree_iter_peek_upto(&iter2, end);
2512 if (k.k && !bkey_err(k)) {
2516 bch2_trans_iter_exit(trans, &iter2);
2518 struct bpos pos = iter->pos;
2520 k = bch2_btree_iter_peek_upto(iter, end);
2521 if (unlikely(bkey_err(k)))
2522 bch2_btree_iter_set_pos(iter, pos);
2527 if (unlikely(bkey_err(k)))
2530 next = k.k ? bkey_start_pos(k.k) : POS_MAX;
2532 if (bkey_lt(iter->pos, next)) {
2533 bkey_init(&iter->k);
2534 iter->k.p = iter->pos;
2536 if (iter->flags & BTREE_ITER_IS_EXTENTS) {
2537 bch2_key_resize(&iter->k,
2538 min_t(u64, KEY_SIZE_MAX,
2539 (next.inode == iter->pos.inode
2543 EBUG_ON(!iter->k.size);
2546 k = (struct bkey_s_c) { &iter->k, NULL };
2550 btree_path_set_should_be_locked(iter->path);
2552 bch2_btree_iter_verify_entry_exit(iter);
2553 bch2_btree_iter_verify(iter);
2554 ret = bch2_btree_iter_verify_ret(iter, k);
2556 return bkey_s_c_err(ret);
2561 struct bkey_s_c bch2_btree_iter_next_slot(struct btree_iter *iter)
2563 if (!bch2_btree_iter_advance(iter))
2564 return bkey_s_c_null;
2566 return bch2_btree_iter_peek_slot(iter);
2569 struct bkey_s_c bch2_btree_iter_prev_slot(struct btree_iter *iter)
2571 if (!bch2_btree_iter_rewind(iter))
2572 return bkey_s_c_null;
2574 return bch2_btree_iter_peek_slot(iter);
2577 /* new transactional stuff: */
2579 #ifdef CONFIG_BCACHEFS_DEBUG
2580 static void btree_trans_verify_sorted_refs(struct btree_trans *trans)
2582 struct btree_path *path;
2585 BUG_ON(trans->nr_sorted != hweight64(trans->paths_allocated));
2587 trans_for_each_path(trans, path) {
2588 BUG_ON(path->sorted_idx >= trans->nr_sorted);
2589 BUG_ON(trans->sorted[path->sorted_idx] != path->idx);
2592 for (i = 0; i < trans->nr_sorted; i++) {
2593 unsigned idx = trans->sorted[i];
2595 EBUG_ON(!(trans->paths_allocated & (1ULL << idx)));
2596 BUG_ON(trans->paths[idx].sorted_idx != i);
2600 static void btree_trans_verify_sorted(struct btree_trans *trans)
2602 struct btree_path *path, *prev = NULL;
2605 if (!bch2_debug_check_iterators)
2608 trans_for_each_path_inorder(trans, path, i) {
2609 if (prev && btree_path_cmp(prev, path) > 0) {
2610 __bch2_dump_trans_paths_updates(trans, true);
2611 panic("trans paths out of order!\n");
2617 static inline void btree_trans_verify_sorted_refs(struct btree_trans *trans) {}
2618 static inline void btree_trans_verify_sorted(struct btree_trans *trans) {}
2621 void __bch2_btree_trans_sort_paths(struct btree_trans *trans)
2623 int i, l = 0, r = trans->nr_sorted, inc = 1;
2626 btree_trans_verify_sorted_refs(trans);
2628 if (trans->paths_sorted)
2632 * Cocktail shaker sort: this is efficient because iterators will be
2638 for (i = inc > 0 ? l : r - 2;
2639 i + 1 < r && i >= l;
2641 if (btree_path_cmp(trans->paths + trans->sorted[i],
2642 trans->paths + trans->sorted[i + 1]) > 0) {
2643 swap(trans->sorted[i], trans->sorted[i + 1]);
2644 trans->paths[trans->sorted[i]].sorted_idx = i;
2645 trans->paths[trans->sorted[i + 1]].sorted_idx = i + 1;
2657 trans->paths_sorted = true;
2659 btree_trans_verify_sorted(trans);
2662 static inline void btree_path_list_remove(struct btree_trans *trans,
2663 struct btree_path *path)
2667 EBUG_ON(path->sorted_idx >= trans->nr_sorted);
2668 #ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
2670 memmove_u64s_down_small(trans->sorted + path->sorted_idx,
2671 trans->sorted + path->sorted_idx + 1,
2672 DIV_ROUND_UP(trans->nr_sorted - path->sorted_idx, 8));
2674 array_remove_item(trans->sorted, trans->nr_sorted, path->sorted_idx);
2676 for (i = path->sorted_idx; i < trans->nr_sorted; i++)
2677 trans->paths[trans->sorted[i]].sorted_idx = i;
2679 path->sorted_idx = U8_MAX;
2682 static inline void btree_path_list_add(struct btree_trans *trans,
2683 struct btree_path *pos,
2684 struct btree_path *path)
2688 path->sorted_idx = pos ? pos->sorted_idx + 1 : trans->nr_sorted;
2690 #ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
2691 memmove_u64s_up_small(trans->sorted + path->sorted_idx + 1,
2692 trans->sorted + path->sorted_idx,
2693 DIV_ROUND_UP(trans->nr_sorted - path->sorted_idx, 8));
2695 trans->sorted[path->sorted_idx] = path->idx;
2697 array_insert_item(trans->sorted, trans->nr_sorted, path->sorted_idx, path->idx);
2700 for (i = path->sorted_idx; i < trans->nr_sorted; i++)
2701 trans->paths[trans->sorted[i]].sorted_idx = i;
2703 btree_trans_verify_sorted_refs(trans);
2706 void bch2_trans_iter_exit(struct btree_trans *trans, struct btree_iter *iter)
2709 bch2_path_put(trans, iter->path,
2710 iter->flags & BTREE_ITER_INTENT);
2711 if (iter->update_path)
2712 bch2_path_put_nokeep(trans, iter->update_path,
2713 iter->flags & BTREE_ITER_INTENT);
2714 if (iter->key_cache_path)
2715 bch2_path_put(trans, iter->key_cache_path,
2716 iter->flags & BTREE_ITER_INTENT);
2718 iter->update_path = NULL;
2719 iter->key_cache_path = NULL;
2722 static inline void bch2_trans_iter_init_inlined(struct btree_trans *trans,
2723 struct btree_iter *iter,
2724 unsigned btree_id, struct bpos pos,
2727 bch2_trans_iter_init_common(trans, iter, btree_id, pos, 0, 0,
2728 bch2_btree_iter_flags(trans, btree_id, flags),
2732 void bch2_trans_iter_init_outlined(struct btree_trans *trans,
2733 struct btree_iter *iter,
2734 unsigned btree_id, struct bpos pos,
2737 bch2_trans_iter_init_common(trans, iter, btree_id, pos, 0, 0,
2738 bch2_btree_iter_flags(trans, btree_id, flags),
2742 void bch2_trans_node_iter_init(struct btree_trans *trans,
2743 struct btree_iter *iter,
2744 enum btree_id btree_id,
2746 unsigned locks_want,
2750 flags |= BTREE_ITER_NOT_EXTENTS;
2751 flags |= __BTREE_ITER_ALL_SNAPSHOTS;
2752 flags |= BTREE_ITER_ALL_SNAPSHOTS;
2754 bch2_trans_iter_init_common(trans, iter, btree_id, pos, locks_want, depth,
2755 __bch2_btree_iter_flags(trans, btree_id, flags),
2758 iter->min_depth = depth;
2760 BUG_ON(iter->path->locks_want < min(locks_want, BTREE_MAX_DEPTH));
2761 BUG_ON(iter->path->level != depth);
2762 BUG_ON(iter->min_depth != depth);
2765 void bch2_trans_copy_iter(struct btree_iter *dst, struct btree_iter *src)
2769 __btree_path_get(src->path, src->flags & BTREE_ITER_INTENT);
2770 if (src->update_path)
2771 __btree_path_get(src->update_path, src->flags & BTREE_ITER_INTENT);
2772 dst->key_cache_path = NULL;
2775 void *__bch2_trans_kmalloc(struct btree_trans *trans, size_t size)
2777 unsigned new_top = trans->mem_top + size;
2778 size_t old_bytes = trans->mem_bytes;
2779 size_t new_bytes = roundup_pow_of_two(new_top);
2783 trans->mem_max = max(trans->mem_max, new_top);
2785 WARN_ON_ONCE(new_bytes > BTREE_TRANS_MEM_MAX);
2787 new_mem = krealloc(trans->mem, new_bytes, GFP_NOFS);
2788 if (!new_mem && new_bytes <= BTREE_TRANS_MEM_MAX) {
2789 new_mem = mempool_alloc(&trans->c->btree_trans_mem_pool, GFP_KERNEL);
2790 new_bytes = BTREE_TRANS_MEM_MAX;
2795 return ERR_PTR(-ENOMEM);
2797 trans->mem = new_mem;
2798 trans->mem_bytes = new_bytes;
2801 trace_and_count(trans->c, trans_restart_mem_realloced, trans, _RET_IP_, new_bytes);
2802 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_mem_realloced));
2805 p = trans->mem + trans->mem_top;
2806 trans->mem_top += size;
2811 static noinline void bch2_trans_reset_srcu_lock(struct btree_trans *trans)
2813 struct bch_fs *c = trans->c;
2814 struct btree_path *path;
2816 trans_for_each_path(trans, path)
2817 if (path->cached && !btree_node_locked(path, 0))
2818 path->l[0].b = ERR_PTR(-BCH_ERR_no_btree_node_srcu_reset);
2820 srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx);
2821 trans->srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
2822 trans->srcu_lock_time = jiffies;
2826 * bch2_trans_begin() - reset a transaction after a interrupted attempt
2827 * @trans: transaction to reset
2829 * While iterating over nodes or updating nodes a attempt to lock a btree node
2830 * may return BCH_ERR_transaction_restart when the trylock fails. When this
2831 * occurs bch2_trans_begin() should be called and the transaction retried.
2833 u32 bch2_trans_begin(struct btree_trans *trans)
2835 struct btree_path *path;
2837 bch2_trans_reset_updates(trans);
2839 trans->restart_count++;
2842 trans_for_each_path(trans, path) {
2843 path->should_be_locked = false;
2846 * If the transaction wasn't restarted, we're presuming to be
2847 * doing something new: dont keep iterators excpt the ones that
2848 * are in use - except for the subvolumes btree:
2850 if (!trans->restarted && path->btree_id != BTREE_ID_subvolumes)
2851 path->preserve = false;
2854 * XXX: we probably shouldn't be doing this if the transaction
2855 * was restarted, but currently we still overflow transaction
2856 * iterators if we do that
2858 if (!path->ref && !path->preserve)
2859 __bch2_path_free(trans, path);
2861 path->preserve = false;
2864 if (!trans->restarted &&
2866 local_clock() - trans->last_begin_time > BTREE_TRANS_MAX_LOCK_HOLD_TIME_NS)) {
2867 bch2_trans_unlock(trans);
2869 bch2_trans_relock(trans);
2872 if (unlikely(time_after(jiffies, trans->srcu_lock_time + msecs_to_jiffies(10))))
2873 bch2_trans_reset_srcu_lock(trans);
2875 trans->last_begin_ip = _RET_IP_;
2876 if (trans->restarted) {
2877 bch2_btree_path_traverse_all(trans);
2878 trans->notrace_relock_fail = false;
2881 trans->last_begin_time = local_clock();
2882 return trans->restart_count;
2885 static void bch2_trans_alloc_paths(struct btree_trans *trans, struct bch_fs *c)
2887 size_t paths_bytes = sizeof(struct btree_path) * BTREE_ITER_MAX;
2888 size_t updates_bytes = sizeof(struct btree_insert_entry) * BTREE_ITER_MAX;
2891 BUG_ON(trans->used_mempool);
2894 p = this_cpu_xchg(c->btree_paths_bufs->path, NULL);
2897 p = mempool_alloc(&trans->c->btree_paths_pool, GFP_NOFS);
2899 trans->paths = p; p += paths_bytes;
2900 trans->updates = p; p += updates_bytes;
2903 const char *bch2_btree_transaction_fns[BCH_TRANSACTIONS_NR];
2905 unsigned bch2_trans_get_fn_idx(const char *fn)
2909 for (i = 0; i < ARRAY_SIZE(bch2_btree_transaction_fns); i++)
2910 if (!bch2_btree_transaction_fns[i] ||
2911 bch2_btree_transaction_fns[i] == fn) {
2912 bch2_btree_transaction_fns[i] = fn;
2916 pr_warn_once("BCH_TRANSACTIONS_NR not big enough!");
2920 void __bch2_trans_init(struct btree_trans *trans, struct bch_fs *c, unsigned fn_idx)
2921 __acquires(&c->btree_trans_barrier)
2923 struct btree_transaction_stats *s;
2925 BUG_ON(lock_class_is_held(&bch2_btree_node_lock_key));
2927 memset(trans, 0, sizeof(*trans));
2929 trans->fn = fn_idx < ARRAY_SIZE(bch2_btree_transaction_fns)
2930 ? bch2_btree_transaction_fns[fn_idx] : NULL;
2931 trans->last_begin_time = local_clock();
2932 trans->fn_idx = fn_idx;
2933 trans->locking_wait.task = current;
2934 trans->journal_replay_not_finished =
2935 !test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags);
2936 closure_init_stack(&trans->ref);
2938 bch2_trans_alloc_paths(trans, c);
2940 s = btree_trans_stats(trans);
2941 if (s && s->max_mem) {
2942 unsigned expected_mem_bytes = roundup_pow_of_two(s->max_mem);
2944 trans->mem = kmalloc(expected_mem_bytes, GFP_KERNEL);
2946 if (!unlikely(trans->mem)) {
2947 trans->mem = mempool_alloc(&c->btree_trans_mem_pool, GFP_KERNEL);
2948 trans->mem_bytes = BTREE_TRANS_MEM_MAX;
2950 trans->mem_bytes = expected_mem_bytes;
2955 trans->nr_max_paths = s->nr_max_paths;
2956 trans->wb_updates_size = s->wb_updates_size;
2959 trans->srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
2960 trans->srcu_lock_time = jiffies;
2962 if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG_TRANSACTIONS)) {
2963 struct btree_trans *pos;
2965 mutex_lock(&c->btree_trans_lock);
2966 list_for_each_entry(pos, &c->btree_trans_list, list) {
2967 if (trans->locking_wait.task->pid < pos->locking_wait.task->pid) {
2968 list_add_tail(&trans->list, &pos->list);
2972 list_add_tail(&trans->list, &c->btree_trans_list);
2974 mutex_unlock(&c->btree_trans_lock);
2978 static void check_btree_paths_leaked(struct btree_trans *trans)
2980 #ifdef CONFIG_BCACHEFS_DEBUG
2981 struct bch_fs *c = trans->c;
2982 struct btree_path *path;
2984 trans_for_each_path(trans, path)
2989 bch_err(c, "btree paths leaked from %s!", trans->fn);
2990 trans_for_each_path(trans, path)
2992 printk(KERN_ERR " btree %s %pS\n",
2993 bch2_btree_ids[path->btree_id],
2994 (void *) path->ip_allocated);
2995 /* Be noisy about this: */
2996 bch2_fatal_error(c);
3000 void bch2_trans_exit(struct btree_trans *trans)
3001 __releases(&c->btree_trans_barrier)
3003 struct btree_insert_entry *i;
3004 struct bch_fs *c = trans->c;
3005 struct btree_transaction_stats *s = btree_trans_stats(trans);
3007 bch2_trans_unlock(trans);
3009 closure_sync(&trans->ref);
3012 s->max_mem = max(s->max_mem, trans->mem_max);
3014 trans_for_each_update(trans, i)
3015 __btree_path_put(i->path, true);
3016 trans->nr_updates = 0;
3018 check_btree_paths_leaked(trans);
3020 if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG_TRANSACTIONS)) {
3021 mutex_lock(&c->btree_trans_lock);
3022 list_del(&trans->list);
3023 mutex_unlock(&c->btree_trans_lock);
3026 srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx);
3028 bch2_journal_preres_put(&c->journal, &trans->journal_preres);
3030 kfree(trans->extra_journal_entries.data);
3032 if (trans->fs_usage_deltas) {
3033 if (trans->fs_usage_deltas->size + sizeof(trans->fs_usage_deltas) ==
3034 REPLICAS_DELTA_LIST_MAX)
3035 mempool_free(trans->fs_usage_deltas,
3036 &c->replicas_delta_pool);
3038 kfree(trans->fs_usage_deltas);
3041 if (trans->mem_bytes == BTREE_TRANS_MEM_MAX)
3042 mempool_free(trans->mem, &c->btree_trans_mem_pool);
3048 * Userspace doesn't have a real percpu implementation:
3050 trans->paths = this_cpu_xchg(c->btree_paths_bufs->path, trans->paths);
3054 mempool_free(trans->paths, &c->btree_paths_pool);
3056 #ifdef CONFIG_BCACHEFS_DEBUG
3057 darray_exit(&trans->last_restarted);
3060 trans->mem = (void *) 0x1;
3061 trans->paths = (void *) 0x1;
3064 static void __maybe_unused
3065 bch2_btree_bkey_cached_common_to_text(struct printbuf *out,
3066 struct btree_bkey_cached_common *b)
3068 struct six_lock_count c = six_lock_counts(&b->lock);
3069 struct task_struct *owner;
3073 owner = READ_ONCE(b->lock.owner);
3074 pid = owner ? owner->pid : 0;
3078 prt_printf(out, "%px %c l=%u %s:", b, b->cached ? 'c' : 'b',
3079 b->level, bch2_btree_ids[b->btree_id]);
3080 bch2_bpos_to_text(out, btree_node_pos(b));
3083 prt_printf(out, " locks %u:%u:%u held by pid %u",
3084 c.n[0], c.n[1], c.n[2], pid);
3087 void bch2_btree_trans_to_text(struct printbuf *out, struct btree_trans *trans)
3089 struct btree_path *path;
3090 struct btree_bkey_cached_common *b;
3091 static char lock_types[] = { 'r', 'i', 'w' };
3094 if (!out->nr_tabstops) {
3095 printbuf_tabstop_push(out, 16);
3096 printbuf_tabstop_push(out, 32);
3099 prt_printf(out, "%i %s\n", trans->locking_wait.task->pid, trans->fn);
3101 trans_for_each_path(trans, path) {
3102 if (!path->nodes_locked)
3105 prt_printf(out, " path %u %c l=%u %s:",
3107 path->cached ? 'c' : 'b',
3109 bch2_btree_ids[path->btree_id]);
3110 bch2_bpos_to_text(out, path->pos);
3113 for (l = 0; l < BTREE_MAX_DEPTH; l++) {
3114 if (btree_node_locked(path, l) &&
3115 !IS_ERR_OR_NULL(b = (void *) READ_ONCE(path->l[l].b))) {
3116 prt_printf(out, " %c l=%u ",
3117 lock_types[btree_node_locked_type(path, l)], l);
3118 bch2_btree_bkey_cached_common_to_text(out, b);
3124 b = READ_ONCE(trans->locking);
3126 prt_printf(out, " blocked for %lluus on",
3127 div_u64(local_clock() - trans->locking_wait.start_time,
3130 prt_printf(out, " %c", lock_types[trans->locking_wait.lock_want]);
3131 bch2_btree_bkey_cached_common_to_text(out, b);
3136 void bch2_fs_btree_iter_exit(struct bch_fs *c)
3138 struct btree_transaction_stats *s;
3140 for (s = c->btree_transaction_stats;
3141 s < c->btree_transaction_stats + ARRAY_SIZE(c->btree_transaction_stats);
3143 kfree(s->max_paths_text);
3144 bch2_time_stats_exit(&s->lock_hold_times);
3147 if (c->btree_trans_barrier_initialized)
3148 cleanup_srcu_struct(&c->btree_trans_barrier);
3149 mempool_exit(&c->btree_trans_mem_pool);
3150 mempool_exit(&c->btree_paths_pool);
3153 int bch2_fs_btree_iter_init(struct bch_fs *c)
3155 struct btree_transaction_stats *s;
3156 unsigned nr = BTREE_ITER_MAX;
3159 for (s = c->btree_transaction_stats;
3160 s < c->btree_transaction_stats + ARRAY_SIZE(c->btree_transaction_stats);
3162 bch2_time_stats_init(&s->lock_hold_times);
3163 mutex_init(&s->lock);
3166 INIT_LIST_HEAD(&c->btree_trans_list);
3167 mutex_init(&c->btree_trans_lock);
3169 ret = mempool_init_kmalloc_pool(&c->btree_paths_pool, 1,
3170 sizeof(struct btree_path) * nr +
3171 sizeof(struct btree_insert_entry) * nr) ?:
3172 mempool_init_kmalloc_pool(&c->btree_trans_mem_pool, 1,
3173 BTREE_TRANS_MEM_MAX) ?:
3174 init_srcu_struct(&c->btree_trans_barrier);
3176 c->btree_trans_barrier_initialized = true;