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_journal_iter.h"
9 #include "btree_key_cache.h"
10 #include "btree_locking.h"
11 #include "btree_update.h"
16 #include "journal_io.h"
21 #include <linux/random.h>
22 #include <linux/prefetch.h>
24 static inline void btree_path_list_remove(struct btree_trans *, struct btree_path *);
25 static inline void btree_path_list_add(struct btree_trans *,
26 btree_path_idx_t, btree_path_idx_t);
28 static inline unsigned long btree_iter_ip_allocated(struct btree_iter *iter)
30 #ifdef TRACK_PATH_ALLOCATED
31 return iter->ip_allocated;
37 static btree_path_idx_t btree_path_alloc(struct btree_trans *, btree_path_idx_t);
38 static void bch2_trans_srcu_lock(struct btree_trans *);
40 static inline int __btree_path_cmp(const struct btree_path *l,
41 enum btree_id r_btree_id,
47 * Must match lock ordering as defined by __bch2_btree_node_lock:
49 return cmp_int(l->btree_id, r_btree_id) ?:
50 cmp_int((int) l->cached, (int) r_cached) ?:
51 bpos_cmp(l->pos, r_pos) ?:
52 -cmp_int(l->level, r_level);
55 static inline int btree_path_cmp(const struct btree_path *l,
56 const struct btree_path *r)
58 return __btree_path_cmp(l, r->btree_id, r->cached, r->pos, r->level);
61 static inline struct bpos bkey_successor(struct btree_iter *iter, struct bpos p)
63 /* Are we iterating over keys in all snapshots? */
64 if (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) {
65 p = bpos_successor(p);
67 p = bpos_nosnap_successor(p);
68 p.snapshot = iter->snapshot;
74 static inline struct bpos bkey_predecessor(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_predecessor(p);
80 p = bpos_nosnap_predecessor(p);
81 p.snapshot = iter->snapshot;
87 static inline struct bpos btree_iter_search_key(struct btree_iter *iter)
89 struct bpos pos = iter->pos;
91 if ((iter->flags & BTREE_ITER_IS_EXTENTS) &&
92 !bkey_eq(pos, POS_MAX))
93 pos = bkey_successor(iter, pos);
97 static inline bool btree_path_pos_before_node(struct btree_path *path,
100 return bpos_lt(path->pos, b->data->min_key);
103 static inline bool btree_path_pos_after_node(struct btree_path *path,
106 return bpos_gt(path->pos, b->key.k.p);
109 static inline bool btree_path_pos_in_node(struct btree_path *path,
112 return path->btree_id == b->c.btree_id &&
113 !btree_path_pos_before_node(path, b) &&
114 !btree_path_pos_after_node(path, b);
117 /* Btree iterator: */
119 #ifdef CONFIG_BCACHEFS_DEBUG
121 static void bch2_btree_path_verify_cached(struct btree_trans *trans,
122 struct btree_path *path)
124 struct bkey_cached *ck;
125 bool locked = btree_node_locked(path, 0);
127 if (!bch2_btree_node_relock(trans, path, 0))
130 ck = (void *) path->l[0].b;
131 BUG_ON(ck->key.btree_id != path->btree_id ||
132 !bkey_eq(ck->key.pos, path->pos));
135 btree_node_unlock(trans, path, 0);
138 static void bch2_btree_path_verify_level(struct btree_trans *trans,
139 struct btree_path *path, unsigned level)
141 struct btree_path_level *l;
142 struct btree_node_iter tmp;
144 struct bkey_packed *p, *k;
145 struct printbuf buf1 = PRINTBUF;
146 struct printbuf buf2 = PRINTBUF;
147 struct printbuf buf3 = PRINTBUF;
150 if (!bch2_debug_check_iterators)
155 locked = btree_node_locked(path, level);
159 bch2_btree_path_verify_cached(trans, path);
163 if (!btree_path_node(path, level))
166 if (!bch2_btree_node_relock_notrace(trans, path, level))
169 BUG_ON(!btree_path_pos_in_node(path, l->b));
171 bch2_btree_node_iter_verify(&l->iter, l->b);
174 * For interior nodes, the iterator will have skipped past deleted keys:
177 ? bch2_btree_node_iter_prev(&tmp, l->b)
178 : bch2_btree_node_iter_prev_all(&tmp, l->b);
179 k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
181 if (p && bkey_iter_pos_cmp(l->b, p, &path->pos) >= 0) {
186 if (k && bkey_iter_pos_cmp(l->b, k, &path->pos) < 0) {
192 btree_node_unlock(trans, path, level);
195 bch2_bpos_to_text(&buf1, path->pos);
198 struct bkey uk = bkey_unpack_key(l->b, p);
200 bch2_bkey_to_text(&buf2, &uk);
202 prt_printf(&buf2, "(none)");
206 struct bkey uk = bkey_unpack_key(l->b, k);
208 bch2_bkey_to_text(&buf3, &uk);
210 prt_printf(&buf3, "(none)");
213 panic("path should be %s key at level %u:\n"
217 msg, level, buf1.buf, buf2.buf, buf3.buf);
220 static void bch2_btree_path_verify(struct btree_trans *trans,
221 struct btree_path *path)
223 struct bch_fs *c = trans->c;
226 EBUG_ON(path->btree_id >= BTREE_ID_NR);
228 for (i = 0; i < (!path->cached ? BTREE_MAX_DEPTH : 1); i++) {
230 BUG_ON(!path->cached &&
231 bch2_btree_id_root(c, path->btree_id)->b->c.level > i);
235 bch2_btree_path_verify_level(trans, path, i);
238 bch2_btree_path_verify_locks(path);
241 void bch2_trans_verify_paths(struct btree_trans *trans)
243 struct btree_path *path;
246 trans_for_each_path(trans, path, iter)
247 bch2_btree_path_verify(trans, path);
250 static void bch2_btree_iter_verify(struct btree_iter *iter)
252 struct btree_trans *trans = iter->trans;
254 BUG_ON(iter->btree_id >= BTREE_ID_NR);
256 BUG_ON(!!(iter->flags & BTREE_ITER_CACHED) != btree_iter_path(trans, iter)->cached);
258 BUG_ON((iter->flags & BTREE_ITER_IS_EXTENTS) &&
259 (iter->flags & BTREE_ITER_ALL_SNAPSHOTS));
261 BUG_ON(!(iter->flags & __BTREE_ITER_ALL_SNAPSHOTS) &&
262 (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) &&
263 !btree_type_has_snapshot_field(iter->btree_id));
265 if (iter->update_path)
266 bch2_btree_path_verify(trans, &trans->paths[iter->update_path]);
267 bch2_btree_path_verify(trans, btree_iter_path(trans, iter));
270 static void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter)
272 BUG_ON((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
273 !iter->pos.snapshot);
275 BUG_ON(!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS) &&
276 iter->pos.snapshot != iter->snapshot);
278 BUG_ON(bkey_lt(iter->pos, bkey_start_pos(&iter->k)) ||
279 bkey_gt(iter->pos, iter->k.p));
282 static int bch2_btree_iter_verify_ret(struct btree_iter *iter, struct bkey_s_c k)
284 struct btree_trans *trans = iter->trans;
285 struct btree_iter copy;
286 struct bkey_s_c prev;
289 if (!bch2_debug_check_iterators)
292 if (!(iter->flags & BTREE_ITER_FILTER_SNAPSHOTS))
295 if (bkey_err(k) || !k.k)
298 BUG_ON(!bch2_snapshot_is_ancestor(trans->c,
302 bch2_trans_iter_init(trans, ©, iter->btree_id, iter->pos,
303 BTREE_ITER_NOPRESERVE|
304 BTREE_ITER_ALL_SNAPSHOTS);
305 prev = bch2_btree_iter_prev(©);
309 ret = bkey_err(prev);
313 if (bkey_eq(prev.k->p, k.k->p) &&
314 bch2_snapshot_is_ancestor(trans->c, iter->snapshot,
315 prev.k->p.snapshot) > 0) {
316 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
318 bch2_bkey_to_text(&buf1, k.k);
319 bch2_bkey_to_text(&buf2, prev.k);
321 panic("iter snap %u\n"
328 bch2_trans_iter_exit(trans, ©);
332 void bch2_assert_pos_locked(struct btree_trans *trans, enum btree_id id,
333 struct bpos pos, bool key_cache)
335 struct btree_path *path;
336 struct trans_for_each_path_inorder_iter iter;
337 struct printbuf buf = PRINTBUF;
339 btree_trans_sort_paths(trans);
341 trans_for_each_path_inorder(trans, path, iter) {
342 int cmp = cmp_int(path->btree_id, id) ?:
343 cmp_int(path->cached, key_cache);
350 if (!btree_node_locked(path, 0) ||
351 !path->should_be_locked)
355 if (bkey_ge(pos, path->l[0].b->data->min_key) &&
356 bkey_le(pos, path->l[0].b->key.k.p))
359 if (bkey_eq(pos, path->pos))
364 bch2_dump_trans_paths_updates(trans);
365 bch2_bpos_to_text(&buf, pos);
367 panic("not locked: %s %s%s\n",
368 bch2_btree_id_str(id), buf.buf,
369 key_cache ? " cached" : "");
374 static inline void bch2_btree_path_verify_level(struct btree_trans *trans,
375 struct btree_path *path, unsigned l) {}
376 static inline void bch2_btree_path_verify(struct btree_trans *trans,
377 struct btree_path *path) {}
378 static inline void bch2_btree_iter_verify(struct btree_iter *iter) {}
379 static inline void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter) {}
380 static inline int bch2_btree_iter_verify_ret(struct btree_iter *iter, struct bkey_s_c k) { return 0; }
384 /* Btree path: fixups after btree updates */
386 static void btree_node_iter_set_set_pos(struct btree_node_iter *iter,
389 struct bkey_packed *k)
391 struct btree_node_iter_set *set;
393 btree_node_iter_for_each(iter, set)
394 if (set->end == t->end_offset) {
395 set->k = __btree_node_key_to_offset(b, k);
396 bch2_btree_node_iter_sort(iter, b);
400 bch2_btree_node_iter_push(iter, b, k, btree_bkey_last(b, t));
403 static void __bch2_btree_path_fix_key_modified(struct btree_path *path,
405 struct bkey_packed *where)
407 struct btree_path_level *l = &path->l[b->c.level];
409 if (where != bch2_btree_node_iter_peek_all(&l->iter, l->b))
412 if (bkey_iter_pos_cmp(l->b, where, &path->pos) < 0)
413 bch2_btree_node_iter_advance(&l->iter, l->b);
416 void bch2_btree_path_fix_key_modified(struct btree_trans *trans,
418 struct bkey_packed *where)
420 struct btree_path *path;
423 trans_for_each_path_with_node(trans, b, path, i) {
424 __bch2_btree_path_fix_key_modified(path, b, where);
425 bch2_btree_path_verify_level(trans, path, b->c.level);
429 static void __bch2_btree_node_iter_fix(struct btree_path *path,
431 struct btree_node_iter *node_iter,
433 struct bkey_packed *where,
434 unsigned clobber_u64s,
437 const struct bkey_packed *end = btree_bkey_last(b, t);
438 struct btree_node_iter_set *set;
439 unsigned offset = __btree_node_key_to_offset(b, where);
440 int shift = new_u64s - clobber_u64s;
441 unsigned old_end = t->end_offset - shift;
442 unsigned orig_iter_pos = node_iter->data[0].k;
443 bool iter_current_key_modified =
444 orig_iter_pos >= offset &&
445 orig_iter_pos <= offset + clobber_u64s;
447 btree_node_iter_for_each(node_iter, set)
448 if (set->end == old_end)
451 /* didn't find the bset in the iterator - might have to readd it: */
453 bkey_iter_pos_cmp(b, where, &path->pos) >= 0) {
454 bch2_btree_node_iter_push(node_iter, b, where, end);
457 /* Iterator is after key that changed */
461 set->end = t->end_offset;
463 /* Iterator hasn't gotten to the key that changed yet: */
468 bkey_iter_pos_cmp(b, where, &path->pos) >= 0) {
470 } else if (set->k < offset + clobber_u64s) {
471 set->k = offset + new_u64s;
472 if (set->k == set->end)
473 bch2_btree_node_iter_set_drop(node_iter, set);
475 /* Iterator is after key that changed */
476 set->k = (int) set->k + shift;
480 bch2_btree_node_iter_sort(node_iter, b);
482 if (node_iter->data[0].k != orig_iter_pos)
483 iter_current_key_modified = true;
486 * When a new key is added, and the node iterator now points to that
487 * key, the iterator might have skipped past deleted keys that should
488 * come after the key the iterator now points to. We have to rewind to
489 * before those deleted keys - otherwise
490 * bch2_btree_node_iter_prev_all() breaks:
492 if (!bch2_btree_node_iter_end(node_iter) &&
493 iter_current_key_modified &&
495 struct bkey_packed *k, *k2, *p;
497 k = bch2_btree_node_iter_peek_all(node_iter, b);
499 for_each_bset(b, t) {
500 bool set_pos = false;
502 if (node_iter->data[0].end == t->end_offset)
505 k2 = bch2_btree_node_iter_bset_pos(node_iter, b, t);
507 while ((p = bch2_bkey_prev_all(b, t, k2)) &&
508 bkey_iter_cmp(b, k, p) < 0) {
514 btree_node_iter_set_set_pos(node_iter,
520 void bch2_btree_node_iter_fix(struct btree_trans *trans,
521 struct btree_path *path,
523 struct btree_node_iter *node_iter,
524 struct bkey_packed *where,
525 unsigned clobber_u64s,
528 struct bset_tree *t = bch2_bkey_to_bset_inlined(b, where);
529 struct btree_path *linked;
532 if (node_iter != &path->l[b->c.level].iter) {
533 __bch2_btree_node_iter_fix(path, b, node_iter, t,
534 where, clobber_u64s, new_u64s);
536 if (bch2_debug_check_iterators)
537 bch2_btree_node_iter_verify(node_iter, b);
540 trans_for_each_path_with_node(trans, b, linked, i) {
541 __bch2_btree_node_iter_fix(linked, b,
542 &linked->l[b->c.level].iter, t,
543 where, clobber_u64s, new_u64s);
544 bch2_btree_path_verify_level(trans, linked, b->c.level);
548 /* Btree path level: pointer to a particular btree node and node iter */
550 static inline struct bkey_s_c __btree_iter_unpack(struct bch_fs *c,
551 struct btree_path_level *l,
553 struct bkey_packed *k)
557 * signal to bch2_btree_iter_peek_slot() that we're currently at
560 u->type = KEY_TYPE_deleted;
561 return bkey_s_c_null;
564 return bkey_disassemble(l->b, k, u);
567 static inline struct bkey_s_c btree_path_level_peek_all(struct bch_fs *c,
568 struct btree_path_level *l,
571 return __btree_iter_unpack(c, l, u,
572 bch2_btree_node_iter_peek_all(&l->iter, l->b));
575 static inline struct bkey_s_c btree_path_level_peek(struct btree_trans *trans,
576 struct btree_path *path,
577 struct btree_path_level *l,
580 struct bkey_s_c k = __btree_iter_unpack(trans->c, l, u,
581 bch2_btree_node_iter_peek(&l->iter, l->b));
583 path->pos = k.k ? k.k->p : l->b->key.k.p;
584 trans->paths_sorted = false;
585 bch2_btree_path_verify_level(trans, path, l - path->l);
589 static inline struct bkey_s_c btree_path_level_prev(struct btree_trans *trans,
590 struct btree_path *path,
591 struct btree_path_level *l,
594 struct bkey_s_c k = __btree_iter_unpack(trans->c, l, u,
595 bch2_btree_node_iter_prev(&l->iter, l->b));
597 path->pos = k.k ? k.k->p : l->b->data->min_key;
598 trans->paths_sorted = false;
599 bch2_btree_path_verify_level(trans, path, l - path->l);
603 static inline bool btree_path_advance_to_pos(struct btree_path *path,
604 struct btree_path_level *l,
607 struct bkey_packed *k;
610 while ((k = bch2_btree_node_iter_peek_all(&l->iter, l->b)) &&
611 bkey_iter_pos_cmp(l->b, k, &path->pos) < 0) {
612 if (max_advance > 0 && nr_advanced >= max_advance)
615 bch2_btree_node_iter_advance(&l->iter, l->b);
622 static inline void __btree_path_level_init(struct btree_path *path,
625 struct btree_path_level *l = &path->l[level];
627 bch2_btree_node_iter_init(&l->iter, l->b, &path->pos);
630 * Iterators to interior nodes should always be pointed at the first non
634 bch2_btree_node_iter_peek(&l->iter, l->b);
637 void bch2_btree_path_level_init(struct btree_trans *trans,
638 struct btree_path *path,
641 BUG_ON(path->cached);
643 EBUG_ON(!btree_path_pos_in_node(path, b));
645 path->l[b->c.level].lock_seq = six_lock_seq(&b->c.lock);
646 path->l[b->c.level].b = b;
647 __btree_path_level_init(path, b->c.level);
650 /* Btree path: fixups after btree node updates: */
652 static void bch2_trans_revalidate_updates_in_node(struct btree_trans *trans, struct btree *b)
654 struct bch_fs *c = trans->c;
655 struct btree_insert_entry *i;
657 trans_for_each_update(trans, i)
659 i->level == b->c.level &&
660 i->btree_id == b->c.btree_id &&
661 bpos_cmp(i->k->k.p, b->data->min_key) >= 0 &&
662 bpos_cmp(i->k->k.p, b->data->max_key) <= 0) {
663 i->old_v = bch2_btree_path_peek_slot(trans->paths + i->path, &i->old_k).v;
665 if (unlikely(trans->journal_replay_not_finished)) {
667 bch2_journal_keys_peek_slot(c, i->btree_id, i->level,
679 * A btree node is being replaced - update the iterator to point to the new
682 void bch2_trans_node_add(struct btree_trans *trans,
683 struct btree_path *path,
686 struct btree_path *prev;
688 BUG_ON(!btree_path_pos_in_node(path, b));
690 while ((prev = prev_btree_path(trans, path)) &&
691 btree_path_pos_in_node(prev, b))
695 path && btree_path_pos_in_node(path, b);
696 path = next_btree_path(trans, path))
697 if (path->uptodate == BTREE_ITER_UPTODATE && !path->cached) {
698 enum btree_node_locked_type t =
699 btree_lock_want(path, b->c.level);
701 if (t != BTREE_NODE_UNLOCKED) {
702 btree_node_unlock(trans, path, b->c.level);
703 six_lock_increment(&b->c.lock, (enum six_lock_type) t);
704 mark_btree_node_locked(trans, path, b->c.level, t);
707 bch2_btree_path_level_init(trans, path, b);
710 bch2_trans_revalidate_updates_in_node(trans, b);
714 * A btree node has been modified in such a way as to invalidate iterators - fix
717 void bch2_trans_node_reinit_iter(struct btree_trans *trans, struct btree *b)
719 struct btree_path *path;
722 trans_for_each_path_with_node(trans, b, path, i)
723 __btree_path_level_init(path, b->c.level);
725 bch2_trans_revalidate_updates_in_node(trans, b);
728 /* Btree path: traverse, set_pos: */
730 static inline int btree_path_lock_root(struct btree_trans *trans,
731 struct btree_path *path,
733 unsigned long trace_ip)
735 struct bch_fs *c = trans->c;
736 struct btree *b, **rootp = &bch2_btree_id_root(c, path->btree_id)->b;
737 enum six_lock_type lock_type;
741 EBUG_ON(path->nodes_locked);
744 b = READ_ONCE(*rootp);
745 path->level = READ_ONCE(b->c.level);
747 if (unlikely(path->level < depth_want)) {
749 * the root is at a lower depth than the depth we want:
750 * got to the end of the btree, or we're walking nodes
751 * greater than some depth and there are no nodes >=
754 path->level = depth_want;
755 for (i = path->level; i < BTREE_MAX_DEPTH; i++)
760 lock_type = __btree_lock_want(path, path->level);
761 ret = btree_node_lock(trans, path, &b->c,
762 path->level, lock_type, trace_ip);
764 if (bch2_err_matches(ret, BCH_ERR_lock_fail_root_changed))
766 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
771 if (likely(b == READ_ONCE(*rootp) &&
772 b->c.level == path->level &&
774 for (i = 0; i < path->level; i++)
775 path->l[i].b = ERR_PTR(-BCH_ERR_no_btree_node_lock_root);
776 path->l[path->level].b = b;
777 for (i = path->level + 1; i < BTREE_MAX_DEPTH; i++)
780 mark_btree_node_locked(trans, path, path->level,
781 (enum btree_node_locked_type) lock_type);
782 bch2_btree_path_level_init(trans, path, b);
786 six_unlock_type(&b->c.lock, lock_type);
791 static int btree_path_prefetch(struct btree_trans *trans, struct btree_path *path)
793 struct bch_fs *c = trans->c;
794 struct btree_path_level *l = path_l(path);
795 struct btree_node_iter node_iter = l->iter;
796 struct bkey_packed *k;
798 unsigned nr = test_bit(BCH_FS_started, &c->flags)
799 ? (path->level > 1 ? 0 : 2)
800 : (path->level > 1 ? 1 : 16);
801 bool was_locked = btree_node_locked(path, path->level);
804 bch2_bkey_buf_init(&tmp);
806 while (nr-- && !ret) {
807 if (!bch2_btree_node_relock(trans, path, path->level))
810 bch2_btree_node_iter_advance(&node_iter, l->b);
811 k = bch2_btree_node_iter_peek(&node_iter, l->b);
815 bch2_bkey_buf_unpack(&tmp, c, l->b, k);
816 ret = bch2_btree_node_prefetch(trans, path, tmp.k, path->btree_id,
821 btree_node_unlock(trans, path, path->level);
823 bch2_bkey_buf_exit(&tmp, c);
827 static int btree_path_prefetch_j(struct btree_trans *trans, struct btree_path *path,
828 struct btree_and_journal_iter *jiter)
830 struct bch_fs *c = trans->c;
833 unsigned nr = test_bit(BCH_FS_started, &c->flags)
834 ? (path->level > 1 ? 0 : 2)
835 : (path->level > 1 ? 1 : 16);
836 bool was_locked = btree_node_locked(path, path->level);
839 bch2_bkey_buf_init(&tmp);
841 while (nr-- && !ret) {
842 if (!bch2_btree_node_relock(trans, path, path->level))
845 bch2_btree_and_journal_iter_advance(jiter);
846 k = bch2_btree_and_journal_iter_peek(jiter);
850 bch2_bkey_buf_reassemble(&tmp, c, k);
851 ret = bch2_btree_node_prefetch(trans, path, tmp.k, path->btree_id,
856 btree_node_unlock(trans, path, path->level);
858 bch2_bkey_buf_exit(&tmp, c);
862 static noinline void btree_node_mem_ptr_set(struct btree_trans *trans,
863 struct btree_path *path,
864 unsigned plevel, struct btree *b)
866 struct btree_path_level *l = &path->l[plevel];
867 bool locked = btree_node_locked(path, plevel);
868 struct bkey_packed *k;
869 struct bch_btree_ptr_v2 *bp;
871 if (!bch2_btree_node_relock(trans, path, plevel))
874 k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
875 BUG_ON(k->type != KEY_TYPE_btree_ptr_v2);
877 bp = (void *) bkeyp_val(&l->b->format, k);
878 bp->mem_ptr = (unsigned long)b;
881 btree_node_unlock(trans, path, plevel);
884 static noinline int btree_node_iter_and_journal_peek(struct btree_trans *trans,
885 struct btree_path *path,
887 struct bkey_buf *out)
889 struct bch_fs *c = trans->c;
890 struct btree_path_level *l = path_l(path);
891 struct btree_and_journal_iter jiter;
895 __bch2_btree_and_journal_iter_init_node_iter(&jiter, c, l->b, l->iter, path->pos);
897 k = bch2_btree_and_journal_iter_peek(&jiter);
899 bch2_bkey_buf_reassemble(out, c, k);
901 if (flags & BTREE_ITER_PREFETCH)
902 ret = btree_path_prefetch_j(trans, path, &jiter);
904 bch2_btree_and_journal_iter_exit(&jiter);
908 static __always_inline int btree_path_down(struct btree_trans *trans,
909 struct btree_path *path,
911 unsigned long trace_ip)
913 struct bch_fs *c = trans->c;
914 struct btree_path_level *l = path_l(path);
916 unsigned level = path->level - 1;
917 enum six_lock_type lock_type = __btree_lock_want(path, level);
921 EBUG_ON(!btree_node_locked(path, path->level));
923 bch2_bkey_buf_init(&tmp);
925 if (unlikely(trans->journal_replay_not_finished)) {
926 ret = btree_node_iter_and_journal_peek(trans, path, flags, &tmp);
930 bch2_bkey_buf_unpack(&tmp, c, l->b,
931 bch2_btree_node_iter_peek(&l->iter, l->b));
933 if (flags & BTREE_ITER_PREFETCH) {
934 ret = btree_path_prefetch(trans, path);
940 b = bch2_btree_node_get(trans, path, tmp.k, level, lock_type, trace_ip);
941 ret = PTR_ERR_OR_ZERO(b);
945 if (likely(!trans->journal_replay_not_finished &&
946 tmp.k->k.type == KEY_TYPE_btree_ptr_v2) &&
947 unlikely(b != btree_node_mem_ptr(tmp.k)))
948 btree_node_mem_ptr_set(trans, path, level + 1, b);
950 if (btree_node_read_locked(path, level + 1))
951 btree_node_unlock(trans, path, level + 1);
953 mark_btree_node_locked(trans, path, level,
954 (enum btree_node_locked_type) lock_type);
956 bch2_btree_path_level_init(trans, path, b);
958 bch2_btree_path_verify_locks(path);
960 bch2_bkey_buf_exit(&tmp, c);
965 static int bch2_btree_path_traverse_all(struct btree_trans *trans)
967 struct bch_fs *c = trans->c;
968 struct btree_path *path;
969 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;
979 trans->last_restarted_ip = 0;
981 trans_for_each_path(trans, path, i)
982 path->should_be_locked = false;
984 btree_trans_sort_paths(trans);
986 bch2_trans_unlock(trans);
989 if (unlikely(trans->memory_allocation_failure)) {
992 closure_init_stack(&cl);
995 ret = bch2_btree_cache_cannibalize_lock(trans, &cl);
1000 /* Now, redo traversals in correct order: */
1002 while (i < trans->nr_sorted) {
1003 btree_path_idx_t idx = trans->sorted[i];
1006 * Traversing a path can cause another path to be added at about
1007 * the same position:
1009 if (trans->paths[idx].uptodate) {
1010 __btree_path_get(&trans->paths[idx], false);
1011 ret = bch2_btree_path_traverse_one(trans, idx, 0, _THIS_IP_);
1012 __btree_path_put(&trans->paths[idx], false);
1014 if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
1015 bch2_err_matches(ret, ENOMEM))
1025 * We used to assert that all paths had been traversed here
1026 * (path->uptodate < BTREE_ITER_NEED_TRAVERSE); however, since
1027 * path->should_be_locked is not set yet, we might have unlocked and
1028 * then failed to relock a path - that's fine.
1031 bch2_btree_cache_cannibalize_unlock(trans);
1033 trans->in_traverse_all = false;
1035 trace_and_count(c, trans_traverse_all, trans, trace_ip);
1039 static inline bool btree_path_check_pos_in_node(struct btree_path *path,
1040 unsigned l, int check_pos)
1042 if (check_pos < 0 && btree_path_pos_before_node(path, path->l[l].b))
1044 if (check_pos > 0 && btree_path_pos_after_node(path, path->l[l].b))
1049 static inline bool btree_path_good_node(struct btree_trans *trans,
1050 struct btree_path *path,
1051 unsigned l, int check_pos)
1053 return is_btree_node(path, l) &&
1054 bch2_btree_node_relock(trans, path, l) &&
1055 btree_path_check_pos_in_node(path, l, check_pos);
1058 static void btree_path_set_level_down(struct btree_trans *trans,
1059 struct btree_path *path,
1064 path->level = new_level;
1066 for (l = path->level + 1; l < BTREE_MAX_DEPTH; l++)
1067 if (btree_lock_want(path, l) == BTREE_NODE_UNLOCKED)
1068 btree_node_unlock(trans, path, l);
1070 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1071 bch2_btree_path_verify(trans, path);
1074 static noinline unsigned __btree_path_up_until_good_node(struct btree_trans *trans,
1075 struct btree_path *path,
1078 unsigned i, l = path->level;
1080 while (btree_path_node(path, l) &&
1081 !btree_path_good_node(trans, path, l, check_pos))
1082 __btree_path_set_level_up(trans, path, l++);
1084 /* If we need intent locks, take them too: */
1086 i < path->locks_want && btree_path_node(path, i);
1088 if (!bch2_btree_node_relock(trans, path, i)) {
1090 __btree_path_set_level_up(trans, path, l++);
1097 static inline unsigned btree_path_up_until_good_node(struct btree_trans *trans,
1098 struct btree_path *path,
1101 return likely(btree_node_locked(path, path->level) &&
1102 btree_path_check_pos_in_node(path, path->level, check_pos))
1104 : __btree_path_up_until_good_node(trans, path, check_pos);
1108 * This is the main state machine for walking down the btree - walks down to a
1111 * Returns 0 on success, -EIO on error (error reading in a btree node).
1113 * On error, caller (peek_node()/peek_key()) must return NULL; the error is
1114 * stashed in the iterator and returned from bch2_trans_exit().
1116 int bch2_btree_path_traverse_one(struct btree_trans *trans,
1117 btree_path_idx_t path_idx,
1119 unsigned long trace_ip)
1121 struct btree_path *path = &trans->paths[path_idx];
1122 unsigned depth_want = path->level;
1123 int ret = -((int) trans->restarted);
1128 if (unlikely(!trans->srcu_held))
1129 bch2_trans_srcu_lock(trans);
1132 * Ensure we obey path->should_be_locked: if it's set, we can't unlock
1133 * and re-traverse the path without a transaction restart:
1135 if (path->should_be_locked) {
1136 ret = bch2_btree_path_relock(trans, path, trace_ip);
1141 ret = bch2_btree_path_traverse_cached(trans, path, flags);
1145 path = &trans->paths[path_idx];
1147 if (unlikely(path->level >= BTREE_MAX_DEPTH))
1150 path->level = btree_path_up_until_good_node(trans, path, 0);
1152 EBUG_ON(btree_path_node(path, path->level) &&
1153 !btree_node_locked(path, path->level));
1156 * Note: path->nodes[path->level] may be temporarily NULL here - that
1157 * would indicate to other code that we got to the end of the btree,
1158 * here it indicates that relocking the root failed - it's critical that
1159 * btree_path_lock_root() comes next and that it can't fail
1161 while (path->level > depth_want) {
1162 ret = btree_path_node(path, path->level)
1163 ? btree_path_down(trans, path, flags, trace_ip)
1164 : btree_path_lock_root(trans, path, depth_want, trace_ip);
1165 if (unlikely(ret)) {
1168 * No nodes at this level - got to the end of
1175 __bch2_btree_path_unlock(trans, path);
1176 path->level = depth_want;
1177 path->l[path->level].b = ERR_PTR(ret);
1182 path->uptodate = BTREE_ITER_UPTODATE;
1184 if (bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted)
1185 panic("ret %s (%i) trans->restarted %s (%i)\n",
1186 bch2_err_str(ret), ret,
1187 bch2_err_str(trans->restarted), trans->restarted);
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 btree_path_idx_t btree_path_clone(struct btree_trans *trans, btree_path_idx_t src,
1212 btree_path_idx_t new = btree_path_alloc(trans, src);
1214 btree_path_copy(trans, trans->paths + new, trans->paths + src);
1215 __btree_path_get(trans->paths + new, intent);
1220 btree_path_idx_t __bch2_btree_path_make_mut(struct btree_trans *trans,
1221 btree_path_idx_t path, bool intent, unsigned long ip)
1223 __btree_path_put(trans->paths + path, intent);
1224 path = btree_path_clone(trans, path, intent);
1225 trans->paths[path].preserve = false;
1229 btree_path_idx_t __must_check
1230 __bch2_btree_path_set_pos(struct btree_trans *trans,
1231 btree_path_idx_t path_idx, struct bpos new_pos,
1232 bool intent, unsigned long ip)
1234 int cmp = bpos_cmp(new_pos, trans->paths[path_idx].pos);
1236 bch2_trans_verify_not_in_restart(trans);
1237 EBUG_ON(!trans->paths[path_idx].ref);
1239 path_idx = bch2_btree_path_make_mut(trans, path_idx, intent, ip);
1241 struct btree_path *path = trans->paths + path_idx;
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 unsigned 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, btree_path_idx_t path)
1319 __bch2_btree_path_unlock(trans, trans->paths + path);
1320 btree_path_list_remove(trans, trans->paths + path);
1321 __clear_bit(path, trans->paths_allocated);
1324 void bch2_path_put(struct btree_trans *trans, btree_path_idx_t path_idx, bool intent)
1326 struct btree_path *path = trans->paths + path_idx, *dup;
1328 if (!__btree_path_put(path, intent))
1331 dup = path->preserve
1332 ? have_path_at_pos(trans, path)
1333 : have_node_at_pos(trans, path);
1335 if (!dup && !(!path->preserve && !is_btree_node(path, path->level)))
1338 if (path->should_be_locked &&
1339 !trans->restarted &&
1340 (!dup || !bch2_btree_path_relock_norestart(trans, dup, _THIS_IP_)))
1344 dup->preserve |= path->preserve;
1345 dup->should_be_locked |= path->should_be_locked;
1348 __bch2_path_free(trans, path_idx);
1351 static void bch2_path_put_nokeep(struct btree_trans *trans, btree_path_idx_t path,
1354 if (!__btree_path_put(trans->paths + path, intent))
1357 __bch2_path_free(trans, path);
1360 void __noreturn bch2_trans_restart_error(struct btree_trans *trans, u32 restart_count)
1362 panic("trans->restart_count %u, should be %u, last restarted by %pS\n",
1363 trans->restart_count, restart_count,
1364 (void *) trans->last_begin_ip);
1367 void __noreturn bch2_trans_in_restart_error(struct btree_trans *trans)
1369 panic("in transaction restart: %s, last restarted by %pS\n",
1370 bch2_err_str(trans->restarted),
1371 (void *) trans->last_restarted_ip);
1375 void bch2_trans_updates_to_text(struct printbuf *buf, struct btree_trans *trans)
1377 struct btree_insert_entry *i;
1379 prt_printf(buf, "transaction updates for %s journal seq %llu",
1380 trans->fn, trans->journal_res.seq);
1382 printbuf_indent_add(buf, 2);
1384 trans_for_each_update(trans, i) {
1385 struct bkey_s_c old = { &i->old_k, i->old_v };
1387 prt_printf(buf, "update: btree=%s cached=%u %pS",
1388 bch2_btree_id_str(i->btree_id),
1390 (void *) i->ip_allocated);
1393 prt_printf(buf, " old ");
1394 bch2_bkey_val_to_text(buf, trans->c, old);
1397 prt_printf(buf, " new ");
1398 bch2_bkey_val_to_text(buf, trans->c, bkey_i_to_s_c(i->k));
1402 for (struct jset_entry *e = trans->journal_entries;
1403 e != btree_trans_journal_entries_top(trans);
1404 e = vstruct_next(e))
1405 bch2_journal_entry_to_text(buf, trans->c, e);
1407 printbuf_indent_sub(buf, 2);
1411 void bch2_dump_trans_updates(struct btree_trans *trans)
1413 struct printbuf buf = PRINTBUF;
1415 bch2_trans_updates_to_text(&buf, trans);
1416 bch2_print_string_as_lines(KERN_ERR, buf.buf);
1417 printbuf_exit(&buf);
1420 static void bch2_btree_path_to_text(struct printbuf *out, struct btree_trans *trans, btree_path_idx_t path_idx)
1422 struct btree_path *path = trans->paths + path_idx;
1424 prt_printf(out, "path: idx %2u ref %u:%u %c %c btree=%s l=%u pos ",
1425 path_idx, path->ref, path->intent_ref,
1426 path->preserve ? 'P' : ' ',
1427 path->should_be_locked ? 'S' : ' ',
1428 bch2_btree_id_str(path->btree_id),
1430 bch2_bpos_to_text(out, path->pos);
1432 prt_printf(out, " locks %u", path->nodes_locked);
1433 #ifdef TRACK_PATH_ALLOCATED
1434 prt_printf(out, " %pS", (void *) path->ip_allocated);
1439 static noinline __cold
1440 void __bch2_trans_paths_to_text(struct printbuf *out, struct btree_trans *trans,
1443 struct trans_for_each_path_inorder_iter iter;
1446 btree_trans_sort_paths(trans);
1448 trans_for_each_path_idx_inorder(trans, iter)
1449 bch2_btree_path_to_text(out, trans, iter.path_idx);
1453 void bch2_trans_paths_to_text(struct printbuf *out, struct btree_trans *trans)
1455 __bch2_trans_paths_to_text(out, trans, false);
1458 static noinline __cold
1459 void __bch2_dump_trans_paths_updates(struct btree_trans *trans, bool nosort)
1461 struct printbuf buf = PRINTBUF;
1463 __bch2_trans_paths_to_text(&buf, trans, nosort);
1464 bch2_trans_updates_to_text(&buf, trans);
1466 bch2_print_string_as_lines(KERN_ERR, buf.buf);
1467 printbuf_exit(&buf);
1471 void bch2_dump_trans_paths_updates(struct btree_trans *trans)
1473 __bch2_dump_trans_paths_updates(trans, false);
1477 static void bch2_trans_update_max_paths(struct btree_trans *trans)
1479 struct btree_transaction_stats *s = btree_trans_stats(trans);
1480 struct printbuf buf = PRINTBUF;
1481 size_t nr = bitmap_weight(trans->paths_allocated, trans->nr_paths);
1486 bch2_trans_paths_to_text(&buf, trans);
1488 if (!buf.allocation_failure) {
1489 mutex_lock(&s->lock);
1490 if (nr > s->nr_max_paths) {
1491 s->nr_max_paths = nr;
1492 swap(s->max_paths_text, buf.buf);
1494 mutex_unlock(&s->lock);
1497 printbuf_exit(&buf);
1499 trans->nr_paths_max = nr;
1503 int __bch2_btree_trans_too_many_iters(struct btree_trans *trans)
1505 if (trace_trans_restart_too_many_iters_enabled()) {
1506 struct printbuf buf = PRINTBUF;
1508 bch2_trans_paths_to_text(&buf, trans);
1509 trace_trans_restart_too_many_iters(trans, _THIS_IP_, buf.buf);
1510 printbuf_exit(&buf);
1513 count_event(trans->c, trans_restart_too_many_iters);
1515 return btree_trans_restart(trans, BCH_ERR_transaction_restart_too_many_iters);
1518 static noinline void btree_paths_realloc(struct btree_trans *trans)
1520 unsigned nr = trans->nr_paths * 2;
1522 void *p = kzalloc(BITS_TO_LONGS(nr) * sizeof(unsigned long) +
1524 sizeof(struct btree_trans_paths) +
1525 nr * sizeof(struct btree_path) +
1526 nr * sizeof(struct btree_insert_entry), GFP_KERNEL|__GFP_NOFAIL);
1528 unsigned long *paths_allocated = p;
1529 p += BITS_TO_LONGS(nr) * sizeof(unsigned long);
1530 struct btree_path *paths = p;
1531 p += nr * sizeof(struct btree_path);
1534 struct btree_insert_entry *updates = p;
1536 *trans_paths_nr(paths) = nr;
1538 memcpy(paths_allocated, trans->paths_allocated, BITS_TO_LONGS(trans->nr_paths) * sizeof(unsigned long));
1539 memcpy(sorted, trans->sorted, trans->nr_sorted);
1540 memcpy(paths, trans->paths, trans->nr_paths * sizeof(struct btree_path));
1541 memcpy(updates, trans->updates, trans->nr_paths * sizeof(struct btree_path));
1543 unsigned long *old = trans->paths_allocated;
1545 rcu_assign_pointer(trans->paths_allocated, paths_allocated);
1546 rcu_assign_pointer(trans->sorted, sorted);
1547 rcu_assign_pointer(trans->paths, paths);
1548 rcu_assign_pointer(trans->updates, updates);
1550 trans->nr_paths = nr;
1552 if (old != trans->_paths_allocated)
1553 kfree_rcu_mightsleep(trans->paths_allocated);
1556 static inline btree_path_idx_t btree_path_alloc(struct btree_trans *trans,
1557 btree_path_idx_t pos)
1559 btree_path_idx_t idx = find_first_zero_bit(trans->paths_allocated, trans->nr_paths);
1561 if (unlikely(idx == trans->nr_paths))
1562 btree_paths_realloc(trans);
1565 * Do this before marking the new path as allocated, since it won't be
1568 if (unlikely(idx > trans->nr_paths_max))
1569 bch2_trans_update_max_paths(trans);
1571 __set_bit(idx, trans->paths_allocated);
1573 struct btree_path *path = &trans->paths[idx];
1575 path->intent_ref = 0;
1576 path->nodes_locked = 0;
1578 btree_path_list_add(trans, pos, idx);
1579 trans->paths_sorted = false;
1583 btree_path_idx_t bch2_path_get(struct btree_trans *trans,
1584 enum btree_id btree_id, struct bpos pos,
1585 unsigned locks_want, unsigned level,
1586 unsigned flags, unsigned long ip)
1588 struct btree_path *path;
1589 bool cached = flags & BTREE_ITER_CACHED;
1590 bool intent = flags & BTREE_ITER_INTENT;
1591 struct trans_for_each_path_inorder_iter iter;
1592 btree_path_idx_t path_pos = 0, path_idx;
1594 bch2_trans_verify_not_in_restart(trans);
1595 bch2_trans_verify_locks(trans);
1597 btree_trans_sort_paths(trans);
1599 trans_for_each_path_inorder(trans, path, iter) {
1600 if (__btree_path_cmp(path,
1607 path_pos = iter.path_idx;
1611 trans->paths[path_pos].cached == cached &&
1612 trans->paths[path_pos].btree_id == btree_id &&
1613 trans->paths[path_pos].level == level) {
1614 __btree_path_get(trans->paths + path_pos, intent);
1615 path_idx = bch2_btree_path_set_pos(trans, path_pos, pos, intent, ip);
1616 path = trans->paths + path_idx;
1618 path_idx = btree_path_alloc(trans, path_pos);
1619 path = trans->paths + path_idx;
1621 __btree_path_get(path, intent);
1623 path->btree_id = btree_id;
1624 path->cached = cached;
1625 path->uptodate = BTREE_ITER_NEED_TRAVERSE;
1626 path->should_be_locked = false;
1627 path->level = level;
1628 path->locks_want = locks_want;
1629 path->nodes_locked = 0;
1630 for (unsigned i = 0; i < ARRAY_SIZE(path->l); i++)
1631 path->l[i].b = ERR_PTR(-BCH_ERR_no_btree_node_init);
1632 #ifdef TRACK_PATH_ALLOCATED
1633 path->ip_allocated = ip;
1635 trans->paths_sorted = false;
1638 if (!(flags & BTREE_ITER_NOPRESERVE))
1639 path->preserve = true;
1641 if (path->intent_ref)
1642 locks_want = max(locks_want, level + 1);
1645 * If the path has locks_want greater than requested, we don't downgrade
1646 * it here - on transaction restart because btree node split needs to
1647 * upgrade locks, we might be putting/getting the iterator again.
1648 * Downgrading iterators only happens via bch2_trans_downgrade(), after
1649 * a successful transaction commit.
1652 locks_want = min(locks_want, BTREE_MAX_DEPTH);
1653 if (locks_want > path->locks_want)
1654 bch2_btree_path_upgrade_noupgrade_sibs(trans, path, locks_want, NULL);
1659 struct bkey_s_c bch2_btree_path_peek_slot(struct btree_path *path, struct bkey *u)
1662 struct btree_path_level *l = path_l(path);
1663 struct bkey_packed *_k;
1666 if (unlikely(!l->b))
1667 return bkey_s_c_null;
1669 EBUG_ON(path->uptodate != BTREE_ITER_UPTODATE);
1670 EBUG_ON(!btree_node_locked(path, path->level));
1672 if (!path->cached) {
1673 _k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
1674 k = _k ? bkey_disassemble(l->b, _k, u) : bkey_s_c_null;
1676 EBUG_ON(k.k && bkey_deleted(k.k) && bpos_eq(k.k->p, path->pos));
1678 if (!k.k || !bpos_eq(path->pos, k.k->p))
1681 struct bkey_cached *ck = (void *) path->l[0].b;
1684 (path->btree_id != ck->key.btree_id ||
1685 !bkey_eq(path->pos, ck->key.pos)));
1686 if (!ck || !ck->valid)
1687 return bkey_s_c_null;
1690 k = bkey_i_to_s_c(ck->k);
1697 return (struct bkey_s_c) { u, NULL };
1700 /* Btree iterators: */
1703 __bch2_btree_iter_traverse(struct btree_iter *iter)
1705 return bch2_btree_path_traverse(iter->trans, iter->path, iter->flags);
1709 bch2_btree_iter_traverse(struct btree_iter *iter)
1711 struct btree_trans *trans = iter->trans;
1714 iter->path = bch2_btree_path_set_pos(trans, iter->path,
1715 btree_iter_search_key(iter),
1716 iter->flags & BTREE_ITER_INTENT,
1717 btree_iter_ip_allocated(iter));
1719 ret = bch2_btree_path_traverse(iter->trans, iter->path, iter->flags);
1723 btree_path_set_should_be_locked(trans->paths + iter->path);
1727 /* Iterate across nodes (leaf and interior nodes) */
1729 struct btree *bch2_btree_iter_peek_node(struct btree_iter *iter)
1731 struct btree_trans *trans = iter->trans;
1732 struct btree *b = NULL;
1735 EBUG_ON(trans->paths[iter->path].cached);
1736 bch2_btree_iter_verify(iter);
1738 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
1742 struct btree_path *path = btree_iter_path(trans, iter);
1743 b = btree_path_node(path, path->level);
1747 BUG_ON(bpos_lt(b->key.k.p, iter->pos));
1749 bkey_init(&iter->k);
1750 iter->k.p = iter->pos = b->key.k.p;
1752 iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p,
1753 iter->flags & BTREE_ITER_INTENT,
1754 btree_iter_ip_allocated(iter));
1755 btree_path_set_should_be_locked(btree_iter_path(trans, iter));
1757 bch2_btree_iter_verify_entry_exit(iter);
1758 bch2_btree_iter_verify(iter);
1766 struct btree *bch2_btree_iter_peek_node_and_restart(struct btree_iter *iter)
1770 while (b = bch2_btree_iter_peek_node(iter),
1771 bch2_err_matches(PTR_ERR_OR_ZERO(b), BCH_ERR_transaction_restart))
1772 bch2_trans_begin(iter->trans);
1777 struct btree *bch2_btree_iter_next_node(struct btree_iter *iter)
1779 struct btree_trans *trans = iter->trans;
1780 struct btree *b = NULL;
1783 EBUG_ON(trans->paths[iter->path].cached);
1784 bch2_trans_verify_not_in_restart(trans);
1785 bch2_btree_iter_verify(iter);
1787 struct btree_path *path = btree_iter_path(trans, iter);
1789 /* already at end? */
1790 if (!btree_path_node(path, path->level))
1794 if (!btree_path_node(path, path->level + 1)) {
1795 btree_path_set_level_up(trans, path);
1799 if (!bch2_btree_node_relock(trans, path, path->level + 1)) {
1800 __bch2_btree_path_unlock(trans, path);
1801 path->l[path->level].b = ERR_PTR(-BCH_ERR_no_btree_node_relock);
1802 path->l[path->level + 1].b = ERR_PTR(-BCH_ERR_no_btree_node_relock);
1803 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1804 trace_and_count(trans->c, trans_restart_relock_next_node, trans, _THIS_IP_, path);
1805 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_relock);
1809 b = btree_path_node(path, path->level + 1);
1811 if (bpos_eq(iter->pos, b->key.k.p)) {
1812 __btree_path_set_level_up(trans, path, path->level++);
1815 * Haven't gotten to the end of the parent node: go back down to
1816 * the next child node
1818 iter->path = bch2_btree_path_set_pos(trans, iter->path,
1819 bpos_successor(iter->pos),
1820 iter->flags & BTREE_ITER_INTENT,
1821 btree_iter_ip_allocated(iter));
1823 path = btree_iter_path(trans, iter);
1824 btree_path_set_level_down(trans, path, iter->min_depth);
1826 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
1830 path = btree_iter_path(trans, iter);
1831 b = path->l[path->level].b;
1834 bkey_init(&iter->k);
1835 iter->k.p = iter->pos = b->key.k.p;
1837 iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p,
1838 iter->flags & BTREE_ITER_INTENT,
1839 btree_iter_ip_allocated(iter));
1840 btree_path_set_should_be_locked(btree_iter_path(trans, iter));
1841 EBUG_ON(btree_iter_path(trans, iter)->uptodate);
1843 bch2_btree_iter_verify_entry_exit(iter);
1844 bch2_btree_iter_verify(iter);
1852 /* Iterate across keys (in leaf nodes only) */
1854 inline bool bch2_btree_iter_advance(struct btree_iter *iter)
1856 struct bpos pos = iter->k.p;
1857 bool ret = !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS
1858 ? bpos_eq(pos, SPOS_MAX)
1859 : bkey_eq(pos, SPOS_MAX));
1861 if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
1862 pos = bkey_successor(iter, pos);
1863 bch2_btree_iter_set_pos(iter, pos);
1867 inline bool bch2_btree_iter_rewind(struct btree_iter *iter)
1869 struct bpos pos = bkey_start_pos(&iter->k);
1870 bool ret = !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS
1871 ? bpos_eq(pos, POS_MIN)
1872 : bkey_eq(pos, POS_MIN));
1874 if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
1875 pos = bkey_predecessor(iter, pos);
1876 bch2_btree_iter_set_pos(iter, pos);
1881 struct bkey_i *__bch2_btree_trans_peek_updates(struct btree_iter *iter)
1883 struct btree_trans *trans = iter->trans;
1884 struct btree_insert_entry *i;
1885 struct bkey_i *ret = NULL;
1887 trans_for_each_update(trans, i) {
1888 if (i->btree_id < iter->btree_id)
1890 if (i->btree_id > iter->btree_id)
1892 if (bpos_lt(i->k->k.p, btree_iter_path(trans, iter)->pos))
1894 if (i->key_cache_already_flushed)
1896 if (!ret || bpos_lt(i->k->k.p, ret->k.p))
1903 static inline struct bkey_i *btree_trans_peek_updates(struct btree_iter *iter)
1905 return iter->flags & BTREE_ITER_WITH_UPDATES
1906 ? __bch2_btree_trans_peek_updates(iter)
1910 static struct bkey_i *bch2_btree_journal_peek(struct btree_trans *trans,
1911 struct btree_iter *iter,
1912 struct bpos end_pos)
1914 struct btree_path *path = btree_iter_path(trans, iter);
1916 return bch2_journal_keys_peek_upto(trans->c, iter->btree_id,
1920 &iter->journal_idx);
1924 struct bkey_s_c btree_trans_peek_slot_journal(struct btree_trans *trans,
1925 struct btree_iter *iter)
1927 struct btree_path *path = btree_iter_path(trans, iter);
1928 struct bkey_i *k = bch2_btree_journal_peek(trans, iter, path->pos);
1932 return bkey_i_to_s_c(k);
1934 return bkey_s_c_null;
1939 struct bkey_s_c btree_trans_peek_journal(struct btree_trans *trans,
1940 struct btree_iter *iter,
1943 struct btree_path *path = btree_iter_path(trans, iter);
1944 struct bkey_i *next_journal =
1945 bch2_btree_journal_peek(trans, iter,
1946 k.k ? k.k->p : path_l(path)->b->key.k.p);
1949 iter->k = next_journal->k;
1950 k = bkey_i_to_s_c(next_journal);
1957 * Checks btree key cache for key at iter->pos and returns it if present, or
1961 struct bkey_s_c btree_trans_peek_key_cache(struct btree_iter *iter, struct bpos pos)
1963 struct btree_trans *trans = iter->trans;
1964 struct bch_fs *c = trans->c;
1969 if ((iter->flags & BTREE_ITER_KEY_CACHE_FILL) &&
1970 bpos_eq(iter->pos, pos))
1971 return bkey_s_c_null;
1973 if (!bch2_btree_key_cache_find(c, iter->btree_id, pos))
1974 return bkey_s_c_null;
1976 if (!iter->key_cache_path)
1977 iter->key_cache_path = bch2_path_get(trans, iter->btree_id, pos,
1978 iter->flags & BTREE_ITER_INTENT, 0,
1979 iter->flags|BTREE_ITER_CACHED|
1980 BTREE_ITER_CACHED_NOFILL,
1983 iter->key_cache_path = bch2_btree_path_set_pos(trans, iter->key_cache_path, pos,
1984 iter->flags & BTREE_ITER_INTENT,
1985 btree_iter_ip_allocated(iter));
1987 ret = bch2_btree_path_traverse(trans, iter->key_cache_path,
1988 iter->flags|BTREE_ITER_CACHED) ?:
1989 bch2_btree_path_relock(trans, btree_iter_path(trans, iter), _THIS_IP_);
1991 return bkey_s_c_err(ret);
1993 btree_path_set_should_be_locked(trans->paths + iter->key_cache_path);
1995 k = bch2_btree_path_peek_slot(trans->paths + iter->key_cache_path, &u);
1996 if (k.k && !bkey_err(k)) {
2003 static struct bkey_s_c __bch2_btree_iter_peek(struct btree_iter *iter, struct bpos search_key)
2005 struct btree_trans *trans = iter->trans;
2006 struct bkey_i *next_update;
2007 struct bkey_s_c k, k2;
2010 EBUG_ON(btree_iter_path(trans, iter)->cached);
2011 bch2_btree_iter_verify(iter);
2014 struct btree_path_level *l;
2016 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2017 iter->flags & BTREE_ITER_INTENT,
2018 btree_iter_ip_allocated(iter));
2020 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2021 if (unlikely(ret)) {
2022 /* ensure that iter->k is consistent with iter->pos: */
2023 bch2_btree_iter_set_pos(iter, iter->pos);
2024 k = bkey_s_c_err(ret);
2028 struct btree_path *path = btree_iter_path(trans, iter);
2031 if (unlikely(!l->b)) {
2032 /* No btree nodes at requested level: */
2033 bch2_btree_iter_set_pos(iter, SPOS_MAX);
2038 btree_path_set_should_be_locked(path);
2040 k = btree_path_level_peek_all(trans->c, l, &iter->k);
2042 if (unlikely(iter->flags & BTREE_ITER_WITH_KEY_CACHE) &&
2044 (k2 = btree_trans_peek_key_cache(iter, k.k->p)).k) {
2048 bch2_btree_iter_set_pos(iter, iter->pos);
2053 if (unlikely(iter->flags & BTREE_ITER_WITH_JOURNAL))
2054 k = btree_trans_peek_journal(trans, iter, k);
2056 next_update = btree_trans_peek_updates(iter);
2059 bpos_le(next_update->k.p,
2060 k.k ? k.k->p : l->b->key.k.p)) {
2061 iter->k = next_update->k;
2062 k = bkey_i_to_s_c(next_update);
2065 if (k.k && bkey_deleted(k.k)) {
2067 * If we've got a whiteout, and it's after the search
2068 * key, advance the search key to the whiteout instead
2069 * of just after the whiteout - it might be a btree
2070 * whiteout, with a real key at the same position, since
2071 * in the btree deleted keys sort before non deleted.
2073 search_key = !bpos_eq(search_key, k.k->p)
2075 : bpos_successor(k.k->p);
2081 } else if (likely(!bpos_eq(l->b->key.k.p, SPOS_MAX))) {
2082 /* Advance to next leaf node: */
2083 search_key = bpos_successor(l->b->key.k.p);
2086 bch2_btree_iter_set_pos(iter, SPOS_MAX);
2092 bch2_btree_iter_verify(iter);
2098 * bch2_btree_iter_peek_upto() - returns first key greater than or equal to
2099 * iterator's current position
2100 * @iter: iterator to peek from
2101 * @end: search limit: returns keys less than or equal to @end
2103 * Returns: key if found, or an error extractable with bkey_err().
2105 struct bkey_s_c bch2_btree_iter_peek_upto(struct btree_iter *iter, struct bpos end)
2107 struct btree_trans *trans = iter->trans;
2108 struct bpos search_key = btree_iter_search_key(iter);
2110 struct bpos iter_pos;
2113 EBUG_ON((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) && bkey_eq(end, POS_MAX));
2115 if (iter->update_path) {
2116 bch2_path_put_nokeep(trans, iter->update_path,
2117 iter->flags & BTREE_ITER_INTENT);
2118 iter->update_path = 0;
2121 bch2_btree_iter_verify_entry_exit(iter);
2124 k = __bch2_btree_iter_peek(iter, search_key);
2127 if (unlikely(bkey_err(k)))
2131 * iter->pos should be mononotically increasing, and always be
2132 * equal to the key we just returned - except extents can
2133 * straddle iter->pos:
2135 if (!(iter->flags & BTREE_ITER_IS_EXTENTS))
2138 iter_pos = bkey_max(iter->pos, bkey_start_pos(k.k));
2140 if (unlikely(!(iter->flags & BTREE_ITER_IS_EXTENTS)
2141 ? bkey_gt(iter_pos, end)
2142 : bkey_ge(iter_pos, end)))
2145 if (iter->update_path &&
2146 !bkey_eq(trans->paths[iter->update_path].pos, k.k->p)) {
2147 bch2_path_put_nokeep(trans, iter->update_path,
2148 iter->flags & BTREE_ITER_INTENT);
2149 iter->update_path = 0;
2152 if ((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
2153 (iter->flags & BTREE_ITER_INTENT) &&
2154 !(iter->flags & BTREE_ITER_IS_EXTENTS) &&
2155 !iter->update_path) {
2156 struct bpos pos = k.k->p;
2158 if (pos.snapshot < iter->snapshot) {
2159 search_key = bpos_successor(k.k->p);
2163 pos.snapshot = iter->snapshot;
2166 * advance, same as on exit for iter->path, but only up
2169 __btree_path_get(trans->paths + iter->path, iter->flags & BTREE_ITER_INTENT);
2170 iter->update_path = iter->path;
2172 iter->update_path = bch2_btree_path_set_pos(trans,
2173 iter->update_path, pos,
2174 iter->flags & BTREE_ITER_INTENT,
2176 ret = bch2_btree_path_traverse(trans, iter->update_path, iter->flags);
2177 if (unlikely(ret)) {
2178 k = bkey_s_c_err(ret);
2184 * We can never have a key in a leaf node at POS_MAX, so
2185 * we don't have to check these successor() calls:
2187 if ((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
2188 !bch2_snapshot_is_ancestor(trans->c,
2191 search_key = bpos_successor(k.k->p);
2195 if (bkey_whiteout(k.k) &&
2196 !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS)) {
2197 search_key = bkey_successor(iter, k.k->p);
2204 iter->pos = iter_pos;
2206 iter->path = bch2_btree_path_set_pos(trans, iter->path, k.k->p,
2207 iter->flags & BTREE_ITER_INTENT,
2208 btree_iter_ip_allocated(iter));
2210 btree_path_set_should_be_locked(btree_iter_path(trans, iter));
2212 if (iter->update_path) {
2213 ret = bch2_btree_path_relock(trans, trans->paths + iter->update_path, _THIS_IP_);
2215 k = bkey_s_c_err(ret);
2217 btree_path_set_should_be_locked(trans->paths + iter->update_path);
2220 if (!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS))
2221 iter->pos.snapshot = iter->snapshot;
2223 ret = bch2_btree_iter_verify_ret(iter, k);
2224 if (unlikely(ret)) {
2225 bch2_btree_iter_set_pos(iter, iter->pos);
2226 k = bkey_s_c_err(ret);
2229 bch2_btree_iter_verify_entry_exit(iter);
2233 bch2_btree_iter_set_pos(iter, end);
2239 * bch2_btree_iter_next() - returns first key greater than iterator's current
2241 * @iter: iterator to peek from
2243 * Returns: key if found, or an error extractable with bkey_err().
2245 struct bkey_s_c bch2_btree_iter_next(struct btree_iter *iter)
2247 if (!bch2_btree_iter_advance(iter))
2248 return bkey_s_c_null;
2250 return bch2_btree_iter_peek(iter);
2254 * bch2_btree_iter_peek_prev() - returns first key less than or equal to
2255 * iterator's current position
2256 * @iter: iterator to peek from
2258 * Returns: key if found, or an error extractable with bkey_err().
2260 struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *iter)
2262 struct btree_trans *trans = iter->trans;
2263 struct bpos search_key = iter->pos;
2265 struct bkey saved_k;
2266 const struct bch_val *saved_v;
2267 btree_path_idx_t saved_path = 0;
2270 EBUG_ON(btree_iter_path(trans, iter)->cached ||
2271 btree_iter_path(trans, iter)->level);
2272 EBUG_ON(iter->flags & BTREE_ITER_WITH_UPDATES);
2274 if (iter->flags & BTREE_ITER_WITH_JOURNAL)
2275 return bkey_s_c_err(-EIO);
2277 bch2_btree_iter_verify(iter);
2278 bch2_btree_iter_verify_entry_exit(iter);
2280 if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2281 search_key.snapshot = U32_MAX;
2284 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2285 iter->flags & BTREE_ITER_INTENT,
2286 btree_iter_ip_allocated(iter));
2288 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2289 if (unlikely(ret)) {
2290 /* ensure that iter->k is consistent with iter->pos: */
2291 bch2_btree_iter_set_pos(iter, iter->pos);
2292 k = bkey_s_c_err(ret);
2296 struct btree_path *path = btree_iter_path(trans, iter);
2298 k = btree_path_level_peek(trans, path, &path->l[0], &iter->k);
2300 ((iter->flags & BTREE_ITER_IS_EXTENTS)
2301 ? bpos_ge(bkey_start_pos(k.k), search_key)
2302 : bpos_gt(k.k->p, search_key)))
2303 k = btree_path_level_prev(trans, path, &path->l[0], &iter->k);
2306 if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) {
2307 if (k.k->p.snapshot == iter->snapshot)
2311 * If we have a saved candidate, and we're no
2312 * longer at the same _key_ (not pos), return
2315 if (saved_path && !bkey_eq(k.k->p, saved_k.p)) {
2316 bch2_path_put_nokeep(trans, iter->path,
2317 iter->flags & BTREE_ITER_INTENT);
2318 iter->path = saved_path;
2325 if (bch2_snapshot_is_ancestor(iter->trans->c,
2329 bch2_path_put_nokeep(trans, saved_path,
2330 iter->flags & BTREE_ITER_INTENT);
2331 saved_path = btree_path_clone(trans, iter->path,
2332 iter->flags & BTREE_ITER_INTENT);
2333 path = btree_iter_path(trans, iter);
2338 search_key = bpos_predecessor(k.k->p);
2342 if (bkey_whiteout(k.k) &&
2343 !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS)) {
2344 search_key = bkey_predecessor(iter, k.k->p);
2345 if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2346 search_key.snapshot = U32_MAX;
2350 btree_path_set_should_be_locked(path);
2352 } else if (likely(!bpos_eq(path->l[0].b->data->min_key, POS_MIN))) {
2353 /* Advance to previous leaf node: */
2354 search_key = bpos_predecessor(path->l[0].b->data->min_key);
2356 /* Start of btree: */
2357 bch2_btree_iter_set_pos(iter, POS_MIN);
2363 EBUG_ON(bkey_gt(bkey_start_pos(k.k), iter->pos));
2365 /* Extents can straddle iter->pos: */
2366 if (bkey_lt(k.k->p, iter->pos))
2369 if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2370 iter->pos.snapshot = iter->snapshot;
2373 bch2_path_put_nokeep(trans, saved_path, iter->flags & BTREE_ITER_INTENT);
2375 bch2_btree_iter_verify_entry_exit(iter);
2376 bch2_btree_iter_verify(iter);
2382 * bch2_btree_iter_prev() - returns first key less than iterator's current
2384 * @iter: iterator to peek from
2386 * Returns: key if found, or an error extractable with bkey_err().
2388 struct bkey_s_c bch2_btree_iter_prev(struct btree_iter *iter)
2390 if (!bch2_btree_iter_rewind(iter))
2391 return bkey_s_c_null;
2393 return bch2_btree_iter_peek_prev(iter);
2396 struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *iter)
2398 struct btree_trans *trans = iter->trans;
2399 struct bpos search_key;
2403 bch2_btree_iter_verify(iter);
2404 bch2_btree_iter_verify_entry_exit(iter);
2405 EBUG_ON(btree_iter_path(trans, iter)->level && (iter->flags & BTREE_ITER_WITH_KEY_CACHE));
2407 /* extents can't span inode numbers: */
2408 if ((iter->flags & BTREE_ITER_IS_EXTENTS) &&
2409 unlikely(iter->pos.offset == KEY_OFFSET_MAX)) {
2410 if (iter->pos.inode == KEY_INODE_MAX)
2411 return bkey_s_c_null;
2413 bch2_btree_iter_set_pos(iter, bpos_nosnap_successor(iter->pos));
2416 search_key = btree_iter_search_key(iter);
2417 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2418 iter->flags & BTREE_ITER_INTENT,
2419 btree_iter_ip_allocated(iter));
2421 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2422 if (unlikely(ret)) {
2423 k = bkey_s_c_err(ret);
2427 if ((iter->flags & BTREE_ITER_CACHED) ||
2428 !(iter->flags & (BTREE_ITER_IS_EXTENTS|BTREE_ITER_FILTER_SNAPSHOTS))) {
2429 struct bkey_i *next_update;
2431 if ((next_update = btree_trans_peek_updates(iter)) &&
2432 bpos_eq(next_update->k.p, iter->pos)) {
2433 iter->k = next_update->k;
2434 k = bkey_i_to_s_c(next_update);
2438 if (unlikely(iter->flags & BTREE_ITER_WITH_JOURNAL) &&
2439 (k = btree_trans_peek_slot_journal(trans, iter)).k)
2442 if (unlikely(iter->flags & BTREE_ITER_WITH_KEY_CACHE) &&
2443 (k = btree_trans_peek_key_cache(iter, iter->pos)).k) {
2446 /* We're not returning a key from iter->path: */
2450 k = bch2_btree_path_peek_slot(trans->paths + iter->path, &iter->k);
2455 struct bpos end = iter->pos;
2457 if (iter->flags & BTREE_ITER_IS_EXTENTS)
2458 end.offset = U64_MAX;
2460 EBUG_ON(btree_iter_path(trans, iter)->level);
2462 if (iter->flags & BTREE_ITER_INTENT) {
2463 struct btree_iter iter2;
2465 bch2_trans_copy_iter(&iter2, iter);
2466 k = bch2_btree_iter_peek_upto(&iter2, end);
2468 if (k.k && !bkey_err(k)) {
2472 bch2_trans_iter_exit(trans, &iter2);
2474 struct bpos pos = iter->pos;
2476 k = bch2_btree_iter_peek_upto(iter, end);
2477 if (unlikely(bkey_err(k)))
2478 bch2_btree_iter_set_pos(iter, pos);
2483 if (unlikely(bkey_err(k)))
2486 next = k.k ? bkey_start_pos(k.k) : POS_MAX;
2488 if (bkey_lt(iter->pos, next)) {
2489 bkey_init(&iter->k);
2490 iter->k.p = iter->pos;
2492 if (iter->flags & BTREE_ITER_IS_EXTENTS) {
2493 bch2_key_resize(&iter->k,
2494 min_t(u64, KEY_SIZE_MAX,
2495 (next.inode == iter->pos.inode
2499 EBUG_ON(!iter->k.size);
2502 k = (struct bkey_s_c) { &iter->k, NULL };
2506 btree_path_set_should_be_locked(btree_iter_path(trans, iter));
2508 bch2_btree_iter_verify_entry_exit(iter);
2509 bch2_btree_iter_verify(iter);
2510 ret = bch2_btree_iter_verify_ret(iter, k);
2512 return bkey_s_c_err(ret);
2517 struct bkey_s_c bch2_btree_iter_next_slot(struct btree_iter *iter)
2519 if (!bch2_btree_iter_advance(iter))
2520 return bkey_s_c_null;
2522 return bch2_btree_iter_peek_slot(iter);
2525 struct bkey_s_c bch2_btree_iter_prev_slot(struct btree_iter *iter)
2527 if (!bch2_btree_iter_rewind(iter))
2528 return bkey_s_c_null;
2530 return bch2_btree_iter_peek_slot(iter);
2533 struct bkey_s_c bch2_btree_iter_peek_and_restart_outlined(struct btree_iter *iter)
2537 while (btree_trans_too_many_iters(iter->trans) ||
2538 (k = bch2_btree_iter_peek_type(iter, iter->flags),
2539 bch2_err_matches(bkey_err(k), BCH_ERR_transaction_restart)))
2540 bch2_trans_begin(iter->trans);
2545 /* new transactional stuff: */
2547 #ifdef CONFIG_BCACHEFS_DEBUG
2548 static void btree_trans_verify_sorted_refs(struct btree_trans *trans)
2550 struct btree_path *path;
2553 BUG_ON(trans->nr_sorted != bitmap_weight(trans->paths_allocated, trans->nr_paths) - 1);
2555 trans_for_each_path(trans, path, i) {
2556 BUG_ON(path->sorted_idx >= trans->nr_sorted);
2557 BUG_ON(trans->sorted[path->sorted_idx] != i);
2560 for (i = 0; i < trans->nr_sorted; i++) {
2561 unsigned idx = trans->sorted[i];
2563 BUG_ON(!test_bit(idx, trans->paths_allocated));
2564 BUG_ON(trans->paths[idx].sorted_idx != i);
2568 static void btree_trans_verify_sorted(struct btree_trans *trans)
2570 struct btree_path *path, *prev = NULL;
2571 struct trans_for_each_path_inorder_iter iter;
2573 if (!bch2_debug_check_iterators)
2576 trans_for_each_path_inorder(trans, path, iter) {
2577 if (prev && btree_path_cmp(prev, path) > 0) {
2578 __bch2_dump_trans_paths_updates(trans, true);
2579 panic("trans paths out of order!\n");
2585 static inline void btree_trans_verify_sorted_refs(struct btree_trans *trans) {}
2586 static inline void btree_trans_verify_sorted(struct btree_trans *trans) {}
2589 void __bch2_btree_trans_sort_paths(struct btree_trans *trans)
2591 int i, l = 0, r = trans->nr_sorted, inc = 1;
2594 btree_trans_verify_sorted_refs(trans);
2596 if (trans->paths_sorted)
2600 * Cocktail shaker sort: this is efficient because iterators will be
2606 for (i = inc > 0 ? l : r - 2;
2607 i + 1 < r && i >= l;
2609 if (btree_path_cmp(trans->paths + trans->sorted[i],
2610 trans->paths + trans->sorted[i + 1]) > 0) {
2611 swap(trans->sorted[i], trans->sorted[i + 1]);
2612 trans->paths[trans->sorted[i]].sorted_idx = i;
2613 trans->paths[trans->sorted[i + 1]].sorted_idx = i + 1;
2625 trans->paths_sorted = true;
2627 btree_trans_verify_sorted(trans);
2630 static inline void btree_path_list_remove(struct btree_trans *trans,
2631 struct btree_path *path)
2635 EBUG_ON(path->sorted_idx >= trans->nr_sorted);
2636 #ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
2638 memmove_u64s_down_small(trans->sorted + path->sorted_idx,
2639 trans->sorted + path->sorted_idx + 1,
2640 DIV_ROUND_UP(trans->nr_sorted - path->sorted_idx, 8));
2642 array_remove_item(trans->sorted, trans->nr_sorted, path->sorted_idx);
2644 for (i = path->sorted_idx; i < trans->nr_sorted; i++)
2645 trans->paths[trans->sorted[i]].sorted_idx = i;
2647 path->sorted_idx = U8_MAX;
2650 static inline void btree_path_list_add(struct btree_trans *trans,
2651 btree_path_idx_t pos,
2652 btree_path_idx_t path_idx)
2654 struct btree_path *path = trans->paths + path_idx;
2657 path->sorted_idx = pos ? trans->paths[pos].sorted_idx + 1 : trans->nr_sorted;
2659 #ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
2660 memmove_u64s_up_small(trans->sorted + path->sorted_idx + 1,
2661 trans->sorted + path->sorted_idx,
2662 DIV_ROUND_UP(trans->nr_sorted - path->sorted_idx, 8));
2664 trans->sorted[path->sorted_idx] = path_idx;
2666 array_insert_item(trans->sorted, trans->nr_sorted, path->sorted_idx, path_idx);
2669 for (i = path->sorted_idx; i < trans->nr_sorted; i++)
2670 trans->paths[trans->sorted[i]].sorted_idx = i;
2672 btree_trans_verify_sorted_refs(trans);
2675 void bch2_trans_iter_exit(struct btree_trans *trans, struct btree_iter *iter)
2677 if (iter->update_path)
2678 bch2_path_put_nokeep(trans, iter->update_path,
2679 iter->flags & BTREE_ITER_INTENT);
2681 bch2_path_put(trans, iter->path,
2682 iter->flags & BTREE_ITER_INTENT);
2683 if (iter->key_cache_path)
2684 bch2_path_put(trans, iter->key_cache_path,
2685 iter->flags & BTREE_ITER_INTENT);
2687 iter->update_path = 0;
2688 iter->key_cache_path = 0;
2692 void bch2_trans_iter_init_outlined(struct btree_trans *trans,
2693 struct btree_iter *iter,
2694 enum btree_id btree_id, struct bpos pos,
2697 bch2_trans_iter_init_common(trans, iter, btree_id, pos, 0, 0,
2698 bch2_btree_iter_flags(trans, btree_id, flags),
2702 void bch2_trans_node_iter_init(struct btree_trans *trans,
2703 struct btree_iter *iter,
2704 enum btree_id btree_id,
2706 unsigned locks_want,
2710 flags |= BTREE_ITER_NOT_EXTENTS;
2711 flags |= __BTREE_ITER_ALL_SNAPSHOTS;
2712 flags |= BTREE_ITER_ALL_SNAPSHOTS;
2714 bch2_trans_iter_init_common(trans, iter, btree_id, pos, locks_want, depth,
2715 __bch2_btree_iter_flags(trans, btree_id, flags),
2718 iter->min_depth = depth;
2720 struct btree_path *path = btree_iter_path(trans, iter);
2721 BUG_ON(path->locks_want < min(locks_want, BTREE_MAX_DEPTH));
2722 BUG_ON(path->level != depth);
2723 BUG_ON(iter->min_depth != depth);
2726 void bch2_trans_copy_iter(struct btree_iter *dst, struct btree_iter *src)
2728 struct btree_trans *trans = src->trans;
2732 __btree_path_get(trans->paths + src->path, src->flags & BTREE_ITER_INTENT);
2733 if (src->update_path)
2734 __btree_path_get(trans->paths + src->update_path, src->flags & BTREE_ITER_INTENT);
2735 dst->key_cache_path = 0;
2738 void *__bch2_trans_kmalloc(struct btree_trans *trans, size_t size)
2740 struct bch_fs *c = trans->c;
2741 unsigned new_top = trans->mem_top + size;
2742 unsigned old_bytes = trans->mem_bytes;
2743 unsigned new_bytes = roundup_pow_of_two(new_top);
2748 WARN_ON_ONCE(new_bytes > BTREE_TRANS_MEM_MAX);
2750 struct btree_transaction_stats *s = btree_trans_stats(trans);
2752 s->max_mem = max(s->max_mem, new_bytes);
2754 new_mem = krealloc(trans->mem, new_bytes, GFP_NOWAIT|__GFP_NOWARN);
2755 if (unlikely(!new_mem)) {
2756 bch2_trans_unlock(trans);
2758 new_mem = krealloc(trans->mem, new_bytes, GFP_KERNEL);
2759 if (!new_mem && new_bytes <= BTREE_TRANS_MEM_MAX) {
2760 new_mem = mempool_alloc(&c->btree_trans_mem_pool, GFP_KERNEL);
2761 new_bytes = BTREE_TRANS_MEM_MAX;
2766 return ERR_PTR(-BCH_ERR_ENOMEM_trans_kmalloc);
2768 trans->mem = new_mem;
2769 trans->mem_bytes = new_bytes;
2771 ret = bch2_trans_relock(trans);
2773 return ERR_PTR(ret);
2776 trans->mem = new_mem;
2777 trans->mem_bytes = new_bytes;
2780 trace_and_count(c, trans_restart_mem_realloced, trans, _RET_IP_, new_bytes);
2781 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_mem_realloced));
2784 p = trans->mem + trans->mem_top;
2785 trans->mem_top += size;
2790 static inline void check_srcu_held_too_long(struct btree_trans *trans)
2792 WARN(trans->srcu_held && time_after(jiffies, trans->srcu_lock_time + HZ * 10),
2793 "btree trans held srcu lock (delaying memory reclaim) for %lu seconds",
2794 (jiffies - trans->srcu_lock_time) / HZ);
2797 void bch2_trans_srcu_unlock(struct btree_trans *trans)
2799 if (trans->srcu_held) {
2800 struct bch_fs *c = trans->c;
2801 struct btree_path *path;
2804 trans_for_each_path(trans, path, i)
2805 if (path->cached && !btree_node_locked(path, 0))
2806 path->l[0].b = ERR_PTR(-BCH_ERR_no_btree_node_srcu_reset);
2808 check_srcu_held_too_long(trans);
2809 srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx);
2810 trans->srcu_held = false;
2814 static void bch2_trans_srcu_lock(struct btree_trans *trans)
2816 if (!trans->srcu_held) {
2817 trans->srcu_idx = srcu_read_lock(&trans->c->btree_trans_barrier);
2818 trans->srcu_lock_time = jiffies;
2819 trans->srcu_held = true;
2824 * bch2_trans_begin() - reset a transaction after a interrupted attempt
2825 * @trans: transaction to reset
2827 * Returns: current restart counter, to be used with trans_was_restarted()
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;
2839 bch2_trans_reset_updates(trans);
2841 trans->restart_count++;
2843 trans->journal_entries = NULL;
2845 trans_for_each_path(trans, path, i) {
2846 path->should_be_locked = false;
2849 * If the transaction wasn't restarted, we're presuming to be
2850 * doing something new: dont keep iterators excpt the ones that
2851 * are in use - except for the subvolumes btree:
2853 if (!trans->restarted && path->btree_id != BTREE_ID_subvolumes)
2854 path->preserve = false;
2857 * XXX: we probably shouldn't be doing this if the transaction
2858 * was restarted, but currently we still overflow transaction
2859 * iterators if we do that
2861 if (!path->ref && !path->preserve)
2862 __bch2_path_free(trans, i);
2864 path->preserve = false;
2867 now = local_clock();
2868 if (!trans->restarted &&
2870 now - trans->last_begin_time > BTREE_TRANS_MAX_LOCK_HOLD_TIME_NS)) {
2871 drop_locks_do(trans, (cond_resched(), 0));
2872 now = local_clock();
2874 trans->last_begin_time = now;
2876 if (unlikely(trans->srcu_held &&
2877 time_after(jiffies, trans->srcu_lock_time + msecs_to_jiffies(10))))
2878 bch2_trans_srcu_unlock(trans);
2880 trans->last_begin_ip = _RET_IP_;
2881 if (trans->restarted) {
2882 bch2_btree_path_traverse_all(trans);
2883 trans->notrace_relock_fail = false;
2886 return trans->restart_count;
2889 const char *bch2_btree_transaction_fns[BCH_TRANSACTIONS_NR];
2891 unsigned bch2_trans_get_fn_idx(const char *fn)
2895 for (i = 0; i < ARRAY_SIZE(bch2_btree_transaction_fns); i++)
2896 if (!bch2_btree_transaction_fns[i] ||
2897 bch2_btree_transaction_fns[i] == fn) {
2898 bch2_btree_transaction_fns[i] = fn;
2902 pr_warn_once("BCH_TRANSACTIONS_NR not big enough!");
2906 struct btree_trans *__bch2_trans_get(struct bch_fs *c, unsigned fn_idx)
2907 __acquires(&c->btree_trans_barrier)
2909 struct btree_trans *trans;
2911 if (IS_ENABLED(__KERNEL__)) {
2912 trans = this_cpu_xchg(c->btree_trans_bufs->trans, NULL);
2914 memset(trans, 0, offsetof(struct btree_trans, list));
2919 trans = mempool_alloc(&c->btree_trans_pool, GFP_NOFS);
2920 memset(trans, 0, sizeof(*trans));
2921 closure_init_stack(&trans->ref);
2923 seqmutex_lock(&c->btree_trans_lock);
2924 if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG)) {
2925 struct btree_trans *pos;
2926 pid_t pid = current->pid;
2928 trans->locking_wait.task = current;
2930 list_for_each_entry(pos, &c->btree_trans_list, list) {
2931 struct task_struct *pos_task = READ_ONCE(pos->locking_wait.task);
2933 * We'd much prefer to be stricter here and completely
2934 * disallow multiple btree_trans in the same thread -
2935 * but the data move path calls bch2_write when we
2936 * already have a btree_trans initialized.
2939 pid == pos_task->pid &&
2940 bch2_trans_locked(pos));
2942 if (pos_task && pid < pos_task->pid) {
2943 list_add_tail(&trans->list, &pos->list);
2948 list_add_tail(&trans->list, &c->btree_trans_list);
2950 seqmutex_unlock(&c->btree_trans_lock);
2953 trans->last_begin_time = local_clock();
2954 trans->fn_idx = fn_idx;
2955 trans->locking_wait.task = current;
2956 trans->journal_replay_not_finished =
2957 unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)) &&
2958 atomic_inc_not_zero(&c->journal_keys.ref);
2959 trans->nr_paths = ARRAY_SIZE(trans->_paths);
2960 trans->paths_allocated = trans->_paths_allocated;
2961 trans->sorted = trans->_sorted;
2962 trans->paths = trans->_paths;
2963 trans->updates = trans->_updates;
2965 *trans_paths_nr(trans->paths) = BTREE_ITER_MAX;
2967 trans->paths_allocated[0] = 1;
2969 if (fn_idx < BCH_TRANSACTIONS_NR) {
2970 trans->fn = bch2_btree_transaction_fns[fn_idx];
2972 struct btree_transaction_stats *s = &c->btree_transaction_stats[fn_idx];
2975 unsigned expected_mem_bytes = roundup_pow_of_two(s->max_mem);
2977 trans->mem = kmalloc(expected_mem_bytes, GFP_KERNEL);
2978 if (likely(trans->mem))
2979 trans->mem_bytes = expected_mem_bytes;
2982 trans->nr_paths_max = s->nr_max_paths;
2983 trans->journal_entries_size = s->journal_entries_size;
2986 trans->srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
2987 trans->srcu_lock_time = jiffies;
2988 trans->srcu_held = true;
2992 static void check_btree_paths_leaked(struct btree_trans *trans)
2994 #ifdef CONFIG_BCACHEFS_DEBUG
2995 struct bch_fs *c = trans->c;
2996 struct btree_path *path;
2999 trans_for_each_path(trans, path, i)
3004 bch_err(c, "btree paths leaked from %s!", trans->fn);
3005 trans_for_each_path(trans, path, i)
3007 printk(KERN_ERR " btree %s %pS\n",
3008 bch2_btree_id_str(path->btree_id),
3009 (void *) path->ip_allocated);
3010 /* Be noisy about this: */
3011 bch2_fatal_error(c);
3015 void bch2_trans_put(struct btree_trans *trans)
3016 __releases(&c->btree_trans_barrier)
3018 struct btree_insert_entry *i;
3019 struct bch_fs *c = trans->c;
3021 bch2_trans_unlock(trans);
3023 trans_for_each_update(trans, i)
3024 __btree_path_put(trans->paths + i->path, true);
3025 trans->nr_updates = 0;
3026 trans->locking_wait.task = NULL;
3028 check_btree_paths_leaked(trans);
3030 if (trans->srcu_held) {
3031 check_srcu_held_too_long(trans);
3032 srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx);
3035 if (trans->fs_usage_deltas) {
3036 if (trans->fs_usage_deltas->size + sizeof(trans->fs_usage_deltas) ==
3037 REPLICAS_DELTA_LIST_MAX)
3038 mempool_free(trans->fs_usage_deltas,
3039 &c->replicas_delta_pool);
3041 kfree(trans->fs_usage_deltas);
3044 if (unlikely(trans->journal_replay_not_finished))
3045 bch2_journal_keys_put(c);
3047 unsigned long *paths_allocated = trans->paths_allocated;
3048 trans->paths_allocated = NULL;
3049 trans->paths = NULL;
3051 if (paths_allocated != trans->_paths_allocated)
3052 kfree_rcu_mightsleep(paths_allocated);
3054 if (trans->mem_bytes == BTREE_TRANS_MEM_MAX)
3055 mempool_free(trans->mem, &c->btree_trans_mem_pool);
3059 /* Userspace doesn't have a real percpu implementation: */
3060 if (IS_ENABLED(__KERNEL__))
3061 trans = this_cpu_xchg(c->btree_trans_bufs->trans, trans);
3064 closure_sync(&trans->ref);
3066 seqmutex_lock(&c->btree_trans_lock);
3067 list_del(&trans->list);
3068 seqmutex_unlock(&c->btree_trans_lock);
3070 mempool_free(trans, &c->btree_trans_pool);
3074 static void __maybe_unused
3075 bch2_btree_bkey_cached_common_to_text(struct printbuf *out,
3076 struct btree_bkey_cached_common *b)
3078 struct six_lock_count c = six_lock_counts(&b->lock);
3079 struct task_struct *owner;
3083 owner = READ_ONCE(b->lock.owner);
3084 pid = owner ? owner->pid : 0;
3088 prt_printf(out, "%px %c l=%u %s:", b, b->cached ? 'c' : 'b',
3089 b->level, bch2_btree_id_str(b->btree_id));
3090 bch2_bpos_to_text(out, btree_node_pos(b));
3093 prt_printf(out, " locks %u:%u:%u held by pid %u",
3094 c.n[0], c.n[1], c.n[2], pid);
3097 void bch2_btree_trans_to_text(struct printbuf *out, struct btree_trans *trans)
3099 struct btree_bkey_cached_common *b;
3100 static char lock_types[] = { 'r', 'i', 'w' };
3101 struct task_struct *task = READ_ONCE(trans->locking_wait.task);
3104 /* before rcu_read_lock(): */
3105 bch2_printbuf_make_room(out, 4096);
3107 if (!out->nr_tabstops) {
3108 printbuf_tabstop_push(out, 16);
3109 printbuf_tabstop_push(out, 32);
3112 prt_printf(out, "%i %s\n", task ? task->pid : 0, trans->fn);
3114 /* trans->paths is rcu protected vs. freeing */
3118 struct btree_path *paths = rcu_dereference(trans->paths);
3122 unsigned long *paths_allocated = trans_paths_allocated(paths);
3124 trans_for_each_path_idx_from(paths_allocated, *trans_paths_nr(paths), idx, 1) {
3125 struct btree_path *path = paths + idx;
3126 if (!path->nodes_locked)
3129 prt_printf(out, " path %u %c l=%u %s:",
3131 path->cached ? 'c' : 'b',
3133 bch2_btree_id_str(path->btree_id));
3134 bch2_bpos_to_text(out, path->pos);
3137 for (l = 0; l < BTREE_MAX_DEPTH; l++) {
3138 if (btree_node_locked(path, l) &&
3139 !IS_ERR_OR_NULL(b = (void *) READ_ONCE(path->l[l].b))) {
3140 prt_printf(out, " %c l=%u ",
3141 lock_types[btree_node_locked_type(path, l)], l);
3142 bch2_btree_bkey_cached_common_to_text(out, b);
3148 b = READ_ONCE(trans->locking);
3150 prt_printf(out, " blocked for %lluus on",
3151 div_u64(local_clock() - trans->locking_wait.start_time,
3154 prt_printf(out, " %c", lock_types[trans->locking_wait.lock_want]);
3155 bch2_btree_bkey_cached_common_to_text(out, b);
3163 void bch2_fs_btree_iter_exit(struct bch_fs *c)
3165 struct btree_transaction_stats *s;
3166 struct btree_trans *trans;
3169 if (c->btree_trans_bufs)
3170 for_each_possible_cpu(cpu) {
3171 struct btree_trans *trans =
3172 per_cpu_ptr(c->btree_trans_bufs, cpu)->trans;
3175 closure_sync(&trans->ref);
3177 seqmutex_lock(&c->btree_trans_lock);
3178 list_del(&trans->list);
3179 seqmutex_unlock(&c->btree_trans_lock);
3183 free_percpu(c->btree_trans_bufs);
3185 trans = list_first_entry_or_null(&c->btree_trans_list, struct btree_trans, list);
3187 panic("%s leaked btree_trans\n", trans->fn);
3189 for (s = c->btree_transaction_stats;
3190 s < c->btree_transaction_stats + ARRAY_SIZE(c->btree_transaction_stats);
3192 kfree(s->max_paths_text);
3193 bch2_time_stats_exit(&s->lock_hold_times);
3196 if (c->btree_trans_barrier_initialized)
3197 cleanup_srcu_struct(&c->btree_trans_barrier);
3198 mempool_exit(&c->btree_trans_mem_pool);
3199 mempool_exit(&c->btree_trans_pool);
3202 void bch2_fs_btree_iter_init_early(struct bch_fs *c)
3204 struct btree_transaction_stats *s;
3206 for (s = c->btree_transaction_stats;
3207 s < c->btree_transaction_stats + ARRAY_SIZE(c->btree_transaction_stats);
3209 bch2_time_stats_init(&s->lock_hold_times);
3210 mutex_init(&s->lock);
3213 INIT_LIST_HEAD(&c->btree_trans_list);
3214 seqmutex_init(&c->btree_trans_lock);
3217 int bch2_fs_btree_iter_init(struct bch_fs *c)
3221 c->btree_trans_bufs = alloc_percpu(struct btree_trans_buf);
3222 if (!c->btree_trans_bufs)
3225 ret = mempool_init_kmalloc_pool(&c->btree_trans_pool, 1,
3226 sizeof(struct btree_trans)) ?:
3227 mempool_init_kmalloc_pool(&c->btree_trans_mem_pool, 1,
3228 BTREE_TRANS_MEM_MAX) ?:
3229 init_srcu_struct(&c->btree_trans_barrier);
3231 c->btree_trans_barrier_initialized = true;