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;
656 trans_for_each_update(trans, i)
658 i->level == b->c.level &&
659 i->btree_id == b->c.btree_id &&
660 bpos_cmp(i->k->k.p, b->data->min_key) >= 0 &&
661 bpos_cmp(i->k->k.p, b->data->max_key) <= 0) {
662 i->old_v = bch2_btree_path_peek_slot(trans->paths + i->path, &i->old_k).v;
664 if (unlikely(trans->journal_replay_not_finished)) {
666 bch2_journal_keys_peek_slot(c, i->btree_id, i->level,
678 * A btree node is being replaced - update the iterator to point to the new
681 void bch2_trans_node_add(struct btree_trans *trans,
682 struct btree_path *path,
685 struct btree_path *prev;
687 BUG_ON(!btree_path_pos_in_node(path, b));
689 while ((prev = prev_btree_path(trans, path)) &&
690 btree_path_pos_in_node(prev, b))
694 path && btree_path_pos_in_node(path, b);
695 path = next_btree_path(trans, path))
696 if (path->uptodate == BTREE_ITER_UPTODATE && !path->cached) {
697 enum btree_node_locked_type t =
698 btree_lock_want(path, b->c.level);
700 if (t != BTREE_NODE_UNLOCKED) {
701 btree_node_unlock(trans, path, b->c.level);
702 six_lock_increment(&b->c.lock, (enum six_lock_type) t);
703 mark_btree_node_locked(trans, path, b->c.level, t);
706 bch2_btree_path_level_init(trans, path, b);
709 bch2_trans_revalidate_updates_in_node(trans, b);
713 * A btree node has been modified in such a way as to invalidate iterators - fix
716 void bch2_trans_node_reinit_iter(struct btree_trans *trans, struct btree *b)
718 struct btree_path *path;
721 trans_for_each_path_with_node(trans, b, path, i)
722 __btree_path_level_init(path, b->c.level);
724 bch2_trans_revalidate_updates_in_node(trans, b);
727 /* Btree path: traverse, set_pos: */
729 static inline int btree_path_lock_root(struct btree_trans *trans,
730 struct btree_path *path,
732 unsigned long trace_ip)
734 struct bch_fs *c = trans->c;
735 struct btree *b, **rootp = &bch2_btree_id_root(c, path->btree_id)->b;
736 enum six_lock_type lock_type;
740 EBUG_ON(path->nodes_locked);
743 b = READ_ONCE(*rootp);
744 path->level = READ_ONCE(b->c.level);
746 if (unlikely(path->level < depth_want)) {
748 * the root is at a lower depth than the depth we want:
749 * got to the end of the btree, or we're walking nodes
750 * greater than some depth and there are no nodes >=
753 path->level = depth_want;
754 for (i = path->level; i < BTREE_MAX_DEPTH; i++)
759 lock_type = __btree_lock_want(path, path->level);
760 ret = btree_node_lock(trans, path, &b->c,
761 path->level, lock_type, trace_ip);
763 if (bch2_err_matches(ret, BCH_ERR_lock_fail_root_changed))
765 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
770 if (likely(b == READ_ONCE(*rootp) &&
771 b->c.level == path->level &&
773 for (i = 0; i < path->level; i++)
774 path->l[i].b = ERR_PTR(-BCH_ERR_no_btree_node_lock_root);
775 path->l[path->level].b = b;
776 for (i = path->level + 1; i < BTREE_MAX_DEPTH; i++)
779 mark_btree_node_locked(trans, path, path->level,
780 (enum btree_node_locked_type) lock_type);
781 bch2_btree_path_level_init(trans, path, b);
785 six_unlock_type(&b->c.lock, lock_type);
790 static int btree_path_prefetch(struct btree_trans *trans, struct btree_path *path)
792 struct bch_fs *c = trans->c;
793 struct btree_path_level *l = path_l(path);
794 struct btree_node_iter node_iter = l->iter;
795 struct bkey_packed *k;
797 unsigned nr = test_bit(BCH_FS_started, &c->flags)
798 ? (path->level > 1 ? 0 : 2)
799 : (path->level > 1 ? 1 : 16);
800 bool was_locked = btree_node_locked(path, path->level);
803 bch2_bkey_buf_init(&tmp);
805 while (nr-- && !ret) {
806 if (!bch2_btree_node_relock(trans, path, path->level))
809 bch2_btree_node_iter_advance(&node_iter, l->b);
810 k = bch2_btree_node_iter_peek(&node_iter, l->b);
814 bch2_bkey_buf_unpack(&tmp, c, l->b, k);
815 ret = bch2_btree_node_prefetch(trans, path, tmp.k, path->btree_id,
820 btree_node_unlock(trans, path, path->level);
822 bch2_bkey_buf_exit(&tmp, c);
826 static int btree_path_prefetch_j(struct btree_trans *trans, struct btree_path *path,
827 struct btree_and_journal_iter *jiter)
829 struct bch_fs *c = trans->c;
832 unsigned nr = test_bit(BCH_FS_started, &c->flags)
833 ? (path->level > 1 ? 0 : 2)
834 : (path->level > 1 ? 1 : 16);
835 bool was_locked = btree_node_locked(path, path->level);
838 bch2_bkey_buf_init(&tmp);
840 while (nr-- && !ret) {
841 if (!bch2_btree_node_relock(trans, path, path->level))
844 bch2_btree_and_journal_iter_advance(jiter);
845 k = bch2_btree_and_journal_iter_peek(jiter);
849 bch2_bkey_buf_reassemble(&tmp, c, k);
850 ret = bch2_btree_node_prefetch(trans, path, tmp.k, path->btree_id,
855 btree_node_unlock(trans, path, path->level);
857 bch2_bkey_buf_exit(&tmp, c);
861 static noinline void btree_node_mem_ptr_set(struct btree_trans *trans,
862 struct btree_path *path,
863 unsigned plevel, struct btree *b)
865 struct btree_path_level *l = &path->l[plevel];
866 bool locked = btree_node_locked(path, plevel);
867 struct bkey_packed *k;
868 struct bch_btree_ptr_v2 *bp;
870 if (!bch2_btree_node_relock(trans, path, plevel))
873 k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
874 BUG_ON(k->type != KEY_TYPE_btree_ptr_v2);
876 bp = (void *) bkeyp_val(&l->b->format, k);
877 bp->mem_ptr = (unsigned long)b;
880 btree_node_unlock(trans, path, plevel);
883 static noinline int btree_node_iter_and_journal_peek(struct btree_trans *trans,
884 struct btree_path *path,
886 struct bkey_buf *out)
888 struct bch_fs *c = trans->c;
889 struct btree_path_level *l = path_l(path);
890 struct btree_and_journal_iter jiter;
894 __bch2_btree_and_journal_iter_init_node_iter(&jiter, c, l->b, l->iter, path->pos);
896 k = bch2_btree_and_journal_iter_peek(&jiter);
898 bch2_bkey_buf_reassemble(out, c, k);
900 if (flags & BTREE_ITER_PREFETCH)
901 ret = btree_path_prefetch_j(trans, path, &jiter);
903 bch2_btree_and_journal_iter_exit(&jiter);
907 static __always_inline int btree_path_down(struct btree_trans *trans,
908 struct btree_path *path,
910 unsigned long trace_ip)
912 struct bch_fs *c = trans->c;
913 struct btree_path_level *l = path_l(path);
915 unsigned level = path->level - 1;
916 enum six_lock_type lock_type = __btree_lock_want(path, level);
920 EBUG_ON(!btree_node_locked(path, path->level));
922 bch2_bkey_buf_init(&tmp);
924 if (unlikely(trans->journal_replay_not_finished)) {
925 ret = btree_node_iter_and_journal_peek(trans, path, flags, &tmp);
929 bch2_bkey_buf_unpack(&tmp, c, l->b,
930 bch2_btree_node_iter_peek(&l->iter, l->b));
932 if (flags & BTREE_ITER_PREFETCH) {
933 ret = btree_path_prefetch(trans, path);
939 b = bch2_btree_node_get(trans, path, tmp.k, level, lock_type, trace_ip);
940 ret = PTR_ERR_OR_ZERO(b);
944 if (likely(!trans->journal_replay_not_finished &&
945 tmp.k->k.type == KEY_TYPE_btree_ptr_v2) &&
946 unlikely(b != btree_node_mem_ptr(tmp.k)))
947 btree_node_mem_ptr_set(trans, path, level + 1, b);
949 if (btree_node_read_locked(path, level + 1))
950 btree_node_unlock(trans, path, level + 1);
952 mark_btree_node_locked(trans, path, level,
953 (enum btree_node_locked_type) lock_type);
955 bch2_btree_path_level_init(trans, path, b);
957 bch2_btree_path_verify_locks(path);
959 bch2_bkey_buf_exit(&tmp, c);
964 static int bch2_btree_path_traverse_all(struct btree_trans *trans)
966 struct bch_fs *c = trans->c;
967 struct btree_path *path;
968 unsigned long trace_ip = _RET_IP_;
972 if (trans->in_traverse_all)
973 return -BCH_ERR_transaction_restart_in_traverse_all;
975 trans->in_traverse_all = true;
977 trans->restarted = 0;
978 trans->last_restarted_ip = 0;
980 trans_for_each_path(trans, path, i)
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(trans, &cl);
999 /* Now, redo traversals in correct order: */
1001 while (i < trans->nr_sorted) {
1002 btree_path_idx_t idx = trans->sorted[i];
1005 * Traversing a path can cause another path to be added at about
1006 * the same position:
1008 if (trans->paths[idx].uptodate) {
1009 __btree_path_get(&trans->paths[idx], false);
1010 ret = bch2_btree_path_traverse_one(trans, idx, 0, _THIS_IP_);
1011 __btree_path_put(&trans->paths[idx], false);
1013 if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
1014 bch2_err_matches(ret, ENOMEM))
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 might have unlocked and
1027 * then failed to relock a path - that's fine.
1030 bch2_btree_cache_cannibalize_unlock(trans);
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 btree_path_idx_t path_idx,
1118 unsigned long trace_ip)
1120 struct btree_path *path = &trans->paths[path_idx];
1121 unsigned depth_want = path->level;
1122 int ret = -((int) trans->restarted);
1127 if (unlikely(!trans->srcu_held))
1128 bch2_trans_srcu_lock(trans);
1131 * Ensure we obey path->should_be_locked: if it's set, we can't unlock
1132 * and re-traverse the path without a transaction restart:
1134 if (path->should_be_locked) {
1135 ret = bch2_btree_path_relock(trans, path, trace_ip);
1140 ret = bch2_btree_path_traverse_cached(trans, path, flags);
1144 path = &trans->paths[path_idx];
1146 if (unlikely(path->level >= BTREE_MAX_DEPTH))
1149 path->level = btree_path_up_until_good_node(trans, path, 0);
1151 EBUG_ON(btree_path_node(path, path->level) &&
1152 !btree_node_locked(path, path->level));
1155 * Note: path->nodes[path->level] may be temporarily NULL here - that
1156 * would indicate to other code that we got to the end of the btree,
1157 * here it indicates that relocking the root failed - it's critical that
1158 * btree_path_lock_root() comes next and that it can't fail
1160 while (path->level > depth_want) {
1161 ret = btree_path_node(path, path->level)
1162 ? btree_path_down(trans, path, flags, trace_ip)
1163 : btree_path_lock_root(trans, path, depth_want, trace_ip);
1164 if (unlikely(ret)) {
1167 * No nodes at this level - got to the end of
1174 __bch2_btree_path_unlock(trans, path);
1175 path->level = depth_want;
1176 path->l[path->level].b = ERR_PTR(ret);
1181 path->uptodate = BTREE_ITER_UPTODATE;
1183 if (bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted)
1184 panic("ret %s (%i) trans->restarted %s (%i)\n",
1185 bch2_err_str(ret), ret,
1186 bch2_err_str(trans->restarted), trans->restarted);
1187 bch2_btree_path_verify(trans, path);
1191 static inline void btree_path_copy(struct btree_trans *trans, struct btree_path *dst,
1192 struct btree_path *src)
1194 unsigned i, offset = offsetof(struct btree_path, pos);
1196 memcpy((void *) dst + offset,
1197 (void *) src + offset,
1198 sizeof(struct btree_path) - offset);
1200 for (i = 0; i < BTREE_MAX_DEPTH; i++) {
1201 unsigned t = btree_node_locked_type(dst, i);
1203 if (t != BTREE_NODE_UNLOCKED)
1204 six_lock_increment(&dst->l[i].b->c.lock, t);
1208 static btree_path_idx_t btree_path_clone(struct btree_trans *trans, btree_path_idx_t src,
1211 btree_path_idx_t new = btree_path_alloc(trans, src);
1212 btree_path_copy(trans, trans->paths + new, trans->paths + src);
1213 __btree_path_get(trans->paths + new, intent);
1218 btree_path_idx_t __bch2_btree_path_make_mut(struct btree_trans *trans,
1219 btree_path_idx_t path, bool intent, unsigned long ip)
1221 __btree_path_put(trans->paths + path, intent);
1222 path = btree_path_clone(trans, path, intent);
1223 trans->paths[path].preserve = false;
1227 btree_path_idx_t __must_check
1228 __bch2_btree_path_set_pos(struct btree_trans *trans,
1229 btree_path_idx_t path_idx, struct bpos new_pos,
1230 bool intent, unsigned long ip)
1232 int cmp = bpos_cmp(new_pos, trans->paths[path_idx].pos);
1234 bch2_trans_verify_not_in_restart(trans);
1235 EBUG_ON(!trans->paths[path_idx].ref);
1237 path_idx = bch2_btree_path_make_mut(trans, path_idx, intent, ip);
1239 struct btree_path *path = trans->paths + path_idx;
1240 path->pos = new_pos;
1241 trans->paths_sorted = false;
1243 if (unlikely(path->cached)) {
1244 btree_node_unlock(trans, path, 0);
1245 path->l[0].b = ERR_PTR(-BCH_ERR_no_btree_node_up);
1246 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1250 unsigned level = btree_path_up_until_good_node(trans, path, cmp);
1252 if (btree_path_node(path, level)) {
1253 struct btree_path_level *l = &path->l[level];
1255 BUG_ON(!btree_node_locked(path, level));
1257 * We might have to skip over many keys, or just a few: try
1258 * advancing the node iterator, and if we have to skip over too
1259 * many keys just reinit it (or if we're rewinding, since that
1263 !btree_path_advance_to_pos(path, l, 8))
1264 bch2_btree_node_iter_init(&l->iter, l->b, &path->pos);
1267 * Iterators to interior nodes should always be pointed at the first non
1270 if (unlikely(level))
1271 bch2_btree_node_iter_peek(&l->iter, l->b);
1274 if (unlikely(level != path->level)) {
1275 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1276 __bch2_btree_path_unlock(trans, path);
1279 bch2_btree_path_verify(trans, path);
1283 /* Btree path: main interface: */
1285 static struct btree_path *have_path_at_pos(struct btree_trans *trans, struct btree_path *path)
1287 struct btree_path *sib;
1289 sib = prev_btree_path(trans, path);
1290 if (sib && !btree_path_cmp(sib, path))
1293 sib = next_btree_path(trans, path);
1294 if (sib && !btree_path_cmp(sib, path))
1300 static struct btree_path *have_node_at_pos(struct btree_trans *trans, struct btree_path *path)
1302 struct btree_path *sib;
1304 sib = prev_btree_path(trans, path);
1305 if (sib && sib->level == path->level && path_l(sib)->b == path_l(path)->b)
1308 sib = next_btree_path(trans, path);
1309 if (sib && sib->level == path->level && path_l(sib)->b == path_l(path)->b)
1315 static inline void __bch2_path_free(struct btree_trans *trans, btree_path_idx_t path)
1317 __bch2_btree_path_unlock(trans, trans->paths + path);
1318 btree_path_list_remove(trans, trans->paths + path);
1319 __clear_bit(path, trans->paths_allocated);
1322 void bch2_path_put(struct btree_trans *trans, btree_path_idx_t path_idx, bool intent)
1324 struct btree_path *path = trans->paths + path_idx, *dup;
1326 if (!__btree_path_put(path, intent))
1329 dup = path->preserve
1330 ? have_path_at_pos(trans, path)
1331 : have_node_at_pos(trans, path);
1333 if (!dup && !(!path->preserve && !is_btree_node(path, path->level)))
1336 if (path->should_be_locked &&
1337 !trans->restarted &&
1338 (!dup || !bch2_btree_path_relock_norestart(trans, dup, _THIS_IP_)))
1342 dup->preserve |= path->preserve;
1343 dup->should_be_locked |= path->should_be_locked;
1346 __bch2_path_free(trans, path_idx);
1349 static void bch2_path_put_nokeep(struct btree_trans *trans, btree_path_idx_t path,
1352 if (!__btree_path_put(trans->paths + path, intent))
1355 __bch2_path_free(trans, path);
1358 void __noreturn bch2_trans_restart_error(struct btree_trans *trans, u32 restart_count)
1360 panic("trans->restart_count %u, should be %u, last restarted by %pS\n",
1361 trans->restart_count, restart_count,
1362 (void *) trans->last_begin_ip);
1365 void __noreturn bch2_trans_in_restart_error(struct btree_trans *trans)
1367 panic("in transaction restart: %s, last restarted by %pS\n",
1368 bch2_err_str(trans->restarted),
1369 (void *) trans->last_restarted_ip);
1373 void bch2_trans_updates_to_text(struct printbuf *buf, struct btree_trans *trans)
1375 prt_printf(buf, "transaction updates for %s journal seq %llu",
1376 trans->fn, trans->journal_res.seq);
1378 printbuf_indent_add(buf, 2);
1380 trans_for_each_update(trans, i) {
1381 struct bkey_s_c old = { &i->old_k, i->old_v };
1383 prt_printf(buf, "update: btree=%s cached=%u %pS",
1384 bch2_btree_id_str(i->btree_id),
1386 (void *) i->ip_allocated);
1389 prt_printf(buf, " old ");
1390 bch2_bkey_val_to_text(buf, trans->c, old);
1393 prt_printf(buf, " new ");
1394 bch2_bkey_val_to_text(buf, trans->c, bkey_i_to_s_c(i->k));
1398 for (struct jset_entry *e = trans->journal_entries;
1399 e != btree_trans_journal_entries_top(trans);
1400 e = vstruct_next(e))
1401 bch2_journal_entry_to_text(buf, trans->c, e);
1403 printbuf_indent_sub(buf, 2);
1407 void bch2_dump_trans_updates(struct btree_trans *trans)
1409 struct printbuf buf = PRINTBUF;
1411 bch2_trans_updates_to_text(&buf, trans);
1412 bch2_print_string_as_lines(KERN_ERR, buf.buf);
1413 printbuf_exit(&buf);
1416 static void bch2_btree_path_to_text(struct printbuf *out, struct btree_trans *trans, btree_path_idx_t path_idx)
1418 struct btree_path *path = trans->paths + path_idx;
1420 prt_printf(out, "path: idx %2u ref %u:%u %c %c btree=%s l=%u pos ",
1421 path_idx, path->ref, path->intent_ref,
1422 path->preserve ? 'P' : ' ',
1423 path->should_be_locked ? 'S' : ' ',
1424 bch2_btree_id_str(path->btree_id),
1426 bch2_bpos_to_text(out, path->pos);
1428 prt_printf(out, " locks %u", path->nodes_locked);
1429 #ifdef TRACK_PATH_ALLOCATED
1430 prt_printf(out, " %pS", (void *) path->ip_allocated);
1435 static noinline __cold
1436 void __bch2_trans_paths_to_text(struct printbuf *out, struct btree_trans *trans,
1439 struct trans_for_each_path_inorder_iter iter;
1442 btree_trans_sort_paths(trans);
1444 trans_for_each_path_idx_inorder(trans, iter)
1445 bch2_btree_path_to_text(out, trans, iter.path_idx);
1449 void bch2_trans_paths_to_text(struct printbuf *out, struct btree_trans *trans)
1451 __bch2_trans_paths_to_text(out, trans, false);
1454 static noinline __cold
1455 void __bch2_dump_trans_paths_updates(struct btree_trans *trans, bool nosort)
1457 struct printbuf buf = PRINTBUF;
1459 __bch2_trans_paths_to_text(&buf, trans, nosort);
1460 bch2_trans_updates_to_text(&buf, trans);
1462 bch2_print_string_as_lines(KERN_ERR, buf.buf);
1463 printbuf_exit(&buf);
1467 void bch2_dump_trans_paths_updates(struct btree_trans *trans)
1469 __bch2_dump_trans_paths_updates(trans, false);
1473 static void bch2_trans_update_max_paths(struct btree_trans *trans)
1475 struct btree_transaction_stats *s = btree_trans_stats(trans);
1476 struct printbuf buf = PRINTBUF;
1477 size_t nr = bitmap_weight(trans->paths_allocated, trans->nr_paths);
1482 bch2_trans_paths_to_text(&buf, trans);
1484 if (!buf.allocation_failure) {
1485 mutex_lock(&s->lock);
1486 if (nr > s->nr_max_paths) {
1487 s->nr_max_paths = nr;
1488 swap(s->max_paths_text, buf.buf);
1490 mutex_unlock(&s->lock);
1493 printbuf_exit(&buf);
1495 trans->nr_paths_max = nr;
1499 int __bch2_btree_trans_too_many_iters(struct btree_trans *trans)
1501 if (trace_trans_restart_too_many_iters_enabled()) {
1502 struct printbuf buf = PRINTBUF;
1504 bch2_trans_paths_to_text(&buf, trans);
1505 trace_trans_restart_too_many_iters(trans, _THIS_IP_, buf.buf);
1506 printbuf_exit(&buf);
1509 count_event(trans->c, trans_restart_too_many_iters);
1511 return btree_trans_restart(trans, BCH_ERR_transaction_restart_too_many_iters);
1514 static noinline void btree_path_overflow(struct btree_trans *trans)
1516 bch2_dump_trans_paths_updates(trans);
1517 bch_err(trans->c, "trans path overflow");
1520 static noinline void btree_paths_realloc(struct btree_trans *trans)
1522 unsigned nr = trans->nr_paths * 2;
1524 void *p = kzalloc(BITS_TO_LONGS(nr) * sizeof(unsigned long) +
1525 sizeof(struct btree_trans_paths) +
1526 nr * sizeof(struct btree_path) +
1527 nr * sizeof(btree_path_idx_t) + 8 +
1528 nr * sizeof(struct btree_insert_entry), GFP_KERNEL|__GFP_NOFAIL);
1530 unsigned long *paths_allocated = p;
1531 memcpy(paths_allocated, trans->paths_allocated, BITS_TO_LONGS(trans->nr_paths) * sizeof(unsigned long));
1532 p += BITS_TO_LONGS(nr) * sizeof(unsigned long);
1534 p += sizeof(struct btree_trans_paths);
1535 struct btree_path *paths = p;
1536 *trans_paths_nr(paths) = nr;
1537 memcpy(paths, trans->paths, trans->nr_paths * sizeof(struct btree_path));
1538 p += nr * sizeof(struct btree_path);
1540 btree_path_idx_t *sorted = p;
1541 memcpy(sorted, trans->sorted, trans->nr_sorted * sizeof(btree_path_idx_t));
1542 p += nr * sizeof(btree_path_idx_t) + 8;
1544 struct btree_insert_entry *updates = p;
1545 memcpy(updates, trans->updates, trans->nr_paths * sizeof(struct btree_insert_entry));
1547 unsigned long *old = trans->paths_allocated;
1549 rcu_assign_pointer(trans->paths_allocated, paths_allocated);
1550 rcu_assign_pointer(trans->paths, paths);
1551 rcu_assign_pointer(trans->sorted, sorted);
1552 rcu_assign_pointer(trans->updates, updates);
1554 trans->nr_paths = nr;
1556 if (old != trans->_paths_allocated)
1557 kfree_rcu_mightsleep(old);
1560 static inline btree_path_idx_t btree_path_alloc(struct btree_trans *trans,
1561 btree_path_idx_t pos)
1563 btree_path_idx_t idx = find_first_zero_bit(trans->paths_allocated, trans->nr_paths);
1565 if (unlikely(idx == trans->nr_paths)) {
1566 if (trans->nr_paths == BTREE_ITER_MAX) {
1567 btree_path_overflow(trans);
1571 btree_paths_realloc(trans);
1575 * Do this before marking the new path as allocated, since it won't be
1578 if (unlikely(idx > trans->nr_paths_max))
1579 bch2_trans_update_max_paths(trans);
1581 __set_bit(idx, trans->paths_allocated);
1583 struct btree_path *path = &trans->paths[idx];
1585 path->intent_ref = 0;
1586 path->nodes_locked = 0;
1588 btree_path_list_add(trans, pos, idx);
1589 trans->paths_sorted = false;
1593 btree_path_idx_t bch2_path_get(struct btree_trans *trans,
1594 enum btree_id btree_id, struct bpos pos,
1595 unsigned locks_want, unsigned level,
1596 unsigned flags, unsigned long ip)
1598 struct btree_path *path;
1599 bool cached = flags & BTREE_ITER_CACHED;
1600 bool intent = flags & BTREE_ITER_INTENT;
1601 struct trans_for_each_path_inorder_iter iter;
1602 btree_path_idx_t path_pos = 0, path_idx;
1604 bch2_trans_verify_not_in_restart(trans);
1605 bch2_trans_verify_locks(trans);
1607 btree_trans_sort_paths(trans);
1609 trans_for_each_path_inorder(trans, path, iter) {
1610 if (__btree_path_cmp(path,
1617 path_pos = iter.path_idx;
1621 trans->paths[path_pos].cached == cached &&
1622 trans->paths[path_pos].btree_id == btree_id &&
1623 trans->paths[path_pos].level == level) {
1624 __btree_path_get(trans->paths + path_pos, intent);
1625 path_idx = bch2_btree_path_set_pos(trans, path_pos, pos, intent, ip);
1626 path = trans->paths + path_idx;
1628 path_idx = btree_path_alloc(trans, path_pos);
1629 path = trans->paths + path_idx;
1631 __btree_path_get(path, intent);
1633 path->btree_id = btree_id;
1634 path->cached = cached;
1635 path->uptodate = BTREE_ITER_NEED_TRAVERSE;
1636 path->should_be_locked = false;
1637 path->level = level;
1638 path->locks_want = locks_want;
1639 path->nodes_locked = 0;
1640 for (unsigned i = 0; i < ARRAY_SIZE(path->l); i++)
1641 path->l[i].b = ERR_PTR(-BCH_ERR_no_btree_node_init);
1642 #ifdef TRACK_PATH_ALLOCATED
1643 path->ip_allocated = ip;
1645 trans->paths_sorted = false;
1648 if (!(flags & BTREE_ITER_NOPRESERVE))
1649 path->preserve = true;
1651 if (path->intent_ref)
1652 locks_want = max(locks_want, level + 1);
1655 * If the path has locks_want greater than requested, we don't downgrade
1656 * it here - on transaction restart because btree node split needs to
1657 * upgrade locks, we might be putting/getting the iterator again.
1658 * Downgrading iterators only happens via bch2_trans_downgrade(), after
1659 * a successful transaction commit.
1662 locks_want = min(locks_want, BTREE_MAX_DEPTH);
1663 if (locks_want > path->locks_want)
1664 bch2_btree_path_upgrade_noupgrade_sibs(trans, path, locks_want, NULL);
1669 struct bkey_s_c bch2_btree_path_peek_slot(struct btree_path *path, struct bkey *u)
1672 struct btree_path_level *l = path_l(path);
1673 struct bkey_packed *_k;
1676 if (unlikely(!l->b))
1677 return bkey_s_c_null;
1679 EBUG_ON(path->uptodate != BTREE_ITER_UPTODATE);
1680 EBUG_ON(!btree_node_locked(path, path->level));
1682 if (!path->cached) {
1683 _k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
1684 k = _k ? bkey_disassemble(l->b, _k, u) : bkey_s_c_null;
1686 EBUG_ON(k.k && bkey_deleted(k.k) && bpos_eq(k.k->p, path->pos));
1688 if (!k.k || !bpos_eq(path->pos, k.k->p))
1691 struct bkey_cached *ck = (void *) path->l[0].b;
1694 (path->btree_id != ck->key.btree_id ||
1695 !bkey_eq(path->pos, ck->key.pos)));
1696 if (!ck || !ck->valid)
1697 return bkey_s_c_null;
1700 k = bkey_i_to_s_c(ck->k);
1707 return (struct bkey_s_c) { u, NULL };
1710 /* Btree iterators: */
1713 __bch2_btree_iter_traverse(struct btree_iter *iter)
1715 return bch2_btree_path_traverse(iter->trans, iter->path, iter->flags);
1719 bch2_btree_iter_traverse(struct btree_iter *iter)
1721 struct btree_trans *trans = iter->trans;
1724 iter->path = bch2_btree_path_set_pos(trans, iter->path,
1725 btree_iter_search_key(iter),
1726 iter->flags & BTREE_ITER_INTENT,
1727 btree_iter_ip_allocated(iter));
1729 ret = bch2_btree_path_traverse(iter->trans, iter->path, iter->flags);
1733 btree_path_set_should_be_locked(trans->paths + iter->path);
1737 /* Iterate across nodes (leaf and interior nodes) */
1739 struct btree *bch2_btree_iter_peek_node(struct btree_iter *iter)
1741 struct btree_trans *trans = iter->trans;
1742 struct btree *b = NULL;
1745 EBUG_ON(trans->paths[iter->path].cached);
1746 bch2_btree_iter_verify(iter);
1748 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
1752 struct btree_path *path = btree_iter_path(trans, iter);
1753 b = btree_path_node(path, path->level);
1757 BUG_ON(bpos_lt(b->key.k.p, iter->pos));
1759 bkey_init(&iter->k);
1760 iter->k.p = iter->pos = b->key.k.p;
1762 iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p,
1763 iter->flags & BTREE_ITER_INTENT,
1764 btree_iter_ip_allocated(iter));
1765 btree_path_set_should_be_locked(btree_iter_path(trans, iter));
1767 bch2_btree_iter_verify_entry_exit(iter);
1768 bch2_btree_iter_verify(iter);
1776 struct btree *bch2_btree_iter_peek_node_and_restart(struct btree_iter *iter)
1780 while (b = bch2_btree_iter_peek_node(iter),
1781 bch2_err_matches(PTR_ERR_OR_ZERO(b), BCH_ERR_transaction_restart))
1782 bch2_trans_begin(iter->trans);
1787 struct btree *bch2_btree_iter_next_node(struct btree_iter *iter)
1789 struct btree_trans *trans = iter->trans;
1790 struct btree *b = NULL;
1793 EBUG_ON(trans->paths[iter->path].cached);
1794 bch2_trans_verify_not_in_restart(trans);
1795 bch2_btree_iter_verify(iter);
1797 struct btree_path *path = btree_iter_path(trans, iter);
1799 /* already at end? */
1800 if (!btree_path_node(path, path->level))
1804 if (!btree_path_node(path, path->level + 1)) {
1805 btree_path_set_level_up(trans, path);
1809 if (!bch2_btree_node_relock(trans, path, path->level + 1)) {
1810 __bch2_btree_path_unlock(trans, path);
1811 path->l[path->level].b = ERR_PTR(-BCH_ERR_no_btree_node_relock);
1812 path->l[path->level + 1].b = ERR_PTR(-BCH_ERR_no_btree_node_relock);
1813 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1814 trace_and_count(trans->c, trans_restart_relock_next_node, trans, _THIS_IP_, path);
1815 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_relock);
1819 b = btree_path_node(path, path->level + 1);
1821 if (bpos_eq(iter->pos, b->key.k.p)) {
1822 __btree_path_set_level_up(trans, path, path->level++);
1825 * Haven't gotten to the end of the parent node: go back down to
1826 * the next child node
1828 iter->path = bch2_btree_path_set_pos(trans, iter->path,
1829 bpos_successor(iter->pos),
1830 iter->flags & BTREE_ITER_INTENT,
1831 btree_iter_ip_allocated(iter));
1833 path = btree_iter_path(trans, iter);
1834 btree_path_set_level_down(trans, path, iter->min_depth);
1836 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
1840 path = btree_iter_path(trans, iter);
1841 b = path->l[path->level].b;
1844 bkey_init(&iter->k);
1845 iter->k.p = iter->pos = b->key.k.p;
1847 iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p,
1848 iter->flags & BTREE_ITER_INTENT,
1849 btree_iter_ip_allocated(iter));
1850 btree_path_set_should_be_locked(btree_iter_path(trans, iter));
1851 EBUG_ON(btree_iter_path(trans, iter)->uptodate);
1853 bch2_btree_iter_verify_entry_exit(iter);
1854 bch2_btree_iter_verify(iter);
1862 /* Iterate across keys (in leaf nodes only) */
1864 inline bool bch2_btree_iter_advance(struct btree_iter *iter)
1866 struct bpos pos = iter->k.p;
1867 bool ret = !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS
1868 ? bpos_eq(pos, SPOS_MAX)
1869 : bkey_eq(pos, SPOS_MAX));
1871 if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
1872 pos = bkey_successor(iter, pos);
1873 bch2_btree_iter_set_pos(iter, pos);
1877 inline bool bch2_btree_iter_rewind(struct btree_iter *iter)
1879 struct bpos pos = bkey_start_pos(&iter->k);
1880 bool ret = !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS
1881 ? bpos_eq(pos, POS_MIN)
1882 : bkey_eq(pos, POS_MIN));
1884 if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
1885 pos = bkey_predecessor(iter, pos);
1886 bch2_btree_iter_set_pos(iter, pos);
1891 void bch2_btree_trans_peek_prev_updates(struct btree_trans *trans, struct btree_iter *iter,
1894 struct bpos end = path_l(btree_iter_path(trans, iter))->b->data->min_key;
1896 trans_for_each_update(trans, i)
1897 if (!i->key_cache_already_flushed &&
1898 i->btree_id == iter->btree_id &&
1899 bpos_le(i->k->k.p, iter->pos) &&
1900 bpos_ge(i->k->k.p, k->k ? k->k->p : end)) {
1902 *k = bkey_i_to_s_c(i->k);
1907 void bch2_btree_trans_peek_updates(struct btree_trans *trans, struct btree_iter *iter,
1910 struct btree_path *path = btree_iter_path(trans, iter);
1911 struct bpos end = path_l(path)->b->key.k.p;
1913 trans_for_each_update(trans, i)
1914 if (!i->key_cache_already_flushed &&
1915 i->btree_id == iter->btree_id &&
1916 bpos_ge(i->k->k.p, path->pos) &&
1917 bpos_le(i->k->k.p, k->k ? k->k->p : end)) {
1919 *k = bkey_i_to_s_c(i->k);
1924 void bch2_btree_trans_peek_slot_updates(struct btree_trans *trans, struct btree_iter *iter,
1927 trans_for_each_update(trans, i)
1928 if (!i->key_cache_already_flushed &&
1929 i->btree_id == iter->btree_id &&
1930 bpos_eq(i->k->k.p, iter->pos)) {
1932 *k = bkey_i_to_s_c(i->k);
1936 static struct bkey_i *bch2_btree_journal_peek(struct btree_trans *trans,
1937 struct btree_iter *iter,
1938 struct bpos end_pos)
1940 struct btree_path *path = btree_iter_path(trans, iter);
1942 return bch2_journal_keys_peek_upto(trans->c, iter->btree_id,
1946 &iter->journal_idx);
1950 struct bkey_s_c btree_trans_peek_slot_journal(struct btree_trans *trans,
1951 struct btree_iter *iter)
1953 struct btree_path *path = btree_iter_path(trans, iter);
1954 struct bkey_i *k = bch2_btree_journal_peek(trans, iter, path->pos);
1958 return bkey_i_to_s_c(k);
1960 return bkey_s_c_null;
1965 struct bkey_s_c btree_trans_peek_journal(struct btree_trans *trans,
1966 struct btree_iter *iter,
1969 struct btree_path *path = btree_iter_path(trans, iter);
1970 struct bkey_i *next_journal =
1971 bch2_btree_journal_peek(trans, iter,
1972 k.k ? k.k->p : path_l(path)->b->key.k.p);
1975 iter->k = next_journal->k;
1976 k = bkey_i_to_s_c(next_journal);
1983 * Checks btree key cache for key at iter->pos and returns it if present, or
1987 struct bkey_s_c btree_trans_peek_key_cache(struct btree_iter *iter, struct bpos pos)
1989 struct btree_trans *trans = iter->trans;
1990 struct bch_fs *c = trans->c;
1995 if ((iter->flags & BTREE_ITER_KEY_CACHE_FILL) &&
1996 bpos_eq(iter->pos, pos))
1997 return bkey_s_c_null;
1999 if (!bch2_btree_key_cache_find(c, iter->btree_id, pos))
2000 return bkey_s_c_null;
2002 if (!iter->key_cache_path)
2003 iter->key_cache_path = bch2_path_get(trans, iter->btree_id, pos,
2004 iter->flags & BTREE_ITER_INTENT, 0,
2005 iter->flags|BTREE_ITER_CACHED|
2006 BTREE_ITER_CACHED_NOFILL,
2009 iter->key_cache_path = bch2_btree_path_set_pos(trans, iter->key_cache_path, pos,
2010 iter->flags & BTREE_ITER_INTENT,
2011 btree_iter_ip_allocated(iter));
2013 ret = bch2_btree_path_traverse(trans, iter->key_cache_path,
2014 iter->flags|BTREE_ITER_CACHED) ?:
2015 bch2_btree_path_relock(trans, btree_iter_path(trans, iter), _THIS_IP_);
2017 return bkey_s_c_err(ret);
2019 btree_path_set_should_be_locked(trans->paths + iter->key_cache_path);
2021 k = bch2_btree_path_peek_slot(trans->paths + iter->key_cache_path, &u);
2022 if (k.k && !bkey_err(k)) {
2029 static struct bkey_s_c __bch2_btree_iter_peek(struct btree_iter *iter, struct bpos search_key)
2031 struct btree_trans *trans = iter->trans;
2032 struct bkey_s_c k, k2;
2035 EBUG_ON(btree_iter_path(trans, iter)->cached);
2036 bch2_btree_iter_verify(iter);
2039 struct btree_path_level *l;
2041 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2042 iter->flags & BTREE_ITER_INTENT,
2043 btree_iter_ip_allocated(iter));
2045 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2046 if (unlikely(ret)) {
2047 /* ensure that iter->k is consistent with iter->pos: */
2048 bch2_btree_iter_set_pos(iter, iter->pos);
2049 k = bkey_s_c_err(ret);
2053 struct btree_path *path = btree_iter_path(trans, iter);
2056 if (unlikely(!l->b)) {
2057 /* No btree nodes at requested level: */
2058 bch2_btree_iter_set_pos(iter, SPOS_MAX);
2063 btree_path_set_should_be_locked(path);
2065 k = btree_path_level_peek_all(trans->c, l, &iter->k);
2067 if (unlikely(iter->flags & BTREE_ITER_WITH_KEY_CACHE) &&
2069 (k2 = btree_trans_peek_key_cache(iter, k.k->p)).k) {
2073 bch2_btree_iter_set_pos(iter, iter->pos);
2078 if (unlikely(iter->flags & BTREE_ITER_WITH_JOURNAL))
2079 k = btree_trans_peek_journal(trans, iter, k);
2081 if (unlikely((iter->flags & BTREE_ITER_WITH_UPDATES) &&
2083 bch2_btree_trans_peek_updates(trans, iter, &k);
2085 if (k.k && bkey_deleted(k.k)) {
2087 * If we've got a whiteout, and it's after the search
2088 * key, advance the search key to the whiteout instead
2089 * of just after the whiteout - it might be a btree
2090 * whiteout, with a real key at the same position, since
2091 * in the btree deleted keys sort before non deleted.
2093 search_key = !bpos_eq(search_key, k.k->p)
2095 : bpos_successor(k.k->p);
2101 } else if (likely(!bpos_eq(l->b->key.k.p, SPOS_MAX))) {
2102 /* Advance to next leaf node: */
2103 search_key = bpos_successor(l->b->key.k.p);
2106 bch2_btree_iter_set_pos(iter, SPOS_MAX);
2112 bch2_btree_iter_verify(iter);
2118 * bch2_btree_iter_peek_upto() - returns first key greater than or equal to
2119 * iterator's current position
2120 * @iter: iterator to peek from
2121 * @end: search limit: returns keys less than or equal to @end
2123 * Returns: key if found, or an error extractable with bkey_err().
2125 struct bkey_s_c bch2_btree_iter_peek_upto(struct btree_iter *iter, struct bpos end)
2127 struct btree_trans *trans = iter->trans;
2128 struct bpos search_key = btree_iter_search_key(iter);
2130 struct bpos iter_pos;
2133 EBUG_ON((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) && bkey_eq(end, POS_MAX));
2135 if (iter->update_path) {
2136 bch2_path_put_nokeep(trans, iter->update_path,
2137 iter->flags & BTREE_ITER_INTENT);
2138 iter->update_path = 0;
2141 bch2_btree_iter_verify_entry_exit(iter);
2144 k = __bch2_btree_iter_peek(iter, search_key);
2147 if (unlikely(bkey_err(k)))
2151 * iter->pos should be mononotically increasing, and always be
2152 * equal to the key we just returned - except extents can
2153 * straddle iter->pos:
2155 if (!(iter->flags & BTREE_ITER_IS_EXTENTS))
2158 iter_pos = bkey_max(iter->pos, bkey_start_pos(k.k));
2160 if (unlikely(!(iter->flags & BTREE_ITER_IS_EXTENTS)
2161 ? bkey_gt(iter_pos, end)
2162 : bkey_ge(iter_pos, end)))
2165 if (iter->update_path &&
2166 !bkey_eq(trans->paths[iter->update_path].pos, k.k->p)) {
2167 bch2_path_put_nokeep(trans, iter->update_path,
2168 iter->flags & BTREE_ITER_INTENT);
2169 iter->update_path = 0;
2172 if ((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
2173 (iter->flags & BTREE_ITER_INTENT) &&
2174 !(iter->flags & BTREE_ITER_IS_EXTENTS) &&
2175 !iter->update_path) {
2176 struct bpos pos = k.k->p;
2178 if (pos.snapshot < iter->snapshot) {
2179 search_key = bpos_successor(k.k->p);
2183 pos.snapshot = iter->snapshot;
2186 * advance, same as on exit for iter->path, but only up
2189 __btree_path_get(trans->paths + iter->path, iter->flags & BTREE_ITER_INTENT);
2190 iter->update_path = iter->path;
2192 iter->update_path = bch2_btree_path_set_pos(trans,
2193 iter->update_path, pos,
2194 iter->flags & BTREE_ITER_INTENT,
2196 ret = bch2_btree_path_traverse(trans, iter->update_path, iter->flags);
2197 if (unlikely(ret)) {
2198 k = bkey_s_c_err(ret);
2204 * We can never have a key in a leaf node at POS_MAX, so
2205 * we don't have to check these successor() calls:
2207 if ((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
2208 !bch2_snapshot_is_ancestor(trans->c,
2211 search_key = bpos_successor(k.k->p);
2215 if (bkey_whiteout(k.k) &&
2216 !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS)) {
2217 search_key = bkey_successor(iter, k.k->p);
2224 iter->pos = iter_pos;
2226 iter->path = bch2_btree_path_set_pos(trans, iter->path, k.k->p,
2227 iter->flags & BTREE_ITER_INTENT,
2228 btree_iter_ip_allocated(iter));
2230 btree_path_set_should_be_locked(btree_iter_path(trans, iter));
2232 if (iter->update_path) {
2233 ret = bch2_btree_path_relock(trans, trans->paths + iter->update_path, _THIS_IP_);
2235 k = bkey_s_c_err(ret);
2237 btree_path_set_should_be_locked(trans->paths + iter->update_path);
2240 if (!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS))
2241 iter->pos.snapshot = iter->snapshot;
2243 ret = bch2_btree_iter_verify_ret(iter, k);
2244 if (unlikely(ret)) {
2245 bch2_btree_iter_set_pos(iter, iter->pos);
2246 k = bkey_s_c_err(ret);
2249 bch2_btree_iter_verify_entry_exit(iter);
2253 bch2_btree_iter_set_pos(iter, end);
2259 * bch2_btree_iter_next() - returns first key greater than iterator's current
2261 * @iter: iterator to peek from
2263 * Returns: key if found, or an error extractable with bkey_err().
2265 struct bkey_s_c bch2_btree_iter_next(struct btree_iter *iter)
2267 if (!bch2_btree_iter_advance(iter))
2268 return bkey_s_c_null;
2270 return bch2_btree_iter_peek(iter);
2274 * bch2_btree_iter_peek_prev() - returns first key less than or equal to
2275 * iterator's current position
2276 * @iter: iterator to peek from
2278 * Returns: key if found, or an error extractable with bkey_err().
2280 struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *iter)
2282 struct btree_trans *trans = iter->trans;
2283 struct bpos search_key = iter->pos;
2285 struct bkey saved_k;
2286 const struct bch_val *saved_v;
2287 btree_path_idx_t saved_path = 0;
2290 EBUG_ON(btree_iter_path(trans, iter)->cached ||
2291 btree_iter_path(trans, iter)->level);
2293 if (iter->flags & BTREE_ITER_WITH_JOURNAL)
2294 return bkey_s_c_err(-EIO);
2296 bch2_btree_iter_verify(iter);
2297 bch2_btree_iter_verify_entry_exit(iter);
2299 if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2300 search_key.snapshot = U32_MAX;
2303 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2304 iter->flags & BTREE_ITER_INTENT,
2305 btree_iter_ip_allocated(iter));
2307 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2308 if (unlikely(ret)) {
2309 /* ensure that iter->k is consistent with iter->pos: */
2310 bch2_btree_iter_set_pos(iter, iter->pos);
2311 k = bkey_s_c_err(ret);
2315 struct btree_path *path = btree_iter_path(trans, iter);
2317 k = btree_path_level_peek(trans, path, &path->l[0], &iter->k);
2319 ((iter->flags & BTREE_ITER_IS_EXTENTS)
2320 ? bpos_ge(bkey_start_pos(k.k), search_key)
2321 : bpos_gt(k.k->p, search_key)))
2322 k = btree_path_level_prev(trans, path, &path->l[0], &iter->k);
2324 if (unlikely((iter->flags & BTREE_ITER_WITH_UPDATES) &&
2326 bch2_btree_trans_peek_prev_updates(trans, iter, &k);
2329 if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) {
2330 if (k.k->p.snapshot == iter->snapshot)
2334 * If we have a saved candidate, and we're no
2335 * longer at the same _key_ (not pos), return
2338 if (saved_path && !bkey_eq(k.k->p, saved_k.p)) {
2339 bch2_path_put_nokeep(trans, iter->path,
2340 iter->flags & BTREE_ITER_INTENT);
2341 iter->path = saved_path;
2348 if (bch2_snapshot_is_ancestor(trans->c,
2352 bch2_path_put_nokeep(trans, saved_path,
2353 iter->flags & BTREE_ITER_INTENT);
2354 saved_path = btree_path_clone(trans, iter->path,
2355 iter->flags & BTREE_ITER_INTENT);
2356 path = btree_iter_path(trans, iter);
2361 search_key = bpos_predecessor(k.k->p);
2365 if (bkey_whiteout(k.k) &&
2366 !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS)) {
2367 search_key = bkey_predecessor(iter, k.k->p);
2368 if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2369 search_key.snapshot = U32_MAX;
2373 btree_path_set_should_be_locked(path);
2375 } else if (likely(!bpos_eq(path->l[0].b->data->min_key, POS_MIN))) {
2376 /* Advance to previous leaf node: */
2377 search_key = bpos_predecessor(path->l[0].b->data->min_key);
2379 /* Start of btree: */
2380 bch2_btree_iter_set_pos(iter, POS_MIN);
2386 EBUG_ON(bkey_gt(bkey_start_pos(k.k), iter->pos));
2388 /* Extents can straddle iter->pos: */
2389 if (bkey_lt(k.k->p, iter->pos))
2392 if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2393 iter->pos.snapshot = iter->snapshot;
2396 bch2_path_put_nokeep(trans, saved_path, iter->flags & BTREE_ITER_INTENT);
2398 bch2_btree_iter_verify_entry_exit(iter);
2399 bch2_btree_iter_verify(iter);
2405 * bch2_btree_iter_prev() - returns first key less than iterator's current
2407 * @iter: iterator to peek from
2409 * Returns: key if found, or an error extractable with bkey_err().
2411 struct bkey_s_c bch2_btree_iter_prev(struct btree_iter *iter)
2413 if (!bch2_btree_iter_rewind(iter))
2414 return bkey_s_c_null;
2416 return bch2_btree_iter_peek_prev(iter);
2419 struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *iter)
2421 struct btree_trans *trans = iter->trans;
2422 struct bpos search_key;
2426 bch2_btree_iter_verify(iter);
2427 bch2_btree_iter_verify_entry_exit(iter);
2428 EBUG_ON(btree_iter_path(trans, iter)->level && (iter->flags & BTREE_ITER_WITH_KEY_CACHE));
2430 /* extents can't span inode numbers: */
2431 if ((iter->flags & BTREE_ITER_IS_EXTENTS) &&
2432 unlikely(iter->pos.offset == KEY_OFFSET_MAX)) {
2433 if (iter->pos.inode == KEY_INODE_MAX)
2434 return bkey_s_c_null;
2436 bch2_btree_iter_set_pos(iter, bpos_nosnap_successor(iter->pos));
2439 search_key = btree_iter_search_key(iter);
2440 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2441 iter->flags & BTREE_ITER_INTENT,
2442 btree_iter_ip_allocated(iter));
2444 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2445 if (unlikely(ret)) {
2446 k = bkey_s_c_err(ret);
2450 if ((iter->flags & BTREE_ITER_CACHED) ||
2451 !(iter->flags & (BTREE_ITER_IS_EXTENTS|BTREE_ITER_FILTER_SNAPSHOTS))) {
2454 if (unlikely((iter->flags & BTREE_ITER_WITH_UPDATES) &&
2455 trans->nr_updates)) {
2456 bch2_btree_trans_peek_slot_updates(trans, iter, &k);
2461 if (unlikely(iter->flags & BTREE_ITER_WITH_JOURNAL) &&
2462 (k = btree_trans_peek_slot_journal(trans, iter)).k)
2465 if (unlikely(iter->flags & BTREE_ITER_WITH_KEY_CACHE) &&
2466 (k = btree_trans_peek_key_cache(iter, iter->pos)).k) {
2469 /* We're not returning a key from iter->path: */
2473 k = bch2_btree_path_peek_slot(trans->paths + iter->path, &iter->k);
2478 struct bpos end = iter->pos;
2480 if (iter->flags & BTREE_ITER_IS_EXTENTS)
2481 end.offset = U64_MAX;
2483 EBUG_ON(btree_iter_path(trans, iter)->level);
2485 if (iter->flags & BTREE_ITER_INTENT) {
2486 struct btree_iter iter2;
2488 bch2_trans_copy_iter(&iter2, iter);
2489 k = bch2_btree_iter_peek_upto(&iter2, end);
2491 if (k.k && !bkey_err(k)) {
2495 bch2_trans_iter_exit(trans, &iter2);
2497 struct bpos pos = iter->pos;
2499 k = bch2_btree_iter_peek_upto(iter, end);
2500 if (unlikely(bkey_err(k)))
2501 bch2_btree_iter_set_pos(iter, pos);
2506 if (unlikely(bkey_err(k)))
2509 next = k.k ? bkey_start_pos(k.k) : POS_MAX;
2511 if (bkey_lt(iter->pos, next)) {
2512 bkey_init(&iter->k);
2513 iter->k.p = iter->pos;
2515 if (iter->flags & BTREE_ITER_IS_EXTENTS) {
2516 bch2_key_resize(&iter->k,
2517 min_t(u64, KEY_SIZE_MAX,
2518 (next.inode == iter->pos.inode
2522 EBUG_ON(!iter->k.size);
2525 k = (struct bkey_s_c) { &iter->k, NULL };
2529 btree_path_set_should_be_locked(btree_iter_path(trans, iter));
2531 bch2_btree_iter_verify_entry_exit(iter);
2532 bch2_btree_iter_verify(iter);
2533 ret = bch2_btree_iter_verify_ret(iter, k);
2535 return bkey_s_c_err(ret);
2540 struct bkey_s_c bch2_btree_iter_next_slot(struct btree_iter *iter)
2542 if (!bch2_btree_iter_advance(iter))
2543 return bkey_s_c_null;
2545 return bch2_btree_iter_peek_slot(iter);
2548 struct bkey_s_c bch2_btree_iter_prev_slot(struct btree_iter *iter)
2550 if (!bch2_btree_iter_rewind(iter))
2551 return bkey_s_c_null;
2553 return bch2_btree_iter_peek_slot(iter);
2556 struct bkey_s_c bch2_btree_iter_peek_and_restart_outlined(struct btree_iter *iter)
2560 while (btree_trans_too_many_iters(iter->trans) ||
2561 (k = bch2_btree_iter_peek_type(iter, iter->flags),
2562 bch2_err_matches(bkey_err(k), BCH_ERR_transaction_restart)))
2563 bch2_trans_begin(iter->trans);
2568 /* new transactional stuff: */
2570 #ifdef CONFIG_BCACHEFS_DEBUG
2571 static void btree_trans_verify_sorted_refs(struct btree_trans *trans)
2573 struct btree_path *path;
2576 BUG_ON(trans->nr_sorted != bitmap_weight(trans->paths_allocated, trans->nr_paths) - 1);
2578 trans_for_each_path(trans, path, i) {
2579 BUG_ON(path->sorted_idx >= trans->nr_sorted);
2580 BUG_ON(trans->sorted[path->sorted_idx] != i);
2583 for (i = 0; i < trans->nr_sorted; i++) {
2584 unsigned idx = trans->sorted[i];
2586 BUG_ON(!test_bit(idx, trans->paths_allocated));
2587 BUG_ON(trans->paths[idx].sorted_idx != i);
2591 static void btree_trans_verify_sorted(struct btree_trans *trans)
2593 struct btree_path *path, *prev = NULL;
2594 struct trans_for_each_path_inorder_iter iter;
2596 if (!bch2_debug_check_iterators)
2599 trans_for_each_path_inorder(trans, path, iter) {
2600 if (prev && btree_path_cmp(prev, path) > 0) {
2601 __bch2_dump_trans_paths_updates(trans, true);
2602 panic("trans paths out of order!\n");
2608 static inline void btree_trans_verify_sorted_refs(struct btree_trans *trans) {}
2609 static inline void btree_trans_verify_sorted(struct btree_trans *trans) {}
2612 void __bch2_btree_trans_sort_paths(struct btree_trans *trans)
2614 int i, l = 0, r = trans->nr_sorted, inc = 1;
2617 btree_trans_verify_sorted_refs(trans);
2619 if (trans->paths_sorted)
2623 * Cocktail shaker sort: this is efficient because iterators will be
2629 for (i = inc > 0 ? l : r - 2;
2630 i + 1 < r && i >= l;
2632 if (btree_path_cmp(trans->paths + trans->sorted[i],
2633 trans->paths + trans->sorted[i + 1]) > 0) {
2634 swap(trans->sorted[i], trans->sorted[i + 1]);
2635 trans->paths[trans->sorted[i]].sorted_idx = i;
2636 trans->paths[trans->sorted[i + 1]].sorted_idx = i + 1;
2648 trans->paths_sorted = true;
2650 btree_trans_verify_sorted(trans);
2653 static inline void btree_path_list_remove(struct btree_trans *trans,
2654 struct btree_path *path)
2656 EBUG_ON(path->sorted_idx >= trans->nr_sorted);
2657 #ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
2659 memmove_u64s_down_small(trans->sorted + path->sorted_idx,
2660 trans->sorted + path->sorted_idx + 1,
2661 DIV_ROUND_UP(trans->nr_sorted - path->sorted_idx,
2662 sizeof(u64) / sizeof(btree_path_idx_t)));
2664 array_remove_item(trans->sorted, trans->nr_sorted, path->sorted_idx);
2666 for (unsigned i = path->sorted_idx; i < trans->nr_sorted; i++)
2667 trans->paths[trans->sorted[i]].sorted_idx = i;
2670 static inline void btree_path_list_add(struct btree_trans *trans,
2671 btree_path_idx_t pos,
2672 btree_path_idx_t path_idx)
2674 struct btree_path *path = trans->paths + path_idx;
2676 path->sorted_idx = pos ? trans->paths[pos].sorted_idx + 1 : trans->nr_sorted;
2678 #ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
2679 memmove_u64s_up_small(trans->sorted + path->sorted_idx + 1,
2680 trans->sorted + path->sorted_idx,
2681 DIV_ROUND_UP(trans->nr_sorted - path->sorted_idx,
2682 sizeof(u64) / sizeof(btree_path_idx_t)));
2684 trans->sorted[path->sorted_idx] = path_idx;
2686 array_insert_item(trans->sorted, trans->nr_sorted, path->sorted_idx, path_idx);
2689 for (unsigned i = path->sorted_idx; i < trans->nr_sorted; i++)
2690 trans->paths[trans->sorted[i]].sorted_idx = i;
2692 btree_trans_verify_sorted_refs(trans);
2695 void bch2_trans_iter_exit(struct btree_trans *trans, struct btree_iter *iter)
2697 if (iter->update_path)
2698 bch2_path_put_nokeep(trans, iter->update_path,
2699 iter->flags & BTREE_ITER_INTENT);
2701 bch2_path_put(trans, iter->path,
2702 iter->flags & BTREE_ITER_INTENT);
2703 if (iter->key_cache_path)
2704 bch2_path_put(trans, iter->key_cache_path,
2705 iter->flags & BTREE_ITER_INTENT);
2707 iter->update_path = 0;
2708 iter->key_cache_path = 0;
2712 void bch2_trans_iter_init_outlined(struct btree_trans *trans,
2713 struct btree_iter *iter,
2714 enum btree_id btree_id, struct bpos pos,
2717 bch2_trans_iter_init_common(trans, iter, btree_id, pos, 0, 0,
2718 bch2_btree_iter_flags(trans, btree_id, flags),
2722 void bch2_trans_node_iter_init(struct btree_trans *trans,
2723 struct btree_iter *iter,
2724 enum btree_id btree_id,
2726 unsigned locks_want,
2730 flags |= BTREE_ITER_NOT_EXTENTS;
2731 flags |= __BTREE_ITER_ALL_SNAPSHOTS;
2732 flags |= BTREE_ITER_ALL_SNAPSHOTS;
2734 bch2_trans_iter_init_common(trans, iter, btree_id, pos, locks_want, depth,
2735 __bch2_btree_iter_flags(trans, btree_id, flags),
2738 iter->min_depth = depth;
2740 struct btree_path *path = btree_iter_path(trans, iter);
2741 BUG_ON(path->locks_want < min(locks_want, BTREE_MAX_DEPTH));
2742 BUG_ON(path->level != depth);
2743 BUG_ON(iter->min_depth != depth);
2746 void bch2_trans_copy_iter(struct btree_iter *dst, struct btree_iter *src)
2748 struct btree_trans *trans = src->trans;
2752 __btree_path_get(trans->paths + src->path, src->flags & BTREE_ITER_INTENT);
2753 if (src->update_path)
2754 __btree_path_get(trans->paths + src->update_path, src->flags & BTREE_ITER_INTENT);
2755 dst->key_cache_path = 0;
2758 void *__bch2_trans_kmalloc(struct btree_trans *trans, size_t size)
2760 struct bch_fs *c = trans->c;
2761 unsigned new_top = trans->mem_top + size;
2762 unsigned old_bytes = trans->mem_bytes;
2763 unsigned new_bytes = roundup_pow_of_two(new_top);
2768 WARN_ON_ONCE(new_bytes > BTREE_TRANS_MEM_MAX);
2770 struct btree_transaction_stats *s = btree_trans_stats(trans);
2772 s->max_mem = max(s->max_mem, new_bytes);
2774 new_mem = krealloc(trans->mem, new_bytes, GFP_NOWAIT|__GFP_NOWARN);
2775 if (unlikely(!new_mem)) {
2776 bch2_trans_unlock(trans);
2778 new_mem = krealloc(trans->mem, new_bytes, GFP_KERNEL);
2779 if (!new_mem && new_bytes <= BTREE_TRANS_MEM_MAX) {
2780 new_mem = mempool_alloc(&c->btree_trans_mem_pool, GFP_KERNEL);
2781 new_bytes = BTREE_TRANS_MEM_MAX;
2786 return ERR_PTR(-BCH_ERR_ENOMEM_trans_kmalloc);
2788 trans->mem = new_mem;
2789 trans->mem_bytes = new_bytes;
2791 ret = bch2_trans_relock(trans);
2793 return ERR_PTR(ret);
2796 trans->mem = new_mem;
2797 trans->mem_bytes = new_bytes;
2800 trace_and_count(c, trans_restart_mem_realloced, trans, _RET_IP_, new_bytes);
2801 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_mem_realloced));
2804 p = trans->mem + trans->mem_top;
2805 trans->mem_top += size;
2810 static inline void check_srcu_held_too_long(struct btree_trans *trans)
2812 WARN(trans->srcu_held && time_after(jiffies, trans->srcu_lock_time + HZ * 10),
2813 "btree trans held srcu lock (delaying memory reclaim) for %lu seconds",
2814 (jiffies - trans->srcu_lock_time) / HZ);
2817 void bch2_trans_srcu_unlock(struct btree_trans *trans)
2819 if (trans->srcu_held) {
2820 struct bch_fs *c = trans->c;
2821 struct btree_path *path;
2824 trans_for_each_path(trans, path, i)
2825 if (path->cached && !btree_node_locked(path, 0))
2826 path->l[0].b = ERR_PTR(-BCH_ERR_no_btree_node_srcu_reset);
2828 check_srcu_held_too_long(trans);
2829 srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx);
2830 trans->srcu_held = false;
2834 static void bch2_trans_srcu_lock(struct btree_trans *trans)
2836 if (!trans->srcu_held) {
2837 trans->srcu_idx = srcu_read_lock(&trans->c->btree_trans_barrier);
2838 trans->srcu_lock_time = jiffies;
2839 trans->srcu_held = true;
2844 * bch2_trans_begin() - reset a transaction after a interrupted attempt
2845 * @trans: transaction to reset
2847 * Returns: current restart counter, to be used with trans_was_restarted()
2849 * While iterating over nodes or updating nodes a attempt to lock a btree node
2850 * may return BCH_ERR_transaction_restart when the trylock fails. When this
2851 * occurs bch2_trans_begin() should be called and the transaction retried.
2853 u32 bch2_trans_begin(struct btree_trans *trans)
2855 struct btree_path *path;
2859 bch2_trans_reset_updates(trans);
2861 trans->restart_count++;
2863 trans->journal_entries = NULL;
2865 trans_for_each_path(trans, path, i) {
2866 path->should_be_locked = false;
2869 * If the transaction wasn't restarted, we're presuming to be
2870 * doing something new: dont keep iterators excpt the ones that
2871 * are in use - except for the subvolumes btree:
2873 if (!trans->restarted && path->btree_id != BTREE_ID_subvolumes)
2874 path->preserve = false;
2877 * XXX: we probably shouldn't be doing this if the transaction
2878 * was restarted, but currently we still overflow transaction
2879 * iterators if we do that
2881 if (!path->ref && !path->preserve)
2882 __bch2_path_free(trans, i);
2884 path->preserve = false;
2887 now = local_clock();
2888 if (!trans->restarted &&
2890 now - trans->last_begin_time > BTREE_TRANS_MAX_LOCK_HOLD_TIME_NS)) {
2891 drop_locks_do(trans, (cond_resched(), 0));
2892 now = local_clock();
2894 trans->last_begin_time = now;
2896 if (unlikely(trans->srcu_held &&
2897 time_after(jiffies, trans->srcu_lock_time + msecs_to_jiffies(10))))
2898 bch2_trans_srcu_unlock(trans);
2900 trans->last_begin_ip = _RET_IP_;
2901 if (trans->restarted) {
2902 bch2_btree_path_traverse_all(trans);
2903 trans->notrace_relock_fail = false;
2906 return trans->restart_count;
2909 const char *bch2_btree_transaction_fns[BCH_TRANSACTIONS_NR];
2911 unsigned bch2_trans_get_fn_idx(const char *fn)
2915 for (i = 0; i < ARRAY_SIZE(bch2_btree_transaction_fns); i++)
2916 if (!bch2_btree_transaction_fns[i] ||
2917 bch2_btree_transaction_fns[i] == fn) {
2918 bch2_btree_transaction_fns[i] = fn;
2922 pr_warn_once("BCH_TRANSACTIONS_NR not big enough!");
2926 struct btree_trans *__bch2_trans_get(struct bch_fs *c, unsigned fn_idx)
2927 __acquires(&c->btree_trans_barrier)
2929 struct btree_trans *trans;
2931 if (IS_ENABLED(__KERNEL__)) {
2932 trans = this_cpu_xchg(c->btree_trans_bufs->trans, NULL);
2934 memset(trans, 0, offsetof(struct btree_trans, list));
2939 trans = mempool_alloc(&c->btree_trans_pool, GFP_NOFS);
2940 memset(trans, 0, sizeof(*trans));
2941 closure_init_stack(&trans->ref);
2943 seqmutex_lock(&c->btree_trans_lock);
2944 if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG)) {
2945 struct btree_trans *pos;
2946 pid_t pid = current->pid;
2948 trans->locking_wait.task = current;
2950 list_for_each_entry(pos, &c->btree_trans_list, list) {
2951 struct task_struct *pos_task = READ_ONCE(pos->locking_wait.task);
2953 * We'd much prefer to be stricter here and completely
2954 * disallow multiple btree_trans in the same thread -
2955 * but the data move path calls bch2_write when we
2956 * already have a btree_trans initialized.
2959 pid == pos_task->pid &&
2960 bch2_trans_locked(pos));
2962 if (pos_task && pid < pos_task->pid) {
2963 list_add_tail(&trans->list, &pos->list);
2968 list_add_tail(&trans->list, &c->btree_trans_list);
2970 seqmutex_unlock(&c->btree_trans_lock);
2973 trans->last_begin_time = local_clock();
2974 trans->fn_idx = fn_idx;
2975 trans->locking_wait.task = current;
2976 trans->journal_replay_not_finished =
2977 unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)) &&
2978 atomic_inc_not_zero(&c->journal_keys.ref);
2979 trans->nr_paths = ARRAY_SIZE(trans->_paths);
2980 trans->paths_allocated = trans->_paths_allocated;
2981 trans->sorted = trans->_sorted;
2982 trans->paths = trans->_paths;
2983 trans->updates = trans->_updates;
2985 *trans_paths_nr(trans->paths) = BTREE_ITER_INITIAL;
2987 trans->paths_allocated[0] = 1;
2989 if (fn_idx < BCH_TRANSACTIONS_NR) {
2990 trans->fn = bch2_btree_transaction_fns[fn_idx];
2992 struct btree_transaction_stats *s = &c->btree_transaction_stats[fn_idx];
2995 unsigned expected_mem_bytes = roundup_pow_of_two(s->max_mem);
2997 trans->mem = kmalloc(expected_mem_bytes, GFP_KERNEL);
2998 if (likely(trans->mem))
2999 trans->mem_bytes = expected_mem_bytes;
3002 trans->nr_paths_max = s->nr_max_paths;
3003 trans->journal_entries_size = s->journal_entries_size;
3006 trans->srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
3007 trans->srcu_lock_time = jiffies;
3008 trans->srcu_held = true;
3012 static void check_btree_paths_leaked(struct btree_trans *trans)
3014 #ifdef CONFIG_BCACHEFS_DEBUG
3015 struct bch_fs *c = trans->c;
3016 struct btree_path *path;
3019 trans_for_each_path(trans, path, i)
3024 bch_err(c, "btree paths leaked from %s!", trans->fn);
3025 trans_for_each_path(trans, path, i)
3027 printk(KERN_ERR " btree %s %pS\n",
3028 bch2_btree_id_str(path->btree_id),
3029 (void *) path->ip_allocated);
3030 /* Be noisy about this: */
3031 bch2_fatal_error(c);
3035 void bch2_trans_put(struct btree_trans *trans)
3036 __releases(&c->btree_trans_barrier)
3038 struct bch_fs *c = trans->c;
3040 bch2_trans_unlock(trans);
3042 trans_for_each_update(trans, i)
3043 __btree_path_put(trans->paths + i->path, true);
3044 trans->nr_updates = 0;
3045 trans->locking_wait.task = NULL;
3047 check_btree_paths_leaked(trans);
3049 if (trans->srcu_held) {
3050 check_srcu_held_too_long(trans);
3051 srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx);
3054 if (trans->fs_usage_deltas) {
3055 if (trans->fs_usage_deltas->size + sizeof(trans->fs_usage_deltas) ==
3056 REPLICAS_DELTA_LIST_MAX)
3057 mempool_free(trans->fs_usage_deltas,
3058 &c->replicas_delta_pool);
3060 kfree(trans->fs_usage_deltas);
3063 if (unlikely(trans->journal_replay_not_finished))
3064 bch2_journal_keys_put(c);
3066 unsigned long *paths_allocated = trans->paths_allocated;
3067 trans->paths_allocated = NULL;
3068 trans->paths = NULL;
3070 if (paths_allocated != trans->_paths_allocated)
3071 kfree_rcu_mightsleep(paths_allocated);
3073 if (trans->mem_bytes == BTREE_TRANS_MEM_MAX)
3074 mempool_free(trans->mem, &c->btree_trans_mem_pool);
3078 /* Userspace doesn't have a real percpu implementation: */
3079 if (IS_ENABLED(__KERNEL__))
3080 trans = this_cpu_xchg(c->btree_trans_bufs->trans, trans);
3083 closure_sync(&trans->ref);
3085 seqmutex_lock(&c->btree_trans_lock);
3086 list_del(&trans->list);
3087 seqmutex_unlock(&c->btree_trans_lock);
3089 mempool_free(trans, &c->btree_trans_pool);
3093 static void __maybe_unused
3094 bch2_btree_bkey_cached_common_to_text(struct printbuf *out,
3095 struct btree_bkey_cached_common *b)
3097 struct six_lock_count c = six_lock_counts(&b->lock);
3098 struct task_struct *owner;
3102 owner = READ_ONCE(b->lock.owner);
3103 pid = owner ? owner->pid : 0;
3107 prt_printf(out, "%px %c l=%u %s:", b, b->cached ? 'c' : 'b',
3108 b->level, bch2_btree_id_str(b->btree_id));
3109 bch2_bpos_to_text(out, btree_node_pos(b));
3112 prt_printf(out, " locks %u:%u:%u held by pid %u",
3113 c.n[0], c.n[1], c.n[2], pid);
3116 void bch2_btree_trans_to_text(struct printbuf *out, struct btree_trans *trans)
3118 struct btree_bkey_cached_common *b;
3119 static char lock_types[] = { 'r', 'i', 'w' };
3120 struct task_struct *task = READ_ONCE(trans->locking_wait.task);
3123 /* before rcu_read_lock(): */
3124 bch2_printbuf_make_room(out, 4096);
3126 if (!out->nr_tabstops) {
3127 printbuf_tabstop_push(out, 16);
3128 printbuf_tabstop_push(out, 32);
3131 prt_printf(out, "%i %s\n", task ? task->pid : 0, trans->fn);
3133 /* trans->paths is rcu protected vs. freeing */
3137 struct btree_path *paths = rcu_dereference(trans->paths);
3141 unsigned long *paths_allocated = trans_paths_allocated(paths);
3143 trans_for_each_path_idx_from(paths_allocated, *trans_paths_nr(paths), idx, 1) {
3144 struct btree_path *path = paths + idx;
3145 if (!path->nodes_locked)
3148 prt_printf(out, " path %u %c l=%u %s:",
3150 path->cached ? 'c' : 'b',
3152 bch2_btree_id_str(path->btree_id));
3153 bch2_bpos_to_text(out, path->pos);
3156 for (l = 0; l < BTREE_MAX_DEPTH; l++) {
3157 if (btree_node_locked(path, l) &&
3158 !IS_ERR_OR_NULL(b = (void *) READ_ONCE(path->l[l].b))) {
3159 prt_printf(out, " %c l=%u ",
3160 lock_types[btree_node_locked_type(path, l)], l);
3161 bch2_btree_bkey_cached_common_to_text(out, b);
3167 b = READ_ONCE(trans->locking);
3169 prt_printf(out, " blocked for %lluus on",
3170 div_u64(local_clock() - trans->locking_wait.start_time,
3173 prt_printf(out, " %c", lock_types[trans->locking_wait.lock_want]);
3174 bch2_btree_bkey_cached_common_to_text(out, b);
3182 void bch2_fs_btree_iter_exit(struct bch_fs *c)
3184 struct btree_transaction_stats *s;
3185 struct btree_trans *trans;
3188 if (c->btree_trans_bufs)
3189 for_each_possible_cpu(cpu) {
3190 struct btree_trans *trans =
3191 per_cpu_ptr(c->btree_trans_bufs, cpu)->trans;
3194 closure_sync(&trans->ref);
3196 seqmutex_lock(&c->btree_trans_lock);
3197 list_del(&trans->list);
3198 seqmutex_unlock(&c->btree_trans_lock);
3202 free_percpu(c->btree_trans_bufs);
3204 trans = list_first_entry_or_null(&c->btree_trans_list, struct btree_trans, list);
3206 panic("%s leaked btree_trans\n", trans->fn);
3208 for (s = c->btree_transaction_stats;
3209 s < c->btree_transaction_stats + ARRAY_SIZE(c->btree_transaction_stats);
3211 kfree(s->max_paths_text);
3212 bch2_time_stats_exit(&s->lock_hold_times);
3215 if (c->btree_trans_barrier_initialized)
3216 cleanup_srcu_struct(&c->btree_trans_barrier);
3217 mempool_exit(&c->btree_trans_mem_pool);
3218 mempool_exit(&c->btree_trans_pool);
3221 void bch2_fs_btree_iter_init_early(struct bch_fs *c)
3223 struct btree_transaction_stats *s;
3225 for (s = c->btree_transaction_stats;
3226 s < c->btree_transaction_stats + ARRAY_SIZE(c->btree_transaction_stats);
3228 bch2_time_stats_init(&s->lock_hold_times);
3229 mutex_init(&s->lock);
3232 INIT_LIST_HEAD(&c->btree_trans_list);
3233 seqmutex_init(&c->btree_trans_lock);
3236 int bch2_fs_btree_iter_init(struct bch_fs *c)
3240 c->btree_trans_bufs = alloc_percpu(struct btree_trans_buf);
3241 if (!c->btree_trans_bufs)
3244 ret = mempool_init_kmalloc_pool(&c->btree_trans_pool, 1,
3245 sizeof(struct btree_trans)) ?:
3246 mempool_init_kmalloc_pool(&c->btree_trans_mem_pool, 1,
3247 BTREE_TRANS_MEM_MAX) ?:
3248 init_srcu_struct(&c->btree_trans_barrier);
3250 c->btree_trans_barrier_initialized = true;