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"
20 #include <linux/random.h>
21 #include <linux/prefetch.h>
23 static inline void btree_path_list_remove(struct btree_trans *, struct btree_path *);
24 static inline void btree_path_list_add(struct btree_trans *, struct btree_path *,
27 static inline unsigned long btree_iter_ip_allocated(struct btree_iter *iter)
29 #ifdef TRACK_PATH_ALLOCATED
30 return iter->ip_allocated;
36 static struct btree_path *btree_path_alloc(struct btree_trans *, struct btree_path *);
38 static inline int __btree_path_cmp(const struct btree_path *l,
39 enum btree_id r_btree_id,
45 * Must match lock ordering as defined by __bch2_btree_node_lock:
47 return cmp_int(l->btree_id, r_btree_id) ?:
48 cmp_int((int) l->cached, (int) r_cached) ?:
49 bpos_cmp(l->pos, r_pos) ?:
50 -cmp_int(l->level, r_level);
53 static inline int btree_path_cmp(const struct btree_path *l,
54 const struct btree_path *r)
56 return __btree_path_cmp(l, r->btree_id, r->cached, r->pos, r->level);
59 static inline struct bpos bkey_successor(struct btree_iter *iter, struct bpos p)
61 /* Are we iterating over keys in all snapshots? */
62 if (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) {
63 p = bpos_successor(p);
65 p = bpos_nosnap_successor(p);
66 p.snapshot = iter->snapshot;
72 static inline struct bpos bkey_predecessor(struct btree_iter *iter, struct bpos p)
74 /* Are we iterating over keys in all snapshots? */
75 if (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) {
76 p = bpos_predecessor(p);
78 p = bpos_nosnap_predecessor(p);
79 p.snapshot = iter->snapshot;
85 static inline struct bpos btree_iter_search_key(struct btree_iter *iter)
87 struct bpos pos = iter->pos;
89 if ((iter->flags & BTREE_ITER_IS_EXTENTS) &&
90 !bkey_eq(pos, POS_MAX))
91 pos = bkey_successor(iter, pos);
95 static inline bool btree_path_pos_before_node(struct btree_path *path,
98 return bpos_lt(path->pos, b->data->min_key);
101 static inline bool btree_path_pos_after_node(struct btree_path *path,
104 return bpos_gt(path->pos, b->key.k.p);
107 static inline bool btree_path_pos_in_node(struct btree_path *path,
110 return path->btree_id == b->c.btree_id &&
111 !btree_path_pos_before_node(path, b) &&
112 !btree_path_pos_after_node(path, b);
115 /* Btree iterator: */
117 #ifdef CONFIG_BCACHEFS_DEBUG
119 static void bch2_btree_path_verify_cached(struct btree_trans *trans,
120 struct btree_path *path)
122 struct bkey_cached *ck;
123 bool locked = btree_node_locked(path, 0);
125 if (!bch2_btree_node_relock(trans, path, 0))
128 ck = (void *) path->l[0].b;
129 BUG_ON(ck->key.btree_id != path->btree_id ||
130 !bkey_eq(ck->key.pos, path->pos));
133 btree_node_unlock(trans, path, 0);
136 static void bch2_btree_path_verify_level(struct btree_trans *trans,
137 struct btree_path *path, unsigned level)
139 struct btree_path_level *l;
140 struct btree_node_iter tmp;
142 struct bkey_packed *p, *k;
143 struct printbuf buf1 = PRINTBUF;
144 struct printbuf buf2 = PRINTBUF;
145 struct printbuf buf3 = PRINTBUF;
148 if (!bch2_debug_check_iterators)
153 locked = btree_node_locked(path, level);
157 bch2_btree_path_verify_cached(trans, path);
161 if (!btree_path_node(path, level))
164 if (!bch2_btree_node_relock_notrace(trans, path, level))
167 BUG_ON(!btree_path_pos_in_node(path, l->b));
169 bch2_btree_node_iter_verify(&l->iter, l->b);
172 * For interior nodes, the iterator will have skipped past deleted keys:
175 ? bch2_btree_node_iter_prev(&tmp, l->b)
176 : bch2_btree_node_iter_prev_all(&tmp, l->b);
177 k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
179 if (p && bkey_iter_pos_cmp(l->b, p, &path->pos) >= 0) {
184 if (k && bkey_iter_pos_cmp(l->b, k, &path->pos) < 0) {
190 btree_node_unlock(trans, path, level);
193 bch2_bpos_to_text(&buf1, path->pos);
196 struct bkey uk = bkey_unpack_key(l->b, p);
198 bch2_bkey_to_text(&buf2, &uk);
200 prt_printf(&buf2, "(none)");
204 struct bkey uk = bkey_unpack_key(l->b, k);
206 bch2_bkey_to_text(&buf3, &uk);
208 prt_printf(&buf3, "(none)");
211 panic("path should be %s key at level %u:\n"
215 msg, level, buf1.buf, buf2.buf, buf3.buf);
218 static void bch2_btree_path_verify(struct btree_trans *trans,
219 struct btree_path *path)
221 struct bch_fs *c = trans->c;
224 EBUG_ON(path->btree_id >= BTREE_ID_NR);
226 for (i = 0; i < (!path->cached ? BTREE_MAX_DEPTH : 1); i++) {
228 BUG_ON(!path->cached &&
229 bch2_btree_id_root(c, path->btree_id)->b->c.level > i);
233 bch2_btree_path_verify_level(trans, path, i);
236 bch2_btree_path_verify_locks(path);
239 void bch2_trans_verify_paths(struct btree_trans *trans)
241 struct btree_path *path;
243 trans_for_each_path(trans, path)
244 bch2_btree_path_verify(trans, path);
247 static void bch2_btree_iter_verify(struct btree_iter *iter)
249 struct btree_trans *trans = iter->trans;
251 BUG_ON(iter->btree_id >= BTREE_ID_NR);
253 BUG_ON(!!(iter->flags & BTREE_ITER_CACHED) != iter->path->cached);
255 BUG_ON((iter->flags & BTREE_ITER_IS_EXTENTS) &&
256 (iter->flags & BTREE_ITER_ALL_SNAPSHOTS));
258 BUG_ON(!(iter->flags & __BTREE_ITER_ALL_SNAPSHOTS) &&
259 (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) &&
260 !btree_type_has_snapshot_field(iter->btree_id));
262 if (iter->update_path)
263 bch2_btree_path_verify(trans, iter->update_path);
264 bch2_btree_path_verify(trans, iter->path);
267 static void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter)
269 BUG_ON((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
270 !iter->pos.snapshot);
272 BUG_ON(!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS) &&
273 iter->pos.snapshot != iter->snapshot);
275 BUG_ON(bkey_lt(iter->pos, bkey_start_pos(&iter->k)) ||
276 bkey_gt(iter->pos, iter->k.p));
279 static int bch2_btree_iter_verify_ret(struct btree_iter *iter, struct bkey_s_c k)
281 struct btree_trans *trans = iter->trans;
282 struct btree_iter copy;
283 struct bkey_s_c prev;
286 if (!bch2_debug_check_iterators)
289 if (!(iter->flags & BTREE_ITER_FILTER_SNAPSHOTS))
292 if (bkey_err(k) || !k.k)
295 BUG_ON(!bch2_snapshot_is_ancestor(trans->c,
299 bch2_trans_iter_init(trans, ©, iter->btree_id, iter->pos,
300 BTREE_ITER_NOPRESERVE|
301 BTREE_ITER_ALL_SNAPSHOTS);
302 prev = bch2_btree_iter_prev(©);
306 ret = bkey_err(prev);
310 if (bkey_eq(prev.k->p, k.k->p) &&
311 bch2_snapshot_is_ancestor(trans->c, iter->snapshot,
312 prev.k->p.snapshot) > 0) {
313 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
315 bch2_bkey_to_text(&buf1, k.k);
316 bch2_bkey_to_text(&buf2, prev.k);
318 panic("iter snap %u\n"
325 bch2_trans_iter_exit(trans, ©);
329 void bch2_assert_pos_locked(struct btree_trans *trans, enum btree_id id,
330 struct bpos pos, bool key_cache)
332 struct btree_path *path;
334 struct printbuf buf = PRINTBUF;
336 btree_trans_sort_paths(trans);
338 trans_for_each_path_inorder(trans, path, idx) {
339 int cmp = cmp_int(path->btree_id, id) ?:
340 cmp_int(path->cached, key_cache);
347 if (!btree_node_locked(path, 0) ||
348 !path->should_be_locked)
352 if (bkey_ge(pos, path->l[0].b->data->min_key) &&
353 bkey_le(pos, path->l[0].b->key.k.p))
356 if (bkey_eq(pos, path->pos))
361 bch2_dump_trans_paths_updates(trans);
362 bch2_bpos_to_text(&buf, pos);
364 panic("not locked: %s %s%s\n",
365 bch2_btree_id_str(id), buf.buf,
366 key_cache ? " cached" : "");
371 static inline void bch2_btree_path_verify_level(struct btree_trans *trans,
372 struct btree_path *path, unsigned l) {}
373 static inline void bch2_btree_path_verify(struct btree_trans *trans,
374 struct btree_path *path) {}
375 static inline void bch2_btree_iter_verify(struct btree_iter *iter) {}
376 static inline void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter) {}
377 static inline int bch2_btree_iter_verify_ret(struct btree_iter *iter, struct bkey_s_c k) { return 0; }
381 /* Btree path: fixups after btree updates */
383 static void btree_node_iter_set_set_pos(struct btree_node_iter *iter,
386 struct bkey_packed *k)
388 struct btree_node_iter_set *set;
390 btree_node_iter_for_each(iter, set)
391 if (set->end == t->end_offset) {
392 set->k = __btree_node_key_to_offset(b, k);
393 bch2_btree_node_iter_sort(iter, b);
397 bch2_btree_node_iter_push(iter, b, k, btree_bkey_last(b, t));
400 static void __bch2_btree_path_fix_key_modified(struct btree_path *path,
402 struct bkey_packed *where)
404 struct btree_path_level *l = &path->l[b->c.level];
406 if (where != bch2_btree_node_iter_peek_all(&l->iter, l->b))
409 if (bkey_iter_pos_cmp(l->b, where, &path->pos) < 0)
410 bch2_btree_node_iter_advance(&l->iter, l->b);
413 void bch2_btree_path_fix_key_modified(struct btree_trans *trans,
415 struct bkey_packed *where)
417 struct btree_path *path;
419 trans_for_each_path_with_node(trans, b, path) {
420 __bch2_btree_path_fix_key_modified(path, b, where);
421 bch2_btree_path_verify_level(trans, path, b->c.level);
425 static void __bch2_btree_node_iter_fix(struct btree_path *path,
427 struct btree_node_iter *node_iter,
429 struct bkey_packed *where,
430 unsigned clobber_u64s,
433 const struct bkey_packed *end = btree_bkey_last(b, t);
434 struct btree_node_iter_set *set;
435 unsigned offset = __btree_node_key_to_offset(b, where);
436 int shift = new_u64s - clobber_u64s;
437 unsigned old_end = t->end_offset - shift;
438 unsigned orig_iter_pos = node_iter->data[0].k;
439 bool iter_current_key_modified =
440 orig_iter_pos >= offset &&
441 orig_iter_pos <= offset + clobber_u64s;
443 btree_node_iter_for_each(node_iter, set)
444 if (set->end == old_end)
447 /* didn't find the bset in the iterator - might have to readd it: */
449 bkey_iter_pos_cmp(b, where, &path->pos) >= 0) {
450 bch2_btree_node_iter_push(node_iter, b, where, end);
453 /* Iterator is after key that changed */
457 set->end = t->end_offset;
459 /* Iterator hasn't gotten to the key that changed yet: */
464 bkey_iter_pos_cmp(b, where, &path->pos) >= 0) {
466 } else if (set->k < offset + clobber_u64s) {
467 set->k = offset + new_u64s;
468 if (set->k == set->end)
469 bch2_btree_node_iter_set_drop(node_iter, set);
471 /* Iterator is after key that changed */
472 set->k = (int) set->k + shift;
476 bch2_btree_node_iter_sort(node_iter, b);
478 if (node_iter->data[0].k != orig_iter_pos)
479 iter_current_key_modified = true;
482 * When a new key is added, and the node iterator now points to that
483 * key, the iterator might have skipped past deleted keys that should
484 * come after the key the iterator now points to. We have to rewind to
485 * before those deleted keys - otherwise
486 * bch2_btree_node_iter_prev_all() breaks:
488 if (!bch2_btree_node_iter_end(node_iter) &&
489 iter_current_key_modified &&
491 struct bkey_packed *k, *k2, *p;
493 k = bch2_btree_node_iter_peek_all(node_iter, b);
495 for_each_bset(b, t) {
496 bool set_pos = false;
498 if (node_iter->data[0].end == t->end_offset)
501 k2 = bch2_btree_node_iter_bset_pos(node_iter, b, t);
503 while ((p = bch2_bkey_prev_all(b, t, k2)) &&
504 bkey_iter_cmp(b, k, p) < 0) {
510 btree_node_iter_set_set_pos(node_iter,
516 void bch2_btree_node_iter_fix(struct btree_trans *trans,
517 struct btree_path *path,
519 struct btree_node_iter *node_iter,
520 struct bkey_packed *where,
521 unsigned clobber_u64s,
524 struct bset_tree *t = bch2_bkey_to_bset_inlined(b, where);
525 struct btree_path *linked;
527 if (node_iter != &path->l[b->c.level].iter) {
528 __bch2_btree_node_iter_fix(path, b, node_iter, t,
529 where, clobber_u64s, new_u64s);
531 if (bch2_debug_check_iterators)
532 bch2_btree_node_iter_verify(node_iter, b);
535 trans_for_each_path_with_node(trans, b, linked) {
536 __bch2_btree_node_iter_fix(linked, b,
537 &linked->l[b->c.level].iter, t,
538 where, clobber_u64s, new_u64s);
539 bch2_btree_path_verify_level(trans, linked, b->c.level);
543 /* Btree path level: pointer to a particular btree node and node iter */
545 static inline struct bkey_s_c __btree_iter_unpack(struct bch_fs *c,
546 struct btree_path_level *l,
548 struct bkey_packed *k)
552 * signal to bch2_btree_iter_peek_slot() that we're currently at
555 u->type = KEY_TYPE_deleted;
556 return bkey_s_c_null;
559 return bkey_disassemble(l->b, k, u);
562 static inline struct bkey_s_c btree_path_level_peek_all(struct bch_fs *c,
563 struct btree_path_level *l,
566 return __btree_iter_unpack(c, l, u,
567 bch2_btree_node_iter_peek_all(&l->iter, l->b));
570 static inline struct bkey_s_c btree_path_level_peek(struct btree_trans *trans,
571 struct btree_path *path,
572 struct btree_path_level *l,
575 struct bkey_s_c k = __btree_iter_unpack(trans->c, l, u,
576 bch2_btree_node_iter_peek(&l->iter, l->b));
578 path->pos = k.k ? k.k->p : l->b->key.k.p;
579 trans->paths_sorted = false;
580 bch2_btree_path_verify_level(trans, path, l - path->l);
584 static inline struct bkey_s_c btree_path_level_prev(struct btree_trans *trans,
585 struct btree_path *path,
586 struct btree_path_level *l,
589 struct bkey_s_c k = __btree_iter_unpack(trans->c, l, u,
590 bch2_btree_node_iter_prev(&l->iter, l->b));
592 path->pos = k.k ? k.k->p : l->b->data->min_key;
593 trans->paths_sorted = false;
594 bch2_btree_path_verify_level(trans, path, l - path->l);
598 static inline bool btree_path_advance_to_pos(struct btree_path *path,
599 struct btree_path_level *l,
602 struct bkey_packed *k;
605 while ((k = bch2_btree_node_iter_peek_all(&l->iter, l->b)) &&
606 bkey_iter_pos_cmp(l->b, k, &path->pos) < 0) {
607 if (max_advance > 0 && nr_advanced >= max_advance)
610 bch2_btree_node_iter_advance(&l->iter, l->b);
617 static inline void __btree_path_level_init(struct btree_path *path,
620 struct btree_path_level *l = &path->l[level];
622 bch2_btree_node_iter_init(&l->iter, l->b, &path->pos);
625 * Iterators to interior nodes should always be pointed at the first non
629 bch2_btree_node_iter_peek(&l->iter, l->b);
632 void bch2_btree_path_level_init(struct btree_trans *trans,
633 struct btree_path *path,
636 BUG_ON(path->cached);
638 EBUG_ON(!btree_path_pos_in_node(path, b));
640 path->l[b->c.level].lock_seq = six_lock_seq(&b->c.lock);
641 path->l[b->c.level].b = b;
642 __btree_path_level_init(path, b->c.level);
645 /* Btree path: fixups after btree node updates: */
647 static void bch2_trans_revalidate_updates_in_node(struct btree_trans *trans, struct btree *b)
649 struct bch_fs *c = trans->c;
650 struct btree_insert_entry *i;
652 trans_for_each_update(trans, i)
654 i->level == b->c.level &&
655 i->btree_id == b->c.btree_id &&
656 bpos_cmp(i->k->k.p, b->data->min_key) >= 0 &&
657 bpos_cmp(i->k->k.p, b->data->max_key) <= 0) {
658 i->old_v = bch2_btree_path_peek_slot(i->path, &i->old_k).v;
660 if (unlikely(trans->journal_replay_not_finished)) {
662 bch2_journal_keys_peek_slot(c, i->btree_id, i->level,
674 * A btree node is being replaced - update the iterator to point to the new
677 void bch2_trans_node_add(struct btree_trans *trans, struct btree *b)
679 struct btree_path *path;
681 trans_for_each_path(trans, path)
682 if (path->uptodate == BTREE_ITER_UPTODATE &&
684 btree_path_pos_in_node(path, b)) {
685 enum btree_node_locked_type t =
686 btree_lock_want(path, b->c.level);
688 if (t != BTREE_NODE_UNLOCKED) {
689 btree_node_unlock(trans, path, b->c.level);
690 six_lock_increment(&b->c.lock, (enum six_lock_type) t);
691 mark_btree_node_locked(trans, path, b->c.level, t);
694 bch2_btree_path_level_init(trans, path, b);
697 bch2_trans_revalidate_updates_in_node(trans, b);
701 * A btree node has been modified in such a way as to invalidate iterators - fix
704 void bch2_trans_node_reinit_iter(struct btree_trans *trans, struct btree *b)
706 struct btree_path *path;
708 trans_for_each_path_with_node(trans, b, path)
709 __btree_path_level_init(path, b->c.level);
711 bch2_trans_revalidate_updates_in_node(trans, b);
714 /* Btree path: traverse, set_pos: */
716 static inline int btree_path_lock_root(struct btree_trans *trans,
717 struct btree_path *path,
719 unsigned long trace_ip)
721 struct bch_fs *c = trans->c;
722 struct btree *b, **rootp = &bch2_btree_id_root(c, path->btree_id)->b;
723 enum six_lock_type lock_type;
727 EBUG_ON(path->nodes_locked);
730 b = READ_ONCE(*rootp);
731 path->level = READ_ONCE(b->c.level);
733 if (unlikely(path->level < depth_want)) {
735 * the root is at a lower depth than the depth we want:
736 * got to the end of the btree, or we're walking nodes
737 * greater than some depth and there are no nodes >=
740 path->level = depth_want;
741 for (i = path->level; i < BTREE_MAX_DEPTH; i++)
746 lock_type = __btree_lock_want(path, path->level);
747 ret = btree_node_lock(trans, path, &b->c,
748 path->level, lock_type, trace_ip);
750 if (bch2_err_matches(ret, BCH_ERR_lock_fail_root_changed))
752 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
757 if (likely(b == READ_ONCE(*rootp) &&
758 b->c.level == path->level &&
760 for (i = 0; i < path->level; i++)
761 path->l[i].b = ERR_PTR(-BCH_ERR_no_btree_node_lock_root);
762 path->l[path->level].b = b;
763 for (i = path->level + 1; i < BTREE_MAX_DEPTH; i++)
766 mark_btree_node_locked(trans, path, path->level,
767 (enum btree_node_locked_type) lock_type);
768 bch2_btree_path_level_init(trans, path, b);
772 six_unlock_type(&b->c.lock, lock_type);
777 static int btree_path_prefetch(struct btree_trans *trans, struct btree_path *path)
779 struct bch_fs *c = trans->c;
780 struct btree_path_level *l = path_l(path);
781 struct btree_node_iter node_iter = l->iter;
782 struct bkey_packed *k;
784 unsigned nr = test_bit(BCH_FS_started, &c->flags)
785 ? (path->level > 1 ? 0 : 2)
786 : (path->level > 1 ? 1 : 16);
787 bool was_locked = btree_node_locked(path, path->level);
790 bch2_bkey_buf_init(&tmp);
792 while (nr-- && !ret) {
793 if (!bch2_btree_node_relock(trans, path, path->level))
796 bch2_btree_node_iter_advance(&node_iter, l->b);
797 k = bch2_btree_node_iter_peek(&node_iter, l->b);
801 bch2_bkey_buf_unpack(&tmp, c, l->b, k);
802 ret = bch2_btree_node_prefetch(trans, path, tmp.k, path->btree_id,
807 btree_node_unlock(trans, path, path->level);
809 bch2_bkey_buf_exit(&tmp, c);
813 static int btree_path_prefetch_j(struct btree_trans *trans, struct btree_path *path,
814 struct btree_and_journal_iter *jiter)
816 struct bch_fs *c = trans->c;
819 unsigned nr = test_bit(BCH_FS_started, &c->flags)
820 ? (path->level > 1 ? 0 : 2)
821 : (path->level > 1 ? 1 : 16);
822 bool was_locked = btree_node_locked(path, path->level);
825 bch2_bkey_buf_init(&tmp);
827 while (nr-- && !ret) {
828 if (!bch2_btree_node_relock(trans, path, path->level))
831 bch2_btree_and_journal_iter_advance(jiter);
832 k = bch2_btree_and_journal_iter_peek(jiter);
836 bch2_bkey_buf_reassemble(&tmp, c, k);
837 ret = bch2_btree_node_prefetch(trans, path, tmp.k, path->btree_id,
842 btree_node_unlock(trans, path, path->level);
844 bch2_bkey_buf_exit(&tmp, c);
848 static noinline void btree_node_mem_ptr_set(struct btree_trans *trans,
849 struct btree_path *path,
850 unsigned plevel, struct btree *b)
852 struct btree_path_level *l = &path->l[plevel];
853 bool locked = btree_node_locked(path, plevel);
854 struct bkey_packed *k;
855 struct bch_btree_ptr_v2 *bp;
857 if (!bch2_btree_node_relock(trans, path, plevel))
860 k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
861 BUG_ON(k->type != KEY_TYPE_btree_ptr_v2);
863 bp = (void *) bkeyp_val(&l->b->format, k);
864 bp->mem_ptr = (unsigned long)b;
867 btree_node_unlock(trans, path, plevel);
870 static noinline int btree_node_iter_and_journal_peek(struct btree_trans *trans,
871 struct btree_path *path,
873 struct bkey_buf *out)
875 struct bch_fs *c = trans->c;
876 struct btree_path_level *l = path_l(path);
877 struct btree_and_journal_iter jiter;
881 __bch2_btree_and_journal_iter_init_node_iter(&jiter, c, l->b, l->iter, path->pos);
883 k = bch2_btree_and_journal_iter_peek(&jiter);
885 bch2_bkey_buf_reassemble(out, c, k);
887 if (flags & BTREE_ITER_PREFETCH)
888 ret = btree_path_prefetch_j(trans, path, &jiter);
890 bch2_btree_and_journal_iter_exit(&jiter);
894 static __always_inline int btree_path_down(struct btree_trans *trans,
895 struct btree_path *path,
897 unsigned long trace_ip)
899 struct bch_fs *c = trans->c;
900 struct btree_path_level *l = path_l(path);
902 unsigned level = path->level - 1;
903 enum six_lock_type lock_type = __btree_lock_want(path, level);
907 EBUG_ON(!btree_node_locked(path, path->level));
909 bch2_bkey_buf_init(&tmp);
911 if (unlikely(trans->journal_replay_not_finished)) {
912 ret = btree_node_iter_and_journal_peek(trans, path, flags, &tmp);
916 bch2_bkey_buf_unpack(&tmp, c, l->b,
917 bch2_btree_node_iter_peek(&l->iter, l->b));
919 if (flags & BTREE_ITER_PREFETCH) {
920 ret = btree_path_prefetch(trans, path);
926 b = bch2_btree_node_get(trans, path, tmp.k, level, lock_type, trace_ip);
927 ret = PTR_ERR_OR_ZERO(b);
931 if (likely(!trans->journal_replay_not_finished &&
932 tmp.k->k.type == KEY_TYPE_btree_ptr_v2) &&
933 unlikely(b != btree_node_mem_ptr(tmp.k)))
934 btree_node_mem_ptr_set(trans, path, level + 1, b);
936 if (btree_node_read_locked(path, level + 1))
937 btree_node_unlock(trans, path, level + 1);
939 mark_btree_node_locked(trans, path, level,
940 (enum btree_node_locked_type) lock_type);
942 bch2_btree_path_level_init(trans, path, b);
944 bch2_btree_path_verify_locks(path);
946 bch2_bkey_buf_exit(&tmp, c);
951 static int bch2_btree_path_traverse_all(struct btree_trans *trans)
953 struct bch_fs *c = trans->c;
954 struct btree_path *path;
955 unsigned long trace_ip = _RET_IP_;
958 if (trans->in_traverse_all)
959 return -BCH_ERR_transaction_restart_in_traverse_all;
961 trans->in_traverse_all = true;
963 trans->restarted = 0;
964 trans->last_restarted_ip = 0;
966 trans_for_each_path(trans, path)
967 path->should_be_locked = false;
969 btree_trans_sort_paths(trans);
971 bch2_trans_unlock(trans);
974 if (unlikely(trans->memory_allocation_failure)) {
977 closure_init_stack(&cl);
980 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
985 /* Now, redo traversals in correct order: */
987 while (i < trans->nr_sorted) {
988 path = trans->paths + trans->sorted[i];
991 * Traversing a path can cause another path to be added at about
994 if (path->uptodate) {
995 __btree_path_get(path, false);
996 ret = bch2_btree_path_traverse_one(trans, path, 0, _THIS_IP_);
997 __btree_path_put(path, false);
999 if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
1000 bch2_err_matches(ret, ENOMEM))
1010 * We used to assert that all paths had been traversed here
1011 * (path->uptodate < BTREE_ITER_NEED_TRAVERSE); however, since
1012 * path->should_be_locked is not set yet, we might have unlocked and
1013 * then failed to relock a path - that's fine.
1016 bch2_btree_cache_cannibalize_unlock(c);
1018 trans->in_traverse_all = false;
1020 trace_and_count(c, trans_traverse_all, trans, trace_ip);
1024 static inline bool btree_path_check_pos_in_node(struct btree_path *path,
1025 unsigned l, int check_pos)
1027 if (check_pos < 0 && btree_path_pos_before_node(path, path->l[l].b))
1029 if (check_pos > 0 && btree_path_pos_after_node(path, path->l[l].b))
1034 static inline bool btree_path_good_node(struct btree_trans *trans,
1035 struct btree_path *path,
1036 unsigned l, int check_pos)
1038 return is_btree_node(path, l) &&
1039 bch2_btree_node_relock(trans, path, l) &&
1040 btree_path_check_pos_in_node(path, l, check_pos);
1043 static void btree_path_set_level_down(struct btree_trans *trans,
1044 struct btree_path *path,
1049 path->level = new_level;
1051 for (l = path->level + 1; l < BTREE_MAX_DEPTH; l++)
1052 if (btree_lock_want(path, l) == BTREE_NODE_UNLOCKED)
1053 btree_node_unlock(trans, path, l);
1055 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1056 bch2_btree_path_verify(trans, path);
1059 static noinline unsigned __btree_path_up_until_good_node(struct btree_trans *trans,
1060 struct btree_path *path,
1063 unsigned i, l = path->level;
1065 while (btree_path_node(path, l) &&
1066 !btree_path_good_node(trans, path, l, check_pos))
1067 __btree_path_set_level_up(trans, path, l++);
1069 /* If we need intent locks, take them too: */
1071 i < path->locks_want && btree_path_node(path, i);
1073 if (!bch2_btree_node_relock(trans, path, i)) {
1075 __btree_path_set_level_up(trans, path, l++);
1082 static inline unsigned btree_path_up_until_good_node(struct btree_trans *trans,
1083 struct btree_path *path,
1086 return likely(btree_node_locked(path, path->level) &&
1087 btree_path_check_pos_in_node(path, path->level, check_pos))
1089 : __btree_path_up_until_good_node(trans, path, check_pos);
1093 * This is the main state machine for walking down the btree - walks down to a
1096 * Returns 0 on success, -EIO on error (error reading in a btree node).
1098 * On error, caller (peek_node()/peek_key()) must return NULL; the error is
1099 * stashed in the iterator and returned from bch2_trans_exit().
1101 int bch2_btree_path_traverse_one(struct btree_trans *trans,
1102 struct btree_path *path,
1104 unsigned long trace_ip)
1106 unsigned depth_want = path->level;
1107 int ret = -((int) trans->restarted);
1112 if (unlikely(!trans->srcu_held))
1113 bch2_trans_srcu_lock(trans);
1116 * Ensure we obey path->should_be_locked: if it's set, we can't unlock
1117 * and re-traverse the path without a transaction restart:
1119 if (path->should_be_locked) {
1120 ret = bch2_btree_path_relock(trans, path, trace_ip);
1125 ret = bch2_btree_path_traverse_cached(trans, path, flags);
1129 if (unlikely(path->level >= BTREE_MAX_DEPTH))
1132 path->level = btree_path_up_until_good_node(trans, path, 0);
1134 EBUG_ON(btree_path_node(path, path->level) &&
1135 !btree_node_locked(path, path->level));
1138 * Note: path->nodes[path->level] may be temporarily NULL here - that
1139 * would indicate to other code that we got to the end of the btree,
1140 * here it indicates that relocking the root failed - it's critical that
1141 * btree_path_lock_root() comes next and that it can't fail
1143 while (path->level > depth_want) {
1144 ret = btree_path_node(path, path->level)
1145 ? btree_path_down(trans, path, flags, trace_ip)
1146 : btree_path_lock_root(trans, path, depth_want, trace_ip);
1147 if (unlikely(ret)) {
1150 * No nodes at this level - got to the end of
1157 __bch2_btree_path_unlock(trans, path);
1158 path->level = depth_want;
1159 path->l[path->level].b = ERR_PTR(ret);
1164 path->uptodate = BTREE_ITER_UPTODATE;
1166 if (bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted)
1167 panic("ret %s (%i) trans->restarted %s (%i)\n",
1168 bch2_err_str(ret), ret,
1169 bch2_err_str(trans->restarted), trans->restarted);
1170 bch2_btree_path_verify(trans, path);
1174 static inline void btree_path_copy(struct btree_trans *trans, struct btree_path *dst,
1175 struct btree_path *src)
1177 unsigned i, offset = offsetof(struct btree_path, pos);
1179 memcpy((void *) dst + offset,
1180 (void *) src + offset,
1181 sizeof(struct btree_path) - offset);
1183 for (i = 0; i < BTREE_MAX_DEPTH; i++) {
1184 unsigned t = btree_node_locked_type(dst, i);
1186 if (t != BTREE_NODE_UNLOCKED)
1187 six_lock_increment(&dst->l[i].b->c.lock, t);
1191 static struct btree_path *btree_path_clone(struct btree_trans *trans, struct btree_path *src,
1194 struct btree_path *new = btree_path_alloc(trans, src);
1196 btree_path_copy(trans, new, src);
1197 __btree_path_get(new, intent);
1202 struct btree_path *__bch2_btree_path_make_mut(struct btree_trans *trans,
1203 struct btree_path *path, bool intent,
1206 __btree_path_put(path, intent);
1207 path = btree_path_clone(trans, path, intent);
1208 path->preserve = false;
1212 struct btree_path * __must_check
1213 __bch2_btree_path_set_pos(struct btree_trans *trans,
1214 struct btree_path *path, struct bpos new_pos,
1215 bool intent, unsigned long ip, int cmp)
1217 bch2_trans_verify_not_in_restart(trans);
1218 EBUG_ON(!path->ref);
1220 path = bch2_btree_path_make_mut(trans, path, intent, ip);
1222 path->pos = new_pos;
1223 trans->paths_sorted = false;
1225 if (unlikely(path->cached)) {
1226 btree_node_unlock(trans, path, 0);
1227 path->l[0].b = ERR_PTR(-BCH_ERR_no_btree_node_up);
1228 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1232 unsigned level = btree_path_up_until_good_node(trans, path, cmp);
1234 if (btree_path_node(path, level)) {
1235 struct btree_path_level *l = &path->l[level];
1237 BUG_ON(!btree_node_locked(path, level));
1239 * We might have to skip over many keys, or just a few: try
1240 * advancing the node iterator, and if we have to skip over too
1241 * many keys just reinit it (or if we're rewinding, since that
1245 !btree_path_advance_to_pos(path, l, 8))
1246 bch2_btree_node_iter_init(&l->iter, l->b, &path->pos);
1249 * Iterators to interior nodes should always be pointed at the first non
1252 if (unlikely(level))
1253 bch2_btree_node_iter_peek(&l->iter, l->b);
1256 if (unlikely(level != path->level)) {
1257 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1258 __bch2_btree_path_unlock(trans, path);
1261 bch2_btree_path_verify(trans, path);
1265 /* Btree path: main interface: */
1267 static struct btree_path *have_path_at_pos(struct btree_trans *trans, struct btree_path *path)
1269 struct btree_path *sib;
1271 sib = prev_btree_path(trans, path);
1272 if (sib && !btree_path_cmp(sib, path))
1275 sib = next_btree_path(trans, path);
1276 if (sib && !btree_path_cmp(sib, path))
1282 static struct btree_path *have_node_at_pos(struct btree_trans *trans, struct btree_path *path)
1284 struct btree_path *sib;
1286 sib = prev_btree_path(trans, path);
1287 if (sib && sib->level == path->level && path_l(sib)->b == path_l(path)->b)
1290 sib = next_btree_path(trans, path);
1291 if (sib && sib->level == path->level && path_l(sib)->b == path_l(path)->b)
1297 static inline void __bch2_path_free(struct btree_trans *trans, struct btree_path *path)
1299 __bch2_btree_path_unlock(trans, path);
1300 btree_path_list_remove(trans, path);
1301 trans->paths_allocated &= ~(1ULL << path->idx);
1304 void bch2_path_put(struct btree_trans *trans, struct btree_path *path, bool intent)
1306 struct btree_path *dup;
1308 EBUG_ON(trans->paths + path->idx != path);
1309 EBUG_ON(!path->ref);
1311 if (!__btree_path_put(path, intent))
1314 dup = path->preserve
1315 ? have_path_at_pos(trans, path)
1316 : have_node_at_pos(trans, path);
1318 if (!dup && !(!path->preserve && !is_btree_node(path, path->level)))
1321 if (path->should_be_locked &&
1322 !trans->restarted &&
1323 (!dup || !bch2_btree_path_relock_norestart(trans, dup, _THIS_IP_)))
1327 dup->preserve |= path->preserve;
1328 dup->should_be_locked |= path->should_be_locked;
1331 __bch2_path_free(trans, path);
1334 static void bch2_path_put_nokeep(struct btree_trans *trans, struct btree_path *path,
1337 EBUG_ON(trans->paths + path->idx != path);
1338 EBUG_ON(!path->ref);
1340 if (!__btree_path_put(path, intent))
1343 __bch2_path_free(trans, path);
1346 void __noreturn bch2_trans_restart_error(struct btree_trans *trans, u32 restart_count)
1348 panic("trans->restart_count %u, should be %u, last restarted by %pS\n",
1349 trans->restart_count, restart_count,
1350 (void *) trans->last_begin_ip);
1353 void __noreturn bch2_trans_in_restart_error(struct btree_trans *trans)
1355 panic("in transaction restart: %s, last restarted by %pS\n",
1356 bch2_err_str(trans->restarted),
1357 (void *) trans->last_restarted_ip);
1361 void bch2_trans_updates_to_text(struct printbuf *buf, struct btree_trans *trans)
1363 struct btree_insert_entry *i;
1364 struct btree_write_buffered_key *wb;
1366 prt_printf(buf, "transaction updates for %s journal seq %llu",
1367 trans->fn, trans->journal_res.seq);
1369 printbuf_indent_add(buf, 2);
1371 trans_for_each_update(trans, i) {
1372 struct bkey_s_c old = { &i->old_k, i->old_v };
1374 prt_printf(buf, "update: btree=%s cached=%u %pS",
1375 bch2_btree_id_str(i->btree_id),
1377 (void *) i->ip_allocated);
1380 prt_printf(buf, " old ");
1381 bch2_bkey_val_to_text(buf, trans->c, old);
1384 prt_printf(buf, " new ");
1385 bch2_bkey_val_to_text(buf, trans->c, bkey_i_to_s_c(i->k));
1389 trans_for_each_wb_update(trans, wb) {
1390 prt_printf(buf, "update: btree=%s wb=1 %pS",
1391 bch2_btree_id_str(wb->btree),
1392 (void *) i->ip_allocated);
1395 prt_printf(buf, " new ");
1396 bch2_bkey_val_to_text(buf, trans->c, bkey_i_to_s_c(&wb->k));
1400 printbuf_indent_sub(buf, 2);
1404 void bch2_dump_trans_updates(struct btree_trans *trans)
1406 struct printbuf buf = PRINTBUF;
1408 bch2_trans_updates_to_text(&buf, trans);
1409 bch2_print_string_as_lines(KERN_ERR, buf.buf);
1410 printbuf_exit(&buf);
1414 void bch2_btree_path_to_text(struct printbuf *out, struct btree_path *path)
1416 prt_printf(out, "path: idx %2u ref %u:%u %c %c btree=%s l=%u pos ",
1417 path->idx, path->ref, path->intent_ref,
1418 path->preserve ? 'P' : ' ',
1419 path->should_be_locked ? 'S' : ' ',
1420 bch2_btree_id_str(path->btree_id),
1422 bch2_bpos_to_text(out, path->pos);
1424 prt_printf(out, " locks %u", path->nodes_locked);
1425 #ifdef TRACK_PATH_ALLOCATED
1426 prt_printf(out, " %pS", (void *) path->ip_allocated);
1431 static noinline __cold
1432 void __bch2_trans_paths_to_text(struct printbuf *out, struct btree_trans *trans,
1435 struct btree_path *path;
1439 btree_trans_sort_paths(trans);
1441 trans_for_each_path_inorder(trans, path, idx)
1442 bch2_btree_path_to_text(out, path);
1446 void bch2_trans_paths_to_text(struct printbuf *out, struct btree_trans *trans)
1448 __bch2_trans_paths_to_text(out, trans, false);
1451 static noinline __cold
1452 void __bch2_dump_trans_paths_updates(struct btree_trans *trans, bool nosort)
1454 struct printbuf buf = PRINTBUF;
1456 __bch2_trans_paths_to_text(&buf, trans, nosort);
1457 bch2_trans_updates_to_text(&buf, trans);
1459 bch2_print_string_as_lines(KERN_ERR, buf.buf);
1460 printbuf_exit(&buf);
1464 void bch2_dump_trans_paths_updates(struct btree_trans *trans)
1466 __bch2_dump_trans_paths_updates(trans, false);
1470 static void bch2_trans_update_max_paths(struct btree_trans *trans)
1472 struct btree_transaction_stats *s = btree_trans_stats(trans);
1473 struct printbuf buf = PRINTBUF;
1478 bch2_trans_paths_to_text(&buf, trans);
1480 if (!buf.allocation_failure) {
1481 mutex_lock(&s->lock);
1482 if (s->nr_max_paths < hweight64(trans->paths_allocated)) {
1483 s->nr_max_paths = trans->nr_max_paths =
1484 hweight64(trans->paths_allocated);
1485 swap(s->max_paths_text, buf.buf);
1487 mutex_unlock(&s->lock);
1490 printbuf_exit(&buf);
1492 trans->nr_max_paths = hweight64(trans->paths_allocated);
1496 int __bch2_btree_trans_too_many_iters(struct btree_trans *trans)
1498 if (trace_trans_restart_too_many_iters_enabled()) {
1499 struct printbuf buf = PRINTBUF;
1501 bch2_trans_paths_to_text(&buf, trans);
1502 trace_trans_restart_too_many_iters(trans, _THIS_IP_, buf.buf);
1503 printbuf_exit(&buf);
1506 count_event(trans->c, trans_restart_too_many_iters);
1508 return btree_trans_restart(trans, BCH_ERR_transaction_restart_too_many_iters);
1511 static noinline void btree_path_overflow(struct btree_trans *trans)
1513 bch2_dump_trans_paths_updates(trans);
1514 panic("trans path overflow\n");
1517 static inline struct btree_path *btree_path_alloc(struct btree_trans *trans,
1518 struct btree_path *pos)
1520 struct btree_path *path;
1523 if (unlikely(trans->paths_allocated ==
1524 ~((~0ULL << 1) << (BTREE_ITER_MAX - 1))))
1525 btree_path_overflow(trans);
1527 idx = __ffs64(~trans->paths_allocated);
1530 * Do this before marking the new path as allocated, since it won't be
1533 if (unlikely(idx > trans->nr_max_paths))
1534 bch2_trans_update_max_paths(trans);
1536 trans->paths_allocated |= 1ULL << idx;
1538 path = &trans->paths[idx];
1541 path->intent_ref = 0;
1542 path->nodes_locked = 0;
1545 btree_path_list_add(trans, pos, path);
1546 trans->paths_sorted = false;
1550 struct btree_path *bch2_path_get(struct btree_trans *trans,
1551 enum btree_id btree_id, struct bpos pos,
1552 unsigned locks_want, unsigned level,
1553 unsigned flags, unsigned long ip)
1555 struct btree_path *path, *path_pos = NULL;
1556 bool cached = flags & BTREE_ITER_CACHED;
1557 bool intent = flags & BTREE_ITER_INTENT;
1560 bch2_trans_verify_not_in_restart(trans);
1561 bch2_trans_verify_locks(trans);
1563 btree_trans_sort_paths(trans);
1565 trans_for_each_path_inorder(trans, path, i) {
1566 if (__btree_path_cmp(path,
1577 path_pos->cached == cached &&
1578 path_pos->btree_id == btree_id &&
1579 path_pos->level == level) {
1580 __btree_path_get(path_pos, intent);
1581 path = bch2_btree_path_set_pos(trans, path_pos, pos, intent, ip);
1583 path = btree_path_alloc(trans, path_pos);
1586 __btree_path_get(path, intent);
1588 path->btree_id = btree_id;
1589 path->cached = cached;
1590 path->uptodate = BTREE_ITER_NEED_TRAVERSE;
1591 path->should_be_locked = false;
1592 path->level = level;
1593 path->locks_want = locks_want;
1594 path->nodes_locked = 0;
1595 for (i = 0; i < ARRAY_SIZE(path->l); i++)
1596 path->l[i].b = ERR_PTR(-BCH_ERR_no_btree_node_init);
1597 #ifdef TRACK_PATH_ALLOCATED
1598 path->ip_allocated = ip;
1600 trans->paths_sorted = false;
1603 if (!(flags & BTREE_ITER_NOPRESERVE))
1604 path->preserve = true;
1606 if (path->intent_ref)
1607 locks_want = max(locks_want, level + 1);
1610 * If the path has locks_want greater than requested, we don't downgrade
1611 * it here - on transaction restart because btree node split needs to
1612 * upgrade locks, we might be putting/getting the iterator again.
1613 * Downgrading iterators only happens via bch2_trans_downgrade(), after
1614 * a successful transaction commit.
1617 locks_want = min(locks_want, BTREE_MAX_DEPTH);
1618 if (locks_want > path->locks_want)
1619 bch2_btree_path_upgrade_noupgrade_sibs(trans, path, locks_want, NULL);
1624 struct bkey_s_c bch2_btree_path_peek_slot(struct btree_path *path, struct bkey *u)
1627 struct btree_path_level *l = path_l(path);
1628 struct bkey_packed *_k;
1631 if (unlikely(!l->b))
1632 return bkey_s_c_null;
1634 EBUG_ON(path->uptodate != BTREE_ITER_UPTODATE);
1635 EBUG_ON(!btree_node_locked(path, path->level));
1637 if (!path->cached) {
1638 _k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
1639 k = _k ? bkey_disassemble(l->b, _k, u) : bkey_s_c_null;
1641 EBUG_ON(k.k && bkey_deleted(k.k) && bpos_eq(k.k->p, path->pos));
1643 if (!k.k || !bpos_eq(path->pos, k.k->p))
1646 struct bkey_cached *ck = (void *) path->l[0].b;
1649 (path->btree_id != ck->key.btree_id ||
1650 !bkey_eq(path->pos, ck->key.pos)));
1651 if (!ck || !ck->valid)
1652 return bkey_s_c_null;
1655 k = bkey_i_to_s_c(ck->k);
1662 return (struct bkey_s_c) { u, NULL };
1665 /* Btree iterators: */
1668 __bch2_btree_iter_traverse(struct btree_iter *iter)
1670 return bch2_btree_path_traverse(iter->trans, iter->path, iter->flags);
1674 bch2_btree_iter_traverse(struct btree_iter *iter)
1678 iter->path = bch2_btree_path_set_pos(iter->trans, iter->path,
1679 btree_iter_search_key(iter),
1680 iter->flags & BTREE_ITER_INTENT,
1681 btree_iter_ip_allocated(iter));
1683 ret = bch2_btree_path_traverse(iter->trans, iter->path, iter->flags);
1687 btree_path_set_should_be_locked(iter->path);
1691 /* Iterate across nodes (leaf and interior nodes) */
1693 struct btree *bch2_btree_iter_peek_node(struct btree_iter *iter)
1695 struct btree_trans *trans = iter->trans;
1696 struct btree *b = NULL;
1699 EBUG_ON(iter->path->cached);
1700 bch2_btree_iter_verify(iter);
1702 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
1706 b = btree_path_node(iter->path, iter->path->level);
1710 BUG_ON(bpos_lt(b->key.k.p, iter->pos));
1712 bkey_init(&iter->k);
1713 iter->k.p = iter->pos = b->key.k.p;
1715 iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p,
1716 iter->flags & BTREE_ITER_INTENT,
1717 btree_iter_ip_allocated(iter));
1718 btree_path_set_should_be_locked(iter->path);
1720 bch2_btree_iter_verify_entry_exit(iter);
1721 bch2_btree_iter_verify(iter);
1729 struct btree *bch2_btree_iter_peek_node_and_restart(struct btree_iter *iter)
1733 while (b = bch2_btree_iter_peek_node(iter),
1734 bch2_err_matches(PTR_ERR_OR_ZERO(b), BCH_ERR_transaction_restart))
1735 bch2_trans_begin(iter->trans);
1740 struct btree *bch2_btree_iter_next_node(struct btree_iter *iter)
1742 struct btree_trans *trans = iter->trans;
1743 struct btree_path *path = iter->path;
1744 struct btree *b = NULL;
1747 bch2_trans_verify_not_in_restart(trans);
1748 EBUG_ON(iter->path->cached);
1749 bch2_btree_iter_verify(iter);
1751 /* already at end? */
1752 if (!btree_path_node(path, path->level))
1756 if (!btree_path_node(path, path->level + 1)) {
1757 btree_path_set_level_up(trans, path);
1761 if (!bch2_btree_node_relock(trans, path, path->level + 1)) {
1762 __bch2_btree_path_unlock(trans, path);
1763 path->l[path->level].b = ERR_PTR(-BCH_ERR_no_btree_node_relock);
1764 path->l[path->level + 1].b = ERR_PTR(-BCH_ERR_no_btree_node_relock);
1765 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1766 trace_and_count(trans->c, trans_restart_relock_next_node, trans, _THIS_IP_, path);
1767 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_relock);
1771 b = btree_path_node(path, path->level + 1);
1773 if (bpos_eq(iter->pos, b->key.k.p)) {
1774 __btree_path_set_level_up(trans, path, path->level++);
1777 * Haven't gotten to the end of the parent node: go back down to
1778 * the next child node
1781 bch2_btree_path_set_pos(trans, path, bpos_successor(iter->pos),
1782 iter->flags & BTREE_ITER_INTENT,
1783 btree_iter_ip_allocated(iter));
1785 btree_path_set_level_down(trans, path, iter->min_depth);
1787 ret = bch2_btree_path_traverse(trans, path, iter->flags);
1791 b = path->l[path->level].b;
1794 bkey_init(&iter->k);
1795 iter->k.p = iter->pos = b->key.k.p;
1797 iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p,
1798 iter->flags & BTREE_ITER_INTENT,
1799 btree_iter_ip_allocated(iter));
1800 btree_path_set_should_be_locked(iter->path);
1801 BUG_ON(iter->path->uptodate);
1803 bch2_btree_iter_verify_entry_exit(iter);
1804 bch2_btree_iter_verify(iter);
1812 /* Iterate across keys (in leaf nodes only) */
1814 inline bool bch2_btree_iter_advance(struct btree_iter *iter)
1816 struct bpos pos = iter->k.p;
1817 bool ret = !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS
1818 ? bpos_eq(pos, SPOS_MAX)
1819 : bkey_eq(pos, SPOS_MAX));
1821 if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
1822 pos = bkey_successor(iter, pos);
1823 bch2_btree_iter_set_pos(iter, pos);
1827 inline bool bch2_btree_iter_rewind(struct btree_iter *iter)
1829 struct bpos pos = bkey_start_pos(&iter->k);
1830 bool ret = !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS
1831 ? bpos_eq(pos, POS_MIN)
1832 : bkey_eq(pos, POS_MIN));
1834 if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
1835 pos = bkey_predecessor(iter, pos);
1836 bch2_btree_iter_set_pos(iter, pos);
1841 struct bkey_i *__bch2_btree_trans_peek_updates(struct btree_iter *iter)
1843 struct btree_insert_entry *i;
1844 struct bkey_i *ret = NULL;
1846 trans_for_each_update(iter->trans, i) {
1847 if (i->btree_id < iter->btree_id)
1849 if (i->btree_id > iter->btree_id)
1851 if (bpos_lt(i->k->k.p, iter->path->pos))
1853 if (i->key_cache_already_flushed)
1855 if (!ret || bpos_lt(i->k->k.p, ret->k.p))
1862 static inline struct bkey_i *btree_trans_peek_updates(struct btree_iter *iter)
1864 return iter->flags & BTREE_ITER_WITH_UPDATES
1865 ? __bch2_btree_trans_peek_updates(iter)
1869 static struct bkey_i *bch2_btree_journal_peek(struct btree_trans *trans,
1870 struct btree_iter *iter,
1871 struct bpos end_pos)
1873 return bch2_journal_keys_peek_upto(trans->c, iter->btree_id,
1877 &iter->journal_idx);
1881 struct bkey_s_c btree_trans_peek_slot_journal(struct btree_trans *trans,
1882 struct btree_iter *iter)
1884 struct bkey_i *k = bch2_btree_journal_peek(trans, iter, iter->path->pos);
1888 return bkey_i_to_s_c(k);
1890 return bkey_s_c_null;
1895 struct bkey_s_c btree_trans_peek_journal(struct btree_trans *trans,
1896 struct btree_iter *iter,
1899 struct bkey_i *next_journal =
1900 bch2_btree_journal_peek(trans, iter,
1901 k.k ? k.k->p : path_l(iter->path)->b->key.k.p);
1904 iter->k = next_journal->k;
1905 k = bkey_i_to_s_c(next_journal);
1912 * Checks btree key cache for key at iter->pos and returns it if present, or
1916 struct bkey_s_c btree_trans_peek_key_cache(struct btree_iter *iter, struct bpos pos)
1918 struct btree_trans *trans = iter->trans;
1919 struct bch_fs *c = trans->c;
1924 if ((iter->flags & BTREE_ITER_KEY_CACHE_FILL) &&
1925 bpos_eq(iter->pos, pos))
1926 return bkey_s_c_null;
1928 if (!bch2_btree_key_cache_find(c, iter->btree_id, pos))
1929 return bkey_s_c_null;
1931 if (!iter->key_cache_path)
1932 iter->key_cache_path = bch2_path_get(trans, iter->btree_id, pos,
1933 iter->flags & BTREE_ITER_INTENT, 0,
1934 iter->flags|BTREE_ITER_CACHED|
1935 BTREE_ITER_CACHED_NOFILL,
1938 iter->key_cache_path = bch2_btree_path_set_pos(trans, iter->key_cache_path, pos,
1939 iter->flags & BTREE_ITER_INTENT,
1940 btree_iter_ip_allocated(iter));
1942 ret = bch2_btree_path_traverse(trans, iter->key_cache_path,
1943 iter->flags|BTREE_ITER_CACHED) ?:
1944 bch2_btree_path_relock(trans, iter->path, _THIS_IP_);
1946 return bkey_s_c_err(ret);
1948 btree_path_set_should_be_locked(iter->key_cache_path);
1950 k = bch2_btree_path_peek_slot(iter->key_cache_path, &u);
1951 if (k.k && !bkey_err(k)) {
1958 static struct bkey_s_c __bch2_btree_iter_peek(struct btree_iter *iter, struct bpos search_key)
1960 struct btree_trans *trans = iter->trans;
1961 struct bkey_i *next_update;
1962 struct bkey_s_c k, k2;
1965 EBUG_ON(iter->path->cached);
1966 bch2_btree_iter_verify(iter);
1969 struct btree_path_level *l;
1971 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
1972 iter->flags & BTREE_ITER_INTENT,
1973 btree_iter_ip_allocated(iter));
1975 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
1976 if (unlikely(ret)) {
1977 /* ensure that iter->k is consistent with iter->pos: */
1978 bch2_btree_iter_set_pos(iter, iter->pos);
1979 k = bkey_s_c_err(ret);
1983 l = path_l(iter->path);
1985 if (unlikely(!l->b)) {
1986 /* No btree nodes at requested level: */
1987 bch2_btree_iter_set_pos(iter, SPOS_MAX);
1992 btree_path_set_should_be_locked(iter->path);
1994 k = btree_path_level_peek_all(trans->c, l, &iter->k);
1996 if (unlikely(iter->flags & BTREE_ITER_WITH_KEY_CACHE) &&
1998 (k2 = btree_trans_peek_key_cache(iter, k.k->p)).k) {
2002 bch2_btree_iter_set_pos(iter, iter->pos);
2007 if (unlikely(iter->flags & BTREE_ITER_WITH_JOURNAL))
2008 k = btree_trans_peek_journal(trans, iter, k);
2010 next_update = btree_trans_peek_updates(iter);
2013 bpos_le(next_update->k.p,
2014 k.k ? k.k->p : l->b->key.k.p)) {
2015 iter->k = next_update->k;
2016 k = bkey_i_to_s_c(next_update);
2019 if (k.k && bkey_deleted(k.k)) {
2021 * If we've got a whiteout, and it's after the search
2022 * key, advance the search key to the whiteout instead
2023 * of just after the whiteout - it might be a btree
2024 * whiteout, with a real key at the same position, since
2025 * in the btree deleted keys sort before non deleted.
2027 search_key = !bpos_eq(search_key, k.k->p)
2029 : bpos_successor(k.k->p);
2035 } else if (likely(!bpos_eq(l->b->key.k.p, SPOS_MAX))) {
2036 /* Advance to next leaf node: */
2037 search_key = bpos_successor(l->b->key.k.p);
2040 bch2_btree_iter_set_pos(iter, SPOS_MAX);
2046 bch2_btree_iter_verify(iter);
2052 * bch2_btree_iter_peek_upto() - returns first key greater than or equal to
2053 * iterator's current position
2054 * @iter: iterator to peek from
2055 * @end: search limit: returns keys less than or equal to @end
2057 * Returns: key if found, or an error extractable with bkey_err().
2059 struct bkey_s_c bch2_btree_iter_peek_upto(struct btree_iter *iter, struct bpos end)
2061 struct btree_trans *trans = iter->trans;
2062 struct bpos search_key = btree_iter_search_key(iter);
2064 struct bpos iter_pos;
2067 EBUG_ON((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) && bkey_eq(end, POS_MAX));
2069 if (iter->update_path) {
2070 bch2_path_put_nokeep(trans, iter->update_path,
2071 iter->flags & BTREE_ITER_INTENT);
2072 iter->update_path = NULL;
2075 bch2_btree_iter_verify_entry_exit(iter);
2078 k = __bch2_btree_iter_peek(iter, search_key);
2081 if (unlikely(bkey_err(k)))
2085 * iter->pos should be mononotically increasing, and always be
2086 * equal to the key we just returned - except extents can
2087 * straddle iter->pos:
2089 if (!(iter->flags & BTREE_ITER_IS_EXTENTS))
2092 iter_pos = bkey_max(iter->pos, bkey_start_pos(k.k));
2094 if (unlikely(!(iter->flags & BTREE_ITER_IS_EXTENTS)
2095 ? bkey_gt(iter_pos, end)
2096 : bkey_ge(iter_pos, end)))
2099 if (iter->update_path &&
2100 !bkey_eq(iter->update_path->pos, k.k->p)) {
2101 bch2_path_put_nokeep(trans, iter->update_path,
2102 iter->flags & BTREE_ITER_INTENT);
2103 iter->update_path = NULL;
2106 if ((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
2107 (iter->flags & BTREE_ITER_INTENT) &&
2108 !(iter->flags & BTREE_ITER_IS_EXTENTS) &&
2109 !iter->update_path) {
2110 struct bpos pos = k.k->p;
2112 if (pos.snapshot < iter->snapshot) {
2113 search_key = bpos_successor(k.k->p);
2117 pos.snapshot = iter->snapshot;
2120 * advance, same as on exit for iter->path, but only up
2123 __btree_path_get(iter->path, iter->flags & BTREE_ITER_INTENT);
2124 iter->update_path = iter->path;
2126 iter->update_path = bch2_btree_path_set_pos(trans,
2127 iter->update_path, pos,
2128 iter->flags & BTREE_ITER_INTENT,
2130 ret = bch2_btree_path_traverse(trans, iter->update_path, iter->flags);
2131 if (unlikely(ret)) {
2132 k = bkey_s_c_err(ret);
2138 * We can never have a key in a leaf node at POS_MAX, so
2139 * we don't have to check these successor() calls:
2141 if ((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
2142 !bch2_snapshot_is_ancestor(trans->c,
2145 search_key = bpos_successor(k.k->p);
2149 if (bkey_whiteout(k.k) &&
2150 !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS)) {
2151 search_key = bkey_successor(iter, k.k->p);
2158 iter->pos = iter_pos;
2160 iter->path = bch2_btree_path_set_pos(trans, iter->path, k.k->p,
2161 iter->flags & BTREE_ITER_INTENT,
2162 btree_iter_ip_allocated(iter));
2164 btree_path_set_should_be_locked(iter->path);
2166 if (iter->update_path) {
2167 ret = bch2_btree_path_relock(trans, iter->update_path, _THIS_IP_);
2169 k = bkey_s_c_err(ret);
2171 btree_path_set_should_be_locked(iter->update_path);
2174 if (!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS))
2175 iter->pos.snapshot = iter->snapshot;
2177 ret = bch2_btree_iter_verify_ret(iter, k);
2178 if (unlikely(ret)) {
2179 bch2_btree_iter_set_pos(iter, iter->pos);
2180 k = bkey_s_c_err(ret);
2183 bch2_btree_iter_verify_entry_exit(iter);
2187 bch2_btree_iter_set_pos(iter, end);
2193 * bch2_btree_iter_next() - returns first key greater than iterator's current
2195 * @iter: iterator to peek from
2197 * Returns: key if found, or an error extractable with bkey_err().
2199 struct bkey_s_c bch2_btree_iter_next(struct btree_iter *iter)
2201 if (!bch2_btree_iter_advance(iter))
2202 return bkey_s_c_null;
2204 return bch2_btree_iter_peek(iter);
2208 * bch2_btree_iter_peek_prev() - returns first key less than or equal to
2209 * iterator's current position
2210 * @iter: iterator to peek from
2212 * Returns: key if found, or an error extractable with bkey_err().
2214 struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *iter)
2216 struct btree_trans *trans = iter->trans;
2217 struct bpos search_key = iter->pos;
2218 struct btree_path *saved_path = NULL;
2220 struct bkey saved_k;
2221 const struct bch_val *saved_v;
2224 EBUG_ON(iter->path->cached || iter->path->level);
2225 EBUG_ON(iter->flags & BTREE_ITER_WITH_UPDATES);
2227 if (iter->flags & BTREE_ITER_WITH_JOURNAL)
2228 return bkey_s_c_err(-EIO);
2230 bch2_btree_iter_verify(iter);
2231 bch2_btree_iter_verify_entry_exit(iter);
2233 if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2234 search_key.snapshot = U32_MAX;
2237 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2238 iter->flags & BTREE_ITER_INTENT,
2239 btree_iter_ip_allocated(iter));
2241 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2242 if (unlikely(ret)) {
2243 /* ensure that iter->k is consistent with iter->pos: */
2244 bch2_btree_iter_set_pos(iter, iter->pos);
2245 k = bkey_s_c_err(ret);
2249 k = btree_path_level_peek(trans, iter->path,
2250 &iter->path->l[0], &iter->k);
2252 ((iter->flags & BTREE_ITER_IS_EXTENTS)
2253 ? bpos_ge(bkey_start_pos(k.k), search_key)
2254 : bpos_gt(k.k->p, search_key)))
2255 k = btree_path_level_prev(trans, iter->path,
2256 &iter->path->l[0], &iter->k);
2259 if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) {
2260 if (k.k->p.snapshot == iter->snapshot)
2264 * If we have a saved candidate, and we're no
2265 * longer at the same _key_ (not pos), return
2268 if (saved_path && !bkey_eq(k.k->p, saved_k.p)) {
2269 bch2_path_put_nokeep(trans, iter->path,
2270 iter->flags & BTREE_ITER_INTENT);
2271 iter->path = saved_path;
2278 if (bch2_snapshot_is_ancestor(iter->trans->c,
2282 bch2_path_put_nokeep(trans, saved_path,
2283 iter->flags & BTREE_ITER_INTENT);
2284 saved_path = btree_path_clone(trans, iter->path,
2285 iter->flags & BTREE_ITER_INTENT);
2290 search_key = bpos_predecessor(k.k->p);
2294 if (bkey_whiteout(k.k) &&
2295 !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS)) {
2296 search_key = bkey_predecessor(iter, k.k->p);
2297 if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2298 search_key.snapshot = U32_MAX;
2303 } else if (likely(!bpos_eq(iter->path->l[0].b->data->min_key, POS_MIN))) {
2304 /* Advance to previous leaf node: */
2305 search_key = bpos_predecessor(iter->path->l[0].b->data->min_key);
2307 /* Start of btree: */
2308 bch2_btree_iter_set_pos(iter, POS_MIN);
2314 EBUG_ON(bkey_gt(bkey_start_pos(k.k), iter->pos));
2316 /* Extents can straddle iter->pos: */
2317 if (bkey_lt(k.k->p, iter->pos))
2320 if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2321 iter->pos.snapshot = iter->snapshot;
2323 btree_path_set_should_be_locked(iter->path);
2326 bch2_path_put_nokeep(trans, saved_path, iter->flags & BTREE_ITER_INTENT);
2328 bch2_btree_iter_verify_entry_exit(iter);
2329 bch2_btree_iter_verify(iter);
2335 * bch2_btree_iter_prev() - returns first key less than iterator's current
2337 * @iter: iterator to peek from
2339 * Returns: key if found, or an error extractable with bkey_err().
2341 struct bkey_s_c bch2_btree_iter_prev(struct btree_iter *iter)
2343 if (!bch2_btree_iter_rewind(iter))
2344 return bkey_s_c_null;
2346 return bch2_btree_iter_peek_prev(iter);
2349 struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *iter)
2351 struct btree_trans *trans = iter->trans;
2352 struct bpos search_key;
2356 bch2_btree_iter_verify(iter);
2357 bch2_btree_iter_verify_entry_exit(iter);
2358 EBUG_ON(iter->path->level && (iter->flags & BTREE_ITER_WITH_KEY_CACHE));
2360 /* extents can't span inode numbers: */
2361 if ((iter->flags & BTREE_ITER_IS_EXTENTS) &&
2362 unlikely(iter->pos.offset == KEY_OFFSET_MAX)) {
2363 if (iter->pos.inode == KEY_INODE_MAX)
2364 return bkey_s_c_null;
2366 bch2_btree_iter_set_pos(iter, bpos_nosnap_successor(iter->pos));
2369 search_key = btree_iter_search_key(iter);
2370 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2371 iter->flags & BTREE_ITER_INTENT,
2372 btree_iter_ip_allocated(iter));
2374 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2375 if (unlikely(ret)) {
2376 k = bkey_s_c_err(ret);
2380 if ((iter->flags & BTREE_ITER_CACHED) ||
2381 !(iter->flags & (BTREE_ITER_IS_EXTENTS|BTREE_ITER_FILTER_SNAPSHOTS))) {
2382 struct bkey_i *next_update;
2384 if ((next_update = btree_trans_peek_updates(iter)) &&
2385 bpos_eq(next_update->k.p, iter->pos)) {
2386 iter->k = next_update->k;
2387 k = bkey_i_to_s_c(next_update);
2391 if (unlikely(iter->flags & BTREE_ITER_WITH_JOURNAL) &&
2392 (k = btree_trans_peek_slot_journal(trans, iter)).k)
2395 if (unlikely(iter->flags & BTREE_ITER_WITH_KEY_CACHE) &&
2396 (k = btree_trans_peek_key_cache(iter, iter->pos)).k) {
2399 /* We're not returning a key from iter->path: */
2403 k = bch2_btree_path_peek_slot(iter->path, &iter->k);
2408 struct bpos end = iter->pos;
2410 if (iter->flags & BTREE_ITER_IS_EXTENTS)
2411 end.offset = U64_MAX;
2413 EBUG_ON(iter->path->level);
2415 if (iter->flags & BTREE_ITER_INTENT) {
2416 struct btree_iter iter2;
2418 bch2_trans_copy_iter(&iter2, iter);
2419 k = bch2_btree_iter_peek_upto(&iter2, end);
2421 if (k.k && !bkey_err(k)) {
2425 bch2_trans_iter_exit(trans, &iter2);
2427 struct bpos pos = iter->pos;
2429 k = bch2_btree_iter_peek_upto(iter, end);
2430 if (unlikely(bkey_err(k)))
2431 bch2_btree_iter_set_pos(iter, pos);
2436 if (unlikely(bkey_err(k)))
2439 next = k.k ? bkey_start_pos(k.k) : POS_MAX;
2441 if (bkey_lt(iter->pos, next)) {
2442 bkey_init(&iter->k);
2443 iter->k.p = iter->pos;
2445 if (iter->flags & BTREE_ITER_IS_EXTENTS) {
2446 bch2_key_resize(&iter->k,
2447 min_t(u64, KEY_SIZE_MAX,
2448 (next.inode == iter->pos.inode
2452 EBUG_ON(!iter->k.size);
2455 k = (struct bkey_s_c) { &iter->k, NULL };
2459 btree_path_set_should_be_locked(iter->path);
2461 bch2_btree_iter_verify_entry_exit(iter);
2462 bch2_btree_iter_verify(iter);
2463 ret = bch2_btree_iter_verify_ret(iter, k);
2465 return bkey_s_c_err(ret);
2470 struct bkey_s_c bch2_btree_iter_next_slot(struct btree_iter *iter)
2472 if (!bch2_btree_iter_advance(iter))
2473 return bkey_s_c_null;
2475 return bch2_btree_iter_peek_slot(iter);
2478 struct bkey_s_c bch2_btree_iter_prev_slot(struct btree_iter *iter)
2480 if (!bch2_btree_iter_rewind(iter))
2481 return bkey_s_c_null;
2483 return bch2_btree_iter_peek_slot(iter);
2486 struct bkey_s_c bch2_btree_iter_peek_and_restart_outlined(struct btree_iter *iter)
2490 while (btree_trans_too_many_iters(iter->trans) ||
2491 (k = bch2_btree_iter_peek_type(iter, iter->flags),
2492 bch2_err_matches(bkey_err(k), BCH_ERR_transaction_restart)))
2493 bch2_trans_begin(iter->trans);
2498 /* new transactional stuff: */
2500 #ifdef CONFIG_BCACHEFS_DEBUG
2501 static void btree_trans_verify_sorted_refs(struct btree_trans *trans)
2503 struct btree_path *path;
2506 BUG_ON(trans->nr_sorted != hweight64(trans->paths_allocated));
2508 trans_for_each_path(trans, path) {
2509 BUG_ON(path->sorted_idx >= trans->nr_sorted);
2510 BUG_ON(trans->sorted[path->sorted_idx] != path->idx);
2513 for (i = 0; i < trans->nr_sorted; i++) {
2514 unsigned idx = trans->sorted[i];
2516 EBUG_ON(!(trans->paths_allocated & (1ULL << idx)));
2517 BUG_ON(trans->paths[idx].sorted_idx != i);
2521 static void btree_trans_verify_sorted(struct btree_trans *trans)
2523 struct btree_path *path, *prev = NULL;
2526 if (!bch2_debug_check_iterators)
2529 trans_for_each_path_inorder(trans, path, i) {
2530 if (prev && btree_path_cmp(prev, path) > 0) {
2531 __bch2_dump_trans_paths_updates(trans, true);
2532 panic("trans paths out of order!\n");
2538 static inline void btree_trans_verify_sorted_refs(struct btree_trans *trans) {}
2539 static inline void btree_trans_verify_sorted(struct btree_trans *trans) {}
2542 void __bch2_btree_trans_sort_paths(struct btree_trans *trans)
2544 int i, l = 0, r = trans->nr_sorted, inc = 1;
2547 btree_trans_verify_sorted_refs(trans);
2549 if (trans->paths_sorted)
2553 * Cocktail shaker sort: this is efficient because iterators will be
2559 for (i = inc > 0 ? l : r - 2;
2560 i + 1 < r && i >= l;
2562 if (btree_path_cmp(trans->paths + trans->sorted[i],
2563 trans->paths + trans->sorted[i + 1]) > 0) {
2564 swap(trans->sorted[i], trans->sorted[i + 1]);
2565 trans->paths[trans->sorted[i]].sorted_idx = i;
2566 trans->paths[trans->sorted[i + 1]].sorted_idx = i + 1;
2578 trans->paths_sorted = true;
2580 btree_trans_verify_sorted(trans);
2583 static inline void btree_path_list_remove(struct btree_trans *trans,
2584 struct btree_path *path)
2588 EBUG_ON(path->sorted_idx >= trans->nr_sorted);
2589 #ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
2591 memmove_u64s_down_small(trans->sorted + path->sorted_idx,
2592 trans->sorted + path->sorted_idx + 1,
2593 DIV_ROUND_UP(trans->nr_sorted - path->sorted_idx, 8));
2595 array_remove_item(trans->sorted, trans->nr_sorted, path->sorted_idx);
2597 for (i = path->sorted_idx; i < trans->nr_sorted; i++)
2598 trans->paths[trans->sorted[i]].sorted_idx = i;
2600 path->sorted_idx = U8_MAX;
2603 static inline void btree_path_list_add(struct btree_trans *trans,
2604 struct btree_path *pos,
2605 struct btree_path *path)
2609 path->sorted_idx = pos ? pos->sorted_idx + 1 : trans->nr_sorted;
2611 #ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
2612 memmove_u64s_up_small(trans->sorted + path->sorted_idx + 1,
2613 trans->sorted + path->sorted_idx,
2614 DIV_ROUND_UP(trans->nr_sorted - path->sorted_idx, 8));
2616 trans->sorted[path->sorted_idx] = path->idx;
2618 array_insert_item(trans->sorted, trans->nr_sorted, path->sorted_idx, path->idx);
2621 for (i = path->sorted_idx; i < trans->nr_sorted; i++)
2622 trans->paths[trans->sorted[i]].sorted_idx = i;
2624 btree_trans_verify_sorted_refs(trans);
2627 void bch2_trans_iter_exit(struct btree_trans *trans, struct btree_iter *iter)
2629 if (iter->update_path)
2630 bch2_path_put_nokeep(trans, iter->update_path,
2631 iter->flags & BTREE_ITER_INTENT);
2633 bch2_path_put(trans, iter->path,
2634 iter->flags & BTREE_ITER_INTENT);
2635 if (iter->key_cache_path)
2636 bch2_path_put(trans, iter->key_cache_path,
2637 iter->flags & BTREE_ITER_INTENT);
2639 iter->update_path = NULL;
2640 iter->key_cache_path = NULL;
2643 void bch2_trans_iter_init_outlined(struct btree_trans *trans,
2644 struct btree_iter *iter,
2645 enum btree_id btree_id, struct bpos pos,
2648 bch2_trans_iter_init_common(trans, iter, btree_id, pos, 0, 0,
2649 bch2_btree_iter_flags(trans, btree_id, flags),
2653 void bch2_trans_node_iter_init(struct btree_trans *trans,
2654 struct btree_iter *iter,
2655 enum btree_id btree_id,
2657 unsigned locks_want,
2661 flags |= BTREE_ITER_NOT_EXTENTS;
2662 flags |= __BTREE_ITER_ALL_SNAPSHOTS;
2663 flags |= BTREE_ITER_ALL_SNAPSHOTS;
2665 bch2_trans_iter_init_common(trans, iter, btree_id, pos, locks_want, depth,
2666 __bch2_btree_iter_flags(trans, btree_id, flags),
2669 iter->min_depth = depth;
2671 BUG_ON(iter->path->locks_want < min(locks_want, BTREE_MAX_DEPTH));
2672 BUG_ON(iter->path->level != depth);
2673 BUG_ON(iter->min_depth != depth);
2676 void bch2_trans_copy_iter(struct btree_iter *dst, struct btree_iter *src)
2680 __btree_path_get(src->path, src->flags & BTREE_ITER_INTENT);
2681 if (src->update_path)
2682 __btree_path_get(src->update_path, src->flags & BTREE_ITER_INTENT);
2683 dst->key_cache_path = NULL;
2686 void *__bch2_trans_kmalloc(struct btree_trans *trans, size_t size)
2688 unsigned new_top = trans->mem_top + size;
2689 size_t old_bytes = trans->mem_bytes;
2690 size_t new_bytes = roundup_pow_of_two(new_top);
2695 trans->mem_max = max(trans->mem_max, new_top);
2697 WARN_ON_ONCE(new_bytes > BTREE_TRANS_MEM_MAX);
2699 new_mem = krealloc(trans->mem, new_bytes, GFP_NOWAIT|__GFP_NOWARN);
2700 if (unlikely(!new_mem)) {
2701 bch2_trans_unlock(trans);
2703 new_mem = krealloc(trans->mem, new_bytes, GFP_KERNEL);
2704 if (!new_mem && new_bytes <= BTREE_TRANS_MEM_MAX) {
2705 new_mem = mempool_alloc(&trans->c->btree_trans_mem_pool, GFP_KERNEL);
2706 new_bytes = BTREE_TRANS_MEM_MAX;
2711 return ERR_PTR(-BCH_ERR_ENOMEM_trans_kmalloc);
2713 trans->mem = new_mem;
2714 trans->mem_bytes = new_bytes;
2716 ret = bch2_trans_relock(trans);
2718 return ERR_PTR(ret);
2721 trans->mem = new_mem;
2722 trans->mem_bytes = new_bytes;
2725 trace_and_count(trans->c, trans_restart_mem_realloced, trans, _RET_IP_, new_bytes);
2726 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_mem_realloced));
2729 p = trans->mem + trans->mem_top;
2730 trans->mem_top += size;
2735 static inline void check_srcu_held_too_long(struct btree_trans *trans)
2737 WARN(trans->srcu_held && time_after(jiffies, trans->srcu_lock_time + HZ * 10),
2738 "btree trans held srcu lock (delaying memory reclaim) for %lu seconds",
2739 (jiffies - trans->srcu_lock_time) / HZ);
2742 void bch2_trans_srcu_unlock(struct btree_trans *trans)
2744 if (trans->srcu_held) {
2745 struct bch_fs *c = trans->c;
2746 struct btree_path *path;
2748 trans_for_each_path(trans, path)
2749 if (path->cached && !btree_node_locked(path, 0))
2750 path->l[0].b = ERR_PTR(-BCH_ERR_no_btree_node_srcu_reset);
2752 check_srcu_held_too_long(trans);
2753 srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx);
2754 trans->srcu_held = false;
2758 void bch2_trans_srcu_lock(struct btree_trans *trans)
2760 if (!trans->srcu_held) {
2761 trans->srcu_idx = srcu_read_lock(&trans->c->btree_trans_barrier);
2762 trans->srcu_lock_time = jiffies;
2763 trans->srcu_held = true;
2768 * bch2_trans_begin() - reset a transaction after a interrupted attempt
2769 * @trans: transaction to reset
2771 * Returns: current restart counter, to be used with trans_was_restarted()
2773 * While iterating over nodes or updating nodes a attempt to lock a btree node
2774 * may return BCH_ERR_transaction_restart when the trylock fails. When this
2775 * occurs bch2_trans_begin() should be called and the transaction retried.
2777 u32 bch2_trans_begin(struct btree_trans *trans)
2779 struct btree_path *path;
2782 bch2_trans_reset_updates(trans);
2784 trans->restart_count++;
2787 trans_for_each_path(trans, path) {
2788 path->should_be_locked = false;
2791 * If the transaction wasn't restarted, we're presuming to be
2792 * doing something new: dont keep iterators excpt the ones that
2793 * are in use - except for the subvolumes btree:
2795 if (!trans->restarted && path->btree_id != BTREE_ID_subvolumes)
2796 path->preserve = false;
2799 * XXX: we probably shouldn't be doing this if the transaction
2800 * was restarted, but currently we still overflow transaction
2801 * iterators if we do that
2803 if (!path->ref && !path->preserve)
2804 __bch2_path_free(trans, path);
2806 path->preserve = false;
2809 now = local_clock();
2810 if (!trans->restarted &&
2812 now - trans->last_begin_time > BTREE_TRANS_MAX_LOCK_HOLD_TIME_NS)) {
2813 drop_locks_do(trans, (cond_resched(), 0));
2814 now = local_clock();
2816 trans->last_begin_time = now;
2818 if (unlikely(trans->srcu_held &&
2819 time_after(jiffies, trans->srcu_lock_time + msecs_to_jiffies(10))))
2820 bch2_trans_srcu_unlock(trans);
2822 trans->last_begin_ip = _RET_IP_;
2823 if (trans->restarted) {
2824 bch2_btree_path_traverse_all(trans);
2825 trans->notrace_relock_fail = false;
2828 return trans->restart_count;
2831 static struct btree_trans *bch2_trans_alloc(struct bch_fs *c)
2833 struct btree_trans *trans;
2835 if (IS_ENABLED(__KERNEL__)) {
2836 trans = this_cpu_xchg(c->btree_trans_bufs->trans, NULL);
2841 trans = mempool_alloc(&c->btree_trans_pool, GFP_NOFS);
2843 * paths need to be zeroed, bch2_check_for_deadlock looks at
2844 * paths in other threads
2846 memset(&trans->paths, 0, sizeof(trans->paths));
2850 const char *bch2_btree_transaction_fns[BCH_TRANSACTIONS_NR];
2852 unsigned bch2_trans_get_fn_idx(const char *fn)
2856 for (i = 0; i < ARRAY_SIZE(bch2_btree_transaction_fns); i++)
2857 if (!bch2_btree_transaction_fns[i] ||
2858 bch2_btree_transaction_fns[i] == fn) {
2859 bch2_btree_transaction_fns[i] = fn;
2863 pr_warn_once("BCH_TRANSACTIONS_NR not big enough!");
2867 struct btree_trans *__bch2_trans_get(struct bch_fs *c, unsigned fn_idx)
2868 __acquires(&c->btree_trans_barrier)
2870 struct btree_trans *trans;
2871 struct btree_transaction_stats *s;
2873 bch2_assert_btree_nodes_not_locked();
2875 trans = bch2_trans_alloc(c);
2877 memset(trans, 0, sizeof(*trans));
2879 trans->fn = fn_idx < ARRAY_SIZE(bch2_btree_transaction_fns)
2880 ? bch2_btree_transaction_fns[fn_idx] : NULL;
2881 trans->last_begin_time = local_clock();
2882 trans->fn_idx = fn_idx;
2883 trans->locking_wait.task = current;
2884 trans->journal_replay_not_finished =
2885 unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)) &&
2886 atomic_inc_not_zero(&c->journal_keys.ref);
2887 closure_init_stack(&trans->ref);
2889 s = btree_trans_stats(trans);
2890 if (s && s->max_mem) {
2891 unsigned expected_mem_bytes = roundup_pow_of_two(s->max_mem);
2893 trans->mem = kmalloc(expected_mem_bytes, GFP_KERNEL);
2895 if (!unlikely(trans->mem)) {
2896 trans->mem = mempool_alloc(&c->btree_trans_mem_pool, GFP_KERNEL);
2897 trans->mem_bytes = BTREE_TRANS_MEM_MAX;
2899 trans->mem_bytes = expected_mem_bytes;
2904 trans->nr_max_paths = s->nr_max_paths;
2905 trans->wb_updates_size = s->wb_updates_size;
2908 trans->srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
2909 trans->srcu_lock_time = jiffies;
2910 trans->srcu_held = true;
2912 if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG_TRANSACTIONS)) {
2913 struct btree_trans *pos;
2915 seqmutex_lock(&c->btree_trans_lock);
2916 list_for_each_entry(pos, &c->btree_trans_list, list) {
2918 * We'd much prefer to be stricter here and completely
2919 * disallow multiple btree_trans in the same thread -
2920 * but the data move path calls bch2_write when we
2921 * already have a btree_trans initialized.
2923 BUG_ON(trans->locking_wait.task->pid == pos->locking_wait.task->pid &&
2924 bch2_trans_locked(pos));
2926 if (trans->locking_wait.task->pid < pos->locking_wait.task->pid) {
2927 list_add_tail(&trans->list, &pos->list);
2931 list_add_tail(&trans->list, &c->btree_trans_list);
2933 seqmutex_unlock(&c->btree_trans_lock);
2939 static void check_btree_paths_leaked(struct btree_trans *trans)
2941 #ifdef CONFIG_BCACHEFS_DEBUG
2942 struct bch_fs *c = trans->c;
2943 struct btree_path *path;
2945 trans_for_each_path(trans, path)
2950 bch_err(c, "btree paths leaked from %s!", trans->fn);
2951 trans_for_each_path(trans, path)
2953 printk(KERN_ERR " btree %s %pS\n",
2954 bch2_btree_id_str(path->btree_id),
2955 (void *) path->ip_allocated);
2956 /* Be noisy about this: */
2957 bch2_fatal_error(c);
2961 void bch2_trans_put(struct btree_trans *trans)
2962 __releases(&c->btree_trans_barrier)
2964 struct btree_insert_entry *i;
2965 struct bch_fs *c = trans->c;
2966 struct btree_transaction_stats *s = btree_trans_stats(trans);
2968 bch2_trans_unlock(trans);
2970 if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG_TRANSACTIONS)) {
2971 seqmutex_lock(&c->btree_trans_lock);
2972 list_del(&trans->list);
2973 seqmutex_unlock(&c->btree_trans_lock);
2976 closure_sync(&trans->ref);
2979 s->max_mem = max(s->max_mem, trans->mem_max);
2981 trans_for_each_update(trans, i)
2982 __btree_path_put(i->path, true);
2983 trans->nr_updates = 0;
2985 check_btree_paths_leaked(trans);
2987 if (trans->srcu_held) {
2988 check_srcu_held_too_long(trans);
2989 srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx);
2992 kfree(trans->extra_journal_entries.data);
2994 if (trans->fs_usage_deltas) {
2995 if (trans->fs_usage_deltas->size + sizeof(trans->fs_usage_deltas) ==
2996 REPLICAS_DELTA_LIST_MAX)
2997 mempool_free(trans->fs_usage_deltas,
2998 &c->replicas_delta_pool);
3000 kfree(trans->fs_usage_deltas);
3003 if (unlikely(trans->journal_replay_not_finished))
3004 bch2_journal_keys_put(c);
3006 if (trans->mem_bytes == BTREE_TRANS_MEM_MAX)
3007 mempool_free(trans->mem, &c->btree_trans_mem_pool);
3011 /* Userspace doesn't have a real percpu implementation: */
3012 if (IS_ENABLED(__KERNEL__))
3013 trans = this_cpu_xchg(c->btree_trans_bufs->trans, trans);
3015 mempool_free(trans, &c->btree_trans_pool);
3018 static void __maybe_unused
3019 bch2_btree_bkey_cached_common_to_text(struct printbuf *out,
3020 struct btree_bkey_cached_common *b)
3022 struct six_lock_count c = six_lock_counts(&b->lock);
3023 struct task_struct *owner;
3027 owner = READ_ONCE(b->lock.owner);
3028 pid = owner ? owner->pid : 0;
3032 prt_printf(out, "%px %c l=%u %s:", b, b->cached ? 'c' : 'b',
3033 b->level, bch2_btree_id_str(b->btree_id));
3034 bch2_bpos_to_text(out, btree_node_pos(b));
3037 prt_printf(out, " locks %u:%u:%u held by pid %u",
3038 c.n[0], c.n[1], c.n[2], pid);
3041 void bch2_btree_trans_to_text(struct printbuf *out, struct btree_trans *trans)
3043 struct btree_path *path;
3044 struct btree_bkey_cached_common *b;
3045 static char lock_types[] = { 'r', 'i', 'w' };
3046 struct task_struct *task = READ_ONCE(trans->locking_wait.task);
3049 if (!out->nr_tabstops) {
3050 printbuf_tabstop_push(out, 16);
3051 printbuf_tabstop_push(out, 32);
3054 prt_printf(out, "%i %s\n", task ? task->pid : 0, trans->fn);
3056 trans_for_each_path_safe(trans, path, idx) {
3057 if (!path->nodes_locked)
3060 prt_printf(out, " path %u %c l=%u %s:",
3062 path->cached ? 'c' : 'b',
3064 bch2_btree_id_str(path->btree_id));
3065 bch2_bpos_to_text(out, path->pos);
3068 for (l = 0; l < BTREE_MAX_DEPTH; l++) {
3069 if (btree_node_locked(path, l) &&
3070 !IS_ERR_OR_NULL(b = (void *) READ_ONCE(path->l[l].b))) {
3071 prt_printf(out, " %c l=%u ",
3072 lock_types[btree_node_locked_type(path, l)], l);
3073 bch2_btree_bkey_cached_common_to_text(out, b);
3079 b = READ_ONCE(trans->locking);
3081 prt_printf(out, " blocked for %lluus on",
3082 div_u64(local_clock() - trans->locking_wait.start_time,
3085 prt_printf(out, " %c", lock_types[trans->locking_wait.lock_want]);
3086 bch2_btree_bkey_cached_common_to_text(out, b);
3091 void bch2_fs_btree_iter_exit(struct bch_fs *c)
3093 struct btree_transaction_stats *s;
3094 struct btree_trans *trans;
3097 trans = list_first_entry_or_null(&c->btree_trans_list, struct btree_trans, list);
3099 panic("%s leaked btree_trans\n", trans->fn);
3101 if (c->btree_trans_bufs)
3102 for_each_possible_cpu(cpu)
3103 kfree(per_cpu_ptr(c->btree_trans_bufs, cpu)->trans);
3104 free_percpu(c->btree_trans_bufs);
3106 for (s = c->btree_transaction_stats;
3107 s < c->btree_transaction_stats + ARRAY_SIZE(c->btree_transaction_stats);
3109 kfree(s->max_paths_text);
3110 bch2_time_stats_exit(&s->lock_hold_times);
3113 if (c->btree_trans_barrier_initialized)
3114 cleanup_srcu_struct(&c->btree_trans_barrier);
3115 mempool_exit(&c->btree_trans_mem_pool);
3116 mempool_exit(&c->btree_trans_pool);
3119 int bch2_fs_btree_iter_init(struct bch_fs *c)
3121 struct btree_transaction_stats *s;
3124 for (s = c->btree_transaction_stats;
3125 s < c->btree_transaction_stats + ARRAY_SIZE(c->btree_transaction_stats);
3127 bch2_time_stats_init(&s->lock_hold_times);
3128 mutex_init(&s->lock);
3131 INIT_LIST_HEAD(&c->btree_trans_list);
3132 seqmutex_init(&c->btree_trans_lock);
3134 c->btree_trans_bufs = alloc_percpu(struct btree_trans_buf);
3135 if (!c->btree_trans_bufs)
3138 ret = mempool_init_kmalloc_pool(&c->btree_trans_pool, 1,
3139 sizeof(struct btree_trans)) ?:
3140 mempool_init_kmalloc_pool(&c->btree_trans_mem_pool, 1,
3141 BTREE_TRANS_MEM_MAX) ?:
3142 init_srcu_struct(&c->btree_trans_barrier);
3144 c->btree_trans_barrier_initialized = true;