1 // SPDX-License-Identifier: GPL-2.0
4 #include "btree_cache.h"
6 #include "btree_iter.h"
7 #include "btree_locking.h"
10 #include <linux/prefetch.h>
11 #include <linux/sched/mm.h>
12 #include <trace/events/bcachefs.h>
14 const char * const bch2_btree_ids[] = {
15 #define x(kwd, val, name) name,
21 void bch2_recalc_btree_reserve(struct bch_fs *c)
23 unsigned i, reserve = 16;
25 if (!c->btree_roots[0].b)
28 for (i = 0; i < BTREE_ID_NR; i++)
29 if (c->btree_roots[i].b)
30 reserve += min_t(unsigned, 1,
31 c->btree_roots[i].b->c.level) * 8;
33 c->btree_cache.reserve = reserve;
36 static inline unsigned btree_cache_can_free(struct btree_cache *bc)
38 return max_t(int, 0, bc->used - bc->reserve);
41 static void __btree_node_data_free(struct bch_fs *c, struct btree *b)
43 EBUG_ON(btree_node_write_in_flight(b));
45 kvpfree(b->data, btree_bytes(c));
51 static void btree_node_data_free(struct bch_fs *c, struct btree *b)
53 struct btree_cache *bc = &c->btree_cache;
55 __btree_node_data_free(c, b);
57 list_move(&b->list, &bc->freed);
60 static int bch2_btree_cache_cmp_fn(struct rhashtable_compare_arg *arg,
63 const struct btree *b = obj;
64 const u64 *v = arg->key;
66 return b->hash_val == *v ? 0 : 1;
69 static const struct rhashtable_params bch_btree_cache_params = {
70 .head_offset = offsetof(struct btree, hash),
71 .key_offset = offsetof(struct btree, hash_val),
72 .key_len = sizeof(u64),
73 .obj_cmpfn = bch2_btree_cache_cmp_fn,
76 static int btree_node_data_alloc(struct bch_fs *c, struct btree *b, gfp_t gfp)
78 BUG_ON(b->data || b->aux_data);
80 b->data = kvpmalloc(btree_bytes(c), gfp);
84 b->aux_data = vmalloc_exec(btree_aux_data_bytes(b), gfp);
86 kvpfree(b->data, btree_bytes(c));
94 static struct btree *__btree_node_mem_alloc(struct bch_fs *c)
96 struct btree *b = kzalloc(sizeof(struct btree), GFP_KERNEL);
100 bkey_btree_ptr_init(&b->key);
101 six_lock_init(&b->c.lock);
102 INIT_LIST_HEAD(&b->list);
103 INIT_LIST_HEAD(&b->write_blocked);
104 b->byte_order = ilog2(btree_bytes(c));
108 static struct btree *btree_node_mem_alloc(struct bch_fs *c)
110 struct btree_cache *bc = &c->btree_cache;
111 struct btree *b = __btree_node_mem_alloc(c);
115 if (btree_node_data_alloc(c, b, GFP_KERNEL)) {
121 list_add(&b->list, &bc->freeable);
125 /* Btree in memory cache - hash table */
127 void bch2_btree_node_hash_remove(struct btree_cache *bc, struct btree *b)
129 rhashtable_remove_fast(&bc->table, &b->hash, bch_btree_cache_params);
131 /* Cause future lookups for this node to fail: */
134 six_lock_wakeup_all(&b->c.lock);
137 int __bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b)
140 b->hash_val = btree_ptr_hash_val(&b->key);
142 return rhashtable_lookup_insert_fast(&bc->table, &b->hash,
143 bch_btree_cache_params);
146 int bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b,
147 unsigned level, enum btree_id id)
154 mutex_lock(&bc->lock);
155 ret = __bch2_btree_node_hash_insert(bc, b);
157 list_add(&b->list, &bc->live);
158 mutex_unlock(&bc->lock);
164 static inline struct btree *btree_cache_find(struct btree_cache *bc,
165 const struct bkey_i *k)
167 u64 v = btree_ptr_hash_val(k);
169 return rhashtable_lookup_fast(&bc->table, &v, bch_btree_cache_params);
173 * this version is for btree nodes that have already been freed (we're not
174 * reaping a real btree node)
176 static int __btree_node_reclaim(struct bch_fs *c, struct btree *b, bool flush)
178 struct btree_cache *bc = &c->btree_cache;
181 lockdep_assert_held(&bc->lock);
183 if (!six_trylock_intent(&b->c.lock))
186 if (!six_trylock_write(&b->c.lock))
187 goto out_unlock_intent;
189 if (btree_node_noevict(b))
192 if (!btree_node_may_write(b))
195 if (btree_node_dirty(b) &&
196 test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags))
199 if (btree_node_dirty(b) ||
200 btree_node_write_in_flight(b) ||
201 btree_node_read_in_flight(b)) {
205 wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
206 TASK_UNINTERRUPTIBLE);
209 * Using the underscore version because we don't want to compact
210 * bsets after the write, since this node is about to be evicted
211 * - unless btree verify mode is enabled, since it runs out of
212 * the post write cleanup:
214 if (bch2_verify_btree_ondisk)
215 bch2_btree_node_write(c, b, SIX_LOCK_intent);
217 __bch2_btree_node_write(c, b, SIX_LOCK_read);
219 /* wait for any in flight btree write */
220 btree_node_wait_on_io(b);
223 if (b->hash_val && !ret)
224 trace_btree_node_reap(c, b);
227 six_unlock_write(&b->c.lock);
229 six_unlock_intent(&b->c.lock);
234 static int btree_node_reclaim(struct bch_fs *c, struct btree *b)
236 return __btree_node_reclaim(c, b, false);
239 static int btree_node_write_and_reclaim(struct bch_fs *c, struct btree *b)
241 return __btree_node_reclaim(c, b, true);
244 static unsigned long bch2_btree_cache_scan(struct shrinker *shrink,
245 struct shrink_control *sc)
247 struct bch_fs *c = container_of(shrink, struct bch_fs,
249 struct btree_cache *bc = &c->btree_cache;
251 unsigned long nr = sc->nr_to_scan;
252 unsigned long can_free;
253 unsigned long touched = 0;
254 unsigned long freed = 0;
257 if (bch2_btree_shrinker_disabled)
260 /* Return -1 if we can't do anything right now */
261 if (sc->gfp_mask & __GFP_FS)
262 mutex_lock(&bc->lock);
263 else if (!mutex_trylock(&bc->lock))
266 flags = memalloc_nofs_save();
269 * It's _really_ critical that we don't free too many btree nodes - we
270 * have to always leave ourselves a reserve. The reserve is how we
271 * guarantee that allocating memory for a new btree node can always
272 * succeed, so that inserting keys into the btree can always succeed and
273 * IO can always make forward progress:
275 nr /= btree_pages(c);
276 can_free = btree_cache_can_free(bc);
277 nr = min_t(unsigned long, nr, can_free);
280 list_for_each_entry_safe(b, t, &bc->freeable, list) {
287 !btree_node_reclaim(c, b)) {
288 btree_node_data_free(c, b);
289 six_unlock_write(&b->c.lock);
290 six_unlock_intent(&b->c.lock);
295 list_for_each_entry_safe(b, t, &bc->live, list) {
300 if (&t->list != &bc->live)
301 list_move_tail(&bc->live, &t->list);
305 if (!btree_node_accessed(b) &&
306 !btree_node_reclaim(c, b)) {
307 /* can't call bch2_btree_node_hash_remove under lock */
309 if (&t->list != &bc->live)
310 list_move_tail(&bc->live, &t->list);
312 btree_node_data_free(c, b);
313 mutex_unlock(&bc->lock);
315 bch2_btree_node_hash_remove(bc, b);
316 six_unlock_write(&b->c.lock);
317 six_unlock_intent(&b->c.lock);
322 if (sc->gfp_mask & __GFP_FS)
323 mutex_lock(&bc->lock);
324 else if (!mutex_trylock(&bc->lock))
328 clear_btree_node_accessed(b);
331 mutex_unlock(&bc->lock);
333 memalloc_nofs_restore(flags);
334 return (unsigned long) freed * btree_pages(c);
337 static unsigned long bch2_btree_cache_count(struct shrinker *shrink,
338 struct shrink_control *sc)
340 struct bch_fs *c = container_of(shrink, struct bch_fs,
342 struct btree_cache *bc = &c->btree_cache;
344 if (bch2_btree_shrinker_disabled)
347 return btree_cache_can_free(bc) * btree_pages(c);
350 void bch2_fs_btree_cache_exit(struct bch_fs *c)
352 struct btree_cache *bc = &c->btree_cache;
356 if (bc->shrink.list.next)
357 unregister_shrinker(&bc->shrink);
359 /* vfree() can allocate memory: */
360 flags = memalloc_nofs_save();
361 mutex_lock(&bc->lock);
363 #ifdef CONFIG_BCACHEFS_DEBUG
365 list_move(&c->verify_data->list, &bc->live);
367 kvpfree(c->verify_ondisk, btree_bytes(c));
370 for (i = 0; i < BTREE_ID_NR; i++)
371 if (c->btree_roots[i].b)
372 list_add(&c->btree_roots[i].b->list, &bc->live);
374 list_splice(&bc->freeable, &bc->live);
376 while (!list_empty(&bc->live)) {
377 b = list_first_entry(&bc->live, struct btree, list);
379 BUG_ON(btree_node_read_in_flight(b) ||
380 btree_node_write_in_flight(b));
382 if (btree_node_dirty(b))
383 bch2_btree_complete_write(c, b, btree_current_write(b));
384 clear_btree_node_dirty(c, b);
386 btree_node_data_free(c, b);
389 BUG_ON(atomic_read(&c->btree_cache.dirty));
391 while (!list_empty(&bc->freed)) {
392 b = list_first_entry(&bc->freed, struct btree, list);
397 mutex_unlock(&bc->lock);
398 memalloc_nofs_restore(flags);
400 if (bc->table_init_done)
401 rhashtable_destroy(&bc->table);
404 int bch2_fs_btree_cache_init(struct bch_fs *c)
406 struct btree_cache *bc = &c->btree_cache;
410 pr_verbose_init(c->opts, "");
412 ret = rhashtable_init(&bc->table, &bch_btree_cache_params);
416 bc->table_init_done = true;
418 bch2_recalc_btree_reserve(c);
420 for (i = 0; i < bc->reserve; i++)
421 if (!btree_node_mem_alloc(c)) {
426 list_splice_init(&bc->live, &bc->freeable);
428 #ifdef CONFIG_BCACHEFS_DEBUG
429 mutex_init(&c->verify_lock);
431 c->verify_ondisk = kvpmalloc(btree_bytes(c), GFP_KERNEL);
432 if (!c->verify_ondisk) {
437 c->verify_data = btree_node_mem_alloc(c);
438 if (!c->verify_data) {
443 list_del_init(&c->verify_data->list);
446 bc->shrink.count_objects = bch2_btree_cache_count;
447 bc->shrink.scan_objects = bch2_btree_cache_scan;
448 bc->shrink.seeks = 4;
449 bc->shrink.batch = btree_pages(c) * 2;
450 ret = register_shrinker(&bc->shrink);
452 pr_verbose_init(c->opts, "ret %i", ret);
456 void bch2_fs_btree_cache_init_early(struct btree_cache *bc)
458 mutex_init(&bc->lock);
459 INIT_LIST_HEAD(&bc->live);
460 INIT_LIST_HEAD(&bc->freeable);
461 INIT_LIST_HEAD(&bc->freed);
465 * We can only have one thread cannibalizing other cached btree nodes at a time,
466 * or we'll deadlock. We use an open coded mutex to ensure that, which a
467 * cannibalize_bucket() will take. This means every time we unlock the root of
468 * the btree, we need to release this lock if we have it held.
470 void bch2_btree_cache_cannibalize_unlock(struct bch_fs *c)
472 struct btree_cache *bc = &c->btree_cache;
474 if (bc->alloc_lock == current) {
475 trace_btree_node_cannibalize_unlock(c);
476 bc->alloc_lock = NULL;
477 closure_wake_up(&bc->alloc_wait);
481 int bch2_btree_cache_cannibalize_lock(struct bch_fs *c, struct closure *cl)
483 struct btree_cache *bc = &c->btree_cache;
484 struct task_struct *old;
486 old = cmpxchg(&bc->alloc_lock, NULL, current);
487 if (old == NULL || old == current)
491 trace_btree_node_cannibalize_lock_fail(c);
495 closure_wait(&bc->alloc_wait, cl);
497 /* Try again, after adding ourselves to waitlist */
498 old = cmpxchg(&bc->alloc_lock, NULL, current);
499 if (old == NULL || old == current) {
501 closure_wake_up(&bc->alloc_wait);
505 trace_btree_node_cannibalize_lock_fail(c);
509 trace_btree_node_cannibalize_lock(c);
513 static struct btree *btree_node_cannibalize(struct bch_fs *c)
515 struct btree_cache *bc = &c->btree_cache;
518 list_for_each_entry_reverse(b, &bc->live, list)
519 if (!btree_node_reclaim(c, b))
523 list_for_each_entry_reverse(b, &bc->live, list)
524 if (!btree_node_write_and_reclaim(c, b))
528 * Rare case: all nodes were intent-locked.
531 WARN_ONCE(1, "btree cache cannibalize failed\n");
536 struct btree *bch2_btree_node_mem_alloc(struct bch_fs *c)
538 struct btree_cache *bc = &c->btree_cache;
540 u64 start_time = local_clock();
543 flags = memalloc_nofs_save();
544 mutex_lock(&bc->lock);
547 * btree_free() doesn't free memory; it sticks the node on the end of
548 * the list. Check if there's any freed nodes there:
550 list_for_each_entry(b, &bc->freeable, list)
551 if (!btree_node_reclaim(c, b))
555 * We never free struct btree itself, just the memory that holds the on
556 * disk node. Check the freed list before allocating a new one:
558 list_for_each_entry(b, &bc->freed, list)
559 if (!btree_node_reclaim(c, b))
565 list_del_init(&b->list);
566 mutex_unlock(&bc->lock);
569 b = __btree_node_mem_alloc(c);
573 BUG_ON(!six_trylock_intent(&b->c.lock));
574 BUG_ON(!six_trylock_write(&b->c.lock));
578 if (btree_node_data_alloc(c, b, __GFP_NOWARN|GFP_KERNEL))
581 mutex_lock(&bc->lock);
583 mutex_unlock(&bc->lock);
586 BUG_ON(btree_node_hashed(b));
587 BUG_ON(btree_node_write_in_flight(b));
594 b->whiteout_u64s = 0;
595 bch2_btree_keys_init(b);
597 bch2_time_stats_update(&c->times[BCH_TIME_btree_node_mem_alloc],
600 memalloc_nofs_restore(flags);
603 mutex_lock(&bc->lock);
606 list_add(&b->list, &bc->freed);
607 six_unlock_write(&b->c.lock);
608 six_unlock_intent(&b->c.lock);
611 /* Try to cannibalize another cached btree node: */
612 if (bc->alloc_lock == current) {
613 b = btree_node_cannibalize(c);
614 list_del_init(&b->list);
615 mutex_unlock(&bc->lock);
617 bch2_btree_node_hash_remove(bc, b);
619 trace_btree_node_cannibalize(c);
623 mutex_unlock(&bc->lock);
624 memalloc_nofs_restore(flags);
625 return ERR_PTR(-ENOMEM);
628 /* Slowpath, don't want it inlined into btree_iter_traverse() */
629 static noinline struct btree *bch2_btree_node_fill(struct bch_fs *c,
630 struct btree_iter *iter,
631 const struct bkey_i *k,
632 enum btree_id btree_id,
634 enum six_lock_type lock_type,
637 struct btree_cache *bc = &c->btree_cache;
640 BUG_ON(level + 1 >= BTREE_MAX_DEPTH);
642 * Parent node must be locked, else we could read in a btree node that's
645 if (iter && !bch2_btree_node_relock(iter, level + 1))
646 return ERR_PTR(-EINTR);
648 b = bch2_btree_node_mem_alloc(c);
652 bkey_copy(&b->key, k);
653 if (bch2_btree_node_hash_insert(bc, b, level, btree_id)) {
654 /* raced with another fill: */
656 /* mark as unhashed... */
659 mutex_lock(&bc->lock);
660 list_add(&b->list, &bc->freeable);
661 mutex_unlock(&bc->lock);
663 six_unlock_write(&b->c.lock);
664 six_unlock_intent(&b->c.lock);
669 * Unlock before doing IO:
671 * XXX: ideally should be dropping all btree node locks here
673 if (iter && btree_node_read_locked(iter, level + 1))
674 btree_node_unlock(iter, level + 1);
676 bch2_btree_node_read(c, b, sync);
678 six_unlock_write(&b->c.lock);
681 six_unlock_intent(&b->c.lock);
685 if (lock_type == SIX_LOCK_read)
686 six_lock_downgrade(&b->c.lock);
691 static int lock_node_check_fn(struct six_lock *lock, void *p)
693 struct btree *b = container_of(lock, struct btree, c.lock);
694 const struct bkey_i *k = p;
696 return b->hash_val == btree_ptr_hash_val(k) ? 0 : -1;
700 * bch_btree_node_get - find a btree node in the cache and lock it, reading it
701 * in from disk if necessary.
703 * If IO is necessary and running under generic_make_request, returns -EAGAIN.
705 * The btree node will have either a read or a write lock held, depending on
706 * the @write parameter.
708 struct btree *bch2_btree_node_get(struct bch_fs *c, struct btree_iter *iter,
709 const struct bkey_i *k, unsigned level,
710 enum six_lock_type lock_type,
711 unsigned long trace_ip)
713 struct btree_cache *bc = &c->btree_cache;
717 EBUG_ON(level >= BTREE_MAX_DEPTH);
719 b = btree_node_mem_ptr(k);
723 b = btree_cache_find(bc, k);
726 * We must have the parent locked to call bch2_btree_node_fill(),
727 * else we could read in a btree node from disk that's been
730 b = bch2_btree_node_fill(c, iter, k, iter->btree_id,
731 level, lock_type, true);
733 /* We raced and found the btree node in the cache */
742 * There's a potential deadlock with splits and insertions into
743 * interior nodes we have to avoid:
745 * The other thread might be holding an intent lock on the node
746 * we want, and they want to update its parent node so they're
747 * going to upgrade their intent lock on the parent node to a
750 * But if we're holding a read lock on the parent, and we're
751 * trying to get the intent lock they're holding, we deadlock.
753 * So to avoid this we drop the read locks on parent nodes when
754 * we're starting to take intent locks - and handle the race.
756 * The race is that they might be about to free the node we
757 * want, and dropping our read lock on the parent node lets them
758 * update the parent marking the node we want as freed, and then
761 * To guard against this, btree nodes are evicted from the cache
762 * when they're freed - and b->hash_val is zeroed out, which we
763 * check for after we lock the node.
765 * Then, bch2_btree_node_relock() on the parent will fail - because
766 * the parent was modified, when the pointer to the node we want
767 * was removed - and we'll bail out:
769 if (btree_node_read_locked(iter, level + 1))
770 btree_node_unlock(iter, level + 1);
772 if (!btree_node_lock(b, k->k.p, level, iter, lock_type,
773 lock_node_check_fn, (void *) k, trace_ip)) {
774 if (b->hash_val != btree_ptr_hash_val(k))
776 return ERR_PTR(-EINTR);
779 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
780 b->c.level != level ||
782 six_unlock_type(&b->c.lock, lock_type);
783 if (bch2_btree_node_relock(iter, level + 1))
786 trace_trans_restart_btree_node_reused(iter->trans->ip);
787 return ERR_PTR(-EINTR);
791 /* XXX: waiting on IO with btree locks held: */
792 wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
793 TASK_UNINTERRUPTIBLE);
795 prefetch(b->aux_data);
797 for_each_bset(b, t) {
798 void *p = (u64 *) b->aux_data + t->aux_data_offset;
800 prefetch(p + L1_CACHE_BYTES * 0);
801 prefetch(p + L1_CACHE_BYTES * 1);
802 prefetch(p + L1_CACHE_BYTES * 2);
805 /* avoid atomic set bit if it's not needed: */
806 if (!btree_node_accessed(b))
807 set_btree_node_accessed(b);
809 if (unlikely(btree_node_read_error(b))) {
810 six_unlock_type(&b->c.lock, lock_type);
811 return ERR_PTR(-EIO);
814 EBUG_ON(b->c.btree_id != iter->btree_id ||
815 BTREE_NODE_LEVEL(b->data) != level ||
816 bkey_cmp(b->data->max_key, k->k.p));
821 struct btree *bch2_btree_node_get_noiter(struct bch_fs *c,
822 const struct bkey_i *k,
823 enum btree_id btree_id,
826 struct btree_cache *bc = &c->btree_cache;
831 EBUG_ON(level >= BTREE_MAX_DEPTH);
833 b = btree_node_mem_ptr(k);
837 b = btree_cache_find(bc, k);
839 b = bch2_btree_node_fill(c, NULL, k, btree_id,
840 level, SIX_LOCK_read, true);
842 /* We raced and found the btree node in the cache */
850 ret = six_lock_read(&b->c.lock, lock_node_check_fn, (void *) k);
854 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
855 b->c.btree_id != btree_id ||
856 b->c.level != level)) {
857 six_unlock_read(&b->c.lock);
862 /* XXX: waiting on IO with btree locks held: */
863 wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
864 TASK_UNINTERRUPTIBLE);
866 prefetch(b->aux_data);
868 for_each_bset(b, t) {
869 void *p = (u64 *) b->aux_data + t->aux_data_offset;
871 prefetch(p + L1_CACHE_BYTES * 0);
872 prefetch(p + L1_CACHE_BYTES * 1);
873 prefetch(p + L1_CACHE_BYTES * 2);
876 /* avoid atomic set bit if it's not needed: */
877 if (!btree_node_accessed(b))
878 set_btree_node_accessed(b);
880 if (unlikely(btree_node_read_error(b))) {
881 six_unlock_read(&b->c.lock);
882 return ERR_PTR(-EIO);
885 EBUG_ON(b->c.btree_id != btree_id ||
886 BTREE_NODE_LEVEL(b->data) != level ||
887 bkey_cmp(b->data->max_key, k->k.p));
892 struct btree *bch2_btree_node_get_sibling(struct bch_fs *c,
893 struct btree_iter *iter,
895 enum btree_node_sibling sib)
897 struct btree_trans *trans = iter->trans;
898 struct btree *parent;
899 struct btree_node_iter node_iter;
900 struct bkey_packed *k;
902 struct btree *ret = NULL;
903 unsigned level = b->c.level;
905 parent = btree_iter_node(iter, level + 1);
910 * There's a corner case where a btree_iter might have a node locked
911 * that is just outside its current pos - when
912 * bch2_btree_iter_set_pos_same_leaf() gets to the end of the node.
914 * But the lock ordering checks in __bch2_btree_node_lock() go off of
915 * iter->pos, not the node's key: so if the iterator is marked as
916 * needing to be traversed, we risk deadlock if we don't bail out here:
918 if (iter->uptodate >= BTREE_ITER_NEED_TRAVERSE)
919 return ERR_PTR(-EINTR);
921 if (!bch2_btree_node_relock(iter, level + 1)) {
922 ret = ERR_PTR(-EINTR);
926 node_iter = iter->l[parent->c.level].iter;
928 k = bch2_btree_node_iter_peek_all(&node_iter, parent);
929 BUG_ON(bkey_cmp_left_packed(parent, k, &b->key.k.p));
931 k = sib == btree_prev_sib
932 ? bch2_btree_node_iter_prev(&node_iter, parent)
933 : (bch2_btree_node_iter_advance(&node_iter, parent),
934 bch2_btree_node_iter_peek(&node_iter, parent));
938 bch2_bkey_unpack(parent, &tmp.k, k);
940 ret = bch2_btree_node_get(c, iter, &tmp.k, level,
941 SIX_LOCK_intent, _THIS_IP_);
943 if (PTR_ERR_OR_ZERO(ret) == -EINTR && !trans->nounlock) {
944 struct btree_iter *linked;
946 if (!bch2_btree_node_relock(iter, level + 1))
950 * We might have got -EINTR because trylock failed, and we're
951 * holding other locks that would cause us to deadlock:
953 trans_for_each_iter(trans, linked)
954 if (btree_iter_lock_cmp(iter, linked) < 0)
955 __bch2_btree_iter_unlock(linked);
957 if (sib == btree_prev_sib)
958 btree_node_unlock(iter, level);
960 ret = bch2_btree_node_get(c, iter, &tmp.k, level,
961 SIX_LOCK_intent, _THIS_IP_);
964 * before btree_iter_relock() calls btree_iter_verify_locks():
966 if (btree_lock_want(iter, level + 1) == BTREE_NODE_UNLOCKED)
967 btree_node_unlock(iter, level + 1);
969 if (!bch2_btree_node_relock(iter, level)) {
970 btree_iter_set_dirty(iter, BTREE_ITER_NEED_RELOCK);
973 six_unlock_intent(&ret->c.lock);
974 ret = ERR_PTR(-EINTR);
978 bch2_trans_relock(trans);
981 if (btree_lock_want(iter, level + 1) == BTREE_NODE_UNLOCKED)
982 btree_node_unlock(iter, level + 1);
984 if (PTR_ERR_OR_ZERO(ret) == -EINTR)
985 bch2_btree_iter_upgrade(iter, level + 2);
987 BUG_ON(!IS_ERR(ret) && !btree_node_locked(iter, level));
989 if (!IS_ERR_OR_NULL(ret)) {
990 struct btree *n1 = ret, *n2 = b;
992 if (sib != btree_prev_sib)
995 BUG_ON(bkey_cmp(bkey_successor(n1->key.k.p),
999 bch2_btree_trans_verify_locks(trans);
1004 void bch2_btree_node_prefetch(struct bch_fs *c, struct btree_iter *iter,
1005 const struct bkey_i *k, unsigned level)
1007 struct btree_cache *bc = &c->btree_cache;
1010 BUG_ON(!btree_node_locked(iter, level + 1));
1011 BUG_ON(level >= BTREE_MAX_DEPTH);
1013 b = btree_cache_find(bc, k);
1017 bch2_btree_node_fill(c, iter, k, iter->btree_id,
1018 level, SIX_LOCK_read, false);
1021 void bch2_btree_node_to_text(struct printbuf *out, struct bch_fs *c,
1024 const struct bkey_format *f = &b->format;
1025 struct bset_stats stats;
1027 memset(&stats, 0, sizeof(stats));
1029 bch2_btree_keys_stats(b, &stats);
1032 "l %u %llu:%llu - %llu:%llu:\n"
1035 b->data->min_key.inode,
1036 b->data->min_key.offset,
1037 b->data->max_key.inode,
1038 b->data->max_key.offset);
1039 bch2_val_to_text(out, c, bkey_i_to_s_c(&b->key));
1041 " format: u64s %u fields %u %u %u %u %u\n"
1042 " unpack fn len: %u\n"
1043 " bytes used %zu/%zu (%zu%% full)\n"
1044 " sib u64s: %u, %u (merge threshold %zu)\n"
1045 " nr packed keys %u\n"
1046 " nr unpacked keys %u\n"
1048 " failed unpacked %zu\n",
1050 f->bits_per_field[0],
1051 f->bits_per_field[1],
1052 f->bits_per_field[2],
1053 f->bits_per_field[3],
1054 f->bits_per_field[4],
1056 b->nr.live_u64s * sizeof(u64),
1057 btree_bytes(c) - sizeof(struct btree_node),
1058 b->nr.live_u64s * 100 / btree_max_u64s(c),
1061 BTREE_FOREGROUND_MERGE_THRESHOLD(c),
1063 b->nr.unpacked_keys,