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 (verify_btree_ondisk(c))
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 (btree_shrinker_disabled(c))
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))
267 * It's _really_ critical that we don't free too many btree nodes - we
268 * have to always leave ourselves a reserve. The reserve is how we
269 * guarantee that allocating memory for a new btree node can always
270 * succeed, so that inserting keys into the btree can always succeed and
271 * IO can always make forward progress:
273 nr /= btree_pages(c);
274 can_free = btree_cache_can_free(bc);
275 nr = min_t(unsigned long, nr, can_free);
278 list_for_each_entry_safe(b, t, &bc->freeable, list) {
285 !btree_node_reclaim(c, b)) {
286 btree_node_data_free(c, b);
287 six_unlock_write(&b->c.lock);
288 six_unlock_intent(&b->c.lock);
293 list_for_each_entry_safe(b, t, &bc->live, list) {
298 if (&t->list != &bc->live)
299 list_move_tail(&bc->live, &t->list);
303 if (!btree_node_accessed(b) &&
304 !btree_node_reclaim(c, b)) {
305 /* can't call bch2_btree_node_hash_remove under lock */
307 if (&t->list != &bc->live)
308 list_move_tail(&bc->live, &t->list);
310 btree_node_data_free(c, b);
311 mutex_unlock(&bc->lock);
313 bch2_btree_node_hash_remove(bc, b);
314 six_unlock_write(&b->c.lock);
315 six_unlock_intent(&b->c.lock);
320 if (sc->gfp_mask & __GFP_FS)
321 mutex_lock(&bc->lock);
322 else if (!mutex_trylock(&bc->lock))
326 clear_btree_node_accessed(b);
329 mutex_unlock(&bc->lock);
331 return (unsigned long) freed * btree_pages(c);
334 static unsigned long bch2_btree_cache_count(struct shrinker *shrink,
335 struct shrink_control *sc)
337 struct bch_fs *c = container_of(shrink, struct bch_fs,
339 struct btree_cache *bc = &c->btree_cache;
341 if (btree_shrinker_disabled(c))
344 return btree_cache_can_free(bc) * btree_pages(c);
347 void bch2_fs_btree_cache_exit(struct bch_fs *c)
349 struct btree_cache *bc = &c->btree_cache;
353 if (bc->shrink.list.next)
354 unregister_shrinker(&bc->shrink);
356 mutex_lock(&bc->lock);
358 #ifdef CONFIG_BCACHEFS_DEBUG
360 list_move(&c->verify_data->list, &bc->live);
362 kvpfree(c->verify_ondisk, btree_bytes(c));
365 for (i = 0; i < BTREE_ID_NR; i++)
366 if (c->btree_roots[i].b)
367 list_add(&c->btree_roots[i].b->list, &bc->live);
369 list_splice(&bc->freeable, &bc->live);
371 while (!list_empty(&bc->live)) {
372 b = list_first_entry(&bc->live, struct btree, list);
374 BUG_ON(btree_node_read_in_flight(b) ||
375 btree_node_write_in_flight(b));
377 if (btree_node_dirty(b))
378 bch2_btree_complete_write(c, b, btree_current_write(b));
379 clear_btree_node_dirty(b);
381 btree_node_data_free(c, b);
384 while (!list_empty(&bc->freed)) {
385 b = list_first_entry(&bc->freed, struct btree, list);
390 mutex_unlock(&bc->lock);
392 if (bc->table_init_done)
393 rhashtable_destroy(&bc->table);
396 int bch2_fs_btree_cache_init(struct bch_fs *c)
398 struct btree_cache *bc = &c->btree_cache;
402 pr_verbose_init(c->opts, "");
404 ret = rhashtable_init(&bc->table, &bch_btree_cache_params);
408 bc->table_init_done = true;
410 bch2_recalc_btree_reserve(c);
412 for (i = 0; i < bc->reserve; i++)
413 if (!btree_node_mem_alloc(c)) {
418 list_splice_init(&bc->live, &bc->freeable);
420 #ifdef CONFIG_BCACHEFS_DEBUG
421 mutex_init(&c->verify_lock);
423 c->verify_ondisk = kvpmalloc(btree_bytes(c), GFP_KERNEL);
424 if (!c->verify_ondisk) {
429 c->verify_data = btree_node_mem_alloc(c);
430 if (!c->verify_data) {
435 list_del_init(&c->verify_data->list);
438 bc->shrink.count_objects = bch2_btree_cache_count;
439 bc->shrink.scan_objects = bch2_btree_cache_scan;
440 bc->shrink.seeks = 4;
441 bc->shrink.batch = btree_pages(c) * 2;
442 register_shrinker(&bc->shrink);
444 pr_verbose_init(c->opts, "ret %i", ret);
448 void bch2_fs_btree_cache_init_early(struct btree_cache *bc)
450 mutex_init(&bc->lock);
451 INIT_LIST_HEAD(&bc->live);
452 INIT_LIST_HEAD(&bc->freeable);
453 INIT_LIST_HEAD(&bc->freed);
457 * We can only have one thread cannibalizing other cached btree nodes at a time,
458 * or we'll deadlock. We use an open coded mutex to ensure that, which a
459 * cannibalize_bucket() will take. This means every time we unlock the root of
460 * the btree, we need to release this lock if we have it held.
462 void bch2_btree_cache_cannibalize_unlock(struct bch_fs *c)
464 struct btree_cache *bc = &c->btree_cache;
466 if (bc->alloc_lock == current) {
467 trace_btree_node_cannibalize_unlock(c);
468 bc->alloc_lock = NULL;
469 closure_wake_up(&bc->alloc_wait);
473 int bch2_btree_cache_cannibalize_lock(struct bch_fs *c, struct closure *cl)
475 struct btree_cache *bc = &c->btree_cache;
476 struct task_struct *old;
478 old = cmpxchg(&bc->alloc_lock, NULL, current);
479 if (old == NULL || old == current)
483 trace_btree_node_cannibalize_lock_fail(c);
487 closure_wait(&bc->alloc_wait, cl);
489 /* Try again, after adding ourselves to waitlist */
490 old = cmpxchg(&bc->alloc_lock, NULL, current);
491 if (old == NULL || old == current) {
493 closure_wake_up(&bc->alloc_wait);
497 trace_btree_node_cannibalize_lock_fail(c);
501 trace_btree_node_cannibalize_lock(c);
505 static struct btree *btree_node_cannibalize(struct bch_fs *c)
507 struct btree_cache *bc = &c->btree_cache;
510 list_for_each_entry_reverse(b, &bc->live, list)
511 if (!btree_node_reclaim(c, b))
515 list_for_each_entry_reverse(b, &bc->live, list)
516 if (!btree_node_write_and_reclaim(c, b))
520 * Rare case: all nodes were intent-locked.
523 WARN_ONCE(1, "btree cache cannibalize failed\n");
528 struct btree *bch2_btree_node_mem_alloc(struct bch_fs *c)
530 struct btree_cache *bc = &c->btree_cache;
532 u64 start_time = local_clock();
535 flags = memalloc_nofs_save();
536 mutex_lock(&bc->lock);
539 * btree_free() doesn't free memory; it sticks the node on the end of
540 * the list. Check if there's any freed nodes there:
542 list_for_each_entry(b, &bc->freeable, list)
543 if (!btree_node_reclaim(c, b))
547 * We never free struct btree itself, just the memory that holds the on
548 * disk node. Check the freed list before allocating a new one:
550 list_for_each_entry(b, &bc->freed, list)
551 if (!btree_node_reclaim(c, b))
557 list_del_init(&b->list);
558 mutex_unlock(&bc->lock);
561 b = __btree_node_mem_alloc(c);
565 BUG_ON(!six_trylock_intent(&b->c.lock));
566 BUG_ON(!six_trylock_write(&b->c.lock));
570 if (btree_node_data_alloc(c, b, __GFP_NOWARN|GFP_KERNEL))
573 mutex_lock(&bc->lock);
575 mutex_unlock(&bc->lock);
578 BUG_ON(btree_node_hashed(b));
579 BUG_ON(btree_node_write_in_flight(b));
586 b->whiteout_u64s = 0;
587 bch2_btree_keys_init(b, &c->expensive_debug_checks);
589 bch2_time_stats_update(&c->times[BCH_TIME_btree_node_mem_alloc],
592 memalloc_nofs_restore(flags);
595 mutex_lock(&bc->lock);
598 list_add(&b->list, &bc->freed);
599 six_unlock_write(&b->c.lock);
600 six_unlock_intent(&b->c.lock);
603 /* Try to cannibalize another cached btree node: */
604 if (bc->alloc_lock == current) {
605 b = btree_node_cannibalize(c);
606 list_del_init(&b->list);
607 mutex_unlock(&bc->lock);
609 bch2_btree_node_hash_remove(bc, b);
611 trace_btree_node_cannibalize(c);
615 mutex_unlock(&bc->lock);
616 memalloc_nofs_restore(flags);
617 return ERR_PTR(-ENOMEM);
620 /* Slowpath, don't want it inlined into btree_iter_traverse() */
621 static noinline struct btree *bch2_btree_node_fill(struct bch_fs *c,
622 struct btree_iter *iter,
623 const struct bkey_i *k,
624 enum btree_id btree_id,
626 enum six_lock_type lock_type,
629 struct btree_cache *bc = &c->btree_cache;
632 BUG_ON(level + 1 >= BTREE_MAX_DEPTH);
634 * Parent node must be locked, else we could read in a btree node that's
637 if (iter && !bch2_btree_node_relock(iter, level + 1))
638 return ERR_PTR(-EINTR);
640 b = bch2_btree_node_mem_alloc(c);
644 bkey_copy(&b->key, k);
645 if (bch2_btree_node_hash_insert(bc, b, level, btree_id)) {
646 /* raced with another fill: */
648 /* mark as unhashed... */
651 mutex_lock(&bc->lock);
652 list_add(&b->list, &bc->freeable);
653 mutex_unlock(&bc->lock);
655 six_unlock_write(&b->c.lock);
656 six_unlock_intent(&b->c.lock);
661 * Unlock before doing IO:
663 * XXX: ideally should be dropping all btree node locks here
665 if (iter && btree_node_read_locked(iter, level + 1))
666 btree_node_unlock(iter, level + 1);
668 bch2_btree_node_read(c, b, sync);
670 six_unlock_write(&b->c.lock);
673 six_unlock_intent(&b->c.lock);
677 if (lock_type == SIX_LOCK_read)
678 six_lock_downgrade(&b->c.lock);
683 static int lock_node_check_fn(struct six_lock *lock, void *p)
685 struct btree *b = container_of(lock, struct btree, c.lock);
686 const struct bkey_i *k = p;
688 return b->hash_val == btree_ptr_hash_val(k) ? 0 : -1;
692 * bch_btree_node_get - find a btree node in the cache and lock it, reading it
693 * in from disk if necessary.
695 * If IO is necessary and running under generic_make_request, returns -EAGAIN.
697 * The btree node will have either a read or a write lock held, depending on
698 * the @write parameter.
700 struct btree *bch2_btree_node_get(struct bch_fs *c, struct btree_iter *iter,
701 const struct bkey_i *k, unsigned level,
702 enum six_lock_type lock_type)
704 struct btree_cache *bc = &c->btree_cache;
708 EBUG_ON(level >= BTREE_MAX_DEPTH);
710 b = btree_node_mem_ptr(k);
714 b = btree_cache_find(bc, k);
717 * We must have the parent locked to call bch2_btree_node_fill(),
718 * else we could read in a btree node from disk that's been
721 b = bch2_btree_node_fill(c, iter, k, iter->btree_id,
722 level, lock_type, true);
724 /* We raced and found the btree node in the cache */
733 * There's a potential deadlock with splits and insertions into
734 * interior nodes we have to avoid:
736 * The other thread might be holding an intent lock on the node
737 * we want, and they want to update its parent node so they're
738 * going to upgrade their intent lock on the parent node to a
741 * But if we're holding a read lock on the parent, and we're
742 * trying to get the intent lock they're holding, we deadlock.
744 * So to avoid this we drop the read locks on parent nodes when
745 * we're starting to take intent locks - and handle the race.
747 * The race is that they might be about to free the node we
748 * want, and dropping our read lock on the parent node lets them
749 * update the parent marking the node we want as freed, and then
752 * To guard against this, btree nodes are evicted from the cache
753 * when they're freed - and b->hash_val is zeroed out, which we
754 * check for after we lock the node.
756 * Then, bch2_btree_node_relock() on the parent will fail - because
757 * the parent was modified, when the pointer to the node we want
758 * was removed - and we'll bail out:
760 if (btree_node_read_locked(iter, level + 1))
761 btree_node_unlock(iter, level + 1);
763 if (!btree_node_lock(b, k->k.p, level, iter, lock_type,
764 lock_node_check_fn, (void *) k)) {
765 if (b->hash_val != btree_ptr_hash_val(k))
767 return ERR_PTR(-EINTR);
770 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
771 b->c.level != level ||
773 six_unlock_type(&b->c.lock, lock_type);
774 if (bch2_btree_node_relock(iter, level + 1))
777 trace_trans_restart_btree_node_reused(iter->trans->ip);
778 return ERR_PTR(-EINTR);
782 /* XXX: waiting on IO with btree locks held: */
783 wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
784 TASK_UNINTERRUPTIBLE);
786 prefetch(b->aux_data);
788 for_each_bset(b, t) {
789 void *p = (u64 *) b->aux_data + t->aux_data_offset;
791 prefetch(p + L1_CACHE_BYTES * 0);
792 prefetch(p + L1_CACHE_BYTES * 1);
793 prefetch(p + L1_CACHE_BYTES * 2);
796 /* avoid atomic set bit if it's not needed: */
797 if (!btree_node_accessed(b))
798 set_btree_node_accessed(b);
800 if (unlikely(btree_node_read_error(b))) {
801 six_unlock_type(&b->c.lock, lock_type);
802 return ERR_PTR(-EIO);
805 EBUG_ON(b->c.btree_id != iter->btree_id ||
806 BTREE_NODE_LEVEL(b->data) != level ||
807 bkey_cmp(b->data->max_key, k->k.p));
812 struct btree *bch2_btree_node_get_noiter(struct bch_fs *c,
813 const struct bkey_i *k,
814 enum btree_id btree_id,
817 struct btree_cache *bc = &c->btree_cache;
822 EBUG_ON(level >= BTREE_MAX_DEPTH);
824 b = btree_node_mem_ptr(k);
828 b = btree_cache_find(bc, k);
830 b = bch2_btree_node_fill(c, NULL, k, btree_id,
831 level, SIX_LOCK_read, true);
833 /* We raced and found the btree node in the cache */
841 ret = six_lock_read(&b->c.lock, lock_node_check_fn, (void *) k);
845 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
846 b->c.btree_id != btree_id ||
847 b->c.level != level)) {
848 six_unlock_read(&b->c.lock);
853 /* XXX: waiting on IO with btree locks held: */
854 wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
855 TASK_UNINTERRUPTIBLE);
857 prefetch(b->aux_data);
859 for_each_bset(b, t) {
860 void *p = (u64 *) b->aux_data + t->aux_data_offset;
862 prefetch(p + L1_CACHE_BYTES * 0);
863 prefetch(p + L1_CACHE_BYTES * 1);
864 prefetch(p + L1_CACHE_BYTES * 2);
867 /* avoid atomic set bit if it's not needed: */
868 if (!btree_node_accessed(b))
869 set_btree_node_accessed(b);
871 if (unlikely(btree_node_read_error(b))) {
872 six_unlock_read(&b->c.lock);
873 return ERR_PTR(-EIO);
876 EBUG_ON(b->c.btree_id != btree_id ||
877 BTREE_NODE_LEVEL(b->data) != level ||
878 bkey_cmp(b->data->max_key, k->k.p));
883 struct btree *bch2_btree_node_get_sibling(struct bch_fs *c,
884 struct btree_iter *iter,
886 enum btree_node_sibling sib)
888 struct btree_trans *trans = iter->trans;
889 struct btree *parent;
890 struct btree_node_iter node_iter;
891 struct bkey_packed *k;
893 struct btree *ret = NULL;
894 unsigned level = b->c.level;
896 parent = btree_iter_node(iter, level + 1);
901 * There's a corner case where a btree_iter might have a node locked
902 * that is just outside its current pos - when
903 * bch2_btree_iter_set_pos_same_leaf() gets to the end of the node.
905 * But the lock ordering checks in __bch2_btree_node_lock() go off of
906 * iter->pos, not the node's key: so if the iterator is marked as
907 * needing to be traversed, we risk deadlock if we don't bail out here:
909 if (iter->uptodate >= BTREE_ITER_NEED_TRAVERSE)
910 return ERR_PTR(-EINTR);
912 if (!bch2_btree_node_relock(iter, level + 1)) {
913 ret = ERR_PTR(-EINTR);
917 node_iter = iter->l[parent->c.level].iter;
919 k = bch2_btree_node_iter_peek_all(&node_iter, parent);
920 BUG_ON(bkey_cmp_left_packed(parent, k, &b->key.k.p));
922 k = sib == btree_prev_sib
923 ? bch2_btree_node_iter_prev(&node_iter, parent)
924 : (bch2_btree_node_iter_advance(&node_iter, parent),
925 bch2_btree_node_iter_peek(&node_iter, parent));
929 bch2_bkey_unpack(parent, &tmp.k, k);
931 ret = bch2_btree_node_get(c, iter, &tmp.k, level,
934 if (PTR_ERR_OR_ZERO(ret) == -EINTR && !trans->nounlock) {
935 struct btree_iter *linked;
937 if (!bch2_btree_node_relock(iter, level + 1))
941 * We might have got -EINTR because trylock failed, and we're
942 * holding other locks that would cause us to deadlock:
944 trans_for_each_iter(trans, linked)
945 if (btree_iter_cmp(iter, linked) < 0)
946 __bch2_btree_iter_unlock(linked);
948 if (sib == btree_prev_sib)
949 btree_node_unlock(iter, level);
951 ret = bch2_btree_node_get(c, iter, &tmp.k, level,
955 * before btree_iter_relock() calls btree_iter_verify_locks():
957 if (btree_lock_want(iter, level + 1) == BTREE_NODE_UNLOCKED)
958 btree_node_unlock(iter, level + 1);
960 if (!bch2_btree_node_relock(iter, level)) {
961 btree_iter_set_dirty(iter, BTREE_ITER_NEED_RELOCK);
964 six_unlock_intent(&ret->c.lock);
965 ret = ERR_PTR(-EINTR);
969 bch2_trans_relock(trans);
972 if (btree_lock_want(iter, level + 1) == BTREE_NODE_UNLOCKED)
973 btree_node_unlock(iter, level + 1);
975 if (PTR_ERR_OR_ZERO(ret) == -EINTR)
976 bch2_btree_iter_upgrade(iter, level + 2);
978 BUG_ON(!IS_ERR(ret) && !btree_node_locked(iter, level));
980 if (!IS_ERR_OR_NULL(ret)) {
981 struct btree *n1 = ret, *n2 = b;
983 if (sib != btree_prev_sib)
986 BUG_ON(bkey_cmp(bkey_successor(n1->key.k.p),
990 bch2_btree_trans_verify_locks(trans);
995 void bch2_btree_node_prefetch(struct bch_fs *c, struct btree_iter *iter,
996 const struct bkey_i *k, unsigned level)
998 struct btree_cache *bc = &c->btree_cache;
1001 BUG_ON(!btree_node_locked(iter, level + 1));
1002 BUG_ON(level >= BTREE_MAX_DEPTH);
1004 b = btree_cache_find(bc, k);
1008 bch2_btree_node_fill(c, iter, k, iter->btree_id,
1009 level, SIX_LOCK_read, false);
1012 void bch2_btree_node_to_text(struct printbuf *out, struct bch_fs *c,
1015 const struct bkey_format *f = &b->format;
1016 struct bset_stats stats;
1018 memset(&stats, 0, sizeof(stats));
1020 bch2_btree_keys_stats(b, &stats);
1023 "l %u %llu:%llu - %llu:%llu:\n"
1026 b->data->min_key.inode,
1027 b->data->min_key.offset,
1028 b->data->max_key.inode,
1029 b->data->max_key.offset);
1030 bch2_val_to_text(out, c, bkey_i_to_s_c(&b->key));
1032 " format: u64s %u fields %u %u %u %u %u\n"
1033 " unpack fn len: %u\n"
1034 " bytes used %zu/%zu (%zu%% full)\n"
1035 " sib u64s: %u, %u (merge threshold %zu)\n"
1036 " nr packed keys %u\n"
1037 " nr unpacked keys %u\n"
1039 " failed unpacked %zu\n",
1041 f->bits_per_field[0],
1042 f->bits_per_field[1],
1043 f->bits_per_field[2],
1044 f->bits_per_field[3],
1045 f->bits_per_field[4],
1047 b->nr.live_u64s * sizeof(u64),
1048 btree_bytes(c) - sizeof(struct btree_node),
1049 b->nr.live_u64s * 100 / btree_max_u64s(c),
1052 BTREE_FOREGROUND_MERGE_THRESHOLD(c),
1054 b->nr.unpacked_keys,