1 // SPDX-License-Identifier: GPL-2.0
5 #include "btree_cache.h"
7 #include "btree_iter.h"
8 #include "btree_locking.h"
15 #include <linux/prefetch.h>
16 #include <linux/sched/mm.h>
18 const char * const bch2_btree_node_flags[] = {
25 void bch2_recalc_btree_reserve(struct bch_fs *c)
27 unsigned i, reserve = 16;
29 if (!c->btree_roots_known[0].b)
32 for (i = 0; i < btree_id_nr_alive(c); i++) {
33 struct btree_root *r = bch2_btree_id_root(c, i);
36 reserve += min_t(unsigned, 1, r->b->c.level) * 8;
39 c->btree_cache.reserve = reserve;
42 static inline unsigned btree_cache_can_free(struct btree_cache *bc)
44 return max_t(int, 0, bc->used - bc->reserve);
47 static void btree_node_to_freedlist(struct btree_cache *bc, struct btree *b)
49 if (b->c.lock.readers)
50 list_move(&b->list, &bc->freed_pcpu);
52 list_move(&b->list, &bc->freed_nonpcpu);
55 static void btree_node_data_free(struct bch_fs *c, struct btree *b)
57 struct btree_cache *bc = &c->btree_cache;
59 EBUG_ON(btree_node_write_in_flight(b));
61 clear_btree_node_just_written(b);
63 kvpfree(b->data, btree_buf_bytes(b));
68 munmap(b->aux_data, btree_aux_data_bytes(b));
74 btree_node_to_freedlist(bc, b);
77 static int bch2_btree_cache_cmp_fn(struct rhashtable_compare_arg *arg,
80 const struct btree *b = obj;
81 const u64 *v = arg->key;
83 return b->hash_val == *v ? 0 : 1;
86 static const struct rhashtable_params bch_btree_cache_params = {
87 .head_offset = offsetof(struct btree, hash),
88 .key_offset = offsetof(struct btree, hash_val),
89 .key_len = sizeof(u64),
90 .obj_cmpfn = bch2_btree_cache_cmp_fn,
93 static int btree_node_data_alloc(struct bch_fs *c, struct btree *b, gfp_t gfp)
95 BUG_ON(b->data || b->aux_data);
97 b->data = kvpmalloc(btree_buf_bytes(b), gfp);
99 return -BCH_ERR_ENOMEM_btree_node_mem_alloc;
101 b->aux_data = kvmalloc(btree_aux_data_bytes(b), gfp);
103 b->aux_data = mmap(NULL, btree_aux_data_bytes(b),
104 PROT_READ|PROT_WRITE|PROT_EXEC,
105 MAP_PRIVATE|MAP_ANONYMOUS, 0, 0);
106 if (b->aux_data == MAP_FAILED)
110 kvpfree(b->data, btree_buf_bytes(b));
112 return -BCH_ERR_ENOMEM_btree_node_mem_alloc;
118 static struct btree *__btree_node_mem_alloc(struct bch_fs *c, gfp_t gfp)
122 b = kzalloc(sizeof(struct btree), gfp);
126 bkey_btree_ptr_init(&b->key);
127 INIT_LIST_HEAD(&b->list);
128 INIT_LIST_HEAD(&b->write_blocked);
129 b->byte_order = ilog2(c->opts.btree_node_size);
133 struct btree *__bch2_btree_node_mem_alloc(struct bch_fs *c)
135 struct btree_cache *bc = &c->btree_cache;
138 b = __btree_node_mem_alloc(c, GFP_KERNEL);
142 if (btree_node_data_alloc(c, b, GFP_KERNEL)) {
147 bch2_btree_lock_init(&b->c, 0);
150 list_add(&b->list, &bc->freeable);
154 /* Btree in memory cache - hash table */
156 void bch2_btree_node_hash_remove(struct btree_cache *bc, struct btree *b)
158 int ret = rhashtable_remove_fast(&bc->table, &b->hash, bch_btree_cache_params);
162 /* Cause future lookups for this node to fail: */
166 int __bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b)
169 b->hash_val = btree_ptr_hash_val(&b->key);
171 return rhashtable_lookup_insert_fast(&bc->table, &b->hash,
172 bch_btree_cache_params);
175 int bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b,
176 unsigned level, enum btree_id id)
183 mutex_lock(&bc->lock);
184 ret = __bch2_btree_node_hash_insert(bc, b);
186 list_add_tail(&b->list, &bc->live);
187 mutex_unlock(&bc->lock);
193 static inline struct btree *btree_cache_find(struct btree_cache *bc,
194 const struct bkey_i *k)
196 u64 v = btree_ptr_hash_val(k);
198 return rhashtable_lookup_fast(&bc->table, &v, bch_btree_cache_params);
202 * this version is for btree nodes that have already been freed (we're not
203 * reaping a real btree node)
205 static int __btree_node_reclaim(struct bch_fs *c, struct btree *b, bool flush)
207 struct btree_cache *bc = &c->btree_cache;
210 lockdep_assert_held(&bc->lock);
212 if (b->flags & ((1U << BTREE_NODE_dirty)|
213 (1U << BTREE_NODE_read_in_flight)|
214 (1U << BTREE_NODE_write_in_flight))) {
216 return -BCH_ERR_ENOMEM_btree_node_reclaim;
218 /* XXX: waiting on IO with btree cache lock held */
219 bch2_btree_node_wait_on_read(b);
220 bch2_btree_node_wait_on_write(b);
223 if (!six_trylock_intent(&b->c.lock))
224 return -BCH_ERR_ENOMEM_btree_node_reclaim;
226 if (!six_trylock_write(&b->c.lock))
227 goto out_unlock_intent;
229 /* recheck under lock */
230 if (b->flags & ((1U << BTREE_NODE_read_in_flight)|
231 (1U << BTREE_NODE_write_in_flight))) {
234 six_unlock_write(&b->c.lock);
235 six_unlock_intent(&b->c.lock);
239 if (btree_node_noevict(b) ||
240 btree_node_write_blocked(b) ||
241 btree_node_will_make_reachable(b))
244 if (btree_node_dirty(b)) {
248 * Using the underscore version because we don't want to compact
249 * bsets after the write, since this node is about to be evicted
250 * - unless btree verify mode is enabled, since it runs out of
251 * the post write cleanup:
253 if (bch2_verify_btree_ondisk)
254 bch2_btree_node_write(c, b, SIX_LOCK_intent,
255 BTREE_WRITE_cache_reclaim);
257 __bch2_btree_node_write(c, b,
258 BTREE_WRITE_cache_reclaim);
260 six_unlock_write(&b->c.lock);
261 six_unlock_intent(&b->c.lock);
265 if (b->hash_val && !ret)
266 trace_and_count(c, btree_cache_reap, c, b);
269 six_unlock_write(&b->c.lock);
271 six_unlock_intent(&b->c.lock);
272 ret = -BCH_ERR_ENOMEM_btree_node_reclaim;
276 static int btree_node_reclaim(struct bch_fs *c, struct btree *b)
278 return __btree_node_reclaim(c, b, false);
281 static int btree_node_write_and_reclaim(struct bch_fs *c, struct btree *b)
283 return __btree_node_reclaim(c, b, true);
286 static unsigned long bch2_btree_cache_scan(struct shrinker *shrink,
287 struct shrink_control *sc)
289 struct bch_fs *c = shrink->private_data;
290 struct btree_cache *bc = &c->btree_cache;
292 unsigned long nr = sc->nr_to_scan;
293 unsigned long can_free = 0;
294 unsigned long freed = 0;
295 unsigned long touched = 0;
297 unsigned long ret = SHRINK_STOP;
298 bool trigger_writes = atomic_read(&bc->dirty) + nr >=
301 if (bch2_btree_shrinker_disabled)
304 mutex_lock(&bc->lock);
305 flags = memalloc_nofs_save();
308 * It's _really_ critical that we don't free too many btree nodes - we
309 * have to always leave ourselves a reserve. The reserve is how we
310 * guarantee that allocating memory for a new btree node can always
311 * succeed, so that inserting keys into the btree can always succeed and
312 * IO can always make forward progress:
314 can_free = btree_cache_can_free(bc);
315 nr = min_t(unsigned long, nr, can_free);
318 list_for_each_entry_safe(b, t, &bc->freeable, list) {
320 * Leave a few nodes on the freeable list, so that a btree split
321 * won't have to hit the system allocator:
331 if (!btree_node_reclaim(c, b)) {
332 btree_node_data_free(c, b);
333 six_unlock_write(&b->c.lock);
334 six_unlock_intent(&b->c.lock);
339 list_for_each_entry_safe(b, t, &bc->live, list) {
342 if (btree_node_accessed(b)) {
343 clear_btree_node_accessed(b);
344 } else if (!btree_node_reclaim(c, b)) {
346 btree_node_data_free(c, b);
348 bch2_btree_node_hash_remove(bc, b);
349 six_unlock_write(&b->c.lock);
350 six_unlock_intent(&b->c.lock);
354 } else if (trigger_writes &&
355 btree_node_dirty(b) &&
356 !btree_node_will_make_reachable(b) &&
357 !btree_node_write_blocked(b) &&
358 six_trylock_read(&b->c.lock)) {
359 list_move(&bc->live, &b->list);
360 mutex_unlock(&bc->lock);
361 __bch2_btree_node_write(c, b, BTREE_WRITE_cache_reclaim);
362 six_unlock_read(&b->c.lock);
365 mutex_lock(&bc->lock);
373 if (&t->list != &bc->live)
374 list_move_tail(&bc->live, &t->list);
376 mutex_unlock(&bc->lock);
379 memalloc_nofs_restore(flags);
380 trace_and_count(c, btree_cache_scan, sc->nr_to_scan, can_free, ret);
384 static unsigned long bch2_btree_cache_count(struct shrinker *shrink,
385 struct shrink_control *sc)
387 struct bch_fs *c = shrink->private_data;
388 struct btree_cache *bc = &c->btree_cache;
390 if (bch2_btree_shrinker_disabled)
393 return btree_cache_can_free(bc);
396 void bch2_fs_btree_cache_exit(struct bch_fs *c)
398 struct btree_cache *bc = &c->btree_cache;
402 shrinker_free(bc->shrink);
404 /* vfree() can allocate memory: */
405 flags = memalloc_nofs_save();
406 mutex_lock(&bc->lock);
409 list_move(&c->verify_data->list, &bc->live);
411 kvpfree(c->verify_ondisk, c->opts.btree_node_size);
413 for (i = 0; i < btree_id_nr_alive(c); i++) {
414 struct btree_root *r = bch2_btree_id_root(c, i);
417 list_add(&r->b->list, &bc->live);
420 list_splice(&bc->freeable, &bc->live);
422 while (!list_empty(&bc->live)) {
423 b = list_first_entry(&bc->live, struct btree, list);
425 BUG_ON(btree_node_read_in_flight(b) ||
426 btree_node_write_in_flight(b));
428 btree_node_data_free(c, b);
431 BUG_ON(!bch2_journal_error(&c->journal) &&
432 atomic_read(&c->btree_cache.dirty));
434 list_splice(&bc->freed_pcpu, &bc->freed_nonpcpu);
436 while (!list_empty(&bc->freed_nonpcpu)) {
437 b = list_first_entry(&bc->freed_nonpcpu, struct btree, list);
439 six_lock_exit(&b->c.lock);
443 mutex_unlock(&bc->lock);
444 memalloc_nofs_restore(flags);
446 if (bc->table_init_done)
447 rhashtable_destroy(&bc->table);
450 int bch2_fs_btree_cache_init(struct bch_fs *c)
452 struct btree_cache *bc = &c->btree_cache;
453 struct shrinker *shrink;
457 ret = rhashtable_init(&bc->table, &bch_btree_cache_params);
461 bc->table_init_done = true;
463 bch2_recalc_btree_reserve(c);
465 for (i = 0; i < bc->reserve; i++)
466 if (!__bch2_btree_node_mem_alloc(c))
469 list_splice_init(&bc->live, &bc->freeable);
471 mutex_init(&c->verify_lock);
473 shrink = shrinker_alloc(0, "%s-btree_cache", c->name);
477 shrink->count_objects = bch2_btree_cache_count;
478 shrink->scan_objects = bch2_btree_cache_scan;
480 shrink->private_data = c;
481 shrinker_register(shrink);
485 return -BCH_ERR_ENOMEM_fs_btree_cache_init;
488 void bch2_fs_btree_cache_init_early(struct btree_cache *bc)
490 mutex_init(&bc->lock);
491 INIT_LIST_HEAD(&bc->live);
492 INIT_LIST_HEAD(&bc->freeable);
493 INIT_LIST_HEAD(&bc->freed_pcpu);
494 INIT_LIST_HEAD(&bc->freed_nonpcpu);
498 * We can only have one thread cannibalizing other cached btree nodes at a time,
499 * or we'll deadlock. We use an open coded mutex to ensure that, which a
500 * cannibalize_bucket() will take. This means every time we unlock the root of
501 * the btree, we need to release this lock if we have it held.
503 void bch2_btree_cache_cannibalize_unlock(struct btree_trans *trans)
505 struct bch_fs *c = trans->c;
506 struct btree_cache *bc = &c->btree_cache;
508 if (bc->alloc_lock == current) {
509 trace_and_count(c, btree_cache_cannibalize_unlock, trans);
510 bc->alloc_lock = NULL;
511 closure_wake_up(&bc->alloc_wait);
515 int bch2_btree_cache_cannibalize_lock(struct btree_trans *trans, struct closure *cl)
517 struct bch_fs *c = trans->c;
518 struct btree_cache *bc = &c->btree_cache;
519 struct task_struct *old;
521 old = cmpxchg(&bc->alloc_lock, NULL, current);
522 if (old == NULL || old == current)
526 trace_and_count(c, btree_cache_cannibalize_lock_fail, trans);
527 return -BCH_ERR_ENOMEM_btree_cache_cannibalize_lock;
530 closure_wait(&bc->alloc_wait, cl);
532 /* Try again, after adding ourselves to waitlist */
533 old = cmpxchg(&bc->alloc_lock, NULL, current);
534 if (old == NULL || old == current) {
536 closure_wake_up(&bc->alloc_wait);
540 trace_and_count(c, btree_cache_cannibalize_lock_fail, trans);
541 return -BCH_ERR_btree_cache_cannibalize_lock_blocked;
544 trace_and_count(c, btree_cache_cannibalize_lock, trans);
548 static struct btree *btree_node_cannibalize(struct bch_fs *c)
550 struct btree_cache *bc = &c->btree_cache;
553 list_for_each_entry_reverse(b, &bc->live, list)
554 if (!btree_node_reclaim(c, b))
558 list_for_each_entry_reverse(b, &bc->live, list)
559 if (!btree_node_write_and_reclaim(c, b))
563 * Rare case: all nodes were intent-locked.
566 WARN_ONCE(1, "btree cache cannibalize failed\n");
571 struct btree *bch2_btree_node_mem_alloc(struct btree_trans *trans, bool pcpu_read_locks)
573 struct bch_fs *c = trans->c;
574 struct btree_cache *bc = &c->btree_cache;
575 struct list_head *freed = pcpu_read_locks
577 : &bc->freed_nonpcpu;
578 struct btree *b, *b2;
579 u64 start_time = local_clock();
582 flags = memalloc_nofs_save();
583 mutex_lock(&bc->lock);
586 * We never free struct btree itself, just the memory that holds the on
587 * disk node. Check the freed list before allocating a new one:
589 list_for_each_entry(b, freed, list)
590 if (!btree_node_reclaim(c, b)) {
591 list_del_init(&b->list);
595 b = __btree_node_mem_alloc(c, GFP_NOWAIT|__GFP_NOWARN);
597 mutex_unlock(&bc->lock);
598 bch2_trans_unlock(trans);
599 b = __btree_node_mem_alloc(c, GFP_KERNEL);
602 mutex_lock(&bc->lock);
605 bch2_btree_lock_init(&b->c, pcpu_read_locks ? SIX_LOCK_INIT_PCPU : 0);
607 BUG_ON(!six_trylock_intent(&b->c.lock));
608 BUG_ON(!six_trylock_write(&b->c.lock));
612 * btree_free() doesn't free memory; it sticks the node on the end of
613 * the list. Check if there's any freed nodes there:
615 list_for_each_entry(b2, &bc->freeable, list)
616 if (!btree_node_reclaim(c, b2)) {
617 swap(b->data, b2->data);
618 swap(b->aux_data, b2->aux_data);
619 btree_node_to_freedlist(bc, b2);
620 six_unlock_write(&b2->c.lock);
621 six_unlock_intent(&b2->c.lock);
625 mutex_unlock(&bc->lock);
627 if (btree_node_data_alloc(c, b, GFP_NOWAIT|__GFP_NOWARN)) {
628 bch2_trans_unlock(trans);
629 if (btree_node_data_alloc(c, b, GFP_KERNEL|__GFP_NOWARN))
633 mutex_lock(&bc->lock);
636 mutex_unlock(&bc->lock);
638 BUG_ON(btree_node_hashed(b));
639 BUG_ON(btree_node_dirty(b));
640 BUG_ON(btree_node_write_in_flight(b));
647 b->whiteout_u64s = 0;
648 bch2_btree_keys_init(b);
649 set_btree_node_accessed(b);
651 bch2_time_stats_update(&c->times[BCH_TIME_btree_node_mem_alloc],
654 memalloc_nofs_restore(flags);
657 mutex_lock(&bc->lock);
659 /* Try to cannibalize another cached btree node: */
660 if (bc->alloc_lock == current) {
661 b2 = btree_node_cannibalize(c);
662 clear_btree_node_just_written(b2);
663 bch2_btree_node_hash_remove(bc, b2);
666 swap(b->data, b2->data);
667 swap(b->aux_data, b2->aux_data);
668 btree_node_to_freedlist(bc, b2);
669 six_unlock_write(&b2->c.lock);
670 six_unlock_intent(&b2->c.lock);
673 list_del_init(&b->list);
676 mutex_unlock(&bc->lock);
678 trace_and_count(c, btree_cache_cannibalize, trans);
682 mutex_unlock(&bc->lock);
683 memalloc_nofs_restore(flags);
684 return ERR_PTR(-BCH_ERR_ENOMEM_btree_node_mem_alloc);
687 /* Slowpath, don't want it inlined into btree_iter_traverse() */
688 static noinline struct btree *bch2_btree_node_fill(struct btree_trans *trans,
689 struct btree_path *path,
690 const struct bkey_i *k,
691 enum btree_id btree_id,
693 enum six_lock_type lock_type,
696 struct bch_fs *c = trans->c;
697 struct btree_cache *bc = &c->btree_cache;
701 BUG_ON(level + 1 >= BTREE_MAX_DEPTH);
703 * Parent node must be locked, else we could read in a btree node that's
706 if (path && !bch2_btree_node_relock(trans, path, level + 1)) {
707 trace_and_count(c, trans_restart_relock_parent_for_fill, trans, _THIS_IP_, path);
708 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_fill_relock));
711 b = bch2_btree_node_mem_alloc(trans, level != 0);
713 if (bch2_err_matches(PTR_ERR_OR_ZERO(b), ENOMEM)) {
714 trans->memory_allocation_failure = true;
715 trace_and_count(c, trans_restart_memory_allocation_failure, trans, _THIS_IP_, path);
716 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_fill_mem_alloc_fail));
722 bkey_copy(&b->key, k);
723 if (bch2_btree_node_hash_insert(bc, b, level, btree_id)) {
724 /* raced with another fill: */
726 /* mark as unhashed... */
729 mutex_lock(&bc->lock);
730 list_add(&b->list, &bc->freeable);
731 mutex_unlock(&bc->lock);
733 six_unlock_write(&b->c.lock);
734 six_unlock_intent(&b->c.lock);
738 set_btree_node_read_in_flight(b);
740 six_unlock_write(&b->c.lock);
741 seq = six_lock_seq(&b->c.lock);
742 six_unlock_intent(&b->c.lock);
744 /* Unlock before doing IO: */
746 bch2_trans_unlock_noassert(trans);
748 bch2_btree_node_read(trans, b, sync);
754 int ret = bch2_trans_relock(trans) ?:
755 bch2_btree_path_relock_intent(trans, path);
757 BUG_ON(!trans->restarted);
762 if (!six_relock_type(&b->c.lock, lock_type, seq)) {
764 trace_and_count(c, trans_restart_relock_after_fill, trans, _THIS_IP_, path);
765 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_relock_after_fill));
771 static noinline void btree_bad_header(struct bch_fs *c, struct btree *b)
773 struct printbuf buf = PRINTBUF;
775 if (c->curr_recovery_pass <= BCH_RECOVERY_PASS_check_allocations)
779 "btree node header doesn't match ptr\n"
780 "btree %s level %u\n"
782 bch2_btree_id_str(b->c.btree_id), b->c.level);
783 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
785 prt_printf(&buf, "\nheader: btree %s level %llu\n"
787 bch2_btree_id_str(BTREE_NODE_ID(b->data)),
788 BTREE_NODE_LEVEL(b->data));
789 bch2_bpos_to_text(&buf, b->data->min_key);
791 prt_printf(&buf, "\nmax ");
792 bch2_bpos_to_text(&buf, b->data->max_key);
794 bch2_fs_inconsistent(c, "%s", buf.buf);
798 static inline void btree_check_header(struct bch_fs *c, struct btree *b)
800 if (b->c.btree_id != BTREE_NODE_ID(b->data) ||
801 b->c.level != BTREE_NODE_LEVEL(b->data) ||
802 !bpos_eq(b->data->max_key, b->key.k.p) ||
803 (b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
804 !bpos_eq(b->data->min_key,
805 bkey_i_to_btree_ptr_v2(&b->key)->v.min_key)))
806 btree_bad_header(c, b);
809 static struct btree *__bch2_btree_node_get(struct btree_trans *trans, struct btree_path *path,
810 const struct bkey_i *k, unsigned level,
811 enum six_lock_type lock_type,
812 unsigned long trace_ip)
814 struct bch_fs *c = trans->c;
815 struct btree_cache *bc = &c->btree_cache;
818 bool need_relock = false;
821 EBUG_ON(level >= BTREE_MAX_DEPTH);
823 b = btree_cache_find(bc, k);
826 * We must have the parent locked to call bch2_btree_node_fill(),
827 * else we could read in a btree node from disk that's been
830 b = bch2_btree_node_fill(trans, path, k, path->btree_id,
831 level, lock_type, true);
834 /* We raced and found the btree node in the cache */
841 if (btree_node_read_locked(path, level + 1))
842 btree_node_unlock(trans, path, level + 1);
844 ret = btree_node_lock(trans, path, &b->c, level, lock_type, trace_ip);
845 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
850 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
851 b->c.level != level ||
853 six_unlock_type(&b->c.lock, lock_type);
854 if (bch2_btree_node_relock(trans, path, level + 1))
857 trace_and_count(c, trans_restart_btree_node_reused, trans, trace_ip, path);
858 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_lock_node_reused));
861 /* avoid atomic set bit if it's not needed: */
862 if (!btree_node_accessed(b))
863 set_btree_node_accessed(b);
866 if (unlikely(btree_node_read_in_flight(b))) {
867 u32 seq = six_lock_seq(&b->c.lock);
869 six_unlock_type(&b->c.lock, lock_type);
870 bch2_trans_unlock(trans);
873 bch2_btree_node_wait_on_read(b);
876 * should_be_locked is not set on this path yet, so we need to
877 * relock it specifically:
879 if (!six_relock_type(&b->c.lock, lock_type, seq))
883 if (unlikely(need_relock)) {
884 ret = bch2_trans_relock(trans) ?:
885 bch2_btree_path_relock_intent(trans, path);
887 six_unlock_type(&b->c.lock, lock_type);
892 prefetch(b->aux_data);
894 for_each_bset(b, t) {
895 void *p = (u64 *) b->aux_data + t->aux_data_offset;
897 prefetch(p + L1_CACHE_BYTES * 0);
898 prefetch(p + L1_CACHE_BYTES * 1);
899 prefetch(p + L1_CACHE_BYTES * 2);
902 if (unlikely(btree_node_read_error(b))) {
903 six_unlock_type(&b->c.lock, lock_type);
904 return ERR_PTR(-EIO);
907 EBUG_ON(b->c.btree_id != path->btree_id);
908 EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
909 btree_check_header(c, b);
915 * bch2_btree_node_get - find a btree node in the cache and lock it, reading it
916 * in from disk if necessary.
918 * @trans: btree transaction object
919 * @path: btree_path being traversed
920 * @k: pointer to btree node (generally KEY_TYPE_btree_ptr_v2)
921 * @level: level of btree node being looked up (0 == leaf node)
922 * @lock_type: SIX_LOCK_read or SIX_LOCK_intent
923 * @trace_ip: ip of caller of btree iterator code (i.e. caller of bch2_btree_iter_peek())
925 * The btree node will have either a read or a write lock held, depending on
926 * the @write parameter.
928 * Returns: btree node or ERR_PTR()
930 struct btree *bch2_btree_node_get(struct btree_trans *trans, struct btree_path *path,
931 const struct bkey_i *k, unsigned level,
932 enum six_lock_type lock_type,
933 unsigned long trace_ip)
935 struct bch_fs *c = trans->c;
940 EBUG_ON(level >= BTREE_MAX_DEPTH);
942 b = btree_node_mem_ptr(k);
945 * Check b->hash_val _before_ calling btree_node_lock() - this might not
946 * be the node we want anymore, and trying to lock the wrong node could
947 * cause an unneccessary transaction restart:
949 if (unlikely(!c->opts.btree_node_mem_ptr_optimization ||
951 b->hash_val != btree_ptr_hash_val(k)))
952 return __bch2_btree_node_get(trans, path, k, level, lock_type, trace_ip);
954 if (btree_node_read_locked(path, level + 1))
955 btree_node_unlock(trans, path, level + 1);
957 ret = btree_node_lock(trans, path, &b->c, level, lock_type, trace_ip);
958 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
963 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
964 b->c.level != level ||
966 six_unlock_type(&b->c.lock, lock_type);
967 if (bch2_btree_node_relock(trans, path, level + 1))
968 return __bch2_btree_node_get(trans, path, k, level, lock_type, trace_ip);
970 trace_and_count(c, trans_restart_btree_node_reused, trans, trace_ip, path);
971 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_lock_node_reused));
974 if (unlikely(btree_node_read_in_flight(b))) {
975 six_unlock_type(&b->c.lock, lock_type);
976 return __bch2_btree_node_get(trans, path, k, level, lock_type, trace_ip);
979 prefetch(b->aux_data);
981 for_each_bset(b, t) {
982 void *p = (u64 *) b->aux_data + t->aux_data_offset;
984 prefetch(p + L1_CACHE_BYTES * 0);
985 prefetch(p + L1_CACHE_BYTES * 1);
986 prefetch(p + L1_CACHE_BYTES * 2);
989 /* avoid atomic set bit if it's not needed: */
990 if (!btree_node_accessed(b))
991 set_btree_node_accessed(b);
993 if (unlikely(btree_node_read_error(b))) {
994 six_unlock_type(&b->c.lock, lock_type);
995 return ERR_PTR(-EIO);
998 EBUG_ON(b->c.btree_id != path->btree_id);
999 EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
1000 btree_check_header(c, b);
1005 struct btree *bch2_btree_node_get_noiter(struct btree_trans *trans,
1006 const struct bkey_i *k,
1007 enum btree_id btree_id,
1011 struct bch_fs *c = trans->c;
1012 struct btree_cache *bc = &c->btree_cache;
1014 struct bset_tree *t;
1017 EBUG_ON(level >= BTREE_MAX_DEPTH);
1019 if (c->opts.btree_node_mem_ptr_optimization) {
1020 b = btree_node_mem_ptr(k);
1025 b = btree_cache_find(bc, k);
1030 b = bch2_btree_node_fill(trans, NULL, k, btree_id,
1031 level, SIX_LOCK_read, true);
1033 /* We raced and found the btree node in the cache */
1038 !bch2_btree_cache_cannibalize_lock(trans, NULL))
1045 ret = btree_node_lock_nopath(trans, &b->c, SIX_LOCK_read, _THIS_IP_);
1046 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1047 return ERR_PTR(ret);
1051 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
1052 b->c.btree_id != btree_id ||
1053 b->c.level != level)) {
1054 six_unlock_read(&b->c.lock);
1059 /* XXX: waiting on IO with btree locks held: */
1060 __bch2_btree_node_wait_on_read(b);
1062 prefetch(b->aux_data);
1064 for_each_bset(b, t) {
1065 void *p = (u64 *) b->aux_data + t->aux_data_offset;
1067 prefetch(p + L1_CACHE_BYTES * 0);
1068 prefetch(p + L1_CACHE_BYTES * 1);
1069 prefetch(p + L1_CACHE_BYTES * 2);
1072 /* avoid atomic set bit if it's not needed: */
1073 if (!btree_node_accessed(b))
1074 set_btree_node_accessed(b);
1076 if (unlikely(btree_node_read_error(b))) {
1077 six_unlock_read(&b->c.lock);
1082 EBUG_ON(b->c.btree_id != btree_id);
1083 EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
1084 btree_check_header(c, b);
1086 bch2_btree_cache_cannibalize_unlock(trans);
1090 int bch2_btree_node_prefetch(struct btree_trans *trans,
1091 struct btree_path *path,
1092 const struct bkey_i *k,
1093 enum btree_id btree_id, unsigned level)
1095 struct bch_fs *c = trans->c;
1096 struct btree_cache *bc = &c->btree_cache;
1099 BUG_ON(trans && !btree_node_locked(path, level + 1));
1100 BUG_ON(level >= BTREE_MAX_DEPTH);
1102 b = btree_cache_find(bc, k);
1106 b = bch2_btree_node_fill(trans, path, k, btree_id,
1107 level, SIX_LOCK_read, false);
1108 return PTR_ERR_OR_ZERO(b);
1111 void bch2_btree_node_evict(struct btree_trans *trans, const struct bkey_i *k)
1113 struct bch_fs *c = trans->c;
1114 struct btree_cache *bc = &c->btree_cache;
1117 b = btree_cache_find(bc, k);
1121 /* not allowed to wait on io with btree locks held: */
1123 /* XXX we're called from btree_gc which will be holding other btree
1126 __bch2_btree_node_wait_on_read(b);
1127 __bch2_btree_node_wait_on_write(b);
1129 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
1130 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
1132 if (btree_node_dirty(b)) {
1133 __bch2_btree_node_write(c, b, BTREE_WRITE_cache_reclaim);
1134 six_unlock_write(&b->c.lock);
1135 six_unlock_intent(&b->c.lock);
1139 BUG_ON(btree_node_dirty(b));
1141 mutex_lock(&bc->lock);
1142 btree_node_data_free(c, b);
1143 bch2_btree_node_hash_remove(bc, b);
1144 mutex_unlock(&bc->lock);
1146 six_unlock_write(&b->c.lock);
1147 six_unlock_intent(&b->c.lock);
1150 const char *bch2_btree_id_str(enum btree_id btree)
1152 return btree < BTREE_ID_NR ? __bch2_btree_ids[btree] : "(unknown)";
1155 void bch2_btree_pos_to_text(struct printbuf *out, struct bch_fs *c, const struct btree *b)
1157 prt_printf(out, "%s level %u/%u\n ",
1158 bch2_btree_id_str(b->c.btree_id),
1160 bch2_btree_id_root(c, b->c.btree_id)->level);
1161 bch2_bkey_val_to_text(out, c, bkey_i_to_s_c(&b->key));
1164 void bch2_btree_node_to_text(struct printbuf *out, struct bch_fs *c, const struct btree *b)
1166 struct bset_stats stats;
1168 memset(&stats, 0, sizeof(stats));
1170 bch2_btree_keys_stats(b, &stats);
1172 prt_printf(out, "l %u ", b->c.level);
1173 bch2_bpos_to_text(out, b->data->min_key);
1174 prt_printf(out, " - ");
1175 bch2_bpos_to_text(out, b->data->max_key);
1176 prt_printf(out, ":\n"
1178 bch2_val_to_text(out, c, bkey_i_to_s_c(&b->key));
1183 bch2_bkey_format_to_text(out, &b->format);
1186 " unpack fn len: %u\n"
1187 " bytes used %zu/%zu (%zu%% full)\n"
1188 " sib u64s: %u, %u (merge threshold %u)\n"
1189 " nr packed keys %u\n"
1190 " nr unpacked keys %u\n"
1192 " failed unpacked %zu\n",
1194 b->nr.live_u64s * sizeof(u64),
1195 btree_buf_bytes(b) - sizeof(struct btree_node),
1196 b->nr.live_u64s * 100 / btree_max_u64s(c),
1199 c->btree_foreground_merge_threshold,
1201 b->nr.unpacked_keys,
1206 void bch2_btree_cache_to_text(struct printbuf *out, const struct bch_fs *c)
1208 prt_printf(out, "nr nodes:\t\t%u\n", c->btree_cache.used);
1209 prt_printf(out, "nr dirty:\t\t%u\n", atomic_read(&c->btree_cache.dirty));
1210 prt_printf(out, "cannibalize lock:\t%p\n", c->btree_cache.alloc_lock);