1 // SPDX-License-Identifier: GPL-2.0
5 #include "btree_cache.h"
7 #include "btree_iter.h"
8 #include "btree_locking.h"
13 #include <linux/prefetch.h>
14 #include <linux/sched/mm.h>
15 #include <trace/events/bcachefs.h>
17 #define BTREE_CACHE_NOT_FREED_INCREMENT(counter) \
19 if (shrinker_counter) \
20 bc->not_freed_##counter++; \
23 const char * const bch2_btree_node_flags[] = {
30 void bch2_recalc_btree_reserve(struct bch_fs *c)
32 unsigned i, reserve = 16;
34 if (!c->btree_roots[0].b)
37 for (i = 0; i < BTREE_ID_NR; i++)
38 if (c->btree_roots[i].b)
39 reserve += min_t(unsigned, 1,
40 c->btree_roots[i].b->c.level) * 8;
42 c->btree_cache.reserve = reserve;
45 static inline unsigned btree_cache_can_free(struct btree_cache *bc)
47 return max_t(int, 0, bc->used - bc->reserve);
50 static void btree_node_to_freedlist(struct btree_cache *bc, struct btree *b)
52 if (b->c.lock.readers)
53 list_move(&b->list, &bc->freed_pcpu);
55 list_move(&b->list, &bc->freed_nonpcpu);
58 static void btree_node_data_free(struct bch_fs *c, struct btree *b)
60 struct btree_cache *bc = &c->btree_cache;
62 EBUG_ON(btree_node_write_in_flight(b));
64 kvpfree(b->data, btree_bytes(c));
69 munmap(b->aux_data, btree_aux_data_bytes(b));
75 btree_node_to_freedlist(bc, b);
78 static int bch2_btree_cache_cmp_fn(struct rhashtable_compare_arg *arg,
81 const struct btree *b = obj;
82 const u64 *v = arg->key;
84 return b->hash_val == *v ? 0 : 1;
87 static const struct rhashtable_params bch_btree_cache_params = {
88 .head_offset = offsetof(struct btree, hash),
89 .key_offset = offsetof(struct btree, hash_val),
90 .key_len = sizeof(u64),
91 .obj_cmpfn = bch2_btree_cache_cmp_fn,
94 static int btree_node_data_alloc(struct bch_fs *c, struct btree *b, gfp_t gfp)
96 BUG_ON(b->data || b->aux_data);
98 b->data = kvpmalloc(btree_bytes(c), gfp);
102 b->aux_data = vmalloc_exec(btree_aux_data_bytes(b), gfp);
104 b->aux_data = mmap(NULL, btree_aux_data_bytes(b),
105 PROT_READ|PROT_WRITE|PROT_EXEC,
106 MAP_PRIVATE|MAP_ANONYMOUS, 0, 0);
107 if (b->aux_data == MAP_FAILED)
111 kvpfree(b->data, btree_bytes(c));
119 static struct btree *__btree_node_mem_alloc(struct bch_fs *c, gfp_t gfp)
123 b = kzalloc(sizeof(struct btree), gfp);
127 bkey_btree_ptr_init(&b->key);
128 __six_lock_init(&b->c.lock, "b->c.lock", &bch2_btree_node_lock_key);
129 #ifdef CONFIG_DEBUG_LOCK_ALLOC
130 lockdep_set_no_check_recursion(&b->c.lock.dep_map);
132 INIT_LIST_HEAD(&b->list);
133 INIT_LIST_HEAD(&b->write_blocked);
134 b->byte_order = ilog2(btree_bytes(c));
138 struct btree *__bch2_btree_node_mem_alloc(struct bch_fs *c)
140 struct btree_cache *bc = &c->btree_cache;
143 b = __btree_node_mem_alloc(c, GFP_KERNEL);
147 if (btree_node_data_alloc(c, b, GFP_KERNEL)) {
153 list_add(&b->list, &bc->freeable);
157 /* Btree in memory cache - hash table */
159 void bch2_btree_node_hash_remove(struct btree_cache *bc, struct btree *b)
161 int ret = rhashtable_remove_fast(&bc->table, &b->hash, bch_btree_cache_params);
165 /* Cause future lookups for this node to fail: */
169 int __bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b)
172 b->hash_val = btree_ptr_hash_val(&b->key);
174 return rhashtable_lookup_insert_fast(&bc->table, &b->hash,
175 bch_btree_cache_params);
178 int bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b,
179 unsigned level, enum btree_id id)
186 mutex_lock(&bc->lock);
187 ret = __bch2_btree_node_hash_insert(bc, b);
189 list_add_tail(&b->list, &bc->live);
190 mutex_unlock(&bc->lock);
196 static inline struct btree *btree_cache_find(struct btree_cache *bc,
197 const struct bkey_i *k)
199 u64 v = btree_ptr_hash_val(k);
201 return rhashtable_lookup_fast(&bc->table, &v, bch_btree_cache_params);
205 * this version is for btree nodes that have already been freed (we're not
206 * reaping a real btree node)
208 static int __btree_node_reclaim(struct bch_fs *c, struct btree *b, bool flush, bool shrinker_counter)
210 struct btree_cache *bc = &c->btree_cache;
213 lockdep_assert_held(&bc->lock);
215 if (b->flags & ((1U << BTREE_NODE_dirty)|
216 (1U << BTREE_NODE_read_in_flight)|
217 (1U << BTREE_NODE_write_in_flight))) {
219 if (btree_node_dirty(b))
220 BTREE_CACHE_NOT_FREED_INCREMENT(dirty);
221 else if (btree_node_read_in_flight(b))
222 BTREE_CACHE_NOT_FREED_INCREMENT(read_in_flight);
223 else if (btree_node_write_in_flight(b))
224 BTREE_CACHE_NOT_FREED_INCREMENT(write_in_flight);
228 /* XXX: waiting on IO with btree cache lock held */
229 bch2_btree_node_wait_on_read(b);
230 bch2_btree_node_wait_on_write(b);
233 if (!six_trylock_intent(&b->c.lock)) {
234 BTREE_CACHE_NOT_FREED_INCREMENT(lock_intent);
238 if (!six_trylock_write(&b->c.lock)) {
239 BTREE_CACHE_NOT_FREED_INCREMENT(lock_write);
240 goto out_unlock_intent;
243 /* recheck under lock */
244 if (b->flags & ((1U << BTREE_NODE_read_in_flight)|
245 (1U << BTREE_NODE_write_in_flight))) {
247 if (btree_node_read_in_flight(b))
248 BTREE_CACHE_NOT_FREED_INCREMENT(read_in_flight);
249 else if (btree_node_write_in_flight(b))
250 BTREE_CACHE_NOT_FREED_INCREMENT(write_in_flight);
253 six_unlock_write(&b->c.lock);
254 six_unlock_intent(&b->c.lock);
258 if (btree_node_noevict(b)) {
259 BTREE_CACHE_NOT_FREED_INCREMENT(noevict);
262 if (btree_node_write_blocked(b)) {
263 BTREE_CACHE_NOT_FREED_INCREMENT(write_blocked);
266 if (btree_node_will_make_reachable(b)) {
267 BTREE_CACHE_NOT_FREED_INCREMENT(will_make_reachable);
271 if (btree_node_dirty(b)) {
273 BTREE_CACHE_NOT_FREED_INCREMENT(dirty);
277 * Using the underscore version because we don't want to compact
278 * bsets after the write, since this node is about to be evicted
279 * - unless btree verify mode is enabled, since it runs out of
280 * the post write cleanup:
282 if (bch2_verify_btree_ondisk)
283 bch2_btree_node_write(c, b, SIX_LOCK_intent,
284 BTREE_WRITE_cache_reclaim);
286 __bch2_btree_node_write(c, b,
287 BTREE_WRITE_cache_reclaim);
289 six_unlock_write(&b->c.lock);
290 six_unlock_intent(&b->c.lock);
294 if (b->hash_val && !ret)
295 trace_and_count(c, btree_cache_reap, c, b);
298 six_unlock_write(&b->c.lock);
300 six_unlock_intent(&b->c.lock);
305 static int btree_node_reclaim(struct bch_fs *c, struct btree *b, bool shrinker_counter)
307 return __btree_node_reclaim(c, b, false, shrinker_counter);
310 static int btree_node_write_and_reclaim(struct bch_fs *c, struct btree *b)
312 return __btree_node_reclaim(c, b, true, false);
315 static unsigned long bch2_btree_cache_scan(struct shrinker *shrink,
316 struct shrink_control *sc)
318 struct bch_fs *c = container_of(shrink, struct bch_fs,
320 struct btree_cache *bc = &c->btree_cache;
322 unsigned long nr = sc->nr_to_scan;
323 unsigned long can_free = 0;
324 unsigned long freed = 0;
325 unsigned long touched = 0;
327 unsigned long ret = SHRINK_STOP;
328 bool trigger_writes = atomic_read(&bc->dirty) + nr >=
331 if (bch2_btree_shrinker_disabled)
334 mutex_lock(&bc->lock);
335 flags = memalloc_nofs_save();
338 * It's _really_ critical that we don't free too many btree nodes - we
339 * have to always leave ourselves a reserve. The reserve is how we
340 * guarantee that allocating memory for a new btree node can always
341 * succeed, so that inserting keys into the btree can always succeed and
342 * IO can always make forward progress:
344 can_free = btree_cache_can_free(bc);
345 nr = min_t(unsigned long, nr, can_free);
348 list_for_each_entry_safe(b, t, &bc->freeable, list) {
350 * Leave a few nodes on the freeable list, so that a btree split
351 * won't have to hit the system allocator:
361 if (!btree_node_reclaim(c, b, true)) {
362 btree_node_data_free(c, b);
363 six_unlock_write(&b->c.lock);
364 six_unlock_intent(&b->c.lock);
370 list_for_each_entry_safe(b, t, &bc->live, list) {
373 if (btree_node_accessed(b)) {
374 clear_btree_node_accessed(b);
375 bc->not_freed_access_bit++;
376 } else if (!btree_node_reclaim(c, b, true)) {
378 btree_node_data_free(c, b);
381 bch2_btree_node_hash_remove(bc, b);
382 six_unlock_write(&b->c.lock);
383 six_unlock_intent(&b->c.lock);
387 } else if (trigger_writes &&
388 btree_node_dirty(b) &&
389 !btree_node_will_make_reachable(b) &&
390 !btree_node_write_blocked(b) &&
391 six_trylock_read(&b->c.lock)) {
392 list_move(&bc->live, &b->list);
393 mutex_unlock(&bc->lock);
394 __bch2_btree_node_write(c, b, BTREE_WRITE_cache_reclaim);
395 six_unlock_read(&b->c.lock);
398 mutex_lock(&bc->lock);
406 if (&t->list != &bc->live)
407 list_move_tail(&bc->live, &t->list);
409 mutex_unlock(&bc->lock);
412 memalloc_nofs_restore(flags);
413 trace_and_count(c, btree_cache_scan, sc->nr_to_scan, can_free, ret);
417 static unsigned long bch2_btree_cache_count(struct shrinker *shrink,
418 struct shrink_control *sc)
420 struct bch_fs *c = container_of(shrink, struct bch_fs,
422 struct btree_cache *bc = &c->btree_cache;
424 if (bch2_btree_shrinker_disabled)
427 return btree_cache_can_free(bc);
430 static void bch2_btree_cache_shrinker_to_text(struct printbuf *out, struct shrinker *shrink)
432 struct bch_fs *c = container_of(shrink, struct bch_fs,
435 bch2_btree_cache_to_text(out, &c->btree_cache);
438 void bch2_fs_btree_cache_exit(struct bch_fs *c)
440 struct btree_cache *bc = &c->btree_cache;
444 if (bc->shrink.list.next)
445 unregister_shrinker(&bc->shrink);
447 /* vfree() can allocate memory: */
448 flags = memalloc_nofs_save();
449 mutex_lock(&bc->lock);
452 list_move(&c->verify_data->list, &bc->live);
454 kvpfree(c->verify_ondisk, btree_bytes(c));
456 for (i = 0; i < BTREE_ID_NR; i++)
457 if (c->btree_roots[i].b)
458 list_add(&c->btree_roots[i].b->list, &bc->live);
460 list_splice(&bc->freeable, &bc->live);
462 while (!list_empty(&bc->live)) {
463 b = list_first_entry(&bc->live, struct btree, list);
465 BUG_ON(btree_node_read_in_flight(b) ||
466 btree_node_write_in_flight(b));
468 if (btree_node_dirty(b))
469 bch2_btree_complete_write(c, b, btree_current_write(b));
470 clear_btree_node_dirty_acct(c, b);
472 btree_node_data_free(c, b);
475 BUG_ON(atomic_read(&c->btree_cache.dirty));
477 list_splice(&bc->freed_pcpu, &bc->freed_nonpcpu);
479 while (!list_empty(&bc->freed_nonpcpu)) {
480 b = list_first_entry(&bc->freed_nonpcpu, struct btree, list);
482 six_lock_pcpu_free(&b->c.lock);
486 mutex_unlock(&bc->lock);
487 memalloc_nofs_restore(flags);
489 if (bc->table_init_done)
490 rhashtable_destroy(&bc->table);
493 int bch2_fs_btree_cache_init(struct bch_fs *c)
495 struct btree_cache *bc = &c->btree_cache;
499 pr_verbose_init(c->opts, "");
501 ret = rhashtable_init(&bc->table, &bch_btree_cache_params);
505 bc->table_init_done = true;
507 bch2_recalc_btree_reserve(c);
509 for (i = 0; i < bc->reserve; i++)
510 if (!__bch2_btree_node_mem_alloc(c)) {
515 list_splice_init(&bc->live, &bc->freeable);
517 mutex_init(&c->verify_lock);
519 bc->shrink.count_objects = bch2_btree_cache_count;
520 bc->shrink.scan_objects = bch2_btree_cache_scan;
521 bc->shrink.to_text = bch2_btree_cache_shrinker_to_text;
522 bc->shrink.seeks = 4;
523 ret = register_shrinker(&bc->shrink, "%s/btree_cache", c->name);
525 pr_verbose_init(c->opts, "ret %i", ret);
529 void bch2_fs_btree_cache_init_early(struct btree_cache *bc)
531 mutex_init(&bc->lock);
532 INIT_LIST_HEAD(&bc->live);
533 INIT_LIST_HEAD(&bc->freeable);
534 INIT_LIST_HEAD(&bc->freed_pcpu);
535 INIT_LIST_HEAD(&bc->freed_nonpcpu);
539 * We can only have one thread cannibalizing other cached btree nodes at a time,
540 * or we'll deadlock. We use an open coded mutex to ensure that, which a
541 * cannibalize_bucket() will take. This means every time we unlock the root of
542 * the btree, we need to release this lock if we have it held.
544 void bch2_btree_cache_cannibalize_unlock(struct bch_fs *c)
546 struct btree_cache *bc = &c->btree_cache;
548 if (bc->alloc_lock == current) {
549 trace_and_count(c, btree_cache_cannibalize_unlock, c);
550 bc->alloc_lock = NULL;
551 closure_wake_up(&bc->alloc_wait);
555 int bch2_btree_cache_cannibalize_lock(struct bch_fs *c, struct closure *cl)
557 struct btree_cache *bc = &c->btree_cache;
558 struct task_struct *old;
560 old = cmpxchg(&bc->alloc_lock, NULL, current);
561 if (old == NULL || old == current)
565 trace_and_count(c, btree_cache_cannibalize_lock_fail, c);
569 closure_wait(&bc->alloc_wait, cl);
571 /* Try again, after adding ourselves to waitlist */
572 old = cmpxchg(&bc->alloc_lock, NULL, current);
573 if (old == NULL || old == current) {
575 closure_wake_up(&bc->alloc_wait);
579 trace_and_count(c, btree_cache_cannibalize_lock_fail, c);
583 trace_and_count(c, btree_cache_cannibalize_lock, c);
587 static struct btree *btree_node_cannibalize(struct bch_fs *c)
589 struct btree_cache *bc = &c->btree_cache;
592 list_for_each_entry_reverse(b, &bc->live, list)
593 if (!btree_node_reclaim(c, b, false))
597 list_for_each_entry_reverse(b, &bc->live, list)
598 if (!btree_node_write_and_reclaim(c, b))
602 * Rare case: all nodes were intent-locked.
605 WARN_ONCE(1, "btree cache cannibalize failed\n");
610 struct btree *bch2_btree_node_mem_alloc(struct bch_fs *c, bool pcpu_read_locks)
612 struct btree_cache *bc = &c->btree_cache;
613 struct list_head *freed = pcpu_read_locks
615 : &bc->freed_nonpcpu;
616 struct btree *b, *b2;
617 u64 start_time = local_clock();
620 flags = memalloc_nofs_save();
621 mutex_lock(&bc->lock);
624 * We never free struct btree itself, just the memory that holds the on
625 * disk node. Check the freed list before allocating a new one:
627 list_for_each_entry(b, freed, list)
628 if (!btree_node_reclaim(c, b, false)) {
629 list_del_init(&b->list);
633 b = __btree_node_mem_alloc(c, __GFP_NOWARN);
635 mutex_unlock(&bc->lock);
636 b = __btree_node_mem_alloc(c, GFP_KERNEL);
639 mutex_lock(&bc->lock);
643 six_lock_pcpu_alloc(&b->c.lock);
645 BUG_ON(!six_trylock_intent(&b->c.lock));
646 BUG_ON(!six_trylock_write(&b->c.lock));
650 * btree_free() doesn't free memory; it sticks the node on the end of
651 * the list. Check if there's any freed nodes there:
653 list_for_each_entry(b2, &bc->freeable, list)
654 if (!btree_node_reclaim(c, b2, false)) {
655 swap(b->data, b2->data);
656 swap(b->aux_data, b2->aux_data);
657 btree_node_to_freedlist(bc, b2);
658 six_unlock_write(&b2->c.lock);
659 six_unlock_intent(&b2->c.lock);
663 mutex_unlock(&bc->lock);
665 if (btree_node_data_alloc(c, b, __GFP_NOWARN|GFP_KERNEL))
668 mutex_lock(&bc->lock);
671 mutex_unlock(&bc->lock);
673 BUG_ON(btree_node_hashed(b));
674 BUG_ON(btree_node_dirty(b));
675 BUG_ON(btree_node_write_in_flight(b));
683 b->whiteout_u64s = 0;
684 bch2_btree_keys_init(b);
685 set_btree_node_accessed(b);
687 bch2_time_stats_update(&c->times[BCH_TIME_btree_node_mem_alloc],
690 memalloc_nofs_restore(flags);
693 mutex_lock(&bc->lock);
695 /* Try to cannibalize another cached btree node: */
696 if (bc->alloc_lock == current) {
697 b2 = btree_node_cannibalize(c);
698 bch2_btree_node_hash_remove(bc, b2);
701 swap(b->data, b2->data);
702 swap(b->aux_data, b2->aux_data);
703 btree_node_to_freedlist(bc, b2);
704 six_unlock_write(&b2->c.lock);
705 six_unlock_intent(&b2->c.lock);
708 list_del_init(&b->list);
711 mutex_unlock(&bc->lock);
713 trace_and_count(c, btree_cache_cannibalize, c);
717 mutex_unlock(&bc->lock);
718 memalloc_nofs_restore(flags);
719 return ERR_PTR(-ENOMEM);
722 /* Slowpath, don't want it inlined into btree_iter_traverse() */
723 static noinline struct btree *bch2_btree_node_fill(struct bch_fs *c,
724 struct btree_trans *trans,
725 struct btree_path *path,
726 const struct bkey_i *k,
727 enum btree_id btree_id,
729 enum six_lock_type lock_type,
732 struct btree_cache *bc = &c->btree_cache;
736 BUG_ON(level + 1 >= BTREE_MAX_DEPTH);
738 * Parent node must be locked, else we could read in a btree node that's
741 if (trans && !bch2_btree_node_relock(trans, path, level + 1)) {
742 trace_and_count(c, trans_restart_relock_parent_for_fill, trans, _THIS_IP_, path);
743 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_fill_relock));
746 b = bch2_btree_node_mem_alloc(c, level != 0);
748 if (trans && b == ERR_PTR(-ENOMEM)) {
749 trans->memory_allocation_failure = true;
750 trace_and_count(c, trans_restart_memory_allocation_failure, trans, _THIS_IP_, path);
751 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_fill_mem_alloc_fail));
757 bkey_copy(&b->key, k);
758 if (bch2_btree_node_hash_insert(bc, b, level, btree_id)) {
759 /* raced with another fill: */
761 /* mark as unhashed... */
764 mutex_lock(&bc->lock);
765 list_add(&b->list, &bc->freeable);
766 mutex_unlock(&bc->lock);
768 six_unlock_write(&b->c.lock);
769 six_unlock_intent(&b->c.lock);
773 set_btree_node_read_in_flight(b);
775 six_unlock_write(&b->c.lock);
776 seq = b->c.lock.state.seq;
777 six_unlock_intent(&b->c.lock);
779 /* Unlock before doing IO: */
781 bch2_trans_unlock(trans);
783 bch2_btree_node_read(c, b, sync);
789 int ret = bch2_trans_relock(trans) ?:
790 bch2_btree_path_relock_intent(trans, path);
792 BUG_ON(!trans->restarted);
797 if (!six_relock_type(&b->c.lock, lock_type, seq)) {
799 trace_and_count(c, trans_restart_relock_after_fill, trans, _THIS_IP_, path);
800 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_relock_after_fill));
806 static noinline void btree_bad_header(struct bch_fs *c, struct btree *b)
808 struct printbuf buf = PRINTBUF;
810 if (!test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags))
814 "btree node header doesn't match ptr\n"
815 "btree %s level %u\n"
817 bch2_btree_ids[b->c.btree_id], b->c.level);
818 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
820 prt_printf(&buf, "\nheader: btree %s level %llu\n"
822 bch2_btree_ids[BTREE_NODE_ID(b->data)],
823 BTREE_NODE_LEVEL(b->data));
824 bch2_bpos_to_text(&buf, b->data->min_key);
826 prt_printf(&buf, "\nmax ");
827 bch2_bpos_to_text(&buf, b->data->max_key);
829 bch2_fs_inconsistent(c, "%s", buf.buf);
833 static inline void btree_check_header(struct bch_fs *c, struct btree *b)
835 if (b->c.btree_id != BTREE_NODE_ID(b->data) ||
836 b->c.level != BTREE_NODE_LEVEL(b->data) ||
837 bpos_cmp(b->data->max_key, b->key.k.p) ||
838 (b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
839 bpos_cmp(b->data->min_key,
840 bkey_i_to_btree_ptr_v2(&b->key)->v.min_key)))
841 btree_bad_header(c, b);
845 * bch_btree_node_get - find a btree node in the cache and lock it, reading it
846 * in from disk if necessary.
848 * If IO is necessary and running under generic_make_request, returns -EAGAIN.
850 * The btree node will have either a read or a write lock held, depending on
851 * the @write parameter.
853 struct btree *bch2_btree_node_get(struct btree_trans *trans, struct btree_path *path,
854 const struct bkey_i *k, unsigned level,
855 enum six_lock_type lock_type,
856 unsigned long trace_ip)
858 struct bch_fs *c = trans->c;
859 struct btree_cache *bc = &c->btree_cache;
864 EBUG_ON(level >= BTREE_MAX_DEPTH);
866 b = btree_node_mem_ptr(k);
869 * Check b->hash_val _before_ calling btree_node_lock() - this might not
870 * be the node we want anymore, and trying to lock the wrong node could
871 * cause an unneccessary transaction restart:
873 if (likely(c->opts.btree_node_mem_ptr_optimization &&
875 b->hash_val == btree_ptr_hash_val(k)))
878 b = btree_cache_find(bc, k);
881 * We must have the parent locked to call bch2_btree_node_fill(),
882 * else we could read in a btree node from disk that's been
885 b = bch2_btree_node_fill(c, trans, path, k, path->btree_id,
886 level, lock_type, true);
888 /* We raced and found the btree node in the cache */
897 * There's a potential deadlock with splits and insertions into
898 * interior nodes we have to avoid:
900 * The other thread might be holding an intent lock on the node
901 * we want, and they want to update its parent node so they're
902 * going to upgrade their intent lock on the parent node to a
905 * But if we're holding a read lock on the parent, and we're
906 * trying to get the intent lock they're holding, we deadlock.
908 * So to avoid this we drop the read locks on parent nodes when
909 * we're starting to take intent locks - and handle the race.
911 * The race is that they might be about to free the node we
912 * want, and dropping our read lock on the parent node lets them
913 * update the parent marking the node we want as freed, and then
916 * To guard against this, btree nodes are evicted from the cache
917 * when they're freed - and b->hash_val is zeroed out, which we
918 * check for after we lock the node.
920 * Then, bch2_btree_node_relock() on the parent will fail - because
921 * the parent was modified, when the pointer to the node we want
922 * was removed - and we'll bail out:
924 if (btree_node_read_locked(path, level + 1))
925 btree_node_unlock(trans, path, level + 1);
927 ret = btree_node_lock(trans, path, &b->c, level, lock_type, trace_ip);
928 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
933 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
934 b->c.level != level ||
936 six_unlock_type(&b->c.lock, lock_type);
937 if (bch2_btree_node_relock(trans, path, level + 1))
940 trace_and_count(c, trans_restart_btree_node_reused, trans, trace_ip, path);
941 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_lock_node_reused));
945 if (unlikely(btree_node_read_in_flight(b))) {
946 u32 seq = b->c.lock.state.seq;
948 six_unlock_type(&b->c.lock, lock_type);
949 bch2_trans_unlock(trans);
951 bch2_btree_node_wait_on_read(b);
954 * should_be_locked is not set on this path yet, so we need to
955 * relock it specifically:
958 int ret = bch2_trans_relock(trans) ?:
959 bch2_btree_path_relock_intent(trans, path);
961 BUG_ON(!trans->restarted);
966 if (!six_relock_type(&b->c.lock, lock_type, seq))
970 prefetch(b->aux_data);
972 for_each_bset(b, t) {
973 void *p = (u64 *) b->aux_data + t->aux_data_offset;
975 prefetch(p + L1_CACHE_BYTES * 0);
976 prefetch(p + L1_CACHE_BYTES * 1);
977 prefetch(p + L1_CACHE_BYTES * 2);
980 /* avoid atomic set bit if it's not needed: */
981 if (!btree_node_accessed(b))
982 set_btree_node_accessed(b);
984 if (unlikely(btree_node_read_error(b))) {
985 six_unlock_type(&b->c.lock, lock_type);
986 return ERR_PTR(-EIO);
989 EBUG_ON(b->c.btree_id != path->btree_id);
990 EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
991 btree_check_header(c, b);
996 struct btree *bch2_btree_node_get_noiter(struct btree_trans *trans,
997 const struct bkey_i *k,
998 enum btree_id btree_id,
1002 struct bch_fs *c = trans->c;
1003 struct btree_cache *bc = &c->btree_cache;
1005 struct bset_tree *t;
1008 EBUG_ON(level >= BTREE_MAX_DEPTH);
1010 if (c->opts.btree_node_mem_ptr_optimization) {
1011 b = btree_node_mem_ptr(k);
1016 b = btree_cache_find(bc, k);
1021 b = bch2_btree_node_fill(c, NULL, NULL, k, btree_id,
1022 level, SIX_LOCK_read, true);
1024 /* We raced and found the btree node in the cache */
1029 !bch2_btree_cache_cannibalize_lock(c, NULL))
1036 ret = btree_node_lock_nopath(trans, &b->c, SIX_LOCK_read);
1037 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1038 return ERR_PTR(ret);
1042 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
1043 b->c.btree_id != btree_id ||
1044 b->c.level != level)) {
1045 six_unlock_read(&b->c.lock);
1050 /* XXX: waiting on IO with btree locks held: */
1051 __bch2_btree_node_wait_on_read(b);
1053 prefetch(b->aux_data);
1055 for_each_bset(b, t) {
1056 void *p = (u64 *) b->aux_data + t->aux_data_offset;
1058 prefetch(p + L1_CACHE_BYTES * 0);
1059 prefetch(p + L1_CACHE_BYTES * 1);
1060 prefetch(p + L1_CACHE_BYTES * 2);
1063 /* avoid atomic set bit if it's not needed: */
1064 if (!btree_node_accessed(b))
1065 set_btree_node_accessed(b);
1067 if (unlikely(btree_node_read_error(b))) {
1068 six_unlock_read(&b->c.lock);
1073 EBUG_ON(b->c.btree_id != btree_id);
1074 EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
1075 btree_check_header(c, b);
1077 bch2_btree_cache_cannibalize_unlock(c);
1081 int bch2_btree_node_prefetch(struct bch_fs *c,
1082 struct btree_trans *trans,
1083 struct btree_path *path,
1084 const struct bkey_i *k,
1085 enum btree_id btree_id, unsigned level)
1087 struct btree_cache *bc = &c->btree_cache;
1090 BUG_ON(trans && !btree_node_locked(path, level + 1));
1091 BUG_ON(level >= BTREE_MAX_DEPTH);
1093 b = btree_cache_find(bc, k);
1097 b = bch2_btree_node_fill(c, trans, path, k, btree_id,
1098 level, SIX_LOCK_read, false);
1099 return PTR_ERR_OR_ZERO(b);
1102 void bch2_btree_node_evict(struct btree_trans *trans, const struct bkey_i *k)
1104 struct bch_fs *c = trans->c;
1105 struct btree_cache *bc = &c->btree_cache;
1108 b = btree_cache_find(bc, k);
1112 /* not allowed to wait on io with btree locks held: */
1114 /* XXX we're called from btree_gc which will be holding other btree
1117 __bch2_btree_node_wait_on_read(b);
1118 __bch2_btree_node_wait_on_write(b);
1120 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
1121 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
1123 if (btree_node_dirty(b)) {
1124 __bch2_btree_node_write(c, b, BTREE_WRITE_cache_reclaim);
1125 six_unlock_write(&b->c.lock);
1126 six_unlock_intent(&b->c.lock);
1130 BUG_ON(btree_node_dirty(b));
1132 mutex_lock(&bc->lock);
1133 btree_node_data_free(c, b);
1134 bch2_btree_node_hash_remove(bc, b);
1135 mutex_unlock(&bc->lock);
1137 six_unlock_write(&b->c.lock);
1138 six_unlock_intent(&b->c.lock);
1141 void bch2_btree_node_to_text(struct printbuf *out, struct bch_fs *c,
1144 const struct bkey_format *f = &b->format;
1145 struct bset_stats stats;
1147 memset(&stats, 0, sizeof(stats));
1149 bch2_btree_keys_stats(b, &stats);
1151 prt_printf(out, "l %u ", b->c.level);
1152 bch2_bpos_to_text(out, b->data->min_key);
1153 prt_printf(out, " - ");
1154 bch2_bpos_to_text(out, b->data->max_key);
1155 prt_printf(out, ":\n"
1157 bch2_val_to_text(out, c, bkey_i_to_s_c(&b->key));
1159 prt_printf(out, "\n"
1160 " format: u64s %u fields %u %u %u %u %u\n"
1161 " unpack fn len: %u\n"
1162 " bytes used %zu/%zu (%zu%% full)\n"
1163 " sib u64s: %u, %u (merge threshold %u)\n"
1164 " nr packed keys %u\n"
1165 " nr unpacked keys %u\n"
1167 " failed unpacked %zu\n",
1169 f->bits_per_field[0],
1170 f->bits_per_field[1],
1171 f->bits_per_field[2],
1172 f->bits_per_field[3],
1173 f->bits_per_field[4],
1175 b->nr.live_u64s * sizeof(u64),
1176 btree_bytes(c) - sizeof(struct btree_node),
1177 b->nr.live_u64s * 100 / btree_max_u64s(c),
1180 c->btree_foreground_merge_threshold,
1182 b->nr.unpacked_keys,
1187 void bch2_btree_cache_to_text(struct printbuf *out, struct btree_cache *bc)
1189 prt_printf(out, "nr nodes:\t\t%u\n", bc->used);
1190 prt_printf(out, "nr dirty:\t\t%u\n", atomic_read(&bc->dirty));
1191 prt_printf(out, "cannibalize lock:\t%p\n", bc->alloc_lock);
1193 prt_printf(out, "freed:\t\t\t\t%u\n", bc->freed);
1194 prt_printf(out, "not freed, dirty:\t\t%u\n", bc->not_freed_dirty);
1195 prt_printf(out, "not freed, write in flight:\t%u\n", bc->not_freed_write_in_flight);
1196 prt_printf(out, "not freed, read in flight:\t%u\n", bc->not_freed_read_in_flight);
1197 prt_printf(out, "not freed, lock intent failed:\t%u\n", bc->not_freed_lock_intent);
1198 prt_printf(out, "not freed, lock write failed:\t%u\n", bc->not_freed_lock_write);
1199 prt_printf(out, "not freed, access bit:\t\t%u\n", bc->not_freed_access_bit);
1200 prt_printf(out, "not freed, no evict failed:\t%u\n", bc->not_freed_noevict);
1201 prt_printf(out, "not freed, write blocked:\t%u\n", bc->not_freed_write_blocked);
1202 prt_printf(out, "not freed, will make reachable:\t%u\n", bc->not_freed_will_make_reachable);