1 // SPDX-License-Identifier: GPL-2.0
5 #include "btree_cache.h"
7 #include "btree_iter.h"
8 #include "btree_locking.h"
14 #include <linux/prefetch.h>
15 #include <linux/sched/mm.h>
16 #include <linux/seq_buf.h>
18 #define BTREE_CACHE_NOT_FREED_INCREMENT(counter) \
20 if (shrinker_counter) \
21 bc->not_freed_##counter++; \
24 const char * const bch2_btree_node_flags[] = {
31 void bch2_recalc_btree_reserve(struct bch_fs *c)
33 unsigned i, reserve = 16;
35 if (!c->btree_roots[0].b)
38 for (i = 0; i < BTREE_ID_NR; i++)
39 if (c->btree_roots[i].b)
40 reserve += min_t(unsigned, 1,
41 c->btree_roots[i].b->c.level) * 8;
43 c->btree_cache.reserve = reserve;
46 static inline unsigned btree_cache_can_free(struct btree_cache *bc)
48 return max_t(int, 0, bc->used - bc->reserve);
51 static void btree_node_to_freedlist(struct btree_cache *bc, struct btree *b)
53 if (b->c.lock.readers)
54 list_move(&b->list, &bc->freed_pcpu);
56 list_move(&b->list, &bc->freed_nonpcpu);
59 static void btree_node_data_free(struct bch_fs *c, struct btree *b)
61 struct btree_cache *bc = &c->btree_cache;
63 EBUG_ON(btree_node_write_in_flight(b));
65 clear_btree_node_just_written(b);
67 kvpfree(b->data, btree_bytes(c));
72 munmap(b->aux_data, btree_aux_data_bytes(b));
78 btree_node_to_freedlist(bc, b);
81 static int bch2_btree_cache_cmp_fn(struct rhashtable_compare_arg *arg,
84 const struct btree *b = obj;
85 const u64 *v = arg->key;
87 return b->hash_val == *v ? 0 : 1;
90 static const struct rhashtable_params bch_btree_cache_params = {
91 .head_offset = offsetof(struct btree, hash),
92 .key_offset = offsetof(struct btree, hash_val),
93 .key_len = sizeof(u64),
94 .obj_cmpfn = bch2_btree_cache_cmp_fn,
97 static int btree_node_data_alloc(struct bch_fs *c, struct btree *b, gfp_t gfp)
99 BUG_ON(b->data || b->aux_data);
101 b->data = kvpmalloc(btree_bytes(c), gfp);
103 return -BCH_ERR_ENOMEM_btree_node_mem_alloc;
105 b->aux_data = kvmalloc(btree_aux_data_bytes(b), gfp);
107 b->aux_data = mmap(NULL, btree_aux_data_bytes(b),
108 PROT_READ|PROT_WRITE|PROT_EXEC,
109 MAP_PRIVATE|MAP_ANONYMOUS, 0, 0);
110 if (b->aux_data == MAP_FAILED)
114 kvpfree(b->data, btree_bytes(c));
116 return -BCH_ERR_ENOMEM_btree_node_mem_alloc;
122 static struct btree *__btree_node_mem_alloc(struct bch_fs *c, gfp_t gfp)
126 b = kzalloc(sizeof(struct btree), gfp);
130 bkey_btree_ptr_init(&b->key);
131 INIT_LIST_HEAD(&b->list);
132 INIT_LIST_HEAD(&b->write_blocked);
133 b->byte_order = ilog2(btree_bytes(c));
137 struct btree *__bch2_btree_node_mem_alloc(struct bch_fs *c)
139 struct btree_cache *bc = &c->btree_cache;
142 b = __btree_node_mem_alloc(c, GFP_KERNEL);
146 if (btree_node_data_alloc(c, b, GFP_KERNEL)) {
151 bch2_btree_lock_init(&b->c, 0);
154 list_add(&b->list, &bc->freeable);
158 /* Btree in memory cache - hash table */
160 void bch2_btree_node_hash_remove(struct btree_cache *bc, struct btree *b)
162 int ret = rhashtable_remove_fast(&bc->table, &b->hash, bch_btree_cache_params);
166 /* Cause future lookups for this node to fail: */
170 int __bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b)
173 b->hash_val = btree_ptr_hash_val(&b->key);
175 return rhashtable_lookup_insert_fast(&bc->table, &b->hash,
176 bch_btree_cache_params);
179 int bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b,
180 unsigned level, enum btree_id id)
187 mutex_lock(&bc->lock);
188 ret = __bch2_btree_node_hash_insert(bc, b);
190 list_add_tail(&b->list, &bc->live);
191 mutex_unlock(&bc->lock);
197 static inline struct btree *btree_cache_find(struct btree_cache *bc,
198 const struct bkey_i *k)
200 u64 v = btree_ptr_hash_val(k);
202 return rhashtable_lookup_fast(&bc->table, &v, bch_btree_cache_params);
206 * this version is for btree nodes that have already been freed (we're not
207 * reaping a real btree node)
209 static int __btree_node_reclaim(struct bch_fs *c, struct btree *b, bool flush, bool shrinker_counter)
211 struct btree_cache *bc = &c->btree_cache;
214 lockdep_assert_held(&bc->lock);
216 if (b->flags & ((1U << BTREE_NODE_dirty)|
217 (1U << BTREE_NODE_read_in_flight)|
218 (1U << BTREE_NODE_write_in_flight))) {
220 if (btree_node_dirty(b))
221 BTREE_CACHE_NOT_FREED_INCREMENT(dirty);
222 else if (btree_node_read_in_flight(b))
223 BTREE_CACHE_NOT_FREED_INCREMENT(read_in_flight);
224 else if (btree_node_write_in_flight(b))
225 BTREE_CACHE_NOT_FREED_INCREMENT(write_in_flight);
226 return -BCH_ERR_ENOMEM_btree_node_reclaim;
229 /* XXX: waiting on IO with btree cache lock held */
230 bch2_btree_node_wait_on_read(b);
231 bch2_btree_node_wait_on_write(b);
234 if (!six_trylock_intent(&b->c.lock)) {
235 BTREE_CACHE_NOT_FREED_INCREMENT(lock_intent);
236 return -BCH_ERR_ENOMEM_btree_node_reclaim;
239 if (!six_trylock_write(&b->c.lock)) {
240 BTREE_CACHE_NOT_FREED_INCREMENT(lock_write);
241 goto out_unlock_intent;
244 /* recheck under lock */
245 if (b->flags & ((1U << BTREE_NODE_read_in_flight)|
246 (1U << BTREE_NODE_write_in_flight))) {
248 if (btree_node_read_in_flight(b))
249 BTREE_CACHE_NOT_FREED_INCREMENT(read_in_flight);
250 else if (btree_node_write_in_flight(b))
251 BTREE_CACHE_NOT_FREED_INCREMENT(write_in_flight);
254 six_unlock_write(&b->c.lock);
255 six_unlock_intent(&b->c.lock);
259 if (btree_node_noevict(b)) {
260 BTREE_CACHE_NOT_FREED_INCREMENT(noevict);
263 if (btree_node_write_blocked(b)) {
264 BTREE_CACHE_NOT_FREED_INCREMENT(write_blocked);
267 if (btree_node_will_make_reachable(b)) {
268 BTREE_CACHE_NOT_FREED_INCREMENT(will_make_reachable);
272 if (btree_node_dirty(b)) {
274 BTREE_CACHE_NOT_FREED_INCREMENT(dirty);
278 * Using the underscore version because we don't want to compact
279 * bsets after the write, since this node is about to be evicted
280 * - unless btree verify mode is enabled, since it runs out of
281 * the post write cleanup:
283 if (bch2_verify_btree_ondisk)
284 bch2_btree_node_write(c, b, SIX_LOCK_intent,
285 BTREE_WRITE_cache_reclaim);
287 __bch2_btree_node_write(c, b,
288 BTREE_WRITE_cache_reclaim);
290 six_unlock_write(&b->c.lock);
291 six_unlock_intent(&b->c.lock);
295 if (b->hash_val && !ret)
296 trace_and_count(c, btree_cache_reap, c, b);
299 six_unlock_write(&b->c.lock);
301 six_unlock_intent(&b->c.lock);
302 ret = -BCH_ERR_ENOMEM_btree_node_reclaim;
306 static int btree_node_reclaim(struct bch_fs *c, struct btree *b, bool shrinker_counter)
308 return __btree_node_reclaim(c, b, false, shrinker_counter);
311 static int btree_node_write_and_reclaim(struct bch_fs *c, struct btree *b)
313 return __btree_node_reclaim(c, b, true, false);
316 static unsigned long bch2_btree_cache_scan(struct shrinker *shrink,
317 struct shrink_control *sc)
319 struct bch_fs *c = container_of(shrink, struct bch_fs,
321 struct btree_cache *bc = &c->btree_cache;
323 unsigned long nr = sc->nr_to_scan;
324 unsigned long can_free = 0;
325 unsigned long freed = 0;
326 unsigned long touched = 0;
328 unsigned long ret = SHRINK_STOP;
329 bool trigger_writes = atomic_read(&bc->dirty) + nr >=
332 if (bch2_btree_shrinker_disabled)
335 mutex_lock(&bc->lock);
336 flags = memalloc_nofs_save();
339 * It's _really_ critical that we don't free too many btree nodes - we
340 * have to always leave ourselves a reserve. The reserve is how we
341 * guarantee that allocating memory for a new btree node can always
342 * succeed, so that inserting keys into the btree can always succeed and
343 * IO can always make forward progress:
345 can_free = btree_cache_can_free(bc);
346 nr = min_t(unsigned long, nr, can_free);
349 list_for_each_entry_safe(b, t, &bc->freeable, list) {
351 * Leave a few nodes on the freeable list, so that a btree split
352 * won't have to hit the system allocator:
362 if (!btree_node_reclaim(c, b, true)) {
363 btree_node_data_free(c, b);
364 six_unlock_write(&b->c.lock);
365 six_unlock_intent(&b->c.lock);
371 list_for_each_entry_safe(b, t, &bc->live, list) {
374 if (btree_node_accessed(b)) {
375 clear_btree_node_accessed(b);
376 bc->not_freed_access_bit++;
377 } else if (!btree_node_reclaim(c, b, true)) {
379 btree_node_data_free(c, b);
382 bch2_btree_node_hash_remove(bc, b);
383 six_unlock_write(&b->c.lock);
384 six_unlock_intent(&b->c.lock);
388 } else if (trigger_writes &&
389 btree_node_dirty(b) &&
390 !btree_node_will_make_reachable(b) &&
391 !btree_node_write_blocked(b) &&
392 six_trylock_read(&b->c.lock)) {
393 list_move(&bc->live, &b->list);
394 mutex_unlock(&bc->lock);
395 __bch2_btree_node_write(c, b, BTREE_WRITE_cache_reclaim);
396 six_unlock_read(&b->c.lock);
399 mutex_lock(&bc->lock);
407 if (&t->list != &bc->live)
408 list_move_tail(&bc->live, &t->list);
410 mutex_unlock(&bc->lock);
413 memalloc_nofs_restore(flags);
414 trace_and_count(c, btree_cache_scan, sc->nr_to_scan, can_free, ret);
418 static unsigned long bch2_btree_cache_count(struct shrinker *shrink,
419 struct shrink_control *sc)
421 struct bch_fs *c = container_of(shrink, struct bch_fs,
423 struct btree_cache *bc = &c->btree_cache;
425 if (bch2_btree_shrinker_disabled)
428 return btree_cache_can_free(bc);
431 static void bch2_btree_cache_shrinker_to_text(struct seq_buf *s, struct shrinker *shrink)
433 struct bch_fs *c = container_of(shrink, struct bch_fs,
436 size_t buflen = seq_buf_get_buf(s, &cbuf);
437 struct printbuf out = PRINTBUF_EXTERN(cbuf, buflen);
439 bch2_btree_cache_to_text(&out, &c->btree_cache);
440 seq_buf_commit(s, out.pos);
443 void bch2_fs_btree_cache_exit(struct bch_fs *c)
445 struct btree_cache *bc = &c->btree_cache;
449 unregister_shrinker(&bc->shrink);
451 /* vfree() can allocate memory: */
452 flags = memalloc_nofs_save();
453 mutex_lock(&bc->lock);
456 list_move(&c->verify_data->list, &bc->live);
458 kvpfree(c->verify_ondisk, btree_bytes(c));
460 for (i = 0; i < BTREE_ID_NR; i++)
461 if (c->btree_roots[i].b)
462 list_add(&c->btree_roots[i].b->list, &bc->live);
464 list_splice(&bc->freeable, &bc->live);
466 while (!list_empty(&bc->live)) {
467 b = list_first_entry(&bc->live, struct btree, list);
469 BUG_ON(btree_node_read_in_flight(b) ||
470 btree_node_write_in_flight(b));
472 if (btree_node_dirty(b))
473 bch2_btree_complete_write(c, b, btree_current_write(b));
474 clear_btree_node_dirty_acct(c, b);
476 btree_node_data_free(c, b);
479 BUG_ON(atomic_read(&c->btree_cache.dirty));
481 list_splice(&bc->freed_pcpu, &bc->freed_nonpcpu);
483 while (!list_empty(&bc->freed_nonpcpu)) {
484 b = list_first_entry(&bc->freed_nonpcpu, struct btree, list);
486 six_lock_exit(&b->c.lock);
490 mutex_unlock(&bc->lock);
491 memalloc_nofs_restore(flags);
493 if (bc->table_init_done)
494 rhashtable_destroy(&bc->table);
497 int bch2_fs_btree_cache_init(struct bch_fs *c)
499 struct btree_cache *bc = &c->btree_cache;
503 pr_verbose_init(c->opts, "");
505 ret = rhashtable_init(&bc->table, &bch_btree_cache_params);
509 bc->table_init_done = true;
511 bch2_recalc_btree_reserve(c);
513 for (i = 0; i < bc->reserve; i++)
514 if (!__bch2_btree_node_mem_alloc(c)) {
515 ret = -BCH_ERR_ENOMEM_fs_btree_cache_init;
519 list_splice_init(&bc->live, &bc->freeable);
521 mutex_init(&c->verify_lock);
523 bc->shrink.count_objects = bch2_btree_cache_count;
524 bc->shrink.scan_objects = bch2_btree_cache_scan;
525 bc->shrink.to_text = bch2_btree_cache_shrinker_to_text;
526 bc->shrink.seeks = 4;
527 ret = register_shrinker(&bc->shrink, "%s/btree_cache", c->name);
529 pr_verbose_init(c->opts, "ret %i", ret);
533 void bch2_fs_btree_cache_init_early(struct btree_cache *bc)
535 mutex_init(&bc->lock);
536 INIT_LIST_HEAD(&bc->live);
537 INIT_LIST_HEAD(&bc->freeable);
538 INIT_LIST_HEAD(&bc->freed_pcpu);
539 INIT_LIST_HEAD(&bc->freed_nonpcpu);
543 * We can only have one thread cannibalizing other cached btree nodes at a time,
544 * or we'll deadlock. We use an open coded mutex to ensure that, which a
545 * cannibalize_bucket() will take. This means every time we unlock the root of
546 * the btree, we need to release this lock if we have it held.
548 void bch2_btree_cache_cannibalize_unlock(struct bch_fs *c)
550 struct btree_cache *bc = &c->btree_cache;
552 if (bc->alloc_lock == current) {
553 trace_and_count(c, btree_cache_cannibalize_unlock, c);
554 bc->alloc_lock = NULL;
555 closure_wake_up(&bc->alloc_wait);
559 int bch2_btree_cache_cannibalize_lock(struct bch_fs *c, struct closure *cl)
561 struct btree_cache *bc = &c->btree_cache;
562 struct task_struct *old;
564 old = cmpxchg(&bc->alloc_lock, NULL, current);
565 if (old == NULL || old == current)
569 trace_and_count(c, btree_cache_cannibalize_lock_fail, c);
570 return -BCH_ERR_ENOMEM_btree_cache_cannibalize_lock;
573 closure_wait(&bc->alloc_wait, cl);
575 /* Try again, after adding ourselves to waitlist */
576 old = cmpxchg(&bc->alloc_lock, NULL, current);
577 if (old == NULL || old == current) {
579 closure_wake_up(&bc->alloc_wait);
583 trace_and_count(c, btree_cache_cannibalize_lock_fail, c);
584 return -BCH_ERR_btree_cache_cannibalize_lock_blocked;
587 trace_and_count(c, btree_cache_cannibalize_lock, c);
591 static struct btree *btree_node_cannibalize(struct bch_fs *c)
593 struct btree_cache *bc = &c->btree_cache;
596 list_for_each_entry_reverse(b, &bc->live, list)
597 if (!btree_node_reclaim(c, b, false))
601 list_for_each_entry_reverse(b, &bc->live, list)
602 if (!btree_node_write_and_reclaim(c, b))
606 * Rare case: all nodes were intent-locked.
609 WARN_ONCE(1, "btree cache cannibalize failed\n");
614 struct btree *bch2_btree_node_mem_alloc(struct btree_trans *trans, bool pcpu_read_locks)
616 struct bch_fs *c = trans->c;
617 struct btree_cache *bc = &c->btree_cache;
618 struct list_head *freed = pcpu_read_locks
620 : &bc->freed_nonpcpu;
621 struct btree *b, *b2;
622 u64 start_time = local_clock();
625 flags = memalloc_nofs_save();
626 mutex_lock(&bc->lock);
629 * We never free struct btree itself, just the memory that holds the on
630 * disk node. Check the freed list before allocating a new one:
632 list_for_each_entry(b, freed, list)
633 if (!btree_node_reclaim(c, b, false)) {
634 list_del_init(&b->list);
638 b = __btree_node_mem_alloc(c, GFP_NOWAIT|__GFP_NOWARN);
640 mutex_unlock(&bc->lock);
641 bch2_trans_unlock(trans);
642 b = __btree_node_mem_alloc(c, GFP_KERNEL);
645 mutex_lock(&bc->lock);
648 bch2_btree_lock_init(&b->c, pcpu_read_locks ? SIX_LOCK_INIT_PCPU : 0);
650 BUG_ON(!six_trylock_intent(&b->c.lock));
651 BUG_ON(!six_trylock_write(&b->c.lock));
655 * btree_free() doesn't free memory; it sticks the node on the end of
656 * the list. Check if there's any freed nodes there:
658 list_for_each_entry(b2, &bc->freeable, list)
659 if (!btree_node_reclaim(c, b2, false)) {
660 swap(b->data, b2->data);
661 swap(b->aux_data, b2->aux_data);
662 btree_node_to_freedlist(bc, b2);
663 six_unlock_write(&b2->c.lock);
664 six_unlock_intent(&b2->c.lock);
668 mutex_unlock(&bc->lock);
670 if (btree_node_data_alloc(c, b, GFP_NOWAIT|__GFP_NOWARN)) {
671 bch2_trans_unlock(trans);
672 if (btree_node_data_alloc(c, b, GFP_KERNEL|__GFP_NOWARN))
676 mutex_lock(&bc->lock);
679 mutex_unlock(&bc->lock);
681 BUG_ON(btree_node_hashed(b));
682 BUG_ON(btree_node_dirty(b));
683 BUG_ON(btree_node_write_in_flight(b));
690 b->whiteout_u64s = 0;
691 bch2_btree_keys_init(b);
692 set_btree_node_accessed(b);
694 bch2_time_stats_update(&c->times[BCH_TIME_btree_node_mem_alloc],
697 memalloc_nofs_restore(flags);
700 mutex_lock(&bc->lock);
702 /* Try to cannibalize another cached btree node: */
703 if (bc->alloc_lock == current) {
704 b2 = btree_node_cannibalize(c);
705 clear_btree_node_just_written(b2);
706 bch2_btree_node_hash_remove(bc, b2);
709 swap(b->data, b2->data);
710 swap(b->aux_data, b2->aux_data);
711 btree_node_to_freedlist(bc, b2);
712 six_unlock_write(&b2->c.lock);
713 six_unlock_intent(&b2->c.lock);
716 list_del_init(&b->list);
719 mutex_unlock(&bc->lock);
721 trace_and_count(c, btree_cache_cannibalize, c);
725 mutex_unlock(&bc->lock);
726 memalloc_nofs_restore(flags);
727 return ERR_PTR(-BCH_ERR_ENOMEM_btree_node_mem_alloc);
730 /* Slowpath, don't want it inlined into btree_iter_traverse() */
731 static noinline struct btree *bch2_btree_node_fill(struct btree_trans *trans,
732 struct btree_path *path,
733 const struct bkey_i *k,
734 enum btree_id btree_id,
736 enum six_lock_type lock_type,
739 struct bch_fs *c = trans->c;
740 struct btree_cache *bc = &c->btree_cache;
744 BUG_ON(level + 1 >= BTREE_MAX_DEPTH);
746 * Parent node must be locked, else we could read in a btree node that's
749 if (path && !bch2_btree_node_relock(trans, path, level + 1)) {
750 trace_and_count(c, trans_restart_relock_parent_for_fill, trans, _THIS_IP_, path);
751 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_fill_relock));
754 b = bch2_btree_node_mem_alloc(trans, level != 0);
756 if (bch2_err_matches(PTR_ERR_OR_ZERO(b), ENOMEM)) {
757 trans->memory_allocation_failure = true;
758 trace_and_count(c, trans_restart_memory_allocation_failure, trans, _THIS_IP_, path);
759 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_fill_mem_alloc_fail));
766 * Btree nodes read in from disk should not have the accessed bit set
767 * initially, so that linear scans don't thrash the cache:
769 clear_btree_node_accessed(b);
771 bkey_copy(&b->key, k);
772 if (bch2_btree_node_hash_insert(bc, b, level, btree_id)) {
773 /* raced with another fill: */
775 /* mark as unhashed... */
778 mutex_lock(&bc->lock);
779 list_add(&b->list, &bc->freeable);
780 mutex_unlock(&bc->lock);
782 six_unlock_write(&b->c.lock);
783 six_unlock_intent(&b->c.lock);
787 set_btree_node_read_in_flight(b);
789 six_unlock_write(&b->c.lock);
790 seq = six_lock_seq(&b->c.lock);
791 six_unlock_intent(&b->c.lock);
793 /* Unlock before doing IO: */
795 bch2_trans_unlock_noassert(trans);
797 bch2_btree_node_read(c, b, sync);
803 int ret = bch2_trans_relock(trans) ?:
804 bch2_btree_path_relock_intent(trans, path);
806 BUG_ON(!trans->restarted);
811 if (!six_relock_type(&b->c.lock, lock_type, seq)) {
813 trace_and_count(c, trans_restart_relock_after_fill, trans, _THIS_IP_, path);
814 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_relock_after_fill));
820 static noinline void btree_bad_header(struct bch_fs *c, struct btree *b)
822 struct printbuf buf = PRINTBUF;
824 if (!test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags))
828 "btree node header doesn't match ptr\n"
829 "btree %s level %u\n"
831 bch2_btree_ids[b->c.btree_id], b->c.level);
832 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
834 prt_printf(&buf, "\nheader: btree %s level %llu\n"
836 bch2_btree_ids[BTREE_NODE_ID(b->data)],
837 BTREE_NODE_LEVEL(b->data));
838 bch2_bpos_to_text(&buf, b->data->min_key);
840 prt_printf(&buf, "\nmax ");
841 bch2_bpos_to_text(&buf, b->data->max_key);
843 bch2_fs_inconsistent(c, "%s", buf.buf);
847 static inline void btree_check_header(struct bch_fs *c, struct btree *b)
849 if (b->c.btree_id != BTREE_NODE_ID(b->data) ||
850 b->c.level != BTREE_NODE_LEVEL(b->data) ||
851 !bpos_eq(b->data->max_key, b->key.k.p) ||
852 (b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
853 !bpos_eq(b->data->min_key,
854 bkey_i_to_btree_ptr_v2(&b->key)->v.min_key)))
855 btree_bad_header(c, b);
858 static struct btree *__bch2_btree_node_get(struct btree_trans *trans, struct btree_path *path,
859 const struct bkey_i *k, unsigned level,
860 enum six_lock_type lock_type,
861 unsigned long trace_ip)
863 struct bch_fs *c = trans->c;
864 struct btree_cache *bc = &c->btree_cache;
867 bool need_relock = false;
870 EBUG_ON(level >= BTREE_MAX_DEPTH);
872 b = btree_cache_find(bc, k);
875 * We must have the parent locked to call bch2_btree_node_fill(),
876 * else we could read in a btree node from disk that's been
879 b = bch2_btree_node_fill(trans, path, k, path->btree_id,
880 level, lock_type, true);
883 /* We raced and found the btree node in the cache */
890 if (btree_node_read_locked(path, level + 1))
891 btree_node_unlock(trans, path, level + 1);
893 ret = btree_node_lock(trans, path, &b->c, level, lock_type, trace_ip);
894 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
899 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
900 b->c.level != level ||
902 six_unlock_type(&b->c.lock, lock_type);
903 if (bch2_btree_node_relock(trans, path, level + 1))
906 trace_and_count(c, trans_restart_btree_node_reused, trans, trace_ip, path);
907 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_lock_node_reused));
910 /* avoid atomic set bit if it's not needed: */
911 if (!btree_node_accessed(b))
912 set_btree_node_accessed(b);
915 if (unlikely(btree_node_read_in_flight(b))) {
916 u32 seq = six_lock_seq(&b->c.lock);
918 six_unlock_type(&b->c.lock, lock_type);
919 bch2_trans_unlock(trans);
922 bch2_btree_node_wait_on_read(b);
925 * should_be_locked is not set on this path yet, so we need to
926 * relock it specifically:
928 if (!six_relock_type(&b->c.lock, lock_type, seq))
932 if (unlikely(need_relock)) {
933 int ret = bch2_trans_relock(trans) ?:
934 bch2_btree_path_relock_intent(trans, path);
936 six_unlock_type(&b->c.lock, lock_type);
941 prefetch(b->aux_data);
943 for_each_bset(b, t) {
944 void *p = (u64 *) b->aux_data + t->aux_data_offset;
946 prefetch(p + L1_CACHE_BYTES * 0);
947 prefetch(p + L1_CACHE_BYTES * 1);
948 prefetch(p + L1_CACHE_BYTES * 2);
951 if (unlikely(btree_node_read_error(b))) {
952 six_unlock_type(&b->c.lock, lock_type);
953 return ERR_PTR(-EIO);
956 EBUG_ON(b->c.btree_id != path->btree_id);
957 EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
958 btree_check_header(c, b);
964 * bch_btree_node_get - find a btree node in the cache and lock it, reading it
965 * in from disk if necessary.
967 * The btree node will have either a read or a write lock held, depending on
968 * the @write parameter.
970 struct btree *bch2_btree_node_get(struct btree_trans *trans, struct btree_path *path,
971 const struct bkey_i *k, unsigned level,
972 enum six_lock_type lock_type,
973 unsigned long trace_ip)
975 struct bch_fs *c = trans->c;
980 EBUG_ON(level >= BTREE_MAX_DEPTH);
982 b = btree_node_mem_ptr(k);
985 * Check b->hash_val _before_ calling btree_node_lock() - this might not
986 * be the node we want anymore, and trying to lock the wrong node could
987 * cause an unneccessary transaction restart:
989 if (unlikely(!c->opts.btree_node_mem_ptr_optimization ||
991 b->hash_val != btree_ptr_hash_val(k)))
992 return __bch2_btree_node_get(trans, path, k, level, lock_type, trace_ip);
994 if (btree_node_read_locked(path, level + 1))
995 btree_node_unlock(trans, path, level + 1);
997 ret = btree_node_lock(trans, path, &b->c, level, lock_type, trace_ip);
998 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1003 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
1004 b->c.level != level ||
1006 six_unlock_type(&b->c.lock, lock_type);
1007 if (bch2_btree_node_relock(trans, path, level + 1))
1008 return __bch2_btree_node_get(trans, path, k, level, lock_type, trace_ip);
1010 trace_and_count(c, trans_restart_btree_node_reused, trans, trace_ip, path);
1011 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_lock_node_reused));
1014 if (unlikely(btree_node_read_in_flight(b))) {
1015 u32 seq = six_lock_seq(&b->c.lock);
1017 six_unlock_type(&b->c.lock, lock_type);
1018 bch2_trans_unlock(trans);
1020 bch2_btree_node_wait_on_read(b);
1023 * should_be_locked is not set on this path yet, so we need to
1024 * relock it specifically:
1027 int ret = bch2_trans_relock(trans) ?:
1028 bch2_btree_path_relock_intent(trans, path);
1030 BUG_ON(!trans->restarted);
1031 return ERR_PTR(ret);
1035 if (!six_relock_type(&b->c.lock, lock_type, seq))
1036 return __bch2_btree_node_get(trans, path, k, level, lock_type, trace_ip);
1039 prefetch(b->aux_data);
1041 for_each_bset(b, t) {
1042 void *p = (u64 *) b->aux_data + t->aux_data_offset;
1044 prefetch(p + L1_CACHE_BYTES * 0);
1045 prefetch(p + L1_CACHE_BYTES * 1);
1046 prefetch(p + L1_CACHE_BYTES * 2);
1049 /* avoid atomic set bit if it's not needed: */
1050 if (!btree_node_accessed(b))
1051 set_btree_node_accessed(b);
1053 if (unlikely(btree_node_read_error(b))) {
1054 six_unlock_type(&b->c.lock, lock_type);
1055 return ERR_PTR(-EIO);
1058 EBUG_ON(b->c.btree_id != path->btree_id);
1059 EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
1060 btree_check_header(c, b);
1065 struct btree *bch2_btree_node_get_noiter(struct btree_trans *trans,
1066 const struct bkey_i *k,
1067 enum btree_id btree_id,
1071 struct bch_fs *c = trans->c;
1072 struct btree_cache *bc = &c->btree_cache;
1074 struct bset_tree *t;
1077 EBUG_ON(level >= BTREE_MAX_DEPTH);
1079 if (c->opts.btree_node_mem_ptr_optimization) {
1080 b = btree_node_mem_ptr(k);
1085 b = btree_cache_find(bc, k);
1090 b = bch2_btree_node_fill(trans, NULL, k, btree_id,
1091 level, SIX_LOCK_read, true);
1093 /* We raced and found the btree node in the cache */
1098 !bch2_btree_cache_cannibalize_lock(c, NULL))
1105 ret = btree_node_lock_nopath(trans, &b->c, SIX_LOCK_read, _THIS_IP_);
1106 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1107 return ERR_PTR(ret);
1111 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
1112 b->c.btree_id != btree_id ||
1113 b->c.level != level)) {
1114 six_unlock_read(&b->c.lock);
1119 /* XXX: waiting on IO with btree locks held: */
1120 __bch2_btree_node_wait_on_read(b);
1122 prefetch(b->aux_data);
1124 for_each_bset(b, t) {
1125 void *p = (u64 *) b->aux_data + t->aux_data_offset;
1127 prefetch(p + L1_CACHE_BYTES * 0);
1128 prefetch(p + L1_CACHE_BYTES * 1);
1129 prefetch(p + L1_CACHE_BYTES * 2);
1132 /* avoid atomic set bit if it's not needed: */
1133 if (!btree_node_accessed(b))
1134 set_btree_node_accessed(b);
1136 if (unlikely(btree_node_read_error(b))) {
1137 six_unlock_read(&b->c.lock);
1142 EBUG_ON(b->c.btree_id != btree_id);
1143 EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
1144 btree_check_header(c, b);
1146 bch2_btree_cache_cannibalize_unlock(c);
1150 int bch2_btree_node_prefetch(struct btree_trans *trans,
1151 struct btree_path *path,
1152 const struct bkey_i *k,
1153 enum btree_id btree_id, unsigned level)
1155 struct bch_fs *c = trans->c;
1156 struct btree_cache *bc = &c->btree_cache;
1159 BUG_ON(trans && !btree_node_locked(path, level + 1));
1160 BUG_ON(level >= BTREE_MAX_DEPTH);
1162 b = btree_cache_find(bc, k);
1166 b = bch2_btree_node_fill(trans, path, k, btree_id,
1167 level, SIX_LOCK_read, false);
1168 return PTR_ERR_OR_ZERO(b);
1171 void bch2_btree_node_evict(struct btree_trans *trans, const struct bkey_i *k)
1173 struct bch_fs *c = trans->c;
1174 struct btree_cache *bc = &c->btree_cache;
1177 b = btree_cache_find(bc, k);
1181 /* not allowed to wait on io with btree locks held: */
1183 /* XXX we're called from btree_gc which will be holding other btree
1186 __bch2_btree_node_wait_on_read(b);
1187 __bch2_btree_node_wait_on_write(b);
1189 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
1190 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
1192 if (btree_node_dirty(b)) {
1193 __bch2_btree_node_write(c, b, BTREE_WRITE_cache_reclaim);
1194 six_unlock_write(&b->c.lock);
1195 six_unlock_intent(&b->c.lock);
1199 BUG_ON(btree_node_dirty(b));
1201 mutex_lock(&bc->lock);
1202 btree_node_data_free(c, b);
1203 bch2_btree_node_hash_remove(bc, b);
1204 mutex_unlock(&bc->lock);
1206 six_unlock_write(&b->c.lock);
1207 six_unlock_intent(&b->c.lock);
1210 void bch2_btree_node_to_text(struct printbuf *out, struct bch_fs *c,
1211 const struct btree *b)
1213 const struct bkey_format *f = &b->format;
1214 struct bset_stats stats;
1216 memset(&stats, 0, sizeof(stats));
1218 bch2_btree_keys_stats(b, &stats);
1220 prt_printf(out, "l %u ", b->c.level);
1221 bch2_bpos_to_text(out, b->data->min_key);
1222 prt_printf(out, " - ");
1223 bch2_bpos_to_text(out, b->data->max_key);
1224 prt_printf(out, ":\n"
1226 bch2_val_to_text(out, c, bkey_i_to_s_c(&b->key));
1228 prt_printf(out, "\n"
1229 " format: u64s %u fields %u %u %u %u %u\n"
1230 " unpack fn len: %u\n"
1231 " bytes used %zu/%zu (%zu%% full)\n"
1232 " sib u64s: %u, %u (merge threshold %u)\n"
1233 " nr packed keys %u\n"
1234 " nr unpacked keys %u\n"
1236 " failed unpacked %zu\n",
1238 f->bits_per_field[0],
1239 f->bits_per_field[1],
1240 f->bits_per_field[2],
1241 f->bits_per_field[3],
1242 f->bits_per_field[4],
1244 b->nr.live_u64s * sizeof(u64),
1245 btree_bytes(c) - sizeof(struct btree_node),
1246 b->nr.live_u64s * 100 / btree_max_u64s(c),
1249 c->btree_foreground_merge_threshold,
1251 b->nr.unpacked_keys,
1256 void bch2_btree_cache_to_text(struct printbuf *out, const struct btree_cache *bc)
1258 prt_printf(out, "nr nodes:\t\t%u\n", bc->used);
1259 prt_printf(out, "nr dirty:\t\t%u\n", atomic_read(&bc->dirty));
1260 prt_printf(out, "cannibalize lock:\t%p\n", bc->alloc_lock);
1262 prt_printf(out, "freed:\t\t\t\t%u\n", bc->freed);
1263 prt_printf(out, "not freed, dirty:\t\t%u\n", bc->not_freed_dirty);
1264 prt_printf(out, "not freed, write in flight:\t%u\n", bc->not_freed_write_in_flight);
1265 prt_printf(out, "not freed, read in flight:\t%u\n", bc->not_freed_read_in_flight);
1266 prt_printf(out, "not freed, lock intent failed:\t%u\n", bc->not_freed_lock_intent);
1267 prt_printf(out, "not freed, lock write failed:\t%u\n", bc->not_freed_lock_write);
1268 prt_printf(out, "not freed, access bit:\t\t%u\n", bc->not_freed_access_bit);
1269 prt_printf(out, "not freed, no evict failed:\t%u\n", bc->not_freed_noevict);
1270 prt_printf(out, "not freed, write blocked:\t%u\n", bc->not_freed_write_blocked);
1271 prt_printf(out, "not freed, will make reachable:\t%u\n", bc->not_freed_will_make_reachable);