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 <linux/seq_buf.h>
16 #include <trace/events/bcachefs.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 kvpfree(b->data, btree_bytes(c));
70 munmap(b->aux_data, btree_aux_data_bytes(b));
76 btree_node_to_freedlist(bc, b);
79 static int bch2_btree_cache_cmp_fn(struct rhashtable_compare_arg *arg,
82 const struct btree *b = obj;
83 const u64 *v = arg->key;
85 return b->hash_val == *v ? 0 : 1;
88 static const struct rhashtable_params bch_btree_cache_params = {
89 .head_offset = offsetof(struct btree, hash),
90 .key_offset = offsetof(struct btree, hash_val),
91 .key_len = sizeof(u64),
92 .obj_cmpfn = bch2_btree_cache_cmp_fn,
95 static int btree_node_data_alloc(struct bch_fs *c, struct btree *b, gfp_t gfp)
97 BUG_ON(b->data || b->aux_data);
99 b->data = kvpmalloc(btree_bytes(c), gfp);
103 b->aux_data = vmalloc_exec(btree_aux_data_bytes(b), gfp);
105 b->aux_data = mmap(NULL, btree_aux_data_bytes(b),
106 PROT_READ|PROT_WRITE|PROT_EXEC,
107 MAP_PRIVATE|MAP_ANONYMOUS, 0, 0);
108 if (b->aux_data == MAP_FAILED)
112 kvpfree(b->data, btree_bytes(c));
120 static struct btree *__btree_node_mem_alloc(struct bch_fs *c, gfp_t gfp)
124 b = kzalloc(sizeof(struct btree), gfp);
128 bkey_btree_ptr_init(&b->key);
129 __six_lock_init(&b->c.lock, "b->c.lock", &bch2_btree_node_lock_key);
130 #ifdef CONFIG_DEBUG_LOCK_ALLOC
131 lockdep_set_no_check_recursion(&b->c.lock.dep_map);
133 INIT_LIST_HEAD(&b->list);
134 INIT_LIST_HEAD(&b->write_blocked);
135 b->byte_order = ilog2(btree_bytes(c));
139 struct btree *__bch2_btree_node_mem_alloc(struct bch_fs *c)
141 struct btree_cache *bc = &c->btree_cache;
144 b = __btree_node_mem_alloc(c, GFP_KERNEL);
148 if (btree_node_data_alloc(c, b, GFP_KERNEL)) {
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);
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);
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);
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 if (bc->shrink.list.next)
450 unregister_shrinker(&bc->shrink);
452 /* vfree() can allocate memory: */
453 flags = memalloc_nofs_save();
454 mutex_lock(&bc->lock);
457 list_move(&c->verify_data->list, &bc->live);
459 kvpfree(c->verify_ondisk, btree_bytes(c));
461 for (i = 0; i < BTREE_ID_NR; i++)
462 if (c->btree_roots[i].b)
463 list_add(&c->btree_roots[i].b->list, &bc->live);
465 list_splice(&bc->freeable, &bc->live);
467 while (!list_empty(&bc->live)) {
468 b = list_first_entry(&bc->live, struct btree, list);
470 BUG_ON(btree_node_read_in_flight(b) ||
471 btree_node_write_in_flight(b));
473 if (btree_node_dirty(b))
474 bch2_btree_complete_write(c, b, btree_current_write(b));
475 clear_btree_node_dirty_acct(c, b);
477 btree_node_data_free(c, b);
480 BUG_ON(atomic_read(&c->btree_cache.dirty));
482 list_splice(&bc->freed_pcpu, &bc->freed_nonpcpu);
484 while (!list_empty(&bc->freed_nonpcpu)) {
485 b = list_first_entry(&bc->freed_nonpcpu, struct btree, list);
487 six_lock_pcpu_free(&b->c.lock);
491 mutex_unlock(&bc->lock);
492 memalloc_nofs_restore(flags);
494 if (bc->table_init_done)
495 rhashtable_destroy(&bc->table);
498 int bch2_fs_btree_cache_init(struct bch_fs *c)
500 struct btree_cache *bc = &c->btree_cache;
504 pr_verbose_init(c->opts, "");
506 ret = rhashtable_init(&bc->table, &bch_btree_cache_params);
510 bc->table_init_done = true;
512 bch2_recalc_btree_reserve(c);
514 for (i = 0; i < bc->reserve; i++)
515 if (!__bch2_btree_node_mem_alloc(c)) {
520 list_splice_init(&bc->live, &bc->freeable);
522 mutex_init(&c->verify_lock);
524 bc->shrink.count_objects = bch2_btree_cache_count;
525 bc->shrink.scan_objects = bch2_btree_cache_scan;
526 bc->shrink.to_text = bch2_btree_cache_shrinker_to_text;
527 bc->shrink.seeks = 4;
528 ret = register_shrinker(&bc->shrink, "%s/btree_cache", c->name);
530 pr_verbose_init(c->opts, "ret %i", ret);
534 void bch2_fs_btree_cache_init_early(struct btree_cache *bc)
536 mutex_init(&bc->lock);
537 INIT_LIST_HEAD(&bc->live);
538 INIT_LIST_HEAD(&bc->freeable);
539 INIT_LIST_HEAD(&bc->freed_pcpu);
540 INIT_LIST_HEAD(&bc->freed_nonpcpu);
544 * We can only have one thread cannibalizing other cached btree nodes at a time,
545 * or we'll deadlock. We use an open coded mutex to ensure that, which a
546 * cannibalize_bucket() will take. This means every time we unlock the root of
547 * the btree, we need to release this lock if we have it held.
549 void bch2_btree_cache_cannibalize_unlock(struct bch_fs *c)
551 struct btree_cache *bc = &c->btree_cache;
553 if (bc->alloc_lock == current) {
554 trace_and_count(c, btree_cache_cannibalize_unlock, c);
555 bc->alloc_lock = NULL;
556 closure_wake_up(&bc->alloc_wait);
560 int bch2_btree_cache_cannibalize_lock(struct bch_fs *c, struct closure *cl)
562 struct btree_cache *bc = &c->btree_cache;
563 struct task_struct *old;
565 old = cmpxchg(&bc->alloc_lock, NULL, current);
566 if (old == NULL || old == current)
570 trace_and_count(c, btree_cache_cannibalize_lock_fail, c);
574 closure_wait(&bc->alloc_wait, cl);
576 /* Try again, after adding ourselves to waitlist */
577 old = cmpxchg(&bc->alloc_lock, NULL, current);
578 if (old == NULL || old == current) {
580 closure_wake_up(&bc->alloc_wait);
584 trace_and_count(c, btree_cache_cannibalize_lock_fail, c);
585 return -BCH_ERR_btree_cache_cannibalize_lock_blocked;
588 trace_and_count(c, btree_cache_cannibalize_lock, c);
592 static struct btree *btree_node_cannibalize(struct bch_fs *c)
594 struct btree_cache *bc = &c->btree_cache;
597 list_for_each_entry_reverse(b, &bc->live, list)
598 if (!btree_node_reclaim(c, b, false))
602 list_for_each_entry_reverse(b, &bc->live, list)
603 if (!btree_node_write_and_reclaim(c, b))
607 * Rare case: all nodes were intent-locked.
610 WARN_ONCE(1, "btree cache cannibalize failed\n");
615 struct btree *bch2_btree_node_mem_alloc(struct bch_fs *c, bool pcpu_read_locks)
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_NOWARN);
640 mutex_unlock(&bc->lock);
641 b = __btree_node_mem_alloc(c, GFP_KERNEL);
644 mutex_lock(&bc->lock);
648 six_lock_pcpu_alloc(&b->c.lock);
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_NOWARN|GFP_KERNEL))
673 mutex_lock(&bc->lock);
676 mutex_unlock(&bc->lock);
678 BUG_ON(btree_node_hashed(b));
679 BUG_ON(btree_node_dirty(b));
680 BUG_ON(btree_node_write_in_flight(b));
687 b->whiteout_u64s = 0;
688 bch2_btree_keys_init(b);
689 set_btree_node_accessed(b);
691 bch2_time_stats_update(&c->times[BCH_TIME_btree_node_mem_alloc],
694 memalloc_nofs_restore(flags);
697 mutex_lock(&bc->lock);
699 /* Try to cannibalize another cached btree node: */
700 if (bc->alloc_lock == current) {
701 b2 = btree_node_cannibalize(c);
702 bch2_btree_node_hash_remove(bc, b2);
705 swap(b->data, b2->data);
706 swap(b->aux_data, b2->aux_data);
707 btree_node_to_freedlist(bc, b2);
708 six_unlock_write(&b2->c.lock);
709 six_unlock_intent(&b2->c.lock);
712 list_del_init(&b->list);
715 mutex_unlock(&bc->lock);
717 trace_and_count(c, btree_cache_cannibalize, c);
721 mutex_unlock(&bc->lock);
722 memalloc_nofs_restore(flags);
723 return ERR_PTR(-ENOMEM);
726 /* Slowpath, don't want it inlined into btree_iter_traverse() */
727 static noinline struct btree *bch2_btree_node_fill(struct bch_fs *c,
728 struct btree_trans *trans,
729 struct btree_path *path,
730 const struct bkey_i *k,
731 enum btree_id btree_id,
733 enum six_lock_type lock_type,
736 struct btree_cache *bc = &c->btree_cache;
740 BUG_ON(level + 1 >= BTREE_MAX_DEPTH);
742 * Parent node must be locked, else we could read in a btree node that's
745 if (trans && !bch2_btree_node_relock(trans, path, level + 1)) {
746 trace_and_count(c, trans_restart_relock_parent_for_fill, trans, _THIS_IP_, path);
747 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_fill_relock));
750 b = bch2_btree_node_mem_alloc(c, level != 0);
752 if (trans && b == ERR_PTR(-ENOMEM)) {
753 trans->memory_allocation_failure = true;
754 trace_and_count(c, trans_restart_memory_allocation_failure, trans, _THIS_IP_, path);
755 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_fill_mem_alloc_fail));
762 * Btree nodes read in from disk should not have the accessed bit set
763 * initially, so that linear scans don't thrash the cache:
765 clear_btree_node_accessed(b);
767 bkey_copy(&b->key, k);
768 if (bch2_btree_node_hash_insert(bc, b, level, btree_id)) {
769 /* raced with another fill: */
771 /* mark as unhashed... */
774 mutex_lock(&bc->lock);
775 list_add(&b->list, &bc->freeable);
776 mutex_unlock(&bc->lock);
778 six_unlock_write(&b->c.lock);
779 six_unlock_intent(&b->c.lock);
783 set_btree_node_read_in_flight(b);
785 six_unlock_write(&b->c.lock);
786 seq = b->c.lock.state.seq;
787 six_unlock_intent(&b->c.lock);
789 /* Unlock before doing IO: */
791 bch2_trans_unlock(trans);
793 bch2_btree_node_read(c, b, sync);
799 int ret = bch2_trans_relock(trans) ?:
800 bch2_btree_path_relock_intent(trans, path);
802 BUG_ON(!trans->restarted);
807 if (!six_relock_type(&b->c.lock, lock_type, seq)) {
809 trace_and_count(c, trans_restart_relock_after_fill, trans, _THIS_IP_, path);
810 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_relock_after_fill));
816 static noinline void btree_bad_header(struct bch_fs *c, struct btree *b)
818 struct printbuf buf = PRINTBUF;
820 if (!test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags))
824 "btree node header doesn't match ptr\n"
825 "btree %s level %u\n"
827 bch2_btree_ids[b->c.btree_id], b->c.level);
828 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
830 prt_printf(&buf, "\nheader: btree %s level %llu\n"
832 bch2_btree_ids[BTREE_NODE_ID(b->data)],
833 BTREE_NODE_LEVEL(b->data));
834 bch2_bpos_to_text(&buf, b->data->min_key);
836 prt_printf(&buf, "\nmax ");
837 bch2_bpos_to_text(&buf, b->data->max_key);
839 bch2_fs_inconsistent(c, "%s", buf.buf);
843 static inline void btree_check_header(struct bch_fs *c, struct btree *b)
845 if (b->c.btree_id != BTREE_NODE_ID(b->data) ||
846 b->c.level != BTREE_NODE_LEVEL(b->data) ||
847 !bpos_eq(b->data->max_key, b->key.k.p) ||
848 (b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
849 !bpos_eq(b->data->min_key,
850 bkey_i_to_btree_ptr_v2(&b->key)->v.min_key)))
851 btree_bad_header(c, b);
854 static struct btree *__bch2_btree_node_get(struct btree_trans *trans, struct btree_path *path,
855 const struct bkey_i *k, unsigned level,
856 enum six_lock_type lock_type,
857 unsigned long trace_ip)
859 struct bch_fs *c = trans->c;
860 struct btree_cache *bc = &c->btree_cache;
865 EBUG_ON(level >= BTREE_MAX_DEPTH);
867 b = btree_cache_find(bc, k);
870 * We must have the parent locked to call bch2_btree_node_fill(),
871 * else we could read in a btree node from disk that's been
874 b = bch2_btree_node_fill(c, trans, path, k, path->btree_id,
875 level, lock_type, true);
877 /* We raced and found the btree node in the cache */
884 if (btree_node_read_locked(path, level + 1))
885 btree_node_unlock(trans, path, level + 1);
887 ret = btree_node_lock(trans, path, &b->c, level, lock_type, trace_ip);
888 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
893 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
894 b->c.level != level ||
896 six_unlock_type(&b->c.lock, lock_type);
897 if (bch2_btree_node_relock(trans, path, level + 1))
900 trace_and_count(c, trans_restart_btree_node_reused, trans, trace_ip, path);
901 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_lock_node_reused));
904 /* avoid atomic set bit if it's not needed: */
905 if (!btree_node_accessed(b))
906 set_btree_node_accessed(b);
909 if (unlikely(btree_node_read_in_flight(b))) {
910 u32 seq = b->c.lock.state.seq;
912 six_unlock_type(&b->c.lock, lock_type);
913 bch2_trans_unlock(trans);
915 bch2_btree_node_wait_on_read(b);
918 * should_be_locked is not set on this path yet, so we need to
919 * relock it specifically:
922 int ret = bch2_trans_relock(trans) ?:
923 bch2_btree_path_relock_intent(trans, path);
925 BUG_ON(!trans->restarted);
930 if (!six_relock_type(&b->c.lock, lock_type, seq))
934 prefetch(b->aux_data);
936 for_each_bset(b, t) {
937 void *p = (u64 *) b->aux_data + t->aux_data_offset;
939 prefetch(p + L1_CACHE_BYTES * 0);
940 prefetch(p + L1_CACHE_BYTES * 1);
941 prefetch(p + L1_CACHE_BYTES * 2);
944 if (unlikely(btree_node_read_error(b))) {
945 six_unlock_type(&b->c.lock, lock_type);
946 return ERR_PTR(-EIO);
949 EBUG_ON(b->c.btree_id != path->btree_id);
950 EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
951 btree_check_header(c, b);
957 * bch_btree_node_get - find a btree node in the cache and lock it, reading it
958 * in from disk if necessary.
960 * The btree node will have either a read or a write lock held, depending on
961 * the @write parameter.
963 struct btree *bch2_btree_node_get(struct btree_trans *trans, struct btree_path *path,
964 const struct bkey_i *k, unsigned level,
965 enum six_lock_type lock_type,
966 unsigned long trace_ip)
968 struct bch_fs *c = trans->c;
973 EBUG_ON(level >= BTREE_MAX_DEPTH);
975 b = btree_node_mem_ptr(k);
978 * Check b->hash_val _before_ calling btree_node_lock() - this might not
979 * be the node we want anymore, and trying to lock the wrong node could
980 * cause an unneccessary transaction restart:
982 if (unlikely(!c->opts.btree_node_mem_ptr_optimization ||
984 b->hash_val != btree_ptr_hash_val(k)))
985 return __bch2_btree_node_get(trans, path, k, level, lock_type, trace_ip);
987 if (btree_node_read_locked(path, level + 1))
988 btree_node_unlock(trans, path, level + 1);
990 ret = btree_node_lock(trans, path, &b->c, level, lock_type, trace_ip);
991 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
996 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
997 b->c.level != level ||
999 six_unlock_type(&b->c.lock, lock_type);
1000 if (bch2_btree_node_relock(trans, path, level + 1))
1001 return __bch2_btree_node_get(trans, path, k, level, lock_type, trace_ip);
1003 trace_and_count(c, trans_restart_btree_node_reused, trans, trace_ip, path);
1004 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_lock_node_reused));
1007 if (unlikely(btree_node_read_in_flight(b))) {
1008 u32 seq = b->c.lock.state.seq;
1010 six_unlock_type(&b->c.lock, lock_type);
1011 bch2_trans_unlock(trans);
1013 bch2_btree_node_wait_on_read(b);
1016 * should_be_locked is not set on this path yet, so we need to
1017 * relock it specifically:
1020 int ret = bch2_trans_relock(trans) ?:
1021 bch2_btree_path_relock_intent(trans, path);
1023 BUG_ON(!trans->restarted);
1024 return ERR_PTR(ret);
1028 if (!six_relock_type(&b->c.lock, lock_type, seq))
1029 return __bch2_btree_node_get(trans, path, k, level, lock_type, trace_ip);
1032 prefetch(b->aux_data);
1034 for_each_bset(b, t) {
1035 void *p = (u64 *) b->aux_data + t->aux_data_offset;
1037 prefetch(p + L1_CACHE_BYTES * 0);
1038 prefetch(p + L1_CACHE_BYTES * 1);
1039 prefetch(p + L1_CACHE_BYTES * 2);
1042 /* avoid atomic set bit if it's not needed: */
1043 if (!btree_node_accessed(b))
1044 set_btree_node_accessed(b);
1046 if (unlikely(btree_node_read_error(b))) {
1047 six_unlock_type(&b->c.lock, lock_type);
1048 return ERR_PTR(-EIO);
1051 EBUG_ON(b->c.btree_id != path->btree_id);
1052 EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
1053 btree_check_header(c, b);
1058 struct btree *bch2_btree_node_get_noiter(struct btree_trans *trans,
1059 const struct bkey_i *k,
1060 enum btree_id btree_id,
1064 struct bch_fs *c = trans->c;
1065 struct btree_cache *bc = &c->btree_cache;
1067 struct bset_tree *t;
1070 EBUG_ON(level >= BTREE_MAX_DEPTH);
1072 if (c->opts.btree_node_mem_ptr_optimization) {
1073 b = btree_node_mem_ptr(k);
1078 b = btree_cache_find(bc, k);
1083 b = bch2_btree_node_fill(c, NULL, NULL, k, btree_id,
1084 level, SIX_LOCK_read, true);
1086 /* We raced and found the btree node in the cache */
1091 !bch2_btree_cache_cannibalize_lock(c, NULL))
1098 ret = btree_node_lock_nopath(trans, &b->c, SIX_LOCK_read, _THIS_IP_);
1099 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1100 return ERR_PTR(ret);
1104 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
1105 b->c.btree_id != btree_id ||
1106 b->c.level != level)) {
1107 six_unlock_read(&b->c.lock);
1112 /* XXX: waiting on IO with btree locks held: */
1113 __bch2_btree_node_wait_on_read(b);
1115 prefetch(b->aux_data);
1117 for_each_bset(b, t) {
1118 void *p = (u64 *) b->aux_data + t->aux_data_offset;
1120 prefetch(p + L1_CACHE_BYTES * 0);
1121 prefetch(p + L1_CACHE_BYTES * 1);
1122 prefetch(p + L1_CACHE_BYTES * 2);
1125 /* avoid atomic set bit if it's not needed: */
1126 if (!btree_node_accessed(b))
1127 set_btree_node_accessed(b);
1129 if (unlikely(btree_node_read_error(b))) {
1130 six_unlock_read(&b->c.lock);
1135 EBUG_ON(b->c.btree_id != btree_id);
1136 EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
1137 btree_check_header(c, b);
1139 bch2_btree_cache_cannibalize_unlock(c);
1143 int bch2_btree_node_prefetch(struct bch_fs *c,
1144 struct btree_trans *trans,
1145 struct btree_path *path,
1146 const struct bkey_i *k,
1147 enum btree_id btree_id, unsigned level)
1149 struct btree_cache *bc = &c->btree_cache;
1152 BUG_ON(trans && !btree_node_locked(path, level + 1));
1153 BUG_ON(level >= BTREE_MAX_DEPTH);
1155 b = btree_cache_find(bc, k);
1159 b = bch2_btree_node_fill(c, trans, path, k, btree_id,
1160 level, SIX_LOCK_read, false);
1161 return PTR_ERR_OR_ZERO(b);
1164 void bch2_btree_node_evict(struct btree_trans *trans, const struct bkey_i *k)
1166 struct bch_fs *c = trans->c;
1167 struct btree_cache *bc = &c->btree_cache;
1170 b = btree_cache_find(bc, k);
1174 /* not allowed to wait on io with btree locks held: */
1176 /* XXX we're called from btree_gc which will be holding other btree
1179 __bch2_btree_node_wait_on_read(b);
1180 __bch2_btree_node_wait_on_write(b);
1182 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
1183 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
1185 if (btree_node_dirty(b)) {
1186 __bch2_btree_node_write(c, b, BTREE_WRITE_cache_reclaim);
1187 six_unlock_write(&b->c.lock);
1188 six_unlock_intent(&b->c.lock);
1192 BUG_ON(btree_node_dirty(b));
1194 mutex_lock(&bc->lock);
1195 btree_node_data_free(c, b);
1196 bch2_btree_node_hash_remove(bc, b);
1197 mutex_unlock(&bc->lock);
1199 six_unlock_write(&b->c.lock);
1200 six_unlock_intent(&b->c.lock);
1203 void bch2_btree_node_to_text(struct printbuf *out, struct bch_fs *c,
1206 const struct bkey_format *f = &b->format;
1207 struct bset_stats stats;
1209 memset(&stats, 0, sizeof(stats));
1211 bch2_btree_keys_stats(b, &stats);
1213 prt_printf(out, "l %u ", b->c.level);
1214 bch2_bpos_to_text(out, b->data->min_key);
1215 prt_printf(out, " - ");
1216 bch2_bpos_to_text(out, b->data->max_key);
1217 prt_printf(out, ":\n"
1219 bch2_val_to_text(out, c, bkey_i_to_s_c(&b->key));
1221 prt_printf(out, "\n"
1222 " format: u64s %u fields %u %u %u %u %u\n"
1223 " unpack fn len: %u\n"
1224 " bytes used %zu/%zu (%zu%% full)\n"
1225 " sib u64s: %u, %u (merge threshold %u)\n"
1226 " nr packed keys %u\n"
1227 " nr unpacked keys %u\n"
1229 " failed unpacked %zu\n",
1231 f->bits_per_field[0],
1232 f->bits_per_field[1],
1233 f->bits_per_field[2],
1234 f->bits_per_field[3],
1235 f->bits_per_field[4],
1237 b->nr.live_u64s * sizeof(u64),
1238 btree_bytes(c) - sizeof(struct btree_node),
1239 b->nr.live_u64s * 100 / btree_max_u64s(c),
1242 c->btree_foreground_merge_threshold,
1244 b->nr.unpacked_keys,
1249 void bch2_btree_cache_to_text(struct printbuf *out, struct btree_cache *bc)
1251 prt_printf(out, "nr nodes:\t\t%u\n", bc->used);
1252 prt_printf(out, "nr dirty:\t\t%u\n", atomic_read(&bc->dirty));
1253 prt_printf(out, "cannibalize lock:\t%p\n", bc->alloc_lock);
1255 prt_printf(out, "freed:\t\t\t\t%u\n", bc->freed);
1256 prt_printf(out, "not freed, dirty:\t\t%u\n", bc->not_freed_dirty);
1257 prt_printf(out, "not freed, write in flight:\t%u\n", bc->not_freed_write_in_flight);
1258 prt_printf(out, "not freed, read in flight:\t%u\n", bc->not_freed_read_in_flight);
1259 prt_printf(out, "not freed, lock intent failed:\t%u\n", bc->not_freed_lock_intent);
1260 prt_printf(out, "not freed, lock write failed:\t%u\n", bc->not_freed_lock_write);
1261 prt_printf(out, "not freed, access bit:\t\t%u\n", bc->not_freed_access_bit);
1262 prt_printf(out, "not freed, no evict failed:\t%u\n", bc->not_freed_noevict);
1263 prt_printf(out, "not freed, write blocked:\t%u\n", bc->not_freed_write_blocked);
1264 prt_printf(out, "not freed, will make reachable:\t%u\n", bc->not_freed_will_make_reachable);