3 #include "btree_cache.h"
4 #include "btree_iter.h"
5 #include "btree_key_cache.h"
6 #include "btree_locking.h"
7 #include "btree_update.h"
10 #include "journal_reclaim.h"
12 #include <linux/sched/mm.h>
13 #include <trace/events/bcachefs.h>
15 static struct kmem_cache *bch2_key_cache;
17 static int bch2_btree_key_cache_cmp_fn(struct rhashtable_compare_arg *arg,
20 const struct bkey_cached *ck = obj;
21 const struct bkey_cached_key *key = arg->key;
23 return cmp_int(ck->key.btree_id, key->btree_id) ?:
24 bpos_cmp(ck->key.pos, key->pos);
27 static const struct rhashtable_params bch2_btree_key_cache_params = {
28 .head_offset = offsetof(struct bkey_cached, hash),
29 .key_offset = offsetof(struct bkey_cached, key),
30 .key_len = sizeof(struct bkey_cached_key),
31 .obj_cmpfn = bch2_btree_key_cache_cmp_fn,
35 inline struct bkey_cached *
36 bch2_btree_key_cache_find(struct bch_fs *c, enum btree_id btree_id, struct bpos pos)
38 struct bkey_cached_key key = {
43 return rhashtable_lookup_fast(&c->btree_key_cache.table, &key,
44 bch2_btree_key_cache_params);
47 static bool bkey_cached_lock_for_evict(struct bkey_cached *ck)
49 if (!six_trylock_intent(&ck->c.lock))
52 if (!six_trylock_write(&ck->c.lock)) {
53 six_unlock_intent(&ck->c.lock);
57 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
58 six_unlock_write(&ck->c.lock);
59 six_unlock_intent(&ck->c.lock);
66 static void bkey_cached_evict(struct btree_key_cache *c,
67 struct bkey_cached *ck)
69 BUG_ON(rhashtable_remove_fast(&c->table, &ck->hash,
70 bch2_btree_key_cache_params));
71 memset(&ck->key, ~0, sizeof(ck->key));
73 atomic_long_dec(&c->nr_keys);
76 static void bkey_cached_free(struct btree_key_cache *bc,
77 struct bkey_cached *ck)
79 struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
81 BUG_ON(test_bit(BKEY_CACHED_DIRTY, &ck->flags));
83 ck->btree_trans_barrier_seq =
84 start_poll_synchronize_srcu(&c->btree_trans_barrier);
86 list_move_tail(&ck->list, &bc->freed);
93 six_unlock_write(&ck->c.lock);
94 six_unlock_intent(&ck->c.lock);
97 static struct bkey_cached *
98 bkey_cached_alloc(struct btree_key_cache *c)
100 struct bkey_cached *ck;
102 ck = kmem_cache_alloc(bch2_key_cache, GFP_NOFS|__GFP_ZERO);
104 INIT_LIST_HEAD(&ck->list);
105 six_lock_init(&ck->c.lock);
106 BUG_ON(!six_trylock_intent(&ck->c.lock));
107 BUG_ON(!six_trylock_write(&ck->c.lock));
114 static struct bkey_cached *
115 bkey_cached_reuse(struct btree_key_cache *c)
117 struct bucket_table *tbl;
118 struct rhash_head *pos;
119 struct bkey_cached *ck;
122 mutex_lock(&c->lock);
123 list_for_each_entry_reverse(ck, &c->freed, list)
124 if (bkey_cached_lock_for_evict(ck)) {
127 mutex_unlock(&c->lock);
130 mutex_unlock(&c->lock);
133 tbl = rht_dereference_rcu(c->table.tbl, &c->table);
134 for (i = 0; i < tbl->size; i++)
135 rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
136 if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
137 bkey_cached_lock_for_evict(ck)) {
138 bkey_cached_evict(c, ck);
148 static struct bkey_cached *
149 btree_key_cache_create(struct btree_key_cache *c,
150 enum btree_id btree_id,
153 struct bkey_cached *ck;
156 ck = bkey_cached_alloc(c);
159 ck = bkey_cached_reuse(c);
161 return ERR_PTR(-ENOMEM);
167 ck->c.btree_id = btree_id;
168 ck->key.btree_id = btree_id;
171 ck->flags = 1U << BKEY_CACHED_ACCESSED;
173 if (unlikely(rhashtable_lookup_insert_fast(&c->table,
175 bch2_btree_key_cache_params))) {
176 /* We raced with another fill: */
178 if (likely(was_new)) {
179 six_unlock_write(&ck->c.lock);
180 six_unlock_intent(&ck->c.lock);
183 mutex_lock(&c->lock);
184 bkey_cached_free(c, ck);
185 mutex_unlock(&c->lock);
191 atomic_long_inc(&c->nr_keys);
193 six_unlock_write(&ck->c.lock);
198 static int btree_key_cache_fill(struct btree_trans *trans,
199 struct btree_path *ck_path,
200 struct bkey_cached *ck)
202 struct btree_iter iter;
204 unsigned new_u64s = 0;
205 struct bkey_i *new_k = NULL;
208 bch2_trans_iter_init(trans, &iter, ck->key.btree_id,
209 ck->key.pos, BTREE_ITER_SLOTS);
210 k = bch2_btree_iter_peek_slot(&iter);
215 if (!bch2_btree_node_relock(trans, ck_path, 0)) {
216 trace_transaction_restart_ip(trans->ip, _THIS_IP_);
217 ret = btree_trans_restart(trans);
222 * bch2_varint_decode can read past the end of the buffer by at
223 * most 7 bytes (it won't be used):
225 new_u64s = k.k->u64s + 1;
227 if (new_u64s > ck->u64s) {
228 new_u64s = roundup_pow_of_two(new_u64s);
229 new_k = kmalloc(new_u64s * sizeof(u64), GFP_NOFS);
237 * XXX: not allowed to be holding read locks when we take a write lock,
240 bch2_btree_node_lock_write(trans, ck_path, ck_path->l[0].b);
247 bkey_reassemble(ck->k, k);
249 bch2_btree_node_unlock_write(trans, ck_path, ck_path->l[0].b);
251 /* We're not likely to need this iterator again: */
252 set_btree_iter_dontneed(&iter);
254 bch2_trans_iter_exit(trans, &iter);
258 static int bkey_cached_check_fn(struct six_lock *lock, void *p)
260 struct bkey_cached *ck = container_of(lock, struct bkey_cached, c.lock);
261 const struct btree_path *path = p;
263 return ck->key.btree_id == path->btree_id &&
264 !bpos_cmp(ck->key.pos, path->pos) ? 0 : -1;
268 int bch2_btree_path_traverse_cached(struct btree_trans *trans, struct btree_path *path,
271 struct bch_fs *c = trans->c;
272 struct bkey_cached *ck;
279 if (bch2_btree_node_relock(trans, path, 0)) {
280 ck = (void *) path->l[0].b;
284 ck = bch2_btree_key_cache_find(c, path->btree_id, path->pos);
286 if (flags & BTREE_ITER_CACHED_NOCREATE) {
291 ck = btree_key_cache_create(&c->btree_key_cache,
292 path->btree_id, path->pos);
293 ret = PTR_ERR_OR_ZERO(ck);
299 mark_btree_node_locked(trans, path, 0, SIX_LOCK_intent);
300 path->locks_want = 1;
302 enum six_lock_type lock_want = __btree_lock_want(path, 0);
304 if (!btree_node_lock(trans, path, (void *) ck, path->pos, 0,
306 bkey_cached_check_fn, path, _THIS_IP_)) {
307 if (!trans->restarted)
310 trace_transaction_restart_ip(trans->ip, _THIS_IP_);
315 if (ck->key.btree_id != path->btree_id ||
316 bpos_cmp(ck->key.pos, path->pos)) {
317 six_unlock_type(&ck->c.lock, lock_want);
321 mark_btree_node_locked(trans, path, 0, lock_want);
324 path->l[0].lock_seq = ck->c.lock.state.seq;
325 path->l[0].b = (void *) ck;
327 if (!ck->valid && !(flags & BTREE_ITER_CACHED_NOFILL)) {
328 if (!path->locks_want &&
329 !__bch2_btree_path_upgrade(trans, path, 1)) {
330 trace_transaction_restart_ip(trans->ip, _THIS_IP_);
331 ret = btree_trans_restart(trans);
335 ret = btree_key_cache_fill(trans, path, ck);
340 if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
341 set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
343 path->uptodate = BTREE_ITER_UPTODATE;
344 BUG_ON(btree_node_locked_type(path, 0) != btree_lock_want(path, 0));
349 btree_node_unlock(path, 0);
350 path->l[0].b = BTREE_ITER_NO_NODE_ERROR;
355 static int btree_key_cache_flush_pos(struct btree_trans *trans,
356 struct bkey_cached_key key,
358 unsigned commit_flags,
361 struct bch_fs *c = trans->c;
362 struct journal *j = &c->journal;
363 struct btree_iter c_iter, b_iter;
364 struct bkey_cached *ck = NULL;
367 bch2_trans_iter_init(trans, &b_iter, key.btree_id, key.pos,
370 bch2_trans_iter_init(trans, &c_iter, key.btree_id, key.pos,
372 BTREE_ITER_CACHED_NOFILL|
373 BTREE_ITER_CACHED_NOCREATE|
375 ret = bch2_btree_iter_traverse(&c_iter);
379 ck = (void *) c_iter.path->l[0].b;
381 (journal_seq && ck->journal.seq != journal_seq))
384 if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
391 * Since journal reclaim depends on us making progress here, and the
392 * allocator/copygc depend on journal reclaim making progress, we need
393 * to be using alloc reserves:
395 ret = bch2_btree_iter_traverse(&b_iter) ?:
396 bch2_trans_update(trans, &b_iter, ck->k,
397 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
398 BTREE_TRIGGER_NORUN) ?:
399 bch2_trans_commit(trans, NULL, NULL,
400 BTREE_INSERT_NOCHECK_RW|
402 BTREE_INSERT_USE_RESERVE|
403 (ck->journal.seq == journal_last_seq(j)
404 ? BTREE_INSERT_JOURNAL_RESERVED
408 bch2_fs_fatal_err_on(ret != -EINTR &&
410 !bch2_journal_error(j), c,
411 "error flushing key cache: %i", ret);
415 bch2_journal_pin_drop(j, &ck->journal);
416 bch2_journal_preres_put(j, &ck->res);
418 BUG_ON(!btree_node_locked(c_iter.path, 0));
421 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
422 clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
423 atomic_long_dec(&c->btree_key_cache.nr_dirty);
427 BUG_ON(!btree_node_intent_locked(c_iter.path, 0));
429 mark_btree_node_unlocked(c_iter.path, 0);
430 c_iter.path->l[0].b = NULL;
432 six_lock_write(&ck->c.lock, NULL, NULL);
434 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
435 clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
436 atomic_long_dec(&c->btree_key_cache.nr_dirty);
439 bkey_cached_evict(&c->btree_key_cache, ck);
441 mutex_lock(&c->btree_key_cache.lock);
442 bkey_cached_free(&c->btree_key_cache, ck);
443 mutex_unlock(&c->btree_key_cache.lock);
446 bch2_trans_iter_exit(trans, &b_iter);
447 bch2_trans_iter_exit(trans, &c_iter);
451 int bch2_btree_key_cache_journal_flush(struct journal *j,
452 struct journal_entry_pin *pin, u64 seq)
454 struct bch_fs *c = container_of(j, struct bch_fs, journal);
455 struct bkey_cached *ck =
456 container_of(pin, struct bkey_cached, journal);
457 struct bkey_cached_key key;
460 int srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
462 six_lock_read(&ck->c.lock, NULL, NULL);
465 if (ck->journal.seq != seq ||
466 !test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
467 six_unlock_read(&ck->c.lock);
470 six_unlock_read(&ck->c.lock);
472 ret = bch2_trans_do(c, NULL, NULL, 0,
473 btree_key_cache_flush_pos(&trans, key, seq,
474 BTREE_INSERT_JOURNAL_RECLAIM, false));
476 srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
482 * Flush and evict a key from the key cache:
484 int bch2_btree_key_cache_flush(struct btree_trans *trans,
485 enum btree_id id, struct bpos pos)
487 struct bch_fs *c = trans->c;
488 struct bkey_cached_key key = { id, pos };
490 /* Fastpath - assume it won't be found: */
491 if (!bch2_btree_key_cache_find(c, id, pos))
494 return btree_key_cache_flush_pos(trans, key, 0, 0, true);
497 bool bch2_btree_insert_key_cached(struct btree_trans *trans,
498 struct btree_path *path,
499 struct bkey_i *insert)
501 struct bch_fs *c = trans->c;
502 struct bkey_cached *ck = (void *) path->l[0].b;
503 bool kick_reclaim = false;
505 BUG_ON(insert->u64s > ck->u64s);
507 if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
510 BUG_ON(jset_u64s(insert->u64s) > trans->journal_preres.u64s);
512 difference = jset_u64s(insert->u64s) - ck->res.u64s;
513 if (difference > 0) {
514 trans->journal_preres.u64s -= difference;
515 ck->res.u64s += difference;
519 bkey_copy(ck->k, insert);
522 if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
523 set_bit(BKEY_CACHED_DIRTY, &ck->flags);
524 atomic_long_inc(&c->btree_key_cache.nr_dirty);
526 if (bch2_nr_btree_keys_need_flush(c))
530 bch2_journal_pin_update(&c->journal, trans->journal_res.seq,
531 &ck->journal, bch2_btree_key_cache_journal_flush);
534 journal_reclaim_kick(&c->journal);
538 #ifdef CONFIG_BCACHEFS_DEBUG
539 void bch2_btree_key_cache_verify_clean(struct btree_trans *trans,
540 enum btree_id id, struct bpos pos)
542 BUG_ON(bch2_btree_key_cache_find(trans->c, id, pos));
546 static unsigned long bch2_btree_key_cache_scan(struct shrinker *shrink,
547 struct shrink_control *sc)
549 struct bch_fs *c = container_of(shrink, struct bch_fs,
550 btree_key_cache.shrink);
551 struct btree_key_cache *bc = &c->btree_key_cache;
552 struct bucket_table *tbl;
553 struct bkey_cached *ck, *t;
554 size_t scanned = 0, freed = 0, nr = sc->nr_to_scan;
555 unsigned start, flags;
558 /* Return -1 if we can't do anything right now */
559 if (sc->gfp_mask & __GFP_FS)
560 mutex_lock(&bc->lock);
561 else if (!mutex_trylock(&bc->lock))
564 srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
565 flags = memalloc_nofs_save();
568 * Newest freed entries are at the end of the list - once we hit one
569 * that's too new to be freed, we can bail out:
571 list_for_each_entry_safe(ck, t, &bc->freed, list) {
572 if (!poll_state_synchronize_srcu(&c->btree_trans_barrier,
573 ck->btree_trans_barrier_seq))
577 kmem_cache_free(bch2_key_cache, ck);
587 tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
588 if (bc->shrink_iter >= tbl->size)
590 start = bc->shrink_iter;
593 struct rhash_head *pos, *next;
595 pos = rht_ptr_rcu(rht_bucket(tbl, bc->shrink_iter));
597 while (!rht_is_a_nulls(pos)) {
598 next = rht_dereference_bucket_rcu(pos->next, tbl, bc->shrink_iter);
599 ck = container_of(pos, struct bkey_cached, hash);
601 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags))
604 if (test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
605 clear_bit(BKEY_CACHED_ACCESSED, &ck->flags);
606 else if (bkey_cached_lock_for_evict(ck)) {
607 bkey_cached_evict(bc, ck);
608 bkey_cached_free(bc, ck);
619 if (bc->shrink_iter >= tbl->size)
621 } while (scanned < nr && bc->shrink_iter != start);
625 memalloc_nofs_restore(flags);
626 srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
627 mutex_unlock(&bc->lock);
632 static unsigned long bch2_btree_key_cache_count(struct shrinker *shrink,
633 struct shrink_control *sc)
635 struct bch_fs *c = container_of(shrink, struct bch_fs,
636 btree_key_cache.shrink);
637 struct btree_key_cache *bc = &c->btree_key_cache;
638 long nr = atomic_long_read(&bc->nr_keys) -
639 atomic_long_read(&bc->nr_dirty);
644 void bch2_fs_btree_key_cache_exit(struct btree_key_cache *bc)
646 struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
647 struct bucket_table *tbl;
648 struct bkey_cached *ck, *n;
649 struct rhash_head *pos;
652 if (bc->shrink.list.next)
653 unregister_shrinker(&bc->shrink);
655 mutex_lock(&bc->lock);
658 tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
659 for (i = 0; i < tbl->size; i++)
660 rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
661 bkey_cached_evict(bc, ck);
662 list_add(&ck->list, &bc->freed);
666 list_for_each_entry_safe(ck, n, &bc->freed, list) {
669 bch2_journal_pin_drop(&c->journal, &ck->journal);
670 bch2_journal_preres_put(&c->journal, &ck->res);
674 kmem_cache_free(bch2_key_cache, ck);
677 BUG_ON(atomic_long_read(&bc->nr_dirty) &&
678 !bch2_journal_error(&c->journal) &&
679 test_bit(BCH_FS_WAS_RW, &c->flags));
680 BUG_ON(atomic_long_read(&bc->nr_keys));
682 mutex_unlock(&bc->lock);
684 if (bc->table_init_done)
685 rhashtable_destroy(&bc->table);
688 void bch2_fs_btree_key_cache_init_early(struct btree_key_cache *c)
690 mutex_init(&c->lock);
691 INIT_LIST_HEAD(&c->freed);
694 int bch2_fs_btree_key_cache_init(struct btree_key_cache *c)
698 ret = rhashtable_init(&c->table, &bch2_btree_key_cache_params);
702 c->table_init_done = true;
705 c->shrink.count_objects = bch2_btree_key_cache_count;
706 c->shrink.scan_objects = bch2_btree_key_cache_scan;
707 return register_shrinker(&c->shrink);
710 void bch2_btree_key_cache_to_text(struct printbuf *out, struct btree_key_cache *c)
712 pr_buf(out, "nr_freed:\t%zu\n", c->nr_freed);
713 pr_buf(out, "nr_keys:\t%zu\n", atomic_long_read(&c->nr_keys));
714 pr_buf(out, "nr_dirty:\t%zu\n", atomic_long_read(&c->nr_dirty));
717 void bch2_btree_key_cache_exit(void)
720 kmem_cache_destroy(bch2_key_cache);
723 int __init bch2_btree_key_cache_init(void)
725 bch2_key_cache = KMEM_CACHE(bkey_cached, 0);