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"
13 #include <linux/sched/mm.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 lockdep_set_novalidate_class(&ck->c.lock);
107 BUG_ON(!six_trylock_intent(&ck->c.lock));
108 BUG_ON(!six_trylock_write(&ck->c.lock));
115 static struct bkey_cached *
116 bkey_cached_reuse(struct btree_key_cache *c)
118 struct bucket_table *tbl;
119 struct rhash_head *pos;
120 struct bkey_cached *ck;
123 mutex_lock(&c->lock);
124 list_for_each_entry_reverse(ck, &c->freed, list)
125 if (bkey_cached_lock_for_evict(ck)) {
128 mutex_unlock(&c->lock);
131 mutex_unlock(&c->lock);
134 tbl = rht_dereference_rcu(c->table.tbl, &c->table);
135 for (i = 0; i < tbl->size; i++)
136 rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
137 if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
138 bkey_cached_lock_for_evict(ck)) {
139 bkey_cached_evict(c, ck);
149 static struct bkey_cached *
150 btree_key_cache_create(struct btree_key_cache *c,
151 enum btree_id btree_id,
154 struct bkey_cached *ck;
157 ck = bkey_cached_alloc(c);
160 ck = bkey_cached_reuse(c);
162 return ERR_PTR(-ENOMEM);
168 ck->c.btree_id = btree_id;
169 ck->key.btree_id = btree_id;
172 ck->flags = 1U << BKEY_CACHED_ACCESSED;
174 if (unlikely(rhashtable_lookup_insert_fast(&c->table,
176 bch2_btree_key_cache_params))) {
177 /* We raced with another fill: */
179 if (likely(was_new)) {
180 six_unlock_write(&ck->c.lock);
181 six_unlock_intent(&ck->c.lock);
184 mutex_lock(&c->lock);
185 bkey_cached_free(c, ck);
186 mutex_unlock(&c->lock);
192 atomic_long_inc(&c->nr_keys);
194 six_unlock_write(&ck->c.lock);
199 static int btree_key_cache_fill(struct btree_trans *trans,
200 struct btree_iter *ck_iter,
201 struct bkey_cached *ck)
203 struct btree_iter *iter;
205 unsigned new_u64s = 0;
206 struct bkey_i *new_k = NULL;
209 iter = bch2_trans_get_iter(trans, ck->key.btree_id,
210 ck->key.pos, BTREE_ITER_SLOTS);
211 k = bch2_btree_iter_peek_slot(iter);
216 if (!bch2_btree_node_relock(ck_iter, 0)) {
217 trace_transaction_restart_ip(trans->ip, _THIS_IP_);
222 if (k.k->u64s > ck->u64s) {
223 new_u64s = roundup_pow_of_two(k.k->u64s);
224 new_k = kmalloc(new_u64s * sizeof(u64), GFP_NOFS);
231 bch2_btree_node_lock_write(ck_iter->l[0].b, ck_iter);
238 bkey_reassemble(ck->k, k);
240 bch2_btree_node_unlock_write(ck_iter->l[0].b, ck_iter);
242 /* We're not likely to need this iterator again: */
243 set_btree_iter_dontneed(trans, iter);
245 bch2_trans_iter_put(trans, iter);
249 static int bkey_cached_check_fn(struct six_lock *lock, void *p)
251 struct bkey_cached *ck = container_of(lock, struct bkey_cached, c.lock);
252 const struct btree_iter *iter = p;
254 return ck->key.btree_id == iter->btree_id &&
255 !bpos_cmp(ck->key.pos, iter->pos) ? 0 : -1;
259 int bch2_btree_iter_traverse_cached(struct btree_iter *iter)
261 struct btree_trans *trans = iter->trans;
262 struct bch_fs *c = trans->c;
263 struct bkey_cached *ck;
268 if (btree_node_locked(iter, 0)) {
269 ck = (void *) iter->l[0].b;
273 ck = bch2_btree_key_cache_find(c, iter->btree_id, iter->pos);
275 if (iter->flags & BTREE_ITER_CACHED_NOCREATE) {
280 ck = btree_key_cache_create(&c->btree_key_cache,
281 iter->btree_id, iter->pos);
282 ret = PTR_ERR_OR_ZERO(ck);
288 mark_btree_node_locked(iter, 0, SIX_LOCK_intent);
289 iter->locks_want = 1;
291 enum six_lock_type lock_want = __btree_lock_want(iter, 0);
293 if (!btree_node_lock((void *) ck, iter->pos, 0, iter, lock_want,
294 bkey_cached_check_fn, iter, _THIS_IP_)) {
295 if (ck->key.btree_id != iter->btree_id ||
296 bpos_cmp(ck->key.pos, iter->pos)) {
300 trace_transaction_restart_ip(trans->ip, _THIS_IP_);
305 if (ck->key.btree_id != iter->btree_id ||
306 bpos_cmp(ck->key.pos, iter->pos)) {
307 six_unlock_type(&ck->c.lock, lock_want);
311 mark_btree_node_locked(iter, 0, lock_want);
314 iter->l[0].lock_seq = ck->c.lock.state.seq;
315 iter->l[0].b = (void *) ck;
317 if (!ck->valid && !(iter->flags & BTREE_ITER_CACHED_NOFILL)) {
318 if (!btree_node_intent_locked(iter, 0))
319 bch2_btree_iter_upgrade(iter, 1);
320 if (!btree_node_intent_locked(iter, 0)) {
321 trace_transaction_restart_ip(trans->ip, _THIS_IP_);
326 ret = btree_key_cache_fill(trans, iter, ck);
331 if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
332 set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
334 iter->uptodate = BTREE_ITER_NEED_PEEK;
336 if (!(iter->flags & BTREE_ITER_INTENT))
337 bch2_btree_iter_downgrade(iter);
338 else if (!iter->locks_want) {
339 if (!__bch2_btree_iter_upgrade(iter, 1))
346 btree_node_unlock(iter, 0);
347 iter->flags |= BTREE_ITER_ERROR;
348 iter->l[0].b = BTREE_ITER_NO_NODE_ERROR;
353 static int btree_key_cache_flush_pos(struct btree_trans *trans,
354 struct bkey_cached_key key,
356 unsigned commit_flags,
359 struct bch_fs *c = trans->c;
360 struct journal *j = &c->journal;
361 struct btree_iter *c_iter = NULL, *b_iter = NULL;
362 struct bkey_cached *ck = NULL;
365 b_iter = bch2_trans_get_iter(trans, key.btree_id, key.pos,
368 c_iter = bch2_trans_get_iter(trans, key.btree_id, key.pos,
370 BTREE_ITER_CACHED_NOFILL|
371 BTREE_ITER_CACHED_NOCREATE|
374 ret = bch2_btree_iter_traverse(c_iter);
378 ck = (void *) c_iter->l[0].b;
380 (journal_seq && ck->journal.seq != journal_seq))
383 if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
389 ret = bch2_btree_iter_traverse(b_iter) ?:
390 bch2_trans_update(trans, b_iter, ck->k, BTREE_TRIGGER_NORUN) ?:
391 bch2_trans_commit(trans, NULL, NULL,
392 BTREE_INSERT_NOUNLOCK|
393 BTREE_INSERT_NOCHECK_RW|
395 (ck->journal.seq == journal_last_seq(j)
396 ? BTREE_INSERT_JOURNAL_RESERVED
407 bch2_fs_fatal_err_on(!bch2_journal_error(j), c,
408 "error flushing key cache: %i", ret);
412 bch2_journal_pin_drop(j, &ck->journal);
413 bch2_journal_preres_put(j, &ck->res);
415 BUG_ON(!btree_node_locked(c_iter, 0));
418 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
419 clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
420 atomic_long_dec(&c->btree_key_cache.nr_dirty);
424 BUG_ON(!btree_node_intent_locked(c_iter, 0));
426 mark_btree_node_unlocked(c_iter, 0);
427 c_iter->l[0].b = NULL;
429 six_lock_write(&ck->c.lock, NULL, NULL);
431 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
432 clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
433 atomic_long_dec(&c->btree_key_cache.nr_dirty);
436 bkey_cached_evict(&c->btree_key_cache, ck);
438 mutex_lock(&c->btree_key_cache.lock);
439 bkey_cached_free(&c->btree_key_cache, ck);
440 mutex_unlock(&c->btree_key_cache.lock);
443 bch2_trans_iter_put(trans, b_iter);
444 bch2_trans_iter_put(trans, c_iter);
448 int bch2_btree_key_cache_journal_flush(struct journal *j,
449 struct journal_entry_pin *pin, u64 seq)
451 struct bch_fs *c = container_of(j, struct bch_fs, journal);
452 struct bkey_cached *ck =
453 container_of(pin, struct bkey_cached, journal);
454 struct bkey_cached_key key;
455 struct btree_trans trans;
458 int srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
460 six_lock_read(&ck->c.lock, NULL, NULL);
463 if (ck->journal.seq != seq ||
464 !test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
465 six_unlock_read(&ck->c.lock);
468 six_unlock_read(&ck->c.lock);
470 bch2_trans_init(&trans, c, 0, 0);
471 ret = btree_key_cache_flush_pos(&trans, key, seq,
472 BTREE_INSERT_JOURNAL_RECLAIM, false);
473 bch2_trans_exit(&trans);
475 srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
481 * Flush and evict a key from the key cache:
483 int bch2_btree_key_cache_flush(struct btree_trans *trans,
484 enum btree_id id, struct bpos pos)
486 struct bch_fs *c = trans->c;
487 struct bkey_cached_key key = { id, pos };
489 /* Fastpath - assume it won't be found: */
490 if (!bch2_btree_key_cache_find(c, id, pos))
493 return btree_key_cache_flush_pos(trans, key, 0, 0, true);
496 bool bch2_btree_insert_key_cached(struct btree_trans *trans,
497 struct btree_iter *iter,
498 struct bkey_i *insert)
500 struct bch_fs *c = trans->c;
501 struct bkey_cached *ck = (void *) iter->l[0].b;
502 bool kick_reclaim = false;
504 BUG_ON(insert->u64s > ck->u64s);
506 if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
509 BUG_ON(jset_u64s(insert->u64s) > trans->journal_preres.u64s);
511 difference = jset_u64s(insert->u64s) - ck->res.u64s;
512 if (difference > 0) {
513 trans->journal_preres.u64s -= difference;
514 ck->res.u64s += difference;
518 bkey_copy(ck->k, insert);
521 if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
522 set_bit(BKEY_CACHED_DIRTY, &ck->flags);
523 atomic_long_inc(&c->btree_key_cache.nr_dirty);
525 if (bch2_nr_btree_keys_need_flush(c))
529 bch2_journal_pin_update(&c->journal, trans->journal_res.seq,
530 &ck->journal, bch2_btree_key_cache_journal_flush);
533 journal_reclaim_kick(&c->journal);
537 #ifdef CONFIG_BCACHEFS_DEBUG
538 void bch2_btree_key_cache_verify_clean(struct btree_trans *trans,
539 enum btree_id id, struct bpos pos)
541 BUG_ON(bch2_btree_key_cache_find(trans->c, id, pos));
545 static unsigned long bch2_btree_key_cache_scan(struct shrinker *shrink,
546 struct shrink_control *sc)
548 struct bch_fs *c = container_of(shrink, struct bch_fs,
549 btree_key_cache.shrink);
550 struct btree_key_cache *bc = &c->btree_key_cache;
551 struct bucket_table *tbl;
552 struct bkey_cached *ck, *t;
553 size_t scanned = 0, freed = 0, nr = sc->nr_to_scan;
554 unsigned start, flags;
557 /* Return -1 if we can't do anything right now */
558 if (sc->gfp_mask & __GFP_FS)
559 mutex_lock(&bc->lock);
560 else if (!mutex_trylock(&bc->lock))
563 srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
564 flags = memalloc_nofs_save();
567 * Newest freed entries are at the end of the list - once we hit one
568 * that's too new to be freed, we can bail out:
570 list_for_each_entry_safe(ck, t, &bc->freed, list) {
571 if (!poll_state_synchronize_srcu(&c->btree_trans_barrier,
572 ck->btree_trans_barrier_seq))
576 kmem_cache_free(bch2_key_cache, ck);
586 tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
587 if (bc->shrink_iter >= tbl->size)
589 start = bc->shrink_iter;
592 struct rhash_head *pos, *next;
594 pos = rht_ptr_rcu(rht_bucket(tbl, bc->shrink_iter));
596 while (!rht_is_a_nulls(pos)) {
597 next = rht_dereference_bucket_rcu(pos->next, tbl, bc->shrink_iter);
598 ck = container_of(pos, struct bkey_cached, hash);
600 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags))
603 if (test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
604 clear_bit(BKEY_CACHED_ACCESSED, &ck->flags);
605 else if (bkey_cached_lock_for_evict(ck)) {
606 bkey_cached_evict(bc, ck);
607 bkey_cached_free(bc, ck);
618 if (bc->shrink_iter >= tbl->size)
620 } while (scanned < nr && bc->shrink_iter != start);
624 memalloc_nofs_restore(flags);
625 srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
626 mutex_unlock(&bc->lock);
631 static unsigned long bch2_btree_key_cache_count(struct shrinker *shrink,
632 struct shrink_control *sc)
634 struct bch_fs *c = container_of(shrink, struct bch_fs,
635 btree_key_cache.shrink);
636 struct btree_key_cache *bc = &c->btree_key_cache;
638 return atomic_long_read(&bc->nr_keys);
641 void bch2_fs_btree_key_cache_exit(struct btree_key_cache *bc)
643 struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
644 struct bucket_table *tbl;
645 struct bkey_cached *ck, *n;
646 struct rhash_head *pos;
649 if (bc->shrink.list.next)
650 unregister_shrinker(&bc->shrink);
652 mutex_lock(&bc->lock);
655 tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
656 for (i = 0; i < tbl->size; i++)
657 rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
658 bkey_cached_evict(bc, ck);
659 list_add(&ck->list, &bc->freed);
663 list_for_each_entry_safe(ck, n, &bc->freed, list) {
666 bch2_journal_pin_drop(&c->journal, &ck->journal);
667 bch2_journal_preres_put(&c->journal, &ck->res);
671 kmem_cache_free(bch2_key_cache, ck);
674 BUG_ON(atomic_long_read(&bc->nr_dirty) && !bch2_journal_error(&c->journal));
675 BUG_ON(atomic_long_read(&bc->nr_keys));
677 mutex_unlock(&bc->lock);
679 if (bc->table_init_done)
680 rhashtable_destroy(&bc->table);
683 void bch2_fs_btree_key_cache_init_early(struct btree_key_cache *c)
685 mutex_init(&c->lock);
686 INIT_LIST_HEAD(&c->freed);
689 int bch2_fs_btree_key_cache_init(struct btree_key_cache *bc)
691 struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
694 ret = rhashtable_init(&bc->table, &bch2_btree_key_cache_params);
698 bc->table_init_done = true;
700 bc->shrink.seeks = 1;
701 bc->shrink.count_objects = bch2_btree_key_cache_count;
702 bc->shrink.scan_objects = bch2_btree_key_cache_scan;
703 return register_shrinker(&bc->shrink, "%s/btree_key_cache", c->name);
706 void bch2_btree_key_cache_to_text(struct printbuf *out, struct btree_key_cache *c)
708 pr_buf(out, "nr_freed:\t%zu\n", c->nr_freed);
709 pr_buf(out, "nr_keys:\t%zu\n", atomic_long_read(&c->nr_keys));
710 pr_buf(out, "nr_dirty:\t%zu\n", atomic_long_read(&c->nr_dirty));
713 void bch2_btree_key_cache_exit(void)
716 kmem_cache_destroy(bch2_key_cache);
719 int __init bch2_btree_key_cache_init(void)
721 bch2_key_cache = KMEM_CACHE(bkey_cached, 0);