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_iter *ck_iter,
200 struct bkey_cached *ck)
202 struct btree_iter *iter;
204 unsigned new_u64s = 0;
205 struct bkey_i *new_k = NULL;
208 iter = bch2_trans_get_iter(trans, 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(ck_iter, 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(ck_iter->l[0].b, ck_iter);
247 bkey_reassemble(ck->k, k);
249 bch2_btree_node_unlock_write(ck_iter->l[0].b, ck_iter);
251 /* We're not likely to need this iterator again: */
252 set_btree_iter_dontneed(trans, iter);
254 bch2_trans_iter_put(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_iter *iter = p;
263 return ck->key.btree_id == iter->btree_id &&
264 !bpos_cmp(ck->key.pos, iter->pos) ? 0 : -1;
268 int bch2_btree_iter_traverse_cached(struct btree_iter *iter)
270 struct btree_trans *trans = iter->trans;
271 struct bch_fs *c = trans->c;
272 struct bkey_cached *ck;
279 if (bch2_btree_node_relock(iter, 0)) {
280 ck = (void *) iter->l[0].b;
284 ck = bch2_btree_key_cache_find(c, iter->btree_id, iter->pos);
286 if (iter->flags & BTREE_ITER_CACHED_NOCREATE) {
291 ck = btree_key_cache_create(&c->btree_key_cache,
292 iter->btree_id, iter->pos);
293 ret = PTR_ERR_OR_ZERO(ck);
299 mark_btree_node_locked(iter, 0, SIX_LOCK_intent);
300 iter->locks_want = 1;
302 enum six_lock_type lock_want = __btree_lock_want(iter, 0);
304 if (!btree_node_lock((void *) ck, iter->pos, 0, iter, lock_want,
305 bkey_cached_check_fn, iter, _THIS_IP_)) {
306 if (!trans->restarted)
309 trace_transaction_restart_ip(trans->ip, _THIS_IP_);
314 if (ck->key.btree_id != iter->btree_id ||
315 bpos_cmp(ck->key.pos, iter->pos)) {
316 six_unlock_type(&ck->c.lock, lock_want);
320 mark_btree_node_locked(iter, 0, lock_want);
323 iter->l[0].lock_seq = ck->c.lock.state.seq;
324 iter->l[0].b = (void *) ck;
326 if (!ck->valid && !(iter->flags & BTREE_ITER_CACHED_NOFILL)) {
327 if (!iter->locks_want &&
328 !!__bch2_btree_iter_upgrade(iter, 1)) {
329 trace_transaction_restart_ip(trans->ip, _THIS_IP_);
330 BUG_ON(!trans->restarted);
335 ret = btree_key_cache_fill(trans, iter, ck);
340 if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
341 set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
343 iter->uptodate = BTREE_ITER_NEED_PEEK;
345 if ((iter->flags & BTREE_ITER_INTENT) &&
346 !bch2_btree_iter_upgrade(iter, 1)) {
347 BUG_ON(!trans->restarted);
351 BUG_ON(!ret && !btree_node_locked(iter, 0));
356 btree_node_unlock(iter, 0);
357 iter->flags |= BTREE_ITER_ERROR;
358 iter->l[0].b = BTREE_ITER_NO_NODE_ERROR;
363 static int btree_key_cache_flush_pos(struct btree_trans *trans,
364 struct bkey_cached_key key,
366 unsigned commit_flags,
369 struct bch_fs *c = trans->c;
370 struct journal *j = &c->journal;
371 struct btree_iter *c_iter = NULL, *b_iter = NULL;
372 struct bkey_cached *ck = NULL;
375 b_iter = bch2_trans_get_iter(trans, key.btree_id, key.pos,
378 c_iter = bch2_trans_get_iter(trans, key.btree_id, key.pos,
380 BTREE_ITER_CACHED_NOFILL|
381 BTREE_ITER_CACHED_NOCREATE|
383 ret = bch2_btree_iter_traverse(c_iter);
387 ck = (void *) c_iter->l[0].b;
389 (journal_seq && ck->journal.seq != journal_seq))
392 if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
399 * Since journal reclaim depends on us making progress here, and the
400 * allocator/copygc depend on journal reclaim making progress, we need
401 * to be using alloc reserves:
403 ret = bch2_btree_iter_traverse(b_iter) ?:
404 bch2_trans_update(trans, b_iter, ck->k,
405 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
406 BTREE_TRIGGER_NORUN) ?:
407 bch2_trans_commit(trans, NULL, NULL,
408 BTREE_INSERT_NOCHECK_RW|
410 BTREE_INSERT_USE_RESERVE|
411 (ck->journal.seq == journal_last_seq(j)
412 ? BTREE_INSERT_JOURNAL_RESERVED
416 bch2_fs_fatal_err_on(ret != -EINTR &&
418 !bch2_journal_error(j), c,
419 "error flushing key cache: %i", ret);
423 bch2_journal_pin_drop(j, &ck->journal);
424 bch2_journal_preres_put(j, &ck->res);
426 BUG_ON(!btree_node_locked(c_iter, 0));
429 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
430 clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
431 atomic_long_dec(&c->btree_key_cache.nr_dirty);
435 BUG_ON(!btree_node_intent_locked(c_iter, 0));
437 mark_btree_node_unlocked(c_iter, 0);
438 c_iter->l[0].b = NULL;
440 six_lock_write(&ck->c.lock, NULL, NULL);
442 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
443 clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
444 atomic_long_dec(&c->btree_key_cache.nr_dirty);
447 bkey_cached_evict(&c->btree_key_cache, ck);
449 mutex_lock(&c->btree_key_cache.lock);
450 bkey_cached_free(&c->btree_key_cache, ck);
451 mutex_unlock(&c->btree_key_cache.lock);
454 bch2_trans_iter_put(trans, b_iter);
455 bch2_trans_iter_put(trans, c_iter);
459 int bch2_btree_key_cache_journal_flush(struct journal *j,
460 struct journal_entry_pin *pin, u64 seq)
462 struct bch_fs *c = container_of(j, struct bch_fs, journal);
463 struct bkey_cached *ck =
464 container_of(pin, struct bkey_cached, journal);
465 struct bkey_cached_key key;
468 int srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
470 six_lock_read(&ck->c.lock, NULL, NULL);
473 if (ck->journal.seq != seq ||
474 !test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
475 six_unlock_read(&ck->c.lock);
478 six_unlock_read(&ck->c.lock);
480 ret = bch2_trans_do(c, NULL, NULL, 0,
481 btree_key_cache_flush_pos(&trans, key, seq,
482 BTREE_INSERT_JOURNAL_RECLAIM, false));
484 srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
490 * Flush and evict a key from the key cache:
492 int bch2_btree_key_cache_flush(struct btree_trans *trans,
493 enum btree_id id, struct bpos pos)
495 struct bch_fs *c = trans->c;
496 struct bkey_cached_key key = { id, pos };
498 /* Fastpath - assume it won't be found: */
499 if (!bch2_btree_key_cache_find(c, id, pos))
502 return btree_key_cache_flush_pos(trans, key, 0, 0, true);
505 bool bch2_btree_insert_key_cached(struct btree_trans *trans,
506 struct btree_iter *iter,
507 struct bkey_i *insert)
509 struct bch_fs *c = trans->c;
510 struct bkey_cached *ck = (void *) iter->l[0].b;
511 bool kick_reclaim = false;
513 BUG_ON(insert->u64s > ck->u64s);
515 if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
518 BUG_ON(jset_u64s(insert->u64s) > trans->journal_preres.u64s);
520 difference = jset_u64s(insert->u64s) - ck->res.u64s;
521 if (difference > 0) {
522 trans->journal_preres.u64s -= difference;
523 ck->res.u64s += difference;
527 bkey_copy(ck->k, insert);
530 if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
531 set_bit(BKEY_CACHED_DIRTY, &ck->flags);
532 atomic_long_inc(&c->btree_key_cache.nr_dirty);
534 if (bch2_nr_btree_keys_need_flush(c))
538 bch2_journal_pin_update(&c->journal, trans->journal_res.seq,
539 &ck->journal, bch2_btree_key_cache_journal_flush);
542 journal_reclaim_kick(&c->journal);
546 #ifdef CONFIG_BCACHEFS_DEBUG
547 void bch2_btree_key_cache_verify_clean(struct btree_trans *trans,
548 enum btree_id id, struct bpos pos)
550 BUG_ON(bch2_btree_key_cache_find(trans->c, id, pos));
554 static unsigned long bch2_btree_key_cache_scan(struct shrinker *shrink,
555 struct shrink_control *sc)
557 struct bch_fs *c = container_of(shrink, struct bch_fs,
558 btree_key_cache.shrink);
559 struct btree_key_cache *bc = &c->btree_key_cache;
560 struct bucket_table *tbl;
561 struct bkey_cached *ck, *t;
562 size_t scanned = 0, freed = 0, nr = sc->nr_to_scan;
563 unsigned start, flags;
566 /* Return -1 if we can't do anything right now */
567 if (sc->gfp_mask & __GFP_FS)
568 mutex_lock(&bc->lock);
569 else if (!mutex_trylock(&bc->lock))
572 srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
573 flags = memalloc_nofs_save();
576 * Newest freed entries are at the end of the list - once we hit one
577 * that's too new to be freed, we can bail out:
579 list_for_each_entry_safe(ck, t, &bc->freed, list) {
580 if (!poll_state_synchronize_srcu(&c->btree_trans_barrier,
581 ck->btree_trans_barrier_seq))
585 kmem_cache_free(bch2_key_cache, ck);
595 tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
596 if (bc->shrink_iter >= tbl->size)
598 start = bc->shrink_iter;
601 struct rhash_head *pos, *next;
603 pos = rht_ptr_rcu(rht_bucket(tbl, bc->shrink_iter));
605 while (!rht_is_a_nulls(pos)) {
606 next = rht_dereference_bucket_rcu(pos->next, tbl, bc->shrink_iter);
607 ck = container_of(pos, struct bkey_cached, hash);
609 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags))
612 if (test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
613 clear_bit(BKEY_CACHED_ACCESSED, &ck->flags);
614 else if (bkey_cached_lock_for_evict(ck)) {
615 bkey_cached_evict(bc, ck);
616 bkey_cached_free(bc, ck);
627 if (bc->shrink_iter >= tbl->size)
629 } while (scanned < nr && bc->shrink_iter != start);
633 memalloc_nofs_restore(flags);
634 srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
635 mutex_unlock(&bc->lock);
640 static unsigned long bch2_btree_key_cache_count(struct shrinker *shrink,
641 struct shrink_control *sc)
643 struct bch_fs *c = container_of(shrink, struct bch_fs,
644 btree_key_cache.shrink);
645 struct btree_key_cache *bc = &c->btree_key_cache;
646 long nr = atomic_long_read(&bc->nr_keys) -
647 atomic_long_read(&bc->nr_dirty);
652 void bch2_fs_btree_key_cache_exit(struct btree_key_cache *bc)
654 struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
655 struct bucket_table *tbl;
656 struct bkey_cached *ck, *n;
657 struct rhash_head *pos;
660 if (bc->shrink.list.next)
661 unregister_shrinker(&bc->shrink);
663 mutex_lock(&bc->lock);
666 tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
667 for (i = 0; i < tbl->size; i++)
668 rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
669 bkey_cached_evict(bc, ck);
670 list_add(&ck->list, &bc->freed);
674 list_for_each_entry_safe(ck, n, &bc->freed, list) {
677 bch2_journal_pin_drop(&c->journal, &ck->journal);
678 bch2_journal_preres_put(&c->journal, &ck->res);
682 kmem_cache_free(bch2_key_cache, ck);
685 BUG_ON(atomic_long_read(&bc->nr_dirty) &&
686 !bch2_journal_error(&c->journal) &&
687 test_bit(BCH_FS_WAS_RW, &c->flags));
688 BUG_ON(atomic_long_read(&bc->nr_keys));
690 mutex_unlock(&bc->lock);
692 if (bc->table_init_done)
693 rhashtable_destroy(&bc->table);
696 void bch2_fs_btree_key_cache_init_early(struct btree_key_cache *c)
698 mutex_init(&c->lock);
699 INIT_LIST_HEAD(&c->freed);
702 int bch2_fs_btree_key_cache_init(struct btree_key_cache *c)
706 ret = rhashtable_init(&c->table, &bch2_btree_key_cache_params);
710 c->table_init_done = true;
713 c->shrink.count_objects = bch2_btree_key_cache_count;
714 c->shrink.scan_objects = bch2_btree_key_cache_scan;
715 return register_shrinker(&c->shrink);
718 void bch2_btree_key_cache_to_text(struct printbuf *out, struct btree_key_cache *c)
720 pr_buf(out, "nr_freed:\t%zu\n", c->nr_freed);
721 pr_buf(out, "nr_keys:\t%zu\n", atomic_long_read(&c->nr_keys));
722 pr_buf(out, "nr_dirty:\t%zu\n", atomic_long_read(&c->nr_dirty));
725 void bch2_btree_key_cache_exit(void)
728 kmem_cache_destroy(bch2_key_cache);
731 int __init bch2_btree_key_cache_init(void)
733 bch2_key_cache = KMEM_CACHE(bkey_cached, 0);