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_);
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);
236 bch2_btree_node_lock_write(ck_iter->l[0].b, ck_iter);
243 bkey_reassemble(ck->k, k);
245 bch2_btree_node_unlock_write(ck_iter->l[0].b, ck_iter);
247 /* We're not likely to need this iterator again: */
248 set_btree_iter_dontneed(trans, iter);
250 bch2_trans_iter_put(trans, iter);
254 static int bkey_cached_check_fn(struct six_lock *lock, void *p)
256 struct bkey_cached *ck = container_of(lock, struct bkey_cached, c.lock);
257 const struct btree_iter *iter = p;
259 return ck->key.btree_id == iter->btree_id &&
260 !bpos_cmp(ck->key.pos, iter->pos) ? 0 : -1;
264 int bch2_btree_iter_traverse_cached(struct btree_iter *iter)
266 struct btree_trans *trans = iter->trans;
267 struct bch_fs *c = trans->c;
268 struct bkey_cached *ck;
275 if (bch2_btree_node_relock(iter, 0)) {
276 ck = (void *) iter->l[0].b;
280 ck = bch2_btree_key_cache_find(c, iter->btree_id, iter->pos);
282 if (iter->flags & BTREE_ITER_CACHED_NOCREATE) {
287 ck = btree_key_cache_create(&c->btree_key_cache,
288 iter->btree_id, iter->pos);
289 ret = PTR_ERR_OR_ZERO(ck);
295 mark_btree_node_locked(iter, 0, SIX_LOCK_intent);
296 iter->locks_want = 1;
298 enum six_lock_type lock_want = __btree_lock_want(iter, 0);
300 if (!btree_node_lock((void *) ck, iter->pos, 0, iter, lock_want,
301 bkey_cached_check_fn, iter, _THIS_IP_)) {
302 if (ck->key.btree_id != iter->btree_id ||
303 bpos_cmp(ck->key.pos, iter->pos)) {
307 trace_transaction_restart_ip(trans->ip, _THIS_IP_);
312 if (ck->key.btree_id != iter->btree_id ||
313 bpos_cmp(ck->key.pos, iter->pos)) {
314 six_unlock_type(&ck->c.lock, lock_want);
318 mark_btree_node_locked(iter, 0, lock_want);
321 iter->l[0].lock_seq = ck->c.lock.state.seq;
322 iter->l[0].b = (void *) ck;
324 if (!ck->valid && !(iter->flags & BTREE_ITER_CACHED_NOFILL)) {
325 if (!btree_node_intent_locked(iter, 0))
326 bch2_btree_iter_upgrade(iter, 1);
327 if (!btree_node_intent_locked(iter, 0)) {
328 trace_transaction_restart_ip(trans->ip, _THIS_IP_);
333 ret = btree_key_cache_fill(trans, iter, ck);
338 if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
339 set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
341 iter->uptodate = BTREE_ITER_NEED_PEEK;
343 if (!(iter->flags & BTREE_ITER_INTENT))
344 bch2_btree_iter_downgrade(iter);
345 else if (!iter->locks_want) {
346 if (!__bch2_btree_iter_upgrade(iter, 1))
353 btree_node_unlock(iter, 0);
354 iter->flags |= BTREE_ITER_ERROR;
355 iter->l[0].b = BTREE_ITER_NO_NODE_ERROR;
360 static int btree_key_cache_flush_pos(struct btree_trans *trans,
361 struct bkey_cached_key key,
363 unsigned commit_flags,
366 struct bch_fs *c = trans->c;
367 struct journal *j = &c->journal;
368 struct btree_iter *c_iter = NULL, *b_iter = NULL;
369 struct bkey_cached *ck = NULL;
372 b_iter = bch2_trans_get_iter(trans, key.btree_id, key.pos,
375 c_iter = bch2_trans_get_iter(trans, key.btree_id, key.pos,
377 BTREE_ITER_CACHED_NOFILL|
378 BTREE_ITER_CACHED_NOCREATE|
381 ret = bch2_btree_iter_traverse(c_iter);
385 ck = (void *) c_iter->l[0].b;
387 (journal_seq && ck->journal.seq != journal_seq))
390 if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
397 * Since journal reclaim depends on us making progress here, and the
398 * allocator/copygc depend on journal reclaim making progress, we need
399 * to be using alloc reserves:
401 ret = bch2_btree_iter_traverse(b_iter) ?:
402 bch2_trans_update(trans, b_iter, ck->k, BTREE_TRIGGER_NORUN) ?:
403 bch2_trans_commit(trans, NULL, NULL,
404 BTREE_INSERT_NOUNLOCK|
405 BTREE_INSERT_NOCHECK_RW|
407 BTREE_INSERT_USE_RESERVE|
408 (ck->journal.seq == journal_last_seq(j)
409 ? BTREE_INSERT_JOURNAL_RESERVED
420 bch2_fs_fatal_err_on(!bch2_journal_error(j), c,
421 "error flushing key cache: %i", ret);
425 bch2_journal_pin_drop(j, &ck->journal);
426 bch2_journal_preres_put(j, &ck->res);
428 BUG_ON(!btree_node_locked(c_iter, 0));
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);
437 BUG_ON(!btree_node_intent_locked(c_iter, 0));
439 mark_btree_node_unlocked(c_iter, 0);
440 c_iter->l[0].b = NULL;
442 six_lock_write(&ck->c.lock, NULL, NULL);
444 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
445 clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
446 atomic_long_dec(&c->btree_key_cache.nr_dirty);
449 bkey_cached_evict(&c->btree_key_cache, ck);
451 mutex_lock(&c->btree_key_cache.lock);
452 bkey_cached_free(&c->btree_key_cache, ck);
453 mutex_unlock(&c->btree_key_cache.lock);
456 bch2_trans_iter_put(trans, b_iter);
457 bch2_trans_iter_put(trans, c_iter);
461 int bch2_btree_key_cache_journal_flush(struct journal *j,
462 struct journal_entry_pin *pin, u64 seq)
464 struct bch_fs *c = container_of(j, struct bch_fs, journal);
465 struct bkey_cached *ck =
466 container_of(pin, struct bkey_cached, journal);
467 struct bkey_cached_key key;
468 struct btree_trans trans;
471 int srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
473 six_lock_read(&ck->c.lock, NULL, NULL);
476 if (ck->journal.seq != seq ||
477 !test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
478 six_unlock_read(&ck->c.lock);
481 six_unlock_read(&ck->c.lock);
483 bch2_trans_init(&trans, c, 0, 0);
484 ret = btree_key_cache_flush_pos(&trans, key, seq,
485 BTREE_INSERT_JOURNAL_RECLAIM, false);
486 bch2_trans_exit(&trans);
488 srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
494 * Flush and evict a key from the key cache:
496 int bch2_btree_key_cache_flush(struct btree_trans *trans,
497 enum btree_id id, struct bpos pos)
499 struct bch_fs *c = trans->c;
500 struct bkey_cached_key key = { id, pos };
502 /* Fastpath - assume it won't be found: */
503 if (!bch2_btree_key_cache_find(c, id, pos))
506 return btree_key_cache_flush_pos(trans, key, 0, 0, true);
509 bool bch2_btree_insert_key_cached(struct btree_trans *trans,
510 struct btree_iter *iter,
511 struct bkey_i *insert)
513 struct bch_fs *c = trans->c;
514 struct bkey_cached *ck = (void *) iter->l[0].b;
515 bool kick_reclaim = false;
517 BUG_ON(insert->u64s > ck->u64s);
519 if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
522 BUG_ON(jset_u64s(insert->u64s) > trans->journal_preres.u64s);
524 difference = jset_u64s(insert->u64s) - ck->res.u64s;
525 if (difference > 0) {
526 trans->journal_preres.u64s -= difference;
527 ck->res.u64s += difference;
531 bkey_copy(ck->k, insert);
534 if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
535 set_bit(BKEY_CACHED_DIRTY, &ck->flags);
536 atomic_long_inc(&c->btree_key_cache.nr_dirty);
538 if (bch2_nr_btree_keys_need_flush(c))
542 bch2_journal_pin_update(&c->journal, trans->journal_res.seq,
543 &ck->journal, bch2_btree_key_cache_journal_flush);
546 journal_reclaim_kick(&c->journal);
550 #ifdef CONFIG_BCACHEFS_DEBUG
551 void bch2_btree_key_cache_verify_clean(struct btree_trans *trans,
552 enum btree_id id, struct bpos pos)
554 BUG_ON(bch2_btree_key_cache_find(trans->c, id, pos));
558 static unsigned long bch2_btree_key_cache_scan(struct shrinker *shrink,
559 struct shrink_control *sc)
561 struct bch_fs *c = container_of(shrink, struct bch_fs,
562 btree_key_cache.shrink);
563 struct btree_key_cache *bc = &c->btree_key_cache;
564 struct bucket_table *tbl;
565 struct bkey_cached *ck, *t;
566 size_t scanned = 0, freed = 0, nr = sc->nr_to_scan;
567 unsigned start, flags;
570 /* Return -1 if we can't do anything right now */
571 if (sc->gfp_mask & __GFP_FS)
572 mutex_lock(&bc->lock);
573 else if (!mutex_trylock(&bc->lock))
576 srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
577 flags = memalloc_nofs_save();
580 * Newest freed entries are at the end of the list - once we hit one
581 * that's too new to be freed, we can bail out:
583 list_for_each_entry_safe(ck, t, &bc->freed, list) {
584 if (!poll_state_synchronize_srcu(&c->btree_trans_barrier,
585 ck->btree_trans_barrier_seq))
589 kmem_cache_free(bch2_key_cache, ck);
599 tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
600 if (bc->shrink_iter >= tbl->size)
602 start = bc->shrink_iter;
605 struct rhash_head *pos, *next;
607 pos = rht_ptr_rcu(rht_bucket(tbl, bc->shrink_iter));
609 while (!rht_is_a_nulls(pos)) {
610 next = rht_dereference_bucket_rcu(pos->next, tbl, bc->shrink_iter);
611 ck = container_of(pos, struct bkey_cached, hash);
613 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags))
616 if (test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
617 clear_bit(BKEY_CACHED_ACCESSED, &ck->flags);
618 else if (bkey_cached_lock_for_evict(ck)) {
619 bkey_cached_evict(bc, ck);
620 bkey_cached_free(bc, ck);
631 if (bc->shrink_iter >= tbl->size)
633 } while (scanned < nr && bc->shrink_iter != start);
637 memalloc_nofs_restore(flags);
638 srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
639 mutex_unlock(&bc->lock);
644 static unsigned long bch2_btree_key_cache_count(struct shrinker *shrink,
645 struct shrink_control *sc)
647 struct bch_fs *c = container_of(shrink, struct bch_fs,
648 btree_key_cache.shrink);
649 struct btree_key_cache *bc = &c->btree_key_cache;
650 long nr = atomic_long_read(&bc->nr_keys) -
651 atomic_long_read(&bc->nr_dirty);
656 void bch2_fs_btree_key_cache_exit(struct btree_key_cache *bc)
658 struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
659 struct bucket_table *tbl;
660 struct bkey_cached *ck, *n;
661 struct rhash_head *pos;
664 if (bc->shrink.list.next)
665 unregister_shrinker(&bc->shrink);
667 mutex_lock(&bc->lock);
670 tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
671 for (i = 0; i < tbl->size; i++)
672 rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
673 bkey_cached_evict(bc, ck);
674 list_add(&ck->list, &bc->freed);
678 list_for_each_entry_safe(ck, n, &bc->freed, list) {
681 bch2_journal_pin_drop(&c->journal, &ck->journal);
682 bch2_journal_preres_put(&c->journal, &ck->res);
686 kmem_cache_free(bch2_key_cache, ck);
689 BUG_ON(atomic_long_read(&bc->nr_dirty) &&
690 !bch2_journal_error(&c->journal) &&
691 test_bit(BCH_FS_WAS_RW, &c->flags));
692 BUG_ON(atomic_long_read(&bc->nr_keys));
694 mutex_unlock(&bc->lock);
696 if (bc->table_init_done)
697 rhashtable_destroy(&bc->table);
700 void bch2_fs_btree_key_cache_init_early(struct btree_key_cache *c)
702 mutex_init(&c->lock);
703 INIT_LIST_HEAD(&c->freed);
706 int bch2_fs_btree_key_cache_init(struct btree_key_cache *c)
710 ret = rhashtable_init(&c->table, &bch2_btree_key_cache_params);
714 c->table_init_done = true;
717 c->shrink.count_objects = bch2_btree_key_cache_count;
718 c->shrink.scan_objects = bch2_btree_key_cache_scan;
719 return register_shrinker(&c->shrink);
722 void bch2_btree_key_cache_to_text(struct printbuf *out, struct btree_key_cache *c)
724 pr_buf(out, "nr_freed:\t%zu\n", c->nr_freed);
725 pr_buf(out, "nr_keys:\t%zu\n", atomic_long_read(&c->nr_keys));
726 pr_buf(out, "nr_dirty:\t%zu\n", atomic_long_read(&c->nr_dirty));
729 void bch2_btree_key_cache_exit(void)
732 kmem_cache_destroy(bch2_key_cache);
735 int __init bch2_btree_key_cache_init(void)
737 bch2_key_cache = KMEM_CACHE(bkey_cached, 0);