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 bch_fs *c,
150 enum btree_id btree_id,
153 struct btree_key_cache *bc = &c->btree_key_cache;
154 struct bkey_cached *ck;
157 ck = bkey_cached_alloc(bc);
160 ck = bkey_cached_reuse(bc);
162 bch_err(c, "error allocating memory for key cache item, btree %s",
163 bch2_btree_ids[btree_id]);
164 return ERR_PTR(-ENOMEM);
169 if (btree_id == BTREE_ID_subvolumes)
170 six_lock_pcpu_alloc(&ck->c.lock);
172 six_lock_pcpu_free(&ck->c.lock);
176 ck->c.btree_id = btree_id;
177 ck->key.btree_id = btree_id;
180 ck->flags = 1U << BKEY_CACHED_ACCESSED;
182 if (unlikely(rhashtable_lookup_insert_fast(&bc->table,
184 bch2_btree_key_cache_params))) {
185 /* We raced with another fill: */
187 if (likely(was_new)) {
188 six_unlock_write(&ck->c.lock);
189 six_unlock_intent(&ck->c.lock);
192 mutex_lock(&bc->lock);
193 bkey_cached_free(bc, ck);
194 mutex_unlock(&bc->lock);
200 atomic_long_inc(&bc->nr_keys);
202 six_unlock_write(&ck->c.lock);
207 static int btree_key_cache_fill(struct btree_trans *trans,
208 struct btree_path *ck_path,
209 struct bkey_cached *ck)
211 struct btree_path *path;
213 unsigned new_u64s = 0;
214 struct bkey_i *new_k = NULL;
218 path = bch2_path_get(trans, ck->key.btree_id,
219 ck->key.pos, 0, 0, 0, _THIS_IP_);
220 ret = bch2_btree_path_traverse(trans, path, 0);
224 k = bch2_btree_path_peek_slot(path, &u);
226 if (!bch2_btree_node_relock(trans, ck_path, 0)) {
227 trace_trans_restart_relock_key_cache_fill(trans->fn,
228 _THIS_IP_, ck_path->btree_id, &ck_path->pos);
229 ret = btree_trans_restart(trans);
234 * bch2_varint_decode can read past the end of the buffer by at
235 * most 7 bytes (it won't be used):
237 new_u64s = k.k->u64s + 1;
240 * Allocate some extra space so that the transaction commit path is less
241 * likely to have to reallocate, since that requires a transaction
244 new_u64s = min(256U, (new_u64s * 3) / 2);
246 if (new_u64s > ck->u64s) {
247 new_u64s = roundup_pow_of_two(new_u64s);
248 new_k = kmalloc(new_u64s * sizeof(u64), GFP_NOFS);
250 bch_err(trans->c, "error allocating memory for key cache key, btree %s u64s %u",
251 bch2_btree_ids[ck->key.btree_id], new_u64s);
258 * XXX: not allowed to be holding read locks when we take a write lock,
261 bch2_btree_node_lock_write(trans, ck_path, ck_path->l[0].b);
268 bkey_reassemble(ck->k, k);
270 bch2_btree_node_unlock_write(trans, ck_path, ck_path->l[0].b);
272 /* We're not likely to need this iterator again: */
273 path->preserve = false;
275 bch2_path_put(trans, path, 0);
279 static int bkey_cached_check_fn(struct six_lock *lock, void *p)
281 struct bkey_cached *ck = container_of(lock, struct bkey_cached, c.lock);
282 const struct btree_path *path = p;
284 return ck->key.btree_id == path->btree_id &&
285 !bpos_cmp(ck->key.pos, path->pos) ? 0 : -1;
289 int bch2_btree_path_traverse_cached(struct btree_trans *trans, struct btree_path *path,
292 struct bch_fs *c = trans->c;
293 struct bkey_cached *ck;
300 if (bch2_btree_node_relock(trans, path, 0)) {
301 ck = (void *) path->l[0].b;
305 ck = bch2_btree_key_cache_find(c, path->btree_id, path->pos);
307 if (flags & BTREE_ITER_CACHED_NOCREATE) {
312 ck = btree_key_cache_create(c, path->btree_id, path->pos);
313 ret = PTR_ERR_OR_ZERO(ck);
319 mark_btree_node_locked(trans, path, 0, SIX_LOCK_intent);
320 path->locks_want = 1;
322 enum six_lock_type lock_want = __btree_lock_want(path, 0);
324 if (!btree_node_lock(trans, path, (void *) ck, path->pos, 0,
326 bkey_cached_check_fn, path, _THIS_IP_)) {
327 if (!trans->restarted)
334 if (ck->key.btree_id != path->btree_id ||
335 bpos_cmp(ck->key.pos, path->pos)) {
336 six_unlock_type(&ck->c.lock, lock_want);
340 mark_btree_node_locked(trans, path, 0, lock_want);
343 path->l[0].lock_seq = ck->c.lock.state.seq;
344 path->l[0].b = (void *) ck;
346 if (!ck->valid && !(flags & BTREE_ITER_CACHED_NOFILL)) {
347 if (!path->locks_want &&
348 !__bch2_btree_path_upgrade(trans, path, 1)) {
349 trace_transaction_restart_ip(trans->fn, _THIS_IP_);
350 ret = btree_trans_restart(trans);
354 ret = btree_key_cache_fill(trans, path, ck);
359 if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
360 set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
362 path->uptodate = BTREE_ITER_UPTODATE;
363 BUG_ON(btree_node_locked_type(path, 0) != btree_lock_want(path, 0));
368 btree_node_unlock(path, 0);
369 path->l[0].b = BTREE_ITER_NO_NODE_ERROR;
374 static int btree_key_cache_flush_pos(struct btree_trans *trans,
375 struct bkey_cached_key key,
377 unsigned commit_flags,
380 struct bch_fs *c = trans->c;
381 struct journal *j = &c->journal;
382 struct btree_iter c_iter, b_iter;
383 struct bkey_cached *ck = NULL;
386 bch2_trans_iter_init(trans, &b_iter, key.btree_id, key.pos,
389 BTREE_ITER_ALL_SNAPSHOTS);
390 bch2_trans_iter_init(trans, &c_iter, key.btree_id, key.pos,
392 BTREE_ITER_CACHED_NOFILL|
393 BTREE_ITER_CACHED_NOCREATE|
395 b_iter.flags &= ~BTREE_ITER_WITH_KEY_CACHE;
397 ret = bch2_btree_iter_traverse(&c_iter);
401 ck = (void *) c_iter.path->l[0].b;
405 if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
413 if (journal_seq && ck->journal.seq != journal_seq)
417 * Since journal reclaim depends on us making progress here, and the
418 * allocator/copygc depend on journal reclaim making progress, we need
419 * to be using alloc reserves:
421 ret = bch2_btree_iter_traverse(&b_iter) ?:
422 bch2_trans_update(trans, &b_iter, ck->k,
423 BTREE_UPDATE_KEY_CACHE_RECLAIM|
424 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
425 BTREE_TRIGGER_NORUN) ?:
426 bch2_trans_commit(trans, NULL, NULL,
427 BTREE_INSERT_NOCHECK_RW|
429 BTREE_INSERT_USE_RESERVE|
430 (ck->journal.seq == journal_last_seq(j)
431 ? JOURNAL_WATERMARK_reserved
435 bch2_fs_fatal_err_on(ret != -EINTR &&
437 !bch2_journal_error(j), c,
438 "error flushing key cache: %i", ret);
442 bch2_journal_pin_drop(j, &ck->journal);
443 bch2_journal_preres_put(j, &ck->res);
445 BUG_ON(!btree_node_locked(c_iter.path, 0));
448 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
449 clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
450 atomic_long_dec(&c->btree_key_cache.nr_dirty);
454 BUG_ON(!btree_node_intent_locked(c_iter.path, 0));
456 mark_btree_node_unlocked(c_iter.path, 0);
457 c_iter.path->l[0].b = NULL;
459 six_lock_write(&ck->c.lock, NULL, NULL);
461 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
462 clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
463 atomic_long_dec(&c->btree_key_cache.nr_dirty);
466 bkey_cached_evict(&c->btree_key_cache, ck);
468 mutex_lock(&c->btree_key_cache.lock);
469 bkey_cached_free(&c->btree_key_cache, ck);
470 mutex_unlock(&c->btree_key_cache.lock);
473 bch2_trans_iter_exit(trans, &b_iter);
474 bch2_trans_iter_exit(trans, &c_iter);
478 int bch2_btree_key_cache_journal_flush(struct journal *j,
479 struct journal_entry_pin *pin, u64 seq)
481 struct bch_fs *c = container_of(j, struct bch_fs, journal);
482 struct bkey_cached *ck =
483 container_of(pin, struct bkey_cached, journal);
484 struct bkey_cached_key key;
487 int srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
489 six_lock_read(&ck->c.lock, NULL, NULL);
492 if (ck->journal.seq != seq ||
493 !test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
494 six_unlock_read(&ck->c.lock);
497 six_unlock_read(&ck->c.lock);
499 ret = bch2_trans_do(c, NULL, NULL, 0,
500 btree_key_cache_flush_pos(&trans, key, seq,
501 BTREE_INSERT_JOURNAL_RECLAIM, false));
503 srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
509 * Flush and evict a key from the key cache:
511 int bch2_btree_key_cache_flush(struct btree_trans *trans,
512 enum btree_id id, struct bpos pos)
514 struct bch_fs *c = trans->c;
515 struct bkey_cached_key key = { id, pos };
517 /* Fastpath - assume it won't be found: */
518 if (!bch2_btree_key_cache_find(c, id, pos))
521 return btree_key_cache_flush_pos(trans, key, 0, 0, true);
524 bool bch2_btree_insert_key_cached(struct btree_trans *trans,
525 struct btree_path *path,
526 struct bkey_i *insert)
528 struct bch_fs *c = trans->c;
529 struct bkey_cached *ck = (void *) path->l[0].b;
530 bool kick_reclaim = false;
532 BUG_ON(insert->u64s > ck->u64s);
534 if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
537 BUG_ON(jset_u64s(insert->u64s) > trans->journal_preres.u64s);
539 difference = jset_u64s(insert->u64s) - ck->res.u64s;
540 if (difference > 0) {
541 trans->journal_preres.u64s -= difference;
542 ck->res.u64s += difference;
546 bkey_copy(ck->k, insert);
549 if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
550 set_bit(BKEY_CACHED_DIRTY, &ck->flags);
551 atomic_long_inc(&c->btree_key_cache.nr_dirty);
553 if (bch2_nr_btree_keys_need_flush(c))
557 bch2_journal_pin_update(&c->journal, trans->journal_res.seq,
558 &ck->journal, bch2_btree_key_cache_journal_flush);
561 journal_reclaim_kick(&c->journal);
565 void bch2_btree_key_cache_drop(struct btree_trans *trans,
566 struct btree_path *path)
568 struct bkey_cached *ck = (void *) path->l[0].b;
572 BUG_ON(test_bit(BKEY_CACHED_DIRTY, &ck->flags));
575 static unsigned long bch2_btree_key_cache_scan(struct shrinker *shrink,
576 struct shrink_control *sc)
578 struct bch_fs *c = container_of(shrink, struct bch_fs,
579 btree_key_cache.shrink);
580 struct btree_key_cache *bc = &c->btree_key_cache;
581 struct bucket_table *tbl;
582 struct bkey_cached *ck, *t;
583 size_t scanned = 0, freed = 0, nr = sc->nr_to_scan;
584 unsigned start, flags;
587 /* Return -1 if we can't do anything right now */
588 if (sc->gfp_mask & __GFP_FS)
589 mutex_lock(&bc->lock);
590 else if (!mutex_trylock(&bc->lock))
593 srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
594 flags = memalloc_nofs_save();
597 * Newest freed entries are at the end of the list - once we hit one
598 * that's too new to be freed, we can bail out:
600 list_for_each_entry_safe(ck, t, &bc->freed, list) {
601 if (!poll_state_synchronize_srcu(&c->btree_trans_barrier,
602 ck->btree_trans_barrier_seq))
606 kmem_cache_free(bch2_key_cache, ck);
616 tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
617 if (bc->shrink_iter >= tbl->size)
619 start = bc->shrink_iter;
622 struct rhash_head *pos, *next;
624 pos = rht_ptr_rcu(rht_bucket(tbl, bc->shrink_iter));
626 while (!rht_is_a_nulls(pos)) {
627 next = rht_dereference_bucket_rcu(pos->next, tbl, bc->shrink_iter);
628 ck = container_of(pos, struct bkey_cached, hash);
630 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags))
633 if (test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
634 clear_bit(BKEY_CACHED_ACCESSED, &ck->flags);
635 else if (bkey_cached_lock_for_evict(ck)) {
636 bkey_cached_evict(bc, ck);
637 bkey_cached_free(bc, ck);
648 if (bc->shrink_iter >= tbl->size)
650 } while (scanned < nr && bc->shrink_iter != start);
654 memalloc_nofs_restore(flags);
655 srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
656 mutex_unlock(&bc->lock);
661 static unsigned long bch2_btree_key_cache_count(struct shrinker *shrink,
662 struct shrink_control *sc)
664 struct bch_fs *c = container_of(shrink, struct bch_fs,
665 btree_key_cache.shrink);
666 struct btree_key_cache *bc = &c->btree_key_cache;
667 long nr = atomic_long_read(&bc->nr_keys) -
668 atomic_long_read(&bc->nr_dirty);
673 void bch2_fs_btree_key_cache_exit(struct btree_key_cache *bc)
675 struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
676 struct bucket_table *tbl;
677 struct bkey_cached *ck, *n;
678 struct rhash_head *pos;
681 if (bc->shrink.list.next)
682 unregister_shrinker(&bc->shrink);
684 mutex_lock(&bc->lock);
687 tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
689 for (i = 0; i < tbl->size; i++)
690 rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
691 bkey_cached_evict(bc, ck);
692 list_add(&ck->list, &bc->freed);
696 list_for_each_entry_safe(ck, n, &bc->freed, list) {
699 bch2_journal_pin_drop(&c->journal, &ck->journal);
700 bch2_journal_preres_put(&c->journal, &ck->res);
704 kmem_cache_free(bch2_key_cache, ck);
707 BUG_ON(atomic_long_read(&bc->nr_dirty) &&
708 !bch2_journal_error(&c->journal) &&
709 test_bit(BCH_FS_WAS_RW, &c->flags));
710 BUG_ON(atomic_long_read(&bc->nr_keys));
712 mutex_unlock(&bc->lock);
714 if (bc->table_init_done)
715 rhashtable_destroy(&bc->table);
718 void bch2_fs_btree_key_cache_init_early(struct btree_key_cache *c)
720 mutex_init(&c->lock);
721 INIT_LIST_HEAD(&c->freed);
724 static void bch2_btree_key_cache_shrinker_to_text(struct printbuf *out, struct shrinker *shrink)
726 struct btree_key_cache *bc =
727 container_of(shrink, struct btree_key_cache, shrink);
729 bch2_btree_key_cache_to_text(out, bc);
732 int bch2_fs_btree_key_cache_init(struct btree_key_cache *c)
736 ret = rhashtable_init(&c->table, &bch2_btree_key_cache_params);
740 c->table_init_done = true;
743 c->shrink.count_objects = bch2_btree_key_cache_count;
744 c->shrink.scan_objects = bch2_btree_key_cache_scan;
745 c->shrink.to_text = bch2_btree_key_cache_shrinker_to_text;
746 return register_shrinker(&c->shrink);
749 void bch2_btree_key_cache_to_text(struct printbuf *out, struct btree_key_cache *c)
751 prt_printf(out, "nr_freed:\t%zu\n", c->nr_freed);
752 prt_printf(out, "nr_keys:\t%lu\n", atomic_long_read(&c->nr_keys));
753 prt_printf(out, "nr_dirty:\t%lu\n", atomic_long_read(&c->nr_dirty));
756 void bch2_btree_key_cache_exit(void)
759 kmem_cache_destroy(bch2_key_cache);
762 int __init bch2_btree_key_cache_init(void)
764 bch2_key_cache = KMEM_CACHE(bkey_cached, 0);