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 bkey_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));
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 list_for_each_entry_reverse(ck, &c->freed, list)
103 if (bkey_cached_lock_for_evict(ck)) {
108 ck = kmem_cache_alloc(bch2_key_cache, GFP_NOFS|__GFP_ZERO);
110 INIT_LIST_HEAD(&ck->list);
111 six_lock_init(&ck->c.lock);
112 BUG_ON(!six_trylock_intent(&ck->c.lock));
113 BUG_ON(!six_trylock_write(&ck->c.lock));
117 list_for_each_entry(ck, &c->clean, list)
118 if (bkey_cached_lock_for_evict(ck)) {
119 bkey_cached_evict(c, ck);
126 static struct bkey_cached *
127 btree_key_cache_create(struct btree_key_cache *c,
128 enum btree_id btree_id,
131 struct bkey_cached *ck;
133 ck = bkey_cached_alloc(c);
135 return ERR_PTR(-ENOMEM);
138 ck->c.btree_id = btree_id;
139 ck->key.btree_id = btree_id;
142 ck->flags = 1U << BKEY_CACHED_ACCESSED;
144 if (rhashtable_lookup_insert_fast(&c->table,
146 bch2_btree_key_cache_params)) {
147 /* We raced with another fill: */
148 bkey_cached_free(c, ck);
154 list_move(&ck->list, &c->clean);
155 six_unlock_write(&ck->c.lock);
160 static int btree_key_cache_fill(struct btree_trans *trans,
161 struct btree_iter *ck_iter,
162 struct bkey_cached *ck)
164 struct btree_iter *iter;
166 unsigned new_u64s = 0;
167 struct bkey_i *new_k = NULL;
170 iter = bch2_trans_get_iter(trans, ck->key.btree_id,
171 ck->key.pos, BTREE_ITER_SLOTS);
172 k = bch2_btree_iter_peek_slot(iter);
177 if (!bch2_btree_node_relock(ck_iter, 0)) {
178 trace_transaction_restart_ip(trans->ip, _THIS_IP_);
183 if (k.k->u64s > ck->u64s) {
184 new_u64s = roundup_pow_of_two(k.k->u64s);
185 new_k = kmalloc(new_u64s * sizeof(u64), GFP_NOFS);
192 bch2_btree_node_lock_write(ck_iter->l[0].b, ck_iter);
199 bkey_reassemble(ck->k, k);
201 bch2_btree_node_unlock_write(ck_iter->l[0].b, ck_iter);
203 /* We're not likely to need this iterator again: */
204 set_btree_iter_dontneed(trans, iter);
206 bch2_trans_iter_put(trans, iter);
210 static int bkey_cached_check_fn(struct six_lock *lock, void *p)
212 struct bkey_cached *ck = container_of(lock, struct bkey_cached, c.lock);
213 const struct btree_iter *iter = p;
215 return ck->key.btree_id == iter->btree_id &&
216 !bkey_cmp(ck->key.pos, iter->pos) ? 0 : -1;
220 int bch2_btree_iter_traverse_cached(struct btree_iter *iter)
222 struct btree_trans *trans = iter->trans;
223 struct bch_fs *c = trans->c;
224 struct bkey_cached *ck;
229 if (btree_node_locked(iter, 0)) {
230 ck = (void *) iter->l[0].b;
234 ck = bch2_btree_key_cache_find(c, iter->btree_id, iter->pos);
236 if (iter->flags & BTREE_ITER_CACHED_NOCREATE) {
241 mutex_lock(&c->btree_key_cache.lock);
242 ck = btree_key_cache_create(&c->btree_key_cache,
243 iter->btree_id, iter->pos);
244 mutex_unlock(&c->btree_key_cache.lock);
246 ret = PTR_ERR_OR_ZERO(ck);
252 mark_btree_node_locked(iter, 0, SIX_LOCK_intent);
253 iter->locks_want = 1;
255 enum six_lock_type lock_want = __btree_lock_want(iter, 0);
257 if (!btree_node_lock((void *) ck, iter->pos, 0, iter, lock_want,
258 bkey_cached_check_fn, iter, _THIS_IP_)) {
259 if (ck->key.btree_id != iter->btree_id ||
260 bkey_cmp(ck->key.pos, iter->pos)) {
264 trace_transaction_restart_ip(trans->ip, _THIS_IP_);
269 if (ck->key.btree_id != iter->btree_id ||
270 bkey_cmp(ck->key.pos, iter->pos)) {
271 six_unlock_type(&ck->c.lock, lock_want);
275 mark_btree_node_locked(iter, 0, lock_want);
278 iter->l[0].lock_seq = ck->c.lock.state.seq;
279 iter->l[0].b = (void *) ck;
281 if (!ck->valid && !(iter->flags & BTREE_ITER_CACHED_NOFILL)) {
282 if (!btree_node_intent_locked(iter, 0))
283 bch2_btree_iter_upgrade(iter, 1);
284 if (!btree_node_intent_locked(iter, 0)) {
285 trace_transaction_restart_ip(trans->ip, _THIS_IP_);
290 ret = btree_key_cache_fill(trans, iter, ck);
295 if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
296 set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
298 iter->uptodate = BTREE_ITER_NEED_PEEK;
300 if (!(iter->flags & BTREE_ITER_INTENT))
301 bch2_btree_iter_downgrade(iter);
302 else if (!iter->locks_want) {
303 if (!__bch2_btree_iter_upgrade(iter, 1))
310 btree_node_unlock(iter, 0);
311 iter->flags |= BTREE_ITER_ERROR;
312 iter->l[0].b = BTREE_ITER_NO_NODE_ERROR;
317 static int btree_key_cache_flush_pos(struct btree_trans *trans,
318 struct bkey_cached_key key,
322 struct bch_fs *c = trans->c;
323 struct journal *j = &c->journal;
324 struct btree_iter *c_iter = NULL, *b_iter = NULL;
325 struct bkey_cached *ck = NULL;
328 b_iter = bch2_trans_get_iter(trans, key.btree_id, key.pos,
331 c_iter = bch2_trans_get_iter(trans, key.btree_id, key.pos,
333 BTREE_ITER_CACHED_NOFILL|
334 BTREE_ITER_CACHED_NOCREATE|
337 ret = bch2_btree_iter_traverse(c_iter);
341 ck = (void *) c_iter->l[0].b;
343 (journal_seq && ck->journal.seq != journal_seq))
346 if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
352 ret = bch2_btree_iter_traverse(b_iter) ?:
353 bch2_trans_update(trans, b_iter, ck->k, BTREE_TRIGGER_NORUN) ?:
354 bch2_trans_commit(trans, NULL, NULL,
355 BTREE_INSERT_NOUNLOCK|
356 BTREE_INSERT_NOCHECK_RW|
358 BTREE_INSERT_JOURNAL_RESERVED|
359 BTREE_INSERT_JOURNAL_RECLAIM);
365 bch2_fs_fatal_err_on(!bch2_journal_error(j), c,
366 "error flushing key cache: %i", ret);
370 bch2_journal_pin_drop(j, &ck->journal);
371 bch2_journal_preres_put(j, &ck->res);
374 mutex_lock(&c->btree_key_cache.lock);
375 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
376 clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
377 c->btree_key_cache.nr_dirty--;
380 list_move_tail(&ck->list, &c->btree_key_cache.clean);
381 mutex_unlock(&c->btree_key_cache.lock);
384 BUG_ON(!btree_node_intent_locked(c_iter, 0));
386 mark_btree_node_unlocked(c_iter, 0);
387 c_iter->l[0].b = NULL;
389 six_lock_write(&ck->c.lock, NULL, NULL);
391 mutex_lock(&c->btree_key_cache.lock);
392 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
393 clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
394 c->btree_key_cache.nr_dirty--;
397 bkey_cached_evict(&c->btree_key_cache, ck);
398 bkey_cached_free(&c->btree_key_cache, ck);
399 mutex_unlock(&c->btree_key_cache.lock);
402 bch2_trans_iter_put(trans, b_iter);
403 bch2_trans_iter_put(trans, c_iter);
407 static void btree_key_cache_journal_flush(struct journal *j,
408 struct journal_entry_pin *pin,
411 struct bch_fs *c = container_of(j, struct bch_fs, journal);
412 struct bkey_cached *ck =
413 container_of(pin, struct bkey_cached, journal);
414 struct bkey_cached_key key;
415 struct btree_trans trans;
417 int srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
419 six_lock_read(&ck->c.lock, NULL, NULL);
422 if (ck->journal.seq != seq ||
423 !test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
424 six_unlock_read(&ck->c.lock);
427 six_unlock_read(&ck->c.lock);
429 bch2_trans_init(&trans, c, 0, 0);
430 btree_key_cache_flush_pos(&trans, key, seq, false);
431 bch2_trans_exit(&trans);
433 srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
437 * Flush and evict a key from the key cache:
439 int bch2_btree_key_cache_flush(struct btree_trans *trans,
440 enum btree_id id, struct bpos pos)
442 struct bch_fs *c = trans->c;
443 struct bkey_cached_key key = { id, pos };
445 /* Fastpath - assume it won't be found: */
446 if (!bch2_btree_key_cache_find(c, id, pos))
449 return btree_key_cache_flush_pos(trans, key, 0, true);
452 bool bch2_btree_insert_key_cached(struct btree_trans *trans,
453 struct btree_iter *iter,
454 struct bkey_i *insert)
456 struct bch_fs *c = trans->c;
457 struct bkey_cached *ck = (void *) iter->l[0].b;
458 bool kick_reclaim = false;
460 BUG_ON(insert->u64s > ck->u64s);
462 if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
465 BUG_ON(jset_u64s(insert->u64s) > trans->journal_preres.u64s);
467 difference = jset_u64s(insert->u64s) - ck->res.u64s;
468 if (difference > 0) {
469 trans->journal_preres.u64s -= difference;
470 ck->res.u64s += difference;
474 bkey_copy(ck->k, insert);
477 if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
478 mutex_lock(&c->btree_key_cache.lock);
479 list_move(&ck->list, &c->btree_key_cache.dirty);
481 set_bit(BKEY_CACHED_DIRTY, &ck->flags);
482 c->btree_key_cache.nr_dirty++;
484 if (bch2_nr_btree_keys_need_flush(c))
487 mutex_unlock(&c->btree_key_cache.lock);
490 bch2_journal_pin_update(&c->journal, trans->journal_res.seq,
491 &ck->journal, btree_key_cache_journal_flush);
494 journal_reclaim_kick(&c->journal);
498 #ifdef CONFIG_BCACHEFS_DEBUG
499 void bch2_btree_key_cache_verify_clean(struct btree_trans *trans,
500 enum btree_id id, struct bpos pos)
502 BUG_ON(bch2_btree_key_cache_find(trans->c, id, pos));
506 static unsigned long bch2_btree_key_cache_scan(struct shrinker *shrink,
507 struct shrink_control *sc)
509 struct bch_fs *c = container_of(shrink, struct bch_fs,
510 btree_key_cache.shrink);
511 struct btree_key_cache *bc = &c->btree_key_cache;
512 struct bkey_cached *ck, *t;
513 size_t scanned = 0, freed = 0, nr = sc->nr_to_scan;
516 /* Return -1 if we can't do anything right now */
517 if (sc->gfp_mask & __GFP_FS)
518 mutex_lock(&bc->lock);
519 else if (!mutex_trylock(&bc->lock))
522 flags = memalloc_nofs_save();
525 * Newest freed entries are at the end of the list - once we hit one
526 * that's too new to be freed, we can bail out:
528 list_for_each_entry_safe(ck, t, &bc->freed, list) {
529 if (!poll_state_synchronize_srcu(&c->btree_trans_barrier,
530 ck->btree_trans_barrier_seq))
534 kmem_cache_free(bch2_key_cache, ck);
543 list_for_each_entry_safe(ck, t, &bc->clean, list) {
544 if (test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
545 clear_bit(BKEY_CACHED_ACCESSED, &ck->flags);
546 else if (bkey_cached_lock_for_evict(ck)) {
547 bkey_cached_evict(bc, ck);
548 bkey_cached_free(bc, ck);
553 if (&t->list != &bc->clean)
554 list_move_tail(&bc->clean, &t->list);
559 memalloc_nofs_restore(flags);
560 mutex_unlock(&bc->lock);
565 static unsigned long bch2_btree_key_cache_count(struct shrinker *shrink,
566 struct shrink_control *sc)
568 struct bch_fs *c = container_of(shrink, struct bch_fs,
569 btree_key_cache.shrink);
570 struct btree_key_cache *bc = &c->btree_key_cache;
575 void bch2_fs_btree_key_cache_exit(struct btree_key_cache *bc)
577 struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
578 struct bkey_cached *ck, *n;
580 if (bc->shrink.list.next)
581 unregister_shrinker(&bc->shrink);
583 mutex_lock(&bc->lock);
584 list_splice(&bc->dirty, &bc->clean);
586 list_for_each_entry_safe(ck, n, &bc->clean, list) {
589 bch2_journal_pin_drop(&c->journal, &ck->journal);
590 bch2_journal_preres_put(&c->journal, &ck->res);
594 kmem_cache_free(bch2_key_cache, ck);
598 BUG_ON(bc->nr_dirty && !bch2_journal_error(&c->journal));
601 list_for_each_entry_safe(ck, n, &bc->freed, list) {
605 kmem_cache_free(bch2_key_cache, ck);
607 mutex_unlock(&bc->lock);
609 if (bc->table_init_done)
610 rhashtable_destroy(&bc->table);
613 void bch2_fs_btree_key_cache_init_early(struct btree_key_cache *c)
615 mutex_init(&c->lock);
616 INIT_LIST_HEAD(&c->freed);
617 INIT_LIST_HEAD(&c->clean);
618 INIT_LIST_HEAD(&c->dirty);
621 int bch2_fs_btree_key_cache_init(struct btree_key_cache *c)
626 c->shrink.count_objects = bch2_btree_key_cache_count;
627 c->shrink.scan_objects = bch2_btree_key_cache_scan;
629 ret = register_shrinker(&c->shrink);
633 ret = rhashtable_init(&c->table, &bch2_btree_key_cache_params);
637 c->table_init_done = true;
641 void bch2_btree_key_cache_to_text(struct printbuf *out, struct btree_key_cache *c)
643 pr_buf(out, "nr_freed:\t%zu\n", c->nr_freed);
644 pr_buf(out, "nr_keys:\t%zu\n", c->nr_keys);
645 pr_buf(out, "nr_dirty:\t%zu\n", c->nr_dirty);
648 void bch2_btree_key_cache_exit(void)
651 kmem_cache_destroy(bch2_key_cache);
654 int __init bch2_btree_key_cache_init(void)
656 bch2_key_cache = KMEM_CACHE(bkey_cached, 0);