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);
173 return PTR_ERR(iter);
175 k = bch2_btree_iter_peek_slot(iter);
178 bch2_trans_iter_put(trans, iter);
182 if (!bch2_btree_node_relock(ck_iter, 0)) {
183 bch2_trans_iter_put(trans, iter);
184 trace_transaction_restart_ip(trans->ip, _THIS_IP_);
188 if (k.k->u64s > ck->u64s) {
189 new_u64s = roundup_pow_of_two(k.k->u64s);
190 new_k = kmalloc(new_u64s * sizeof(u64), GFP_NOFS);
192 bch2_trans_iter_put(trans, iter);
197 bch2_btree_node_lock_write(ck_iter->l[0].b, ck_iter);
204 bkey_reassemble(ck->k, k);
206 bch2_btree_node_unlock_write(ck_iter->l[0].b, ck_iter);
208 /* We're not likely to need this iterator again: */
209 bch2_trans_iter_free(trans, iter);
214 static int bkey_cached_check_fn(struct six_lock *lock, void *p)
216 struct bkey_cached *ck = container_of(lock, struct bkey_cached, c.lock);
217 const struct btree_iter *iter = p;
219 return ck->key.btree_id == iter->btree_id &&
220 !bkey_cmp(ck->key.pos, iter->pos) ? 0 : -1;
224 int bch2_btree_iter_traverse_cached(struct btree_iter *iter)
226 struct btree_trans *trans = iter->trans;
227 struct bch_fs *c = trans->c;
228 struct bkey_cached *ck;
233 if (btree_node_locked(iter, 0)) {
234 ck = (void *) iter->l[0].b;
238 ck = bch2_btree_key_cache_find(c, iter->btree_id, iter->pos);
240 if (iter->flags & BTREE_ITER_CACHED_NOCREATE) {
245 mutex_lock(&c->btree_key_cache.lock);
246 ck = btree_key_cache_create(&c->btree_key_cache,
247 iter->btree_id, iter->pos);
248 mutex_unlock(&c->btree_key_cache.lock);
250 ret = PTR_ERR_OR_ZERO(ck);
256 mark_btree_node_locked(iter, 0, SIX_LOCK_intent);
257 iter->locks_want = 1;
259 enum six_lock_type lock_want = __btree_lock_want(iter, 0);
261 if (!btree_node_lock((void *) ck, iter->pos, 0, iter, lock_want,
262 bkey_cached_check_fn, iter, _THIS_IP_)) {
263 if (ck->key.btree_id != iter->btree_id ||
264 bkey_cmp(ck->key.pos, iter->pos)) {
268 trace_transaction_restart_ip(trans->ip, _THIS_IP_);
273 if (ck->key.btree_id != iter->btree_id ||
274 bkey_cmp(ck->key.pos, iter->pos)) {
275 six_unlock_type(&ck->c.lock, lock_want);
279 mark_btree_node_locked(iter, 0, lock_want);
282 iter->l[0].lock_seq = ck->c.lock.state.seq;
283 iter->l[0].b = (void *) ck;
285 if (!ck->valid && !(iter->flags & BTREE_ITER_CACHED_NOFILL)) {
286 if (!btree_node_intent_locked(iter, 0))
287 bch2_btree_iter_upgrade(iter, 1);
288 if (!btree_node_intent_locked(iter, 0)) {
289 trace_transaction_restart_ip(trans->ip, _THIS_IP_);
294 ret = btree_key_cache_fill(trans, iter, ck);
299 if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
300 set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
302 iter->uptodate = BTREE_ITER_NEED_PEEK;
303 bch2_btree_iter_downgrade(iter);
307 btree_node_unlock(iter, 0);
308 iter->flags |= BTREE_ITER_ERROR;
309 iter->l[0].b = BTREE_ITER_NO_NODE_ERROR;
314 static int btree_key_cache_flush_pos(struct btree_trans *trans,
315 struct bkey_cached_key key,
319 struct bch_fs *c = trans->c;
320 struct journal *j = &c->journal;
321 struct btree_iter *c_iter = NULL, *b_iter = NULL;
322 struct bkey_cached *ck;
325 b_iter = bch2_trans_get_iter(trans, key.btree_id, key.pos,
328 ret = PTR_ERR_OR_ZERO(b_iter);
332 c_iter = bch2_trans_get_iter(trans, key.btree_id, key.pos,
334 BTREE_ITER_CACHED_NOFILL|
335 BTREE_ITER_CACHED_NOCREATE|
337 ret = PTR_ERR_OR_ZERO(c_iter);
341 ret = bch2_btree_iter_traverse(c_iter);
345 ck = (void *) c_iter->l[0].b;
347 (journal_seq && ck->journal.seq != journal_seq))
350 if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
356 ret = bch2_btree_iter_traverse(b_iter) ?:
357 bch2_trans_update(trans, b_iter, ck->k, BTREE_TRIGGER_NORUN) ?:
358 bch2_trans_commit(trans, NULL, NULL,
359 BTREE_INSERT_NOUNLOCK|
360 BTREE_INSERT_NOCHECK_RW|
362 BTREE_INSERT_USE_RESERVE|
363 BTREE_INSERT_USE_ALLOC_RESERVE|
364 BTREE_INSERT_JOURNAL_RESERVED|
365 BTREE_INSERT_JOURNAL_RECLAIM);
370 BUG_ON(ret && !bch2_journal_error(j));
375 bch2_journal_pin_drop(j, &ck->journal);
376 bch2_journal_preres_put(j, &ck->res);
379 mutex_lock(&c->btree_key_cache.lock);
380 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
381 clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
382 c->btree_key_cache.nr_dirty--;
385 list_move_tail(&ck->list, &c->btree_key_cache.clean);
386 mutex_unlock(&c->btree_key_cache.lock);
389 BUG_ON(!btree_node_intent_locked(c_iter, 0));
391 mark_btree_node_unlocked(c_iter, 0);
392 c_iter->l[0].b = NULL;
394 six_lock_write(&ck->c.lock, NULL, NULL);
396 mutex_lock(&c->btree_key_cache.lock);
397 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
398 clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
399 c->btree_key_cache.nr_dirty--;
402 bkey_cached_evict(&c->btree_key_cache, ck);
403 bkey_cached_free(&c->btree_key_cache, ck);
404 mutex_unlock(&c->btree_key_cache.lock);
407 bch2_trans_iter_put(trans, b_iter);
408 bch2_trans_iter_put(trans, c_iter);
412 static void btree_key_cache_journal_flush(struct journal *j,
413 struct journal_entry_pin *pin,
416 struct bch_fs *c = container_of(j, struct bch_fs, journal);
417 struct bkey_cached *ck =
418 container_of(pin, struct bkey_cached, journal);
419 struct bkey_cached_key key;
420 struct btree_trans trans;
422 int srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
424 six_lock_read(&ck->c.lock, NULL, NULL);
427 if (ck->journal.seq != seq ||
428 !test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
429 six_unlock_read(&ck->c.lock);
432 six_unlock_read(&ck->c.lock);
434 bch2_trans_init(&trans, c, 0, 0);
435 btree_key_cache_flush_pos(&trans, key, seq, false);
436 bch2_trans_exit(&trans);
438 srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
442 * Flush and evict a key from the key cache:
444 int bch2_btree_key_cache_flush(struct btree_trans *trans,
445 enum btree_id id, struct bpos pos)
447 struct bch_fs *c = trans->c;
448 struct bkey_cached_key key = { id, pos };
450 /* Fastpath - assume it won't be found: */
451 if (!bch2_btree_key_cache_find(c, id, pos))
454 return btree_key_cache_flush_pos(trans, key, 0, true);
457 bool bch2_btree_insert_key_cached(struct btree_trans *trans,
458 struct btree_iter *iter,
459 struct bkey_i *insert)
461 struct bch_fs *c = trans->c;
462 struct bkey_cached *ck = (void *) iter->l[0].b;
463 bool kick_reclaim = false;
465 BUG_ON(insert->u64s > ck->u64s);
467 if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
470 BUG_ON(jset_u64s(insert->u64s) > trans->journal_preres.u64s);
472 difference = jset_u64s(insert->u64s) - ck->res.u64s;
473 if (difference > 0) {
474 trans->journal_preres.u64s -= difference;
475 ck->res.u64s += difference;
479 bkey_copy(ck->k, insert);
482 if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
483 mutex_lock(&c->btree_key_cache.lock);
484 list_move(&ck->list, &c->btree_key_cache.dirty);
486 set_bit(BKEY_CACHED_DIRTY, &ck->flags);
487 c->btree_key_cache.nr_dirty++;
489 if (bch2_nr_btree_keys_need_flush(c))
492 mutex_unlock(&c->btree_key_cache.lock);
495 bch2_journal_pin_update(&c->journal, trans->journal_res.seq,
496 &ck->journal, btree_key_cache_journal_flush);
499 journal_reclaim_kick(&c->journal);
503 #ifdef CONFIG_BCACHEFS_DEBUG
504 void bch2_btree_key_cache_verify_clean(struct btree_trans *trans,
505 enum btree_id id, struct bpos pos)
507 BUG_ON(bch2_btree_key_cache_find(trans->c, id, pos));
511 static unsigned long bch2_btree_key_cache_scan(struct shrinker *shrink,
512 struct shrink_control *sc)
514 struct bch_fs *c = container_of(shrink, struct bch_fs,
515 btree_key_cache.shrink);
516 struct btree_key_cache *bc = &c->btree_key_cache;
517 struct bkey_cached *ck, *t;
518 size_t scanned = 0, freed = 0, nr = sc->nr_to_scan;
521 /* Return -1 if we can't do anything right now */
522 if (sc->gfp_mask & __GFP_FS)
523 mutex_lock(&bc->lock);
524 else if (!mutex_trylock(&bc->lock))
527 flags = memalloc_nofs_save();
530 * Newest freed entries are at the end of the list - once we hit one
531 * that's too new to be freed, we can bail out:
533 list_for_each_entry_safe(ck, t, &bc->freed, list) {
534 if (!poll_state_synchronize_srcu(&c->btree_trans_barrier,
535 ck->btree_trans_barrier_seq))
539 kmem_cache_free(bch2_key_cache, ck);
548 list_for_each_entry_safe(ck, t, &bc->clean, list) {
549 if (test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
550 clear_bit(BKEY_CACHED_ACCESSED, &ck->flags);
551 else if (bkey_cached_lock_for_evict(ck)) {
552 bkey_cached_evict(bc, ck);
553 bkey_cached_free(bc, ck);
558 if (&t->list != &bc->clean)
559 list_move_tail(&bc->clean, &t->list);
564 memalloc_nofs_restore(flags);
565 mutex_unlock(&bc->lock);
570 static unsigned long bch2_btree_key_cache_count(struct shrinker *shrink,
571 struct shrink_control *sc)
573 struct bch_fs *c = container_of(shrink, struct bch_fs,
574 btree_key_cache.shrink);
575 struct btree_key_cache *bc = &c->btree_key_cache;
580 void bch2_fs_btree_key_cache_exit(struct btree_key_cache *bc)
582 struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
583 struct bkey_cached *ck, *n;
585 if (bc->shrink.list.next)
586 unregister_shrinker(&bc->shrink);
588 mutex_lock(&bc->lock);
589 list_splice(&bc->dirty, &bc->clean);
591 list_for_each_entry_safe(ck, n, &bc->clean, list) {
592 bch2_journal_pin_drop(&c->journal, &ck->journal);
593 bch2_journal_preres_put(&c->journal, &ck->res);
597 kmem_cache_free(bch2_key_cache, ck);
601 BUG_ON(bc->nr_dirty && !bch2_journal_error(&c->journal));
604 list_for_each_entry_safe(ck, n, &bc->freed, list) {
606 kmem_cache_free(bch2_key_cache, ck);
608 mutex_unlock(&bc->lock);
610 if (bc->table_init_done)
611 rhashtable_destroy(&bc->table);
614 void bch2_fs_btree_key_cache_init_early(struct btree_key_cache *c)
616 mutex_init(&c->lock);
617 INIT_LIST_HEAD(&c->freed);
618 INIT_LIST_HEAD(&c->clean);
619 INIT_LIST_HEAD(&c->dirty);
622 int bch2_fs_btree_key_cache_init(struct btree_key_cache *c)
627 c->shrink.count_objects = bch2_btree_key_cache_count;
628 c->shrink.scan_objects = bch2_btree_key_cache_scan;
630 ret = register_shrinker(&c->shrink);
634 ret = rhashtable_init(&c->table, &bch2_btree_key_cache_params);
638 c->table_init_done = true;
642 void bch2_btree_key_cache_to_text(struct printbuf *out, struct btree_key_cache *c)
644 pr_buf(out, "nr_freed:\t%zu\n", c->nr_freed);
645 pr_buf(out, "nr_keys:\t%zu\n", c->nr_keys);
646 pr_buf(out, "nr_dirty:\t%zu\n", c->nr_dirty);
649 void bch2_btree_key_cache_exit(void)
652 kmem_cache_destroy(bch2_key_cache);
655 int __init bch2_btree_key_cache_init(void)
657 bch2_key_cache = KMEM_CACHE(bkey_cached, 0);