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);
166 if (btree_id == BTREE_ID_subvolumes)
167 six_lock_pcpu_alloc(&ck->c.lock);
169 six_lock_pcpu_free(&ck->c.lock);
172 ck->c.btree_id = btree_id;
173 ck->key.btree_id = btree_id;
176 ck->flags = 1U << BKEY_CACHED_ACCESSED;
178 if (unlikely(rhashtable_lookup_insert_fast(&c->table,
180 bch2_btree_key_cache_params))) {
181 /* We raced with another fill: */
183 if (likely(was_new)) {
184 six_unlock_write(&ck->c.lock);
185 six_unlock_intent(&ck->c.lock);
188 mutex_lock(&c->lock);
189 bkey_cached_free(c, ck);
190 mutex_unlock(&c->lock);
196 atomic_long_inc(&c->nr_keys);
198 six_unlock_write(&ck->c.lock);
203 static int btree_key_cache_fill(struct btree_trans *trans,
204 struct btree_path *ck_path,
205 struct bkey_cached *ck)
207 struct btree_iter iter;
209 unsigned new_u64s = 0;
210 struct bkey_i *new_k = NULL;
213 bch2_trans_iter_init(trans, &iter, ck->key.btree_id,
214 ck->key.pos, BTREE_ITER_SLOTS);
215 k = bch2_btree_iter_peek_slot(&iter);
220 if (!bch2_btree_node_relock(trans, ck_path, 0)) {
221 trace_transaction_restart_ip(trans->ip, _THIS_IP_);
222 ret = btree_trans_restart(trans);
227 * bch2_varint_decode can read past the end of the buffer by at
228 * most 7 bytes (it won't be used):
230 new_u64s = k.k->u64s + 1;
232 if (new_u64s > ck->u64s) {
233 new_u64s = roundup_pow_of_two(new_u64s);
234 new_k = kmalloc(new_u64s * sizeof(u64), GFP_NOFS);
242 * XXX: not allowed to be holding read locks when we take a write lock,
245 bch2_btree_node_lock_write(trans, ck_path, ck_path->l[0].b);
252 bkey_reassemble(ck->k, k);
254 bch2_btree_node_unlock_write(trans, ck_path, ck_path->l[0].b);
256 /* We're not likely to need this iterator again: */
257 set_btree_iter_dontneed(&iter);
259 bch2_trans_iter_exit(trans, &iter);
263 static int bkey_cached_check_fn(struct six_lock *lock, void *p)
265 struct bkey_cached *ck = container_of(lock, struct bkey_cached, c.lock);
266 const struct btree_path *path = p;
268 return ck->key.btree_id == path->btree_id &&
269 !bpos_cmp(ck->key.pos, path->pos) ? 0 : -1;
273 int bch2_btree_path_traverse_cached(struct btree_trans *trans, struct btree_path *path,
276 struct bch_fs *c = trans->c;
277 struct bkey_cached *ck;
284 if (bch2_btree_node_relock(trans, path, 0)) {
285 ck = (void *) path->l[0].b;
289 ck = bch2_btree_key_cache_find(c, path->btree_id, path->pos);
291 if (flags & BTREE_ITER_CACHED_NOCREATE) {
296 ck = btree_key_cache_create(&c->btree_key_cache,
297 path->btree_id, path->pos);
298 ret = PTR_ERR_OR_ZERO(ck);
304 mark_btree_node_locked(path, 0, SIX_LOCK_intent);
305 path->locks_want = 1;
307 enum six_lock_type lock_want = __btree_lock_want(path, 0);
309 if (!btree_node_lock(trans, path, (void *) ck, path->pos, 0,
311 bkey_cached_check_fn, path, _THIS_IP_)) {
312 if (!trans->restarted)
315 trace_transaction_restart_ip(trans->ip, _THIS_IP_);
320 if (ck->key.btree_id != path->btree_id ||
321 bpos_cmp(ck->key.pos, path->pos)) {
322 six_unlock_type(&ck->c.lock, lock_want);
326 mark_btree_node_locked(path, 0, lock_want);
329 path->l[0].lock_seq = ck->c.lock.state.seq;
330 path->l[0].b = (void *) ck;
332 if (!ck->valid && !(flags & BTREE_ITER_CACHED_NOFILL)) {
333 if (!path->locks_want &&
334 !__bch2_btree_path_upgrade(trans, path, 1)) {
335 trace_transaction_restart_ip(trans->ip, _THIS_IP_);
336 ret = btree_trans_restart(trans);
340 ret = btree_key_cache_fill(trans, path, ck);
345 if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
346 set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
348 path->uptodate = BTREE_ITER_UPTODATE;
349 BUG_ON(btree_node_locked_type(path, 0) != btree_lock_want(path, 0));
354 btree_node_unlock(path, 0);
355 path->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, b_iter;
369 struct bkey_cached *ck = NULL;
372 bch2_trans_iter_init(trans, &b_iter, key.btree_id, key.pos,
375 BTREE_ITER_ALL_SNAPSHOTS);
376 bch2_trans_iter_init(trans, &c_iter, key.btree_id, key.pos,
378 BTREE_ITER_CACHED_NOFILL|
379 BTREE_ITER_CACHED_NOCREATE|
381 ret = bch2_btree_iter_traverse(&c_iter);
385 ck = (void *) c_iter.path->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,
403 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
404 BTREE_TRIGGER_NORUN) ?:
405 bch2_trans_commit(trans, NULL, NULL,
406 BTREE_INSERT_NOCHECK_RW|
408 BTREE_INSERT_USE_RESERVE|
409 (ck->journal.seq == journal_last_seq(j)
410 ? BTREE_INSERT_JOURNAL_RESERVED
414 bch2_fs_fatal_err_on(ret != -EINTR &&
416 !bch2_journal_error(j), c,
417 "error flushing key cache: %i", ret);
421 bch2_journal_pin_drop(j, &ck->journal);
422 bch2_journal_preres_put(j, &ck->res);
424 BUG_ON(!btree_node_locked(c_iter.path, 0));
427 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
428 clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
429 atomic_long_dec(&c->btree_key_cache.nr_dirty);
433 BUG_ON(!btree_node_intent_locked(c_iter.path, 0));
435 mark_btree_node_unlocked(c_iter.path, 0);
436 c_iter.path->l[0].b = NULL;
438 six_lock_write(&ck->c.lock, NULL, NULL);
440 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
441 clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
442 atomic_long_dec(&c->btree_key_cache.nr_dirty);
445 bkey_cached_evict(&c->btree_key_cache, ck);
447 mutex_lock(&c->btree_key_cache.lock);
448 bkey_cached_free(&c->btree_key_cache, ck);
449 mutex_unlock(&c->btree_key_cache.lock);
452 bch2_trans_iter_exit(trans, &b_iter);
453 bch2_trans_iter_exit(trans, &c_iter);
457 int bch2_btree_key_cache_journal_flush(struct journal *j,
458 struct journal_entry_pin *pin, u64 seq)
460 struct bch_fs *c = container_of(j, struct bch_fs, journal);
461 struct bkey_cached *ck =
462 container_of(pin, struct bkey_cached, journal);
463 struct bkey_cached_key key;
466 int srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
468 six_lock_read(&ck->c.lock, NULL, NULL);
471 if (ck->journal.seq != seq ||
472 !test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
473 six_unlock_read(&ck->c.lock);
476 six_unlock_read(&ck->c.lock);
478 ret = bch2_trans_do(c, NULL, NULL, 0,
479 btree_key_cache_flush_pos(&trans, key, seq,
480 BTREE_INSERT_JOURNAL_RECLAIM, false));
482 srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
488 * Flush and evict a key from the key cache:
490 int bch2_btree_key_cache_flush(struct btree_trans *trans,
491 enum btree_id id, struct bpos pos)
493 struct bch_fs *c = trans->c;
494 struct bkey_cached_key key = { id, pos };
496 /* Fastpath - assume it won't be found: */
497 if (!bch2_btree_key_cache_find(c, id, pos))
500 return btree_key_cache_flush_pos(trans, key, 0, 0, true);
503 bool bch2_btree_insert_key_cached(struct btree_trans *trans,
504 struct btree_path *path,
505 struct bkey_i *insert)
507 struct bch_fs *c = trans->c;
508 struct bkey_cached *ck = (void *) path->l[0].b;
509 bool kick_reclaim = false;
511 BUG_ON(insert->u64s > ck->u64s);
513 if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
516 BUG_ON(jset_u64s(insert->u64s) > trans->journal_preres.u64s);
518 difference = jset_u64s(insert->u64s) - ck->res.u64s;
519 if (difference > 0) {
520 trans->journal_preres.u64s -= difference;
521 ck->res.u64s += difference;
525 bkey_copy(ck->k, insert);
528 if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
529 set_bit(BKEY_CACHED_DIRTY, &ck->flags);
530 atomic_long_inc(&c->btree_key_cache.nr_dirty);
532 if (bch2_nr_btree_keys_need_flush(c))
536 bch2_journal_pin_update(&c->journal, trans->journal_res.seq,
537 &ck->journal, bch2_btree_key_cache_journal_flush);
540 journal_reclaim_kick(&c->journal);
544 #ifdef CONFIG_BCACHEFS_DEBUG
545 void bch2_btree_key_cache_verify_clean(struct btree_trans *trans,
546 enum btree_id id, struct bpos pos)
548 BUG_ON(bch2_btree_key_cache_find(trans->c, id, pos));
552 static unsigned long bch2_btree_key_cache_scan(struct shrinker *shrink,
553 struct shrink_control *sc)
555 struct bch_fs *c = container_of(shrink, struct bch_fs,
556 btree_key_cache.shrink);
557 struct btree_key_cache *bc = &c->btree_key_cache;
558 struct bucket_table *tbl;
559 struct bkey_cached *ck, *t;
560 size_t scanned = 0, freed = 0, nr = sc->nr_to_scan;
561 unsigned start, flags;
564 /* Return -1 if we can't do anything right now */
565 if (sc->gfp_mask & __GFP_FS)
566 mutex_lock(&bc->lock);
567 else if (!mutex_trylock(&bc->lock))
570 srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
571 flags = memalloc_nofs_save();
574 * Newest freed entries are at the end of the list - once we hit one
575 * that's too new to be freed, we can bail out:
577 list_for_each_entry_safe(ck, t, &bc->freed, list) {
578 if (!poll_state_synchronize_srcu(&c->btree_trans_barrier,
579 ck->btree_trans_barrier_seq))
583 kmem_cache_free(bch2_key_cache, ck);
593 tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
594 if (bc->shrink_iter >= tbl->size)
596 start = bc->shrink_iter;
599 struct rhash_head *pos, *next;
601 pos = rht_ptr_rcu(rht_bucket(tbl, bc->shrink_iter));
603 while (!rht_is_a_nulls(pos)) {
604 next = rht_dereference_bucket_rcu(pos->next, tbl, bc->shrink_iter);
605 ck = container_of(pos, struct bkey_cached, hash);
607 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags))
610 if (test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
611 clear_bit(BKEY_CACHED_ACCESSED, &ck->flags);
612 else if (bkey_cached_lock_for_evict(ck)) {
613 bkey_cached_evict(bc, ck);
614 bkey_cached_free(bc, ck);
625 if (bc->shrink_iter >= tbl->size)
627 } while (scanned < nr && bc->shrink_iter != start);
631 memalloc_nofs_restore(flags);
632 srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
633 mutex_unlock(&bc->lock);
638 static unsigned long bch2_btree_key_cache_count(struct shrinker *shrink,
639 struct shrink_control *sc)
641 struct bch_fs *c = container_of(shrink, struct bch_fs,
642 btree_key_cache.shrink);
643 struct btree_key_cache *bc = &c->btree_key_cache;
644 long nr = atomic_long_read(&bc->nr_keys) -
645 atomic_long_read(&bc->nr_dirty);
650 void bch2_fs_btree_key_cache_exit(struct btree_key_cache *bc)
652 struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
653 struct bucket_table *tbl;
654 struct bkey_cached *ck, *n;
655 struct rhash_head *pos;
658 if (bc->shrink.list.next)
659 unregister_shrinker(&bc->shrink);
661 mutex_lock(&bc->lock);
664 tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
666 for (i = 0; i < tbl->size; i++)
667 rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
668 bkey_cached_evict(bc, ck);
669 list_add(&ck->list, &bc->freed);
673 list_for_each_entry_safe(ck, n, &bc->freed, list) {
676 bch2_journal_pin_drop(&c->journal, &ck->journal);
677 bch2_journal_preres_put(&c->journal, &ck->res);
681 kmem_cache_free(bch2_key_cache, ck);
684 BUG_ON(atomic_long_read(&bc->nr_dirty) &&
685 !bch2_journal_error(&c->journal) &&
686 test_bit(BCH_FS_WAS_RW, &c->flags));
687 BUG_ON(atomic_long_read(&bc->nr_keys));
689 mutex_unlock(&bc->lock);
691 if (bc->table_init_done)
692 rhashtable_destroy(&bc->table);
695 void bch2_fs_btree_key_cache_init_early(struct btree_key_cache *c)
697 mutex_init(&c->lock);
698 INIT_LIST_HEAD(&c->freed);
701 int bch2_fs_btree_key_cache_init(struct btree_key_cache *c)
705 ret = rhashtable_init(&c->table, &bch2_btree_key_cache_params);
709 c->table_init_done = true;
712 c->shrink.count_objects = bch2_btree_key_cache_count;
713 c->shrink.scan_objects = bch2_btree_key_cache_scan;
714 return register_shrinker(&c->shrink);
717 void bch2_btree_key_cache_to_text(struct printbuf *out, struct btree_key_cache *c)
719 pr_buf(out, "nr_freed:\t%zu\n", c->nr_freed);
720 pr_buf(out, "nr_keys:\t%zu\n", atomic_long_read(&c->nr_keys));
721 pr_buf(out, "nr_dirty:\t%zu\n", atomic_long_read(&c->nr_dirty));
724 void bch2_btree_key_cache_exit(void)
727 kmem_cache_destroy(bch2_key_cache);
730 int __init bch2_btree_key_cache_init(void)
732 bch2_key_cache = KMEM_CACHE(bkey_cached, 0);