1 // SPDX-License-Identifier: GPL-2.0
4 #include "btree_cache.h"
5 #include "btree_iter.h"
6 #include "btree_key_cache.h"
7 #include "btree_locking.h"
8 #include "btree_update.h"
12 #include "journal_reclaim.h"
15 #include <linux/sched/mm.h>
17 static inline bool btree_uses_pcpu_readers(enum btree_id id)
19 return id == BTREE_ID_subvolumes;
22 static struct kmem_cache *bch2_key_cache;
24 static int bch2_btree_key_cache_cmp_fn(struct rhashtable_compare_arg *arg,
27 const struct bkey_cached *ck = obj;
28 const struct bkey_cached_key *key = arg->key;
30 return ck->key.btree_id != key->btree_id ||
31 !bpos_eq(ck->key.pos, key->pos);
34 static const struct rhashtable_params bch2_btree_key_cache_params = {
35 .head_offset = offsetof(struct bkey_cached, hash),
36 .key_offset = offsetof(struct bkey_cached, key),
37 .key_len = sizeof(struct bkey_cached_key),
38 .obj_cmpfn = bch2_btree_key_cache_cmp_fn,
42 inline struct bkey_cached *
43 bch2_btree_key_cache_find(struct bch_fs *c, enum btree_id btree_id, struct bpos pos)
45 struct bkey_cached_key key = {
50 return rhashtable_lookup_fast(&c->btree_key_cache.table, &key,
51 bch2_btree_key_cache_params);
54 static bool bkey_cached_lock_for_evict(struct bkey_cached *ck)
56 if (!six_trylock_intent(&ck->c.lock))
59 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
60 six_unlock_intent(&ck->c.lock);
64 if (!six_trylock_write(&ck->c.lock)) {
65 six_unlock_intent(&ck->c.lock);
72 static void bkey_cached_evict(struct btree_key_cache *c,
73 struct bkey_cached *ck)
75 BUG_ON(rhashtable_remove_fast(&c->table, &ck->hash,
76 bch2_btree_key_cache_params));
77 memset(&ck->key, ~0, sizeof(ck->key));
79 atomic_long_dec(&c->nr_keys);
82 static void bkey_cached_free(struct btree_key_cache *bc,
83 struct bkey_cached *ck)
85 struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
87 BUG_ON(test_bit(BKEY_CACHED_DIRTY, &ck->flags));
89 ck->btree_trans_barrier_seq =
90 start_poll_synchronize_srcu(&c->btree_trans_barrier);
92 if (ck->c.lock.readers)
93 list_move_tail(&ck->list, &bc->freed_pcpu);
95 list_move_tail(&ck->list, &bc->freed_nonpcpu);
96 atomic_long_inc(&bc->nr_freed);
102 six_unlock_write(&ck->c.lock);
103 six_unlock_intent(&ck->c.lock);
107 static void __bkey_cached_move_to_freelist_ordered(struct btree_key_cache *bc,
108 struct bkey_cached *ck)
110 struct bkey_cached *pos;
112 list_for_each_entry_reverse(pos, &bc->freed_nonpcpu, list) {
113 if (ULONG_CMP_GE(ck->btree_trans_barrier_seq,
114 pos->btree_trans_barrier_seq)) {
115 list_move(&ck->list, &pos->list);
120 list_move(&ck->list, &bc->freed_nonpcpu);
124 static void bkey_cached_move_to_freelist(struct btree_key_cache *bc,
125 struct bkey_cached *ck)
127 BUG_ON(test_bit(BKEY_CACHED_DIRTY, &ck->flags));
129 if (!ck->c.lock.readers) {
131 struct btree_key_cache_freelist *f;
135 f = this_cpu_ptr(bc->pcpu_freed);
137 if (f->nr < ARRAY_SIZE(f->objs)) {
138 f->objs[f->nr++] = ck;
144 mutex_lock(&bc->lock);
146 f = this_cpu_ptr(bc->pcpu_freed);
148 while (f->nr > ARRAY_SIZE(f->objs) / 2) {
149 struct bkey_cached *ck2 = f->objs[--f->nr];
151 __bkey_cached_move_to_freelist_ordered(bc, ck2);
155 __bkey_cached_move_to_freelist_ordered(bc, ck);
156 mutex_unlock(&bc->lock);
159 mutex_lock(&bc->lock);
160 list_move_tail(&ck->list, &bc->freed_nonpcpu);
161 mutex_unlock(&bc->lock);
164 mutex_lock(&bc->lock);
165 list_move_tail(&ck->list, &bc->freed_pcpu);
166 mutex_unlock(&bc->lock);
170 static void bkey_cached_free_fast(struct btree_key_cache *bc,
171 struct bkey_cached *ck)
173 struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
175 ck->btree_trans_barrier_seq =
176 start_poll_synchronize_srcu(&c->btree_trans_barrier);
178 list_del_init(&ck->list);
179 atomic_long_inc(&bc->nr_freed);
185 bkey_cached_move_to_freelist(bc, ck);
187 six_unlock_write(&ck->c.lock);
188 six_unlock_intent(&ck->c.lock);
191 static struct bkey_cached *
192 bkey_cached_alloc(struct btree_trans *trans, struct btree_path *path,
195 struct bch_fs *c = trans->c;
196 struct btree_key_cache *bc = &c->btree_key_cache;
197 struct bkey_cached *ck = NULL;
198 bool pcpu_readers = btree_uses_pcpu_readers(path->btree_id);
203 struct btree_key_cache_freelist *f;
206 f = this_cpu_ptr(bc->pcpu_freed);
208 ck = f->objs[--f->nr];
212 mutex_lock(&bc->lock);
214 f = this_cpu_ptr(bc->pcpu_freed);
216 while (!list_empty(&bc->freed_nonpcpu) &&
217 f->nr < ARRAY_SIZE(f->objs) / 2) {
218 ck = list_last_entry(&bc->freed_nonpcpu, struct bkey_cached, list);
219 list_del_init(&ck->list);
220 f->objs[f->nr++] = ck;
223 ck = f->nr ? f->objs[--f->nr] : NULL;
225 mutex_unlock(&bc->lock);
228 mutex_lock(&bc->lock);
229 if (!list_empty(&bc->freed_nonpcpu)) {
230 ck = list_last_entry(&bc->freed_nonpcpu, struct bkey_cached, list);
231 list_del_init(&ck->list);
233 mutex_unlock(&bc->lock);
236 mutex_lock(&bc->lock);
237 if (!list_empty(&bc->freed_pcpu)) {
238 ck = list_last_entry(&bc->freed_pcpu, struct bkey_cached, list);
239 list_del_init(&ck->list);
241 mutex_unlock(&bc->lock);
245 ret = btree_node_lock_nopath(trans, &ck->c, SIX_LOCK_intent, _THIS_IP_);
247 bkey_cached_move_to_freelist(bc, ck);
251 path->l[0].b = (void *) ck;
252 path->l[0].lock_seq = six_lock_seq(&ck->c.lock);
253 mark_btree_node_locked(trans, path, 0, BTREE_NODE_INTENT_LOCKED);
255 ret = bch2_btree_node_lock_write(trans, path, &ck->c);
257 btree_node_unlock(trans, path, 0);
258 bkey_cached_move_to_freelist(bc, ck);
265 ck = allocate_dropping_locks(trans, ret,
266 kmem_cache_zalloc(bch2_key_cache, _gfp));
268 kmem_cache_free(bch2_key_cache, ck);
275 INIT_LIST_HEAD(&ck->list);
276 bch2_btree_lock_init(&ck->c, pcpu_readers ? SIX_LOCK_INIT_PCPU : 0);
279 BUG_ON(!six_trylock_intent(&ck->c.lock));
280 BUG_ON(!six_trylock_write(&ck->c.lock));
285 static struct bkey_cached *
286 bkey_cached_reuse(struct btree_key_cache *c)
288 struct bucket_table *tbl;
289 struct rhash_head *pos;
290 struct bkey_cached *ck;
293 mutex_lock(&c->lock);
295 tbl = rht_dereference_rcu(c->table.tbl, &c->table);
296 for (i = 0; i < tbl->size; i++)
297 rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
298 if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
299 bkey_cached_lock_for_evict(ck)) {
300 bkey_cached_evict(c, ck);
307 mutex_unlock(&c->lock);
311 static struct bkey_cached *
312 btree_key_cache_create(struct btree_trans *trans, struct btree_path *path)
314 struct bch_fs *c = trans->c;
315 struct btree_key_cache *bc = &c->btree_key_cache;
316 struct bkey_cached *ck;
317 bool was_new = false;
319 ck = bkey_cached_alloc(trans, path, &was_new);
324 ck = bkey_cached_reuse(bc);
326 bch_err(c, "error allocating memory for key cache item, btree %s",
327 bch2_btree_ids[path->btree_id]);
328 return ERR_PTR(-BCH_ERR_ENOMEM_btree_key_cache_create);
331 mark_btree_node_locked(trans, path, 0, BTREE_NODE_INTENT_LOCKED);
335 ck->c.btree_id = path->btree_id;
336 ck->key.btree_id = path->btree_id;
337 ck->key.pos = path->pos;
339 ck->flags = 1U << BKEY_CACHED_ACCESSED;
341 if (unlikely(rhashtable_lookup_insert_fast(&bc->table,
343 bch2_btree_key_cache_params))) {
344 /* We raced with another fill: */
346 if (likely(was_new)) {
347 six_unlock_write(&ck->c.lock);
348 six_unlock_intent(&ck->c.lock);
351 bkey_cached_free_fast(bc, ck);
354 mark_btree_node_locked(trans, path, 0, BTREE_NODE_UNLOCKED);
358 atomic_long_inc(&bc->nr_keys);
360 six_unlock_write(&ck->c.lock);
365 static int btree_key_cache_fill(struct btree_trans *trans,
366 struct btree_path *ck_path,
367 struct bkey_cached *ck)
369 struct btree_iter iter;
371 unsigned new_u64s = 0;
372 struct bkey_i *new_k = NULL;
375 k = bch2_bkey_get_iter(trans, &iter, ck->key.btree_id, ck->key.pos,
376 BTREE_ITER_KEY_CACHE_FILL|
377 BTREE_ITER_CACHED_NOFILL);
382 if (!bch2_btree_node_relock(trans, ck_path, 0)) {
383 trace_and_count(trans->c, trans_restart_relock_key_cache_fill, trans, _THIS_IP_, ck_path);
384 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_fill);
389 * bch2_varint_decode can read past the end of the buffer by at
390 * most 7 bytes (it won't be used):
392 new_u64s = k.k->u64s + 1;
395 * Allocate some extra space so that the transaction commit path is less
396 * likely to have to reallocate, since that requires a transaction
399 new_u64s = min(256U, (new_u64s * 3) / 2);
401 if (new_u64s > ck->u64s) {
402 new_u64s = roundup_pow_of_two(new_u64s);
403 new_k = kmalloc(new_u64s * sizeof(u64), GFP_NOWAIT|__GFP_NOWARN);
405 bch2_trans_unlock(trans);
407 new_k = kmalloc(new_u64s * sizeof(u64), GFP_KERNEL);
409 bch_err(trans->c, "error allocating memory for key cache key, btree %s u64s %u",
410 bch2_btree_ids[ck->key.btree_id], new_u64s);
411 ret = -BCH_ERR_ENOMEM_btree_key_cache_fill;
415 if (!bch2_btree_node_relock(trans, ck_path, 0)) {
417 trace_and_count(trans->c, trans_restart_relock_key_cache_fill, trans, _THIS_IP_, ck_path);
418 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_fill);
422 ret = bch2_trans_relock(trans);
430 ret = bch2_btree_node_lock_write(trans, ck_path, &ck_path->l[0].b->c);
442 bkey_reassemble(ck->k, k);
444 bch2_btree_node_unlock_write(trans, ck_path, ck_path->l[0].b);
446 /* We're not likely to need this iterator again: */
447 set_btree_iter_dontneed(&iter);
449 bch2_trans_iter_exit(trans, &iter);
454 bch2_btree_path_traverse_cached_slowpath(struct btree_trans *trans, struct btree_path *path,
457 struct bch_fs *c = trans->c;
458 struct bkey_cached *ck;
465 if (bch2_btree_node_relock_notrace(trans, path, 0)) {
466 ck = (void *) path->l[0].b;
470 ck = bch2_btree_key_cache_find(c, path->btree_id, path->pos);
472 ck = btree_key_cache_create(trans, path);
473 ret = PTR_ERR_OR_ZERO(ck);
479 mark_btree_node_locked(trans, path, 0, BTREE_NODE_INTENT_LOCKED);
480 path->locks_want = 1;
482 enum six_lock_type lock_want = __btree_lock_want(path, 0);
484 ret = btree_node_lock(trans, path, (void *) ck, 0,
485 lock_want, _THIS_IP_);
486 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
491 if (ck->key.btree_id != path->btree_id ||
492 !bpos_eq(ck->key.pos, path->pos)) {
493 six_unlock_type(&ck->c.lock, lock_want);
497 mark_btree_node_locked(trans, path, 0,
498 (enum btree_node_locked_type) lock_want);
501 path->l[0].lock_seq = six_lock_seq(&ck->c.lock);
502 path->l[0].b = (void *) ck;
504 path->uptodate = BTREE_ITER_UPTODATE;
506 if (!ck->valid && !(flags & BTREE_ITER_CACHED_NOFILL)) {
508 * Using the underscore version because we haven't set
509 * path->uptodate yet:
511 if (!path->locks_want &&
512 !__bch2_btree_path_upgrade(trans, path, 1)) {
513 trace_and_count(trans->c, trans_restart_key_cache_upgrade, trans, _THIS_IP_);
514 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_upgrade);
518 ret = btree_key_cache_fill(trans, path, ck);
522 ret = bch2_btree_path_relock(trans, path, _THIS_IP_);
526 path->uptodate = BTREE_ITER_UPTODATE;
529 if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
530 set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
532 BUG_ON(btree_node_locked_type(path, 0) != btree_lock_want(path, 0));
533 BUG_ON(path->uptodate);
537 path->uptodate = BTREE_ITER_NEED_TRAVERSE;
538 if (!bch2_err_matches(ret, BCH_ERR_transaction_restart)) {
539 btree_node_unlock(trans, path, 0);
540 path->l[0].b = ERR_PTR(ret);
545 int bch2_btree_path_traverse_cached(struct btree_trans *trans, struct btree_path *path,
548 struct bch_fs *c = trans->c;
549 struct bkey_cached *ck;
552 EBUG_ON(path->level);
556 if (bch2_btree_node_relock_notrace(trans, path, 0)) {
557 ck = (void *) path->l[0].b;
561 ck = bch2_btree_key_cache_find(c, path->btree_id, path->pos);
563 return bch2_btree_path_traverse_cached_slowpath(trans, path, flags);
565 enum six_lock_type lock_want = __btree_lock_want(path, 0);
567 ret = btree_node_lock(trans, path, (void *) ck, 0,
568 lock_want, _THIS_IP_);
569 EBUG_ON(ret && !bch2_err_matches(ret, BCH_ERR_transaction_restart));
574 if (ck->key.btree_id != path->btree_id ||
575 !bpos_eq(ck->key.pos, path->pos)) {
576 six_unlock_type(&ck->c.lock, lock_want);
580 mark_btree_node_locked(trans, path, 0,
581 (enum btree_node_locked_type) lock_want);
584 path->l[0].lock_seq = six_lock_seq(&ck->c.lock);
585 path->l[0].b = (void *) ck;
588 return bch2_btree_path_traverse_cached_slowpath(trans, path, flags);
590 if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
591 set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
593 path->uptodate = BTREE_ITER_UPTODATE;
595 EBUG_ON(btree_node_locked_type(path, 0) != btree_lock_want(path, 0));
600 static int btree_key_cache_flush_pos(struct btree_trans *trans,
601 struct bkey_cached_key key,
603 unsigned commit_flags,
606 struct bch_fs *c = trans->c;
607 struct journal *j = &c->journal;
608 struct btree_iter c_iter, b_iter;
609 struct bkey_cached *ck = NULL;
612 bch2_trans_iter_init(trans, &b_iter, key.btree_id, key.pos,
615 BTREE_ITER_ALL_SNAPSHOTS);
616 bch2_trans_iter_init(trans, &c_iter, key.btree_id, key.pos,
619 b_iter.flags &= ~BTREE_ITER_WITH_KEY_CACHE;
621 ret = bch2_btree_iter_traverse(&c_iter);
625 ck = (void *) c_iter.path->l[0].b;
629 if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
637 if (journal_seq && ck->journal.seq != journal_seq)
641 * Since journal reclaim depends on us making progress here, and the
642 * allocator/copygc depend on journal reclaim making progress, we need
643 * to be using alloc reserves:
645 ret = bch2_btree_iter_traverse(&b_iter) ?:
646 bch2_trans_update(trans, &b_iter, ck->k,
647 BTREE_UPDATE_KEY_CACHE_RECLAIM|
648 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
649 BTREE_TRIGGER_NORUN) ?:
650 bch2_trans_commit(trans, NULL, NULL,
651 BTREE_INSERT_NOCHECK_RW|
653 (ck->journal.seq == journal_last_seq(j)
654 ? BCH_WATERMARK_reclaim
658 bch2_fs_fatal_err_on(ret &&
659 !bch2_err_matches(ret, BCH_ERR_transaction_restart) &&
660 !bch2_err_matches(ret, BCH_ERR_journal_reclaim_would_deadlock) &&
661 !bch2_journal_error(j), c,
662 "error flushing key cache: %s", bch2_err_str(ret));
666 bch2_journal_pin_drop(j, &ck->journal);
667 bch2_journal_preres_put(j, &ck->res);
669 BUG_ON(!btree_node_locked(c_iter.path, 0));
672 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
673 clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
674 atomic_long_dec(&c->btree_key_cache.nr_dirty);
677 struct btree_path *path2;
679 trans_for_each_path(trans, path2)
680 if (path2 != c_iter.path)
681 __bch2_btree_path_unlock(trans, path2);
683 bch2_btree_node_lock_write_nofail(trans, c_iter.path, &ck->c);
685 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
686 clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
687 atomic_long_dec(&c->btree_key_cache.nr_dirty);
690 mark_btree_node_locked_noreset(c_iter.path, 0, BTREE_NODE_UNLOCKED);
691 bkey_cached_evict(&c->btree_key_cache, ck);
692 bkey_cached_free_fast(&c->btree_key_cache, ck);
695 bch2_trans_iter_exit(trans, &b_iter);
696 bch2_trans_iter_exit(trans, &c_iter);
700 int bch2_btree_key_cache_journal_flush(struct journal *j,
701 struct journal_entry_pin *pin, u64 seq)
703 struct bch_fs *c = container_of(j, struct bch_fs, journal);
704 struct bkey_cached *ck =
705 container_of(pin, struct bkey_cached, journal);
706 struct bkey_cached_key key;
707 struct btree_trans *trans = bch2_trans_get(c);
708 int srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
711 btree_node_lock_nopath_nofail(trans, &ck->c, SIX_LOCK_read);
714 if (ck->journal.seq != seq ||
715 !test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
716 six_unlock_read(&ck->c.lock);
720 if (ck->seq != seq) {
721 bch2_journal_pin_update(&c->journal, ck->seq, &ck->journal,
722 bch2_btree_key_cache_journal_flush);
723 six_unlock_read(&ck->c.lock);
726 six_unlock_read(&ck->c.lock);
728 ret = commit_do(trans, NULL, NULL, 0,
729 btree_key_cache_flush_pos(trans, key, seq,
730 BTREE_INSERT_JOURNAL_RECLAIM, false));
732 srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
734 bch2_trans_put(trans);
739 * Flush and evict a key from the key cache:
741 int bch2_btree_key_cache_flush(struct btree_trans *trans,
742 enum btree_id id, struct bpos pos)
744 struct bch_fs *c = trans->c;
745 struct bkey_cached_key key = { id, pos };
747 /* Fastpath - assume it won't be found: */
748 if (!bch2_btree_key_cache_find(c, id, pos))
751 return btree_key_cache_flush_pos(trans, key, 0, 0, true);
754 bool bch2_btree_insert_key_cached(struct btree_trans *trans,
756 struct btree_insert_entry *insert_entry)
758 struct bch_fs *c = trans->c;
759 struct bkey_cached *ck = (void *) insert_entry->path->l[0].b;
760 struct bkey_i *insert = insert_entry->k;
761 bool kick_reclaim = false;
763 BUG_ON(insert->k.u64s > ck->u64s);
765 if (likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY))) {
768 BUG_ON(jset_u64s(insert->k.u64s) > trans->journal_preres.u64s);
770 difference = jset_u64s(insert->k.u64s) - ck->res.u64s;
771 if (difference > 0) {
772 trans->journal_preres.u64s -= difference;
773 ck->res.u64s += difference;
777 bkey_copy(ck->k, insert);
780 if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
781 EBUG_ON(test_bit(BCH_FS_CLEAN_SHUTDOWN, &c->flags));
782 set_bit(BKEY_CACHED_DIRTY, &ck->flags);
783 atomic_long_inc(&c->btree_key_cache.nr_dirty);
785 if (bch2_nr_btree_keys_need_flush(c))
790 * To minimize lock contention, we only add the journal pin here and
791 * defer pin updates to the flush callback via ->seq. Be careful not to
792 * update ->seq on nojournal commits because we don't want to update the
793 * pin to a seq that doesn't include journal updates on disk. Otherwise
794 * we risk losing the update after a crash.
796 * The only exception is if the pin is not active in the first place. We
797 * have to add the pin because journal reclaim drives key cache
798 * flushing. The flush callback will not proceed unless ->seq matches
799 * the latest pin, so make sure it starts with a consistent value.
801 if (!(insert_entry->flags & BTREE_UPDATE_NOJOURNAL) ||
802 !journal_pin_active(&ck->journal)) {
803 ck->seq = trans->journal_res.seq;
805 bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
806 &ck->journal, bch2_btree_key_cache_journal_flush);
809 journal_reclaim_kick(&c->journal);
813 void bch2_btree_key_cache_drop(struct btree_trans *trans,
814 struct btree_path *path)
816 struct bch_fs *c = trans->c;
817 struct bkey_cached *ck = (void *) path->l[0].b;
822 * We just did an update to the btree, bypassing the key cache: the key
823 * cache key is now stale and must be dropped, even if dirty:
825 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
826 clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
827 atomic_long_dec(&c->btree_key_cache.nr_dirty);
828 bch2_journal_pin_drop(&c->journal, &ck->journal);
834 static unsigned long bch2_btree_key_cache_scan(struct shrinker *shrink,
835 struct shrink_control *sc)
837 struct bch_fs *c = shrink->private_data;
838 struct btree_key_cache *bc = &c->btree_key_cache;
839 struct bucket_table *tbl;
840 struct bkey_cached *ck, *t;
841 size_t scanned = 0, freed = 0, nr = sc->nr_to_scan;
842 unsigned start, flags;
845 mutex_lock(&bc->lock);
846 srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
847 flags = memalloc_nofs_save();
850 * Newest freed entries are at the end of the list - once we hit one
851 * that's too new to be freed, we can bail out:
853 list_for_each_entry_safe(ck, t, &bc->freed_nonpcpu, list) {
854 if (!poll_state_synchronize_srcu(&c->btree_trans_barrier,
855 ck->btree_trans_barrier_seq))
859 six_lock_exit(&ck->c.lock);
860 kmem_cache_free(bch2_key_cache, ck);
861 atomic_long_dec(&bc->nr_freed);
869 list_for_each_entry_safe(ck, t, &bc->freed_pcpu, list) {
870 if (!poll_state_synchronize_srcu(&c->btree_trans_barrier,
871 ck->btree_trans_barrier_seq))
875 six_lock_exit(&ck->c.lock);
876 kmem_cache_free(bch2_key_cache, ck);
877 atomic_long_dec(&bc->nr_freed);
886 tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
887 if (bc->shrink_iter >= tbl->size)
889 start = bc->shrink_iter;
892 struct rhash_head *pos, *next;
894 pos = rht_ptr_rcu(rht_bucket(tbl, bc->shrink_iter));
896 while (!rht_is_a_nulls(pos)) {
897 next = rht_dereference_bucket_rcu(pos->next, tbl, bc->shrink_iter);
898 ck = container_of(pos, struct bkey_cached, hash);
900 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags))
903 if (test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
904 clear_bit(BKEY_CACHED_ACCESSED, &ck->flags);
905 else if (bkey_cached_lock_for_evict(ck)) {
906 bkey_cached_evict(bc, ck);
907 bkey_cached_free(bc, ck);
918 if (bc->shrink_iter >= tbl->size)
920 } while (scanned < nr && bc->shrink_iter != start);
924 memalloc_nofs_restore(flags);
925 srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
926 mutex_unlock(&bc->lock);
931 static unsigned long bch2_btree_key_cache_count(struct shrinker *shrink,
932 struct shrink_control *sc)
934 struct bch_fs *c = shrink->private_data;
935 struct btree_key_cache *bc = &c->btree_key_cache;
936 long nr = atomic_long_read(&bc->nr_keys) -
937 atomic_long_read(&bc->nr_dirty);
942 void bch2_fs_btree_key_cache_exit(struct btree_key_cache *bc)
944 struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
945 struct bucket_table *tbl;
946 struct bkey_cached *ck, *n;
947 struct rhash_head *pos;
954 shrinker_free(bc->shrink);
956 mutex_lock(&bc->lock);
959 * The loop is needed to guard against racing with rehash:
961 while (atomic_long_read(&bc->nr_keys)) {
963 tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
965 for (i = 0; i < tbl->size; i++)
966 rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
967 bkey_cached_evict(bc, ck);
968 list_add(&ck->list, &items);
974 for_each_possible_cpu(cpu) {
975 struct btree_key_cache_freelist *f =
976 per_cpu_ptr(bc->pcpu_freed, cpu);
978 for (i = 0; i < f->nr; i++) {
980 list_add(&ck->list, &items);
985 list_splice(&bc->freed_pcpu, &items);
986 list_splice(&bc->freed_nonpcpu, &items);
988 mutex_unlock(&bc->lock);
990 list_for_each_entry_safe(ck, n, &items, list) {
993 bch2_journal_pin_drop(&c->journal, &ck->journal);
994 bch2_journal_preres_put(&c->journal, &ck->res);
998 six_lock_exit(&ck->c.lock);
999 kmem_cache_free(bch2_key_cache, ck);
1002 if (atomic_long_read(&bc->nr_dirty) &&
1003 !bch2_journal_error(&c->journal) &&
1004 test_bit(BCH_FS_WAS_RW, &c->flags))
1005 panic("btree key cache shutdown error: nr_dirty nonzero (%li)\n",
1006 atomic_long_read(&bc->nr_dirty));
1008 if (atomic_long_read(&bc->nr_keys))
1009 panic("btree key cache shutdown error: nr_keys nonzero (%li)\n",
1010 atomic_long_read(&bc->nr_keys));
1012 if (bc->table_init_done)
1013 rhashtable_destroy(&bc->table);
1015 free_percpu(bc->pcpu_freed);
1018 void bch2_fs_btree_key_cache_init_early(struct btree_key_cache *c)
1020 mutex_init(&c->lock);
1021 INIT_LIST_HEAD(&c->freed_pcpu);
1022 INIT_LIST_HEAD(&c->freed_nonpcpu);
1025 int bch2_fs_btree_key_cache_init(struct btree_key_cache *bc)
1027 struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
1028 struct shrinker *shrink;
1031 bc->pcpu_freed = alloc_percpu(struct btree_key_cache_freelist);
1032 if (!bc->pcpu_freed)
1033 return -BCH_ERR_ENOMEM_fs_btree_cache_init;
1036 if (rhashtable_init(&bc->table, &bch2_btree_key_cache_params))
1037 return -BCH_ERR_ENOMEM_fs_btree_cache_init;
1039 bc->table_init_done = true;
1041 shrink = shrinker_alloc(0, "%s/btree_key_cache", c->name);
1043 return -BCH_ERR_ENOMEM_fs_btree_cache_init;
1044 bc->shrink = shrink;
1046 shrink->count_objects = bch2_btree_key_cache_count;
1047 shrink->scan_objects = bch2_btree_key_cache_scan;
1048 shrink->private_data = c;
1049 shrinker_register(shrink);
1053 void bch2_btree_key_cache_to_text(struct printbuf *out, struct btree_key_cache *c)
1055 prt_printf(out, "nr_freed:\t%lu", atomic_long_read(&c->nr_freed));
1057 prt_printf(out, "nr_keys:\t%lu", atomic_long_read(&c->nr_keys));
1059 prt_printf(out, "nr_dirty:\t%lu", atomic_long_read(&c->nr_dirty));
1063 void bch2_btree_key_cache_exit(void)
1065 kmem_cache_destroy(bch2_key_cache);
1068 int __init bch2_btree_key_cache_init(void)
1070 bch2_key_cache = KMEM_CACHE(bkey_cached, SLAB_RECLAIM_ACCOUNT);
1071 if (!bch2_key_cache)