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);
96 list_move_tail(&ck->list, &bc->freed_nonpcpu);
97 bc->nr_freed_nonpcpu++;
99 atomic_long_inc(&bc->nr_freed);
105 six_unlock_write(&ck->c.lock);
106 six_unlock_intent(&ck->c.lock);
110 static void __bkey_cached_move_to_freelist_ordered(struct btree_key_cache *bc,
111 struct bkey_cached *ck)
113 struct bkey_cached *pos;
115 bc->nr_freed_nonpcpu++;
117 list_for_each_entry_reverse(pos, &bc->freed_nonpcpu, list) {
118 if (ULONG_CMP_GE(ck->btree_trans_barrier_seq,
119 pos->btree_trans_barrier_seq)) {
120 list_move(&ck->list, &pos->list);
125 list_move(&ck->list, &bc->freed_nonpcpu);
129 static void bkey_cached_move_to_freelist(struct btree_key_cache *bc,
130 struct bkey_cached *ck)
132 BUG_ON(test_bit(BKEY_CACHED_DIRTY, &ck->flags));
134 if (!ck->c.lock.readers) {
136 struct btree_key_cache_freelist *f;
140 f = this_cpu_ptr(bc->pcpu_freed);
142 if (f->nr < ARRAY_SIZE(f->objs)) {
143 f->objs[f->nr++] = ck;
149 mutex_lock(&bc->lock);
151 f = this_cpu_ptr(bc->pcpu_freed);
153 while (f->nr > ARRAY_SIZE(f->objs) / 2) {
154 struct bkey_cached *ck2 = f->objs[--f->nr];
156 __bkey_cached_move_to_freelist_ordered(bc, ck2);
160 __bkey_cached_move_to_freelist_ordered(bc, ck);
161 mutex_unlock(&bc->lock);
164 mutex_lock(&bc->lock);
165 list_move_tail(&ck->list, &bc->freed_nonpcpu);
166 bc->nr_freed_nonpcpu++;
167 mutex_unlock(&bc->lock);
170 mutex_lock(&bc->lock);
171 list_move_tail(&ck->list, &bc->freed_pcpu);
172 mutex_unlock(&bc->lock);
176 static void bkey_cached_free_fast(struct btree_key_cache *bc,
177 struct bkey_cached *ck)
179 struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
181 ck->btree_trans_barrier_seq =
182 start_poll_synchronize_srcu(&c->btree_trans_barrier);
184 list_del_init(&ck->list);
185 atomic_long_inc(&bc->nr_freed);
191 bkey_cached_move_to_freelist(bc, ck);
193 six_unlock_write(&ck->c.lock);
194 six_unlock_intent(&ck->c.lock);
197 static struct bkey_cached *
198 bkey_cached_alloc(struct btree_trans *trans, struct btree_path *path,
201 struct bch_fs *c = trans->c;
202 struct btree_key_cache *bc = &c->btree_key_cache;
203 struct bkey_cached *ck = NULL;
204 bool pcpu_readers = btree_uses_pcpu_readers(path->btree_id);
209 struct btree_key_cache_freelist *f;
212 f = this_cpu_ptr(bc->pcpu_freed);
214 ck = f->objs[--f->nr];
218 mutex_lock(&bc->lock);
220 f = this_cpu_ptr(bc->pcpu_freed);
222 while (!list_empty(&bc->freed_nonpcpu) &&
223 f->nr < ARRAY_SIZE(f->objs) / 2) {
224 ck = list_last_entry(&bc->freed_nonpcpu, struct bkey_cached, list);
225 list_del_init(&ck->list);
226 bc->nr_freed_nonpcpu--;
227 f->objs[f->nr++] = ck;
230 ck = f->nr ? f->objs[--f->nr] : NULL;
232 mutex_unlock(&bc->lock);
235 mutex_lock(&bc->lock);
236 if (!list_empty(&bc->freed_nonpcpu)) {
237 ck = list_last_entry(&bc->freed_nonpcpu, struct bkey_cached, list);
238 list_del_init(&ck->list);
239 bc->nr_freed_nonpcpu--;
241 mutex_unlock(&bc->lock);
244 mutex_lock(&bc->lock);
245 if (!list_empty(&bc->freed_pcpu)) {
246 ck = list_last_entry(&bc->freed_pcpu, struct bkey_cached, list);
247 list_del_init(&ck->list);
249 mutex_unlock(&bc->lock);
253 ret = btree_node_lock_nopath(trans, &ck->c, SIX_LOCK_intent, _THIS_IP_);
255 bkey_cached_move_to_freelist(bc, ck);
259 path->l[0].b = (void *) ck;
260 path->l[0].lock_seq = six_lock_seq(&ck->c.lock);
261 mark_btree_node_locked(trans, path, 0, BTREE_NODE_INTENT_LOCKED);
263 ret = bch2_btree_node_lock_write(trans, path, &ck->c);
265 btree_node_unlock(trans, path, 0);
266 bkey_cached_move_to_freelist(bc, ck);
273 ck = allocate_dropping_locks(trans, ret,
274 kmem_cache_zalloc(bch2_key_cache, _gfp));
276 kmem_cache_free(bch2_key_cache, ck);
283 INIT_LIST_HEAD(&ck->list);
284 bch2_btree_lock_init(&ck->c, pcpu_readers ? SIX_LOCK_INIT_PCPU : 0);
287 BUG_ON(!six_trylock_intent(&ck->c.lock));
288 BUG_ON(!six_trylock_write(&ck->c.lock));
293 static struct bkey_cached *
294 bkey_cached_reuse(struct btree_key_cache *c)
296 struct bucket_table *tbl;
297 struct rhash_head *pos;
298 struct bkey_cached *ck;
301 mutex_lock(&c->lock);
303 tbl = rht_dereference_rcu(c->table.tbl, &c->table);
304 for (i = 0; i < tbl->size; i++)
305 rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
306 if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
307 bkey_cached_lock_for_evict(ck)) {
308 bkey_cached_evict(c, ck);
315 mutex_unlock(&c->lock);
319 static struct bkey_cached *
320 btree_key_cache_create(struct btree_trans *trans, struct btree_path *path)
322 struct bch_fs *c = trans->c;
323 struct btree_key_cache *bc = &c->btree_key_cache;
324 struct bkey_cached *ck;
325 bool was_new = false;
327 ck = bkey_cached_alloc(trans, path, &was_new);
332 ck = bkey_cached_reuse(bc);
334 bch_err(c, "error allocating memory for key cache item, btree %s",
335 bch2_btree_id_str(path->btree_id));
336 return ERR_PTR(-BCH_ERR_ENOMEM_btree_key_cache_create);
339 mark_btree_node_locked(trans, path, 0, BTREE_NODE_INTENT_LOCKED);
343 ck->c.btree_id = path->btree_id;
344 ck->key.btree_id = path->btree_id;
345 ck->key.pos = path->pos;
347 ck->flags = 1U << BKEY_CACHED_ACCESSED;
349 if (unlikely(rhashtable_lookup_insert_fast(&bc->table,
351 bch2_btree_key_cache_params))) {
352 /* We raced with another fill: */
354 if (likely(was_new)) {
355 six_unlock_write(&ck->c.lock);
356 six_unlock_intent(&ck->c.lock);
359 bkey_cached_free_fast(bc, ck);
362 mark_btree_node_locked(trans, path, 0, BTREE_NODE_UNLOCKED);
366 atomic_long_inc(&bc->nr_keys);
368 six_unlock_write(&ck->c.lock);
373 static int btree_key_cache_fill(struct btree_trans *trans,
374 struct btree_path *ck_path,
375 struct bkey_cached *ck)
377 struct btree_iter iter;
379 unsigned new_u64s = 0;
380 struct bkey_i *new_k = NULL;
383 k = bch2_bkey_get_iter(trans, &iter, ck->key.btree_id, ck->key.pos,
384 BTREE_ITER_KEY_CACHE_FILL|
385 BTREE_ITER_CACHED_NOFILL);
390 if (!bch2_btree_node_relock(trans, ck_path, 0)) {
391 trace_and_count(trans->c, trans_restart_relock_key_cache_fill, trans, _THIS_IP_, ck_path);
392 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_fill);
397 * bch2_varint_decode can read past the end of the buffer by at
398 * most 7 bytes (it won't be used):
400 new_u64s = k.k->u64s + 1;
403 * Allocate some extra space so that the transaction commit path is less
404 * likely to have to reallocate, since that requires a transaction
407 new_u64s = min(256U, (new_u64s * 3) / 2);
409 if (new_u64s > ck->u64s) {
410 new_u64s = roundup_pow_of_two(new_u64s);
411 new_k = kmalloc(new_u64s * sizeof(u64), GFP_NOWAIT|__GFP_NOWARN);
413 bch2_trans_unlock(trans);
415 new_k = kmalloc(new_u64s * sizeof(u64), GFP_KERNEL);
417 bch_err(trans->c, "error allocating memory for key cache key, btree %s u64s %u",
418 bch2_btree_id_str(ck->key.btree_id), new_u64s);
419 ret = -BCH_ERR_ENOMEM_btree_key_cache_fill;
423 if (!bch2_btree_node_relock(trans, ck_path, 0)) {
425 trace_and_count(trans->c, trans_restart_relock_key_cache_fill, trans, _THIS_IP_, ck_path);
426 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_fill);
430 ret = bch2_trans_relock(trans);
438 ret = bch2_btree_node_lock_write(trans, ck_path, &ck_path->l[0].b->c);
450 bkey_reassemble(ck->k, k);
452 bch2_btree_node_unlock_write(trans, ck_path, ck_path->l[0].b);
454 /* We're not likely to need this iterator again: */
455 set_btree_iter_dontneed(&iter);
457 bch2_trans_iter_exit(trans, &iter);
462 bch2_btree_path_traverse_cached_slowpath(struct btree_trans *trans, struct btree_path *path,
465 struct bch_fs *c = trans->c;
466 struct bkey_cached *ck;
473 if (bch2_btree_node_relock_notrace(trans, path, 0)) {
474 ck = (void *) path->l[0].b;
478 ck = bch2_btree_key_cache_find(c, path->btree_id, path->pos);
480 ck = btree_key_cache_create(trans, path);
481 ret = PTR_ERR_OR_ZERO(ck);
487 mark_btree_node_locked(trans, path, 0, BTREE_NODE_INTENT_LOCKED);
488 path->locks_want = 1;
490 enum six_lock_type lock_want = __btree_lock_want(path, 0);
492 ret = btree_node_lock(trans, path, (void *) ck, 0,
493 lock_want, _THIS_IP_);
494 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
499 if (ck->key.btree_id != path->btree_id ||
500 !bpos_eq(ck->key.pos, path->pos)) {
501 six_unlock_type(&ck->c.lock, lock_want);
505 mark_btree_node_locked(trans, path, 0,
506 (enum btree_node_locked_type) lock_want);
509 path->l[0].lock_seq = six_lock_seq(&ck->c.lock);
510 path->l[0].b = (void *) ck;
512 path->uptodate = BTREE_ITER_UPTODATE;
514 if (!ck->valid && !(flags & BTREE_ITER_CACHED_NOFILL)) {
516 * Using the underscore version because we haven't set
517 * path->uptodate yet:
519 if (!path->locks_want &&
520 !__bch2_btree_path_upgrade(trans, path, 1, NULL)) {
521 trace_and_count(trans->c, trans_restart_key_cache_upgrade, trans, _THIS_IP_);
522 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_upgrade);
526 ret = btree_key_cache_fill(trans, path, ck);
530 ret = bch2_btree_path_relock(trans, path, _THIS_IP_);
534 path->uptodate = BTREE_ITER_UPTODATE;
537 if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
538 set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
540 BUG_ON(btree_node_locked_type(path, 0) != btree_lock_want(path, 0));
541 BUG_ON(path->uptodate);
545 path->uptodate = BTREE_ITER_NEED_TRAVERSE;
546 if (!bch2_err_matches(ret, BCH_ERR_transaction_restart)) {
547 btree_node_unlock(trans, path, 0);
548 path->l[0].b = ERR_PTR(ret);
553 int bch2_btree_path_traverse_cached(struct btree_trans *trans, struct btree_path *path,
556 struct bch_fs *c = trans->c;
557 struct bkey_cached *ck;
560 EBUG_ON(path->level);
564 if (bch2_btree_node_relock_notrace(trans, path, 0)) {
565 ck = (void *) path->l[0].b;
569 ck = bch2_btree_key_cache_find(c, path->btree_id, path->pos);
571 return bch2_btree_path_traverse_cached_slowpath(trans, path, flags);
573 enum six_lock_type lock_want = __btree_lock_want(path, 0);
575 ret = btree_node_lock(trans, path, (void *) ck, 0,
576 lock_want, _THIS_IP_);
577 EBUG_ON(ret && !bch2_err_matches(ret, BCH_ERR_transaction_restart));
582 if (ck->key.btree_id != path->btree_id ||
583 !bpos_eq(ck->key.pos, path->pos)) {
584 six_unlock_type(&ck->c.lock, lock_want);
588 mark_btree_node_locked(trans, path, 0,
589 (enum btree_node_locked_type) lock_want);
592 path->l[0].lock_seq = six_lock_seq(&ck->c.lock);
593 path->l[0].b = (void *) ck;
596 return bch2_btree_path_traverse_cached_slowpath(trans, path, flags);
598 if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
599 set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
601 path->uptodate = BTREE_ITER_UPTODATE;
603 EBUG_ON(btree_node_locked_type(path, 0) != btree_lock_want(path, 0));
608 static int btree_key_cache_flush_pos(struct btree_trans *trans,
609 struct bkey_cached_key key,
611 unsigned commit_flags,
614 struct bch_fs *c = trans->c;
615 struct journal *j = &c->journal;
616 struct btree_iter c_iter, b_iter;
617 struct bkey_cached *ck = NULL;
620 bch2_trans_iter_init(trans, &b_iter, key.btree_id, key.pos,
623 BTREE_ITER_ALL_SNAPSHOTS);
624 bch2_trans_iter_init(trans, &c_iter, key.btree_id, key.pos,
627 b_iter.flags &= ~BTREE_ITER_WITH_KEY_CACHE;
629 ret = bch2_btree_iter_traverse(&c_iter);
633 ck = (void *) c_iter.path->l[0].b;
637 if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
645 if (journal_seq && ck->journal.seq != journal_seq)
648 trans->journal_res.seq = ck->journal.seq;
651 * If we're at the end of the journal, we really want to free up space
652 * in the journal right away - we don't want to pin that old journal
653 * sequence number with a new btree node write, we want to re-journal
656 if (ck->journal.seq == journal_last_seq(j))
657 commit_flags |= BCH_WATERMARK_reclaim;
659 commit_flags |= BCH_TRANS_COMMIT_no_journal_res;
661 ret = bch2_btree_iter_traverse(&b_iter) ?:
662 bch2_trans_update(trans, &b_iter, ck->k,
663 BTREE_UPDATE_KEY_CACHE_RECLAIM|
664 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
665 BTREE_TRIGGER_NORUN) ?:
666 bch2_trans_commit(trans, NULL, NULL,
667 BCH_TRANS_COMMIT_no_check_rw|
668 BCH_TRANS_COMMIT_no_enospc|
671 bch2_fs_fatal_err_on(ret &&
672 !bch2_err_matches(ret, BCH_ERR_transaction_restart) &&
673 !bch2_err_matches(ret, BCH_ERR_journal_reclaim_would_deadlock) &&
674 !bch2_journal_error(j), c,
675 "error flushing key cache: %s", bch2_err_str(ret));
679 bch2_journal_pin_drop(j, &ck->journal);
681 BUG_ON(!btree_node_locked(c_iter.path, 0));
684 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
685 clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
686 atomic_long_dec(&c->btree_key_cache.nr_dirty);
689 struct btree_path *path2;
691 trans_for_each_path(trans, path2)
692 if (path2 != c_iter.path)
693 __bch2_btree_path_unlock(trans, path2);
695 bch2_btree_node_lock_write_nofail(trans, c_iter.path, &ck->c);
697 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
698 clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
699 atomic_long_dec(&c->btree_key_cache.nr_dirty);
702 mark_btree_node_locked_noreset(c_iter.path, 0, BTREE_NODE_UNLOCKED);
703 bkey_cached_evict(&c->btree_key_cache, ck);
704 bkey_cached_free_fast(&c->btree_key_cache, ck);
707 bch2_trans_iter_exit(trans, &b_iter);
708 bch2_trans_iter_exit(trans, &c_iter);
712 int bch2_btree_key_cache_journal_flush(struct journal *j,
713 struct journal_entry_pin *pin, u64 seq)
715 struct bch_fs *c = container_of(j, struct bch_fs, journal);
716 struct bkey_cached *ck =
717 container_of(pin, struct bkey_cached, journal);
718 struct bkey_cached_key key;
719 struct btree_trans *trans = bch2_trans_get(c);
720 int srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
723 btree_node_lock_nopath_nofail(trans, &ck->c, SIX_LOCK_read);
726 if (ck->journal.seq != seq ||
727 !test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
728 six_unlock_read(&ck->c.lock);
732 if (ck->seq != seq) {
733 bch2_journal_pin_update(&c->journal, ck->seq, &ck->journal,
734 bch2_btree_key_cache_journal_flush);
735 six_unlock_read(&ck->c.lock);
738 six_unlock_read(&ck->c.lock);
740 ret = lockrestart_do(trans,
741 btree_key_cache_flush_pos(trans, key, seq,
742 BCH_TRANS_COMMIT_journal_reclaim, false));
744 srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
746 bch2_trans_put(trans);
751 * Flush and evict a key from the key cache:
753 int bch2_btree_key_cache_flush(struct btree_trans *trans,
754 enum btree_id id, struct bpos pos)
756 struct bch_fs *c = trans->c;
757 struct bkey_cached_key key = { id, pos };
759 /* Fastpath - assume it won't be found: */
760 if (!bch2_btree_key_cache_find(c, id, pos))
763 return btree_key_cache_flush_pos(trans, key, 0, 0, true);
766 bool bch2_btree_insert_key_cached(struct btree_trans *trans,
768 struct btree_insert_entry *insert_entry)
770 struct bch_fs *c = trans->c;
771 struct bkey_cached *ck = (void *) insert_entry->path->l[0].b;
772 struct bkey_i *insert = insert_entry->k;
773 bool kick_reclaim = false;
775 BUG_ON(insert->k.u64s > ck->u64s);
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 scanned += bc->nr_freed_nonpcpu;
855 list_for_each_entry_safe(ck, t, &bc->freed_nonpcpu, list) {
856 if (!poll_state_synchronize_srcu(&c->btree_trans_barrier,
857 ck->btree_trans_barrier_seq))
861 six_lock_exit(&ck->c.lock);
862 kmem_cache_free(bch2_key_cache, ck);
863 atomic_long_dec(&bc->nr_freed);
865 bc->nr_freed_nonpcpu--;
871 scanned += bc->nr_freed_pcpu;
873 list_for_each_entry_safe(ck, t, &bc->freed_pcpu, list) {
874 if (!poll_state_synchronize_srcu(&c->btree_trans_barrier,
875 ck->btree_trans_barrier_seq))
879 six_lock_exit(&ck->c.lock);
880 kmem_cache_free(bch2_key_cache, ck);
881 atomic_long_dec(&bc->nr_freed);
890 tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
891 if (bc->shrink_iter >= tbl->size)
893 start = bc->shrink_iter;
896 struct rhash_head *pos, *next;
898 pos = rht_ptr_rcu(rht_bucket(tbl, bc->shrink_iter));
900 while (!rht_is_a_nulls(pos)) {
901 next = rht_dereference_bucket_rcu(pos->next, tbl, bc->shrink_iter);
902 ck = container_of(pos, struct bkey_cached, hash);
904 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags))
907 if (test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
908 clear_bit(BKEY_CACHED_ACCESSED, &ck->flags);
909 else if (bkey_cached_lock_for_evict(ck)) {
910 bkey_cached_evict(bc, ck);
911 bkey_cached_free(bc, ck);
922 if (bc->shrink_iter >= tbl->size)
924 } while (scanned < nr && bc->shrink_iter != start);
928 memalloc_nofs_restore(flags);
929 srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
930 mutex_unlock(&bc->lock);
935 static unsigned long bch2_btree_key_cache_count(struct shrinker *shrink,
936 struct shrink_control *sc)
938 struct bch_fs *c = shrink->private_data;
939 struct btree_key_cache *bc = &c->btree_key_cache;
940 long nr = atomic_long_read(&bc->nr_keys) -
941 atomic_long_read(&bc->nr_dirty);
946 void bch2_fs_btree_key_cache_exit(struct btree_key_cache *bc)
948 struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
949 struct bucket_table *tbl;
950 struct bkey_cached *ck, *n;
951 struct rhash_head *pos;
958 shrinker_free(bc->shrink);
960 mutex_lock(&bc->lock);
963 * The loop is needed to guard against racing with rehash:
965 while (atomic_long_read(&bc->nr_keys)) {
967 tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
969 for (i = 0; i < tbl->size; i++)
970 rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
971 bkey_cached_evict(bc, ck);
972 list_add(&ck->list, &items);
978 for_each_possible_cpu(cpu) {
979 struct btree_key_cache_freelist *f =
980 per_cpu_ptr(bc->pcpu_freed, cpu);
982 for (i = 0; i < f->nr; i++) {
984 list_add(&ck->list, &items);
989 BUG_ON(list_count_nodes(&bc->freed_pcpu) != bc->nr_freed_pcpu);
990 BUG_ON(list_count_nodes(&bc->freed_nonpcpu) != bc->nr_freed_nonpcpu);
992 list_splice(&bc->freed_pcpu, &items);
993 list_splice(&bc->freed_nonpcpu, &items);
995 mutex_unlock(&bc->lock);
997 list_for_each_entry_safe(ck, n, &items, list) {
1000 bch2_journal_pin_drop(&c->journal, &ck->journal);
1002 list_del(&ck->list);
1004 six_lock_exit(&ck->c.lock);
1005 kmem_cache_free(bch2_key_cache, ck);
1008 if (atomic_long_read(&bc->nr_dirty) &&
1009 !bch2_journal_error(&c->journal) &&
1010 test_bit(BCH_FS_was_rw, &c->flags))
1011 panic("btree key cache shutdown error: nr_dirty nonzero (%li)\n",
1012 atomic_long_read(&bc->nr_dirty));
1014 if (atomic_long_read(&bc->nr_keys))
1015 panic("btree key cache shutdown error: nr_keys nonzero (%li)\n",
1016 atomic_long_read(&bc->nr_keys));
1018 if (bc->table_init_done)
1019 rhashtable_destroy(&bc->table);
1021 free_percpu(bc->pcpu_freed);
1024 void bch2_fs_btree_key_cache_init_early(struct btree_key_cache *c)
1026 mutex_init(&c->lock);
1027 INIT_LIST_HEAD(&c->freed_pcpu);
1028 INIT_LIST_HEAD(&c->freed_nonpcpu);
1031 int bch2_fs_btree_key_cache_init(struct btree_key_cache *bc)
1033 struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
1034 struct shrinker *shrink;
1037 bc->pcpu_freed = alloc_percpu(struct btree_key_cache_freelist);
1038 if (!bc->pcpu_freed)
1039 return -BCH_ERR_ENOMEM_fs_btree_cache_init;
1042 if (rhashtable_init(&bc->table, &bch2_btree_key_cache_params))
1043 return -BCH_ERR_ENOMEM_fs_btree_cache_init;
1045 bc->table_init_done = true;
1047 shrink = shrinker_alloc(0, "%s-btree_key_cache", c->name);
1049 return -BCH_ERR_ENOMEM_fs_btree_cache_init;
1050 bc->shrink = shrink;
1052 shrink->count_objects = bch2_btree_key_cache_count;
1053 shrink->scan_objects = bch2_btree_key_cache_scan;
1054 shrink->private_data = c;
1055 shrinker_register(shrink);
1059 void bch2_btree_key_cache_to_text(struct printbuf *out, struct btree_key_cache *c)
1061 prt_printf(out, "nr_freed:\t%lu", atomic_long_read(&c->nr_freed));
1063 prt_printf(out, "nr_keys:\t%lu", atomic_long_read(&c->nr_keys));
1065 prt_printf(out, "nr_dirty:\t%lu", atomic_long_read(&c->nr_dirty));
1069 void bch2_btree_key_cache_exit(void)
1071 kmem_cache_destroy(bch2_key_cache);
1074 int __init bch2_btree_key_cache_init(void)
1076 bch2_key_cache = KMEM_CACHE(bkey_cached, SLAB_RECLAIM_ACCOUNT);
1077 if (!bch2_key_cache)