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>
16 #include <linux/seq_buf.h>
18 static inline bool btree_uses_pcpu_readers(enum btree_id id)
20 return id == BTREE_ID_subvolumes;
23 static struct kmem_cache *bch2_key_cache;
25 static int bch2_btree_key_cache_cmp_fn(struct rhashtable_compare_arg *arg,
28 const struct bkey_cached *ck = obj;
29 const struct bkey_cached_key *key = arg->key;
31 return ck->key.btree_id != key->btree_id ||
32 !bpos_eq(ck->key.pos, key->pos);
35 static const struct rhashtable_params bch2_btree_key_cache_params = {
36 .head_offset = offsetof(struct bkey_cached, hash),
37 .key_offset = offsetof(struct bkey_cached, key),
38 .key_len = sizeof(struct bkey_cached_key),
39 .obj_cmpfn = bch2_btree_key_cache_cmp_fn,
43 inline struct bkey_cached *
44 bch2_btree_key_cache_find(struct bch_fs *c, enum btree_id btree_id, struct bpos pos)
46 struct bkey_cached_key key = {
51 return rhashtable_lookup_fast(&c->btree_key_cache.table, &key,
52 bch2_btree_key_cache_params);
55 static bool bkey_cached_lock_for_evict(struct bkey_cached *ck)
57 if (!six_trylock_intent(&ck->c.lock))
60 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
61 six_unlock_intent(&ck->c.lock);
65 if (!six_trylock_write(&ck->c.lock)) {
66 six_unlock_intent(&ck->c.lock);
73 static void bkey_cached_evict(struct btree_key_cache *c,
74 struct bkey_cached *ck)
76 BUG_ON(rhashtable_remove_fast(&c->table, &ck->hash,
77 bch2_btree_key_cache_params));
78 memset(&ck->key, ~0, sizeof(ck->key));
80 atomic_long_dec(&c->nr_keys);
83 static void bkey_cached_free(struct btree_key_cache *bc,
84 struct bkey_cached *ck)
86 struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
88 BUG_ON(test_bit(BKEY_CACHED_DIRTY, &ck->flags));
90 ck->btree_trans_barrier_seq =
91 start_poll_synchronize_srcu(&c->btree_trans_barrier);
93 if (ck->c.lock.readers) {
94 list_move_tail(&ck->list, &bc->freed_pcpu);
97 list_move_tail(&ck->list, &bc->freed_nonpcpu);
98 bc->nr_freed_nonpcpu++;
100 atomic_long_inc(&bc->nr_freed);
106 six_unlock_write(&ck->c.lock);
107 six_unlock_intent(&ck->c.lock);
111 static void __bkey_cached_move_to_freelist_ordered(struct btree_key_cache *bc,
112 struct bkey_cached *ck)
114 struct bkey_cached *pos;
116 bc->nr_freed_nonpcpu++;
118 list_for_each_entry_reverse(pos, &bc->freed_nonpcpu, list) {
119 if (ULONG_CMP_GE(ck->btree_trans_barrier_seq,
120 pos->btree_trans_barrier_seq)) {
121 list_move(&ck->list, &pos->list);
126 list_move(&ck->list, &bc->freed_nonpcpu);
130 static void bkey_cached_move_to_freelist(struct btree_key_cache *bc,
131 struct bkey_cached *ck)
133 BUG_ON(test_bit(BKEY_CACHED_DIRTY, &ck->flags));
135 if (!ck->c.lock.readers) {
137 struct btree_key_cache_freelist *f;
141 f = this_cpu_ptr(bc->pcpu_freed);
143 if (f->nr < ARRAY_SIZE(f->objs)) {
144 f->objs[f->nr++] = ck;
150 mutex_lock(&bc->lock);
152 f = this_cpu_ptr(bc->pcpu_freed);
154 while (f->nr > ARRAY_SIZE(f->objs) / 2) {
155 struct bkey_cached *ck2 = f->objs[--f->nr];
157 __bkey_cached_move_to_freelist_ordered(bc, ck2);
161 __bkey_cached_move_to_freelist_ordered(bc, ck);
162 mutex_unlock(&bc->lock);
165 mutex_lock(&bc->lock);
166 list_move_tail(&ck->list, &bc->freed_nonpcpu);
167 bc->nr_freed_nonpcpu++;
168 mutex_unlock(&bc->lock);
171 mutex_lock(&bc->lock);
172 list_move_tail(&ck->list, &bc->freed_pcpu);
173 mutex_unlock(&bc->lock);
177 static void bkey_cached_free_fast(struct btree_key_cache *bc,
178 struct bkey_cached *ck)
180 struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
182 ck->btree_trans_barrier_seq =
183 start_poll_synchronize_srcu(&c->btree_trans_barrier);
185 list_del_init(&ck->list);
186 atomic_long_inc(&bc->nr_freed);
192 bkey_cached_move_to_freelist(bc, ck);
194 six_unlock_write(&ck->c.lock);
195 six_unlock_intent(&ck->c.lock);
198 static struct bkey_cached *
199 bkey_cached_alloc(struct btree_trans *trans, struct btree_path *path,
202 struct bch_fs *c = trans->c;
203 struct btree_key_cache *bc = &c->btree_key_cache;
204 struct bkey_cached *ck = NULL;
205 bool pcpu_readers = btree_uses_pcpu_readers(path->btree_id);
210 struct btree_key_cache_freelist *f;
213 f = this_cpu_ptr(bc->pcpu_freed);
215 ck = f->objs[--f->nr];
219 mutex_lock(&bc->lock);
221 f = this_cpu_ptr(bc->pcpu_freed);
223 while (!list_empty(&bc->freed_nonpcpu) &&
224 f->nr < ARRAY_SIZE(f->objs) / 2) {
225 ck = list_last_entry(&bc->freed_nonpcpu, struct bkey_cached, list);
226 list_del_init(&ck->list);
227 bc->nr_freed_nonpcpu--;
228 f->objs[f->nr++] = ck;
231 ck = f->nr ? f->objs[--f->nr] : NULL;
233 mutex_unlock(&bc->lock);
236 mutex_lock(&bc->lock);
237 if (!list_empty(&bc->freed_nonpcpu)) {
238 ck = list_last_entry(&bc->freed_nonpcpu, struct bkey_cached, list);
239 list_del_init(&ck->list);
240 bc->nr_freed_nonpcpu--;
242 mutex_unlock(&bc->lock);
245 mutex_lock(&bc->lock);
246 if (!list_empty(&bc->freed_pcpu)) {
247 ck = list_last_entry(&bc->freed_pcpu, struct bkey_cached, list);
248 list_del_init(&ck->list);
250 mutex_unlock(&bc->lock);
254 ret = btree_node_lock_nopath(trans, &ck->c, SIX_LOCK_intent, _THIS_IP_);
256 bkey_cached_move_to_freelist(bc, ck);
260 path->l[0].b = (void *) ck;
261 path->l[0].lock_seq = six_lock_seq(&ck->c.lock);
262 mark_btree_node_locked(trans, path, 0, BTREE_NODE_INTENT_LOCKED);
264 ret = bch2_btree_node_lock_write(trans, path, &ck->c);
266 btree_node_unlock(trans, path, 0);
267 bkey_cached_move_to_freelist(bc, ck);
274 ck = allocate_dropping_locks(trans, ret,
275 kmem_cache_zalloc(bch2_key_cache, _gfp));
277 kmem_cache_free(bch2_key_cache, ck);
284 INIT_LIST_HEAD(&ck->list);
285 bch2_btree_lock_init(&ck->c, pcpu_readers ? SIX_LOCK_INIT_PCPU : 0);
288 BUG_ON(!six_trylock_intent(&ck->c.lock));
289 BUG_ON(!six_trylock_write(&ck->c.lock));
294 static struct bkey_cached *
295 bkey_cached_reuse(struct btree_key_cache *c)
297 struct bucket_table *tbl;
298 struct rhash_head *pos;
299 struct bkey_cached *ck;
302 mutex_lock(&c->lock);
304 tbl = rht_dereference_rcu(c->table.tbl, &c->table);
305 for (i = 0; i < tbl->size; i++)
306 rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
307 if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
308 bkey_cached_lock_for_evict(ck)) {
309 bkey_cached_evict(c, ck);
316 mutex_unlock(&c->lock);
320 static struct bkey_cached *
321 btree_key_cache_create(struct btree_trans *trans, struct btree_path *path)
323 struct bch_fs *c = trans->c;
324 struct btree_key_cache *bc = &c->btree_key_cache;
325 struct bkey_cached *ck;
326 bool was_new = false;
328 ck = bkey_cached_alloc(trans, path, &was_new);
333 ck = bkey_cached_reuse(bc);
335 bch_err(c, "error allocating memory for key cache item, btree %s",
336 bch2_btree_id_str(path->btree_id));
337 return ERR_PTR(-BCH_ERR_ENOMEM_btree_key_cache_create);
340 mark_btree_node_locked(trans, path, 0, BTREE_NODE_INTENT_LOCKED);
344 ck->c.btree_id = path->btree_id;
345 ck->key.btree_id = path->btree_id;
346 ck->key.pos = path->pos;
348 ck->flags = 1U << BKEY_CACHED_ACCESSED;
350 if (unlikely(rhashtable_lookup_insert_fast(&bc->table,
352 bch2_btree_key_cache_params))) {
353 /* We raced with another fill: */
355 if (likely(was_new)) {
356 six_unlock_write(&ck->c.lock);
357 six_unlock_intent(&ck->c.lock);
360 bkey_cached_free_fast(bc, ck);
363 mark_btree_node_locked(trans, path, 0, BTREE_NODE_UNLOCKED);
367 atomic_long_inc(&bc->nr_keys);
369 six_unlock_write(&ck->c.lock);
374 static int btree_key_cache_fill(struct btree_trans *trans,
375 struct btree_path *ck_path,
376 struct bkey_cached *ck)
378 struct btree_iter iter;
380 unsigned new_u64s = 0;
381 struct bkey_i *new_k = NULL;
384 k = bch2_bkey_get_iter(trans, &iter, ck->key.btree_id, ck->key.pos,
385 BTREE_ITER_KEY_CACHE_FILL|
386 BTREE_ITER_CACHED_NOFILL);
391 if (!bch2_btree_node_relock(trans, ck_path, 0)) {
392 trace_and_count(trans->c, trans_restart_relock_key_cache_fill, trans, _THIS_IP_, ck_path);
393 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_fill);
398 * bch2_varint_decode can read past the end of the buffer by at
399 * most 7 bytes (it won't be used):
401 new_u64s = k.k->u64s + 1;
404 * Allocate some extra space so that the transaction commit path is less
405 * likely to have to reallocate, since that requires a transaction
408 new_u64s = min(256U, (new_u64s * 3) / 2);
410 if (new_u64s > ck->u64s) {
411 new_u64s = roundup_pow_of_two(new_u64s);
412 new_k = kmalloc(new_u64s * sizeof(u64), GFP_NOWAIT|__GFP_NOWARN);
414 bch2_trans_unlock(trans);
416 new_k = kmalloc(new_u64s * sizeof(u64), GFP_KERNEL);
418 bch_err(trans->c, "error allocating memory for key cache key, btree %s u64s %u",
419 bch2_btree_id_str(ck->key.btree_id), new_u64s);
420 ret = -BCH_ERR_ENOMEM_btree_key_cache_fill;
424 if (!bch2_btree_node_relock(trans, ck_path, 0)) {
426 trace_and_count(trans->c, trans_restart_relock_key_cache_fill, trans, _THIS_IP_, ck_path);
427 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_fill);
431 ret = bch2_trans_relock(trans);
439 ret = bch2_btree_node_lock_write(trans, ck_path, &ck_path->l[0].b->c);
451 bkey_reassemble(ck->k, k);
453 bch2_btree_node_unlock_write(trans, ck_path, ck_path->l[0].b);
455 /* We're not likely to need this iterator again: */
456 set_btree_iter_dontneed(&iter);
458 bch2_trans_iter_exit(trans, &iter);
463 bch2_btree_path_traverse_cached_slowpath(struct btree_trans *trans, struct btree_path *path,
466 struct bch_fs *c = trans->c;
467 struct bkey_cached *ck;
474 if (bch2_btree_node_relock_notrace(trans, path, 0)) {
475 ck = (void *) path->l[0].b;
479 ck = bch2_btree_key_cache_find(c, path->btree_id, path->pos);
481 ck = btree_key_cache_create(trans, path);
482 ret = PTR_ERR_OR_ZERO(ck);
488 mark_btree_node_locked(trans, path, 0, BTREE_NODE_INTENT_LOCKED);
489 path->locks_want = 1;
491 enum six_lock_type lock_want = __btree_lock_want(path, 0);
493 ret = btree_node_lock(trans, path, (void *) ck, 0,
494 lock_want, _THIS_IP_);
495 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
500 if (ck->key.btree_id != path->btree_id ||
501 !bpos_eq(ck->key.pos, path->pos)) {
502 six_unlock_type(&ck->c.lock, lock_want);
506 mark_btree_node_locked(trans, path, 0,
507 (enum btree_node_locked_type) lock_want);
510 path->l[0].lock_seq = six_lock_seq(&ck->c.lock);
511 path->l[0].b = (void *) ck;
513 path->uptodate = BTREE_ITER_UPTODATE;
515 if (!ck->valid && !(flags & BTREE_ITER_CACHED_NOFILL)) {
517 * Using the underscore version because we haven't set
518 * path->uptodate yet:
520 if (!path->locks_want &&
521 !__bch2_btree_path_upgrade(trans, path, 1, NULL)) {
522 trace_and_count(trans->c, trans_restart_key_cache_upgrade, trans, _THIS_IP_);
523 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_upgrade);
527 ret = btree_key_cache_fill(trans, path, ck);
531 ret = bch2_btree_path_relock(trans, path, _THIS_IP_);
535 path->uptodate = BTREE_ITER_UPTODATE;
538 if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
539 set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
541 BUG_ON(btree_node_locked_type(path, 0) != btree_lock_want(path, 0));
542 BUG_ON(path->uptodate);
546 path->uptodate = BTREE_ITER_NEED_TRAVERSE;
547 if (!bch2_err_matches(ret, BCH_ERR_transaction_restart)) {
548 btree_node_unlock(trans, path, 0);
549 path->l[0].b = ERR_PTR(ret);
554 int bch2_btree_path_traverse_cached(struct btree_trans *trans, struct btree_path *path,
557 struct bch_fs *c = trans->c;
558 struct bkey_cached *ck;
561 EBUG_ON(path->level);
565 if (bch2_btree_node_relock_notrace(trans, path, 0)) {
566 ck = (void *) path->l[0].b;
570 ck = bch2_btree_key_cache_find(c, path->btree_id, path->pos);
572 return bch2_btree_path_traverse_cached_slowpath(trans, path, flags);
574 enum six_lock_type lock_want = __btree_lock_want(path, 0);
576 ret = btree_node_lock(trans, path, (void *) ck, 0,
577 lock_want, _THIS_IP_);
578 EBUG_ON(ret && !bch2_err_matches(ret, BCH_ERR_transaction_restart));
583 if (ck->key.btree_id != path->btree_id ||
584 !bpos_eq(ck->key.pos, path->pos)) {
585 six_unlock_type(&ck->c.lock, lock_want);
589 mark_btree_node_locked(trans, path, 0,
590 (enum btree_node_locked_type) lock_want);
593 path->l[0].lock_seq = six_lock_seq(&ck->c.lock);
594 path->l[0].b = (void *) ck;
597 return bch2_btree_path_traverse_cached_slowpath(trans, path, flags);
599 if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
600 set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
602 path->uptodate = BTREE_ITER_UPTODATE;
604 EBUG_ON(btree_node_locked_type(path, 0) != btree_lock_want(path, 0));
609 static int btree_key_cache_flush_pos(struct btree_trans *trans,
610 struct bkey_cached_key key,
612 unsigned commit_flags,
615 struct bch_fs *c = trans->c;
616 struct journal *j = &c->journal;
617 struct btree_iter c_iter, b_iter;
618 struct bkey_cached *ck = NULL;
621 bch2_trans_iter_init(trans, &b_iter, key.btree_id, key.pos,
624 BTREE_ITER_ALL_SNAPSHOTS);
625 bch2_trans_iter_init(trans, &c_iter, key.btree_id, key.pos,
628 b_iter.flags &= ~BTREE_ITER_WITH_KEY_CACHE;
630 ret = bch2_btree_iter_traverse(&c_iter);
634 ck = (void *) c_iter.path->l[0].b;
638 if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
646 if (journal_seq && ck->journal.seq != journal_seq)
649 trans->journal_res.seq = ck->journal.seq;
652 * If we're at the end of the journal, we really want to free up space
653 * in the journal right away - we don't want to pin that old journal
654 * sequence number with a new btree node write, we want to re-journal
657 if (ck->journal.seq == journal_last_seq(j))
658 commit_flags |= BCH_WATERMARK_reclaim;
660 commit_flags |= BCH_TRANS_COMMIT_no_journal_res;
662 ret = bch2_btree_iter_traverse(&b_iter) ?:
663 bch2_trans_update(trans, &b_iter, ck->k,
664 BTREE_UPDATE_KEY_CACHE_RECLAIM|
665 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
666 BTREE_TRIGGER_NORUN) ?:
667 bch2_trans_commit(trans, NULL, NULL,
668 BCH_TRANS_COMMIT_no_check_rw|
669 BCH_TRANS_COMMIT_no_enospc|
672 bch2_fs_fatal_err_on(ret &&
673 !bch2_err_matches(ret, BCH_ERR_transaction_restart) &&
674 !bch2_err_matches(ret, BCH_ERR_journal_reclaim_would_deadlock) &&
675 !bch2_journal_error(j), c,
676 "error flushing key cache: %s", bch2_err_str(ret));
680 bch2_journal_pin_drop(j, &ck->journal);
682 BUG_ON(!btree_node_locked(c_iter.path, 0));
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 struct btree_path *path2;
692 trans_for_each_path(trans, path2)
693 if (path2 != c_iter.path)
694 __bch2_btree_path_unlock(trans, path2);
696 bch2_btree_node_lock_write_nofail(trans, c_iter.path, &ck->c);
698 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
699 clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
700 atomic_long_dec(&c->btree_key_cache.nr_dirty);
703 mark_btree_node_locked_noreset(c_iter.path, 0, BTREE_NODE_UNLOCKED);
704 bkey_cached_evict(&c->btree_key_cache, ck);
705 bkey_cached_free_fast(&c->btree_key_cache, ck);
708 bch2_trans_iter_exit(trans, &b_iter);
709 bch2_trans_iter_exit(trans, &c_iter);
713 int bch2_btree_key_cache_journal_flush(struct journal *j,
714 struct journal_entry_pin *pin, u64 seq)
716 struct bch_fs *c = container_of(j, struct bch_fs, journal);
717 struct bkey_cached *ck =
718 container_of(pin, struct bkey_cached, journal);
719 struct bkey_cached_key key;
720 struct btree_trans *trans = bch2_trans_get(c);
721 int srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
724 btree_node_lock_nopath_nofail(trans, &ck->c, SIX_LOCK_read);
727 if (ck->journal.seq != seq ||
728 !test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
729 six_unlock_read(&ck->c.lock);
733 if (ck->seq != seq) {
734 bch2_journal_pin_update(&c->journal, ck->seq, &ck->journal,
735 bch2_btree_key_cache_journal_flush);
736 six_unlock_read(&ck->c.lock);
739 six_unlock_read(&ck->c.lock);
741 ret = commit_do(trans, NULL, NULL, 0,
742 btree_key_cache_flush_pos(trans, key, seq,
743 BCH_TRANS_COMMIT_journal_reclaim, false));
745 srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
747 bch2_trans_put(trans);
752 * Flush and evict a key from the key cache:
754 int bch2_btree_key_cache_flush(struct btree_trans *trans,
755 enum btree_id id, struct bpos pos)
757 struct bch_fs *c = trans->c;
758 struct bkey_cached_key key = { id, pos };
760 /* Fastpath - assume it won't be found: */
761 if (!bch2_btree_key_cache_find(c, id, pos))
764 return btree_key_cache_flush_pos(trans, key, 0, 0, true);
767 bool bch2_btree_insert_key_cached(struct btree_trans *trans,
769 struct btree_insert_entry *insert_entry)
771 struct bch_fs *c = trans->c;
772 struct bkey_cached *ck = (void *) insert_entry->path->l[0].b;
773 struct bkey_i *insert = insert_entry->k;
774 bool kick_reclaim = false;
776 BUG_ON(insert->k.u64s > ck->u64s);
778 bkey_copy(ck->k, insert);
781 if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
782 EBUG_ON(test_bit(BCH_FS_CLEAN_SHUTDOWN, &c->flags));
783 set_bit(BKEY_CACHED_DIRTY, &ck->flags);
784 atomic_long_inc(&c->btree_key_cache.nr_dirty);
786 if (bch2_nr_btree_keys_need_flush(c))
791 * To minimize lock contention, we only add the journal pin here and
792 * defer pin updates to the flush callback via ->seq. Be careful not to
793 * update ->seq on nojournal commits because we don't want to update the
794 * pin to a seq that doesn't include journal updates on disk. Otherwise
795 * we risk losing the update after a crash.
797 * The only exception is if the pin is not active in the first place. We
798 * have to add the pin because journal reclaim drives key cache
799 * flushing. The flush callback will not proceed unless ->seq matches
800 * the latest pin, so make sure it starts with a consistent value.
802 if (!(insert_entry->flags & BTREE_UPDATE_NOJOURNAL) ||
803 !journal_pin_active(&ck->journal)) {
804 ck->seq = trans->journal_res.seq;
806 bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
807 &ck->journal, bch2_btree_key_cache_journal_flush);
810 journal_reclaim_kick(&c->journal);
814 void bch2_btree_key_cache_drop(struct btree_trans *trans,
815 struct btree_path *path)
817 struct bch_fs *c = trans->c;
818 struct bkey_cached *ck = (void *) path->l[0].b;
823 * We just did an update to the btree, bypassing the key cache: the key
824 * cache key is now stale and must be dropped, even if dirty:
826 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
827 clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
828 atomic_long_dec(&c->btree_key_cache.nr_dirty);
829 bch2_journal_pin_drop(&c->journal, &ck->journal);
835 static unsigned long bch2_btree_key_cache_scan(struct shrinker *shrink,
836 struct shrink_control *sc)
838 struct bch_fs *c = shrink->private_data;
839 struct btree_key_cache *bc = &c->btree_key_cache;
840 struct bucket_table *tbl;
841 struct bkey_cached *ck, *t;
842 size_t scanned = 0, freed = 0, nr = sc->nr_to_scan;
843 unsigned start, flags;
846 mutex_lock(&bc->lock);
847 srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
848 flags = memalloc_nofs_save();
851 * Newest freed entries are at the end of the list - once we hit one
852 * that's too new to be freed, we can bail out:
854 scanned += bc->nr_freed_nonpcpu;
856 list_for_each_entry_safe(ck, t, &bc->freed_nonpcpu, list) {
857 if (!poll_state_synchronize_srcu(&c->btree_trans_barrier,
858 ck->btree_trans_barrier_seq))
862 six_lock_exit(&ck->c.lock);
863 kmem_cache_free(bch2_key_cache, ck);
864 atomic_long_dec(&bc->nr_freed);
866 bc->nr_freed_nonpcpu--;
872 scanned += bc->nr_freed_pcpu;
874 list_for_each_entry_safe(ck, t, &bc->freed_pcpu, list) {
875 if (!poll_state_synchronize_srcu(&c->btree_trans_barrier,
876 ck->btree_trans_barrier_seq))
880 six_lock_exit(&ck->c.lock);
881 kmem_cache_free(bch2_key_cache, ck);
882 atomic_long_dec(&bc->nr_freed);
891 tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
892 if (bc->shrink_iter >= tbl->size)
894 start = bc->shrink_iter;
897 struct rhash_head *pos, *next;
899 pos = rht_ptr_rcu(rht_bucket(tbl, bc->shrink_iter));
901 while (!rht_is_a_nulls(pos)) {
902 next = rht_dereference_bucket_rcu(pos->next, tbl, bc->shrink_iter);
903 ck = container_of(pos, struct bkey_cached, hash);
905 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags))
908 if (test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
909 clear_bit(BKEY_CACHED_ACCESSED, &ck->flags);
910 else if (bkey_cached_lock_for_evict(ck)) {
911 bkey_cached_evict(bc, ck);
912 bkey_cached_free(bc, ck);
923 if (bc->shrink_iter >= tbl->size)
925 } while (scanned < nr && bc->shrink_iter != start);
929 memalloc_nofs_restore(flags);
930 srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
931 mutex_unlock(&bc->lock);
936 static unsigned long bch2_btree_key_cache_count(struct shrinker *shrink,
937 struct shrink_control *sc)
939 struct bch_fs *c = shrink->private_data;
940 struct btree_key_cache *bc = &c->btree_key_cache;
941 long nr = atomic_long_read(&bc->nr_keys) -
942 atomic_long_read(&bc->nr_dirty);
947 void bch2_fs_btree_key_cache_exit(struct btree_key_cache *bc)
949 struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
950 struct bucket_table *tbl;
951 struct bkey_cached *ck, *n;
952 struct rhash_head *pos;
959 shrinker_free(bc->shrink);
961 mutex_lock(&bc->lock);
964 * The loop is needed to guard against racing with rehash:
966 while (atomic_long_read(&bc->nr_keys)) {
968 tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
970 for (i = 0; i < tbl->size; i++)
971 rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
972 bkey_cached_evict(bc, ck);
973 list_add(&ck->list, &items);
979 for_each_possible_cpu(cpu) {
980 struct btree_key_cache_freelist *f =
981 per_cpu_ptr(bc->pcpu_freed, cpu);
983 for (i = 0; i < f->nr; i++) {
985 list_add(&ck->list, &items);
990 BUG_ON(list_count_nodes(&bc->freed_pcpu) != bc->nr_freed_pcpu);
991 BUG_ON(list_count_nodes(&bc->freed_nonpcpu) != bc->nr_freed_nonpcpu);
993 list_splice(&bc->freed_pcpu, &items);
994 list_splice(&bc->freed_nonpcpu, &items);
996 mutex_unlock(&bc->lock);
998 list_for_each_entry_safe(ck, n, &items, list) {
1001 bch2_journal_pin_drop(&c->journal, &ck->journal);
1003 list_del(&ck->list);
1005 six_lock_exit(&ck->c.lock);
1006 kmem_cache_free(bch2_key_cache, ck);
1009 if (atomic_long_read(&bc->nr_dirty) &&
1010 !bch2_journal_error(&c->journal) &&
1011 test_bit(BCH_FS_WAS_RW, &c->flags))
1012 panic("btree key cache shutdown error: nr_dirty nonzero (%li)\n",
1013 atomic_long_read(&bc->nr_dirty));
1015 if (atomic_long_read(&bc->nr_keys))
1016 panic("btree key cache shutdown error: nr_keys nonzero (%li)\n",
1017 atomic_long_read(&bc->nr_keys));
1019 if (bc->table_init_done)
1020 rhashtable_destroy(&bc->table);
1022 free_percpu(bc->pcpu_freed);
1025 void bch2_fs_btree_key_cache_init_early(struct btree_key_cache *c)
1027 mutex_init(&c->lock);
1028 INIT_LIST_HEAD(&c->freed_pcpu);
1029 INIT_LIST_HEAD(&c->freed_nonpcpu);
1032 static void bch2_btree_key_cache_shrinker_to_text(struct seq_buf *s, struct shrinker *shrink)
1034 struct bch_fs *c = shrink->private_data;
1035 struct btree_key_cache *bc = &c->btree_key_cache;
1037 size_t buflen = seq_buf_get_buf(s, &cbuf);
1038 struct printbuf out = PRINTBUF_EXTERN(cbuf, buflen);
1040 bch2_btree_key_cache_to_text(&out, bc);
1041 seq_buf_commit(s, out.pos);
1044 int bch2_fs_btree_key_cache_init(struct btree_key_cache *bc)
1046 struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
1047 struct shrinker *shrink;
1050 bc->pcpu_freed = alloc_percpu(struct btree_key_cache_freelist);
1051 if (!bc->pcpu_freed)
1052 return -BCH_ERR_ENOMEM_fs_btree_cache_init;
1055 if (rhashtable_init(&bc->table, &bch2_btree_key_cache_params))
1056 return -BCH_ERR_ENOMEM_fs_btree_cache_init;
1058 bc->table_init_done = true;
1060 shrink = shrinker_alloc(0, "%s-btree_key_cache", c->name);
1062 return -BCH_ERR_ENOMEM_fs_btree_cache_init;
1063 bc->shrink = shrink;
1065 shrink->count_objects = bch2_btree_key_cache_count;
1066 shrink->scan_objects = bch2_btree_key_cache_scan;
1067 shrink->to_text = bch2_btree_key_cache_shrinker_to_text;
1068 shrink->private_data = c;
1069 shrinker_register(shrink);
1073 void bch2_btree_key_cache_to_text(struct printbuf *out, struct btree_key_cache *c)
1075 prt_printf(out, "nr_freed:\t%lu", atomic_long_read(&c->nr_freed));
1077 prt_printf(out, "nr_keys:\t%lu", atomic_long_read(&c->nr_keys));
1079 prt_printf(out, "nr_dirty:\t%lu", atomic_long_read(&c->nr_dirty));
1083 void bch2_btree_key_cache_exit(void)
1085 kmem_cache_destroy(bch2_key_cache);
1088 int __init bch2_btree_key_cache_init(void)
1090 bch2_key_cache = KMEM_CACHE(bkey_cached, SLAB_RECLAIM_ACCOUNT);
1091 if (!bch2_key_cache)