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);
96 list_move_tail(&ck->list, &bc->freed_nonpcpu);
97 atomic_long_inc(&bc->nr_freed);
103 six_unlock_write(&ck->c.lock);
104 six_unlock_intent(&ck->c.lock);
108 static void __bkey_cached_move_to_freelist_ordered(struct btree_key_cache *bc,
109 struct bkey_cached *ck)
111 struct bkey_cached *pos;
113 list_for_each_entry_reverse(pos, &bc->freed_nonpcpu, list) {
114 if (ULONG_CMP_GE(ck->btree_trans_barrier_seq,
115 pos->btree_trans_barrier_seq)) {
116 list_move(&ck->list, &pos->list);
121 list_move(&ck->list, &bc->freed_nonpcpu);
125 static void bkey_cached_move_to_freelist(struct btree_key_cache *bc,
126 struct bkey_cached *ck)
128 BUG_ON(test_bit(BKEY_CACHED_DIRTY, &ck->flags));
130 if (!ck->c.lock.readers) {
132 struct btree_key_cache_freelist *f;
136 f = this_cpu_ptr(bc->pcpu_freed);
138 if (f->nr < ARRAY_SIZE(f->objs)) {
139 f->objs[f->nr++] = ck;
145 mutex_lock(&bc->lock);
147 f = this_cpu_ptr(bc->pcpu_freed);
149 while (f->nr > ARRAY_SIZE(f->objs) / 2) {
150 struct bkey_cached *ck2 = f->objs[--f->nr];
152 __bkey_cached_move_to_freelist_ordered(bc, ck2);
156 __bkey_cached_move_to_freelist_ordered(bc, ck);
157 mutex_unlock(&bc->lock);
160 mutex_lock(&bc->lock);
161 list_move_tail(&ck->list, &bc->freed_nonpcpu);
162 mutex_unlock(&bc->lock);
165 mutex_lock(&bc->lock);
166 list_move_tail(&ck->list, &bc->freed_pcpu);
167 mutex_unlock(&bc->lock);
171 static void bkey_cached_free_fast(struct btree_key_cache *bc,
172 struct bkey_cached *ck)
174 struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
176 ck->btree_trans_barrier_seq =
177 start_poll_synchronize_srcu(&c->btree_trans_barrier);
179 list_del_init(&ck->list);
180 atomic_long_inc(&bc->nr_freed);
186 bkey_cached_move_to_freelist(bc, ck);
188 six_unlock_write(&ck->c.lock);
189 six_unlock_intent(&ck->c.lock);
192 static struct bkey_cached *
193 bkey_cached_alloc(struct btree_trans *trans, struct btree_path *path,
196 struct bch_fs *c = trans->c;
197 struct btree_key_cache *bc = &c->btree_key_cache;
198 struct bkey_cached *ck = NULL;
199 bool pcpu_readers = btree_uses_pcpu_readers(path->btree_id);
204 struct btree_key_cache_freelist *f;
207 f = this_cpu_ptr(bc->pcpu_freed);
209 ck = f->objs[--f->nr];
213 mutex_lock(&bc->lock);
215 f = this_cpu_ptr(bc->pcpu_freed);
217 while (!list_empty(&bc->freed_nonpcpu) &&
218 f->nr < ARRAY_SIZE(f->objs) / 2) {
219 ck = list_last_entry(&bc->freed_nonpcpu, struct bkey_cached, list);
220 list_del_init(&ck->list);
221 f->objs[f->nr++] = ck;
224 ck = f->nr ? f->objs[--f->nr] : NULL;
226 mutex_unlock(&bc->lock);
229 mutex_lock(&bc->lock);
230 if (!list_empty(&bc->freed_nonpcpu)) {
231 ck = list_last_entry(&bc->freed_nonpcpu, struct bkey_cached, list);
232 list_del_init(&ck->list);
234 mutex_unlock(&bc->lock);
237 mutex_lock(&bc->lock);
238 if (!list_empty(&bc->freed_pcpu)) {
239 ck = list_last_entry(&bc->freed_pcpu, struct bkey_cached, list);
240 list_del_init(&ck->list);
242 mutex_unlock(&bc->lock);
246 ret = btree_node_lock_nopath(trans, &ck->c, SIX_LOCK_intent, _THIS_IP_);
248 bkey_cached_move_to_freelist(bc, ck);
252 path->l[0].b = (void *) ck;
253 path->l[0].lock_seq = six_lock_seq(&ck->c.lock);
254 mark_btree_node_locked(trans, path, 0, BTREE_NODE_INTENT_LOCKED);
256 ret = bch2_btree_node_lock_write(trans, path, &ck->c);
258 btree_node_unlock(trans, path, 0);
259 bkey_cached_move_to_freelist(bc, ck);
266 ck = allocate_dropping_locks(trans, ret,
267 kmem_cache_zalloc(bch2_key_cache, _gfp));
269 kmem_cache_free(bch2_key_cache, ck);
276 INIT_LIST_HEAD(&ck->list);
277 bch2_btree_lock_init(&ck->c, pcpu_readers ? SIX_LOCK_INIT_PCPU : 0);
280 BUG_ON(!six_trylock_intent(&ck->c.lock));
281 BUG_ON(!six_trylock_write(&ck->c.lock));
286 static struct bkey_cached *
287 bkey_cached_reuse(struct btree_key_cache *c)
289 struct bucket_table *tbl;
290 struct rhash_head *pos;
291 struct bkey_cached *ck;
294 mutex_lock(&c->lock);
296 tbl = rht_dereference_rcu(c->table.tbl, &c->table);
297 for (i = 0; i < tbl->size; i++)
298 rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
299 if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
300 bkey_cached_lock_for_evict(ck)) {
301 bkey_cached_evict(c, ck);
308 mutex_unlock(&c->lock);
312 static struct bkey_cached *
313 btree_key_cache_create(struct btree_trans *trans, struct btree_path *path)
315 struct bch_fs *c = trans->c;
316 struct btree_key_cache *bc = &c->btree_key_cache;
317 struct bkey_cached *ck;
318 bool was_new = false;
320 ck = bkey_cached_alloc(trans, path, &was_new);
325 ck = bkey_cached_reuse(bc);
327 bch_err(c, "error allocating memory for key cache item, btree %s",
328 bch2_btree_id_str(path->btree_id));
329 return ERR_PTR(-BCH_ERR_ENOMEM_btree_key_cache_create);
332 mark_btree_node_locked(trans, path, 0, BTREE_NODE_INTENT_LOCKED);
336 ck->c.btree_id = path->btree_id;
337 ck->key.btree_id = path->btree_id;
338 ck->key.pos = path->pos;
340 ck->flags = 1U << BKEY_CACHED_ACCESSED;
342 if (unlikely(rhashtable_lookup_insert_fast(&bc->table,
344 bch2_btree_key_cache_params))) {
345 /* We raced with another fill: */
347 if (likely(was_new)) {
348 six_unlock_write(&ck->c.lock);
349 six_unlock_intent(&ck->c.lock);
352 bkey_cached_free_fast(bc, ck);
355 mark_btree_node_locked(trans, path, 0, BTREE_NODE_UNLOCKED);
359 atomic_long_inc(&bc->nr_keys);
361 six_unlock_write(&ck->c.lock);
366 static int btree_key_cache_fill(struct btree_trans *trans,
367 struct btree_path *ck_path,
368 struct bkey_cached *ck)
370 struct btree_iter iter;
372 unsigned new_u64s = 0;
373 struct bkey_i *new_k = NULL;
376 k = bch2_bkey_get_iter(trans, &iter, ck->key.btree_id, ck->key.pos,
377 BTREE_ITER_KEY_CACHE_FILL|
378 BTREE_ITER_CACHED_NOFILL);
383 if (!bch2_btree_node_relock(trans, ck_path, 0)) {
384 trace_and_count(trans->c, trans_restart_relock_key_cache_fill, trans, _THIS_IP_, ck_path);
385 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_fill);
390 * bch2_varint_decode can read past the end of the buffer by at
391 * most 7 bytes (it won't be used):
393 new_u64s = k.k->u64s + 1;
396 * Allocate some extra space so that the transaction commit path is less
397 * likely to have to reallocate, since that requires a transaction
400 new_u64s = min(256U, (new_u64s * 3) / 2);
402 if (new_u64s > ck->u64s) {
403 new_u64s = roundup_pow_of_two(new_u64s);
404 new_k = kmalloc(new_u64s * sizeof(u64), GFP_NOWAIT|__GFP_NOWARN);
406 bch2_trans_unlock(trans);
408 new_k = kmalloc(new_u64s * sizeof(u64), GFP_KERNEL);
410 bch_err(trans->c, "error allocating memory for key cache key, btree %s u64s %u",
411 bch2_btree_id_str(ck->key.btree_id), new_u64s);
412 ret = -BCH_ERR_ENOMEM_btree_key_cache_fill;
416 if (!bch2_btree_node_relock(trans, ck_path, 0)) {
418 trace_and_count(trans->c, trans_restart_relock_key_cache_fill, trans, _THIS_IP_, ck_path);
419 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_fill);
423 ret = bch2_trans_relock(trans);
431 ret = bch2_btree_node_lock_write(trans, ck_path, &ck_path->l[0].b->c);
443 bkey_reassemble(ck->k, k);
445 bch2_btree_node_unlock_write(trans, ck_path, ck_path->l[0].b);
447 /* We're not likely to need this iterator again: */
448 set_btree_iter_dontneed(&iter);
450 bch2_trans_iter_exit(trans, &iter);
455 bch2_btree_path_traverse_cached_slowpath(struct btree_trans *trans, struct btree_path *path,
458 struct bch_fs *c = trans->c;
459 struct bkey_cached *ck;
466 if (bch2_btree_node_relock_notrace(trans, path, 0)) {
467 ck = (void *) path->l[0].b;
471 ck = bch2_btree_key_cache_find(c, path->btree_id, path->pos);
473 ck = btree_key_cache_create(trans, path);
474 ret = PTR_ERR_OR_ZERO(ck);
480 mark_btree_node_locked(trans, path, 0, BTREE_NODE_INTENT_LOCKED);
481 path->locks_want = 1;
483 enum six_lock_type lock_want = __btree_lock_want(path, 0);
485 ret = btree_node_lock(trans, path, (void *) ck, 0,
486 lock_want, _THIS_IP_);
487 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
492 if (ck->key.btree_id != path->btree_id ||
493 !bpos_eq(ck->key.pos, path->pos)) {
494 six_unlock_type(&ck->c.lock, lock_want);
498 mark_btree_node_locked(trans, path, 0,
499 (enum btree_node_locked_type) lock_want);
502 path->l[0].lock_seq = six_lock_seq(&ck->c.lock);
503 path->l[0].b = (void *) ck;
505 path->uptodate = BTREE_ITER_UPTODATE;
507 if (!ck->valid && !(flags & BTREE_ITER_CACHED_NOFILL)) {
509 * Using the underscore version because we haven't set
510 * path->uptodate yet:
512 if (!path->locks_want &&
513 !__bch2_btree_path_upgrade(trans, path, 1, NULL)) {
514 trace_and_count(trans->c, trans_restart_key_cache_upgrade, trans, _THIS_IP_);
515 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_upgrade);
519 ret = btree_key_cache_fill(trans, path, ck);
523 ret = bch2_btree_path_relock(trans, path, _THIS_IP_);
527 path->uptodate = BTREE_ITER_UPTODATE;
530 if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
531 set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
533 BUG_ON(btree_node_locked_type(path, 0) != btree_lock_want(path, 0));
534 BUG_ON(path->uptodate);
538 path->uptodate = BTREE_ITER_NEED_TRAVERSE;
539 if (!bch2_err_matches(ret, BCH_ERR_transaction_restart)) {
540 btree_node_unlock(trans, path, 0);
541 path->l[0].b = ERR_PTR(ret);
546 int bch2_btree_path_traverse_cached(struct btree_trans *trans, struct btree_path *path,
549 struct bch_fs *c = trans->c;
550 struct bkey_cached *ck;
553 EBUG_ON(path->level);
557 if (bch2_btree_node_relock_notrace(trans, path, 0)) {
558 ck = (void *) path->l[0].b;
562 ck = bch2_btree_key_cache_find(c, path->btree_id, path->pos);
564 return bch2_btree_path_traverse_cached_slowpath(trans, path, flags);
566 enum six_lock_type lock_want = __btree_lock_want(path, 0);
568 ret = btree_node_lock(trans, path, (void *) ck, 0,
569 lock_want, _THIS_IP_);
570 EBUG_ON(ret && !bch2_err_matches(ret, BCH_ERR_transaction_restart));
575 if (ck->key.btree_id != path->btree_id ||
576 !bpos_eq(ck->key.pos, path->pos)) {
577 six_unlock_type(&ck->c.lock, lock_want);
581 mark_btree_node_locked(trans, path, 0,
582 (enum btree_node_locked_type) lock_want);
585 path->l[0].lock_seq = six_lock_seq(&ck->c.lock);
586 path->l[0].b = (void *) ck;
589 return bch2_btree_path_traverse_cached_slowpath(trans, path, flags);
591 if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
592 set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
594 path->uptodate = BTREE_ITER_UPTODATE;
596 EBUG_ON(btree_node_locked_type(path, 0) != btree_lock_want(path, 0));
601 static int btree_key_cache_flush_pos(struct btree_trans *trans,
602 struct bkey_cached_key key,
604 unsigned commit_flags,
607 struct bch_fs *c = trans->c;
608 struct journal *j = &c->journal;
609 struct btree_iter c_iter, b_iter;
610 struct bkey_cached *ck = NULL;
613 bch2_trans_iter_init(trans, &b_iter, key.btree_id, key.pos,
616 BTREE_ITER_ALL_SNAPSHOTS);
617 bch2_trans_iter_init(trans, &c_iter, key.btree_id, key.pos,
620 b_iter.flags &= ~BTREE_ITER_WITH_KEY_CACHE;
622 ret = bch2_btree_iter_traverse(&c_iter);
626 ck = (void *) c_iter.path->l[0].b;
630 if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
638 if (journal_seq && ck->journal.seq != journal_seq)
642 * Since journal reclaim depends on us making progress here, and the
643 * allocator/copygc depend on journal reclaim making progress, we need
644 * to be using alloc reserves:
646 ret = bch2_btree_iter_traverse(&b_iter) ?:
647 bch2_trans_update(trans, &b_iter, ck->k,
648 BTREE_UPDATE_KEY_CACHE_RECLAIM|
649 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
650 BTREE_TRIGGER_NORUN) ?:
651 bch2_trans_commit(trans, NULL, NULL,
652 BTREE_INSERT_NOCHECK_RW|
654 (ck->journal.seq == journal_last_seq(j)
655 ? BCH_WATERMARK_reclaim
659 bch2_fs_fatal_err_on(ret &&
660 !bch2_err_matches(ret, BCH_ERR_transaction_restart) &&
661 !bch2_err_matches(ret, BCH_ERR_journal_reclaim_would_deadlock) &&
662 !bch2_journal_error(j), c,
663 "error flushing key cache: %s", bch2_err_str(ret));
667 bch2_journal_pin_drop(j, &ck->journal);
668 bch2_journal_preres_put(j, &ck->res);
670 BUG_ON(!btree_node_locked(c_iter.path, 0));
673 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
674 clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
675 atomic_long_dec(&c->btree_key_cache.nr_dirty);
678 struct btree_path *path2;
680 trans_for_each_path(trans, path2)
681 if (path2 != c_iter.path)
682 __bch2_btree_path_unlock(trans, path2);
684 bch2_btree_node_lock_write_nofail(trans, c_iter.path, &ck->c);
686 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
687 clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
688 atomic_long_dec(&c->btree_key_cache.nr_dirty);
691 mark_btree_node_locked_noreset(c_iter.path, 0, BTREE_NODE_UNLOCKED);
692 bkey_cached_evict(&c->btree_key_cache, ck);
693 bkey_cached_free_fast(&c->btree_key_cache, ck);
696 bch2_trans_iter_exit(trans, &b_iter);
697 bch2_trans_iter_exit(trans, &c_iter);
701 int bch2_btree_key_cache_journal_flush(struct journal *j,
702 struct journal_entry_pin *pin, u64 seq)
704 struct bch_fs *c = container_of(j, struct bch_fs, journal);
705 struct bkey_cached *ck =
706 container_of(pin, struct bkey_cached, journal);
707 struct bkey_cached_key key;
708 struct btree_trans *trans = bch2_trans_get(c);
709 int srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
712 btree_node_lock_nopath_nofail(trans, &ck->c, SIX_LOCK_read);
715 if (ck->journal.seq != seq ||
716 !test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
717 six_unlock_read(&ck->c.lock);
721 if (ck->seq != seq) {
722 bch2_journal_pin_update(&c->journal, ck->seq, &ck->journal,
723 bch2_btree_key_cache_journal_flush);
724 six_unlock_read(&ck->c.lock);
727 six_unlock_read(&ck->c.lock);
729 ret = commit_do(trans, NULL, NULL, 0,
730 btree_key_cache_flush_pos(trans, key, seq,
731 BTREE_INSERT_JOURNAL_RECLAIM, false));
733 srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
735 bch2_trans_put(trans);
740 * Flush and evict a key from the key cache:
742 int bch2_btree_key_cache_flush(struct btree_trans *trans,
743 enum btree_id id, struct bpos pos)
745 struct bch_fs *c = trans->c;
746 struct bkey_cached_key key = { id, pos };
748 /* Fastpath - assume it won't be found: */
749 if (!bch2_btree_key_cache_find(c, id, pos))
752 return btree_key_cache_flush_pos(trans, key, 0, 0, true);
755 bool bch2_btree_insert_key_cached(struct btree_trans *trans,
757 struct btree_insert_entry *insert_entry)
759 struct bch_fs *c = trans->c;
760 struct bkey_cached *ck = (void *) insert_entry->path->l[0].b;
761 struct bkey_i *insert = insert_entry->k;
762 bool kick_reclaim = false;
764 BUG_ON(insert->k.u64s > ck->u64s);
766 if (likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY))) {
769 BUG_ON(jset_u64s(insert->k.u64s) > trans->journal_preres.u64s);
771 difference = jset_u64s(insert->k.u64s) - ck->res.u64s;
772 if (difference > 0) {
773 trans->journal_preres.u64s -= difference;
774 ck->res.u64s += difference;
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 = container_of(shrink, struct bch_fs,
839 btree_key_cache.shrink);
840 struct btree_key_cache *bc = &c->btree_key_cache;
841 struct bucket_table *tbl;
842 struct bkey_cached *ck, *t;
843 size_t scanned = 0, freed = 0, nr = sc->nr_to_scan;
844 unsigned start, flags;
847 mutex_lock(&bc->lock);
848 srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
849 flags = memalloc_nofs_save();
852 * Newest freed entries are at the end of the list - once we hit one
853 * that's too new to be freed, we can bail out:
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);
871 list_for_each_entry_safe(ck, t, &bc->freed_pcpu, list) {
872 if (!poll_state_synchronize_srcu(&c->btree_trans_barrier,
873 ck->btree_trans_barrier_seq))
877 six_lock_exit(&ck->c.lock);
878 kmem_cache_free(bch2_key_cache, ck);
879 atomic_long_dec(&bc->nr_freed);
888 tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
889 if (bc->shrink_iter >= tbl->size)
891 start = bc->shrink_iter;
894 struct rhash_head *pos, *next;
896 pos = rht_ptr_rcu(rht_bucket(tbl, bc->shrink_iter));
898 while (!rht_is_a_nulls(pos)) {
899 next = rht_dereference_bucket_rcu(pos->next, tbl, bc->shrink_iter);
900 ck = container_of(pos, struct bkey_cached, hash);
902 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags))
905 if (test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
906 clear_bit(BKEY_CACHED_ACCESSED, &ck->flags);
907 else if (bkey_cached_lock_for_evict(ck)) {
908 bkey_cached_evict(bc, ck);
909 bkey_cached_free(bc, ck);
920 if (bc->shrink_iter >= tbl->size)
922 } while (scanned < nr && bc->shrink_iter != start);
926 memalloc_nofs_restore(flags);
927 srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
928 mutex_unlock(&bc->lock);
933 static unsigned long bch2_btree_key_cache_count(struct shrinker *shrink,
934 struct shrink_control *sc)
936 struct bch_fs *c = container_of(shrink, struct bch_fs,
937 btree_key_cache.shrink);
938 struct btree_key_cache *bc = &c->btree_key_cache;
939 long nr = atomic_long_read(&bc->nr_keys) -
940 atomic_long_read(&bc->nr_dirty);
945 void bch2_fs_btree_key_cache_exit(struct btree_key_cache *bc)
947 struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
948 struct bucket_table *tbl;
949 struct bkey_cached *ck, *n;
950 struct rhash_head *pos;
957 unregister_shrinker(&bc->shrink);
959 mutex_lock(&bc->lock);
962 * The loop is needed to guard against racing with rehash:
964 while (atomic_long_read(&bc->nr_keys)) {
966 tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
968 for (i = 0; i < tbl->size; i++)
969 rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
970 bkey_cached_evict(bc, ck);
971 list_add(&ck->list, &items);
977 for_each_possible_cpu(cpu) {
978 struct btree_key_cache_freelist *f =
979 per_cpu_ptr(bc->pcpu_freed, cpu);
981 for (i = 0; i < f->nr; i++) {
983 list_add(&ck->list, &items);
988 list_splice(&bc->freed_pcpu, &items);
989 list_splice(&bc->freed_nonpcpu, &items);
991 mutex_unlock(&bc->lock);
993 list_for_each_entry_safe(ck, n, &items, list) {
996 bch2_journal_pin_drop(&c->journal, &ck->journal);
997 bch2_journal_preres_put(&c->journal, &ck->res);
1001 six_lock_exit(&ck->c.lock);
1002 kmem_cache_free(bch2_key_cache, ck);
1005 if (atomic_long_read(&bc->nr_dirty) &&
1006 !bch2_journal_error(&c->journal) &&
1007 test_bit(BCH_FS_WAS_RW, &c->flags))
1008 panic("btree key cache shutdown error: nr_dirty nonzero (%li)\n",
1009 atomic_long_read(&bc->nr_dirty));
1011 if (atomic_long_read(&bc->nr_keys))
1012 panic("btree key cache shutdown error: nr_keys nonzero (%li)\n",
1013 atomic_long_read(&bc->nr_keys));
1015 if (bc->table_init_done)
1016 rhashtable_destroy(&bc->table);
1018 free_percpu(bc->pcpu_freed);
1021 void bch2_fs_btree_key_cache_init_early(struct btree_key_cache *c)
1023 mutex_init(&c->lock);
1024 INIT_LIST_HEAD(&c->freed_pcpu);
1025 INIT_LIST_HEAD(&c->freed_nonpcpu);
1028 static void bch2_btree_key_cache_shrinker_to_text(struct seq_buf *s, struct shrinker *shrink)
1030 struct btree_key_cache *bc =
1031 container_of(shrink, struct btree_key_cache, shrink);
1033 size_t buflen = seq_buf_get_buf(s, &cbuf);
1034 struct printbuf out = PRINTBUF_EXTERN(cbuf, buflen);
1036 bch2_btree_key_cache_to_text(&out, bc);
1037 seq_buf_commit(s, out.pos);
1040 int bch2_fs_btree_key_cache_init(struct btree_key_cache *bc)
1042 struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
1045 bc->pcpu_freed = alloc_percpu(struct btree_key_cache_freelist);
1046 if (!bc->pcpu_freed)
1047 return -BCH_ERR_ENOMEM_fs_btree_cache_init;
1050 if (rhashtable_init(&bc->table, &bch2_btree_key_cache_params))
1051 return -BCH_ERR_ENOMEM_fs_btree_cache_init;
1053 bc->table_init_done = true;
1055 bc->shrink.seeks = 0;
1056 bc->shrink.count_objects = bch2_btree_key_cache_count;
1057 bc->shrink.scan_objects = bch2_btree_key_cache_scan;
1058 bc->shrink.to_text = bch2_btree_key_cache_shrinker_to_text;
1059 if (register_shrinker(&bc->shrink, "%s-btree_key_cache", c->name))
1060 return -BCH_ERR_ENOMEM_fs_btree_cache_init;
1064 void bch2_btree_key_cache_to_text(struct printbuf *out, struct btree_key_cache *c)
1066 prt_printf(out, "nr_freed:\t%lu", atomic_long_read(&c->nr_freed));
1068 prt_printf(out, "nr_keys:\t%lu", atomic_long_read(&c->nr_keys));
1070 prt_printf(out, "nr_dirty:\t%lu", atomic_long_read(&c->nr_dirty));
1074 void bch2_btree_key_cache_exit(void)
1076 kmem_cache_destroy(bch2_key_cache);
1079 int __init bch2_btree_key_cache_init(void)
1081 bch2_key_cache = KMEM_CACHE(bkey_cached, SLAB_RECLAIM_ACCOUNT);
1082 if (!bch2_key_cache)