#include <linux/sched/mm.h>
#include <trace/events/bcachefs.h>
+static struct kmem_cache *bch2_key_cache;
+
static int bch2_btree_key_cache_cmp_fn(struct rhashtable_compare_arg *arg,
const void *obj)
{
const struct bkey_cached_key *key = arg->key;
return cmp_int(ck->key.btree_id, key->btree_id) ?:
- bkey_cmp(ck->key.pos, key->pos);
+ bpos_cmp(ck->key.pos, key->pos);
}
static const struct rhashtable_params bch2_btree_key_cache_params = {
bch2_btree_key_cache_params));
memset(&ck->key, ~0, sizeof(ck->key));
- c->nr_keys--;
+ atomic_long_dec(&c->nr_keys);
}
static void bkey_cached_free(struct btree_key_cache *bc,
{
struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
+ BUG_ON(test_bit(BKEY_CACHED_DIRTY, &ck->flags));
+
ck->btree_trans_barrier_seq =
start_poll_synchronize_srcu(&c->btree_trans_barrier);
- list_move(&ck->list, &bc->freed);
+ list_move_tail(&ck->list, &bc->freed);
+ bc->nr_freed++;
kfree(ck->k);
ck->k = NULL;
{
struct bkey_cached *ck;
- list_for_each_entry(ck, &c->freed, list)
- if (bkey_cached_lock_for_evict(ck))
- return ck;
+ ck = kmem_cache_alloc(bch2_key_cache, GFP_NOFS|__GFP_ZERO);
+ if (likely(ck)) {
+ INIT_LIST_HEAD(&ck->list);
+ six_lock_init(&ck->c.lock);
+ BUG_ON(!six_trylock_intent(&ck->c.lock));
+ BUG_ON(!six_trylock_write(&ck->c.lock));
+ return ck;
+ }
+
+ return NULL;
+}
+
+static struct bkey_cached *
+bkey_cached_reuse(struct btree_key_cache *c)
+{
+ struct bucket_table *tbl;
+ struct rhash_head *pos;
+ struct bkey_cached *ck;
+ unsigned i;
- list_for_each_entry(ck, &c->clean, list)
+ mutex_lock(&c->lock);
+ list_for_each_entry_reverse(ck, &c->freed, list)
if (bkey_cached_lock_for_evict(ck)) {
- bkey_cached_evict(c, ck);
+ c->nr_freed--;
+ list_del(&ck->list);
+ mutex_unlock(&c->lock);
return ck;
}
+ mutex_unlock(&c->lock);
- ck = kzalloc(sizeof(*ck), GFP_NOFS);
- if (!ck)
- return NULL;
-
- INIT_LIST_HEAD(&ck->list);
- six_lock_init(&ck->c.lock);
- BUG_ON(!six_trylock_intent(&ck->c.lock));
- BUG_ON(!six_trylock_write(&ck->c.lock));
+ rcu_read_lock();
+ tbl = rht_dereference_rcu(c->table.tbl, &c->table);
+ for (i = 0; i < tbl->size; i++)
+ rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
+ if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
+ bkey_cached_lock_for_evict(ck)) {
+ bkey_cached_evict(c, ck);
+ rcu_read_unlock();
+ return ck;
+ }
+ }
+ rcu_read_unlock();
- return ck;
+ return NULL;
}
static struct bkey_cached *
-btree_key_cache_create(struct btree_key_cache *c,
+btree_key_cache_create(struct bch_fs *c,
enum btree_id btree_id,
struct bpos pos)
{
+ struct btree_key_cache *bc = &c->btree_key_cache;
struct bkey_cached *ck;
+ bool was_new = true;
- ck = bkey_cached_alloc(c);
- if (!ck)
- return ERR_PTR(-ENOMEM);
+ ck = bkey_cached_alloc(bc);
+
+ if (unlikely(!ck)) {
+ ck = bkey_cached_reuse(bc);
+ if (unlikely(!ck)) {
+ bch_err(c, "error allocating memory for key cache item, btree %s",
+ bch2_btree_ids[btree_id]);
+ return ERR_PTR(-ENOMEM);
+ }
+
+ was_new = false;
+ } else {
+ if (btree_id == BTREE_ID_subvolumes)
+ six_lock_pcpu_alloc(&ck->c.lock);
+ else
+ six_lock_pcpu_free(&ck->c.lock);
+ }
ck->c.level = 0;
ck->c.btree_id = btree_id;
ck->key.btree_id = btree_id;
ck->key.pos = pos;
ck->valid = false;
+ ck->flags = 1U << BKEY_CACHED_ACCESSED;
- BUG_ON(ck->flags);
-
- if (rhashtable_lookup_insert_fast(&c->table,
+ if (unlikely(rhashtable_lookup_insert_fast(&bc->table,
&ck->hash,
- bch2_btree_key_cache_params)) {
+ bch2_btree_key_cache_params))) {
/* We raced with another fill: */
- bkey_cached_free(c, ck);
+
+ if (likely(was_new)) {
+ six_unlock_write(&ck->c.lock);
+ six_unlock_intent(&ck->c.lock);
+ kfree(ck);
+ } else {
+ mutex_lock(&bc->lock);
+ bkey_cached_free(bc, ck);
+ mutex_unlock(&bc->lock);
+ }
+
return NULL;
}
- c->nr_keys++;
+ atomic_long_inc(&bc->nr_keys);
- list_move(&ck->list, &c->clean);
six_unlock_write(&ck->c.lock);
return ck;
}
static int btree_key_cache_fill(struct btree_trans *trans,
- struct btree_iter *ck_iter,
+ struct btree_path *ck_path,
struct bkey_cached *ck)
{
- struct btree_iter *iter;
+ struct btree_path *path;
struct bkey_s_c k;
unsigned new_u64s = 0;
struct bkey_i *new_k = NULL;
+ struct bkey u;
int ret;
- iter = bch2_trans_get_iter(trans, ck->key.btree_id,
- ck->key.pos, BTREE_ITER_SLOTS);
- if (IS_ERR(iter))
- return PTR_ERR(iter);
+ path = bch2_path_get(trans, ck->key.btree_id,
+ ck->key.pos, 0, 0, 0, _THIS_IP_);
+ ret = bch2_btree_path_traverse(trans, path, 0);
+ if (ret)
+ goto err;
- k = bch2_btree_iter_peek_slot(iter);
- ret = bkey_err(k);
- if (ret) {
- bch2_trans_iter_put(trans, iter);
- return ret;
- }
+ k = bch2_btree_path_peek_slot(path, &u);
- if (!bch2_btree_node_relock(ck_iter, 0)) {
- bch2_trans_iter_put(trans, iter);
- trace_transaction_restart_ip(trans->ip, _THIS_IP_);
- return -EINTR;
+ if (!bch2_btree_node_relock(trans, ck_path, 0)) {
+ trace_trans_restart_relock_key_cache_fill(trans->fn,
+ _THIS_IP_, ck_path->btree_id, &ck_path->pos);
+ ret = btree_trans_restart(trans);
+ goto err;
}
- if (k.k->u64s > ck->u64s) {
- new_u64s = roundup_pow_of_two(k.k->u64s);
+ /*
+ * bch2_varint_decode can read past the end of the buffer by at
+ * most 7 bytes (it won't be used):
+ */
+ new_u64s = k.k->u64s + 1;
+
+ /*
+ * Allocate some extra space so that the transaction commit path is less
+ * likely to have to reallocate, since that requires a transaction
+ * restart:
+ */
+ new_u64s = min(256U, (new_u64s * 3) / 2);
+
+ if (new_u64s > ck->u64s) {
+ new_u64s = roundup_pow_of_two(new_u64s);
new_k = kmalloc(new_u64s * sizeof(u64), GFP_NOFS);
if (!new_k) {
- bch2_trans_iter_put(trans, iter);
- return -ENOMEM;
+ bch_err(trans->c, "error allocating memory for key cache key, btree %s u64s %u",
+ bch2_btree_ids[ck->key.btree_id], new_u64s);
+ ret = -ENOMEM;
+ goto err;
}
}
- bch2_btree_node_lock_write(ck_iter->l[0].b, ck_iter);
+ /*
+ * XXX: not allowed to be holding read locks when we take a write lock,
+ * currently
+ */
+ bch2_btree_node_lock_write(trans, ck_path, ck_path->l[0].b);
if (new_k) {
kfree(ck->k);
ck->u64s = new_u64s;
bkey_reassemble(ck->k, k);
ck->valid = true;
- bch2_btree_node_unlock_write(ck_iter->l[0].b, ck_iter);
+ bch2_btree_node_unlock_write(trans, ck_path, ck_path->l[0].b);
/* We're not likely to need this iterator again: */
- bch2_trans_iter_free(trans, iter);
-
- return 0;
+ path->preserve = false;
+err:
+ bch2_path_put(trans, path, 0);
+ return ret;
}
static int bkey_cached_check_fn(struct six_lock *lock, void *p)
{
struct bkey_cached *ck = container_of(lock, struct bkey_cached, c.lock);
- const struct btree_iter *iter = p;
+ const struct btree_path *path = p;
- return ck->key.btree_id == iter->btree_id &&
- !bkey_cmp(ck->key.pos, iter->pos) ? 0 : -1;
+ return ck->key.btree_id == path->btree_id &&
+ !bpos_cmp(ck->key.pos, path->pos) ? 0 : -1;
}
__flatten
-int bch2_btree_iter_traverse_cached(struct btree_iter *iter)
+int bch2_btree_path_traverse_cached(struct btree_trans *trans, struct btree_path *path,
+ unsigned flags)
{
- struct btree_trans *trans = iter->trans;
struct bch_fs *c = trans->c;
struct bkey_cached *ck;
int ret = 0;
- BUG_ON(iter->level);
+ BUG_ON(path->level);
- if (btree_node_locked(iter, 0)) {
- ck = (void *) iter->l[0].b;
+ path->l[1].b = NULL;
+
+ if (bch2_btree_node_relock(trans, path, 0)) {
+ ck = (void *) path->l[0].b;
goto fill;
}
retry:
- ck = bch2_btree_key_cache_find(c, iter->btree_id, iter->pos);
+ ck = bch2_btree_key_cache_find(c, path->btree_id, path->pos);
if (!ck) {
- if (iter->flags & BTREE_ITER_CACHED_NOCREATE) {
- iter->l[0].b = NULL;
+ if (flags & BTREE_ITER_CACHED_NOCREATE) {
+ path->l[0].b = NULL;
return 0;
}
- mutex_lock(&c->btree_key_cache.lock);
- ck = btree_key_cache_create(&c->btree_key_cache,
- iter->btree_id, iter->pos);
- mutex_unlock(&c->btree_key_cache.lock);
-
+ ck = btree_key_cache_create(c, path->btree_id, path->pos);
ret = PTR_ERR_OR_ZERO(ck);
if (ret)
goto err;
if (!ck)
goto retry;
- mark_btree_node_locked(iter, 0, SIX_LOCK_intent);
- iter->locks_want = 1;
+ mark_btree_node_locked(trans, path, 0, SIX_LOCK_intent);
+ path->locks_want = 1;
} else {
- enum six_lock_type lock_want = __btree_lock_want(iter, 0);
+ enum six_lock_type lock_want = __btree_lock_want(path, 0);
- if (!btree_node_lock((void *) ck, iter->pos, 0, iter, lock_want,
- bkey_cached_check_fn, iter, _THIS_IP_)) {
- if (ck->key.btree_id != iter->btree_id ||
- bkey_cmp(ck->key.pos, iter->pos)) {
+ if (!btree_node_lock(trans, path, (void *) ck, path->pos, 0,
+ lock_want,
+ bkey_cached_check_fn, path, _THIS_IP_)) {
+ if (!trans->restarted)
goto retry;
- }
- trace_transaction_restart_ip(trans->ip, _THIS_IP_);
ret = -EINTR;
goto err;
}
- if (ck->key.btree_id != iter->btree_id ||
- bkey_cmp(ck->key.pos, iter->pos)) {
+ if (ck->key.btree_id != path->btree_id ||
+ bpos_cmp(ck->key.pos, path->pos)) {
six_unlock_type(&ck->c.lock, lock_want);
goto retry;
}
- mark_btree_node_locked(iter, 0, lock_want);
+ mark_btree_node_locked(trans, path, 0, lock_want);
}
- iter->l[0].lock_seq = ck->c.lock.state.seq;
- iter->l[0].b = (void *) ck;
+ path->l[0].lock_seq = ck->c.lock.state.seq;
+ path->l[0].b = (void *) ck;
fill:
- if (!ck->valid && !(iter->flags & BTREE_ITER_CACHED_NOFILL)) {
- if (!btree_node_intent_locked(iter, 0))
- bch2_btree_iter_upgrade(iter, 1);
- if (!btree_node_intent_locked(iter, 0)) {
- trace_transaction_restart_ip(trans->ip, _THIS_IP_);
- ret = -EINTR;
+ if (!ck->valid && !(flags & BTREE_ITER_CACHED_NOFILL)) {
+ if (!path->locks_want &&
+ !__bch2_btree_path_upgrade(trans, path, 1)) {
+ trace_transaction_restart_ip(trans->fn, _THIS_IP_);
+ ret = btree_trans_restart(trans);
goto err;
}
- ret = btree_key_cache_fill(trans, iter, ck);
+ ret = btree_key_cache_fill(trans, path, ck);
if (ret)
goto err;
}
- iter->uptodate = BTREE_ITER_NEED_PEEK;
- bch2_btree_iter_downgrade(iter);
+ if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
+ set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
+
+ path->uptodate = BTREE_ITER_UPTODATE;
+ BUG_ON(btree_node_locked_type(path, 0) != btree_lock_want(path, 0));
+
return ret;
err:
if (ret != -EINTR) {
- btree_node_unlock(iter, 0);
- iter->flags |= BTREE_ITER_ERROR;
- iter->l[0].b = BTREE_ITER_NO_NODE_ERROR;
+ btree_node_unlock(path, 0);
+ path->l[0].b = BTREE_ITER_NO_NODE_ERROR;
}
return ret;
}
static int btree_key_cache_flush_pos(struct btree_trans *trans,
struct bkey_cached_key key,
u64 journal_seq,
+ unsigned commit_flags,
bool evict)
{
struct bch_fs *c = trans->c;
struct journal *j = &c->journal;
- struct btree_iter *c_iter = NULL, *b_iter = NULL;
- struct bkey_cached *ck;
+ struct btree_iter c_iter, b_iter;
+ struct bkey_cached *ck = NULL;
int ret;
- b_iter = bch2_trans_get_iter(trans, key.btree_id, key.pos,
- BTREE_ITER_SLOTS|
- BTREE_ITER_INTENT);
- ret = PTR_ERR_OR_ZERO(b_iter);
+ bch2_trans_iter_init(trans, &b_iter, key.btree_id, key.pos,
+ BTREE_ITER_SLOTS|
+ BTREE_ITER_INTENT|
+ BTREE_ITER_ALL_SNAPSHOTS);
+ bch2_trans_iter_init(trans, &c_iter, key.btree_id, key.pos,
+ BTREE_ITER_CACHED|
+ BTREE_ITER_CACHED_NOFILL|
+ BTREE_ITER_CACHED_NOCREATE|
+ BTREE_ITER_INTENT);
+ b_iter.flags &= ~BTREE_ITER_WITH_KEY_CACHE;
+
+ ret = bch2_btree_iter_traverse(&c_iter);
if (ret)
goto out;
- c_iter = bch2_trans_get_iter(trans, key.btree_id, key.pos,
- BTREE_ITER_CACHED|
- BTREE_ITER_CACHED_NOFILL|
- BTREE_ITER_CACHED_NOCREATE|
- BTREE_ITER_INTENT);
- ret = PTR_ERR_OR_ZERO(c_iter);
- if (ret)
- goto out;
-retry:
- ret = bch2_btree_iter_traverse(c_iter);
- if (ret)
- goto err;
-
- ck = (void *) c_iter->l[0].b;
- if (!ck ||
- (journal_seq && ck->journal.seq != journal_seq))
+ ck = (void *) c_iter.path->l[0].b;
+ if (!ck)
goto out;
if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
- if (!evict)
- goto out;
- goto evict;
+ if (evict)
+ goto evict;
+ goto out;
}
- ret = bch2_btree_iter_traverse(b_iter) ?:
- bch2_trans_update(trans, b_iter, ck->k, BTREE_TRIGGER_NORUN) ?:
+ BUG_ON(!ck->valid);
+
+ if (journal_seq && ck->journal.seq != journal_seq)
+ goto out;
+
+ /*
+ * Since journal reclaim depends on us making progress here, and the
+ * allocator/copygc depend on journal reclaim making progress, we need
+ * to be using alloc reserves:
+ * */
+ ret = bch2_btree_iter_traverse(&b_iter) ?:
+ bch2_trans_update(trans, &b_iter, ck->k,
+ BTREE_UPDATE_KEY_CACHE_RECLAIM|
+ BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
+ BTREE_TRIGGER_NORUN) ?:
bch2_trans_commit(trans, NULL, NULL,
- BTREE_INSERT_NOUNLOCK|
BTREE_INSERT_NOCHECK_RW|
BTREE_INSERT_NOFAIL|
BTREE_INSERT_USE_RESERVE|
- BTREE_INSERT_USE_ALLOC_RESERVE|
- BTREE_INSERT_JOURNAL_RESERVED|
- BTREE_INSERT_JOURNAL_RECLAIM);
-err:
- if (ret == -EINTR)
- goto retry;
-
- BUG_ON(ret && !bch2_journal_error(j));
-
- if (ret)
+ (ck->journal.seq == journal_last_seq(j)
+ ? JOURNAL_WATERMARK_reserved
+ : 0)|
+ commit_flags);
+ if (ret) {
+ bch2_fs_fatal_err_on(ret != -EINTR &&
+ ret != -EAGAIN &&
+ !bch2_journal_error(j), c,
+ "error flushing key cache: %i", ret);
goto out;
+ }
bch2_journal_pin_drop(j, &ck->journal);
bch2_journal_preres_put(j, &ck->res);
+ BUG_ON(!btree_node_locked(c_iter.path, 0));
+
if (!evict) {
- mutex_lock(&c->btree_key_cache.lock);
if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
- c->btree_key_cache.nr_dirty--;
+ atomic_long_dec(&c->btree_key_cache.nr_dirty);
}
-
- list_move_tail(&ck->list, &c->btree_key_cache.clean);
- mutex_unlock(&c->btree_key_cache.lock);
} else {
evict:
- BUG_ON(!btree_node_intent_locked(c_iter, 0));
+ BUG_ON(!btree_node_intent_locked(c_iter.path, 0));
- mark_btree_node_unlocked(c_iter, 0);
- c_iter->l[0].b = NULL;
+ mark_btree_node_unlocked(c_iter.path, 0);
+ c_iter.path->l[0].b = NULL;
six_lock_write(&ck->c.lock, NULL, NULL);
- mutex_lock(&c->btree_key_cache.lock);
if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
- c->btree_key_cache.nr_dirty--;
+ atomic_long_dec(&c->btree_key_cache.nr_dirty);
}
bkey_cached_evict(&c->btree_key_cache, ck);
+
+ mutex_lock(&c->btree_key_cache.lock);
bkey_cached_free(&c->btree_key_cache, ck);
mutex_unlock(&c->btree_key_cache.lock);
}
out:
- bch2_trans_iter_put(trans, b_iter);
- bch2_trans_iter_put(trans, c_iter);
+ bch2_trans_iter_exit(trans, &b_iter);
+ bch2_trans_iter_exit(trans, &c_iter);
return ret;
}
-static void btree_key_cache_journal_flush(struct journal *j,
- struct journal_entry_pin *pin,
- u64 seq)
+int bch2_btree_key_cache_journal_flush(struct journal *j,
+ struct journal_entry_pin *pin, u64 seq)
{
struct bch_fs *c = container_of(j, struct bch_fs, journal);
struct bkey_cached *ck =
container_of(pin, struct bkey_cached, journal);
struct bkey_cached_key key;
- struct btree_trans trans;
+ int ret = 0;
int srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
}
six_unlock_read(&ck->c.lock);
- bch2_trans_init(&trans, c, 0, 0);
- btree_key_cache_flush_pos(&trans, key, seq, false);
- bch2_trans_exit(&trans);
+ ret = bch2_trans_do(c, NULL, NULL, 0,
+ btree_key_cache_flush_pos(&trans, key, seq,
+ BTREE_INSERT_JOURNAL_RECLAIM, false));
unlock:
srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
+
+ return ret;
}
/*
if (!bch2_btree_key_cache_find(c, id, pos))
return 0;
- return btree_key_cache_flush_pos(trans, key, 0, true);
+ return btree_key_cache_flush_pos(trans, key, 0, 0, true);
}
bool bch2_btree_insert_key_cached(struct btree_trans *trans,
- struct btree_iter *iter,
+ struct btree_path *path,
struct bkey_i *insert)
{
struct bch_fs *c = trans->c;
- struct bkey_cached *ck = (void *) iter->l[0].b;
+ struct bkey_cached *ck = (void *) path->l[0].b;
+ bool kick_reclaim = false;
BUG_ON(insert->u64s > ck->u64s);
ck->valid = true;
if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
- mutex_lock(&c->btree_key_cache.lock);
- list_move(&ck->list, &c->btree_key_cache.dirty);
-
set_bit(BKEY_CACHED_DIRTY, &ck->flags);
- c->btree_key_cache.nr_dirty++;
- mutex_unlock(&c->btree_key_cache.lock);
+ atomic_long_inc(&c->btree_key_cache.nr_dirty);
+
+ if (bch2_nr_btree_keys_need_flush(c))
+ kick_reclaim = true;
}
bch2_journal_pin_update(&c->journal, trans->journal_res.seq,
- &ck->journal, btree_key_cache_journal_flush);
- return true;
-}
+ &ck->journal, bch2_btree_key_cache_journal_flush);
-#ifdef CONFIG_BCACHEFS_DEBUG
-void bch2_btree_key_cache_verify_clean(struct btree_trans *trans,
- enum btree_id id, struct bpos pos)
-{
- BUG_ON(bch2_btree_key_cache_find(trans->c, id, pos));
+ if (kick_reclaim)
+ journal_reclaim_kick(&c->journal);
+ return true;
}
-#endif
static unsigned long bch2_btree_key_cache_scan(struct shrinker *shrink,
struct shrink_control *sc)
struct bch_fs *c = container_of(shrink, struct bch_fs,
btree_key_cache.shrink);
struct btree_key_cache *bc = &c->btree_key_cache;
+ struct bucket_table *tbl;
struct bkey_cached *ck, *t;
size_t scanned = 0, freed = 0, nr = sc->nr_to_scan;
- unsigned flags;
+ unsigned start, flags;
+ int srcu_idx;
/* Return -1 if we can't do anything right now */
if (sc->gfp_mask & __GFP_FS)
else if (!mutex_trylock(&bc->lock))
return -1;
+ srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
flags = memalloc_nofs_save();
+ /*
+ * Newest freed entries are at the end of the list - once we hit one
+ * that's too new to be freed, we can bail out:
+ */
list_for_each_entry_safe(ck, t, &bc->freed, list) {
+ if (!poll_state_synchronize_srcu(&c->btree_trans_barrier,
+ ck->btree_trans_barrier_seq))
+ break;
+
+ list_del(&ck->list);
+ kmem_cache_free(bch2_key_cache, ck);
+ bc->nr_freed--;
scanned++;
+ freed++;
+ }
- if (poll_state_synchronize_srcu(&c->btree_trans_barrier,
- ck->btree_trans_barrier_seq)) {
- list_del(&ck->list);
- kfree(ck);
- freed++;
- }
+ if (scanned >= nr)
+ goto out;
- if (scanned >= nr)
- goto out;
- }
+ rcu_read_lock();
+ tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
+ if (bc->shrink_iter >= tbl->size)
+ bc->shrink_iter = 0;
+ start = bc->shrink_iter;
- list_for_each_entry_safe(ck, t, &bc->clean, list) {
- scanned++;
+ do {
+ struct rhash_head *pos, *next;
- if (bkey_cached_lock_for_evict(ck)) {
- bkey_cached_evict(bc, ck);
- bkey_cached_free(bc, ck);
- }
+ pos = rht_ptr_rcu(rht_bucket(tbl, bc->shrink_iter));
+
+ while (!rht_is_a_nulls(pos)) {
+ next = rht_dereference_bucket_rcu(pos->next, tbl, bc->shrink_iter);
+ ck = container_of(pos, struct bkey_cached, hash);
- if (scanned >= nr) {
- if (&t->list != &bc->clean)
- list_move_tail(&bc->clean, &t->list);
- goto out;
+ if (test_bit(BKEY_CACHED_DIRTY, &ck->flags))
+ goto next;
+
+ if (test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
+ clear_bit(BKEY_CACHED_ACCESSED, &ck->flags);
+ else if (bkey_cached_lock_for_evict(ck)) {
+ bkey_cached_evict(bc, ck);
+ bkey_cached_free(bc, ck);
+ }
+
+ scanned++;
+ if (scanned >= nr)
+ break;
+next:
+ pos = next;
}
- }
+
+ bc->shrink_iter++;
+ if (bc->shrink_iter >= tbl->size)
+ bc->shrink_iter = 0;
+ } while (scanned < nr && bc->shrink_iter != start);
+
+ rcu_read_unlock();
out:
memalloc_nofs_restore(flags);
+ srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
mutex_unlock(&bc->lock);
return freed;
struct bch_fs *c = container_of(shrink, struct bch_fs,
btree_key_cache.shrink);
struct btree_key_cache *bc = &c->btree_key_cache;
+ long nr = atomic_long_read(&bc->nr_keys) -
+ atomic_long_read(&bc->nr_dirty);
- return bc->nr_keys;
+ return max(0L, nr);
}
void bch2_fs_btree_key_cache_exit(struct btree_key_cache *bc)
{
struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
+ struct bucket_table *tbl;
struct bkey_cached *ck, *n;
+ struct rhash_head *pos;
+ unsigned i;
if (bc->shrink.list.next)
unregister_shrinker(&bc->shrink);
mutex_lock(&bc->lock);
- list_splice(&bc->dirty, &bc->clean);
- list_for_each_entry_safe(ck, n, &bc->clean, list) {
+ rcu_read_lock();
+ tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
+ if (tbl)
+ for (i = 0; i < tbl->size; i++)
+ rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
+ bkey_cached_evict(bc, ck);
+ list_add(&ck->list, &bc->freed);
+ }
+ rcu_read_unlock();
+
+ list_for_each_entry_safe(ck, n, &bc->freed, list) {
+ cond_resched();
+
bch2_journal_pin_drop(&c->journal, &ck->journal);
bch2_journal_preres_put(&c->journal, &ck->res);
+ list_del(&ck->list);
kfree(ck->k);
- kfree(ck);
- bc->nr_keys--;
+ kmem_cache_free(bch2_key_cache, ck);
}
- BUG_ON(bc->nr_dirty && !bch2_journal_error(&c->journal));
- BUG_ON(bc->nr_keys);
+ BUG_ON(atomic_long_read(&bc->nr_dirty) &&
+ !bch2_journal_error(&c->journal) &&
+ test_bit(BCH_FS_WAS_RW, &c->flags));
+ BUG_ON(atomic_long_read(&bc->nr_keys));
- list_for_each_entry_safe(ck, n, &bc->freed, list)
- kfree(ck);
mutex_unlock(&bc->lock);
- rhashtable_destroy(&bc->table);
+ if (bc->table_init_done)
+ rhashtable_destroy(&bc->table);
}
void bch2_fs_btree_key_cache_init_early(struct btree_key_cache *c)
{
mutex_init(&c->lock);
INIT_LIST_HEAD(&c->freed);
- INIT_LIST_HEAD(&c->clean);
- INIT_LIST_HEAD(&c->dirty);
}
int bch2_fs_btree_key_cache_init(struct btree_key_cache *c)
{
+ int ret;
+
+ ret = rhashtable_init(&c->table, &bch2_btree_key_cache_params);
+ if (ret)
+ return ret;
+
+ c->table_init_done = true;
+
+ c->shrink.seeks = 1;
c->shrink.count_objects = bch2_btree_key_cache_count;
c->shrink.scan_objects = bch2_btree_key_cache_scan;
-
- return register_shrinker(&c->shrink) ?:
- rhashtable_init(&c->table, &bch2_btree_key_cache_params);
+ return register_shrinker(&c->shrink);
}
void bch2_btree_key_cache_to_text(struct printbuf *out, struct btree_key_cache *c)
{
- struct bucket_table *tbl;
- struct bkey_cached *ck;
- struct rhash_head *pos;
- size_t i;
+ pr_buf(out, "nr_freed:\t%zu\n", c->nr_freed);
+ pr_buf(out, "nr_keys:\t%lu\n", atomic_long_read(&c->nr_keys));
+ pr_buf(out, "nr_dirty:\t%lu\n", atomic_long_read(&c->nr_dirty));
+}
- mutex_lock(&c->lock);
- tbl = rht_dereference_rcu(c->table.tbl, &c->table);
+void bch2_btree_key_cache_exit(void)
+{
+ if (bch2_key_cache)
+ kmem_cache_destroy(bch2_key_cache);
+}
- for (i = 0; i < tbl->size; i++) {
- rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
- pr_buf(out, "%s:",
- bch2_btree_ids[ck->key.btree_id]);
- bch2_bpos_to_text(out, ck->key.pos);
+int __init bch2_btree_key_cache_init(void)
+{
+ bch2_key_cache = KMEM_CACHE(bkey_cached, 0);
+ if (!bch2_key_cache)
+ return -ENOMEM;
- if (test_bit(BKEY_CACHED_DIRTY, &ck->flags))
- pr_buf(out, " journal seq %llu", ck->journal.seq);
- pr_buf(out, "\n");
- }
- }
- mutex_unlock(&c->lock);
+ return 0;
}
projects / bcachefs-tools-debian / blobdiff