+// SPDX-License-Identifier: GPL-2.0
#include "bcachefs.h"
#include "btree_cache.h"
#include "journal_reclaim.h"
#include <linux/sched/mm.h>
+#include <linux/seq_buf.h>
#include <trace/events/bcachefs.h>
static inline bool btree_uses_pcpu_readers(enum btree_id id)
const struct bkey_cached *ck = obj;
const struct bkey_cached_key *key = arg->key;
- return cmp_int(ck->key.btree_id, key->btree_id) ?:
- bpos_cmp(ck->key.pos, key->pos);
+ return ck->key.btree_id != key->btree_id ||
+ !bpos_eq(ck->key.pos, key->pos);
}
static const struct rhashtable_params bch2_btree_key_cache_params = {
if (!six_trylock_intent(&ck->c.lock))
return false;
- if (!six_trylock_write(&ck->c.lock)) {
+ if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
six_unlock_intent(&ck->c.lock);
return false;
}
- if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
- six_unlock_write(&ck->c.lock);
+ if (!six_trylock_write(&ck->c.lock)) {
six_unlock_intent(&ck->c.lock);
return false;
}
six_unlock_intent(&ck->c.lock);
}
+#ifdef __KERNEL__
+static void __bkey_cached_move_to_freelist_ordered(struct btree_key_cache *bc,
+ struct bkey_cached *ck)
+{
+ struct bkey_cached *pos;
+
+ list_for_each_entry_reverse(pos, &bc->freed_nonpcpu, list) {
+ if (ULONG_CMP_GE(ck->btree_trans_barrier_seq,
+ pos->btree_trans_barrier_seq)) {
+ list_move(&ck->list, &pos->list);
+ return;
+ }
+ }
+
+ list_move(&ck->list, &bc->freed_nonpcpu);
+}
+#endif
+
static void bkey_cached_move_to_freelist(struct btree_key_cache *bc,
struct bkey_cached *ck)
{
- struct btree_key_cache_freelist *f;
- bool freed = false;
-
BUG_ON(test_bit(BKEY_CACHED_DIRTY, &ck->flags));
if (!ck->c.lock.readers) {
+#ifdef __KERNEL__
+ struct btree_key_cache_freelist *f;
+ bool freed = false;
+
preempt_disable();
f = this_cpu_ptr(bc->pcpu_freed);
while (f->nr > ARRAY_SIZE(f->objs) / 2) {
struct bkey_cached *ck2 = f->objs[--f->nr];
- list_move_tail(&ck2->list, &bc->freed_nonpcpu);
+ __bkey_cached_move_to_freelist_ordered(bc, ck2);
}
preempt_enable();
- list_move_tail(&ck->list, &bc->freed_nonpcpu);
+ __bkey_cached_move_to_freelist_ordered(bc, ck);
mutex_unlock(&bc->lock);
}
+#else
+ mutex_lock(&bc->lock);
+ list_move_tail(&ck->list, &bc->freed_nonpcpu);
+ mutex_unlock(&bc->lock);
+#endif
} else {
mutex_lock(&bc->lock);
list_move_tail(&ck->list, &bc->freed_pcpu);
}
static struct bkey_cached *
-bkey_cached_alloc(struct btree_trans *trans, struct btree_path *path)
+bkey_cached_alloc(struct btree_trans *trans, struct btree_path *path,
+ bool *was_new)
{
struct bch_fs *c = trans->c;
struct btree_key_cache *bc = &c->btree_key_cache;
struct bkey_cached *ck = NULL;
- struct btree_key_cache_freelist *f;
bool pcpu_readers = btree_uses_pcpu_readers(path->btree_id);
+ int ret;
if (!pcpu_readers) {
+#ifdef __KERNEL__
+ struct btree_key_cache_freelist *f;
+
preempt_disable();
f = this_cpu_ptr(bc->pcpu_freed);
if (f->nr)
preempt_enable();
mutex_unlock(&bc->lock);
}
+#else
+ mutex_lock(&bc->lock);
+ if (!list_empty(&bc->freed_nonpcpu)) {
+ ck = list_last_entry(&bc->freed_nonpcpu, struct bkey_cached, list);
+ list_del_init(&ck->list);
+ }
+ mutex_unlock(&bc->lock);
+#endif
} else {
mutex_lock(&bc->lock);
if (!list_empty(&bc->freed_pcpu)) {
if (ck) {
int ret;
- ret = btree_node_lock_nopath(trans, &ck->c, SIX_LOCK_intent);
+ ret = btree_node_lock_nopath(trans, &ck->c, SIX_LOCK_intent, _THIS_IP_);
if (unlikely(ret)) {
bkey_cached_move_to_freelist(bc, ck);
return ERR_PTR(ret);
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, "b->c.lock", &bch2_btree_node_lock_key);
- if (pcpu_readers)
- six_lock_pcpu_alloc(&ck->c.lock);
+ ck = kmem_cache_zalloc(bch2_key_cache, GFP_NOWAIT|__GFP_NOWARN);
+ if (likely(ck))
+ goto init;
- ck->c.cached = true;
- BUG_ON(!six_trylock_intent(&ck->c.lock));
- BUG_ON(!six_trylock_write(&ck->c.lock));
- return ck;
+ bch2_trans_unlock(trans);
+
+ ck = kmem_cache_zalloc(bch2_key_cache, GFP_KERNEL);
+
+ ret = bch2_trans_relock(trans);
+ if (ret) {
+ kmem_cache_free(bch2_key_cache, ck);
+ return ERR_PTR(ret);
}
- return NULL;
+ if (!ck)
+ return NULL;
+init:
+ INIT_LIST_HEAD(&ck->list);
+ bch2_btree_lock_init(&ck->c);
+ if (pcpu_readers)
+ six_lock_pcpu_alloc(&ck->c.lock);
+
+ ck->c.cached = true;
+ BUG_ON(!six_trylock_intent(&ck->c.lock));
+ BUG_ON(!six_trylock_write(&ck->c.lock));
+ *was_new = true;
+ return ck;
}
static struct bkey_cached *
struct bkey_cached *ck;
unsigned i;
+ mutex_lock(&c->lock);
rcu_read_lock();
tbl = rht_dereference_rcu(c->table.tbl, &c->table);
for (i = 0; i < tbl->size; i++)
if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
bkey_cached_lock_for_evict(ck)) {
bkey_cached_evict(c, ck);
- rcu_read_unlock();
- return ck;
+ goto out;
}
}
+ ck = NULL;
+out:
rcu_read_unlock();
-
- return NULL;
+ mutex_unlock(&c->lock);
+ return ck;
}
static struct bkey_cached *
struct bch_fs *c = trans->c;
struct btree_key_cache *bc = &c->btree_key_cache;
struct bkey_cached *ck;
- bool was_new = true;
+ bool was_new = false;
- ck = bkey_cached_alloc(trans, path);
- if (unlikely(IS_ERR(ck)))
+ ck = bkey_cached_alloc(trans, path, &was_new);
+ if (IS_ERR(ck))
return ck;
if (unlikely(!ck)) {
if (unlikely(!ck)) {
bch_err(c, "error allocating memory for key cache item, btree %s",
bch2_btree_ids[path->btree_id]);
- return ERR_PTR(-ENOMEM);
+ return ERR_PTR(-BCH_ERR_ENOMEM_btree_key_cache_create);
}
mark_btree_node_locked(trans, path, 0, SIX_LOCK_intent);
- was_new = false;
} else {
if (path->btree_id == BTREE_ID_subvolumes)
six_lock_pcpu_alloc(&ck->c.lock);
if (likely(was_new)) {
six_unlock_write(&ck->c.lock);
six_unlock_intent(&ck->c.lock);
- mark_btree_node_locked(trans, path, 0, BTREE_NODE_UNLOCKED);
kfree(ck);
} else {
bkey_cached_free_fast(bc, ck);
}
+ mark_btree_node_locked(trans, path, 0, BTREE_NODE_UNLOCKED);
return NULL;
}
struct btree_path *ck_path,
struct bkey_cached *ck)
{
- struct btree_path *path;
+ struct btree_iter iter;
struct bkey_s_c k;
unsigned new_u64s = 0;
struct bkey_i *new_k = NULL;
- struct bkey u;
int ret;
- 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);
+ bch2_trans_iter_init(trans, &iter, ck->key.btree_id, ck->key.pos,
+ BTREE_ITER_KEY_CACHE_FILL|
+ BTREE_ITER_CACHED_NOFILL);
+ k = bch2_btree_iter_peek_slot(&iter);
+ ret = bkey_err(k);
if (ret)
goto err;
- k = bch2_btree_path_peek_slot(path, &u);
-
if (!bch2_btree_node_relock(trans, ck_path, 0)) {
trace_and_count(trans->c, trans_restart_relock_key_cache_fill, trans, _THIS_IP_, ck_path);
- ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_raced);
+ ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_fill);
goto err;
}
if (new_u64s > ck->u64s) {
new_u64s = roundup_pow_of_two(new_u64s);
- new_k = kmalloc(new_u64s * sizeof(u64), GFP_NOFS);
+ new_k = kmalloc(new_u64s * sizeof(u64), GFP_NOWAIT|__GFP_NOWARN);
if (!new_k) {
- 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_trans_unlock(trans);
+
+ new_k = kmalloc(new_u64s * sizeof(u64), GFP_KERNEL);
+ if (!new_k) {
+ 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 = -BCH_ERR_ENOMEM_btree_key_cache_fill;
+ goto err;
+ }
+
+ if (!bch2_btree_node_relock(trans, ck_path, 0)) {
+ kfree(new_k);
+ trace_and_count(trans->c, trans_restart_relock_key_cache_fill, trans, _THIS_IP_, ck_path);
+ ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_fill);
+ goto err;
+ }
+
+ ret = bch2_trans_relock(trans);
+ if (ret) {
+ kfree(new_k);
+ goto err;
+ }
}
}
bch2_btree_node_unlock_write(trans, ck_path, ck_path->l[0].b);
/* We're not likely to need this iterator again: */
- path->preserve = false;
+ set_btree_iter_dontneed(&iter);
err:
- bch2_path_put(trans, path, 0);
+ bch2_trans_iter_exit(trans, &iter);
return ret;
}
-noinline static int
+static noinline int
bch2_btree_path_traverse_cached_slowpath(struct btree_trans *trans, struct btree_path *path,
unsigned flags)
{
path->l[1].b = NULL;
- if (bch2_btree_node_relock(trans, path, 0)) {
+ if (bch2_btree_node_relock_notrace(trans, path, 0)) {
ck = (void *) path->l[0].b;
goto fill;
}
BUG_ON(ret);
if (ck->key.btree_id != path->btree_id ||
- bpos_cmp(ck->key.pos, path->pos)) {
+ !bpos_eq(ck->key.pos, path->pos)) {
six_unlock_type(&ck->c.lock, lock_want);
goto retry;
}
path->l[0].lock_seq = ck->c.lock.state.seq;
path->l[0].b = (void *) ck;
fill:
- if (!ck->valid) {
+ path->uptodate = BTREE_ITER_UPTODATE;
+
+ if (!ck->valid && !(flags & BTREE_ITER_CACHED_NOFILL)) {
/*
* Using the underscore version because we haven't set
* path->uptodate yet:
ret = btree_key_cache_fill(trans, path, ck);
if (ret)
goto err;
+
+ ret = bch2_btree_path_relock(trans, path, _THIS_IP_);
+ if (ret)
+ goto err;
+
+ path->uptodate = BTREE_ITER_UPTODATE;
}
if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
- path->uptodate = BTREE_ITER_UPTODATE;
- BUG_ON(!ck->valid);
BUG_ON(btree_node_locked_type(path, 0) != btree_lock_want(path, 0));
+ BUG_ON(path->uptodate);
return ret;
err:
+ path->uptodate = BTREE_ITER_NEED_TRAVERSE;
if (!bch2_err_matches(ret, BCH_ERR_transaction_restart)) {
btree_node_unlock(trans, path, 0);
path->l[0].b = ERR_PTR(ret);
path->l[1].b = NULL;
- if (bch2_btree_node_relock(trans, path, 0)) {
+ if (bch2_btree_node_relock_notrace(trans, path, 0)) {
ck = (void *) path->l[0].b;
goto fill;
}
return ret;
if (ck->key.btree_id != path->btree_id ||
- bpos_cmp(ck->key.pos, path->pos)) {
+ !bpos_eq(ck->key.pos, path->pos)) {
six_unlock_type(&ck->c.lock, lock_want);
goto retry;
}
* 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|
six_unlock_read(&ck->c.lock);
goto unlock;
}
+
+ if (ck->seq != seq) {
+ bch2_journal_pin_update(&c->journal, ck->seq, &ck->journal,
+ bch2_btree_key_cache_journal_flush);
+ six_unlock_read(&ck->c.lock);
+ goto unlock;
+ }
six_unlock_read(&ck->c.lock);
ret = commit_do(&trans, NULL, NULL, 0,
}
bool bch2_btree_insert_key_cached(struct btree_trans *trans,
- struct btree_path *path,
- struct bkey_i *insert)
+ unsigned flags,
+ struct btree_insert_entry *insert_entry)
{
struct bch_fs *c = trans->c;
- struct bkey_cached *ck = (void *) path->l[0].b;
+ struct bkey_cached *ck = (void *) insert_entry->path->l[0].b;
+ struct bkey_i *insert = insert_entry->k;
bool kick_reclaim = false;
- BUG_ON(insert->u64s > ck->u64s);
+ BUG_ON(insert->k.u64s > ck->u64s);
- if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
+ if (likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY))) {
int difference;
- BUG_ON(jset_u64s(insert->u64s) > trans->journal_preres.u64s);
+ BUG_ON(jset_u64s(insert->k.u64s) > trans->journal_preres.u64s);
- difference = jset_u64s(insert->u64s) - ck->res.u64s;
+ difference = jset_u64s(insert->k.u64s) - ck->res.u64s;
if (difference > 0) {
trans->journal_preres.u64s -= difference;
ck->res.u64s += difference;
kick_reclaim = true;
}
- bch2_journal_pin_update(&c->journal, trans->journal_res.seq,
- &ck->journal, bch2_btree_key_cache_journal_flush);
+ /*
+ * To minimize lock contention, we only add the journal pin here and
+ * defer pin updates to the flush callback via ->seq. Be careful not to
+ * update ->seq on nojournal commits because we don't want to update the
+ * pin to a seq that doesn't include journal updates on disk. Otherwise
+ * we risk losing the update after a crash.
+ *
+ * The only exception is if the pin is not active in the first place. We
+ * have to add the pin because journal reclaim drives key cache
+ * flushing. The flush callback will not proceed unless ->seq matches
+ * the latest pin, so make sure it starts with a consistent value.
+ */
+ if (!(insert_entry->flags & BTREE_UPDATE_NOJOURNAL) ||
+ !journal_pin_active(&ck->journal)) {
+ ck->seq = trans->journal_res.seq;
+ }
+ bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
+ &ck->journal, bch2_btree_key_cache_journal_flush);
if (kick_reclaim)
journal_reclaim_kick(&c->journal);
unsigned start, flags;
int srcu_idx;
- /* Return -1 if we can't do anything right now */
- if (sc->gfp_mask & __GFP_FS)
- mutex_lock(&bc->lock);
- else if (!mutex_trylock(&bc->lock))
- return -1;
-
+ mutex_lock(&bc->lock);
srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
flags = memalloc_nofs_save();
struct bucket_table *tbl;
struct bkey_cached *ck, *n;
struct rhash_head *pos;
+ LIST_HEAD(items);
unsigned i;
+#ifdef __KERNEL__
int cpu;
+#endif
if (bc->shrink.list.next)
unregister_shrinker(&bc->shrink);
mutex_lock(&bc->lock);
- 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_nonpcpu);
- }
- rcu_read_unlock();
+ /*
+ * The loop is needed to guard against racing with rehash:
+ */
+ while (atomic_long_read(&bc->nr_keys)) {
+ 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, &items);
+ }
+ rcu_read_unlock();
+ }
+#ifdef __KERNEL__
for_each_possible_cpu(cpu) {
struct btree_key_cache_freelist *f =
per_cpu_ptr(bc->pcpu_freed, cpu);
for (i = 0; i < f->nr; i++) {
ck = f->objs[i];
- list_add(&ck->list, &bc->freed_nonpcpu);
+ list_add(&ck->list, &items);
}
}
+#endif
- list_splice(&bc->freed_pcpu, &bc->freed_nonpcpu);
+ list_splice(&bc->freed_pcpu, &items);
+ list_splice(&bc->freed_nonpcpu, &items);
- list_for_each_entry_safe(ck, n, &bc->freed_nonpcpu, list) {
+ mutex_unlock(&bc->lock);
+
+ list_for_each_entry_safe(ck, n, &items, list) {
cond_resched();
bch2_journal_pin_drop(&c->journal, &ck->journal);
kmem_cache_free(bch2_key_cache, ck);
}
- 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));
+ if (atomic_long_read(&bc->nr_dirty) &&
+ !bch2_journal_error(&c->journal) &&
+ test_bit(BCH_FS_WAS_RW, &c->flags))
+ panic("btree key cache shutdown error: nr_dirty nonzero (%li)\n",
+ atomic_long_read(&bc->nr_dirty));
- mutex_unlock(&bc->lock);
+ if (atomic_long_read(&bc->nr_keys))
+ panic("btree key cache shutdown error: nr_keys nonzero (%li)\n",
+ atomic_long_read(&bc->nr_keys));
if (bc->table_init_done)
rhashtable_destroy(&bc->table);
INIT_LIST_HEAD(&c->freed_nonpcpu);
}
-static void bch2_btree_key_cache_shrinker_to_text(struct printbuf *out, struct shrinker *shrink)
+static void bch2_btree_key_cache_shrinker_to_text(struct seq_buf *s, struct shrinker *shrink)
{
struct btree_key_cache *bc =
container_of(shrink, struct btree_key_cache, shrink);
+ char *cbuf;
+ size_t buflen = seq_buf_get_buf(s, &cbuf);
+ struct printbuf out = PRINTBUF_EXTERN(cbuf, buflen);
- bch2_btree_key_cache_to_text(out, bc);
+ bch2_btree_key_cache_to_text(&out, bc);
+ seq_buf_commit(s, out.pos);
}
int bch2_fs_btree_key_cache_init(struct btree_key_cache *bc)
{
struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
- int ret;
+#ifdef __KERNEL__
bc->pcpu_freed = alloc_percpu(struct btree_key_cache_freelist);
if (!bc->pcpu_freed)
- return -ENOMEM;
+ return -BCH_ERR_ENOMEM_fs_btree_cache_init;
+#endif
- ret = rhashtable_init(&bc->table, &bch2_btree_key_cache_params);
- if (ret)
- return ret;
+ if (rhashtable_init(&bc->table, &bch2_btree_key_cache_params))
+ return -BCH_ERR_ENOMEM_fs_btree_cache_init;
bc->table_init_done = true;
- bc->shrink.seeks = 1;
+ bc->shrink.seeks = 0;
bc->shrink.count_objects = bch2_btree_key_cache_count;
bc->shrink.scan_objects = bch2_btree_key_cache_scan;
bc->shrink.to_text = bch2_btree_key_cache_shrinker_to_text;
- return register_shrinker(&bc->shrink, "%s/btree_key_cache", c->name);
+ if (register_shrinker(&bc->shrink, "%s/btree_key_cache", c->name))
+ return -BCH_ERR_ENOMEM_fs_btree_cache_init;
+ return 0;
}
void bch2_btree_key_cache_to_text(struct printbuf *out, struct btree_key_cache *c)
{
- prt_printf(out, "nr_freed:\t%zu\n", atomic_long_read(&c->nr_freed));
- prt_printf(out, "nr_keys:\t%lu\n", atomic_long_read(&c->nr_keys));
- prt_printf(out, "nr_dirty:\t%lu\n", atomic_long_read(&c->nr_dirty));
+ prt_printf(out, "nr_freed:\t%zu", atomic_long_read(&c->nr_freed));
+ prt_newline(out);
+ prt_printf(out, "nr_keys:\t%lu", atomic_long_read(&c->nr_keys));
+ prt_newline(out);
+ prt_printf(out, "nr_dirty:\t%lu", atomic_long_read(&c->nr_dirty));
+ prt_newline(out);
}
void bch2_btree_key_cache_exit(void)
{
- if (bch2_key_cache)
- kmem_cache_destroy(bch2_key_cache);
+ kmem_cache_destroy(bch2_key_cache);
}
int __init bch2_btree_key_cache_init(void)
projects / bcachefs-tools-debian / blobdiff