for (i = 0; i < BTREE_ID_NR; i++)
if (c->btree_roots[i].b)
reserve += min_t(unsigned, 1,
- c->btree_roots[i].b->level) * 8;
+ c->btree_roots[i].b->c.level) * 8;
c->btree_cache.reserve = reserve;
}
const struct btree *b = obj;
const u64 *v = arg->key;
- return PTR_HASH(&b->key) == *v ? 0 : 1;
+ return b->hash_val == *v ? 0 : 1;
}
static const struct rhashtable_params bch_btree_cache_params = {
.head_offset = offsetof(struct btree, hash),
- .key_offset = offsetof(struct btree, key.v),
- .key_len = sizeof(struct bch_extent_ptr),
+ .key_offset = offsetof(struct btree, hash_val),
+ .key_len = sizeof(u64),
.obj_cmpfn = bch2_btree_cache_cmp_fn,
};
-static void btree_node_data_alloc(struct bch_fs *c, struct btree *b, gfp_t gfp)
+static int __btree_node_data_alloc(struct bch_fs *c, struct btree *b, gfp_t gfp)
{
- struct btree_cache *bc = &c->btree_cache;
+ BUG_ON(b->data || b->aux_data);
b->data = kvpmalloc(btree_bytes(c), gfp);
if (!b->data)
- goto err;
+ return -ENOMEM;
- if (bch2_btree_keys_alloc(b, btree_page_order(c), gfp))
- goto err;
+ if (bch2_btree_keys_alloc(b, btree_page_order(c), gfp)) {
+ kvpfree(b->data, btree_bytes(c));
+ b->data = NULL;
+ return -ENOMEM;
+ }
- bc->used++;
- list_move(&b->list, &bc->freeable);
- return;
-err:
- kvpfree(b->data, btree_bytes(c));
- b->data = NULL;
- list_move(&b->list, &bc->freed);
+ return 0;
+}
+
+static void btree_node_data_alloc(struct bch_fs *c, struct btree *b, gfp_t gfp)
+{
+ struct btree_cache *bc = &c->btree_cache;
+
+ if (!__btree_node_data_alloc(c, b, gfp)) {
+ bc->used++;
+ list_move(&b->list, &bc->freeable);
+ } else {
+ list_move(&b->list, &bc->freed);
+ }
}
static struct btree *btree_node_mem_alloc(struct bch_fs *c, gfp_t gfp)
return NULL;
bkey_btree_ptr_init(&b->key);
- six_lock_init(&b->lock);
+ six_lock_init(&b->c.lock);
INIT_LIST_HEAD(&b->list);
INIT_LIST_HEAD(&b->write_blocked);
rhashtable_remove_fast(&bc->table, &b->hash, bch_btree_cache_params);
/* Cause future lookups for this node to fail: */
- PTR_HASH(&b->key) = 0;
+ b->hash_val = 0;
}
int __bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b)
{
+ BUG_ON(b->hash_val);
+ b->hash_val = btree_ptr_hash_val(&b->key);
+
return rhashtable_lookup_insert_fast(&bc->table, &b->hash,
bch_btree_cache_params);
}
{
int ret;
- b->level = level;
- b->btree_id = id;
+ b->c.level = level;
+ b->c.btree_id = id;
mutex_lock(&bc->lock);
ret = __bch2_btree_node_hash_insert(bc, b);
static inline struct btree *btree_cache_find(struct btree_cache *bc,
const struct bkey_i *k)
{
- return rhashtable_lookup_fast(&bc->table, &PTR_HASH(k),
- bch_btree_cache_params);
+ u64 v = btree_ptr_hash_val(k);
+
+ return rhashtable_lookup_fast(&bc->table, &v, bch_btree_cache_params);
}
/*
lockdep_assert_held(&bc->lock);
- if (!six_trylock_intent(&b->lock))
+ if (!six_trylock_intent(&b->c.lock))
return -ENOMEM;
- if (!six_trylock_write(&b->lock))
+ if (!six_trylock_write(&b->c.lock))
goto out_unlock_intent;
if (btree_node_noevict(b))
btree_node_wait_on_io(b);
}
out:
- if (PTR_HASH(&b->key) && !ret)
+ if (b->hash_val && !ret)
trace_btree_node_reap(c, b);
return ret;
out_unlock:
- six_unlock_write(&b->lock);
+ six_unlock_write(&b->c.lock);
out_unlock_intent:
- six_unlock_intent(&b->lock);
+ six_unlock_intent(&b->c.lock);
ret = -ENOMEM;
goto out;
}
return SHRINK_STOP;
/* Return -1 if we can't do anything right now */
- if (sc->gfp_mask & __GFP_IO)
+ if (sc->gfp_mask & __GFP_FS)
mutex_lock(&bc->lock);
else if (!mutex_trylock(&bc->lock))
return -1;
if (++i > 3 &&
!btree_node_reclaim(c, b)) {
btree_node_data_free(c, b);
- six_unlock_write(&b->lock);
- six_unlock_intent(&b->lock);
+ six_unlock_write(&b->c.lock);
+ six_unlock_intent(&b->c.lock);
freed++;
}
}
mutex_unlock(&bc->lock);
bch2_btree_node_hash_remove(bc, b);
- six_unlock_write(&b->lock);
- six_unlock_intent(&b->lock);
+ six_unlock_write(&b->c.lock);
+ six_unlock_intent(&b->c.lock);
if (freed >= nr)
goto out;
*/
list_for_each_entry(b, &bc->freeable, list)
if (!btree_node_reclaim(c, b))
- goto out_unlock;
+ goto got_node;
/*
* We never free struct btree itself, just the memory that holds the on
* disk node. Check the freed list before allocating a new one:
*/
list_for_each_entry(b, &bc->freed, list)
- if (!btree_node_reclaim(c, b)) {
- btree_node_data_alloc(c, b, __GFP_NOWARN|GFP_NOIO);
- if (b->data)
- goto out_unlock;
+ if (!btree_node_reclaim(c, b))
+ goto got_node;
- six_unlock_write(&b->lock);
- six_unlock_intent(&b->lock);
+ b = NULL;
+got_node:
+ if (b)
+ list_del_init(&b->list);
+ mutex_unlock(&bc->lock);
+
+ if (!b) {
+ b = kzalloc(sizeof(struct btree), GFP_KERNEL);
+ if (!b)
goto err;
- }
- b = btree_node_mem_alloc(c, __GFP_NOWARN|GFP_NOIO);
- if (!b)
- goto err;
+ bkey_btree_ptr_init(&b->key);
+ six_lock_init(&b->c.lock);
+ INIT_LIST_HEAD(&b->list);
+ INIT_LIST_HEAD(&b->write_blocked);
+
+ BUG_ON(!six_trylock_intent(&b->c.lock));
+ BUG_ON(!six_trylock_write(&b->c.lock));
+ }
+
+ if (!b->data) {
+ if (__btree_node_data_alloc(c, b, __GFP_NOWARN|GFP_KERNEL))
+ goto err;
+
+ mutex_lock(&bc->lock);
+ bc->used++;
+ mutex_unlock(&bc->lock);
+ }
- BUG_ON(!six_trylock_intent(&b->lock));
- BUG_ON(!six_trylock_write(&b->lock));
-out_unlock:
BUG_ON(btree_node_hashed(b));
BUG_ON(btree_node_write_in_flight(b));
-
- list_del_init(&b->list);
- mutex_unlock(&bc->lock);
- memalloc_nofs_restore(flags);
out:
b->flags = 0;
b->written = 0;
b->sib_u64s[0] = 0;
b->sib_u64s[1] = 0;
b->whiteout_u64s = 0;
- b->uncompacted_whiteout_u64s = 0;
bch2_btree_keys_init(b, &c->expensive_debug_checks);
bch2_time_stats_update(&c->times[BCH_TIME_btree_node_mem_alloc],
start_time);
+ memalloc_nofs_restore(flags);
return b;
err:
+ mutex_lock(&bc->lock);
+
+ if (b) {
+ list_add(&b->list, &bc->freed);
+ six_unlock_write(&b->c.lock);
+ six_unlock_intent(&b->c.lock);
+ }
+
/* Try to cannibalize another cached btree node: */
if (bc->alloc_lock == current) {
b = btree_node_cannibalize(c);
}
mutex_unlock(&bc->lock);
+ memalloc_nofs_restore(flags);
return ERR_PTR(-ENOMEM);
}
static noinline struct btree *bch2_btree_node_fill(struct bch_fs *c,
struct btree_iter *iter,
const struct bkey_i *k,
+ enum btree_id btree_id,
unsigned level,
enum six_lock_type lock_type,
bool sync)
struct btree_cache *bc = &c->btree_cache;
struct btree *b;
+ BUG_ON(level + 1 >= BTREE_MAX_DEPTH);
/*
* Parent node must be locked, else we could read in a btree node that's
* been freed:
*/
- BUG_ON(!btree_node_locked(iter, level + 1));
- BUG_ON(level >= BTREE_MAX_DEPTH);
+ if (iter && !bch2_btree_node_relock(iter, level + 1))
+ return ERR_PTR(-EINTR);
b = bch2_btree_node_mem_alloc(c);
if (IS_ERR(b))
return b;
bkey_copy(&b->key, k);
- if (bch2_btree_node_hash_insert(bc, b, level, iter->btree_id)) {
+ if (bch2_btree_node_hash_insert(bc, b, level, btree_id)) {
/* raced with another fill: */
/* mark as unhashed... */
- PTR_HASH(&b->key) = 0;
+ b->hash_val = 0;
mutex_lock(&bc->lock);
list_add(&b->list, &bc->freeable);
mutex_unlock(&bc->lock);
- six_unlock_write(&b->lock);
- six_unlock_intent(&b->lock);
+ six_unlock_write(&b->c.lock);
+ six_unlock_intent(&b->c.lock);
return NULL;
}
/*
- * If the btree node wasn't cached, we can't drop our lock on
- * the parent until after it's added to the cache - because
- * otherwise we could race with a btree_split() freeing the node
- * we're trying to lock.
+ * Unlock before doing IO:
*
- * But the deadlock described below doesn't exist in this case,
- * so it's safe to not drop the parent lock until here:
+ * XXX: ideally should be dropping all btree node locks here
*/
- if (btree_node_read_locked(iter, level + 1))
+ if (iter && btree_node_read_locked(iter, level + 1))
btree_node_unlock(iter, level + 1);
bch2_btree_node_read(c, b, sync);
- six_unlock_write(&b->lock);
+ six_unlock_write(&b->c.lock);
if (!sync) {
- six_unlock_intent(&b->lock);
+ six_unlock_intent(&b->c.lock);
return NULL;
}
if (lock_type == SIX_LOCK_read)
- six_lock_downgrade(&b->lock);
+ six_lock_downgrade(&b->c.lock);
return b;
}
+static int lock_node_check_fn(struct six_lock *lock, void *p)
+{
+ struct btree *b = container_of(lock, struct btree, c.lock);
+ const struct bkey_i *k = p;
+
+ return b->hash_val == btree_ptr_hash_val(k) ? 0 : -1;
+}
+
/**
* bch_btree_node_get - find a btree node in the cache and lock it, reading it
* in from disk if necessary.
struct btree *b;
struct bset_tree *t;
- /*
- * XXX: locking optimization
- *
- * we can make the locking looser here - caller can drop lock on parent
- * node before locking child node (and potentially blocking): we just
- * have to have bch2_btree_node_fill() call relock on the parent and
- * return -EINTR if that fails
- */
- EBUG_ON(!btree_node_locked(iter, level + 1));
EBUG_ON(level >= BTREE_MAX_DEPTH);
+
+ b = btree_node_mem_ptr(k);
+ if (b)
+ goto lock_node;
retry:
b = btree_cache_find(bc, k);
if (unlikely(!b)) {
* else we could read in a btree node from disk that's been
* freed:
*/
- b = bch2_btree_node_fill(c, iter, k, level, lock_type, true);
+ b = bch2_btree_node_fill(c, iter, k, iter->btree_id,
+ level, lock_type, true);
/* We raced and found the btree node in the cache */
if (!b)
if (IS_ERR(b))
return b;
} else {
+lock_node:
/*
* There's a potential deadlock with splits and insertions into
* interior nodes we have to avoid:
* free it:
*
* To guard against this, btree nodes are evicted from the cache
- * when they're freed - and PTR_HASH() is zeroed out, which we
+ * when they're freed - and b->hash_val is zeroed out, which we
* check for after we lock the node.
*
* Then, bch2_btree_node_relock() on the parent will fail - because
if (btree_node_read_locked(iter, level + 1))
btree_node_unlock(iter, level + 1);
- if (!btree_node_lock(b, k->k.p, level, iter, lock_type))
+ if (!btree_node_lock(b, k->k.p, level, iter, lock_type,
+ lock_node_check_fn, (void *) k)) {
+ if (b->hash_val != btree_ptr_hash_val(k))
+ goto retry;
return ERR_PTR(-EINTR);
+ }
- if (unlikely(PTR_HASH(&b->key) != PTR_HASH(k) ||
- b->level != level ||
+ if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
+ b->c.level != level ||
race_fault())) {
- six_unlock_type(&b->lock, lock_type);
+ six_unlock_type(&b->c.lock, lock_type);
if (bch2_btree_node_relock(iter, level + 1))
goto retry;
}
}
+ /* XXX: waiting on IO with btree locks held: */
+ wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
+ TASK_UNINTERRUPTIBLE);
+
+ prefetch(b->aux_data);
+
+ for_each_bset(b, t) {
+ void *p = (u64 *) b->aux_data + t->aux_data_offset;
+
+ prefetch(p + L1_CACHE_BYTES * 0);
+ prefetch(p + L1_CACHE_BYTES * 1);
+ prefetch(p + L1_CACHE_BYTES * 2);
+ }
+
+ /* avoid atomic set bit if it's not needed: */
+ if (!btree_node_accessed(b))
+ set_btree_node_accessed(b);
+
+ if (unlikely(btree_node_read_error(b))) {
+ six_unlock_type(&b->c.lock, lock_type);
+ return ERR_PTR(-EIO);
+ }
+
+ EBUG_ON(b->c.btree_id != iter->btree_id ||
+ BTREE_NODE_LEVEL(b->data) != level ||
+ bkey_cmp(b->data->max_key, k->k.p));
+
+ return b;
+}
+
+struct btree *bch2_btree_node_get_noiter(struct bch_fs *c,
+ const struct bkey_i *k,
+ enum btree_id btree_id,
+ unsigned level)
+{
+ struct btree_cache *bc = &c->btree_cache;
+ struct btree *b;
+ struct bset_tree *t;
+ int ret;
+
+ EBUG_ON(level >= BTREE_MAX_DEPTH);
+
+ b = btree_node_mem_ptr(k);
+ if (b)
+ goto lock_node;
+retry:
+ b = btree_cache_find(bc, k);
+ if (unlikely(!b)) {
+ b = bch2_btree_node_fill(c, NULL, k, btree_id,
+ level, SIX_LOCK_read, true);
+
+ /* We raced and found the btree node in the cache */
+ if (!b)
+ goto retry;
+
+ if (IS_ERR(b))
+ return b;
+ } else {
+lock_node:
+ ret = six_lock_read(&b->c.lock, lock_node_check_fn, (void *) k);
+ if (ret)
+ goto retry;
+
+ if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
+ b->c.btree_id != btree_id ||
+ b->c.level != level)) {
+ six_unlock_read(&b->c.lock);
+ goto retry;
+ }
+ }
+
+ /* XXX: waiting on IO with btree locks held: */
wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
TASK_UNINTERRUPTIBLE);
}
/* avoid atomic set bit if it's not needed: */
- if (btree_node_accessed(b))
+ if (!btree_node_accessed(b))
set_btree_node_accessed(b);
if (unlikely(btree_node_read_error(b))) {
- six_unlock_type(&b->lock, lock_type);
+ six_unlock_read(&b->c.lock);
return ERR_PTR(-EIO);
}
- EBUG_ON(b->btree_id != iter->btree_id ||
+ EBUG_ON(b->c.btree_id != btree_id ||
BTREE_NODE_LEVEL(b->data) != level ||
bkey_cmp(b->data->max_key, k->k.p));
struct bkey_packed *k;
BKEY_PADDED(k) tmp;
struct btree *ret = NULL;
- unsigned level = b->level;
+ unsigned level = b->c.level;
parent = btree_iter_node(iter, level + 1);
if (!parent)
return NULL;
+ /*
+ * There's a corner case where a btree_iter might have a node locked
+ * that is just outside its current pos - when
+ * bch2_btree_iter_set_pos_same_leaf() gets to the end of the node.
+ *
+ * But the lock ordering checks in __bch2_btree_node_lock() go off of
+ * iter->pos, not the node's key: so if the iterator is marked as
+ * needing to be traversed, we risk deadlock if we don't bail out here:
+ */
+ if (iter->uptodate >= BTREE_ITER_NEED_TRAVERSE)
+ return ERR_PTR(-EINTR);
+
if (!bch2_btree_node_relock(iter, level + 1)) {
ret = ERR_PTR(-EINTR);
goto out;
}
- node_iter = iter->l[parent->level].iter;
+ node_iter = iter->l[parent->c.level].iter;
k = bch2_btree_node_iter_peek_all(&node_iter, parent);
BUG_ON(bkey_cmp_left_packed(parent, k, &b->key.k.p));
btree_iter_set_dirty(iter, BTREE_ITER_NEED_RELOCK);
if (!IS_ERR(ret)) {
- six_unlock_intent(&ret->lock);
+ six_unlock_intent(&ret->c.lock);
ret = ERR_PTR(-EINTR);
}
}
if (sib != btree_prev_sib)
swap(n1, n2);
- BUG_ON(bkey_cmp(btree_type_successor(n1->btree_id,
- n1->key.k.p),
+ BUG_ON(bkey_cmp(bkey_successor(n1->key.k.p),
n2->data->min_key));
}
if (b)
return;
- bch2_btree_node_fill(c, iter, k, level, SIX_LOCK_read, false);
+ bch2_btree_node_fill(c, iter, k, iter->btree_id,
+ level, SIX_LOCK_read, false);
}
void bch2_btree_node_to_text(struct printbuf *out, struct bch_fs *c,
pr_buf(out,
"l %u %llu:%llu - %llu:%llu:\n"
" ptrs: ",
- b->level,
+ b->c.level,
b->data->min_key.inode,
b->data->min_key.offset,
b->data->max_key.inode,