Btrfs: fix the deadlock between the transaction start/attach and commit
Now btrfs_commit_transaction() does this
ret = btrfs_run_ordered_operations(root, 0)
which async flushes all inodes on the ordered operations list, it introduced
a deadlock that transaction-start task, transaction-commit task and the flush
workers waited for each other.
(See the following URL to get the detail
http://marc.info/?l=linux-btrfs&m=136070705732646&w=2)
As we know, if ->in_commit is set, it means someone is committing the
current transaction, we should not try to join it if we are not JOIN
or JOIN_NOLOCK, wait is the best choice for it. In this way, we can avoid
the above problem. In this way, there is another benefit: there is no new
transaction handle to block the transaction which is on the way of commit,
once we set ->in_commit.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
diff --git a/fs/btrfs/transaction.c b/fs/btrfs/transaction.c
index 425d5b5..5767ea1 100644
--- a/fs/btrfs/transaction.c
+++ b/fs/btrfs/transaction.c
@@ -50,6 +50,14 @@
root->commit_root = btrfs_root_node(root);
}
+static inline int can_join_transaction(struct btrfs_transaction *trans,
+ int type)
+{
+ return !(trans->in_commit &&
+ type != TRANS_JOIN &&
+ type != TRANS_JOIN_NOLOCK);
+}
+
/*
* either allocate a new transaction or hop into the existing one
*/
@@ -85,6 +93,10 @@
spin_unlock(&fs_info->trans_lock);
return cur_trans->aborted;
}
+ if (!can_join_transaction(cur_trans, type)) {
+ spin_unlock(&fs_info->trans_lock);
+ return -EBUSY;
+ }
atomic_inc(&cur_trans->use_count);
atomic_inc(&cur_trans->num_writers);
cur_trans->num_joined++;
@@ -360,8 +372,11 @@
do {
ret = join_transaction(root, type);
- if (ret == -EBUSY)
+ if (ret == -EBUSY) {
wait_current_trans(root);
+ if (unlikely(type == TRANS_ATTACH))
+ ret = -ENOENT;
+ }
} while (ret == -EBUSY);
if (ret < 0) {