Today, I ran 4 rounds of this testcase and they all passed. So, I'm thinking maybe it's the CPU scheduler problem on KVM cluster, which is also happened for corosync... jbd2 is always the first to report hung warning; we can see this 2 lines (reverse order): kernel: [<ffffffff81704cac>] schedule+0x3c/0x90 kernel: [<ffffffff813053f0>] jbd2_journal_commit_transaction+0x250/0x1920 In what situation, jbd2_journal_commit_transaction() will give up the cpu by schedule(), and never get scheduled from then? Oh, it sounds crazy... Anyway, i'd like keep this low priority so far, and look into further