On Wed, Oct 10, 2018 at 04:43:27PM +0300, Adrian Hunter wrote: > Hi > > I have a case on a v4.14 kernel where the EXT4 journal commit disables > preemption for 30ms due to jbd2_clear_buffer_revoked_flags(). That in turn > disables preemption on other CPUs as they come to spin waiting for the same > lock. The side-effect of that is that it periodically blocks high priority > tasks from running. > > I see jbd2_clear_buffer_revoked_flags() iterating 32768 times calling > __find_get_block(). > > Is there any way to make jbd2_clear_buffer_revoked_flags() take less time, > or move its work out from under write_lock(&journal->j_state_lock)? Hmm.... I'd have to look a bit more carefully and then run some tests, but I *think* we can drop the j_state_lock at the beginning of JBD2 commit phase 1, and then grab it again right before we set commit_transaction->t_state to T_FLUSH. That should be safe because while the transaction state is T_LOCKED, we can't start any new handles, so there can't be any new blocks added to the revoke table. Can you give that a try and see whether that solves your priority inversion problem? Cheers, - Ted
