On Tue, Nov 06, 2018 at 11:22:30AM +0100, Jan Kara wrote: >> So the buffer is on BJ_Shadow list while the assertion in > jbd2_journal_dirty_metadata() expects it to be in BJ_Metadata list. This is > really weird as we have also checked that jh->b_transaction == > handle->h_transaction so the transaction couldn't have passed to commit > phase... Oh, I see, the code in start_this_handle() got racy with the > removal of j_state_lock protection from journal_commit_transaction() so now > transaction can start even though there are handles outstanding! I'll think > about the best solution for this. Thanks for report! Thanks for the analysis! I finished the bisection last night and it was too late for me to dive into how this was going on. I should have realized this before I had suggested the approach in the patch. The original complaint which Andrian made was that the long hold times of j_state_lock at the beginning of the commit. What he didn't mention was what the other "high priority tasks" were blocked on, but they were almost certainly start_this_handle. And that's fundamental; when we are trying to at the beginning of the commit process is waiting for the outstanding handles to close; and so we can't let new handles start. What we can do is to try to decrease the handle hold times. This is why we track the handle type and we have the jbd2_handle_stats tracepoint. If we can find that handles of a particular type and a particular line number are the ones which are taking more time than other handles, we can try to make them run faster; for example, by pre-reading blocks aren't in the buffer cache before starting the handle. The other thing which we can probably do is to for truncate handles, if we notice that current_transaction->t_state == T_LOCKED, we can suspend the truncation activity and close the handle, and then resume it (which will block until a new transaction is ready to be started). - Ted
