Re: [PATCH 5/7] jbd2: Don't call __bforget() unnecessarily

["Theodore Y. Ts'o" <tytso@mit.edu>]

On Fri, Aug 09, 2019 at 02:42:31PM +0200, Jan Kara wrote:
> @@ -1660,10 +1660,9 @@ int jbd2_journal_forget (handle_t *handle, struct buffer_head *bh)
>  		__jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
>  		spin_unlock(&journal->j_list_lock);
>  	}
> -
> +drop:
>  	jbd_unlock_bh_state(bh);
>  	__brelse(bh);
> -drop:
>  	if (drop_reserve) {
>  		/* no need to reserve log space for this block -bzzz */
>  		handle->h_buffer_credits++;

(Workflow observation: this is an example where Gerrit can be *so*
*much* *better* than e-mail review at times; sometimes, you *really*
want to see more context than what e-mail affords you.)

After this patch, the resulting code looks like this:

-----
drop:
	jbd_unlock_bh_state(bh);
	__brelse(bh);
	if (drop_reserve) {
		/* no need to reserve log space for this block -bzzz */
		handle->h_buffer_credits++;
	}
	return err;

not_jbd:
	jbd_unlock_bh_state(bh);
	__bforget(bh);
	goto drop;
----

And we still have a case we jump to not_jbd, at which point hilarity
will ensue.

This is cleaned up in the following patch in this sequence, but this
leaves us in a not-great state if we are ever bisecting.

					- Ted