Re: [PATCH] jbd2: Optimize jbd2_log_do_checkpoint() a bit

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

On Tue, Sep 02, 2014 at 07:18:30PM +0200, Jan Kara wrote:
> When we discover written out buffer in transaction checkpoint list we
> don't have to recheck validity of a transaction. Either this is the last
> buffer in a transaction - and then we are done - or this isn't and then
> we can just take another buffer from the checkpoint list without
> dropping j_list_lock.
> 
> Signed-off-by: Jan Kara <jack@suse.cz>
> ---
>  fs/jbd2/checkpoint.c | 13 ++++++++-----
>  1 file changed, 8 insertions(+), 5 deletions(-)
> 
> diff --git a/fs/jbd2/checkpoint.c b/fs/jbd2/checkpoint.c
> index 993a187527f3..3722e2e53638 100644
> --- a/fs/jbd2/checkpoint.c
> +++ b/fs/jbd2/checkpoint.c
> @@ -343,12 +343,15 @@ restart:
>  		if (!buffer_dirty(bh)) {
>  			if (unlikely(buffer_write_io_error(bh)) && !result)
>  				result = -EIO;
> -			get_bh(bh);
>  			BUFFER_TRACE(bh, "remove from checkpoint");
> -			__jbd2_journal_remove_checkpoint(jh);
> -			spin_unlock(&journal->j_list_lock);
> -			__brelse(bh);

Currently, all of the places which call
__jbd2_jouranl_remove_checkpoint(jh) are doing so with an elevated
b_count.  For example, see __try_to_free_cp_buf().

After doing a lot of desk checking, I can't see any reason for holding
the elevanted b_count, so I think it should be to remove it, but then
we can simplify the other uses __try_to_free_cp_buf().  For example,
in the loop that I folded from __wait_cp_io, we could drop the done
variable and change:

		done = __jbd2_journal_remove_checkpoint(jh);
		__brelse(bh);

to this:

		__brelse(bh);
		if (__jbd2_journal_remove_checkpoint(jh))
			break;

How much testing have you done of this optimization?  I'm tempted to
try nuking all of the elevated b_counts around the call to
__jbd2_journal_remove_checkpoint(), and then doing a test to see if
anything blows up.

Cheers,

					- Ted