[PATCH] jbd2: Speedup jbd2_journal_dirty_metadata()

[PATCH] jbd2: Speedup jbd2_journal_dirty_metadata()

[Jan Kara]

From: Jan Kara <jack@suse.cz>

It is often the case that we mark buffer as having dirty metadata when
the buffer is already in that state (frequent for bitmaps, inode table
blocks, superblock). Thus it is unnecessary to contend on grabbing
journal head reference and bh_state lock. Avoid that by checking whether
any modification to the buffer is needed before grabbing any locks or
references.

Signed-off-by: Jan Kara <jack@suse.cz>
---
 fs/jbd2/transaction.c | 38 ++++++++++++++++++++++++++++++++------
 1 file changed, 32 insertions(+), 6 deletions(-)

  Ted, this is the speedup patch with fixed assertion. It survived several
hours of beating with xfstests generic/269. Ming Lei, can you verify this
patch passes testing on your machine as well? Thanks!

diff --git a/fs/jbd2/transaction.c b/fs/jbd2/transaction.c
index f3d06174b051..546ce0e76605 100644
--- a/fs/jbd2/transaction.c
+++ b/fs/jbd2/transaction.c
@@ -1280,8 +1280,6 @@ void jbd2_buffer_abort_trigger(struct journal_head *jh,
 	triggers->t_abort(triggers, jh2bh(jh));
 }
 
-
-
 /**
  * int jbd2_journal_dirty_metadata() -  mark a buffer as containing dirty metadata
  * @handle: transaction to add buffer to.
@@ -1314,12 +1312,41 @@ int jbd2_journal_dirty_metadata(handle_t *handle, struct buffer_head *bh)
 
 	if (is_handle_aborted(handle))
 		return -EROFS;
-	journal = transaction->t_journal;
-	jh = jbd2_journal_grab_journal_head(bh);
-	if (!jh) {
+	if (!buffer_jbd(bh)) {
 		ret = -EUCLEAN;
 		goto out;
 	}
+	/*
+	 * We don't grab jh reference here since the buffer must be part
+	 * of the running transaction.
+	 */
+	jh = bh2jh(bh);
+	/*
+	 * This and the following assertions are unreliable since we may see jh
+	 * in inconsistent state unless we grab bh_state lock. But this is
+	 * crutial to catch bugs so let's do a reliable check until the
+	 * lockless handling is fully proven.
+	 */
+	if (jh->b_transaction != transaction &&
+	    jh->b_next_transaction != transaction) {
+		jbd_lock_bh_state(bh);
+		J_ASSERT_JH(jh, jh->b_transaction == transaction ||
+				jh->b_next_transaction == transaction);
+		jbd_unlock_bh_state(bh);
+	}
+	if (jh->b_modified == 1) {
+		/* If it's in our transaction it must be in BJ_Metadata list. */
+		if (jh->b_transaction == transaction &&
+		    jh->b_jlist != BJ_Metadata) {
+			jbd_lock_bh_state(bh);
+			J_ASSERT_JH(jh, jh->b_transaction != transaction ||
+					jh->b_jlist == BJ_Metadata);
+			jbd_unlock_bh_state(bh);
+		}
+		goto out;
+	}
+
+	journal = transaction->t_journal;
 	jbd_debug(5, "journal_head %p\n", jh);
 	JBUFFER_TRACE(jh, "entry");
 
@@ -1410,7 +1437,6 @@ int jbd2_journal_dirty_metadata(handle_t *handle, struct buffer_head *bh)
 	spin_unlock(&journal->j_list_lock);
 out_unlock_bh:
 	jbd_unlock_bh_state(bh);
-	jbd2_journal_put_journal_head(jh);
 out:
 	JBUFFER_TRACE(jh, "exit");
 	return ret;
-- 
2.1.4

Re: [PATCH] jbd2: Speedup jbd2_journal_dirty_metadata()

[Theodore Ts'o]

On Fri, Jul 03, 2015 at 07:37:12PM +0200, Jan Kara wrote:
> From: Jan Kara <jack@suse.cz>
> 
> It is often the case that we mark buffer as having dirty metadata when
> the buffer is already in that state (frequent for bitmaps, inode table
> blocks, superblock). Thus it is unnecessary to contend on grabbing
> journal head reference and bh_state lock. Avoid that by checking whether
> any modification to the buffer is needed before grabbing any locks or
> references.
> 
> Signed-off-by: Jan Kara <jack@suse.cz>

Applied, thanks.

					- Ted