Fast commit performance can be optimized if commit thread doesn't wait for ongoing fast commits to complete until the transaction enters T_FLUSH state. Document this optimization. Suggested-by: Jan Kara <jack@xxxxxxx> Signed-off-by: Harshad Shirwadkar <harshadshirwadkar@xxxxxxxxx> --- fs/jbd2/commit.c | 9 +++++++++ 1 file changed, 9 insertions(+) diff --git a/fs/jbd2/commit.c b/fs/jbd2/commit.c index ec516490cb35..b121d7d434c6 100644 --- a/fs/jbd2/commit.c +++ b/fs/jbd2/commit.c @@ -450,6 +450,15 @@ void jbd2_journal_commit_transaction(journal_t *journal) schedule(); write_lock(&journal->j_state_lock); finish_wait(&journal->j_fc_wait, &wait); + /* + * TODO: by blocking fast commits here, we are increasing + * fsync() latency slightly. Strictly speaking, we don't need + * to block fast commits until the transaction enters T_FLUSH + * state. So an optimization is possible where we block new fast + * commits here and wait for existing ones to complete + * just before we enter T_FLUSH. That way, the existing fast + * commits and this full commit can proceed parallely. + */ } write_unlock(&journal->j_state_lock); -- 2.29.1.341.ge80a0c044ae-goog
