jbd: Refine commit writeout logic

Currently we write out all journal buffers in WRITE_SYNC mode. This improves
performance for fsync heavy workloads but hinders performance when writes
are mostly asynchronous, most noticably it slows down readers and users
complain about slow desktop response etc.

So submit writes as asynchronous in the normal case and only submit writes as
WRITE_SYNC if we detect someone is waiting for current transaction commit.

I've gathered some numbers to back this change. The first is the read latency
test. It measures time to read 1 MB after several seconds of sleeping in
presence of streaming writes.

Top 10 times (out of 90) in us:
Before		After
2131586		697473
1709932		557487
1564598		535642
1480462		347573
1478579		323153
1408496		222181
1388960		181273
1329565		181070
1252486		172832
1223265		172278

Average:
619377		82180

So the improvement in both maximum and average latency is massive.

I've measured fsync throughput by:
fs_mark -n 100 -t 1 -s 16384 -d /mnt/fsync/ -S 1 -L 4

in presence of streaming reader. The numbers (fsyncs/s) are:
Before		After
9.9		6.3
6.8		6.0
6.3		6.2
5.8		6.1

So fsync performance seems unharmed by this change.

Signed-off-by: Jan Kara <jack@suse.cz>

3 files changed, 9 insertions, 5 deletions

diff --git a/fs/jbd/commit.c b/fs/jbd/commit.c
index f2b9a571f4cf..9d31e6a39205 100644
--- a/fs/jbd/commit.c
+++ b/fs/jbd/commit.c
@@ -298,6 +298,7 @@ void journal_commit_transaction(journal_t *journal)
 	int tag_flag;
 	int i;
 	struct blk_plug plug;
+	int write_op = WRITE;
 
 	/*
 	 * First job: lock down the current transaction and wait for
@@ -413,13 +414,16 @@ void journal_commit_transaction(journal_t *journal)
 
 	jbd_debug (3, "JBD: commit phase 2\n");
 
+	if (tid_geq(journal->j_commit_waited, commit_transaction->t_tid))
+		write_op = WRITE_SYNC;
+
 	/*
 	 * Now start flushing things to disk, in the order they appear
 	 * on the transaction lists.  Data blocks go first.
 	 */
 	blk_start_plug(&plug);
 	err = journal_submit_data_buffers(journal, commit_transaction,
-					  WRITE_SYNC);
+					  write_op);
 	blk_finish_plug(&plug);
 
 	/*
@@ -478,7 +482,7 @@ void journal_commit_transaction(journal_t *journal)
 
 	blk_start_plug(&plug);
 
-	journal_write_revoke_records(journal, commit_transaction, WRITE_SYNC);
+	journal_write_revoke_records(journal, commit_transaction, write_op);
 
 	/*
 	 * If we found any dirty or locked buffers, then we should have
@@ -649,7 +653,7 @@ start_journal_io:
 				clear_buffer_dirty(bh);
 				set_buffer_uptodate(bh);
 				bh->b_end_io = journal_end_buffer_io_sync;
-				submit_bh(WRITE_SYNC, bh);
+				submit_bh(write_op, bh);
 			}
 			cond_resched();
 
diff --git a/fs/jbd/journal.c b/fs/jbd/journal.c
index 0971e9217808..2047fd77bf38 100644
--- a/fs/jbd/journal.c
+++ b/fs/jbd/journal.c
@@ -563,6 +563,8 @@ int log_wait_commit(journal_t *journal, tid_t tid)
 	spin_unlock(&journal->j_state_lock);
 #endif
 	spin_lock(&journal->j_state_lock);
+	if (!tid_geq(journal->j_commit_waited, tid))
+		journal->j_commit_waited = tid;
 	while (tid_gt(tid, journal->j_commit_sequence)) {
 		jbd_debug(1, "JBD: want %d, j_commit_sequence=%d\n",
 				  tid, journal->j_commit_sequence);
diff --git a/fs/jbd/transaction.c b/fs/jbd/transaction.c
index b2a7e5244e39..febc10db5ced 100644
--- a/fs/jbd/transaction.c
+++ b/fs/jbd/transaction.c
@@ -1433,8 +1433,6 @@ int journal_stop(handle_t *handle)
 		}
 	}
 
-	if (handle->h_sync)
-		transaction->t_synchronous_commit = 1;
 	current->journal_info = NULL;
 	spin_lock(&journal->j_state_lock);
 	spin_lock(&transaction->t_handle_lock);