drbd: introduce drbd_recv_header_maybe_unplug

Recently, drbd_recv_header() was changed to potentially
implicitly "unplug" the backend device(s), in case there
is currently nothing to receive.

Be more explicit about it: re-introduce the original drbd_recv_header(),
and introduce a new drbd_recv_header_maybe_unplug() for use by the
receiver "main loop".

Using explicit plugging via blk_start_plug(); blk_finish_plug();
really helps the io-scheduler of the backend with merging requests.

Wrap the receiver "main loop" with such a plug.
Also catch unplug events on the Primary,
and try to propagate.

This is performance relevant.  Without this, if the receiving side does
not merge requests, number of IOPS on the peer can me significantly
higher than IOPS on the Primary, and can easily become the bottleneck.

Together, both changes should help to reduce the number of IOPS
as seen on the backend of the receiving side, by increasing
the chance of merging mergable requests, without trading latency
for more throughput.

Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

1 files changed, 44 insertions, 3 deletions

diff --git a/drivers/block/drbd/drbd_receiver.c b/drivers/block/drbd/drbd_receiver.c
index ece6e5d7dc3f2..1b3f439a3b237 100644
--- a/drivers/block/drbd/drbd_receiver.c
+++ b/drivers/block/drbd/drbd_receiver.c
@@ -1194,6 +1194,14 @@ static int decode_header(struct drbd_connection *connection, void *header, struc
 	return 0;
 }
 
+static void drbd_unplug_all_devices(struct drbd_connection *connection)
+{
+	if (current->plug == &connection->receiver_plug) {
+		blk_finish_plug(&connection->receiver_plug);
+		blk_start_plug(&connection->receiver_plug);
+	} /* else: maybe just schedule() ?? */
+}
+
 static int drbd_recv_header(struct drbd_connection *connection, struct packet_info *pi)
 {
 	void *buffer = connection->data.rbuf;
@@ -1209,6 +1217,36 @@ static int drbd_recv_header(struct drbd_connection *connection, struct packet_in
 	return err;
 }
 
+static int drbd_recv_header_maybe_unplug(struct drbd_connection *connection, struct packet_info *pi)
+{
+	void *buffer = connection->data.rbuf;
+	unsigned int size = drbd_header_size(connection);
+	int err;
+
+	err = drbd_recv_short(connection->data.socket, buffer, size, MSG_NOSIGNAL|MSG_DONTWAIT);
+	if (err != size) {
+		/* If we have nothing in the receive buffer now, to reduce
+		 * application latency, try to drain the backend queues as
+		 * quickly as possible, and let remote TCP know what we have
+		 * received so far. */
+		if (err == -EAGAIN) {
+			drbd_tcp_quickack(connection->data.socket);
+			drbd_unplug_all_devices(connection);
+		}
+		if (err > 0) {
+			buffer += err;
+			size -= err;
+		}
+		err = drbd_recv_all_warn(connection, buffer, size);
+		if (err)
+			return err;
+	}
+
+	err = decode_header(connection, connection->data.rbuf, pi);
+	connection->last_received = jiffies;
+
+	return err;
+}
 /* This is blkdev_issue_flush, but asynchronous.
  * We want to submit to all component volumes in parallel,
  * then wait for all completions.
@@ -4882,8 +4920,8 @@ static void drbdd(struct drbd_connection *connection)
 		struct data_cmd const *cmd;
 
 		drbd_thread_current_set_cpu(&connection->receiver);
-		update_receiver_timing_details(connection, drbd_recv_header);
-		if (drbd_recv_header(connection, &pi))
+		update_receiver_timing_details(connection, drbd_recv_header_maybe_unplug);
+		if (drbd_recv_header_maybe_unplug(connection, &pi))
 			goto err_out;
 
 		cmd = &drbd_cmd_handler[pi.cmd];
@@ -5375,8 +5413,11 @@ int drbd_receiver(struct drbd_thread *thi)
 		}
 	} while (h == 0);
 
-	if (h > 0)
+	if (h > 0) {
+		blk_start_plug(&connection->receiver_plug);
 		drbdd(connection);
+		blk_finish_plug(&connection->receiver_plug);
+	}
 
 	conn_disconnect(connection);