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>

drivers/block/drbd/drbd_int.h

@@ -745,6 +745,8 @@ struct drbd_connection {
 	unsigned current_tle_writes;	/* writes seen within this tl epoch */
 
 	unsigned long last_reconnect_jif;
+	/* empty member on older kernels without blk_start_plug() */
+	struct blk_plug receiver_plug;
 	struct drbd_thread receiver;
 	struct drbd_thread worker;
 	struct drbd_thread ack_receiver;
@@ -1131,7 +1133,8 @@ extern void conn_send_sr_reply(struct drbd_connection *connection, enum drbd_sta
 extern int drbd_send_rs_deallocated(struct drbd_peer_device *, struct drbd_peer_request *);
 extern void drbd_backing_dev_free(struct drbd_device *device, struct drbd_backing_dev *ldev);
 extern void drbd_device_cleanup(struct drbd_device *device);
-void drbd_print_uuids(struct drbd_device *device, const char *text);
+extern void drbd_print_uuids(struct drbd_device *device, const char *text);
+extern void drbd_queue_unplug(struct drbd_device *device);
 
 extern void conn_md_sync(struct drbd_connection *connection);
 extern void drbd_md_write(struct drbd_device *device, void *buffer);

drivers/block/drbd/drbd_main.c

@@ -1952,6 +1952,19 @@ static void drbd_release(struct gendisk *gd, fmode_t mode)
 	mutex_unlock(&drbd_main_mutex);
 }
 
+/* need to hold resource->req_lock */
+void drbd_queue_unplug(struct drbd_device *device)
+{
+	if (device->state.pdsk >= D_INCONSISTENT && device->state.conn >= C_CONNECTED) {
+		D_ASSERT(device, device->state.role == R_PRIMARY);
+		if (test_and_clear_bit(UNPLUG_REMOTE, &device->flags)) {
+			drbd_queue_work_if_unqueued(
+				&first_peer_device(device)->connection->sender_work,
+				&device->unplug_work);
+		}
+	}
+}
+
 static void drbd_set_defaults(struct drbd_device *device)
 {
 	/* Beware! The actual layout differs

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);
 

drivers/block/drbd/drbd_req.c

@@ -1279,6 +1279,56 @@ static bool may_do_writes(struct drbd_device *device)
 	return s.disk == D_UP_TO_DATE || s.pdsk == D_UP_TO_DATE;
 }
 
+struct drbd_plug_cb {
+	struct blk_plug_cb cb;
+	struct drbd_request *most_recent_req;
+	/* do we need more? */
+};
+
+static void drbd_unplug(struct blk_plug_cb *cb, bool from_schedule)
+{
+	struct drbd_plug_cb *plug = container_of(cb, struct drbd_plug_cb, cb);
+	struct drbd_resource *resource = plug->cb.data;
+	struct drbd_request *req = plug->most_recent_req;
+
+	if (!req)
+		return;
+
+	spin_lock_irq(&resource->req_lock);
+	/* In case the sender did not process it yet, raise the flag to
+	 * have it followed with P_UNPLUG_REMOTE just after. */
+	req->rq_state |= RQ_UNPLUG;
+	/* but also queue a generic unplug */
+	drbd_queue_unplug(req->device);
+	spin_unlock_irq(&resource->req_lock);
+	kref_put(&req->kref, drbd_req_destroy);
+}
+
+static struct drbd_plug_cb* drbd_check_plugged(struct drbd_resource *resource)
+{
+	/* A lot of text to say
+	 * return (struct drbd_plug_cb*)blk_check_plugged(); */
+	struct drbd_plug_cb *plug;
+	struct blk_plug_cb *cb = blk_check_plugged(drbd_unplug, resource, sizeof(*plug));
+
+	if (cb)
+		plug = container_of(cb, struct drbd_plug_cb, cb);
+	else
+		plug = NULL;
+	return plug;
+}
+
+static void drbd_update_plug(struct drbd_plug_cb *plug, struct drbd_request *req)
+{
+	struct drbd_request *tmp = plug->most_recent_req;
+	/* Will be sent to some peer.
+	 * Remember to tag it with UNPLUG_REMOTE on unplug */
+	kref_get(&req->kref);
+	plug->most_recent_req = req;
+	if (tmp)
+		kref_put(&tmp->kref, drbd_req_destroy);
+}
+
 static void drbd_send_and_submit(struct drbd_device *device, struct drbd_request *req)
 {
 	struct drbd_resource *resource = device->resource;
@@ -1287,6 +1337,8 @@ static void drbd_send_and_submit(struct drbd_device *device, struct drbd_request
 	bool no_remote = false;
 	bool submit_private_bio = false;
 
+	struct drbd_plug_cb *plug = drbd_check_plugged(resource);
+
 	spin_lock_irq(&resource->req_lock);
 	if (rw == WRITE) {
 		/* This may temporarily give up the req_lock,
@@ -1351,6 +1403,9 @@ static void drbd_send_and_submit(struct drbd_device *device, struct drbd_request
 			no_remote = true;
 	}
 
+	if (plug != NULL && no_remote == false)
+		drbd_update_plug(plug, req);
+
 	/* If it took the fast path in drbd_request_prepare, add it here.
 	 * The slow path has added it already. */
 	if (list_empty(&req->req_pending_master_completion))

drivers/block/drbd/drbd_req.h

@@ -212,6 +212,11 @@ enum drbd_req_state_bits {
 	/* Should call drbd_al_complete_io() for this request... */
 	__RQ_IN_ACT_LOG,
 
+	/* This was the most recent request during some blk_finish_plug()
+	 * or its implicit from-schedule equivalent.
+	 * We may use it as hint to send a P_UNPLUG_REMOTE */
+	__RQ_UNPLUG,
+
 	/* The peer has sent a retry ACK */
 	__RQ_POSTPONED,
 
@@ -249,6 +254,7 @@ enum drbd_req_state_bits {
 #define RQ_WSAME           (1UL << __RQ_WSAME)
 #define RQ_UNMAP           (1UL << __RQ_UNMAP)
 #define RQ_IN_ACT_LOG      (1UL << __RQ_IN_ACT_LOG)
+#define RQ_UNPLUG          (1UL << __RQ_UNPLUG)
 #define RQ_POSTPONED	   (1UL << __RQ_POSTPONED)
 #define RQ_COMPLETION_SUSP (1UL << __RQ_COMPLETION_SUSP)
 #define RQ_EXP_RECEIVE_ACK (1UL << __RQ_EXP_RECEIVE_ACK)

drivers/block/drbd/drbd_worker.c

@@ -1382,18 +1382,22 @@ static int drbd_send_barrier(struct drbd_connection *connection)
 	return conn_send_command(connection, sock, P_BARRIER, sizeof(*p), NULL, 0);
 }
 
+static int pd_send_unplug_remote(struct drbd_peer_device *pd)
+{
+	struct drbd_socket *sock = &pd->connection->data;
+	if (!drbd_prepare_command(pd, sock))
+		return -EIO;
+	return drbd_send_command(pd, sock, P_UNPLUG_REMOTE, 0, NULL, 0);
+}
+
 int w_send_write_hint(struct drbd_work *w, int cancel)
 {
 	struct drbd_device *device =
 		container_of(w, struct drbd_device, unplug_work);
-	struct drbd_socket *sock;
 
 	if (cancel)
 		return 0;
-	sock = &first_peer_device(device)->connection->data;
-	if (!drbd_prepare_command(first_peer_device(device), sock))
-		return -EIO;
-	return drbd_send_command(first_peer_device(device), sock, P_UNPLUG_REMOTE, 0, NULL, 0);
+	return pd_send_unplug_remote(first_peer_device(device));
 }
 
 static void re_init_if_first_write(struct drbd_connection *connection, unsigned int epoch)
@@ -1455,6 +1459,7 @@ int w_send_dblock(struct drbd_work *w, int cancel)
 	struct drbd_device *device = req->device;
 	struct drbd_peer_device *const peer_device = first_peer_device(device);
 	struct drbd_connection *connection = peer_device->connection;
+	bool do_send_unplug = req->rq_state & RQ_UNPLUG;
 	int err;
 
 	if (unlikely(cancel)) {
@@ -1470,6 +1475,9 @@ int w_send_dblock(struct drbd_work *w, int cancel)
 	err = drbd_send_dblock(peer_device, req);
 	req_mod(req, err ? SEND_FAILED : HANDED_OVER_TO_NETWORK);
 
+	if (do_send_unplug && !err)
+		pd_send_unplug_remote(peer_device);
+
 	return err;
 }
 
@@ -1484,6 +1492,7 @@ int w_send_read_req(struct drbd_work *w, int cancel)
 	struct drbd_device *device = req->device;
 	struct drbd_peer_device *const peer_device = first_peer_device(device);
 	struct drbd_connection *connection = peer_device->connection;
+	bool do_send_unplug = req->rq_state & RQ_UNPLUG;
 	int err;
 
 	if (unlikely(cancel)) {
@@ -1501,6 +1510,9 @@ int w_send_read_req(struct drbd_work *w, int cancel)
 
 	req_mod(req, err ? SEND_FAILED : HANDED_OVER_TO_NETWORK);
 
+	if (do_send_unplug && !err)
+		pd_send_unplug_remote(peer_device);
+
 	return err;
 }