net: sent udp open messages
The receiver goes directly into recvfrom() when it starts up,
which greatly skews the first submission latency (since it's
measure from when it enters recvfrom() and until we finally
get the first packet from the client). Add a file open
handshake to avoid this issue.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
diff --git a/engines/net.c b/engines/net.c
index 373821b..fb554a4 100644
--- a/engines/net.c
+++ b/engines/net.c
@@ -42,7 +42,8 @@
enum {
FIO_LINK_CLOSE = 0x89,
- FIO_LINK_CLOSE_MAGIC = 0x6c696e6b,
+ FIO_LINK_OPEN_CLOSE_MAGIC = 0x6c696e6b,
+ FIO_LINK_OPEN = 0x98,
FIO_TYPE_TCP = 1,
FIO_TYPE_UDP = 2,
@@ -329,7 +330,7 @@
return 0;
msg = io_u->xfer_buf;
- if (ntohl(msg->magic) != FIO_LINK_CLOSE_MAGIC)
+ if (ntohl(msg->magic) != FIO_LINK_OPEN_CLOSE_MAGIC)
return 0;
if (ntohl(msg->cmd) != FIO_LINK_CLOSE)
return 0;
@@ -503,21 +504,6 @@
return 1;
}
-static int fio_netio_open_file(struct thread_data *td, struct fio_file *f)
-{
- int ret;
- struct netio_options *o = td->eo;
-
- if (o->listen)
- ret = fio_netio_accept(td, f);
- else
- ret = fio_netio_connect(td, f);
-
- if (ret)
- f->fd = -1;
- return ret;
-}
-
static void fio_netio_udp_close(struct thread_data *td, struct fio_file *f)
{
struct netio_data *nd = td->io_ops->data;
@@ -525,7 +511,7 @@
struct sockaddr *to = (struct sockaddr *) &nd->addr;
int ret;
- msg.magic = htonl(FIO_LINK_CLOSE_MAGIC);
+ msg.magic = htonl(FIO_LINK_OPEN_CLOSE_MAGIC);
msg.cmd = htonl(FIO_LINK_CLOSE);
ret = sendto(f->fd, &msg, sizeof(msg), MSG_WAITALL, to,
@@ -548,6 +534,84 @@
return generic_close_file(td, f);
}
+static int fio_netio_udp_recv_open(struct thread_data *td, struct fio_file *f)
+{
+ struct netio_data *nd = td->io_ops->data;
+ struct udp_close_msg msg;
+ struct sockaddr *to = (struct sockaddr *) &nd->addr;
+ fio_socklen_t len = sizeof(nd->addr);
+ int ret;
+
+ ret = recvfrom(f->fd, &msg, sizeof(msg), MSG_WAITALL, to, &len);
+ if (ret < 0) {
+ td_verror(td, errno, "sendto udp link open");
+ return ret;
+ }
+
+ if (ntohl(msg.magic) != FIO_LINK_OPEN_CLOSE_MAGIC ||
+ ntohl(msg.cmd) != FIO_LINK_OPEN) {
+ log_err("fio: bad udp open magic %x/%x\n", ntohl(msg.magic),
+ ntohl(msg.cmd));
+ return -1;
+ }
+
+ return 0;
+}
+
+static int fio_netio_udp_send_open(struct thread_data *td, struct fio_file *f)
+{
+ struct netio_data *nd = td->io_ops->data;
+ struct udp_close_msg msg;
+ struct sockaddr *to = (struct sockaddr *) &nd->addr;
+ int ret;
+
+ msg.magic = htonl(FIO_LINK_OPEN_CLOSE_MAGIC);
+ msg.cmd = htonl(FIO_LINK_OPEN);
+
+ ret = sendto(f->fd, &msg, sizeof(msg), MSG_WAITALL, to,
+ sizeof(nd->addr));
+ if (ret < 0) {
+ td_verror(td, errno, "sendto udp link open");
+ return ret;
+ }
+
+ return 0;
+}
+
+static int fio_netio_open_file(struct thread_data *td, struct fio_file *f)
+{
+ int ret;
+ struct netio_options *o = td->eo;
+
+ if (o->listen)
+ ret = fio_netio_accept(td, f);
+ else
+ ret = fio_netio_connect(td, f);
+
+ if (ret) {
+ f->fd = -1;
+ return ret;
+ }
+
+ if (o->proto == FIO_TYPE_UDP) {
+ if (td_write(td))
+ ret = fio_netio_udp_send_open(td, f);
+ else {
+ int state;
+
+ state = td->runstate;
+ td_set_runstate(td, TD_SETTING_UP);
+ ret = fio_netio_udp_recv_open(td, f);
+ td_set_runstate(td, state);
+ }
+ }
+
+ if (ret)
+ fio_netio_close_file(td, f);
+
+ return ret;
+}
+
static int fio_netio_setup_connect_inet(struct thread_data *td,
const char *host, unsigned short port)
{