Skip to contentSkip to conversation

char-socket: disable reconnect timer in the sync connect

PATCH In review Wang Xin 2 participants 2 messages Dec 21, 2020 – Dec 23, 2020

lore
Share & tools
lore

Wang Xin

author
#0 lore raw

PATCH 1/1 char-socket: disable reconnect timer in the sync connect

From: suruifeng <[email protected]>

The qio_channel_socket_connect_sync maybe called twice if the openvswitchd restart during we attaching a vhost-user nic.

-> call trace 1:
net_vhost_user_init
tcp_chr_wait_connected //loop call sync connect until socekt connected
tcp_chr_connect_client_sync //return, but socekt state still disconnected
qio_channel_socket_connect_sync //socket connect sucess
tcp_chr_new_client
tcp_chr_connect
qemu_chr_be_event
net_vhost_user_event //CHR_EVENT_OPENED
vhost_user_start
tcp_chr_write //Broken Pipe, as peer restart
tcp_chr_disconnect_locked //disconnect & reconnect timer create

-> call trace 2:
socket_reconnect_timeout //timeout, and the peer restart just finished
tcp_chr_connect_client_async //concurrent with tcp_chr_connect_client_sync
tcp_chr_connect_client_task
qio_channel_socket_connect_sync //try connect same socket

This patch disabled tcp reconnect timer when we try to connect in synchronous mode, it seems to work.

chardev/char-socket.c +12 -1
index 213a4c8dd0..da1befca9e 100644
@@ -80,6 +80,7 @@ struct SocketChardev {
 
     bool is_websock;
 
+    bool async_reconnect_disable;
     GSource *reconnect_timer;
     int64_t reconnect_time;
     bool connect_err_reported;
@@ -506,7 +507,9 @@ static void tcp_chr_disconnect_locked(Chardev *chr)
     if (emit_close) {
         qemu_chr_be_event(chr, CHR_EVENT_CLOSED);
     }
-    if (s->reconnect_time && !s->reconnect_timer) {
+    if (s->reconnect_time &&
+        !s->reconnect_timer &&
+        !s->async_reconnect_disable) {
         qemu_chr_socket_restart_timer(chr);
     }
 }
@@ -954,15 +957,23 @@ static int tcp_chr_connect_client_sync(Chardev *chr, Error **errp)
 {
     SocketChardev *s = SOCKET_CHARDEV(chr);
     QIOChannelSocket *sioc = qio_channel_socket_new();
+
+    s->async_reconnect_disable = true;
     tcp_chr_change_state(s, TCP_CHARDEV_STATE_CONNECTING);
     tcp_chr_set_client_ioc_name(chr, sioc);
     if (qio_channel_socket_connect_sync(sioc, s->addr, errp) < 0) {
         tcp_chr_change_state(s, TCP_CHARDEV_STATE_DISCONNECTED);
         object_unref(OBJECT(sioc));
+        s->async_reconnect_disable = false;
         return -1;
     }
     tcp_chr_new_client(chr, sioc);
     object_unref(OBJECT(sioc));
+    s->async_reconnect_disable = false;
+
+    if (s->state != TCP_CHARDEV_STATE_CONNECTED) {
+        return -1;
+    }
     return 0;
 }
 
Signature · 1 line
2.26.0.windows.1
#1 lore raw

Hi

Quoted 20 lines · Wang Xin

From: suruifeng <[email protected]>

The qio_channel_socket_connect_sync maybe called twice if the openvswitchd restart during we attaching a vhost-user nic.

-> call trace 1:
net_vhost_user_init
tcp_chr_wait_connected //loop call sync connect until socekt connected
tcp_chr_connect_client_sync //return, but socekt state still
disconnected
qio_channel_socket_connect_sync //socket connect sucess
tcp_chr_new_client
tcp_chr_connect
qemu_chr_be_event
net_vhost_user_event //CHR_EVENT_OPENED
vhost_user_start
tcp_chr_write //Broken Pipe, as peer restart
tcp_chr_disconnect_locked //disconnect & reconnect
timer create

Quoted 5 lines

-> call trace 2:
socket_reconnect_timeout //timeout, and the peer restart just finished
tcp_chr_connect_client_async //concurrent with
tcp_chr_connect_client_sync

What do you mean by "concurrent with tcp_chr_connect_client_sync"? Are we talking about threads? If not, could you provide the full backtrace?

tcp_chr_connect_client_task

Quoted 1 line

qio_channel_socket_connect_sync //try connect same socket

This patch disabled tcp reconnect timer when we try to connect in

Quoted 2 lines

synchronous mode,

it seems to work.

Quoted 28 lines
chardev/char-socket.c +4 -1
index 213a4c8dd0..da1befca9e 100644
@@ -80,6 +80,7 @@ struct SocketChardev {
 
     bool is_websock;
 
+    bool async_reconnect_disable;
     GSource *reconnect_timer;
     int64_t reconnect_time;
     bool connect_err_reported;
@@ -506,7 +507,9 @@ static void tcp_chr_disconnect_locked(Chardev *chr)
     if (emit_close) {
         qemu_chr_be_event(chr, CHR_EVENT_CLOSED);
     }
-    if (s->reconnect_time && !s->reconnect_timer) {
+    if (s->reconnect_time &&
+        !s->reconnect_timer &&
+        !s->async_reconnect_disable) {
         qemu_chr_socket_restart_timer(chr);
     }
 }

chardev/char-socket.c static int tcp_chr_connect_client_sync(Chardev *chr, · line 963

Quoted 7 lines
chardev/char-socket.c +2 -0
@@ -954,15 +957,23 @@ static int tcp_chr_connect_client_sync(Chardev *chr,
 Error **errp)
 {
     SocketChardev *s = SOCKET_CHARDEV(chr);
     QIOChannelSocket *sioc = qio_channel_socket_new();
+
+    s->async_reconnect_disable = true;
 

Instead of having a new field, we could probably zero s->reconnect_timer temporarily.

     tcp_chr_change_state(s, TCP_CHARDEV_STATE_CONNECTING);
Quoted 5 lines
chardev/char-socket.c +1 -0
     tcp_chr_set_client_ioc_name(chr, sioc);
     if (qio_channel_socket_connect_sync(sioc, s->addr, errp) < 0) {
         tcp_chr_change_state(s, TCP_CHARDEV_STATE_DISCONNECTED);
         object_unref(OBJECT(sioc));
+        s->async_reconnect_disable = false;

chardev/char-socket.c static int tcp_chr_connect_client_sync(Chardev *chr, Error **errp) · line 979

Quoted 10 lines
chardev/char-socket.c +5 -0
         return -1;
     }
     tcp_chr_new_client(chr, sioc);
     object_unref(OBJECT(sioc));
+    s->async_reconnect_disable = false;
+
+    if (s->state != TCP_CHARDEV_STATE_CONNECTED) {
+        return -1;
+    }
 

How is this related?

     return 0;
Quoted 6 lines
chardev/char-socket.c +0 -1
 }
 
--

2.26.0.windows.1