epoll 断联之后，可以继续重连

关联测试 : docs/kernel/fs/epoll/epoll-reconnect.c

Socket disconnected (EPOLLERR/EPOLLHUP), attempting to reconnect...
Socket disconnected (EPOLLERR/EPOLLHUP), attempting to reconnect...
Socket disconnected (EPOLLERR/EPOLLHUP), attempting to reconnect...
Socket disconnected (EPOLLERR/EPOLLHUP), attempting to reconnect...
Socket disconnected (EPOLLERR/EPOLLHUP), attempting to reconnect...
Socket disconnected (EPOLLERR/EPOLLHUP), attempting to reconnect...
Socket disconnected (EPOLLERR/EPOLLHUP), attempting to reconnect...
Socket disconnected (EPOLLERR/EPOLLHUP), attempting to reconnect...
Socket disconnected (EPOLLERR/EPOLLHUP), attempting to reconnect...
client can connect to a broken unix domain socket
client can connect to a broken unix domain socket
client reconnect failed
client reconnect failed
client reconnect failed
client reconnect failed
client reconnect failed
client reconnect failed
client reconnect failed
Server process terminated

关闭 socket 的基本流程为:

exit_files()                                      →  kernel/exit.c
    └─> __fput()                                  →  fs/file_table.c
        └─> sock_close()                          →  net/socket.c
            └─> sock_release()                    →  net/socket.c
                └─> sk->sk_prot->release()        →  对于 AF_UNIX: unix_release() → net/unix/af_unix.c:1330
                    └─> unix_release_sock(sk)
                        ├─> unix_state_lock(peer_sk)
                        ├─> peer_sk->sk_state = TCP_CLOSE
                        ├─> unix_state_unlock(peer_sk)
                        └─> sk_wake_async(peer_sk, SOCK_WAKE_WAITD, POLLHUP)   ← 关键！
                            └─> sock_def_readable()         →  net/core/sock.c:2430
                                ├─> __wake_up_sync(&sk->sk_sleep, TASK_NORMAL, 1)
                                └─> wake_up_interruptible_poll(&wq_head, EPOLLIN | EPOLLHUP)

unix_release_sock 中会调用 unix_remove_bsd_socket(sk) ，将 socket 从 hash table 中移除掉。

如果重连，connect syscall 会调用到 unix_find_socket_byinode 中，会由于没有查询到，那么 connect syscall 失败。

当 server 关闭的时候，会自动关闭掉所有的 socekt :

1. server 的 listen socket ，通过 connect listen socket 来建立连接，获取 socket 通信
2. 和 client 通信的 socket ，这个会在 unix_release_sock 中调用 sk_wake_async ，让 client 的 epoll 接受到信号。

server 关闭 socket 的顺序是未定义的， 如果不去先 close server listen socket ，当 server 关闭掉 和 client 通信的 socket ， 那么 client 就会收到 server 挂掉的消息，然后发起重连。

解决办法:

1. 对于可控重连，在 server 结束的时候，关闭掉 server socket
2. 对于不可控重连，qemu 并不会存在立刻来连接，而是会等待一段时间，qemu 不会使用 epoll 来监听:

    ``reconnect-ms`` sets the timeout for reconnecting on non-server
    sockets when the remote end goes away. qemu will delay this many
    milliseconds and then attempt to reconnect. Zero disables reconnecting,
    and is the default.
   

本站所有文章转发 CSDN 将按侵权追究法律责任，其它情况随意。