lighttpd源码分析1:网络模型
拿到lighttpd的源码就迫不及待的想去掉繁杂的皮肉以窥其简单的网络模型框架。我们平常所写的TCP网络服务程序离不开这样的步骤:新建socket ——》将socket绑定到某个地址——》侦听客户端连接——》accept获取已连接socket——》读写已连接socket。Lighttpd不外如此。
lighttpd使用的是TCP预先派生子进程,每一个子进程各自accept的服务器设计范式,或者叫watcher-worker模型,关于各种网络程序设计范式在unix网络编程一书中有详细描述。整个程序的入口函数在server.c文件中,在main函数开始部分是各种繁杂的初始化工作,现在暂且略过,直接看到重点代码:
/*当是以root用户运行程序时,调用network_init函数*/ if(i_am_root) { … ... /* we need root-perms for port < 1024 */ if (0 != network_init(srv)) { plugins_free(srv); server_free(srv); return -1; } … ... }
network_init定义在network.c中,起初也是各种初始化工作,最后调用network_server_init,我们假设运行平台是ipv4(代码针对不同socket类型有不同的流程,为了化繁为简,只看ipv4流程),看下network_server_init流程的关键代码:
… … /*这是在创建socket*/ if (srv_socket->fd == -1) { srv_socket->addr.plain.sa_family = AF_INET; if (-1 == (srv_socket->fd = socket(srv_socket->addr.plain.sa_family, SOCK_STREAM, IPPROTO_TCP))) { log_error_write(srv, __FILE__, __LINE__, "ss", "socket failed:", strerror(errno)); goto error_free_socket; } } … … /*这是在初始化socket地址*/ case AF_INET: memset(&srv_socket->addr, 0, sizeof(struct sockaddr_in)); srv_socket->addr.ipv4.sin_family = AF_INET; if (host == NULL) { srv_socket->addr.ipv4.sin_addr.s_addr = htonl(INADDR_ANY); } else { struct hostent *he; if (NULL == (he = gethostbyname(host))) { log_error_write(srv, __FILE__, __LINE__, "sds", "gethostbyname failed: ", h_errno, host); goto error_free_socket; } if (he->h_addrtype != AF_INET) { log_error_write(srv, __FILE__, __LINE__, "sd", "addr-type != AF_INET: ", he->h_addrtype); goto error_free_socket; } if (he->h_length != sizeof(struct in_addr)) { log_error_write(srv, __FILE__, __LINE__, "sd", "addr-length != sizeof(in_addr): ", he->h_length); goto error_free_socket; } memcpy(&(srv_socket->addr.ipv4.sin_addr.s_addr), he->h_addr_list[0], he->h_length); } srv_socket->addr.ipv4.sin_port = htons(port); addr_len = sizeof(struct sockaddr_in); break; … … /*这是在绑定socket地址*/ if (0 != bind(srv_socket->fd, (struct sockaddr *) &(srv_socket->addr), addr_len)) { switch(srv_socket->addr.plain.sa_family) { case AF_UNIX: log_error_write(srv, __FILE__, __LINE__, "sds", "can't bind to socket:", host, strerror(errno)); break; default: log_error_write(srv, __FILE__, __LINE__, "ssds", "can't bind to port:", host, port, strerror(errno)); break; } goto error_free_socket; } … … /*这是在侦听*/ if (-1 == listen(srv_socket->fd, 128 * 8)) { log_error_write(srv, __FILE__, __LINE__, "ss", "listen failed: ", strerror(errno)); goto error_free_socket; } … …
一直到此处,lighttpd走的都是我们熟悉的流程。再回到main函数,来看下main中最重要的部分:
… ... /*父进程是watcher,fork出许多worker子进程,当子进程个数达到上限时,父进程进入等待*/ /*直到有子进程退出,父进程在while循环中运行中,一旦跳出while循环程序也结束了*/ /*子进程fork出老后跳出while,也就是后面代码都是子进程的流程。*/ /* start watcher and workers */ num_childs = srv->srvconf.max_worker; if (num_childs > 0) { int child = 0; while (!child && !srv_shutdown && !graceful_shutdown) { if (num_childs > 0) { switch (fork()) { case -1: return -1; case 0: child = 1; break; default: num_childs--; break; } } else { int status; if (-1 != wait(&status)) { /** * one of our workers went away */ num_childs++; } else { switch (errno) { case EINTR: /** * if we receive a SIGHUP we have to close our logs ourself as we don't * have the mainloop who can help us here */ if (handle_sig_hup) { handle_sig_hup = 0; log_error_cycle(srv); /** * forward to all procs in the process-group * * we also send it ourself */ if (!forwarded_sig_hup) { forwarded_sig_hup = 1; kill(0, SIGHUP); } } break; default: break; } } } } /** * for the parent this is the exit-point */ if (!child) { /** * kill all children too */ if (graceful_shutdown) { kill(0, SIGINT); } else if (srv_shutdown) { kill(0, SIGTERM); } log_error_close(srv); network_close(srv); connections_free(srv); plugins_free(srv); server_free(srv); return 0; } } … …
到此,我们知道父进程在固定端口上监听后预先fork了一定数量的子进程,子进程将会做什么呢?按照本文开头描述的应该是accept后读写socket了吧!看接下的代码是否如此:
… … /*fdevent系统的初始化,fdevent在lighttpd中主要处理各种IO事件,lighttpd采用的*/ /*是reactor模式,也就是多路复用加非阻塞式IO,而多路复用在各种平台上有差异,fdevent*/ /*通过OO的方法封装了各个不同实现,以使得代码中可以使用统一的接口*/ if (NULL == (srv->ev = fdevent_init(srv, srv->max_fds + 1, srv->event_handler))) { log_error_write(srv, __FILE__, __LINE__, "s", "fdevent_init failed"); return -1; } /*注册srv中保存的socket到fdevent中*/ /* * kqueue() is called here, select resets its internals, * all server sockets get their handlers * * */ if (0 != network_register_fdevents(srv)) { plugins_free(srv); network_close(srv); server_free(srv); return -1; } … …
函数network_register_fdevents在network.c中定义,代码如下:
int network_register_fdevents(server *srv) { size_t i; /*清除fdevent的IO句柄,如同select的FD_ZERO清除fd set*/ if (-1 == fdevent_reset(srv->ev)) { return -1; } /* register fdevents after reset */ for (i = 0; i < srv->srv_sockets.used; i++) { server_socket *srv_socket = srv->srv_sockets.ptr[i]; //注册回调函数 //一旦srv_socket->fd就绪,则触发函数 network_server_handle_fdevent fdevent_register(srv->ev, srv_socket->fd, network_server_handle_fdevent, srv_socket); //告诉fdevent观察srv_socket->fd,一旦可读,则调用相应回调函数。 fdevent_event_set(srv->ev, &(srv_socket->fde_ndx), srv_socket->fd, FDEVENT_IN); } return 0; }
这里的srv_socket->fd其实就是之前创建的监听套接字,至此,我们假设有一个客户连接请求过来,这时子进程的srv_socket->fd 可读,回调函数network_server_handle_fdevent被调用:
static handler_t network_server_handle_fdevent(server *srv, void *context, int revents) { … ... /* accept()s at most 100 connections directly * * we jump out after 100 to give the waiting connections a chance */ for (loops = 0; loops < 100 && NULL != (con = connection_accept(srv, srv_socket)); loops++) { handler_t r; connection_state_machine(srv, con); switch(r = plugins_call_handle_joblist(srv, con)) { case HANDLER_FINISHED: case HANDLER_GO_ON: break; default: log_error_write(srv, __FILE__, __LINE__, "d", r); break; } } return HANDLER_GO_ON; } connection_accept在connections.c中定义,代码简化为如下: … … //获取已连接套接字 if (-1 == (cnt = accept(srv_socket->fd, (struct sockaddr *) &cnt_addr, &cnt_len))) { switch (errno) { case EAGAIN: #if EWOULDBLOCK != EAGAIN case EWOULDBLOCK: #endif case EINTR: /* we were stopped _before_ we had a connection */ case ECONNABORTED: /* this is a FreeBSD thingy */ /* we were stopped _after_ we had a connection */ break; case EMFILE: /* out of fds */ break; default: log_error_write(srv, __FILE__, __LINE__, "ssd", "accept failed:", strerror(errno), errno); } return NULL; } … … con->fd = cnt; con->fde_ndx = -1; //在fdevent中注册已连接socket : con->fd的回调函数connection_handle_fdevent fdevent_register(srv->ev, con->fd, connection_handle_fdevent, con); … … //设置一些属性,比如将con->fd设置为非阻塞的 if (-1 == (fdevent_fcntl_set(srv->ev, con->fd))) { log_error_write(srv, __FILE__, __LINE__, "ss", "fcntl failed: ", strerror(errno)); return NULL; } … …
分析到了这个地方,lighttpd的网络模型框架大致清楚了,正如文首所述,它和所有网络服务器程序一样都要走socket->bind->listen->accept流程,更具体的说,它使用了预先创建子进程,各子进程各自accept的范式,在UNIX网络编程中说这种范式会有accept惊群的问题,即当监听套接字可读,所有accept的子进程都会醒过来,但是只有一个进程获得已连接套接字,所有进程都唤醒是没有必要的,这样影响效率。对于这个问题,lighttpd似乎并没有处理。但是在新的linux内核中已经不存在accept惊群现象了。不过对于多路复用函数如select,epoll仍然存在类似问题,而代码里时常是先调epoll(select),再accept,lighttpd就是如此,因此还是会有新的惊群现象需要处理。如果不是我遗漏了的话,我没有发现lighttpd有相关代码对此进行处理,而nginx却有相关处理。