Nginx 因其出色的处理并发连接的能力,越来越多地作为一个反向代理服务器被使用。所谓反向代理,即把Nginx置于用户接入的最前端,监听用户发来的请求,并把 它们转发给相应的后端服务器来处理具体的请求。后端服务器可以是缓存服务器(如Squid)或是处理动态/静态请求的服务器(如apache/Nginx /lighttpd),在这里不作深入讨论。本文对Nginx在upstream服务器的设置和"proxy_pass"的功能在代码层面进行分析,并讨 论在实际生产环境的出现的问题。
upstream模块主要有两个命令,"upstream"和"server"。在配置文件里面,大概的结构是这样的:
- http{
- #...
- upstream backend_name{
- server xx.xx.xx.xx:xx weight=2 max_fails=3;
- server www.xxx.com weight=1;
- server unix://xxx/xxx;
- #...
- }
- #...
- }
其中在"http"命令被解析的时候,会调用ngx_http_block()函数,而在这个函数里面会再次解析配置文件,从而发现"upstream"命令,调用它的set(),即ngx_http_upstream()函数。这个函数是这样的:
- ngx_http_upstream(...)
- {
- //获取upstream服务器组的名字,即upstream命令的参数
- //...
- //这个函数初始化一个ngx_http_upstream_srv_conf_t类型,并存入umcf->upstreams数组里面
- //注意:这个函数还有一个作用,就是在"proxy_pass"的时候从umcf->upstreams找出match的upstream服务器组(不知道为啥要放在同一个函数里><)
- uscf = ngx_http_upstream_add(...);
- ctx = ngx_pcalloc(cf->pool, sizeof(ngx_http_conf_ctx_t));
- //继承main_conf
- http_ctx = cf->ctx;
- ctx->main_conf = http_ctx->main_conf;
- //创建新的srv_conf
- ctx->srv_conf = ngx_pcalloc(cf->pool, sizeof(void *) * ngx_http_max_module);
- //把ngx_http_upstream_add()创建的ngx_http_upstream_srv_conf_t存入srv_conf
- ctx->srv_conf[ngx_http_upstream_module.ctx_index] = uscf;
- //虽然uscf本身是一个srv_conf,但是它的成员中也有srv_conf,指向全局的srv_conf
- uscf->srv_conf = ctx->srv_conf;
- //创建loc_conf
- ctx->loc_conf = ngx_pcalloc(cf->pool, sizeof(void *) * ngx_http_max_module);
- //for each NGX_HTTP_MODULE
- for (;;){
- //if this module has create_srv_conf() ...
- mconf = module->create_srv_conf(cf);
- ctx->srv_conf[ngx_modules[m]->ctx_index] = mconf;
- //if this module has create_loc_conf() ...
- mconf = module->create_loc_conf(cf);
- ctx->loc_conf[ngx_modules[m]->ctx_index] = mconf;
- }
- cf->ctx = ctx;
- cf->cmd_type = NGX_HTTP_UPS_CONF;
- //继续解析server的配置
- rv = ngx_conf_parse(cf, NULL);
- }
继续解析"server"的配置行,每行调用一次:
- ngx_http_upstream_server(...)
- {
- ngx_http_upstream_server_t *us;
- us = ngx_array_push(uscf->servers);
- //获取第一个参数,即url
- u.url = value[1];
- u.default_port = 80;
- ngx_parse_url(cf->pool, &u);
- //解析"server"后面的其他参数
- //...
- //存入us
- us->addrs = u.addrs;
- us->naddrs = u.naddrs;
- us->weight = weight;
- us->max_fails = max_fails;
- us->fail_timeout = fail_timeout;
- }
在ngx_http_core_module的server {...} block的 /location {...} block里面,如果有"proxy_pass"指令,就会调用它的set(),即ngx_http_proxy_pass():
- ngx_http_proxy_pass(...)
- {
- ngx_http_proxy_loc_conf_t *plcf = conf;
- //获取当前的location,即在哪个location配置的"proxy_pass"指令
- clcf = ngx_http_conf_get_module_loc_conf(cf, ngx_http_core_module);
- // 设置loc的handler,这个clcf->handler会在ngx_http_update_location_config()里面赋予 r->content_handler,从而在NGX_HTTP_CONTENT_PHASE里面调用这个handler,即 ngx_http_proxy_handler。
- clcf->handler = ngx_http_proxy_handler;
- //如果"proxy_pass"第一个参数(即url)里面的变量
- //...
- //根据http/https设置port和add(位移)
- //...
- u.url.len = url->len - add;//设置url的长度,除去http://(https://)
- u.url.data = url->data + add;//设置url,除去http://(https://),比如原来是"http://backend1",现在就是"backend1"
- u.default_port = port;//默认port
- u.uri_part = 1;
- u.no_resolve = 1;//不要resolve这个url的域名
- //这儿从已经定义的upstream服务器组里面找到名字match的组
- plcf->upstream.upstream = ngx_http_upstream_add(cf, &u, 0);
- //设置plcf的成员
- //...
- //设置location的名字
- plcf->location = clcf->name;
- //...
- }
那么,当有请求访问到特定的location的时候(假设这个location配置了proxy_pass指令),跟其他请求一样,会调用各个 phase的checker和handler,到了NGX_HTTP_CONTENT_PHASE的checker,即 ngx_http_core_content_phase()的时候,会调用r->content_handler(r),即
ngx_http_proxy_handler
。
- ngx_http_proxy_handler(r)
- {
- ngx_http_upstream_t *u;
- //生成一个新的ngx_http_upstream_t,赋值给r->upstream
- //...
- ngx_http_upstream_create(r);
- //设置r->ctx[ngx_http_proxy_module.ctx_index] = ctx
- ctx = ngx_pcalloc(r->pool, sizeof(ngx_http_proxy_ctx_t));
- ngx_http_set_ctx(r, ctx, ngx_http_proxy_module);
- //获取ngx_http_proxy_loc_conf_t
- plcf = ngx_http_get_module_loc_conf(r, ngx_http_proxy_module);
- //...
- //获取upstream server的信息,如标签、服务器ip、port等
- u->conf = &plcf->upstream;
- //生成发送到上游服务器的请求缓冲(或者一条缓冲链)
- u->create_request = ngx_http_proxy_create_request;
- //在后端服务器被重置的情况下(在create_request被第二次调用之前)被调用
- u->reinit_request = ngx_http_proxy_reinit_request;
- //处理上游服务器回复的第一个bit,时常是保存一个指向上游回复负载的指针
- u->process_header = ngx_http_proxy_process_status_line;
- //在客户端放弃请求的时候被调用
- u->abort_request = ngx_http_proxy_abort_request;
- //在Nginx完成从上游服务器读入回复以后被调用
- u->finalize_request = ngx_http_proxy_finalize_request;
- //...
- u->pipe = ngx_pcalloc(r->pool, sizeof(ngx_event_pipe_t));
- u->pipe->input_filter = ngx_event_pipe_copy_input_filter;
- ngx_http_read_client_request_body(r, ngx_http_upstream_init);
- //没有special response
- return NGX_DONE;//-4
- }
ngx_http_read_client_request_body()里面会调用post_handler()即ngx_http_upstream_init():
- ngx_http_upstream_init(r)
- {
- //...
- //如果是edge trigger事件机制(epoll/kqueue),添加一个NGX_WRITE_EVENT
- //这个事件用来检查与用户的连接是否断开
- ngx_add_event(c->write, NGX_WRITE_EVENT, NGX_CLEAR_EVENT);
- ngx_http_upstream_init_request(r);
- }
- ngx_http_upstream_init_request(r)
- {
- ngx_http_upstream_t *u;
- u = r->upstream;
- //设置r->read_event_handler和r->write_event_handler为事件调用函数
- r->read_event_handler = ngx_http_upstream_rd_check_broken_connection;
- r->write_event_handler = ngx_http_upstream_wr_check_broken_connection;
- //把用户来的request内容(如请求行,headers)放到r->upstream的成员变量中
- //如r->upstream->url存放请求的url
- //把请求(包括请求行,headers)放入r->upstream->request_bufs这个buf chain
- u->create_request(r);
- clcf = ngx_http_get_module_loc_conf(r, ngx_http_core_module);
- //初始化u->output(ngx_output_chain_ctx_s类型)
- u->output.alignment = clcf->directio_alignment;
- u->output.pool = r->pool;
- u->output.bufs.num = 1;
- u->output.bufs.size = clcf->client_body_buffer_size;
- u->output.output_filter = ngx_chain_writer;
- u->output.filter_ctx = &u->writer;
- u->writer.pool = r->pool;
- //创建state(ngx_http_upstream_state_t类型)
- //...
- //添加一个cleanup处理方法
- ngx_http_cleanup_add(r,0);
- cln->handler = ngx_http_upstream_cleanup;
- cln->data = r;
- u->cleanup = &cln->handler;
- //获取ngx_http_upstream_srv_conf_t
- uscf = u->conf->upstream;
- ngx_http_upstream_connect(r,u);
- }
- ngx_http_upstream_connect(ngx_http_request_t *r, ngx_http_upstream_t *u)
- {
- u->state = ngx_array_push(r->upstream_states);
- //获取时间,存入u->state
- //...
- //创建一个socket,调用ngx_get_connection()获取一个空闲的connection并初始化
- //把这个connection添加到事件监听,并bind()以及connect()到peer
- //不等建立连接,直接返回,一般返回值EINPROGRESS,表示正在进行中
- rc = ngx_event_connect_peer(&u->peer);
- c = u->peer.connection;
- c->data = r;
- c->write->handler = ngx_http_upstream_handler;
- c->read->handler = ngx_http_upstream_handler;
- u->write_event_handler = ngx_http_upstream_send_request_handler;
- u->read_event_handler = ngx_http_upstream_process_header;
- c->sendfile &= r->connection->sendfile;
- u->output.sendfile = c->sendfile;
- //init or reinit the ngx_output_chain() and ngx_chain_writer() contexts
- //...
- //rc == NGX_AGAIN,连接进行中
- ngx_add_timer(c->write, u->conf->connect_timeout);
- //rc == NGX_OK,连接成功
- ngx_http_upstream_send_request();
- }
返回之后就进入了ngx_http_finalize_request() --> ngx_http_finalize_connection() --> ngx_http_close_request(),但是不会关闭客户端的连接。然后触发一个EPOLLOUT事件(发生在和用户的连 接),进入ngx_http_request_handler()。这个事件是在ngx_http_upstream_init()里面添加的。而这个事 件处理函数在ngx_http_process_request()里面设置。这个函数很简单:
- ngx_http_request_handler(ngx_event_t *ev)
- {
- c = ev->data;
- r = c->data;
- //if...
- r->write_event_handler(r);
- //else...
- r->read_event_handler(r);
- //subrequest...
- ngx_http_run_posted_requests(c);
- }
ngx_http_upstream_init_request()里面设置了r->read_event_handler 和r->write_event_handler,它们都指向 ngx_http_upstream_check_broken_connection,那么在上面的函数里,就会调用这个函数。它主要是检查与用户的连 接是否断开。
- ngx_http_upstream_check_broken_connection(r,ev)
- {
- c = r->connection;
- u = r->upstream;
- n = recv(c->fd, buf, 1, MSG_PEEK);
- //...
- //正常err应该是NGX_EAGAIN(11)
- err = ngx_socket_errno;
- //...
- }
处理完之后,另外一个事件也同时触发(在epoll处理事件循环里),是之前跟upstream建立的连接的写事件(向upstream服务器 发送)。调用c->write->handler即ngx_http_upstream_handler()。这个是在之前 ngx_http_upstream_connect()里面设置的:
- ngx_http_upstream_handler(ev)
- {
- c = ev->data;
- r = c->data;
- u = r->upstream;
- c = r->connection;
- //if...
- u->write_event_handler(r, u);
- //else...
- u->read_event_handler(r, u);
- //subrequest...
- ngx_http_run_posted_requests(c);
- }
从
ngx_http_upstream_connect(),我们知道
u->write_event_handler = ngx_http_upstream_send_request_handler,这个函数主要调用了ngx_http_upstream_send_request()函数:
- ngx_http_upstream_send_request(r,u)
- {
- c = u->peer.connection;
- //向后端发送请求
- rc = ngx_output_chain(&u->output, u->request_sent ? NULL : u->request_bufs);
- u->write_event_handler = ngx_http_upstream_dummy_handler;
- }
当后端(upstream)服务器处理完请求之后,会发回响应,这个时候事件被触发。再次调用
ngx_http_upstream_handler(),这次调用
u->read_event_handler,即
ngx_http_upstream_process_header():
- ngx_http_upstream_process_header(r,u)
- {
- //...
- for ( ;; ) {
- n = c->recv(c, u->buffer.last, u->buffer.end - u->buffer.last);
- //u->process_header = ngx_http_proxy_process_status_line
- //还调用了ngx_http_proxy_process_header(),处理回复的headers
- rc = u->process_header(r);
- }
- ngx_http_upstream_process_headers();
- //如果subrequest_in_memory == 0
- ngx_http_upstream_send_response(r, u);
- //else...
- //...
- }
- ngx_http_upstream_send_response(r,u)
- {
- //调用ngx_http_top_header_filter即ngx_http_header_filter()
- //ngx_http_header_filter()中会调用ngx_http_write_filter()
- //ngx_http_write_filter()遍历所有chain,然后输出所有数据
- rc = ngx_http_send_header(r);
- //...
- u->read_event_handler = ngx_http_upstream_process_upstream;
- r->write_event_handler = ngx_http_upstream_process_downstream;
- ngx_http_upstream_process_upstream(r, u);
- }
- ngx_http_upstream_process_upstream(r,u)
- {
- c = u->peer.connection;
- //如果没有timeout
- //
- ngx_event_pipe(u->pipe, 0);
- }