随笔-156  评论-223  文章-30  trackbacks-0
   nginx的域名解析器使用已连接udp(收发前先调用ngx_udp_connect)发送dns查询、接收dns响应,如上篇tcp异步连接所讲,iocp需要先投递udp的接收操作,才能引发接收完成的事件,因此要对域名解析器和udp异步接收作些改进。

发送后投递
    dns查询由ngx_resolver_send_query函数实现,定义在core/ngx_resolver.c中。
 1static ngx_int_t ngx_resolver_send_query(ngx_resolver_t *r, ngx_resolver_node_t *rn)
 2{
 3    
 4     if (rn->naddrs == (u_short) -1{
 5        n = ngx_send(uc->connection, rn->query, rn->qlen);
 6    
 7  }

 8
 9#if (NGX_HAVE_INET6)
10  if (rn->query6 && rn->naddrs6 == (u_short) -1{
11      n = ngx_send(uc->connection, rn->query6, rn->qlen);
12    
13  }

14#endif
15
16#if (NGX_WIN32) 
17    if (ngx_event_flags & NGX_USE_IOCP_EVENT){
18        uc->connection->read->ready = 1;
19        ngx_resolver_read_response(uc->connection->read);
20}

21#endif
22
23  return NGX_OK;
24}
    当nginx用于代理连接上游服务器前,要先解析域名,首次调用链为:ngx_http_upstream_init_request->ngx_resolver_name->ngx_resolver_name_locked->ngx_resolver_send_query;若5s(单次超时)后还没收到dns响应,则再发送1次查询,调用链为:ngx_resolver_resend_handler->ngx_resolver_resend->ngx_resolver_send_query,如此反复,直到收到响应或30s(默认总超时)后不再发送查询。它调用ngx_send发送dns查询,16行~21行代码为笔者添加,ngx_resolver_read_response函数用于接收并分析dns响应报文,它会调用到下面的ngx_udp_overlapped_wsarecv函数。

异步接收
   由ngx_udp_overlapped_wsarecv函数实现,定义在os/win32/ngx_udp_wsarecv.c中。
 1ssize_t ngx_udp_overlapped_wsarecv(ngx_connection_t *c, u_char *buf, size_t size)
 2{
 3    int             flags, rc;
 4    WSABUF          wsabuf;
 5    ngx_err_t       err;
 6    ngx_event_t    *rev;
 7    WSAOVERLAPPED  *ovlp;
 8    u_long             bytes;
 9    
10    rev = c->read;
11    
12    if (!rev->ready) {
13      ngx_log_error(NGX_LOG_ALERT, c->log, 0"ngx_udp_overlapped_wsarecv second wsa post");
14      return NGX_AGAIN;
15  }

16    
17  if (rev->complete) {
18   if (ngx_event_flags & NGX_USE_IOCP_EVENT) {
19       if (rev->ovlp.error && rev->ovlp.error != ERROR_MORE_DATA{
20           ngx_connection_error(c, rev->ovlp.error, "ngx_udp_overlapped_wsarecv() failed");
21           return NGX_ERROR;
22       }

23   }

24   
25   rev->complete = 0;
26  }

27     
28    ovlp = NULL;
29    wsabuf.buf = (CHAR *) buf;
30    wsabuf.len = (ULONG) size;
31    flags = 0;
32    
33retry:    
34    rc = WSARecv(c->fd, &wsabuf, 1, (DWORD*)&bytes, (LPDWORD)&flags, ovlp, NULL);
35        
36    if (rc == -1{
37         rev->ready = 0;
38         err = ngx_socket_errno;
39    
40       if (err == WSA_IO_PENDING) {
41             return NGX_AGAIN;
42         }

43    
44       if (err == WSAEWOULDBLOCK) {
45           if (ngx_event_flags & NGX_USE_IOCP_EVENT) {                
46                 rev->ovlp.type = NGX_IOCP_IO;
47                 ovlp = (WSAOVERLAPPED *)&rev->ovlp;
48                 ngx_memzero(ovlp, sizeof(WSAOVERLAPPED));
49                
50                 wsabuf.buf = NULL;
51                 wsabuf.len = 0;
52                 flags = MSG_PEEK;
53                
54                goto retry;            
55            }

56            
57            return NGX_AGAIN;            
58        }

59            
60        ngx_connection_error(c, err, "ngx_udp_overlapped_wsarecv() failed");    
61        rev->error = 1;
62    
63        return NGX_ERROR;
64    }

65    
66    if ((ngx_event_flags & NGX_USE_IOCP_EVENT) && ovlp) {
67        rev->ready = 0;
68        return NGX_AGAIN;
69    }

70    
71    return bytes;
72}
   先以非阻塞方式接收,若发生WSAWOULDBLOCK错误,则使用MSG_PEEK标志投递一个0字节的重叠接收操作,当dns响应返回时发生完成事件,会再次进入ngx_resolver_read_response而调用到该函数,此时rev->complete为1,rev->ovlp.error为ERROR_MORE_DATA(GetQueuedCompletionStatus返回的错误),由于使用了MSG_PEEK,因此数据还在接收缓冲区中,要忽略ERROR_MORE_DATA而继续接收,这时就能成功了。不管WSARecv返回WSA_IO_PENDING错误还是成功,iocp都会得到完成通知,所以这里当重叠操作投递成功时,返回NGX_AGAIN,便于在回调内统一处理。
posted on 2015-06-25 17:01 春秋十二月 阅读(6215) 评论(0)  编辑 收藏 引用 所属分类: Opensrc

只有注册用户登录后才能发表评论。
网站导航: 博客园   IT新闻   BlogJava   博问   Chat2DB   管理