nginx的域名解析器使用已連接udp（收發(fā)前先調(diào)用ngx_udp_connect）發(fā)送dns查詢、接收dns響應(yīng)，如上篇
tcp異步連接所講，iocp需要先投遞udp的接收操作，才能引發(fā)接收完成的事件，因此要對(duì)域名解析器和udp異步接收作些改進(jìn)。
發(fā)送后投遞
dns查詢由ngx_resolver_send_query函數(shù)實(shí)現(xiàn)，定義在core/ngx_resolver.c中。
1
static 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
}
當(dāng)nginx用于代理連接上游服務(wù)器前，要先解析域名，首次調(diào)用鏈為：
ngx_http_upstream_init_request->
ngx_resolver_name->
ngx_resolver_name_locked->
ngx_resolver_send_query；若5s（單次超時(shí)）后還沒(méi)收到dns響應(yīng)，則再發(fā)送1次查詢，調(diào)用鏈為：
ngx_resolver_resend_handler->
ngx_resolver_resend->
ngx_resolver_send_query，如此反復(fù)，直到收到響應(yīng)或30s（默認(rèn)總超時(shí)）后不再發(fā)送查詢。它調(diào)用ngx_send發(fā)送dns查詢，16行~21行代碼為筆者添加，ngx_resolver_read_response函數(shù)用于接收并分析dns響應(yīng)報(bào)文，它會(huì)調(diào)用到下面的ngx_udp_overlapped_wsarecv函數(shù)。
異步接收
由ngx_udp_overlapped_wsarecv函數(shù)實(shí)現(xiàn)，定義在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，若發(fā)生WSAWOULDBLOCK錯(cuò)誤，則使用MSG_PEEK標(biāo)志投遞一個(gè)0字節(jié)的重疊接收操作，當(dāng)dns響應(yīng)返回時(shí)發(fā)生完成事件，會(huì)再次進(jìn)入ngx_resolver_read_response而調(diào)用到該函數(shù)，此時(shí)rev->complete為1，rev->ovlp.error為ERROR_MORE_DATA（GetQueuedCompletionStatus返回的錯(cuò)誤），由于使用了MSG_PEEK，因此數(shù)據(jù)還在接收緩沖區(qū)中，要忽略ERROR_MORE_DATA而繼續(xù)接收，這時(shí)就能成功了。不管WSARecv返回WSA_IO_PENDING錯(cuò)誤還是成功，iocp都會(huì)得到完成通知，所以這里當(dāng)重疊操作投遞成功時(shí)，返回NGX_AGAIN，便于在回調(diào)內(nèi)統(tǒng)一處理。
posted on 2015-06-25 17:01
春秋十二月 閱讀(6259)
評(píng)論(0) 編輯 收藏 引用 所屬分類:
Opensrc