本文主要是介绍Linux下使用ZMQ实践之与libevent结合,希望对大家解决编程问题提供一定的参考价值,需要的开发者们随着小编来一起学习吧!
1. 前言
最近考虑到一个问题,项目中有同时处理socket、zeromq的逻辑需求,想通过libevent(I/O服用)一块将zmq-socket的事件也放一个线程中处理。
网上了解了一些实现,大部分都是通过将zmq的sockfd
拿到,加入libevent_dispatch
中一并处理,但是存在问题是实时性不够,写法不对导致丢事件的情况。
上述的根本原因是:zeromq是通过底层独立线程完成socket层面的收发,拷贝到内存队列供给上层业务使用。强行监听socket事件,但实际数据获取还是从内存队列拿取,肯定是存在时效性的问题。
所以,本文尝试一种方法,服务端这块直接剖析zeromq协议zmtp
,不要zmq子线程、内存队列,直接从socket层面收取数据;客户端仍然兼容libzmq
常规的写法。
2. 相关知识
根据 zmtp官方说明 的解读:
协议流程主要为:打招呼greeting
、握手handshake
、通信traffic
;
通信内容主要分为命令command
、消息message
;
2.1 Greeting格式
Greeting 又可以细分为:签名signature
、版本号version
、机器码mechanism
、类型as-server
、扩展filter
;主要是zmq考虑到了多种协议版本、多种协议类型,考虑的字段信息就多了。
; The greeting announces the protocol detailsgreeting = signature version mechanism as-server fillersignature = %xFF padding %x7F
padding = 8OCTET ; Not significantversion = version-major version-minor
version-major = %x03
version-minor = %x00; The mechanism is a null padded string
mechanism = 20mechanism-char
mechanism-char = "A"-"Z" | DIGIT| "-" | "_" | "." | "+" | %x0; Is the peer acting as server?
as-server = %x00 | %x01; The filler extends the greeting to 64 octets
filler = 31%x00 ; 31 zero octets
2.1 Handshake格式
Handshake相比Greeting简单多了,一个command搞定
; The handshake consists of at least one command
; The actual grammar depends on the security mechanism
handshake = 1*command
2.2 Command格式
Command提供了长命令、短命令进行选择;与Message不同的是,Command数据域body
还拆分出命令名称Command-name
和命令数据Command-data
,这块主要用到的地方是包含在数据帧中,配合Message进行信息控制交互;
; A command is a single long or short frame
command = command-size command-body
command-size = %x04 short-size | %x06 long-size
short-size = OCTET ; Body is 0 to 255 octets
long-size = 8OCTET ; Body is 0 to 2^63-1 octets
command-body = command-name command-data
command-name = OCTET 1*255command-name-char
command-name-char = ALPHA
command-data = *OCTET
2.3 Message格式
Message内部其实提供了两种类型(长消息、短消息)、两种标记(独立报文、组合报文)进行灵活选择
; A message is one or more frames
message = *message-more message-last
message-more = ( %x01 short-size | %x03 long-size ) message-body
message-last = ( %x00 short-size | %x02 long-size ) message-body
message-body = *OCTET
3. 实践
通过zmtp
协议的指导,我们选取一个方向来实践一下:push-pull
协议、libevent::buffer_event
;
3.1 连接握手
这次先看一下主函数,有个大概的印象:socket连接__do_connect
,zmq握手__do_handshake
,回调函数拉数据on_recv
。
int main(int argc, char *argv[])
{int res = -1;struct bufferevent *bev = NULL;struct event_base *base = NULL;struct msg_ctx mctx = {0};mctx.evbuf = evbuffer_new();__do_reset(&mctx);if (argc < 2) {printf("%s <address>\n", argv[0], argv[1]);exit(EXIT_FAILURE);}printf("Program start...\n");base = event_base_new();assert(base);bev = __do_connect(base, argv[1]);res = __do_handshark(bev);assert(res == 0);bufferevent_setwatermark(bev, EV_READ, 0, 0);bufferevent_setcb(bev, on_recv, NULL, NULL, &mctx);bufferevent_enable(bev, EV_READ);event_base_dispatch(base);printf("Program quit...\n");sleep(5);return EXIT_SUCCESS;
}
本端实现的是pull的功能,需要连接到push,然后进行拉取数据;
socket建立连接使用的常规的bufferevent_socket_connect
的方法
static struct bufferevent *__do_connect(struct event_base *base, const char *paddr)
{int res = -1;int dstlen = sizeof(struct sockaddr_storage);struct bufferevent *bev = NULL;struct sockaddr_storage dst;res = evutil_parse_sockaddr_port(paddr, (struct sockaddr *)&dst, &dstlen);assert(0 == res);bev = bufferevent_socket_new(base, -1, BEV_OPT_CLOSE_ON_FREE);assert(bev);res = bufferevent_socket_connect(bev, (struct sockaddr *)&dst, dstlen);assert(0 == res);return bev;
}
下来准备看握手过程的实现,先看一下结构体的定义是怎么跟zmtp
结合的;这里有个地方注意一下,zmtp3.0与2.0协议格式还不一致,所以先制定约束:使用v3协议,不做v2的兼容。
/** ZMTP 3.0* connection = greeting handshake traffic** This is the 3.0 greeting (64 bytes)* greeting = signature version mechanism as-server filler*/
struct zmtp_greeting {char signature [10]; // %xFF padding %x7Fchar version [2]; // version-major %x03, version-minor %x00char mechanism [20]; // The mechanism is a null padded stringchar as_server [1]; // Is the peer acting as serverchar filler [31]; // The filler extends the greeting to 64 octets
};
从第2小节可以看出,Message和Command是类似的,这里我们全抽象为:flags、size、data(柔性数组)
/** A message is one or more frames** message = *message-more message-last* message-more = ( %x01 short-size | %x03 long-size ) message-body* message-last = ( %x00 short-size | %x02 long-size ) message-body* message-body = *OCTET** short-size = OCTET ; Body is 0 to 255 octets* long-size = 8OCTET ; Body is 0 to 2^63-1 octets**/
struct zmtp_msg_shdr {u8 flags; // Must be zerou8 size; // Size, 0 to 255 bytesu8 data [0]; // Message data
};struct zmtp_msg_lhdr {u8 flags; // Must be zerou64 size; // Size, 0 to 255 bytesu8 data [0]; // Message data
};
结构体准备好了,我们看一个握手的快速实现,这里说的快速
是指使用阻塞同步的形式完成;另外一种常规的做法是使用状态机+事件回调的方式完成。
static int __do_handshark(struct bufferevent *bev)
{int res = -1;int fd = bufferevent_getfd(bev);struct zmtp_greeting gt = ZMTP_GREETING_INIT;char buffer[SIZE_LINE_NORMAL];struct zmtp_msg_shdr *phead = (struct zmtp_msg_shdr *)buffer;char pull_data[] = ZMTP_HANDSHAKE_PULL;struct zmtp_msg_shdr pull_head = {.flags = ZMTP_FLAGS_SCMD,.size = sizeof(pull_data) - 1,};/* Greeting */res = sdk_tcp_send_nbytes(fd, >, sizeof(struct zmtp_greeting), TIMEO);assert(0 == res);res = sdk_tcp_recv_nbytes(fd, >, sizeof(struct zmtp_greeting), TIMEO);assert(0 == res);printf("Greeting done...\n");/* Handshark */res = sdk_tcp_recv_nbytes(fd, phead, sizeof(struct zmtp_msg_shdr), TIMEO);assert(0 == res);assert(0x04 == phead->flags);res = sdk_tcp_recv_nbytes(fd, phead->data, phead->size, TIMEO);assert(0 == res);printf("Handsharke size: %d\n\n", phead->size);DSP_TOTAL(phead->data, phead->size);res = sdk_tcp_send_nbytes(fd, &pull_head, sizeof(struct zmtp_msg_shdr), TIMEO);assert(0 == res);res = sdk_tcp_send_nbytes(fd, &pull_data, pull_head.size, TIMEO);assert(0 == res);printf("Handsharke done...\n");return 0;
}
上述其实可以看到,报文填充非常暴力,通过宏直接将握手信息全部填充进去了。
#define ZMTP_GREETING_INIT { \{ 0xFF, 0, 0, 0, 0, 0, 0, 0, 1, 0x7F }, \{ 3, 0 }, \{ 'N', 'U', 'L', 'L', 0 }, \{ 0 }, \{ 0 } \
}#define ZMTP_HANDSHAKE_PULL { \0x05, \'R', 'E', 'A', 'D', 'Y', \0x0b, \'S', 'o', 'c', 'k', 'e', 't', '-', 'T', 'y', 'p', 'e', \0x00, 0x00, 0x00, 0x04, \'P', 'U', 'L', 'L', '0'\
}
3.2 数据拉取
数据拉取过程,由于Message
数据域是不定长的,所以我们通过状态机
的形式剖报文
void on_recv(struct bufferevent *bev, void *args)
{int res = 0;struct msg_ctx *mctx = (struct msg_ctx *)args;/* Et mode */while (1) {size_t length = evbuffer_get_length(bufferevent_get_input(bev));if (length == 0) {break;}printf("Evbuffer length: %u\n", length);switch (mctx->status) {case STATUS_MSG_HEAD:if (0 == __do_head_parser(bev, mctx)) {/* state transition */mctx->status = STATUS_MSG_BODY;printf("Head done, data: %u\n", mctx->size);}break;case STATUS_MSG_BODY:if (0 == __do_body_parser(bev, mctx)) {/* state transition */mctx->status = STATUS_MSG_HEAD;printf("Body done, data: %u\n", mctx->size);}break;default:assert(0);break;}}return;
}
Message头部信息,我们主要需要区分是长消息、还是短消息;是独立消息、还是组合消息。
enum {ZMTP_FLAGS_SCMD = 0x04,ZMTP_FLAGS_LCMD = 0x06,ZMTP_FLAGS_SMSG_MORE = 0x01,ZMTP_FLAGS_LMSG_MORE = 0x03,ZMTP_FLAGS_SMSG_LAST = 0x00,ZMTP_FLAGS_LMSG_LAST = 0x02,
};static int __do_head_parser(struct bufferevent *bev, struct msg_ctx *mctx)
{if (mctx->flags == 0xFF) {bufferevent_read(bev, &mctx->flags, sizeof(u8));printf("Head flag: %x\n", mctx->flags);}if (mctx->size == 0x00) {size_t expect = 0;size_t length = evbuffer_get_length(bufferevent_get_input(bev));switch (mctx->flags) {case ZMTP_FLAGS_SMSG_MORE:case ZMTP_FLAGS_SMSG_LAST:expect = sizeof(u8);printf("Body short\n");break;case ZMTP_FLAGS_LMSG_MORE:case ZMTP_FLAGS_LMSG_LAST:expect = sizeof(u64);printf("Body long\n");break;default:assert(0);break;}if (length < expect) {printf("Retry\n");return RETRY;}size_t rlen = bufferevent_read(bev, &mctx->size, expect);assert(rlen == expect);if (expect == sizeof(u64)) {mctx->size = ntohll(mctx->size);}assert(mctx->size > 0);}return 0;
}
然后是数据域状态下的数据读取了,这个主要考虑缓冲消息到evbuffer
中,消息收全了最后才调用__do_something
和__do_reset
函数
static int __do_body_parser(struct bufferevent *bev, struct msg_ctx *mctx)
{char buffer[SIZE_LINE_LONG];size_t rmax = _MIN(sizeof(buffer), mctx->size - evbuffer_get_length(mctx->evbuf));size_t rlen = bufferevent_read(bev, buffer, rmax);int res = evbuffer_add(mctx->evbuf, buffer, rlen);assert(res == 0);if (evbuffer_get_length(mctx->evbuf) >= mctx->size) {printf("Body done, length: %u/%u\n",evbuffer_get_length(mctx->evbuf), mctx->size);__do_something(mctx);__do_reset(mctx);return 0;}printf("Retry\n");return RETRY;
}static void __do_reset(struct msg_ctx *mctx)
{mctx->status = STATUS_MSG_HEAD;mctx->flags = 0xFF;mctx->size = 0x00;evbuffer_drain(mctx->evbuf, evbuffer_get_length(mctx->evbuf));
}static void __do_something(struct msg_ctx *mctx)
{printf("-- 0x%x\n", mctx->flags);printf("-- %d\n", mctx->size);while (evbuffer_get_length(mctx->evbuf) > 0) {char buffer[SIZE_LINE_LONG];ssize_t rlen = evbuffer_remove(mctx->evbuf, buffer, sizeof(buffer));printf("-- %s\n", buffer);}
}
3.3 运行结果
短消息的拉取:
./pullx 127.0.0.1:5555Program start...
Greeting done...
Handsharke size: 260000 05 52 45 41 44 59 0b 53 - 6f 63 6b 65 74 2d 54 79 .READY.Socket-Ty
0010 70 65 00 00 00 04 50 55 - 53 48 ** ** ** ** ** ** pe....PUSH
Handsharke done...
Evbuffer length: 14
Head flag: 0
Body short
Head done, data: 12
Evbuffer length: 12
Body done, length: 12/12
-- 0x0
-- 12
-- Data- -000
Body done, data: 0
Evbuffer length: 14
Head flag: 0
Body short
Head done, data: 12
Evbuffer length: 12
Body done, length: 12/12
-- 0x0
-- 12
-- Data- -001
Body done, data: 0
长消息的拉取:
./pullx 127.0.0.1:5555
Program start...
Greeting done...
Handsharke size: 260000 05 52 45 41 44 59 0b 53 - 6f 63 6b 65 74 2d 54 79 .READY.Socket-Ty
0010 70 65 00 00 00 04 50 55 - 53 48 ** ** ** ** ** ** pe....PUSH
Handsharke done...
Evbuffer length: 1033
Head flag: 2
Body long
Head done, data: 1024
Evbuffer length: 1024
Body done, length: 1024/1024
-- 0x2
-- 1024
-- Data- -000
Body done, data: 0
Evbuffer length: 1033
Head flag: 2
Body long
Head done, data: 1024
Evbuffer length: 1024
Body done, length: 1024/1024
-- 0x2
-- 1024
-- Data- -001
Body done, data: 0
Evbuffer length: 1033
Head flag: 2
Body long
Head done, data: 1024
Evbuffer length: 1024
Body done, length: 1024/1024
4. 结论
通过本次试验,验证了直接剖取zmtp
方法的可行性;但是要到工程中实践,还得思考几个问题:
- push-pull协议简单,但router-dealer、pub-sub模式、加密、认证协议的剖析会更加复杂;
- socket维护的问题,需要考虑断线重连、重新握手的问题;
- 协议版本的兼容性;
这篇关于Linux下使用ZMQ实践之与libevent结合的文章就介绍到这儿,希望我们推荐的文章对编程师们有所帮助!