c/c++背景開發學習筆記 2.1.2 reactor原理與實作，實作百萬級并發

reactor

把epoll傳回的fd讀寫事件進行封裝，并為每種事件設定回調函數，把所有關注的fd以及對應的事件存儲在一個資料結構裡，與epoll内部的紅黑樹的節點形成一一對應的關系。epoll傳回時使用data.ptr得到我們資料結構中對應的entry，再進行處理

對fd的封裝如下

typedef int (*NCALLBACK)(int fd, int events, void *arg);

struct ntyevent {
	int fd;
	int events;                     //監聽的事件：EPOLLIN或EPOLLOUT
	void *arg;                      //傳給回調函數的額外參數（reactor指針）
	NCALLBACK callback;             //回調函數
	
	int used;                       //目前entry是有有效
	char buffer[BUFFER_LENGTH];     //讀寫共用一個buffer，因為目前監聽的事件隻能是讀或寫，不會兩個都有
	int length;                     //buffer中有效資料的長度
	long last_active;               //目前fd上一次活躍的事件
	int sticky;                     //如果置為1，則fd不會因為長時間不活躍而被關閉（用于listenfd）
};

reactor資料結構的定義如下

struct ntyreactor {
	int epfd;
	int block_num;
	struct ntyevent **events; //array of ntyevent *, length = block_num, each block has 1024 ntyevent
};

events是一個指針數組，可擴容。每個元素指向一個1024長度的ntyevent數組

整個伺服器大緻有如下函數

int accept_cb(int fd, int events, void *arg);
int recv_cb(int fd, int events, void *arg);
int send_cb(int fd, int events, void *arg);

void nty_event_update(struct ntyevent *ev, int fd, NCALLBACK callback, void *arg);
int nty_event_add(int epfd, int events, struct ntyevent *ev);
int nty_event_del(int epfd, struct ntyevent *ev);

int start_listen(short port);
int ntyreactor_init(struct ntyreactor *reactor);
struct ntyevent *ntyreactor_get_event(struct ntyreactor *reactor, int fd);
int ntyreactor_destory(struct ntyreactor *reactor);
int ntyreactor_addlistener(struct ntyreactor *reactor, int listenfd, NCALLBACK acceptor);
int ntyreactor_run(struct ntyreactor *reactor);

先看main函數

#define BUFFER_LENGTH		4096
#define EVENTS_BLOCK_SIZE	1024
#define EVENT_BATCH_SIZE    1024
#define SERVER_PORT			8888
#define PORT_COUNT          100
#define CLIENT_TIMEOUT      15

int main(int argc, char *argv[]) {

	unsigned short port = SERVER_PORT;
	if (argc == 2) {
		port = atoi(argv[1]);
	}

	struct ntyreactor *reactor = (struct ntyreactor*)calloc(1, sizeof(struct ntyreactor));
	ntyreactor_init(reactor);
	
	int i;
	for(i = 0; i < PORT_COUNT; i++) {
		int listenfd = start_listen(port + i);
		ntyreactor_addlistener(reactor, listenfd, accept_cb);
	}
	ntyreactor_run(reactor);

	ntyreactor_destory(reactor);
	free(reactor);
	return 0;
}

PORT_COUNT設為100時，程式監聽從8888開始的100個端口，每個端口至少能接受1萬個port，是以連接配接數能達到一百萬。（因為測試的用戶端數量比較少，如果不多監聽一些端口，會導緻（src-ip, src-port, dst-ip, dst-port) 四元組的數量不夠用）

其中

ntyreactor_run

是程式主循環

int ntyreactor_run(struct ntyreactor *reactor) {
	if (reactor == NULL) return -1;
	if (reactor->epfd < 0) return -1;
	if (reactor->events == NULL) return -1;
	
	struct epoll_event events[EVENT_BATCH_SIZE];
	
	int block_idx = 0, check_interval = 10, i, cur; // check every 10 loops

	while (1) {
#if CHECK_TIMEOUT
		cur++;
		if (cur % check_interval == 9) {
			if (block_idx >= reactor->block_num) {
				block_idx = 0;
			}
			ntyreactor_check_timeout(reactor, block_idx++);
		}
#endif
		int nready = epoll_wait(reactor->epfd, events, EVENT_BATCH_SIZE, 1000);
		if (nready < 0) {
			printf("epoll_wait error\n");
			continue;
		}

		for (i = 0; i < nready; i++) {

			struct ntyevent *ev = (struct ntyevent*)events[i].data.ptr;
			int err = 0;
			if ((events[i].events & EPOLLIN) && (ev->events & EPOLLIN)) {
				err = ev->callback(ev->fd, events[i].events, ev->arg);
			}
			if ((events[i].events & EPOLLOUT) && (ev->events & EPOLLOUT)) {
				err = ev->callback(ev->fd, events[i].events, ev->arg);
			}
			
		}

	}
}

主循環中對于每個監聽到的事件直接調用回調函數

accept_cb把每個新連接配接添加到epoll和reactor中監控

int accept_cb(int fd, int events, void *arg) {

	struct ntyreactor *reactor = (struct ntyreactor*)arg;
	if (reactor == NULL) return -1;

	struct sockaddr_in client_addr;
	socklen_t len = sizeof(client_addr);

	int clientfd;

	if ((clientfd = accept(fd, (struct sockaddr*)&client_addr, &len)) == -1) {
		if (errno != EAGAIN && errno != EINTR) {
			
		}
		printf("accept: %s\n", strerror(errno));
		return -1;
	}

	int i = 0;
	struct ntyevent *event = ntyreactor_get_event(reactor, clientfd);
	if (event == NULL) return -1;

	do {

		// 設定clientfd非阻塞
		if (fcntl(clientfd, F_SETFL, O_NONBLOCK) < 0) {
			printf("%s: fcntl nonblocking failed, %d\n", __func__, EVENTS_BLOCK_SIZE);
			break;
		}

		nty_event_update(event, clientfd, recv_cb, reactor);
		nty_event_add(reactor->epfd, EPOLLIN, event);

	} while (0);

	printf("new connect [%s:%d][time:%ld], pos[%d]\n", 
		inet_ntoa(client_addr.sin_addr), ntohs(client_addr.sin_port), event->last_active, i);

	return 0;

}

recv_cb 和 send_cb分别負責讀和寫

這裡要提到水準觸發和邊沿觸發的問題

水準觸發是隻要條件為真就一直觸發（例如隻要buffer裡資料還沒讀完，epoll就一直觸發EPOLLIN），而邊沿觸發隻在條件從假變為真時觸發一次

通常，資料量較大時使用水準觸發（一次讀不完buffer當中所有資料時）。資料較小是可以使用邊沿觸發。邊沿觸發每次必須用while循環讀取所有資料。

這段代碼使用的是水準觸發。

int recv_cb(int fd, int events, void *arg) {

	struct ntyreactor *reactor = (struct ntyreactor*)arg;
	struct ntyevent *ev =  ntyreactor_get_event(reactor, fd);
	if(ev == NULL) return -1;

	int len = recv(fd, ev->buffer, BUFFER_LENGTH, 0);
	nty_event_del(reactor->epfd, ev);

	if (len > 0) {
		ev->length = len;
		ev->buffer[len] = '\0';
		printf("C[%d]:%s\n", fd, ev->buffer);

		nty_event_update(ev, fd, send_cb, reactor);
		nty_event_add(reactor->epfd, EPOLLOUT, ev);
	} else if (len == 0) {
		close(ev->fd);
		printf("[fd=%d] closed\n", fd);
	} else {
		close(ev->fd);
		printf("recv[fd=%d] error[%d]:%s\n", fd, errno, strerror(errno));
	}

	return len;
}


int send_cb(int fd, int events, void *arg) {

	struct ntyreactor *reactor = (struct ntyreactor*)arg;
	struct ntyevent *ev =  ntyreactor_get_event(reactor, fd);
	if(ev == NULL) return -1;

	int len = send(fd, ev->buffer, ev->length, 0);
	if (len > 0) {
		printf("send[fd=%d], [%d]%s\n", fd, len, ev->buffer);

		nty_event_del(reactor->epfd, ev);
		nty_event_update(ev, fd, recv_cb, reactor);
		nty_event_add(reactor->epfd, EPOLLIN, ev);
		
	} else {

		close(ev->fd);

		nty_event_del(reactor->epfd, ev);
		printf("send[fd=%d] error %s\n", fd, strerror(errno));

	}

	return len;
}

reactor自動擴容通過ntyreactor_get_event實作

//輸入fd，傳回對應的ntyevent
struct ntyevent *ntyreactor_get_event(struct ntyreactor *reactor, int fd) {
	int b = fd / EVENTS_BLOCK_SIZE;
	int i = fd % EVENTS_BLOCK_SIZE;
	if(b >= reactor->block_num) {
		int new_block_num = reactor->block_num == 0 ? 1 : 2 * reactor->block_num;
		while(new_block_num <= b) new_block_num <<= 1;
		struct ntyevent **new_event_blocks = (struct ntyevent **)realloc(
			reactor->events, new_block_num * (sizeof(struct ntyevent *))
		);
		if(new_event_blocks == NULL) {
			printf("cannot allocate block in %s for fd = %d\n", __func__, fd);
			return NULL;
		}
		memset(
			new_event_blocks + reactor->block_num, 0,
			(new_block_num - reactor->block_num) * sizeof(struct ntyevent*));
		reactor->block_num = new_block_num;
		reactor->events = new_event_blocks;
	}
	
	if(reactor->events[b] == NULL) {
		reactor->events[b] = (struct ntyevent *)calloc(EVENTS_BLOCK_SIZE, sizeof(struct ntyevent));
		if (reactor->events[b] == NULL) {
			printf("cannot allocate block in %s for fd = %d\n", __func__, fd);
			return NULL;
		}
	}

	return &reactor->events[b][i];
}

使用 ntyreactor_check_timeout自動檢查block_idx中長時間不活躍的fd并關閉連接配接

int ntyreactor_check_timeout(struct ntyreactor *reactor, int block_idx) {
	if(!reactor || !reactor->events || block_idx >= reactor->block_num) return 0;
	struct ntyevent *block = reactor->events[block_idx];
	if (!block) return 0;

	long now = time(NULL);
	int i;
	for(i = 0; i < EVENTS_BLOCK_SIZE; i++) {
		if(!block[i].used || block[i].sticky) continue;
		long duration = now - block[i].last_active;
		if (duration >= 60) {
			close(block[i].fd);
			printf("[fd=%d] timeout\n", block[i].fd);
			nty_event_del(reactor->epfd, &block[i]);
		}
	}
}

nty_event_* 這幾個函數負責把fd和監聽的event加入到epoll和reactor中，并設定回調函數

void nty_event_update(struct ntyevent *ev, int fd, NCALLBACK callback, void *arg) {
	ev->fd = fd;
	ev->callback = callback;
	ev->events = 0;
	ev->arg = arg;
	ev->last_active = time(NULL);
	ev->sticky = 0;
}

// add/update ev on epfd with events
int nty_event_add(int epfd, int events, struct ntyevent *ev) {

	struct epoll_event ep_ev = {0, {0}};
	ep_ev.data.ptr = ev;
	ep_ev.events = ev->events = events;

	int op;
	if (ev->used == 1) {
		op = EPOLL_CTL_MOD;
	} else {
		op = EPOLL_CTL_ADD;
		ev->used = 1;
	}

	if (epoll_ctl(epfd, op, ev->fd, &ep_ev) < 0) {
		printf("event add failed [fd=%d], events[%d]\n", ev->fd, events);
		return -1;
	}

	return 0;
}

// remove ev from epfd
int nty_event_del(int epfd, struct ntyevent *ev) {

	struct epoll_event ep_ev = {0, {0}};

	if (ev->used != 1) {
		return -1;
	}

	ep_ev.data.ptr = ev;
	ev->used = 0;
	epoll_ctl(epfd, EPOLL_CTL_DEL, ev->fd, NULL);

	return 0;
}

還有其他一些函數一并寫在這裡

int start_listen(short port) {

	int fd = socket(AF_INET, SOCK_STREAM, 0);
	fcntl(fd, F_SETFL, O_NONBLOCK);

	struct sockaddr_in server_addr;
	memset(&server_addr, 0, sizeof(server_addr));
	server_addr.sin_family = AF_INET;
	server_addr.sin_addr.s_addr = htonl(INADDR_ANY);
	server_addr.sin_port = htons(port);

	bind(fd, (struct sockaddr*)&server_addr, sizeof(server_addr));

	if (listen(fd, 20) < 0) {
		printf("listen failed : %s\n", strerror(errno));
	}

	printf("listening to port %hd with fd = %d\n", port, fd);
	return fd;
}


int ntyreactor_init(struct ntyreactor *reactor) {

	if (reactor == NULL) return -1;
	memset(reactor, 0, sizeof(struct ntyreactor));

	reactor->epfd = epoll_create(1);
	if (reactor->epfd <= 0) {
		printf("create epfd in %s err %s\n", __func__, strerror(errno));
		return -2;
	}

	reactor->events = (struct ntyevent **)calloc(1, sizeof(struct ntyevent *));
	if (reactor->events == NULL) {
		printf("cannot allocate events block in %s\n", __func__);
		return -3;
	}
	reactor->block_num = 1;
	
	return 0;
}


int ntyreactor_destory(struct ntyreactor *reactor) {
	close(reactor->epfd);
	int i;
	for(i = 0; i < reactor->block_num; i++) {
		free(reactor->events[i]);
	}
	free(reactor->events);
}



int ntyreactor_addlistener(struct ntyreactor *reactor, int listenfd, NCALLBACK acceptor) {

	if (reactor == NULL) return -1;
	if (reactor->events == NULL) return -1;

	struct ntyevent *event = ntyreactor_get_event(reactor, listenfd);
	if (event == NULL) return -1;

	nty_event_update(event, listenfd, acceptor, reactor);
	event->sticky = 1; //don't close listen fd on timeout
	nty_event_add(reactor->epfd, EPOLLIN, event);

	return 0;
}

其他注意事項

為了達到百萬連接配接數，我們還要增加系統的最大open files

使用

ulimit -a

檢視open files

用-n設定，如果不行，就更改

sudo vim /etc/security/limits.conf

這個

末尾加上這兩行

*                soft    nofile          1048576
*                hard    nofile          1048576

并且關掉shell并重新登入

此時應該就行了

并且可能還需要更改核心tcp收發緩存的大小

可以參考這裡

參考資料

[1] 零聲教育 Linux C/C++後端伺服器架構開發 2.1網絡開發

c/c++背景開發學習筆記 2.1.2 reactor原理與實作，實作百萬級并發

reactor

其他注意事項

參考資料

繼續閱讀

while 循環、do- while 循環和 for 循環之間的那點事C語言自學之三種循環比較

httpd服務的部署、啟動、配置和簡單優化一、部署二、啟動三、配置檔案

配置網頁内容通路

手動安裝Intel network I217-LM網卡的Linux驅動

禁止ubuntu系統彈出報錯界面

Ubuntu Linux下Apache的配置檔案

結構體：typedef與struct的差別

QR碼編碼原理三（日本漢字和中文編碼）

samba伺服器的功能

AOP程式設計_Android優雅權限架構(1)概念基礎，2021金三銀四前言正文大綱正文

【Linux】UDP廣播封包接收速率問題

express 搭建背景建構工程目錄結構

Linux裝置模型（中）之上層容器

PowerPC平台 Linux移植三

hdu7108哈希