我们知道，数据流从网络到用户态应用程序，需要经历的阶段，如下图黑色线部分
就是从网卡拷贝到内核，再从内核拷贝到用户态程序，这其中会经历两次拷贝。如果时大量客户端C10M时，就会造成性能的瓶颈。
所以，就有了用户态协议栈的诞生，也就是说，原本内核完成了协议栈的功能，被移动到用户程序，那此时，用户程序如何从网卡取数据呢？
有很多种实现方式，比如通过netmap，将网卡的数据映射到内存，则用户态协议栈只需要从内存中解析并实现协议栈的功能，则可以实现用户态的协议栈的应用程序。减少了两次拷贝的开销。
这也就是实现了零拷贝，因为网卡到内存的数据，是直接通过DMA技术来实现，绕过了CPU。
如何将网卡数据映射到内存？这里使用netmap的开源软件，https://github.com/luigirizzo/netmap
如何安装和编译，这个软件依赖了内核，所以，首先要找到自己的系统内核版本：
https://github.com/luigirizzo/netmap比如我的 3.10.0-1160.62.1.el7.x86_64，上谷歌搜索，找到下面的这个包，然后安装对应的内核的rpm包kernel-devel-3.10.0-1160.62.1.el7.x86_64.rpm，然后通过.configure,再make和make install
就能成功安装netmap，这里只是大概的讲解，基本上网上都能搜索到，还有最好把网卡改名位eth0
编译完成后，就能发现有netmap.ko 的文件。
启动内核模块，insmod netmap.ko
netmap的代码实现首先，我们先要了解一包网络数据是怎么组成的。
以太网头的组成：对应的数据结构定义：
#define ETH_ALEN 6
struct ethhdr {
unsigned char h_dest[ETH_ALEN];
unsigned char h_source[ETH_ALEN];
unsigned short h_proto;
};
IP头定义
代码实现：
struct iphdr {
unsigned char version;
unsigned char tos;
unsigned short tot_len;
unsigned short id;
unsigned short flag_off;
unsigned char ttl;
unsigned char protocol;
unsigned short check;
unsigned int saddr;
unsigned int daddr;
};
UDP头
struct udphdr {
unsigned short source;
unsigned short dest;
unsigned short len;
unsigned short check;
};
剩下的就是整个UDP包
struct udppkt {
struct ethhdr eh;
struct iphdr ip;
struct udphdr udp;
unsigned char body[0];
};
其中用到了柔性数组，body[0]，这是在工程中如果作为结构体通信的常用技术手段，为啥这里不用指针，因为指针与指向的地址的内存不连续。这里应该很好理解。
然后，我们来看怎么通过netmap实现对数据包的解析。
主程序实现int main()
{
struct ethhdr* eh;
struct pollfd pfd = { 0 };
struct nm_pkthdr h;
unsigned char* stream = NULL;
struct nm_desc* nmr = nm_open("netmap:eth0", NULL, 0, NULL);
if (nmr == NULL) {
printf("Error openn");
return -1;
}
pfd.fd = nmr->fd;
pfd.events = POLLIN;
while (1) {
int ret = poll(&pfd, 1, -1);
if (ret < 0) continue;
if (pfd.revents & POLLIN) {
stream = nm_nextpkt(nmr, &h);
eh = (struct ethhdr*)stream;
if (ntohs(eh->h_proto) == PROTO_IP) {
struct udppkt* udp = (struct udppkt*)stream;
if (udp->ip.protocol == PROTO_UDP) {
struct in_addr addr;
addr.s_addr = udp->ip.saddr;
int udp_length = ntohs(udp->udp.len);
printf("%s:%d:length:%d, ip_len:%d --> ", inet_ntoa(addr), udp->udp.source,
udp_length, ntohs(udp->ip.tot_len));
printf("udp --> ");
for (int i = 0; i < udp_length - 8; i++)
{
printf("%c", udp->body[i]);
}
printf("n");
}
}
}
}
}
从程序看上，首先，通过nm_open("netmap:eth0", NULL, 0, NULL)获取nm_desc的对象
然后poll等待对应的fd上有数据，stream = nm_nextpkt(nmr, &h);
nm_nextpkt，可见，netmap应该是为我们的UDP数据获取了一块类似内存池的分配机制，我们只要直接从其中取出数据既可。
eh = (struct ethhdr*)stream;
从取出的stream中获取到以太网头，然后匹配对应的协议，如果是IP协议才处理
struct udppkt* udp = (struct udppkt*)stream;
然后转换为UDP包，然后解析出对应的协议。
测试结果：这简单了获取了udp的数据包。并打印出来，其实还是有bug的，没有处理arp和ICMP协议，会导致中间发了一会就收不到。
完整的实现代码#include#include #include #include #include #include #define NETMAP_WITH_LIBS #include #pragma pack(1) #define ETH_ALEN 6 #define PROTO_IP 0x0800 #define PROTO_ARP 0x0806 #define PROTO_UDP 17 #define PROTO_ICMP 1 #define PROTO_IGMP 2 struct ethhdr { unsigned char h_dest[ETH_ALEN]; unsigned char h_source[ETH_ALEN]; unsigned short h_proto; }; struct iphdr { unsigned char version; unsigned char tos; unsigned short tot_len; unsigned short id; unsigned short flag_off; unsigned char ttl; unsigned char protocol; unsigned short check; unsigned int saddr; unsigned int daddr; }; struct udphdr { unsigned short source; unsigned short dest; unsigned short len; unsigned short check; }; struct udppkt { struct ethhdr eh; struct iphdr ip; struct udphdr udp; unsigned char body[128]; }; struct arphdr { unsigned short h_type; unsigned short h_proto; unsigned char h_addrlen; unsigned char protolen; unsigned short oper; unsigned char smac[ETH_ALEN]; unsigned int sip; unsigned char dmac[ETH_ALEN]; unsigned int dip; }; struct arppkt { struct ethhdr eh; struct arphdr arp; }; struct icmphdr { unsigned char type; unsigned char code; unsigned short check; unsigned short identifier; unsigned short seq; unsigned char data[32]; }; struct icmppkt { struct ethhdr eh; struct iphdr ip; struct icmphdr icmp; }; void print_mac(unsigned char *mac) { int i = 0; for (i = 0;i < ETH_ALEN-1;i ++) { printf("%02x:", mac[i]); } printf("%02x", mac[i]); } void print_ip(unsigned char *ip) { int i = 0; for (i = 0;i < 3;i ++) { printf("%d.", ip[i]); } printf("%d", ip[i]); } void print_arp(struct arppkt *arp) { print_mac(arp->eh.h_dest); printf(" "); print_mac(arp->eh.h_source); printf(" "); printf("0x%04x ", ntohs(arp->eh.h_proto)); printf(" "); } int str2mac(char *mac, char *str) { char *p = str; unsigned char value = 0x0; int i = 0; while (p != ' ') { if (*p == ':') { mac[i++] = value; value = 0x0; } else { unsigned char temp = *p; if (temp <= '9' && temp >= '0') { temp -= '0'; } else if (temp <= 'f' && temp >= 'a') { temp -= 'a'; temp += 10; } else if (temp <= 'F' && temp >= 'A') { temp -= 'A'; temp += 10; } else { break; } value <<= 4; value |= temp; } p ++; } mac[i] = value; return 0; } void echo_arp_pkt(struct arppkt *arp, struct arppkt *arp_rt, char *hmac) { memcpy(arp_rt, arp, sizeof(struct arppkt)); memcpy(arp_rt->eh.h_dest, arp->eh.h_source, ETH_ALEN); str2mac(arp_rt->eh.h_source, hmac); arp_rt->eh.h_proto = arp->eh.h_proto; arp_rt->arp.h_addrlen = 6; arp_rt->arp.protolen = 4; arp_rt->arp.oper = htons(2); str2mac(arp_rt->arp.smac, hmac); arp_rt->arp.sip = arp->arp.dip; memcpy(arp_rt->arp.dmac, arp->arp.smac, ETH_ALEN); arp_rt->arp.dip = arp->arp.sip; } void echo_udp_pkt(struct udppkt *udp, struct udppkt *udp_rt) { memcpy(udp_rt, udp, sizeof(struct udppkt)); memcpy(udp_rt->eh.h_dest, udp->eh.h_source, ETH_ALEN); memcpy(udp_rt->eh.h_source, udp->eh.h_dest, ETH_ALEN); udp_rt->ip.saddr = udp->ip.daddr; udp_rt->ip.daddr = udp->ip.saddr; udp_rt->udp.source = udp->udp.dest; udp_rt->udp.dest = udp->udp.source; } unsigned short in_cksum(unsigned short *addr, int len) { register int nleft = len; register unsigned short *w = addr; register int sum = 0; unsigned short answer = 0; while (nleft > 1) { sum += *w++; nleft -= 2; } if (nleft == 1) { *(u_char *)(&answer) = *(u_char *)w ; sum += answer; } sum = (sum >> 16) + (sum & 0xffff); sum += (sum >> 16); answer = ~sum; return (answer); } void echo_icmp_pkt(struct icmppkt *icmp, struct icmppkt *icmp_rt) { memcpy(icmp_rt, icmp, sizeof(struct icmppkt)); icmp_rt->icmp.type = 0x0; // icmp_rt->icmp.code = 0x0; // icmp_rt->icmp.check = 0x0; icmp_rt->ip.saddr = icmp->ip.daddr; icmp_rt->ip.daddr = icmp->ip.saddr; memcpy(icmp_rt->eh.h_dest, icmp->eh.h_source, ETH_ALEN); memcpy(icmp_rt->eh.h_source, icmp->eh.h_dest, ETH_ALEN); icmp_rt->icmp.check = in_cksum((unsigned short*)&icmp_rt->icmp, sizeof(struct icmphdr)); } int main() { struct ethhdr *eh; struct pollfd pfd = {0}; struct nm_pkthdr h; unsigned char *stream = NULL; struct nm_desc *nmr = nm_open("netmap:eth0", NULL, 0, NULL); if (nmr == NULL) { printf("Error openn"); return -1; } pfd.fd = nmr->fd; pfd.events = POLLIN; while (1) { int ret = poll(&pfd, 1, -1); if (ret < 0) continue; if (pfd.revents & POLLIN) { stream = nm_nextpkt(nmr, &h); eh = (struct ethhdr*)stream; if (ntohs(eh->h_proto) == PROTO_IP) { struct udppkt *udp = (struct udppkt*)stream; if (udp->ip.protocol == PROTO_UDP) { struct in_addr addr; addr.s_addr = udp->ip.saddr; int udp_length = ntohs(udp->udp.len); printf("%s:%d:length:%d, ip_len:%d --> ", inet_ntoa(addr), udp->udp.source, udp_length, ntohs(udp->ip.tot_len)); udp->body[udp_length-8] = ' '; printf("udp --> %sn", udp->body); #if 1 struct udppkt udp_rt; echo_udp_pkt(udp, &udp_rt); nm_inject(nmr, &udp_rt, sizeof(struct udppkt)); #endif #if 0 } else if (udp->ip.protocol == PROTO_ICMP) { struct icmppkt *icmp = (struct icmppkt*)stream; printf("icmp ---------- --> %d, %xn", icmp->icmp.type, icmp->icmp.check); if (icmp->icmp.type == 0x08) { struct icmppkt icmp_rt = {0}; echo_icmp_pkt(icmp, &icmp_rt); //printf("icmp check %xn", icmp_rt.icmp.check); nm_inject(nmr, &icmp_rt, sizeof(struct icmppkt)); } #endif } else if (udp->ip.protocol == PROTO_IGMP) { } else { printf("other ip packet"); } #if 0 } else if (ntohs(eh->h_proto) == PROTO_ARP) { struct arppkt *arp = (struct arppkt *)stream; struct arppkt arp_rt; if (arp->arp.dip == inet_addr("192.168.2.217")) { echo_arp_pkt(arp, &arp_rt, "00:50:56:33:1c:ca"); nm_inject(nmr, &arp_rt, sizeof(struct arppkt)); } #endif } } } }
