TCP/IP Checksum 吐槽
算法原理:
假定 output[2] 为输出结果,input[n]为待计算校验和的内存块。
1)所有奇数位[0,2,4……] byte 累加进 结果的奇数位内存 output[0],如果溢出,则进位给偶数位的 output[1];
2)所有偶数位[1,3,5……] byte 累加进 结果的偶数位内存 output[1],如果溢出,则进位给奇数位的 output[0];
3)最后对 output[2] 求反码即可
示例代码
#!/usr/bin/env python # -*- coding: utf-8 -*- import struct import sys def ip_cksum(s): a = 0 b = 0 # 偶数序号的 unsigned char 互相累加 for i in xrange(0, len(s), 2): a += struct.unpack('B', s[i])[0] # 奇数序号的 unsigned char 互相累加 for i in xrange(1, len(s), 2): b += struct.unpack('B', s[i])[0] # 缩小值为 unsigned char while a > 256 or b > 256: b += a/256 # a 超过 byte 的部分进位给 b a = a%256 a += b/256 # b 超过 byte 的部分进位给 a b = b%256 # 取反 a = ~a & 0xff b = ~b & 0xff # 校验和作为字符串 v = chr(a) + chr(b) # 校验和作为 unsigned short v = struct.unpack('H', v)[0] return v if __name__ == '__main__': for i in sys.argv[1:]: print ip_cksum(i)
关于TCP/IP 校验和计算的代码,网上很多,但不少都有些问题,这里作一番简单分析
1.最尾部 byte 处理依赖机序
来自 http://locklessinc.com/articles/tcp_checksum/ 的 C 代码片段:
1 unsigned short checksum1(const char *buf, unsigned size) 2 { 3 unsigned sum = 0; 4 int i; 5 6 /* Accumulate checksum */ 7 for (i = 0; i < size - 1; i += 2) 8 { 9 unsigned short word16 = *(unsigned short *) &buf[i]; 10 sum += word16; 11 } 12 13 /* Handle odd-sized case */ 14 if (size & 1) 15 { 16 unsigned short word16 = (unsigned char) buf[i]; 17 sum += word16; 18 } 19 20 /* Fold to get the ones-complement result */ 21 while (sum >> 16) sum = (sum & 0xFFFF)+(sum >> 16); 22 23 /* Invert to get the negative in ones-complement arithmetic */ 24 return ~sum; 25 }
注意第16行,对于buffer 长度非偶数情况的处理, 导致此代码只可在 Little-Endian (如x86) 机器上运行。只需对最后一个 byte 补一个’\0'的 byte,凑够两个 byte 然后转为 unsinged short 相加即可。
2.多内存块的计算
来自 python 网络包创建、解析库 dpkt 的代码 dpkt.py
1 try: 2 import dnet 3 def in_cksum_add(s, buf): 4 return dnet.ip_cksum_add(buf, s) 5 def in_cksum_done(s): 6 return socket.ntohs(dnet.ip_cksum_carry(s)) 7 except ImportError: 8 import array 9 def in_cksum_add(s, buf): 10 n = len(buf) 11 cnt = (n / 2) * 2 12 a = array.array('H', buf[:cnt]) 13 if cnt != n: 14 a.append(struct.unpack('H', buf[-1] + '\x00')[0]) 15 return s + sum(a) 16 def in_cksum_done(s): 17 s = (s >> 16) + (s & 0xffff) 18 s += (s >> 16) 19 return socket.ntohs(~s & 0xffff)
它这里会有两个实现,一个是调用dnet库的实现(见2-6行),一个是用python自己实现的版本(见8-19行)。
dnet 库是 C 实现的一个库,但和 dpkt 库是同一个作者,这里都有一个共同的问题:对于 in_cksum_add 进的内存块,如果为奇数长度,则尾部会追加一个byte '\x00' (见14行),这里就导致了问题。其实呢,尾部的那个 byte 应该留给下一个接下来的内存块一起计算,当且仅当所有的内存块都处理完毕(即 in_cksum_done 时),多余一个 byte 时才该追加 byte '\x00'。
3.经典的实现
来自 wireshark 的 in_cksum.c
1 /* 2 * Checksum routine for Internet Protocol family headers (Portable Version). 3 * 4 * This routine is very heavily used in the network 5 * code and should be modified for each CPU to be as fast as possible. 6 */ 7 8 #define ADDCARRY(x) {if ((x) > 65535) (x) -= 65535;} 9 #define REDUCE {l_util.l = sum; sum = l_util.s[0] + l_util.s[1]; ADDCARRY(sum);} 10 11 int 12 in_cksum(const vec_t *vec, int veclen) 13 { 14 register const guint16 *w; 15 register int sum = 0; 16 register int mlen = 0; 17 int byte_swapped = 0; 18 19 union { 20 guint8 c[2]; 21 guint16 s; 22 } s_util; 23 union { 24 guint16 s[2]; 25 guint32 l; 26 } l_util; 27 28 for (; veclen != 0; vec++, veclen--) { 29 if (vec->len == 0) 30 continue; 31 w = (const guint16 *)(const void *)vec->ptr; 32 if (mlen == -1) { 33 /* 34 * The first byte of this chunk is the continuation 35 * of a word spanning between this chunk and the 36 * last chunk. 37 * 38 * s_util.c[0] is already saved when scanning previous 39 * chunk. 40 */ 41 s_util.c[1] = *(const guint8 *)w; 42 sum += s_util.s; 43 w = (const guint16 *)(const void *)((const guint8 *)w + 1); 44 mlen = vec->len - 1; 45 } else 46 mlen = vec->len; 47 /* 48 * Force to even boundary. 49 */ 50 if ((1 & (unsigned long) w) && (mlen > 0)) { 51 REDUCE; 52 sum <<= 8; 53 s_util.c[0] = *(const guint8 *)w; 54 w = (const guint16 *)(const void *)((const guint8 *)w + 1); 55 mlen--; 56 byte_swapped = 1; 57 } 58 /* 59 * Unroll the loop to make overhead from 60 * branches &c small. 61 */ 62 while ((mlen -= 32) >= 0) { 63 sum += w[0]; sum += w[1]; sum += w[2]; sum += w[3]; 64 sum += w[4]; sum += w[5]; sum += w[6]; sum += w[7]; 65 sum += w[8]; sum += w[9]; sum += w[10]; sum += w[11]; 66 sum += w[12]; sum += w[13]; sum += w[14]; sum += w[15]; 67 w += 16; 68 } 69 mlen += 32; 70 while ((mlen -= 8) >= 0) { 71 sum += w[0]; sum += w[1]; sum += w[2]; sum += w[3]; 72 w += 4; 73 } 74 mlen += 8; 75 if (mlen == 0 && byte_swapped == 0) 76 continue; 77 REDUCE; 78 while ((mlen -= 2) >= 0) { 79 sum += *w++; 80 } 81 if (byte_swapped) { 82 REDUCE; 83 sum <<= 8; 84 byte_swapped = 0; 85 if (mlen == -1) { 86 s_util.c[1] = *(const guint8 *)w; 87 sum += s_util.s; 88 mlen = 0; 89 } else 90 mlen = -1; 91 } else if (mlen == -1) 92 s_util.c[0] = *(const guint8 *)w; 93 } 94 if (mlen == -1) { 95 /* The last mbuf has odd # of bytes. Follow the 96 standard (the odd byte may be shifted left by 8 bits 97 or not as determined by endian-ness of the machine) */ 98 s_util.c[1] = 0; 99 sum += s_util.s; 100 } 101 REDUCE; 102 return (~sum & 0xffff); 103 }
1)92行是当前内存块还余一个 byte ,则会 s_util 等待下个内存卡再处理——恰当的处理前面提到的第二个问题
2)94行是所有内存块处理完毕后,对尾部最后一个 byte 的处理 ——恰当的处理了前面提到的第一个问题
3)看点:指针非对齐的情况下处理
50行会先将未对其的1个 byte 暂存,这样可迫使指针对齐,但又为了让同奇位、同偶位内存相加,所以使 sum<<8;81行,如果前面sum是已经左移过的,则再次 sum<<8,让sum回归最初的奇偶次序
注:REDUCE 宏实现的功能是将大于 short 的值(即大于65535)转化为 short 能表示的值.
posted on 2013-06-22 22:45 JesseFang 阅读(2257) 评论(0) 编辑 收藏 举报