[4de817d] | 1 | /* |
---|
| 2 | * Checksum routine for Internet Protocol family headers. |
---|
| 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 | * This implementation is 386 version. |
---|
| 8 | * |
---|
| 9 | * $Id$ |
---|
| 10 | */ |
---|
| 11 | |
---|
[eb0d7ae] | 12 | #include <stdio.h> /* for puts */ |
---|
| 13 | |
---|
[4de817d] | 14 | #undef ADDCARRY |
---|
| 15 | #define ADDCARRY(x) if ((x) > 0xffff) (x) -= 0xffff |
---|
| 16 | #define REDUCE {sum = (sum & 0xffff) + (sum >> 16); ADDCARRY(sum);} |
---|
| 17 | |
---|
| 18 | /* |
---|
| 19 | * Thanks to gcc we don't have to guess |
---|
| 20 | * which registers contain sum & w. |
---|
| 21 | */ |
---|
[d50c0d2] | 22 | #define ADD(n) __asm__ volatile \ |
---|
| 23 | ("addl " #n "(%2), %0" : "=r" (sum) : "0" (sum), "r" (w)) |
---|
| 24 | #define ADDC(n) __asm__ volatile \ |
---|
| 25 | ("adcl " #n "(%2), %0" : "=r" (sum) : "0" (sum), "r" (w)) |
---|
| 26 | #define LOAD(n) __asm__ volatile \ |
---|
| 27 | ("movb " #n "(%1), %0" : "=r" (junk) : "r" (w)) |
---|
| 28 | #define MOP __asm__ volatile \ |
---|
| 29 | ("adcl $0, %0" : "=r" (sum) : "0" (sum)) |
---|
[4de817d] | 30 | |
---|
| 31 | int |
---|
| 32 | in_cksum(m, len) |
---|
| 33 | register struct mbuf *m; |
---|
| 34 | register int len; |
---|
| 35 | { |
---|
| 36 | register u_short *w; |
---|
| 37 | register unsigned sum = 0; |
---|
| 38 | register int mlen = 0; |
---|
| 39 | int byte_swapped = 0; |
---|
| 40 | union { char c[2]; u_short s; } su; |
---|
| 41 | |
---|
| 42 | for (;m && len; m = m->m_next) { |
---|
| 43 | if (m->m_len == 0) |
---|
| 44 | continue; |
---|
| 45 | w = mtod(m, u_short *); |
---|
| 46 | if (mlen == -1) { |
---|
| 47 | /* |
---|
| 48 | * The first byte of this mbuf is the continuation |
---|
| 49 | * of a word spanning between this mbuf and the |
---|
| 50 | * last mbuf. |
---|
| 51 | */ |
---|
| 52 | |
---|
| 53 | /* su.c[0] is already saved when scanning previous |
---|
| 54 | * mbuf. sum was REDUCEd when we found mlen == -1 |
---|
| 55 | */ |
---|
| 56 | su.c[1] = *(u_char *)w; |
---|
| 57 | sum += su.s; |
---|
| 58 | w = (u_short *)((char *)w + 1); |
---|
| 59 | mlen = m->m_len - 1; |
---|
| 60 | len--; |
---|
| 61 | } else |
---|
| 62 | mlen = m->m_len; |
---|
| 63 | if (len < mlen) |
---|
| 64 | mlen = len; |
---|
| 65 | len -= mlen; |
---|
| 66 | /* |
---|
| 67 | * Force to long boundary so we do longword aligned |
---|
| 68 | * memory operations |
---|
| 69 | */ |
---|
| 70 | if (3 & (int) w) { |
---|
| 71 | REDUCE; |
---|
| 72 | if ((1 & (int) w) && (mlen > 0)) { |
---|
| 73 | sum <<= 8; |
---|
| 74 | su.c[0] = *(char *)w; |
---|
| 75 | w = (u_short *)((char *)w + 1); |
---|
| 76 | mlen--; |
---|
| 77 | byte_swapped = 1; |
---|
| 78 | } |
---|
| 79 | if ((2 & (int) w) && (mlen >= 2)) { |
---|
| 80 | sum += *w++; |
---|
| 81 | mlen -= 2; |
---|
| 82 | } |
---|
| 83 | } |
---|
| 84 | /* |
---|
| 85 | * Advance to a 486 cache line boundary. |
---|
| 86 | */ |
---|
| 87 | if (4 & (int) w && mlen >= 4) { |
---|
| 88 | ADD(0); |
---|
| 89 | MOP; |
---|
| 90 | w += 2; |
---|
| 91 | mlen -= 4; |
---|
| 92 | } |
---|
| 93 | if (8 & (int) w && mlen >= 8) { |
---|
| 94 | ADD(0); |
---|
| 95 | ADDC(4); |
---|
| 96 | MOP; |
---|
| 97 | w += 4; |
---|
| 98 | mlen -= 8; |
---|
| 99 | } |
---|
| 100 | /* |
---|
| 101 | * Do as much of the checksum as possible 32 bits at at time. |
---|
| 102 | * In fact, this loop is unrolled to make overhead from |
---|
| 103 | * branches &c small. |
---|
| 104 | */ |
---|
| 105 | mlen -= 1; |
---|
| 106 | while ((mlen -= 32) >= 0) { |
---|
| 107 | u_char junk; |
---|
| 108 | /* |
---|
| 109 | * Add with carry 16 words and fold in the last |
---|
| 110 | * carry by adding a 0 with carry. |
---|
| 111 | * |
---|
| 112 | * The early ADD(16) and the LOAD(32) are to load |
---|
| 113 | * the next 2 cache lines in advance on 486's. The |
---|
| 114 | * 486 has a penalty of 2 clock cycles for loading |
---|
| 115 | * a cache line, plus whatever time the external |
---|
| 116 | * memory takes to load the first word(s) addressed. |
---|
| 117 | * These penalties are unavoidable. Subsequent |
---|
| 118 | * accesses to a cache line being loaded (and to |
---|
| 119 | * other external memory?) are delayed until the |
---|
| 120 | * whole load finishes. These penalties are mostly |
---|
| 121 | * avoided by not accessing external memory for |
---|
| 122 | * 8 cycles after the ADD(16) and 12 cycles after |
---|
| 123 | * the LOAD(32). The loop terminates when mlen |
---|
| 124 | * is initially 33 (not 32) to guaranteed that |
---|
| 125 | * the LOAD(32) is within bounds. |
---|
| 126 | */ |
---|
| 127 | ADD(16); |
---|
| 128 | ADDC(0); |
---|
| 129 | ADDC(4); |
---|
| 130 | ADDC(8); |
---|
| 131 | ADDC(12); |
---|
| 132 | LOAD(32); |
---|
| 133 | ADDC(20); |
---|
| 134 | ADDC(24); |
---|
| 135 | ADDC(28); |
---|
| 136 | MOP; |
---|
| 137 | w += 16; |
---|
| 138 | } |
---|
| 139 | mlen += 32 + 1; |
---|
| 140 | if (mlen >= 32) { |
---|
| 141 | ADD(16); |
---|
| 142 | ADDC(0); |
---|
| 143 | ADDC(4); |
---|
| 144 | ADDC(8); |
---|
| 145 | ADDC(12); |
---|
| 146 | ADDC(20); |
---|
| 147 | ADDC(24); |
---|
| 148 | ADDC(28); |
---|
| 149 | MOP; |
---|
| 150 | w += 16; |
---|
| 151 | mlen -= 32; |
---|
| 152 | } |
---|
| 153 | if (mlen >= 16) { |
---|
| 154 | ADD(0); |
---|
| 155 | ADDC(4); |
---|
| 156 | ADDC(8); |
---|
| 157 | ADDC(12); |
---|
| 158 | MOP; |
---|
| 159 | w += 8; |
---|
| 160 | mlen -= 16; |
---|
| 161 | } |
---|
| 162 | if (mlen >= 8) { |
---|
| 163 | ADD(0); |
---|
| 164 | ADDC(4); |
---|
| 165 | MOP; |
---|
| 166 | w += 4; |
---|
| 167 | mlen -= 8; |
---|
| 168 | } |
---|
| 169 | if (mlen == 0 && byte_swapped == 0) |
---|
| 170 | continue; /* worth 1% maybe ?? */ |
---|
| 171 | REDUCE; |
---|
| 172 | while ((mlen -= 2) >= 0) { |
---|
| 173 | sum += *w++; |
---|
| 174 | } |
---|
| 175 | if (byte_swapped) { |
---|
| 176 | sum <<= 8; |
---|
| 177 | byte_swapped = 0; |
---|
| 178 | if (mlen == -1) { |
---|
| 179 | su.c[1] = *(char *)w; |
---|
| 180 | sum += su.s; |
---|
| 181 | mlen = 0; |
---|
| 182 | } else |
---|
| 183 | mlen = -1; |
---|
| 184 | } else if (mlen == -1) |
---|
| 185 | /* |
---|
| 186 | * This mbuf has odd number of bytes. |
---|
| 187 | * There could be a word split betwen |
---|
| 188 | * this mbuf and the next mbuf. |
---|
| 189 | * Save the last byte (to prepend to next mbuf). |
---|
| 190 | */ |
---|
| 191 | su.c[0] = *(char *)w; |
---|
| 192 | } |
---|
| 193 | |
---|
| 194 | if (len) |
---|
[eb0d7ae] | 195 | puts("cksum: out of data"); |
---|
[4de817d] | 196 | if (mlen == -1) { |
---|
| 197 | /* The last mbuf has odd # of bytes. Follow the |
---|
| 198 | standard (the odd byte is shifted left by 8 bits) */ |
---|
| 199 | su.c[1] = 0; |
---|
| 200 | sum += su.s; |
---|
| 201 | } |
---|
| 202 | REDUCE; |
---|
| 203 | return (~sum & 0xffff); |
---|
| 204 | } |
---|