1 /*- 2 * SPDX-License-Identifier: BSD-3-Clause 3 * 4 * Copyright (c) 1990 The Regents of the University of California. 5 * All rights reserved. 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 1. Redistributions of source code must retain the above copyright 11 * notice, this list of conditions and the following disclaimer. 12 * 2. Redistributions in binary form must reproduce the above copyright 13 * notice, this list of conditions and the following disclaimer in the 14 * documentation and/or other materials provided with the distribution. 15 * 3. Neither the name of the University nor the names of its contributors 16 * may be used to endorse or promote products derived from this software 17 * without specific prior written permission. 18 * 19 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 20 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 21 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 22 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 23 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 24 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 25 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 26 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 27 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 28 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 29 * SUCH DAMAGE. 30 * 31 * from tahoe: in_cksum.c 1.2 86/01/05 32 */ 33 34 #include <sys/param.h> 35 #include <sys/systm.h> 36 #include <sys/mbuf.h> 37 38 #include <netinet/in.h> 39 #include <netinet/in_systm.h> 40 #include <netinet/ip.h> 41 42 #include <machine/in_cksum.h> 43 44 /* 45 * Checksum routine for Internet Protocol family headers. 46 * 47 * This routine is very heavily used in the network 48 * code and should be modified for each CPU to be as fast as possible. 49 * 50 * This implementation is 386 version. 51 */ 52 53 #undef ADDCARRY 54 #define ADDCARRY(x) if ((x) > 0xffff) (x) -= 0xffff 55 #define REDUCE {sum = (sum & 0xffff) + (sum >> 16); ADDCARRY(sum);} 56 57 /* 58 * These asm statements require __volatile because they pass information 59 * via the condition codes. GCC does not currently provide a way to specify 60 * the condition codes as an input or output operand. 61 * 62 * The LOAD macro below is effectively a prefetch into cache. GCC will 63 * load the value into a register but will not use it. Since modern CPUs 64 * reorder operations, this will generally take place in parallel with 65 * other calculations. 66 */ 67 u_short 68 in_cksum_skip(struct mbuf *m, int len, int skip) 69 { 70 u_short *w; 71 unsigned sum = 0; 72 int mlen = 0; 73 int byte_swapped = 0; 74 union { char c[2]; u_short s; } su; 75 76 len -= skip; 77 for (; skip && m; m = m->m_next) { 78 if (m->m_len > skip) { 79 mlen = m->m_len - skip; 80 w = (u_short *)(mtod(m, u_char *) + skip); 81 goto skip_start; 82 } else { 83 skip -= m->m_len; 84 } 85 } 86 87 for (;m && len; m = m->m_next) { 88 if (m->m_len == 0) 89 continue; 90 w = mtod(m, u_short *); 91 if (mlen == -1) { 92 /* 93 * The first byte of this mbuf is the continuation 94 * of a word spanning between this mbuf and the 95 * last mbuf. 96 */ 97 98 /* su.c[0] is already saved when scanning previous 99 * mbuf. sum was REDUCEd when we found mlen == -1 100 */ 101 su.c[1] = *(u_char *)w; 102 sum += su.s; 103 w = (u_short *)((char *)w + 1); 104 mlen = m->m_len - 1; 105 len--; 106 } else 107 mlen = m->m_len; 108 skip_start: 109 if (len < mlen) 110 mlen = len; 111 len -= mlen; 112 /* 113 * Force to long boundary so we do longword aligned 114 * memory operations 115 */ 116 if (3 & (int) w) { 117 REDUCE; 118 if ((1 & (int) w) && (mlen > 0)) { 119 sum <<= 8; 120 su.c[0] = *(char *)w; 121 w = (u_short *)((char *)w + 1); 122 mlen--; 123 byte_swapped = 1; 124 } 125 if ((2 & (int) w) && (mlen >= 2)) { 126 sum += *w++; 127 mlen -= 2; 128 } 129 } 130 /* 131 * Advance to a 486 cache line boundary. 132 */ 133 if (4 & (int) w && mlen >= 4) { 134 __asm __volatile ( 135 "addl %1, %0\n" 136 "adcl $0, %0" 137 : "+r" (sum) 138 : "g" (((const u_int32_t *)w)[0]) 139 ); 140 w += 2; 141 mlen -= 4; 142 } 143 if (8 & (int) w && mlen >= 8) { 144 __asm __volatile ( 145 "addl %1, %0\n" 146 "adcl %2, %0\n" 147 "adcl $0, %0" 148 : "+r" (sum) 149 : "g" (((const u_int32_t *)w)[0]), 150 "g" (((const u_int32_t *)w)[1]) 151 ); 152 w += 4; 153 mlen -= 8; 154 } 155 /* 156 * Do as much of the checksum as possible 32 bits at at time. 157 * In fact, this loop is unrolled to make overhead from 158 * branches &c small. 159 */ 160 mlen -= 1; 161 while ((mlen -= 32) >= 0) { 162 /* 163 * Add with carry 16 words and fold in the last 164 * carry by adding a 0 with carry. 165 * 166 * The early ADD(16) and the LOAD(32) are to load 167 * the next 2 cache lines in advance on 486's. The 168 * 486 has a penalty of 2 clock cycles for loading 169 * a cache line, plus whatever time the external 170 * memory takes to load the first word(s) addressed. 171 * These penalties are unavoidable. Subsequent 172 * accesses to a cache line being loaded (and to 173 * other external memory?) are delayed until the 174 * whole load finishes. These penalties are mostly 175 * avoided by not accessing external memory for 176 * 8 cycles after the ADD(16) and 12 cycles after 177 * the LOAD(32). The loop terminates when mlen 178 * is initially 33 (not 32) to guaranteed that 179 * the LOAD(32) is within bounds. 180 */ 181 __asm __volatile ( 182 "addl %1, %0\n" 183 "adcl %2, %0\n" 184 "adcl %3, %0\n" 185 "adcl %4, %0\n" 186 "adcl %5, %0\n" 187 "mov %6, %%eax\n" 188 "adcl %7, %0\n" 189 "adcl %8, %0\n" 190 "adcl %9, %0\n" 191 "adcl $0, %0" 192 : "+r" (sum) 193 : "g" (((const u_int32_t *)w)[4]), 194 "g" (((const u_int32_t *)w)[0]), 195 "g" (((const u_int32_t *)w)[1]), 196 "g" (((const u_int32_t *)w)[2]), 197 "g" (((const u_int32_t *)w)[3]), 198 "g" (((const u_int32_t *)w)[8]), 199 "g" (((const u_int32_t *)w)[5]), 200 "g" (((const u_int32_t *)w)[6]), 201 "g" (((const u_int32_t *)w)[7]) 202 : "eax" 203 ); 204 w += 16; 205 } 206 mlen += 32 + 1; 207 if (mlen >= 32) { 208 __asm __volatile ( 209 "addl %1, %0\n" 210 "adcl %2, %0\n" 211 "adcl %3, %0\n" 212 "adcl %4, %0\n" 213 "adcl %5, %0\n" 214 "adcl %6, %0\n" 215 "adcl %7, %0\n" 216 "adcl %8, %0\n" 217 "adcl $0, %0" 218 : "+r" (sum) 219 : "g" (((const u_int32_t *)w)[4]), 220 "g" (((const u_int32_t *)w)[0]), 221 "g" (((const u_int32_t *)w)[1]), 222 "g" (((const u_int32_t *)w)[2]), 223 "g" (((const u_int32_t *)w)[3]), 224 "g" (((const u_int32_t *)w)[5]), 225 "g" (((const u_int32_t *)w)[6]), 226 "g" (((const u_int32_t *)w)[7]) 227 ); 228 w += 16; 229 mlen -= 32; 230 } 231 if (mlen >= 16) { 232 __asm __volatile ( 233 "addl %1, %0\n" 234 "adcl %2, %0\n" 235 "adcl %3, %0\n" 236 "adcl %4, %0\n" 237 "adcl $0, %0" 238 : "+r" (sum) 239 : "g" (((const u_int32_t *)w)[0]), 240 "g" (((const u_int32_t *)w)[1]), 241 "g" (((const u_int32_t *)w)[2]), 242 "g" (((const u_int32_t *)w)[3]) 243 ); 244 w += 8; 245 mlen -= 16; 246 } 247 if (mlen >= 8) { 248 __asm __volatile ( 249 "addl %1, %0\n" 250 "adcl %2, %0\n" 251 "adcl $0, %0" 252 : "+r" (sum) 253 : "g" (((const u_int32_t *)w)[0]), 254 "g" (((const u_int32_t *)w)[1]) 255 ); 256 w += 4; 257 mlen -= 8; 258 } 259 if (mlen == 0 && byte_swapped == 0) 260 continue; /* worth 1% maybe ?? */ 261 REDUCE; 262 while ((mlen -= 2) >= 0) { 263 sum += *w++; 264 } 265 if (byte_swapped) { 266 sum <<= 8; 267 byte_swapped = 0; 268 if (mlen == -1) { 269 su.c[1] = *(char *)w; 270 sum += su.s; 271 mlen = 0; 272 } else 273 mlen = -1; 274 } else if (mlen == -1) 275 /* 276 * This mbuf has odd number of bytes. 277 * There could be a word split between 278 * this mbuf and the next mbuf. 279 * Save the last byte (to prepend to next mbuf). 280 */ 281 su.c[0] = *(char *)w; 282 } 283 284 if (len) 285 printf("%s: out of data by %d\n", __func__, len); 286 if (mlen == -1) { 287 /* The last mbuf has odd # of bytes. Follow the 288 standard (the odd byte is shifted left by 8 bits) */ 289 su.c[1] = 0; 290 sum += su.s; 291 } 292 REDUCE; 293 return (~sum & 0xffff); 294 } 295