in_cksum.c revision 126891
1/*- 2 * Copyright (c) 1990 The Regents of the University of California. 3 * All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice, this list of conditions and the following disclaimer. 10 * 2. Redistributions in binary form must reproduce the above copyright 11 * notice, this list of conditions and the following disclaimer in the 12 * documentation and/or other materials provided with the distribution. 13 * 3. All advertising materials mentioning features or use of this software 14 * must display the following acknowledgement: 15 * This product includes software developed by the University of 16 * California, Berkeley and its contributors. 17 * 4. Neither the name of the University nor the names of its contributors 18 * may be used to endorse or promote products derived from this software 19 * without specific prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 22 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 23 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 24 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 25 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 26 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 27 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 28 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 29 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 30 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 31 * SUCH DAMAGE. 32 * 33 * from tahoe: in_cksum.c 1.2 86/01/05 34 * from: @(#)in_cksum.c 1.3 (Berkeley) 1/19/91 35 */ 36 37#include <sys/cdefs.h> 38__FBSDID("$FreeBSD: head/sys/i386/i386/in_cksum.c 126891 2004-03-12 21:45:33Z trhodes $"); 39 40/* 41 * MPsafe: alfred 42 */ 43#include <sys/param.h> 44#include <sys/systm.h> 45#include <sys/mbuf.h> 46 47#include <netinet/in.h> 48#include <netinet/in_systm.h> 49#include <netinet/ip.h> 50 51#include <machine/in_cksum.h> 52 53/* 54 * Checksum routine for Internet Protocol family headers. 55 * 56 * This routine is very heavily used in the network 57 * code and should be modified for each CPU to be as fast as possible. 58 * 59 * This implementation is 386 version. 60 */ 61 62#undef ADDCARRY 63#define ADDCARRY(x) if ((x) > 0xffff) (x) -= 0xffff 64#if !defined(__GNUC__) || defined(__INTEL_COMPILER) 65/* non gcc parts stolen from sys/alpha/alpha/in_cksum.c */ 66#define REDUCE32 \ 67 { \ 68 q_util.q = sum; \ 69 sum = q_util.s[0] + q_util.s[1] + q_util.s[2] + q_util.s[3]; \ 70 } 71#define REDUCE16 \ 72 { \ 73 q_util.q = sum; \ 74 l_util.l = q_util.s[0] + q_util.s[1] + q_util.s[2] + q_util.s[3]; \ 75 sum = l_util.s[0] + l_util.s[1]; \ 76 ADDCARRY(sum); \ 77 } 78#endif 79#define REDUCE {sum = (sum & 0xffff) + (sum >> 16); ADDCARRY(sum);} 80 81#if !defined(__GNUC__) || defined(__INTEL_COMPILER) 82static const u_int32_t in_masks[] = { 83 /*0 bytes*/ /*1 byte*/ /*2 bytes*/ /*3 bytes*/ 84 0x00000000, 0x000000FF, 0x0000FFFF, 0x00FFFFFF, /* offset 0 */ 85 0x00000000, 0x0000FF00, 0x00FFFF00, 0xFFFFFF00, /* offset 1 */ 86 0x00000000, 0x00FF0000, 0xFFFF0000, 0xFFFF0000, /* offset 2 */ 87 0x00000000, 0xFF000000, 0xFF000000, 0xFF000000, /* offset 3 */ 88}; 89 90union l_util { 91 u_int16_t s[2]; 92 u_int32_t l; 93}; 94union q_util { 95 u_int16_t s[4]; 96 u_int32_t l[2]; 97 u_int64_t q; 98}; 99 100static u_int64_t 101in_cksumdata(const u_int32_t *lw, int len) 102{ 103 u_int64_t sum = 0; 104 u_int64_t prefilled; 105 int offset; 106 union q_util q_util; 107 108 if ((3 & (long) lw) == 0 && len == 20) { 109 sum = (u_int64_t) lw[0] + lw[1] + lw[2] + lw[3] + lw[4]; 110 REDUCE32; 111 return sum; 112 } 113 114 if ((offset = 3 & (long) lw) != 0) { 115 const u_int32_t *masks = in_masks + (offset << 2); 116 lw = (u_int32_t *) (((long) lw) - offset); 117 sum = *lw++ & masks[len >= 3 ? 3 : len]; 118 len -= 4 - offset; 119 if (len <= 0) { 120 REDUCE32; 121 return sum; 122 } 123 } 124#if 0 125 /* 126 * Force to cache line boundary. 127 */ 128 offset = 32 - (0x1f & (long) lw); 129 if (offset < 32 && len > offset) { 130 len -= offset; 131 if (4 & offset) { 132 sum += (u_int64_t) lw[0]; 133 lw += 1; 134 } 135 if (8 & offset) { 136 sum += (u_int64_t) lw[0] + lw[1]; 137 lw += 2; 138 } 139 if (16 & offset) { 140 sum += (u_int64_t) lw[0] + lw[1] + lw[2] + lw[3]; 141 lw += 4; 142 } 143 } 144#endif 145 /* 146 * access prefilling to start load of next cache line. 147 * then add current cache line 148 * save result of prefilling for loop iteration. 149 */ 150 prefilled = lw[0]; 151 while ((len -= 32) >= 4) { 152 u_int64_t prefilling = lw[8]; 153 sum += prefilled + lw[1] + lw[2] + lw[3] 154 + lw[4] + lw[5] + lw[6] + lw[7]; 155 lw += 8; 156 prefilled = prefilling; 157 } 158 if (len >= 0) { 159 sum += prefilled + lw[1] + lw[2] + lw[3] 160 + lw[4] + lw[5] + lw[6] + lw[7]; 161 lw += 8; 162 } else { 163 len += 32; 164 } 165 while ((len -= 16) >= 0) { 166 sum += (u_int64_t) lw[0] + lw[1] + lw[2] + lw[3]; 167 lw += 4; 168 } 169 len += 16; 170 while ((len -= 4) >= 0) { 171 sum += (u_int64_t) *lw++; 172 } 173 len += 4; 174 if (len > 0) 175 sum += (u_int64_t) (in_masks[len] & *lw); 176 REDUCE32; 177 return sum; 178} 179 180u_short 181in_addword(u_short a, u_short b) 182{ 183 u_int64_t sum = a + b; 184 185 ADDCARRY(sum); 186 return (sum); 187} 188 189u_short 190in_pseudo(u_int32_t a, u_int32_t b, u_int32_t c) 191{ 192 u_int64_t sum; 193 union q_util q_util; 194 union l_util l_util; 195 196 sum = (u_int64_t) a + b + c; 197 REDUCE16; 198 return (sum); 199} 200 201u_short 202in_cksum_skip(struct mbuf *m, int len, int skip) 203{ 204 u_int64_t sum = 0; 205 int mlen = 0; 206 int clen = 0; 207 caddr_t addr; 208 union q_util q_util; 209 union l_util l_util; 210 211 len -= skip; 212 for (; skip && m; m = m->m_next) { 213 if (m->m_len > skip) { 214 mlen = m->m_len - skip; 215 addr = mtod(m, caddr_t) + skip; 216 goto skip_start; 217 } else { 218 skip -= m->m_len; 219 } 220 } 221 222 for (; m && len; m = m->m_next) { 223 if (m->m_len == 0) 224 continue; 225 mlen = m->m_len; 226 addr = mtod(m, caddr_t); 227skip_start: 228 if (len < mlen) 229 mlen = len; 230 if ((clen ^ (long) addr) & 1) 231 sum += in_cksumdata((const u_int32_t *)addr, mlen) << 8; 232 else 233 sum += in_cksumdata((const u_int32_t *)addr, mlen); 234 235 clen += mlen; 236 len -= mlen; 237 } 238 REDUCE16; 239 return (~sum & 0xffff); 240} 241 242u_int in_cksum_hdr(const struct ip *ip) 243{ 244 u_int64_t sum = in_cksumdata((const u_int32_t *)ip, sizeof(struct ip)); 245 union q_util q_util; 246 union l_util l_util; 247 248 REDUCE16; 249 return (~sum & 0xffff); 250} 251#else 252 253/* 254 * These asm statements require __volatile because they pass information 255 * via the condition codes. GCC does not currently provide a way to specify 256 * the condition codes as an input or output operand. 257 * 258 * The LOAD macro below is effectively a prefetch into cache. GCC will 259 * load the value into a register but will not use it. Since modern CPUs 260 * reorder operations, this will generally take place in parallel with 261 * other calculations. 262 */ 263#define ADD(n) __asm __volatile \ 264 ("addl %1, %0" : "+r" (sum) : \ 265 "g" (((const u_int32_t *)w)[n / 4])) 266#define ADDC(n) __asm __volatile \ 267 ("adcl %1, %0" : "+r" (sum) : \ 268 "g" (((const u_int32_t *)w)[n / 4])) 269#define LOAD(n) __asm __volatile \ 270 ("" : : "r" (((const u_int32_t *)w)[n / 4])) 271#define MOP __asm __volatile \ 272 ("adcl $0, %0" : "+r" (sum)) 273 274u_short 275in_cksum_skip(m, len, skip) 276 struct mbuf *m; 277 int len; 278 int skip; 279{ 280 register u_short *w; 281 register unsigned sum = 0; 282 register int mlen = 0; 283 int byte_swapped = 0; 284 union { char c[2]; u_short s; } su; 285 286 len -= skip; 287 for (; skip && m; m = m->m_next) { 288 if (m->m_len > skip) { 289 mlen = m->m_len - skip; 290 w = (u_short *)(mtod(m, u_char *) + skip); 291 goto skip_start; 292 } else { 293 skip -= m->m_len; 294 } 295 } 296 297 for (;m && len; m = m->m_next) { 298 if (m->m_len == 0) 299 continue; 300 w = mtod(m, u_short *); 301 if (mlen == -1) { 302 /* 303 * The first byte of this mbuf is the continuation 304 * of a word spanning between this mbuf and the 305 * last mbuf. 306 */ 307 308 /* su.c[0] is already saved when scanning previous 309 * mbuf. sum was REDUCEd when we found mlen == -1 310 */ 311 su.c[1] = *(u_char *)w; 312 sum += su.s; 313 w = (u_short *)((char *)w + 1); 314 mlen = m->m_len - 1; 315 len--; 316 } else 317 mlen = m->m_len; 318skip_start: 319 if (len < mlen) 320 mlen = len; 321 len -= mlen; 322 /* 323 * Force to long boundary so we do longword aligned 324 * memory operations 325 */ 326 if (3 & (int) w) { 327 REDUCE; 328 if ((1 & (int) w) && (mlen > 0)) { 329 sum <<= 8; 330 su.c[0] = *(char *)w; 331 w = (u_short *)((char *)w + 1); 332 mlen--; 333 byte_swapped = 1; 334 } 335 if ((2 & (int) w) && (mlen >= 2)) { 336 sum += *w++; 337 mlen -= 2; 338 } 339 } 340 /* 341 * Advance to a 486 cache line boundary. 342 */ 343 if (4 & (int) w && mlen >= 4) { 344 ADD(0); 345 MOP; 346 w += 2; 347 mlen -= 4; 348 } 349 if (8 & (int) w && mlen >= 8) { 350 ADD(0); 351 ADDC(4); 352 MOP; 353 w += 4; 354 mlen -= 8; 355 } 356 /* 357 * Do as much of the checksum as possible 32 bits at at time. 358 * In fact, this loop is unrolled to make overhead from 359 * branches &c small. 360 */ 361 mlen -= 1; 362 while ((mlen -= 32) >= 0) { 363 /* 364 * Add with carry 16 words and fold in the last 365 * carry by adding a 0 with carry. 366 * 367 * The early ADD(16) and the LOAD(32) are to load 368 * the next 2 cache lines in advance on 486's. The 369 * 486 has a penalty of 2 clock cycles for loading 370 * a cache line, plus whatever time the external 371 * memory takes to load the first word(s) addressed. 372 * These penalties are unavoidable. Subsequent 373 * accesses to a cache line being loaded (and to 374 * other external memory?) are delayed until the 375 * whole load finishes. These penalties are mostly 376 * avoided by not accessing external memory for 377 * 8 cycles after the ADD(16) and 12 cycles after 378 * the LOAD(32). The loop terminates when mlen 379 * is initially 33 (not 32) to guaranteed that 380 * the LOAD(32) is within bounds. 381 */ 382 ADD(16); 383 ADDC(0); 384 ADDC(4); 385 ADDC(8); 386 ADDC(12); 387 LOAD(32); 388 ADDC(20); 389 ADDC(24); 390 ADDC(28); 391 MOP; 392 w += 16; 393 } 394 mlen += 32 + 1; 395 if (mlen >= 32) { 396 ADD(16); 397 ADDC(0); 398 ADDC(4); 399 ADDC(8); 400 ADDC(12); 401 ADDC(20); 402 ADDC(24); 403 ADDC(28); 404 MOP; 405 w += 16; 406 mlen -= 32; 407 } 408 if (mlen >= 16) { 409 ADD(0); 410 ADDC(4); 411 ADDC(8); 412 ADDC(12); 413 MOP; 414 w += 8; 415 mlen -= 16; 416 } 417 if (mlen >= 8) { 418 ADD(0); 419 ADDC(4); 420 MOP; 421 w += 4; 422 mlen -= 8; 423 } 424 if (mlen == 0 && byte_swapped == 0) 425 continue; /* worth 1% maybe ?? */ 426 REDUCE; 427 while ((mlen -= 2) >= 0) { 428 sum += *w++; 429 } 430 if (byte_swapped) { 431 sum <<= 8; 432 byte_swapped = 0; 433 if (mlen == -1) { 434 su.c[1] = *(char *)w; 435 sum += su.s; 436 mlen = 0; 437 } else 438 mlen = -1; 439 } else if (mlen == -1) 440 /* 441 * This mbuf has odd number of bytes. 442 * There could be a word split betwen 443 * this mbuf and the next mbuf. 444 * Save the last byte (to prepend to next mbuf). 445 */ 446 su.c[0] = *(char *)w; 447 } 448 449 if (len) 450 printf("%s: out of data by %d\n", __func__, len); 451 if (mlen == -1) { 452 /* The last mbuf has odd # of bytes. Follow the 453 standard (the odd byte is shifted left by 8 bits) */ 454 su.c[1] = 0; 455 sum += su.s; 456 } 457 REDUCE; 458 return (~sum & 0xffff); 459} 460#endif 461