1 /*- 2 * Copyright (c) 2010-2011 Juniper Networks, Inc. 3 * All rights reserved. 4 * 5 * This software was developed by Robert N. M. Watson under contract 6 * to Juniper Networks, Inc. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 20 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 27 * SUCH DAMAGE. 28 */ 29 30 #include <sys/cdefs.h> 31 32 __FBSDID("$FreeBSD$"); 33 34 #include "opt_inet6.h" 35 #include "opt_pcbgroup.h" 36 37 #ifndef PCBGROUP 38 #error "options RSS depends on options PCBGROUP" 39 #endif 40 41 #include <sys/param.h> 42 #include <sys/mbuf.h> 43 #include <sys/socket.h> 44 #include <sys/priv.h> 45 #include <sys/kernel.h> 46 #include <sys/smp.h> 47 #include <sys/sysctl.h> 48 #include <sys/sbuf.h> 49 50 #include <net/if.h> 51 #include <net/if_var.h> 52 #include <net/netisr.h> 53 #include <net/rss_config.h> 54 55 #include <netinet/in.h> 56 #include <netinet/in_pcb.h> 57 #include <netinet/in_rss.h> 58 #include <netinet/in_var.h> 59 60 /* for software rss hash support */ 61 #include <netinet/ip.h> 62 #include <netinet/tcp.h> 63 #include <netinet/udp.h> 64 65 /* 66 * Hash an IPv4 2-tuple. 67 */ 68 uint32_t 69 rss_hash_ip4_2tuple(struct in_addr src, struct in_addr dst) 70 { 71 uint8_t data[sizeof(src) + sizeof(dst)]; 72 u_int datalen; 73 74 datalen = 0; 75 bcopy(&src, &data[datalen], sizeof(src)); 76 datalen += sizeof(src); 77 bcopy(&dst, &data[datalen], sizeof(dst)); 78 datalen += sizeof(dst); 79 return (rss_hash(datalen, data)); 80 } 81 82 /* 83 * Hash an IPv4 4-tuple. 84 */ 85 uint32_t 86 rss_hash_ip4_4tuple(struct in_addr src, u_short srcport, struct in_addr dst, 87 u_short dstport) 88 { 89 uint8_t data[sizeof(src) + sizeof(dst) + sizeof(srcport) + 90 sizeof(dstport)]; 91 u_int datalen; 92 93 datalen = 0; 94 bcopy(&src, &data[datalen], sizeof(src)); 95 datalen += sizeof(src); 96 bcopy(&dst, &data[datalen], sizeof(dst)); 97 datalen += sizeof(dst); 98 bcopy(&srcport, &data[datalen], sizeof(srcport)); 99 datalen += sizeof(srcport); 100 bcopy(&dstport, &data[datalen], sizeof(dstport)); 101 datalen += sizeof(dstport); 102 return (rss_hash(datalen, data)); 103 } 104 105 /* 106 * Calculate an appropriate ipv4 2-tuple or 4-tuple given the given 107 * IPv4 source/destination address, UDP or TCP source/destination ports 108 * and the protocol type. 109 * 110 * The protocol code may wish to do a software hash of the given 111 * tuple. This depends upon the currently configured RSS hash types. 112 * 113 * This assumes that the packet in question isn't a fragment. 114 * 115 * It also assumes the packet source/destination address 116 * are in "incoming" packet order (ie, source is "far" address.) 117 */ 118 int 119 rss_proto_software_hash_v4(struct in_addr s, struct in_addr d, 120 u_short sp, u_short dp, int proto, 121 uint32_t *hashval, uint32_t *hashtype) 122 { 123 uint32_t hash; 124 125 /* 126 * Next, choose the hash type depending upon the protocol 127 * identifier. 128 */ 129 if ((proto == IPPROTO_TCP) && 130 (rss_gethashconfig() & RSS_HASHTYPE_RSS_TCP_IPV4)) { 131 hash = rss_hash_ip4_4tuple(s, sp, d, dp); 132 *hashval = hash; 133 *hashtype = M_HASHTYPE_RSS_TCP_IPV4; 134 return (0); 135 } else if ((proto == IPPROTO_UDP) && 136 (rss_gethashconfig() & RSS_HASHTYPE_RSS_UDP_IPV4)) { 137 hash = rss_hash_ip4_4tuple(s, sp, d, dp); 138 *hashval = hash; 139 *hashtype = M_HASHTYPE_RSS_UDP_IPV4; 140 return (0); 141 } else if (rss_gethashconfig() & RSS_HASHTYPE_RSS_IPV4) { 142 /* RSS doesn't hash on other protocols like SCTP; so 2-tuple */ 143 hash = rss_hash_ip4_2tuple(s, d); 144 *hashval = hash; 145 *hashtype = M_HASHTYPE_RSS_IPV4; 146 return (0); 147 } 148 149 /* No configured available hashtypes! */ 150 RSS_DEBUG("no available hashtypes!\n"); 151 return (-1); 152 } 153 154 /* 155 * Calculate an appropriate ipv4 2-tuple or 4-tuple given the given 156 * IPv4 source/destination address, UDP or TCP source/destination ports 157 * and the protocol type. 158 * 159 * The protocol code may wish to do a software hash of the given 160 * tuple. This depends upon the currently configured RSS hash types. 161 * 162 * It assumes the packet source/destination address 163 * are in "outgoing" packet order (ie, destination is "far" address.) 164 */ 165 uint32_t 166 xps_proto_software_hash_v4(struct in_addr s, struct in_addr d, 167 u_short sp, u_short dp, int proto, uint32_t *hashtype) 168 { 169 uint32_t hash; 170 171 /* 172 * Next, choose the hash type depending upon the protocol 173 * identifier. 174 */ 175 if ((proto == IPPROTO_TCP) && 176 (rss_gethashconfig() & RSS_HASHTYPE_RSS_TCP_IPV4)) { 177 hash = rss_hash_ip4_4tuple(d, dp, s, sp); 178 *hashtype = M_HASHTYPE_RSS_TCP_IPV4; 179 return (hash); 180 } else if ((proto == IPPROTO_UDP) && 181 (rss_gethashconfig() & RSS_HASHTYPE_RSS_UDP_IPV4)) { 182 hash = rss_hash_ip4_4tuple(d, dp, s, sp); 183 *hashtype = M_HASHTYPE_RSS_UDP_IPV4; 184 return (hash); 185 } else if (rss_gethashconfig() & RSS_HASHTYPE_RSS_IPV4) { 186 /* RSS doesn't hash on other protocols like SCTP; so 2-tuple */ 187 hash = rss_hash_ip4_2tuple(d, s); 188 *hashtype = M_HASHTYPE_RSS_IPV4; 189 return (hash); 190 } 191 192 *hashtype = M_HASHTYPE_NONE; 193 return (0); 194 } 195 196 /* 197 * Do a software calculation of the RSS for the given mbuf. 198 * 199 * This is typically used by the input path to recalculate the RSS after 200 * some form of packet processing (eg de-capsulation, IP fragment reassembly.) 201 * 202 * dir is the packet direction - RSS_HASH_PKT_INGRESS for incoming and 203 * RSS_HASH_PKT_EGRESS for outgoing. 204 * 205 * Returns 0 if a hash was done, -1 if no hash was done, +1 if 206 * the mbuf already had a valid RSS flowid. 207 * 208 * This function doesn't modify the mbuf. It's up to the caller to 209 * assign flowid/flowtype as appropriate. 210 */ 211 int 212 rss_mbuf_software_hash_v4(const struct mbuf *m, int dir, uint32_t *hashval, 213 uint32_t *hashtype) 214 { 215 const struct ip *ip; 216 const struct tcphdr *th; 217 const struct udphdr *uh; 218 uint32_t flowid; 219 uint32_t flowtype; 220 uint8_t proto; 221 int iphlen; 222 int is_frag = 0; 223 224 /* 225 * XXX For now this only handles hashing on incoming mbufs. 226 */ 227 if (dir != RSS_HASH_PKT_INGRESS) { 228 RSS_DEBUG("called on EGRESS packet!\n"); 229 return (-1); 230 } 231 232 /* 233 * First, validate that the mbuf we have is long enough 234 * to have an IPv4 header in it. 235 */ 236 if (m->m_pkthdr.len < (sizeof(struct ip))) { 237 RSS_DEBUG("short mbuf pkthdr\n"); 238 return (-1); 239 } 240 if (m->m_len < (sizeof(struct ip))) { 241 RSS_DEBUG("short mbuf len\n"); 242 return (-1); 243 } 244 245 /* Ok, let's dereference that */ 246 ip = mtod(m, struct ip *); 247 proto = ip->ip_p; 248 iphlen = ip->ip_hl << 2; 249 250 /* 251 * If this is a fragment then it shouldn't be four-tuple 252 * hashed just yet. Once it's reassembled into a full 253 * frame it should be re-hashed. 254 */ 255 if (ip->ip_off & htons(IP_MF | IP_OFFMASK)) 256 is_frag = 1; 257 258 /* 259 * If the mbuf flowid/flowtype matches the packet type, 260 * and we don't support the 4-tuple version of the given protocol, 261 * then signal to the owner that it can trust the flowid/flowtype 262 * details. 263 * 264 * This is a little picky - eg, if TCPv4 / UDPv4 hashing 265 * is supported but we got a TCP/UDP frame only 2-tuple hashed, 266 * then we shouldn't just "trust" the 2-tuple hash. We need 267 * a 4-tuple hash. 268 */ 269 flowid = m->m_pkthdr.flowid; 270 flowtype = M_HASHTYPE_GET(m); 271 272 if (flowtype != M_HASHTYPE_NONE) { 273 switch (proto) { 274 case IPPROTO_UDP: 275 if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_UDP_IPV4) && 276 (flowtype == M_HASHTYPE_RSS_UDP_IPV4) && 277 (is_frag == 0)) { 278 return (1); 279 } 280 /* 281 * Only allow 2-tuple for UDP frames if we don't also 282 * support 4-tuple for UDP. 283 */ 284 if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_IPV4) && 285 ((rss_gethashconfig() & RSS_HASHTYPE_RSS_UDP_IPV4) == 0) && 286 flowtype == M_HASHTYPE_RSS_IPV4) { 287 return (1); 288 } 289 break; 290 case IPPROTO_TCP: 291 if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_TCP_IPV4) && 292 (flowtype == M_HASHTYPE_RSS_TCP_IPV4) && 293 (is_frag == 0)) { 294 return (1); 295 } 296 /* 297 * Only allow 2-tuple for TCP frames if we don't also 298 * support 2-tuple for TCP. 299 */ 300 if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_IPV4) && 301 ((rss_gethashconfig() & RSS_HASHTYPE_RSS_TCP_IPV4) == 0) && 302 flowtype == M_HASHTYPE_RSS_IPV4) { 303 return (1); 304 } 305 break; 306 default: 307 if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_IPV4) && 308 flowtype == M_HASHTYPE_RSS_IPV4) { 309 return (1); 310 } 311 break; 312 } 313 } 314 315 /* 316 * Decode enough information to make a hash decision. 317 * 318 * XXX TODO: does the hardware hash on 4-tuple if IP 319 * options are present? 320 */ 321 if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_TCP_IPV4) && 322 (proto == IPPROTO_TCP) && 323 (is_frag == 0)) { 324 if (m->m_len < iphlen + sizeof(struct tcphdr)) { 325 RSS_DEBUG("short TCP frame?\n"); 326 return (-1); 327 } 328 th = (const struct tcphdr *)((c_caddr_t)ip + iphlen); 329 return rss_proto_software_hash_v4(ip->ip_src, ip->ip_dst, 330 th->th_sport, 331 th->th_dport, 332 proto, 333 hashval, 334 hashtype); 335 } else if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_UDP_IPV4) && 336 (proto == IPPROTO_UDP) && 337 (is_frag == 0)) { 338 uh = (const struct udphdr *)((c_caddr_t)ip + iphlen); 339 if (m->m_len < iphlen + sizeof(struct udphdr)) { 340 RSS_DEBUG("short UDP frame?\n"); 341 return (-1); 342 } 343 return rss_proto_software_hash_v4(ip->ip_src, ip->ip_dst, 344 uh->uh_sport, 345 uh->uh_dport, 346 proto, 347 hashval, 348 hashtype); 349 } else if (rss_gethashconfig() & RSS_HASHTYPE_RSS_IPV4) { 350 /* Default to 2-tuple hash */ 351 return rss_proto_software_hash_v4(ip->ip_src, ip->ip_dst, 352 0, /* source port */ 353 0, /* destination port */ 354 0, /* IPPROTO_IP */ 355 hashval, 356 hashtype); 357 } else { 358 RSS_DEBUG("no available hashtypes!\n"); 359 return (-1); 360 } 361 } 362 363 /* 364 * Similar to rss_m2cpuid, but designed to be used by the IP NETISR 365 * on incoming frames. 366 * 367 * If an existing RSS hash exists and it matches what the configured 368 * hashing is, then use it. 369 * 370 * If there's an existing RSS hash but the desired hash is different, 371 * or if there's no useful RSS hash, then calculate it via 372 * the software path. 373 * 374 * XXX TODO: definitely want statistics here! 375 */ 376 struct mbuf * 377 rss_soft_m2cpuid_v4(struct mbuf *m, uintptr_t source, u_int *cpuid) 378 { 379 uint32_t hash_val, hash_type; 380 int ret; 381 382 M_ASSERTPKTHDR(m); 383 384 ret = rss_mbuf_software_hash_v4(m, RSS_HASH_PKT_INGRESS, 385 &hash_val, &hash_type); 386 if (ret > 0) { 387 /* mbuf has a valid hash already; don't need to modify it */ 388 *cpuid = rss_hash2cpuid(m->m_pkthdr.flowid, M_HASHTYPE_GET(m)); 389 } else if (ret == 0) { 390 /* hash was done; update */ 391 m->m_pkthdr.flowid = hash_val; 392 M_HASHTYPE_SET(m, hash_type); 393 *cpuid = rss_hash2cpuid(m->m_pkthdr.flowid, M_HASHTYPE_GET(m)); 394 } else { /* ret < 0 */ 395 /* no hash was done */ 396 *cpuid = NETISR_CPUID_NONE; 397 } 398 return (m); 399 } 400