sys/netinet/in_rss.c

   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  * Do a software calculation of the RSS for the given mbuf.
 156  *
 157  * This is typically used by the input path to recalculate the RSS after
 158  * some form of packet processing (eg de-capsulation, IP fragment reassembly.)
 159  *
 160  * dir is the packet direction - RSS_HASH_PKT_INGRESS for incoming and
 161  * RSS_HASH_PKT_EGRESS for outgoing.
 162  *
 163  * Returns 0 if a hash was done, -1 if no hash was done, +1 if
 164  * the mbuf already had a valid RSS flowid.
 165  *
 166  * This function doesn't modify the mbuf.  It's up to the caller to
 167  * assign flowid/flowtype as appropriate.
 168  */
 169 int
 170 rss_mbuf_software_hash_v4(const struct mbuf *m, int dir, uint32_t *hashval,
 171     uint32_t *hashtype)
 172 {
 173         const struct ip *ip;
 174         const struct tcphdr *th;
 175         const struct udphdr *uh;
 176         uint32_t flowid;
 177         uint32_t flowtype;
 178         uint8_t proto;
 179         int iphlen;
 180         int is_frag = 0;
 181
 182         /*
 183          * XXX For now this only handles hashing on incoming mbufs.
 184          */
 185         if (dir != RSS_HASH_PKT_INGRESS) {
 186                 RSS_DEBUG("called on EGRESS packet!\n");
 187                 return (-1);
 188         }
 189
 190         /*
 191          * First, validate that the mbuf we have is long enough
 192          * to have an IPv4 header in it.
 193          */
 194         if (m->m_pkthdr.len < (sizeof(struct ip))) {
 195                 RSS_DEBUG("short mbuf pkthdr\n");
 196                 return (-1);
 197         }
 198         if (m->m_len < (sizeof(struct ip))) {
 199                 RSS_DEBUG("short mbuf len\n");
 200                 return (-1);
 201         }
 202
 203         /* Ok, let's dereference that */
 204         ip = mtod(m, struct ip *);
 205         proto = ip->ip_p;
 206         iphlen = ip->ip_hl << 2;
 207
 208         /*
 209          * If this is a fragment then it shouldn't be four-tuple
 210          * hashed just yet.  Once it's reassembled into a full
 211          * frame it should be re-hashed.
 212          */
 213         if (ip->ip_off & htons(IP_MF | IP_OFFMASK))
 214                 is_frag = 1;
 215
 216         /*
 217          * If the mbuf flowid/flowtype matches the packet type,
 218          * and we don't support the 4-tuple version of the given protocol,
 219          * then signal to the owner that it can trust the flowid/flowtype
 220          * details.
 221          *
 222          * This is a little picky - eg, if TCPv4 / UDPv4 hashing
 223          * is supported but we got a TCP/UDP frame only 2-tuple hashed,
 224          * then we shouldn't just "trust" the 2-tuple hash.  We need
 225          * a 4-tuple hash.
 226          */
 227         flowid = m->m_pkthdr.flowid;
 228         flowtype = M_HASHTYPE_GET(m);
 229
 230         if (flowtype != M_HASHTYPE_NONE) {
 231                 switch (proto) {
 232                 case IPPROTO_UDP:
 233                         if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_UDP_IPV4) &&
 234                             (flowtype == M_HASHTYPE_RSS_UDP_IPV4) &&
 235                             (is_frag == 0)) {
 236                                 return (1);
 237                         }
 238                         /*
 239                          * Only allow 2-tuple for UDP frames if we don't also
 240                          * support 4-tuple for UDP.
 241                          */
 242                         if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_IPV4) &&
 243                             ((rss_gethashconfig() & RSS_HASHTYPE_RSS_UDP_IPV4) == 0) &&
 244                             flowtype == M_HASHTYPE_RSS_IPV4) {
 245                                 return (1);
 246                         }
 247                         break;
 248                 case IPPROTO_TCP:
 249                         if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_TCP_IPV4) &&
 250                             (flowtype == M_HASHTYPE_RSS_TCP_IPV4) &&
 251                             (is_frag == 0)) {
 252                                 return (1);
 253                         }
 254                         /*
 255                          * Only allow 2-tuple for TCP frames if we don't also
 256                          * support 2-tuple for TCP.
 257                          */
 258                         if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_IPV4) &&
 259                             ((rss_gethashconfig() & RSS_HASHTYPE_RSS_TCP_IPV4) == 0) &&
 260                             flowtype == M_HASHTYPE_RSS_IPV4) {
 261                                 return (1);
 262                         }
 263                         break;
 264                 default:
 265                         if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_IPV4) &&
 266                             flowtype == M_HASHTYPE_RSS_IPV4) {
 267                                 return (1);
 268                         }
 269                         break;
 270                 }
 271         }
 272
 273         /*
 274          * Decode enough information to make a hash decision.
 275          *
 276          * XXX TODO: does the hardware hash on 4-tuple if IP
 277          *    options are present?
 278          */
 279         if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_TCP_IPV4) &&
 280             (proto == IPPROTO_TCP) &&
 281             (is_frag == 0)) {
 282                 if (m->m_len < iphlen + sizeof(struct tcphdr)) {
 283                         RSS_DEBUG("short TCP frame?\n");
 284                         return (-1);
 285                 }
 286                 th = (const struct tcphdr *)((c_caddr_t)ip + iphlen);
 287                 return rss_proto_software_hash_v4(ip->ip_src, ip->ip_dst,
 288                     th->th_sport,
 289                     th->th_dport,
 290                     proto,
 291                     hashval,
 292                     hashtype);
 293         } else if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_UDP_IPV4) &&
 294             (proto == IPPROTO_UDP) &&
 295             (is_frag == 0)) {
 296                 uh = (const struct udphdr *)((c_caddr_t)ip + iphlen);
 297                 if (m->m_len < iphlen + sizeof(struct udphdr)) {
 298                         RSS_DEBUG("short UDP frame?\n");
 299                         return (-1);
 300                 }
 301                 return rss_proto_software_hash_v4(ip->ip_src, ip->ip_dst,
 302                     uh->uh_sport,
 303                     uh->uh_dport,
 304                     proto,
 305                     hashval,
 306                     hashtype);
 307         } else if (rss_gethashconfig() & RSS_HASHTYPE_RSS_IPV4) {
 308                 /* Default to 2-tuple hash */
 309                 return rss_proto_software_hash_v4(ip->ip_src, ip->ip_dst,
 310                     0,  /* source port */
 311                     0,  /* destination port */
 312                     0,  /* IPPROTO_IP */
 313                     hashval,
 314                     hashtype);
 315         } else {
 316                 RSS_DEBUG("no available hashtypes!\n");
 317                 return (-1);
 318         }
 319 }
 320
 321 /*
 322  * Similar to rss_m2cpuid, but designed to be used by the IP NETISR
 323  * on incoming frames.
 324  *
 325  * If an existing RSS hash exists and it matches what the configured
 326  * hashing is, then use it.
 327  *
 328  * If there's an existing RSS hash but the desired hash is different,
 329  * or if there's no useful RSS hash, then calculate it via
 330  * the software path.
 331  *
 332  * XXX TODO: definitely want statistics here!
 333  */
 334 struct mbuf *
 335 rss_soft_m2cpuid_v4(struct mbuf *m, uintptr_t source, u_int *cpuid)
 336 {
 337         uint32_t hash_val, hash_type;
 338         int ret;
 339
 340         M_ASSERTPKTHDR(m);
 341
 342         ret = rss_mbuf_software_hash_v4(m, RSS_HASH_PKT_INGRESS,
 343             &hash_val, &hash_type);
 344         if (ret > 0) {
 345                 /* mbuf has a valid hash already; don't need to modify it */
 346                 *cpuid = rss_hash2cpuid(m->m_pkthdr.flowid, M_HASHTYPE_GET(m));
 347         } else if (ret == 0) {
 348                 /* hash was done; update */
 349                 m->m_pkthdr.flowid = hash_val;
 350                 M_HASHTYPE_SET(m, hash_type);
 351                 *cpuid = rss_hash2cpuid(m->m_pkthdr.flowid, M_HASHTYPE_GET(m));
 352         } else { /* ret < 0 */
 353                 /* no hash was done */
 354                 *cpuid = NETISR_CPUID_NONE;
 355         }
 356         return (m);
 357 }