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 49 #include <net/if.h> 50 #include <net/if_var.h> 51 #include <net/netisr.h> 52 53 #include <netinet/in.h> 54 #include <netinet/in_pcb.h> 55 #include <netinet/in_rss.h> 56 #include <netinet/in_var.h> 57 #include <netinet/toeplitz.h> 58 59 /*- 60 * Operating system parts of receiver-side scaling (RSS), which allows 61 * network cards to direct flows to particular receive queues based on hashes 62 * of header tuples. This implementation aligns RSS buckets with connection 63 * groups at the TCP/IP layer, so each bucket is associated with exactly one 64 * group. As a result, the group lookup structures (and lock) should have an 65 * effective affinity with exactly one CPU. 66 * 67 * Network device drivers needing to configure RSS will query this framework 68 * for parameters, such as the current RSS key, hashing policies, number of 69 * bits, and indirection table mapping hashes to buckets and CPUs. They may 70 * provide their own supplementary information, such as queue<->CPU bindings. 71 * It is the responsibility of the network device driver to inject packets 72 * into the stack on as close to the right CPU as possible, if playing by RSS 73 * rules. 74 * 75 * TODO: 76 * 77 * - Synchronization for rss_key and other future-configurable parameters. 78 * - Event handler drivers can register to pick up RSS configuration changes. 79 * - Should we allow rss_basecpu to be configured? 80 * - Randomize key on boot. 81 * - IPv6 support. 82 * - Statistics on how often there's a misalignment between hardware 83 * placement and pcbgroup expectations. 84 */ 85 86 SYSCTL_NODE(_net_inet, OID_AUTO, rss, CTLFLAG_RW, 0, "Receive-side steering"); 87 88 /* 89 * Toeplitz is the only required hash function in the RSS spec, so use it by 90 * default. 91 */ 92 static u_int rss_hashalgo = RSS_HASH_TOEPLITZ; 93 SYSCTL_INT(_net_inet_rss, OID_AUTO, hashalgo, CTLFLAG_RD, &rss_hashalgo, 0, 94 "RSS hash algorithm"); 95 TUNABLE_INT("net.inet.rss.hashalgo", &rss_hashalgo); 96 97 /* 98 * Size of the indirection table; at most 128 entries per the RSS spec. We 99 * size it to at least 2 times the number of CPUs by default to allow useful 100 * rebalancing. If not set explicitly with a loader tunable, we tune based 101 * on the number of CPUs present. 102 * 103 * XXXRW: buckets might be better to use for the tunable than bits. 104 */ 105 static u_int rss_bits; 106 SYSCTL_INT(_net_inet_rss, OID_AUTO, bits, CTLFLAG_RD, &rss_bits, 0, 107 "RSS bits"); 108 TUNABLE_INT("net.inet.rss.bits", &rss_bits); 109 110 static u_int rss_mask; 111 SYSCTL_INT(_net_inet_rss, OID_AUTO, mask, CTLFLAG_RD, &rss_mask, 0, 112 "RSS mask"); 113 114 static const u_int rss_maxbits = RSS_MAXBITS; 115 SYSCTL_INT(_net_inet_rss, OID_AUTO, maxbits, CTLFLAG_RD, 116 __DECONST(int *, &rss_maxbits), 0, "RSS maximum bits"); 117 118 /* 119 * RSS's own count of the number of CPUs it could be using for processing. 120 * Bounded to 64 by RSS constants. 121 */ 122 static u_int rss_ncpus; 123 SYSCTL_INT(_net_inet_rss, OID_AUTO, ncpus, CTLFLAG_RD, &rss_ncpus, 0, 124 "Number of CPUs available to RSS"); 125 126 #define RSS_MAXCPUS (1 << (RSS_MAXBITS - 1)) 127 static const u_int rss_maxcpus = RSS_MAXCPUS; 128 SYSCTL_INT(_net_inet_rss, OID_AUTO, maxcpus, CTLFLAG_RD, 129 __DECONST(int *, &rss_maxcpus), 0, "RSS maximum CPUs that can be used"); 130 131 /* 132 * Variable exists just for reporting rss_bits in a user-friendly way. 133 */ 134 static u_int rss_buckets; 135 SYSCTL_INT(_net_inet_rss, OID_AUTO, buckets, CTLFLAG_RD, &rss_buckets, 0, 136 "RSS buckets"); 137 138 /* 139 * Base CPU number; devices will add this to all CPU numbers returned by the 140 * RSS indirection table. Currently unmodifable in FreeBSD. 141 */ 142 static const u_int rss_basecpu; 143 SYSCTL_INT(_net_inet_rss, OID_AUTO, basecpu, CTLFLAG_RD, 144 __DECONST(int *, &rss_basecpu), 0, "RSS base CPU"); 145 146 /* 147 * RSS secret key, intended to prevent attacks on load-balancing. Its 148 * effectiveness may be limited by algorithm choice and available entropy 149 * during the boot. 150 * 151 * XXXRW: And that we don't randomize it yet! 152 * 153 * XXXRW: This default is actually the default key from Chelsio T3 cards, as 154 * it offers reasonable distribution, unlike all-0 keys which always 155 * generate a hash of 0 (upsettingly). 156 */ 157 static uint8_t rss_key[RSS_KEYSIZE] = { 158 0x43, 0xa3, 0x8f, 0xb0, 0x41, 0x67, 0x25, 0x3d, 159 0x25, 0x5b, 0x0e, 0xc2, 0x6d, 0x5a, 0x56, 0xda, 160 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 161 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 162 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 163 }; 164 165 /* 166 * RSS hash->CPU table, which maps hashed packet headers to particular CPUs. 167 * Drivers may supplement this table with a seperate CPU<->queue table when 168 * programming devices. 169 */ 170 struct rss_table_entry { 171 uint8_t rte_cpu; /* CPU affinity of bucket. */ 172 }; 173 static struct rss_table_entry rss_table[RSS_TABLE_MAXLEN]; 174 175 static void 176 rss_init(__unused void *arg) 177 { 178 u_int i; 179 180 /* 181 * Validate tunables, coerce to sensible values. 182 */ 183 switch (rss_hashalgo) { 184 case RSS_HASH_TOEPLITZ: 185 case RSS_HASH_NAIVE: 186 break; 187 188 default: 189 printf("%s: invalid RSS hashalgo %u, coercing to %u", 190 __func__, rss_hashalgo, RSS_HASH_TOEPLITZ); 191 rss_hashalgo = RSS_HASH_TOEPLITZ; 192 } 193 194 /* 195 * Count available CPUs. 196 * 197 * XXXRW: Note incorrect assumptions regarding contiguity of this set 198 * elsewhere. 199 */ 200 rss_ncpus = 0; 201 for (i = 0; i <= mp_maxid; i++) { 202 if (CPU_ABSENT(i)) 203 continue; 204 rss_ncpus++; 205 } 206 if (rss_ncpus > RSS_MAXCPUS) 207 rss_ncpus = RSS_MAXCPUS; 208 209 /* 210 * Tune RSS table entries to be no less than 2x the number of CPUs 211 * -- unless we're running uniprocessor, in which case there's not 212 * much point in having buckets to rearrange for load-balancing! 213 */ 214 if (rss_ncpus > 1) { 215 if (rss_bits == 0) 216 rss_bits = fls(rss_ncpus - 1) + 1; 217 218 /* 219 * Microsoft limits RSS table entries to 128, so apply that 220 * limit to both auto-detected CPU counts and user-configured 221 * ones. 222 */ 223 if (rss_bits == 0 || rss_bits > RSS_MAXBITS) { 224 printf("%s: RSS bits %u not valid, coercing to %u", 225 __func__, rss_bits, RSS_MAXBITS); 226 rss_bits = RSS_MAXBITS; 227 } 228 229 /* 230 * Figure out how many buckets to use; warn if less than the 231 * number of configured CPUs, although this is not a fatal 232 * problem. 233 */ 234 rss_buckets = (1 << rss_bits); 235 if (rss_buckets < rss_ncpus) 236 printf("%s: WARNING: rss_buckets (%u) less than " 237 "rss_ncpus (%u)\n", __func__, rss_buckets, 238 rss_ncpus); 239 rss_mask = rss_buckets - 1; 240 } else { 241 rss_bits = 0; 242 rss_buckets = 1; 243 rss_mask = 0; 244 } 245 246 /* 247 * Set up initial CPU assignments: round-robin by default. 248 * 249 * XXXRW: Need a mapping to non-contiguous IDs here. 250 */ 251 for (i = 0; i < rss_buckets; i++) 252 rss_table[i].rte_cpu = i % rss_ncpus; 253 254 /* 255 * Randomize rrs_key. 256 * 257 * XXXRW: Not yet. If nothing else, will require an rss_isbadkey() 258 * loop to check for "bad" RSS keys. 259 */ 260 } 261 SYSINIT(rss_init, SI_SUB_SOFTINTR, SI_ORDER_SECOND, rss_init, NULL); 262 263 static uint32_t 264 rss_naive_hash(u_int keylen, const uint8_t *key, u_int datalen, 265 const uint8_t *data) 266 { 267 uint32_t v; 268 u_int i; 269 270 v = 0; 271 for (i = 0; i < keylen; i++) 272 v += key[i]; 273 for (i = 0; i < datalen; i++) 274 v += data[i]; 275 return (v); 276 } 277 278 static uint32_t 279 rss_hash(u_int datalen, const uint8_t *data) 280 { 281 282 switch (rss_hashalgo) { 283 case RSS_HASH_TOEPLITZ: 284 return (toeplitz_hash(sizeof(rss_key), rss_key, datalen, 285 data)); 286 287 case RSS_HASH_NAIVE: 288 return (rss_naive_hash(sizeof(rss_key), rss_key, datalen, 289 data)); 290 291 default: 292 panic("%s: unsupported/unknown hashalgo %d", __func__, 293 rss_hashalgo); 294 } 295 } 296 297 /* 298 * Hash an IPv4 2-tuple. 299 */ 300 uint32_t 301 rss_hash_ip4_2tuple(struct in_addr src, struct in_addr dst) 302 { 303 uint8_t data[sizeof(src) + sizeof(dst)]; 304 u_int datalen; 305 306 datalen = 0; 307 bcopy(&src, &data[datalen], sizeof(src)); 308 datalen += sizeof(src); 309 bcopy(&dst, &data[datalen], sizeof(dst)); 310 datalen += sizeof(dst); 311 return (rss_hash(datalen, data)); 312 } 313 314 /* 315 * Hash an IPv4 4-tuple. 316 */ 317 uint32_t 318 rss_hash_ip4_4tuple(struct in_addr src, u_short srcport, struct in_addr dst, 319 u_short dstport) 320 { 321 uint8_t data[sizeof(src) + sizeof(dst) + sizeof(srcport) + 322 sizeof(dstport)]; 323 u_int datalen; 324 325 datalen = 0; 326 bcopy(&src, &data[datalen], sizeof(src)); 327 datalen += sizeof(src); 328 bcopy(&dst, &data[datalen], sizeof(dst)); 329 datalen += sizeof(dst); 330 bcopy(&srcport, &data[datalen], sizeof(srcport)); 331 datalen += sizeof(srcport); 332 bcopy(&dstport, &data[datalen], sizeof(dstport)); 333 datalen += sizeof(dstport); 334 return (rss_hash(datalen, data)); 335 } 336 337 #ifdef INET6 338 /* 339 * Hash an IPv6 2-tuple. 340 */ 341 uint32_t 342 rss_hash_ip6_2tuple(struct in6_addr src, struct in6_addr dst) 343 { 344 uint8_t data[sizeof(src) + sizeof(dst)]; 345 u_int datalen; 346 347 datalen = 0; 348 bcopy(&src, &data[datalen], sizeof(src)); 349 datalen += sizeof(src); 350 bcopy(&dst, &data[datalen], sizeof(dst)); 351 datalen += sizeof(dst); 352 return (rss_hash(datalen, data)); 353 } 354 355 /* 356 * Hash an IPv6 4-tuple. 357 */ 358 uint32_t 359 rss_hash_ip6_4tuple(struct in6_addr src, u_short srcport, 360 struct in6_addr dst, u_short dstport) 361 { 362 uint8_t data[sizeof(src) + sizeof(dst) + sizeof(srcport) + 363 sizeof(dstport)]; 364 u_int datalen; 365 366 datalen = 0; 367 bcopy(&src, &data[datalen], sizeof(src)); 368 datalen += sizeof(src); 369 bcopy(&dst, &data[datalen], sizeof(dst)); 370 datalen += sizeof(dst); 371 bcopy(&srcport, &data[datalen], sizeof(srcport)); 372 datalen += sizeof(srcport); 373 bcopy(&dstport, &data[datalen], sizeof(dstport)); 374 datalen += sizeof(dstport); 375 return (rss_hash(datalen, data)); 376 } 377 #endif /* INET6 */ 378 379 /* 380 * Query the number of RSS bits in use. 381 */ 382 u_int 383 rss_getbits(void) 384 { 385 386 return (rss_bits); 387 } 388 389 /* 390 * Query the RSS bucket associated with an RSS hash. 391 */ 392 u_int 393 rss_getbucket(u_int hash) 394 { 395 396 return (hash & rss_mask); 397 } 398 399 /* 400 * Query the RSS CPU associated with an RSS bucket. 401 */ 402 u_int 403 rss_getcpu(u_int bucket) 404 { 405 406 return (rss_table[bucket].rte_cpu); 407 } 408 409 /* 410 * netisr CPU affinity lookup routine for use by protocols. 411 */ 412 struct mbuf * 413 rss_m2cpuid(struct mbuf *m, uintptr_t source, u_int *cpuid) 414 { 415 416 M_ASSERTPKTHDR(m); 417 418 switch (M_HASHTYPE_GET(m)) { 419 case M_HASHTYPE_RSS_IPV4: 420 case M_HASHTYPE_RSS_TCP_IPV4: 421 *cpuid = rss_getcpu(rss_getbucket(m->m_pkthdr.flowid)); 422 return (m); 423 424 default: 425 *cpuid = NETISR_CPUID_NONE; 426 return (m); 427 } 428 } 429 430 /* 431 * Query the RSS hash algorithm. 432 */ 433 u_int 434 rss_gethashalgo(void) 435 { 436 437 return (rss_hashalgo); 438 } 439 440 /* 441 * Query the current RSS key; likely to be used by device drivers when 442 * configuring hardware RSS. Caller must pass an array of size RSS_KEYSIZE. 443 * 444 * XXXRW: Perhaps we should do the accept-a-length-and-truncate thing? 445 */ 446 void 447 rss_getkey(uint8_t *key) 448 { 449 450 bcopy(rss_key, key, sizeof(rss_key)); 451 } 452 453 /* 454 * Query the number of buckets; this may be used by both network device 455 * drivers, which will need to populate hardware shadows of the software 456 * indirection table, and the network stack itself (such as when deciding how 457 * many connection groups to allocate). 458 */ 459 u_int 460 rss_getnumbuckets(void) 461 { 462 463 return (rss_buckets); 464 } 465 466 /* 467 * Query the number of CPUs in use by RSS; may be useful to device drivers 468 * trying to figure out how to map a larger number of CPUs into a smaller 469 * number of receive queues. 470 */ 471 u_int 472 rss_getnumcpus(void) 473 { 474 475 return (rss_ncpus); 476 } 477 478 /* 479 * XXXRW: Confirm that sysctl -a won't dump this keying material, don't want 480 * it appearing in debugging output unnecessarily. 481 */ 482 static int 483 sysctl_rss_key(SYSCTL_HANDLER_ARGS) 484 { 485 uint8_t temp_rss_key[RSS_KEYSIZE]; 486 int error; 487 488 error = priv_check(req->td, PRIV_NETINET_HASHKEY); 489 if (error) 490 return (error); 491 492 bcopy(rss_key, temp_rss_key, sizeof(temp_rss_key)); 493 error = sysctl_handle_opaque(oidp, temp_rss_key, 494 sizeof(temp_rss_key), req); 495 if (error) 496 return (error); 497 if (req->newptr != NULL) { 498 /* XXXRW: Not yet. */ 499 return (EINVAL); 500 } 501 return (0); 502 } 503 SYSCTL_PROC(_net_inet_rss, OID_AUTO, key, 504 CTLTYPE_OPAQUE | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0, sysctl_rss_key, 505 "", "RSS keying material"); 506