1 /* 2 * Copyright 1994, 1995 Massachusetts Institute of Technology 3 * 4 * Permission to use, copy, modify, and distribute this software and 5 * its documentation for any purpose and without fee is hereby 6 * granted, provided that both the above copyright notice and this 7 * permission notice appear in all copies, that both the above 8 * copyright notice and this permission notice appear in all 9 * supporting documentation, and that the name of M.I.T. not be used 10 * in advertising or publicity pertaining to distribution of the 11 * software without specific, written prior permission. M.I.T. makes 12 * no representations about the suitability of this software for any 13 * purpose. It is provided "as is" without express or implied 14 * warranty. 15 * 16 * THIS SOFTWARE IS PROVIDED BY M.I.T. ``AS IS''. M.I.T. DISCLAIMS 17 * ALL EXPRESS OR IMPLIED WARRANTIES WITH REGARD TO THIS SOFTWARE, 18 * INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF 19 * MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE. IN NO EVENT 20 * SHALL M.I.T. BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 21 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 22 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF 23 * USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND 24 * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 25 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 26 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 27 * SUCH DAMAGE. 28 * 29 * $FreeBSD: src/sys/netinet/in_rmx.c,v 1.37.2.3 2002/08/09 14:49:23 ru Exp $ 30 * $DragonFly: src/sys/netinet/in_rmx.c,v 1.14 2006/04/11 06:59:34 dillon Exp $ 31 */ 32 33 /* 34 * This code does two things necessary for the enhanced TCP metrics to 35 * function in a useful manner: 36 * 1) It marks all non-host routes as `cloning', thus ensuring that 37 * every actual reference to such a route actually gets turned 38 * into a reference to a host route to the specific destination 39 * requested. 40 * 2) When such routes lose all their references, it arranges for them 41 * to be deleted in some random collection of circumstances, so that 42 * a large quantity of stale routing data is not kept in kernel memory 43 * indefinitely. See in_rtqtimo() below for the exact mechanism. 44 */ 45 46 #include "opt_carp.h" 47 48 #include <sys/param.h> 49 #include <sys/systm.h> 50 #include <sys/kernel.h> 51 #include <sys/sysctl.h> 52 #include <sys/socket.h> 53 #include <sys/mbuf.h> 54 #include <sys/syslog.h> 55 #include <sys/globaldata.h> 56 #include <sys/thread2.h> 57 58 #include <net/if.h> 59 #include <net/route.h> 60 #include <net/if_var.h> 61 #ifdef CARP 62 #include <net/if_types.h> 63 #endif 64 #include <net/netmsg2.h> 65 #include <net/netisr2.h> 66 #include <netinet/in.h> 67 #include <netinet/in_var.h> 68 #include <netinet/ip_var.h> 69 #include <netinet/ip_flow.h> 70 71 #define RTPRF_EXPIRING RTF_PROTO3 /* set on routes we manage */ 72 73 struct in_rtq_pcpu { 74 struct radix_node_head *rnh; 75 76 struct callout timo_ch; 77 struct netmsg_base timo_nmsg; 78 79 time_t lastdrain; 80 int draining; 81 struct netmsg_base drain_nmsg; 82 } __cachealign; 83 84 static void in_rtqtimo(void *); 85 86 static struct in_rtq_pcpu in_rtq_pcpu[MAXCPU]; 87 88 /* 89 * Do what we need to do when inserting a route. 90 */ 91 static struct radix_node * 92 in_addroute(const void *key, const void *mask, struct radix_node_head *head, 93 struct radix_node *nodes) 94 { 95 struct rtentry *rt = (struct rtentry *)nodes; 96 struct sockaddr_in *sin = (struct sockaddr_in *)rt_key(rt); 97 struct radix_node *ret; 98 struct in_ifaddr_container *iac; 99 struct in_ifaddr *ia; 100 101 /* 102 * For IP, mark routes to multicast addresses as such, because 103 * it's easy to do and might be useful (but this is much more 104 * dubious since it's so easy to inspect the address). 105 * 106 * For IP, all unicast non-host routes are automatically cloning. 107 */ 108 if (IN_MULTICAST(ntohl(sin->sin_addr.s_addr))) 109 rt->rt_flags |= RTF_MULTICAST; 110 111 if (!(rt->rt_flags & (RTF_HOST | RTF_CLONING | RTF_MULTICAST))) 112 rt->rt_flags |= RTF_PRCLONING; 113 114 /* 115 * Try to set RTF_BROADCAST or RTF_LOCAL for a host route. 116 * 117 * Skip this process if a host route already has RTF_LOCAL set, 118 * for example by ifa_maintain_loopback_route(). 119 * 120 * For host routes, we make sure that RTF_BROADCAST is set for 121 * anything that looks like a broadcast address. This way, we can 122 * avoid an expensive call to in_broadcast() in ip_output() most of 123 * the time (because the route passed to ip_output() is almost always 124 * a host route). 125 * 126 * For local routes, we set RTF_LOCAL to allow various shortcuts. 127 * 128 * A cloned network route will point to one of several possible 129 * addresses if an interface has aliases and must be repointed back to 130 * the correct address or arp_rtrequest() will not properly detect the 131 * local IP. 132 */ 133 if ((rt->rt_flags & (RTF_HOST | RTF_LOCAL)) == RTF_HOST) { 134 if (in_broadcast(sin->sin_addr, rt->rt_ifp)) { 135 rt->rt_flags |= RTF_BROADCAST; 136 } else if (satosin(rt->rt_ifa->ifa_addr)->sin_addr.s_addr == 137 sin->sin_addr.s_addr) { 138 rt->rt_flags |= RTF_LOCAL; 139 } else { 140 LIST_FOREACH(iac, INADDR_HASH(sin->sin_addr.s_addr), 141 ia_hash) { 142 ia = iac->ia; 143 if (sin->sin_addr.s_addr == 144 ia->ia_addr.sin_addr.s_addr) { 145 rt->rt_flags |= RTF_LOCAL; 146 IFAREF(&ia->ia_ifa); 147 IFAFREE(rt->rt_ifa); 148 rt->rt_ifa = &ia->ia_ifa; 149 rt->rt_ifp = rt->rt_ifa->ifa_ifp; 150 break; 151 } 152 } 153 } 154 } 155 156 if (rt->rt_rmx.rmx_mtu == 0 && 157 !(rt->rt_rmx.rmx_locks & RTV_MTU) && 158 rt->rt_ifp != NULL) 159 rt->rt_rmx.rmx_mtu = rt->rt_ifp->if_mtu; 160 161 ret = rn_addroute(key, mask, head, nodes); 162 if (ret == NULL && (rt->rt_flags & RTF_HOST)) { 163 struct rtentry *oldrt; 164 165 /* 166 * We are trying to add a host route, but can't. 167 * Find out if it is because of an ARP entry and 168 * delete it if so. 169 */ 170 oldrt = rtpurelookup((struct sockaddr *)sin); 171 if (oldrt != NULL) { 172 --oldrt->rt_refcnt; 173 if ((oldrt->rt_flags & RTF_LLINFO) && 174 (oldrt->rt_flags & RTF_HOST) && 175 oldrt->rt_gateway && 176 oldrt->rt_gateway->sa_family == AF_LINK) { 177 rtrequest(RTM_DELETE, rt_key(oldrt), 178 oldrt->rt_gateway, rt_mask(oldrt), 179 oldrt->rt_flags, NULL); 180 ret = rn_addroute(key, mask, head, nodes); 181 } 182 } 183 } 184 185 /* 186 * If the new route has been created successfully, and it is 187 * not a multicast/broadcast or cloned route, then we will 188 * have to flush the ipflow. Otherwise, we may end up using 189 * the wrong route. 190 */ 191 if (ret != NULL && 192 (rt->rt_flags & 193 (RTF_MULTICAST | RTF_BROADCAST | RTF_WASCLONED)) == 0) 194 ipflow_flush_oncpu(); 195 return ret; 196 } 197 198 /* 199 * This code is the inverse of in_closeroute: on first reference, if we 200 * were managing the route, stop doing so and set the expiration timer 201 * back off again. 202 */ 203 static struct radix_node * 204 in_matchroute(const void *key, struct radix_node_head *head) 205 { 206 struct radix_node *rn = rn_match(key, head); 207 struct rtentry *rt = (struct rtentry *)rn; 208 209 if (rt != NULL && rt->rt_refcnt == 0) { /* this is first reference */ 210 if (rt->rt_flags & RTPRF_EXPIRING) { 211 rt->rt_flags &= ~RTPRF_EXPIRING; 212 rt->rt_rmx.rmx_expire = 0; 213 } 214 } 215 return rn; 216 } 217 218 static int rtq_reallyold = 60*60; /* one hour is ``really old'' */ 219 SYSCTL_INT(_net_inet_ip, IPCTL_RTEXPIRE, rtexpire, CTLFLAG_RW, 220 &rtq_reallyold , 0, 221 "Default expiration time on cloned routes"); 222 223 static int rtq_minreallyold = 10; /* never automatically crank down to less */ 224 SYSCTL_INT(_net_inet_ip, IPCTL_RTMINEXPIRE, rtminexpire, CTLFLAG_RW, 225 &rtq_minreallyold , 0, 226 "Minimum time to attempt to hold onto cloned routes"); 227 228 static int rtq_toomany = 128; /* 128 cached routes is ``too many'' */ 229 SYSCTL_INT(_net_inet_ip, IPCTL_RTMAXCACHE, rtmaxcache, CTLFLAG_RW, 230 &rtq_toomany , 0, "Upper limit on cloned routes"); 231 232 /* 233 * On last reference drop, mark the route as belong to us so that it can be 234 * timed out. 235 */ 236 static void 237 in_closeroute(struct radix_node *rn, struct radix_node_head *head) 238 { 239 struct rtentry *rt = (struct rtentry *)rn; 240 241 if (!(rt->rt_flags & RTF_UP)) 242 return; /* prophylactic measures */ 243 244 if ((rt->rt_flags & (RTF_LLINFO | RTF_HOST)) != RTF_HOST) 245 return; 246 247 if ((rt->rt_flags & (RTF_WASCLONED | RTPRF_EXPIRING)) != RTF_WASCLONED) 248 return; 249 250 /* 251 * As requested by David Greenman: 252 * If rtq_reallyold is 0, just delete the route without 253 * waiting for a timeout cycle to kill it. 254 */ 255 if (rtq_reallyold != 0) { 256 rt->rt_flags |= RTPRF_EXPIRING; 257 rt->rt_rmx.rmx_expire = time_uptime + rtq_reallyold; 258 } else { 259 /* 260 * Remove route from the radix tree, but defer deallocation 261 * until we return to rtfree(). 262 */ 263 rtrequest(RTM_DELETE, rt_key(rt), rt->rt_gateway, rt_mask(rt), 264 rt->rt_flags, &rt); 265 } 266 } 267 268 struct rtqk_arg { 269 struct radix_node_head *rnh; 270 int draining; 271 int killed; 272 int found; 273 int updating; 274 time_t nextstop; 275 }; 276 277 /* 278 * Get rid of old routes. When draining, this deletes everything, even when 279 * the timeout is not expired yet. When updating, this makes sure that 280 * nothing has a timeout longer than the current value of rtq_reallyold. 281 */ 282 static int 283 in_rtqkill(struct radix_node *rn, void *rock) 284 { 285 struct rtqk_arg *ap = rock; 286 struct rtentry *rt = (struct rtentry *)rn; 287 int err; 288 289 if (rt->rt_flags & RTPRF_EXPIRING) { 290 ap->found++; 291 if (ap->draining || rt->rt_rmx.rmx_expire <= time_uptime) { 292 if (rt->rt_refcnt > 0) 293 panic("rtqkill route really not free"); 294 295 err = rtrequest(RTM_DELETE, rt_key(rt), rt->rt_gateway, 296 rt_mask(rt), rt->rt_flags, NULL); 297 if (err) 298 log(LOG_WARNING, "in_rtqkill: error %d\n", err); 299 else 300 ap->killed++; 301 } else { 302 if (ap->updating && 303 (int)(rt->rt_rmx.rmx_expire - time_uptime) > 304 rtq_reallyold) { 305 rt->rt_rmx.rmx_expire = time_uptime + 306 rtq_reallyold; 307 } 308 ap->nextstop = lmin(ap->nextstop, 309 rt->rt_rmx.rmx_expire); 310 } 311 } 312 313 return 0; 314 } 315 316 #define RTQ_TIMEOUT 60*10 /* run no less than once every ten minutes */ 317 static int rtq_timeout = RTQ_TIMEOUT; 318 319 /* 320 * NOTE: 321 * 'last_adjusted_timeout' and 'rtq_reallyold' are _not_ read-only, and 322 * could be changed by all CPUs. However, they are changed at so low 323 * frequency that we could ignore the cache trashing issue and take them 324 * as read-mostly. 325 */ 326 static void 327 in_rtqtimo_dispatch(netmsg_t nmsg) 328 { 329 struct rtqk_arg arg; 330 struct timeval atv; 331 static time_t last_adjusted_timeout = 0; 332 struct in_rtq_pcpu *pcpu = &in_rtq_pcpu[mycpuid]; 333 struct radix_node_head *rnh = pcpu->rnh; 334 335 ASSERT_NETISR_NCPUS(mycpuid); 336 337 /* Reply ASAP */ 338 crit_enter(); 339 lwkt_replymsg(&nmsg->lmsg, 0); 340 crit_exit(); 341 342 arg.found = arg.killed = 0; 343 arg.rnh = rnh; 344 arg.nextstop = time_uptime + rtq_timeout; 345 arg.draining = arg.updating = 0; 346 rnh->rnh_walktree(rnh, in_rtqkill, &arg); 347 348 /* 349 * Attempt to be somewhat dynamic about this: 350 * If there are ``too many'' routes sitting around taking up space, 351 * then crank down the timeout, and see if we can't make some more 352 * go away. However, we make sure that we will never adjust more 353 * than once in rtq_timeout seconds, to keep from cranking down too 354 * hard. 355 */ 356 if ((arg.found - arg.killed > rtq_toomany) && 357 (int)(time_uptime - last_adjusted_timeout) >= rtq_timeout && 358 rtq_reallyold > rtq_minreallyold) { 359 rtq_reallyold = 2*rtq_reallyold / 3; 360 if (rtq_reallyold < rtq_minreallyold) { 361 rtq_reallyold = rtq_minreallyold; 362 } 363 364 last_adjusted_timeout = time_uptime; 365 #ifdef DIAGNOSTIC 366 log(LOG_DEBUG, "in_rtqtimo: adjusted rtq_reallyold to %d\n", 367 rtq_reallyold); 368 #endif 369 arg.found = arg.killed = 0; 370 arg.updating = 1; 371 rnh->rnh_walktree(rnh, in_rtqkill, &arg); 372 } 373 374 atv.tv_usec = 0; 375 atv.tv_sec = arg.nextstop - time_uptime; 376 if ((int)atv.tv_sec < 1) { /* time shift safety */ 377 atv.tv_sec = 1; 378 arg.nextstop = time_uptime + atv.tv_sec; 379 } 380 if ((int)atv.tv_sec > rtq_timeout) { /* time shift safety */ 381 atv.tv_sec = rtq_timeout; 382 arg.nextstop = time_uptime + atv.tv_sec; 383 } 384 callout_reset(&pcpu->timo_ch, tvtohz_high(&atv), in_rtqtimo, NULL); 385 } 386 387 static void 388 in_rtqtimo(void *arg __unused) 389 { 390 int cpuid = mycpuid; 391 struct lwkt_msg *lmsg = &in_rtq_pcpu[cpuid].timo_nmsg.lmsg; 392 393 crit_enter(); 394 if (lmsg->ms_flags & MSGF_DONE) 395 lwkt_sendmsg_oncpu(netisr_cpuport(cpuid), lmsg); 396 crit_exit(); 397 } 398 399 static void 400 in_rtqdrain_oncpu(struct in_rtq_pcpu *pcpu) 401 { 402 struct radix_node_head *rnh = rt_tables[mycpuid][AF_INET]; 403 struct rtqk_arg arg; 404 405 ASSERT_NETISR_NCPUS(mycpuid); 406 407 arg.found = arg.killed = 0; 408 arg.rnh = rnh; 409 arg.nextstop = 0; 410 arg.draining = 1; 411 arg.updating = 0; 412 rnh->rnh_walktree(rnh, in_rtqkill, &arg); 413 414 pcpu->lastdrain = time_uptime; 415 } 416 417 static void 418 in_rtqdrain_dispatch(netmsg_t nmsg) 419 { 420 struct in_rtq_pcpu *pcpu = &in_rtq_pcpu[mycpuid]; 421 422 /* Reply ASAP */ 423 crit_enter(); 424 lwkt_replymsg(&nmsg->lmsg, 0); 425 crit_exit(); 426 427 in_rtqdrain_oncpu(pcpu); 428 pcpu->draining = 0; 429 } 430 431 static void 432 in_rtqdrain_ipi(void *arg __unused) 433 { 434 int cpu = mycpuid; 435 struct lwkt_msg *msg = &in_rtq_pcpu[cpu].drain_nmsg.lmsg; 436 437 crit_enter(); 438 if (msg->ms_flags & MSGF_DONE) 439 lwkt_sendmsg_oncpu(netisr_cpuport(cpu), msg); 440 crit_exit(); 441 } 442 443 void 444 in_rtqdrain(void) 445 { 446 cpumask_t mask; 447 int cpu; 448 449 CPUMASK_ASSBMASK(mask, netisr_ncpus); 450 CPUMASK_ANDMASK(mask, smp_active_mask); 451 452 cpu = mycpuid; 453 if (IN_NETISR_NCPUS(cpu)) { 454 in_rtqdrain_oncpu(&in_rtq_pcpu[cpu]); 455 CPUMASK_NANDBIT(mask, cpu); 456 } 457 458 for (cpu = 0; cpu < netisr_ncpus; ++cpu) { 459 struct in_rtq_pcpu *pcpu = &in_rtq_pcpu[cpu]; 460 461 if (!CPUMASK_TESTBIT(mask, cpu)) 462 continue; 463 464 if (pcpu->draining || pcpu->lastdrain == time_uptime) { 465 /* Just drained or is draining; skip this cpu. */ 466 CPUMASK_NANDBIT(mask, cpu); 467 continue; 468 } 469 pcpu->draining = 1; 470 } 471 472 if (CPUMASK_TESTNZERO(mask)) 473 lwkt_send_ipiq_mask(mask, in_rtqdrain_ipi, NULL); 474 } 475 476 /* 477 * Initialize our routing tree. 478 */ 479 int 480 in_inithead(void **head, int off) 481 { 482 struct radix_node_head *rnh; 483 struct in_rtq_pcpu *pcpu; 484 int cpuid = mycpuid; 485 486 rnh = *head; 487 KKASSERT(rnh == rt_tables[cpuid][AF_INET]); 488 489 if (!rn_inithead(&rnh, rn_cpumaskhead(cpuid), off)) 490 return 0; 491 492 *head = rnh; 493 rnh->rnh_addaddr = in_addroute; 494 rnh->rnh_matchaddr = in_matchroute; 495 rnh->rnh_close = in_closeroute; 496 497 pcpu = &in_rtq_pcpu[cpuid]; 498 pcpu->rnh = rnh; 499 callout_init_mp(&pcpu->timo_ch); 500 netmsg_init(&pcpu->timo_nmsg, NULL, &netisr_adone_rport, MSGF_PRIORITY, 501 in_rtqtimo_dispatch); 502 netmsg_init(&pcpu->drain_nmsg, NULL, &netisr_adone_rport, MSGF_PRIORITY, 503 in_rtqdrain_dispatch); 504 505 in_rtqtimo(NULL); /* kick off timeout first time */ 506 return 1; 507 } 508 509 /* 510 * This zaps old routes when the interface goes down or interface 511 * address is deleted. In the latter case, it deletes static routes 512 * that point to this address. If we don't do this, we may end up 513 * using the old address in the future. The ones we always want to 514 * get rid of are things like ARP entries, since the user might down 515 * the interface, walk over to a completely different network, and 516 * plug back in. 517 * 518 * in_ifadown() is typically called when an interface is being brought 519 * down. We must iterate through all per-cpu route tables and clean 520 * them up. 521 */ 522 struct in_ifadown_arg { 523 struct radix_node_head *rnh; 524 struct ifaddr *ifa; 525 int del; 526 }; 527 528 static int 529 in_ifadownkill(struct radix_node *rn, void *xap) 530 { 531 struct in_ifadown_arg *ap = xap; 532 struct rtentry *rt = (struct rtentry *)rn; 533 int err; 534 535 if (rt->rt_ifa == ap->ifa && 536 (ap->del || !(rt->rt_flags & RTF_STATIC))) { 537 /* 538 * We need to disable the automatic prune that happens 539 * in this case in rtrequest() because it will blow 540 * away the pointers that rn_walktree() needs in order 541 * continue our descent. We will end up deleting all 542 * the routes that rtrequest() would have in any case, 543 * so that behavior is not needed there. 544 */ 545 rt->rt_flags &= ~(RTF_CLONING | RTF_PRCLONING); 546 err = rtrequest(RTM_DELETE, rt_key(rt), rt->rt_gateway, 547 rt_mask(rt), rt->rt_flags, NULL); 548 if (err) 549 log(LOG_WARNING, "in_ifadownkill: error %d\n", err); 550 } 551 return 0; 552 } 553 554 struct netmsg_ifadown { 555 struct netmsg_base base; 556 struct ifaddr *ifa; 557 int del; 558 }; 559 560 static void 561 in_ifadown_dispatch(netmsg_t msg) 562 { 563 struct netmsg_ifadown *rmsg = (void *)msg; 564 struct radix_node_head *rnh; 565 struct ifaddr *ifa = rmsg->ifa; 566 struct in_ifadown_arg arg; 567 int cpu; 568 569 cpu = mycpuid; 570 ASSERT_NETISR_NCPUS(cpu); 571 572 arg.rnh = rnh = rt_tables[cpu][AF_INET]; 573 arg.ifa = ifa; 574 arg.del = rmsg->del; 575 rnh->rnh_walktree(rnh, in_ifadownkill, &arg); 576 ifa->ifa_flags &= ~IFA_ROUTE; 577 578 netisr_forwardmsg(&msg->base, cpu + 1); 579 } 580 581 int 582 in_ifadown_force(struct ifaddr *ifa, int delete) 583 { 584 struct netmsg_ifadown msg; 585 586 if (ifa->ifa_addr->sa_family != AF_INET) 587 return 1; 588 589 /* 590 * XXX individual requests are not independantly chained, 591 * which means that the per-cpu route tables will not be 592 * consistent in the middle of the operation. If routes 593 * related to the interface are manipulated while we are 594 * doing this the inconsistancy could trigger a panic. 595 */ 596 netmsg_init(&msg.base, NULL, &curthread->td_msgport, MSGF_PRIORITY, 597 in_ifadown_dispatch); 598 msg.ifa = ifa; 599 msg.del = delete; 600 netisr_domsg_global(&msg.base); 601 602 return 0; 603 } 604 605 int 606 in_ifadown(struct ifaddr *ifa, int delete) 607 { 608 #ifdef CARP 609 if (ifa->ifa_ifp->if_type == IFT_CARP) 610 return 0; 611 #endif 612 return in_ifadown_force(ifa, delete); 613 } 614