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_rtqtimo_ctx { 74 struct callout timo_ch; 75 struct netmsg_base timo_nmsg; 76 struct radix_node_head *timo_rnh; 77 } __cachealign; 78 79 static void in_rtqtimo(void *); 80 81 static struct in_rtqtimo_ctx in_rtqtimo_context[MAXCPU]; 82 83 /* 84 * Do what we need to do when inserting a route. 85 */ 86 static struct radix_node * 87 in_addroute(char *key, char *mask, struct radix_node_head *head, 88 struct radix_node *treenodes) 89 { 90 struct rtentry *rt = (struct rtentry *)treenodes; 91 struct sockaddr_in *sin = (struct sockaddr_in *)rt_key(rt); 92 struct radix_node *ret; 93 struct in_ifaddr_container *iac; 94 struct in_ifaddr *ia; 95 96 /* 97 * For IP, mark routes to multicast addresses as such, because 98 * it's easy to do and might be useful (but this is much more 99 * dubious since it's so easy to inspect the address). 100 * 101 * For IP, all unicast non-host routes are automatically cloning. 102 */ 103 if (IN_MULTICAST(ntohl(sin->sin_addr.s_addr))) 104 rt->rt_flags |= RTF_MULTICAST; 105 106 if (!(rt->rt_flags & (RTF_HOST | RTF_CLONING | RTF_MULTICAST))) 107 rt->rt_flags |= RTF_PRCLONING; 108 109 /* 110 * For host routes, we make sure that RTF_BROADCAST 111 * is set for anything that looks like a broadcast address. 112 * This way, we can avoid an expensive call to in_broadcast() 113 * in ip_output() most of the time (because the route passed 114 * to ip_output() is almost always a host route). 115 * 116 * For local routes we set RTF_LOCAL allowing various shortcuts. 117 * 118 * A cloned network route will point to one of several possible 119 * addresses if an interface has aliases and must be repointed 120 * back to the correct address or arp_rtrequest() will not properly 121 * detect the local ip. 122 */ 123 if (rt->rt_flags & RTF_HOST) { 124 if (in_broadcast(sin->sin_addr, rt->rt_ifp)) { 125 rt->rt_flags |= RTF_BROADCAST; 126 } else if (satosin(rt->rt_ifa->ifa_addr)->sin_addr.s_addr == 127 sin->sin_addr.s_addr) { 128 rt->rt_flags |= RTF_LOCAL; 129 } else { 130 LIST_FOREACH(iac, INADDR_HASH(sin->sin_addr.s_addr), 131 ia_hash) { 132 ia = iac->ia; 133 if (sin->sin_addr.s_addr == 134 ia->ia_addr.sin_addr.s_addr) { 135 rt->rt_flags |= RTF_LOCAL; 136 IFAREF(&ia->ia_ifa); 137 IFAFREE(rt->rt_ifa); 138 rt->rt_ifa = &ia->ia_ifa; 139 rt->rt_ifp = rt->rt_ifa->ifa_ifp; 140 break; 141 } 142 } 143 } 144 } 145 146 if (rt->rt_rmx.rmx_mtu != 0 && !(rt->rt_rmx.rmx_locks & RTV_MTU) && 147 rt->rt_ifp != NULL) 148 rt->rt_rmx.rmx_mtu = rt->rt_ifp->if_mtu; 149 150 ret = rn_addroute(key, mask, head, treenodes); 151 if (ret == NULL && (rt->rt_flags & RTF_HOST)) { 152 struct rtentry *oldrt; 153 154 /* 155 * We are trying to add a host route, but can't. 156 * Find out if it is because of an ARP entry and 157 * delete it if so. 158 */ 159 oldrt = rtpurelookup((struct sockaddr *)sin); 160 if (oldrt != NULL) { 161 --oldrt->rt_refcnt; 162 if ((oldrt->rt_flags & RTF_LLINFO) && 163 (oldrt->rt_flags & RTF_HOST) && 164 oldrt->rt_gateway && 165 oldrt->rt_gateway->sa_family == AF_LINK) { 166 rtrequest(RTM_DELETE, rt_key(oldrt), 167 oldrt->rt_gateway, rt_mask(oldrt), 168 oldrt->rt_flags, NULL); 169 ret = rn_addroute(key, mask, head, treenodes); 170 } 171 } 172 } 173 174 /* 175 * If the new route has been created successfully, and it is 176 * not a multicast/broadcast or cloned route, then we will 177 * have to flush the ipflow. Otherwise, we may end up using 178 * the wrong route. 179 */ 180 if (ret != NULL && 181 (rt->rt_flags & 182 (RTF_MULTICAST | RTF_BROADCAST | RTF_WASCLONED)) == 0) { 183 ipflow_flush_oncpu(); 184 } 185 return ret; 186 } 187 188 /* 189 * This code is the inverse of in_closeroute: on first reference, if we 190 * were managing the route, stop doing so and set the expiration timer 191 * back off again. 192 */ 193 static struct radix_node * 194 in_matchroute(char *key, struct radix_node_head *head) 195 { 196 struct radix_node *rn = rn_match(key, head); 197 struct rtentry *rt = (struct rtentry *)rn; 198 199 if (rt != NULL && rt->rt_refcnt == 0) { /* this is first reference */ 200 if (rt->rt_flags & RTPRF_EXPIRING) { 201 rt->rt_flags &= ~RTPRF_EXPIRING; 202 rt->rt_rmx.rmx_expire = 0; 203 } 204 } 205 return rn; 206 } 207 208 static int rtq_reallyold = 60*60; /* one hour is ``really old'' */ 209 SYSCTL_INT(_net_inet_ip, IPCTL_RTEXPIRE, rtexpire, CTLFLAG_RW, 210 &rtq_reallyold , 0, 211 "Default expiration time on cloned routes"); 212 213 static int rtq_minreallyold = 10; /* never automatically crank down to less */ 214 SYSCTL_INT(_net_inet_ip, IPCTL_RTMINEXPIRE, rtminexpire, CTLFLAG_RW, 215 &rtq_minreallyold , 0, 216 "Minimum time to attempt to hold onto cloned routes"); 217 218 static int rtq_toomany = 128; /* 128 cached routes is ``too many'' */ 219 SYSCTL_INT(_net_inet_ip, IPCTL_RTMAXCACHE, rtmaxcache, CTLFLAG_RW, 220 &rtq_toomany , 0, "Upper limit on cloned routes"); 221 222 /* 223 * On last reference drop, mark the route as belong to us so that it can be 224 * timed out. 225 */ 226 static void 227 in_closeroute(struct radix_node *rn, struct radix_node_head *head) 228 { 229 struct rtentry *rt = (struct rtentry *)rn; 230 231 if (!(rt->rt_flags & RTF_UP)) 232 return; /* prophylactic measures */ 233 234 if ((rt->rt_flags & (RTF_LLINFO | RTF_HOST)) != RTF_HOST) 235 return; 236 237 if ((rt->rt_flags & (RTF_WASCLONED | RTPRF_EXPIRING)) != RTF_WASCLONED) 238 return; 239 240 /* 241 * As requested by David Greenman: 242 * If rtq_reallyold is 0, just delete the route without 243 * waiting for a timeout cycle to kill it. 244 */ 245 if (rtq_reallyold != 0) { 246 rt->rt_flags |= RTPRF_EXPIRING; 247 rt->rt_rmx.rmx_expire = time_uptime + rtq_reallyold; 248 } else { 249 /* 250 * Remove route from the radix tree, but defer deallocation 251 * until we return to rtfree(). 252 */ 253 rtrequest(RTM_DELETE, rt_key(rt), rt->rt_gateway, rt_mask(rt), 254 rt->rt_flags, &rt); 255 } 256 } 257 258 struct rtqk_arg { 259 struct radix_node_head *rnh; 260 int draining; 261 int killed; 262 int found; 263 int updating; 264 time_t nextstop; 265 }; 266 267 /* 268 * Get rid of old routes. When draining, this deletes everything, even when 269 * the timeout is not expired yet. When updating, this makes sure that 270 * nothing has a timeout longer than the current value of rtq_reallyold. 271 */ 272 static int 273 in_rtqkill(struct radix_node *rn, void *rock) 274 { 275 struct rtqk_arg *ap = rock; 276 struct rtentry *rt = (struct rtentry *)rn; 277 int err; 278 279 if (rt->rt_flags & RTPRF_EXPIRING) { 280 ap->found++; 281 if (ap->draining || rt->rt_rmx.rmx_expire <= time_uptime) { 282 if (rt->rt_refcnt > 0) 283 panic("rtqkill route really not free"); 284 285 err = rtrequest(RTM_DELETE, rt_key(rt), rt->rt_gateway, 286 rt_mask(rt), rt->rt_flags, NULL); 287 if (err) 288 log(LOG_WARNING, "in_rtqkill: error %d\n", err); 289 else 290 ap->killed++; 291 } else { 292 if (ap->updating && 293 (int)(rt->rt_rmx.rmx_expire - time_uptime) > 294 rtq_reallyold) { 295 rt->rt_rmx.rmx_expire = time_uptime + 296 rtq_reallyold; 297 } 298 ap->nextstop = lmin(ap->nextstop, 299 rt->rt_rmx.rmx_expire); 300 } 301 } 302 303 return 0; 304 } 305 306 #define RTQ_TIMEOUT 60*10 /* run no less than once every ten minutes */ 307 static int rtq_timeout = RTQ_TIMEOUT; 308 309 /* 310 * NOTE: 311 * 'last_adjusted_timeout' and 'rtq_reallyold' are _not_ read-only, and 312 * could be changed by all CPUs. However, they are changed at so low 313 * frequency that we could ignore the cache trashing issue and take them 314 * as read-mostly. 315 */ 316 static void 317 in_rtqtimo_dispatch(netmsg_t nmsg) 318 { 319 struct rtqk_arg arg; 320 struct timeval atv; 321 static time_t last_adjusted_timeout = 0; 322 struct in_rtqtimo_ctx *ctx = &in_rtqtimo_context[mycpuid]; 323 struct radix_node_head *rnh = ctx->timo_rnh; 324 325 /* Reply ASAP */ 326 crit_enter(); 327 lwkt_replymsg(&nmsg->lmsg, 0); 328 crit_exit(); 329 330 arg.found = arg.killed = 0; 331 arg.rnh = rnh; 332 arg.nextstop = time_uptime + rtq_timeout; 333 arg.draining = arg.updating = 0; 334 rnh->rnh_walktree(rnh, in_rtqkill, &arg); 335 336 /* 337 * Attempt to be somewhat dynamic about this: 338 * If there are ``too many'' routes sitting around taking up space, 339 * then crank down the timeout, and see if we can't make some more 340 * go away. However, we make sure that we will never adjust more 341 * than once in rtq_timeout seconds, to keep from cranking down too 342 * hard. 343 */ 344 if ((arg.found - arg.killed > rtq_toomany) && 345 (int)(time_uptime - last_adjusted_timeout) >= rtq_timeout && 346 rtq_reallyold > rtq_minreallyold) { 347 rtq_reallyold = 2*rtq_reallyold / 3; 348 if (rtq_reallyold < rtq_minreallyold) { 349 rtq_reallyold = rtq_minreallyold; 350 } 351 352 last_adjusted_timeout = time_uptime; 353 #ifdef DIAGNOSTIC 354 log(LOG_DEBUG, "in_rtqtimo: adjusted rtq_reallyold to %d\n", 355 rtq_reallyold); 356 #endif 357 arg.found = arg.killed = 0; 358 arg.updating = 1; 359 rnh->rnh_walktree(rnh, in_rtqkill, &arg); 360 } 361 362 atv.tv_usec = 0; 363 atv.tv_sec = arg.nextstop - time_uptime; 364 if ((int)atv.tv_sec < 1) { /* time shift safety */ 365 atv.tv_sec = 1; 366 arg.nextstop = time_uptime + atv.tv_sec; 367 } 368 if ((int)atv.tv_sec > rtq_timeout) { /* time shift safety */ 369 atv.tv_sec = rtq_timeout; 370 arg.nextstop = time_uptime + atv.tv_sec; 371 } 372 callout_reset(&ctx->timo_ch, tvtohz_high(&atv), in_rtqtimo, NULL); 373 } 374 375 static void 376 in_rtqtimo(void *arg __unused) 377 { 378 int cpuid = mycpuid; 379 struct lwkt_msg *lmsg = &in_rtqtimo_context[cpuid].timo_nmsg.lmsg; 380 381 crit_enter(); 382 if (lmsg->ms_flags & MSGF_DONE) 383 lwkt_sendmsg_oncpu(netisr_cpuport(cpuid), lmsg); 384 crit_exit(); 385 } 386 387 void 388 in_rtqdrain(void) 389 { 390 struct radix_node_head *rnh = rt_tables[mycpuid][AF_INET]; 391 struct rtqk_arg arg; 392 393 arg.found = arg.killed = 0; 394 arg.rnh = rnh; 395 arg.nextstop = 0; 396 arg.draining = 1; 397 arg.updating = 0; 398 crit_enter(); 399 rnh->rnh_walktree(rnh, in_rtqkill, &arg); 400 crit_exit(); 401 } 402 403 /* 404 * Initialize our routing tree. 405 */ 406 int 407 in_inithead(void **head, int off) 408 { 409 struct radix_node_head *rnh; 410 struct in_rtqtimo_ctx *ctx; 411 int cpuid = mycpuid; 412 413 if (!rn_inithead(head, rn_cpumaskhead(cpuid), off)) 414 return 0; 415 416 if (head != (void **)&rt_tables[cpuid][AF_INET]) /* BOGUS! */ 417 return 1; /* only do this for the real routing table */ 418 419 rnh = *head; 420 rnh->rnh_addaddr = in_addroute; 421 rnh->rnh_matchaddr = in_matchroute; 422 rnh->rnh_close = in_closeroute; 423 424 ctx = &in_rtqtimo_context[cpuid]; 425 ctx->timo_rnh = rnh; 426 callout_init_mp(&ctx->timo_ch); 427 netmsg_init(&ctx->timo_nmsg, NULL, &netisr_adone_rport, 0, 428 in_rtqtimo_dispatch); 429 430 in_rtqtimo(NULL); /* kick off timeout first time */ 431 return 1; 432 } 433 434 /* 435 * This zaps old routes when the interface goes down or interface 436 * address is deleted. In the latter case, it deletes static routes 437 * that point to this address. If we don't do this, we may end up 438 * using the old address in the future. The ones we always want to 439 * get rid of are things like ARP entries, since the user might down 440 * the interface, walk over to a completely different network, and 441 * plug back in. 442 * 443 * in_ifadown() is typically called when an interface is being brought 444 * down. We must iterate through all per-cpu route tables and clean 445 * them up. 446 */ 447 struct in_ifadown_arg { 448 struct radix_node_head *rnh; 449 struct ifaddr *ifa; 450 int del; 451 }; 452 453 static int 454 in_ifadownkill(struct radix_node *rn, void *xap) 455 { 456 struct in_ifadown_arg *ap = xap; 457 struct rtentry *rt = (struct rtentry *)rn; 458 int err; 459 460 if (rt->rt_ifa == ap->ifa && 461 (ap->del || !(rt->rt_flags & RTF_STATIC))) { 462 /* 463 * We need to disable the automatic prune that happens 464 * in this case in rtrequest() because it will blow 465 * away the pointers that rn_walktree() needs in order 466 * continue our descent. We will end up deleting all 467 * the routes that rtrequest() would have in any case, 468 * so that behavior is not needed there. 469 */ 470 rt->rt_flags &= ~(RTF_CLONING | RTF_PRCLONING); 471 err = rtrequest(RTM_DELETE, rt_key(rt), rt->rt_gateway, 472 rt_mask(rt), rt->rt_flags, NULL); 473 if (err) 474 log(LOG_WARNING, "in_ifadownkill: error %d\n", err); 475 } 476 return 0; 477 } 478 479 struct netmsg_ifadown { 480 struct netmsg_base base; 481 struct ifaddr *ifa; 482 int del; 483 }; 484 485 static void 486 in_ifadown_dispatch(netmsg_t msg) 487 { 488 struct netmsg_ifadown *rmsg = (void *)msg; 489 struct radix_node_head *rnh; 490 struct ifaddr *ifa = rmsg->ifa; 491 struct in_ifadown_arg arg; 492 int nextcpu, cpu; 493 494 cpu = mycpuid; 495 496 arg.rnh = rnh = rt_tables[cpu][AF_INET]; 497 arg.ifa = ifa; 498 arg.del = rmsg->del; 499 rnh->rnh_walktree(rnh, in_ifadownkill, &arg); 500 ifa->ifa_flags &= ~IFA_ROUTE; 501 502 nextcpu = cpu + 1; 503 if (nextcpu < ncpus) 504 lwkt_forwardmsg(netisr_cpuport(nextcpu), &rmsg->base.lmsg); 505 else 506 lwkt_replymsg(&rmsg->base.lmsg, 0); 507 } 508 509 int 510 in_ifadown_force(struct ifaddr *ifa, int delete) 511 { 512 struct netmsg_ifadown msg; 513 514 if (ifa->ifa_addr->sa_family != AF_INET) 515 return 1; 516 517 /* 518 * XXX individual requests are not independantly chained, 519 * which means that the per-cpu route tables will not be 520 * consistent in the middle of the operation. If routes 521 * related to the interface are manipulated while we are 522 * doing this the inconsistancy could trigger a panic. 523 */ 524 netmsg_init(&msg.base, NULL, &curthread->td_msgport, MSGF_PRIORITY, 525 in_ifadown_dispatch); 526 msg.ifa = ifa; 527 msg.del = delete; 528 rt_domsg_global(&msg.base); 529 530 return 0; 531 } 532 533 int 534 in_ifadown(struct ifaddr *ifa, int delete) 535 { 536 #ifdef CARP 537 if (ifa->ifa_ifp->if_type == IFT_CARP) 538 return 0; 539 #endif 540 return in_ifadown_force(ifa, delete); 541 } 542