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 <sys/param.h> 47 #include <sys/systm.h> 48 #include <sys/kernel.h> 49 #include <sys/sysctl.h> 50 #include <sys/socket.h> 51 #include <sys/mbuf.h> 52 #include <sys/syslog.h> 53 #include <sys/globaldata.h> 54 #include <sys/thread2.h> 55 56 #include <net/if.h> 57 #include <net/route.h> 58 #include <netinet/in.h> 59 #include <netinet/in_var.h> 60 #include <netinet/ip_var.h> 61 62 #define RTPRF_EXPIRING RTF_PROTO3 /* set on routes we manage */ 63 64 static struct callout in_rtqtimo_ch[MAXCPU]; 65 66 /* 67 * Do what we need to do when inserting a route. 68 */ 69 static struct radix_node * 70 in_addroute(char *key, char *mask, struct radix_node_head *head, 71 struct radix_node *treenodes) 72 { 73 struct rtentry *rt = (struct rtentry *)treenodes; 74 struct sockaddr_in *sin = (struct sockaddr_in *)rt_key(rt); 75 struct radix_node *ret; 76 77 /* 78 * For IP, all unicast non-host routes are automatically cloning. 79 */ 80 if (IN_MULTICAST(ntohl(sin->sin_addr.s_addr))) 81 rt->rt_flags |= RTF_MULTICAST; 82 83 if (!(rt->rt_flags & (RTF_HOST | RTF_CLONING | RTF_MULTICAST))) 84 rt->rt_flags |= RTF_PRCLONING; 85 86 /* 87 * A little bit of help for both IP output and input: 88 * For host routes, we make sure that RTF_BROADCAST 89 * is set for anything that looks like a broadcast address. 90 * This way, we can avoid an expensive call to in_broadcast() 91 * in ip_output() most of the time (because the route passed 92 * to ip_output() is almost always a host route). 93 * 94 * We also do the same for local addresses, with the thought 95 * that this might one day be used to speed up ip_input(). 96 * 97 * We also 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). (This 100 * is done above.) 101 */ 102 if (rt->rt_flags & RTF_HOST) { 103 if (in_broadcast(sin->sin_addr, rt->rt_ifp)) { 104 rt->rt_flags |= RTF_BROADCAST; 105 } else { 106 if (satosin(rt->rt_ifa->ifa_addr)->sin_addr.s_addr 107 == sin->sin_addr.s_addr) 108 rt->rt_flags |= RTF_LOCAL; 109 } 110 } 111 112 if (rt->rt_rmx.rmx_mtu != 0 && !(rt->rt_rmx.rmx_locks & RTV_MTU) && 113 rt->rt_ifp != NULL) 114 rt->rt_rmx.rmx_mtu = rt->rt_ifp->if_mtu; 115 116 ret = rn_addroute(key, mask, head, treenodes); 117 if (ret == NULL && rt->rt_flags & RTF_HOST) { 118 struct rtentry *oldrt; 119 120 /* 121 * We are trying to add a host route, but can't. 122 * Find out if it is because of an ARP entry and 123 * delete it if so. 124 */ 125 oldrt = rtpurelookup((struct sockaddr *)sin); 126 if (oldrt != NULL) { 127 --oldrt->rt_refcnt; 128 if (oldrt->rt_flags & RTF_LLINFO && 129 oldrt->rt_flags & RTF_HOST && 130 oldrt->rt_gateway && 131 oldrt->rt_gateway->sa_family == AF_LINK) { 132 rtrequest(RTM_DELETE, rt_key(oldrt), 133 oldrt->rt_gateway, rt_mask(oldrt), 134 oldrt->rt_flags, NULL); 135 ret = rn_addroute(key, mask, head, treenodes); 136 } 137 } 138 } 139 140 /* 141 * If the new route created successfully, and we are forwarding, 142 * and there is a cached route, free it. Otherwise, we may end 143 * up using the wrong route. 144 */ 145 if (ret != NULL && ipforwarding && 146 ipforward_rt[mycpuid].ro_rt != NULL) { 147 RTFREE(ipforward_rt[mycpuid].ro_rt); 148 ipforward_rt[mycpuid].ro_rt = NULL; 149 } 150 151 return ret; 152 } 153 154 /* 155 * This code is the inverse of in_closeroute: on first reference, if we 156 * were managing the route, stop doing so and set the expiration timer 157 * back off again. 158 */ 159 static struct radix_node * 160 in_matchroute(char *key, struct radix_node_head *head) 161 { 162 struct radix_node *rn = rn_match(key, head); 163 struct rtentry *rt = (struct rtentry *)rn; 164 165 if (rt != NULL && rt->rt_refcnt == 0) { /* this is first reference */ 166 if (rt->rt_flags & RTPRF_EXPIRING) { 167 rt->rt_flags &= ~RTPRF_EXPIRING; 168 rt->rt_rmx.rmx_expire = 0; 169 } 170 } 171 return rn; 172 } 173 174 static int rtq_reallyold = 60*60; /* one hour is ``really old'' */ 175 SYSCTL_INT(_net_inet_ip, IPCTL_RTEXPIRE, rtexpire, CTLFLAG_RW, 176 &rtq_reallyold , 0, 177 "Default expiration time on cloned routes"); 178 179 static int rtq_minreallyold = 10; /* never automatically crank down to less */ 180 SYSCTL_INT(_net_inet_ip, IPCTL_RTMINEXPIRE, rtminexpire, CTLFLAG_RW, 181 &rtq_minreallyold , 0, 182 "Minimum time to attempt to hold onto cloned routes"); 183 184 static int rtq_toomany = 128; /* 128 cached routes is ``too many'' */ 185 SYSCTL_INT(_net_inet_ip, IPCTL_RTMAXCACHE, rtmaxcache, CTLFLAG_RW, 186 &rtq_toomany , 0, "Upper limit on cloned routes"); 187 188 /* 189 * On last reference drop, mark the route as belong to us so that it can be 190 * timed out. 191 */ 192 static void 193 in_closeroute(struct radix_node *rn, struct radix_node_head *head) 194 { 195 struct rtentry *rt = (struct rtentry *)rn; 196 197 if (!(rt->rt_flags & RTF_UP)) 198 return; /* prophylactic measures */ 199 200 if ((rt->rt_flags & (RTF_LLINFO | RTF_HOST)) != RTF_HOST) 201 return; 202 203 if ((rt->rt_flags & (RTF_WASCLONED | RTPRF_EXPIRING)) != RTF_WASCLONED) 204 return; 205 206 /* 207 * As requested by David Greenman: 208 * If rtq_reallyold is 0, just delete the route without 209 * waiting for a timeout cycle to kill it. 210 */ 211 if (rtq_reallyold != 0) { 212 rt->rt_flags |= RTPRF_EXPIRING; 213 rt->rt_rmx.rmx_expire = time_second + rtq_reallyold; 214 } else { 215 /* 216 * Remove route from the radix tree, but defer deallocation 217 * until we return to rtfree(). 218 */ 219 rtrequest(RTM_DELETE, rt_key(rt), rt->rt_gateway, rt_mask(rt), 220 rt->rt_flags, &rt); 221 } 222 } 223 224 struct rtqk_arg { 225 struct radix_node_head *rnh; 226 int draining; 227 int killed; 228 int found; 229 int updating; 230 time_t nextstop; 231 }; 232 233 /* 234 * Get rid of old routes. When draining, this deletes everything, even when 235 * the timeout is not expired yet. When updating, this makes sure that 236 * nothing has a timeout longer than the current value of rtq_reallyold. 237 */ 238 static int 239 in_rtqkill(struct radix_node *rn, void *rock) 240 { 241 struct rtqk_arg *ap = rock; 242 struct rtentry *rt = (struct rtentry *)rn; 243 int err; 244 245 if (rt->rt_flags & RTPRF_EXPIRING) { 246 ap->found++; 247 if (ap->draining || rt->rt_rmx.rmx_expire <= time_second) { 248 if (rt->rt_refcnt > 0) 249 panic("rtqkill route really not free"); 250 251 err = rtrequest(RTM_DELETE, rt_key(rt), rt->rt_gateway, 252 rt_mask(rt), rt->rt_flags, NULL); 253 if (err) 254 log(LOG_WARNING, "in_rtqkill: error %d\n", err); 255 else 256 ap->killed++; 257 } else { 258 if (ap->updating && 259 (rt->rt_rmx.rmx_expire - time_second > 260 rtq_reallyold)) { 261 rt->rt_rmx.rmx_expire = time_second + 262 rtq_reallyold; 263 } 264 ap->nextstop = lmin(ap->nextstop, 265 rt->rt_rmx.rmx_expire); 266 } 267 } 268 269 return 0; 270 } 271 272 #define RTQ_TIMEOUT 60*10 /* run no less than once every ten minutes */ 273 static int rtq_timeout = RTQ_TIMEOUT; 274 275 static void 276 in_rtqtimo(void *rock) 277 { 278 struct radix_node_head *rnh = rock; 279 struct rtqk_arg arg; 280 struct timeval atv; 281 static time_t last_adjusted_timeout = 0; 282 283 arg.found = arg.killed = 0; 284 arg.rnh = rnh; 285 arg.nextstop = time_second + rtq_timeout; 286 arg.draining = arg.updating = 0; 287 crit_enter(); 288 rnh->rnh_walktree(rnh, in_rtqkill, &arg); 289 crit_exit(); 290 291 /* 292 * Attempt to be somewhat dynamic about this: 293 * If there are ``too many'' routes sitting around taking up space, 294 * then crank down the timeout, and see if we can't make some more 295 * go away. However, we make sure that we will never adjust more 296 * than once in rtq_timeout seconds, to keep from cranking down too 297 * hard. 298 */ 299 if ((arg.found - arg.killed > rtq_toomany) && 300 (time_second - last_adjusted_timeout >= rtq_timeout) && 301 rtq_reallyold > rtq_minreallyold) { 302 rtq_reallyold = 2*rtq_reallyold / 3; 303 if (rtq_reallyold < rtq_minreallyold) { 304 rtq_reallyold = rtq_minreallyold; 305 } 306 307 last_adjusted_timeout = time_second; 308 #ifdef DIAGNOSTIC 309 log(LOG_DEBUG, "in_rtqtimo: adjusted rtq_reallyold to %d\n", 310 rtq_reallyold); 311 #endif 312 arg.found = arg.killed = 0; 313 arg.updating = 1; 314 crit_enter(); 315 rnh->rnh_walktree(rnh, in_rtqkill, &arg); 316 crit_exit(); 317 } 318 319 atv.tv_usec = 0; 320 atv.tv_sec = arg.nextstop - time_second; 321 callout_reset(&in_rtqtimo_ch[mycpuid], tvtohz_high(&atv), in_rtqtimo, 322 rock); 323 } 324 325 void 326 in_rtqdrain(void) 327 { 328 struct radix_node_head *rnh = rt_tables[mycpuid][AF_INET]; 329 struct rtqk_arg arg; 330 331 arg.found = arg.killed = 0; 332 arg.rnh = rnh; 333 arg.nextstop = 0; 334 arg.draining = 1; 335 arg.updating = 0; 336 crit_enter(); 337 rnh->rnh_walktree(rnh, in_rtqkill, &arg); 338 crit_exit(); 339 } 340 341 /* 342 * Initialize our routing tree. 343 */ 344 int 345 in_inithead(void **head, int off) 346 { 347 struct radix_node_head *rnh; 348 349 if (!rn_inithead(head, off)) 350 return 0; 351 352 if (head != (void **)&rt_tables[mycpuid][AF_INET]) /* BOGUS! */ 353 return 1; /* only do this for the real routing table */ 354 355 rnh = *head; 356 rnh->rnh_addaddr = in_addroute; 357 rnh->rnh_matchaddr = in_matchroute; 358 rnh->rnh_close = in_closeroute; 359 callout_init(&in_rtqtimo_ch[mycpuid]); 360 in_rtqtimo(rnh); /* kick off timeout first time */ 361 return 1; 362 } 363 364 /* 365 * This zaps old routes when the interface goes down or interface 366 * address is deleted. In the latter case, it deletes static routes 367 * that point to this address. If we don't do this, we may end up 368 * using the old address in the future. The ones we always want to 369 * get rid of are things like ARP entries, since the user might down 370 * the interface, walk over to a completely different network, and 371 * plug back in. 372 * 373 * in_ifadown() is typically called when an interface is being brought 374 * down. We must iterate through all per-cpu route tables and clean 375 * them up. 376 */ 377 struct in_ifadown_arg { 378 struct radix_node_head *rnh; 379 struct ifaddr *ifa; 380 int del; 381 }; 382 383 static int 384 in_ifadownkill(struct radix_node *rn, void *xap) 385 { 386 struct in_ifadown_arg *ap = xap; 387 struct rtentry *rt = (struct rtentry *)rn; 388 int err; 389 390 if (rt->rt_ifa == ap->ifa && 391 (ap->del || !(rt->rt_flags & RTF_STATIC))) { 392 /* 393 * We need to disable the automatic prune that happens 394 * in this case in rtrequest() because it will blow 395 * away the pointers that rn_walktree() needs in order 396 * continue our descent. We will end up deleting all 397 * the routes that rtrequest() would have in any case, 398 * so that behavior is not needed there. 399 */ 400 rt->rt_flags &= ~(RTF_CLONING | RTF_PRCLONING); 401 err = rtrequest(RTM_DELETE, rt_key(rt), rt->rt_gateway, 402 rt_mask(rt), rt->rt_flags, NULL); 403 if (err) 404 log(LOG_WARNING, "in_ifadownkill: error %d\n", err); 405 } 406 return 0; 407 } 408 409 int 410 in_ifadown(struct ifaddr *ifa, int delete) 411 { 412 struct in_ifadown_arg arg; 413 struct radix_node_head *rnh; 414 int origcpu; 415 int cpu; 416 417 if (ifa->ifa_addr->sa_family != AF_INET) 418 return 1; 419 420 /* 421 * XXX individual requests are not independantly chained, 422 * which means that the per-cpu route tables will not be 423 * consistent in the middle of the operation. If routes 424 * related to the interface are manipulated while we are 425 * doing this the inconsistancy could trigger a panic. 426 */ 427 origcpu = mycpuid; 428 for (cpu = 0; cpu < ncpus2; cpu++) { 429 lwkt_migratecpu(cpu); 430 431 arg.rnh = rnh = rt_tables[cpu][AF_INET]; 432 arg.ifa = ifa; 433 arg.del = delete; 434 rnh->rnh_walktree(rnh, in_ifadownkill, &arg); 435 ifa->ifa_flags &= ~IFA_ROUTE; 436 } 437 lwkt_migratecpu(origcpu); 438 return 0; 439 } 440 441