xref: /dragonfly/sys/netinet/in_rmx.c (revision d5f516c3) 
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 1.5 2004/09/16 23:14:29 joerg 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 
54 #include <net/if.h>
55 #include <net/route.h>
56 #include <netinet/in.h>
57 #include <netinet/in_var.h>
58 #include <netinet/ip_var.h>
59 
60 static struct callout	in_rtqtimo_ch;
61 
62 extern int	in_inithead (void **head, int off);
63 
64 #define RTPRF_OURS		RTF_PROTO3	/* set on routes we manage */
65 
66 /*
67  * Do what we need to do when inserting a route.
68  */
69 static struct radix_node *
70 in_addroute(void *v_arg, void *n_arg, 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 	/*
88 	 * A little bit of help for both IP output and input:
89 	 *   For host routes, we make sure that RTF_BROADCAST
90 	 *   is set for anything that looks like a broadcast address.
91 	 *   This way, we can avoid an expensive call to in_broadcast()
92 	 *   in ip_output() most of the time (because the route passed
93 	 *   to ip_output() is almost always a host route).
94 	 *
95 	 *   We also do the same for local addresses, with the thought
96 	 *   that this might one day be used to speed up ip_input().
97 	 *
98 	 * We also mark routes to multicast addresses as such, because
99 	 * it's easy to do and might be useful (but this is much more
100 	 * dubious since it's so easy to inspect the address).  (This
101 	 * is done above.)
102 	 */
103 	if (rt->rt_flags & RTF_HOST) {
104 		if (in_broadcast(sin->sin_addr, rt->rt_ifp)) {
105 			rt->rt_flags |= RTF_BROADCAST;
106 		} else {
107 			if (satosin(rt->rt_ifa->ifa_addr)->sin_addr.s_addr
108 			    == sin->sin_addr.s_addr)
109 				rt->rt_flags |= RTF_LOCAL;
110 		}
111 	}
112 
113 	if (!rt->rt_rmx.rmx_mtu && !(rt->rt_rmx.rmx_locks & RTV_MTU)
114 	    && rt->rt_ifp)
115 		rt->rt_rmx.rmx_mtu = rt->rt_ifp->if_mtu;
116 
117 	ret = rn_addroute(v_arg, n_arg, head, treenodes);
118 	if (ret == NULL && rt->rt_flags & RTF_HOST) {
119 		struct rtentry *rt2;
120 		/*
121 		 * We are trying to add a host route, but can't.
122 		 * Find out if it is because of an
123 		 * ARP entry and delete it if so.
124 		 */
125 		rt2 = rtalloc1((struct sockaddr *)sin, 0,
126 				RTF_CLONING | RTF_PRCLONING);
127 		if (rt2) {
128 			if (rt2->rt_flags & RTF_LLINFO &&
129 				rt2->rt_flags & RTF_HOST &&
130 				rt2->rt_gateway &&
131 				rt2->rt_gateway->sa_family == AF_LINK) {
132 				rtrequest(RTM_DELETE,
133 					  (struct sockaddr *)rt_key(rt2),
134 					  rt2->rt_gateway,
135 					  rt_mask(rt2), rt2->rt_flags, 0);
136 				ret = rn_addroute(v_arg, n_arg, head,
137 					treenodes);
138 			}
139 			RTFREE(rt2);
140 		}
141 	}
142 
143 	/*
144 	 * If the new route created successfully, and we are forwarding,
145 	 * and there is a cached route, free it.  Otherwise, we may end
146 	 * up using the wrong route.
147 	 */
148 	if (ret != NULL && ipforwarding && ipforward_rt.ro_rt) {
149 		RTFREE(ipforward_rt.ro_rt);
150 		ipforward_rt.ro_rt = 0;
151 	}
152 
153 	return ret;
154 }
155 
156 /*
157  * This code is the inverse of in_clsroute: on first reference, if we
158  * were managing the route, stop doing so and set the expiration timer
159  * back off again.
160  */
161 static struct radix_node *
162 in_matroute(void *v_arg, struct radix_node_head *head)
163 {
164 	struct radix_node *rn = rn_match(v_arg, head);
165 	struct rtentry *rt = (struct rtentry *)rn;
166 
167 	if(rt && rt->rt_refcnt == 0) { /* this is first reference */
168 		if(rt->rt_flags & RTPRF_OURS) {
169 			rt->rt_flags &= ~RTPRF_OURS;
170 			rt->rt_rmx.rmx_expire = 0;
171 		}
172 	}
173 	return rn;
174 }
175 
176 static int rtq_reallyold = 60*60;
177 	/* one hour is ``really old'' */
178 SYSCTL_INT(_net_inet_ip, IPCTL_RTEXPIRE, rtexpire, CTLFLAG_RW,
179     &rtq_reallyold , 0,
180     "Default expiration time on dynamically learned routes");
181 
182 static int rtq_minreallyold = 10;
183 	/* never automatically crank down to less */
184 SYSCTL_INT(_net_inet_ip, IPCTL_RTMINEXPIRE, rtminexpire, CTLFLAG_RW,
185     &rtq_minreallyold , 0,
186     "Minimum time to attempt to hold onto dynamically learned routes");
187 
188 static int rtq_toomany = 128;
189 	/* 128 cached routes is ``too many'' */
190 SYSCTL_INT(_net_inet_ip, IPCTL_RTMAXCACHE, rtmaxcache, CTLFLAG_RW,
191     &rtq_toomany , 0, "Upper limit on dynamically learned routes");
192 
193 /*
194  * On last reference drop, mark the route as belong to us so that it can be
195  * timed out.
196  */
197 static void
198 in_clsroute(struct radix_node *rn, struct radix_node_head *head)
199 {
200 	struct rtentry *rt = (struct rtentry *)rn;
201 
202 	if(!(rt->rt_flags & RTF_UP))
203 		return;		/* prophylactic measures */
204 
205 	if((rt->rt_flags & (RTF_LLINFO | RTF_HOST)) != RTF_HOST)
206 		return;
207 
208 	if((rt->rt_flags & (RTF_WASCLONED | RTPRF_OURS))
209 	   != RTF_WASCLONED)
210 		return;
211 
212 	/*
213 	 * As requested by David Greenman:
214 	 * If rtq_reallyold is 0, just delete the route without
215 	 * waiting for a timeout cycle to kill it.
216 	 */
217 	if(rtq_reallyold != 0) {
218 		rt->rt_flags |= RTPRF_OURS;
219 		rt->rt_rmx.rmx_expire = time_second + rtq_reallyold;
220 	} else {
221 		rtrequest(RTM_DELETE,
222 			  (struct sockaddr *)rt_key(rt),
223 			  rt->rt_gateway, rt_mask(rt),
224 			  rt->rt_flags, 0);
225 	}
226 }
227 
228 struct rtqk_arg {
229 	struct radix_node_head *rnh;
230 	int draining;
231 	int killed;
232 	int found;
233 	int updating;
234 	time_t nextstop;
235 };
236 
237 /*
238  * Get rid of old routes.  When draining, this deletes everything, even when
239  * the timeout is not expired yet.  When updating, this makes sure that
240  * nothing has a timeout longer than the current value of rtq_reallyold.
241  */
242 static int
243 in_rtqkill(struct radix_node *rn, void *rock)
244 {
245 	struct rtqk_arg *ap = rock;
246 	struct rtentry *rt = (struct rtentry *)rn;
247 	int err;
248 
249 	if(rt->rt_flags & RTPRF_OURS) {
250 		ap->found++;
251 
252 		if(ap->draining || rt->rt_rmx.rmx_expire <= time_second) {
253 			if(rt->rt_refcnt > 0)
254 				panic("rtqkill route really not free");
255 
256 			err = rtrequest(RTM_DELETE,
257 					(struct sockaddr *)rt_key(rt),
258 					rt->rt_gateway, rt_mask(rt),
259 					rt->rt_flags, 0);
260 			if(err) {
261 				log(LOG_WARNING, "in_rtqkill: error %d\n", err);
262 			} else {
263 				ap->killed++;
264 			}
265 		} else {
266 			if(ap->updating
267 			   && (rt->rt_rmx.rmx_expire - time_second
268 			       > rtq_reallyold)) {
269 				rt->rt_rmx.rmx_expire = time_second
270 					+ rtq_reallyold;
271 			}
272 			ap->nextstop = lmin(ap->nextstop,
273 					    rt->rt_rmx.rmx_expire);
274 		}
275 	}
276 
277 	return 0;
278 }
279 
280 #define RTQ_TIMEOUT	60*10	/* run no less than once every ten minutes */
281 static int rtq_timeout = RTQ_TIMEOUT;
282 
283 static void
284 in_rtqtimo(void *rock)
285 {
286 	struct radix_node_head *rnh = rock;
287 	struct rtqk_arg arg;
288 	struct timeval atv;
289 	static time_t last_adjusted_timeout = 0;
290 	int s;
291 
292 	arg.found = arg.killed = 0;
293 	arg.rnh = rnh;
294 	arg.nextstop = time_second + rtq_timeout;
295 	arg.draining = arg.updating = 0;
296 	s = splnet();
297 	rnh->rnh_walktree(rnh, in_rtqkill, &arg);
298 	splx(s);
299 
300 	/*
301 	 * Attempt to be somewhat dynamic about this:
302 	 * If there are ``too many'' routes sitting around taking up space,
303 	 * then crank down the timeout, and see if we can't make some more
304 	 * go away.  However, we make sure that we will never adjust more
305 	 * than once in rtq_timeout seconds, to keep from cranking down too
306 	 * hard.
307 	 */
308 	if((arg.found - arg.killed > rtq_toomany)
309 	   && (time_second - last_adjusted_timeout >= rtq_timeout)
310 	   && rtq_reallyold > rtq_minreallyold) {
311 		rtq_reallyold = 2*rtq_reallyold / 3;
312 		if(rtq_reallyold < rtq_minreallyold) {
313 			rtq_reallyold = rtq_minreallyold;
314 		}
315 
316 		last_adjusted_timeout = time_second;
317 #ifdef DIAGNOSTIC
318 		log(LOG_DEBUG, "in_rtqtimo: adjusted rtq_reallyold to %d\n",
319 		    rtq_reallyold);
320 #endif
321 		arg.found = arg.killed = 0;
322 		arg.updating = 1;
323 		s = splnet();
324 		rnh->rnh_walktree(rnh, in_rtqkill, &arg);
325 		splx(s);
326 	}
327 
328 	atv.tv_usec = 0;
329 	atv.tv_sec = arg.nextstop - time_second;
330 	callout_reset(&in_rtqtimo_ch, tvtohz_high(&atv), in_rtqtimo, rock);
331 }
332 
333 void
334 in_rtqdrain(void)
335 {
336 	struct radix_node_head *rnh = rt_tables[AF_INET];
337 	struct rtqk_arg arg;
338 	int s;
339 	arg.found = arg.killed = 0;
340 	arg.rnh = rnh;
341 	arg.nextstop = 0;
342 	arg.draining = 1;
343 	arg.updating = 0;
344 	s = splnet();
345 	rnh->rnh_walktree(rnh, in_rtqkill, &arg);
346 	splx(s);
347 }
348 
349 /*
350  * Initialize our routing tree.
351  */
352 int
353 in_inithead(void **head, int off)
354 {
355 	struct radix_node_head *rnh;
356 
357 	if(!rn_inithead(head, off))
358 		return 0;
359 
360 	if(head != (void **)&rt_tables[AF_INET]) /* BOGUS! */
361 		return 1;	/* only do this for the real routing table */
362 
363 	rnh = *head;
364 	rnh->rnh_addaddr = in_addroute;
365 	rnh->rnh_matchaddr = in_matroute;
366 	rnh->rnh_close = in_clsroute;
367 	callout_init(&in_rtqtimo_ch);
368 	in_rtqtimo(rnh);	/* kick off timeout first time */
369 	return 1;
370 }
371 
372 
373 /*
374  * This zaps old routes when the interface goes down or interface
375  * address is deleted.  In the latter case, it deletes static routes
376  * that point to this address.  If we don't do this, we may end up
377  * using the old address in the future.  The ones we always want to
378  * get rid of are things like ARP entries, since the user might down
379  * the interface, walk over to a completely different network, and
380  * plug back in.
381  */
382 struct in_ifadown_arg {
383 	struct radix_node_head *rnh;
384 	struct ifaddr *ifa;
385 	int del;
386 };
387 
388 static int
389 in_ifadownkill(struct radix_node *rn, void *xap)
390 {
391 	struct in_ifadown_arg *ap = xap;
392 	struct rtentry *rt = (struct rtentry *)rn;
393 	int err;
394 
395 	if (rt->rt_ifa == ap->ifa &&
396 	    (ap->del || !(rt->rt_flags & RTF_STATIC))) {
397 		/*
398 		 * We need to disable the automatic prune that happens
399 		 * in this case in rtrequest() because it will blow
400 		 * away the pointers that rn_walktree() needs in order
401 		 * continue our descent.  We will end up deleting all
402 		 * the routes that rtrequest() would have in any case,
403 		 * so that behavior is not needed there.
404 		 */
405 		rt->rt_flags &= ~(RTF_CLONING | RTF_PRCLONING);
406 		err = rtrequest(RTM_DELETE, (struct sockaddr *)rt_key(rt),
407 				rt->rt_gateway, rt_mask(rt), rt->rt_flags, 0);
408 		if (err) {
409 			log(LOG_WARNING, "in_ifadownkill: error %d\n", err);
410 		}
411 	}
412 	return 0;
413 }
414 
415 int
416 in_ifadown(struct ifaddr *ifa, int delete)
417 {
418 	struct in_ifadown_arg arg;
419 	struct radix_node_head *rnh;
420 
421 	if (ifa->ifa_addr->sa_family != AF_INET)
422 		return 1;
423 
424 	arg.rnh = rnh = rt_tables[AF_INET];
425 	arg.ifa = ifa;
426 	arg.del = delete;
427 	rnh->rnh_walktree(rnh, in_ifadownkill, &arg);
428 	ifa->ifa_flags &= ~IFA_ROUTE;
429 	return 0;
430 }
431