xref: /freebsd/sys/netinet/in_rmx.c (revision a8445737e740901f5f2c8d24c12ef7fc8b00134e)
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  *	$Id: in_rmx.c,v 1.34 1998/03/30 09:52:46 phk Exp $
30  */
31 
32 /*
33  * This code does two things necessary for the enhanced TCP metrics to
34  * function in a useful manner:
35  *  1) It marks all non-host routes as `cloning', thus ensuring that
36  *     every actual reference to such a route actually gets turned
37  *     into a reference to a host route to the specific destination
38  *     requested.
39  *  2) When such routes lose all their references, it arranges for them
40  *     to be deleted in some random collection of circumstances, so that
41  *     a large quantity of stale routing data is not kept in kernel memory
42  *     indefinitely.  See in_rtqtimo() below for the exact mechanism.
43  */
44 
45 #include <sys/param.h>
46 #include <sys/systm.h>
47 #include <sys/kernel.h>
48 #include <sys/sysctl.h>
49 #include <sys/socket.h>
50 #include <sys/mbuf.h>
51 #include <sys/syslog.h>
52 
53 #include <net/if.h>
54 #include <net/route.h>
55 #include <netinet/in.h>
56 #include <netinet/in_var.h>
57 
58 extern int	in_inithead __P((void **head, int off));
59 
60 #define RTPRF_OURS		RTF_PROTO3	/* set on routes we manage */
61 
62 /*
63  * Do what we need to do when inserting a route.
64  */
65 static struct radix_node *
66 in_addroute(void *v_arg, void *n_arg, struct radix_node_head *head,
67 	    struct radix_node *treenodes)
68 {
69 	struct rtentry *rt = (struct rtentry *)treenodes;
70 	struct sockaddr_in *sin = (struct sockaddr_in *)rt_key(rt);
71 	struct radix_node *ret;
72 
73 	/*
74 	 * For IP, all unicast non-host routes are automatically cloning.
75 	 */
76 	if(IN_MULTICAST(ntohl(sin->sin_addr.s_addr)))
77 		rt->rt_flags |= RTF_MULTICAST;
78 
79 	if(!(rt->rt_flags & (RTF_HOST | RTF_CLONING | RTF_MULTICAST))) {
80 		rt->rt_flags |= RTF_PRCLONING;
81 	}
82 
83 	/*
84 	 * A little bit of help for both IP output and input:
85 	 *   For host routes, we make sure that RTF_BROADCAST
86 	 *   is set for anything that looks like a broadcast address.
87 	 *   This way, we can avoid an expensive call to in_broadcast()
88 	 *   in ip_output() most of the time (because the route passed
89 	 *   to ip_output() is almost always a host route).
90 	 *
91 	 *   We also do the same for local addresses, with the thought
92 	 *   that this might one day be used to speed up ip_input().
93 	 *
94 	 * We also mark routes to multicast addresses as such, because
95 	 * it's easy to do and might be useful (but this is much more
96 	 * dubious since it's so easy to inspect the address).  (This
97 	 * is done above.)
98 	 */
99 	if (rt->rt_flags & RTF_HOST) {
100 		if (in_broadcast(sin->sin_addr, rt->rt_ifp)) {
101 			rt->rt_flags |= RTF_BROADCAST;
102 		} else {
103 #define satosin(sa) ((struct sockaddr_in *)sa)
104 			if (satosin(rt->rt_ifa->ifa_addr)->sin_addr.s_addr
105 			    == sin->sin_addr.s_addr)
106 				rt->rt_flags |= RTF_LOCAL;
107 #undef satosin
108 		}
109 	}
110 
111 	if (!rt->rt_rmx.rmx_mtu && !(rt->rt_rmx.rmx_locks & RTV_MTU)
112 	    && rt->rt_ifp)
113 		rt->rt_rmx.rmx_mtu = rt->rt_ifp->if_mtu;
114 
115 	ret = rn_addroute(v_arg, n_arg, head, treenodes);
116 	if (ret == NULL && rt->rt_flags & RTF_HOST) {
117 		struct rtentry *rt2;
118 		/*
119 		 * We are trying to add a host route, but can't.
120 		 * Find out if it is because of an
121 		 * ARP entry and delete it if so.
122 		 */
123 		rt2 = rtalloc1((struct sockaddr *)sin, 0,
124 				RTF_CLONING | RTF_PRCLONING);
125 		if (rt2) {
126 			if (rt2->rt_flags & RTF_LLINFO &&
127 				rt2->rt_flags & RTF_HOST &&
128 				rt2->rt_gateway &&
129 				rt2->rt_gateway->sa_family == AF_LINK) {
130 				rtrequest(RTM_DELETE,
131 					  (struct sockaddr *)rt_key(rt2),
132 					  rt2->rt_gateway,
133 					  rt_mask(rt2), rt2->rt_flags, 0);
134 				ret = rn_addroute(v_arg, n_arg, head,
135 					treenodes);
136 			}
137 			RTFREE(rt2);
138 		}
139 	}
140 	return ret;
141 }
142 
143 /*
144  * This code is the inverse of in_clsroute: on first reference, if we
145  * were managing the route, stop doing so and set the expiration timer
146  * back off again.
147  */
148 static struct radix_node *
149 in_matroute(void *v_arg, struct radix_node_head *head)
150 {
151 	struct radix_node *rn = rn_match(v_arg, head);
152 	struct rtentry *rt = (struct rtentry *)rn;
153 
154 	if(rt && rt->rt_refcnt == 0) { /* this is first reference */
155 		if(rt->rt_flags & RTPRF_OURS) {
156 			rt->rt_flags &= ~RTPRF_OURS;
157 			rt->rt_rmx.rmx_expire = 0;
158 		}
159 	}
160 	return rn;
161 }
162 
163 static int rtq_reallyold = 60*60;
164 	/* one hour is ``really old'' */
165 SYSCTL_INT(_net_inet_ip, IPCTL_RTEXPIRE, rtexpire,
166 	CTLFLAG_RW, &rtq_reallyold , 0, "");
167 
168 static int rtq_minreallyold = 10;
169 	/* never automatically crank down to less */
170 SYSCTL_INT(_net_inet_ip, IPCTL_RTMINEXPIRE, rtminexpire,
171 	CTLFLAG_RW, &rtq_minreallyold , 0, "");
172 
173 static int rtq_toomany = 128;
174 	/* 128 cached routes is ``too many'' */
175 SYSCTL_INT(_net_inet_ip, IPCTL_RTMAXCACHE, rtmaxcache,
176 	CTLFLAG_RW, &rtq_toomany , 0, "");
177 
178 
179 /*
180  * On last reference drop, mark the route as belong to us so that it can be
181  * timed out.
182  */
183 static void
184 in_clsroute(struct radix_node *rn, struct radix_node_head *head)
185 {
186 	struct rtentry *rt = (struct rtentry *)rn;
187 
188 	if(!(rt->rt_flags & RTF_UP))
189 		return;		/* prophylactic measures */
190 
191 	if((rt->rt_flags & (RTF_LLINFO | RTF_HOST)) != RTF_HOST)
192 		return;
193 
194 	if((rt->rt_flags & (RTF_WASCLONED | RTPRF_OURS))
195 	   != RTF_WASCLONED)
196 		return;
197 
198 	/*
199 	 * As requested by David Greenman:
200 	 * If rtq_reallyold is 0, just delete the route without
201 	 * waiting for a timeout cycle to kill it.
202 	 */
203 	if(rtq_reallyold != 0) {
204 		rt->rt_flags |= RTPRF_OURS;
205 		rt->rt_rmx.rmx_expire = time_second + rtq_reallyold;
206 	} else {
207 		rtrequest(RTM_DELETE,
208 			  (struct sockaddr *)rt_key(rt),
209 			  rt->rt_gateway, rt_mask(rt),
210 			  rt->rt_flags, 0);
211 	}
212 }
213 
214 struct rtqk_arg {
215 	struct radix_node_head *rnh;
216 	int draining;
217 	int killed;
218 	int found;
219 	int updating;
220 	time_t nextstop;
221 };
222 
223 /*
224  * Get rid of old routes.  When draining, this deletes everything, even when
225  * the timeout is not expired yet.  When updating, this makes sure that
226  * nothing has a timeout longer than the current value of rtq_reallyold.
227  */
228 static int
229 in_rtqkill(struct radix_node *rn, void *rock)
230 {
231 	struct rtqk_arg *ap = rock;
232 	struct rtentry *rt = (struct rtentry *)rn;
233 	int err;
234 
235 	if(rt->rt_flags & RTPRF_OURS) {
236 		ap->found++;
237 
238 		if(ap->draining || rt->rt_rmx.rmx_expire <= time_second) {
239 			if(rt->rt_refcnt > 0)
240 				panic("rtqkill route really not free");
241 
242 			err = rtrequest(RTM_DELETE,
243 					(struct sockaddr *)rt_key(rt),
244 					rt->rt_gateway, rt_mask(rt),
245 					rt->rt_flags, 0);
246 			if(err) {
247 				log(LOG_WARNING, "in_rtqkill: error %d\n", err);
248 			} else {
249 				ap->killed++;
250 			}
251 		} else {
252 			if(ap->updating
253 			   && (rt->rt_rmx.rmx_expire - time_second
254 			       > rtq_reallyold)) {
255 				rt->rt_rmx.rmx_expire = time_second
256 					+ rtq_reallyold;
257 			}
258 			ap->nextstop = lmin(ap->nextstop,
259 					    rt->rt_rmx.rmx_expire);
260 		}
261 	}
262 
263 	return 0;
264 }
265 
266 #define RTQ_TIMEOUT	60*10	/* run no less than once every ten minutes */
267 static int rtq_timeout = RTQ_TIMEOUT;
268 
269 static void
270 in_rtqtimo(void *rock)
271 {
272 	struct radix_node_head *rnh = rock;
273 	struct rtqk_arg arg;
274 	struct timeval atv;
275 	static time_t last_adjusted_timeout = 0;
276 	int s;
277 
278 	arg.found = arg.killed = 0;
279 	arg.rnh = rnh;
280 	arg.nextstop = time_second + rtq_timeout;
281 	arg.draining = arg.updating = 0;
282 	s = splnet();
283 	rnh->rnh_walktree(rnh, in_rtqkill, &arg);
284 	splx(s);
285 
286 	/*
287 	 * Attempt to be somewhat dynamic about this:
288 	 * If there are ``too many'' routes sitting around taking up space,
289 	 * then crank down the timeout, and see if we can't make some more
290 	 * go away.  However, we make sure that we will never adjust more
291 	 * than once in rtq_timeout seconds, to keep from cranking down too
292 	 * hard.
293 	 */
294 	if((arg.found - arg.killed > rtq_toomany)
295 	   && (time_second - last_adjusted_timeout >= rtq_timeout)
296 	   && rtq_reallyold > rtq_minreallyold) {
297 		rtq_reallyold = 2*rtq_reallyold / 3;
298 		if(rtq_reallyold < rtq_minreallyold) {
299 			rtq_reallyold = rtq_minreallyold;
300 		}
301 
302 		last_adjusted_timeout = time_second;
303 #ifdef DIAGNOSTIC
304 		log(LOG_DEBUG, "in_rtqtimo: adjusted rtq_reallyold to %d\n",
305 		    rtq_reallyold);
306 #endif
307 		arg.found = arg.killed = 0;
308 		arg.updating = 1;
309 		s = splnet();
310 		rnh->rnh_walktree(rnh, in_rtqkill, &arg);
311 		splx(s);
312 	}
313 
314 	atv.tv_usec = 0;
315 	atv.tv_sec = arg.nextstop - time_second;
316 	timeout(in_rtqtimo, rock, tvtohz(&atv));
317 }
318 
319 void
320 in_rtqdrain(void)
321 {
322 	struct radix_node_head *rnh = rt_tables[AF_INET];
323 	struct rtqk_arg arg;
324 	int s;
325 	arg.found = arg.killed = 0;
326 	arg.rnh = rnh;
327 	arg.nextstop = 0;
328 	arg.draining = 1;
329 	arg.updating = 0;
330 	s = splnet();
331 	rnh->rnh_walktree(rnh, in_rtqkill, &arg);
332 	splx(s);
333 }
334 
335 /*
336  * Initialize our routing tree.
337  */
338 int
339 in_inithead(void **head, int off)
340 {
341 	struct radix_node_head *rnh;
342 
343 	if(!rn_inithead(head, off))
344 		return 0;
345 
346 	if(head != (void **)&rt_tables[AF_INET]) /* BOGUS! */
347 		return 1;	/* only do this for the real routing table */
348 
349 	rnh = *head;
350 	rnh->rnh_addaddr = in_addroute;
351 	rnh->rnh_matchaddr = in_matroute;
352 	rnh->rnh_close = in_clsroute;
353 	in_rtqtimo(rnh);	/* kick off timeout first time */
354 	return 1;
355 }
356 
357 
358 /*
359  * This zaps old routes when the interface goes down.
360  * Currently it doesn't delete static routes; there are
361  * arguments one could make for both behaviors.  For the moment,
362  * we will adopt the Principle of Least Surprise and leave them
363  * alone (with the knowledge that this will not be enough for some
364  * people).  The ones we really want to get rid of are things like ARP
365  * entries, since the user might down the interface, walk over to a completely
366  * different network, and plug back in.
367  */
368 struct in_ifadown_arg {
369 	struct radix_node_head *rnh;
370 	struct ifaddr *ifa;
371 };
372 
373 static int
374 in_ifadownkill(struct radix_node *rn, void *xap)
375 {
376 	struct in_ifadown_arg *ap = xap;
377 	struct rtentry *rt = (struct rtentry *)rn;
378 	int err;
379 
380 	if (rt->rt_ifa == ap->ifa && !(rt->rt_flags & RTF_STATIC)) {
381 		/*
382 		 * We need to disable the automatic prune that happens
383 		 * in this case in rtrequest() because it will blow
384 		 * away the pointers that rn_walktree() needs in order
385 		 * continue our descent.  We will end up deleting all
386 		 * the routes that rtrequest() would have in any case,
387 		 * so that behavior is not needed there.
388 		 */
389 		rt->rt_flags &= ~RTF_PRCLONING;
390 		err = rtrequest(RTM_DELETE, (struct sockaddr *)rt_key(rt),
391 				rt->rt_gateway, rt_mask(rt), rt->rt_flags, 0);
392 		if (err) {
393 			log(LOG_WARNING, "in_ifadownkill: error %d\n", err);
394 		}
395 	}
396 	return 0;
397 }
398 
399 int
400 in_ifadown(struct ifaddr *ifa)
401 {
402 	struct in_ifadown_arg arg;
403 	struct radix_node_head *rnh;
404 
405 	if (ifa->ifa_addr->sa_family != AF_INET)
406 		return 1;
407 
408 	arg.rnh = rnh = rt_tables[AF_INET];
409 	arg.ifa = ifa;
410 	rnh->rnh_walktree(rnh, in_ifadownkill, &arg);
411 	ifa->ifa_flags &= ~IFA_ROUTE;
412 	return 0;
413 }
414