xref: /freebsd/sys/netinet/in_rmx.c (revision 9336e0699bda8a301cd2bfa37106b6ec5e32012e)
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 
30 /*
31  * This code does two things necessary for the enhanced TCP metrics to
32  * function in a useful manner:
33  *  1) It marks all non-host routes as `cloning', thus ensuring that
34  *     every actual reference to such a route actually gets turned
35  *     into a reference to a host route to the specific destination
36  *     requested.
37  *  2) When such routes lose all their references, it arranges for them
38  *     to be deleted in some random collection of circumstances, so that
39  *     a large quantity of stale routing data is not kept in kernel memory
40  *     indefinitely.  See in_rtqtimo() below for the exact mechanism.
41  */
42 
43 #include <sys/cdefs.h>
44 __FBSDID("$FreeBSD$");
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/callout.h>
54 
55 #include <net/if.h>
56 #include <net/route.h>
57 #include <netinet/in.h>
58 #include <netinet/in_var.h>
59 #include <netinet/ip_var.h>
60 
61 extern int	in_inithead(void **head, int off);
62 
63 #define RTPRF_OURS		RTF_PROTO3	/* set on routes we manage */
64 
65 /*
66  * Do what we need to do when inserting a route.
67  */
68 static struct radix_node *
69 in_addroute(void *v_arg, void *n_arg, struct radix_node_head *head,
70     struct radix_node *treenodes)
71 {
72 	struct rtentry *rt = (struct rtentry *)treenodes;
73 	struct sockaddr_in *sin = (struct sockaddr_in *)rt_key(rt);
74 	struct radix_node *ret;
75 
76 	/*
77 	 * A little bit of help for both IP output and input:
78 	 *   For host routes, we make sure that RTF_BROADCAST
79 	 *   is set for anything that looks like a broadcast address.
80 	 *   This way, we can avoid an expensive call to in_broadcast()
81 	 *   in ip_output() most of the time (because the route passed
82 	 *   to ip_output() is almost always a host route).
83 	 *
84 	 *   We also do the same for local addresses, with the thought
85 	 *   that this might one day be used to speed up ip_input().
86 	 *
87 	 * We also mark routes to multicast addresses as such, because
88 	 * it's easy to do and might be useful (but this is much more
89 	 * dubious since it's so easy to inspect the address).
90 	 */
91 	if (rt->rt_flags & RTF_HOST) {
92 		if (in_broadcast(sin->sin_addr, rt->rt_ifp)) {
93 			rt->rt_flags |= RTF_BROADCAST;
94 		} else if (satosin(rt->rt_ifa->ifa_addr)->sin_addr.s_addr ==
95 		    sin->sin_addr.s_addr) {
96 			rt->rt_flags |= RTF_LOCAL;
97 		}
98 	}
99 	if (IN_MULTICAST(ntohl(sin->sin_addr.s_addr)))
100 		rt->rt_flags |= RTF_MULTICAST;
101 
102 	if (!rt->rt_rmx.rmx_mtu && rt->rt_ifp)
103 		rt->rt_rmx.rmx_mtu = rt->rt_ifp->if_mtu;
104 
105 	ret = rn_addroute(v_arg, n_arg, head, treenodes);
106 	if (ret == NULL && rt->rt_flags & RTF_HOST) {
107 		struct rtentry *rt2;
108 		/*
109 		 * We are trying to add a host route, but can't.
110 		 * Find out if it is because of an
111 		 * ARP entry and delete it if so.
112 		 */
113 		rt2 = rtalloc1((struct sockaddr *)sin, 0, RTF_CLONING);
114 		if (rt2) {
115 			if (rt2->rt_flags & RTF_LLINFO &&
116 			    rt2->rt_flags & RTF_HOST &&
117 			    rt2->rt_gateway &&
118 			    rt2->rt_gateway->sa_family == AF_LINK) {
119 				rtexpunge(rt2);
120 				RTFREE_LOCKED(rt2);
121 				ret = rn_addroute(v_arg, n_arg, head,
122 						  treenodes);
123 			} else
124 				RTFREE_LOCKED(rt2);
125 		}
126 	}
127 
128 	return ret;
129 }
130 
131 /*
132  * This code is the inverse of in_clsroute: on first reference, if we
133  * were managing the route, stop doing so and set the expiration timer
134  * back off again.
135  */
136 static struct radix_node *
137 in_matroute(void *v_arg, struct radix_node_head *head)
138 {
139 	struct radix_node *rn = rn_match(v_arg, head);
140 	struct rtentry *rt = (struct rtentry *)rn;
141 
142 	/*XXX locking? */
143 	if (rt && rt->rt_refcnt == 0) {		/* this is first reference */
144 		if (rt->rt_flags & RTPRF_OURS) {
145 			rt->rt_flags &= ~RTPRF_OURS;
146 			rt->rt_rmx.rmx_expire = 0;
147 		}
148 	}
149 	return rn;
150 }
151 
152 static int rtq_reallyold = 60*60;		/* one hour is "really old" */
153 SYSCTL_INT(_net_inet_ip, IPCTL_RTEXPIRE, rtexpire, CTLFLAG_RW,
154     &rtq_reallyold, 0, "Default expiration time on dynamically learned routes");
155 
156 static int rtq_minreallyold = 10;  /* never automatically crank down to less */
157 SYSCTL_INT(_net_inet_ip, IPCTL_RTMINEXPIRE, rtminexpire, CTLFLAG_RW,
158     &rtq_minreallyold, 0,
159     "Minimum time to attempt to hold onto dynamically learned routes");
160 
161 static int rtq_toomany = 128;		/* 128 cached routes is "too many" */
162 SYSCTL_INT(_net_inet_ip, IPCTL_RTMAXCACHE, rtmaxcache, CTLFLAG_RW,
163     &rtq_toomany, 0, "Upper limit on dynamically learned routes");
164 
165 /*
166  * On last reference drop, mark the route as belong to us so that it can be
167  * timed out.
168  */
169 static void
170 in_clsroute(struct radix_node *rn, struct radix_node_head *head)
171 {
172 	struct rtentry *rt = (struct rtentry *)rn;
173 
174 	RT_LOCK_ASSERT(rt);
175 
176 	if (!(rt->rt_flags & RTF_UP))
177 		return;			/* prophylactic measures */
178 
179 	if ((rt->rt_flags & (RTF_LLINFO | RTF_HOST)) != RTF_HOST)
180 		return;
181 
182 	if (rt->rt_flags & RTPRF_OURS)
183 		return;
184 
185 	if (!(rt->rt_flags & (RTF_WASCLONED | RTF_DYNAMIC)))
186 		return;
187 
188 	/*
189 	 * If rtq_reallyold is 0, just delete the route without
190 	 * waiting for a timeout cycle to kill it.
191 	 */
192 	if (rtq_reallyold != 0) {
193 		rt->rt_flags |= RTPRF_OURS;
194 		rt->rt_rmx.rmx_expire = time_uptime + rtq_reallyold;
195 	} else {
196 		rtexpunge(rt);
197 	}
198 }
199 
200 struct rtqk_arg {
201 	struct radix_node_head *rnh;
202 	int draining;
203 	int killed;
204 	int found;
205 	int updating;
206 	time_t nextstop;
207 };
208 
209 /*
210  * Get rid of old routes.  When draining, this deletes everything, even when
211  * the timeout is not expired yet.  When updating, this makes sure that
212  * nothing has a timeout longer than the current value of rtq_reallyold.
213  */
214 static int
215 in_rtqkill(struct radix_node *rn, void *rock)
216 {
217 	struct rtqk_arg *ap = rock;
218 	struct rtentry *rt = (struct rtentry *)rn;
219 	int err;
220 
221 	if (rt->rt_flags & RTPRF_OURS) {
222 		ap->found++;
223 
224 		if (ap->draining || rt->rt_rmx.rmx_expire <= time_uptime) {
225 			if (rt->rt_refcnt > 0)
226 				panic("rtqkill route really not free");
227 
228 			err = rtrequest(RTM_DELETE,
229 					(struct sockaddr *)rt_key(rt),
230 					rt->rt_gateway, rt_mask(rt),
231 					rt->rt_flags, 0);
232 			if (err) {
233 				log(LOG_WARNING, "in_rtqkill: error %d\n", err);
234 			} else {
235 				ap->killed++;
236 			}
237 		} else {
238 			if (ap->updating &&
239 			    (rt->rt_rmx.rmx_expire - time_uptime >
240 			     rtq_reallyold)) {
241 				rt->rt_rmx.rmx_expire =
242 				    time_uptime + rtq_reallyold;
243 			}
244 			ap->nextstop = lmin(ap->nextstop,
245 					    rt->rt_rmx.rmx_expire);
246 		}
247 	}
248 
249 	return 0;
250 }
251 
252 #define RTQ_TIMEOUT	60*10	/* run no less than once every ten minutes */
253 static int rtq_timeout = RTQ_TIMEOUT;
254 static struct callout rtq_timer;
255 
256 static void
257 in_rtqtimo(void *rock)
258 {
259 	struct radix_node_head *rnh = rock;
260 	struct rtqk_arg arg;
261 	struct timeval atv;
262 	static time_t last_adjusted_timeout = 0;
263 
264 	arg.found = arg.killed = 0;
265 	arg.rnh = rnh;
266 	arg.nextstop = time_uptime + rtq_timeout;
267 	arg.draining = arg.updating = 0;
268 	RADIX_NODE_HEAD_LOCK(rnh);
269 	rnh->rnh_walktree(rnh, in_rtqkill, &arg);
270 	RADIX_NODE_HEAD_UNLOCK(rnh);
271 
272 	/*
273 	 * Attempt to be somewhat dynamic about this:
274 	 * If there are ``too many'' routes sitting around taking up space,
275 	 * then crank down the timeout, and see if we can't make some more
276 	 * go away.  However, we make sure that we will never adjust more
277 	 * than once in rtq_timeout seconds, to keep from cranking down too
278 	 * hard.
279 	 */
280 	if ((arg.found - arg.killed > rtq_toomany) &&
281 	    (time_uptime - last_adjusted_timeout >= rtq_timeout) &&
282 	    rtq_reallyold > rtq_minreallyold) {
283 		rtq_reallyold = 2 * rtq_reallyold / 3;
284 		if (rtq_reallyold < rtq_minreallyold) {
285 			rtq_reallyold = rtq_minreallyold;
286 		}
287 
288 		last_adjusted_timeout = time_uptime;
289 #ifdef DIAGNOSTIC
290 		log(LOG_DEBUG, "in_rtqtimo: adjusted rtq_reallyold to %d\n",
291 		    rtq_reallyold);
292 #endif
293 		arg.found = arg.killed = 0;
294 		arg.updating = 1;
295 		RADIX_NODE_HEAD_LOCK(rnh);
296 		rnh->rnh_walktree(rnh, in_rtqkill, &arg);
297 		RADIX_NODE_HEAD_UNLOCK(rnh);
298 	}
299 
300 	atv.tv_usec = 0;
301 	atv.tv_sec = arg.nextstop - time_uptime;
302 	callout_reset(&rtq_timer, tvtohz(&atv), in_rtqtimo, rock);
303 }
304 
305 void
306 in_rtqdrain(void)
307 {
308 	struct radix_node_head *rnh = rt_tables[AF_INET];
309 	struct rtqk_arg arg;
310 
311 	arg.found = arg.killed = 0;
312 	arg.rnh = rnh;
313 	arg.nextstop = 0;
314 	arg.draining = 1;
315 	arg.updating = 0;
316 	RADIX_NODE_HEAD_LOCK(rnh);
317 	rnh->rnh_walktree(rnh, in_rtqkill, &arg);
318 	RADIX_NODE_HEAD_UNLOCK(rnh);
319 }
320 
321 /*
322  * Initialize our routing tree.
323  */
324 int
325 in_inithead(void **head, int off)
326 {
327 	struct radix_node_head *rnh;
328 
329 	if (!rn_inithead(head, off))
330 		return 0;
331 
332 	if (head != (void **)&rt_tables[AF_INET])	/* BOGUS! */
333 		return 1;	/* only do this for the real routing table */
334 
335 	rnh = *head;
336 	rnh->rnh_addaddr = in_addroute;
337 	rnh->rnh_matchaddr = in_matroute;
338 	rnh->rnh_close = in_clsroute;
339 	callout_init(&rtq_timer, CALLOUT_MPSAFE);
340 	in_rtqtimo(rnh);	/* kick off timeout first time */
341 	return 1;
342 }
343 
344 /*
345  * This zaps old routes when the interface goes down or interface
346  * address is deleted.  In the latter case, it deletes static routes
347  * that point to this address.  If we don't do this, we may end up
348  * using the old address in the future.  The ones we always want to
349  * get rid of are things like ARP entries, since the user might down
350  * the interface, walk over to a completely different network, and
351  * plug back in.
352  */
353 struct in_ifadown_arg {
354 	struct radix_node_head *rnh;
355 	struct ifaddr *ifa;
356 	int del;
357 };
358 
359 static int
360 in_ifadownkill(struct radix_node *rn, void *xap)
361 {
362 	struct in_ifadown_arg *ap = xap;
363 	struct rtentry *rt = (struct rtentry *)rn;
364 
365 	RT_LOCK(rt);
366 	if (rt->rt_ifa == ap->ifa &&
367 	    (ap->del || !(rt->rt_flags & RTF_STATIC))) {
368 		/*
369 		 * We need to disable the automatic prune that happens
370 		 * in this case in rtrequest() because it will blow
371 		 * away the pointers that rn_walktree() needs in order
372 		 * continue our descent.  We will end up deleting all
373 		 * the routes that rtrequest() would have in any case,
374 		 * so that behavior is not needed there.
375 		 */
376 		rt->rt_flags &= ~RTF_CLONING;
377 		rtexpunge(rt);
378 	}
379 	RT_UNLOCK(rt);
380 	return 0;
381 }
382 
383 int
384 in_ifadown(struct ifaddr *ifa, int delete)
385 {
386 	struct in_ifadown_arg arg;
387 	struct radix_node_head *rnh;
388 
389 	if (ifa->ifa_addr->sa_family != AF_INET)
390 		return 1;
391 
392 	arg.rnh = rnh = rt_tables[AF_INET];
393 	arg.ifa = ifa;
394 	arg.del = delete;
395 	RADIX_NODE_HEAD_LOCK(rnh);
396 	rnh->rnh_walktree(rnh, in_ifadownkill, &arg);
397 	RADIX_NODE_HEAD_UNLOCK(rnh);
398 	ifa->ifa_flags &= ~IFA_ROUTE;		/* XXXlocking? */
399 	return 0;
400 }
401