in_rmx.c revision 15652 
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 1.24 1996/04/26 18:31:41 wollman 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/queue.h>
50#include <sys/socket.h>
51#include <sys/socketvar.h>
52#include <sys/mbuf.h>
53#include <sys/syslog.h>
54
55#include <net/if.h>
56#include <net/route.h>
57#include <netinet/in.h>
58#include <netinet/in_systm.h>
59#include <netinet/in_var.h>
60
61#include <netinet/ip.h>
62#include <netinet/ip_var.h>
63
64#include <netinet/tcp.h>
65#include <netinet/tcp_seq.h>
66#include <netinet/tcp_timer.h>
67#include <netinet/tcp_var.h>
68
69extern int	in_inithead __P((void **head, int off));
70
71#define RTPRF_OURS		RTF_PROTO3	/* set on routes we manage */
72
73/*
74 * Do what we need to do when inserting a route.
75 */
76static struct radix_node *
77in_addroute(void *v_arg, void *n_arg, struct radix_node_head *head,
78	    struct radix_node *treenodes)
79{
80	struct rtentry *rt = (struct rtentry *)treenodes;
81	struct sockaddr_in *sin = (struct sockaddr_in *)rt_key(rt);
82	struct radix_node *ret;
83
84	/*
85	 * For IP, all unicast non-host routes are automatically cloning.
86	 */
87	if(IN_MULTICAST(ntohl(sin->sin_addr.s_addr)))
88		rt->rt_flags |= RTF_MULTICAST;
89
90	if(!(rt->rt_flags & (RTF_HOST | RTF_CLONING | RTF_MULTICAST))) {
91		rt->rt_flags |= RTF_PRCLONING;
92	}
93
94	/*
95	 * A little bit of help for both IP output and input:
96	 *   For host routes, we make sure that RTF_BROADCAST
97	 *   is set for anything that looks like a broadcast address.
98	 *   This way, we can avoid an expensive call to in_broadcast()
99	 *   in ip_output() most of the time (because the route passed
100	 *   to ip_output() is almost always a host route).
101	 *
102	 *   We also do the same for local addresses, with the thought
103	 *   that this might one day be used to speed up ip_input().
104	 *
105	 * We also mark routes to multicast addresses as such, because
106	 * it's easy to do and might be useful (but this is much more
107	 * dubious since it's so easy to inspect the address).  (This
108	 * is done above.)
109	 */
110	if (rt->rt_flags & RTF_HOST) {
111		if (in_broadcast(sin->sin_addr, rt->rt_ifp)) {
112			rt->rt_flags |= RTF_BROADCAST;
113		} else {
114#define satosin(sa) ((struct sockaddr_in *)sa)
115			if (satosin(rt->rt_ifa->ifa_addr)->sin_addr.s_addr
116			    == sin->sin_addr.s_addr)
117				rt->rt_flags |= RTF_LOCAL;
118#undef satosin
119		}
120	}
121
122	/*
123	 * We also specify a send and receive pipe size for every
124	 * route added, to help TCP a bit.  TCP doesn't actually
125	 * want a true pipe size, which would be prohibitive in memory
126	 * costs and is hard to compute anyway; it simply uses these
127	 * values to size its buffers.  So, we fill them in with the
128	 * same values that TCP would have used anyway, and allow the
129	 * installing program or the link layer to override these values
130	 * as it sees fit.  This will hopefully allow TCP more
131	 * opportunities to save its ssthresh value.
132	 */
133	if (!rt->rt_rmx.rmx_sendpipe && !(rt->rt_rmx.rmx_locks & RTV_SPIPE))
134		rt->rt_rmx.rmx_sendpipe = tcp_sendspace;
135
136	if (!rt->rt_rmx.rmx_recvpipe && !(rt->rt_rmx.rmx_locks & RTV_RPIPE))
137		rt->rt_rmx.rmx_recvpipe = tcp_recvspace;
138
139	if (!rt->rt_rmx.rmx_mtu && !(rt->rt_rmx.rmx_locks & RTV_MTU)
140	    && rt->rt_ifp)
141		rt->rt_rmx.rmx_mtu = rt->rt_ifp->if_mtu;
142
143	ret = rn_addroute(v_arg, n_arg, head, treenodes);
144	if (ret == NULL && rt->rt_flags & RTF_HOST) {
145		struct rtentry *rt2;
146		/*
147		 * We are trying to add a host route, but can't.
148		 * Find out if it is because of an
149		 * ARP entry and delete it if so.
150		 */
151		rt2 = rtalloc1((struct sockaddr *)sin, 0,
152				RTF_CLONING | RTF_PRCLONING);
153		if (rt2) {
154			if (rt2->rt_flags & RTF_LLINFO &&
155				rt2->rt_flags & RTF_HOST &&
156				rt2->rt_gateway &&
157				rt2->rt_gateway->sa_family == AF_LINK) {
158				rtrequest(RTM_DELETE,
159					  (struct sockaddr *)rt_key(rt2),
160					  rt2->rt_gateway,
161					  rt_mask(rt2), rt2->rt_flags, 0);
162				ret = rn_addroute(v_arg, n_arg, head,
163					treenodes);
164			}
165			RTFREE(rt2);
166		}
167	}
168	return ret;
169}
170
171/*
172 * This code is the inverse of in_clsroute: on first reference, if we
173 * were managing the route, stop doing so and set the expiration timer
174 * back off again.
175 */
176static struct radix_node *
177in_matroute(void *v_arg, struct radix_node_head *head)
178{
179	struct radix_node *rn = rn_match(v_arg, head);
180	struct rtentry *rt = (struct rtentry *)rn;
181
182	if(rt && rt->rt_refcnt == 0) { /* this is first reference */
183		if(rt->rt_flags & RTPRF_OURS) {
184			rt->rt_flags &= ~RTPRF_OURS;
185			rt->rt_rmx.rmx_expire = 0;
186		}
187	}
188	return rn;
189}
190
191static int rtq_reallyold = 60*60;
192	/* one hour is ``really old'' */
193SYSCTL_INT(_net_inet_ip, IPCTL_RTEXPIRE, rtexpire,
194	CTLFLAG_RW, &rtq_reallyold , 0, "");
195
196static int rtq_minreallyold = 10;
197	/* never automatically crank down to less */
198SYSCTL_INT(_net_inet_ip, IPCTL_RTMINEXPIRE, rtminexpire,
199	CTLFLAG_RW, &rtq_minreallyold , 0, "");
200
201static int rtq_toomany = 128;
202	/* 128 cached routes is ``too many'' */
203SYSCTL_INT(_net_inet_ip, IPCTL_RTMAXCACHE, rtmaxcache,
204	CTLFLAG_RW, &rtq_toomany , 0, "");
205
206
207/*
208 * On last reference drop, mark the route as belong to us so that it can be
209 * timed out.
210 */
211static void
212in_clsroute(struct radix_node *rn, struct radix_node_head *head)
213{
214	struct rtentry *rt = (struct rtentry *)rn;
215
216	if(!(rt->rt_flags & RTF_UP))
217		return;		/* prophylactic measures */
218
219	if((rt->rt_flags & (RTF_LLINFO | RTF_HOST)) != RTF_HOST)
220		return;
221
222	if((rt->rt_flags & (RTF_WASCLONED | RTPRF_OURS))
223	   != RTF_WASCLONED)
224		return;
225
226	/*
227	 * As requested by David Greenman:
228	 * If rtq_reallyold is 0, just delete the route without
229	 * waiting for a timeout cycle to kill it.
230	 */
231	if(rtq_reallyold != 0) {
232		rt->rt_flags |= RTPRF_OURS;
233		rt->rt_rmx.rmx_expire = time.tv_sec + rtq_reallyold;
234	} else {
235		rtrequest(RTM_DELETE,
236			  (struct sockaddr *)rt_key(rt),
237			  rt->rt_gateway, rt_mask(rt),
238			  rt->rt_flags, 0);
239	}
240}
241
242struct rtqk_arg {
243	struct radix_node_head *rnh;
244	int draining;
245	int killed;
246	int found;
247	int updating;
248	time_t nextstop;
249};
250
251/*
252 * Get rid of old routes.  When draining, this deletes everything, even when
253 * the timeout is not expired yet.  When updating, this makes sure that
254 * nothing has a timeout longer than the current value of rtq_reallyold.
255 */
256static int
257in_rtqkill(struct radix_node *rn, void *rock)
258{
259	struct rtqk_arg *ap = rock;
260	struct rtentry *rt = (struct rtentry *)rn;
261	int err;
262
263	if(rt->rt_flags & RTPRF_OURS) {
264		ap->found++;
265
266		if(ap->draining || rt->rt_rmx.rmx_expire <= time.tv_sec) {
267			if(rt->rt_refcnt > 0)
268				panic("rtqkill route really not free");
269
270			err = rtrequest(RTM_DELETE,
271					(struct sockaddr *)rt_key(rt),
272					rt->rt_gateway, rt_mask(rt),
273					rt->rt_flags, 0);
274			if(err) {
275				log(LOG_WARNING, "in_rtqkill: error %d\n", err);
276			} else {
277				ap->killed++;
278			}
279		} else {
280			if(ap->updating
281			   && (rt->rt_rmx.rmx_expire - time.tv_sec
282			       > rtq_reallyold)) {
283				rt->rt_rmx.rmx_expire = time.tv_sec
284					+ rtq_reallyold;
285			}
286			ap->nextstop = lmin(ap->nextstop,
287					    rt->rt_rmx.rmx_expire);
288		}
289	}
290
291	return 0;
292}
293
294#define RTQ_TIMEOUT	60*10	/* run no less than once every ten minutes */
295static int rtq_timeout = RTQ_TIMEOUT;
296
297static void
298in_rtqtimo(void *rock)
299{
300	struct radix_node_head *rnh = rock;
301	struct rtqk_arg arg;
302	struct timeval atv;
303	static time_t last_adjusted_timeout = 0;
304	int s;
305
306	arg.found = arg.killed = 0;
307	arg.rnh = rnh;
308	arg.nextstop = time.tv_sec + rtq_timeout;
309	arg.draining = arg.updating = 0;
310	s = splnet();
311	rnh->rnh_walktree(rnh, in_rtqkill, &arg);
312	splx(s);
313
314	/*
315	 * Attempt to be somewhat dynamic about this:
316	 * If there are ``too many'' routes sitting around taking up space,
317	 * then crank down the timeout, and see if we can't make some more
318	 * go away.  However, we make sure that we will never adjust more
319	 * than once in rtq_timeout seconds, to keep from cranking down too
320	 * hard.
321	 */
322	if((arg.found - arg.killed > rtq_toomany)
323	   && (time.tv_sec - last_adjusted_timeout >= rtq_timeout)
324	   && rtq_reallyold > rtq_minreallyold) {
325		rtq_reallyold = 2*rtq_reallyold / 3;
326		if(rtq_reallyold < rtq_minreallyold) {
327			rtq_reallyold = rtq_minreallyold;
328		}
329
330		last_adjusted_timeout = time.tv_sec;
331		log(LOG_DEBUG, "in_rtqtimo: adjusted rtq_reallyold to %d\n",
332		    rtq_reallyold);
333		arg.found = arg.killed = 0;
334		arg.updating = 1;
335		s = splnet();
336		rnh->rnh_walktree(rnh, in_rtqkill, &arg);
337		splx(s);
338	}
339
340	atv.tv_usec = 0;
341	atv.tv_sec = arg.nextstop;
342	timeout(in_rtqtimo, rock, hzto(&atv));
343}
344
345void
346in_rtqdrain(void)
347{
348	struct radix_node_head *rnh = rt_tables[AF_INET];
349	struct rtqk_arg arg;
350	int s;
351	arg.found = arg.killed = 0;
352	arg.rnh = rnh;
353	arg.nextstop = 0;
354	arg.draining = 1;
355	arg.updating = 0;
356	s = splnet();
357	rnh->rnh_walktree(rnh, in_rtqkill, &arg);
358	splx(s);
359}
360
361/*
362 * Initialize our routing tree.
363 */
364int
365in_inithead(void **head, int off)
366{
367	struct radix_node_head *rnh;
368
369	if(!rn_inithead(head, off))
370		return 0;
371
372	if(head != (void **)&rt_tables[AF_INET]) /* BOGUS! */
373		return 1;	/* only do this for the real routing table */
374
375	rnh = *head;
376	rnh->rnh_addaddr = in_addroute;
377	rnh->rnh_matchaddr = in_matroute;
378	rnh->rnh_close = in_clsroute;
379	in_rtqtimo(rnh);	/* kick off timeout first time */
380	return 1;
381}
382
383