in_rmx.c revision 15395 
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 1.23 1996/01/23 05:15:30 fenner 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(!(rt->rt_flags & (RTF_HOST | RTF_CLONING))) {
88		if(!IN_MULTICAST(ntohl(sin->sin_addr.s_addr))) {
89			rt->rt_flags |= RTF_PRCLONING;
90		}
91	}
92
93	/*
94	 * We also specify a send and receive pipe size for every
95	 * route added, to help TCP a bit.  TCP doesn't actually
96	 * want a true pipe size, which would be prohibitive in memory
97	 * costs and is hard to compute anyway; it simply uses these
98	 * values to size its buffers.  So, we fill them in with the
99	 * same values that TCP would have used anyway, and allow the
100	 * installing program or the link layer to override these values
101	 * as it sees fit.  This will hopefully allow TCP more
102	 * opportunities to save its ssthresh value.
103	 */
104	if (!rt->rt_rmx.rmx_sendpipe && !(rt->rt_rmx.rmx_locks & RTV_SPIPE))
105		rt->rt_rmx.rmx_sendpipe = tcp_sendspace;
106
107	if (!rt->rt_rmx.rmx_recvpipe && !(rt->rt_rmx.rmx_locks & RTV_RPIPE))
108		rt->rt_rmx.rmx_recvpipe = tcp_recvspace;
109
110	if (!rt->rt_rmx.rmx_mtu && !(rt->rt_rmx.rmx_locks & RTV_MTU)
111	    && rt->rt_ifp)
112		rt->rt_rmx.rmx_mtu = rt->rt_ifp->if_mtu;
113
114	ret = rn_addroute(v_arg, n_arg, head, treenodes);
115	if (ret == NULL && rt->rt_flags & RTF_HOST) {
116		struct rtentry *rt2;
117		/*
118		 * We are trying to add a host route, but can't.
119		 * Find out if it is because of an
120		 * ARP entry and delete it if so.
121		 */
122		rt2 = rtalloc1((struct sockaddr *)sin, 0,
123				RTF_CLONING | RTF_PRCLONING);
124		if (rt2) {
125			if (rt2->rt_flags & RTF_LLINFO &&
126				rt2->rt_flags & RTF_HOST &&
127				rt2->rt_gateway &&
128				rt2->rt_gateway->sa_family == AF_LINK) {
129				rtrequest(RTM_DELETE,
130					  (struct sockaddr *)rt_key(rt2),
131					  rt2->rt_gateway,
132					  rt_mask(rt2), rt2->rt_flags, 0);
133				ret = rn_addroute(v_arg, n_arg, head,
134					treenodes);
135			}
136			RTFREE(rt2);
137		}
138	}
139	return ret;
140}
141
142/*
143 * This code is the inverse of in_clsroute: on first reference, if we
144 * were managing the route, stop doing so and set the expiration timer
145 * back off again.
146 */
147static struct radix_node *
148in_matroute(void *v_arg, struct radix_node_head *head)
149{
150	struct radix_node *rn = rn_match(v_arg, head);
151	struct rtentry *rt = (struct rtentry *)rn;
152
153	if(rt && rt->rt_refcnt == 0) { /* this is first reference */
154		if(rt->rt_flags & RTPRF_OURS) {
155			rt->rt_flags &= ~RTPRF_OURS;
156			rt->rt_rmx.rmx_expire = 0;
157		}
158	}
159	return rn;
160}
161
162static int rtq_reallyold = 60*60;
163	/* one hour is ``really old'' */
164SYSCTL_INT(_net_inet_ip, IPCTL_RTEXPIRE, rtexpire,
165	CTLFLAG_RW, &rtq_reallyold , 0, "");
166
167static int rtq_minreallyold = 10;
168	/* never automatically crank down to less */
169SYSCTL_INT(_net_inet_ip, IPCTL_RTMINEXPIRE, rtminexpire,
170	CTLFLAG_RW, &rtq_minreallyold , 0, "");
171
172static int rtq_toomany = 128;
173	/* 128 cached routes is ``too many'' */
174SYSCTL_INT(_net_inet_ip, IPCTL_RTMAXCACHE, rtmaxcache,
175	CTLFLAG_RW, &rtq_toomany , 0, "");
176
177
178/*
179 * On last reference drop, mark the route as belong to us so that it can be
180 * timed out.
181 */
182static void
183in_clsroute(struct radix_node *rn, struct radix_node_head *head)
184{
185	struct rtentry *rt = (struct rtentry *)rn;
186
187	if(!(rt->rt_flags & RTF_UP))
188		return;		/* prophylactic measures */
189
190	if((rt->rt_flags & (RTF_LLINFO | RTF_HOST)) != RTF_HOST)
191		return;
192
193	if((rt->rt_flags & (RTF_WASCLONED | RTPRF_OURS))
194	   != RTF_WASCLONED)
195		return;
196
197	/*
198	 * As requested by David Greenman:
199	 * If rtq_reallyold is 0, just delete the route without
200	 * waiting for a timeout cycle to kill it.
201	 */
202	if(rtq_reallyold != 0) {
203		rt->rt_flags |= RTPRF_OURS;
204		rt->rt_rmx.rmx_expire = time.tv_sec + rtq_reallyold;
205	} else {
206		rtrequest(RTM_DELETE,
207			  (struct sockaddr *)rt_key(rt),
208			  rt->rt_gateway, rt_mask(rt),
209			  rt->rt_flags, 0);
210	}
211}
212
213struct rtqk_arg {
214	struct radix_node_head *rnh;
215	int draining;
216	int killed;
217	int found;
218	int updating;
219	time_t nextstop;
220};
221
222/*
223 * Get rid of old routes.  When draining, this deletes everything, even when
224 * the timeout is not expired yet.  When updating, this makes sure that
225 * nothing has a timeout longer than the current value of rtq_reallyold.
226 */
227static int
228in_rtqkill(struct radix_node *rn, void *rock)
229{
230	struct rtqk_arg *ap = rock;
231	struct rtentry *rt = (struct rtentry *)rn;
232	int err;
233
234	if(rt->rt_flags & RTPRF_OURS) {
235		ap->found++;
236
237		if(ap->draining || rt->rt_rmx.rmx_expire <= time.tv_sec) {
238			if(rt->rt_refcnt > 0)
239				panic("rtqkill route really not free");
240
241			err = rtrequest(RTM_DELETE,
242					(struct sockaddr *)rt_key(rt),
243					rt->rt_gateway, rt_mask(rt),
244					rt->rt_flags, 0);
245			if(err) {
246				log(LOG_WARNING, "in_rtqkill: error %d\n", err);
247			} else {
248				ap->killed++;
249			}
250		} else {
251			if(ap->updating
252			   && (rt->rt_rmx.rmx_expire - time.tv_sec
253			       > rtq_reallyold)) {
254				rt->rt_rmx.rmx_expire = time.tv_sec
255					+ rtq_reallyold;
256			}
257			ap->nextstop = lmin(ap->nextstop,
258					    rt->rt_rmx.rmx_expire);
259		}
260	}
261
262	return 0;
263}
264
265#define RTQ_TIMEOUT	60*10	/* run no less than once every ten minutes */
266static int rtq_timeout = RTQ_TIMEOUT;
267
268static void
269in_rtqtimo(void *rock)
270{
271	struct radix_node_head *rnh = rock;
272	struct rtqk_arg arg;
273	struct timeval atv;
274	static time_t last_adjusted_timeout = 0;
275	int s;
276
277	arg.found = arg.killed = 0;
278	arg.rnh = rnh;
279	arg.nextstop = time.tv_sec + rtq_timeout;
280	arg.draining = arg.updating = 0;
281	s = splnet();
282	rnh->rnh_walktree(rnh, in_rtqkill, &arg);
283	splx(s);
284
285	/*
286	 * Attempt to be somewhat dynamic about this:
287	 * If there are ``too many'' routes sitting around taking up space,
288	 * then crank down the timeout, and see if we can't make some more
289	 * go away.  However, we make sure that we will never adjust more
290	 * than once in rtq_timeout seconds, to keep from cranking down too
291	 * hard.
292	 */
293	if((arg.found - arg.killed > rtq_toomany)
294	   && (time.tv_sec - last_adjusted_timeout >= rtq_timeout)
295	   && rtq_reallyold > rtq_minreallyold) {
296		rtq_reallyold = 2*rtq_reallyold / 3;
297		if(rtq_reallyold < rtq_minreallyold) {
298			rtq_reallyold = rtq_minreallyold;
299		}
300
301		last_adjusted_timeout = time.tv_sec;
302		log(LOG_DEBUG, "in_rtqtimo: adjusted rtq_reallyold to %d\n",
303		    rtq_reallyold);
304		arg.found = arg.killed = 0;
305		arg.updating = 1;
306		s = splnet();
307		rnh->rnh_walktree(rnh, in_rtqkill, &arg);
308		splx(s);
309	}
310
311	atv.tv_usec = 0;
312	atv.tv_sec = arg.nextstop;
313	timeout(in_rtqtimo, rock, hzto(&atv));
314}
315
316void
317in_rtqdrain(void)
318{
319	struct radix_node_head *rnh = rt_tables[AF_INET];
320	struct rtqk_arg arg;
321	int s;
322	arg.found = arg.killed = 0;
323	arg.rnh = rnh;
324	arg.nextstop = 0;
325	arg.draining = 1;
326	arg.updating = 0;
327	s = splnet();
328	rnh->rnh_walktree(rnh, in_rtqkill, &arg);
329	splx(s);
330}
331
332/*
333 * Initialize our routing tree.
334 */
335int
336in_inithead(void **head, int off)
337{
338	struct radix_node_head *rnh;
339
340	if(!rn_inithead(head, off))
341		return 0;
342
343	if(head != (void **)&rt_tables[AF_INET]) /* BOGUS! */
344		return 1;	/* only do this for the real routing table */
345
346	rnh = *head;
347	rnh->rnh_addaddr = in_addroute;
348	rnh->rnh_matchaddr = in_matroute;
349	rnh->rnh_close = in_clsroute;
350	in_rtqtimo(rnh);	/* kick off timeout first time */
351	return 1;
352}
353
354