xref: /netbsd-src/sys/netinet/in_selsrc.c (revision 404fbe5fb94ca1e054339640cabb2801ce52dd30) 
1 /*	$NetBSD: in_selsrc.c,v 1.6 2007/12/04 10:33:11 dyoung Exp $	*/
2 
3 /*-
4  * Copyright (c) 2005 David Young.  All rights reserved.
5  *
6  * This code was written by David Young.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 3. All advertising materials mentioning features or use of this software
17  *    must display the following acknowledgement:
18  *	This product includes software developed by David Young.
19  * 4. The name of David Young may not be used to endorse or promote
20  *    products derived from this software without specific prior
21  *    written permission.
22  *
23  * THIS SOFTWARE IS PROVIDED BY DAVID YOUNG ``AS IS'' AND ANY
24  * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO,
25  * THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A
26  * PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
27  * FOUNDATION OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
28  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
29  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
31  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
32  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
33  * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
34  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
35  */
36 
37 #include <sys/cdefs.h>
38 __KERNEL_RCSID(0, "$NetBSD: in_selsrc.c,v 1.6 2007/12/04 10:33:11 dyoung Exp $");
39 
40 #include "opt_inet.h"
41 #include "opt_inet_conf.h"
42 
43 #include <lib/libkern/libkern.h>
44 
45 #include <sys/param.h>
46 #include <sys/ioctl.h>
47 #include <sys/errno.h>
48 #include <sys/malloc.h>
49 #include <sys/socket.h>
50 #include <sys/socketvar.h>
51 #include <sys/sysctl.h>
52 #include <sys/systm.h>
53 #include <sys/proc.h>
54 #include <sys/syslog.h>
55 
56 #include <net/if.h>
57 
58 #include <net/if_ether.h>
59 
60 #include <netinet/in_systm.h>
61 #include <netinet/in.h>
62 #include <netinet/in_var.h>
63 #include <netinet/ip.h>
64 #include <netinet/ip_var.h>
65 #include <netinet/in_ifattach.h>
66 #include <netinet/in_pcb.h>
67 #include <netinet/if_inarp.h>
68 #include <netinet/ip_mroute.h>
69 #include <netinet/igmp_var.h>
70 #include <netinet/in_selsrc.h>
71 
72 #ifdef INET
73 struct score_src_name {
74 	const char		*sn_name;
75 	const in_score_src_t	sn_score_src;
76 };
77 
78 static const struct sysctlnode *in_domifattach_sysctl(struct in_ifsysctl *);
79 static int in_preference(const struct in_addr *, int, int,
80     const struct in_addr *);
81 static int in_index(const struct in_addr *, int, int, const struct in_addr *);
82 static int in_matchlen(const struct in_addr *, int, int,
83     const struct in_addr *);
84 static int in_match_category(const struct in_addr *, int, int,
85     const struct in_addr *);
86 static size_t in_get_selectsrc(const struct in_ifselsrc *, char *,
87     const size_t);
88 static int in_set_selectsrc(struct in_ifselsrc *, char *buf);
89 static int in_sysctl_selectsrc(SYSCTLFN_PROTO);
90 static in_score_src_t name_to_score_src(const char *);
91 static const char *score_src_to_name(const in_score_src_t);
92 static void in_score(const in_score_src_t *, int *, int *,
93     const struct in_addr *, int, int, const struct in_addr *);
94 
95 static const struct score_src_name score_src_names[] = {
96 	  {"same-category", in_match_category}
97 	, {"common-prefix-len", in_matchlen}
98 	, {"index", in_index}
99 	, {"preference", in_preference}
100 	, {NULL, NULL}
101 };
102 
103 static const struct in_ifselsrc initial_iss = { 0, {NULL} };
104 
105 static struct in_ifselsrc default_iss = { 0, {in_index} };
106 
107 #ifdef GETIFA_DEBUG
108 int in_selsrc_debug = 0;
109 #endif /* GETIFA_DEBUG */
110 
111 SYSCTL_SETUP(sysctl_selectsrc_setup, "sysctl selectsrc subtree setup")
112 {
113 	int rc;
114 	const struct sysctlnode *rnode, *cnode;
115 
116 	if ((rc = sysctl_createv(clog, 0, NULL, &rnode,
117 	    CTLFLAG_PERMANENT, CTLTYPE_NODE, "net",
118 	    NULL, NULL, 0, NULL, 0, CTL_NET, CTL_EOL)) != 0) {
119 		printf("%s: could not create net, rc = %d\n", __func__, rc);
120 		return;
121 	}
122 	if ((rc = sysctl_createv(clog, 0, NULL, &rnode,
123 	    CTLFLAG_PERMANENT, CTLTYPE_NODE, "inet",
124 	    NULL, NULL, 0, NULL, 0, CTL_NET, PF_INET, CTL_EOL)) != 0) {
125 		printf("%s: could not create net.inet, rc = %d\n", __func__,
126 		    rc);
127 		return;
128 	}
129 	if ((rc = sysctl_createv(clog, 0, NULL, &rnode,
130 	    CTLFLAG_PERMANENT, CTLTYPE_NODE, "ip",
131 	    NULL, NULL, 0, NULL, 0,
132 	    CTL_NET, PF_INET, IPPROTO_IP, CTL_EOL)) != 0) {
133 		printf("%s: could not create net.inet.ip, rc = %d\n", __func__,
134 		    rc);
135 		return;
136 	}
137 	if ((rc = sysctl_createv(clog, 0, NULL, &rnode,
138 	    CTLFLAG_PERMANENT, CTLTYPE_NODE, "selectsrc",
139 	    NULL, NULL, 0, NULL, 0,
140 	    CTL_NET, PF_INET, IPPROTO_IP, CTL_CREATE, CTL_EOL)) != 0) {
141 		printf("%s: could not create net.inet.ip.selectsrc, "
142 		       "rc = %d\n", __func__, rc);
143 		return;
144 	}
145 #ifdef GETIFA_DEBUG
146 	if ((rc = sysctl_createv(clog, 0, &rnode, &cnode,
147 	    CTLFLAG_PERMANENT|CTLFLAG_READWRITE, CTLTYPE_INT, "debug",
148 	    SYSCTL_DESCR("enable source-selection debug messages"),
149 	    NULL, 0, &in_selsrc_debug, 0, CTL_CREATE, CTL_EOL)) != 0) {
150 		printf("%s: could not create net.inet.ip.selectsrc.debug, "
151 		       "rc = %d\n", __func__, rc);
152 		return;
153 	}
154 #endif /* GETIFA_DEBUG */
155 	if ((rc = sysctl_createv(clog, 0, &rnode, &cnode,
156 	    CTLFLAG_READWRITE, CTLTYPE_STRING, "default",
157 	    SYSCTL_DESCR("default source selection policy"),
158 	    in_sysctl_selectsrc, 0, &default_iss, IN_SELECTSRC_LEN,
159 	    CTL_CREATE, CTL_EOL)) != 0) {
160 		printf(
161 		    "%s: could not create net.inet.ip.selectsrc.default (%d)\n",
162 		    __func__, rc);
163 		return;
164 	}
165 }
166 
167 /*
168  * Score by address preference: prefer addresses with higher preference
169  * number.  Preference numbers are assigned with ioctl SIOCSIFADDRPREF.
170  */
171 static int
172 in_preference(const struct in_addr *src, int preference,
173     int idx, const struct in_addr *dst)
174 {
175 	return preference;
176 }
177 
178 /*
179  * Score by address "index": prefer addresses nearer the head of
180  * the ifaddr list.
181  */
182 static int
183 in_index(const struct in_addr *src, int preference, int idx,
184     const struct in_addr *dst)
185 {
186 	return -idx;
187 }
188 
189 /*
190  * Length of longest common prefix of src and dst.
191  *
192  * (Derived from in6_matchlen.)
193  */
194 static int
195 in_matchlen(const struct in_addr *src, int preference,
196     int idx, const struct in_addr *dst)
197 {
198 	int match = 0;
199 	const uint8_t *s = (const uint8_t *)src, *d = (const uint8_t *)dst;
200 	const uint8_t *lim = s + 4;
201 	uint_fast8_t r = 0;
202 
203 	while (s < lim && (r = (*d++ ^ *s++)) == 0)
204 		match += 8;
205 
206 	if (s == lim)
207 		return match;
208 
209 	while ((r & 0x80) == 0) {
210 		match++;
211 		r <<= 1;
212 	}
213 	return match;
214 }
215 
216 static enum in_category
217 in_categorize(const struct in_addr *s)
218 {
219 	if (IN_ANY_LOCAL(s->s_addr))
220 		return IN_CATEGORY_LINKLOCAL;
221 	else if (IN_PRIVATE(s->s_addr))
222 		return IN_CATEGORY_PRIVATE;
223 	else
224 		return IN_CATEGORY_OTHER;
225 }
226 
227 static int
228 in_match_category(const struct in_addr *src, int preference,
229     int idx, const struct in_addr *dst)
230 {
231 	enum in_category dst_c = in_categorize(dst),
232 	                 src_c = in_categorize(src);
233 #ifdef GETIFA_DEBUG
234 	if (in_selsrc_debug) {
235 		printf("%s: dst %#08" PRIx32 " categ %d, src %#08" PRIx32
236 		    " categ %d\n", __func__, ntohl(dst->s_addr), dst_c,
237 		    ntohl(src->s_addr), src_c);
238 	}
239 #endif /* GETIFA_DEBUG */
240 
241 	if (dst_c == src_c)
242 		return 2;
243 	else if (dst_c == IN_CATEGORY_LINKLOCAL && src_c == IN_CATEGORY_PRIVATE)
244 		return 1;
245 	else if (dst_c == IN_CATEGORY_PRIVATE && src_c == IN_CATEGORY_LINKLOCAL)
246 		return 1;
247 	else if (dst_c == IN_CATEGORY_OTHER && src_c == IN_CATEGORY_PRIVATE)
248 		return 1;
249 	else
250 		return 0;
251 }
252 
253 static void
254 in_score(const in_score_src_t *score_src, int *score, int *scorelenp,
255     const struct in_addr *src, int preference, int idx,
256     const struct in_addr *dst)
257 {
258 	int i;
259 
260 	for (i = 0; i < IN_SCORE_SRC_MAX && score_src[i] != NULL; i++)
261 		score[i] = (*score_src[i])(src, preference, idx, dst);
262 	if (scorelenp != NULL)
263 		*scorelenp = i;
264 }
265 
266 static int
267 in_score_cmp(int *score1, int *score2, int scorelen)
268 {
269 	int i;
270 
271 	for (i = 0; i < scorelen; i++) {
272 		if (score1[i] == score2[i])
273 			continue;
274 		return score1[i] - score2[i];
275 	}
276 	return 0;
277 }
278 
279 #ifdef GETIFA_DEBUG
280 static void
281 in_score_println(int *score, int scorelen)
282 {
283 	int i;
284 	const char *delim = "[";
285 
286 	for (i = 0; i < scorelen; i++) {
287 		printf("%s%d", delim, score[i]);
288 		delim = ", ";
289 	}
290 	printf("]\n");
291 }
292 #endif /* GETIFA_DEBUG */
293 
294 /* Scan the interface addresses on the interface ifa->ifa_ifp for
295  * the source address that best matches the destination, dst0,
296  * according to the source address-selection policy for this
297  * interface.  If there is no better match than `ifa', return `ifa'.
298  * Otherwise, return the best address.
299  *
300  * Note that in_getifa is called after the kernel has decided which
301  * output interface to use (ifa->ifa_ifp), and in_getifa will not
302  * scan an address belonging to any other interface.
303  */
304 struct ifaddr *
305 in_getifa(struct ifaddr *ifa, const struct sockaddr *dst0)
306 {
307 	const in_score_src_t *score_src;
308 	int idx, scorelen;
309 	const struct sockaddr_in *dst, *src;
310 	struct ifaddr *alt_ifa, *best_ifa;
311 	struct ifnet *ifp;
312 	struct in_ifsysctl *isc;
313 	struct in_ifselsrc *iss;
314 	int best_score[IN_SCORE_SRC_MAX], score[IN_SCORE_SRC_MAX];
315 
316 	if (ifa->ifa_addr->sa_family != AF_INET ||
317 	    dst0 == NULL || dst0->sa_family != AF_INET) {	/* Possible. */
318 		ifa->ifa_seqno = NULL;
319 		return ifa;
320 	}
321 
322 	ifp = ifa->ifa_ifp;
323 	isc = (struct in_ifsysctl *)ifp->if_afdata[AF_INET];
324 	if (isc != NULL && isc->isc_selsrc != NULL &&
325 	    isc->isc_selsrc->iss_score_src[0] != NULL)
326 		iss = isc->isc_selsrc;
327 	else
328 		iss = &default_iss;
329 	score_src = &iss->iss_score_src[0];
330 
331 	dst = (const struct sockaddr_in *)dst0;
332 
333 	best_ifa = ifa;
334 
335 	/* Find out the index of this ifaddr. */
336 	idx = 0;
337 	IFADDR_FOREACH(alt_ifa, ifa->ifa_ifp) {
338 		if (alt_ifa == best_ifa)
339 			break;
340 		idx++;
341 	}
342 	in_score(score_src, best_score, &scorelen, &IA_SIN(best_ifa)->sin_addr,
343 	    best_ifa->ifa_preference, idx, &dst->sin_addr);
344 
345 #ifdef GETIFA_DEBUG
346 	if (in_selsrc_debug) {
347 		printf("%s: enter dst %#" PRIx32 " src %#" PRIx32 " score ",
348 		    __func__, ntohl(dst->sin_addr.s_addr),
349 		    ntohl(satosin(best_ifa->ifa_addr)->sin_addr.s_addr));
350 		in_score_println(best_score, scorelen);
351 	}
352 #endif /* GETIFA_DEBUG */
353 
354 	idx = -1;
355 	IFADDR_FOREACH(alt_ifa, ifa->ifa_ifp) {
356 		++idx;
357 		src = IA_SIN(alt_ifa);
358 
359 		if (alt_ifa == ifa || src->sin_family != AF_INET)
360 			continue;
361 
362 		in_score(score_src, score, NULL, &src->sin_addr,
363 		         alt_ifa->ifa_preference, idx, &dst->sin_addr);
364 
365 #ifdef GETIFA_DEBUG
366 		if (in_selsrc_debug) {
367 			printf("%s: src %#" PRIx32 " score ", __func__,
368 			    ntohl(src->sin_addr.s_addr));
369 			in_score_println(score, scorelen);
370 		}
371 #endif /* GETIFA_DEBUG */
372 
373 		if (in_score_cmp(score, best_score, scorelen) > 0) {
374 			(void)memcpy(best_score, score, sizeof(best_score));
375 			best_ifa = alt_ifa;
376 		}
377 	}
378 #ifdef GETIFA_DEBUG
379 	if (in_selsrc_debug) {
380 		printf("%s: choose src %#" PRIx32 " score ", __func__,
381 		    ntohl(IA_SIN(best_ifa)->sin_addr.s_addr));
382 		in_score_println(best_score, scorelen);
383 	}
384 #endif /* GETIFA_DEBUG */
385 
386 	best_ifa->ifa_seqno = &iss->iss_seqno;
387 	return best_ifa;
388 }
389 
390 static in_score_src_t
391 name_to_score_src(const char *name)
392 {
393 	int i;
394 
395 	for (i = 0; score_src_names[i].sn_name != NULL; i++) {
396 		if (strcmp(score_src_names[i].sn_name, name) == 0)
397 			return score_src_names[i].sn_score_src;
398 	}
399 	return NULL;
400 }
401 
402 static const char *
403 score_src_to_name(const in_score_src_t score_src)
404 {
405 	int i;
406 	for (i = 0; score_src_names[i].sn_name != NULL; i++) {
407 		if (score_src == score_src_names[i].sn_score_src)
408 			return score_src_names[i].sn_name;
409 	}
410 	return "<unknown>";
411 }
412 
413 static size_t
414 in_get_selectsrc(const struct in_ifselsrc *iss, char *buf0,
415     const size_t buflen0)
416 {
417 	int i, rc;
418 	char *buf = buf0;
419 	const char *delim;
420 	size_t buflen = buflen0;
421 
422 	KASSERT(buflen >= 1);
423 
424 	for (delim = "", i = 0;
425 	     i < IN_SCORE_SRC_MAX && iss->iss_score_src[i] != NULL;
426 	     delim = ",", i++) {
427 		rc = snprintf(buf, buflen, "%s%s",
428 		    delim, score_src_to_name(iss->iss_score_src[i]));
429 		if (rc == -1)
430 			return buflen0 - buflen;
431 		if (rc >= buflen)
432 			return buflen0 + rc - buflen;
433 		buf += rc;
434 		buflen -= rc;
435 	}
436 	if (buf == buf0)
437 		*buf++ = '\0';
438 	return buf - buf0;
439 }
440 
441 static int
442 in_set_selectsrc(struct in_ifselsrc *iss, char *buf)
443 {
444 	int i, s;
445 	char *next = buf;
446 	const char *name;
447 	in_score_src_t score_src;
448 	in_score_src_t scorers[IN_SCORE_SRC_MAX];
449 
450 	memset(&scorers, 0, sizeof(scorers));
451 	for (i = 0;
452 	     (name = strsep(&next, ",")) != NULL && i < IN_SCORE_SRC_MAX;
453 	     i++) {
454 		if (strcmp(name, "") == 0)
455 			break;
456 		if ((score_src = name_to_score_src(name)) == NULL)
457 			return EINVAL;
458 		scorers[i] = score_src;
459 	}
460 	if (i == IN_SCORE_SRC_MAX && name != NULL)
461 		return EFBIG;
462 	s = splnet();
463 	(void)memcpy(iss->iss_score_src, scorers, sizeof(iss->iss_score_src));
464         /* If iss affects a specific interface that used to use
465          * the default policy, increase the sequence number on the
466          * default policy, forcing routes that cache a source
467          * (rt_ifa) found by the default policy to refresh their
468          * cache.
469 	 */
470 	if (iss != &default_iss && iss->iss_score_src[0] == NULL &&
471 	    scorers[0] != NULL)
472 		default_iss.iss_seqno++;
473 	iss->iss_seqno++;
474 	splx(s);
475 	return 0;
476 }
477 
478 /*
479  * sysctl helper routine for net.inet.ip.interfaces.<interface>.selectsrc.
480  * Pulls the old value out as a human-readable string, interprets
481  * and records the new value.
482  */
483 static int
484 in_sysctl_selectsrc(SYSCTLFN_ARGS)
485 {
486 	char policy[IN_SELECTSRC_LEN];
487 	int error;
488 	struct sysctlnode node;
489 	struct in_ifselsrc *iss;
490 
491 	node = *rnode;
492 	iss = (struct in_ifselsrc *)node.sysctl_data;
493 	if (oldp != NULL &&
494 	    (error = in_get_selectsrc(iss, policy, sizeof(policy))) >= sizeof(policy))
495 		return error;
496 	node.sysctl_data = &policy[0];
497 	error = sysctl_lookup(SYSCTLFN_CALL(&node));
498 	if (error || newp == NULL)
499 		return (error);
500 
501 	return in_set_selectsrc(iss, policy);
502 }
503 
504 static const struct sysctlnode *
505 in_domifattach_sysctl(struct in_ifsysctl *isc)
506 {
507 	int rc;
508 	const struct sysctlnode *rnode;
509 
510 	if ((rc = sysctl_createv(&isc->isc_log, 0, NULL, &rnode,
511 	                         CTLFLAG_READWRITE, CTLTYPE_NODE,
512 				 "interfaces", NULL,
513 				 NULL, 0, NULL, 0,
514 				 CTL_NET, PF_INET, IPPROTO_IP, CTL_CREATE,
515 				 CTL_EOL)) != 0) {
516 		printf("%s: could not create net.inet.ip.interfaces, rc = %d\n",
517 		    __func__, rc);
518 		return NULL;
519 	}
520 	if ((rc = sysctl_createv(&isc->isc_log, 0, &rnode, &rnode,
521 	                         CTLFLAG_READWRITE, CTLTYPE_NODE,
522 				 isc->isc_ifp->if_xname,
523 				 SYSCTL_DESCR("interface ip options"),
524 				 NULL, 0, NULL, 0, CTL_CREATE, CTL_EOL)) != 0) {
525 		printf("%s: could not create net.inet.ip.interfaces.%s, "
526 		       "rc = %d\n", __func__, isc->isc_ifp->if_xname, rc);
527 		goto err;
528 	}
529 	if ((rc = sysctl_createv(&isc->isc_log, 0, &rnode, &rnode,
530 	                         CTLFLAG_READWRITE, CTLTYPE_STRING,
531 				 "selectsrc",
532 				 SYSCTL_DESCR("source selection policy"),
533 				 in_sysctl_selectsrc, 0,
534 				 isc->isc_selsrc, IN_SELECTSRC_LEN,
535 				 CTL_CREATE, CTL_EOL)) != 0) {
536 		printf(
537 		    "%s: could not create net.inet.ip.%s.selectsrc, rc = %d\n",
538 		    __func__, isc->isc_ifp->if_xname, rc);
539 		goto err;
540 	}
541 	return rnode;
542 err:
543 	sysctl_teardown(&isc->isc_log);
544 	return NULL;
545 }
546 
547 void *
548 in_domifattach(struct ifnet *ifp)
549 {
550 	struct in_ifsysctl *isc;
551 	struct in_ifselsrc *iss;
552 
553 	isc = (struct in_ifsysctl *)malloc(sizeof(*isc), M_IFADDR,
554 	    M_WAITOK | M_ZERO);
555 
556 	iss = (struct in_ifselsrc *)malloc(sizeof(*iss), M_IFADDR,
557 	    M_WAITOK | M_ZERO);
558 
559 	memcpy(&iss->iss_score_src[0], &initial_iss.iss_score_src[0],
560 	    MIN(sizeof(iss->iss_score_src), sizeof(initial_iss.iss_score_src)));
561 
562 	isc->isc_ifp = ifp;
563 	isc->isc_selsrc = iss;
564 
565 	if (in_domifattach_sysctl(isc) == NULL)
566 		goto err;
567 
568 	return isc;
569 err:
570 	free(iss, M_IFADDR);
571 	free(isc, M_IFADDR);
572 	return NULL;
573 }
574 
575 void
576 in_domifdetach(struct ifnet *ifp, void *aux)
577 {
578 	struct in_ifsysctl *isc;
579 	struct in_ifselsrc *iss;
580 
581 	if (aux == NULL)
582 		return;
583 	isc = (struct in_ifsysctl *)aux;
584 	iss = isc->isc_selsrc;
585 	sysctl_teardown(&isc->isc_log);
586 	free(isc, M_IFADDR);
587 	free(iss, M_IFADDR);
588 }
589 #endif /* INET */
590