xref: /netbsd-src/sys/netinet/ip_flow.c (revision 267197ec1eebfcb9810ea27a89625b6ddf68e3e7) 
1 /*	$NetBSD: ip_flow.c,v 1.51 2008/01/04 23:28:07 dyoung Exp $	*/
2 
3 /*-
4  * Copyright (c) 1998 The NetBSD Foundation, Inc.
5  * All rights reserved.
6  *
7  * This code is derived from software contributed to The NetBSD Foundation
8  * by the 3am Software Foundry ("3am").  It was developed by Matt Thomas.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 3. All advertising materials mentioning features or use of this software
19  *    must display the following acknowledgement:
20  *	This product includes software developed by the NetBSD
21  *	Foundation, Inc. and its contributors.
22  * 4. Neither the name of The NetBSD Foundation nor the names of its
23  *    contributors may be used to endorse or promote products derived
24  *    from this software without specific prior written permission.
25  *
26  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
27  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
28  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
29  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
30  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
31  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
32  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
33  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
34  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
35  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
36  * POSSIBILITY OF SUCH DAMAGE.
37  */
38 
39 #include <sys/cdefs.h>
40 __KERNEL_RCSID(0, "$NetBSD: ip_flow.c,v 1.51 2008/01/04 23:28:07 dyoung Exp $");
41 
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/malloc.h>
45 #include <sys/mbuf.h>
46 #include <sys/domain.h>
47 #include <sys/protosw.h>
48 #include <sys/socket.h>
49 #include <sys/socketvar.h>
50 #include <sys/errno.h>
51 #include <sys/time.h>
52 #include <sys/kernel.h>
53 #include <sys/pool.h>
54 #include <sys/sysctl.h>
55 
56 #include <net/if.h>
57 #include <net/if_dl.h>
58 #include <net/route.h>
59 #include <net/pfil.h>
60 
61 #include <netinet/in.h>
62 #include <netinet/in_systm.h>
63 #include <netinet/ip.h>
64 #include <netinet/in_pcb.h>
65 #include <netinet/in_var.h>
66 #include <netinet/ip_var.h>
67 
68 /*
69  * Similar code is very well commented in netinet6/ip6_flow.c
70  */
71 
72 POOL_INIT(ipflow_pool, sizeof(struct ipflow), 0, 0, 0, "ipflowpl", NULL,
73     IPL_NET);
74 
75 LIST_HEAD(ipflowhead, ipflow);
76 
77 #define	IPFLOW_TIMER		(5 * PR_SLOWHZ)
78 #define	IPFLOW_DEFAULT_HASHSIZE	(1 << IPFLOW_HASHBITS)
79 
80 static struct ipflowhead *ipflowtable = NULL;
81 static struct ipflowhead ipflowlist;
82 static int ipflow_inuse;
83 
84 #define	IPFLOW_INSERT(bucket, ipf) \
85 do { \
86 	LIST_INSERT_HEAD((bucket), (ipf), ipf_hash); \
87 	LIST_INSERT_HEAD(&ipflowlist, (ipf), ipf_list); \
88 } while (/*CONSTCOND*/ 0)
89 
90 #define	IPFLOW_REMOVE(ipf) \
91 do { \
92 	LIST_REMOVE((ipf), ipf_hash); \
93 	LIST_REMOVE((ipf), ipf_list); \
94 } while (/*CONSTCOND*/ 0)
95 
96 #ifndef IPFLOW_MAX
97 #define	IPFLOW_MAX		256
98 #endif
99 int ip_maxflows = IPFLOW_MAX;
100 int ip_hashsize = IPFLOW_DEFAULT_HASHSIZE;
101 
102 static size_t
103 ipflow_hash(const struct ip *ip)
104 {
105 	size_t hash = ip->ip_tos;
106 	size_t idx;
107 
108 	for (idx = 0; idx < 32; idx += IPFLOW_HASHBITS) {
109 		hash += (ip->ip_dst.s_addr >> (32 - idx)) +
110 		    (ip->ip_src.s_addr >> idx);
111 	}
112 
113 	return hash & (ip_hashsize-1);
114 }
115 
116 static struct ipflow *
117 ipflow_lookup(const struct ip *ip)
118 {
119 	size_t hash;
120 	struct ipflow *ipf;
121 
122 	hash = ipflow_hash(ip);
123 
124 	LIST_FOREACH(ipf, &ipflowtable[hash], ipf_hash) {
125 		if (ip->ip_dst.s_addr == ipf->ipf_dst.s_addr
126 		    && ip->ip_src.s_addr == ipf->ipf_src.s_addr
127 		    && ip->ip_tos == ipf->ipf_tos)
128 			break;
129 	}
130 	return ipf;
131 }
132 
133 int
134 ipflow_init(int table_size)
135 {
136 	struct ipflowhead *new_table;
137 	size_t i;
138 
139 	new_table = (struct ipflowhead *)malloc(sizeof(struct ipflowhead) *
140 	    table_size, M_RTABLE, M_NOWAIT);
141 
142 	if (new_table == NULL)
143 		return 1;
144 
145 	if (ipflowtable != NULL)
146 		free(ipflowtable, M_RTABLE);
147 
148 	ipflowtable = new_table;
149 	ip_hashsize = table_size;
150 
151 	LIST_INIT(&ipflowlist);
152 	for (i = 0; i < ip_hashsize; i++)
153 		LIST_INIT(&ipflowtable[i]);
154 
155 	return 0;
156 }
157 
158 int
159 ipflow_fastforward(struct mbuf *m)
160 {
161 	struct ip *ip;
162 	struct ip ip_store;
163 	struct ipflow *ipf;
164 	struct rtentry *rt;
165 	const struct sockaddr *dst;
166 	int error;
167 	int iplen;
168 
169 	/*
170 	 * Are we forwarding packets?  Big enough for an IP packet?
171 	 */
172 	if (!ipforwarding || ipflow_inuse == 0 || m->m_len < sizeof(struct ip))
173 		return 0;
174 
175 	/*
176 	 * Was packet received as a link-level multicast or broadcast?
177 	 * If so, don't try to fast forward..
178 	 */
179 	if ((m->m_flags & (M_BCAST|M_MCAST)) != 0)
180 		return 0;
181 
182 	/*
183 	 * IP header with no option and valid version and length
184 	 */
185 	if (IP_HDR_ALIGNED_P(mtod(m, const void *)))
186 		ip = mtod(m, struct ip *);
187 	else {
188 		memcpy(&ip_store, mtod(m, const void *), sizeof(ip_store));
189 		ip = &ip_store;
190 	}
191 	iplen = ntohs(ip->ip_len);
192 	if (ip->ip_v != IPVERSION || ip->ip_hl != (sizeof(struct ip) >> 2) ||
193 	    iplen < sizeof(struct ip) || iplen > m->m_pkthdr.len)
194 		return 0;
195 	/*
196 	 * Find a flow.
197 	 */
198 	if ((ipf = ipflow_lookup(ip)) == NULL)
199 		return 0;
200 
201 	/*
202 	 * Verify the IP header checksum.
203 	 */
204 	switch (m->m_pkthdr.csum_flags &
205 		((m->m_pkthdr.rcvif->if_csum_flags_rx & M_CSUM_IPv4) |
206 		 M_CSUM_IPv4_BAD)) {
207 	case M_CSUM_IPv4|M_CSUM_IPv4_BAD:
208 		return (0);
209 
210 	case M_CSUM_IPv4:
211 		/* Checksum was okay. */
212 		break;
213 
214 	default:
215 		/* Must compute it ourselves. */
216 		if (in_cksum(m, sizeof(struct ip)) != 0)
217 			return (0);
218 		break;
219 	}
220 
221 	/*
222 	 * Route and interface still up?
223 	 */
224 	if ((rt = rtcache_validate(&ipf->ipf_ro)) == NULL ||
225 	    (rt->rt_ifp->if_flags & IFF_UP) == 0)
226 		return 0;
227 
228 	/*
229 	 * Packet size OK?  TTL?
230 	 */
231 	if (m->m_pkthdr.len > rt->rt_ifp->if_mtu || ip->ip_ttl <= IPTTLDEC)
232 		return 0;
233 
234 	/*
235 	 * Clear any in-bound checksum flags for this packet.
236 	 */
237 	m->m_pkthdr.csum_flags = 0;
238 
239 	/*
240 	 * Everything checks out and so we can forward this packet.
241 	 * Modify the TTL and incrementally change the checksum.
242 	 *
243 	 * This method of adding the checksum works on either endian CPU.
244 	 * If htons() is inlined, all the arithmetic is folded; otherwise
245 	 * the htons()s are combined by CSE due to the const attribute.
246 	 *
247 	 * Don't bother using HW checksumming here -- the incremental
248 	 * update is pretty fast.
249 	 */
250 	ip->ip_ttl -= IPTTLDEC;
251 	if (ip->ip_sum >= (u_int16_t) ~htons(IPTTLDEC << 8))
252 		ip->ip_sum -= ~htons(IPTTLDEC << 8);
253 	else
254 		ip->ip_sum += htons(IPTTLDEC << 8);
255 
256 	/*
257 	 * Done modifying the header; copy it back, if necessary.
258 	 *
259 	 * XXX Use m_copyback_cow(9) here? --dyoung
260 	 */
261 	if (IP_HDR_ALIGNED_P(mtod(m, void *)) == 0)
262 		memcpy(mtod(m, void *), &ip_store, sizeof(ip_store));
263 
264 	/*
265 	 * Trim the packet in case it's too long..
266 	 */
267 	if (m->m_pkthdr.len > iplen) {
268 		if (m->m_len == m->m_pkthdr.len) {
269 			m->m_len = iplen;
270 			m->m_pkthdr.len = iplen;
271 		} else
272 			m_adj(m, iplen - m->m_pkthdr.len);
273 	}
274 
275 	/*
276 	 * Send the packet on it's way.  All we can get back is ENOBUFS
277 	 */
278 	ipf->ipf_uses++;
279 	PRT_SLOW_ARM(ipf->ipf_timer, IPFLOW_TIMER);
280 
281 	if (rt->rt_flags & RTF_GATEWAY)
282 		dst = rt->rt_gateway;
283 	else
284 		dst = rtcache_getdst(&ipf->ipf_ro);
285 
286 	if ((error = (*rt->rt_ifp->if_output)(rt->rt_ifp, m, dst, rt)) != 0) {
287 		if (error == ENOBUFS)
288 			ipf->ipf_dropped++;
289 		else
290 			ipf->ipf_errors++;
291 	}
292 	return 1;
293 }
294 
295 static void
296 ipflow_addstats(struct ipflow *ipf)
297 {
298 	struct rtentry *rt;
299 
300 	if ((rt = rtcache_validate(&ipf->ipf_ro)) != NULL)
301 		rt->rt_use += ipf->ipf_uses;
302 	ipstat.ips_cantforward += ipf->ipf_errors + ipf->ipf_dropped;
303 	ipstat.ips_total += ipf->ipf_uses;
304 	ipstat.ips_forward += ipf->ipf_uses;
305 	ipstat.ips_fastforward += ipf->ipf_uses;
306 }
307 
308 static void
309 ipflow_free(struct ipflow *ipf)
310 {
311 	int s;
312 	/*
313 	 * Remove the flow from the hash table (at elevated IPL).
314 	 * Once it's off the list, we can deal with it at normal
315 	 * network IPL.
316 	 */
317 	s = splnet();
318 	IPFLOW_REMOVE(ipf);
319 	splx(s);
320 	ipflow_addstats(ipf);
321 	rtcache_free(&ipf->ipf_ro);
322 	ipflow_inuse--;
323 	s = splnet();
324 	pool_put(&ipflow_pool, ipf);
325 	splx(s);
326 }
327 
328 struct ipflow *
329 ipflow_reap(int just_one)
330 {
331 	while (just_one || ipflow_inuse > ip_maxflows) {
332 		struct ipflow *ipf, *maybe_ipf = NULL;
333 		int s;
334 
335 		ipf = LIST_FIRST(&ipflowlist);
336 		while (ipf != NULL) {
337 			/*
338 			 * If this no longer points to a valid route
339 			 * reclaim it.
340 			 */
341 			if (rtcache_validate(&ipf->ipf_ro) == NULL)
342 				goto done;
343 			/*
344 			 * choose the one that's been least recently
345 			 * used or has had the least uses in the
346 			 * last 1.5 intervals.
347 			 */
348 			if (maybe_ipf == NULL ||
349 			    ipf->ipf_timer < maybe_ipf->ipf_timer ||
350 			    (ipf->ipf_timer == maybe_ipf->ipf_timer &&
351 			     ipf->ipf_last_uses + ipf->ipf_uses <
352 			         maybe_ipf->ipf_last_uses +
353 			         maybe_ipf->ipf_uses))
354 				maybe_ipf = ipf;
355 			ipf = LIST_NEXT(ipf, ipf_list);
356 		}
357 		ipf = maybe_ipf;
358 	    done:
359 		/*
360 		 * Remove the entry from the flow table.
361 		 */
362 		s = splnet();
363 		IPFLOW_REMOVE(ipf);
364 		splx(s);
365 		ipflow_addstats(ipf);
366 		rtcache_free(&ipf->ipf_ro);
367 		if (just_one)
368 			return ipf;
369 		pool_put(&ipflow_pool, ipf);
370 		ipflow_inuse--;
371 	}
372 	return NULL;
373 }
374 
375 void
376 ipflow_slowtimo(void)
377 {
378 	struct rtentry *rt;
379 	struct ipflow *ipf, *next_ipf;
380 
381 	for (ipf = LIST_FIRST(&ipflowlist); ipf != NULL; ipf = next_ipf) {
382 		next_ipf = LIST_NEXT(ipf, ipf_list);
383 		if (PRT_SLOW_ISEXPIRED(ipf->ipf_timer) ||
384 		    (rt = rtcache_validate(&ipf->ipf_ro)) == NULL) {
385 			ipflow_free(ipf);
386 		} else {
387 			ipf->ipf_last_uses = ipf->ipf_uses;
388 			rt->rt_use += ipf->ipf_uses;
389 			ipstat.ips_total += ipf->ipf_uses;
390 			ipstat.ips_forward += ipf->ipf_uses;
391 			ipstat.ips_fastforward += ipf->ipf_uses;
392 			ipf->ipf_uses = 0;
393 		}
394 	}
395 }
396 
397 void
398 ipflow_create(const struct route *ro, struct mbuf *m)
399 {
400 	const struct ip *const ip = mtod(m, const struct ip *);
401 	struct ipflow *ipf;
402 	size_t hash;
403 	int s;
404 
405 	/*
406 	 * Don't create cache entries for ICMP messages.
407 	 */
408 	if (ip_maxflows == 0 || ip->ip_p == IPPROTO_ICMP)
409 		return;
410 	/*
411 	 * See if an existing flow struct exists.  If so remove it from it's
412 	 * list and free the old route.  If not, try to malloc a new one
413 	 * (if we aren't at our limit).
414 	 */
415 	ipf = ipflow_lookup(ip);
416 	if (ipf == NULL) {
417 		if (ipflow_inuse >= ip_maxflows) {
418 			ipf = ipflow_reap(1);
419 		} else {
420 			s = splnet();
421 			ipf = pool_get(&ipflow_pool, PR_NOWAIT);
422 			splx(s);
423 			if (ipf == NULL)
424 				return;
425 			ipflow_inuse++;
426 		}
427 		memset(ipf, 0, sizeof(*ipf));
428 	} else {
429 		s = splnet();
430 		IPFLOW_REMOVE(ipf);
431 		splx(s);
432 		ipflow_addstats(ipf);
433 		rtcache_free(&ipf->ipf_ro);
434 		ipf->ipf_uses = ipf->ipf_last_uses = 0;
435 		ipf->ipf_errors = ipf->ipf_dropped = 0;
436 	}
437 
438 	/*
439 	 * Fill in the updated information.
440 	 */
441 	rtcache_copy(&ipf->ipf_ro, ro);
442 	ipf->ipf_dst = ip->ip_dst;
443 	ipf->ipf_src = ip->ip_src;
444 	ipf->ipf_tos = ip->ip_tos;
445 	PRT_SLOW_ARM(ipf->ipf_timer, IPFLOW_TIMER);
446 	ipf->ipf_start = time_uptime;
447 	/*
448 	 * Insert into the approriate bucket of the flow table.
449 	 */
450 	hash = ipflow_hash(ip);
451 	s = splnet();
452 	IPFLOW_INSERT(&ipflowtable[hash], ipf);
453 	splx(s);
454 }
455 
456 int
457 ipflow_invalidate_all(int new_size)
458 {
459 	struct ipflow *ipf, *next_ipf;
460 	int s, error;
461 
462 	error = 0;
463 	s = splnet();
464 	for (ipf = LIST_FIRST(&ipflowlist); ipf != NULL; ipf = next_ipf) {
465 		next_ipf = LIST_NEXT(ipf, ipf_list);
466 		ipflow_free(ipf);
467 	}
468 
469 	if (new_size)
470 		error = ipflow_init(new_size);
471 	splx(s);
472 
473 	return error;
474 }
475