xref: /netbsd-src/sys/netinet/ip_flow.c (revision 8b0f9554ff8762542c4defc4f70e1eb76fb508fa) 
1 /*	$NetBSD: ip_flow.c,v 1.48 2007/08/20 19:42:34 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 1.48 2007/08/20 19:42:34 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(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(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, ip_store;
162 	struct ipflow *ipf;
163 	struct rtentry *rt;
164 	const struct sockaddr *dst;
165 	int error;
166 	int iplen;
167 
168 	/*
169 	 * Are we forwarding packets?  Big enough for an IP packet?
170 	 */
171 	if (!ipforwarding || ipflow_inuse == 0 || m->m_len < sizeof(struct ip))
172 		return 0;
173 
174 	/*
175 	 * Was packet received as a link-level multicast or broadcast?
176 	 * If so, don't try to fast forward..
177 	 */
178 	if ((m->m_flags & (M_BCAST|M_MCAST)) != 0)
179 		return 0;
180 
181 	/*
182 	 * IP header with no option and valid version and length
183 	 */
184 	if (IP_HDR_ALIGNED_P(mtod(m, void *)))
185 		ip = mtod(m, struct ip *);
186 	else {
187 		memcpy(&ip_store, mtod(m, void *), sizeof(ip_store));
188 		ip = &ip_store;
189 	}
190 	iplen = ntohs(ip->ip_len);
191 	if (ip->ip_v != IPVERSION || ip->ip_hl != (sizeof(struct ip) >> 2) ||
192 	    iplen < sizeof(struct ip) || iplen > m->m_pkthdr.len)
193 		return 0;
194 	/*
195 	 * Find a flow.
196 	 */
197 	if ((ipf = ipflow_lookup(ip)) == NULL)
198 		return 0;
199 
200 	/*
201 	 * Verify the IP header checksum.
202 	 */
203 	switch (m->m_pkthdr.csum_flags &
204 		((m->m_pkthdr.rcvif->if_csum_flags_rx & M_CSUM_IPv4) |
205 		 M_CSUM_IPv4_BAD)) {
206 	case M_CSUM_IPv4|M_CSUM_IPv4_BAD:
207 		return (0);
208 
209 	case M_CSUM_IPv4:
210 		/* Checksum was okay. */
211 		break;
212 
213 	default:
214 		/* Must compute it ourselves. */
215 		if (in_cksum(m, sizeof(struct ip)) != 0)
216 			return (0);
217 		break;
218 	}
219 
220 	/*
221 	 * Route and interface still up?
222 	 */
223 	if (rtcache_down(&ipf->ipf_ro) || (rt = ipf->ipf_ro.ro_rt) == NULL ||
224 	    (rt->rt_ifp->if_flags & IFF_UP) == 0)
225 		return 0;
226 
227 	/*
228 	 * Packet size OK?  TTL?
229 	 */
230 	if (m->m_pkthdr.len > rt->rt_ifp->if_mtu || ip->ip_ttl <= IPTTLDEC)
231 		return 0;
232 
233 	/*
234 	 * Clear any in-bound checksum flags for this packet.
235 	 */
236 	m->m_pkthdr.csum_flags = 0;
237 
238 	/*
239 	 * Everything checks out and so we can forward this packet.
240 	 * Modify the TTL and incrementally change the checksum.
241 	 *
242 	 * This method of adding the checksum works on either endian CPU.
243 	 * If htons() is inlined, all the arithmetic is folded; otherwise
244 	 * the htons()s are combined by CSE due to the const attribute.
245 	 *
246 	 * Don't bother using HW checksumming here -- the incremental
247 	 * update is pretty fast.
248 	 */
249 	ip->ip_ttl -= IPTTLDEC;
250 	if (ip->ip_sum >= (u_int16_t) ~htons(IPTTLDEC << 8))
251 		ip->ip_sum -= ~htons(IPTTLDEC << 8);
252 	else
253 		ip->ip_sum += htons(IPTTLDEC << 8);
254 
255 	/*
256 	 * Done modifying the header; copy it back, if necessary.
257 	 */
258 	if (IP_HDR_ALIGNED_P(mtod(m, void *)) == 0)
259 		memcpy(mtod(m, void *), &ip_store, sizeof(ip_store));
260 
261 	/*
262 	 * Trim the packet in case it's too long..
263 	 */
264 	if (m->m_pkthdr.len > iplen) {
265 		if (m->m_len == m->m_pkthdr.len) {
266 			m->m_len = iplen;
267 			m->m_pkthdr.len = iplen;
268 		} else
269 			m_adj(m, iplen - m->m_pkthdr.len);
270 	}
271 
272 	/*
273 	 * Send the packet on it's way.  All we can get back is ENOBUFS
274 	 */
275 	ipf->ipf_uses++;
276 	PRT_SLOW_ARM(ipf->ipf_timer, IPFLOW_TIMER);
277 
278 	if (rt->rt_flags & RTF_GATEWAY)
279 		dst = rt->rt_gateway;
280 	else
281 		dst = rtcache_getdst(&ipf->ipf_ro);
282 
283 	if ((error = (*rt->rt_ifp->if_output)(rt->rt_ifp, m, dst, rt)) != 0) {
284 		if (error == ENOBUFS)
285 			ipf->ipf_dropped++;
286 		else
287 			ipf->ipf_errors++;
288 	}
289 	return 1;
290 }
291 
292 static void
293 ipflow_addstats(struct ipflow *ipf)
294 {
295 	if (!rtcache_down(&ipf->ipf_ro) && ipf->ipf_ro.ro_rt != NULL)
296 		ipf->ipf_ro.ro_rt->rt_use += ipf->ipf_uses;
297 	ipstat.ips_cantforward += ipf->ipf_errors + ipf->ipf_dropped;
298 	ipstat.ips_total += ipf->ipf_uses;
299 	ipstat.ips_forward += ipf->ipf_uses;
300 	ipstat.ips_fastforward += ipf->ipf_uses;
301 }
302 
303 static void
304 ipflow_free(struct ipflow *ipf)
305 {
306 	int s;
307 	/*
308 	 * Remove the flow from the hash table (at elevated IPL).
309 	 * Once it's off the list, we can deal with it at normal
310 	 * network IPL.
311 	 */
312 	s = splnet();
313 	IPFLOW_REMOVE(ipf);
314 	splx(s);
315 	ipflow_addstats(ipf);
316 	rtcache_free(&ipf->ipf_ro);
317 	ipflow_inuse--;
318 	s = splnet();
319 	pool_put(&ipflow_pool, ipf);
320 	splx(s);
321 }
322 
323 struct ipflow *
324 ipflow_reap(int just_one)
325 {
326 	while (just_one || ipflow_inuse > ip_maxflows) {
327 		struct ipflow *ipf, *maybe_ipf = NULL;
328 		int s;
329 
330 		ipf = LIST_FIRST(&ipflowlist);
331 		while (ipf != NULL) {
332 			/*
333 			 * If this no longer points to a valid route
334 			 * reclaim it.
335 			 */
336 			if (rtcache_down(&ipf->ipf_ro) ||
337 			    ipf->ipf_ro.ro_rt == NULL)
338 				goto done;
339 			/*
340 			 * choose the one that's been least recently
341 			 * used or has had the least uses in the
342 			 * last 1.5 intervals.
343 			 */
344 			if (maybe_ipf == NULL ||
345 			    ipf->ipf_timer < maybe_ipf->ipf_timer ||
346 			    (ipf->ipf_timer == maybe_ipf->ipf_timer &&
347 			     ipf->ipf_last_uses + ipf->ipf_uses <
348 			         maybe_ipf->ipf_last_uses +
349 			         maybe_ipf->ipf_uses))
350 				maybe_ipf = ipf;
351 			ipf = LIST_NEXT(ipf, ipf_list);
352 		}
353 		ipf = maybe_ipf;
354 	    done:
355 		/*
356 		 * Remove the entry from the flow table.
357 		 */
358 		s = splnet();
359 		IPFLOW_REMOVE(ipf);
360 		splx(s);
361 		ipflow_addstats(ipf);
362 		rtcache_free(&ipf->ipf_ro);
363 		if (just_one)
364 			return ipf;
365 		pool_put(&ipflow_pool, ipf);
366 		ipflow_inuse--;
367 	}
368 	return NULL;
369 }
370 
371 void
372 ipflow_slowtimo(void)
373 {
374 	struct ipflow *ipf, *next_ipf;
375 
376 	for (ipf = LIST_FIRST(&ipflowlist); ipf != NULL; ipf = next_ipf) {
377 		next_ipf = LIST_NEXT(ipf, ipf_list);
378 		if (PRT_SLOW_ISEXPIRED(ipf->ipf_timer) ||
379 		    rtcache_down(&ipf->ipf_ro) || ipf->ipf_ro.ro_rt == NULL) {
380 			ipflow_free(ipf);
381 		} else {
382 			ipf->ipf_last_uses = ipf->ipf_uses;
383 			ipf->ipf_ro.ro_rt->rt_use += ipf->ipf_uses;
384 			ipstat.ips_total += ipf->ipf_uses;
385 			ipstat.ips_forward += ipf->ipf_uses;
386 			ipstat.ips_fastforward += ipf->ipf_uses;
387 			ipf->ipf_uses = 0;
388 		}
389 	}
390 }
391 
392 void
393 ipflow_create(const struct route *ro, struct mbuf *m)
394 {
395 	struct ip *const ip = mtod(m, struct ip *);
396 	struct ipflow *ipf;
397 	size_t hash;
398 	int s;
399 
400 	/*
401 	 * Don't create cache entries for ICMP messages.
402 	 */
403 	if (ip_maxflows == 0 || ip->ip_p == IPPROTO_ICMP)
404 		return;
405 	/*
406 	 * See if an existing flow struct exists.  If so remove it from it's
407 	 * list and free the old route.  If not, try to malloc a new one
408 	 * (if we aren't at our limit).
409 	 */
410 	ipf = ipflow_lookup(ip);
411 	if (ipf == NULL) {
412 		if (ipflow_inuse >= ip_maxflows) {
413 			ipf = ipflow_reap(1);
414 		} else {
415 			s = splnet();
416 			ipf = pool_get(&ipflow_pool, PR_NOWAIT);
417 			splx(s);
418 			if (ipf == NULL)
419 				return;
420 			ipflow_inuse++;
421 		}
422 		memset(ipf, 0, sizeof(*ipf));
423 	} else {
424 		s = splnet();
425 		IPFLOW_REMOVE(ipf);
426 		splx(s);
427 		ipflow_addstats(ipf);
428 		rtcache_free(&ipf->ipf_ro);
429 		ipf->ipf_uses = ipf->ipf_last_uses = 0;
430 		ipf->ipf_errors = ipf->ipf_dropped = 0;
431 	}
432 
433 	/*
434 	 * Fill in the updated information.
435 	 */
436 	rtcache_copy(&ipf->ipf_ro, ro);
437 	ipf->ipf_dst = ip->ip_dst;
438 	ipf->ipf_src = ip->ip_src;
439 	ipf->ipf_tos = ip->ip_tos;
440 	PRT_SLOW_ARM(ipf->ipf_timer, IPFLOW_TIMER);
441 	ipf->ipf_start = time_uptime;
442 	/*
443 	 * Insert into the approriate bucket of the flow table.
444 	 */
445 	hash = ipflow_hash(ip);
446 	s = splnet();
447 	IPFLOW_INSERT(&ipflowtable[hash], ipf);
448 	splx(s);
449 }
450 
451 int
452 ipflow_invalidate_all(int new_size)
453 {
454 	struct ipflow *ipf, *next_ipf;
455 	int s, error;
456 
457 	error = 0;
458 	s = splnet();
459 	for (ipf = LIST_FIRST(&ipflowlist); ipf != NULL; ipf = next_ipf) {
460 		next_ipf = LIST_NEXT(ipf, ipf_list);
461 		ipflow_free(ipf);
462 	}
463 
464 	if (new_size)
465 		error = ipflow_init(new_size);
466 	splx(s);
467 
468 	return error;
469 }
470