xref: /freebsd/sys/netgraph/netflow/netflow.c (revision 79775f8f1be8dc6ce2b705276b7626cb24347f7d)
1 /*-
2  * Copyright (c) 2004-2005 Gleb Smirnoff <glebius@FreeBSD.org>
3  * Copyright (c) 2001-2003 Roman V. Palagin <romanp@unshadow.net>
4  * All rights reserved.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25  * SUCH DAMAGE.
26  *
27  * $SourceForge: netflow.c,v 1.41 2004/09/05 11:41:10 glebius Exp $
28  */
29 
30 static const char rcs_id[] =
31     "@(#) $FreeBSD$";
32 
33 #include <sys/param.h>
34 #include <sys/kernel.h>
35 #include <sys/limits.h>
36 #include <sys/mbuf.h>
37 #include <sys/syslog.h>
38 #include <sys/systm.h>
39 #include <sys/socket.h>
40 
41 #include <machine/atomic.h>
42 
43 #include <net/if.h>
44 #include <net/route.h>
45 #include <netinet/in.h>
46 #include <netinet/in_systm.h>
47 #include <netinet/ip.h>
48 #include <netinet/tcp.h>
49 #include <netinet/udp.h>
50 
51 #include <netgraph/ng_message.h>
52 #include <netgraph/netgraph.h>
53 
54 #include <netgraph/netflow/netflow.h>
55 #include <netgraph/netflow/ng_netflow.h>
56 
57 #define	NBUCKETS	(65536)		/* must be power of 2 */
58 
59 /* This hash is for TCP or UDP packets. */
60 #define FULL_HASH(addr1, addr2, port1, port2)	\
61 	(((addr1 ^ (addr1 >> 16) ^ 		\
62 	htons(addr2 ^ (addr2 >> 16))) ^ 	\
63 	port1 ^ htons(port2)) &			\
64 	(NBUCKETS - 1))
65 
66 /* This hash is for all other IP packets. */
67 #define ADDR_HASH(addr1, addr2)			\
68 	((addr1 ^ (addr1 >> 16) ^ 		\
69 	htons(addr2 ^ (addr2 >> 16))) &		\
70 	(NBUCKETS - 1))
71 
72 /* Macros to shorten logical constructions */
73 /* XXX: priv must exist in namespace */
74 #define	INACTIVE(fle)	(time_uptime - fle->f.last > priv->info.nfinfo_inact_t)
75 #define	AGED(fle)	(time_uptime - fle->f.first > priv->info.nfinfo_act_t)
76 #define	ISFREE(fle)	(fle->f.packets == 0)
77 
78 /*
79  * 4 is a magical number: statistically number of 4-packet flows is
80  * bigger than 5,6,7...-packet flows by an order of magnitude. Most UDP/ICMP
81  * scans are 1 packet (~ 90% of flow cache). TCP scans are 2-packet in case
82  * of reachable host and 4-packet otherwise.
83  */
84 #define	SMALL(fle)	(fle->f.packets <= 4)
85 
86 /*
87  * Cisco uses milliseconds for uptime. Bad idea, since it overflows
88  * every 48+ days. But we will do same to keep compatibility. This macro
89  * does overflowable multiplication to 1000.
90  */
91 #define	MILLIUPTIME(t)	(((t) << 9) +	/* 512 */	\
92 			 ((t) << 8) +	/* 256 */	\
93 			 ((t) << 7) +	/* 128 */	\
94 			 ((t) << 6) +	/* 64  */	\
95 			 ((t) << 5) +	/* 32  */	\
96 			 ((t) << 3))	/* 8   */
97 
98 MALLOC_DECLARE(M_NETFLOW_HASH);
99 MALLOC_DEFINE(M_NETFLOW_HASH, "netflow_hash", "NetFlow hash");
100 
101 static int export_add(item_p, struct flow_entry *);
102 static int export_send(priv_p, item_p, int flags);
103 
104 /* Generate hash for a given flow record. */
105 static __inline uint32_t
106 ip_hash(struct flow_rec *r)
107 {
108 	switch (r->r_ip_p) {
109 	case IPPROTO_TCP:
110 	case IPPROTO_UDP:
111 		return FULL_HASH(r->r_src.s_addr, r->r_dst.s_addr,
112 		    r->r_sport, r->r_dport);
113 	default:
114 		return ADDR_HASH(r->r_src.s_addr, r->r_dst.s_addr);
115 	}
116 }
117 
118 /* This is callback from uma(9), called on alloc. */
119 static int
120 uma_ctor_flow(void *mem, int size, void *arg, int how)
121 {
122 	priv_p priv = (priv_p )arg;
123 
124 	if (atomic_load_acq_32(&priv->info.nfinfo_used) >= CACHESIZE)
125 		return (ENOMEM);
126 
127 	atomic_add_32(&priv->info.nfinfo_used, 1);
128 
129 	return (0);
130 }
131 
132 /* This is callback from uma(9), called on free. */
133 static void
134 uma_dtor_flow(void *mem, int size, void *arg)
135 {
136 	priv_p priv = (priv_p )arg;
137 
138 	atomic_subtract_32(&priv->info.nfinfo_used, 1);
139 }
140 
141 /*
142  * Detach export datagram from priv, if there is any.
143  * If there is no, allocate a new one.
144  */
145 static item_p
146 get_export_dgram(priv_p priv)
147 {
148 	item_p	item = NULL;
149 
150 	mtx_lock(&priv->export_mtx);
151 	if (priv->export_item != NULL) {
152 		item = priv->export_item;
153 		priv->export_item = NULL;
154 	}
155 	mtx_unlock(&priv->export_mtx);
156 
157 	if (item == NULL) {
158 		struct netflow_v5_export_dgram *dgram;
159 		struct mbuf *m;
160 
161 		m = m_getcl(M_DONTWAIT, MT_DATA, M_PKTHDR);
162 		if (m == NULL)
163 			return (NULL);
164 		item = ng_package_data(m, NG_NOFLAGS);
165 		if (item == NULL)
166 			return (NULL);
167 		dgram = mtod(m, struct netflow_v5_export_dgram *);
168 		dgram->header.count = 0;
169 		dgram->header.version = htons(NETFLOW_V5);
170 
171 	}
172 
173 	return (item);
174 }
175 
176 /*
177  * Re-attach incomplete datagram back to priv.
178  * If there is already another one, then send incomplete. */
179 static void
180 return_export_dgram(priv_p priv, item_p item, int flags)
181 {
182 	/*
183 	 * It may happen on SMP, that some thread has already
184 	 * put its item there, in this case we bail out and
185 	 * send what we have to collector.
186 	 */
187 	mtx_lock(&priv->export_mtx);
188 	if (priv->export_item == NULL) {
189 		priv->export_item = item;
190 		mtx_unlock(&priv->export_mtx);
191 	} else {
192 		mtx_unlock(&priv->export_mtx);
193 		export_send(priv, item, flags);
194 	}
195 }
196 
197 /*
198  * The flow is over. Call export_add() and free it. If datagram is
199  * full, then call export_send().
200  */
201 static __inline void
202 expire_flow(priv_p priv, item_p *item, struct flow_entry *fle, int flags)
203 {
204 	if (*item == NULL)
205 		*item = get_export_dgram(priv);
206 	if (*item == NULL) {
207 		atomic_add_32(&priv->info.nfinfo_export_failed, 1);
208 		uma_zfree_arg(priv->zone, fle, priv);
209 		return;
210 	}
211 	if (export_add(*item, fle) > 0) {
212 		export_send(priv, *item, flags);
213 		*item = NULL;
214 	}
215 	uma_zfree_arg(priv->zone, fle, priv);
216 }
217 
218 /* Get a snapshot of node statistics */
219 void
220 ng_netflow_copyinfo(priv_p priv, struct ng_netflow_info *i)
221 {
222 	/* XXX: atomic */
223 	memcpy((void *)i, (void *)&priv->info, sizeof(priv->info));
224 }
225 
226 /*
227  * Insert a record into defined slot.
228  *
229  * First we get for us a free flow entry, then fill in all
230  * possible fields in it.
231  *
232  * TODO: consider dropping hash mutex while filling in datagram,
233  * as this was done in previous version. Need to test & profile
234  * to be sure.
235  */
236 static __inline int
237 hash_insert(priv_p priv, struct flow_hash_entry  *hsh, struct flow_rec *r,
238 	int plen, uint8_t tcp_flags)
239 {
240 	struct flow_entry *fle;
241 	struct sockaddr_in sin;
242 	struct rtentry *rt;
243 
244 	mtx_assert(&hsh->mtx, MA_OWNED);
245 
246 	fle = uma_zalloc_arg(priv->zone, priv, M_NOWAIT);
247 	if (fle == NULL) {
248 		atomic_add_32(&priv->info.nfinfo_alloc_failed, 1);
249 		return (ENOMEM);
250 	}
251 
252 	/*
253 	 * Now fle is totally ours. It is detached from all lists,
254 	 * we can safely edit it.
255 	 */
256 
257 	bcopy(r, &fle->f.r, sizeof(struct flow_rec));
258 	fle->f.bytes = plen;
259 	fle->f.packets = 1;
260 	fle->f.tcp_flags = tcp_flags;
261 
262 	fle->f.first = fle->f.last = time_uptime;
263 
264 	/*
265 	 * First we do route table lookup on destination address. So we can
266 	 * fill in out_ifx, dst_mask, nexthop, and dst_as in future releases.
267 	 */
268 	bzero(&sin, sizeof(sin));
269 	sin.sin_len = sizeof(struct sockaddr_in);
270 	sin.sin_family = AF_INET;
271 	sin.sin_addr = fle->f.r.r_dst;
272 	/* XXX MRT 0 as a default.. need the m here to get fib */
273 	rt = rtalloc1_fib((struct sockaddr *)&sin, 0, RTF_CLONING, 0);
274 	if (rt != NULL) {
275 		fle->f.fle_o_ifx = rt->rt_ifp->if_index;
276 
277 		if (rt->rt_flags & RTF_GATEWAY &&
278 		    rt->rt_gateway->sa_family == AF_INET)
279 			fle->f.next_hop =
280 			    ((struct sockaddr_in *)(rt->rt_gateway))->sin_addr;
281 
282 		if (rt_mask(rt))
283 			fle->f.dst_mask = bitcount32(((struct sockaddr_in *)
284 			    rt_mask(rt))->sin_addr.s_addr);
285 		else if (rt->rt_flags & RTF_HOST)
286 			/* Give up. We can't determine mask :( */
287 			fle->f.dst_mask = 32;
288 
289 		RTFREE_LOCKED(rt);
290 	}
291 
292 	/* Do route lookup on source address, to fill in src_mask. */
293 	bzero(&sin, sizeof(sin));
294 	sin.sin_len = sizeof(struct sockaddr_in);
295 	sin.sin_family = AF_INET;
296 	sin.sin_addr = fle->f.r.r_src;
297 	/* XXX MRT 0 as a default  revisit.  need the mbuf for fib*/
298 	rt = rtalloc1_fib((struct sockaddr *)&sin, 0, RTF_CLONING, 0);
299 	if (rt != NULL) {
300 		if (rt_mask(rt))
301 			fle->f.src_mask = bitcount32(((struct sockaddr_in *)
302 			    rt_mask(rt))->sin_addr.s_addr);
303 		else if (rt->rt_flags & RTF_HOST)
304 			/* Give up. We can't determine mask :( */
305 			fle->f.src_mask = 32;
306 
307 		RTFREE_LOCKED(rt);
308 	}
309 
310 	/* Push new flow at the and of hash. */
311 	TAILQ_INSERT_TAIL(&hsh->head, fle, fle_hash);
312 
313 	return (0);
314 }
315 
316 
317 /*
318  * Non-static functions called from ng_netflow.c
319  */
320 
321 /* Allocate memory and set up flow cache */
322 int
323 ng_netflow_cache_init(priv_p priv)
324 {
325 	struct flow_hash_entry	*hsh;
326 	int i;
327 
328 	/* Initialize cache UMA zone. */
329 	priv->zone = uma_zcreate("NetFlow cache", sizeof(struct flow_entry),
330 	    uma_ctor_flow, uma_dtor_flow, NULL, NULL, UMA_ALIGN_CACHE, 0);
331 	uma_zone_set_max(priv->zone, CACHESIZE);
332 
333 	/* Allocate hash. */
334 	MALLOC(priv->hash, struct flow_hash_entry *,
335 	    NBUCKETS * sizeof(struct flow_hash_entry),
336 	    M_NETFLOW_HASH, M_WAITOK | M_ZERO);
337 
338 	if (priv->hash == NULL) {
339 		uma_zdestroy(priv->zone);
340 		return (ENOMEM);
341 	}
342 
343 	/* Initialize hash. */
344 	for (i = 0, hsh = priv->hash; i < NBUCKETS; i++, hsh++) {
345 		mtx_init(&hsh->mtx, "hash mutex", NULL, MTX_DEF);
346 		TAILQ_INIT(&hsh->head);
347 	}
348 
349 	mtx_init(&priv->export_mtx, "export dgram lock", NULL, MTX_DEF);
350 
351 	return (0);
352 }
353 
354 /* Free all flow cache memory. Called from node close method. */
355 void
356 ng_netflow_cache_flush(priv_p priv)
357 {
358 	struct flow_entry	*fle, *fle1;
359 	struct flow_hash_entry	*hsh;
360 	item_p			item = NULL;
361 	int i;
362 
363 	/*
364 	 * We are going to free probably billable data.
365 	 * Expire everything before freeing it.
366 	 * No locking is required since callout is already drained.
367 	 */
368 	for (hsh = priv->hash, i = 0; i < NBUCKETS; hsh++, i++)
369 		TAILQ_FOREACH_SAFE(fle, &hsh->head, fle_hash, fle1) {
370 			TAILQ_REMOVE(&hsh->head, fle, fle_hash);
371 			expire_flow(priv, &item, fle, NG_QUEUE);
372 		}
373 
374 	if (item != NULL)
375 		export_send(priv, item, NG_QUEUE);
376 
377 	uma_zdestroy(priv->zone);
378 
379 	/* Destroy hash mutexes. */
380 	for (i = 0, hsh = priv->hash; i < NBUCKETS; i++, hsh++)
381 		mtx_destroy(&hsh->mtx);
382 
383 	/* Free hash memory. */
384 	if (priv->hash)
385 		FREE(priv->hash, M_NETFLOW_HASH);
386 
387 	mtx_destroy(&priv->export_mtx);
388 }
389 
390 /* Insert packet from into flow cache. */
391 int
392 ng_netflow_flow_add(priv_p priv, struct ip *ip, iface_p iface,
393 	struct ifnet *ifp)
394 {
395 	register struct flow_entry	*fle, *fle1;
396 	struct flow_hash_entry		*hsh;
397 	struct flow_rec		r;
398 	item_p			item = NULL;
399 	int			hlen, plen;
400 	int			error = 0;
401 	uint8_t			tcp_flags = 0;
402 
403 	/* Try to fill flow_rec r */
404 	bzero(&r, sizeof(r));
405 	/* check version */
406 	if (ip->ip_v != IPVERSION)
407 		return (EINVAL);
408 
409 	/* verify min header length */
410 	hlen = ip->ip_hl << 2;
411 
412 	if (hlen < sizeof(struct ip))
413 		return (EINVAL);
414 
415 	r.r_src = ip->ip_src;
416 	r.r_dst = ip->ip_dst;
417 
418 	/* save packet length */
419 	plen = ntohs(ip->ip_len);
420 
421 	r.r_ip_p = ip->ip_p;
422 	r.r_tos = ip->ip_tos;
423 
424 	/* Configured in_ifx overrides mbuf's */
425 	if (iface->info.ifinfo_index == 0) {
426 		if (ifp != NULL)
427 			r.r_i_ifx = ifp->if_index;
428 	} else
429 		r.r_i_ifx = iface->info.ifinfo_index;
430 
431 	/*
432 	 * XXX NOTE: only first fragment of fragmented TCP, UDP and
433 	 * ICMP packet will be recorded with proper s_port and d_port.
434 	 * Following fragments will be recorded simply as IP packet with
435 	 * ip_proto = ip->ip_p and s_port, d_port set to zero.
436 	 * I know, it looks like bug. But I don't want to re-implement
437 	 * ip packet assebmling here. Anyway, (in)famous trafd works this way -
438 	 * and nobody complains yet :)
439 	 */
440 	if ((ip->ip_off & htons(IP_OFFMASK)) == 0)
441 		switch(r.r_ip_p) {
442 		case IPPROTO_TCP:
443 		{
444 			register struct tcphdr *tcp;
445 
446 			tcp = (struct tcphdr *)((caddr_t )ip + hlen);
447 			r.r_sport = tcp->th_sport;
448 			r.r_dport = tcp->th_dport;
449 			tcp_flags = tcp->th_flags;
450 			break;
451 		}
452 			case IPPROTO_UDP:
453 			r.r_ports = *(uint32_t *)((caddr_t )ip + hlen);
454 			break;
455 		}
456 
457 	/* Update node statistics. XXX: race... */
458 	priv->info.nfinfo_packets ++;
459 	priv->info.nfinfo_bytes += plen;
460 
461 	/* Find hash slot. */
462 	hsh = &priv->hash[ip_hash(&r)];
463 
464 	mtx_lock(&hsh->mtx);
465 
466 	/*
467 	 * Go through hash and find our entry. If we encounter an
468 	 * entry, that should be expired, purge it. We do a reverse
469 	 * search since most active entries are first, and most
470 	 * searches are done on most active entries.
471 	 */
472 	TAILQ_FOREACH_REVERSE_SAFE(fle, &hsh->head, fhead, fle_hash, fle1) {
473 		if (bcmp(&r, &fle->f.r, sizeof(struct flow_rec)) == 0)
474 			break;
475 		if ((INACTIVE(fle) && SMALL(fle)) || AGED(fle)) {
476 			TAILQ_REMOVE(&hsh->head, fle, fle_hash);
477 			expire_flow(priv, &item, fle, NG_QUEUE);
478 			atomic_add_32(&priv->info.nfinfo_act_exp, 1);
479 		}
480 	}
481 
482 	if (fle) {			/* An existent entry. */
483 
484 		fle->f.bytes += plen;
485 		fle->f.packets ++;
486 		fle->f.tcp_flags |= tcp_flags;
487 		fle->f.last = time_uptime;
488 
489 		/*
490 		 * We have the following reasons to expire flow in active way:
491 		 * - it hit active timeout
492 		 * - a TCP connection closed
493 		 * - it is going to overflow counter
494 		 */
495 		if (tcp_flags & TH_FIN || tcp_flags & TH_RST || AGED(fle) ||
496 		    (fle->f.bytes >= (UINT_MAX - IF_MAXMTU)) ) {
497 			TAILQ_REMOVE(&hsh->head, fle, fle_hash);
498 			expire_flow(priv, &item, fle, NG_QUEUE);
499 			atomic_add_32(&priv->info.nfinfo_act_exp, 1);
500 		} else {
501 			/*
502 			 * It is the newest, move it to the tail,
503 			 * if it isn't there already. Next search will
504 			 * locate it quicker.
505 			 */
506 			if (fle != TAILQ_LAST(&hsh->head, fhead)) {
507 				TAILQ_REMOVE(&hsh->head, fle, fle_hash);
508 				TAILQ_INSERT_TAIL(&hsh->head, fle, fle_hash);
509 			}
510 		}
511 	} else				/* A new flow entry. */
512 		error = hash_insert(priv, hsh, &r, plen, tcp_flags);
513 
514 	mtx_unlock(&hsh->mtx);
515 
516 	if (item != NULL)
517 		return_export_dgram(priv, item, NG_QUEUE);
518 
519 	return (error);
520 }
521 
522 /*
523  * Return records from cache to userland.
524  *
525  * TODO: matching particular IP should be done in kernel, here.
526  */
527 int
528 ng_netflow_flow_show(priv_p priv, uint32_t last, struct ng_mesg *resp)
529 {
530 	struct flow_hash_entry *hsh;
531 	struct flow_entry *fle;
532 	struct ngnf_flows *data;
533 	int i;
534 
535 	data = (struct ngnf_flows *)resp->data;
536 	data->last = 0;
537 	data->nentries = 0;
538 
539 	/* Check if this is a first run */
540 	if (last == 0) {
541 		hsh = priv->hash;
542 		i = 0;
543 	} else {
544 		if (last > NBUCKETS-1)
545 			return (EINVAL);
546 		hsh = priv->hash + last;
547 		i = last;
548 	}
549 
550 	/*
551 	 * We will transfer not more than NREC_AT_ONCE. More data
552 	 * will come in next message.
553 	 * We send current hash index to userland, and userland should
554 	 * return it back to us. Then, we will restart with new entry.
555 	 *
556 	 * The resulting cache snapshot is inaccurate for the
557 	 * following reasons:
558 	 *  - we skip locked hash entries
559 	 *  - we bail out, if someone wants our entry
560 	 *  - we skip rest of entry, when hit NREC_AT_ONCE
561 	 */
562 	for (; i < NBUCKETS; hsh++, i++) {
563 		if (mtx_trylock(&hsh->mtx) == 0)
564 			continue;
565 
566 		TAILQ_FOREACH(fle, &hsh->head, fle_hash) {
567 			if (hsh->mtx.mtx_lock & MTX_CONTESTED)
568 				break;
569 
570 			bcopy(&fle->f, &(data->entries[data->nentries]),
571 			    sizeof(fle->f));
572 			data->nentries++;
573 			if (data->nentries == NREC_AT_ONCE) {
574 				mtx_unlock(&hsh->mtx);
575 				if (++i < NBUCKETS)
576 					data->last = i;
577 				return (0);
578 			}
579 		}
580 		mtx_unlock(&hsh->mtx);
581 	}
582 
583 	return (0);
584 }
585 
586 /* We have full datagram in privdata. Send it to export hook. */
587 static int
588 export_send(priv_p priv, item_p item, int flags)
589 {
590 	struct mbuf *m = NGI_M(item);
591 	struct netflow_v5_export_dgram *dgram = mtod(m,
592 					struct netflow_v5_export_dgram *);
593 	struct netflow_v5_header *header = &dgram->header;
594 	struct timespec ts;
595 	int error = 0;
596 
597 	/* Fill mbuf header. */
598 	m->m_len = m->m_pkthdr.len = sizeof(struct netflow_v5_record) *
599 	   header->count + sizeof(struct netflow_v5_header);
600 
601 	/* Fill export header. */
602 	header->sys_uptime = htonl(MILLIUPTIME(time_uptime));
603 	getnanotime(&ts);
604 	header->unix_secs  = htonl(ts.tv_sec);
605 	header->unix_nsecs = htonl(ts.tv_nsec);
606 	header->engine_type = 0;
607 	header->engine_id = 0;
608 	header->pad = 0;
609 	header->flow_seq = htonl(atomic_fetchadd_32(&priv->flow_seq,
610 	    header->count));
611 	header->count = htons(header->count);
612 
613 	if (priv->export != NULL)
614 		NG_FWD_ITEM_HOOK_FLAGS(error, item, priv->export, flags);
615 	else
616 		NG_FREE_ITEM(item);
617 
618 	return (error);
619 }
620 
621 
622 /* Add export record to dgram. */
623 static int
624 export_add(item_p item, struct flow_entry *fle)
625 {
626 	struct netflow_v5_export_dgram *dgram = mtod(NGI_M(item),
627 					struct netflow_v5_export_dgram *);
628 	struct netflow_v5_header *header = &dgram->header;
629 	struct netflow_v5_record *rec;
630 
631 	rec = &dgram->r[header->count];
632 	header->count ++;
633 
634 	KASSERT(header->count <= NETFLOW_V5_MAX_RECORDS,
635 	    ("ng_netflow: export too big"));
636 
637 	/* Fill in export record. */
638 	rec->src_addr = fle->f.r.r_src.s_addr;
639 	rec->dst_addr = fle->f.r.r_dst.s_addr;
640 	rec->next_hop = fle->f.next_hop.s_addr;
641 	rec->i_ifx    = htons(fle->f.fle_i_ifx);
642 	rec->o_ifx    = htons(fle->f.fle_o_ifx);
643 	rec->packets  = htonl(fle->f.packets);
644 	rec->octets   = htonl(fle->f.bytes);
645 	rec->first    = htonl(MILLIUPTIME(fle->f.first));
646 	rec->last     = htonl(MILLIUPTIME(fle->f.last));
647 	rec->s_port   = fle->f.r.r_sport;
648 	rec->d_port   = fle->f.r.r_dport;
649 	rec->flags    = fle->f.tcp_flags;
650 	rec->prot     = fle->f.r.r_ip_p;
651 	rec->tos      = fle->f.r.r_tos;
652 	rec->dst_mask = fle->f.dst_mask;
653 	rec->src_mask = fle->f.src_mask;
654 
655 	/* Not supported fields. */
656 	rec->src_as = rec->dst_as = 0;
657 
658 	if (header->count == NETFLOW_V5_MAX_RECORDS)
659 		return (1); /* end of datagram */
660 	else
661 		return (0);
662 }
663 
664 /* Periodic flow expiry run. */
665 void
666 ng_netflow_expire(void *arg)
667 {
668 	struct flow_entry	*fle, *fle1;
669 	struct flow_hash_entry	*hsh;
670 	priv_p			priv = (priv_p )arg;
671 	item_p			item = NULL;
672 	uint32_t		used;
673 	int			i;
674 
675 	/*
676 	 * Going through all the cache.
677 	 */
678 	for (hsh = priv->hash, i = 0; i < NBUCKETS; hsh++, i++) {
679 		/*
680 		 * Skip entries, that are already being worked on.
681 		 */
682 		if (mtx_trylock(&hsh->mtx) == 0)
683 			continue;
684 
685 		used = atomic_load_acq_32(&priv->info.nfinfo_used);
686 		TAILQ_FOREACH_SAFE(fle, &hsh->head, fle_hash, fle1) {
687 			/*
688 			 * Interrupt thread wants this entry!
689 			 * Quick! Quick! Bail out!
690 			 */
691 			if (hsh->mtx.mtx_lock & MTX_CONTESTED)
692 				break;
693 
694 			/*
695 			 * Don't expire aggressively while hash collision
696 			 * ratio is predicted small.
697 			 */
698 			if (used <= (NBUCKETS*2) && !INACTIVE(fle))
699 				break;
700 
701 			if ((INACTIVE(fle) && (SMALL(fle) ||
702 			    (used > (NBUCKETS*2)))) || AGED(fle)) {
703 				TAILQ_REMOVE(&hsh->head, fle, fle_hash);
704 				expire_flow(priv, &item, fle, NG_NOFLAGS);
705 				used--;
706 				atomic_add_32(&priv->info.nfinfo_inact_exp, 1);
707 			}
708 		}
709 		mtx_unlock(&hsh->mtx);
710 	}
711 
712 	if (item != NULL)
713 		return_export_dgram(priv, item, NG_NOFLAGS);
714 
715 	/* Schedule next expire. */
716 	callout_reset(&priv->exp_callout, (1*hz), &ng_netflow_expire,
717 	    (void *)priv);
718 }
719