2 * Copyright (c) 2004-2005 Gleb Smirnoff <glebius@FreeBSD.org>
3 * Copyright (c) 2001-2003 Roman V. Palagin <romanp@unshadow.net>
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
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.
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
27 * $SourceForge: netflow.c,v 1.41 2004/09/05 11:41:10 glebius Exp $
28 * $FreeBSD: src/sys/netgraph/netflow/netflow.c,v 1.29 2008/05/09 23:02:57 julian Exp $
29 * $DragonFly: src/sys/netgraph7/netflow/netflow.c,v 1.2 2008/06/26 23:05:40 dillon Exp $
32 #include <sys/param.h>
33 #include <sys/kernel.h>
34 #include <sys/limits.h>
36 #include <sys/syslog.h>
37 #include <sys/systm.h>
38 #include <sys/socket.h>
40 #include <machine/atomic.h>
43 #include <net/route.h>
44 #include <netinet/in.h>
45 #include <netinet/in_systm.h>
46 #include <netinet/ip.h>
47 #include <netinet/tcp.h>
48 #include <netinet/udp.h>
50 #include "ng_message.h"
53 #include "netflow/netflow.h"
54 #include "netflow/ng_netflow.h"
56 #define NBUCKETS (65536) /* must be power of 2 */
58 /* This hash is for TCP or UDP packets. */
59 #define FULL_HASH(addr1, addr2, port1, port2) \
60 (((addr1 ^ (addr1 >> 16) ^ \
61 htons(addr2 ^ (addr2 >> 16))) ^ \
62 port1 ^ htons(port2)) & \
65 /* This hash is for all other IP packets. */
66 #define ADDR_HASH(addr1, addr2) \
67 ((addr1 ^ (addr1 >> 16) ^ \
68 htons(addr2 ^ (addr2 >> 16))) & \
71 /* Macros to shorten logical constructions */
72 /* XXX: priv must exist in namespace */
73 #define INACTIVE(fle) (time_uptime - fle->f.last > priv->info.nfinfo_inact_t)
74 #define AGED(fle) (time_uptime - fle->f.first > priv->info.nfinfo_act_t)
75 #define ISFREE(fle) (fle->f.packets == 0)
78 * 4 is a magical number: statistically number of 4-packet flows is
79 * bigger than 5,6,7...-packet flows by an order of magnitude. Most UDP/ICMP
80 * scans are 1 packet (~ 90% of flow cache). TCP scans are 2-packet in case
81 * of reachable host and 4-packet otherwise.
83 #define SMALL(fle) (fle->f.packets <= 4)
86 * Cisco uses milliseconds for uptime. Bad idea, since it overflows
87 * every 48+ days. But we will do same to keep compatibility. This macro
88 * does overflowable multiplication to 1000.
90 #define MILLIUPTIME(t) (((t) << 9) + /* 512 */ \
91 ((t) << 8) + /* 256 */ \
92 ((t) << 7) + /* 128 */ \
93 ((t) << 6) + /* 64 */ \
94 ((t) << 5) + /* 32 */ \
97 MALLOC_DECLARE(M_NETFLOW_HASH);
98 MALLOC_DEFINE(M_NETFLOW_HASH, "netflow_hash", "NetFlow hash");
100 static int export_add(item_p, struct flow_entry *);
101 static int export_send(priv_p, item_p, int flags);
103 /* Generate hash for a given flow record. */
104 static __inline uint32_t
105 ip_hash(struct flow_rec *r)
110 return FULL_HASH(r->r_src.s_addr, r->r_dst.s_addr,
111 r->r_sport, r->r_dport);
113 return ADDR_HASH(r->r_src.s_addr, r->r_dst.s_addr);
117 /* This is callback from uma(9), called on alloc. */
119 uma_ctor_flow(void *mem, int size, void *arg, int how)
121 priv_p priv = (priv_p )arg;
123 if (atomic_load_acq_32(&priv->info.nfinfo_used) >= CACHESIZE)
126 atomic_add_32(&priv->info.nfinfo_used, 1);
131 /* This is callback from uma(9), called on free. */
133 uma_dtor_flow(void *mem, int size, void *arg)
135 priv_p priv = (priv_p )arg;
137 atomic_subtract_32(&priv->info.nfinfo_used, 1);
141 * Detach export datagram from priv, if there is any.
142 * If there is no, allocate a new one.
145 get_export_dgram(priv_p priv)
149 mtx_lock(&priv->export_mtx);
150 if (priv->export_item != NULL) {
151 item = priv->export_item;
152 priv->export_item = NULL;
154 mtx_unlock(&priv->export_mtx);
157 struct netflow_v5_export_dgram *dgram;
160 m = m_getcl(MB_DONTWAIT, MT_DATA, M_PKTHDR);
163 item = ng_package_data(m, NG_NOFLAGS);
166 dgram = mtod(m, struct netflow_v5_export_dgram *);
167 dgram->header.count = 0;
168 dgram->header.version = htons(NETFLOW_V5);
176 * Re-attach incomplete datagram back to priv.
177 * If there is already another one, then send incomplete. */
179 return_export_dgram(priv_p priv, item_p item, int flags)
182 * It may happen on SMP, that some thread has already
183 * put its item there, in this case we bail out and
184 * send what we have to collector.
186 mtx_lock(&priv->export_mtx);
187 if (priv->export_item == NULL) {
188 priv->export_item = item;
189 mtx_unlock(&priv->export_mtx);
191 mtx_unlock(&priv->export_mtx);
192 export_send(priv, item, flags);
197 * The flow is over. Call export_add() and free it. If datagram is
198 * full, then call export_send().
201 expire_flow(priv_p priv, item_p *item, struct flow_entry *fle, int flags)
204 *item = get_export_dgram(priv);
206 atomic_add_32(&priv->info.nfinfo_export_failed, 1);
207 uma_zfree_arg(priv->zone, fle, priv);
210 if (export_add(*item, fle) > 0) {
211 export_send(priv, *item, flags);
214 uma_zfree_arg(priv->zone, fle, priv);
217 /* Get a snapshot of node statistics */
219 ng_netflow_copyinfo(priv_p priv, struct ng_netflow_info *i)
222 memcpy((void *)i, (void *)&priv->info, sizeof(priv->info));
226 * Insert a record into defined slot.
228 * First we get for us a free flow entry, then fill in all
229 * possible fields in it.
231 * TODO: consider dropping hash mutex while filling in datagram,
232 * as this was done in previous version. Need to test & profile
236 hash_insert(priv_p priv, struct flow_hash_entry *hsh, struct flow_rec *r,
237 int plen, uint8_t tcp_flags)
239 struct flow_entry *fle;
240 struct sockaddr_in sin;
243 mtx_assert(&hsh->mtx, MA_OWNED);
245 fle = uma_zalloc_arg(priv->zone, priv, M_WAITOK | M_NULLOK);
247 atomic_add_32(&priv->info.nfinfo_alloc_failed, 1);
252 * Now fle is totally ours. It is detached from all lists,
253 * we can safely edit it.
256 bcopy(r, &fle->f.r, sizeof(struct flow_rec));
259 fle->f.tcp_flags = tcp_flags;
261 fle->f.first = fle->f.last = time_uptime;
264 * First we do route table lookup on destination address. So we can
265 * fill in out_ifx, dst_mask, nexthop, and dst_as in future releases.
267 bzero(&sin, sizeof(sin));
268 sin.sin_len = sizeof(struct sockaddr_in);
269 sin.sin_family = AF_INET;
270 sin.sin_addr = fle->f.r.r_dst;
271 /* XXX MRT 0 as a default.. need the m here to get fib */
272 rt = rtalloc1_fib((struct sockaddr *)&sin, 0, RTF_CLONING, 0);
274 fle->f.fle_o_ifx = rt->rt_ifp->if_index;
276 if (rt->rt_flags & RTF_GATEWAY &&
277 rt->rt_gateway->sa_family == AF_INET)
279 ((struct sockaddr_in *)(rt->rt_gateway))->sin_addr;
282 fle->f.dst_mask = bitcount32(((struct sockaddr_in *)
283 rt_mask(rt))->sin_addr.s_addr);
284 else if (rt->rt_flags & RTF_HOST)
285 /* Give up. We can't determine mask :( */
286 fle->f.dst_mask = 32;
291 /* Do route lookup on source address, to fill in src_mask. */
292 bzero(&sin, sizeof(sin));
293 sin.sin_len = sizeof(struct sockaddr_in);
294 sin.sin_family = AF_INET;
295 sin.sin_addr = fle->f.r.r_src;
296 /* XXX MRT 0 as a default revisit. need the mbuf for fib*/
297 rt = rtalloc1_fib((struct sockaddr *)&sin, 0, RTF_CLONING, 0);
300 fle->f.src_mask = bitcount32(((struct sockaddr_in *)
301 rt_mask(rt))->sin_addr.s_addr);
302 else if (rt->rt_flags & RTF_HOST)
303 /* Give up. We can't determine mask :( */
304 fle->f.src_mask = 32;
309 /* Push new flow at the and of hash. */
310 TAILQ_INSERT_TAIL(&hsh->head, fle, fle_hash);
317 * Non-static functions called from ng_netflow.c
320 /* Allocate memory and set up flow cache */
322 ng_netflow_cache_init(priv_p priv)
324 struct flow_hash_entry *hsh;
327 /* Initialize cache UMA zone. */
328 priv->zone = uma_zcreate("NetFlow cache", sizeof(struct flow_entry),
329 uma_ctor_flow, uma_dtor_flow, NULL, NULL, UMA_ALIGN_CACHE, 0);
330 uma_zone_set_max(priv->zone, CACHESIZE);
333 priv->hash = kmalloc(NBUCKETS * sizeof(struct flow_hash_entry),
334 M_NETFLOW_HASH, M_WAITOK | M_ZERO);
336 if (priv->hash == NULL) {
337 uma_zdestroy(priv->zone);
341 /* Initialize hash. */
342 for (i = 0, hsh = priv->hash; i < NBUCKETS; i++, hsh++) {
343 mtx_init(&hsh->mtx, "hash mutex", NULL, MTX_DEF);
344 TAILQ_INIT(&hsh->head);
347 mtx_init(&priv->export_mtx, "export dgram lock", NULL, MTX_DEF);
352 /* Free all flow cache memory. Called from node close method. */
354 ng_netflow_cache_flush(priv_p priv)
356 struct flow_entry *fle, *fle1;
357 struct flow_hash_entry *hsh;
362 * We are going to free probably billable data.
363 * Expire everything before freeing it.
364 * No locking is required since callout is already drained.
366 for (hsh = priv->hash, i = 0; i < NBUCKETS; hsh++, i++)
367 TAILQ_FOREACH_SAFE(fle, &hsh->head, fle_hash, fle1) {
368 TAILQ_REMOVE(&hsh->head, fle, fle_hash);
369 expire_flow(priv, &item, fle, NG_QUEUE);
373 export_send(priv, item, NG_QUEUE);
375 uma_zdestroy(priv->zone);
377 /* Destroy hash mutexes. */
378 for (i = 0, hsh = priv->hash; i < NBUCKETS; i++, hsh++)
379 mtx_destroy(&hsh->mtx);
381 /* Free hash memory. */
383 kfree(priv->hash, M_NETFLOW_HASH);
385 mtx_destroy(&priv->export_mtx);
388 /* Insert packet from into flow cache. */
390 ng_netflow_flow_add(priv_p priv, struct ip *ip, iface_p iface,
393 register struct flow_entry *fle, *fle1;
394 struct flow_hash_entry *hsh;
399 uint8_t tcp_flags = 0;
401 /* Try to fill flow_rec r */
402 bzero(&r, sizeof(r));
404 if (ip->ip_v != IPVERSION)
407 /* verify min header length */
408 hlen = ip->ip_hl << 2;
410 if (hlen < sizeof(struct ip))
413 r.r_src = ip->ip_src;
414 r.r_dst = ip->ip_dst;
416 /* save packet length */
417 plen = ntohs(ip->ip_len);
420 r.r_tos = ip->ip_tos;
422 /* Configured in_ifx overrides mbuf's */
423 if (iface->info.ifinfo_index == 0) {
425 r.r_i_ifx = ifp->if_index;
427 r.r_i_ifx = iface->info.ifinfo_index;
430 * XXX NOTE: only first fragment of fragmented TCP, UDP and
431 * ICMP packet will be recorded with proper s_port and d_port.
432 * Following fragments will be recorded simply as IP packet with
433 * ip_proto = ip->ip_p and s_port, d_port set to zero.
434 * I know, it looks like bug. But I don't want to re-implement
435 * ip packet assebmling here. Anyway, (in)famous trafd works this way -
436 * and nobody complains yet :)
438 if ((ip->ip_off & htons(IP_OFFMASK)) == 0)
442 register struct tcphdr *tcp;
444 tcp = (struct tcphdr *)((caddr_t )ip + hlen);
445 r.r_sport = tcp->th_sport;
446 r.r_dport = tcp->th_dport;
447 tcp_flags = tcp->th_flags;
451 r.r_ports = *(uint32_t *)((caddr_t )ip + hlen);
455 /* Update node statistics. XXX: race... */
456 priv->info.nfinfo_packets ++;
457 priv->info.nfinfo_bytes += plen;
459 /* Find hash slot. */
460 hsh = &priv->hash[ip_hash(&r)];
465 * Go through hash and find our entry. If we encounter an
466 * entry, that should be expired, purge it. We do a reverse
467 * search since most active entries are first, and most
468 * searches are done on most active entries.
470 TAILQ_FOREACH_REVERSE_SAFE(fle, &hsh->head, fhead, fle_hash, fle1) {
471 if (bcmp(&r, &fle->f.r, sizeof(struct flow_rec)) == 0)
473 if ((INACTIVE(fle) && SMALL(fle)) || AGED(fle)) {
474 TAILQ_REMOVE(&hsh->head, fle, fle_hash);
475 expire_flow(priv, &item, fle, NG_QUEUE);
476 atomic_add_32(&priv->info.nfinfo_act_exp, 1);
480 if (fle) { /* An existent entry. */
482 fle->f.bytes += plen;
484 fle->f.tcp_flags |= tcp_flags;
485 fle->f.last = time_uptime;
488 * We have the following reasons to expire flow in active way:
489 * - it hit active timeout
490 * - a TCP connection closed
491 * - it is going to overflow counter
493 if (tcp_flags & TH_FIN || tcp_flags & TH_RST || AGED(fle) ||
494 (fle->f.bytes >= (UINT_MAX - IF_MAXMTU)) ) {
495 TAILQ_REMOVE(&hsh->head, fle, fle_hash);
496 expire_flow(priv, &item, fle, NG_QUEUE);
497 atomic_add_32(&priv->info.nfinfo_act_exp, 1);
500 * It is the newest, move it to the tail,
501 * if it isn't there already. Next search will
504 if (fle != TAILQ_LAST(&hsh->head, fhead)) {
505 TAILQ_REMOVE(&hsh->head, fle, fle_hash);
506 TAILQ_INSERT_TAIL(&hsh->head, fle, fle_hash);
509 } else /* A new flow entry. */
510 error = hash_insert(priv, hsh, &r, plen, tcp_flags);
512 mtx_unlock(&hsh->mtx);
515 return_export_dgram(priv, item, NG_QUEUE);
521 * Return records from cache to userland.
523 * TODO: matching particular IP should be done in kernel, here.
526 ng_netflow_flow_show(priv_p priv, uint32_t last, struct ng_mesg *resp)
528 struct flow_hash_entry *hsh;
529 struct flow_entry *fle;
530 struct ngnf_flows *data;
533 data = (struct ngnf_flows *)resp->data;
537 /* Check if this is a first run */
542 if (last > NBUCKETS-1)
544 hsh = priv->hash + last;
549 * We will transfer not more than NREC_AT_ONCE. More data
550 * will come in next message.
551 * We send current hash index to userland, and userland should
552 * return it back to us. Then, we will restart with new entry.
554 * The resulting cache snapshot is inaccurate for the
556 * - we skip locked hash entries
557 * - we bail out, if someone wants our entry
558 * - we skip rest of entry, when hit NREC_AT_ONCE
560 for (; i < NBUCKETS; hsh++, i++) {
561 if (mtx_trylock(&hsh->mtx) == 0)
564 TAILQ_FOREACH(fle, &hsh->head, fle_hash) {
565 if (mtx_contested(&hsh->mtx)
568 bcopy(&fle->f, &(data->entries[data->nentries]),
571 if (data->nentries == NREC_AT_ONCE) {
572 mtx_unlock(&hsh->mtx);
578 mtx_unlock(&hsh->mtx);
584 /* We have full datagram in privdata. Send it to export hook. */
586 export_send(priv_p priv, item_p item, int flags)
588 struct mbuf *m = NGI_M(item);
589 struct netflow_v5_export_dgram *dgram = mtod(m,
590 struct netflow_v5_export_dgram *);
591 struct netflow_v5_header *header = &dgram->header;
595 /* Fill mbuf header. */
596 m->m_len = m->m_pkthdr.len = sizeof(struct netflow_v5_record) *
597 header->count + sizeof(struct netflow_v5_header);
599 /* Fill export header. */
600 header->sys_uptime = htonl(MILLIUPTIME(time_uptime));
602 header->unix_secs = htonl(ts.tv_sec);
603 header->unix_nsecs = htonl(ts.tv_nsec);
604 header->engine_type = 0;
605 header->engine_id = 0;
607 header->flow_seq = htonl(atomic_fetchadd_32(&priv->flow_seq,
609 header->count = htons(header->count);
611 if (priv->export != NULL)
612 NG_FWD_ITEM_HOOK_FLAGS(error, item, priv->export, flags);
620 /* Add export record to dgram. */
622 export_add(item_p item, struct flow_entry *fle)
624 struct netflow_v5_export_dgram *dgram = mtod(NGI_M(item),
625 struct netflow_v5_export_dgram *);
626 struct netflow_v5_header *header = &dgram->header;
627 struct netflow_v5_record *rec;
629 rec = &dgram->r[header->count];
632 KASSERT(header->count <= NETFLOW_V5_MAX_RECORDS,
633 ("ng_netflow: export too big"));
635 /* Fill in export record. */
636 rec->src_addr = fle->f.r.r_src.s_addr;
637 rec->dst_addr = fle->f.r.r_dst.s_addr;
638 rec->next_hop = fle->f.next_hop.s_addr;
639 rec->i_ifx = htons(fle->f.fle_i_ifx);
640 rec->o_ifx = htons(fle->f.fle_o_ifx);
641 rec->packets = htonl(fle->f.packets);
642 rec->octets = htonl(fle->f.bytes);
643 rec->first = htonl(MILLIUPTIME(fle->f.first));
644 rec->last = htonl(MILLIUPTIME(fle->f.last));
645 rec->s_port = fle->f.r.r_sport;
646 rec->d_port = fle->f.r.r_dport;
647 rec->flags = fle->f.tcp_flags;
648 rec->prot = fle->f.r.r_ip_p;
649 rec->tos = fle->f.r.r_tos;
650 rec->dst_mask = fle->f.dst_mask;
651 rec->src_mask = fle->f.src_mask;
653 /* Not supported fields. */
654 rec->src_as = rec->dst_as = 0;
656 if (header->count == NETFLOW_V5_MAX_RECORDS)
657 return (1); /* end of datagram */
662 /* Periodic flow expiry run. */
664 ng_netflow_expire(void *arg)
666 struct flow_entry *fle, *fle1;
667 struct flow_hash_entry *hsh;
668 priv_p priv = (priv_p )arg;
674 * Going through all the cache.
676 for (hsh = priv->hash, i = 0; i < NBUCKETS; hsh++, i++) {
678 * Skip entries, that are already being worked on.
680 if (mtx_trylock(&hsh->mtx) == 0)
683 used = atomic_load_acq_32(&priv->info.nfinfo_used);
684 TAILQ_FOREACH_SAFE(fle, &hsh->head, fle_hash, fle1) {
686 * Interrupt thread wants this entry!
687 * Quick! Quick! Bail out!
689 if (mtx_contested(&hsh->mtx))
693 * Don't expire aggressively while hash collision
694 * ratio is predicted small.
696 if (used <= (NBUCKETS*2) && !INACTIVE(fle))
699 if ((INACTIVE(fle) && (SMALL(fle) ||
700 (used > (NBUCKETS*2)))) || AGED(fle)) {
701 TAILQ_REMOVE(&hsh->head, fle, fle_hash);
702 expire_flow(priv, &item, fle, NG_NOFLAGS);
704 atomic_add_32(&priv->info.nfinfo_inact_exp, 1);
707 mtx_unlock(&hsh->mtx);
711 return_export_dgram(priv, item, NG_NOFLAGS);
713 /* Schedule next expire. */
714 callout_reset(&priv->exp_callout, (1*hz), &ng_netflow_expire,