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 $
31 #include <sys/param.h>
32 #include <sys/kernel.h>
33 #include <sys/limits.h>
35 #include <sys/syslog.h>
36 #include <sys/systm.h>
37 #include <sys/socket.h>
39 #include <machine/atomic.h>
42 #include <net/route.h>
43 #include <netinet/in.h>
44 #include <netinet/in_systm.h>
45 #include <netinet/ip.h>
46 #include <netinet/tcp.h>
47 #include <netinet/udp.h>
49 #include "ng_message.h"
52 #include "netflow/netflow.h"
53 #include "netflow/ng_netflow.h"
55 #define NBUCKETS (65536) /* must be power of 2 */
57 /* This hash is for TCP or UDP packets. */
58 #define FULL_HASH(addr1, addr2, port1, port2) \
59 (((addr1 ^ (addr1 >> 16) ^ \
60 htons(addr2 ^ (addr2 >> 16))) ^ \
61 port1 ^ htons(port2)) & \
64 /* This hash is for all other IP packets. */
65 #define ADDR_HASH(addr1, addr2) \
66 ((addr1 ^ (addr1 >> 16) ^ \
67 htons(addr2 ^ (addr2 >> 16))) & \
70 /* Macros to shorten logical constructions */
71 /* XXX: priv must exist in namespace */
72 #define INACTIVE(fle) (time_uptime - fle->f.last > priv->info.nfinfo_inact_t)
73 #define AGED(fle) (time_uptime - fle->f.first > priv->info.nfinfo_act_t)
74 #define ISFREE(fle) (fle->f.packets == 0)
77 * 4 is a magical number: statistically number of 4-packet flows is
78 * bigger than 5,6,7...-packet flows by an order of magnitude. Most UDP/ICMP
79 * scans are 1 packet (~ 90% of flow cache). TCP scans are 2-packet in case
80 * of reachable host and 4-packet otherwise.
82 #define SMALL(fle) (fle->f.packets <= 4)
85 * Cisco uses milliseconds for uptime. Bad idea, since it overflows
86 * every 48+ days. But we will do same to keep compatibility. This macro
87 * does overflowable multiplication to 1000.
89 #define MILLIUPTIME(t) (((t) << 9) + /* 512 */ \
90 ((t) << 8) + /* 256 */ \
91 ((t) << 7) + /* 128 */ \
92 ((t) << 6) + /* 64 */ \
93 ((t) << 5) + /* 32 */ \
96 MALLOC_DECLARE(M_NETFLOW_HASH);
97 MALLOC_DEFINE(M_NETFLOW_HASH, "netflow_hash", "NetFlow hash");
99 static int export_add(item_p, struct flow_entry *);
100 static int export_send(priv_p, item_p, int flags);
102 /* Generate hash for a given flow record. */
103 static __inline uint32_t
104 ip_hash(struct flow_rec *r)
109 return FULL_HASH(r->r_src.s_addr, r->r_dst.s_addr,
110 r->r_sport, r->r_dport);
112 return ADDR_HASH(r->r_src.s_addr, r->r_dst.s_addr);
116 /* This is callback from uma(9), called on alloc. */
118 uma_ctor_flow(void *mem, int size, void *arg, int how)
120 priv_p priv = (priv_p )arg;
122 if (atomic_load_acq_32(&priv->info.nfinfo_used) >= CACHESIZE)
125 atomic_add_32(&priv->info.nfinfo_used, 1);
130 /* This is callback from uma(9), called on free. */
132 uma_dtor_flow(void *mem, int size, void *arg)
134 priv_p priv = (priv_p )arg;
136 atomic_subtract_32(&priv->info.nfinfo_used, 1);
140 * Detach export datagram from priv, if there is any.
141 * If there is no, allocate a new one.
144 get_export_dgram(priv_p priv)
148 mtx_lock(&priv->export_mtx);
149 if (priv->export_item != NULL) {
150 item = priv->export_item;
151 priv->export_item = NULL;
153 mtx_unlock(&priv->export_mtx);
156 struct netflow_v5_export_dgram *dgram;
159 m = m_getcl(MB_DONTWAIT, MT_DATA, M_PKTHDR);
162 item = ng_package_data(m, NG_NOFLAGS);
165 dgram = mtod(m, struct netflow_v5_export_dgram *);
166 dgram->header.count = 0;
167 dgram->header.version = htons(NETFLOW_V5);
175 * Re-attach incomplete datagram back to priv.
176 * If there is already another one, then send incomplete. */
178 return_export_dgram(priv_p priv, item_p item, int flags)
181 * It may happen on SMP, that some thread has already
182 * put its item there, in this case we bail out and
183 * send what we have to collector.
185 mtx_lock(&priv->export_mtx);
186 if (priv->export_item == NULL) {
187 priv->export_item = item;
188 mtx_unlock(&priv->export_mtx);
190 mtx_unlock(&priv->export_mtx);
191 export_send(priv, item, flags);
196 * The flow is over. Call export_add() and free it. If datagram is
197 * full, then call export_send().
200 expire_flow(priv_p priv, item_p *item, struct flow_entry *fle, int flags)
203 *item = get_export_dgram(priv);
205 atomic_add_32(&priv->info.nfinfo_export_failed, 1);
206 uma_zfree_arg(priv->zone, fle, priv);
209 if (export_add(*item, fle) > 0) {
210 export_send(priv, *item, flags);
213 uma_zfree_arg(priv->zone, fle, priv);
216 /* Get a snapshot of node statistics */
218 ng_netflow_copyinfo(priv_p priv, struct ng_netflow_info *i)
221 memcpy((void *)i, (void *)&priv->info, sizeof(priv->info));
225 * Insert a record into defined slot.
227 * First we get for us a free flow entry, then fill in all
228 * possible fields in it.
230 * TODO: consider dropping hash mutex while filling in datagram,
231 * as this was done in previous version. Need to test & profile
235 hash_insert(priv_p priv, struct flow_hash_entry *hsh, struct flow_rec *r,
236 int plen, uint8_t tcp_flags)
238 struct flow_entry *fle;
239 struct sockaddr_in sin;
242 KKASSERT(mtx_owned(&hsh->mtx));
244 fle = uma_zalloc_arg(priv->zone, priv, M_WAITOK | M_NULLOK);
246 atomic_add_32(&priv->info.nfinfo_alloc_failed, 1);
251 * Now fle is totally ours. It is detached from all lists,
252 * we can safely edit it.
255 bcopy(r, &fle->f.r, sizeof(struct flow_rec));
258 fle->f.tcp_flags = tcp_flags;
260 fle->f.first = fle->f.last = time_uptime;
263 * First we do route table lookup on destination address. So we can
264 * fill in out_ifx, dst_mask, nexthop, and dst_as in future releases.
266 bzero(&sin, sizeof(sin));
267 sin.sin_len = sizeof(struct sockaddr_in);
268 sin.sin_family = AF_INET;
269 sin.sin_addr = fle->f.r.r_dst;
270 /* XXX MRT 0 as a default.. need the m here to get fib */
271 rt = rtalloc1_fib((struct sockaddr *)&sin, 0, RTF_CLONING, 0);
273 fle->f.fle_o_ifx = rt->rt_ifp->if_index;
275 if (rt->rt_flags & RTF_GATEWAY &&
276 rt->rt_gateway->sa_family == AF_INET)
278 ((struct sockaddr_in *)(rt->rt_gateway))->sin_addr;
281 fle->f.dst_mask = bitcount32(((struct sockaddr_in *)
282 rt_mask(rt))->sin_addr.s_addr);
283 else if (rt->rt_flags & RTF_HOST)
284 /* Give up. We can't determine mask :( */
285 fle->f.dst_mask = 32;
290 /* Do route lookup on source address, to fill in src_mask. */
291 bzero(&sin, sizeof(sin));
292 sin.sin_len = sizeof(struct sockaddr_in);
293 sin.sin_family = AF_INET;
294 sin.sin_addr = fle->f.r.r_src;
295 /* XXX MRT 0 as a default revisit. need the mbuf for fib*/
296 rt = rtalloc1_fib((struct sockaddr *)&sin, 0, RTF_CLONING, 0);
299 fle->f.src_mask = bitcount32(((struct sockaddr_in *)
300 rt_mask(rt))->sin_addr.s_addr);
301 else if (rt->rt_flags & RTF_HOST)
302 /* Give up. We can't determine mask :( */
303 fle->f.src_mask = 32;
308 /* Push new flow at the and of hash. */
309 TAILQ_INSERT_TAIL(&hsh->head, fle, fle_hash);
316 * Non-static functions called from ng_netflow.c
319 /* Allocate memory and set up flow cache */
321 ng_netflow_cache_init(priv_p priv)
323 struct flow_hash_entry *hsh;
326 /* Initialize cache UMA zone. */
327 priv->zone = uma_zcreate("NetFlow cache", sizeof(struct flow_entry),
328 uma_ctor_flow, uma_dtor_flow, NULL, NULL, UMA_ALIGN_CACHE, 0);
329 uma_zone_set_max(priv->zone, CACHESIZE);
332 priv->hash = kmalloc(NBUCKETS * sizeof(struct flow_hash_entry),
333 M_NETFLOW_HASH, M_WAITOK | M_ZERO);
335 /* Initialize hash. */
336 for (i = 0, hsh = priv->hash; i < NBUCKETS; i++, hsh++) {
337 mtx_init(&hsh->mtx, "hash mutex", NULL, MTX_DEF);
338 TAILQ_INIT(&hsh->head);
341 mtx_init(&priv->export_mtx, "export dgram lock", NULL, MTX_DEF);
346 /* Free all flow cache memory. Called from node close method. */
348 ng_netflow_cache_flush(priv_p priv)
350 struct flow_entry *fle, *fle1;
351 struct flow_hash_entry *hsh;
356 * We are going to free probably billable data.
357 * Expire everything before freeing it.
358 * No locking is required since callout is already drained.
360 for (hsh = priv->hash, i = 0; i < NBUCKETS; hsh++, i++)
361 TAILQ_FOREACH_SAFE(fle, &hsh->head, fle_hash, fle1) {
362 TAILQ_REMOVE(&hsh->head, fle, fle_hash);
363 expire_flow(priv, &item, fle, NG_QUEUE);
367 export_send(priv, item, NG_QUEUE);
369 uma_zdestroy(priv->zone);
371 /* Destroy hash mutexes. */
372 for (i = 0, hsh = priv->hash; i < NBUCKETS; i++, hsh++)
373 mtx_destroy(&hsh->mtx);
375 /* Free hash memory. */
377 kfree(priv->hash, M_NETFLOW_HASH);
379 mtx_destroy(&priv->export_mtx);
382 /* Insert packet from into flow cache. */
384 ng_netflow_flow_add(priv_p priv, struct ip *ip, iface_p iface,
387 register struct flow_entry *fle, *fle1;
388 struct flow_hash_entry *hsh;
393 uint8_t tcp_flags = 0;
395 /* Try to fill flow_rec r */
396 bzero(&r, sizeof(r));
398 if (ip->ip_v != IPVERSION)
401 /* verify min header length */
402 hlen = ip->ip_hl << 2;
404 if (hlen < sizeof(struct ip))
407 r.r_src = ip->ip_src;
408 r.r_dst = ip->ip_dst;
410 /* save packet length */
411 plen = ntohs(ip->ip_len);
414 r.r_tos = ip->ip_tos;
416 /* Configured in_ifx overrides mbuf's */
417 if (iface->info.ifinfo_index == 0) {
419 r.r_i_ifx = ifp->if_index;
421 r.r_i_ifx = iface->info.ifinfo_index;
424 * XXX NOTE: only first fragment of fragmented TCP, UDP and
425 * ICMP packet will be recorded with proper s_port and d_port.
426 * Following fragments will be recorded simply as IP packet with
427 * ip_proto = ip->ip_p and s_port, d_port set to zero.
428 * I know, it looks like bug. But I don't want to re-implement
429 * ip packet assebmling here. Anyway, (in)famous trafd works this way -
430 * and nobody complains yet :)
432 if ((ip->ip_off & htons(IP_OFFMASK)) == 0)
436 register struct tcphdr *tcp;
438 tcp = (struct tcphdr *)((caddr_t )ip + hlen);
439 r.r_sport = tcp->th_sport;
440 r.r_dport = tcp->th_dport;
441 tcp_flags = tcp->th_flags;
445 r.r_ports = *(uint32_t *)((caddr_t )ip + hlen);
449 /* Update node statistics. XXX: race... */
450 priv->info.nfinfo_packets ++;
451 priv->info.nfinfo_bytes += plen;
453 /* Find hash slot. */
454 hsh = &priv->hash[ip_hash(&r)];
459 * Go through hash and find our entry. If we encounter an
460 * entry, that should be expired, purge it. We do a reverse
461 * search since most active entries are first, and most
462 * searches are done on most active entries.
464 TAILQ_FOREACH_REVERSE_SAFE(fle, &hsh->head, fhead, fle_hash, fle1) {
465 if (bcmp(&r, &fle->f.r, sizeof(struct flow_rec)) == 0)
467 if ((INACTIVE(fle) && SMALL(fle)) || AGED(fle)) {
468 TAILQ_REMOVE(&hsh->head, fle, fle_hash);
469 expire_flow(priv, &item, fle, NG_QUEUE);
470 atomic_add_32(&priv->info.nfinfo_act_exp, 1);
474 if (fle) { /* An existent entry. */
476 fle->f.bytes += plen;
478 fle->f.tcp_flags |= tcp_flags;
479 fle->f.last = time_uptime;
482 * We have the following reasons to expire flow in active way:
483 * - it hit active timeout
484 * - a TCP connection closed
485 * - it is going to overflow counter
487 if (tcp_flags & TH_FIN || tcp_flags & TH_RST || AGED(fle) ||
488 (fle->f.bytes >= (UINT_MAX - IF_MAXMTU)) ) {
489 TAILQ_REMOVE(&hsh->head, fle, fle_hash);
490 expire_flow(priv, &item, fle, NG_QUEUE);
491 atomic_add_32(&priv->info.nfinfo_act_exp, 1);
494 * It is the newest, move it to the tail,
495 * if it isn't there already. Next search will
498 if (fle != TAILQ_LAST(&hsh->head, fhead)) {
499 TAILQ_REMOVE(&hsh->head, fle, fle_hash);
500 TAILQ_INSERT_TAIL(&hsh->head, fle, fle_hash);
503 } else /* A new flow entry. */
504 error = hash_insert(priv, hsh, &r, plen, tcp_flags);
506 mtx_unlock(&hsh->mtx);
509 return_export_dgram(priv, item, NG_QUEUE);
515 * Return records from cache to userland.
517 * TODO: matching particular IP should be done in kernel, here.
520 ng_netflow_flow_show(priv_p priv, uint32_t last, struct ng_mesg *resp)
522 struct flow_hash_entry *hsh;
523 struct flow_entry *fle;
524 struct ngnf_flows *data;
527 data = (struct ngnf_flows *)resp->data;
531 /* Check if this is a first run */
536 if (last > NBUCKETS-1)
538 hsh = priv->hash + last;
543 * We will transfer not more than NREC_AT_ONCE. More data
544 * will come in next message.
545 * We send current hash index to userland, and userland should
546 * return it back to us. Then, we will restart with new entry.
548 * The resulting cache snapshot is inaccurate for the
550 * - we skip locked hash entries
551 * - we bail out, if someone wants our entry
552 * - we skip rest of entry, when hit NREC_AT_ONCE
554 for (; i < NBUCKETS; hsh++, i++) {
555 if (mtx_trylock(&hsh->mtx) == 0)
558 TAILQ_FOREACH(fle, &hsh->head, fle_hash) {
559 if (mtx_contested(&hsh->mtx)
562 bcopy(&fle->f, &(data->entries[data->nentries]),
565 if (data->nentries == NREC_AT_ONCE) {
566 mtx_unlock(&hsh->mtx);
572 mtx_unlock(&hsh->mtx);
578 /* We have full datagram in privdata. Send it to export hook. */
580 export_send(priv_p priv, item_p item, int flags)
582 struct mbuf *m = NGI_M(item);
583 struct netflow_v5_export_dgram *dgram = mtod(m,
584 struct netflow_v5_export_dgram *);
585 struct netflow_v5_header *header = &dgram->header;
589 /* Fill mbuf header. */
590 m->m_len = m->m_pkthdr.len = sizeof(struct netflow_v5_record) *
591 header->count + sizeof(struct netflow_v5_header);
593 /* Fill export header. */
594 header->sys_uptime = htonl(MILLIUPTIME(time_uptime));
596 header->unix_secs = htonl(ts.tv_sec);
597 header->unix_nsecs = htonl(ts.tv_nsec);
598 header->engine_type = 0;
599 header->engine_id = 0;
601 header->flow_seq = htonl(atomic_fetchadd_32(&priv->flow_seq,
603 header->count = htons(header->count);
605 if (priv->export != NULL)
606 NG_FWD_ITEM_HOOK_FLAGS(error, item, priv->export, flags);
614 /* Add export record to dgram. */
616 export_add(item_p item, struct flow_entry *fle)
618 struct netflow_v5_export_dgram *dgram = mtod(NGI_M(item),
619 struct netflow_v5_export_dgram *);
620 struct netflow_v5_header *header = &dgram->header;
621 struct netflow_v5_record *rec;
623 rec = &dgram->r[header->count];
626 KASSERT(header->count <= NETFLOW_V5_MAX_RECORDS,
627 ("ng_netflow: export too big"));
629 /* Fill in export record. */
630 rec->src_addr = fle->f.r.r_src.s_addr;
631 rec->dst_addr = fle->f.r.r_dst.s_addr;
632 rec->next_hop = fle->f.next_hop.s_addr;
633 rec->i_ifx = htons(fle->f.fle_i_ifx);
634 rec->o_ifx = htons(fle->f.fle_o_ifx);
635 rec->packets = htonl(fle->f.packets);
636 rec->octets = htonl(fle->f.bytes);
637 rec->first = htonl(MILLIUPTIME(fle->f.first));
638 rec->last = htonl(MILLIUPTIME(fle->f.last));
639 rec->s_port = fle->f.r.r_sport;
640 rec->d_port = fle->f.r.r_dport;
641 rec->flags = fle->f.tcp_flags;
642 rec->prot = fle->f.r.r_ip_p;
643 rec->tos = fle->f.r.r_tos;
644 rec->dst_mask = fle->f.dst_mask;
645 rec->src_mask = fle->f.src_mask;
647 /* Not supported fields. */
648 rec->src_as = rec->dst_as = 0;
650 if (header->count == NETFLOW_V5_MAX_RECORDS)
651 return (1); /* end of datagram */
656 /* Periodic flow expiry run. */
658 ng_netflow_expire(void *arg)
660 struct flow_entry *fle, *fle1;
661 struct flow_hash_entry *hsh;
662 priv_p priv = (priv_p )arg;
668 * Going through all the cache.
670 for (hsh = priv->hash, i = 0; i < NBUCKETS; hsh++, i++) {
672 * Skip entries, that are already being worked on.
674 if (mtx_trylock(&hsh->mtx) == 0)
677 used = atomic_load_acq_32(&priv->info.nfinfo_used);
678 TAILQ_FOREACH_SAFE(fle, &hsh->head, fle_hash, fle1) {
680 * Interrupt thread wants this entry!
681 * Quick! Quick! Bail out!
683 if (mtx_contested(&hsh->mtx))
687 * Don't expire aggressively while hash collision
688 * ratio is predicted small.
690 if (used <= (NBUCKETS*2) && !INACTIVE(fle))
693 if ((INACTIVE(fle) && (SMALL(fle) ||
694 (used > (NBUCKETS*2)))) || AGED(fle)) {
695 TAILQ_REMOVE(&hsh->head, fle, fle_hash);
696 expire_flow(priv, &item, fle, NG_NOFLAGS);
698 atomic_add_32(&priv->info.nfinfo_inact_exp, 1);
701 mtx_unlock(&hsh->mtx);
705 return_export_dgram(priv, item, NG_NOFLAGS);
707 /* Schedule next expire. */
708 callout_reset(&priv->exp_callout, (1*hz), &ng_netflow_expire,