1 /*-
2 * Copyright (c) 2014 Gleb Smirnoff <glebius@FreeBSD.org>
3 * Copyright (c) 2008-2010, BitGravity Inc.
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 are met:
8 *
9 * 1. Redistributions of source code must retain the above copyright notice,
10 * this list of conditions and the following disclaimer.
11 *
12 * 2. Neither the name of the BitGravity Corporation nor the names of its
13 * contributors may be used to endorse or promote products derived from
14 * this software without specific prior written permission.
15 *
16 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
17 * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19 * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
20 * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
21 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
22 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
23 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
24 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
25 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
26 * POSSIBILITY OF SUCH DAMAGE.
27 */
28
29 #include "opt_route.h"
30 #include "opt_mpath.h"
31 #include "opt_ddb.h"
32 #include "opt_inet.h"
33 #include "opt_inet6.h"
34
35 #include <sys/cdefs.h>
36 __FBSDID("$FreeBSD$");
37
38 #include <sys/param.h>
39 #include <sys/types.h>
40 #include <sys/bitstring.h>
41 #include <sys/condvar.h>
42 #include <sys/callout.h>
43 #include <sys/hash.h>
44 #include <sys/kernel.h>
45 #include <sys/kthread.h>
46 #include <sys/limits.h>
47 #include <sys/malloc.h>
48 #include <sys/mbuf.h>
49 #include <sys/pcpu.h>
50 #include <sys/proc.h>
51 #include <sys/queue.h>
52 #include <sys/sbuf.h>
53 #include <sys/sched.h>
54 #include <sys/smp.h>
55 #include <sys/socket.h>
56 #include <sys/syslog.h>
57 #include <sys/sysctl.h>
58 #include <vm/uma.h>
59
60 #include <net/if.h>
61 #include <net/if_llatbl.h>
62 #include <net/if_var.h>
63 #include <net/route.h>
64 #include <net/flowtable.h>
65 #include <net/vnet.h>
66
67 #include <netinet/in.h>
68 #include <netinet/in_systm.h>
69 #include <netinet/in_var.h>
70 #include <netinet/if_ether.h>
71 #include <netinet/ip.h>
72 #ifdef INET6
73 #include <netinet/ip6.h>
74 #endif
75 #ifdef FLOWTABLE_HASH_ALL
76 #include <netinet/tcp.h>
77 #include <netinet/udp.h>
78 #include <netinet/sctp.h>
79 #endif
80
81 #include <ddb/ddb.h>
82
83 #ifdef FLOWTABLE_HASH_ALL
84 #define KEY_PORTS (sizeof(uint16_t) * 2)
85 #define KEY_ADDRS 2
86 #else
87 #define KEY_PORTS 0
88 #define KEY_ADDRS 1
89 #endif
90
91 #ifdef INET6
92 #define KEY_ADDR_LEN sizeof(struct in6_addr)
93 #else
94 #define KEY_ADDR_LEN sizeof(struct in_addr)
95 #endif
96
97 #define KEYLEN ((KEY_ADDR_LEN * KEY_ADDRS + KEY_PORTS) / sizeof(uint32_t))
98
99 struct flentry {
100 uint32_t f_hash; /* hash flowing forward */
101 uint32_t f_key[KEYLEN]; /* address(es and ports) */
102 uint32_t f_uptime; /* uptime at last access */
103 uint16_t f_fibnum; /* fib index */
104 #ifdef FLOWTABLE_HASH_ALL
105 uint8_t f_proto; /* protocol */
106 uint8_t f_flags; /* stale? */
107 #define FL_STALE 1
108 #endif
109 SLIST_ENTRY(flentry) f_next; /* pointer to collision entry */
110 struct rtentry *f_rt; /* rtentry for flow */
111 struct llentry *f_lle; /* llentry for flow */
112 };
113 #undef KEYLEN
114
115 SLIST_HEAD(flist, flentry);
116 /* Make sure we can use pcpu_zone_ptr for struct flist. */
117 CTASSERT(sizeof(struct flist) == sizeof(void *));
118
119 struct flowtable {
120 counter_u64_t *ft_stat;
121 int ft_size;
122 /*
123 * ft_table is a malloc(9)ed array of pointers. Pointers point to
124 * memory from UMA_ZONE_PCPU zone.
125 * ft_masks is per-cpu pointer itself. Each instance points
126 * to a malloc(9)ed bitset, that is private to corresponding CPU.
127 */
128 struct flist **ft_table;
129 bitstr_t **ft_masks;
130 bitstr_t *ft_tmpmask;
131 };
132
133 #define FLOWSTAT_ADD(ft, name, v) \
134 counter_u64_add((ft)->ft_stat[offsetof(struct flowtable_stat, name) / sizeof(uint64_t)], (v))
135 #define FLOWSTAT_INC(ft, name) FLOWSTAT_ADD(ft, name, 1)
136
137 static struct proc *flowcleanerproc;
138 static uint32_t flow_hashjitter;
139
140 static struct cv flowclean_f_cv;
141 static struct cv flowclean_c_cv;
142 static struct mtx flowclean_lock;
143 static uint32_t flowclean_cycles;
144
145 /*
146 * TODO:
147 * - add sysctls to resize && flush flow tables
148 * - Add per flowtable sysctls for statistics and configuring timeouts
149 * - add saturation counter to rtentry to support per-packet load-balancing
150 * add flag to indicate round-robin flow, add list lookup from head
151 for flows
152 * - add sysctl / device node / syscall to support exporting and importing
153 * of flows with flag to indicate that a flow was imported so should
154 * not be considered for auto-cleaning
155 * - support explicit connection state (currently only ad-hoc for DSR)
156 * - idetach() cleanup for options VIMAGE builds.
157 */
158 #ifdef INET
159 static VNET_DEFINE(struct flowtable, ip4_ft);
160 #define V_ip4_ft VNET(ip4_ft)
161 #endif
162 #ifdef INET6
163 static VNET_DEFINE(struct flowtable, ip6_ft);
164 #define V_ip6_ft VNET(ip6_ft)
165 #endif
166
167 static uma_zone_t flow_zone;
168
169 static VNET_DEFINE(int, flowtable_enable) = 1;
170 #define V_flowtable_enable VNET(flowtable_enable)
171
172 static SYSCTL_NODE(_net, OID_AUTO, flowtable, CTLFLAG_RD, NULL,
173 "flowtable");
174 SYSCTL_INT(_net_flowtable, OID_AUTO, enable, CTLFLAG_VNET | CTLFLAG_RW,
175 &VNET_NAME(flowtable_enable), 0, "enable flowtable caching.");
176 SYSCTL_UMA_MAX(_net_flowtable, OID_AUTO, maxflows, CTLFLAG_RW,
177 &flow_zone, "Maximum number of flows allowed");
178
179 static MALLOC_DEFINE(M_FTABLE, "flowtable", "flowtable hashes and bitstrings");
180
181 static struct flentry *
182 flowtable_lookup_common(struct flowtable *, uint32_t *, int, uint32_t);
183
184 #ifdef INET
185 static struct flentry *
flowtable_lookup_ipv4(struct mbuf * m,struct route * ro)186 flowtable_lookup_ipv4(struct mbuf *m, struct route *ro)
187 {
188 struct flentry *fle;
189 struct sockaddr_in *sin;
190 struct ip *ip;
191 uint32_t fibnum;
192 #ifdef FLOWTABLE_HASH_ALL
193 uint32_t key[3];
194 int iphlen;
195 uint16_t sport, dport;
196 uint8_t proto;
197 #endif
198
199 ip = mtod(m, struct ip *);
200
201 if (ip->ip_src.s_addr == ip->ip_dst.s_addr ||
202 (ntohl(ip->ip_dst.s_addr) >> IN_CLASSA_NSHIFT) == IN_LOOPBACKNET ||
203 (ntohl(ip->ip_src.s_addr) >> IN_CLASSA_NSHIFT) == IN_LOOPBACKNET)
204 return (NULL);
205
206 fibnum = M_GETFIB(m);
207
208 #ifdef FLOWTABLE_HASH_ALL
209 iphlen = ip->ip_hl << 2;
210 proto = ip->ip_p;
211
212 switch (proto) {
213 case IPPROTO_TCP: {
214 struct tcphdr *th;
215
216 th = (struct tcphdr *)((char *)ip + iphlen);
217 sport = th->th_sport;
218 dport = th->th_dport;
219 if (th->th_flags & (TH_RST|TH_FIN))
220 fibnum |= (FL_STALE << 24);
221 break;
222 }
223 case IPPROTO_UDP: {
224 struct udphdr *uh;
225
226 uh = (struct udphdr *)((char *)ip + iphlen);
227 sport = uh->uh_sport;
228 dport = uh->uh_dport;
229 break;
230 }
231 case IPPROTO_SCTP: {
232 struct sctphdr *sh;
233
234 sh = (struct sctphdr *)((char *)ip + iphlen);
235 sport = sh->src_port;
236 dport = sh->dest_port;
237 /* XXXGL: handle stale? */
238 break;
239 }
240 default:
241 sport = dport = 0;
242 break;
243 }
244
245 key[0] = ip->ip_dst.s_addr;
246 key[1] = ip->ip_src.s_addr;
247 key[2] = (dport << 16) | sport;
248 fibnum |= proto << 16;
249
250 fle = flowtable_lookup_common(&V_ip4_ft, key, 3 * sizeof(uint32_t),
251 fibnum);
252
253 #else /* !FLOWTABLE_HASH_ALL */
254
255 fle = flowtable_lookup_common(&V_ip4_ft, (uint32_t *)&ip->ip_dst,
256 sizeof(struct in_addr), fibnum);
257
258 #endif /* FLOWTABLE_HASH_ALL */
259
260 if (fle == NULL)
261 return (NULL);
262
263 sin = (struct sockaddr_in *)&ro->ro_dst;
264 sin->sin_family = AF_INET;
265 sin->sin_len = sizeof(*sin);
266 sin->sin_addr = ip->ip_dst;
267
268 return (fle);
269 }
270 #endif /* INET */
271
272 #ifdef INET6
273 /*
274 * PULLUP_TO(len, p, T) makes sure that len + sizeof(T) is contiguous,
275 * then it sets p to point at the offset "len" in the mbuf. WARNING: the
276 * pointer might become stale after other pullups (but we never use it
277 * this way).
278 */
279 #define PULLUP_TO(_len, p, T) \
280 do { \
281 int x = (_len) + sizeof(T); \
282 if ((m)->m_len < x) \
283 return (NULL); \
284 p = (mtod(m, char *) + (_len)); \
285 } while (0)
286
287 #define TCP(p) ((struct tcphdr *)(p))
288 #define SCTP(p) ((struct sctphdr *)(p))
289 #define UDP(p) ((struct udphdr *)(p))
290
291 static struct flentry *
flowtable_lookup_ipv6(struct mbuf * m,struct route * ro)292 flowtable_lookup_ipv6(struct mbuf *m, struct route *ro)
293 {
294 struct flentry *fle;
295 struct sockaddr_in6 *sin6;
296 struct ip6_hdr *ip6;
297 uint32_t fibnum;
298 #ifdef FLOWTABLE_HASH_ALL
299 uint32_t key[9];
300 void *ulp;
301 int hlen;
302 uint16_t sport, dport;
303 u_short offset;
304 uint8_t proto;
305 #else
306 uint32_t key[4];
307 #endif
308
309 ip6 = mtod(m, struct ip6_hdr *);
310 if (in6_localaddr(&ip6->ip6_dst))
311 return (NULL);
312
313 fibnum = M_GETFIB(m);
314
315 #ifdef FLOWTABLE_HASH_ALL
316 hlen = sizeof(struct ip6_hdr);
317 proto = ip6->ip6_nxt;
318 offset = sport = dport = 0;
319 ulp = NULL;
320 while (ulp == NULL) {
321 switch (proto) {
322 case IPPROTO_ICMPV6:
323 case IPPROTO_OSPFIGP:
324 case IPPROTO_PIM:
325 case IPPROTO_CARP:
326 case IPPROTO_ESP:
327 case IPPROTO_NONE:
328 ulp = ip6;
329 break;
330 case IPPROTO_TCP:
331 PULLUP_TO(hlen, ulp, struct tcphdr);
332 dport = TCP(ulp)->th_dport;
333 sport = TCP(ulp)->th_sport;
334 if (TCP(ulp)->th_flags & (TH_RST|TH_FIN))
335 fibnum |= (FL_STALE << 24);
336 break;
337 case IPPROTO_SCTP:
338 PULLUP_TO(hlen, ulp, struct sctphdr);
339 dport = SCTP(ulp)->src_port;
340 sport = SCTP(ulp)->dest_port;
341 /* XXXGL: handle stale? */
342 break;
343 case IPPROTO_UDP:
344 PULLUP_TO(hlen, ulp, struct udphdr);
345 dport = UDP(ulp)->uh_dport;
346 sport = UDP(ulp)->uh_sport;
347 break;
348 case IPPROTO_HOPOPTS: /* RFC 2460 */
349 PULLUP_TO(hlen, ulp, struct ip6_hbh);
350 hlen += (((struct ip6_hbh *)ulp)->ip6h_len + 1) << 3;
351 proto = ((struct ip6_hbh *)ulp)->ip6h_nxt;
352 ulp = NULL;
353 break;
354 case IPPROTO_ROUTING: /* RFC 2460 */
355 PULLUP_TO(hlen, ulp, struct ip6_rthdr);
356 hlen += (((struct ip6_rthdr *)ulp)->ip6r_len + 1) << 3;
357 proto = ((struct ip6_rthdr *)ulp)->ip6r_nxt;
358 ulp = NULL;
359 break;
360 case IPPROTO_FRAGMENT: /* RFC 2460 */
361 PULLUP_TO(hlen, ulp, struct ip6_frag);
362 hlen += sizeof (struct ip6_frag);
363 proto = ((struct ip6_frag *)ulp)->ip6f_nxt;
364 offset = ((struct ip6_frag *)ulp)->ip6f_offlg &
365 IP6F_OFF_MASK;
366 ulp = NULL;
367 break;
368 case IPPROTO_DSTOPTS: /* RFC 2460 */
369 PULLUP_TO(hlen, ulp, struct ip6_hbh);
370 hlen += (((struct ip6_hbh *)ulp)->ip6h_len + 1) << 3;
371 proto = ((struct ip6_hbh *)ulp)->ip6h_nxt;
372 ulp = NULL;
373 break;
374 case IPPROTO_AH: /* RFC 2402 */
375 PULLUP_TO(hlen, ulp, struct ip6_ext);
376 hlen += (((struct ip6_ext *)ulp)->ip6e_len + 2) << 2;
377 proto = ((struct ip6_ext *)ulp)->ip6e_nxt;
378 ulp = NULL;
379 break;
380 default:
381 PULLUP_TO(hlen, ulp, struct ip6_ext);
382 break;
383 }
384 }
385
386 bcopy(&ip6->ip6_dst, &key[0], sizeof(struct in6_addr));
387 bcopy(&ip6->ip6_src, &key[4], sizeof(struct in6_addr));
388 key[8] = (dport << 16) | sport;
389 fibnum |= proto << 16;
390
391 fle = flowtable_lookup_common(&V_ip6_ft, key, 9 * sizeof(uint32_t),
392 fibnum);
393 #else /* !FLOWTABLE_HASH_ALL */
394 bcopy(&ip6->ip6_dst, &key[0], sizeof(struct in6_addr));
395 fle = flowtable_lookup_common(&V_ip6_ft, key, sizeof(struct in6_addr),
396 fibnum);
397 #endif /* FLOWTABLE_HASH_ALL */
398
399 if (fle == NULL)
400 return (NULL);
401
402 sin6 = (struct sockaddr_in6 *)&ro->ro_dst;
403 sin6->sin6_family = AF_INET6;
404 sin6->sin6_len = sizeof(*sin6);
405 bcopy(&ip6->ip6_dst, &sin6->sin6_addr, sizeof(struct in6_addr));
406
407 return (fle);
408 }
409 #endif /* INET6 */
410
411 static bitstr_t *
flowtable_mask(struct flowtable * ft)412 flowtable_mask(struct flowtable *ft)
413 {
414
415 /*
416 * flowtable_free_stale() calls w/o critical section, but
417 * with sched_bind(). Since pointer is stable throughout
418 * ft lifetime, it is safe, otherwise...
419 *
420 * CRITICAL_ASSERT(curthread);
421 */
422
423 return (*(bitstr_t **)zpcpu_get(ft->ft_masks));
424 }
425
426 static struct flist *
flowtable_list(struct flowtable * ft,uint32_t hash)427 flowtable_list(struct flowtable *ft, uint32_t hash)
428 {
429
430 CRITICAL_ASSERT(curthread);
431 return (zpcpu_get(ft->ft_table[hash % ft->ft_size]));
432 }
433
434 static int
flow_stale(struct flowtable * ft,struct flentry * fle,int maxidle)435 flow_stale(struct flowtable *ft, struct flentry *fle, int maxidle)
436 {
437
438 if (((fle->f_rt->rt_flags & RTF_UP) == 0) ||
439 (fle->f_rt->rt_ifp == NULL) ||
440 !RT_LINK_IS_UP(fle->f_rt->rt_ifp) ||
441 (fle->f_lle->la_flags & LLE_VALID) == 0)
442 return (1);
443
444 if (time_uptime - fle->f_uptime > maxidle)
445 return (1);
446
447 #ifdef FLOWTABLE_HASH_ALL
448 if (fle->f_flags & FL_STALE)
449 return (1);
450 #endif
451
452 return (0);
453 }
454
455 static int
flow_full(void)456 flow_full(void)
457 {
458 int count, max;
459
460 count = uma_zone_get_cur(flow_zone);
461 max = uma_zone_get_max(flow_zone);
462
463 return (count > (max - (max >> 3)));
464 }
465
466 static int
flow_matches(struct flentry * fle,uint32_t * key,int keylen,uint32_t fibnum)467 flow_matches(struct flentry *fle, uint32_t *key, int keylen, uint32_t fibnum)
468 {
469 #ifdef FLOWTABLE_HASH_ALL
470 uint8_t proto;
471
472 proto = (fibnum >> 16) & 0xff;
473 fibnum &= 0xffff;
474 #endif
475
476 CRITICAL_ASSERT(curthread);
477
478 /* Microoptimization for IPv4: don't use bcmp(). */
479 if (((keylen == sizeof(uint32_t) && (fle->f_key[0] == key[0])) ||
480 (bcmp(fle->f_key, key, keylen) == 0)) &&
481 fibnum == fle->f_fibnum &&
482 #ifdef FLOWTABLE_HASH_ALL
483 proto == fle->f_proto &&
484 #endif
485 (fle->f_rt->rt_flags & RTF_UP) &&
486 fle->f_rt->rt_ifp != NULL &&
487 (fle->f_lle->la_flags & LLE_VALID))
488 return (1);
489
490 return (0);
491 }
492
493 static struct flentry *
flowtable_insert(struct flowtable * ft,uint32_t hash,uint32_t * key,int keylen,uint32_t fibnum0)494 flowtable_insert(struct flowtable *ft, uint32_t hash, uint32_t *key,
495 int keylen, uint32_t fibnum0)
496 {
497 #ifdef INET6
498 struct route_in6 sro6;
499 #endif
500 #ifdef INET
501 struct route sro;
502 #endif
503 struct route *ro = NULL;
504 struct rtentry *rt;
505 struct lltable *lt = NULL;
506 struct llentry *lle;
507 struct sockaddr_storage *l3addr;
508 struct ifnet *ifp;
509 struct flist *flist;
510 struct flentry *fle, *iter;
511 bitstr_t *mask;
512 uint16_t fibnum = fibnum0;
513 #ifdef FLOWTABLE_HASH_ALL
514 uint8_t proto;
515
516 proto = (fibnum0 >> 16) & 0xff;
517 fibnum = fibnum0 & 0xffff;
518 #endif
519
520 /*
521 * This bit of code ends up locking the
522 * same route 3 times (just like ip_output + ether_output)
523 * - at lookup
524 * - in rt_check when called by arpresolve
525 * - dropping the refcount for the rtentry
526 *
527 * This could be consolidated to one if we wrote a variant
528 * of arpresolve with an rt_check variant that expected to
529 * receive the route locked
530 */
531 #ifdef INET
532 if (ft == &V_ip4_ft) {
533 struct sockaddr_in *sin;
534
535 ro = &sro;
536 bzero(&sro.ro_dst, sizeof(sro.ro_dst));
537
538 sin = (struct sockaddr_in *)&sro.ro_dst;
539 sin->sin_family = AF_INET;
540 sin->sin_len = sizeof(*sin);
541 sin->sin_addr.s_addr = key[0];
542 }
543 #endif
544 #ifdef INET6
545 if (ft == &V_ip6_ft) {
546 struct sockaddr_in6 *sin6;
547
548 ro = (struct route *)&sro6;
549 sin6 = &sro6.ro_dst;
550
551 bzero(sin6, sizeof(*sin6));
552 sin6->sin6_family = AF_INET6;
553 sin6->sin6_len = sizeof(*sin6);
554 bcopy(key, &sin6->sin6_addr, sizeof(struct in6_addr));
555 }
556 #endif
557
558 ro->ro_rt = NULL;
559 #ifdef RADIX_MPATH
560 rtalloc_mpath_fib(ro, hash, fibnum);
561 #else
562 rtalloc_ign_fib(ro, 0, fibnum);
563 #endif
564 if (ro->ro_rt == NULL)
565 return (NULL);
566
567 rt = ro->ro_rt;
568 ifp = rt->rt_ifp;
569
570 if (ifp->if_flags & (IFF_POINTOPOINT | IFF_LOOPBACK)) {
571 RTFREE(rt);
572 return (NULL);
573 }
574
575 #ifdef INET
576 if (ft == &V_ip4_ft)
577 lt = LLTABLE(ifp);
578 #endif
579 #ifdef INET6
580 if (ft == &V_ip6_ft)
581 lt = LLTABLE6(ifp);
582 #endif
583
584 if (rt->rt_flags & RTF_GATEWAY)
585 l3addr = (struct sockaddr_storage *)rt->rt_gateway;
586 else
587 l3addr = (struct sockaddr_storage *)&ro->ro_dst;
588 lle = llentry_alloc(ifp, lt, l3addr);
589
590 if (lle == NULL) {
591 RTFREE(rt);
592 return (NULL);
593 }
594
595 /* Don't insert the entry if the ARP hasn't yet finished resolving. */
596 if ((lle->la_flags & LLE_VALID) == 0) {
597 RTFREE(rt);
598 LLE_FREE(lle);
599 FLOWSTAT_INC(ft, ft_fail_lle_invalid);
600 return (NULL);
601 }
602
603 fle = uma_zalloc(flow_zone, M_NOWAIT | M_ZERO);
604 if (fle == NULL) {
605 RTFREE(rt);
606 LLE_FREE(lle);
607 return (NULL);
608 }
609
610 fle->f_hash = hash;
611 bcopy(key, &fle->f_key, keylen);
612 fle->f_rt = rt;
613 fle->f_lle = lle;
614 fle->f_fibnum = fibnum;
615 fle->f_uptime = time_uptime;
616 #ifdef FLOWTABLE_HASH_ALL
617 fle->f_proto = proto;
618 fle->f_flags = fibnum0 >> 24;
619 #endif
620
621 critical_enter();
622 mask = flowtable_mask(ft);
623 flist = flowtable_list(ft, hash);
624
625 if (SLIST_EMPTY(flist)) {
626 bit_set(mask, (hash % ft->ft_size));
627 SLIST_INSERT_HEAD(flist, fle, f_next);
628 goto skip;
629 }
630
631 /*
632 * find end of list and make sure that we were not
633 * preempted by another thread handling this flow
634 */
635 SLIST_FOREACH(iter, flist, f_next) {
636 KASSERT(iter->f_hash % ft->ft_size == hash % ft->ft_size,
637 ("%s: wrong hash", __func__));
638 if (flow_matches(iter, key, keylen, fibnum)) {
639 /*
640 * We probably migrated to an other CPU after
641 * lookup in flowtable_lookup_common() failed.
642 * It appeared that this CPU already has flow
643 * entry.
644 */
645 iter->f_uptime = time_uptime;
646 #ifdef FLOWTABLE_HASH_ALL
647 iter->f_flags |= fibnum >> 24;
648 #endif
649 critical_exit();
650 FLOWSTAT_INC(ft, ft_collisions);
651 uma_zfree(flow_zone, fle);
652 return (iter);
653 }
654 }
655
656 SLIST_INSERT_HEAD(flist, fle, f_next);
657 skip:
658 critical_exit();
659 FLOWSTAT_INC(ft, ft_inserts);
660
661 return (fle);
662 }
663
664 int
flowtable_lookup(sa_family_t sa,struct mbuf * m,struct route * ro)665 flowtable_lookup(sa_family_t sa, struct mbuf *m, struct route *ro)
666 {
667 struct flentry *fle;
668 struct llentry *lle;
669
670 if (V_flowtable_enable == 0)
671 return (ENXIO);
672
673 switch (sa) {
674 #ifdef INET
675 case AF_INET:
676 fle = flowtable_lookup_ipv4(m, ro);
677 break;
678 #endif
679 #ifdef INET6
680 case AF_INET6:
681 fle = flowtable_lookup_ipv6(m, ro);
682 break;
683 #endif
684 default:
685 panic("%s: sa %d", __func__, sa);
686 }
687
688 if (fle == NULL)
689 return (EHOSTUNREACH);
690
691 if (M_HASHTYPE_GET(m) == M_HASHTYPE_NONE) {
692 M_HASHTYPE_SET(m, M_HASHTYPE_OPAQUE_HASH);
693 m->m_pkthdr.flowid = fle->f_hash;
694 }
695
696 ro->ro_rt = fle->f_rt;
697 ro->ro_flags |= RT_NORTREF;
698 lle = fle->f_lle;
699 if (lle != NULL && (lle->la_flags & LLE_VALID))
700 ro->ro_lle = lle; /* share ref with fle->f_lle */
701
702 return (0);
703 }
704
705 static struct flentry *
flowtable_lookup_common(struct flowtable * ft,uint32_t * key,int keylen,uint32_t fibnum)706 flowtable_lookup_common(struct flowtable *ft, uint32_t *key, int keylen,
707 uint32_t fibnum)
708 {
709 struct flist *flist;
710 struct flentry *fle;
711 uint32_t hash;
712
713 FLOWSTAT_INC(ft, ft_lookups);
714
715 hash = jenkins_hash32(key, keylen / sizeof(uint32_t), flow_hashjitter);
716
717 critical_enter();
718 flist = flowtable_list(ft, hash);
719 SLIST_FOREACH(fle, flist, f_next) {
720 KASSERT(fle->f_hash % ft->ft_size == hash % ft->ft_size,
721 ("%s: wrong hash", __func__));
722 if (flow_matches(fle, key, keylen, fibnum)) {
723 fle->f_uptime = time_uptime;
724 #ifdef FLOWTABLE_HASH_ALL
725 fle->f_flags |= fibnum >> 24;
726 #endif
727 critical_exit();
728 FLOWSTAT_INC(ft, ft_hits);
729 return (fle);
730 }
731 }
732 critical_exit();
733
734 FLOWSTAT_INC(ft, ft_misses);
735
736 return (flowtable_insert(ft, hash, key, keylen, fibnum));
737 }
738
739 static void
flowtable_alloc(struct flowtable * ft)740 flowtable_alloc(struct flowtable *ft)
741 {
742 int i;
743
744 ft->ft_table = malloc(ft->ft_size * sizeof(struct flist),
745 M_FTABLE, M_WAITOK);
746 for (int i = 0; i < ft->ft_size; i++)
747 ft->ft_table[i] = uma_zalloc(pcpu_zone_ptr, M_WAITOK | M_ZERO);
748
749 ft->ft_masks = uma_zalloc(pcpu_zone_ptr, M_WAITOK);
750 CPU_FOREACH(i) {
751 bitstr_t **b;
752
753 b = zpcpu_get_cpu(ft->ft_masks, i);
754 *b = bit_alloc(ft->ft_size, M_FTABLE, M_WAITOK);
755 }
756 ft->ft_tmpmask = bit_alloc(ft->ft_size, M_FTABLE, M_WAITOK);
757 }
758
759 static void
flowtable_free_stale(struct flowtable * ft,struct rtentry * rt,int maxidle)760 flowtable_free_stale(struct flowtable *ft, struct rtentry *rt, int maxidle)
761 {
762 struct flist *flist, freelist;
763 struct flentry *fle, *fle1, *fleprev;
764 bitstr_t *mask, *tmpmask;
765 int curbit, tmpsize;
766
767 SLIST_INIT(&freelist);
768 mask = flowtable_mask(ft);
769 tmpmask = ft->ft_tmpmask;
770 tmpsize = ft->ft_size;
771 memcpy(tmpmask, mask, ft->ft_size/8);
772 curbit = 0;
773 fleprev = NULL; /* pacify gcc */
774 /*
775 * XXX Note to self, bit_ffs operates at the byte level
776 * and thus adds gratuitous overhead
777 */
778 bit_ffs(tmpmask, ft->ft_size, &curbit);
779 while (curbit != -1) {
780 if (curbit >= ft->ft_size || curbit < -1) {
781 log(LOG_ALERT,
782 "warning: bad curbit value %d \n",
783 curbit);
784 break;
785 }
786
787 FLOWSTAT_INC(ft, ft_free_checks);
788
789 critical_enter();
790 flist = flowtable_list(ft, curbit);
791 #ifdef DIAGNOSTIC
792 if (SLIST_EMPTY(flist) && curbit > 0) {
793 log(LOG_ALERT,
794 "warning bit=%d set, but no fle found\n",
795 curbit);
796 }
797 #endif
798 SLIST_FOREACH_SAFE(fle, flist, f_next, fle1) {
799 if (rt != NULL && fle->f_rt != rt) {
800 fleprev = fle;
801 continue;
802 }
803 if (!flow_stale(ft, fle, maxidle)) {
804 fleprev = fle;
805 continue;
806 }
807
808 if (fle == SLIST_FIRST(flist))
809 SLIST_REMOVE_HEAD(flist, f_next);
810 else
811 SLIST_REMOVE_AFTER(fleprev, f_next);
812 SLIST_INSERT_HEAD(&freelist, fle, f_next);
813 }
814 if (SLIST_EMPTY(flist))
815 bit_clear(mask, curbit);
816 critical_exit();
817
818 bit_clear(tmpmask, curbit);
819 bit_ffs(tmpmask, tmpsize, &curbit);
820 }
821
822 SLIST_FOREACH_SAFE(fle, &freelist, f_next, fle1) {
823 FLOWSTAT_INC(ft, ft_frees);
824 if (fle->f_rt != NULL)
825 RTFREE(fle->f_rt);
826 if (fle->f_lle != NULL)
827 LLE_FREE(fle->f_lle);
828 uma_zfree(flow_zone, fle);
829 }
830 }
831
832 static void
flowtable_clean_vnet(struct flowtable * ft,struct rtentry * rt,int maxidle)833 flowtable_clean_vnet(struct flowtable *ft, struct rtentry *rt, int maxidle)
834 {
835 int i;
836
837 CPU_FOREACH(i) {
838 if (smp_started == 1) {
839 thread_lock(curthread);
840 sched_bind(curthread, i);
841 thread_unlock(curthread);
842 }
843
844 flowtable_free_stale(ft, rt, maxidle);
845
846 if (smp_started == 1) {
847 thread_lock(curthread);
848 sched_unbind(curthread);
849 thread_unlock(curthread);
850 }
851 }
852 }
853
854 void
flowtable_route_flush(sa_family_t sa,struct rtentry * rt)855 flowtable_route_flush(sa_family_t sa, struct rtentry *rt)
856 {
857 struct flowtable *ft;
858
859 switch (sa) {
860 #ifdef INET
861 case AF_INET:
862 ft = &V_ip4_ft;
863 break;
864 #endif
865 #ifdef INET6
866 case AF_INET6:
867 ft = &V_ip6_ft;
868 break;
869 #endif
870 default:
871 panic("%s: sa %d", __func__, sa);
872 }
873
874 flowtable_clean_vnet(ft, rt, 0);
875 }
876
877 static void
flowtable_cleaner(void)878 flowtable_cleaner(void)
879 {
880 VNET_ITERATOR_DECL(vnet_iter);
881 struct thread *td;
882
883 if (bootverbose)
884 log(LOG_INFO, "flowtable cleaner started\n");
885 td = curthread;
886 while (1) {
887 uint32_t flowclean_freq, maxidle;
888
889 /*
890 * The maximum idle time, as well as frequency are arbitrary.
891 */
892 if (flow_full())
893 maxidle = 5;
894 else
895 maxidle = 30;
896
897 VNET_LIST_RLOCK();
898 VNET_FOREACH(vnet_iter) {
899 CURVNET_SET(vnet_iter);
900 #ifdef INET
901 flowtable_clean_vnet(&V_ip4_ft, NULL, maxidle);
902 #endif
903 #ifdef INET6
904 flowtable_clean_vnet(&V_ip6_ft, NULL, maxidle);
905 #endif
906 CURVNET_RESTORE();
907 }
908 VNET_LIST_RUNLOCK();
909
910 if (flow_full())
911 flowclean_freq = 4*hz;
912 else
913 flowclean_freq = 20*hz;
914 mtx_lock(&flowclean_lock);
915 thread_lock(td);
916 sched_prio(td, PPAUSE);
917 thread_unlock(td);
918 flowclean_cycles++;
919 cv_broadcast(&flowclean_f_cv);
920 cv_timedwait(&flowclean_c_cv, &flowclean_lock, flowclean_freq);
921 mtx_unlock(&flowclean_lock);
922 }
923 }
924
925 static void
flowtable_flush(void * unused __unused)926 flowtable_flush(void *unused __unused)
927 {
928 uint64_t start;
929
930 mtx_lock(&flowclean_lock);
931 start = flowclean_cycles;
932 while (start == flowclean_cycles) {
933 cv_broadcast(&flowclean_c_cv);
934 cv_wait(&flowclean_f_cv, &flowclean_lock);
935 }
936 mtx_unlock(&flowclean_lock);
937 }
938
939 static struct kproc_desc flow_kp = {
940 "flowcleaner",
941 flowtable_cleaner,
942 &flowcleanerproc
943 };
944 SYSINIT(flowcleaner, SI_SUB_KTHREAD_IDLE, SI_ORDER_ANY, kproc_start, &flow_kp);
945
946 static int
flowtable_get_size(char * name)947 flowtable_get_size(char *name)
948 {
949 int size;
950
951 if (TUNABLE_INT_FETCH(name, &size)) {
952 if (size < 256)
953 size = 256;
954 if (!powerof2(size)) {
955 printf("%s must be power of 2\n", name);
956 size = 2048;
957 }
958 } else {
959 /*
960 * round up to the next power of 2
961 */
962 size = 1 << fls((1024 + maxusers * 64) - 1);
963 }
964
965 return (size);
966 }
967
968 static void
flowtable_init(const void * unused __unused)969 flowtable_init(const void *unused __unused)
970 {
971
972 flow_hashjitter = arc4random();
973
974 flow_zone = uma_zcreate("flows", sizeof(struct flentry),
975 NULL, NULL, NULL, NULL, (64-1), UMA_ZONE_MAXBUCKET);
976 uma_zone_set_max(flow_zone, 1024 + maxusers * 64 * mp_ncpus);
977
978 cv_init(&flowclean_c_cv, "c_flowcleanwait");
979 cv_init(&flowclean_f_cv, "f_flowcleanwait");
980 mtx_init(&flowclean_lock, "flowclean lock", NULL, MTX_DEF);
981 EVENTHANDLER_REGISTER(ifnet_departure_event, flowtable_flush, NULL,
982 EVENTHANDLER_PRI_ANY);
983 }
984 SYSINIT(flowtable_init, SI_SUB_PROTO_BEGIN, SI_ORDER_FIRST,
985 flowtable_init, NULL);
986
987 #ifdef INET
988 static SYSCTL_NODE(_net_flowtable, OID_AUTO, ip4, CTLFLAG_RD, NULL,
989 "Flowtable for IPv4");
990
991 static VNET_PCPUSTAT_DEFINE(struct flowtable_stat, ip4_ftstat);
992 VNET_PCPUSTAT_SYSINIT(ip4_ftstat);
993 VNET_PCPUSTAT_SYSUNINIT(ip4_ftstat);
994 SYSCTL_VNET_PCPUSTAT(_net_flowtable_ip4, OID_AUTO, stat, struct flowtable_stat,
995 ip4_ftstat, "Flowtable statistics for IPv4 "
996 "(struct flowtable_stat, net/flowtable.h)");
997
998 static void
flowtable_init_vnet_v4(const void * unused __unused)999 flowtable_init_vnet_v4(const void *unused __unused)
1000 {
1001
1002 V_ip4_ft.ft_size = flowtable_get_size("net.flowtable.ip4.size");
1003 V_ip4_ft.ft_stat = VNET(ip4_ftstat);
1004 flowtable_alloc(&V_ip4_ft);
1005 }
1006 VNET_SYSINIT(ft_vnet_v4, SI_SUB_PROTO_IFATTACHDOMAIN, SI_ORDER_ANY,
1007 flowtable_init_vnet_v4, NULL);
1008 #endif /* INET */
1009
1010 #ifdef INET6
1011 static SYSCTL_NODE(_net_flowtable, OID_AUTO, ip6, CTLFLAG_RD, NULL,
1012 "Flowtable for IPv6");
1013
1014 static VNET_PCPUSTAT_DEFINE(struct flowtable_stat, ip6_ftstat);
1015 VNET_PCPUSTAT_SYSINIT(ip6_ftstat);
1016 VNET_PCPUSTAT_SYSUNINIT(ip6_ftstat);
1017 SYSCTL_VNET_PCPUSTAT(_net_flowtable_ip6, OID_AUTO, stat, struct flowtable_stat,
1018 ip6_ftstat, "Flowtable statistics for IPv6 "
1019 "(struct flowtable_stat, net/flowtable.h)");
1020
1021 static void
flowtable_init_vnet_v6(const void * unused __unused)1022 flowtable_init_vnet_v6(const void *unused __unused)
1023 {
1024
1025 V_ip6_ft.ft_size = flowtable_get_size("net.flowtable.ip6.size");
1026 V_ip6_ft.ft_stat = VNET(ip6_ftstat);
1027 flowtable_alloc(&V_ip6_ft);
1028 }
1029 VNET_SYSINIT(flowtable_init_vnet_v6, SI_SUB_PROTO_IFATTACHDOMAIN, SI_ORDER_ANY,
1030 flowtable_init_vnet_v6, NULL);
1031 #endif /* INET6 */
1032
1033 #ifdef DDB
1034 static bitstr_t *
flowtable_mask_pcpu(struct flowtable * ft,int cpuid)1035 flowtable_mask_pcpu(struct flowtable *ft, int cpuid)
1036 {
1037
1038 return (zpcpu_get_cpu(*ft->ft_masks, cpuid));
1039 }
1040
1041 static struct flist *
flowtable_list_pcpu(struct flowtable * ft,uint32_t hash,int cpuid)1042 flowtable_list_pcpu(struct flowtable *ft, uint32_t hash, int cpuid)
1043 {
1044
1045 return (zpcpu_get_cpu(&ft->ft_table[hash % ft->ft_size], cpuid));
1046 }
1047
1048 static void
flow_show(struct flowtable * ft,struct flentry * fle)1049 flow_show(struct flowtable *ft, struct flentry *fle)
1050 {
1051 int idle_time;
1052 int rt_valid, ifp_valid;
1053 volatile struct rtentry *rt;
1054 struct ifnet *ifp = NULL;
1055 uint32_t *hashkey = fle->f_key;
1056
1057 idle_time = (int)(time_uptime - fle->f_uptime);
1058 rt = fle->f_rt;
1059 rt_valid = rt != NULL;
1060 if (rt_valid)
1061 ifp = rt->rt_ifp;
1062 ifp_valid = ifp != NULL;
1063
1064 #ifdef INET
1065 if (ft == &V_ip4_ft) {
1066 char daddr[4*sizeof "123"];
1067 #ifdef FLOWTABLE_HASH_ALL
1068 char saddr[4*sizeof "123"];
1069 uint16_t sport, dport;
1070 #endif
1071
1072 inet_ntoa_r(*(struct in_addr *) &hashkey[0], daddr);
1073 #ifdef FLOWTABLE_HASH_ALL
1074 inet_ntoa_r(*(struct in_addr *) &hashkey[1], saddr);
1075 dport = ntohs((uint16_t)(hashkey[2] >> 16));
1076 sport = ntohs((uint16_t)(hashkey[2] & 0xffff));
1077 db_printf("%s:%d->%s:%d", saddr, sport, daddr, dport);
1078 #else
1079 db_printf("%s ", daddr);
1080 #endif
1081 }
1082 #endif /* INET */
1083 #ifdef INET6
1084 if (ft == &V_ip6_ft) {
1085 #ifdef FLOWTABLE_HASH_ALL
1086 db_printf("\n\tkey=%08x:%08x:%08x%08x:%08x:%08x%08x:%08x:%08x",
1087 hashkey[0], hashkey[1], hashkey[2],
1088 hashkey[3], hashkey[4], hashkey[5],
1089 hashkey[6], hashkey[7], hashkey[8]);
1090 #else
1091 db_printf("\n\tkey=%08x:%08x:%08x ",
1092 hashkey[0], hashkey[1], hashkey[2]);
1093 #endif
1094 }
1095 #endif /* INET6 */
1096
1097 db_printf("hash=%08x idle_time=%03d"
1098 "\n\tfibnum=%02d rt=%p",
1099 fle->f_hash, idle_time, fle->f_fibnum, fle->f_rt);
1100
1101 #ifdef FLOWTABLE_HASH_ALL
1102 if (fle->f_flags & FL_STALE)
1103 db_printf(" FL_STALE ");
1104 #endif
1105 if (rt_valid) {
1106 if (rt->rt_flags & RTF_UP)
1107 db_printf(" RTF_UP ");
1108 }
1109 if (ifp_valid) {
1110 if (ifp->if_flags & IFF_LOOPBACK)
1111 db_printf(" IFF_LOOPBACK ");
1112 if (ifp->if_flags & IFF_UP)
1113 db_printf(" IFF_UP ");
1114 if (ifp->if_flags & IFF_POINTOPOINT)
1115 db_printf(" IFF_POINTOPOINT ");
1116 }
1117 db_printf("\n");
1118 }
1119
1120 static void
flowtable_show(struct flowtable * ft,int cpuid)1121 flowtable_show(struct flowtable *ft, int cpuid)
1122 {
1123 int curbit = 0;
1124 bitstr_t *mask, *tmpmask;
1125
1126 if (cpuid != -1)
1127 db_printf("cpu: %d\n", cpuid);
1128 mask = flowtable_mask_pcpu(ft, cpuid);
1129 tmpmask = ft->ft_tmpmask;
1130 memcpy(tmpmask, mask, ft->ft_size/8);
1131 /*
1132 * XXX Note to self, bit_ffs operates at the byte level
1133 * and thus adds gratuitous overhead
1134 */
1135 bit_ffs(tmpmask, ft->ft_size, &curbit);
1136 while (curbit != -1) {
1137 struct flist *flist;
1138 struct flentry *fle;
1139
1140 if (curbit >= ft->ft_size || curbit < -1) {
1141 db_printf("warning: bad curbit value %d \n",
1142 curbit);
1143 break;
1144 }
1145
1146 flist = flowtable_list_pcpu(ft, curbit, cpuid);
1147
1148 SLIST_FOREACH(fle, flist, f_next)
1149 flow_show(ft, fle);
1150 bit_clear(tmpmask, curbit);
1151 bit_ffs(tmpmask, ft->ft_size, &curbit);
1152 }
1153 }
1154
1155 static void
flowtable_show_vnet(struct flowtable * ft)1156 flowtable_show_vnet(struct flowtable *ft)
1157 {
1158
1159 int i;
1160
1161 CPU_FOREACH(i)
1162 flowtable_show(ft, i);
1163 }
1164
DB_SHOW_COMMAND(flowtables,db_show_flowtables)1165 DB_SHOW_COMMAND(flowtables, db_show_flowtables)
1166 {
1167 VNET_ITERATOR_DECL(vnet_iter);
1168
1169 VNET_FOREACH(vnet_iter) {
1170 CURVNET_SET(vnet_iter);
1171 #ifdef VIMAGE
1172 db_printf("vnet %p\n", vnet_iter);
1173 #endif
1174 #ifdef INET
1175 printf("IPv4:\n");
1176 flowtable_show_vnet(&V_ip4_ft);
1177 #endif
1178 #ifdef INET6
1179 printf("IPv6:\n");
1180 flowtable_show_vnet(&V_ip6_ft);
1181 #endif
1182 CURVNET_RESTORE();
1183 }
1184 }
1185 #endif
1186