]> CyberLeo.Net >> Repos - FreeBSD/releng/10.2.git/blob - sys/net/flowtable.c
- Copy stable/10@285827 to releng/10.2 in preparation for 10.2-RC1
[FreeBSD/releng/10.2.git] / sys / net / flowtable.c
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_VNET_INT(_net_flowtable, OID_AUTO, enable, 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 *
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 *
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 *
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 *
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
435 flow_stale(struct flowtable *ft, struct flentry *fle, int maxidle)
436 {
437
438         if (((fle->f_rt->rt_flags & RTF_HOST) &&
439             ((fle->f_rt->rt_flags & (RTF_UP)) != (RTF_UP))) ||
440             (fle->f_rt->rt_ifp == NULL) ||
441             !RT_LINK_IS_UP(fle->f_rt->rt_ifp) ||
442             (fle->f_lle->la_flags & LLE_VALID) == 0)
443                 return (1);
444
445         if (time_uptime - fle->f_uptime > maxidle)
446                 return (1);
447
448 #ifdef FLOWTABLE_HASH_ALL
449         if (fle->f_flags & FL_STALE)
450                 return (1);
451 #endif
452
453         return (0);
454 }
455
456 static int
457 flow_full(void)
458 {
459         int count, max;
460
461         count = uma_zone_get_cur(flow_zone);
462         max = uma_zone_get_max(flow_zone);
463
464         return (count > (max - (max >> 3)));
465 }
466
467 static int
468 flow_matches(struct flentry *fle, uint32_t *key, int keylen, uint32_t fibnum)
469 {
470 #ifdef FLOWTABLE_HASH_ALL
471         uint8_t proto;
472
473         proto = (fibnum >> 16) & 0xff;
474         fibnum &= 0xffff;
475 #endif
476
477         CRITICAL_ASSERT(curthread);
478
479         /* Microoptimization for IPv4: don't use bcmp(). */
480         if (((keylen == sizeof(uint32_t) && (fle->f_key[0] != key[0])) ||
481             (bcmp(fle->f_key, key, keylen) == 0)) &&
482             fibnum == fle->f_fibnum &&
483 #ifdef FLOWTABLE_HASH_ALL
484             proto == fle->f_proto &&
485 #endif
486             (fle->f_rt->rt_flags & RTF_UP) &&
487             fle->f_rt->rt_ifp != NULL &&
488             (fle->f_lle->la_flags & LLE_VALID))
489                 return (1);
490
491         return (0);
492 }
493
494 static struct flentry *
495 flowtable_insert(struct flowtable *ft, uint32_t hash, uint32_t *key,
496     int keylen, uint32_t fibnum0)
497 {
498 #ifdef INET6
499         struct route_in6 sro6;
500 #endif
501 #ifdef INET
502         struct route sro;
503 #endif
504         struct route *ro = NULL;
505         struct rtentry *rt;
506         struct lltable *lt = NULL;
507         struct llentry *lle;
508         struct sockaddr_storage *l3addr;
509         struct ifnet *ifp;
510         struct flist *flist;
511         struct flentry *fle, *iter;
512         bitstr_t *mask;
513         uint16_t fibnum = fibnum0;
514 #ifdef FLOWTABLE_HASH_ALL
515         uint8_t proto;
516
517         proto = (fibnum0 >> 16) & 0xff;
518         fibnum = fibnum0 & 0xffff;
519 #endif
520
521         /*
522          * This bit of code ends up locking the
523          * same route 3 times (just like ip_output + ether_output)
524          * - at lookup
525          * - in rt_check when called by arpresolve
526          * - dropping the refcount for the rtentry
527          *
528          * This could be consolidated to one if we wrote a variant
529          * of arpresolve with an rt_check variant that expected to
530          * receive the route locked
531          */
532 #ifdef INET
533         if (ft == &V_ip4_ft) {
534                 struct sockaddr_in *sin;
535
536                 ro = &sro;
537                 bzero(&sro.ro_dst, sizeof(sro.ro_dst));
538
539                 sin = (struct sockaddr_in *)&sro.ro_dst;
540                 sin->sin_family = AF_INET;
541                 sin->sin_len = sizeof(*sin);
542                 sin->sin_addr.s_addr = key[0];
543         }
544 #endif
545 #ifdef INET6
546         if (ft == &V_ip6_ft) {
547                 struct sockaddr_in6 *sin6;
548
549                 ro = (struct route *)&sro6;
550                 sin6 = &sro6.ro_dst;
551
552                 bzero(sin6, sizeof(*sin6));
553                 sin6->sin6_family = AF_INET6;
554                 sin6->sin6_len = sizeof(*sin6);
555                 bcopy(key, &sin6->sin6_addr, sizeof(struct in6_addr));
556         }
557 #endif
558
559         ro->ro_rt = NULL;
560 #ifdef RADIX_MPATH
561         rtalloc_mpath_fib(ro, hash, fibnum);
562 #else
563         rtalloc_ign_fib(ro, 0, fibnum);
564 #endif
565         if (ro->ro_rt == NULL)
566                 return (NULL);
567
568         rt = ro->ro_rt;
569         ifp = rt->rt_ifp;
570
571         if (ifp->if_flags & (IFF_POINTOPOINT | IFF_LOOPBACK)) {
572                 RTFREE(rt);
573                 return (NULL);
574         }
575
576 #ifdef INET
577         if (ft == &V_ip4_ft)
578                 lt = LLTABLE(ifp);
579 #endif
580 #ifdef INET6
581         if (ft == &V_ip6_ft)
582                 lt = LLTABLE6(ifp);
583 #endif
584
585         if (rt->rt_flags & RTF_GATEWAY)
586                 l3addr = (struct sockaddr_storage *)rt->rt_gateway;
587         else
588                 l3addr = (struct sockaddr_storage *)&ro->ro_dst;
589         lle = llentry_alloc(ifp, lt, l3addr);
590
591         if (lle == NULL) {
592                 RTFREE(rt);
593                 return (NULL);
594         }
595
596         /* Don't insert the entry if the ARP hasn't yet finished resolving. */
597         if ((lle->la_flags & LLE_VALID) == 0) {
598                 RTFREE(rt);
599                 LLE_FREE(lle);
600                 FLOWSTAT_INC(ft, ft_fail_lle_invalid);
601                 return (NULL);
602         }
603
604         fle = uma_zalloc(flow_zone, M_NOWAIT | M_ZERO);
605         if (fle == NULL) {
606                 RTFREE(rt);
607                 LLE_FREE(lle);
608                 return (NULL);
609         }
610
611         fle->f_hash = hash;
612         bcopy(key, &fle->f_key, keylen);
613         fle->f_rt = rt;
614         fle->f_lle = lle;
615         fle->f_fibnum = fibnum;
616         fle->f_uptime = time_uptime;
617 #ifdef FLOWTABLE_HASH_ALL
618         fle->f_proto = proto;
619         fle->f_flags = fibnum0 >> 24;
620 #endif
621
622         critical_enter();
623         mask = flowtable_mask(ft);
624         flist = flowtable_list(ft, hash);
625
626         if (SLIST_EMPTY(flist)) {
627                 bit_set(mask, (hash % ft->ft_size));
628                 SLIST_INSERT_HEAD(flist, fle, f_next);
629                 goto skip;
630         }
631
632         /*
633          * find end of list and make sure that we were not
634          * preempted by another thread handling this flow
635          */
636         SLIST_FOREACH(iter, flist, f_next) {
637                 KASSERT(iter->f_hash % ft->ft_size == hash % ft->ft_size,
638                     ("%s: wrong hash", __func__));
639                 if (flow_matches(iter, key, keylen, fibnum)) {
640                         /*
641                          * We probably migrated to an other CPU after
642                          * lookup in flowtable_lookup_common() failed.
643                          * It appeared that this CPU already has flow
644                          * entry.
645                          */
646                         iter->f_uptime = time_uptime;
647 #ifdef FLOWTABLE_HASH_ALL
648                         iter->f_flags |= fibnum >> 24;
649 #endif
650                         critical_exit();
651                         FLOWSTAT_INC(ft, ft_collisions);
652                         uma_zfree(flow_zone, fle);
653                         return (iter);
654                 }
655         }
656
657         SLIST_INSERT_HEAD(flist, fle, f_next);
658 skip:
659         critical_exit();
660         FLOWSTAT_INC(ft, ft_inserts);
661
662         return (fle);
663 }
664
665 int
666 flowtable_lookup(sa_family_t sa, struct mbuf *m, struct route *ro)
667 {
668         struct flentry *fle;
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);
693                 m->m_pkthdr.flowid = fle->f_hash;
694         }
695
696         ro->ro_rt = fle->f_rt;
697         ro->ro_lle = fle->f_lle;
698         ro->ro_flags |= RT_NORTREF;
699
700         return (0);
701 }
702
703 static struct flentry *
704 flowtable_lookup_common(struct flowtable *ft, uint32_t *key, int keylen,
705     uint32_t fibnum)
706 {
707         struct flist *flist;
708         struct flentry *fle;
709         uint32_t hash;
710
711         FLOWSTAT_INC(ft, ft_lookups);
712
713         hash = jenkins_hash32(key, keylen / sizeof(uint32_t), flow_hashjitter);
714
715         critical_enter();
716         flist = flowtable_list(ft, hash);
717         SLIST_FOREACH(fle, flist, f_next) {
718                 KASSERT(fle->f_hash % ft->ft_size == hash % ft->ft_size,
719                     ("%s: wrong hash", __func__));
720                 if (flow_matches(fle, key, keylen, fibnum)) {
721                         fle->f_uptime = time_uptime;
722 #ifdef FLOWTABLE_HASH_ALL
723                         fle->f_flags |= fibnum >> 24;
724 #endif
725                         critical_exit();
726                         FLOWSTAT_INC(ft, ft_hits);
727                         return (fle);
728                 }
729         }
730         critical_exit();
731
732         FLOWSTAT_INC(ft, ft_misses);
733
734         return (flowtable_insert(ft, hash, key, keylen, fibnum));
735 }
736
737 /*
738  * used by the bit_alloc macro
739  */
740 #define calloc(count, size) malloc((count)*(size), M_FTABLE, M_WAITOK | M_ZERO)
741 static void
742 flowtable_alloc(struct flowtable *ft)
743 {
744
745         ft->ft_table = malloc(ft->ft_size * sizeof(struct flist),
746             M_FTABLE, M_WAITOK);
747         for (int i = 0; i < ft->ft_size; i++)
748                 ft->ft_table[i] = uma_zalloc(pcpu_zone_ptr, M_WAITOK | M_ZERO);
749
750         ft->ft_masks = uma_zalloc(pcpu_zone_ptr, M_WAITOK);
751         for (int i = 0; i < mp_ncpus; i++) {
752                 bitstr_t **b;
753
754                 b = zpcpu_get_cpu(ft->ft_masks, i);
755                 *b = bit_alloc(ft->ft_size);
756         }
757         ft->ft_tmpmask = bit_alloc(ft->ft_size);
758 }
759 #undef calloc
760
761 static void
762 flowtable_free_stale(struct flowtable *ft, struct rtentry *rt, int maxidle)
763 {
764         struct flist *flist, freelist;
765         struct flentry *fle, *fle1, *fleprev;
766         bitstr_t *mask, *tmpmask;
767         int curbit, tmpsize;
768
769         SLIST_INIT(&freelist);
770         mask = flowtable_mask(ft);
771         tmpmask = ft->ft_tmpmask;
772         tmpsize = ft->ft_size;
773         memcpy(tmpmask, mask, ft->ft_size/8);
774         curbit = 0;
775         fleprev = NULL; /* pacify gcc */
776         /*
777          * XXX Note to self, bit_ffs operates at the byte level
778          * and thus adds gratuitous overhead
779          */
780         bit_ffs(tmpmask, ft->ft_size, &curbit);
781         while (curbit != -1) {
782                 if (curbit >= ft->ft_size || curbit < -1) {
783                         log(LOG_ALERT,
784                             "warning: bad curbit value %d \n",
785                             curbit);
786                         break;
787                 }
788
789                 FLOWSTAT_INC(ft, ft_free_checks);
790
791                 critical_enter();
792                 flist = flowtable_list(ft, curbit);
793 #ifdef DIAGNOSTIC
794                 if (SLIST_EMPTY(flist) && curbit > 0) {
795                         log(LOG_ALERT,
796                             "warning bit=%d set, but no fle found\n",
797                             curbit);
798                 }
799 #endif
800                 SLIST_FOREACH_SAFE(fle, flist, f_next, fle1) {
801                         if (rt != NULL && fle->f_rt != rt) {
802                                 fleprev = fle;
803                                 continue;
804                         }
805                         if (!flow_stale(ft, fle, maxidle)) {
806                                 fleprev = fle;
807                                 continue;
808                         }
809
810                         if (fle == SLIST_FIRST(flist))
811                                 SLIST_REMOVE_HEAD(flist, f_next);
812                         else
813                                 SLIST_REMOVE_AFTER(fleprev, f_next);
814                         SLIST_INSERT_HEAD(&freelist, fle, f_next);
815                 }
816                 if (SLIST_EMPTY(flist))
817                         bit_clear(mask, curbit);
818                 critical_exit();
819
820                 bit_clear(tmpmask, curbit);
821                 tmpmask += (curbit / 8);
822                 tmpsize -= (curbit / 8) * 8;
823                 bit_ffs(tmpmask, tmpsize, &curbit);
824         }
825
826         SLIST_FOREACH_SAFE(fle, &freelist, f_next, fle1) {
827                 FLOWSTAT_INC(ft, ft_frees);
828                 if (fle->f_rt != NULL)
829                         RTFREE(fle->f_rt);
830                 if (fle->f_lle != NULL)
831                         LLE_FREE(fle->f_lle);
832                 uma_zfree(flow_zone, fle);
833         }
834 }
835
836 static void
837 flowtable_clean_vnet(struct flowtable *ft, struct rtentry *rt, int maxidle)
838 {
839         int i;
840
841         CPU_FOREACH(i) {
842                 if (smp_started == 1) {
843                         thread_lock(curthread);
844                         sched_bind(curthread, i);
845                         thread_unlock(curthread);
846                 }
847
848                 flowtable_free_stale(ft, rt, maxidle);
849
850                 if (smp_started == 1) {
851                         thread_lock(curthread);
852                         sched_unbind(curthread);
853                         thread_unlock(curthread);
854                 }
855         }
856 }
857
858 void
859 flowtable_route_flush(sa_family_t sa, struct rtentry *rt)
860 {
861         struct flowtable *ft;
862
863         switch (sa) {
864 #ifdef INET
865         case AF_INET:
866                 ft = &V_ip4_ft;
867                 break;
868 #endif
869 #ifdef INET6
870         case AF_INET6:
871                 ft = &V_ip6_ft;
872                 break;
873 #endif
874         default:
875                 panic("%s: sa %d", __func__, sa);
876         }
877
878         flowtable_clean_vnet(ft, rt, 0);
879 }
880
881 static void
882 flowtable_cleaner(void)
883 {
884         VNET_ITERATOR_DECL(vnet_iter);
885         struct thread *td;
886
887         if (bootverbose)
888                 log(LOG_INFO, "flowtable cleaner started\n");
889         td = curthread;
890         while (1) {
891                 uint32_t flowclean_freq, maxidle;
892
893                 /*
894                  * The maximum idle time, as well as frequency are arbitrary.
895                  */
896                 if (flow_full())
897                         maxidle = 5;
898                 else
899                         maxidle = 30;
900
901                 VNET_LIST_RLOCK();
902                 VNET_FOREACH(vnet_iter) {
903                         CURVNET_SET(vnet_iter);
904 #ifdef INET
905                         flowtable_clean_vnet(&V_ip4_ft, NULL, maxidle);
906 #endif
907 #ifdef INET6
908                         flowtable_clean_vnet(&V_ip6_ft, NULL, maxidle);
909 #endif
910                         CURVNET_RESTORE();
911                 }
912                 VNET_LIST_RUNLOCK();
913
914                 if (flow_full())
915                         flowclean_freq = 4*hz;
916                 else
917                         flowclean_freq = 20*hz;
918                 mtx_lock(&flowclean_lock);
919                 thread_lock(td);
920                 sched_prio(td, PPAUSE);
921                 thread_unlock(td);
922                 flowclean_cycles++;
923                 cv_broadcast(&flowclean_f_cv);
924                 cv_timedwait(&flowclean_c_cv, &flowclean_lock, flowclean_freq);
925                 mtx_unlock(&flowclean_lock);
926         }
927 }
928
929 static void
930 flowtable_flush(void *unused __unused)
931 {
932         uint64_t start;
933
934         mtx_lock(&flowclean_lock);
935         start = flowclean_cycles;
936         while (start == flowclean_cycles) {
937                 cv_broadcast(&flowclean_c_cv);
938                 cv_wait(&flowclean_f_cv, &flowclean_lock);
939         }
940         mtx_unlock(&flowclean_lock);
941 }
942
943 static struct kproc_desc flow_kp = {
944         "flowcleaner",
945         flowtable_cleaner,
946         &flowcleanerproc
947 };
948 SYSINIT(flowcleaner, SI_SUB_KTHREAD_IDLE, SI_ORDER_ANY, kproc_start, &flow_kp);
949
950 static int
951 flowtable_get_size(char *name)
952 {
953         int size;
954
955         if (TUNABLE_INT_FETCH(name, &size)) {
956                 if (size < 256)
957                         size = 256;
958                 if (!powerof2(size)) {
959                         printf("%s must be power of 2\n", name);
960                         size = 2048;
961                 }
962         } else {
963                 /*
964                  * round up to the next power of 2
965                  */
966                 size = 1 << fls((1024 + maxusers * 64) - 1);
967         }
968
969         return (size);
970 }
971
972 static void
973 flowtable_init(const void *unused __unused)
974 {
975
976         flow_hashjitter = arc4random();
977
978         flow_zone = uma_zcreate("flows", sizeof(struct flentry),
979             NULL, NULL, NULL, NULL, (64-1), UMA_ZONE_MAXBUCKET);
980         uma_zone_set_max(flow_zone, 1024 + maxusers * 64 * mp_ncpus);
981
982         cv_init(&flowclean_c_cv, "c_flowcleanwait");
983         cv_init(&flowclean_f_cv, "f_flowcleanwait");
984         mtx_init(&flowclean_lock, "flowclean lock", NULL, MTX_DEF);
985         EVENTHANDLER_REGISTER(ifnet_departure_event, flowtable_flush, NULL,
986             EVENTHANDLER_PRI_ANY);
987 }
988 SYSINIT(flowtable_init, SI_SUB_PROTO_BEGIN, SI_ORDER_FIRST,
989     flowtable_init, NULL);
990
991 #ifdef INET
992 static SYSCTL_NODE(_net_flowtable, OID_AUTO, ip4, CTLFLAG_RD, NULL,
993     "Flowtable for IPv4");
994
995 static VNET_PCPUSTAT_DEFINE(struct flowtable_stat, ip4_ftstat);
996 VNET_PCPUSTAT_SYSINIT(ip4_ftstat);
997 VNET_PCPUSTAT_SYSUNINIT(ip4_ftstat);
998 SYSCTL_VNET_PCPUSTAT(_net_flowtable_ip4, OID_AUTO, stat, struct flowtable_stat,
999     ip4_ftstat, "Flowtable statistics for IPv4 "
1000     "(struct flowtable_stat, net/flowtable.h)");
1001
1002 static void
1003 flowtable_init_vnet_v4(const void *unused __unused)
1004 {
1005
1006         V_ip4_ft.ft_size = flowtable_get_size("net.flowtable.ip4.size");
1007         V_ip4_ft.ft_stat = VNET(ip4_ftstat);
1008         flowtable_alloc(&V_ip4_ft);
1009 }
1010 VNET_SYSINIT(ft_vnet_v4, SI_SUB_PROTO_IFATTACHDOMAIN, SI_ORDER_ANY,
1011     flowtable_init_vnet_v4, NULL);
1012 #endif /* INET */
1013
1014 #ifdef INET6
1015 static SYSCTL_NODE(_net_flowtable, OID_AUTO, ip6, CTLFLAG_RD, NULL,
1016     "Flowtable for IPv6");
1017
1018 static VNET_PCPUSTAT_DEFINE(struct flowtable_stat, ip6_ftstat);
1019 VNET_PCPUSTAT_SYSINIT(ip6_ftstat);
1020 VNET_PCPUSTAT_SYSUNINIT(ip6_ftstat);
1021 SYSCTL_VNET_PCPUSTAT(_net_flowtable_ip6, OID_AUTO, stat, struct flowtable_stat,
1022     ip6_ftstat, "Flowtable statistics for IPv6 "
1023     "(struct flowtable_stat, net/flowtable.h)");
1024
1025 static void
1026 flowtable_init_vnet_v6(const void *unused __unused)
1027 {
1028
1029         V_ip6_ft.ft_size = flowtable_get_size("net.flowtable.ip6.size");
1030         V_ip6_ft.ft_stat = VNET(ip6_ftstat);
1031         flowtable_alloc(&V_ip6_ft);
1032 }
1033 VNET_SYSINIT(flowtable_init_vnet_v6, SI_SUB_PROTO_IFATTACHDOMAIN, SI_ORDER_ANY,
1034     flowtable_init_vnet_v6, NULL);
1035 #endif /* INET6 */
1036
1037 #ifdef DDB
1038 static bitstr_t *
1039 flowtable_mask_pcpu(struct flowtable *ft, int cpuid)
1040 {
1041
1042         return (zpcpu_get_cpu(*ft->ft_masks, cpuid));
1043 }
1044
1045 static struct flist *
1046 flowtable_list_pcpu(struct flowtable *ft, uint32_t hash, int cpuid)
1047 {
1048
1049         return (zpcpu_get_cpu(&ft->ft_table[hash % ft->ft_size], cpuid));
1050 }
1051
1052 static void
1053 flow_show(struct flowtable *ft, struct flentry *fle)
1054 {
1055         int idle_time;
1056         int rt_valid, ifp_valid;
1057         volatile struct rtentry *rt;
1058         struct ifnet *ifp = NULL;
1059         uint32_t *hashkey = fle->f_key;
1060
1061         idle_time = (int)(time_uptime - fle->f_uptime);
1062         rt = fle->f_rt;
1063         rt_valid = rt != NULL;
1064         if (rt_valid)
1065                 ifp = rt->rt_ifp;
1066         ifp_valid = ifp != NULL;
1067
1068 #ifdef INET
1069         if (ft == &V_ip4_ft) {
1070                 char daddr[4*sizeof "123"];
1071 #ifdef FLOWTABLE_HASH_ALL
1072                 char saddr[4*sizeof "123"];
1073                 uint16_t sport, dport;
1074 #endif
1075
1076                 inet_ntoa_r(*(struct in_addr *) &hashkey[0], daddr);
1077 #ifdef FLOWTABLE_HASH_ALL
1078                 inet_ntoa_r(*(struct in_addr *) &hashkey[1], saddr);
1079                 dport = ntohs((uint16_t)(hashkey[2] >> 16));
1080                 sport = ntohs((uint16_t)(hashkey[2] & 0xffff));
1081                 db_printf("%s:%d->%s:%d", saddr, sport, daddr, dport);
1082 #else
1083                 db_printf("%s ", daddr);
1084 #endif
1085         }
1086 #endif /* INET */
1087 #ifdef INET6
1088         if (ft == &V_ip6_ft) {
1089 #ifdef FLOWTABLE_HASH_ALL
1090                 db_printf("\n\tkey=%08x:%08x:%08x%08x:%08x:%08x%08x:%08x:%08x",
1091                     hashkey[0], hashkey[1], hashkey[2],
1092                     hashkey[3], hashkey[4], hashkey[5],
1093                     hashkey[6], hashkey[7], hashkey[8]);
1094 #else
1095                 db_printf("\n\tkey=%08x:%08x:%08x ",
1096                     hashkey[0], hashkey[1], hashkey[2]);
1097 #endif
1098         }
1099 #endif /* INET6 */
1100
1101         db_printf("hash=%08x idle_time=%03d"
1102             "\n\tfibnum=%02d rt=%p",
1103             fle->f_hash, idle_time, fle->f_fibnum, fle->f_rt);
1104
1105 #ifdef FLOWTABLE_HASH_ALL
1106         if (fle->f_flags & FL_STALE)
1107                 db_printf(" FL_STALE ");
1108 #endif
1109         if (rt_valid) {
1110                 if (rt->rt_flags & RTF_UP)
1111                         db_printf(" RTF_UP ");
1112         }
1113         if (ifp_valid) {
1114                 if (ifp->if_flags & IFF_LOOPBACK)
1115                         db_printf(" IFF_LOOPBACK ");
1116                 if (ifp->if_flags & IFF_UP)
1117                         db_printf(" IFF_UP ");
1118                 if (ifp->if_flags & IFF_POINTOPOINT)
1119                         db_printf(" IFF_POINTOPOINT ");
1120         }
1121         db_printf("\n");
1122 }
1123
1124 static void
1125 flowtable_show(struct flowtable *ft, int cpuid)
1126 {
1127         int curbit = 0;
1128         bitstr_t *mask, *tmpmask;
1129
1130         if (cpuid != -1)
1131                 db_printf("cpu: %d\n", cpuid);
1132         mask = flowtable_mask_pcpu(ft, cpuid);
1133         tmpmask = ft->ft_tmpmask;
1134         memcpy(tmpmask, mask, ft->ft_size/8);
1135         /*
1136          * XXX Note to self, bit_ffs operates at the byte level
1137          * and thus adds gratuitous overhead
1138          */
1139         bit_ffs(tmpmask, ft->ft_size, &curbit);
1140         while (curbit != -1) {
1141                 struct flist *flist;
1142                 struct flentry *fle;
1143
1144                 if (curbit >= ft->ft_size || curbit < -1) {
1145                         db_printf("warning: bad curbit value %d \n",
1146                             curbit);
1147                         break;
1148                 }
1149
1150                 flist = flowtable_list_pcpu(ft, curbit, cpuid);
1151
1152                 SLIST_FOREACH(fle, flist, f_next)
1153                         flow_show(ft, fle);
1154                 bit_clear(tmpmask, curbit);
1155                 bit_ffs(tmpmask, ft->ft_size, &curbit);
1156         }
1157 }
1158
1159 static void
1160 flowtable_show_vnet(struct flowtable *ft)
1161 {
1162
1163         int i;
1164
1165         CPU_FOREACH(i)
1166                 flowtable_show(ft, i);
1167 }
1168
1169 DB_SHOW_COMMAND(flowtables, db_show_flowtables)
1170 {
1171         VNET_ITERATOR_DECL(vnet_iter);
1172
1173         VNET_FOREACH(vnet_iter) {
1174                 CURVNET_SET(vnet_iter);
1175 #ifdef VIMAGE
1176                 db_printf("vnet %p\n", vnet_iter);
1177 #endif
1178 #ifdef INET
1179                 printf("IPv4:\n");
1180                 flowtable_show_vnet(&V_ip4_ft);
1181 #endif
1182 #ifdef INET6
1183                 printf("IPv6:\n");
1184                 flowtable_show_vnet(&V_ip6_ft);
1185 #endif
1186                 CURVNET_RESTORE();
1187         }
1188 }
1189 #endif