sys/net/flowtable.c

   1 /**************************************************************************
   2
   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
  30 #include "opt_route.h"
  31 #include "opt_mpath.h"
  32 #include "opt_ddb.h"
  33 #include "opt_inet.h"
  34 #include "opt_inet6.h"
  35
  36 #include <sys/cdefs.h>
  37 __FBSDID("$FreeBSD$");
  38
  39 #include <sys/param.h>
  40 #include <sys/types.h>
  41 #include <sys/bitstring.h>
  42 #include <sys/condvar.h>
  43 #include <sys/callout.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/proc.h>
  50 #include <sys/sbuf.h>
  51 #include <sys/sched.h>
  52 #include <sys/smp.h>
  53 #include <sys/socket.h>
  54 #include <sys/syslog.h>
  55 #include <sys/sysctl.h>
  56
  57 #include <net/if.h>
  58 #include <net/if_llatbl.h>
  59 #include <net/if_var.h>
  60 #include <net/route.h>
  61 #include <net/flowtable.h>
  62 #include <net/vnet.h>
  63
  64 #include <netinet/in.h>
  65 #include <netinet/in_systm.h>
  66 #include <netinet/in_var.h>
  67 #include <netinet/if_ether.h>
  68 #include <netinet/ip.h>
  69 #ifdef INET6
  70 #include <netinet/ip6.h>
  71 #endif
  72 #include <netinet/tcp.h>
  73 #include <netinet/udp.h>
  74 #include <netinet/sctp.h>
  75
  76 #include <libkern/jenkins.h>
  77 #include <ddb/ddb.h>
  78
  79 struct ipv4_tuple {
  80         uint16_t        ip_sport;       /* source port */
  81         uint16_t        ip_dport;       /* destination port */
  82         in_addr_t       ip_saddr;       /* source address */
  83         in_addr_t       ip_daddr;       /* destination address */
  84 };
  85
  86 union ipv4_flow {
  87         struct ipv4_tuple ipf_ipt;
  88         uint32_t        ipf_key[3];
  89 };
  90
  91 struct ipv6_tuple {
  92         uint16_t        ip_sport;       /* source port */
  93         uint16_t        ip_dport;       /* destination port */
  94         struct in6_addr ip_saddr;       /* source address */
  95         struct in6_addr ip_daddr;       /* destination address */
  96 };
  97
  98 union ipv6_flow {
  99         struct ipv6_tuple ipf_ipt;
 100         uint32_t        ipf_key[9];
 101 };
 102
 103 struct flentry {
 104         volatile uint32_t       f_fhash;        /* hash flowing forward */
 105         uint16_t                f_flags;        /* flow flags */
 106         uint8_t                 f_pad;
 107         uint8_t                 f_proto;        /* protocol */
 108         uint32_t                f_fibnum;       /* fib index */
 109         uint32_t                f_uptime;       /* uptime at last access */
 110         struct flentry          *f_next;        /* pointer to collision entry */
 111         volatile struct rtentry *f_rt;          /* rtentry for flow */
 112         volatile struct llentry *f_lle;         /* llentry for flow */
 113 };
 114
 115 struct flentry_v4 {
 116         struct flentry  fl_entry;
 117         union ipv4_flow fl_flow;
 118 };
 119
 120 struct flentry_v6 {
 121         struct flentry  fl_entry;
 122         union ipv6_flow fl_flow;
 123 };
 124
 125 #define fl_fhash        fl_entry.fl_fhash
 126 #define fl_flags        fl_entry.fl_flags
 127 #define fl_proto        fl_entry.fl_proto
 128 #define fl_uptime       fl_entry.fl_uptime
 129 #define fl_rt           fl_entry.fl_rt
 130 #define fl_lle          fl_entry.fl_lle
 131
 132 #define SECS_PER_HOUR           3600
 133 #define SECS_PER_DAY            (24*SECS_PER_HOUR)
 134
 135 #define SYN_IDLE                300
 136 #define UDP_IDLE                300
 137 #define FIN_WAIT_IDLE           600
 138 #define TCP_IDLE                SECS_PER_DAY
 139
 140
 141 typedef void fl_lock_t(struct flowtable *, uint32_t);
 142 typedef void fl_rtalloc_t(struct route *, uint32_t, u_int);
 143
 144 union flentryp {
 145         struct flentry          **global;
 146         struct flentry          **pcpu[MAXCPU];
 147 };
 148
 149 struct flowtable_stats {
 150         uint64_t        ft_collisions;
 151         uint64_t        ft_allocated;
 152         uint64_t        ft_misses;
 153         uint64_t        ft_max_depth;
 154         uint64_t        ft_free_checks;
 155         uint64_t        ft_frees;
 156         uint64_t        ft_hits;
 157         uint64_t        ft_lookups;
 158 } __aligned(CACHE_LINE_SIZE);
 159
 160 struct flowtable {
 161         struct  flowtable_stats ft_stats[MAXCPU];
 162         int             ft_size;
 163         int             ft_lock_count;
 164         uint32_t        ft_flags;
 165         char            *ft_name;
 166         fl_lock_t       *ft_lock;
 167         fl_lock_t       *ft_unlock;
 168         fl_rtalloc_t    *ft_rtalloc;
 169         /*
 170          * XXX need to pad out
 171          */
 172         struct mtx      *ft_locks;
 173         union flentryp  ft_table;
 174         bitstr_t        *ft_masks[MAXCPU];
 175         bitstr_t        *ft_tmpmask;
 176         struct flowtable *ft_next;
 177
 178         uint32_t        ft_count __aligned(CACHE_LINE_SIZE);
 179         uint32_t        ft_udp_idle __aligned(CACHE_LINE_SIZE);
 180         uint32_t        ft_fin_wait_idle;
 181         uint32_t        ft_syn_idle;
 182         uint32_t        ft_tcp_idle;
 183         boolean_t       ft_full;
 184 } __aligned(CACHE_LINE_SIZE);
 185
 186 static struct proc *flowcleanerproc;
 187 static VNET_DEFINE(struct flowtable *, flow_list_head);
 188 static VNET_DEFINE(uint32_t, flow_hashjitter);
 189 static VNET_DEFINE(uma_zone_t, flow_ipv4_zone);
 190 static VNET_DEFINE(uma_zone_t, flow_ipv6_zone);
 191
 192 #define V_flow_list_head        VNET(flow_list_head)
 193 #define V_flow_hashjitter       VNET(flow_hashjitter)
 194 #define V_flow_ipv4_zone        VNET(flow_ipv4_zone)
 195 #define V_flow_ipv6_zone        VNET(flow_ipv6_zone)
 196
 197
 198 static struct cv        flowclean_cv;
 199 static struct mtx       flowclean_lock;
 200 static uint32_t         flowclean_cycles;
 201 static uint32_t         flowclean_freq;
 202
 203 #ifdef FLOWTABLE_DEBUG
 204 #define FLDPRINTF(ft, flags, fmt, ...)          \
 205 do {                                            \
 206         if ((ft)->ft_flags & (flags))           \
 207                 printf((fmt), __VA_ARGS__);     \
 208 } while (0);                                    \
 209
 210 #else
 211 #define FLDPRINTF(ft, flags, fmt, ...)
 212
 213 #endif
 214
 215
 216 /*
 217  * TODO:
 218  * - Make flowtable stats per-cpu, aggregated at sysctl call time,
 219  *   to avoid extra cache evictions caused by incrementing a shared
 220  *   counter
 221  * - add sysctls to resize && flush flow tables
 222  * - Add per flowtable sysctls for statistics and configuring timeouts
 223  * - add saturation counter to rtentry to support per-packet load-balancing
 224  *   add flag to indicate round-robin flow, add list lookup from head
 225      for flows
 226  * - add sysctl / device node / syscall to support exporting and importing
 227  *   of flows with flag to indicate that a flow was imported so should
 228  *   not be considered for auto-cleaning
 229  * - support explicit connection state (currently only ad-hoc for DSR)
 230  * - idetach() cleanup for options VIMAGE builds.
 231  */
 232 VNET_DEFINE(int, flowtable_enable) = 1;
 233 static VNET_DEFINE(int, flowtable_debug);
 234 static VNET_DEFINE(int, flowtable_syn_expire) = SYN_IDLE;
 235 static VNET_DEFINE(int, flowtable_udp_expire) = UDP_IDLE;
 236 static VNET_DEFINE(int, flowtable_fin_wait_expire) = FIN_WAIT_IDLE;
 237 static VNET_DEFINE(int, flowtable_tcp_expire) = TCP_IDLE;
 238 static VNET_DEFINE(int, flowtable_nmbflows);
 239 static VNET_DEFINE(int, flowtable_ready) = 0;
 240
 241 #define V_flowtable_enable              VNET(flowtable_enable)
 242 #define V_flowtable_debug               VNET(flowtable_debug)
 243 #define V_flowtable_syn_expire          VNET(flowtable_syn_expire)
 244 #define V_flowtable_udp_expire          VNET(flowtable_udp_expire)
 245 #define V_flowtable_fin_wait_expire     VNET(flowtable_fin_wait_expire)
 246 #define V_flowtable_tcp_expire          VNET(flowtable_tcp_expire)
 247 #define V_flowtable_nmbflows            VNET(flowtable_nmbflows)
 248 #define V_flowtable_ready               VNET(flowtable_ready)
 249
 250 SYSCTL_NODE(_net_inet, OID_AUTO, flowtable, CTLFLAG_RD, NULL, "flowtable");
 251 SYSCTL_VNET_INT(_net_inet_flowtable, OID_AUTO, debug, CTLFLAG_RW,
 252     &VNET_NAME(flowtable_debug), 0, "print debug info.");
 253 SYSCTL_VNET_INT(_net_inet_flowtable, OID_AUTO, enable, CTLFLAG_RW,
 254     &VNET_NAME(flowtable_enable), 0, "enable flowtable caching.");
 255
 256 /*
 257  * XXX This does not end up updating timeouts at runtime
 258  * and only reflects the value for the last table added :-/
 259  */
 260 SYSCTL_VNET_INT(_net_inet_flowtable, OID_AUTO, syn_expire, CTLFLAG_RW,
 261     &VNET_NAME(flowtable_syn_expire), 0,
 262     "seconds after which to remove syn allocated flow.");
 263 SYSCTL_VNET_INT(_net_inet_flowtable, OID_AUTO, udp_expire, CTLFLAG_RW,
 264     &VNET_NAME(flowtable_udp_expire), 0,
 265     "seconds after which to remove flow allocated to UDP.");
 266 SYSCTL_VNET_INT(_net_inet_flowtable, OID_AUTO, fin_wait_expire, CTLFLAG_RW,
 267     &VNET_NAME(flowtable_fin_wait_expire), 0,
 268     "seconds after which to remove a flow in FIN_WAIT.");
 269 SYSCTL_VNET_INT(_net_inet_flowtable, OID_AUTO, tcp_expire, CTLFLAG_RW,
 270     &VNET_NAME(flowtable_tcp_expire), 0,
 271     "seconds after which to remove flow allocated to a TCP connection.");
 272
 273
 274 /*
 275  * Maximum number of flows that can be allocated of a given type.
 276  *
 277  * The table is allocated at boot time (for the pure caching case
 278  * there is no reason why this could not be changed at runtime)
 279  * and thus (currently) needs to be set with a tunable.
 280  */
 281 static int
 282 sysctl_nmbflows(SYSCTL_HANDLER_ARGS)
 283 {
 284         int error, newnmbflows;
 285
 286         newnmbflows = V_flowtable_nmbflows;
 287         error = sysctl_handle_int(oidp, &newnmbflows, 0, req);
 288         if (error == 0 && req->newptr) {
 289                 if (newnmbflows > V_flowtable_nmbflows) {
 290                         V_flowtable_nmbflows = newnmbflows;
 291                         uma_zone_set_max(V_flow_ipv4_zone,
 292                             V_flowtable_nmbflows);
 293                         uma_zone_set_max(V_flow_ipv6_zone,
 294                             V_flowtable_nmbflows);
 295                 } else
 296                         error = EINVAL;
 297         }
 298         return (error);
 299 }
 300 SYSCTL_VNET_PROC(_net_inet_flowtable, OID_AUTO, nmbflows,
 301     CTLTYPE_INT|CTLFLAG_RW, 0, 0, sysctl_nmbflows, "IU",
 302     "Maximum number of flows allowed");
 303
 304
 305
 306 #define FS_PRINT(sb, field)     sbuf_printf((sb), "\t%s: %jd\n", #field, fs->ft_##field)
 307
 308 static void
 309 fs_print(struct sbuf *sb, struct flowtable_stats *fs)
 310 {
 311
 312         FS_PRINT(sb, collisions);
 313         FS_PRINT(sb, allocated);
 314         FS_PRINT(sb, misses);
 315         FS_PRINT(sb, max_depth);
 316         FS_PRINT(sb, free_checks);
 317         FS_PRINT(sb, frees);
 318         FS_PRINT(sb, hits);
 319         FS_PRINT(sb, lookups);
 320 }
 321
 322 static void
 323 flowtable_show_stats(struct sbuf *sb, struct flowtable *ft)
 324 {
 325         int i;
 326         struct flowtable_stats fs, *pfs;
 327
 328         if (ft->ft_flags & FL_PCPU) {
 329                 bzero(&fs, sizeof(fs));
 330                 pfs = &fs;
 331                 for (i = 0; i <= mp_maxid; i++) {
 332                         if (CPU_ABSENT(i))
 333                                 continue;
 334                         pfs->ft_collisions  += ft->ft_stats[i].ft_collisions;
 335                         pfs->ft_allocated   += ft->ft_stats[i].ft_allocated;
 336                         pfs->ft_misses      += ft->ft_stats[i].ft_misses;
 337                         pfs->ft_free_checks += ft->ft_stats[i].ft_free_checks;
 338                         pfs->ft_frees       += ft->ft_stats[i].ft_frees;
 339                         pfs->ft_hits        += ft->ft_stats[i].ft_hits;
 340                         pfs->ft_lookups     += ft->ft_stats[i].ft_lookups;
 341                         if (ft->ft_stats[i].ft_max_depth > pfs->ft_max_depth)
 342                                 pfs->ft_max_depth = ft->ft_stats[i].ft_max_depth;
 343                 }
 344         } else {
 345                 pfs = &ft->ft_stats[0];
 346         }
 347         fs_print(sb, pfs);
 348 }
 349
 350 static int
 351 sysctl_flowtable_stats(SYSCTL_HANDLER_ARGS)
 352 {
 353         struct flowtable *ft;
 354         struct sbuf *sb;
 355         int error;
 356
 357         sb = sbuf_new(NULL, NULL, 64*1024, SBUF_FIXEDLEN);
 358
 359         ft = V_flow_list_head;
 360         while (ft != NULL) {
 361                 sbuf_printf(sb, "\ntable name: %s\n", ft->ft_name);
 362                 flowtable_show_stats(sb, ft);
 363                 ft = ft->ft_next;
 364         }
 365         sbuf_finish(sb);
 366         error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
 367         sbuf_delete(sb);
 368
 369         return (error);
 370 }
 371 SYSCTL_VNET_PROC(_net_inet_flowtable, OID_AUTO, stats, CTLTYPE_STRING|CTLFLAG_RD,
 372     NULL, 0, sysctl_flowtable_stats, "A", "flowtable statistics");
 373
 374
 375 #ifndef RADIX_MPATH
 376 static void
 377 in_rtalloc_ign_wrapper(struct route *ro, uint32_t hash, u_int fibnum)
 378 {
 379
 380         rtalloc_ign_fib(ro, 0, fibnum);
 381 }
 382 #endif
 383
 384 static void
 385 flowtable_global_lock(struct flowtable *table, uint32_t hash)
 386 {
 387         int lock_index = (hash)&(table->ft_lock_count - 1);
 388
 389         mtx_lock(&table->ft_locks[lock_index]);
 390 }
 391
 392 static void
 393 flowtable_global_unlock(struct flowtable *table, uint32_t hash)
 394 {
 395         int lock_index = (hash)&(table->ft_lock_count - 1);
 396
 397         mtx_unlock(&table->ft_locks[lock_index]);
 398 }
 399
 400 static void
 401 flowtable_pcpu_lock(struct flowtable *table, uint32_t hash)
 402 {
 403
 404         critical_enter();
 405 }
 406
 407 static void
 408 flowtable_pcpu_unlock(struct flowtable *table, uint32_t hash)
 409 {
 410
 411         critical_exit();
 412 }
 413
 414 #define FL_ENTRY_INDEX(table, hash)((hash) % (table)->ft_size)
 415 #define FL_ENTRY(table, hash) *flowtable_entry((table), (hash))
 416 #define FL_ENTRY_LOCK(table, hash)  (table)->ft_lock((table), (hash))
 417 #define FL_ENTRY_UNLOCK(table, hash) (table)->ft_unlock((table), (hash))
 418
 419 #define FL_STALE        (1<<8)
 420 #define FL_IPV6         (1<<9)
 421 #define FL_OVERWRITE    (1<<10)
 422
 423 void
 424 flow_invalidate(struct flentry *fle)
 425 {
 426
 427         fle->f_flags |= FL_STALE;
 428 }
 429
 430 static __inline int
 431 proto_to_flags(uint8_t proto)
 432 {
 433         int flag;
 434
 435         switch (proto) {
 436         case IPPROTO_TCP:
 437                 flag = FL_TCP;
 438                 break;
 439         case IPPROTO_SCTP:
 440                 flag = FL_SCTP;
 441                 break;
 442         case IPPROTO_UDP:
 443                 flag = FL_UDP;
 444                 break;
 445         default:
 446                 flag = 0;
 447                 break;
 448         }
 449
 450         return (flag);
 451 }
 452
 453 static __inline int
 454 flags_to_proto(int flags)
 455 {
 456         int proto, protoflags;
 457
 458         protoflags = flags & (FL_TCP|FL_SCTP|FL_UDP);
 459         switch (protoflags) {
 460         case FL_TCP:
 461                 proto = IPPROTO_TCP;
 462                 break;
 463         case FL_SCTP:
 464                 proto = IPPROTO_SCTP;
 465                 break;
 466         case FL_UDP:
 467                 proto = IPPROTO_UDP;
 468                 break;
 469         default:
 470                 proto = 0;
 471                 break;
 472         }
 473         return (proto);
 474 }
 475
 476 #ifdef INET
 477 #ifdef FLOWTABLE_DEBUG
 478 static void
 479 ipv4_flow_print_tuple(int flags, int proto, struct sockaddr_in *ssin,
 480     struct sockaddr_in *dsin)
 481 {
 482         char saddr[4*sizeof "123"], daddr[4*sizeof "123"];
 483
 484         if (flags & FL_HASH_ALL) {
 485                 inet_ntoa_r(ssin->sin_addr, saddr);
 486                 inet_ntoa_r(dsin->sin_addr, daddr);
 487                 printf("proto=%d %s:%d->%s:%d\n",
 488                     proto, saddr, ntohs(ssin->sin_port), daddr,
 489                     ntohs(dsin->sin_port));
 490         } else {
 491                 inet_ntoa_r(*(struct in_addr *) &dsin->sin_addr, daddr);
 492                 printf("proto=%d %s\n", proto, daddr);
 493         }
 494
 495 }
 496 #endif
 497
 498 static int
 499 ipv4_mbuf_demarshal(struct flowtable *ft, struct mbuf *m,
 500     struct sockaddr_in *ssin, struct sockaddr_in *dsin, uint16_t *flags)
 501 {
 502         struct ip *ip;
 503         uint8_t proto;
 504         int iphlen;
 505         struct tcphdr *th;
 506         struct udphdr *uh;
 507         struct sctphdr *sh;
 508         uint16_t sport, dport;
 509
 510         proto = sport = dport = 0;
 511         ip = mtod(m, struct ip *);
 512         dsin->sin_family = AF_INET;
 513         dsin->sin_len = sizeof(*dsin);
 514         dsin->sin_addr = ip->ip_dst;
 515         ssin->sin_family = AF_INET;
 516         ssin->sin_len = sizeof(*ssin);
 517         ssin->sin_addr = ip->ip_src;
 518
 519         proto = ip->ip_p;
 520         if ((*flags & FL_HASH_ALL) == 0) {
 521                 FLDPRINTF(ft, FL_DEBUG_ALL, "skip port check flags=0x%x ",
 522                     *flags);
 523                 goto skipports;
 524         }
 525
 526         iphlen = ip->ip_hl << 2; /* XXX options? */
 527
 528         switch (proto) {
 529         case IPPROTO_TCP:
 530                 th = (struct tcphdr *)((caddr_t)ip + iphlen);
 531                 sport = th->th_sport;
 532                 dport = th->th_dport;
 533                 if ((*flags & FL_HASH_ALL) &&
 534                     (th->th_flags & (TH_RST|TH_FIN)))
 535                         *flags |= FL_STALE;
 536         break;
 537         case IPPROTO_UDP:
 538                 uh = (struct udphdr *)((caddr_t)ip + iphlen);
 539                 sport = uh->uh_sport;
 540                 dport = uh->uh_dport;
 541         break;
 542         case IPPROTO_SCTP:
 543                 sh = (struct sctphdr *)((caddr_t)ip + iphlen);
 544                 sport = sh->src_port;
 545                 dport = sh->dest_port;
 546         break;
 547         default:
 548                 FLDPRINTF(ft, FL_DEBUG_ALL, "proto=0x%x not supported\n", proto);
 549                 return (ENOTSUP);
 550                 /* no port - hence not a protocol we care about */
 551                 break;
 552
 553         }
 554
 555 skipports:
 556         *flags |= proto_to_flags(proto);
 557         ssin->sin_port = sport;
 558         dsin->sin_port = dport;
 559         return (0);
 560 }
 561
 562 static uint32_t
 563 ipv4_flow_lookup_hash_internal(
 564         struct sockaddr_in *ssin, struct sockaddr_in *dsin,
 565             uint32_t *key, uint16_t flags)
 566 {
 567         uint16_t sport, dport;
 568         uint8_t proto;
 569         int offset = 0;
 570
 571         if ((V_flowtable_enable == 0) || (V_flowtable_ready == 0))
 572                 return (0);
 573         proto = flags_to_proto(flags);
 574         sport = dport = key[2] = key[1] = key[0] = 0;
 575         if ((ssin != NULL) && (flags & FL_HASH_ALL)) {
 576                 key[1] = ssin->sin_addr.s_addr;
 577                 sport = ssin->sin_port;
 578         }
 579         if (dsin != NULL) {
 580                 key[2] = dsin->sin_addr.s_addr;
 581                 dport = dsin->sin_port;
 582         }
 583         if (flags & FL_HASH_ALL) {
 584                 ((uint16_t *)key)[0] = sport;
 585                 ((uint16_t *)key)[1] = dport;
 586         } else
 587                 offset = V_flow_hashjitter + proto;
 588
 589         return (jenkins_hashword(key, 3, offset));
 590 }
 591
 592 static struct flentry *
 593 flowtable_lookup_mbuf4(struct flowtable *ft, struct mbuf *m)
 594 {
 595         struct sockaddr_storage ssa, dsa;
 596         uint16_t flags;
 597         struct sockaddr_in *dsin, *ssin;
 598
 599         dsin = (struct sockaddr_in *)&dsa;
 600         ssin = (struct sockaddr_in *)&ssa;
 601         bzero(dsin, sizeof(*dsin));
 602         bzero(ssin, sizeof(*ssin));
 603         flags = ft->ft_flags;
 604         if (ipv4_mbuf_demarshal(ft, m, ssin, dsin, &flags) != 0)
 605                 return (NULL);
 606
 607         return (flowtable_lookup(ft, &ssa, &dsa, M_GETFIB(m), flags));
 608 }
 609
 610 void
 611 flow_to_route(struct flentry *fle, struct route *ro)
 612 {
 613         uint32_t *hashkey = NULL;
 614         struct sockaddr_in *sin;
 615
 616         sin = (struct sockaddr_in *)&ro->ro_dst;
 617         sin->sin_family = AF_INET;
 618         sin->sin_len = sizeof(*sin);
 619         hashkey = ((struct flentry_v4 *)fle)->fl_flow.ipf_key;
 620         sin->sin_addr.s_addr = hashkey[2];
 621         ro->ro_rt = __DEVOLATILE(struct rtentry *, fle->f_rt);
 622         ro->ro_lle = __DEVOLATILE(struct llentry *, fle->f_lle);
 623 }
 624 #endif /* INET */
 625
 626 #ifdef INET6
 627 /*
 628  * PULLUP_TO(len, p, T) makes sure that len + sizeof(T) is contiguous,
 629  * then it sets p to point at the offset "len" in the mbuf. WARNING: the
 630  * pointer might become stale after other pullups (but we never use it
 631  * this way).
 632  */
 633 #define PULLUP_TO(_len, p, T)                                           \
 634 do {                                                                    \
 635         int x = (_len) + sizeof(T);                                     \
 636         if ((m)->m_len < x) {                                           \
 637                 goto receive_failed;                                    \
 638         }                                                               \
 639         p = (mtod(m, char *) + (_len));                                 \
 640 } while (0)
 641
 642 #define TCP(p)          ((struct tcphdr *)(p))
 643 #define SCTP(p)         ((struct sctphdr *)(p))
 644 #define UDP(p)          ((struct udphdr *)(p))
 645
 646 static int
 647 ipv6_mbuf_demarshal(struct flowtable *ft, struct mbuf *m,
 648     struct sockaddr_in6 *ssin6, struct sockaddr_in6 *dsin6, uint16_t *flags)
 649 {
 650         struct ip6_hdr *ip6;
 651         uint8_t proto;
 652         int hlen;
 653         uint16_t src_port, dst_port;
 654         u_short offset;
 655         void *ulp;
 656
 657         offset = hlen = src_port = dst_port = 0;
 658         ulp = NULL;
 659         ip6 = mtod(m, struct ip6_hdr *);
 660         hlen = sizeof(struct ip6_hdr);
 661         proto = ip6->ip6_nxt;
 662
 663         if ((*flags & FL_HASH_ALL) == 0)
 664                 goto skipports;
 665
 666         while (ulp == NULL) {
 667                 switch (proto) {
 668                 case IPPROTO_ICMPV6:
 669                 case IPPROTO_OSPFIGP:
 670                 case IPPROTO_PIM:
 671                 case IPPROTO_CARP:
 672                 case IPPROTO_ESP:
 673                 case IPPROTO_NONE:
 674                         ulp = ip6;
 675                         break;
 676                 case IPPROTO_TCP:
 677                         PULLUP_TO(hlen, ulp, struct tcphdr);
 678                         dst_port = TCP(ulp)->th_dport;
 679                         src_port = TCP(ulp)->th_sport;
 680                         if ((*flags & FL_HASH_ALL) &&
 681                             (TCP(ulp)->th_flags & (TH_RST|TH_FIN)))
 682                                 *flags |= FL_STALE;
 683                         break;
 684                 case IPPROTO_SCTP:
 685                         PULLUP_TO(hlen, ulp, struct sctphdr);
 686                         src_port = SCTP(ulp)->src_port;
 687                         dst_port = SCTP(ulp)->dest_port;
 688                         break;
 689                 case IPPROTO_UDP:
 690                         PULLUP_TO(hlen, ulp, struct udphdr);
 691                         dst_port = UDP(ulp)->uh_dport;
 692                         src_port = UDP(ulp)->uh_sport;
 693                         break;
 694                 case IPPROTO_HOPOPTS:   /* RFC 2460 */
 695                         PULLUP_TO(hlen, ulp, struct ip6_hbh);
 696                         hlen += (((struct ip6_hbh *)ulp)->ip6h_len + 1) << 3;
 697                         proto = ((struct ip6_hbh *)ulp)->ip6h_nxt;
 698                         ulp = NULL;
 699                         break;
 700                 case IPPROTO_ROUTING:   /* RFC 2460 */
 701                         PULLUP_TO(hlen, ulp, struct ip6_rthdr);
 702                         hlen += (((struct ip6_rthdr *)ulp)->ip6r_len + 1) << 3;
 703                         proto = ((struct ip6_rthdr *)ulp)->ip6r_nxt;
 704                         ulp = NULL;
 705                         break;
 706                 case IPPROTO_FRAGMENT:  /* RFC 2460 */
 707                         PULLUP_TO(hlen, ulp, struct ip6_frag);
 708                         hlen += sizeof (struct ip6_frag);
 709                         proto = ((struct ip6_frag *)ulp)->ip6f_nxt;
 710                         offset = ((struct ip6_frag *)ulp)->ip6f_offlg &
 711                             IP6F_OFF_MASK;
 712                         ulp = NULL;
 713                         break;
 714                 case IPPROTO_DSTOPTS:   /* RFC 2460 */
 715                         PULLUP_TO(hlen, ulp, struct ip6_hbh);
 716                         hlen += (((struct ip6_hbh *)ulp)->ip6h_len + 1) << 3;
 717                         proto = ((struct ip6_hbh *)ulp)->ip6h_nxt;
 718                         ulp = NULL;
 719                         break;
 720                 case IPPROTO_AH:        /* RFC 2402 */
 721                         PULLUP_TO(hlen, ulp, struct ip6_ext);
 722                         hlen += (((struct ip6_ext *)ulp)->ip6e_len + 2) << 2;
 723                         proto = ((struct ip6_ext *)ulp)->ip6e_nxt;
 724                         ulp = NULL;
 725                         break;
 726                 default:
 727                         PULLUP_TO(hlen, ulp, struct ip6_ext);
 728                         break;
 729                 }
 730         }
 731
 732         if (src_port == 0) {
 733         receive_failed:
 734                 return (ENOTSUP);
 735         }
 736
 737 skipports:
 738         dsin6->sin6_family = AF_INET6;
 739         dsin6->sin6_len = sizeof(*dsin6);
 740         dsin6->sin6_port = dst_port;
 741         memcpy(&dsin6->sin6_addr, &ip6->ip6_dst, sizeof(struct in6_addr));
 742
 743         ssin6->sin6_family = AF_INET6;
 744         ssin6->sin6_len = sizeof(*ssin6);
 745         ssin6->sin6_port = src_port;
 746         memcpy(&ssin6->sin6_addr, &ip6->ip6_src, sizeof(struct in6_addr));
 747         *flags |= proto_to_flags(proto);
 748
 749         return (0);
 750 }
 751
 752 #define zero_key(key)           \
 753 do {                            \
 754         key[0] = 0;             \
 755         key[1] = 0;             \
 756         key[2] = 0;             \
 757         key[3] = 0;             \
 758         key[4] = 0;             \
 759         key[5] = 0;             \
 760         key[6] = 0;             \
 761         key[7] = 0;             \
 762         key[8] = 0;             \
 763 } while (0)
 764
 765 static uint32_t
 766 ipv6_flow_lookup_hash_internal(
 767         struct sockaddr_in6 *ssin6, struct sockaddr_in6 *dsin6,
 768             uint32_t *key, uint16_t flags)
 769 {
 770         uint16_t sport, dport;
 771         uint8_t proto;
 772         int offset = 0;
 773
 774         if ((V_flowtable_enable == 0) || (V_flowtable_ready == 0))
 775                 return (0);
 776
 777         proto = flags_to_proto(flags);
 778         zero_key(key);
 779         sport = dport = 0;
 780         if (dsin6 != NULL) {
 781                 memcpy(&key[1], &dsin6->sin6_addr, sizeof(struct in6_addr));
 782                 dport = dsin6->sin6_port;
 783         }
 784         if ((ssin6 != NULL) && (flags & FL_HASH_ALL)) {
 785                 memcpy(&key[5], &ssin6->sin6_addr, sizeof(struct in6_addr));
 786                 sport = ssin6->sin6_port;
 787         }
 788         if (flags & FL_HASH_ALL) {
 789                 ((uint16_t *)key)[0] = sport;
 790                 ((uint16_t *)key)[1] = dport;
 791         } else
 792                 offset = V_flow_hashjitter + proto;
 793
 794         return (jenkins_hashword(key, 9, offset));
 795 }
 796
 797 static struct flentry *
 798 flowtable_lookup_mbuf6(struct flowtable *ft, struct mbuf *m)
 799 {
 800         struct sockaddr_storage ssa, dsa;
 801         struct sockaddr_in6 *dsin6, *ssin6;
 802         uint16_t flags;
 803
 804         dsin6 = (struct sockaddr_in6 *)&dsa;
 805         ssin6 = (struct sockaddr_in6 *)&ssa;
 806         bzero(dsin6, sizeof(*dsin6));
 807         bzero(ssin6, sizeof(*ssin6));
 808         flags = ft->ft_flags;
 809
 810         if (ipv6_mbuf_demarshal(ft, m, ssin6, dsin6, &flags) != 0)
 811                 return (NULL);
 812
 813         return (flowtable_lookup(ft, &ssa, &dsa, M_GETFIB(m), flags));
 814 }
 815
 816 void
 817 flow_to_route_in6(struct flentry *fle, struct route_in6 *ro)
 818 {
 819         uint32_t *hashkey = NULL;
 820         struct sockaddr_in6 *sin6;
 821
 822         sin6 = (struct sockaddr_in6 *)&ro->ro_dst;
 823
 824         sin6->sin6_family = AF_INET6;
 825         sin6->sin6_len = sizeof(*sin6);
 826         hashkey = ((struct flentry_v6 *)fle)->fl_flow.ipf_key;
 827         memcpy(&sin6->sin6_addr, &hashkey[5], sizeof (struct in6_addr));
 828         ro->ro_rt = __DEVOLATILE(struct rtentry *, fle->f_rt);
 829         ro->ro_lle = __DEVOLATILE(struct llentry *, fle->f_lle);
 830
 831 }
 832 #endif /* INET6 */
 833
 834 static bitstr_t *
 835 flowtable_mask(struct flowtable *ft)
 836 {
 837         bitstr_t *mask;
 838
 839         if (ft->ft_flags & FL_PCPU)
 840                 mask = ft->ft_masks[curcpu];
 841         else
 842                 mask = ft->ft_masks[0];
 843
 844         return (mask);
 845 }
 846
 847 static struct flentry **
 848 flowtable_entry(struct flowtable *ft, uint32_t hash)
 849 {
 850         struct flentry **fle;
 851         int index = (hash % ft->ft_size);
 852
 853         if (ft->ft_flags & FL_PCPU) {
 854                 KASSERT(&ft->ft_table.pcpu[curcpu][0] != NULL, ("pcpu not set"));
 855                 fle = &ft->ft_table.pcpu[curcpu][index];
 856         } else {
 857                 KASSERT(&ft->ft_table.global[0] != NULL, ("global not set"));
 858                 fle = &ft->ft_table.global[index];
 859         }
 860
 861         return (fle);
 862 }
 863
 864 static int
 865 flow_stale(struct flowtable *ft, struct flentry *fle)
 866 {
 867         time_t idle_time;
 868
 869         if ((fle->f_fhash == 0)
 870             || ((fle->f_rt->rt_flags & RTF_HOST) &&
 871                 ((fle->f_rt->rt_flags & (RTF_UP))
 872                     != (RTF_UP)))
 873             || (fle->f_rt->rt_ifp == NULL)
 874             || !RT_LINK_IS_UP(fle->f_rt->rt_ifp))
 875                 return (1);
 876
 877         idle_time = time_uptime - fle->f_uptime;
 878
 879         if ((fle->f_flags & FL_STALE) ||
 880             ((fle->f_flags & (TH_SYN|TH_ACK|TH_FIN)) == 0
 881                 && (idle_time > ft->ft_udp_idle)) ||
 882             ((fle->f_flags & TH_FIN)
 883                 && (idle_time > ft->ft_fin_wait_idle)) ||
 884             ((fle->f_flags & (TH_SYN|TH_ACK)) == TH_SYN
 885                 && (idle_time > ft->ft_syn_idle)) ||
 886             ((fle->f_flags & (TH_SYN|TH_ACK)) == (TH_SYN|TH_ACK)
 887                 && (idle_time > ft->ft_tcp_idle)) ||
 888             ((fle->f_rt->rt_flags & RTF_UP) == 0 ||
 889                 (fle->f_rt->rt_ifp == NULL)))
 890                 return (1);
 891
 892         return (0);
 893 }
 894
 895 static void
 896 flowtable_set_hashkey(struct flentry *fle, uint32_t *key)
 897 {
 898         uint32_t *hashkey;
 899         int i, nwords;
 900
 901         if (fle->f_flags & FL_IPV6) {
 902                 nwords = 9;
 903                 hashkey = ((struct flentry_v4 *)fle)->fl_flow.ipf_key;
 904         } else {
 905                 nwords = 3;
 906                 hashkey = ((struct flentry_v6 *)fle)->fl_flow.ipf_key;
 907         }
 908
 909         for (i = 0; i < nwords; i++)
 910                 hashkey[i] = key[i];
 911 }
 912
 913 static struct flentry *
 914 flow_alloc(struct flowtable *ft)
 915 {
 916         struct flentry *newfle;
 917         uma_zone_t zone;
 918
 919         newfle = NULL;
 920         zone = (ft->ft_flags & FL_IPV6) ? V_flow_ipv6_zone : V_flow_ipv4_zone;
 921
 922         newfle = uma_zalloc(zone, M_NOWAIT | M_ZERO);
 923         if (newfle != NULL)
 924                 atomic_add_int(&ft->ft_count, 1);
 925         return (newfle);
 926 }
 927
 928 static void
 929 flow_free(struct flentry *fle, struct flowtable *ft)
 930 {
 931         uma_zone_t zone;
 932
 933         zone = (ft->ft_flags & FL_IPV6) ? V_flow_ipv6_zone : V_flow_ipv4_zone;
 934         atomic_add_int(&ft->ft_count, -1);
 935         uma_zfree(zone, fle);
 936 }
 937
 938 static int
 939 flow_full(struct flowtable *ft)
 940 {
 941         boolean_t full;
 942         uint32_t count;
 943
 944         full = ft->ft_full;
 945         count = ft->ft_count;
 946
 947         if (full && (count < (V_flowtable_nmbflows - (V_flowtable_nmbflows >> 3))))
 948                 ft->ft_full = FALSE;
 949         else if (!full && (count > (V_flowtable_nmbflows - (V_flowtable_nmbflows >> 5))))
 950                 ft->ft_full = TRUE;
 951
 952         if (full && !ft->ft_full) {
 953                 flowclean_freq = 4*hz;
 954                 if ((ft->ft_flags & FL_HASH_ALL) == 0)
 955                         ft->ft_udp_idle = ft->ft_fin_wait_idle =
 956                             ft->ft_syn_idle = ft->ft_tcp_idle = 5;
 957                 cv_broadcast(&flowclean_cv);
 958         } else if (!full && ft->ft_full) {
 959                 flowclean_freq = 20*hz;
 960                 if ((ft->ft_flags & FL_HASH_ALL) == 0)
 961                         ft->ft_udp_idle = ft->ft_fin_wait_idle =
 962                             ft->ft_syn_idle = ft->ft_tcp_idle = 30;
 963         }
 964
 965         return (ft->ft_full);
 966 }
 967
 968 static int
 969 flowtable_insert(struct flowtable *ft, uint32_t hash, uint32_t *key,
 970     uint32_t fibnum, struct route *ro, uint16_t flags)
 971 {
 972         struct flentry *fle, *fletail, *newfle, **flep;
 973         struct flowtable_stats *fs = &ft->ft_stats[curcpu];
 974         int depth;
 975         bitstr_t *mask;
 976         uint8_t proto;
 977
 978         newfle = flow_alloc(ft);
 979         if (newfle == NULL)
 980                 return (ENOMEM);
 981
 982         newfle->f_flags |= (flags & FL_IPV6);
 983         proto = flags_to_proto(flags);
 984
 985         FL_ENTRY_LOCK(ft, hash);
 986         mask = flowtable_mask(ft);
 987         flep = flowtable_entry(ft, hash);
 988         fletail = fle = *flep;
 989
 990         if (fle == NULL) {
 991                 bit_set(mask, FL_ENTRY_INDEX(ft, hash));
 992                 *flep = fle = newfle;
 993                 goto skip;
 994         }
 995
 996         depth = 0;
 997         fs->ft_collisions++;
 998         /*
 999          * find end of list and make sure that we were not
1000          * preempted by another thread handling this flow
1001          */
1002         while (fle != NULL) {
1003                 if (fle->f_fhash == hash && !flow_stale(ft, fle)) {
1004                         /*
1005                          * there was either a hash collision
1006                          * or we lost a race to insert
1007                          */
1008                         FL_ENTRY_UNLOCK(ft, hash);
1009                         flow_free(newfle, ft);
1010
1011                         if (flags & FL_OVERWRITE)
1012                                 goto skip;
1013                         return (EEXIST);
1014                 }
1015                 /*
1016                  * re-visit this double condition XXX
1017                  */
1018                 if (fletail->f_next != NULL)
1019                         fletail = fle->f_next;
1020
1021                 depth++;
1022                 fle = fle->f_next;
1023         }
1024
1025         if (depth > fs->ft_max_depth)
1026                 fs->ft_max_depth = depth;
1027         fletail->f_next = newfle;
1028         fle = newfle;
1029 skip:
1030         flowtable_set_hashkey(fle, key);
1031
1032         fle->f_proto = proto;
1033         fle->f_rt = ro->ro_rt;
1034         fle->f_lle = ro->ro_lle;
1035         fle->f_fhash = hash;
1036         fle->f_fibnum = fibnum;
1037         fle->f_uptime = time_uptime;
1038         FL_ENTRY_UNLOCK(ft, hash);
1039         return (0);
1040 }
1041
1042 int
1043 kern_flowtable_insert(struct flowtable *ft,
1044     struct sockaddr_storage *ssa, struct sockaddr_storage *dsa,
1045     struct route *ro, uint32_t fibnum, int flags)
1046 {
1047         uint32_t key[9], hash;
1048
1049         flags = (ft->ft_flags | flags | FL_OVERWRITE);
1050         hash = 0;
1051
1052 #ifdef INET
1053         if (ssa->ss_family == AF_INET)
1054                 hash = ipv4_flow_lookup_hash_internal((struct sockaddr_in *)ssa,
1055                     (struct sockaddr_in *)dsa, key, flags);
1056 #endif
1057 #ifdef INET6
1058         if (ssa->ss_family == AF_INET6)
1059                 hash = ipv6_flow_lookup_hash_internal((struct sockaddr_in6 *)ssa,
1060                     (struct sockaddr_in6 *)dsa, key, flags);
1061 #endif
1062         if (ro->ro_rt == NULL || ro->ro_lle == NULL)
1063                 return (EINVAL);
1064
1065         FLDPRINTF(ft, FL_DEBUG,
1066             "kern_flowtable_insert: key=%x:%x:%x hash=%x fibnum=%d flags=%x\n",
1067             key[0], key[1], key[2], hash, fibnum, flags);
1068         return (flowtable_insert(ft, hash, key, fibnum, ro, flags));
1069 }
1070
1071 static int
1072 flowtable_key_equal(struct flentry *fle, uint32_t *key)
1073 {
1074         uint32_t *hashkey;
1075         int i, nwords;
1076
1077         if (fle->f_flags & FL_IPV6) {
1078                 nwords = 9;
1079                 hashkey = ((struct flentry_v4 *)fle)->fl_flow.ipf_key;
1080         } else {
1081                 nwords = 3;
1082                 hashkey = ((struct flentry_v6 *)fle)->fl_flow.ipf_key;
1083         }
1084
1085         for (i = 0; i < nwords; i++)
1086                 if (hashkey[i] != key[i])
1087                         return (0);
1088
1089         return (1);
1090 }
1091
1092 struct flentry *
1093 flowtable_lookup_mbuf(struct flowtable *ft, struct mbuf *m, int af)
1094 {
1095         struct flentry *fle = NULL;
1096
1097 #ifdef INET
1098         if (af == AF_INET)
1099                 fle = flowtable_lookup_mbuf4(ft, m);
1100 #endif
1101 #ifdef INET6
1102         if (af == AF_INET6)
1103                 fle = flowtable_lookup_mbuf6(ft, m);
1104 #endif
1105         if (fle != NULL && m != NULL && (m->m_flags & M_FLOWID) == 0) {
1106                 m->m_flags |= M_FLOWID;
1107                 m->m_pkthdr.flowid = fle->f_fhash;
1108         }
1109         return (fle);
1110 }
1111
1112 struct flentry *
1113 flowtable_lookup(struct flowtable *ft, struct sockaddr_storage *ssa,
1114     struct sockaddr_storage *dsa, uint32_t fibnum, int flags)
1115 {
1116         uint32_t key[9], hash;
1117         struct flentry *fle;
1118         struct flowtable_stats *fs = &ft->ft_stats[curcpu];
1119         uint8_t proto = 0;
1120         int error = 0;
1121         struct rtentry *rt;
1122         struct llentry *lle;
1123         struct route sro, *ro;
1124         struct route_in6 sro6;
1125
1126         sro.ro_rt = sro6.ro_rt = NULL;
1127         sro.ro_lle = sro6.ro_lle = NULL;
1128         ro = NULL;
1129         hash = 0;
1130         flags |= ft->ft_flags;
1131         proto = flags_to_proto(flags);
1132 #ifdef INET
1133         if (ssa->ss_family == AF_INET) {
1134                 struct sockaddr_in *ssin, *dsin;
1135
1136                 ro = &sro;
1137                 memcpy(&ro->ro_dst, dsa, sizeof(struct sockaddr_in));
1138                 /*
1139                  * The harvested source and destination addresses
1140                  * may contain port information if the packet is
1141                  * from a transport protocol (e.g. TCP/UDP). The
1142                  * port field must be cleared before performing
1143                  * a route lookup.
1144                  */
1145                 ((struct sockaddr_in *)&ro->ro_dst)->sin_port = 0;
1146                 dsin = (struct sockaddr_in *)dsa;
1147                 ssin = (struct sockaddr_in *)ssa;
1148                 if ((dsin->sin_addr.s_addr == ssin->sin_addr.s_addr) ||
1149                     (ntohl(dsin->sin_addr.s_addr) >> IN_CLASSA_NSHIFT) == IN_LOOPBACKNET ||
1150                     (ntohl(ssin->sin_addr.s_addr) >> IN_CLASSA_NSHIFT) == IN_LOOPBACKNET)
1151                         return (NULL);
1152
1153                 hash = ipv4_flow_lookup_hash_internal(ssin, dsin, key, flags);
1154         }
1155 #endif
1156 #ifdef INET6
1157         if (ssa->ss_family == AF_INET6) {
1158                 struct sockaddr_in6 *ssin6, *dsin6;
1159
1160                 ro = (struct route *)&sro6;
1161                 memcpy(&sro6.ro_dst, dsa,
1162                     sizeof(struct sockaddr_in6));
1163                 ((struct sockaddr_in6 *)&ro->ro_dst)->sin6_port = 0;
1164                 dsin6 = (struct sockaddr_in6 *)dsa;
1165                 ssin6 = (struct sockaddr_in6 *)ssa;
1166
1167                 flags |= FL_IPV6;
1168                 hash = ipv6_flow_lookup_hash_internal(ssin6, dsin6, key, flags);
1169         }
1170 #endif
1171         /*
1172          * Ports are zero and this isn't a transmit cache
1173          * - thus not a protocol for which we need to keep
1174          * state
1175          * FL_HASH_ALL => key[0] != 0 for TCP || UDP || SCTP
1176          */
1177         if (hash == 0 || (key[0] == 0 && (ft->ft_flags & FL_HASH_ALL)))
1178                 return (NULL);
1179
1180         fs->ft_lookups++;
1181         FL_ENTRY_LOCK(ft, hash);
1182         if ((fle = FL_ENTRY(ft, hash)) == NULL) {
1183                 FL_ENTRY_UNLOCK(ft, hash);
1184                 goto uncached;
1185         }
1186 keycheck:
1187         rt = __DEVOLATILE(struct rtentry *, fle->f_rt);
1188         lle = __DEVOLATILE(struct llentry *, fle->f_lle);
1189         if ((rt != NULL)
1190             && fle->f_fhash == hash
1191             && flowtable_key_equal(fle, key)
1192             && (proto == fle->f_proto)
1193             && (fibnum == fle->f_fibnum)
1194             && (rt->rt_flags & RTF_UP)
1195             && (rt->rt_ifp != NULL)) {
1196                 fs->ft_hits++;
1197                 fle->f_uptime = time_uptime;
1198                 fle->f_flags |= flags;
1199                 FL_ENTRY_UNLOCK(ft, hash);
1200                 return (fle);
1201         } else if (fle->f_next != NULL) {
1202                 fle = fle->f_next;
1203                 goto keycheck;
1204         }
1205         FL_ENTRY_UNLOCK(ft, hash);
1206 uncached:
1207         if (flags & FL_NOAUTO || flow_full(ft))
1208                 return (NULL);
1209
1210         fs->ft_misses++;
1211         /*
1212          * This bit of code ends up locking the
1213          * same route 3 times (just like ip_output + ether_output)
1214          * - at lookup
1215          * - in rt_check when called by arpresolve
1216          * - dropping the refcount for the rtentry
1217          *
1218          * This could be consolidated to one if we wrote a variant
1219          * of arpresolve with an rt_check variant that expected to
1220          * receive the route locked
1221          */
1222
1223 #ifdef INVARIANTS
1224         if ((ro->ro_dst.sa_family != AF_INET) &&
1225             (ro->ro_dst.sa_family != AF_INET6))
1226                 panic("sa_family == %d\n", ro->ro_dst.sa_family);
1227 #endif
1228
1229         ft->ft_rtalloc(ro, hash, fibnum);
1230         if (ro->ro_rt == NULL)
1231                 error = ENETUNREACH;
1232         else {
1233                 struct llentry *lle = NULL;
1234                 struct sockaddr_storage *l3addr;
1235                 struct rtentry *rt = ro->ro_rt;
1236                 struct ifnet *ifp = rt->rt_ifp;
1237
1238                 if (ifp->if_flags & (IFF_POINTOPOINT | IFF_LOOPBACK)) {
1239                         RTFREE(rt);
1240                         ro->ro_rt = NULL;
1241                         return (NULL);
1242                 }
1243 #ifdef INET6
1244                 if (ssa->ss_family == AF_INET6) {
1245                         struct sockaddr_in6 *dsin6;
1246
1247                         dsin6 = (struct sockaddr_in6 *)dsa;
1248                         if (in6_localaddr(&dsin6->sin6_addr)) {
1249                                 RTFREE(rt);
1250                                 ro->ro_rt = NULL;
1251                                 return (NULL);
1252                         }
1253
1254                         if (rt->rt_flags & RTF_GATEWAY)
1255                                 l3addr = (struct sockaddr_storage *)rt->rt_gateway;
1256
1257                         else
1258                                 l3addr = (struct sockaddr_storage *)&ro->ro_dst;
1259                         llentry_update(&lle, LLTABLE6(ifp), l3addr, ifp);
1260                 }
1261 #endif
1262 #ifdef INET
1263                 if (ssa->ss_family == AF_INET) {
1264                         if (rt->rt_flags & RTF_GATEWAY)
1265                                 l3addr = (struct sockaddr_storage *)rt->rt_gateway;
1266                         else
1267                                 l3addr = (struct sockaddr_storage *)&ro->ro_dst;
1268                         llentry_update(&lle, LLTABLE(ifp), l3addr, ifp);
1269                 }
1270
1271 #endif
1272                 ro->ro_lle = lle;
1273
1274                 if (lle == NULL) {
1275                         RTFREE(rt);
1276                         ro->ro_rt = NULL;
1277                         return (NULL);
1278                 }
1279                 error = flowtable_insert(ft, hash, key, fibnum, ro, flags);
1280
1281                 if (error) {
1282                         RTFREE(rt);
1283                         LLE_FREE(lle);
1284                         ro->ro_rt = NULL;
1285                         ro->ro_lle = NULL;
1286                 }
1287         }
1288
1289         return ((error) ? NULL : fle);
1290 }
1291
1292 /*
1293  * used by the bit_alloc macro
1294  */
1295 #define calloc(count, size) malloc((count)*(size), M_DEVBUF, M_WAITOK|M_ZERO)
1296
1297 struct flowtable *
1298 flowtable_alloc(char *name, int nentry, int flags)
1299 {
1300         struct flowtable *ft, *fttail;
1301         int i;
1302
1303         if (V_flow_hashjitter == 0)
1304                 V_flow_hashjitter = arc4random();
1305
1306         KASSERT(nentry > 0, ("nentry must be > 0, is %d\n", nentry));
1307
1308         ft = malloc(sizeof(struct flowtable),
1309             M_RTABLE, M_WAITOK | M_ZERO);
1310
1311         ft->ft_name = name;
1312         ft->ft_flags = flags;
1313         ft->ft_size = nentry;
1314 #ifdef RADIX_MPATH
1315         ft->ft_rtalloc = rtalloc_mpath_fib;
1316 #else
1317         ft->ft_rtalloc = in_rtalloc_ign_wrapper;
1318 #endif
1319         if (flags & FL_PCPU) {
1320                 ft->ft_lock = flowtable_pcpu_lock;
1321                 ft->ft_unlock = flowtable_pcpu_unlock;
1322
1323                 for (i = 0; i <= mp_maxid; i++) {
1324                         ft->ft_table.pcpu[i] =
1325                             malloc(nentry*sizeof(struct flentry *),
1326                                 M_RTABLE, M_WAITOK | M_ZERO);
1327                         ft->ft_masks[i] = bit_alloc(nentry);
1328                 }
1329         } else {
1330                 ft->ft_lock_count = 2*(powerof2(mp_maxid + 1) ? (mp_maxid + 1):
1331                     (fls(mp_maxid + 1) << 1));
1332
1333                 ft->ft_lock = flowtable_global_lock;
1334                 ft->ft_unlock = flowtable_global_unlock;
1335                 ft->ft_table.global =
1336                             malloc(nentry*sizeof(struct flentry *),
1337                                 M_RTABLE, M_WAITOK | M_ZERO);
1338                 ft->ft_locks = malloc(ft->ft_lock_count*sizeof(struct mtx),
1339                                 M_RTABLE, M_WAITOK | M_ZERO);
1340                 for (i = 0; i < ft->ft_lock_count; i++)
1341                         mtx_init(&ft->ft_locks[i], "flow", NULL, MTX_DEF|MTX_DUPOK);
1342
1343                 ft->ft_masks[0] = bit_alloc(nentry);
1344         }
1345         ft->ft_tmpmask = bit_alloc(nentry);
1346
1347         /*
1348          * In the local transmit case the table truly is
1349          * just a cache - so everything is eligible for
1350          * replacement after 5s of non-use
1351          */
1352         if (flags & FL_HASH_ALL) {
1353                 ft->ft_udp_idle = V_flowtable_udp_expire;
1354                 ft->ft_syn_idle = V_flowtable_syn_expire;
1355                 ft->ft_fin_wait_idle = V_flowtable_fin_wait_expire;
1356                 ft->ft_tcp_idle = V_flowtable_fin_wait_expire;
1357         } else {
1358                 ft->ft_udp_idle = ft->ft_fin_wait_idle =
1359                     ft->ft_syn_idle = ft->ft_tcp_idle = 30;
1360
1361         }
1362
1363         /*
1364          * hook in to the cleaner list
1365          */
1366         if (V_flow_list_head == NULL)
1367                 V_flow_list_head = ft;
1368         else {
1369                 fttail = V_flow_list_head;
1370                 while (fttail->ft_next != NULL)
1371                         fttail = fttail->ft_next;
1372                 fttail->ft_next = ft;
1373         }
1374
1375         return (ft);
1376 }
1377
1378 /*
1379  * The rest of the code is devoted to garbage collection of expired entries.
1380  * It is a new additon made necessary by the switch to dynamically allocating
1381  * flow tables.
1382  *
1383  */
1384 static void
1385 fle_free(struct flentry *fle, struct flowtable *ft)
1386 {
1387         struct rtentry *rt;
1388         struct llentry *lle;
1389
1390         rt = __DEVOLATILE(struct rtentry *, fle->f_rt);
1391         lle = __DEVOLATILE(struct llentry *, fle->f_lle);
1392         RTFREE(rt);
1393         LLE_FREE(lle);
1394         flow_free(fle, ft);
1395 }
1396
1397 static void
1398 flowtable_free_stale(struct flowtable *ft, struct rtentry *rt)
1399 {
1400         int curbit = 0, count;
1401         struct flentry *fle,  **flehead, *fleprev;
1402         struct flentry *flefreehead, *flefreetail, *fletmp;
1403         bitstr_t *mask, *tmpmask;
1404         struct flowtable_stats *fs = &ft->ft_stats[curcpu];
1405
1406         flefreehead = flefreetail = NULL;
1407         mask = flowtable_mask(ft);
1408         tmpmask = ft->ft_tmpmask;
1409         memcpy(tmpmask, mask, ft->ft_size/8);
1410         /*
1411          * XXX Note to self, bit_ffs operates at the byte level
1412          * and thus adds gratuitous overhead
1413          */
1414         bit_ffs(tmpmask, ft->ft_size, &curbit);
1415         while (curbit != -1) {
1416                 if (curbit >= ft->ft_size || curbit < -1) {
1417                         log(LOG_ALERT,
1418                             "warning: bad curbit value %d \n",
1419                             curbit);
1420                         break;
1421                 }
1422
1423                 FL_ENTRY_LOCK(ft, curbit);
1424                 flehead = flowtable_entry(ft, curbit);
1425                 fle = fleprev = *flehead;
1426
1427                 fs->ft_free_checks++;
1428 #ifdef DIAGNOSTIC
1429                 if (fle == NULL && curbit > 0) {
1430                         log(LOG_ALERT,
1431                             "warning bit=%d set, but no fle found\n",
1432                             curbit);
1433                 }
1434 #endif
1435                 while (fle != NULL) {
1436                         if (rt != NULL) {
1437                                 if (__DEVOLATILE(struct rtentry *, fle->f_rt) != rt) {
1438                                         fleprev = fle;
1439                                         fle = fle->f_next;
1440                                         continue;
1441                                 }
1442                         } else if (!flow_stale(ft, fle)) {
1443                                 fleprev = fle;
1444                                 fle = fle->f_next;
1445                                 continue;
1446                         }
1447                         /*
1448                          * delete head of the list
1449                          */
1450                         if (fleprev == *flehead) {
1451                                 fletmp = fleprev;
1452                                 if (fle == fleprev) {
1453                                         fleprev = *flehead = fle->f_next;
1454                                 } else
1455                                         fleprev = *flehead = fle;
1456                                 fle = fle->f_next;
1457                         } else {
1458                                 /*
1459                                  * don't advance fleprev
1460                                  */
1461                                 fletmp = fle;
1462                                 fleprev->f_next = fle->f_next;
1463                                 fle = fleprev->f_next;
1464                         }
1465
1466                         if (flefreehead == NULL)
1467                                 flefreehead = flefreetail = fletmp;
1468                         else {
1469                                 flefreetail->f_next = fletmp;
1470                                 flefreetail = fletmp;
1471                         }
1472                         fletmp->f_next = NULL;
1473                 }
1474                 if (*flehead == NULL)
1475                         bit_clear(mask, curbit);
1476                 FL_ENTRY_UNLOCK(ft, curbit);
1477                 bit_clear(tmpmask, curbit);
1478                 bit_ffs(tmpmask, ft->ft_size, &curbit);
1479         }
1480         count = 0;
1481         while ((fle = flefreehead) != NULL) {
1482                 flefreehead = fle->f_next;
1483                 count++;
1484                 fs->ft_frees++;
1485                 fle_free(fle, ft);
1486         }
1487         if (V_flowtable_debug && count)
1488                 log(LOG_DEBUG, "freed %d flow entries\n", count);
1489 }
1490
1491 void
1492 flowtable_route_flush(struct flowtable *ft, struct rtentry *rt)
1493 {
1494         int i;
1495
1496         if (ft->ft_flags & FL_PCPU) {
1497                 for (i = 0; i <= mp_maxid; i++) {
1498                         if (CPU_ABSENT(i))
1499                                 continue;
1500
1501                         if (smp_started == 1) {
1502                                 thread_lock(curthread);
1503                                 sched_bind(curthread, i);
1504                                 thread_unlock(curthread);
1505                         }
1506
1507                         flowtable_free_stale(ft, rt);
1508
1509                         if (smp_started == 1) {
1510                                 thread_lock(curthread);
1511                                 sched_unbind(curthread);
1512                                 thread_unlock(curthread);
1513                         }
1514                 }
1515         } else {
1516                 flowtable_free_stale(ft, rt);
1517         }
1518 }
1519
1520 static void
1521 flowtable_clean_vnet(void)
1522 {
1523         struct flowtable *ft;
1524         int i;
1525
1526         ft = V_flow_list_head;
1527         while (ft != NULL) {
1528                 if (ft->ft_flags & FL_PCPU) {
1529                         for (i = 0; i <= mp_maxid; i++) {
1530                                 if (CPU_ABSENT(i))
1531                                         continue;
1532
1533                                 if (smp_started == 1) {
1534                                         thread_lock(curthread);
1535                                         sched_bind(curthread, i);
1536                                         thread_unlock(curthread);
1537                                 }
1538
1539                                 flowtable_free_stale(ft, NULL);
1540
1541                                 if (smp_started == 1) {
1542                                         thread_lock(curthread);
1543                                         sched_unbind(curthread);
1544                                         thread_unlock(curthread);
1545                                 }
1546                         }
1547                 } else {
1548                         flowtable_free_stale(ft, NULL);
1549                 }
1550                 ft = ft->ft_next;
1551         }
1552 }
1553
1554 static void
1555 flowtable_cleaner(void)
1556 {
1557         VNET_ITERATOR_DECL(vnet_iter);
1558
1559         if (bootverbose)
1560                 log(LOG_INFO, "flowtable cleaner started\n");
1561         while (1) {
1562                 VNET_LIST_RLOCK();
1563                 VNET_FOREACH(vnet_iter) {
1564                         CURVNET_SET(vnet_iter);
1565                         flowtable_clean_vnet();
1566                         CURVNET_RESTORE();
1567                 }
1568                 VNET_LIST_RUNLOCK();
1569
1570                 flowclean_cycles++;
1571                 /*
1572                  * The 10 second interval between cleaning checks
1573                  * is arbitrary
1574                  */
1575                 mtx_lock(&flowclean_lock);
1576                 cv_broadcast(&flowclean_cv);
1577                 cv_timedwait(&flowclean_cv, &flowclean_lock, flowclean_freq);
1578                 mtx_unlock(&flowclean_lock);
1579         }
1580 }
1581
1582 static void
1583 flowtable_flush(void *unused __unused)
1584 {
1585         uint64_t start;
1586
1587         mtx_lock(&flowclean_lock);
1588         start = flowclean_cycles;
1589         while (start == flowclean_cycles) {
1590                 cv_broadcast(&flowclean_cv);
1591                 cv_wait(&flowclean_cv, &flowclean_lock);
1592         }
1593         mtx_unlock(&flowclean_lock);
1594 }
1595
1596 static struct kproc_desc flow_kp = {
1597         "flowcleaner",
1598         flowtable_cleaner,
1599         &flowcleanerproc
1600 };
1601 SYSINIT(flowcleaner, SI_SUB_KTHREAD_IDLE, SI_ORDER_ANY, kproc_start, &flow_kp);
1602
1603 static void
1604 flowtable_init_vnet(const void *unused __unused)
1605 {
1606
1607         V_flowtable_nmbflows = 1024 + maxusers * 64 * mp_ncpus;
1608         V_flow_ipv4_zone = uma_zcreate("ip4flow", sizeof(struct flentry_v4),
1609             NULL, NULL, NULL, NULL, 64, UMA_ZONE_MAXBUCKET);
1610         V_flow_ipv6_zone = uma_zcreate("ip6flow", sizeof(struct flentry_v6),
1611             NULL, NULL, NULL, NULL, 64, UMA_ZONE_MAXBUCKET);
1612         uma_zone_set_max(V_flow_ipv4_zone, V_flowtable_nmbflows);
1613         uma_zone_set_max(V_flow_ipv6_zone, V_flowtable_nmbflows);
1614         V_flowtable_ready = 1;
1615 }
1616 VNET_SYSINIT(flowtable_init_vnet, SI_SUB_SMP, SI_ORDER_ANY,
1617     flowtable_init_vnet, NULL);
1618
1619 static void
1620 flowtable_init(const void *unused __unused)
1621 {
1622
1623         cv_init(&flowclean_cv, "flowcleanwait");
1624         mtx_init(&flowclean_lock, "flowclean lock", NULL, MTX_DEF);
1625         EVENTHANDLER_REGISTER(ifnet_departure_event, flowtable_flush, NULL,
1626             EVENTHANDLER_PRI_ANY);
1627         flowclean_freq = 20*hz;
1628 }
1629 SYSINIT(flowtable_init, SI_SUB_SMP, SI_ORDER_MIDDLE,
1630     flowtable_init, NULL);
1631
1632
1633 #ifdef VIMAGE
1634 static void
1635 flowtable_uninit(const void *unused __unused)
1636 {
1637
1638         V_flowtable_ready = 0;
1639         uma_zdestroy(V_flow_ipv4_zone);
1640         uma_zdestroy(V_flow_ipv6_zone);
1641 }
1642
1643 VNET_SYSUNINIT(flowtable_uninit, SI_SUB_KTHREAD_INIT, SI_ORDER_ANY,
1644     flowtable_uninit, NULL);
1645 #endif
1646
1647 #ifdef DDB
1648 static uint32_t *
1649 flowtable_get_hashkey(struct flentry *fle)
1650 {
1651         uint32_t *hashkey;
1652
1653         if (fle->f_flags & FL_IPV6)
1654                 hashkey = ((struct flentry_v4 *)fle)->fl_flow.ipf_key;
1655         else
1656                 hashkey = ((struct flentry_v6 *)fle)->fl_flow.ipf_key;
1657
1658         return (hashkey);
1659 }
1660
1661 static bitstr_t *
1662 flowtable_mask_pcpu(struct flowtable *ft, int cpuid)
1663 {
1664         bitstr_t *mask;
1665
1666         if (ft->ft_flags & FL_PCPU)
1667                 mask = ft->ft_masks[cpuid];
1668         else
1669                 mask = ft->ft_masks[0];
1670
1671         return (mask);
1672 }
1673
1674 static struct flentry **
1675 flowtable_entry_pcpu(struct flowtable *ft, uint32_t hash, int cpuid)
1676 {
1677         struct flentry **fle;
1678         int index = (hash % ft->ft_size);
1679
1680         if (ft->ft_flags & FL_PCPU) {
1681                 fle = &ft->ft_table.pcpu[cpuid][index];
1682         } else {
1683                 fle = &ft->ft_table.global[index];
1684         }
1685
1686         return (fle);
1687 }
1688
1689 static void
1690 flow_show(struct flowtable *ft, struct flentry *fle)
1691 {
1692         int idle_time;
1693         int rt_valid, ifp_valid;
1694         uint16_t sport, dport;
1695         uint32_t *hashkey;
1696         char saddr[4*sizeof "123"], daddr[4*sizeof "123"];
1697         volatile struct rtentry *rt;
1698         struct ifnet *ifp = NULL;
1699
1700         idle_time = (int)(time_uptime - fle->f_uptime);
1701         rt = fle->f_rt;
1702         rt_valid = rt != NULL;
1703         if (rt_valid)
1704                 ifp = rt->rt_ifp;
1705         ifp_valid = ifp != NULL;
1706         hashkey = flowtable_get_hashkey(fle);
1707         if (fle->f_flags & FL_IPV6)
1708                 goto skipaddr;
1709
1710         inet_ntoa_r(*(struct in_addr *) &hashkey[2], daddr);
1711         if (ft->ft_flags & FL_HASH_ALL) {
1712                 inet_ntoa_r(*(struct in_addr *) &hashkey[1], saddr);
1713                 sport = ntohs(((uint16_t *)hashkey)[0]);
1714                 dport = ntohs(((uint16_t *)hashkey)[1]);
1715                 db_printf("%s:%d->%s:%d",
1716                     saddr, sport, daddr,
1717                     dport);
1718         } else
1719                 db_printf("%s ", daddr);
1720
1721 skipaddr:
1722         if (fle->f_flags & FL_STALE)
1723                 db_printf(" FL_STALE ");
1724         if (fle->f_flags & FL_TCP)
1725                 db_printf(" FL_TCP ");
1726         if (fle->f_flags & FL_UDP)
1727                 db_printf(" FL_UDP ");
1728         if (rt_valid) {
1729                 if (rt->rt_flags & RTF_UP)
1730                         db_printf(" RTF_UP ");
1731         }
1732         if (ifp_valid) {
1733                 if (ifp->if_flags & IFF_LOOPBACK)
1734                         db_printf(" IFF_LOOPBACK ");
1735                 if (ifp->if_flags & IFF_UP)
1736                         db_printf(" IFF_UP ");
1737                 if (ifp->if_flags & IFF_POINTOPOINT)
1738                         db_printf(" IFF_POINTOPOINT ");
1739         }
1740         if (fle->f_flags & FL_IPV6)
1741                 db_printf("\n\tkey=%08x:%08x:%08x%08x:%08x:%08x%08x:%08x:%08x",
1742                     hashkey[0], hashkey[1], hashkey[2],
1743                     hashkey[3], hashkey[4], hashkey[5],
1744                     hashkey[6], hashkey[7], hashkey[8]);
1745         else
1746                 db_printf("\n\tkey=%08x:%08x:%08x ",
1747                     hashkey[0], hashkey[1], hashkey[2]);
1748         db_printf("hash=%08x idle_time=%03d"
1749             "\n\tfibnum=%02d rt=%p",
1750             fle->f_fhash, idle_time, fle->f_fibnum, fle->f_rt);
1751         db_printf("\n");
1752 }
1753
1754 static void
1755 flowtable_show(struct flowtable *ft, int cpuid)
1756 {
1757         int curbit = 0;
1758         struct flentry *fle,  **flehead;
1759         bitstr_t *mask, *tmpmask;
1760
1761         if (cpuid != -1)
1762                 db_printf("cpu: %d\n", cpuid);
1763         mask = flowtable_mask_pcpu(ft, cpuid);
1764         tmpmask = ft->ft_tmpmask;
1765         memcpy(tmpmask, mask, ft->ft_size/8);
1766         /*
1767          * XXX Note to self, bit_ffs operates at the byte level
1768          * and thus adds gratuitous overhead
1769          */
1770         bit_ffs(tmpmask, ft->ft_size, &curbit);
1771         while (curbit != -1) {
1772                 if (curbit >= ft->ft_size || curbit < -1) {
1773                         db_printf("warning: bad curbit value %d \n",
1774                             curbit);
1775                         break;
1776                 }
1777
1778                 flehead = flowtable_entry_pcpu(ft, curbit, cpuid);
1779                 fle = *flehead;
1780
1781                 while (fle != NULL) {
1782                         flow_show(ft, fle);
1783                         fle = fle->f_next;
1784                         continue;
1785                 }
1786                 bit_clear(tmpmask, curbit);
1787                 bit_ffs(tmpmask, ft->ft_size, &curbit);
1788         }
1789 }
1790
1791 static void
1792 flowtable_show_vnet(void)
1793 {
1794         struct flowtable *ft;
1795         int i;
1796
1797         ft = V_flow_list_head;
1798         while (ft != NULL) {
1799                 printf("name: %s\n", ft->ft_name);
1800                 if (ft->ft_flags & FL_PCPU) {
1801                         for (i = 0; i <= mp_maxid; i++) {
1802                                 if (CPU_ABSENT(i))
1803                                         continue;
1804                                 flowtable_show(ft, i);
1805                         }
1806                 } else {
1807                         flowtable_show(ft, -1);
1808                 }
1809                 ft = ft->ft_next;
1810         }
1811 }
1812
1813 DB_SHOW_COMMAND(flowtables, db_show_flowtables)
1814 {
1815         VNET_ITERATOR_DECL(vnet_iter);
1816
1817         VNET_FOREACH(vnet_iter) {
1818                 CURVNET_SET(vnet_iter);
1819                 flowtable_show_vnet();
1820                 CURVNET_RESTORE();
1821         }
1822 }
1823 #endif