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