The Design and Implementation of the FreeBSD Operating System, Second Edition
Now available: The Design and Implementation of the FreeBSD Operating System (Second Edition)


[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]

FreeBSD/Linux Kernel Cross Reference
sys/netinet/tcp_syncache.c

Version: -  FREEBSD  -  FREEBSD-12-STABLE  -  FREEBSD-12-0  -  FREEBSD-11-STABLE  -  FREEBSD-11-2  -  FREEBSD-11-1  -  FREEBSD-11-0  -  FREEBSD-10-STABLE  -  FREEBSD-10-4  -  FREEBSD-10-3  -  FREEBSD-10-2  -  FREEBSD-10-1  -  FREEBSD-10-0  -  FREEBSD-9-STABLE  -  FREEBSD-9-3  -  FREEBSD-9-2  -  FREEBSD-9-1  -  FREEBSD-9-0  -  FREEBSD-8-STABLE  -  FREEBSD-8-4  -  FREEBSD-8-3  -  FREEBSD-8-2  -  FREEBSD-8-1  -  FREEBSD-8-0  -  FREEBSD-7-STABLE  -  FREEBSD-7-4  -  FREEBSD-7-3  -  FREEBSD-7-2  -  FREEBSD-7-1  -  FREEBSD-7-0  -  FREEBSD-6-STABLE  -  FREEBSD-6-4  -  FREEBSD-6-3  -  FREEBSD-6-2  -  FREEBSD-6-1  -  FREEBSD-6-0  -  FREEBSD-5-STABLE  -  FREEBSD-5-5  -  FREEBSD-5-4  -  FREEBSD-5-3  -  FREEBSD-5-2  -  FREEBSD-5-1  -  FREEBSD-5-0  -  FREEBSD-4-STABLE  -  FREEBSD-3-STABLE  -  FREEBSD22  -  linux-2.6  -  linux-2.4.22  -  MK83  -  MK84  -  PLAN9  -  DFBSD  -  NETBSD  -  NETBSD5  -  NETBSD4  -  NETBSD3  -  NETBSD20  -  OPENBSD  -  xnu-517  -  xnu-792  -  xnu-792.6.70  -  xnu-1228  -  xnu-1456.1.26  -  xnu-1699.24.8  -  xnu-2050.18.24  -  OPENSOLARIS  -  minix-3-1-1 
SearchContext: -  none  -  3  -  10 

    1 /*-
    2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
    3  *
    4  * Copyright (c) 2001 McAfee, Inc.
    5  * Copyright (c) 2006,2013 Andre Oppermann, Internet Business Solutions AG
    6  * All rights reserved.
    7  *
    8  * This software was developed for the FreeBSD Project by Jonathan Lemon
    9  * and McAfee Research, the Security Research Division of McAfee, Inc. under
   10  * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the
   11  * DARPA CHATS research program. [2001 McAfee, Inc.]
   12  *
   13  * Redistribution and use in source and binary forms, with or without
   14  * modification, are permitted provided that the following conditions
   15  * are met:
   16  * 1. Redistributions of source code must retain the above copyright
   17  *    notice, this list of conditions and the following disclaimer.
   18  * 2. Redistributions in binary form must reproduce the above copyright
   19  *    notice, this list of conditions and the following disclaimer in the
   20  *    documentation and/or other materials provided with the distribution.
   21  *
   22  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
   23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
   24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
   25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
   26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
   27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
   28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
   29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
   30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
   31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
   32  * SUCH DAMAGE.
   33  */
   34 
   35 #include <sys/cdefs.h>
   36 __FBSDID("$FreeBSD: stable/12/sys/netinet/tcp_syncache.c 341504 2018-12-04 22:46:58Z tuexen $");
   37 
   38 #include "opt_inet.h"
   39 #include "opt_inet6.h"
   40 #include "opt_ipsec.h"
   41 #include "opt_pcbgroup.h"
   42 
   43 #include <sys/param.h>
   44 #include <sys/systm.h>
   45 #include <sys/hash.h>
   46 #include <sys/refcount.h>
   47 #include <sys/kernel.h>
   48 #include <sys/sysctl.h>
   49 #include <sys/limits.h>
   50 #include <sys/lock.h>
   51 #include <sys/mutex.h>
   52 #include <sys/malloc.h>
   53 #include <sys/mbuf.h>
   54 #include <sys/proc.h>           /* for proc0 declaration */
   55 #include <sys/random.h>
   56 #include <sys/socket.h>
   57 #include <sys/socketvar.h>
   58 #include <sys/syslog.h>
   59 #include <sys/ucred.h>
   60 
   61 #include <sys/md5.h>
   62 #include <crypto/siphash/siphash.h>
   63 
   64 #include <vm/uma.h>
   65 
   66 #include <net/if.h>
   67 #include <net/if_var.h>
   68 #include <net/route.h>
   69 #include <net/vnet.h>
   70 
   71 #include <netinet/in.h>
   72 #include <netinet/in_kdtrace.h>
   73 #include <netinet/in_systm.h>
   74 #include <netinet/ip.h>
   75 #include <netinet/in_var.h>
   76 #include <netinet/in_pcb.h>
   77 #include <netinet/ip_var.h>
   78 #include <netinet/ip_options.h>
   79 #ifdef INET6
   80 #include <netinet/ip6.h>
   81 #include <netinet/icmp6.h>
   82 #include <netinet6/nd6.h>
   83 #include <netinet6/ip6_var.h>
   84 #include <netinet6/in6_pcb.h>
   85 #endif
   86 #include <netinet/tcp.h>
   87 #include <netinet/tcp_fastopen.h>
   88 #include <netinet/tcp_fsm.h>
   89 #include <netinet/tcp_seq.h>
   90 #include <netinet/tcp_timer.h>
   91 #include <netinet/tcp_var.h>
   92 #include <netinet/tcp_syncache.h>
   93 #ifdef INET6
   94 #include <netinet6/tcp6_var.h>
   95 #endif
   96 #ifdef TCP_OFFLOAD
   97 #include <netinet/toecore.h>
   98 #endif
   99 
  100 #include <netipsec/ipsec_support.h>
  101 
  102 #include <machine/in_cksum.h>
  103 
  104 #include <security/mac/mac_framework.h>
  105 
  106 VNET_DEFINE_STATIC(int, tcp_syncookies) = 1;
  107 #define V_tcp_syncookies                VNET(tcp_syncookies)
  108 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_VNET | CTLFLAG_RW,
  109     &VNET_NAME(tcp_syncookies), 0,
  110     "Use TCP SYN cookies if the syncache overflows");
  111 
  112 VNET_DEFINE_STATIC(int, tcp_syncookiesonly) = 0;
  113 #define V_tcp_syncookiesonly            VNET(tcp_syncookiesonly)
  114 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_VNET | CTLFLAG_RW,
  115     &VNET_NAME(tcp_syncookiesonly), 0,
  116     "Use only TCP SYN cookies");
  117 
  118 VNET_DEFINE_STATIC(int, functions_inherit_listen_socket_stack) = 1;
  119 #define V_functions_inherit_listen_socket_stack \
  120     VNET(functions_inherit_listen_socket_stack)
  121 SYSCTL_INT(_net_inet_tcp, OID_AUTO, functions_inherit_listen_socket_stack,
  122     CTLFLAG_VNET | CTLFLAG_RW,
  123     &VNET_NAME(functions_inherit_listen_socket_stack), 0,
  124     "Inherit listen socket's stack");
  125 
  126 #ifdef TCP_OFFLOAD
  127 #define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL)
  128 #endif
  129 
  130 static void      syncache_drop(struct syncache *, struct syncache_head *);
  131 static void      syncache_free(struct syncache *);
  132 static void      syncache_insert(struct syncache *, struct syncache_head *);
  133 static int       syncache_respond(struct syncache *, struct syncache_head *,
  134                     const struct mbuf *, int);
  135 static struct    socket *syncache_socket(struct syncache *, struct socket *,
  136                     struct mbuf *m);
  137 static void      syncache_timeout(struct syncache *sc, struct syncache_head *sch,
  138                     int docallout);
  139 static void      syncache_timer(void *);
  140 
  141 static uint32_t  syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t,
  142                     uint8_t *, uintptr_t);
  143 static tcp_seq   syncookie_generate(struct syncache_head *, struct syncache *);
  144 static struct syncache
  145                 *syncookie_lookup(struct in_conninfo *, struct syncache_head *,
  146                     struct syncache *, struct tcphdr *, struct tcpopt *,
  147                     struct socket *);
  148 static void      syncookie_reseed(void *);
  149 #ifdef INVARIANTS
  150 static int       syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
  151                     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
  152                     struct socket *lso);
  153 #endif
  154 
  155 /*
  156  * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies.
  157  * 3 retransmits corresponds to a timeout of 3 * (1 + 2 + 4 + 8) == 45 seconds,
  158  * the odds are that the user has given up attempting to connect by then.
  159  */
  160 #define SYNCACHE_MAXREXMTS              3
  161 
  162 /* Arbitrary values */
  163 #define TCP_SYNCACHE_HASHSIZE           512
  164 #define TCP_SYNCACHE_BUCKETLIMIT        30
  165 
  166 VNET_DEFINE_STATIC(struct tcp_syncache, tcp_syncache);
  167 #define V_tcp_syncache                  VNET(tcp_syncache)
  168 
  169 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, CTLFLAG_RW, 0,
  170     "TCP SYN cache");
  171 
  172 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
  173     &VNET_NAME(tcp_syncache.bucket_limit), 0,
  174     "Per-bucket hash limit for syncache");
  175 
  176 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
  177     &VNET_NAME(tcp_syncache.cache_limit), 0,
  178     "Overall entry limit for syncache");
  179 
  180 SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET,
  181     &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache");
  182 
  183 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN,
  184     &VNET_NAME(tcp_syncache.hashsize), 0,
  185     "Size of TCP syncache hashtable");
  186 
  187 static int
  188 sysctl_net_inet_tcp_syncache_rexmtlimit_check(SYSCTL_HANDLER_ARGS)
  189 {
  190         int error;
  191         u_int new;
  192 
  193         new = V_tcp_syncache.rexmt_limit;
  194         error = sysctl_handle_int(oidp, &new, 0, req);
  195         if ((error == 0) && (req->newptr != NULL)) {
  196                 if (new > TCP_MAXRXTSHIFT)
  197                         error = EINVAL;
  198                 else
  199                         V_tcp_syncache.rexmt_limit = new;
  200         }
  201         return (error);
  202 }
  203 
  204 SYSCTL_PROC(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit,
  205     CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW,
  206     &VNET_NAME(tcp_syncache.rexmt_limit), 0,
  207     sysctl_net_inet_tcp_syncache_rexmtlimit_check, "UI",
  208     "Limit on SYN/ACK retransmissions");
  209 
  210 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1;
  211 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail,
  212     CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0,
  213     "Send reset on socket allocation failure");
  214 
  215 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
  216 
  217 #define SCH_LOCK(sch)           mtx_lock(&(sch)->sch_mtx)
  218 #define SCH_UNLOCK(sch)         mtx_unlock(&(sch)->sch_mtx)
  219 #define SCH_LOCK_ASSERT(sch)    mtx_assert(&(sch)->sch_mtx, MA_OWNED)
  220 
  221 /*
  222  * Requires the syncache entry to be already removed from the bucket list.
  223  */
  224 static void
  225 syncache_free(struct syncache *sc)
  226 {
  227 
  228         if (sc->sc_ipopts)
  229                 (void) m_free(sc->sc_ipopts);
  230         if (sc->sc_cred)
  231                 crfree(sc->sc_cred);
  232 #ifdef MAC
  233         mac_syncache_destroy(&sc->sc_label);
  234 #endif
  235 
  236         uma_zfree(V_tcp_syncache.zone, sc);
  237 }
  238 
  239 void
  240 syncache_init(void)
  241 {
  242         int i;
  243 
  244         V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
  245         V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
  246         V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
  247         V_tcp_syncache.hash_secret = arc4random();
  248 
  249         TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
  250             &V_tcp_syncache.hashsize);
  251         TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
  252             &V_tcp_syncache.bucket_limit);
  253         if (!powerof2(V_tcp_syncache.hashsize) ||
  254             V_tcp_syncache.hashsize == 0) {
  255                 printf("WARNING: syncache hash size is not a power of 2.\n");
  256                 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
  257         }
  258         V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1;
  259 
  260         /* Set limits. */
  261         V_tcp_syncache.cache_limit =
  262             V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit;
  263         TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
  264             &V_tcp_syncache.cache_limit);
  265 
  266         /* Allocate the hash table. */
  267         V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize *
  268             sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO);
  269 
  270 #ifdef VIMAGE
  271         V_tcp_syncache.vnet = curvnet;
  272 #endif
  273 
  274         /* Initialize the hash buckets. */
  275         for (i = 0; i < V_tcp_syncache.hashsize; i++) {
  276                 TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket);
  277                 mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head",
  278                          NULL, MTX_DEF);
  279                 callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer,
  280                          &V_tcp_syncache.hashbase[i].sch_mtx, 0);
  281                 V_tcp_syncache.hashbase[i].sch_length = 0;
  282                 V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache;
  283                 V_tcp_syncache.hashbase[i].sch_last_overflow =
  284                     -(SYNCOOKIE_LIFETIME + 1);
  285         }
  286 
  287         /* Create the syncache entry zone. */
  288         V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache),
  289             NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
  290         V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone,
  291             V_tcp_syncache.cache_limit);
  292 
  293         /* Start the SYN cookie reseeder callout. */
  294         callout_init(&V_tcp_syncache.secret.reseed, 1);
  295         arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0);
  296         arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0);
  297         callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz,
  298             syncookie_reseed, &V_tcp_syncache);
  299 }
  300 
  301 #ifdef VIMAGE
  302 void
  303 syncache_destroy(void)
  304 {
  305         struct syncache_head *sch;
  306         struct syncache *sc, *nsc;
  307         int i;
  308 
  309         /*
  310          * Stop the re-seed timer before freeing resources.  No need to
  311          * possibly schedule it another time.
  312          */
  313         callout_drain(&V_tcp_syncache.secret.reseed);
  314 
  315         /* Cleanup hash buckets: stop timers, free entries, destroy locks. */
  316         for (i = 0; i < V_tcp_syncache.hashsize; i++) {
  317 
  318                 sch = &V_tcp_syncache.hashbase[i];
  319                 callout_drain(&sch->sch_timer);
  320 
  321                 SCH_LOCK(sch);
  322                 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc)
  323                         syncache_drop(sc, sch);
  324                 SCH_UNLOCK(sch);
  325                 KASSERT(TAILQ_EMPTY(&sch->sch_bucket),
  326                     ("%s: sch->sch_bucket not empty", __func__));
  327                 KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0",
  328                     __func__, sch->sch_length));
  329                 mtx_destroy(&sch->sch_mtx);
  330         }
  331 
  332         KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0,
  333             ("%s: cache_count not 0", __func__));
  334 
  335         /* Free the allocated global resources. */
  336         uma_zdestroy(V_tcp_syncache.zone);
  337         free(V_tcp_syncache.hashbase, M_SYNCACHE);
  338 }
  339 #endif
  340 
  341 /*
  342  * Inserts a syncache entry into the specified bucket row.
  343  * Locks and unlocks the syncache_head autonomously.
  344  */
  345 static void
  346 syncache_insert(struct syncache *sc, struct syncache_head *sch)
  347 {
  348         struct syncache *sc2;
  349 
  350         SCH_LOCK(sch);
  351 
  352         /*
  353          * Make sure that we don't overflow the per-bucket limit.
  354          * If the bucket is full, toss the oldest element.
  355          */
  356         if (sch->sch_length >= V_tcp_syncache.bucket_limit) {
  357                 KASSERT(!TAILQ_EMPTY(&sch->sch_bucket),
  358                         ("sch->sch_length incorrect"));
  359                 sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head);
  360                 sch->sch_last_overflow = time_uptime;
  361                 syncache_drop(sc2, sch);
  362                 TCPSTAT_INC(tcps_sc_bucketoverflow);
  363         }
  364 
  365         /* Put it into the bucket. */
  366         TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash);
  367         sch->sch_length++;
  368 
  369 #ifdef TCP_OFFLOAD
  370         if (ADDED_BY_TOE(sc)) {
  371                 struct toedev *tod = sc->sc_tod;
  372 
  373                 tod->tod_syncache_added(tod, sc->sc_todctx);
  374         }
  375 #endif
  376 
  377         /* Reinitialize the bucket row's timer. */
  378         if (sch->sch_length == 1)
  379                 sch->sch_nextc = ticks + INT_MAX;
  380         syncache_timeout(sc, sch, 1);
  381 
  382         SCH_UNLOCK(sch);
  383 
  384         TCPSTATES_INC(TCPS_SYN_RECEIVED);
  385         TCPSTAT_INC(tcps_sc_added);
  386 }
  387 
  388 /*
  389  * Remove and free entry from syncache bucket row.
  390  * Expects locked syncache head.
  391  */
  392 static void
  393 syncache_drop(struct syncache *sc, struct syncache_head *sch)
  394 {
  395 
  396         SCH_LOCK_ASSERT(sch);
  397 
  398         TCPSTATES_DEC(TCPS_SYN_RECEIVED);
  399         TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
  400         sch->sch_length--;
  401 
  402 #ifdef TCP_OFFLOAD
  403         if (ADDED_BY_TOE(sc)) {
  404                 struct toedev *tod = sc->sc_tod;
  405 
  406                 tod->tod_syncache_removed(tod, sc->sc_todctx);
  407         }
  408 #endif
  409 
  410         syncache_free(sc);
  411 }
  412 
  413 /*
  414  * Engage/reengage time on bucket row.
  415  */
  416 static void
  417 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout)
  418 {
  419         int rexmt;
  420 
  421         if (sc->sc_rxmits == 0)
  422                 rexmt = TCPTV_RTOBASE;
  423         else
  424                 TCPT_RANGESET(rexmt, TCPTV_RTOBASE * tcp_syn_backoff[sc->sc_rxmits],
  425                     tcp_rexmit_min, TCPTV_REXMTMAX);
  426         sc->sc_rxttime = ticks + rexmt;
  427         sc->sc_rxmits++;
  428         if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) {
  429                 sch->sch_nextc = sc->sc_rxttime;
  430                 if (docallout)
  431                         callout_reset(&sch->sch_timer, sch->sch_nextc - ticks,
  432                             syncache_timer, (void *)sch);
  433         }
  434 }
  435 
  436 /*
  437  * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
  438  * If we have retransmitted an entry the maximum number of times, expire it.
  439  * One separate timer for each bucket row.
  440  */
  441 static void
  442 syncache_timer(void *xsch)
  443 {
  444         struct syncache_head *sch = (struct syncache_head *)xsch;
  445         struct syncache *sc, *nsc;
  446         int tick = ticks;
  447         char *s;
  448 
  449         CURVNET_SET(sch->sch_sc->vnet);
  450 
  451         /* NB: syncache_head has already been locked by the callout. */
  452         SCH_LOCK_ASSERT(sch);
  453 
  454         /*
  455          * In the following cycle we may remove some entries and/or
  456          * advance some timeouts, so re-initialize the bucket timer.
  457          */
  458         sch->sch_nextc = tick + INT_MAX;
  459 
  460         TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) {
  461                 /*
  462                  * We do not check if the listen socket still exists
  463                  * and accept the case where the listen socket may be
  464                  * gone by the time we resend the SYN/ACK.  We do
  465                  * not expect this to happens often. If it does,
  466                  * then the RST will be sent by the time the remote
  467                  * host does the SYN/ACK->ACK.
  468                  */
  469                 if (TSTMP_GT(sc->sc_rxttime, tick)) {
  470                         if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc))
  471                                 sch->sch_nextc = sc->sc_rxttime;
  472                         continue;
  473                 }
  474                 if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) {
  475                         if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
  476                                 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, "
  477                                     "giving up and removing syncache entry\n",
  478                                     s, __func__);
  479                                 free(s, M_TCPLOG);
  480                         }
  481                         syncache_drop(sc, sch);
  482                         TCPSTAT_INC(tcps_sc_stale);
  483                         continue;
  484                 }
  485                 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
  486                         log(LOG_DEBUG, "%s; %s: Response timeout, "
  487                             "retransmitting (%u) SYN|ACK\n",
  488                             s, __func__, sc->sc_rxmits);
  489                         free(s, M_TCPLOG);
  490                 }
  491 
  492                 syncache_respond(sc, sch, NULL, TH_SYN|TH_ACK);
  493                 TCPSTAT_INC(tcps_sc_retransmitted);
  494                 syncache_timeout(sc, sch, 0);
  495         }
  496         if (!TAILQ_EMPTY(&(sch)->sch_bucket))
  497                 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick,
  498                         syncache_timer, (void *)(sch));
  499         CURVNET_RESTORE();
  500 }
  501 
  502 /*
  503  * Find an entry in the syncache.
  504  * Returns always with locked syncache_head plus a matching entry or NULL.
  505  */
  506 static struct syncache *
  507 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
  508 {
  509         struct syncache *sc;
  510         struct syncache_head *sch;
  511         uint32_t hash;
  512 
  513         /*
  514          * The hash is built on foreign port + local port + foreign address.
  515          * We rely on the fact that struct in_conninfo starts with 16 bits
  516          * of foreign port, then 16 bits of local port then followed by 128
  517          * bits of foreign address.  In case of IPv4 address, the first 3
  518          * 32-bit words of the address always are zeroes.
  519          */
  520         hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5,
  521             V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask;
  522 
  523         sch = &V_tcp_syncache.hashbase[hash];
  524         *schp = sch;
  525         SCH_LOCK(sch);
  526 
  527         /* Circle through bucket row to find matching entry. */
  528         TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash)
  529                 if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie,
  530                     sizeof(struct in_endpoints)) == 0)
  531                         break;
  532 
  533         return (sc);    /* Always returns with locked sch. */
  534 }
  535 
  536 /*
  537  * This function is called when we get a RST for a
  538  * non-existent connection, so that we can see if the
  539  * connection is in the syn cache.  If it is, zap it.
  540  * If required send a challenge ACK.
  541  */
  542 void
  543 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th, struct mbuf *m)
  544 {
  545         struct syncache *sc;
  546         struct syncache_head *sch;
  547         char *s = NULL;
  548 
  549         sc = syncache_lookup(inc, &sch);        /* returns locked sch */
  550         SCH_LOCK_ASSERT(sch);
  551 
  552         /*
  553          * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags.
  554          * See RFC 793 page 65, section SEGMENT ARRIVES.
  555          */
  556         if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) {
  557                 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
  558                         log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or "
  559                             "FIN flag set, segment ignored\n", s, __func__);
  560                 TCPSTAT_INC(tcps_badrst);
  561                 goto done;
  562         }
  563 
  564         /*
  565          * No corresponding connection was found in syncache.
  566          * If syncookies are enabled and possibly exclusively
  567          * used, or we are under memory pressure, a valid RST
  568          * may not find a syncache entry.  In that case we're
  569          * done and no SYN|ACK retransmissions will happen.
  570          * Otherwise the RST was misdirected or spoofed.
  571          */
  572         if (sc == NULL) {
  573                 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
  574                         log(LOG_DEBUG, "%s; %s: Spurious RST without matching "
  575                             "syncache entry (possibly syncookie only), "
  576                             "segment ignored\n", s, __func__);
  577                 TCPSTAT_INC(tcps_badrst);
  578                 goto done;
  579         }
  580 
  581         /*
  582          * If the RST bit is set, check the sequence number to see
  583          * if this is a valid reset segment.
  584          *
  585          * RFC 793 page 37:
  586          *   In all states except SYN-SENT, all reset (RST) segments
  587          *   are validated by checking their SEQ-fields.  A reset is
  588          *   valid if its sequence number is in the window.
  589          *
  590          * RFC 793 page 69:
  591          *   There are four cases for the acceptability test for an incoming
  592          *   segment:
  593          *
  594          * Segment Receive  Test
  595          * Length  Window
  596          * ------- -------  -------------------------------------------
  597          *    0       0     SEG.SEQ = RCV.NXT
  598          *    0      >0     RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
  599          *   >0       0     not acceptable
  600          *   >0      >0     RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
  601          *               or RCV.NXT =< SEG.SEQ+SEG.LEN-1 < RCV.NXT+RCV.WND
  602          *
  603          * Note that when receiving a SYN segment in the LISTEN state,
  604          * IRS is set to SEG.SEQ and RCV.NXT is set to SEG.SEQ+1, as
  605          * described in RFC 793, page 66.
  606          */
  607         if ((SEQ_GEQ(th->th_seq, sc->sc_irs + 1) &&
  608             SEQ_LT(th->th_seq, sc->sc_irs + 1 + sc->sc_wnd)) ||
  609             (sc->sc_wnd == 0 && th->th_seq == sc->sc_irs + 1)) {
  610                 if (V_tcp_insecure_rst ||
  611                     th->th_seq == sc->sc_irs + 1) {
  612                         syncache_drop(sc, sch);
  613                         if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
  614                                 log(LOG_DEBUG,
  615                                     "%s; %s: Our SYN|ACK was rejected, "
  616                                     "connection attempt aborted by remote "
  617                                     "endpoint\n",
  618                                     s, __func__);
  619                         TCPSTAT_INC(tcps_sc_reset);
  620                 } else {
  621                         TCPSTAT_INC(tcps_badrst);
  622                         /* Send challenge ACK. */
  623                         if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
  624                                 log(LOG_DEBUG, "%s; %s: RST with invalid "
  625                                     " SEQ %u != NXT %u (+WND %u), "
  626                                     "sending challenge ACK\n",
  627                                     s, __func__,
  628                                     th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
  629                         syncache_respond(sc, sch, m, TH_ACK);
  630                 }
  631         } else {
  632                 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
  633                         log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != "
  634                             "NXT %u (+WND %u), segment ignored\n",
  635                             s, __func__,
  636                             th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
  637                 TCPSTAT_INC(tcps_badrst);
  638         }
  639 
  640 done:
  641         if (s != NULL)
  642                 free(s, M_TCPLOG);
  643         SCH_UNLOCK(sch);
  644 }
  645 
  646 void
  647 syncache_badack(struct in_conninfo *inc)
  648 {
  649         struct syncache *sc;
  650         struct syncache_head *sch;
  651 
  652         sc = syncache_lookup(inc, &sch);        /* returns locked sch */
  653         SCH_LOCK_ASSERT(sch);
  654         if (sc != NULL) {
  655                 syncache_drop(sc, sch);
  656                 TCPSTAT_INC(tcps_sc_badack);
  657         }
  658         SCH_UNLOCK(sch);
  659 }
  660 
  661 void
  662 syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq)
  663 {
  664         struct syncache *sc;
  665         struct syncache_head *sch;
  666 
  667         sc = syncache_lookup(inc, &sch);        /* returns locked sch */
  668         SCH_LOCK_ASSERT(sch);
  669         if (sc == NULL)
  670                 goto done;
  671 
  672         /* If the sequence number != sc_iss, then it's a bogus ICMP msg */
  673         if (ntohl(th_seq) != sc->sc_iss)
  674                 goto done;
  675 
  676         /*
  677          * If we've rertransmitted 3 times and this is our second error,
  678          * we remove the entry.  Otherwise, we allow it to continue on.
  679          * This prevents us from incorrectly nuking an entry during a
  680          * spurious network outage.
  681          *
  682          * See tcp_notify().
  683          */
  684         if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) {
  685                 sc->sc_flags |= SCF_UNREACH;
  686                 goto done;
  687         }
  688         syncache_drop(sc, sch);
  689         TCPSTAT_INC(tcps_sc_unreach);
  690 done:
  691         SCH_UNLOCK(sch);
  692 }
  693 
  694 /*
  695  * Build a new TCP socket structure from a syncache entry.
  696  *
  697  * On success return the newly created socket with its underlying inp locked.
  698  */
  699 static struct socket *
  700 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
  701 {
  702         struct tcp_function_block *blk;
  703         struct inpcb *inp = NULL;
  704         struct socket *so;
  705         struct tcpcb *tp;
  706         int error;
  707         char *s;
  708 
  709         INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
  710 
  711         /*
  712          * Ok, create the full blown connection, and set things up
  713          * as they would have been set up if we had created the
  714          * connection when the SYN arrived.  If we can't create
  715          * the connection, abort it.
  716          */
  717         so = sonewconn(lso, 0);
  718         if (so == NULL) {
  719                 /*
  720                  * Drop the connection; we will either send a RST or
  721                  * have the peer retransmit its SYN again after its
  722                  * RTO and try again.
  723                  */
  724                 TCPSTAT_INC(tcps_listendrop);
  725                 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
  726                         log(LOG_DEBUG, "%s; %s: Socket create failed "
  727                             "due to limits or memory shortage\n",
  728                             s, __func__);
  729                         free(s, M_TCPLOG);
  730                 }
  731                 goto abort2;
  732         }
  733 #ifdef MAC
  734         mac_socketpeer_set_from_mbuf(m, so);
  735 #endif
  736 
  737         inp = sotoinpcb(so);
  738         inp->inp_inc.inc_fibnum = so->so_fibnum;
  739         INP_WLOCK(inp);
  740         /*
  741          * Exclusive pcbinfo lock is not required in syncache socket case even
  742          * if two inpcb locks can be acquired simultaneously:
  743          *  - the inpcb in LISTEN state,
  744          *  - the newly created inp.
  745          *
  746          * In this case, an inp cannot be at same time in LISTEN state and
  747          * just created by an accept() call.
  748          */
  749         INP_HASH_WLOCK(&V_tcbinfo);
  750 
  751         /* Insert new socket into PCB hash list. */
  752         inp->inp_inc.inc_flags = sc->sc_inc.inc_flags;
  753 #ifdef INET6
  754         if (sc->sc_inc.inc_flags & INC_ISIPV6) {
  755                 inp->inp_vflag &= ~INP_IPV4;
  756                 inp->inp_vflag |= INP_IPV6;
  757                 inp->in6p_laddr = sc->sc_inc.inc6_laddr;
  758         } else {
  759                 inp->inp_vflag &= ~INP_IPV6;
  760                 inp->inp_vflag |= INP_IPV4;
  761 #endif
  762                 inp->inp_laddr = sc->sc_inc.inc_laddr;
  763 #ifdef INET6
  764         }
  765 #endif
  766 
  767         /*
  768          * If there's an mbuf and it has a flowid, then let's initialise the
  769          * inp with that particular flowid.
  770          */
  771         if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
  772                 inp->inp_flowid = m->m_pkthdr.flowid;
  773                 inp->inp_flowtype = M_HASHTYPE_GET(m);
  774         }
  775 
  776         /*
  777          * Install in the reservation hash table for now, but don't yet
  778          * install a connection group since the full 4-tuple isn't yet
  779          * configured.
  780          */
  781         inp->inp_lport = sc->sc_inc.inc_lport;
  782         if ((error = in_pcbinshash_nopcbgroup(inp)) != 0) {
  783                 /*
  784                  * Undo the assignments above if we failed to
  785                  * put the PCB on the hash lists.
  786                  */
  787 #ifdef INET6
  788                 if (sc->sc_inc.inc_flags & INC_ISIPV6)
  789                         inp->in6p_laddr = in6addr_any;
  790                 else
  791 #endif
  792                         inp->inp_laddr.s_addr = INADDR_ANY;
  793                 inp->inp_lport = 0;
  794                 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
  795                         log(LOG_DEBUG, "%s; %s: in_pcbinshash failed "
  796                             "with error %i\n",
  797                             s, __func__, error);
  798                         free(s, M_TCPLOG);
  799                 }
  800                 INP_HASH_WUNLOCK(&V_tcbinfo);
  801                 goto abort;
  802         }
  803 #ifdef INET6
  804         if (inp->inp_vflag & INP_IPV6PROTO) {
  805                 struct inpcb *oinp = sotoinpcb(lso);
  806 
  807                 /*
  808                  * Inherit socket options from the listening socket.
  809                  * Note that in6p_inputopts are not (and should not be)
  810                  * copied, since it stores previously received options and is
  811                  * used to detect if each new option is different than the
  812                  * previous one and hence should be passed to a user.
  813                  * If we copied in6p_inputopts, a user would not be able to
  814                  * receive options just after calling the accept system call.
  815                  */
  816                 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
  817                 if (oinp->in6p_outputopts)
  818                         inp->in6p_outputopts =
  819                             ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
  820         }
  821 
  822         if (sc->sc_inc.inc_flags & INC_ISIPV6) {
  823                 struct in6_addr laddr6;
  824                 struct sockaddr_in6 sin6;
  825 
  826                 sin6.sin6_family = AF_INET6;
  827                 sin6.sin6_len = sizeof(sin6);
  828                 sin6.sin6_addr = sc->sc_inc.inc6_faddr;
  829                 sin6.sin6_port = sc->sc_inc.inc_fport;
  830                 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
  831                 laddr6 = inp->in6p_laddr;
  832                 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
  833                         inp->in6p_laddr = sc->sc_inc.inc6_laddr;
  834                 if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6,
  835                     thread0.td_ucred, m)) != 0) {
  836                         inp->in6p_laddr = laddr6;
  837                         if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
  838                                 log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed "
  839                                     "with error %i\n",
  840                                     s, __func__, error);
  841                                 free(s, M_TCPLOG);
  842                         }
  843                         INP_HASH_WUNLOCK(&V_tcbinfo);
  844                         goto abort;
  845                 }
  846                 /* Override flowlabel from in6_pcbconnect. */
  847                 inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
  848                 inp->inp_flow |= sc->sc_flowlabel;
  849         }
  850 #endif /* INET6 */
  851 #if defined(INET) && defined(INET6)
  852         else
  853 #endif
  854 #ifdef INET
  855         {
  856                 struct in_addr laddr;
  857                 struct sockaddr_in sin;
  858 
  859                 inp->inp_options = (m) ? ip_srcroute(m) : NULL;
  860                 
  861                 if (inp->inp_options == NULL) {
  862                         inp->inp_options = sc->sc_ipopts;
  863                         sc->sc_ipopts = NULL;
  864                 }
  865 
  866                 sin.sin_family = AF_INET;
  867                 sin.sin_len = sizeof(sin);
  868                 sin.sin_addr = sc->sc_inc.inc_faddr;
  869                 sin.sin_port = sc->sc_inc.inc_fport;
  870                 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero));
  871                 laddr = inp->inp_laddr;
  872                 if (inp->inp_laddr.s_addr == INADDR_ANY)
  873                         inp->inp_laddr = sc->sc_inc.inc_laddr;
  874                 if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin,
  875                     thread0.td_ucred, m)) != 0) {
  876                         inp->inp_laddr = laddr;
  877                         if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
  878                                 log(LOG_DEBUG, "%s; %s: in_pcbconnect failed "
  879                                     "with error %i\n",
  880                                     s, __func__, error);
  881                                 free(s, M_TCPLOG);
  882                         }
  883                         INP_HASH_WUNLOCK(&V_tcbinfo);
  884                         goto abort;
  885                 }
  886         }
  887 #endif /* INET */
  888 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
  889         /* Copy old policy into new socket's. */
  890         if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0)
  891                 printf("syncache_socket: could not copy policy\n");
  892 #endif
  893         INP_HASH_WUNLOCK(&V_tcbinfo);
  894         tp = intotcpcb(inp);
  895         tcp_state_change(tp, TCPS_SYN_RECEIVED);
  896         tp->iss = sc->sc_iss;
  897         tp->irs = sc->sc_irs;
  898         tcp_rcvseqinit(tp);
  899         tcp_sendseqinit(tp);
  900         blk = sototcpcb(lso)->t_fb;
  901         if (V_functions_inherit_listen_socket_stack && blk != tp->t_fb) {
  902                 /*
  903                  * Our parents t_fb was not the default,
  904                  * we need to release our ref on tp->t_fb and 
  905                  * pickup one on the new entry.
  906                  */
  907                 struct tcp_function_block *rblk;
  908                 
  909                 rblk = find_and_ref_tcp_fb(blk);
  910                 KASSERT(rblk != NULL,
  911                     ("cannot find blk %p out of syncache?", blk));
  912                 if (tp->t_fb->tfb_tcp_fb_fini)
  913                         (*tp->t_fb->tfb_tcp_fb_fini)(tp, 0);
  914                 refcount_release(&tp->t_fb->tfb_refcnt);
  915                 tp->t_fb = rblk;
  916                 /*
  917                  * XXXrrs this is quite dangerous, it is possible
  918                  * for the new function to fail to init. We also
  919                  * are not asking if the handoff_is_ok though at
  920                  * the very start thats probalbly ok.
  921                  */
  922                 if (tp->t_fb->tfb_tcp_fb_init) {
  923                         (*tp->t_fb->tfb_tcp_fb_init)(tp);
  924                 }
  925         }               
  926         tp->snd_wl1 = sc->sc_irs;
  927         tp->snd_max = tp->iss + 1;
  928         tp->snd_nxt = tp->iss + 1;
  929         tp->rcv_up = sc->sc_irs + 1;
  930         tp->rcv_wnd = sc->sc_wnd;
  931         tp->rcv_adv += tp->rcv_wnd;
  932         tp->last_ack_sent = tp->rcv_nxt;
  933 
  934         tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY);
  935         if (sc->sc_flags & SCF_NOOPT)
  936                 tp->t_flags |= TF_NOOPT;
  937         else {
  938                 if (sc->sc_flags & SCF_WINSCALE) {
  939                         tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
  940                         tp->snd_scale = sc->sc_requested_s_scale;
  941                         tp->request_r_scale = sc->sc_requested_r_scale;
  942                 }
  943                 if (sc->sc_flags & SCF_TIMESTAMP) {
  944                         tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
  945                         tp->ts_recent = sc->sc_tsreflect;
  946                         tp->ts_recent_age = tcp_ts_getticks();
  947                         tp->ts_offset = sc->sc_tsoff;
  948                 }
  949 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
  950                 if (sc->sc_flags & SCF_SIGNATURE)
  951                         tp->t_flags |= TF_SIGNATURE;
  952 #endif
  953                 if (sc->sc_flags & SCF_SACK)
  954                         tp->t_flags |= TF_SACK_PERMIT;
  955         }
  956 
  957         if (sc->sc_flags & SCF_ECN)
  958                 tp->t_flags |= TF_ECN_PERMIT;
  959 
  960         /*
  961          * Set up MSS and get cached values from tcp_hostcache.
  962          * This might overwrite some of the defaults we just set.
  963          */
  964         tcp_mss(tp, sc->sc_peer_mss);
  965 
  966         /*
  967          * If the SYN,ACK was retransmitted, indicate that CWND to be
  968          * limited to one segment in cc_conn_init().
  969          * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits.
  970          */
  971         if (sc->sc_rxmits > 1)
  972                 tp->snd_cwnd = 1;
  973 
  974 #ifdef TCP_OFFLOAD
  975         /*
  976          * Allow a TOE driver to install its hooks.  Note that we hold the
  977          * pcbinfo lock too and that prevents tcp_usr_accept from accepting a
  978          * new connection before the TOE driver has done its thing.
  979          */
  980         if (ADDED_BY_TOE(sc)) {
  981                 struct toedev *tod = sc->sc_tod;
  982 
  983                 tod->tod_offload_socket(tod, sc->sc_todctx, so);
  984         }
  985 #endif
  986         /*
  987          * Copy and activate timers.
  988          */
  989         tp->t_keepinit = sototcpcb(lso)->t_keepinit;
  990         tp->t_keepidle = sototcpcb(lso)->t_keepidle;
  991         tp->t_keepintvl = sototcpcb(lso)->t_keepintvl;
  992         tp->t_keepcnt = sototcpcb(lso)->t_keepcnt;
  993         tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp));
  994 
  995         TCPSTAT_INC(tcps_accepts);
  996         return (so);
  997 
  998 abort:
  999         INP_WUNLOCK(inp);
 1000 abort2:
 1001         if (so != NULL)
 1002                 soabort(so);
 1003         return (NULL);
 1004 }
 1005 
 1006 /*
 1007  * This function gets called when we receive an ACK for a
 1008  * socket in the LISTEN state.  We look up the connection
 1009  * in the syncache, and if its there, we pull it out of
 1010  * the cache and turn it into a full-blown connection in
 1011  * the SYN-RECEIVED state.
 1012  *
 1013  * On syncache_socket() success the newly created socket
 1014  * has its underlying inp locked.
 1015  */
 1016 int
 1017 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
 1018     struct socket **lsop, struct mbuf *m)
 1019 {
 1020         struct syncache *sc;
 1021         struct syncache_head *sch;
 1022         struct syncache scs;
 1023         char *s;
 1024 
 1025         /*
 1026          * Global TCP locks are held because we manipulate the PCB lists
 1027          * and create a new socket.
 1028          */
 1029         INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
 1030         KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK,
 1031             ("%s: can handle only ACK", __func__));
 1032 
 1033         sc = syncache_lookup(inc, &sch);        /* returns locked sch */
 1034         SCH_LOCK_ASSERT(sch);
 1035 
 1036 #ifdef INVARIANTS
 1037         /*
 1038          * Test code for syncookies comparing the syncache stored
 1039          * values with the reconstructed values from the cookie.
 1040          */
 1041         if (sc != NULL)
 1042                 syncookie_cmp(inc, sch, sc, th, to, *lsop);
 1043 #endif
 1044 
 1045         if (sc == NULL) {
 1046                 /*
 1047                  * There is no syncache entry, so see if this ACK is
 1048                  * a returning syncookie.  To do this, first:
 1049                  *  A. Check if syncookies are used in case of syncache
 1050                  *     overflows
 1051                  *  B. See if this socket has had a syncache entry dropped in
 1052                  *     the recent past. We don't want to accept a bogus
 1053                  *     syncookie if we've never received a SYN or accept it
 1054                  *     twice.
 1055                  *  C. check that the syncookie is valid.  If it is, then
 1056                  *     cobble up a fake syncache entry, and return.
 1057                  */
 1058                 if (!V_tcp_syncookies) {
 1059                         SCH_UNLOCK(sch);
 1060                         if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
 1061                                 log(LOG_DEBUG, "%s; %s: Spurious ACK, "
 1062                                     "segment rejected (syncookies disabled)\n",
 1063                                     s, __func__);
 1064                         goto failed;
 1065                 }
 1066                 if (!V_tcp_syncookiesonly &&
 1067                     sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) {
 1068                         SCH_UNLOCK(sch);
 1069                         if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
 1070                                 log(LOG_DEBUG, "%s; %s: Spurious ACK, "
 1071                                     "segment rejected (no syncache entry)\n",
 1072                                     s, __func__);
 1073                         goto failed;
 1074                 }
 1075                 bzero(&scs, sizeof(scs));
 1076                 sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop);
 1077                 SCH_UNLOCK(sch);
 1078                 if (sc == NULL) {
 1079                         if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
 1080                                 log(LOG_DEBUG, "%s; %s: Segment failed "
 1081                                     "SYNCOOKIE authentication, segment rejected "
 1082                                     "(probably spoofed)\n", s, __func__);
 1083                         goto failed;
 1084                 }
 1085 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
 1086                 /* If received ACK has MD5 signature, check it. */
 1087                 if ((to->to_flags & TOF_SIGNATURE) != 0 &&
 1088                     (!TCPMD5_ENABLED() ||
 1089                     TCPMD5_INPUT(m, th, to->to_signature) != 0)) {
 1090                         /* Drop the ACK. */
 1091                         if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
 1092                                 log(LOG_DEBUG, "%s; %s: Segment rejected, "
 1093                                     "MD5 signature doesn't match.\n",
 1094                                     s, __func__);
 1095                                 free(s, M_TCPLOG);
 1096                         }
 1097                         TCPSTAT_INC(tcps_sig_err_sigopt);
 1098                         return (-1); /* Do not send RST */
 1099                 }
 1100 #endif /* TCP_SIGNATURE */
 1101         } else {
 1102 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
 1103                 /*
 1104                  * If listening socket requested TCP digests, check that
 1105                  * received ACK has signature and it is correct.
 1106                  * If not, drop the ACK and leave sc entry in th cache,
 1107                  * because SYN was received with correct signature.
 1108                  */
 1109                 if (sc->sc_flags & SCF_SIGNATURE) {
 1110                         if ((to->to_flags & TOF_SIGNATURE) == 0) {
 1111                                 /* No signature */
 1112                                 TCPSTAT_INC(tcps_sig_err_nosigopt);
 1113                                 SCH_UNLOCK(sch);
 1114                                 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
 1115                                         log(LOG_DEBUG, "%s; %s: Segment "
 1116                                             "rejected, MD5 signature wasn't "
 1117                                             "provided.\n", s, __func__);
 1118                                         free(s, M_TCPLOG);
 1119                                 }
 1120                                 return (-1); /* Do not send RST */
 1121                         }
 1122                         if (!TCPMD5_ENABLED() ||
 1123                             TCPMD5_INPUT(m, th, to->to_signature) != 0) {
 1124                                 /* Doesn't match or no SA */
 1125                                 SCH_UNLOCK(sch);
 1126                                 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
 1127                                         log(LOG_DEBUG, "%s; %s: Segment "
 1128                                             "rejected, MD5 signature doesn't "
 1129                                             "match.\n", s, __func__);
 1130                                         free(s, M_TCPLOG);
 1131                                 }
 1132                                 return (-1); /* Do not send RST */
 1133                         }
 1134                 }
 1135 #endif /* TCP_SIGNATURE */
 1136                 /*
 1137                  * Pull out the entry to unlock the bucket row.
 1138                  * 
 1139                  * NOTE: We must decrease TCPS_SYN_RECEIVED count here, not
 1140                  * tcp_state_change().  The tcpcb is not existent at this
 1141                  * moment.  A new one will be allocated via syncache_socket->
 1142                  * sonewconn->tcp_usr_attach in TCPS_CLOSED state, then
 1143                  * syncache_socket() will change it to TCPS_SYN_RECEIVED.
 1144                  */
 1145                 TCPSTATES_DEC(TCPS_SYN_RECEIVED);
 1146                 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
 1147                 sch->sch_length--;
 1148 #ifdef TCP_OFFLOAD
 1149                 if (ADDED_BY_TOE(sc)) {
 1150                         struct toedev *tod = sc->sc_tod;
 1151 
 1152                         tod->tod_syncache_removed(tod, sc->sc_todctx);
 1153                 }
 1154 #endif
 1155                 SCH_UNLOCK(sch);
 1156         }
 1157 
 1158         /*
 1159          * Segment validation:
 1160          * ACK must match our initial sequence number + 1 (the SYN|ACK).
 1161          */
 1162         if (th->th_ack != sc->sc_iss + 1) {
 1163                 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
 1164                         log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment "
 1165                             "rejected\n", s, __func__, th->th_ack, sc->sc_iss);
 1166                 goto failed;
 1167         }
 1168 
 1169         /*
 1170          * The SEQ must fall in the window starting at the received
 1171          * initial receive sequence number + 1 (the SYN).
 1172          */
 1173         if (SEQ_LEQ(th->th_seq, sc->sc_irs) ||
 1174             SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
 1175                 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
 1176                         log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment "
 1177                             "rejected\n", s, __func__, th->th_seq, sc->sc_irs);
 1178                 goto failed;
 1179         }
 1180 
 1181         /*
 1182          * If timestamps were not negotiated during SYN/ACK they
 1183          * must not appear on any segment during this session.
 1184          */
 1185         if (!(sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) {
 1186                 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
 1187                         log(LOG_DEBUG, "%s; %s: Timestamp not expected, "
 1188                             "segment rejected\n", s, __func__);
 1189                 goto failed;
 1190         }
 1191 
 1192         /*
 1193          * If timestamps were negotiated during SYN/ACK they should
 1194          * appear on every segment during this session.
 1195          * XXXAO: This is only informal as there have been unverified
 1196          * reports of non-compliants stacks.
 1197          */
 1198         if ((sc->sc_flags & SCF_TIMESTAMP) && !(to->to_flags & TOF_TS)) {
 1199                 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
 1200                         log(LOG_DEBUG, "%s; %s: Timestamp missing, "
 1201                             "no action\n", s, __func__);
 1202                         free(s, M_TCPLOG);
 1203                         s = NULL;
 1204                 }
 1205         }
 1206 
 1207         *lsop = syncache_socket(sc, *lsop, m);
 1208 
 1209         if (*lsop == NULL)
 1210                 TCPSTAT_INC(tcps_sc_aborted);
 1211         else
 1212                 TCPSTAT_INC(tcps_sc_completed);
 1213 
 1214 /* how do we find the inp for the new socket? */
 1215         if (sc != &scs)
 1216                 syncache_free(sc);
 1217         return (1);
 1218 failed:
 1219         if (sc != NULL && sc != &scs)
 1220                 syncache_free(sc);
 1221         if (s != NULL)
 1222                 free(s, M_TCPLOG);
 1223         *lsop = NULL;
 1224         return (0);
 1225 }
 1226 
 1227 static void
 1228 syncache_tfo_expand(struct syncache *sc, struct socket **lsop, struct mbuf *m,
 1229     uint64_t response_cookie)
 1230 {
 1231         struct inpcb *inp;
 1232         struct tcpcb *tp;
 1233         unsigned int *pending_counter;
 1234 
 1235         /*
 1236          * Global TCP locks are held because we manipulate the PCB lists
 1237          * and create a new socket.
 1238          */
 1239         INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
 1240 
 1241         pending_counter = intotcpcb(sotoinpcb(*lsop))->t_tfo_pending;
 1242         *lsop = syncache_socket(sc, *lsop, m);
 1243         if (*lsop == NULL) {
 1244                 TCPSTAT_INC(tcps_sc_aborted);
 1245                 atomic_subtract_int(pending_counter, 1);
 1246         } else {
 1247                 soisconnected(*lsop);
 1248                 inp = sotoinpcb(*lsop);
 1249                 tp = intotcpcb(inp);
 1250                 tp->t_flags |= TF_FASTOPEN;
 1251                 tp->t_tfo_cookie.server = response_cookie;
 1252                 tp->snd_max = tp->iss;
 1253                 tp->snd_nxt = tp->iss;
 1254                 tp->t_tfo_pending = pending_counter;
 1255                 TCPSTAT_INC(tcps_sc_completed);
 1256         }
 1257 }
 1258 
 1259 /*
 1260  * Given a LISTEN socket and an inbound SYN request, add
 1261  * this to the syn cache, and send back a segment:
 1262  *      <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
 1263  * to the source.
 1264  *
 1265  * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
 1266  * Doing so would require that we hold onto the data and deliver it
 1267  * to the application.  However, if we are the target of a SYN-flood
 1268  * DoS attack, an attacker could send data which would eventually
 1269  * consume all available buffer space if it were ACKed.  By not ACKing
 1270  * the data, we avoid this DoS scenario.
 1271  *
 1272  * The exception to the above is when a SYN with a valid TCP Fast Open (TFO)
 1273  * cookie is processed and a new socket is created.  In this case, any data
 1274  * accompanying the SYN will be queued to the socket by tcp_input() and will
 1275  * be ACKed either when the application sends response data or the delayed
 1276  * ACK timer expires, whichever comes first.
 1277  */
 1278 int
 1279 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
 1280     struct inpcb *inp, struct socket **lsop, struct mbuf *m, void *tod,
 1281     void *todctx)
 1282 {
 1283         struct tcpcb *tp;
 1284         struct socket *so;
 1285         struct syncache *sc = NULL;
 1286         struct syncache_head *sch;
 1287         struct mbuf *ipopts = NULL;
 1288         u_int ltflags;
 1289         int win, ip_ttl, ip_tos;
 1290         char *s;
 1291         int rv = 0;
 1292 #ifdef INET6
 1293         int autoflowlabel = 0;
 1294 #endif
 1295 #ifdef MAC
 1296         struct label *maclabel;
 1297 #endif
 1298         struct syncache scs;
 1299         struct ucred *cred;
 1300         uint64_t tfo_response_cookie;
 1301         unsigned int *tfo_pending = NULL;
 1302         int tfo_cookie_valid = 0;
 1303         int tfo_response_cookie_valid = 0;
 1304 
 1305         INP_WLOCK_ASSERT(inp);                  /* listen socket */
 1306         KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN,
 1307             ("%s: unexpected tcp flags", __func__));
 1308 
 1309         /*
 1310          * Combine all so/tp operations very early to drop the INP lock as
 1311          * soon as possible.
 1312          */
 1313         so = *lsop;
 1314         KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so));
 1315         tp = sototcpcb(so);
 1316         cred = crhold(so->so_cred);
 1317 
 1318 #ifdef INET6
 1319         if ((inc->inc_flags & INC_ISIPV6) &&
 1320             (inp->inp_flags & IN6P_AUTOFLOWLABEL))
 1321                 autoflowlabel = 1;
 1322 #endif
 1323         ip_ttl = inp->inp_ip_ttl;
 1324         ip_tos = inp->inp_ip_tos;
 1325         win = so->sol_sbrcv_hiwat;
 1326         ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE));
 1327 
 1328         if (V_tcp_fastopen_server_enable && IS_FASTOPEN(tp->t_flags) &&
 1329             (tp->t_tfo_pending != NULL) &&
 1330             (to->to_flags & TOF_FASTOPEN)) {
 1331                 /*
 1332                  * Limit the number of pending TFO connections to
 1333                  * approximately half of the queue limit.  This prevents TFO
 1334                  * SYN floods from starving the service by filling the
 1335                  * listen queue with bogus TFO connections.
 1336                  */
 1337                 if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <=
 1338                     (so->sol_qlimit / 2)) {
 1339                         int result;
 1340 
 1341                         result = tcp_fastopen_check_cookie(inc,
 1342                             to->to_tfo_cookie, to->to_tfo_len,
 1343                             &tfo_response_cookie);
 1344                         tfo_cookie_valid = (result > 0);
 1345                         tfo_response_cookie_valid = (result >= 0);
 1346                 }
 1347 
 1348                 /*
 1349                  * Remember the TFO pending counter as it will have to be
 1350                  * decremented below if we don't make it to syncache_tfo_expand().
 1351                  */
 1352                 tfo_pending = tp->t_tfo_pending;
 1353         }
 1354 
 1355         /* By the time we drop the lock these should no longer be used. */
 1356         so = NULL;
 1357         tp = NULL;
 1358 
 1359 #ifdef MAC
 1360         if (mac_syncache_init(&maclabel) != 0) {
 1361                 INP_WUNLOCK(inp);
 1362                 goto done;
 1363         } else
 1364                 mac_syncache_create(maclabel, inp);
 1365 #endif
 1366         if (!tfo_cookie_valid)
 1367                 INP_WUNLOCK(inp);
 1368 
 1369         /*
 1370          * Remember the IP options, if any.
 1371          */
 1372 #ifdef INET6
 1373         if (!(inc->inc_flags & INC_ISIPV6))
 1374 #endif
 1375 #ifdef INET
 1376                 ipopts = (m) ? ip_srcroute(m) : NULL;
 1377 #else
 1378                 ipopts = NULL;
 1379 #endif
 1380 
 1381 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
 1382         /*
 1383          * If listening socket requested TCP digests, check that received
 1384          * SYN has signature and it is correct. If signature doesn't match
 1385          * or TCP_SIGNATURE support isn't enabled, drop the packet.
 1386          */
 1387         if (ltflags & TF_SIGNATURE) {
 1388                 if ((to->to_flags & TOF_SIGNATURE) == 0) {
 1389                         TCPSTAT_INC(tcps_sig_err_nosigopt);
 1390                         goto done;
 1391                 }
 1392                 if (!TCPMD5_ENABLED() ||
 1393                     TCPMD5_INPUT(m, th, to->to_signature) != 0)
 1394                         goto done;
 1395         }
 1396 #endif  /* TCP_SIGNATURE */
 1397         /*
 1398          * See if we already have an entry for this connection.
 1399          * If we do, resend the SYN,ACK, and reset the retransmit timer.
 1400          *
 1401          * XXX: should the syncache be re-initialized with the contents
 1402          * of the new SYN here (which may have different options?)
 1403          *
 1404          * XXX: We do not check the sequence number to see if this is a
 1405          * real retransmit or a new connection attempt.  The question is
 1406          * how to handle such a case; either ignore it as spoofed, or
 1407          * drop the current entry and create a new one?
 1408          */
 1409         sc = syncache_lookup(inc, &sch);        /* returns locked entry */
 1410         SCH_LOCK_ASSERT(sch);
 1411         if (sc != NULL) {
 1412                 if (tfo_cookie_valid)
 1413                         INP_WUNLOCK(inp);
 1414                 TCPSTAT_INC(tcps_sc_dupsyn);
 1415                 if (ipopts) {
 1416                         /*
 1417                          * If we were remembering a previous source route,
 1418                          * forget it and use the new one we've been given.
 1419                          */
 1420                         if (sc->sc_ipopts)
 1421                                 (void) m_free(sc->sc_ipopts);
 1422                         sc->sc_ipopts = ipopts;
 1423                 }
 1424                 /*
 1425                  * Update timestamp if present.
 1426                  */
 1427                 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS))
 1428                         sc->sc_tsreflect = to->to_tsval;
 1429                 else
 1430                         sc->sc_flags &= ~SCF_TIMESTAMP;
 1431 #ifdef MAC
 1432                 /*
 1433                  * Since we have already unconditionally allocated label
 1434                  * storage, free it up.  The syncache entry will already
 1435                  * have an initialized label we can use.
 1436                  */
 1437                 mac_syncache_destroy(&maclabel);
 1438 #endif
 1439                 TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
 1440                 /* Retransmit SYN|ACK and reset retransmit count. */
 1441                 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) {
 1442                         log(LOG_DEBUG, "%s; %s: Received duplicate SYN, "
 1443                             "resetting timer and retransmitting SYN|ACK\n",
 1444                             s, __func__);
 1445                         free(s, M_TCPLOG);
 1446                 }
 1447                 if (syncache_respond(sc, sch, m, TH_SYN|TH_ACK) == 0) {
 1448                         sc->sc_rxmits = 0;
 1449                         syncache_timeout(sc, sch, 1);
 1450                         TCPSTAT_INC(tcps_sndacks);
 1451                         TCPSTAT_INC(tcps_sndtotal);
 1452                 }
 1453                 SCH_UNLOCK(sch);
 1454                 goto donenoprobe;
 1455         }
 1456 
 1457         if (tfo_cookie_valid) {
 1458                 bzero(&scs, sizeof(scs));
 1459                 sc = &scs;
 1460                 goto skip_alloc;
 1461         }
 1462 
 1463         sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
 1464         if (sc == NULL) {
 1465                 /*
 1466                  * The zone allocator couldn't provide more entries.
 1467                  * Treat this as if the cache was full; drop the oldest
 1468                  * entry and insert the new one.
 1469                  */
 1470                 TCPSTAT_INC(tcps_sc_zonefail);
 1471                 if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) {
 1472                         sch->sch_last_overflow = time_uptime;
 1473                         syncache_drop(sc, sch);
 1474                 }
 1475                 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
 1476                 if (sc == NULL) {
 1477                         if (V_tcp_syncookies) {
 1478                                 bzero(&scs, sizeof(scs));
 1479                                 sc = &scs;
 1480                         } else {
 1481                                 SCH_UNLOCK(sch);
 1482                                 if (ipopts)
 1483                                         (void) m_free(ipopts);
 1484                                 goto done;
 1485                         }
 1486                 }
 1487         }
 1488 
 1489 skip_alloc:
 1490         if (!tfo_cookie_valid && tfo_response_cookie_valid)
 1491                 sc->sc_tfo_cookie = &tfo_response_cookie;
 1492 
 1493         /*
 1494          * Fill in the syncache values.
 1495          */
 1496 #ifdef MAC
 1497         sc->sc_label = maclabel;
 1498 #endif
 1499         sc->sc_cred = cred;
 1500         cred = NULL;
 1501         sc->sc_ipopts = ipopts;
 1502         bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
 1503 #ifdef INET6
 1504         if (!(inc->inc_flags & INC_ISIPV6))
 1505 #endif
 1506         {
 1507                 sc->sc_ip_tos = ip_tos;
 1508                 sc->sc_ip_ttl = ip_ttl;
 1509         }
 1510 #ifdef TCP_OFFLOAD
 1511         sc->sc_tod = tod;
 1512         sc->sc_todctx = todctx;
 1513 #endif
 1514         sc->sc_irs = th->th_seq;
 1515         sc->sc_iss = arc4random();
 1516         sc->sc_flags = 0;
 1517         sc->sc_flowlabel = 0;
 1518 
 1519         /*
 1520          * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN].
 1521          * win was derived from socket earlier in the function.
 1522          */
 1523         win = imax(win, 0);
 1524         win = imin(win, TCP_MAXWIN);
 1525         sc->sc_wnd = win;
 1526 
 1527         if (V_tcp_do_rfc1323) {
 1528                 /*
 1529                  * A timestamp received in a SYN makes
 1530                  * it ok to send timestamp requests and replies.
 1531                  */
 1532                 if (to->to_flags & TOF_TS) {
 1533                         sc->sc_tsreflect = to->to_tsval;
 1534                         sc->sc_flags |= SCF_TIMESTAMP;
 1535                         sc->sc_tsoff = tcp_new_ts_offset(inc);
 1536                 }
 1537                 if (to->to_flags & TOF_SCALE) {
 1538                         int wscale = 0;
 1539 
 1540                         /*
 1541                          * Pick the smallest possible scaling factor that
 1542                          * will still allow us to scale up to sb_max, aka
 1543                          * kern.ipc.maxsockbuf.
 1544                          *
 1545                          * We do this because there are broken firewalls that
 1546                          * will corrupt the window scale option, leading to
 1547                          * the other endpoint believing that our advertised
 1548                          * window is unscaled.  At scale factors larger than
 1549                          * 5 the unscaled window will drop below 1500 bytes,
 1550                          * leading to serious problems when traversing these
 1551                          * broken firewalls.
 1552                          *
 1553                          * With the default maxsockbuf of 256K, a scale factor
 1554                          * of 3 will be chosen by this algorithm.  Those who
 1555                          * choose a larger maxsockbuf should watch out
 1556                          * for the compatibility problems mentioned above.
 1557                          *
 1558                          * RFC1323: The Window field in a SYN (i.e., a <SYN>
 1559                          * or <SYN,ACK>) segment itself is never scaled.
 1560                          */
 1561                         while (wscale < TCP_MAX_WINSHIFT &&
 1562                             (TCP_MAXWIN << wscale) < sb_max)
 1563                                 wscale++;
 1564                         sc->sc_requested_r_scale = wscale;
 1565                         sc->sc_requested_s_scale = to->to_wscale;
 1566                         sc->sc_flags |= SCF_WINSCALE;
 1567                 }
 1568         }
 1569 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
 1570         /*
 1571          * If listening socket requested TCP digests, flag this in the
 1572          * syncache so that syncache_respond() will do the right thing
 1573          * with the SYN+ACK.
 1574          */
 1575         if (ltflags & TF_SIGNATURE)
 1576                 sc->sc_flags |= SCF_SIGNATURE;
 1577 #endif  /* TCP_SIGNATURE */
 1578         if (to->to_flags & TOF_SACKPERM)
 1579                 sc->sc_flags |= SCF_SACK;
 1580         if (to->to_flags & TOF_MSS)
 1581                 sc->sc_peer_mss = to->to_mss;   /* peer mss may be zero */
 1582         if (ltflags & TF_NOOPT)
 1583                 sc->sc_flags |= SCF_NOOPT;
 1584         if ((th->th_flags & (TH_ECE|TH_CWR)) && V_tcp_do_ecn)
 1585                 sc->sc_flags |= SCF_ECN;
 1586 
 1587         if (V_tcp_syncookies)
 1588                 sc->sc_iss = syncookie_generate(sch, sc);
 1589 #ifdef INET6
 1590         if (autoflowlabel) {
 1591                 if (V_tcp_syncookies)
 1592                         sc->sc_flowlabel = sc->sc_iss;
 1593                 else
 1594                         sc->sc_flowlabel = ip6_randomflowlabel();
 1595                 sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK;
 1596         }
 1597 #endif
 1598         SCH_UNLOCK(sch);
 1599 
 1600         if (tfo_cookie_valid) {
 1601                 syncache_tfo_expand(sc, lsop, m, tfo_response_cookie);
 1602                 /* INP_WUNLOCK(inp) will be performed by the caller */
 1603                 rv = 1;
 1604                 goto tfo_expanded;
 1605         }
 1606 
 1607         TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
 1608         /*
 1609          * Do a standard 3-way handshake.
 1610          */
 1611         if (syncache_respond(sc, sch, m, TH_SYN|TH_ACK) == 0) {
 1612                 if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs)
 1613                         syncache_free(sc);
 1614                 else if (sc != &scs)
 1615                         syncache_insert(sc, sch);   /* locks and unlocks sch */
 1616                 TCPSTAT_INC(tcps_sndacks);
 1617                 TCPSTAT_INC(tcps_sndtotal);
 1618         } else {
 1619                 if (sc != &scs)
 1620                         syncache_free(sc);
 1621                 TCPSTAT_INC(tcps_sc_dropped);
 1622         }
 1623         goto donenoprobe;
 1624 
 1625 done:
 1626         TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
 1627 donenoprobe:
 1628         if (m) {
 1629                 *lsop = NULL;
 1630                 m_freem(m);
 1631         }
 1632         /*
 1633          * If tfo_pending is not NULL here, then a TFO SYN that did not
 1634          * result in a new socket was processed and the associated pending
 1635          * counter has not yet been decremented.  All such TFO processing paths
 1636          * transit this point.
 1637          */
 1638         if (tfo_pending != NULL)
 1639                 tcp_fastopen_decrement_counter(tfo_pending);
 1640 
 1641 tfo_expanded:
 1642         if (cred != NULL)
 1643                 crfree(cred);
 1644 #ifdef MAC
 1645         if (sc == &scs)
 1646                 mac_syncache_destroy(&maclabel);
 1647 #endif
 1648         return (rv);
 1649 }
 1650 
 1651 /*
 1652  * Send SYN|ACK or ACK to the peer.  Either in response to a peer's segment,
 1653  * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL.
 1654  */
 1655 static int
 1656 syncache_respond(struct syncache *sc, struct syncache_head *sch,
 1657     const struct mbuf *m0, int flags)
 1658 {
 1659         struct ip *ip = NULL;
 1660         struct mbuf *m;
 1661         struct tcphdr *th = NULL;
 1662         int optlen, error = 0;  /* Make compiler happy */
 1663         u_int16_t hlen, tlen, mssopt;
 1664         struct tcpopt to;
 1665 #ifdef INET6
 1666         struct ip6_hdr *ip6 = NULL;
 1667 #endif
 1668         hlen =
 1669 #ifdef INET6
 1670                (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) :
 1671 #endif
 1672                 sizeof(struct ip);
 1673         tlen = hlen + sizeof(struct tcphdr);
 1674 
 1675         /* Determine MSS we advertize to other end of connection. */
 1676         mssopt = max(tcp_mssopt(&sc->sc_inc), V_tcp_minmss);
 1677 
 1678         /* XXX: Assume that the entire packet will fit in a header mbuf. */
 1679         KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN,
 1680             ("syncache: mbuf too small"));
 1681 
 1682         /* Create the IP+TCP header from scratch. */
 1683         m = m_gethdr(M_NOWAIT, MT_DATA);
 1684         if (m == NULL)
 1685                 return (ENOBUFS);
 1686 #ifdef MAC
 1687         mac_syncache_create_mbuf(sc->sc_label, m);
 1688 #endif
 1689         m->m_data += max_linkhdr;
 1690         m->m_len = tlen;
 1691         m->m_pkthdr.len = tlen;
 1692         m->m_pkthdr.rcvif = NULL;
 1693 
 1694 #ifdef INET6
 1695         if (sc->sc_inc.inc_flags & INC_ISIPV6) {
 1696                 ip6 = mtod(m, struct ip6_hdr *);
 1697                 ip6->ip6_vfc = IPV6_VERSION;
 1698                 ip6->ip6_nxt = IPPROTO_TCP;
 1699                 ip6->ip6_src = sc->sc_inc.inc6_laddr;
 1700                 ip6->ip6_dst = sc->sc_inc.inc6_faddr;
 1701                 ip6->ip6_plen = htons(tlen - hlen);
 1702                 /* ip6_hlim is set after checksum */
 1703                 ip6->ip6_flow &= ~IPV6_FLOWLABEL_MASK;
 1704                 ip6->ip6_flow |= sc->sc_flowlabel;
 1705 
 1706                 th = (struct tcphdr *)(ip6 + 1);
 1707         }
 1708 #endif
 1709 #if defined(INET6) && defined(INET)
 1710         else
 1711 #endif
 1712 #ifdef INET
 1713         {
 1714                 ip = mtod(m, struct ip *);
 1715                 ip->ip_v = IPVERSION;
 1716                 ip->ip_hl = sizeof(struct ip) >> 2;
 1717                 ip->ip_len = htons(tlen);
 1718                 ip->ip_id = 0;
 1719                 ip->ip_off = 0;
 1720                 ip->ip_sum = 0;
 1721                 ip->ip_p = IPPROTO_TCP;
 1722                 ip->ip_src = sc->sc_inc.inc_laddr;
 1723                 ip->ip_dst = sc->sc_inc.inc_faddr;
 1724                 ip->ip_ttl = sc->sc_ip_ttl;
 1725                 ip->ip_tos = sc->sc_ip_tos;
 1726 
 1727                 /*
 1728                  * See if we should do MTU discovery.  Route lookups are
 1729                  * expensive, so we will only unset the DF bit if:
 1730                  *
 1731                  *      1) path_mtu_discovery is disabled
 1732                  *      2) the SCF_UNREACH flag has been set
 1733                  */
 1734                 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
 1735                        ip->ip_off |= htons(IP_DF);
 1736 
 1737                 th = (struct tcphdr *)(ip + 1);
 1738         }
 1739 #endif /* INET */
 1740         th->th_sport = sc->sc_inc.inc_lport;
 1741         th->th_dport = sc->sc_inc.inc_fport;
 1742 
 1743         if (flags & TH_SYN)
 1744                 th->th_seq = htonl(sc->sc_iss);
 1745         else
 1746                 th->th_seq = htonl(sc->sc_iss + 1);
 1747         th->th_ack = htonl(sc->sc_irs + 1);
 1748         th->th_off = sizeof(struct tcphdr) >> 2;
 1749         th->th_x2 = 0;
 1750         th->th_flags = flags;
 1751         th->th_win = htons(sc->sc_wnd);
 1752         th->th_urp = 0;
 1753 
 1754         if ((flags & TH_SYN) && (sc->sc_flags & SCF_ECN)) {
 1755                 th->th_flags |= TH_ECE;
 1756                 TCPSTAT_INC(tcps_ecn_shs);
 1757         }
 1758 
 1759         /* Tack on the TCP options. */
 1760         if ((sc->sc_flags & SCF_NOOPT) == 0) {
 1761                 to.to_flags = 0;
 1762 
 1763                 if (flags & TH_SYN) {
 1764                         to.to_mss = mssopt;
 1765                         to.to_flags = TOF_MSS;
 1766                         if (sc->sc_flags & SCF_WINSCALE) {
 1767                                 to.to_wscale = sc->sc_requested_r_scale;
 1768                                 to.to_flags |= TOF_SCALE;
 1769                         }
 1770                         if (sc->sc_flags & SCF_SACK)
 1771                                 to.to_flags |= TOF_SACKPERM;
 1772 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
 1773                         if (sc->sc_flags & SCF_SIGNATURE)
 1774                                 to.to_flags |= TOF_SIGNATURE;
 1775 #endif
 1776                         if (sc->sc_tfo_cookie) {
 1777                                 to.to_flags |= TOF_FASTOPEN;
 1778                                 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
 1779                                 to.to_tfo_cookie = sc->sc_tfo_cookie;
 1780                                 /* don't send cookie again when retransmitting response */
 1781                                 sc->sc_tfo_cookie = NULL;
 1782                         }
 1783                 }
 1784                 if (sc->sc_flags & SCF_TIMESTAMP) {
 1785                         to.to_tsval = sc->sc_tsoff + tcp_ts_getticks();
 1786                         to.to_tsecr = sc->sc_tsreflect;
 1787                         to.to_flags |= TOF_TS;
 1788                 }
 1789                 optlen = tcp_addoptions(&to, (u_char *)(th + 1));
 1790 
 1791                 /* Adjust headers by option size. */
 1792                 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
 1793                 m->m_len += optlen;
 1794                 m->m_pkthdr.len += optlen;
 1795 #ifdef INET6
 1796                 if (sc->sc_inc.inc_flags & INC_ISIPV6)
 1797                         ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen);
 1798                 else
 1799 #endif
 1800                         ip->ip_len = htons(ntohs(ip->ip_len) + optlen);
 1801 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
 1802                 if (sc->sc_flags & SCF_SIGNATURE) {
 1803                         KASSERT(to.to_flags & TOF_SIGNATURE,
 1804                             ("tcp_addoptions() didn't set tcp_signature"));
 1805 
 1806                         /* NOTE: to.to_signature is inside of mbuf */
 1807                         if (!TCPMD5_ENABLED() ||
 1808                             TCPMD5_OUTPUT(m, th, to.to_signature) != 0) {
 1809                                 m_freem(m);
 1810                                 return (EACCES);
 1811                         }
 1812                 }
 1813 #endif
 1814         } else
 1815                 optlen = 0;
 1816 
 1817         M_SETFIB(m, sc->sc_inc.inc_fibnum);
 1818         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
 1819         /*
 1820          * If we have peer's SYN and it has a flowid, then let's assign it to
 1821          * our SYN|ACK.  ip6_output() and ip_output() will not assign flowid
 1822          * to SYN|ACK due to lack of inp here.
 1823          */
 1824         if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) {
 1825                 m->m_pkthdr.flowid = m0->m_pkthdr.flowid;
 1826                 M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0));
 1827         }
 1828 #ifdef INET6
 1829         if (sc->sc_inc.inc_flags & INC_ISIPV6) {
 1830                 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
 1831                 th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen,
 1832                     IPPROTO_TCP, 0);
 1833                 ip6->ip6_hlim = in6_selecthlim(NULL, NULL);
 1834 #ifdef TCP_OFFLOAD
 1835                 if (ADDED_BY_TOE(sc)) {
 1836                         struct toedev *tod = sc->sc_tod;
 1837 
 1838                         error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
 1839 
 1840                         return (error);
 1841                 }
 1842 #endif
 1843                 TCP_PROBE5(send, NULL, NULL, ip6, NULL, th);
 1844                 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL);
 1845         }
 1846 #endif
 1847 #if defined(INET6) && defined(INET)
 1848         else
 1849 #endif
 1850 #ifdef INET
 1851         {
 1852                 m->m_pkthdr.csum_flags = CSUM_TCP;
 1853                 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
 1854                     htons(tlen + optlen - hlen + IPPROTO_TCP));
 1855 #ifdef TCP_OFFLOAD
 1856                 if (ADDED_BY_TOE(sc)) {
 1857                         struct toedev *tod = sc->sc_tod;
 1858 
 1859                         error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
 1860 
 1861                         return (error);
 1862                 }
 1863 #endif
 1864                 TCP_PROBE5(send, NULL, NULL, ip, NULL, th);
 1865                 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL);
 1866         }
 1867 #endif
 1868         return (error);
 1869 }
 1870 
 1871 /*
 1872  * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks
 1873  * that exceed the capacity of the syncache by avoiding the storage of any
 1874  * of the SYNs we receive.  Syncookies defend against blind SYN flooding
 1875  * attacks where the attacker does not have access to our responses.
 1876  *
 1877  * Syncookies encode and include all necessary information about the
 1878  * connection setup within the SYN|ACK that we send back.  That way we
 1879  * can avoid keeping any local state until the ACK to our SYN|ACK returns
 1880  * (if ever).  Normally the syncache and syncookies are running in parallel
 1881  * with the latter taking over when the former is exhausted.  When matching
 1882  * syncache entry is found the syncookie is ignored.
 1883  *
 1884  * The only reliable information persisting the 3WHS is our initial sequence
 1885  * number ISS of 32 bits.  Syncookies embed a cryptographically sufficient
 1886  * strong hash (MAC) value and a few bits of TCP SYN options in the ISS
 1887  * of our SYN|ACK.  The MAC can be recomputed when the ACK to our SYN|ACK
 1888  * returns and signifies a legitimate connection if it matches the ACK.
 1889  *
 1890  * The available space of 32 bits to store the hash and to encode the SYN
 1891  * option information is very tight and we should have at least 24 bits for
 1892  * the MAC to keep the number of guesses by blind spoofing reasonably high.
 1893  *
 1894  * SYN option information we have to encode to fully restore a connection:
 1895  * MSS: is imporant to chose an optimal segment size to avoid IP level
 1896  *   fragmentation along the path.  The common MSS values can be encoded
 1897  *   in a 3-bit table.  Uncommon values are captured by the next lower value
 1898  *   in the table leading to a slight increase in packetization overhead.
 1899  * WSCALE: is necessary to allow large windows to be used for high delay-
 1900  *   bandwidth product links.  Not scaling the window when it was initially
 1901  *   negotiated is bad for performance as lack of scaling further decreases
 1902  *   the apparent available send window.  We only need to encode the WSCALE
 1903  *   we received from the remote end.  Our end can be recalculated at any
 1904  *   time.  The common WSCALE values can be encoded in a 3-bit table.
 1905  *   Uncommon values are captured by the next lower value in the table
 1906  *   making us under-estimate the available window size halving our
 1907  *   theoretically possible maximum throughput for that connection.
 1908  * SACK: Greatly assists in packet loss recovery and requires 1 bit.
 1909  * TIMESTAMP and SIGNATURE is not encoded because they are permanent options
 1910  *   that are included in all segments on a connection.  We enable them when
 1911  *   the ACK has them.
 1912  *
 1913  * Security of syncookies and attack vectors:
 1914  *
 1915  * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod)
 1916  * together with the gloabl secret to make it unique per connection attempt.
 1917  * Thus any change of any of those parameters results in a different MAC output
 1918  * in an unpredictable way unless a collision is encountered.  24 bits of the
 1919  * MAC are embedded into the ISS.
 1920  *
 1921  * To prevent replay attacks two rotating global secrets are updated with a
 1922  * new random value every 15 seconds.  The life-time of a syncookie is thus
 1923  * 15-30 seconds.
 1924  *
 1925  * Vector 1: Attacking the secret.  This requires finding a weakness in the
 1926  * MAC itself or the way it is used here.  The attacker can do a chosen plain
 1927  * text attack by varying and testing the all parameters under his control.
 1928  * The strength depends on the size and randomness of the secret, and the
 1929  * cryptographic security of the MAC function.  Due to the constant updating
 1930  * of the secret the attacker has at most 29.999 seconds to find the secret
 1931  * and launch spoofed connections.  After that he has to start all over again.
 1932  *
 1933  * Vector 2: Collision attack on the MAC of a single ACK.  With a 24 bit MAC
 1934  * size an average of 4,823 attempts are required for a 50% chance of success
 1935  * to spoof a single syncookie (birthday collision paradox).  However the
 1936  * attacker is blind and doesn't know if one of his attempts succeeded unless
 1937  * he has a side channel to interfere success from.  A single connection setup
 1938  * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets.
 1939  * This many attempts are required for each one blind spoofed connection.  For
 1940  * every additional spoofed connection he has to launch another N attempts.
 1941  * Thus for a sustained rate 100 spoofed connections per second approximately
 1942  * 1,800,000 packets per second would have to be sent.
 1943  *
 1944  * NB: The MAC function should be fast so that it doesn't become a CPU
 1945  * exhaustion attack vector itself.
 1946  *
 1947  * References:
 1948  *  RFC4987 TCP SYN Flooding Attacks and Common Mitigations
 1949  *  SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996
 1950  *   http://cr.yp.to/syncookies.html    (overview)
 1951  *   http://cr.yp.to/syncookies/archive (details)
 1952  *
 1953  *
 1954  * Schematic construction of a syncookie enabled Initial Sequence Number:
 1955  *  0        1         2         3
 1956  *  12345678901234567890123456789012
 1957  * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP|
 1958  *
 1959  *  x 24 MAC (truncated)
 1960  *  W  3 Send Window Scale index
 1961  *  M  3 MSS index
 1962  *  S  1 SACK permitted
 1963  *  P  1 Odd/even secret
 1964  */
 1965 
 1966 /*
 1967  * Distribution and probability of certain MSS values.  Those in between are
 1968  * rounded down to the next lower one.
 1969  * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011]
 1970  *                            .2%  .3%   5%    7%    7%    20%   15%   45%
 1971  */
 1972 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 };
 1973 
 1974 /*
 1975  * Distribution and probability of certain WSCALE values.  We have to map the
 1976  * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3
 1977  * bits based on prevalence of certain values.  Where we don't have an exact
 1978  * match for are rounded down to the next lower one letting us under-estimate
 1979  * the true available window.  At the moment this would happen only for the
 1980  * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer
 1981  * and window size).  The absence of the WSCALE option (no scaling in either
 1982  * direction) is encoded with index zero.
 1983  * [WSCALE values histograms, Allman, 2012]
 1984  *                            X 10 10 35  5  6 14 10%   by host
 1985  *                            X 11  4  5  5 18 49  3%   by connections
 1986  */
 1987 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 };
 1988 
 1989 /*
 1990  * Compute the MAC for the SYN cookie.  SIPHASH-2-4 is chosen for its speed
 1991  * and good cryptographic properties.
 1992  */
 1993 static uint32_t
 1994 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags,
 1995     uint8_t *secbits, uintptr_t secmod)
 1996 {
 1997         SIPHASH_CTX ctx;
 1998         uint32_t siphash[2];
 1999 
 2000         SipHash24_Init(&ctx);
 2001         SipHash_SetKey(&ctx, secbits);
 2002         switch (inc->inc_flags & INC_ISIPV6) {
 2003 #ifdef INET
 2004         case 0:
 2005                 SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr));
 2006                 SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr));
 2007                 break;
 2008 #endif
 2009 #ifdef INET6
 2010         case INC_ISIPV6:
 2011                 SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr));
 2012                 SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr));
 2013                 break;
 2014 #endif
 2015         }
 2016         SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport));
 2017         SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport));
 2018         SipHash_Update(&ctx, &irs, sizeof(irs));
 2019         SipHash_Update(&ctx, &flags, sizeof(flags));
 2020         SipHash_Update(&ctx, &secmod, sizeof(secmod));
 2021         SipHash_Final((u_int8_t *)&siphash, &ctx);
 2022 
 2023         return (siphash[0] ^ siphash[1]);
 2024 }
 2025 
 2026 static tcp_seq
 2027 syncookie_generate(struct syncache_head *sch, struct syncache *sc)
 2028 {
 2029         u_int i, secbit, wscale;
 2030         uint32_t iss, hash;
 2031         uint8_t *secbits;
 2032         union syncookie cookie;
 2033 
 2034         SCH_LOCK_ASSERT(sch);
 2035 
 2036         cookie.cookie = 0;
 2037 
 2038         /* Map our computed MSS into the 3-bit index. */
 2039         for (i = nitems(tcp_sc_msstab) - 1;
 2040              tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0;
 2041              i--)
 2042                 ;
 2043         cookie.flags.mss_idx = i;
 2044 
 2045         /*
 2046          * Map the send window scale into the 3-bit index but only if
 2047          * the wscale option was received.
 2048          */
 2049         if (sc->sc_flags & SCF_WINSCALE) {
 2050                 wscale = sc->sc_requested_s_scale;
 2051                 for (i = nitems(tcp_sc_wstab) - 1;
 2052                     tcp_sc_wstab[i] > wscale && i > 0;
 2053                      i--)
 2054                         ;
 2055                 cookie.flags.wscale_idx = i;
 2056         }
 2057 
 2058         /* Can we do SACK? */
 2059         if (sc->sc_flags & SCF_SACK)
 2060                 cookie.flags.sack_ok = 1;
 2061 
 2062         /* Which of the two secrets to use. */
 2063         secbit = sch->sch_sc->secret.oddeven & 0x1;
 2064         cookie.flags.odd_even = secbit;
 2065 
 2066         secbits = sch->sch_sc->secret.key[secbit];
 2067         hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits,
 2068             (uintptr_t)sch);
 2069 
 2070         /*
 2071          * Put the flags into the hash and XOR them to get better ISS number
 2072          * variance.  This doesn't enhance the cryptographic strength and is
 2073          * done to prevent the 8 cookie bits from showing up directly on the
 2074          * wire.
 2075          */
 2076         iss = hash & ~0xff;
 2077         iss |= cookie.cookie ^ (hash >> 24);
 2078 
 2079         TCPSTAT_INC(tcps_sc_sendcookie);
 2080         return (iss);
 2081 }
 2082 
 2083 static struct syncache *
 2084 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch, 
 2085     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
 2086     struct socket *lso)
 2087 {
 2088         uint32_t hash;
 2089         uint8_t *secbits;
 2090         tcp_seq ack, seq;
 2091         int wnd, wscale = 0;
 2092         union syncookie cookie;
 2093 
 2094         SCH_LOCK_ASSERT(sch);
 2095 
 2096         /*
 2097          * Pull information out of SYN-ACK/ACK and revert sequence number
 2098          * advances.
 2099          */
 2100         ack = th->th_ack - 1;
 2101         seq = th->th_seq - 1;
 2102 
 2103         /*
 2104          * Unpack the flags containing enough information to restore the
 2105          * connection.
 2106          */
 2107         cookie.cookie = (ack & 0xff) ^ (ack >> 24);
 2108 
 2109         /* Which of the two secrets to use. */
 2110         secbits = sch->sch_sc->secret.key[cookie.flags.odd_even];
 2111 
 2112         hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch);
 2113 
 2114         /* The recomputed hash matches the ACK if this was a genuine cookie. */
 2115         if ((ack & ~0xff) != (hash & ~0xff))
 2116                 return (NULL);
 2117 
 2118         /* Fill in the syncache values. */
 2119         sc->sc_flags = 0;
 2120         bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
 2121         sc->sc_ipopts = NULL;
 2122         
 2123         sc->sc_irs = seq;
 2124         sc->sc_iss = ack;
 2125 
 2126         switch (inc->inc_flags & INC_ISIPV6) {
 2127 #ifdef INET
 2128         case 0:
 2129                 sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl;
 2130                 sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos;
 2131                 break;
 2132 #endif
 2133 #ifdef INET6
 2134         case INC_ISIPV6:
 2135                 if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL)
 2136                         sc->sc_flowlabel = sc->sc_iss & IPV6_FLOWLABEL_MASK;
 2137                 break;
 2138 #endif
 2139         }
 2140 
 2141         sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx];
 2142 
 2143         /* We can simply recompute receive window scale we sent earlier. */
 2144         while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max)
 2145                 wscale++;
 2146 
 2147         /* Only use wscale if it was enabled in the orignal SYN. */
 2148         if (cookie.flags.wscale_idx > 0) {
 2149                 sc->sc_requested_r_scale = wscale;
 2150                 sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx];
 2151                 sc->sc_flags |= SCF_WINSCALE;
 2152         }
 2153 
 2154         wnd = lso->sol_sbrcv_hiwat;
 2155         wnd = imax(wnd, 0);
 2156         wnd = imin(wnd, TCP_MAXWIN);
 2157         sc->sc_wnd = wnd;
 2158 
 2159         if (cookie.flags.sack_ok)
 2160                 sc->sc_flags |= SCF_SACK;
 2161 
 2162         if (to->to_flags & TOF_TS) {
 2163                 sc->sc_flags |= SCF_TIMESTAMP;
 2164                 sc->sc_tsreflect = to->to_tsval;
 2165                 sc->sc_tsoff = tcp_new_ts_offset(inc);
 2166         }
 2167 
 2168         if (to->to_flags & TOF_SIGNATURE)
 2169                 sc->sc_flags |= SCF_SIGNATURE;
 2170 
 2171         sc->sc_rxmits = 0;
 2172 
 2173         TCPSTAT_INC(tcps_sc_recvcookie);
 2174         return (sc);
 2175 }
 2176 
 2177 #ifdef INVARIANTS
 2178 static int
 2179 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
 2180     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
 2181     struct socket *lso)
 2182 {
 2183         struct syncache scs, *scx;
 2184         char *s;
 2185 
 2186         bzero(&scs, sizeof(scs));
 2187         scx = syncookie_lookup(inc, sch, &scs, th, to, lso);
 2188 
 2189         if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL)
 2190                 return (0);
 2191 
 2192         if (scx != NULL) {
 2193                 if (sc->sc_peer_mss != scx->sc_peer_mss)
 2194                         log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n",
 2195                             s, __func__, sc->sc_peer_mss, scx->sc_peer_mss);
 2196 
 2197                 if (sc->sc_requested_r_scale != scx->sc_requested_r_scale)
 2198                         log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n",
 2199                             s, __func__, sc->sc_requested_r_scale,
 2200                             scx->sc_requested_r_scale);
 2201 
 2202                 if (sc->sc_requested_s_scale != scx->sc_requested_s_scale)
 2203                         log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n",
 2204                             s, __func__, sc->sc_requested_s_scale,
 2205                             scx->sc_requested_s_scale);
 2206 
 2207                 if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK))
 2208                         log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__);
 2209         }
 2210 
 2211         if (s != NULL)
 2212                 free(s, M_TCPLOG);
 2213         return (0);
 2214 }
 2215 #endif /* INVARIANTS */
 2216 
 2217 static void
 2218 syncookie_reseed(void *arg)
 2219 {
 2220         struct tcp_syncache *sc = arg;
 2221         uint8_t *secbits;
 2222         int secbit;
 2223 
 2224         /*
 2225          * Reseeding the secret doesn't have to be protected by a lock.
 2226          * It only must be ensured that the new random values are visible
 2227          * to all CPUs in a SMP environment.  The atomic with release
 2228          * semantics ensures that.
 2229          */
 2230         secbit = (sc->secret.oddeven & 0x1) ? 0 : 1;
 2231         secbits = sc->secret.key[secbit];
 2232         arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0);
 2233         atomic_add_rel_int(&sc->secret.oddeven, 1);
 2234 
 2235         /* Reschedule ourself. */
 2236         callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz);
 2237 }
 2238 
 2239 /*
 2240  * Exports the syncache entries to userland so that netstat can display
 2241  * them alongside the other sockets.  This function is intended to be
 2242  * called only from tcp_pcblist.
 2243  *
 2244  * Due to concurrency on an active system, the number of pcbs exported
 2245  * may have no relation to max_pcbs.  max_pcbs merely indicates the
 2246  * amount of space the caller allocated for this function to use.
 2247  */
 2248 int
 2249 syncache_pcblist(struct sysctl_req *req, int max_pcbs, int *pcbs_exported)
 2250 {
 2251         struct xtcpcb xt;
 2252         struct syncache *sc;
 2253         struct syncache_head *sch;
 2254         int count, error, i;
 2255 
 2256         for (count = 0, error = 0, i = 0; i < V_tcp_syncache.hashsize; i++) {
 2257                 sch = &V_tcp_syncache.hashbase[i];
 2258                 SCH_LOCK(sch);
 2259                 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
 2260                         if (count >= max_pcbs) {
 2261                                 SCH_UNLOCK(sch);
 2262                                 goto exit;
 2263                         }
 2264                         if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0)
 2265                                 continue;
 2266                         bzero(&xt, sizeof(xt));
 2267                         xt.xt_len = sizeof(xt);
 2268                         if (sc->sc_inc.inc_flags & INC_ISIPV6)
 2269                                 xt.xt_inp.inp_vflag = INP_IPV6;
 2270                         else
 2271                                 xt.xt_inp.inp_vflag = INP_IPV4;
 2272                         bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc,
 2273                             sizeof (struct in_conninfo));
 2274                         xt.t_state = TCPS_SYN_RECEIVED;
 2275                         xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP;
 2276                         xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket);
 2277                         xt.xt_inp.xi_socket.so_type = SOCK_STREAM;
 2278                         xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING;
 2279                         error = SYSCTL_OUT(req, &xt, sizeof xt);
 2280                         if (error) {
 2281                                 SCH_UNLOCK(sch);
 2282                                 goto exit;
 2283                         }
 2284                         count++;
 2285                 }
 2286                 SCH_UNLOCK(sch);
 2287         }
 2288 exit:
 2289         *pcbs_exported = count;
 2290         return error;
 2291 }

Cache object: 0c1ca7c7c3fbd1cabc93ebdd2f1ac65b


[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]


This page is part of the FreeBSD/Linux Linux Kernel Cross-Reference, and was automatically generated using a modified version of the LXR engine.