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: head/sys/netinet/tcp_syncache.c 345458 2019-03-23 21:36:59Z 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 with default values of
  158  * tcp_rexmit_initial * (             1 +
  159  *                       tcp_backoff[1] +
  160  *                       tcp_backoff[2] +
  161  *                       tcp_backoff[3]) + 3 * tcp_rexmit_slop,
  162  * 1000 ms * (1 + 2 + 4 + 8) +  3 * 200 ms = 15600 ms,
  163  * the odds are that the user has given up attempting to connect by then.
  164  */
  165 #define SYNCACHE_MAXREXMTS              3
  166 
  167 /* Arbitrary values */
  168 #define TCP_SYNCACHE_HASHSIZE           512
  169 #define TCP_SYNCACHE_BUCKETLIMIT        30
  170 
  171 VNET_DEFINE_STATIC(struct tcp_syncache, tcp_syncache);
  172 #define V_tcp_syncache                  VNET(tcp_syncache)
  173 
  174 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, CTLFLAG_RW, 0,
  175     "TCP SYN cache");
  176 
  177 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
  178     &VNET_NAME(tcp_syncache.bucket_limit), 0,
  179     "Per-bucket hash limit for syncache");
  180 
  181 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
  182     &VNET_NAME(tcp_syncache.cache_limit), 0,
  183     "Overall entry limit for syncache");
  184 
  185 SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET,
  186     &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache");
  187 
  188 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN,
  189     &VNET_NAME(tcp_syncache.hashsize), 0,
  190     "Size of TCP syncache hashtable");
  191 
  192 static int
  193 sysctl_net_inet_tcp_syncache_rexmtlimit_check(SYSCTL_HANDLER_ARGS)
  194 {
  195         int error;
  196         u_int new;
  197 
  198         new = V_tcp_syncache.rexmt_limit;
  199         error = sysctl_handle_int(oidp, &new, 0, req);
  200         if ((error == 0) && (req->newptr != NULL)) {
  201                 if (new > TCP_MAXRXTSHIFT)
  202                         error = EINVAL;
  203                 else
  204                         V_tcp_syncache.rexmt_limit = new;
  205         }
  206         return (error);
  207 }
  208 
  209 SYSCTL_PROC(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit,
  210     CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW,
  211     &VNET_NAME(tcp_syncache.rexmt_limit), 0,
  212     sysctl_net_inet_tcp_syncache_rexmtlimit_check, "UI",
  213     "Limit on SYN/ACK retransmissions");
  214 
  215 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1;
  216 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail,
  217     CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0,
  218     "Send reset on socket allocation failure");
  219 
  220 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache");
  221 
  222 #define SCH_LOCK(sch)           mtx_lock(&(sch)->sch_mtx)
  223 #define SCH_UNLOCK(sch)         mtx_unlock(&(sch)->sch_mtx)
  224 #define SCH_LOCK_ASSERT(sch)    mtx_assert(&(sch)->sch_mtx, MA_OWNED)
  225 
  226 /*
  227  * Requires the syncache entry to be already removed from the bucket list.
  228  */
  229 static void
  230 syncache_free(struct syncache *sc)
  231 {
  232 
  233         if (sc->sc_ipopts)
  234                 (void) m_free(sc->sc_ipopts);
  235         if (sc->sc_cred)
  236                 crfree(sc->sc_cred);
  237 #ifdef MAC
  238         mac_syncache_destroy(&sc->sc_label);
  239 #endif
  240 
  241         uma_zfree(V_tcp_syncache.zone, sc);
  242 }
  243 
  244 void
  245 syncache_init(void)
  246 {
  247         int i;
  248 
  249         V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
  250         V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT;
  251         V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS;
  252         V_tcp_syncache.hash_secret = arc4random();
  253 
  254         TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize",
  255             &V_tcp_syncache.hashsize);
  256         TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit",
  257             &V_tcp_syncache.bucket_limit);
  258         if (!powerof2(V_tcp_syncache.hashsize) ||
  259             V_tcp_syncache.hashsize == 0) {
  260                 printf("WARNING: syncache hash size is not a power of 2.\n");
  261                 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE;
  262         }
  263         V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1;
  264 
  265         /* Set limits. */
  266         V_tcp_syncache.cache_limit =
  267             V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit;
  268         TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit",
  269             &V_tcp_syncache.cache_limit);
  270 
  271         /* Allocate the hash table. */
  272         V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize *
  273             sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO);
  274 
  275 #ifdef VIMAGE
  276         V_tcp_syncache.vnet = curvnet;
  277 #endif
  278 
  279         /* Initialize the hash buckets. */
  280         for (i = 0; i < V_tcp_syncache.hashsize; i++) {
  281                 TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket);
  282                 mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head",
  283                          NULL, MTX_DEF);
  284                 callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer,
  285                          &V_tcp_syncache.hashbase[i].sch_mtx, 0);
  286                 V_tcp_syncache.hashbase[i].sch_length = 0;
  287                 V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache;
  288                 V_tcp_syncache.hashbase[i].sch_last_overflow =
  289                     -(SYNCOOKIE_LIFETIME + 1);
  290         }
  291 
  292         /* Create the syncache entry zone. */
  293         V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache),
  294             NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
  295         V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone,
  296             V_tcp_syncache.cache_limit);
  297 
  298         /* Start the SYN cookie reseeder callout. */
  299         callout_init(&V_tcp_syncache.secret.reseed, 1);
  300         arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0);
  301         arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0);
  302         callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz,
  303             syncookie_reseed, &V_tcp_syncache);
  304 }
  305 
  306 #ifdef VIMAGE
  307 void
  308 syncache_destroy(void)
  309 {
  310         struct syncache_head *sch;
  311         struct syncache *sc, *nsc;
  312         int i;
  313 
  314         /*
  315          * Stop the re-seed timer before freeing resources.  No need to
  316          * possibly schedule it another time.
  317          */
  318         callout_drain(&V_tcp_syncache.secret.reseed);
  319 
  320         /* Cleanup hash buckets: stop timers, free entries, destroy locks. */
  321         for (i = 0; i < V_tcp_syncache.hashsize; i++) {
  322 
  323                 sch = &V_tcp_syncache.hashbase[i];
  324                 callout_drain(&sch->sch_timer);
  325 
  326                 SCH_LOCK(sch);
  327                 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc)
  328                         syncache_drop(sc, sch);
  329                 SCH_UNLOCK(sch);
  330                 KASSERT(TAILQ_EMPTY(&sch->sch_bucket),
  331                     ("%s: sch->sch_bucket not empty", __func__));
  332                 KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0",
  333                     __func__, sch->sch_length));
  334                 mtx_destroy(&sch->sch_mtx);
  335         }
  336 
  337         KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0,
  338             ("%s: cache_count not 0", __func__));
  339 
  340         /* Free the allocated global resources. */
  341         uma_zdestroy(V_tcp_syncache.zone);
  342         free(V_tcp_syncache.hashbase, M_SYNCACHE);
  343 }
  344 #endif
  345 
  346 /*
  347  * Inserts a syncache entry into the specified bucket row.
  348  * Locks and unlocks the syncache_head autonomously.
  349  */
  350 static void
  351 syncache_insert(struct syncache *sc, struct syncache_head *sch)
  352 {
  353         struct syncache *sc2;
  354 
  355         SCH_LOCK(sch);
  356 
  357         /*
  358          * Make sure that we don't overflow the per-bucket limit.
  359          * If the bucket is full, toss the oldest element.
  360          */
  361         if (sch->sch_length >= V_tcp_syncache.bucket_limit) {
  362                 KASSERT(!TAILQ_EMPTY(&sch->sch_bucket),
  363                         ("sch->sch_length incorrect"));
  364                 sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head);
  365                 sch->sch_last_overflow = time_uptime;
  366                 syncache_drop(sc2, sch);
  367                 TCPSTAT_INC(tcps_sc_bucketoverflow);
  368         }
  369 
  370         /* Put it into the bucket. */
  371         TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash);
  372         sch->sch_length++;
  373 
  374 #ifdef TCP_OFFLOAD
  375         if (ADDED_BY_TOE(sc)) {
  376                 struct toedev *tod = sc->sc_tod;
  377 
  378                 tod->tod_syncache_added(tod, sc->sc_todctx);
  379         }
  380 #endif
  381 
  382         /* Reinitialize the bucket row's timer. */
  383         if (sch->sch_length == 1)
  384                 sch->sch_nextc = ticks + INT_MAX;
  385         syncache_timeout(sc, sch, 1);
  386 
  387         SCH_UNLOCK(sch);
  388 
  389         TCPSTATES_INC(TCPS_SYN_RECEIVED);
  390         TCPSTAT_INC(tcps_sc_added);
  391 }
  392 
  393 /*
  394  * Remove and free entry from syncache bucket row.
  395  * Expects locked syncache head.
  396  */
  397 static void
  398 syncache_drop(struct syncache *sc, struct syncache_head *sch)
  399 {
  400 
  401         SCH_LOCK_ASSERT(sch);
  402 
  403         TCPSTATES_DEC(TCPS_SYN_RECEIVED);
  404         TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
  405         sch->sch_length--;
  406 
  407 #ifdef TCP_OFFLOAD
  408         if (ADDED_BY_TOE(sc)) {
  409                 struct toedev *tod = sc->sc_tod;
  410 
  411                 tod->tod_syncache_removed(tod, sc->sc_todctx);
  412         }
  413 #endif
  414 
  415         syncache_free(sc);
  416 }
  417 
  418 /*
  419  * Engage/reengage time on bucket row.
  420  */
  421 static void
  422 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout)
  423 {
  424         int rexmt;
  425 
  426         if (sc->sc_rxmits == 0)
  427                 rexmt = tcp_rexmit_initial;
  428         else
  429                 TCPT_RANGESET(rexmt,
  430                     tcp_rexmit_initial * tcp_backoff[sc->sc_rxmits],
  431                     tcp_rexmit_min, TCPTV_REXMTMAX);
  432         sc->sc_rxttime = ticks + rexmt;
  433         sc->sc_rxmits++;
  434         if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) {
  435                 sch->sch_nextc = sc->sc_rxttime;
  436                 if (docallout)
  437                         callout_reset(&sch->sch_timer, sch->sch_nextc - ticks,
  438                             syncache_timer, (void *)sch);
  439         }
  440 }
  441 
  442 /*
  443  * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted.
  444  * If we have retransmitted an entry the maximum number of times, expire it.
  445  * One separate timer for each bucket row.
  446  */
  447 static void
  448 syncache_timer(void *xsch)
  449 {
  450         struct syncache_head *sch = (struct syncache_head *)xsch;
  451         struct syncache *sc, *nsc;
  452         int tick = ticks;
  453         char *s;
  454 
  455         CURVNET_SET(sch->sch_sc->vnet);
  456 
  457         /* NB: syncache_head has already been locked by the callout. */
  458         SCH_LOCK_ASSERT(sch);
  459 
  460         /*
  461          * In the following cycle we may remove some entries and/or
  462          * advance some timeouts, so re-initialize the bucket timer.
  463          */
  464         sch->sch_nextc = tick + INT_MAX;
  465 
  466         TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) {
  467                 /*
  468                  * We do not check if the listen socket still exists
  469                  * and accept the case where the listen socket may be
  470                  * gone by the time we resend the SYN/ACK.  We do
  471                  * not expect this to happens often. If it does,
  472                  * then the RST will be sent by the time the remote
  473                  * host does the SYN/ACK->ACK.
  474                  */
  475                 if (TSTMP_GT(sc->sc_rxttime, tick)) {
  476                         if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc))
  477                                 sch->sch_nextc = sc->sc_rxttime;
  478                         continue;
  479                 }
  480                 if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) {
  481                         if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
  482                                 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, "
  483                                     "giving up and removing syncache entry\n",
  484                                     s, __func__);
  485                                 free(s, M_TCPLOG);
  486                         }
  487                         syncache_drop(sc, sch);
  488                         TCPSTAT_INC(tcps_sc_stale);
  489                         continue;
  490                 }
  491                 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
  492                         log(LOG_DEBUG, "%s; %s: Response timeout, "
  493                             "retransmitting (%u) SYN|ACK\n",
  494                             s, __func__, sc->sc_rxmits);
  495                         free(s, M_TCPLOG);
  496                 }
  497 
  498                 syncache_respond(sc, sch, NULL, TH_SYN|TH_ACK);
  499                 TCPSTAT_INC(tcps_sc_retransmitted);
  500                 syncache_timeout(sc, sch, 0);
  501         }
  502         if (!TAILQ_EMPTY(&(sch)->sch_bucket))
  503                 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick,
  504                         syncache_timer, (void *)(sch));
  505         CURVNET_RESTORE();
  506 }
  507 
  508 /*
  509  * Find an entry in the syncache.
  510  * Returns always with locked syncache_head plus a matching entry or NULL.
  511  */
  512 static struct syncache *
  513 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp)
  514 {
  515         struct syncache *sc;
  516         struct syncache_head *sch;
  517         uint32_t hash;
  518 
  519         /*
  520          * The hash is built on foreign port + local port + foreign address.
  521          * We rely on the fact that struct in_conninfo starts with 16 bits
  522          * of foreign port, then 16 bits of local port then followed by 128
  523          * bits of foreign address.  In case of IPv4 address, the first 3
  524          * 32-bit words of the address always are zeroes.
  525          */
  526         hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5,
  527             V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask;
  528 
  529         sch = &V_tcp_syncache.hashbase[hash];
  530         *schp = sch;
  531         SCH_LOCK(sch);
  532 
  533         /* Circle through bucket row to find matching entry. */
  534         TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash)
  535                 if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie,
  536                     sizeof(struct in_endpoints)) == 0)
  537                         break;
  538 
  539         return (sc);    /* Always returns with locked sch. */
  540 }
  541 
  542 /*
  543  * This function is called when we get a RST for a
  544  * non-existent connection, so that we can see if the
  545  * connection is in the syn cache.  If it is, zap it.
  546  * If required send a challenge ACK.
  547  */
  548 void
  549 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th, struct mbuf *m)
  550 {
  551         struct syncache *sc;
  552         struct syncache_head *sch;
  553         char *s = NULL;
  554 
  555         sc = syncache_lookup(inc, &sch);        /* returns locked sch */
  556         SCH_LOCK_ASSERT(sch);
  557 
  558         /*
  559          * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags.
  560          * See RFC 793 page 65, section SEGMENT ARRIVES.
  561          */
  562         if (th->th_flags & (TH_ACK|TH_SYN|TH_FIN)) {
  563                 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
  564                         log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or "
  565                             "FIN flag set, segment ignored\n", s, __func__);
  566                 TCPSTAT_INC(tcps_badrst);
  567                 goto done;
  568         }
  569 
  570         /*
  571          * No corresponding connection was found in syncache.
  572          * If syncookies are enabled and possibly exclusively
  573          * used, or we are under memory pressure, a valid RST
  574          * may not find a syncache entry.  In that case we're
  575          * done and no SYN|ACK retransmissions will happen.
  576          * Otherwise the RST was misdirected or spoofed.
  577          */
  578         if (sc == NULL) {
  579                 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
  580                         log(LOG_DEBUG, "%s; %s: Spurious RST without matching "
  581                             "syncache entry (possibly syncookie only), "
  582                             "segment ignored\n", s, __func__);
  583                 TCPSTAT_INC(tcps_badrst);
  584                 goto done;
  585         }
  586 
  587         /*
  588          * If the RST bit is set, check the sequence number to see
  589          * if this is a valid reset segment.
  590          *
  591          * RFC 793 page 37:
  592          *   In all states except SYN-SENT, all reset (RST) segments
  593          *   are validated by checking their SEQ-fields.  A reset is
  594          *   valid if its sequence number is in the window.
  595          *
  596          * RFC 793 page 69:
  597          *   There are four cases for the acceptability test for an incoming
  598          *   segment:
  599          *
  600          * Segment Receive  Test
  601          * Length  Window
  602          * ------- -------  -------------------------------------------
  603          *    0       0     SEG.SEQ = RCV.NXT
  604          *    0      >0     RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
  605          *   >0       0     not acceptable
  606          *   >0      >0     RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND
  607          *               or RCV.NXT =< SEG.SEQ+SEG.LEN-1 < RCV.NXT+RCV.WND
  608          *
  609          * Note that when receiving a SYN segment in the LISTEN state,
  610          * IRS is set to SEG.SEQ and RCV.NXT is set to SEG.SEQ+1, as
  611          * described in RFC 793, page 66.
  612          */
  613         if ((SEQ_GEQ(th->th_seq, sc->sc_irs + 1) &&
  614             SEQ_LT(th->th_seq, sc->sc_irs + 1 + sc->sc_wnd)) ||
  615             (sc->sc_wnd == 0 && th->th_seq == sc->sc_irs + 1)) {
  616                 if (V_tcp_insecure_rst ||
  617                     th->th_seq == sc->sc_irs + 1) {
  618                         syncache_drop(sc, sch);
  619                         if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
  620                                 log(LOG_DEBUG,
  621                                     "%s; %s: Our SYN|ACK was rejected, "
  622                                     "connection attempt aborted by remote "
  623                                     "endpoint\n",
  624                                     s, __func__);
  625                         TCPSTAT_INC(tcps_sc_reset);
  626                 } else {
  627                         TCPSTAT_INC(tcps_badrst);
  628                         /* Send challenge ACK. */
  629                         if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
  630                                 log(LOG_DEBUG, "%s; %s: RST with invalid "
  631                                     " SEQ %u != NXT %u (+WND %u), "
  632                                     "sending challenge ACK\n",
  633                                     s, __func__,
  634                                     th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
  635                         syncache_respond(sc, sch, m, TH_ACK);
  636                 }
  637         } else {
  638                 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
  639                         log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != "
  640                             "NXT %u (+WND %u), segment ignored\n",
  641                             s, __func__,
  642                             th->th_seq, sc->sc_irs + 1, sc->sc_wnd);
  643                 TCPSTAT_INC(tcps_badrst);
  644         }
  645 
  646 done:
  647         if (s != NULL)
  648                 free(s, M_TCPLOG);
  649         SCH_UNLOCK(sch);
  650 }
  651 
  652 void
  653 syncache_badack(struct in_conninfo *inc)
  654 {
  655         struct syncache *sc;
  656         struct syncache_head *sch;
  657 
  658         sc = syncache_lookup(inc, &sch);        /* returns locked sch */
  659         SCH_LOCK_ASSERT(sch);
  660         if (sc != NULL) {
  661                 syncache_drop(sc, sch);
  662                 TCPSTAT_INC(tcps_sc_badack);
  663         }
  664         SCH_UNLOCK(sch);
  665 }
  666 
  667 void
  668 syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq)
  669 {
  670         struct syncache *sc;
  671         struct syncache_head *sch;
  672 
  673         sc = syncache_lookup(inc, &sch);        /* returns locked sch */
  674         SCH_LOCK_ASSERT(sch);
  675         if (sc == NULL)
  676                 goto done;
  677 
  678         /* If the sequence number != sc_iss, then it's a bogus ICMP msg */
  679         if (ntohl(th_seq) != sc->sc_iss)
  680                 goto done;
  681 
  682         /*
  683          * If we've rertransmitted 3 times and this is our second error,
  684          * we remove the entry.  Otherwise, we allow it to continue on.
  685          * This prevents us from incorrectly nuking an entry during a
  686          * spurious network outage.
  687          *
  688          * See tcp_notify().
  689          */
  690         if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) {
  691                 sc->sc_flags |= SCF_UNREACH;
  692                 goto done;
  693         }
  694         syncache_drop(sc, sch);
  695         TCPSTAT_INC(tcps_sc_unreach);
  696 done:
  697         SCH_UNLOCK(sch);
  698 }
  699 
  700 /*
  701  * Build a new TCP socket structure from a syncache entry.
  702  *
  703  * On success return the newly created socket with its underlying inp locked.
  704  */
  705 static struct socket *
  706 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m)
  707 {
  708         struct tcp_function_block *blk;
  709         struct inpcb *inp = NULL;
  710         struct socket *so;
  711         struct tcpcb *tp;
  712         int error;
  713         char *s;
  714 
  715         INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
  716 
  717         /*
  718          * Ok, create the full blown connection, and set things up
  719          * as they would have been set up if we had created the
  720          * connection when the SYN arrived.  If we can't create
  721          * the connection, abort it.
  722          */
  723         so = sonewconn(lso, 0);
  724         if (so == NULL) {
  725                 /*
  726                  * Drop the connection; we will either send a RST or
  727                  * have the peer retransmit its SYN again after its
  728                  * RTO and try again.
  729                  */
  730                 TCPSTAT_INC(tcps_listendrop);
  731                 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
  732                         log(LOG_DEBUG, "%s; %s: Socket create failed "
  733                             "due to limits or memory shortage\n",
  734                             s, __func__);
  735                         free(s, M_TCPLOG);
  736                 }
  737                 goto abort2;
  738         }
  739 #ifdef MAC
  740         mac_socketpeer_set_from_mbuf(m, so);
  741 #endif
  742 
  743         inp = sotoinpcb(so);
  744         inp->inp_inc.inc_fibnum = so->so_fibnum;
  745         INP_WLOCK(inp);
  746         /*
  747          * Exclusive pcbinfo lock is not required in syncache socket case even
  748          * if two inpcb locks can be acquired simultaneously:
  749          *  - the inpcb in LISTEN state,
  750          *  - the newly created inp.
  751          *
  752          * In this case, an inp cannot be at same time in LISTEN state and
  753          * just created by an accept() call.
  754          */
  755         INP_HASH_WLOCK(&V_tcbinfo);
  756 
  757         /* Insert new socket into PCB hash list. */
  758         inp->inp_inc.inc_flags = sc->sc_inc.inc_flags;
  759 #ifdef INET6
  760         if (sc->sc_inc.inc_flags & INC_ISIPV6) {
  761                 inp->inp_vflag &= ~INP_IPV4;
  762                 inp->inp_vflag |= INP_IPV6;
  763                 inp->in6p_laddr = sc->sc_inc.inc6_laddr;
  764         } else {
  765                 inp->inp_vflag &= ~INP_IPV6;
  766                 inp->inp_vflag |= INP_IPV4;
  767 #endif
  768                 inp->inp_laddr = sc->sc_inc.inc_laddr;
  769 #ifdef INET6
  770         }
  771 #endif
  772 
  773         /*
  774          * If there's an mbuf and it has a flowid, then let's initialise the
  775          * inp with that particular flowid.
  776          */
  777         if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
  778                 inp->inp_flowid = m->m_pkthdr.flowid;
  779                 inp->inp_flowtype = M_HASHTYPE_GET(m);
  780         }
  781 
  782         /*
  783          * Install in the reservation hash table for now, but don't yet
  784          * install a connection group since the full 4-tuple isn't yet
  785          * configured.
  786          */
  787         inp->inp_lport = sc->sc_inc.inc_lport;
  788         if ((error = in_pcbinshash_nopcbgroup(inp)) != 0) {
  789                 /*
  790                  * Undo the assignments above if we failed to
  791                  * put the PCB on the hash lists.
  792                  */
  793 #ifdef INET6
  794                 if (sc->sc_inc.inc_flags & INC_ISIPV6)
  795                         inp->in6p_laddr = in6addr_any;
  796                 else
  797 #endif
  798                         inp->inp_laddr.s_addr = INADDR_ANY;
  799                 inp->inp_lport = 0;
  800                 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
  801                         log(LOG_DEBUG, "%s; %s: in_pcbinshash failed "
  802                             "with error %i\n",
  803                             s, __func__, error);
  804                         free(s, M_TCPLOG);
  805                 }
  806                 INP_HASH_WUNLOCK(&V_tcbinfo);
  807                 goto abort;
  808         }
  809 #ifdef INET6
  810         if (inp->inp_vflag & INP_IPV6PROTO) {
  811                 struct inpcb *oinp = sotoinpcb(lso);
  812 
  813                 /*
  814                  * Inherit socket options from the listening socket.
  815                  * Note that in6p_inputopts are not (and should not be)
  816                  * copied, since it stores previously received options and is
  817                  * used to detect if each new option is different than the
  818                  * previous one and hence should be passed to a user.
  819                  * If we copied in6p_inputopts, a user would not be able to
  820                  * receive options just after calling the accept system call.
  821                  */
  822                 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS;
  823                 if (oinp->in6p_outputopts)
  824                         inp->in6p_outputopts =
  825                             ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT);
  826         }
  827 
  828         if (sc->sc_inc.inc_flags & INC_ISIPV6) {
  829                 struct in6_addr laddr6;
  830                 struct sockaddr_in6 sin6;
  831 
  832                 sin6.sin6_family = AF_INET6;
  833                 sin6.sin6_len = sizeof(sin6);
  834                 sin6.sin6_addr = sc->sc_inc.inc6_faddr;
  835                 sin6.sin6_port = sc->sc_inc.inc_fport;
  836                 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0;
  837                 laddr6 = inp->in6p_laddr;
  838                 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
  839                         inp->in6p_laddr = sc->sc_inc.inc6_laddr;
  840                 if ((error = in6_pcbconnect_mbuf(inp, (struct sockaddr *)&sin6,
  841                     thread0.td_ucred, m)) != 0) {
  842                         inp->in6p_laddr = laddr6;
  843                         if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
  844                                 log(LOG_DEBUG, "%s; %s: in6_pcbconnect failed "
  845                                     "with error %i\n",
  846                                     s, __func__, error);
  847                                 free(s, M_TCPLOG);
  848                         }
  849                         INP_HASH_WUNLOCK(&V_tcbinfo);
  850                         goto abort;
  851                 }
  852                 /* Override flowlabel from in6_pcbconnect. */
  853                 inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
  854                 inp->inp_flow |= sc->sc_flowlabel;
  855         }
  856 #endif /* INET6 */
  857 #if defined(INET) && defined(INET6)
  858         else
  859 #endif
  860 #ifdef INET
  861         {
  862                 struct in_addr laddr;
  863                 struct sockaddr_in sin;
  864 
  865                 inp->inp_options = (m) ? ip_srcroute(m) : NULL;
  866                 
  867                 if (inp->inp_options == NULL) {
  868                         inp->inp_options = sc->sc_ipopts;
  869                         sc->sc_ipopts = NULL;
  870                 }
  871 
  872                 sin.sin_family = AF_INET;
  873                 sin.sin_len = sizeof(sin);
  874                 sin.sin_addr = sc->sc_inc.inc_faddr;
  875                 sin.sin_port = sc->sc_inc.inc_fport;
  876                 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero));
  877                 laddr = inp->inp_laddr;
  878                 if (inp->inp_laddr.s_addr == INADDR_ANY)
  879                         inp->inp_laddr = sc->sc_inc.inc_laddr;
  880                 if ((error = in_pcbconnect_mbuf(inp, (struct sockaddr *)&sin,
  881                     thread0.td_ucred, m)) != 0) {
  882                         inp->inp_laddr = laddr;
  883                         if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) {
  884                                 log(LOG_DEBUG, "%s; %s: in_pcbconnect failed "
  885                                     "with error %i\n",
  886                                     s, __func__, error);
  887                                 free(s, M_TCPLOG);
  888                         }
  889                         INP_HASH_WUNLOCK(&V_tcbinfo);
  890                         goto abort;
  891                 }
  892         }
  893 #endif /* INET */
  894 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
  895         /* Copy old policy into new socket's. */
  896         if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0)
  897                 printf("syncache_socket: could not copy policy\n");
  898 #endif
  899         INP_HASH_WUNLOCK(&V_tcbinfo);
  900         tp = intotcpcb(inp);
  901         tcp_state_change(tp, TCPS_SYN_RECEIVED);
  902         tp->iss = sc->sc_iss;
  903         tp->irs = sc->sc_irs;
  904         tcp_rcvseqinit(tp);
  905         tcp_sendseqinit(tp);
  906         blk = sototcpcb(lso)->t_fb;
  907         if (V_functions_inherit_listen_socket_stack && blk != tp->t_fb) {
  908                 /*
  909                  * Our parents t_fb was not the default,
  910                  * we need to release our ref on tp->t_fb and 
  911                  * pickup one on the new entry.
  912                  */
  913                 struct tcp_function_block *rblk;
  914                 
  915                 rblk = find_and_ref_tcp_fb(blk);
  916                 KASSERT(rblk != NULL,
  917                     ("cannot find blk %p out of syncache?", blk));
  918                 if (tp->t_fb->tfb_tcp_fb_fini)
  919                         (*tp->t_fb->tfb_tcp_fb_fini)(tp, 0);
  920                 refcount_release(&tp->t_fb->tfb_refcnt);
  921                 tp->t_fb = rblk;
  922                 /*
  923                  * XXXrrs this is quite dangerous, it is possible
  924                  * for the new function to fail to init. We also
  925                  * are not asking if the handoff_is_ok though at
  926                  * the very start thats probalbly ok.
  927                  */
  928                 if (tp->t_fb->tfb_tcp_fb_init) {
  929                         (*tp->t_fb->tfb_tcp_fb_init)(tp);
  930                 }
  931         }               
  932         tp->snd_wl1 = sc->sc_irs;
  933         tp->snd_max = tp->iss + 1;
  934         tp->snd_nxt = tp->iss + 1;
  935         tp->rcv_up = sc->sc_irs + 1;
  936         tp->rcv_wnd = sc->sc_wnd;
  937         tp->rcv_adv += tp->rcv_wnd;
  938         tp->last_ack_sent = tp->rcv_nxt;
  939 
  940         tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY);
  941         if (sc->sc_flags & SCF_NOOPT)
  942                 tp->t_flags |= TF_NOOPT;
  943         else {
  944                 if (sc->sc_flags & SCF_WINSCALE) {
  945                         tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE;
  946                         tp->snd_scale = sc->sc_requested_s_scale;
  947                         tp->request_r_scale = sc->sc_requested_r_scale;
  948                 }
  949                 if (sc->sc_flags & SCF_TIMESTAMP) {
  950                         tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP;
  951                         tp->ts_recent = sc->sc_tsreflect;
  952                         tp->ts_recent_age = tcp_ts_getticks();
  953                         tp->ts_offset = sc->sc_tsoff;
  954                 }
  955 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
  956                 if (sc->sc_flags & SCF_SIGNATURE)
  957                         tp->t_flags |= TF_SIGNATURE;
  958 #endif
  959                 if (sc->sc_flags & SCF_SACK)
  960                         tp->t_flags |= TF_SACK_PERMIT;
  961         }
  962 
  963         if (sc->sc_flags & SCF_ECN)
  964                 tp->t_flags |= TF_ECN_PERMIT;
  965 
  966         /*
  967          * Set up MSS and get cached values from tcp_hostcache.
  968          * This might overwrite some of the defaults we just set.
  969          */
  970         tcp_mss(tp, sc->sc_peer_mss);
  971 
  972         /*
  973          * If the SYN,ACK was retransmitted, indicate that CWND to be
  974          * limited to one segment in cc_conn_init().
  975          * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits.
  976          */
  977         if (sc->sc_rxmits > 1)
  978                 tp->snd_cwnd = 1;
  979 
  980 #ifdef TCP_OFFLOAD
  981         /*
  982          * Allow a TOE driver to install its hooks.  Note that we hold the
  983          * pcbinfo lock too and that prevents tcp_usr_accept from accepting a
  984          * new connection before the TOE driver has done its thing.
  985          */
  986         if (ADDED_BY_TOE(sc)) {
  987                 struct toedev *tod = sc->sc_tod;
  988 
  989                 tod->tod_offload_socket(tod, sc->sc_todctx, so);
  990         }
  991 #endif
  992         /*
  993          * Copy and activate timers.
  994          */
  995         tp->t_keepinit = sototcpcb(lso)->t_keepinit;
  996         tp->t_keepidle = sototcpcb(lso)->t_keepidle;
  997         tp->t_keepintvl = sototcpcb(lso)->t_keepintvl;
  998         tp->t_keepcnt = sototcpcb(lso)->t_keepcnt;
  999         tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp));
 1000 
 1001         TCPSTAT_INC(tcps_accepts);
 1002         return (so);
 1003 
 1004 abort:
 1005         INP_WUNLOCK(inp);
 1006 abort2:
 1007         if (so != NULL)
 1008                 soabort(so);
 1009         return (NULL);
 1010 }
 1011 
 1012 /*
 1013  * This function gets called when we receive an ACK for a
 1014  * socket in the LISTEN state.  We look up the connection
 1015  * in the syncache, and if its there, we pull it out of
 1016  * the cache and turn it into a full-blown connection in
 1017  * the SYN-RECEIVED state.
 1018  *
 1019  * On syncache_socket() success the newly created socket
 1020  * has its underlying inp locked.
 1021  */
 1022 int
 1023 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
 1024     struct socket **lsop, struct mbuf *m)
 1025 {
 1026         struct syncache *sc;
 1027         struct syncache_head *sch;
 1028         struct syncache scs;
 1029         char *s;
 1030 
 1031         /*
 1032          * Global TCP locks are held because we manipulate the PCB lists
 1033          * and create a new socket.
 1034          */
 1035         INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
 1036         KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK,
 1037             ("%s: can handle only ACK", __func__));
 1038 
 1039         sc = syncache_lookup(inc, &sch);        /* returns locked sch */
 1040         SCH_LOCK_ASSERT(sch);
 1041 
 1042 #ifdef INVARIANTS
 1043         /*
 1044          * Test code for syncookies comparing the syncache stored
 1045          * values with the reconstructed values from the cookie.
 1046          */
 1047         if (sc != NULL)
 1048                 syncookie_cmp(inc, sch, sc, th, to, *lsop);
 1049 #endif
 1050 
 1051         if (sc == NULL) {
 1052                 /*
 1053                  * There is no syncache entry, so see if this ACK is
 1054                  * a returning syncookie.  To do this, first:
 1055                  *  A. Check if syncookies are used in case of syncache
 1056                  *     overflows
 1057                  *  B. See if this socket has had a syncache entry dropped in
 1058                  *     the recent past. We don't want to accept a bogus
 1059                  *     syncookie if we've never received a SYN or accept it
 1060                  *     twice.
 1061                  *  C. check that the syncookie is valid.  If it is, then
 1062                  *     cobble up a fake syncache entry, and return.
 1063                  */
 1064                 if (!V_tcp_syncookies) {
 1065                         SCH_UNLOCK(sch);
 1066                         if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
 1067                                 log(LOG_DEBUG, "%s; %s: Spurious ACK, "
 1068                                     "segment rejected (syncookies disabled)\n",
 1069                                     s, __func__);
 1070                         goto failed;
 1071                 }
 1072                 if (!V_tcp_syncookiesonly &&
 1073                     sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) {
 1074                         SCH_UNLOCK(sch);
 1075                         if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
 1076                                 log(LOG_DEBUG, "%s; %s: Spurious ACK, "
 1077                                     "segment rejected (no syncache entry)\n",
 1078                                     s, __func__);
 1079                         goto failed;
 1080                 }
 1081                 bzero(&scs, sizeof(scs));
 1082                 sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop);
 1083                 SCH_UNLOCK(sch);
 1084                 if (sc == NULL) {
 1085                         if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
 1086                                 log(LOG_DEBUG, "%s; %s: Segment failed "
 1087                                     "SYNCOOKIE authentication, segment rejected "
 1088                                     "(probably spoofed)\n", s, __func__);
 1089                         goto failed;
 1090                 }
 1091 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
 1092                 /* If received ACK has MD5 signature, check it. */
 1093                 if ((to->to_flags & TOF_SIGNATURE) != 0 &&
 1094                     (!TCPMD5_ENABLED() ||
 1095                     TCPMD5_INPUT(m, th, to->to_signature) != 0)) {
 1096                         /* Drop the ACK. */
 1097                         if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
 1098                                 log(LOG_DEBUG, "%s; %s: Segment rejected, "
 1099                                     "MD5 signature doesn't match.\n",
 1100                                     s, __func__);
 1101                                 free(s, M_TCPLOG);
 1102                         }
 1103                         TCPSTAT_INC(tcps_sig_err_sigopt);
 1104                         return (-1); /* Do not send RST */
 1105                 }
 1106 #endif /* TCP_SIGNATURE */
 1107         } else {
 1108 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
 1109                 /*
 1110                  * If listening socket requested TCP digests, check that
 1111                  * received ACK has signature and it is correct.
 1112                  * If not, drop the ACK and leave sc entry in th cache,
 1113                  * because SYN was received with correct signature.
 1114                  */
 1115                 if (sc->sc_flags & SCF_SIGNATURE) {
 1116                         if ((to->to_flags & TOF_SIGNATURE) == 0) {
 1117                                 /* No signature */
 1118                                 TCPSTAT_INC(tcps_sig_err_nosigopt);
 1119                                 SCH_UNLOCK(sch);
 1120                                 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
 1121                                         log(LOG_DEBUG, "%s; %s: Segment "
 1122                                             "rejected, MD5 signature wasn't "
 1123                                             "provided.\n", s, __func__);
 1124                                         free(s, M_TCPLOG);
 1125                                 }
 1126                                 return (-1); /* Do not send RST */
 1127                         }
 1128                         if (!TCPMD5_ENABLED() ||
 1129                             TCPMD5_INPUT(m, th, to->to_signature) != 0) {
 1130                                 /* Doesn't match or no SA */
 1131                                 SCH_UNLOCK(sch);
 1132                                 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
 1133                                         log(LOG_DEBUG, "%s; %s: Segment "
 1134                                             "rejected, MD5 signature doesn't "
 1135                                             "match.\n", s, __func__);
 1136                                         free(s, M_TCPLOG);
 1137                                 }
 1138                                 return (-1); /* Do not send RST */
 1139                         }
 1140                 }
 1141 #endif /* TCP_SIGNATURE */
 1142                 /*
 1143                  * Pull out the entry to unlock the bucket row.
 1144                  * 
 1145                  * NOTE: We must decrease TCPS_SYN_RECEIVED count here, not
 1146                  * tcp_state_change().  The tcpcb is not existent at this
 1147                  * moment.  A new one will be allocated via syncache_socket->
 1148                  * sonewconn->tcp_usr_attach in TCPS_CLOSED state, then
 1149                  * syncache_socket() will change it to TCPS_SYN_RECEIVED.
 1150                  */
 1151                 TCPSTATES_DEC(TCPS_SYN_RECEIVED);
 1152                 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash);
 1153                 sch->sch_length--;
 1154 #ifdef TCP_OFFLOAD
 1155                 if (ADDED_BY_TOE(sc)) {
 1156                         struct toedev *tod = sc->sc_tod;
 1157 
 1158                         tod->tod_syncache_removed(tod, sc->sc_todctx);
 1159                 }
 1160 #endif
 1161                 SCH_UNLOCK(sch);
 1162         }
 1163 
 1164         /*
 1165          * Segment validation:
 1166          * ACK must match our initial sequence number + 1 (the SYN|ACK).
 1167          */
 1168         if (th->th_ack != sc->sc_iss + 1) {
 1169                 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
 1170                         log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment "
 1171                             "rejected\n", s, __func__, th->th_ack, sc->sc_iss);
 1172                 goto failed;
 1173         }
 1174 
 1175         /*
 1176          * The SEQ must fall in the window starting at the received
 1177          * initial receive sequence number + 1 (the SYN).
 1178          */
 1179         if (SEQ_LEQ(th->th_seq, sc->sc_irs) ||
 1180             SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) {
 1181                 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
 1182                         log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment "
 1183                             "rejected\n", s, __func__, th->th_seq, sc->sc_irs);
 1184                 goto failed;
 1185         }
 1186 
 1187         /*
 1188          * If timestamps were not negotiated during SYN/ACK they
 1189          * must not appear on any segment during this session.
 1190          */
 1191         if (!(sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) {
 1192                 if ((s = tcp_log_addrs(inc, th, NULL, NULL)))
 1193                         log(LOG_DEBUG, "%s; %s: Timestamp not expected, "
 1194                             "segment rejected\n", s, __func__);
 1195                 goto failed;
 1196         }
 1197 
 1198         /*
 1199          * If timestamps were negotiated during SYN/ACK they should
 1200          * appear on every segment during this session.
 1201          * XXXAO: This is only informal as there have been unverified
 1202          * reports of non-compliants stacks.
 1203          */
 1204         if ((sc->sc_flags & SCF_TIMESTAMP) && !(to->to_flags & TOF_TS)) {
 1205                 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) {
 1206                         log(LOG_DEBUG, "%s; %s: Timestamp missing, "
 1207                             "no action\n", s, __func__);
 1208                         free(s, M_TCPLOG);
 1209                         s = NULL;
 1210                 }
 1211         }
 1212 
 1213         *lsop = syncache_socket(sc, *lsop, m);
 1214 
 1215         if (*lsop == NULL)
 1216                 TCPSTAT_INC(tcps_sc_aborted);
 1217         else
 1218                 TCPSTAT_INC(tcps_sc_completed);
 1219 
 1220 /* how do we find the inp for the new socket? */
 1221         if (sc != &scs)
 1222                 syncache_free(sc);
 1223         return (1);
 1224 failed:
 1225         if (sc != NULL && sc != &scs)
 1226                 syncache_free(sc);
 1227         if (s != NULL)
 1228                 free(s, M_TCPLOG);
 1229         *lsop = NULL;
 1230         return (0);
 1231 }
 1232 
 1233 static void
 1234 syncache_tfo_expand(struct syncache *sc, struct socket **lsop, struct mbuf *m,
 1235     uint64_t response_cookie)
 1236 {
 1237         struct inpcb *inp;
 1238         struct tcpcb *tp;
 1239         unsigned int *pending_counter;
 1240 
 1241         /*
 1242          * Global TCP locks are held because we manipulate the PCB lists
 1243          * and create a new socket.
 1244          */
 1245         INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
 1246 
 1247         pending_counter = intotcpcb(sotoinpcb(*lsop))->t_tfo_pending;
 1248         *lsop = syncache_socket(sc, *lsop, m);
 1249         if (*lsop == NULL) {
 1250                 TCPSTAT_INC(tcps_sc_aborted);
 1251                 atomic_subtract_int(pending_counter, 1);
 1252         } else {
 1253                 soisconnected(*lsop);
 1254                 inp = sotoinpcb(*lsop);
 1255                 tp = intotcpcb(inp);
 1256                 tp->t_flags |= TF_FASTOPEN;
 1257                 tp->t_tfo_cookie.server = response_cookie;
 1258                 tp->snd_max = tp->iss;
 1259                 tp->snd_nxt = tp->iss;
 1260                 tp->t_tfo_pending = pending_counter;
 1261                 TCPSTAT_INC(tcps_sc_completed);
 1262         }
 1263 }
 1264 
 1265 /*
 1266  * Given a LISTEN socket and an inbound SYN request, add
 1267  * this to the syn cache, and send back a segment:
 1268  *      <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK>
 1269  * to the source.
 1270  *
 1271  * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN.
 1272  * Doing so would require that we hold onto the data and deliver it
 1273  * to the application.  However, if we are the target of a SYN-flood
 1274  * DoS attack, an attacker could send data which would eventually
 1275  * consume all available buffer space if it were ACKed.  By not ACKing
 1276  * the data, we avoid this DoS scenario.
 1277  *
 1278  * The exception to the above is when a SYN with a valid TCP Fast Open (TFO)
 1279  * cookie is processed and a new socket is created.  In this case, any data
 1280  * accompanying the SYN will be queued to the socket by tcp_input() and will
 1281  * be ACKed either when the application sends response data or the delayed
 1282  * ACK timer expires, whichever comes first.
 1283  */
 1284 int
 1285 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
 1286     struct inpcb *inp, struct socket **lsop, struct mbuf *m, void *tod,
 1287     void *todctx)
 1288 {
 1289         struct tcpcb *tp;
 1290         struct socket *so;
 1291         struct syncache *sc = NULL;
 1292         struct syncache_head *sch;
 1293         struct mbuf *ipopts = NULL;
 1294         u_int ltflags;
 1295         int win, ip_ttl, ip_tos;
 1296         char *s;
 1297         int rv = 0;
 1298 #ifdef INET6
 1299         int autoflowlabel = 0;
 1300 #endif
 1301 #ifdef MAC
 1302         struct label *maclabel;
 1303 #endif
 1304         struct syncache scs;
 1305         struct ucred *cred;
 1306         uint64_t tfo_response_cookie;
 1307         unsigned int *tfo_pending = NULL;
 1308         int tfo_cookie_valid = 0;
 1309         int tfo_response_cookie_valid = 0;
 1310 
 1311         INP_WLOCK_ASSERT(inp);                  /* listen socket */
 1312         KASSERT((th->th_flags & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN,
 1313             ("%s: unexpected tcp flags", __func__));
 1314 
 1315         /*
 1316          * Combine all so/tp operations very early to drop the INP lock as
 1317          * soon as possible.
 1318          */
 1319         so = *lsop;
 1320         KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so));
 1321         tp = sototcpcb(so);
 1322         cred = crhold(so->so_cred);
 1323 
 1324 #ifdef INET6
 1325         if ((inc->inc_flags & INC_ISIPV6) &&
 1326             (inp->inp_flags & IN6P_AUTOFLOWLABEL))
 1327                 autoflowlabel = 1;
 1328 #endif
 1329         ip_ttl = inp->inp_ip_ttl;
 1330         ip_tos = inp->inp_ip_tos;
 1331         win = so->sol_sbrcv_hiwat;
 1332         ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE));
 1333 
 1334         if (V_tcp_fastopen_server_enable && IS_FASTOPEN(tp->t_flags) &&
 1335             (tp->t_tfo_pending != NULL) &&
 1336             (to->to_flags & TOF_FASTOPEN)) {
 1337                 /*
 1338                  * Limit the number of pending TFO connections to
 1339                  * approximately half of the queue limit.  This prevents TFO
 1340                  * SYN floods from starving the service by filling the
 1341                  * listen queue with bogus TFO connections.
 1342                  */
 1343                 if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <=
 1344                     (so->sol_qlimit / 2)) {
 1345                         int result;
 1346 
 1347                         result = tcp_fastopen_check_cookie(inc,
 1348                             to->to_tfo_cookie, to->to_tfo_len,
 1349                             &tfo_response_cookie);
 1350                         tfo_cookie_valid = (result > 0);
 1351                         tfo_response_cookie_valid = (result >= 0);
 1352                 }
 1353 
 1354                 /*
 1355                  * Remember the TFO pending counter as it will have to be
 1356                  * decremented below if we don't make it to syncache_tfo_expand().
 1357                  */
 1358                 tfo_pending = tp->t_tfo_pending;
 1359         }
 1360 
 1361         /* By the time we drop the lock these should no longer be used. */
 1362         so = NULL;
 1363         tp = NULL;
 1364 
 1365 #ifdef MAC
 1366         if (mac_syncache_init(&maclabel) != 0) {
 1367                 INP_WUNLOCK(inp);
 1368                 goto done;
 1369         } else
 1370                 mac_syncache_create(maclabel, inp);
 1371 #endif
 1372         if (!tfo_cookie_valid)
 1373                 INP_WUNLOCK(inp);
 1374 
 1375         /*
 1376          * Remember the IP options, if any.
 1377          */
 1378 #ifdef INET6
 1379         if (!(inc->inc_flags & INC_ISIPV6))
 1380 #endif
 1381 #ifdef INET
 1382                 ipopts = (m) ? ip_srcroute(m) : NULL;
 1383 #else
 1384                 ipopts = NULL;
 1385 #endif
 1386 
 1387 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
 1388         /*
 1389          * If listening socket requested TCP digests, check that received
 1390          * SYN has signature and it is correct. If signature doesn't match
 1391          * or TCP_SIGNATURE support isn't enabled, drop the packet.
 1392          */
 1393         if (ltflags & TF_SIGNATURE) {
 1394                 if ((to->to_flags & TOF_SIGNATURE) == 0) {
 1395                         TCPSTAT_INC(tcps_sig_err_nosigopt);
 1396                         goto done;
 1397                 }
 1398                 if (!TCPMD5_ENABLED() ||
 1399                     TCPMD5_INPUT(m, th, to->to_signature) != 0)
 1400                         goto done;
 1401         }
 1402 #endif  /* TCP_SIGNATURE */
 1403         /*
 1404          * See if we already have an entry for this connection.
 1405          * If we do, resend the SYN,ACK, and reset the retransmit timer.
 1406          *
 1407          * XXX: should the syncache be re-initialized with the contents
 1408          * of the new SYN here (which may have different options?)
 1409          *
 1410          * XXX: We do not check the sequence number to see if this is a
 1411          * real retransmit or a new connection attempt.  The question is
 1412          * how to handle such a case; either ignore it as spoofed, or
 1413          * drop the current entry and create a new one?
 1414          */
 1415         sc = syncache_lookup(inc, &sch);        /* returns locked entry */
 1416         SCH_LOCK_ASSERT(sch);
 1417         if (sc != NULL) {
 1418                 if (tfo_cookie_valid)
 1419                         INP_WUNLOCK(inp);
 1420                 TCPSTAT_INC(tcps_sc_dupsyn);
 1421                 if (ipopts) {
 1422                         /*
 1423                          * If we were remembering a previous source route,
 1424                          * forget it and use the new one we've been given.
 1425                          */
 1426                         if (sc->sc_ipopts)
 1427                                 (void) m_free(sc->sc_ipopts);
 1428                         sc->sc_ipopts = ipopts;
 1429                 }
 1430                 /*
 1431                  * Update timestamp if present.
 1432                  */
 1433                 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS))
 1434                         sc->sc_tsreflect = to->to_tsval;
 1435                 else
 1436                         sc->sc_flags &= ~SCF_TIMESTAMP;
 1437 #ifdef MAC
 1438                 /*
 1439                  * Since we have already unconditionally allocated label
 1440                  * storage, free it up.  The syncache entry will already
 1441                  * have an initialized label we can use.
 1442                  */
 1443                 mac_syncache_destroy(&maclabel);
 1444 #endif
 1445                 TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
 1446                 /* Retransmit SYN|ACK and reset retransmit count. */
 1447                 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) {
 1448                         log(LOG_DEBUG, "%s; %s: Received duplicate SYN, "
 1449                             "resetting timer and retransmitting SYN|ACK\n",
 1450                             s, __func__);
 1451                         free(s, M_TCPLOG);
 1452                 }
 1453                 if (syncache_respond(sc, sch, m, TH_SYN|TH_ACK) == 0) {
 1454                         sc->sc_rxmits = 0;
 1455                         syncache_timeout(sc, sch, 1);
 1456                         TCPSTAT_INC(tcps_sndacks);
 1457                         TCPSTAT_INC(tcps_sndtotal);
 1458                 }
 1459                 SCH_UNLOCK(sch);
 1460                 goto donenoprobe;
 1461         }
 1462 
 1463         if (tfo_cookie_valid) {
 1464                 bzero(&scs, sizeof(scs));
 1465                 sc = &scs;
 1466                 goto skip_alloc;
 1467         }
 1468 
 1469         sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
 1470         if (sc == NULL) {
 1471                 /*
 1472                  * The zone allocator couldn't provide more entries.
 1473                  * Treat this as if the cache was full; drop the oldest
 1474                  * entry and insert the new one.
 1475                  */
 1476                 TCPSTAT_INC(tcps_sc_zonefail);
 1477                 if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) {
 1478                         sch->sch_last_overflow = time_uptime;
 1479                         syncache_drop(sc, sch);
 1480                 }
 1481                 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO);
 1482                 if (sc == NULL) {
 1483                         if (V_tcp_syncookies) {
 1484                                 bzero(&scs, sizeof(scs));
 1485                                 sc = &scs;
 1486                         } else {
 1487                                 SCH_UNLOCK(sch);
 1488                                 if (ipopts)
 1489                                         (void) m_free(ipopts);
 1490                                 goto done;
 1491                         }
 1492                 }
 1493         }
 1494 
 1495 skip_alloc:
 1496         if (!tfo_cookie_valid && tfo_response_cookie_valid)
 1497                 sc->sc_tfo_cookie = &tfo_response_cookie;
 1498 
 1499         /*
 1500          * Fill in the syncache values.
 1501          */
 1502 #ifdef MAC
 1503         sc->sc_label = maclabel;
 1504 #endif
 1505         sc->sc_cred = cred;
 1506         cred = NULL;
 1507         sc->sc_ipopts = ipopts;
 1508         bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
 1509 #ifdef INET6
 1510         if (!(inc->inc_flags & INC_ISIPV6))
 1511 #endif
 1512         {
 1513                 sc->sc_ip_tos = ip_tos;
 1514                 sc->sc_ip_ttl = ip_ttl;
 1515         }
 1516 #ifdef TCP_OFFLOAD
 1517         sc->sc_tod = tod;
 1518         sc->sc_todctx = todctx;
 1519 #endif
 1520         sc->sc_irs = th->th_seq;
 1521         sc->sc_iss = arc4random();
 1522         sc->sc_flags = 0;
 1523         sc->sc_flowlabel = 0;
 1524 
 1525         /*
 1526          * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN].
 1527          * win was derived from socket earlier in the function.
 1528          */
 1529         win = imax(win, 0);
 1530         win = imin(win, TCP_MAXWIN);
 1531         sc->sc_wnd = win;
 1532 
 1533         if (V_tcp_do_rfc1323) {
 1534                 /*
 1535                  * A timestamp received in a SYN makes
 1536                  * it ok to send timestamp requests and replies.
 1537                  */
 1538                 if (to->to_flags & TOF_TS) {
 1539                         sc->sc_tsreflect = to->to_tsval;
 1540                         sc->sc_flags |= SCF_TIMESTAMP;
 1541                         sc->sc_tsoff = tcp_new_ts_offset(inc);
 1542                 }
 1543                 if (to->to_flags & TOF_SCALE) {
 1544                         int wscale = 0;
 1545 
 1546                         /*
 1547                          * Pick the smallest possible scaling factor that
 1548                          * will still allow us to scale up to sb_max, aka
 1549                          * kern.ipc.maxsockbuf.
 1550                          *
 1551                          * We do this because there are broken firewalls that
 1552                          * will corrupt the window scale option, leading to
 1553                          * the other endpoint believing that our advertised
 1554                          * window is unscaled.  At scale factors larger than
 1555                          * 5 the unscaled window will drop below 1500 bytes,
 1556                          * leading to serious problems when traversing these
 1557                          * broken firewalls.
 1558                          *
 1559                          * With the default maxsockbuf of 256K, a scale factor
 1560                          * of 3 will be chosen by this algorithm.  Those who
 1561                          * choose a larger maxsockbuf should watch out
 1562                          * for the compatibility problems mentioned above.
 1563                          *
 1564                          * RFC1323: The Window field in a SYN (i.e., a <SYN>
 1565                          * or <SYN,ACK>) segment itself is never scaled.
 1566                          */
 1567                         while (wscale < TCP_MAX_WINSHIFT &&
 1568                             (TCP_MAXWIN << wscale) < sb_max)
 1569                                 wscale++;
 1570                         sc->sc_requested_r_scale = wscale;
 1571                         sc->sc_requested_s_scale = to->to_wscale;
 1572                         sc->sc_flags |= SCF_WINSCALE;
 1573                 }
 1574         }
 1575 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
 1576         /*
 1577          * If listening socket requested TCP digests, flag this in the
 1578          * syncache so that syncache_respond() will do the right thing
 1579          * with the SYN+ACK.
 1580          */
 1581         if (ltflags & TF_SIGNATURE)
 1582                 sc->sc_flags |= SCF_SIGNATURE;
 1583 #endif  /* TCP_SIGNATURE */
 1584         if (to->to_flags & TOF_SACKPERM)
 1585                 sc->sc_flags |= SCF_SACK;
 1586         if (to->to_flags & TOF_MSS)
 1587                 sc->sc_peer_mss = to->to_mss;   /* peer mss may be zero */
 1588         if (ltflags & TF_NOOPT)
 1589                 sc->sc_flags |= SCF_NOOPT;
 1590         if ((th->th_flags & (TH_ECE|TH_CWR)) && V_tcp_do_ecn)
 1591                 sc->sc_flags |= SCF_ECN;
 1592 
 1593         if (V_tcp_syncookies)
 1594                 sc->sc_iss = syncookie_generate(sch, sc);
 1595 #ifdef INET6
 1596         if (autoflowlabel) {
 1597                 if (V_tcp_syncookies)
 1598                         sc->sc_flowlabel = sc->sc_iss;
 1599                 else
 1600                         sc->sc_flowlabel = ip6_randomflowlabel();
 1601                 sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK;
 1602         }
 1603 #endif
 1604         SCH_UNLOCK(sch);
 1605 
 1606         if (tfo_cookie_valid) {
 1607                 syncache_tfo_expand(sc, lsop, m, tfo_response_cookie);
 1608                 /* INP_WUNLOCK(inp) will be performed by the caller */
 1609                 rv = 1;
 1610                 goto tfo_expanded;
 1611         }
 1612 
 1613         TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
 1614         /*
 1615          * Do a standard 3-way handshake.
 1616          */
 1617         if (syncache_respond(sc, sch, m, TH_SYN|TH_ACK) == 0) {
 1618                 if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs)
 1619                         syncache_free(sc);
 1620                 else if (sc != &scs)
 1621                         syncache_insert(sc, sch);   /* locks and unlocks sch */
 1622                 TCPSTAT_INC(tcps_sndacks);
 1623                 TCPSTAT_INC(tcps_sndtotal);
 1624         } else {
 1625                 if (sc != &scs)
 1626                         syncache_free(sc);
 1627                 TCPSTAT_INC(tcps_sc_dropped);
 1628         }
 1629         goto donenoprobe;
 1630 
 1631 done:
 1632         TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
 1633 donenoprobe:
 1634         if (m) {
 1635                 *lsop = NULL;
 1636                 m_freem(m);
 1637         }
 1638         /*
 1639          * If tfo_pending is not NULL here, then a TFO SYN that did not
 1640          * result in a new socket was processed and the associated pending
 1641          * counter has not yet been decremented.  All such TFO processing paths
 1642          * transit this point.
 1643          */
 1644         if (tfo_pending != NULL)
 1645                 tcp_fastopen_decrement_counter(tfo_pending);
 1646 
 1647 tfo_expanded:
 1648         if (cred != NULL)
 1649                 crfree(cred);
 1650 #ifdef MAC
 1651         if (sc == &scs)
 1652                 mac_syncache_destroy(&maclabel);
 1653 #endif
 1654         return (rv);
 1655 }
 1656 
 1657 /*
 1658  * Send SYN|ACK or ACK to the peer.  Either in response to a peer's segment,
 1659  * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL.
 1660  */
 1661 static int
 1662 syncache_respond(struct syncache *sc, struct syncache_head *sch,
 1663     const struct mbuf *m0, int flags)
 1664 {
 1665         struct ip *ip = NULL;
 1666         struct mbuf *m;
 1667         struct tcphdr *th = NULL;
 1668         int optlen, error = 0;  /* Make compiler happy */
 1669         u_int16_t hlen, tlen, mssopt;
 1670         struct tcpopt to;
 1671 #ifdef INET6
 1672         struct ip6_hdr *ip6 = NULL;
 1673 #endif
 1674         hlen =
 1675 #ifdef INET6
 1676                (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) :
 1677 #endif
 1678                 sizeof(struct ip);
 1679         tlen = hlen + sizeof(struct tcphdr);
 1680 
 1681         /* Determine MSS we advertize to other end of connection. */
 1682         mssopt = max(tcp_mssopt(&sc->sc_inc), V_tcp_minmss);
 1683 
 1684         /* XXX: Assume that the entire packet will fit in a header mbuf. */
 1685         KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN,
 1686             ("syncache: mbuf too small"));
 1687 
 1688         /* Create the IP+TCP header from scratch. */
 1689         m = m_gethdr(M_NOWAIT, MT_DATA);
 1690         if (m == NULL)
 1691                 return (ENOBUFS);
 1692 #ifdef MAC
 1693         mac_syncache_create_mbuf(sc->sc_label, m);
 1694 #endif
 1695         m->m_data += max_linkhdr;
 1696         m->m_len = tlen;
 1697         m->m_pkthdr.len = tlen;
 1698         m->m_pkthdr.rcvif = NULL;
 1699 
 1700 #ifdef INET6
 1701         if (sc->sc_inc.inc_flags & INC_ISIPV6) {
 1702                 ip6 = mtod(m, struct ip6_hdr *);
 1703                 ip6->ip6_vfc = IPV6_VERSION;
 1704                 ip6->ip6_nxt = IPPROTO_TCP;
 1705                 ip6->ip6_src = sc->sc_inc.inc6_laddr;
 1706                 ip6->ip6_dst = sc->sc_inc.inc6_faddr;
 1707                 ip6->ip6_plen = htons(tlen - hlen);
 1708                 /* ip6_hlim is set after checksum */
 1709                 ip6->ip6_flow &= ~IPV6_FLOWLABEL_MASK;
 1710                 ip6->ip6_flow |= sc->sc_flowlabel;
 1711 
 1712                 th = (struct tcphdr *)(ip6 + 1);
 1713         }
 1714 #endif
 1715 #if defined(INET6) && defined(INET)
 1716         else
 1717 #endif
 1718 #ifdef INET
 1719         {
 1720                 ip = mtod(m, struct ip *);
 1721                 ip->ip_v = IPVERSION;
 1722                 ip->ip_hl = sizeof(struct ip) >> 2;
 1723                 ip->ip_len = htons(tlen);
 1724                 ip->ip_id = 0;
 1725                 ip->ip_off = 0;
 1726                 ip->ip_sum = 0;
 1727                 ip->ip_p = IPPROTO_TCP;
 1728                 ip->ip_src = sc->sc_inc.inc_laddr;
 1729                 ip->ip_dst = sc->sc_inc.inc_faddr;
 1730                 ip->ip_ttl = sc->sc_ip_ttl;
 1731                 ip->ip_tos = sc->sc_ip_tos;
 1732 
 1733                 /*
 1734                  * See if we should do MTU discovery.  Route lookups are
 1735                  * expensive, so we will only unset the DF bit if:
 1736                  *
 1737                  *      1) path_mtu_discovery is disabled
 1738                  *      2) the SCF_UNREACH flag has been set
 1739                  */
 1740                 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0))
 1741                        ip->ip_off |= htons(IP_DF);
 1742 
 1743                 th = (struct tcphdr *)(ip + 1);
 1744         }
 1745 #endif /* INET */
 1746         th->th_sport = sc->sc_inc.inc_lport;
 1747         th->th_dport = sc->sc_inc.inc_fport;
 1748 
 1749         if (flags & TH_SYN)
 1750                 th->th_seq = htonl(sc->sc_iss);
 1751         else
 1752                 th->th_seq = htonl(sc->sc_iss + 1);
 1753         th->th_ack = htonl(sc->sc_irs + 1);
 1754         th->th_off = sizeof(struct tcphdr) >> 2;
 1755         th->th_x2 = 0;
 1756         th->th_flags = flags;
 1757         th->th_win = htons(sc->sc_wnd);
 1758         th->th_urp = 0;
 1759 
 1760         if ((flags & TH_SYN) && (sc->sc_flags & SCF_ECN)) {
 1761                 th->th_flags |= TH_ECE;
 1762                 TCPSTAT_INC(tcps_ecn_shs);
 1763         }
 1764 
 1765         /* Tack on the TCP options. */
 1766         if ((sc->sc_flags & SCF_NOOPT) == 0) {
 1767                 to.to_flags = 0;
 1768 
 1769                 if (flags & TH_SYN) {
 1770                         to.to_mss = mssopt;
 1771                         to.to_flags = TOF_MSS;
 1772                         if (sc->sc_flags & SCF_WINSCALE) {
 1773                                 to.to_wscale = sc->sc_requested_r_scale;
 1774                                 to.to_flags |= TOF_SCALE;
 1775                         }
 1776                         if (sc->sc_flags & SCF_SACK)
 1777                                 to.to_flags |= TOF_SACKPERM;
 1778 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
 1779                         if (sc->sc_flags & SCF_SIGNATURE)
 1780                                 to.to_flags |= TOF_SIGNATURE;
 1781 #endif
 1782                         if (sc->sc_tfo_cookie) {
 1783                                 to.to_flags |= TOF_FASTOPEN;
 1784                                 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
 1785                                 to.to_tfo_cookie = sc->sc_tfo_cookie;
 1786                                 /* don't send cookie again when retransmitting response */
 1787                                 sc->sc_tfo_cookie = NULL;
 1788                         }
 1789                 }
 1790                 if (sc->sc_flags & SCF_TIMESTAMP) {
 1791                         to.to_tsval = sc->sc_tsoff + tcp_ts_getticks();
 1792                         to.to_tsecr = sc->sc_tsreflect;
 1793                         to.to_flags |= TOF_TS;
 1794                 }
 1795                 optlen = tcp_addoptions(&to, (u_char *)(th + 1));
 1796 
 1797                 /* Adjust headers by option size. */
 1798                 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
 1799                 m->m_len += optlen;
 1800                 m->m_pkthdr.len += optlen;
 1801 #ifdef INET6
 1802                 if (sc->sc_inc.inc_flags & INC_ISIPV6)
 1803                         ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen);
 1804                 else
 1805 #endif
 1806                         ip->ip_len = htons(ntohs(ip->ip_len) + optlen);
 1807 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
 1808                 if (sc->sc_flags & SCF_SIGNATURE) {
 1809                         KASSERT(to.to_flags & TOF_SIGNATURE,
 1810                             ("tcp_addoptions() didn't set tcp_signature"));
 1811 
 1812                         /* NOTE: to.to_signature is inside of mbuf */
 1813                         if (!TCPMD5_ENABLED() ||
 1814                             TCPMD5_OUTPUT(m, th, to.to_signature) != 0) {
 1815                                 m_freem(m);
 1816                                 return (EACCES);
 1817                         }
 1818                 }
 1819 #endif
 1820         } else
 1821                 optlen = 0;
 1822 
 1823         M_SETFIB(m, sc->sc_inc.inc_fibnum);
 1824         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
 1825         /*
 1826          * If we have peer's SYN and it has a flowid, then let's assign it to
 1827          * our SYN|ACK.  ip6_output() and ip_output() will not assign flowid
 1828          * to SYN|ACK due to lack of inp here.
 1829          */
 1830         if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) {
 1831                 m->m_pkthdr.flowid = m0->m_pkthdr.flowid;
 1832                 M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0));
 1833         }
 1834 #ifdef INET6
 1835         if (sc->sc_inc.inc_flags & INC_ISIPV6) {
 1836                 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
 1837                 th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen,
 1838                     IPPROTO_TCP, 0);
 1839                 ip6->ip6_hlim = in6_selecthlim(NULL, NULL);
 1840 #ifdef TCP_OFFLOAD
 1841                 if (ADDED_BY_TOE(sc)) {
 1842                         struct toedev *tod = sc->sc_tod;
 1843 
 1844                         error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
 1845 
 1846                         return (error);
 1847                 }
 1848 #endif
 1849                 TCP_PROBE5(send, NULL, NULL, ip6, NULL, th);
 1850                 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL);
 1851         }
 1852 #endif
 1853 #if defined(INET6) && defined(INET)
 1854         else
 1855 #endif
 1856 #ifdef INET
 1857         {
 1858                 m->m_pkthdr.csum_flags = CSUM_TCP;
 1859                 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
 1860                     htons(tlen + optlen - hlen + IPPROTO_TCP));
 1861 #ifdef TCP_OFFLOAD
 1862                 if (ADDED_BY_TOE(sc)) {
 1863                         struct toedev *tod = sc->sc_tod;
 1864 
 1865                         error = tod->tod_syncache_respond(tod, sc->sc_todctx, m);
 1866 
 1867                         return (error);
 1868                 }
 1869 #endif
 1870                 TCP_PROBE5(send, NULL, NULL, ip, NULL, th);
 1871                 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL);
 1872         }
 1873 #endif
 1874         return (error);
 1875 }
 1876 
 1877 /*
 1878  * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks
 1879  * that exceed the capacity of the syncache by avoiding the storage of any
 1880  * of the SYNs we receive.  Syncookies defend against blind SYN flooding
 1881  * attacks where the attacker does not have access to our responses.
 1882  *
 1883  * Syncookies encode and include all necessary information about the
 1884  * connection setup within the SYN|ACK that we send back.  That way we
 1885  * can avoid keeping any local state until the ACK to our SYN|ACK returns
 1886  * (if ever).  Normally the syncache and syncookies are running in parallel
 1887  * with the latter taking over when the former is exhausted.  When matching
 1888  * syncache entry is found the syncookie is ignored.
 1889  *
 1890  * The only reliable information persisting the 3WHS is our initial sequence
 1891  * number ISS of 32 bits.  Syncookies embed a cryptographically sufficient
 1892  * strong hash (MAC) value and a few bits of TCP SYN options in the ISS
 1893  * of our SYN|ACK.  The MAC can be recomputed when the ACK to our SYN|ACK
 1894  * returns and signifies a legitimate connection if it matches the ACK.
 1895  *
 1896  * The available space of 32 bits to store the hash and to encode the SYN
 1897  * option information is very tight and we should have at least 24 bits for
 1898  * the MAC to keep the number of guesses by blind spoofing reasonably high.
 1899  *
 1900  * SYN option information we have to encode to fully restore a connection:
 1901  * MSS: is imporant to chose an optimal segment size to avoid IP level
 1902  *   fragmentation along the path.  The common MSS values can be encoded
 1903  *   in a 3-bit table.  Uncommon values are captured by the next lower value
 1904  *   in the table leading to a slight increase in packetization overhead.
 1905  * WSCALE: is necessary to allow large windows to be used for high delay-
 1906  *   bandwidth product links.  Not scaling the window when it was initially
 1907  *   negotiated is bad for performance as lack of scaling further decreases
 1908  *   the apparent available send window.  We only need to encode the WSCALE
 1909  *   we received from the remote end.  Our end can be recalculated at any
 1910  *   time.  The common WSCALE values can be encoded in a 3-bit table.
 1911  *   Uncommon values are captured by the next lower value in the table
 1912  *   making us under-estimate the available window size halving our
 1913  *   theoretically possible maximum throughput for that connection.
 1914  * SACK: Greatly assists in packet loss recovery and requires 1 bit.
 1915  * TIMESTAMP and SIGNATURE is not encoded because they are permanent options
 1916  *   that are included in all segments on a connection.  We enable them when
 1917  *   the ACK has them.
 1918  *
 1919  * Security of syncookies and attack vectors:
 1920  *
 1921  * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod)
 1922  * together with the gloabl secret to make it unique per connection attempt.
 1923  * Thus any change of any of those parameters results in a different MAC output
 1924  * in an unpredictable way unless a collision is encountered.  24 bits of the
 1925  * MAC are embedded into the ISS.
 1926  *
 1927  * To prevent replay attacks two rotating global secrets are updated with a
 1928  * new random value every 15 seconds.  The life-time of a syncookie is thus
 1929  * 15-30 seconds.
 1930  *
 1931  * Vector 1: Attacking the secret.  This requires finding a weakness in the
 1932  * MAC itself or the way it is used here.  The attacker can do a chosen plain
 1933  * text attack by varying and testing the all parameters under his control.
 1934  * The strength depends on the size and randomness of the secret, and the
 1935  * cryptographic security of the MAC function.  Due to the constant updating
 1936  * of the secret the attacker has at most 29.999 seconds to find the secret
 1937  * and launch spoofed connections.  After that he has to start all over again.
 1938  *
 1939  * Vector 2: Collision attack on the MAC of a single ACK.  With a 24 bit MAC
 1940  * size an average of 4,823 attempts are required for a 50% chance of success
 1941  * to spoof a single syncookie (birthday collision paradox).  However the
 1942  * attacker is blind and doesn't know if one of his attempts succeeded unless
 1943  * he has a side channel to interfere success from.  A single connection setup
 1944  * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets.
 1945  * This many attempts are required for each one blind spoofed connection.  For
 1946  * every additional spoofed connection he has to launch another N attempts.
 1947  * Thus for a sustained rate 100 spoofed connections per second approximately
 1948  * 1,800,000 packets per second would have to be sent.
 1949  *
 1950  * NB: The MAC function should be fast so that it doesn't become a CPU
 1951  * exhaustion attack vector itself.
 1952  *
 1953  * References:
 1954  *  RFC4987 TCP SYN Flooding Attacks and Common Mitigations
 1955  *  SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996
 1956  *   http://cr.yp.to/syncookies.html    (overview)
 1957  *   http://cr.yp.to/syncookies/archive (details)
 1958  *
 1959  *
 1960  * Schematic construction of a syncookie enabled Initial Sequence Number:
 1961  *  0        1         2         3
 1962  *  12345678901234567890123456789012
 1963  * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP|
 1964  *
 1965  *  x 24 MAC (truncated)
 1966  *  W  3 Send Window Scale index
 1967  *  M  3 MSS index
 1968  *  S  1 SACK permitted
 1969  *  P  1 Odd/even secret
 1970  */
 1971 
 1972 /*
 1973  * Distribution and probability of certain MSS values.  Those in between are
 1974  * rounded down to the next lower one.
 1975  * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011]
 1976  *                            .2%  .3%   5%    7%    7%    20%   15%   45%
 1977  */
 1978 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 };
 1979 
 1980 /*
 1981  * Distribution and probability of certain WSCALE values.  We have to map the
 1982  * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3
 1983  * bits based on prevalence of certain values.  Where we don't have an exact
 1984  * match for are rounded down to the next lower one letting us under-estimate
 1985  * the true available window.  At the moment this would happen only for the
 1986  * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer
 1987  * and window size).  The absence of the WSCALE option (no scaling in either
 1988  * direction) is encoded with index zero.
 1989  * [WSCALE values histograms, Allman, 2012]
 1990  *                            X 10 10 35  5  6 14 10%   by host
 1991  *                            X 11  4  5  5 18 49  3%   by connections
 1992  */
 1993 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 };
 1994 
 1995 /*
 1996  * Compute the MAC for the SYN cookie.  SIPHASH-2-4 is chosen for its speed
 1997  * and good cryptographic properties.
 1998  */
 1999 static uint32_t
 2000 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags,
 2001     uint8_t *secbits, uintptr_t secmod)
 2002 {
 2003         SIPHASH_CTX ctx;
 2004         uint32_t siphash[2];
 2005 
 2006         SipHash24_Init(&ctx);
 2007         SipHash_SetKey(&ctx, secbits);
 2008         switch (inc->inc_flags & INC_ISIPV6) {
 2009 #ifdef INET
 2010         case 0:
 2011                 SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr));
 2012                 SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr));
 2013                 break;
 2014 #endif
 2015 #ifdef INET6
 2016         case INC_ISIPV6:
 2017                 SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr));
 2018                 SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr));
 2019                 break;
 2020 #endif
 2021         }
 2022         SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport));
 2023         SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport));
 2024         SipHash_Update(&ctx, &irs, sizeof(irs));
 2025         SipHash_Update(&ctx, &flags, sizeof(flags));
 2026         SipHash_Update(&ctx, &secmod, sizeof(secmod));
 2027         SipHash_Final((u_int8_t *)&siphash, &ctx);
 2028 
 2029         return (siphash[0] ^ siphash[1]);
 2030 }
 2031 
 2032 static tcp_seq
 2033 syncookie_generate(struct syncache_head *sch, struct syncache *sc)
 2034 {
 2035         u_int i, secbit, wscale;
 2036         uint32_t iss, hash;
 2037         uint8_t *secbits;
 2038         union syncookie cookie;
 2039 
 2040         SCH_LOCK_ASSERT(sch);
 2041 
 2042         cookie.cookie = 0;
 2043 
 2044         /* Map our computed MSS into the 3-bit index. */
 2045         for (i = nitems(tcp_sc_msstab) - 1;
 2046              tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0;
 2047              i--)
 2048                 ;
 2049         cookie.flags.mss_idx = i;
 2050 
 2051         /*
 2052          * Map the send window scale into the 3-bit index but only if
 2053          * the wscale option was received.
 2054          */
 2055         if (sc->sc_flags & SCF_WINSCALE) {
 2056                 wscale = sc->sc_requested_s_scale;
 2057                 for (i = nitems(tcp_sc_wstab) - 1;
 2058                     tcp_sc_wstab[i] > wscale && i > 0;
 2059                      i--)
 2060                         ;
 2061                 cookie.flags.wscale_idx = i;
 2062         }
 2063 
 2064         /* Can we do SACK? */
 2065         if (sc->sc_flags & SCF_SACK)
 2066                 cookie.flags.sack_ok = 1;
 2067 
 2068         /* Which of the two secrets to use. */
 2069         secbit = sch->sch_sc->secret.oddeven & 0x1;
 2070         cookie.flags.odd_even = secbit;
 2071 
 2072         secbits = sch->sch_sc->secret.key[secbit];
 2073         hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits,
 2074             (uintptr_t)sch);
 2075 
 2076         /*
 2077          * Put the flags into the hash and XOR them to get better ISS number
 2078          * variance.  This doesn't enhance the cryptographic strength and is
 2079          * done to prevent the 8 cookie bits from showing up directly on the
 2080          * wire.
 2081          */
 2082         iss = hash & ~0xff;
 2083         iss |= cookie.cookie ^ (hash >> 24);
 2084 
 2085         TCPSTAT_INC(tcps_sc_sendcookie);
 2086         return (iss);
 2087 }
 2088 
 2089 static struct syncache *
 2090 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch, 
 2091     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
 2092     struct socket *lso)
 2093 {
 2094         uint32_t hash;
 2095         uint8_t *secbits;
 2096         tcp_seq ack, seq;
 2097         int wnd, wscale = 0;
 2098         union syncookie cookie;
 2099 
 2100         SCH_LOCK_ASSERT(sch);
 2101 
 2102         /*
 2103          * Pull information out of SYN-ACK/ACK and revert sequence number
 2104          * advances.
 2105          */
 2106         ack = th->th_ack - 1;
 2107         seq = th->th_seq - 1;
 2108 
 2109         /*
 2110          * Unpack the flags containing enough information to restore the
 2111          * connection.
 2112          */
 2113         cookie.cookie = (ack & 0xff) ^ (ack >> 24);
 2114 
 2115         /* Which of the two secrets to use. */
 2116         secbits = sch->sch_sc->secret.key[cookie.flags.odd_even];
 2117 
 2118         hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch);
 2119 
 2120         /* The recomputed hash matches the ACK if this was a genuine cookie. */
 2121         if ((ack & ~0xff) != (hash & ~0xff))
 2122                 return (NULL);
 2123 
 2124         /* Fill in the syncache values. */
 2125         sc->sc_flags = 0;
 2126         bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo));
 2127         sc->sc_ipopts = NULL;
 2128         
 2129         sc->sc_irs = seq;
 2130         sc->sc_iss = ack;
 2131 
 2132         switch (inc->inc_flags & INC_ISIPV6) {
 2133 #ifdef INET
 2134         case 0:
 2135                 sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl;
 2136                 sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos;
 2137                 break;
 2138 #endif
 2139 #ifdef INET6
 2140         case INC_ISIPV6:
 2141                 if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL)
 2142                         sc->sc_flowlabel = sc->sc_iss & IPV6_FLOWLABEL_MASK;
 2143                 break;
 2144 #endif
 2145         }
 2146 
 2147         sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx];
 2148 
 2149         /* We can simply recompute receive window scale we sent earlier. */
 2150         while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max)
 2151                 wscale++;
 2152 
 2153         /* Only use wscale if it was enabled in the orignal SYN. */
 2154         if (cookie.flags.wscale_idx > 0) {
 2155                 sc->sc_requested_r_scale = wscale;
 2156                 sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx];
 2157                 sc->sc_flags |= SCF_WINSCALE;
 2158         }
 2159 
 2160         wnd = lso->sol_sbrcv_hiwat;
 2161         wnd = imax(wnd, 0);
 2162         wnd = imin(wnd, TCP_MAXWIN);
 2163         sc->sc_wnd = wnd;
 2164 
 2165         if (cookie.flags.sack_ok)
 2166                 sc->sc_flags |= SCF_SACK;
 2167 
 2168         if (to->to_flags & TOF_TS) {
 2169                 sc->sc_flags |= SCF_TIMESTAMP;
 2170                 sc->sc_tsreflect = to->to_tsval;
 2171                 sc->sc_tsoff = tcp_new_ts_offset(inc);
 2172         }
 2173 
 2174         if (to->to_flags & TOF_SIGNATURE)
 2175                 sc->sc_flags |= SCF_SIGNATURE;
 2176 
 2177         sc->sc_rxmits = 0;
 2178 
 2179         TCPSTAT_INC(tcps_sc_recvcookie);
 2180         return (sc);
 2181 }
 2182 
 2183 #ifdef INVARIANTS
 2184 static int
 2185 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch,
 2186     struct syncache *sc, struct tcphdr *th, struct tcpopt *to,
 2187     struct socket *lso)
 2188 {
 2189         struct syncache scs, *scx;
 2190         char *s;
 2191 
 2192         bzero(&scs, sizeof(scs));
 2193         scx = syncookie_lookup(inc, sch, &scs, th, to, lso);
 2194 
 2195         if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL)
 2196                 return (0);
 2197 
 2198         if (scx != NULL) {
 2199                 if (sc->sc_peer_mss != scx->sc_peer_mss)
 2200                         log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n",
 2201                             s, __func__, sc->sc_peer_mss, scx->sc_peer_mss);
 2202 
 2203                 if (sc->sc_requested_r_scale != scx->sc_requested_r_scale)
 2204                         log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n",
 2205                             s, __func__, sc->sc_requested_r_scale,
 2206                             scx->sc_requested_r_scale);
 2207 
 2208                 if (sc->sc_requested_s_scale != scx->sc_requested_s_scale)
 2209                         log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n",
 2210                             s, __func__, sc->sc_requested_s_scale,
 2211                             scx->sc_requested_s_scale);
 2212 
 2213                 if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK))
 2214                         log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__);
 2215         }
 2216 
 2217         if (s != NULL)
 2218                 free(s, M_TCPLOG);
 2219         return (0);
 2220 }
 2221 #endif /* INVARIANTS */
 2222 
 2223 static void
 2224 syncookie_reseed(void *arg)
 2225 {
 2226         struct tcp_syncache *sc = arg;
 2227         uint8_t *secbits;
 2228         int secbit;
 2229 
 2230         /*
 2231          * Reseeding the secret doesn't have to be protected by a lock.
 2232          * It only must be ensured that the new random values are visible
 2233          * to all CPUs in a SMP environment.  The atomic with release
 2234          * semantics ensures that.
 2235          */
 2236         secbit = (sc->secret.oddeven & 0x1) ? 0 : 1;
 2237         secbits = sc->secret.key[secbit];
 2238         arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0);
 2239         atomic_add_rel_int(&sc->secret.oddeven, 1);
 2240 
 2241         /* Reschedule ourself. */
 2242         callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz);
 2243 }
 2244 
 2245 /*
 2246  * Exports the syncache entries to userland so that netstat can display
 2247  * them alongside the other sockets.  This function is intended to be
 2248  * called only from tcp_pcblist.
 2249  *
 2250  * Due to concurrency on an active system, the number of pcbs exported
 2251  * may have no relation to max_pcbs.  max_pcbs merely indicates the
 2252  * amount of space the caller allocated for this function to use.
 2253  */
 2254 int
 2255 syncache_pcblist(struct sysctl_req *req, int max_pcbs, int *pcbs_exported)
 2256 {
 2257         struct xtcpcb xt;
 2258         struct syncache *sc;
 2259         struct syncache_head *sch;
 2260         int count, error, i;
 2261 
 2262         for (count = 0, error = 0, i = 0; i < V_tcp_syncache.hashsize; i++) {
 2263                 sch = &V_tcp_syncache.hashbase[i];
 2264                 SCH_LOCK(sch);
 2265                 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) {
 2266                         if (count >= max_pcbs) {
 2267                                 SCH_UNLOCK(sch);
 2268                                 goto exit;
 2269                         }
 2270                         if (cr_cansee(req->td->td_ucred, sc->sc_cred) != 0)
 2271                                 continue;
 2272                         bzero(&xt, sizeof(xt));
 2273                         xt.xt_len = sizeof(xt);
 2274                         if (sc->sc_inc.inc_flags & INC_ISIPV6)
 2275                                 xt.xt_inp.inp_vflag = INP_IPV6;
 2276                         else
 2277                                 xt.xt_inp.inp_vflag = INP_IPV4;
 2278                         bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc,
 2279                             sizeof (struct in_conninfo));
 2280                         xt.t_state = TCPS_SYN_RECEIVED;
 2281                         xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP;
 2282                         xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket);
 2283                         xt.xt_inp.xi_socket.so_type = SOCK_STREAM;
 2284                         xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING;
 2285                         error = SYSCTL_OUT(req, &xt, sizeof xt);
 2286                         if (error) {
 2287                                 SCH_UNLOCK(sch);
 2288                                 goto exit;
 2289                         }
 2290                         count++;
 2291                 }
 2292                 SCH_UNLOCK(sch);
 2293         }
 2294 exit:
 2295         *pcbs_exported = count;
 2296         return error;
 2297 }

Cache object: 669079f9742b2d4a9d31e13e8b7873a1


[ 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.