]> CyberLeo.Net >> Repos - FreeBSD/FreeBSD.git/blob - sys/netinet6/in6_pcb.c
Merge commit '93bf91b4012a28610672d2266366dfa0a663b70f' into HEAD
[FreeBSD/FreeBSD.git] / sys / netinet6 / in6_pcb.c
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (C) 1995, 1996, 1997, and 1998 WIDE Project.
5  * Copyright (c) 2010-2011 Juniper Networks, Inc.
6  * All rights reserved.
7  *
8  * Portions of this software were developed by Robert N. M. Watson under
9  * contract to Juniper Networks, Inc.
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  * 3. Neither the name of the project nor the names of its contributors
20  *    may be used to endorse or promote products derived from this software
21  *    without specific prior written permission.
22  *
23  * THIS SOFTWARE IS PROVIDED BY THE PROJECT AND CONTRIBUTORS ``AS IS'' AND
24  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
25  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
26  * ARE DISCLAIMED.  IN NO EVENT SHALL THE PROJECT OR CONTRIBUTORS BE LIABLE
27  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
28  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
29  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
30  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
31  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
32  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33  * SUCH DAMAGE.
34  *
35  *      $KAME: in6_pcb.c,v 1.31 2001/05/21 05:45:10 jinmei Exp $
36  */
37
38 /*-
39  * Copyright (c) 1982, 1986, 1991, 1993
40  *      The Regents of the University of California.  All rights reserved.
41  *
42  * Redistribution and use in source and binary forms, with or without
43  * modification, are permitted provided that the following conditions
44  * are met:
45  * 1. Redistributions of source code must retain the above copyright
46  *    notice, this list of conditions and the following disclaimer.
47  * 2. Redistributions in binary form must reproduce the above copyright
48  *    notice, this list of conditions and the following disclaimer in the
49  *    documentation and/or other materials provided with the distribution.
50  * 3. Neither the name of the University nor the names of its contributors
51  *    may be used to endorse or promote products derived from this software
52  *    without specific prior written permission.
53  *
54  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
55  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
56  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
57  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
58  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
59  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
60  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
61  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
62  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
63  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
64  * SUCH DAMAGE.
65  *
66  *      @(#)in_pcb.c    8.2 (Berkeley) 1/4/94
67  */
68
69 #include <sys/cdefs.h>
70 __FBSDID("$FreeBSD$");
71
72 #include "opt_inet.h"
73 #include "opt_inet6.h"
74 #include "opt_ipsec.h"
75 #include "opt_route.h"
76 #include "opt_rss.h"
77
78 #include <sys/hash.h>
79 #include <sys/param.h>
80 #include <sys/systm.h>
81 #include <sys/malloc.h>
82 #include <sys/mbuf.h>
83 #include <sys/domain.h>
84 #include <sys/protosw.h>
85 #include <sys/socket.h>
86 #include <sys/socketvar.h>
87 #include <sys/sockio.h>
88 #include <sys/errno.h>
89 #include <sys/time.h>
90 #include <sys/priv.h>
91 #include <sys/proc.h>
92 #include <sys/jail.h>
93
94 #include <vm/uma.h>
95
96 #include <net/if.h>
97 #include <net/if_var.h>
98 #include <net/if_llatbl.h>
99 #include <net/if_types.h>
100 #include <net/route.h>
101 #include <net/route/nhop.h>
102
103 #include <netinet/in.h>
104 #include <netinet/in_var.h>
105 #include <netinet/in_systm.h>
106 #include <netinet/tcp_var.h>
107 #include <netinet/ip6.h>
108 #include <netinet/ip_var.h>
109
110 #include <netinet6/ip6_var.h>
111 #include <netinet6/nd6.h>
112 #include <netinet/in_pcb.h>
113 #include <netinet/in_pcb_var.h>
114 #include <netinet6/in6_pcb.h>
115 #include <netinet6/in6_fib.h>
116 #include <netinet6/scope6_var.h>
117
118 int
119 in6_pcbsetport(struct in6_addr *laddr, struct inpcb *inp, struct ucred *cred)
120 {
121         struct socket *so = inp->inp_socket;
122         u_int16_t lport = 0;
123         int error, lookupflags = 0;
124 #ifdef INVARIANTS
125         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
126 #endif
127
128         INP_WLOCK_ASSERT(inp);
129         INP_HASH_WLOCK_ASSERT(pcbinfo);
130
131         error = prison_local_ip6(cred, laddr,
132             ((inp->inp_flags & IN6P_IPV6_V6ONLY) != 0));
133         if (error)
134                 return(error);
135
136         /* XXX: this is redundant when called from in6_pcbbind */
137         if ((so->so_options & (SO_REUSEADDR|SO_REUSEPORT|SO_REUSEPORT_LB)) == 0)
138                 lookupflags = INPLOOKUP_WILDCARD;
139
140         inp->inp_flags |= INP_ANONPORT;
141
142         error = in_pcb_lport(inp, NULL, &lport, cred, lookupflags);
143         if (error != 0)
144                 return (error);
145
146         inp->inp_lport = lport;
147         if (in_pcbinshash(inp) != 0) {
148                 inp->in6p_laddr = in6addr_any;
149                 inp->inp_lport = 0;
150                 return (EAGAIN);
151         }
152
153         return (0);
154 }
155
156 int
157 in6_pcbbind(struct inpcb *inp, struct sockaddr *nam,
158     struct ucred *cred)
159 {
160         struct socket *so = inp->inp_socket;
161         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)NULL;
162         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
163         u_short lport = 0;
164         int error, lookupflags = 0;
165         int reuseport = (so->so_options & SO_REUSEPORT);
166
167         /*
168          * XXX: Maybe we could let SO_REUSEPORT_LB set SO_REUSEPORT bit here
169          * so that we don't have to add to the (already messy) code below.
170          */
171         int reuseport_lb = (so->so_options & SO_REUSEPORT_LB);
172
173         INP_WLOCK_ASSERT(inp);
174         INP_HASH_WLOCK_ASSERT(pcbinfo);
175
176         if (inp->inp_lport || !IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
177                 return (EINVAL);
178         if ((so->so_options & (SO_REUSEADDR|SO_REUSEPORT|SO_REUSEPORT_LB)) == 0)
179                 lookupflags = INPLOOKUP_WILDCARD;
180         if (nam == NULL) {
181                 if ((error = prison_local_ip6(cred, &inp->in6p_laddr,
182                     ((inp->inp_flags & IN6P_IPV6_V6ONLY) != 0))) != 0)
183                         return (error);
184         } else {
185                 sin6 = (struct sockaddr_in6 *)nam;
186                 KASSERT(sin6->sin6_family == AF_INET6,
187                     ("%s: invalid address family for %p", __func__, sin6));
188                 KASSERT(sin6->sin6_len == sizeof(*sin6),
189                     ("%s: invalid address length for %p", __func__, sin6));
190
191                 if ((error = sa6_embedscope(sin6, V_ip6_use_defzone)) != 0)
192                         return(error);
193
194                 if ((error = prison_local_ip6(cred, &sin6->sin6_addr,
195                     ((inp->inp_flags & IN6P_IPV6_V6ONLY) != 0))) != 0)
196                         return (error);
197
198                 lport = sin6->sin6_port;
199                 if (IN6_IS_ADDR_MULTICAST(&sin6->sin6_addr)) {
200                         /*
201                          * Treat SO_REUSEADDR as SO_REUSEPORT for multicast;
202                          * allow compepte duplication of binding if
203                          * SO_REUSEPORT is set, or if SO_REUSEADDR is set
204                          * and a multicast address is bound on both
205                          * new and duplicated sockets.
206                          */
207                         if ((so->so_options & (SO_REUSEADDR|SO_REUSEPORT)) != 0)
208                                 reuseport = SO_REUSEADDR|SO_REUSEPORT;
209                         /*
210                          * XXX: How to deal with SO_REUSEPORT_LB here?
211                          * Treat same as SO_REUSEPORT for now.
212                          */
213                         if ((so->so_options &
214                             (SO_REUSEADDR|SO_REUSEPORT_LB)) != 0)
215                                 reuseport_lb = SO_REUSEADDR|SO_REUSEPORT_LB;
216                 } else if (!IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
217                         struct epoch_tracker et;
218                         struct ifaddr *ifa;
219
220                         sin6->sin6_port = 0;            /* yech... */
221                         NET_EPOCH_ENTER(et);
222                         if ((ifa = ifa_ifwithaddr((struct sockaddr *)sin6)) ==
223                             NULL &&
224                             (inp->inp_flags & INP_BINDANY) == 0) {
225                                 NET_EPOCH_EXIT(et);
226                                 return (EADDRNOTAVAIL);
227                         }
228
229                         /*
230                          * XXX: bind to an anycast address might accidentally
231                          * cause sending a packet with anycast source address.
232                          * We should allow to bind to a deprecated address, since
233                          * the application dares to use it.
234                          */
235                         if (ifa != NULL &&
236                             ((struct in6_ifaddr *)ifa)->ia6_flags &
237                             (IN6_IFF_ANYCAST|IN6_IFF_NOTREADY|IN6_IFF_DETACHED)) {
238                                 NET_EPOCH_EXIT(et);
239                                 return (EADDRNOTAVAIL);
240                         }
241                         NET_EPOCH_EXIT(et);
242                 }
243                 if (lport) {
244                         struct inpcb *t;
245
246                         /* GROSS */
247                         if (ntohs(lport) <= V_ipport_reservedhigh &&
248                             ntohs(lport) >= V_ipport_reservedlow &&
249                             priv_check_cred(cred, PRIV_NETINET_RESERVEDPORT))
250                                 return (EACCES);
251                         if (!IN6_IS_ADDR_MULTICAST(&sin6->sin6_addr) &&
252                             priv_check_cred(inp->inp_cred, PRIV_NETINET_REUSEPORT) != 0) {
253                                 t = in6_pcblookup_local(pcbinfo,
254                                     &sin6->sin6_addr, lport,
255                                     INPLOOKUP_WILDCARD, cred);
256                                 if (t &&
257                                     ((inp->inp_flags2 & INP_BINDMULTI) == 0) &&
258                                     (so->so_type != SOCK_STREAM ||
259                                      IN6_IS_ADDR_UNSPECIFIED(&t->in6p_faddr)) &&
260                                     (!IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr) ||
261                                      !IN6_IS_ADDR_UNSPECIFIED(&t->in6p_laddr) ||
262                                      (t->inp_flags2 & INP_REUSEPORT) ||
263                                      (t->inp_flags2 & INP_REUSEPORT_LB) == 0) &&
264                                     (inp->inp_cred->cr_uid !=
265                                      t->inp_cred->cr_uid))
266                                         return (EADDRINUSE);
267
268                                 /*
269                                  * If the socket is a BINDMULTI socket, then
270                                  * the credentials need to match and the
271                                  * original socket also has to have been bound
272                                  * with BINDMULTI.
273                                  */
274                                 if (t && (! in_pcbbind_check_bindmulti(inp, t)))
275                                         return (EADDRINUSE);
276
277 #ifdef INET
278                                 if ((inp->inp_flags & IN6P_IPV6_V6ONLY) == 0 &&
279                                     IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
280                                         struct sockaddr_in sin;
281
282                                         in6_sin6_2_sin(&sin, sin6);
283                                         t = in_pcblookup_local(pcbinfo,
284                                             sin.sin_addr, lport,
285                                             INPLOOKUP_WILDCARD, cred);
286                                         if (t &&
287                                             ((inp->inp_flags2 & INP_BINDMULTI) == 0) &&
288                                             (so->so_type != SOCK_STREAM ||
289                                              ntohl(t->inp_faddr.s_addr) ==
290                                               INADDR_ANY) &&
291                                             (inp->inp_cred->cr_uid !=
292                                              t->inp_cred->cr_uid))
293                                                 return (EADDRINUSE);
294
295                                         if (t && (! in_pcbbind_check_bindmulti(inp, t)))
296                                                 return (EADDRINUSE);
297                                 }
298 #endif
299                         }
300                         t = in6_pcblookup_local(pcbinfo, &sin6->sin6_addr,
301                             lport, lookupflags, cred);
302                         if (t && (reuseport & inp_so_options(t)) == 0 &&
303                             (reuseport_lb & inp_so_options(t)) == 0) {
304                                 return (EADDRINUSE);
305                         }
306 #ifdef INET
307                         if ((inp->inp_flags & IN6P_IPV6_V6ONLY) == 0 &&
308                             IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) {
309                                 struct sockaddr_in sin;
310
311                                 in6_sin6_2_sin(&sin, sin6);
312                                 t = in_pcblookup_local(pcbinfo, sin.sin_addr,
313                                    lport, lookupflags, cred);
314                                 if (t &&
315                                     (reuseport & inp_so_options(t)) == 0 &&
316                                     (reuseport_lb & inp_so_options(t)) == 0 &&
317                                     (ntohl(t->inp_laddr.s_addr) != INADDR_ANY ||
318                                         (t->inp_vflag & INP_IPV6PROTO) != 0)) {
319                                         return (EADDRINUSE);
320                                 }
321                         }
322 #endif
323                 }
324                 inp->in6p_laddr = sin6->sin6_addr;
325         }
326         if (lport == 0) {
327                 if ((error = in6_pcbsetport(&inp->in6p_laddr, inp, cred)) != 0) {
328                         /* Undo an address bind that may have occurred. */
329                         inp->in6p_laddr = in6addr_any;
330                         return (error);
331                 }
332         } else {
333                 inp->inp_lport = lport;
334                 if (in_pcbinshash(inp) != 0) {
335                         inp->in6p_laddr = in6addr_any;
336                         inp->inp_lport = 0;
337                         return (EAGAIN);
338                 }
339         }
340         return (0);
341 }
342
343 /*
344  *   Transform old in6_pcbconnect() into an inner subroutine for new
345  *   in6_pcbconnect(): Do some validity-checking on the remote
346  *   address (in mbuf 'nam') and then determine local host address
347  *   (i.e., which interface) to use to access that remote host.
348  *
349  *   This preserves definition of in6_pcbconnect(), while supporting a
350  *   slightly different version for T/TCP.  (This is more than
351  *   a bit of a kludge, but cleaning up the internal interfaces would
352  *   have forced minor changes in every protocol).
353  */
354 static int
355 in6_pcbladdr(struct inpcb *inp, struct sockaddr_in6 *sin6,
356     struct in6_addr *plocal_addr6)
357 {
358         int error = 0;
359         int scope_ambiguous = 0;
360         struct in6_addr in6a;
361         struct epoch_tracker et;
362
363         INP_WLOCK_ASSERT(inp);
364         INP_HASH_WLOCK_ASSERT(inp->inp_pcbinfo);        /* XXXRW: why? */
365
366         if (sin6->sin6_port == 0)
367                 return (EADDRNOTAVAIL);
368
369         if (sin6->sin6_scope_id == 0 && !V_ip6_use_defzone)
370                 scope_ambiguous = 1;
371         if ((error = sa6_embedscope(sin6, V_ip6_use_defzone)) != 0)
372                 return(error);
373
374         if (!CK_STAILQ_EMPTY(&V_in6_ifaddrhead)) {
375                 /*
376                  * If the destination address is UNSPECIFIED addr,
377                  * use the loopback addr, e.g ::1.
378                  */
379                 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
380                         sin6->sin6_addr = in6addr_loopback;
381         }
382         if ((error = prison_remote_ip6(inp->inp_cred, &sin6->sin6_addr)) != 0)
383                 return (error);
384
385         NET_EPOCH_ENTER(et);
386         error = in6_selectsrc_socket(sin6, inp->in6p_outputopts,
387             inp, inp->inp_cred, scope_ambiguous, &in6a, NULL);
388         NET_EPOCH_EXIT(et);
389         if (error)
390                 return (error);
391
392         /*
393          * Do not update this earlier, in case we return with an error.
394          *
395          * XXX: this in6_selectsrc_socket result might replace the bound local
396          * address with the address specified by setsockopt(IPV6_PKTINFO).
397          * Is it the intended behavior?
398          */
399         *plocal_addr6 = in6a;
400
401         /*
402          * Don't do pcblookup call here; return interface in
403          * plocal_addr6
404          * and exit to caller, that will do the lookup.
405          */
406
407         return (0);
408 }
409
410 /*
411  * Outer subroutine:
412  * Connect from a socket to a specified address.
413  * Both address and port must be specified in argument sin.
414  * If don't have a local address for this socket yet,
415  * then pick one.
416  */
417 int
418 in6_pcbconnect_mbuf(struct inpcb *inp, struct sockaddr *nam,
419     struct ucred *cred, struct mbuf *m, bool rehash)
420 {
421         struct inpcbinfo *pcbinfo = inp->inp_pcbinfo;
422         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)nam;
423         struct sockaddr_in6 laddr6;
424         int error;
425
426         KASSERT(sin6->sin6_family == AF_INET6,
427             ("%s: invalid address family for %p", __func__, sin6));
428         KASSERT(sin6->sin6_len == sizeof(*sin6),
429             ("%s: invalid address length for %p", __func__, sin6));
430
431         bzero(&laddr6, sizeof(laddr6));
432         laddr6.sin6_family = AF_INET6;
433
434         INP_WLOCK_ASSERT(inp);
435         INP_HASH_WLOCK_ASSERT(pcbinfo);
436
437 #ifdef ROUTE_MPATH
438         if (CALC_FLOWID_OUTBOUND) {
439                 uint32_t hash_type, hash_val;
440
441                 hash_val = fib6_calc_software_hash(&inp->in6p_laddr,
442                     &sin6->sin6_addr, 0, sin6->sin6_port,
443                     inp->inp_socket->so_proto->pr_protocol, &hash_type);
444                 inp->inp_flowid = hash_val;
445                 inp->inp_flowtype = hash_type;
446         }
447 #endif
448         /*
449          * Call inner routine, to assign local interface address.
450          * in6_pcbladdr() may automatically fill in sin6_scope_id.
451          */
452         if ((error = in6_pcbladdr(inp, sin6, &laddr6.sin6_addr)) != 0)
453                 return (error);
454
455         if (in6_pcblookup_hash_locked(pcbinfo, &sin6->sin6_addr,
456                                sin6->sin6_port,
457                               IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)
458                               ? &laddr6.sin6_addr : &inp->in6p_laddr,
459                               inp->inp_lport, 0, NULL, M_NODOM) != NULL) {
460                 return (EADDRINUSE);
461         }
462         if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) {
463                 if (inp->inp_lport == 0) {
464                         /*
465                          * rehash was required to be true in the past for
466                          * this case; retain that convention.  However,
467                          * we now call in_pcb_lport_dest rather than
468                          * in6_pcbbind; the former does not insert into
469                          * the hash table, the latter does.  Change rehash
470                          * to false to do the in_pcbinshash below.
471                          */
472                         KASSERT(rehash == true,
473                             ("Rehashing required for unbound inps"));
474                         rehash = false;
475                         error = in_pcb_lport_dest(inp,
476                             (struct sockaddr *) &laddr6, &inp->inp_lport,
477                             (struct sockaddr *) sin6, sin6->sin6_port, cred,
478                             INPLOOKUP_WILDCARD);
479                         if (error)
480                                 return (error);
481                 }
482                 inp->in6p_laddr = laddr6.sin6_addr;
483         }
484         inp->in6p_faddr = sin6->sin6_addr;
485         inp->inp_fport = sin6->sin6_port;
486         /* update flowinfo - draft-itojun-ipv6-flowlabel-api-00 */
487         inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
488         if (inp->inp_flags & IN6P_AUTOFLOWLABEL)
489                 inp->inp_flow |=
490                     (htonl(ip6_randomflowlabel()) & IPV6_FLOWLABEL_MASK);
491
492         if (rehash) {
493                 in_pcbrehash(inp);
494         } else {
495                 in_pcbinshash(inp);
496         }
497
498         return (0);
499 }
500
501 int
502 in6_pcbconnect(struct inpcb *inp, struct sockaddr *nam, struct ucred *cred)
503 {
504
505         return (in6_pcbconnect_mbuf(inp, nam, cred, NULL, true));
506 }
507
508 void
509 in6_pcbdisconnect(struct inpcb *inp)
510 {
511
512         INP_WLOCK_ASSERT(inp);
513         INP_HASH_WLOCK_ASSERT(inp->inp_pcbinfo);
514
515         bzero((caddr_t)&inp->in6p_faddr, sizeof(inp->in6p_faddr));
516         inp->inp_fport = 0;
517         /* clear flowinfo - draft-itojun-ipv6-flowlabel-api-00 */
518         inp->inp_flow &= ~IPV6_FLOWLABEL_MASK;
519         in_pcbrehash(inp);
520 }
521
522 struct sockaddr *
523 in6_sockaddr(in_port_t port, struct in6_addr *addr_p)
524 {
525         struct sockaddr_in6 *sin6;
526
527         sin6 = malloc(sizeof *sin6, M_SONAME, M_WAITOK);
528         bzero(sin6, sizeof *sin6);
529         sin6->sin6_family = AF_INET6;
530         sin6->sin6_len = sizeof(*sin6);
531         sin6->sin6_port = port;
532         sin6->sin6_addr = *addr_p;
533         (void)sa6_recoverscope(sin6); /* XXX: should catch errors */
534
535         return (struct sockaddr *)sin6;
536 }
537
538 struct sockaddr *
539 in6_v4mapsin6_sockaddr(in_port_t port, struct in_addr *addr_p)
540 {
541         struct sockaddr_in sin;
542         struct sockaddr_in6 *sin6_p;
543
544         bzero(&sin, sizeof sin);
545         sin.sin_family = AF_INET;
546         sin.sin_len = sizeof(sin);
547         sin.sin_port = port;
548         sin.sin_addr = *addr_p;
549
550         sin6_p = malloc(sizeof *sin6_p, M_SONAME,
551                 M_WAITOK);
552         in6_sin_2_v4mapsin6(&sin, sin6_p);
553
554         return (struct sockaddr *)sin6_p;
555 }
556
557 int
558 in6_getsockaddr(struct socket *so, struct sockaddr **nam)
559 {
560         struct inpcb *inp;
561         struct in6_addr addr;
562         in_port_t port;
563
564         inp = sotoinpcb(so);
565         KASSERT(inp != NULL, ("in6_getsockaddr: inp == NULL"));
566
567         INP_RLOCK(inp);
568         port = inp->inp_lport;
569         addr = inp->in6p_laddr;
570         INP_RUNLOCK(inp);
571
572         *nam = in6_sockaddr(port, &addr);
573         return 0;
574 }
575
576 int
577 in6_getpeeraddr(struct socket *so, struct sockaddr **nam)
578 {
579         struct inpcb *inp;
580         struct in6_addr addr;
581         in_port_t port;
582
583         inp = sotoinpcb(so);
584         KASSERT(inp != NULL, ("in6_getpeeraddr: inp == NULL"));
585
586         INP_RLOCK(inp);
587         port = inp->inp_fport;
588         addr = inp->in6p_faddr;
589         INP_RUNLOCK(inp);
590
591         *nam = in6_sockaddr(port, &addr);
592         return 0;
593 }
594
595 int
596 in6_mapped_sockaddr(struct socket *so, struct sockaddr **nam)
597 {
598         struct  inpcb *inp;
599         int     error;
600
601         inp = sotoinpcb(so);
602         KASSERT(inp != NULL, ("in6_mapped_sockaddr: inp == NULL"));
603
604 #ifdef INET
605         if ((inp->inp_vflag & (INP_IPV4 | INP_IPV6)) == INP_IPV4) {
606                 error = in_getsockaddr(so, nam);
607                 if (error == 0)
608                         in6_sin_2_v4mapsin6_in_sock(nam);
609         } else
610 #endif
611         {
612                 /* scope issues will be handled in in6_getsockaddr(). */
613                 error = in6_getsockaddr(so, nam);
614         }
615
616         return error;
617 }
618
619 int
620 in6_mapped_peeraddr(struct socket *so, struct sockaddr **nam)
621 {
622         struct  inpcb *inp;
623         int     error;
624
625         inp = sotoinpcb(so);
626         KASSERT(inp != NULL, ("in6_mapped_peeraddr: inp == NULL"));
627
628 #ifdef INET
629         if ((inp->inp_vflag & (INP_IPV4 | INP_IPV6)) == INP_IPV4) {
630                 error = in_getpeeraddr(so, nam);
631                 if (error == 0)
632                         in6_sin_2_v4mapsin6_in_sock(nam);
633         } else
634 #endif
635         /* scope issues will be handled in in6_getpeeraddr(). */
636         error = in6_getpeeraddr(so, nam);
637
638         return error;
639 }
640
641 /*
642  * Pass some notification to all connections of a protocol
643  * associated with address dst.  The local address and/or port numbers
644  * may be specified to limit the search.  The "usual action" will be
645  * taken, depending on the ctlinput cmd.  The caller must filter any
646  * cmds that are uninteresting (e.g., no error in the map).
647  * Call the protocol specific routine (if any) to report
648  * any errors for each matching socket.
649  */
650 static bool
651 inp_match6(const struct inpcb *inp, void *v __unused)
652 {
653
654         return ((inp->inp_vflag & INP_IPV6) != 0);
655 }
656
657 void
658 in6_pcbnotify(struct inpcbinfo *pcbinfo, struct sockaddr_in6 *sa6_dst,
659     u_int fport_arg, const struct sockaddr_in6 *src, u_int lport_arg,
660     int errno, void *cmdarg,
661     struct inpcb *(*notify)(struct inpcb *, int))
662 {
663         struct inpcb_iterator inpi = INP_ITERATOR(pcbinfo, INPLOOKUP_WLOCKPCB,
664             inp_match6, NULL);
665         struct inpcb *inp;
666         struct sockaddr_in6 sa6_src;
667         u_short fport = fport_arg, lport = lport_arg;
668         u_int32_t flowinfo;
669
670         if (IN6_IS_ADDR_UNSPECIFIED(&sa6_dst->sin6_addr))
671                 return;
672
673         /*
674          * note that src can be NULL when we get notify by local fragmentation.
675          */
676         sa6_src = (src == NULL) ? sa6_any : *src;
677         flowinfo = sa6_src.sin6_flowinfo;
678
679         while ((inp = inp_next(&inpi)) != NULL) {
680                 INP_WLOCK_ASSERT(inp);
681                 /*
682                  * If the error designates a new path MTU for a destination
683                  * and the application (associated with this socket) wanted to
684                  * know the value, notify.
685                  * XXX: should we avoid to notify the value to TCP sockets?
686                  */
687                 if (errno == EMSGSIZE && cmdarg != NULL)
688                         ip6_notify_pmtu(inp, sa6_dst, *(uint32_t *)cmdarg);
689
690                 /*
691                  * Detect if we should notify the error. If no source and
692                  * destination ports are specified, but non-zero flowinfo and
693                  * local address match, notify the error. This is the case
694                  * when the error is delivered with an encrypted buffer
695                  * by ESP. Otherwise, just compare addresses and ports
696                  * as usual.
697                  */
698                 if (lport == 0 && fport == 0 && flowinfo &&
699                     inp->inp_socket != NULL &&
700                     flowinfo == (inp->inp_flow & IPV6_FLOWLABEL_MASK) &&
701                     IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, &sa6_src.sin6_addr))
702                         goto do_notify;
703                 else if (!IN6_ARE_ADDR_EQUAL(&inp->in6p_faddr,
704                                              &sa6_dst->sin6_addr) ||
705                          inp->inp_socket == 0 ||
706                          (lport && inp->inp_lport != lport) ||
707                          (!IN6_IS_ADDR_UNSPECIFIED(&sa6_src.sin6_addr) &&
708                           !IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr,
709                                               &sa6_src.sin6_addr)) ||
710                          (fport && inp->inp_fport != fport)) {
711                         continue;
712                 }
713
714           do_notify:
715                 if (notify)
716                         (*notify)(inp, errno);
717         }
718 }
719
720 /*
721  * Lookup a PCB based on the local address and port.  Caller must hold the
722  * hash lock.  No inpcb locks or references are acquired.
723  */
724 struct inpcb *
725 in6_pcblookup_local(struct inpcbinfo *pcbinfo, struct in6_addr *laddr,
726     u_short lport, int lookupflags, struct ucred *cred)
727 {
728         struct inpcb *inp;
729         int matchwild = 3, wildcard;
730
731         KASSERT((lookupflags & ~(INPLOOKUP_WILDCARD)) == 0,
732             ("%s: invalid lookup flags %d", __func__, lookupflags));
733
734         INP_HASH_LOCK_ASSERT(pcbinfo);
735
736         if ((lookupflags & INPLOOKUP_WILDCARD) == 0) {
737                 struct inpcbhead *head;
738                 /*
739                  * Look for an unconnected (wildcard foreign addr) PCB that
740                  * matches the local address and port we're looking for.
741                  */
742                 head = &pcbinfo->ipi_hashbase[INP_PCBHASH_WILD(lport,
743                     pcbinfo->ipi_hashmask)];
744                 CK_LIST_FOREACH(inp, head, inp_hash) {
745                         /* XXX inp locking */
746                         if ((inp->inp_vflag & INP_IPV6) == 0)
747                                 continue;
748                         if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) &&
749                             IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr) &&
750                             inp->inp_lport == lport) {
751                                 /* Found. */
752                                 if (prison_equal_ip6(cred->cr_prison,
753                                     inp->inp_cred->cr_prison))
754                                         return (inp);
755                         }
756                 }
757                 /*
758                  * Not found.
759                  */
760                 return (NULL);
761         } else {
762                 struct inpcbporthead *porthash;
763                 struct inpcbport *phd;
764                 struct inpcb *match = NULL;
765                 /*
766                  * Best fit PCB lookup.
767                  *
768                  * First see if this local port is in use by looking on the
769                  * port hash list.
770                  */
771                 porthash = &pcbinfo->ipi_porthashbase[INP_PCBPORTHASH(lport,
772                     pcbinfo->ipi_porthashmask)];
773                 CK_LIST_FOREACH(phd, porthash, phd_hash) {
774                         if (phd->phd_port == lport)
775                                 break;
776                 }
777                 if (phd != NULL) {
778                         /*
779                          * Port is in use by one or more PCBs. Look for best
780                          * fit.
781                          */
782                         CK_LIST_FOREACH(inp, &phd->phd_pcblist, inp_portlist) {
783                                 wildcard = 0;
784                                 if (!prison_equal_ip6(cred->cr_prison,
785                                     inp->inp_cred->cr_prison))
786                                         continue;
787                                 /* XXX inp locking */
788                                 if ((inp->inp_vflag & INP_IPV6) == 0)
789                                         continue;
790                                 if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr))
791                                         wildcard++;
792                                 if (!IN6_IS_ADDR_UNSPECIFIED(
793                                         &inp->in6p_laddr)) {
794                                         if (IN6_IS_ADDR_UNSPECIFIED(laddr))
795                                                 wildcard++;
796                                         else if (!IN6_ARE_ADDR_EQUAL(
797                                             &inp->in6p_laddr, laddr))
798                                                 continue;
799                                 } else {
800                                         if (!IN6_IS_ADDR_UNSPECIFIED(laddr))
801                                                 wildcard++;
802                                 }
803                                 if (wildcard < matchwild) {
804                                         match = inp;
805                                         matchwild = wildcard;
806                                         if (matchwild == 0)
807                                                 break;
808                                 }
809                         }
810                 }
811                 return (match);
812         }
813 }
814
815 static bool
816 in6_multi_match(const struct inpcb *inp, void *v __unused)
817 {
818
819         if ((inp->inp_vflag & INP_IPV6) && inp->in6p_moptions != NULL)
820                 return (true);
821         else
822                 return (false);
823 }
824
825 void
826 in6_pcbpurgeif0(struct inpcbinfo *pcbinfo, struct ifnet *ifp)
827 {
828         struct inpcb_iterator inpi = INP_ITERATOR(pcbinfo, INPLOOKUP_RLOCKPCB,
829             in6_multi_match, NULL);
830         struct inpcb *inp;
831         struct in6_multi *inm;
832         struct in6_mfilter *imf;
833         struct ip6_moptions *im6o;
834
835         IN6_MULTI_LOCK_ASSERT();
836
837         while ((inp = inp_next(&inpi)) != NULL) {
838                 INP_RLOCK_ASSERT(inp);
839
840                 im6o = inp->in6p_moptions;
841                 /*
842                  * Unselect the outgoing ifp for multicast if it
843                  * is being detached.
844                  */
845                 if (im6o->im6o_multicast_ifp == ifp)
846                         im6o->im6o_multicast_ifp = NULL;
847                 /*
848                  * Drop multicast group membership if we joined
849                  * through the interface being detached.
850                  */
851 restart:
852                 IP6_MFILTER_FOREACH(imf, &im6o->im6o_head) {
853                         if ((inm = imf->im6f_in6m) == NULL)
854                                 continue;
855                         if (inm->in6m_ifp != ifp)
856                                 continue;
857                         ip6_mfilter_remove(&im6o->im6o_head, imf);
858                         in6_leavegroup_locked(inm, NULL);
859                         ip6_mfilter_free(imf);
860                         goto restart;
861                 }
862         }
863 }
864
865 /*
866  * Check for alternatives when higher level complains
867  * about service problems.  For now, invalidate cached
868  * routing information.  If the route was created dynamically
869  * (by a redirect), time to try a default gateway again.
870  */
871 void
872 in6_losing(struct inpcb *inp)
873 {
874
875         RO_INVALIDATE_CACHE(&inp->inp_route6);
876 }
877
878 /*
879  * After a routing change, flush old routing
880  * and allocate a (hopefully) better one.
881  */
882 struct inpcb *
883 in6_rtchange(struct inpcb *inp, int errno __unused)
884 {
885
886         RO_INVALIDATE_CACHE(&inp->inp_route6);
887         return inp;
888 }
889
890 static bool
891 in6_pcblookup_lb_numa_match(const struct inpcblbgroup *grp, int domain)
892 {
893         return (domain == M_NODOM || domain == grp->il_numa_domain);
894 }
895
896 static struct inpcb *
897 in6_pcblookup_lbgroup(const struct inpcbinfo *pcbinfo,
898     const struct in6_addr *laddr, uint16_t lport, const struct in6_addr *faddr,
899     uint16_t fport, int lookupflags, uint8_t domain)
900 {
901         const struct inpcblbgrouphead *hdr;
902         struct inpcblbgroup *grp;
903         struct inpcblbgroup *jail_exact, *jail_wild, *local_exact, *local_wild;
904
905         INP_HASH_LOCK_ASSERT(pcbinfo);
906
907         hdr = &pcbinfo->ipi_lbgrouphashbase[
908             INP_PCBPORTHASH(lport, pcbinfo->ipi_lbgrouphashmask)];
909
910         /*
911          * Search for an LB group match based on the following criteria:
912          * - prefer jailed groups to non-jailed groups
913          * - prefer exact source address matches to wildcard matches
914          * - prefer groups bound to the specified NUMA domain 
915          */
916         jail_exact = jail_wild = local_exact = local_wild = NULL;
917         CK_LIST_FOREACH(grp, hdr, il_list) {
918                 bool injail;
919
920 #ifdef INET
921                 if (!(grp->il_vflag & INP_IPV6))
922                         continue;
923 #endif
924                 if (grp->il_lport != lport)
925                         continue;
926
927                 injail = prison_flag(grp->il_cred, PR_IP6) != 0;
928                 if (injail && prison_check_ip6_locked(grp->il_cred->cr_prison,
929                     laddr) != 0)
930                         continue;
931
932                 if (IN6_ARE_ADDR_EQUAL(&grp->il6_laddr, laddr)) {
933                         if (injail) {
934                                 jail_exact = grp;
935                                 if (in6_pcblookup_lb_numa_match(grp, domain))
936                                         /* This is a perfect match. */
937                                         goto out;
938                         } else if (local_exact == NULL ||
939                             in6_pcblookup_lb_numa_match(grp, domain)) {
940                                 local_exact = grp;
941                         }
942                 } else if (IN6_IS_ADDR_UNSPECIFIED(&grp->il6_laddr) &&
943                     (lookupflags & INPLOOKUP_WILDCARD) != 0) {
944                         if (injail) {
945                                 if (jail_wild == NULL ||
946                                     in6_pcblookup_lb_numa_match(grp, domain))
947                                         jail_wild = grp;
948                         } else if (local_wild == NULL ||
949                             in6_pcblookup_lb_numa_match(grp, domain)) {
950                                 local_wild = grp;
951                         }
952                 }
953         }
954
955         if (jail_exact != NULL)
956                 grp = jail_exact;
957         else if (jail_wild != NULL)
958                 grp = jail_wild;
959         else if (local_exact != NULL)
960                 grp = local_exact;
961         else
962                 grp = local_wild;
963         if (grp == NULL)
964                 return (NULL);
965 out:
966         return (grp->il_inp[INP6_PCBLBGROUP_PKTHASH(faddr, lport, fport) %
967             grp->il_inpcnt]);
968 }
969
970 /*
971  * Lookup PCB in hash list.  Used in in_pcb.c as well as here.
972  */
973 struct inpcb *
974 in6_pcblookup_hash_locked(struct inpcbinfo *pcbinfo, struct in6_addr *faddr,
975     u_int fport_arg, struct in6_addr *laddr, u_int lport_arg,
976     int lookupflags, struct ifnet *ifp, uint8_t numa_domain)
977 {
978         struct inpcbhead *head;
979         struct inpcb *inp, *tmpinp;
980         u_short fport = fport_arg, lport = lport_arg;
981
982         KASSERT((lookupflags & ~(INPLOOKUP_WILDCARD)) == 0,
983             ("%s: invalid lookup flags %d", __func__, lookupflags));
984
985         INP_HASH_LOCK_ASSERT(pcbinfo);
986
987         /*
988          * First look for an exact match.
989          */
990         tmpinp = NULL;
991         head = &pcbinfo->ipi_hashbase[INP6_PCBHASH(faddr, lport, fport,
992             pcbinfo->ipi_hashmask)];
993         CK_LIST_FOREACH(inp, head, inp_hash) {
994                 /* XXX inp locking */
995                 if ((inp->inp_vflag & INP_IPV6) == 0)
996                         continue;
997                 if (IN6_ARE_ADDR_EQUAL(&inp->in6p_faddr, faddr) &&
998                     IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr) &&
999                     inp->inp_fport == fport &&
1000                     inp->inp_lport == lport) {
1001                         /*
1002                          * XXX We should be able to directly return
1003                          * the inp here, without any checks.
1004                          * Well unless both bound with SO_REUSEPORT?
1005                          */
1006                         if (prison_flag(inp->inp_cred, PR_IP6))
1007                                 return (inp);
1008                         if (tmpinp == NULL)
1009                                 tmpinp = inp;
1010                 }
1011         }
1012         if (tmpinp != NULL)
1013                 return (tmpinp);
1014
1015         /*
1016          * Then look for a wildcard match, if requested.
1017          */
1018         if ((lookupflags & INPLOOKUP_WILDCARD) != 0) {
1019                 struct inpcb *local_wild = NULL, *local_exact = NULL;
1020                 struct inpcb *jail_wild = NULL;
1021                 int injail;
1022
1023                 /*
1024                  * First see if an LB group matches the request before scanning
1025                  * all sockets on this port.
1026                  */
1027                 inp = in6_pcblookup_lbgroup(pcbinfo, laddr, lport, faddr,
1028                     fport, lookupflags, numa_domain);
1029                 if (inp != NULL)
1030                         return (inp);
1031
1032                 /*
1033                  * Order of socket selection - we always prefer jails.
1034                  *      1. jailed, non-wild.
1035                  *      2. jailed, wild.
1036                  *      3. non-jailed, non-wild.
1037                  *      4. non-jailed, wild.
1038                  */
1039                 head = &pcbinfo->ipi_hashbase[INP_PCBHASH_WILD(lport,
1040                     pcbinfo->ipi_hashmask)];
1041                 CK_LIST_FOREACH(inp, head, inp_hash) {
1042                         /* XXX inp locking */
1043                         if ((inp->inp_vflag & INP_IPV6) == 0)
1044                                 continue;
1045
1046                         if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) ||
1047                             inp->inp_lport != lport) {
1048                                 continue;
1049                         }
1050
1051                         injail = prison_flag(inp->inp_cred, PR_IP6);
1052                         if (injail) {
1053                                 if (prison_check_ip6_locked(
1054                                     inp->inp_cred->cr_prison, laddr) != 0)
1055                                         continue;
1056                         } else {
1057                                 if (local_exact != NULL)
1058                                         continue;
1059                         }
1060
1061                         if (IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr)) {
1062                                 if (injail)
1063                                         return (inp);
1064                                 else
1065                                         local_exact = inp;
1066                         } else if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) {
1067                                 if (injail)
1068                                         jail_wild = inp;
1069                                 else
1070                                         local_wild = inp;
1071                         }
1072                 } /* LIST_FOREACH */
1073
1074                 if (jail_wild != NULL)
1075                         return (jail_wild);
1076                 if (local_exact != NULL)
1077                         return (local_exact);
1078                 if (local_wild != NULL)
1079                         return (local_wild);
1080         } /* if ((lookupflags & INPLOOKUP_WILDCARD) != 0) */
1081
1082         /*
1083          * Not found.
1084          */
1085         return (NULL);
1086 }
1087
1088 /*
1089  * Lookup PCB in hash list, using pcbinfo tables.  This variation locks the
1090  * hash list lock, and will return the inpcb locked (i.e., requires
1091  * INPLOOKUP_LOCKPCB).
1092  */
1093 static struct inpcb *
1094 in6_pcblookup_hash(struct inpcbinfo *pcbinfo, struct in6_addr *faddr,
1095     u_int fport, struct in6_addr *laddr, u_int lport, int lookupflags,
1096     struct ifnet *ifp, uint8_t numa_domain)
1097 {
1098         struct inpcb *inp;
1099
1100         smr_enter(pcbinfo->ipi_smr);
1101         inp = in6_pcblookup_hash_locked(pcbinfo, faddr, fport, laddr, lport,
1102             lookupflags & INPLOOKUP_WILDCARD, ifp, numa_domain);
1103         if (inp != NULL) {
1104                 if (__predict_false(inp_smr_lock(inp,
1105                     (lookupflags & INPLOOKUP_LOCKMASK)) == false))
1106                         inp = NULL;
1107         } else
1108                 smr_exit(pcbinfo->ipi_smr);
1109
1110         return (inp);
1111 }
1112
1113 /*
1114  * Public inpcb lookup routines, accepting a 4-tuple, and optionally, an mbuf
1115  * from which a pre-calculated hash value may be extracted.
1116  */
1117 struct inpcb *
1118 in6_pcblookup(struct inpcbinfo *pcbinfo, struct in6_addr *faddr, u_int fport,
1119     struct in6_addr *laddr, u_int lport, int lookupflags, struct ifnet *ifp)
1120 {
1121
1122         KASSERT((lookupflags & ~INPLOOKUP_MASK) == 0,
1123             ("%s: invalid lookup flags %d", __func__, lookupflags));
1124         KASSERT((lookupflags & (INPLOOKUP_RLOCKPCB | INPLOOKUP_WLOCKPCB)) != 0,
1125             ("%s: LOCKPCB not set", __func__));
1126
1127         return (in6_pcblookup_hash(pcbinfo, faddr, fport, laddr, lport,
1128             lookupflags, ifp, M_NODOM));
1129 }
1130
1131 struct inpcb *
1132 in6_pcblookup_mbuf(struct inpcbinfo *pcbinfo, struct in6_addr *faddr,
1133     u_int fport, struct in6_addr *laddr, u_int lport, int lookupflags,
1134     struct ifnet *ifp, struct mbuf *m)
1135 {
1136
1137         KASSERT((lookupflags & ~INPLOOKUP_MASK) == 0,
1138             ("%s: invalid lookup flags %d", __func__, lookupflags));
1139         KASSERT((lookupflags & (INPLOOKUP_RLOCKPCB | INPLOOKUP_WLOCKPCB)) != 0,
1140             ("%s: LOCKPCB not set", __func__));
1141
1142         return (in6_pcblookup_hash(pcbinfo, faddr, fport, laddr, lport,
1143             lookupflags, ifp, m->m_pkthdr.numa_domain));
1144 }
1145
1146 void
1147 init_sin6(struct sockaddr_in6 *sin6, struct mbuf *m, int srcordst)
1148 {
1149         struct ip6_hdr *ip;
1150
1151         ip = mtod(m, struct ip6_hdr *);
1152         bzero(sin6, sizeof(*sin6));
1153         sin6->sin6_len = sizeof(*sin6);
1154         sin6->sin6_family = AF_INET6;
1155         sin6->sin6_addr = srcordst ? ip->ip6_dst : ip->ip6_src;
1156
1157         (void)sa6_recoverscope(sin6); /* XXX: should catch errors... */
1158
1159         return;
1160 }