kernel - Change tcp keepalive options from ms to seconds (DISRUPTIVE) (2)
[dragonfly.git] / sys / netinet / tcp_usrreq.c
1 /*
2  * Copyright (c) 2003, 2004 Jeffrey M. Hsu.  All rights reserved.
3  * Copyright (c) 2003, 2004 The DragonFly Project.  All rights reserved.
4  *
5  * This code is derived from software contributed to The DragonFly Project
6  * by Jeffrey M. Hsu.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 3. Neither the name of The DragonFly Project nor the names of its
17  *    contributors may be used to endorse or promote products derived
18  *    from this software without specific, prior written permission.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
21  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
22  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
23  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
24  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
25  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
26  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
27  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
28  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
29  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
30  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  */
33
34 /*
35  * Copyright (c) 1982, 1986, 1988, 1993
36  *      The Regents of the University of California.  All rights reserved.
37  *
38  * Redistribution and use in source and binary forms, with or without
39  * modification, are permitted provided that the following conditions
40  * are met:
41  * 1. Redistributions of source code must retain the above copyright
42  *    notice, this list of conditions and the following disclaimer.
43  * 2. Redistributions in binary form must reproduce the above copyright
44  *    notice, this list of conditions and the following disclaimer in the
45  *    documentation and/or other materials provided with the distribution.
46  * 3. Neither the name of the University nor the names of its contributors
47  *    may be used to endorse or promote products derived from this software
48  *    without specific prior written permission.
49  *
50  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
51  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
52  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
53  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
54  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
55  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
56  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
57  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
58  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
59  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
60  * SUCH DAMAGE.
61  *
62  *      From: @(#)tcp_usrreq.c  8.2 (Berkeley) 1/3/94
63  * $FreeBSD: src/sys/netinet/tcp_usrreq.c,v 1.51.2.17 2002/10/11 11:46:44 ume Exp $
64  */
65
66 #include "opt_inet.h"
67 #include "opt_inet6.h"
68 #include "opt_tcpdebug.h"
69
70 #include <sys/param.h>
71 #include <sys/systm.h>
72 #include <sys/kernel.h>
73 #include <sys/malloc.h>
74 #include <sys/sysctl.h>
75 #include <sys/globaldata.h>
76 #include <sys/thread.h>
77
78 #include <sys/mbuf.h>
79 #ifdef INET6
80 #include <sys/domain.h>
81 #endif /* INET6 */
82 #include <sys/socket.h>
83 #include <sys/socketvar.h>
84 #include <sys/socketops.h>
85 #include <sys/protosw.h>
86 #include <sys/jail.h>
87
88 #include <sys/msgport2.h>
89 #include <sys/socketvar2.h>
90
91 #include <net/if.h>
92 #include <net/netisr.h>
93 #include <net/route.h>
94
95 #include <net/netmsg2.h>
96 #include <net/netisr2.h>
97
98 #include <netinet/in.h>
99 #include <netinet/in_systm.h>
100 #ifdef INET6
101 #include <netinet/ip6.h>
102 #endif
103 #include <netinet/in_pcb.h>
104 #ifdef INET6
105 #include <netinet6/in6_pcb.h>
106 #endif
107 #include <netinet/in_var.h>
108 #include <netinet/ip_var.h>
109 #ifdef INET6
110 #include <netinet6/ip6_var.h>
111 #include <netinet6/tcp6_var.h>
112 #endif
113 #include <netinet/tcp.h>
114 #include <netinet/tcp_fsm.h>
115 #include <netinet/tcp_seq.h>
116 #include <netinet/tcp_timer.h>
117 #include <netinet/tcp_timer2.h>
118 #include <netinet/tcp_var.h>
119 #include <netinet/tcpip.h>
120 #ifdef TCPDEBUG
121 #include <netinet/tcp_debug.h>
122 #endif
123 #include <machine/limits.h>
124
125 /*
126  * Limits for TCP_KEEP* options (we will adopt the same limits that linux
127  * uses).
128  */
129 #define MAXKEEPALIVE            32767
130 #define MAXKEEPCNT              127
131
132 /*
133  * TCP protocol interface to socket abstraction.
134  */
135 extern  char *tcpstates[];      /* XXX ??? */
136
137 static int      tcp_attach (struct socket *, struct pru_attach_info *);
138 static void     tcp_connect (netmsg_t msg);
139 #ifdef INET6
140 static void     tcp6_connect (netmsg_t msg);
141 static int      tcp6_connect_oncpu(struct tcpcb *tp, int flags,
142                                 struct mbuf **mp,
143                                 struct sockaddr_in6 *sin6,
144                                 struct in6_addr *addr6);
145 #endif /* INET6 */
146 static struct tcpcb *
147                 tcp_disconnect (struct tcpcb *);
148 static struct tcpcb *
149                 tcp_usrclosed (struct tcpcb *);
150
151 #ifdef TCPDEBUG
152 #define TCPDEBUG0       int ostate = 0
153 #define TCPDEBUG1()     ostate = tp ? tp->t_state : 0
154 #define TCPDEBUG2(req)  if (tp && (so->so_options & SO_DEBUG)) \
155                                 tcp_trace(TA_USER, ostate, tp, 0, 0, req)
156 #else
157 #define TCPDEBUG0
158 #define TCPDEBUG1()
159 #define TCPDEBUG2(req)
160 #endif
161
162 /*
163  * For some ill optimized programs, which try to use TCP_NOPUSH
164  * to improve performance, will have small amount of data sits
165  * in the sending buffer.  These small amount of data will _not_
166  * be pushed into the network until more data are written into
167  * the socket or the socket write side is shutdown.
168  */ 
169 static int      tcp_disable_nopush = 1;
170 SYSCTL_INT(_net_inet_tcp, OID_AUTO, disable_nopush, CTLFLAG_RW,
171     &tcp_disable_nopush, 0, "TCP_NOPUSH socket option will have no effect");
172
173 /*
174  * Allocate socket buffer space.
175  */
176 static int
177 tcp_usr_preattach(struct socket *so, int proto __unused,
178     struct pru_attach_info *ai)
179 {
180         int error;
181
182         if (so->so_snd.ssb_hiwat == 0 || so->so_rcv.ssb_hiwat == 0) {
183                 error = soreserve(so, tcp_sendspace, tcp_recvspace,
184                                   ai->sb_rlimit);
185                 if (error)
186                         return (error);
187         }
188         atomic_set_int(&so->so_rcv.ssb_flags, SSB_AUTOSIZE);
189         atomic_set_int(&so->so_snd.ssb_flags, SSB_AUTOSIZE | SSB_PREALLOC);
190
191         return 0;
192 }
193
194 /*
195  * TCP attaches to socket via pru_attach(), reserving space,
196  * and an internet control block.  This socket may move to
197  * other CPU later when we bind/connect.
198  */
199 static void
200 tcp_usr_attach(netmsg_t msg)
201 {
202         struct socket *so = msg->base.nm_so;
203         struct pru_attach_info *ai = msg->attach.nm_ai;
204         int error;
205         struct inpcb *inp;
206         struct tcpcb *tp = NULL;
207         TCPDEBUG0;
208
209         inp = so->so_pcb;
210         KASSERT(inp == NULL, ("tcp socket attached"));
211         TCPDEBUG1();
212
213         error = tcp_attach(so, ai);
214         if (error)
215                 goto out;
216
217         if ((so->so_options & SO_LINGER) && so->so_linger == 0)
218                 so->so_linger = TCP_LINGERTIME;
219         tp = sototcpcb(so);
220 out:
221         TCPDEBUG2(PRU_ATTACH);
222         lwkt_replymsg(&msg->lmsg, error);
223 }
224
225 /*
226  * pru_detach() detaches the TCP protocol from the socket.
227  * If the protocol state is non-embryonic, then can't
228  * do this directly: have to initiate a pru_disconnect(),
229  * which may finish later; embryonic TCB's can just
230  * be discarded here.
231  */
232 static void
233 tcp_usr_detach(netmsg_t msg)
234 {
235         struct socket *so = msg->base.nm_so;
236         int error = 0;
237         struct inpcb *inp;
238         struct tcpcb *tp;
239         TCPDEBUG0;
240
241         inp = so->so_pcb;
242
243         /*
244          * If the inp is already detached or never attached, it may have
245          * been due to an async close or async attach failure.  Just return
246          * as if no error occured.
247          */
248         if (inp) {
249                 tp = intotcpcb(inp);
250                 KASSERT(tp != NULL, ("tcp_usr_detach: tp is NULL"));
251                 TCPDEBUG1();
252                 tp = tcp_disconnect(tp);
253                 TCPDEBUG2(PRU_DETACH);
254         }
255         lwkt_replymsg(&msg->lmsg, error);
256 }
257
258 /*
259  * NOTE: ignore_error is non-zero for certain disconnection races
260  * which we want to silently allow, otherwise close() may return
261  * an unexpected error.
262  *
263  * NOTE: The variables (msg) and (tp) are assumed.
264  */
265 #define COMMON_START(so, inp, ignore_error)                     \
266         TCPDEBUG0;                                              \
267                                                                 \
268         inp = so->so_pcb;                                       \
269         do {                                                    \
270                 if (inp == NULL) {                              \
271                         error = ignore_error ? 0 : EINVAL;      \
272                         tp = NULL;                              \
273                         goto out;                               \
274                 }                                               \
275                 tp = intotcpcb(inp);                            \
276                 TCPDEBUG1();                                    \
277         } while(0)
278
279 #define COMMON_END1(req, noreply)                               \
280         out: do {                                               \
281                 TCPDEBUG2(req);                                 \
282                 if (!(noreply))                                 \
283                         lwkt_replymsg(&msg->lmsg, error);       \
284                 return;                                         \
285         } while(0)
286
287 #define COMMON_END(req)         COMMON_END1((req), 0)
288
289 static void
290 tcp_sosetport(struct lwkt_msg *msg, lwkt_port_t port)
291 {
292         sosetport(((struct netmsg_base *)msg)->nm_so, port);
293 }
294
295 /*
296  * Give the socket an address.
297  */
298 static void
299 tcp_usr_bind(netmsg_t msg)
300 {
301         struct socket *so = msg->bind.base.nm_so;
302         struct sockaddr *nam = msg->bind.nm_nam;
303         struct thread *td = msg->bind.nm_td;
304         int error = 0;
305         struct inpcb *inp;
306         struct tcpcb *tp;
307         struct sockaddr_in *sinp;
308         lwkt_port_t port0 = netisr_cpuport(0);
309
310         COMMON_START(so, inp, 0);
311
312         /*
313          * Must check for multicast addresses and disallow binding
314          * to them.
315          */
316         sinp = (struct sockaddr_in *)nam;
317         if (sinp->sin_family == AF_INET &&
318             IN_MULTICAST(ntohl(sinp->sin_addr.s_addr))) {
319                 error = EAFNOSUPPORT;
320                 goto out;
321         }
322
323         /*
324          * Check "already bound" here (in_pcbbind() does the same check
325          * though), so we don't forward a connected socket to netisr0,
326          * which would panic in the following in_pcbunlink().
327          */
328         if (inp->inp_lport != 0 || inp->inp_laddr.s_addr != INADDR_ANY) {
329                 error = EINVAL; /* already bound */
330                 goto out;
331         }
332
333         /*
334          * Use netisr0 to serialize in_pcbbind(), so that pru_detach and
335          * pru_bind for different sockets on the same local port could be
336          * properly ordered.  The original race is illustrated here for
337          * reference.
338          *
339          * s1 = socket();
340          * bind(s1, *.PORT);
341          * close(s1);  <----- asynchronous
342          * s2 = socket();
343          * bind(s2, *.PORT);
344          *
345          * All will expect bind(s2, *.PORT) to succeed.  However, it will
346          * fail, if following sequence happens due to random socket initial
347          * msgport and asynchronous close(2):
348          *
349          *    netisrN                  netisrM
350          *       :                        :
351          *       :                    pru_bind(s2) [*.PORT is used by s1]
352          *  pru_detach(s1)                :
353          */
354         if (&curthread->td_msgport != port0) {
355                 lwkt_msg_t lmsg = &msg->bind.base.lmsg;
356
357                 KASSERT((msg->bind.nm_flags & PRUB_RELINK) == 0,
358                     ("already asked to relink"));
359
360                 in_pcbunlink(so->so_pcb, &tcbinfo[mycpuid]);
361                 msg->bind.nm_flags |= PRUB_RELINK;
362
363                 TCP_STATE_MIGRATE_START(tp);
364
365                 /* See the related comment in tcp_connect() */
366                 lwkt_setmsg_receipt(lmsg, tcp_sosetport);
367                 lwkt_forwardmsg(port0, lmsg);
368                 /* msg invalid now */
369                 return;
370         }
371         KASSERT(so->so_port == port0, ("so_port is not netisr0"));
372
373         if (msg->bind.nm_flags & PRUB_RELINK) {
374                 msg->bind.nm_flags &= ~PRUB_RELINK;
375                 TCP_STATE_MIGRATE_END(tp);
376                 in_pcblink(so->so_pcb, &tcbinfo[mycpuid]);
377         }
378         KASSERT(inp->inp_pcbinfo == &tcbinfo[0], ("pcbinfo is not tcbinfo0"));
379
380         error = in_pcbbind(inp, nam, td);
381         if (error)
382                 goto out;
383
384         COMMON_END(PRU_BIND);
385 }
386
387 #ifdef INET6
388
389 static void
390 tcp6_usr_bind(netmsg_t msg)
391 {
392         struct socket *so = msg->bind.base.nm_so;
393         struct sockaddr *nam = msg->bind.nm_nam;
394         struct thread *td = msg->bind.nm_td;
395         int error = 0;
396         struct inpcb *inp;
397         struct tcpcb *tp;
398         struct sockaddr_in6 *sin6p;
399
400         COMMON_START(so, inp, 0);
401
402         /*
403          * Must check for multicast addresses and disallow binding
404          * to them.
405          */
406         sin6p = (struct sockaddr_in6 *)nam;
407         if (sin6p->sin6_family == AF_INET6 &&
408             IN6_IS_ADDR_MULTICAST(&sin6p->sin6_addr)) {
409                 error = EAFNOSUPPORT;
410                 goto out;
411         }
412         error = in6_pcbbind(inp, nam, td);
413         if (error)
414                 goto out;
415         COMMON_END(PRU_BIND);
416 }
417 #endif /* INET6 */
418
419 struct netmsg_inswildcard {
420         struct netmsg_base      base;
421         struct inpcb            *nm_inp;
422 };
423
424 static void
425 in_pcbinswildcardhash_handler(netmsg_t msg)
426 {
427         struct netmsg_inswildcard *nm = (struct netmsg_inswildcard *)msg;
428         int cpu = mycpuid, nextcpu;
429
430         in_pcbinswildcardhash_oncpu(nm->nm_inp, &tcbinfo[cpu]);
431
432         nextcpu = cpu + 1;
433         if (nextcpu < netisr_ncpus)
434                 lwkt_forwardmsg(netisr_cpuport(nextcpu), &nm->base.lmsg);
435         else
436                 lwkt_replymsg(&nm->base.lmsg, 0);
437 }
438
439 /*
440  * Prepare to accept connections.
441  */
442 static void
443 tcp_usr_listen(netmsg_t msg)
444 {
445         struct socket *so = msg->listen.base.nm_so;
446         struct thread *td = msg->listen.nm_td;
447         int error = 0;
448         struct inpcb *inp;
449         struct tcpcb *tp;
450         struct netmsg_inswildcard nm;
451         lwkt_port_t port0 = netisr_cpuport(0);
452
453         COMMON_START(so, inp, 0);
454
455         if (&curthread->td_msgport != port0) {
456                 lwkt_msg_t lmsg = &msg->listen.base.lmsg;
457
458                 KASSERT((msg->listen.nm_flags & PRUL_RELINK) == 0,
459                     ("already asked to relink"));
460
461                 in_pcbunlink(so->so_pcb, &tcbinfo[mycpuid]);
462                 msg->listen.nm_flags |= PRUL_RELINK;
463
464                 TCP_STATE_MIGRATE_START(tp);
465
466                 /* See the related comment in tcp_connect() */
467                 lwkt_setmsg_receipt(lmsg, tcp_sosetport);
468                 lwkt_forwardmsg(port0, lmsg);
469                 /* msg invalid now */
470                 return;
471         }
472         KASSERT(so->so_port == port0, ("so_port is not netisr0"));
473
474         if (msg->listen.nm_flags & PRUL_RELINK) {
475                 msg->listen.nm_flags &= ~PRUL_RELINK;
476                 TCP_STATE_MIGRATE_END(tp);
477                 in_pcblink(so->so_pcb, &tcbinfo[mycpuid]);
478         }
479         KASSERT(inp->inp_pcbinfo == &tcbinfo[0], ("pcbinfo is not tcbinfo0"));
480
481         if (tp->t_flags & TF_LISTEN)
482                 goto out;
483
484         if (inp->inp_lport == 0) {
485                 error = in_pcbbind(inp, NULL, td);
486                 if (error)
487                         goto out;
488         }
489
490         TCP_STATE_CHANGE(tp, TCPS_LISTEN);
491         tp->t_flags |= TF_LISTEN;
492         tp->tt_msg = NULL; /* Catch any invalid timer usage */
493
494         /*
495          * Create tcpcb per-cpu port cache
496          *
497          * NOTE:
498          * This _must_ be done before installing this inpcb into
499          * wildcard hash.
500          */
501         tcp_pcbport_create(tp);
502
503         if (netisr_ncpus > 1) {
504                 /*
505                  * Put this inpcb into wildcard hash on other cpus.
506                  */
507                 ASSERT_INP_NOTINHASH(inp);
508                 netmsg_init(&nm.base, NULL, &curthread->td_msgport,
509                             MSGF_PRIORITY, in_pcbinswildcardhash_handler);
510                 nm.nm_inp = inp;
511                 lwkt_domsg(netisr_cpuport(1), &nm.base.lmsg, 0);
512         }
513         in_pcbinswildcardhash(inp);
514         COMMON_END(PRU_LISTEN);
515 }
516
517 #ifdef INET6
518
519 static void
520 tcp6_usr_listen(netmsg_t msg)
521 {
522         struct socket *so = msg->listen.base.nm_so;
523         struct thread *td = msg->listen.nm_td;
524         int error = 0;
525         struct inpcb *inp;
526         struct tcpcb *tp;
527         struct netmsg_inswildcard nm;
528
529         COMMON_START(so, inp, 0);
530
531         if (tp->t_flags & TF_LISTEN)
532                 goto out;
533
534         if (inp->inp_lport == 0) {
535                 error = in6_pcbbind(inp, NULL, td);
536                 if (error)
537                         goto out;
538         }
539
540         TCP_STATE_CHANGE(tp, TCPS_LISTEN);
541         tp->t_flags |= TF_LISTEN;
542         tp->tt_msg = NULL; /* Catch any invalid timer usage */
543
544         /*
545          * Create tcpcb per-cpu port cache
546          *
547          * NOTE:
548          * This _must_ be done before installing this inpcb into
549          * wildcard hash.
550          */
551         tcp_pcbport_create(tp);
552
553         if (netisr_ncpus > 1) {
554                 /*
555                  * Put this inpcb into wildcard hash on other cpus.
556                  */
557                 KKASSERT(so->so_port == netisr_cpuport(0));
558                 ASSERT_NETISR0;
559                 KKASSERT(inp->inp_pcbinfo == &tcbinfo[0]);
560                 ASSERT_INP_NOTINHASH(inp);
561
562                 netmsg_init(&nm.base, NULL, &curthread->td_msgport,
563                             MSGF_PRIORITY, in_pcbinswildcardhash_handler);
564                 nm.nm_inp = inp;
565                 lwkt_domsg(netisr_cpuport(1), &nm.base.lmsg, 0);
566         }
567         in_pcbinswildcardhash(inp);
568         COMMON_END(PRU_LISTEN);
569 }
570 #endif /* INET6 */
571
572 /*
573  * Initiate connection to peer.
574  * Create a template for use in transmissions on this connection.
575  * Enter SYN_SENT state, and mark socket as connecting.
576  * Start keep-alive timer, and seed output sequence space.
577  * Send initial segment on connection.
578  */
579 static void
580 tcp_usr_connect(netmsg_t msg)
581 {
582         struct socket *so = msg->connect.base.nm_so;
583         struct sockaddr *nam = msg->connect.nm_nam;
584         struct thread *td = msg->connect.nm_td;
585         int error = 0;
586         struct inpcb *inp;
587         struct tcpcb *tp;
588         struct sockaddr_in *sinp;
589
590         ASSERT_NETISR_NCPUS(mycpuid);
591
592         COMMON_START(so, inp, 0);
593
594         /*
595          * Must disallow TCP ``connections'' to multicast addresses.
596          */
597         sinp = (struct sockaddr_in *)nam;
598         if (sinp->sin_family == AF_INET
599             && IN_MULTICAST(ntohl(sinp->sin_addr.s_addr))) {
600                 error = EAFNOSUPPORT;
601                 goto out;
602         }
603
604         if (!prison_remote_ip(td, (struct sockaddr*)sinp)) {
605                 error = EAFNOSUPPORT; /* IPv6 only jail */
606                 goto out;
607         }
608
609         tcp_connect(msg);
610         /* msg is invalid now */
611         return;
612 out:
613         if (msg->connect.nm_m) {
614                 m_freem(msg->connect.nm_m);
615                 msg->connect.nm_m = NULL;
616         }
617         if (msg->connect.nm_flags & PRUC_HELDTD)
618                 lwkt_rele(td);
619         if (error && (msg->connect.nm_flags & PRUC_ASYNC)) {
620                 so->so_error = error;
621                 soisdisconnected(so);
622         }
623         lwkt_replymsg(&msg->lmsg, error);
624 }
625
626 #ifdef INET6
627
628 static void
629 tcp6_usr_connect(netmsg_t msg)
630 {
631         struct socket *so = msg->connect.base.nm_so;
632         struct sockaddr *nam = msg->connect.nm_nam;
633         struct thread *td = msg->connect.nm_td;
634         int error = 0;
635         struct inpcb *inp;
636         struct tcpcb *tp;
637         struct sockaddr_in6 *sin6p;
638
639         ASSERT_NETISR_NCPUS(mycpuid);
640
641         COMMON_START(so, inp, 0);
642
643         /*
644          * Must disallow TCP ``connections'' to multicast addresses.
645          */
646         sin6p = (struct sockaddr_in6 *)nam;
647         if (sin6p->sin6_family == AF_INET6
648             && IN6_IS_ADDR_MULTICAST(&sin6p->sin6_addr)) {
649                 error = EAFNOSUPPORT;
650                 goto out;
651         }
652
653         if (!prison_remote_ip(td, nam)) {
654                 error = EAFNOSUPPORT; /* IPv4 only jail */
655                 goto out;
656         }
657
658         /* Reject v4-mapped address */
659         if (IN6_IS_ADDR_V4MAPPED(&sin6p->sin6_addr)) {
660                 error = EADDRNOTAVAIL;
661                 goto out;
662         }
663
664         inp->inp_inc.inc_isipv6 = 1;
665         tcp6_connect(msg);
666         /* msg is invalid now */
667         return;
668 out:
669         if (msg->connect.nm_m) {
670                 m_freem(msg->connect.nm_m);
671                 msg->connect.nm_m = NULL;
672         }
673         lwkt_replymsg(&msg->lmsg, error);
674 }
675
676 #endif /* INET6 */
677
678 /*
679  * Initiate disconnect from peer.
680  * If connection never passed embryonic stage, just drop;
681  * else if don't need to let data drain, then can just drop anyways,
682  * else have to begin TCP shutdown process: mark socket disconnecting,
683  * drain unread data, state switch to reflect user close, and
684  * send segment (e.g. FIN) to peer.  Socket will be really disconnected
685  * when peer sends FIN and acks ours.
686  *
687  * SHOULD IMPLEMENT LATER PRU_CONNECT VIA REALLOC TCPCB.
688  */
689 static void
690 tcp_usr_disconnect(netmsg_t msg)
691 {
692         struct socket *so = msg->disconnect.base.nm_so;
693         int error = 0;
694         struct inpcb *inp;
695         struct tcpcb *tp;
696
697         COMMON_START(so, inp, 1);
698         tp = tcp_disconnect(tp);
699         COMMON_END(PRU_DISCONNECT);
700 }
701
702 /*
703  * Accept a connection.  Essentially all the work is
704  * done at higher levels; just return the address
705  * of the peer, storing through addr.
706  */
707 static void
708 tcp_usr_accept(netmsg_t msg)
709 {
710         struct socket *so = msg->accept.base.nm_so;
711         struct sockaddr **nam = msg->accept.nm_nam;
712         int error = 0;
713         struct inpcb *inp;
714         struct tcpcb *tp = NULL;
715         TCPDEBUG0;
716
717         inp = so->so_pcb;
718         if (so->so_state & SS_ISDISCONNECTED) {
719                 error = ECONNABORTED;
720                 goto out;
721         }
722         if (inp == NULL) {
723                 error = EINVAL;
724                 goto out;
725         }
726
727         tp = intotcpcb(inp);
728         TCPDEBUG1();
729         in_setpeeraddr(so, nam);
730         COMMON_END(PRU_ACCEPT);
731 }
732
733 #ifdef INET6
734 static void
735 tcp6_usr_accept(netmsg_t msg)
736 {
737         struct socket *so = msg->accept.base.nm_so;
738         struct sockaddr **nam = msg->accept.nm_nam;
739         int error = 0;
740         struct inpcb *inp;
741         struct tcpcb *tp = NULL;
742         TCPDEBUG0;
743
744         inp = so->so_pcb;
745
746         if (so->so_state & SS_ISDISCONNECTED) {
747                 error = ECONNABORTED;
748                 goto out;
749         }
750         if (inp == NULL) {
751                 error = EINVAL;
752                 goto out;
753         }
754         tp = intotcpcb(inp);
755         TCPDEBUG1();
756         in6_setpeeraddr(so, nam);
757         COMMON_END(PRU_ACCEPT);
758 }
759 #endif /* INET6 */
760
761 /*
762  * Mark the connection as being incapable of further output.
763  */
764 static void
765 tcp_usr_shutdown(netmsg_t msg)
766 {
767         struct socket *so = msg->shutdown.base.nm_so;
768         int error = 0;
769         struct inpcb *inp;
770         struct tcpcb *tp;
771
772         COMMON_START(so, inp, 0);
773         socantsendmore(so);
774         tp = tcp_usrclosed(tp);
775         if (tp)
776                 error = tcp_output(tp);
777         COMMON_END(PRU_SHUTDOWN);
778 }
779
780 /*
781  * After a receive, possibly send window update to peer.
782  */
783 static void
784 tcp_usr_rcvd(netmsg_t msg)
785 {
786         struct socket *so = msg->rcvd.base.nm_so;
787         int error = 0, noreply = 0;
788         struct inpcb *inp;
789         struct tcpcb *tp;
790
791         COMMON_START(so, inp, 0);
792
793         if (msg->rcvd.nm_pru_flags & PRUR_ASYNC) {
794                 noreply = 1;
795                 so_async_rcvd_reply(so);
796         }
797         tcp_output(tp);
798
799         COMMON_END1(PRU_RCVD, noreply);
800 }
801
802 /*
803  * Do a send by putting data in output queue and updating urgent
804  * marker if URG set.  Possibly send more data.  Unlike the other
805  * pru_*() routines, the mbuf chains are our responsibility.  We
806  * must either enqueue them or free them.  The other pru_* routines
807  * generally are caller-frees.
808  */
809 static void
810 tcp_usr_send(netmsg_t msg)
811 {
812         struct socket *so = msg->send.base.nm_so;
813         int flags = msg->send.nm_flags;
814         struct mbuf *m = msg->send.nm_m;
815         int error = 0;
816         struct inpcb *inp;
817         struct tcpcb *tp;
818         TCPDEBUG0;
819
820         KKASSERT(msg->send.nm_control == NULL);
821         KKASSERT(msg->send.nm_addr == NULL);
822         KKASSERT((flags & PRUS_FREEADDR) == 0);
823
824         inp = so->so_pcb;
825
826         if (inp == NULL) {
827                 /*
828                  * OOPS! we lost a race, the TCP session got reset after
829                  * we checked SS_CANTSENDMORE, eg: while doing uiomove or a
830                  * network interrupt in the non-critical section of sosend().
831                  */
832                 m_freem(m);
833                 error = ECONNRESET;     /* XXX EPIPE? */
834                 tp = NULL;
835                 TCPDEBUG1();
836                 goto out;
837         }
838         tp = intotcpcb(inp);
839         TCPDEBUG1();
840
841 #ifdef foo
842         /*
843          * This is no longer necessary, since:
844          * - sosendtcp() has already checked it for us
845          * - It does not work with asynchronized send
846          */
847
848         /*
849          * Don't let too much OOB data build up
850          */
851         if (flags & PRUS_OOB) {
852                 if (ssb_space(&so->so_snd) < -512) {
853                         m_freem(m);
854                         error = ENOBUFS;
855                         goto out;
856                 }
857         }
858 #endif
859
860         /*
861          * Pump the data into the socket.
862          */
863         if (m) {
864                 ssb_appendstream(&so->so_snd, m);
865                 sowwakeup(so);
866         }
867         if (flags & PRUS_OOB) {
868                 /*
869                  * According to RFC961 (Assigned Protocols),
870                  * the urgent pointer points to the last octet
871                  * of urgent data.  We continue, however,
872                  * to consider it to indicate the first octet
873                  * of data past the urgent section.
874                  * Otherwise, snd_up should be one lower.
875                  */
876                 tp->snd_up = tp->snd_una + so->so_snd.ssb_cc;
877                 tp->t_flags |= TF_FORCE;
878                 error = tcp_output(tp);
879                 tp->t_flags &= ~TF_FORCE;
880         } else {
881                 if (flags & PRUS_EOF) {
882                         /*
883                          * Close the send side of the connection after
884                          * the data is sent.
885                          */
886                         socantsendmore(so);
887                         tp = tcp_usrclosed(tp);
888                 }
889                 if (tp != NULL && !tcp_output_pending(tp)) {
890                         if (flags & PRUS_MORETOCOME)
891                                 tp->t_flags |= TF_MORETOCOME;
892                         error = tcp_output_fair(tp);
893                         if (flags & PRUS_MORETOCOME)
894                                 tp->t_flags &= ~TF_MORETOCOME;
895                 }
896         }
897         COMMON_END1((flags & PRUS_OOB) ? PRU_SENDOOB :
898                    ((flags & PRUS_EOF) ? PRU_SEND_EOF : PRU_SEND),
899                    (flags & PRUS_NOREPLY));
900 }
901
902 /*
903  * NOTE: (so) is referenced from soabort*() and netmsg_pru_abort()
904  *       will sofree() it when we return.
905  */
906 static void
907 tcp_usr_abort(netmsg_t msg)
908 {
909         struct socket *so = msg->abort.base.nm_so;
910         int error = 0;
911         struct inpcb *inp;
912         struct tcpcb *tp;
913
914         COMMON_START(so, inp, 1);
915         tp = tcp_drop(tp, ECONNABORTED);
916         COMMON_END(PRU_ABORT);
917 }
918
919 /*
920  * Receive out-of-band data.
921  */
922 static void
923 tcp_usr_rcvoob(netmsg_t msg)
924 {
925         struct socket *so = msg->rcvoob.base.nm_so;
926         struct mbuf *m = msg->rcvoob.nm_m;
927         int flags = msg->rcvoob.nm_flags;
928         int error = 0;
929         struct inpcb *inp;
930         struct tcpcb *tp;
931
932         COMMON_START(so, inp, 0);
933         if ((so->so_oobmark == 0 &&
934              (so->so_state & SS_RCVATMARK) == 0) ||
935             so->so_options & SO_OOBINLINE ||
936             tp->t_oobflags & TCPOOB_HADDATA) {
937                 error = EINVAL;
938                 goto out;
939         }
940         if ((tp->t_oobflags & TCPOOB_HAVEDATA) == 0) {
941                 error = EWOULDBLOCK;
942                 goto out;
943         }
944         m->m_len = 1;
945         *mtod(m, caddr_t) = tp->t_iobc;
946         if ((flags & MSG_PEEK) == 0)
947                 tp->t_oobflags ^= (TCPOOB_HAVEDATA | TCPOOB_HADDATA);
948         COMMON_END(PRU_RCVOOB);
949 }
950
951 static void
952 tcp_usr_savefaddr(struct socket *so, const struct sockaddr *faddr)
953 {
954         in_savefaddr(so, faddr);
955 }
956
957 #ifdef INET6
958 static void
959 tcp6_usr_savefaddr(struct socket *so, const struct sockaddr *faddr)
960 {
961         in6_savefaddr(so, faddr);
962 }
963 #endif
964
965 static int
966 tcp_usr_preconnect(struct socket *so, const struct sockaddr *nam,
967     struct thread *td __unused)
968 {
969         const struct sockaddr_in *sinp;
970
971         sinp = (const struct sockaddr_in *)nam;
972         if (sinp->sin_family == AF_INET &&
973             IN_MULTICAST(ntohl(sinp->sin_addr.s_addr)))
974                 return EAFNOSUPPORT;
975
976         soisconnecting(so);
977         return 0;
978 }
979
980 /* xxx - should be const */
981 struct pr_usrreqs tcp_usrreqs = {
982         .pru_abort = tcp_usr_abort,
983         .pru_accept = tcp_usr_accept,
984         .pru_attach = tcp_usr_attach,
985         .pru_bind = tcp_usr_bind,
986         .pru_connect = tcp_usr_connect,
987         .pru_connect2 = pr_generic_notsupp,
988         .pru_control = in_control_dispatch,
989         .pru_detach = tcp_usr_detach,
990         .pru_disconnect = tcp_usr_disconnect,
991         .pru_listen = tcp_usr_listen,
992         .pru_peeraddr = in_setpeeraddr_dispatch,
993         .pru_rcvd = tcp_usr_rcvd,
994         .pru_rcvoob = tcp_usr_rcvoob,
995         .pru_send = tcp_usr_send,
996         .pru_sense = pru_sense_null,
997         .pru_shutdown = tcp_usr_shutdown,
998         .pru_sockaddr = in_setsockaddr_dispatch,
999         .pru_sosend = sosendtcp,
1000         .pru_soreceive = sorecvtcp,
1001         .pru_savefaddr = tcp_usr_savefaddr,
1002         .pru_preconnect = tcp_usr_preconnect,
1003         .pru_preattach = tcp_usr_preattach
1004 };
1005
1006 #ifdef INET6
1007 struct pr_usrreqs tcp6_usrreqs = {
1008         .pru_abort = tcp_usr_abort,
1009         .pru_accept = tcp6_usr_accept,
1010         .pru_attach = tcp_usr_attach,
1011         .pru_bind = tcp6_usr_bind,
1012         .pru_connect = tcp6_usr_connect,
1013         .pru_connect2 = pr_generic_notsupp,
1014         .pru_control = in6_control_dispatch,
1015         .pru_detach = tcp_usr_detach,
1016         .pru_disconnect = tcp_usr_disconnect,
1017         .pru_listen = tcp6_usr_listen,
1018         .pru_peeraddr = in6_setpeeraddr_dispatch,
1019         .pru_rcvd = tcp_usr_rcvd,
1020         .pru_rcvoob = tcp_usr_rcvoob,
1021         .pru_send = tcp_usr_send,
1022         .pru_sense = pru_sense_null,
1023         .pru_shutdown = tcp_usr_shutdown,
1024         .pru_sockaddr = in6_setsockaddr_dispatch,
1025         .pru_sosend = sosendtcp,
1026         .pru_soreceive = sorecvtcp,
1027         .pru_savefaddr = tcp6_usr_savefaddr
1028 };
1029 #endif /* INET6 */
1030
1031 static int
1032 tcp_connect_oncpu(struct tcpcb *tp, int flags, struct mbuf *m,
1033                   const struct sockaddr_in *sin, struct sockaddr_in *if_sin,
1034                   uint16_t hash)
1035 {
1036         struct inpcb *inp = tp->t_inpcb, *oinp;
1037         struct socket *so = inp->inp_socket;
1038         struct route *ro = &inp->inp_route;
1039
1040         KASSERT(inp->inp_pcbinfo == &tcbinfo[mycpu->gd_cpuid],
1041             ("pcbinfo mismatch"));
1042
1043         oinp = in_pcblookup_hash(inp->inp_pcbinfo,
1044                                  sin->sin_addr, sin->sin_port,
1045                                  (inp->inp_laddr.s_addr != INADDR_ANY ?
1046                                   inp->inp_laddr : if_sin->sin_addr),
1047                                 inp->inp_lport, 0, NULL);
1048         if (oinp != NULL) {
1049                 m_freem(m);
1050                 return (EADDRINUSE);
1051         }
1052         if (inp->inp_laddr.s_addr == INADDR_ANY)
1053                 inp->inp_laddr = if_sin->sin_addr;
1054         KASSERT(inp->inp_faddr.s_addr == sin->sin_addr.s_addr,
1055             ("faddr mismatch for reconnect"));
1056         KASSERT(inp->inp_fport == sin->sin_port,
1057             ("fport mismatch for reconnect"));
1058         in_pcbinsconnhash(inp);
1059
1060         inp->inp_flags |= INP_HASH;
1061         inp->inp_hashval = hash;
1062
1063         /*
1064          * We are now on the inpcb's owner CPU, if the cached route was
1065          * freed because the rtentry's owner CPU is not the current CPU
1066          * (e.g. in tcp_connect()), then we try to reallocate it here with
1067          * the hope that a rtentry may be cloned from a RTF_PRCLONING
1068          * rtentry.
1069          */
1070         if (!(inp->inp_socket->so_options & SO_DONTROUTE) && /*XXX*/
1071             ro->ro_rt == NULL) {
1072                 bzero(&ro->ro_dst, sizeof(struct sockaddr_in));
1073                 ro->ro_dst.sa_family = AF_INET;
1074                 ro->ro_dst.sa_len = sizeof(struct sockaddr_in);
1075                 ((struct sockaddr_in *)&ro->ro_dst)->sin_addr =
1076                         sin->sin_addr;
1077                 rtalloc(ro);
1078         }
1079
1080         /*
1081          * Now that no more errors can occur, change the protocol processing
1082          * port to the current thread (which is the correct thread).
1083          *
1084          * Create TCP timer message now; we are on the tcpcb's owner
1085          * CPU/thread.
1086          */
1087         tcp_create_timermsg(tp, &curthread->td_msgport);
1088
1089         /*
1090          * Compute window scaling to request.  Use a larger scaling then
1091          * needed for the initial receive buffer in case the receive buffer
1092          * gets expanded.
1093          */
1094         if (tp->request_r_scale < TCP_MIN_WINSHIFT)
1095                 tp->request_r_scale = TCP_MIN_WINSHIFT;
1096         while (tp->request_r_scale < TCP_MAX_WINSHIFT &&
1097                (TCP_MAXWIN << tp->request_r_scale) < so->so_rcv.ssb_hiwat
1098         ) {
1099                 tp->request_r_scale++;
1100         }
1101
1102         soisconnecting(so);
1103         tcpstat.tcps_connattempt++;
1104         TCP_STATE_CHANGE(tp, TCPS_SYN_SENT);
1105         tcp_callout_reset(tp, tp->tt_keep, tp->t_keepinit, tcp_timer_keep);
1106         tp->iss = tcp_new_isn(tp);
1107         tcp_sendseqinit(tp);
1108         if (m) {
1109                 ssb_appendstream(&so->so_snd, m);
1110                 m = NULL;
1111                 if (flags & PRUS_OOB)
1112                         tp->snd_up = tp->snd_una + so->so_snd.ssb_cc;
1113         }
1114
1115         /*
1116          * Close the send side of the connection after
1117          * the data is sent if flagged.
1118          */
1119         if ((flags & (PRUS_OOB|PRUS_EOF)) == PRUS_EOF) {
1120                 socantsendmore(so);
1121                 tp = tcp_usrclosed(tp);
1122         }
1123         return (tcp_output(tp));
1124 }
1125
1126 /*
1127  * Common subroutine to open a TCP connection to remote host specified
1128  * by struct sockaddr_in in mbuf *nam.  Call in_pcbbind to assign a local
1129  * port number if needed.  Call in_pcbladdr to do the routing and to choose
1130  * a local host address (interface).
1131  * Initialize connection parameters and enter SYN-SENT state.
1132  */
1133 static void
1134 tcp_connect(netmsg_t msg)
1135 {
1136         struct socket *so = msg->connect.base.nm_so;
1137         struct sockaddr *nam = msg->connect.nm_nam;
1138         struct thread *td = msg->connect.nm_td;
1139         struct sockaddr_in *sin = (struct sockaddr_in *)nam;
1140         struct sockaddr_in *if_sin = NULL;
1141         struct inpcb *inp;
1142         struct tcpcb *tp;
1143         int error;
1144         uint16_t hash;
1145         lwkt_port_t port;
1146
1147         COMMON_START(so, inp, 0);
1148
1149         /*
1150          * Reconnect our pcb if we have to
1151          */
1152         if (msg->connect.nm_flags & PRUC_RECONNECT) {
1153                 msg->connect.nm_flags &= ~PRUC_RECONNECT;
1154                 TCP_STATE_MIGRATE_END(tp);
1155                 in_pcblink(so->so_pcb, &tcbinfo[mycpu->gd_cpuid]);
1156         } else {
1157                 if (inp->inp_faddr.s_addr != INADDR_ANY) {
1158                         kprintf("inpcb %p, double-connect race\n", inp);
1159                         error = EISCONN;
1160                         if (so->so_state & SS_ISCONNECTING)
1161                                 error = EALREADY;
1162                         goto out;
1163                 }
1164                 KASSERT(inp->inp_fport == 0, ("invalid fport"));
1165         }
1166
1167         /*
1168          * Select local port, if it is not yet selected.
1169          */
1170         if (inp->inp_lport == 0) {
1171                 KKASSERT(inp->inp_laddr.s_addr == INADDR_ANY);
1172
1173                 error = in_pcbladdr(inp, nam, &if_sin, td);
1174                 if (error)
1175                         goto out;
1176                 inp->inp_laddr.s_addr = if_sin->sin_addr.s_addr;
1177                 msg->connect.nm_flags |= PRUC_HASLADDR;
1178
1179                 /*
1180                  * Install faddr/fport earlier, so that when this
1181                  * inpcb is installed on to the lport hash, the
1182                  * 4-tuple contains correct value.
1183                  *
1184                  * NOTE: The faddr/fport will have to be installed
1185                  * after the in_pcbladdr(), which may change them.
1186                  */
1187                 inp->inp_faddr = sin->sin_addr;
1188                 inp->inp_fport = sin->sin_port;
1189
1190                 error = in_pcbbind_remote(inp, nam, td);
1191                 if (error)
1192                         goto out;
1193         }
1194
1195         if ((msg->connect.nm_flags & PRUC_HASLADDR) == 0) {
1196                 /*
1197                  * Rarely used path:
1198                  * This inpcb was bound before this connect.
1199                  */
1200                 error = in_pcbladdr(inp, nam, &if_sin, td);
1201                 if (error)
1202                         goto out;
1203
1204                 /*
1205                  * Save or refresh the faddr/fport, since they may
1206                  * be changed by in_pcbladdr().
1207                  */
1208                 inp->inp_faddr = sin->sin_addr;
1209                 inp->inp_fport = sin->sin_port;
1210         }
1211 #ifdef INVARIANTS
1212         else {
1213                 KASSERT(inp->inp_faddr.s_addr == sin->sin_addr.s_addr,
1214                     ("faddr mismatch for reconnect"));
1215                 KASSERT(inp->inp_fport == sin->sin_port,
1216                     ("fport mismatch for reconnect"));
1217         }
1218 #endif
1219         KKASSERT(inp->inp_socket == so);
1220
1221         hash = tcp_addrhash(sin->sin_addr.s_addr, sin->sin_port,
1222                             (inp->inp_laddr.s_addr != INADDR_ANY ?
1223                              inp->inp_laddr.s_addr : if_sin->sin_addr.s_addr),
1224                             inp->inp_lport);
1225         port = netisr_hashport(hash);
1226
1227         if (port != &curthread->td_msgport) {
1228                 lwkt_msg_t lmsg = &msg->connect.base.lmsg;
1229
1230                 /*
1231                  * in_pcbladdr() may have allocated a route entry for us
1232                  * on the current CPU, but we need a route entry on the
1233                  * inpcb's owner CPU, so free it here.
1234                  */
1235                 in_pcbresetroute(inp);
1236
1237                 /*
1238                  * We are moving the protocol processing port the socket
1239                  * is on, we have to unlink here and re-link on the
1240                  * target cpu.
1241                  */
1242                 in_pcbunlink(so->so_pcb, &tcbinfo[mycpu->gd_cpuid]);
1243                 msg->connect.nm_flags |= PRUC_RECONNECT;
1244                 msg->connect.base.nm_dispatch = tcp_connect;
1245
1246                 TCP_STATE_MIGRATE_START(tp);
1247
1248                 /*
1249                  * Use message put done receipt to change this socket's
1250                  * so_port, i.e. _after_ this message was put onto the
1251                  * target netisr's msgport but _before_ the message could
1252                  * be pulled from the target netisr's msgport, so that:
1253                  * - The upper half (socket code) will not see the new
1254                  *   msgport before this message reaches the new msgport
1255                  *   and messages for this socket will be ordered.
1256                  * - This message will see the new msgport, when its
1257                  *   handler is called in the target netisr.
1258                  *
1259                  * NOTE:
1260                  * We MUST use messege put done receipt to change this
1261                  * socket's so_port:
1262                  * If we changed the so_port in this netisr after the
1263                  * lwkt_forwardmsg (so messages for this socket will be
1264                  * ordered) and changed the so_port in the target netisr
1265                  * at the very beginning of this message's handler, we
1266                  * would suffer so_port overwritten race, given this
1267                  * message might be forwarded again.
1268                  *
1269                  * NOTE:
1270                  * This mechanism depends on that the netisr's msgport
1271                  * is spin msgport (currently it is :).
1272                  *
1273                  * If the upper half saw the new msgport before this
1274                  * message reached the target netisr's msgport, the
1275                  * messages sent from the upper half could reach the new
1276                  * msgport before this message, thus there would be
1277                  * message reordering.  The worst case could be soclose()
1278                  * saw the new msgport and the detach message could reach
1279                  * the new msgport before this message, i.e. the inpcb
1280                  * could have been destroyed when this message was still
1281                  * pending on or on its way to the new msgport.  Other
1282                  * weird cases could also happen, e.g. inpcb->inp_pcbinfo,
1283                  * since we have unlinked this inpcb from the current
1284                  * pcbinfo first.
1285                  */
1286                 lwkt_setmsg_receipt(lmsg, tcp_sosetport);
1287                 lwkt_forwardmsg(port, lmsg);
1288                 /* msg invalid now */
1289                 return;
1290         } else if (msg->connect.nm_flags & PRUC_HELDTD) {
1291                 /*
1292                  * The original thread is no longer needed; release it.
1293                  */
1294                 lwkt_rele(td);
1295                 msg->connect.nm_flags &= ~PRUC_HELDTD;
1296         }
1297         error = tcp_connect_oncpu(tp, msg->connect.nm_sndflags,
1298                                   msg->connect.nm_m, sin, if_sin, hash);
1299         msg->connect.nm_m = NULL;
1300 out:
1301         if (msg->connect.nm_m) {
1302                 m_freem(msg->connect.nm_m);
1303                 msg->connect.nm_m = NULL;
1304         }
1305         if (msg->connect.nm_flags & PRUC_HELDTD)
1306                 lwkt_rele(td);
1307         if (error && (msg->connect.nm_flags & PRUC_ASYNC)) {
1308                 so->so_error = error;
1309                 soisdisconnected(so);
1310         }
1311         lwkt_replymsg(&msg->connect.base.lmsg, error);
1312         /* msg invalid now */
1313 }
1314
1315 #ifdef INET6
1316
1317 static void
1318 tcp6_connect(netmsg_t msg)
1319 {
1320         struct tcpcb *tp;
1321         struct socket *so = msg->connect.base.nm_so;
1322         struct sockaddr *nam = msg->connect.nm_nam;
1323         struct thread *td = msg->connect.nm_td;
1324         struct inpcb *inp;
1325         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)nam;
1326         struct in6_addr *addr6;
1327         lwkt_port_t port;
1328         int error;
1329
1330         COMMON_START(so, inp, 0);
1331
1332         /*
1333          * Reconnect our pcb if we have to
1334          */
1335         if (msg->connect.nm_flags & PRUC_RECONNECT) {
1336                 msg->connect.nm_flags &= ~PRUC_RECONNECT;
1337                 TCP_STATE_MIGRATE_END(tp);
1338                 in_pcblink(so->so_pcb, &tcbinfo[mycpu->gd_cpuid]);
1339         }
1340
1341         /*
1342          * Bind if we have to
1343          */
1344         if (inp->inp_lport == 0) {
1345                 error = in6_pcbbind(inp, NULL, td);
1346                 if (error)
1347                         goto out;
1348         }
1349
1350         /*
1351          * Cannot simply call in_pcbconnect, because there might be an
1352          * earlier incarnation of this same connection still in
1353          * TIME_WAIT state, creating an ADDRINUSE error.
1354          */
1355         error = in6_pcbladdr(inp, nam, &addr6, td);
1356         if (error)
1357                 goto out;
1358
1359         port = tcp6_addrport(); /* XXX hack for now, always cpu0 */
1360
1361         if (port != &curthread->td_msgport) {
1362                 lwkt_msg_t lmsg = &msg->connect.base.lmsg;
1363
1364                 /*
1365                  * in_pcbladdr() may have allocated a route entry for us
1366                  * on the current CPU, but we need a route entry on the
1367                  * inpcb's owner CPU, so free it here.
1368                  */
1369                 in_pcbresetroute(inp);
1370
1371                 in_pcbunlink(so->so_pcb, &tcbinfo[mycpu->gd_cpuid]);
1372                 msg->connect.nm_flags |= PRUC_RECONNECT;
1373                 msg->connect.base.nm_dispatch = tcp6_connect;
1374
1375                 TCP_STATE_MIGRATE_START(tp);
1376
1377                 /* See the related comment in tcp_connect() */
1378                 lwkt_setmsg_receipt(lmsg, tcp_sosetport);
1379                 lwkt_forwardmsg(port, lmsg);
1380                 /* msg invalid now */
1381                 return;
1382         }
1383         error = tcp6_connect_oncpu(tp, msg->connect.nm_sndflags,
1384                                    &msg->connect.nm_m, sin6, addr6);
1385         /* nm_m may still be intact */
1386 out:
1387         if (msg->connect.nm_m) {
1388                 m_freem(msg->connect.nm_m);
1389                 msg->connect.nm_m = NULL;
1390         }
1391         lwkt_replymsg(&msg->connect.base.lmsg, error);
1392         /* msg invalid now */
1393 }
1394
1395 static int
1396 tcp6_connect_oncpu(struct tcpcb *tp, int flags, struct mbuf **mp,
1397                    struct sockaddr_in6 *sin6, struct in6_addr *addr6)
1398 {
1399         struct mbuf *m = *mp;
1400         struct inpcb *inp = tp->t_inpcb;
1401         struct socket *so = inp->inp_socket;
1402         struct inpcb *oinp;
1403
1404         /*
1405          * Cannot simply call in_pcbconnect, because there might be an
1406          * earlier incarnation of this same connection still in
1407          * TIME_WAIT state, creating an ADDRINUSE error.
1408          */
1409         oinp = in6_pcblookup_hash(inp->inp_pcbinfo,
1410                                   &sin6->sin6_addr, sin6->sin6_port,
1411                                   (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr) ?
1412                                       addr6 : &inp->in6p_laddr),
1413                                   inp->inp_lport,  0, NULL);
1414         if (oinp)
1415                 return (EADDRINUSE);
1416
1417         if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr))
1418                 inp->in6p_laddr = *addr6;
1419         inp->in6p_faddr = sin6->sin6_addr;
1420         inp->inp_fport = sin6->sin6_port;
1421         if ((sin6->sin6_flowinfo & IPV6_FLOWINFO_MASK) != 0)
1422                 inp->in6p_flowinfo = sin6->sin6_flowinfo;
1423         in_pcbinsconnhash(inp);
1424
1425         /*
1426          * Now that no more errors can occur, change the protocol processing
1427          * port to the current thread (which is the correct thread).
1428          *
1429          * Create TCP timer message now; we are on the tcpcb's owner
1430          * CPU/thread.
1431          */
1432         tcp_create_timermsg(tp, &curthread->td_msgport);
1433
1434         /* Compute window scaling to request.  */
1435         if (tp->request_r_scale < TCP_MIN_WINSHIFT)
1436                 tp->request_r_scale = TCP_MIN_WINSHIFT;
1437         while (tp->request_r_scale < TCP_MAX_WINSHIFT &&
1438             (TCP_MAXWIN << tp->request_r_scale) < so->so_rcv.ssb_hiwat) {
1439                 tp->request_r_scale++;
1440         }
1441
1442         soisconnecting(so);
1443         tcpstat.tcps_connattempt++;
1444         TCP_STATE_CHANGE(tp, TCPS_SYN_SENT);
1445         tcp_callout_reset(tp, tp->tt_keep, tp->t_keepinit, tcp_timer_keep);
1446         tp->iss = tcp_new_isn(tp);
1447         tcp_sendseqinit(tp);
1448         if (m) {
1449                 ssb_appendstream(&so->so_snd, m);
1450                 *mp = NULL;
1451                 if (flags & PRUS_OOB)
1452                         tp->snd_up = tp->snd_una + so->so_snd.ssb_cc;
1453         }
1454
1455         /*
1456          * Close the send side of the connection after
1457          * the data is sent if flagged.
1458          */
1459         if ((flags & (PRUS_OOB|PRUS_EOF)) == PRUS_EOF) {
1460                 socantsendmore(so);
1461                 tp = tcp_usrclosed(tp);
1462         }
1463         return (tcp_output(tp));
1464 }
1465
1466 #endif /* INET6 */
1467
1468 /*
1469  * The new sockopt interface makes it possible for us to block in the
1470  * copyin/out step (if we take a page fault).  Taking a page fault while
1471  * in a critical section is probably a Bad Thing.  (Since sockets and pcbs
1472  * both now use TSM, there probably isn't any need for this function to 
1473  * run in a critical section any more.  This needs more examination.)
1474  */
1475 void
1476 tcp_ctloutput(netmsg_t msg)
1477 {
1478         struct socket *so = msg->base.nm_so;
1479         struct sockopt *sopt = msg->ctloutput.nm_sopt;
1480         struct thread *td = NULL;
1481         int     error, opt, optval, opthz;
1482         struct  inpcb *inp;
1483         struct  tcpcb *tp;
1484
1485         if (msg->ctloutput.nm_flags & PRCO_HELDTD)
1486                 td = sopt->sopt_td;
1487
1488         error = 0;
1489         inp = so->so_pcb;
1490         if (inp == NULL) {
1491                 error = ECONNRESET;
1492                 goto done;
1493         }
1494         tp = intotcpcb(inp);
1495
1496         /* Get socket's owner cpuid hint */
1497         if (sopt->sopt_level == SOL_SOCKET &&
1498             sopt->sopt_dir == SOPT_GET &&
1499             sopt->sopt_name == SO_CPUHINT) {
1500                 if (tp->t_flags & TF_LISTEN) {
1501                         /*
1502                          * Listen sockets owner cpuid is always 0,
1503                          * which does not make sense if SO_REUSEPORT
1504                          * is not set.
1505                          *
1506                          * NOTE: inp_lgrpindex is _not_ assigned in jail.
1507                          */
1508                         if ((so->so_options & SO_REUSEPORT) &&
1509                             inp->inp_lgrpindex >= 0)
1510                                 optval = inp->inp_lgrpindex % netisr_ncpus;
1511                         else
1512                                 optval = -1; /* no hint */
1513                 } else {
1514                         optval = mycpuid;
1515                 }
1516                 soopt_from_kbuf(sopt, &optval, sizeof(optval));
1517                 goto done;
1518         }
1519
1520         if (sopt->sopt_level != IPPROTO_TCP) {
1521                 if (sopt->sopt_level == IPPROTO_IP) {
1522                         switch (sopt->sopt_name) {
1523                         case IP_MULTICAST_IF:
1524                         case IP_MULTICAST_VIF:
1525                         case IP_MULTICAST_TTL:
1526                         case IP_MULTICAST_LOOP:
1527                         case IP_ADD_MEMBERSHIP:
1528                         case IP_DROP_MEMBERSHIP:
1529                                 /*
1530                                  * Multicast does not make sense on
1531                                  * TCP sockets.
1532                                  */
1533                                 error = EOPNOTSUPP;
1534                                 goto done;
1535                         }
1536                 }
1537 #ifdef INET6
1538                 if (INP_CHECK_SOCKAF(so, AF_INET6))
1539                         ip6_ctloutput_dispatch(msg);
1540                 else
1541 #endif /* INET6 */
1542                 ip_ctloutput(msg);
1543                 /* msg invalid now */
1544                 if (td != NULL)
1545                         lwkt_rele(td);
1546                 return;
1547         }
1548
1549         switch (sopt->sopt_dir) {
1550         case SOPT_SET:
1551                 error = soopt_to_kbuf(sopt, &optval, sizeof optval,
1552                                       sizeof optval);
1553                 if (error)
1554                         break;
1555                 switch (sopt->sopt_name) {
1556                 case TCP_FASTKEEP:
1557                         if (optval > 0)
1558                                 tp->t_keepidle = tp->t_keepintvl;
1559                         else
1560                                 tp->t_keepidle = tcp_keepidle;
1561                         tcp_timer_keep_activity(tp, 0);
1562                         break;
1563 #ifdef TCP_SIGNATURE
1564                 case TCP_SIGNATURE_ENABLE:
1565                         if (tp->t_state == TCPS_CLOSED) {
1566                                 /*
1567                                  * This is the only safe state that this
1568                                  * option could be changed.  Some segments
1569                                  * could already have been sent in other
1570                                  * states.
1571                                  */
1572                                 if (optval > 0)
1573                                         tp->t_flags |= TF_SIGNATURE;
1574                                 else
1575                                         tp->t_flags &= ~TF_SIGNATURE;
1576                         } else {
1577                                 error = EOPNOTSUPP;
1578                         }
1579                         break;
1580 #endif /* TCP_SIGNATURE */
1581                 case TCP_NODELAY:
1582                 case TCP_NOOPT:
1583                         switch (sopt->sopt_name) {
1584                         case TCP_NODELAY:
1585                                 opt = TF_NODELAY;
1586                                 break;
1587                         case TCP_NOOPT:
1588                                 opt = TF_NOOPT;
1589                                 break;
1590                         default:
1591                                 opt = 0; /* dead code to fool gcc */
1592                                 break;
1593                         }
1594
1595                         if (optval)
1596                                 tp->t_flags |= opt;
1597                         else
1598                                 tp->t_flags &= ~opt;
1599                         break;
1600
1601                 case TCP_NOPUSH:
1602                         if (tcp_disable_nopush)
1603                                 break;
1604                         if (optval)
1605                                 tp->t_flags |= TF_NOPUSH;
1606                         else {
1607                                 tp->t_flags &= ~TF_NOPUSH;
1608                                 error = tcp_output(tp);
1609                         }
1610                         break;
1611
1612                 case TCP_MAXSEG:
1613                         /*
1614                          * Must be between 0 and maxseg.  If the requested
1615                          * maxseg is too small to satisfy the desired minmss,
1616                          * pump it up (silently so sysctl modifications of
1617                          * minmss do not create unexpected program failures).
1618                          * Handle degenerate cases.
1619                          */
1620                         if (optval > 0 && optval <= tp->t_maxseg) {
1621                                 if (optval + 40 < tcp_minmss) {
1622                                         optval = tcp_minmss - 40;
1623                                         if (optval < 0)
1624                                                 optval = 1;
1625                                 }
1626                                 tp->t_maxseg = optval;
1627                         } else {
1628                                 error = EINVAL;
1629                         }
1630                         break;
1631
1632                 case TCP_KEEPINIT:
1633                 case TCP_KEEPIDLE:
1634                 case TCP_KEEPINTVL:
1635                         if (optval < 1 || optval > MAXKEEPALIVE) {
1636                                 error = EINVAL;
1637                                 break;
1638                         }
1639                         opthz = optval * hz;
1640
1641                         switch (sopt->sopt_name) {
1642                         case TCP_KEEPINIT:
1643                                 tp->t_keepinit = opthz;
1644                                 break;
1645                         case TCP_KEEPIDLE:
1646                                 tp->t_keepidle = opthz;
1647                                 tcp_timer_keep_activity(tp, 0);
1648                                 break;
1649                         case TCP_KEEPINTVL:
1650                                 tp->t_keepintvl = opthz;
1651                                 tp->t_maxidle = tp->t_keepintvl * tp->t_keepcnt;
1652                                 break;
1653                         }
1654                         break;
1655
1656                 case TCP_KEEPCNT:
1657                         if (optval < 1 || optval > MAXKEEPCNT) {
1658                                 error = EINVAL;
1659                                 break;
1660                         }
1661                         tp->t_keepcnt = optval;
1662                         tp->t_maxidle = tp->t_keepintvl * tp->t_keepcnt;
1663                         break;
1664
1665                 default:
1666                         error = ENOPROTOOPT;
1667                         break;
1668                 }
1669                 break;
1670
1671         case SOPT_GET:
1672                 switch (sopt->sopt_name) {
1673 #ifdef TCP_SIGNATURE
1674                 case TCP_SIGNATURE_ENABLE:
1675                         optval = (tp->t_flags & TF_SIGNATURE) ? 1 : 0;
1676                         break;
1677 #endif /* TCP_SIGNATURE */
1678                 case TCP_NODELAY:
1679                         optval = tp->t_flags & TF_NODELAY;
1680                         break;
1681                 case TCP_MAXSEG:
1682                         optval = tp->t_maxseg;
1683                         break;
1684                 case TCP_NOOPT:
1685                         optval = tp->t_flags & TF_NOOPT;
1686                         break;
1687                 case TCP_NOPUSH:
1688                         optval = tp->t_flags & TF_NOPUSH;
1689                         break;
1690                 case TCP_KEEPINIT:
1691                         optval = tp->t_keepinit / hz;
1692                         break;
1693                 case TCP_KEEPIDLE:
1694                         optval = tp->t_keepidle / hz;
1695                         break;
1696                 case TCP_KEEPINTVL:
1697                         optval = tp->t_keepintvl / hz;
1698                         break;
1699                 case TCP_KEEPCNT:
1700                         optval = tp->t_keepcnt;
1701                         break;
1702                 default:
1703                         error = ENOPROTOOPT;
1704                         break;
1705                 }
1706                 if (error == 0)
1707                         soopt_from_kbuf(sopt, &optval, sizeof optval);
1708                 break;
1709         }
1710 done:
1711         if (td != NULL)
1712                 lwkt_rele(td);
1713         lwkt_replymsg(&msg->lmsg, error);
1714 }
1715
1716 struct netmsg_tcp_ctloutput {
1717         struct netmsg_pr_ctloutput ctloutput;
1718         struct sockopt          sopt;
1719         int                     sopt_val;
1720 };
1721
1722 /*
1723  * Allocate netmsg_pr_ctloutput for asynchronous tcp_ctloutput.
1724  */
1725 struct netmsg_pr_ctloutput *
1726 tcp_ctloutmsg(struct sockopt *sopt)
1727 {
1728         struct netmsg_tcp_ctloutput *msg;
1729         int flags = 0, error;
1730
1731         KASSERT(sopt->sopt_dir == SOPT_SET, ("not from ctloutput"));
1732
1733         /* Only small set of options allows asynchronous setting. */
1734         if (sopt->sopt_level != IPPROTO_TCP)
1735                 return NULL;
1736         switch (sopt->sopt_name) {
1737         case TCP_NODELAY:
1738         case TCP_NOOPT:
1739         case TCP_NOPUSH:
1740         case TCP_FASTKEEP:
1741                 break;
1742         default:
1743                 return NULL;
1744         }
1745
1746         msg = kmalloc(sizeof(*msg), M_LWKTMSG, M_WAITOK | M_NULLOK);
1747         if (msg == NULL) {
1748                 /* Fallback to synchronous tcp_ctloutput */
1749                 return NULL;
1750         }
1751
1752         /* Save the sockopt */
1753         msg->sopt = *sopt;
1754
1755         /* Fixup the sopt.sopt_val ptr */
1756         error = sooptcopyin(sopt, &msg->sopt_val,
1757             sizeof(msg->sopt_val), sizeof(msg->sopt_val));
1758         if (error) {
1759                 kfree(msg, M_LWKTMSG);
1760                 return NULL;
1761         }
1762         msg->sopt.sopt_val = &msg->sopt_val;
1763
1764         /* Hold the current thread */
1765         if (msg->sopt.sopt_td != NULL) {
1766                 flags |= PRCO_HELDTD;
1767                 lwkt_hold(msg->sopt.sopt_td);
1768         }
1769
1770         msg->ctloutput.nm_flags = flags;
1771         msg->ctloutput.nm_sopt = &msg->sopt;
1772
1773         return &msg->ctloutput;
1774 }
1775
1776 /*
1777  * tcp_sendspace and tcp_recvspace are the default send and receive window
1778  * sizes, respectively.  These are obsolescent (this information should
1779  * be set by the route).
1780  *
1781  * Use a default that does not require tcp window scaling to be turned
1782  * on.  Individual programs or the administrator can increase the default.
1783  */
1784 u_long  tcp_sendspace = 57344;  /* largest multiple of PAGE_SIZE < 64k */
1785 SYSCTL_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_RW,
1786     &tcp_sendspace , 0, "Maximum outgoing TCP datagram size");
1787 u_long  tcp_recvspace = 57344;  /* largest multiple of PAGE_SIZE < 64k */
1788 SYSCTL_INT(_net_inet_tcp, TCPCTL_RECVSPACE, recvspace, CTLFLAG_RW,
1789     &tcp_recvspace , 0, "Maximum incoming TCP datagram size");
1790
1791 /*
1792  * Attach TCP protocol to socket, allocating internet protocol control
1793  * block, tcp control block, buffer space, and entering CLOSED state.
1794  */
1795 static int
1796 tcp_attach(struct socket *so, struct pru_attach_info *ai)
1797 {
1798         struct inpcb *inp;
1799         int error;
1800         int cpu;
1801 #ifdef INET6
1802         boolean_t isipv6 = INP_CHECK_SOCKAF(so, AF_INET6);
1803 #endif
1804
1805         if (ai != NULL) {
1806                 error = tcp_usr_preattach(so, 0 /* don't care */, ai);
1807                 if (error)
1808                         return (error);
1809         } else {
1810                 /* Post attach; do nothing */
1811         }
1812
1813         cpu = mycpu->gd_cpuid;
1814
1815         /*
1816          * Set the default pcbinfo.  This will likely change when we
1817          * bind/connect.
1818          */
1819         error = in_pcballoc(so, &tcbinfo[cpu]);
1820         if (error)
1821                 return (error);
1822         inp = so->so_pcb;
1823 #ifdef INET6
1824         if (isipv6)
1825                 inp->in6p_hops = -1;    /* use kernel default */
1826 #endif
1827         tcp_newtcpcb(inp);
1828         /* Keep a reference for asynchronized pru_rcvd */
1829         soreference(so);
1830         return (0);
1831 }
1832
1833 /*
1834  * Initiate (or continue) disconnect.
1835  * If embryonic state, just send reset (once).
1836  * If in ``let data drain'' option and linger null, just drop.
1837  * Otherwise (hard), mark socket disconnecting and drop
1838  * current input data; switch states based on user close, and
1839  * send segment to peer (with FIN).
1840  */
1841 static struct tcpcb *
1842 tcp_disconnect(struct tcpcb *tp)
1843 {
1844         struct socket *so = tp->t_inpcb->inp_socket;
1845
1846         if (tp->t_state < TCPS_ESTABLISHED) {
1847                 tp = tcp_close(tp);
1848         } else if ((so->so_options & SO_LINGER) && so->so_linger == 0) {
1849                 tp = tcp_drop(tp, 0);
1850         } else {
1851                 lwkt_gettoken(&so->so_rcv.ssb_token);
1852                 soisdisconnecting(so);
1853                 sbflush(&so->so_rcv.sb);
1854                 tp = tcp_usrclosed(tp);
1855                 if (tp)
1856                         tcp_output(tp);
1857                 lwkt_reltoken(&so->so_rcv.ssb_token);
1858         }
1859         return (tp);
1860 }
1861
1862 /*
1863  * User issued close, and wish to trail through shutdown states:
1864  * if never received SYN, just forget it.  If got a SYN from peer,
1865  * but haven't sent FIN, then go to FIN_WAIT_1 state to send peer a FIN.
1866  * If already got a FIN from peer, then almost done; go to LAST_ACK
1867  * state.  In all other cases, have already sent FIN to peer (e.g.
1868  * after PRU_SHUTDOWN), and just have to play tedious game waiting
1869  * for peer to send FIN or not respond to keep-alives, etc.
1870  * We can let the user exit from the close as soon as the FIN is acked.
1871  */
1872 static struct tcpcb *
1873 tcp_usrclosed(struct tcpcb *tp)
1874 {
1875
1876         switch (tp->t_state) {
1877
1878         case TCPS_CLOSED:
1879         case TCPS_LISTEN:
1880                 TCP_STATE_CHANGE(tp, TCPS_CLOSED);
1881                 tp = tcp_close(tp);
1882                 break;
1883
1884         case TCPS_SYN_SENT:
1885         case TCPS_SYN_RECEIVED:
1886                 tp->t_flags |= TF_NEEDFIN;
1887                 break;
1888
1889         case TCPS_ESTABLISHED:
1890                 TCP_STATE_CHANGE(tp, TCPS_FIN_WAIT_1);
1891                 break;
1892
1893         case TCPS_CLOSE_WAIT:
1894                 TCP_STATE_CHANGE(tp, TCPS_LAST_ACK);
1895                 break;
1896         }
1897         if (tp && tp->t_state >= TCPS_FIN_WAIT_2) {
1898                 soisdisconnected(tp->t_inpcb->inp_socket);
1899                 /* To prevent the connection hanging in FIN_WAIT_2 forever. */
1900                 if (tp->t_state == TCPS_FIN_WAIT_2) {
1901                         tcp_callout_reset(tp, tp->tt_2msl, tp->t_maxidle,
1902                             tcp_timer_2msl);
1903                 }
1904         }
1905         return (tp);
1906 }