2 * Copyright (c) 1982, 1986, 1989, 1991, 1993
3 * The Regents of the University of California. All rights reserved.
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
8 * 1. Redistributions of source code must retain the above copyright
9 * notice, this list of conditions and the following disclaimer.
10 * 2. Redistributions in binary form must reproduce the above copyright
11 * notice, this list of conditions and the following disclaimer in the
12 * documentation and/or other materials provided with the distribution.
13 * 3. Neither the name of the University nor the names of its contributors
14 * may be used to endorse or promote products derived from this software
15 * without specific prior written permission.
17 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29 * From: @(#)uipc_usrreq.c 8.3 (Berkeley) 1/4/94
30 * $FreeBSD: src/sys/kern/uipc_usrreq.c,v 1.54.2.10 2003/03/04 17:28:09 nectar Exp $
33 #include <sys/param.h>
34 #include <sys/systm.h>
35 #include <sys/kernel.h>
36 #include <sys/domain.h>
37 #include <sys/fcntl.h>
38 #include <sys/malloc.h> /* XXX must be before <sys/file.h> */
41 #include <sys/filedesc.h>
43 #include <sys/nlookup.h>
44 #include <sys/protosw.h>
45 #include <sys/socket.h>
46 #include <sys/socketvar.h>
47 #include <sys/resourcevar.h>
49 #include <sys/mount.h>
50 #include <sys/sysctl.h>
52 #include <sys/unpcb.h>
53 #include <sys/vnode.h>
54 #include <sys/kern_syscall.h>
55 #include <sys/taskqueue.h>
57 #include <sys/file2.h>
58 #include <sys/spinlock2.h>
59 #include <sys/socketvar2.h>
60 #include <sys/msgport2.h>
63 * Unix communications domain.
67 * rethink name space problems
68 * need a proper out-of-band
72 * Unix domain sockets GC.
74 * It was originally designed to address following three cases:
75 * 1) Receiving unix domain socket can not accept the rights, e.g.
76 * when the so_rcv is full.
77 * 2) Caller of recvmsg(2) does not pass buffer to receive rights.
78 * 3) Unix domain sockets loop reference, e.g. s1 is on s2.so_rcv,
79 * while s2 on s1.so_rcv.
81 * Code under UNP_GC_ALLFILES is intended to address all above three
82 * cases. However, 1) was addressed a long time ago in uipc_send()
83 * (we inheritted the fix from FreeBSD when DragonFly forked). 2)
84 * was addressed in soreceive() by git-e62cfe62. 3) is the only
85 * case that needs GC. The new code (!UNP_GC_ALLFILES) addresses
86 * case 3) in the following way:
87 * - Record the struct file in unpcb, if the Unix domain socket is
88 * passed as one of the rights.
89 * - At GC time, only unpcbs are scanned, and only Unix domain sockets
90 * that are still used as rights are potential GC targets.
93 #define UNP_DETACHED UNP_PRIVATE1
94 #define UNP_CONNECTING UNP_PRIVATE2
95 #define UNP_DROPPED UNP_PRIVATE3
96 #define UNP_MARKER UNP_PRIVATE4
98 #define UNPGC_REF 0x1 /* unpcb has external ref. */
99 #define UNPGC_DEAD 0x2 /* unpcb might be dead. */
100 #define UNPGC_SCANNED 0x4 /* Has been scanned. */
102 #define UNP_GCFILE_MAX 256
104 /* For unp_internalize() and unp_externalize() */
105 CTASSERT(sizeof(struct file *) >= sizeof(int));
107 #define UNP_ISATTACHED(unp) \
108 ((unp) != NULL && ((unp)->unp_flags & UNP_DETACHED) == 0)
111 #define UNP_ASSERT_TOKEN_HELD(unp) \
112 ASSERT_LWKT_TOKEN_HELD(lwkt_token_pool_lookup((unp)))
113 #else /* !INVARIANTS */
114 #define UNP_ASSERT_TOKEN_HELD(unp)
115 #endif /* INVARIANTS */
117 struct unp_defdiscard {
118 SLIST_ENTRY(unp_defdiscard) next;
121 SLIST_HEAD(unp_defdiscard_list, unp_defdiscard);
123 TAILQ_HEAD(unpcb_qhead, unpcb);
124 struct unp_global_head {
125 struct unpcb_qhead list;
129 static MALLOC_DEFINE(M_UNPCB, "unpcb", "unpcb struct");
130 static unp_gen_t unp_gencnt;
132 static struct unp_global_head unp_stream_head;
133 static struct unp_global_head unp_dgram_head;
134 static struct unp_global_head unp_seqpkt_head;
136 static struct unp_global_head * const unp_heads[] =
137 { &unp_stream_head, &unp_dgram_head, &unp_seqpkt_head, NULL };
139 static struct lwkt_token unp_token = LWKT_TOKEN_INITIALIZER(unp_token);
140 static struct taskqueue *unp_taskqueue;
142 static struct unp_defdiscard_list unp_defdiscard_head;
143 static struct spinlock unp_defdiscard_spin;
144 static struct task unp_defdiscard_task;
146 static struct sockaddr sun_noname = { sizeof(sun_noname), AF_LOCAL };
148 static int unp_attach (struct socket *, struct pru_attach_info *);
149 static void unp_detach (struct unpcb *);
150 static int unp_bind (struct unpcb *,struct sockaddr *, struct thread *);
151 static int unp_connect (struct socket *,struct sockaddr *,
153 static void unp_disconnect(struct unpcb *, int);
154 static void unp_shutdown (struct unpcb *);
155 static void unp_gc(void *, int);
156 #ifdef UNP_GC_ALLFILES
157 static int unp_gc_clearmarks(struct file *, void *);
158 static int unp_gc_checkmarks(struct file *, void *);
159 static int unp_gc_checkrefs(struct file *, void *);
160 static void unp_mark(struct file *, void *data);
162 static void unp_scan (struct mbuf *, void (*)(struct file *, void *),
164 static void unp_discard (struct file *, void *);
165 static int unp_internalize (struct mbuf *, struct thread *);
166 static int unp_listen (struct unpcb *, struct thread *);
167 static void unp_fp_externalize(struct lwp *lp, struct file *fp, int fd,
169 static int unp_find_lockref(struct sockaddr *nam, struct thread *td,
170 short type, struct unpcb **unp_ret);
171 static int unp_connect_pair(struct unpcb *unp, struct unpcb *unp2);
172 static void unp_drop(struct unpcb *unp, int error);
173 static void unp_defdiscard_taskfunc(void *, int);
175 static int unp_rights; /* file descriptors in flight */
176 static struct lwkt_token unp_rights_token =
177 LWKT_TOKEN_INITIALIZER(unp_rights_token);
178 static struct task unp_gc_task;
179 static struct unpcb *unp_gc_marker;
181 SYSCTL_DECL(_net_local);
182 SYSCTL_INT(_net_local, OID_AUTO, inflight, CTLFLAG_RD, &unp_rights, 0,
183 "File descriptors in flight");
186 * SMP Considerations:
188 * Since unp_token will be automaticly released upon execution of
189 * blocking code, we need to reference unp_conn before any possible
190 * blocking code to prevent it from being ripped behind our back.
192 * Any adjustment to unp->unp_conn requires both the global unp_token
193 * AND the per-unp token (lwkt_token_pool_lookup(unp)) to be held.
195 * Any access to so_pcb to obtain unp requires the pool token for
200 unp_reference(struct unpcb *unp)
202 /* 0->1 transition will not work */
203 KKASSERT(unp->unp_refcnt > 0);
204 atomic_add_int(&unp->unp_refcnt, 1);
208 unp_free(struct unpcb *unp)
210 KKASSERT(unp->unp_refcnt > 0);
211 if (atomic_fetchadd_int(&unp->unp_refcnt, -1) == 1)
215 static __inline struct unpcb *
216 unp_getsocktoken(struct socket *so)
221 * The unp pointer is invalid until we verify that it is
222 * good by re-checking so_pcb AFTER obtaining the token.
224 while ((unp = so->so_pcb) != NULL) {
225 lwkt_getpooltoken(unp);
226 if (unp == so->so_pcb)
228 lwkt_relpooltoken(unp);
234 unp_reltoken(struct unpcb *unp)
237 lwkt_relpooltoken(unp);
241 unp_setflags(struct unpcb *unp, int flags)
243 atomic_set_int(&unp->unp_flags, flags);
247 unp_clrflags(struct unpcb *unp, int flags)
249 atomic_clear_int(&unp->unp_flags, flags);
252 static __inline struct unp_global_head *
253 unp_globalhead(short type)
257 return &unp_stream_head;
259 return &unp_dgram_head;
261 return &unp_seqpkt_head;
263 panic("unknown socket type %d", type);
267 static __inline struct unpcb *
268 unp_fp2unpcb(struct file *fp)
272 if (fp->f_type != DTYPE_SOCKET)
279 if (so->so_proto->pr_domain != &localdomain)
286 unp_add_right(struct file *fp)
290 ASSERT_LWKT_TOKEN_HELD(&unp_rights_token);
291 KASSERT(fp->f_count > 0, ("invalid f_count %d", fp->f_count));
293 unp = unp_fp2unpcb(fp);
303 unp_del_right(struct file *fp)
307 ASSERT_LWKT_TOKEN_HELD(&unp_rights_token);
308 KASSERT(fp->f_count > 0, ("invalid f_count %d", fp->f_count));
310 unp = unp_fp2unpcb(fp);
312 KASSERT(unp->unp_msgcount > 0,
313 ("invalid unp msgcount %d", unp->unp_msgcount));
315 if (unp->unp_msgcount == 0)
323 * NOTE: (so) is referenced from soabort*() and netmsg_pru_abort()
324 * will sofree() it when we return.
327 uipc_abort(netmsg_t msg)
332 lwkt_gettoken(&unp_token);
333 unp = unp_getsocktoken(msg->base.nm_so);
335 if (UNP_ISATTACHED(unp)) {
336 unp_drop(unp, ECONNABORTED);
343 lwkt_reltoken(&unp_token);
345 lwkt_replymsg(&msg->lmsg, error);
349 uipc_accept(netmsg_t msg)
354 lwkt_gettoken(&unp_token);
355 unp = unp_getsocktoken(msg->base.nm_so);
357 if (!UNP_ISATTACHED(unp)) {
360 struct unpcb *unp2 = unp->unp_conn;
363 * Pass back name of connected socket,
364 * if it was bound and we are still connected
365 * (our peer may have closed already!).
367 if (unp2 && unp2->unp_addr) {
369 *msg->accept.nm_nam = dup_sockaddr(
370 (struct sockaddr *)unp2->unp_addr);
373 *msg->accept.nm_nam = dup_sockaddr(&sun_noname);
379 lwkt_reltoken(&unp_token);
381 lwkt_replymsg(&msg->lmsg, error);
385 uipc_attach(netmsg_t msg)
389 lwkt_gettoken(&unp_token);
391 KASSERT(msg->base.nm_so->so_pcb == NULL, ("double unp attach"));
392 error = unp_attach(msg->base.nm_so, msg->attach.nm_ai);
394 lwkt_reltoken(&unp_token);
395 lwkt_replymsg(&msg->lmsg, error);
399 uipc_bind(netmsg_t msg)
404 lwkt_gettoken(&unp_token);
405 unp = unp_getsocktoken(msg->base.nm_so);
407 if (UNP_ISATTACHED(unp))
408 error = unp_bind(unp, msg->bind.nm_nam, msg->bind.nm_td);
413 lwkt_reltoken(&unp_token);
415 lwkt_replymsg(&msg->lmsg, error);
419 uipc_connect(netmsg_t msg)
423 error = unp_connect(msg->base.nm_so, msg->connect.nm_nam,
425 lwkt_replymsg(&msg->lmsg, error);
429 uipc_connect2(netmsg_t msg)
433 error = unp_connect2(msg->connect2.nm_so1, msg->connect2.nm_so2,
434 msg->connect2.nm_cred);
435 lwkt_replymsg(&msg->lmsg, error);
438 /* control is EOPNOTSUPP */
441 uipc_detach(netmsg_t msg)
446 lwkt_gettoken(&unp_token);
447 unp = unp_getsocktoken(msg->base.nm_so);
449 if (UNP_ISATTACHED(unp)) {
457 lwkt_reltoken(&unp_token);
459 lwkt_replymsg(&msg->lmsg, error);
463 uipc_disconnect(netmsg_t msg)
468 lwkt_gettoken(&unp_token);
469 unp = unp_getsocktoken(msg->base.nm_so);
471 if (UNP_ISATTACHED(unp)) {
472 unp_disconnect(unp, 0);
479 lwkt_reltoken(&unp_token);
481 lwkt_replymsg(&msg->lmsg, error);
485 uipc_listen(netmsg_t msg)
490 lwkt_gettoken(&unp_token);
491 unp = unp_getsocktoken(msg->base.nm_so);
493 if (!UNP_ISATTACHED(unp) || unp->unp_vnode == NULL)
496 error = unp_listen(unp, msg->listen.nm_td);
499 lwkt_reltoken(&unp_token);
501 lwkt_replymsg(&msg->lmsg, error);
505 uipc_peeraddr(netmsg_t msg)
510 lwkt_gettoken(&unp_token);
511 unp = unp_getsocktoken(msg->base.nm_so);
513 if (!UNP_ISATTACHED(unp)) {
515 } else if (unp->unp_conn && unp->unp_conn->unp_addr) {
516 struct unpcb *unp2 = unp->unp_conn;
519 *msg->peeraddr.nm_nam = dup_sockaddr(
520 (struct sockaddr *)unp2->unp_addr);
525 * XXX: It seems that this test always fails even when
526 * connection is established. So, this else clause is
527 * added as workaround to return PF_LOCAL sockaddr.
529 *msg->peeraddr.nm_nam = dup_sockaddr(&sun_noname);
534 lwkt_reltoken(&unp_token);
536 lwkt_replymsg(&msg->lmsg, error);
540 uipc_rcvd(netmsg_t msg)
542 struct unpcb *unp, *unp2;
548 * so_pcb is only modified with both the global and the unp
551 so = msg->base.nm_so;
552 unp = unp_getsocktoken(so);
554 if (!UNP_ISATTACHED(unp)) {
559 switch (so->so_type) {
561 panic("uipc_rcvd DGRAM?");
565 if (unp->unp_conn == NULL)
567 unp2 = unp->unp_conn; /* protected by pool token */
570 * Because we are transfering mbufs directly to the
571 * peer socket we have to use SSB_STOP on the sender
572 * to prevent it from building up infinite mbufs.
574 * As in several places in this module w ehave to ref unp2
575 * to ensure that it does not get ripped out from under us
576 * if we block on the so2 token or in sowwakeup().
578 so2 = unp2->unp_socket;
580 lwkt_gettoken(&so2->so_rcv.ssb_token);
581 if (so->so_rcv.ssb_cc < so2->so_snd.ssb_hiwat &&
582 so->so_rcv.ssb_mbcnt < so2->so_snd.ssb_mbmax
584 atomic_clear_int(&so2->so_snd.ssb_flags, SSB_STOP);
588 lwkt_reltoken(&so2->so_rcv.ssb_token);
592 panic("uipc_rcvd unknown socktype");
598 lwkt_replymsg(&msg->lmsg, error);
601 /* pru_rcvoob is EOPNOTSUPP */
604 uipc_send(netmsg_t msg)
606 struct unpcb *unp, *unp2;
609 struct mbuf *control;
613 so = msg->base.nm_so;
614 control = msg->send.nm_control;
618 * so_pcb is only modified with both the global and the unp
621 so = msg->base.nm_so;
622 unp = unp_getsocktoken(so);
624 if (!UNP_ISATTACHED(unp)) {
629 if (msg->send.nm_flags & PRUS_OOB) {
634 wakeup_start_delayed();
636 if (control && (error = unp_internalize(control, msg->send.nm_td)))
639 switch (so->so_type) {
642 struct sockaddr *from;
644 if (msg->send.nm_addr) {
649 lwkt_gettoken(&unp_token);
650 error = unp_find_lockref(msg->send.nm_addr,
651 msg->send.nm_td, so->so_type, &unp2);
653 lwkt_reltoken(&unp_token);
658 * unp2 is locked and referenced.
660 * We could unlock unp2 now, since it was checked
664 lwkt_reltoken(&unp_token);
666 if (unp->unp_conn == NULL) {
670 unp2 = unp->unp_conn;
673 /* NOTE: unp2 is referenced. */
674 so2 = unp2->unp_socket;
677 * Include creds if the receive side wants them, even if
678 * the send side did not send them.
680 if (so2->so_options & SO_PASSCRED) {
683 struct cmsgcred cred;
687 while ((ncon = *mp) != NULL) {
688 cm = mtod(ncon, struct cmsghdr *);
689 if (cm->cmsg_type == SCM_CREDS &&
690 cm->cmsg_level == SOL_SOCKET)
695 ncon = sbcreatecontrol(&cred, sizeof(cred),
696 SCM_CREDS, SOL_SOCKET);
697 unp_internalize(ncon, msg->send.nm_td);
703 from = (struct sockaddr *)unp->unp_addr;
707 lwkt_gettoken(&so2->so_rcv.ssb_token);
708 if (ssb_appendaddr(&so2->so_rcv, from, m, control)) {
715 lwkt_reltoken(&so2->so_rcv.ssb_token);
723 /* Connect if not connected yet. */
725 * Note: A better implementation would complain
726 * if not equal to the peer's address.
728 if (unp->unp_conn == NULL) {
729 if (msg->send.nm_addr) {
730 error = unp_connect(so,
738 * unp_conn still could be NULL, even if the
739 * above unp_connect() succeeds; since the
740 * current unp's token could be released due
741 * to blocking operations after unp_conn is
744 if (unp->unp_conn == NULL) {
749 if (so->so_state & SS_CANTSENDMORE) {
754 unp2 = unp->unp_conn;
755 KASSERT(unp2 != NULL, ("unp is not connected"));
756 so2 = unp2->unp_socket;
761 * Send to paired receive port, and then reduce
762 * send buffer hiwater marks to maintain backpressure.
765 lwkt_gettoken(&so2->so_rcv.ssb_token);
767 if (ssb_appendcontrol(&so2->so_rcv, m, control)) {
771 } else if (so->so_type == SOCK_SEQPACKET) {
772 sbappendrecord(&so2->so_rcv.sb, m);
775 sbappend(&so2->so_rcv.sb, m);
780 * Because we are transfering mbufs directly to the
781 * peer socket we have to use SSB_STOP on the sender
782 * to prevent it from building up infinite mbufs.
784 if (so2->so_rcv.ssb_cc >= so->so_snd.ssb_hiwat ||
785 so2->so_rcv.ssb_mbcnt >= so->so_snd.ssb_mbmax
787 atomic_set_int(&so->so_snd.ssb_flags, SSB_STOP);
789 lwkt_reltoken(&so2->so_rcv.ssb_token);
796 panic("uipc_send unknown socktype");
800 * SEND_EOF is equivalent to a SEND followed by a SHUTDOWN.
802 if (msg->send.nm_flags & PRUS_EOF) {
807 if (control && error != 0)
808 unp_dispose(control);
811 wakeup_end_delayed();
817 lwkt_replymsg(&msg->lmsg, error);
824 uipc_sense(netmsg_t msg)
831 so = msg->base.nm_so;
832 sb = msg->sense.nm_stat;
835 * so_pcb is only modified with both the global and the unp
838 unp = unp_getsocktoken(so);
840 if (!UNP_ISATTACHED(unp)) {
845 sb->st_blksize = so->so_snd.ssb_hiwat;
850 lwkt_replymsg(&msg->lmsg, error);
854 uipc_shutdown(netmsg_t msg)
861 * so_pcb is only modified with both the global and the unp
864 so = msg->base.nm_so;
865 unp = unp_getsocktoken(so);
867 if (UNP_ISATTACHED(unp)) {
876 lwkt_replymsg(&msg->lmsg, error);
880 uipc_sockaddr(netmsg_t msg)
886 * so_pcb is only modified with both the global and the unp
889 unp = unp_getsocktoken(msg->base.nm_so);
891 if (UNP_ISATTACHED(unp)) {
893 *msg->sockaddr.nm_nam =
894 dup_sockaddr((struct sockaddr *)unp->unp_addr);
902 lwkt_replymsg(&msg->lmsg, error);
905 struct pr_usrreqs uipc_usrreqs = {
906 .pru_abort = uipc_abort,
907 .pru_accept = uipc_accept,
908 .pru_attach = uipc_attach,
909 .pru_bind = uipc_bind,
910 .pru_connect = uipc_connect,
911 .pru_connect2 = uipc_connect2,
912 .pru_control = pr_generic_notsupp,
913 .pru_detach = uipc_detach,
914 .pru_disconnect = uipc_disconnect,
915 .pru_listen = uipc_listen,
916 .pru_peeraddr = uipc_peeraddr,
917 .pru_rcvd = uipc_rcvd,
918 .pru_rcvoob = pr_generic_notsupp,
919 .pru_send = uipc_send,
920 .pru_sense = uipc_sense,
921 .pru_shutdown = uipc_shutdown,
922 .pru_sockaddr = uipc_sockaddr,
923 .pru_sosend = sosend,
924 .pru_soreceive = soreceive
928 uipc_ctloutput(netmsg_t msg)
931 struct sockopt *sopt;
935 so = msg->base.nm_so;
936 sopt = msg->ctloutput.nm_sopt;
938 lwkt_gettoken(&unp_token);
939 unp = unp_getsocktoken(so);
941 if (!UNP_ISATTACHED(unp)) {
946 switch (sopt->sopt_dir) {
948 switch (sopt->sopt_name) {
950 if (unp->unp_flags & UNP_HAVEPC)
951 soopt_from_kbuf(sopt, &unp->unp_peercred,
952 sizeof(unp->unp_peercred));
954 if (so->so_type == SOCK_STREAM)
956 else if (so->so_type == SOCK_SEQPACKET)
975 lwkt_reltoken(&unp_token);
977 lwkt_replymsg(&msg->lmsg, error);
981 * Both send and receive buffers are allocated PIPSIZ bytes of buffering
982 * for stream sockets, although the total for sender and receiver is
983 * actually only PIPSIZ.
985 * Datagram sockets really use the sendspace as the maximum datagram size,
986 * and don't really want to reserve the sendspace. Their recvspace should
987 * be large enough for at least one max-size datagram plus address.
989 * We want the local send/recv space to be significant larger then lo0's
992 * We no longer need to worry about avoiding the windows scaling option.
993 * Programs which use unix domain sockets expect larger defaults these days.
998 static u_long unpst_sendspace = PIPSIZ;
999 static u_long unpst_recvspace = PIPSIZ;
1000 static u_long unpdg_sendspace = PIPSIZ; /* really max datagram size */
1001 static u_long unpdg_recvspace = PIPSIZ;
1002 static u_long unpsp_sendspace = PIPSIZ; /* really max datagram size */
1003 static u_long unpsp_recvspace = PIPSIZ;
1005 SYSCTL_DECL(_net_local_stream);
1006 SYSCTL_DECL(_net_local_dgram);
1007 SYSCTL_DECL(_net_local_seqpacket);
1009 SYSCTL_ULONG(_net_local_stream, OID_AUTO, sendspace, CTLFLAG_RW,
1010 &unpst_sendspace, 0, "Size of stream socket send buffer");
1011 SYSCTL_ULONG(_net_local_stream, OID_AUTO, recvspace, CTLFLAG_RW,
1012 &unpst_recvspace, 0, "Size of stream socket receive buffer");
1014 SYSCTL_ULONG(_net_local_dgram, OID_AUTO, maxdgram, CTLFLAG_RW,
1015 &unpdg_sendspace, 0, "Max datagram socket size");
1016 SYSCTL_ULONG(_net_local_dgram, OID_AUTO, recvspace, CTLFLAG_RW,
1017 &unpdg_recvspace, 0, "Size of datagram socket receive buffer");
1019 SYSCTL_ULONG(_net_local_seqpacket, OID_AUTO, maxseqpacket, CTLFLAG_RW,
1020 &unpsp_sendspace, 0, "Default seqpacket send space.");
1021 SYSCTL_ULONG(_net_local_seqpacket, OID_AUTO, recvspace, CTLFLAG_RW,
1022 &unpsp_recvspace, 0, "Default seqpacket receive space.");
1026 unp_attach(struct socket *so, struct pru_attach_info *ai)
1028 struct unp_global_head *head;
1032 lwkt_gettoken(&unp_token);
1034 if (so->so_snd.ssb_hiwat == 0 || so->so_rcv.ssb_hiwat == 0) {
1035 switch (so->so_type) {
1037 error = soreserve(so, unpst_sendspace, unpst_recvspace,
1041 error = soreserve(so, unpdg_sendspace, unpdg_recvspace,
1044 case SOCK_SEQPACKET:
1045 error = soreserve(so, unpsp_sendspace, unpsp_recvspace,
1049 panic("unp_attach");
1056 * In order to support sendfile we have to set either SSB_STOPSUPP
1057 * or SSB_PREALLOC. Unix domain sockets use the SSB_STOP flow
1058 * control mechanism.
1060 if (so->so_type == SOCK_STREAM) {
1061 atomic_set_int(&so->so_rcv.ssb_flags, SSB_STOPSUPP);
1062 atomic_set_int(&so->so_snd.ssb_flags, SSB_STOPSUPP);
1065 unp = kmalloc(sizeof(*unp), M_UNPCB, M_WAITOK | M_ZERO | M_NULLOK);
1070 unp->unp_refcnt = 1;
1071 unp->unp_gencnt = ++unp_gencnt;
1072 LIST_INIT(&unp->unp_refs);
1073 unp->unp_socket = so;
1074 unp->unp_rvnode = ai->fd_rdir; /* jail cruft XXX JH */
1075 so->so_pcb = (caddr_t)unp;
1078 head = unp_globalhead(so->so_type);
1079 TAILQ_INSERT_TAIL(&head->list, unp, unp_link);
1083 lwkt_reltoken(&unp_token);
1088 unp_detach(struct unpcb *unp)
1092 lwkt_gettoken(&unp_token);
1093 lwkt_getpooltoken(unp);
1095 so = unp->unp_socket;
1097 unp->unp_gencnt = ++unp_gencnt;
1098 if (unp->unp_vnode) {
1099 unp->unp_vnode->v_socket = NULL;
1100 vrele(unp->unp_vnode);
1101 unp->unp_vnode = NULL;
1103 soisdisconnected(so);
1104 KKASSERT(so->so_pcb == unp);
1105 so->so_pcb = NULL; /* both tokens required */
1106 unp->unp_socket = NULL;
1108 lwkt_relpooltoken(unp);
1109 lwkt_reltoken(&unp_token);
1113 KASSERT(unp->unp_conn == NULL, ("unp is still connected"));
1114 KASSERT(LIST_EMPTY(&unp->unp_refs), ("unp still has references"));
1117 kfree(unp->unp_addr, M_SONAME);
1118 kfree(unp, M_UNPCB);
1121 taskqueue_enqueue(unp_taskqueue, &unp_gc_task);
1125 unp_bind(struct unpcb *unp, struct sockaddr *nam, struct thread *td)
1127 struct proc *p = td->td_proc;
1128 struct sockaddr_un *soun = (struct sockaddr_un *)nam;
1132 struct nlookupdata nd;
1133 char buf[SOCK_MAXADDRLEN];
1135 ASSERT_LWKT_TOKEN_HELD(&unp_token);
1136 UNP_ASSERT_TOKEN_HELD(unp);
1138 if (unp->unp_vnode != NULL)
1141 namelen = soun->sun_len - offsetof(struct sockaddr_un, sun_path);
1144 strncpy(buf, soun->sun_path, namelen);
1145 buf[namelen] = 0; /* null-terminate the string */
1146 error = nlookup_init(&nd, buf, UIO_SYSSPACE,
1147 NLC_LOCKVP | NLC_CREATE | NLC_REFDVP);
1149 error = nlookup(&nd);
1150 if (error == 0 && nd.nl_nch.ncp->nc_vp != NULL)
1152 if (error == 0 && nd.nl_dvp == NULL) /* e.g. bind <mountpt> */
1158 vattr.va_type = VSOCK;
1159 vattr.va_mode = (ACCESSPERMS & ~p->p_fd->fd_cmask);
1160 error = VOP_NCREATE(&nd.nl_nch, nd.nl_dvp, &vp, nd.nl_cred, &vattr);
1162 if (unp->unp_vnode == NULL) {
1163 vp->v_socket = unp->unp_socket;
1164 unp->unp_vnode = vp;
1165 unp->unp_addr = (struct sockaddr_un *)dup_sockaddr(nam);
1168 vput(vp); /* late race */
1178 unp_connect(struct socket *so, struct sockaddr *nam, struct thread *td)
1180 struct unpcb *unp, *unp2;
1181 int error, flags = 0;
1183 lwkt_gettoken(&unp_token);
1185 unp = unp_getsocktoken(so);
1186 if (!UNP_ISATTACHED(unp)) {
1191 if ((unp->unp_flags & UNP_CONNECTING) || unp->unp_conn != NULL) {
1196 flags = UNP_CONNECTING;
1197 unp_setflags(unp, flags);
1199 error = unp_find_lockref(nam, td, so->so_type, &unp2);
1204 * unp2 is locked and referenced.
1207 if (so->so_proto->pr_flags & PR_CONNREQUIRED) {
1208 struct socket *so2, *so3;
1211 so2 = unp2->unp_socket;
1212 if (!(so2->so_options & SO_ACCEPTCONN) ||
1213 /* listen is not completed yet */
1214 !(unp2->unp_flags & UNP_HAVEPCCACHED) ||
1215 (so3 = sonewconn_faddr(so2, 0, NULL,
1216 TRUE /* keep ref */)) == NULL) {
1217 error = ECONNREFUSED;
1220 /* so3 has a socket reference. */
1222 unp3 = unp_getsocktoken(so3);
1223 if (!UNP_ISATTACHED(unp3)) {
1226 * Already aborted; we only need to drop the
1227 * socket reference held by sonewconn_faddr().
1230 error = ECONNREFUSED;
1233 unp_reference(unp3);
1236 * unp3 is locked and referenced.
1240 * Release so3 socket reference held by sonewconn_faddr().
1241 * Since we have referenced unp3, neither unp3 nor so3 will
1242 * be destroyed here.
1246 if (unp2->unp_addr != NULL) {
1247 unp3->unp_addr = (struct sockaddr_un *)
1248 dup_sockaddr((struct sockaddr *)unp2->unp_addr);
1252 * unp_peercred management:
1254 * The connecter's (client's) credentials are copied
1255 * from its process structure at the time of connect()
1258 cru2x(td->td_proc->p_ucred, &unp3->unp_peercred);
1259 unp_setflags(unp3, UNP_HAVEPC);
1261 * The receiver's (server's) credentials are copied
1262 * from the unp_peercred member of socket on which the
1263 * former called listen(); unp_listen() cached that
1264 * process's credentials at that time so we can use
1267 KASSERT(unp2->unp_flags & UNP_HAVEPCCACHED,
1268 ("unp_connect: listener without cached peercred"));
1269 memcpy(&unp->unp_peercred, &unp2->unp_peercred,
1270 sizeof(unp->unp_peercred));
1271 unp_setflags(unp, UNP_HAVEPC);
1273 error = unp_connect_pair(unp, unp3);
1275 soabort_direct(so3);
1277 /* Done with unp3 */
1281 error = unp_connect_pair(unp, unp2);
1288 unp_clrflags(unp, flags);
1291 lwkt_reltoken(&unp_token);
1296 * Connect two unix domain sockets together.
1298 * NOTE: Semantics for any change to unp_conn requires that the per-unp
1299 * pool token also be held.
1302 unp_connect2(struct socket *so, struct socket *so2, struct ucred *cred)
1304 struct unpcb *unp, *unp2;
1307 lwkt_gettoken(&unp_token);
1308 if (so2->so_type != so->so_type) {
1309 lwkt_reltoken(&unp_token);
1310 return (EPROTOTYPE);
1312 unp = unp_getsocktoken(so);
1313 unp2 = unp_getsocktoken(so2);
1315 cru2x(cred, &unp->unp_peercred);
1316 cru2x(cred, &unp2->unp_peercred);
1317 unp_setflags(unp, UNP_HAVEPC);
1318 unp_setflags(unp2, UNP_HAVEPC);
1320 if (!UNP_ISATTACHED(unp)) {
1324 if (!UNP_ISATTACHED(unp2)) {
1325 error = ECONNREFUSED;
1329 if (unp->unp_conn != NULL) {
1333 if ((so->so_type == SOCK_STREAM || so->so_type == SOCK_SEQPACKET) &&
1334 unp2->unp_conn != NULL) {
1339 error = unp_connect_pair(unp, unp2);
1343 lwkt_reltoken(&unp_token);
1348 * Disconnect a unix domain socket pair.
1350 * NOTE: Semantics for any change to unp_conn requires that the per-unp
1351 * pool token also be held.
1354 unp_disconnect(struct unpcb *unp, int error)
1356 struct socket *so = unp->unp_socket;
1359 ASSERT_LWKT_TOKEN_HELD(&unp_token);
1360 UNP_ASSERT_TOKEN_HELD(unp);
1363 so->so_error = error;
1365 while ((unp2 = unp->unp_conn) != NULL) {
1366 lwkt_getpooltoken(unp2);
1367 if (unp2 == unp->unp_conn)
1369 lwkt_relpooltoken(unp2);
1373 /* unp2 is locked. */
1375 KASSERT((unp2->unp_flags & UNP_DROPPED) == 0, ("unp2 was dropped"));
1377 unp->unp_conn = NULL;
1379 switch (so->so_type) {
1381 LIST_REMOVE(unp, unp_reflink);
1382 soclrstate(so, SS_ISCONNECTED);
1386 case SOCK_SEQPACKET:
1388 * Keep a reference before clearing the unp_conn
1389 * to avoid racing uipc_detach()/uipc_abort() in
1392 unp_reference(unp2);
1393 KASSERT(unp2->unp_conn == unp, ("unp_conn mismatch"));
1394 unp2->unp_conn = NULL;
1396 soisdisconnected(so);
1397 soisdisconnected(unp2->unp_socket);
1403 lwkt_relpooltoken(unp2);
1408 unp_abort(struct unpcb *unp)
1410 lwkt_gettoken(&unp_token);
1412 lwkt_reltoken(&unp_token);
1417 prison_unpcb(struct thread *td, struct unpcb *unp)
1423 if ((p = td->td_proc) == NULL)
1425 if (!p->p_ucred->cr_prison)
1427 if (p->p_fd->fd_rdir == unp->unp_rvnode)
1433 unp_pcblist(SYSCTL_HANDLER_ARGS)
1435 struct unp_global_head *head = arg1;
1437 struct unpcb *unp, *marker;
1439 KKASSERT(curproc != NULL);
1442 * The process of preparing the PCB list is too time-consuming and
1443 * resource-intensive to repeat twice on every request.
1445 if (req->oldptr == NULL) {
1447 req->oldidx = (n + n/8) * sizeof(struct xunpcb);
1451 if (req->newptr != NULL)
1454 marker = kmalloc(sizeof(*marker), M_UNPCB, M_WAITOK | M_ZERO);
1455 marker->unp_flags |= UNP_MARKER;
1457 lwkt_gettoken(&unp_token);
1463 TAILQ_INSERT_HEAD(&head->list, marker, unp_link);
1464 while ((unp = TAILQ_NEXT(marker, unp_link)) != NULL && i < n) {
1467 TAILQ_REMOVE(&head->list, marker, unp_link);
1468 TAILQ_INSERT_AFTER(&head->list, unp, marker, unp_link);
1470 if (unp->unp_flags & UNP_MARKER)
1472 if (prison_unpcb(req->td, unp))
1475 xu.xu_len = sizeof(xu);
1480 * unp->unp_addr and unp->unp_conn are protected by
1481 * unp_token. So if we want to get rid of unp_token
1482 * or reduce the coverage of unp_token, care must be
1485 if (unp->unp_addr) {
1486 bcopy(unp->unp_addr, &xu.xu_addr,
1487 unp->unp_addr->sun_len);
1489 if (unp->unp_conn && unp->unp_conn->unp_addr) {
1490 bcopy(unp->unp_conn->unp_addr,
1492 unp->unp_conn->unp_addr->sun_len);
1494 bcopy(unp, &xu.xu_unp, sizeof(*unp));
1495 sotoxsocket(unp->unp_socket, &xu.xu_socket);
1497 /* NOTE: This could block and temporarily release unp_token */
1498 error = SYSCTL_OUT(req, &xu, sizeof(xu));
1503 TAILQ_REMOVE(&head->list, marker, unp_link);
1505 lwkt_reltoken(&unp_token);
1507 kfree(marker, M_UNPCB);
1511 SYSCTL_PROC(_net_local_dgram, OID_AUTO, pcblist, CTLFLAG_RD,
1512 &unp_dgram_head, 0, unp_pcblist, "S,xunpcb",
1513 "List of active local datagram sockets");
1514 SYSCTL_PROC(_net_local_stream, OID_AUTO, pcblist, CTLFLAG_RD,
1515 &unp_stream_head, 0, unp_pcblist, "S,xunpcb",
1516 "List of active local stream sockets");
1517 SYSCTL_PROC(_net_local_seqpacket, OID_AUTO, pcblist, CTLFLAG_RD,
1518 &unp_seqpkt_head, 0, unp_pcblist, "S,xunpcb",
1519 "List of active local seqpacket sockets");
1522 unp_shutdown(struct unpcb *unp)
1526 if ((unp->unp_socket->so_type == SOCK_STREAM ||
1527 unp->unp_socket->so_type == SOCK_SEQPACKET) &&
1528 unp->unp_conn != NULL && (so = unp->unp_conn->unp_socket)) {
1537 lwkt_gettoken(&unp_token);
1538 lwkt_reltoken(&unp_token);
1543 unp_externalize(struct mbuf *rights, int flags)
1545 struct thread *td = curthread;
1546 struct proc *p = td->td_proc; /* XXX */
1547 struct lwp *lp = td->td_lwp;
1548 struct cmsghdr *cm = mtod(rights, struct cmsghdr *);
1553 int newfds = (cm->cmsg_len - (CMSG_DATA(cm) - (u_char *)cm))
1554 / sizeof(struct file *);
1557 lwkt_gettoken(&unp_rights_token);
1560 * if the new FD's will not fit, then we free them all
1562 if (!fdavail(p, newfds)) {
1563 rp = (struct file **)CMSG_DATA(cm);
1564 for (i = 0; i < newfds; i++) {
1567 * zero the pointer before calling unp_discard,
1568 * since it may end up in unp_gc()..
1571 unp_discard(fp, NULL);
1573 lwkt_reltoken(&unp_rights_token);
1578 * now change each pointer to an fd in the global table to
1579 * an integer that is the index to the local fd table entry
1580 * that we set up to point to the global one we are transferring.
1581 * Since the sizeof(struct file *) is bigger than or equal to
1582 * the sizeof(int), we do it in forward order. In that case,
1583 * an integer will always come in the same place or before its
1584 * corresponding struct file pointer.
1586 * Hold revoke_token in 'shared' mode, so that we won't miss
1587 * the FREVOKED update on fps being externalized (fsetfd).
1589 lwkt_gettoken_shared(&revoke_token);
1590 fdp = (int *)CMSG_DATA(cm);
1591 rp = (struct file **)CMSG_DATA(cm);
1592 for (i = 0; i < newfds; i++) {
1593 if (fdalloc(p, 0, &f)) {
1597 * Previous fdavail() can't garantee
1598 * fdalloc() success due to SMP race.
1599 * Just clean up and return the same
1600 * error value as if fdavail() failed.
1602 lwkt_reltoken(&revoke_token);
1604 /* Close externalized files */
1605 for (j = 0; j < i; j++)
1607 /* Discard the rest of internal files */
1608 for (; i < newfds; i++)
1609 unp_discard(rp[i], NULL);
1610 /* Wipe out the control message */
1611 for (i = 0; i < newfds; i++)
1614 lwkt_reltoken(&unp_rights_token);
1618 unp_fp_externalize(lp, fp, f, flags);
1621 lwkt_reltoken(&revoke_token);
1623 lwkt_reltoken(&unp_rights_token);
1626 * Adjust length, in case sizeof(struct file *) and sizeof(int)
1629 cm->cmsg_len = CMSG_LEN(newfds * sizeof(int));
1630 rights->m_len = cm->cmsg_len;
1636 unp_fp_externalize(struct lwp *lp, struct file *fp, int fd, int flags)
1639 struct filedesc *fdp = lp->lwp_proc->p_fd;
1642 if (fp->f_flag & FREVOKED) {
1646 kprintf("Warning: revoked fp exiting unix socket\n");
1647 error = falloc(lp, &fx, NULL);
1649 if (flags & MSG_CMSG_CLOEXEC)
1650 fdp->fd_files[fd].fileflags |= UF_EXCLOSE;
1651 fsetfd(fdp, fx, fd);
1654 fsetfd(fdp, NULL, fd);
1657 if (flags & MSG_CMSG_CLOEXEC)
1658 fdp->fd_files[fd].fileflags |= UF_EXCLOSE;
1659 fsetfd(fdp, fp, fd);
1669 TAILQ_INIT(&unp_stream_head.list);
1670 TAILQ_INIT(&unp_dgram_head.list);
1671 TAILQ_INIT(&unp_seqpkt_head.list);
1673 SLIST_INIT(&unp_defdiscard_head);
1674 spin_init(&unp_defdiscard_spin, "unpdisc");
1675 TASK_INIT(&unp_defdiscard_task, 0, unp_defdiscard_taskfunc, NULL);
1678 * This implies that only one gc can be in-progress at any
1681 TASK_INIT(&unp_gc_task, 0, unp_gc, NULL);
1683 unp_gc_marker = kmalloc(sizeof(*unp_gc_marker), M_UNPCB,
1685 unp_gc_marker->unp_flags |= UNP_MARKER;
1688 * Create taskqueue for defered discard, and stick it to
1691 unp_taskqueue = taskqueue_create("unp_taskq", M_WAITOK,
1692 taskqueue_thread_enqueue, &unp_taskqueue);
1693 taskqueue_start_threads(&unp_taskqueue, 1, TDPRI_KERN_DAEMON,
1694 ncpus - 1, "unp taskq");
1698 unp_internalize(struct mbuf *control, struct thread *td)
1700 struct proc *p = td->td_proc;
1701 struct filedesc *fdescp;
1702 struct cmsghdr *cm = mtod(control, struct cmsghdr *);
1706 struct cmsgcred *cmcred;
1714 * Make sure the message is reasonable, and either CREDS or RIGHTS.
1716 * NOTE: overall message length does not have to be aligned, but the
1719 if ((cm->cmsg_type != SCM_RIGHTS && cm->cmsg_type != SCM_CREDS) ||
1720 cm->cmsg_level != SOL_SOCKET ||
1721 control->m_len < sizeof(*cm) || /* control too small */
1722 cm->cmsg_len < sizeof(*cm) || /* cmsg_len too small */
1723 cm->cmsg_len > control->m_len) { /* cmsg_len too big */
1728 * Fill in credential information.
1730 if (cm->cmsg_type == SCM_CREDS) {
1731 cmcred = (struct cmsgcred *)CMSG_DATA(cm);
1732 cmcred->cmcred_pid = p->p_pid;
1733 cmcred->cmcred_uid = p->p_ucred->cr_ruid;
1734 cmcred->cmcred_gid = p->p_ucred->cr_rgid;
1735 cmcred->cmcred_euid = p->p_ucred->cr_uid;
1736 cmcred->cmcred_ngroups = MIN(p->p_ucred->cr_ngroups,
1738 for (i = 0; i < cmcred->cmcred_ngroups; i++)
1739 cmcred->cmcred_groups[i] = p->p_ucred->cr_groups[i];
1744 * cmsghdr may not be aligned, do not allow calculation(s) to
1747 * Data must be aligned but the data length does not have to be.
1749 * If there are multiple headers (XXX not supported) then the
1750 * next header will be aligned after the end of the possibly
1753 if (cm->cmsg_len < CMSG_LEN(0)) {
1757 oldfds = (cm->cmsg_len - CMSG_LEN(0)) / sizeof(int);
1760 * Now replace the integer FDs with pointers to
1761 * the associated global file table entry..
1762 * Allocate a bigger buffer as necessary. But if an cluster is not
1763 * enough, return E2BIG.
1765 newlen = CMSG_LEN(oldfds * sizeof(struct file *));
1766 if (newlen > MCLBYTES)
1768 if (newlen - control->m_len > M_TRAILINGSPACE(control)) {
1769 if (control->m_flags & M_EXT)
1771 MCLGET(control, M_WAITOK);
1773 /* copy the data to the cluster */
1774 memcpy(mtod(control, char *), cm, cm->cmsg_len);
1775 cm = mtod(control, struct cmsghdr *);
1778 lwkt_gettoken(&unp_rights_token);
1781 spin_lock_shared(&fdescp->fd_spin);
1784 * check that all the FDs passed in refer to legal OPEN files
1785 * If not, reject the entire operation.
1787 fdp = (int *)CMSG_DATA(cm);
1788 for (i = 0; i < oldfds; i++) {
1790 if ((unsigned)fd >= fdescp->fd_nfiles ||
1791 fdescp->fd_files[fd].fp == NULL) {
1795 if (fdescp->fd_files[fd].fp->f_type == DTYPE_KQUEUE) {
1802 * Adjust length, in case sizeof(struct file *) and sizeof(int)
1805 cm->cmsg_len = newlen;
1806 control->m_len = CMSG_ALIGN(newlen);
1809 * Transform the file descriptors into struct file pointers.
1810 * Since the sizeof(struct file *) is bigger than or equal to
1811 * the sizeof(int), we do it in reverse order so that the int
1812 * won't get trashed until we're done.
1814 fdp = (int *)CMSG_DATA(cm) + oldfds - 1;
1815 rp = (struct file **)CMSG_DATA(cm) + oldfds - 1;
1816 for (i = 0; i < oldfds; i++) {
1817 fp = fdescp->fd_files[*fdp--].fp;
1824 spin_unlock_shared(&fdescp->fd_spin);
1825 lwkt_reltoken(&unp_rights_token);
1829 #ifdef UNP_GC_ALLFILES
1832 * Garbage collect in-transit file descriptors that get lost due to
1833 * loops (i.e. when a socket is sent to another process over itself,
1834 * and more complex situations).
1836 * NOT MPSAFE - TODO socket flush code and maybe fdrop. Rest is MPSAFE.
1839 struct unp_gc_info {
1840 struct file **extra_ref;
1841 struct file *locked_fp;
1848 unp_gc(void *arg __unused, int pending __unused)
1850 struct unp_gc_info info;
1854 lwkt_gettoken(&unp_rights_token);
1857 * Before going through all this, set all FDs to be NOT defered
1858 * and NOT externally accessible (not marked). During the scan
1859 * a fd can be marked externally accessible but we may or may not
1860 * be able to immediately process it (controlled by FDEFER).
1862 * If we loop sleep a bit. The complexity of the topology can cause
1863 * multiple loops. Also failure to acquire the socket's so_rcv
1864 * token can cause us to loop.
1866 allfiles_scan_exclusive(unp_gc_clearmarks, NULL);
1869 allfiles_scan_exclusive(unp_gc_checkmarks, &info);
1871 tsleep(&info, 0, "gcagain", 1);
1872 } while (info.defer);
1875 * We grab an extra reference to each of the file table entries
1876 * that are not otherwise accessible and then free the rights
1877 * that are stored in messages on them.
1879 * The bug in the orginal code is a little tricky, so I'll describe
1880 * what's wrong with it here.
1882 * It is incorrect to simply unp_discard each entry for f_msgcount
1883 * times -- consider the case of sockets A and B that contain
1884 * references to each other. On a last close of some other socket,
1885 * we trigger a gc since the number of outstanding rights (unp_rights)
1886 * is non-zero. If during the sweep phase the gc code unp_discards,
1887 * we end up doing a (full) fdrop on the descriptor. A fdrop on A
1888 * results in the following chain. Closef calls soo_close, which
1889 * calls soclose. Soclose calls first (through the switch
1890 * uipc_usrreq) unp_detach, which re-invokes unp_gc. Unp_gc simply
1891 * returns because the previous instance had set unp_gcing, and
1892 * we return all the way back to soclose, which marks the socket
1893 * with SS_NOFDREF, and then calls sofree. Sofree calls sorflush
1894 * to free up the rights that are queued in messages on the socket A,
1895 * i.e., the reference on B. The sorflush calls via the dom_dispose
1896 * switch unp_dispose, which unp_scans with unp_discard. This second
1897 * instance of unp_discard just calls fdrop on B.
1899 * Well, a similar chain occurs on B, resulting in a sorflush on B,
1900 * which results in another fdrop on A. Unfortunately, A is already
1901 * being closed, and the descriptor has already been marked with
1902 * SS_NOFDREF, and soclose panics at this point.
1904 * Here, we first take an extra reference to each inaccessible
1905 * descriptor. Then, we call sorflush ourself, since we know
1906 * it is a Unix domain socket anyhow. After we destroy all the
1907 * rights carried in messages, we do a last fdrop to get rid
1908 * of our extra reference. This is the last close, and the
1909 * unp_detach etc will shut down the socket.
1911 * 91/09/19, bsy@cs.cmu.edu
1913 info.extra_ref = kmalloc(256 * sizeof(struct file *), M_FILE, M_WAITOK);
1914 info.maxindex = 256;
1921 allfiles_scan_exclusive(unp_gc_checkrefs, &info);
1924 * For each FD on our hit list, do the following two things
1926 for (i = info.index, fpp = info.extra_ref; --i >= 0; ++fpp) {
1927 struct file *tfp = *fpp;
1928 if (tfp->f_type == DTYPE_SOCKET && tfp->f_data != NULL)
1929 sorflush((struct socket *)(tfp->f_data));
1931 for (i = info.index, fpp = info.extra_ref; --i >= 0; ++fpp)
1933 } while (info.index == info.maxindex);
1935 kfree((caddr_t)info.extra_ref, M_FILE);
1937 lwkt_reltoken(&unp_rights_token);
1941 * MPSAFE - NOTE: filehead list and file pointer spinlocked on entry
1944 unp_gc_checkrefs(struct file *fp, void *data)
1946 struct unp_gc_info *info = data;
1948 if (fp->f_count == 0)
1950 if (info->index == info->maxindex)
1954 * If all refs are from msgs, and it's not marked accessible
1955 * then it must be referenced from some unreachable cycle
1956 * of (shut-down) FDs, so include it in our
1957 * list of FDs to remove
1959 if (fp->f_count == fp->f_msgcount && !(fp->f_flag & FMARK)) {
1960 info->extra_ref[info->index++] = fp;
1967 * MPSAFE - NOTE: filehead list and file pointer spinlocked on entry
1970 unp_gc_clearmarks(struct file *fp, void *data __unused)
1972 atomic_clear_int(&fp->f_flag, FMARK | FDEFER);
1977 * MPSAFE - NOTE: filehead list and file pointer spinlocked on entry
1980 unp_gc_checkmarks(struct file *fp, void *data)
1982 struct unp_gc_info *info = data;
1986 * If the file is not open, skip it. Make sure it isn't marked
1987 * defered or we could loop forever, in case we somehow race
1990 if (fp->f_count == 0) {
1991 if (fp->f_flag & FDEFER)
1992 atomic_clear_int(&fp->f_flag, FDEFER);
1996 * If we already marked it as 'defer' in a
1997 * previous pass, then try process it this time
2000 if (fp->f_flag & FDEFER) {
2001 atomic_clear_int(&fp->f_flag, FDEFER);
2004 * if it's not defered, then check if it's
2005 * already marked.. if so skip it
2007 if (fp->f_flag & FMARK)
2010 * If all references are from messages
2011 * in transit, then skip it. it's not
2012 * externally accessible.
2014 if (fp->f_count == fp->f_msgcount)
2017 * If it got this far then it must be
2018 * externally accessible.
2020 atomic_set_int(&fp->f_flag, FMARK);
2024 * either it was defered, or it is externally
2025 * accessible and not already marked so.
2026 * Now check if it is possibly one of OUR sockets.
2028 if (fp->f_type != DTYPE_SOCKET ||
2029 (so = (struct socket *)fp->f_data) == NULL) {
2032 if (so->so_proto->pr_domain != &localdomain ||
2033 !(so->so_proto->pr_flags & PR_RIGHTS)) {
2038 * So, Ok, it's one of our sockets and it IS externally accessible
2039 * (or was defered). Now we look to see if we hold any file
2040 * descriptors in its message buffers. Follow those links and mark
2041 * them as accessible too.
2043 * We are holding multiple spinlocks here, if we cannot get the
2044 * token non-blocking defer until the next loop.
2046 info->locked_fp = fp;
2047 if (lwkt_trytoken(&so->so_rcv.ssb_token)) {
2048 unp_scan(so->so_rcv.ssb_mb, unp_mark, info);
2049 lwkt_reltoken(&so->so_rcv.ssb_token);
2051 atomic_set_int(&fp->f_flag, FDEFER);
2058 * Mark visibility. info->defer is recalculated on every pass.
2061 unp_mark(struct file *fp, void *data)
2063 struct unp_gc_info *info = data;
2065 if ((fp->f_flag & FMARK) == 0) {
2067 atomic_set_int(&fp->f_flag, FMARK | FDEFER);
2068 } else if (fp->f_flag & FDEFER) {
2073 #else /* !UNP_GC_ALLFILES */
2076 * They are thread local and do not require explicit synchronization.
2078 static int unp_marked;
2079 static int unp_unreachable;
2082 unp_accessable(struct file *fp, void *data __unused)
2086 if ((unp = unp_fp2unpcb(fp)) == NULL)
2088 if (unp->unp_gcflags & UNPGC_REF)
2090 unp->unp_gcflags &= ~UNPGC_DEAD;
2091 unp->unp_gcflags |= UNPGC_REF;
2096 unp_gc_process(struct unpcb *unp)
2100 /* Already processed. */
2101 if (unp->unp_gcflags & UNPGC_SCANNED)
2106 * Check for a socket potentially in a cycle. It must be in a
2107 * queue as indicated by msgcount, and this must equal the file
2108 * reference count. Note that when msgcount is 0 the file is NULL.
2110 if ((unp->unp_gcflags & UNPGC_REF) == 0 && fp &&
2111 unp->unp_msgcount != 0 && fp->f_count == unp->unp_msgcount) {
2112 unp->unp_gcflags |= UNPGC_DEAD;
2118 * Mark all sockets we reference with RIGHTS.
2120 if (UNP_ISATTACHED(unp)) {
2121 struct signalsockbuf *ssb = &unp->unp_socket->so_rcv;
2124 lwkt_gettoken(&ssb->ssb_token);
2126 * unp_token would be temporarily dropped, if getting
2127 * so_rcv token blocks, so we need to check unp state
2130 if (UNP_ISATTACHED(unp))
2131 unp_scan(ssb->ssb_mb, unp_accessable, NULL);
2132 lwkt_reltoken(&ssb->ssb_token);
2133 unp->unp_gcflags |= UNPGC_SCANNED;
2136 unp->unp_gcflags |= UNPGC_SCANNED;
2141 unp_gc(void *arg __unused, int pending __unused)
2143 struct unp_global_head *head;
2144 int h, filemax, fileidx, filetot;
2145 struct file **unref;
2148 lwkt_gettoken(&unp_rights_token);
2149 lwkt_gettoken(&unp_token);
2152 * First clear all gc flags from previous runs.
2154 for (h = 0; unp_heads[h] != NULL; ++h) {
2156 * NOTE: This loop does not block, so it is safe
2157 * to use TAILQ_FOREACH here.
2159 head = unp_heads[h];
2160 TAILQ_FOREACH(unp, &head->list, unp_link)
2161 unp->unp_gcflags = 0;
2165 * Scan marking all reachable sockets with UNPGC_REF. Once a socket
2166 * is reachable all of the sockets it references are reachable.
2167 * Stop the scan once we do a complete loop without discovering
2168 * a new reachable socket.
2171 unp_unreachable = 0;
2173 for (h = 0; unp_heads[h] != NULL; ++h) {
2174 head = unp_heads[h];
2175 TAILQ_INSERT_HEAD(&head->list, unp_gc_marker, unp_link);
2176 while ((unp = TAILQ_NEXT(unp_gc_marker, unp_link))
2178 TAILQ_REMOVE(&head->list, unp_gc_marker,
2180 TAILQ_INSERT_AFTER(&head->list, unp,
2181 unp_gc_marker, unp_link);
2183 if (unp->unp_flags & UNP_MARKER)
2185 unp_gc_process(unp);
2187 TAILQ_REMOVE(&head->list, unp_gc_marker, unp_link);
2189 } while (unp_marked);
2191 if (unp_unreachable == 0)
2195 * We grab an extra reference to each of the file table entries
2196 * that are not otherwise accessible and then free the rights
2197 * that are stored in messages on them.
2199 * The bug in the orginal code is a little tricky, so I'll describe
2200 * what's wrong with it here.
2202 * It is incorrect to simply unp_discard each entry for f_msgcount
2203 * times -- consider the case of sockets A and B that contain
2204 * references to each other. On a last close of some other socket,
2205 * we trigger a gc since the number of outstanding rights (unp_rights)
2206 * is non-zero. If during the sweep phase the gc code unp_discards,
2207 * we end up doing a (full) fdrop on the descriptor. A fdrop on A
2208 * results in the following chain. Closef calls soo_close, which
2209 * calls soclose. Soclose calls first (through the switch
2210 * uipc_usrreq) unp_detach, which re-invokes unp_gc. Unp_gc simply
2211 * returns because the previous instance had set unp_gcing, and
2212 * we return all the way back to soclose, which marks the socket
2213 * with SS_NOFDREF, and then calls sofree. Sofree calls sorflush
2214 * to free up the rights that are queued in messages on the socket A,
2215 * i.e., the reference on B. The sorflush calls via the dom_dispose
2216 * switch unp_dispose, which unp_scans with unp_discard. This second
2217 * instance of unp_discard just calls fdrop on B.
2219 * Well, a similar chain occurs on B, resulting in a sorflush on B,
2220 * which results in another fdrop on A. Unfortunately, A is already
2221 * being closed, and the descriptor has already been marked with
2222 * SS_NOFDREF, and soclose panics at this point.
2224 * Here, we first take an extra reference to each inaccessible
2225 * descriptor. Then, we call sorflush ourself, since we know
2226 * it is a Unix domain socket anyhow. After we destroy all the
2227 * rights carried in messages, we do a last fdrop to get rid
2228 * of our extra reference. This is the last close, and the
2229 * unp_detach etc will shut down the socket.
2231 * 91/09/19, bsy@cs.cmu.edu
2234 filemax = unp_unreachable;
2235 if (filemax > UNP_GCFILE_MAX)
2236 filemax = UNP_GCFILE_MAX;
2237 unref = kmalloc(filemax * sizeof(struct file *), M_TEMP, M_WAITOK);
2244 * Iterate looking for sockets which have been specifically
2245 * marked as as unreachable and store them locally.
2248 for (h = 0; unp_heads[h] != NULL; ++h) {
2250 * NOTE: This loop does not block, so it is safe
2251 * to use TAILQ_FOREACH here.
2253 head = unp_heads[h];
2254 TAILQ_FOREACH(unp, &head->list, unp_link) {
2257 if ((unp->unp_gcflags & UNPGC_DEAD) == 0)
2259 unp->unp_gcflags &= ~UNPGC_DEAD;
2262 if (unp->unp_msgcount == 0 || fp == NULL ||
2263 fp->f_count != unp->unp_msgcount)
2267 KASSERT(fileidx < filemax,
2268 ("invalid fileidx %d, filemax %d",
2270 unref[fileidx++] = fp;
2272 KASSERT(filetot < unp_unreachable,
2273 ("invalid filetot %d and "
2274 "unp_unreachable %d",
2275 filetot, unp_unreachable));
2278 if (fileidx == filemax ||
2279 filetot == unp_unreachable)
2285 * For each Unix domain socket on our hit list, do the
2286 * following two things.
2288 for (i = 0; i < fileidx; ++i)
2289 sorflush(unref[i]->f_data);
2290 for (i = 0; i < fileidx; ++i)
2292 } while (fileidx == filemax && filetot < unp_unreachable);
2293 kfree(unref, M_TEMP);
2295 lwkt_reltoken(&unp_token);
2296 lwkt_reltoken(&unp_rights_token);
2299 #endif /* UNP_GC_ALLFILES */
2302 * Dispose of the fp's stored in a mbuf.
2304 * The dds loop can cause additional fps to be entered onto the
2305 * list while it is running, flattening out the operation and avoiding
2306 * a deep kernel stack recursion.
2309 unp_dispose(struct mbuf *m)
2311 lwkt_gettoken(&unp_rights_token);
2313 unp_scan(m, unp_discard, NULL);
2314 lwkt_reltoken(&unp_rights_token);
2318 unp_listen(struct unpcb *unp, struct thread *td)
2320 struct proc *p = td->td_proc;
2322 ASSERT_LWKT_TOKEN_HELD(&unp_token);
2323 UNP_ASSERT_TOKEN_HELD(unp);
2326 cru2x(p->p_ucred, &unp->unp_peercred);
2327 unp_setflags(unp, UNP_HAVEPCCACHED);
2332 unp_scan(struct mbuf *m0, void (*op)(struct file *, void *), void *data)
2341 for (m = m0; m; m = m->m_next) {
2342 if (m->m_type == MT_CONTROL &&
2343 m->m_len >= sizeof(*cm)) {
2344 cm = mtod(m, struct cmsghdr *);
2345 if (cm->cmsg_level != SOL_SOCKET ||
2346 cm->cmsg_type != SCM_RIGHTS)
2348 qfds = (cm->cmsg_len - CMSG_LEN(0)) /
2350 rp = (struct file **)CMSG_DATA(cm);
2351 for (i = 0; i < qfds; i++)
2353 break; /* XXX, but saves time */
2361 * Discard a fp previously held in a unix domain socket mbuf. To
2362 * avoid blowing out the kernel stack due to contrived chain-reactions
2363 * we may have to defer the operation to a dedicated taskqueue.
2365 * Caller holds unp_rights_token.
2368 unp_discard(struct file *fp, void *data __unused)
2371 if (unp_fp2unpcb(fp) != NULL) {
2372 struct unp_defdiscard *d;
2375 * This fp is a Unix domain socket itself and fdrop()
2376 * it here directly may cause deep unp_discard()
2377 * recursion, so the fdrop() is defered to the
2378 * dedicated taskqueue.
2380 d = kmalloc(sizeof(*d), M_UNPCB, M_WAITOK);
2383 spin_lock(&unp_defdiscard_spin);
2384 SLIST_INSERT_HEAD(&unp_defdiscard_head, d, next);
2385 spin_unlock(&unp_defdiscard_spin);
2387 taskqueue_enqueue(unp_taskqueue, &unp_defdiscard_task);
2389 /* This fp is not a Unix domain socket */
2396 * unp_token must be held before calling this function to avoid name
2397 * resolution and v_socket accessing races, especially racing against
2401 * For anyone caring about unconnected Unix domain socket sending
2402 * performance, other approach could be taken...
2405 unp_find_lockref(struct sockaddr *nam, struct thread *td, short type,
2406 struct unpcb **unp_ret)
2408 struct proc *p = td->td_proc;
2409 struct sockaddr_un *soun = (struct sockaddr_un *)nam;
2410 struct vnode *vp = NULL;
2414 struct nlookupdata nd;
2415 char buf[SOCK_MAXADDRLEN];
2417 ASSERT_LWKT_TOKEN_HELD(&unp_token);
2421 len = nam->sa_len - offsetof(struct sockaddr_un, sun_path);
2426 strncpy(buf, soun->sun_path, len);
2429 error = nlookup_init(&nd, buf, UIO_SYSSPACE, NLC_FOLLOW);
2431 error = nlookup(&nd);
2433 error = cache_vget(&nd.nl_nch, nd.nl_cred, LK_EXCLUSIVE, &vp);
2440 if (vp->v_type != VSOCK) {
2444 error = VOP_EACCESS(vp, VWRITE, p->p_ucred);
2449 error = ECONNREFUSED;
2452 if (so->so_type != type) {
2457 /* Lock this unp. */
2458 unp = unp_getsocktoken(so);
2459 if (!UNP_ISATTACHED(unp)) {
2461 error = ECONNREFUSED;
2464 /* And keep this unp referenced. */
2477 unp_connect_pair(struct unpcb *unp, struct unpcb *unp2)
2479 struct socket *so = unp->unp_socket;
2480 struct socket *so2 = unp2->unp_socket;
2482 ASSERT_LWKT_TOKEN_HELD(&unp_token);
2483 UNP_ASSERT_TOKEN_HELD(unp);
2484 UNP_ASSERT_TOKEN_HELD(unp2);
2486 KASSERT(so->so_type == so2->so_type,
2487 ("socket type mismatch, so %d, so2 %d", so->so_type, so2->so_type));
2489 if (!UNP_ISATTACHED(unp))
2491 if (!UNP_ISATTACHED(unp2))
2492 return ECONNREFUSED;
2494 KASSERT(unp->unp_conn == NULL, ("unp is already connected"));
2495 unp->unp_conn = unp2;
2497 switch (so->so_type) {
2499 LIST_INSERT_HEAD(&unp2->unp_refs, unp, unp_reflink);
2504 case SOCK_SEQPACKET:
2505 KASSERT(unp2->unp_conn == NULL, ("unp2 is already connected"));
2506 unp2->unp_conn = unp;
2512 panic("unp_connect_pair: unknown socket type %d", so->so_type);
2518 unp_drop(struct unpcb *unp, int error)
2520 struct unp_global_head *head;
2523 ASSERT_LWKT_TOKEN_HELD(&unp_token);
2524 UNP_ASSERT_TOKEN_HELD(unp);
2526 KASSERT((unp->unp_flags & (UNP_DETACHED | UNP_DROPPED)) == 0,
2527 ("unp is dropped"));
2529 /* Mark this unp as detached. */
2530 unp_setflags(unp, UNP_DETACHED);
2532 /* Remove this unp from the global unp list. */
2533 head = unp_globalhead(unp->unp_socket->so_type);
2534 KASSERT(head->count > 0, ("invalid unp count"));
2535 TAILQ_REMOVE(&head->list, unp, unp_link);
2538 /* Disconnect all. */
2539 unp_disconnect(unp, error);
2540 while ((unp2 = LIST_FIRST(&unp->unp_refs)) != NULL) {
2541 lwkt_getpooltoken(unp2);
2542 unp_disconnect(unp2, ECONNRESET);
2543 lwkt_relpooltoken(unp2);
2545 unp_setflags(unp, UNP_DROPPED);
2547 /* Try freeing this unp. */
2552 unp_defdiscard_taskfunc(void *arg __unused, int pending __unused)
2554 struct unp_defdiscard *d;
2556 spin_lock(&unp_defdiscard_spin);
2557 while ((d = SLIST_FIRST(&unp_defdiscard_head)) != NULL) {
2558 SLIST_REMOVE_HEAD(&unp_defdiscard_head, next);
2559 spin_unlock(&unp_defdiscard_spin);
2564 spin_lock(&unp_defdiscard_spin);
2566 spin_unlock(&unp_defdiscard_spin);