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 };
147 static ino_t unp_ino = 1; /* prototype for fake inode numbers */
149 static int unp_attach (struct socket *, struct pru_attach_info *);
150 static void unp_detach (struct unpcb *);
151 static int unp_bind (struct unpcb *,struct sockaddr *, struct thread *);
152 static int unp_connect (struct socket *,struct sockaddr *,
154 static void unp_disconnect(struct unpcb *, int);
155 static void unp_shutdown (struct unpcb *);
156 static void unp_gc(void *, int);
157 #ifdef UNP_GC_ALLFILES
158 static int unp_gc_clearmarks(struct file *, void *);
159 static int unp_gc_checkmarks(struct file *, void *);
160 static int unp_gc_checkrefs(struct file *, void *);
161 static void unp_mark(struct file *, void *data);
163 static void unp_scan (struct mbuf *, void (*)(struct file *, void *),
165 static void unp_discard (struct file *, void *);
166 static int unp_internalize (struct mbuf *, struct thread *);
167 static int unp_listen (struct unpcb *, struct thread *);
168 static void unp_fp_externalize(struct lwp *lp, struct file *fp, int fd,
170 static int unp_find_lockref(struct sockaddr *nam, struct thread *td,
171 short type, struct unpcb **unp_ret);
172 static int unp_connect_pair(struct unpcb *unp, struct unpcb *unp2);
173 static void unp_drop(struct unpcb *unp, int error);
174 static void unp_defdiscard_taskfunc(void *, int);
176 static int unp_rights; /* file descriptors in flight */
177 static struct lwkt_token unp_rights_token =
178 LWKT_TOKEN_INITIALIZER(unp_rights_token);
179 static struct task unp_gc_task;
180 static struct unpcb *unp_gc_marker;
182 SYSCTL_DECL(_net_local);
183 SYSCTL_INT(_net_local, OID_AUTO, inflight, CTLFLAG_RD, &unp_rights, 0,
184 "File descriptors in flight");
187 * SMP Considerations:
189 * Since unp_token will be automaticly released upon execution of
190 * blocking code, we need to reference unp_conn before any possible
191 * blocking code to prevent it from being ripped behind our back.
193 * Any adjustment to unp->unp_conn requires both the global unp_token
194 * AND the per-unp token (lwkt_token_pool_lookup(unp)) to be held.
196 * Any access to so_pcb to obtain unp requires the pool token for
201 unp_reference(struct unpcb *unp)
203 /* 0->1 transition will not work */
204 KKASSERT(unp->unp_refcnt > 0);
205 atomic_add_int(&unp->unp_refcnt, 1);
209 unp_free(struct unpcb *unp)
211 KKASSERT(unp->unp_refcnt > 0);
212 if (atomic_fetchadd_int(&unp->unp_refcnt, -1) == 1)
216 static __inline struct unpcb *
217 unp_getsocktoken(struct socket *so)
222 * The unp pointer is invalid until we verify that it is
223 * good by re-checking so_pcb AFTER obtaining the token.
225 while ((unp = so->so_pcb) != NULL) {
226 lwkt_getpooltoken(unp);
227 if (unp == so->so_pcb)
229 lwkt_relpooltoken(unp);
235 unp_reltoken(struct unpcb *unp)
238 lwkt_relpooltoken(unp);
242 unp_setflags(struct unpcb *unp, int flags)
244 atomic_set_int(&unp->unp_flags, flags);
248 unp_clrflags(struct unpcb *unp, int flags)
250 atomic_clear_int(&unp->unp_flags, flags);
253 static __inline struct unp_global_head *
254 unp_globalhead(short type)
258 return &unp_stream_head;
260 return &unp_dgram_head;
262 return &unp_seqpkt_head;
264 panic("unknown socket type %d", type);
268 static __inline struct unpcb *
269 unp_fp2unpcb(struct file *fp)
273 if (fp->f_type != DTYPE_SOCKET)
280 if (so->so_proto->pr_domain != &localdomain)
287 unp_add_right(struct file *fp)
291 ASSERT_LWKT_TOKEN_HELD(&unp_rights_token);
292 KASSERT(fp->f_count > 0, ("invalid f_count %d", fp->f_count));
294 unp = unp_fp2unpcb(fp);
304 unp_del_right(struct file *fp)
308 ASSERT_LWKT_TOKEN_HELD(&unp_rights_token);
309 KASSERT(fp->f_count > 0, ("invalid f_count %d", fp->f_count));
311 unp = unp_fp2unpcb(fp);
313 KASSERT(unp->unp_msgcount > 0,
314 ("invalid unp msgcount %d", unp->unp_msgcount));
316 if (unp->unp_msgcount == 0)
324 * NOTE: (so) is referenced from soabort*() and netmsg_pru_abort()
325 * will sofree() it when we return.
328 uipc_abort(netmsg_t msg)
333 lwkt_gettoken(&unp_token);
334 unp = unp_getsocktoken(msg->base.nm_so);
336 if (UNP_ISATTACHED(unp)) {
337 unp_drop(unp, ECONNABORTED);
344 lwkt_reltoken(&unp_token);
346 lwkt_replymsg(&msg->lmsg, error);
350 uipc_accept(netmsg_t msg)
355 lwkt_gettoken(&unp_token);
356 unp = unp_getsocktoken(msg->base.nm_so);
358 if (!UNP_ISATTACHED(unp)) {
361 struct unpcb *unp2 = unp->unp_conn;
364 * Pass back name of connected socket,
365 * if it was bound and we are still connected
366 * (our peer may have closed already!).
368 if (unp2 && unp2->unp_addr) {
370 *msg->accept.nm_nam = dup_sockaddr(
371 (struct sockaddr *)unp2->unp_addr);
374 *msg->accept.nm_nam = dup_sockaddr(&sun_noname);
380 lwkt_reltoken(&unp_token);
382 lwkt_replymsg(&msg->lmsg, error);
386 uipc_attach(netmsg_t msg)
390 lwkt_gettoken(&unp_token);
392 KASSERT(msg->base.nm_so->so_pcb == NULL, ("double unp attach"));
393 error = unp_attach(msg->base.nm_so, msg->attach.nm_ai);
395 lwkt_reltoken(&unp_token);
396 lwkt_replymsg(&msg->lmsg, error);
400 uipc_bind(netmsg_t msg)
405 lwkt_gettoken(&unp_token);
406 unp = unp_getsocktoken(msg->base.nm_so);
408 if (UNP_ISATTACHED(unp))
409 error = unp_bind(unp, msg->bind.nm_nam, msg->bind.nm_td);
414 lwkt_reltoken(&unp_token);
416 lwkt_replymsg(&msg->lmsg, error);
420 uipc_connect(netmsg_t msg)
424 error = unp_connect(msg->base.nm_so, msg->connect.nm_nam,
426 lwkt_replymsg(&msg->lmsg, error);
430 uipc_connect2(netmsg_t msg)
434 error = unp_connect2(msg->connect2.nm_so1, msg->connect2.nm_so2);
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 from = (struct sockaddr *)unp->unp_addr;
681 lwkt_gettoken(&so2->so_rcv.ssb_token);
682 if (ssb_appendaddr(&so2->so_rcv, from, m, control)) {
689 lwkt_reltoken(&so2->so_rcv.ssb_token);
697 /* Connect if not connected yet. */
699 * Note: A better implementation would complain
700 * if not equal to the peer's address.
702 if (unp->unp_conn == NULL) {
703 if (msg->send.nm_addr) {
704 error = unp_connect(so,
712 * unp_conn still could be NULL, even if the
713 * above unp_connect() succeeds; since the
714 * current unp's token could be released due
715 * to blocking operations after unp_conn is
718 if (unp->unp_conn == NULL) {
723 if (so->so_state & SS_CANTSENDMORE) {
728 unp2 = unp->unp_conn;
729 KASSERT(unp2 != NULL, ("unp is not connected"));
730 so2 = unp2->unp_socket;
735 * Send to paired receive port, and then reduce
736 * send buffer hiwater marks to maintain backpressure.
739 lwkt_gettoken(&so2->so_rcv.ssb_token);
741 if (ssb_appendcontrol(&so2->so_rcv, m, control)) {
745 } else if (so->so_type == SOCK_SEQPACKET) {
746 sbappendrecord(&so2->so_rcv.sb, m);
749 sbappend(&so2->so_rcv.sb, m);
754 * Because we are transfering mbufs directly to the
755 * peer socket we have to use SSB_STOP on the sender
756 * to prevent it from building up infinite mbufs.
758 if (so2->so_rcv.ssb_cc >= so->so_snd.ssb_hiwat ||
759 so2->so_rcv.ssb_mbcnt >= so->so_snd.ssb_mbmax
761 atomic_set_int(&so->so_snd.ssb_flags, SSB_STOP);
763 lwkt_reltoken(&so2->so_rcv.ssb_token);
770 panic("uipc_send unknown socktype");
774 * SEND_EOF is equivalent to a SEND followed by a SHUTDOWN.
776 if (msg->send.nm_flags & PRUS_EOF) {
781 if (control && error != 0)
782 unp_dispose(control);
785 wakeup_end_delayed();
791 lwkt_replymsg(&msg->lmsg, error);
798 uipc_sense(netmsg_t msg)
805 so = msg->base.nm_so;
806 sb = msg->sense.nm_stat;
809 * so_pcb is only modified with both the global and the unp
812 unp = unp_getsocktoken(so);
814 if (!UNP_ISATTACHED(unp)) {
819 sb->st_blksize = so->so_snd.ssb_hiwat;
821 if (unp->unp_ino == 0) { /* make up a non-zero inode number */
822 unp->unp_ino = atomic_fetchadd_long(&unp_ino, 1);
823 if (__predict_false(unp->unp_ino == 0))
824 unp->unp_ino = atomic_fetchadd_long(&unp_ino, 1);
826 sb->st_ino = unp->unp_ino;
830 lwkt_replymsg(&msg->lmsg, error);
834 uipc_shutdown(netmsg_t msg)
841 * so_pcb is only modified with both the global and the unp
844 so = msg->base.nm_so;
845 unp = unp_getsocktoken(so);
847 if (UNP_ISATTACHED(unp)) {
856 lwkt_replymsg(&msg->lmsg, error);
860 uipc_sockaddr(netmsg_t msg)
866 * so_pcb is only modified with both the global and the unp
869 unp = unp_getsocktoken(msg->base.nm_so);
871 if (UNP_ISATTACHED(unp)) {
873 *msg->sockaddr.nm_nam =
874 dup_sockaddr((struct sockaddr *)unp->unp_addr);
882 lwkt_replymsg(&msg->lmsg, error);
885 struct pr_usrreqs uipc_usrreqs = {
886 .pru_abort = uipc_abort,
887 .pru_accept = uipc_accept,
888 .pru_attach = uipc_attach,
889 .pru_bind = uipc_bind,
890 .pru_connect = uipc_connect,
891 .pru_connect2 = uipc_connect2,
892 .pru_control = pr_generic_notsupp,
893 .pru_detach = uipc_detach,
894 .pru_disconnect = uipc_disconnect,
895 .pru_listen = uipc_listen,
896 .pru_peeraddr = uipc_peeraddr,
897 .pru_rcvd = uipc_rcvd,
898 .pru_rcvoob = pr_generic_notsupp,
899 .pru_send = uipc_send,
900 .pru_sense = uipc_sense,
901 .pru_shutdown = uipc_shutdown,
902 .pru_sockaddr = uipc_sockaddr,
903 .pru_sosend = sosend,
904 .pru_soreceive = soreceive
908 uipc_ctloutput(netmsg_t msg)
911 struct sockopt *sopt;
915 so = msg->base.nm_so;
916 sopt = msg->ctloutput.nm_sopt;
918 lwkt_gettoken(&unp_token);
919 unp = unp_getsocktoken(so);
921 if (!UNP_ISATTACHED(unp)) {
926 switch (sopt->sopt_dir) {
928 switch (sopt->sopt_name) {
930 if (unp->unp_flags & UNP_HAVEPC)
931 soopt_from_kbuf(sopt, &unp->unp_peercred,
932 sizeof(unp->unp_peercred));
934 if (so->so_type == SOCK_STREAM)
936 else if (so->so_type == SOCK_SEQPACKET)
955 lwkt_reltoken(&unp_token);
957 lwkt_replymsg(&msg->lmsg, error);
961 * Both send and receive buffers are allocated PIPSIZ bytes of buffering
962 * for stream sockets, although the total for sender and receiver is
963 * actually only PIPSIZ.
965 * Datagram sockets really use the sendspace as the maximum datagram size,
966 * and don't really want to reserve the sendspace. Their recvspace should
967 * be large enough for at least one max-size datagram plus address.
969 * We want the local send/recv space to be significant larger then lo0's
975 static u_long unpst_sendspace = PIPSIZ;
976 static u_long unpst_recvspace = PIPSIZ;
977 static u_long unpdg_sendspace = 2*1024; /* really max datagram size */
978 static u_long unpdg_recvspace = 4*1024;
980 SYSCTL_DECL(_net_local_seqpacket);
981 SYSCTL_DECL(_net_local_stream);
982 SYSCTL_INT(_net_local_stream, OID_AUTO, sendspace, CTLFLAG_RW,
983 &unpst_sendspace, 0, "Size of stream socket send buffer");
984 SYSCTL_INT(_net_local_stream, OID_AUTO, recvspace, CTLFLAG_RW,
985 &unpst_recvspace, 0, "Size of stream socket receive buffer");
987 SYSCTL_DECL(_net_local_dgram);
988 SYSCTL_INT(_net_local_dgram, OID_AUTO, maxdgram, CTLFLAG_RW,
989 &unpdg_sendspace, 0, "Max datagram socket size");
990 SYSCTL_INT(_net_local_dgram, OID_AUTO, recvspace, CTLFLAG_RW,
991 &unpdg_recvspace, 0, "Size of datagram socket receive buffer");
994 unp_attach(struct socket *so, struct pru_attach_info *ai)
996 struct unp_global_head *head;
1000 lwkt_gettoken(&unp_token);
1002 if (so->so_snd.ssb_hiwat == 0 || so->so_rcv.ssb_hiwat == 0) {
1003 switch (so->so_type) {
1005 case SOCK_SEQPACKET:
1006 error = soreserve(so, unpst_sendspace, unpst_recvspace,
1011 error = soreserve(so, unpdg_sendspace, unpdg_recvspace,
1016 panic("unp_attach");
1023 * In order to support sendfile we have to set either SSB_STOPSUPP
1024 * or SSB_PREALLOC. Unix domain sockets use the SSB_STOP flow
1025 * control mechanism.
1027 if (so->so_type == SOCK_STREAM) {
1028 atomic_set_int(&so->so_rcv.ssb_flags, SSB_STOPSUPP);
1029 atomic_set_int(&so->so_snd.ssb_flags, SSB_STOPSUPP);
1032 unp = kmalloc(sizeof(*unp), M_UNPCB, M_WAITOK | M_ZERO | M_NULLOK);
1037 unp->unp_refcnt = 1;
1038 unp->unp_gencnt = ++unp_gencnt;
1039 LIST_INIT(&unp->unp_refs);
1040 unp->unp_socket = so;
1041 unp->unp_rvnode = ai->fd_rdir; /* jail cruft XXX JH */
1042 so->so_pcb = (caddr_t)unp;
1045 head = unp_globalhead(so->so_type);
1046 TAILQ_INSERT_TAIL(&head->list, unp, unp_link);
1050 lwkt_reltoken(&unp_token);
1055 unp_detach(struct unpcb *unp)
1059 lwkt_gettoken(&unp_token);
1060 lwkt_getpooltoken(unp);
1062 so = unp->unp_socket;
1064 unp->unp_gencnt = ++unp_gencnt;
1065 if (unp->unp_vnode) {
1066 unp->unp_vnode->v_socket = NULL;
1067 vrele(unp->unp_vnode);
1068 unp->unp_vnode = NULL;
1070 soisdisconnected(so);
1071 KKASSERT(so->so_pcb == unp);
1072 so->so_pcb = NULL; /* both tokens required */
1073 unp->unp_socket = NULL;
1075 lwkt_relpooltoken(unp);
1076 lwkt_reltoken(&unp_token);
1080 KASSERT(unp->unp_conn == NULL, ("unp is still connected"));
1081 KASSERT(LIST_EMPTY(&unp->unp_refs), ("unp still has references"));
1084 kfree(unp->unp_addr, M_SONAME);
1085 kfree(unp, M_UNPCB);
1088 taskqueue_enqueue(unp_taskqueue, &unp_gc_task);
1092 unp_bind(struct unpcb *unp, struct sockaddr *nam, struct thread *td)
1094 struct proc *p = td->td_proc;
1095 struct sockaddr_un *soun = (struct sockaddr_un *)nam;
1099 struct nlookupdata nd;
1100 char buf[SOCK_MAXADDRLEN];
1102 ASSERT_LWKT_TOKEN_HELD(&unp_token);
1103 UNP_ASSERT_TOKEN_HELD(unp);
1105 if (unp->unp_vnode != NULL)
1108 namelen = soun->sun_len - offsetof(struct sockaddr_un, sun_path);
1111 strncpy(buf, soun->sun_path, namelen);
1112 buf[namelen] = 0; /* null-terminate the string */
1113 error = nlookup_init(&nd, buf, UIO_SYSSPACE,
1114 NLC_LOCKVP | NLC_CREATE | NLC_REFDVP);
1116 error = nlookup(&nd);
1117 if (error == 0 && nd.nl_nch.ncp->nc_vp != NULL)
1123 vattr.va_type = VSOCK;
1124 vattr.va_mode = (ACCESSPERMS & ~p->p_fd->fd_cmask);
1125 error = VOP_NCREATE(&nd.nl_nch, nd.nl_dvp, &vp, nd.nl_cred, &vattr);
1127 if (unp->unp_vnode == NULL) {
1128 vp->v_socket = unp->unp_socket;
1129 unp->unp_vnode = vp;
1130 unp->unp_addr = (struct sockaddr_un *)dup_sockaddr(nam);
1133 vput(vp); /* late race */
1143 unp_connect(struct socket *so, struct sockaddr *nam, struct thread *td)
1145 struct unpcb *unp, *unp2;
1146 int error, flags = 0;
1148 lwkt_gettoken(&unp_token);
1150 unp = unp_getsocktoken(so);
1151 if (!UNP_ISATTACHED(unp)) {
1156 if ((unp->unp_flags & UNP_CONNECTING) || unp->unp_conn != NULL) {
1161 flags = UNP_CONNECTING;
1162 unp_setflags(unp, flags);
1164 error = unp_find_lockref(nam, td, so->so_type, &unp2);
1169 * unp2 is locked and referenced.
1172 if (so->so_proto->pr_flags & PR_CONNREQUIRED) {
1173 struct socket *so2, *so3;
1176 so2 = unp2->unp_socket;
1177 if (!(so2->so_options & SO_ACCEPTCONN) ||
1178 (so3 = sonewconn_faddr(so2, 0, NULL,
1179 TRUE /* keep ref */)) == NULL) {
1180 error = ECONNREFUSED;
1183 /* so3 has a socket reference. */
1185 unp3 = unp_getsocktoken(so3);
1186 if (!UNP_ISATTACHED(unp3)) {
1189 * Already aborted; we only need to drop the
1190 * socket reference held by sonewconn_faddr().
1193 error = ECONNREFUSED;
1196 unp_reference(unp3);
1199 * unp3 is locked and referenced.
1203 * Release so3 socket reference held by sonewconn_faddr().
1204 * Since we have referenced unp3, neither unp3 nor so3 will
1205 * be destroyed here.
1209 if (unp2->unp_addr != NULL) {
1210 unp3->unp_addr = (struct sockaddr_un *)
1211 dup_sockaddr((struct sockaddr *)unp2->unp_addr);
1215 * unp_peercred management:
1217 * The connecter's (client's) credentials are copied
1218 * from its process structure at the time of connect()
1221 cru2x(td->td_proc->p_ucred, &unp3->unp_peercred);
1222 unp_setflags(unp3, UNP_HAVEPC);
1224 * The receiver's (server's) credentials are copied
1225 * from the unp_peercred member of socket on which the
1226 * former called listen(); unp_listen() cached that
1227 * process's credentials at that time so we can use
1230 KASSERT(unp2->unp_flags & UNP_HAVEPCCACHED,
1231 ("unp_connect: listener without cached peercred"));
1232 memcpy(&unp->unp_peercred, &unp2->unp_peercred,
1233 sizeof(unp->unp_peercred));
1234 unp_setflags(unp, UNP_HAVEPC);
1236 error = unp_connect_pair(unp, unp3);
1238 soabort_direct(so3);
1240 /* Done with unp3 */
1244 error = unp_connect_pair(unp, unp2);
1251 unp_clrflags(unp, flags);
1254 lwkt_reltoken(&unp_token);
1259 * Connect two unix domain sockets together.
1261 * NOTE: Semantics for any change to unp_conn requires that the per-unp
1262 * pool token also be held.
1265 unp_connect2(struct socket *so, struct socket *so2)
1267 struct unpcb *unp, *unp2;
1270 lwkt_gettoken(&unp_token);
1271 if (so2->so_type != so->so_type) {
1272 lwkt_reltoken(&unp_token);
1273 return (EPROTOTYPE);
1275 unp = unp_getsocktoken(so);
1276 unp2 = unp_getsocktoken(so2);
1278 if (!UNP_ISATTACHED(unp)) {
1282 if (!UNP_ISATTACHED(unp2)) {
1283 error = ECONNREFUSED;
1287 if (unp->unp_conn != NULL) {
1291 if ((so->so_type == SOCK_STREAM || so->so_type == SOCK_SEQPACKET) &&
1292 unp2->unp_conn != NULL) {
1297 error = unp_connect_pair(unp, unp2);
1301 lwkt_reltoken(&unp_token);
1306 * Disconnect a unix domain socket pair.
1308 * NOTE: Semantics for any change to unp_conn requires that the per-unp
1309 * pool token also be held.
1312 unp_disconnect(struct unpcb *unp, int error)
1314 struct socket *so = unp->unp_socket;
1317 ASSERT_LWKT_TOKEN_HELD(&unp_token);
1318 UNP_ASSERT_TOKEN_HELD(unp);
1321 so->so_error = error;
1323 while ((unp2 = unp->unp_conn) != NULL) {
1324 lwkt_getpooltoken(unp2);
1325 if (unp2 == unp->unp_conn)
1327 lwkt_relpooltoken(unp2);
1331 /* unp2 is locked. */
1333 KASSERT((unp2->unp_flags & UNP_DROPPED) == 0, ("unp2 was dropped"));
1335 unp->unp_conn = NULL;
1337 switch (so->so_type) {
1339 LIST_REMOVE(unp, unp_reflink);
1340 soclrstate(so, SS_ISCONNECTED);
1344 case SOCK_SEQPACKET:
1346 * Keep a reference before clearing the unp_conn
1347 * to avoid racing uipc_detach()/uipc_abort() in
1350 unp_reference(unp2);
1351 KASSERT(unp2->unp_conn == unp, ("unp_conn mismatch"));
1352 unp2->unp_conn = NULL;
1354 soisdisconnected(so);
1355 soisdisconnected(unp2->unp_socket);
1361 lwkt_relpooltoken(unp2);
1366 unp_abort(struct unpcb *unp)
1368 lwkt_gettoken(&unp_token);
1370 lwkt_reltoken(&unp_token);
1375 prison_unpcb(struct thread *td, struct unpcb *unp)
1381 if ((p = td->td_proc) == NULL)
1383 if (!p->p_ucred->cr_prison)
1385 if (p->p_fd->fd_rdir == unp->unp_rvnode)
1391 unp_pcblist(SYSCTL_HANDLER_ARGS)
1393 struct unp_global_head *head = arg1;
1395 struct unpcb *unp, *marker;
1397 KKASSERT(curproc != NULL);
1400 * The process of preparing the PCB list is too time-consuming and
1401 * resource-intensive to repeat twice on every request.
1403 if (req->oldptr == NULL) {
1405 req->oldidx = (n + n/8) * sizeof(struct xunpcb);
1409 if (req->newptr != NULL)
1412 marker = kmalloc(sizeof(*marker), M_UNPCB, M_WAITOK | M_ZERO);
1413 marker->unp_flags |= UNP_MARKER;
1415 lwkt_gettoken(&unp_token);
1421 TAILQ_INSERT_HEAD(&head->list, marker, unp_link);
1422 while ((unp = TAILQ_NEXT(marker, unp_link)) != NULL && i < n) {
1425 TAILQ_REMOVE(&head->list, marker, unp_link);
1426 TAILQ_INSERT_AFTER(&head->list, unp, marker, unp_link);
1428 if (unp->unp_flags & UNP_MARKER)
1430 if (prison_unpcb(req->td, unp))
1433 xu.xu_len = sizeof(xu);
1438 * unp->unp_addr and unp->unp_conn are protected by
1439 * unp_token. So if we want to get rid of unp_token
1440 * or reduce the coverage of unp_token, care must be
1443 if (unp->unp_addr) {
1444 bcopy(unp->unp_addr, &xu.xu_addr,
1445 unp->unp_addr->sun_len);
1447 if (unp->unp_conn && unp->unp_conn->unp_addr) {
1448 bcopy(unp->unp_conn->unp_addr,
1450 unp->unp_conn->unp_addr->sun_len);
1452 bcopy(unp, &xu.xu_unp, sizeof(*unp));
1453 sotoxsocket(unp->unp_socket, &xu.xu_socket);
1455 /* NOTE: This could block and temporarily release unp_token */
1456 error = SYSCTL_OUT(req, &xu, sizeof(xu));
1461 TAILQ_REMOVE(&head->list, marker, unp_link);
1463 lwkt_reltoken(&unp_token);
1465 kfree(marker, M_UNPCB);
1469 SYSCTL_PROC(_net_local_dgram, OID_AUTO, pcblist, CTLFLAG_RD,
1470 &unp_dgram_head, 0, unp_pcblist, "S,xunpcb",
1471 "List of active local datagram sockets");
1472 SYSCTL_PROC(_net_local_stream, OID_AUTO, pcblist, CTLFLAG_RD,
1473 &unp_stream_head, 0, unp_pcblist, "S,xunpcb",
1474 "List of active local stream sockets");
1475 SYSCTL_PROC(_net_local_seqpacket, OID_AUTO, pcblist, CTLFLAG_RD,
1476 &unp_seqpkt_head, 0, unp_pcblist, "S,xunpcb",
1477 "List of active local seqpacket sockets");
1480 unp_shutdown(struct unpcb *unp)
1484 if ((unp->unp_socket->so_type == SOCK_STREAM ||
1485 unp->unp_socket->so_type == SOCK_SEQPACKET) &&
1486 unp->unp_conn != NULL && (so = unp->unp_conn->unp_socket)) {
1495 lwkt_gettoken(&unp_token);
1496 lwkt_reltoken(&unp_token);
1501 unp_externalize(struct mbuf *rights, int flags)
1503 struct thread *td = curthread;
1504 struct proc *p = td->td_proc; /* XXX */
1505 struct lwp *lp = td->td_lwp;
1506 struct cmsghdr *cm = mtod(rights, struct cmsghdr *);
1511 int newfds = (cm->cmsg_len - (CMSG_DATA(cm) - (u_char *)cm))
1512 / sizeof(struct file *);
1515 lwkt_gettoken(&unp_rights_token);
1518 * if the new FD's will not fit, then we free them all
1520 if (!fdavail(p, newfds)) {
1521 rp = (struct file **)CMSG_DATA(cm);
1522 for (i = 0; i < newfds; i++) {
1525 * zero the pointer before calling unp_discard,
1526 * since it may end up in unp_gc()..
1529 unp_discard(fp, NULL);
1531 lwkt_reltoken(&unp_rights_token);
1536 * now change each pointer to an fd in the global table to
1537 * an integer that is the index to the local fd table entry
1538 * that we set up to point to the global one we are transferring.
1539 * Since the sizeof(struct file *) is bigger than or equal to
1540 * the sizeof(int), we do it in forward order. In that case,
1541 * an integer will always come in the same place or before its
1542 * corresponding struct file pointer.
1544 * Hold revoke_token in 'shared' mode, so that we won't miss
1545 * the FREVOKED update on fps being externalized (fsetfd).
1547 lwkt_gettoken_shared(&revoke_token);
1548 fdp = (int *)CMSG_DATA(cm);
1549 rp = (struct file **)CMSG_DATA(cm);
1550 for (i = 0; i < newfds; i++) {
1551 if (fdalloc(p, 0, &f)) {
1555 * Previous fdavail() can't garantee
1556 * fdalloc() success due to SMP race.
1557 * Just clean up and return the same
1558 * error value as if fdavail() failed.
1560 lwkt_reltoken(&revoke_token);
1562 /* Close externalized files */
1563 for (j = 0; j < i; j++)
1565 /* Discard the rest of internal files */
1566 for (; i < newfds; i++)
1567 unp_discard(rp[i], NULL);
1568 /* Wipe out the control message */
1569 for (i = 0; i < newfds; i++)
1572 lwkt_reltoken(&unp_rights_token);
1576 unp_fp_externalize(lp, fp, f, flags);
1579 lwkt_reltoken(&revoke_token);
1581 lwkt_reltoken(&unp_rights_token);
1584 * Adjust length, in case sizeof(struct file *) and sizeof(int)
1587 cm->cmsg_len = CMSG_LEN(newfds * sizeof(int));
1588 rights->m_len = cm->cmsg_len;
1594 unp_fp_externalize(struct lwp *lp, struct file *fp, int fd, int flags)
1597 struct filedesc *fdp = lp->lwp_proc->p_fd;
1600 if (fp->f_flag & FREVOKED) {
1604 kprintf("Warning: revoked fp exiting unix socket\n");
1605 error = falloc(lp, &fx, NULL);
1607 if (flags & MSG_CMSG_CLOEXEC)
1608 fdp->fd_files[fd].fileflags |= UF_EXCLOSE;
1609 fsetfd(fdp, fx, fd);
1612 fsetfd(fdp, NULL, fd);
1615 if (flags & MSG_CMSG_CLOEXEC)
1616 fdp->fd_files[fd].fileflags |= UF_EXCLOSE;
1617 fsetfd(fdp, fp, fd);
1627 TAILQ_INIT(&unp_stream_head.list);
1628 TAILQ_INIT(&unp_dgram_head.list);
1629 TAILQ_INIT(&unp_seqpkt_head.list);
1631 SLIST_INIT(&unp_defdiscard_head);
1632 spin_init(&unp_defdiscard_spin, "unpdisc");
1633 TASK_INIT(&unp_defdiscard_task, 0, unp_defdiscard_taskfunc, NULL);
1636 * This implies that only one gc can be in-progress at any
1639 TASK_INIT(&unp_gc_task, 0, unp_gc, NULL);
1641 unp_gc_marker = kmalloc(sizeof(*unp_gc_marker), M_UNPCB,
1643 unp_gc_marker->unp_flags |= UNP_MARKER;
1646 * Create taskqueue for defered discard, and stick it to
1649 unp_taskqueue = taskqueue_create("unp_taskq", M_WAITOK,
1650 taskqueue_thread_enqueue, &unp_taskqueue);
1651 taskqueue_start_threads(&unp_taskqueue, 1, TDPRI_KERN_DAEMON,
1652 ncpus - 1, "unp taskq");
1656 unp_internalize(struct mbuf *control, struct thread *td)
1658 struct proc *p = td->td_proc;
1659 struct filedesc *fdescp;
1660 struct cmsghdr *cm = mtod(control, struct cmsghdr *);
1664 struct cmsgcred *cmcred;
1671 if ((cm->cmsg_type != SCM_RIGHTS && cm->cmsg_type != SCM_CREDS) ||
1672 cm->cmsg_level != SOL_SOCKET ||
1673 CMSG_ALIGN(cm->cmsg_len) != control->m_len)
1677 * Fill in credential information.
1679 if (cm->cmsg_type == SCM_CREDS) {
1680 cmcred = (struct cmsgcred *)CMSG_DATA(cm);
1681 cmcred->cmcred_pid = p->p_pid;
1682 cmcred->cmcred_uid = p->p_ucred->cr_ruid;
1683 cmcred->cmcred_gid = p->p_ucred->cr_rgid;
1684 cmcred->cmcred_euid = p->p_ucred->cr_uid;
1685 cmcred->cmcred_ngroups = MIN(p->p_ucred->cr_ngroups,
1687 for (i = 0; i < cmcred->cmcred_ngroups; i++)
1688 cmcred->cmcred_groups[i] = p->p_ucred->cr_groups[i];
1693 * cmsghdr may not be aligned, do not allow calculation(s) to
1696 if (cm->cmsg_len < CMSG_LEN(0))
1699 oldfds = (cm->cmsg_len - CMSG_LEN(0)) / sizeof(int);
1702 * Now replace the integer FDs with pointers to
1703 * the associated global file table entry..
1704 * Allocate a bigger buffer as necessary. But if an cluster is not
1705 * enough, return E2BIG.
1707 newlen = CMSG_LEN(oldfds * sizeof(struct file *));
1708 if (newlen > MCLBYTES)
1710 if (newlen - control->m_len > M_TRAILINGSPACE(control)) {
1711 if (control->m_flags & M_EXT)
1713 MCLGET(control, M_WAITOK);
1714 if (!(control->m_flags & M_EXT))
1717 /* copy the data to the cluster */
1718 memcpy(mtod(control, char *), cm, cm->cmsg_len);
1719 cm = mtod(control, struct cmsghdr *);
1722 lwkt_gettoken(&unp_rights_token);
1725 spin_lock_shared(&fdescp->fd_spin);
1728 * check that all the FDs passed in refer to legal OPEN files
1729 * If not, reject the entire operation.
1731 fdp = (int *)CMSG_DATA(cm);
1732 for (i = 0; i < oldfds; i++) {
1734 if ((unsigned)fd >= fdescp->fd_nfiles ||
1735 fdescp->fd_files[fd].fp == NULL) {
1739 if (fdescp->fd_files[fd].fp->f_type == DTYPE_KQUEUE) {
1746 * Adjust length, in case sizeof(struct file *) and sizeof(int)
1749 cm->cmsg_len = newlen;
1750 control->m_len = CMSG_ALIGN(newlen);
1753 * Transform the file descriptors into struct file pointers.
1754 * Since the sizeof(struct file *) is bigger than or equal to
1755 * the sizeof(int), we do it in reverse order so that the int
1756 * won't get trashed until we're done.
1758 fdp = (int *)CMSG_DATA(cm) + oldfds - 1;
1759 rp = (struct file **)CMSG_DATA(cm) + oldfds - 1;
1760 for (i = 0; i < oldfds; i++) {
1761 fp = fdescp->fd_files[*fdp--].fp;
1768 spin_unlock_shared(&fdescp->fd_spin);
1769 lwkt_reltoken(&unp_rights_token);
1773 #ifdef UNP_GC_ALLFILES
1776 * Garbage collect in-transit file descriptors that get lost due to
1777 * loops (i.e. when a socket is sent to another process over itself,
1778 * and more complex situations).
1780 * NOT MPSAFE - TODO socket flush code and maybe fdrop. Rest is MPSAFE.
1783 struct unp_gc_info {
1784 struct file **extra_ref;
1785 struct file *locked_fp;
1792 unp_gc(void *arg __unused, int pending __unused)
1794 struct unp_gc_info info;
1798 lwkt_gettoken(&unp_rights_token);
1801 * Before going through all this, set all FDs to be NOT defered
1802 * and NOT externally accessible (not marked). During the scan
1803 * a fd can be marked externally accessible but we may or may not
1804 * be able to immediately process it (controlled by FDEFER).
1806 * If we loop sleep a bit. The complexity of the topology can cause
1807 * multiple loops. Also failure to acquire the socket's so_rcv
1808 * token can cause us to loop.
1810 allfiles_scan_exclusive(unp_gc_clearmarks, NULL);
1813 allfiles_scan_exclusive(unp_gc_checkmarks, &info);
1815 tsleep(&info, 0, "gcagain", 1);
1816 } while (info.defer);
1819 * We grab an extra reference to each of the file table entries
1820 * that are not otherwise accessible and then free the rights
1821 * that are stored in messages on them.
1823 * The bug in the orginal code is a little tricky, so I'll describe
1824 * what's wrong with it here.
1826 * It is incorrect to simply unp_discard each entry for f_msgcount
1827 * times -- consider the case of sockets A and B that contain
1828 * references to each other. On a last close of some other socket,
1829 * we trigger a gc since the number of outstanding rights (unp_rights)
1830 * is non-zero. If during the sweep phase the gc code unp_discards,
1831 * we end up doing a (full) fdrop on the descriptor. A fdrop on A
1832 * results in the following chain. Closef calls soo_close, which
1833 * calls soclose. Soclose calls first (through the switch
1834 * uipc_usrreq) unp_detach, which re-invokes unp_gc. Unp_gc simply
1835 * returns because the previous instance had set unp_gcing, and
1836 * we return all the way back to soclose, which marks the socket
1837 * with SS_NOFDREF, and then calls sofree. Sofree calls sorflush
1838 * to free up the rights that are queued in messages on the socket A,
1839 * i.e., the reference on B. The sorflush calls via the dom_dispose
1840 * switch unp_dispose, which unp_scans with unp_discard. This second
1841 * instance of unp_discard just calls fdrop on B.
1843 * Well, a similar chain occurs on B, resulting in a sorflush on B,
1844 * which results in another fdrop on A. Unfortunately, A is already
1845 * being closed, and the descriptor has already been marked with
1846 * SS_NOFDREF, and soclose panics at this point.
1848 * Here, we first take an extra reference to each inaccessible
1849 * descriptor. Then, we call sorflush ourself, since we know
1850 * it is a Unix domain socket anyhow. After we destroy all the
1851 * rights carried in messages, we do a last fdrop to get rid
1852 * of our extra reference. This is the last close, and the
1853 * unp_detach etc will shut down the socket.
1855 * 91/09/19, bsy@cs.cmu.edu
1857 info.extra_ref = kmalloc(256 * sizeof(struct file *), M_FILE, M_WAITOK);
1858 info.maxindex = 256;
1865 allfiles_scan_exclusive(unp_gc_checkrefs, &info);
1868 * For each FD on our hit list, do the following two things
1870 for (i = info.index, fpp = info.extra_ref; --i >= 0; ++fpp) {
1871 struct file *tfp = *fpp;
1872 if (tfp->f_type == DTYPE_SOCKET && tfp->f_data != NULL)
1873 sorflush((struct socket *)(tfp->f_data));
1875 for (i = info.index, fpp = info.extra_ref; --i >= 0; ++fpp)
1877 } while (info.index == info.maxindex);
1879 kfree((caddr_t)info.extra_ref, M_FILE);
1881 lwkt_reltoken(&unp_rights_token);
1885 * MPSAFE - NOTE: filehead list and file pointer spinlocked on entry
1888 unp_gc_checkrefs(struct file *fp, void *data)
1890 struct unp_gc_info *info = data;
1892 if (fp->f_count == 0)
1894 if (info->index == info->maxindex)
1898 * If all refs are from msgs, and it's not marked accessible
1899 * then it must be referenced from some unreachable cycle
1900 * of (shut-down) FDs, so include it in our
1901 * list of FDs to remove
1903 if (fp->f_count == fp->f_msgcount && !(fp->f_flag & FMARK)) {
1904 info->extra_ref[info->index++] = fp;
1911 * MPSAFE - NOTE: filehead list and file pointer spinlocked on entry
1914 unp_gc_clearmarks(struct file *fp, void *data __unused)
1916 atomic_clear_int(&fp->f_flag, FMARK | FDEFER);
1921 * MPSAFE - NOTE: filehead list and file pointer spinlocked on entry
1924 unp_gc_checkmarks(struct file *fp, void *data)
1926 struct unp_gc_info *info = data;
1930 * If the file is not open, skip it. Make sure it isn't marked
1931 * defered or we could loop forever, in case we somehow race
1934 if (fp->f_count == 0) {
1935 if (fp->f_flag & FDEFER)
1936 atomic_clear_int(&fp->f_flag, FDEFER);
1940 * If we already marked it as 'defer' in a
1941 * previous pass, then try process it this time
1944 if (fp->f_flag & FDEFER) {
1945 atomic_clear_int(&fp->f_flag, FDEFER);
1948 * if it's not defered, then check if it's
1949 * already marked.. if so skip it
1951 if (fp->f_flag & FMARK)
1954 * If all references are from messages
1955 * in transit, then skip it. it's not
1956 * externally accessible.
1958 if (fp->f_count == fp->f_msgcount)
1961 * If it got this far then it must be
1962 * externally accessible.
1964 atomic_set_int(&fp->f_flag, FMARK);
1968 * either it was defered, or it is externally
1969 * accessible and not already marked so.
1970 * Now check if it is possibly one of OUR sockets.
1972 if (fp->f_type != DTYPE_SOCKET ||
1973 (so = (struct socket *)fp->f_data) == NULL) {
1976 if (so->so_proto->pr_domain != &localdomain ||
1977 !(so->so_proto->pr_flags & PR_RIGHTS)) {
1982 * So, Ok, it's one of our sockets and it IS externally accessible
1983 * (or was defered). Now we look to see if we hold any file
1984 * descriptors in its message buffers. Follow those links and mark
1985 * them as accessible too.
1987 * We are holding multiple spinlocks here, if we cannot get the
1988 * token non-blocking defer until the next loop.
1990 info->locked_fp = fp;
1991 if (lwkt_trytoken(&so->so_rcv.ssb_token)) {
1992 unp_scan(so->so_rcv.ssb_mb, unp_mark, info);
1993 lwkt_reltoken(&so->so_rcv.ssb_token);
1995 atomic_set_int(&fp->f_flag, FDEFER);
2002 * Mark visibility. info->defer is recalculated on every pass.
2005 unp_mark(struct file *fp, void *data)
2007 struct unp_gc_info *info = data;
2009 if ((fp->f_flag & FMARK) == 0) {
2011 atomic_set_int(&fp->f_flag, FMARK | FDEFER);
2012 } else if (fp->f_flag & FDEFER) {
2017 #else /* !UNP_GC_ALLFILES */
2020 * They are thread local and do not require explicit synchronization.
2022 static int unp_marked;
2023 static int unp_unreachable;
2026 unp_accessable(struct file *fp, void *data __unused)
2030 if ((unp = unp_fp2unpcb(fp)) == NULL)
2032 if (unp->unp_gcflags & UNPGC_REF)
2034 unp->unp_gcflags &= ~UNPGC_DEAD;
2035 unp->unp_gcflags |= UNPGC_REF;
2040 unp_gc_process(struct unpcb *unp)
2044 /* Already processed. */
2045 if (unp->unp_gcflags & UNPGC_SCANNED)
2050 * Check for a socket potentially in a cycle. It must be in a
2051 * queue as indicated by msgcount, and this must equal the file
2052 * reference count. Note that when msgcount is 0 the file is NULL.
2054 if ((unp->unp_gcflags & UNPGC_REF) == 0 && fp &&
2055 unp->unp_msgcount != 0 && fp->f_count == unp->unp_msgcount) {
2056 unp->unp_gcflags |= UNPGC_DEAD;
2062 * Mark all sockets we reference with RIGHTS.
2064 if (UNP_ISATTACHED(unp)) {
2065 struct signalsockbuf *ssb = &unp->unp_socket->so_rcv;
2068 lwkt_gettoken(&ssb->ssb_token);
2070 * unp_token would be temporarily dropped, if getting
2071 * so_rcv token blocks, so we need to check unp state
2074 if (UNP_ISATTACHED(unp))
2075 unp_scan(ssb->ssb_mb, unp_accessable, NULL);
2076 lwkt_reltoken(&ssb->ssb_token);
2077 unp->unp_gcflags |= UNPGC_SCANNED;
2080 unp->unp_gcflags |= UNPGC_SCANNED;
2085 unp_gc(void *arg __unused, int pending __unused)
2087 struct unp_global_head *head;
2088 int h, filemax, fileidx, filetot;
2089 struct file **unref;
2092 lwkt_gettoken(&unp_rights_token);
2093 lwkt_gettoken(&unp_token);
2096 * First clear all gc flags from previous runs.
2098 for (h = 0; unp_heads[h] != NULL; ++h) {
2100 * NOTE: This loop does not block, so it is safe
2101 * to use TAILQ_FOREACH here.
2103 head = unp_heads[h];
2104 TAILQ_FOREACH(unp, &head->list, unp_link)
2105 unp->unp_gcflags = 0;
2109 * Scan marking all reachable sockets with UNPGC_REF. Once a socket
2110 * is reachable all of the sockets it references are reachable.
2111 * Stop the scan once we do a complete loop without discovering
2112 * a new reachable socket.
2115 unp_unreachable = 0;
2117 for (h = 0; unp_heads[h] != NULL; ++h) {
2118 head = unp_heads[h];
2119 TAILQ_INSERT_HEAD(&head->list, unp_gc_marker, unp_link);
2120 while ((unp = TAILQ_NEXT(unp_gc_marker, unp_link))
2122 TAILQ_REMOVE(&head->list, unp_gc_marker,
2124 TAILQ_INSERT_AFTER(&head->list, unp,
2125 unp_gc_marker, unp_link);
2127 if (unp->unp_flags & UNP_MARKER)
2129 unp_gc_process(unp);
2131 TAILQ_REMOVE(&head->list, unp_gc_marker, unp_link);
2133 } while (unp_marked);
2135 if (unp_unreachable == 0)
2139 * We grab an extra reference to each of the file table entries
2140 * that are not otherwise accessible and then free the rights
2141 * that are stored in messages on them.
2143 * The bug in the orginal code is a little tricky, so I'll describe
2144 * what's wrong with it here.
2146 * It is incorrect to simply unp_discard each entry for f_msgcount
2147 * times -- consider the case of sockets A and B that contain
2148 * references to each other. On a last close of some other socket,
2149 * we trigger a gc since the number of outstanding rights (unp_rights)
2150 * is non-zero. If during the sweep phase the gc code unp_discards,
2151 * we end up doing a (full) fdrop on the descriptor. A fdrop on A
2152 * results in the following chain. Closef calls soo_close, which
2153 * calls soclose. Soclose calls first (through the switch
2154 * uipc_usrreq) unp_detach, which re-invokes unp_gc. Unp_gc simply
2155 * returns because the previous instance had set unp_gcing, and
2156 * we return all the way back to soclose, which marks the socket
2157 * with SS_NOFDREF, and then calls sofree. Sofree calls sorflush
2158 * to free up the rights that are queued in messages on the socket A,
2159 * i.e., the reference on B. The sorflush calls via the dom_dispose
2160 * switch unp_dispose, which unp_scans with unp_discard. This second
2161 * instance of unp_discard just calls fdrop on B.
2163 * Well, a similar chain occurs on B, resulting in a sorflush on B,
2164 * which results in another fdrop on A. Unfortunately, A is already
2165 * being closed, and the descriptor has already been marked with
2166 * SS_NOFDREF, and soclose panics at this point.
2168 * Here, we first take an extra reference to each inaccessible
2169 * descriptor. Then, we call sorflush ourself, since we know
2170 * it is a Unix domain socket anyhow. After we destroy all the
2171 * rights carried in messages, we do a last fdrop to get rid
2172 * of our extra reference. This is the last close, and the
2173 * unp_detach etc will shut down the socket.
2175 * 91/09/19, bsy@cs.cmu.edu
2178 filemax = unp_unreachable;
2179 if (filemax > UNP_GCFILE_MAX)
2180 filemax = UNP_GCFILE_MAX;
2181 unref = kmalloc(filemax * sizeof(struct file *), M_TEMP, M_WAITOK);
2188 * Iterate looking for sockets which have been specifically
2189 * marked as as unreachable and store them locally.
2192 for (h = 0; unp_heads[h] != NULL; ++h) {
2194 * NOTE: This loop does not block, so it is safe
2195 * to use TAILQ_FOREACH here.
2197 head = unp_heads[h];
2198 TAILQ_FOREACH(unp, &head->list, unp_link) {
2201 if ((unp->unp_gcflags & UNPGC_DEAD) == 0)
2203 unp->unp_gcflags &= ~UNPGC_DEAD;
2206 if (unp->unp_msgcount == 0 || fp == NULL ||
2207 fp->f_count != unp->unp_msgcount)
2211 KASSERT(fileidx < filemax,
2212 ("invalid fileidx %d, filemax %d",
2214 unref[fileidx++] = fp;
2216 KASSERT(filetot < unp_unreachable,
2217 ("invalid filetot %d and "
2218 "unp_unreachable %d",
2219 filetot, unp_unreachable));
2222 if (fileidx == filemax ||
2223 filetot == unp_unreachable)
2229 * For each Unix domain socket on our hit list, do the
2230 * following two things.
2232 for (i = 0; i < fileidx; ++i)
2233 sorflush(unref[i]->f_data);
2234 for (i = 0; i < fileidx; ++i)
2236 } while (fileidx == filemax && filetot < unp_unreachable);
2237 kfree(unref, M_TEMP);
2239 lwkt_reltoken(&unp_token);
2240 lwkt_reltoken(&unp_rights_token);
2243 #endif /* UNP_GC_ALLFILES */
2246 * Dispose of the fp's stored in a mbuf.
2248 * The dds loop can cause additional fps to be entered onto the
2249 * list while it is running, flattening out the operation and avoiding
2250 * a deep kernel stack recursion.
2253 unp_dispose(struct mbuf *m)
2255 lwkt_gettoken(&unp_rights_token);
2257 unp_scan(m, unp_discard, NULL);
2258 lwkt_reltoken(&unp_rights_token);
2262 unp_listen(struct unpcb *unp, struct thread *td)
2264 struct proc *p = td->td_proc;
2266 ASSERT_LWKT_TOKEN_HELD(&unp_token);
2267 UNP_ASSERT_TOKEN_HELD(unp);
2270 cru2x(p->p_ucred, &unp->unp_peercred);
2271 unp_setflags(unp, UNP_HAVEPCCACHED);
2276 unp_scan(struct mbuf *m0, void (*op)(struct file *, void *), void *data)
2285 for (m = m0; m; m = m->m_next) {
2286 if (m->m_type == MT_CONTROL &&
2287 m->m_len >= sizeof(*cm)) {
2288 cm = mtod(m, struct cmsghdr *);
2289 if (cm->cmsg_level != SOL_SOCKET ||
2290 cm->cmsg_type != SCM_RIGHTS)
2292 qfds = (cm->cmsg_len - CMSG_LEN(0)) /
2294 rp = (struct file **)CMSG_DATA(cm);
2295 for (i = 0; i < qfds; i++)
2297 break; /* XXX, but saves time */
2305 * Discard a fp previously held in a unix domain socket mbuf. To
2306 * avoid blowing out the kernel stack due to contrived chain-reactions
2307 * we may have to defer the operation to a dedicated taskqueue.
2309 * Caller holds unp_rights_token.
2312 unp_discard(struct file *fp, void *data __unused)
2315 if (unp_fp2unpcb(fp) != NULL) {
2316 struct unp_defdiscard *d;
2319 * This fp is a Unix domain socket itself and fdrop()
2320 * it here directly may cause deep unp_discard()
2321 * recursion, so the fdrop() is defered to the
2322 * dedicated taskqueue.
2324 d = kmalloc(sizeof(*d), M_UNPCB, M_WAITOK);
2327 spin_lock(&unp_defdiscard_spin);
2328 SLIST_INSERT_HEAD(&unp_defdiscard_head, d, next);
2329 spin_unlock(&unp_defdiscard_spin);
2331 taskqueue_enqueue(unp_taskqueue, &unp_defdiscard_task);
2333 /* This fp is not a Unix domain socket */
2340 * unp_token must be held before calling this function to avoid name
2341 * resolution and v_socket accessing races, especially racing against
2345 * For anyone caring about unconnected Unix domain socket sending
2346 * performance, other approach could be taken...
2349 unp_find_lockref(struct sockaddr *nam, struct thread *td, short type,
2350 struct unpcb **unp_ret)
2352 struct proc *p = td->td_proc;
2353 struct sockaddr_un *soun = (struct sockaddr_un *)nam;
2354 struct vnode *vp = NULL;
2358 struct nlookupdata nd;
2359 char buf[SOCK_MAXADDRLEN];
2361 ASSERT_LWKT_TOKEN_HELD(&unp_token);
2365 len = nam->sa_len - offsetof(struct sockaddr_un, sun_path);
2370 strncpy(buf, soun->sun_path, len);
2373 error = nlookup_init(&nd, buf, UIO_SYSSPACE, NLC_FOLLOW);
2375 error = nlookup(&nd);
2377 error = cache_vget(&nd.nl_nch, nd.nl_cred, LK_EXCLUSIVE, &vp);
2384 if (vp->v_type != VSOCK) {
2388 error = VOP_EACCESS(vp, VWRITE, p->p_ucred);
2393 error = ECONNREFUSED;
2396 if (so->so_type != type) {
2401 /* Lock this unp. */
2402 unp = unp_getsocktoken(so);
2403 if (!UNP_ISATTACHED(unp)) {
2405 error = ECONNREFUSED;
2408 /* And keep this unp referenced. */
2421 unp_connect_pair(struct unpcb *unp, struct unpcb *unp2)
2423 struct socket *so = unp->unp_socket;
2424 struct socket *so2 = unp2->unp_socket;
2426 ASSERT_LWKT_TOKEN_HELD(&unp_token);
2427 UNP_ASSERT_TOKEN_HELD(unp);
2428 UNP_ASSERT_TOKEN_HELD(unp2);
2430 KASSERT(so->so_type == so2->so_type,
2431 ("socket type mismatch, so %d, so2 %d", so->so_type, so2->so_type));
2433 if (!UNP_ISATTACHED(unp))
2435 if (!UNP_ISATTACHED(unp2))
2436 return ECONNREFUSED;
2438 KASSERT(unp->unp_conn == NULL, ("unp is already connected"));
2439 unp->unp_conn = unp2;
2441 switch (so->so_type) {
2443 LIST_INSERT_HEAD(&unp2->unp_refs, unp, unp_reflink);
2448 case SOCK_SEQPACKET:
2449 KASSERT(unp2->unp_conn == NULL, ("unp2 is already connected"));
2450 unp2->unp_conn = unp;
2456 panic("unp_connect_pair: unknown socket type %d", so->so_type);
2462 unp_drop(struct unpcb *unp, int error)
2464 struct unp_global_head *head;
2467 ASSERT_LWKT_TOKEN_HELD(&unp_token);
2468 UNP_ASSERT_TOKEN_HELD(unp);
2470 KASSERT((unp->unp_flags & (UNP_DETACHED | UNP_DROPPED)) == 0,
2471 ("unp is dropped"));
2473 /* Mark this unp as detached. */
2474 unp_setflags(unp, UNP_DETACHED);
2476 /* Remove this unp from the global unp list. */
2477 head = unp_globalhead(unp->unp_socket->so_type);
2478 KASSERT(head->count > 0, ("invalid unp count"));
2479 TAILQ_REMOVE(&head->list, unp, unp_link);
2482 /* Disconnect all. */
2483 unp_disconnect(unp, error);
2484 while ((unp2 = LIST_FIRST(&unp->unp_refs)) != NULL) {
2485 lwkt_getpooltoken(unp2);
2486 unp_disconnect(unp2, ECONNRESET);
2487 lwkt_relpooltoken(unp2);
2489 unp_setflags(unp, UNP_DROPPED);
2491 /* Try freeing this unp. */
2496 unp_defdiscard_taskfunc(void *arg __unused, int pending __unused)
2498 struct unp_defdiscard *d;
2500 spin_lock(&unp_defdiscard_spin);
2501 while ((d = SLIST_FIRST(&unp_defdiscard_head)) != NULL) {
2502 SLIST_REMOVE_HEAD(&unp_defdiscard_head, next);
2503 spin_unlock(&unp_defdiscard_spin);
2508 spin_lock(&unp_defdiscard_spin);
2510 spin_unlock(&unp_defdiscard_spin);