1 /* 2 * Copyright (c) 1982, 1986, 1989, 1991, 1993 3 * The Regents of the University of California. All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 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. 16 * 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 27 * SUCH DAMAGE. 28 * 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 $ 31 */ 32 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> */ 39 #include <sys/proc.h> 40 #include <sys/file.h> 41 #include <sys/filedesc.h> 42 #include <sys/mbuf.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> 48 #include <sys/stat.h> 49 #include <sys/mount.h> 50 #include <sys/sysctl.h> 51 #include <sys/un.h> 52 #include <sys/unpcb.h> 53 #include <sys/vnode.h> 54 #include <sys/kern_syscall.h> 55 #include <sys/taskqueue.h> 56 57 #include <sys/file2.h> 58 #include <sys/spinlock2.h> 59 #include <sys/socketvar2.h> 60 #include <sys/msgport2.h> 61 62 /* 63 * Unix communications domain. 64 * 65 * TODO: 66 * RDM 67 * rethink name space problems 68 * need a proper out-of-band 69 * lock pushdown 70 * 71 * 72 * Unix domain sockets GC. 73 * 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. 80 * 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. 91 */ 92 93 #define UNP_DETACHED UNP_PRIVATE1 94 #define UNP_CONNECTING UNP_PRIVATE2 95 #define UNP_DROPPED UNP_PRIVATE3 96 #define UNP_MARKER UNP_PRIVATE4 97 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. */ 101 102 #define UNP_GCFILE_MAX 256 103 104 /* For unp_internalize() and unp_externalize() */ 105 CTASSERT(sizeof(struct file *) >= sizeof(int)); 106 107 #define UNP_ISATTACHED(unp) \ 108 ((unp) != NULL && ((unp)->unp_flags & UNP_DETACHED) == 0) 109 110 #ifdef INVARIANTS 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 */ 116 117 struct unp_defdiscard { 118 SLIST_ENTRY(unp_defdiscard) next; 119 struct file *fp; 120 }; 121 SLIST_HEAD(unp_defdiscard_list, unp_defdiscard); 122 123 TAILQ_HEAD(unpcb_qhead, unpcb); 124 struct unp_global_head { 125 struct unpcb_qhead list; 126 int count; 127 }; 128 129 static MALLOC_DEFINE(M_UNPCB, "unpcb", "unpcb struct"); 130 static unp_gen_t unp_gencnt; 131 132 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; 135 136 static struct unp_global_head * const unp_heads[] = 137 { &unp_stream_head, &unp_dgram_head, &unp_seqpkt_head, NULL }; 138 139 static struct lwkt_token unp_token = LWKT_TOKEN_INITIALIZER(unp_token); 140 static struct taskqueue *unp_taskqueue; 141 142 static struct unp_defdiscard_list unp_defdiscard_head; 143 static struct spinlock unp_defdiscard_spin; 144 static struct task unp_defdiscard_task; 145 146 static struct sockaddr sun_noname = { sizeof(sun_noname), AF_LOCAL }; 147 148 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 *, 152 struct thread *); 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); 161 #endif 162 static void unp_scan (struct mbuf *, void (*)(struct file *, void *), 163 void *data); 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, 168 int flags); 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); 174 175 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; 180 181 SYSCTL_DECL(_net_local); 182 SYSCTL_INT(_net_local, OID_AUTO, inflight, CTLFLAG_RD, &unp_rights, 0, 183 "File descriptors in flight"); 184 185 /* 186 * SMP Considerations: 187 * 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. 191 * 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. 194 * 195 * Any access to so_pcb to obtain unp requires the pool token for 196 * unp to be held. 197 */ 198 199 static __inline void 200 unp_reference(struct unpcb *unp) 201 { 202 /* 0->1 transition will not work */ 203 KKASSERT(unp->unp_refcnt > 0); 204 atomic_add_int(&unp->unp_refcnt, 1); 205 } 206 207 static __inline void 208 unp_free(struct unpcb *unp) 209 { 210 KKASSERT(unp->unp_refcnt > 0); 211 if (atomic_fetchadd_int(&unp->unp_refcnt, -1) == 1) 212 unp_detach(unp); 213 } 214 215 static __inline struct unpcb * 216 unp_getsocktoken(struct socket *so) 217 { 218 struct unpcb *unp; 219 220 /* 221 * The unp pointer is invalid until we verify that it is 222 * good by re-checking so_pcb AFTER obtaining the token. 223 */ 224 while ((unp = so->so_pcb) != NULL) { 225 lwkt_getpooltoken(unp); 226 if (unp == so->so_pcb) 227 break; 228 lwkt_relpooltoken(unp); 229 } 230 return unp; 231 } 232 233 static __inline void 234 unp_reltoken(struct unpcb *unp) 235 { 236 if (unp != NULL) 237 lwkt_relpooltoken(unp); 238 } 239 240 static __inline void 241 unp_setflags(struct unpcb *unp, int flags) 242 { 243 atomic_set_int(&unp->unp_flags, flags); 244 } 245 246 static __inline void 247 unp_clrflags(struct unpcb *unp, int flags) 248 { 249 atomic_clear_int(&unp->unp_flags, flags); 250 } 251 252 static __inline struct unp_global_head * 253 unp_globalhead(short type) 254 { 255 switch (type) { 256 case SOCK_STREAM: 257 return &unp_stream_head; 258 case SOCK_DGRAM: 259 return &unp_dgram_head; 260 case SOCK_SEQPACKET: 261 return &unp_seqpkt_head; 262 default: 263 panic("unknown socket type %d", type); 264 } 265 } 266 267 static __inline struct unpcb * 268 unp_fp2unpcb(struct file *fp) 269 { 270 struct socket *so; 271 272 if (fp->f_type != DTYPE_SOCKET) 273 return NULL; 274 275 so = fp->f_data; 276 if (so == NULL) 277 return NULL; 278 279 if (so->so_proto->pr_domain != &localdomain) 280 return NULL; 281 282 return so->so_pcb; 283 } 284 285 static __inline void 286 unp_add_right(struct file *fp) 287 { 288 struct unpcb *unp; 289 290 ASSERT_LWKT_TOKEN_HELD(&unp_rights_token); 291 KASSERT(fp->f_count > 0, ("invalid f_count %d", fp->f_count)); 292 293 unp = unp_fp2unpcb(fp); 294 if (unp != NULL) { 295 unp->unp_fp = fp; 296 unp->unp_msgcount++; 297 } 298 fp->f_msgcount++; 299 unp_rights++; 300 } 301 302 static __inline void 303 unp_del_right(struct file *fp) 304 { 305 struct unpcb *unp; 306 307 ASSERT_LWKT_TOKEN_HELD(&unp_rights_token); 308 KASSERT(fp->f_count > 0, ("invalid f_count %d", fp->f_count)); 309 310 unp = unp_fp2unpcb(fp); 311 if (unp != NULL) { 312 KASSERT(unp->unp_msgcount > 0, 313 ("invalid unp msgcount %d", unp->unp_msgcount)); 314 unp->unp_msgcount--; 315 if (unp->unp_msgcount == 0) 316 unp->unp_fp = NULL; 317 } 318 fp->f_msgcount--; 319 unp_rights--; 320 } 321 322 /* 323 * NOTE: (so) is referenced from soabort*() and netmsg_pru_abort() 324 * will sofree() it when we return. 325 */ 326 static void 327 uipc_abort(netmsg_t msg) 328 { 329 struct unpcb *unp; 330 int error; 331 332 lwkt_gettoken(&unp_token); 333 unp = unp_getsocktoken(msg->base.nm_so); 334 335 if (UNP_ISATTACHED(unp)) { 336 unp_drop(unp, ECONNABORTED); 337 error = 0; 338 } else { 339 error = EINVAL; 340 } 341 342 unp_reltoken(unp); 343 lwkt_reltoken(&unp_token); 344 345 lwkt_replymsg(&msg->lmsg, error); 346 } 347 348 static void 349 uipc_accept(netmsg_t msg) 350 { 351 struct unpcb *unp; 352 int error; 353 354 lwkt_gettoken(&unp_token); 355 unp = unp_getsocktoken(msg->base.nm_so); 356 357 if (!UNP_ISATTACHED(unp)) { 358 error = EINVAL; 359 } else { 360 struct unpcb *unp2 = unp->unp_conn; 361 362 /* 363 * Pass back name of connected socket, 364 * if it was bound and we are still connected 365 * (our peer may have closed already!). 366 */ 367 if (unp2 && unp2->unp_addr) { 368 unp_reference(unp2); 369 *msg->accept.nm_nam = dup_sockaddr( 370 (struct sockaddr *)unp2->unp_addr); 371 unp_free(unp2); 372 } else { 373 *msg->accept.nm_nam = dup_sockaddr(&sun_noname); 374 } 375 error = 0; 376 } 377 378 unp_reltoken(unp); 379 lwkt_reltoken(&unp_token); 380 381 lwkt_replymsg(&msg->lmsg, error); 382 } 383 384 static void 385 uipc_attach(netmsg_t msg) 386 { 387 int error; 388 389 lwkt_gettoken(&unp_token); 390 391 KASSERT(msg->base.nm_so->so_pcb == NULL, ("double unp attach")); 392 error = unp_attach(msg->base.nm_so, msg->attach.nm_ai); 393 394 lwkt_reltoken(&unp_token); 395 lwkt_replymsg(&msg->lmsg, error); 396 } 397 398 static void 399 uipc_bind(netmsg_t msg) 400 { 401 struct unpcb *unp; 402 int error; 403 404 lwkt_gettoken(&unp_token); 405 unp = unp_getsocktoken(msg->base.nm_so); 406 407 if (UNP_ISATTACHED(unp)) 408 error = unp_bind(unp, msg->bind.nm_nam, msg->bind.nm_td); 409 else 410 error = EINVAL; 411 412 unp_reltoken(unp); 413 lwkt_reltoken(&unp_token); 414 415 lwkt_replymsg(&msg->lmsg, error); 416 } 417 418 static void 419 uipc_connect(netmsg_t msg) 420 { 421 int error; 422 423 error = unp_connect(msg->base.nm_so, msg->connect.nm_nam, 424 msg->connect.nm_td); 425 lwkt_replymsg(&msg->lmsg, error); 426 } 427 428 static void 429 uipc_connect2(netmsg_t msg) 430 { 431 int error; 432 433 error = unp_connect2(msg->connect2.nm_so1, msg->connect2.nm_so2); 434 lwkt_replymsg(&msg->lmsg, error); 435 } 436 437 /* control is EOPNOTSUPP */ 438 439 static void 440 uipc_detach(netmsg_t msg) 441 { 442 struct unpcb *unp; 443 int error; 444 445 lwkt_gettoken(&unp_token); 446 unp = unp_getsocktoken(msg->base.nm_so); 447 448 if (UNP_ISATTACHED(unp)) { 449 unp_drop(unp, 0); 450 error = 0; 451 } else { 452 error = EINVAL; 453 } 454 455 unp_reltoken(unp); 456 lwkt_reltoken(&unp_token); 457 458 lwkt_replymsg(&msg->lmsg, error); 459 } 460 461 static void 462 uipc_disconnect(netmsg_t msg) 463 { 464 struct unpcb *unp; 465 int error; 466 467 lwkt_gettoken(&unp_token); 468 unp = unp_getsocktoken(msg->base.nm_so); 469 470 if (UNP_ISATTACHED(unp)) { 471 unp_disconnect(unp, 0); 472 error = 0; 473 } else { 474 error = EINVAL; 475 } 476 477 unp_reltoken(unp); 478 lwkt_reltoken(&unp_token); 479 480 lwkt_replymsg(&msg->lmsg, error); 481 } 482 483 static void 484 uipc_listen(netmsg_t msg) 485 { 486 struct unpcb *unp; 487 int error; 488 489 lwkt_gettoken(&unp_token); 490 unp = unp_getsocktoken(msg->base.nm_so); 491 492 if (!UNP_ISATTACHED(unp) || unp->unp_vnode == NULL) 493 error = EINVAL; 494 else 495 error = unp_listen(unp, msg->listen.nm_td); 496 497 unp_reltoken(unp); 498 lwkt_reltoken(&unp_token); 499 500 lwkt_replymsg(&msg->lmsg, error); 501 } 502 503 static void 504 uipc_peeraddr(netmsg_t msg) 505 { 506 struct unpcb *unp; 507 int error; 508 509 lwkt_gettoken(&unp_token); 510 unp = unp_getsocktoken(msg->base.nm_so); 511 512 if (!UNP_ISATTACHED(unp)) { 513 error = EINVAL; 514 } else if (unp->unp_conn && unp->unp_conn->unp_addr) { 515 struct unpcb *unp2 = unp->unp_conn; 516 517 unp_reference(unp2); 518 *msg->peeraddr.nm_nam = dup_sockaddr( 519 (struct sockaddr *)unp2->unp_addr); 520 unp_free(unp2); 521 error = 0; 522 } else { 523 /* 524 * XXX: It seems that this test always fails even when 525 * connection is established. So, this else clause is 526 * added as workaround to return PF_LOCAL sockaddr. 527 */ 528 *msg->peeraddr.nm_nam = dup_sockaddr(&sun_noname); 529 error = 0; 530 } 531 532 unp_reltoken(unp); 533 lwkt_reltoken(&unp_token); 534 535 lwkt_replymsg(&msg->lmsg, error); 536 } 537 538 static void 539 uipc_rcvd(netmsg_t msg) 540 { 541 struct unpcb *unp, *unp2; 542 struct socket *so; 543 struct socket *so2; 544 int error; 545 546 /* 547 * so_pcb is only modified with both the global and the unp 548 * pool token held. 549 */ 550 so = msg->base.nm_so; 551 unp = unp_getsocktoken(so); 552 553 if (!UNP_ISATTACHED(unp)) { 554 error = EINVAL; 555 goto done; 556 } 557 558 switch (so->so_type) { 559 case SOCK_DGRAM: 560 panic("uipc_rcvd DGRAM?"); 561 /*NOTREACHED*/ 562 case SOCK_STREAM: 563 case SOCK_SEQPACKET: 564 if (unp->unp_conn == NULL) 565 break; 566 unp2 = unp->unp_conn; /* protected by pool token */ 567 568 /* 569 * Because we are transfering mbufs directly to the 570 * peer socket we have to use SSB_STOP on the sender 571 * to prevent it from building up infinite mbufs. 572 * 573 * As in several places in this module w ehave to ref unp2 574 * to ensure that it does not get ripped out from under us 575 * if we block on the so2 token or in sowwakeup(). 576 */ 577 so2 = unp2->unp_socket; 578 unp_reference(unp2); 579 lwkt_gettoken(&so2->so_rcv.ssb_token); 580 if (so->so_rcv.ssb_cc < so2->so_snd.ssb_hiwat && 581 so->so_rcv.ssb_mbcnt < so2->so_snd.ssb_mbmax 582 ) { 583 atomic_clear_int(&so2->so_snd.ssb_flags, SSB_STOP); 584 585 sowwakeup(so2); 586 } 587 lwkt_reltoken(&so2->so_rcv.ssb_token); 588 unp_free(unp2); 589 break; 590 default: 591 panic("uipc_rcvd unknown socktype"); 592 /*NOTREACHED*/ 593 } 594 error = 0; 595 done: 596 unp_reltoken(unp); 597 lwkt_replymsg(&msg->lmsg, error); 598 } 599 600 /* pru_rcvoob is EOPNOTSUPP */ 601 602 static void 603 uipc_send(netmsg_t msg) 604 { 605 struct unpcb *unp, *unp2; 606 struct socket *so; 607 struct socket *so2; 608 struct mbuf *control; 609 struct mbuf *m; 610 int error = 0; 611 612 so = msg->base.nm_so; 613 control = msg->send.nm_control; 614 m = msg->send.nm_m; 615 616 /* 617 * so_pcb is only modified with both the global and the unp 618 * pool token held. 619 */ 620 so = msg->base.nm_so; 621 unp = unp_getsocktoken(so); 622 623 if (!UNP_ISATTACHED(unp)) { 624 error = EINVAL; 625 goto release; 626 } 627 628 if (msg->send.nm_flags & PRUS_OOB) { 629 error = EOPNOTSUPP; 630 goto release; 631 } 632 633 wakeup_start_delayed(); 634 635 if (control && (error = unp_internalize(control, msg->send.nm_td))) 636 goto release; 637 638 switch (so->so_type) { 639 case SOCK_DGRAM: 640 { 641 struct sockaddr *from; 642 643 if (msg->send.nm_addr) { 644 if (unp->unp_conn) { 645 error = EISCONN; 646 break; 647 } 648 lwkt_gettoken(&unp_token); 649 error = unp_find_lockref(msg->send.nm_addr, 650 msg->send.nm_td, so->so_type, &unp2); 651 if (error) { 652 lwkt_reltoken(&unp_token); 653 break; 654 } 655 /* 656 * NOTE: 657 * unp2 is locked and referenced. 658 * 659 * We could unlock unp2 now, since it was checked 660 * and referenced. 661 */ 662 unp_reltoken(unp2); 663 lwkt_reltoken(&unp_token); 664 } else { 665 if (unp->unp_conn == NULL) { 666 error = ENOTCONN; 667 break; 668 } 669 unp2 = unp->unp_conn; 670 unp_reference(unp2); 671 } 672 /* NOTE: unp2 is referenced. */ 673 so2 = unp2->unp_socket; 674 675 if (unp->unp_addr) 676 from = (struct sockaddr *)unp->unp_addr; 677 else 678 from = &sun_noname; 679 680 lwkt_gettoken(&so2->so_rcv.ssb_token); 681 if (ssb_appendaddr(&so2->so_rcv, from, m, control)) { 682 sorwakeup(so2); 683 m = NULL; 684 control = NULL; 685 } else { 686 error = ENOBUFS; 687 } 688 lwkt_reltoken(&so2->so_rcv.ssb_token); 689 690 unp_free(unp2); 691 break; 692 } 693 694 case SOCK_STREAM: 695 case SOCK_SEQPACKET: 696 /* Connect if not connected yet. */ 697 /* 698 * Note: A better implementation would complain 699 * if not equal to the peer's address. 700 */ 701 if (unp->unp_conn == NULL) { 702 if (msg->send.nm_addr) { 703 error = unp_connect(so, 704 msg->send.nm_addr, 705 msg->send.nm_td); 706 if (error) 707 break; /* XXX */ 708 } 709 /* 710 * NOTE: 711 * unp_conn still could be NULL, even if the 712 * above unp_connect() succeeds; since the 713 * current unp's token could be released due 714 * to blocking operations after unp_conn is 715 * assigned. 716 */ 717 if (unp->unp_conn == NULL) { 718 error = ENOTCONN; 719 break; 720 } 721 } 722 if (so->so_state & SS_CANTSENDMORE) { 723 error = EPIPE; 724 break; 725 } 726 727 unp2 = unp->unp_conn; 728 KASSERT(unp2 != NULL, ("unp is not connected")); 729 so2 = unp2->unp_socket; 730 731 unp_reference(unp2); 732 733 /* 734 * Send to paired receive port, and then reduce 735 * send buffer hiwater marks to maintain backpressure. 736 * Wake up readers. 737 */ 738 lwkt_gettoken(&so2->so_rcv.ssb_token); 739 if (control) { 740 if (ssb_appendcontrol(&so2->so_rcv, m, control)) { 741 control = NULL; 742 m = NULL; 743 } 744 } else if (so->so_type == SOCK_SEQPACKET) { 745 sbappendrecord(&so2->so_rcv.sb, m); 746 m = NULL; 747 } else { 748 sbappend(&so2->so_rcv.sb, m); 749 m = NULL; 750 } 751 752 /* 753 * Because we are transfering mbufs directly to the 754 * peer socket we have to use SSB_STOP on the sender 755 * to prevent it from building up infinite mbufs. 756 */ 757 if (so2->so_rcv.ssb_cc >= so->so_snd.ssb_hiwat || 758 so2->so_rcv.ssb_mbcnt >= so->so_snd.ssb_mbmax 759 ) { 760 atomic_set_int(&so->so_snd.ssb_flags, SSB_STOP); 761 } 762 lwkt_reltoken(&so2->so_rcv.ssb_token); 763 sorwakeup(so2); 764 765 unp_free(unp2); 766 break; 767 768 default: 769 panic("uipc_send unknown socktype"); 770 } 771 772 /* 773 * SEND_EOF is equivalent to a SEND followed by a SHUTDOWN. 774 */ 775 if (msg->send.nm_flags & PRUS_EOF) { 776 socantsendmore(so); 777 unp_shutdown(unp); 778 } 779 780 if (control && error != 0) 781 unp_dispose(control); 782 release: 783 unp_reltoken(unp); 784 wakeup_end_delayed(); 785 786 if (control) 787 m_freem(control); 788 if (m) 789 m_freem(m); 790 lwkt_replymsg(&msg->lmsg, error); 791 } 792 793 /* 794 * MPSAFE 795 */ 796 static void 797 uipc_sense(netmsg_t msg) 798 { 799 struct unpcb *unp; 800 struct socket *so; 801 struct stat *sb; 802 int error; 803 804 so = msg->base.nm_so; 805 sb = msg->sense.nm_stat; 806 807 /* 808 * so_pcb is only modified with both the global and the unp 809 * pool token held. 810 */ 811 unp = unp_getsocktoken(so); 812 813 if (!UNP_ISATTACHED(unp)) { 814 error = EINVAL; 815 goto done; 816 } 817 818 sb->st_blksize = so->so_snd.ssb_hiwat; 819 sb->st_dev = NOUDEV; 820 error = 0; 821 done: 822 unp_reltoken(unp); 823 lwkt_replymsg(&msg->lmsg, error); 824 } 825 826 static void 827 uipc_shutdown(netmsg_t msg) 828 { 829 struct socket *so; 830 struct unpcb *unp; 831 int error; 832 833 /* 834 * so_pcb is only modified with both the global and the unp 835 * pool token held. 836 */ 837 so = msg->base.nm_so; 838 unp = unp_getsocktoken(so); 839 840 if (UNP_ISATTACHED(unp)) { 841 socantsendmore(so); 842 unp_shutdown(unp); 843 error = 0; 844 } else { 845 error = EINVAL; 846 } 847 848 unp_reltoken(unp); 849 lwkt_replymsg(&msg->lmsg, error); 850 } 851 852 static void 853 uipc_sockaddr(netmsg_t msg) 854 { 855 struct unpcb *unp; 856 int error; 857 858 /* 859 * so_pcb is only modified with both the global and the unp 860 * pool token held. 861 */ 862 unp = unp_getsocktoken(msg->base.nm_so); 863 864 if (UNP_ISATTACHED(unp)) { 865 if (unp->unp_addr) { 866 *msg->sockaddr.nm_nam = 867 dup_sockaddr((struct sockaddr *)unp->unp_addr); 868 } 869 error = 0; 870 } else { 871 error = EINVAL; 872 } 873 874 unp_reltoken(unp); 875 lwkt_replymsg(&msg->lmsg, error); 876 } 877 878 struct pr_usrreqs uipc_usrreqs = { 879 .pru_abort = uipc_abort, 880 .pru_accept = uipc_accept, 881 .pru_attach = uipc_attach, 882 .pru_bind = uipc_bind, 883 .pru_connect = uipc_connect, 884 .pru_connect2 = uipc_connect2, 885 .pru_control = pr_generic_notsupp, 886 .pru_detach = uipc_detach, 887 .pru_disconnect = uipc_disconnect, 888 .pru_listen = uipc_listen, 889 .pru_peeraddr = uipc_peeraddr, 890 .pru_rcvd = uipc_rcvd, 891 .pru_rcvoob = pr_generic_notsupp, 892 .pru_send = uipc_send, 893 .pru_sense = uipc_sense, 894 .pru_shutdown = uipc_shutdown, 895 .pru_sockaddr = uipc_sockaddr, 896 .pru_sosend = sosend, 897 .pru_soreceive = soreceive 898 }; 899 900 void 901 uipc_ctloutput(netmsg_t msg) 902 { 903 struct socket *so; 904 struct sockopt *sopt; 905 struct unpcb *unp; 906 int error = 0; 907 908 so = msg->base.nm_so; 909 sopt = msg->ctloutput.nm_sopt; 910 911 lwkt_gettoken(&unp_token); 912 unp = unp_getsocktoken(so); 913 914 if (!UNP_ISATTACHED(unp)) { 915 error = EINVAL; 916 goto done; 917 } 918 919 switch (sopt->sopt_dir) { 920 case SOPT_GET: 921 switch (sopt->sopt_name) { 922 case LOCAL_PEERCRED: 923 if (unp->unp_flags & UNP_HAVEPC) 924 soopt_from_kbuf(sopt, &unp->unp_peercred, 925 sizeof(unp->unp_peercred)); 926 else { 927 if (so->so_type == SOCK_STREAM) 928 error = ENOTCONN; 929 else if (so->so_type == SOCK_SEQPACKET) 930 error = ENOTCONN; 931 else 932 error = EINVAL; 933 } 934 break; 935 default: 936 error = EOPNOTSUPP; 937 break; 938 } 939 break; 940 case SOPT_SET: 941 default: 942 error = EOPNOTSUPP; 943 break; 944 } 945 946 done: 947 unp_reltoken(unp); 948 lwkt_reltoken(&unp_token); 949 950 lwkt_replymsg(&msg->lmsg, error); 951 } 952 953 /* 954 * Both send and receive buffers are allocated PIPSIZ bytes of buffering 955 * for stream sockets, although the total for sender and receiver is 956 * actually only PIPSIZ. 957 * 958 * Datagram sockets really use the sendspace as the maximum datagram size, 959 * and don't really want to reserve the sendspace. Their recvspace should 960 * be large enough for at least one max-size datagram plus address. 961 * 962 * We want the local send/recv space to be significant larger then lo0's 963 * mtu of 16384. 964 * 965 * We no longer need to worry about avoiding the windows scaling option. 966 * Programs which use unix domain sockets expect larger defaults these days. 967 */ 968 #ifndef PIPSIZ 969 #define PIPSIZ 65536 970 #endif 971 static u_long unpst_sendspace = PIPSIZ; 972 static u_long unpst_recvspace = PIPSIZ; 973 static u_long unpdg_sendspace = PIPSIZ; /* really max datagram size */ 974 static u_long unpdg_recvspace = PIPSIZ; 975 static u_long unpsp_sendspace = PIPSIZ; /* really max datagram size */ 976 static u_long unpsp_recvspace = PIPSIZ; 977 978 SYSCTL_DECL(_net_local_stream); 979 SYSCTL_DECL(_net_local_dgram); 980 SYSCTL_DECL(_net_local_seqpacket); 981 982 SYSCTL_ULONG(_net_local_stream, OID_AUTO, sendspace, CTLFLAG_RW, 983 &unpst_sendspace, 0, "Size of stream socket send buffer"); 984 SYSCTL_ULONG(_net_local_stream, OID_AUTO, recvspace, CTLFLAG_RW, 985 &unpst_recvspace, 0, "Size of stream socket receive buffer"); 986 987 SYSCTL_ULONG(_net_local_dgram, OID_AUTO, maxdgram, CTLFLAG_RW, 988 &unpdg_sendspace, 0, "Max datagram socket size"); 989 SYSCTL_ULONG(_net_local_dgram, OID_AUTO, recvspace, CTLFLAG_RW, 990 &unpdg_recvspace, 0, "Size of datagram socket receive buffer"); 991 992 SYSCTL_ULONG(_net_local_seqpacket, OID_AUTO, maxseqpacket, CTLFLAG_RW, 993 &unpsp_sendspace, 0, "Default seqpacket send space."); 994 SYSCTL_ULONG(_net_local_seqpacket, OID_AUTO, recvspace, CTLFLAG_RW, 995 &unpsp_recvspace, 0, "Default seqpacket receive space."); 996 997 998 static int 999 unp_attach(struct socket *so, struct pru_attach_info *ai) 1000 { 1001 struct unp_global_head *head; 1002 struct unpcb *unp; 1003 int error; 1004 1005 lwkt_gettoken(&unp_token); 1006 1007 if (so->so_snd.ssb_hiwat == 0 || so->so_rcv.ssb_hiwat == 0) { 1008 switch (so->so_type) { 1009 case SOCK_STREAM: 1010 error = soreserve(so, unpst_sendspace, unpst_recvspace, 1011 ai->sb_rlimit); 1012 break; 1013 case SOCK_DGRAM: 1014 error = soreserve(so, unpdg_sendspace, unpdg_recvspace, 1015 ai->sb_rlimit); 1016 break; 1017 case SOCK_SEQPACKET: 1018 error = soreserve(so, unpsp_sendspace, unpsp_recvspace, 1019 ai->sb_rlimit); 1020 break; 1021 default: 1022 panic("unp_attach"); 1023 } 1024 if (error) 1025 goto failed; 1026 } 1027 1028 /* 1029 * In order to support sendfile we have to set either SSB_STOPSUPP 1030 * or SSB_PREALLOC. Unix domain sockets use the SSB_STOP flow 1031 * control mechanism. 1032 */ 1033 if (so->so_type == SOCK_STREAM) { 1034 atomic_set_int(&so->so_rcv.ssb_flags, SSB_STOPSUPP); 1035 atomic_set_int(&so->so_snd.ssb_flags, SSB_STOPSUPP); 1036 } 1037 1038 unp = kmalloc(sizeof(*unp), M_UNPCB, M_WAITOK | M_ZERO | M_NULLOK); 1039 if (unp == NULL) { 1040 error = ENOBUFS; 1041 goto failed; 1042 } 1043 unp->unp_refcnt = 1; 1044 unp->unp_gencnt = ++unp_gencnt; 1045 LIST_INIT(&unp->unp_refs); 1046 unp->unp_socket = so; 1047 unp->unp_rvnode = ai->fd_rdir; /* jail cruft XXX JH */ 1048 so->so_pcb = (caddr_t)unp; 1049 soreference(so); 1050 1051 head = unp_globalhead(so->so_type); 1052 TAILQ_INSERT_TAIL(&head->list, unp, unp_link); 1053 head->count++; 1054 error = 0; 1055 failed: 1056 lwkt_reltoken(&unp_token); 1057 return error; 1058 } 1059 1060 static void 1061 unp_detach(struct unpcb *unp) 1062 { 1063 struct socket *so; 1064 1065 lwkt_gettoken(&unp_token); 1066 lwkt_getpooltoken(unp); 1067 1068 so = unp->unp_socket; 1069 1070 unp->unp_gencnt = ++unp_gencnt; 1071 if (unp->unp_vnode) { 1072 unp->unp_vnode->v_socket = NULL; 1073 vrele(unp->unp_vnode); 1074 unp->unp_vnode = NULL; 1075 } 1076 soisdisconnected(so); 1077 KKASSERT(so->so_pcb == unp); 1078 so->so_pcb = NULL; /* both tokens required */ 1079 unp->unp_socket = NULL; 1080 1081 lwkt_relpooltoken(unp); 1082 lwkt_reltoken(&unp_token); 1083 1084 sofree(so); 1085 1086 KASSERT(unp->unp_conn == NULL, ("unp is still connected")); 1087 KASSERT(LIST_EMPTY(&unp->unp_refs), ("unp still has references")); 1088 1089 if (unp->unp_addr) 1090 kfree(unp->unp_addr, M_SONAME); 1091 kfree(unp, M_UNPCB); 1092 1093 if (unp_rights) 1094 taskqueue_enqueue(unp_taskqueue, &unp_gc_task); 1095 } 1096 1097 static int 1098 unp_bind(struct unpcb *unp, struct sockaddr *nam, struct thread *td) 1099 { 1100 struct proc *p = td->td_proc; 1101 struct sockaddr_un *soun = (struct sockaddr_un *)nam; 1102 struct vnode *vp; 1103 struct vattr vattr; 1104 int error, namelen; 1105 struct nlookupdata nd; 1106 char buf[SOCK_MAXADDRLEN]; 1107 1108 ASSERT_LWKT_TOKEN_HELD(&unp_token); 1109 UNP_ASSERT_TOKEN_HELD(unp); 1110 1111 if (unp->unp_vnode != NULL) 1112 return EINVAL; 1113 1114 namelen = soun->sun_len - offsetof(struct sockaddr_un, sun_path); 1115 if (namelen <= 0) 1116 return EINVAL; 1117 strncpy(buf, soun->sun_path, namelen); 1118 buf[namelen] = 0; /* null-terminate the string */ 1119 error = nlookup_init(&nd, buf, UIO_SYSSPACE, 1120 NLC_LOCKVP | NLC_CREATE | NLC_REFDVP); 1121 if (error == 0) 1122 error = nlookup(&nd); 1123 if (error == 0 && nd.nl_nch.ncp->nc_vp != NULL) 1124 error = EADDRINUSE; 1125 if (error) 1126 goto done; 1127 1128 VATTR_NULL(&vattr); 1129 vattr.va_type = VSOCK; 1130 vattr.va_mode = (ACCESSPERMS & ~p->p_fd->fd_cmask); 1131 error = VOP_NCREATE(&nd.nl_nch, nd.nl_dvp, &vp, nd.nl_cred, &vattr); 1132 if (error == 0) { 1133 if (unp->unp_vnode == NULL) { 1134 vp->v_socket = unp->unp_socket; 1135 unp->unp_vnode = vp; 1136 unp->unp_addr = (struct sockaddr_un *)dup_sockaddr(nam); 1137 vn_unlock(vp); 1138 } else { 1139 vput(vp); /* late race */ 1140 error = EINVAL; 1141 } 1142 } 1143 done: 1144 nlookup_done(&nd); 1145 return (error); 1146 } 1147 1148 static int 1149 unp_connect(struct socket *so, struct sockaddr *nam, struct thread *td) 1150 { 1151 struct unpcb *unp, *unp2; 1152 int error, flags = 0; 1153 1154 lwkt_gettoken(&unp_token); 1155 1156 unp = unp_getsocktoken(so); 1157 if (!UNP_ISATTACHED(unp)) { 1158 error = EINVAL; 1159 goto failed; 1160 } 1161 1162 if ((unp->unp_flags & UNP_CONNECTING) || unp->unp_conn != NULL) { 1163 error = EISCONN; 1164 goto failed; 1165 } 1166 1167 flags = UNP_CONNECTING; 1168 unp_setflags(unp, flags); 1169 1170 error = unp_find_lockref(nam, td, so->so_type, &unp2); 1171 if (error) 1172 goto failed; 1173 /* 1174 * NOTE: 1175 * unp2 is locked and referenced. 1176 */ 1177 1178 if (so->so_proto->pr_flags & PR_CONNREQUIRED) { 1179 struct socket *so2, *so3; 1180 struct unpcb *unp3; 1181 1182 so2 = unp2->unp_socket; 1183 if (!(so2->so_options & SO_ACCEPTCONN) || 1184 /* listen is not completed yet */ 1185 !(unp2->unp_flags & UNP_HAVEPCCACHED) || 1186 (so3 = sonewconn_faddr(so2, 0, NULL, 1187 TRUE /* keep ref */)) == NULL) { 1188 error = ECONNREFUSED; 1189 goto done; 1190 } 1191 /* so3 has a socket reference. */ 1192 1193 unp3 = unp_getsocktoken(so3); 1194 if (!UNP_ISATTACHED(unp3)) { 1195 unp_reltoken(unp3); 1196 /* 1197 * Already aborted; we only need to drop the 1198 * socket reference held by sonewconn_faddr(). 1199 */ 1200 sofree(so3); 1201 error = ECONNREFUSED; 1202 goto done; 1203 } 1204 unp_reference(unp3); 1205 /* 1206 * NOTE: 1207 * unp3 is locked and referenced. 1208 */ 1209 1210 /* 1211 * Release so3 socket reference held by sonewconn_faddr(). 1212 * Since we have referenced unp3, neither unp3 nor so3 will 1213 * be destroyed here. 1214 */ 1215 sofree(so3); 1216 1217 if (unp2->unp_addr != NULL) { 1218 unp3->unp_addr = (struct sockaddr_un *) 1219 dup_sockaddr((struct sockaddr *)unp2->unp_addr); 1220 } 1221 1222 /* 1223 * unp_peercred management: 1224 * 1225 * The connecter's (client's) credentials are copied 1226 * from its process structure at the time of connect() 1227 * (which is now). 1228 */ 1229 cru2x(td->td_proc->p_ucred, &unp3->unp_peercred); 1230 unp_setflags(unp3, UNP_HAVEPC); 1231 /* 1232 * The receiver's (server's) credentials are copied 1233 * from the unp_peercred member of socket on which the 1234 * former called listen(); unp_listen() cached that 1235 * process's credentials at that time so we can use 1236 * them now. 1237 */ 1238 KASSERT(unp2->unp_flags & UNP_HAVEPCCACHED, 1239 ("unp_connect: listener without cached peercred")); 1240 memcpy(&unp->unp_peercred, &unp2->unp_peercred, 1241 sizeof(unp->unp_peercred)); 1242 unp_setflags(unp, UNP_HAVEPC); 1243 1244 error = unp_connect_pair(unp, unp3); 1245 if (error) 1246 soabort_direct(so3); 1247 1248 /* Done with unp3 */ 1249 unp_free(unp3); 1250 unp_reltoken(unp3); 1251 } else { 1252 error = unp_connect_pair(unp, unp2); 1253 } 1254 done: 1255 unp_free(unp2); 1256 unp_reltoken(unp2); 1257 failed: 1258 if (flags) 1259 unp_clrflags(unp, flags); 1260 unp_reltoken(unp); 1261 1262 lwkt_reltoken(&unp_token); 1263 return (error); 1264 } 1265 1266 /* 1267 * Connect two unix domain sockets together. 1268 * 1269 * NOTE: Semantics for any change to unp_conn requires that the per-unp 1270 * pool token also be held. 1271 */ 1272 int 1273 unp_connect2(struct socket *so, struct socket *so2) 1274 { 1275 struct unpcb *unp, *unp2; 1276 int error; 1277 1278 lwkt_gettoken(&unp_token); 1279 if (so2->so_type != so->so_type) { 1280 lwkt_reltoken(&unp_token); 1281 return (EPROTOTYPE); 1282 } 1283 unp = unp_getsocktoken(so); 1284 unp2 = unp_getsocktoken(so2); 1285 1286 if (!UNP_ISATTACHED(unp)) { 1287 error = EINVAL; 1288 goto done; 1289 } 1290 if (!UNP_ISATTACHED(unp2)) { 1291 error = ECONNREFUSED; 1292 goto done; 1293 } 1294 1295 if (unp->unp_conn != NULL) { 1296 error = EISCONN; 1297 goto done; 1298 } 1299 if ((so->so_type == SOCK_STREAM || so->so_type == SOCK_SEQPACKET) && 1300 unp2->unp_conn != NULL) { 1301 error = EISCONN; 1302 goto done; 1303 } 1304 1305 error = unp_connect_pair(unp, unp2); 1306 done: 1307 unp_reltoken(unp2); 1308 unp_reltoken(unp); 1309 lwkt_reltoken(&unp_token); 1310 return (error); 1311 } 1312 1313 /* 1314 * Disconnect a unix domain socket pair. 1315 * 1316 * NOTE: Semantics for any change to unp_conn requires that the per-unp 1317 * pool token also be held. 1318 */ 1319 static void 1320 unp_disconnect(struct unpcb *unp, int error) 1321 { 1322 struct socket *so = unp->unp_socket; 1323 struct unpcb *unp2; 1324 1325 ASSERT_LWKT_TOKEN_HELD(&unp_token); 1326 UNP_ASSERT_TOKEN_HELD(unp); 1327 1328 if (error) 1329 so->so_error = error; 1330 1331 while ((unp2 = unp->unp_conn) != NULL) { 1332 lwkt_getpooltoken(unp2); 1333 if (unp2 == unp->unp_conn) 1334 break; 1335 lwkt_relpooltoken(unp2); 1336 } 1337 if (unp2 == NULL) 1338 return; 1339 /* unp2 is locked. */ 1340 1341 KASSERT((unp2->unp_flags & UNP_DROPPED) == 0, ("unp2 was dropped")); 1342 1343 unp->unp_conn = NULL; 1344 1345 switch (so->so_type) { 1346 case SOCK_DGRAM: 1347 LIST_REMOVE(unp, unp_reflink); 1348 soclrstate(so, SS_ISCONNECTED); 1349 break; 1350 1351 case SOCK_STREAM: 1352 case SOCK_SEQPACKET: 1353 /* 1354 * Keep a reference before clearing the unp_conn 1355 * to avoid racing uipc_detach()/uipc_abort() in 1356 * other thread. 1357 */ 1358 unp_reference(unp2); 1359 KASSERT(unp2->unp_conn == unp, ("unp_conn mismatch")); 1360 unp2->unp_conn = NULL; 1361 1362 soisdisconnected(so); 1363 soisdisconnected(unp2->unp_socket); 1364 1365 unp_free(unp2); 1366 break; 1367 } 1368 1369 lwkt_relpooltoken(unp2); 1370 } 1371 1372 #ifdef notdef 1373 void 1374 unp_abort(struct unpcb *unp) 1375 { 1376 lwkt_gettoken(&unp_token); 1377 unp_free(unp); 1378 lwkt_reltoken(&unp_token); 1379 } 1380 #endif 1381 1382 static int 1383 prison_unpcb(struct thread *td, struct unpcb *unp) 1384 { 1385 struct proc *p; 1386 1387 if (td == NULL) 1388 return (0); 1389 if ((p = td->td_proc) == NULL) 1390 return (0); 1391 if (!p->p_ucred->cr_prison) 1392 return (0); 1393 if (p->p_fd->fd_rdir == unp->unp_rvnode) 1394 return (0); 1395 return (1); 1396 } 1397 1398 static int 1399 unp_pcblist(SYSCTL_HANDLER_ARGS) 1400 { 1401 struct unp_global_head *head = arg1; 1402 int error, i, n; 1403 struct unpcb *unp, *marker; 1404 1405 KKASSERT(curproc != NULL); 1406 1407 /* 1408 * The process of preparing the PCB list is too time-consuming and 1409 * resource-intensive to repeat twice on every request. 1410 */ 1411 if (req->oldptr == NULL) { 1412 n = head->count; 1413 req->oldidx = (n + n/8) * sizeof(struct xunpcb); 1414 return 0; 1415 } 1416 1417 if (req->newptr != NULL) 1418 return EPERM; 1419 1420 marker = kmalloc(sizeof(*marker), M_UNPCB, M_WAITOK | M_ZERO); 1421 marker->unp_flags |= UNP_MARKER; 1422 1423 lwkt_gettoken(&unp_token); 1424 1425 n = head->count; 1426 i = 0; 1427 error = 0; 1428 1429 TAILQ_INSERT_HEAD(&head->list, marker, unp_link); 1430 while ((unp = TAILQ_NEXT(marker, unp_link)) != NULL && i < n) { 1431 struct xunpcb xu; 1432 1433 TAILQ_REMOVE(&head->list, marker, unp_link); 1434 TAILQ_INSERT_AFTER(&head->list, unp, marker, unp_link); 1435 1436 if (unp->unp_flags & UNP_MARKER) 1437 continue; 1438 if (prison_unpcb(req->td, unp)) 1439 continue; 1440 1441 xu.xu_len = sizeof(xu); 1442 xu.xu_unpp = unp; 1443 1444 /* 1445 * NOTE: 1446 * unp->unp_addr and unp->unp_conn are protected by 1447 * unp_token. So if we want to get rid of unp_token 1448 * or reduce the coverage of unp_token, care must be 1449 * taken. 1450 */ 1451 if (unp->unp_addr) { 1452 bcopy(unp->unp_addr, &xu.xu_addr, 1453 unp->unp_addr->sun_len); 1454 } 1455 if (unp->unp_conn && unp->unp_conn->unp_addr) { 1456 bcopy(unp->unp_conn->unp_addr, 1457 &xu.xu_caddr, 1458 unp->unp_conn->unp_addr->sun_len); 1459 } 1460 bcopy(unp, &xu.xu_unp, sizeof(*unp)); 1461 sotoxsocket(unp->unp_socket, &xu.xu_socket); 1462 1463 /* NOTE: This could block and temporarily release unp_token */ 1464 error = SYSCTL_OUT(req, &xu, sizeof(xu)); 1465 if (error) 1466 break; 1467 ++i; 1468 } 1469 TAILQ_REMOVE(&head->list, marker, unp_link); 1470 1471 lwkt_reltoken(&unp_token); 1472 1473 kfree(marker, M_UNPCB); 1474 return error; 1475 } 1476 1477 SYSCTL_PROC(_net_local_dgram, OID_AUTO, pcblist, CTLFLAG_RD, 1478 &unp_dgram_head, 0, unp_pcblist, "S,xunpcb", 1479 "List of active local datagram sockets"); 1480 SYSCTL_PROC(_net_local_stream, OID_AUTO, pcblist, CTLFLAG_RD, 1481 &unp_stream_head, 0, unp_pcblist, "S,xunpcb", 1482 "List of active local stream sockets"); 1483 SYSCTL_PROC(_net_local_seqpacket, OID_AUTO, pcblist, CTLFLAG_RD, 1484 &unp_seqpkt_head, 0, unp_pcblist, "S,xunpcb", 1485 "List of active local seqpacket sockets"); 1486 1487 static void 1488 unp_shutdown(struct unpcb *unp) 1489 { 1490 struct socket *so; 1491 1492 if ((unp->unp_socket->so_type == SOCK_STREAM || 1493 unp->unp_socket->so_type == SOCK_SEQPACKET) && 1494 unp->unp_conn != NULL && (so = unp->unp_conn->unp_socket)) { 1495 socantrcvmore(so); 1496 } 1497 } 1498 1499 #ifdef notdef 1500 void 1501 unp_drain(void) 1502 { 1503 lwkt_gettoken(&unp_token); 1504 lwkt_reltoken(&unp_token); 1505 } 1506 #endif 1507 1508 int 1509 unp_externalize(struct mbuf *rights, int flags) 1510 { 1511 struct thread *td = curthread; 1512 struct proc *p = td->td_proc; /* XXX */ 1513 struct lwp *lp = td->td_lwp; 1514 struct cmsghdr *cm = mtod(rights, struct cmsghdr *); 1515 int *fdp; 1516 int i; 1517 struct file **rp; 1518 struct file *fp; 1519 int newfds = (cm->cmsg_len - (CMSG_DATA(cm) - (u_char *)cm)) 1520 / sizeof(struct file *); 1521 int f; 1522 1523 lwkt_gettoken(&unp_rights_token); 1524 1525 /* 1526 * if the new FD's will not fit, then we free them all 1527 */ 1528 if (!fdavail(p, newfds)) { 1529 rp = (struct file **)CMSG_DATA(cm); 1530 for (i = 0; i < newfds; i++) { 1531 fp = *rp; 1532 /* 1533 * zero the pointer before calling unp_discard, 1534 * since it may end up in unp_gc().. 1535 */ 1536 *rp++ = NULL; 1537 unp_discard(fp, NULL); 1538 } 1539 lwkt_reltoken(&unp_rights_token); 1540 return (EMSGSIZE); 1541 } 1542 1543 /* 1544 * now change each pointer to an fd in the global table to 1545 * an integer that is the index to the local fd table entry 1546 * that we set up to point to the global one we are transferring. 1547 * Since the sizeof(struct file *) is bigger than or equal to 1548 * the sizeof(int), we do it in forward order. In that case, 1549 * an integer will always come in the same place or before its 1550 * corresponding struct file pointer. 1551 * 1552 * Hold revoke_token in 'shared' mode, so that we won't miss 1553 * the FREVOKED update on fps being externalized (fsetfd). 1554 */ 1555 lwkt_gettoken_shared(&revoke_token); 1556 fdp = (int *)CMSG_DATA(cm); 1557 rp = (struct file **)CMSG_DATA(cm); 1558 for (i = 0; i < newfds; i++) { 1559 if (fdalloc(p, 0, &f)) { 1560 int j; 1561 1562 /* 1563 * Previous fdavail() can't garantee 1564 * fdalloc() success due to SMP race. 1565 * Just clean up and return the same 1566 * error value as if fdavail() failed. 1567 */ 1568 lwkt_reltoken(&revoke_token); 1569 1570 /* Close externalized files */ 1571 for (j = 0; j < i; j++) 1572 kern_close(fdp[j]); 1573 /* Discard the rest of internal files */ 1574 for (; i < newfds; i++) 1575 unp_discard(rp[i], NULL); 1576 /* Wipe out the control message */ 1577 for (i = 0; i < newfds; i++) 1578 rp[i] = NULL; 1579 1580 lwkt_reltoken(&unp_rights_token); 1581 return (EMSGSIZE); 1582 } 1583 fp = rp[i]; 1584 unp_fp_externalize(lp, fp, f, flags); 1585 fdp[i] = f; 1586 } 1587 lwkt_reltoken(&revoke_token); 1588 1589 lwkt_reltoken(&unp_rights_token); 1590 1591 /* 1592 * Adjust length, in case sizeof(struct file *) and sizeof(int) 1593 * differs. 1594 */ 1595 cm->cmsg_len = CMSG_LEN(newfds * sizeof(int)); 1596 rights->m_len = cm->cmsg_len; 1597 1598 return (0); 1599 } 1600 1601 static void 1602 unp_fp_externalize(struct lwp *lp, struct file *fp, int fd, int flags) 1603 { 1604 if (lp) { 1605 struct filedesc *fdp = lp->lwp_proc->p_fd; 1606 1607 KKASSERT(fd >= 0); 1608 if (fp->f_flag & FREVOKED) { 1609 struct file *fx; 1610 int error; 1611 1612 kprintf("Warning: revoked fp exiting unix socket\n"); 1613 error = falloc(lp, &fx, NULL); 1614 if (error == 0) { 1615 if (flags & MSG_CMSG_CLOEXEC) 1616 fdp->fd_files[fd].fileflags |= UF_EXCLOSE; 1617 fsetfd(fdp, fx, fd); 1618 fdrop(fx); 1619 } else { 1620 fsetfd(fdp, NULL, fd); 1621 } 1622 } else { 1623 if (flags & MSG_CMSG_CLOEXEC) 1624 fdp->fd_files[fd].fileflags |= UF_EXCLOSE; 1625 fsetfd(fdp, fp, fd); 1626 } 1627 } 1628 unp_del_right(fp); 1629 fdrop(fp); 1630 } 1631 1632 void 1633 unp_init(void) 1634 { 1635 TAILQ_INIT(&unp_stream_head.list); 1636 TAILQ_INIT(&unp_dgram_head.list); 1637 TAILQ_INIT(&unp_seqpkt_head.list); 1638 1639 SLIST_INIT(&unp_defdiscard_head); 1640 spin_init(&unp_defdiscard_spin, "unpdisc"); 1641 TASK_INIT(&unp_defdiscard_task, 0, unp_defdiscard_taskfunc, NULL); 1642 1643 /* 1644 * This implies that only one gc can be in-progress at any 1645 * given moment. 1646 */ 1647 TASK_INIT(&unp_gc_task, 0, unp_gc, NULL); 1648 1649 unp_gc_marker = kmalloc(sizeof(*unp_gc_marker), M_UNPCB, 1650 M_WAITOK | M_ZERO); 1651 unp_gc_marker->unp_flags |= UNP_MARKER; 1652 1653 /* 1654 * Create taskqueue for defered discard, and stick it to 1655 * the last CPU. 1656 */ 1657 unp_taskqueue = taskqueue_create("unp_taskq", M_WAITOK, 1658 taskqueue_thread_enqueue, &unp_taskqueue); 1659 taskqueue_start_threads(&unp_taskqueue, 1, TDPRI_KERN_DAEMON, 1660 ncpus - 1, "unp taskq"); 1661 } 1662 1663 static int 1664 unp_internalize(struct mbuf *control, struct thread *td) 1665 { 1666 struct proc *p = td->td_proc; 1667 struct filedesc *fdescp; 1668 struct cmsghdr *cm = mtod(control, struct cmsghdr *); 1669 struct file **rp; 1670 struct file *fp; 1671 int i, fd, *fdp; 1672 struct cmsgcred *cmcred; 1673 int oldfds; 1674 u_int newlen; 1675 int error; 1676 1677 KKASSERT(p); 1678 1679 if ((cm->cmsg_type != SCM_RIGHTS && cm->cmsg_type != SCM_CREDS) || 1680 cm->cmsg_level != SOL_SOCKET || 1681 CMSG_ALIGN(cm->cmsg_len) != control->m_len) 1682 return EINVAL; 1683 1684 /* 1685 * Fill in credential information. 1686 */ 1687 if (cm->cmsg_type == SCM_CREDS) { 1688 cmcred = (struct cmsgcred *)CMSG_DATA(cm); 1689 cmcred->cmcred_pid = p->p_pid; 1690 cmcred->cmcred_uid = p->p_ucred->cr_ruid; 1691 cmcred->cmcred_gid = p->p_ucred->cr_rgid; 1692 cmcred->cmcred_euid = p->p_ucred->cr_uid; 1693 cmcred->cmcred_ngroups = MIN(p->p_ucred->cr_ngroups, 1694 CMGROUP_MAX); 1695 for (i = 0; i < cmcred->cmcred_ngroups; i++) 1696 cmcred->cmcred_groups[i] = p->p_ucred->cr_groups[i]; 1697 return 0; 1698 } 1699 1700 /* 1701 * cmsghdr may not be aligned, do not allow calculation(s) to 1702 * go negative. 1703 */ 1704 if (cm->cmsg_len < CMSG_LEN(0)) 1705 return EINVAL; 1706 1707 oldfds = (cm->cmsg_len - CMSG_LEN(0)) / sizeof(int); 1708 1709 /* 1710 * Now replace the integer FDs with pointers to 1711 * the associated global file table entry.. 1712 * Allocate a bigger buffer as necessary. But if an cluster is not 1713 * enough, return E2BIG. 1714 */ 1715 newlen = CMSG_LEN(oldfds * sizeof(struct file *)); 1716 if (newlen > MCLBYTES) 1717 return E2BIG; 1718 if (newlen - control->m_len > M_TRAILINGSPACE(control)) { 1719 if (control->m_flags & M_EXT) 1720 return E2BIG; 1721 MCLGET(control, M_WAITOK); 1722 if (!(control->m_flags & M_EXT)) 1723 return ENOBUFS; 1724 1725 /* copy the data to the cluster */ 1726 memcpy(mtod(control, char *), cm, cm->cmsg_len); 1727 cm = mtod(control, struct cmsghdr *); 1728 } 1729 1730 lwkt_gettoken(&unp_rights_token); 1731 1732 fdescp = p->p_fd; 1733 spin_lock_shared(&fdescp->fd_spin); 1734 1735 /* 1736 * check that all the FDs passed in refer to legal OPEN files 1737 * If not, reject the entire operation. 1738 */ 1739 fdp = (int *)CMSG_DATA(cm); 1740 for (i = 0; i < oldfds; i++) { 1741 fd = *fdp++; 1742 if ((unsigned)fd >= fdescp->fd_nfiles || 1743 fdescp->fd_files[fd].fp == NULL) { 1744 error = EBADF; 1745 goto done; 1746 } 1747 if (fdescp->fd_files[fd].fp->f_type == DTYPE_KQUEUE) { 1748 error = EOPNOTSUPP; 1749 goto done; 1750 } 1751 } 1752 1753 /* 1754 * Adjust length, in case sizeof(struct file *) and sizeof(int) 1755 * differs. 1756 */ 1757 cm->cmsg_len = newlen; 1758 control->m_len = CMSG_ALIGN(newlen); 1759 1760 /* 1761 * Transform the file descriptors into struct file pointers. 1762 * Since the sizeof(struct file *) is bigger than or equal to 1763 * the sizeof(int), we do it in reverse order so that the int 1764 * won't get trashed until we're done. 1765 */ 1766 fdp = (int *)CMSG_DATA(cm) + oldfds - 1; 1767 rp = (struct file **)CMSG_DATA(cm) + oldfds - 1; 1768 for (i = 0; i < oldfds; i++) { 1769 fp = fdescp->fd_files[*fdp--].fp; 1770 *rp-- = fp; 1771 fhold(fp); 1772 unp_add_right(fp); 1773 } 1774 error = 0; 1775 done: 1776 spin_unlock_shared(&fdescp->fd_spin); 1777 lwkt_reltoken(&unp_rights_token); 1778 return error; 1779 } 1780 1781 #ifdef UNP_GC_ALLFILES 1782 1783 /* 1784 * Garbage collect in-transit file descriptors that get lost due to 1785 * loops (i.e. when a socket is sent to another process over itself, 1786 * and more complex situations). 1787 * 1788 * NOT MPSAFE - TODO socket flush code and maybe fdrop. Rest is MPSAFE. 1789 */ 1790 1791 struct unp_gc_info { 1792 struct file **extra_ref; 1793 struct file *locked_fp; 1794 int defer; 1795 int index; 1796 int maxindex; 1797 }; 1798 1799 static void 1800 unp_gc(void *arg __unused, int pending __unused) 1801 { 1802 struct unp_gc_info info; 1803 struct file **fpp; 1804 int i; 1805 1806 lwkt_gettoken(&unp_rights_token); 1807 1808 /* 1809 * Before going through all this, set all FDs to be NOT defered 1810 * and NOT externally accessible (not marked). During the scan 1811 * a fd can be marked externally accessible but we may or may not 1812 * be able to immediately process it (controlled by FDEFER). 1813 * 1814 * If we loop sleep a bit. The complexity of the topology can cause 1815 * multiple loops. Also failure to acquire the socket's so_rcv 1816 * token can cause us to loop. 1817 */ 1818 allfiles_scan_exclusive(unp_gc_clearmarks, NULL); 1819 do { 1820 info.defer = 0; 1821 allfiles_scan_exclusive(unp_gc_checkmarks, &info); 1822 if (info.defer) 1823 tsleep(&info, 0, "gcagain", 1); 1824 } while (info.defer); 1825 1826 /* 1827 * We grab an extra reference to each of the file table entries 1828 * that are not otherwise accessible and then free the rights 1829 * that are stored in messages on them. 1830 * 1831 * The bug in the orginal code is a little tricky, so I'll describe 1832 * what's wrong with it here. 1833 * 1834 * It is incorrect to simply unp_discard each entry for f_msgcount 1835 * times -- consider the case of sockets A and B that contain 1836 * references to each other. On a last close of some other socket, 1837 * we trigger a gc since the number of outstanding rights (unp_rights) 1838 * is non-zero. If during the sweep phase the gc code unp_discards, 1839 * we end up doing a (full) fdrop on the descriptor. A fdrop on A 1840 * results in the following chain. Closef calls soo_close, which 1841 * calls soclose. Soclose calls first (through the switch 1842 * uipc_usrreq) unp_detach, which re-invokes unp_gc. Unp_gc simply 1843 * returns because the previous instance had set unp_gcing, and 1844 * we return all the way back to soclose, which marks the socket 1845 * with SS_NOFDREF, and then calls sofree. Sofree calls sorflush 1846 * to free up the rights that are queued in messages on the socket A, 1847 * i.e., the reference on B. The sorflush calls via the dom_dispose 1848 * switch unp_dispose, which unp_scans with unp_discard. This second 1849 * instance of unp_discard just calls fdrop on B. 1850 * 1851 * Well, a similar chain occurs on B, resulting in a sorflush on B, 1852 * which results in another fdrop on A. Unfortunately, A is already 1853 * being closed, and the descriptor has already been marked with 1854 * SS_NOFDREF, and soclose panics at this point. 1855 * 1856 * Here, we first take an extra reference to each inaccessible 1857 * descriptor. Then, we call sorflush ourself, since we know 1858 * it is a Unix domain socket anyhow. After we destroy all the 1859 * rights carried in messages, we do a last fdrop to get rid 1860 * of our extra reference. This is the last close, and the 1861 * unp_detach etc will shut down the socket. 1862 * 1863 * 91/09/19, bsy@cs.cmu.edu 1864 */ 1865 info.extra_ref = kmalloc(256 * sizeof(struct file *), M_FILE, M_WAITOK); 1866 info.maxindex = 256; 1867 1868 do { 1869 /* 1870 * Look for matches 1871 */ 1872 info.index = 0; 1873 allfiles_scan_exclusive(unp_gc_checkrefs, &info); 1874 1875 /* 1876 * For each FD on our hit list, do the following two things 1877 */ 1878 for (i = info.index, fpp = info.extra_ref; --i >= 0; ++fpp) { 1879 struct file *tfp = *fpp; 1880 if (tfp->f_type == DTYPE_SOCKET && tfp->f_data != NULL) 1881 sorflush((struct socket *)(tfp->f_data)); 1882 } 1883 for (i = info.index, fpp = info.extra_ref; --i >= 0; ++fpp) 1884 fdrop(*fpp); 1885 } while (info.index == info.maxindex); 1886 1887 kfree((caddr_t)info.extra_ref, M_FILE); 1888 1889 lwkt_reltoken(&unp_rights_token); 1890 } 1891 1892 /* 1893 * MPSAFE - NOTE: filehead list and file pointer spinlocked on entry 1894 */ 1895 static int 1896 unp_gc_checkrefs(struct file *fp, void *data) 1897 { 1898 struct unp_gc_info *info = data; 1899 1900 if (fp->f_count == 0) 1901 return(0); 1902 if (info->index == info->maxindex) 1903 return(-1); 1904 1905 /* 1906 * If all refs are from msgs, and it's not marked accessible 1907 * then it must be referenced from some unreachable cycle 1908 * of (shut-down) FDs, so include it in our 1909 * list of FDs to remove 1910 */ 1911 if (fp->f_count == fp->f_msgcount && !(fp->f_flag & FMARK)) { 1912 info->extra_ref[info->index++] = fp; 1913 fhold(fp); 1914 } 1915 return(0); 1916 } 1917 1918 /* 1919 * MPSAFE - NOTE: filehead list and file pointer spinlocked on entry 1920 */ 1921 static int 1922 unp_gc_clearmarks(struct file *fp, void *data __unused) 1923 { 1924 atomic_clear_int(&fp->f_flag, FMARK | FDEFER); 1925 return(0); 1926 } 1927 1928 /* 1929 * MPSAFE - NOTE: filehead list and file pointer spinlocked on entry 1930 */ 1931 static int 1932 unp_gc_checkmarks(struct file *fp, void *data) 1933 { 1934 struct unp_gc_info *info = data; 1935 struct socket *so; 1936 1937 /* 1938 * If the file is not open, skip it. Make sure it isn't marked 1939 * defered or we could loop forever, in case we somehow race 1940 * something. 1941 */ 1942 if (fp->f_count == 0) { 1943 if (fp->f_flag & FDEFER) 1944 atomic_clear_int(&fp->f_flag, FDEFER); 1945 return(0); 1946 } 1947 /* 1948 * If we already marked it as 'defer' in a 1949 * previous pass, then try process it this time 1950 * and un-mark it 1951 */ 1952 if (fp->f_flag & FDEFER) { 1953 atomic_clear_int(&fp->f_flag, FDEFER); 1954 } else { 1955 /* 1956 * if it's not defered, then check if it's 1957 * already marked.. if so skip it 1958 */ 1959 if (fp->f_flag & FMARK) 1960 return(0); 1961 /* 1962 * If all references are from messages 1963 * in transit, then skip it. it's not 1964 * externally accessible. 1965 */ 1966 if (fp->f_count == fp->f_msgcount) 1967 return(0); 1968 /* 1969 * If it got this far then it must be 1970 * externally accessible. 1971 */ 1972 atomic_set_int(&fp->f_flag, FMARK); 1973 } 1974 1975 /* 1976 * either it was defered, or it is externally 1977 * accessible and not already marked so. 1978 * Now check if it is possibly one of OUR sockets. 1979 */ 1980 if (fp->f_type != DTYPE_SOCKET || 1981 (so = (struct socket *)fp->f_data) == NULL) { 1982 return(0); 1983 } 1984 if (so->so_proto->pr_domain != &localdomain || 1985 !(so->so_proto->pr_flags & PR_RIGHTS)) { 1986 return(0); 1987 } 1988 1989 /* 1990 * So, Ok, it's one of our sockets and it IS externally accessible 1991 * (or was defered). Now we look to see if we hold any file 1992 * descriptors in its message buffers. Follow those links and mark 1993 * them as accessible too. 1994 * 1995 * We are holding multiple spinlocks here, if we cannot get the 1996 * token non-blocking defer until the next loop. 1997 */ 1998 info->locked_fp = fp; 1999 if (lwkt_trytoken(&so->so_rcv.ssb_token)) { 2000 unp_scan(so->so_rcv.ssb_mb, unp_mark, info); 2001 lwkt_reltoken(&so->so_rcv.ssb_token); 2002 } else { 2003 atomic_set_int(&fp->f_flag, FDEFER); 2004 ++info->defer; 2005 } 2006 return (0); 2007 } 2008 2009 /* 2010 * Mark visibility. info->defer is recalculated on every pass. 2011 */ 2012 static void 2013 unp_mark(struct file *fp, void *data) 2014 { 2015 struct unp_gc_info *info = data; 2016 2017 if ((fp->f_flag & FMARK) == 0) { 2018 ++info->defer; 2019 atomic_set_int(&fp->f_flag, FMARK | FDEFER); 2020 } else if (fp->f_flag & FDEFER) { 2021 ++info->defer; 2022 } 2023 } 2024 2025 #else /* !UNP_GC_ALLFILES */ 2026 2027 /* 2028 * They are thread local and do not require explicit synchronization. 2029 */ 2030 static int unp_marked; 2031 static int unp_unreachable; 2032 2033 static void 2034 unp_accessable(struct file *fp, void *data __unused) 2035 { 2036 struct unpcb *unp; 2037 2038 if ((unp = unp_fp2unpcb(fp)) == NULL) 2039 return; 2040 if (unp->unp_gcflags & UNPGC_REF) 2041 return; 2042 unp->unp_gcflags &= ~UNPGC_DEAD; 2043 unp->unp_gcflags |= UNPGC_REF; 2044 unp_marked++; 2045 } 2046 2047 static void 2048 unp_gc_process(struct unpcb *unp) 2049 { 2050 struct file *fp; 2051 2052 /* Already processed. */ 2053 if (unp->unp_gcflags & UNPGC_SCANNED) 2054 return; 2055 fp = unp->unp_fp; 2056 2057 /* 2058 * Check for a socket potentially in a cycle. It must be in a 2059 * queue as indicated by msgcount, and this must equal the file 2060 * reference count. Note that when msgcount is 0 the file is NULL. 2061 */ 2062 if ((unp->unp_gcflags & UNPGC_REF) == 0 && fp && 2063 unp->unp_msgcount != 0 && fp->f_count == unp->unp_msgcount) { 2064 unp->unp_gcflags |= UNPGC_DEAD; 2065 unp_unreachable++; 2066 return; 2067 } 2068 2069 /* 2070 * Mark all sockets we reference with RIGHTS. 2071 */ 2072 if (UNP_ISATTACHED(unp)) { 2073 struct signalsockbuf *ssb = &unp->unp_socket->so_rcv; 2074 2075 unp_reference(unp); 2076 lwkt_gettoken(&ssb->ssb_token); 2077 /* 2078 * unp_token would be temporarily dropped, if getting 2079 * so_rcv token blocks, so we need to check unp state 2080 * here again. 2081 */ 2082 if (UNP_ISATTACHED(unp)) 2083 unp_scan(ssb->ssb_mb, unp_accessable, NULL); 2084 lwkt_reltoken(&ssb->ssb_token); 2085 unp->unp_gcflags |= UNPGC_SCANNED; 2086 unp_free(unp); 2087 } else { 2088 unp->unp_gcflags |= UNPGC_SCANNED; 2089 } 2090 } 2091 2092 static void 2093 unp_gc(void *arg __unused, int pending __unused) 2094 { 2095 struct unp_global_head *head; 2096 int h, filemax, fileidx, filetot; 2097 struct file **unref; 2098 struct unpcb *unp; 2099 2100 lwkt_gettoken(&unp_rights_token); 2101 lwkt_gettoken(&unp_token); 2102 2103 /* 2104 * First clear all gc flags from previous runs. 2105 */ 2106 for (h = 0; unp_heads[h] != NULL; ++h) { 2107 /* 2108 * NOTE: This loop does not block, so it is safe 2109 * to use TAILQ_FOREACH here. 2110 */ 2111 head = unp_heads[h]; 2112 TAILQ_FOREACH(unp, &head->list, unp_link) 2113 unp->unp_gcflags = 0; 2114 } 2115 2116 /* 2117 * Scan marking all reachable sockets with UNPGC_REF. Once a socket 2118 * is reachable all of the sockets it references are reachable. 2119 * Stop the scan once we do a complete loop without discovering 2120 * a new reachable socket. 2121 */ 2122 do { 2123 unp_unreachable = 0; 2124 unp_marked = 0; 2125 for (h = 0; unp_heads[h] != NULL; ++h) { 2126 head = unp_heads[h]; 2127 TAILQ_INSERT_HEAD(&head->list, unp_gc_marker, unp_link); 2128 while ((unp = TAILQ_NEXT(unp_gc_marker, unp_link)) 2129 != NULL) { 2130 TAILQ_REMOVE(&head->list, unp_gc_marker, 2131 unp_link); 2132 TAILQ_INSERT_AFTER(&head->list, unp, 2133 unp_gc_marker, unp_link); 2134 2135 if (unp->unp_flags & UNP_MARKER) 2136 continue; 2137 unp_gc_process(unp); 2138 } 2139 TAILQ_REMOVE(&head->list, unp_gc_marker, unp_link); 2140 } 2141 } while (unp_marked); 2142 2143 if (unp_unreachable == 0) 2144 goto done; 2145 2146 /* 2147 * We grab an extra reference to each of the file table entries 2148 * that are not otherwise accessible and then free the rights 2149 * that are stored in messages on them. 2150 * 2151 * The bug in the orginal code is a little tricky, so I'll describe 2152 * what's wrong with it here. 2153 * 2154 * It is incorrect to simply unp_discard each entry for f_msgcount 2155 * times -- consider the case of sockets A and B that contain 2156 * references to each other. On a last close of some other socket, 2157 * we trigger a gc since the number of outstanding rights (unp_rights) 2158 * is non-zero. If during the sweep phase the gc code unp_discards, 2159 * we end up doing a (full) fdrop on the descriptor. A fdrop on A 2160 * results in the following chain. Closef calls soo_close, which 2161 * calls soclose. Soclose calls first (through the switch 2162 * uipc_usrreq) unp_detach, which re-invokes unp_gc. Unp_gc simply 2163 * returns because the previous instance had set unp_gcing, and 2164 * we return all the way back to soclose, which marks the socket 2165 * with SS_NOFDREF, and then calls sofree. Sofree calls sorflush 2166 * to free up the rights that are queued in messages on the socket A, 2167 * i.e., the reference on B. The sorflush calls via the dom_dispose 2168 * switch unp_dispose, which unp_scans with unp_discard. This second 2169 * instance of unp_discard just calls fdrop on B. 2170 * 2171 * Well, a similar chain occurs on B, resulting in a sorflush on B, 2172 * which results in another fdrop on A. Unfortunately, A is already 2173 * being closed, and the descriptor has already been marked with 2174 * SS_NOFDREF, and soclose panics at this point. 2175 * 2176 * Here, we first take an extra reference to each inaccessible 2177 * descriptor. Then, we call sorflush ourself, since we know 2178 * it is a Unix domain socket anyhow. After we destroy all the 2179 * rights carried in messages, we do a last fdrop to get rid 2180 * of our extra reference. This is the last close, and the 2181 * unp_detach etc will shut down the socket. 2182 * 2183 * 91/09/19, bsy@cs.cmu.edu 2184 */ 2185 2186 filemax = unp_unreachable; 2187 if (filemax > UNP_GCFILE_MAX) 2188 filemax = UNP_GCFILE_MAX; 2189 unref = kmalloc(filemax * sizeof(struct file *), M_TEMP, M_WAITOK); 2190 2191 filetot = 0; 2192 do { 2193 int i; 2194 2195 /* 2196 * Iterate looking for sockets which have been specifically 2197 * marked as as unreachable and store them locally. 2198 */ 2199 fileidx = 0; 2200 for (h = 0; unp_heads[h] != NULL; ++h) { 2201 /* 2202 * NOTE: This loop does not block, so it is safe 2203 * to use TAILQ_FOREACH here. 2204 */ 2205 head = unp_heads[h]; 2206 TAILQ_FOREACH(unp, &head->list, unp_link) { 2207 struct file *fp; 2208 2209 if ((unp->unp_gcflags & UNPGC_DEAD) == 0) 2210 continue; 2211 unp->unp_gcflags &= ~UNPGC_DEAD; 2212 2213 fp = unp->unp_fp; 2214 if (unp->unp_msgcount == 0 || fp == NULL || 2215 fp->f_count != unp->unp_msgcount) 2216 continue; 2217 fhold(fp); 2218 2219 KASSERT(fileidx < filemax, 2220 ("invalid fileidx %d, filemax %d", 2221 fileidx, filemax)); 2222 unref[fileidx++] = fp; 2223 2224 KASSERT(filetot < unp_unreachable, 2225 ("invalid filetot %d and " 2226 "unp_unreachable %d", 2227 filetot, unp_unreachable)); 2228 ++filetot; 2229 2230 if (fileidx == filemax || 2231 filetot == unp_unreachable) 2232 goto dogc; 2233 } 2234 } 2235 dogc: 2236 /* 2237 * For each Unix domain socket on our hit list, do the 2238 * following two things. 2239 */ 2240 for (i = 0; i < fileidx; ++i) 2241 sorflush(unref[i]->f_data); 2242 for (i = 0; i < fileidx; ++i) 2243 fdrop(unref[i]); 2244 } while (fileidx == filemax && filetot < unp_unreachable); 2245 kfree(unref, M_TEMP); 2246 done: 2247 lwkt_reltoken(&unp_token); 2248 lwkt_reltoken(&unp_rights_token); 2249 } 2250 2251 #endif /* UNP_GC_ALLFILES */ 2252 2253 /* 2254 * Dispose of the fp's stored in a mbuf. 2255 * 2256 * The dds loop can cause additional fps to be entered onto the 2257 * list while it is running, flattening out the operation and avoiding 2258 * a deep kernel stack recursion. 2259 */ 2260 void 2261 unp_dispose(struct mbuf *m) 2262 { 2263 lwkt_gettoken(&unp_rights_token); 2264 if (m) 2265 unp_scan(m, unp_discard, NULL); 2266 lwkt_reltoken(&unp_rights_token); 2267 } 2268 2269 static int 2270 unp_listen(struct unpcb *unp, struct thread *td) 2271 { 2272 struct proc *p = td->td_proc; 2273 2274 ASSERT_LWKT_TOKEN_HELD(&unp_token); 2275 UNP_ASSERT_TOKEN_HELD(unp); 2276 2277 KKASSERT(p); 2278 cru2x(p->p_ucred, &unp->unp_peercred); 2279 unp_setflags(unp, UNP_HAVEPCCACHED); 2280 return (0); 2281 } 2282 2283 static void 2284 unp_scan(struct mbuf *m0, void (*op)(struct file *, void *), void *data) 2285 { 2286 struct mbuf *m; 2287 struct file **rp; 2288 struct cmsghdr *cm; 2289 int i; 2290 int qfds; 2291 2292 while (m0) { 2293 for (m = m0; m; m = m->m_next) { 2294 if (m->m_type == MT_CONTROL && 2295 m->m_len >= sizeof(*cm)) { 2296 cm = mtod(m, struct cmsghdr *); 2297 if (cm->cmsg_level != SOL_SOCKET || 2298 cm->cmsg_type != SCM_RIGHTS) 2299 continue; 2300 qfds = (cm->cmsg_len - CMSG_LEN(0)) / 2301 sizeof(void *); 2302 rp = (struct file **)CMSG_DATA(cm); 2303 for (i = 0; i < qfds; i++) 2304 (*op)(*rp++, data); 2305 break; /* XXX, but saves time */ 2306 } 2307 } 2308 m0 = m0->m_nextpkt; 2309 } 2310 } 2311 2312 /* 2313 * Discard a fp previously held in a unix domain socket mbuf. To 2314 * avoid blowing out the kernel stack due to contrived chain-reactions 2315 * we may have to defer the operation to a dedicated taskqueue. 2316 * 2317 * Caller holds unp_rights_token. 2318 */ 2319 static void 2320 unp_discard(struct file *fp, void *data __unused) 2321 { 2322 unp_del_right(fp); 2323 if (unp_fp2unpcb(fp) != NULL) { 2324 struct unp_defdiscard *d; 2325 2326 /* 2327 * This fp is a Unix domain socket itself and fdrop() 2328 * it here directly may cause deep unp_discard() 2329 * recursion, so the fdrop() is defered to the 2330 * dedicated taskqueue. 2331 */ 2332 d = kmalloc(sizeof(*d), M_UNPCB, M_WAITOK); 2333 d->fp = fp; 2334 2335 spin_lock(&unp_defdiscard_spin); 2336 SLIST_INSERT_HEAD(&unp_defdiscard_head, d, next); 2337 spin_unlock(&unp_defdiscard_spin); 2338 2339 taskqueue_enqueue(unp_taskqueue, &unp_defdiscard_task); 2340 } else { 2341 /* This fp is not a Unix domain socket */ 2342 fdrop(fp); 2343 } 2344 } 2345 2346 /* 2347 * NOTE: 2348 * unp_token must be held before calling this function to avoid name 2349 * resolution and v_socket accessing races, especially racing against 2350 * the unp_detach(). 2351 * 2352 * NOTE: 2353 * For anyone caring about unconnected Unix domain socket sending 2354 * performance, other approach could be taken... 2355 */ 2356 static int 2357 unp_find_lockref(struct sockaddr *nam, struct thread *td, short type, 2358 struct unpcb **unp_ret) 2359 { 2360 struct proc *p = td->td_proc; 2361 struct sockaddr_un *soun = (struct sockaddr_un *)nam; 2362 struct vnode *vp = NULL; 2363 struct socket *so; 2364 struct unpcb *unp; 2365 int error, len; 2366 struct nlookupdata nd; 2367 char buf[SOCK_MAXADDRLEN]; 2368 2369 ASSERT_LWKT_TOKEN_HELD(&unp_token); 2370 2371 *unp_ret = NULL; 2372 2373 len = nam->sa_len - offsetof(struct sockaddr_un, sun_path); 2374 if (len <= 0) { 2375 error = EINVAL; 2376 goto failed; 2377 } 2378 strncpy(buf, soun->sun_path, len); 2379 buf[len] = 0; 2380 2381 error = nlookup_init(&nd, buf, UIO_SYSSPACE, NLC_FOLLOW); 2382 if (error == 0) 2383 error = nlookup(&nd); 2384 if (error == 0) 2385 error = cache_vget(&nd.nl_nch, nd.nl_cred, LK_EXCLUSIVE, &vp); 2386 nlookup_done(&nd); 2387 if (error) { 2388 vp = NULL; 2389 goto failed; 2390 } 2391 2392 if (vp->v_type != VSOCK) { 2393 error = ENOTSOCK; 2394 goto failed; 2395 } 2396 error = VOP_EACCESS(vp, VWRITE, p->p_ucred); 2397 if (error) 2398 goto failed; 2399 so = vp->v_socket; 2400 if (so == NULL) { 2401 error = ECONNREFUSED; 2402 goto failed; 2403 } 2404 if (so->so_type != type) { 2405 error = EPROTOTYPE; 2406 goto failed; 2407 } 2408 2409 /* Lock this unp. */ 2410 unp = unp_getsocktoken(so); 2411 if (!UNP_ISATTACHED(unp)) { 2412 unp_reltoken(unp); 2413 error = ECONNREFUSED; 2414 goto failed; 2415 } 2416 /* And keep this unp referenced. */ 2417 unp_reference(unp); 2418 2419 /* Done! */ 2420 *unp_ret = unp; 2421 error = 0; 2422 failed: 2423 if (vp != NULL) 2424 vput(vp); 2425 return error; 2426 } 2427 2428 static int 2429 unp_connect_pair(struct unpcb *unp, struct unpcb *unp2) 2430 { 2431 struct socket *so = unp->unp_socket; 2432 struct socket *so2 = unp2->unp_socket; 2433 2434 ASSERT_LWKT_TOKEN_HELD(&unp_token); 2435 UNP_ASSERT_TOKEN_HELD(unp); 2436 UNP_ASSERT_TOKEN_HELD(unp2); 2437 2438 KASSERT(so->so_type == so2->so_type, 2439 ("socket type mismatch, so %d, so2 %d", so->so_type, so2->so_type)); 2440 2441 if (!UNP_ISATTACHED(unp)) 2442 return EINVAL; 2443 if (!UNP_ISATTACHED(unp2)) 2444 return ECONNREFUSED; 2445 2446 KASSERT(unp->unp_conn == NULL, ("unp is already connected")); 2447 unp->unp_conn = unp2; 2448 2449 switch (so->so_type) { 2450 case SOCK_DGRAM: 2451 LIST_INSERT_HEAD(&unp2->unp_refs, unp, unp_reflink); 2452 soisconnected(so); 2453 break; 2454 2455 case SOCK_STREAM: 2456 case SOCK_SEQPACKET: 2457 KASSERT(unp2->unp_conn == NULL, ("unp2 is already connected")); 2458 unp2->unp_conn = unp; 2459 soisconnected(so); 2460 soisconnected(so2); 2461 break; 2462 2463 default: 2464 panic("unp_connect_pair: unknown socket type %d", so->so_type); 2465 } 2466 return 0; 2467 } 2468 2469 static void 2470 unp_drop(struct unpcb *unp, int error) 2471 { 2472 struct unp_global_head *head; 2473 struct unpcb *unp2; 2474 2475 ASSERT_LWKT_TOKEN_HELD(&unp_token); 2476 UNP_ASSERT_TOKEN_HELD(unp); 2477 2478 KASSERT((unp->unp_flags & (UNP_DETACHED | UNP_DROPPED)) == 0, 2479 ("unp is dropped")); 2480 2481 /* Mark this unp as detached. */ 2482 unp_setflags(unp, UNP_DETACHED); 2483 2484 /* Remove this unp from the global unp list. */ 2485 head = unp_globalhead(unp->unp_socket->so_type); 2486 KASSERT(head->count > 0, ("invalid unp count")); 2487 TAILQ_REMOVE(&head->list, unp, unp_link); 2488 head->count--; 2489 2490 /* Disconnect all. */ 2491 unp_disconnect(unp, error); 2492 while ((unp2 = LIST_FIRST(&unp->unp_refs)) != NULL) { 2493 lwkt_getpooltoken(unp2); 2494 unp_disconnect(unp2, ECONNRESET); 2495 lwkt_relpooltoken(unp2); 2496 } 2497 unp_setflags(unp, UNP_DROPPED); 2498 2499 /* Try freeing this unp. */ 2500 unp_free(unp); 2501 } 2502 2503 static void 2504 unp_defdiscard_taskfunc(void *arg __unused, int pending __unused) 2505 { 2506 struct unp_defdiscard *d; 2507 2508 spin_lock(&unp_defdiscard_spin); 2509 while ((d = SLIST_FIRST(&unp_defdiscard_head)) != NULL) { 2510 SLIST_REMOVE_HEAD(&unp_defdiscard_head, next); 2511 spin_unlock(&unp_defdiscard_spin); 2512 2513 fdrop(d->fp); 2514 kfree(d, M_UNPCB); 2515 2516 spin_lock(&unp_defdiscard_spin); 2517 } 2518 spin_unlock(&unp_defdiscard_spin); 2519 } 2520