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 static ino_t unp_ino = 1; /* prototype for fake inode numbers */ 148 149 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 *, 153 struct thread *); 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); 162 #endif 163 static void unp_scan (struct mbuf *, void (*)(struct file *, void *), 164 void *data); 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, 169 int flags); 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); 175 176 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; 181 182 SYSCTL_DECL(_net_local); 183 SYSCTL_INT(_net_local, OID_AUTO, inflight, CTLFLAG_RD, &unp_rights, 0, 184 "File descriptors in flight"); 185 186 /* 187 * SMP Considerations: 188 * 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. 192 * 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. 195 * 196 * Any access to so_pcb to obtain unp requires the pool token for 197 * unp to be held. 198 */ 199 200 static __inline void 201 unp_reference(struct unpcb *unp) 202 { 203 /* 0->1 transition will not work */ 204 KKASSERT(unp->unp_refcnt > 0); 205 atomic_add_int(&unp->unp_refcnt, 1); 206 } 207 208 static __inline void 209 unp_free(struct unpcb *unp) 210 { 211 KKASSERT(unp->unp_refcnt > 0); 212 if (atomic_fetchadd_int(&unp->unp_refcnt, -1) == 1) 213 unp_detach(unp); 214 } 215 216 static __inline struct unpcb * 217 unp_getsocktoken(struct socket *so) 218 { 219 struct unpcb *unp; 220 221 /* 222 * The unp pointer is invalid until we verify that it is 223 * good by re-checking so_pcb AFTER obtaining the token. 224 */ 225 while ((unp = so->so_pcb) != NULL) { 226 lwkt_getpooltoken(unp); 227 if (unp == so->so_pcb) 228 break; 229 lwkt_relpooltoken(unp); 230 } 231 return unp; 232 } 233 234 static __inline void 235 unp_reltoken(struct unpcb *unp) 236 { 237 if (unp != NULL) 238 lwkt_relpooltoken(unp); 239 } 240 241 static __inline void 242 unp_setflags(struct unpcb *unp, int flags) 243 { 244 atomic_set_int(&unp->unp_flags, flags); 245 } 246 247 static __inline void 248 unp_clrflags(struct unpcb *unp, int flags) 249 { 250 atomic_clear_int(&unp->unp_flags, flags); 251 } 252 253 static __inline struct unp_global_head * 254 unp_globalhead(short type) 255 { 256 switch (type) { 257 case SOCK_STREAM: 258 return &unp_stream_head; 259 case SOCK_DGRAM: 260 return &unp_dgram_head; 261 case SOCK_SEQPACKET: 262 return &unp_seqpkt_head; 263 default: 264 panic("unknown socket type %d", type); 265 } 266 } 267 268 static __inline struct unpcb * 269 unp_fp2unpcb(struct file *fp) 270 { 271 struct socket *so; 272 273 if (fp->f_type != DTYPE_SOCKET) 274 return NULL; 275 276 so = fp->f_data; 277 if (so == NULL) 278 return NULL; 279 280 if (so->so_proto->pr_domain != &localdomain) 281 return NULL; 282 283 return so->so_pcb; 284 } 285 286 static __inline void 287 unp_add_right(struct file *fp) 288 { 289 struct unpcb *unp; 290 291 ASSERT_LWKT_TOKEN_HELD(&unp_rights_token); 292 KASSERT(fp->f_count > 0, ("invalid f_count %d", fp->f_count)); 293 294 unp = unp_fp2unpcb(fp); 295 if (unp != NULL) { 296 unp->unp_fp = fp; 297 unp->unp_msgcount++; 298 } 299 fp->f_msgcount++; 300 unp_rights++; 301 } 302 303 static __inline void 304 unp_del_right(struct file *fp) 305 { 306 struct unpcb *unp; 307 308 ASSERT_LWKT_TOKEN_HELD(&unp_rights_token); 309 KASSERT(fp->f_count > 0, ("invalid f_count %d", fp->f_count)); 310 311 unp = unp_fp2unpcb(fp); 312 if (unp != NULL) { 313 KASSERT(unp->unp_msgcount > 0, 314 ("invalid unp msgcount %d", unp->unp_msgcount)); 315 unp->unp_msgcount--; 316 if (unp->unp_msgcount == 0) 317 unp->unp_fp = NULL; 318 } 319 fp->f_msgcount--; 320 unp_rights--; 321 } 322 323 /* 324 * NOTE: (so) is referenced from soabort*() and netmsg_pru_abort() 325 * will sofree() it when we return. 326 */ 327 static void 328 uipc_abort(netmsg_t msg) 329 { 330 struct unpcb *unp; 331 int error; 332 333 lwkt_gettoken(&unp_token); 334 unp = unp_getsocktoken(msg->base.nm_so); 335 336 if (UNP_ISATTACHED(unp)) { 337 unp_drop(unp, ECONNABORTED); 338 error = 0; 339 } else { 340 error = EINVAL; 341 } 342 343 unp_reltoken(unp); 344 lwkt_reltoken(&unp_token); 345 346 lwkt_replymsg(&msg->lmsg, error); 347 } 348 349 static void 350 uipc_accept(netmsg_t msg) 351 { 352 struct unpcb *unp; 353 int error; 354 355 lwkt_gettoken(&unp_token); 356 unp = unp_getsocktoken(msg->base.nm_so); 357 358 if (!UNP_ISATTACHED(unp)) { 359 error = EINVAL; 360 } else { 361 struct unpcb *unp2 = unp->unp_conn; 362 363 /* 364 * Pass back name of connected socket, 365 * if it was bound and we are still connected 366 * (our peer may have closed already!). 367 */ 368 if (unp2 && unp2->unp_addr) { 369 unp_reference(unp2); 370 *msg->accept.nm_nam = dup_sockaddr( 371 (struct sockaddr *)unp2->unp_addr); 372 unp_free(unp2); 373 } else { 374 *msg->accept.nm_nam = dup_sockaddr(&sun_noname); 375 } 376 error = 0; 377 } 378 379 unp_reltoken(unp); 380 lwkt_reltoken(&unp_token); 381 382 lwkt_replymsg(&msg->lmsg, error); 383 } 384 385 static void 386 uipc_attach(netmsg_t msg) 387 { 388 int error; 389 390 lwkt_gettoken(&unp_token); 391 392 KASSERT(msg->base.nm_so->so_pcb == NULL, ("double unp attach")); 393 error = unp_attach(msg->base.nm_so, msg->attach.nm_ai); 394 395 lwkt_reltoken(&unp_token); 396 lwkt_replymsg(&msg->lmsg, error); 397 } 398 399 static void 400 uipc_bind(netmsg_t msg) 401 { 402 struct unpcb *unp; 403 int error; 404 405 lwkt_gettoken(&unp_token); 406 unp = unp_getsocktoken(msg->base.nm_so); 407 408 if (UNP_ISATTACHED(unp)) 409 error = unp_bind(unp, msg->bind.nm_nam, msg->bind.nm_td); 410 else 411 error = EINVAL; 412 413 unp_reltoken(unp); 414 lwkt_reltoken(&unp_token); 415 416 lwkt_replymsg(&msg->lmsg, error); 417 } 418 419 static void 420 uipc_connect(netmsg_t msg) 421 { 422 int error; 423 424 error = unp_connect(msg->base.nm_so, msg->connect.nm_nam, 425 msg->connect.nm_td); 426 lwkt_replymsg(&msg->lmsg, error); 427 } 428 429 static void 430 uipc_connect2(netmsg_t msg) 431 { 432 int error; 433 434 error = unp_connect2(msg->connect2.nm_so1, msg->connect2.nm_so2); 435 lwkt_replymsg(&msg->lmsg, error); 436 } 437 438 /* control is EOPNOTSUPP */ 439 440 static void 441 uipc_detach(netmsg_t msg) 442 { 443 struct unpcb *unp; 444 int error; 445 446 lwkt_gettoken(&unp_token); 447 unp = unp_getsocktoken(msg->base.nm_so); 448 449 if (UNP_ISATTACHED(unp)) { 450 unp_drop(unp, 0); 451 error = 0; 452 } else { 453 error = EINVAL; 454 } 455 456 unp_reltoken(unp); 457 lwkt_reltoken(&unp_token); 458 459 lwkt_replymsg(&msg->lmsg, error); 460 } 461 462 static void 463 uipc_disconnect(netmsg_t msg) 464 { 465 struct unpcb *unp; 466 int error; 467 468 lwkt_gettoken(&unp_token); 469 unp = unp_getsocktoken(msg->base.nm_so); 470 471 if (UNP_ISATTACHED(unp)) { 472 unp_disconnect(unp, 0); 473 error = 0; 474 } else { 475 error = EINVAL; 476 } 477 478 unp_reltoken(unp); 479 lwkt_reltoken(&unp_token); 480 481 lwkt_replymsg(&msg->lmsg, error); 482 } 483 484 static void 485 uipc_listen(netmsg_t msg) 486 { 487 struct unpcb *unp; 488 int error; 489 490 lwkt_gettoken(&unp_token); 491 unp = unp_getsocktoken(msg->base.nm_so); 492 493 if (!UNP_ISATTACHED(unp) || unp->unp_vnode == NULL) 494 error = EINVAL; 495 else 496 error = unp_listen(unp, msg->listen.nm_td); 497 498 unp_reltoken(unp); 499 lwkt_reltoken(&unp_token); 500 501 lwkt_replymsg(&msg->lmsg, error); 502 } 503 504 static void 505 uipc_peeraddr(netmsg_t msg) 506 { 507 struct unpcb *unp; 508 int error; 509 510 lwkt_gettoken(&unp_token); 511 unp = unp_getsocktoken(msg->base.nm_so); 512 513 if (!UNP_ISATTACHED(unp)) { 514 error = EINVAL; 515 } else if (unp->unp_conn && unp->unp_conn->unp_addr) { 516 struct unpcb *unp2 = unp->unp_conn; 517 518 unp_reference(unp2); 519 *msg->peeraddr.nm_nam = dup_sockaddr( 520 (struct sockaddr *)unp2->unp_addr); 521 unp_free(unp2); 522 error = 0; 523 } else { 524 /* 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. 528 */ 529 *msg->peeraddr.nm_nam = dup_sockaddr(&sun_noname); 530 error = 0; 531 } 532 533 unp_reltoken(unp); 534 lwkt_reltoken(&unp_token); 535 536 lwkt_replymsg(&msg->lmsg, error); 537 } 538 539 static void 540 uipc_rcvd(netmsg_t msg) 541 { 542 struct unpcb *unp, *unp2; 543 struct socket *so; 544 struct socket *so2; 545 int error; 546 547 /* 548 * so_pcb is only modified with both the global and the unp 549 * pool token held. 550 */ 551 so = msg->base.nm_so; 552 unp = unp_getsocktoken(so); 553 554 if (!UNP_ISATTACHED(unp)) { 555 error = EINVAL; 556 goto done; 557 } 558 559 switch (so->so_type) { 560 case SOCK_DGRAM: 561 panic("uipc_rcvd DGRAM?"); 562 /*NOTREACHED*/ 563 case SOCK_STREAM: 564 case SOCK_SEQPACKET: 565 if (unp->unp_conn == NULL) 566 break; 567 unp2 = unp->unp_conn; /* protected by pool token */ 568 569 /* 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. 573 * 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(). 577 */ 578 so2 = unp2->unp_socket; 579 unp_reference(unp2); 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 583 ) { 584 atomic_clear_int(&so2->so_snd.ssb_flags, SSB_STOP); 585 586 sowwakeup(so2); 587 } 588 lwkt_reltoken(&so2->so_rcv.ssb_token); 589 unp_free(unp2); 590 break; 591 default: 592 panic("uipc_rcvd unknown socktype"); 593 /*NOTREACHED*/ 594 } 595 error = 0; 596 done: 597 unp_reltoken(unp); 598 lwkt_replymsg(&msg->lmsg, error); 599 } 600 601 /* pru_rcvoob is EOPNOTSUPP */ 602 603 static void 604 uipc_send(netmsg_t msg) 605 { 606 struct unpcb *unp, *unp2; 607 struct socket *so; 608 struct socket *so2; 609 struct mbuf *control; 610 struct mbuf *m; 611 int error = 0; 612 613 so = msg->base.nm_so; 614 control = msg->send.nm_control; 615 m = msg->send.nm_m; 616 617 /* 618 * so_pcb is only modified with both the global and the unp 619 * pool token held. 620 */ 621 so = msg->base.nm_so; 622 unp = unp_getsocktoken(so); 623 624 if (!UNP_ISATTACHED(unp)) { 625 error = EINVAL; 626 goto release; 627 } 628 629 if (msg->send.nm_flags & PRUS_OOB) { 630 error = EOPNOTSUPP; 631 goto release; 632 } 633 634 wakeup_start_delayed(); 635 636 if (control && (error = unp_internalize(control, msg->send.nm_td))) 637 goto release; 638 639 switch (so->so_type) { 640 case SOCK_DGRAM: 641 { 642 struct sockaddr *from; 643 644 if (msg->send.nm_addr) { 645 if (unp->unp_conn) { 646 error = EISCONN; 647 break; 648 } 649 lwkt_gettoken(&unp_token); 650 error = unp_find_lockref(msg->send.nm_addr, 651 msg->send.nm_td, so->so_type, &unp2); 652 if (error) { 653 lwkt_reltoken(&unp_token); 654 break; 655 } 656 /* 657 * NOTE: 658 * unp2 is locked and referenced. 659 * 660 * We could unlock unp2 now, since it was checked 661 * and referenced. 662 */ 663 unp_reltoken(unp2); 664 lwkt_reltoken(&unp_token); 665 } else { 666 if (unp->unp_conn == NULL) { 667 error = ENOTCONN; 668 break; 669 } 670 unp2 = unp->unp_conn; 671 unp_reference(unp2); 672 } 673 /* NOTE: unp2 is referenced. */ 674 so2 = unp2->unp_socket; 675 676 if (unp->unp_addr) 677 from = (struct sockaddr *)unp->unp_addr; 678 else 679 from = &sun_noname; 680 681 lwkt_gettoken(&so2->so_rcv.ssb_token); 682 if (ssb_appendaddr(&so2->so_rcv, from, m, control)) { 683 sorwakeup(so2); 684 m = NULL; 685 control = NULL; 686 } else { 687 error = ENOBUFS; 688 } 689 lwkt_reltoken(&so2->so_rcv.ssb_token); 690 691 unp_free(unp2); 692 break; 693 } 694 695 case SOCK_STREAM: 696 case SOCK_SEQPACKET: 697 /* Connect if not connected yet. */ 698 /* 699 * Note: A better implementation would complain 700 * if not equal to the peer's address. 701 */ 702 if (unp->unp_conn == NULL) { 703 if (msg->send.nm_addr) { 704 error = unp_connect(so, 705 msg->send.nm_addr, 706 msg->send.nm_td); 707 if (error) 708 break; /* XXX */ 709 } 710 /* 711 * NOTE: 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 716 * assigned. 717 */ 718 if (unp->unp_conn == NULL) { 719 error = ENOTCONN; 720 break; 721 } 722 } 723 if (so->so_state & SS_CANTSENDMORE) { 724 error = EPIPE; 725 break; 726 } 727 728 unp2 = unp->unp_conn; 729 KASSERT(unp2 != NULL, ("unp is not connected")); 730 so2 = unp2->unp_socket; 731 732 unp_reference(unp2); 733 734 /* 735 * Send to paired receive port, and then reduce 736 * send buffer hiwater marks to maintain backpressure. 737 * Wake up readers. 738 */ 739 lwkt_gettoken(&so2->so_rcv.ssb_token); 740 if (control) { 741 if (ssb_appendcontrol(&so2->so_rcv, m, control)) { 742 control = NULL; 743 m = NULL; 744 } 745 } else if (so->so_type == SOCK_SEQPACKET) { 746 sbappendrecord(&so2->so_rcv.sb, m); 747 m = NULL; 748 } else { 749 sbappend(&so2->so_rcv.sb, m); 750 m = NULL; 751 } 752 753 /* 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. 757 */ 758 if (so2->so_rcv.ssb_cc >= so->so_snd.ssb_hiwat || 759 so2->so_rcv.ssb_mbcnt >= so->so_snd.ssb_mbmax 760 ) { 761 atomic_set_int(&so->so_snd.ssb_flags, SSB_STOP); 762 } 763 lwkt_reltoken(&so2->so_rcv.ssb_token); 764 sorwakeup(so2); 765 766 unp_free(unp2); 767 break; 768 769 default: 770 panic("uipc_send unknown socktype"); 771 } 772 773 /* 774 * SEND_EOF is equivalent to a SEND followed by a SHUTDOWN. 775 */ 776 if (msg->send.nm_flags & PRUS_EOF) { 777 socantsendmore(so); 778 unp_shutdown(unp); 779 } 780 781 if (control && error != 0) 782 unp_dispose(control); 783 release: 784 unp_reltoken(unp); 785 wakeup_end_delayed(); 786 787 if (control) 788 m_freem(control); 789 if (m) 790 m_freem(m); 791 lwkt_replymsg(&msg->lmsg, error); 792 } 793 794 /* 795 * MPSAFE 796 */ 797 static void 798 uipc_sense(netmsg_t msg) 799 { 800 struct unpcb *unp; 801 struct socket *so; 802 struct stat *sb; 803 int error; 804 805 so = msg->base.nm_so; 806 sb = msg->sense.nm_stat; 807 808 /* 809 * so_pcb is only modified with both the global and the unp 810 * pool token held. 811 */ 812 unp = unp_getsocktoken(so); 813 814 if (!UNP_ISATTACHED(unp)) { 815 error = EINVAL; 816 goto done; 817 } 818 819 sb->st_blksize = so->so_snd.ssb_hiwat; 820 sb->st_dev = NOUDEV; 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); 825 } 826 sb->st_ino = unp->unp_ino; 827 error = 0; 828 done: 829 unp_reltoken(unp); 830 lwkt_replymsg(&msg->lmsg, error); 831 } 832 833 static void 834 uipc_shutdown(netmsg_t msg) 835 { 836 struct socket *so; 837 struct unpcb *unp; 838 int error; 839 840 /* 841 * so_pcb is only modified with both the global and the unp 842 * pool token held. 843 */ 844 so = msg->base.nm_so; 845 unp = unp_getsocktoken(so); 846 847 if (UNP_ISATTACHED(unp)) { 848 socantsendmore(so); 849 unp_shutdown(unp); 850 error = 0; 851 } else { 852 error = EINVAL; 853 } 854 855 unp_reltoken(unp); 856 lwkt_replymsg(&msg->lmsg, error); 857 } 858 859 static void 860 uipc_sockaddr(netmsg_t msg) 861 { 862 struct unpcb *unp; 863 int error; 864 865 /* 866 * so_pcb is only modified with both the global and the unp 867 * pool token held. 868 */ 869 unp = unp_getsocktoken(msg->base.nm_so); 870 871 if (UNP_ISATTACHED(unp)) { 872 if (unp->unp_addr) { 873 *msg->sockaddr.nm_nam = 874 dup_sockaddr((struct sockaddr *)unp->unp_addr); 875 } 876 error = 0; 877 } else { 878 error = EINVAL; 879 } 880 881 unp_reltoken(unp); 882 lwkt_replymsg(&msg->lmsg, error); 883 } 884 885 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 905 }; 906 907 void 908 uipc_ctloutput(netmsg_t msg) 909 { 910 struct socket *so; 911 struct sockopt *sopt; 912 struct unpcb *unp; 913 int error = 0; 914 915 so = msg->base.nm_so; 916 sopt = msg->ctloutput.nm_sopt; 917 918 lwkt_gettoken(&unp_token); 919 unp = unp_getsocktoken(so); 920 921 if (!UNP_ISATTACHED(unp)) { 922 error = EINVAL; 923 goto done; 924 } 925 926 switch (sopt->sopt_dir) { 927 case SOPT_GET: 928 switch (sopt->sopt_name) { 929 case LOCAL_PEERCRED: 930 if (unp->unp_flags & UNP_HAVEPC) 931 soopt_from_kbuf(sopt, &unp->unp_peercred, 932 sizeof(unp->unp_peercred)); 933 else { 934 if (so->so_type == SOCK_STREAM) 935 error = ENOTCONN; 936 else if (so->so_type == SOCK_SEQPACKET) 937 error = ENOTCONN; 938 else 939 error = EINVAL; 940 } 941 break; 942 default: 943 error = EOPNOTSUPP; 944 break; 945 } 946 break; 947 case SOPT_SET: 948 default: 949 error = EOPNOTSUPP; 950 break; 951 } 952 953 done: 954 unp_reltoken(unp); 955 lwkt_reltoken(&unp_token); 956 957 lwkt_replymsg(&msg->lmsg, error); 958 } 959 960 /* 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. 964 * 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. 968 * 969 * We want the local send/recv space to be significant larger then lo0's 970 * mtu of 16384. 971 * 972 * We no longer need to worry about avoiding the windows scaling option. 973 * Programs which use unix domain sockets expect larger defaults these days. 974 */ 975 #ifndef PIPSIZ 976 #define PIPSIZ 65536 977 #endif 978 static u_long unpst_sendspace = PIPSIZ; 979 static u_long unpst_recvspace = PIPSIZ; 980 static u_long unpdg_sendspace = PIPSIZ; /* really max datagram size */ 981 static u_long unpdg_recvspace = PIPSIZ; 982 static u_long unpsp_sendspace = PIPSIZ; /* really max datagram size */ 983 static u_long unpsp_recvspace = PIPSIZ; 984 985 SYSCTL_DECL(_net_local_stream); 986 SYSCTL_DECL(_net_local_dgram); 987 SYSCTL_DECL(_net_local_seqpacket); 988 989 SYSCTL_ULONG(_net_local_stream, OID_AUTO, sendspace, CTLFLAG_RW, 990 &unpst_sendspace, 0, "Size of stream socket send buffer"); 991 SYSCTL_ULONG(_net_local_stream, OID_AUTO, recvspace, CTLFLAG_RW, 992 &unpst_recvspace, 0, "Size of stream socket receive buffer"); 993 994 SYSCTL_ULONG(_net_local_dgram, OID_AUTO, maxdgram, CTLFLAG_RW, 995 &unpdg_sendspace, 0, "Max datagram socket size"); 996 SYSCTL_ULONG(_net_local_dgram, OID_AUTO, recvspace, CTLFLAG_RW, 997 &unpdg_recvspace, 0, "Size of datagram socket receive buffer"); 998 999 SYSCTL_ULONG(_net_local_seqpacket, OID_AUTO, maxseqpacket, CTLFLAG_RW, 1000 &unpsp_sendspace, 0, "Default seqpacket send space."); 1001 SYSCTL_ULONG(_net_local_seqpacket, OID_AUTO, recvspace, CTLFLAG_RW, 1002 &unpsp_recvspace, 0, "Default seqpacket receive space."); 1003 1004 1005 static int 1006 unp_attach(struct socket *so, struct pru_attach_info *ai) 1007 { 1008 struct unp_global_head *head; 1009 struct unpcb *unp; 1010 int error; 1011 1012 lwkt_gettoken(&unp_token); 1013 1014 if (so->so_snd.ssb_hiwat == 0 || so->so_rcv.ssb_hiwat == 0) { 1015 switch (so->so_type) { 1016 case SOCK_STREAM: 1017 error = soreserve(so, unpst_sendspace, unpst_recvspace, 1018 ai->sb_rlimit); 1019 break; 1020 case SOCK_DGRAM: 1021 error = soreserve(so, unpdg_sendspace, unpdg_recvspace, 1022 ai->sb_rlimit); 1023 break; 1024 case SOCK_SEQPACKET: 1025 error = soreserve(so, unpsp_sendspace, unpsp_recvspace, 1026 ai->sb_rlimit); 1027 break; 1028 default: 1029 panic("unp_attach"); 1030 } 1031 if (error) 1032 goto failed; 1033 } 1034 1035 /* 1036 * In order to support sendfile we have to set either SSB_STOPSUPP 1037 * or SSB_PREALLOC. Unix domain sockets use the SSB_STOP flow 1038 * control mechanism. 1039 */ 1040 if (so->so_type == SOCK_STREAM) { 1041 atomic_set_int(&so->so_rcv.ssb_flags, SSB_STOPSUPP); 1042 atomic_set_int(&so->so_snd.ssb_flags, SSB_STOPSUPP); 1043 } 1044 1045 unp = kmalloc(sizeof(*unp), M_UNPCB, M_WAITOK | M_ZERO | M_NULLOK); 1046 if (unp == NULL) { 1047 error = ENOBUFS; 1048 goto failed; 1049 } 1050 unp->unp_refcnt = 1; 1051 unp->unp_gencnt = ++unp_gencnt; 1052 LIST_INIT(&unp->unp_refs); 1053 unp->unp_socket = so; 1054 unp->unp_rvnode = ai->fd_rdir; /* jail cruft XXX JH */ 1055 so->so_pcb = (caddr_t)unp; 1056 soreference(so); 1057 1058 head = unp_globalhead(so->so_type); 1059 TAILQ_INSERT_TAIL(&head->list, unp, unp_link); 1060 head->count++; 1061 error = 0; 1062 failed: 1063 lwkt_reltoken(&unp_token); 1064 return error; 1065 } 1066 1067 static void 1068 unp_detach(struct unpcb *unp) 1069 { 1070 struct socket *so; 1071 1072 lwkt_gettoken(&unp_token); 1073 lwkt_getpooltoken(unp); 1074 1075 so = unp->unp_socket; 1076 1077 unp->unp_gencnt = ++unp_gencnt; 1078 if (unp->unp_vnode) { 1079 unp->unp_vnode->v_socket = NULL; 1080 vrele(unp->unp_vnode); 1081 unp->unp_vnode = NULL; 1082 } 1083 soisdisconnected(so); 1084 KKASSERT(so->so_pcb == unp); 1085 so->so_pcb = NULL; /* both tokens required */ 1086 unp->unp_socket = NULL; 1087 1088 lwkt_relpooltoken(unp); 1089 lwkt_reltoken(&unp_token); 1090 1091 sofree(so); 1092 1093 KASSERT(unp->unp_conn == NULL, ("unp is still connected")); 1094 KASSERT(LIST_EMPTY(&unp->unp_refs), ("unp still has references")); 1095 1096 if (unp->unp_addr) 1097 kfree(unp->unp_addr, M_SONAME); 1098 kfree(unp, M_UNPCB); 1099 1100 if (unp_rights) 1101 taskqueue_enqueue(unp_taskqueue, &unp_gc_task); 1102 } 1103 1104 static int 1105 unp_bind(struct unpcb *unp, struct sockaddr *nam, struct thread *td) 1106 { 1107 struct proc *p = td->td_proc; 1108 struct sockaddr_un *soun = (struct sockaddr_un *)nam; 1109 struct vnode *vp; 1110 struct vattr vattr; 1111 int error, namelen; 1112 struct nlookupdata nd; 1113 char buf[SOCK_MAXADDRLEN]; 1114 1115 ASSERT_LWKT_TOKEN_HELD(&unp_token); 1116 UNP_ASSERT_TOKEN_HELD(unp); 1117 1118 if (unp->unp_vnode != NULL) 1119 return EINVAL; 1120 1121 namelen = soun->sun_len - offsetof(struct sockaddr_un, sun_path); 1122 if (namelen <= 0) 1123 return EINVAL; 1124 strncpy(buf, soun->sun_path, namelen); 1125 buf[namelen] = 0; /* null-terminate the string */ 1126 error = nlookup_init(&nd, buf, UIO_SYSSPACE, 1127 NLC_LOCKVP | NLC_CREATE | NLC_REFDVP); 1128 if (error == 0) 1129 error = nlookup(&nd); 1130 if (error == 0 && nd.nl_nch.ncp->nc_vp != NULL) 1131 error = EADDRINUSE; 1132 if (error) 1133 goto done; 1134 1135 VATTR_NULL(&vattr); 1136 vattr.va_type = VSOCK; 1137 vattr.va_mode = (ACCESSPERMS & ~p->p_fd->fd_cmask); 1138 error = VOP_NCREATE(&nd.nl_nch, nd.nl_dvp, &vp, nd.nl_cred, &vattr); 1139 if (error == 0) { 1140 if (unp->unp_vnode == NULL) { 1141 vp->v_socket = unp->unp_socket; 1142 unp->unp_vnode = vp; 1143 unp->unp_addr = (struct sockaddr_un *)dup_sockaddr(nam); 1144 vn_unlock(vp); 1145 } else { 1146 vput(vp); /* late race */ 1147 error = EINVAL; 1148 } 1149 } 1150 done: 1151 nlookup_done(&nd); 1152 return (error); 1153 } 1154 1155 static int 1156 unp_connect(struct socket *so, struct sockaddr *nam, struct thread *td) 1157 { 1158 struct unpcb *unp, *unp2; 1159 int error, flags = 0; 1160 1161 lwkt_gettoken(&unp_token); 1162 1163 unp = unp_getsocktoken(so); 1164 if (!UNP_ISATTACHED(unp)) { 1165 error = EINVAL; 1166 goto failed; 1167 } 1168 1169 if ((unp->unp_flags & UNP_CONNECTING) || unp->unp_conn != NULL) { 1170 error = EISCONN; 1171 goto failed; 1172 } 1173 1174 flags = UNP_CONNECTING; 1175 unp_setflags(unp, flags); 1176 1177 error = unp_find_lockref(nam, td, so->so_type, &unp2); 1178 if (error) 1179 goto failed; 1180 /* 1181 * NOTE: 1182 * unp2 is locked and referenced. 1183 */ 1184 1185 if (so->so_proto->pr_flags & PR_CONNREQUIRED) { 1186 struct socket *so2, *so3; 1187 struct unpcb *unp3; 1188 1189 so2 = unp2->unp_socket; 1190 if (!(so2->so_options & SO_ACCEPTCONN) || 1191 /* listen is not completed yet */ 1192 !(unp2->unp_flags & UNP_HAVEPCCACHED) || 1193 (so3 = sonewconn_faddr(so2, 0, NULL, 1194 TRUE /* keep ref */)) == NULL) { 1195 error = ECONNREFUSED; 1196 goto done; 1197 } 1198 /* so3 has a socket reference. */ 1199 1200 unp3 = unp_getsocktoken(so3); 1201 if (!UNP_ISATTACHED(unp3)) { 1202 unp_reltoken(unp3); 1203 /* 1204 * Already aborted; we only need to drop the 1205 * socket reference held by sonewconn_faddr(). 1206 */ 1207 sofree(so3); 1208 error = ECONNREFUSED; 1209 goto done; 1210 } 1211 unp_reference(unp3); 1212 /* 1213 * NOTE: 1214 * unp3 is locked and referenced. 1215 */ 1216 1217 /* 1218 * Release so3 socket reference held by sonewconn_faddr(). 1219 * Since we have referenced unp3, neither unp3 nor so3 will 1220 * be destroyed here. 1221 */ 1222 sofree(so3); 1223 1224 if (unp2->unp_addr != NULL) { 1225 unp3->unp_addr = (struct sockaddr_un *) 1226 dup_sockaddr((struct sockaddr *)unp2->unp_addr); 1227 } 1228 1229 /* 1230 * unp_peercred management: 1231 * 1232 * The connecter's (client's) credentials are copied 1233 * from its process structure at the time of connect() 1234 * (which is now). 1235 */ 1236 cru2x(td->td_proc->p_ucred, &unp3->unp_peercred); 1237 unp_setflags(unp3, UNP_HAVEPC); 1238 /* 1239 * The receiver's (server's) credentials are copied 1240 * from the unp_peercred member of socket on which the 1241 * former called listen(); unp_listen() cached that 1242 * process's credentials at that time so we can use 1243 * them now. 1244 */ 1245 KASSERT(unp2->unp_flags & UNP_HAVEPCCACHED, 1246 ("unp_connect: listener without cached peercred")); 1247 memcpy(&unp->unp_peercred, &unp2->unp_peercred, 1248 sizeof(unp->unp_peercred)); 1249 unp_setflags(unp, UNP_HAVEPC); 1250 1251 error = unp_connect_pair(unp, unp3); 1252 if (error) 1253 soabort_direct(so3); 1254 1255 /* Done with unp3 */ 1256 unp_free(unp3); 1257 unp_reltoken(unp3); 1258 } else { 1259 error = unp_connect_pair(unp, unp2); 1260 } 1261 done: 1262 unp_free(unp2); 1263 unp_reltoken(unp2); 1264 failed: 1265 if (flags) 1266 unp_clrflags(unp, flags); 1267 unp_reltoken(unp); 1268 1269 lwkt_reltoken(&unp_token); 1270 return (error); 1271 } 1272 1273 /* 1274 * Connect two unix domain sockets together. 1275 * 1276 * NOTE: Semantics for any change to unp_conn requires that the per-unp 1277 * pool token also be held. 1278 */ 1279 int 1280 unp_connect2(struct socket *so, struct socket *so2) 1281 { 1282 struct unpcb *unp, *unp2; 1283 int error; 1284 1285 lwkt_gettoken(&unp_token); 1286 if (so2->so_type != so->so_type) { 1287 lwkt_reltoken(&unp_token); 1288 return (EPROTOTYPE); 1289 } 1290 unp = unp_getsocktoken(so); 1291 unp2 = unp_getsocktoken(so2); 1292 1293 if (!UNP_ISATTACHED(unp)) { 1294 error = EINVAL; 1295 goto done; 1296 } 1297 if (!UNP_ISATTACHED(unp2)) { 1298 error = ECONNREFUSED; 1299 goto done; 1300 } 1301 1302 if (unp->unp_conn != NULL) { 1303 error = EISCONN; 1304 goto done; 1305 } 1306 if ((so->so_type == SOCK_STREAM || so->so_type == SOCK_SEQPACKET) && 1307 unp2->unp_conn != NULL) { 1308 error = EISCONN; 1309 goto done; 1310 } 1311 1312 error = unp_connect_pair(unp, unp2); 1313 done: 1314 unp_reltoken(unp2); 1315 unp_reltoken(unp); 1316 lwkt_reltoken(&unp_token); 1317 return (error); 1318 } 1319 1320 /* 1321 * Disconnect a unix domain socket pair. 1322 * 1323 * NOTE: Semantics for any change to unp_conn requires that the per-unp 1324 * pool token also be held. 1325 */ 1326 static void 1327 unp_disconnect(struct unpcb *unp, int error) 1328 { 1329 struct socket *so = unp->unp_socket; 1330 struct unpcb *unp2; 1331 1332 ASSERT_LWKT_TOKEN_HELD(&unp_token); 1333 UNP_ASSERT_TOKEN_HELD(unp); 1334 1335 if (error) 1336 so->so_error = error; 1337 1338 while ((unp2 = unp->unp_conn) != NULL) { 1339 lwkt_getpooltoken(unp2); 1340 if (unp2 == unp->unp_conn) 1341 break; 1342 lwkt_relpooltoken(unp2); 1343 } 1344 if (unp2 == NULL) 1345 return; 1346 /* unp2 is locked. */ 1347 1348 KASSERT((unp2->unp_flags & UNP_DROPPED) == 0, ("unp2 was dropped")); 1349 1350 unp->unp_conn = NULL; 1351 1352 switch (so->so_type) { 1353 case SOCK_DGRAM: 1354 LIST_REMOVE(unp, unp_reflink); 1355 soclrstate(so, SS_ISCONNECTED); 1356 break; 1357 1358 case SOCK_STREAM: 1359 case SOCK_SEQPACKET: 1360 /* 1361 * Keep a reference before clearing the unp_conn 1362 * to avoid racing uipc_detach()/uipc_abort() in 1363 * other thread. 1364 */ 1365 unp_reference(unp2); 1366 KASSERT(unp2->unp_conn == unp, ("unp_conn mismatch")); 1367 unp2->unp_conn = NULL; 1368 1369 soisdisconnected(so); 1370 soisdisconnected(unp2->unp_socket); 1371 1372 unp_free(unp2); 1373 break; 1374 } 1375 1376 lwkt_relpooltoken(unp2); 1377 } 1378 1379 #ifdef notdef 1380 void 1381 unp_abort(struct unpcb *unp) 1382 { 1383 lwkt_gettoken(&unp_token); 1384 unp_free(unp); 1385 lwkt_reltoken(&unp_token); 1386 } 1387 #endif 1388 1389 static int 1390 prison_unpcb(struct thread *td, struct unpcb *unp) 1391 { 1392 struct proc *p; 1393 1394 if (td == NULL) 1395 return (0); 1396 if ((p = td->td_proc) == NULL) 1397 return (0); 1398 if (!p->p_ucred->cr_prison) 1399 return (0); 1400 if (p->p_fd->fd_rdir == unp->unp_rvnode) 1401 return (0); 1402 return (1); 1403 } 1404 1405 static int 1406 unp_pcblist(SYSCTL_HANDLER_ARGS) 1407 { 1408 struct unp_global_head *head = arg1; 1409 int error, i, n; 1410 struct unpcb *unp, *marker; 1411 1412 KKASSERT(curproc != NULL); 1413 1414 /* 1415 * The process of preparing the PCB list is too time-consuming and 1416 * resource-intensive to repeat twice on every request. 1417 */ 1418 if (req->oldptr == NULL) { 1419 n = head->count; 1420 req->oldidx = (n + n/8) * sizeof(struct xunpcb); 1421 return 0; 1422 } 1423 1424 if (req->newptr != NULL) 1425 return EPERM; 1426 1427 marker = kmalloc(sizeof(*marker), M_UNPCB, M_WAITOK | M_ZERO); 1428 marker->unp_flags |= UNP_MARKER; 1429 1430 lwkt_gettoken(&unp_token); 1431 1432 n = head->count; 1433 i = 0; 1434 error = 0; 1435 1436 TAILQ_INSERT_HEAD(&head->list, marker, unp_link); 1437 while ((unp = TAILQ_NEXT(marker, unp_link)) != NULL && i < n) { 1438 struct xunpcb xu; 1439 1440 TAILQ_REMOVE(&head->list, marker, unp_link); 1441 TAILQ_INSERT_AFTER(&head->list, unp, marker, unp_link); 1442 1443 if (unp->unp_flags & UNP_MARKER) 1444 continue; 1445 if (prison_unpcb(req->td, unp)) 1446 continue; 1447 1448 xu.xu_len = sizeof(xu); 1449 xu.xu_unpp = unp; 1450 1451 /* 1452 * NOTE: 1453 * unp->unp_addr and unp->unp_conn are protected by 1454 * unp_token. So if we want to get rid of unp_token 1455 * or reduce the coverage of unp_token, care must be 1456 * taken. 1457 */ 1458 if (unp->unp_addr) { 1459 bcopy(unp->unp_addr, &xu.xu_addr, 1460 unp->unp_addr->sun_len); 1461 } 1462 if (unp->unp_conn && unp->unp_conn->unp_addr) { 1463 bcopy(unp->unp_conn->unp_addr, 1464 &xu.xu_caddr, 1465 unp->unp_conn->unp_addr->sun_len); 1466 } 1467 bcopy(unp, &xu.xu_unp, sizeof(*unp)); 1468 sotoxsocket(unp->unp_socket, &xu.xu_socket); 1469 1470 /* NOTE: This could block and temporarily release unp_token */ 1471 error = SYSCTL_OUT(req, &xu, sizeof(xu)); 1472 if (error) 1473 break; 1474 ++i; 1475 } 1476 TAILQ_REMOVE(&head->list, marker, unp_link); 1477 1478 lwkt_reltoken(&unp_token); 1479 1480 kfree(marker, M_UNPCB); 1481 return error; 1482 } 1483 1484 SYSCTL_PROC(_net_local_dgram, OID_AUTO, pcblist, CTLFLAG_RD, 1485 &unp_dgram_head, 0, unp_pcblist, "S,xunpcb", 1486 "List of active local datagram sockets"); 1487 SYSCTL_PROC(_net_local_stream, OID_AUTO, pcblist, CTLFLAG_RD, 1488 &unp_stream_head, 0, unp_pcblist, "S,xunpcb", 1489 "List of active local stream sockets"); 1490 SYSCTL_PROC(_net_local_seqpacket, OID_AUTO, pcblist, CTLFLAG_RD, 1491 &unp_seqpkt_head, 0, unp_pcblist, "S,xunpcb", 1492 "List of active local seqpacket sockets"); 1493 1494 static void 1495 unp_shutdown(struct unpcb *unp) 1496 { 1497 struct socket *so; 1498 1499 if ((unp->unp_socket->so_type == SOCK_STREAM || 1500 unp->unp_socket->so_type == SOCK_SEQPACKET) && 1501 unp->unp_conn != NULL && (so = unp->unp_conn->unp_socket)) { 1502 socantrcvmore(so); 1503 } 1504 } 1505 1506 #ifdef notdef 1507 void 1508 unp_drain(void) 1509 { 1510 lwkt_gettoken(&unp_token); 1511 lwkt_reltoken(&unp_token); 1512 } 1513 #endif 1514 1515 int 1516 unp_externalize(struct mbuf *rights, int flags) 1517 { 1518 struct thread *td = curthread; 1519 struct proc *p = td->td_proc; /* XXX */ 1520 struct lwp *lp = td->td_lwp; 1521 struct cmsghdr *cm = mtod(rights, struct cmsghdr *); 1522 int *fdp; 1523 int i; 1524 struct file **rp; 1525 struct file *fp; 1526 int newfds = (cm->cmsg_len - (CMSG_DATA(cm) - (u_char *)cm)) 1527 / sizeof(struct file *); 1528 int f; 1529 1530 lwkt_gettoken(&unp_rights_token); 1531 1532 /* 1533 * if the new FD's will not fit, then we free them all 1534 */ 1535 if (!fdavail(p, newfds)) { 1536 rp = (struct file **)CMSG_DATA(cm); 1537 for (i = 0; i < newfds; i++) { 1538 fp = *rp; 1539 /* 1540 * zero the pointer before calling unp_discard, 1541 * since it may end up in unp_gc().. 1542 */ 1543 *rp++ = NULL; 1544 unp_discard(fp, NULL); 1545 } 1546 lwkt_reltoken(&unp_rights_token); 1547 return (EMSGSIZE); 1548 } 1549 1550 /* 1551 * now change each pointer to an fd in the global table to 1552 * an integer that is the index to the local fd table entry 1553 * that we set up to point to the global one we are transferring. 1554 * Since the sizeof(struct file *) is bigger than or equal to 1555 * the sizeof(int), we do it in forward order. In that case, 1556 * an integer will always come in the same place or before its 1557 * corresponding struct file pointer. 1558 * 1559 * Hold revoke_token in 'shared' mode, so that we won't miss 1560 * the FREVOKED update on fps being externalized (fsetfd). 1561 */ 1562 lwkt_gettoken_shared(&revoke_token); 1563 fdp = (int *)CMSG_DATA(cm); 1564 rp = (struct file **)CMSG_DATA(cm); 1565 for (i = 0; i < newfds; i++) { 1566 if (fdalloc(p, 0, &f)) { 1567 int j; 1568 1569 /* 1570 * Previous fdavail() can't garantee 1571 * fdalloc() success due to SMP race. 1572 * Just clean up and return the same 1573 * error value as if fdavail() failed. 1574 */ 1575 lwkt_reltoken(&revoke_token); 1576 1577 /* Close externalized files */ 1578 for (j = 0; j < i; j++) 1579 kern_close(fdp[j]); 1580 /* Discard the rest of internal files */ 1581 for (; i < newfds; i++) 1582 unp_discard(rp[i], NULL); 1583 /* Wipe out the control message */ 1584 for (i = 0; i < newfds; i++) 1585 rp[i] = NULL; 1586 1587 lwkt_reltoken(&unp_rights_token); 1588 return (EMSGSIZE); 1589 } 1590 fp = rp[i]; 1591 unp_fp_externalize(lp, fp, f, flags); 1592 fdp[i] = f; 1593 } 1594 lwkt_reltoken(&revoke_token); 1595 1596 lwkt_reltoken(&unp_rights_token); 1597 1598 /* 1599 * Adjust length, in case sizeof(struct file *) and sizeof(int) 1600 * differs. 1601 */ 1602 cm->cmsg_len = CMSG_LEN(newfds * sizeof(int)); 1603 rights->m_len = cm->cmsg_len; 1604 1605 return (0); 1606 } 1607 1608 static void 1609 unp_fp_externalize(struct lwp *lp, struct file *fp, int fd, int flags) 1610 { 1611 if (lp) { 1612 struct filedesc *fdp = lp->lwp_proc->p_fd; 1613 1614 KKASSERT(fd >= 0); 1615 if (fp->f_flag & FREVOKED) { 1616 struct file *fx; 1617 int error; 1618 1619 kprintf("Warning: revoked fp exiting unix socket\n"); 1620 error = falloc(lp, &fx, NULL); 1621 if (error == 0) { 1622 if (flags & MSG_CMSG_CLOEXEC) 1623 fdp->fd_files[fd].fileflags |= UF_EXCLOSE; 1624 fsetfd(fdp, fx, fd); 1625 fdrop(fx); 1626 } else { 1627 fsetfd(fdp, NULL, fd); 1628 } 1629 } else { 1630 if (flags & MSG_CMSG_CLOEXEC) 1631 fdp->fd_files[fd].fileflags |= UF_EXCLOSE; 1632 fsetfd(fdp, fp, fd); 1633 } 1634 } 1635 unp_del_right(fp); 1636 fdrop(fp); 1637 } 1638 1639 void 1640 unp_init(void) 1641 { 1642 TAILQ_INIT(&unp_stream_head.list); 1643 TAILQ_INIT(&unp_dgram_head.list); 1644 TAILQ_INIT(&unp_seqpkt_head.list); 1645 1646 SLIST_INIT(&unp_defdiscard_head); 1647 spin_init(&unp_defdiscard_spin, "unpdisc"); 1648 TASK_INIT(&unp_defdiscard_task, 0, unp_defdiscard_taskfunc, NULL); 1649 1650 /* 1651 * This implies that only one gc can be in-progress at any 1652 * given moment. 1653 */ 1654 TASK_INIT(&unp_gc_task, 0, unp_gc, NULL); 1655 1656 unp_gc_marker = kmalloc(sizeof(*unp_gc_marker), M_UNPCB, 1657 M_WAITOK | M_ZERO); 1658 unp_gc_marker->unp_flags |= UNP_MARKER; 1659 1660 /* 1661 * Create taskqueue for defered discard, and stick it to 1662 * the last CPU. 1663 */ 1664 unp_taskqueue = taskqueue_create("unp_taskq", M_WAITOK, 1665 taskqueue_thread_enqueue, &unp_taskqueue); 1666 taskqueue_start_threads(&unp_taskqueue, 1, TDPRI_KERN_DAEMON, 1667 ncpus - 1, "unp taskq"); 1668 } 1669 1670 static int 1671 unp_internalize(struct mbuf *control, struct thread *td) 1672 { 1673 struct proc *p = td->td_proc; 1674 struct filedesc *fdescp; 1675 struct cmsghdr *cm = mtod(control, struct cmsghdr *); 1676 struct file **rp; 1677 struct file *fp; 1678 int i, fd, *fdp; 1679 struct cmsgcred *cmcred; 1680 int oldfds; 1681 u_int newlen; 1682 int error; 1683 1684 KKASSERT(p); 1685 1686 if ((cm->cmsg_type != SCM_RIGHTS && cm->cmsg_type != SCM_CREDS) || 1687 cm->cmsg_level != SOL_SOCKET || 1688 CMSG_ALIGN(cm->cmsg_len) != control->m_len) 1689 return EINVAL; 1690 1691 /* 1692 * Fill in credential information. 1693 */ 1694 if (cm->cmsg_type == SCM_CREDS) { 1695 cmcred = (struct cmsgcred *)CMSG_DATA(cm); 1696 cmcred->cmcred_pid = p->p_pid; 1697 cmcred->cmcred_uid = p->p_ucred->cr_ruid; 1698 cmcred->cmcred_gid = p->p_ucred->cr_rgid; 1699 cmcred->cmcred_euid = p->p_ucred->cr_uid; 1700 cmcred->cmcred_ngroups = MIN(p->p_ucred->cr_ngroups, 1701 CMGROUP_MAX); 1702 for (i = 0; i < cmcred->cmcred_ngroups; i++) 1703 cmcred->cmcred_groups[i] = p->p_ucred->cr_groups[i]; 1704 return 0; 1705 } 1706 1707 /* 1708 * cmsghdr may not be aligned, do not allow calculation(s) to 1709 * go negative. 1710 */ 1711 if (cm->cmsg_len < CMSG_LEN(0)) 1712 return EINVAL; 1713 1714 oldfds = (cm->cmsg_len - CMSG_LEN(0)) / sizeof(int); 1715 1716 /* 1717 * Now replace the integer FDs with pointers to 1718 * the associated global file table entry.. 1719 * Allocate a bigger buffer as necessary. But if an cluster is not 1720 * enough, return E2BIG. 1721 */ 1722 newlen = CMSG_LEN(oldfds * sizeof(struct file *)); 1723 if (newlen > MCLBYTES) 1724 return E2BIG; 1725 if (newlen - control->m_len > M_TRAILINGSPACE(control)) { 1726 if (control->m_flags & M_EXT) 1727 return E2BIG; 1728 MCLGET(control, M_WAITOK); 1729 if (!(control->m_flags & M_EXT)) 1730 return ENOBUFS; 1731 1732 /* copy the data to the cluster */ 1733 memcpy(mtod(control, char *), cm, cm->cmsg_len); 1734 cm = mtod(control, struct cmsghdr *); 1735 } 1736 1737 lwkt_gettoken(&unp_rights_token); 1738 1739 fdescp = p->p_fd; 1740 spin_lock_shared(&fdescp->fd_spin); 1741 1742 /* 1743 * check that all the FDs passed in refer to legal OPEN files 1744 * If not, reject the entire operation. 1745 */ 1746 fdp = (int *)CMSG_DATA(cm); 1747 for (i = 0; i < oldfds; i++) { 1748 fd = *fdp++; 1749 if ((unsigned)fd >= fdescp->fd_nfiles || 1750 fdescp->fd_files[fd].fp == NULL) { 1751 error = EBADF; 1752 goto done; 1753 } 1754 if (fdescp->fd_files[fd].fp->f_type == DTYPE_KQUEUE) { 1755 error = EOPNOTSUPP; 1756 goto done; 1757 } 1758 } 1759 1760 /* 1761 * Adjust length, in case sizeof(struct file *) and sizeof(int) 1762 * differs. 1763 */ 1764 cm->cmsg_len = newlen; 1765 control->m_len = CMSG_ALIGN(newlen); 1766 1767 /* 1768 * Transform the file descriptors into struct file pointers. 1769 * Since the sizeof(struct file *) is bigger than or equal to 1770 * the sizeof(int), we do it in reverse order so that the int 1771 * won't get trashed until we're done. 1772 */ 1773 fdp = (int *)CMSG_DATA(cm) + oldfds - 1; 1774 rp = (struct file **)CMSG_DATA(cm) + oldfds - 1; 1775 for (i = 0; i < oldfds; i++) { 1776 fp = fdescp->fd_files[*fdp--].fp; 1777 *rp-- = fp; 1778 fhold(fp); 1779 unp_add_right(fp); 1780 } 1781 error = 0; 1782 done: 1783 spin_unlock_shared(&fdescp->fd_spin); 1784 lwkt_reltoken(&unp_rights_token); 1785 return error; 1786 } 1787 1788 #ifdef UNP_GC_ALLFILES 1789 1790 /* 1791 * Garbage collect in-transit file descriptors that get lost due to 1792 * loops (i.e. when a socket is sent to another process over itself, 1793 * and more complex situations). 1794 * 1795 * NOT MPSAFE - TODO socket flush code and maybe fdrop. Rest is MPSAFE. 1796 */ 1797 1798 struct unp_gc_info { 1799 struct file **extra_ref; 1800 struct file *locked_fp; 1801 int defer; 1802 int index; 1803 int maxindex; 1804 }; 1805 1806 static void 1807 unp_gc(void *arg __unused, int pending __unused) 1808 { 1809 struct unp_gc_info info; 1810 struct file **fpp; 1811 int i; 1812 1813 lwkt_gettoken(&unp_rights_token); 1814 1815 /* 1816 * Before going through all this, set all FDs to be NOT defered 1817 * and NOT externally accessible (not marked). During the scan 1818 * a fd can be marked externally accessible but we may or may not 1819 * be able to immediately process it (controlled by FDEFER). 1820 * 1821 * If we loop sleep a bit. The complexity of the topology can cause 1822 * multiple loops. Also failure to acquire the socket's so_rcv 1823 * token can cause us to loop. 1824 */ 1825 allfiles_scan_exclusive(unp_gc_clearmarks, NULL); 1826 do { 1827 info.defer = 0; 1828 allfiles_scan_exclusive(unp_gc_checkmarks, &info); 1829 if (info.defer) 1830 tsleep(&info, 0, "gcagain", 1); 1831 } while (info.defer); 1832 1833 /* 1834 * We grab an extra reference to each of the file table entries 1835 * that are not otherwise accessible and then free the rights 1836 * that are stored in messages on them. 1837 * 1838 * The bug in the orginal code is a little tricky, so I'll describe 1839 * what's wrong with it here. 1840 * 1841 * It is incorrect to simply unp_discard each entry for f_msgcount 1842 * times -- consider the case of sockets A and B that contain 1843 * references to each other. On a last close of some other socket, 1844 * we trigger a gc since the number of outstanding rights (unp_rights) 1845 * is non-zero. If during the sweep phase the gc code unp_discards, 1846 * we end up doing a (full) fdrop on the descriptor. A fdrop on A 1847 * results in the following chain. Closef calls soo_close, which 1848 * calls soclose. Soclose calls first (through the switch 1849 * uipc_usrreq) unp_detach, which re-invokes unp_gc. Unp_gc simply 1850 * returns because the previous instance had set unp_gcing, and 1851 * we return all the way back to soclose, which marks the socket 1852 * with SS_NOFDREF, and then calls sofree. Sofree calls sorflush 1853 * to free up the rights that are queued in messages on the socket A, 1854 * i.e., the reference on B. The sorflush calls via the dom_dispose 1855 * switch unp_dispose, which unp_scans with unp_discard. This second 1856 * instance of unp_discard just calls fdrop on B. 1857 * 1858 * Well, a similar chain occurs on B, resulting in a sorflush on B, 1859 * which results in another fdrop on A. Unfortunately, A is already 1860 * being closed, and the descriptor has already been marked with 1861 * SS_NOFDREF, and soclose panics at this point. 1862 * 1863 * Here, we first take an extra reference to each inaccessible 1864 * descriptor. Then, we call sorflush ourself, since we know 1865 * it is a Unix domain socket anyhow. After we destroy all the 1866 * rights carried in messages, we do a last fdrop to get rid 1867 * of our extra reference. This is the last close, and the 1868 * unp_detach etc will shut down the socket. 1869 * 1870 * 91/09/19, bsy@cs.cmu.edu 1871 */ 1872 info.extra_ref = kmalloc(256 * sizeof(struct file *), M_FILE, M_WAITOK); 1873 info.maxindex = 256; 1874 1875 do { 1876 /* 1877 * Look for matches 1878 */ 1879 info.index = 0; 1880 allfiles_scan_exclusive(unp_gc_checkrefs, &info); 1881 1882 /* 1883 * For each FD on our hit list, do the following two things 1884 */ 1885 for (i = info.index, fpp = info.extra_ref; --i >= 0; ++fpp) { 1886 struct file *tfp = *fpp; 1887 if (tfp->f_type == DTYPE_SOCKET && tfp->f_data != NULL) 1888 sorflush((struct socket *)(tfp->f_data)); 1889 } 1890 for (i = info.index, fpp = info.extra_ref; --i >= 0; ++fpp) 1891 fdrop(*fpp); 1892 } while (info.index == info.maxindex); 1893 1894 kfree((caddr_t)info.extra_ref, M_FILE); 1895 1896 lwkt_reltoken(&unp_rights_token); 1897 } 1898 1899 /* 1900 * MPSAFE - NOTE: filehead list and file pointer spinlocked on entry 1901 */ 1902 static int 1903 unp_gc_checkrefs(struct file *fp, void *data) 1904 { 1905 struct unp_gc_info *info = data; 1906 1907 if (fp->f_count == 0) 1908 return(0); 1909 if (info->index == info->maxindex) 1910 return(-1); 1911 1912 /* 1913 * If all refs are from msgs, and it's not marked accessible 1914 * then it must be referenced from some unreachable cycle 1915 * of (shut-down) FDs, so include it in our 1916 * list of FDs to remove 1917 */ 1918 if (fp->f_count == fp->f_msgcount && !(fp->f_flag & FMARK)) { 1919 info->extra_ref[info->index++] = fp; 1920 fhold(fp); 1921 } 1922 return(0); 1923 } 1924 1925 /* 1926 * MPSAFE - NOTE: filehead list and file pointer spinlocked on entry 1927 */ 1928 static int 1929 unp_gc_clearmarks(struct file *fp, void *data __unused) 1930 { 1931 atomic_clear_int(&fp->f_flag, FMARK | FDEFER); 1932 return(0); 1933 } 1934 1935 /* 1936 * MPSAFE - NOTE: filehead list and file pointer spinlocked on entry 1937 */ 1938 static int 1939 unp_gc_checkmarks(struct file *fp, void *data) 1940 { 1941 struct unp_gc_info *info = data; 1942 struct socket *so; 1943 1944 /* 1945 * If the file is not open, skip it. Make sure it isn't marked 1946 * defered or we could loop forever, in case we somehow race 1947 * something. 1948 */ 1949 if (fp->f_count == 0) { 1950 if (fp->f_flag & FDEFER) 1951 atomic_clear_int(&fp->f_flag, FDEFER); 1952 return(0); 1953 } 1954 /* 1955 * If we already marked it as 'defer' in a 1956 * previous pass, then try process it this time 1957 * and un-mark it 1958 */ 1959 if (fp->f_flag & FDEFER) { 1960 atomic_clear_int(&fp->f_flag, FDEFER); 1961 } else { 1962 /* 1963 * if it's not defered, then check if it's 1964 * already marked.. if so skip it 1965 */ 1966 if (fp->f_flag & FMARK) 1967 return(0); 1968 /* 1969 * If all references are from messages 1970 * in transit, then skip it. it's not 1971 * externally accessible. 1972 */ 1973 if (fp->f_count == fp->f_msgcount) 1974 return(0); 1975 /* 1976 * If it got this far then it must be 1977 * externally accessible. 1978 */ 1979 atomic_set_int(&fp->f_flag, FMARK); 1980 } 1981 1982 /* 1983 * either it was defered, or it is externally 1984 * accessible and not already marked so. 1985 * Now check if it is possibly one of OUR sockets. 1986 */ 1987 if (fp->f_type != DTYPE_SOCKET || 1988 (so = (struct socket *)fp->f_data) == NULL) { 1989 return(0); 1990 } 1991 if (so->so_proto->pr_domain != &localdomain || 1992 !(so->so_proto->pr_flags & PR_RIGHTS)) { 1993 return(0); 1994 } 1995 1996 /* 1997 * So, Ok, it's one of our sockets and it IS externally accessible 1998 * (or was defered). Now we look to see if we hold any file 1999 * descriptors in its message buffers. Follow those links and mark 2000 * them as accessible too. 2001 * 2002 * We are holding multiple spinlocks here, if we cannot get the 2003 * token non-blocking defer until the next loop. 2004 */ 2005 info->locked_fp = fp; 2006 if (lwkt_trytoken(&so->so_rcv.ssb_token)) { 2007 unp_scan(so->so_rcv.ssb_mb, unp_mark, info); 2008 lwkt_reltoken(&so->so_rcv.ssb_token); 2009 } else { 2010 atomic_set_int(&fp->f_flag, FDEFER); 2011 ++info->defer; 2012 } 2013 return (0); 2014 } 2015 2016 /* 2017 * Mark visibility. info->defer is recalculated on every pass. 2018 */ 2019 static void 2020 unp_mark(struct file *fp, void *data) 2021 { 2022 struct unp_gc_info *info = data; 2023 2024 if ((fp->f_flag & FMARK) == 0) { 2025 ++info->defer; 2026 atomic_set_int(&fp->f_flag, FMARK | FDEFER); 2027 } else if (fp->f_flag & FDEFER) { 2028 ++info->defer; 2029 } 2030 } 2031 2032 #else /* !UNP_GC_ALLFILES */ 2033 2034 /* 2035 * They are thread local and do not require explicit synchronization. 2036 */ 2037 static int unp_marked; 2038 static int unp_unreachable; 2039 2040 static void 2041 unp_accessable(struct file *fp, void *data __unused) 2042 { 2043 struct unpcb *unp; 2044 2045 if ((unp = unp_fp2unpcb(fp)) == NULL) 2046 return; 2047 if (unp->unp_gcflags & UNPGC_REF) 2048 return; 2049 unp->unp_gcflags &= ~UNPGC_DEAD; 2050 unp->unp_gcflags |= UNPGC_REF; 2051 unp_marked++; 2052 } 2053 2054 static void 2055 unp_gc_process(struct unpcb *unp) 2056 { 2057 struct file *fp; 2058 2059 /* Already processed. */ 2060 if (unp->unp_gcflags & UNPGC_SCANNED) 2061 return; 2062 fp = unp->unp_fp; 2063 2064 /* 2065 * Check for a socket potentially in a cycle. It must be in a 2066 * queue as indicated by msgcount, and this must equal the file 2067 * reference count. Note that when msgcount is 0 the file is NULL. 2068 */ 2069 if ((unp->unp_gcflags & UNPGC_REF) == 0 && fp && 2070 unp->unp_msgcount != 0 && fp->f_count == unp->unp_msgcount) { 2071 unp->unp_gcflags |= UNPGC_DEAD; 2072 unp_unreachable++; 2073 return; 2074 } 2075 2076 /* 2077 * Mark all sockets we reference with RIGHTS. 2078 */ 2079 if (UNP_ISATTACHED(unp)) { 2080 struct signalsockbuf *ssb = &unp->unp_socket->so_rcv; 2081 2082 unp_reference(unp); 2083 lwkt_gettoken(&ssb->ssb_token); 2084 /* 2085 * unp_token would be temporarily dropped, if getting 2086 * so_rcv token blocks, so we need to check unp state 2087 * here again. 2088 */ 2089 if (UNP_ISATTACHED(unp)) 2090 unp_scan(ssb->ssb_mb, unp_accessable, NULL); 2091 lwkt_reltoken(&ssb->ssb_token); 2092 unp->unp_gcflags |= UNPGC_SCANNED; 2093 unp_free(unp); 2094 } else { 2095 unp->unp_gcflags |= UNPGC_SCANNED; 2096 } 2097 } 2098 2099 static void 2100 unp_gc(void *arg __unused, int pending __unused) 2101 { 2102 struct unp_global_head *head; 2103 int h, filemax, fileidx, filetot; 2104 struct file **unref; 2105 struct unpcb *unp; 2106 2107 lwkt_gettoken(&unp_rights_token); 2108 lwkt_gettoken(&unp_token); 2109 2110 /* 2111 * First clear all gc flags from previous runs. 2112 */ 2113 for (h = 0; unp_heads[h] != NULL; ++h) { 2114 /* 2115 * NOTE: This loop does not block, so it is safe 2116 * to use TAILQ_FOREACH here. 2117 */ 2118 head = unp_heads[h]; 2119 TAILQ_FOREACH(unp, &head->list, unp_link) 2120 unp->unp_gcflags = 0; 2121 } 2122 2123 /* 2124 * Scan marking all reachable sockets with UNPGC_REF. Once a socket 2125 * is reachable all of the sockets it references are reachable. 2126 * Stop the scan once we do a complete loop without discovering 2127 * a new reachable socket. 2128 */ 2129 do { 2130 unp_unreachable = 0; 2131 unp_marked = 0; 2132 for (h = 0; unp_heads[h] != NULL; ++h) { 2133 head = unp_heads[h]; 2134 TAILQ_INSERT_HEAD(&head->list, unp_gc_marker, unp_link); 2135 while ((unp = TAILQ_NEXT(unp_gc_marker, unp_link)) 2136 != NULL) { 2137 TAILQ_REMOVE(&head->list, unp_gc_marker, 2138 unp_link); 2139 TAILQ_INSERT_AFTER(&head->list, unp, 2140 unp_gc_marker, unp_link); 2141 2142 if (unp->unp_flags & UNP_MARKER) 2143 continue; 2144 unp_gc_process(unp); 2145 } 2146 TAILQ_REMOVE(&head->list, unp_gc_marker, unp_link); 2147 } 2148 } while (unp_marked); 2149 2150 if (unp_unreachable == 0) 2151 goto done; 2152 2153 /* 2154 * We grab an extra reference to each of the file table entries 2155 * that are not otherwise accessible and then free the rights 2156 * that are stored in messages on them. 2157 * 2158 * The bug in the orginal code is a little tricky, so I'll describe 2159 * what's wrong with it here. 2160 * 2161 * It is incorrect to simply unp_discard each entry for f_msgcount 2162 * times -- consider the case of sockets A and B that contain 2163 * references to each other. On a last close of some other socket, 2164 * we trigger a gc since the number of outstanding rights (unp_rights) 2165 * is non-zero. If during the sweep phase the gc code unp_discards, 2166 * we end up doing a (full) fdrop on the descriptor. A fdrop on A 2167 * results in the following chain. Closef calls soo_close, which 2168 * calls soclose. Soclose calls first (through the switch 2169 * uipc_usrreq) unp_detach, which re-invokes unp_gc. Unp_gc simply 2170 * returns because the previous instance had set unp_gcing, and 2171 * we return all the way back to soclose, which marks the socket 2172 * with SS_NOFDREF, and then calls sofree. Sofree calls sorflush 2173 * to free up the rights that are queued in messages on the socket A, 2174 * i.e., the reference on B. The sorflush calls via the dom_dispose 2175 * switch unp_dispose, which unp_scans with unp_discard. This second 2176 * instance of unp_discard just calls fdrop on B. 2177 * 2178 * Well, a similar chain occurs on B, resulting in a sorflush on B, 2179 * which results in another fdrop on A. Unfortunately, A is already 2180 * being closed, and the descriptor has already been marked with 2181 * SS_NOFDREF, and soclose panics at this point. 2182 * 2183 * Here, we first take an extra reference to each inaccessible 2184 * descriptor. Then, we call sorflush ourself, since we know 2185 * it is a Unix domain socket anyhow. After we destroy all the 2186 * rights carried in messages, we do a last fdrop to get rid 2187 * of our extra reference. This is the last close, and the 2188 * unp_detach etc will shut down the socket. 2189 * 2190 * 91/09/19, bsy@cs.cmu.edu 2191 */ 2192 2193 filemax = unp_unreachable; 2194 if (filemax > UNP_GCFILE_MAX) 2195 filemax = UNP_GCFILE_MAX; 2196 unref = kmalloc(filemax * sizeof(struct file *), M_TEMP, M_WAITOK); 2197 2198 filetot = 0; 2199 do { 2200 int i; 2201 2202 /* 2203 * Iterate looking for sockets which have been specifically 2204 * marked as as unreachable and store them locally. 2205 */ 2206 fileidx = 0; 2207 for (h = 0; unp_heads[h] != NULL; ++h) { 2208 /* 2209 * NOTE: This loop does not block, so it is safe 2210 * to use TAILQ_FOREACH here. 2211 */ 2212 head = unp_heads[h]; 2213 TAILQ_FOREACH(unp, &head->list, unp_link) { 2214 struct file *fp; 2215 2216 if ((unp->unp_gcflags & UNPGC_DEAD) == 0) 2217 continue; 2218 unp->unp_gcflags &= ~UNPGC_DEAD; 2219 2220 fp = unp->unp_fp; 2221 if (unp->unp_msgcount == 0 || fp == NULL || 2222 fp->f_count != unp->unp_msgcount) 2223 continue; 2224 fhold(fp); 2225 2226 KASSERT(fileidx < filemax, 2227 ("invalid fileidx %d, filemax %d", 2228 fileidx, filemax)); 2229 unref[fileidx++] = fp; 2230 2231 KASSERT(filetot < unp_unreachable, 2232 ("invalid filetot %d and " 2233 "unp_unreachable %d", 2234 filetot, unp_unreachable)); 2235 ++filetot; 2236 2237 if (fileidx == filemax || 2238 filetot == unp_unreachable) 2239 goto dogc; 2240 } 2241 } 2242 dogc: 2243 /* 2244 * For each Unix domain socket on our hit list, do the 2245 * following two things. 2246 */ 2247 for (i = 0; i < fileidx; ++i) 2248 sorflush(unref[i]->f_data); 2249 for (i = 0; i < fileidx; ++i) 2250 fdrop(unref[i]); 2251 } while (fileidx == filemax && filetot < unp_unreachable); 2252 kfree(unref, M_TEMP); 2253 done: 2254 lwkt_reltoken(&unp_token); 2255 lwkt_reltoken(&unp_rights_token); 2256 } 2257 2258 #endif /* UNP_GC_ALLFILES */ 2259 2260 /* 2261 * Dispose of the fp's stored in a mbuf. 2262 * 2263 * The dds loop can cause additional fps to be entered onto the 2264 * list while it is running, flattening out the operation and avoiding 2265 * a deep kernel stack recursion. 2266 */ 2267 void 2268 unp_dispose(struct mbuf *m) 2269 { 2270 lwkt_gettoken(&unp_rights_token); 2271 if (m) 2272 unp_scan(m, unp_discard, NULL); 2273 lwkt_reltoken(&unp_rights_token); 2274 } 2275 2276 static int 2277 unp_listen(struct unpcb *unp, struct thread *td) 2278 { 2279 struct proc *p = td->td_proc; 2280 2281 ASSERT_LWKT_TOKEN_HELD(&unp_token); 2282 UNP_ASSERT_TOKEN_HELD(unp); 2283 2284 KKASSERT(p); 2285 cru2x(p->p_ucred, &unp->unp_peercred); 2286 unp_setflags(unp, UNP_HAVEPCCACHED); 2287 return (0); 2288 } 2289 2290 static void 2291 unp_scan(struct mbuf *m0, void (*op)(struct file *, void *), void *data) 2292 { 2293 struct mbuf *m; 2294 struct file **rp; 2295 struct cmsghdr *cm; 2296 int i; 2297 int qfds; 2298 2299 while (m0) { 2300 for (m = m0; m; m = m->m_next) { 2301 if (m->m_type == MT_CONTROL && 2302 m->m_len >= sizeof(*cm)) { 2303 cm = mtod(m, struct cmsghdr *); 2304 if (cm->cmsg_level != SOL_SOCKET || 2305 cm->cmsg_type != SCM_RIGHTS) 2306 continue; 2307 qfds = (cm->cmsg_len - CMSG_LEN(0)) / 2308 sizeof(void *); 2309 rp = (struct file **)CMSG_DATA(cm); 2310 for (i = 0; i < qfds; i++) 2311 (*op)(*rp++, data); 2312 break; /* XXX, but saves time */ 2313 } 2314 } 2315 m0 = m0->m_nextpkt; 2316 } 2317 } 2318 2319 /* 2320 * Discard a fp previously held in a unix domain socket mbuf. To 2321 * avoid blowing out the kernel stack due to contrived chain-reactions 2322 * we may have to defer the operation to a dedicated taskqueue. 2323 * 2324 * Caller holds unp_rights_token. 2325 */ 2326 static void 2327 unp_discard(struct file *fp, void *data __unused) 2328 { 2329 unp_del_right(fp); 2330 if (unp_fp2unpcb(fp) != NULL) { 2331 struct unp_defdiscard *d; 2332 2333 /* 2334 * This fp is a Unix domain socket itself and fdrop() 2335 * it here directly may cause deep unp_discard() 2336 * recursion, so the fdrop() is defered to the 2337 * dedicated taskqueue. 2338 */ 2339 d = kmalloc(sizeof(*d), M_UNPCB, M_WAITOK); 2340 d->fp = fp; 2341 2342 spin_lock(&unp_defdiscard_spin); 2343 SLIST_INSERT_HEAD(&unp_defdiscard_head, d, next); 2344 spin_unlock(&unp_defdiscard_spin); 2345 2346 taskqueue_enqueue(unp_taskqueue, &unp_defdiscard_task); 2347 } else { 2348 /* This fp is not a Unix domain socket */ 2349 fdrop(fp); 2350 } 2351 } 2352 2353 /* 2354 * NOTE: 2355 * unp_token must be held before calling this function to avoid name 2356 * resolution and v_socket accessing races, especially racing against 2357 * the unp_detach(). 2358 * 2359 * NOTE: 2360 * For anyone caring about unconnected Unix domain socket sending 2361 * performance, other approach could be taken... 2362 */ 2363 static int 2364 unp_find_lockref(struct sockaddr *nam, struct thread *td, short type, 2365 struct unpcb **unp_ret) 2366 { 2367 struct proc *p = td->td_proc; 2368 struct sockaddr_un *soun = (struct sockaddr_un *)nam; 2369 struct vnode *vp = NULL; 2370 struct socket *so; 2371 struct unpcb *unp; 2372 int error, len; 2373 struct nlookupdata nd; 2374 char buf[SOCK_MAXADDRLEN]; 2375 2376 ASSERT_LWKT_TOKEN_HELD(&unp_token); 2377 2378 *unp_ret = NULL; 2379 2380 len = nam->sa_len - offsetof(struct sockaddr_un, sun_path); 2381 if (len <= 0) { 2382 error = EINVAL; 2383 goto failed; 2384 } 2385 strncpy(buf, soun->sun_path, len); 2386 buf[len] = 0; 2387 2388 error = nlookup_init(&nd, buf, UIO_SYSSPACE, NLC_FOLLOW); 2389 if (error == 0) 2390 error = nlookup(&nd); 2391 if (error == 0) 2392 error = cache_vget(&nd.nl_nch, nd.nl_cred, LK_EXCLUSIVE, &vp); 2393 nlookup_done(&nd); 2394 if (error) { 2395 vp = NULL; 2396 goto failed; 2397 } 2398 2399 if (vp->v_type != VSOCK) { 2400 error = ENOTSOCK; 2401 goto failed; 2402 } 2403 error = VOP_EACCESS(vp, VWRITE, p->p_ucred); 2404 if (error) 2405 goto failed; 2406 so = vp->v_socket; 2407 if (so == NULL) { 2408 error = ECONNREFUSED; 2409 goto failed; 2410 } 2411 if (so->so_type != type) { 2412 error = EPROTOTYPE; 2413 goto failed; 2414 } 2415 2416 /* Lock this unp. */ 2417 unp = unp_getsocktoken(so); 2418 if (!UNP_ISATTACHED(unp)) { 2419 unp_reltoken(unp); 2420 error = ECONNREFUSED; 2421 goto failed; 2422 } 2423 /* And keep this unp referenced. */ 2424 unp_reference(unp); 2425 2426 /* Done! */ 2427 *unp_ret = unp; 2428 error = 0; 2429 failed: 2430 if (vp != NULL) 2431 vput(vp); 2432 return error; 2433 } 2434 2435 static int 2436 unp_connect_pair(struct unpcb *unp, struct unpcb *unp2) 2437 { 2438 struct socket *so = unp->unp_socket; 2439 struct socket *so2 = unp2->unp_socket; 2440 2441 ASSERT_LWKT_TOKEN_HELD(&unp_token); 2442 UNP_ASSERT_TOKEN_HELD(unp); 2443 UNP_ASSERT_TOKEN_HELD(unp2); 2444 2445 KASSERT(so->so_type == so2->so_type, 2446 ("socket type mismatch, so %d, so2 %d", so->so_type, so2->so_type)); 2447 2448 if (!UNP_ISATTACHED(unp)) 2449 return EINVAL; 2450 if (!UNP_ISATTACHED(unp2)) 2451 return ECONNREFUSED; 2452 2453 KASSERT(unp->unp_conn == NULL, ("unp is already connected")); 2454 unp->unp_conn = unp2; 2455 2456 switch (so->so_type) { 2457 case SOCK_DGRAM: 2458 LIST_INSERT_HEAD(&unp2->unp_refs, unp, unp_reflink); 2459 soisconnected(so); 2460 break; 2461 2462 case SOCK_STREAM: 2463 case SOCK_SEQPACKET: 2464 KASSERT(unp2->unp_conn == NULL, ("unp2 is already connected")); 2465 unp2->unp_conn = unp; 2466 soisconnected(so); 2467 soisconnected(so2); 2468 break; 2469 2470 default: 2471 panic("unp_connect_pair: unknown socket type %d", so->so_type); 2472 } 2473 return 0; 2474 } 2475 2476 static void 2477 unp_drop(struct unpcb *unp, int error) 2478 { 2479 struct unp_global_head *head; 2480 struct unpcb *unp2; 2481 2482 ASSERT_LWKT_TOKEN_HELD(&unp_token); 2483 UNP_ASSERT_TOKEN_HELD(unp); 2484 2485 KASSERT((unp->unp_flags & (UNP_DETACHED | UNP_DROPPED)) == 0, 2486 ("unp is dropped")); 2487 2488 /* Mark this unp as detached. */ 2489 unp_setflags(unp, UNP_DETACHED); 2490 2491 /* Remove this unp from the global unp list. */ 2492 head = unp_globalhead(unp->unp_socket->so_type); 2493 KASSERT(head->count > 0, ("invalid unp count")); 2494 TAILQ_REMOVE(&head->list, unp, unp_link); 2495 head->count--; 2496 2497 /* Disconnect all. */ 2498 unp_disconnect(unp, error); 2499 while ((unp2 = LIST_FIRST(&unp->unp_refs)) != NULL) { 2500 lwkt_getpooltoken(unp2); 2501 unp_disconnect(unp2, ECONNRESET); 2502 lwkt_relpooltoken(unp2); 2503 } 2504 unp_setflags(unp, UNP_DROPPED); 2505 2506 /* Try freeing this unp. */ 2507 unp_free(unp); 2508 } 2509 2510 static void 2511 unp_defdiscard_taskfunc(void *arg __unused, int pending __unused) 2512 { 2513 struct unp_defdiscard *d; 2514 2515 spin_lock(&unp_defdiscard_spin); 2516 while ((d = SLIST_FIRST(&unp_defdiscard_head)) != NULL) { 2517 SLIST_REMOVE_HEAD(&unp_defdiscard_head, next); 2518 spin_unlock(&unp_defdiscard_spin); 2519 2520 fdrop(d->fp); 2521 kfree(d, M_UNPCB); 2522 2523 spin_lock(&unp_defdiscard_spin); 2524 } 2525 spin_unlock(&unp_defdiscard_spin); 2526 } 2527